Count-Aware Expansion Order for Sparse-Data Unfolding and Rare-Event Surrogates

Many physics measurements are limited by how few events were seen. The noise in a count of $N$ events is about $\sqrt{N}$, so the relative error is $1/\sqrt{N}$. In this regime, two different machine-learning tools quietly make the same fixed choice: how many terms to keep in a series expansion. Moment Unfolding recovers the true distribution of a collider observable from detector-blurred data by fitting its first $n$ moments, meaning its mean, its spread, and so on. RESOLVE builds a cheap stand-in for an expensive binary-star simulator by fitting a polynomial of fixed degree. Both papers fix the number of terms ahead of time. Moment Unfolding explicitly reports that adding moments captures more shape but ruins the precision of each estimate when data are sparse. RESOLVE shows that higher-order polynomials are needed so that secondary physics is not drowned out by the dominant parameter.

We propose to let the data choose the expansion order. The rule adds a term only when the evidence for that term clearly exceeds its own counting noise, and it uses a stricter bar the more candidate terms it has tried, the same way a particle-physics search corrects for looking in many places at once. The result is an order that differs by observable, by kinematic region, and by parameter direction. Sparse tails get short expansions and well-populated regions get longer ones. The rule's only setting is a target rate of falsely included terms, which a physicist can state directly; nothing is tuned by hand. The question is whether this gives correctly calibrated error bars at lower cost than the obvious alternatives. Those alternatives are cross-validation, sparsity penalties, and simply picking the best fixed order.

Core Claim

In Poisson-limited regimes, choosing the expansion order is itself a small hypothesis search, and it should be treated as one. When the order is selected adaptively, with a threshold set by a trial-factor correction and a two-threshold rule so that terms do not flip in and out, we claim three things. Coverage of the reported uncertainties will stay near nominal across a wide range of $N$. Point error will be no worse than the best fixed order, which can only be found in hindsight. And the rule will cost far fewer fits than cross-validating over per-direction orders.

The new element is not stepwise selection in general. It is the combination of three things. The noise floor comes from effective event counts, not from a Gaussian residual. The number of candidate terms enters the threshold the way the look-elsewhere effect enters a bump hunt. And the rule is applied inside two specific estimators that currently fix their order by hand.

Relation to Prior Work

Two papers supply both the problem and the testbeds.

Moment Unfolding reweights simulated events with a maximum-entropy function, $g(z) \propto \exp(-\sum_{a=1}^{n} \beta_a z^a)$. It trains this function adversarially against a classifier and recovers truth-level moments about 100 times faster than the iterative OmniFold method. It leaves $n$ static and names the resulting trade-off between bias and variance as unresolved.

RESOLVE denoises hit-or-miss low-fidelity runs of the COMPAS binary-evolution simulator with a conditional neural process. It then corrects them with 15 high-fidelity runs through a multi-fidelity Bayesian polynomial chaos expansion of fixed order 4. It reports coverage of 74%, 97%, and 100% at 1, 2, and 3 standard deviations. It also shows that Gaussian-process surrogates flatline on secondary parameters such as natal kick dispersion.

The two papers pull in opposite directions. RESOLVE argues for more terms and Moment Unfolding warns against them. Neither offers a way to decide between these per region. We do not change either estimator; we only replace its fixed order with a rule.

Direction 1 also motivates the threshold. The Telescope Array hotspot analysis (Telescope Array anisotropy) is a case where accounting for multiple tries deflates a local excess below discovery level. Low-significance results such as the ψ(2S) branching fractions (psi(2S) branching fractions) are known to swing under small changes to the data. An order-selection rule that ignores how many terms it tried is the expansion-fitting version of an uncorrected bump hunt, and it should show the same instability.

Direction 2 constrains the design more than it supplies it. The theory-guided benchmarking study (theory-guided benchmarking of (1+λ) EA variants) found that a self-adjusting evolutionary algorithm with proven asymptotic guarantees was the worst performer at practical problem sizes. The ACM work on balancing gradient and subproblem time (ACM) flags noisy adaptive signals as an open vulnerability. Both are warnings: adaptation driven by noisy feedback can oscillate or simply cost more than it saves.

We borrow the remedy from the nonsmooth nonconvex trust-region method. That method accepts or rejects steps by comparing actual against predicted progress, with separate expansion and contraction thresholds and explicit safeguards. Our rule uses the same two-threshold structure. Unlike that method, our acceptance test is set against counting noise rather than a deterministic model.

The Rule

Suppose the current expansion has a set of active terms. A candidate term $k$ is one more moment, or one more polynomial degree along one input parameter. For each candidate we fit the expanded model and estimate its coefficient $\hat\beta_k$. We estimate the coefficient's standard error $\hat\sigma_k$ by Poisson bootstrap, which resamples each event with a Poisson(1) weight, and we use the effective count $N_{\mathrm{eff}} = (\sum w_i)^2/\sum w_i^2$ of the reweighted sample so that weights are not mistaken for data. The term is admitted when

$$ \frac{|\hat\beta_k|}{\hat\sigma_k} > \tau(m), \qquad \tau(m) = \Phi^{-1}\!\left(1 - \frac{\alpha}{2m}\right). $$

Here $m$ is the number of candidate terms tested so far, $\alpha$ is the target probability of admitting any spurious term, and $\Phi^{-1}$ is the inverse standard normal distribution function. This form is the standard Šidák/Bonferroni-style trial correction applied to a two-sided test.

An active term is removed only if its ratio falls below $\tau(m)/2$. This gap is the hysteresis that prevents flip-flopping under resampling. The search stops when no candidate passes, or when a cost cap is reached.

