acceptodds
Under review as a conference paper at ICLR 2027

When Fisher Information Mispredicts Prior Sensitivity: Model Misspecification, Not Noise

Abstract

Physics-based inverse models resolve what the measurements cannot determine by adding a prior, so part of every recovered quantity comes from the prior rather than from the object measured. Fisher information is the standard tool for predicting how large that part is. We show that the prediction fails, characterise when, and give the mechanism. For a maximum a posteriori fit whose Fisher matrix has a near-null direction , the omitted curvature term multiplies every parameter's prior sensitivity by a single scalar, , in which the parameter index cancels. The sign of that scalar is the correlation between the fit residual and the model's curvature along . That correlation vanishes for noise independent of the model, which is what the standard justification assumes. On 4,608 clinical MRI exams it does not vanish, and two nulls built from the cohort's own residuals reject both halves of the assumption. The Fisher prediction is then too large by 29–38%, exceeds the measured sensitivity in to of exams, and loses the ordering across exams for . Because the error occupies one eigendirection, it can be repaired without computing the curvature: shifting a single eigenvalue recovers 92–99% of the full-Hessian correction at two gradient evaluations, whatever the parameter count. Both the failure and the repair reproduce on public diffusion MRI from another laboratory and on outdoor photovoltaic modules, and both vanish on transit photometry, where the residual is genuinely noise. The same substitution underlies Laplace approximations, influence functions and Fisher-based continual learning. Measured on a neural network, with weight decay as the prior, Fisher over-predicts how far the weights follow a shifted prior by to times across seven widths, and the exact curvature recovers it to . On the clinical images, self-supervised networks differing only in random seed disagree to more along the unconstrained direction than across it. What the measurements cannot determine, the training run decides.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Fisher information, Laplace approximation, Gauss-Newton approximation, prior sensitivity, uncertainty quantification, model misspecification, identifiability, inverse problems

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_ysT7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how a nonlinear maximum a posteriori estimate changes when the centre of its Gaussian prior is shifted. It contrasts the Fisher/Gauss–Newton prediction with the exact local sensitivity. The main argument is that, when the Fisher matrix has one weakly constrained direction, the omitted curvature along that direction can substantially alter prior sensitivity. The authors propose a rank-one correction that shifts the corresponding Fisher eigenvalue using a directional curvature estimate.

Critical strengths
  1. A concrete and practically relevant diagnostic. Perturbing the prior centre and refitting directly tests the quantity being predicted. This provides a clear empirical comparison between Fisher-based sensitivity, exact curvature, and observed displacement.

  2. Substantial empirical evidence in the main application. The clinical cohort is large, and the paper reports effect sizes, patient-resampled confidence intervals, exam-level consistency, and robustness across instruments and prior settings. The improvement in T1 rank correlation from 0.419 to 0.915 is especially compelling: the discrepancy affects ordering as well as magnitude.

  3. A potentially useful inexpensive correction. Estimating curvature along a single weak direction is attractive when that direction dominates sensitivity. Recovering 92–99% of the full-Hessian correction in the clinical setting provides meaningful evidence for this approach.

  4. Breadth and disclosure of limitations. The public-data experiments strengthen the case beyond one private cohort. The authors also report settings where the rank-one correction fails, acknowledge that it does not improve influence functions, and disclose that it is slower than the full approach at four parameters. These qualifications improve the credibility of the empirical presentation.

Critical weaknesses
  1. The central approximation requires additional assumptions. Equation 6 gives the sensitivity ratio as approximately 1−v⊤Cv/λ1-v^\top Cv/\lambda. This requires small directional curvature relative to λ\lambda, in addition to the stated spectral-gap assumptions. Yet the reported v⊤Cv/λv^\top Cv/\lambda is 0.655 for clinical MRI and exceeds 1 in several other experiments, making the first-order approximation inaccurate or even negative. Appendix C acknowledges this and supplies the more appropriate rational expression, but the abstract and main theoretical presentation still foreground the first-order result. More fundamentally, the gap f1≪λ≪f2f_1\ll\lambda\ll f_2 alone does not justify neglecting the other blocks of CC. In the Fisher eigenbasis, the effective weak-direction curvature includes the Schur-complement term f1+λ+C11−C1⊥(F⊥+λI+C⊥⊥)−1C⊥1f_1+\lambda+C_{11}-C_{1\perp}(F_\perp+\lambda I+C_{\perp\perp})^{-1}C_{\perp1}. Bounds on cross-direction coupling and the constrained block are needed to justify dropping this term. The rank-one approximation is empirically successful in some settings, but its stated theoretical conditions are incomplete.

  2. The controls do not isolate misspecification from fitting and regularization. The residual and curvature are evaluated at a fitted parameter that depends on the observations and prior. Even with a correctly specified model and zero-mean observation noise, the fitted residual need not be independent of curvature or have zero mean under regularization. The proposed nulls also require stronger assumptions than stated. Zero mean does not imply sign symmetry or a 50% probability of either sign. Voxel permutation requires exchangeability assumptions and may disrupt anatomical structure and spatial dependence. Consequently, these tests demonstrate structured residual–curvature alignment more directly than they establish its cause. A controlled synthetic experiment should vary model misspecification while holding acquisition, noise, prior, and optimization fixed. Matched correctly specified simulations would be particularly valuable for supporting the title’s “not noise” claim.

  3. Several claims about neural-network curvature are incorrect. Section 4.4 states that a ReLU network has C=0C=0 almost everywhere because it is piecewise linear. ReLU networks are piecewise linear in their inputs, but generally not jointly piecewise linear in all trainable weights. Mixed second derivatives across layers can remain nonzero within a fixed activation region. The assertion that near interpolation necessarily produces indefinite posterior curvature is also problematic. At exact interpolation, the residual curvature term vanishes, so H=F⪰0H=F\succeq0, and H+λIH+\lambda I is positive definite for λ>0\lambda>0. Indefinite curvature at returned solutions should therefore prompt investigation of convergence and optimization, rather than be attributed generally to interpolation.

  4. The network admission criteria risk selecting favourable cases. Appendix D admits runs partly based on observed displacement being aligned with vv, positive curvature, and measured sensitivity lying in (0,1](0,1]. These criteria condition the evaluation on properties related to the proposed explanation. Moreover, the relevant minimum condition is H+λI≻0H+\lambda I\succ0, rather than necessarily H≻0H\succ0. A positive-definite posterior Hessian can coexist with negative likelihood curvature and prior sensitivity exceeding one. The manuscript should clarify its notation, justify exclusions independently of measured outcomes, and report results for all runs.

