Matrix-Valued Transport on Orbit Complexes

Many transport problems carry both a symmetry and a matrix at every point. A diffusion tensor field in a crystal-like material, the quantum state of an electron hopping between lattice sites, and a spin density on a periodic structure all fit this description. Optimal mass transport, the problem of finding the cheapest way to reshape one distribution into another, now has matrix-valued versions built for such data. They are solved on the full grid or graph and ignore the symmetry. The one symmetry-aware transport solver in the literature treats every orbit as identical and every density as a scalar. That assumption fails exactly where symmetric problems are most interesting: at special points such as mirror lines, rotation centers, and high-symmetry lattice sites. There a subgroup of the symmetry fixes the point, and that subgroup acts on the matrix stored there.

This note proposes an exact reduction. A transport problem between two symmetric matrix-valued densities can be solved on the quotient cell complex, meaning one copy of each symmetry-inequivalent cell. Each cell stores a short list of small matrix blocks, one block per irreducible representation of the subgroup fixing that cell. The blocks on neighboring cells are coupled by the induction and restriction maps between those subgroups. This coupling data is already known to topologists as the Bredon coefficient system. The equivariant K-homology literature has computed its degree-zero homology for $SL(3,\mathbb{Z})$, for Coxeter groups, and for 3D hyperbolic reflection groups. The proposal therefore offers two things. The first is a reduced solver that is exact for symmetric data, including stabilized cells, and should be much cheaper than the full problem. The second is a checkable prediction: on site-and-bond networks, the number of independent "sector masses" that any transport must conserve equals a Bredon homology rank that others have already computed. For $SL(3,\mathbb{Z})$ that number is 8, the number of conjugacy classes of finite-order elements.

Core Claim

Let a discrete group $G$ act properly on a cell complex $X$, and let each point carry a positive semidefinite matrix on which stabilizers act unitarily. For a cell $\sigma$ with stabilizer $G_\sigma$, Schur's lemma forces a $G$-invariant density to be block diagonal there: $\rho(\sigma) = \bigoplus_{\pi} (I_{\dim \pi} \otimes \rho_\pi)$. The sum runs over the irreducible representations $\pi$ of $G_\sigma$, and each $\rho_\pi$ is a small positive semidefinite matrix whose size is the multiplicity of $\pi$.

The first claim concerns the reduced equation. The dynamic, Benamou–Brenier-style matrix transport of Chen, Georgiou and Tannenbaum and Ryu et al. restricts to $G$-invariant densities. It then becomes a vector/matrix transport on the orbit complex $X/G$ whose continuity equation reads

$$\partial_t \rho_v + \sum_{e \ni v} \mathrm{Ind}_{G_e}^{G_v}(J_e) = \mathcal{L}_v(\rho_v).$$

Here $\rho_v$ is the tuple of blocks at vertex orbit $v$, and $J_e$ is the matrix flux carried on edge orbit $e$ in the blocks of $G_e$. The operator $\mathrm{Ind}_{G_e}^{G_v}$ maps edge blocks to vertex blocks by the induction branching isometries. $\mathcal{L}_v$ is the within-block quantum (Lindblad-type) mixing already present in those papers. Symmetry forbids any mixing across different irreducible blocks at a site.

This equation has the same shape as the vector transport of Chen, Georgiou and Tannenbaum, in which a "mutation graph" says which internal components may exchange mass. The proposal's step is to take that coupling structure from group theory rather than design it by hand. The first claim is that the reduced distance equals the full distance divided by the orbit count, up to solver tolerance.

The second claim concerns the site network. Here transport runs only along the bonds of the 1-skeleton, as in lattice hopping models. The linear functionals of block traces that every admissible flux leaves unchanged are then dual to the cokernel of $\partial_1$ in the Bredon chain complex with complex representation coefficients. So their number equals the rank of Bredon $H_0$. Published calculations predict this count without any numerics: 8 for $SL(3,\mathbb{Z})$ and 16 for $GL(3,\mathbb{Z})$ (Sánchez-García), the spherical-subset formulas for right-angled and even Coxeter groups (Sánchez-García), and $cf(\Gamma)$ for 3D hyperbolic reflection groups (Lafont, Ortiz, Rahm and Sánchez-García). A transport between two symmetric states is feasible only if these charges match.

