betaFeedback

WorksPhysics and Astronomy

A Fast Spectral Formulation of the Multiscale Proper Orthogonal Decomposition

Abstract

Multiscale Proper Orthogonal Decomposition (mPOD) decomposes fluid flows into energy-optimal modes within prescribed frequency bands by combining Proper Orthogonal Decomposition with a multiresolution analysis (MRA). In its classical formulation, mPOD relies on a filter bank of finite impulse response (FIR) filters, enabling lossless reconstruction while mitigating Gibbs oscillations and temporal ringing. However, the smooth transition bands required for this purpose introduce partial spectral overlap between adjacent scales and require, for each band, the solution of an eigenvalue problem spanning the full temporal dimension. This work introduces a fast spectral formulation of the mPOD that substantially reduces the computational cost. The proposed approach replaces time-domain FIR filters with compact spectral masks enforcing strictly disjoint frequency supports, thereby exactly decoupling the problem across scales. This leads to a block-diagonal correlation operator in spectral space, so that each band can be treated independently. The resulting eigenvalue problems reduce to small systems whose size depends on the number of active frequencies per band rather than the full time dimension. The approach is validated on a synthetic dataset highlighting spectral windowing effects and on experimental particle image velocimetry (PIV) data of a cylinder wake at Reynolds number \mbox{Re} \approx 5000. In both cases, the proposed formulation accurately recovers the modal structures and singular values of the classical mPOD while reducing the computational cost by orders of magnitude.

Introduction

Data-driven modal decompositions are key tools for identifying coherent structures and interpretable patterns in fluid flow data, enabling diagnostics and reduced-order modeling [1, 2, 3].

A taxonomy of such methods can be organized along a spectrum between two competing objectives: energy optimality, i.e., the ability to capture the largest fraction of the dataset’s variance with a truncated basis, and spectral purity, i.e., the localization of each mode within a prescribed frequency band. At these two extremes sit POD [4, 5, 6] and DMD [7, 8], respectively. The classical trade-off between the two has motivated the development of hybrid decompositions such as the spectral POD proposed by Sieber et al. [9], the frequency-based formulations of the POD by Towne et al. [10], independently referred to as spectral POD, and the multiscale POD (mPOD) by Mendez et al. [11].

Unlike other hybrid decompositions, the mPOD gives the user direct control over this trade-off by explicitly partitioning the frequency axis into prescribed bands and providing an energy-optimal decomposition within each. This yields modes that are simultaneously energy-ranked and spectrally constrained, enabling consistent separation of multiscale and transient dynamics while retaining the convergence properties of classical POD. This flexibility made the mPOD relevant for a wide range of problems, including turbulent motion analysis [12], controlled cylinder wakes [13], transient dynamics [14], jet wiping and gas–liquid coupling instabilities [15], cavitating orifice flows [16], propeller wakes [17], swirled and stratified flames [18], swirling spray flames near lean blowout [19], flow-induced vibration problems [20], compressor corner separation [21], spray characterization [22], sweeping-jet flow control [23] or flows over low-aspect-ratio plates [24] to mention some examples.

The spectral constraints of mPOD were first defined via wavelet-based multiresolution analysis [25] and later reformulated using filter banks [11, 3], as implemented in the open-source toolbox MODULO [26, 27]. In this approach, scale separation is achieved through time-domain FIR filters with smooth transition bands, minimizing spectral leakage and temporal ringing but precluding strict separation between adjacent bands and requiring a full eigenvalue problem per band, sized by the number of temporal snapshots rather than band content. This cost scales steeply with dataset size and dominates the computation for large problems.

Rapid growth in dataset size has spurred efficient implementations of data decomposition methods, including randomized SVD [28], streaming spectral POD [29], and tensor-accelerated DMD variants [30, 31]. In this context, we propose a fast spectral variant of multiscale Proper Orthogonal Decomposition (mPOD) that greatly reduces computational cost.

Our approach departs from the classical formulation by strictly separating spectral supports. While the idea of tackling the eigenvalue problem in the frequency domain can be traced back to the early spectral formulations of the POD [32, 33], the present compact formulation, restricted to non-zero spectral content, is, to the authors’ knowledge, the first systematic framework for multiscale decomposition of this kind.

The remainder of this paper is organized as follows. Section2 briefly recalls the classical formulation of the mPOD and highlights its computational bottlenecks. Section 3 introduces the proposed spectral formulation, detailing the construction of the compact operators and the resulting fast algorithm, proposed in two variants. Section 4 verifies the approach on both synthetic and experimental datasets and assesses its computational performance compared to the classical mPOD. Finally, concluding remarks and perspectives are provided in Section 5.

The classical Multiscale POD (mPOD)

We denote D∈Rns×nt\mathbf{D} \in\mathbb{R}^{n_s \times n_t} the snapshot matrix collecting ntn_t temporal realizations of a spatial field sampled at nsn_s spatial locations. The mPOD proceeds in three steps: (i) the snapshot matrix is partitioned into scale-specific contributions Dm\mathbf{D}_m using a predefined filter bank, (ii) a POD is performed on each scale independently, and (iii) the resulting modes are assembled into a single global factorization. This yields a decomposition of the form

D=∑m=1nMDm=∑m=1nMΦmΣmΨmT=ΦMΣMΨMT,(1)\mathbf{D} = \sum_{m=1}^{n_M} \mathbf{D}_m = \sum_{m=1}^{n_M} \mathbf{\Phi}_m \mathbf{\Sigma}_m \mathbf{\Psi}_m^{\mathsf{T}} = \mathbf{\Phi}_M \mathbf{\Sigma}_M \mathbf{\Psi}_M^{\mathsf{T}}, \tag*{(1)}

where (⋅)T(\cdot)^{\mathsf{T}} denotes transposition, Ψm∈Rnt×rm\mathbf{\Psi}_m \in\mathbb{R}^{n_t \times r_m} contains the rmr_m temporal modes associated with scale mm, Φm∈Rns×rm\mathbf{\Phi}_m \in\mathbb{R}^{n_s \times r_m} the corresponding spatial modes, and Σm∈Rrm×rm\mathbf{\Sigma}_m \in\mathbb{R}^{r_m \times r_m} is a diagonal matrix of singular values. The temporal basis of the mPOD is obtained by concatenating the temporal structures from all scales,

ΨM=[Ψ1,Ψ2,…,ΨnM]∈Rnt×R,(2)\mathbf{\Psi}_M = [\mathbf{\Psi}_1,\mathbf{\Psi}_2,\ldots,\mathbf{\Psi}_{n_M}] \in\mathbb{R}^{n_t \times R}, \tag*{(2)}

where R=∑m=1nMrmR = \sum_{m=1}^{n_M} r_m. The associated spatial structures ΦM\mathbf{\Phi}_M and amplitudes ΣM\mathbf{\Sigma}_M are then obtained by projection and normalization, as in the standard snapshot POD [5] (see also Mendez [3]).

The partition of the snapshot matrix into scale contributions Dm\mathbf{D}_m defines a multi-resolution analysis of the snapshot matrix (see [34]). Each scale is identified by a temporal filter. In the time domain, the filtering corresponds to a convolution along the temporal direction,

Dm=DHm,\mathbf{D}_m = \mathbf{D}\mathbf{H}_m,

where Hm∈Rnt×nt\mathbf{H}_m \in\mathbb{R}^{n_t \times n_t} collects the impulse responses of the filter associated with scale mm (see [3]).

Equivalently, filtering can be performed in the frequency domain. Denoting by ΨF∈Cnt×nt\mathbf{\Psi}_F \in\mathbb{C}^{n_t \times n_t} the discrete Fourier matrix for the temporal direction, such that the temporal Fourier coefficients of the data read D^=DΨF\widehat{\mathbf{D}} = \mathbf{D}\mathbf{\Psi}_F, with Ψ‾F\overline{\mathbf{\Psi}}_F denoting a complex conjugation of ΨF\mathbf{\Psi}_F, and denoting by H^m\widehat{\mathbf{H}}_m the diagonal matrix collecting the transfer function of the filter associated with scale mm, the scale contribution reads

Dm=D Ψ‾F H^m ΨF.(3)\mathbf{D}_m = \mathbf{D}\,\overline{\mathbf{\Psi}}_F\,\widehat{\mathbf{H}}_m\,\mathbf{\Psi}_F. \tag*{(3)}

In (3) and in the following, matrix multiplications involving ΨF\mathbf{\Psi}_F and Ψ‾F\overline{\mathbf{\Psi}}_F are only formal—in practice, these operations correspond to inverse and forward discrete Fourier transforms implemented using fft and ifft routines.

