Many datasets are compared with similarity measures that are not true distances. Examples include dynamic time warping (DTW, which aligns two time series by stretching them in time), edit distance between peptide sequences, and Hamming distances between ballots. When you take such a similarity matrix and try to place the points in a geometric space, some directions come out with negative squared length. A pseudo-Euclidean space allows exactly this: it has ordinary "positive" axes plus "negative" axes along which squared distances subtract. Work on learning with these indefinite similarities has argued that the negative directions carry real structure rather than noise. The clearest case is a sensor-presence task, where keeping them lifted accuracy from 0.79 for nearest-neighbour DTW to 0.98 (Multi-perspective embedding for non-metric time series classification).
Synthetic data is judged with tools that cannot see those directions. They either assume a positive-definite geometry, or they take the full indefinite similarity and collapse it into one number in which positive and negative mismatches can cancel. We propose to embed real and synthetic samples together in a pseudo-Euclidean space and to measure the real-versus-synthetic discrepancy separately on the positive and the negative axes. The result is a two-number fidelity signature. Our claim is testable. A generator can match the positive part while getting the negative part wrong. When that happens, the negative-side discrepancy predicts whether a model trained on the synthetic data fails on real data, and existing training-free metrics miss it.
Core Claim
Let $P$ be the real distribution and $Q$ the synthetic one. Pool $n$ real and $m$ synthetic samples, compute their pairwise dissimilarity in the native measure (DTW, alignment score, Hamming), double-center the squared dissimilarities as in classical multidimensional scaling, and eigendecompose. This gives each sample coordinates $x_+$ on the axes with positive eigenvalues and $x_-$ on the axes with negative eigenvalues. Write $\bar{x}^P_\pm$ and $\bar{x}^Q_\pm$ for the real and synthetic means on each side. The quantity one would get by plugging the indefinite kernel into maximum mean discrepancy (MMD, the standard kernel two-sample statistic) is
$$\mathrm{MMD}^2_{\text{ind}} = D_+ - D_-, \qquad D_\pm = \lVert \bar{x}^P_\pm - \bar{x}^Q_\pm \rVert^2 .$$
The pooled $\mathrm{MMD}^2_{\text{ind}}$ can be near zero while $P \neq Q$, because a positive-side error can be offset by a negative-side error. Clipping the negative eigenvalues to zero, which is the usual fix, simply reports $D_+$. We report $(D_+, D_-)$, together with a second-order version that compares the covariance spectra of real and synthetic coordinates on each side. Each statistic gets its own permutation null, built by reshuffling the real/synthetic labels and re-embedding. The prediction is concrete. In domains where the negative spectrum carries structure, $D_-$ explains variation in downstream train-on-synthetic, test-on-real (TSTR) performance that $D_+$, clipped MMD, and subspace metrics do not.
One caveat shapes the design. In an indefinite space, splitting into positive and negative parts is not unique in general. We define the split canonically through the pooled eigendecomposition, which is the same choice classical MDS makes. We then treat its stability under subsampling as something to measure, not something to assume.
Relation to Prior Work
Three papers motivate this work by showing the gap. A Geometry-Aware Metric for Mode Collapse in Time Series Generative Models introduced DMD-GEN. It decomposes trajectories into linear dynamic modes, treats the resulting subspaces as points on a Grassmannian (the space of linear subspaces), and compares real and synthetic sets by optimal transport. That geometry is Euclidean throughout, so it cannot register warping-induced indefiniteness. Our method replaces the linear-mode subspace with the native non-metric dissimilarity. Learning Real-Life Approval Elections showed that simple synthetic election models miss the multimodal structure of real participatory-budgeting elections, and it scored fidelity with voter-anonymous Hamming distance. It left open why some elections stay hard to learn, and it never looked at the spectrum. On the Equivalency, Substitutability, and Flexibility of Synthetic Data found that behavioural complexity, not photorealism, decides whether synthetic tracking data helps. Its evidence was that multi-group synthetic crowds raised DanceTrack HOTA from 68.8 to 73.5, while single-person synthetic data hurt. But it could only reach this conclusion by training a full tracker, which is the cost a training-free metric should remove.
Four papers supply the method. The case that negative eigenvalues are structural and that clipping, flipping, or shifting them distorts the data comes from Learning in indefinite proximity spaces - recent trends. We also reuse its warning about $O(N^3)$ cost. The Nyström landmark approximation (computing the embedding from a small set of reference points) that keeps negative eigenvalues comes from Multi-perspective embedding for non-metric time series classification. Multi-Dimensional Scaling on Groups proved that for rankings and binary hypercubes under Hamming distance the MDS eigenvectors are group characters. This gives one testbed where the spectrum is known exactly, so we can check the estimator against ground truth. Lifting Weak Supervision To Structured Prediction is the closest use of pseudo-Euclidean embeddings next to synthetic data. It embeds finite label metrics isometrically to recover the error rates of noisy labelling functions, and it uses synthetic Mallows models only as a testbed. It never asks whether a synthetic set is faithful, which is our whole question.
Research Program
The first component is the estimator. It computes the pooled pseudo-Euclidean embedding with Nyström landmarks drawn from both real and synthetic samples, then produces $D_+$, $D_-$, the per-side covariance-spectrum distances, and permutation p-values. On the hypercube and $S_n$ testbeds, where Multi-Dimensional Scaling on Groups gives the exact spectrum, we can also break the signature down by group character. A mismatch then becomes interpretable, for example "the generator misses pairwise co-approval structure."
The second component is a controlled study of when the signature matters. We build synthetic sets that match the real positive coordinates exactly but have distorted negative coordinates (rescaled, reflected, or collapsed). We also build the reverse case, and a cancelling case in which $D_+ \approx D_-$. These sets are mapped back to dissimilarities and fed to every metric and every downstream learner.
