Spectral Envelope Divergence

• Spectral Envelope Divergence is a frequency-domain metric that compares real and synthetic categorical sequences by capturing their strongest latent periodic structures.
• It is computed as a normalized L2 distance between the mean spectral envelopes of real and synthetic data, ensuring accurate reflection of cyclical and seasonal patterns.
• SED serves as both an evaluation metric and a differentiable loss in relational time series synthesis, offering improved benchmarks over traditional time-domain methods.

Spectral Envelope Divergence (SED) is a frequency-domain discrepancy metric for categorical time series. It compares the mean spectral envelopes of real and synthetic categorical sequences, where a spectral envelope summarizes the strongest latent periodic structure obtainable by optimally scaling the categories. As a named metric, SED is introduced for relational time-series synthesis in "Sequential RC-TGAN: Generating Relational Time Series with Spectral Envelope Loss" (Gueye et al., 30 Jun 2026). In that formulation, SED is lower-is-better: small values indicate preservation of latent cyclical and seasonal structure, whereas large values indicate distorted or mismatched spectral structure.

1. Formal definition

For a categorical attribute $d_j$, Spectral Envelope Divergence is defined as

$$\mathcal{D}_{env}(d_j) = \frac{\left\| \bar{\lambda}^{(real)}(\cdot; d_j) - \bar{\lambda}^{(synth)}(\cdot; d_j) \right\|_2}{K-1},$$

where $\bar{\lambda}^{(real)}$ and $\bar{\lambda}^{(synth)}$ are the mean spectral envelopes of the real and synthetic sequences, and $K$ is the number of categories (Gueye et al., 30 Jun 2026).

The global metric averages this quantity over all categorical attributes: $$\overline{\mathcal{D}_{env}} = \frac{1}{J}\sum_{j=1}^{J}\mathcal{D}_{env}(d_j),$$ with $J$ the number of categorical features. The metric is explicitly described as a frequency-domain discrepancy measure for categorical time series in relational databases. Its intended use is evaluation of whether synthetic data preserves the same latent cyclical and seasonal structure as the source data, rather than merely matching local transitions or marginal category frequencies.

This definition makes SED a comparison between functions over frequency, not a pointwise disagreement between labels. The object being compared is the spectral envelope curve itself, averaged over sequences associated with a feature.

2. Spectral-envelope foundation

SED is built directly on Spectral Envelope Theory. Let $X_t$ be a stationary categorical time series taking values in $\{a_0,\dots,a_{K-1}\}$, and let

$$\beta = (\beta_0,\dots,\beta_{K-1})^\prime \in \mathbb{R}^K$$

be a scaling vector assigning a numerical value to each category. The transformed process $$\mathcal{D}_{env}(d_j) = \frac{\left\| \bar{\lambda}^{(real)}(\cdot; d_j) - \bar{\lambda}^{(synth)}(\cdot; d_j) \right\|_2}{K-1},$$0 takes the value $$\mathcal{D}_{env}(d_j) = \frac{\left\| \bar{\lambda}^{(real)}(\cdot; d_j) - \bar{\lambda}^{(synth)}(\cdot; d_j) \right\|_2}{K-1},$$1 whenever $$\mathcal{D}_{env}(d_j) = \frac{\left\| \bar{\lambda}^{(real)}(\cdot; d_j) - \bar{\lambda}^{(synth)}(\cdot; d_j) \right\|_2}{K-1},$$2. The spectral envelope is then defined by

$$\mathcal{D}_{env}(d_j) = \frac{\left\| \bar{\lambda}^{(real)}(\cdot; d_j) - \bar{\lambda}^{(synth)}(\cdot; d_j) \right\|_2}{K-1},$$3

Using the one-hot encoded process $$\mathcal{D}_{env}(d_j) = \frac{\left\| \bar{\lambda}^{(real)}(\cdot; d_j) - \bar{\lambda}^{(synth)}(\cdot; d_j) \right\|_2}{K-1},$$4, with spectral density matrix $$\mathcal{D}_{env}(d_j) = \frac{\left\| \bar{\lambda}^{(real)}(\cdot; d_j) - \bar{\lambda}^{(synth)}(\cdot; d_j) \right\|_2}{K-1},$$5 and variance matrix $$\mathcal{D}_{env}(d_j) = \frac{\left\| \bar{\lambda}^{(real)}(\cdot; d_j) - \bar{\lambda}^{(synth)}(\cdot; d_j) \right\|_2}{K-1},$$6, the same quantity is written as

$$\mathcal{D}_{env}(d_j) = \frac{\left\| \bar{\lambda}^{(real)}(\cdot; d_j) - \bar{\lambda}^{(synth)}(\cdot; d_j) \right\|_2}{K-1},$$7

This is a generalized Rayleigh quotient, and its solution is the largest eigenvalue of the appropriately normalized spectral matrix (Gueye et al., 30 Jun 2026).

