---
title: Generalised Singular Spectrum (GSS)
url: https://www.emergentmind.com/topics/generalised-singular-spectrum-gss
type: topic
---

# Generalised Singular Spectrum (GSS)

Searching arXiv for the cited papers on Generalised Singular Spectrum and related SSA extensions.
Generalised Singular Spectrum (GSS) is a context-dependent term centered on singular-value-based structure extraction. In the SSA lineage, it is best understood as a broad generalization of classical Singular Spectrum Analysis in which the standard embedding–decomposition–grouping–reconstruction scheme is extended from univariate Hankel trajectory matrices to richer structured objects, including multivariate, multidimensional, shaped, functional, interval-valued, projected, and non-Hermitian constructions [1907.02579, 1309.5050, 1906.05232, 2011.03872, 2008.03413]. In a distinct 2025 usage, the term denotes the complex quantity
\[
\varepsilon_j(k)=\sigma_j(k)e^{-i\Arg \langle u_j(k),v_j(k)\rangle},
\]
derived from the singular value decomposition of a non-Hermitian dynamical matrix and interpreted as a band structure for driven-dissipative systems [2509.19433]. The term is therefore not fully standardized, but in both lineages it preserves a common core: structured data are embedded into a singular-spectrum object, decomposed into elementary singular components, and then interpreted through reconstruction, filtering, or band topology.

## 1. Conceptual scope

Within the SSA family, the most general abstract formulation is a four-step scheme: embed the input object by an operator \(T\) into a structured trajectory matrix from a set \(H\), decompose that trajectory matrix into rank-one elementary matrices, group them, and project grouped matrices back to \(H\) before applying \(T^{-1}\) [1907.02579]. In this sense, “generalised singular spectrum” denotes not a single algorithm but a class of methods obtained by changing the embedding geometry, the structured matrix class, the decomposition operator, or the reconstruction rule while retaining the singular-spectrum logic.

The multidimensional/image literature makes this generalization particularly explicit. There, classical 1D SSA is extended in at least five senses: generalized embedding from 1D lag vectors to multidimensional moving-window patches, multidimensional lag-covariance PCA, adaptive filter-bank interpretation of eigenspaces, exact reconstruction through completeness of eigenvectors, and low-rank/noise separation by grouping or truncating components [1505.01599]. This broader framing treats GSS less as a fixed named formalism than as a family resemblance across SSA extensions.

## 2. Classical SSA substrate

Classical SSA starts from a univariate series \(X=(x_1,\dots,x_N)\), chooses a window length \(L\), and forms the Hankel trajectory matrix \(\mathbf X=\mathcal T(X)\) with \(K=N-L+1\) columns. Its singular value decomposition is
\[
\mathbf X=\sum_{m=1}^d \sqrt{\lambda_m}\,U_mV_m^T,
\]
where \((\sqrt{\lambda_m},U_m,V_m)\) are the eigentriples, and grouped reconstruction is obtained by summing selected rank-one terms and projecting back to the Hankel class [1907.02579]. In signal-extraction form, this can be written compactly as
\[
\widetilde S=\mathcal T^{-1}\Pi_{\mathcal H}\Pi_r\mathcal T X.
\]

A complementary interpretation, foundational for many GSS variants, views SSA as a parallel FIR filter bank. The \(m\)-th eigenvector defines an analysis filter
\[
H_m(z)=\sum_{k=0}^{M-1}u_{km}z^{-k}
\]
and a synthesis filter
\[
F_m(z)=\frac{1}{M}\sum_{k=0}^{M-1}u_{km}z^k,
\]
with branch transfer function
\[
T_m(z)=H_m(z)F_m(z).
\]
Because the branch impulse response is symmetric, each branch is zero-phase, and the full bank satisfies
\[
\sum_m T_m(e^{j\omega})=1,
\]
which is an exact-reconstruction partition of unity [1807.10679]. In the 1D filtering interpretation emphasized in the multidimensional SSA literature, each mode contributes the zero-phase transfer \(|\hat v^{(k)}_\alpha|^2\), with completeness
\[
\frac{1}{K}\sum_{k=1}^K |\hat v^{(k)}_\alpha|^2 =1,
\]
so the singular spectrum is simultaneously an eigenspectrum, a filter bank, and a reconstruction system [1505.01599].

