acceptodds
Under review as a conference paper at ICLR 2027

QDSB: Quantized Diffusion Schrödinger Bridges

Abstract

Learning generative models in settings where the source and target distributions are only specified through unpaired samples is gaining in importance. Here, one frequently-used model are Schrödinger bridges (SB), which represent the most likely evolution between both endpoint distributions. To accelerate training, simulation-free SBs avoid the path simulation of the original SB models. However, learning simulation-free SBs requires paired data; a coupling of the source and target samples is obtained as the solution of the entropic optimal transport (OT) problem. As obtaining the optimal global coupling is infeasible in many practical cases, the entropic OT problem is iteratively solved on minibatches instead. Still, the repeated cost remains substantial and the locality can distort the global transport geometry. We propose quantized diffusion Schrödinger bridges (QDSB), which compute the endpoint coupling on anchor-quantized endpoint distributions and lift the resulting plan back to original data points through cell-wise sampling. We show that the regularized optimal coupling is stable w.r.t. anchor quantization, with an error controlled by the quality of the anchor approximation. In real-world experiments, QDSB matches the sample quality of existing baselines, requiring substantially less time. Code and data are available at github.com/anon1248/qdsb.

Primary area:
generative models
Keywords:
generative modeling, diffusion models, Schrödinger bridges, optimal transport, quantization

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

This paper proposes QDSB, an anchor-based approximation to endpoint coupling for simulation-free schrodinger bridge training from unpaired samples. Each endpoint dataset is partitioned into nearest-anchor cells using farthest-first anchor selection. A transport plan is computed between anchors weighted by their cell populations and training pairs are obtained by sampling an anchor pair followed by original data points within the corresponding cells. This preserves the empirical endpoint marginals while moving transport computation outside the inner training loop. The plan can be periodically refreshed.

The paper presents quantization-based stability bounds and evaluates the method on three synthetic tasks, three single cell datasets with held out intermediate timepoints, and qualitative FFHQ adult-to-child translation. The quantitative results report substantially shorter training times than SF2M with generally comparable MMD, while the image experiment compares outputs under a fixed training-time budget.

Critical strengths
  1. Practical acceleration of a relevant training bottleneck. QDSB amortizes endpoint transport computation across many training updates. In Table 1 it is reported an approximately 1.95x shorter training time than SF2M in the toy suite (205 versus 105 seconds) and 4.30x shorter training time in the cellular suite (318 versus 74 seconds), with generally comparable final MMD. Evaluating quality at fixed time budgets is appropriate for the paper's efficiency objective. These results support the potential practical value of the method.

  2. A simple coupling construction that preserves empirical marginals. Weighting anchors by cell populations and sampling original endpoints within matched cells preserves the empirical source and target marginals while retaining raw-data supervision. This makes the method a straightforward modification of SF2M rather than requiring a new training objective or model architecture. The five-seed quantitative evaluation and released reproduction code are additional strengths.

Critical weaknesses
  1. The stated coupling-stability guarantee does not measure differences between joint couplings. In Assumption 3.1, for finite a, the paper defines Wa′(π,π′)=(Wa(r0,r0′)a+Wa(r1,r1′)a)1/a.W'_a(\pi,\pi')=\left(W_a(r_0,r'_0)^a+W_a(r_1,r'_1)^a\right)^{1/a}.

This quantity depends only on the marginals. It is therefore zero for any two couplings with identical marginals, even when their pairings differ substantially. Consequently, Assumption 3.2 is trivial under this definition, and Theorem 3.3(iii) does not establish the claimed control of joint coupling error.

The admissibility inequality also fails for squared euclidean cost under this definition. Consider source and target marginals that are both uniform on {-1, 1}. One coupling assigns probability 1/2 to each of the pairs (-1, -1) and (1, 1); another assigns probability 1/2 to each of (-1, 1) and (1, -1). Both couplings have identical marginals, so the defined distance is zero, but their expected squared transport costs are respectively 0 and 4. Thus, the displayed admissibility inequality cannot hold for this example.

