Organizations: Korea Institute for Advanced Study, Seoul, Republic of Korea · Department of Statistics, Sungkyunkwan University, Seoul, Republic of Korea · Department of Mathematics, University of Seoul, Seoul, Republic of Korea
Functional data such as time series, trajectories, and fields are increasingly central in machine learning and are naturally modeled in infinite-dimensional Hilbert spaces. We study Semi-dual Neural Optimal Transport (SNOT) for learning transport maps between probability measures on separable Hilbert spaces. A critical challenge in this setting is the spurious solution problem: a neural map can globally optimize the SNOT objective while failing to recover an optimal transport map. We show that this failure arises from non-regular source measures, which make the inner minimization in the semi-dual objective non-identifiable. We prove that regular source measures, defined via Gaussian null sets, restore inner-minimizer uniqueness and recover the Monge map. For singular sources, we introduce Gaussian smoothing and establish a necessary and sufficient kernel condition for smoothing to restore regularity. We further prove plan-level consistency: as smoothing vanishes, every Wasserstein accumulation point of the induced optimal plans solves the original Kantorovich problem. Experiments on synthetic singular transports and real-world unpaired time-series imputation benchmarks validate the theory, achieving the best performance on several benchmarks and competitive performance on the remaining ones.
Figures & tables
Figure 1: Comparison of transport maps learned by vanilla HiSNOT (top row) and HiSNOT with annealed smoothing (bottom row). The vanilla model exhibits spurious solutions in singular cases (Perpendicular, One-to-Many, Grid), while smoothing improves agreement with the target distribution in these finite-mode experiments.
Smoothing
Perpendicular
Parallel
One-to-Many
Grid
Dcost
Dtarget
Dcost
Dtarget
Dcost
Dtarget
Dcost
Dtarget
✗
0.101
0.249
0.001
0.000
0.051
0.288
0.007
0.031
✓
0.038
0.006
0.003
0.001
0.001
0.004
0.007
0.008
Table 1: Quantitative comparison of HiSNOT performance with and without Gaussian smoothing across synthetic scenarios. We report the transport cost error Dcost(↓) and target distribution error Dtarget(↓) . Smoothing ( ✓ ) reduces the reported errors in three of the four tested scenarios.
Figure 2: Smoothing directions in the effective two-mode space. Left: Smoothing the second mode makes the source absolutely continuous in H2 . Right: Smoothing only the first mode leaves a singular kernel marginal; the trained map exhibits a spurious-solution pattern.
Datasets
ETTh1
ETTh2
ETTm1
ETTm2
Exchange
Methods
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
MSE
MAE
Transformer [ 33 ]
0.338
0.410
0.282
0.383
0.085
0.196
0.056
0.161
0.320
0.264
DLinear [ 40 ]
0.187
0.307
0.138
0.262
0.131
0.250
0.107
0.227
0.288
0.238
TimesNet [ 37 ]
0.343
0.415
0.147
0.278
0.086
0.209
0.093
0.220
0.372
0.308
FreTS [ 39 ]
0.233
0.355
0.159
0.278
0.061
0.168
0.040
0.136
0.229
0.189
PatchTST [ 27 ]
0.216
0.338
0.144
0.272
0.055
0.153
0.033
0.122
0.227
0.188
Table 2: Unpaired time-series imputation performance (MSE and MAE) across five datasets. Results are averaged over missing ratios of 0.5 and 0.7 . The bold and underlined values denote the best and second-best results. All baseline results are taken from Wang et al. [35] .
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 3: Illustrations of non-unique optimal transport maps in the perpendicular example.
Window size
Stride
HiSNOT ( W2↓ )
SNOT ( W2↓ )
720
1
9.632±0.201
9.341±0.341
360
1
5.009±0.161
5.133±0.120
240
1
3.543±0.031
3.817±0.010
240
2
3.687±0.232
4.426±0.080
240
3
3.681±0.143
4.783±0.120
120
4
2.055±0.126
2.383±0.085
Appendix
Table 4: Function-level generation on Exchange. Evaluation uses 1,000 generated and 1,000 target windows. The ± terms are reported dispersion values; their definition and replication count are unavailable in the provided experiment record. Bold marks the lower reported central W2 value within each row.
Active smoothing modes
Kernel condition in H5
Dtarget↓
e1
Violated
0.2063
e1,e2
Satisfied
0.0276
e1,e2,e3
Satisfied
0.0300
e1,e2,e3,e4
Satisfied
0.0249
Appendix
Table 5: Smoothing-support ablation in H5 . The source occupies modes e3,e4,e5 with a joint coefficient density. The condition column records whether its unsmoothed projection is absolutely continuous in the finite-dimensional covariance kernel. Replication counts and dispersion estimates are unavailable in the provided experiment record.
εmax
Dtarget↓
0.0
0.288
0.2
0.004
0.5
0.004
0.6
0.003
0.8
0.004
1.0
0.004
Appendix
Table 6: Initial smoothing-level ablation on One-to-Many. The remaining annealing protocol is fixed. Replication counts and dispersion estimates are unavailable in the provided experiment record.
εmin
Dtarget↓
0.02
0.008
0.06
0.003
0.10
0.004
0.14
0.004
0.18
0.004
0.20
0.004
Appendix
Table 7: Terminal smoothing-level ablation. The initial level is fixed at εmax=0.5 . Replication counts and dispersion estimates are unavailable in the provided experiment record.