acceptodds
Under review as a conference paper at ICLR 2027

Learning High-Order Equilibrium Dynamics with Sinkhorn-Constrained Heterogeneous Priors

Abstract

Deep equilibrium models (DEMs) provide an effective framework for solving inverse problems by modeling the reconstruction process as finding the fixed point of a learned operator. However, they are fundamentally limited by the dilemma between expressivity and stability, i.e., incorporating more expressive or heterogeneous priors violating the contractivity required for convergence. The implicit equilibrium solver could suffer from unstable solving dynamics or degenerated solutions by simply integrating multiple priors that represent complementary image statistics, even well captured by varying neural architectures. To address this challenge, we propose a manifold-constrained hyper-connected deep equilibrium model (MHC-DEM) that achieves multi-prior fusion with a structured transport process. Specifically, we construct a learnable interaction matrix via Sinkhorn normalization that lies on the Birkhoff polytope to project a query distribution into convex fusion weights, and consequently, enforce doubly-stochastic constraints on prior interactions to simultaneously ensure Lipschitz constraints with bounded Lipschitz constants and prevent collapse to a single dominant prior. Furthermore, we design a high-order equilibrium solver based on the stable multi-prior fusion mechanism. We introduce an inertial second-order update in an augmented state space to accelerate convergence, and embed a short inner optimization trajectory within each equilibrium step to perform multi-stage spatial refinement with non-shared parameters. We demonstrate in theory that the proposed multi-prior fusion preserves the contractivity property of neural network based operator represented by MHC-DEM. Extensive experiments show that MHC-DEM achieves consistent gains over existing methods across diverse inverse problems, including image restoration and interferometric imaging, with stable convergence guarantees.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Deep Equilibrium Models, Multi-Prior Fusion, Manifold-Constrained Optimization, Image Restoration

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_wPHMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes incorporating multiple priors into an iterative solver, specifically deep equilibrium models (DEM). The first modification is to add momentum to the outer layer of deep equilibrium network dynamics, and replace the inner layer with a standard multi-step iterative solver. The second modification is within the iterative solver: to combine multiple network residuals together in a Lipschitz-preserving way, using doubly stochastic matrices as a learnable update. Such a method allows for more expressive residuals while (theoretically) preserving convergence of DEM methods, under standard assumptions on the residuals. Experiments are reasonable.

Critical strengths
  • Interesting and effective method to combine multiple priors into the DEM framework. Clear motivation as to why to do so, and procedure is sufficiently clearly detailed in the main text and appendix.
Critical weaknesses
  • The motivation of the doubly stochastic matrices is entirely unclear, my understanding is that it simply works in practice. There is first a learnable logit matrix LL, and then another learnable query vector qq (on the simplex, which is advocated against later in the page!), then the so-called ``fusion weight'' (i.e. weights to combine the residuals) is given by ΠBNLq\Pi_{B_N} L q where ΠBN\Pi_{B_N} is projection onto the space of doubly stochastic matrices via Sinkhorn. Why split the training?
    • The paragraph on "anti-collapse" is not convincing, neither is the claim in Appendix B.2.
    • In Appendix E.4.4, since the relative importance of the softmax and Sinkhorn methods are similar, surely this could also be interpreted as softmax training much faster than Sinkhorn? If you were to continue training, would the Sinkhorn fusion weights tend towards collapse? i.e. how sensitive are these weights to changes in learning rate?
  • Black hole experiment seems a bit unfair. Any comparison against DEM methods which allows for this additional adaptation after the model? What model is used for PnP-DM?
    • Similar concerns for the imaging problems: are the retrained networks (Appendix C) also applied for the compared baselines?
Questions for authors
  • What exactly is being trained? Are the prior models themselves (e.g. DnCNN) being finetuned between the iterations as well, or is it just the stepsize/momentum/logit matrix? Please add some clarification in the main text instead of deferring to appendix.
  • Please add a sentence in Section 3.1 saying that the theory is first presented when FinnerF_{inner} is one step, then for TT step.
  • Should reference and clarify the differences with [CRDP26] which is also DEQ with inertia (in the inner solver).
  • References should be in brackets (maybe \citep?), also there are several instances of "Eq. Equation (xx)".