Eckstein and Nutz's cited result instead distinguishes wasserstein distance between joint measures on the product space from the distance between tuples of marginals. This seems to be a repairable formulation error, but the corrected definition, assumptions, and application must be supplied because coupling stability is a central contribution. The paper should also explicitly connect the corrected guarantee to the lifted raw-data coupling used for training, including the effect of within-cell sampling.

  1. The released experiments do not use the regularized coupling analyzed by the theory. Equation (2) and Theorem 3.3 concern entropic OT with positive regularization, with τ=2σ2\tau=2\sigma^2 in the Brownian-reference construction. However, appendix d specifies an exact OT solver, the released toy and cellular result configurations specify ot_method: "exact", and the implementation maps this setting to pot.emd, rather than the separate entropic Sinkhorn branch.

This supports an empirical acceleration result for unregularized OT based training, but does not establish that the same efficiency-quality trade-off holds for the entropic coupling analyzed in the paper. The displayed stability bound cannot simply be applied at τ=0\tau=0. This mismatch affects my recommendation because the main experiments do not test the version of the method covered by the theory.

Overall, I lean toward weak rejection despite the promising practical results. I would reconsider if the joint-distance formulation is repaired and the authors establish that the efficiency-quality conclusion holds for the positively regularized coupling described in the method.

Questions for authors
  1. Clarification: Is Wa′W'_a intended to denote wasserstein distance between joint measures on the product space, rather than the marginal-only quantity defined in Assumption 3.1? Please provide the corrected formulation and explain which assumptions and constants justify applying the cited stability theorem to your setting.

  2. Connection to the training algorithm: What bound follows for the lifted raw-data coupling qancq_{\mathrm{anc}} relative to the original empirical entropic optimal coupling, after accounting for within-cell sampling?

  3. Clarification: Were the reported toy and cellular results generated using unregularized EMD, as indicated by the released configurations? If so, how should these experiments be interpreted relative to Equation (2) and the Brownian-reference Schrodinger bridge claim?

  4. Validation that could change my recommendation: Can you provide a focused comparison showing whether QDSB retains its efficiency-quality advantage over minibatch SF2M when both use entropic OT with τ=2σ2\tau=2\sigma^2? Please include anchor construction and refresh costs in the reported timing. This directly tests the method analyzed in the paper rather than suggesting a separate future application.

3 comments

We thank the reviewer for their time and effort invested and answer the open questions below.

1 Clarification on Wa′W_a': Indeed, as the reviewer has correctly guessed, Wa′W_a' in Assumption 3.1 and Wb′W_b' in Assumption 3.2 were erroneous; we meant the (usual) non-prime version, which is in line with what Eckstein & Nutz (2022) use, what is required for using their results, and what we actually employ in the proofs of the article (see also our answer 1.1 to reviewer nsM6). We thank the reviewer for pointing this out and have updated the manuscript accordingly.

2 Bound on qancq_\mathrm{anc}: It can be shown that qancq_\mathrm{anc} satisfies the bound

Wb(qanc,πq0,q1τ)≤21+1/b−1/aΔa+Cb(2LΔa/τ)1/(2b),W_b(q_{\mathrm{anc}},\pi^\tau_{q_0,q_1}) \le 2^{1+1/b-1/a}\Delta_a + C_b\big(2L\Delta_a/\tau\big)^{1/(2b)},

which shows that Δa\Delta_a controls the error w.r.t. qancq_\mathrm{anc} as well, and for which we include the proof in our answer 1.2 to reviewer nsM6 and the revised version of the article.

We thank the reviewer for this nice suggestion, which we have added as Theorem 3.3 (iv) in the revised version of the manuscript.

3 Unregularized EMD: We confirm that the reported toy and cellular results used the exact option, which calls unregularized pot.emd. The released code also supports sinkhorn, however. We will make this distinction explicit in the manuscript.

