acceptodds
Under review as a conference paper at ICLR 2027

Constructive Rationalization: Learning from Collective Revealed Preferences

Abstract

This paper presents a revealed preference approach for learning the collective consumption behavior. We introduce the Constructive Rationalization Method (CRM), which approximates the real market via a surrogate market of artificial consumers, called androids, with easy-to-compute demand functions. CRM uses observed aggregate demand and adds artificial consumers on the fly, while redistributing wealth under an empirical risk minimization principle. Unlike classical revealed preference approaches, CRM provides guarantees on the generalization risk for learning the aggregate demand function, while relying little on individual rationality. As an application, CRM can be used to provide reliable predictions for collective consumption behavior. Specifically, we show how to apply CRM to approximate allocations that are proportionally fair.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
rationalization, revealed preferences, computational economics, rademacher complexity, preference learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_sDRDOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary
  • The paper looks at learning aggregate consumption behavior from historical price-aggregate-demand observations, without observing individual consumers or their utilities
  • They propose the Constructive Rationalization Method (CRM), with theoretical analysis and empirical evaluations on synthetic markets.
Critical strengths
  • Paper studies an important and less explored problem of learning collective consumption behavior using only aggregate price and demand data. The surrogate market idea is original and useful because it produces a model that can be used both for prediction and for fair allocation, without recovering each person’s utility.
Critical weaknesses
  • The writing is very difficult to understand (but also, admittedly I'm not in the area).
  • I'm confused by the inconsistency (correct me if wrong) relating to the convergence theorem and experiments. The theorem studies full-data Frank-Wolfe with a specific update and measures error against the best solution of each search step. The experiments use mini-batches, fully corrective updates, and L-BFGS on a nonconvex problem. But L-BFGS does not tell us how far it is from the best solution, the error term in the theorem cannot be evaluated for the reported runs right?
Questions for authors
  • See weaknesses
2 comments
AuthorsOpenReview ↗

Response to Reviewer sDRD

We appreciate your feedback on the presentation and on how the convergence analysis relates to the algorithm used in our experiments. We will carefully address your concerns below.

W1: Readability.

We sincerely thank you for carefully evaluating our paper and apologize for the parts that were difficult to follow. Your comments, together with those of Reviewer BTTP, have helped us identify concrete ways to improve the introduction and clarify the technical presentation. We will address these issues in the revised manuscript.

W2: Connection between Theorem 4.2 and the experiments.

We sincerely thank you for your careful reading and the opportunity to clarify how the theory and experiments complement each other.

(i) Choice of Frank-Wolfe methods. We present an analysis for the smoothed, full-data Frank–Wolfe method in Theorem 4.2 and Appendix E.3. In our experiments, we use fully corrective FCFW with mini batches (i.e., stochastic FCFW, SFCFW) instead of full samples for efficiency purposes. Note that the SFCFW methods are well studied in the literature, so we did not present the convergence analysis see, for example, Fully-Corrective Frank-Wolfe in [1] and Stochastic Frank–Wolfe in [2].

(ii) Choice of solving SEP. For our empirical scheme, Theorem 4.2 tracks the error from approximate SEP solutions through ETE_T, which is accumulated from the errors by SEP. For CES androids, if we use a branch-and-bound method (by Gurobi), it returns a feasible value vv and a global upper bound UU, so U−vU-v bounds the SEP error; while the local method (by L-BFGS) could supply a feasible solution without this global certificate. In the solver, we provide both methods. In the paper, Tables 6–7 assess the local SEP solver against Gurobi and shows that L-BFGS matches Gurobi's best feasible value to the reported precision on 19 of 20 instances.

Gurobi reference for L-BFGSMean relative gapMaximum relative gap
Best feasible value0.15%2.96%
Final upper bound0.15%2.97%

Table 8 explains this computational choice. At batch size 10, Gurobi already requires over 15,000 seconds of SEP time per run, versus about 11 seconds for L-BFGS. Hence, in the CRM L-BFGS is used for most of the time unless the progress is slow. We appreciate your suggestion and will further clarify this distinction in the manuscript.

