acceptodds
Under review as a conference paper at ICLR 2027

Why Diffusion Sampling Can Ignore Ambient Dimension: Rate-Distortion Bounds in KL and TV

Abstract

Diffusion models sample high-dimensional data in few steps. Yet existing discretization bounds scale with an intrinsic dimension of the data or, for Gaussian mixtures, with the squared logarithm of the number of components. We show that what matters instead is the information the data carry at the target resolution. For the posterior-mean exponential integrator with steps, the time-discretization error in total variation is . Here is the mutual information between the data and their Gaussian observation at the terminal signal-to-noise ratio, and is the log-SNR range. The bound is linear in , whereas Pinsker's inequality applied to information-theoretic KL bounds gives only . We bound by the rate-distortion function of the data. Up to a constant, it is at most the metric entropy of the support, and for Gaussian mixtures it is at most the label entropy, so our bound improves on both. The implication: as far as discretization error is concerned, diffusion in pixel space does as well as diffusion in latent space, whatever the number of pixels.

Primary area:
learning theory
Keywords:
Diffusion Model, Information Theory, Score-based Generative Models

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_qB2kOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper provides a novel direct TV bound for diffusion sampling with respect to discretization, using a mutual information quantity. The role of this quantity, as well as several applications of this result, were explored.

Critical strengths
  1. Clear exposition: The paper identifies previous bounds on the sampling error with respect to discretization and successfully sets its' contribution apart. Several applications of their theoretical result were also discussed. The role that the mutual information term IΓI_{\Gamma} plays is illustrated very well in the comparison with intrinsic dimension and metric entropy.

  2. Novel theoretical contribution: The paper's main result, Theorem 2, appears to be novel and eliminates the crudeness of several previous approaches in direct TV bounds. The surprising aspect to me is that the proof strategy is concise and perhaps clever.

Critical weaknesses
  1. Broader implications are somewhat limited. Although I acknowledge and support the paper's main theoretical contribution, I think that the broader consequences for diffusion models are limited. The paper touched on this aspect through the comparison between diffusions in latent space or pixel space. As the paper also identified, there can still be discretization advantages of using latent diffusion if the upper bounds are loose. I would like to clarify that this is the reason that I lean towards weak acceptance rather than clear acceptance (while still supporting the paper's contribution generally).
Questions for authors

I have no questions or concerns.

3 comments

We thank the reviewer for the positive assessment. Our main objective is to close the gap between theory and practice: existing discretization bounds scale with a dimension of the data, yet recent pixel-space models with suitable network design, such as SiD2 (Hoogeboom et al., 2025), JiT (Li & He, 2026) and PixNerd (Wang et al., 2026), approach the quality of latent diffusion in a space about 48 times larger. This suggests that sampling cost is governed by the structure of the data, not its dimension. Our bound makes this precise through one quantity, the mutual information IΓI_\Gamma. It recovers and improves the complexities of earlier analyses: the intrinsic dimension for data near a low-dimensional set (Theorem 3(c)), the metric entropy of the support plus half a nat (Corollary 3), and log⁡Kmix\log K_{\mathrm{mix}} instead of log⁡2(KmixN)\log^2(K_{\mathrm{mix}}N) for Gaussian mixtures (Corollary 6, Theorem H.1). It is strictly smaller whenever the data spread their mass unevenly.

We found the reviewer's point insightful: our comparison rests on upper bounds, and latent diffusion could keep a discretization advantage if they were loose. We address this with new experiments on real data and an extension of the theory.

1. New experiments on real data (MNIST and CIFAR-10). Total variation cannot be computed for real images, so we measure the discretization error with two quantities that our theory controls. The first is FID. The second is a coupled upper bound on W22W_2^2: for data in a bounded set of diameter BB, such as images, W22≤B2TVW_2^2 \le B^2 \mathrm{TV}, so our total-variation bound also controls W22W_2^2.

