betaFeedback

WorksPhysics and Astronomy

Spectral Analysis of Jet Substructure with Neural Networks: Boosted Higgs Case

Abstract

Jets from boosted heavy particles have a typical angular scale which can be used to distinguish them from QCD jets. We introduce a machine learning strategy for jet substructure analysis using a spectral function on the angular scale. The angular spectrum allows us to scan energy deposits over the angle between a pair of particles in a highly visual way. We set up an artificial neural network (ANN) to find out characteristic shapes of the spectra of the jets from heavy particle decays. By taking the Higgs jets and QCD jets as examples, we show that the ANN of the angular spectrum input has similar performance to existing taggers. In addition, some improvement is seen when additional extra radiations occur. Notably, the new algorithm automatically combines the information of the multi-point correlations in the jet.

Introduction

At multi TeV pppp colliders such as the LHC, boosted heavy particles can be produced and form a single collimated cluster of particles. Such a localized cluster is distinguished from a QCD jet from a hard quark or gluon by the substructures of the cluster [1]. For this purpose, consistent definitions of substructures of jets have been studied extensively. There are various methods for identifying the jet substructures, such as strategies based on cluster decomposition [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] and shape variables [11, 12, 13, 14, 15, 16, 17]. These methods focus on different features of jet substructures to maximize the discrimination power. For the case of Higgs, W, and Z boson decaying hadronically into two quarks, a critical feature is a two-prong substructure inside. Because the key features depend on the nature of the parent particle of a jet, there are several frameworks that can be applied to jets [18, 19, 20, 21, 22, 23].

We propose a spectral analysis in order to identify originating partons of a jet using a spectral function inspired by the proton nuclear magnetic (1H^{1}\mathrm{H}-NMR) spectroscopy of organic molecules. The organic molecules consist mostly of a carbon skeleton and hydrogen atoms. These substructures can be identified by the 1H^{1}\mathrm{H}-NMR spectrum, which records the resonant frequency of the hydrogen nuclei under an external magnetic field. The resonant frequency depends on an induced magnetic field generated by the rest of the molecular substructure; hence, the molecular structures can be determined by investigating the spectrum, as shown in Figure 1. In particular, we use shift and splitting of the resonance frequency, where the big shift comes from the electron cloud, whose density is determined by the skeletal structure, and small splittings are from the electromagnetic interaction between the hydrogen nuclei. Similarly, we develop a spectral function for jet substructure study, where the big structure of the spectrum is made from initiating hard partons while small structures come from QCD interaction from the hard partons. The spectral function that we propose is similar to the angular structure function [18, 24, 25]. The resulting spectrum contains useful information for identifying the nature of a given jet.

Figure 1. Molecular structure of ethanol and its 1H^{1}\mathrm{H}-NMR spectrum. The intensity, location and splitting of peaks allow us to identify the original molecular structure. The chemical shift is the resonant frequency of a hydrogen nucleus relative to a reference frequency.

An artificial neural network (ANN) is a useful model for analyzing the spectrum. Jet substructure analyses based on ANN are gaining attention recently and have been studied in various contexts. The analyses are categorized mainly into two groups with different inputs. One group utilizes special-purpose observables and uses ANN to identify a correlation between substructures [19, 21, 26] similar to analyses with a boosted decision tree [27, 28]. The other group uses the jet constituents directly and uses ANN to find out particular substructures in a jet. This group is again categorized into two subgroups depending on how you interpret jet constituents. One can interpret jet constituents as an image [29] and use image recognition techniques [30, 31, 32, 33, 34, 35]. The other strategies are to interpret a jet as a sequence of data, such as clustering sequence of the jet algorithms [36, 37, 38, 39, 40, 41], and utilize ANN for sequential data analysis [42, 43, 44] Our approach is different from these approaches, namely our network is requested to analyze an event-by-event spectrum of a jet. This approach reduces the inputs of the ANN significantly but it can still learn characteristic non-local correlations in a jet from a heavy particle like the non-local neural networks for video classification [45]. We will show that our approach improves the separation between Higgs jets and QCD jets in natural manner.

This paper is organized as follows. In § 2, we define a spectral function S2(R)S_2(R) and describe its nature. We also explain the setup of our Monte-Carlo simulations. In § 3, we show a cut-based classification of the Higgs jet vs QCD jet using the spectrum S2(R)S_2(R) and compare the performance with the ratio of the enegy correlation function, D2D_2 [16], for two-prong substructures. In § 4, we introduce the spectral analysis of jet substructure with ANN. § 5 is devoted for summary and discussions.

A Spectral Function for Jet Substructure

The jet substructure analysis has similarities with the 1H^{1}\mathrm{H}-NMR analysis of organic molecular structures. An organic molecule has a carbon skeleton surrounded by hydrogens, which 1H^{1}\mathrm{H}-NMR analyzes to find out the topological skeleton. On the other hand, a jet of our interest arises from hard partons, which can be regarded as a topological skeleton of the momentum distribution of jet constituents. The jet constituents radiated from hard partons are something analogous to the hydrogens in a organic molecule.

In the 1H^{1}\mathrm{H}-NMR analysis, we record interactions between hydrogen nuclei and the rest of the molecular structure by their resonant frequencies to measure the molecular structure. Likewise, we focus on correlations between pairs of the constituents of the jet based on distance to identify the originating partons. A popular choice of the distance is angular distance Rij=ηij2+ϕij2R_{ij} = \sqrt{\eta_{ij}^{2} + \phi_{ij}^{2}}, where ηij\eta_{ij} is pseudorapidity and ϕij\phi_{ij} is azimuthal angle between the jet constituents ii and jj. Hence, we define a binned spectral function of angular distance RR as follows,

S2(R;ΔR)=1ΔR∑i,j∈jetRij∈[R,R+ΔR)pT,ipT,j,(1)S_2(R;\Delta R) = \frac{1}{\Delta R} \sum_{\substack{i,j \in\mathrm{jet} \\ R_{ij} \in[R,R+\Delta R)}} p_{T,i}p_{T,j}, \tag*{(1)}

where ΔR\Delta R is a bin width, and pT,ip_{T,i} and pT,jp_{T,j} are transverse momenta of jet constituents ii and jj. In a continuum limit ΔR→0\Delta R \to0,1 this binned spectral function turns into,

S2(R)=∫dR⃗1dR⃗2 PT(R⃗1)PT(R⃗2)⋅δ(R−R12),(2)S_2(R) = \int d\vec{R}_1 d\vec{R}_2\, P_T(\vec{R}_1)P_T(\vec{R}_2)\cdot\delta(R-R_{12}), \tag*{(2)}
PT(R⃗)=∑i∈jetpT,iδ(R⃗−R⃗i)(3)P_T(\vec{R}) = \sum_{i\in\mathrm{jet}} p_{T,i}\delta(\vec{R}-\vec{R}_i) \tag*{(3)}

where dR⃗PT(R⃗)d\vec{R}P_T(\vec{R}) is a pTp_T sum of constituents in a neighborhood dR⃗d\vec{R} of R⃗\vec{R}, and δ(x)\delta(x) is the Dirac δ\delta function. Integrating S2(R)S_2(R) over the bin range [R,R+ΔR)[R,R+\Delta R) returns the binned spectral function,

S2(R;ΔR)=1ΔR∫RR+ΔRdR S2(R).(4)S_2(R;\Delta R) = \frac{1}{\Delta R}\int_R^{R+\Delta R}dR\,S_2(R). \tag*{(4)}

The two-point correlation spectral function may be easily generalized to three-point or multi-point correlation spectral function, analogous to the energy correlation functions [15] and the energy flow polynomials [22]. However, those generalizations are out of the scope of this paper.

The spectral function is infrared and collinear (IRC) safe, namely invariant under soft and collinear radiations. If the IRC safety is not satisfied, Kinoshita-Lee-Nauenberg theorem [47, 48] is not applicable, and the resulting spectrum is hard to be estimated from perturbative QCD calculations. Soft radiation does not change S2(R)S_2(R) because the soft radiation has a zero transverse momentum, which has no impact on PT(R⃗)P_T(\vec{R}) as well as S2(R)S_2(R). Collinear radiation does not change S2(R)S_2(R) because the products stay at the original R⃗\vec{R} coordinate. The momenta of the products are added together at R⃗\vec{R}; therefore, PT(R⃗)P_T(\vec{R}) and S2(R)S_2(R) are invariant.