References

[1] Jaggi, M.: Revisiting Frank-Wolfe: Projection-Free Sparse Convex Optimization, Proceedings of the 30th International Conference on Machine Learning, PMLR 28(1), 427–435, https://proceedings.mlr.press/v28/jaggi13.html, (2013).

[2] Hazan, E., and Luo, H. Variance-Reduced and Projection-Free Stochastic Optimization. ICML, https://proceedings.mlr.press/v48/hazana16.html, (2016).

thank you for your response. i maintain my positive score.

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

The paper studies learning aggregate demand from historical price–demand observations without access to individual preferences or demand records. It constructs a surrogate market by selecting artificial consumer types and allocating budgets among them to fit aggregate expenditure shares. The framework is implemented using established convex optimization methods. The paper provides generalization and convergence guarantees, analyzes the computational complexity of finding new consumer types, and studies the expressive power and limitations of the surrogate class. Synthetic experiments evaluate demand prediction and an application to approximately fair allocation.

Critical strengths
  1. The theoretical treatment is generally careful and provides useful problem-specific contributions. Although much of the learning and optimization machinery is established, the paper develops a coherent analysis of the proposed market-learning formulation. In particular, the oracle complexity results and the approximation limitations provide useful information about what the framework can and cannot achieve.

  2. The expenditure-share representation is a useful modeling choice. Working with the response function γ\gamma gives a bounded, simplex-valued representation, whereas quantity demand can become unbounded near zero prices. This normalization facilitates the mixture formulation and its statistical analysis, and is a thoughtful aspect of the approach.

Critical weaknesses
  1. The connection between the theoretical observations and realistic market data needs further clarification. Budget exhaustion is natural within the Fisher-market model, but observed expenditure in an application may not cover the full relevant budget, and recorded transactions may differ from unconstrained demand. It is therefore unclear how the total-budget normalization should be implemented and interpreted when the total budget is not separately observed. The synthetic experiments do not resolve this issue. A clearer account of the required data and budget interpretation would strengthen the practical motivation and establish the scope of the proposed applications.

  2. The presentation is overly technical in places, particularly in the Introduction. Section 1.1 develops Afriat–Varian inequalities, utility reconstruction, proportional fairness, Nash welfare, and equilibrium conditions before giving a sufficiently intuitive account of the proposed construction. The exposition would be easier to follow if it first established the basic market setting, the available aggregate data, and the learning objective, followed by a short explanation of how artificial consumer types form a surrogate market. Some formal background could be moved to the preliminaries or related work.

Questions for authors

The normalization uses the total budget W(p)W(p), which can be recovered as p⋅d(p)p\cdot d(p) under budget exhaustion. How should the method be applied when W(p)W(p) is unknown and observed purchases do not exhaust the relevant budget, for example because of spending or savings outside the modeled goods? Can observed expenditure instead be treated as an effective budget for normalization, and which learning and allocation guarantees would remain applicable under that interpretation?

1 comment
AuthorsOpenReview ↗

Response to Reviewer BTTP

We want to thank you for pointing out these valuable comments. We will carefully address your concerns below.

W1: The connection between the theoretical observations and realistic market data needs further clarification...

We address W1 and the related question together below.

Thank you for highlighting this issue. As you have mentioned, with prices and aggregate demand, we can only compute "observed" expenditure S=p⊤qS=p^\top q. The relevant budget need not represent buyers’ entire wealth: it may be a budget allocated specifically to the "modeled goods". If these individual budgets are fixed and fully spent, the existing learning and allocation guarantees apply under their stated assumptions.

When spending on the modeled goods instead varies because buyers save or spend elsewhere, normalizing by S>0S>0 produces shares of "observed spending". This interpretation is particularly relevant to fair resource allocation, where participants may receive budgets or credits dedicated to the resources being allocated, with limited or no outside use. In such settings, budget exhaustion is a natural modeling assumption, and observed expenditure can serve as the relevant within-market budget.

