Anchoring Heavy-Ion Quark/Gluon Demixing to Proton-Proton Jets

When quarks and gluons cross the quark-gluon plasma (QGP), the hot droplet of deconfined matter formed in heavy-ion collisions, they lose energy. Theory predicts gluons should lose more than quarks, roughly by the ratio of their color charges. Testing this requires separating the two kinds of jet in data without trusting a simulation to label them. The existing data-driven method treats each jet sample (for example photon+jet versus dijet events) as a different blend of two underlying "topics", quark-like and gluon-like, and unmixes them. In heavy-ion data it breaks down where it matters most. Photon+jet samples are small, and the unmixing then systematically overestimates the gluon fraction, as its authors report (Data-driven quark- and gluon-jet modification in heavy-ion collisions). Meanwhile, proton-proton collisions supply millions of well-understood, unquenched jets whose quark and gluon topics have already been extracted from real data (Disentangling Quarks and Gluons with CMS Open Data).

Speaker recognition solved a structurally identical problem twenty years ago. A well-trained "universal background model" (UBM) is fitted on everyone. Each new speaker's model is then obtained by maximum a posteriori (MAP) adaptation, which means nudging the background model's means toward the new data, with each mean moving only as far as the local evidence supports. We propose to treat proton-proton quark and gluon topics as the background model and the heavy-ion topics as the adapted target. The proton-proton jets used as the anchor are matched by energy rank rather than by measured energy, because a quenched 100 GeV jet started life more energetic. The adapted topics then double as a per-jet quark-versus-gluon likelihood-ratio discriminator in heavy-ion events. That discriminator builds a third, purer sample, which in turn makes the unmixing better determined.

Core Claim

Two conditions together turn small-sample heavy-ion quark/gluon demixing from a biased fit into a stable measurement. First, the prior must be the right proton-proton population: jets selected at matched quantile momentum $p_T^{\mathrm{quant}}$, so prior and target describe the same initial partons. Second, adaptation must be restricted to component means with evidence-weighted shrinkage. The claim is falsifiable on three counts. On JEWEL pseudo-experiments at Run-4-like statistics, the gluon-fraction bias should shrink relative to unanchored fits. Injected medium modifications should still be recovered rather than shrunk away. And the method's advantage should disappear if the anchor is taken at fixed measured $p_T$.

Research Program

The method has three roles, one per direction. The physics of energy loss defines the problem and the correct reference sample. MAP adaptation is the estimator. Signal-to-background discrimination is both a downstream use and a feedback loop.

The prior is built as follows. Jets come from PYTHIA proton-proton photon+jet and dijet samples, and in a real-data variant from the CMS 2011A Open Data. We describe each jet by a low-dimensional feature vector: constituent multiplicity (dequantized), groomed radius $R_g$, and $m/p_T$. We then run standard topic demixing in proton-proton collisions, where statistics are ample, as in Jet Topics: Disentangling Quarks and Gluons at Colliders. Each resulting topic distribution is fitted as a Gaussian mixture with $K$ components. This step is done separately in bins of $p_T^{\mathrm{quant}}$. The quantile construction of Sorting Out Quenched Jets maps each heavy-ion jet's energy rank to the proton-proton jet of equal rank. That paper shows substructure modifications look much smaller once jets are compared at matched initial energy. This is exactly why the prior must be indexed this way: a fixed-$p_T$ prior would encode a different parton population and bias the adaptation.

The adaptation runs on the two heavy-ion samples $s\in\{\gamma\text{+jet}, \text{dijet}\}$, which share topics $t\in\{q,g\}$ but differ in quark fraction $f_s$. Expectation-maximization alternates between two steps. The first assigns each jet soft responsibilities over (topic, component). The second updates each component mean $\mu_{tm}$ by the classical relevance rule:

$$\hat{\mu}_{tm} = \frac{n_{tm}}{n_{tm}+r}\,\bar{x}_{tm} + \frac{r}{n_{tm}+r}\,\mu^{pp}_{tm},$$

Here $n_{tm}$ is the summed responsibility of heavy-ion jets for component $m$ of topic $t$, $\bar{x}_{tm}$ is their responsibility-weighted mean feature vector, $\mu^{pp}_{tm}$ is the proton-proton prior mean, and $r$ is the relevance factor. The fractions $f_s$ are updated freely. Covariances and within-topic weights stay at their proton-proton values, following the consistent finding that mean-only adaptation beats full adaptation when data are scarce (Statistical model migration in speaker recognition; Adapted Gaussian models for image classification). The medium modification of each topic is read off as the adapted-minus-prior mean shifts at fixed $p_T^{\mathrm{quant}}$.

The design has a known identifiability risk. With shapes partly frozen, a common mean shift can partially trade off against the fractions. We will map this degeneracy explicitly with profile likelihoods over $f_s$. The relevance factor $r$ is not hand-tuned. It is chosen by held-out likelihood on independent JEWEL pseudo-experiments that the evaluation never touches, and we report sensitivity across a decade of $r$.

