acceptodds
Under review as a conference paper at ICLR 2027

Posterior Entrances: Reuse the Information, Preserve the Sampling Law

Abstract

Recovering an image from incomplete measurements requires useful reconstructions under a finite inference budget. Yet generating posterior samples repeatedly purchases learned-prior information at related states. We ask whether this information can be reused without changing the prescribed sampling transition. Posterior Entrances (PE) couples two operations: certify a retained first-order reference under the current context, then correct its proposal with live queries. A history-conditioned rejection identity preserves the ideal block conditional and finite-horizon state law; a movement-based refresh bound and a sufficient break-even condition account for both avoided anchor work and added correction. Across two trained priors and four model–observation configurations, paired-median fresh-FORS/PE ratios at sixteen sweeps are 1.354–1.369 for total queries and 1.135–1.296 for pipeline time. In the representative profile, the complete query ledger shows a net saving of 80.34 queried states per endpoint and 22.85% of complete pipeline time. Site-resolved diagnostics evaluate 512-endpoint estimates. A separate photographic entrance ensemble, E32, demonstrates fixed-budget reconstruction from missing pixels and spatial averages. The method makes reference lifetime a controllable inference resource while preserving the prescribed transition.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Diffusion models, Bayesian inverse problems, Posterior sampling, Denoising means, Total variation bounds, Query complexity

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xQCROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes two methods to accerate samplers : PE for a particular structure of potential and E32 for image restauration. The first proposed method is show to keep the same finite-horizon state law than a reference method (FORS). Then, it is experimentally tested to sample a synthetic potential. The second method is tested for image restoration including inpainting and deblurring.

Critical strengths
  • The proposed method obtains an acceleration around 20% that is very interesting in many practical applications.
Critical weaknesses

Major :

  • The two methods seem unrelated except the context of accelerating samplers.
  • The two methods are very specific. For PE, it solves the problem of sampling from a specific structure of sampler with restrictive constraint on curvature (e.g. S_u needs to be convex and smooth). E32 relies on a very specific methodology : two run of PC, two run of DDRM and 32 steps. These choices are not clearly motivated, explained and challenged.

Minor :

  • line 38 : Please introduce the term of E32.
  • line 47 : The cost of 92.63% is for which example ? In which situation ? It is a very specific number in a specific situation for a general introduction.
  • line 49 : What do you mean exactly by "80.34 queries" ?
  • line 85 : Please introduce the term FORS at the first occurence.
  • line 130 : Which are the dimension of zz and uu ?
  • line 133 : a, b lives in which interval ?
Questions for authors
  • Could you extend your study for PE outside convex setting for SuS_u ?
  • Could you explain the different choice for the E32 method ?
  • What is the conceptual link between PE and E32 ? Could you apply PE for image posterior sampling ?
3 comments
AuthorsOpenReview ↗

Response to Reviewer xQCR

We thank the reviewer for the careful reading and for recognizing the practical value of the acceleration. The revision focuses on PE. E32 is presented separately as an approximate diffusion study in Appendix I; the main paper's PE claims do not depend on it. References below are to the revised manuscript unless marked otherwise.

Can PE extend beyond convex potentials and support image posterior sampling?

The nonlinear potential need not be convex, but the directly sampled block conditional must be strongly log-concave. This remains a substantive restriction. For the queried potential (SuS_u in the original formulation, now ϕu\phi_u), the revised theorem requires a globally LL-Lipschitz gradient, a valid reference-error bound, and H≻(L+γ)IH\succ(L+\gamma)I. Nonzero reference error requires γ>0\gamma>0; zero error permits γ=0\gamma=0 (Theorem 1; Appendix B.2). Gaussian augmentation lets this precision condition apply to an augmented conditional while preserving the original posterior as the marginal target (Appendix D.2). The unaugmented posterior need not be globally log-concave.

For image posterior sampling, Appendix D provides a learned 243–128–64–1 RGB patch model, with coefficient-derived curvature and gradient-transport certificates and block updates up to 243 coordinates. Its tanh nonlinear component is bounded and nonconstant, so it is not globally convex (Eq. D.3). This construction does not certify a pretrained diffusion network. The original digit potential was also learned from data (original Section 4.1).

