Fermi National Accelerator Laboratory


FERMILAB-PUB-95/191-T
July 14, 1995
(corrected October 30, 1995)




MEASURING MULTIJET STRUCTURE
OF HADRONIC ENERGY FLOW

OR

WHAT IS A JET


Fyodor V. Tkachov

Fermi National Accelerator Laboratory
P.O.Box 500, Batavia, IL 60510, USA

and

Institute for Nuclear Research of the Russian Academy of Sciences
Moscow 117312, RUSSIA


Abstract Ambiguities of jet algorithms are reinterpreted as instability wrt small variations of input.
Optimal stability occurs for observables possessing property of calorimetric continuity (C-continuity)
predetermined by kinematical structure of calorimetric detectors. The so-called C-correlators form a
 19 Jan 1996 basic class of such observables and fit naturally into QFT framework, allowing systematic theoretical
studies. A few rules generate other C-continuous observables. The resulting C-algebra correctly quanti-
fies any feature of multijet structure such as the "number of jets" and mass spectra of "multijet sub-
states". The new observables are physically equivalent to traditional ones but can be computed from fi-
nal states bypassing jet algorithms which reemerge as a tool of approximate computation of C-obser-
vables from data with all ambiguities under analytical control and an optimal recombination criterion
minimizing approximation errors. Quick preview on next page






Permanent address.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 2






Quick preview  The jet algorithms are a useful approximation tool
because computing, say, C (m) (= the ideal 3-jet
 This work gives an answer to the question "what 3j
is a jet?" The answer is a purely "kinematical" one mass spectrum) from the jet pattern of an event is
(which most experts seem to be ill-prepared for), and easier than from the raw event because there are
is based on an analysis of "kinematics" of calorimet- many fewer jets than particles. It is also clear how the
ric detectors, "kinematical" aspect of QFT observ- approximation can be improved: one should simply
C
ables, and error enhancement properties of data proc- compute the new observable (m) from jet patterns
3j
essing algorithms. The answer is as follows. for as small ycut as possible -- or directly from raw
"Jets" and "jet finding algorithms" are simply data if one has enough computer resources to do so.
tools of approximate description of data (in a con-  The notorious "ambiguities of jet definition" are
crete sense explained below). Their physical signifi- then, essentially, the inevitable approximation errors
cance does not go deeper than the leading order QCD in Eq. 0.2. They are gotten rid of by letting y
cut 0 .
predictions (which in many cases is deep enough, of  The formalism of C-algebra is a purely kinemati-
course). As any approximation, it is only useful as cal one: it cannot help one e.g. to determine the
long as one is satisfied with its precision.
 observable with which a new particle is best seen
Improving upon that approximation does not against background. But it can help one to squeeze
mean finding a "perfect jet algorithm". The notion the most out of data once such an observable is found.
of a perfect jet algorithm is physically ill-conceived For instance, C (m) is, by construction, less sensi-
because of the non-deterministic quantum nature of 3j
hadronization. tive to data errors and statistical fluctuations than
Instead, the "ideal solution" is based on the for- old (m).
3j
malism of C-algebra. The C-algebra is a family of  Although the perfect jet algorithm is a fiction, an
"well-behaved" observables together with a few tran- "optimal" one is not: it is an algorithm that mini-
slation rules that allow one to reproduce any conven- mizes the error in 0.2 for a given y
tional observable defined via jet algorithms. For in- cut for the observ-
ables such as C (m) . Criteria for such an "optimal
stance, take a conventional distribution of events with 3j
respect to, say, invariant masses of 3-jet substates, preclustering" can be found analytically in an explicit
old (m) (e.g. with a peak due to a new particle at form (see Sec. 7).
3j  The results of this work should not be interpreted
some value of mass). One can write down an observ- in the sense that "one should not count jets", nor that
able from C-algebra -- call it C (m) -- that is an
3j one should throw out the old code and start implem-
"ideal solution" which old (m) approximates in the
3j enting the ideal solution of C-algebra. What is being
following sense. One can compute the new observable suggested is that various features of the ideal solution
from the jet pattern of an event (using any reasonable should be incrementally built into the existing proce-
jet algorithm) and the result will be, by construction, dures depending on: the computer resources avail-
similar to old (m): able, the required precision of results, and the level of
3j understanding achieved. The ways to improve the
C (m) old (m). 0.1
3j jet pattern 3j conventional data processing are:
-- Regularization of (all) hard cuts (Sec.2.5; perhaps
Note that the r.h.s. cannot be computed otherwise including spline-based schemes described in Sec.15
than from the jet pattern. The new observable, on the instead of the bin-type ones).
contrary, is defined in such a way that it can also be -- One optimal preclustering (Sec.7) instead of all
computed directly from the raw event (i.e. the event conventional jet algorithms (both recombination and
prior to application of a jet algorithm) and the result cone-type).
will be as follows: -- Replacing conventional observables with their C-
C ( ) = C
m (m) + O(y ). 0.2
3j raw event 3j jet pattern cut analogues (for this one has to learn to translate phy-
sics into the language of C-algebra, which should be
(In the case of cone-type algorithms the error is O(R) possible with the examples in Secs.1013).
where R is the jet cone radius.) -- Exploiting the fully explicit and essentially simple
To emphasize: the distribution C (m) is physical-
3j analytical form of C-observables to develop more
ly equivalent to the conventional 3-jet mass distribu- flexible and precise computational schemes than
tion and can be computed without actually identifying what is possible with the conventional rather rigid
individual jets. A similar translation can be accom- algorithms.
plished for any other jet-related quantity.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 3






How to read this paper

The subject of jet-related measurements concerns both theorists and experimentalists (the latter more
than the former), and there are different obstacles that may make reading this paper difficult for differ-
ent categories of readers. One is that the philosophy behind the new theory deviates rather considerably
from the conventional attitude to jet counting. Moreover, the "philosophical" arguments are mingled
with a reasoning that uses unusual logical patterns of general topology (convergences and continuities)
that seem to be unfamiliar to a majority of physicists (perhaps because they are so fundamental that are
usually taken for granted). This could not have been avoided -- the essential uniqueness of the pre-
sented construction cannot be appreciated otherwise.
Unfamiliar functional formalism of abstract measures etc. used to describe the mathematical nature
of energy flow and C-continuous observables may also be difficult for the uninitiated. This could not
have been avoided either: The mathematical description of stability of data processing algorithms in
calorimetric measurements with respect to errors in the language of a -weak convergence in a space of
linear functionals is not an invented hypothesis; it is a straightforward formalization of what is.1
However, a data processing expert not seeking to understand the mathematical subtleties in complete
detail but willing to give the new formalism a try in a realistic problem, may concentrate on the discus-
sion of the role of continuity of observables in Sec. 2.5, and the QFT derivation of C-correlators in
Sec. 6.4. For practical purposes, one needs the definitions and notations of Sec. 4 and the formulas of
Sec. 5 (discrete variants only) while the details of the reasoning of Sec. 3 may be ignored. Then one can
turn directly to the discussion of applications in Secs. 813. The issues of numerical work with spectral
discriminators discussed in Secs. 15 and 12 cannot be skipped.
The material of Sec.7, Secs. 89, and Secs.1012 can be studied in any order.
Finally, it should be noted that the validity of the prescription to regularize one's cuts as a matter of
routine (Sec. 2.5) extends beyond the measurements in high energy physics. Similarly, Secs.12 and 15
describe a rather universal scheme based on linear splines for computation of various distributions ob-
tained from irregular "stochastic" approximations such as typically obtained from a sample of events
generated by a quantum process. The scheme is hardly more complex, but less sensitive to data errors
and statistical fluctuations, than the conventional bin-type algorithms.

Numbering and cross references

I hope the reader will find it convenient that Fig.10.24 is located between items 10.23 and 10.25 --
irrespective of what the latter are, equations or subsection headings. The universal two-level numbering
scheme (with right-aligned numbers conveniently not distracting one's attention) offers an easy way of
finding cross references (as with page number references) while remaining coupled to the logical struc-
ture of the text. Sub- (and subsub-) section headings are treated along with proposition headings rather
like labels in the text (with a liberating effect that one no longer needs to conform to pedantic conven-
tions, e.g., about whether or not a section may be split in the middle by a single subsection heading).
For that reason they do not have a separate numbering from equations, figures and tables -- it is not
clear what such a separate numbering might be good for except a convenience of uncomputerized
typographer.






1 If one needs a proof of this fact other than the arguments of the present paper one may note that the definition
of calorimetric continuity underwent a noticeable evolution [16], [17], [15] with the original version [16] rather
different in form from, but essentially equivalent to, the final one [15].


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 4






Acknowledgments

This work in its final form owes so much to a continuous intensive interaction (which I enjoyed
hugely) with the world high energy physics community that justice must be paid to those whose inputs
influenced this work:
W. T .Giele introduced me into the theory of jets in 1991 at FERMILAB when my first doubts con-
cerning jet algorithms were raised in connection with the problem of higher order calculations in the
QCD theory of jets. The review talks of S. D. Ellis impressed upon me the fundamental importance of
the problem ("What is a jet?" [6]). A.V. Radyushkin offered a crucial encouragement after hearing me
out at CEBAF in 1992. I thank Z. Kunszt and ETH for hospitality during the Workshop on New Tech-
niques for Calculating Higher Order QCD Corrections (ETH, Zrich, December 1992) -- its intensely
informative atmosphere catalyzed this work (a key formula for the jet-number discriminators material-
ized on my way home from ETH; I am indebted to a Swiss train attendant who rescued my notes).
S. Catani supplied a useful list of bibliography (any omissions are my fault though), as did R. K. Ellis
who also provided a stimulating criticism at an early stage of this work.
Very numerous and lively conversations took place with the participants -- both theorists and ex-
perimentalists -- of: a seminar of the Division of Experimental Physics of INR RAS (Moscow, 1993),
the 1993 ITEP Winter School, the VIII and IX Workshops on High Energy Physics (INP MSU, Zveni-
gorod, 1993 and 1994), the XVI and XVII Seminars on High Energy Physics (IHEP, Protvino, 1993
and 1994), a seminar of the Moscow group of the H1 collaboration (1993), a theory seminar at CERN
(1993), a theory seminar at Cornell University (1994), a seminar at NORDITA (1995), the Moscow
session of AINHEP-95 (INP MSU, 1995), a theory seminar at FERMILAB (1995).
Stimulating and/or useful were interactions with B. A. Arbuzov, Yu. Arestov, W. A. Bardeen,
Yu. Bashmakov, A.V.Batunin, E. E. Boos, F. Cuypers, L. Dudko, F. Dydak, R. K. Ellis, L. Frankfurt,
I. F.Ginzburg, B. L. Ioffe, K. Kato, A.B. Kaydalov, V. D. Khovansky, A. Klatchko, V. Kopeliovich,
M. Krawczyk, Z. Kunszt, Y. Kurihara, E. Kushnirenko, A. I. Lebedev, D. Lehner, G. P. Lepage,
L. Lnnblad, D. Nesic, Kh. S. Nirov, T. A.Osipova, D. Perret-Gallix, Referees A, B and C of Physical
Review Letters, M. A. Samuel, E. K. Shabalina, R. Shanidze, D.V. Shirkov, A. Shtannikov, T. Sjstrand,
I. K.Sobolev, N. Sotnikova, B. Straub, M. Strikman, I. I. Tkachev, A. A.Tyapkin, E. Verlinde,
B. R.Webber, S. Wolfram and S.Youssef. I thank B. B. Levtchenko for some numerical checks,
A. S. Barabash for providing information on how calorimetric detectors actually work [72], and
N. A. Sveshnikov for many discussions of the QFT aspect of the C-algebra.
Parts of this work were performed during my visits to ETH (1993); Penn State University (1994);
and NORDITA (1995) -- I thank for hospitality, respectively, Z. Kunszt; J. C. Collins and H. Grotch;
L. Brink, P. Hoyer and P. Osland. At various space-time points this work was supported in parts by
(chronologically) the International Science Foundation (the George Soros Fund; grants MP9000 and
MP9300), the U.S. Department of Energy (grant DE-FG02-90ER-40577), the Russian Fund for Basic
Research (grant 95-02-05794).
This work was completed at FERMILAB: I thank its Theory Group for their kind hospitality, and
members of the D0 all-jets top group (S. Ahn, H. Prosper, and D. A. Stewart Jr.) for a thorough discus-
sion of the C-algebra in a practical context; I owe special thanks to D. A. Stewart Jr. for explanations of
the data acquisition system of the D0 detector and many comments, and to N. Sotnikova for much help
with preparation of the paper.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 5






TABLE OF CONTENT Comparison with the L2 formalism 4.29
Correct continuous limit 4.31
Introduction and overview 1 Calorimetrically continuous (C-continuous)
The jet paradigm 1.1 observables 5
Ambiguities of jet algorithms 1.3 C-correlators: basic C-continuous observables 5.1
Purpose of this work 1.4 Linear functions of energy flows 5.2
Understanding the problem 1.5 Bilinear scalar functions of energy flows 5.8
Measurements or a modeling of dynamics? 1.6 Multilinear C-correlators 5.15
C-correlators and C-algebra 1.7 Regularity restriction from IR safety 5.19
C-continuity 1.9 Comparison with EEC observables 5.20
Expressing physics in the language of C-correlators. C-correlators and "physical information" 5.21
C-algebra 1.10 C-algebra 5.22
C-algebra and jet algorithms 1.11 Compositions of C-continuous functions 5.23
"Linguistic" restrictions of C-algebra 1.13 Filters and substates 5.25
"Ambiguities" of jet finding algorithms 2 Integration over a parameter 5.29
Jet algorithms and precision measurements 2.1 Minimization with respect to a parameter 5.31
Theoretical aspects: jet algorithms vs. QFT 2.2 Differential C-continuous observables 5.35
Instabilities of jet algorithms 2.3 C-correlators and Quantum Field Theory 6
Discontinuities and data errors 2.5 C-correlators from the point of view of QFT 6.1
A simple 1-dimensional model 2.6 An alternative derivation of C-correlators 6.4
Taking into account measurement errors 2.8 Fragmentation invariance of C-correlators 6.6
Integral error for a hard cut 2.12 C-continuity vs. IR safety 6.9
Integral error for continuous weight 2.15 A continuous and perturbatively IR safe function
Comparison and discussion 2.21 that is not C-continuous 6.13
Instability regions of jet algorithms 2.22 Aspects of theoretical calculations 6.16
Geometry of instability regions Analytical calculations of C-correlators 6.17
in the continuum of final states 2.23 Power corrections 6.18
Effects of the shape of (P) 2.26
Summary 2.30 Computing C-continuous observables from data 7
Mathematics of measurements and continuity 3 Understanding the problem 7.1
Sources of optimizations 7.2
Data processing and errors 3.1 Criterion of optimization 7.3
A mathematical subtlety 3.5
Soft particles 7.4
Data errors and convergence 3.7 Expanding in energies of soft particles 7.5
Reminder: length and real numbers 3.8 Estimating the error 7.10
Convergence in terms of resolution of measurement Errors due to missing energy 7.14
device 3.9 Missing energy for jet-number discriminators 7.15
Complex measurement installations 3.10 Preclustering 7.20
Definition of convergence 3.11 On masslessness of pseudoparticles 7.22
Convergence in terms of open neighborhoods 3.13 Estimating errors induced by recombinations 7.23
Summary 3.14 Modifications for general C-correlators 7.36
Multiparticle states, energy flows and The "optimal" preclustering prescription 7.38
calorimetric detectors 4 Comments on the definition of 7.41
ab
Comparison with the conventional algorithms 7.43
Multiparticle states and their energy flows 4.1 Preclustering n 1 7.47
Parametrization of particles' directions 4.2 Accumulation of errors 7.55
Angular distance between directions: "angular
separation" 4.5 Applications to measuring the "number of jets" 8
Calorimetric information and energy flow 4.8 "Jet counting" with C-correlators 8.1
Lorentz invariance 4.11 Understanding the problem 8.2
Detectors 4.14 Defining multijet states in terms of C-continuity 8.3
Data records 4.16 Explicit formulas for jet-number discriminators 8.4
Energy flows as "abstract measures" 4.19 Normalization 8.12
Calorimetric convergence (C-convergence). Some special values 8.17
When are two energy flows close? 4.22 Understanding the jet-number discriminators 9
C-convergence and collinear fragmentations 4.24 Qualitative behavior of J 9.1
m
Correctness of the scheme 4.27 Estimating J via a crude clustering 9.2
m


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 6






A typical picture of values 9.3 Computation of r (s ; R
N i ) for one event 12.8
Understanding the decrease at large m 9.5 Adding a contribution to r (s ; R) 12.9
N i
Simple examples 9.6 The folding trick 12.11
Two pure jets 9.7
Adding one soft particle 9.11 Advanced options 13
2 hard + 2 soft particles 9.18 Modifications for hadronic reactions,
J never see spurious jets 9.22 recombination-style 13.1
m
Width of jets 9.23 Special reference frames for DIS 13.2
Suppressing "forward jets" 13.5
J vs. jet counting 9.25
m
Counting jets with J 9.27 Modifications for hadron-hadron reactions,
m rapidity-inclusive-style 13.7
Instability of the conventional jet counting 9.28 Using rapidity: cylinder replaces unit sphere 13.8
J vs. n -jet fractions 9.29
m Defining angular separation 13.13
Applications to mass measurements. Transverse energy 13.18
Simplest spectral discriminator 10 C-correlators 13.21
1-jet spectral discriminator 10.1 Modifications of the optimal preclustering 13.23
Modifications of the jet-number discriminators 13.24
Elementary mass detector 10.2 Modifications of spectral discriminators 13.25
Definition 10.6 Jet distributions with respect to E 13.26
T
Simple properties of 10.9
1 jet (s; R) Further options for C-continuous observables 13.29
Small mass detector ( R 0) 10.10 Differential observables 13.30
Large mass detector ( R 1) 10.12 Example: using additional dynamical information
Flow of spectral weight 10.14 to enhance signal 13.32
Uniform distribution of energy over the sphere 10.15 Higher differential C-continuous observables 13.34
Spectral density is localized
within Summary and conclusions 14
0 s smax S 10.17
R tot
Zero-mass -spike for 0 < R < 1 10.18 Objections 14.1
Jets and -spikes 10.21 Higher computing resources 14.2
Events with a single jet 10.22 Ambiguities 14.3
Events with several well-separated jets 10.25 Psychological objections 14.4
From jet finding algorithms to C-algebra 14.5
-spikes corresponding to groups of jets 10.27
Spurious -spikes 10.28 Appendix. Abstract measures on [ ,
0 ]
1 15
Summary. Evolution of a -spike 10.29 Continuous functions. Constructive aspects 15.1
Realistic example: X jet + "muon" 10.30 Linear splines 15.3
Smearing and accumulation of -spikes 10.33 Measures on [ ,
0 ]
1 15.8
General spectral discriminators 11 Examples of functionals 15.9
Mass detectors and spectral discriminators 11.1 Strong assumption 15.11
Changing the spectral variable 11.4 What are the most general linear functionals
Continuous component of ( ;
s P) 11.8 on C0? 15.12
-functional components of ( ;sP) 11.11 Continuity of measures 15.14
Measures as additive functions of subsets 15.15
Multijet spectral discriminators 11.13 General structure of measures 15.16
Composite mass detectors 11.14
Example. Spectral discriminator for 2-jet Convergence of sequences of measures 15.17
substates 11.19 Quality of approximations 15.18
Qualitative behavior of 11.24 Approximate description 15.19
2 jet ( ;
s R)
Comparison with the conventional practice 11.25 Measures represented by continuous functions 15.20
-functions 15.21
Discussion 11.28 Constructing regular approximations 15.23
Practical computation of spectral converges to in the sense of measures 15.27
N
discriminators 12 r (s ) are valid observables per se 15.28
N i
The data 12.1 Regularization 15.29
The grid of masses 12.2
The grid of R 12.4 Appendix. General theory of abstract
Discretization of integrations 12.5 measures 16

General scenario. Regularization 12.7 References 17


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 7






Introduction and overview 1

The jet paradigm 1.1

The jet paradigm is the foundation of the modern high-energy physics. It is based on the experimental
evidence for hadron jets [1] and a QCD-based theoretical picture [2] of the hadronic energy flow in the
final state inheriting the shape of partonic energy flow in the underlying hard process. The association
of each jet with a hard parton [3], [4] is the qualitative basis for comparison of data with theory. Its
straightforward quantitative implementation leads to the so-called jet finding algorithms. For each final
state (event) P, such an algorithm computes what may be called its jet pattern Q -- i.e. the total num-
ber of jets in P and the 4-momentum of each jet:

jet finding algorithm
P = p , p ,K p K
{ 1 2 }
Q = Q ,Q , Q
N { 1 2 N } , 1.2
part jets


where pi are the 4-momenta of the particles in the final state while Qj are the 4-momenta of the result-
ing jets. While the event may have O(100) particles, a majority of events have only a few jets. There-
fore, the jet pattern Q has typically much fewer 4-momenta than the original event P and can be studied
much easier. Moreover, such a reduction of information brings out the physics of the process being
studied: For instance, if an unstable virtual particle decayed into quarks and gluons that hadronized into
jets, the invariant mass of the corresponding group of jets is close to that of the decayed particle.
On the use of such algorithms the current practice of measurements in high-energy physics is
founded.i

Ambiguities of jet algorithms 1.3

Unfortunately, hadron jets have a finite angular width and irregular shape. So when their angular
separation is not large enough they are hard or impossible to resolve. Then the answer (the jet pattern Q
on the r.h.s. of 1.2) depends on inessential details of jet finding algorithms. As a result, the ambiguities
of jet definition remain a subject of the ongoing discussion ("What is a jet?" [6], [8]). Moreover, as the
research moves on to physical problems with more stringent precision requirements, the ambiguities of
jet definition manifest themselves quite tangibly: Recent results of Monte Carlo modeling indicate that
the dominant uncertainty in the determination of the top quark mass at the LHC is expected to be due to
the ambiguities of jet definition [9].

Purpose of this work 1.4

The purpose of this work is to reexamine the problem of ambiguities of jet algorithms in a systematic
fashion from the point of view of first principles of physical measurements. I maintain that a clarifica-
tion of the issues of kinematics and measurements must be achieved before one will be in a position to
discuss dynamics in a completely meaningful fashion. For instance, the qualitative physical notion of
length (of, say, sticks) is known to be quantified by real numbers, and the continuity of the latter plays a
crucial role in the analysis of measurement errors.
What is the precise mathematical analogue of the qualitative physical notion of energy flow of colli-
sion events? And what is the corresponding continuity that would allow one to control effects of meas-
urement errors in a systematic fashion? (Cf. Table 4.28 below.)
The focus on the aspect of measurements and kinematics constitutes the key difference between this
work and the conventional approach that emphasizes the dynamics of jets.ii


i For an introduction to the uses of hadronic jets in experimental studies of the Standard Model see e.g. [5].
ii This fact might ignite an ideological discussion. As a preemptive measure, I've made an attempt to clarify
certain implicit assumptions behind the conventional point of view on jet counting. This resulted in


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 8






Understanding the problem 1.5

The physical problems where jets are involved can be divided into two classes:
 In the problems of exploratory type one may be satisfied with observing an effect at a minimal
sufficient level of statistical significance. This is usually the first stage in studies of any new class of
physical phenomena. For instance, in what may be called descriptive theory of hadronic jets one studies
the dynamics of jets as such (a good review mostly devoted to this sort of physics is [10]). Here one is
mostly interested in qualitative effects that occur already in the leading order of perturbation theory (cf.
the effects of QCD coherence [11]). Another example is the search for top quark [12], [13].
 Precision measurements comprise quantitative studies of the Standard Model (for a review see
e.g. [5]) such as precision measurements of 2
S (Q ) , the mass and width of the top quark etc. In such
problems jets are only an auxiliary intermediate notion.

Clearly, the two classes of problems differ with respect to the requirements on the data processing al-
gorithms employed. In the first case, one would like to have simple and fast algorithms that allow one to
represent the corresponding qualitative physical features in a most direct and visual manner even, per-
haps, at the expense of numerical precision.i On the other hand, in applications of the precision meas-
urement class the overriding concern is the highest precision attainable for a given event sample.
Note that even in the exploratory problems one may encounter a situation with a low signal/back-
ground ratio so that even to establish existence of an effect one may need to squeeze every bit of physi-
cal information out of data, which may be impossible with crude methods.

Measurements or a modeling of dynamics? 1.6

One may notice that jet algorithms were invented in the context of the problems of exploratory typeii
and their systematic use for data processing in the precision measurement applications may not be ac-
cepted uncritically.
Indeed, theoretical research into jet definition algorithms has traditionally been performed in the con-
text of studies of hadronization. The latter is a non-perturbative phenomenon, is ill-understood, and pre-
sents a fascinating field of study. As a result, the problem of adequate numerical description of mutlijet
structure has come to be regarded as a "physics issue" that concerns the "underlying jet dynamics" ra-
ther than "kinematics" and "geometry"iii, and that instead "the main point is that one would like to iden-
tify a class of <...> observables that are in closer correspondence with the underlying jet dynamics".iv
The importance of understanding hadronization and jet dynamics is indisputable. But in the context of
precision measurements such a point of view makes hardly more sense than requiring that rulers and
clocks reflect the "underlying dynamics" of gravity when one studies the laws governing trajectories of
falling bodies. The rulers and clocks do reflect properties of the external world -- but only in the kine-
matical aspect: first and foremost, we want our rulers to be straight and clocks, precise.
Of course, there is a valid point in the "dynamical" argument, too. Indeed, some a priori information
on the underlying dynamics helps one to perform measurements where they are most informative (e.g.

"philosophical" arguments that are somewhat lengthy and which a pragmatist may not approve of. The subject
of hadron jets, however, is fraught with prejudice (as I discovered in many conversations with both theorists
and experimentalists), and it is hard to discuss unarticulated attitudes in brief. I may add that all "philosophi-
cal" arguments of this paper are more or less direct answers to the questions raised and attitudes exhibited, by
the many physicists with whom I discussed these matters.
i For instance, one may need to find -- very much by trial and error -- an observable with respect to which to
perform cuts so as to suppress background [12], [13].
ii Cf. [2] as well as the fact that the recombination algorithms of a modern type were invented in the context of
Monte Carlo event generators [14].
iii which are, no doubt, lowly subjects that have nothing to do with The Profound Physics.
iv The quotations are from the report of Referee A on the first version of [15].


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 9






choosing objects of convenient weight to drop from a tower of convenient height and inclination; or by
restricting a search for a new particle to events satisfying certain cuts).
But the kinematical requirement of correctness of measurements (including the data processing stage)
is a matter of firstest concern when the precision of results is at stake. Therefore, it makes sense to ex-
amine the problem of studying hadronic energy flow as a problem of measurement, and to try -- instead
of attempting to solve the mathematically ill-posed problem of inverting hadronization -- to find
"straight rulers" for measuring the multijet structure of multihadron final states.
One may further notice that defining individual jets is only an auxiliary intermediate step in any pre-
cision measurement application. A good deal of its attractiveness is due to simplicity of the jet pattern
one has to handle at the final stage (when the parameters one measures are actually obtained) and the
simplicity of the conventional jet algorithms. Simplicity and speed, however, are not the matters of pri-
mary concern in precision measurements -- the precision is. It makes sense, therefore, to ask which al-
gorithms may be best for that particular purpose. Having found such algorithms, one can then proceed
to optimize them while being fully aware of all the tradeoffs involved.
Then the crucial first step is to subject to a scrutiny the key implicit assumption already present in the
question "what is a jet?" -- the assumption that the data processing of the form 1.2 is the only way to
quantify the multijet structure. The theory of calorimetric observables for measuring multijet structure
outlined in [15], [16], [17], [18], [19], [20], [21] and systematically developed in this paper proves the
assumption to be wrong.

C-correlators and C-algebra 1.7

I argue that the fundamental means (the "straight rulers") to measure the multijet structure are not the
jet algorithms 1.2 but the so-called C-correlatorsi -- i.e. the observables of the following general form:

F P = K K
m ( ) E E f ( $p , , $p ), 1.8
i i
K i i m i i
1 m 1 m 1 m

where Ei and $i
p are the energy and direction of the i -th particle (or calorimeter cell), each summation

runs over all particles in the event (all cells that lit up), m is any positive integer, and fm is any suffi-
ciently smooth symmetric function of its m arguments.
Special cases of such correlators have been well-known for a long time, e.g. the special sequences of
observables studied in [22], [23]. The energy-energy correlators of [24] are also closely related to the
C-correlators but correspond to discontinuous fm of a special form. However, it has never been realized
that the following two properties of the functions 1.8 ensure them a very special role in studies of multi-
jet structure:
First, the observables 1.8 have a form of multiparticle correlators that is natural from the point of
view of Quantum Field Theory. This is further discussed in Secs. 2.2 and 6 (see also [54]).
Second, they possess special stability properties with respect to data errors -- a stability that can be
briefly described as follows.

C-continuity 1.9

Consider any observable 1.8, any two events and any calorimetric detector installation. If the two
events P and P (whatever their numbers of particles etc.) are seen by the detector installation as the
same, i.e. P P within data errors, then the corresponding two values of the observable are guaran-
teed to stay close, F P
( ) F P
( )
m m , with the difference depending on the observable and vanishing to-
gether with data errors. Note that with jet algorithms this is not always the case: two such events may
yield different numbers of jets -- and a difference of two unequal integer numbers cannot be made less
than 1.


i C from "calorimetric". The prefix will be widely used below.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 10






The described stability is a form of continuity (we call it C-continuity; its precise description is given
below in Secs. 45). Its importance is due to the fact that, given the same measurement errors in the in-
put data the error of an observable that is continuous in the above sense is, in general, less than the error
for the corresponding cut-type observables (Secs. 2.5 and 3; see also [41]).
It is also important to understand that the role of C-continuity is not limited to the issues of data er-
rors as such but concerns other types of small variations as well, such as higher order corrections, com-
putational approximations, minor programming variations, etc.
The above two properties of C-correlators allow one to say that the collection of values of all C-cor-
relators constitutes the entire physical information about an event. Once one has understood that, one
realizes that any aspect of multijet structure that can be meaningfully quantified and measured, allows
an expression in the language of C-correlators.

Expressing physics in the language of C-correlators. C-algebra 1.10

The next question is, how to translate the qualitative phenomenon of jets into numbers using only C-
correlators? The answer is not obvious because their energy dependence is rigidly fixed (cf. 1.8). But
once the central role of C-correlators is understood (in a sense, C-correlators is all one can "see" about
events using calorimetric detectors from the point of view of QFT), it becomes clear that such a transla-
tion must be possible.
-





The translation is based on a few rules that allow one to organize values of various C-correlators into
new observables that (i) inherit the property of C-continuity and (ii) describe various qualitative features
of multijet structure one happens to be interested in. The expressive power of the resulting C-algebra of
observables proves to be sufficient for the purposes of precision measurement applications.
There are two main applications for jet finding algorithms that can be classified as precision meas-
urements. Both usually involve a classification of the events with respect to the number of jets ( Njets
in 1.2).
The first application may be described as "jet counting". The idea here is to count the fraction of
events for each number of jets in the final state. One then obtains what is called n-jet fractions (usually
denoted as n jets ). They provide a more detailed information than the total cross section and prove to
be greatly useful both for studying QCD and the Standard Modeli. This is due to their simplicity and
their more or less direct connection with hard subprocesses (each additional jet is associated with an
additional radiated parton). For instance, the 3-jet fraction in e+e- annihilation in the lowest order of
perturbation theory receives contributions from radiation of a gluon from the quark-antiquark pair.
Therefore, it is directly sensitive to the coupling constant S and can be used to extract the latter from
experimental data. It turns out that the physical information about the "number of jets" in the final state
can be equivalently represented in terms of a special sequence of C-correlators 1.8 (the so-called jet-
number discriminators Jm introduced in [16], [15]; see also Sec.8 below).
The other class of applications involves studies of particles that decay into jets, which requires inves-
tigation of mass spectra of multijet substates. Here one normally selects events with a given number of
jets corresponding to the partonic subprocess where the particle is expected to be seen and then studies
invariant masses of groups of jets/partons in a more or less exhaustive fashion (i.e. considering all
combinations of two jets if the particle decays into two partons). An example is the top search in the
purely hadronic channel [25] where one would have to select 6-jet events and then identify 3-jet sub-
states that may be decay products of the top quark. The observables that can be used in such a context
are the so-called spectral discriminators ([18], [21] and Sec. 10 below). These are somewhat more
complex than the C-correlators 1.8 but use the latter as elementary building blocks. Spectral discrimina-
tors contain more information about each event than multijet mass distributions obtained via jet algo-


i For instance, a major part of the proceedings [7] is devoted to such applications.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 11






rithms, and allow one, in principle, to measure the signal from a virtual particle that decays into jets
even in situations where the jets are too wide to be resolved unambiguously by the conventional algo-
rithms. If the higher computer requirements of the new formalism prove not to be prohibitive, this may
allow one to bypass the limitations of the conventional jet algorithms (cf. Sec.1.3).