Assuming perfect spectral separation between filters, that is,

∑m=1nMH^m=I,H^mH^ℓ=0,m≠ℓ,(4)\sum_{m=1}^{n_M} \widehat{\mathbf{H}}_m = \mathbf{I}, \qquad\widehat{\mathbf{H}}_m \widehat{\mathbf{H}}_\ell= 0,\quad m \ne\ell, \tag*{(4)}

each frequency is assigned to a single scale.

Therefore, each matrix Dm\mathbf{D}_m has rank rm≤n(m)r_m \le n^{(m)}, where n(m)n^{(m)} denotes the number of Fourier coefficients retained in the mm-th scale. For real-valued data, Fourier coefficients occur in complex-conjugate pairs, so the number of independent degrees of freedom is generally smaller than n(m)n^{(m)}.

If (4) holds, the temporal modes belonging to different scales are mutually orthogonal and ΨM\mathbf{\Psi}_M forms an orthonormal temporal basis—the factorization (1) is a summation of Singular Value Decompositions (SVDs) of the filtered blocks. Accordingly, the spatial modes Φm\mathbf{\Phi}_m are orthonormal within each scale by construction of the SVD, but are not, in general, orthogonal across different scales.

In practice, preserving the orthogonality of the temporal basis ΨM\mathbf{\Psi}_M requires additional processing. The assumption of perfect spectral separation is not enforced because filters with excessively sharp transfer functions would produce long impulse responses and spurious oscillations in the time domain. For this reason, practical filter banks employ smooth transition bands, which lead to partial spectral overlap between neighboring filters. To restore orthogonality, the traditional mPOD algorithm uses a QR factorization ΨM=QR\mathbf{\Psi}_M = \mathbf{Q}\mathbf{R}, to polish the assembled temporal basis, i.e. replacing ΨM←Q\mathbf{\Psi}_M \leftarrow\mathbf{Q}.

Case nt≪nsn_t \ll n_s – temporal correlation approach

When the number of temporal samples is much smaller than the number of spatial degrees of freedom, it is convenient to construct the decomposition from the temporal correlation matrix K=DTD∈Rnt×nt\mathbf{K} = \mathbf{D}^{\mathsf{T}}\mathbf{D} \in\mathbb{R}^{n_t \times n_t}, where T\mathsf{T} denotes the Hermitian transpose. This leads to eigenvalue problems of size nt×ntn_t \times n_t.

The multiscale decomposition is obtained by neglecting cross terms between scales (which should equal zero in case of perfect spectral separation). That is

K=DTD=(∑m=1nMDm)T(∑m=1nMDm)=∑m=1nMDmTDm+∑ℓ≠mDℓTDm≈∑m=1nMDmTDm=∑m=1nMKm.(5)\begin{aligned} \mathbf{K} = \mathbf{D}^{\mathsf{T}}\mathbf{D} = \left(\sum_{m=1}^{n_M}\mathbf{D}_m\right)^{\mathsf{T}}\left(\sum_{m=1}^{n_M}\mathbf{D}_m\right) \\ &= \sum_{m=1}^{n_M}\mathbf{D}_m^{\mathsf{T}}\mathbf{D}_m + \sum_{\ell\ne m}\mathbf{D}_\ell^{\mathsf{T}}\mathbf{D}_m \approx\sum_{m=1}^{n_M}\mathbf{D}_m^{\mathsf{T}}\mathbf{D}_m = \sum_{m=1}^{n_M}\mathbf{K}_m. \tag*{(5)} \end{aligned}

Introducing (3), one obtains

Km=(DΨ‾FH^mΨF)†(DΨ‾FH^mΨF)=Ψ‾FH^mΨFKFΨF†H^mΨF=Ψ‾FH^mKFH^mΨF=Ψ‾FKF,mΨF,(6)\mathbf{K}_m=\left(\mathbf{D}\overline{\mathbf{\Psi}}_{\mathrm{F}}\widehat{\mathbf{H}}_m\mathbf{\Psi}_{\mathrm{F}}\right)^\dagger\left(\mathbf{D}\overline{\mathbf{\Psi}}_{\mathrm{F}}\widehat{\mathbf{H}}_m\mathbf{\Psi}_{\mathrm{F}}\right) = \overline{\mathbf{\Psi}}_{\mathrm{F}}\widehat{\mathbf{H}}_m\mathbf{\Psi}_{\mathrm{F}}\mathbf{K}_{\mathrm{F}}\mathbf{\Psi}_{\mathrm{F}}^\dagger\widehat{\mathbf{H}}_m\mathbf{\Psi}_{\mathrm{F}} = \overline{\mathbf{\Psi}}_{\mathrm{F}}\widehat{\mathbf{H}}_m\mathbf{K}_{\mathrm{F}}\widehat{\mathbf{H}}_m\mathbf{\Psi}_{\mathrm{F}} = \overline{\mathbf{\Psi}}_{\mathrm{F}}\mathbf{K}_{\mathrm{F},m}\mathbf{\Psi}_{\mathrm{F}}, \tag*{(6)}

assuming real-valued transfer functions (corresponding to noncausal, zero-phase filtering), and having introduced the spectral representation of the temporal correlation matrix

KF=ΨFKΨ‾F,(7)\mathbf{K}_{\mathrm{F}}=\mathbf{\Psi}_{\mathrm{F}}\mathbf{K}\overline{\mathbf{\Psi}}_{\mathrm{F}}, \tag*{(7)}

and its filtered version KF,m=H^mKFH^m\mathbf{K}_{\mathrm{F},m}=\widehat{\mathbf{H}}_m\mathbf{K}_{\mathrm{F}}\widehat{\mathbf{H}}_m.

These spectral representations correspond to applying the discrete Fourier transform along both temporal indices of the correlation matrix (see [11]).

The temporal structures are eigenvectors of the covariance matrix at each scale, that is

KmΨm=ΨmΣm2.(8)\mathbf{K}_m\mathbf{\Psi}_m=\mathbf{\Psi}_m\mathbf{\Sigma}_m^2. \tag*{(8)}

Algorithm 1 summarizes the classic mPOD algorithm proposed in [11]. To set the stage for the proposed spectral implementation, it is worth noticing that the definition in (7) leads to a similarity transform KF,m=ΨFKmΨ‾F\mathbf{K}_{\mathrm{F},m}=\mathbf{\Psi}_{\mathrm{F}}\mathbf{K}_m\overline{\mathbf{\Psi}}_{\mathrm{F}} and hence the eigenvalue problem (8) can be related to the eigenvalue problem for KF,m\mathbf{K}_{\mathrm{F},m} [3]:

KF,mΨ^m=Ψ^mΣm2,(9)\mathbf{K}_{\mathrm{F},m}\widehat{\mathbf{\Psi}}_m=\widehat{\mathbf{\Psi}}_m\mathbf{\Sigma}_m^2, \tag*{(9)}

where Ψ^m=ΨFΨm\widehat{\mathbf{\Psi}}_m=\mathbf{\Psi}_{\mathrm{F}}\mathbf{\Psi}_m is the Fourier representation of the temporal modes. The modes in physical space are recovered via

Ψm=Ψ‾FΨ^m.(10)\mathbf{\Psi}_m=\overline{\mathbf{\Psi}}_{\mathrm{F}}\widehat{\mathbf{\Psi}}_m. \tag*{(10)}

Equations (9) and (10) could be used instead of (8) in line 6 of Algorithm 1. However, since both involve diagonalization of matrices of size nt×ntn_t \times n_t, there are currently no benefits in operating in the spectral domain.

Algorithm 1 Classical mPOD via temporal-correlation
1: Input: data matrix D\mathbf{D}, frequency splitting vector FVF_V
2: Construct the filter bank {Hm}m=1nM\{\mathbf{H}_m\}_{m=1}^{n_M} from FVF_V
3: Compute temporal correlation matrix K=D†D\mathbf{K}=\mathbf{D}^\dagger\mathbf{D}
4: for m=1,…,nMm=1,\ldots,n_M do
5: Compute filtered correlation matrix Km=Hm†KHm\mathbf{K}_m=\mathbf{H}_m^\dagger\mathbf{K}\mathbf{H}_m
6: Solve eigenvalue problem KmΨm=ΨmΣm2\mathbf{K}_m\mathbf{\Psi}_m=\mathbf{\Psi}_m\mathbf{\Sigma}_m^2
7: end for
8: Assemble temporal basis ΨM0=[Ψ1,…,ΨnM]\mathbf{\Psi}_M^0=[\mathbf{\Psi}_1,\ldots,\mathbf{\Psi}_{n_M}]
9: Enforce orthonormality ΨM0=QR,ΨM=Q\mathbf{\Psi}_M^0=\mathbf{Q}\mathbf{R},\qquad\mathbf{\Psi}_M=\mathbf{Q}
10: Compute spatial modes ΦM=DΨMΣM−1\mathbf{\Phi}_M=\mathbf{D}\mathbf{\Psi}_M\mathbf{\Sigma}_M^{-1}

Table 1.