[CRDP26] Clerc, A., Renaud, M., De Seneville, B. D., & Papadakis, N. (2026). i-DEQ: A stable inertial deep equilibrium model for image restoration. arXiv preprint arXiv:2605.19705.

3 comments

We thank the reviewer for the insightful comments.

Q1. Motivation of the Doubly-Stochastic Fusion Mechanism

We thank the reviewer for questioning the motivation and optimization behavior of the doubly-stochastic fusion mechanism. We clarify that the two learnable components are jointly optimized end-to-end, rather than trained in separate stages. Specifically, the fusion mechanism is defined as

\mathbf{q}=\operatorname{Softmax}(\mathbf{v}),\quad \mathbf{W}=\operatorname{Sinkhorn}\left(\exp(\mathbf{L}/\tau)\right),$$ where $\mathbf{q}$ represents the base preference over heterogeneous priors, while $\mathbf{W}$ models their structured interactions. The factorization is not intended to increase the representational capacity of the final fusion vector: a direct simplex-valued parameterization can also represent any valid fusion vector. Instead, it introduces a structured parameterization that separates the base preference from the interaction-induced redistribution of prior contributions. This provides an additional inductive bias on how the prior weights are formed. Importantly, the doubly-stochastic constraint has a precise mathematical implication beyond preserving the simplex. Since $\mathbf{W}$ is doubly stochastic, the fused distribution satisfies $$\mathbf{w}=\mathbf{W}\mathbf{q}\prec\mathbf{q},$$ where $\prec$ denotes majorization. Consequently, $$\max_i w_i\leq\max_j q_j,\qquad H(\mathbf{w})\geq H(\mathbf{q}),$$ where $H(\cdot)$ is the Shannon entropy. Thus, the doubly-stochastic transformation cannot make the fused weights more concentrated than the base preference distribution. In contrast, column-stochasticity alone only preserves the sum of the weights and does not ensure this property. For example, a column-stochastic matrix can map a uniform distribution to a one-hot vector. We emphasize, however, that this property is a relative concentration constraint, not a guarantee that the final weights are uniformly distributed or immune to asymptotic collapse: if $\mathbf{q}$ itself approaches a simplex vertex, the fused distribution may also become concentrated. To examine the contribution of this structured constraint, we compare four fusion strategies under identical experimental settings: **Table R5. Comparison of reconstruction performance and weight concentration under different prior fusion strategies.** | Fusion Strategy | PSNR (dB) | Max Weight | | :------------------------ | :------------------: | :--------: | | Softmax (Query-only) | $30.96 \pm 0.16$ | 0.97 | | Column-stochastic + Query | $31.22 \pm 0.35$ | 0.85 | | Uniform Weight | $31.45 \pm 0.10$ | 0.34 | | Sinkhorn (Ours) | **$32.91 \pm 0.32$** | 0.51 | Sinkhorn fusion outperforms Softmax, column-stochastic, and uniform fusion by 1.95, 1.69, and 1.46 dB, respectively. Notably, uniform fusion produces a more balanced weight distribution than Sinkhorn but achieves lower reconstruction performance. This indicates that preventing weight concentration alone does not explain the performance gain. Rather, the results suggest that the structured, learnable interaction matrix enables more effective coordination among heterogeneous priors, while its doubly-stochastic constraint limits excessive concentration in the resulting fusion weights. We have therefore revised the corresponding statements in Section 2.3 and Appendix B.2 to distinguish this mathematical concentration property from the empirical observation of balanced prior contributions.

We further track the fusion weights throughout training to investigate whether the difference between Softmax and Sinkhorn is merely due to slower convergence: Both strategies preserve the same ranking of prior contributions, with DnCNN receiving the largest weight. However, their weight distributions evolve differently: Softmax progressively concentrates almost all the weight on DnCNN, reaching approximately 98%, whereas Sinkhorn maintains a substantially more distributed allocation throughout the extended training. The similar ranking therefore does not imply equivalent fusion behavior, nor does it establish that Sinkhorn merely converges more slowly. Instead, the two parameterizations lead to markedly different concentration patterns.

