acceptodds
Under review as a conference paper at ICLR 2027

Stable Approximate Canonicalization: Stability Obstructions, Selector Ambiguity, and Data-Dependent Trade-offs

13 views

Abstract

Canonicalization obtains invariant predictors by selecting one representative from each symmetry orbit, but a continuous exact selection is often topologically impossible. We show that approximation does not remove this obstruction for free. On any compact obstructed domain, deterministic canonicalizers whose invariance and orbit errors vanish must have unbounded Lipschitz constants. For free finite-group actions on compact connected smooth manifolds, the optimal expected orbit error of an exactly invariant -Lipschitz relaxation is under every density bounded above and below. An antipodally symmetric population fitted to TUM VI orientations supports this rate: over the reported fit range , an explicit quaternion relaxation has a fitted log–log slope , with times the error approaching a nonzero plateau. We then show that continuous relaxations can localize their error near construction-specific selector-ambiguity strata, making data mass near those strata the relevant practical quantity. Experiments on LiDAR tracks, molecules, histopathology, and particle clouds show that occupancy, representation instability, and downstream harm are distinct. These results turn a qualitative obstruction into a quantitative, data-dependent account of when canonicalization is useful and what it costs

Primary area:
learning theory
Keywords:
canonicalization, equivariance, invariant representations, symmetry breaking, group actions, Lipschitz continuity, topological obstruction, approximation theory, orbit selection

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_QobFOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies canonicalization, i.e., the idea to obtain invariance of a prediction to a group action, by considering the orbit of any input, and feeding a fixed member of the orbit into an ordinary predictor. In particular, it considers the case where a continuous selection of a fixed member of the orbit is not possible (referred to as an \textit{obstructed} domain). It shows that Lipschitz constants of an approximation to such a canonicalization must have an unbounded Lipschitz constant as the approximation becomes more invariant to the group action and the distance of the approximation to the true orbit goes to zero. Numerical results support the presented theory, and highlight that there is an inherent trade-off between stability, precision of the approximate canonicalization, and computational costs (by considering/averaging multiple inputs).

Critical strengths

The paper studies an interesting and important problem from a theoretical perspective and discovers an interesting trade-off that is subsequently investigated in several numerical experiments. The work deduces concrete recommendations for future works on approximate canonicalization to make the inherent trade-off visible. Thus, it could be of interest to the community due to novel theoretical insights as well as for its concrete recommendations.

Critical weaknesses

I found this work very difficult to read. It quickly dives into the topic, requires strong prior knowledge on the respective mathematical objects, and does not really preserve a big picture with concrete examples of groups and subsets of metric spaces that lead to an obstruction. Considering that the manuscript did not fully use the available nine pages, the big picture for non-expert readers could have been more pronounced. While this is partially ok for a theory-heavy work, it is unfortunate that not even the experimental section is easily readable and self contained: The group actions associated with the five dataset considered are not explained in the main paper, "protocol details" are in the appendix only, sections dive into technical aspects without an introduction (e.g. Sec. 5.3 starts with "For PCA alignment, we normalize eigenvalue gaps by total variance" without introducing the dataset or task the current section is studying), and neither Table 1, nor Fig. 3, nor Fig. 4 are explained or even referred to in the text.

Questions for authors

Can the authors please provide a bigger picture, e.g. with the help of a concrete example, and introduce the numerical experiments conducted?

This question/suggestion should be viewed in the light of my low confidence about the overall assessment, i.e., if researchers more familiar with the topic find the paper well written and easy to follow, feel free to ignore the above point. I would, however, consider a reference and explanation of all tables and figures as very important in any case. In addition, to my mind, the main paper should always be self contained and thus fully understandable without considering the appendix.

1 comment

We thank the reviewer for the constructive suggestions concerning accessibility, the overall narrative, and the presentation of the experiments. We agree that the original manuscript introduced mathematical concepts too quickly and did not sufficiently explain the experimental settings and their relationship to the theory in the main text. We have substantially reorganized the presentation while preserving the original mathematical results and cross-domain investigation.

