In my paper Muon Dynamics as a Spectral Wasserstein Flow, I connect a matrix optimizer to optimal transport by changing how a distribution’s motion is measured.
Muon replaces the nonzero singular values of its momentum matrix by ones. Ignoring momentum and stochasticity, the compact SVD \(G_k=\nabla F(X_k)=U_k\Sigma_k V_k^\top\) gives the update direction
\[X_{k+1}=X_k-\eta_k U_kV_k^\top.\]The spectral Wasserstein distance uses a norm \(\gamma\) on symmetric matrices, monotone in the positive semidefinite order. For probability measures on \(\mathbb R^d\) with finite second moments, define
\[\begin{aligned} \Sigma_\pi&=\int(y-x)(y-x)^\top\,\mathrm d\pi(x,y),\\ \mathsf W_\gamma(\mu,\nu)^2 &=\inf_{\pi\in\Pi(\mu,\nu)}\gamma(\Sigma_\pi). \end{aligned}\]Here \(\Pi(\mu,\nu)\) denotes all couplings. The trace choice \(\gamma(S)=\operatorname{tr}(S)\) recovers \(\mathsf W_2\); the operator norm \(\gamma(S)=\|S\|_{\mathrm{op}}\) measures the largest directional mean-square displacement.
The connection comes from the infinitesimal cost \(\gamma(\int vv^\top\,\mathrm d\mu)\). For particles \(\mu_X=n^{-1}\sum_i\delta_{x_i}\) stacked as rows of \(X\), the operator choice gives \(n^{-1}\|\dot X\|_{\mathrm{op}}^2\). For a loss \(\mathcal F\) and \(F(X)=n\mathcal F(\mu_X)\), steepest descent reads
\[\begin{aligned} G&=\nabla F(X)=U\Sigma V^\top,\\ \dot X&=-\|G\|_*UV^\top, \end{aligned}\]where \(\|G\|_*=\sum_i\sigma_i\) is the nuclear norm. This is Muon’s polar direction with a scalar speed factor: the paper’s idealized continuous-time, zero-momentum mean-field dynamics is a \(\mathsf W_\gamma\) gradient flow.