  5. Finite shifts are conflated with local derivatives and prior attribution. The exact formula predicts an infinitesimal derivative, whereas the main clinical experiment uses a 25% prior-centre shift. The appendix reports changes of up to approximately 10% with shift size, comparable to or larger than the remaining discrepancy after exact correction. A sequence of shrinking, preferably central, perturbations would distinguish nonlinear path effects from numerical or theoretical error. In addition, local sensitivity to the prior centre is not generally the fraction of a fitted value “coming from the prior.” That interpretation is justified in the linearized decomposition, but needs qualification for nonlinear fits without known ground truth.

  6. The broader machine-learning claims exceed the demonstrated scope. The network experiments concern small, single-hidden-layer regressors fitted with a specialized optimizer. The rank-one correction does not help influence functions, and the Laplace experiment does not establish improved calibration. Likewise, testing a diagonal Fisher matrix on prior sensitivity does not establish the effect on continual-learning performance. The distinction between Fisher information and Hessian curvature is already established in the literature, including work cited by the authors. The contribution would be clearer if framed around the empirical sensitivity diagnostic and its conditional correction, with more precise positioning against existing curvature analyses.

Questions for authors

Please see the weakness part.

3 comments
AuthorsOpenReview ↗

Response to Reviewer ysT7 (1/2): You Were Right on All Three Theory Points; Conditions Now Stated, Coupling Measured, and One Claim Narrowed

Thank you for this review. Your three theoretical objections were all correct, and acting on them changed the paper's claims rather than just its wording. We also appreciate your noting that the paper "reports settings where the rank-one correction fails" and that these qualifications "improve the credibility of the empirical presentation".

W1. Equation 6 gives the sensitivity ratio as approximately 1−v⊤Cv/λ1 - v^\top C v/\lambda. This requires small directional curvature relative to λ\lambda... the gap f1≪λ≪f2f_1 \ll \lambda \ll f_2 alone does not justify neglecting the other blocks of C. Bounds on cross-direction coupling and the constrained block are needed to justify dropping this term.

You are right that we never stated the conditions, and right again that the gap does not bound what we dropped. Both are fixed.

  1. The equation ladder is reordered. The exact rational form is now equation 7, the first-order form is equation 8 with its condition printed in the display, and the abstract quotes the exact one. The violation is not marginal: at v⊤Cv/λ≈+6v^\top C v/\lambda \approx +6 on photovoltaics, the first-order form predicts to 62% where the rational form predicts to 1.0%.

  2. We measured the coupling instead of bounding it (full table in our reply to 22am). It is 0.0000 of the kept curvature where the gap holds, and up to 168% where it does not.

  3. Your Schur point is sharp and it narrows one of our claims. The gap bounds the stiff Fisher block but says nothing about the stiff curvature block, which equation 6 also assumes is unmoved. We added a second diagnostic, equation 24, η⊥=∥(F⊥+λI)−1/2C⊥⊥(F⊥+λI)−1/2∥2\eta_\perp = \lVert (F_\perp + \lambda I)^{-1/2} C_{\perp\perp} (F_\perp + \lambda I)^{-1/2} \rVert_2, the stiff block's curvature against its own posterior curvature. It is 0.0247 median on the photovoltaic fits against 4.50 on the networks, separating the arms the same way the gap does.

  4. The claim this narrows. Our consequence (ii), that a parameter with no weight on the weak direction (vj=0v_j = 0) is untouched, assumes that diagnostic is negligible. In a constructed case with a gap of 10−410^{-4} and the weak-direction curvature exactly zero, so the rank-one repair predicts 1 and makes no change, the exact ratio still falls from 0.9998 to 0.1730 as a parameter's loading on the weak direction falls. Consequence (i) and the scalar are unaffected. The appendix now states consequence (ii) with this diagnostic as its condition.

W3. Section 4.4 states that a ReLU network has C = 0 almost everywhere because it is piecewise linear... The assertion that near interpolation necessarily produces indefinite posterior curvature is also problematic.

Both claims were wrong and both are gone.

  1. ReLU curvature is the same order as tanh, not zero. A ReLU network is linear in each layer's weights separately but not jointly, so only the blocks containing the activation's second derivative vanish. On the Airfoil fit the curvature norm is 3261.5 for ReLU against 5100.5 for tanh, with the first-layer block exactly zero and the cross-layer block carrying 2306.2. We use tanh because the second derivative must exist everywhere.

  2. You are right about interpolation. At exact interpolation the residual term vanishes and the data Hessian is positive semidefinite. The Yacht fits hold negative data curvature and a positive-definite posterior at the same time, which is ordinary at a MAP solution; 22am independently drove the gradient norm to 1.6×10−131.6\times10^{-13} with the signs unchanged. Section 4.6 now says this instead of attributing the exclusions to interpolation.

In the revised PDF (changes in blue): the equation ladder in Section 3.1, the new Appendix C diagnostic, and the corrected Section 4.4 and 4.6. Our second reply covers your remaining three points. We hope this answers W1 and W3, and we would welcome your view on whether the narrowed claims are now stated correctly.

AuthorsOpenReview ↗

Response to Reviewer ysT7 (2/2): Matched Control Built and It Changed Our Title; Admission Criteria Fixed, Central Shifts Run, Claims Narrowed to What We Measured

W2. The controls do not isolate misspecification from fitting and regularization... A controlled synthetic experiment should vary model misspecification while holding acquisition, noise, prior, and optimization fixed. Matched correctly specified simulations would be particularly valuable for supporting the title's "not noise" claim.

We built exactly the experiment you specify, and it changed our title.

  1. The effect tracks alignment, not cause. Varying misspecification with acquisition, noise, prior and optimisation fixed, the residual grows by five orders of magnitude while the Fisher error shrinks (table in our reply to 22am). The structural reason is that at a weak-prior MAP the normal equations push injected misfit where the curvature direction has no component.

  2. The "not noise" half survives. A correct-model noiseless fit over-predicts by 32%, parameter-independent to 0.2%, with no noise anywhere in it.

  3. The title now claims only what the tests establish. It reads curvature-correlated residuals, not noise. As you put it, the tests demonstrate structured residual-curvature alignment more directly than they establish its cause, so the title now names the alignment.

W4. Appendix D admits runs partly based on observed displacement being aligned with v, positive curvature, and measured sensitivity lying in (0, 1]. These criteria condition the evaluation on properties related to the proposed explanation.

  1. The sensitivity bound is removed from admission, and the posterior test replaced the data-Hessian test. You are right that the alignment test still conditions on a measured outcome, so we report both sets: the median exact ratio is 1.0003 with the filter and 1.0003 without, over all 77 runs, with only the tail moving. That test catches refits that left the stationary point the implicit function theorem describes, visible as a 1/delta divergence.

W5. The exact formula predicts an infinitesimal derivative, whereas the main clinical experiment uses a 25% prior-centre shift... In addition, local sensitivity to the prior centre is not generally the fraction of a fitted value "coming from the prior."

  1. You identified the dominant error term. Differencing centrally at the same 25% shift moves measured/Hessian from 1.061 / 0.963 / 1.018 to 1.011 / 1.012 / 1.001. About 1% survives, flat across a 16-fold range.

