Is That Peak Real? Discovery Tests for Spectral Features in Analytic Continuation

Many simulations of strongly interacting quantum systems can only compute correlation functions in "imaginary time", a mathematical rotation that makes the numerics stable. The quantity physicists actually want is the spectral function: how much response the system has at each real frequency, which shows where particles, resonances and decay thresholds sit. Recovering it from imaginary-time data is called analytic continuation, and it is badly ill-posed, because the map from spectrum to data smooths away almost all fine structure. Current methods, from maximum-entropy fits to neural networks, return one plausible spectrum. Nobody then asks the question a collider physicist asks of every bump: at what significance does the data force this peak to exist?

We propose to treat "is there an isolated particle peak, or only a smooth continuum?" as a signal-versus-background hypothesis test over the whole set of spectra consistent with the data. We will build the test and calibrate it against exact ground truth produced by functional renormalization group flow equations that are analytically continued to real frequency. These flows compute spectral functions directly on the real axis, so they supply realistic spectra whose peaks, thresholds and continua are known by construction. The organizing idea comes from quantum complexity theory. There, a quantum advice state matters only through its acceptance statistics on allowed tests, and an untrusted object can be checked by a trusted classical description. We apply that logic to spectra: the reconstruction is the untrusted proposal, and the well-posed forward map is the trusted check.

Core Claim

Post-hoc analytic continuation is naturally a relational problem in the complexity-theory sense: many outputs are valid for a given input. Statements about features, such as "there is a pole below the two-particle threshold", are decision problems. They should be answered by asking whether every valid output contains the feature, not by inspecting one reconstruction.

This distinction comes from A Qubit, a Coin, and an Advice String Walk Into a Relational Problem, which separates relational from decision tasks for quantum advice. It is sharpened by A Full Characterization of Quantum Advice, which shows that quantum advice is equivalent to trusted classical advice plus an untrusted state that the verifier checks. We borrow only the logic, not the theorems. An untrusted spectrum is acceptable if the trusted forward kernel maps it back onto the data within noise, and a feature is "discovered" only if no acceptable spectrum lacks it.

Concretely, let $G$ be the measured imaginary-time correlator with noise covariance $C$, let $K$ be the forward kernel, and let $\rho \ge 0$ be a spectrum. Let $\mathcal{C}_0$ be the convex set of "background-only" spectra, those with no isolated weight in a window $W$ below a threshold. The test statistic is

$$ q = \min_{\rho \in \mathcal{C}_0} \chi^2(\rho) - \min_{\rho \ge 0} \chi^2(\rho), \qquad \chi^2(\rho) = (K\rho - G)^\top C^{-1} (K\rho - G). $$

Both minimizations are convex programs, so $q$ is exactly computable. It is the nonparametric analogue of the profile-likelihood ratio used in collider bump hunts. Its null distribution will be calibrated with pseudo-experiments. Our hypothesis is that $q$ yields calibrated significances where single-reconstruction methods systematically report peaks that the data do not require, and miss peaks that it does.

Research Program

The ground-truth generator reuses the method of Self-consistent spectral functions in the O(N) model from the functional renormalization group. That work solves the continued flows on a complex momentum grid, uses a mass shift to keep a strip of the plane free of poles introduced by the cutoff, and compares three approximation levels. In order of increasing completeness these are the local potential approximation (LPA/LPA'), momentum-dependent propagators (PMOM), and momentum-dependent propagators and vertices (PVMOM). Only PVMOM resolves the threshold at which the σ meson decays into two pions.

We take this sensitivity to the approximation as a feature rather than a nuisance. Different approximation levels produce physically different spectra from one theory, so they give natural pairs of hypotheses: a stable pole versus a resonance above threshold.

Finite temperature spectral functions in the O(N) model adds a second pair. In that work, simpler approximations produce artificially stable particles at finite temperature. This is exactly the kind of spurious "signal" a discovery test should refuse to certify.

Both works describe the continued flow approach of Ralf-Arno Tripolt et al. as bypassing ill-posed inversion. We use the flows only as a source of truth and make no claim that they replace continuation of data from lattice or quantum Monte Carlo simulations, where only imaginary-time data exist.

From each ground-truth spectrum we compute imaginary-time data through the forward kernel, which is well-posed. We then add noise at controlled levels. The noise includes Gaussian noise, correlated noise, and heavy-tailed noise, the last because Noise enhanced neural networks for analytic continuation reports that its method degrades when test noise differs from training noise.