4 Efficiency-quality advantage of QDSB: The reported wall-clock times in the paper already include every step of Algorithm 1: initial anchor selection, assignment of all samples to cells, OT construction, training, and subsequent refreshes. To test the positively regularized variant directly, we reran the three tasks of Table 1(a) with both methods using Sinkhorn and τ=2σ2=0.125\tau=2\sigma^2=0.125 (σ=0.25\sigma=0.25). Model, data, and optimizer settings are unchanged; QDSB uses 256 farthest-first anchors refreshed every 100 epochs. Timings include all anchor selection, cell assignment, OT, and refresh costs, with complete evaluation excluded for both methods. We report MMD (mean ±\pm standard deviation, five seeds) on an Apple M1 Pro CPU using one thread. Both methods use the same double-precision Sinkhorn solver with tolerance 10−610^{-6}. QDSB obtains lower mean MMD on every task at both budgets.

Method8G→M (10 s)G→M (10 s)G→8G (10 s)8G→M (60 s)G→M (60 s)G→8G (60 s)
QDSB0.019±0.0060.019\pm0.0060.011±0.0070.011\pm0.0070.009±0.0190.009\pm0.0190.017±0.0130.017\pm0.0130.003±0.0060.003\pm0.0060.011±0.0160.011\pm0.016
SF2M0.083±0.0050.083\pm0.0050.058±0.0040.058\pm0.0040.063±0.0070.063\pm0.0070.049±0.0080.049\pm0.0080.023±0.0050.023\pm0.0050.043±0.0080.043\pm0.008

We hope this answers all the open questions.

Thank you for the detailed response and the additional experiments.

The corrected Assumption 3.1, now stated in terms of the Wasserstein distance between couplings on the product space, resolves my main theoretical concern, and the new Theorem 3.3(iv) directly answers my question about the lifted coupling used for training. I checked the proof of (iv) and it looks correct to me.

The Sinkhorn comparison with τ = 2σ² = 0.125 is exactly the experiment I was hoping for, and it is reassuring that QDSB retains its advantage under the regularized coupling covered by the theory, with all anchor and refresh costs included.

One request: the revised PDF does not yet reflect the exact-versus-entropic distinction. Appendix D still describes "an exact numerical OT solver" for the toy and single-cell experiments, the main text does not state that Table 1 uses unregularized OT, and the Sinkhorn results are not included. Please add both the explicit statement and the Sinkhorn table (e.g., in Appendix E) to the final version, since they are what connects the experiments to the theory.

With these points addressed, I am raising my score.

We again thank the reviewer for their very constructive feedback, which we have now fully incorporated into the article. See in particular Appendix E and the changes (colored blue) in the experiment setup (Section 4).

Reviewer_nsM6OpenReview ↗
Rating
3(+1)
Confidence
1(−2)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Quantized Diffusion Schrödinger Bridges (QDSB), an endpoint-coupling approximation for simulation-free Schrödinger bridge training. The method quantizes source and target datasets using farthest-first anchors, computes transport between the weighted anchor distributions, and samples original endpoints from matched anchor cells. Reusing this coupling reduces the overhead of solving OT repeatedly during training while preserving the empirical endpoint marginals. The paper presents quantization stability bounds derived from existing regularized OT results. Experiments on synthetic distributions, single-cell datasets, and latent-space image translation report improved training-time efficiency with broadly comparable sample quality.

Critical strengths
  1. A practical coupling construction that preserves the empirical marginals. Weighting anchors by their cell masses and sampling original points within matched cells avoids restricting neural-network supervision to the anchors themselves. This provides a simple way to reduce coupling costs while retaining the full endpoint datasets.

  2. Promising improvements under wall-clock comparisons. Table 1 reports comparable final MMD to SF2M with training times reduced from 205 to 105 seconds on the toy suite and from 318 to 74 seconds on the single-cell suite. Reporting both fixed-time performance and results over five seeds provides useful evidence for the method’s practical value.

Critical weaknesses
  1. The stability analysis needs correction. Assumption 3.1 defines coupling distance using only the marginals, assigning zero distance to distinct couplings with identical marginals. This contradicts its claimed applicability to squared Euclidean costs and undermines the stated guarantee.

  2. The experiments do not isolate the proposed contribution. Comparisons against cached OT pairings and alternative anchor selection with matched refresh schedules are missing. Thus, the gains attributable to anchor geometry remain unclear. High-dimensional evaluation is limited to qualitative results from one seed.