Table R6. Evolution of prior fusion weights during training for Softmax and Sinkhorn-based fusion.

EpochSoftmax DnCNNSoftmax DRUNetSoftmax RestormerSinkhorn DnCNNSinkhorn DRUNetSinkhorn Restormer
133.533.233.333.433.333.3
5094.54.21.350.831.018.2
10097.22.40.451.230.917.9
20097.81.90.351.931.216.9
30097.71.90.451.730.617.7

We also examine the sensitivity to the learning rate by varying the base learning rate while keeping the remaining settings unchanged: Across the tested learning-rate range, the maximum Sinkhorn weight remains around 0.51--0.52, and the reconstruction performance varies only slightly. In comparison, Softmax consistently exhibits a highly concentrated distribution, with its maximum weight remaining around 0.96--0.97. These results indicate that the observed difference is not specific to the default learning rate. Learning rates beyond the tested range lead to training-loss divergence. Overall, the additional analyses clarify that the role of the doubly-stochastic constraint is not to guarantee uniform prior usage. It imposes a principled majorization constraint on the fused distribution, while the learnable interaction matrix provides a structured mechanism for adapting the relative contributions of heterogeneous priors. The improved reconstruction performance is therefore associated with the combination of structured interactions and constrained weight redistribution, rather than weight balancing alone.

Table R7. Sensitivity of fusion weights and reconstruction performance to different learning rates.

Learning RatePSNR (dB)Softmax Max WeightSinkhorn Max Weight
0.1×32.94±0.290.960.52
0.5×32.85±0.300.970.52
1×32.91±0.320.970.51
2×32.80±0.210.970.51

Q2. Experimental Fairness

Black-hole imaging. To address the concern about additional prior adaptation, we introduce DE-Prox and DenseDiff as additional equilibrium-based baselines. Both use exactly the same pretrained denoisers and follow the same preprocessing and experimental protocols as MHC-DEM. The results on synthetic M87 data are:

Table R8. Quantitative comparison of black-hole image reconstruction on synthetic M87 data.

MethodMSE (×10⁻⁴)Corr.
DE-Prox1.890.974
DenseDiff1.680.978
PnP-DM3.520.936
MHC-DEM1.320.996

MHC-DEM achieves the best MSE and correlation, demonstrating that its improvement is not merely due to the use of pretrained denoisers or additional prior adaptation.

Regarding PnP-DM, we follow its original configuration. It employs an EDM-parameterized diffusion denoiser trained via denoising score matching, which differs fundamentally from the pretrained discriminative restoration networks used in MHC-DEM. Replacing its denoiser with our prior networks would alter its original formulation; therefore, we retain the original configuration.

Image deblurring. For PnP-based methods, including GS-PnP, DE-Prox, ELDER, and ProxPnP-DRS, we replace their denoisers with the same pretrained DRUNet used in MHC-DEM, ensuring a consistent prior-network basis. For the other baselines, including DPS, DiffIR, DAPS, and DenseDiff, we use the same training data as MHC-DEM while retaining the original configurations and hyperparameters reported in their respective papers. These experimental details, together with the additional M87 results, will be included in the revised manuscript.

Q3. Training and Learnable Parameters

The default heterogeneous prior bank consists of three pretrained denoisers: DnCNN, DRUNet, and Restormer. These denoisers are first trained independently on image restoration datasets and then kept frozen throughout MHC-DEM training, without updating their parameters during either the inner refinement steps or the outer equilibrium iterations. SwinIR is additionally considered in the ablation experiments.

The learnable parameters of MHC-DEM include: (i) non-shared inner-stage step sizes ηt\eta_t for t=0,...,T−1t = 0, ..., T-1; (ii) non-shared interaction logits LtL_t for t=0,...,T−1t = 0, ..., T-1, which determine the structured prior interactions through Sinkhorn fusion; and (iii) non-shared query logits vtv_t for t=0,...,T−1t = 0, ..., T-1, from which the stage-specific query distributions are obtained as qt=Softmax(vt)q_t = Softmax(v_t). These parameters are jointly optimized using the equilibrium training objective and implicit differentiation. In addition, the temporal momentum coefficient is parameterized as a shared learnable scalar, β=Sigmoid(b)\beta = Sigmoid(b). Since the inertial term vanishes at equilibrium, the equilibrium solution is independent of β\beta, and thus implicit differentiation does not provide gradients for this parameter. Instead, bb is optimized using an auxiliary finite-trajectory objective. This separation preserves the stationary fixed-point formulation required by DEQ training while allowing the momentum coefficient to be learned.

