acceptodds
Under review as a conference paper at ICLR 2027

Discrete Diffusion Models are Minimax Optimal Under Sparsity

Abstract

Discrete diffusion models have emerged as a leading alternative to autoregressive models for generating text and other categorical data, yet their theoretical foundations remain less developed than those of their continuous counterparts. For continuous diffusion models, a prominent line of theory suggests that these models can adapt to low-dimensional structure in the data, thereby improving statistical efficiency and generalization. In this work, we investigate the discrete analogue of this phenomenon by studying the statistical rates of discrete diffusion models under sparsity assumptions. By analyzing a nonparametric estimator derived from a regularized score-entropy objective, we establish minimax-optimal rates in the sparse regime. The proposed regularization can be viewed as an adaptation of label smoothing techniques introduced for classification, and does not require the knowledge of the sparsity parameter. This shows that discrete diffusion models can adapt intrinsically to sparsity.

Primary area:
generative models
Keywords:
discrete diffusion models, statistical learning theory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_PV5rOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This work studies whether discrete diffusion models can learn efficiently when only a small fraction of all possible sequences occur in the data. The paper introduces a smoothing method that assigns probability to unseen sequences, where the probability is determined by the number of distinct training sequences. The paper shows that masked and uniform diffusion can achieve statistically optimal estimation rates without knowing the true number of possible sequences.

Critical strengths
  1. Useful theoretical insight.

This work shows that both masked and uniform diffusion can learn sparse distributions under ideal conditions. This helps clarify what models can theoretically achieve.

Critical weaknesses
  1. Limited practical significance.

The paper does not show why diffusion is beneficial or whether trained neural networks achieve the same guarantees. The practical benefit remains unclear.

Questions for authors
  1. How can these results guide the training of diffusion models in the real world?
  2. Is there any evidence that the proposed regularization improves diffusion models trained on real-world datasets?
2 comments

We thank you for your review. The object of this paper is the theoretical study of discrete diffusion models, and we focused on the case of sparse distribution estimation, a class that seems realistic for many structured discrete data. Valid sentences, syntactically correct programs, valid molecules and functional protein sequences are all a vanishing fraction of the strings of the same length.

We agree that why diffusion is beneficial in practice, and whether trained neural networks achieve the same guarantees, are important questions (we are adding a result in which, for masked diffusion, we also show that the minimax rate is attained by an explicit ReLU--softmax denoiser whose size grows linearly with the sparsity, see our answer to Reviewer H4TK, although this construction does not cover SGD training). Our results address a question that comes before them: what discrete diffusion models can achieve statistically, through their own training objective and sampler. A trained network that approximately minimizes LML^M therefore targets a minimax-optimal law, and any remaining gap is due to approximation or optimization. Quantifying this gap for specific architectures is a natural next step, which we listed as future work.

Although our work does not aim to beat any benchmark, we believe it can help to better design these models. For example, it shows that the standard, unregularized objective is minimized by the empirical score, whose sampler only reproduces the training sequences and assigns zero probability to every unseen one. Regarding your questions, we answer point by point below.

  1. Our results suggest the following.
  • Regularize the objective rather than the output: LML^M is implemented as a simple data augmentation, replacing each training sequence by a uniformly random one with probability γ/(n+γ)\gamma/(n+\gamma), and it applies to both masked and uniform diffusion.
  • The amount of regularization can be chosen without knowing the sparsity of the data: γ=Dn\gamma=D_n, the number of distinct training sequences, attains the minimax rate adaptively (Corollary 4).
  • The guarantees are most informative when training sequences repeat, i.e.\ in data-constrained regimes and for short, highly structured sequences.
  • Since masked and uniform diffusion attain the same minimax rate, the minimax risk of the estimation problem cannot explain the differences observed in practice between the two kernels; these must come from approximation, optimization or sampling (see also our answer to Reviewer H4TK).
  1. We refer the reviewer to our answer to Question 5 of Reviewer 7xtj above.

Thank you for clarifying the training recipe and theoretical scope.

My practical-significance concern remains unresolved, and related concerns were also raised by Reviewers H4TK and 7xtj.