To prevent overclaiming, one point needs stating up front. In the continuum regime, generic points have trivial stabilizer and connect every sector through the interior. Total trace is then the only conserved quantity, and the Bredon count applies only to the site network.

Research Program

The theoretical work has three parts. We will first prove the exact reduction lemma. The proof checks that symmetric critical points of the full action are critical points of the reduced action. It also checks that the anti-commutator and logarithmic-mean multiplications of Chen et al. commute with the stabilizer action, so that they preserve the block structure.

We will then prove the conservation statement on networks. Over the reals only ranks matter, so this part is linear algebra once the induction couplings are fixed.

Finally, we will state where degree concentration enters and where it does not. For real-valued densities, integral torsion is irrelevant. It becomes relevant only in a quantized variant, where block masses are integer occupation numbers. Hidden conserved quantities modulo $n$ could then appear if $H_0$ had torsion. The total-unimodularity "vertex block" certificate of Lafont et al. rules this out for their groups. We will test integer reachability as a secondary hypothesis, not a headline result.

The computational work builds the reduced solver in two versions. The first is a primal–dual hybrid gradient (PDHG) solver with the closed-form proximal steps of Ryu et al.. The second is an inexact sequential quadratic programming solver in the style of the SQP algorithm. In the SQP version, each weighted Poisson solve becomes a solve with a block sheaf Laplacian on $X/G$.

We will generate the coupling data automatically from cell stabilizers and character tables. GAP already produced such tables for the Bredon calculations in Sánchez-García. The cell complexes come from Soulé's retract for $SL(3,\mathbb{Z})$ and from Davis complexes for Coxeter groups.

The $K(\pi,1)$ direction plays a bounded role, as a limiting case. Artin groups are torsion-free, so every stabilizer on the Salvetti complex (the standard finite cell model for these groups) is trivial. The method must then collapse to ordinary graph transport with a single block. Paolini and Salvetti proved asphericity for affine Artin groups, which guarantees that this complex computes group homology. The collapse is therefore a clean correctness check on a family of infinite groups. We do not expect new phenomena there, and we say so.

Baselines

The naive reference point is the unreduced problem: full matrix-valued transport on the whole grid or a finite patch of $X$, with symmetric source and target data, solved by the GPU PDHG method of Ryu et al.. That paper reports about 13.2 seconds for matrix transport on a $256 \times 256$ grid, about 1.36 seconds for the vector case, and a 50x to 150x speedup over serial CPU. We will rerun it on identical data, hardware, and stopping tolerance. It serves both as ground truth for correctness and as the cost to beat. Beating it means matching its transport distance to within solver tolerance (relative error below $10^{-6}$) at a wall-clock cost that falls roughly in proportion to the orbit count.

The second-order reference is the inexact SQP solver of the SQP paper. It reports 35 iterations at density contrast 10 and 62 at contrast 50 on $64 \times 64 \times 40$ grids. We will report iteration counts at matched contrast and tolerance, full versus reduced. The reduced problem should not need more iterations, and any change in how iterations grow with contrast will be reported.

The strongest existing symmetry-aware competitor is SG-ROT from Optimal Transport with Symmetry Groups. It detects orbits from distance patterns and solves on $X/G$, reporting about 1.8 seconds versus over 37 seconds for unreduced transport on 14,400 points, with agreement to about $10^{-16}$. We will apply it in the only way it applies: scalar mass per orbit, with no stabilizer blocks, which gives it the same orbit data and the same tuning budget. Beating it means that the proposal stays exact at stabilized cells while SG-ROT's distances and interpolations measurably deviate from ground truth there, at comparable cost.

The key ablation keeps everything in the reduced solver except the induction couplings. These are replaced either by identity maps (plain reflecting boundaries on the fundamental domain) or by a single block per cell. The ablation isolates whether the Bredon coefficient system itself is what buys exactness.

Why not just…?

Why not just run the full matrix solver on symmetric data? This works for finite or periodic domains, and it is our ground truth there. It wastes a factor equal to the orbit count, and in a transport solver that factor multiplies every Poisson solve. For infinite groups such as $SL(3,\mathbb{Z})$ or hyperbolic reflection groups, no finite patch is invariant. Truncation breaks the symmetry at the patch boundary, so the full solver is not even an option in that regime. The experiments measure both the cost gap and the truncation bias.

