Simultaneous Confidence Bands for Sliced Wasserstein Profiles When Dimension Exceeds Sample Size

When two high-dimensional datasets differ, a practitioner usually wants more than a yes/no test. Two tumor subtypes measured on 207 genes are a good example: the practitioner wants to know along which few directions the distributions differ and by how much, with error bars that hold simultaneously. Optimal transport gives a geometric way to measure such differences. The Wasserstein distance is the minimum cost of moving one distribution's mass onto the other's. Its one-dimensional version is cheap: project both samples onto a line, sort them, and compare. But every existing limit theorem for transport statistics fixes the dimension $d$ while the sample size $n$ grows. High-dimensional central limit theory, the work showing that maxima of many correlated sums stay approximately Gaussian even when the number of coordinates grows exponentially in $n$, has never been turned around to target transport statistics.

This note proposes that link. We build simultaneous confidence bands for the 1-Wasserstein distance between the two samples' projections onto every direction in a large, sparse dictionary (for instance all coordinates and all coordinate pairs, so $K \approx d^2/2$ directions). The bands are calibrated by a multiplier bootstrap and come with a Gaussian approximation theorem in which dimension enters only through $\log K$. Two things are new. First, the band is valid whether or not a given direction actually differs. This matters because 1D transport statistics are non-Gaussian along directions where the two projections coincide, and existing delta-method arguments break exactly there. Second, we show that restricting to sparse directions is necessary, not a convenience: the maximum over all directions on the sphere stops concentrating once $d$ is comparable to $n$.

Core Claim

Fix a dictionary $U$ of $K$ unit directions chosen without looking at the data. For each $u \in U$, let $F_{n,u}$ and $G_{m,u}$ be the empirical distribution functions of $u^\top X_i$ ($n$ samples from $P$) and $u^\top Y_j$ ($m$ samples from $Q$), and let $F_u, G_u$ be their population versions. The projected distance is $\hat W_u = \int |F_{n,u}(t) - G_{m,u}(t)|\,dt$. By the reverse triangle inequality, for every $u$ simultaneously,

$$ |\hat W_u - W_u| \le \int \big| (F_{n,u}-F_u)(t) - (G_{m,u}-G_u)(t) \big|\, dt , $$

where $W_u$ is the population projected distance. The right side does not depend on whether $W_u$ is zero, so a quantile of its maximum over $u$ gives an honest simultaneous band. We estimate that quantile with the multiplier bootstrap

$$ M^* = \max_{u \in U} \int \Big| \tfrac{1}{n}\textstyle\sum_i \xi_i\big(1\{u^\top X_i \le t\} - F_{n,u}(t)\big) - \tfrac{1}{m}\sum_j \eta_j\big(1\{u^\top Y_j \le t\} - G_{m,u}(t)\big) \Big|\, dt , $$

where $\xi_i, \eta_j$ are independent standard normal multipliers drawn given the data. With $c^*_\alpha$ the conditional $(1-\alpha)$ quantile of $M^*$, the band $\hat W_u \pm c^*_\alpha$ is claimed to cover all $W_u$ at level $1-\alpha$, up to an error of order $(\log K)^a / n^b$ for constants $a, b>0$ that the theory will pin down. The band width should scale like $\sqrt{\log K / n}$, because a maximum of $K$ approximately Gaussian quantities grows like $\sqrt{\log K}$. For an $s$-sparse dictionary $\log K \approx s \log d$, so $d$ can grow polynomially or faster in $n$. Inverting the band at zero also gives a global two-sample test.

Where This Sits

The fixed-dimension transport limit theory supplies the object we study. Central limit theorems for empirical transport cost (del Barrio and Loubes; del Barrio, González-Sanz and Loubes) linearize the cost around stable dual potentials, but only when $P \ne Q$ and $d$ is fixed. Xi and Niles-Weed prove that the sliced Wasserstein process indexed by the whole sphere converges to a Gaussian process and validate the empirical bootstrap. That proof relies on covering numbers of the sphere that grow like $(1/\varepsilon)^d$, so it says nothing once $d = d_n \to \infty$.