C-algebra and jet algorithms 1.11

It is remarkable that the relation between the two approaches to jet measurements -- the one based on
jet algorithms 1.2 and the other on the C-correlators 1.8 -- is precisely that between an approximation
trick and an exact answer. Indeed, it is quite natural to regard the jet pattern Q on the r.h.s. of 1.2 as an
approximate representation of the final state P on the l.h.s. But what is an exact interpretation of the
adjective "approximate"? The answer is, in the sense that the values of any C-correlator 1.8 for the
r.h.s. and the l.h.s. are numerically close, and what is known as the jet resolution parameter ycut in the
context of jet algorithms is simply a parameter that controls the approximation error:

F P
( ) F (Q +
m m ) O(y )
cut . 1.12

If one requires that the approximation error be minimal one obtains a criterion for an "optimal" jet al-
gorithm ([18], [19], [20] and Sec. 7 below). Such a criterion turns out to combine features of the jet
definitions currently in use. Then the ambiguities of the conventional jet algorithms are naturally inter-
preted as approximation errors which it becomes possible to control (and, in principle, eliminate) if one
systematically employs the C-correlators 1.8 (and other C-continuous observables correctly constructed
from C-correlators) to quantify the multijet structure.

"Linguistic" restrictions of C-algebra 1.13

It should be emphasized that there are questions about multijet structure of an event that can not be
asked in the precise language of C-algebra. By construction, the linguistic restrictions of this sort reflect
the intrinsic limitations of the measurement procedures based on the use of finite precision calorimetric
detectors. For instance, it is incorrect to ask how many jets a given event has, and what their 4-momenta
are. Nevertheless, as was already mentioned, it is possible to define observables that are physical equi-
valents of the conventional n-jet fractions and of the mass spectra of "multijet substates". Moreover,
since jet algorithms retain the role of an approximation trick for computation of C-continuous observ-
ables, the concepts of "jets" and their "number" retain their value within the precision of such an ap-
proximation. (Recall e.g. quantum mechanics where it is, strictly speaking, forbidden to ask what was
the exact trajectory of an electron. However, if the electron's motion is quasiclassical, the question be-
comes meaningful to the same degree to which the quasiclassical approximation is acceptable.)

Plan of the paper 1.14

The theory of jet measurements described in this paper consists -- like any other theory -- of four
parts: philosophy (a critique of jet algorithms and a discussion of a special role of continuity of observ-
ables in precision measurements, Secs. 13); formalism (elucidation of the precise mathematical nature
of energy flow and the construction of C-algebra, Secs. 46); computational methods (esp. the optimal
jet clustering criterion, Sec. 7; also Secs. 12, 15 and 6.16); applications (a description of various quan-
tities of physical interest in the language of C-algebra, Secs. 813).
In Sec. 2 the ambiguities of the conventional jet algorithms are subjected to a scrutiny. It is then ar-
gued that instead of ambiguities one should talk about instabilities of the corresponding numerical pro-
cedures, and "instability" is, essentially, just another term for "discontinuity". We consider a simple
one-dimensional example and demonstrate how a use of continuous weights instead of hard cuts sup-
presses statistical data errors. (This is a key ideological point of the present work.) Then we consider the
geometry of discontinuities and the associated instability regions in the continuum of multiparticle final


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 12






states. We investigate the potential numerical effects of instabilities taking into account specifics of
QCD. A conclusion is that instabilities are a fundamental feature of any algorithm that attempts to de-
scribe a final state in terms of an integer number of jets.
Once the role of continuity of observables for minimization of errors is understood, it should be
clarified which particular form of continuity is the right one. The point here is that in infinitely dimen-
sional spaces (such as the continuum of events/final states/energy flows), many radically different con-
tinuities are possible. This issue is addressed in Sec. 3. It is pointed out that a prerequisite to defining
continuity of observables is a definition of a convergence (topology) in the space of final states. We
show that the appropriate convergence is determined uniquely by the structure of data errors of a par-
ticular class of measurement installations.
In Sec. 4 the general scenario is specialized to the case of multimodule calorimetric detectors, and the
so-called C-convergence is described. This completes a precise mathematical interpretation for the phy-
sical notion of "energy flow" which is central in the discussion of jet-related measurements. In Sec. 5 the
observables possessing the corresponding property of C-continuity are described. They are shown to
constitute a sort of algebra (the C-algebra). The basisi of the latter consists of C-correlators. Other
physically useful C-continuous observables can be constructed using a few rules. In Sec. 6 the C-cor-
relators are discussed from the point of view of Quantum Field Theory. It is argued that they fit natu-
rally into the general framework of QFT. In particular, an alternative derivation of C-correlators is pre-
sented. The derivation starts with a QFT-compatible correlator-type observable and makes use of the
condition of fragmentation invariance extended to take into account finite energy and angular resolution
of calorimeters in the spirit of the preceding discussion of the role of continuity. Prospects for theoreti-
cal studies and calculations of C-correlators are briefly discussed.
Since C-correlators play such a central role in the new theory, Sec. 7 addresses the issue of their op-
timal computation from data. First of all, expansions in masses of soft particles are possible -- an op-
tion that has no analog in the conventional data processing (except in the form of the problem of the
missing energy, which is discussed in Sec. 7.14). Furthermore, a stability with respect to almost collin-
ear fragmentations allows one to employ an approximation trick ("preclustering") that is similar to con-
ventional jet algorithms (cf. the discussion in Sec. 1.11). An optimal preclustering criterion is derived in
a purely analytical fashion from the well-defined requirement of minimizing approximation errors. The
new criterion is shown to possess features of the popular jet definitions but is more general in that it al-
lows a simultaneous clustering of more than 2 particles into one.
Secs. 813 are devoted to phenomenological applications, and we show how to express physics in the
language of C-continuous observables (C-correlators and their derivatives). In Sec. 8 a special sequence
of C-correlators is derived (the so-called jet-number discriminators Jm ) that are physically equivalent to
the conventional n-jet fractions. The properties of the jet-number discriminators are considered in Sec. 9.
Sec. 10 introduces and studies properties of a spectral discriminator. The latter are C-continuous ob-
servables for measuring mass spectra of multijet substates without identifying individual jets, which
allows one to avoid instabilities of the conventional algorithms. It is demonstrated that a presence of
isolated clusters of particles results in the so-called -spikes -- -functional contributions to spectral
discriminators. -spikes are a universal feature of spectral discriminators and, after averaging over all
final states, result in enhancements that signal presence of new particles. Sec. 11 defines more complex
spectral discriminators that allow one to better focus on multijet substates. Sec. 12 summarizes an al-
gorithm of computation of spectral discriminators from data. Sec. 13 describes the options available in
the formalism to enhance signal/background ratio by taking into account a priori dynamical information.
It also describes modifications needed for the case of hadronic collisions.
Sec. 14 contains a summary and concluding remarks. The two appendices (Secs. 15 and 16) provide
background information on abstract measures.


i A non-linear one.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 13






"Ambiguities" of jet finding algorithms 2

The jet algorithms currently in use come in two flavors: cone-type and recombination (alternatively
called, respectively, clustering [10] and successive combination [26]). The cone-type algorithms define
a jet as a group of particles that belong to a cone of certain angular radius [2], [27]. They involve a free
parameter (the radius of jets R) and an optimization [2] or an iterative procedure [27] to determine jet
axes, as well as a rather arbitrary prescription to deal with overlapping cones. In the case of recombina-
tion algorithmsi, the underlying idea was, apparently, to invert the hadronization. They also involve a
free parameter (the so-called jet resolution parameter y < ycut <<
cut , 0 1) and proceed in an iterative
fashion, taking at each step two particles and deciding (using a criterion that involves ycut as a cut)
whether or not to combine them into a jet.
For the purposes of our discussion, the details of jet algorithms are less important than the structure
of the results they yield. Such an algorithm is applied to a multiparticle state and computes its jet pat-
tern (cf. 1.2). Note that the jet pattern contains an implicit dependence on the control parameter of the
algorithm ( ycut or R). There is no criterion to fix the control parameter; the conventional practice is to
perform computations for an interval of its values. Note that the control parameter has a physical sig-
nificance in the conventional schemes that becomes manifest when one studies mass spectra of substates
consisting of a fixed number of jets. This is because the number of such substates and their invariant
masses can be completely different for different values of the control parameter.

Jet finding algorithms and precision measurements 2.1

Once one has realized what were the context in which the conventional jet algorithms were invented
and the implicit assumptions behind them, it is no longer self-evident that they should be an adequate
tool for the needs of precision measurement applications.

Theoretical aspects: jet algorithms vs. QFT 2.2

Precision measurements imply a comparison with high-quality theoretical calculations. For system-
atic theoretical calculations to be possible (including various kinds of corrections), they should be per-
formed within the formalism of Quantum Field Theory. The first objection to jet finding algorithms is
that they do not fit well into the QFT framework.ii
A practical manifestation of this circumstance is that jet algorithms are hard to study theoretically:
For instance, analytical calculations of theoretical predictions are impossible even in the simplest case
of the cross section for e+e- 3 jets in the leading order of perturbation theory, which is to be con-
trasted with the purely analytical next-next-to-leading order calculations of the total cross section [33]
made possible by the algorithms [34], [35], [36] that exploit the structure of multiloop Feynman dia-
grams -- the structure predetermined by QFT -- to the fullest extent. Moreover, a theoretical study of
such algorithms seems to be rather involved (Sudakov resummations etc.; cf. [31]). Also, it is not clear
how to approach the issue of power corrections that have been argued to be numerically important in jet
physics [37].
On the other hand, one can notice that the ensemble of multihadron final states is a system with a



i Luclus [14] and its derivatives JADE [26], Durham [27] and Geneva [28]; for a review and comparison see
[28], [29].
ii This argument is sometimes underrated. Recall, however, that QFT is a very tightly knit construction; it
combines quantum mechanics, special relativity, and the experimental fact that particle interactions occur via
exchange of quanta; and it does so in a practically unique way -- which is expressed most clearly in the con-
struction of perturbative QFT in [30]. It thus summarizes a huge body of experimental knowledge about the
Universe. Therefore, such a "purely theoretical" objection may not be dismissed lightly.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 14






varying number of particles. The appropriate theoretical language to describe such systems is in terms
of multiparticle correlators (see any systematic textbook on QFT, e.g. [38], or statistical mechanics, e.g.
[39]). Next one notices that "jettiness" is but a special sort of geometrical correlations between parti-
cles' momenta. Then one is driven towards a logical conclusion that it should be possible to describe the
multijet structure in terms of multiparticle correlators, which would provide a direct connection to the
formalism of QFT. We will see that such a description is indeed possible.

Instabilities of jet algorithms 2.3

Let us now turn to the key idea behind jet algorithms, which is to reconstruct the pattern of underly-
ing hard partons' momenta by, in effect, inverting the hadronization. The first observation is that the
evolution of a hard parton state into hadrons is neither classical nor deterministic. Its inversion, there-
fore, is bound to be a mathematically ill-posed problem. This means that such an inversion may not be
always stable with respect to small changes of input data.i
Instability of an algorithm means that there are large changes in the output when the input is changed
a little. The simplest reason for that is a discontinuity of the algorithm with respect to input data. For
instance, the number of jets (whatever algorithm is used to define it) is an integer-valued function of
events, and such a function cannot be continuous in any non-pathological sense on a connected contin-
uum, i.e. a continuum in which any two points can be connected by a continuous curve (as is the case
with multiparticle final states where any such state can be continuously deformed into any other). In
such a situation the maximal output error due to input errors is of order of the discontinuity -- irre-
spective of the size of the input errors (cf. Fig. 2.11 below). Although the size of the region in the con-
tinuum of final states where the instability occurs does get smaller as the errors decrease so that the cu-
mulative integral error also decreases, nevertheless the numerical estimates of Sec. 2.5 demonstrate that
a well-behaved continuous algorithm may have a considerable advantage over (and is never worse than)
discontinuous algorithms in this respect.
Another form of instability of jet algorithms -- also connected with their discontinuity -- affects jets'
4-momenta. Recall that masses of particles that decay into jets are reconstructed from the invariant
masses of the corresponding multijet substates. Now a jet algorithm enforces a representation of the fi-
nal state by a few jets' 4-momenta in spite of the jets' non-zero angular width and overlaps. In the latter
case different jet algorithms will never assign particles to jets in exactly the same way, and the resulting
jet pattern will be different. Accordingly, invariant masses of, say, pairs of jets will be measured with an
uncertainty due to the ambiguity of jet definition. This phenomenon is behind the fact already mentioned
that the error of the top mass determination at the LHC is expected to be dominated by uncertainties due
to ambiguities of jet algorithms [9].
The above effect takes place even in the absence of data errors. Moreover, since (non-negligible) data
errors are always present, they would influence the results of different algorithms in a different way
(sometimes causing, say, a recombination algorithm to combine a pair of particles when it would not do
so in the absence of data errors, or vice versa), and the resulting ambiguity increases. In fact, there takes
place an enhancement of data errors.
To understand this point recall that some sort of optimization (maximization, minimization) proce-
dure is explicit in some algorithms (e.g. the scheme described in [2]), and the iterative procedures of the
popular algorithms of both recombination and cone type, may also be formally thought of as implement-



i A non-iterative -- and presumably more stable -- version of recombination algorithms is described in [38].
However, its clustering criterion is rather ad hoc, and it also does not eliminate the instability with respect to
jets' 4-momenta that leads to ambiguities in mass measurements. It should best be considered as a shell in
which various clustering criteria can be used (cf. the analytical criterion for n 1 clustering of the "optimal"
algorithm described in Sec. 7).


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 15






ing some such implicit optimization. Therefore, consider the following situation as a toy modeli. Sup-
pose one deals with a function f (x) of a real argument x, represented by a collection of pairs
f (x ), x
i i both elements of which are known with errors. Suppose a theory predicts that the function is
a -function smeared by unknown small corrections, but one nevertheless expects that the function re-
tains a form of a peak, and the shape of the function is controlled by a parameter A which one wishes to
determine (this is analogous to determining some Standard Model parameter from hadronic final states).
Then one may attempt to extract A by finding the position of the maximum of the function.
Further suppose that one does so in a straightforward manner -- e.g. by choosing one value of the
argument that gives a maximal function value. (This is similar to trying to represent a final state by a
finite number of jets via some iterative or optimization procedure.) Then the resulting error is deter-
mined not only by data errors but is also enhanced by the width of the peak near its top within the mar-
gin of data errors as shown in the following figure:

f (x)





x 2.4

where the crosses are the measured points f (x ), x
i i with error bars and the horizontal double arrow
represents the interval of uncertainty in the position of the maximum -- an enhanced error inevitably in-
herited by the resulting value of A.
It is clear that the uncertainty is a result of one's using a too crude representation for f (x, A) with a

single -peak. Now suppose one performs some more calculations and computes not just one number
(the position of the peak) but the entire shape, and fits the theoretical curve against experimental data.
Then the above enhancement of errors would not occur because the shape would be held better in place
by the slopes. Correspondingly, the determination of A would be more precise.
The above model illustrates the mechanism of how the instabilities of jet algorithms are avoided when
one studies mass spectra of multijet substates using spectral discriminators of Secs. 1011.

Discontinuities and data errors 2.5

Let us study numerical effects of discontinuities on the integral error of the result. The mathematical
mechanism considered here is very general and not specific to jet physics. For clarity, we consider a
simple example with a 1-dimensional continuum of final states but the resulting estimate remains valid
in a general situation.

A simple 1-dimensional model 2.6

One deals with a continuum of final states P. We take each P to be a point from the interval [0,1].
(P can be regarded as a parameter with respect to which a cut is being imposed on a sample of events.
A more realistic description of hadronic final states is given in Sec. 4.) The events are generated accord-
ing to some probability distribution (P) (determined by the S-matrix). Its form is controlled by the pa-
rameters one usually wants to extract from experimental data, while various physical phenomena mani-
fest themselves through its qualitative behavior (bumps etc.).
There are two typical examples of such phenomena: (i) Virtual production of a particle which is


i motivated by the comparison in Sec.11.25 of our spectral discriminators with the conventional procedures.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 16






manifested via an enhancement of the probability distribution (P) , say, near P ~ 1 but not near P ~ 0.
(ii) A dynamical mechanism of the underlying theory that, say, suppresses the production of events near
P ~ 1. (For instance, P ~ 0 could correspond to 2-jet events while the region P ~ 1, to 3-jet events.) In
either case, the events P ~ 1 correspond to a physical "feature" one wishes to quantify. The conventional
way of describing such a feature is in terms of a "region" in phase space. So one introduces a cut in a
more or less arbitrary fashion (say, P = 1 ) to describe the region, and counts the fraction of events that
2
fall within the region defined by the cut. But the position of the cut is ill-defined in either case -- in the
first example, because the virtual particle has a finite width; in the second example, because any 3-jet
event can be continuously deformed into a 2-jet one and there is no unambiguous criterion to separate 2-
jet events from the 3-jet ones. It is clear, therefore, that a hard cut does not correspond to any physical
reality -- a fact made manifest by the ambiguity of its placement.
We have come to an important point. The above argument offers a physical reason for why a conti-
nuous weight varying between 0 at P = 0 and 1 at P = 1 should be a better tool to quantify that feature
than a hard cut. Such a weight can be thought of as representing the degree to which a final state re-
sembles those that exhibit the feature most clearlyi -- the "weight" of the feature carried by the final
state.ii If one computes and adds up the values of the weight for all events (normalizing by the total
number of events) then the result represents the integral "weight" of the physical feature one studies in a
given sample of events.
Of course there is an ambiguity of choosing a weight. But it is often possible to invoke additional
considerations to reduce the ambiguity to a minimum. Such considerations can be theoretical -- cf. the
construction of jet-number discriminators in Sec. 8 where the weight is built of "natural" scalar products
analogous to those from which matrix elements consist. Or such considerations could be empirical --
one could e.g. choose a linear "regularization" of the weight to minimize effects of data errors etc. (cf.
the analysis of errors below and Sec. 2.21; also cf. the fact that control parameters of jet algorithms are
chosen to minimize distortions of mass spectra due to hadronization).
Let us return to the example. A quantum observable yields a number for a given reaction. It is defined
by specifying a function that returns a number F P
( ) for each final state P. Then the value of the ob-
servable on the statistical ensemble of final states is the mean value

F = P P F P
z d ( ) ( ). 2.7

Choosing F P
( ) to take values 1 and 0 within and outside the chosen region of phase space is equivalent
to counting the fraction of events that fall into the region described by a hard cut.
The case of a continuous weight discussed above corresponds to F P
( ) that is continuous, bounded
(which assumption does not incur much loss of physical generality), and is such that 0 F P
( ) 1
(which can be always achieved by a normalization).
From the point of view of numerical mathematics (recall e.g. various schemes of numerical integra-
tion), it is perfectly obvious that different properties of continuity should result in a different numerical
sensitivity of the two schemes to errors and approximations.

Taking into account measurement errors 2.8

In practice, the ideal formula 2.7 is distorted in several ways (e.g. the integration is replaced with a
summation etc.). But here we are interested only in the data errors. The latter can be taken into account
~
as follows. For each ideal final state P, the measurement device "sees" another final state P distorted



i or those from which it differs most.
ii This interpretation was influenced by a probabilistic interpretation of the jet-number discriminators suggested
by F. Dydak.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 17






~ ~
by data errors. The state P is seen with the probability h ( ,
P P), where describes the error intervals~
and is determined by the detector hardware. The same data processing algorithm F is then applied to P ,
and one obtains the following result instead of 2.7:

~ ~ ~ ~
F = P
d P
d P
( ) h P
( ,P) F P
( )
P
d P
( ) F P
zz z ( ) , 2.9
exp

where
~ ~ ~ ~
F (P) = P h ( ,
P P)  (
F P)
zd . 2.10
It is natural to assume that the measurement errors are distributed continuously, then Eq. 2.10 defines a
function that is continuous everywhere -- even if F has discontinuities. Fig. 2.11 illustrates the differ-
~
ence between how F (P)
fluctuates around F P
( ) in the cases when the latter represents a hard cut and a
continuous weight.

w w
hard reg r

enhanced /r
error

P P
2.11

One sees that however small , the fluctuations of the values of F around a discontinuity are of the same
order as the discontinuity itself. Therefore, one can talk about an instability region of the observable
with respect to data errors around the points where the discontinuity occurs. In such instability regions
small data errors cause the values of the observable to fluctuate within an interval that does not vanish
as 0 .
The qualitative difference of the two pictures 2.11 translates into a difference between how the meas-
urement errors from many events accumulate when the integration is performed. The key point here is
that measurement errors for different events are independent, so it is the corresponding 2 that are
added rather than (should the latter be true, the discontinuity would have played no role). As a result,
the integral statistical error is quite different in the two cases.
(It may be pointed out here that the notion of "continuity" is unambiguous only in our simple one-
dimensional example. Defining an appropriate continuity in the realistic case of multihadron final states
is less straightforward; cf. Sec. 3.10.)

Integral error for a hard cut 2.12

Suppose one decides to count the number of the events above the cut. The exact result is given by

1dP P
z ( )whard P(), 2.13
0

where the weight whard (P) corresponds to the hard cut and is 1 or 0 depending on whether P is above
the cut or not.
We neglect statistical fluctuations of the density of generated events but assume that their positions
within [ ,
0 ]
1 are measured with an error of order O(). For simplicity suppose (P) ~ const , so that the
generated events are distributed more or less uniformly. Then the measurement errors for events that oc-
cur close to the cut (their fraction is O()) will induce fluctuations of order O( )
1 for the value of
whard (P) because of the discontinuity. Then the induced variance for 2.13 is

2 = O( 2
)  O 1
( ) = O(). 2.14
hard


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 18






Integral error for continuous weight 2.15

Now suppose one quantifies the same physical feature using a continuous weight (= regularized cut),
e.g. one that linearly interpolates between 0 and 1 within a subinterval of length r:

,
1 for P
R -1/ 2 > r / ,
2

w r
(P) = S| -1 + 1 P, for - r / 2 P - 1/ 2 r
reg / 2, 2.16
2 r
T| 0, for P - 1/ 2 < -r / .
2

Then one computes

1dP P
z ( )wreg P(). 2.17
0

We take r > ; otherwise the situation is equivalent to the hard cut.
The integral error for 2.17 is induced by the events whose weight is computed with errors due to er-
rors in their position on the horizontal axis. The fraction of such events is O(r) . The variance of the
value of wreg for each such event is
O() 1 2 2
O
 = e j. 2.18
r r2

Since measurement errors for different events are independent, the total variance for 2.17 is estimated
by Eq. 2.18 times the fraction of such events which is O(r) , i.e.

2 2
= e j
reg O . 2.19
r

For r = O() , this degenerates into 2.14, as expected.
The net effect is that the resulting error interval in the regularized case is suppressed as compared
with the hard cut case by

F
hard = r
O
HG IKJ . 2.20
reg


Comparison and discussion 2.21

(i) From 2.20 it is clear that if the ambiguity in defining the physical feature is larger than measure-
ment errors then it is generally advantageous to quantify that feature using a continuous weight instead
of a hard cut. Of course, both measurements and theoretical calculations should be done with the same
weights.

(ii) Note that one may choose the position of the cut so as to optimize the signal/background ratio.
With a regularized cut of the above type, one has an extra parameter with respect to which to optimize.
In the worst case the optimum would be reached for the hard cut (r = 0), but it is clear that such situa-
tions are exceptional. This means that with a regularized cut one is never at a disadvantage as compared
with a hard cut as far as the resulting errors are concerned. The simplest prescription is to make the
weight linearly interpolate between 0 and 1 over the interval of uncertainty of cut's placement; the end
points of such an interval presumably correspond to some physical reasons why one would not like to
place the cut above and below certain points. The exact size of the interval should not usually be very
important but it should not be less than the size of errors.

(iii) Strictly speaking, the above suppression of errors is due to not just simple continuity but a some-
what stronger regularity. In our example it was existence of a bounded (even if discontinuous) first de-


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 19






rivative. A Hlder's condition would also be sufficient as well as some other types of regularity. In
practice such a stronger regularity is usually ensured automatically (if one does not aim to construct a
counterexample). On the other hand, in more complex cases there are many continuities possible
(because the space of final states is infinitely dimensional), each one determining the corresponding dif-
ferentiability etc. The real difficulty is to choose the correct form of continuity. This is why in subse-
quent sections we will be talking only about continuity without mentioning differentiability etc.

(iv) It is easy to see that the suppression factor 2.20 for integral errors that is induced by a regulariza-
tion remains operative in a general case (not necessarily one-dimensional). It is sufficient to interpret
as the fraction of events that fall within the instability region of the hard cut and r, as the fraction of
events that fall within the interpolation region of the regularized cut provided the variation of the weight
in the region is more or less uniform.

(v) Take into account a small additional computational cost of implementing a regularized cut. Then
it may be suggested that one should regularize one's cuts as a matter of routine -- unless the compu-
tational overhead proves to be prohibitive or other sources of errors clearly dominate. This conclusion
remains valid irrespective of the physical nature of the problem.

(vi) From the standard formulas of the theory of Monte Carlo integration, one can see that regularizing
the cut also reduces the variance of the purely statistical error (assuming the regularization is such that
the mean value remains roughly the same). This point is discussed in more detail in [41].

Instability regions of jet algorithms 2.22

The above toy model ignored two effects: the geometry of cuts in a multidimensional case and the
shape of the probability distribution (P) . The examples presented below -- although not constituting a
proof -- indicate that in the case of QCD the two effects tend to conspire in an unfavorable fashion.
We consider the two aspects in turn.

Geometry of instability regions in the continuum of final states 2.23

The geometry of the cut plays a role in the enhancement of errors because the more compact the
boundary corresponding to the cut, the less its volume (for a fixed magnitude of data errors), and the
less the error due to instabilities. Unfortunately, the situation with the boundaries between n-jet regions
for different n seems to be quite opposite. A few examples will illustrate this point.
A jet finding algorithm splits the continuum of final states into "n-jet regions":

1-jet region


2-jet region


3-jet region 2.24

Consider a 1-jet state P1 that consists of one particle with the total energy E going in the direction
= 0. Also consider a continuous family of final states P,n, each consisting of n particles carrying
equal shares of the total energy E and going in the directions described by
= ( - )
1 2 , =
i i for all
n i
i . As varies, P,n describes a continuous curve in the continuum { }
P .

Let n = 2. For = 0 , P,2 is equivalent to the 1-jet state P1. For = one has two pure jets. Ob-
2
viously, there is an intermediate value = when the curve P,n crosses the boundary between 1-
and 2-jet regions. Because of measurement errors there is an interval of around where the jet al-


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 20






gorithm cannot count jets stably -- a region of instability that separates 1- and 2-jet regions (cf.
Fig. 2.24). The thickness of this region is determined by experimental errors and the jet algorithm used.
Let n = 3 . For = 0 , P,3 is again equivalent to the 1-jet state P1 while for = one now has
2
three pure jets. But what happens in between? It is not difficult to understand that even in the absence of
experimental errors, there is an interval < < where the algorithm would see 2 jets. But which
of the three particles are combined into a jet -- and therefore the resulting 4-momenta -- depends on the
order in which the algorithm treats the particles and on minor programming details. It is also clear that
to regard the states P
,3 for < < as having two jets does not make much sense. Experimental
errors make the situation worse. Thus there exists an instability region that is adjacent simultaneously to
1-, 2- and 3-jet regions, in which the conventional jet counting is rather meaningless. This is also reflec-
ted in the fact that even if one forces the algorithm to "count" jets, the resulting jets' 4-momenta cannot
be determined in stably. Such an instability region is cross hatched in Fig.2.24.
Similar instability regions occur for higher n. Consider the following configuration:


2.25

As varies from 0 to the 2-jet region directly connects to the 4-jet region (unlike the connection be-
2
tween 1-jet and 3-jet regions considered above). However, the connection point is also adjacent to the 3-
jet region so that there is an instability region around that point in which the number of jets fluctuates
between 2 and 4. A similar direct connection exists e.g. between 3-jet and 6-jet regions.
So, the geometry of n-jet regions and boundaries between them as well as the instability regions, be-
come rather intricate: The n-particle subspace of { }
P consists of interlaced m-jet regions for all m n,
with many boundary points adjacent to more than two such regions.

Effects of the shape of (P) 2.26

Consider the n-jet fractions defined using any conventional jet finding algorithm:

n jets = d ( ) f
z P P n P(), 2.27

where f P P ha n
n ( ) ( s je )
ts is 1 or 0 depending on whether or not the condition in the argument of
is satisfied. In particular,

n
jets = 1 . 2.28
n=1,2K

The number of jets is determined incorrectly for events from instability regions, which results in a redis-
tribution of events between n-jet fractions for different n, i.e. a smearing between them. The smearing
affects not only pairs of adjacent n-jet fractions, say 2- and 3-jet fractions, but also 2- and 4-jet ones
(cf. Fig. 2.25), etc.
Furthermore, recall that within QCD radiation of each additional parton involves a factor S so that

n+1jets = n
O )  jets
( S . 2.29

This means that the smearing affects quantities that may differ by an order of magnitude: 1% of 2-jet
events incorrectly identified as having 3 jets may mean an O(10%) error in the 3-jet fraction which
would affect the determination of + -
S from (e e 3 je )
ts . There is also a smearing between, say,
3
3 jets and 6 jets in which case the two quantities may differ by O( s ) ~O(1000) although it is hard to

estimate the contribution of the corresponding instability region.
Another feature of QCD is that the probability of additional parton radiation is enhanced in collinear


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 21






regions by collinear singularities. Therefore radiation of an additional jet is enhanced in configurations
where it is closer to, and harder to differentiate from, the other jets.
It is clear that the smearing affects most significantly the n-jet fractions for higher n. The obvious
geometric reason is that the surface of the unit sphere remains 4 so that angular distances between jets
must decrease while jets on average tend to be softer (less energy per jet), therefore, wider. On the other
hand, the collinear singularities make it more likely that for a larger number of jets at least one pair of
jets has a smaller angular distance.
Numerical importance of such effects in a realistic situation has already been discussed (cf. 1.3).

Summary 2.30

(i) Mathematically, instead of "ambiguities" of jet algorithms one should talk about their instabilities
resulting from their discontinuity.

(ii) The discontinuity of jet finding algorithms manifests itself through various instabilities. One is the
enhancement (as compared to the case of continuous weights) of data errors near the boundaries sepa-
rating regions corresponding to events with different jet numbers. This results in a smearing between
such regions. Its numerical effect is enhanced by the intricate geometry of the n-jet regions in the con-
tinuum of all final states and by the specific form of the probability distribution in QCD. It can be
eliminated neither by varying the jet resolution parameter ycut, nor by increasing statistics (for a given
precision of detectors). It is more important for smaller ycut, lower energies and larger numbers of jets.
Another manifestation of instabilities is a systematic error in determining jets' 4-momenta. In the latter
case the numerical effects of instabilities are expected to remain significant in physical problems of
much interest even at the energies of LHC [9].

(iii) The enhancement of statistical measurement errors by discontinuities can be eliminated by using
continuous weights (regularized cuts) instead of hard cuts in order to quantify the physical features one
studies. Moreover, the use of continuous weights here conforms better to the physical reality of absence
of boundaries between events with different numbers of jets.

(iv) It can be suggested that in high precision/low signal situations one should regularize cuts as a
matter of routine -- irrespective of the physical nature of the problem (jets or not; see Sec. 2.21).

(v) The term "ambiguities" used in connection with the problems of jet algorithms, conveys a wrong
impression that the ambiguities may perhaps be fixed by invoking additional considerations. It should be
emphasized that this is not possible because one actually deals with instabilities that are due to disconti-
nuities that cannot be eliminated as a matter of principle: A mapping of any connected continuum (such
as the space of final states ) into integer numbers (such as classification of events according to their
"number of jets") cannot be continuous in any non-pathological sense. The discontinuity, therefore, is a
fundamental intrinsic property of any jet finding algorithm. It is impossible to eliminate the effects of
"ambiguities" of such algorithms without abandoning them altogether as a primary data processing tool
in jet-related measurements.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 22






Mathematics of measurements and continuity 3

Everything that happens to experimental data prior to the moment when they are confronted with the-
oretical numbers constitutes a physical measurement. In high energy physics this involves two stages:
 First one records the raw data from a detector installation. In the context of precision measurement
problems, the data thus obtained constitute a result of a physically meaningful measurement to no
higher degree than, say, a digital photograph used as an intermediate representation to measure the
length of a stick.i
 At the second stage a reduction of information is performed, i.e. extraction of a "physical meaning"
from the raw data. This involves a rather complex manipulation of numbers with a computer.ii

The above definition of physical measurement is an expression of an attitude which is appropriate for
the problems of precision measurement class: It makes one focus on what is the primary concern of any
measurement, namely, the precision of the final results.