1. A concrete running example now introduces the problem

Section 1 now explains the stability obstruction through quaternion sign selection before introducing the formal mathematical framework. Two opposite unit quaternions, qq and −q-q, encode the same physical 3D rotation. Choosing one sign produces an invariant, orbit-valued representation away from the selection boundary, but nearby inputs crossing that boundary can yield distant outputs. Figure 1 illustrates both the discontinuity and the alternative of continuously fading the representation toward zero. The latter removes the jump while sacrificing orbit membership.

This example introduces invariance, orbit fidelity, and stability in concrete terms and reappears in Sections 4 and 5. We believe it makes the central question accessible without requiring prior familiarity with quotient spaces or topological sections.

2. The experimental section is now more self-contained

We have reorganized Section 5 so that the main text explains the purpose, symmetry action, and principal findings of each study. Table 1 explicitly identifies the physical meaning of the group action for all five datasets and summarizes the evidence obtained from each. The experimental subsections now follow a clear progression:

  • Sections 5.2–5.3 (TUM VI, antipodal sign symmetry C2C_2 on S3S^3): Test the inverse-Lipschitz scaling and quantitatively evaluate the localization bound using calibrated quaternion constructions.
  • Section 5.4 (Waymo SO(2)SO(2), SPICE/MD17 SO(3)SO(3), JetClass SnS_n): Ask which selector-ambiguity mechanisms occur frequently across different real-world populations.
  • Section 5.5 (Waymo LiDAR crops, planar rotations SO(2)SO(2)): Investigate which occupied ambiguities cause actual temporal frame tearing, collapse, or retrieval degradation.
  • Section 5.6 (PatchCamelyon histopathology, exact image quarter turns C4C_4): Compare canonicalization, continuous relaxations, augmentation, and group averaging under strictly controlled training.
  • Section 5.7 (SPICE molecules SO(3)SO(3) and JetClass particles SnS_n): Examine downstream consequences in molecular force prediction and particle-cloud sorting.

Essential definitions of representation sensitivity, training controls, evaluation settings, and uncertainty are now provided in the main text, with detailed protocols and audit tables retained in the appendix. We also clarify distinctions that were previously insufficiently explained, including physical pose changes versus canonical-object changes and molecular energy invariance versus force equivariance.

3. Figures and tables are now explicitly introduced and interpreted

We have revised the presentation so that every main-text figure and table is referenced and explained in context:

  • Figure 1 illustrates the canonicalization obstruction and its continuous off-orbit relaxation on the quaternion sphere.
  • Figure 2 connects the quantitative theory to experiments: panel (a) evaluates inverse-budget scaling (L=1/tL = 1/t), while panel (b) compares measured localization error with its calibrated metric bound.
  • Figure 3 compares ambiguity-score occupancy across populations and motivates the subsequent analysis of actual failures.
  • Figure 4 presents the joint trade-off among orbit fidelity, representation sensitivity, prediction quality (AUROC), and computational cost on PatchCamelyon.

Tables 1 and 2 are likewise introduced and discussed in the main text, with Table 2 explicitly distinguishing orbit fidelity from invariance. The main results and their interpretations no longer require consulting the appendix, while complete reproducibility details remain available there.

We appreciate the reviewer's emphasis on communicating the big picture. The revised narrative now proceeds from a concrete obstruction, through its quantitative characterization, to calibrated evidence and data-dependent consequences, making the contribution accessible to a broad machine learning audience.

Reviewer_8ceCOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper quantifies the trade-off between stability and orbit fidelity in canonicalization, proving that optimal expected distortion scales as 1/L1/L under specified assumptions, where LL is the allowed Lipschitz constant. Experiments show that practical consequences depend on the data distribution and downstream task: smoother representations can improve or worsen prediction.

Critical strengths
  1. Matching upper and lower bounds establish the optimal1/L expected distortion rate under clearly stated assumptions, turning a qualitative impossibility result into a quantitative stability–fidelity trade-off.

  2. The paper distinguishes ambiguity frequency, representation instability, and downstream prediction error. Experiments across diverse domains demonstrate why these quantities must be evaluated separately when judging the usefulness of canonicalization.

Critical weaknesses

None

Questions for authors
  1. In the real-data downstream comparisons, were the backbone architectures, training budgets, data splits, and tuning procedures matched across canonicalization methods? These controls determine how confidently the reported performance differences can be attributed to canonicalization.

