Hankel Matrix

A Hankel matrix is a matrix with constant anti-diagonals — every entry depends only on the sum of its row and column indices:

The structure encodes a single sequence — each value appears once along an anti-diagonal, and the whole matrix is parameterized by that sequence.

The construction

Given a length- signal and a window length , the trajectory matrix is an Hankel matrix (with ) whose columns are successive windows of the signal:

This is the data structure that bridges 1D time-series analysis and 2D linear algebra. Once you have a matrix, you can apply SVD, PCA, factorization, low-rank approximation — all the tools that need a matrix input.

Why this matters in ML and signal processing

Hankel matrices show up wherever someone wants to apply matrix methods to a sequence:

• Singular Spectrum Analysis (SSA) — builds the trajectory Hankel matrix, runs SVD, separates trend / oscillations / noise from the singular value spectrum, then reconstructs each component via diagonal averaging. The Hankel structure is what makes the reconstruction step well-defined: you average each anti-diagonal back into a 1D signal.
• System identification — the Ho-Kalman algorithm recovers a linear state-space model from impulse response data by SVD on a Hankel matrix of measurements. Same trick: 1D → 2D via Hankel, then linear algebra.
• Matrix completion for time series — recommend missing values in a Hankel structure rather than an arbitrary matrix, exploiting the shared-entry constraint as an inductive bias.

The rank structure

The most useful property of a Hankel matrix built from a signal is that rank reveals structure:

• A constant signal → rank 1.
• A single pure sinusoid → rank 2.
• A sum of pure sinusoids → rank (each gives two singular values, for cosine + sine basis).
• A linear trend → rank 2 (constant + linear basis).
• White noise → full rank, , with singular values that spread roughly uniformly.

This is why looking at the singular value spectrum of a trajectory Hankel matrix separates signal from noise. Large singular values capture low-rank deterministic components (trends, oscillations); the long tail of small singular values is the noise. Choosing a rank- truncation amounts to picking which structural components to keep — see SSA for the full machinery.

Diagonal averaging — going back to a signal

After truncating the SVD to components, the result is a rank- matrix that’s no longer exactly Hankel — different positions on the same anti-diagonal will hold slightly different values. To recover a 1D signal, you average each anti-diagonal:

where is the -th anti-diagonal. This is the projection of a general matrix back onto the Hankel manifold; it’s also the closed-form minimum-Frobenius-norm projection if you want a clean mathematical justification.

The full Hankel → SVD → truncate → diagonal-average pipeline is the spine of every SSA-style decomposition. Hankel structure isn’t decorative — it’s the structural assumption that makes the whole approach work.