Symmetry-Resolved Multi-Marginal Barycenters That Keep Their Holes

Suppose you have several point clouds that each sample the same kind of shape: a ring, a torus, the conformation space of a molecule. They were collected independently, so their coordinates don't line up. You want one "average" shape that summarizes them all. The standard tool for this is a Gromov-Wasserstein barycenter. Gromov-Wasserstein (GW) transport matches two datasets by comparing their internal distance patterns rather than their coordinates. Its multi-marginal version couples all N datasets jointly and outputs a consensus distance matrix. Our hypothesis is that on shapes with holes and symmetries, this average fills the holes in.

The mechanism is simple. A circle can be matched to another circle equally well under every rotation and reflection. A regularized or relaxed solver therefore returns a blend of many rotated matchings instead of one. Averaging distances over that blend pulls every pair of points toward the same typical distance, and the loop disappears. With more marginals, each one brings its own independent symmetry ambiguity, so we expect the collapse to get worse as N grows. We propose to test this failure mode directly. We then fix it with a symmetry-resolved multi-marginal barycenter. It attaches to every point intrinsic angle-valued coordinates computed from each dataset's own long-lived loops, and it jointly solves for one small symmetry correction (a rotation and a possible flip) per dataset. The expected payoff is that the transport plan becomes nearly deterministic and the consensus keeps the topology of its inputs.

Core Claim

Our claim has two parts, and both are falsifiable. The first is diagnostic. Relaxed multi-marginal GW barycenters of datasets that share non-trivial topology lose persistent loops and voids. This loss grows with the entropic regularization strength, with the symmetry of the shape, and with the number of marginals N. Plain pairwise topology metrics cannot attribute the loss to the averaging step, because they never see the coupling.

The second part is constructive. If the coupling is fused with topology-derived angle coordinates, and each marginal's residual symmetry is solved for explicitly, the barycenter recovers the correct persistent homology. It also recovers ground-truth correspondences, and the method stays tractable at the scales where current multi-marginal GW solvers run.

The problem sits between the two directions. Multi-Marginal Gromov-Wasserstein Transport and Barycenters supplies the method we build on. It proves that a free-support GW barycenter is equivalent to a single multi-marginal GW problem, gives a tight bi-convex relaxation and a multi-marginal Sinkhorn solver, and fuses in labels to keep semantic parts matched, as in its camel experiment. That paper never measures homology, and its fused variant needs externally supplied labels.

The topology-comparison line (Geometry Score, Topology Distance, Manifold Topology Divergence (MTop), and the Intrinsic Multi-Scale Distance) is strictly pairwise. It diagnoses but does not produce a consensus:

The motivation for suspecting hole filling comes from Implications of data topology for deep generative models. It shows that models scoring well on Wasserstein distance can still put mass inside a torus hole. That is evidence that transport objectives are blind to topology. It is not evidence about barycenters specifically, which is why our first step is diagnostic.

The Mechanism and the Fix

The theory motivates the construction. On the Existence of Monge Maps for the Gromov-Wasserstein Problem shows that even two-marginal GW with squared-distance cost can require optimal plans that split mass onto two graphs rather than one map. A general condition for Monge solutions in the multi-marginal optimal transport problem and Multi-marginal optimal transport on Riemannian manifolds give conditions for deterministic multi-marginal plans, but for linear costs rather than GW's quadratic one.

For symmetric inputs such as uniform circles, every symmetry-transformed plan is equally optimal. Entropic solvers therefore prefer their average. We will make this precise for $N$ uniform circles in a short analytic result. As the regularization grows, the barycenter distance matrix converges to the average of the input distance matrices over the symmetry group. For the circle that average is a constant off-diagonal matrix, which has no persistent first homology. The form follows directly from the fact that entropic plans are strictly positive and invariant under the shared symmetry.

