Robust Low-Rank Spectroscopy for Noisy Lattice QCD Correlators

Lattice QCD computes hadron properties by simulating the strong force on a grid of spacetime points. The basic output is a correlator: a measurement of how a state created at one time overlaps with a state destroyed later. Each hadron state contributes a decaying exponential to this signal, and the decay rate is the hadron's energy. The trouble is that the noise grows exponentially relative to the signal as time increases, and much faster for systems with many quarks, such as nucleons and pairs of nucleons. The research file names this as a central obstacle to multi-baryon calculations. Excited states are usually read off from the late-time correlator, where only the lightest states survive, and that is exactly where the signal is buried.

This note proposes treating the whole stack of correlator matrices, across all time separations, as one structured low-rank factorization. The model is a shared set of operator overlaps times a diagonal of decaying exponentials. We would fit it with alternating minimization, which repeatedly solves for one block of unknowns while holding the others fixed. The fit would use a robust loss that resists the heavy-tailed noise typical of these correlators. In two sentences: the early, cleaner time slices already constrain the overlaps, and once the overlaps are pinned down, the energies can be extracted from far less late-time data. Alternating minimization makes the robust version of this fit cheap and well-behaved when it starts from the standard spectral estimate.

Core Claim

For an $N\times N$ matrix of correlators $C_{ij}(t)$ built from $N$ interpolating operators, the spectral decomposition is

$$ C_{ij}(t) \approx \sum_{n=1}^{r} Z_{in} Z_{jn}\, e^{-E_n t}, $$

where $Z_{in}$ is how strongly operator $i$ creates state $n$, $E_n$ is that state's energy, and $r$ is the number of states retained. We fit $Z$ and $E$ jointly over all time slices by minimizing

$$ \sum_{t}\sum_{i\le j} \rho_\delta\!\left(\frac{\bar C_{ij}(t) - \sum_n Z_{in}Z_{jn}e^{-E_n t}}{\sigma_{ij}(t)}\right). $$

Here $\bar C_{ij}(t)$ is the ensemble estimate of the correlator, $\sigma_{ij}(t)$ is its error estimate, and $\rho_\delta$ is a Huber-type loss with threshold $\delta$. A Huber loss is quadratic for small residuals and linear for large ones. As $\delta\to\infty$ the objective becomes ordinary weighted least squares.

The structure splits cleanly. With $E$ fixed, the $Z$-step is a robust regression, solved by iteratively reweighted least squares. With $Z$ fixed, the problem separates into $r$ one-dimensional problems, one per energy. We initialize from the generalized eigenvalue solution, which is the standard spectral method in lattice spectroscopy. This mirrors the result from matrix completion that alternating minimization converges when started near the truth (Asymptotic Convergence Rate of Alternating Minimization for Rank One Matrix Completion proves this for the rank-one case). That paper supplies the method template. Our proposal differs in that the factors are physically constrained: the overlaps are shared across time, the middle factor is a positive decaying diagonal, and the noise is heavy-tailed and grows with time rather than coming from missing entries.

The claim we can actually test is narrow. On problems with a known spectrum, the robust joint fit recovers the first few excited-state energies with smaller error than generalized-eigenvalue and Gaussian multi-exponential fits, using the same correlator data, particularly when late time slices are discarded. We do not claim that alternating minimization as an optimizer is the source of the gain. An ablation below separates the optimizer from the loss.

Research Program

The physics motivation comes from the open problems in the file. Lattice studies of two-nucleon scattering (Two-nucleon higher partial-wave scattering from lattice QCD), nuclear parity violation (Nuclear Parity Violation from Lattice QCD), and the nucleon axial coupling (Nucleon axial coupling from Lattice QCD) all depend on separating ground states from excited-state contamination in noisy baryon correlators. Finite-volume analyses (N. Ishii et al.) need the finite-volume energy levels themselves as inputs, so better level extraction feeds directly into scattering phase shifts. These papers motivate the work. None of them proposes the estimator.