Section 5.3 reports a separate learned-prior CUDA workload study (Profile A; Appendix G), covering masking, deblurring, 4× super-resolution, and denoising up to 512×512, with three cohorts per task and 128 scenes per cell. With matched burn-in and retained-sample counts among the exact configurations, PE-Full achieves 2.18–2.40× speedups over Fresh FORS and 1.53–1.68× over Fixed + path in complete warm runtime. Fixed + path uses a fixed reference with path bounds, and all twelve adjusted runtime-ratio intervals against it lie below one (Table 3).

For deblurring in cohort A, Figure 3 reports mean times over 256 four-chain bundles:

Time component (seconds)Fresh FORSPE-Full
Gradient evaluations508.47141.59
Bound computation1.4149.56
Complete warm runtime657.73277.13

Complete time also includes proposals and other device and host work. Gradient savings outweigh the added bound-computation cost, giving a 57.9% reduction in complete warm runtime.

How are PE and E32 related, and why this E32 configuration?

Both reuse earlier computation: PE retains gradients, certifies their current error, and uses an exact correction, whereas E32 retains observation-informed diffusion states for approximate continuation. Their guarantees are distinct.

E32 denotes a four-branch ensemble with a common 32-step continuation. Two PC and two DDRM prefixes give both initialization methods equal representation. Each prefix uses 224 evaluations; after noise alignment, every branch uses the same 32-step suffix, with independent randomness and no further observation updates (Figure I.1). The split allocates most computation to incorporating observations and gives each branch the same continuation budget: 4(224+32)=1,024 evaluated states per pool, or 2,048 for the two independently initialized scoring pools.

Presentation and notation.

FORS is now introduced as first-order rejection sampling in the abstract and Section 1. Table 1 compares reference policies, Figure 1 illustrates the procedure, and Appendix A provides notation and a worked example.

The 92.63% figure is the model-evaluation share of Fresh pipeline time for the 8,192-channel digit model on image 1020, whose complete pipeline takes 11.052 seconds per 128-output workload. The 80.34-query saving is 304.20−223.86 mean gradient queries per endpoint, including acquisition and evaluated correction (Section 5.1); its fractional value reflects averaging.

The block variables satisfy z,r,δ∈Rkz,r,\delta\in\mathbb R^k and H,P∈Rk×kH,P\in\mathbb R^{k\times k}. If uu contains only the remaining coordinates of a dd-dimensional state, u∈Rd−ku\in\mathbb R^{d-k}; observations and auxiliary variables may also enter the context. The digit conditional uses k=4k=4 and two neighboring context coordinates. Under the original assumptions, a∈[0,λmin⁡(H))a\in[0,\lambda_{\min}(H)) and b∈[0,∞)b\in[0,\infty), with 0⪯∇2Su⪯aI0\preceq\nabla^2S_u\preceq aI and 0⪯∇2V⪯bI0\preceq\nabla^2V\preceq bI.

I thank the authors for their response. My second main concern on the specific caractere of the proposed method has not been answered. Moreover, they are many modifications of the text which might need a new complete careful reading. Therefore, I keep my negative score and suggest to resubmit.

AuthorsOpenReview ↗

Follow-up Response to Reviewer xQCR

Thank you for clarifying your concern about specificity. We distinguish PE's applicability conditions from E32's exploratory design choices.

PE is defined by certificates, not a particular potential. Theorem 1 requires a globally Lipschitz gradient, a valid reference-error bound, and positive proposal precision—not convexity of the nonlinear residual. Appendix D implements this construction for a trained 243–128–64–1 tanh potential with a nonconvex residual, demonstrating learned image-patch posterior sampling beyond log-cosh. The directly sampled conditional must still be strongly log-concave; Gaussian augmentation can supply the required precision while preserving the original posterior as its marginal.

The observed savings are not confined to one retention setting. Appendix C.4 varies the reuse allowance, proposal slack, and refresh trigger. All seven settings on the tested wide-model target reduce both queries and pipeline time: Fresh-FORS/PE ratios are 1.168–1.426 for queries and 1.068–1.318 for time. Under the theorem's conditions, these choices change computational cost, not the returned conditional law. This provides sensitivity evidence within the tested setting, rather than a claim of universal speedup.

E32 has a design rationale, but its configuration is not established as necessary or optimal. Two PC and two DDRM prefixes give equal representation to the initialization methods; 32 steps specify a shared continuation budget. Available controls compare prefix origins under a common suffix and common versus native suffixes from identical entries. The single-image 16-step control shows no improvement from the common suffix; these controls do not establish an optimal branch allocation or robustness to suffix length. The connection to PE is computational reuse, not a shared correctness guarantee: E32 does not implement or validate exact PE.E32 is therefore retained only as an exploratory study in Appendix I. We are also willing to remove the E32 study entirely to keep the paper focused on PE; doing so would leave PE's method, guarantees, and supporting experiments unchanged.