We will clarify the initialization, parameter-sharing strategy, and optimization procedures of these components in the main text.

Q4. Differences from i-DEQ [CRDP26]

We thank the reviewer for pointing out the relevance of the concurrent work i-DEQ [CRDP26]. Both i-DEQ and MHC-DEM incorporate inertial information into DEQ dynamics, and thus share the general idea of exploiting historical iterates to improve the convergence behavior of equilibrium models. However, they differ in the placement and role of the inertial mechanism, as well as its integration into the stability analysis.

First, the inertia is introduced at different levels of the equilibrium computation. In i-DEQ, the inertial mechanism is incorporated into the inner solver to accelerate the fixed-point iterations. In contrast, MHC-DEM introduces inertia exclusively into the outer temporal equilibrium dynamics, while deliberately keeping the nested spatial refinement steps free of momentum. This separation distinguishes spatial reconstruction refinement from temporal information propagation across equilibrium iterations.

Second, MHC-DEM employs a learnable shared momentum coefficient, parameterized as β=Sigmoid(b)\beta=Sigmoid(b), which is optimized through an auxiliary finite-trajectory objective. This allows the strength of temporal inertia to be adapted during training while preserving a stationary fixed-point mapping.

Third, our Theorem 1 explicitly incorporates the effects of heterogeneous prior fusion and temporal inertia into a unified stability condition:

+2\alpha+2\alpha\beta<1$$ Here, the terms involving $\alpha$ and $\alpha\beta$ jointly characterize the contributions of the hyper-connected prior interaction and the inertial coupling, respectively. Thus, the stability analysis accounts for their combined effects within the proposed equilibrium dynamics. In summary, although both methods exploit inertia in DEQ models, i-DEQ introduces inertia into the inner fixed-point solver, whereas MHC-DEM places a learnable inertial mechanism in the outer equilibrium dynamics, together with a momentum-free nested spatial refinement. We will add this discussion and the corresponding reference to better clarify the distinction from [CRDP26]. **Q5. Text Revisions** In Section 3.1, we will clarify that the theoretical analysis first establishes the stability condition for a single inner refinement step ($T=1$), isolating the effects of prior fusion and temporal inertia, and then extends to the general case of $T$ non-shared inner steps, where the stability condition accounts for the cumulative expansion of the composite operator. We will also standardize citations using bracketed reference commands (e.g., \citep{}) and correct equation references throughout the manuscript, eliminating redundant expressions such as “Eq. Equation (xx).”
Reviewer_xuF6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MHC-DEM for imaging inverse problems, combining Sinkhorn-based fusion of heterogeneous denoisers, nested spatial refinement, and inertial equilibrium updates. Experiments on deblurring, inpainting, and interferometric imaging report improvements in reconstruction quality and convergence speed.

Critical strengths
  1. Combining complementary denoisers is well motivated, and the reported results demonstrate practical benefits.

  2. The appendix includes useful ablations on fusion strategies, refinement depth, and momentum, including entropy and load-balancing baselines.

Critical weaknesses

1、 Limited technical novelty. Although applying the MHC idea to combine multiple denoisers is a worthwhile direction to explore in image denoising, the technical novelty requires further discussion. The authors should clarify the specific technical advances of the proposed fusion mechanism over existing MHC designs and conventional weighted ensembling.

2、 Parameter-matched comparison. The experiments would benefit from comparison with a single denoiser whose parameter count is comparable to the total parameter count of the ensemble. This would help clarify whether the proposed ensemble offers advantages beyond simply increasing model capacity.

3、 Comparison with lightweight adaptive gating. Given the motivation of exploiting complementary priors, a useful baseline would be a lightweight network that predicts fusion weights from the input observation. Such a comparison would help establish whether Sinkhorn-based fusion offers advantages over a simpler input-adaptive weighting mechanism.