  • Pixel versus latent diffusion on MNIST (Section 7, Figure 1; Section I). We train the same diffusion model, with the same budget and the same sampler, in pixel space (d=784d = 784) and in latent spaces of two autoencoders (d=256d = 256 and d=128d = 128) and PCA (d=256d = 256). Each model is compared with its own samples at a large number of steps, so that training and reconstruction errors cancel and only the discretization error remains. In both measures, the discretization error in pixel space is close to that of the latent models: in MNIST-FID, pixel space needs 0.88–1.12× the steps of the latent models, whereas a dimension-based account predicts 3.1–6.1×. In the coupled bound on W22W_2^2, the four representations stay within a factor of about 1.5 of each other at every number of steps. Empirically, latent diffusion has no significant discretization advantage over pixel diffusion.
  • σ-aware sampling improves practical samplers (Section I, Figure 3). Our analysis suggests absorbing a Gaussian-like component of scale σ\sigma into the diffusion time, τk=tk+σ2\tau_k = t_k + \sigma^2, which shortens the log-SNR range LL of the discretization term L/KL/K. This is a one-line change that needs no retraining. On CIFAR-10, with a pretrained EDM model and σ=0.01\sigma = 0.01, the 50-step FID improves from 8.71 to 7.40 with one fewer network evaluation, and the coupled bound on W22W_2^2 decreases in proportion to LL (W22/L≈2.64W_2^2/L \approx 2.64 for every σ\sigma), as Theorem 2 predicts. On MNIST, the discretization error of the pixel model decreases at every KK (MNIST-FID 48.9 → 30.0 at K=10K = 10).

So the theory is now backed by experiments with trained networks.

2. A measurable complexity. Our bound depends on the data only through the mutual information IΓI_\Gamma, and this quantity can be measured. By the I–MMSE identity, IΓ=12∫0Γmmse(γ) dγI_\Gamma = \tfrac12\int_0^\Gamma \mathrm{mmse}(\gamma)\,d\gamma, and the denoising error of any trained model is never below the MMSE, so a pretrained denoiser gives an upper bound on IΓI_\Gamma. For ImageNet 256×256 (d=196,608d = 196{,}608), using a pretrained pixel-space model (ADM), we obtain IΓ≤13,600I_\Gamma \le 13{,}600 nats at noise variance δ≈0.01\delta \approx 0.01 per pixel, about 15 times smaller than the dimension. For δ≥0.215\delta \ge 0.215, IΓI_\Gamma is below 2,480 nats, the entropy bound of the LlamaGen VQ-16 tokenizer's code (Section I, Table 6). We should emphasize that this estimation is only upper bound and true value can be much smaller.

3. Removing the resolution restriction: approximate Gaussian mixtures. To make the theory independent of the target resolution, we extend it to every distribution that can be approximated by an isotropic Gaussian mixture, the setting of Li, Cai and Wei (2026) (Theorem H.1; Corollary 8 applies it to a code and its decoder). The discretization term becomes LK6.1I(J;X0G)\tfrac{L}{K} 6.1 I(J; X_0^G), in any dimension, where I(J;X0G)I(J; X_0^G) is the mutual information between the mixture label and the data; dd enters only through the approximation term. This rate is better than the best previously known result (Li et al., 2026), whose discretization term is log⁡2(KmixN)log⁡2N/N\log^2(K_{\mathrm{mix}}N)\log^2 N/N:

  • Mutual information instead of the number of components: I(J;X0G)≤H(J)≤log⁡KmixI(J; X_0^G) \le H(J) \le \log K_{\mathrm{mix}}, and it is much smaller when components overlap or weights are unbalanced.
  • Our worst case is the square root of theirs: log⁡Kmix\log K_{\mathrm{mix}} against log⁡2(KmixN)≥log⁡2Kmix\log^2(K_{\mathrm{mix}}N) \ge \log^2 K_{\mathrm{mix}}.
  • Weaker assumptions: only a second moment E∥μJ∥2<∞\mathbb{E}\|\mu_J\|^2 < \infty, instead of a uniform bound on all component means.
  • Infinite mixtures: countably many components are allowed, where any bound in KmixK_{\mathrm{mix}} is vacuous.
  • Any geometric grid with K≥2LK \ge 2L and explicit constants, instead of a specific schedule with unspecified constants.

For more details, see Section H and Table 4.

We hope these results make the practical implications of our analysis concrete, and we are happy to answer any further questions.

Thank you so much for addressing my concerns. I will retain my positive score for acceptance.