The evidence addressing specificity is concentrated in Theorem 1 (applicability conditions), Appendix D (a learned construction beyond log-cosh), and Appendix C.4 (parameter sensitivity). The central contribution is certified gradient reuse with measured computational savings, while preserving the prescribed conditional distribution and, under Corollary 2's matched conditions, the finite-horizon law of the outer sampler.

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

This paper asks if prior evaluations can be reused for first order (denoiser/gradient) queries at previous states of image reconstruction posterior sampling, as opposed to doing a fresh evaluation with each change in the conditioning context? The key technique proposed, Posterior Entrances (PE), occurs within a random-scan block Gibbs sampler: a saved reference gradient may be used provided that (i) its deviation from the current gradient is certifiable (bound, ∥Du(r)−D^∥≤E\|D_u(r)-\widehat D\| \le E), (ii) it is used in a Gaussian proposal, and (iii) a two-phase randomized acceptance (with a known-term prefilter, and then the Poisson correction of FORS Chen et al. 2026, broadened to incorporate the allowance EE) yields an exact draw to the current block conditional. (This is expressed by Theorem 1: it guarantees preservation. It yields equivalence in finite-horizon state law to new-reference FORS). Also presented are refresh bound and criterion for net query-saving (Proposition 2).

Experimenting with a restricted 64d log-cosh prior on 8x8 (two width, four model-observation) digits finds 1.35--1.37 fresh-FORS/PE total queries, 1.14--1.30 fresh-FORS/PE in pipeline time. Another photographic experiment (E32) combining PC+DDRM diffusion prefixes with a common 32-step stochastic suffix from a pretrained ImageNet diffusion model finds lower energy scores on 100 images (than several comparators) for a fixed denoiser-state budget.

Critical strengths
  1. Correct, simple way to recycle stale 1st order info without modifying the target. Certified allowance makes it into both acceptance steps, you can use a ref grad from some prior state and the draw you accept is the exact current block conditional regardless of the policy that picked the ref. Small, but a legitimate FORS stretch. The main technical contribution.

  2. Careful cost accounting in the trained study. It is the same conditional engine for the fresh FORS control. So you can isolate the reference policy effect. You charge both anchor and live-correction queries. You surface wall clock pipe time in addition to queries.

Critical weaknesses
  1. The approach is developed and tested only on a toy problem built to satisfy its assumptions. The real photography experiment is not using the approach. PE requires global bound aa on curvature and context constant LDL_D on learned prior. Here they are available only because the prior is a specially restricted 64-dimensional log-cosh model with ring neighbours and parametrisation of that enforces a<0.45a<0.45 which yields strong log-concavity and unimodality of the prior. The real photography experiment, E32, uses a pretrained (non-certified, non-correction, non-exact) diffusion model. It's a different and approximate (diffusion prefixes then common suffix) approach which the paper itself accepts by calling it "operational" (Section 5.3) relationship between them. Therefore the paper does not demonstrate any applicability of PE for practical image prior. And photography gains are not due to PE. This gap between theory-inspired toy and real example is the major component in my rating.

  2. Data in the studies are limited. The improvements in PE are modest. Relative to paper's own fresh reference baseline. query ratio: 1.35 to 1.37, pipeline time ratio: 1.14 to 1.30 (12% to 23% less time) on 64-dim problems with three test images. Comparison to common samplers (MALA, HMC, random-walk Metropolis) is supposed to be "in accompanying file", but not presented. E32: Table 4 shows 5 favourable cells picked from 15 contrast family. In figure 4, DAPS and TGD reconstructions appear to be failed runs (DAPS is uniform brown image). So, the 75%/49% improvements on them are not informative. Only energy score is given (not common restoration PSNR, SSIM, LPIPS). Separation of effect of PC/DDRM prefix mix vs. common suffix is missing.

Questions for authors
  1. Can you get (or at least can we bound) the certificate quantities (aa, LDL_D, GfG_f) of a typical learned prior e.g. a diffusion/score network at a fixed noise level? Otherwise, what needs to be done to use PE outside of the constrained log-cosh setting? Did you attempt to use PE in the photo setting? This is the most important part of the review.