Data processing and errors 3.1

The problem, therefore, boils down to studying how the properties of the algorithms employed for re-
duction of information interact with data errors etc.iii As follows from the discussion of a one-dimensi-
onal example in Sec. 2.5, the integral output errors for a quantum observable are minimized if the func-
tion that defines the observable is chosen sufficiently regular -- as a minimum, continuousiv.
The crux of the matter is that continuity is a simple notion only for numeric functions of a finite
number of numeric arguments. But the number of particles in final states in high energy physics ex-
periments cannot be meaningfully restricted (the actual numbers are O(100) ), so the continuum of mul-
tihadron final states should be regarded as infinitely dimensional. In an infinitely dimensional space
many radically nonequivalent continuities are possible (see any textbook on functional analysis, e.g.
[43], [44]).
For instance, consider the infinitely dimensional space of ordinary continuous functions; the Fourier
series for such a function converges to it in the L2 topology but need not converge to anything at all in
the uniform sense (i.e. in the so-called C0 topology; cf. Sec. 15). An appropriate convergence in each
case is determined by a particular application and cannot be postulated a priori.
So, the problem is to choose an appropriate continuity for functions defined on final states. There is a
concrete criterion to make the choice: the desired continuity should ensure a minimization of integral er-
rors in accordance with the conclusions of Sec. 2.21. Obviously, the structure of data errors of a particu-
lar class of detectors should play a crucial role here.
Consider the difference between the ideal expression of an observable, Eq. 2.7, and its version that
takes into account measurement errors, Eq. 2.9 (we consider here the most general case making no as-
sumptions about the dimensionality of the continuum of P):

~
F - F = P
d P
( )  F P
( ) - F P
z ( ) . 3.2
exp



i Of course, this depends on the concrete problem: if one aims at modeling data in full detail (as is the case,
e.g., with weather forecasts) then such digitized data are immediately compared with theoretical predictions,
and are to be regarded as the final results of measurement.
ii Inclusion of a mathematical algorithm into measuring system is a familiar concept, e.g., in spectroscopy and
other related fields; cf. the textbook [40].
iii Although we are speaking about data errors, similar arguments apply to other sources of uncertainties such as
unknown corrections (logarithmic and power) in theoretical studies as well as minor variations of program im-
plementations.
iv More stringent types of regularity -- differentiability etc. -- require that a continuity be defined first.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 23






The integration over the events P with the weight can be simply interpreted as a summation over a
sample of events. Then the first term in the square brackets can be thought of as the value of the obser-
vable F for the event P distorted by the detector errors ( describes their magnitude). To make things
more transparent consider the following analogue of 3.2:

F - F = P
d P
( )  F P
( ) - F P
z ( ) , 3.3
exp


where P is the distorted event. We have seen in Sec.2.5 that the integral error is minimized if the fluc-
tuations of the values of F induced by errors in the argument vanish together with the magnitude of er-
rors, which can be expressed as follows:

F P
( ) - F P
( )
0 for any P
. 3.4
0

This is nothing but a form of continuity, and it will be discussed later on in more detail.

A mathematical subtlety 3.5

Strictly speaking one should require that F satisfy a somewhat more stringent condition:

max (
F P ) - (
F P)

0. 3.6
P 0

If P were real numbers (or finite dimensional vectors) then Eq. 3.4 would simply describe continuity of
the function F P
( ) while Eq. 3.6, its uniform continuity. The latter does not, in general, follow automati-
cally from the former -- but it does if the region in which the function is defined is compact (e.g. a
sphere). In finite dimensional situations compactness of a region is equivalent to its being bounded and
closed [45]. In infinitely dimensional spaces the issues of compactness and uniform continuity are more
tricky.
A mathematician will note that the required compactness in our case follows from the standard Ba-
nach-Alaoglu theorem (see e.g. [44]). Indeed, we will see (Sec. 4) that P in our case are interpreted as
measures on the unit sphere while the continuity expressed by 3.4 will be seen to be the continuity with
respect to the *-weak topology in the space of P. Furthermore, the integral of any P over the entire unit
sphere with unit weight is limited by a constant (the total energy of the colliding particles). According to
the Banach-Alaoglu theorem the set of such P is compact. Then Eq. 3.4 automatically implies 3.6.
On the other hand, a reader to whom *-weak topologies mean nothing need not worry much about the
reasoning in the preceding paragraph. Indeed, in the final respect our objective is to find a class of ob-
servables that could serve as an alternative to jet algorithms. The condition 3.4 is restrictive enough for
that purpose -- it will allow us in Sec. 5 to derive the so-called C-correlators that: form a rather narrow
class; allow one to express practically any physical feature; allow an alternative derivation (Sec. 6).
Whether or not the resulting formalism can become a viable alternative to jet algorithms is a complex
issue that can only be decided by practice, and we simply accept 3.4 as a basic requirement which our
observables should satisfy.

Data errors and convergence 3.7

Recall that to define continuity of a function one first has to define convergence of its argumentsi. In
our case, the arguments are final states P that form a space that we denote as { }
P . Suppose we have de-
fined which sequences of final states P P
n { } are considered as "convergent". Then the observable
F P
( ) is continuous (with respect to the specified convergence in { }
P ) if its values F P
( )
n form a conver-


i The standard mathematical term to denote what we call convergence is "topology" [44]. Note that experimen-
talists speak about "topology of an event" (= geometry of the event) which is a completely different thing.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 24






gent numerical sequence for any sequence Pn that converges in the specified sense.

Reminder: length and real numbers 3.8

It may be helpful to begin the discussion of connection of measurement devices, errors, and conver-
gences by briefly reviewing the issue in the simple case of length measurements.
Recall the formal construction of real numbers. One begins with rational numbers and defines the so-
called Cauchy sequences of rational numbers that have the following property: If rn is any such sequen-
ce then for any N one has r - r < 1
n n for all sufficiently large n and n . (The l.h.s. of the criterion
N
is the usual distance between two numbers, i.e. the modulus of difference.) Such sequences are declared
to be "convergent" (with respect to the chosen criterion). The real numbers are then defined as ideali
objects that correspond to such sequences: One says that each Cauchy sequence converges to a real
number. Two different convergent sequences an and bn are said to be equivalent and represent the same
real number if the sequence obtained by combining the terms of the two, a , b , a , b , a , b K
1 1 2 2 3 3 , is again
a convergent sequence. The construction is called completion of the space of rational numbers with re-
spect to the defined convergence. Arithmetic operations are extended to the ideal numbers via continuity.
One can see that the construction of real numbers is connected with length measurements as follows.ii
Imagine a stick whose length L one wants to measure. One takes a ruler graduated every 1/ 1
N part of,
say, meter. One aligns its zero end with one end of the stick and finds the two adjacent marks on the
ruler between which the other end happens to be. One counts the marks and obtains a pair of rational
numbers r, r
1 1. One chooses either of the two or any rational number in between depending on a con-
vention. Denote the chosen number as r1. One says that r1 represents L within the precision
= /N = r- r
1 1 1 1 2 . Next one takes a more precise ruler, i.e. graduated every 1/N2 -th part of the meter
with N > N
2 1 and repeats the procedure. One obtains another interval r , r
2 2 and a rational number r2
which is said to represent L within precision =
2 1/N2, etc. It is an experimental (sic!) fact that all the
intervals r , r
n n overlap.iii This is reflected in that whatever the convention for choosing rn , all such se-
quences happen to be convergent and equivalent to one another in the sense of the preceding paragraph.
This should be compared with the construction of the so-called p-adic numbers (for each prime p
there exists a continuum of such numbers). The latter are also defined as ideal numbers represented as
"limits" of Cauchy sequences of rational numbers except for one difference: the "distance" -- and there-
fore the notion of convergence -- is chosen differently from the case of real numbers (for details see
[47]). Despite the difference, the p-adic numbers share many properties with the real numbers, e.g.
arithmetic operations remain continuous in the space of p-adic numbers. The p-adic numbers are by no
means a pathology: They emerged naturally in connection with certain fundamental problems of the the-
ory of numbers (the latter has vital applications to cryptography and digital signature systems on which,
e.g., modern electronic banking is increasingly dependent [48]).
Lastly, it is only rational numbers that are actually used in practice, and one might wonder whether
choosing a particular idealization makes any difference. In particular, what makes us stick to the real
numbers in theoretical constructions? The answer should by now be obvious: it is the fact that their
continuity properties directly formalize the structure of measurement errors of the concrete measurement
procedures that prove to be useful in a particular kind of applications. Similarly, we will see that there
is just one convergence in the space of final states/energy flows that formalizes the structure of errors of
calorimetric detectors.


i The ideal nature of real numbers is to be contrasted with the fact that rational numbers exist in a very tangible
form as finite sequences of digits on paper or bits in computer memory.
ii One would normally discuss such measurements while taking for granted that length is represented by a real
number. We wish to make no such assumption, which explains a somewhat pedantic exposition.
iii The complication of statistical errors is ignored here for simplicity.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 25






Convergence in terms of resolution of measurement device 3.9

In what follows it will be useful to reformulate the definition of Cauchy sequences in terms of meas-
urement devices in such a way as to make it easily generalizable to more complex situations. The new
definition is as follows: A sequence of measurements ln is convergent in the sense determined by meas-
urement devices (e.g. rulers) if whatever the resolution of a ruler, ln cannot be distinguished using that
ruler for all sufficiently large n.
The definition remains valid after the completion that results in the space of ideal lengths/real num-
bers: A sequence of lengths/real numbers ln converges if whatever the resolution of the ruler, ln can-
not be distinguished using that ruler for all sufficiently large n.

Complex measurement installations 3.10

In general, a complex measurement installation -- e.g. those used in high energy physics -- consists
of more than one elementary detector module each yielding a (rational) number. The total number of
modules need not be fixed, i.e. it may vary depending on a particular installation. But it is always finite.
An example is a typical high-energy physics detector consisting of many detector modules. Each
measurement from such an installation (corresponding to one event) yields a data record -- a finite ar-
ray of numbers organized in a certain way. Denote such a data record as P. Consider a particular class
C of measurement installations (e.g. calorimetric detectors). The detector records P from all possible
installations of the class C can be regarded as elements of a mathematical space (denoted { }
P ). The
structure of installations of the class C reflected in the structure of errors of the records P determines a
convergence in { }
P .

Definition of convergence 3.11

First define the convergence in { }
P following the pattern of Sec. 3.9: A sequence of data records Pn is
called convergent in the sense of C if whatever detector installation from this class is used, the ele-
ments of the sequence cannot be distinguished by that installation for all sufficiently large n. (We will
also use the term C -convergent and C -convergence in such cases.)
Let us now take into account that the number of elementary detector modules in any such installation
is finite. Denote the modules of the installation as mi. Consider any one such module mi; let mi ( n
P ) be
the number it measured for Pn. The module mi has a finite precision, and the numbers mi ( n
P ) become
indistinguishable within the precision of m =
i starting with, say, n Ni. Then the entire installation (the
collection of all m =
i ) will not "see" the difference between Pn starting from n max N
i i . One notices
that because the number of elementary modules is always finite, the convergence in the sense of C may
be equivalently defined using only elementary modules: One says that the sequence Pn converges in the
sense of C if for any elementary module m from measurement installations from the class C the num-
bers m( n
P ) become indistinguishable for all sufficiently large n -- whatever the precision of m.
Finally, one can reformulate the definition in such a way that the precision of modules is not mention-
ed: The sequence Pn converges in the sense of C if for any elementary module m there exists the limit

lim m (P ) < +
n . 3.12
n

This is the definition we adopt in what follows.
It remains to note that the above definition is completely general, for the very nature of measurement
is such that only a finite number of "detector modules" can be involved in any instance of measurement
process.
A concretization of this scheme to the case of multihadron final states and calorimetric detectors to-
gether with precise descriptions of P, m, { }
P and the resulting "C-convergence" is presented in Sec. 4.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 26






Convergence in terms of open neighborhoods 3.13

A mathematician may note that the above scheme is easily transformed into the usual mathematical
definition of a topology. Indeed, the "C-topology" in the space { }
P that is equivalent to the above C-
convergence is exactly the weakest topology in which all numeric functions m(P) corresponding to all
allowed elementary detectors are continuous. Note also that in applications { }
P is always separable. In
the concrete situation considered below in Sec. 4, { }
P consists of linear functionals on a linear space
whose elements correspond to elementary detector modules m. Then the C-topology is a variant of the
well-known -weak topology [44]. In our special case we deal with a linear space of measures, and the
convergence is also known as the weak convergence of measures [49].

Summary 3.14


(i) Integral errors of data processing in complex measurements -- the errors that are due to measure-
ment errors in the raw data -- are minimized if one uses sufficiently regular continuous observables.
(This is also true of the purely statistical fluctuations; cf. [41].)

(ii) To define the required continuity of observables one must specify a notion of convergence in the
continuum of mathematical images of physical objects one deals with.

(iii) The required convergence is determined by the structure of measurement errors in the raw data,
eventually, by the kinematic structure of the measurement device.

(iv) A natural completion construction results in an ideal representation of the physical objects one
measures as some sort of ideal mathematical objects (e.g. length as a real number).

(v) The notion of convergence that is specific to multimodule detectors can be characterized in terms
of elementary detector modules (Eq. 3.12). (The resulting convergence happens to be a variant of the so-
called *-weak topologies that are well-known in mathematics.)

The plan of our construction is now clear: First, we should precisely describe the objects we wish to
measure (i.e. events/multiparticle states/energy flows) as well as the elementary calorimetric detector
modules. Then we will be in a position to describe the convergence determined by the calorimetric detec-
tors following the above scheme. This will give us a complete formalization of what "energy flow" is
from mathematical point of view.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 27






Multiparticle states, energy flows and calorimetric detectors 4

In this section we apply the scenario described in the preceding section to the special case of multi-
module calorimetric detectors in order to arrive at a precise mathematical description of energy flow in-
cluding the convergence defined by such detectors.
As was explained e.g. in [26], there are two different geometric frameworks in which jets are studied
that correspond to two different physical situations. The first case is the e+e- annihilation experiments
where the hadronic system as a whole is at rest in the laboratory reference frame and one aims to em-
phasize rotational symmetry. Then it is natural to regard the detector modules as covering a sphere.
The other case corresponds to hadronic collisions where one emphasizes invariance with respect to
boosts along the beam direction. One then deals with a cylindrical geometry. In what follows we con-
sider the spherically symmetric case. Our notations are mostly independent of a particular parametriza-
tion, so only a few formulas are affected by modifications needed for the case of hadronic collisions dis-
cussed in Sec. 13.

Multiparticle states and their energy flows 4.1

One deals with multiparticle final states produced in collisions at a fixed point within a detector
where we place the origin of the coordinate system. For each particle in such a state, a detector installa-
tion measures its characteristics. Calorimetric detectors measure its energy E 0 and the direction in
which it goes (denoted as $p ).

Parametrization of directions $p 4.2

A simple way is to represent the direction $ 2
p as a unit 3-vector $
p = 1, which is a point of S2, the
unit sphere in 3-dimensional Euclidean space. One can parametrize it, e.g., with the two standard an-
gles, and . But what really fixes the spherical geometry in our context is how one defines integration
over $p . If is a set of directions then for the spherical geometry one defines:

d $p surface o
z = f . 4.3


In terms of the angles and :
z 2
d $p ( $p ) = z sind . 4.4
0 z d ( ,)
0

Angular distance between directions: "angular separation" 4.5

Another quantity that depends on whether one uses spherical or cylindrical geometry is the measure
of distance between two directions corresponding to two particles. One such measure used in the context
of spherical geometry is the angle between the two directions:

=
ij arccos $ $
i
p p j . 4.6

It turns out, however, that another quantity appears in the expressions on a regular basis -- the quantity
we call angular separation and denote as ij systematically throughout this work. Its specific expres-
sion in terms of angles etc. depends on a particular kinematic situation. For the spherically symmetric
case of e+e- hadrons the angular separation is defined as follows:

= 1 - cos = 1- $ $ 1 2
p p d for small i
ij ij i j ij ~ . 4.7
2 ij ij

The considerations that went into the definition 4.7 are as follows:


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 28






(i) For small angular distances between the two directions $ $
p and p
i j the factor ij should be useable

as the angular factor in the optimal preclustering criterion 7.39.

(ii) For large angular distances, it should be useable as a natural building block for the jet-number dis-
criminators Jm defined in Sec.8 (cf. Eq.8.8).
(iii) It should be as geometrically natural as possible for a given kinematical framework. In particular,
it should conform to the structure of typical factors in matrix elements; this facilitates theoretical study
of, e.g., Jm that play a central role in the C-algebra of observables described in subsequent sections.

The formulas expressed in terms of the angular separation remain valid in the case of cylindrical
kinematics of hadron-hadron collisions -- all one has to do is to redefine ij appropriately; cf. Sec.13.

Calorimetric information and energy flow 4.8

We will distinguish calorimetric information (i.e. E and $p for all particles constituting a given final
state) from other types of information. If one retains only calorimetric information about the final state,
then one obtains what is called energy flow of that state. It is the ways to measure the geometric shape
of the energy flow that is our object of study. All the non-calorimetric information is then treated as
additional parameters (e.g. momenta of "high- pT muons" etc.).
There is some arbitrariness in separating what one will treat as calorimetric information: If one be-
lieves one can identify an intermediate particle from purely calorimetric data with enough confidence
and precision, one may be willing to treat it as a non-calorimetric information. We are interested in
situations where not all data can be treated as non-calorimetric, and it is the calorimetric information
that we focus upon.
Thus, a final state P is characterized by its energy flow (calorimetric information) and perhaps some
other parameters which we will treat as implicit. For simplicity of notation, we will not distinguish the
multiparticle state from its energy flow.i
A final state as seen by an ideal calorimetric detector (with perfect energy and angular resolution) is
represented as

P = kE , $p p =
1 K k p
= , , . 4.9
part p
i i i N i i

Calorimetric detectors do not distinguish types of particles, so the states are not affected by permuta-
tions of pi .
It is convenient to have a notation for the state P consisting of all particles from other states, e.g. Pa
and Pb. Then we write

P = Pa Pb . 4.10

Lorentz invariance 4.11

The concrete formulas below are motivated by the idealization that is usual in high energy experi-
ments, namely, that all particles are massless. If such an idealization is valid, then the 3-momentum of a
particle is 2
p = Ep
$ , and the 4-momentum is p = (E, p), p = 0 . A Lorentz-covariant expression for the
energy is

E = pP , P2 = 1, 4.12



i Strictly speaking, energy flows are elements of the factor space of the space of final states with respect to the
equivalence relation of fragmentation invariance. Drawing such a distinction explicitly would be too cumber-
some.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 29






where P is a 4-vector whose rest frame defines the coordinate system. Then a Lorentz covariant ex-
pression in the P's rest frame for the angles 0 ij between pairs of particles can be obtained from
the following relation:

p p = E E - p p = E E (1 - p$ p$ ) = E E (1 -
i j i j i j i j i j i j cos ij ). 4.13

The concept of jet is not Lorentz-invariant, and the calorimetric data (energies and directions of par-
ticles) are specific to the detector rest frame. In some cases one may wish to look at an event in a refer-
ence frame other than the detector rest frame [50]. This entails performing a Lorentz transformation that
makes use of the masslessness assumption. But although Lorentz invariance provides motivations for
fixing details in concrete formulas for observables, it is an extrinsic concept for the issues we address in
the theory of calorimetric observables. Thus, the 4-momentum introduced above is just a convenient la-
bel, while the basic objects are the energy and direction, both immediately measurable and subject to er-
rors. This means that studying corrections due to non-zero masses is relegated to the theory department
where it properly belongs. Note that such corrections can be studied in the general context of power cor-
rections. This is further discussed in Sec. 6.16.

Detectors 4.14

A realistic detector installation consists of a large but finite number of elementary calorimeter mod-
ules. Each such module may be represented by a continuous function ( $p ) that takes the values betwe-
en 0 and 1 and describes the local efficiency of the module, and the energy it measures for a state P is

N
P, part i
E ( $p ) . 4.15
i= i
1


Data records 4.16

Let the c-th module of the detector be described by c ( $p). Then what the entire detector sees about
the final state P is the collection of numbers

E = P, , c =
c c ,
1 K . 4.17

The modules are chosen to have small angular sizes, and one assumes that all the information one needs
is the approximate position $
qc of the calorimeter module. Then a typical data record is a finite array of
the form
kE p
c , $
qc . 4.18
c

Given sufficient energy and angular resolutions of individual detector modules, a physicist regards such
a data record as an adequate representation of the exact energy flow 4.9.
All possible different detector records from all imaginable detector installations form a mathematical
infinitely dimensional space (= the space of data records). Formally, multiparticle states 4.9 with finite
numbers of particles are elements of that space.
I emphasize that when translated into the language of mathematics, the implicit physical convention
that a detector record is "close within detector resolution" to the corresponding physical multiparticle
state, becomes an axiom that determines one of many mathematically possible forms of convergence in
the space of data records -- a convergence predetermined by the structure of measuring devices
(calorimetric detectors in the present case). This is a central point to which we will soon return.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 30






Energy flows as "abstract measures" 4.19

The structure of 4.15 suggest the following mathematical interpretation: Each multiparticle final state
P as seen by all imaginable calorimetric detectors is represented by a sum of -functions:

N
P = P( $ part
p) = E . 4.20
i= ( $, $
p p )
i i
1

To interpret this expression correctly, recall that, by definition, the -function ( $, $
p p )
i acquires a con-
crete numerical meaning only in integrals with continuous functions:
z d$p ($p)($, $
p p )
i ( $i
p ). 4.21
S2

Correspondingly, Eq. 4.20 represents the collection of all values of integrals with all possible continu-
ous functions (the values of such integrals reduce to 4.15).
In mathematical language, the objects represented by 4.20 are linear functionals defined on all con-
tinuous functions on S2 -- for each such function they define a number (given by 4.15). Such objects
are called measures on S2 . To avoid confusion with physical measurements, we will use the term
abstract measure to denote such objects. (For a background mathematical information on abstract
measures see Secs. 1516.)
For applications it would be sufficient to have in view the two formulas 4.20 and 4.21.

Calorimetric convergence (C-convergence). When are two energy flows close? 4.22

We have come to the key point of our study, namely, a mathematical description of closeness of en-
ergy flows of multiparticle states. This is achieved by introducing a convergence in the space of all en-
ergy flows -- the convergence determined by the measurement devices (cf. Secs. 3.73.14).
Suppose one has a sequence Pn of multiparticle states (abstract measures, or linear functionals).
We say that it converges to a multiparticle state P if for any detector module described by a continuous
function , the values of energies measured by for Pn converge to the energy measured by for P:

lim P , = ,
n P for any allowed . 4.23
n

The reasoning of Sec. 3.10 shows that Pn will then be indistinguishable for any detector installation
consisting of a finite number of detector modules for all sufficiently large n. We call this calorimetric
convergence (C-convergence for short).
Mathematically, this is a special case of the general notion of -weak topology in a space of linear
functionals (Sec. 3.13). Such topologies cannot, in general, be usefully described by a simple single-
valued distance or norm, which makes them seem somewhat amorphous for applications. Nevertheless,
the above definition is precise enough to allow one to find constructive ways to deal with the situation.

C-convergence and collinear fragmentations 4.24

Let us explain the meaning of C-convergence with a few examples. One easily verifies from the
definition that an energy flow changes C-continuously under any of the following modifications:

(i) Continuous variations in the energies of the particles constituting the event (energies are never
known precisely).

(ii) Adding any number of arbitrarily directed particles with the total energy going to zero (and the
number and directions of particles may change arbitrarily in the process). Indeed, soft particles may es-
cape undetected and their total number is unknown.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 31






(iii) Continuous variations of the angular parameters of particles in the event. This is because particles'
directions are determined up to the size of calorimeter modules.

(iv) Almost-collinear fragmentations. This can be regarded as a combination of an exact fragmenta-
tion and (iii). An exact fragmentation i consists in a replacement of a parent particle by any number of
collinear fragments,

E , $ E , $ ;K; E
a a b , $
p p pb , 4.25

where
E = E +K+ E p =K
, $ = $p = $
a b a b p . 4.26

This of course does not change the energy flow 4.20. If after such a fragmentation one changes the di-
rections of the fragments a little, one obtains an almost-collinear fragmentation.

Correctness of the scheme 4.27

We have introduced a mathematical model for energy flow via a somewhat loose reasoning using the
physical notions of particle etc. that are extrinsic with respect to the measurement process as such. We
should verify that no inconsistency was introduced thereby. To this end, it remains to recall the simple
mathematical fact that any measure can be approximated in the *-weak topology by finite sums of -
functions, and the space of abstract measures is complete in that topology (cf. Secs. 1516).
The data record 4.18 can be viewed as a finite sum of -functions on the sphere (the energies meas-
ured by individual detector cells are the coefficients, and their positions are the points where the -
functions are localized). The completion of the space of all data records with respect to the C-
convergence restores abstract measures as ideal representations of energy flow. Another aspect of this is
that we never see the final state as such (except in Monte Carlo modeling) but only the energies meas-
ured by calorimetric cells; in an operational sense, a final state is a collection of the data records pro-
duced by all possible calorimetric detectors.
The physical meaning of the C-convergence clarifies in a broader context of the discussion of Sec. 3.
It may be helpful to represent the analogy between calorimetric measurements and the measurements of
length as follows:


























length energy flow






                                              


                                                 


















ruler calorimetric detector






                                              

                                                 












rational nu
m
ber

data record

























usual convergence C-convergence













of rational numbers (
*-weak topology)




























real number abstract measure 4.28













i The notion of fragmentation was introduced in [2] and formalized in [50], where the adjective "exact" was not
used because data errors were not considered.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 32






Comparison with the L2 formalism 4.29

As was already emphasized, the space of all energy flows is infinitely dimensional, and there are
many nonequivalent continuities in such spaces. Choosing a wrong one may be rather misleading. To
appreciate the subtlety of the problem, recall the large scale study described in [22] where an idealiza-
tion of energy flows as ordinary functions on S2 was adopted. To measure closeness of energy flows,
the following norm was chosen:

1/2
def I
P 2
2 = F d $p P
z ( $p ) . 4.30
L H K
S2

The resulting L2 topology is familiar from courses of quantum mechanics, is very similar to norms in
finite-dimensional Euclidean spaces, is fairly easy to deal with, and allows a number of sophisticated
constructions, e.g. expansions in spherical harmonics, summations involving 3-j symbols, etc.
Unfortunately, such superficial analogies have strictly nothing to do with physics. Indeed, it is well-
known (and is rather obvious) that the L2 norm is a poor tool for comparing shapes, which is exactly
what one does in classifying events with respect to the geometry of energy flow. In particular, the L2
norm of the energy flow that corresponds to a finite number of particles, Eq. 4.20, is infinite.
One reason why the research of [22] was led astray was an unmotivated transition to "continuous
limit" (interpreted there as a replacement of the energy flow with an ordinary function) prior to specify-
ing the notion of convergence with respect to which the limit is understood.

Correct continuous limit 4.31

Our interpretation allows taking infinite -- and even continuous -- sums in 4.20, which can be done
as long as the total energy remains finite. The limiting procedures implied thereby are to be understood
as applied to each value of the functional representing energy flow, on each continuous function. For
instance, consider a continuous sum,

P( $p ) = E ( $, $
p p )
zd , 4.32

where E is, e.g., a continuous function of and the integration may run, e.g., over a non-zero length
arc on S2 , or over a part of S2 with non-zero surface. Then by definition one has
P,
zd E ($p ), 4.33
where the integral is defined in the usual sense.
Eq. 4.33 remains well-defined as long as is continuous and the total energy is finite:

d
z E<. 4.34
In particular, the definition of C-convergence remains valid in the case of more general energy flows
that cannot be associated with states with finite number of particles (e.g. in QED).


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 33






Calorimetrically continuous (C-continuous) observables 5

In the preceding Sec. 4 we established the mathematical nature of energy flow as abstract measure on
the unit sphere, together with a correct convergence (the C-convergence) that corresponds to the struc-
ture of errors of multimodule calorimetric detectors. But that was only an intermediate step. What one
actually needs is a practical description of functions on final states/energy flows that are continuous
with respect to the C-convergence of their arguments. In this section we describe a large class of such
functions. (The special role of continuous observables was discussed in Sec. 2.)
Recalling that an (ideal) energy flow P is an abstract measure (= a linear functional on continuous
functions on the unit sphere; Sec. 4.19), one cannot miss the mathematical subtlety of the situation: an
observable F P
( ) is a numeric-valued function defined on linear functionals defined on continuous func-
tions defined on the unit sphere. There seems to exist no systematic mathematical study of the structure
of non-linear functions on functional spaces in the constructive aspect.
Nevertheless, it turns out possible to point out a useful basic class of such functions in our case (C-
correlators; Sec. 5.1, esp. Sec. 5.15). C-correlators also turn out to have a natural QFT interpretation
(discussed in Sec. 6.1). The latter property opens a prospect of their systematic theoretical study.
Moreover, as described in Sec. 5.22, C-correlators can be organized into more sophisticated C-
continuous observables (e.g. the spectral discriminators introduced in Sec. 13) thus forming an algebra
(C-algebra) in terms of which any physical feature can be correctly expressed.

C-correlators: basic C-continuous observables 5.1

Linear functions of energy flows 5.2

Abstract measures P are defined as linear functionals P on continuous functions , formally rep-
resented as

P = p P p
z d$ ($) ($p). 5.3
S2

Fix one such function ( $p) and consider the expression P as a function of P:

F P
( ) = P
. 5.4

It is a tautology to say that this is a C-continuous function of P. Indeed, recall that the C-convergence
was defined in Sec. 4.22 by requiring continuity of functions of exactly this form with special varying
between 0 and 1. However, linearity ensures that the expression 5.4 with arbitrary continuous are also
continuous.
If P describes a finite number of particles (4.9 and 4.20) or a detector record 4.18, then Eq. 5.4 be-
comes

N
F P part
( ) = i
E ( $p ) . 5.5
i= i
1

Note that this expression is obviously invariant with respect to exact fragmentations 4.25. Equally obvi-
ously, small variations of angles and energies as well as almost collinear fragmentations (Sec. 4.24) re-
sult in small variations of F .
Examples 5.6

Let 0 ( $p ) 1 be a continuous function on the unit sphere. It can be thought of as describing the
local acceptance of a detector cell. Then the expression 5.5 gives the energy deposited in that cell.
A trivial special case (( $p) = 1) is the total energy of the event:


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 34






def N
E part
= E P
( ) = E

tot . 5.7
i i
=1

Less trivial examples are the multipole (non-spherically symmetric) moments used as elementary build-
ing blocks in the constructions of [22].

Bilinear scalar functions of energy flows 5.8

A standard mathematical construction is to take a tensor product of two or more measures ([45],
sec. IV.8). Consider a direct product S  S
2 2 of two unit spheres (its points are parametrized by a pair
of unit 3-vectors, $
ap , $
p) S  S p p
2 2 or, equivalently, by four angles). Let ( $ , $ ) be a continuous
function on S  S
2 2 . Then the well-defined expression

( $p) $
= dpP ( $
p)( $p , $
p)
z 5.9
S

is a continuous function on the unit sphere and one can perform another integration with P :
z d$pP($p)($p) . 5.10
S

This defines a measure on S  S
2 2 that can be denoted as

P  P ( $p, $p) = P( $p)  P( $p). 5.11

Correctness of the construction -- the required continuity properties etc. -- is verified in detail in Theo-
rem 73 of [45]. In particular, the result is independent of the order of integrations.
In particular, one can take P = P = P. Applying the arguments of Sec.5.2 to the measure P  P
thus constructed, one obtains a C-continuous function that is bilinear in energy flow:

F P
( ) $ $
= dp dp P( $p)P
( $p)( $p , $
p)
z . 5.12
S zS
2 2

If P has a discrete form of 4.20 (cf. also 4.18), then one obtains:

F P
(  P
) = E E ( $p , $p ). 5.13
ij i j i j


Without loss of generality, one takes to be symmetric in its two arguments.
For instance, the total invariant massi of an event P is given (under the usual assumption of
masslessness of all particles) by the following special case of 5.13:

def
S = S P
( ) = E E
(1 - $ $
tot p p ). 5.14
ij i j i j


Multilinear C-correlators 5.15

More generally, let

f K
m ( $
p , , $p
1 m ) 0 5.16

be a continuous symmetric function of m unit vectors (i.e. a function on a direct product of m copies of
the unit sphere, Sm ; also, the assumption of positivity results in no loss of generality because one can
2
always write f = f + - f - 
m m m with fm 0 ). Then the following expression defines a C-continuous ob-
servable:


i We are always talking about invariant mass squared unless explicitly stated otherwise.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 35






F P
( ) $ K $
= dp dp P( $p ) K P
z K
m ( $p ) f ( $p , , $p ). 5.17
Sm 1 m 1 m m 1 m
2

The discrete analogue is

F P = K K
m ( ) E E f ( $p , , $p ). 5.18
i i
K i i m i i
1 m 1 m 1 m

For example, a sequence of observables of this form with specially chosen fm quantify the feature
known as the "number of jets" (the jet-number discriminators of Sec. 8).

