acceptodds
Under review as a conference paper at ICLR 2027

Multi-Anchor LOT: from Local Transport Maps to Global Wasserstein Geometry

Abstract

Linearized optimal transport (LOT) provides Euclidean representations of distributions whose coordinate distances approximate the Wasserstein distance. However, linearization at a single reference cannot generally preserve the nonflat geometry of Wasserstein space. On regular Wasserstein models, we quantify this limitation through a local distortion bound governed by curvature and squared reference distance. This analysis motivates Multi-Anchor Linearized Optimal Transport (MALOT) for recovering global Wasserstein geometry from local transport maps. Starting from a multi-anchor Riemannian metric that gives greater weight to nearby references, we derive a symmetric two-point approximation with endpoint-dependent weights on squared LOT separations. Radial completion corrects each discrete coordinate block to preserve its reference radius. We prove that the resulting dissimilarities converge uniformly to the Wasserstein distance as the anchor covering radius and weight-localization error vanish, and establish coverage guarantees for farthest-point sampling. Simulations demonstrate improved recovery over LOT and decreasing error with finer anchor coverage; real-data studies evaluate distance recovery and downstream prediction.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Linearized Optimal Transport, Riemannian Geometry, Wasserstein Space

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

The paper addresses the dependence of linearized optimal transport (LOT) on its reference measure. A single reference keeps each target's distance to the reference exact, but the distance between two targets can be distorted when they are far from the reference, and Prop. 1 makes this precise locally through curvature. MALOT uses several references (anchors) chosen by farthest point sampling. For each pair of measures it averages the squared per-anchor LOT distances with softmax weights that favor anchors close to either measure, so the weights change from pair to pair. For discrete plans that split mass, radial completion appends one scalar per anchor so that each block keeps the exact distance to its anchor. Theorem 2 bounds the error by the distances of the two measures to their nearest anchors plus a temperature term, for any finitely supported measures with exact reference plans, and Theorem 1 gives a smooth Riemannian counterpart. The validation measures stress against exact W2 on synthetic families and seven real datasets, and downstream classification on three of them, with single-reference LOT, concatenated LOT and sliced Wasserstein as baselines.

Critical strengths
  1. The idea is simple and, as far as I know, new. Previous works with several LOT references stack them into one long vector, while here each pair mostly relies on the references that are close to it. The four-point example in App. A.5 is a nice way to show why any fixed embedding has to distort W2 somewhere. The method only needs OT between the anchors and the data, and radial completion is a cheap and useful fix for discrete plans. This is the main reason I think the paper has value, since the dependence on the reference is a well known weakness of LOT and the idea is easy to adopt.
  2. The gains over single-reference LOT are consistent and carefully measured. Errors are computed against W2 itself without rescaling, the real data splits are by date, object or run, and the oracle experiment in Sec. 5.2 gives concatenation and MALOT the same references, so it isolates the effect of the weighting. Without radial completion, MALOT reduces the error of single-reference LOT on all seven real datasets, by roughly 10% to 33%, and the paper also reports the cases where things do not work.
Critical weaknesses
  1. The theory does not really explain why the method works. Theorem 2 only relies on each block having the correct distance to its anchor, and then on the triangle inequality, as the authors say themselves in App. A. So the same bound would hold for a method that keeps only the distances to the anchors and throws away the transport maps. The paper actually has such a method, the anchor-radius control in Table 2, and it is much worse than MALOT. So what makes MALOT work seems to be the transport directions, and the theory says nothing about them. The bound is also loose in my opinion. On the crescents with 8 anchors it is larger than every distance in the data. Finally, the same kind of bound for a reference-based OT distance already appears in Mialon et al. (ICLR 2021, Lemma 3.1), along with an analysis of several references, and this work is not cited.

  2. I think the experiments do not show that averaging over the anchors near the two endpoints is the right way to use them. When the plans from an anchor are maps, as for Gaussians or point clouds of the same size, each per-anchor LOT distance can only overestimate W2. In that case, simply taking the smallest of these distances can never do worse than the MALOT average, and there is nothing to tune. This is the setting of the rotating shapes experiment in Sec. 5.2 and of Stanford96. Using only the anchor closest to one of the two measures is another simple option. Aside from the anchor-radius control, none of these is compared, so the experiments mainly show that MALOT beats uniform concatenation. The evidence on the real datasets also seems limited. Radial completion, which is part of the default method, makes the error worse on several datasets, the accuracy gains over concatenation are small and given without error bars (on Fingering the gain is about one test frame, and concatenation is even better on ETH-80), and there are no runtimes.

Questions for authors
  1. What does Theorem 2 add compared to Lemma 3.1 of Mialon et al. (2021) and their analysis of several references? And which variant do you recommend by default, with or without radial completion?
  2. How does MALOT compare with the simpler options above, using the same anchors, i.e. the smallest per-anchor distance when the plans are maps, and the single closest anchor? Could you also give the spread across splits and some runtimes?
  3. I think a result that actually uses the transport directions would make the paper much stronger, for example a faster rate when the maps from the anchors are smooth.
6 comments

We are grateful for your assessment of the original coverage guarantee and for drawing our attention to Mialon et al. (2021). Your comments helped us clarify what the anchor radii guarantee and how transport directions improve the analysis of individual pairs.

Weakness 1. Coverage, transport directions, and prior work

What the original bound establishes. We agree that the original guarantee relies on radial exactness, the endpoint weights, and the triangle inequality. The same argument applies to scalar anchor radii, so it establishes uniform recovery as the covering radius and τlog⁡K\tau\log K vanish. The proof leaves open how directional information affects a given pair. On the eight-anchor crescents, the uniform bound for radially completed MALOT is 1.3700, above the exact W2W_2 diameter of 1.1544. In the revised manuscript, we have retained this guarantee as Lemma 3 (Appendix A.2, Eq. (17)) and made its scope explicit.

How the revised bound uses directions. We have revised Theorem 2 (Section 4.1, Eq. (13)) to use the alignment of the LOT displacements. Write Wij=W2(μi,μj)W_{ij}=W_2(\mu_i,\mu_j) and rki=W2(ρk,μi)r_{ki}=W_2(\rho_k,\mu_i). At a point drawn from anchor kk, independently sample the two targets conditional on that point using the two optimal reference plans. This gluing gives a feasible target-pair coupling. With barycentric displacement blocks ϕki\phi_{ki} and ϕkj\phi_{kj}, its squared cost is

Gk2=rki2+rkj2−2⟨ϕki,ϕkj⟩.G_k^2=r_{ki}^2+r_{kj}^2-2\langle\phi_{ki},\phi_{kj}\rangle.

The inner product measures how the displacements align. Feasibility gives Wij≤GkW_{ij}\le G_k, and the L2L^2 triangle inequality gives Gk≤rki+rkjG_k\le r_{ki}+r_{kj}. Therefore the directional upper endpoint Uijdir=min⁡kGkU_{ij}^{\mathrm{dir}}=\min_k G_k can tighten the radius-only upper endpoint Uijrad=min⁡k(rki+rkj)U_{ij}^{\mathrm{rad}}=\min_k(r_{ki}+r_{kj}). Both intervals have the lower endpoint ℓij=max⁡k∣rki−rkj∣\ell_{ij}=\max_k|r_{ki}-r_{kj}|. Theorem 2 bounds the absolute error of either MALOT estimate DD by

∣D−Wij∣≤max⁡∣D−ℓij∣,∣D−Uijdir∣.|D-W_{ij}|\le\max\\{|D-\ell_{ij}|,|D-U_{ij}^{\mathrm{dir}}|\\}.

The bound uses the actual endpoint-weighted estimate and can improve its certificate even with the same anchor bank and weights. It is at most the corresponding radius-interval bound. For completed blocks it is also at most the original pair-specific coverage bound. For uncompleted blocks, the corresponding comparison with the coverage bound includes the computable difference between the two MALOT estimates. The certificate uses the reference plans already computed for MALOT and requires no additional target-pair transport solve (Appendix A.6).

What can be recovered at fixed coverage. To show what directions add when the radii agree, we have included a centered eight-atom example in Section 4.1 and Appendix A.6. Each distribution has the law of a uniformly random sign vector multiplied coordinatewise by positive scales. The target scales lie on a circle in a plane perpendicular to the line of anchor scales. Every target has the same radius vector at covering radius r>0r > 0, but distances between distinct targets range over (0,2r](0,2r]. Any estimator based solely on the two radius vectors therefore has worst-case absolute error at least rr. The optimal plans match signs, so every LOT separation equals the exact target-pair distance. Both MALOT variants and uniform LOT aggregation recover all the distances in this construction.

Bounds on the crescents. Your comparison with the data diameter underscored the importance of checking how conservative the bound is. We have added a comparison in Appendix D.3 of the maximum pairwise bounds and the observed maximum absolute errors at K=8K=8.

LOT blocksCoverage boundRadius intervalDirectional intervalObserved error
Completed1.08440.85810.43600.0938
Uncompletedn/a0.85900.43230.0929

For the uncompleted estimator used in the simulations, the directional interval strictly tightens the certificate on 902 of 2,775 pairs. Against exact W2W_2, uncompleted MALOT has smaller observed absolute error than the pivot midpoint on 1,433 of these pairs. The maximum certificate remains substantially above the observed maximum error. In the separate fixed-weight control, replacing exact anchor-radius differences with the barycentric LOT separations reduces raw stress from 0.3549 to 0.0325 (Table 6).

We have added citations to Mialon et al. (2021) in the Introduction and Appendix A.6. Their shared-reference coupling is the basis for the directional upper endpoint. Our response to Question 1 explains how their bounds relate to the interval certificate for the endpoint-weighted estimator.

We appreciate your suggestions to compare Min-LOT and the single closest anchor. These controls help us assess the effect of endpoint weighting, and your observations about radial completion, accuracy, and runtime have helped us identify where the real-data evaluation needs further evidence.

Weakness 2. Matched baselines and real-data evidence

When Min-LOT has a pointwise advantage. Your common-source argument shows why the minimum is a strong baseline for deterministic maps. We agree that it applies to Stanford96's Gaussian transports and the equal-mass assignment plans in the original rotating-shape simulation. When the two reference transports are deterministic maps, applying them to the same anchor draw yields a feasible coupling. The per-anchor LOT separation is therefore at least the true distance W=W2(μi,μj)W=W_2(\mu_i,\mu_j). On the same bank, taking the minimum gives no larger pointwise absolute error than a weighted root mean square of the same separations. The minimum requires no temperature.

What changes under split plans. In the revised manuscript, we have updated the rotating multi-cluster comparison (Section 5.2) to use compressed anchors with eight equally weighted atoms and targets with 80 equally weighted atoms. The reference plans necessarily split mass in this setting. A barycentric block retains the conditional mean displacement, so for glued-coupling distance GkG_k and uncompleted LOT separation LkL_k,

Lk2−W2=(Gk2−W2)−(Vki+Vkj),L_k^2-W^2=(G_k^2-W^2)-(V_{ki}+V_{kj}),

where VkiV_{ki} and VkjV_{kj} are the splitting variances. The variance term can exceed the coupling gap even though Gk≥WG_k\ge W. With radial completion, the subtracted term is 2VkiVkj2\sqrt{V_{ki}V_{kj}}, so a split-plan LOT separation can still underestimate WW. Min-LOT selects the smallest separation, even if it is an underestimate. Let MM be that minimum for the chosen block variant. On those same blocks, MALOT has strictly smaller absolute error than Min-LOT exactly when M<DMALOT<2W−MM < D_{\mathrm{MALOT}} < 2W-M.

Same-bank comparison with Min-LOT. We have added Min-LOT and Nearest-one at matched anchors and budgets (Tables 3–5, Appendices D.1–D.2). We use uncompleted blocks and compute raw stress against exact W2W_2 before embedding. On the rotating multi-cluster data, stress is 0.1094 for Min-LOT and 0.0508 for MALOT on the FPS bank, and 0.1141 versus 0.0595 on the oracle bank. Min-LOT underestimates 6,965 of 7,140 FPS pairs (97.55%) and 6,961 oracle pairs (97.49%). MALOT has strictly smaller absolute error on 6,498 FPS pairs and 6,791 oracle pairs. On the crescents, MALOT lowers full-pair stress at K=8,16,24K=8,16,24, including 0.0581 to 0.0173 at K=24K=24. The full-pair results nearly tie at K=4K=4, and Min-LOT gives lower stress on the 2,485 pairs with neither endpoint an anchor (0.0596 versus 0.0644).

Same-bank comparison with the single closest anchor. Nearest-one selects the anchor closest to either endpoint and uses that chart for both targets. Hard endpoint averages the squared separations from the two endpoints' nearest charts, distributing each endpoint's weight uniformly across ties. At positive temperature, MALOT can retain nearby charts beyond the closest ones. Comparing Nearest-one with Hard endpoint assesses the effect of including both endpoints' nearest charts. Comparing Hard endpoint with MALOT on the same blocks then assesses positive temperature. We have included an example in Appendix A.7 where the nearest chart gives 10, a nearby chart gives the exact distance of 8, and positive-temperature MALOT moves the estimate to 9.345.

Hard endpoint and MALOT give lower full-pair stress than Nearest-one in all six matched settings. On the K=24K=24 crescents the values are 0.0185 for Hard endpoint, 0.0173 for MALOT, and 0.0326 for Nearest-one. Hard endpoint has lower full-pair stress than MALOT at crescent K=4,8K=4,8, while MALOT has lower stress at K=16,24K=16,24 and on both multi-cluster banks. On the 2,211 non-anchor pairs at K=8K=8, Nearest-one has lower stress than MALOT (0.0326 versus 0.0342), and MALOT has lower maximum absolute error (0.0929 versus 0.1441).

Real-data follow-up. We are adding same-anchor comparisons with Min-LOT and Nearest-one on Stanford96 and oracle Concat LOT at matched anchors and KK on real datasets. The current Table 1 shows mixed effects of radial completion and small classification accuracy differences. We are checking the effect of radial completion across datasets, split-level accuracy on Fingering and ETH-80, and wall-clock costs by stage. We will share an update promptly once these real-data results are available.

We are grateful for your questions on prior bounds, split variability, and local rates. They helped us clarify the scope of our results.

Question 1. Relation to Mialon et al. and the recommended radial setting

Your comparison with Mialon et al. helped us clarify Theorem 2. For unregularized plans, their Lemma 3.1 bounds the single-reference glued distance error by twice the smaller anchor radius. Appendix B.2 gives a pairwise bound and a dataset-wide mean squared error bound for the uniform RMS of glued distances. In the revised manuscript, we have used this gluing to obtain a directional upper endpoint for W2W_2. With a radius-based lower endpoint, it bounds either endpoint-weighted MALOT variant's error (Section 4.1 and Appendix A.6). Lemma 3 gives the uniform coverage and temperature guarantee for radially exact blocks.

Your question about the default helped us distinguish radial exactness in the coverage theorem from its empirical effect. We have specified uncompleted barycentric LOT blocks as the empirical default (Section 3.2, Algorithm 1). With optimal plans, completion restores exact anchor radii. It has no effect for deterministic maps and adds a nonnegative term to squared MALOT distances under split plans (Appendix C.1, Eq. (28)). It can reduce an underestimate or increase an overestimate. Holding banks, plans, and weights fixed on crescents, it raises stress from 0.0325 to 0.0383 at K=8K=8 and lowers it from 0.0173 to 0.0046 at K=24K=24 (Appendix D.3). In the current Table 1, completion raises stress on Caulobacter and ETH-80 and slightly lowers it on Fingering. Further real-data checks are in progress.

Question 2. Same-bank comparisons, variability across splits, and runtime

Your request helped us assess split variability and computational cost. On the same bank with deterministic maps, Min-LOT is pointwise at least as accurate as MALOT. With uncompleted blocks, MALOT has lower full-pair raw stress than Min-LOT and Nearest-one in all six matched settings, with non-anchor exceptions (Weakness 2, Tables 3–5). The matched Stanford96 comparisons are in progress.

We have added a five-fold held-out evaluation for each simulation (Appendix G, Table 10). Crescent rotations of each base shape stay together, and each multi-cluster fold holds out two targets per family. We select FPS anchors and calibrate the temperature on training data. The oracle uses fixed generating prototypes. Methods share anchors and uncompleted blocks within each fold, with test-pair raw stress against exact W2W_2.

Multi-cluster bankMALOTMin-LOTNearest-oneHard endpoint
FPS0.0508 ± 0.00440.1116 ± 0.00470.0853 ± 0.00600.0613 ± 0.0019
Oracle0.0584 ± 0.00350.1127 ± 0.00510.0812 ± 0.00230.0622 ± 0.0028

These are means and sample standard deviations. The training sets overlap, so the spread describes variation across these folds. On crescents, Min-LOT and Nearest-one have the lowest mean stress at K=4K=4 and 88, respectively. MALOT has the lowest mean at K=16,24K=16,24.

We have reported setup, matrix assembly, and optional full classical MDS separately for both simulations (Table 2). In a separate single-threaded K=24K=24 crescent benchmark using the same stored radii and blocks, median assembly takes 4.58 ms for Nearest-one, 40.95 ms for Hard endpoint, and 41.40 ms for MALOT (five interleaved runs, Appendix C.3). This benchmark excludes setup and MDS.

In Table 1, Fingering accuracy is 93.99% for MALOT versus 93.23% for Concat LOT. On ETH-80 it is 82.07% for Concat LOT versus 81.98% for uncompleted MALOT. We are checking split-level accuracy and timing real-data runs by stage. We will share the results promptly once available.

Question 3. A local recovery rate that uses transport directions

We appreciate your suggestion to quantify the benefit of directions under smooth maps. We have added a finite-pair bound for Eq. (10) in Appendix B.6. Under the smooth exact-map and geometric hypotheses of Theorem 1, we integrate LOT differential distortion along a piecewise smooth minimizing Wasserstein geodesic contained in the analysis region and apply the endpoint weights. For W=W2(μ,ν)>0W=W_2(\mu,\nu) > 0 and ζ=h2+τlog⁡K\zeta=h^2+\tau\log K, Eq. (25) gives

1≤DR(μ,ν)W≤1+A(ζ+Wζ+W2/3).1\le\frac{D_{\mathcal R}(\mu,\nu)}{W}\le\sqrt{1+A(\zeta+W\sqrt{\zeta}+W^2/3)}.

With bounded geometric constants, W=O(h)W=O(h) and τlog⁡K=O(h2)\tau\log K=O(h^2) give O(h2)O(h^2) relative and O(h3)O(h^3) absolute error. At the same temperature scale, Lemma 3 gives a general O(h)O(h) absolute bound for radially exact blocks.

We have also added a finite-plan ablation (Appendix D.3) that replaces each LOT separation by the absolute difference of its two block norms. Removing the angular term at fixed norms, plans, anchors, and weights raises raw stress at every crescent budget, from 0.0325 to 0.3636 at K=8K=8.

Reviewer_PGXDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the reference-dependent local distortion bound of single-reference LOT, which is governed by curvature and the squared reference distance (Prop 1). This limitation motivates the use of multiple anchors to localize the tangent approximation. Specifically, MALOT combines the quadratic forms induced by anchor-specific LOT maps on a common target tangent space, favoring anchors near either endpoint.

Starting from a multi-anchor Riemannian metric, MALOT yields a pairwise dissimilarity that satisfies symmetry, non-negativity, and zero self-distance. Radial completion preserves the reference-to-target distance using a single scalar correction, while FPS constructs an anchor bank where coverage is most needed. By reusing the reference transports at the two targets, MALOT avoids target-to-target OT computations.

This paper proves that the resulting dissimilarities converge uniformly to the true Wasserstein distance as the anchor covering radius and weight-localization error vanish (Thm 2), and MALOT-based MDS converges to exact Wasserstein MDS up to orthogonal alignment as the distance error vanishes (Prop 3).

Critical strengths
  1. Practicality and computational efficiency MALOT effectively balances geometric locality with computational efficiency. By reusing nKnK reference transports, it obtains all pairwise dissimilarities, proximity weights, radial completion, and FPS coverage updates without any pairwise target-to-target OT. Moreover, out-of-sample projection is computationally lightweight: embedding a new distribution via fixed Gower projection requires only KK reference transports and its MALOT distances to the training set. In particular, the radial-completion construction is simple but useful. Without requiring extra OT solves, it exploits the variance decomposition of the barycentric projection and its residual conditional variance, it ensures that the discrete LOT block preserves the exact reference radius by appending a single scalar to the representation vector.

  2. A theoretical bridge from anchor budget to recovery error, supported by controlled simulation Theorem 2 establishes a non-asymptotic recovery bound for MALOT under a finite anchor budget. Combined with the 2-approximation guarantee of FPS for KK-center clustering, this provides an end-to-end theoretical guarantee that connects the anchor budget KK and covering radius to the metric recovery error. Corollary 1 also accounts for compressed anchors, inexact feasible plans, and observation errors. In the rotating multi-cluster experiments, MALOT outperforms both single-reference LOT and oracle Concat LOT, validating the efficacy of combining multi-anchor coverage with endpoint-dependent weighting.

Critical weaknesses

The advantages of MALOT over other multi-reference OT methods are not fully established in the current manuscript.

  1. The finite-anchor recovery guarantee (Thm 2) does not capture MALOT-specific advantages. The core foundation of Theorem 2 is Lemma 2, which does not rely on any MALOT-specific properties. Actually, the authors note (Appendix A.1) that analogous guarantees can be achieved by any radially exact representation when at least one endpoint is close to an anchor. Specifically, using solely the exact reference costs assumed in Theorem 2, the following pivot estimator can be constructed: LBij:=max⁡k∣rki−rkj∣,UBij:=min⁡k(rki+rkj),Mij:=LBij+UBij2  (i≠j),Mii:=0.LB_{ij}:=\max_k|r_{ki}-r_{kj}|, \quad UB_{ij}:=\min_k(r_{ki}+r_{kj}), \quad M_{ij}:=\frac{LB_{ij}+UB_{ij}}{2} \ \ (i\neq j),\quad M_{ii}:=0. This pivot midpoint already enjoys the same one-nearby-endpoint coverage property as Theorem 2, with the sharper bound ∣Mij−Wij∣≤min⁡(hi,hj)|M_{ij}-W_{ij}|\le\min(h_i,h_j) versus 2(hi2+hj2)/2+τlog⁡K2\sqrt{(h_i^2+h_j^2)/2+\tau\log K} in Eq. (13). The distinctive advantage of MALOT is therefore not endpoint coverage itself, but its use of directional LOT information beyond the anchor radii. Although the authors mention the benefit (Appendix A.1) of directional displacement information, the current analysis does not yet show—either theoretically or empirically—how this directional information improves upon a simpler radial-only pivot estimator. Furthermore, the connection between the Riemannian construction and the finite-budget MALOT estimator remains unclear. Theorem 1 establishes recovery for the intrinsic metric, and Proposition 2 shows infinitesimal consistency of the two-point rule, whereas the finite-pair guarantee in Theorem 2 relies only on radial exactness. Consequently, there seems to be a gap between the Riemannian motivation in Section 3.1 and the finite-budget guarantee in Theorem 2: it remains unclear how the Riemannian structure itself benefits the discrete estimator beyond generic anchor coverage.

  2. The current empirical evaluation is insufficient to demonstrate MALOT-specific advantages. Under the same reference anchor bank and budget KK, additional comparisons against the following baselines are needed to isolate the contribution of endpoint-dependent weights on squared LOT separations:

  • The pivot estimator defined above;
  • A nearest-anchor LOT baseline that uses only each endpoint’s nearest anchor, i.e., the τ→0\tau\to 0 limit of MALOT;
  • A Landmark-MDS-type embedding based only on the same anchor distances. Additionally, the performance of oracle Concat LOT with the same bank and KK as MALOT should be evaluated in the real-data analysis. This would disentangle the effects of reference quality, multiple anchors, and endpoint-dependent weighting. In particular, demonstrating the superior performance of MALOT in the small-KK, coarse-coverage regime is essential to validate the claim that directional displacement information “can improve approximation before the covering radius is small” (Appendix A.1). A wall-clock runtime comparison would also be beneficial. While Appendix C.3 provides asymptotic computational complexity of LOT, MALOT, Sinkhorn, and Sliced Wasserstein, MALOT requires KK times as many reference OT solves as LOT, and full MDS incurs an O(n3)O(n^3) cost. An empirical wall-clock tradeoff analysis would more convincingly demonstrate the practical utility of MALOT.
Questions for authors
  1. I would be glad to hear the authors’ own view of where MALOT makes a genuine difference compared to other multi-reference OT methods and simpler reference-based estimators. Theoretical or empirical evidence supporting such an advantage would raise my assessment.

  2. What is meant by “smooth exact LOT map” in Section 3.1? I read it as the Brenier map of Section 2.2 (consistent with Hk=L2(ρk;Rd)H_k=L^2(\rho_k;\mathbb{R}^d), with "exact Brenier-map LOT" in Theorem 2, and with Appendix B.1); please correct me if a weaker notion is intended. Under this reading, the gluing argument gives ∣ϕρ(μ)−ϕρ(ν)∣L2(ρ)≥W2(μ,ν)\\|\phi_\rho(\mu)-\phi_\rho(\nu)\\|_{L^2(\rho)}\ge W_2(\mu,\nu) globally, hence gρk⪰gWg_{\rho_k}\succeq g_W for every anchor, so Theorem 1 appears to hold one-sidedly for all δ≥0\delta \ge 0 without the condition δ<1\delta<1.

I would appreciate clarification on whether this reading is correct, and how the authors view the distinct technical contribution of Theorem 1 beyond combining Proposition 1 with the Gibbs inequality.

7 comments

We are grateful for your careful analysis of the finite-anchor guarantee. Your pivot construction and question about the geometric argument have helped us clarify what anchor coverage, transport directions, and smooth geometry each contribute to finite-pair recovery.

Weakness 1. Finite-anchor recovery and the role of directions

Pivot and coverage. In the revised manuscript, we have added the pivot comparison in Appendix A.6. As your argument shows, the pivot midpoint has a sharper coverage bound. Let rki=W2(ρk,μi)r_{ki}=W_2(\rho_k,\mu_i) and hi=min⁡krkih_i=\min_k r_{ki}. The triangle and reverse triangle inequalities put Wij=W2(μi,μj)W_{ij}=W_2(\mu_i,\mu_j) in every interval [∣rki−rkj∣,rki+rkj][|r_{ki}-r_{kj}|,r_{ki}+r_{kj}]. Their intersection has endpoints ℓij=max⁡k∣rki−rkj∣,Uijrad=min⁡k(rki+rkj).\ell_{ij}=\max_k|r_{ki}-r_{kj}|,\qquad U_{ij}^{\mathrm{rad}}=\min_k(r_{ki}+r_{kj}).

The midpoint Mij=(ℓij+Uijrad)/2M_{ij}=(\ell_{ij}+U_{ij}^{\mathrm{rad}})/2 has error at most half the intersection width. Each anchor interval has width 2min⁡(rki,rkj)2\min(r_{ki},r_{kj}), so ∣Mij−Wij∣≤min⁡(hi,hj)|M_{ij}-W_{ij}|\le\min(h_i,h_j). The original MALOT proof uses radial exactness and triangle inequalities. We have retained its uniform recovery guarantee in Lemma 3 (Appendix A.2). It applies as coverage and τlog⁡K\tau\log K vanish.

What directions add. We have revised Theorem 2 to use the directional information in the reference plans. With optimal reference plans, drawing independently from the two target laws conditional on the same anchor draw yields a target-pair coupling with squared cost

Gk2=rki2+rkj2−2⟨ϕki,ϕkj⟩,G_k^2=r_{ki}^2+r_{kj}^2-2\langle\phi_{ki},\phi_{kj}\rangle,

where ϕki\phi_{ki} and ϕkj\phi_{kj} are the barycentric displacement blocks. Feasibility and the triangle inequality give Wij≤Gk≤rki+rkjW_{ij}\le G_k\le r_{ki}+r_{kj}. Theorem 2 replaces the radius-only upper endpoint by min⁡kGk\min_k G_k. The larger of the two absolute differences between the MALOT estimate and the refined interval endpoints bounds its error. This calculation uses the stored reference plans and applies to both block variants. For completed blocks the resulting bound is at most the original pair-specific coverage bound. For uncompleted blocks the comparison includes the difference between the two estimates (§4.1, Eq. (13), and Appendix A.6).

In the centered eight-atom example, the targets vary in scale along a circle of radius rr and share identical anchor-radius vectors. True distances for distinct pairs range over (0,2r](0,2r], so any rule using only the two radius vectors has worst-case error at least rr. Each LOT separation equals the true distance, and both MALOT and uniform Concat LOT recover every pair at this positive covering radius. Appendix A.6 also gives a translation example where the pivot returns rr for all distinct pairs and attains that lower bound.

At coarse coverage. We have added a pairwise comparison in Appendix D.3. On the K=4K=4 crescents, using directions tightens the error bound for uncompleted MALOT on 1,015 of 2,775 pairs. Against exact W2W_2, MALOT has strictly smaller absolute error than the pivot on 1,724 pairs. Our response to Weakness 2 gives the same-bank stress and maximum-error comparisons.

From smooth geometry to finite pairs. For a distinct target approaching an off-anchor target, the pivot midpoint tends to the latter's positive nearest-anchor distance as the true distance vanishes (Appendix A.6). Proposition 2 gives a linear leading-order distance for the smooth exact-map MALOT rule along a curve with nonzero initial velocity. We have made the connection to finite-pair recovery explicit in Appendix B.6. For a pair with W=W2(μ,ν)>0W=W_2(\mu,\nu)>0 under Theorem 1, we integrate the referencewise speed bounds along a piecewise smooth minimizing Wasserstein geodesic contained in the analysis region and apply the endpoint weights. With covering radius hh and ζ=h2+τlog⁡K\zeta=h^2+\tau\log K, Eq. (25) gives

\le \sqrt{1+A\bigl(\zeta+W\sqrt{\zeta}+W^2/3\bigr)}.$$ For uniformly bounded geometric constant $A$, separation $W=O(h)$ and $\tau\log K=O(h^2)$ give $O(h^2)$ relative error and $O(h^3)$ absolute error. This rate concerns smooth exact maps. Lemma 3 and Theorem 2 give the separate coverage and computable pair-specific guarantees for finite reference plans.

We appreciate your thoughtful suggestions for matched baselines and a wall-clock comparison. The matched controls have helped us assess when directions and positive temperature improve distance recovery. The timing breakdown has made the costs of each stage clearer.

Weakness 2. Matched baselines under coarse coverage

Matched comparisons. In the revised manuscript, we have added comparisons with the pivot midpoint, Landmark MDS, and Hard endpoint in Sections 5.1–5.2 and Appendix D, alongside medoid LOT, Concat LOT, and Min-LOT. The multi-reference estimators share anchors and exact target-to-anchor costs within each bank and budget. The crescents use nested FPS banks at K=4,8,16,24K=4,8,16,24. The rotating multi-cluster simulation uses two K=12K=12 banks, one FPS and one oracle, with eight-atom anchors. The LOT methods use uncompleted barycentric blocks. Landmark MDS fits on exact inter-anchor distances and projects targets from their exact distances to the anchors. We measure raw stress against exact W2W_2 on the recovered distances before any common display embedding.

Hard endpoint implements the τ↓0\tau\downarrow0 limit of MALOT. Each endpoint assigns uniform weight to its nearest anchors, including ties. The rule averages the two endpoint weight vectors and compares both targets in each selected chart. Appendix A.7 gives a uniform 2h2h error bound for Hard endpoint with optimal plans and radially exact blocks as coverage improves. It also illustrates a discontinuity at a nearest-anchor boundary with the bank held fixed.

Distance recovery. At crescent K=4K=4, MALOT reduces raw stress from 0.1730 for the pivot to 0.0619 and maximum absolute error from 0.4685 to 0.1419. Among the 2,485 pairs with neither endpoint an anchor, raw stress is 0.1835 for pivot and 0.0644 for MALOT (Appendix D.2). At the same budget, holding the endpoint weights fixed and replacing the LOT separations with scalar radius differences raises raw stress to 0.5167 (Appendix D.3, Table 6).

MALOT and Hard endpoint give lower full-pair raw stress than pivot and Landmark MDS at every reported budget and on both multi-cluster banks. MALOT also gives lower stress than Concat LOT and Min-LOT in these full-pair comparisons. Hard endpoint gives lower raw stress than MALOT at crescent K=4,8K=4,8. MALOT gives lower raw stress at K=16,24K=16,24 and on both multi-cluster banks, and lower maximum absolute error than Hard endpoint at crescent K=4,16,24K=4,16,24 (Tables 3–5). On the multi-cluster oracle bank, Landmark MDS has lower maximum absolute error than MALOT (0.0930 versus 0.1302), alongside its higher raw stress (Appendix D.1).

Wall-clock costs. We have added a staged timing comparison in Table 2 and Appendix C.3. With anchor identities fixed, it reports reference setup, distance-matrix assembly, and optional full classical MDS separately. For the weighted crescents, we measured reference transport and exact W2W_2 in separate fresh passes and report medians of five postprocessing runs. For the multi-cluster simulation, we report medians of five full runs after warm-up.

On the crescents, MALOT takes 28.19 s at K=4K=4 and 122.74 s at K=24K=24 for setup plus distance assembly, compared with 14.49 s for medoid LOT and 158.90 s for exact W2W_2. At K=24K=24, assembly takes 39.01 ms for both MALOT and Hard endpoint, 62.48 ms for Concat LOT, and under 1 ms for pivot and Landmark MDS. On the multi-cluster FPS and oracle banks, MALOT takes 180.91 and 183.33 ms for setup plus assembly, compared with 14.85 ms for medoid LOT and 1,611.03 ms for exact W2W_2. Reference setup dominates MALOT's total. The matched anchor-distance controls are charged the common reference setup, including LOT block construction. Landmark fitting is included in that baseline's assembly stage. The optional full classical MDS step takes 0.25–0.76 ms for the 75 crescents and 0.65–1.90 ms for the 120 multi-cluster targets. These times describe the tested sizes. Appendix C.3 gives the scaling of full MDS.

Your recommendation to match the anchor banks and KK is also important for interpreting the real-data comparison. We are evaluating oracle Concat LOT on real datasets under these matched conditions and will share the results promptly once available.

We appreciate your careful questions about where MALOT offers a practical advantage and the assumptions underlying its geometric analysis. They have helped us explain the role of endpoint weighting and state the smooth exact-map guarantee more precisely.

Question 1. Conditions for improved distance recovery

Compared with radius-only estimators. In the revised manuscript, we have added examples in Appendix A.6 showing that LOT directions can recover distances at positive covering radius when distinct pairs have identical anchor-radius data. Both MALOT and Concat LOT are exact in those constructions. Table 6 separately shows the gain from retaining barycentric displacement directions at fixed weights on the crescents.

Compared with fixed LOT representations. Endpoint weighting reduces the influence of far anchors. Theorem 1 controls the weighted mean squared radius by h2+τlog⁡Kh^2+\tau\log K. Uniform Concat LOT instead uses the unweighted mean squared radius in the corresponding geometric bound (Appendix B.6). Appendix A.5 gives a strict separation on four rotating two-atom targets. Every fixed Hilbert representation, including single-reference LOT and fixed Concat LOT, incurs at least 13.29% maximum relative error. With those targets as anchors, radially exact blocks, and sufficiently small positive τ\tau, Lemma 3 gives a smaller bound on MALOT's pairwise dissimilarities before MDS.

Compared with Hard endpoint and Min-LOT. At a nearest-anchor boundary, Appendix A.7 gives one-sided Hard endpoint limits of 11 and 2\sqrt{2} while the positive-temperature estimate is continuous in the example. In a separate three-anchor example, the true distance is 8. Hard endpoint returns 10, Concat LOT returns 9.381, and MALOT returns 9.345 by retaining weight on an accurate nearby chart. Min-LOT is exact on that pair. With deterministic maps, common-source coupling makes every same-anchor LOT separation an upper bound on W2W_2, so the minimum gives pointwise error no larger than a same-bank weighted RMS. With split plans, averaging displacements within an anchor atom can shrink a block separation below W2W_2. One such block can determine the minimum. MALOT weights anchor proximity and can limit the influence of that block.

In the K=12K=12 rotating multi-cluster experiment, MALOT gives raw stress 0.0508 on the FPS bank. Single-reference LOT, Concat LOT, Min-LOT, and Hard endpoint give 0.2347, 0.1844, 0.1094, and 0.0610, respectively. MALOT also gives the lowest stress on the oracle bank (0.0595, Table 3). Min-LOT underestimates about 97.5% of the 7,140 distances on each bank. Against exact W2W_2, MALOT has strictly smaller absolute error than Min-LOT on 6,498 FPS pairs and 6,791 oracle pairs (Appendix D.1). Comparing each baseline on the same 7,140 pairs per bank with tolerance 10−1010^{-10}, MALOT's strict win rates against single-reference LOT, Concat LOT, Min-LOT, and Hard endpoint are 78.32%, 96.53%, 91.01%, and 68.96% on the FPS bank, and 74.97%, 93.73%, 95.11%, and 70.52% on the oracle bank. The crescent results in Table 4 show how the comparison with Hard endpoint varies by budget.

Question 2. Smooth exact LOT maps and Theorem 1

Your observation about the common-source coupling helped us state the lower speed bound independently of the additional assumptions for the upper bound. We have clarified in Section 3.1 that the exact Brenier displacement ϕρ(μ)=Tρ→μ−id\phi_{\rho}(\mu)=T_{\rho\to\mu}-\mathrm{id} from an absolutely continuous reference ρ\rho is assumed to vary smoothly with the target in L2(ρ;Rd)L^2(\rho;\mathbb R^d). Local agreement with the Riemannian logarithm is a separate assumption for the upper bound (Appendix B.1). Using both maps on a common draw from ρk\rho_k gives a coupling and hence

W22(μ,ν)≤∣ϕρk(μ)−ϕρk(ν)∣L2(ρk)2.W_2^2(\mu,\nu)\le\\|\phi_{\rho_k}(\mu)-\phi_{\rho_k}(\nu)\\|_{L^2(\rho_k)}^2.

Taking the smooth-curve limit and using equality of model and ambient Wasserstein speeds yields gρk⪰gWg_{\rho_k}\succeq g_W. Averaging gives gR⪰gWg_{\mathcal R}\succeq g_W and positive definiteness for every τ>0\tau>0. We have revised Theorem 1 to make this lower bound explicit. The distortion bound holds for all finite δ\delta furnished by the theorem, including δ≥1\delta\ge1.

For the upper bound, Proposition 1 controls a chart near its anchor, and the far-anchor derivative bound controls charts that still receive positive weight. Gibbs localization bounds the mean squared radius at each point, even as the nearby anchors change along a path. Comparing lengths along the same paths in UU gives the intrinsic distance ratio in Eq. (9), bounded above by 1+A(h2+τlog⁡K)\sqrt{1+A(h^2+\tau\log K)}. Its denominator equals ambient W2W_2 when a piecewise smooth Wasserstein minimizing geodesic lies in UU (Section 3.1 and Appendix B.3).

Reviewer_6D2gOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Linearized optimal transport (LOT) is a method for embedding a set of distributions into a linear space by picking an anchor point, computing the Monge map that takes this anchor to the distribution and mapping each distribution to its Monge map. This linearization gives an embedding where Euclidean distances approximate W_2 distances in the original space.

This manuscript proposes an extension of the LOT approach that uses multiple anchor points that are combined via local kernel weighting. It describes the method, provides theory that shows that the error is controlled by the maximal distance to the nearest anchor, and presents several simulations on synthetic and real data sets that demonstrate improved performance relative to standard LOT.

Critical strengths

The method is properly motivated and easy to understand. It is tested on both synthetic and real data sets and demonstrates good performance.

The paper includes approximation bounds with formal proofs.

Overall, the paper is well-structured and well-written. There is a slight LLM-scent to parts of the text here and there, which I personally find unpleasant, but it is clear that the authors have made extensive edits.

Critical weaknesses
  • The method presented in the paper is closely related to Khurana et al., (2023) - what the authors call "Concat LOT". This paper is cited but I think it deserves a better exposition and contrast against the proposed method. In particular, if I understand correctly, if one takes \tau equal to infinity then the methods are essentially the same and so Algorithm 1 can be viewed as an extension. On the real data sets, the accuracy of MALOT is not much better than Concat-LOT, so it seems like the unequal weighting does not help too much on these examples. I would call the results comparable.

  • The writing could be improved to make the theoretical results clearer to an ML audience by including introductory sentences and intuition. The paper would have more impact if it did not assume that the reader has a solid background in Riemannian geometry. Additional points regarding writing: the authors need to make sure that all metrics are described in the main text. For example the "accuracy" metric of Table 1 only appears in the supplementary. Theorem 1 should list all the conditions clearly and not just state "Under the preceding conditions"

  • Algorithm 1 uses farthest-point selection for the next anchor. This is not really motivated and we know from K-means clustering that this is not optimal. There is no comparison against variants such as random selection or K-means++ style selection..

Questions for authors
  • Can you also test a different anchor selection method? e.g. try random selection and/or K-means++.

  • How did you select the data sets for the real experiments? Please explain you selection process.

  • I am confused by the fact that the theoretical results do not require compactness or a bound on the reach of the manifold. It seems like the global C_0 and B are somehow accounting for these. The text only says that C_0 is a universal constant but I would appreciate a clearer discussion relating the universal constants to standard properties of the manifold like reach, radius of curvature, etc. Section B.2 sets C_0=4 in the proof of proposition 1 which makes me even more confused.

2 comments

We appreciate your careful reading, which helped us clarify the link to Concat LOT, the geometric bounds, and anchor selection.

Weakness 1. Relation to Concat LOT and classification accuracy

Your observation helped us clarify the Concat LOT limit. In the revised manuscript, we have expanded Section 1 to explain how Khurana et al. (2023, Section 5) use a fixed product of LOT maps and an unnormalized ℓ2\ell_2 product distance to study class separation for sheared distributions. Our Concat LOT baseline scales the concatenated barycentric LOT blocks by 1/K1/\sqrt{K}, for KK anchors. In MALOT, the two endpoints determine each anchor's weight.

We have made the limit for Algorithm 1's aggregation step explicit in Section 3.2. With fixed anchors and reference plans, the endpoint weights and their pairwise averages in Eq. (11) tend to 1/K1/K as τ→∞\tau\to\infty. Without radial completion, this recovers the Concat LOT distance DijConcatD_{ij}^{\mathrm{Concat}} on the same barycentric LOT blocks. With radial completion, the limit is the normalized product distance on the completed blocks. Let Dijrad(τ)D_{ij}^{\mathrm{rad}}(\tau) denote MALOT with radial completion and VkiV_{ki} the residual variance of target ii at anchor kk. The squared limit is

= \bigl(D_{ij}^{\mathrm{Concat}}\bigr)^2 + \frac{1}{K}\sum_{k=1}^{K}\bigl(\sqrt{V_{ki}}-\sqrt{V_{kj}}\bigr)^2.$$ The completed limit agrees with Concat LOT for a given pair exactly when $V_{ki}=V_{kj}$ for every anchor, as happens when both reference plans are deterministic at every anchor. We agree that the classification accuracies in Table 1 are comparable. For MALOT without radial completion and Concat LOT, respectively, they are 87.48% and 87.16% on Caulobacter, 81.98% and 82.07% on ETH-80, and 93.99% and 93.23% on Fingering. These values reflect complete downstream procedures, including their reference choices and representations. For example, on one ETH-80 split Concat LOT uses four anchors and MALOT uses sixteen (Appendix E.1). Additional real-data comparisons using matched anchor banks and budgets are in progress, and we will provide an update promptly once the results are available. ## Weakness 2. Theory exposition and evaluation metrics Your guidance helped us make the geometry more accessible to machine learning readers. We have revised Section 2.2 to interpret the LOT quadratic form as squared speed in reference coordinates and to link LOT distortion to how close a minimizing Wasserstein path stays to the reference in a compatible chart. Theorem 1 (Section 3.1) states the smooth Wasserstein setting and the assumption of smooth exact LOT maps from absolutely continuous anchors. A common-source coupling gives the lower speed bound and positive definiteness of the blend for every $\tau>0$. The uniform upper bound additionally uses compatible normal charts, a curvature bound, a derivative bound for distant anchors and finite coverage. We have explained after Theorem 1 how the nearest-anchor cost and the $\tau\log K$ term bound the weighted squared radius. The speed comparison then bounds the lengths of the same paths, even when the nearest anchor changes along a path. The intrinsic Wasserstein distance equals ambient $W_2$ when a minimizing geodesic stays in the analysis region $U$. Section 4 introduces the finite-support bound. For a radially exact anchor comparison, both the estimated and true distances lie between the absolute difference and the sum of the two anchor radii. The width of that interval is twice the smaller radius, so one nearby endpoint suffices to control the error for that anchor. We have defined the reported metrics in Section 5. Raw stress is the square root of the ratio of summed squared distance errors to summed squared exact distances, with no fitted rescaling. Spearman correlation compares the ranks of the distances, and maximum absolute error is taken over the evaluated pairs. ACC is the percentage of correctly classified test observations in a split. Table 1 gives equal-weight averages of split-level raw stress and ACC. ## Weakness 3. Rationale for farthest-point sampling We appreciate your question about the objective in Algorithm 1. In Section 3.3, we have explained that farthest-point sampling (FPS) targets the largest distance to the nearest anchor, the discrete $K$-center objective that enters the uniform recovery bound in Lemma 3 (Appendix A.2). The standard $k$-means objective sums squared assignment distances. With exact $W_2$ distances and uncompressed anchors, FPS attains a covering radius at most twice the optimal discrete $K$-center radius on the candidate pool. Appendix A.3 gives the corresponding bound with anchor compression. Appendix F compares FPS with uniform sampling and Wasserstein $D^2$ seeding (Question 1).

We appreciate your questions, which helped us strengthen the anchor-selection comparison and clarify the dataset choices and geometric assumptions.

Question 1. Alternative anchor selectors

In the revised manuscript, we have added an ablation (Appendix F) comparing farthest-point sampling (FPS), uniform sampling without replacement, and Wasserstein D2D^2 seeding. The last adapts kk-means++ initialization by sampling each next anchor in proportion to its squared Wasserstein distance to the nearest selected anchor. We fix the first anchor, budget, and temperature within each setting. We report deterministic FPS results and means and sample standard deviations over 20 seeds per randomized rule.

For the 75 rotated crescent laws, we evaluate all 2,775 unordered pairs with barycentric LOT blocks without radial completion. FPS gives a smaller covering radius than either randomized mean at every budget (Table 8). The mean radius is smaller for D2D^2 seeding than for uniform sampling at K=8,16,24K=8,16,24. At K=4K=4, the uniform and D2D^2 means are 0.6382 and 0.6515, respectively. FPS gives the lowest raw stress and maximum absolute error at K=4,8K=4,8. The lowest means are 0.01609 and 0.05883 for D2D^2 at K=16K=16, and 0.01363 and 0.05375 for uniform sampling at K=24K=24.

Table 9 reports the existing Fingering ablation. In three run-level splits, we select from training frames, compress anchors to 128 atoms, and use radially completed LOT blocks. Distances are evaluated on the same 64 test pairs per split, and classification uses all test frames. FPS gives a smaller training covering radius than either randomized mean in each split, and D2D^2 seeding gives a smaller mean radius than uniform sampling. Against the randomized means, FPS gives the lowest maximum absolute error in four of seven settings. Fingering's split-averaged raw stress ranges from 0.00463 to 0.00534 across selectors, with uniform sampling lowest. Mean test ACC ranges from 93.84% to 93.99%.

Question 2. Dataset selection

Your question highlighted the need to explain dataset selection. We have explained our choices in Section 5.3 and documented eligibility in Appendices E.1 and E.3. We selected the three Table 1 cases from a broader exploratory evaluation of public datasets. Each supports recovery against exact W2W_2 on the stated representation and a classification task with group-level holdouts. Caulobacter links membrane fluorescence to recorded cell division, with splits by acquisition date. We excluded one candidate date because imaging and division records lacked reliable cell-ID matches. ETH-80 complements the rotating-shape simulation with 41 contour views of each of 80 objects. All views of a physical object remain in the same partition. Fingering links occupancy images to independently measured volumetric breakthrough, with splits by experimental run. Eligibility requires foreground-fraction checks and excludes frames in the published breakthrough measurement-error band.

Question 3. Geometric assumptions and constants

We appreciate your question about the geometric assumptions. We have clarified the uniform chart and derivative conditions in Theorem 1 and discussed reach in Appendix B.1. Around each anchor, the compatible chart must contain the exponential image of the closed tangent ball of radius r0r_0. It must also contain every target in UU whose ambient W2W_2 distance to that anchor is at most r0r_0. This keeps ambient-close targets on the anchor's local branch. Classical Euclidean reach concerns uniqueness of nearest-point projection.

The admissible C0=4C_0=4 in Proposition 1 comes from the Jacobi-field estimate on the range κr2≤1/4\kappa r^2\le 1/4 (Appendix B.2). For κ>0\kappa>0, curvature sets the length scale κ−1/2\kappa^{-1/2}, and the condition κr02≤1/4\kappa r_0^2\le 1/4 requires r0≤1/(2κ)r_0\le 1/(2\sqrt{\kappa}). The same radius must satisfy the chart conditions above. For target–anchor distances above r0r_0, 1+B1+B bounds the squared operator norm of the LOT differential from the Wasserstein tangent norm to the reference L2L^2 norm. On compact regular regions, joint smoothness under isometric identifications suffices to bound this norm. Appendix B.4 illustrates the conditions for Gaussian measures with covariance eigenvalues bounded above and away from zero.

In Appendix B.3, we have shown how the near- and far-anchor estimates enter A=max⁡4κ,(1+B)/r02A=\max\\{4\kappa,(1+B)/r_0^2\\}. The far-anchor term attains the maximum when κr02≤1/4\kappa r_0^2\le 1/4. Curvature still constrains r0r_0 and controls the local distortion estimate. The finite-budget bound uses finite coverage and uniform regularity. For asymptotic recovery, r0,κ,Br_0,\kappa,B are common across anchor banks, hK→0h_K\to0, and τKlog⁡K→0\tau_K\log K\to0. Total boundedness in W2W_2 supplies arbitrarily fine finite covers, while the chart and derivative bounds must hold along the chosen sequence.

Comments on the paper

1 comment
Reviewer_6D2gOpenReview ↗

Score upgrade

The authors have done an excellent job by posting an extensive revision and answering my questions. All of my concerns are addressed and since no critical issues were raised by the other referees I am upgrading my rating to Accept.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.