Questions for authors

See weakness.

2 comments

We thank the reviewer for these insightful comments.

Q1. Technical Novelty

Existing MHC designs primarily employ learnable hyper-connections to establish information exchange and aggregation across network branches or feature streams. Conventional weighted ensembling, on the other hand, typically combines multiple denoisers through directly learned weights, e.g.,

r_{\mathrm{fused}}(z)=\sum_{i=1}^{K}w_i(D_i(z)-z).$$ These formulations provide flexible feature aggregation or direct weight-based fusion, respectively. Our method further introduces an explicit interaction structure among heterogeneous denoising priors, allowing their contributions to be jointly determined by both learnable prior preferences and structured prior interactions. Specifically, we parameterize the fusion weights as $$W_t=\operatorname{Sinkhorn}\left(\exp(L_t/\tau)\right), \qquad q_t=\operatorname{Softmax}(v_t), \qquad w_t=W_tq_t,$$ where $q_t$ represents the learnable base preference over different priors, and $W_t$, constrained to the Birkhoff polytope, captures their structured interactions. Both components are jointly optimized end-to-end. This formulation introduces a structured transformation of the base preference, rather than directly learning the final fusion weights. Moreover, since $W_t$ is doubly stochastic and $q_t$ lies on the probability simplex, the resulting $w_t$ also belongs to the probability simplex. Consequently, the fused prior is a convex combination of individual denoiser residuals, yielding a Lipschitz bound independent of the number of priors under the assumptions of our analysis. The structured interaction parameterization also provides a more balanced fusion mechanism, as empirically demonstrated by the fusion ablations and weight-evolution results in Tables 8. Beyond the fusion parameterization itself, our method integrates the fused denoiser residuals directly into the inner reconstruction updates of the DEQ model. Thus, heterogeneous priors continuously participate in the iterative reconstruction dynamics and jointly determine the equilibrium solution, rather than being combined only through a one-time output-level aggregation. **Q2. Parameter-matched Comparison** To distinguish the benefits of multi-prior fusion from increased model capacity, we construct a wider DRUNet by expanding its channel width. All configurations employ learnable hyper-connections. The wider DRUNet serves as a single-denoiser capacity control. As shown in Table R4, increasing DRUNet from 32M to 74M parameters improves PSNR by only 0.31 dB. In contrast, our three-denoiser configuration achieves 32.91 dB with only 59M parameters, improving over the original DRUNet by 1.25 dB and outperforming the wider single-denoiser baseline by 0.94 dB. Furthermore, adding SwinIR reduces PSNR to 32.32 dB despite increasing the parameter count to 71M. These results demonstrate that performance gains arise primarily from effective fusion of complementary priors, rather than simply increasing model capacity or the number of denoisers. **Table R4. Performance comparison under different denoiser configurations.** Denoiser Configuration | #Params (M) | PSNR | ΔPSNR | Time (ms) ----------------------------------------|-------------|-------|--------|--------- DRUNet | 32 | 31.66 | -- | 517 Wider DRUNet | 74 | 31.97 | +0.31 | 756 DnCNN + DRUNet | 33 | 32.65 | +0.99 | 548 DnCNN + DRUNet + Restormer | 59 | 32.91 | +1.25 | 591 DnCNN + DRUNet + Restormer + SwinIR | 71 | 32.32 | +0.66 | 672 **Q3. Comparison with Lightweight Adaptive Gating** We further introduce a lightweight gating baseline within MHC-DEM. A three-layer CNN (16/32/64 channels), followed by global average pooling and a two-layer MLP (hidden dimension 32), predicts input-dependent global fusion weights: $$w_{\mathrm{gate}}(y)=\operatorname{Softmax}(G_\phi(y)).$$ These weights fuse denoiser residuals and remain fixed throughout reconstruction. The gating network is jointly trained with the reconstruction model, while all other settings remain unchanged. The resulting PSNRs are 29.43±0.35 dB (Gated), 30.96±0.16 dB (Softmax), and 32.91±0.32 dB (Ours). Our method outperforms Gated and Softmax by 3.48 dB and 1.95 dB, respectively. This suggests that global input-adaptive weighting is insufficient to fully exploit complementary priors. In contrast, Sinkhorn-based fusion introduces structured interactions among priors through a doubly-stochastic transformation, enabling more effective prior integration.
Reviewer_xuF6OpenReview ↗

Response to rebuttal

Thank you for your detailed response. Your clarifications have fully addressed my concerns, and I will maintain my original score.