The explicit neural construction uses weights computed from sample counts and does not establish guarantees for networks learned through SGD. Also, the response describes promising text8 behavior but does not yet provide quantitative results.

As the practical benefit cannot be assessed, which is my main concern, I maintain my score.

Reviewer_H4TKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the statistical efficiency of discrete diffusion models under sparsity. The authors introduce a regularized score-entropy estimator that can be interpreted as context-dependent label smoothing in masked diffusion. The resulting score corresponds to a smoothed empirical distribution. The paper derives expected and high-probability KL bounds, shows that the masked reverse sampler exactly recovers this estimator, and extends the analysis to uniform diffusion with an additional finite-horizon error. With the adaptive choice γ=Dn\gamma=D_n, the paper argues that both masked and uniform diffusion achieve sparse minimax-optimal rates. Synthetic experiments are provided to illustrate the effect of the regularization and its use in neural score training.

Critical strengths

A main strength of the paper is the clean connection between regularized discrete diffusion training and classical smoothed distribution estimation. In particular, the interpretation of the masked score as context-dependent label smoothing and the characterization of the resulting reverse sampler are conceptually interesting.

I also appreciate that the paper studies both masked and uniform corruption within the same framework and provides explicit statistical guarantees for the terminal distributions.

Critical weaknesses
  1. For me the novelty of the main statistical result relative to Mourtada (2026) is somehow unclear. The adaptive estimator with γ=Dn\gamma=D_n, as well as closely related sparse and effective-support KL guarantees, already appear in Mourtada (2026). The main new component here seems to be showing how this estimator can be represented through a discrete diffusion score and reverse sampler. I think the current presentation makes the statistical contribution appear broader than it is. This is important because the sparse minimax result is presented as one of the central contributions of the paper.

  2. If I am interpreting the result correctly, I think the claim that "the results rule out a separation between masked and uniform diffusion" appears too broad. The minimax lower bound is over arbitrary distribution estimators and is independent of the corruption kernel or how the score is learned. From my reading, it seems that the paper constructs the exact score of essentially the same smoothed estimator under both kernels, so obtaining the same minimax distribution-estimation rate does not show that masked and uniform diffusion are equally difficult to learn from finite data with a restricted model. This distinction is particularly important given prior work showing different score-learning behavior for the two kernels. It appears to me that the result supports a narrower statement about unrestricted minimax distribution estimation, rather than a general absence of statistical separation between masked and uniform diffusion.

Questions for authors
  1. I would like the authors to clarify which parts of the sparse minimax analysis are genuinely new relative to Mourtada (2026). In particular, if one takes Mourtada's adaptive smoothed estimator and its KL guarantees as given, what statistical result in the present paper remains novel?

  2. It would be helpful if the authors could clarify the intended meaning of the statement that the results "rule out a separation between masked and uniform diffusion." Is this statement meant only at the level of unrestricted minimax distribution-estimation risk, or is it also intended to say something about the difficulty of learning the score with a finite neural model?

  3. If I understand the transformer experiment correctly, all training sequences are distinct, so Dn=nD_n=n. The proposed adaptive choice would therefore give γ=n\gamma=n, corresponding to a smoothing mixture weight of 1/21/2. It would be helpful if the authors could report results using exactly this adaptive choice? In particular, for n=8192n=8192, the reported grid does not seem to include γ=8192\gamma=8192.

  4. Since the theoretical comparison is between masked and uniform diffusion, I think the empirical section would be stronger with a neural experiment comparing both kernels under the proposed regularization. I may have overlooked such a comparison, but my reading is that the current experiments use only the masked kernel.

2 comments

We thank the reviewer for the suggestions and the time spent in reviewing our paper. We answer to your questions in order in the following. Below, Mourtada (2026) refers to the paper cited in the original submitted version, Q to "Question" and W to "Critical Weaknesses"

Contributions compared with Mourtada (2026) (Q1 and W1).

While our work builds on the results of Mourtada (2026), we believe that their adaptation to our setting is far from straightforward and constitutes the main contribution of our paper. More broadly, a substantial part of theoretical machine learning research consists of building on established mathematical results and developing new techniques to apply them to problems for which they were not originally designed. The novelty of such contributions lies not necessarily in introducing entirely new theoretical tools, but in overcoming the mathematical and methodological challenges involved in making rigorous, nontrivial connections to existing mathematical results. We believe that our work follows this perspective. Let us clarify our contributions more precisely.