From collider practice we import two safeguards: calibration by pseudo-experiments, and an explicit check that the test does not "sculpt" peaks into a featureless background. The sculpting check mirrors the same-sign control channel used in Isolating Unisolated Upsilons with Anomaly Detection in CMS Open Data. That analysis is our methodological model for discovery claims. We do not use its machine-learning normalizing flows, which are unrelated to renormalization group flows.

Baselines

The naive reference point is to run maximum-entropy continuation and declare a peak whenever the reconstructed spectrum has a local maximum in $W$ whose height exceeds a fixed fraction of the continuum. This is what is often done in practice. It shows how often eye-ball discovery is wrong.

The strongest existing reconstruction method in the file is the noise-trained neural network of Noise enhanced neural networks for analytic continuation. It recovers sharp multi-peak structure and continua that maximum entropy misses on the transverse field Ising model. We found no reported numerical error for it on Lacuna, so we will reproduce it ourselves. We will train it on spectra drawn from the same family as the test set, with the same noise models, and give it the same tuning budget as our method. To make it a fair discovery method, its peak calls will get bootstrap error bars obtained by retraining and resampling the noise.

The strongest competitor on the actual question is a parametric bump hunt: fit "continuum only" versus "continuum plus Lorentzian pole" models to $G$ and compute a likelihood-ratio significance. This is the direct transplant of collider practice. We will run it with the same toys, the same noise covariance, and continuum models of comparable flexibility.

The ablation that isolates our key ingredient replaces the constraint set $\mathcal{C}_0$ with a single best-fit background spectrum. It answers whether optimizing over the full set of consistent spectra, rather than one representative, is what produces calibration.

For physics context, the flow solutions themselves report pion curvature and pole masses agreeing within about 3% in the higher-order PVMOM scheme (Self-consistent spectral functions in the O(N) model from the functional renormalization group). That figure sets the scale of approximation-level differences our test must resolve.

Beating the baselines means two things at matched noise. First, empirical false-discovery rates match nominal levels, for example at most 5% at a nominal 5% test, on continuum-only truths. Second, power to detect true poles is at least that of the calibrated neural network and parametric methods.

Why not just…?

Why not just put bootstrap error bars on a neural-network or maximum-entropy reconstruction? Bootstrap captures how the reconstruction moves under noise. It does not capture the directions in which spectra can change while leaving the data unchanged, and those are what make continuation ill-posed. A regularized method sits at one point in that set however many times it is resampled. The ablation and the bootstrap neural network baseline test this directly. If bootstrap intervals turn out calibrated on the O(N) hypothesis pairs, the convex-set machinery is unnecessary, and we will report that.

Why not just fit a parametric pole-plus-continuum model and run a standard likelihood-ratio test? This is the real threat, and it may win whenever the continuum shape is known. Its weakness is that significance depends on the assumed continuum. The flow solutions provide continua with threshold and Landau-cut structure that a smooth parametric background may mismodel, producing spurious significance. We will measure both methods' calibration under correctly specified and misspecified continua. If the parametric test stays calibrated under realistic misspecification, the right conclusion is to recommend it, with our method as a robustness check.

Why not just skip continuation and compute spectra with continued flow equations directly? For theories where a trustworthy flow approximation exists, one should. But the approximation-dependent artifacts documented in Finite temperature spectral functions in the O(N) model show that flows answer "what does this approximation predict". The cost of carrying full vertex momentum dependence to gauge theories, noted in the file, means many systems will keep producing only imaginary-time data. Our test answers "what does the data require", which is the complementary question.

Proposed Experiments

The first experiment uses the vacuum σ meson from the O(N) flows. The PVMOM spectrum, with its two-pion threshold, is the truth. The hypotheses are "isolated pole below threshold" versus "broad weight above threshold". Using the PMOM and LPA' spectra as alternative truths, we will trace significance against noise level for all methods.

The second experiment repeats this at finite temperature. The artificially stable particle from the simpler approximation serves as a decoy, and the fuller approximation serves as truth. A method that certifies the decoy fails.

The third is a null test. Spectra with continuum only, including sharp threshold edges, must not yield certified peaks above nominal rates. This is the analogue of the same-sign control channel.

The fourth transfers the test to the transverse field Ising model setting of Noise enhanced neural networks for analytic continuation. The question is whether its reported sharp edges survive as certified features at realistic Monte Carlo noise.

The decisive outcome is a curve of nominal versus empirical false-discovery rate across noise types. If our statistic is calibrated where the bootstrap neural network and the parametric fit under misspecified continua are not, the claim holds. If the parametric fit is equally calibrated, the contribution shrinks to a diagnostic, and we will say so.