The discriminator follows directly from the adapted topics. Each heavy-ion jet gets a score $\Lambda(x)=\log p(x\mid\hat\theta_q)-\log p(x\mid\hat\theta_g)$, the same log-likelihood-ratio test used in UBM speaker verification. Cutting on $\Lambda$ inside the dijet sample yields a quark-enriched third sample. Demixing works best when samples have very different fractions, so we re-run the adaptation with three samples and check whether the fraction estimates tighten. The step is iterated at most once to avoid self-reinforcing bias. All evaluation uses JEWEL 2.1.0 Pb+Pb with PYTHIA 6 references at 0–10% centrality, downsampled to Run-4 statistics as in Brewer, Thaler and Turner, so per-jet parton truth is available.

Baselines

The do-nothing reference is the published pipeline of Data-driven quark- and gluon-jet modification in heavy-ion collisions: DEMIX unmixing plus a skew-normal MCMC fit on the heavy-ion samples alone. It is also the strongest existing data-driven method for this exact question, since nothing else in the file extracts heavy-ion quark/gluon topics without simulation templates. We will rerun it on identical downsampled JEWEL samples, matching the number of pseudo-experiments and giving its MCMC the same compute budget as our expectation-maximization plus $r$-selection. The paper reports qualitatively that gluon fractions are overestimated at low statistics. No reported numerical bias was found on Lacuna for DEMIX on JEWEL photon+jet, so we will measure it ourselves. Beating this baseline means a lower mean absolute error on $f_s$ against JEWEL truth, together with coverage of injected modifications that is no worse.

The second baseline freezes the topics entirely. It uses the proton-proton topics at $p_T^{\mathrm{quant}}$ unchanged and fits only $f_s$, which corresponds to $r\to\infty$. It will probably have the lowest variance but should be biased whenever the medium really modifies topics. Our method must match its variance where modifications are small and avoid its bias where they are large.

The third baseline is supervised. A classifier is trained on JEWEL parton labels. For reference, End-to-End Jet Classification of Quarks and Gluons with the CMS Open Data reports ROC AUC 0.8076 for a supervised network separating proton-proton quark and gluon jets. We will compare AUCs of $\Lambda$ and a supervised gradient-boosted classifier on identical heavy-ion features. The supervised model is an upper bound that relies on the generator, not a competitor we must beat. Getting within a few AUC points without using labels would count as success.

Two ablations isolate the ingredients. The first is identical MAP adaptation with the prior taken at fixed measured $p_T$, which isolates the quantile anchoring. The second is full adaptation of means, covariances and weights, which isolates the mean-only restriction.

Why not just…?

Why not just put an informative proton-proton prior on the existing skew-normal MCMC fit? This is the most serious threat. It is essentially MAP with a parametric prior, and if it does as well, the contribution reduces to "anchor at $p_T^{\mathrm{quant}}$". We include it as a baseline, with priors centered on proton-proton skew-normal fits at the same $p_T^{\mathrm{quant}}$ and widths tuned on the same held-out pseudo-experiments. Our expected advantage is specific: per-component, evidence-dependent shrinkage keeps the poorly populated multiplicity tails near the prior, while the bulk moves freely. A single global prior width cannot do both. If the comparison shows no gain, we will report that the anchoring, not the mixture machinery, carries the effect.

Why not just use the proton-proton topics as fixed templates? This works only if quenching leaves topic shapes unchanged. Brewer, Thaler and Turner find distinct in-medium modifications of the extracted topics, so fixed templates should misattribute shape changes to fraction changes. The injected-modification test below measures this bias directly.

Why not just wait for more heavy-ion photon+jet data? The bias in the file already appears at projected Run-4 statistics, and photon+jet cross sections set the limit. More luminosity shrinks the problem without removing it. Even so, we will report how our gain decays as the sample size grows, since MAP should converge to the unanchored fit as $n_{tm}\gg r$.

Proposed Experiments

The first experiment is a closure test of fraction bias versus sample size. Across hundreds of downsampled JEWEL pseudo-experiments, we plot the error on $f_s$ for every baseline and ablation against heavy-ion photon+jet sample size.

The second, and the decisive check, guards against the main danger of any prior: shrinking toward "no modification". We generate pseudo-data by reweighting JEWEL so that gluon multiplicity is shifted by known amounts. The test is whether the adapted mean shifts recover the injected shift with correct confidence-interval coverage, especially in sparse tails. Good fraction accuracy paired with suppressed modification would falsify the proposal.

The third experiment compares quantile and fixed-$p_T$ anchoring. If the fixed-$p_T$ prior performs as well, the energy-loss reasoning behind the design is wrong for this observable.

The fourth experiment tests the discriminator loop. We report the AUC of $\Lambda$ against the supervised upper bound, and whether adding the $\Lambda$-enriched third sample reduces the variance of the fractions without shifting their means.

Finally, as a real-data sanity check, we apply the proton-proton prior construction to CMS 2011A Open Data. We require it to reproduce the published quark fractions of 70.8% (forward) and 56.1% (central) reported by Komiske et al. before trusting it as an anchor.