Cutoff Shopping in Finite-Energy Sum Rules

Finite-energy sum rules are exact identities from complex analysis that tie the low-energy resonance region of a hadron scattering amplitude to its smooth high-energy behavior. Physicists use them as a lie detector. If a low-energy model of the amplitude, fitted to resonance data, integrates to something inconsistent with what high-energy data imply, the model is suspect. This test has real consequences: it has been used to flag the η-MAID description of η photoproduction as deficient, and it has exposed a sign mismatch in one invariant amplitude. Every such verdict, however, rests on choices the analyst makes after seeing the data: the energy cutoff $N$ at which the integral stops, the power $n$ of energy used to weight the integral, and the momentum-transfer bins inspected. Each choice moves the agreement up or down. When high-energy data are scarce and counting-noise limited, as they are here, those choices can manufacture both consistency and inconsistency.

This note proposes treating the analyst's cutoff choice as a stopping problem. As the cutoff increases, the gap between the two sides of the sum rule evolves like a random walk, and "where to stop the integral" is a stopping time on that walk. That is the same mathematical object that cryptographers bound when they ask how much an adversary can bias a shared coin flip by choosing when to walk away from the protocol. We will reuse their linear-programming duality bounds to build a sum-rule consistency test whose significance stays valid under any cutoff-selection strategy. We will then apply it to the published photoproduction disputes and ask how much additional high-energy data would be needed to settle them.

Core Claim

Let $D(N)$ be the sum-rule discrepancy at cutoff $N$. It equals the low-energy integral $\int_0^N \mathrm{Im}\,A(\nu)\nu^n d\nu$ of the model amplitude $A$ over energy variable $\nu$, minus the high-energy Regge prediction $\beta N^{\alpha+n+1}/(\alpha+n+1)$. Here $\alpha$ and $\beta$ are the intercept and strength of the exchanged family of particles, fitted from high-energy data. Under the null hypothesis that the low-energy model is correct, the increments of $D$ between successive energy bins are approximately mean-zero, because the Regge term is the local average of the resonances (What is duality? supplies this averaging picture). Their size is set by bin-wise model and statistical uncertainty. The quantity an honest test must control is the worst-case false-verdict rate

$$V = \sup_{\tau} \; \sup_{P \in \mathcal{P}} \; \Pr_P\big[\,|D(N_\tau)|/\sigma(N_\tau) \text{ crosses the decision threshold}\,\big],$$

where $\tau$ ranges over every cutoff-selection rule that may depend on the curve seen so far, $\sigma$ is the reported uncertainty, and $\mathcal{P}$ is a class of increment laws defined only by bounded variance and bounded increments. The class is not a single trusted noise model.

Fair Coin Flipping: Tighter Analysis and the Many-Party Case shows how to bound exactly this kind of supremum. It casts an aborting adversary's strategy as a stopping time in an "online binomial game", writes the maximum bias as a linear program, and upper-bounds it with a feasible dual solution. We claim that the same dual construction yields calibrated, distribution-robust p-values for sum-rule consistency. We further claim that, applied to published analyses, these p-values will change the evidential status of at least some claimed agreements or disagreements.

Relation to Prior Work

Three papers motivate the problem. Analyticity constraints for hadron amplitudes: Going high to heal low-energy issues extracts effective Regge strengths from partial wave analyses, which are fits that decompose low-energy data into angular-momentum components. It finds that SAID agrees with high-energy π⁰ data while η-MAID fails in the forward region $-t<0.25$ GeV². Finite-energy sum rules in eta photoproduction off a nucleon reports the $A_4$ sign mismatch and shows that the integral below the physical threshold depends on how multipoles are analytically continued. Structure of pion photoproduction amplitudes traces sum-rule differences to conflicting resonance quantum-number assignments. All three present agreement by inspecting curves across cutoffs and bins. None attaches a significance that accounts for those choices. We take their observables as given and add the missing calibration.

Two papers supply the method. The coin-flipping LP analysis above provides the primal/dual stopping formulation. Its open problems, the $2^t$ degradation and the $\sqrt{\log m}$ factor, matter here only through the analogous question of whether our dual bound is tight. An Optimally Fair Coin Toss achieves bias $O(1/r)$ by hiding a random threshold round, and its explicit attack reaches about $1/(4r)$. That attack serves as a template for constructing near-worst-case cutoff strategies that test tightness from below. When the state of the walk becomes high-dimensional (several invariant amplitudes, several moments $n$, several $t$ bins jointly), we borrow the pure-dual neural approach of DeepMartingale: Duality of the Optimal Stopping Problem with Expressivity. It parameterizes a martingale with a network and minimizes a dual upper bound, with network size polynomial in dimension up to $d=100$. That paper assumes continuous Brownian noise, so we must adapt it to discrete, bounded-increment walks.

The rare-event literature constrains the design. Branching fractions for psi(2S) -> gamma eta' and gamma eta and the Telescope Array hotspot analysis (Indications of intermediate-scale anisotropy…) document how roughly 2σ effects dissolve with more data and how scanning erodes significance. Our cutoff scan is a look-elsewhere effect in a dimension that physicists rarely count. The high-energy side of the sum rule sits in the regime where statistical error dominates systematic error. Measurement of the proton-air cross section with Telescope Array's Middle Drum detector… shows that regime in a hadronic measurement, so the Regge fit uncertainty scales like $1/\sqrt{N_{\text{events}}}$.