  2. Network is trained to denoise unit-variance Gaussian noise and the prior score is defined as Dϑ(x)−xD_\vartheta(x)-x but at the denoising optimum that's the score of the noise-smoothed data density (i.e. tweedie), not the clean digit density (which would also explain the benign prior). Is this the intended prior?

  3. For E32, please report all 15 contrasts along with the config and budget for each comparator (in particular why are the outputs from DAPS/TGD what they are in Figure 4), and report standard restoration metrics (PSNR, SSIM, LPIPS). Ablations of the prefix mixture with respect to the shared suffix would also be useful.

  4. How does PE in this paper compare to MALA or HMC at equal query cost? equal wall-clock time?

4 comments
AuthorsOpenReview ↗

Response to Reviewer ygvV (1/2)

We thank the reviewer for recognizing the conditional-preservation result and the complete cost accounting. Your central concern is whether PE applies directly to learned image priors and delivers meaningful computational gains. The revised manuscript addresses both questions directly.

PE on learned image priors and larger image workloads.

Yes. In the revised manuscript, we apply PE directly to a learned 243–128–64–1 image-patch potential with coefficient-derived certificates (Appendix D). Separately, Section 5.3 reports CUDA wall-clock measurements under a learned prior for masking, deblurring, 4× super-resolution, and denoising, at resolutions up to 512×512. Each task has three cohorts, including one with context drift, and 128 scenes per cell.

Across all twelve task–cohort cells, PE-Full achieves the following gains:

ComparatorPE-Full resultComparison protocol
Fresh FORS2.18–2.40× speedupComplete warm runtime; matched burn-in and retained lengths.
Fixed + path1.53–1.68× speedupComplete warm runtime against an initial reference with path bounds; matched burn-in and retained lengths.
MALA41.07–1.19× ESS/sWall-clock efficiency under Appendix G.2's normalization and method-specific budgets.

All twelve adjusted runtime-ratio intervals relative to Fixed + path lie below one (Table 3). Figure 3 shows how the savings survive implementation overhead: for deblurring in cohort A, mean gradient time falls from 508.47 to 141.59 seconds, while bound-computation time rises from 1.41 to 49.56 seconds. Mean complete warm time falls from 657.73 to 277.13 seconds, a 57.9% reduction. These are means over 256 four-chain bundles. The reduction in gradient-evaluation time more than offsets the additional cost of computing tighter bounds.

How are the learned-model certificates obtained?

The patch model provides a concrete construction beyond the constrained log-cosh family. Write its two-hidden-layer tanh potential as ψ(x)=f(W1x+b1)\psi(x)=f(W_1x+b_1). The network coefficients give the directional curvature bound ∣v⊤∇2ψ(x)v∣≤Lhid∥W1v∥2.|v^\top\nabla^2\psi(x)v|\le L_{\mathrm{hid}}\|W_1v\|^2. For block BB, let s=W1x+b1s=W_1x+b_1 and DB(s)=W1,B⊤∇f(s)D_B(s)=W_{1,B}^\top\nabla f(s). A stored preactivation sjs_j supplies the gradient-transport bound

\le L_{\mathrm{hid}}\|s-s_j\|\,\|W_{1,B}\delta\|.$$ Equations (D.4)–(D.6) specify the coefficient-derived constant, transport bound, and positive proposal precision. Gaussian augmentation preserves the target marginal while supplying the required conditional precision. At the specified reference points on six 9×9 RGB patches from three photographs, the directional matrix bound reduces the expected full-block curvature contribution to the Poisson count from **66.3–78.1** under the scalar majorant to **0.257–0.303** (Table D.2). On the specified patch, blocks of 27, 81, and 243 coordinates yield **69.9, 150.3, and 291.5 minimum-observable ESS/s**, respectively (Table D.3). Appendix D establishes computable certificates and usable block sizes; the separate CUDA study evaluates complete workload cost. Appendix H adds matched local-energy image comparisons with **55.0–56.7% fewer field queries** (Figure H.3). **What would be required for a pretrained diffusion or score network?** PE requires a specified potential, certified curvature and transport bounds for its actual gradient field, and positive proposal precision. The nonlinear component need not be convex. The directly sampled conditionals must satisfy the precision condition, which may be obtained through Gaussian augmentation; this does not require the original unaugmented posterior to be globally log-concave. For the exact smoothed density of $X_0+\sigma Z$, with $\sigma>0$ and independent standard Gaussian $Z$, $$\nabla^2\log p_\sigma(x) =\sigma^{-4}\operatorname{Cov}(X_0\mid X_0+\sigma Z=x)-\sigma^{-2}I.$$ If the clean-data support lies in a ball of radius $R$, the Hessian norm is bounded by $\sigma^{-2}+R^2\sigma^{-4}$, and the cross-block derivative norm by $R^2\sigma^{-4}$. These provide spatial curvature and transport bounds at a fixed noise level. Their growth at low noise may require strong augmentation and affect outer-chain mixing. An approximate pretrained score network needs certificates for its own field; the ideal-density identity alone does not provide them.
AuthorsOpenReview ↗

Response to Reviewer ygvV (2/2)

Is the noise-smoothed density the intended prior?