  • For masked diffusion, the score factorizes into a time-dependent factor and a context-dependent classifier (11). Building on this, in Lemma 1 we show that, among label-smoothed scores, the proper ones are exactly those with εc\varepsilon_c as in (14), and that they are the scores induced by μ^0(γ)\hat\mu_0^{(\gamma)}.
  • The regularized score is the minimizer of a training objective, the mixed loss LML^M, which is the score-entropy loss under the mixture with the uniform law. For masked diffusion, this shows that context-dependent label smoothing is equivalent to regularization by the uniform loss (Lemma 9); the mixed loss and its explicit minimizer (24),(25) apply to any kernel (Lemma 18), while the label-smoothing interpretation is specific to the masked case.
  • We show that the masked reverse sampler started from the fully masked state outputs μ^0(γ)\hat\mu^{(\gamma)}_0 exactly (Lemma 4). Note that this is non-trivial: the sampler starts from the fully masked state rather than μ^T\hat\mu_T, and a label-smoothed denoiser is in general not a proper score. Lemmas 1 and 4 show that, for the smoothing (14), the output is nevertheless exact for every TT, whereas uniform diffusion pays an extra 1/n1/n (Lemma 5).
  • For uniform diffusion, no classifier factorization is available, but the mixed loss is still minimized by the exact score of μ^0(γ)\hat\mu_0^{(\gamma)} (Lemma 18). The only additional error comes from initializing the sampler at the uniform distribution instead of μ^T\hat\mu_T: we show that, for σˉT≥log⁡(dKn)\bar\sigma_T \ge \log(dKn), KL(μdata ∣ ν^)≤KL(μdata ∣ μ^0(γ))+1/n\mathrm{KL}(\mu_{\mathrm{data}} \,|\, \hat\nu) \le \mathrm{KL}(\mu_{\mathrm{data}} \,|\, \hat\mu_0^{(\gamma)}) + 1/n (Lemma 5), so the same rates hold (Theorem 2).
  • We extended Mourtada's final bounds, which are stated for specific smoothing levels (in particular the adaptive choice γ=Dn\gamma=D_n), to hold for an arbitrary data-dependent γ\gamma (Lemma 2), with a single high-probability event valid simultaneously for all γ\gamma (Lemma 3), and we identified a range of smoothing levels attaining the sparse minimax rate (Corollary 2). This is motivated by how, in the diffusion model setting, hyperparameters are tuned based on experimental results and not directly as a function of the dataset.

To conclude, we believe that both in the introduction and throughout the theoretical developments, we clearly explain how our work relates to Mourtada (2026) and explicitly identify which results are borrowed from or closely related to previous work.

We appreciate that our discussion of the connections with Mourtada (2026) may not have sufficiently emphasized the additional developments required in our setting. While our approach builds on these existing results, their adaptation involves substantial theoretical and methodological challenges that constitute the main contributions of our work. We thank the reviewer for bringing this point to our attention and will revise the manuscript to better highlight these contributions and clarify how our approach differs from and extends Mourtada (2026) .

We hope that our response has convinced the reviewer that our work represents an important contribution beyond Mourtada (2026) . Should any concerns remain, we would be happy to provide further details clarifying how our approach differs from and extends this reference.

The comparison between the masked and uniform setting (Q2 & W2).

We understand the concern about the sentence ``the results rule out a separation between masked and uniform diffusion'' in the conclusion paragraph.

The statement is intended at the level of the unrestricted minimax distribution-estimation risk over ll sparse distribution: in the Conclusion it is qualified as holding ``at the level of the minimax risk over sparse distributions'', and we explicitly leave open whether the gap observed in practice comes from approximation or optimization of the score. We agree that the short form can be misread and so we will remove it from the conclusion. However, for masked diffusion we can now go beyond unrestricted estimators: we can construct an explicit ReLU--softmax denoiser of size O(dl+dK(dlog⁡K+log⁡2n)(d/ε)1/4)O\big(dl+dK(d\log K+\log 2n)(d/\varepsilon)^{1/4}\big), linear in the sparsity, whose terminal law νθ\nu_\theta satisfies E KL(μdata∣νθ)≲llog⁡(eKd/l)/n+ε\mathbb E\,\mathrm{KL}(\mu_{\mathrm{data}}|\nu_\theta)\lesssim l\log(eK^d/l)/n+\varepsilon, i.e. it attains the minimax rate. Its weights are computed from the sample counts, so this does not cover SGD training. The uniform case remains open; we will add the construction in the appendix and state this explicitly.

