acceptodds
Under review as a conference paper at ICLR 2027

Random Subspace Descent for Probability Flows

Abstract

Wasserstein gradient flows connect sampling and generative modeling with optimization over probability measures. Kernel-based flows such as MMD descent, Sobolev descent, and Laplacian Adjusted Wasserstein Gradient Descent (LAWGD) yield interacting particle algorithms, but their descent directions can be expensive to compute. Random features and slicing reduce this cost, but many slicing algorithms ignore the geometry of the flow and need not decrease the original objective. We approximate the flows by solving their defining variational problem over a random span of ridge potentials. This preserves the original objective and geometry, reduces the update to a Gram solve, and yields an exact decomposition of captured and lost dissipation. Particle experiments demonstrate the advantage of projection over naive slicing.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Optimal Transport, Gradient Flows, Subspace Descent, Coordinate Descent, Flow Matching

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rhsSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a random subspace descent framework for probability flows to reduce the computational cost of high-dimensional distribution optimization. By projecting descent directions onto randomly sampled subspaces while preserving the original objective and geometry, the method provides theoretical guarantees and demonstrates promising empirical results across several probability flows.

Critical strengths

The proposed method is well-motivated and technically sound. The overall framework is natural and clearly presented, with a reasonably comprehensive theoretical analysis supporting its effectiveness. Although the underlying idea is relatively straightforward, the formulation is mathematically principled, and the reported results demonstrate its potential benefits.

Critical weaknesses
  1. Random subspace methods have been extensively studied in the optimization literature. Therefore, the core technical idea of this paper is not particularly novel. Nevertheless, extending subspace-based optimization to probability flows is a natural and reasonable direction, and the authors provide a relatively comprehensive theoretical analysis. Despite some concerns regarding novelty, I still consider this a technically solid and meaningful contribution.
  2. The manuscript appears somewhat unfinished in several places. For example, the generative modeling experiments lack qualitative visualizations of the generated images, which would help readers better understand the practical improvements. There are also noticeable editorial issues, such as missing appendix references in Section 2.2. These issues give the impression that the manuscript would benefit from another careful round of revision and proofreading.
  3. Although the experimental results are generally positive, the evaluation of the method on practical AI applications remains somewhat limited. Additional experiments on representative real-world AI tasks would help demonstrate the broader applicability and practical value of the proposed framework.
Questions for authors
  1. What are the key technical distinctions between the proposed framework and existing subspace-based optimization methods, particularly those involving variational projections or functional gradient approximations?
  2. The theoretical analysis suggests that convergence is influenced by the dissipation captured by the random subspace and the associated projection loss. Can this dissipation loss be reasonably estimated or characterized in the practical problems considered in the experiments?
  3. The paper aims to reduce computational costs, but the experiments do not seem to provide sufficient empirical evidence of computational efficiency. Could the authors provide runtime comparisons or other computational measurements to validate the claimed efficiency gains?
  4. The current baselines mainly focus on comparing projected and sliced variants of MMD, Sobolev descent, and LAWGD. Could the authors include at least one existing computationally efficient approximation method as an additional baseline to better demonstrate the competitiveness of the proposed approach?

Overall, I consider this a reasonably solid paper with a natural methodological design and meaningful theoretical analysis. My main concerns are the relatively limited novelty and the insufficient experimental validation, along with several presentation issues that suggest the manuscript would benefit from further refinement. Nevertheless, I appreciate the technical foundation of this work and believe these concerns are potentially addressable. I would be willing to increase my score if the authors can provide convincing clarifications and strengthen the experimental evaluation during the rebuttal process.

1 comment

We thank the reviewer for these comments. We address Critical Weaknesses (CW) and Questions (Q) below.