Regularity restriction from IR safety 5.19

In general one would prefer some sort of a more stringent regularity rather than a mere continuity (cf.
(iii), Sec. 2.21). For instance, the requirement of IR safety [2] implies that the observables should be e.g.
Hlder continuous [51] (it is sufficient that they have bounded first derivatives; cf. Sec. 6.9 below). For
C-correlators, it is sufficient to impose appropriate restrictions onto their component functions fm in
5.18. In the cases of practical interest that we will consider fm happen to be infinitely differentiable.

Comparison with EEC observables 5.20

A sequence of multiparticle correlators of a special form similar to 5.18 was introduced in [24] (the
antenna pattern, the energy-energy correlators etc.; cf. the discussion of their phenomenological appli-
cations in [37]). However, those correlators are defined with discontinuous angular functions in the
form of cuts (taking values 0 and 1). Therefore, they are, strictly speaking, not C-continuous.

C-correlators and "physical information" 5.21

It is convenient to call the observables of the special form 5.17, 5.18 C-correlators. They play a
central role in our theory:
 They are directly connected with the underlying QFT formalism (Sec. 6.1).
 One can express any interesting physical feature in terms of them (Secs. 5.22 and 813).

It therefore makes sense to say that the collection of values of all C-correlators on a given final state
constitutes the entire physical information about that state.
Note that any discontinuous observable can be approximated (in an appropriate integral sense with
respect to integration over all final states) by C-continuous ones. From such a viewpoint, no physical
information is lost due to the restriction of C-continuity.

C-algebra 5.22

By C-algebra we understand a collection of C-continuous observables built from C-correlators fol-
lowing a few simple operations listed below that preserve C-continuity. An example involving integra-
tion over a parameter already occurred in the construction of bilinear C-continuous functions. The basic
operations provide a sufficient flexibility to allow one to express jet-related physics in terms of observ-
ables from the C-algebra. In particular, the spectral discriminators (Sesc. 1011) are built this way.
In view of this, whether or not the C-algebra comprises all possible C-continuous observables becomes
an issue of somewhat academic interest and will not be discussed.
Note that limiting procedures may violate C-continuity (cf. the counterexample in Sec. 6.9), so some
(minimal) care has to be exercised.

Compositions of C-continuous functions 5.23

Obviously, finite algebraic combinations (linear combinations and products) of C-continuous func-
tions yield again C-continuous functions.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 36






More generally, if a C-continuous F P
( ) takes values in a region D, and f (z) is a continuous function
of z D, then f ( (
F P)) is C-continuous. For instance, exp (
F P) is C-continuous.
Division is allowed provided the denominator can not become infinitesimally small for the final states
one works with. In particular, one can divide the basic observable 5.18 by m-th power of the total en-
ergy 5.7. Then one only has to deal with energy fractions instead of absolute energies of the particles.
(We will use this option in the definition of jet-number discriminators in Sec. 8. This is said to reduce
systematic errors due to undetected soft particles.)
A generalization of the above constructions is as follows. Let H(P, F ) be a function of two argu-
ments, the energy flow and a numeric parameter; suppose it is continuous in its pair of arguments (in
the sense of C-convergence in its first argument, and in the numeric sense in the second argument). Let
F P
( ) be another C-continuous function taking values of the same nature as the parameter. Then the
composition

F P
( ) = H(P, F P
( )) 5.24

is C-continuous. This construction is extended straightforwardly to any finite number of arguments of
both types and to the case when the second argument has a non-numeric nature (e.g. is a measure).

Filters and substates 5.25

One often wishes to study certain substates of a final state P (e.g. when searching for a particle that
decays into jets). It would then be natural to treat the substate as a final state in its own right and com-
pute observables for it (e.g. its invariant mass).
The first question is how to specify a group of particles within our formalism, i.e. using the language
of C-observables. A simple and natural way to do so is with a continuous function 0 ( $p ) 1 which
we call filter. Then the subset consists of those particles for which ( $p ) >
i 0 ; we say that the resulting
substate is "filtered" by . The energies of the particles of the substate are taken to be E
i ( $pi ) with
their directions unaffected. Denote the substate thus obtained as

def
o P = kE p
i ( $i
p ) , $i
p 5.26
K
i= ,
1


(cf. Eq. 4.9). The filtering

P oP 5.27

is a C-continuous operation: C-convergent sequences of final states become C-convergent sequences of
filtered substates (which, formally, are also elements of the space of final states).
One can take a composition of the mapping 5.27 with any C-continuous observable F P
( ). The result,
F( o P), is C-continuous.
For instance, the invariant mass of the substate o P is given by the following expression:

S( o P) $ $
= E E 1c- p p h ( $
p ) ( $
p ). 5.28
i, j i j i j i j


Similarly, the expression J o P
3 ( ) measures how well the substate can be characterized as having no
less than 3 jets. (The observables Jm for measuring the "number of jets" are studied in Secs.89.)

Integration over a parameter 5.29

Another construction that preserves C-continuity is integration of a C-continuous observable over a
parameter (which may be multidimensional):


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 37






d
z H P(, ). 5.30

In practice, it is sufficient to require that the domain of variation of is compact, and that | (
H ,
P )| is
bounded by a constant independent of P (cf. the Lebesgue convergence theorem; see e.g. Theorem 34
of [45]).

Minimization with respect to a parameter 5.31

The integration considered above is the most useful example of a construction that involves a limiting
procedure but, nevertheless, yields a C-continuous function. Another useful example is a minimization
(maximization) with respect to a parameter. If the parameter is discrete and takes values from a finite
set then the result,

min a F (P), F (P) f
1 2 , 5.32

is again a C-continuous function. (Extension to more than two functions is obvious.)
With a parameter taking an infinite number of values (e.g. a continuous one) the situation is more
tricky because, in general, limiting procedures may violate continuity even in the case of ordinary con-
tinuous functions of a real argument. (An example is discussed in Sec. 6.9.) However, the precautions
one has to take to avoid such pathologies prove neither difficult to understand nor too restrictive.
Consider the following expression:

min H( ,
P ) , 5.33


where is a continuous parameter varying in a connected compact region, e.g. an interval with both
ends included, a sphere, etc. To functions of this class belong the standard observables known as sphe-
rocity, acomplanarity, etc. (see [10] for a complete list and references). Of course, the function H must
be continuous in its pair of arguments but that is not enough. The two useful cases when C-continuity of
5.33 is ensured are as follows.
The first (obvious) case is when H is a composition of an ordinary continuous function of two real
arguments, h(x, ), and a C-continuous function F P
( ),

H P
( , ) = h F
( P
( ), ) , 5.34

and min (
h x, )
is a continuous function of x.
The second case is when the angular function fm of a C-correlator 5.18 depends on . The depend-
ence should be such that, roughly speaking, the rate of variation of fm with respect to its angular argu-
ments is independent of . For instance, may describe rotations of fm as a whole (as is the case with
spherocity etc.). But one should avoid the cases when is, say, the radius of the region within which fm
is localized, and is allowed to go to zero (cf. Sec.6.13, esp. Eq.6.15 and the remarks thereafter).

Differential C-continuous observables 5.35

It is a common practice to consider differential distributions of events with respect to an observable.
Then one considers expressions of the following form:
z dP P()(z- A P( )), 5.36
where integration is over all final states P against the probability (P) defined by the S-matrix (cf. 2.7).
Such a construction is also possible with C-continuous observables, in particular, with C-correlators.
The expression 5.36, however, is too simple to be really useful. Indeed, even to describe the feature


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 38






known as "number of jets" one properly needs an infinite sequence of scalar observables (Sec. 8).
Consider a C-continuous observable F P
( , ) that depends on a parameter and takes values from the
interval [ ,
0 ]
1 (which can be achieved by appropriate normalization). Consider the expression
z dP P()zd(z-F P(,)). 5.37
One can regard this construction as an "extension of phase space", P P, and defining a differential
observable on the "extended events". If one introduces a normalized weight depending on (cf.
Sec. 13.30) then the analogy with 5.36 is complete.
Eq. 5.37 is equivalent to measuring the observable defined by the following "function" on final states:

f (z) = (z - F( ,
P ))
zd . 5.38

f (z) is not necessarily a continuous function of z but may, in general, contain singular -functional
components (see the examples in Sec. 11.11). Therefore it is essential to understand its mathematical
nature (in order e.g. to correctly treat such issues as approximations and data errors):
A -function is defined by integrals with continuous functions. The same is true for 5.38: Let (z) be
continuous for z [ ,
0 ]
1 . Then z d (
F P
( , )) is C-continuous. It can also be represented as follows:

1
d (F P
( , )
) dz
z z (z) f(z). 5.39
0

The net effect is that the object f (z) defined by 5.38 acquires a numerical meaning after integration
with a continuous function. Mathematically, this means that f (z) is an "abstract measure" on z [ ,
0 ]
1
for each final state P (for definitions see Sec. 15). This means that Eq. 5.38 represents a measure-valued
C-continuous observable.
It may be worthwhile to emphasize that the expression 5.38 is, in fact, a shorthand notation for the
collection of 5.39 for all allowed . Therefore, the convergence of values of 5.38 is to be understood as
numerical convergence of all expressions 5.39 -- without, however, requiring any correlation of the rate
of convergence for different .
Measure-valued observables emerge most naturally in the context of studying spectral properties of
multijet substates. Recall e.g. that the spectral density of a quantum propagator is a measure (a single -
function in the case of a free particle). More generally, spectral densities of self-adjoint operators (e.g.
quantum mechanical Hamiltonians) are also, in general, measures, etc. Measures are singled out from
among all distributions by the fact that it makes sense to talk about their positivity -- which is exactly
why they occur in spectral problems.
The occurrence of measure-valued observables, however, has nothing to do whatsoever with the fact
that the energy flow of an event is also interpreted as a measure on the unit sphere. Even if all energy
flows were ordinary continuous functions, the spectral observables (e.g. the spectral discriminators in-
troduced in Secs. 1011) could contain singular -functional components. Vice versa, a spectral discri-
minator for an energy flow corresponding to a few isolated particles (and, consequently, represented as
a sum of -functions on the unit sphere) may happen to be a continuous function.
The above construction becomes particularly useful in combinations with the filtering of Sec. 5.25 --
see Sec.11.1. Further options for construction of C-continuous observables are discussed in Sec. 13.29.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 39






C-correlators and Quantum Field Theory 6

To compare results of precision measurements with theoretical predictions, one wishes the latter to be
as precise and reliable as the former. A high quality of theoretical calculations can only be achieved
within the systematic formalism of Quantum Field Theory. The C-correlators 5.18 fit well into the
framework of QFT. Moreover, the form 5.18 for jet-related observables can be easily obtained if one
follows the very basic guidelines of QFT together with minimal continuity requirements motivated by
the conclusions of Sec. 2.21. That fact strengthens the central position of C-correlators in the theory of
jets and opens a prospect for advances in the study of various kinds of corrections -- logarithmic per-
turbative higher-order terms, power-suppressed non-perturbative contributions, etc.

C-correlators from the point of view of QFT 6.1

A fundamental ingredient of QFT is the fact that interactions of elementary particles occur via emis-
sion/absorption of quanta. Therefore, one deals here with systems with a varying number of particles.
Much thought was given to this, and a general, systematic, and well-studied formal scheme for analysis
of such systems is provided by the formalism of secondary quantization. Then one works with two-,
three-, ... particle correlators, each "particle" of a correlator corresponding to an operator that probes a
particular physical feature (e.g. energy density etc.)i Moreover, one can make a stronger statement,
namely, that if a complex physical feature cannot be expressed in terms of multiparticle correlators, the
observables indigenous to QFT, then it cannot be a correct observable at all. Recall in this respect also
that vacuum averages of products of field operators -- the Wightman functions -- carry a complete in-
formation on QFT models (the Wightman's reconstruction theorem; see e.g. [52]).
It is not difficult to realize that any C-correlator 5.18 has exactly the following form of a correlator of
Bose-Einstein type:

m! 1
k kn
E KE f K K K
( $p , , $p , , $p , , $p ). 6.2
K 1
i i m
n 1
i 1
i i i
n n
k ! k !
1 K K 1
i < <inN k
part + +kn=
1 1 m n 1
k times kn times

Because we deal with quantum averages expressed via matrix elements squared rather than wave func-
tions, the Fermi statistics does not occur. The corresponding expression in operator terms has the fol-
lowing general structure:

p , ,
K p O p ,K, p
1 N 1 N ,
part part


O = d(q )K d(q )  j(q )K j(q )  f K
z z
1 m 1 m m ( $1
q , $
qm ) ,


j(q) = E a+ (q) a(q) . 6.3
q


( is the standard 1-particle phase space measure.) Such an operator interpretation leaves open a pos-
sibility of their investigation using methods of QFT. An example of such a study is [53]. Ref. [54] re-
ports a representation of 6.3 in terms of the energy-momentum tensor.
The above QFT form of C-correlators (cf. also the results of [54]) is to be appreciated in view of the
fact that, as follows from the results of Secs. 813, it is theoretically sufficient to study C-correlators
because other jet-related observables can be expressed in their terms.



i See e.g. [36], [37] for a detailed discussion of such issues. The treatment of [37] particularly emphasizes the
role of correlators in studies of multiparticle systems. Note that here again we are dealing with a purely
"kinematical" aspect that has never been properly addressed in the theory of jets.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 40






An alternative derivation of C-correlators 6.4

The derivation of C-correlators in the form 5.17 did not use assumptions about the underlying theory.
Then we showed that they have a natural QFT interpretation (Sec. 6.1). Let us now approach the issue
from the opposite end: Let us begin with a multiparticle correlator, impose the restriction of fragmenta-
tion invariance 4.25, and obtain an observable of the form 5.18. In fact, we will use a somewhat
stronger form of fragmentation invariance including a requirement of continuity inspired by the analysis
of Sec. 2.5. This and the obligatory correlator form of observables are the two ingredients that were
lacking in [2] and [51].
An m-particle correlator has the following form with an arbitrary correlator function fm:

F P = K
m ( ) f ( p , , p ). 6.5
i i
K m i i
1 m 1 m


Consider for simplicity the case m = 2 . First, for an integer n the fragmentation invariance yields
F (np) = F K
2 2 ( ,
p ,
p , p), where the argument on the r.h.s. is a state with n equal particles. From this and

6.5 one obtains f (np, np) = n2 f ( p, p
2 2 ) . From this and a similar restriction with k instead of n one ob-
tains a similar restriction with a rational r = n / k instead of integer n. The condition that the observ-
ables should be insensitive to adding soft particles results in continuity with respect to energies whence
follows a similar scaling for any real r. Then f ( p, p) = E2 f
2 2 ( $ , $
p p). In a similar way one obtains

f ( p, nq) = nf ( p, q
2 2 ) for any integer n, etc. -- until the desired form of energy dependence is obtained.
Imposing a simplest requirement of continuity with respect to the angular variables (= minimal sensitiv-
ity to almost collinear fragmentations, as motivated by the analysis of Sec. 2.5) results in the restriction
of continuity of the angular function.

Fragmentation invariance of C-correlators 6.6

The combinatorial structure of C-correlators can be studied most easily with the help of the following
representation in terms of functional derivatives:

Eq.5.18 = 1 m $ K $
D dq dq f ( $ K $
q , , q ) ( $
q )K
( $
q ), 6.7
m! z 1 m m 1 m 1 m

where
N
D = part E . 6.8
i= i
1 ( $ip )

Now all the dependence on the particle content of the state is localized within D. In particular, the
fragmentation invariance of 5.18 and 6.2 (with respect to exact fragmentations 4.25) is an obvious
consequence of the linearity of D in particles' energies.

C-continuity vs. IR safety 6.9

An interesting general issue is how C-continuity relates to the IR safety, the notion familiar from the
QCD theory of hadron jets [2]. The issue of IR safety is usually discussed in the context of perturbative
QCD where final states consist of a finite number of partons (quarks and gluons). Let F P
( ) be an arbi-
trary function on final states. Define its component functions Fn by restricting F to states with finite
numbers of particles, as follows:

F P
( ) = F ( p , ,
K p )
n 1 n , 6.10

where for simplicity each of the n particles of the final state is represented by its 3-momentum.
Since the multiparticle state does not change if the particles are permuted, each Fn must be symmet-
ric in its arguments. An observable that can be measured by calorimeters has to be [2], as a minimum,


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 41






fragmentation invariant, i.e. invariant with respect to exact collinear fragmentations, whence [51]:

F ( p , ,
K p ) F ( p , ,
K
= (
+ - z)p ,zp ), 0 z
n 1 n n 1 1 1 n n 1. 6.11

This condition of fragmentation invariance is, of course, well-known from the studies of IR safety in the
context of QCD [2], [51]. In particular, if a QCD observable is to be IR finite for a multiparticle (quark
and gluon) final state order by order in perturbation theory, then in addition to fragmentation invariance,
the functions Fn should satisfy some minimum regularity requirements [51]. It is sufficient that each Fn
has bounded (not necessarily continuous) derivatives with respect to its arguments:

f (x) - f (y) M x - y , 6.12

for any sufficiently close pair x, y. Fragmentation invariance plus the above regularity amount to IR
safety in the usual sense.
However, such conditions impose no restriction on how fast the functions Fn vary as n . In par-
ticular, there is no limit on how fast the derivatives of Fn grow as n . In the notations of 6.12, this
means that M may depend on n, and M may grow arbitrarily fast as n .

A continuous and perturbatively IR safe function that is not C-continuous 6.13

Here is an example of a fragmentation invariant function on final states whose component functions
vary arbitrarily fast. Consider the following function:

F P = F P

bad ( ) ( ), 6.14
m= m
2

with Fm defined by 5.18, and
min ,
1 / z
f ( $p ,K, $p ) = cmin ,1 /z
h 1
m 1 m ij m , 6.15
1i< jm z
where >
ij is the angle between $
pi and $p j and zm 0 . The function 6.15 nullifies when any pair of its
arguments is collinear so that the series in 6.14 is truncated, therefore well-defined, and satisfies 6.12, in
each order or perturbation theory. It follows that it is perturbatively IR safe.
To see that it is not necessarily C-continuous, consider a one-particle final state P = {E, $}
p . One can

see that F P
bad ( ) = 0 . Now fix an infinite sequence of states Pn such that each of them (i) has the total
energy exactly equal to E; (ii) consists of n pairwise non-collinear particles with directions localized
within a cone of angular radius n-1 around $p . Then P P
n in the sense of C-continuity. However, the
values F P
bad ( )
n are expressed in terms of F m n
m , , and choosing zm 0 for m fast enough,
one can ensure that the values F P
bad ( )
n do not converge to F P
bad ( ) = 0 .
Although one is not likely to encounter an infinite sum of C-correlators like 6.14 in applications, the
moral of the example is that manipulations involving limiting procedures may violate C-continuity. This
may be the case e.g. with maximization/minimization with respect to a parameter which affects the rate
of variation of the functions describing angular dependences (cf. the dependence on zm in 6.15).


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 42






Aspects of theoretical calculations 6.16

The two general issues to be considered here are standard perturbative calculations and the problem
of power corrections. One should also distinguish two types of observables: the basic class of relatively
simple C-correlators (Sec. 5.1; the most important example of C-correlators are the jet-number discrimi-
nators, Sec. 8) and the complex observables such as spectral discriminators (Sec. 10).

Analytical calculations of C-correlators 6.17

Consider theoretical calculations of such observables as the jet-number discriminators of Sec. 8. In
general, if one deals with hadronic processes and non-4 detector geometries, the perturbative calcula-
tions should follow the standard scheme: first one evaluates the matrix elements, and then one performs
integrations (via Monte Carlo) over the phase space. However, the highly regular analytical structure of
jet-number discriminators opens a possibility of their direct analytical evaluation -- at least for the
kinematically simplest but fundamental annihilation process e+e- hadrons.
Indeed, the expressions of the jet-number discriminators Jm , Eq.8.8, differ from the case of total
cross sections only by the weights that have to be inserted into the phase space integrals. The weights
are built of scalar products (cf. 13.4), and their form is such that calculations of a large number of dia-
grams is greatly simplified: due to masslessness, 2 2
p p = ( p + p
i j i j ) , and if partons pi and p j resulted

from a decay of another virtual parton, the denominator of the propagator of the latter is canceled.
There are still powers of energies in the denominator (the factors ( p P)-1
i in 13.4) but their dependence
on partons' momenta is linear which in some cases simplifies calculations. Of course, there are also dia-
grams in which the above cancellation does not occur, but nevertheless the situation here is rather less
difficult than in the case of other shape observables that can only be handled numerically.
Another nice feature here is that since there are no phase space cutoffs involved, the standard renor-
malization group equations are sufficient to perform resummation of logarithms -- neither Sudakov re-
summations are necessary, nor factorization theorems much different from those for total cross sections
[51], [55].
It is interesting to compare the case of jet-number discriminators Jm with the case of total cross
section + -
tot (e e hadrons) which is nothing but J2 (up to a numerical coefficient). For this
quantity, analytical calculations were pushed through next-next-to-leading order [33]. That feat was
made possible by the calculational methods [34], [35], [36] that reduced the calculation to a rather me-
chanical feeding of the corresponding diagrams into a computer [56]. Unfortunately, the methods of
[34], [35], [36] are based in an essential way on the possibility to Wick-rotate the unitarity diagrams in
the case of + -
tot (e e hadrons) ~ J2 into Euclidean region. But that is impossible (at least in a
straightforward manner) for J >
m , m 2, so the algorithms of [34], [35], [36] cannot be employed.
Therefore, although the cancellations mentioned above actually make calculation of quite a few dia-
grams for J > + -
m , m 2, easier than for J2 ~ (e e )
tot hadrons , nevertheless one should not expect
the NNL-order calculations to be doable in an entirely analytical fashion. But a combination of analyti-
cal and numerical techniques (cf. e.g. [57]) may work although the calculations remain very hard.
Note that for the purposes of precision measurement of s, it should probably be amply sufficient to
have NNL corrections to J3 (i.e. three terms in the QCD expansion in s). The NL corrections for
J4 are of about the same calculational complexity while the leading term for J5 should actually be
simpler. Notice that the QCD expansion for J + -
m , m > 2 in the case of e e hadrons starts at
O m
( -2
s ) ; the values of J4 and J5 can be used as a check for the rate of convergence of perturba-
tion series.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 43






Power corrections 6.18

As was discussed in [37], power corrections in the physics of jets seem to be numerically important
(although, strictly speaking, it remains to be seen whether they will have the same numerical signifi-
cance for, say, the jet-number discriminators). On the other hand, a theoretical study of hadronization
effects reduces simply to (i) computing logarithmic corrections (via higher order perturbative contribu-
tions as discussed above) and (ii) studying and estimating power corrections to Jm . Our formalism of-
fers an option for a theoretical study of power corrections.
Recall the case of e+e- hadrons where one has J
2 tot , and where the structure of tot in-
cluding power corrections is well known. In this case tot can be connected via a dispersion relation
with the vacuum average of a chronological product of two currents, so that its asymptotic behavior is
connected to the Wilson operator product expansion [58]. Then power corrections emerge as vacuum
averages of local operators of the Wilson expansion (the so-called vacuum condensates) with perturba-
tively calculable coefficients. This was used for phenomenological purposes in [59].
From a theoretical viewpoint, it was demonstrated in [60], [61] that such vacuum condensates are di-
rectly related to soft singularities of the expansion of the correlator in quark masses. Moreover, the
structure of condensates can be determined explicitly even within perturbation theory [60]. The vacuum
condensates are perturbatively uncalculable but they may be studied e.g. using the methods of lattice
QCD. On the other hand, the coefficients with which the condensates enter the expression for the cross
section, and which contain all the dependence on the energy of the process, are perturbatively calculable
(cf. e.g. the two-loop calculations of [63]).
A similar procedure can be used to study the structure of Jm for m > 2: One would start with the
corresponding perturbative expressions (a sum over Feynman diagrams with massive quarks etc.), and
perform the expansion in powers and logarithms of quark masses. The non-analytic contributions asso-
ciated with IR singularities (soft and collinear) would be organized into some kind of operators (cf. the
formalism and results of [61]).
Note that the presence of phase space weights in the expressions for jet-number discriminators does
not allow one to perform the Wick rotation and reduce the integrals to Euclidean ones in the usual way.
Therefore, one would need a non-Euclidean extension of the method of asymptotic operation used in
[61]. Such an extension is feasible [62], [64].


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 44






Computing C-continuous observables from data 7

In the preceding section we discussed theoretical calculability of C-correlators. In this section we
consider the problem of their optimal computation from data.i The point is that although the defining
formula 5.18 is simple, the volume of arithmetic involved may be rather large. For instance, for a final
state that lit up 200 calorimeter modules, computation of the 5-th jet-number discriminator J5 (Eq.8.8)
involves adding ~3109 terms, each containing 5 energy and 10 angular factors. Therefore, a straight-
forward approach may be unacceptable. However, one should bear in mind the following:
 Such large numbers of particles in the final state are more typical for future colliders such as
LHC, and by then computing power will become many times cheaper.
 The quality and amount of information from C-observables is higher than in the case of the con-
ventional processing. The resulting safety margin leaves room for approximations.
 The very regular analytic form of C-correlators 5.18 allows a number of optimizations.

In what follows we concentrate on purely analytical optimizations and ignore programming tricks
(e.g. parallelization). Also, an ideal optimization method may depend, in general, on whether the event
one deals with is "fuzzy" or has needle-like jets, etc., so that a sophisticated computational scheme may
depend on a concrete data sample. Therefore, only some typical options are described below.

Understanding the problem 7.1

We are going to investigate optimal ways to compute the value F P
( ) of a C-correlator 5.18 on one
given event P. The event is represented by a data record of the form 4.18, and each calorimeter module
is treated as a particle. The correctness of this is ensured by the C-continuity by definition of the latter.

Sources of optimizations 7.2

In the case of C-observables there are three (groups of) properties that can be made use of for optimi-
zation:

1) The C-continuity (recall that it determined the energy dependence of C-correlators 5.18) which
comes in two flavors (cf. Sec. 4.24):

1a) Stability (i.e. continuous variation) with respect to almost collinear fragmentations. The small pa-
rameter that can be utilized here is a small angle between the fragments.

1b) Stability with respect to adding soft particles, i.e. analyticity in particles' energies. The small pa-
rameter here may be taken to be the total energy fraction of soft particles.

2) Information on a concrete form of the angular dependence, i.e. on the function f K
m ( $
p , $
1 m
p ) .

3) Information on the structure of the event P.

The properties 2) and 3), we only touch upon. The main focus below is on the C-continuity.

Criterion of optimization 7.3

Unlike the conventional paradigm based on jet finding algorithms where observables are defined in
terms of the output of such algorithms, in our theory observables are defined directly in terms of un-
processed events (final states). Therefore, the issue of what an observable should be is cleanly separated
from how it is to be computed. The first advantage of such a separation is that once the observable is
explicitly defined, one is free to use any tricks in order to compute it (including jet algorithms). The sec-
ond advantage is that one now has a clear and unambiguous criterion to choose among optimization
tricks: The criterion is the numeric quality of the resulting approximation.

i There is no need to consider general C-continuous observables here because they are built from C-correlators.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 45






Soft particles 7.4

We begin by considering the property 1b) from Sec. 7.2. We will first show how the analyticity of C-
correlators at zero particles' energies can be exploited to obtain expansions in energies of soft particles.
Then we consider an important special case -- the problem of estimating effects of undetected particles
(the so-called missing energy; Sec. 7.14). The very fact that the latter problem can be meaningfully ap-
proached in the framework of the new formalism is a clear demonstration of its analytical superiority
over the conventional scheme.

Expanding in energies of soft particles 7.5

The expression 5.18 allows one to perform Taylor expansion in energies of particles. Divide all par-
ticles of the event into two groups, "soft" and "hard":

P = Psoft Phard . 7.6

Then one can expand 5.18 as follows (e.g. using 6.7):

F P
( ) F P
( ) F( )
1 P
( ;P ) F(2)
= K
hard + hard soft + P
( ;P )
hard soft + , 7.7

where F P
( )
hard is the value of the observable F computed on the state Phard that consists of only "hard"
particles, and

F( )
1 P
( ;P ) = m E KE E f ( $p , ,
K $
p )
i i , 7.8
m
hard soft 1 1
- i m
m 1
i im
i , ,
K i P
i P

1 m 1
- hard m soft


F(2) P m m 1
( ;P ( )
) = - E KE E E f ( $p , ,
K $
p )
i i . 7.9
m
hard soft 2 1 -2 im 1
- i m
m 1
i im
i , i
K P
i ,i P

1 m-2 hard m 1
- m soft

To see the computational savings, assume the numbers of hard and soft particles are equal. Then
computing just the first term in 7.7, i.e. F P
( )
hard , involves 2m times fewer terms than the complete cal-
culation. Taking into account the first correction doubles the number of terms, which still means 2m 1
-

times fewer terms than the complete calculation.

Estimating the error 7.10

As a simple example, consider the error involved in retaining only the first term F P
( )
hard on the
r.h.s. of 7.7. The error can be estimated from the second term given by 7.8. There is a spectrum of pos-
sibilities for writing a bound for it. The following inequality seems to be adequate:

F( )
1 P P m E P F[ ]
( ; ) ( ) 1 P
( )
hard soft miss hard hard , 7.11

where the C-correlator E is defined in 5.7,

( soft
= E P )
miss , 7.12
E( h
P ard )

and F[ ]
1 P
( )
hard is the value on the hard subsystem Phard of the following C-correlator:

F[ ]
1 P [ ]
( ) 1
= E KE f ( $p , ,
K $
p ),
1
j jm 1
- m 1
- 1
i im 1
-
j , ,
K j P
m
1 1
-
[ ]
1
f ( $p , ,
K $
p ) f K 7.13
m m- = sup ( $p , , $p , $p).
1
- 1 1 m 1 m 1
-
$
p


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 46






It is assumed here that the angular function fm is non-negative; cf. Sec.5.15.
Similar estimates can be obtained for higher terms.

Errors due to missing energy 7.14

The factor m in 7.11 indicates that the higher C-correlators are increasingly more sensitive to the er-
rors of such an approximation. In particular, they are increasingly more sensitive to errors due to unde-
tected particles.
To estimate the error due to missing energy, i.e. the particles that escaped undetected, one simply
uses 7.11, treating all missing particles as "soft" irrespective of their actual energies. (The estimate 7.11
does not assume that the individual particles are soft in any sense.) The missing energy enters the r.h.s.
of 7.11 only via . All other quantities are directly measurable.

Missing energy for jet-number discriminators 7.15

The above estimates did not employ any concrete information on the angular function fm. Let us
show how such information can be made use of. As an example, we will estimate the effect of missing
energy for the case of the jet-number discriminators 8.8. (Their concrete physical meaning is of no im-
portance here.) We are going to obtain estimates of a type somewhat different from 7.11 for the defect
defined as follows:

def
def J P
( ) = J P
( ) - J ( >
hard
P
m m m ) 0 . 7.16

It is important here that all terms in the expression for J P
m ( ) 8.8 are non-negative.
The reasoning below will be in terms of energy fractions rather than absolute energies (cf. the nor-
malization in 8.8).
Assume that the total energy fraction of missing particles soft (it is always less than miss , Eq.7.12)
is distributed uniformly over the unit sphere and their total number is large. Then an estimate can be
obtained by replacing the summation over the soft particles by an integration over the unit sphere. For
definiteness, consider the 3rd jet-number discriminator J P
3 ( ) . The resulting expression is as follows:

def 27 1 2
J P =
i
E Ej ij  + i
p p j + O
3( ) 1
( $ $ ) ( ) , 7.17
4 soft 3 soft
i< j

where the summation is restricted to "hard" particles. The factor in parentheses is between 2/3 and 4/3.
If one simply estimates it by 1 (numerical experimentation may suggest a different value), one obtains:

def 2
J P + O
3 ( ) 3.4 soft ( ) . 7.18
soft

Similarly,

def 2
J P
( ) 5 3
. + O
soft J ( h
P
4 3 ard ) ( ). 7.19
soft

One could use the expressions thus obtained as corrections to the numbers computed from data in order
to reduce the effects of missing energy.