The main risk is bias, and it is a real one. A robust loss estimates something closer to the typical value of the correlator than its mean. Heavy-tailed correlator distributions have different typical values and means, and the physics lives in the mean. The program therefore treats $\delta$ as a dial with bias on one side and variance on the other. For each $\delta$ we measure bias and variance against a known truth, and we test a bootstrap bias correction that subtracts the shift between the robust and least-squares fits estimated on resampled ensembles.

The second component is choosing the number of states $r$. Too few states pushes excited-state contamination into the ground state. Too many makes the fit unstable. We would select $r$ by cross-validation over held-out gauge configurations, borrowing the held-out-entry logic of matrix completion. The selection would be checked against the known number of resolvable states in the test problems.

Baselines

The naive reference point is the effective-mass plateau: the logarithm of the ratio of a single diagonal correlator at neighboring times, averaged over a window where it looks flat. It shows how much any multi-operator method gains at all.

The strongest existing competitor is the generalized eigenvalue method. It solves $C(t)v=\lambda C(t_0)v$ at fixed reference time $t_0$ and reads energies off the decay of each eigenvalue. It is the standard tool, it is already a low-rank method, and it is the thing our method must beat. We would run it with the same operator basis and the same configurations, and scan $t_0$ and the fit window over a grid of the same size as our tuning grid for $\delta$ and $r$.

The second strong competitor is a correlated multi-exponential least-squares fit of exactly our model, solved jointly by Levenberg–Marquardt with the same initialization. It is our method with $\delta=\infty$ and a different optimizer.

Two ablations isolate the ingredients. Setting $\delta=\infty$ inside our alternating minimization measures the effect of the robust loss. Comparing alternating minimization with Levenberg–Marquardt on the identical robust objective measures the effect of the optimizer.

We found no reported result on Lacuna for excited-state energy error of any of these methods on a shared benchmark. All numbers will therefore be produced within this study under equal configuration counts and equal tuning budgets. Beating the strongest baseline means a lower root-mean-square error on the first and second excited-state energies against the known truth, at matched ground-state bias, across at least three noise levels. A gain that appears only at one noise level would not count.

Why not just…?

Why not just use the generalized eigenvalue method with a bigger operator basis? Enlarging the basis is the standard remedy, and it may close much of the gap. The trouble is that it requires the eigenvalues to be fitted at times late enough for the $t_0$ approximation to hold, which is where noise dominates. Our experiments include a large-basis generalized-eigenvalue run as a separate arm. If it matches the robust fit, that is a negative result we will report.

Why not just run a correlated least-squares fit of the same model? This is the most serious threat. If the gain comes only from fitting jointly across times, a standard nonlinear least-squares fit already does that. The $\delta=\infty$ ablation answers this directly. If robustness adds nothing at matched bias, the contribution collapses to existing practice, and we would redirect the work toward the cross-validated choice of the number of states.

Why not just use a robust estimator of the ensemble mean, then any standard fit? Applying a robust average configuration by configuration ignores the shared-overlap structure. It also introduces the same typical-value-versus-mean bias without the structural constraint that partially absorbs it. We include this two-stage pipeline as an extra arm, with the same bias correction applied.

Proposed Experiments

The first testbed is synthetic correlator matrices with a planted spectrum of four to six states and planted overlaps. Noise is drawn from log-normal and Gaussian families, with the signal-to-noise ratio decaying exponentially in time at controlled rates, ranging from meson-like to baryon-like. Here truth is exact, and bias versus $\delta$ can be mapped precisely.

The second testbed is the lattice Schwinger model, quantum electrodynamics in one space and one time dimension. Its spectrum can be computed independently by exact diagonalization of the Hamiltonian on small lattices, and it is the same model family studied in Meson thermalization with a hot medium in the open Schwinger model. Monte Carlo correlators from this model test the method on real gauge-field noise with known answers.

The third testbed is nucleon and two-nucleon correlators from small four-dimensional QCD ensembles, where truth is unknown. There we test consistency rather than accuracy. Energies from truncated time ranges should agree with energies from the full range, and finite-volume levels should yield phase shifts consistent with existing two-nucleon analyses.

The decisive check comes from truncating the maximum time slice used in the fit. If the robust joint fit is genuinely using the early-time information, its error should grow more slowly than the generalized eigenvalue method's error as late times are removed. If both degrade at the same rate, the core claim is false.