Within this framework, $$\mathcal{D}_{env}(d_j) = \frac{\left\| \bar{\lambda}^{(real)}(\cdot; d_j) - \bar{\lambda}^{(synth)}(\cdot; d_j) \right\|_2}{K-1},$$8 acts as an envelope over all normalized spectra obtainable from category scalings. Peaks in $$\mathcal{D}_{env}(d_j) = \frac{\left\| \bar{\lambda}^{(real)}(\cdot; d_j) - \bar{\lambda}^{(synth)}(\cdot; d_j) \right\|_2}{K-1},$$9 reveal hidden cycles or periodicities; a flat envelope indicates more white-noise-like behavior. SED therefore compares not raw categorical trajectories, but the strongest periodic structure recoverable from those trajectories under optimal category scaling.

3. Computation and analytical properties

The computation of SED proceeds in three layers: sequence-level envelope estimation, feature-level averaging, and feature-aggregated comparison. For each categorical sequence, one first computes $\bar{\lambda}^{(real)}$0 by one-hot encoding the categorical process, estimating $\bar{\lambda}^{(real)}$1 and $\bar{\lambda}^{(real)}$2, and solving the generalized eigenvalue problem. For a categorical feature $\bar{\lambda}^{(real)}$3, the real-data mean envelope is then

$\bar{\lambda}^{(real)}$4

with an analogous definition for synthetic data. In training, the same averaging is performed over mini-batches: $\bar{\lambda}^{(real)}$5 again with a synthetic counterpart (Gueye et al., 30 Jun 2026).

The envelope discrepancy is the $\bar{\lambda}^{(real)}$6 distance

$\bar{\lambda}^{(real)}$7

Dividing by $\bar{\lambda}^{(real)}$8 yields the normalized per-feature SED. The stated purpose of this normalization is scale normalization reflecting the categorical dimensionality constraint.

The metric is supported by two analytical properties. First, the paper states a continuity lemma: every spectral envelope in $\bar{\lambda}^{(real)}$9 is continuous on $\bar{\lambda}^{(synth)}$0. Second, it proves norm bounds

$\bar{\lambda}^{(synth)}$1

The paper uses these facts to argue that the envelope is a well-behaved function in $\bar{\lambda}^{(synth)}$2 and that SED is mathematically stable (Gueye et al., 30 Jun 2026).

4. Role in relational time-series synthesis

SED is not only an evaluation metric; it is also aligned with a differentiable training objective in Sequential RC-TGAN. The model introduces a spectral loss

$\bar{\lambda}^{(synth)}$3

where the categorical component is

$\bar{\lambda}^{(synth)}$4

The evaluation SED uses the same $\bar{\lambda}^{(synth)}$5-based envelope comparison, but with normalization and averaging across categorical attributes (Gueye et al., 30 Jun 2026).

This setup is motivated by a specific failure mode of standard methods on categorical relational time series. One-hot encoding makes categories orthogonal and equally distant, so models trained only in the time domain may reproduce local transitions while missing periodicity, seasonality, harmonic structure, and long-range cyclical behavior. SED is intended to measure precisely this missing global structure.

The same paper places SED beside Spectral Density Divergence (SDD), forming a unified frequency-domain evaluation family. For continuous attributes $\bar{\lambda}^{(synth)}$6,

$\bar{\lambda}^{(synth)}$7

and

$\bar{\lambda}^{(synth)}$8

The stated distinction is categorical series $\bar{\lambda}^{(synth)}$9 SED and continuous series $K$0 SDD.

Because spectral envelope theory is described as inherently designed for categorical sequences, the paper extends its frequency-domain regularization to continuous variables through Variational Gaussian Mixture discretization. For each continuous attribute $K$1, a VGM is fit, the most probable component is selected by

$K$2

and an intra-mode normalized scalar is retained as

$K$3

This produces a discrete mode sequence that can be analyzed with spectral envelope theory while still supporting reconstruction of the continuous value.

5. Benchmarks, exact envelopes, and empirical behavior

A distinctive feature of the SED framework is the use of categorical benchmarks with known theoretical spectral envelopes. The paper simulates stationary first-order Markov chains, letting $K$4 denote the one-hot encoding of the categorical state and $K$5 the transition matrix. It then gives

$K$6

where $K$7 is the stationary distribution and $K$8. For circulant $K$9, the spectral envelope is analytically tractable: $$\overline{\mathcal{D}_{env}} = \frac{1}{J}\sum_{j=1}^{J}\mathcal{D}_{env}(d_j),$$0 where the eigenvalues are $$\overline{\mathcal{D}_{env}} = \frac{1}{J}\sum_{j=1}^{J}\mathcal{D}_{env}(d_j),$$1 (Gueye et al., 30 Jun 2026).

