acceptodds
Under review as a conference paper at ICLR 2027

CertINR: Certified Verification and Compression of Implicit Neural Fields

Abstract

Implicit neural representations (INRs) use neural networks to represent continuous functions over spatial or spatiotemporal coordinates. After compression, the decoded network can differ from the reference in ways that evaluations at sampled coordinates may miss. We introduce CertINR, a framework with two coupled components: a verifier that provides provable upper bounds on the maximum reference–decoded error over the continuous domain, and a certificate-aware compressor that uses these bounds to repair and select compressed representations. The key verification idea is to analyze the reference and decoded networks jointly at the same coordinates, so that variation shared by the two networks cancels before their difference is bounded. We realize this idea through paired Taylor bounds that preserve this cancellation through nonlinear layers and admit refinement-based certification guarantees. Alongside paired Taylor, we develop complementary global spectral and local residual/Jacobian certificates and combine all three in a verification portfolio. On the compression side, CertINR searches a pool of compressed candidates, uses the local error models to repair the final affine layer, verifies the actual decoded streams, and returns the smallest representation that satisfies the target error tolerance. Experiments on analytic and pretrained sinusoidal fields show that no single certificate dominates across network regimes: paired Taylor is strongest on shallow analytic models, whereas spectral bounds are stronger on larger pretrained models. The CertINR bounds also resolve cases left uncertified by general-purpose neural-network verifiers. On six pretrained signed distance fields (SDFs), CertINR achieves 41.21% mean per-shape storage savings at tolerance 0.01. More generally, we show that for SDFs, a certified uniform error preserves inside–outside decisions and, under explicit gradient conditions, can also guarantee topology preservation and bound surface displacement.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
implicit neural representations, neural fields, certified compression, neural network verification, signed distance fields, reliable AI

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_U2K7OpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper "CertINR: Certified Verification and Compression of Implicit Neural Fields" `certify' the error of compressed INRs. Specifically, it computes an upper bound on the maximum absolute error over the entire continuous coordinate domain between the original INR and a compressed INR decoded with the same architecture. The compressed representation is certified when this bound is no greater than a given error tolerance.

The authors additionally propose a correction method for this upper bound, showing that the compressed INR provides a better representation.

Critical strengths
  • The paper addresses a meaningful verification problem. It distinguishes errors measured at sampled coordinates from a uniform error guarantee over the continuous domain. This distinction is relevant to assessing the reliability of compressed INRs.

  • The paired analysis effectively exploits cancellation between the reference and decoded networks. On the small-scale benchmark at an error tolerance of 0.01, paired Taylor certifies 63 of 288 pairs, compared with 36 for the residual/Jacobian bound. This demonstrates a concrete improvement within the evaluated setting.

  • The implementation carefully connects certification to the compressed representation. It accounts for decoding roundoff and re-verifies the stored and decoded network after head repair. Storage is also measured using actual byte counts, including codec metadata.

Critical weaknesses

This paper has several weaknesses. Considering them together, I recommend rejection. They are as follows.

  • The paper does not clearly present a vision for its theoretical contribution.

The validity of the presented formulas should be distinguished from the significance of the theoretical contribution. The paper provides proofs of its bounds and remainder terms. However, it is unclear what additional contribution these proofs offer over existing methods.

Compared with other work on error bounds, such as HiTaB [1], it is difficult to identify the conditions under which the paper’s main methodology guarantees tighter bounds. A theoretical paper proposing such bounds should clearly address the following: 1) How strong are the assumptions? 2) What are the constraints, and what exactly do they affect? 3) How much reduction in the bound is guaranteed compared with other methods?

The paper states its assumptions and analyzes the effects of some constraints. However, it does not sufficiently establish the conditions under which improvements over existing methods are guaranteed or quantify those improvements. From this perspective, the paper’s theoretical contribution is limited.

  • It is quite difficult to regard this paper as practically useful.

In particular, the computation requires two networks and a specified error tolerance. This leads to a simple question: “Why not directly compute the difference between the outputs of the two networks at the required coordinate density?” The paper does not sufficiently answer this question from a practical perspective.

It is also unclear how certification performed during compression translates into benefits during subsequent decoding and inference. Fundamentally, it is difficult to understand the significance, for engineering applications, of guaranteeing an error bound without directly evaluating every coordinate. A concrete application demonstrating this value is needed. Furthermore, the implementation and experiments are limited to scalar-output SIRENs. Practical usefulness for other INR architectures therefore remains unverified.