  2. What precisely is the “hard exact C4C_4" representation used for PatchCamelyon, and how is its representation distance computed? Does it select one transformed image or retain/aggregate multiple orbit elements? This affects how you should interpret its reported stability relative to the single-representative obstruction.

2 comments

Thank you for the positive assessment of our theoretical contribution and for highlighting the distinction between ambiguity frequency, representation instability, and downstream performance. Your questions motivated a strictly controlled PatchCamelyon study and selector-directed sensitivity probes, now added to the manuscript alongside clarified scopes for our cross-domain mechanism studies.

Q1. Were architectures, training budgets, data splits, and tuning procedures matched?

In the initial submission, cross-domain studies served as mechanism diagnostics rather than strictly paired causal benchmarks (e.g., unpaired initializations and differing backbone details or subset sampling). We added a strictly controlled five-seed PatchCamelyon study (Section 5.6, Table 2, Figure 4; protocol in Appendix B.6) enforcing:

  • Architecture & Initialization: Identical CNN (578,978 trainable parameters) across all 15 configurations, with identical initial weights within each seed (verified via saved hashes).
  • Training, Validation & Test Splits: Within each seed, the same 40,000 training patches, batch-index sequence, and 5,000 updates; fixed 8,000-patch validation set for update-count pilot and relaxation operating-point selection; fixed balanced 32,754-patch test set from the official PatchCamelyon test split.
  • Stability Evaluation: Same 4,096 test patches, shared perturbation probes, and equal-budget per-method searches over r∈{0.1,0.3,1}r \in \lbrace 0.1, 0.3, 1 \rbrace. We report r=0.1r = 0.1 as headline radius because seed-0 diagnostics showed branch-switch saturation at r=1r = 1 (96.7%96.7\% vs. 21.8%21.8\% at r=0.1r = 0.1; used only for reporting, not model selection; all radii retained).
  • Replication & Compute: Five paired seeds; measured inference latency (four-pass averaging is explicitly not compute-matched to one-pass methods).

An added frozen-encoder control separates preprocessing from retraining effects: separately trained pipelines shift AUROC vs. raw input by −0.0058±0.0044-0.0058 \pm 0.0044, whereas fixing each seed's augmentation-trained encoder and changing only test-time preprocessing gives −0.0010±0.0011-0.0010 \pm 0.0011. Thus, we do not attribute the full retrained gap to canonicalization alone. Other real-data domains are retained explicitly as descriptive mechanism studies (Sections 5.5, 5.7, Appendix B), clarifying Waymo's annotated box-heading reference, JetClass's pTp_T-sorted inputs and unmasked padding, and the 256-conformer single-molecule SPICE force-sensitivity subset.

Q2. What exactly is the hard C4C_4 representation, and how is its distance measured?

The hard C4C_4 canonicalizer selects one transformed image rather than retaining or averaging multiple orbit elements. Let RR be the exact quarter-turn pixel permutation, xˉ\bar{x} the channel-mean image, and uk=Rkxˉu_k = R^k \bar{x}. The orientation scores are

sk(x)=(Dxuk)2‾−(Dyuk)2‾+(Dxuk)3‾,s_k(x) = \overline{(D_x u_k)^2} - \overline{(D_y u_k)^2} + \overline{(D_x u_k)^3},

where DxD_x, DyD_y are central pixel differences and the overline denotes spatial averaging. The canonicalized image is

C(x)=Rk∗(x)x,k∗(x)=arg⁡max⁡ksk(x).C(x) = R^{k^{\ast}(x)} x, \qquad k^{\ast}(x) = \arg\max_k s_k(x).

Scores cyclically reindex under quarter turns, making C(x)C(x) invariant wherever the maximizer is unique. While first-index tie-breaking is not globally invariant, zero exact top-score ties occur on the 32,754 test patches, and enumerating all four rotations yields zero measured invariance defect.

