acceptodds
Under review as a conference paper at ICLR 2027

Timescale Identification from -Step Transition Operators: The Structural Entropy Spectrum

Abstract

Agents and world models discretise experience into a transition graph at one fixed step, yet behaviour has structure at several timescales. Which step should the graph be built at, and which timescales are present? Multiscale tools (Markov stability, the map equation at Markov time, implied timescales of Markov state models) sweep a time parameter but attach no significance statement to the scales they report. We make the step count k an argument of a clustering objective on the k-step operator P^k and calibrate it against a matched null: chains whose stationary flow matrix has identical margins and dwell (self-loop) mass. Our theory uses structural entropy (SE), a two-level code-length objective: the structural entropy spectrum SE(k), its null SE_null(k) and the calibrated spectrum ΔSE(k) = SE_null(k) − SE(k). We prove that the k-dependence factors through an exact identity, SI(k;C) = Σ_a F_aa(k) log2(1/V_a) (k-step module persistence × module code length), prove that when metastable blocks have stationary volume < 1/2 the metastable partition is the unique global optimizer of the idealized in-window operator, and sketch a Simon–Ando extension to nearly uncoupled chains on a window whose endpoints scale as 1/(1−λ) for the fast and slow eigenvalue clusters, so the optimal partition path coarsens through a planted hierarchy in order. On a 36-state three-level chain, the path recovers both planted levels in order at calibrated effect sizes up to 1.8 bits, although at this fixture's modest scale separation ΔSE(k) shows no plateaus; a single-scale expander control is featureless; the spectrum is estimable from a single trajectory. On latent transition graphs from a pixel-based JEPA world model, the one-step graph's calibrated effect falls below the preregistered 0.1-bit floor, while the spectrum peaks reproducibly at k* = 11–16 with 1.6–1.9 bits across five training seeds; a self-loop control shows that this structure survives removal of dwell but that the peak location is a dwell-inflated step count (two state changes). The calibration is not SE-specific: wrapped around the map equation it passes the planted-hierarchy, single-scale and structureless controls, and around Markov stability all but the expander control; both recover two balanced blocks, SE's boundary case, and the scale boundaries move with the objective. All decisive experiments were preregistered with pass/fail gates and are reported verbatim.

Primary area:
learning on time series and dynamical systems
Keywords:
structural entropy, Markov chains, timescale identification, metastability, spectral methods, world models

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_t7wGOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper solves a problem of multi-scale clustering that arise in Markov-chains as they evolve in time. A chain may remain within small clusters over short times, while larger groups of clusters emerge over longer times. The authors identify these groupings by applying a structural entropy clustering objective to the kk-step transition matrix PkP^k, varying kk. They compare the resulting scores against randomized chains that preserve the stationary distribution and self-loop probabilities, to assess whether the detected structure exceeds this baseline.

Critical strengths

None

Critical weaknesses

The paper is not very well written. There is a lot of superfluous technical detail that makes it very hard to follow. I suspect that large parts of it are entirely AI generated. While the results seem to be interesting, the paper is almost unreadable.

Questions for authors
  1. What is reasoning behind subtracting the null baseline from the structural entropy? I understand that this allows the authors to interpret the calibrated quantity as a kind of additional compression induced by the markov chain when compared to the null chain. But, is this subtraction required for clustering?

  2. What precisely is meant by an identified “timescale” in this framework? I understand that changing (k) reveals different optimal partitions. However, the locations at which these partitions change depend on the clustering objective. How should we distinguish a physical relaxation timescale of the Markov chain from a scale selected by the particular compression objective?

  3. What motivates the choice of structural entropy over other clustering objectives? The paper demonstrates that the same null calibration also works with the map equation, which additionally recovers the balanced two-block example where structural entropy fails. Does structural entropy offer a particular advantage for identifying dynamical organization, or is the main contribution the calibration framework, independently of the chosen objective?

1 comment
AuthorsOpenReview ↗

On the writing, and answers to your three questions

Thank you for the review, and for three questions that go to the centre of the paper.