Two benchmark processes are defined. The Noisy Cyclic Process (NCP) is

$$\overline{\mathcal{D}_{env}} = \frac{1}{J}\sum_{j=1}^{J}\mathcal{D}_{env}(d_j),$$2

with $$\overline{\mathcal{D}_{env}} = \frac{1}{J}\sum_{j=1}^{J}\mathcal{D}_{env}(d_j),$$3. As $$\overline{\mathcal{D}_{env}} = \frac{1}{J}\sum_{j=1}^{J}\mathcal{D}_{env}(d_j),$$4, it approaches a near-deterministic cycle with sharp harmonic peaks; as $$\overline{\mathcal{D}_{env}} = \frac{1}{J}\sum_{j=1}^{J}\mathcal{D}_{env}(d_j),$$5 decreases, phase noise broadens the peaks. The Symmetric Sticky Process (SSP) is

$$\overline{\mathcal{D}_{env}} = \frac{1}{J}\sum_{j=1}^{J}\mathcal{D}_{env}(d_j),$$6

with $$\overline{\mathcal{D}_{env}} = \frac{1}{J}\sum_{j=1}^{J}\mathcal{D}_{env}(d_j),$$7, and its spectral envelope simplifies to

$$\overline{\mathcal{D}_{env}} = \frac{1}{J}\sum_{j=1}^{J}\mathcal{D}_{env}(d_j),$$8

As $$\overline{\mathcal{D}_{env}} = \frac{1}{J}\sum_{j=1}^{J}\mathcal{D}_{env}(d_j),$$9, this produces a sharp peak at $J$0; as $J$1, the spectrum becomes white-noise-like.

These chains are embedded into the child table of a relational database, with $J$2 stored in the parent table. The reported setup uses 100 parent entities, one $J$3 per parent, and one 10,000-step child sequence per parent, with $J$4, $J$5, and $J$6. Baselines include SDV, ClavaDDPM, DoppelGANger, and TimeGAN. On these simulated benchmarks, Seq. RC-TGAN reports substantially lower SED than the baselines. For NCP, the reported SED values are $J$7 at $J$8, $J$9 at $X_t$0, and $X_t$1 at $X_t$2, with reported relative improvements of 45.8%, 43.3%, and 28.2% over the second-best baselines. For SSP, the reported SED values are $X_t$3, $X_t$4, and $X_t$5, with reported improvements of 32.7%, 52.8%, and 37.48% (Gueye et al., 30 Jun 2026).

The paper interprets low SED as correct preservation of cyclicity, seasonality, and latent periodic structure. It emphasizes two evaluation criteria: frequency localization, meaning peaks at the true theoretical frequencies, and spectral purity, meaning the absence of hallucinated periodic artifacts. It also contrasts SED with MSE(ACF), noting that good autocorrelation error can still coincide with failure to represent seasonal structure.

6. Relation to neighboring concepts and terminological ambiguity

The term “spectral envelope” has multiple meanings across different research areas, and most adjacent literatures do not define Spectral Envelope Divergence as a named quantity.

Context	Spectral-envelope object	Status of SED
Singing synthesis	Frame-wise acoustic target represented by 60 log Mel-frequency spectral coefficients	No dedicated spectral envelope divergence is defined (Bous et al., 2019)
NMF source separation	LPC-estimated instrument envelope used as a constraint in excitation-filter decomposition	No separate quantity literally called spectral envelope divergence (Park et al., 2018)
Harmonic analysis	Weak$X_t$6 closure $X_t$7 of squared moduli of trigonometric polynomials on the circle group	No divergence, distance, or comparison functional is defined (Lawton, 2012)

In singing synthesis, spectral envelopes are acoustic targets predicted frame-by-frame in a vocoder pipeline. The reported training loss is mean squared error on log amplitudes, with additional experiments involving mean absolute error and Sobolev norms, while evaluation is perceptual rather than based on a specialized envelope-distance measure (Bous et al., 2019). In instrument-sound separation using NMF, the optimized divergence is KL divergence between the observed spectrogram and its reconstruction, and envelope similarity is imposed indirectly through convex-combination updates that replace or average the envelope component of basis vectors; the paper explicitly states that it does not define a separate quantity literally called “spectral envelope divergence” (Park et al., 2018).

A more fundamental ambiguity arises in harmonic analysis, where the “spectral envelope” of a subset $X_t$8 is the weak$X_t$9 closure of normalized squared trigonometric polynomials with frequencies in $\{a_0,\dots,a_{K-1}\}$0. That setting concerns probability measures on the circle group, symbolic dynamics, and the Kadison–Singer program, and again no divergence between spectral envelopes is introduced (Lawton, 2012).

There are also nearby but distinct notions of “spectral divergence.” "Multivariate Spectral Estimation based on the concept of Optimal Prediction" defines a prediction-based divergence family between multivariate spectral densities, reducing in limit cases to multivariate Itakura–Saito and Kullback–Leibler-type divergences (Zorzi, 2014). "The noncommutative residue, divergence theorems and the spectral geometry functional" defines residue-based spectral divergence functionals built from commutators with Dirac-type operators (Wang et al., 22 Jun 2025). These are terminologically related, but neither concerns categorical time-series spectral envelopes in the sense of SED.

This suggests a narrow and domain-specific meaning for Spectral Envelope Divergence: within the cited literature, it is specifically a normalized $\{a_0,\dots,a_{K-1}\}$1 discrepancy between real and synthetic mean spectral envelopes for categorical time series, introduced to assess whether synthetic relational sequences preserve the latent periodic skeleton of the original data (Gueye et al., 30 Jun 2026).