Collider searches for new physics often fit data with an effective field theory (EFT). An EFT is a low-energy approximation that summarizes the effects of unknown heavy particles as a short list of correction terms, each with a coefficient to be measured. These corrections grow with collision energy, so the most energetic events carry the most information about the coefficients. Those same events are where the approximation breaks down. Analysts handle this conflict by hand: they pick one energy cutoff, discard everything above it, and fit the rest. This proposal treats that choice as a two-objective optimization. One objective is how tightly the coefficients are measured. The other is how much the result can be trusted. The selection can cut on several observables at once, because different observables leave the EFT's range of validity at very different energies.
The opportunity is concrete. Stress tests against full heavy-particle models show that a linear EFT prediction for associated Higgs production turns negative near $m_{Vh}\approx 600$ GeV. Rate errors in high-energy tails reach 30–70% (Extending the limits of Higgs effective theory; Pushing Higgs effective theory to its limits). The same work finds that the leading-jet transverse momentum $p_{T,j1}$ stays well described across a broad range. A single cutoff on one energy variable cannot exploit that difference. A selection shaped jointly over several observables can, and two-objective search is the natural tool for finding it.
Core Claim
We claim that a validity measure computable without knowing the underlying heavy-particle model can serve as a second objective. That measure is the shift in fitted coefficients when the squared dimension-6 term is added to the linear prediction, expressed in units of the statistical error. Searching over multi-observable selection regions against this measure and against expected precision traces a trade-off curve, or Pareto front: the set of selections where neither objective can improve without the other getting worse.
We further claim that this front is close to the curve one would obtain with oracle knowledge of the true model's bias. We also claim it strictly dominates the curve obtained by sweeping a single energy cutoff. Both claims are falsifiable on simulated pseudo-data generated from full models whose answers are known.
The contribution is the formulation and its validation, not a new optimizer. Prior work motivates the problem but never closes it. The validity studies (Pushing Higgs effective theory to its limits; Extending the limits of Higgs effective theory) diagnose where truncation fails but leave the cutoff as a manual choice. Simulation-based inference (Constraining Effective Field Theories with Machine Learning; MadMiner) maximizes information alone, which pulls sensitivity toward the very tails where the EFT fails. The closest optimization work in physics (Bayesian optimization in ab initio nuclear physics) calibrates nuclear EFT parameters against a single combined goodness-of-fit score, with no validity objective at all.
Research Program
The physics setting is weak boson fusion Higgs production with decay to four leptons, plus associated $Vh$ production. These are the benchmarks shared by the validity studies and the MadMiner papers.
The key enabler is reweighting. Using MadMiner's morphing (rewriting predictions at any coefficient value as a fixed combination of a few simulated benchmark points) and its per-event extraction of the joint likelihood ratio and score from MadGraph matrix elements, we generate one large event sample. That sample stores, for every event, its standard-model weight, its linear-interference weight and its squared-dimension-6 weight. Evaluating any candidate selection region then reduces to reweighting a fixed sample, so each evaluation takes seconds.
For a region $R$ we compute two quantities from the restricted sample. The first is the Fisher information $I_R$, the curvature of the expected log-likelihood that sets the best achievable error bars. The precision objective is $$f_1(R)=\big[\det I_R^{-1}\big]^{1/(2k)},$$ the geometric-mean expected error over the $k$ fitted coefficients. The validity objective is $$f_2(R)=\max_j \frac{|\hat c_j^{\,\mathrm{lin+quad}}(R)-\hat c_j^{\,\mathrm{lin}}(R)|}{\sigma_j(R)},$$ where $\hat c_j$ are coefficients fitted to Asimov pseudo-data (the idealized dataset equal to its expectation) at a reference coefficient point, and $\sigma_j$ is the statistical error. This uses the common heuristic that the gap between linear and squared dimension-6 truncations estimates the size of the omitted dimension-8 effects. Whether that heuristic actually holds is what the experiments test, rather than something we assume.
Regions are defined by four to six thresholds. We use upper cuts on $m_{Vh}$ or on the dijet-system energy, upper cuts on $p_{T,j1}$, a window in $\Delta\phi_{jj}$, and optionally an upper cut on the learned SALLY score, so the selection can act on the summary statistic an analyst would actually use.
The choice of search method follows from the multi-objective literature, which mostly constrains our design rather than supplying it. Linear weighting of objectives cannot reach concave parts of a front (A practical guide to multi-objective reinforcement learning and planning), so we do not rely on it. Weighted Chebyshev scalarization, which minimizes the largest weighted distance to an ideal point, reaches concave regions (Multi-Objective Constrained Optimization for Energy Applications via Tree Ensembles). NSGA-II's known degradation appears above two objectives (A Crowding Distance That Provably Solves the Difficulties of the NSGA-II in Many-Objective Optimization), so it is adequate here. Because evaluations are cheap, we run NSGA-II and a Chebyshev sweep, and we check both against exhaustive grids in two and three dimensions. Expensive-evaluation methods such as MORBO (Multi-Objective Bayesian Optimization over High-Dimensional Search Spaces) are only needed if we extend to selections whose evaluation requires re-simulation, and we treat that as an extension.