Choices of γ\gamma in experiment (Q3).

We have since updated all experiments to consider a range of values for γ\gamma, taken as fractions of the adaptive choice DnD_n. We agree that this would be more helpful, in particular, for understanding the independent role of nn and γ\gamma in the outcome of the experiments. Moreover, the new grid includes γ=Dn=n\gamma=D_n=n (e.g. γ=8192\gamma=8192 for n=8192n=8192).

Experiments for the uniform kernel (Q4 & W2).

We intended to focus our work on the masked case, hence why the experiments consider this case only. However, we agree with the reviewer that the uniform setting is of interest, particularly for comparing the nature of smoothing and the mixed loss between the two settings. We are now including the results for this experiment in the appendix, which will be available later during the discussion period.

Reviewer_7xtjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose a method for regularizing scores in the setting of masked discrete diffusion. Building on the view that the score in masked diffusion can be factorized into a time dependent scalar and a context-conditioned classifier, they suggest a context-dependent label smoothing. They then connect this regularized score to a score-entropy loss with a regularizer term based on a uniform distribution. Next, they give bounds on the KL between the data distribution and smoothed distribution induced by the regularized score. They adapt existing results on discrete density estimation to derive upper bounds on KL divergence (both in expectation and tail probability) show that this regularized score gives a minimax rate optimal estimate of the data distribution. However, this is only true assuming the data distribution is sparse, and that the parameter γ\gamma governing the regularization strength has been chosen properly to match the sparsity. Limited experiments, test the validity of the theory, and whether the regularized estimates of the data distribution can be useful in practice.

Critical strengths

1- Connecting the minimax optimality results for discrete density estimation to discrete diffusion is an interesting idea that highlights the promising potential of discrete diffusion models (that are commonly under-appreciated compared to their autoregressive counterparts) from a theoretical perspective.

2- The paper is theoretically well-motivated, as naive bounds that disregard the sparsity of discrete data distributions are typically not too useful for theoretical or practical insight. Furthermore the experimental results, although limited, seem to suggest that the theoretical results have practical implications.

3- Although the paper is somewhat dense and hard to follow in parts, I appreciate the authors attempting to give intuition where they can, for example after defining the effective support size keffk_eff.

Critical weaknesses

1- The point of the mixed loss LML_M is unclear to me. My understanding is that it leads to the label-smoothed score, which in turn induces the regularized distribution μ^(γ)\hat \mu^{(\gamma)}. If this is the case, then what is the point of this loss since we can already implement the label-smoothed score without retraining? If it is not the case, then could the authors explain it more clearly?

2- With both of the experiment settings, the connection to the theory is not particularly well-explained. For example, I understand from Figure 1, that the KL to the true data distribution (sometimes) has a U-shaped dependence on γ\gamma. Is there anything beyond this to glean from the first set of experiments? Since ll is known, is the true minimizer γ=l\gamma = l? Also the purpose of the experiments where learning with mixed losses is done is not fully clear. (See weakness 1)

3- Section 3 and 4 are quite dense and hard to read. Relegating some of the Lemmas and results that are not crucial to understanding the main theorems to the appendix to leave more room for intuition building could dramatically improve its readability.(e.g. Lemma 4, or merging Lemmas 2 and 3)

4- Although I find the theory motivating and insightful, at least one or two more realistic experiments (even if on small models), for example on text datasets, would go a long way to convince me of the potential practical implications of the theory.

Questions for authors

1- Related to strength 1, a discussion on whether any similar results exist for autoregressive models would help better contextualize this work in the literature, and potentially further highlight the significance of the current work. Could the authors briefly talk about this?