The writing. You are right that the paper is hard to read, and Reviewer mGEB says the same of the experimental sections. Those sections put several numbers into most sentences, label the experiments with codes instead of saying what they test, and keep detail in the main text that belongs in an appendix. On AI: yes, AI tools were used to draft and edit parts of the text. The AI Use Statement before the references lists every use, including help with proofs and code, and we are responsible for all of it. It also says every proof was checked line by line; Lemma 5 was nevertheless wrong as stated (found by Reviewer bjbx, corrected in our reply there).

1. Why subtract the null, and is it needed for clustering? It is not needed for clustering. The partition at each step count kk is found by minimising structural entropy on PkP^k alone (Definition 1); the null plays no part in choosing it. The subtraction answers a different question: is this partition worth reporting at this kk? The raw saving cannot answer that, because it is positive even for a chain with no structure: a walk that lands in the same module twice by chance still saves code length (Lemma 3; about 0.53 bits for three equal modules). The null chains have the same stationary distribution and so the same chance level, which cancels in the difference. The difference tends to zero once the chain and its null chains have mixed, and it stayed below 0.1 bits on the two structureless graphs of the main text (Sections 4.2 and 4.6); the one with heavy self-loops reached 0.10–0.19 bits once they were removed (Appendix D). Figure 1(d) is the example: at k=1024k = 1024 the optimizer still returns the three planted super-blocks, but the calibrated value is 0.00, so that scale is not reported.

2. What is a "timescale" here? On the planted chain it is a range of step counts over which one partition is the optimum and significant (it beats every null draw, and the null average by at least 0.1 bits). On the world-model graphs, where the partition coarsens smoothly and no range is stable, we used the word for the step count at which the calibrated value peaks (Section 4.6). Neither is a relaxation time. Relaxation times belong to the chain alone and are read from its eigenvalues, with no clustering objective involved (0.9, 10.7 and 56.9 steps for the planted chain with its 0.01 teleport, Section 4.1). You are right that where the partition changes also depends on the objective. On that chain structural entropy keeps the nine fine clusters up to k=11k = 11 and has handed over to the three super-blocks by k=23k = 23, that is, at 1.03 to 2.1 times the middle relaxation time; Markov stability and the map equation keep the nine only to k=4k = 4–66 (Sections 4.1 and 4.5). So the boundaries lie, on this chain, within a small factor of the relaxation times, but they are not equal to them, and the factor depends on the objective. All three objectives return the two planted partitions, in the same order. To tell the two apart: the eigenvalues give the relaxation times; the spectrum adds which partition goes with each range of steps and whether it is significant.

3. Why structural entropy? The calibration is the main contribution, and it is not specific to structural entropy. The paper's own comparison shows this (Section 4.5, Table 2): under the same null the map equation passes every test in that table, including two equal blocks, where structural entropy fails by construction (5/5 against 0/5); Markov stability fails the expander control (3/5) and did not pass from trajectories of up to 3,000 steps; it passes the rest. We did not show that structural entropy is better at finding dynamical structure; in one case it did better than both alternatives (the planted chain with self-loops removed, Appendix D). We used it for two reasons. It already builds state abstractions in reinforcement learning (Zeng et al., 2023) and had no notion of timescale. And it allows exact statements: its saving equals kk-step persistence times module code length (Lemma 1), which explains why it holds the fine level longer and gives, for an idealized chain with no leakage between blocks, a sharp condition for the blocks to be the optimum (each below half the stationary mass; Lemma 4, Remark 1).

We have not uploaded a revised PDF: a rewrite, a full proof and further checks of the null are more than we can do properly before the deadline, so the corrections are stated in these replies.

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

This manuscript is about trying to identify time scales kk on which Markov chains show interesting structure of their states. At a rough, hand-waving level, the notion of interesting structure here is something like "there are clusters of states in which the chain stays persistently, with rarer transitions to other clusters". This is an idea with a long past behind it (*). There are two, conceptually distinct, innovations here:

  1. A specific test statistic ("structural entropy") is used to assess whether there is, in fact, interesting structure revealed by some partition of the states.
  2. Assessment of statistical significance against a null model, namely a random Markov chain which preserves both the invariant distribution and the self-transition probabilities (**).