The fix gives the solver something that breaks the symmetry but is intrinsic to each dataset. For each input space $(X_i, d_i, \mu_i)$ we compute persistent cohomology. For each long-lived one-dimensional class $k$ we extract an angle-valued coordinate $\theta_{i,k}: X_i \to S^1$, the standard circular-coordinate construction. These angles are well defined only up to a rotation-and-flip $g_{i,k} \in O(2)$ and up to a relabeling of classes across datasets. That residual ambiguity is exactly the shape's symmetry, but now it is a handful of discrete-plus-one-dimensional parameters rather than a mixture hidden in the plan. We solve

$$ \min_{D,\ \phi,\ \{\pi_i\},\ \{g_{i}\}} \ \sum_{i=1}^N \lambda_i \Big[ \sum_{x,x',y,y'} \big(d_i(x,x') - D(y,y')\big)^2 \pi_i(x,y)\pi_i(x',y') \;+\; \alpha \sum_{x,y} \big\| g_i \cdot \theta_i(x) - \phi(y) \big\|_{S^1}^2 \, \pi_i(x,y) \Big]. $$

Here $D$ is the consensus distance matrix on a fixed support $\{y\}$, and $\phi(y)$ is the consensus angle coordinate. $\pi_i$ couples dataset $i$ to the consensus, $\lambda_i$ are barycentric weights, $\alpha$ trades off geometry against topology, and $\|\cdot\|_{S^1}$ is arc distance. The fixed-support "dummy marginal" formulation from the multi-marginal GW paper lets this run as $N$ couplings to the consensus rather than an $n^N$ tensor.

We alternate among three steps: fused GW Sinkhorn updates for the plans, closed-form circular alignment for each $g_i$, and averaging for $D$ and $\phi$. Class relabeling is resolved by enumeration, because data with more than three long-lived loops is rare in the target benchmarks. For two-dimensional voids we use the same scheme with a sphere-valued coordinate where available. Otherwise we fall back to a persistence penalty on the output, described below as an ablation.

Research Program

The work has three stages.

The first stage is diagnostic. We build families of $N$ point clouds with known topology from the synthetic benchmarks in Implications of data topology for deep generative models: circles, the torus-ellipse images, the rotating-jar sphere, and the cyclooctane Klein-bottle-plus-sphere space. Each cloud is resampled, noised, and put through a random isometry and a mild smooth deformation. We add the 2D shapes (spades, hearts, camel) and the 40-particle rotation data used by the multi-marginal GW paper. We then run standard multi-marginal GW barycenters while sweeping regularization, N from 2 to 16, and the degree of symmetry, which we break by making circles slightly elliptical. On each output we measure the persistence of the longest-lived loop and void. This stage confirms or refutes the hole-filling hypothesis.

The second stage is the constructive method above, plus the circle result.

The third stage tests whether the correspondences, not just the consensus, are better. On synthetic data we know the true parameterization of every sample, so correspondence accuracy and three-way cycle consistency can be measured exactly. Cycle consistency means that composing the dataset-1-to-2 and 2-to-3 matchings reproduces the 1-to-3 matching. Transitivity is the property Hierarchical Multi-Marginal Optimal Transport for Network Alignment shows pairwise chaining violates on graphs. That paper motivates the joint formulation but works on discrete networks with anchor nodes, not on point clouds with holes.

Baselines

The naive reference is the pooled cloud. We take the union of all $N$ inputs after aligning each to the first by pairwise GW and compute its persistence. This shows what topology survives with no consensus step at all.

The strongest existing method is plain multi-marginal GW barycentering with the Sinkhorn solver and fixed-support construction of Multi-Marginal Gromov-Wasserstein Transport and Barycenters. It is the only method in the file that produces a joint N-way consensus of metric spaces. We run it on the identical support size, marginals, and wall-clock budget. Its regularization schedule gets the same tuning grid as ours, selected on held-out random seeds. We also give it the favorable setting of entropy annealing with multiple restarts, keeping the lowest-cost run.

We found no reported number on Lacuna for the persistent homology of multi-marginal GW barycenters, and none for correspondence accuracy on these synthetic families. The paper reports its particle and camel results qualitatively. We will therefore establish the baseline numbers ourselves and report them before our method's.