Yes. At the unrestricted population optimum, the unit-noise denoising objective gives D∗(y)−y=∇log⁡p1(y)D^*(y)-y=\nabla\log p_1(y) for the noise-smoothed data density p1p_1. The finite, constrained network defines the fitted prior pϑ(x)∝exp⁡{−∥x∥2/2+Sϑ(x)},p_\vartheta(x)\propto\exp\{-\|x\|^2/2+S_\vartheta(x)\}, which is the prior used with the observation model in Appendix C. As you note, L≈0.335<1L\approx0.335<1 makes it strongly log-concave and unimodal. We use this controlled benchmark to isolate the computational effect of reuse. PE preserves the exact block conditionals of the fitted posterior; it does not require the network to attain the unrestricted denoising optimum.

MALA/HMC comparisons at equal query cost or wall-clock time.

The submitted supplementary material reports results for all four digit configurations and eight samplers. The MALA and HMC implementations use the same first-order, mean-only interface. The table reports wall-clock seconds to the first saved checkpoint whose 97.5th bootstrap percentile of standardized site-moment error is at most 0.10, measured against independently computed quadrature references. Each estimate combines 512 endpoints; runtime is summed over four 128-endpoint pipeline runs.

Channels / imageFresh FORSPEMALAHMC
8,192 / 10209.8778.32423.11048.992
8,192 / 15989.8148.93025.33524.120
8,192 / 369.83710.17223.36635.524
1,024 / 10201.8141.6343.6317.053

PE reaches this diagnostic threshold sooner than MALA and HMC in all four configurations. The submitted supplementary material also includes results for Static, IMH, RWM, and energy-enabled RWM, along with all saved thresholds and noncrossing cases. The comparison quantifies the cost of reaching a common estimation diagnostic at saved checkpoints; it does not compare terminal error under identical hard query or wall-clock caps.

For the image workloads, the 7.0–19.0% higher ESS/s than MALA4 is a wall-clock efficiency result. The 7.6–16.3% lower complete warm runtime uses method-specific retained budgets; MALA1 has a larger retained budget than MALA4, explaining its longer total runtime. Corollary 2 gives the same outer-trajectory law for exact implementations under matched target, initialization, augmentation, and update schedules. With the same observable and recording rule, their theoretical ESS per retained draw is identical. Appendix G.2 uses a common scene–bundle normalization for exact implementations and separate reported MH/MALA statistics.

E32 comparisons, reconstruction quality, and ablations.

E32 is an approximate diffusion-state continuation study, presented separately in Appendix I. Table I.1 distinguishes five primary comparisons on 100 paired photographs from ten complementary comparisons under another protocol, including results favoring DDRM and TGD for random missingness.

E32 uses two PC and two DDRM prefixes of 224 evaluations, followed by a common 32-step suffix: 1,024 evaluated states per four-output pool, or 2,048 for the two scoring pools.

The E32 study evaluates photographic reconstruction using energy scores, assessing the quality of the generated reconstruction distributions. Its computational budget explicitly includes both prefix construction and continuation, making the cost of the complete E32 pipeline transparent. These results provide complementary evidence on the performance of the full reconstruction procedure. The central contribution of PE is supported directly by computable certificates for a learned image potential, preservation of the target conditional law, and measured runtime savings across the evaluated image workloads.

AuthorsOpenReview ↗

Additional Response to Reviewer ygvV

We report post-hoc analyses of existing runs, weights, and outputs.

Learned-model certificates.

The v13 image model combines a Gaussian quadratic potential with a 9×99\times9 RGB neural-patch residual. Recomputation from saved weights gives residual-curvature bounds of 0.296266 for 12-coordinate blocks and 0.908151 for 48-coordinate blocks, on both 16×1616\times16 and 32×3232\times32 layouts. Appendix D separately demonstrates PE with certified transport and positive proposal precision. These establish computable certificates beyond log-cosh, not for arbitrary pretrained diffusion networks.