The IRC safety of the spectral function can be understood easily by explicit examples such as a jet with a single constituent. Suppose that aa is the only jet constituent. This jet has the only angular scale Raa=0R_{aa}=0 and its binned spectrum is given as follows,

S2(R;ΔR)=1ΔR{pT,a2if R=0,0if R≠0.(5)S_2(R;\Delta R) = \frac{1}{\Delta R} \begin{cases} p_{T,a}^2 & \text{if } R=0,\\ 0 & \text{if } R\ne0. \end{cases} \tag*{(5)}

Invariance of this S2(R;ΔR)S_2(R;\Delta R) under soft radiation is trivial, and hence, we consider collinear splitting of aa. Suppose that bb and cc are the jet constituents from the collinear splitting of aa. Then all the pairs bbbb, bcbc, cbcb, and cccc have the same angular scale Rij=0R_{ij}=0. If bb and cc carry zz and 1−z1-z fraction of the transverse momentum pT,ap_{T,a}, S2(R;ΔR)S_2(R;\Delta R) turns into

S2(R;ΔR)=1ΔR{z2pT,a2+z(1−z)pT,a2+(1−z)zpT,a2+(1−z)2pT,a2if R=0,0if R≠0,(6)S_2(R;\Delta R) = \frac{1}{\Delta R} \begin{cases} z^2p_{T,a}^2+z(1-z)p_{T,a}^2+(1-z)zp_{T,a}^2+(1-z)^2p_{T,a}^2 & \text{if } R=0,\\ 0 & \text{if } R\ne0, \end{cases} \tag*{(6)}
=1ΔR{pT,a2if R=0,0if R≠0.(7)= \frac{1}{\Delta R} \begin{cases} p_{T,a}^2 & \text{if } R=0,\\ 0 & \text{if } R\ne0. \end{cases} \tag*{(7)}

Therefore, the binned spectrum is IRC safe. Note that summing the autocorrelation term pT,i2p_{T,i}^2 in Eq. (1) is necessary to achieve IRC safety at R=0R=0, because the crossing term z(1−z)pT,a2z(1-z)p_{T,a}^2 after the splitting is originated from the autocorrelation term pT,a2p_{T,a}^2.

We show another example of a jet with two constituents aa and bb, so that it has a non-zero spectrum at R=Rab>0R=R_{ab}>0. Now the binned spectrum has cross-correlation terms at non-zero angular scale,

S2(R;ΔR)=1ΔR{pT,a2+pT,b2if R=0,0otherwise,+1ΔR{2pT,apT,bif Rab∈[R,R+ΔR),0otherwise.(8)S_2(R;\Delta R) = \frac{1}{\Delta R} \begin{cases} p_{T,a}^2+p_{T,b}^2 & \text{if } R=0,\\ 0 & \text{otherwise}, \end{cases} +\frac{1}{\Delta R} \begin{cases} 2p_{T,a}p_{T,b} & \text{if } R_{ab}\in[R,R+\Delta R),\\ 0 & \text{otherwise}. \end{cases} \tag*{(8)}

Suppose that a collinear splitting of bb produces two partons with transverse momenta zpT,bzp_{T,b} and (1−z)pT,b(1-z)p_{T,b} respectively. Then the binned spectrum turns into

S2(R;ΔR)=1ΔR{pT,a2+[z2pT,b2+z(1−z)pT,b2+(1−z)zpT,b2+(1−z)2pT,b2]if R=0,0otherwise,+1ΔR{2zpT,apT,b+2(1−z)pT,apT,bif Rab∈[R,R+ΔR),0otherwise.(9)S_2(R;\Delta R)=\frac{1}{\Delta R}\begin{cases} p_{T,a}^2+\left[z^2p_{T,b}^2+z(1-z)p_{T,b}^2+(1-z)zp_{T,b}^2+(1-z)^2p_{T,b}^2\right] & \text{if }R=0,\\ 0 & \text{otherwise,}\end{cases}+\frac{1}{\Delta R}\begin{cases}2zp_{T,a}p_{T,b}+2(1-z)p_{T,a}p_{T,b} & \text{if }R_{ab}\in[R,R+\Delta R),\\ 0 & \text{otherwise.}\end{cases} \tag*{(9)}
=1ΔR{pT,a2+pT,b2if R=0,0otherwise,+1ΔR{2pT,apT,bif Rab∈[R,R+ΔR),0otherwise.(10)=\frac{1}{\Delta R}\begin{cases}p_{T,a}^2+p_{T,b}^2 & \text{if }R=0,\\ 0 & \text{otherwise,}\end{cases}+\frac{1}{\Delta R}\begin{cases}2p_{T,a}p_{T,b} & \text{if }R_{ab}\in[R,R+\Delta R),\\ 0 & \text{otherwise.}\end{cases} \tag*{(10)}

Therefore, the binned spectrum is IRC safe. In general, this IRC safety is achieved by the bilinear term of pT,ip_{T,i} and pT,jp_{T,j} in Eq. (1) like the other jet substructure variables directly built from jet constituents [11, 12, 13, 15, 16, 17, 18, 22].

The IRC safety of the spectral function S2(R)S_2(R) is also understood in the context of CC-correlators [22, 46]. The S2(R)S_2(R) is a special case of CC-correlators with an unbounded non-smooth angular weighting function f2(p^1,p^2)=δ(R−R12)f_2(\hat{p}_1,\hat{p}_2)=\delta(R-R_{12}). If we replace the Dirac δ\delta function to a bounded smooth function, for example, δ(x)→(∣a∣π)−1e−(x/a)2\delta(x)\rightarrow(|a|\sqrt{\pi})^{-1}e^{-(x/a)^2}, the Taylor expansion of δ(x)\delta(x) transforms S2(R)S_2(R) into a series of IRC safe energy flow polynomials [22] with two vertices. The series converges to S2(R)S_2(R) in the limit a→0a\rightarrow0, and the IRC safety of the spectral function is understood asymptotically.

Note that the spectral function S2(R)S_2(R) is a basis of bilinear CC-correlators F2F_2 [46] with an angular weighting function f2(Rij)f_2(R_{ij}) of the angular distance RijR_{ij},

F2=∫dR⃗1dR⃗2 PT(R⃗1)PT(R⃗2)⋅f2(R12)=∫0∞dR S2(R)f2(R).(11)F_2=\int d\vec{R}_1d\vec{R}_2\,P_T(\vec{R}_1)P_T(\vec{R}_2)\cdot f_2(R_{12})=\int_0^\infty dR\,S_2(R)f_2(R). \tag*{(11)}

For example, the zeroth and the second moment of S2(R)S_2(R) are the one-point and the two-point energy correlation functions [15], which are approximately the transverse momentum pT,jetp_{T,\mathrm{jet}} and the mass mjetm_{\mathrm{jet}} of the jet respectively,

∫0∞dR S2(R)=(∑i∈jetpT,i)2≈pT,jet2;(12)\int_0^\infty dR\,S_2(R)=\left(\sum_{i\in\mathrm{jet}}p_{T,i}\right)^2\approx p_{T,\mathrm{jet}}^2; \tag*{(12)}
∫0∞dR S2(R)R2=∑i,j∈jetpT,ipT,jRij2≈2mjet2.(13)\int_0^\infty dR\,S_2(R)R^2=\sum_{i,j\in\mathrm{jet}}p_{T,i}p_{T,j}R_{ij}^2\approx2m_{\mathrm{jet}}^2. \tag*{(13)}

These integrals help to interpret S2(R)S_2(R). The spectral densities S2(R)S_2(R) and S2(R)R2S_2(R)R^2 measure contribution to pT,jet2p_{T,\mathrm{jet}}^2 and mjet2m_{\mathrm{jet}}^2 from the pairs of jet constituents at the angular scale RR, respectively.

We perform a Monte Carlo study of Higgs jets vs QCD jets classification using the spectrum S2(R)S_2(R). We generate pp→Zjpp\rightarrow Zj events and pp→Zhpp\rightarrow Zh events followed by h→bbˉh\rightarrow b\bar{b}, and use the leading jet of the events as training sample of one prong and two prong jets respectively. Each sample is generated at the leading order in QCD using MadGraph5_aMC@NLO 2.6.1 [49] with parton distribution function (PDF) set NNPDF 2.3 LO at αS(mZ)=0.130\alpha_S(m_Z)=0.130 [50]. ZZ bosons are forced to decay into neutrinos so that they are not detected. The events are showered and hadronized by Pythia 8.226 [51] with Monash tune [52]. We include effects of underlying events such as multi-parton interaction and beam remnant treatment but we do not take pile-ups into account.

