Mass-Conditional Christoffel Functions for Bump Hunts Without Sculpting

Particle physicists hunting rare new particles often train an "anomaly score" on ordinary background events and keep only the events that score as unusual. If that score secretly depends on the quantity being searched over, such as the invariant mass of a jet or the energy of an event, the selection reshapes the smooth background spectrum. It can then manufacture a fake peak, an effect called mass sculpting. This proposal builds the anomaly score from a polynomial moment matrix, the table of averages of all products of features up to some degree, and lets that table vary smoothly with mass. Mass enters the score as a conditioning variable rather than an input. The construction comes with an exact algebraic identity: the average score of background events is the same at every mass. Making the varying table valid everywhere requires a semidefinite program, an optimization over positive semidefinite matrices. Its structure is unusual, and both generic fast solvers and low-rank factorizations handle it poorly.

The score is a higher-degree generalization of the Mahalanobis distance. At degree one it reduces to the familiar ellipsoid. At degree two or three it shrink-wraps curved, multi-cluster background shapes, and it grows exponentially fast for points outside the background support. The open question is whether such a score can match deep decorrelated detectors on standard jet benchmarks while sculpting less. A second open question is whether the semidefinite structure of "positive at every mass" can be solved at the sizes this requires.

Core Claim

Let $x \in \mathbb{R}^n$ be a low-dimensional event representation, $m$ the nuisance mass, and $v_d(x)$ the vector of the $s=\binom{n+d}{d}$ polynomial basis functions of degree at most $d$. We use an orthogonal-polynomial basis on standardized features to limit ill-conditioning. The classical empirical Christoffel function scores a point by $v_d(x)^T M_d^{-1} v_d(x)$, with $M_d$ the moment matrix of the background. Sorting out typicality with the inverse moment matrix SOS polynomial introduced this score as an outlier detector and showed that degree one is exactly Mahalanobis distance. The Christoffel function review and the sparsified Christoffel function establish its support dichotomy: polynomial growth inside the support and exponential growth outside. We replace the single matrix with a polynomial matrix in mass, $M(m)=\sum_{k=0}^{K} M_k m^k$, fit to estimate $\mathbb{E}[v_d(x)v_d(x)^T \mid m]$. The score is then

$$Q(x \mid m) = v_d(x)^T M(m)^{-1} v_d(x), \qquad \mathbb{E}\left[Q(x\mid m)\,\middle|\, m\right] = \operatorname{tr}\left(M(m)^{-1}\,\mathbb{E}[v_d v_d^T\mid m]\right) = s .$$

The identity on the right holds at every mass whenever $M(m)$ equals the true conditional moment matrix, so first-moment decorrelation is built in rather than learned through an adversary. Sculpting is governed by tail quantiles, not means, so the identity alone does not guarantee a flat spectrum after a tight cut. Whether conditioning the full moment matrix also flattens the tails in practice is the central empirical question.

The same construction generalizes a trick detector physicists already use. In Multisite event discrimination for the MAJORANA Demonstrator, the ratio of peak current to energy is corrected by a quadratic fit in energy across 100 keV to 3 MeV. That correction is in effect a degree-one statistic normalized by a degree-two polynomial in the nuisance variable. Our score is the multivariate, higher-degree version of that correction.

Research Program

The estimation step is where semidefinite programming becomes necessary rather than decorative. An unconstrained least-squares fit of $M(m)$ to the per-event outer products is cheap, but it can lose positive definiteness between or beyond the training masses. The score inverts $M(m)$, so a near-singular direction anywhere in the mass window produces exploding, mass-dependent scores. Those scores are exactly a sculpting mechanism. We therefore solve

$$\min_{M_0,\dots,M_K} \sum_{i=1}^N \left\| v_d(x_i)v_d(x_i)^T - M(m_i)\right\|_F^2 \quad \text{s.t.} \quad M(m) - \varepsilon I = S_0(m) + (m-a)(b-m)\,S_1(m)\ \ \forall m,$$

where $[a,b]$ is the search window, $\varepsilon>0$ is a certified eigenvalue floor, and $S_0, S_1$ are matrix sum-of-squares polynomials, each represented by a positive semidefinite Gram matrix. The constraint certifies $M(m) \succeq \varepsilon I$ on the whole interval, not just on a grid. The objective depends on the data only through the sums $\sum_i m_i^k v_d(x_i)v_d(x_i)^T$, so one pass over the events suffices and $N$ drops out of the solve. This is the same single-pass property exploited by the streaming detectors in Leveraging the Christoffel function for outlier detection in data streams.