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

The paper bounds the time-discretization error of the posterior-mean exponential integrator, the Euler scheme of stochastic localization, by a single quantity: the mutual information between the data and its Gaussian observation at the final noise level. The KL bound follows Reeves & Pfister (2025). The main new result is a direct total-variation bound that shrinks linearly with the step size, rather than as its square root as Pinsker's inequality would give. A rate–distortion argument then bounds this mutual information through any code of the data, recovering bounds via code entropy, metric entropy, mixture label entropy, and VQ tokenizers. From this the paper concludes that pixel-space diffusion has the same discretization complexity as latent diffusion "whatever the number of pixels". Exact-TV experiments on one-dimensional mixtures and product-structured targets with exact scores show the error tracking the mutual information rather than the dimension.

Critical strengths
  1. The exact reverse step is a Gaussian mixture over the posterior, and the sampler replaces it with a single Gaussian with the same mean. A pointwise Jensen bound cancels the first-order error, and a one-step drop in estimation error converts the remainder into mutual information. No density, typical-set, or truncation argument is needed. The result improves the step count for a target TV accuracy from inverse-square to inverse, and the technique looks reusable.

  2. The rate–distortion bound replaces intrinsic dimension and covering numbers. It is never worse than the metric entropy plus half a nat, strictly better when probability mass is spread unevenly, and it covers singular data, which Gaussian mixtures cannot approximate in total variation. This makes the framework more general than existing dimension-free analyses.

Critical weaknesses

The claim that pixel-space diffusion "pays only for the latent, whatever the number of pixels" holds only at a resolution where the target is dominated by noise. Corollary 5 runs the sampler to a noise level at which each pixel's noise variance equals the code's total reconstruction error over all pixels. The output therefore carries roughly d times more noise than a decoded latent sample, so the comparison with latent diffusion is not like-for-like. At a matched resolution, the dimension re-enters whenever the reconstruction error is spread across pixels, as it typically is for real tokenizers. The paper's own per-pixel experiment shows this: 919 nats of mutual information and a TV of 0.51 at d = 1024. Since the abstract and Remark 1 present this comparison as the paper's main message, it needs to be restated, even though the bounds themselves appear correct.

Questions for authors
  1. Is my reading correct that the target noise level sets each pixel's noise variance equal to the total reconstruction error? If so, could you restate Corollary 5 and Remark 1 at the resolution where total noise energy matches the code's error, and use Corollary 4 to estimate the residual term for a trained tokenizer at that resolution?

  2. With the experimental settings, the leading term of the TV bound is about 0.5 times the mutual information. It is therefore vacuous (above 1) for most points in Fig. 1(a) and for every family in Fig. 1(b), including the low-rank family, whose actual TV is 0.027. The observed ratio of TV to mutual information is about 0.04. Is the linear dependence on mutual information tight in any regime? In the per-pixel family, TV grows like the square root of the dimension, and hence of the mutual information; does that suggest the true dependence is sublinear?

2 comments

We really appreciate your careful and constructive feedback. Your reading is correct: at δ=D\delta = D, each pixel carries noise variance equal to the code's total reconstruction error, so the comparison with a decoded latent sample is not like-for-like. We address this concern in two ways.

1. A finer resolution. Theorem 3(b), with log⁡(1+x)≤x\log(1+x) \le x and Γ=1/δ\Gamma = 1/\delta, gives in one line

IΓ≤H(Z)+d2log⁡(1+ΓDd)≤H(Z)+D2δ≤2H(Z)whenever δ≥D2H(Z).I_\Gamma \le H(Z) + \tfrac{d}{2}\log\big(1 + \tfrac{\Gamma D}{d}\big) \le H(Z) + \tfrac{D}{2\delta} \le 2H(Z) \quad \text{whenever } \delta \ge \tfrac{D}{2H(Z)}.