I think that, of these two, the null model (2) is actually more interesting than the "structural entropy" test statistic. SE isn't bad, but it is not obviously more elegant or convenient than a range of other information-theoretic measures, and in particular it does not seem to work much better, or much differently, than the "map equation" of Rosvall and Bergstrom (see especially lines 358--369). The idea of comparing against a random Markov chain which, however, is tweaked to match the real chain's invariant distribution and self-transition rates is however more novel, and has potentially broader applications. It does, however, raise two questions:

a. Why match both the invariant distribution and the self-transition rates? How different would things be with random Markov chains that only matched one attribute or the other? (I realize that the second question might only be answerable with new experiments, which are not reasonable to ask for here, but some clear motivation for the choice of invariants does not seem like too much to ask.) b. How, precisely, are the entries in the Φ\Phi matrix randomized? I have diligently read the MS. multiple times, to try to determine this, without success. It's possible that the distributioon is "uniform on all matrices with the given marginals and diagonals", but if so that should be stated explicitly, and justified.

As is usual with LLM-written papers, the writing is needlessly unclear. The experimental sections, in particular, are almost impossible to follow.

*: Much longer than one would know from the references given here.

**: What the MS says (lines 155-157) is that Φnull\Phi_{\mathrm{null}} matches the row sums, columns sums, and diagonal of Φ\Phi, the matrix which gives the stationary joint distribution over successive time-steps. Since the row sums and column sums are both π\pi, the invariant distribution, and Φaa=πaPaa\Phi_{aa} = \pi_a P_{aa}, the diagonal entries of PnullP_{\mathrm{null}} must equal those of PP.

Critical strengths

See above.

Critical weaknesses

See above.

Questions for authors

See above.

1 comment
AuthorsOpenReview ↗

How the null chains are drawn, why both constraints, and structural entropy versus the map equation

Thank you for the careful reading. Your footnote (**) is exactly right, and we agree with your judgement of where the contribution lies.

(b) How the entries of Φ\Phi are randomized. You could not find it because it is not in the paper: Definition 1 gives only the constraints. It is not a uniform draw over the matrices with those margins and that diagonal. The procedure is in the code supplement and belongs in Definition 1:

  1. Start from the observed Φ\Phi (every null draw starts from it again).
  2. Pick two off-diagonal cells that currently hold positive flow, (a,b)(a,b) and (c,d)(c,d), independently and uniformly at random.
  3. If they share a row or a column, or if a=da=d or c=bc=b (the move would put flow on the diagonal), do nothing.
  4. Otherwise move m=min⁡(Φab,Φcd)m=\min(\Phi_{a b},\Phi_{c d}) out of those two cells and into (a,d)(a,d) and (c,b)(c,b).
  5. Repeat steps 2–4 for 20 proposals per off-diagonal cell, skipped ones included, then set Pnull=diag(π)−1ΦnullP_{\rm null}={\rm diag}(\pi)^{-1}\Phi_{\rm null}.

Every move keeps each row sum, each column sum and the whole diagonal. A null draw is therefore the state of this random walk after a fixed number of steps from the observed matrix. We have not characterised its limiting distribution, we do not claim uniformity, and we have no convergence diagnostic for it. (Limitation (iii) concerns how fast each null chain mixes over states, not this sampler.) So each p-value is the rank of the data among these draws, not an exact test under a stated law, and another ensemble with the same margins and diagonal could give different values. A move empties at least one cell and can fill empty ones, so the zero pattern is not preserved (with the 0.01 teleport, Φ\Phi has no zeros to begin with). We agree that the law has to be stated and justified in the paper itself.

(a) Why match both. The stationary distribution fixes H1H_1 and the volume of every module, and with them the saving that any partition gets by chance (Lemma 3). With π\pi matched, that chance part is the same in the real and the null chain and cancels, so the calibrated value tends to zero once the chain and its null chains have mixed. The self-transition rates are matched because of the data we had in mind: in the world-model graphs 81–88% of one-step transitions are self-loops (Section 1). A null that kept π\pi but spread the diagonal mass over other cells would be far less "sticky" than the data, and we expect almost any partition to beat it at small kk simply because states persist. Our question was what structure there is beyond margins and dwell, so both are held fixed.