Ground truth comes from full heavy-particle models taken from the stress-test papers: the heavy vector triplet, the scalar singlet and the two-Higgs-doublet model, each matched to the EFT with $v$-improved matching. That matching treats the Higgs vacuum value correctly when relating the heavy model to the EFT, and it reproduces weak boson fusion to 1–10% according to Pushing Higgs effective theory to its limits. We simulate pseudo-data from each full model across a range of heavy masses. For every selection on a front we fit the EFT, then record the true bias of the fitted coefficients against the matched values, the interval width, and the frequentist coverage.
Baselines
The naive reference point is to use all events with a linear dimension-6 fit and no cut. It gives the tightest intervals and the worst bias, and it marks one end of any front.
The strongest existing practice is the single-cutoff sweep. We scan a hard upper cut on $m_{Vh}$ (and separately on $p_{T,j1}$), using the same event sample, the same fits and the same pseudo-data, and trace the resulting curve of precision against true bias. This is what analysts do and what the validity papers implicitly recommend. Because it is a one-parameter slice of our selection space, the comparison is exactly fair. No reported result found on Lacuna for a precision-versus-bias curve from a cutoff sweep. The nearest reported figures concern information loss, not validity. Histogram analyses need 90% more data than RASCAL, equivalent to a 16% smaller reach in the new-physics scale $\Lambda$ (Constraining Effective Field Theories with Machine Learning), and about three times more data in $h\to\tau\tau$ (New Ideas for Effective Higgs Measurements).
We beat the sweep if, at matched true bias (for instance below $0.5\sigma$ on every coefficient across all three full models), our selection gives intervals at least 15% narrower. Equivalently, the hypervolume (the area of objective space a front dominates) of our front, computed with the true-bias objective, must exceed the sweep's beyond seed-to-seed variation.
A second strong competitor uses all events with a linear-plus-squared fit and no cut. This non-linear analysis restores positivity and extends agreement to about $m_\xi/2$ (Extending the limits of Higgs effective theory), where $m_\xi$ is the mass of the heavy mediator. It is a single point that we place on the same plot.
The decisive ablation runs the identical optimization twice: once with the proxy $f_2$, and once with an oracle $f_2$ equal to the true bias under one full model. The gap between the proxy front and the oracle front measures how much the UV-agnostic proxy costs. A second ablation restricts the selection to a single observable at a time, which isolates whether the multi-observable shape is what helps.
Why not just…?
Why not just sweep one energy cutoff and read off the curve? This is the real threat. If the bias is controlled almost entirely by one energy variable, a multi-dimensional search adds nothing. The motivating evidence points the other way: $m_{Vh}$ tails fail early while $p_{T,j1}$ stays robust (Extending the limits of Higgs effective theory), so validity is not a function of one variable. Still, that is suggestive rather than decisive. The single-observable ablation settles it directly. If the multi-observable front does not dominate the best single-cutoff curve on true bias, we report that cutoff sweeps are sufficient, and that is a useful negative result.
Why not just include the squared dimension-6 terms everywhere and skip the cut? Squared terms fix the negative-rate pathology, but the same paper shows agreement only up to roughly half the mediator mass. Beyond that, dimension-8 terms and on-shell production dominate, and the squared terms cannot model them. Our experiments include this fit as a baseline point. If its true bias is already below threshold across all three full models at the masses tested, the cut problem is moot in that regime, and we will say so.
Why not just add a validity penalty to the chi-squared and minimize one number? With two objectives and cheap evaluations, a weight sweep is easy to run. The difficulty is that it misses concave stretches of the front (A practical guide to multi-objective reinforcement learning and planning). We run the linear sweep alongside the Chebyshev and NSGA-II fronts. If the front turns out to be convex, the methods agree and the simpler one wins. Nothing in the claim depends on an elaborate optimizer.
Proposed Experiments
The first check tests whether the proxy works. For each full model and heavy mass, we compute the Spearman rank correlation between $f_2$ and the true bias across all selections on the front. We require correlation above 0.7 in at least two of the three models. We also require that the proxy front sit within 10% of the oracle front in hypervolume. If the proxy fails, the method fails, and we would publish where and why the linear-versus-squared gap stops tracking the omitted dimension-8 effects.
The second check is dominance over the cutoff sweep, at matched true bias, on both weak boson fusion and $Vh$, as defined in the baselines.
The third check is transfer to unseen models. A selection chosen using only the proxy must keep coverage within 5 points of nominal on the held-out full models, including the vector triplet, whose angular correlations scalar-mediator simplified models fail to reproduce (Extending the limits of Higgs effective theory).
The fourth check repeats the comparison with the SALLY score as one of the cut variables. This tests whether learned high-information statistics can be made validity-aware rather than simply pulled into the tails.
References
- Extending the limits of Higgs effective theory
- Pushing Higgs effective theory to its limits
- Constraining Effective Field Theories with Machine Learning
- MadMiner: Machine Learning-Based Inference for Particle Physics
- Bayesian optimization in ab initio nuclear physics
- A practical guide to multi-objective reinforcement learning and planning
- Multi-Objective Constrained Optimization for Energy Applications via Tree Ensembles
- A Crowding Distance That Provably Solves the Difficulties of the NSGA-II in Many-Objective Optimization
- Multi-Objective Bayesian Optimization over High-Dimensional Search Spaces
- New Ideas for Effective Higgs Measurements