To model the choice between purchases and retained money explicitly, a possible extension introduces an outside option, as in quasilinear Fisher markets [1,2]. Constructing the corresponding full-budget shares requires aggregate total-budget information, or equivalently aggregate expenditure outside the modeled goods and retained money. The augmented target is

\frac{1}{W(p)} \begin{bmatrix} \operatorname{diag}(p)q\\ W(p)-S \end{bmatrix}.$$ This construction requires no individual-level budget observations. CRM can then fit these augmented shares following $\tilde g$. A separate issue arises when stockouts or rationing constrain purchases. In that case, observed sales may differ from demand, and recovering demand requires availability information and a model of these constraints. We will clarify these data requirements and distinguish prediction of observed spending shares from the budget assumptions required for our allocation guarantee. ### **W2: The presentation is overly technical in places, particularly in the Introduction...** Thank you for the suggestion on improvement of presentation. We do find presenting the problem setting rigorously challenging. We will adjust the introduction in the revision. ### References [1] Liao, L., Gao, Y., Kroer, C.: Statistical inference for Fisher market equilibrium, http://arxiv.org/abs/2209.15422, (2025) [2] Murray, R., Kroer, C., Peysakhovich, A., Shah, P.: Robust market equilibria with uncertain preferences, http://arxiv.org/abs/1912.04867, (2019)
Reviewer_LDJ7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the Constructive Rationalization Method (CRM), which learns aggregate demand from observed price-demand pairs by constructing a surrogate market consisting of artificial consumers, called androids. Instead of recovering individual preferences, CRM incrementally selects androids from a prescribed utility class and optimizes their wealth allocations to approximate the observed aggregate expenditure shares.

The paper formulates this learning problem as a convex optimization problem and develops Frank–Wolfe-based algorithms for constructing the surrogate market. It establishes generalization and convergence guarantees, investigates the computational complexity of the separation oracle, and studies the representational capabilities and limitations of different android classes.

Numerical experiments demonstrate the predictive performance of CRM on synthetic CES and generalized CES markets. The paper also applies the learned surrogate market to approximate proportionally fair allocations by computing market-clearing prices and then querying each real consumer's demand once.

Critical strengths
  1. Interesting technical approach. Constructing a surrogate market through the incremental addition of artificial consumers is an interesting way to approximate aggregate demand without identifying individual preferences. The connection to convex optimization, column generation, and Frank–Wolfe methods is appealing.

  2. Substantial theoretical analysis. The paper provides several results concerning generalization, convergence, computational hardness, and approximation capabilities. In particular, the analysis of inexact separation oracles and the investigation of misspecification errors are valuable contributions.

  3. Potential applications to large-scale markets. Learning aggregate demand without estimating each consumer's utility function may be useful in markets with many participants. The application to approximate proportionally fair allocation is interesting, as it combines historical aggregate observations with a single demand query per participant at deployment.

Critical weaknesses

1. Lack of comparisons with alternative demand-learning methods.

My main concern is the absence of theoretical and/or experimental comparisons with alternative approaches for learning aggregate demand.

The experiments compare different variants and implementations of CRM, but do not establish whether the proposed method offers advantages over other demand-prediction methods.

For example, natural baselines could include:

  • Standard neural network regressors for predicting aggregate expenditure shares, with appropriate output constraints.
  • Fixed-size mixtures of CES demand functions, whose parameters and mixture weights are jointly estimated rather than constructed incrementally.

These comparisons would help clarify the benefits of the proposed constructive approach in terms of prediction accuracy, sample efficiency, and computational cost.

This issue is particularly relevant because the basic experiments generate markets from CES consumers and then learn their aggregate demand using mixtures of CES androids. Thus, the true aggregate demand belongs to the model class under the constant-wealth setting. While the experiments demonstrate that CRM can learn such markets, they provide limited evidence of its advantages over competing approaches.

The generalized CES experiments partially address model misspecification, but comparisons with alternative estimators remain important.

2. The practical scope of the proposed approach needs to be clarified.

The paper motivates CRM as a tool for decision-making in real markets. However, it remains unclear which practical decisions can be supported by the learned aggregate demand function.

I do not think that CRM necessarily needs to support the full range of structural estimation considered in econometric demand estimation. Recovering individual preferences or predicting the effects of interventions on individual wealth distributions may be unnecessary for the paper's intended applications. Nevertheless, the authors should more clearly identify the market environments and decisions for which learning the aggregate price-response function is sufficient.