CW1 — Novelty relative to random subspace optimization. We agree the Euclidean idea is classical and the appendix now says so directly, citing that literature. What changes is the setting: the "subspace" is a space of potentials under a non-Euclidean geometry, so the Gram is the flow's own metric rather than a restricted Hessian UᵀHU; the sampling law preconditions an infinite-dimensional space, with no finite-dimensional analogue; and the captured/lost split is exact in that geometry. We also claim the exact captured/lost decomposition with its trace bound, and three instantiations one of which needs no score, while conceding that the Gram solve and the projection identity themselves are standard. (Sec. 1; App. B, C.)

CW2 — Manuscript appears unfinished. Thanks reviewer for the careful reading. Both specifics fixed. The missing appendix pointers in Section 2.2 are restored; the unresolved references elsewhere pointed into a commented-out appendix, and the manuscript now has zero undefined references. We have also proofread the revised sections. Qualitative sample grids for all four superclasses are added, with real images alongside the sliced and projected arms. (Sec. 2.2; App. I.3, Fig. 3.)

CW3 — Limited evaluation on practical AI applications. We acknowledge this as a scope limitation and prefer to state it. The CIFAR-100 latent coupling task remains our only real-data application. This revision adds depth rather than breadth: seed variability, two coupling baselines, a mechanism ablation isolating which part of the method works, end-to-end timings, sensitivity studies on the regularizer and probe budget, and qualitative samples.

Q1 — Distinctions from existing subspace / variational-projection methods. See CW1.

Q2 — Can the dissipation loss be estimated in the practical problems? Yes, and we now report it rather than leaving it theoretical. Along the reported CIFAR-100 trajectories we assemble the Gram system both for the P sampled potentials and for a rich reference dictionary of 8192 potentials drawn from the same law, and report the captured fraction:

captured fractionP = 128P = 512P = 2048
MMD0.030.120.40
Sobolev0.160.110.31
LAWGD0.160.110.34

A 512-d trial space captures roughly a tenth of what an 8192-d one does — the regime the trace bound predicts — which sharpens the claim: the method helps not because the subspace is nearly complete but because the fraction it captures is spent well. Two caveats are stated in the text: the reference is itself a finite-dictionary estimate, hence a lower bound, and each column follows its own trajectory. We additionally report the companion quantity, the fraction of the available one-step decrease captured by each solve rule at a fixed dictionary: 0.14–0.40 for uniform slicing at its best global scale against 1 for the Gram solve. (App. I.2.)

Q3 — Runtime comparisons. Now reported, and split into the claim we can support and the one we cannot. We withdraw the general efficiency claim on the image task. CIFAR-100 coupling takes 3.6–5.8 s for uniform slicing against 22.9–39.5 s for the projection, downstream training identical at 16.2 s; per projected iteration the cubic solve is only 14–24 %, so cost is dominated by estimating the objective gradient, not by the projection. The cost discussion claims savings only where the accounting supports them: dense Sobolev descent, and LAWGD vs a full Poisson solve. We now also measure that saving. Timing full against projected Sobolev descent to a common objective (16-d target, m = 16, five seeds), the per-step saving grows with the feature count d_ψ as predicted — 1.4× at d_ψ = 768 to 7.8× at 3072 — but projection needs ~3× more steps, so the net speedup is 1.5–2.5× and appears only above d_ψ ≈ 10³; below that the projected arm is slower end to end. This also explains the CIFAR-100 timings: that study runs far below the crossover. (Sec. 3.3; App. D.)

Q4 — An existing efficient approximation as a baseline. Added: minibatch-OT coupling at three batch sizes, the standard coupling baseline for flow matching, alongside an independent-pairing control. GlobFID on the four groups is 21.9/8.1/16.9/15.3 for independent pairing and 16.4/5.8/15.4/14.3 for minibatch OT at batch 512, against 8.9/3.6/11.5/7.3 for the best flow. Minibatch OT improves monotonically with batch size but remains behind every flow coupling on every group, by 1.3–1.8× (worst flow per group). Uniform slicing, our existing comparison, is itself the standard cheap approximation in this family. (Sec. 5.4; App. I.)