  • The presentation lacks clarity and focus.

The assumptions are scattered across the paper rather than clearly organized, making the theoretical analysis difficult to follow. The paper also presents multiple claims without clearly identifying the central contribution. These issues obscure its main message.

[1] Entesari, Taha, and Mahyar Fazlyab. "Hierarchical End-to-End Taylor Bounds for Complete Neural Network Verification." arXiv preprint arXiv:2605.10621 (2026).

Questions for authors
  • Could the authors state the paper’s central claim in a single sentence?

  • How would the proposed method apply to INR architectures other than SIREN?

  • Would the authors validate the proposed bound on INRs compressed using existing methods, such as [2]?

[2] Strümpler, Yannick, et al. "Implicit neural representations for image compression." European conference on computer vision. Cham: Springer Nature Switzerland, 2022.

2 comments

Thank you for the constructive review. We have clarified the central claim and strengthened the theoretical and practical evidence while retaining the original framework.

Central claim. CertINR couples continuous-domain verification with certificate-aware compression to repair candidates when beneficial and select the smallest certified representation in a searched pool. We certify the uniform error relative to a fixed reference INR, including at coordinates chosen after compression.

What the theory establishes. The key question is how verification responds to both compression perturbations and spatial refinement. Under the stated local exact-arithmetic assumptions, we derive a residual-remainder bound proportional to parameter mismatch and quadratic in cell radius, so the certificate gap is at most

2Cρ∥r∥12,2C\rho\|r\|_1^2,

where ρ measures parameter mismatch and r is the cell half-width vector. This yields explicit sufficient refinement conditions for candidates whose true error lies strictly below the tolerance. We establish this rate for paired Taylor and a specified fused affine-fixed construction on the sine residual graph; under stated curvature conditions, a specified two-branch affine-fixed verifier retains a positive gap even for identical networks. The result therefore explains a concrete advantage of preserving compression dependence during propagation. It does not order complete methods: in our matched comparisons, affine-fixed and CROWN bounds on the residual graph are generally tighter than paired Taylor. The main assumptions table and Residual-form verification appendix state the conditions, constants, and separate requirement that numerical allowances fit within the available error slack.

For HiTaB, we now give an exact comparison of the first-order components. With common exact center values and gradients and a valid gradient-Lipschitz constant, paired Taylor gives the smaller bound precisely when its layerwise remainder is smaller than the quadratic Lipschitz remainder. This directly answers when an improvement is guaranteed, while keeping the comparison distinct from HiTaB's complete higher-order portfolio. See Relation to HiTaB and architectural scope for the identity and assumptions.

Why certification is useful after compression. Dense sampling answers the error question at the evaluated points. A reusable INR may later be queried at adaptive or off-grid coordinates, which are not fixed during compression. CertINR's uniform certificate covers these subsequent queries without requiring access to the reference network each time.

Both networks are used offline. At deployment, the decoded-only client combines an outward-rounded evaluation with the stored certificate, returns a sign when the resulting reference-value interval excludes zero, and otherwise abstains. On six pretrained shapes, it answers 99.52% of uniform queries and 37.62% of floating-point-guided adaptive-trace queries, with every returned sign agreeing with a conclusive reference interval. The streams and certificate metadata together save 29.66% in pooled bytes using spectral certificates. This demonstrates reuse of a reference-relative guarantee with abstention, not an inference-speed improvement; timing and decoded memory are reported separately in Reference-relative sign-query deployment.

Other architectures and existing compression methods. The construction extends to smooth activations with bounded derivatives on the reachable intervals and a Lipschitz second derivative. We validate this extension on tanh INRs and additionally evaluate RGB fields compressed using COIN author code. These studies extend the evidence beyond scalar SIRENs while making the architectural requirements explicit.

The COIN study verifies the mathematical fields defined by exported FP32/FP16 coefficients. At tolerance 0.1, the added affine-fixed comparison certifies two of three pairs on the finest cover; the original portfolio leaves all three unresolved, and a coarse cover favors paired Taylor. Numerical FP16 execution requires separate rounding analysis. COIN provides an existing-codec integration example, while the full Strümpler et al. pipeline remains outside the completed experiments. Scope of external-codec integration explains its positional-encoding, decoding, and metadata requirements; Supplementary validation with tanh fields and Validation with COIN-compressed RGB fields give the protocols and results.

Finally, the main text now brings together the assumptions, all three original certificate definitions, and principal quantitative results.