Finally, we simulate detector response using Delphes 3.3.3 [53] with their default ATLAS configuration. Jets are reconstructed from calorimeter towers using anti-kTk_T algorithm [41] with a jet radius parameter Rjet=1.0R_{\mathrm{jet}}=1.0 implemented in fastjet 3.3.0 [54, 55]. We study substructures of the leading jets with pT,jet∈[300,400]p_{T,\mathrm{jet}}\in[300,400] GeV and mjet∈[100,150]m_{\mathrm{jet}}\in[100,150]. The characteristic angle between two bb quarks from the boosted Higgs boson is then Rbbˉ≳2mh/pT,jet≈0.83R_{b\bar{b}} \gtrsim2m_h/p_{T,\mathrm{jet}} \approx0.83. Hence, the choice of the jet radius is enough to catch parton showers from those two bb quarks efficiently. For ZhZh events, we additionally require that at least there is one bb parton whose momentum in matrix element level is located within R≤1R \le1 from the jet center to remove events with hard initial state radiations. After these preselections, we have 256691 ZhZh and ZjZj events for training ANN. When we test ANN, we use a testing sample generated independently to the training sample.

We show typical pixelated jet images and spectra S2(R;0.1)S_2(R; 0.1) of a Higgs boson jet and a quark jet in Figure 2. For the boosted Higgs boson, the S2(R;0.1)S_2(R; 0.1) distribution has two prominent peaks at R=0R=0 and RbbˉR_{b\bar{b}} which correspond to autocorrelation and cross-correlation of two bb partons respectively,

Pixelated jet images and two-point spectra for a typical Higgs jet and quark jet

Figure 2. The pixelated jet image (left), two-point spectrum S2(R;0.1)S_2(R; 0.1) (center), and the trimmed spectrum S2(R;0.1)S_2(R; 0.1) (right) of a typical Higgs jet (top) and a typical quark jet (bottom). The sum of the bins of each spectrum is normalized to 1. The red reversed triangles, blue triangles, and green plus symbols in the pixelated jet images indicate direction of light quarks, bb quarks and gluons obtained from the matrix element including order αS\alpha_S radiations. We show the Higgs-like probabilities of the jets, ph(YX)p_h(Y_X) of an ANN model with various input XX, defined in § 4, in the central plots.

pT(R⃗)=pT,b⋅δ(R⃗−R⃗b)+pT,bˉ⋅δ(R⃗−R⃗bˉ),(14)p_T(\vec{R}) = p_{T,b}\cdot\delta(\vec{R}-\vec{R}_b) + p_{T,\bar{b}}\cdot\delta(\vec{R}-\vec{R}_{\bar{b}}), \tag*{(14)}
S2(R)=(pT,b2+pT,bˉ2)⋅δ(R)+2pT,bpT,bˉ⋅δ(R−Rbbˉ).S_2(R) = \left(p_{T,b}^2+p_{T,\bar{b}}^2\right)\cdot\delta(R) + 2p_{T,b}p_{T,\bar{b}}\cdot\delta(R-R_{b\bar{b}}).

Because the Higgs boson decays spherically, the two peaks tend to have comparable intensities, pT,b2+pT,bˉ2∼2pT,bpT,bˉp_{T,b}^2+p_{T,\bar{b}}^2 \sim2p_{T,b}p_{T,\bar{b}}, as shown in Figure 2. Parton shower develops along the partons. Each splitting of a parton is characterized by the angle between daughter partons and their momenta. The spectral density S2(R)S_2(R) sums up those individual splittings. The peaks in RR is smeared by the parton shower and hadronization, but it does not change the initial radiation pattern. In Figure 2, we also show the quark jet spectrum. The spectrum does not have distinctive peaks significantly. Instead, it is gradually decreasing as RR increases.

Note that the spectral function S2(R)S_2(R) is not completely independent to the angular structure function ΔG(R)\Delta G(R) in [18]. We have a relation between S2(R)S_2(R) and ΔG(R)\Delta G(R) as follows,

G(R)=∫0RdR′S2(R′)R′2∫0∞dR′S2(R′)R′2.(15)\mathcal{G}(R)=\frac{\int_0^R dR' S_2(R') R'^{2}}{\int_0^\infty dR' S_2(R') R'^{2}}. \tag*{(15)}
ΔG(R)=dlog⁡G(R)dlog⁡R=R⋅S2(R)R2∫0RdR′S2(R′)R′2.(16)\Delta G(R)=\frac{d\log\mathcal{G}(R)}{d\log R}=\frac{R\cdot S_2(R)R^2}{\int_0^R dR' S_2(R')R'^{2}}. \tag*{(16)}

This ΔG(R)\Delta G(R) is a Higuchi’s fractal dimension [56] of G(R)\mathcal{G}(R) which measures irregularity of G(R)\mathcal{G}(R) over RR. QCD jets have a uniform ΔG(R)\Delta G(R) distribution on average because of approximate scale invariance of QCD [18, 24, 25]. On the other hand, G(R)\mathcal{G}(R) of the multi-prong jet shows sharp peaks at some angular scales [18]. Hence, a number of peaks and peak heights in ΔG(R)\Delta G(R) can be used as a classifier of the jets.

In the following sections, we will show two Higgs jet classifiers using the spectrum S2(R)S_2(R). In § 3, we show a cut-based analysis using S2(R)S_2(R). In § 4, we introduce neural network classifiers using S2(R)S_2(R).

Cut-Based Spectral Analysis

Before presenting our neural network classifier for the Higgs jets, we show a cut-based analysis using S2(R)S_2(R) to get a quick insight on the spectrum. Namely, we introduce a ratio of the activity on the characteristic angular scale R^bbˉ\widehat{R}_{b\bar{b}} of a Higgs jet and that of the surrounding angular scales,

RS2=∫R^bbˉmin⁡[aR^bbˉ,Rjet]dR S2(R)∫0aR^bbˉdR S2(R)+∫min⁡[a′R^bbˉ,Rjet]∞dR S2(R),(17)R_{S_2}=\frac{\int_{\widehat{R}_{b\bar{b}}}^{\min[a\widehat{R}_{b\bar{b}},R_{\mathrm{jet}}]}dR\,S_2(R)}{\int_0^{a\widehat{R}_{b\bar{b}}}dR\,S_2(R)+\int_{\min[a'\widehat{R}_{b\bar{b}},R_{\mathrm{jet}}]}^\infty dR\,S_2(R)}, \tag*{(17)}

where R^bbˉ=2⋅125 GeV/pT,jet\widehat{R}_{b\bar{b}}=2\cdot125\ \mathrm{GeV}/p_{T,\mathrm{jet}}, a=0.75a=0.75, and a′=1.25a'=1.25. For the upper boundary, we take the minimum between the boundary a′R^bbˉa'\widehat{R}_{b\bar{b}} and the jet radius RjetR_{\mathrm{jet}}. The angular scale beyond RjetR_{\mathrm{jet}} is mainly covered by large angle radiations rather than soft and collinear radiations from the bb partons. The bb partons from the Higgs boson do not emit parton shower in large angle because the whole system is color neutral. Therefore, we restrict the upper bound of the integral in the numerator up to RjetR_{\mathrm{jet}} while include the integral beyond the upper bound to the denominator.

The probability that a QCD jet emits another hard parton at R^bbˉ\widehat{R}_{b\bar{b}} is small; hence, this ratio works as a classifier. Even if a QCD jet accidentally has substructure from parton splittings, the ratio of the momenta of two partons are different from that of Higgs jets. It is typically small,

RS2=2pT,ipT,jpT,i2+pT,j2=2z(1−z)1−2z(1−z),(18)R_{S_2}=\frac{2p_{T,i}p_{T,j}}{p_{T,i}^{2}+p_{T,j}^{2}}=\frac{2z(1-z)}{1-2z(1-z)}, \tag*{(18)}

where zz is a fraction of momentum of the one of the partons, i.e., pT,i=zpT,q/gp_{T,i}=zp_{T,q/g} and pT,j=(1−z)pT,q/gp_{T,j}=(1-z)p_{T,q/g}. The fraction zz tends to be much smaller than 1 and suppresses the numerator. In contrast, the phase-space of the Higgs boson decay is symmetric; hence, the two terms in Eq. (14) are in a similar order and RS2R_{S_2} is approximately 1. Therefore, the typical Higgs jets and QCD jets have different RS2R_{S_2} values as in Figure 3 and the ratio works as a classifier.