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

This paper proposes MHC-DEM, a deep equilibrium framework combining Sinkhorn-constrained heterogeneous prior fusion, second-order inertial dynamics, and nested spatial refinement. The method aims to improve reconstruction quality while maintaining equilibrium stability. Experiments on image deblurring, inpainting, and black hole interferometric imaging demonstrate promising performance.

Critical strengths
  1. The integration of heterogeneous priors, Sinkhorn-based fusion, and inertial equilibrium dynamics is well motivated. The framework provides a meaningful approach to improving the expressiveness of deep equilibrium models.
  2. The method achieves strong reconstruction results across different inverse problems. The fusion comparisons, convergence analysis, and refinement-depth ablations provide useful experimental evidence.
Critical weaknesses
  1. The convergence analysis relies on strong contractivity conditions and 1-Lipschitz denoisers. These assumptions may not hold for the ill-posed and nonlinear inverse problems considered, limiting the applicability of the theoretical guarantees.
  2. Although Sinkhorn fusion empirically produces more balanced weights, doubly stochastic constraints alone do not necessarily prevent highly concentrated fusion distributions. The claimed theoretical advantage over conventional simplex-based fusion requires further clarification.
Questions for authors

Q1. Could the authors clarify how the convergence guarantees apply to rank-deficient and nonlinear inverse problems, where the required contractivity conditions may not hold?

Q2. Does the doubly stochastic constraint theoretically prevent prior collapse when the query distribution is learnable? Could the authors distinguish this guarantee from empirical observations?

Q3. How is the 1-Lipschitz assumption verified for the heterogeneous denoisers used in the experiments?

Q4. Are the Sinkhorn fusion weights conditioned on the current reconstruction, or are they fixed after training for each inner step?

Q5. Could the authors clarify how iteration-dependent momentum coefficients preserve the fixed-point formulation and implicit differentiation?

Q6. Since real EHT observations lack ground-truth images, could the authors justify using a reference reconstruction as ground truth and provide measurement-domain evaluations?

Q7. Could the authors provide more detailed compute- and parameter-matched comparisons to distinguish the benefits of structured fusion from increased model capacity?

Q8. Could the authors further clarify the methodological differences between the proposed high-order dynamics and existing inertial or densely connected equilibrium models?

3 comments

We thank the reviewer for the insightful comments.

Q1. Convergence for Rank-deficient and Nonlinear Inverse Problems

Theorem 1 establishes a sufficient stability condition for the proposed equilibrium dynamics:

+2\alpha+2\alpha\beta_k<1.$$ This condition jointly accounts for the data-consistency operator, spatial prior correction, and temporal momentum. For nonlinear inverse problems, the condition can be interpreted locally around an equilibrium, with stability depending on the local curvature of the data-fidelity term. We acknowledge that rank deficiency or nonlinearity does not automatically guarantee this condition. Thus, Theorem 1 provides a sufficient condition rather than a universal guarantee for arbitrary inverse problems. Nevertheless, our EHT experiments demonstrate the practical applicability of MHC-DEM to a challenging nonlinear imaging problem involving visibility amplitudes, closure phases, and log closure amplitudes. We will clarify this distinction in the revised manuscript. **Q2. Doubly Stochastic Constraint and Prior Collapse** Our fusion mechanism is defined as $$W=\operatorname{Sinkhorn}(\exp(L/\tau)),\quad q=\operatorname{Softmax}(v),\quad w=Wq,$$ where $W$ is doubly stochastic. For any probability vector $q$, the fused weights satisfy $$w\geq 0,\qquad \mathbf{1}^{\top}w =\mathbf{1}^{\top}Wq =\mathbf{1}^{\top}q=1.$$ However, column-stochasticity alone is sufficient to preserve the simplex constraint. Therefore, the additional row-stochastic constraint does not theoretically guarantee the prevention of prior collapse, particularly when $q$ is learnable. Instead, it imposes a more structured transformation that empirically mitigates weight concentration. To verify this, we compare four fusion strategies in Table R1. Sinkhorn fusion reduces weight concentration compared with Softmax-only and column-stochastic fusion, while achieving the best PSNR. These results support its empirical effectiveness. **Table R1. Ablation study of different prior fusion strategies.** | Fusion Strategy | PSNR (dB) | Maximum Weight | | :------------------------ | :---------------------: | :------------: | | Query only (Softmax) | $30.96\pm0.16$ | 0.97 | | Column-stochastic + Query | $31.22\pm0.35$ | 0.85 | | Uniform Weight | $31.45\pm0.10$ | 0.34 | | Sinkhorn (Ours) | $\mathbf{32.91\pm0.32}$ | 0.51 | **Q3. Verification of the 1-Lipschitz Assumption** In our implementation, spectral normalization (SN) is employed to control the operator norms of heterogeneous denoiser networks and promote their 1-Lipschitz regularity. Additionally, Fig. 15 reports the estimated spectral radius of the equilibrium Jacobian under different momentum configurations. The learned configuration yields a spectral radius below one, providing empirical evidence of local stability of the overall equilibrium dynamics. We acknowledge that SN is a practical means of controlling network regularity, rather than an exact numerical verification that every heterogeneous denoiser is strictly 1-Lipschitz. The 1-Lipschitz condition in Assumption 1 is a theoretical regularity assumption used to derive the stability bound of the fused prior operator. Related regularity assumptions have also been adopted in denoising-based inverse solvers, including Ryu et al. (2019), who use real spectral normalization, and Xue et al. (2022), who establish convergence under a demi-Lipschitz condition. **Q4. Whether Fusion Weights Depend on the Current Reconstruction** The Sinkhorn fusion weights are fixed after training for each inner step, rather than dynamically conditioned on the current reconstruction. Specifically, each inner step $t$ has its own transport logits $L_t$ and query logits $v_t$, which are optimized during training and remain unchanged during inference. The fusion weights are computed as $$W_t=\operatorname{Sinkhorn}(\exp(L_t/\tau)),\quad q_t=\operatorname{Softmax}(v_t),\quad w_t=W_tq_t.$$ Thus, the fusion coefficients are step-specific but fixed during inference. Nevertheless, the fused prior residual is reconstruction-dependent, since each denoiser residual $D_i(z^{(t)})-z^{(t)}$ is evaluated using the current inner-step state.

Q5. Iteration-dependent Momentum and Implicit Differentiation

We apologize for the notational inconsistency. The momentum coefficient is not iteration-dependent; it is a shared learnable scalar across equilibrium iterations, defined as βk=Sigmoid⁡(b)\beta_k=\operatorname{Sigmoid}(b), where bb is optimized during training. We will correct this notation throughout the revised manuscript. This parameter sharing ensures that the augmented dynamics form a stationary mapping uk+1=Tθ(uk;y)u_{k+1}=T_\theta(u_k;y), preserving the fixed-point formulation required by DEQ models.

At equilibrium, zk=zk−1=z⋆z_k=z_{k-1}=z^\star, and thus the inertial term αβ(zk−zk−1)\alpha\beta(z_k-z_{k-1}) vanishes. Consequently, the equilibrium equation reduces to the spatial fixed-point equation, meaning that momentum affects the convergence trajectory but not the equilibrium solution. Therefore, implicit differentiation remains applicable to the spatial model parameters. Since the equilibrium solution is independent of β\beta, the shared momentum parameter is instead optimized using an auxiliary finite-trajectory objective.

Q6. Reference Reconstruction and Measurement-domain Evaluation for EHT

Real EHT observations do not provide access to the exact underlying sky brightness distribution. We therefore use the official EHT Collaboration SR1 reconstruction obtained using CLEAN as a physically plausible reference, rather than treating it as absolute ground truth. Table 5 reports image-domain MSE and Pearson correlation with this reference, while Figs. 10-12 provide visual comparisons, residual maps, edge structures, and radial intensity profiles.

