acceptodds
Under review as a conference paper at ICLR 2027

A Rate Spectrum for Moving Proximal Centers in Sequence-Form Games

Abstract

Modern solvers for two-player zero-sum extensive-form games deploy the current iterate as the policy, so the last-iterate question is when it converges. In the regularized counterfactual regret family, convergence comes from a reference strategy. The iterate is regularized toward a reference, which removes the cycling of unregularized dynamics at the price of pulling the solution away from the Nash set. On the Euclidean carrier a fixed magnet converges linearly to a regularized equilibrium at distance from the Nash set, for proximal weight . Periodic re-centering and hard resets reach exact equilibria at asymptotic or sublinear rates. Between these ends lies the interior speed , for which no finite-time rate is known for this update class. All three move one reference at different speeds, so their common-carrier proxies are instances of one schedule coordinate, the prox-center update spectrum, and one analysis covers the carrier-matched family. A constant speed inside an explicit window gives linear last-iterate convergence to an exact Nash equilibrium of the original game; the bound is sharp at the endpoint. Summable schedules freeze at the computable resolvent of the center's limit, with a closed-form positive floor at interior equilibria. At interior equilibria the asymptotic slope is a closed form in the speed, the proximal weight, and the game's skew spectrum, and its mode eigenvalues match measured slopes to three significant figures. On Leduc, whose equilibrium set is a manifold, every practical speed lies outside the estimated window and the finite-horizon decay at the faster speeds is fitted by a power law. Linear decay reaches a saddle gap of on Kuhn, and the Leduc power-law exponent is –. Across twenty-eight carrier-matched entries the moving center reaches a saddle gap below wherever any other method does. On Leduc it leads periodic re-centering by – at k iterations.

Primary area:
reinforcement learning
Keywords:
imperfect-information games, sequence-form games, last-iterate, counterfactual regret minimization

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

This paper gave a general framework for the anchoring algorithms in extensive-form games, and showed a last-iterate linear convergence rate to the Nash equilibrium.

Critical strengths
  • The algorithm achieves linear convergence to the NE.
  • The algorithm generalizes different anchoring algorithms in EFGs.
Critical weaknesses

The algorithm depends on a coefficient γ\gamma, which can be arbitrarily small. Hence, the convergence rate in the paper can be arbitrarily bad. This dependence also appears in previous papers that achieve linear convergence, such as [1][2]. A discussion about the relationship between these slope constants would be good.

[1] Wei, Chen-Yu, et al. "Linear last-iterate convergence in constrained saddle-point optimization." arXiv preprint arXiv:2006.09517 (2020).

[2] Liu, Mingyang, et al. "The power of regularization in solving extensive-form games." arXiv preprint arXiv:2206.09495 (2022).

Questions for authors

Could you please add these baseline algorithms? I would like to see if this anchoring framework performs better than previous algorithms.

  • PCFR+
  • Regularized algorithms such as regularized optimistic mirror descent
4 comments

We thank the reviewer for the careful reading and for two concrete requests, both of which improved the paper.


Weaknesses

The algorithm depends on a coefficient γ\gamma, which can be arbitrarily small. Hence, the convergence rate in the paper can be arbitrarily bad. This dependence also appears in previous papers that achieve linear convergence, such as [1][2]. A discussion about the relationship between these slope constants would be good.

Thank you for this valuable suggestion. We have added the requested discussion to the revision (Section 4, "Relation to earlier constants", and Appendix I). In short, the constant you refer to and our γ\gamma are the same quantity.

(1) Wei et al. [1]. For bilinear games on polytopes, their SP-MS condition (Definition 1, exponent 00) is exactly dist(z,Z⋆)≤1C dist(0,F(z)+NZ(z))\mathrm{dist}(z, Z^\star) \le \frac{1}{C}\,\mathrm{dist}(0, F(z)+N_Z(z)), which is metric subregularity with modulus γ=1/C\gamma = 1/C. Consequently, an arbitrarily small CC corresponds to an arbitrarily large γ\gamma in our notation (Lemma 4.3). The identity follows directly: for a polytope ZZ, the supremum in SP-MS equals the norm of the projection of −F(z)-F(z) onto the tangent cone of ZZ at zz. Because the tangent and normal cones are polar to each other, this equals dist(0,F(z)+NZ(z))\mathrm{dist}(0, F(z)+N_Z(z)).