  2. Your wording point stands too. Section 3 now calls equation 2 a linearisation and defines the phrase to that order, and the appendix states that the measured object is the derivative of the estimate with respect to the prior centre, defined for any twice-differentiable fit at a non-degenerate stationary point.

W6. The broader machine-learning claims exceed the demonstrated scope.

  1. We have adopted your framing. The contributions had claimed the failure and its repair reproduce on influence functions, where the repair in fact recovers -0.02%. That overreach is gone. The Discussion now takes each case separately: influence functions reorder (0.628 to 0.881) and the repair does not transfer; Laplace is a measured null; for elastic weight consolidation we claim only that a diagonal Fisher cannot see a mixed near-null direction.

  2. On your point that the Fisher-versus-Hessian distinction is already established, you are right that we cited none of the local prior-sensitivity literature. Related work now cites Giordano et al. (2018) on linear response for variational covariances and local robustness, and Immer et al. (2021) on the Gauss-Newton substitution as a local linearisation. Against that, what we add is the sign of the dropped term, its measurement at clinical scale, and a scalar that corrects it without the Hessian.

In the revised PDF (changes in blue): the title, Section 2's positioning, Section 3's wording on the linearisation, Section 4.1's central-difference results, and the rewritten Discussion.

Your review changed what this paper claims, and we are grateful for that. We hope the narrowed claims and the new controls address your concerns, and if so we would be grateful if you would consider raising your score.

AuthorsOpenReview ↗

Follow-up to Reviewer ysT7: Checking Whether Anything Remains Open Before the Discussion Period Closes

Dear Reviewer ysT7,

With the discussion period closing on 19 November, we wanted to check whether anything remains open on your side.

Acting on your review, we reordered the equation ladder so the exact scalar leads and the first-order form carries its condition, removed both incorrect network claims, replaced the data-Hessian admission test with the posterior test, built the controlled misspecification experiment you specified, and changed the title: that experiment showed the effect follows the residual's alignment with curvature rather than its cause, so "model misspecification" was not the right word. Your Schur point also led us to a second diagnostic, η⊥\eta_\perp, which narrowed one of our claims rather than supporting it.

We genuinely appreciate the care that review took, and we would like to use the remaining time well. If any of W1 to W6 still feels unresolved, or if something we changed has raised a new question, please tell us and we will do our best to address it before the deadline.

Reviewer_9GdwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper try to answer the question of what is the importance of measurement compare to the prior for the reconstructed parameters in the inverse problem. That is a solid question as we try to solve an ill posed inverse problem question and for recovering the parameters, we need to use prior to select the best answers from the null space. Fisher information can estimate which parameters are identifiable and how strongly a reconstruction depends on its prior. The authors argue that Fisher information can poorly estimate this prior dependence, especially when the forward model is nonlinear, Hessian shows itself here, and the residual contains systematic model errors rather than only measurement noise.

The authors showed with some assumptions that Fisher-based approximations neglect a residual-dependent second-order curvature term in the objective's Hessian. Although the term is often assumed to be negligible, the authors show that in the experiments it can remain significant along directions poorly constrained by measurements. They adjust the Fisher matrix along its weakest eigenvector using a directional curvature measurement. Their approach attempts to recover the accuracy of the full Hessian without constructing it.

Critical strengths

The paper addresses a fundamental issue in inverse problems: Its distinction between measurement and prior sensitivity is valuable. The way they experiment a Fisher-based approximations can be wrong metric for evaluation this concern is an important findings for non-linear problems.

The authors propose a rank-one correction to the Fisher information matrix that accounts for the neglected residual-dependent curvature along the weakest identifiable direction. This provides an efficient approximation to the full-Hessian sensitivity without explicitly computing the entire Hessian.

The paper provides extensive experiments across MRI, photovoltaic systems, and neural networks that support its theoretical claims. On 4,608 clinical MRI examinations, Fisher information overestimates prior sensitivity by 29–38%, while the proposed rank-one correction recovers 92–99% of the full-Hessian improvement.

Critical weaknesses

First of all, please if you put the equation in the abstract, define all of the parameters there. I do not have any information of \lambda and C there. The \hat{x} is obvious in the Eq1 but you should define explicitly what is that.

The method is limited to complexity of null space: Many modern inverse problems have a large nullspace, especially when the number of unknowns substantially exceeds the number of measurements. The authors already demonstrate that their rank-one method recovers only 31–60% of the full-Hessian correction with two null directions. For practical tomography, compressive imaging, or high-dimensional reconstruction, this limitation may be substantial.

The neural-network experiments do not generalizable to broad deep-learning applicability: The principal neural-network curvature experiments use relatively small tanh networks with 71–351 parameters.That is not enough to establish that the method is useful for modern deep networks. Moreover, the paper states that ReLU networks have zero relevant curvature almost everywhere because they are piecewise linear. This requires correction: a multilayer ReLU network is piecewise linear in its input, but generally not in all its weights jointly.

Computation claim: The authors highlight that the directional curvature can be evaluated using two gradient evaluations, independently of the parameter dimension. That is true for evaluating a directional second derivative once the direction is known, but the full procedure also requires identifying the weakest eigenvector of F and applying the corrected inverse.

Finding this eigenvector can be expensive in large neural networks, especially when the relevant eigenvalue is near zero and the nullspace is high-dimensional.The paper acknowledges that its rank-one procedure is slower than the full-Hessian route for the four-parameter clinical fits.

Questions for authors

Have the authors considered extending the method to a rank-(k) correction? How would its computational complexity and accuracy change as the nullspace dimension increases?

The proposed correction modifies only the smallest eigenvalue of the Fisher matrix. However, the residual-dependent curvature term may also introduce interactions between different eigendirections. Can the authors provide theoretical error bounds for the rank-one approximation or experimentally evaluate the effect of off-diagonal curvature terms?

Could the authors report the total computational cost, including eigendecomposition, and compare it against full-Hessian and Hessian-vector-product approaches, especially for higher-dimensional problems?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 9Gdw: Thank You for the ReLU Correction; Abstract Fixed, Off-Diagonal Curvature Measured, and Full Cost Reported

Thank you for these three questions; each one produced a measurement we did not have. Your ReLU correction was right and the claim is gone (the measured replacement is in our reply to ysT7), and λ\lambda, CC and x^\hat{x} are now all defined in the abstract at first use.

Q1. Have the authors considered extending the method to a rank-(k) correction? How would its computational complexity and accuracy change as the nullspace dimension increases?

  1. We do not have one, and the obstacle is structural rather than computational. Widening the hidden layer from 10 to 160 units takes the number of Fisher eigenvalues below λ\lambda from 2 to 470, and the ratio of the two smallest from 2.16 to 0.88. The bottom of the spectrum becomes a cluster, so "the k weakest eigenvectors" stops being well defined before k is large enough to help. Our higher-codimension numbers are consistent with that: 31-60% recovery on two directions, 34-48% on three.