For example, large markets with approximately price-taking consumers may provide a compelling application. When individual participants have negligible influence on market outcomes, predicting aggregate demand responses to exogenously specified prices may be sufficient for market-clearing price determination or resource allocation, provided that preferences, budgets, and other relevant market conditions remain stable.

Such a setting could provide a stronger practical motivation for CRM without requiring identification of individual preferences.

It would also be useful to clarify the distinction between predicting aggregate demand under alternative prices and conducting more general structural counterfactual analyses. A comparison with structural demand models, such as BLP-type approaches, might be informative in a suitably designed common setting focused on price interventions, although their standard market environments differ substantially from the divisible-goods markets studied here.

3. The incremental value of constructing a surrogate market is not fully demonstrated.

The application to proportionally fair allocation is interesting, and the paper provides theoretical welfare guarantees under appropriate assumptions. However, the surrogate market primarily serves to compute an approximate market-clearing price. The final allocation is then constructed using actual demand reports from individual participants. This raises a natural question: could a conventional aggregate-demand predictor, combined with a suitable equilibrium-finding procedure, achieve comparable welfare performance?

The paper would be considerably strengthened by comparing alternative demand-learning methods not only in prediction accuracy but also in their downstream performance, including equilibrium accuracy, Nash-welfare loss, and computational efficiency.

Questions for authors
  1. Have the authors considered comparisons with fixed-size CES mixture models or standard neural network regressors? What practical advantages do they expect CRM to provide over these approaches?

  2. What are the main real-world decision-making scenarios in which learning the aggregate price-response function, without recovering individual preferences, is sufficient? Could the authors clarify the role of large-market or price-taking assumptions in these applications?

  3. For the proportionally fair allocation experiments, could alternative aggregate-demand predictors be used to compute approximate equilibrium prices? If so, what are the expected advantages of CRM in terms of solution quality or computational efficiency?

  4. To what extent can the learned surrogate market reliably predict demand under price interventions outside the training distribution? What assumptions are required to interpret these predictions as responses to actual price changes rather than statistical associations?

3 comments
AuthorsOpenReview ↗

Response to LDJ7 (Part 1 of 2) — W1–W3, Q2, and Q4

We want to thank you for these insightful comments. We will carefully address W1–W3, Q2, and Q4 below and answer Q1 and Q3 together in Part 2 because both questions concern the same methods.

Weaknesses

W1: Lack of comparisons with alternative demand-learning methods.

Thank you for asking for this comparison. We want to kindly remind that we have included a comparison in the appendix (Appendix I.2.4), due to limited space. Specifically, Table 9 compares CRM with Simplex-MLP and Fixed-size CES dictionaries on constant-wealth CES and GES markets. We leave further details in our response to Q1 in Part 2.

W2: The practical scope of the proposed approach needs to be clarified.

Thank you for asking us to clarify the scope. Since this also relates to Q2, we kindly refer you to the response of Q2.

W3: The incremental value of constructing a surrogate market is not fully demonstrated.

Thank you for raising this point. In section I.4, we present Table 10 to show the welfare loss when applying the CRM. To address your concern, we extend this table to compare with Simplex-MLP and Fixed-size CES market in the same fashion of Table 9. We leave further details in our response to Q3 in Part 2.

Questions

Q2 What are the main real-world decision-making scenarios ...

Thank you for asking us to clarify the practical use of aggregate demand, the price-taking assumption, and the relationship to structural demand estimation. We will clarify this practical scope and the distinction from structural counterfactual analysis in the revision.

(i) Practical applications and price taking. We agree that large markets with approximately price-taking consumers provide a natural setting for CRM. Such settings can arise in cloud computing, communication networks, online advertising, and electricity demand management [1], when resources can be treated as divisible and each participant has negligible influence on prices.

(ii) Comparison with BLP. As you mentioned, BLP-type models [2] and CRM can both use aggregate observations. BLP estimates structural demand for differentiated, discrete-choice products and addresses price endogeneity through identifying assumptions and instruments. CRM constructs a surrogate divisible-goods market to approximate aggregate price responses under stable market conditions. Therefore, BLP typically needs additional types of data to identify broader counterfactuals involving changes in preferences or wealth distributions. In comparison, CRM focuses on price-demand pairs and can be seen as an alternative approach to conduct price interventions and prediction tasks.