Both rates then degrade in the same way. In Theorem 8 of Wei et al., the contraction is governed by C5∝η2C2=η2/γ2C_5 \propto \eta^2 C^2 = \eta^2/\gamma^2, while in our Lemma 4.3 it is governed by 1−ρJ≈1/(2γ2τ2)1-\rho_J \approx 1/(2\gamma^2\tau^2) for large γ\gamma. The γ−2\gamma^{-2} dependence is a property of the game's geometry shared by exact linear-rate results; it is not specific to the moving center.

(2) Liu et al. [2]. Although the linear rate (1−ητ)t(1-\eta\tau)^t of Reg-DOMD (their Theorem 4.1) does not involve γ\gamma, it describes convergence toward the regularized equilibrium, which lies at a distance of O(τ)O(\tau) from the Nash set. Reaching an exact equilibrium requires annealing τ\tau, at which point the guarantee becomes sublinear (their Theorem 4.2).

Therefore, the two results sit at different points along a single trade-off: a γ\gamma-free rate to a biased point, or a γ\gamma-dependent rate to an exact equilibrium without annealing.

(3) Is the rate still informative when γ\gamma is large? We conducted a new experiment on a family of games where γ\gamma is known in closed form, γ=2/ε\gamma = 2/\varepsilon (new Table 6, Appendix I). As γ\gamma increases by 30×30\times, the rate decreases by 907×907\times, and our prediction matches the measured rate to within 0.12%0.12\% with no fitted constant:

ε\varepsilonγ\gammapredicted ratemeasured ratemeasured / predicted
0.36.671.2262×10−31.2262\times10^{-3}1.2262×10−31.2262\times10^{-3}1.0000
0.120.01.3530×10−41.3530\times10^{-4}1.3532×10−41.3532\times10^{-4}1.0001
0.0366.71.2168×10−51.2168\times10^{-5}1.2182×10−51.2182\times10^{-5}1.0012
0.012001.3519×10−61.3519\times10^{-6}1.3519×10−61.3519\times10^{-6}1.0000

(Rates are per iteration, at η=0.2\eta = 0.2, τ=0.5\tau = 0.5, β=0.05\beta = 0.05; the prediction is the exact slow eigenvalue of the mode system of Section 6.) Thus, the slowdown for large γ\gamma reflects the true behavior of the dynamics tracked by our theory, rather than looseness in the bound.

Questions

Could you please add these baseline algorithms? I would like to see if this anchoring framework performs better than previous algorithms.

Thank you for this helpful request. We evaluated both baselines across all four games and added the results to Table 2 of the revision. All entries use the paper's metric: the saddle gap of the last iterate, with the first iteration falling below 10−610^{-6} shown in parentheses (a dash means it is not reached within the horizon).

MethodKuhnLeducLiar's Dice (4-sided)Liar's Dice (5-sided)
moving center, β=0.05\beta=0.05 (ours)8.4×10−128.4\times10^{-12} (5200)2.30×10−42.30\times10^{-4} (–)10−1610^{-16} (4600)10−1610^{-16} (8000)
re-center K=20K=201.5×10−111.5\times10^{-11} (6400)2.12×10−42.12\times10^{-4} (–)10−1610^{-16} (3600)10−1610^{-16} (6400)
CFR+5.9×10−35.9\times10^{-3} (–)3.4×10−33.4\times10^{-3} (–)10−1610^{-16} (200)10−1610^{-16} (100)
RTCFR+10−1610^{-16} (400)1.7×10−141.7\times10^{-14} (6000)10−1610^{-16} (200)10−1610^{-16} (300)
Adp-RTCFR+10−1610^{-16} (400)1.7×10−141.7\times10^{-14} (6000)10−1610^{-16} (200)10−1610^{-16} (100)
PCFR+ (new)10−1610^{-16} (400)10−1610^{-16} (22400)10−1610^{-16} (200)10−1610^{-16} (100)
Reg-DOMD, annealed (new)1.2×10−151.2\times10^{-15} (38800)1.3×10−21.3\times10^{-2} (–)6.1×10−66.1\times10^{-6} (–)1.5×10−11.5\times10^{-1} (–)