## 3. Generalizations within the SSA family

A first axis of generalization changes the embedding geometry. MSSA replaces one Hankel trajectory matrix by a stacked Hankel construction
\[
T_{\mathrm{MSSA}}(X)=[\mathbf X^{(1)}:\cdots:\mathbf X^{(s)}],
\]
2D-SSA replaces it by a Hankel-block-Hankel trajectory matrix built from \(L_x\times L_y\) moving windows, and Shaped 2D-SSA generalizes both the data support and the window support to arbitrary shapes \(\mathfrak N\) and \(\mathfrak L\), producing quasi-Hankel matrices and serving as a unifying implementation basis for MSSA and related extensions [1309.5050]. In this framework, generalized singular-spectrum methods differ chiefly by the structure of \(T(X)\).

A second axis changes the observation space itself. Functional SSA lifts the observations to the Hilbert space \(\mathbb H=\mathcal L^2([0,1])\), embeds them into lagged vectors in \(\mathbb H^L\), and replaces matrix SVD by the operator SVD
\[
\mathcal X=\sum_{i=1}^{\infty}\sqrt{\lambda_i}\,v_i\otimes \psi_i
\]
of a compact trajectory operator \(\mathcal X:\mathbb R^K\to\mathbb H^L\) [1906.05232]. Interval-Valued SSA instead embeds interval series into Hankel matrices of ordered pairs, uses a symbolic covariance matrix with entries
\[
s_{jj'}=\frac{1}{6}\sum_{i=1}^k \left\{ 2aa+ab+ba+2bb \right\},
\]
reconstructs by interval diagonal averaging, and interprets grouped components as interval trendlines, cycles, or noise [2011.03872].

A third axis changes the decomposition operator or introduces prior structure. SSA with projection inserts row and column projectors \(\Pi_{\mathrm{row}}\) and \(\Pi_{\mathrm{col}}\) before residual SVD and is therefore semi-nonparametric; it is especially effective for polynomial trend extraction, particularly linear trends [1507.05286]. Non-Hermitian SSA replaces the Hermitian covariance eigenproblem by the generalized pencil
\[
\Gamma_1 v=\hat\lambda\,\Gamma_0 v,
\]
and produces transformed coordinates
\[
Z_{k,j}=\lambda_j^k c_j + \epsilon \eta_{k,j},
\]
so that each retained coordinate is approximately a single exponential rather than an orthogonal variance-maximizing mixture [2008.03413]. A forecasting-oriented extension, called General SSA, leaves the decomposition stage unchanged but replaces the fixed linear recurrent formula by the state-dependent recurrence
\[
y_t=\phi_1(Y_{t-1})y_{t-1}+\cdots+\phi_{L-1}(Y_{t-1})y_{t-(L-1)},
\]
to address structural breaks [1605.02188].

## 4. Multidimensional filtering and image-domain GSS

The image-domain literature gives one of the clearest operational realizations of generalized singular spectrum. For an image
\[
A=\{a_{ij}\}, \qquad 0\le i\le M-1,\quad 0\le j\le N-1,
\]
an \(m\times n\) moving window with \(K=mn\) generates a trajectory matrix by sliding the window over the lattice. The eigenvectors of the lag-covariance matrix
\[
C=\frac{X^TX}{L-K+1}
\]
are arranged into adaptive 2D filters \(F^{(k)}\), and the decomposition is implemented as a two-step filtering system: forward filtering \(a_{ij}\mapsto b_{ij}^{(k)}\), reverse filtering by the point-symmetric filter \(b_{ij}^{(k)}\mapsto d_{ij}^{(k)}\), and exact synthesis
\[
a_{ij}=\frac{1}{K}\sum_{k=1}^K d_{ij}^{(k)}.
\]
In the Fourier domain, each component contributes a nonnegative zero-phase weighting, and completeness becomes
\[
\frac{1}{K}\sum_{k=1}^K |\hat V_{\alpha\beta}^{(k)}|^2=1
\]
at every frequency [1505.01599].