So it suffices to take δ=D/(2H(Z))\delta = D/(2H(Z))(you can see it as error per bit), which is 2H(Z)2H(Z) times finer than δ=D\delta = D, and the guarantee remains independent of dd. Keeping the logarithm gives the exact threshold D/(d(e2cH(Z)/d−1))D/\big(d(e^{2cH(Z)/d}-1)\big) for IΓ≤(1+c)H(Z)I_\Gamma \le (1+c)H(Z), and D/(s(e2cH(Z)/s−1))D/\big(s(e^{2cH(Z)/s}-1)\big) for a residual of rank ss, which is exponentially small in H(Z)/sH(Z)/s. We have revised Corollary 5, Remark 1, the abstract and the conclusion accordingly.

2. Removing the resolution restriction. To remove the resolution restriction altogether, we leverage the proof technique of the Gaussian-mixture analysis of Li, Cai and Wei (2026). We now obtain convergence for every distribution that can be approximated by an isotropic Gaussian mixture (Theorem H.1). Applied to a code and its decoder (Corollary 8): if εae=TV(L(X0),L(g(Z)+σG0))\varepsilon_{ae} = \mathrm{TV}\big(\mathcal{L}(X_0), \mathcal{L}(g(Z) + \sigma G_0)\big) and the sampler runs with τk=tk+σ2\tau_k = t_k + \sigma^2 down to δ=0\delta = 0, then

TV(L(X0),P^0)≤LK(6.1 I(Z;g(Z)+σG0)+2.2 RT+σ2)+12Sh (d+2log⁡(dK)) εae+εae+2eapx+R4(T+σ2)+8.4Sh(log⁡(dK)+1)K,\mathrm{TV}\big(\mathcal{L}(X_0), \widehat P_0\big) \le \frac{L}{K}\Big(6.1\, I\big(Z; g(Z)+\sigma G_0\big) + 2.2\,\frac{R}{T+\sigma^2}\Big) + 12\sqrt{S_h\,(d + 2\log(dK))\,\varepsilon_{ae}} + \varepsilon_{ae} + \sqrt{2 e_{\mathrm{apx}}} + \sqrt{\frac{R}{4(T+\sigma^2)}} + \frac{8.4\sqrt{S_h(\log(dK)+1)}}{K},

with I(Z;g(Z)+σG0)≤H(Z)I(Z; g(Z) + \sigma G_0) \le H(Z) and Sh≤4(L+1+log⁡(1/h))S_h \le 4(L + 1 + \log(1/h)). The target is the data itself, at full resolution: the discretization term depends only on the code, in any dimension, and dd enters only through the approximation term. For comparison, the best previous result (Li et al., 2026) is

TV≲log⁡2(KmixN)log⁡2NN+εscorelog⁡N+d εmix log⁡3/2(dN).\mathrm{TV} \lesssim \frac{\log^2(K_{\mathrm{mix}}N)\log^2 N}{N} + \varepsilon_{\mathrm{score}}\sqrt{\log N} + \sqrt{d\,\varepsilon_{\mathrm{mix}}}\,\log^{3/2}(dN).

Our bound has the same d ε\sqrt{d\,\varepsilon} approximation scale and improves on theirs in several ways:

  • Mutual information instead of the number of components: our complexity is I(J;X0G)≤H(J)≤log⁡KmixI(J; X_0^G) \le H(J) \le \log K_{\mathrm{mix}}. It is much smaller than H(J)H(J) when components overlap, and H(J)H(J) is much smaller than log⁡Kmix\log K_{\mathrm{mix}} when weights are unbalanced.
  • Our worst case is the square root of theirs: their dependence on the mixture, log⁡2(KmixN)≥log⁡2Kmix\log^2(K_{\mathrm{mix}}N) \ge \log^2 K_{\mathrm{mix}}, is the square of our worst case log⁡Kmix\log K_{\mathrm{mix}}.
  • Only a second moment: we need E∥μJ∥2<∞\mathbb{E}\|\mu_J\|^2 < \infty, whereas they require a uniform bound max⁡k∥μk∥≤NcR\max_k \|\mu_k\| \le N^{c_R} on all component means.
  • Infinite codes: since our bound is in terms of information, it covers countably infinite codes with finite entropy, where any bound in KmixK_{\mathrm{mix}} is vacuous.
  • Any geometric grid with K≥2LK \ge 2L and explicit constants, instead of a specific schedule with unspecified constants.