What changes with only one constraint: we have not run a null that matches only π\pi or only the diagonal, so we cannot give numbers. The nearest evidence in the paper is a structureless expander given the self-loop mass of the world-model graphs (0.85): against the two-constraint null it reaches 0.02–0.05 bits, below the 0.1-bit floor, where the world-model graphs reach 1.60–1.93 (Sections 4.5 and 4.6). So heavy dwell alone produced only a small effect there for structural entropy (the map equation exceeded its floor in one of the five instances, Table 2). That says nothing about what a one-constraint null would report.

Structural entropy versus the map equation. We agree that it was not shown to be better, and Table 2 is the evidence: under the same null the map equation passes every test in that table, and it recovers two equal blocks where structural entropy cannot (5/5 against 0/5). The two do differ in where the partition changes: structural entropy keeps the fine level to k=11k=11, the map equation to k=4k=4–66. The null calibration is the main contribution. Structural entropy is the objective we started from and analysed in most detail (the identity of Lemma 1; the sharp condition of Lemma 4 on the idealized operator). It did better than both alternatives in one case (self-loops removed, Appendix D) and is the only one that fails on two equal blocks.

History (*) and writing. Both points are fair: the history is thin and the text is hard to follow. The related work goes back only to Simon and Ando (1961), names Courtois without a reference entry, and credits the wider literature on metastability and on aggregating Markov chains only through a few citations. On the writing: AI tools were used to draft and edit parts of the text; the AI Use Statement before the references lists every use, including help with proofs and code, and we are responsible for all of it. It also says every proof was checked line by line; Lemma 5 was nevertheless wrong as stated (see our reply to Reviewer bjbx). The experimental sections are dense and labelled with codes where they should be plain.

We have not uploaded a revised PDF: a rewrite, a full proof and further checks of the null are more than we can do properly before the deadline, so the corrections are stated in these replies.

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

My understanding of this paper is that the authors tried to identify useful timescales by clustering the k-step transition operator of a Markov chain. It aims to optimize a two-level structural entropy objective at each k, then compares the result with randomized chains that preserve the stationary flow margins and per-state self-loop mass. The resulting calibrated spectrum and partition path were then used to report significant scales.

The authors then derived an identity that related structural-information savings to module persistence and code length, and then gave an idealized recovery result for metastable blocks with stationary volume below one half.

They then sketched an extension to nearly uncoupled chains. The experiments conducted in the paper covered cases like planted chains, several controls, estimation from a trajectory, comparisons with other clustering objectives, and latent graphs from a world model.

Critical strengths
  1. I find that the proposed persistence-code-length identity useful. It helps to makes clear what structural entropy rewards, and why even an independent chain can have positive raw structural-information savings. Also, comparing the optimized score with a matched reference seems like a a sensible approach to that issue to me. The connection to code-length-weighted Markov stability also helps explain why the objectives choose different partitions, including the balanced two-block boundary case in the paper.

  2. The experiments conducted include useful checks beyond a single successful example. On the planted chain the proposed method recovers the nine-cluster and three-super-block partitions in order, and full recovery from a trajectory holds in all five seeds from 3,000 steps. I also appreciate the reported failures and limitations in the paper, where the plateau test fails its tolerance and the expander result depends on the effect-size floor. These details were helpful in clarifying the boundary of the proposed method.

Critical weaknesses
  1. I think a full perturbation argument for the main global-optimizer guarantee woud be needed. The idealized result in Lemma 4 in the paper is useful, but Theorem 1 seems to depend on transporting that result to a nearly uncoupled chain. Then Lemma 5 is stated for "any two stochastic" matrices, with its corresponding proof fixing the reference weights and module volumes. If each chain uses its own stationary distribution, then the bound would fail.

For example, with singleton modules at k=1, P=[[0.99,0.01],[0.01,0.99]] and Q=[[0.99,0.01],[0.02,0.98]] have stationary distributions (1/2,1/2) and (2/3,1/3). Their SI values would differ by approximately 0.08617, but the proposed bound is only 0.02 using P's minimum stationary mass, or 0.03170 using the smaller mass across both chains.

If a common fixed reference distribution is intended, this example does not contradict that version of the bound. The authors would need to state this clearly and show how it applies to the idealized operator in Theorem 1, including the claimed O(epsilon) stationary perturbation and the objective-gap argument that preserves the optimizer.

