We study prediction of nonlinear audit functionals from scalar observations of a shared hidden table state. For Schur-rational operators, local Gaussian projections yield a common sparse kernel for overlapping tasks, repeated row identities, and cross-row dependence. Shared-state residual samples extend this kernel through a positive-semidefinite feature representation. Positive matrix integration and sample-value enclosures connect the original inputs to the actual saved hybrid prediction. A sorted calibration-role perturbation bound then certifies its complete finite-pool pinball risk. Independent complete-loss selection chooses the ridge, score scale, and residual weight, while task-rank calibration supplies an upper endpoint. Evaluation uses twenty new reference contexts per dataset at two task sizes, giving 400 conditions across ten benchmark populations. The mean dataset complete-loss ratios are 0.9977 to the degree-two kernel, 0.9961 to matched complete-MC shrinkage, and 0.9911 to the retrospective minimum over 174 public configurations. The last ratio has a paired reference-bootstrap interval of [0.9898,0.9927]. On 40 profiles at coefficient tolerance 10−7, Gamma quadrature has a geometric mean cold-time ratio of .0646 to strengthened Arb/Petras integration.
Figures & tables
Dataset
K2
Matched MC
Public min
Gaussian
0.999877
0.9959
0.9927
Heteroscedastic
0.999773
0.9973
0.9925
Heavy-tailed
0.982489
0.9884
0.9908
Nonlinear
0.999966
0.9969
0.9941
Concrete
1.000782
0.9984
0.9846
Airfoil
0.995314
0.9957
0.9882
Table 1: Residual-rule complete-loss ratios on twenty new reference contexts per dataset. Each dataset averages original-unit loss across both task sizes and all contexts. Public min is the retrospective minimum over 174 public configurations.
Figure 1: Learning from scalar source observations on the fixed ten-dataset development panel. Complete risk is normalized by the zero-source analytic-mean risk within each dataset and then averaged. Full MC uses the matched 8192-state shrinkage rule. The selection budget is 512 outcomes at every source budget.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 2: Matched original-input certificate cost. Each point is one of 40 profiles and the black bars give geometric means. Both engines meet the same mean, coefficient-radius, and residual-energy contracts.
Figure 3: Frozen score-scale choices for the earlier degree-two rule on references 5301–5320 and both task sizes. Each choice uses the independent selection pool.
Figure 4: Multiscale subsampling on all twenty development conditions. Left, the mean subset risk divided by the complete full-pool risk. Right, bVarB(DB)/(N−b) divided by the delete-one jackknife statistic. Points show conditions and bars show medians and interquartile ranges. Each estimated subset variance uses 256 nested paired subsamples.
Figure 5: Covariance sampling variation on every development condition. Each point represents one reference and task-size condition and is estimated from four independent MC streams.
Figure 6: Direct covariance-control comparison on the complete development panel. Left, complete-loss ratios with paired hierarchical bootstrap intervals. Right, every condition’s ratio of covariance sampling variation, estimated separately for source and query–source blocks.
Figure 7: Complete source-observation curves for all ten datasets at development reference 4301. Each panel uses that dataset’s zero-source analytic-mean risk as its unit. The horizontal scale is linear from zero to 80 observations and logarithmic with base two above 80.
Figure 8: Task-size strata and the residual covariance comparison on the same fixed development panel. The two learning panels average dataset ratios separately at each task size.