CDT Reduces Modes for Transport-Dominated Conservative PDEs

Eulerian reduced-order models struggle when a PDE solution mostly moves rather than changes shape. A small linear basis in physical coordinates must approximate many shifted profiles, so singular values decay slowly and low-rank reconstructions smear fronts or shocks. This paper studies that failure mode for one-dimensional conservative PDEs and proposes a transport-coordinate fix: apply the cumulative distribution transform (CDT), run POD/SVD in transformed coordinates, then invert the transform back to physical density space.

The contribution is partly theoretical and partly numerical. The paper does not claim a new time integrator or a full online ROM pipeline. Its main claim is narrower: for nonnegative, equal-mass scalar states, CDT coordinates expose low-dimensional structure that standard Eulerian POD hides.

Core Contribution

The key mathematical move is to represent a density through the monotone transport map from a fixed reference density. In one dimension, the CDT turns the 2-Wasserstein distance between densities into a weighted $L^2$ distance in transform space. Spatial translations become affine shifts, which is exactly the geometry that transport-dominated conservation laws produce.

Figure 2 illustrates the paper's central picture: snapshots that trace a curved, hard-to-compress path in native space become much closer to a low-dimensional set after the CDT.

The cleanest result is for linear transport. If $u(x,t)=u_0(x-At)$, the transformed trajectory lies in the span of the transformed initial datum and the constant function. In the paper's Kolmogorov-width language, the transformed solution manifold has zero width for every $n \geq 2$. That is the reason POD needs so few modes after the coordinate change.

Technical Approach

The numerical method, CDT-POD, is simple. The authors collect high-fidelity snapshots into a matrix $U$, transform each column with respect to a fixed reference density $r$, form a transformed snapshot matrix $\widehat{U}$, compute a rank-$k$ POD/SVD approximation in CDT space, and reconstruct physical snapshots with the inverse CDT. The native baseline is the usual rank-$k$ SVD approximation of $U$.

The comparison is subtle. Native POD minimizes Euclidean reconstruction error in the sampled physical coordinates. CDT-POD minimizes an error in transform space, where the norm corresponds to a one-dimensional Wasserstein metric. After inverse transformation, the physical reconstruction error is not guaranteed to be the direct optimum for the native $L^2$ metric. The paper handles this by reporting both singular-value decay in transform space and physical-space reconstruction errors after inversion.

The theory extends beyond pure translation. For scalar hyperbolic conservation laws, the authors prove two bounds in CDT coordinates: an $O(n^{-1})$ Kolmogorov-width estimate that remains meaningful for entropy solutions after shock formation, and a sharper $O(n^{-2})$ pre-shock estimate under stronger smoothness assumptions. For conservative advection-diffusion, the CDT trajectory stays within $O(\sqrt{DT})$ of the pure-transport plane, with sharper $O(D^2T^2)$ estimates under extra regularity or away from initial layers. As $D \to 0$, the linear-transport zero-2-width behavior is recovered.

Results and Analysis

The experiments test inviscid Burgers, traffic flow, Buckley-Leverett, and advection-diffusion equations. Across the hyperbolic examples, the figures compare rank-2 and rank-5 reconstructions, time-dependent $L^2$ errors, singular-value decay, and global $L^2$ error versus number of retained modes. The recurring pattern is that CDT singular values drop faster than native singular values, and the rank-2 or rank-5 CDT reconstructions preserve advected peaks and fronts better than Eulerian POD.

Figure 5 shows the Burgers case, including the shock time, rank-2 and rank-5 reconstructions, and the separation between native-space and transform-space singular-value decay.

The advection-diffusion section is useful because it probes the boundary of the method. The authors use the computational domain $(x,t)\in[0,30]\times[0,1.5]$, 1500 spatial points, and 200 output times. For one case, $A=5$ and $D=0.5$; for a three-regime comparison, they test pure advection $(A,D)=(5,0)$, pure diffusion $(0,1)$, and combined advection-diffusion $(5,1)$. CDT-POD is strongest when advection dominates. For pure diffusion, it is roughly on par with native POD, and the transform-inverse error floor remains visible.

The evidence supports the paper's main claim for 1D conservative, mass-preserving, displacement-heavy dynamics. It does not establish that CDT-POD is a drop-in replacement for all ROM settings. The method depends on admissible monotone transport maps, nonnegative equal-mass states, and a reliable inverse CDT; when reduced coordinates break monotonicity, post-processing or constraints may be needed.