The computational cost is dominated by two operations: the construction of the temporal correlation matrix K=D†D\mathbf{K}=\mathbf{D}^\dagger\mathbf{D}, which requires O(nsnt2)O(n_s n_t^2) operations, and the repeated solution of the eigenvalue problems of size nt×ntn_t \times n_t, which requires O(nt3)O(n_t^3) per scale. Therefore, the overall complexity is

O(nsnt2+nMnt3).(11)O(n_s n_t^2+n_M n_t^3). \tag*{(11)}

Case ns≪ntn_s \ll n_t– spatial correlation approach

When the number of spatial degrees of freedom is much smaller than the number of temporal samples, it is more convenient to construct the decomposition directly from the matrices Dm\mathbf{D}_m in (3) through their singular value decomposition. This can be computed by diagonalizing the spatial correlation matrices Cm=DmDm†∈Rns×ns\mathbf{C}_m=\mathbf{D}_m\mathbf{D}_m^\dagger\in\mathbb{R}^{n_s\times n_s}, which leads to nMn_M eigenvalue problems of size ns×nsn_s \times n_s. An efficient spectral formulation to this approach, not presented in [11], is detailed below.

Introducing (3), one obtains

Cm=(DΨ‾FH^mΨF)(DΨ‾FH^mΨF)†=DΨ‾FH^mΨFΨF†H^m†ΨFD†=DΨ‾FH^m2ΨFD†=D^H^m2D^†.(12)\mathbf{C}_m=\left(\mathbf{D}\overline{\mathbf{\Psi}}_{\mathrm{F}}\widehat{\mathbf{H}}_m\mathbf{\Psi}_{\mathrm{F}}\right)\left(\mathbf{D}\overline{\mathbf{\Psi}}_{\mathrm{F}}\widehat{\mathbf{H}}_m\mathbf{\Psi}_{\mathrm{F}}\right)^\dagger = \mathbf{D}\overline{\mathbf{\Psi}}_{\mathrm{F}}\widehat{\mathbf{H}}_m\mathbf{\Psi}_{\mathrm{F}}\mathbf{\Psi}_{\mathrm{F}}^\dagger\widehat{\mathbf{H}}_m^\dagger\mathbf{\Psi}_{\mathrm{F}}\mathbf{D}^\dagger = \mathbf{D}\overline{\mathbf{\Psi}}_{\mathrm{F}}\widehat{\mathbf{H}}_m^2\mathbf{\Psi}_{\mathrm{F}}\mathbf{D}^\dagger = \widehat{\mathbf{D}}\widehat{\mathbf{H}}_m^2\widehat{\mathbf{D}}^\dagger. \tag*{(12)}

(12) shows that the spatial correlation matrix at scale mm can be assembled directly in the spectral domain, without explicitly forming the filtered signal in the time domain. Once the eigenvalue problem

CmΦm=ΦmΣm2(13)\mathbf{C}_m\mathbf{\Phi}_m=\mathbf{\Phi}_m\mathbf{\Sigma}_m^2 \tag*{(13)}

has been solved, the temporal modes can be recovered from the filtered spectral data through

Ψ^m=D^m†ΦmΣm−1,Ψm=Ψ‾FΨ^m.(14)\widehat{\mathbf{\Psi}}_m=\widehat{\mathbf{D}}_m^\dagger\mathbf{\Phi}_m\mathbf{\Sigma}_m^{-1},\qquad \mathbf{\Psi}_m=\overline{\mathbf{\Psi}}_{\mathrm{F}}\widehat{\mathbf{\Psi}}_m. \tag*{(14)}

Algorithm 2 summarizes the corresponding mPOD implementation.

Algorithm 2 Classical mPOD via spatial correlation
1: Input: data matrix D\mathbf{D}, frequency splitting vector FVF_V
2: Construct the filter bank {Hm}m=1nM\{\mathbf{H}_m\}_{m=1}^{n_M} from FVF_V
3: Compute the Fourier-transformed data matrix D^=DΨ‾F\widehat{\mathbf{D}}=\mathbf{D}\overline{\mathbf{\Psi}}_{\mathrm{F}}
4: for m=1,…,nMm=1,\ldots,n_M do
5: Filter the data and get D^m=D^H^m\widehat{\mathbf{D}}_m=\widehat{\mathbf{D}}\widehat{\mathbf{H}}_m
6: Compute the spatial correlation matrix Cm=D^mD^m†\mathbf{C}_m=\widehat{\mathbf{D}}_m\widehat{\mathbf{D}}_m^\dagger
7: Solve eigenvalue problem CmΦm=ΦmΣm2\mathbf{C}_m\mathbf{\Phi}_m=\mathbf{\Phi}_m\mathbf{\Sigma}_m^2
8: Compute spectra Ψ^m=D^m†ΦmΣm−1\widehat{\mathbf{\Psi}}_m=\widehat{\mathbf{D}}_m^\dagger\mathbf{\Phi}_m\mathbf{\Sigma}_m^{-1}
9: Recover temporal modes Ψm=Ψ‾FΨ^m\mathbf{\Psi}_m=\overline{\mathbf{\Psi}}_{\mathrm{F}}\widehat{\mathbf{\Psi}}_m
10: end for
11: Assemble temporal basis ΨM0=[Ψ1,…,ΨnM]\mathbf{\Psi}_M^0=[\mathbf{\Psi}_1,\ldots,\mathbf{\Psi}_{n_M}]
12: Enforce orthonormality ΨM0=QR,ΨM=Q\mathbf{\Psi}_M^0=\mathbf{Q}\mathbf{R},\qquad\mathbf{\Psi}_M=\mathbf{Q}
13: Compute spatial modes ΦM=DΨMΣM−1\mathbf{\Phi}_M=\mathbf{D}\mathbf{\Psi}_M\mathbf{\Sigma}_M^{-1}

Table 2.

The computational cost of this formulation is dominated by three operations: the Fourier transform of the data matrix, which is O(nsntlog⁡nt)O(n_s n_t\log n_t), the construction of the filtered spatial correlation matrices, which is O(ns2nt)O(n_s^2 n_t) for each frequency band, and the repeated solution of the eigenvalue problems of size ns×nsn_s \times n_s, which is O(nMns3)O(n_M n_s^3). The total complexity of this approach is thus

O(nsntlog⁡nt+nMns2nt+nMns3).(15)O(n_s n_t\log n_t+n_M n_s^2 n_t+n_M n_s^3). \tag*{(15)}

The fast (spectral) mPOD

The classical mPOD employs smooth transfer functions Hm(f)H_m(f) designed to approximate a partition of unity while limiting Gibbs oscillations and temporal ringing. However, the resulting overlap between neighboring bands prevents strictly compact spectral support, so the matrices Dm\mathbf{D}_m and Km\mathbf{K}_m are not exactly low-rank, even if most of their energy is concentrated in a subspace of dimension close to n(m)n^{(m)}.

In the present work, we deliberately adopt the opposite design philosophy in order to unlock major computational gains for large datasets. Instead of constructing time-domain FIR filters, we define compact spectral masks Mm(f)M_m(f) directly in the frequency domain, enforcing strictly disjoint spectral supports while relaxing the partition-of-unity condition.

Let FV={f1,…,fnM−1}F_V=\{f_1,\ldots,f_{n_M-1}\} denote the frequency splitting vector defining the spectral bands. For each band mm, the mask Mm(f)M_m(f) is constructed to be unity within the core of the interval [fm−1,fm][f_{m-1},f_m] and to smoothly taper to zero near the band boundaries. The tapering is implemented through cosine windows of width δm\delta_m placed inside the nominal band, so that Mm(f)=0M_m(f)=0 for ∣f∣<fm−1|f|<f_{m-1} and ∣f∣>fm|f|>f_m, while smoothly reaching unity within the band interior.

