Field Guide · algorithm

Also known as: ESPRIT, Estimation of Signal Parameters via Rotational Invariance Techniques

ESPRIT (Estimation of Signal Parameters via Rotational Invariance Techniques) is a subspace direction-of-arrival estimator that, like MUSIC, separates a signal subspace from noise — but instead of scanning a spectrum for peaks it solves for the arrival angles algebraically by exploiting a shift structure built into the array.1 Two identical subarrays separated by a fixed displacement see the same sources with only a phase rotation between them, and that rotation encodes the angles directly.

subarray 1 subarray 2 (shifted Δ) signalsubspaces rotation Ψeig(Ψ) → θ eigenvalue phases = arrival angles — no grid search
ESPRIT pairs two displacement-shifted subarrays; the rotation matrix relating their signal subspaces has eigenvalues whose phases are the arrival angles, found in closed form.

How it works

Split the array into two subarrays that are identical up to a known translation Δ. Both observe the same signal subspace Eₛ (from an eigendecomposition of the array covariance), but the second subarray’s copy is the first multiplied by a diagonal matrix of phase factors e^{jωΔsinθ} — one per source. That means the two subarray slices of Eₛ are linked by a single invertible rotation matrix Ψ:

  • Estimate the covariance and its signal subspace Eₛ.
  • Partition Eₛ into the two subarray rows and solve (typically total least squares) for Ψ such that Eₛ₁ Ψ ≈ Eₛ₂.
  • Take the eigenvalues of Ψ. They lie on the unit circle, and each one’s phase maps straight to an arrival angle (or, in the temporal version, a frequency).

There is no angle grid and no peak search — the parameters drop out of an eigenvalue computation, which is both faster and free of the resolution-versus-grid-spacing trade-off that a scanned spectrum has.

Contrast with MUSIC

MUSIC and ESPRIT reach the same signal/noise subspace split; they differ in the last step. MUSIC searches a pseudo-spectrum over all candidate angles, needs an accurately calibrated steering vector a(θ) for every angle, and pays for a fine grid. ESPRIT computes the angles from the shift structure, so it needs no calibrated array manifold and no search, making it cheaper and more robust to calibration error — at the cost of requiring the array to have that specific translational (doublet) geometry. MUSIC works with arbitrary array shapes; ESPRIT trades that flexibility for speed and closed-form estimates.

Relevance to SDR

ESPRIT is used for direction finding, radar and channel-parameter estimation, and its temporal form gives high-resolution frequency and delay estimates in channel sounders and OFDM channel estimation. Like MUSIC it assumes a coherent multi-element antenna array with per-element sampling — hardware a single-front-end receiver like GopherTrunk does not have, so GT performs no DOA estimation. It appears here as the search-free counterpart to MUSIC in the array-processing toolkit of the wider RF world.

Sources

  1. ESPRIT — Wikipedia, on rotational-invariance subspace DOA estimation without a spectral search. 

See also