Preclustering 7.20

Let us turn to the property 1a). It can be directly utilized to derive an important approximation that
consists in recombining (groups of) particles of a given final state into one provided this affects the ob-
servables to be computed only within specified errors. We call this preclustering, and it bears a resem-
blance to the conventional jet algorithms. But there are also differences which it is important to
elucidate.
The preclustering is simply a computational approximation trick aimed at reducing the amount of


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 47






arithmetic needed to compute C-continuous observables by exploiting the property of C-continuity
shared by all C-continuous observables.i Its numerical effect, however, depends on the form of a con-
crete observable; for some observables it works better than for others. Our preclustering has a mathe-
matically well-defined purpose and involves a parameter (see below) that directly controls the resulting
approximation errors. Therefore, all arbitrariness is to be judged against the well-defined criterion of
whether or not one can achieve the desired precision for specific observables using available computer
resources.
The small parameter which the preclustering exploits is the small angle between almost collinear
fragments. However, the specific range of what constitutes "small" angles will be seen to depend on the
energies of the particles. This should be no surprise because the energy dependence of C-correlators is
fixed and known (cf. 5.18).
Because we will be considering here C-correlators that are uniform functions of energies, we can take
all particles' energies Ei to be energy fractions rather than absolute energies:

E E / E
i i tot . 7.21

Such an assumption entails no loss of generality.

On masslessness of pseudoparticles 7.22

Before we turn to formulas, the following subtlety should be emphasized. The preclustering replaces
one energy flow (as defined in our formalism) with another object of exactly the same type. Therefore,
the pseudoparticles that emerge in our case (and that are analogous to the protojets of the conventional
algorithms) are also to be formally interpreted as massless. But one should not try to assign a profound
physical meaning to this fact because we are simply dealing with a computational trick here.
On the other hand, in the conventional algorithms protojets may emerge with non-zero masses. Incor-
porating non-zero pseudoparticles' masses within our formalism would be equivalent to constructing
approximations for an observable not in terms of the same observable -- which is the case that we are
considering -- but in terms of a different one. This may be an interesting option, but we only concen-
trate on the optimizations that do not require a detailed knowledge of the structure of observables. In
that sense, our formulas are universal.
One also notices that the recombination is defined unambiguously only for infinitesimal angles be-
tween particles (cf. below). The energy and momentum conservation of the conventional algorithms can
be regarded as ways to extend the recombination criteria to finite angular separations while preserving
maximum information about the event.

Estimating errors induced by recombinations 7.23

Let us begin by considering the simplest C-correlators that are linear in energies. This is meaningful
because the general C-correlators 5.18 can be obtained from these using algebraic combinations and
appropriate limiting procedures (Sec. 5.22). We will see that the criterion we are going to derive remains
essentially the same in the general case.
Consider the following correlator:

F P
( ) = E f ( $p )
. 7.24
i i i

On a state P = p P with one particle, p, singled out, the correlator becomes

F( p P) = E f ( $p ) + E f ( $p )
. 7.25
i i i


i In fact, experimentalists routinely use a similar preclustering of the raw data except that jet-algorithm-based
observables are computed in the end instead of C-continuous observables, and using criteria different from the
optimal one explained below in this section. I thank D. A. Stewart, Jr. for explaining this to me.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 48






Replace p with two particles p and p
a b :

F( p p P) = E f ( $p ) + E f ( $p ) + E f ( $p )

a b a a b b . 7.26
i i i

To compare 7.26 and 7.25, consider their difference:

F( p p P) - F( p P) = E f ( $p ) + E f ( $p ) - E f ( $p )
a b a a b b . 7.27

At this point, it is convenient to formally extend the function f to non-unit 3-vectors by imposing the
condition

f ( $) = f ( $), $
q q for any q. 7.28

Assuming that the angles between each of the fragments and the initial particle p are small, we are go-
ing to Taylor-expand the expression 7.27 in $ $
p and p
a b around $p , using the following formula:
f ( $ 1 2
p ) = f ( $
p) $
+ p f ( $p) + ( $
p ) f ( $p) +K
# # # . 7.29
2

(Because one deals with vector arguments here, the products are tensorial.) We obtain:

(E + E - E) f ( $p)
a b
+ ( $ $
E p + E p ) f ( $p)
a a b b
+ E ( $
2 2 1
p ) + E ( $
p ) f ( $p) + K
a a b b 7.30
2

Requiring that the first two lines nullify, we obtain:

E + E = E
a b , 7.31

E $ = E $p + E $
p a a b b
p . 7.32

The latter relation cannot be satisfied by three unit vectors -- but that is of little import here: First, the
relation is only meant to be exact for infinitesimal angles, and its extension to finite angles involves an
arbitrariness that cannot be avoided in principle. Second, the trick of defining f for non-unit vectors,
Eq. 7.28, offers a simple variant of such an extension, and it allows one to simply enforce a correct
normalization whenever necessary.
The remainder is bounded by the following expression:

1 2 2
E | $
p | +E | $
p |  M , 7.33
2 a a b b f

where M
f 0 involves a maximal value of 2nd order derivatives of f (we will not need its precise
form). Notice that
E E
$ b a
p = ( $ $
p - p ) , = - -
a a b $p ( $ $
p p )
b a b . 7.34
E E
Finally:
E E
F( p p P) - F( p P) a b
 1 | $ $
- |2
p p M
a b . 7.35
2 a b f
E

Modifications for general C-correlators 7.36

The above reasoning is extended in a straightforward manner to the case of the general C-correlators
5.18. (This is done easiest using the representation 6.7.) One obtains
-
E E m 1
F( P) - F( P) a b
 1 | $ $
p - p |2 k (k )
p p p  E M , 7.37
2
a b a b
E f
k =0


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 49






where M(k) are constants that depend on second order derivatives of f . Because E < 1, this reduces to
f

7.35 (with a different M f ). Therefore, the form of dependence on energies and angle in 7.35 is com-
pletely general -- as was to be expected in view of the construction of multilinear observables from lin-
ear ones described in Sec. 5.8.

The optimal preclustering prescription 7.38

Now suppose we wish to simplify computation of some C-continuous observables via reducing the
number of particles by combining a pair of particles, Ea, $pa and Eb, $b
p , into one, E, $
p . From 7.35 one

sees that the criterion to combine two particles into one is

E E
a b  < y
ab cut , 7.39
E

where

=
def 1 | $p - $ 2
p | = 1 -
ab a b cos ab , 7.40
2

and the parameter ycut effectively controls the error induced thereby. The actual error differs from ycut
by a factor which depends on the observable and is best determined via numerical experiments.
The energy of the new particle is determined from 7.31. Its direction is found from 7.32 (with the
normalization enforced as described after 7.32).
Recall that one deals with energy fractions in 7.39 (cf. Eq. 7.21).

Comments on the definition of ab 7.41

The optimal preclustering criterion 7.39 has been rephrased, as compared with the underlying ine-
quality 7.35, in terms of the angular separation ab defined in Sec.4.5. Eq.7.40 involves an extension
to large ab which in principle is arbitrary, whereas the bound 7.35 was derived only for infinitesimal
ab. The considerations behind the definition 7.40 were discussed in Sec.4.5.
It should be clearly understood that such an extension, theoretically speaking, can be chosen arbitrar-
ily as far as preclustering is concerned. Indeed, whereas in the conventional approach the radius of jets
(or ycut which effectively controls it) is physically significant (because e.g. reconstructed masses of
multijet substates -- therefore, masses of new particles -- depend on it), in our formalism the preclus-
tering is only an approximation trick (e.g. the spectral discriminators described in Secs. 1011 will have
a bump due to a new particle at the right value of invariant mass irrespective of whether or not the pre-
clustering is used). Therefore, if the arbitrariness in the definition of the angular separation turns out to
be numerically important for the chosen ycut, one should conclude that the approximation errors due to
the preclustering are out of control and ycut should be reduced.
Another point one should have in view is that the vector norm on the r.h.s. of 7.35 can be any vector
norm, not necessarily the Euclidean one. This means that any alternative ab such that

c < < c , for some c , >
1 ab ab 2 ab 1 2 0 , 7.42

is allowed. The preclustering criterion 7.39 (and also the criteria discussed in Sec. 7.47) based on such
ab instead of ab is equivalent to the one based on ab but may result, e.g., in a faster code (cf.
Sec. 13.13).
In general, the overriding consideration for the choice of the angular separation used in the precluster-
ing algorithms is computational simplicity.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 50






Comparison with the conventional algorithms 7.43

It is interesting to treat the above "optimal" algorithm as a conventional jet algorithm because one
can use the criterion 7.39 in a Luclus-type iterative recombination scheme [14]. (Note, however, that a
better way is to use the n 1 variant of preclustering described in Sec.7.47.) Various recombination
algorithms currently in use were discussed in detail in [30]. Our criterion 7.39 directly compares with
the JADE criterion [28]:
E E  < y
a b ab cut , 7.44

and the Geneva criterion [30]:

E E
a b  < y
ab cut . 7.45
E2

Recall that the energy squared in the denominator of 7.45 was introduced [30] to avoid problems with
soft particles encountered by the JADE algorithm; its role is to provide a repulsion that ensures, in par-
ticular, that no "lumps" (i.e. spurious jets consisting of widely separated soft particles) are produced.
Since there was no a priory criterion to choose the form of such a factor, the squared energy of the pair
was taken to match the dimensionality of the numerator. Note that we use energy fractions so the l.h.s.
of the optimal criterion 7.39 is made dimensionless by Etot in the denominator rather than an additional
E as in the Geneva criterion 7.45.
Note also that the Luclus criterion (eq. (38) of [14]) is equivalent to E E E-1  < y
a b ab cut at small
angles but involves absolute (unnormalized) energies.
From the point of view of the errors induced in the "physical information" (i.e. the collection of all C-
correlators; cf. Sec. 5.21), neither Geneva nor Luclus criteria err because they overestimate the errors
( E-2 E -
> 1 for E < 1, and >
ab ab for small angles, respectively). But for the same reason, they
are not optimal: both are overcautions with certain pairs of particles.
For two soft widely separated particles (cf. sec. 2 of [30]) our criterion results in a cutoff of the form

minaE , E f ~ y
1 2 cut , 7.46

which resembles the Durham algorithm [29] except that the latter involves squares of energy fractions.
Because C-continuous observables are IR finite, our algorithm does not give rise to IR divergences
if used for jet counting in a conventional way. This can be verified directly (cf. the reasoning in sec. 2
of [30]).

Preclustering n 1 7.47

Because C-continuity allows fragmentations into any number of particles, it is natural to extend the
above preclustering to more than two particles. The modifications of the reasoning of Sec. 7.23 are as
follows. One compares the values of C-correlators on two multiparticle states, p P and ( p )
a a P.
One forms their difference and Taylor-expands it in angular variables. One finds that the leading terms
cancel if

E = E
, $ E
$
p p . 7.48
a a a a a

The remainder is bounded by an expression of the form 7.33. From 7.48 one finds:

$ $ $ 1
p = p - p = E ( $ $
p
- p )
a a , 7.49
E ba b a b

2
-
1 2 (E E )
| $
a
p |  max
a . 7.50
2 2 ab
E b


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 51






One finds:
E ( 2
a E - E )
F(( p ) a
a a P) - F( p P) max
ab  M f . 7.51
E2 b
a
The corresponding clustering criterion is as follows:

2
a
E (E a
E ) max ab y
- cut . 7.52
E2 b
a

This is more precise than the following simpler but less optimal version described in [19], [20]:

E (E - E )
max a a
 2
ab y
2 cut . 7.53
,
a b E
a

Note that M f in 7.51 is the same as in 7.35. Therefore, the numerical value of the cutoff parameter here
is to be taken the same as in the case of two particles. (For two particles, Eqs. 7.52 and 7.53 coincide
with 7.39.)
An even simpler and cruder version is as follows:

max 
ab E ycut . 7.54
,
a b


This is because E - E E
a .
The advantage of the simpler version 7.54 is that it is easy to accumulate the particles to be clustered
iteratively: One may, e.g., begin with an energetic particle and proceed by adding other particles -- re-
jecting particles that are too energetic -- until the threshold is achieved. In the end, one may revert to
the more precise formula 7.53 or 7.52, which may be needed to take into account accumulation of errors
correctly (cf. the discussion in Sec. 7.55 below).
The simplicity of 7.54 reminds one of the cone-type algorithms traditionally used in hadron collisions
(for a review see [10]). In particular, the quantity max ab corresponds to the angular diameter of the
a,b

jet (also notice the presence of an interesting energy factor E which is the energy of the protojet normal-
ized by the total energy of the event). However, in our case one worries neither about overlapping cones
(as with cone-type algorithms) nor about irregularity of jets (as in the case of recombination algorithms)
because the errors induced by preclustering are under control by analytical means, so the issue of jet
shapes is irrelevant.
I emphasize that there can be no restriction on the geometric form of the cone spanned by the parti-
cles being clustered as long as the above analytical criteria are met. For instance, two particles that are
too energetic to be recombined into one can, nevertheless, attract all softer particles around them even
from behind each other and even if the angular distances of the latter from the former are much larger
than the distance between the energetic particles. The only criterion of clustering is the numerical quality
of the resulting approximation.
Lastly, it might also be possible to develop variants of preclustering with more than one resulting
pseudoparticles (cf. the 3 2 scheme of [65]). Since elimination of a particle in such a scheme would
take into account more information about the final state, one might be able to achieve a better quality of
the resulting approximation (e.g. eliminate or reduce the quadratic terms in 7.30).


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 52






Accumulation of errors 7.55

Errors induced by several (say, n) instances of preclustering -- no matter how many particles are re-
combined in each, and irrespective of whether or not the particles affected are pseudoparticlesi -- are
accumulated. For example, if all n instances of preclustering have been done so that the error induced
by each is ~ y 
cut , then the cumulative error might be estimated by ~ n ycut . (This issue should be
further studied via numerical experimenting.)
But whatever law governs the accumulation of errors, the very fact of their accumulation implies that
an actual preclustering algorithm should, perhaps, somehow take the fact into account, e.g. start with a
preset allowance Y > 0 which is decreased after each instance of preclustering until it reaches zero. How
much of Y is used up in successive instances of preclustering (whose corresponding ycut may all be dif-
ferent), is up to the programmer's ingenuity. Note also that Y may be chosen differently for different
events (the harder the event for computations, the larger Y one may be willing to allocate to it, and vice
versa). Optimal criteria for this depend on concrete samples of events.
Note that the values of ycut (and Y) used in the preclustering should typically be smaller than in the
case of conventional algorithms.
Finally, it should be noted that we have only derived a criterion for an optimal preclustering. A com-
plete algorithm should efficiently find the clusters of particles to combine into a protojet. The simplest
variant is to use the code for a usual recombination (2 1) algorithm and replace the criterion by the
above optimal one. One could also use more sophisticated schemes similar to [40]. In general, different
implementations may not be identical and may have to be fine-tuned for a particular application.






i Note that the n 1 recombinations may eliminate the need in iterative preclustering (i.e. such in which pseu-
doparticles are further clustered along with the unclustered particles from the initial multiparticle state).


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 53






Applications to measuring the "number of jets" 8

There are two main uses for the quantity known as the "number of jets" of an event. It can be used
either to compute relative fractions of events with a given number of jets -- the observables called n-jet
fractions -- or as a tag to select events when searching for new particles. In this section we derive a
special sequence of C-correlators -- the jet-number discriminators [15] -- that can be used similarly.
Their properties will be studied in Sec. 9 while an example of their uses as building blocks in more
complex observables can be found in Sec. 13.32. Since jet counting depends on the kinematics of the
concrete reaction, here we consider the simplest case of annihilation e+e- hadrons with a full 4
detector. Modifications for hadronic collisions are discussed in Sec. 13.

"Jet counting" with C-correlators 8.1

Understanding the problem 8.2

We would like to construct C-continuous observables that quantify the qualitative feature of final
states known as the number of jets. A jet, qualitatively, is a spray of particles radiated within a small
solid angle (in a given reference frame) and carrying a substantial fraction of the total energy of the
event. The conventional approach consists in a straightforward formalization of this description which
results in a use of cuts and an integer-valued observable (the "number of jets").
First of all, it should be emphasized that the linguistic restrictions of the C-algebra (which, by con-
struction, reflect the limitations of multimodule calorimetric detectors) do not allow one, as a matter of
principle, to write down an expression for the "number of jets" because the latter is an integer number
and the only integer-valued C-continuous observables are constants on the entire collection of final
states. This, however, is in perfect agreement with the fact that the "number of jets" is an ambiguous
notion for a non-negligible number of final states. On the other hand, one should draw a distinction be-
tween a qualitative feature of the physical phenonema one studies and its numerical expression in terms
of an underlying theory. Such an expression need not be a straightforward definition of what one feels
one sees but it must conform to the kinematical ("linguistic") requirements of the theory. For instance,
the "position" of an electron is desribed by a vector in Hilbert space -- a notion not quite exactly intui-
tive but precise. The liberties one allows oneself to take with kinematical/"linguistic" restrictions are in-
versely proportional to the precision one aims to attain.
It proved possible to construct a sequencei of C-correlators of a special form (the jet-number dis-
criminators [15]) that quantify the "number of jets" without actually identifying individual jets. The
purpose of the detailed derivation presented below is to present concrete motivations for every element
of their construction in order to show that it is essentially unique. (In the simplest and cleanest situation
of e+e- annihilation into hadrons the qualifier "essentially" seems to be superfluous.)

Defining multijet states in terms of C-continuity 8.3

The following assumption deserves to be explicated: An "m-jet state" is one that is "similar" to a
state which contains exactly m energetic particles with large angular separation. One cannot fail to see
that the crucial point here is what "similar" means, exactly. It is natural to express it in mathematical
terms using the C-convergence: at a qualitative level, two states are "similar" if they are indistinguish-
able by calorimetric detectors with poor enough energy and angular resolutions.
Unfortunately, the C-convergence cannot be expressed in terms of a useful distance function. There-
fore, there is no single numerical criterion to measure the above "similarity". But fortunately, the wis-


i The sequence might be thought of as an infinitely dimensional vector. Recall that some physical quantities are
described by 4-vectors, some by 6-component antisymmetric 4-tensors, etc. The "number of jets" happens to be
correctly described by an infinite sequence of scalar continuous components.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 54






dom of general topology [46] tells us that the most general option to discriminate the elements with a
particular property in spaces with unusual convergences is to employ continuousi functions that take
special values (e.g. 0) on those elements. The number of the functions used in such a comparison and
their form depend on the specifics of the particular problem as well as practical expediency. A natural
thought would then be to construct a function that takes, say, the value 0 on m-jet states, and is appre-
ciably different from zero on all states that cannot be described as "similar" to m-jet states.
Unfortunately, the notion of "m-jet state" is ill-defined (if one wishes to avoid cuts, as we do). But
fortunately, a little further thought reveals that the definition of m-jet states does have an unambiguous
aspect, namely, that a state consisting of exactly m particles cannot have more than m jets. The latter
property ("cannot have more than m jets") is shared by all states consisting of no more than m particles,
and the collection of all such states is defined unambiguously. Any state that is sufficiently close to this
collection (in the sense of C-convergence) shares the same property -- it cannot have more than m jets
(within the precision of "sufficiently").
Therefore, the problem of "jet counting" has transformed into that of constructing C-continuous
functions that take a special value (zero) only on all multiparticle states with less than a certain number
of particles.
Going back for a moment to the conventional jet counting, one can see that instead of one integer-
valued function ("number of jets") one could use a sequence of step functions: each such function
should take the value 0 on all states with less than a certain number of jets, and the value 1 on all other
states. A sequence of such functions (m = 1 2
, ,K) would do the job of jet counting just fine. The func-
tions we are going to construct can be regarded as C-continuous weights that replace such step func-
tions in accordance with the philosophy of Sec. 2.5.
Among all C-continuous functions, it is natural first to try the basic ones -- i.e. the C-correlators
5.18 -- as candidates for that role.

Explicit formulas for jet-number discriminators 8.4

Consider a C-correlator 5.18 that is exactly 0 on any state with less than m particles. Then
f p p K = K
m ( $, $, ) 0 . Then f p p
m ( $ , $ , )
1 2 should contain a nullifying factor (cf. Theorem 119 in [45] which
states that if a sufficiently smooth function f (v) is zero on a manifold described by the equation
(v) = 0 then f (v) = (v) f (v) where f (v) is also a smooth function). We choose the nullifying factor
to be = -
12 1 cos12, an object that has already occurred in the study of optimal preclustering,
Eq. 7.34.
The nullifying factor 1 - cosij is a rotationally invariant function that is analytic everywhere on the
unit sphere and is simply connected to scalar products, which makes it a perfect choice. It also seems to
be the only reasonable choice -- apart from raising it to some power, a complication so substantialii that
one should have very good reasons to justify it.
Symmetry requires a similar factor for each pair of arguments of f :
~
f ( $p ,K, $p ) f K
= 
m m ij m ( $
p , , $p
1 1 m ) , 8.5
1i< jm
where
= 1 - cos = 1 - $ $

ij ij i
p p j . 8.6
In what follows it will be important that
ij ( ij ) is non-negative, monotonic and smoothly interpo-
lates between (0) = 0 and () < +.

i with respect to the concrete convergence defined for the elements of that space.
ii from the point of view of analytical calculations. This is because of a combinatorial blow up of the number of
terms at intermediate stages; cf. the analytical calculation of normalizations in Sec. 8.12.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 55






~
To fix fm, it is natural to check out the simplest option first, so set
~
f =
m 1. 8.7

Since the perceived jet structure is independent of the energy scale (which is reflected in the fact that the
corresponding amplitudes scale up to logarithmic scaling violating corrections), it is natural to make the
observable dimensionless by introducing a factor E-m
tot , which is equivalent to using energy fractions
instead of absolute energies.
One arrives at the following expressions for the jet-number discriminators [15]:

J P -m
= K K K
tot p p p p =
m ( ) Nm E i
E i
E jm ( $i , , $i ), jm( $ , , $m ) 8.8
1 1 1 ij .
m m
i1<K i
<
1 i< jm
m

The summation here runs over all selections of m different particles from all the particles of the final
state P. The terms with equal angular arguments are absent here as compared with 5.18 because they
are nullified by the 's.
The normalization Nm in 8.8 can be chosen so that the discriminators always vary between 0 and 1
(see Sec. 8.12):

0 J P
( ) 1, for all and P
m m . 8.9

For m = 1, 2 the discriminators are trivial:

J P -1
N Etot
1( ) = E
1 1, 8.10
i i

s
J P 2
= -
N E
tot E E -
2 ( ) 2 1
( cos ) , 8.11
i< j i j ij E2
tot

where s is the total invariant mass of the event. In the center of mass reference frame, J P
2 ( ) 1.

Normalization 8.12

The jet-number discriminators 8.8 are always bounded:
F I
J P -m
K K
tot
m ( ) Nm E G i
E i
E sup j p p 8.13
1 J 1
m ( $ , , $m ) const.
m
H < < K
1 K K
i i $
p , , $
m 1 pm

In general, Nm can only be found numerically because they depend on the geometry of the experimental
setup (e.g. whether or not one deals with a full 4 detector, etc.). For the case of e+e- hadrons and
full 4 geometry, numeric experiments show that the maximal value is reached on the configuration
Psym with the energy uniformly spread over the unit sphere (as well as on a few highly symmetric con-
figurations; cf. Sec. 8.17). It can be regarded as a limit of final states consisting of N particles with
equal energy E = N -1 N
i uniformly distributed over the unit sphere so that i zS i
4 d $p as N .
2
Then one can choose Nm from the condition

J Psym
m ( )
= 1 . 8.14

The angular integrals are done using the following formula:

1 1
z d$p ($ $pq )K($ $pq )= K , 8.15
K ($ $
q q )   $
1 2k ( $
q - q )
4 1 2 2 1 2
2
S 3 5 1
( + 2k k k
)
where the summation runs over all non-equivalent decompositions of 2k objects into k pairs
(decompositions differing by the order of pairs or the order of objects in each pair are treated as


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 56






equivalent). One finds:

m 1 2 3 4 5 6

N 27 9375 455625
m 1 2 36 ... 8.16
4 32 128

Note that a very large number of terms is generated here at intermediate stages of analytical calcula-
tions.i,ii

Some special values 8.17

To illustrate how the discriminators approach their maximum (= 1), here are the values of J Psym
d i
m n

on some highly symmetric configurations of particles with equal energies ( n = 4 corresponds to vertices
of a tetrahedron inscribed into the unit sphere; n = 6 corresponds to an octahedron; the case n = is
defined in 8.12):

J J J J K
1 2 3 4

Psym 1 0 0 0 0
1

Psym 1 1 0 0 0
2

Psym 27
1 1 0 84
. 0 0
3 32

Psym 64
1 1 1 0 79
. 0
4 81

Psym 1 1 1 1 K
6

K K K K K K

Psym 1 1 1 1 1
8.18

(Recall that, up to normalizations, J1 and J2 are the total energy and the invariant mass of the state, re-
spectively. Actually, J P
1( ) 1 in the spherically symmetric case, and the corresponding column is in-
cluded for completeness.) The symmetric states have the largest angular distances between the particles.
The table shows that the uniformity of the angular spread together with the uniformity of energy distri-
bution are more important factors for achieving the maximal value than the total number of particles.
Note that the state Psym must be considered as having an infinite rather than zero number of jets.
Note also the following experimental fact: all global maxima of the jet-number discriminators (i.e. the
states on which at least one Jm besides J1 takes the value 1) correspond to the following eight configu-
rations: the five Platonic solids (the tetrahedron Psym, the octahedron Psym, the cube Psym, the icosahe-
4 6 8

dron Psym, the dodecahedron Psym; see e.g. [67]) as well as the sphere P
12 20 sym and the two configurations
that can be regarded as degenerate regular polyhedra, namely, the `dipole' Psym and the `dihedron'
2

Psym. A failure to include the latter three degenerate configurations into the analysis might explain the
3
limited (~ 5% ) precision of an earlier purely classical attempt [68] to find a kinematical explanation of
the well-known profound significance of the Platonic solids.





i I failed to find a way to evaluate N7 with my copy of FORM-2 [64].
ii I thank B.B.Levtchenko for some numerical checks of the normalizations.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 57






Understanding the jet-number discriminators 9

For definiteness, we consider here the standard jet-number discriminators Jm exactly as defined in
Sec. 8.4 for the 4 geometry parametrized in terms of two angles (the unit sphere; modifications to the
cylindrical geometry of hadron-hadron collisions will be discussed in Sec. 13). But we allow the total
3-momentum of the states to be non-zero. The qualitative features observed in the examples remain
valid in more general cases.

Qualitative behavior of Jm 9.1

We will first explain how to obtain crude estimates for Jm , and then discuss in detail various features
of their qualitative behavior. Examples for states with a few particles are presented in Sec. 9.6.

Estimating Jm via a crude clustering 9.2

The jet-number discriminators -- as any other C-correlators -- are fragmentation invariant
(Sec. 6.6). Therefore to obtain a rough estimate for the values of the discriminators on a given final state
it is sufficient to replace each jet with one particle carrying the same total energy and going in the direc-
tion of the jet and drop soft particles. (A more refined version of this procedure can be used to optimize
computation of C-correlators from data; see Sec. 7.)

A typical picture of values 9.3

The values of Jm for a typical final state P with 3 jets are roughly as follows :

Jm
1



ycut

0 1 2 3 4 5 m 9.4

The white circles represent the values of discriminators for the 3-particle state P0 obtained by recombin-
ing particles from each jet into one, as described in Sec. 9.2. P0 can be interpreted as the parton state
prior to hadronization; then P is the corresponding hadronic state. As a rule, J P
( ) > J P
m m ( )
0 as a re-
sult of fragmentation. The C-continuity ensures that the closer (in the sense of C-convergence) the final
hadronic state to the underlying parton state, the less the upward drift of Jm during fragmentation.
In particular, Jm ~ 0 for m larger than the perceived number of jets; the non-zero values (the tail at
large m) are due to fragmentation (almost collinear radiation from hard partons as well as "drops of
glue" between jets).
Note, however, that since the maximum is reached on some highly symmetric configurations (cf.
8.17), fragmentation of the latter results in a downward (rather than upward) shift of the values of Jm .
But such cases are exceptional.

Understanding the decrease at large m i 9.5

The monotonic decrease at large m is explained as follows. Let M and be, respectively, the per-
ceived number of distinct jets in a state and their average (small) angular width. The state may also


i I thank B.Straub for this question.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 58






contain a "soft" component that consists of particles with small total energy soft spread between the
jets. Then one can see from the explicit expression 8.8 that for m > M , each term in the sum for Jm is
suppressed by additional powers of 2 and soft (cf. the examples below). The larger m, the larger the
number of such factors, which explains the decrease.
Numerical experiments show that the decrease of J <
m is a universal feature even for M m
~ .i,ii

Simple examples 9.6

As explained above, up to a small upward shift and a non-zero tail at large m, the values of Jm are
roughly the same as for the states with particles clustered into the corresponding jets. Thus, it is useful
to consider states with just a few particles, each representing a jet, which we do below. One-jet states
being too simple to be interesting, we begin directly with two jets.

Two pure jets 9.7

Here one deals with two hard widely separated particles:
P = k , $ p
p ; , $p , >> ,
, >>
2 1 1 2 2 1 2 0 12 0
h E E E . 9.8

Our usual notation is = 1 - cos = 1 - $ $

ij ij i
p p j . Due to the normalization of Jm all energies are to be
compared with Etot , the total energy of all particles of the event. One has:

E E
J P 1 2
( ) = 4 , J P
( ) = J P = K =
2 2 ( ) 0.
h 9.9
2 12 3 2h 4 2h
Etot

One can see that the conditions 9.8 that ensure that we deal with two jets are equivalent to one condi-
tion, namely, J P >>
2 ( 2 ) 0
h . Vice versa, it is sufficient to require that

J P <
2 ( 2 )
h ycut 9.10

with a small ycut to obtain a state with one jet. (For a comparison of our formalism and conventional jet
algorithms see Sec. 7. Here we only note that the resemblance of the above to the Geneva criterion,
Eq. 7.45, is due to the fact that there are just two particles so that E E1 + E
tot 2 in this simple case.)

Adding one soft particle 9.11

Here one deals with a state

P P
k , $ p
p ,
2h 1s 2h 3 3 3 1
E
+ = << ,2 , 9.12

that can be shown as follows:


E 12
2

E1
3 9.13

One has ( E E1 + E
hard 2 ):



i I thank B.B.Levchenko for numerical checks of this property.
ii The relationship between jet clustering algorithms and event shape measures, including the monotonicity, is
discussed in [65].


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 59






2
E
J P
( ) J P hard a f
2 2h 1s 2 ( 2 ) 2
h E E
+ = + + , 9.14
2 3 1 13 2 23
Etot
E2
J P 27
( ) J P hard 3 13 23
3 2 1 ( )
h+ =  
s , 9.15
8 2 2h E2 E
tot tot

J P
m ( 2h s ) , m
1 0 4
+ . 9.16

Eq. 9.14 illustrates the fact that one can take into account soft particles via an expansion in their energy
(cf. Eq. 9.21 and Sec. 7.4). Eq. 9.15 shows that if the soft particle becomes almost collinear to any of the
hard ones then the restriction on its energy is relaxed.
For completeness, here is the expression of J3 on a state with exactly three hard particles:

J P 27
= -3
3( 3 )
h E E E E . 9.17
4 tot 1 2 3 12 13 23


2 hard + 2 soft particles 9.18

As a further example, add another soft particle:

P P
k , $ p
p ; , $
p ,
2h 2s 2h 3 3 4 4 3,4 1
E
+ = << ,2. 9.19

This can be depicted as follows:

E
2 12
E1

3 4 9.20

Consider again the most interesting case of the third discriminator:

2
E
hard +
J P 27
( ) J P 3 13 23 4 14 24 2
+
3 2 + =  
2 ( ) (
O ).
h s 9.21
8 2 2h 2
E E
tot tot

This expression should be considered in view of what was said in Sec. 9.2. In particular, we wish to see
whether the pair of widely separated soft particles could imitate a jet by pushing the value of
J P
3 ( 2 +2 )
h s well above 0. Comparing 9.21 with 9.15 and taking into account the fact that ij const,
one can see that this does not occur as long as the total energy fraction of the soft particles remains
bounded by a small value.