Figure 3. Histogram of RS2R_{S_2} of the Higgs jets and the QCD jets in the training sample. The bin height is the probability distribution function (PDF) of the given bin.

To compare the performance of RS2R_{S_2} with other observables, we consider an observable D2D_2 [16] for the two-prong substructure identification. The variable D2D_2 is defined by a ratio of two-point and three-point energy correlation functions e2βe_2^\beta and e3βe_3^\beta as follows,

e2β=1pT,jet2∑i,j∈jeti<jpT,ipT,jRijβ,(19)e_2^\beta= \frac{1}{p_{T,\mathrm{jet}}^2}\sum_{\substack{i,j\in\mathrm{jet}\\i<j}}p_{T,i}p_{T,j}R_{ij}^\beta, \tag*{(19)}
e3β=1pT,jet3∑i,j,k∈jeti<j<kpT,ipT,jpT,kRijβRjkβRkiβ,(20)e_3^\beta= \frac{1}{p_{T,\mathrm{jet}}^3}\sum_{\substack{i,j,k\in\mathrm{jet}\\i<j<k}}p_{T,i}p_{T,j}p_{T,k}R_{ij}^\beta R_{jk}^\beta R_{ki}^\beta, \tag*{(20)}
D2β=e3β(e2β)3,(21)D_2^\beta= \frac{e_3^\beta}{\left(e_2^\beta\right)^3}, \tag*{(21)}

where the summations of e2βe_2^\beta and e3βe_3^\beta run over all jet constituents. We consider angular exponents β=0,5,1,2\beta= 0, 5, 1, 2 and 44 for further discussion, but we focus on β=2\beta= 2 when we discuss analyses with a single D2D_2. The Higgs jets have a small D2D_2 value because e3βe_3^\beta is suppressed by collinear and soft radiations while e2βe_2^\beta is large because the pairs of jet constituents with Rij∼RbbˉR_{ij} \sim R_{b\bar{b}} dominate. The QCD jets do not have such suppression, and hence, the Higgs jets and the QCD jets cover different regions of D2D_2.

Since RS2R_{S_2} and D2β=2D_2^{\beta=2} are sensitive to the two-prong substructure, moderate correlation between them is expected. We show histograms of (D2β=2,RS2)\left(D_2^{\beta=2}, R_{S_2}\right) of the training sample, in Figure 4. The Higgs jets prefer small D2D_2 and large RS2R_{S_2}, while the QCD jets prefer large D2D_2 and small RS2R_{S_2}. Hence, there is correlation.

Figure 4. Histograms of (D2β=2D_2^{\beta=2}, RS2R_{S_2}) of the Higgs jets and the QCD jets in the training sample.

We show the receiver operating characteristic (ROC) curves of RS2R_{S_2} and D2β=2D_2^{\beta=2} in Figure 5 to compare the classification performance. Since RS2R_{S_2} and D2β=2D_2^{\beta=2} are correlated, they show similar performance. At the Higgs tagging efficiency 0.2 (0.4), the QCD jet mistag rate is 0.0506 (0.135) for D2β=2D_2^{\beta=2} and 0.0525 (0.145) for RS2R_{S_2}.

Figure 5. ROC curves of D2D_2 and RS2R_{S_2} for classifying Higgs jets and QCD jets.

Note that RS2R_{S_2} is not a unique classifier. However, building the sophisticated variable from S2(R)S_2(R) is not the scope of this paper. Instead, we will let the neural network build an optimized variable from S2(R)S_2(R) for the Higgs jet classification.

Spectral Analysis with Artificial Neural Networks

We now feed the event-by-event binned spectra S2(R;ΔR)S_2(R;\Delta R) to our ANN and build a neural network classifier between the Higgs jets and QCD jets. First, we prepare an equal number of Higgs jets and QCD jets to avoid overfitting from unbalanced data. We use TFLearn [57] with backend TensorFlow [58] for the ANN analysis. An input set we consider includes S2(R;0.1)S_2(R;0.1) up to angular scale R<2R<2,

{xi}S2={pT,jet,mjet,S2(0;0.1),⋯ ,S2(1.9;0.1)}.(22)\{x_i\}_{S_2}=\{p_{T,\mathrm{jet}},m_{\mathrm{jet}},S_2(0;0.1),\cdots,S_2(1.9;0.1)\}. \tag*{(22)}

Note that R=2R=2 is the diameter of our jet definition. All the input data {xi}\{x_i\} are standardized, i.e., xi→(xi−xˉi)/σ(xi)x_i \to(x_i-\bar{x}_i)/\sigma(x_i), where xˉi\bar{x}_i and σ(xi)\sigma(x_i) are the mean and the standard deviation of xix_i of the whole training sample including both Higgs jets and QCD jets. The network is configured with four hidden layers having (400, 300, 200, 100) nodes with the ReLU activation functions, f(x)=max⁡(0,x)f(x)=\max(0,x), and an output layer with two nodes having the softmax activations which map inputs to a Higgs-like score yS2y_{S_2}. To avoid overtraining, we insert dropout layers [59] with rate 20% between each hidden layer. The network is trained by Adam optimizer [60] with learning rate 0.001, β1=0.99\beta_1=0.99 and β2=0.999\beta_2=0.999 minimizing a categorical cross-entropy as a loss function,

L=1Nevents∑{−log⁡[yS2({xi})]Higgs jets,−log⁡[1−yS2({xi})]QCD jets,(23)L=\frac{1}{N_{\mathrm{events}}}\sum\begin{cases}-\log\left[y_{S_2}(\{x_i\})\right] & \text{Higgs jets},\\ -\log\left[1-y_{S_2}(\{x_i\})\right] & \text{QCD jets},\end{cases} \tag*{(23)}

where NeventsN_{\mathrm{events}} is the number of events in the training event set. We call this network as NS2\mathcal{N}_{S_2}. In the trained network, Higgs jets have scores near 1, while QCD jets have scores near 0. We validate NS2\mathcal{N}_{S_2} using the testing samples.

We compare the performance of NS2\mathcal{N}_{S_2} with that of a network trained with D2D_2. We prepare another neural network which maps following inputs to the Higgs-like score yD2y_{D_2},

{xi}D2={pT,jet,mjet,D2β=2}.(24)\{x_i\}_{D_2}=\{p_{T,\mathrm{jet}},m_{\mathrm{jet}},D_2^{\beta=2}\}. \tag*{(24)}

Again, the input data is normalized to [0,1][0,1], i.e., xi→(xi−min⁡xi)/(max⁡xi−min⁡xi)x_i\to(x_i-\min x_i)/(\max x_i-\min x_i), where max⁡xi\max x_i and min⁡xi\min x_i are the maximum and the minimum of xix_i in the training sample respectively. We use smaller hidden layers (100, 100) ReLU nodes because the number of inputs is smaller. The other ANN setups are identical to the S2(R)S_2(R) analysis. We call this network as ND2\mathcal{N}_{D_2}.

In our approach, we do not use individual pixels as inputs; therefore, the binned spectrum is affected by both soft and hard calorimeter activities. To make ANN learn a hierarchy between soft and hard radiations, variables after jet trimming [61] are useful. To obtain trimmed quantities, we first reconstruct kTk_T subjets [37, 38] with Rsub=0.2R_{\mathrm{sub}}=0.2 from constituents of the jet and remove subjets having transverse momentum pT,subjet<fcut⋅pT,jetp_{T,\mathrm{subjet}}<f_{\mathrm{cut}}\cdot p_{T,\mathrm{jet}}, where fcut=0.05f_{\mathrm{cut}}=0.05. In the right panel of Figure 2, we show typical two-point correlation spectra of the trimmed jet constituents, S2,tr(R)S_{2,\mathrm{tr}}(R), of a Higgs jet and a QCD jet. The spectra S2(R)S_2(R) before trimming are shown in the central panel. The two-prong substructure of a Higgs jet is consisted by hard activities, and the double peak structure appears both in S2(R)S_2(R) and S2,tr(R)S_{2,\mathrm{tr}}(R). On the other hand, the spectrum of a QCD jet is significantly changed after trimming, which means that soft activities dominate the spectrum S2(R)S_2(R). This shows a difference between S2(R)S_2(R) and S2,tr(R)S_{2,\mathrm{tr}}(R) contains useful information for the classification.