Questions for authors
  1. Clarification of the theoretical guarantee: Is (W_a') intended to denote Wasserstein distance between joint measures on the product space? Please provide the corrected assumptions and explain how the resulting guarantee applies to the lifted coupling (q_{\mathrm{anc}}) actually used for training, rather than only to the anchor-supported plan.

  2. Clarification of the implemented objective: Appendix D describes an “exact numerical OT solver,” whereas the method and theory concern entropically regularized OT. Which solver and regularization coefficient are used in each experiment? Is (\tau=2\sigma^2)? If the experiments use unregularized OT, how do the guarantees and Brownian Schrödinger bridge interpretation apply? Please also reconcile the unsquared cost defined in Section 3.1 with the squared Euclidean cost used experimentally.

  3. Evidence needed to interpret the speedup: Do reported times include initial anchor selection, assignment of all samples to cells, OT construction, and subsequent refreshes? How does QDSB compare with cached pairings or alternative anchor selection under matched refresh schedules and complete preprocessing-inclusive timing?

  4. Evidence for coupling fidelity: On a tractable synthetic problem, can the authors compare the lifted coupling or induced intermediate distributions against a full-data entropic OT reference? Distributional MMD alone does not establish that the intended transport geometry is recovered.

3 comments

We thank the reviewer for their time and effort invested. We answer the open questions below.

1.1 Corrected assumption: We thank the reviewer for pointing this out. The equation in Assumption 3.1 is incorrect as stated. Mirroring Eckstein & Nutz (2022, Definition 3.3), which our results build upon, we require instead the following notion: For a∈[1,∞]a\in[1,\infty] and r0,r1,r0′,r1′∈Pa(Rd)r_0,r_1,r_0',r_1' \in \mathcal P_a(\mathbb R^d) fixed, a cost function c:Rd×Rd→[0,∞)c : \mathbb R^d\times \mathbb R^d \to [0,\infty) [satisfying sufficient regularity conditions; see Section 2 of Eckstein & Nutz (2022); we have included these details in the revised version of our submitted work] is admissible if there exists L≥0L\ge 0 such that for all π∈Π(r0,r1)\pi\in\Pi(r_0,r_1) and π′∈Π(r0′,r1′)\pi'\in\Pi(r_0',r_1'), it holds that

∣∫c d(π−π′)∣≤LWa(π,π′).\biggl|\int c \thinspace \mathrm d(\pi - \pi')\biggr| \le L W_a(\pi,\pi').

We have also updated Theorem B.2 to agree with its original statement in Eckstein & Nutz (2022). Importantly, our results (Theorem 3.3 and Lemma 3.4) continue to hold with this change. Therefore, besides replacing Wa′W'_a and Wb′W'_b by WaW_a and WbW_b, this important fix does not influence the take-aways of our results.

1.2 Guarantee on qancq_{\mathrm{anc}}: In the setting of Theorem 3.3, it holds that (proved below)

Wb(qanc,πq0,q1τ)≤21+1/b−1/aΔa+Cb(2LΔa/τ)1/(2b),W_b(q_{\mathrm{anc}},\pi^\tau_{q_0,q_1}) \le 2^{1+1/b-1/a}\Delta_a + C_b\big(2L\Delta_a/\tau\big)^{1/(2b)},

showing that a small Δa\Delta_a ensures that the coupling used for training and the entropic OT plan of the full data are close.

We thank the reviewer for this nice suggestion, which we have added as Theorem 3.3 (iv) to the revised version of the manuscript. The proof is as follows and also included in the revised version of the manuscript. Let F(x0,x1):=(T0(x0),T1(x1))F(x_0,x_1) := (T_0(x_0), T_1(x_1)). Since F♯qanc=πq~0,q~1τF_{\sharp} q_{\mathrm{anc}} = \pi^\tau_{\tilde q_0,\tilde q_1}, we use the same idea as in the proof of Theorem 3.3(i). Indeed, let Φ(x0,x1):=(x0,x1,T0(x0),T1(x1))\Phi(x_0,x_1) := (x_0, x_1, T_0(x_0), T_1(x_1)) and γ:=Φ♯qanc∈Π(qanc,πq~0,q~1τ)\gamma := \Phi_{\sharp} q_{\mathrm{anc}} \in \Pi(q_{\mathrm{anc}}, \pi^\tau_{\tilde q_0,\tilde q_1}). Then