Importantly, our reconstruction is directly guided by the observed measurements through the data-fidelity objective:

=\alpha_{d1}\chi_{\mathrm{amp}}^2(z) +\alpha_{d2}\chi_{\mathrm{cp}}^2(z) +\alpha_{d3}\chi_{\mathrm{logcamp}}^2(z).$$ The corresponding analytical gradients are computed using the Ehtim library and incorporated into the reconstruction updates. We focus on reference-based image-domain evaluation rather than reporting separate measurement-domain residual scores, since the three measurement types have different statistical characteristics and scales, requiring additional calibration for meaningful interpretation. We have also added DE-Prox and DenseDiff as DEM-based baselines on synthetic M87 data in Table R2, where ground-truth images are available. Both use the same pretrained denoisers and follow the same preprocessing and experimental protocols as MHC-DEM. MHC-DEM achieves the best MSE and correlation among these methods, providing an additional quantitative comparison under controlled conditions with known ground truth. **Table R2. Quantitative comparison of black-hole image reconstruction on synthetic M87 data.** Method | MSE (×10⁻⁴) | Corr. -------------|-------------|------ DE-Prox | 1.89 | 0.974 DenseDiff | 1.68 | 0.978 PnP-DM | 3.52 | 0.936 MHC-DEM | 1.32 | 0.996 **Q7. Comparisons with Increased Model Capacity and Computation** To distinguish the benefits of structured fusion from increased model capacity, we compare different denoiser configurations and introduce a wider DRUNet as a single-denoiser capacity control in Table R3. All configurations employ learnable hyper-connections. **Table R3. Performance comparison under different denoiser configurations.** Denoiser Configuration | #Params (M) | PSNR (dB) | ΔPSNR | Time (ms) ----------------------------------------|-------------|-----------|--------|--------- DRUNet | 32 | 31.66 | -- | 517 Wider DRUNet | 74 | 31.97 | +0.31 | 756 DnCNN + DRUNet | 33 | 32.65 | +0.99 | 548 DnCNN + DRUNet + Restormer | 59 | 32.91 | +1.25 | 591 DnCNN + DRUNet + Restormer + SwinIR | 71 | 32.32 | +0.66 | 672 Increasing DRUNet from 32M to 74M parameters improves PSNR by only 0.31 dB, while increasing inference time from 517 to 756 ms. In contrast, the three-denoiser configuration achieves 32.91 dB with 59M parameters and 591 ms inference time, outperforming the wider DRUNet by 0.94 dB with fewer parameters and lower computation. Adding SwinIR further increases the parameter count to 71M but reduces PSNR to 32.32 dB. These results indicate that the improvement is not simply due to increased model capacity or the number of denoisers, but rather the effective exploitation of complementary priors through structured fusion. We have added these comparisons to the revised manuscript.

Q8. Differences from Existing Inertial and Densely Connected Equilibrium Models

MHC-DEM does not simply introduce a conventional momentum term; it integrates temporal inertial dynamics, structured multi-prior fusion, and nested spatial refinement within a unified equilibrium framework. Unlike conventional inertial methods such as Heavy-Ball iterations, which incorporate historical state differences as momentum, MHC-DEM formulates temporal dynamics in an augmented state space. This formulation enables an explicit stability analysis that jointly accounts for data fidelity, heterogeneous prior fusion, and temporal inertia. Unlike densely connected equilibrium models such as DenseDiff, which aggregate information from multiple historical iterations through learned dense connections, MHC-DEM employs an explicit second-order temporal hyper-connection based on two consecutive equilibrium states. Furthermore, its nested spatial refinement updates

=z^{(t)}-\eta_t\nabla\mathcal{E}(z^{(t)};y) +\alpha v_{\mathrm{fused}}(z^{(t)};L_t)$$ progressively refine the reconstruction using non-shared stage-wise parameters, while temporal inertia operates at the outer equilibrium level. This separates inner spatial refinement from outer temporal evolution. Our ablation and convergence experiments (Table 3 and Figs. 4, 14, and 15) further validate the contributions of these components.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.