Dear Authors,

I appreciate the effort and honesty you have put into preparing the rebuttal in response to my review.

While the rebuttal and revised manuscript have addressed some of my original concerns, the revision itself raises new concerns:

  • The manuscript has been changed with extensive revisions (34 to 70 pages).

  • Given the limited time remaining in the discussion period, I do not believe it is feasible to thoroughly review these substantial changes and provide a fair reassessment.

  • More importantly, the central claim appears to have shifted, and the additional material has made the paper's main contribution even less clear and focused.

I will summarize my remaining concern regarding the rebuttal as follows.

W1/Q1: Central Claim and Theoretical Contribution

The revised central claim emphasizes selecting the smallest certified representation. However, as the revised manuscript itself states, this optimality is only "relative to the searched pool and available certificates." The main method 1) evaluate every candidate in a predefined finite pool and then 2) take the minimum storage among those certified. Once all candidates have been verified, selecting the smallest is a trivial argmin over a finite set. The guarantee of "smallest" therefore adds no theoretical content beyond the verification step itself, so any theoretical novelty must lie in the proposed verifier.

Recommendation

I believe that there is a potentially valuable paper here. However, the paper needs to be rebuilt around what the evidence now supports rather than revised within the discussion period.

For these reasons, I lower my rating and encourage the authors to reorganize and resubmit the work.

Best regards, Reviewer

Reviewer_y9PwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

When a neural field is compressed, its error is usually measured at a few sampled points. This can miss a small region where the compressed network changes sign, which changes the shape for a signed distance field. The paper proposes CertINR, which proves an upper bound on the error between the original and compressed networks at every point in the domain, and then keeps the smallest compressed version whose bound is below a tolerance. The main tool, called paired Taylor, expands both networks around the same point so the parts they share cancel out. The paper adds two simpler bounds, an optional re-fit of the last layer, and a proof that a small enough error keeps the surface topology. The authors report certified savings on six pretrained SDFs.

Critical strengths
  1. The problem is well motivated and considered to be a real gap
  2. The theory appears sound.
  3. The paper is broad in what it checks, covering verification, compression, head repair, and geometry, with comparisons to three general-purpose verifiers.
Critical weaknesses
  1. The authors list the paired Taylor bound as the main contribution, but it looks identical to affine arithmetic (Sharp and Jacobson 2022), which keeps a center, a slope per input symbol and a remainder, cancels shared terms under subtraction, and is already recommended in that paper for neural implicits with a rule for sine. The paper cites that work but does not compare against it. Please provide a clear distinction against this method.
  2. The authors claim the new certificates enable certified compression, but looking at Table 10 and Table 11, the spectral and residual/Jacobian bounds beat paired Taylor on all 48 pretrained pairs and the spectral bound certifies every selected stream, which might indicate the 41% savings come from the standard norm bound rather than from the paired certificate.
  3. As per Table 4, Table 6 and Table 8, head repair and the topology guarantee are not shown to be useful at scale. Repair saves under 6% on nine 353-parameter networks and raises the true error in most of the cases, and the topology check fails on two of three tori and was never run on the pretrained shapes, which limits the practical significance of these two contributions.
  4. The main text carries the claims and theory but almost none of the evidence. Most quantitative results and two of the three certificates are only in the appendix, so the main paper cannot be assessed on its own.
Questions for authors
  1. What does paired Taylor add beyond affine-fixed arithmetic applied to the residual graph with the same cell schedule, on both the analytic and pretrained pairs?
  2. How would the authors show that the compression savings depend on the paired certificate rather than on the spectral bound?
  3. Does head repair reduce storage or true error on any pretrained model, and does the topology check certify any of the six pretrained shapes?
2 comments

Thank you for the detailed review. We have clarified the methodological distinction, added the requested matched comparisons and pretrained checks, and moved the certificate definitions and principal results into the main text.

What paired Taylor adds. We credit Sharp and Jacobson (2022) for the affine representation, shared-symbol cancellation, and sine-envelope rules. Paired Taylor uses exact center values and Jacobians and directly bounds the fused activation difference; affine-fixed propagates activation envelopes with interval-dependent slopes and condensed remainders. These are distinct propagation rules within a shared affine framework.