Because $\hat\sigma_k$ shrinks roughly as $1/\sqrt{N_{\mathrm{eff}}}$, the admitted order should grow with $N$. A fixed true signal in the $k$-th term becomes detectable once $N_{\mathrm{eff}}$ exceeds roughly $(\tau\sigma_1/\beta_k)^2$, where $\sigma_1$ is the per-event spread of that coefficient's estimator. This prediction is directly testable.

For Moment Unfolding, a "term" is the next moment for a given observable and transverse-momentum bin. For RESOLVE, a "term" is a degree increase along one of the inputs: metallicity, kick dispersion, or common-envelope efficiency. The result is an anisotropic expansion, with different degrees along different inputs, rather than a single global order of 4.

Research Program

We will implement the rule as a thin wrapper around the published Moment Unfolding training loop and around a reimplementation of the RESOLVE multi-fidelity expansion. We will then study its behavior in three settings.

The first setting is synthetic. Truth distributions are built to have exactly $k$ nonzero maximum-entropy exponents, and detector smearing is known. Here we can check directly whether the false-inclusion rate matches $\alpha$ and whether the admitted order tracks the predicted growth with $N$.

The second is the jet-substructure setting of Moment Unfolding at 14 TeV, using jet mass, charge, width, and groomed momentum fraction. We subsample to event counts ranging from a few hundred to the full sample, and we use many independent pseudo-experiments at each count.

The third is COMPAS with RESOLVE's budget of 1,000 low-fidelity and 15 high-fidelity runs. Here the question is whether an anisotropic order keeps the recovered non-monotonic natal-kick dependence while improving or preserving coverage.

Baselines

The do-nothing reference is the published fixed order. For RESOLVE this is order 4, whose coverage of 74%, 97%, and 100% at 1, 2, and 3 standard deviations is the strongest reported number on this question (RESOLVE). For Moment Unfolding it is the fixed $n$ used in the paper. Precision at the level of individual moments is not available: no reported result found on Lacuna for moment-level precision or coverage of Moment Unfolding on jet substructure; the file reports only the roughly 100-fold speedup over OmniFold. We will therefore rerun the fixed-order configuration ourselves on identical pseudo-experiments.

The strongest fair competitor is K-fold cross-validated order selection, run in two forms. One searches a single global order. The other is greedy forward selection per direction scored by held-out loss. Both get the same pseudo-experiments, the same bootstrap budget, and the same cap on total model fits as our rule.

The second competitor is an L1 sparsity penalty. For RESOLVE it is applied to the full expansion coefficients. For Moment Unfolding it is applied to the exponents $\beta_a$ up to a generous maximum order. The penalty strength is chosen by cross-validation under the same fit budget.

The upper reference is an oracle fixed order. It is the best single order at each $N$, chosen using truth, and it is available only in simulation. Matching it without access to truth is the target.

Ablations isolate the two key ingredients. One uses a constant threshold, with no trial correction. The other drops hysteresis. A third ablation runs a global rather than per-direction rule.

Our method beats the field if three conditions hold. Its 1-standard-deviation coverage must lie within a few points of 68% across all $N$ while the fixed order drifts away from 68%. Its root-mean-square error on truth moments must be within noise of the oracle order. And it must use fewer fits than cross-validation for equal or better coverage. Against RESOLVE specifically, we need coverage at least as close to nominal as the reported 74% and 97% figures, with the natal-kick trend still recovered.

Why not just…?

Why not just grid-search one good fixed order offline? On full-statistics simulation this is cheap. But the core empirical fact in Moment Unfolding is that the best order depends on $N$, so an order tuned at one count is wrong in the sparse tails. The oracle-order baseline measures exactly how much a single order loses across counts and bins. If the oracle order is nearly flat in $N$, the adaptive rule is unnecessary, and the experiments will say so.

Why not just cross-validate the order? This is the real threat. Cross-validation is standard for choosing the degree of a polynomial chaos expansion. Our argument against it is specific, and it is an argument to be tested, not assumed. With a few hundred events or 15 high-fidelity runs, each fold removes a meaningful fraction of the data. That makes the held-out score itself Poisson-noisy and biases selection toward small orders. Searching per-direction orders multiplies the number of fits. If cross-validation matches our coverage and error at equal fit budget, we will report the contribution as a calibration study of cross-validation in counting regimes rather than as a new selector.

Why not just add a sparsity penalty and let it zero out terms? An L1 penalty does shrink unneeded terms. However, its strength is one more tuned meta-parameter, of exactly the kind the ACM authors note becomes unstable under noisy signals (ACM). Shrinkage also biases the very moments being measured. Our rule replaces that strength with a stated false-inclusion rate $\alpha$. The comparison of coverage under the penalty baseline will show whether this matters in practice.

Won't adaptation on noisy counts just oscillate, as in the evolutionary-algorithm results? The concern is legitimate, given the poor finite-size showing of the self-adjusting variant in theory-guided benchmarking. The hysteresis ablation and the measured overhead in fits address it directly. We will report the variance of the selected order across bootstrap replicas, and we will treat high order-flip rates as a failure.

Decisive Checks

Three results would settle the proposal.

First, on synthetic truths with known nonzero terms, the empirical rate of admitting a spurious term should be close to $\alpha$ at every $N$. The constant-threshold ablation should exceed it by a margin that grows with the number of candidates tried. That would confirm that the look-elsewhere correction is doing the work.

Second, on jet substructure across subsampled $N$, the selected order should rise with $N$, and the coverage curve should stay flat near nominal where the fixed order's curve bends.

Third, on COMPAS under RESOLVE's 1,000/15 budget, the anisotropic expansion should keep a high degree along kick dispersion. It should not let that dependence collapse, and it should match or improve the reported coverage of 74%, 97%, and 100% at 1, 2, and 3 standard deviations.

If cross-validation passes all three at equal cost, the selector is not needed, and that is a result worth publishing for these communities.