PCFR+. Its last iterate reaches the floating-point floor on all four games. It falls below 10−610^{-6} at the same iteration as the RTCFR+ family on Kuhn and both Liar's Dice games, and 3.7×3.7\times later on Leduc (2240022400 vs 60006000). Note that RTCFR+ moves its reference by periodic hard resets (βt∈{0,1}\beta_t \in \{0,1\} in our family), in its native geometry; on 5-sided Liar's Dice, RTCFR+ itself enters at 300300 and Adp-RTCFR+ at 100100. Moreover, the averaged strategies of PCFR+ reach this threshold no earlier (2400/82800/3400/37002400/82800/3400/3700 with linear averaging, and 400/32000/600/500400/32000/600/500 with quadratic averaging).

Regularized optimistic mirror descent (Reg-DOMD, Liu et al.). With a fixed τ=0.5\tau = 0.5, this method converges to the regularized equilibrium, and the gap remains at 0.51/4.58/1.56/1.660.51 / 4.58 / 1.56 / 1.66 across the four games. To ensure a strong baseline, we annealed τ\tau (halving it every KK iterations) and tuned over 3232 configurations of KK and step size. It reaches 10−610^{-6} only on Kuhn (earliest at iteration 2040020400, compared to 52005200 for the moving center). On the other three games, its best final gaps are 1.3×10−21.3\times10^{-2} (Leduc), 6.1×10−66.1\times10^{-6} (4-sided Liar's Dice), and 1.5×10−11.5\times10^{-1} (5-sided), compared to 2.3×10−42.3\times10^{-4}, 10−1610^{-16}, and 10−1610^{-16} for the moving center. The moving center achieves these gaps using one constant β\beta and requires no annealing schedule to tune.

Protocol. All methods are deterministic, eliminating seed variance. The moving center uses one setting, (η,τ,β)=(0.2,0.5,0.05)(\eta, \tau, \beta) = (0.2, 0.5, 0.05), across all four games without per-game tuning. RTCFR+ uses its published hyperparameters, PCFR+ has no tunable parameters, and only Reg-DOMD is tuned per game.

Summary of the comparison. RTCFR+ and Adp-RTCFR+, which update their reference via hard resets, reach 10−610^{-6} on Leduc 3.7×3.7\times earlier than PCFR+, while Adp-RTCFR+ ties PCFR+ on the other three games. The Euclidean moving center (the variant covered by our theorems) reaches 10−610^{-6} earlier than tuned Reg-DOMD on Kuhn and attains smaller gaps on the other three games, whereas PCFR+ converges earlier than it, as the revised Table 2 shows side by side. Within the Euclidean family, the moving center uses a single speed. Against the four re-centering periods, it ranks first on Kuhn and second on the other three games, whereas the period K=200K = 200 finishes 32×32\times above it on Leduc (these findings are now listed explicitly in Section 8). The paper's main contribution is the rate theory for this family, and these baselines place it in context.


Thank you again for your time and feedback. We hope the added discussion of the constants and the results for the two baselines address your concerns. We are happy to run further comparisons if needed.

Thank you for your response! I will keep my score.

We sincerely thank the reviewer for your careful reading and positive evaluation of our paper and response.

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

This paper studies last-iterate convergence in two-player zero-sum extensive-form games through a unified framework for moving proximal centers. The authors introduce a schedule parameter that interpolates between fixed reference strategies, continuously moving centers, and hard resets. By analyzing a coupled recursion for the iterate-tracking error and the center error, the paper establishes a sufficient condition for linear convergence to an exact Nash equilibrium under constant center speeds. It further characterizes summable and vanishing schedules and derives local convergence-rate expressions in terms of the game's skew spectrum. Experiments on matrix games, Kuhn poker, Leduc poker, and Liar's Dice are used to examine the predicted regimes and compare the proposed dynamics with several baselines.

The paper addresses an interesting theoretical question, and the analysis is generally well organized. My main concerns are the precise novelty relative to existing moving-proximal-center methods, and the extent to which the theoretical results translate to native extensive-form game solvers.

