Does the Singular-Value Damping Rule for Message Passing Hold on Real MRI Scanners?

Message-passing algorithms reconstruct a hidden signal from indirect measurements by repeatedly exchanging estimates between the measurements and the unknowns. Approximate message passing (AMP) and its generalized form (GAMP) are the standard examples. On random, well-behaved measurement matrices they are fast and accurate. On the structured matrices that real instruments produce, they often oscillate or blow up. The usual fix is damping: each new estimate is blended with the previous one instead of replacing it. In practice the amount of damping is picked by trial and error.

A recent theory, On the convergence of approximate message passing with arbitrary matrices, proposes a principled replacement. It says the required damping is set by one number, $\kappa(\mathbf{A})$, the largest squared singular value of the measurement matrix divided by the average squared singular value. For Gaussian-prior GAMP it proves convergence exactly when a damping-dependent stability function exceeds $\kappa(\mathbf{A})$. That theorem has only been checked on synthetic random matrices. This note proposes the missing test on real physical operators: the undersampled single-coil and multi-coil Fourier operators of the fastMRI benchmark. It also asks whether a damping level fixed in advance from $\kappa$, computed with no tuning, can match hand-tuned and adaptive damping on real reconstructions.

Core Claim

We claim that $\kappa(\mathbf{A})$ is cheap to compute for MRI operators, which undercuts the file's objection that the rule needs a prohibitive singular value decomposition. We further claim that $\kappa$ can be turned into an a priori damping setting and then tested for whether it predicts the observed divergence boundary.

The MRI forward operator is the product of three pieces: a sampling mask $M$ that keeps only some frequencies, a Fourier transform $F$, and coil sensitivity maps $S$, giving $\mathbf{A} = MFS$. The average squared singular value is $\|\mathbf{A}\|_F^2/n$. This follows in closed form from the mask fraction and the sensitivity maps. The peak value takes roughly twenty FFT-based power iterations.

The case gives a sharp, falsifiable prediction. For single-coil Cartesian sampling, $MF$ has orthonormal rows, so $\kappa = 1$ and the theory predicts that Gaussian GAMP needs no damping. Multi-coil operators have $\kappa > 1$, and the theory predicts a damping threshold that shifts with coil geometry and acceleration. Either outcome is informative. If the rule holds on real operators, it replaces tuning with a principled setting. If it fails, the failure localizes what the theory misses. The key evidence here is that MRI Image Recovery using Damped Denoising Vector AMP reports that plain vector AMP (VAMP) diverges under Cartesian masks even though $\kappa = 1$. That points to the prior or denoiser, not the operator spectrum, as the source of that failure.

Research Program

The work proceeds in three stages. Each stage relaxes one assumption of the theorem.

Stage one: the proven regime on real operators. We keep the theorem's own setting, a Gaussian prior with quadratic cost, and replace only the matrix with real fastMRI operators. The operators are single-coil and multi-coil, knee and brain, at 4x and 8x acceleration, under Cartesian and point masks, as defined in fastMRI: An Open Dataset and Benchmarks for Accelerated MRI. For each slice we find the empirical critical damping by bisection over $(\theta_s, \theta_x)$, the damping weights on the measurement-side and signal-side updates. We then compare it with the threshold predicted from that slice's $\kappa$. Synthetic matrices with matched $\kappa$ serve as a control, replicating the original paper's result so that any gap can be attributed to real-operator structure.

Stage two: priors outside the theorem. We move to priors the theorem does not cover. The first is an L1 penalty on wavelet coefficients, a sparsity prior with non-quadratic cost for which the theory offers only the local condition $\theta_s\theta_x\|\tilde{\mathbf{A}}\|_2^2 < 1$ on the row- and column-normalized matrix. The second is a learned DnCNN denoiser. For each, we measure how far the true boundary departs from the quadratic prediction. We then test whether a single safety margin, fit once on the 30-image tuning split, corrects the gap across operators.

Stage three: the practical question. We package $\kappa$-based damping as a tuning-free rule. We compare it against static, adaptive, amplitude-damped and non-message-passing solvers on image quality, failure rate, iterations and GPU wall-clock time. The rule's cost is about twenty extra FFTs per operator.

Baselines

Undamped algorithms. The naive reference point is plain GAMP and plain VAMP with no damping. They define the failure mode: the MAD-GAMP paper reports that undamped GAMP sits at 0 dB NMSE on hard synthetic matrices, and the DD-VAMP paper reports that undamped VAMP diverges under Cartesian masks.