MALA/HMC and common-budget accuracy.

Both baselines randomly update four-coordinate blocks of the 64-dimensional target. MALA uses a preconditioned Gaussian proposal with forward/reverse proposal-density correction; HMC uses four leapfrog steps and includes kinetic-energy changes in acceptance. Both compute potential differences for acceptance via conditional-mean queries. Step sizes are selected on two independent tuning seeds using jump distance per unit time. Across 32 jobs, target-aggregated acceptance rates are 46.17–46.61% for MALA and 91.04–95.42% for HMC.

At a common cap of 32,768 conditional-mean state evaluations, we use each method's last saved checkpoint within the cap, pooling 512 endpoints across four seeds. Standardized RMS error covers first and second moments of all 64 coordinates. Across four configurations, PE-CACHE errors are 0.0434–0.0456, versus 0.0499–0.0556 for Fresh, 0.1836–0.2241 for MALA, and 0.1921–0.2334 for HMC. PE has the lowest observed error in all four configurations. Actual query consumption can differ below the cap. Costs were recomputed; error estimates and intervals are unchanged from the archive. Earlier timings measure time-to-diagnostic.

Original E32 contrasts and restoration metrics.

Reanalysis of 1,800 records (100 photographs × three operators × six methods) recovers all 18 method–operator means and the original 15 paired contrasts without pooling protocols. Under 15-test Holm correction, E32 is significantly better in 9 comparisons, significantly worse in 2, with no significant difference in 4. Five meet the preregistered composite criterion, distinct from significance alone.

The 5,600 saved PNGs cover five original contrasts. PSNR was recomputed from pixels; SSIM and LPIPS were reaggregated from per-image records. For random 70% missingness, E32-minus-PC per-output mean differences are +2.388 dB PSNR, +0.1143 SSIM, and −0.1662 LPIPS. These descriptive saved-subset results use 100 paired source photographs, not independent PNGs.

DAPS/TGD reconstruction diagnostics.

For DAPS under 4×44\times4 averaging, observations constrain block means, not within-block detail. Analysis of the original step sizes and schedule shows that 100 inner updates per stage reduce conditional-mean error in block-mean directions by only 0.92–3.08%. This identifies undercorrection of observed block means as a concrete cause of the original coarse-scale distortion.

On image 49109, uniformly increasing the step size reduces block-mean RMS error from 0.535 to 0.114, but increases RMS error in unobserved within-block directions from 0.137 to 0.556, producing severe color noise. Preconditioning retains stronger block-mean correction while restoring the original within-block step size and noise scale, achieving 22.036 dB on the fixed first output. The diagnosis concerns two configuration-specific failure modes, not an intrinsic limitation of DAPS.

For TGD, disabling stage resampling worsens all four checked averaging cases (mean energy-score change +0.009047). For random 70% missingness, the mean change is −0.000353, with two of four cases worsening. These quantify score effects but do not explain a specific visual artifact.

Prefix and suffix controls.

With the same 32-step suffix, 1,600 outputs from 100 photographs and two degradations give PC-origin/DDRM-origin PSNRs of 18.862/23.641 dB for random missingness and 22.395/23.498 dB for 4×44\times4 averaging. This compares prefix origins, not the benefit of mixing them.

On the image, the same four initial states and 960 total state evaluations give common/native 16-step suffix PSNRs of 21.169/21.191 dB and LPIPS of 0.41563/0.41400. The common suffix is slightly worse on this image. These controls neither constitute a full cross-image factorial ablation nor establish compute parity across E32 comparators.

E32 remains a separate approximate reconstruction procedure. PE's theory guarantees preservation of the block-conditional law and, under Corollary 2's matched conditions, the outer sampling law. Its certificates and common-cap comparisons support certified gradient reuse to reduce computation.

I thank the authors for their detailed response to my review and the revision made to the paper in response which has improved it. In particular, the certified patch network in Appendix D is a good extension of the experiments beyond the 8x8 digital model. I will raise my rating of the paper. However, my main concern about whether the method works on a practical learned image prior is only partly addressed. The new patch network experiment shows that it is possible to calculate the bounds for a small learned network but without any baselines for comparison. Furthermore, GPU image study doesn't state which priors it used.