Jm never see spurious jets 9.22

This issue was discussed in the literature [31], [30] in connection with the fact that some recombina-
tion algorithms (e.g. JADE) suffer from the problem of counting "lumps" instead of jets. A lump is a
state consisting of soft particles with large relative angles but erroneously identified as a jet. It is inter-
esting to consider whether the problem of lumps reemerges in the context of the jet-number discrimina-
tors (and C-correlators in general). If this were so, then the values of C-continuous observables would
be affected by "lumps" in a numerically significant way.
Recall that the energy dependence of C-correlators in general -- and Jm in particular -- is analytic,
which means that an expansion in powers of soft energies is always possible. Moreover, it is always
possible (cf. the examples 9.14, 9.15, and 9.21) to obtain estimates for the terms that are linear in ener-
gies of the soft particles (as well as for higher terms) by a power of soft (the total energy of soft parti-
cles) times a factor that is independent of the geometry of the state. We see that the problem of lumps in


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 60






the context of C-correlators reduces to that of determining a correct measure of "softness". The latter
turns out to be given by the total energy fraction of the soft particles: The values of C-correlators are
guaranteed (by their C-continuity) to remain within a given small error interval if the total energy frac-
tion of the soft particles that are added to (or eliminated from) the state, is bounded by the correspond-
ing small value .

Width of jetsi 9.23

The problem of the width of jets in the conventional approach arises due to the fact that observables
such as mass spectra of multijet substates depend on the control parameters of jet finding algorithms
which (parameters) correspond -- directly or indirectly -- to the angular width of jets. Since in the pre-
sent formalism all observables are redefined in such a way that the intermediate representation in terms
of jets is avoided, the issue does not emerge in such an interpretation. Nevertheless, as a further example
of expressive power of the C-algebra, it is interesting to find a C-continuous measure for what could be
called "width of jets".
The preceding examples motivate the use of ratios of jet-number discriminators for that purpose.
Indeed, consider the ratios

J P
( )
W P m
= +1
m ( ) . 9.24
J P
m ( )

From the explicit expressions 8.8 and the examples above one can see that if there are M m distinct
jets in the event then the above ratios are suppressed by powers of angular distances squared between
particles in the same jet, and/or by powers of energy fractions of soft particles. The latter fact together
with the dependence of the ratios on the global geometry of the event, does not allow one to interpret
them straightforwardly in terms of average width of jets.ii Dynamical width, therefore, is a better name
for such observables. Their concrete form is chosen solely from the consideration of simplicity (one
might e.g. consider taking square root of the r.h.s.).
The averaging over events is done after evaluating the ratios, which ensures that Wm carry new in-
formation as compared with the jet-number discriminators Jm . In particular, the sensitivity of Wm
to hadronization is different.

Jm vs. jet counting 9.25

The conventional jet counting assigns an integer number of jets to each event and classifies the events
accordingly. Fix a multiparticle state P and consider any jet counting algorithm A that yields an integer
number of jets N y
A ( ; )
cut P for each ycut ; N y
A ( ; )
cut P should decrease monotonically as ycut 0:






i I thank E.Kushnirenko for suggesting this problem.
ii Note that there is actually no point in trying to reproduce some arbitrary -- however visual -- definition of
"jet width". Human eye was created by Nature for purposes other then studying multiparticle systems. "Physical
meaning" is not the same as semiclassical visualization. The correct apparatus of "vision" in high-energy jet
physics is C-correlators. This also answers an objection sometimes raised that a jet-number discriminator takes
different values on different states that "apparently" have the same number of jets. Just because one can "see"
that the states have the same "number of jets", can hardly be regarded as a truly physical reason to count them
with the same weight in an observable that is supposed to be sensitive to "jettiness".


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 61






N m
A


5

4
4

3 JA
2
2

1

0
y A
J
cut m 9.26

(The hashed areas correspond to experimental and theoretical uncertainties.) One sees from the figure
that one could, in theory, restore a sequence of jet-number discriminators JA P
m ( ) similar to J P
m ( ).
Therefore, the information contents of JA P
m ( ) and N y
A ( ; )
cut P for one event are essentially equivalent.
The difference emerges when one performs an averaging over many events and takes into account errors
(Sec. 9.28).
It is, of course, hardly possible to find simple expressions for JA P
m ( ) for the popular algorithms. Our
J P
m ( ) are singled out by the transparency of their analytical structure.

Counting jets with Jm 9.27

If the physical contents of the conventional jet counting and the discriminators are equivalent, then it
should be possible to use the latter for jet counting of a conventional kind. Let us discuss this briefly.
Actually, it is interesting to consider a jet-counting-type procedure not for its own sake but rather in
connection with the composite observables discussed in Sec. 13 as an auxiliary tool for their approxi-
mate computation.
Recall that the state P contains exactly M particles (up to exactly collinear fragmentations) if and
only if J P >
m ( ) 0 for all m M but not for m M . This follows from the construction of the discrimi-
nators Jm . Therefore, J P
m ( ) >> 0 indicates that there are no less than m jets in P, while J P
m ( ) ~ 0
implies that there are less than m jets in P.
Due to fragmentation one has to introduce a small "resolution parameter" y >
cut 0 . Then the state P
is said to have M jets if J P = +
m ( ) < ycut for m M 1 but not for m M (cf. Fig. 9.4). The jet counting
now proceeds in a sieve-like manneri: one first computes J P
2 ( ) (in the center of mass frame one starts
with J P
3 ( ) because J P
2 ( ) 1). If J P
2 ( ) < ycut then P contains just one jet. Otherwise one computes
J P
3 ( ) . If J P
3 ( ) < ycut then one deals with two jets... etc.
In general there is no reason why the cut should be the same for all m, so one would (and probably
should) use a sequence of cuts ycut,m -- i.e. a separate parameter for each m.

Instability of the conventional jet counting 9.28

Fig. 9.26 is an illustration of the difference with respect to experimental data errors and the unknown
higher order corrections between the conventional jet counting, on the one hand, and the jet-number dis-
criminators, on the other hand. (From the discussion in Sec. 9.25 it is clear that our conclusions will be
valid for any jet counting algorithm of a conventional type.) Suppose the errors are purely statistical and
distributed with a given variance. In the case of jet-number discriminators, if the statistics is increased,
the statistical error of the results goes to zero.

i Following the custom of giving jet counting algorithms geographic names (Geneva, Durham), the procedure
being described was called `Moscow sieve' in [16].


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 62






The conventional jet counting corresponds to putting events into different bins depending on their jet
number. Due to errors, some events would be assigned a wrong number of jets and go into a wrong bin.
Therefore, each bin would have a certain fraction of events from other bins, depending on the width of
the error intervals only -- increasing the statistics cannot help. Thus, a statistical error is transformed
into a systematic one. Although such a transformation occurs for any non-linear function of a random
variable, the analysis of Sec. 2.5 shows that the effect is expected to be alleviated in the case of C-
correlators as compared with the discontinuous mappings of conventional algorithms.i
Note that changing the jet resolution parameter ycut does not help one to get rid of such a smearing.
Lastly, the conventional jet counting is particularly sensitive to the errors at large m / small jet resolu-
tions. One source of such errors is the so-called Sudakov radiation of almost collinear partons. The
above analysis suggests that the importance of Sudakov effects within the framework of the conven-
tional jet counting is an artifact due to the instability of the latter.

Jm vs. n-jet fractions 9.29

Recall the interpretation of continuous observables as weights measuring the content of a particular
physical feature in final states (Sec. 2.6). Also recall that Jm is a continuous analogue of a step function
that would take the value 0 on final states with less than m jets and 1 otherwise. It follows that the aver-
age value of Jm over all events is naturally interpreted as the "weight" of the m -jet component in the
entire ensemble of final states. Then the average value Jm is a natural replacement for
m jets m+ j
1 ets
+ +K . The quantity m Jm is then naturally interpreted as the average multiplicity
of jets.





Take a Break....






i It was noticed by F.Dydak that for jet discriminators, the energy factors in each term in the sum are independ-
ent measurements from independent detector modules, and their errors are also independent. As a result, there
is no systematic shift due to nonlinearity with respect to individual energy measurements. The property is also
true for the C-correlators from which the spectral discriminators are built (see below).


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 63






Applications to mass measurements.
Simplest spectral discriminator 10

In the preceding sections we have constructed a correct numerical description of the feature of multi-
hadron final states usually referred to as the "number of jets". The description involves only continuous
shape observables from the C-algebra (the sequence of jet-number discriminators J =
m , m ,
1 ,
2 3K)
while identification of individual jets is avoided. Yet we have seen that the physical content of the new
observables remaines equivalent to that of the conventional n-jet fractions. What did change was the
mathematical expression of that physical content.
Now we are going to perform a similar transformation for the other important class of applications of
jet algorithms, namely, searches for new particles based on studying invariant masses of multijet sub-
states that resulted from the decays of those particles. We are going to reexpress in the correct language
of C-algebra the observables such as invariant mass distributions of n-jet substates. Needless to say,
identification of individual jets will be avoided.
In this section we define and study in detail spectral discriminators that accomplish that task in the
toy model of a particle decaying into one jet and, as an immediate extension, in the more realistic case of
decays into 1 jet + "muon" (Sec. 10.30; by "muon" we mean any particle that can be reliably indenti-
fied, using calorimeters or not, including its energy and direction). A generalization to the case of multi-
jet substates and the corresponding computational procedures will be described, respectively, in Secs. 11
and 12.

1-jet spectral discriminator 10.1

Let us begin with a very simple situation where one studies masses of just one-jet substates. Much of
the reasoning will remain the same in the general case.
Suppose there is a particle X that decays into one jet. Then the invariant mass of the particles consti-
tuting the jet is equal to the mass of X. For simplicity, assume first that X is the only particle produced
in the collision. Then the final state consists of just one jet from X. How can one determine the mass of
X using the tools we have at our disposal?
We have seen (Sec. 5.25) that to select a substate from a final state within our formalism one uses the
so-called filters -- continuous functions interpolating between 0 and 1 in different parts of the unit
sphere. For each filter one obtains a substate o P, and evaluates its invariant mass S( o P) ,
Eq. 5.28. (It is convenient to say in such a case that one uses a mass detector described by the filter .
Since one can analyze other properties of the substate selected by the filter, one can talk about ET -
detectors, J3-detectors etc.)

Elementary mass detector 10.2

In order to be able to select a jet substate, consider a mass detector centered at the point $
q with angu-

lar size O(R) (measured in units of the angular separation ; cf. Secs.4.5 and 7.41). The corresponding
filter is defined as follows:

=
$ p , 10.3
; ( $ ) ( / 2 )
q R qp R

where qp is the angular separation between the directions $q and $p ,

(d) = 1 if d ,
1
= 0 if d 2,
= 2 - d otherwise , 10.4


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 64






and the factor 2 is introduced into the r.h.s. of 10.3 to make R vary between 0 and 1.
The continuous linear interpolation between 1 and 0 in 10.4 regulates the cut in accordance with the
philosophy of Sec. 2.15. The concrete form is chosen from considerations of simplicity. Note that for
R = 1 the filter is identically equal to 1: $; ( $p )
q R=
1 1.
One can, of course, determine the mass of X by simply looking at an event, seeing what the position
and size of the jet are (i.e. $
q and R ), and putting the mass detector described by 10.3 over the jet, after

which one obtains the invariant mass. But such a procedure involves a step that cannot be expressed in
the language of C-observables; namely, there is no recipe to choose one position and size for the filter
from the family 10.3.
Anyhow, choosing one such filter involves a comparison between all filters. Therefore, consider all
filters 10.3, and for each position $
q and size R , compute the invariant mass according to 5.28:


S( $
q, R) S ( o P
$ 0 . 10.5
q, R )

Let us examine the properties of this expression as a function of $
q and R in order to determine what

sort of information it can yield.
For all mass detectors that cover the jet completely, S( $,
q R) is the same and equal to the jet mass.

This can only occur for R larger than the jet's size. The larger R, the larger the number of the positions
$
q for which the measured mass is equal to that of the jet. On the other hand, for R smaller than the jet's

size, the detector and the jet overlap only partially, and S( $
q, R) yields values that are spread below the

jet mass in a more or less continuous fashion.

Definition 10.6

We are thus naturally led to consider the distribution of the values of S( $,
q R) over the real half axis

parametrized by a variable, say, s. A straightforward way to do so is to consider the following expres-
sion:

def 1
- $
= a
q - q )
z
1 jet (s; )
R Z d s ( $
S , R) , 10.7

which we call one-jet spectral discriminator.
The dependence on R carries important information about the width of the jet, so it would be a good
idea to leave it as a free parameter -- one can always integrate over it later. (Notice certain parallels
between the above logic and the cone-type jet algorithms. For a more detailed comparison see
Sec. 11.25.)
It is convenient to choose the normalization in 10.7 so that
+
z ds jet(s;R)=
1 1. 10.8
0

In the spherically symmetric case Z = 4 .

Simple properties of 1 jet (s; )
R 10.9


For brevity, we will be writing (s; )
R = (s; )
R
1 jet in what follows.

Small mass detector ( R 0 ) 10.10

Infinitesimal mass detectors measure infinitesimal mass so that all spectral weight is near the origin:

(s; )
R (s) . 10.11
R0


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 65






Large mass detector ( R 1) 10.12

When the mass detector covers the entire sphere -- which corresponds to R 1 -- all the spectral
weight is concentrated at the point s = Stot , the total invariant mass of the state:
(s; )
R (s - S )
tot . 10.13
R1

Flow of spectral weight 10.14

Obviously, the regions of small and large R are physically uninformative. But as R varies from 0 to
1, the spectral discriminator undergoes an evolution that can be described as a flow of spectral weight
from s = 0 to s = Stot . Indeed, the normalization of the spectral discriminator 10.8 does not depend on
R, and each position of the mass detector contributes an infinitesimal unit of spectral weight, the total
number of such units being fixed and independent of R. As R increases a little, so does, in general, the
mass measured by the mass detector in the same position. This means that the infinitesimal weight from
that position is shifted a little towards higher s. Ultimately, all the spectral weight concentrated initially
at s = 0 is collected at the point s = Stot . But what happens in between?
We will see that localized clusters of particles in the event manifest themselves as `barriers' for the
flow of spectral weight that take a form of isolated -functions (-spikes) in spectral discriminators, but
first we consider the opposite extreme case.

Uniform distribution of energy over the sphere 10.15

In this degenerate case the spectral flow is simple. In any position, the mass detector measures the
same invariant mass, so that all spectral density is localized at the same value s = SR which is a mono-
tonic function of R. The spectral flow in this case is visualized in the following figure:

1
R = 0 R = 1
R




0 S S s 0 S S s
R tot R tot 10.16

The left figure shows three positions of the -spike corresponding to three values of R; the height of the
-spike does not change as R varies. The right figure shows the trajectory of the -spike on the sR
plane.
On the other hand, for a non-uniform energy distribution different positions of the mass detector
measure different values of s for a fixed R. Therefore, the spectral weight is spread over s for a fixed R
(a horizontal spread along each line R = const in the right figure in 10.16). In what follows we study
the qualitative features of spectral flow with emphasis on the case of "extremely non-uniform" energy
distribution, i.e. when the final state contains a few clear well-separated jets.

Spectral density is localized within 0 s smax S
R tot 10.17

This is obvious. Moreover, the boundary value smax monotonically increases with R.
R

Zero-mass -spike for 0 < R < 1 10.18

This is a simple illustration of some of the features encountered in less trivial cases.
If R is small enough, there will be many positions $
q of the mass detector where it will measure zero


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 66






mass -- because the mass detector will be covering just one or no particles. Denote as R the collec-
0
tion of all such positions, and let |R| be its normalized surface:
0

| R | 1 $
z d
0 4 R q. 10.19
0


If |R|> 0, the discriminator (s; )
R contains a contribution of the form |R| (s) . It is clear that the
0 0
coefficient decreases,

|R| 0 , 10.20
0 R 1


monotonically for all R, starting from the initial value 1.

Jets and -spikes 10.21

Events with a single jet 10.22

Suppose the event contains just one jet of particles. If all particles are strictly collinear then mass de-
tectors always measure zero mass, and (s; )
R (s). Therefore suppose there is a non-zero angular
spread of particles of the jet. Its invariant mass Sj is the total invariant mass of the state Stot . If the
mass detector is not large enough to cover the entire jet then the mass of the latter will be underesti-
mated in all positions of the mass detector, and the spread of (s; )
R does not extend far enough to the
right to reach the boundary s = Sj .
But it comes closer as R increases and reaches s = S -
j when R reaches a critical value Rj (which

roughly corresponds to the angular size of the jet). At R = R-j there is just one position of the mass de-
tector when it measures the correct mass.
Once R passed the critical value, there will be many positions around the jet (the wider the mass de-
tector, the larger the number of such positions) the correct mass is measured. Let Rj be the part of the
unit sphere consisting of all positions of the mass detector in which it measures S < -
j . Then for R Rj the
spectral discriminator contains an isolated -function,

|R| (s - S ) . 10.23
j j

The weight |R| R -
j (= normalized surface of j ) grows monotonically from 0 to 1 as R varies from Rj

to 1 while the position of the spike on the s-axis does not change and no spectral weight moves above it.
Evolution of the spectral discriminator in this case as a function of R can be visualized as follows:


0 < - < <
~ R R ~ R
j R ~ 1




0 s
S 0 S 0 S
tot tot tot 10.24

The fat vertical lines represent the -spikes. The shape of the continuous component depends on the
distribution of energy within the jet.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 67






Events with several well-separated jets 10.25

Let the index j label the jets of the event, so that Sj are their masses. The situation is similar to the
single jet case in the sense that a contribution of the form 10.23 will show up for R > R-j . However,
since there are other jets in the event, then starting from some value Rj (dependent upon the distance
from the jet j to other jets) there will be fewer positions in which the mass detector covers only the jet
j , and there will also be an increasing number of positions in which the mass detector feels particles

from other jets and measures masses larger than Sj. This means that for R R
j ~
> j the spectral weight
accumulated at s = Sj will start to spill over to the right from that point. Finally, there will be another
critical value R+ > +
j such that for R Rj the mass detector is too large to cover only the jet j -- then the
-spike at s = Sj completely disappears and all its spectral weight is pushed above its position:

      





      
- < +       
R < R R < <
j ~ j R R R
j ~ j +       
R       
      
j       
      
      





R       
j       
      





-       
      





R       
j       
      





s
0 S 0 S s
j j Sj 10.26

The two left figures show two points in the evolution of the spectral discriminator. The right figure
shows the spectral flow as a density on the sR plane. The vertical fat line represents the -spike; its
variable width corresponds to |R|
j , the size of the coefficient of the -function.

If there occurs an overlap of a few -spikes, the picture will be more chaotic. This may often happen
in the 1-jet case because all jets may be expected to have roughly the same size determined by the dy-
namics of hadronization -- but not in the case of 2 -jet substates whose invariant masses are scattered
over a much larger interval.

-spikes corresponding to groups of jets 10.27
As the size R of the mass detector is increased, it may cover more than one jet. Such a cluster of jets
will also show up as a -spike in (s; )
R -- but for larger R than for individual jets.
It is not difficult to realize that choosing a better shape for the mass detector (e.g. a `sum' of two
elementary detectors Eq. 11.15; cf. Sec. 11.14) would allow one to focus much better on clusters of jets
-- and therefore on particles that decay into jets. This is exactly the idea behind the definition of more
complex spectral discriminators in Sec. 11.13.

Spurious -spikes 10.28
In certain exceptional ideal situations (e.g. a wide jet with energy uniformly distributed over the jet
cone) the mass detector for smaller R may have a freedom of movement within the jet while measuring
the same mass (which is a fraction of the jet mass S -
j and approaches the latter as R Rj from below).
Then there will be a spurious -spike. Such exact spikes, however, are practically improbable and
should rather be expected to materialize as an enhancement advancing towards Sj from the left for

R < R-
~ j .
Minijets will be represented by minispikes for small values of R.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 68






Summary. Evolution of a -spike 10.29

We have seen that the presence of a well-separated jet in the final state is manifested in the spectral
discriminator as a -functional contribution (-spike) with a specific dependence on R. This dependence
can be visualized as a flow of spectral weight as R changes. The flow is in the direction of larger s. The
jet corresponds to a barrier for the spectral flow at s = Sj .
The relevant parameters are the jet's invariant mass S - +

j and the three values R R R
j j j (cf.

Fig. 10.26). As R approaches R-j from below, the continuous component of the spectral density extends
to the right and approaches s = S - =
j . When R passes Rj , there develops a -spike at s Sj , and the spike
increases as R increases from R-
j to Rj -- the spectral weight accumulates at the barrier. When R ex-

ceeds Rj, the accumulated spectral density spills over the barrier and begins to spread continuously to
the right of the barrier while the -spike at s = S +
j shrinks. At last when R passes Rj , the -spike at the

barrier disappears and is replaced by a distribution above s = Sj .
The characteristic form of the -spike (shown as a density in the rightmost figure 10.26) is expected
to persist in the more general cases of spectral discriminators for multijet substates where its position
will correspond to invariant masses of multijet substates.

Realistic example: X jet + "muon" 10.30

This case is actually very similar to the simplest 1-jet spectral discriminator. Suppose the particle X
decays into one jet and a muon . (By "muon" we mean any particle that can be reliably indentified,
using calorimeters or not, including its energy and direction.) Let E, $
p be the muon's energy and di-

rection. The muon is treated as a non-calorimetric "external parameter" (recall the remarks in Sec. 4.8).
It is sufficient to make only one change in the formulas of Sec. 10.1, namely, for each filter one should
"measure" the total invariant mass of the hadronic substate selected by the filter (as usual, the hadrons'
energies Ei enter all expressions with the weights $; ( $p )
q R i ) and the muon -- irrespective of whether

the muon's direction is covered by the filter or not:

S( $
q, R) = 2  1 -
$ ( $p ) $ ( $p ) E E ( $ $
p p )
q;R i q;R j i j i j

1 i< j N

+2 E  1 -
$ ( $p ) E ( $ $
p p ). 10.31
q;R i i  i

1 i N

Then one defines:

def -1 $
= q ( - q )
z
1 jet+ (s ; R) Z d s S( $, )
R . 10.32

Because the structure of this expression is similar to 10.7, the only difference from the simple 1-jet casei
is that the position of -spikes is shifted to higher s. The characteristic shape of the density distribution
in the rightmost figure 10.26 is not affected.

Smearing and accumulation of -spikes 10.33
In actuality, the -spikes from X's decays will get smeared into a more or less wide bumps for vari-
ous reasons (e.g. averaging over all final states and non-zero width of the particles that decay into jets;
overlaps of -spikes; the numerical procedures described in Sec.12 that require one to deal with spe-


i apart from the unphysical region of very small R which is left for an interested reader to investigate as an ex-
ercise.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 69






cially chosen continuous replacements for exact spectral discriminators). But apart from non-zero
width, the qualitative behavior of the smeared -spikes is expected, due to C-continuity, to retain a re-
semblance to the right figure in 10.26 as long as there remains a resemblance to the ideal final states for
which the behavior of -spikes was established.
Suppose now the sample of events one studies contains both QCD background jets and jets from the
decay of X. (The reasoning is the same whether one considers the decay into just one jet or jet +
"muon".) Then averaging of the discriminator over all events results in an accumulation of -spikes
from X at about the same value of s while the background jets are spread in a more or less uniform
fashion over a wide interval of s. Therefore, the signal from X will appear as a bump against a continu-
ous background. To enhance the signal, one may integrate over R but then the potentially important in-
formation would be lost (the two tails in the right figure in 10.26). Whether or not the characteristic S-
shaped density distribution on the sR plane will be preserved after averaging, depends on how clean
and well-isolated the jets from X are, etc.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 70






General spectral discriminators 11

In Sec.10 we have studied the simplest spectral discriminator for 1-jet substates, (s; )
R (s; )
R
1 jet .

We have found that, for a single event, it may contain isolated -functions (-spikes). Each such -spike
corresponds to an isolated jet -- or a cluster of jets if the -spike occurs for larger R. We have also ob-
served that modifying the configuration of the mass detectors employed should allow one to focus better
on multijet substates. The purpose of this section is to discuss such modifications in more detail. First in
Sec. 11.1 we consider a general definition of spectral discriminators which is a direct extension from the
1-jet case. Then in Sec. 11.13 we present a simple concrete version for n-jet substates in the spherically
symmetric case of e+e- hadrons. Modifications for hadronic collisions will be considered in Sec.13.

Mass detectors and spectral discriminators 11.1

One considers a family of filters describing a family of substates o P (the continuous parame-
ter describes the configuration -- size and shape -- of the filters; in general, contains several scalar
components, see examples below). Each filter is thought of as corresponding to an ideal mass detec-
tor that measures the invariant mass of the substate it selects, S( P)
o ; cf. Eq. 5.28.i Then one con-
structs, following the pattern of Sec. 5.35, a differential observable with respect to the invariant masses:

( ;sP) = Z-1
d s - (
S
z o
c P)h. 11.2

This is a continuous sum of -functions and does not, in general, reduce to an ordinary function (cf.
Sec. 11.11).
Observables 11.2 will be referred to as spectral discriminators. Their normalization can be defined
so that
+ds (s;P) 1
z = . 11.3
0

Changing the spectral variable 11.4

If one chooses to work with true masses M = s instead of s, the -function in 11.2 is transformed
as follows:

s - ( o ) = (2M)-
S P 1 M - S( o
c h c P)h. 11.5

On the other hand, spectral discriminators are measures with respect to s and one should take into ac-
count a Jacobian:

(s; )
R s
d = 2M (M 2; )
R M
d . 11.6

The net effect is that the extra factors cancel out:

(M;P) = Z-1
d M - (
S
z o
c P)h. 11.7

One may wish to make a non-linear change of the variable so as to focus better on a particular range
of s. This is because the algorithms described in Sec. 12 use a uniform discretization of s and a non-
uniform discretization can be best achieved via a non-linear change of variable.






i Note that one can study characteristics of multijet substates other than masses in a completely similar manner.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 71






Continuous component of ( ;
s P) 11.8

Recall that in 11.2 is a continuous parameter that describes the shape (configuration) of the filter
( $p). In general, has several scalar components, and runs over a multidimensional manifold (e.g.
a direct product of unit spheres; cf. 11.14). To study the structure of 11.2 consider the equation

s = S P
( ; )
11.9

with given fixed s and P, and the unknown.
In a generic case, its solutions with respect to form a hypersurface s in . For such s, ( ;
s P) is an
ordinary function, and its values can be represented in the following form:
1
( ;
s P) ~ (
S P;
z -
d ) , 11.10
s

where d is the infinitesimal element of the hypersurface s (in natural units), and is the corre-
sponding gradient operator. The formula means that the larger s and the slower the variation of
S P
( ; )
on s , the larger ( ;
s P) .

-functional components of ( ;sP) 11.11

If, on the other hand, the values of that are solutions of the equation 11.9 for s = s constitute a part
of the manifold with a non-zero hypervolume
| | d >
z 0 , 11.12


then ( ;
s P) contains a contribution (s - s) . Such a situation occurs, e.g., if for some value = 0
the mass detector described by completely covers a cluster of particles (i.e. ( $p ) = 1 for each
0 0 j

particle of the group), and for small variations of in all directions around 0 the cluster remains cov-
ered so that the measured mass S P
( ; )
stays the same.
There is little one can say a priori about the continuous component -- its shape depends on the details
of shape and geometry of the final state. But about -spikes, a qualitative information can be obtained
by purely analytical means -- exactly as was done in Sec. 10: All the conclusions about -spikes and
their evolution made there remain valid in the general case. In particular, the characteristic shape of
the density on the s-R plane (Fig. 10.26) persists in the general case (cf. below Sec. 11.24).

Multijet spectral discriminators 11.13

Composite mass detectors 11.14

A composite mass detector can be regarded as consisting of two or more elementary mass detectors
described by 10.3. For instance, consider two elementary mass detectors of the same size R, centered at
$
q1 and $
q2 , respectively. The two functions are $ ; ( $p ) p
q1 R and $ ; ( $ )
q2 R . We combine them as follows:
=
$ ( $p ) ( $p ) . 11.15
, $ ; $
R ; $
R ; ( $
p )
q q q q
1 2 1 2 R

The operation takes two or more numbers as operands, and its exact form is not importanti as long as
it has the following properties:

(i) continuity (the result is a continuous function);

(ii) it should give the result from the interval [ ,
0 ]
1 if the operands are from that interval;

(iii) commutativity (the result is independent of the order of operands).



i To insist on a "physical" interpretation here would be inappropriate.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 72






The simplest choice is:

a b = max(a, b). 11.16

One can combine more than two mass detectors, e.g.:
=
$ ( $p ) ( $p ) ( $p ) , 11.17
, $ , $ ; $
R ; $
R ; $
R ; ( $
p )
q q q q q q
1 2 3 1 2 3 R

with the result independent of the order.
So, a composite mass detector is characterized by:

(i) n, the number of its constituent elementary detectors; (heuristically, this corresponds to the number
of jets in the substates the detector probes)

(ii) their positions (configuration) $ $
q , , q
1 K n . This is in terms of Sec. 11.1 and the manifold is the
direct product of n unit spheres; it can be parametrized e.g. by 2n angles;

(iii) their size R i (it is not included into "configuration" because we want to keep it a free parameter);

(iv) the filter:

(n) ( $p ) = ( $p ) L
( $p ). 11.18
$
K $
q , , $
q ;R q ; $
R q ;
1 n 1 n R


Example. Spectral discriminator for 2-jet substates 11.19

The following filter selects two clusters of particles:

( $ ) = (2)
p ( $p ) . 11.20
$
q , $
q ;
1 2 R

Evaluate

S( $
q , $
q ; R) (
S o P)
1 2 11.21

according to the definition 5.28. This is the invariant mass of the pair of clusters selected by . The
simplest 2-jet spectral discriminator is defined as follows (cf. 10.7 and 11.2):

def -1 $ $
= a f
1
q 2
q -
z z
2 jets (s; R) Z d d s S( $q , $q ; )
R
1 2 . 11.22

The normalization is such that
+ds (s;R) 1
z 2 jets = . 11.23
0

In the spherically symmetric case Z = (4 )2
.
Generalizations to 3-, 4-...jet spectral discriminators as well as to the cases such as 2 jets + "muon"
are straightforward.

Qualitative behavior of 2 jet ( ;
s )
R 11.24

In this case the mass detector is a `sum' of two elementary `modules', and its configuration is de-
scribed by the pair $ , $
q q
1 2 (four angles). An infinitesimal unit of spectral weight is now associated with
each such configuration.
A -spike occurs whenever the final state contains a pair of well isolated jets, and the size of elemen-
tary detectors is large enough to cover them; then the two modules can be moved independently around
the corresponding jets without changing the measured mass, which corresponds to the situation of
Sec. 11.11. The reasoning of Sec. 10.21 is repeated almost verbatim, and one concludes that the evolu-



i We only consider the case when all elementary mass detectors constituting a composite one have the same
size. The complication of different sizes would have to be well justified.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 73






tion of the resulting -spike as R changes follows the same qualitative pattern as in the 1-jet case
(Sec. 10.29). In particular, the -spike on the sR plot has a similar form with a characteristic vertical
line (-functional component) and two wider tails (cf. the right figure in 10.26). Of course, Sj is now
the invariant mass of the pair of jets.
If the final state has more than two jets, then 2 jet ( ;
s )
R exhibits a -spike for each pair of jets; e.g.

for a state with four jets, one would have 6 -spikes. They will be positioned on the s-axis at the invari-
ant masses of the corresponding 2-jet substates.
The remarks of Secs. 10.33 concerning smearing of -spikes in realistic situations and searches for
particles decaying into a certain number of jets, remain valid in the most general case.
A new point here as compared with the 1-jet case is as follows. Suppose the final state contains a pair
of jets that are not well separated from each other. Nevertheless, 2 jet ( ;
s )
R would exhibit a -spike at s

equal to the invariant mass of the pair. The contributions to the -spike come from those configurations
of the mass detector when the two elementary modules partially overlap while covering both jets. The
problem of separating the two jets does not arise here.