Q2. Can the authors provide theoretical error bounds for the rank-one approximation or experimentally evaluate the effect of off-diagonal curvature terms?

  1. We measured it, and the answer splits cleanly on the spectral gap. The dropped coupling is 0.0000 of the kept curvature on the photovoltaic fits (λ/f2=6.95×10−4\lambda/f_2 = 6.95\times10^{-4}) and runs from −1.678-1.678 to +0.375 on the networks (λ/f2=2.098\lambda/f_2 = 2.098). The rank-one error follows: 0.0009 at the 90th percentile where the gap holds, 0.275 where it does not.

  2. The stiff block needed its own diagnostic, which we added as equation 24, η⊥=∥(F⊥+λI)−1/2C⊥⊥(F⊥+λI)−1/2∥2\eta_\perp = \lVert (F_\perp + \lambda I)^{-1/2} C_{\perp\perp} (F_\perp + \lambda I)^{-1/2} \rVert_2: median 0.0247 on the photovoltaic fits against 4.50 on the networks. It narrows one of our claims; the reply to ysT7 gives the detail.

Q3. Could the authors report the total computational cost, including eigendecomposition, and compare it against full-Hessian and Hessian-vector-product approaches, especially for higher-dimensional problems?

Table R4. Per-fit cost on the Airfoil network, including the eigendecomposition.

hpdense eigLanczosfull Hessiandirectional deriv.eigs below λ\lambdacos(Lanczos, dense)
10711.6 ms3.5 ms52 ms0.14 ms20.614
4028112 ms11 ms775 ms0.45 ms120.055
1601121434 ms36 ms13,621 ms1.8 ms4700.004
  1. The matrix-free route scales but does not recover the direction, which is the more important finding. The cosine against the true eigenvector falls to 0.004, and more iterations do not help even under a spectral transformation that moves the small end to the fast end. The weakest eigenvector is not separated, so no Krylov method isolates it, and the repair fails at its premise before cost becomes the question.

  2. Where the premise holds the cost is negligible: at p = 4 on the cohort and p = 5 on the photovoltaic fits, the dense decomposition takes under two milliseconds. The appendix now scopes the method to low-dimensional parameter spaces with an isolated null direction, and names the count of eigenvalues below λ\lambda as the first number to check. For the large-nullspace problems you name, we would not recommend the rank-one repair, and on modern deep networks our evidence does not establish applicability.

In the revised PDF (changes in blue): the abstract now defines every symbol, Section 4.4 carries the corrected ReLU statement, and Appendices C and E carry the new diagnostics and the cost table above.

Thank you again for the ReLU correction in particular; it was a real error. We hope these answers address your three questions, and if they do we would be grateful if you would consider raising your score.

Thanks for the new experiment and detailed response. I have some questions regarding to understand deeply and correct me if I am wrong. Q1: The additional experiments confirm that the rank-one correction becomes unreliable as the dimensionality and number of weakly constrained directions increase. If that is true, then that leads to reduce the applicability of the model to other experiments as we have more than one weak direction. If so, do you proceed any error analysis of the method and how error propagate over the different number of poorly constrained directions. Q2:In lower dimension you can calculate Hessian with less computation that the revised fisher information and in higher dimensions your method have computationally better while leading to rank one assumption. Is there a setting where the rank-one correction is both sufficiently accurate and computationally more efficient than an exact-Hessian?

My remaining concern is about the significance and applicability of the proposed methodological contribution

AuthorsOpenReview ↗

Response to Reviewer 9Gdw (1/2): Your Question Prompted a Controlled Sweep, and It Identified What Actually Governs the Error

Thank you for your reply! This is a question we should have answered in the submission, and running it properly taught us something. Our codimension numbers (92–99% / 31–60% / 34–48%) come from three MRI acquisition arms, and as Section 4 and Appendix H note, those arms are not a controlled series: dropping the ADC anchor changes which parameter is unconstrained as well as how many. Yours is the right question to ask of that evidence, so we built the controlled version, a single network widened from 71 to 1121 parameters, which moves the number of eigenvalues of FF below λ\lambda while holding the dataset, optimiser and prior fixed.

Q1. The additional experiments confirm that the rank-one correction becomes unreliable as the dimensionality and number of weakly constrained directions increase... do you proceed any error analysis of the method and how error propagate over the different number of poorly constrained directions.

  1. The error analysis is already in closed form, and the count does not appear in it. Equation 6, the mixture form, keeps every direction: the ratio is (vj2/(f1+v⊤Cv+λ)+Rj)/(vj2/(f1+λ)+Rj)(v_j^2/(f_1 + v^\top C v + \lambda) + R_j)/(v_j^2/(f_1+\lambda) + R_j), where RjR_j sums over all the remaining directions. The rank-one repair is equation 6 with RjR_j frozen at its Fisher value, so its error is exactly the RjR_j mismatch. The number of weak directions enters only through that term, which is why it does not predict the error on its own.

  2. At a fixed count of one weak direction, the error spans the whole range. Two parameters, a gap of λ/f2=10−4\lambda/f_2 = 10^{-4}, and CC acting only in the stiff block so v⊤Cvv^\top C v is zero to machine precision: the repair predicts exactly 1.0000 in every row. Only the share of the parameter's leak carried by vv changes.

Table R5. One weak direction throughout; only the vv-loading varies. Four of six rows; the two omitted sit on the same monotone curve, at 1.1% and 74.3%.

share of leak on vvexact ratiorank-one predictsrank-one error
0.99970.99981.00000.0%
0.75290.79411.000020.6%
0.52310.60261.000039.7%
0.00760.17301.000082.7%
  1. Across widths the error is not monotone in the count either. One weak direction gives the worst error of the four admitted fits; eight directions give a better one than two.

Table R6. Airfoil network widened, λ=10−2\lambda = 10^{-2}, same fit throughout.

ppeigs below λ\lambdaλ/f2\lambda/f_2Fisherrank-oneexactadmitted
7122.590.2941.0591.007yes
14110.820.4871.2180.995yes
28186.680.3321.0841.002yes
5614321.80.4781.2421.001yes
1121468—2.8424.5412.443no
  1. The last row answers "what happens as dimensionality grows", and it repays a careful read. At p=1121p = 1121 the measurement itself stops being well posed: f1f_1 and f2f_2 are both −10−11-10^{-11}, zero to machine precision, so there is no isolated near-null direction and λ/f2\lambda/f_2 is meaningless. The refit no longer tracks the direction we perturbed, with cos⁡(Δw,v)=0.056\cos(\Delta w, v) = 0.056 against 0.920.92 to 0.9990.999 at every smaller width, and the admission test of Appendix D rejects it. The exact curvature misses by a factor of 2.4 there for the same reason the repair does: nothing is being measured. We report the row rather than drop it, because it is where the method's premise fails, but it is not evidence about the rank-one approximation.