We then prepare three networks ND2+tr\mathcal{N}_{D_2+\mathrm{tr}}, ND2β+tr\mathcal{N}_{D_2^\beta+\mathrm{tr}} and NS2+tr\mathcal{N}_{S_2+\mathrm{tr}} taking inputs {xi}D2+tr\{x_i\}_{D_2+\mathrm{tr}}, {xi}D2β+tr\{x_i\}_{D_2^\beta+\mathrm{tr}} and {xi}S2+tr\{x_i\}_{S_2+\mathrm{tr}} respectively,

{xi}D2+tr={xi}D2∪{pT,jet,tr,mjet,tr,D2,trβ=2},(25)\{x_i\}_{D_2+\mathrm{tr}}=\{x_i\}_{D_2}\cup\{p_{T,\mathrm{jet,tr}},m_{\mathrm{jet,tr}},D_{2,\mathrm{tr}}^{\beta=2}\}, \tag*{(25)}
{xi}D2β+tr={xi}D2+tr∪{D2β=0.5,D2β=1,D2β=2,D2β=4,D2,trβ=0.5,D2,trβ=1,D2,trβ=2,D2,trβ=4},(26)\{x_i\}_{D_2^\beta+\mathrm{tr}}=\{x_i\}_{D_2+\mathrm{tr}}\cup\{D_2^{\beta=0.5},D_2^{\beta=1},D_2^{\beta=2},D_2^{\beta=4},D_{2,\mathrm{tr}}^{\beta=0.5},D_{2,\mathrm{tr}}^{\beta=1},D_{2,\mathrm{tr}}^{\beta=2},D_{2,\mathrm{tr}}^{\beta=4}\}, \tag*{(26)}
{xi}S2+tr={xi}S2∪{pT,jet,tr,mjet,tr,S2,tr(0.0:0.1),⋯ ,S2,tr(1.9:0.1)},(27)\{x_i\}_{S_2+\mathrm{tr}}=\{x_i\}_{S_2}\cup\{p_{T,\mathrm{jet,tr}},m_{\mathrm{jet,tr}},S_{2,\mathrm{tr}}(0.0:0.1),\cdots,S_{2,\mathrm{tr}}(1.9:0.1)\}, \tag*{(27)}

where the variables calculated after trimming have a subscript tr\mathrm{tr}. The other ANN setups are the same as that of ND2\mathcal{N}_{D_2} and NS2\mathcal{N}_{S_2}. The networks ND2+tr\mathcal{N}_{D_2+\mathrm{tr}}, ND2β+tr\mathcal{N}_{D_2^\beta+\mathrm{tr}} and NS2+tr\mathcal{N}_{S_2+\mathrm{tr}} give us Higgs-like scores yD2+try_{D_2+\mathrm{tr}}, yD2β+try_{D_2^\beta+\mathrm{tr}} and yS2+try_{S_2+\mathrm{tr}}. Note that ND2β+tr\mathcal{N}_{D_2^\beta+\mathrm{tr}} takes the D2βD_2^\beta with various angular exponents β=0.5\beta=0.5, 1, 2, and 4 as inputs.

To compare the information contained in D2D_2 and S2(R)S_2(R), we show the ROC curves of our ANN analyses with ND2+tr\mathcal{N}_{D_2+\mathrm{tr}}, ND2+tr\mathcal{N}_{D_2+\mathrm{tr}}, ND2β+tr\mathcal{N}_{D_2^\beta+\mathrm{tr}}, NS2\mathcal{N}_{S_2}, and NS2+tr\mathcal{N}_{S_2+\mathrm{tr}} in Figure 6. The ROC curves show that the ANN with S2(R;ΔR)S_2(R;\Delta R) rejects more QCD jets for a fixed Higgs jet efficiency. At the Higgs tagging efficiency 0.4 (0.2), the QCD jet mistag rate of NS2+tr\mathcal{N}_{S_2+\mathrm{tr}}, which shows the best performance among the classifiers, is 0.0246 (0.0807). The mistag rate is reduced by 51.4% (40.2%) compared to that of the cut-based analysis of RS2R_{S_2} while the mistag rate is still reduced by 17.1% (10.5%) compared to that of ND2β+tr\mathcal{N}_{D_2^{\beta}+tr}. This is expected because NS2+tr\mathcal{N}_{S_2+tr} uses two-point energy correlation from S2(R;ΔR)S_2(R;\Delta R) and infers three-point energy correlation from correlations between different angular scales. For example, S2(R;ΔR)S_2(R;\Delta R) of a three-prong jet having angular scales R1R_1, R2R_2, and R3R_3 has three peaks away from R=0R=0. The intensity of each peak gives a three-point energy correlation function, e3β≈pT,jetβ⋅(ΔR)βS2(R1;ΔR)S2(R2;ΔR)S2(R3;ΔR)R1βR2βR3βe_3^\beta\approx p_{T,\mathrm{jet}}^\beta\cdot\sqrt{(\Delta R)^\beta S_2(R_1;\Delta R)S_2(R_2;\Delta R)S_2(R_3;\Delta R)R_1^\beta R_2^\beta R_3^\beta}. Hence, NS2\mathcal{N}_{S_2} and NS2+tr\mathcal{N}_{S_2+tr} have better discrimination power than ND2+tr\mathcal{N}_{D_2+tr}, ND2+tr\mathcal{N}_{D_2+tr} and ND2β+tr\mathcal{N}_{D_2^\beta+tr}. Also, adding trimmed observables allows ANN to learn hard and soft substructures separately. Hence, the ANN solve degeneracy in the variables before trimming and reject QCD jets better. One interesting feature is that ND2β+tr\mathcal{N}_{D_2^\beta+tr} suppress QCD jets about as equal as NS2\mathcal{N}_{S_2} in the region of high tagging efficiency, but the difference between ND2β+tr\mathcal{N}_{D_2^\beta+tr} and NS2\mathcal{N}_{S_2} is large in the region of low tagging efficiency. This gap in the ROC curves implies that S2(R)S_2(R) has additional information, which we will be discussed in the later part.

Figure 6. ROC curves of ANN classifiers for Higgs jets and QCD jets with inputs Eq. (22), Eq. (24), Eq. (25), Eq. (26), and Eq. (27). We overlay the cut-based ROC curves of RS2R_{S_2} shown in Figure 5 as a gray dashed line for a reference.

Note that the ANN inputs include pT,jetp_{T,\mathrm{jet}}; as a result, the ANN partially use the jet pTp_T distribution of Higgs jets and QCD jets for the classification. To remove the pTp_T dependence, we may resample the training set whose signal and background pTp_T distribution is same, or use a weighted loss function for an imbalanced training set, or train adversarial neural networks [62, 63]. For the comparison of the ANN’s with different inputs, we do not need to use these techniques because pT,jetp_{T,\mathrm{jet}} is a common input for all the ANN analyses. We have to pay attention to the bias from pT,jetp_{T,\mathrm{jet}} distribution of signal and background when this analysis is applied to the experimental data.

