Measuring how much the strange quark contributes to the proton's mass comes down to a large linear-algebra problem. One must estimate the trace of the inverse of the lattice Dirac operator, a sparse complex matrix with tens of millions of rows that describes how quarks propagate on a discretized spacetime grid. Practitioners do this with random probe vectors. A common way to reduce the variance is "deflation": compute the handful of eigenvectors with the smallest eigenvalues, treat that part exactly, and estimate only the remainder stochastically. Deflation is usually done conservatively. Eigenvectors are converged to high precision because a small gap between neighbouring low eigenvalues seems to make the computed subspace unstable, and textbook perturbation bounds such as Davis–Kahan blow up as that gap shrinks. A recent line of random-matrix work tells a different story. It models noise as a continuous random walk on matrices (Dyson Brownian motion, in which eigenvalues behave like mutually repelling particles), and it shows that a rank-$k$ subspace is stable whenever the gap at the rank-$k$ boundary is large. Clustering among the top $k$ eigenvalues does not matter. Those results were built for privacy-preserving PCA. There they carry an awkward requirement: the noise must be complex Hermitian, the $\beta=2$ ensemble.
This proposal rests on one observation. The Hermitian Dirac operator of lattice QCD is already complex Hermitian, and the errors a mixed-precision eigensolver makes on it are complex. The setting where the Dyson-Brownian-motion guarantees are strongest, and most artificial in privacy, is the native setting of the strange-quark loop. The claim is that deflation subspaces for disconnected strange loops can be computed cheaply and loosely, and that the deflation rank should be placed at a large spectral gap rather than fixed in advance. Bias can be kept at zero by computing the deflated part through the subspace projection rather than through approximate eigenvalues. The leftover variance should follow a boundary-gap formula, not the smallest internal gap. A further question is whether any of this beats the current best variance-reduction recipe, which uses no deflation at all, and the proposal tests that directly.
Core Claim
Write $Q = \gamma_5 D$ for the Hermitian Dirac operator at the strange-quark mass and $A = Q^{-1}$ (or $\Gamma$-inserted variants). Let $\tilde P$ be the projector onto an approximate low-mode subspace of rank $k$. The loop is split as
$$\mathrm{Tr}[A] = \sum_{i=1}^{k} \tilde v_i^\dagger A \tilde v_i + \mathrm{Tr}[A(1-\tilde P)],$$
where the $\tilde v_i$ are the computed eigenvectors. If the first term is evaluated with actual solves $A\tilde v_i$ rather than with approximate inverse eigenvalues $1/\tilde\lambda_i$, the split is exact for any orthonormal $\tilde v_i$. Eigensolver error then never creates bias. It only changes how much variance the stochastic second term retains. That retained variance is a Frobenius-norm quantity: for complex unit-modulus noise it is the squared off-diagonal Frobenius norm of $A(1-\tilde P)$. This is precisely the norm in which diffusion-based analyses are tight.
Suppose the eigensolver error behaves like a small complex Hermitian perturbation of scale $\sigma$. Then the analysis in Private Low-Rank Approximation for Covariance Matrices, Dyson Brownian Motion, and Eigenvalue-Gap Bounds for Gaussian Perturbations predicts that the subspace error is governed by a sum of inverse-squared gaps. In our setting, only gaps that cross the cutoff should enter:
$$\|\tilde P - P\|_F^2 \;\approx\; \sigma^2 \sum_{i \le k < j} \frac{1}{(\lambda_j - \lambda_i)^2},$$
with $\lambda_i$ the exact eigenvalues and $P$ the exact projector. The prediction we test is that leaked variance tracks this boundary sum. Internal near-degeneracies among the deflated modes should be irrelevant. That has two practical consequences. The rank $k$ should be chosen where $\lambda_{k+1}-\lambda_k$ is large, and eigensolver tolerance can be relaxed until the boundary sum, not the smallest gap, becomes the bottleneck.
Where the Ingredients Come From
The method comes from the Dyson-Brownian-motion line of work. Re-Analyze Gauss treats Gaussian noise addition as a matrix diffusion and integrates the eigenvector stochastic differential equations. It gains a $\sqrt{k}$ factor in Frobenius error over static bounds, but it needs gaps between every adjacent pair of top eigenvalues. Private Covariance Approximation and Eigenvalue-Gap Bounds for Complex Gaussian Perturbations removes the internal-gap requirement, but only for complex ($\beta=2$) noise, which in privacy means artificially complexifying real data. Singular Subspace Perturbation Bounds via Rectangular Random Matrix Diffusions extends the technique to rectangular matrices. All of these assume an i.i.d. random matrix added to a deterministic one, and none touches a structured sparse operator. The research file flags that application as unexplored. We reuse their technique and their gap-sum quantity. Their $\beta=2$ restriction becomes an advantage here.