Q4 To what extent can the learned surrogate market...

This is an interesting question. We will clarify what our prediction guarantee covers and what additional assumptions are needed to assess the effect of a price intervention.

(i) Scope of prediction. Our generalization bound guarantees average prediction accuracy for fresh prices drawn from the training distribution. For new price vectors, the surrogate market provides economically structured predictions through utility maximization, rather than unconstrained statistical extrapolation. Their reliability depends on the extent of distribution shift and the validity of the underlying preference and wealth specifications [3].

(ii) Assumptions for price interventions. Interpreting predicted demand changes as intervention responses requires

  • stable consumer preferences,
  • stable market composition, and
  • a valid specification of wealth adjustments.

For observational data, price variations should be exogenous conditional on observed market conditions, or an appropriate identification strategy should be employed to address price endogeneity. These are standard considerations in structural demand analysis [2].

References

[1] Ito, K., Ida, T., Tanaka, M.: Selection on welfare gains: Experimental evidence from electricity plan choice, American Economic Review, 113(11), 2937–2973, (2023).

[2] Berry, S., Levinsohn, J., Pakes, A.: Automobile prices in market equilibrium, Econometrica, 63(4), 841–890, (1995).

[3] Varian, H. R.: The nonparametric approach to demand analysis, Econometrica, 50(4), 945–973, (1982).

AuthorsOpenReview ↗

Response to LDJ7 (Part 2 of 2) — Q1 and Q3

Q1 Have the authors considered comparisons with fixed-size CES mixture...

Since the question also relates to Question 3, we combine the question and response here.

Q3 For the proportionally fair allocation experiments, could alternative aggregate-demand predictors ...

We sincerely thank the reviewer for these constructive suggestions. We addressed some of these comparisons in the appendix I.2.4 with two alternative predictors

  • fixed-size CES mixture models (the fixed CES dictionary)
  • and a standard neural network regressor (the Simplex-MLP)

For your convenience, we want to highlight a few results below. After that, we provide a few remarks on the advantages of our method.

(i) Baseline comparisons. In Table 9, we test several sizes for each baseline. The fixed CES dictionaries have 25, 50, or 100 atoms (275, 550, or 1,100 atom parameters), and the Simplex-MLPs have 618, 1,738, or 5,514 trainable parameters. The table below shows the best tested size of each baseline method by mean test error. Errors are mean ± SD over five seeds.

MarketMethodParametersTrain L∞L_\inftyTest L∞L_\infty
CESCRM-SFCFW948 ± 136(3.91±4.50)×10−4\mathbf{(3.91\pm4.50)\times10^{-4}}(4.57±4.72)×10−4\mathbf{(4.57\pm4.72)\times10^{-4}}
CESSimplex-MLP5,514(1.27±0.23)×10−3(1.27\pm0.23)\times10^{-3}(1.75±0.43)×10−3(1.75\pm0.43)\times10^{-3}
CESFixed-size CES1,100(1.490±0.884)×10−2(1.490\pm0.884)\times10^{-2}(1.516±0.926)×10−2(1.516\pm0.926)\times10^{-2}
GESCRM-SFCFW2,517 ± 142(2.007±0.023)×10−4\mathbf{(2.007\pm0.023)\times10^{-4}}(1.837±0.117)×10−2\mathbf{(1.837\pm0.117)\times10^{-2}}
GESSimplex-MLP5,514(2.39±0.27)×10−4(2.39\pm0.27)\times10^{-4}(2.832±0.421)×10−2(2.832\pm0.421)\times10^{-2}
GESFixed-size CES1,100(3.159±0.179)×10−1(3.159\pm0.179)\times10^{-1}(3.116±0.072)×10−1(3.116\pm0.072)\times10^{-1}

The results show that CRM outperforms these baseline methods. While MLPs may also provide good results, they may use significantly more parameters and have larger test errors.

(ii) Allocation quality. Table 10 reports CRM's NSW losses; here we complement those results with the worst individual utility loss.