Your practical question (Corollary 4 for a trained tokenizer). Our sampling guarantees (Theorems 1–2) depend on the data only through the mutual information IΓI_\Gamma. Corollaries 4–5 make this mutual information interpretable, by bounding it through the entropy of a code and what the code misses. For real images, however, IΓI_\Gamma itself can be upper bounded directly: by the I–MMSE identity, IΓ=12∫0Γmmse(γ) dγI_\Gamma = \tfrac12\int_0^\Gamma \mathrm{mmse}(\gamma)\,d\gamma, and the denoising error of any model is never below the MMSE. Using a pretrained pixel-space diffusion model (ADM) on ImageNet 256×256 validation images (d=196,608d = 196{,}608), we computed this upper bound and compared it with the entropy bound of the LlamaGen VQ-16 tokenizer, H(Z)≤2,480H(Z) \le 2{,}480 nats (Section I, Table 6):

  • at the tokenizer's matched resolution δ=D/d≈0.01\delta = D/d \approx 0.01, IΓ≤13,600I_\Gamma \le 13{,}600 nats, about 15 times smaller than the dimension d=196,608d = 196{,}608 (about 0.07 nats per pixel);
  • IΓ≤2H(Z)I_\Gamma \le 2H(Z) for δ≥0.07\delta \ge 0.07.

These values are upper bounds; the true IΓI_\Gamma can only be smaller.

Tightness. As stated in our limitations, we do not expect the linear dependence on IΓI_\Gamma to be tight in general (Section G). For small II, the exact TV is linear in II with a common slope across families (TV/I\mathrm{TV}/I between 0.036 and 0.041). For larger II the dependence is sublinear, as you suggest: the TV saturates, and in the per-pixel family IΓI_\Gamma grows like dd while the TV grows like d\sqrt{d}. Obtaining a sublinear dependence in general is an interesting direction for future work.

We hope these changes address your concerns, and we are happy to clarify anything further.

Reviewer_SRnWOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the sampling complexity of diffusion models via the information the data perspective. It gives error bounds for data distributions with finite second moments, allowing the step count to depend on the information retained in noisy observations rather than directly on ambient dimension. A main contribution is a direct total-variation analysis with better step-size dependence than converting the KL bound. The paper connects this analysis to rate–distortion theory and illustrates the conclusions in controlled experiments. It also uses countable latent codes to compare pixel-space and latent-space diffusion.

Critical strengths
  1. The direct TV result is a meaningful theoretical contribution. It applies to general data with finite second moments and gives a better dependence on the sampling step size than one would obtain by converting the KL bound.

  2. The connection to compressibility gives a useful explanation of when ambient dimension can be misleading. The bounds capture differences in probability weights and residual spectra that support dimension alone would miss.

Critical weaknesses
  1. The pixel-versus-latent comparison is harder to clearly understand. It would be better to conduct experiments (as popular diffusion models using latent space) to support the results.

  2. The form of Lemma 1 is similar to the form of the Girsanov Theorem. It would be better to discuss the novelty of the technique.

Questions for authors
  1. It would be better to conduct experiments on the simple real datasets (for example, MINIST, CELEBA, ImageNet) to show the bound (performance) depends on the information instead of data low dimensioanl scture.
1 comment

We thank the reviewer for the positive assessment and the constructive suggestions.

1. Pixel-space versus latent-space diffusion. Our claim is that the discretization error is governed by the information the data carry at the target resolution, not by the dimension of the representation. Corollary 5 makes this precise: if a code ZZ has total reconstruction error DD, then for every terminal noise level δ≥D/(2H(Z))\delta \ge D/(2H(Z)) we have IΓ≤2H(Z)I_\Gamma \le 2H(Z), independently of the number of pixels.

To illustrate this, we conducted new experiments with trained networks on MNIST (Section 7, Figure 1, and Section I). We train the same diffusion model, with the same budget and the same sampler, in pixel space (d=784d = 784) and in three latent spaces: two autoencoders (d=256d = 256 and d=128d = 128) and PCA (d=256d = 256). To isolate the discretization error, each model is compared with its own samples at a large number of steps, so that training and reconstruction errors cancel. The discretization error is nearly the same in pixel space and in the latent spaces: to reach a given error, pixel space needs 0.88–1.12× the steps of the latent models, whereas a dimension-based account would predict 3.1–6.1×. Halving the latent dimension from 256 to 128 does not change the number of steps either.