  2. Both results follow from point 1, and together they sharpen the claim. On the MRI arms the step from one to more than one null direction is real. On a widened network the count moves without tracking the error. Both are consistent with the closed form: what governs the repair is how much of a parameter's leak rides on vv. Changing the codimension on the MRI arms changes that; widening a network does not change it monotonically. From four admitted widths we would not want to fit a predictor, so we report λ/f2\lambda/f_2 and the count together and let a reader check their own problem against both.

We hope this answers Q1. Our second comment takes up Q2 and the significance question.

Reviewer_22amOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how a nonlinear maximum a posteriori estimate changes when the centre of its Gaussian prior moves. It compares the Fisher/Gauss-Newton prediction MF=λ(F+λI)−1M_F=\lambda(F+\lambda I)^{-1} with the local derivative MH=λ(H+λI)−1M_H=\lambda(H+\lambda I)^{-1}, where the data Hessian H=F+CH=F+C includes residual-weighted model curvature. When one weak Fisher direction dominates and the remaining curvature terms are negligible, the paper approximates this correction by shifting one eigenvalue. On 4,608 clinical MRI exams, the reported prior-induced displacements are smaller than the Fisher predictions, and restoring curvature substantially improves T1 ordering across exams. Further experiments cover diffusion MRI, photovoltaic curves, transit photometry, regression networks and influence functions. The proposed explanation is that structured fitted residuals produce a systematic curvature correction, which the paper attributes to model misspecification.

Critical strengths
  1. The clinical experiment measures an interpretable effect and shows a meaningful failure of ordering.

Moving the prior centre and refitting the same observations is a direct way to measure the dependence of an estimate on that prior. Table 2 gives measured-to-Fisher ratios of 0.706, 0.636 and 0.617 for PD, T1 and T2, with patient-resampled intervals and the fraction of exams below prediction. The implementation confirms that the bootstrap operates at the patient level. The consistency across 4,608 exams makes the discrepancy worth investigating.

The strongest result is the reported T1 rank correlation, which rises from 0.419 with Fisher to 0.915 with the full Hessian. A scale correction cannot restore lost ordering. The paper also reports the remaining disagreement after the Hessian correction, instead of implying exact agreement for a finite prior shift. These details give the main MRI result substance.

  1. Several numerical checks support the curvature calculation, and the paper reports limits of the cheap repair.

The inspected release contains useful tests, including a correct-model noiseless MRI case and an explicit noise-versus-mismatch comparison for an exponential model. All 64 selected core test functions executed in the audit passed. In a separate execution of the released Airfoil configuration with h=10h=10 and seed = 0, the measured-to-Fisher, measured-to-rank-one and measured-to-full-Hessian ratios were 0.294061, 1.059296 and 1.006727, matching the rounded values in Table 3. This is a limited reproduction, but it supports the numerical behaviour of that example.

The reported limits are also informative. The rank-one repair recovers 92–99% of the full-Hessian correction in the clinical one-null-direction setting, but substantially less in the higher-codimension arms. It recovers essentially none of the full-Hessian improvement in influence ranking. The acknowledgement that the scalar route is slower at p=4p=4 is appropriate; fewer gradient evaluations are not presented as a demonstrated runtime gain there.

Critical weaknesses
  1. The network admission rule rejects valid local MAP fits and therefore does not support the stated exclusion of Yacht results.

This issue can be traced to the implementation. In physharm/analysis/net_arm.py:33–50, curvature() differentiates the data-loss gradient, so its HH excludes the prior. The exact sensitivity correctly uses A=H+λIA=H+\lambda I at line 91. However, well_posed at lines 146–147 requires the minimum eigenvalue of HH itself to be positive and the measured sensitivity to lie in (0,1](0,1]. A nondegenerate local MAP minimum requires positive posterior curvature H+λIH+\lambda I; the data Hessian can have negative curvature that the prior compensates. Positive posterior curvature also permits a directional sensitivity above one.

Executing the released Yacht configuration with h=10h=10, seed = 0, λ=0.01\lambda=0.01, δ=0.05\delta=0.05 and 3,000 iterations gives:

Minimum data-Hessian eigenvalue: −0.00206012-0.00206012.

Minimum posterior-Hessian eigenvalue: +0.00793988+0.00793988.

Measured sensitivity: 1.031848251.03184825.

Full-Hessian prediction: 1.031216551.03121655.

Displacement alignment with vv: 0.9855180.985518.

The measured-to-full-Hessian ratio is 1.000613. The stiff-direction control also passes, yet the implementation returns well_posed = False and admit = False. Further stationary-point refinement reduces the gradient norm to approximately 1.6×10−131.6\times10^{-13} while retaining positive posterior curvature. Changing the Hessian differencing step from 10−410^{-4} to 10−610^{-6} preserves the signs of both eigenvalues. This is strong evidence that the exclusion is caused by the criterion, rather than a convergence or differencing failure.

This configuration has f2λ≈0.00190\frac{f_2}{\lambda}\approx0.00190, so it does not satisfy the isolated-weak-direction assumption λ≪f2\lambda\ll f_2. It is not a counterexample to the rank-one approximation within its stated domain. It does establish that the local sensitivity is well defined in a released Yacht configuration and accurately predicted by the full Hessian. Table 1 and Section 4.6 should therefore not attribute all Yacht exclusions to an indefinite posterior or an undefined sensitivity. Posterior validity, spectral-gap applicability and the additional alignment restrictions should be reported separately.

The retained-run accounting also needs repair. The current driver generates 16 dataset runs per dataset, but paper_iclr/make_tables.py:37–40 hardcodes “0 of 8 admitted” whenever a dataset has no admitted runs. Supplying 16 excluded records reproduces that output. The original run artifacts were absent from the inspected release, so this does not establish the denominator used in the original manuscript run. It does establish that the current generator cannot verify that denominator. The counts and directional summaries should be regenerated from the actual records after correcting the posterior check. Without the full rerun, I cannot quantify how much the headline network result changes.

  1. The mechanism claim and some validation claims extend beyond what the implemented comparisons establish.

The code provides a particularly relevant control. tests/test_prior_shift.py:46–58 explicitly tests a correctly specified, noiseless MRI render with truth away from the prior centre. Running a small instance of this fixture with a 0.01 prior shift gives measured-to-voxel-Fisher ratios of approximately 0.760, 0.755 and 0.755 for PD, T1 and T2. The corresponding measured-to-full-Hessian ratios are approximately 1.004, 0.998 and 0.996. Thus forward-model misspecification is not necessary for this type of Fisher overprediction in the released MRI setup. Prior-induced fitted residuals can produce it as well.