Reviewer_SEc6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper treats evaluated first-order information as a reusable sampling resource. Posterior Entrances uses a certificate to decide whether to retain a reference, current-context correction that preserves the ideal conditional transition, and a movement-triggered retention policy with a net-query benefit condition that charges live correction. E32 extends the entrance-to-continuation design to photographic reconstruction by handing observation-informed native diffusion states to a common stochastic suffix under a fixed budget.

Validation uses a trained constrained model on optical digits at two widths, reporting lower paired-median query and pipeline time, with a representative 80.34-query and 22.85% time reduction, plus sitewise estimation and retention-policy sensitivity. The photographic study uses 100 ImageNet images with two degradation settings and reports improved energy scores over selected comparators.

Critical strengths

The certificate-based reuse mechanism is a notable strength: it provides a preservation result for the ideal conditional transition under stated assumptions and a net-query benefit condition that explicitly charges live correction and refresh work. This is consequential because it supports assessing whether retained first-order information pays under a complete accounting, rather than assuming reuse is beneficial. The matched trained-model evaluation is also relevant: it connects reference acquisition, live correction, pipeline time, and sitewise estimation quality, and the photographic study explicitly charges both entrance construction and continuation under a fixed budget. This is consequential because it provides operationally grounded evidence for the claimed reuse–correction trade-off.

Critical weaknesses

The scope of experimental validation is relatively narrow: the trained study uses a 1,797-image 8×8 optical-digit collection with 297 test images, and the reported configurations rely on a small number of held-out images (1020, 1598, and 36; the 1,024-channel model uses only 1020). This limits how strongly the query and pipeline-time savings support broader conditional geometries.

The photographic validation uses 100 ImageNet validation photographs with simulated degradation operators (random 70% missing observations and 4×4 averaging), not real degraded captures; the main text reports only five comparison cells satisfying a composite improvement criterion, with broader comparison context placed in a companion record. This limits independent assessment of the full comparison and of behavior under real acquisition conditions.

Questions for authors

What is PE’s precise increment over fresh FORS under the same engine—retention certificate and movement-triggered refresh only, or also accuracy/step-complexity gains—given the 64-coordinate trained setting? How does E32 differ from DAPS, TGD, and Spin in noise scheduling and data-consistency placement, and do the gains come from entrance construction, the common suffix, or budget accounting? What deployment setting is intended, and does the observed query/time reduction extend to repeated conditional draws under fixed or slowly changing observations? Future-work suggestion: Test retention and E32 on real degraded captures and beyond one constrained architecture and narrow held-out set.

1 comment
AuthorsOpenReview ↗

Response to Reviewer SEc6

We thank the reviewer for recognizing the value of preserving the conditional sampling law while accounting for acquisition, correction, and refresh costs.

What is PE's precise contribution beyond fresh FORS?

PE reduces the cost of conditional updates without changing their sampling law. The original paper introduced certified gradient retention, movement-triggered refresh, and complete query accounting. Theorem 1 and Corollary 2 establish preservation under matched initialization, augmentation, and update schedules. The comparison with Fresh FORS therefore isolates the computational benefit of reuse: PE lowers the cost of completing the same number of updates while preserving their joint law under the matched protocol.

The revision adds correction-work monotonicity and context-coverage results. Proposition 4 shows that intersecting valid directional intervals preserves the acceptance probability and cannot increase expected evaluated corrections at a fixed proposal, including screening and early stopping. Holding the proposal and reference bank fixed, the controlled experiment reduces corrections from 0.6855 to 0.5972 per draw, or 12.9% (Section 5.2). Proposition 7 explains the additional acquisition savings from revisiting covered contexts.

Does reuse help with repeated draws under fixed or slowly changing observations?

PE targets settings in which gradient evaluations are expensive relative to computing valid transport bounds. Even when observations remain fixed, updates to neighboring blocks change the conditioning context. Slowly changing observations can also be accommodated when the updated conditional satisfies the certificate and proposal conditions, with the known terms and reference-error bounds updated accordingly.

The controlled adaptive-context test in Table F.1 reports total queries over 4,096 calls, averaged over four seeds:

Context sequenceFresh queriesBest-of-four queriesReduction
Nearby contexts5,552.501,713.0069.1%
Four recurring, separated contexts5,560.251,468.2573.6%

In the nearby-context experiment, reuse avoids 4,094 gradient acquisitions at the cost of 254.5 additional corrections; 53.9% of calls reuse a reference with nonzero certified error. In the separated case, four acquisitions cover four recurring contexts with zero-error certificates. These results show that reuse remains beneficial with nonzero certified error and avoids repeated acquisition when contexts recur. Proposition 7 formalizes coverage under its fixed-anchor and storage conditions.