The network ND2+tr\mathcal{N}_{D_2+tr} uses a small number of inputs; therefore, it is easy to check the ANN outputs of the input parameters mjetm_{\mathrm{jet}}, pT,jetp_{T,\mathrm{jet}}, and D2D_2. We show distributions of the events in the Higgs-like probability ph(YD2+tr)p_h(Y_{D_2+tr}) and one of the inputs {xi}D2+tr\{x_i\}_{D_2+tr} plane in Figure 7. The Higgs-like probability of NX\mathcal{N}_X, ph(YX)p_h(Y_X), is defined by the probability of getting a score yXy_X less than YXY_X in the Higgs jet samples,

Figure 7. The distributions of Higgs jets and QCD jets in the Higgs-like probability ph(YD2+tr)p_h(Y_{D_2+tr}) and inputs in {xi}D2+tr\{x_i\}_{D_2+tr} planes. For each pair of figures, a left panel is the distribution of Higgs jets and a right panel is the distribution of QCD jets.

ph(YX)=Pr⁡(yX<YX∣Higgs jets),(28)p_h(Y_X) = \operatorname{Pr}(y_X < Y_X \mid\text{Higgs jets}), \tag*{(28)}

where Pr⁡(C∣H)\operatorname{Pr}(C\mid H) represents the conditional probability of CC given HH. A large ph(YX)p_h(Y_X) means that the given jet is more Higgs-like in NXN_X. Figure 7 shows that ND2+trN_{D_2+tr} tries to select jets having mjetm_{\mathrm{jet}} around 125 GeV up to energy loss, and small D2D_2 for capturing two-prong jets. Note that the ND2+trN_{D_2+tr} is trained by the QCD jets and Higgs jets with mh=125m_h = 125 GeV. The mjetm_{\mathrm{jet}} and trimmed mjetm_{\mathrm{jet}} of Higgs jet is tend to be higher (lower) than the mass of Higgs boson and a jet with mjet∼120m_{\mathrm{jet}} \sim120 GeV is regarded most likely as a Higgs jet. These shifts of the mass from the input Higgs mass are due to contamination of other activities, or large angle radiations from bb jets. The mistagged QCD jets with high ph(YD2+tr)p_h(Y_{D_2+tr}) cover similar phase-space, where D2,trD_{2,tr} is small, mjet,tr∼115m_{\mathrm{jet,tr}} \sim115 GeV, and mjet∼125m_{\mathrm{jet}} \sim125 GeV. Meanwhile, the mjetm_{\mathrm{jet}} and mjet,trm_{\mathrm{jet,tr}} distributions of QCD jets in ph(YD2+tr)<0.2p_h(Y_{D_2+tr}) < 0.2 and ph(YD2+tr)<0.3p_h(Y_{D_2+tr}) < 0.3 in Figure 7 show relatively low probability for mjet∼130m_{\mathrm{jet}} \sim130 GeV and mjet,tr>100m_{\mathrm{jet,tr}} > 100 GeV, respectively. This means, ND2+trN_{D_2+tr} labels events in the particular mass window as Higgs jet; hence, ND2+trN_{D_2+tr} rejects QCD jets in a similar fashion that a cut-based approach rejects QCD events.

We compare the Higgs-like probabilities in ND2+trN_{D_2+tr} and NS2+trN_{S_2+tr} in Figure 8 to find out the origin of improvement. The events are widely spreading around the line ph(YD2+tr)=ph(YS2+tr)p_h(Y_{D_2+tr}) = p_h(Y_{S_2+tr}), which means that those two analyses have different selection criteria. To quantify the residual of the anticorrelation of ND2+trN_{D_2+tr} and NS2+trN_{S_2+tr}, we show the fractions of the events in the upper triangular region ph(YD2+tr)>ph(YS2+tr)p_h(Y_{D_{2+tr}}) > p_h(Y_{S_{2+tr}}) and the lower triangular region ph(YD2+tr)<ph(YS2+tr)p_h(Y_{D_{2+tr}}) < p_h(Y_{S_{2+tr}}). For Higgs jets, the lower triangular region contains more events compared with the upper triangular region, 51.3% of the total events. For QCD jets, the lower triangular region contains less events, 43.1%. Hence, NS2+tr\mathcal{N}_{S_{2+tr}} improves signal and background ratio S/BS/B from ND2+tr\mathcal{N}_{D_{2+tr}}.

Figure 8. The distributions of Higgs-like probability in ph(YS2+tr)p_h(Y_{S_{2}+tr}) and ph(YD2+tr)p_h(Y_{D_{2}+tr}) planes. The numbers in upper left and lower right area represent the fraction of events in the area.

To figure out how NS2+tr\mathcal{N}_{S_{2+tr}} accepts more Higgs jets while rejecting more QCD jets compared to ND2+tr\mathcal{N}_{D_{2+tr}}, we will show three examples of the events located on the off-diagonal regions in Figure 8. We show a Higgs jet in Figure 9, which is Higgs-like in NS2+tr\mathcal{N}_{S_{2+tr}} but regarded as a QCD jet in ND2+tr\mathcal{N}_{D_{2+tr}}, ph(YD2+tr)=19.6%p_h(Y_{D_{2+tr}}) = 19.6\% and ph(YS2+tr)=80.7%p_h(Y_{S_{2+tr}}) = 80.7\%. This jet has a moderate wide-angle radiation on top of two-prong substructure which increases D2D_2 significantly. Remind that a Higgs jet originates from a color singlet particle while a QCD jet originates from a colored parton. Such wide-angle radiation is easily generated from a colored parton compared to a color singlet particle. ND2+tr\mathcal{N}_{D_{2+tr}} is distracted by a large D2D_2 and assigns this jet as a QCD jet even though the jet has small trimmed D2D_2. NS2+tr\mathcal{N}_{S_{2+tr}} must have determined the jet as Higgs-like from the information of microscopic radiation patterns in S2(R)S_2(R) which shows a clear double peak structure. Figure 10 shows D2D_2 and D2,trD_{2,tr} distributions in events having ph(YD2,tr)<30%p_h(Y_{D_{2,tr}}) < 30\% and ph(YS2,tr)>70%p_h(Y_{S_{2,tr}}) > 70\%. We can see that some, but not all, events with large D2D_2 but small D2,trD_{2,tr} fall into this region.

Pixelated Higgs jet image and two-point spectrum

Figure 9. The pixelated jet image (left), and two-point spectrum S2(R;0.1)S_{2}(R;0.1) (right) of a Higgs jet. The event is Higgs-like in the neural networks with S2(R)S_{2}(R), NS2\mathcal{N}_{S_{2}} and NS2+tr\mathcal{N}_{S_{2+tr}}, but it is not in those with D2D_{2}, ND2\mathcal{N}_{D_{2}} and ND2+tr\mathcal{N}_{D_{2+tr}}. The sum of the bins of the spectrum is normalized to 1. The blue triangles and green plus symbols in the pixelated jet images indicate direction of bb quarks and gluons obtained from the matrix element including order αS\alpha_{S} radiations. We show the Higgs-like probability ph(YX)p_h(Y_X) in an ANN model NX\mathcal{N}_X defined by Eq. (28).

Figure 10. Distributions of Higgs jets (left) and QCD jets (right) in D2D_2 and D2,trD_{2,tr}, classified as Higgs jets in NS2+tr\mathcal{N}_{S_2+tr} but QCD jets in ND2+tr\mathcal{N}_{D_2+tr}. We select events with Higgs-like probabilities ph(YD2,tr)<30%p_h(Y_{D_2,tr}) < 30\% and ph(YS2,tr)>70%p_h(Y_{S_2,tr}) > 70\%.

The second example in Figure 11 is a jet classified as a Higgs jet in ND2+tr\mathcal{N}_{D_{2+tr}} but categorized as a

The pixelated jet image and two-point spectrum of a Higgs jet

Figure 11. The Pixelated jet image (left), and two-point spectrum S2(R;0.1)S_2(R; 0.1) (right) of a Higgs jet. The event is Higgs-like in the neural networks with D2D_2, ND2\mathcal{N}_{D_2} and ND2+tr\mathcal{N}_{D_2+tr}, but it is not in those with S2(R)S_2(R), NS2\mathcal{N}_{S_2} and NS2+tr\mathcal{N}_{S_2+tr}. The sum of the bins of the spectrum is normalized to 1. The red triangles and green plus symbols in the pixelated jet images indicate direction of light quarks and gluons obtained from the matrix element including order αS\alpha_S radiations. We show the Higgs-like probability ph(YX)p_h(Y_X) in an ANN model NX\mathcal{N}_X defined by Eq. (28).

QCD jet in NS2+tr\mathcal{N}_{S_2+tr}. This jet has evident two-prong substructure, and hence, ND2+tr\mathcal{N}_{D_2+tr} classifies this jet as a Higgs jet. However, the two subjets are asymmetric in pTp_T. Such events appear frequently in QCD jet samples. We did not give subjet momenta to ND2+tr\mathcal{N}_{D_2+tr}, and the ANN classify the jet as a Higgs jet. In contrast, S2(R)S_2(R) knows the pTp_T asymmetry by comparing the peak intensity; see Eq. (14). Hence, NS2+tr\mathcal{N}_{S_2+tr} avoids these pTp_T asymmetric events which appear often among QCD jets while NS2+tr\mathcal{N}_{S_2+tr} finds the cut on the subjet pTp_T from the training samples. In the mass-drop tagger [1], the events with asymmetric pTp_T subjets are removed by a cut.