The third component is an analysis of natural generators. It asks whether $D_-$ adds predictive power for TSTR loss beyond the existing metrics. The test is a regression of TSTR loss on all metrics jointly, evaluated across generators and datasets with leave-one-dataset-out validation.
Baselines
The do-nothing reference is TSTR itself: train the downstream model on synthetic data, test it on real data. It is the ground truth we try to predict cheaply, not a competitor. A training-free metric has value only if it ranks generators the way TSTR does. In the cheapest naive variant, the "metric" is the raw mean dissimilarity between synthetic and real samples. Any proposal must beat this ranking.
The strongest existing training-free competitor on time series is DMD-GEN (A Geometry-Aware Metric for Mode Collapse in Time Series Generative Models). We run it with its published procedure on exactly the same real and synthetic samples from TimeGAN, TimeVAE and DiffusionTS on Sines, Stock, Energy and ETTh. We give it the same sample budget and tune its hyperparameters on the same held-out dataset. The file reports no numeric DMD-GEN values for these benchmarks (only which generator scored best), so the comparison is on agreement with TSTR rankings: rank correlation and partial $R^2$. Beating it means a higher leave-one-dataset-out rank correlation with TSTR, with a confidence interval over generator seeds that excludes DMD-GEN's.
On elections, the strongest competitor is voter-anonymous Hamming distance from Learning Real-Life Approval Elections. That paper reports a Pearson correlation of 0.993 between model log-likelihood and va-ham distance across Pabulib elections. Because va-ham already tracks likelihood so closely, we will not claim to beat it on that axis. The claim to test is narrower: on the elections that paper calls hard to learn, $D_-$ or the per-character signature separates the 4-component mixture from real data where va-ham shows a small gap.
Two kernel baselines isolate the key ingredient. Clipped MMD (negative eigenvalues set to zero, i.e. $D_+$ alone) and pooled indefinite MMD ($D_+ - D_-$) are computed from the same embedding. Neither uses the separation. If either matches the two-sided signature on the controlled and natural studies, the separation is not doing the work.
Why not just…?
Why not just clip the negative eigenvalues and use ordinary MMD or kernel distances? This is the standard fix, and Learning in indefinite proximity spaces - recent trends argues it discards real structure. The more direct evidence is that downstream accuracy depends on the negative part, as in the 0.79 versus 0.98 presence-detection result. Still, "learners use it" does not prove that "fidelity metrics need it." The controlled study settles this. Clipped MMD is blind to negative-only distortions by construction. The open question is whether such distortions occur in real generators, which the natural-generator regression answers.
Why not just run a two-sample test directly on the raw DTW or Hamming dissimilarities, for example a nearest-neighbour test? This is the most serious threat. It needs no embedding and no eigendecomposition, it has no basis-choice ambiguity, and it may match our detection power. We run it as a baseline on every experiment. If it equals the two-sided signature in both detection power and TSTR prediction, then our contribution shrinks to a diagnostic: the signature says which side of the geometry is wrong, which a nearest-neighbour test cannot. We will report the result in those terms rather than claim a better detector.
Why not just use the indefinite kernel in MMD without splitting? It is the same computation minus one subtraction, so it is the natural ablation. The cancelling construction is designed to show its failure mode. If cancellation never appears in natural generators, the split adds little except interpretability, and we will say so.
Proposed Experiments
The time-series study uses the DMD-GEN benchmarks with DTW dissimilarity. It also adds the multivariate presence-detection data from Multi-perspective embedding for non-metric time series classification, which we generate synthetically with the same three generator families. The downstream learners are NN-DTW and MPE-cGMLVQ, so TSTR is measured both with a learner that keeps the negative part explicitly and with one that uses raw distances.
The structured-data study uses Pabulib elections with 1-, 2- and 4-component approval models and Impartial Culture as generators. Exact group spectra on the hypercube serve as a check on the Nyström estimator. The sequence study uses the AMPScan and DeepAMP peptide sets from Krein support vector machine classification of antimicrobial peptides. Its synthetic peptides come from simple mutation and shuffling generators, and the downstream learner is Kreĭn-SVM, whose reported real-data test AUCs are 0.967 on AMPScan and 0.863 on DeepAMP and which serve as reference points. If M³Act's simple and multi-group subsets are obtainable, a tracking check asks whether the signature computed on trajectory DTW ranks multi-group above single-person against DanceTrack trajectories. That is the ordering the 68.8 to 73.5 HOTA result implies.
Three checks are decisive. First, on the controlled negative-only distortions, $D_-$ must detect them at nominal permutation-test size, while clipped MMD and DMD-GEN do not, and the distortions must actually lower TSTR for the negative-aware learner. If they do not lower TSTR, the negative axes do not matter for fidelity. Second, across natural generators, adding $D_-$ to a regression that already contains $D_+$, DMD-GEN, and the raw-dissimilarity nearest-neighbour test must significantly raise held-out prediction of TSTR loss. Third, the sign and magnitude of $D_-$ must stay stable across Nyström landmark draws and subsample sizes. If they do not, the basis dependence of the split undermines the metric, and we will report that as a negative result.
References
- Multi-perspective embedding for non-metric time series classification
- A Geometry-Aware Metric for Mode Collapse in Time Series Generative Models
- On the Equivalency, Substitutability, and Flexibility of Synthetic Data
- Learning in indefinite proximity spaces - recent trends
- Multi-Dimensional Scaling on Groups
- Lifting Weak Supervision To Structured Prediction
- Krein support vector machine classification of antimicrobial peptides