Beating the baseline means three things: statistically lower bottleneck distance between the barycenter's persistence diagram and the ground-truth diagram, higher correspondence accuracy and cycle consistency, and no worse GW distortion to the inputs. The margin must widen with N, as the mechanism predicts.

A second competitor is the persistence-diagram barycenter of Wasserstein Dictionaries of Persistence Diagrams and the merge-tree barycenters of Wasserstein Distances, Geodesics and Barycenters of Merge Trees. These average topological summaries directly. They set the ceiling on diagram accuracy but produce no consensus space and no correspondences.

Three ablations isolate the key ingredient:

  • Random features. We replace the circular coordinates with random angle-valued features of the same dimension. This tests whether any symmetry breaking suffices.
  • Fixed symmetry corrections. We fix every $g_i$ to the identity. This tests whether the explicit symmetry solve matters.
  • Persistence penalty. We drop the fusion term and instead add a differentiable persistence penalty on $D$, which rewards long-lived loops in the output.

Why not just…?

Why not just chain pairwise GW to a reference and average? Chaining is cheaper, and with a well-chosen reference it may preserve loops, because each pairwise plan is less entangled. Its weakness is that errors and symmetry choices in each pairwise step are independent, so correspondences through the reference need not be transitive. Hierarchical Multi-Marginal Optimal Transport for Network Alignment reports a Hits@10 gain of at least 12.0% for joint over pairwise alignment on graphs. That is supporting evidence from a different setting, not proof for manifolds. We include chaining as a baseline and measure cycle consistency directly.

Why not just lower the entropy, or anneal it with restarts? This is the most serious threat. If near-unregularized solvers settle into one symmetric branch, the fix may be unnecessary. Two things can go wrong with that route. At small entropy, relaxed GW solvers get stuck in poor local minima. And rounding a mixed plan can pick different branches for different points, which tears the loop instead of filling it. Our diagnostic sweep reports both effects. If annealed restarts match our method across N and symmetry levels, we will present the contribution as the diagnosis plus the circle result, and drop the claim for the construction.

Why not just average persistence diagrams? Diagram barycenters get the Betti numbers right almost by construction. They cannot say which point of dataset 3 corresponds to which point of dataset 7, nor produce a shape to sample from or visualize. We report their diagram accuracy as a ceiling, and our method should approach it while also delivering correspondences.

Why not just fuse with ordinary spectral or curvature features, as the fused multi-marginal GW variant does with labels? Features such as heat-kernel signatures, or the curvature quantities in Ollivier-Ricci Curvature for Hypergraphs, are invariant under the shape's symmetries. On a uniform circle they are constant, so they cannot break the tie. We test this by including heat-kernel-signature fusion as an ablation. The prediction is that it helps on asymmetric shapes and fails on symmetric ones.

Proposed Experiments

The first decisive check is the diagnostic curve. For uniform circles, tori, and the jar sphere, we plot longest-loop persistence of the plain multi-marginal GW barycenter against N and against regularization strength. If persistence does not fall, the premise is wrong and we stop.

The second check is the symmetry dose-response. Making circles progressively elliptical should shrink plain multi-marginal GW's hole-filling and shrink our method's advantage with it. That pattern would confirm that symmetry is the cause, rather than something else we happen to fix.

The third check is the head-to-head at matched compute on all synthetic families, cyclooctane, and the 2D shapes and rotating particles from the multi-marginal GW paper. We score diagram bottleneck distance to ground truth, correspondence accuracy, triple cycle consistency, and GW distortion.

The fourth check is external evaluation of the consensus. We decode the barycenter by multidimensional scaling into a point cloud. We then compare that cloud to fresh held-out samples of the true manifold using MTop's cross-barcode, chosen because on a noise and mode-dropping CIFAR-10 benchmark Manifold Topology Divergence reports a Kendall-tau of 0.89 with disturbance level versus 0.36 for Geometry Score. We also use IMD as a second, rotation-invariant check.

The last check is a stress test on unbalanced inputs, where some marginals miss part of the shape. Tightness of the relaxation is open in that setting, per Multi-Marginal Gromov-Wasserstein Transport and Barycenters. We report any failures there as limitations, not as wins.