Because the transition regions are contained within each band rather than extending across adjacent bands, neighboring masks share exact zero crossings at the splitting frequencies. Therefore, Mℓ(f)Mm(f)=0M_\ell(f)M_m(f)=0, if ℓ≠m\ell\ne m, and the filtered matrices Dm\mathbf{D}_m and Km\mathbf{K}_m contain at most n(m)n^{(m)} active spectral components. Hence, rank⁡(Dm)\operatorname{rank}(\mathbf{D}_m), rank⁡(Km)≤n(m)\operatorname{rank}(\mathbf{K}_m)\le n^{(m)}. This compact-support property enables a reformulation of the mPOD in which each scale can be treated independently on a reduced spectral subspace. The structure of this reduction is made explicit by examining the sparsity patterns of the filtered spectral data matrices D^m\widehat{\mathbf{D}}_m and their associated correlation matrices KF,m\mathbf{K}_{\mathcal{F},m}.

Sparsity patterns and compact structures

The sparsity patterns in D^m\widehat{\mathbf{D}}_m and KF,m\mathbf{K}_{\mathcal{F},m} depend on the position of the passband within the spectrum and fall into three canonical configurations. Since D∈Rns×nt\mathbf{D}\in\mathbb{R}^{n_s\times n_t} and K=D†D∈Rnt×nt\mathbf{K}=\mathbf{D}^{\dagger}\mathbf{D}\in\mathbb{R}^{n_t\times n_t} are real-valued, their Fourier representations are conjugate symmetric with respect to positive and negative frequencies. As a result, the spectral blocks associated with opposite frequency intervals are not independent, but related by complex conjugation (up to a possible permutation of the frequency ordering).

For the first scale, corresponding to the low-frequency contribution, only the coefficients in a neighborhood of zero frequency are retained. In a frequency ordering centred around f=0f=0, the filtered spectral data matrix takes the form

D^1=[0D^L0],(16)\widehat{\mathbf{D}}_1=\begin{bmatrix}0&\widehat{\mathbf{D}}_{\mathcal{L}}&0\end{bmatrix}, \tag*{(16)}

where D^L∈Cns×n(1)\widehat{\mathbf{D}}_{\mathcal{L}}\in\mathbb{C}^{n_s\times n^{(1)}} collects the active low-frequency coefficients and the zero blocks represent frequency intervals where the spectral mask vanishes. The size of these blocks is determined by the bandwidth of the active region, so that each matrix D^m\widehat{\mathbf{D}}_m contains at most n(m)n^{(m)} nonzero columns (up to possible degeneracies in the data) corresponding to the retained frequencies, while the remaining nt−n(m)n_t-n^{(m)} columns are identically zero.

The corresponding filtered spectral correlation matrix is

KF,1=[0000KF,L0000],(17)\mathbf{K}_{\mathcal{F},1}= \begin{bmatrix} 0&0&0\\ 0&\mathbf{K}_{\mathcal{F},\mathcal{L}}&0\\ 0&0&0 \end{bmatrix}, \tag*{(17)}

with KF,L=D^L†D^L\mathbf{K}_{\mathcal{F},\mathcal{L}}=\widehat{\mathbf{D}}_{\mathcal{L}}^{\dagger}\widehat{\mathbf{D}}_{\mathcal{L}} the compact version of the large scale contribution.

For an intermediate scale m=2,…,nM−1m=2,\ldots,n_M-1, corresponding to a band-pass contribution, the retained coefficients occupy two symmetric intervals around the zero frequency. The filtered spectral data matrix takes the form

D^m=[0D^B(−)0D^B(+)0],(18)\widehat{\mathbf{D}}_m= \begin{bmatrix} 0&\widehat{\mathbf{D}}_{\mathcal{B}}^{(-)}&0&\widehat{\mathbf{D}}_{\mathcal{B}}^{(+)}&0 \end{bmatrix}, \tag*{(18)}

where D^B(−)\widehat{\mathbf{D}}_{\mathcal{B}}^{(-)} and D^B(+)\widehat{\mathbf{D}}_{\mathcal{B}}^{(+)} correspond to the negative- and positive-frequency contributions, respectively, and satisfy

D^B(+)=D^B(−)‾(up to column ordering).(19)\widehat{\mathbf{D}}_{\mathcal{B}}^{(+)}=\overline{\widehat{\mathbf{D}}_{\mathcal{B}}^{(-)}}\quad\text{(up to column ordering)}. \tag*{(19)}

The filtered spectral correlation matrix then has the structure

KF,m=[000000KF,B(1)0KF,B(2)0000000KF,B(2)†0KF,B(3)000000],(20)\mathbf{K}_{\mathcal{F},m}= \begin{bmatrix} 0&0&0&0&0\\ 0&\mathbf{K}_{\mathcal{F},\mathcal{B}}^{(1)}&0&\mathbf{K}_{\mathcal{F},\mathcal{B}}^{(2)}&0\\ 0&0&0&0&0\\ 0&\mathbf{K}_{\mathcal{F},\mathcal{B}}^{(2)\dagger}&0&\mathbf{K}_{\mathcal{F},\mathcal{B}}^{(3)}&0\\ 0&0&0&0&0 \end{bmatrix}, \tag*{(20)}

with

KF,B=[KF,B(1)KF,B(2)KF,B(2)†KF,B(3)]=D^B†D^B,(21)\mathbf{K}_{\mathcal{F},\mathcal{B}}= \begin{bmatrix} \mathbf{K}_{\mathcal{F},\mathcal{B}}^{(1)}&\mathbf{K}_{\mathcal{F},\mathcal{B}}^{(2)}\\ \mathbf{K}_{\mathcal{F},\mathcal{B}}^{(2)\dagger}&\mathbf{K}_{\mathcal{F},\mathcal{B}}^{(3)} \end{bmatrix} =\widehat{\mathbf{D}}_{\mathcal{B}}^{\dagger}\widehat{\mathbf{D}}_{\mathcal{B}}, \tag*{(21)}

the compact version. Moreover, because of the conjugate symmetry of the Fourier coefficients of real-valued signals, the diagonal blocks KF,B(1)\mathbf{K}_{\mathcal{F},\mathcal{B}}^{(1)} and KF,B(3)\mathbf{K}_{\mathcal{F},\mathcal{B}}^{(3)} are themselves related by complex conjugation, up to a possible permutation induced by frequency ordering. Finally, for the last scale, corresponding to the high-pass contribution, the retained coefficients occupy the outer portions of the spectrum. The filtered spectral data matrix takes the form

D^nM=[D^H(−)0D^H(+)],(22)\widehat{\mathbf{D}}_{n_M}= \begin{bmatrix} \widehat{\mathbf{D}}_{\mathcal{H}}^{(-)}&0&\widehat{\mathbf{D}}_{\mathcal{H}}^{(+)} \end{bmatrix}, \tag*{(22)}

where D^H(−)\widehat{\mathbf{D}}_{\mathcal{H}}^{(-)} and D^H(+)\widehat{\mathbf{D}}_{\mathcal{H}}^{(+)} denote the negative- and positive-frequency contributions, respectively, and satisfy

D^H(+)=D^H(−)‾(up to column ordering).(23)\widehat{\mathbf{D}}_{\mathcal{H}}^{(+)}=\overline{\widehat{\mathbf{D}}_{\mathcal{H}}^{(-)}}\quad\text{(up to column ordering)}. \tag*{(23)}

The filtered spectral correlation matrix is

KF,nM=[KF,H(1)0KF,H(2)000KF,H(2)†0KF,H(3)],(24)\mathbf{K}_{\mathcal{F},n_M}= \begin{bmatrix} \mathbf{K}_{\mathcal{F},\mathcal{H}}^{(1)}&0&\mathbf{K}_{\mathcal{F},\mathcal{H}}^{(2)}\\ 0&0&0\\ \mathbf{K}_{\mathcal{F},\mathcal{H}}^{(2)\dagger}&0&\mathbf{K}_{\mathcal{F},\mathcal{H}}^{(3)} \end{bmatrix}, \tag*{(24)}

with

KF,H=[KF,H(1)KF,H(2)KF,H(2)†KF,H(3)]=D^H†D^H,(25)\mathbf{K}_{\mathcal{F},\mathcal{H}}= \begin{bmatrix} \mathbf{K}_{\mathcal{F},\mathcal{H}}^{(1)}&\mathbf{K}_{\mathcal{F},\mathcal{H}}^{(2)}\\ \mathbf{K}_{\mathcal{F},\mathcal{H}}^{(2)\dagger}&\mathbf{K}_{\mathcal{F},\mathcal{H}}^{(3)} \end{bmatrix} =\widehat{\mathbf{D}}_{\mathcal{H}}^{\dagger}\widehat{\mathbf{D}}_{\mathcal{H}}, \tag*{(25)}

recalling that the conjugate symmetry of the Fourier coefficients of real-valued signals ensures that the diagonal blocks KF,H(1)\mathbf{K}_{\mathcal{F},\mathcal{H}}^{(1)} and KF,H(3)\mathbf{K}_{\mathcal{F},\mathcal{H}}^{(3)} are related by complex conjugation, up to a permutation induced by frequency ordering.