This fixture has a nonzero fitted residual, so it does not contradict the conditional argument for zero-mean residuals independent of curvature. The distinction is between the measurement noise at the true parameters and the residual after regularized fitting. Section 3.2 acknowledges dependence on the fit, and Section 4.2 explicitly allows a prior that prevents an exact fit to supply the residual. The title and conclusion should reflect those qualifications. If “misspecification” is intended to include this prior-induced residual, that broader definition should be explicit. Otherwise, a matched MRI control that varies forward-model discrepancy while retaining the priors and refitting procedure is needed to support the narrower attribution. The exponential-model noise-versus-mismatch control already exists and passes; it remains unclear how far that result settles the attribution in the MRI setting.

The voxel-level residual nulls also need explicit symmetry and exchangeability assumptions. physharm/analysis/prior_shift.py:240–282 permutes residual columns globally and flips signs per voxel, without using spatial blocks. The spatial sampling and dependence need to be described so the reader can assess those assumptions. The absence of blocking alone does not establish that the reported clinical null results are invalid.

Two plotting paths also need clearer descriptions. In Figure 5, experiments/pv_runs.py:75–93 fills measured_ratio with (MH)jj(MF)jj\frac{(M_H){jj}}{(M_F){jj}}. These points compare the rank-one expression with a full-Hessian derivative; they are not independent coordinate-wise perturb-and-refit measurements. The axis already identifies the matrix ratio, but the legend and accompanying discussion describe it as a measurement. Independent photovoltaic refits do exist for the top-level vv-direction measurement in physharm/analysis/pv_arm.py:91–92. The correction is to identify which validation each result supplies. In Figure 6, paper_iclr/make_figures.py:81–93 varies the photovoltaic prior precision while keeping the fitted FF and CC fixed. The network curve instead uses models refitted at each precision. The photovoltaic curve is a valid conditional calculation at fixed geometry, but it does not test whether the same direction of error persists after changing the prior and refitting.

Finally, Appendix C explicitly requires negligible action of CC outside the weak-direction block. That condition deserves a quantitative diagnostic alongside the reported approximation accuracy. In the weak/stiff basis, write the posterior Hessian as

\begin{pmatrix} a+c & b^\top\ b & D \end{pmatrix}, \qquad a=f_1+\lambda,\quad c=v^\top C v,$$ with $D$ the posterior block on the remaining directions. Its inverse along $v$ is $$v^\top(H+\lambda I)^{-1}v =\frac{1}{a+c-b^\top D^{-1}b}.$$ The one-eigenvalue shift omits the coupling term. Reporting its size relative to $a+c$, together with $\lambda/f_2$ and the per-fit approximation error, would help explain the T2 tail and identify where the cheap repair is reliable.
Questions for authors

After replacing the data-Hessian positivity test with a posterior-Hessian test and separating it from the sensitivity-bound and approximation restrictions, how many of the network runs remain valid and how do their Fisher and full-Hessian comparisons change? Please include the positive-posterior Yacht cases, report overlapping rejection reasons, and derive all denominators from the saved records. This would determine whether the broad directional conclusion survives the corrected selection.

How should the correct-model noiseless MRI fixture be reconciled with the title's attribution to model misspecification? Can a matched MRI control separate the contribution of the prior from that of a controlled forward-model discrepancy, or should the conclusion explicitly cover prior-induced residual structure? Please retain the distinction between raw measurement noise and fitted residuals and state the symmetry and spatial exchangeability assumptions of the residual nulls.

Can Figure 5 and its associated accuracy statement distinguish full-Hessian derivative comparisons from independent prior-shift refits, and can Figure 6 distinguish fixed-geometry photovoltaic calculations from the refitted network sweep? If the claim is robustness after changing the prior and refitting, what results support it for the photovoltaic arm?

Across the evaluated fits, how large are the omitted weak-to-stiff coupling terms, and do they explain the approximation's outliers? For the remaining clinical full-Hessian ratios of 1.061, 0.963 and 1.018, do smaller central prior shifts converge toward one? These checks would separate approximation error from the finite-displacement versus point-derivative comparison.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 22am (1/3): Thank You for Running Our Code; Both Bugs Are Fixed, All 48 Configurations Now Admitted, and the Conclusion Holds

Thank you for running the released configurations and auditing the implementation. Both defects you traced are real, and both are fixed. We also appreciate your noting that "the consistency across 4,608 exams makes the discrepancy worth investigating" and that the T1T_1 rank correlation rising from 0.419 to 0.915 is a result "a scale correction cannot restore".

Q1. After replacing the data-Hessian positivity test with a posterior-Hessian test and separating it from the sensitivity-bound and approximation restrictions, how many of the network runs remain valid and how do their Fisher and full-Hessian comparisons change? Please include the positive-posterior Yacht cases, report overlapping rejection reasons, and derive all denominators from the saved records.

Thank you for catching this! Your diagnosis was exactly right, including the line numbers. We have made all three changes you asked for.

  1. All 48 configurations are now admitted on the posterior test. well_posed now tests λmin⁡(H+λI)>0\lambda_{\min}(H + \lambda I) > 0. Every one of the 48 has a positive-definite posterior, so this test now rejects none. The sensitivity bound was removed from admission entirely; it is still reported, but never used to admit, because conditioning on a measured outcome is what ysT7 objected to.

  2. Yacht goes from 0 admitted runs to 3, and your numbers reproduce ours to eight significant figures. The three admitted fits give Fisher ratios 1.033 / 1.643 / 1.022 and exact ratios 1.001 / 1.042 / 0.998.

  3. Overlapping rejection reasons, from the saved records. Over 77 runs: the stiff-direction test rejects 0, the posterior test rejects 0, and the alignment test rejects 18, leaving 59. The spectral gap of equation 5 holds in 12 of 77 and is now reported separately. The hardcoded "0 of 8" denominator you flagged is gone.

Table R1. Corrected network arm, per dataset. Values are median ratios of measured to predicted; 1.000 is perfect.

armadmittedFisherrank-oneexact
airfoil16 of 160.3391.1161.002
wine12 of 160.5671.2951.000
yacht3 of 161.0331.1851.001
  1. The conclusion does not depend on the selection. Over all 77 runs with no admission test at all, the median exact ratio is 1.0003 and the median Fisher ratio 0.487; over the 59 admitted they are 1.0003 and 0.387. The alignment test only improves the tail, where the 90th percentile of the exact error falls from 0.134 to 0.048. Both sets are now in the appendix, unfiltered first.

  2. Yacht also found a real contradiction in our own appendix, which we are grateful for. Appendix I claimed the direction survives every setting tested, with the curvature term positive and Fisher over-predicting. Yacht breaks both halves: Fisher under-predicts there, and one admitted run has a negative curvature term (-0.186). Appendix I now states this exception explicitly instead of claiming a universal direction.

In short, the corrected selection admits more runs, not fewer, and the broad directional conclusion survives it. Table 3 is unchanged, with h = 10 still at 0.294 / 1.059 / 1.007, matching the values you reproduced exactly.