The third example in Figure 12 is the case where only NS2+tr\mathcal{N}_{S_2+tr}, which takes into account trimmed S2(R)S_2(R), classifies the jet as a Higgs jet. This jet has a two-prong substructure but deeply buried in radiations compared to Figure 11. As a result, D2D_2 is large, and S2(R)S_2(R) spectrum is falling toward high angular scale RR as in Figure 2. Trimming helps NS2+tr\mathcal{N}_{S_2+tr} this time because NS2+tr\mathcal{N}_{S_2+tr} recognizes hard and soft substructure separately by comparing S2(R)S_2(R) and S2,tr(R)S_{2,tr}(R). Trimming does not change the tail of S2(R)S_2(R) distribution, which means the substructure at large RR is hard.

Pixelated jet image, two-point spectrum, and trimmed spectrum of a Higgs jet

Figure 12. The Pixelated jet image (left), two-point spectrum S2(R;0.1)S_2(R; 0.1) (center), and trimmed spectrum S2(R;0.1)S_2(R; 0.1) (right) of a Higgs jet which is Higgs-like only in the NS2+tr\mathcal{N}_{S_2+\mathrm{tr}}. The sum of the bins of each spectrum is normalized to 1. The blue triangles and green plus symbols in the pixelated jet images indicate direction of bb quarks and gluons obtained from the matrix element including order αS\alpha_S radiations. We show the Higgs-like probability ph(YX)p_h(Y_X) in an ANN model NX\mathcal{N}_X defined by Eq. (28).

Discussion and Conclusion

In this paper, we have introduced a spectral analysis of jet substructure with the artificial neural networks (ANN). Unlike the other ANN approach, our algorithm use the spectral function S2(R)S_2(R) constructed from pTp_T and RR of the pair of particles in the jet. The spectrum S2(R)S_2(R) is useful in describing substructures with large angular separation by relatively small inputs. ANN can learn non-local correlations in jets from the spectrum. To show this, we have constructed ANN from S2(R)S_2(R), NS2\mathcal{N}_{S_2}, and compare it with ANN from D2D_2, ND2\mathcal{N}_{D_2}. We have shown that NS2\mathcal{N}_{S_2} discriminates between boosted Higgs jets and QCD jets with better performance compared to ND2\mathcal{N}_{D_2}. Introducing trimming to S2(R)S_2(R) further helps ANN separate hard and soft substructures, and the ANN with trimmed observable outperforms the ANN without trimming. The improvement comes from the better handling of the cases with radiation from bb parton or with contamination of other hadronic activities.

The improvement we observe is not large, because D2D_2 catches the two-prong substructure of the Higgs jet efficiently, but ANN analysis with S2(R)S_2(R) has much wider application. One of the merits of NS2\mathcal{N}_{S_2} and NS2+tr\mathcal{N}_{S_2+\mathrm{tr}} is that the analyses automatically take care of radiations from the bb jet. Note that the existence of radiation has to be taken care of even in the original mass drop tagger by [1]. The S2(R)S_2(R) has information on three-point correlation and higher simultaneously and additional selections are not required. Consequently, S2(R)S_2(R) can be used in a cascade decay of a heavy particle, especially the top quark. We also note that S2(R)S_2(R) is sensitive to the color of the boosted heavy particle. We will show in a separate publication that a modified NS2\mathcal{N}_{S_2} discriminate color octet resonance and color singlet resonance efficiently [64].

Acknowledgments

The authors would like to thank Amit Chakraborty, Patrick T. Komiske, Devdatta Majumder, and Tilman Plehn for useful discussions. This work was supported by the Grant-in-Aid for Scientific Research on Scientific Research B (No.16H03991, 17H02878 [MMN]) and Innovative Areas (16H06492 [MMN]), and by World Premier International Research Center Initiative (WPI Initiative), MEXT, Japan. The work of SHL was supported in part by MEXT KAKENHI Grant Number JP16K21730.