Fast mPOD algorithms and computational cost

The compact-support structure described above enables the construction of the mPOD through reduced eigenvalue problems. In practice, this requires assembling the bandwise spectral correlation matrices KF,mK_{\mathcal{F},m}, which can be achieved through two alternative implementation routes.

In the first route, one follows the method-of-snapshots philosophy and works from the temporal correlation matrix K=D†D\mathbf{K} = \mathbf{D}^{\dagger}\mathbf{D}. The matrix K\mathbf{K} is transformed into the spectral domain to obtain KFK_{\mathcal{F}}, after which the mask MmM_m is applied to extract the contribution of the mm-th band. This route is natural when nt≪nsn_t \ll n_s, since the correlation matrix is then much smaller than the data matrix.

In the second route, one first transforms the data matrix into the spectral domain, obtaining D^=DΨF\widehat{\mathbf{D}} = \mathbf{D}\mathbf{\Psi}_{\mathcal{F}}, and then applies the mask directly to the Fourier coefficients of the data. The filtered spectral correlation matrix is then assembled as KF,m=D^m†D^mK_{\mathcal{F},m} = \widehat{\mathbf{D}}_m^{\dagger}\widehat{\mathbf{D}}_m. This route avoids the construction of the full matrix KFK_{\mathcal{F}} and becomes advantageous when the spectral bands are narrow, i.e. when n(m)≪ntn^{(m)} \ll n_t, since all operations are restricted to the reduced spectral subspaces associated with the active frequencies.

The two implementations are equivalent in the ideal compact-support setting, because the spectral masks enforce strictly disjoint frequency supports, so that no cross-band interactions are present and the filtering operation commutes with the construction of the correlation matrix. The overall procedure is summarized in Algorithm 3, while the two implementations of ComputeBandMatrix are detailed in Algorithm 4.

Importantly, the eigenvalue problem is not solved on the full matrix KF,m∈Cnt×ntK_{\mathcal{F},m} \in\mathbb{C}^{n_t \times n_t}, but only on its compact nonzero blocks KF,c∈Cn(m)×n(m)K_{\mathcal{F},c} \in\mathbb{C}^{n^{(m)} \times n^{(m)}} introduced in the previous subsection. This reduction is justified by the block-structured results in Appendix Appendix A, which show that the nonzero eigenvalues and eigenvectors of KF,mK_{\mathcal{F},m} are entirely determined by those of its compact block. Therefore, each scale leads to an independent eigenvalue problem whose size is determined solely by the number of active frequencies n(m)n^{(m)}, rather than the full temporal dimension ntn_t.

Algorithm 3 Fast spectral mPOD

  1. Input: data matrix D\mathbf{D}, frequency splitting vector FVF_V

  2. Construct the spectral masks {Mm(f)}m=1nM\{M_m(f)\}_{m=1}^{n_M} from FVF_V.

  3. for m=1,…,nMm = 1,\ldots,n_M do

  4. Compute KF,m=ComputeBandMatrix(D,Mm)K_{\mathcal{F},m} = \mathrm{ComputeBandMatrix}(\mathbf{D},M_m)

  5. Extract compact block KF,cK_{\mathcal{F},c}

  6. Solve reduced eigenvalue problem KF,cΨ^m=Ψ^mΣ^m2K_{\mathcal{F},c}\widehat{\mathbf{\Psi}}_m = \widehat{\mathbf{\Psi}}_m\widehat{\mathbf{\Sigma}}_m^2

  7. Recover temporal modes Ψm=ΨFΨ^m\mathbf{\Psi}_m = \mathbf{\Psi}_{\mathcal{F}}\widehat{\mathbf{\Psi}}_m

  8. end for

  9. Assemble temporal basis ΨM=[Ψ1,…,ΨnM]\mathbf{\Psi}_M = [\mathbf{\Psi}_1,\ldots,\mathbf{\Psi}_{n_M}]

  10. Compute spatial modes ΦM=DΨMΣM−1\mathbf{\Phi}_M = \mathbf{D}\mathbf{\Psi}_M\mathbf{\Sigma}_M^{-1}

The computational cost of the fast mPOD formulation is dominated by the construction of bandwise spectral matrices and by the reduced eigendecompositions. In all cases, extracting compact blocks from KF,mK_{\mathcal{F},m} is negligible compared to the matrix products and eigensolves.

Algorithm 4 ComputeBandMatrix(D,Mm\mathbf{D},M_m)

  1. Input: data matrix D\mathbf{D}, mask MmM_m of scale mm

  2. if nt≪nsn_t \ll n_s then

  3. Compute temporal correlation matrix K=D†D\mathbf{K} = \mathbf{D}^{\dagger}\mathbf{D}

  4. Compute spectral representation KF=ΨF†KΨFK_{\mathcal{F}} = \mathbf{\Psi}_{\mathcal{F}}^{\dagger}\mathbf{K}\mathbf{\Psi}_{\mathcal{F}}

  5. Return KF,m=MmKFMmK_{\mathcal{F},m} = M_m K_{\mathcal{F}} M_m

  6. else

  7. Compute Fourier-transformed data D^=DΨF\widehat{\mathbf{D}} = \mathbf{D}\mathbf{\Psi}_{\mathcal{F}}

  8. Extract filtered spectral data block D^m=D^Mm\widehat{\mathbf{D}}_m = \widehat{\mathbf{D}}M_m

  9. Return KF,m=D^m†D^mK_{\mathcal{F},m} = \widehat{\mathbf{D}}_m^{\dagger}\widehat{\mathbf{D}}_m

  10. end if

For the correlation-based route (lines 4–6 in Algorithm 4), the dominant steps are the construction of the temporal correlation matrix, with cost O(nsnt2)O(n_s n_t^2), the spectral transformation of K\mathbf{K}, with cost O(nt2log⁡nt)O(n_t^2 \log n_t), and the reduced eigendecompositions, with total cost O(∑m=1nM(n(m))3)O\left(\sum_{m=1}^{n_M}(n^{(m)})^3\right). The total complexity is therefore

O(nsnt2+nt2log⁡nt+∑m=1nM(n(m))3).(26)O\left(n_s n_t^2+n_t^2\log n_t+\sum_{m=1}^{n_M}(n^{(m)})^3\right). \tag*{(26)}

For the data-based route (lines 8–10 in Algorithm 4), the Fourier transform of the data matrix costs O(nsntlog⁡nt)O(n_s n_t\log n_t), while the construction of the bandwise matrices KF,mK_{\mathcal{F},m}, which costs O(∑m=1nMns(n(m))2)O\left(\sum_{m=1}^{n_M}n_s(n^{(m)})^2\right). Adding the reduced eigendecompositions gives the total complexity

O(nsntlog⁡nt+∑m=1nMns(n(m))2+∑m=1nM(n(m))3).(27)O\left(n_s n_t\log n_t+\sum_{m=1}^{n_M}n_s(n^{(m)})^2+\sum_{m=1}^{n_M}(n^{(m)})^3\right). \tag*{(27)}

A similar acceleration can be applied to the spatial-correlation formulation based on Cm=DmDm†\mathbf{C}_m = \mathbf{D}_m\mathbf{D}_m^{\dagger}. In this case, the compact-support structure allows one to compute Cm=D^mD^m†\mathbf{C}_m = \widehat{\mathbf{D}}_m\widehat{\mathbf{D}}_m^{\dagger} so that only the n(m)n^{(m)} active spectral coefficients are involved. This reduces the cost of assembling Cm\mathbf{C}_m from O(ns2nt)O(n_s^2 n_t) to O(ns2n(m))O(n_s^2 n^{(m)}) per band. However, the size of the associated eigenvalue problems remains ns×nsn_s \times n_s, so that the overall computational gain is less pronounced than in the temporal-correlation formulation.