Critical strengths
  1. This paper establishes a good analysis framework for moving proximal centers, derives an explicit sufficient convergence window, and proves that under specific conditions, the final iteration linearly converges to an exact Nash equilibrium.
  2. This paper characterizes the convergence behavior under different center update schedules, and derive the asymptotic convergence rate by combining the local spectral structure of game operators.
Critical weaknesses

I am not very familiar with the author's research field, so I cannot make a very accurate weak point judgment. However, I am concerned about:

  1. Will the computational cost increase if the algorithm proposed by the author moves at the center of each iteration?
  2. In the Leduc experiment in Table 2, at 200k iterations, the moving center method was inferior to the periodic reset center method, but after being extended to 500k, it exceeded the comparison method again. Can this inconsistency advantage be better explained?
Questions for authors

If the author answers the above weaknesses, I think it will make me have stronger confidence in this paper.

1 comment

We thank the reviewer for the positive assessment of the paper.


Weaknesses

  1. Will the computational cost increase if the algorithm proposed by the author moves at the center of each iteration?

Thank you for this insightful question. We measured this overhead directly and added the results to the revision (Section 8, "Per-iteration cost").

Moving the center requires just one convex combination of two vectors, μt+1=(1−β)μt+βzt+1\mu_{t+1} = (1-\beta)\mu_t + \beta z_{t+1}. This operation takes 1.71.7--16.716.7 microseconds per step across the four games, accounting for only 0.005%0.005\%--1.5%1.5\% of the total step cost. The cost of a full step is 0.910.91--1.06×1.06\times that of periodic re-centering with K=20K = 20 or K=50K = 50. That is, moving the center at every iteration costs approximately the same per step as moving it every 2020 or 5050 iterations.

Full step, moving center / periodic re-centeringKuhnLeducLiar's Dice (4-sided)Liar's Dice (5-sided)
vs K=20K = 200.961.000.911.02
vs K=50K = 501.021.030.991.06

  1. In the Leduc experiment in Table 2, at 200k iterations, the moving center method was inferior to the periodic reset center method, but after being extended to 500k, it exceeded the comparison method again. Can this inconsistency advantage be better explained?

Thank you for your close reading of Table 2. The two orderings are consistent, and we have stated the mechanism in the revision (Section 8).

Why periodic resets lead early. A hard reset sets the center to the current iterate every KK steps, immediately removing the accumulated center error. Because the moving center reduces this error gradually, it trails at short horizons: at 200200k, the gaps are 2.30×10−42.30\times10^{-4} (moving, β=0.05\beta = 0.05) and 2.12×10−42.12\times10^{-4} (K=20K = 20), a 9%9\% difference.

Why the moving center leads later. The moving center achieves faster tail decay. Over 100100k--500500k, the decay rate fitted to the running minimum is 7.28×10−67.28\times10^{-6} per iteration for the moving center, compared to 6.33×10−66.33\times10^{-6} for K=20K = 20. Given these different slopes, the two curves cross; the running minima cross at 280280k, between the two horizons you compared.

At 500500k, the moving center reaches 2.12×10−52.12\times10^{-5} while K=20K = 20 is at 4.32×10−54.32\times10^{-5}, a 2.0×2.0\times lead. Thus, this reflects a crossover between an early advantage and a faster tail, and the relative ordering depends on the horizon in a predictable way.


My main concerns are the precise novelty relative to existing moving-proximal-center methods, and the extent to which the theoretical results translate to native extensive-form game solvers.

We also briefly address the two points raised in your summary.

Novelty relative to existing moving-center methods. Existing analyses cover the two extremes: a fixed reference (linear rate, but to a regularized equilibrium) and periodic resets (exact equilibria, with asymptotic or sublinear rates). For the intermediate regime, where the reference moves by a fraction βt∈(0,1)\beta_t \in (0,1) at every step, no finite-time rate to an exact equilibrium was previously known. Our window theorem (Theorem 4.5) provides the first such rate, with a sharp endpoint.