The physics motivation comes from the strangeness literature. Indirect Feynman–Hellmann determinations take the slope of the nucleon mass with respect to the strange mass. They remain loose: Sigma term and strangeness content of the nucleon reports $y_N \simeq 0.1(2)(1)$, consistent with zero, and Sigma term and strangeness content of octet baryons reports $y_N = 0.20(7)(+13)$. Part of the reason is the narrow strange-mass lever arm. Direct disconnected-loop evaluations avoid that lever-arm problem but are noise-limited. Evaluation of disconnected quark loops for hadron structure using GPUs ranks the one-end trick combined with the truncated solver method (many cheap low-precision solves corrected by a few high-precision ones, hereafter TSM) as the most suitable method for $\sigma$-terms and axial charges. The proposal tries to improve the direct route specifically.
One paper constrains the design. Eigenvalue Attraction shows that eigenvalues of real non-Hermitian matrices can attract rather than repel. The repulsion picture therefore cannot be applied to the non-Hermitian Wilson operator $D$ itself. We work only with the Hermitian $Q$, or with $D^\dagger D$ for twisted-mass fermions. In both cases the spectrum is real and repulsion is the relevant dynamics. The repulsion strength is an empirical question that we check rather than assume.
Research Program
The first stage is diagnostic and inexpensive. On the B55.32 ensemble ($N_f=2+1+1$ twisted mass, $32^3\times 64$, $a\approx 0.082$ fm, $M_\pi\approx 370$ MeV) used in the disconnected-loop paper, we compute the lowest few hundred eigenpairs of $D_W^\dagger D_W$ to high precision on a subset of configurations. For twisted-mass valence quarks, $D^\dagger D = D_W^\dagger D_W + \mu^2$, so the same eigenvectors serve light and strange quarks. Only the spectral shift differs, and this shift makes the strange operator better conditioned. Using these eigenpairs, we measure the nearest-neighbour spacing distribution of the low modes and test whether it shows quadratic small-gap suppression. That suppression is what the $\beta=2$ analysis relies on.
The second stage tests the error model. We rerun the eigensolver in single and half precision and with loosened tolerances. For each run we measure the actual projector error $\|\tilde P - P\|_F$ and check two things. The first is whether the error matrix, expressed in the exact eigenbasis, looks like isotropic complex noise; the alternative is that it concentrates near particular modes. The second is whether the error scales with the boundary gap sum above or with the minimum internal gap. This stage is the scientific core. It is where the diffusion picture either earns its place or fails.
The third stage is theory, aimed at a bound proved for the actual estimator rather than for low-rank approximation. We adapt the Itô-calculus argument from the covariance papers to bound the leaked variance of the deflated estimator in terms of the boundary gap sum and $\sigma$. The perturbation is complex Hermitian, so the $\beta=2$ gap lower bounds apply without modification. The open spectral-norm question from that literature is not needed, because the target quantity is Frobenius.
The fourth stage turns the analysis into a procedure. Rank $k$ is chosen at the largest relative gap within a compute budget. Eigensolver tolerance is set as loose as the boundary-sum prediction allows. The deflated term is evaluated with solves, keeping it unbiased. This deflation is combined with the one-end trick and TSM, and the result is used to compute the strange scalar loop (for $\sigma_s$) and the strange axial loop (for $g_A^s$).
Baselines
The naive reference point is plain stochastic trace estimation with random probe vectors and high-precision solves, with no deflation and no variance reduction. It fixes the cost-per-error scale. The GPU disconnected-loop paper reports that naive estimation needs an impractical number of inversions. We run it on the same configurations and the same number of GPU-hours as everything else.
The strongest existing competitor is one-end trick plus TSM without deflation, as implemented in modified QUDA in Evaluation of disconnected quark loops for hadron structure using GPUs. That paper reports cost reductions of a factor of 2–3 for light quarks and a two-fold error reduction for the strange contribution to the nucleon axial charge. It also ranks this combination as best for $\sigma$-terms. No reported absolute error on the strange scalar loop on B55.32 was found on Lacuna. We therefore reproduce this baseline ourselves, with its number of low-precision and high-precision solves tuned on a held-out configuration subset, using the same tuning budget as our method. Our method beats it only if it reduces the variance of the strange scalar loop, or of the extracted $\sigma_s$, at equal total GPU-hours, with eigensolver cost included. A reduction of at least 1.5 times in error at equal cost is the threshold we would consider meaningful.
The third baseline is conventional exact-precision deflation at a fixed rank chosen by budget rather than by gap, combined with one-end plus TSM. No reported result was found on Lacuna for deflated strange loops on this ensemble, so we run it ourselves at matched cost.