&W_b(q_{\mathrm{anc}}, \pi^\tau_{\tilde q_0,\tilde q_1}) \cr &\quad= \inf_{\eta \in \Pi(q_{\mathrm{anc}}, \pi^\tau_{\tilde q_0,\tilde q_1})} \Bigl( \int d_b\bigl((x_0,x_1),(y_0,y_1)\bigr)^b \thinspace \mathrm d\eta(x_0,x_1,y_0,y_1) \Bigr)^{1/b} && \text{(by definition)} \cr &\quad\le \Bigl( \int d_b\bigl((x_0,x_1),(y_0,y_1)\bigr)^b \thinspace \mathrm d\gamma(x_0,x_1,y_0,y_1) \Bigr)^{1/b} \cr &\quad= \Bigl( \int \Vert x_0-y_0\Vert^b \thinspace \mathrm d\gamma(x_0,x_1,y_0,y_1) + \int \Vert x_1-y_1\Vert^b \thinspace \mathrm d\gamma(x_0,x_1,y_0,y_1) \Bigr)^{1/b} && \text{(definition of $d_b$, linearity)} \cr &\quad= \Bigl( \int \Vert x_0-T_0(x_0)\Vert^b \thinspace \mathrm dq_{\mathrm{anc}}(x_0,x_1) + \int \Vert x_1-T_1(x_1)\Vert^b \thinspace \mathrm dq_{\mathrm{anc}}(x_0,x_1) \Bigr)^{1/b} && \text{(change of variables)} \cr &\quad= \Bigl( \int \Vert x_0-T_0(x_0)\Vert^b \thinspace \mathrm dq_0(x_0) + \int \Vert x_1-T_1(x_1)\Vert^b \thinspace \mathrm dq_1(x_1) \Bigr)^{1/b} && \text{(marginals of $q_{\mathrm{anc}}$)} \cr &\quad\le \bigl( \varepsilon_0^b + \varepsilon_1^b \bigr)^{1/b} && \text{(Jensen, $b \le a$)} \cr &\quad\le 2^{1/b-1/a} \thinspace \Delta_a && \text{(power-mean inequality)}. \end{aligned}$$ The claim then follows by combining this bound, the triangle inequality, and Theorem 3.3 (iii).

2 Implemented objective: Regarding the squared/unsquared cost: In Section 3.1, cc was used both for the distance and for the transport cost. We renamed the distance to d(x,y):=∥x−y∥d(x,y) := \Vert x-y\Vert and kept cc for the cost. The connection between the Schrödinger bridge and entropic OT with τ=2σ2\tau = 2\sigma^2 in (1) requires the squared Euclidean cost, which is the cost used in all experiments. We also note that, by Example 3.4 of Eckstein & Nutz (2022), this cost satisfies the corrected Assumption 3.1.

3 Speedup: Indeed, the reported wall-clock times include every step of Algorithm 1: initial anchor selection, assignment of all samples to cells, OT construction, training, and subsequent refreshes. The computational saving comes from reusing the anchor coupling across gradient updates. We additionally compare farthest-first with uniformly random anchors drawn without replacement, fixing k=256k=256 and refreshing both every 100 epochs. With all other model, data, OT, optimizer, and sampling settings held fixed, the following MMDs use the same CPU timing protocol (Apple M1 Pro, one thread), including all coupling construction costs. We report mean ±\pm standard deviation over five seeds. While the results are close, farthest-first traversal obtains lower or equal mean MMD more consistently.