Grid-tuned static damping. This is the practitioner's default. We select one $(\theta_s,\theta_x)$ pair by grid search on the 30-image tuning split, using the same split and the same number of trial runs the spectral rule's safety margin gets. The spectral rule "beats" it if it matches NMSE within 0.1 dB while never diverging on test slices where the tuned value does. It also counts as a win if it transfers to unseen acceleration or coil settings without retuning.

MAD-GAMP. Adaptive damping and mean removal for the generalized approximate message passing algorithm raises damping whenever the objective rises and relaxes it when the objective falls. Its authors report it to be 2 to 5 times faster than sequential swept GAMP in wall-clock time, though only on synthetic matrices. It is the strongest competitor that needs no spectral information, so it directly tests whether knowing $\kappa$ adds anything. We run it with its published adjustment rule on identical operators and hardware. Beating it means fewer iterations to reach 0.1 dB of final NMSE at equal failure rate. The cost-evaluation overhead is counted in wall-clock time.

DD-VAMP++. From MRI Image Recovery using Damped Denoising Vector AMP, this method damps the square roots of variance terms and starts with a Peaceman-Rachford ADMM (ADMM-PR) warm-up. It is the strongest reported damped message-passing result on real MRI. On 19 test knee slices of 128×128 pixels under a 4x Cartesian mask it reports −25.475 dB NMSE and 0.939 SSIM, against −24.694 dB NMSE for the best-tuned plug-and-play ADMM (PnP-ADMM). We reproduce that exact split first. Beating it means matching −25.475 dB with the same DnCNN denoiser and no warm-up phase, or keeping the gain at full resolution and on multi-coil data. For damped AMP on multi-coil fastMRI there is no reported result found on Lacuna, so there we compare only against our own reproductions.

Non-message-passing solvers. PnP-ADMM and primal-dual hybrid gradient (PDHG) are included with the same denoiser and a tuning budget matched to the grid search.

Ablation. The key ablation replaces the measured $\kappa$ with $\kappa = 1$ (the "assume i.i.d." setting), holding everything else fixed. This isolates whether the spectrum actually matters.

Why not just…?

Why not just grid-search the damping on a tuning set? This is the real threat. On a single fixed protocol it may do as well as the spectral rule. The spectral rule's claimed advantage is transfer. It should keep working when the coil count, acceleration or mask changes, because $\kappa$ moves with those settings and a fixed tuned value does not. Stage three evaluates the tuned value on held-out protocols to settle this. If grid search transfers equally well, we will report that the theory's value is explanatory rather than practical.

Why not just use MAD-GAMP, which adapts without any singular values? Adaptive damping reacts only after the objective has already gone up. Its reported speedups were measured against sequential swept GAMP, not against a well-chosen fixed damping, and only on synthetic matrices. The head-to-head comparison of iterations and wall-clock time on real operators decides this. A hybrid is the natural fallback if neither wins outright: start MAD-GAMP at the $\kappa$-predicted damping.

Why not just use ADMM or conjugate gradient and skip message passing? For the Gaussian prior of stage one, conjugate gradient solves the problem exactly. That stage exists to test the theory, not to win the reconstruction task. For learned denoisers, the DD-VAMP paper found that message passing beats tuned PnP-ADMM by about 0.8 dB NMSE. It also notes that ADMM needs manual tuning of its penalty parameter, while VAMP estimates noise levels on its own. We keep ADMM-PR and PDHG in every table so that any practical loss is visible.

Decisive Checks

The first check is the $\kappa = 1$ single-coil Cartesian case. Gaussian GAMP should converge undamped there. If it diverges on real masks, the theorem's premise fails on real operators and that result is the headline.

The second check asks whether, on multi-coil operators, the predicted threshold falls within a factor of 1.2 of the bisected empirical threshold on at least 90% of slices. This is the hurdle we set to call the rule predictive.

The third check asks whether a single safety margin, fit on 30 tuning images, keeps the wavelet and DnCNN variants convergent across all four acceleration-and-coil settings.

The fourth check asks whether the tuning-free rule matches DD-VAMP++ on the published 19-slice split and holds up at full resolution.

Success on the first two establishes that the singular-value theory survives contact with real scanners. Failure at stage two, with success at stage one, would show exactly where damping stops being a spectral matter and becomes a property of the prior.