Our new analysis explains why preserving the compression perturbation through nonlinear propagation matters. Under explicit local exact-arithmetic assumptions, the remainder bound scales with the parameter mismatch and the squared cell radius. We establish this property for paired Taylor and a specified fused affine-fixed construction on the sine residual graph, and prove a separation from a specified two-branch affine-fixed verifier that propagates condensed remainders independently. The separation concerns this specified verifier; it does not extend automatically to other relational verifiers, CROWN, or differential zonotopes. Thus, the theory identifies the conditions that preserve cancellation through the network, while recognizing that another suitably constructed verifier can share this benefit. Residual-form verification gives the assumptions, constants, and distinction between exact-arithmetic rates and numerical enclosures.

What the matched evidence shows. On the same residual graph and 512-cell schedule, affine-fixed gives tighter bounds on all 288 analytic and 46 of 48 pretrained pairs. On the 288 original analytic candidates at tolerance 0.01, paired Taylor certifies 63, fused affine-fixed 76, and the specified two-branch baseline zero. On the matched tanh pool at tolerance 0.001, the reconstructed residual CROWN and fused affine-fixed implementations certify 22 and 20 of 36 candidates, respectively, compared with 18 for paired Taylor. Paired Taylor's complementary advantage appears when verification cost affects the compression search. In the same tanh study, at tolerance 0.03 and one second per reference, it certifies 33 candidates across the nine searches, compared with 27 for affine-fixed and 31 for residual CROWN, and selects fewer pooled bytes. Other budgets and tolerances favor the tighter alternatives. We therefore present paired Taylor as a useful portfolio component whose value depends on the verification budget and candidate regime. Matched affine-fixed comparison and Matched smooth-activation and perturbation comparisons report the full bounds, outcomes, and timing results.

Which certificates produce the storage gains? CertINR selects the smallest candidate certified by its portfolio. The original 41.21% mean per-shape pretrained saving is supported by the spectral component; the revised main text now states this attribution directly. On the analytic pool, adding paired Taylor to spectral + residual/Jacobian reduces selected storage by 4.36% at tolerance 0.01. The affine-fixed-augmented portfolio selects the same bytes with or without paired Taylor. Together, these ablations separate the demonstrated benefit of certified compression from the incremental contribution of each verifier.

Repair and topology at larger scale. Repair targets storage subject to a certified error tolerance. On the analytic models, it reduces mean selected storage by 5.84% relative to adaptive sampled minimax at tolerance 0.02. Sampled reconstruction quality is a separate outcome: paired-head repair has the larger sampled maximum in 11 of 18 equal-byte comparisons, which we report alongside MSE and the certified bounds. Across all 48 pretrained candidates, post-repair verification retains the original streams, leaving storage and reference-relative errors unchanged. This establishes where the current repair procedure provides a demonstrated gain and where the unrepaired candidate remains preferable.

We also ran the topology checks on all six pretrained shapes. The checks remain unresolved within the stated budget, whereas the analytic study certifies 20 of 27 configurations. An unresolved sufficient check leaves topology undecided; it does not invalidate the uniform error certificate or the associated margin-based sign guarantee. The revision now separates these guarantees clearly and identifies the remaining boundary and gradient checks in the detailed geometric evaluation.

Finally, the main paper now defines all three original certificates and presents the principal comparison and attribution tables.

Reviewer_y9PwOpenReview ↗

Reply to the authors

Thank you for the explanation.

The feedback has helped clear some of the questions. However, I agree with reviewer U2K7 that there have been substaintial changes to the original manuscipt and the appendices to review in the short discussion period. Furthermore there is unanimous agreement of the reviewers on rating (1-2). Keeping these in mind I will stay with the original score of 2 for the paper.

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

The study tries find the smallest compressed representation of a searched candidate pool keeping the certified upper bound of the maximum reference-decode error over the domain below a prescribed tolerance. However, it is not ideal to evaluate the absolute compression error only at sample coordinates, since there might be points between samples that give a larger error. Therefore, they formulate the optimisation problem as,

min⁡sR(s)subject toU(s;Ω)≤ϵ\min_{s} R(s) \quad \text{subject to} \quad U(s; \Omega) \le \epsilon

such that the maximum compression error in the domain falls below the upper bound, forcing a candidate to require a guarantee over every coordinate in Ω. The relationship between the reference and the compressed version is preserved by taking the difference between two networks before bounding. One-Lipschitz property of sine is used to show that the difference after activation cannot exceed the difference before activation. To calculate the complementary bounds, they cover the region with spatial boxes,