Anchor selection8G→M (10 s)G→M (10 s)G→8G (10 s)8G→M (60 s)G→M (60 s)G→8G (60 s)
Farthest-first0.018±0.0080.018\pm0.0080.014±0.0090.014\pm0.0090.011±0.0180.011\pm0.0180.007±0.0100.007\pm0.0100.002±0.0060.002\pm0.0060.010±0.0170.010\pm0.017
Random anchors0.018±0.0120.018\pm0.0120.014±0.0120.014\pm0.0120.015±0.0160.015\pm0.0160.024±0.0160.024\pm0.0160.005±0.0050.005\pm0.0050.011±0.0170.011\pm0.017

4 Coupling fidelity: We agree that target-distribution MMD alone does not establish coupling fidelity. We therefore compared SF2M and QDSB with full-data entropic OT on 8Gaussians-to-Moons, using 2048 samples per endpoint, 256 anchors for QDSB, minibatch size 256 for SF2M, and the same squared Euclidean cost with τ=2σ2=0.125\tau = 2\sigma^2 = 0.125. Across five seeds, the empirical joint endpoint W2W_2 is 0.342±0.0270.342\pm0.027 for QDSB and 0.541±0.0680.541\pm0.068 for SF2M (mean ±\pm standard deviation); reference resampling gives 0.290±0.0170.290\pm0.017. We also compared their Brownian-bridge marginals at every t=0,0.01,…,1t=0,0.01,\ldots,1. Distances use 1024 endpoint pairs per method, with common source samples and conditional target draws to reduce sampling noise. This directly tests the coupling approximation without neural-network fitting. We include the plot that features these results as Figure 7 in the revised version.

We hope this answers all the open questions.

Reviewer_nsM6OpenReview ↗

concerns have been resolved

Thanks to the authors' rebuttal. My technical concerns have been resolved. I will raise my score to 3.

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

The paper proposes a new method for reducing the cost of estimating the OT coupling by computing the OT plan between the quantised anchors. The error analysis presented in the paper ties the error between the true and quantised OT plans to the quantisation error can in turn be be controlled through the quantisation algorithm (the authors propose using farthest-first traversal). Additionally, the method is further supported by the numerical experiments.

Critical strengths
  1. The paper addresses the important question of designing fast algorithms for computing accurate OT couplings for flow matching models, diffusion models and Schrödinger bridges.
  2. The algorithm proposed in the paper is a promising approach of computing the OT coupling quickly. The authors carefully study its theoretical properties and derive an error bound on the quantised OT plan.
Critical weaknesses

The major weakness of the paper lies in showing how the results scale with dimensionality. Although the results in Table 1 suggest that QDSB converges faster than other methods, these experiments are mostly low-dimensional. I am curious whether the same effects can be observed on images and in the latent space of a pretrained VAE. I appreciate the results shown in Figure 4; however, I venture to suggest performing some numerical experiments as well. For example, it would be interesting to compute the path KL divergence for different methods to measure how accurately the SB is approximated. I think that the paper could benefit from experiments on high-dimensional datasets, specifically because computing the OT coupling is costly in high dimensions. Thus, the real advantage of the method would be its ability to couple the marginals when the use of classical OT methods is cost-ineffective. In this light, I would also suggest comparing the proposed algorithm to the recent work of Mousavi-Hosseini et al. (2026), which addresses a similar issue.

Questions for authors

I have several questions regarding the proposed algorithm and its analysis:

  1. What motivates the choice of farthest-first traversal as the algorithm for finding anchors. I assume that any clustering algorithm could be used. Moreover, in that case it would not be necessary to refresh the sets A0A_0 and A1A_1. Furthermore, the motivation behind refreshing A0A_0 and A1A_1 is not quite clear. I would suggest numerically testing whether the refreshing makes any difference at all.
  2. Currently, the proposed method uses hard anchor assignment, that is, each point is assigned to only one anchor. Would it be easy to modify the algorithm to incorporate the uncertainty of the anchor assignment, so that, when a point is close to two or more anchors, this would somehow be taken into account?
  3. What does the constant CbC_b in Theorem 3.3 look like? Is it dimension-independent? If it is, does this imply that the proposed algorithm is robust to increasing dimensionality? If not, is there any way to quantify how the algorithm's performance changes as the dimension grows?
3 comments

We thank the reviewer for the time and effort invested. Below we answer the questions in detail.