In the revised PDF (changes in blue): Section 4.6 and Appendix D carry the corrected admission rule and the unfiltered results, and Appendix I now states the yacht exception. Thank you again for the audit; it found things no amount of rereading would have. We would welcome your view on whether this settles the selection question.

AuthorsOpenReview ↗

Response to Reviewer 22am (2/3): The Noiseless Fixture You Ran Changed Our Title; Matched Control Added, Figures Relabelled, Coupling Measured

Q2. How should the correct-model noiseless MRI fixture be reconciled with the title's attribution to model misspecification? Can a matched MRI control separate the contribution of the prior from that of a controlled forward-model discrepancy, or should the conclusion explicitly cover prior-induced residual structure?

You were right, and we changed the title because of this. Your reading of that fixture is correct, and we reproduce your numbers exactly.

  1. Your measurement replicates. We get measured-to-Fisher ratios of 0.7577 / 0.7567 / 0.7581 for PD / T1T_1 / T2T_2 against the 0.760 / 0.755 / 0.755 you measured. We quote the reciprocal, so our "32%" and your "24%" are the same number. The two Fisher forms agree to six digits, so the gap is not an artifact of a nominal-parameter Fisher against a fitted-point Hessian.

  2. The matched control you asked for now exists, varying forward-model discrepancy with acquisition, noise, prior and optimiser held fixed. It gave a cleaner answer than we expected: the effect follows the residual's alignment with curvature, not its size and not its origin.

Table R2. Matched control, eight measurements against four parameters.

injected gainresidual normcos(r, g)meas/Fishermeas/Hessian
0.006.3e-6-0.3400.9551.0003
0.201.7e-1-0.0000.9640.9998
0.806.9e-1-0.0000.9871.0001
  1. The residual grows by five orders of magnitude while the Fisher error shrinks. At a weak-prior MAP the normal equations push injected misfit into the orthogonal complement of the Jacobian's range, where the curvature direction has no component. Only the part the prior returns to that range drives the effect. This is why the title now claims an alignment rather than a cause.

  2. Arm A1 cannot host this control at all: it has three measurements against four parameters with full row rank, so the model represents any observation and an injected gain of 0.80 leaves a residual of 2.1×10−42.1\times10^{-4}. Every residual at a converged A1 fit is prior-induced.

  3. On the nulls, you are right and we now state the assumptions. The permutation null assumes voxel exchangeability, which anatomy violates, and we do not have the spatially blocked version that would fix it; the sign-flip null assumes symmetry about zero, stronger than zero mean. The exam-level counts need neither, since exams are independent.

In short, the control you asked for is what retitled the paper: the effect follows the residual's alignment with curvature rather than its origin.

In the revised PDF (changes in blue): the title, Section 3.2 on the sign condition, and Section 4.2. Your reading of that noiseless fixture is the single comment that changed what this paper claims, and we are grateful for it. Our next comment covers your questions on the figures and the omitted coupling.

AuthorsOpenReview ↗

Response to Reviewer 22am (3/3): Both Figures Relabelled, the Omitted Coupling Measured, and Central Shifts Do Converge Toward One

Continuing from our previous comment, this covers your questions on the two figures and on the omitted coupling.

Q3. Can Figure 5 and its associated accuracy statement distinguish full-Hessian derivative comparisons from independent prior-shift refits, and can Figure 6 distinguish fixed-geometry photovoltaic calculations from the refitted network sweep?

Thank you for catching this mislabelling.

  1. Figure 5's field is renamed. pv_runs.py filled measured_ratio with a ratio of two curvature matrices. It is now exact_ratio in the script, the figure builder and the artifact, and the caption says the panel validates the closed form against exact linear algebra.

  2. Figure 6's two curves are now separated in the caption, and we say plainly that nothing supports persistence after refitting for the photovoltaic arm, rather than letting a shared axis imply it.

Q4. Across the evaluated fits, how large are the omitted weak-to-stiff coupling terms, and do they explain the approximation's outliers? For the remaining clinical full-Hessian ratios of 1.061, 0.963 and 1.018, do smaller central prior shifts converge toward one?

  1. We measured the coupling using your Schur expression, now equation 23. The separation falls exactly on the spectral gap.

Table R3. Omitted coupling against the gap.

λ/f2\lambda/f_2coupling / keptrank-1 error
photovoltaics, 60 curves6.95e-40.0000 (max 0.003)0.0009 p90
network, 16 fits2.098+0.0864 (-1.678 to +0.375)0.275 p90
  1. Yes, central shifts converge toward one, and you identified why. Most of the residue was the one-sided estimator's O(δ)O(\delta) bias. Differencing centrally at the same shift moves measured/Hessian from 1.061 / 0.963 / 1.018 to 1.011 / 1.012 / 1.001, and measured/rank-one to 1.000 / 1.009 / 1.000, which is what Section 4.1 now reports. Across a 16-fold sweep the central values are flat to a maximum drift of 0.0065.

In short, both corrections are more accurate than Section 4.1 originally reported, and the rank-one repair is within 0.9% on every parameter.

In the revised PDF (changes in blue): the Figure 5 and 6 captions, Section 4.1's central-difference results, and the new Appendix C diagnostics built on your Schur expression.

Thank you again for the depth of this audit. We hope these three comments answer all four questions, and we would be glad to run anything further during the discussion period. If you feel your concerns are resolved, we would be grateful if you would consider raising your score.

AuthorsOpenReview ↗

Follow-up to Reviewer 22am: Anything Further We Can Run Before the Deadline?

Dear Reviewer 22am,

The discussion period closes on 19 November, and we wanted to check whether anything is still open.

Acting on your questions we corrected the admission rule, rebuilt the clinical cohort from source imaging and re-ran it with central differencing, relabelled Figures 5 and 6, and measured the omitted coupling using your Schur expression.

We are grateful for the time your audit must have taken. If any of Q1 to Q4 is incompletely answered, or you would like a further measurement, please say so and we will run it before the deadline.

Reviewer_aANvOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This article makes a contribution to the field of misspecification in model fitting. Specifically proposes a novel method to quantify the contribution of the prior with respect to that of the data attachment term (if we set up the AI algorithm as an MAP problem . Their formalsim, proposed in section 3, is focused on problems stated as a regression problem and the probabilsitic model has a Gaussian prior. The main argument is that taking into account the curvature of the fitting problem, usually omitted through approximations and work on the Fisher information matrix for misspecification, provides valuable information of the contribution of the prior to the regression prediction. Furthermore they propose a correction term which is cheap and needs few gradient evaluations and no Hessian evaluations to estimate more than 92%, and up to 99%, of the curvature term.

The authors show the effect of their contribution in 7 datasets and different learned regression functions.