That is the part I cannot yet verify from the sketch. Proposition 1 needs a qualification too: the reversible, teleported chain [[0.005,0.995],[0.995,0.005]] has eigenvalue -0.99 and singleton SI(k)-chance=0.5*(-0.99)^k. The expansion remains valid, but reversibility alone does not justify the subsequent interpretation as decaying steps and plateaus. The authors already noted odd/even effects in Pin Pad, however the formal statement should need to handle those modes more explicitly.

  1. Also, I think the statistical calibration needs a more precise definition, especially since this is a central addition to earlier Markov-time sweeps. Definition 1 gives the constraints on the null, but the paper does not seem to specify a probability law over admissible weighted flow matrices or a sampling algorithm. I suspect different ensembles can preserve the same margins and diagonal. It would be great to know which ensemble the empirical p-value refers to, including whether off-diagonal support or sparsity is preserved. This may be documented in the code supplement, but the statistical definition needs to be stated in the paper clearly.

I would also like the authors to separate null-relative significance from the decision that a graph has meaningful dynamical scales. Appendices D/F show that halving the SE floor from 0.1 to 0.05 bits makes the expander gate pass in only 1/5 seeds, even with 500 nulls and correction across k. This however does not by itself invalidate the p-values: an effect-size threshold is a legitimate additional rule. But it means the chosen floor is doing important work, and I am not entirely clear how a user should choose it on a new graph. For G4, the self-loop controls establish structure beyond dwell, but do not check whether P^k reproduces directly observed lag-k transitions. The authors acknowledged the reset episodes and nearly static scene, and the Pin Pad peak captures lattice parity. I therefore read these as useful diagnostics of fitted chains, with the link to a useful observation stride still to be validated.

Questions for authors

Several technical questions - see my comments above also:

  1. Is Lemma 5 in the paper defined with a common fixed reference distribution or not? Please can the authors make the reference measure explicit in Theorem 1 and do a fuller argument of the stationary-perturbation and optimizer-gap. Also, what additional spectral assumptions support Proposition 1's decay interpretation?

  2. Additionally, I would like to learn What the exact probability law and sampler is for the matched null, and does it preserve off-diagonal support or sparsity? If the sampler uses MCMC, what supports adequate sampling? How should the effect-size floor be selected on a new graph, and what null hypothesis does the resulting decision test?

  3. Have the authors compared P^k with directly observed lag-k transitions within episodes, particularly around the world-model peak? A Chapman-Kolmogorov check, with uncertainty, could hlep distinguish a timescale of the fitted operator from one supported by the observed latent dynamics.

1 comment
AuthorsOpenReview ↗

Lemma 5 corrected, the null's sampling law, and the lag-k check we did not run

Thank you for checking the mathematics. Both counterexamples are correct.

Question 1.

Lemma 5. As stated ("any two stochastic P,P~P,\tilde P") it is false when each chain is weighted by its own stationary distribution (we get the same 0.086 against 0.02). The proof shows the version with one fixed reference: fix π\pi and V=A⊤πV=A^\top\pi, and for row-stochastic MM let S(M;C)=∑alog⁡2(1/Va)∑i∈aπi∑j∈aMijS(M;C)=\sum_a\log_2(1/V_a)\sum_{i\in a}\pi_i\sum_{j\in a}M_{i j}. Then

∣S(M;C)−S(M~;C)∣≤log⁡2(1/πmin⁡)max⁡i∥Mi⋅−M~i⋅∥1.\lvert S(M;C)-S(\tilde M;C)\rvert\le\log_2(1/\pi_{\min})\max_i\lVert M_{i\cdot}-\tilde M_{i\cdot}\rVert_1 .

Theorem 1 uses the stationary distribution π\pi of the coupled chain PP as the reference for both operators. ΠR\Pi_R is built from that same π\pi, restricted to each block and renormalised (hence Faa∞=∑bVa∩b2/VbF^\infty_{a a}=\sum_b V_{a\cap b}^2/V_b in Section 3.3), so π\pi is stationary for ΠR\Pi_R too and the two-reference case does not arise. The paper writes π(b)\pi^{(b)} both for these conditionals and for the uncoupled blocks' stationary distributions.