The resulting program has a specific shape. With $s=165$ (eight features at degree three) and $K=4$, the Gram matrices have side about $3s \approx 500$. There are roughly $s^2(2K+1)/2 \approx 1.2\times 10^5$ equality constraints, of the same order as the matrix side squared. That is the "tall dense" regime in which Solving SDP Faster: A Robust IPM Framework attains $O(m^\omega\log(1/\epsilon))$. It is also the regime in which Burer-Monteiro factorization degenerates. The no-spurious-minima guarantee of The non-convex Burer-Monteiro approach works on smooth semidefinite programs needs rank $p(p+1)/2 > m$, which here forces $p$ close to the full side. That paper's guarantees also exclude the inequality-style structure we have.

What generic analyses ignore is that our constraint map is a Kronecker product of a small Hankel-type operator in the mass-polynomial index with the identity on the $s\times s$ entries. The Newton normal equations therefore decompose into $O(s^2)$ independent small systems coupled only through the Gram blocks. We will derive and implement an interior-point step that uses this structure and compare its cost with the generic bound of A Faster Interior Point Method for Semidefinite Programming, whose $mn^2$ input-reading term our structure avoids. We will also compare against a first-order alternative, the conditional gradient homotopy method. The hierarchical-matrix moment relaxation solver shows structured moment SDPs can be compressed in other settings. It does not treat matrix-valued polynomial positivity in a nuisance variable.

The representation $x$ needs to be low-dimensional, because $s$ grows combinatorially. Raw $50\times 50$ jet images are out of reach, and the research file notes moment-SOS methods stall above about 20 variables (Optimal Data Fitting: A Moment Approach). We follow the protocol of Robust Anomaly Detection for Particle Physics Using Multi-Background Representation Learning. That work trains a classifier on QCD versus W/Z backgrounds and scores events by Mahalanobis distance on its representation (nurd-md), a degree-one moment model. We project that representation to $n \in \{4,8,12\}$ dimensions and replace the ellipsoid with the mass-conditional Christoffel score at $d\in\{1,2,3\}$. A second, physics-native instantiation uses the four handcrafted topological variables of PandaX-III topological discrimination with energy as the nuisance. There the dimension is tiny and the question is purely whether multivariate correlations, which that paper says go unexploited, add rejection without distorting the spectrum near the decay endpoint.

Baselines

The naive reference point is the unconditioned Christoffel score of Sorting out typicality at the same degree and representation, computed from one global moment matrix that ignores mass. It shows how much sculpting a higher-degree moment score produces when nothing is done about it. Beating it means a substantially lower post-selection Jensen-Shannon divergence (JSD) of the background mass spectrum at equal or better area under the ROC curve (AUROC).

The strongest existing methods are the NuRD models of Robust Anomaly Detection for Particle Physics Using Multi-Background Representation Learning on the hls4ml jet dataset. NuRD distills a background classifier with a mutual-information adversary so that its scores are independent of jet mass. That paper reports AUROC 0.914 and significance improvement 2.32 for nurd-ml, and JSD 0.118 for nurd-md, against a VAE baseline at AUROC 0.881 and JSD 0.255. These are the numbers to beat. The file does not state the cut at which JSD was measured, so we will rerun both NuRD variants and the VAE in our pipeline. All methods will share the same splits, the same representation network, the same background efficiency for the JSD cut, and the same hyperparameter budget, set as a fixed number of configurations for each method. Our method beats NuRD if, at matched JSD, it matches nurd-ml AUROC within run-to-run noise or exceeds it. Equally acceptable is AUROC at least that of nurd-md with clearly lower JSD. Matching only nurd-md at degree one would show nothing new.