B={c+y∣∣yi∣≤ri}B = \{ c + y \mid \lvert y_i \rvert \le r_i \}

Global spectral bound controls how each layer can amplify parameter errors using matrix spectral norms, whereas the local residual/Jacobian bound starts with the error at the center and then adds the largest amount by which that error could change within the box. Finally, the paired Taylor bound describes local residuals with central values and central gradients, while accounting for the remaining nonlinear behaviour separately. Then, they take the minimum of these 3 bounds for each box and take the maximum of the cell bounds over the domain. Finally, they optionally perform head-repair to repair the final affine layer while keeping the hidden layers fixed. Sign preservation is introduced for signed distance fields when the reference field magnitude is larger than the certified error bound. Additionally, they derive guarantees for topology preservation and surface displacement under additional gradient, regularity, and boundary conditions. 288 compressed pairs from 9 small networks trained using analytic SDFs are used for validation along with 48 compressed streams from 6 pretrained SDFs. Ablations of paired propagation, comparisons with head repair objectives, GenBaB, CROWN, α-CROWN are presented in the paper. At a tolerance of 0.01, they report a 41.21% mean per-shape storage saving on the pretrained experiments, corresponding to a 30.01% reduction on pooled bytes relative to FP32 copies.

Critical strengths

Introduction of paired Taylor bounds improves certification by analysing the reference and the compressed network together: Cancellation between shared components is preserved avoiding unnecessarily large error bounds. Sufficient experimental evidence is reported involving certifying 63 compressed pairs at tolerance 0.01 and compared with 36 using residual/Jacobian bounds. The contribution is theoretically justified properly and is supported by controlled comparisons.

Head-repair uses error bounds to improve compression: Here, linear programming is used to adjust the output layer keeping compressed hidden layers fixed. The resulting storage reduction is properly reported on 9 analytic models (5.84%) and compared with adaptive sampled minimax at tolerance 0.02. This demonstrates that the certificates guide compression and produce smaller representation that meet the required error tolerance.

Critical weaknesses

The added benefit of paired Taylor bounds and head repair is not demonstrated on the pretrained benchmark: The minimum of the spectral and residual/Jacobian bounds is tighter than the paired Taylor bounds across all 48 pretrained streams. Spectral bounds alone account for every portfolio acceptance at the tested tolerances, and these experiments retain the reference output heads. Also, authors explicitly mention that spectral bounds account for every portfolio acceptance at the tested tolerances. Thus, paired Taylor bounds do not add accepted compressed streams. In addition, head repair is not applied in these pretrained experiments. Even though the pretrained experiments show certified compression, they do not provide evidence for additional certified storage gains from paired Taylor bounds or head repair. Spectral bounds have already certified every accepted stream. Furthermore, the head repair is not evaluated on this benchmark. Therefore the value added from these components for certified storage on larger pretrained networks are not sufficiently justified.

The comparisons with CROWN and α-CROWN do not fully isolate the verification method’s advantage: While CROWN and α-CROWN are analysed by treating the entire domain as one box, paired Taylor subdivides the domain to smaller boxes. Within a large box, since network outputs and internal variations can vary significantly, it might produce conservative upper bounds, while within a small box, due to restricted variation, tighter bounds can be produced. The reported comparison does not fully separate those effects. A matched comparison would make the evidence considerably stronger. While appendix M describes cancellation to CROWN and α-CROWN through residual computation graphs, it still analyses a single box, leaving the spatial subdivision difference unsolved.

Questions for authors

Additional validation: Do the advantages of introducing paired Taylor persist when CROWN and α-CROWN use the residual computation graph and the same spatial partitions on the existing analytic test pairs? Please consider reporting the resulting bounds, certification outcomes, and total runtimes.

1 comment

Thank you for the constructive review. We agree that component-level storage attribution and matched spatial refinement are important for assessing the method. We therefore added a certificate-only storage ablation, evaluated head repair across all six pretrained reference models, and ran CROWN and α-CROWN on the same spatial partitions as paired Taylor.

Attribution of storage savings. The new ablation holds the candidate pools and heads fixed to isolate each certificate's contribution. On the analytic pool at tolerance 0.01, spectral + residual/Jacobian selects 4,955 bytes, and adding paired Taylor reduces this to 4,739 bytes, a 4.36% reduction. For comparison, adding affine-fixed to spectral + residual/Jacobian selects 4,587 bytes; the same selection is retained when paired Taylor also joins that augmented portfolio. These results quantify the storage benefit of the local certificates and distinguish the original portfolio from the additional affine-fixed comparison.