Verification and Performance Assessment

This section verifies that the proposed formulation reproduces the results of the classical mPOD while significantly reducing computational cost. Section 4.1 compares the scale-separation obtained with the classical FIR-based filters and the proposed spectral-mask formulation using a synthetic test case. Section 4.2 demonstrates the performance of the method on a canonical fluid-dynamics dataset consisting of the wake of a circular cylinder at Reynolds number Re⁡≈5000\operatorname{Re} \approx5000. Finally, Section 4.3 analyzes the computational scaling of the proposed algorithm and compares it with the classical mPOD implementation.

Synthetic test: FIR filters vs spectral masks

This test compares the mask-based fast mPOD approach to the classical FIR filter-based approach on an artificial dataset. The test has three main objectives: (i) to compare the shapes of the individual masks Mm(f)M_m(f) with the amplitudes of the FIR filter responses ∣H^m(f)∣|\widehat{H}_m(f)|, (ii) to evaluate the reconstruction properties of the masks and filters, i.e. ∑m=1nMMm(f)\sum_{m=1}^{n_M} M_m(f) and ∑m=1nMH^m(f)\sum_{m=1}^{n_M} \widehat{H}_m(f), and (iii) to assess their performance in mitigating the Gibbs phenomenon. For this last objective, three approaches are compared: (i) the FIR filter-based formulation as in Algorithm 1, (ii) the mask-based approach as in Algorithm 3, and (iii) a variant of the mask-based approach with infinitely sharp band boundaries.

The test is based on the mPOD analysis of an artificial dataset with nt=6000n_t = 6000, which was constructed to pose a challenging problem for Fourier-based approaches, thereby promoting the appearance of the Gibbs phenomenon. The temporal evolution shapes used to create the test data are shown in Figure 1, with the sine waves superimposed on the slopes having frequencies of f1/fs=1/60f_1/f_s = 1/60 and f2/fs=1/30f_2/f_s = 1/30, respectively. The mPOD analysis was performed using the frequency splitting vector FV/fs=[1/120,3/120,5/120,10/120]F_V/f_s = [1/120, 3/120, 5/120, 10/120] with the taper region size set to δ/fs=1/300\delta/f_s = 1/300 for all scales and the FIR filter order set to 1001 for all scales.

Figure 1. Temporal functions used to assemble the test dataset. Test mode 1 (ΨT1\Psi_{T1}) is nonzero only in the range temporal indices i∈(0,2000)i \in(0, 2000) with the slope defined as (i−500)/1200(i - 500)/1200 and the superimposed sine wave with f1/fs=1/60f_1/f_s = 1/60. Test mode 2 (ΨT2\Psi_{T2}) is nonzero only in the range temporal indices i∈(4000,6000)i \in(4000, 6000) with the slope defined as (4000−i)/800(4000 - i)/800 and the superimposed sine wave with f2/fs=1/30f_2/f_s = 1/30. The third test mode (ΨT3\Psi_{T3}) is a single wavelet in the interval i∈(2200,3800)i \in(2200, 3800).

As shown in Figure 2, the mask exhibits a taper shape similar to that of the FIR filter amplitude. However, unlike the FIR filters, the taper of the mask is entirely confined within the corresponding frequency band, as described in Section 3. This ensures perfect spectral separation, but at the expense of the reconstruction property, as illustrated in Figure 3.

Figure 2. Mask and the magnitude of the low-pass FIR filter frequency response of the first frequency band.

Figure 3. Reconstruction property of the masks and FIR filters. The drops in the mask-based approach around the frequency block boundaries due to the combination of smooth taper and disjoint spectral supports of the masks are clearly visible.

The impact of this trade-off on the Gibbs phenomenon is illustrated in Figure 4. The mask-based approach exhibits localized oscillations near temporal discontinuities, which are slightly more pronounced than in the FIR-based formulation due to the absence of spectral overlap. However, these oscillations remain significantly smaller than those obtained with sharp spectral truncation, demonstrating that the smooth tapering within each band effectively mitigates spurious ringing

Figure 4. mPOD modes from the decomposition of the artificial dataset to showcase the Gibbs phenomenon. All 5 significant modes together with the relevant details are shown. The figure is structured as follows: (a) Mode 1, (b) Mode 2, (c) Mode 3, (d) Mode 4, and (e) Mode 5. The frequency split of the mPOD allowed for isolation of the sine waves while retaining the temporal localization of those waves.

while preserving strict spectral separation.

A classic experimental benchmark

The experimental dataset used in this study is taken from [35], where the canonical case of a cylinder in crossflow is investigated using time-resolved particle image velocimetry (TR-PIV). This configuration, being one of the standard benchmarks in fluid dynamics, is adopted here to assess the performance of the proposed method. For a complete description of the experimental setup and measurement procedure, the reader is referred to [35].

Measurements were conducted in a wind tunnel with a closed test section of a 250×250 mm2250 \times250\ \mathrm{mm}^2 cross-section. The freestream velocity was Uin=5 m/sU_{in} = 5\ \mathrm{m/s} and the cylinder diameter was d=15 mmd = 15\ \mathrm{mm}, yielding a Reynolds number Re≈5000\mathrm{Re} \approx5000 and a fundamental vortex shedding frequency of f1H≈70 Hzf_{1H} \approx70\ \mathrm{Hz}. The experiments were performed in air at ambient conditions (ν=1.5⋅10−5 m2/s\nu= 1.5 \cdot10^{-5}\ \mathrm{m^2/s}, ρ=1.2 kg/m3\rho= 1.2\ \mathrm{kg/m^3}), with the cylinder oriented transversely to the flow and spanning the entire test section.

The velocity fields were acquired in a plane normal to the cylinder axis using TR-PIV with a sampling frequency of fs=2f_s = 2 kHz. The region of interest has a spatial resolution of nx=63n_x = 63 and ny=36n_y = 36 points in the xx and yy directions, respectively, with two velocity components per point, resulting in a total spatial dimension of ns=4536n_s = 4536. A total of 2 seconds of data were recorded, corresponding to nt=4000n_t = 4000 snapshots.

The mPOD processing was based on the frequency splitting vector FV=[35,105,175,245]FV = [35, 105, 175, 245] Hz, with the last scale (above 245 Hz) being discarded. This places the band boundaries between successive harmonic frequencies. The FIR filter order for the classical mPOD was set to 501 for all scales, while the taper width for the fast mPOD was set to δ=f1H/10=7\delta= f_{1H}/10 = 7 Hz. In all cases, the algorithm was configured to extract the 10 most energetic modes. Figure 5 presents the spatial structures and corresponding spectra of the modes associated with the first, second, and third harmonic frequencies.

mPOD modes and corresponding mode spectra for the first, second, and third harmonic frequencies

Figure 5. mPOD modes corresponding to first (a, b), second (c, d), and third (e, f) harmonic frequency, (a, c, e) magnitude with line integral convolution (LIC) streamlines for visualizing the flow patterns, (b, d, f) corresponding mode spectra

Figure 6 shows the absolute differences between the classical and fast mPOD results, displayed on a logarithmic color scale. Overall, the fast mPOD accurately reproduces the spatial structures obtained with the classical formulation, with relative differences on the order of a few percent. These discrepancies are predominantly confined to regions of low fluctuation intensity, where the flow does not exhibit coherent structures and the modal amplitudes are small.

Absolute differences between FIR implementation and fast mPOD modes

Figure 6. Absolute differences between FIR implementation and fast mPOD modes. Modes correspond to (a) first, (b) second, (c) third harmonic frequency.

In contrast, the dominant coherent structures associated with the vortex shedding are recovered almost identically. This indicates that the proposed formulation preserves the relevant content of the decomposition, while the observed differences mainly affect low-energy components. The remaining discrepancies can be attributed to two closely related factors: (i) the absence of spectral overlap between adjacent bands in the fast formulation, and (ii) the QR orthogonalization step in the classical mPOD, which slightly modifies the temporal modes to enforce orthogonality, with a corresponding propagation to the spatial modes.

Figure 7 compares the convergence behavior of the different decompositions in terms of singular value decay. The results show that mPOD retains the rapid convergence characteristics of POD while enforcing spectral separation across scales. In particular, the decay of the modal amplitudes closely follows that of POD, indicating that the energy-optimality of the decomposition is largely preserved within each band. By contrast, the DFT exhibits a much slower decay, as expected from its purely frequency-based nature, which does not provide an optimal energy ranking of the modes. This results highlights the ability of mPOD to combine spectral localization with efficient energy capture, even in flows with strongly periodic dynamics.

Figure 7. Convergence of different data decomposition methods on the cylinder-in-crossflow dataset.