Two ablations isolate the key ingredients. The first is binned conditioning: a separate Christoffel matrix per mass bin with eigenvalue clipping. It tests whether a smooth polynomial matrix with a certified floor matters. The second is the same polynomial-in-mass least-squares fit with pointwise eigenvalue clipping in place of the semidefinite constraint. It tests whether the SDP itself earns its cost. A degree sweep at fixed conditioning tests whether going beyond Mahalanobis helps. For the solver question, the baselines are a generic off-the-shelf interior-point solver and Burer-Monteiro with the Riemannian staircase on a penalized version of the same program. We compare wall-clock time and memory to a fixed duality gap.

Why not just…?

Why not just compute a Christoffel score in mass bins and quantile-normalize within each bin? This is the most serious threat, because per-bin quantile flattening decorrelates the score by construction wherever bins are well populated. Its weakness is the mass tails, where bins are sparse, per-bin moment matrices are singular, and the quantile maps are noisy. A resonance search scans exactly those regions. The decisive experiment subsamples the background training set by factors of 10 and 100 and measures JSD and false-bump rate in the tails. If the binned method matches us at all sample sizes, the polynomial-in-mass model offers no advantage. We would then report the contribution as a conditioned higher-degree score, with the SDP reduced to a solver study.

Why not just fit $M(m)$ by least squares and clip negative eigenvalues? This is cheaper and may suffice. Clipping is applied pointwise, so it gives no guarantee between evaluation points, and the clipped matrix can change non-smoothly with mass. Through the inverse, that produces mass-dependent score jumps. The clipping ablation measures this directly. It counts how often the unconstrained fit loses definiteness inside the window and compares the score variance across mass. If the two variants are indistinguishable, we drop the claim that certification matters for physics.

Why not just use NuRD's deep decorrelated model, which already works? NuRD reaches the best reported AUROC, but its decorrelation is enforced by an adversary during training and is approximate. The VAE comparison in the same paper shows unconstrained deep scores sculpt badly, with JSD 0.255. In the LHC Olympics 2020, most methods missed the Black Box 3 signal, whose characteristics deviated from simulated training. Our score is closed-form given the representation, has an exact conditional-mean identity, and needs no adversary tuning. The fair test is head-to-head on identical representations. If NuRD dominates at every JSD level, the moment approach adds no value on this benchmark.

Why not just stay at degree one, which is cheap? On KDD Cup 99, Sorting out typicality found that moving from degree one to three improved precision-recall on multimodal data. The hls4ml background is a mixture of QCD and W/Z jets, which is plausibly multimodal in representation space. The degree sweep settles this. If degree two and three give no AUROC gain over degree one at matched JSD, the combinatorial cost is unjustified.

Proposed Experiments

The primary experiment runs on the hls4ml LHC jet dataset with QCD and W/Z as background and top jets as the out-of-distribution signal. The search window is set to the jet-mass range used by the NuRD paper. We report AUROC, significance improvement, and post-selection mass JSD at several background efficiencies, so sculpting is shown as a curve rather than a single point. The core check is whether degree two or three with mass conditioning lands on or beyond the NuRD AUROC-versus-JSD frontier, which the reported numbers anchor at 0.914 AUROC (nurd-ml) and 0.118 JSD (nurd-md).

The second experiment uses the LHC Olympics R&D dataset and Black Boxes 1–3. It runs a bump hunt on the selected spectrum and records whether a spurious peak appears in background-only Black Box 2, whether the Black Box 1 resonance is recovered, and how the method behaves on Black Box 3. The file reports no quantitative scores for these boxes, only which methods detected the resonances.

The third experiment uses PandaX-III-style topology variables simulated in the REST framework, with energy as the nuisance. It tests whether the conditioned multivariate score improves on cut-based rejection at fixed signal efficiency without distorting the spectrum near the endpoint. The reference point is the reported seven orders of magnitude of suppression at above 20% efficiency. No reported result was found in the research file for a multivariate method on this simulation.

The solver study measures time and memory to a fixed gap as $s$ grows from about 35 to about 1000 and $K$ from 2 to 6. It compares the Kronecker-structured interior-point step, a generic solver, conditional-gradient homotopy, and Burer-Monteiro. It also records the rank Burer-Monteiro needs, to test the predicted rank inflation.

The project fails in a clean, informative way if the binned or clipped ablation matches the full method on tail JSD. In that case the useful result is the conditioned higher-degree score itself, without the SDP.