Representation sensitivity uses ordinary Euclidean distance between canonicalized images—with no minimization over group rotations (orbit fidelity is evaluated separately):

S^robj(x)=max⁡δ∈Ar(x)∥C(x+δ)−C(x)∥2∥δ∥2.\widehat{S}_{r}^{\mathrm{obj}}(x) = \max_{\delta \in \mathcal{A}_{r}(x)} \frac{\Vert C(x+\delta) - C(x) \Vert_{2}}{\Vert \delta \Vert_{2}}.

The submitted p95 of 1.001.00 came from random-direction probing at r=1r = 1: when the selected rotation does not change, the quotient equals 11 because quarter-turn permutation is an isometry, missing branch-switch discontinuities. Adding selector-directed probes toward ties reveals selection changes on 21.8%21.8\% of patches at r=0.1r = 0.1 (raising hard-representation p95 to ≈521\approx 521) and 96.7%96.7\% at r=1r = 1.

Importantly, this large representation-sensitivity gap does not translate into a comparable prediction-robustness gap: under the full finite search (with per-method PGD), prediction p95 is 3.543.54 for raw input vs. 3.683.68 for hard selection. The revised manuscript now explicitly distinguishes the single-image representation, orbit fidelity, selector-switch sensitivity, and downstream prediction behavior.

We appreciate these questions, which led to sharper definitions and substantially stronger controls.

Thank you! These address my comments sufficiently.

Reviewer_9XGLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the trade-offs among invariance, orbit preservation, and Lipschitz stability in symmetry canonicalization. Building on established topological obstructions to continuous exact canonicalization, the authors further investigate the error levels that approximate canonicalization can achieve under stability constraints. The paper first proves that, on a compact domain admitting no continuous exact canonicalization map, the Lipschitz constants of approximate maps cannot remain uniformly bounded if both their invariance and orbit errors converge to zero. It then establishes a Θ(1/L) decay rate for the optimal expected orbit error of exactly invariant maps under free finite-group actions satisfying specific regularity conditions. In addition, the paper proposes a localized continuous relaxation and controls its expected orbit error through the probability mass of the data distribution near the selector-ambiguity set. Experiments on TUM VI, Waymo, SPICE/MD17, PatchCamelyon, and JetClass examine the relationships among stability, ambiguity in representative selection, and downstream task performance. Although the paper has some mathematical merit, I have substantial reservations about the significance of its overall contribution, the practical scope of its theoretical results, and the adequacy of its experimental validation.

Critical strengths

Strength 1: Quantitative error analysis with matching upper and lower bounds for a specific class of symmetry groups. The paper's main strength is that, under free finite-group actions and suitable regularity assumptions, it establishes a Θ(1/L) decay rate for the optimal expected orbit error as a function of the Lipschitz constant. This result provides not only an error upper bound for a particular algorithm but also a lower bound applicable to the entire class of feasible maps, giving it clear mathematical significance. However, this contribution primarily quantifies the cost of approximation in the presence of known topological obstructions; its practical impact still requires further demonstration.

Strength 2: Distinguishing different sources of canonicalization instability and their downstream effects. The paper emphasizes that ambiguity-region occupancy, representation sensitivity, and final task performance are three distinct quantities, and demonstrates across multiple datasets that their relationships need not be straightforward. This analysis helps prevent representation stability from being directly equated with predictive performance and is a practically useful aspect of the paper.

Critical weaknesses

Weakness 1: The theoretical contribution has a narrow scope, and its significance is not sufficiently established. The paper's most important theoretical result is the Θ(1/L) rate for the optimal expected error. However, this result relies on relatively strong conditions, including free finite-group actions, compact connected smooth manifolds, and a data density bounded below by a positive constant and above by a finite constant. These conditions make the theoretical analysis tractable but also substantially restrict its applicability. In particular, practical geometric learning problems often involve continuous groups such as SO(2) and SO(3), nontrivial stabilizers, or nonuniform, low-dimensional data distributions. More importantly, the paper's main theoretical framework does not change the fundamental understanding established by prior work: in the presence of topological obstructions, deterministic exact canonicalization cannot simultaneously maintain continuity and the required invariance. The paper further characterizes this phenomenon through a quantitative error rate, but does not sufficiently explain why this rate is of decisive importance for practical machine learning problems. For example, the paper establishes neither new prediction-error bounds nor generalization-error bounds, and does not demonstrate that its theoretical results can directly guide the design of learning methods with better task performance. Thus, although the optimal error rate is mathematically nontrivial, I find that its broader significance and incremental contribution relative to existing theory have not been sufficiently established.