Computational complexity and scaling

Computational complexity was assessed on synthetic test cases using random data matrices and a uniform frequency partition. The dependence of computation time on the number of temporal snapshots was analyzed for four configurations, obtained by combining two spatial dimensions, ns=3000n_s = 3000 and ns=12000n_s = 12000, with two numbers of frequency bands, nM=5n_M = 5 and nM=10n_M = 10. The bands were uniformly distributed over the frequency range, yielding n(m)≈nt/nMn^{(m)} \approx n_t/n_M for all mm. For each configuration, the number of snapshots was varied in the range nt∈[5⋅102,2⋅103,4⋅103,6⋅103,8⋅103,104,1.2⋅104]n_t \in[5 \cdot10^2, 2 \cdot10^3, 4 \cdot10^3, 6 \cdot10^3, 8 \cdot10^3, 10^4, 1.2 \cdot10^4]. The results shown in Figure 8 correspond to averages over four realizations, computed using the MODULО toolbox (version 2.1.6) on Intel Xeon Gold 3.0 GHz processors using 12 cores.

Figure 8. Computation times for (a) ns=3000n_s = 3000 and (b) ns=12000n_s = 12000. Approaches according to Algorithm 1, correlation-based version of Algorithm 3, and data-based version of Algorithm 3 are compared. In both (a) and (b), it can be seen that the correlation-based approach is good for slender data matrices, while the data-based approach is best suited for short and fat data matrices.

As shown in Figure 8, the fast mPOD achieves speedups of up to two orders of magnitude compared to the classical formulation. Moreover, a distinctive scaling behavior of the fast mPOD is observed: the computational cost decreases as the number of frequency bands nMn_M increases. This behavior follows directly from the reduced eigenvalue problems, whose dimensions scale as n(m)≈nt/nMn^{(m)} \approx n_t/n_M. Substituting this relation into the complexity estimates in equations 26 and 27 yields contributions proportional to 1/nM1/n_M and 1/nM21/n_M^2, respectively.

This trend contrasts with the classical mPOD, whose computational cost increases linearly with nMn_M, as shown in equation 11. Therefore, while increasing the number of scales typically leads to higher costs in the classical formulation, it instead results in cheaper computations in the fast formulation due to the progressive reduction of the spectral subspaces.

Finally, comparing the two implementations of the fast mPOD shows that the correlation-based approach is most efficient for tall matrices (nt≪nsn_t \ll n_s), whereas the data-based approach performs best for short, wide matrices. This behavior is consistent with the theoretical complexity estimates and reflects the different matrix dimensions involved in each formulation.

Conclusion

This work revisits the multiscale Proper Orthogonal Decomposition (mPOD) with a focus on its computational complexity and the identification of its main bottlenecks. In the classical formulation based on FIR filters, the scale correlation matrices concentrate their energy within prescribed frequency bands but are not strictly low-rank, which necessitates solving full-order eigenvalue problems at each scale.

To overcome this limitation, we introduced a fast spectral formulation of the mPOD based on masks with compact spectral support. By enforcing strictly disjoint frequency bands, the proposed approach yields scale-specific correlation matrices that are exactly low-rank and admits a block structure that allows the eigenvalue problems to be solved in reduced spectral subspaces. This results in a significant reduction of the computational cost.

The method was assessed on both synthetic and experimental datasets. The synthetic test case highlighted the trade-off between spectral separation and Gibbs mitigation: while the fast mPOD exhibits slightly stronger oscillations than the FIR-based formulation, it remains far superior to sharp spectral truncation. The experimental test on a cylinder wake at Reynolds Re⁡≈5000\operatorname{Re} \approx5000 demonstrated that the fast mPOD accurately reproduces the spatial structures and spectral content of the classical formulation, with only minor differences confined to lowenergy regions. Finally, the computational complexity study confirmed the theoretical scaling predictions and showed that the proposed method achieves speedups of up to two orders of magnitude for the investigated dataset. Overall, the results demonstrate that the fast mPOD preserves the interpretability and convergence properties of the classical formulation while drastically reducing computational cost. This opens the way to efficient multiscale analysis of large-scale datasets.

Acknowledgments

This work was carried out as part of the Short Training Program (STP) of Marek Belda at the von Karman Institute. Marek Belda acknowledges the institutional support RVO:61388998 and the financial support provided by the Grant Agency of the Czech Technical University in Prague, grant No. SGS25/127/OHK2/3T/12. Martin Isoz acknowledges the financial support provided by the Ministry of Education, Youth, and Sports of the Czech Republic via the project No. CZ.02.01.01/00/23_020/0008501 (METEX), co-funded by the European Union. M. A. Mendez is supported by the European Research Council (ERC, grant agreement No 101165479 RE-TWIST StG). Views and opinions expressed are however those of the authors only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them. L. Schena is supported by Fonds Wetenschappelijk Onderzoek (FWO), Proj. Number 1S67925N.

Results for block-structured matrices

This appendix collects the algebraic results used in Section 3 to justify the reduction of eigenvalue problems associated with block-sparse spectral correlation matrices.

Theorem A1. Let

A=[000000A10A20000000A30A4000000]∈CN×N,A = \begin{bmatrix} 0 & 0 & 0 & 0 & 0 \\ 0 & A_1 & 0 & A_2 & 0 \\ 0 & 0 & 0 & 0 & 0 \\ 0 & A_3 & 0 & A_4 & 0 \\ 0 & 0 & 0 & 0 & 0 \end{bmatrix} \in\mathbb{C}^{N \times N},

where the nonzero blocks do not overlap and are arranged so that

Ac=[A1A2A3A4]∈Cn×nA_c = \begin{bmatrix} A_1 & A_2 \\ A_3 & A_4 \end{bmatrix} \in\mathbb{C}^{n \times n}

is well defined. Then the nonzero eigenvalues of AA coincide with the eigenvalues of AcA_c.

Proof. The characteristic polynomial of AA factorizes as

det⁡(A−λI)=λN−ndet⁡(Ac−λI),(28)\det(A-\lambda I)=\lambda^{N-n}\det(A_c-\lambda I), \tag*{(28)}

obtained by expanding the determinant along the zero-padded rows and columns. Hence, the spectrum of AA consists of the spectrum of AcA_c together with N−nN-n additional zero eigenvalues. □

Theorem A2. Let λ≠0\lambda\ne0 be an eigenvalue of AA in (A.1), with associated eigenvector

v=[v1v2v3v4v5].v = \begin{bmatrix} v_1 \\ v_2 \\ v_3 \\ v_4 \\ v_5 \end{bmatrix}.

Then v1=v3=v5=0v_1=v_3=v_5=0, and the nonzero part of vv is obtained from the eigenvector

vc=[v2v4]v_c = \begin{bmatrix} v_2 \\ v_4 \end{bmatrix}

of AcA_c associated with the same eigenvalue λ\lambda.

Proof. From (A−λI)v=0(A-\lambda I)v=0, the rows corresponding to the zero blocks give

−λv1=0,−λv3=0,−λv5=0.-\lambda v_1=0,\qquad-\lambda v_3=0,\qquad-\lambda v_5=0.

Since λ≠0\lambda\ne0, it follows that

v1=v3=v5=0.v_1=v_3=v_5=0.

The remaining equations reduce to

\begin{bmatrix} A_1-\lambda I & A_2 \\ A_3 & A_4-\lambda I \end{bmatrix} \begin{bmatrix} v_2 \\ v_4 \end{bmatrix} =(A_c-\lambda I)v_c=0, \tag*{(29)}\end{bmatrix}

so vcv_c is an eigenvector of AcA_c, and vv is recovered by zero-padding. □

Theorem A3. Let A,B∈CN×NA,B\in\mathbb{C}^{N\times N} be two block-structured matrices of the form (A.1), with nonzero blocks supported on disjoint sets of rows and columns. Define

C=A+B.C=A+B.

Then:

  1. every nonzero eigenpair of AA or BB is also an eigenpair of CC;

  2. the nonzero spectrum of CC is the union of the nonzero spectra of AA and BB.

Proof. Let (λ,v)(\lambda,v) be a nonzero eigenpair of AA. By Theorem A2, the support of vv is confined to the rows and columns occupied by the nonzero blocks of AA. Since BB is supported on a disjoint set of rows and columns, one has Bv=0Bv=0. Therefore

Cv=(A+B)v=Av=λv.Cv=(A+B)v=Av=\lambda v.