Reviewer_14QpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper approximates maximum mean discrepancy (MMD), Sobolev, and Laplacian Adjusted Wasserstein Gradient Descent (LAWGD) through Gram fitting of random ridge potentials, preserving the objective and geometry. With exact moments, it analyzes dissipation capture and convergence via average projections and a random-feature norm. Experiments cover Gaussian targets, high-frequency ripples, multiscale rings, and CIFAR-100.

Critical strengths

I think the framework clearly separates objective, geometry, and trial space. Under its stated assumptions, Eq. (8) is well motivated, and the trace bound explains why finite random subspaces cannot uniformly capture dissipation in infinite dimensions. Appendix G includes full-feature references, repeated runs, and matched-time two-dimensional comparisons, supporting joint fitting over naive uniform slicing in the tested settings.

Critical weaknesses

W1. The key real-data experiment does not sufficiently specify the implemented method or evaluation protocol. I think, section 5.4 does not explain the CIFAR-100 target score required by Eq.9 and Algorithm 2 which is really important to readers. Latent samples and whitening do not determine ∇V. The score or generator used, its training, and its correspondence to the stated LAWGD geometry remain unclear. Appendix D.5 assumes an available score without identifying its source in this experiment. The GLOBFID extractor, splits, sample counts, repeated runs, uncertainty estimates, and step-size selection set are insufficiently specified. These omissions hinder interpretation of Table 3 and verification of the claimed LAWGD implementation and generative benefits.

W2. The structure-preservation and convergence results are not sufficiently connected to the implemented particle algorithm. The guarantees assume population flows and exact moments, while Algorithm 1 uses estimated moments, regularization, and finite steps. Appendix D.1 calls fitted dissipation a diagnostic rather than a descent certificate. A complete particle theory is unnecessary, but the structure-preservation claim needs explicit, verifiable conditions or direct error and descent diagnostics for the reported experiments. Theorem 4.3 assumes discrete descent and either uniform dissipation capture or a random-feature norm bound; general LAWGD descent remains open. Example 4.4 requires an orthogonal product decomposition, one active factor, and a small bounded perturbation. The nonincrease claim in lines 317–319 should be limited to Proposition E.7’s fixed-operator model. This does not cover general flows with time-varying geometries and sampling laws. My concern is the theory’s practical scope, not the projection identity or conditional theorem.

W3. The substantive theoretical and algorithmic advances over the closest prior work remain insufficiently articulated. The quadratic fit, Gram solve, and Eq. (8) use standard projection principles. Appendix A.1 identifies related work by Dong et al. (2023), Wang et al. (2024), and Xu et al. (2025). The additional insights beyond random trial spaces and unified notation need clearer explanation. The sampling analysis and unification have value; simple proofs need not lack novelty. As far as I am concerned, the theoretical increment and under-specified empirical support do not yet justify acceptance.

W4. The existing baselines and cost analysis do not establish the broader practical advantages claimed. Gram fitting changes scaling, normalization, and directional coupling together. Backtracking, matched-time comparisons, and step-size searches partly address this, but scale-calibrated slicing or a diagonal-Gram control on an existing task would better isolate the mechanism. Appendix D.4 retains the O(N(N + M)d) witness cost for general-kernel MMD, while fixed-feature MMD already avoids a large solve. LAWGD costs exclude score evaluations and marginal proposal solves. The two-dimensional timings do not establish end-to-end efficiency for high-dimensional or image tasks.

W5. Please fix unresolved references on pages 3, 18, and 20, and qualify “capture the full dissipation” in Section 6 consistently with Eq. (8).

Missing real-data implementation details, limited links between theory, particle experiments, and efficiency evidence, and unclear advances over prior work leave central claims insufficiently supported. I recommend rejection but would reconsider concrete evidence addressing these concerns.

Questions for authors

Q1. LAWGD implementation and CIFAR-100 evaluation. How are the CIFAR-100 target score or generator, Gram matrix, and right-hand side obtained? Please specify the GLOBFID extractor, splits, sample counts, step-size selection, and repeated runs.