We use the same Table 10 setting (10 goods, 1,000 agents, constant budgets, 300 training observations). The baselines use the Table 9 MLP architecture and training settings (two 64-unit hidden layers) and 100 fixed-size CES atoms. We use matched price initializations and a 10−1010^{-10} expenditure-share clearing tolerance, then query true individual demands and rescale each good to supply.

Worst individual utility loss: ΔNSW−=max⁡i[1−ui(x^i)/ui(xiref)]\Delta_{\text{NSW}}^-=\max_i[1-u_i(\hat x_i)/u_i(x_i^{\mathrm{ref}})] (in %; five-seed mean). For CES, the reference is the NSW-optimal allocation x∗x^*. For GES, we use the competitive-equilibrium allocation xCEx^{\mathrm{CE}} as the reference, computed at market-clearing prices obtained from the true demand.

MarketCRM-SFCFWSimplex-MLPFixed-size CES
CES2.49×10−2\mathbf{2.49\times10^{-2}}3.86×10−23.86\times10^{-2}7.85×10−27.85\times10^{-2}
GES6.74×10−1\mathbf{6.74\times10^{-1}}7.63×10−17.63\times10^{-1}1.06×1011.06\times10^{1}

Our results suggest that a surrogate fitted to aggregate demand can help limit individual utility losses. With the same allocation rule for all methods, CRM’s prices led to the lowest average loss for the worst-affected buyer in both settings.

(iii) Practical advantages. As we show that CRM can have better performance in prediction error and welfare loss, numerically, in practice, we believe the approach may have a few features that are friendly to further applications. For example,

  • Firstly, CRM can be very flexible. If the current estimates are not good enough, CRM safely increases the market size without worrying about generalization error. At the same time, CRM can represent a potentially large population through a smaller collection of artificial consumers. However, for neural network based methods, increasing the model complexity may deteriorate the performance.

  • Secondly, when budgets are constant in CRM, we can compute its clearing prices by convex optimization approaches in polynomial time. While if Simplex-MLP (or similar neural network approaches) is used, we need a reliable algorithm for a fixed-point problem, which can generally be PPAD-hard.

  • Moreover, the learned market can be reused to investigate different resource capacities while holding preferences and budgets fixed. This supports computing candidate clearing prices for capacity-planning scenarios.

These findings highlight CRM’s practical value in demand prediction and downstream allocation. We thank the reviewer for encouraging us to clarify these benefits.

AuthorsOpenReview ↗

Your feedback would be appreciated

Dear Reviewer LDJ7,

We sincerely thank you again for your valuable feedback. We have carefully addressed each of your points and made revisions accordingly. As the author–reviewer discussion period is coming to a close, we would greatly appreciate it if you could review our responses and let us know if there are any remaining concerns.

In particular, we would like to highlight two key updates in the revised manuscript:

  • Q1: We have moved the prediction comparison against Simplex-MLP and Fixed Random CES to the main text (§5.1, Table 3, p. 9).
  • Q3: We have added a comparison in §5.2 (Table 5, p. 10), evaluating each predictor by computing prices under the same allocation rule and reporting the worst-case buyer utility loss in each market.

Once again, thank you for your time and support in reviewing our work. We look forward to hearing from you soon

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions

We sincerely thank the reviewers for their thoughtful comments and suggestions. We have made the following revisions:

1. More accessible baseline comparisons (§5.1). Following the comment on alternative demand-learning methods by reviewer LDJ7, we moved the existing comparisons with Fixed Random CES and Simplex-MLP from the appendix to the main text (Table 3), making the prediction results easier to find.

2. Additional downstream comparisons (§5.2). Following the comment on additional downstream comparisons by reviewer LDJ7, we moved the existing Nash-welfare results to the main text (Table 4) and added a downstream comparison of worst individual utility loss across the three methods (Table 5).

3. Clearer introduction and background (§§1–2). Following the comments on readability by reviewers BTTP and sDRD , we now introduce the CRM construction before the formal background and have moved the individual-rationalization and fair-allocation background to the preliminaries.

We appreciate the reviewers’ feedback, which have helped us make the presentation clearer and explain the experimental evaluation more fully.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.