We claim this is not just a gap in the proof but a real obstruction. The supremum over all directions is a supremum over half-space indicators, a class whose complexity grows linearly in $d$. Standard empirical-process bounds then suggest the full-sphere max-sliced statistic is of order $\sqrt{d/n}$ even when $P = Q$. Once $d \ge n+m-1$, generic pooled points can be split by a hyperplane, so the statistic stays bounded away from zero under the null. Restricting to sparse directions is the transport analogue of the "$s$-sparsely convex sets" that Chernozhukov, Chetverikov and Kato use to keep Gaussian approximation valid in high dimensions. That paper, together with the 2013 multiplier bootstrap paper, supplies our method. The anti-concentration bounds for Gaussian maxima supply the step that turns closeness of smoothed functionals into closeness of quantiles.

Two pieces of recent work motivate the proposal by leaving the gap open. Buzun, Shvetsov and Dylov connect the areas in the opposite direction: they use transport as a proof tool to bound Gaussian approximation of sums. They do not do inference on a transport quantity. Tran and Schreuder give a minimax-optimal permutation test built on sliced Wasserstein. However, that work gives no limiting distribution, does not address sparse alternatives, and, being a permutation test, produces no confidence statements under the alternative.

The closest work is Neural Wasserstein Two-Sample Tests. It also aggregates many projected Wasserstein statistics through a maximum and calibrates them with a Gaussian limit. But it learns its projections on half the data, conditions on that half, and only tests a global null. We use the full sample and a fixed dictionary, so no conditioning is needed. We also target simultaneous intervals for every direction, not a single p-value.

Two constraints shape the design. The entropic limit theorems of Goldfeld and coauthors show that transport statistics change scaling under the null, with the Sinkhorn divergence converging at rate $1/n$ rather than $1/\sqrt{n}$. This is why a per-direction delta-method band will misbehave at directions with $W_u = 0$, and it is why we bound the error with the triangle inequality instead. The heavy-tail results in the 2017 high-dimensional CLT show rates slowing to $n^{-1/6}$ with only polynomial moments. We therefore expect, and will test, a cost under heavy tails.

Research Program

The theoretical work has three parts.

The first is a Gaussian approximation theorem for the maximum $M = \max_{u} \int |Z_u(t)|\,dt$ of the centered empirical difference process. This is a supremum over the function class $\{x \mapsto f(u^\top x): u \in U, f \text{ 1-Lipschitz on a bounded interval}\}$. Its covering-number entropy splits into $\log K$ plus a one-dimensional term that does not depend on $d$. We will extend the smoothing-plus-interpolation argument of Chernozhukov, Chetverikov and Kato from finite maxima to this class, keeping explicit the dependence on $\log K$ and on tails of the projections. We will also prove that the multiplier bootstrap is consistent at the same rate.

The second part is the matching negative result: a lower bound showing that the full-sphere maximum has null magnitude of order $\sqrt{d/n}$. This makes the necessity of sparsity precise.

The third part is a refinement. For directions where the honest band already excludes zero, the projected distributions differ almost everywhere and the delta method applies. A second-stage band can then be sharpened using the linearization of the projected cost along those directions, following the dual-potential arguments of del Barrio and coauthors. Selection effects from the first stage will be controlled by a union bound across stages.

On the computational side, projections and sorts are computed once per direction. Each bootstrap draw is then a cumulative sum of multipliers in sorted order, so it costs $O(n+m)$ per direction. On a GPU this makes coordinate-pair dictionaries with $K$ near $10^6$ feasible at $d$ in the low thousands.

Baselines

The naive reference point is per-direction pointwise intervals with no multiplicity correction: the bootstrap quantile computed separately for each $u$. It will look sharp and should undercover badly as $K$ grows. A second naive reference is a Bonferroni band built from one-dimensional Dvoretzky–Kiefer–Wolfowitz inequalities for each projection. It is valid but ignores correlation across directions. We beat it if our bands are uniformly narrower at equal coverage. We expect the gap to be largest for pair dictionaries, where neighboring directions are strongly correlated.

The strongest existing inferential method for this object is the empirical bootstrap validated by Xi and Niles-Weed, applied to the same sparse dictionary and the same maximum statistic. It is a real competitor, because Efron-type and multiplier bootstraps often behave similarly. We run it with the same dictionary, the same 1000 resamples and the same hardware. Here we claim theoretical cover in the growing-$d$ regime and lower per-draw cost (no re-sorting). We do not necessarily claim better coverage, and we will report honestly if the two methods match.

For global testing power, the strongest competitors are the permutation sliced Wasserstein test of Tran and Schreuder, run with $L \asymp n$ random directions as their theory requires, and the Neural Wasserstein test. We also run MMDAgg (Schrab et al.) and the AutoML test (Kübler et al.) with their default tuning budgets. On TCGA, the Neural Wasserstein paper reports $p < 0.001$ for LGG versus GBM. For Models A–D, Ball versus Sphere and the MNIST shifts, the file gives only curves, so there is no reported numeric power result found on Lacuna for any of these baselines. We will rerun all of them ourselves under identical splits and level $\alpha = 0.05$. We count as "beating" Neural Wasserstein matching its power on sparse alternatives while using the full sample.

Two ablations isolate the key ingredients. Replacing the triangle-inequality band with per-direction delta-method Gaussian intervals tests whether robustness to degeneracy matters. Replacing the sparse dictionary with an approximate full-sphere maximum (projected gradient ascent from many random starts) tests whether sparsity matters.

Why not just…?

Why not just use the permutation sliced Wasserstein test? For a yes/no global test it is a genuine threat. Permutation calibration is exact in finite samples at any dimension. A permutation-calibrated maximum over our dictionary would also control level under the full null. We therefore do not claim a better global test, and we expect rough power parity there. What permutation cannot do is produce intervals under the alternative. Once $P \ne Q$ the pooled labels are no longer exchangeable, so permuting them no longer reproduces the null distribution of the statistic. Permutation therefore says nothing about how large $W_u$ is or which directions differ. Our calibration experiments compare band coverage directly against what a permutation-inverted procedure would give.

Why not just apply the coordinate-wise high-dimensional CLT to mean differences? That test, from the 2013 multiplier bootstrap paper, sees only shifts in the marginal means. Models B–D in Neural Wasserstein Two-Sample Tests involve variance changes and non-Gaussian differences with matched means, and the mean-difference test is blind to these by construction. It also cannot detect dependence changes that only pair directions reveal. It is included in the power study, where we expect it to win on pure mean shifts and fail elsewhere.

Why not just learn the best projection on half the data, as Neural Wasserstein does? Learning can find dense, nonlinear directions that a fixed sparse dictionary misses. If the true difference is dense, we will lose. But splitting halves the sample used for the test, and the conditional theory gives no simultaneous statement across directions. Our dense-alternative simulations will show where the crossover lies, and we will report it.

Why not just use entropic or kernel max-sliced distances? The entropic limit theory of Goldfeld and coauthors holds only for fixed regularization, so it estimates a biased surrogate of the Wasserstein distance. Kernel max-sliced Wasserstein (its ICML 2025 analysis) has no limiting distribution at all and requires an $n\times n$ semidefinite relaxation. Neither yields interpretable per-direction intervals in the original coordinates.

Proposed Experiments

Calibration is the decisive check. We simulate under three settings: sub-Gaussian data, Student-$t_5$ data, and the spiked transport models of Xi and Niles-Weed, in which distributions differ only on a two-dimensional subspace and so most directions are degenerate. We vary $d \in \{60, 200, 500, 2000, 5000\}$ and $n \in \{100, 400\}$. These values mirror the $p=5000$, $n\in\{100,400\}$ setting where Chernozhukov et al. report P-P alignment (no numeric distance reported). We use coordinate and pair dictionaries and measure simultaneous coverage of all $W_u$ and the median band width. The proposal is falsified if coverage at $d=5000$, $n=100$ falls more than a few points below nominal under sub-Gaussian data. We will report degradation under $t_5$ as a measured cost, not hide it.

A second simulation shows the full-sphere obstruction. We will plot the null rejection rate of the full-sphere maximum against $d/n$, together with the null magnitude of that statistic, to check the predicted $\sqrt{d/n}$ growth.

Power and localization are tested on reconstructed versions of Models A–D at $d \in \{60, 200, 500\}$, on Ball versus Sphere, and on MNIST downsampled to $7\times 7$, comparing against all baselines above.

The real-data showcase is the TCGA methylation comparison: 511 LGG versus 150 GBM samples on 207 genes. We report the global p-value against Neural Wasserstein's $p<0.001$. The real deliverable there is the short list of genes and gene pairs whose intervals exclude zero, at a controlled simultaneous level, using every sample.