Why not just use SG-ROT-style orbit reduction? The method in Optimal Transport with Symmetry Groups assumes orbits that all look alike and a scalar mass. Its own scope says nothing about matrix densities or cells whose stabilizers differ, such as $S_4$ at one vertex and $C_2 \times C_2$ on an edge in Soulé's complex. We expect it to be exact on free orbits and wrong at stabilized cells. The mirror-line and rotation-center errors in the wallpaper-group experiment below will show whether that expectation holds. If SG-ROT turns out to be exact whenever the data happen to be scalar multiples of the identity, we will report that as the regime where our extra machinery buys nothing.

Why not just diagonalize each density and transport the eigenvalues? The block decomposition here is forced by symmetry, not chosen per point. Inside a block, eigenbases still rotate along the path. Chen et al. show that densities along an optimal matrix path generally do not commute across times, so eigenvalue transport loses the rotational part of the cost. Including eigenvalue transport as a baseline quantifies that loss on our data.

Why not just compute the conserved charges numerically, as the nullspace of the discretized divergence? This is a real threat to the second claim, and we will do it anyway, because it is the check. The Bredon calculation adds two things. It gives the count before building anything. It also gives a basis indexed by conjugacy classes of finite-order elements, which makes the charges interpretable. And it applies to the infinite complex rather than to a truncated patch whose boundary can add spurious kernel. The solver claim does not depend on this point. If numerical nullspaces are always cheaper and never mislead, the counting result shrinks to a consistency check, and we will say that.

Why not just work over the rationals and ignore torsion? For real-valued densities, this is correct. Ranks are all that matter, and we do not claim that total unimodularity helps the continuous solver. Integral structure matters only in the quantized, integer-occupation variant. There we test whether torsion-freeness, certified as in Lafont et al., coincides with the absence of hidden mod-$n$ reachability obstructions.

Proposed Experiments

The first experiment uses continuum wallpaper groups and has ground truth. We will generate $256 \times 256$ periodic fields of $3 \times 3$ diffusion-tensor-like densities, following the setup of Ryu et al. but symmetrized under three Euclidean triangle reflection groups. The first has all labels even, so Bredon homology is concentrated in degree zero (Sánchez-García). The other two contain odd labels, so concentration fails. Each problem is solved three ways: full PDHG, SG-ROT-style scalar reduction, and our reduced solver on one fundamental triangle with dihedral blocks at the corners and $C_2$ blocks on the mirrors.

This experiment is decisive for the first claim. The reduced distance should match the full distance to solver tolerance for all three groups. The ablation with identity couplings should show errors concentrated on the mirrors and corners. The odd-label groups tell us whether nonzero degree-one homology shows up as slow, zero-cost circulating modes in the primal–dual iterations. We treat that as an open question, not a prediction.

The second experiment uses site networks for groups that have published Bredon calculations: $SL(3,\mathbb{Z})$ and $GL(3,\mathbb{Z})$ on Soulé's complex, right-angled and even Coxeter groups on Davis complexes, and a cocompact 3D hyperbolic reflection group from Lafont et al.. We will build the reduced bond network and compute the dimension of its conserved-functional space numerically, then compare it with the predicted ranks: 8, 16, the spherical-subset formulas, and $cf(\Gamma)$. We will then pick source and target states that agree in total trace but differ in exactly one predicted charge. The reduced solver should report infeasibility. Scalar orbit reduction should instead return a finite, and wrong, distance.

As a free-action control, the same pipeline runs on the Salvetti complexes of small affine Artin groups. Every stalk must collapse to one block, and the charge count must equal the number of connected components.

For all experiments we report distance relative error against ground truth wherever it exists, wall-clock time, and iteration counts at contrasts 10 and 50 so they are comparable to the SQP benchmark. We also report predicted versus measured charge counts and the size of charge violations by each baseline. Two outcomes would refute the proposal: reduced distances that deviate from full distances beyond tolerance, or measured charge counts that differ from the Bredon ranks. Either would mean the induction-coupled continuity equation is not the correct reduction.