An empirical investigation of the open problem: which local minima are reached, and do the recoverability theorem's feasibility assumptions hold at those minima?
Riemannian AmbientFlow minimizes a combined objective with a variational lower bound and geometric regularization:
The recoverability theorem requires three feasibility assumptions:
The open question: do gradient-based optimizers find local minima that satisfy all three conditions?
Unit circle S1 parameterized by f*(z) = (cos z, sin z). Intrinsic dim d=1, ambient dim D=2.
Unit sphere S2 via inverse stereographic projection. Intrinsic dim d=2, ambient dim D=3.
Helix f*(t) = (cos t, sin t, t/2π). Intrinsic dim d=1, ambient dim D=3.
Objective value across 10 random initializations for each λ. High variance indicates multiple distinct local minima.
Standard deviation of the converged objective across starts. Higher spread indicates more distinct local minima.
| λ | Circle (std) | Sphere (std) | Helix (std) |
|---|---|---|---|
| 0.00 | 1.837 | 0.160 | 2.083 |
| 0.01 | 1.208 | 0.128 | 1.823 |
| 0.05 | 1.208 | 0.117 | 1.379 |
| 0.10 | 1.624 | 0.133 | 0.058 |
| 0.50 | 1.198 | 0.143 | 0.031 |
| 1.00 | 1.202 | 0.187 | 1.391 |
| 2.00 | 1.606 | 0.281 | 0.048 |
Aggregate feasibility score combining data matching (F1), posterior matching (F2), and geometric constraint (F3).
Breaking down the aggregate score into its three components reveals the fundamental trade-off.
Distribution of directional second derivatives at converged critical points. All positive curvature confirms genuine local minima.
| Manifold | λ | Min Eigenvalue | Max Eigenvalue | Mean | Negative Dirs |
|---|---|---|---|---|---|
| Circle | 0.0 | 12.28 | 461.67 | 210.71 | 0/50 |
| Circle | 0.1 | 10.83 | 476.92 | 229.86 | 0/50 |
| Circle | 0.5 | 10.45 | 580.49 | 256.59 | 0/50 |
| Circle | 1.0 | 11.52 | 671.04 | 279.47 | 0/50 |
| Sphere | 0.0 | 8.18 | 76.64 | 36.81 | 0/50 |
| Sphere | 0.1 | 8.69 | 75.44 | 37.65 | 0/50 |
| Sphere | 0.5 | 7.08 | 70.88 | 33.68 | 0/50 |
| Sphere | 1.0 | 62.76 | 859.35 | 375.30 | 0/50 |
| Helix | 0.0 | 141.03 | 1870.47 | 878.69 | 0/50 |
| Helix | 0.1 | 81.96 | 982.71 | 396.46 | 0/50 |
| Helix | 0.5 | 124.54 | 1512.43 | 614.73 | 0/50 |
| Helix | 1.0 | 326.87 | 3896.46 | 1648.25 | 0/50 |
Tracking a single local minimum as λ increases from 0 to 2. Reveals smooth deformation without bifurcation.
Comparing the Riemannian geometry (pullback metric trace) of learned vs. ground-truth diffeomorphisms.
| Manifold | λ | Tr(Gθ(0)) | Tr(G*(0)) | Ratio |
|---|---|---|---|---|
| Circle | 0.0 | 0.450 | 1.000 | 0.450 |
| Circle | 0.1 | 0.405 | 1.000 | 0.405 |
| Circle | 1.0 | 0.261 | 1.000 | 0.261 |
| Sphere | 0.0 | 1.010 | 8.000 | 0.126 |
| Sphere | 0.1 | 0.825 | 8.000 | 0.103 |
| Sphere | 1.0 | 0.515 | 8.000 | 0.064 |
| Helix | 0.0 | 0.416 | 1.025 | 0.406 |
| Helix | 0.1 | 0.388 | 1.025 | 0.378 |
| Helix | 1.0 | 0.263 | 1.025 | 0.256 |
These findings suggest that closing the gap between the theoretical recoverability guarantee and practical optimization will require:
Based on the open problem from: Diepeveen et al., "Riemannian AmbientFlow: Towards Simultaneous Manifold Learning and Generative Modeling from Corrupted Data," arXiv:2601.18728, January 2026.