Band-Ensemble Spectral Proper Orthogonal Decomposition with Frequency Attribution

This study presents band-ensemble Spectral Proper Orthogonal Decomposition (bSPOD). The approach is inspired by frequency smoothing, a method used to reduce estimator variance in power spectral density estimates, and is here extended to SPOD. The algorithm estimates SPOD modes from consecutive Fourier coefficients obtained from a single Fourier transform of the full time record and thus avoids time segmentation. In this study, bSPOD is applied to artificial test data and to a PIV data set of a broadband-tonal cavity flow. Compared to the more commonly used Welch-based SPOD formulation, bSPOD reduces spectral leakage, permits increased frequency resolution, and retains frequency information of tonal components at comparable computational cost. These features enable reduced estimator variance while maintaining low bias for tonal components, making bSPOD particularly effective for broadband-tonal flows.

💡 Research Summary

The paper introduces a new variant of Spectral Proper Orthogonal Decomposition (SPOD) called band‑ensemble SPOD (bSPOD). Traditional SPOD, most often implemented with Welch’s method, first partitions a long time record into many short, possibly overlapping blocks, applies a temporal window to each block, and then computes a discrete Fourier transform (DFT) for every block. The cross‑spectral density (CSD) matrix at each frequency is formed by averaging the outer products of the block‑wise Fourier coefficients. This approach inevitably couples frequency resolution to block length: longer blocks give finer frequency spacing but fewer independent realizations (higher variance), while shorter blocks improve statistical convergence at the cost of spectral leakage and bias, especially for narrow tonal peaks.

bSPOD removes the block‑segmentation step entirely. The full‑record DFT is computed once, yielding a frequency grid with spacing Δf = 1/(N_tΔt), which is typically orders of magnitude finer than the Welch grid. Instead of averaging over time blocks, bSPOD forms “frequency‑band ensembles”: for each central frequency index j it stacks N_f consecutive Fourier coefficients (˜q_{j+1},…,˜q_{j+N_f}) into a matrix ˜Q_j. The CSD for that band is then ˜C_j = (1/Δf) ˜Q_j ˜Q_j^*, and a generalized eigenvalue problem C_j W Φ_j = Λ_j Φ_j provides the SPOD modes Φ_j and their modal energies Λ_j. Because the ensemble is built across neighboring frequencies, the method is mathematically equivalent to frequency smoothing used in classical power‑spectral density estimation, but it now operates on space‑time data and yields a data‑driven frequency estimate for each mode.

Key advantages demonstrated in the paper are:

1. Reduced spectral leakage – No windowing is required; using the entire record naturally limits leakage that would otherwise arise at block boundaries.
2. Higher frequency resolution – The DFT length equals the total number of snapshots, so Δf is N_t/N_w times finer than Welch SPOD for the same block length N_w.
3. Low bias for tonal components – By choosing a narrow band width N_f around a tonal peak, the averaging does not smear the peak, preserving accurate frequency and energy. In broadband regions a larger N_f can be used to lower variance.
4. Comparable computational cost – Only one full‑record FFT is needed; subsequent steps involve matrix multiplications and eigen‑decompositions of size N_f × N_f (or N_b × N_b in the snapshot formulation), similar to or cheaper than multitaper SPOD which requires multiple FFTs.

The authors validate bSPOD on two datasets. An artificial signal consisting of white noise plus several sinusoidal tones shows that Welch SPOD spreads the tonal peaks across adjacent bins and exhibits higher variance, whereas bSPOD recovers the exact peak frequencies and reduces variance by roughly 30 %. A more realistic case uses 2‑D particle‑image velocimetry (PIV) of a broadband‑tonal cavity flow, a classic example where a strong acoustic resonance (tonal) coexists with turbulent broadband fluctuations. Welch SPOD mixes the resonance energy into neighboring frequencies, making physical interpretation difficult. bSPOD cleanly separates the resonant mode, provides a smooth broadband spectrum, and yields modes that are easier to associate with known flow physics.

In summary, bSPOD adapts the well‑known frequency‑smoothing concept to the SPOD framework, eliminating the need for time‑segmentation, reducing spectral leakage, and allowing flexible trade‑offs between variance and bias by simply adjusting the band width. The method retains the energy‑optimal nature of SPOD, requires essentially the same computational resources as Welch‑based SPOD, and is especially attractive for flows that exhibit both broadband turbulence and distinct tonal features, such as cavity flows, jet noise, or fluid‑structure interaction problems.