For periodic images and rectangular windows, the lag-covariance matrix is bisymmetric:
\[
C_{i,j}=C_{K-i+1,K-j+1}.
\]
Hence it commutes with the exchange matrix \(J\), and nondegenerate eigenvectors satisfy
\[
Jv^{(k)}=\pm v^{(k)}.
\]
For square windows this yields centrosymmetric or skew-centrosymmetric filters. In a \(3\times 3\) Taylor expansion, symmetric filters cancel odd-order terms and behave like even-order differential operators, while antisymmetric filters cancel even-order terms and behave like odd-order differential operators. The dominant filter is therefore a smoother, intermediate filters act as directional edge enhancers or second-derivative detail operators, and small-eigenvalue filters behave as high-pass or noise filters. In the noisy \(256\times256\) Lenna example with an \(11\times11\) window, the RMS error
\[
D(\ell)=\sqrt{ \frac{1}{MN} \sum_{i,j} \left( a_{ij} - \frac{1}{K}\sum_{k=1}^{\ell} d^{(k)}_{ij} \right)^2 }
\]
decreases until \(\ell=37\) and then increases, indicating that components \(k>37\) mainly reintroduce noise. This image-domain interpretation is one of the strongest arguments for viewing generalized SSA as an adaptive filter-bank theory rather than only as matrix factorization [1505.01599].

## 5. GSS as a non-Hermitian band structure

A 2025 work introduces the generalised singular spectrum as a specific complex band variable for driven-dissipative non-Hermitian systems. Starting from the Bloch dynamical matrix
\[
H(k)=\sum_j \sigma_j(k)\ket{u_j(k)}\bra{v_j(k)},
\]
it defines
\[
\phi_j(k)=\Arg\braket{u_j(k)}{v_j(k)},
\qquad
\varepsilon_j(k)=\sigma_j(k)e^{-i\phi_j(k)}.
\]
The motivation is physical rather than purely algebraic: in such systems the scattering response
\[
S(\omega)=\mathbb 1-i\sqrt{\gamma}\,(\omega\mathbb 1-H)^{-1}\sqrt{\gamma}
\]
is governed more directly by singular values and singular vectors than by the ordinary complex eigenvalue spectrum, especially in the presence of non-normality and the non-Hermitian skin effect. In the normal-matrix limit, the GSS reduces to the familiar band structure [2509.19433].

Within that framework, the GSS supports both point-gap and line-gap topology. The band-resolved point-gap invariant is
\[
\nu_j=\frac{1}{2\pi i}\int_0^{2\pi} dk \left[\bra{u_j(k)}\partial_k\ket{u_j(k)}-\bra{v_j(k)}\partial_k\ket{v_j(k)}\right],
\]
which is also the winding of the \(j\)-th GSS band around the origin. Under additional particle-hole symmetry of the doubled Hermitian matrix, one can define the \(\mathbb Z_2\)-quantized line-gap invariant
\[
\mathcal W_j=\frac{1}{\pi i}\int_0^{2\pi} dk \left[\bra{u_j(k)}\partial_k\ket{u_j(k)}+\bra{v_j(k)}\partial_k\ket{v_j(k)}\right] \mod 2,
\]
and the gap invariant
\[
\Lambda_\ell = \mathcal W + \sum_{j=1}^{\ell}\mathcal W_j \mod 2.
\]
The associated bulk-boundary correspondences distinguish two types of boundary modes: nonzero \(\nu=\sum_j \nu_j\) yields singular zero modes with left and right singular vectors localized at opposite boundaries and thus directional amplification, whereas nonzero \(\Lambda_\ell\) yields finite-singular-value Hermitian-like edge modes with co-localized singular vectors and boundary-local response. The paper illustrates these two regimes with a 1D non-Hermitian SSH model and with a 2D non-Hermitian BBH-type model showing corner-to-corner amplification [2509.19433].

## 6. Standardization, ambiguities, and related usages

In the SSA literature, “Generalised Singular Spectrum” is not yet a settled proper noun. One survey of SSA extensions explicitly does not provide evidence that “GSS” is an established standard name synonymous with any one modification, and the image-decomposition work that is often read as strongly GSS-like does not formulate itself under that label [1907.02579, 1505.01599]. The term therefore functions more reliably as an umbrella description than as a universally fixed method name.

A further complication is acronym overload. In long-context language modeling, GSS denotes Gated State Spaces rather than Generalised Singular Spectrum [2606.16093]. In numerical linear algebra, GSS denotes generalized shift-splitting in saddle-point preconditioning [1605.05818]. This suggests that any technical use of “GSS” requires an explicit local definition. Within the SSA family itself, the practical bottleneck is often no longer decomposition but grouping: automated identification methods based on low-frequency concentration and on angle regularity of singular-vector pairs have been proposed for SSA and extended to MSSA and 2D-SSA [2302.08993]. A plausible implication is that, as singular-spectrum methods generalize across data types and operator geometries, automated grouping becomes part of the effective definition of a usable GSS pipeline.

Source: https://www.emergentmind.com/topics/generalised-singular-spectrum-gss