To cover a more general case, we also extend our results to data whose distribution can be approximated by a Gaussian mixture, the setting of Li et al. (2026) (Theorem H.1). The discretization term then depends only on the mixture's information I(J;X0G)≤H(J)I(J; X_0^G) \le H(J), in any dimension, improving their log⁡2(KmixN)\log^2(K_{\mathrm{mix}}N) dependence. Applied to a code and its decoder, this gives a pixel-space guarantee at full resolution (Corollary 8).

2. Lemma 1 and Girsanov's theorem. Girsanov's theorem is used only in Theorem 1, the KL bound, which follows the approach of Reeves & Pfister (2025). Our main contribution is the total-variation bound (Theorem 2), and its proof does not use Girsanov's theorem at all: it does not go through a KL bound between path laws, which would only give h IΓ\sqrt{h\,I_\Gamma} via Pinsker's inequality. The proof technique is entirely different. It compares one step of the sampler with one exact step directly, through a pointwise lower bound on their density ratio (Lemma 2), a one-step MMSE-drop inequality (Lemma 3), and a killed-chain argument (Section C). This gives a bound linear in the step size hh.

We hope these clarifications address the reviewer's concerns, and we are happy to answer any further questions.

Comments on the paper

1 comment

We thank all reviewers for their careful and constructive reviews. As the discussion period ends today, we summarize how the revised PDF addresses the three main concerns raised.

1. Experiments on real data (Reviewers SRnW and rRtn)

  • Reviewer SRnW asked for experiments comparing pixel-space and latent diffusion. We trained the same diffusion model, with the same budget and sampler, on MNIST in pixel space (d=784d = 784) and in three latent spaces (two autoencoders with d=256d = 256 and d=128d = 128, and PCA). Measured by MNIST-FID and by a coupled upper bound on W22W_2^2, the discretization error is nearly the same in all representations: pixel space needs 0.88–1.12 times the steps of the latent models, against 3.1–6.1 times for a dimension-based account (Section 7, Figure 1, Section I). We also clarified that our main total-variation result (Theorem 2) does not use Girsanov's theorem.

  • Reviewer rRtn asked us to evaluate our bound for a trained tokenizer. We measured an upper bound on IΓI_{\Gamma} for ImageNet 256×256 directly, through the I–MMSE identity and the denoising error of a pretrained pixel-space model (Section I, Table 6). At the tokenizer's matched resolution, IΓ≤I_{\Gamma} \le 13,600 nats, about 15 times smaller than the dimension dd = 196,608.

2. Practical implications (Reviewer qB2k)

Although our contribution is theoretical, the analysis suggests a simple change to the sampler: absorbing a Gaussian-like component of the data into the diffusion time (σ\sigma-aware sampling). With a pretrained EDM model on CIFAR-10, this one-line change improves the 50-step FID from 8.71 to 7.40 with one fewer network evaluation; on MNIST it lowers the discretization error at every number of steps (Section I, Figure 3).

3. The resolution in the pixel-versus-latent comparison (Reviewer rRtn)

Reviewer rRtn raised a fair point about the resolution at which Corollary 5 compares pixel and latent diffusion. We addressed it in two ways:

  • A finer resolution: pixel-space sampling satisfies IΓ≤2H(Z)I_{\Gamma} \le 2H(Z) for every δ≥D/(2H(Z))\delta \ge D/(2H(Z)), where H(Z)H(Z) is the entropy of the code and DD its total reconstruction error. This is 2H(Z)2H(Z) times finer than before and independent of dd, with exact and rank-dependent thresholds (revised Corollary 5 and Remark 1).

  • No resolution restriction for Gaussian-mixture approximations: we extended the theory to every distribution that can be approximated by an isotropic Gaussian mixture (Theorem H.1, Corollary 8). The discretization term now depends on the mutual information between the data and the mixture label, which is at most its Shannon entropy H(J)H(J), instead of the squared logarithm of the number of components in Li et al. (2026). It also requires only a second moment and allows infinitely many components (Section H, Table 4).

We would be grateful if the reviewers could take these revisions into account in their final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.