Two ablations isolate the key ingredients. The first keeps the rank $k$ chosen by the gap rule but restores tight eigensolver tolerance. This separates the savings from loose eigenvectors from the savings from rank selection. The second keeps loose eigenvectors but places $k$ deliberately inside a cluster of eigenvalues. If the boundary-gap theory is right, this placement should sharply increase leaked variance even at the same rank and tolerance.
For the bounds themselves, we compare three predictors of measured leaked variance on identical spectra: Davis–Kahan, the contour-bootstrapping spectral-norm bound of Spectral Perturbation Bounds for Low-Rank Approximation with Applications to Privacy, and the diffusion-based boundary sum. On census and KDD-Cup data, that paper's bound outperformed the classical Eckart–Young–Mirsky baseline by orders of magnitude; the file gives no exact values.
Why not just…?
Why not just use the static contour-bootstrapping bound instead of a diffusion argument? This is the closest competitor and a real one. The contour bound of Spectral Perturbation Bounds for Low-Rank Approximation with Applications to Privacy also depends on the gap at the cutoff, scaling as $\|E\|\lambda_p/\delta_p$. It may support the same rank-selection rule. The difference lies in the norm. Leaked variance is a Frobenius quantity, and a spectral-norm bound converted to Frobenius norm picks up a factor of roughly $\sqrt{k}$ in the projector error. That conversion is the same looseness the diffusion analysis removed in Re-Analyze Gauss. Stage two settles this empirically. If the contour bound tracks measured leaked variance as closely as the boundary sum across a range of $k$, the diffusion machinery adds nothing practical, and we would report that.
Why not just keep one-end plus TSM and skip deflation? This may well be the right answer for strange quarks. The strange mass shifts $D^\dagger D$ by $\mu_s^2$, which compresses the low-mode advantage that makes deflation valuable for light quarks. The disconnected-loop paper already found one-end plus TSM best without deflation. The head-to-head at equal GPU-hours is the decisive experiment. If deflation does not pay for itself at the strange mass, the gap-rule procedure still transfers to the light-quark loops, where the conditioning argument is stronger. The light loops on the same eigenvectors are a secondary target.
Why not just get $\sigma_s$ from Feynman–Hellmann mass derivatives and avoid disconnected loops? The indirect results in the file remain at or near consistency with zero: $\sigma_{\bar ssN}=49(37)(26)$ MeV in Decay constants and sigma terms from the lattice., and $y_N\simeq 0.1(2)(1)$. Part of the reason is the narrow strange-mass lever arm. A direct determination is complementary, and the two can be cross-checked. Our ensemble has an unphysical pion mass, so we claim cost-per-error gains for the direct method, not a better physical $y_N$.
Why not just converge the eigenvectors tightly and stop worrying? Tight convergence removes the error model but costs eigensolver time. That cost competes directly with probe vectors in the same budget. The first ablation measures how much of the gain comes from loosening tolerance. If that share is negligible, the claim reduces to the rank-selection rule alone.
Decisive Checks
The proposal stands or falls on three measurements on B55.32. First, the low-mode spacing distribution of $D_W^\dagger D_W$ must show the small-gap suppression assumed by the $\beta=2$ theory. If spacings are instead Poisson-like, the repulsion premise fails. Second, the projector error from low-precision eigensolves must track the boundary inverse-gap sum better than the minimum-gap and contour predictors. The cluster-placed rank ablation must show the predicted variance jump. Third, at equal GPU-hours including eigensolves, gap-rule loose deflation combined with one-end plus TSM must reduce the error on the strange scalar and axial loops relative to the reproduced one-end plus TSM baseline. If the first two checks hold and the third fails, the result is a validated perturbation theory for lattice deflation that pays off for light quarks but not strange ones. That outcome is worth reporting, and it would redirect the method.
References
- Private Low-Rank Approximation for Covariance Matrices, Dyson Brownian Motion, and Eigenvalue-Gap Bounds for Gaussian Perturbations
- Re-Analyze Gauss: Bounds for Private Matrix Approximation via Dyson Brownian Motion
- Private Covariance Approximation and Eigenvalue-Gap Bounds for Complex Gaussian Perturbations
- Singular Subspace Perturbation Bounds via Rectangular Random Matrix Diffusions
- Sigma term and strangeness content of the nucleon
- Sigma term and strangeness content of octet baryons
- Evaluation of disconnected quark loops for hadron structure using GPUs
- Eigenvalue Attraction
- Spectral Perturbation Bounds for Low-Rank Approximation with Applications to Privacy
- Decay constants and sigma terms from the lattice.