Weakness 2: There is a substantial disconnect between the experiments and the theoretical contribution, with insufficient evidence of practical value. The paper covers multiple real-world datasets, but experimental breadth cannot substitute for direct validation of the core theoretical claims. The TUM VI experiment reports a decay exponent close to the theoretically predicted -1, but primarily examines a prespecified family of quaternion relaxation maps under a fitted symmetric probability distribution. It verifies the consistency of this construction with the theoretical prediction, but does not directly validate optimality over all admissible maps. Meanwhile, the Waymo, SPICE/MD17, PatchCamelyon, and JetClass experiments mainly reveal instability in existing canonicalization or continuous relaxation methods and its effects on downstream tasks. These observations are useful, but do not sufficiently demonstrate how the paper's new theory leads to algorithms that outperform existing methods. In particular, the proposed localized relaxation obtains continuity by allowing outputs to deviate from the original orbit. This mechanism is intuitively straightforward, and the paper does not demonstrate that it can systematically outperform other smoothing, group-averaging, multi-frame, or equivariant learning methods on practical tasks. Furthermore, the ambiguity scores used in the paper lack a complete numerical calibration to the metric-neighborhood distances in the theory. Consequently, the theoretical upper bounds involving probability mass near ambiguity regions have not received direct quantitative experimental validation. Overall, the experiments do more to illustrate potential problems with canonicalization than to establish that the paper offers a practically effective new solution. This weakens its overall contribution as a machine learning paper.

Questions for authors

Q1. What is the essential difference between the theoretical innovation here and prior work? Prior work has already established topological obstructions to continuous exact canonicalization under certain group actions. Beyond quantifying this conclusion as a Θ(1/L) error rate under specific conditions, does the paper provide theoretical insights that can substantively change algorithm design? Could the authors give a concrete learning problem in which this optimal rate helps select a model or canonicalization method in a way that cannot be inferred from existing impossibility results alone?

Q2. What guarantees does the main theorem actually provide for practical problems involving continuous groups? Several applications in the paper involve continuous rotation groups, whereas the strongest matching upper and lower bounds apply only to free finite-group actions. For these continuous-group experiments, could the authors identify which specific conclusions are rigorously supported by the main theorem and which are only empirical analogies or mechanism analyses? If the main applications do not satisfy the theorem's assumptions, how do the authors assess the practical guidance that this theoretical result offers for those applications?

Q3. What verifiable practical advantages does the new construction offer over existing methods? The localized continuous relaxation can reduce discontinuities in canonicalization, but also allows outputs to deviate from the original group orbit. Could the authors provide direct controlled comparisons of their relaxation with existing soft canonicalization, group averaging, or multi-frame methods, using the same encoder, training budget, and task setting? I would particularly like to see the joint trade-off among improved stability, orbit distortion, and downstream predictive performance, rather than changes in only one of these metrics.

1 comment

Thank you for recognizing the matching bounds and the distinction between ambiguity occupancy, representation instability, and downstream harm. We revised the manuscript to clarify what our quantitative theory adds beyond impossibility results, show its design implications, and strengthen links to experiments, focusing on the optimal stability–fidelity trade-off rather than a new architecture.

Q1. What does the quantitative result add beyond prior impossibility results, and how can it guide design?

Prior work asks whether a continuous exact canonicalizer exists; we ask: given a finite stability budget, what is the smallest achievable expected orbit distortion? Theorem 2 gives matching Θ(1/L)\Theta(1/L) bounds for exactly invariant ambient-valued maps under free finite-group actions and bounded densities:

  • Optimality & Achievability: The lower bound holds for all admissible maps regardless of algorithm, and an explicit construction achieves this rate.
  • Population dependence: Geometry and density govern the trade-off constants.

Unlike qualitative impossibility, this rules out infeasible stability–fidelity pairs. In Section 3 (Theorem 3, antipodal circle), requiring orbit error ≤0.01\leq 0.01 and L≤20L \leq 20 gives ηinv(20)≥121π≈0.01516>0.01,\eta^{\mathrm{inv}}(20) \geq \frac{1}{21\pi} \approx 0.01516 > 0.01, so no admissible map satisfies both (the bound requires L≥30.83L \geq 30.83; our explicit construction achieves it at L=31.84L = 31.84). These thresholds apply to the circle model; our theory constrains representation design, not generalization or accuracy.

Q2. Which guarantees extend to continuous-group applications, and how is the theory validated?

Revised Sections 3–5 and Table 1 separate three levels of evidence:

  1. Mathematical optimality: Theorems 2–3 cover free finite-group actions on compact connected smooth manifolds under bounded densities; they do not establish a matching 1/L1/L rate for continuous groups (SO(2)SO(2), SO(3)SO(3)) or varying stabilizers.
  2. Construction behavior & calibrated bounds: On smooth antipodally symmetric populations fitted to TUM VI orientations, Figure 2a evaluates the clip map against its exact Lipschitz constant L=1/tL = 1/t, giving log–log slope −0.9917-0.9917 (refining our initial fit using the looser bound 1+1/t1 + 1/t); optimality over all admissible maps follows analytically. To address uncalibrated ambiguity scores, we numerically evaluate the localization bound: for the quaternion selector, D(q)=arcsin⁡∣qw∣D(q) = \arcsin |q_w| and s(q)=∣qw∣s(q) = |q_w| (11-Lipschitz), giving E d(ft(q),{q,−q})≤P(D<2t)≤P(∣qw∣<2t).\mathbb{E}\, d(f_t(q), \{q, -q\}) \leq P(D < 2t) \leq P(|q_w| < 2t). At t=1/32t = 1/32, expected orbit error is 0.041620.04162 on the TUM VI population vs. 0.235270.23527 on an ambiguity-concentrated population—a 5.7-fold difference for the same map and cutoff (bounds: 0.055200.05520 and 0.310850.31085; TUM VI error-to-bound ratio 0.7540.754, near the 3/43/4 limit; Figure 2b).
  3. Downstream diagnostics: Sections 5.4–5.7 (Waymo, molecules, histopathology, particle clouds) use ambiguity scores as diagnostics rather than calibrated distances; the general localization theorem bounds the specified map under its assumptions without guaranteeing a matching rate or Lipschitz conditioning.

Q3. What practical advantages do the controlled comparisons demonstrate?

We added a controlled 5-seed PatchCamelyon study comparing hard selection, soft canonicalization, localized relaxation, augmentation, and image/logit/feature averaging using identical CNNs, paired seeds/batches, update counts, fixed splits, validation-tuned operating points, and measured latency (Section 5.6, Table 2, Figure 4):

  • Hard selection: preserves orbit fidelity, with high representation sensitivity (selector-directed p95 ≈521\approx 521, AUROC 0.90170.9017).
  • Soft mixing (ε=1\varepsilon = 1): lowers p95 to 51.751.7, with relative orbit error 0.1910.191 and AUROC 0.85080.8508 (validation-tuned: 0.8983±0.00630.8983 \pm 0.0063).
  • Localized score-gap relaxation (t=0.005t = 0.005): does not improve the trade-off (p95 ≈999\approx 999, orbit error 0.0330.033, AUROC 0.89600.8960).
  • Logit averaging: reaches AUROC 0.92470.9247, at 3.79×3.79\times inference latency.

Why is this score-gap map steep on PatchCamelyon while localized constructions are well-conditioned in our calibrated quaternion fade (L=1+t−2L = \sqrt{1 + t^{-2}}, Figure 2b) and synthetic SO(3)SO(3) audit (Appendix C, Table 10: t=0.05t = 0.05 cuts fixed-radius p95 from 259.8259.8 to 2.192.19)? Per Section 4 (Proposition 2, Corollary 3), cutoff continuity alone does not yield O(1/t)O(1/t) Lipschitz conditioning unless the ambiguity score has controlled Lipschitz regularity (CjC_j), and localization bounds orbit distortion rather than accuracy. Rather than claiming superiority over averaging or equivariant models, we show how stability, fidelity, accuracy, and cost trade off jointly.

We hope these revisions address your concerns and welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.