Comparison with the conventional practice 11.25

The spectral distribution from, say, 2 jet ( ;
s )
R after averaging over all final states is similar to mass

distributions for 2-jet substates obtained within the conventional approach. But there are also important
differences.
The comparison is easiest with the cone-type algorithms. In that case, the conventional approach
consists, essentially, in choosing one position for an elementary mass detector per each jet, for any given
R. After that is done, one value of invariant mass is computed per each 2-jet substate. Denote as S R
n ( )
the resulting values of invariant mass, where n numerates the 2-jet substates. One then computes the
distribution of the masses for the entire sample of events. This is exactly the same as if one replaced our
2 jet( ;s )
R with the following expression:

conv 1
( ;
s )
R = Z - as - S ( )
R f, 11.26
2 jet n n

where n enumerates the 2-jet substates as determined by the jet algorithm used. (The normalization
factor Z here is not the same as in Eq. 11.22.) Performing an averaging of conv over all events is ex-
2 jet

actly equivalent to computing the invariant mass distribution of 2-jet substates. (The fact that we deal
with invariant masses squared is not important; cf. Sec. 11.4.)
Comparing Eqs. 11.26 and 11.22 allows one to see exactly what is the difference between the con-
ventional mass distributions and our spectral discriminators.
Graphically, using the conventional conv instead of our
2 jet 2 jet is equivalent to replacing the density

distribution in the right figure of 10.26 with a line parametrized by

s = S (R
n ). 11.27

Now if jets are well separated, the curves 11.27 would have vertical segments that pass exactly over
(parts of) the -spikes of the exact 2 jet ( ;
s )
R (provided energy and momentum conservation is re-

spected when jets' 4-momenta are determined). But even then the corresponding -functions enter into
the r.h.s. of 11.26 with equal coefficients, which is not the case with the exact 2 jet ( ;
s )
R . In other

words, even in the best case important information about the event is lost. Moreover, in the conventional
approach the coefficients of -functions on the r.h.s. of 11.26 are independent of R. So if the curve
11.27 has no vertical segments due to ill-defined jets it is hard to choose one value for the value of the
mass. On the other hand, the -spikes (and the resulting bumps) of 2 jet ( ;
s )
R have a variable height


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 74






even for one event, which information can be used to better pinpoint the X's mass.

But the worst problem is encountered when the jets are not well-defined and the -spikes are smeared
and/or overlapi. Then a conventional algorithm attempts to represent a continuous distribution with a
few -functions; graphically, the smeared -spike of the right figure 10.26 is replaced with a curve. If
the continuous distribution has narrow peaks then such a representation may be meaningful. But if the
peaks are not narrow, then whatever method one uses to position the -functions, one cannot do that
stably. For instance, if one attempts to determine a weighted average point for the position of the curve
for each R, then the result depends on how one separates the bump for this pair of jets from other simi-
lar bumps, which involves an ambiguity. Alternatively, if one attempts to draw the line along the maxi-
mum of the bump, such a procedure is unstable against measurement errors (cf. the discussion in
Sec. 2.3, esp. Fig. 2.4). In either case, there occur "ambiguities" resulting in an enhancement of errors in
the determination of the mass of the particle one searches for.

On the other hand, when one computes a spectral discriminator, more information about each event is
involved: One computes a spectral weight distribution from each event which depends in a well-defined
C-continuous manner on the final state, and one fits the entire spectral weight distribution against theo-
retical predictions. At no point in data processing does one have to make unstable choices, and no in-
stabilities occur.

Discussion 11.28


(i) The pure QCD background contribution to spectral discriminators reflects the QCD dynamics of
production of high-mass virtual partons -- in fact, the structure of QCD matrix elements. Therefore,
measuring spectral discriminators (perhaps, modified by appropriate weights; cf. Sec. 13.30) may be
used as tests of QCD.
(ii) It is sufficient to perform the processing for the values of R only within an interval of typical
widths of jets resulting from the decay of X. An integration over an appropriate range of R may make
the bump more prominent against the continuous background, especially its maximum.
(iii) The overall picture (-spikes and their evolution on the s-R plane) in more complex cases (e.g. in
the case of 3-jet substates, or mixed cases like 2 jets + "muon" etc.) will be similar to what was de-
scribed in Sec. 10.29.
(iv) Recall the example of Sec. 10.30 (jet + "muon"). It does not matter what are, say, the experimental
cuts for the muon: one may simply sum contributions from all events -- whatever the energies of the
muon. The position of the bump relative to s should normally not be affected much by this.
(v) Similarly, if the geometry of the entire detector installation is such that it covers only part of the
sphere around the collision point, then the integrations over the unit sphere would have to be corre-
spondingly restricted. Of course, theoretical predictions would have to take such things into account.
(vi) One does not have to determine the exact number of jets in each event: all events are processed, in
principle, in the same wayii. Nor does one have to identify jets. Even the events with X, in which indi-
vidual jets are hard or impossible to resolve in a conventional way, will contribute their share to the
bump at s = SX in the spectral discriminator.
(vii) From the point of view of theoretical predictions, higher order corrections may prove important for
a precise description of the characteristic tails in the right figure 10.26. It may well turn out that to ex-
tract the most from spectral discriminators one would need to include such corrections into theoretical
predictions. Therefore, whatever discrepancies may remain between theoretical predictions and experi-
mental curves, they are either due to experimental data errors or to unknown theoretical higher order
corrections -- but never to "ambiguities" of jet definition.

i We are talking about one event here; the smearing due to averaging over many events is a separate issue.
ii A realistic code may involve optimizations such that different events would be treated differently; cf. Sec. 7.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 75






Practical computation of spectral discriminators 12

This section summarizes an algorithm to compute a spectral discriminator from data. The scheme de-
scribed is insensitive to which concrete spectral discriminator one deals with (except that the integration
over configurations of `mass detectors' depends on how many -- and what type of -- parameters each
configuration is characterized by). It also generalizes to other differential observables (including higher
differential observables considered in Sec. 13.34) in a straightforward manner. The two realistic exam-
ples are jet + "muon" (Sec. 10.30) and two jets (Sec. 11.19). The exact discriminator is denoted as (s; )
R .
Note that a good part of the described scheme (the notion of a continuous regularization; the use of
linear splines; the folding trick) might be of interest whenever one computes a differential cross section
etc. from a random sample of events with more or less significant data errors (in addition to the limita-
tion of finite statistics). This is because the scheme systematically takes into account the arguments of
Sec. 2.5 so that all cuts are regularized and linear splines are used instead of the more conventional bin-
type algorithms etc. Whether or not such a "politically correct" scheme would lead to a noticeable im-
provement of results depends on many factors (e.g. the size of statistical data errors and the desired
precision of resultss). However, since additional computing costs required are minor one may consider
using the continuous splines-based scenario instead of the bin-type algorithms as a matter of routine.
In the case of spectral discriminators there are -- in addition to data errors -- also approximation er-
rors due to discretized integrations and an approximate representation of -spikes for each event. So
here one has an additional motivation for the use of splines instead of bin-type algorithms.

The data 12.1

One deals with a finite sample of events. Each event is represented by a finite list of physical
"particles" which in practice are the physical calorimeter modules lit up for that event. The length of the
list depends on the event. Each "particle" then is a non-negative energy Ei and a unit 3D vector
(direction) $pi ; the latter may be represented e.g. by three Cartesian coordinates or by two angles i,i .
In the jet + "muon" case one also has the muon's energy E and direction $p .
The grid of masses 12.2

One has to choose the interval of invariant masses s to work with (concerning other parametrizations
see Sec. 11.4). Theoretically, it is [ ,
0 S ]
max -- from zero to the maximal invariant mass of the events in
the sample.
One might be tempted to truncate it at both ends. But that would affect the results of the folding pro-
cedure described below in Sec. 12.11. The folding trick involves a smearing between adjacent values of
s , so if a truncation were introduced, then the validity interval of the results would get narrower after
each folding, which may be undesirable.
Anyhow, the interval of s should be divided into many equali subintervals. It does not seem to make
sense to choose the length of a subinterval to be much less than the error in computation of the invariant
masses of filtered substates.ii However, memory permitting (CPU resources are not affected), one may
choose it to be, say, half that -- the folding trick of Sec. 12.11 allows one to double it any number of
times.iii
In what follows, s i = 0 KN = =
i , , , denotes the boundaries of subintervals; s0 0 and s S
N max . Also,


i Non-uniformity can be introduced via a change of the mass variable; cf. Sec.11.4. One might wish to do this
to focus better on a particular mass range.
ii Remember that there are not only data errors but also errors due to discretization of integrations over unit
spheres etc.
iii If the number of subintervals N is chosen as a power of 2 then the foldings can be done all the way down to
one subinterval. In general, if N is proportional to 2t , then t foldings are possible.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 76






s s +1 - s
i i 12.3

is independent of i .

The grid of R 12.4

The maximal theoretically possible interval for R is [ ,
0 ]
1 (Sec. 10.2). However, it does not seem to

make sense to go beyond R = 1 (which corresponds to the elementary mass detector covering half the
2
sphere; even that value may be too large). Because there is no interaction in the algorithm between dif-
ferent values of R, one can choose any values to work with and employ optimizations that allow simul-
taneous computation for several values of R.
Note, however, that one would like to determine the value(s) of R for which the bump corresponding
to the particle one searches for is the narrowest. This should roughly correspond to the average angular
size of the resulting jets.

Discretization of integrations 12.5

The definition of spectral discriminators involves an integration over all configurations of mass detec-
tors. Each such configuration is characterized by positions of the corresponding elementary mass detec-
tors. In practice, one would choose a finite number of positions $
qi and distribute them over the unit
sphere uniformlyi to form a dense grid. Then configurations of mass detectors can be labeled by a dis-
crete index (say, ).
Denote S S( o
R P)
, , the mass measured by the -th mass detector. Then

(s; )
R Z-1 - =
(s
S ), Z 1, 12.6

where the normalization factor Z ensures validity of 10.8.
It should be remembered that even in the continuum limit for the integration in 10.7, the exact spec-
tral discriminator need not be a continuous function and may contain -functional contributions
(Sec. 11.8).
If one defines the discriminator with special weights (cf. Sec. 13.32), those weights have to be in-
serted into the summands of 12.6.

General scenario. Regularization 12.7

Numerical work with objects such as spectral discriminators that involve -functional contributions is
based on the use of a regularization (Sec. 15.29). One computes a sequence of arrays r (s ; R
N i ) from the
available data, and then chooses (empirically) an optimal value N = ~
N which is neither too large nor too
small. It should not be too large so that the stochastic irregularities are sufficiently smeared, and it
should not be too small so that the signal from the particle one searches for is seen. Whether such a bal-
ance can be achieved depends on the size of the event sample and the precision of integrations. (Recall
that any pair of jets in a final state -- in the case of 2-jet spectral discriminator; any one jet in the
jet + muon case; etc. -- contributes a bump to the spectral discriminator. But only the bumps that corre-
spond to a particle decaying into jets are added up instead of being smeared away after averaging over
many events.)
Practically, in view of the simplicity of the folding trick (Sec. 12.11) one computes from data one ar-



i Uniformity is, strictly speaking, not necessary and is assumed for simplicity. With a non-uniform distribution,
one should introduce appropriate weights in all the formulas (e.g. via systematic mappings of coordinates). This
may be used as an optimization trick; for instance, one could take more points around the peaks of the energy
flow due to energetic jets -- similarly to the adaptive integration routines.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 77






~
ray r (s ; R
N i ) for a large enough N and then finds an optimal N using foldings while, say, watching the
results on computer screen.
It should be emphasized that each element of the array r (s ; R
N i ) is a correct observable in its own
right; given infinite statistics and computer resources, it could be computed with infinite precision; it
accumulates contributions from all events as is usual with any quantum mechanical observable.

Computation of r (s ; R
N i ) for one event 12.8

The event is represented by the data described in 12.1.
One runs a loop over all configurations of the mass detector (each configuration is labeled by ). At
each step one runs a loop over the chosen values of R. At each step one computes ( $p )
i ( $p )
;R i
and then the invariant mass S S( o P) . The obtained value S is used to add a contribution to
r (s ; R
N i ) .

Adding a contribution to r (s ; R
N i ) 12.9

Each configuration of mass detector adds a unit of spectral weight to the spectral discriminator; the
contribution is located at s = S computed previously for each R. In terms of r (s ; R
N i ) , one does the
following.
One determines i such that s S s
i i+1, and redefines

s +1 - S
r (s ; R) r (s ; R i
+
N i N i ) ,
s
S - s
r (s ; R)
+1 r (s ;R i
+1 +
N i N i ) . 12.10
s

At this point all the loops are CONTINUE'd.
In the end one divides r (s ; R
N i ) by Z , the total number of points used to discretize integration(s) over
the unit sphere(s) (cf. 12.6).

The folding trick 12.11

This is based on the observation that linear splines for the number of subdivisions N / 2 can be ob-
tained from the splines for the number of subdivision N in the following simple way:

h (s - s ) = 1 h (s - s ) 1 1
4 2 2 -1 + h (s - s ) + h (s - s
N i N i N i N i ).
2 2 2 4 2 2 +1 12.12

The formula for the array r (s ; R
N i ) is a replica from 12.12:

r (s ; R) = 1 r (s ; R) 1 1
4 2 2 -1 + r (s ;R) + r (s ; R
N i N i N i N i )
2 2 2 4 2 2 +1 . 12.13

It is implied that whenever an index goes beyond its boundaries, the value returned is zero.
Note that the array on the l.h.s. has (almost) twice as few elements as the array on the r.h.s., and the
corresponding values si are twice as far apart.
The folding trick effectively halves N without complete recalculations, and without making any fur-
ther approximations.
A concrete implementation of the algorithm need not follow the above description in every detail.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 78






Advanced options 13

There are several ways in which the formalism can be extended. We have so far been considering the
simplest kinematical context of the c.m.s. annihilation e+e- hadrons. Now we turn to modifications
needed to adapt the formalism to the case of hadronic reactions. As was discussed in [26], there are two
points of view on jet counting in the case of hadronic initial states. One is to modify the algorithms de-
veloped for e+e- hadrons in a straightforward manner [69]. The other is to emphasize the specifics
of hadronic reactions, most notably, the invariance with respect to boosts in the direction of the colliding
beams and the inclusive nature of the corresponding jet-related observables [26]. Two simple modifica-
tions in the spirit of the first point of view are discussed in Sec. 13.1. Accommodating the ideas of the
second approach requires more involved and extensive modifications of the formalism, as discussed in
Sec. 13.7. Lastly, in Sec. 13.29 we discuss the options available for constructing more complex C-
continuous observables that use additional dynamical information that may be available about the reac-
tion one studied in order to enhance the signal. The examples we consider are motivated by the top
search experiments at FERMILAB [12], [13].

Modifications for hadronic reactions, recombination-style 13.1

Special reference frames for DIS 13.2

In the case of deeply inelastic lepton-nucleon scattering it may be desirable to define jets in special
reference frames [50]. To achieve this, we first rewrite particles' energies and the factors ij defined in
8.6 in a covariant form. Define P = ( ,
1 0) and

p = (E , 2
p ), p = E $
p , p =
i i i i i i 0. 13.3
i

Then

E = p P, = ( p p )( p P -
) 1 ( p P -1
i i ij i j i j ) . 13.4

Now to make e.g. the jet-number discriminators (Sec. 8) "count jets" in any other reference frame, it is
sufficient to choose a 4-vector such that P2 = 1 and its rest frame is the desired reference frame. Other
observables (such as spectral discriminators of Sec. 11) are modified similarly.

Suppressing "forward jets" 13.5

In the case of hadrons in the initial state one may wish to modify the jet-number discriminators 8.8 to
suppress contributions from spectator partons (forward jets). Such modifications are easily done by
analogy with how the recombination algorithms are modified in such cases (cf. [50]). For instance, it
may be sufficient to introduce into jm (see 8.8) the factor

= 1 - 2
i cos i 13.6

per each particle, where i is the angle between the particle's direction and the beam axis.

Modifications for hadron-hadron reactions, rapidity-inclusive-style 13.7

Using rapidity: cylinder replaces unit sphere 13.8

Recall that in the spherically symmetric case of e+e- hadrons we represented particles' directions
$
p by unit 3-vectors, i.e. points of the unit sphere parametrized e.g. with and . In hadron-hadron col-
lisions it proves advantageous to replace with pseudorapidity
= ln cot (cf. [10], [26]), and to work
2


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 79






with transverse components of particles' 3-momenta. This is equivalent to representing the directions $p

in terms of and a unit 2-vectori $pT orthogonal to the beam axis:

$ $
a
p p , f
T ( , ). 13.9

This means that, instead of the unit sphere, one now deals with a cylinder, a direct product of a unit
circle (the collection of all $pT) and a real axis corresponding to . This is equivalent to taking out the
two polar points of the sphere and stretching it to  .
$
pT can also be referred to as transverse direction. Note that
$ $ a f
p p = cos cos -
T,a T,b ab a b . 13.10

On the other hand, to avoid confusion it is reasonable to retain the interpretation of the scalar product of
two directions as cosine of the angle between them:

$ $
p p = cos
a b ab . 13.11

The natural integration measure is modified as follows. If is a part of the cylinder 13.9 then Eq.4.3
is still valid but "surface" is interpreted differently; in place of 4.4 one now has
z 2
$
p ( $
p ) = z $
p . 13.12
S1 z+ ($
T
p , ) = +
d d d d d
T ( , )
- z0 z-

Defining angular separation 13.13

The definition of angular separation for the spherically symmetric case was based on the considera-
tions discussed in Sec. 4.5. For the cylindrical geometry and small angular distances one has, instead of
Eq. 4.7,

1 | $p - $ 2
p | = 1 | $
p
T, - $ 2
p |
T, + 2 = 1
( - cos ) + 1 2
ab a b a b , 13.14
2 2 ab ab 2 ab

where = -
ab a b . This definition retains a maximal resemblance both to the spherically symmetric
version, Eq. 4.7, and, for small ab, to the conventional measure of angular distance used in cone-type
jet definitions:

1 2 1 2 2
R = + . 13.15
2 ab 2 ab ab

Note, however, that the Euclidean form of 13.15 (and 13.14 for small ab) is completely ad hoc. It may
have had some sense in the context of conventional algorithms where it might have been expected that
jets should rather have smooth regular shapes (but even then the issue of the form of jets is rather murky
[26]). As was remarked in Sec. 7.41, in the context of the new formalism the preclustering algorithm is
expected to be mostly used for smaller angular resolutions so that the issue of the protojet shapes is
simply irrelevant as long as the induced errors are kept sufficiently small (cf. Sec. 7.41). On the other
hand, the considerations of computational efficiency may induce one to use a simpler definition, e.g.

= a f 2
ab max | ab|,| ab| . 13.16

There are many possible variants, and the choice may depend e.g. on the specific form of representation
of the information about angular positions of detector cells of a concrete detector installation.



i Instead of which one could use the angle . I prefer the coordinateless vector notation as a more natural one.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 80






The theoretical considerations that fixed the definition of angular separation in the spherically sym-
metric case of e+e- hadrons (ease of analytical study of the jet-number discriminators) seem less
important in the hadronic case because analytical calculations are hardly possible here anyway. How-
ever, should one wish to take this into account, one may opt for the following definition which corre-
sponds to one of the angular distances discussed in [69]:

= cosh -
ab ab cos ab . 13.17

This expression is said to conform to the structure of eikonal factors in QCD matrix elements [69].
Concerning the arbitrariness of the definition of angular separation, the following should be under-
stood. The choice of angular separation is of little consequence for the preclustering of Sec. 7 (because
sensitivity to a particular choice means that the chosen value of ycut is too large; cf. Sec.7.41) and the
choice may vary depending on a particular implementation of the algorithm. In the definition of the jet-
number discriminators, however, it may be desirable to adhere to one definition from the very beginning.
Eventually, some form of an "accord" may perhaps be needed to fix the definition of ab (perhaps, a
suitable topic for SNOWMASS'96). Because considerations other than computational simplicity do not
seem to carry sufficient weight in the case of hadron-hadron collisions, the definition 13.16 appears to
be a strong favourite.
The reasoning below is independent of the particular choice of the angular separation ab .

Transverse energy 13.18

Another modification in the case of hadronic reactions is that, instead of the total energy of the i -th
particle Ei it is considered more natural in this context to use its transverse energy [26],

def
E = E
i,T i sin i . 13.19

The ordinary energy should be replaced with the transverse energy,

E E
i i,T , 13.20

in all the formulas and interpretations of the preceding sections. Thus, the "energy flow" is now an ab-
stract measure on the cylinder $p 
T whose values are transvers energy, etc. Practically, all one has to
do is to reinterpret Eqs. 4.20, 4.21, etc. accordingly.

C-correlators 13.21

The general formula for C-correlators 5.18 remains true with the above reinterpretations. However,
since the two poles of the unit sphere have been removed, the required continuity of the angular func-
tions ( fm in 5.18; remember that it is now expressed in terms of instead of ) does not now concern
the "points" =  (formerly = 0, ). This effect is seen e.g. in the formula for the C-correlator that
computes invariant mass, Eq. 5.14, which now becomes

(1 - $ $
i
p p j )
S = S P
( ) = E E

tot , 13.22
ij i,T j,T |sin
i sin j |


with the angular function that is singular as =  . However, the total energy of the event is bounded
by a constant so that the behavior of the "transverse energy flow" Ei,T at  suppresses the sin-
gularities of the angular functions ( Ei,T depends on via i). Such singularities are, therefore, spurious.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 81






Modifications of the optimal preclustering 13.23

The reasoning of Sec. 7 remains valid with the reinterpretation of Eqs. 13.9, 13.20 and any of the
definitions of the angular separation ab discussed in Sec.13.13. It should be emphasized that for
smaller ycut one should opt for the computationally simplest among all equivalent definitions.

Modifications of the jet-number discriminators Jm 13.24
The formula 8.8 and the reasoning that led to it remain valid in the new context provided one uses a
new definition for the angular separation ab and all energies are transverse energies. Since the had-
ronic case is kinematically more complex, it is no longer possible to present analytical formulas for,
e.g., normalizations of Jm , which will have to be determined numerically for each choice of ab . Other
properties (such as monotonic decrease for large m etc.; see Sec. 9) are retained independently of this
choice.

Modifications of spectral discriminators 13.25

All one has to do is the following:
 Reinterpret the angular separation ab in the definition of the elementary filter Eq.10.3 according to
Sec. 13.13.
 Recompute the normalization factors for all spectral discriminators, Eqs.10.7, 10.32, 11.22.
 Restrict each integration over $q (position of an elementary mass detector) to the interval of pseu-
dorapidity corresponding to the actual experimental data.

The key qualitative features of spectral discriminators -- -spikes, their evolution etc. -- are independ-
ent of the particular kinematic context.

Jet distributions with respect to ET 13.26

As was clarified in [26] there are physical arguments in the case of hadron-hadron collisions to prefer
inclusive observables such as 1-jet inclusive differential cross section d / d T
E (used e.g. for precision
determination of S [70]). Such an observable can be translated into the language of C-algebra in a
most natural manner. In fact, the relevant construction is a simple analogue of spectral discriminators.
Define:

N
E P def part
=
T ( ) i
E T
, . 13.27
i=1

Consider the 1-jet spectral discriminator 10.7 with kinematic modifications described above, and use the
observable 13.27 in place of the invariant mass. One obtains:

def $
= a f
q -
z
1 jet ( T
E ; R) d T
E ET ( $,
q )
R . 13.28

The meaning of this expression is similar to spectral discriminators already considered: Each clean jet in
the final state gives rise to a -spike in 1 jet ( T
E ; )
R for an appropriate range of R. Averaging over all

events yields a continuous distribution that is analogous to d / d T
E in the high- ET region. The analogy
would be rather complete if all -spikes were of the same height. This, however, is not so, and the -
spikes have different height depending on the structure of the surrounding event. This means that there
is some dynamical information blended into 1 jet ( T
E ; )
R as compared with the simple counting of

events in the case of d / d T
E . But this can hardly be considered a drawback -- an advantage rather --
given the intrinsically natural manner of how it is accomplished.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 82






Further options for C-continuous observables 13.29

Even within the restrictions of C-algebra there still is a considerable freedom of choice to allow one
to take into account the dynamical information that may be available for a particular experimental
situation (e.g. from a preliminary study of the data using conventional methods). The general options
outlined below should help one with that.

Generalized differential C-continuous observables 13.30

In practice it is convenient to have in view a rather wide class of C-continuous observables of the
following form:

F P
( ) d ( ) G( o P)
z c -H( o
s P)
h, 13.31

where F, G and H are, typically, C-correlators (other C-continuous observables are also allowed),
and are two families of filters parametrized by which describes the geometry of the filters -- in fact,
the structure of substates the observable probes (cf. Sec. 11.13), H describes the physical feature with
respect to which the distribution is studied (e.g. the invariant mass of the substate or its total transverse
energy ET ), while F, G and may be used to enhance the signal (an example is given in Sec.13.32).
The C-continuity of the construct 13.31 follows from the results of Sec. 5.

Example: using additional dynamical information to enhance signal 13.32

Consider a search for particle X that decays into two jets, for which one uses 2-jet spectral discrimi-
nator 11.22 expecting to see a bump signaling a virtual presence of X. Let us show how additional dy-
namical information may allow one to enhance the signal/background ratio by modifying the definition
11.22 along the line of 13.31.
For instance, suppose the jets resulting from the decay are expected to be accompanied by no less
than, say, 3 jets. Then the signal can be enhanced by multiplying each contribution to the spectral dis-
criminator (the integrand of 11.22) by J ( o
$ P
3 ) (= G in 13.31) where
1
q , $2
q ;R


-
$ ( $p ) 1 , 13.33
, $ ; $
R i , $ ; ( $
p )
q q q q
1 2 1 2 R i

and $q , $q ;
1 2 R is the same filter that describes the mass detector used in the definition of the spectral dis-
criminator (cf. 11.20). The point is that the filter $q , $q ;
1 2 R defines a substate that is complementary to
the one being tested by $q , $q ;
1 2 R . If the complementary substate has less than 3 jets then
J ( o
$ P
3 ) suppresses the contribution (the properties of the observables J
1
q , $2
q ;R m are described in

Secs. 8 and 9). This would affect the normalization 11.23 but that is of little consequence as long as the
desired suppression is effected. Note that such a factor can be raised, without spoiling the C-continuity,
to any positive power which may make the suppression sharper (perhaps, at the cost of decreased sta-
bility with respect to data errors and statistical fluctuations).
Similarly, suppose one expects that 2-jet events cannot contain the particle one searches for (i.e. X is
always accompanied by at least another jet). Then it is natural to modify the expression of spectral dis-
criminator by simply multiplying the contribution to it from each event P by (a power of) the ordinary
3-jet discriminator J P
3 ( ) (= F in Eq. 13.31).
One could also use information about expected angular distribution of the jets in the final state to in-
troduce angular weights (= in Eq.13.31).
The above scheme is generalized in a straightforward way to the case of, say, particles that decay into
3-jets and are mostly produced in 6-jet events (which corresponds to the top search at the FERMILAB


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 83






Tevatron in the purely hadronic channel [25]).

Higher differential C-continuous observables 13.34

Similarly to differential cross sections, one can define differential C-continuous observables with re-
spect to more than one parameter. For this, it is sufficient to insert additional -functions into the inte-
grand of 13.31.
As an example, consider the Lego plots in Fig. 4 of [13]. Their (simplified) C-continuous analogue
would be as follows. One considers a 3-jet spectral discriminator and inserts an additional -function,

$ $ $ $ $ $ $
c a fh c a fh
3j 2j ( 3j
s , s2j ; R) d 1
q dq2 d 3
q s3j S 1
q , q2 , $3
q ; R s2j S 1
q ,
- = - -
z q2 , $3
q ; R , 13.35

where

S $ $
a f
q , q , $
q ; R d
S o Pi a f d
S o Pi
1 2 3 = $ $ , S $ $
q , q , $
q ; R
1 2 3 = min $ $ . 13.36
q ,q , $
q ;R q ,q ;R
1 2 3 a,b= a b
,
1 2,3
ab

The function S is the minimal invariant mass among pairs of jets from the three jets selected by the
filter in S .
Transition to m = s1/2
nj can be performed as explained in Sec. 11.4.
nj


The above examples demonstrate that the expressive power of C-algebra is sufficient for practically
any application of the precision measurement class.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 84






Summary and conclusions 14

The results of this work can be summed up as follows:
 A systematic analysis allows one to reinterpret the "ambiguities" of the conventional jet finding
algorithms as instabilities caused by the intrinsic discontinuity of such algorithms. Such instabilities
are inherited in the form of systematic errors by any observable defined on the basis of such algorithms.
 The instabilities are eliminated if observables are chosen to be continuous in an appropriate sense.
A proper continuity of observables is determined uniquely by the structure of measurement errors of
multimodule calorimetric detectors. It can be described as a stability with respect to "almost collinear"
fragmentations that cannot be distinguished by a finite precision calorimetric detector. The fact that
real-life calorimetric detectors have finite energy and angular resolutions is a key consideration that has
been missing in the discussion of jet-related fragmentation-invariant observables: The resulting continu-
ity (which we call calorimetric continuity, or C-continuity) restricts the structure of observables rather
severely -- yet leaving enough freedom to allow one to describe any jets-related physics in the language
of such observables.
 The C-continuity ensures that observables possess optimal properties with respect to data errors.
On the other hand, a use of continuous weights instead of hard cuts corresponds better to the physical
reality of absence of any internal boundaries separating final states with different "numbers of jets" in
the continuum of all possible final states.
 A rather wide family of C-continuous observables can be constructed within the framework of the
so-called C-algebra. The C-algebra contains:
-- A special class of observables (the so-called C-correlators) that have a rigid analytical structure;
in particular, their dependence on particles' energies is fixed. The most important among them are:
(transverse) energy; invariant mass; a sequence J = 1 2 K
m , m , , , of the jet-number discriminators
(Secs. 8, 9 and 13.24). (Note that the construction of the latter differs slightly for different kinematical
situations such as e+e- hadrons, deeply inelastic lepton-nucleon scattering or hadron-hadron colli-
sions; cf. Sec. 13.)
-- A set of rules to construct new C-continuous observables from those already available (Secs. 5.22
and 13.29).
-- A set of rules to translate the conventional observables based on jet algorithms into the language of
C-algebra (the filtering to select "multijet substates", Secs. 5.25, 10.2, 11.14; the "spectral" construc-
tion to describe distributions of properties -- e.g. masses -- of "multijet substates", Secs. 10.1
and 11.1).
 The C-correlators that form a basis of C-algebra have the form of multiparticle correlators and,
therefore, fit tightly and naturally into the framework of Quantum Field Theory (they can be rewritten
in terms of the energy-momentum tensor [54]). It is fair to say that any property that can be meaning-
fully studied within QFT (which summarizes a vast body of experimental knowledge) should be ex-
pressible in the form of multiparticle correlators. Since there is, really, little reason to expect that QFT
may break down at the energies of the current and planned colliders, one concludes that the taking into
account of "the kinematical QFT aspect" is the other key ingredient that has been missing from the dis-
cussion of jet-related observables. The fact that the C-correlators fit naturally into the QFT framework
opens a prospect for higher quality theoretical predictions in the physics of jets.
 It should come as no surprise that because only kinematical restrictions went into the condition of
C-continuity, it proves possible to express any physics conventionally studied via jet finding algorithms,
in the language of C-continuous observables. With such observables, at no stage does one have to iden-
tify individual jets. As a result, the problem of instabilities of jet algorithms in its current form simply
vanishes.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 85






 In the new formalism, jet algorithms retain the role of an approximation trick for faster computation
of such observables from data (preclustering). The jet resolution parameters of conventional algorithms
become parameters that control the corresponding approximation errors. In particular, there is one
algorithm (Sec. 7.20) that is optimal from the point of view of minimization of such approximation
errors. Such an "optimal preclustering" happens to possess features of various conventional jet algo-
rithms; in particular, it allows one to recombine any number of particles into one (or more) protojet(s)
using well-defined analytical criteria (Eq. 7.39 and Sec. 7.47).
 The explicit -- and rather simple -- analytical structure of the observables from C-algebra results in
a much greater flexibility as regards construction of approximation tricks etc. than in the case of the
conventional scheme. One example is the expansions in energies of soft particles (Sec. 7.4; cf. the ana-
lysis of errors due to missing energy in Sec. 7.15). Other tricks will undoubtedly be found in the context
of conrete applications.
 The advantages of the new formalism -- a lesser sensitivity to data errors, absence of instabilities,
advantages for theoretical studies including calculations, computational flexibility -- should be expec-
ted to be least negligible in the applications that can be characterized as precision measurements, i.e.
whenever the quality of both experimental numbers and theoretical predictions is important, including
low-signal situations.