Critical strengths

The main strengths are as follows:

  1. A clear, simple formulation. The misspecification issue in terms of the Fisher information matrix and a correction term is very general and can be checked.
  2. The authors made an effort to show the widespread applicability of their contribution. They tested several datasets coming from very different fields (medical imaging, photovoltaic cells, for instance). Furthermore, the authors have carefully designed the nulls for each analysis.
  3. The proposed repair to the lack of curvature term is cheap. Meaning that two extra gradient evaluations are enough to estimate most of the required correction term. This is an interesting finding which also provides a user of the term with a method that has much more chance of being numerically stable than those that would have required Hessian computations.
  4. If this becomes a widespread tool, it has the potential of being applicable across the board in parameter estimation issues.
Critical weaknesses
  • The article is really hard to follow. It requires a much simpler way to ease the reader into the regression problem and its relationship with the Fisher information matrix. Even if this is a very old and standard formalism, the way it's introduced in the article puts a high bar on its conference-wide readability
  • The scalar claim in the headline is mistated. In eq. 6 give 1 − v^TCv/λ, which needs |vTCvv^TCv| << λ. The paper’s own values (0.66 to 6.2) make it negative for three of five arms. The restated form in eq. 22 is seems to be much better, and 4.1, 4.4 and 4.6 cite eq. 6 when they mean it, it probably should be 22.
  • The scope and the context are narrow. The theory covers one near-null direction not mentioning what happens with other near-null ones, especially if all are close to the noise floor. This is based on scalar prior precision and a Gaussian prior, and magnitudes that depend on λ.
Questions for authors

My main questions are

  • What apens with the other near-null directions if these exist or are closed to the noise floor and corresponding magnitude? Is there a way to aggregate their contributions for leak analysis ?
  • Can you maybe ease more the reader into the larger and more formal framing of the problem?
3 comments
AuthorsOpenReview ↗

Response to Reviewer aANv: You Were Right About Equation 6; Section 3 Rewritten to Ease the Reader In, and the Scalar Corrected Throughout

Thank you for both points. The readability one was an extremely important comment to us as we value being able to communicate the ideas to a conference-wide audience.

W. The scalar claim in the headline is misstated. In eq. 6 give 1−v⊤Cv/λ1 - v^\top C v/\lambda, which needs ∣v⊤Cv∣≪λ|v^\top C v| \ll \lambda. The paper's own values (0.66 to 6.2) make it negative for three of five arms. The restated form in eq. 22 seems to be much better, and 4.1, 4.4 and 4.6 cite eq. 6 when they mean it.

Your reading is correct and we've made the following changes.

  1. The ladder is reordered. In the revision, equation 6 is the mixture form, equation 7 is the exact rational scalar (your eq. 22), and equation 8 is the first-order form with its condition printed in the display. The abstract now quotes the exact scalar.

  2. You were right about 4.1, 4.4 and 4.6, and in 4.4 the expression we had quoted was the rational form sitting under the wrong equation number. All three are corrected.

Q1. What [happens] with the other near-null directions if these exist or are close[] to the noise floor and corresponding magnitude? Is there a way to aggregate their contributions for leak analysis?

  1. The theory already aggregates them; the cheap repair is what stops at one. Equation 6, the mixture form, keeps every direction, and we validate it against exact linear algebra on the photovoltaic fits at a median relative error of 1.0%.

  2. The rank-one repair degrades predictably as the near-null end becomes a cluster. As a network widens, the Fisher eigenvalues below λ\lambda go from 2 to 470 and the ratio of the two smallest falls from 2.16 to 0.88. The repair recovers 92-99% at codimension one, 31-60% at two, 34-48% at three. The two numbers that say when it applies are the gap ratio and the count of eigenvalues below λ\lambda, and both are now reported.

Q2. Can you maybe ease more the reader into the larger and more formal framing of the problem?

  1. Section 3 now opens with the entire mechanism in one unknown and one measurement, before any vectors appear. For a scalar fit under a Gaussian prior, differentiating the MAP condition with respect to the prior centre gives dx^/dx0=λ/(F+C+λ)d\hat{x}/dx_0 = \lambda/(F + C + \lambda), where CC is the residual times the model's second derivative. Fisher predicts the same expression with C dropped. The two agree at a perfect fit or a linear model and come apart as soon as a nonzero residual multiplies a curved model, worst when F is small and the prior is doing the work. One sentence then says the rest of the paper is this with the parameter as a vector.

  2. We paid for the space by moving the influence-function estimator and its figure to the appendix, so the body keeps the result without growing.

In the revised PDF (changes in blue): the new opening of Section 3, the reordered equations 6 to 8, the corrected Sections 4.1, 4.4 and 4.6, and the abstract.

Your readability comment prompted several large changes in this revision. We hope the new Section 3 opening is easier to follow, and we would genuinely welcome your view on whether it works, since you are the reader it was written for. If it addresses your concern, we would be grateful if you would consider raising your score.

The authors have addressed most of my questions. Still the contribution is quite hard to follow. I will though rise my score.

Thank you for raising your score. Your statement the contribution is still hard to follow is worth knowing, and we will keep working on the presentation for the camera-ready, we appreciate all your feedback thus far!

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Thank You All; Three Implementation Errors Found, All Fixed, and the Main Conclusion Holds

We thank all four reviewers for the care in these reviews. Reviewer 22am ran our released code and audited the implementation line by line, and that audit turned up real errors. We are grateful for it.

Three errors, all now fixed, and the headline conclusion survives all of them.

  1. We tested the wrong matrix in the network admission rule (22am, ysT7). Our well_posed check asked for a positive-definite data Hessian when a MAP minimum only needs a positive definite posterior. Corrected, all 48 configurations pass, and Yacht goes from 0 admitted runs to 3. 22am's reproduction matches ours to eight significant figures.

  2. We led with the wrong scalar (aANv, ysT7). The main text foregrounded the first-order form whose condition our own numbers violate. The exact rational form is now equation 7 in the main text, the first-order form is equation 8 with its condition printed next to it, and the abstract quotes the exact one.

  3. Our ReLU claim was wrong (9Gdw, ysT7). You are both right that a ReLU network is not jointly piecewise linear in its weights. Measured on the Airfoil fit, the curvature norm is 3261.5 for ReLU against 5100.5 for tanh, the same order, not zero.

What changed in the paper. The title is now Fisher information mispredicts prior sensitivity: curvature-correlated residuals, not noise, because the matched control that 22am and ysT7 both asked for showed the effect follows the residual's alignment with curvature, not its cause. The revision adds eight experiments. One claim is narrowed rather than repaired: ysT7's Schur point was right, and a new diagnostic (equation 24) shows our consequence (ii) needs a condition we had not stated.

The revised PDF is uploaded. Every change is marked in blue. The anonymised repository carries all eight experiments.

We hope these revisions address your concerns, and we are glad to run anything further during the discussion period. If you feel we have resolved your points, we would be grateful if you would consider raising your score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.