Transfer to native solvers. Our theorems hold for the Euclidean members of the family, and the paper notes that no rate transfers to the native geometries. Instead, the native solvers are compared empirically in the same table, and our revision makes this distinction explicit in the introduction.


If the author answers the above weaknesses, I think it will make me have stronger confidence in this paper.

Thank you for your time and feedback. We hope our two answers increase your confidence in the paper. We are happy to answer any further questions.

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

The paper studies a reanchoring method in two-player zero-sum extensive form games that is parameterized by β\beta, which control the way there anchor changes at each iteration. The main result is that this method, termed prox-center update spectrum, is a linear last iterate convergence to an exact Nash equilibrium for a constant and suitable selected β\beta. Moreover, the authors show (non-linear) convergence for a summable sequence βt\beta_t under an additional step-size condition and linear albeit game dependent convergence for games with a unique fully mixed equilibrium. The paper also includes some toy examples that corroborate the theoretical findings.

Critical strengths

None

Critical weaknesses

Unfortunately, I found the writing of the paper very hard to parse.

Rebuttal Update: during the rebuttal the authors have implemented significant changes in the presentation of the paper.

Questions for authors

In the experimental section, the authors presents iteration counts and no wall-clock times. I understand that the counts are needed to support the theory but could they also briefly comment on how their method compares in wall-clock time against the variants of CFR? I would like to understand if the method can be applied in practice.

3 comments

We thank the reviewer for the time spent on our paper and for the candid assessment of its presentation, which led to a substantially clearer revision.


Unfortunately, I found the writing of the paper very hard to parse.

We sincerely apologize that the paper was difficult to read. Rereading the submission with your helpful feedback in mind, we identified the cause: the abstract and introduction used custom terms before they were defined. We have revised the paper accordingly (changes in blue):

  • Abstract and introduction rewritten in plain language. They now state the problem, the update rule, and the three main results without specialized vocabulary.
  • Coined terms replaced throughout. The term "carrier" is now "Euclidean geometry", "common-carrier proxy" is "Euclidean counterpart", "prox-center update spectrum" is "moving-center family", and "magnet" is "reference".
  • Algorithm 1 added (Section 3, page 3): presents the entire method in two lines.
  • Terminology table added (Table 9, Appendix I).
  • Experiments restructured around three questions. Section 8 now opens by stating the three experimental conclusions, and the findings from Table 2 are listed explicitly.

For your convenience, we summarize the paper in three sentences. Regularized solvers pull each iterate toward a reference strategy; a fixed reference converges quickly but to a biased point, whereas periodically resetting the reference removes this bias but was known to converge only at asymptotic or sublinear rates. We study the update rule "after each step, move the reference a fraction β\beta toward the new iterate". We prove that a constant β\beta within an explicit window yields linear last-iterate convergence to an exact Nash equilibrium, and we characterize the behavior under other schedules.

We hope the revised presentation makes our contribution easier to assess, and we would be grateful if you could review the updated abstract and introduction.


In the experimental section, the authors presents iteration counts and no wall-clock times. I understand that the counts are needed to support the theory but could they also briefly comment on how their method compares in wall-clock time against the variants of CFR? I would like to understand if the method can be applied in practice.

Thank you for this insightful question. We have measured the per-iteration wall-clock time (single-threaded) and added a paragraph with these results to Section 8 of the revised paper. Per iteration, our method is 55--23×23\times cheaper than the CFR variants (CFR+, PCFR+, RTCFR+, Adp-RTCFR+) on Leduc and Liar's Dice. Because the CFR variants require fewer iterations, their total time to reach a saddle gap of 10−610^{-6} remains shorter; nonetheless, our method reaches this target in approximately 0.90.9 s on Kuhn and in 44--3939 s on the two Liar's Dice games.

On practicality. These two lines of work offer complementary strengths. CFR is designed for practical scalability: it decomposes the problem and minimizes regret locally at each information set, making it highly efficient on large game trees. The strength of our method is its rate: it guarantees a last-iterate linear rate to an exact Nash equilibrium with an explicit window for the speed, whereas the classical guarantee of CFR covers the average strategy at a sublinear rate. We therefore do not seek to replace tuned CFR variants. Practical solvers such as RTCFR+ already combine CFR with a periodically reset reference, and our results establish finite-time rates for the Euclidean counterparts of this family, including continuous movement, for which no rate was previously known.