2- My understanding is that training with LML_M yields the label-smoothed score, which in turn induces the regularized distribution μ^(γ)\hat\mu^{(\gamma)}. If so, what does LML_M add, given that the label-smoothed score can be implemented directly from a standard trained score without retraining? If this understanding is wrong, could the authors clarify what LML_M achieves that post-hoc smoothing cannot? Also what do the experiments that train with mixed losses demonstrate that the experiments using the label-smoothed score directly do not? Is there a measurable difference between the two approaches?

3- Figure 1 shows that the KL divergence to the true data distribution can have a U-shaped dependence on γ\gamma. Beyond this, what should the reader take away from these experiments? In particular, since ll is known in this setting, is the optimal value γ=l\gamma = l, and do the empirical minimizers agree with what the theory predicts?

4- Could the authors move non-essential results (for example Lemma 4) to the appendix and possibly merge Lemmas 2 and 3, using the freed space for intuition about the main theorems and how they connect to the experiments?

5- Could the authors include one or two more realistic experiments, even on small models, for example on text datasets, to show that the theoretical predictions, such as the existence of a beneficial γ\gamma, carry over to practical settings?

If these questions and weaknesses are sufficiently addressed I am open to raising my score as I find the contribution of the paper interesting!

2 comments

We thank the reviewer for their interesting questions and suggestions which we now address. We will provide an updated version of the paper later into the rebuttal period and throughout the following, numbering refers to the original submitted version. Below, Mourtada (2026) refers to the paper cited in the original submitted version, Q to "Question" and W to "Critical Weaknesses"

Results for AR models (Q1).

To the best of our knowledge, minimax guarantees in the sparse setting are not available for AR models, nor, before this work, for discrete diffusion models. Extending this type of analysis to AR models is an interesting direction which we will suggest as future work in the updated version.

The purpose of the mixed loss LML_M (Q2 and W1).

Your understanding is correct for an unrestricted minimizer: in that case LM\mathcal{L}^M and post-hoc smoothing via equation (15) yield the same score s^(γ)\hat s^{(\gamma)}. However, they differ in computational cost at sample time, as post-hoc smoothing requires the counts ncn_c of every context visited during sampling which requires iterating over the entire dataset at each sampler step. On the other hand, training with LM\mathcal{L}^M acts as a form of data augmentation with no additional cost at sampling. LM\mathcal{L}^M acts as a bridge between the smoothing method we develop in our theory and how discrete diffusion models are used and trained in practice (typically without access to the dataset at sample-time). In the experiments, we investigate if LM\mathcal{L}^M might provide some generalization benefit before convergence (and therefore, before its relationship to post-hoc smoothing) which is shown in Figure 2. We agree with the reviewer that a comparison of post-hoc smoothing versus the introduction of LM\mathcal{L}^M during training is interesting and we will add this comparison.

Meaning of experiments and optimality of γ\gamma (Q3 and W2).

Figure 1 computes KL(μdata∣ν^)\mathrm{KL}(\mu_{\mathrm{data}}|\hat\nu) in closed form (Lemma 4), isolating the statistical error. As predicted by Theorem 1, the error is U-shaped in γ\gamma (the theory predicts a tradeoff between missing mass keffoLγ/nk^o_{\mathrm{eff}}L_\gamma/n vs.\ bias γ/n\gamma/n); the error scales with ll rather than KdK^d; and the adaptive choice γ=Dn\gamma=D_n lies close to the minimizer. We will use the additional space in the updated version of the paper to provide a longer explanation of this figure. To answer the second question: the choice γ=l\gamma = l is not optimal for every μdata\mu_{\mathrm{data}}. Corollary 2 shows that any γ∈[c0l, C0 lΛl]\gamma \in [c_0 l,\, C_0\, l\Lambda_l] achieves the upper bound lΛl/nl\Lambda_l/n, which matches the lower bound of Mourtada (2026) up to universal constants. In other words, γ∼l\gamma \sim l is a worst-case choice over the class of ll sparse distributions, and not necessarily the optimal one for a given distribution. Indeed, for deterministic γ\gamma, the numerator of the bound in (18), keff+γ+keffo(μdata,n/8) log⁡eKdγ,k_{\mathrm{eff}} + \gamma + k^o_{\mathrm{eff}}(\mu_{\mathrm{data}}, n/8)\,\log\frac{eK^d}{\gamma}, is minimized at γ=keffo(μdata,n/8)\gamma = k^o_{\mathrm{eff}}(\mu_{\mathrm{data}}, n/8), which can be much smaller than ll and is unknown in general. This is precisely why we propose the adaptive choice γ=Dn\gamma = D_n: since DnD_n concentrates around keff≤lk_{\mathrm{eff}} \le l (Lemma 17), it attains the minimax rate without knowledge of ll or keffok^o_{\mathrm{eff}}. We thank you for pointing it out: we have clarified this discussion in the revised version.