A separate twelve-seed controlled study provides direct evidence of wall-clock savings. The best-of-four/fixed-reference runtime ratio is 0.561 [0.500, 0.629] on recurring contexts; the best-of-four/single-reference ratio is 0.692 [0.586, 0.817] on rapidly changing contexts. The intervals are adjusted across 112 comparisons (Section 5.2; Appendix E).

Broader architectures and acquisition settings.

Appendix D derives certificates from the coefficients of a learned 243–128–64–1 patch potential, evaluates six 9×9 RGB patches, and supports blocks up to 243 coordinates. The separate CUDA study covers four image restoration tasks, three cohorts per task, and 128 scenes per cell. With matched retained-sample and burn-in budgets, PE-Full achieves 2.18–2.40× speedups over Fresh and 1.53–1.68× over Fixed + path in complete warm runtime (Table 3). The gains hold against both fresh gradient acquisition and a fixed reference equipped with path bounds.

E32's schedule, data consistency, and source of gains.

E32 incorporates observations through two PC and two DDRM prefixes, each using 224 evaluations. After noise alignment, a shared 32-step stochastic suffix follows a geometric schedule ending at 0.01, without further observation updates. The cost is 1,024 evaluated states per four-output pool, or 2,048 for the two scoring pools (Appendix I).

The local visual runs use measured-subspace mobility for DAPS, operator-aware covariance for TGD's averaging updates, and paper-aligned schedules with finite-gradient averaging updates for Spin. Table I.1 separates five primary and ten complementary comparisons, including outcomes favoring DDRM and TGD. Without component ablations, these results compare complete configurations; they do not isolate prefix, mixture, or suffix effects.

Table 1 and Figure 1 clarify the reference policies and sampling sequence; Appendix A provides notation and a worked example.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank the reviewers for recognizing the preservation of the conditional sampling law and our complete cost accounting (SEc6 and ygvV), as well as the practical relevance of the acceleration (xQCR). The central questions concern the scope of certification, evidence beyond the digit model, and the relationship between PE and E32. The revision focuses the main paper on PE and addresses these questions through additional theoretical results and direct PE experiments. E32 is presented separately in Appendix I. References below are to the revised manuscript unless the original submission is explicitly identified.

PE reuses previously evaluated gradients across changing conditionals while preserving the target conditional law. The revision strengthens this contribution beyond single-reference retention: Proposition 4 proves that intersecting valid directional bounds preserves the acceptance probability and cannot increase the expected number of evaluated correction queries at a fixed proposal, including screening and early stopping. The corresponding controlled experiment reduces evaluated corrections by 12.9% (Section 5.2). Adaptive-context experiments demonstrate 69.1% and 73.6% fewer total queries for nearby and recurring contexts, respectively, accounting for both gradient acquisition and correction (Appendix F).

To demonstrate applicability to learned priors, Appendix D derives curvature and gradient-transport certificates from the coefficients of a trained 243–128–64–1 image-patch potential and evaluates block updates of up to 243 coordinates. The nonlinear component need not be convex. The positive-precision requirement applies to the directly sampled block conditional, and Gaussian augmentation can supply that precision while preserving the original target marginal. These conditionals remain strongly log-concave; the construction does not certify the pretrained ADM network used by E32.

A separate CUDA study evaluates four image restoration tasks, three cohorts per task, and 128 scenes per cell. With matched burn-in and retained-sample budgets for the exact controls, PE-Full achieves 2.18–2.40× speedups over Fresh FORS and 1.53–1.68× over Fixed + path, which retains an initial reference and uses path bounds. These complete warm-runtime measurements include the cost of computing and maintaining the bounds. Table 3 also reports ESS/s ratios of 1.07–1.19 versus MALA4 under the normalization in Appendix G.2. Together, these results show that PE preserves the sampling law while reducing measured runtime on the evaluated image workloads.

The submitted supplementary material includes comparisons across all four digit configurations and eight samplers, including MALA and HMC. We summarize the key results in the individual responses. To improve readability, Sections 2–4 now progress from the conditional target to the sampling construction and cost analysis; Table 1 compares reference policies, Figure 1 illustrates the algorithm, and Appendix A provides notation and a worked example.

AuthorsOpenReview ↗

Follow-up on Our Response

We would be grateful for your thoughts on whether our response has addressed your concerns. We would be happy to discuss any remaining questions and provide further clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.