While the authors admit to have AI assistance to improve the prose of the paper, I think that at some parts they have done it to such an extent that it makes the paper difficult to read.

We take this concern seriously. In accordance with ICLR policy, our submission discloses our use of AI tools in the "AI use statement": they were used to polish the prose and to write and run the experiment code. We reviewed all AI-assisted work against the underlying theorems, numbers, and figures, and we take full responsibility for the content. We sincerely apologize that the polished prose proved harder to read, and we have revised the abstract, introduction, and terminology as described above.


Thank you for your review. Your comments have led to a clearer paper. We are glad to clarify any passage that is still unclear.

I would like to thank the authors for reading my review. I have read the revised version and found the presentation much clearer, so I have updated my review as follows: I increased the presentation score from 1 to 2, added an update comment in the critical weaknesses and ethical concerns sections, and updated my overall score to weak acceptance. I still think the presentation can be improved even further. For example, instead of mentioning the proximal weight, you could mention that using a fixed reference causes the methods to converge to a biased point rather than the actual equilibrium (as you did in your very concise explanation in the rebuttal). You could also avoid mentioning the treeplex early on in the introduction. Overall, however, as indicated by my score updates, I am happy with the changes.

I would also like to clarify one more thing: I hope that my comments about the writing in the submitted version have not offended the authors. Writing is also a matter of personal preference, as evidenced by the different scores across the reviewers, and perhaps the original version simply did not align with mine. In any case, I apologize to the authors if I caused any offense.

Thank you very much for reading our revision and updating your assessment. We took no offense whatsoever. Your earlier comments were specific, fair, and genuinely helped improve the clarity of our paper.

We have adopted both of your additional suggestions in the revised manuscript:

  • Biased point instead of the proximal weight. The abstract and introduction now state that a fixed reference converges to a biased point rather than an equilibrium of the original game. The weight τ\tau is introduced only in Section 3, alongside its definition.
  • No treeplex in the introduction. This term now first appears in Section 3 together with its definition.
  • Applying the same principle elsewhere. Following your suggestion, we also removed the term "treeplex" from the related work (Section 2), explained τ\tau in words where it is defined (Section 3), and restated the conclusion in plain terms.

Thank you again for your thoughtful engagement and constructive feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all three reviewers for their careful reading and constructive comments. We have uploaded a revised paper with all changes marked in blue. This note summarizes our new experiments and revisions.

New experiments (all requested by reviewers):

  1. PCFR+ and regularized optimistic mirror descent as additional baselines on all four games (Reviewer NYPU).
  2. Per-iteration wall-clock time against CFR+, PCFR+, RTCFR+, and Adp-RTCFR+ (Reviewer jeCg).
  3. Per-iteration cost of moving the center (Reviewer rBBi).
  4. Rate versus the constant γ\gamma on a family of games where γ\gamma is known in closed form (Reviewer NYPU).

Summary of changes in the revised paper (all marked in blue):

  • Abstract and Introduction: Rewritten in plain terms. The update rule, the three main results, and the experimental conclusions are now stated without specialized vocabulary.
  • Terminology (whole paper): Replaced "carrier" with "Euclidean geometry", "common-carrier proxy" with "Euclidean counterpart", "prox-center update spectrum" with "moving-center family", and "magnet" with "reference".
  • Section 2: Moved the most relevant related work (optimistic last-iterate methods, regularized sequence-form solvers, and other continuously updated references) from the appendix into the main text.
  • Section 3: Added Algorithm 1, which states the method in two lines.
  • Section 4: Added a new paragraph, "Relation to earlier constants", relating γ\gamma to the constants of Wei et al. and Liu et al.
  • Section 8: Now opens with the three experimental conclusions. It also adds PCFR+ and Reg-DOMD rows to Table 2, an explicit list of the table's findings, a new paragraph on per-iteration cost, and an explanation of the Leduc crossover between 200200k and 500500k. Figure 1 is enlarged.
  • Appendix I (new): Added baseline protocols, the γ\gamma experiment (Table 8), and a terminology table (Table 9).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.