Simplifying exposition of theory (Q4 & W3).

We merged Lemmas 2 and 3 and moved Corollary 1 to the appendix, and used the additional space to expand on the intuition behind Definition 1 and Theorem 1, and to explain better the connection to the experiments. We kept Lemma 4 in the main text, since it is the bridge between the regularized score and the statistical rates: it shows that the masked sampler outputs μ^0(γ)\hat\mu_0^{(\gamma)} exactly. However, we have shortened the statement to make it more suitable for the main body.

More realistic experiments (Q5 & W4).

We thank the reviewer for their suggestion which has motivated us to include a text and Sudoku experiment which will be available later into the rebuttal period. We use the transformer of Section~5 trained on character 12-grams from text8 over many epochs. With γ=0\gamma=0 the held-out NELBO deteriorates as the model memorizes, while γ>0\gamma>0 mitigates this. We also include an experiment on the Sudoku dataset, a natural test since valid grids are a vanishing fraction of the 9819^{81} sequences and validity can be checked exactly. We believe it is a well-suited test of whether a beneficial γ\gamma exists in practice, because it matches the structural assumption of the paper. We will include experiments for text8 on 256-grams, though we will not be able to provide this during the rebuttal period and so we will include it in the camera-ready version.

I thank the authors for their detailed rebuttal and some of my concerns are partially addressed. I am inclined to raise my score, however this is dependent on the changes that the authors claim to have made in the paper and I will reserve my final judgment until I see the revised version with the clarity/readability improvements and new results promised by the authors in their rebuttal. I would urge the authors to upload the revised pdf soon so further potential clarifying concerns can be resolved during the discussion period.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Updated version

Dear reviewers,

We uploaded the new version of the paper. In the following we summarised the new results.

  1. First, we proved that for masked diffusion, the minimax-optimal label-smoothed score is computed exactly by a ReLU network with an ℓ1\ell^1 head and O(dK(Dn+dK))O\big(dK(D_n+dK)\big) parameters, Dn≤min⁡(n,l)D_n\le\min(n,l) (Proposition 1). Unlike the softmax version announced earlier, the ℓ1\ell^1 head makes this exact, so the sampler's law is exactly μ^0(γ)\hat\mu_0^{(\gamma)}. Moreover, every global minimizer of the regularized loss over this class induces the same score (Corollary 7): with γ=Dn\gamma=D_n, it is minimax optimal, leaving optimization as he only open source of error. The uniform case remains open.
  2. For the sake of clarity, we merged the old Lemmas 2 and 3, moved Corollary 1 to the appendix, and made Lemma 4 shorter. Moreover, we provided intuitions of the concepts, like in Definition 1 and of the result in Theorem 1 immediately after their statement.
  3. We rewrote parts of the introduction, like the contributions, to make our work clear and concise and we rewrote the conclusion, adding our new results and leaving open questions not touched in this paper like minimax optimal guarantees for AR models.
  4. We have included new experiments on a more realistic text setting as well as a sudoku setting. We rewrote the experiments section, trying to explain better connections to the theory. We have also added some new experiments to the appendix comparing post hoc smoothing to training with the mixed loss, as well as a short study on the uniform kernel in the Markov chain setting (section D.5). Due to the limited time and slower convergence (see section D.5) we have not had time to run the uniform model on the new settings of text8 and Sudoku. We will have this prepared for the camera ready version.

We hope that this addresses your remaining concerns and that you consider increasing your score in light of this.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.