Q2. Fixed objectives versus refreshed probes, and descent guarantees. Please distinguish objective features, trial potentials, and refreshed probes. Do Section 5.2’s probes estimate a fixed objective or change it? Which particle-level descent properties are guaranteed rather than only empirical?

Q3. Novelty relative to the closest methods. What advances beyond established variational projection principles does the paper offer relative to Appendix A.1’s closest methods, and how do these inform implementation?

Q4. Mechanism ablation and end-to-end cost. Could calibrated slicing or diagonal-Gram fitting isolate normalization effects on one existing task? Do timings include score evaluation, proposal construction or solving, and Gram assembly?

Q5. Validity of the density transformation used for KL evaluation. For Appendix G.1’s KL evaluation, what ensures that Tₙ(x) = x − ηₙ∇ûₙ(x) is globally one-to-one, as required by log ρₙ₊₁(Tₙ(x)) = log ρₙ(x) − log |det DTₙ(x)|? Exact Jacobians and objective decrease do not establish global invertibility. Multiple preimages require summing their density contributions. Appendix F.2’s proof concerns a special one-dimensional LAWGD update and does not directly cover these two-dimensional experiments. Does the implementation enforce a global Hessian or Lipschitz-based step-size bound, or another sufficient condition? If it checks only objective decrease or sampled Jacobians, how is density propagation justified? Please explain how multiple preimages are handled and whether the LAWGD KL results in Figures 2–4 are affected.

1 comment

We thank the reviewer for these comments. Two changed what the paper claims, and the density-transformation question identified a gap we have closed. We address Critical Weaknesses (CW) and Questions (Q) below.

W1 — Under-specified real-data experiment; where is the score? The reviewer is right that latent samples and whitening do not determine ∇V; none is used. The LAWGD arm is score-free: no score model is trained and none is available. Its profiles are marginal Poisson proposals, but the combination uses the kinetic Gram, not the LAWGD Gram, which would need the score. It is a valid instance of the projection, to which the decomposition applies, but not KL descent. Protocol: we specify the autoencoder, the 256-d whitened latent, the velocity network trained identically for every arm, 100 Euler steps, 5000 samples, the GlobFID extractor (128-d CNN features on the group's classes), and the step-size rule. Table 3 gives the seed-paired difference and its sem over 3–10 seeds (per-arm values in Table 7), plus independent-pairing and minibatch-OT baselines; projection improves GlobFID on 8 of 12 cells, 7 resolved at |t| > 2. (Sec. 5.4; App. I.)

W2 — Theory not connected to the particle algorithm. The non-increase claim is now restricted to the fixed-operator model of its proposition, not flows with time-varying geometry. We add the diagnostics the reviewer prefers: along the reported trajectories we assemble the Gram system for the P potentials and a rich 8192-potential reference, reporting the captured fraction (≈0.11 at P = 512, ≈0.31–0.40 at P = 2048) and each solve rule's per-step decrease (W4). Guaranteed at particle level: with exact moments the fit satisfies bᵀa ≥ 0. Not guaranteed: monotone decrease under estimated moments. (Sec. 4; App. I.2.)

W3 — Advances over the closest prior work. We concede the standard parts: the quadratic fit, the Gram solve and the dissipation identity follow from standard projection principles, and Dong et al. (2023) share the fitting principle. Our increment is approximating only the trial space while preserving objective and geometry, and the average-projector analysis of the sampling law, which converts the identity into a rate. (Sec. 1; App. B.)