On the pretrained benchmark at tolerance 0.01, spectral verification alone and the original three-certificate portfolio both accept 28 of 48 candidates and select 587,154 bytes. Thus, the reported 41.21% mean per-shape saving, or 30.01% in pooled bytes, is supported by the spectral component. The paired Taylor storage contribution is demonstrated on the analytic pool; pretrained selections remain unchanged when it is included. We make this attribution explicit in the main experimental tables.

Head repair. The analytic study demonstrates how certificate-aware repair can improve the storage objective: at tolerance 0.02, complete paired-head search reduces mean selected storage by 5.84% relative to adaptive sampled minimax. To examine its behavior on larger networks, we additionally evaluate all three repair methods on all 48 compressed bodies from the six pretrained reference models. All 288 LP solves succeed across 144 repair proposals. Verification reverts 143 serialized proposals with larger portfolio bounds, and the remaining proposal is byte-identical to its unrepaired stream. Every method therefore retains the original pretrained streams and their storage and reference-relative errors. This experiment documents how the retention rule operates across the full pretrained pool, while the analytic study supplies the demonstrated storage gain. We report sampled reconstruction errors separately from the certified bounds optimized by repair.

Matched CROWN comparison. To separate verifier tightness from spatial subdivision, we evaluate all nine original analytic comparison pairs at tolerance 0.02 using identical decoded coefficients and covers of 1, 8, 64, and 512 cells. At 512 cells, CROWN and α-CROWN each certify eight pairs, compared with two for standalone paired Taylor; both CROWN variants give tighter bounds on all nine pairs, and common point checks prove the same remaining violation. Median full-process times through all four grids are 2.637 s for CROWN, 68.476 s for α-CROWN, and 0.456 s for paired Taylor. Taken together, the matched results characterize the observed trade-off between bound tightness and verification time, including the lower measured runtime of paired Taylor in these implementations. We distinguish the official solvers' float64 bounds from the outward-rounded certificates and point checks.

We also examine this trade-off beyond sine activations. On the same 36 tanh candidates and shared residual graph, affine-fixed and validated residual CROWN certify 20 and 22 candidates at tolerance 0.001, respectively, compared with 18 for paired Taylor. To complement these comparisons, the separate deadline study identifies settings in which paired Taylor's lower computation cost yields more certificates within the available budget, as well as settings favoring the tighter alternatives. The tanh CROWN results use our newly written validated implementation; α-CROWN belongs to the original SIREN comparison and was not evaluated in the tanh study.

The main experimental tables now present the matched comparisons and component attributions directly. Full protocols and results appear in Certificate-only storage ablation, Pretrained head-repair experiment, Matched spatial partitions, and Matched smooth-activation and perturbation comparisons.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Changes in the Revised Manuscript

We thank the reviewers for their careful reading and constructive suggestions. In this revision, we expand the theoretical analysis, add matched baseline comparisons, and bring the certificate definitions and principal quantitative results into the main text. Throughout these changes, we retain CertINR's original framework: we couple continuous-domain verification with certificate-aware compression, use the resulting bounds to repair candidates when beneficial, and select the smallest certified representation in the searched pool relative to a fixed reference INR. We retain the original spectral + residual/Jacobian + paired Taylor portfolio and separately evaluate an affine-fixed-augmented portfolio and CROWN baselines.

The revision makes three aspects of this framework clearer:

  1. Contribution. We characterize how joint residual propagation retains dependence on the compression perturbation. A new appendix establishes a difference-proportional transfer theorem, verifies its conditions for paired Taylor and a specified fused affine-fixed construction on the sine residual graph, and proves a separation from a specified two-branch affine-fixed verifier. We state the local exact-arithmetic assumptions and distinguish these results from the numerical enclosures used in implementation.
  2. Comparison. We compare verifiers on matched residual graphs and spatial partitions, and add a study with fixed computation budgets. These experiments characterize complementary advantages in bound tightness and verification cost: affine-fixed and CROWN generally produce tighter bounds, while paired Taylor obtains more certificates in some short-budget settings. We report the complete comparisons across tolerances and deadlines.
  3. Component contributions and scope. The storage ablations identify the contribution of each certificate across network regimes. Paired Taylor improves analytic selections over spectral plus residual/Jacobian, while spectral verification supports the original pretrained savings. We also extend repair and topology checks to the pretrained models and add tanh, COIN, and downstream sign-query studies.

We preserve the core verification-and-compression framework, place the detailed proofs and additional experiments in the appendices, and describe the protocols and supporting evidence. The replies below address each reviewer's questions directly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.