1.1 Choosing farthest-first traversal: The choice of using the farthest-first traversal algorithm is motivated by our Lemma 3.4. Indeed, the lemma shows that it is desirable to select anchors that guarantee a good coverage of the endpoint distributions, corresponding to the metric kk-center problem. As the farthest-first traversal algorithm is known to approximately solve this problem, it is our algorithm of choice. We further elaborate the reasoning in the paragraph starting at line 245; see also Theorem B.3.

1.2 Motivation for refreshing A0A_0 and A1A_1: We agree with the reviewer that refreshing has no theoretical benefit: While Theorem 3.3 and Lemma 3.4 hold for every anchor set produced by farthest-first traversal, we include the step to reduce the effect of an unlucky initial anchor choice.

2 Soft anchor assignment: We can imagine an algorithm where the distribution given an anchor is not uniform but distance-dependent, that is, samples are drawn by taking their distances into account. We thank the reviewer for this very nice idea, which could lead to very interesting future work.

3 Constant CbC_b: We first note that Assumption 3.2 is supposed to require that there exists a Cb>0C_b > 0 such that

W_b(\pi',\pi) \le C_b \operatorname{KL}(\pi\Vert\pi')^{1/(2b)} \tag{$\ast$}

holds for all π,π′∈Π(q0,q1)\pi,\pi'\in \Pi(q_0,q_1) [corresponding to (Iq)(I_q) in Eckstein & Nutz (2022); the notation Wb′W_b' in Assumption 3.2 of our submitted work is erroneous and we have updated the manuscript accordingly].

To answer the question: In general, CbC_b in (∗)(\ast) depends on the dimension. For, if the marginals are q0=q1=N(0d,Id)q_0 = q_1 = \mathcal{N}(0_d,I_d), and the couplings are π=N(02d,Σ)\pi = \mathcal N(0_{2d},\Sigma) (0<ρ<10 < \rho < 1 fixed) and π′=N(02d,Σ′)\pi' = \mathcal N(0_{2d},\Sigma'), where

Σ=(IdρIdρIdId),Σ′=(Id0d×d0d×dId),\Sigma=\begin{pmatrix} I_d & \rho I_d \cr \rho I_d & I_d \end{pmatrix}, \qquad \Sigma'=\begin{pmatrix} I_d & 0_{d\times d} \cr 0_{d\times d} & I_d \end{pmatrix},

then, by the known expressions for Gaussians of the 2-Wasserstein distance (with Euclidean norm) and of the Kullback–Leibler divergence, one has, respectively, that

W^2_2(\pi',\pi) &= \operatorname{trace}\bigl(\Sigma+\Sigma' - 2\bigl((\Sigma')^{1/2}\Sigma(\Sigma')^{1/2}\bigr)^{1/2}\bigr) = 4d-2d(\sqrt{1+\rho} + \sqrt{1-\rho}) = \Theta(d), \cr \operatorname{KL}(\pi\Vert\pi') &= \frac12\biggl(\operatorname{trace}\bigl((\Sigma')^{-1}\Sigma\bigr)-2d +\ln \frac{\det(\Sigma')}{\det(\Sigma)}\biggr) = -\frac d2\ln(1-\rho^2) = \Theta(d). \end{aligned}$$ Combining both, we have that $$\frac{W_2(\pi',\pi)}{\operatorname{KL}(\pi\Vert\pi')^{1/4}} = \Theta(d^{1/4}),$$ which shows that $C_b$ cannot be chosen independently of $d$ in general. We hope this answers all the open questions.

I would like to thank the authors the answers. It is still not clear to me if the proposed method scales well with dimensionality, especially in light of the constant CbC_b depending on dd explicitly. Therefore, I choose to maintain my score.

We thank the reviewer for acknowledging our rebuttal and for reconsidering their score.

With the rebuttal closing soon, we would like to emphasize that our image translation experiments feature d=512d=512 and that the proposed method shows good results there. While we consider dd fixed in the manuscript, one expects that the problem difficulty increases with an increase of the (effective) dimension, requiring more data.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.