W4 — Baselines and cost. We ran both controls — scale-calibrated slicing and a diagonal-Gram fit — holding the Gram and right-hand side fixed, varying only the final solve. Fraction of available one-step decrease captured over 400 iterations: 0.14–0.40 / 0.91–0.99 / 1. Coupling adds the last few percent on this dictionary, where patch directions are nearly orthogonal; where they cannot be (m = 16–24, d = 2) projection wins 9 of 18 configurations by up to 5.2× (new figure). Cost: we accept the accounting and report end-to-end image-scale timings — 3.6–5.8 s sliced vs 22.9–39.5 s projected, identical 16.2 s training. Per iteration at P = 512 the cubic solve is only 14–24 %; load assembly dominates and covers proposal solves and Gram assembly (no score is evaluated). We also test the saving we do claim: timing full against projected Sobolev descent to a common objective, the per-step saving grows with d_ψ (1.4×→7.8×) but ~3× more steps are needed, giving net 1.5–2.5× only above d_ψ ≈ 10³ (App. D). (Sec. 5.4; App. I.2.)

W5 — References and "capture the full dissipation". Both fixed. The unresolved references pointed into a commented-out appendix. The Section 6 phrase is qualified: the Gram solve captures all dissipation available within the sampled trial space, in general a strict fraction of the full dissipation. (Sec. 6.)

Q1/Q3/Q4. See W1, W3, W4. Q2. A new paragraph distinguishes the three: objective features ψ are fixed and define the functional; trial potentials span U and are redrawn at each refresh; probes are an independent set used only to estimate the load b. Probes are redrawn each step but re-estimate a fixed population quantity with fresh randomness; they do not change the objective. (Sec. 3; App. E.1.)

Q5 — Density transformation. The reviewer is right, and the gap is real: in a synthetic check at step size 1, 33 of 200 particles had positive Jacobian determinant while the Jacobian was not positive definite — where the propagated density is wrong. The condition η·λ_max(∇²û) < 1 makes the map a diffeomorphism and is computable exactly for ridge dictionaries. We evaluated it on every accepted step of 270 runs spanning both 2-D experiments, anisotropies κ ∈ {1, 25, 100}, nine profile families, laws and rules: it holds everywhere, 0 % of steps violate it, max margin 0.9997, so the KL results are unaffected. The determinant test sufficed only because d = 2, where it fails to imply positive definiteness only for steps strictly larger than the λ_max crossing it already catches; above two dimensions that argument fails, so we state the condition and recommend enforcing it. (App. J.1, Eq. 24.)

Reviewer_jT2zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes random-subspace approximations of probability flows. It samples ridge potentials and solves the original flow’s variational problem over their span through a Gram system. At the population level, the exact solution is an orthogonal projection in the flow’s native geometry, with an exact decomposition of captured and lost dissipation. The framework covers MMD descent, Sobolev descent, and LAWGD. The authors study how sampling affects progress using an average projector and a random-feature norm, derive conditional convergence rates, and evaluate the method on synthetic distributions and a CIFAR-100 latent-space coupling task.

Critical strengths
  1. The distinction between the objective, geometry, and trial space is clear. It helps explain why approximating a flow is different from changing its objective or geometry through slicing.
  2. The framework covers MMD, Sobolev, and LAWGD flows with concrete implementations. For LAWGD, it avoids explicitly estimating the density ratio or solving the full Poisson equation.
  3. The theory explains the limits of finite-dimensional random subspaces and how the sampling law affects the dissipation they capture. The translation examples also connect the method to familiar subspace optimization methods and show how sampling and geometry affect conditioning.
Critical weaknesses
  1. Novelty and computational benefit need clearer justification. The Gram system and Pythagorean identity follow directly from quadratic variational projection, and related fitting principles appear in Dong et al. (2023) [1]. The unification and sampling analysis are useful, but their advance over these foundations needs sharper explanation. Computational savings are also unclear for MMD. The general-kernel estimator in Appendix D.4 retains the O(N(N+M)d)O(N(N+M)d) cost of evaluating the full witness gradient, then adds projection overhead. For fixed-feature MMD, the full potential ψ⊤δμ\psi^\top\delta_\mu already requires no linear solve, so projection again adds cost. The clearest savings concern dense Sobolev descent when m≪dψm\ll d_\psi, and LAWGD relative to solving the full Poisson problem.

  2. The baselines do not isolate the value of joint Gram fitting. Since the population dissipation is b⊤ab^\top a, both a=ba=b and diagonal fitting aj=bj/(Gjj+ε)a_j=b_j/(G_{jj}+\varepsilon) already guarantee nonnegative dissipation. These cheaper alternatives would test whether accounting for interactions between potentials is worth the additional cost. The main experiments compare primarily against uniform slicing. Its normalization needs clearer documentation, particularly in Figure 1, where scalar velocity scaling can affect comparisons against flow time. Tuned diagonal and rescaled-slicing baselines under matched budgets would strengthen the evidence.

  3. The guarantees remain conditional and do not cover the practical algorithm. Theorem 4.3 assumes discrete descent, a full-flow dissipation bound, and quantitative sampling conditions. General discrete descent for LAWGD remains open. Example 4.4 verifies a special case in which sampling the single active factor recovers the full potential, rather than a partial approximation. The assumptions are not established for the practical dictionaries or the experimental trajectories. Estimated moments and regularization introduce further uncertainty. Results on sensitivity to particle count, reference-sample size, and regularization would help assess practical reliability.

  4. The CIFAR-100 experiment lacks essential details. Section 5.4 does not explain how LAWGD obtains the target score from the latent data. The GLOBFID feature extractor, the model architectures, the evaluation split, and the step-size selection criterion also need clarification. Table 3 reports no variability across runs and no standard coupling baseline, such as independent or minibatch OT pairs. Without these, improving 10 of 12 comparisons does not establish the broader usefulness of the couplings. Table 2 similarly reports small absolute W2W_2 gains without uncertainty estimates.

[1] Hanze Dong, Xi Wang, Yong Lin, and Tong Zhang. Particle-based variational inference with preconditioned functional gradient flow. ICLR 2023.

Questions for authors
  1. Which results provide the main advance over existing variational projection methods, such as Dong et al. (2023) [1]? For MMD specifically, where does the computational saving arise if assembling bb retains the cost of evaluating the full witness gradient?
  2. How does joint Gram fitting compare with diagonal fitting and scalar-rescaled slicing under tuned step sizes and matched computational budgets? These comparisons would help distinguish the benefit of interactions between potentials from differences in velocity scaling.
  3. Can the conditions of Theorem 4.3 be verified for the practical dictionaries and experimental trajectories, beyond the special construction in Example 4.4? How do finite-sample estimation and regularization affect objective decrease in practice, without the objective-based backtracking used in Appendix G?
  4. For CIFAR-100, how is the latent target score obtained, and how are GLOBFID, the evaluation split, and step-size selection defined? Could you report variability across runs and compare with independent or minibatch OT couplings? Variability estimates would also help assess the small gains in Table 2.

[1] Hanze Dong, Xi Wang, Yong Lin, and Tong Zhang. Particle-based variational inference with preconditioned functional gradient flow. ICLR 2023.

2 comments

We thank the reviewer for these comments. We address Critical Weaknesses (CW) and Questions (Q) below.

CW1 — Novelty and computational benefit. The introduction now states that the Gram system and Pythagorean identity are standard consequences of Hilbert-space projection, that Dong et al. (2023) fit a descent direction by an equivalent principle, and that random subspace descent is classical in optimization. Our increment: approximating only the trial space, leaving objective and geometry intact (slicing changes the objective, preconditioning the geometry); the exact captured/lost decomposition with its trace bound; and the average-projector analysis of the sampling law, which converts the identity into a rate. On cost we adopt the reviewer's accounting — general-kernel MMD retains the full witness cost, fixed-feature MMD needs no solve — so we drop any implied MMD saving and restrict the claim to dense Sobolev and LAWGD, now measured rather than asserted (Q2). (Sec. 1, 3.3; App. B, C, D.)

CW2 — Baselines do not isolate joint Gram fitting. Implemented as specified: three arms share an identical Gram and right-hand side, differing only in the final solve — scale-calibrated slicing (s = 1ᵀb/1ᵀG1), a diagonal fit (aⱼ = bⱼ/Gⱼⱼ), and the joint solve. Scored on the fraction of available one-step decrease ½bᵀG†b captured over 400 iterations: 0.14–0.40 / 0.91–0.99 / 1 — uniform averaging discards 60–86 %, the mechanism behind the synthetic sections, here in high dimension with estimated moments. Coupling adds the last few percent here, where patch directions in a 256-d latent are nearly orthogonal; where they cannot be (m = 16–24, d = 2) projection wins 9 of 18 configurations, by up to 5.2×. Normalization is load-bearing: without it every projected arm loses to slicing on the first superclass (+0.72/+1.60/+1.97 GlobFID) while the sliced arm barely moves (9.10 → 9.25). Figure 1's caption now documents it. (Sec. 5.4; App. I.2; Fig. 2.)

CW3 — Conditional guarantees. Scoped: the non-increase statement is restricted to the fixed-operator model of its proposition, general discrete descent for LAWGD is stated as open, and the product example is a special case. In place of a particle theory we verify Theorem 4.3(a)'s hypothesis directly on the reported trajectories and sweep the regularizer and probe budget (Q3). Limitation: particle count and reference-sample size are fixed at the full group size. (Sec. 4; App. I.2.)

CW4/Q4 — CIFAR-100 detail, variability, baselines. The protocol is now given in full — autoencoder, 256-d whitened latent, the velocity network trained identically for every arm, sampling, and the GlobFID extractor (128-d CNN features on the group's classes, not Inception) — together with the step-size rule: each arm at its optimum on the shared grid. On the latent score: there is none. The LAWGD arm combines marginal Poisson profiles with the kinetic Gram, not the score-based Gram: a valid instance of the projection, but not KL descent. The score-based form is used in the synthetic sections. Table 3 reports the seed-paired difference and its sem over 3–10 seeds (per-arm values in Table 7): projection improves GlobFID on 8 of 12 cells, 7 resolved at |t| > 2, concentrated on the two groups of similar species. Baselines: independent pairing 21.9/8.1/16.9/15.3 and minibatch OT 16.4/5.8/15.4/14.3 vs 8.9/3.6/11.5/7.3 for the best flow; every flow beats both on every group (1.4–2.4×, 1.3–1.8×). (Sec. 5.4, Table 3; App. I.)

Q1. See CW1. Q2. Step sizes are tuned per arm; against slicing projection costs ≈6.9× more, but given that budget as extra iterations the sliced arm is flat or worse, so the ranking at matched wall-clock is that at matched iterations (App. I.2). Against the full solve — the comparison the cost claim is about — projection is measurably cheaper beyond a crossover near d_ψ ≈ 10³: timed to a common objective, net 1.66× at d_ψ = 1536 and 2.54× at 3072; below it projection is slower (0.48× at 768). CIFAR-100 runs far below that crossover, which is why the projected coupling costs more there (App. D, Table 8).

Q3. Partly, theorem 4.3(a) needs a uniform lower bound α on the captured dissipation fraction, and α is measurable along our trajectories: against a rich 8192-potential reference it is positive at every step, min α ≈ 0.09–0.10 at P = 512 and 0.25–0.36 at P = 2048 for all three flows, min and median within a factor 1.5 — stable, not just positive. With κ = 1 exact for LAWGD, ακ is pinned for the reported runs. Caveats: this verifies the hypothesis a posteriori, and Δ⋆ is a finite-dictionary estimate, so the measured α upper-bounds the truth. On descent without objective-based backtracking: the CIFAR-100 runs use none — fixed step with renormalization — and ε is inert across four decades, so neither finite-sample estimation nor ε governs the decrease there (Sec. 4; App. I.2).

I thank the authors for the detailed response and the new experiments. That said, the results mostly narrow the paper's claims. The diagonal fit already gets 91-99% of the available decrease on CIFAR-100, the projection is slower than slicing on the image task, and the CIFAR-100 "LAWGD" arm turns out not to be KL descent. I will keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.