Research Program

The first component is a calibration engine. We discretize cutoffs on the energy-bin grid of each partial wave analysis and define the walk $D$ jointly across amplitudes, moments and $t$ bins. We write the linear program over stopping rules with the class $\mathcal{P}$ as constraints, and solve its dual exactly for low-dimensional cases and with a DeepMartingale-style network for the joint case. The output is an upper bound on $V$, which we convert into an adaptivity-robust p-value.

The second component is a tightness check. Following the Moran–Naor–Segev attack construction, we build explicit greedy and dynamic-programming cutoff strategies. Their achieved false-verdict rate gives a lower bound. The gap between this lower bound and the dual upper bound tells us whether the robust test is needlessly conservative.

The third component is a reanalysis of the SAID π⁰ agreement, the η-MAID forward-region failure, and the $A_4$ mismatch. For each we report naive and robust significance and an "evidence budget". The evidence budget is the high-energy exposure needed, using Poisson scaling of the Regge-fit error, for the robust test to separate competing low-energy models at 3σ and 5σ. We use the 9 GeV beam-asymmetry predictions of the η paper as the target observable. Where Regge-cut corrections at larger $-t$ set a floor that more data cannot beat, the budget reports that floor instead of an event count.

Baselines

The naive reference point is current practice. It uses a single cutoff, chosen after inspection, with the nominal χ² or pull of LHS against RHS and no correction for choice. This is how the cited sum-rule papers report agreement, so beating it means showing that its false-verdict rate exceeds nominal on simulations with known truth.

The simple multiplicity correction is a Bonferroni adjustment over the full grid of cutoffs, moments and bins. It is valid under any selection rule but ignores the strong correlation between neighboring cutoffs. We expect it to be overly conservative. The robust test beats it if it reaches the same guaranteed error rate with higher power.

The strongest competitor is a parametric max-statistic bootstrap. We resample the partial-wave multipoles from their reported uncertainties, refit the Regge side on resampled high-energy data, and calibrate the maximum of $|D(N)|/\sigma(N)$ over the grid. Because the analyst sees the whole curve at once, this is exactly calibrated whenever the resampling model is correct. That makes it the real threat. It receives the same data, the same grid and an equal compute budget, measured in simulation draws versus dual-optimization steps. The robust test beats it only if it keeps nominal coverage when the noise model is misspecified, for example through correlated multipole errors, heavier tails, or unreported continuation uncertainty below threshold, while the bootstrap undercovers. It must also lose little power when the model is correct. No reported result found on Lacuna for calibrated coverage or significance of any sum-rule consistency test, so this baseline's numbers will be produced here. For the stopping-bias side, the reference figures from the file are the Moran–Naor–Segev attack bias of about $1/(4r)$ against Cleve's lower bound of $1/(8r+2)$. These anchor the tightness check on the analogous binary game before we move to the sum-rule walk.

The key ablation replaces the class $\mathcal{P}$ in the dual with the single true increment law. This should reproduce the bootstrap. Any remaining difference is then attributable to robustness, not to the solver.

Why not just…?

Why not just use the max-statistic bootstrap? If partial-wave uncertainties were complete and correctly specified, it would suffice, and the misspecification experiments are designed to show whether that holds. The cited papers give reasons to doubt it. Subthreshold continuation introduces model-dependent artifacts that no reported error captures (Finite-energy sum rules in eta photoproduction off a nucleon), and models disagree because of discrete quantum-number assignments rather than smooth noise (Structure of pion photoproduction amplitudes). If the bootstrap with inflated errors covers as well as the dual bound under every misspecification we test, we will report that result. The contribution then shrinks to the reanalysis and the evidence budgets, which use the bootstrap.

Why not just pre-register one cutoff? A single committed $N$ is unbiased. It is wasteful, however, because the useful range of Regge validity is not known in advance, and there is rarely a credible commitment device for reanalyses of public partial-wave fits. We will measure its power against the robust test on synthetic amplitudes. We will also test a hidden randomized cutoff in the spirit of An Optimally Fair Coin Toss as a middle option.

Why not just fold the cutoff into a global fit as a nuisance parameter? Global fits handle uncertainty in Regge parameters but not the selection of which curves and bins the analyst chooses to report. The reanalysis will check whether the η-MAID verdict of Analyticity constraints for hadron amplitudes survives once that selection is counted.

Proposed Experiments

The decisive synthetic benchmark uses amplitudes built to satisfy duality exactly, with resonances on linear trajectories as in A model for pion-pion scattering in large-N QCD. We inject a known defect into one low-energy model and Poisson-sample high-energy data at several exposures. The metrics are the size of the test under adaptive cutoff strategies, its power to detect the injected defect, and the dual-minus-attack gap. The real-data experiments apply all four tests to the π⁰ (SAID) and η (η-MAID) photoproduction comparisons. They report significance before and after correction for each published disagreement, along with evidence budgets.

Three outcomes would falsify the approach. The first is if adaptive cutoff choice barely inflates error on realistic walks, meaning the naive and robust tests agree. The second is if the bootstrap remains calibrated under every misspecification. The third is if the dual bound sits so far above the attack lower bound that the robust test has no power.