The O(ε)O(\varepsilon) step enters once, in ∥ΠRˉ−ΠR∥∞=O(ε)\lVert\Pi_{\bar R}-\Pi_R\rVert_\infty=O(\varepsilon), where ΠRˉ\Pi_{\bar R} is built from the uncoupled blocks' own stationary distributions: π\pi conditioned on a block is O(ε)O(\varepsilon) from that block's stationary distribution (Simon and Ando, 1961; Meyer, 1989, as cited in the paper). The constant depends on the conditioning of each block and is not tracked; Limitation (i) calls the constants explicit, and they are not.

Optimizer gap. The paper uses γ(V)\gamma(V) without defining it: it is the smallest value of S(ΠR;R)−S(ΠR;C)S(\Pi_R;R)-S(\Pi_R;C) over C≠RC\ne R, positive by Lemma 4. For each such CC, S(M;R)−S(M;C)S(M;R)-S(M;C) is again linear in the rows of MM, with coefficients of absolute value at most πilog⁡2(1/πmin⁡)\pi_i\log_2(1/\pi_{\min}), so between ΠR\Pi_R and PkP^k it changes by at most 3δlog⁡2(1/πmin⁡)<γ(V)3\delta\log_2(1/\pi_{\min})\lt\gamma(V) and RR stays the unique optimizer.

This is not a full proof: c1,c2c_1,c_2 are not worked out and the multi-level case is numerical only. Theorem 1 should be read as proved for the idealized operator and sketched for nearly uncoupled chains, as the abstract words it.

Proposition 1. Agreed. The expansion holds for every reversible chain, but "a nonnegative mixture" and the steps-and-plateaus reading need λi≥0\lambda_i\ge 0, or must be read on even kk only. A negative eigenvalue gives an alternating term, as in your two-state chain; the checkerboard partition at k=2k=2 in the Pin Pad room (Section 4.7) is consistent with such a mode. The statement should carry the assumption.

Question 2.

Law and sampler. The paper gives only the constraints; the procedure is in the code supplement and belongs in Definition 1 (our reply to mGEB lists the steps). Each draw starts from the observed Φ\Phi and repeatedly picks two positive off-diagonal cells (a,b)(a,b), (c,d)(c,d) at random and moves min⁡(Φab,Φcd)\min(\Phi_{a b},\Phi_{c d}) from them to (a,d)(a,d) and (c,b)(c,b) (pairs sharing a row or column, or with a=da=d or c=bc=b, are skipped), with 20 proposals per off-diagonal cell. A null draw is thus the state of a random walk after a fixed number of steps from the data. We have not characterised its limiting law, do not claim uniformity, and have no convergence diagnostic; the 500-draw rerun (Appendix D) concerns the number of draws, not the mixing of each. So each p-value is the rank of the data among these draws, not an exact test under a stated law.

Support is not preserved: a move empties at least one cell and can fill empty ones.

What is tested. The p-value asks whether the chain's optimized score at step kk could be a draw from the rewired chains with the same π\pi and self-loops. It does not test "no dynamical scales": the structureless expander differs from its nulls by little (peak 0.049–0.094 bits) but consistently, and without the floor it beats all 20 draws at 6 to 9 of the 22 step counts in every instance. The reported decision adds an effect-size floor (0.1 bits; added to the p-value rule before the reported runs, after trial runs of a test fixture holding the planted chain and the same expander chain as the first reported instance), so it is a rule, not a test of a stated hypothesis. We have no rule for choosing the floor on a new graph. It sat above the two structureless controls of the main text and carries the expander control (Limitation ix); the one with heavy self-loops reaches 0.10–0.19 bits once they are removed (Appendix D).

Question 3. No. We have not compared PkP^k with observed lag-kk transitions within episodes, so the peak at k∗=11k^\ast=11–1616 is a property of the fitted one-step chain. It is the right test, and your reading of what the paper shows is correct (Limitations vii and xii).

We have not uploaded a revised PDF: a rewrite, a full proof and further checks of the null are more than we can do properly before the deadline, so the corrections are stated in these replies.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.