References

  1. [1]J. M. Butterworth, A. R. Davison, M. Rubin and G. P. Salam, Jet substructure as a new Higgs search channel at the LHC, Phys. Rev. Lett. 100 (2008) 242001, [0802.2470].DOI
  2. [2]J. Thaler and L.-T. Wang, Strategies to Identify Boosted Tops, JHEP 07 (2008) 092, [0806.0023].DOI
  3. [3]D. E. Kaplan, K. Rehermann, M. D. Schwartz and B. Tweedie, Top Tagging: A Method for Identifying Boosted Hadronically Decaying Top Quarks, Phys. Rev. Lett. 101 (2008) 142001, [0806.0848].pubmed.ncbi.nlm.nih.gov/18851520
  4. [4]T. Plehn, G. P. Salam and M. Spannowsky, Fat Jets for a Light Higgs, Phys. Rev. Lett. 104 (2010) 111801, [0910.5472].DOI
  5. [5]T. Plehn, M. Spannowsky, M. Takeuchi and D. Zerwas, Stop Reconstruction with Tagged Tops, JHEP 10 (2010) 078, [1006.2833].DOI
  6. [6]D. E. Soper and M. Spannowsky, Finding physics signals with shower deconstruction, Phys. Rev. D84 (2011) 074002, [1102.3480].DOI
  7. [7]D. E. Soper and M. Spannowsky, Finding top quarks with shower deconstruction, Phys. Rev. D87 (2013) 054012, [1211.3140].DOI
  8. [8]M. Dasgupta, A. Fregoso, S. Marzani and G. P. Salam, Towards an understanding of jet substructure, JHEP 09 (2013) 029, [1307.0007].
  9. [9]D. E. Soper and M. Spannowsky, Finding physics signals with event deconstruction, Phys. Rev. D89 (2014) 094005, [1402.1189].DOI
  10. [10]A. J. Larkoski, S. Marzani, G. Soyez and J. Thaler, Soft Drop, JHEP 05 (2014) 146, [1402.2657].
  11. [11]J. Gallicchio and M. D. Schwartz, Seeing in Color: Jet Superstructure, Phys. Rev. Lett. 105 (2010) 022001, [1001.5027].DOI
  12. [12]J. Thaler and K. Van Tilburg, Identifying Boosted Objects with N-subjettiness, JHEP 03 (2011) 015, [1011.2268].DOI
  13. [13]J. Gallicchio and M. D. Schwartz, Quark and Gluon Tagging at the LHC, Phys. Rev. Lett. 107 (2011) 172001, [1106.3076].DOI
  14. [14]Y.-T. Chien, Telescoping jets: Probing hadronic event structure with multiple R ’s, Phys. Rev. D90 (2014) 054008, [1304.5240].DOI
  15. [15]A. J. Larkoski, G. P. Salam and J. Thaler, Energy Correlation Functions for Jet Substructure, JHEP 06 (2013) 108, [1305.0007].DOI
  16. [16]A. J. Larkoski, I. Moult and D. Neill, Power Counting to Better Jet Observables, JHEP 12 (2014) 009, [1409.6298].
  17. [17]I. Moult, L. Necib and J. Thaler, New Angles on Energy Correlation Functions, JHEP 12 (2016) 153, [1609.07483].DOI
  18. [18]M. Jankowiak and A. J. Larkoski, Jet Substructure Without Trees, JHEP 06 (2011) 057, [1104.1646].DOI
  19. [19]K. Datta and A. Larkoski, How Much Information is in a Jet?, JHEP 06 (2017) 073, [1704.08249].DOI
  20. [20]A. Chakraborty, A. M. Iyer and T. S. Roy, A Framework for Finding Anomalous Objects at the LHC, Nucl. Phys. B932 (2018) 439–470, [1707.07084].DOI
  21. [21]J. A. Aguilar-Saavedra, J. H. Collins and R. K. Mishra, A generic anti-QCD jet tagger, JHEP 11 (2017) 163, [1709.01087].DOI
  22. [22]P. T. Komiske, E. M. Metodiev and J. Thaler, Energy flow polynomials: A complete linear basis for jet substructure, JHEP 04 (2018) 013, [1712.07124].DOI
  23. [23]Y.-T. Chien and R. Kunnawalkam Elayavalli, Probing heavy ion collisions using quark and gluon jet substructure, 1803.03589.arxiv.org/abs/1803.03589
  24. [24]M. Jankowiak and A. J. Larkoski, Angular Scaling in Jets, JHEP 04 (2012) 039, [1201.2688].DOI
  25. [25]A. J. Larkoski, QCD Analysis of the Scale-Invariance of Jets, Phys. Rev. D86 (2012) 054004, [1207.1437].DOI
  26. [26]K. Datta and A. J. Larkoski, Novel Jet Observables from Machine Learning, JHEP 03 (2018) 086, [1710.01305].DOI
  27. [27]B. Bhattacherjee, S. Mukhopadhyay, M. M. Nojiri, Y. Sakaki and B. R. Webber, Quark-gluon discrimination in the search for gluino pair production at the LHC, JHEP 01 (2017) 044, [1609.08781].
  28. [28]Y.-T. Chien, A. Emerman, S.-C. Hsu, S. Meehan and Z. Montague, Telescoping jet substructure, 1711.11041.DOI
  29. [29]J. Cogan, M. Kagan, E. Strauss and A. Schwartzman, Jet-Images: Computer Vision Inspired Techniques for Jet Tagging, JHEP 02 (2015) 118, [1407.5675].DOI
  30. [30]L. G. Almeida, M. Backović, M. Cliche, S. J. Lee and M. Perelstein, Playing Tag with ANN: Boosted Top Identification with Pattern Recognition, JHEP 07 (2015) 086, [1501.05968].arxiv.org/abs/1501.05968
  31. [31]L. de Oliveira, M. Kagan, L. Mackey, B. Nachman and A. Schwartzman, Jet-images — deep learning edition, JHEP 07 (2016) 069, [1511.05190].DOI
  32. [32]P. Baldi, K. Bauer, C. Eng, P. Sadowski and D. Whiteson, Jet Substructure Classification in High-Energy Physics with Deep Neural Networks, Phys. Rev. D93 (2016) 094034, [1603.09349].DOI
  33. [33]P. T. Komiske, E. M. Metodiev and M. D. Schwartz, Deep learning in color: towards automated quark/gluon jet discrimination, JHEP 01 (2017) 110, [1612.01551].DOI
  34. [34]G. Kasieczka, T. Plehn, M. Russell and T. Schell, Deep-learning Top Taggers or The End of QCD?, JHEP 05 (2017) 006, [1701.08784].
  35. [35]S. Choi, S. J. Lee and M. Perelstein, Infrared Safety of a Neural-Net Top Tagging Algorithm, 1806.01263.DOI
  36. [36]JADE collaboration, W. Bartel et al., Experimental Studies on Multi-Jet Production in e+ e- Annihilation at PETRA Energies, Z. Phys. C33 (1986) 23.
  37. [37]S. Catani, Y. L. Dokshitzer, M. H. Seymour and B. R. Webber, Longitudinally invariant K_t clustering algorithms for hadron hadron collisions, Nucl. Phys. B406 (1993) 187–224.
  38. [38]S. D. Ellis and D. E. Soper, Successive combination jet algorithm for hadron collisions, Phys. Rev. D48 (1993) 3160–3166, [hep-ph/9305266].DOI
  39. [39]Y. L. Dokshitzer, G. D. Leder, S. Moretti and B. R. Webber, Better jet clustering algorithms, JHEP 08 (1997) 001, [hep-ph/9707323].
  40. [40]M. Wobisch and T. Wengler, Hadronization corrections to jet cross-sections in deep inelastic scattering, in Monte Carlo generators for HERA physics. Proceedings, Workshop, Hamburg, Germany, 1998-1999, 1998. hep-ph/9907280.
  41. [41]M. Cacciari, G. P. Salam and G. Soyez, The Anti-k(t) jet clustering algorithm, JHEP 04 (2008) 063, [0802.1189].DOI
  42. [42]G. Louppe, K. Cho, C. Becot and K. Cranmer, QCD-Aware Recursive Neural Networks for Jet Physics, 1702.00748.
  43. [43]T. Cheng, Recursive Neural Networks in Quark/Gluon Tagging, Comput. Softw. Big Sci. 2 (2018) 3, [1711.02633].DOI
  44. [44]S. Egan, W. Fedorko, A. Lister, J. Pearkes and C. Gay, Long Short-Term Memory (LSTM) networks with jet constituents for boosted top tagging at the LHC, 1711.09059.arxiv.org/abs/1711.09059
  45. [45]X. Wang, R. Girshick, A. Gupta and K. He, Non-local neural networks, CVPR (2018) , [1711.07971].arxiv.org/abs/1711.07971
  46. [46]F. V. Tkachov, Measuring multi - jet structure of hadronic energy flow or What is a jet?, Int. J. Mod. Phys. A12 (1997) 5411–5529, [hep-ph/9601308].DOI
  47. [47]T. Kinoshita, Mass singularities of Feynman amplitudes, J. Math. Phys. 3 (1962) 650–677.
  48. [48]T. D. Lee and M. Nauenberg, Degenerate Systems and Mass Singularities, Phys. Rev. 133 (1964) B1549–B1562.
  49. [49]J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer et al., The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton shower simulations, JHEP 07 (2014) 079, [1405.0301].DOI
  50. [50]R. D. Ball et al., Parton distributions with LHC data, Nucl. Phys. B867 (2013) 244–289, [1207.1303].DOI
  51. [51]T. Sjöstrand, S. Ask, J. R. Christiansen, R. Corke, N. Desai, P. Ilten et al., An Introduction to PYTHIA 8.2, Comput. Phys. Commun. 191 (2015) 159–177, [1410.3012].DOI
  52. [52]P. Skands, S. Carrazza and J. Rojo, Tuning PYTHIA 8.1: the Monash 2013 Tune, Eur. Phys. J. C74 (2014) 3024, [1404.5630].DOI
  53. [53]DELPHES 3 collaboration, J. de Favereau, C. Delaere, P. Demin, A. Giammanco, V. Lemaitre, A. Mertens et al., DELPHES 3, A modular framework for fast simulation of a generic collider experiment, JHEP 02 (2014) 057, [1307.6346].
  54. [54]M. Cacciari, G. P. Salam and G. Soyez, FastJet User Manual, Eur. Phys. J. C72 (2012) 1896, [1111.6097].DOI
  55. [55]M. Cacciari and G. P. Salam, Dispelling the N3 myth for the kt jet-finder, Phys. Lett. B641 (2006) 57–61, [hep-ph/0512210].DOI
  56. [56]T. Higuchi, Approach to an irregular time series on the basis of the fractal theory, Physica D: Nonlinear Phenomena 31 (1988) 277 - 283.DOI
  57. [57]A. Damien et al., “TFLearn.” https://github.com/tflearn/tflearn, 2016.
  58. [58]M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro et al., TensorFlow: Large-scale machine learning on heterogeneous systems, 2015.arxiv.org/abs/1603.04467
  59. [59]N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever and R. Salakhutdinov, Dropout: A simple way to prevent neural networks from overfitting, Journal of Machine Learning Research 15 (2014) 1929–1958.
  60. [60]D. P. Kingma and J. Ba, Adam: A method for stochastic optimization, in The 3rd International Conference for Learning Representations, 2014. 1412.6980.arxiv.org/abs/1412.6980
  61. [61]D. Krohn, J. Thaler and L.-T. Wang, Jet Trimming, JHEP 02 (2010) 084, [0912.1342].DOI
  62. [62]G. Louppe, M. Kagan and K. Cranmer, Learning to Pivot with Adversarial Networks, 1611.01046.
  63. [63]C. Shimmin, P. Sadowski, P. Baldi, E. Weik, D. Whiteson, E. Goul et al., Decorrelated Jet Substructure Tagging using Adversarial Neural Networks, Phys. Rev. D96 (2017) 074034, [1703.03507].DOI
  64. [64]A. Chakraborty, S. H. Lim and M. M. Nojiri, Spectral analysis of color charge in two-prong jets with neural networks, work in progress.

Paper details

vlnc:W_q5exqjkv
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 8

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents