Sliced Wasserstein distances compare two datasets by projecting both onto many random lines, then measuring on each line how far the one-dimensional histograms must be moved to match. They are cheap to compute and do not degrade as the data dimension grows, so they are widely used to compare generative models, point clouds, and cell populations. Sometimes the question is only whether two datasets differ. Often it is harder: how far apart are they, with an error bar, and is model A really closer to the data than model B? The asymptotic theory behind these error bars requires every projected distribution to have connected support, meaning no gaps. Real data routinely breaks this. Point clouds, count data, and any dataset with clusters or atoms project to one-dimensional distributions full of gaps. No one has checked whether the resulting confidence intervals are still calibrated.
We propose a smoothed sliced Wasserstein statistic that removes the connected-support requirement exactly, together with an empirical study of calibration on real benchmarks. The mechanism comes from the Fourier slice theorem, the fact that the one-dimensional Fourier transform of a projection is a line through the multi-dimensional Fourier transform. Taken in its version for generalized functions, this theorem implies that blurring the data with a Gaussian and then projecting gives the same result as projecting and then blurring in one dimension. This holds for any distribution, including point masses and measures on lower-dimensional sets. After the blur, every projected distribution has a positive density on the whole line, so the gap problem disappears by construction. The smoothed one-dimensional transport can also be computed exactly from Gaussian-mixture distribution functions, with no random noise injected. The core claims are empirical and falsifiable. The first is that standard sliced-Wasserstein bootstrap intervals under-cover on real discrete data. The second is that the exactly smoothed statistic restores nominal coverage at a small, quantifiable bias.
What is known, and where the gap is
Three directions meet here, and they play different roles.
The first direction supplies the analytic tool. Distributional Extension and Invertibility of the k-Plane Transform and Its Dual extends the transform that integrates over $k$-dimensional planes to tempered distributions, which are generalized functions that include point masses. It proves a Fourier slice theorem valid for every tempered distribution, covering the Radon ($k=d-1$) and line ($k=1$) cases. We reuse that identity. We do not need its harder invertibility results, which concern the dual backprojection operator. The companion Function-Space Optimality of Neural Architectures with Multivariate Nonlinearities builds variational theory on the same transforms. Both papers state "Datasets: not discussed." Our proposal gives this machinery its first concrete empirical use: a statistic whose validity on singular data we then measure.
The second direction supplies the inference theory and its limits. Distributional Convergence of the Sliced Wasserstein Process views the per-direction transport costs as a random function on the sphere of directions. It shows this function converges to a Gaussian process, which justifies the bootstrap for standard, max-, and distributional sliced Wasserstein distances. The proof assumes compact supports and connected one-dimensional projections, since gaps make the optimal one-dimensional map non-unique and break the linearization. This is precisely the assumption we remove.
Central limit theorems for empirical transportation cost in general dimension similarly requires positive densities and $(4+\delta)$ moments, which excludes discrete and manifold data. Limit distribution theory for smooth p-Wasserstein distances shows that Gaussian smoothing restores limit theory for full-dimensional transport. We borrow its tail and moment arguments, but apply smoothing after slicing, where the transport problem is one-dimensional and cheap.
Minimax-Optimal Two-Sample Test with Sliced Wasserstein shows permutation tests with sliced Wasserstein are minimax-optimal, provided the number of directions $L$ grows in proportion to $n$. That result constrains our design in two ways. It tells us permutation testing of "no difference" already works, so our target is interval estimation when the distributions differ. It also tells us that direction sampling must be part of the uncertainty accounting.
The closest bridging work is Fourier Sliced-Wasserstein Embedding for Multisets and Measures. It slices measures, takes a cosine transform of each one-dimensional quantile function, and proves the resulting embedding approximates sliced Wasserstein distance. It was evaluated on ModelNet40 and single-cell RNA-seq. It uses a Fourier transform of quantile functions, not the Fourier slice theorem for the measure itself, and it offers no confidence intervals or limit theory. We differ on both counts: our Fourier-side object is the characteristic function of the data, and our product is calibrated inference, not an embedding.
The third direction, the empirical validation gap, is the evaluation lens. The sliced-Wasserstein limit theory above was validated on synthetic data meeting its own assumptions. We ask what happens on data that violates them.
Core Claim
Write $P$ and $Q$ for the two data distributions, $u$ for a unit direction, $u_\sharp P$ for the distribution of $\langle u, X\rangle$ with $X\sim P$, and $N_\sigma$ for a centered Gaussian with standard deviation $\sigma$ (in $d$ dimensions on the left, one dimension on the right). The Fourier slice theorem states that the characteristic function of $u_\sharp P$ at frequency $t$ equals the characteristic function of $P$ at $tu$. Multiplying both sides by the Gaussian's characteristic function gives
$$u_\sharp (P * N_\sigma) = (u_\sharp P) * N_\sigma, \qquad \text{and likewise } A_\sharp(P*N_\sigma) = (A_\sharp P)*N_\sigma^{(k)}$$
for any $k$-frame $A$. Here $N_\sigma^{(k)}$ is the $k$-dimensional Gaussian. The identity holds for every tempered distribution, so it covers atoms and measures supported on surfaces.
We define the smoothed sliced statistic as
$$SW_{p,\sigma}^p(P,Q) = \int_{\mathbb{S}^{d-1}} W_p^p\big((u_\sharp P)*N_\sigma,\ (u_\sharp Q)*N_\sigma\big)\, d\sigma(u).$$
Each smoothed projection has an everywhere-positive density, which is what the linearization in the sliced-process proof needs. We therefore conjecture the following. For any $P$ and $Q$ with finite moments of order slightly above $2p$, and any fixed $\sigma>0$, the empirical smoothed process converges to a Gaussian process on the sphere. The bootstrap would then be valid with no connected-support or compactness assumption. The price is a bias relative to unsmoothed $SW_p$, which is at most of order $\sigma$ for $p=1$ because Gaussian convolution moves each point by at most its noise.
For empirical data, $(u_\sharp P_n)*N_\sigma$ is an explicit mixture of $n$ Gaussians. Its distribution and quantile functions can be evaluated to machine precision. We therefore compute the smoothed statistic exactly rather than by adding random jitter to the samples.
Research Program
The theoretical part has one target: proving the conjectured Gaussian-process limit for $SW_{p,\sigma}$ at $P\neq Q$, for $p\in\{1,2\}$ and $k\in\{1,2\}$. The proof would adapt the functional delta method of the sliced-process paper. Compactness would be replaced by moment conditions of the kind used for smooth Wasserstein limits. Connectedness would be replaced by the strictly positive smoothed density. We will also derive a combined bootstrap that resamples both data points and directions, and characterize when $L$ directions suffice relative to $n$, motivated by the $L\propto n$ requirement in the minimax testing paper. If the proof fails for $p=2$ with heavy tails, that is a reportable boundary, not a reason to drop the empirical study.
The empirical part is a coverage audit. A "population" is a large fixed pool of data. Examples are all ModelNet40 point clouds in two classes, all cells of two types in the RNA-seq data, or all MNIST digits of two classes. The true distance is computed on the full pool with many directions. We then repeatedly subsample $n$ points per side, build a nominal 95% interval with each method, and record how often it covers the pool value, along with interval width and wall-clock time.
These pools are discrete by construction, so their projections always have gaps. This is the regime the existing theory excludes. We also measure a simple diagnostic, the largest gap in each projected empirical distribution relative to its spread, to test whether under-coverage tracks gap size as the theory predicts. Model comparison is the downstream use case. On MNIST, we will ask whether the method correctly ranks two corrupted versions of the data by their distance to the clean pool.
Baselines
The naive reference point is the plug-in unsmoothed sliced Wasserstein distance with the empirical bootstrap justified by Distributional Convergence of the Sliced Wasserstein Process. It is what practitioners would run today. It uses the same samples, the same $L$ directions, and the same number of bootstrap replicates as our method. Our method beats it if its coverage is within two Monte Carlo standard errors of 95% wherever the naive method falls measurably below, at comparable runtime.
The strongest existing competitor for interval estimation on discrete data is the Sinkhorn divergence, the entropically regularized transport cost. Its limit distributions are derived in Limit theorems for entropic optimal transport maps and Sinkhorn divergence, and its parametric rate holds without density assumptions according to Statistical bounds for entropic optimal transport: sample complexity and the central limit theorem. We will tune its regularization strength over a grid of the same size as our grid over $\sigma$, and give it the same time budget. Beating it means matching its coverage with narrower intervals or substantially lower time at the dimensions of ModelNet40 and RNA-seq. Its target quantity differs from ours, so we compare coverage of each method's own estimand, plus agreement in ranking on the MNIST model-comparison task.
For power and ranking comparisons we also run the projected and kernel-projected Wasserstein tests from Two-sample Test using Projected Wasserstein Distance and Two-Sample Test with Kernel Projected Wasserstein Distance, which those papers evaluated on MNIST and MSRC-12. We also run the Euclidean distance between Fourier Sliced-Wasserstein embeddings with a naive bootstrap. The research file contains no reported numbers for these methods on these benchmarks: no reported result found on Lacuna for coverage, power, or accuracy on ModelNet40, RNA-seq, MNIST, or MSRC-12 for any of these baselines. We will rerun all of them under our protocol rather than quote figures.
Two ablations isolate the key ingredient. The first replaces exact Gaussian-mixture smoothing with random jitter added to the samples, the same thing in expectation but with extra randomness. This tells us whether exact computation matters for calibration. The second sweeps $\sigma$ from zero upward to trace the trade-off between coverage and bias against the unsmoothed population distance. We also compare $k=1$ with $k=2$ slices to see whether planes reduce the number of directions needed.
Why not just…?
Why not just add Gaussian noise to the data and run ordinary sliced Wasserstein? By the identity above, this targets the same population quantity, so the question is fair. The difference is in the estimator. Injected noise adds a second source of randomness that the standard bootstrap does not account for, and it requires extra draws to control. Exact convolution removes that randomness at no asymptotic cost. The jitter ablation settles this directly. If jitter matches exact smoothing in coverage and width, we will report that the Fourier-slice route buys only computational convenience.
Why not just use permutation tests, which are valid in finite samples regardless of support? For the question "do $P$ and $Q$ differ," permutation tests are the right tool and are already minimax-optimal per Minimax-Optimal Two-Sample Test with Sliced Wasserstein. We do not compete there. Permutation gives no interval for the size of a nonzero distance and no way to compare two models' distances to data. Those are the targets of our coverage audit.
Why not just use Sinkhorn divergence, whose inference already covers discrete data? This is the real threat, and it is why Sinkhorn is the primary baseline. Its limit theory does not need connected supports. If its intervals are calibrated and comparably cheap on our benchmarks, the case for smoothed slicing reduces to cost and to keeping the sliced geometry practitioners already use. The coverage and timing comparison will show which holds. We will state plainly if Sinkhorn dominates.
Why not just use a kernel distance on characteristic functions, since the slice theorem moves everything to the Fourier side? Unweighted Fourier-side distances of this kind lose the transport interpretation: they do not measure how far mass moves. Our statistic stays a Wasserstein cost on each smoothed slice and recovers sliced Wasserstein as $\sigma\to 0$. The MNIST ranking task tests whether this geometric fidelity changes conclusions in practice.
Decisive Checks
Three outcomes would decide the proposal. The first is whether naive sliced-Wasserstein bootstrap intervals measurably under-cover on the discrete pools of ModelNet40, RNA-seq, and MNIST, with under-coverage increasing with the gap diagnostic. If they cover correctly, the theoretical gap is harmless in practice, which is itself worth reporting under the validation-gap lens. The second is whether exact smoothing restores 95% coverage at a $\sigma$ small enough that the bias against unsmoothed sliced Wasserstein is below the interval width. The third is whether this holds at lower cost than Sinkhorn intervals with matched coverage. A positive result on all three would give the first empirically validated inference procedure built on the distributional Fourier slice theorem. A negative result on the third would still show which existing tool practitioners should use instead.
References
- Distributional Extension and Invertibility of the \(\boldsymbol{k}\)-Plane Transform and Its Dual
- Function-Space Optimality of Neural Architectures with Multivariate Nonlinearities
- Distributional Convergence of the Sliced Wasserstein Process
- Central limit theorems for empirical transportation cost in general dimension
- Limit distribution theory for smooth p-Wasserstein distances
- Minimax-Optimal Two-Sample Test with Sliced Wasserstein
- Fourier Sliced-Wasserstein Embedding for Multisets and Measures
- Limit theorems for entropic optimal transport maps and Sinkhorn divergence
- Statistical bounds for entropic optimal transport: sample complexity and the central limit theorem
- Two-sample Test using Projected Wasserstein Distance
- Two-Sample Test with Kernel Projected Wasserstein Distance