Objections 14.1

Higher computing resources 14.2

One objection to the new formalism is that C-continuous observables require more computer resour-
ces to compute them from raw data. That, however, is a purely technical matter: First, the regular ana-
lytical structure of new observables allows many analytical and programming optimizations. Second,
little additional computer power seems to be needed for computing new observables for the values of
control parameters of jet algorithms that (values) are currently employed. As more computing power
becomes available (which happens very fast), more precise computations can be done (see below).

Ambiguities 14.3

One may argue that there are still ambiguities to fix in the definitions of various C-continuous obser-
vables (e.g. the shape of elementary mass detectors in the construction of spectral discriminators in
Secs. 10 and 11; the choice of angular separation in the hadron-hadron case Sec. 13.13, etc.). The an-
swer to this objection is that it was not ambiguities per se that were the problem with the conventional
jet algorithms, but their enhancement by instabilities. Once the cause of the enhancement is eliminated,
all one has to worry about is a consistency of the definitions used in theoretical and experimental calcu-
lations.i

Psychological objections 14.4

More difficult to cope with (in a sense) are objections of psychological nature, mostly due to an un-
derstandable inertia of thought.
One such objection is that a jet finding algorithm, if used in a consistent manner by theorists and
experimentalists, is a well-defined observable. Of course, one can call any number computed from data
an observable. However, it stands to physical reason that defining a physically correct observable must
take into account errors of measurement and calculation; a measure of stability against such errors must
be incorporated into the definition; and one must be aware of the consequences of the choices one makes


i Of course, having got rid of the systematic errors due to "ambiguities of jet algorithms" one may discover that
what one really needs is more theoretical higher order corrections, or a higher precision of one's calorimeters,
or a greater bandwidth of one's data acquisition system -- but that would be an entirely different situation.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 86






and of the available options. Instability of jet algorithms is a self-inflicted woe, and there is no reason
why it should not be eliminated whenever technically possible and physically necessary.
On the other hand, it is not clear how one can rationally argue against the point of view that the jet
pattern, i.e. the number and 4-momenta of jets for a final state (cf. Eq. 1.2), is only an approximate
description of that final state. One may or may not be happy with such an approximation. In the former
case, one can simply ignore the C-algebrai. In the latter case, however, one would like to know how to
improve upon the approximation. The new formalism offers a systematic answer to this question.
Another psychological difficulty is that some effort is needed in order to learn to think in terms of the
new formalism rather than keep interpreting and judging it (sometimes with misleading conclusions) in
terms of the old paradigm based on a naive definition of jets, which the new formalism is meant to su-
persede. The habit of thinking in terms of jets is deeply ingrained, in part due to their highly visual char-
acter. Physical theories, however, are not always entirely intuitive, and physical meaning -- contrary to
a wide-spread implicit belief -- is not the same as quasiclassical visualization. Although the jet/parton
picture remains a valid and greatly useful approximation, it is only a first approximation to the precise
QFT-compatible description based on the C-algebra.

From jet finding algorithms to C-algebra 14.5

A transition to the systematic use of C-continuous observables -- in situations where the additional
computational resources required by the new formalism would be justified by higher precision require-
ments -- could be performed gradually. Indeed, the conventional data processing consists of two steps:
first, a jet finding algorithm computes a jet pattern for each event (Eq. 1.2); second, an observable is
computed from jet patterns.
The new formalism allows one to translate any such observable into the language of C-continuous
functions (C-correlators, spectral discriminators, etc.). The difficulty here is that the translation is not
a mechanical procedure. But once one has worked through typical examples (Secs.813), it should not
be particularly difficult. Because the framework of C-algebra incorporates only kinematical restrictions,
any truly observable physical featureii can be adequately expressed in the language of C-continuous
observables.
Since jet algorithms are treated as an approximation trick (= preclustering) within the framework of
C-algebra, one can start by simply replacing the old observables with new C-continuous ones -- with-
out changing the jet algorithmiii -- and reprocess the same events in order to compute the new observ-
ables (recall e.g. that only O(100) events were used to discover the top quark [12], [13]). This allows
one to test the algorithms with a limited but meaningful sample of events.
Next, since C-continuous observables retain their meaning and qualitative behavior even if one deals
directly with the original events instead of their jet patterns, one can choose a smaller value for the jet
resolution ycut (or the jet cone radius R) and repeat the entire computation. This means that fewer dis-
tortions would be introduced into the computed observables but more computing power would be
needed. Continuing in this fashion one should be able to see how far one is able and willing to go with
the available computer resources.
Therefore, although a prospective transition to the use of C-algebra for quantitative description
of physical phenomena involving hadron jets in situations where the limitations of the conventional
approach are too restrictive -- although such a transition does represent a change of direction, it does
not mean a complete break with the tradition of jet finding algorithms, and can be accomplished in
an evolutionary manner.



i As did the Indiana State House of Representatives in 1887 when it decreed that = 3 [70].
ii Not to be confused with an expression of such a feature in the language of jet algorithms.
iii Eventually, the optimal preclustering will have to be used but the switch can be done any time.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 87






Appendix. Abstract measures on [0,1] 15

An excellent source on abstract measures with emphasis on the functional point of view is [45]. For
numeric handling of measures one only needs to put together a few tricks that are well-known in applied
mathematics. In this appendix only measures on the unit interval are discussed. The primary purpose is
to explain the computational issues associated with spectral discriminators.

Continuous functions. Constructive aspects 15.1

The notion of (abstract) measure is derivative from that of continuous function. Consider continuous
functions defined in the unit interval 0 s 1, and the continuity extends to the end points. Define the
C 0-distancei between two such functions:

def
f - g 0 = sup f (s) - g(s
C ) . 15.2
0s1
We say that f approximates g in the sense of C 0 if the C 0-distance 15.2 is sufficiently small.

Linear splines 15.3

Split the interval [ ,
0 ]
1 into N equal subintervals; the boundaries between them are the N + 1 points
~
s = i / N, i = ,
0 ,
1 KN =
i . Fix N + 1 numbers f i 0 1 K N
i , , , , . Consider a function f (s) such that it takes
~
the values f =
i at the points si , f (s ) f
i i , and interpolates linearly in between:
~
f (s) = N f (s +1 - s) + f (s
+1 - s ) , s s s
i i i i i i+1. 15.4

We call such functions linear splines. The following two facts are important here: (a) Linear splines are
perfectly constructive objects: they can be represented as arrays [0: N] of floating point numbers.
(b) One can use them to approximately represent any continuous function.
~
Indeed, suppose f = f (s
i i ) . By choosing N large enough, f (s) can be made to represent f (s) with
any precision for all s simultaneously (i.e. in the sense of C0):
~ ~
f - f < , or f (s) - f (s) < for all 0 s 1. 15.5
C0

The linear splines play the same role for continuous functions as the floating point numbers do for real
numbers. Note that if N is fixed, different continuous functions will be represented with a different C 0-
precision.
A convenient representation for 15.4 is as follows. Define:

h(s) = max(0,1 |
- s| ,) h (s) = Nh(Ns), lim h (s) = (s
N N ) . 15.6
N
Then
~ N
f (s) 1 h (s
- s ) f (s ) . 15.7
N i N i i
=0

Measures on [ ,
0 ]
1 15.8

Imagine a subroutine (called, say, ) that accepts as an argument a variable length array f [0: N] of
floating point numbers, and returns one number as a result, ( f ) . Since the array f can be regarded as
representing a linear spline, the subroutine defines a function on the space of all linear splines. To dis-


i C0 is a standard mathematical notation for things pertaining to continuous functions when the latter adjective
is inconvenient as in the case of C 0-distance. The superscript 0 indicates that one talks about continuity of the
function itself only; C 1 would mean that there are continuous first order derivatives, etc. The distance is C 0 be-
cause if a sequence of continuous functions converges in the sense of C 0-distance, then the limit is a function
that is automatically continuous (i.e. also C 0).


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 88






tinguish linguistically the ordinary functions and these new functions that have ordinary functions as ar-
guments, the new sort of functions are called functionals. Thus we say that is a functional on linear
splines.

Examples of functionals 15.9

(1) ( f ) = surface under f (s)g(s) where g(s) is an integrable function (the algorithm here involves
iterations to perform integration).
(2) ( f ) = the value f ( / )
5 (can be computed with arbitrary precision).
(3) ( f ) = surface under sin f (s) .
(4) Arrange all rational numbers from [ ,
0 ]
1 into a sequence r 1 2 1 3
n (e.g. 1 , , , , K). Then
2 3 3 4 4
1
( f ) = f (r )
n . 15.10
n2
n= ,
1 2,K

The series is convergent because f is bounded. But the convergence is slow enough to ensure that one
has to deal with long sequences of rational numbers to attain high precision.
The subroutines (1), (2) and (4) have an additional property: their results are linear with respect to
their argument (f ).
Furthermore, take a sequence fn of linear spline functions that approximates a given continuous
function f in the sense of C 0. What can one say about the sequence ( f )
n ? In general, nothing. So, one
has to make the following ...

Strong assumption 15.11

For any sequence of linear spline functions fn that converges in the sense of C0 to some continuous
function f , the sequence of numbers ( f )
n should converge in the usual numerical sense, and the result
should be the same for any sequence that approximates f . (This is true in all the above examples.) Then
the subroutine effectively defines a functional on any continuous function. In other words, for any
continuous function f one can compute ( f ) to any accuracy by choosing a linear spline approximation
~ ~
f that is sufficiently close to f in the sense of C 0, and evaluating ( f ).
I emphasize that such a subroutine need not give meaningful results if a different type of closeness for
its argument f is considered. For instance, if one computes the values of for a partial Fourier series for
f , then -- whatever the precision of computations -- the resulting numerical sequence will not, in gen-
eral, converge. Thus, choosing a wrong type of approximation/convergence (in the case of Fourier se-
ries, it is the convergence "in the sense of L2") would render computations meaningless.

What are the most general linear functionals on C0? 15.12

It is exactly such functionals that are called measures.i The crucial requirements are (i) that measures
are linear functionals, and (ii) that they are defined on the entire space C 0.
A convenient syntactic convention is to write a measure as an integral:

1
( f ) = s
d (s) f (s)
z . 15.13
0

This emphasizes the fact that the notion of measure is an extension of the notion of ordinary function
(because an integrable function (s) defines a measure according to 15.13). But it may prompt one to
handle a measure as one would a continuous function, which may lead to erroneous results.

Continuity of measures 15.14

Since we are dealing with measures in a numerical way the issues of errors -- therefore, continuity
-- are central. In standard textbooks one usually requires of measures to satisfy certain continuity prop-


i In the main text they are called "abstract measures" to avoid interference with physical measurements.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 89






erties. However, a remarkable recent theorem [73] ensures that measures defined as above automatically
possess those properties.i More precisely, any measure is automatically continuous in the following
sense: If a sequence of continuous functions fn converges in the sense of C0 then the numbers ( f )
n
converge in the usual numerical sense.

Measures as additive functions of subsets 15.15

In older textbooks and in all advanced textbooks on the theory of probability where measures are
routinely studied (e.g. [49]), one defines measures differently -- as additive functions on subsets. For
instance, if one takes a subset S of the interval [ ,
0 ]
1 to consist of a sequence of non-intersecting inter-
vals, then the total length of S is a measure: the length of a union of two such non-intersecting subsets is
a sum of their lengths. After that one embarks upon an agonizing study of which subsets are measur-
ableii etc., most of which has nothing to do with constructive mathematics. The only textbook I know of
that discusses both definitions systematically with the functional definition as a primary one, is [45].
The modern definition in terms of linear functionals is better suited to our needs: Our measures emerge
directly as sums of -functions (cf. 10.7).

General structure of measures 15.16

In general, a measure is a sum of two components: one is an integrable function (this is usually a
continuous function with a few integrable singularities -- e.g. s-1/2 ); the other is localized on a zero
length subset of [ ,
0 ]
1 (this is usually a discrete -- perhaps infinite -- sum of -functions (cf. 15.10).
A typical spectral discriminator is a sum of a continuous component and a finite number of -
functions (Sec. 11.1).

Convergence of sequences of measures 15.17

This is a central issue because we want to perform computations with measures. This means that our
expressions/subroutines for measures will be only approximations, and if we want them to be indeed
approximations , we should understand precisely what convergence means in the case of measures.
From among different types of convergence of measures we need the one that is natural in our prob-
lem. We call it convergence in the sense of measures.iii
Consider a sequence of measures n. Suppose it is such that for any continuous function f their val-
ues n ( f ) form a convergent sequence of numbers. The limit defines a functional that is automatically
linear. Since we require convergence for each continuous function, the resulting functional is defined on
each continuous function and is, therefore, a measure automatically. Wright's theorem (see Sec. 15.14
above) guarantees continuity of the resulting measure with respect to variations of its argument (in the
sense of C 0).
The rates of convergence for different f are not correlated. This means that there is, in general, no
single number to characterize precision of an approximation in the sense of measures.iv
A rule of thumb is, the faster f changes, the slower the convergence. This has an important practical
consequence:


i The theorem is valid if one replaces the notorious Axiom of Choice that allows one to prove "existence" of all
sorts of pathological counterexamples, with the so-called Axiom of Determinateness due to Myczelski and Ste-
inhaus (the only readable account I am aware of is [73]) which offers a more adequate formalization of con-
structive aspects of mathematics.
ii The modern answer (the Myczelski-Sverczkowski theorem; cf. [73]) is, all. It is valid in the same context as
the Wright theorem mentioned above.
iii In the context of functional analysis it is a special case of the so-called *-weak topology [42]. In the context of
the theory of measure proper it is often called simply weak convergence [47].
iv Mathematically, this means that the *-weak topology is in general not metrizable [42]; even when it is (e.g. in
the case of normalized measures on a compact set), the corresponding distance function is not too useful.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 90






Quality of approximations 15.18

From the above one can infer that if a measure is known approximately, then numerical errors of
( f ) differ for different f. In particular, for some f the errors may be judged acceptable, for others, too
large. Spectral discriminators are constructed by accumulating statistically data from many events. As
the number of events in the data sample is increased, the spectral discriminator obtained is expected to
converge to the "true" one. But with a limited statistics, on which f the obtained precision will be
(un)acceptable?
There is no concrete numerical answer to this question but a rule of thumb is that the "softer" the
shape of f , the better the precision. For instance, it is easier to obtain a high precision for the jet-number
discriminators Jm (Sec. 8.8) for smaller m.

Approximate description 15.19

Now we turn to practical issues of how to construct and manipulate approximations for a given
measure. First, we have to agree on what is "approximate". In all cases we encounter the meaning is
dictated by the problem, and is as follows. A sequence of measures n converges to a given one, if for
any continuous test function f one has convergence
n ( f ) ( f ) in the usual numerical sense. Note
that the rate of convergence depends on f . Practically, one chooses a sufficiently rich finite set of test
functions f -
i , and measures closeness of n to by the differences n ( i
f ) ( if ) .

Measures represented by continuous functions 15.20

For measures represented by continuous functions, one may use linear splines. Note that a sequence
of continuous functions that converges in the sense if C 0, also converges in the sense of measures. Lin-
ear splines approximate a continuous function in the sense of C 0, therefore, will automatically result in
a satisfactory approximation in the sense of measures.
Unfortunately, this obvious method is inapplicable in our case because spectral discriminators emerge
as sums of -functions -- not as continuous functions.
-functions 15.21
The fact is, any measure can be approximated to arbitrary precision in the sense of measures by finite
linear combinations of -functions (cf. 15.26). A well-known case when this occurs is a Monte Carlo
evaluation of an observable 2.7. Then one approximates the expression 2.7 by a finite sum, ~ (
F P )
i i ,
which is equivalent to approximating the probability measure with a sum of -functions as follows:
(P) ~ i ( ,
P P )
i . 15.22

In terms of computer data structures, such a linear combination is just a finite-length list of records,
each containing two fields: the coefficient of the -function, and its location. For measures on [ ,
0 ]
1 , the
location is just a number from this interval.

Constructing regular approximations 15.23

The problem is that in our case the locations of -functions are irregularly scattered over the unit in-
terval so that visualization and comparison are difficult. Moreover, fluctuations due to errors may be
unacceptably large. So, a more regular representation is needed.
Recall that a measure is a collection of its values on continuous functions. Approximate a continuous
~ ~
test function f by a linear spline f
N , Eqs. 15.4 and 15.7, where f f
N as N in exactly the sense
we need (i.e. in the sense of C0 that ensures convergence of values of any measure ). This fact allows
one to approximately represent:

1 1 ~ N
ds (s) f (s) ds (s) f (s 1
) r (s ) f (s
z z
N ). 15.24
N
0 i N i i
0 =0

The explicit expression for rN is as follows:


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 91






1
r (s) s
d (
s)h (s - s
z
N N ). 15.25
0

This is a continuous function of s but only a finite array of its values r (s
N i ) is actually used to repre-
sent (cf. 15.24) -- for a better approximation one takes a larger N' and the new array r s
N ( i ) consists
of values of a different continuous function.
The r.h.s. of 15.24 is exactly the value on f of a sum of -functions located at si , with weights equal
to r (s
N i ) . So, we have obtained the following approximation:
N
(s) (s) 1 r (s )
-
N N i (s s ). 15.26
N i i
=0

N converges to in the sense of measures 15.27
This is ensured by construction.
An important fact is that whereas was any measure -- perhaps, a sum of -functions scattered ir-
regularly within [ ,
0 ]
1 -- the -functions of N are distributed in a regular mann*er, independently of
what was. Therefore, the array of numbers r (s ), i = ,
0 ,
1 ,
K N
N i is a convenient approximate represen-
tation for a measure, to be interpreted according to 15.26. Choosing N large enough, one can make the
approximation as precise as needed.

r (s
N i ) are valid observables per se 15.28
An advantage of the above construction is that the numbers r (s
N i ) are simply the values of on test
functions h (s - s
N i ) -- without any approximations involved. This means that if is a spectral dis-
criminator obtained by collecting statistics from many events, then r (s
N i ) are perfectly valid scalar-
valued observables per se. In particular, if is known precisely (e.g. from infinite statistics experiments
with no systematic errors etc.) then they are also known precisely. In other words, the "measured" val-
ues for r (s
N i ) (i.e. the values computed from a finite sample of data) will converge to the ideal values
as the statistics is increased, and can be used to compare experimental data with theory.
However, statistical fluctuations (including those due to data errors) are smaller for smaller N (cf.
Sec. 2.21).

Regularization 15.29

This is a fundamental notion formalized and systematically developed for applied problems by Tik-
honov and his school [75]. While any exact spectral discriminator (s) is a continuous function after
~
averaging over all final states its approximate representations that emerge due to a discretized inte-
gration over in 11.2 are rather irregular sums of -functions. Although in principle ~
converges to
in the sense of measures (when the event sample and the precision of integrations are increased), one
would like to obtain a more uniform approximation. This is achieved as follows. For each approxima-
tion ~
, one constructs the sequence ~N as described in Sec.15.23. (~N are practically represented by
the corresponding arrays ~
r (s ) = ~
N i .) Then one chooses N N as large as possible -- yet not too large so
~
that the irregularities (due to the stochastic nature of the approximation ) would not be too manifest.
~
The result ~ (equivalently, the array ~
r~ ) is the required regularization. As ~
is made more precise
N N ~ ~
(e.g. by increasing the event sample), the optimal N goes to , and ~ ( ~
or r~ ) . Precise mathemati-
N N
~
cal criteria for choosing the optimal N are unknown at the time of this writing so it is best to proceed in
an empirical fashion. A convenient tool for that is the folding trick of Sec. 12.11.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 92






Appendix. General theory of abstract measures 16

The definitions and facts listed below constitute a minimum of information on abstract measures pre-
sented for completeness' sake. A reader inexperienced in this sort of mathematics may find it difficult to
understand it; in that case, one should consult the material of Sec. 15 as well as an expert mathemati-
cian. A much more complete source is the excellent advanced textbook [45]; general functional-analytic
aspects are treated with elegance in [44].
Note that standard expositions (like [45]) are grappling with difficulties due to an unfortunate axio-
matization of infinite constructions (pathologies due to the Axiom Of Choice). These have nothing to do
with practical mathematics and are ignored below. (See the remarks and references in Sec. 15.8.)
A) Consider a finite-dimensional smooth manifold M. In our case this can be a Euclidean vector
space, a sphere, a cylinder, or a direct product of a finite number of these.
B) Test function on M is a continuous numeric-valued function on M such that it takes non-zero val-
ues only within a compact subregion K of M; i.e. K is such that it does not stretch to infinity in any di-
rection. If M is a (direct product of) sphere(s) than it has no infinite directions and the restriction does
not apply.
C) Measure on M is a linear functional defined on all test functions. (This definition is sufficient for
all practical purposes.) Two generic examples: (i) a linear combination of -functions; (ii) an integrable
function.
D) Measures form a linear space, i.e. one can take their linear combinations.
E) Convergence of measures can be defined in several ways. The one we need is the most natural one
(sometimes called weak convergence of measures; it is a special case of the *-weak topology in spaces
of linear functionals; we call it convergence in the sense of measures), and is as follows: A sequence of
measures converges if their values on any test function form a sequence that converges in the usual nu-
meric sense.
F) Such convergence is, in general, not metrizable, i.e. cannot be described by a single distance func-
tion. This is possible in the case of a compact M (e.g. a sphere) but even then such a distance function is
not practically useful.
G) The linear space of measures is closed with respect to the convergence in the sense of measures.
H) Any measure can be approximated (with respect to convergence in the sense of measures) with
continuous functions (e.g. multidimensional analogues of linear splines). This can also be done with fi-
nite linear combinations of -functions. These two facts are the basis of how measures are to be handled
in numerical applications.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 93






References 17 Russia). Singapore: World Scientific (to be
published).
[1] G. Hanson et al.: Phys. Rev. Lett. 35 (1975) [21] F. V. Tkachov: in Proc. AINHEP-95 (April
1609. 1995, Pisa, Italy). Singapore: World Scientific
[2] G. Sterman and S. Weinberg: Phys. Rev. Lett. (to be published).
39 (1977) 1436. [22] G. C. Fox and S.Wolfram: Phys. Rev. Lett. 41
[3] S. D. Drell, D. Levy, and T.-M. Yan: Phys. (1979) 1581; Nucl. Phys. B149 (1979) 413;
Rev. 187 (1969) 2159; Phys. Rev. D1 (1970) Phys. Lett. B82 (1979) 134.
1035; Phys. Rev. D1 (1970) 1617. [23] G. C. Fox, T.-Y. Tse, and S. Wolfram: Nucl.
[4] J. Ellis, M. K.Gaillard and G.G. Ross: Phys. B165 (1980) 80.
Nucl.Phys. B111 (1976) 253. [24] C. L. Basham, L. S. Brown, S. D. Ellis, and
[5] S. D. Ellis: preprint NSF-ITP-88-55 (lectures at S. T. Love: Phys. Rev. Lett. 41 (1978) 1585;
The 1987 Theoretical Advanced Study Insti- Phys. Rev. D17 (1978) 2298; Phys. Rev. D19
tute, July 1987, St. John's College, Santa Fe, (1979) 2018.
New Mexico). [25] CDF Collaboration: preprint FERMILAB-
[6] S. D. Ellis: in [7]. CONF-94/152-E CDF, June 1994.
[7] Z. Kunszt, ed.: Proc. New Techniques for Cal- [26] S. D. Ellis and D. Soper: Phys. Rev. D48 (1993)
culating Higher Order QCD Corrections 3160.
(1992, 1618 December 1992, ETH, Zrich). [27] S. D. Ellis, J. Huth, N. Wainer, K. Meier,
Zrich: ETH, 1992. N. Hadley, D. Soper, M. Greco: in Research Di-
[8] S. D. Ellis: preprint CERN-TH.6861/93, Ge- rections for the Decade, Snowmass 1990. Sin-
neva, 1993 (talk at The XXVIIIth Rencontres gapore: World Scientific, 1992.
de Moriond "QCD and High Energy Hadronic [28] JADE collaboration (S. Bethke et al.): Phys.
Interactions", 1993). Lett. B213 (1988) 235.
[9] F. Dydak: talk at The IX Int. Workshop on [29] Y. L. Dokshitzer: contribution to The Work-
High Energy Physics (September 1994, shop on Jets at LEP and HERA, Durham, De-
Zvenigorod, Russia). cember, 1990; S. Catani, Y.L. Dokshitzer,
[10] R. Barlow: Rep. Prog. Phys. 36 (1993) 1067. M. Olsson, G. Turnock, and B. R. Webber:
[11] Yu. L. Dokshitzer, V. A. Khoze, S. I. Troyan, Phys. Lett. B269 (1991) 432.
and A. H. Mueller: Rev. Mod. Phys. 60 (1988) [30] S. Bethke, Z. Kunszt, D. E. Soper, and
373. W. J. Stirling: Nucl. Phys. B370 (1992) 310.
[12] CDF Collaboration (F.Abe et al.): Phys. Rev. [31] S. Catani: preprint CERN-TH.6281/91, Ge-
Lett. 74 (1995) 2626. neva, 1991 (invited talk at The 17th Workshop
[13] D0 Collaboration (S.Abachi et al.): Phys. Rev. on the INFN Eloisatron Project, QCD at 200
Lett. 74 (1995) 2632. TeV, Erice, June 1991).
[14] T. Sjstrand: Comp. Phys. Commun. 28 (1983) [32] N. N. Bogoliubov and D. V. Shirkov: Introduc-
229. tion to the Theory of Quantized Fields. 1st ed.
[15] F. V. Tkachov: Phys. Rev. Lett. 73 (1994) 2405 New York: Interscience, 1959.
(editorial and typesetter's errata in print). [33] S. G. Gorishny et al.: Phys. Lett. B259 (1991)
[16] F. V. Tkachov: preprint INR-F5T/93-01, Mos- 144; L. R. Surguladze and M. A. Samuel: Phys.
cow, 1993 (talk at The ITEP Winter School on Rev. Lett. 66 (1991) 560.
High Energy Physics, February 1993). [34] A. A.Vladimirov and D.V. Shirkov: Sov. Phys.
[17] F. V. Tkachov: in Proc. VIII Int. Workshop on Uspekhi 22 (1979) 860; K. G. Chetyrkin and
High Energy Physics (September 1993, F. V. Tkachov: Phys. Lett. B114 (1982) 340.
Zvenigorod, Russia). INP MSU: Moscow, [35] K.G. Chetyrkin, A. L. Kataev, and
1994. F. V. Tkachov: Nucl. Phys. B174 (1980) 345.
[18] F. V. Tkachov: in Proc. IX Int. Workshop on [36] F. V. Tkachov: Phys. Lett. B100 (1981) 65;
High Energy Physics (September 1994 K. G. Chetyrkin and F. V. Tkachov: Nucl. Phys.
Zvenigorod, Russia). INP MSU: Moscow, B192 (1981) 159; F.V. Tkachov: Theor. Math.
1995. Physics 56 (1983) 350.
[19] F. V. Tkachov: preprint NORDITA-95/14P, [37] B. R. Webber: in [7].
Copenhagen, 1995. [38] S. S. Schweber: An Introduction to Relativistic
[20] F. V. Tkachov: in Proc. IX Int. School Particles Quantum Field Theory. Evanston, Illinois:
and Cosmology (April 1995, Baksan Valley, Row, Peterson and Co., 1961.


F
. V
. Tkachov / Measuring multijet structure... FERMILAB-PUB-95/191-T (corrected
10/30/95
) 94






[39] N. N. Bogoliubov and N. N. Bogoliubov, Jr.: [57] Z. Kunszt and D. E. Soper: Phys. Rev. D46
Introduction to Quantum Statistical Mechan- (1992) 192.
ics. Moscow: NAUKA, 1984 (in Russian). [58] For a review see J. C. Collins: Renormalization.
[40] S. Youssef: Comp. Phys. Commun. 45 (1987) Cambridge: Cambridge University Press,
423. 1984.
[41] F. V. Tkachov: in Proc. X Int. Workshop on [59] M. A. Shifman, A. I. Vainshtein, and
High Energy Physics (INP MSU, Zvenigorod, V. I. Zakharov: Nucl. Phys. B147 (1979) 448;
Russia, September 1995, to be published). L. J. Reinders, H. Rubinstein, and S. Yazaki:
[42] Yu.P.Pyt'ev: Mathematical Methods of Inter- Phys. Reports 127 (1985) 1.
preting Experiment. Moscow: VYSSHAYA [60] F. V. Tkachov: Phys. Lett. B125 (1983) 85.
SHKOLA, 1989 (in Russian). [61] F. V. Tkachov: Phys. Lett. B124 (1983) 212;
[43] M. Reed and B. Simon: Methods of Modern Int. J. Mod. Phys. A8 (1993) 2047;
Mathematical Physics. I: Functional Analysis. G. B. Pivovarov and F. V. Tkachov: Int. J. Mod.
New York: Academic, 1972. Phys. A8 (1993) 2241; see also [62].
[44] W. Rudin: Functional Analysis. New York: [62] F. V. Tkachov: in [7]; Sov. J. Part. Nucl. 25
McGraw-Hill, 1973. (1994) 649; Sov. J. Nucl. Phys. 56 (1993) 180.
[45] L. Schwartz: Analyse Mathmatique. Vol. 1. [63] L. R. Surguladze and F. V. Tkachov: Nucl.
Paris: Hermann, 1967. Phys. B331 (1990) 35.
[46] J. L. Kelley: General Topology. Princeton: [64] J. C. Collins and F. V. Tkachov: Phys. Lett.
D. Van Nostrand, 1955. B294 (1992) 403.
[47] N. Koblitz: p-adic Numbers, p-adic Analysis [65] L. Lnnblad: preprint DESY 92-181.
and Zeta-Functions. 2nd ed. New York: [66] J. A. M. Vermaseren: Symbolic Manipulation
Springer Verlag, 1984. with FORM. Amsterdam: Computer Algebra
[48] B. Schneier: BYTE 18 (no.12) (1993) 309. Nederland, 1991.
[49] J. Lamperti: Probability. A Survey of the [67] Eucklidis Elementa, ed. J. L.Heiberg, Lipsae:
Mathematical Theory. New York, 1966. Teubner , 18831888; Liber XIII.
[50] B. R.Webber: J. Phys. G19 (1993) 1567. [68] Johannis Kepleri Astronomi Opera Omnia, ed.
[51] G. Sterman: Phys. Rev. D17 (1978) 2789; Frisch Ch. Francofurti ad Moenum et Erlan-
Phys. Rev. D17 (1978) 2773; Phys. Rev. D19 gae: Heyder et Zimmer, 1871.
(1979) 3135. [69] S. Catani, Yu. L. Dokshitzer, M. H. Seymour
[52] R. F. Streater and A. S.Wightman: PCT, Spin and B. R.Webber: Nucl. Phys. B406 (1993)
and Statistics, and All That. New-York: Ben- 187.
jamin, 1964; N. N. Bogoliubov, A. A. Logunov, [70] W. T.Giele, E.W. N.Glover and J.Yu: preprint
A. I. Oksak, and I. T. Todorov: General Prin- FERMILAB-PUB-95/127-T.
ciples of Quantum Field Theory. Dordrecht: [71] P. Beckmann: A History of Pi. 4th ed., Boul-
Kluwer, 1990. der, Colorado: Golem Press, 1977.
[53] F. R. Ore and G. Sterman: Nucl. Phys. B165 [72] A. S. Barabash and A. I. Bolozdynia: Liquid
(1980) 93. Ionization Detectors. Moscow: ENERGOA-
[54] N. A. Sveshnikov and F. V. Tkachov: in Proc. X TOMIZDAT, 1993 (in Russian).
Int. Workshop on High Energy Physics (INP [73] J. D. M. Wright: Bull. Amer. Math. Soc. 79
MSU, Zvenigorod, Russia, September 1995, to (1973) 1247.
be published). [74] V.G. Kanovey: Axiom of Choice and Axiom of
[55] G. Tiktopoulos: Nucl. Phys. B147 (1979) 371. Determinateness. Moscow: NAUKA, 1984 (in
[56] S. G. Gorishny, S. A. Larin, L. R. Surguladze, Russian); Y. M. Moschovakis: Descriptive Set
and F. V. Tkachov: Comp. Phys. Commun. 55 Theory. Amsterdam: North Holland, 1980.
(1989) 381; S. A. Larin, F. V. Tkachov, and [75] A. N. Tikhonov and V. Y. Arsenin: Methods of
J. A. M. Vermaseren: preprint NIKHEF-H/91- Solving Ill-Posed Problems. 3rd ed. Moscow:
18, 1991. NAUKA, 1986 (in Russian).