The same argument holds for eigenpairs of BB. Since the supports are disjoint, the nonzero invariant subspaces do not interact, and the nonzero spectrum of CC is exactly the union of the nonzero spectra of AA and BB. □

References

  1. [1]K. Taira, S. L. Brunton, S. T. M. Dawson, C. W. Rowley, T. Colonius, B. J. McKeon, O. T. Schmidt, S. Gordeyev, V. Theofilis, L. S. Ukeiley, Modal analysis of fluid flows: An overview, AIAA Journal 55 (2017) 4013–4041. doi:10.2514/1.j056060.DOI
  2. [2]B. Begiashvili, N. Groun, J. Garicano-Mena, S. Le Clainche, E. Valero, Data-driven modal decomposition methods as feature detection techniques for flow problems: A critical assessment, Physics of Fluids 35 (2023). doi:10.1063/5.0142102.DOI
  3. [3]M. Mendez, Generalized and Multiscale Modal Analysis, Cambridge University Press, 2023, p. 153–181.arxiv.org/abs/2208.12630
  4. [4]J. L. Lumley, Stochastic Tools in Turbulence, DOVER PUBN INC, 1970.
  5. [5]L. Sirovich, Turbulence and the dynamics of coherent structures, part III: Dynamics and scaling, Quarterly of Applied Mathematics 45 (1987) 583–590.DOI
  6. [6]S. Dawson, The Proper Orthogonal Decomposition, Cambridge University Press, 2023, p. 117–132.DOI
  7. [7]P. J. Schmid, Dynamic mode decomposition of numerical and experimental data, Journal of Fluid Mechanics 656 (2010) 5–28. doi:10.1017/S00221120100001217.
  8. [8]J. H. Tu, C. W. Rowley, D. M. Luchtenburg, S. L. Brunton, J. N. Kutz, On dynamic mode decomposition: theory and applications, Journal of Computational Dynamics 1 (2014) 391–421. doi:10.3934/jcd.2014.1.391.DOI
  9. [9]M. Sieber, C. O. Paschereit, K. Oberleithner, Spectral proper orthogonal decomposition, Journal of Fluid Mechanics 792 (2016) 798–828. doi:10.1017/jfm.2016.103.DOI
  10. [10]A. Towne, O. T. Schmidt, T. Colonius, Spectral proper orthogonal decomposition and its relationship to dynamic mode decomposition and resolvent analysis, Journal of Fluid Mechanics 847 (2018) 821–867. doi:10.1017/jfm.2018.283.DOI
  11. [11]M. A. Mendez, M. Balabane, J.-M. Buchlin, Multi-scale proper orthogonal decomposition of complex fluid flows, Journal of Fluid Mechanics 870 (2019) 988–1036. doi:10.1017/jfm.2019.212.DOI
  12. [12]C. Chi, D. Thévenin, A. J. Smits, S. Wolligandt, H. Theisel, Identification and analysis of very-large-scale turbulent motions using multiscale proper orthogonal decomposition, Phys. Rev. Fluids 7 (2022) 084603. doi:10.1103/PhysRevFluids.7.084603.DOI
  13. [13]Y. Zhong, H. Ma, J. Guo, Multi-scale proper orthogonal decomposition (mpod) analysis of vortex evolution and viscous dissipation in a circular-cylinder wake controlled by parallel symmetric jets, Ocean Engineering 289 (2023) 116280. doi:10.1016/j.oceaneng.2023.116280.DOI
  14. [14]M. A. Mendez, D. Hess, B. B. Watz, J.-M. Buchlin, Multiscale proper orthogonal decomposition (mpod) of tr-piv data—a case study on stationary and transient cylinder wake flows, Measurement Science and Technology 31 (2020) 094014. doi:10.1088/1361-6501/ab82be.DOI
  15. [15]D. Barreiro-Villaverde, A. Gosset, M. Lema, M. A. Mendez, On the coupling instability of a gas jet impinging on a liquid film, Journal of Fluid Mechanics 992 (2024) A11. URL: https://www.cambridge.org/core/product/identifier/S0022112024005536/type/journal_article. doi:10.1017/jfm.2024.553.
  16. [16]C. Esposito, M. Mendez, J. Steelant, M. Vetrano, Spectral and modal analysis of a cavitating flow through an orifice, Experimental Thermal and Fluid Science 121 (2021) 110251. doi:10.1016/j.expthermflusci.2020.110251.DOI
  17. [17]H. Huang, M. Li, L. Wang, Study on the propeller wake dynamics using multiscale proper orthogonal decomposition, Physics of Fluids 37 (2025). doi:10.1063/5.0243607.DOI
  18. [18]A. Procacci, M. M. Kamal, M. A. Mendez, S. Hochgreb, A. Coussemient, A. Parente, Multi-scale proper orthogonal decomposition analysis of instabilities in swirled and stratified flames, Physics of Fluids 34 (2022). doi:10.1063/5.0127956.
  19. [19]M. G. De Giorgi, P. Di Gloria, D. Fontanarosa, A. Ficarella, Advanced multiscale model and frequency analysis of swirling spray flame near to lean blowout, Case Studies in Thermal Engineering 60 (2024) 104651. doi:10.1016/j.csite.2024.104651.DOI
  20. [20]M. J. Janocha, M. C. Ong, G. Yin, Large eddy simulations and modal decomposition analysis of flow past a cylinder subject to flow-induced vibration, Physics of Fluids 34 (2022). doi:10.1063/5.0084966.DOI
  21. [21]Y. Liu, W. Zhong, Y. Tang, Impact of dominant flow structures on turbulence energy in compressor corner separation, AIAA Journal 63 (2025) 5355–5370. doi:10.2514/1.J065420.DOI
  22. [22]R. Zhou, C. Shen, Experimental study on the spray characteristics of a pintle injector element, Acta Astronautica 194 (2022) 255–262. doi:10.1016/j.actaastro.2022.02.019.DOI
  23. [23]P. Wang, J. Song, J. Cai, Y. Liu, X. Wen, Transient evolution of flow structures induced by sweeping jets on a circular cylinder via multi-scale proper orthogonal decomposition (mpod), Aerospace Science and Technology 168 (2026) 111341. doi:10.1016/j.ast.2025.111341.
  24. [24]Y. Zhu, J. Liu, J. Wang, Reversal of nonlinear energy transfer direction in flows over low-aspect-ratio plates, Journal of Fluid Mechanics 1010 (2025). doi:10.1017/jfm.2025.228.DOI
  25. [25]M. Mendez, M. Scelzo, J.-M. Buchlin, Multiscale modal analysis of an oscillating impinging gas jet, Experimental Thermal and Fluid Science 91 (2018) 256–276. doi:10.1016/j.expthermflusci.2017.10.032.DOI
  26. [26]D. Ninni, M. A. Mendez, Modulo: A software for multiscale proper orthogonal decomposition of data, SoftwareX 12 (2020) 100622. doi:10.1016/j.softx.2020.100622.DOI
  27. [27]R. Poletti, L. Schena, D. Ninni, M. A. Mendez, Modulo: A python toolbox for data-driven modal decomposition, Journal of Open Source Software 9 (2024) 6753. doi:10.21105/joss.06753.DOI
  28. [28]D. Palitta, S. Portaro, Row-aware randomized svd with applications, 2025. URL: https://arxiv.org/abs/2408.04503. doi:10.48550/arXiv.2408.04503. arXiv:2408.04503.
  29. [29]O. T. Schmidt, A. Towne, An efficient streaming algorithm for spectral proper orthogonal decomposition, Computer Physics Communications 237 (2019) 98–109. doi:10.1016/j.cpc.2018.11.009.DOI
  30. [30]K. Li, S. Utyuzhnikov, Tensor train-based higher-order dynamic mode decomposition for dynamical systems, Mathematics (2023). doi:10.3390/math11081809.DOI
  31. [31]Z. He, M. Hu, Y. Lou, C. Chen, Tensor dynamic mode decomposition, 2025. URL: https://arxiv.org/abs/2508.02627. doi:10.48550/arXiv.2508.02627. arXiv:2508.02627.
  32. [32]W. K. George, Insight into the dynamics of coherent structures from a proper orthogonal decomposition, in: Proceedings of the International Centre for Heat and Mass Transfer, 1988, p. 469 – 487.
  33. [33]M. N. Glauser, S. J. Leib, W. K. George, Coherent structures in the axisymmetric turbulent jet mixing layer, in: F. Durst, B. E. Launder, J. L. Lumley, F. W. Schmidt, J. H. Whitelaw (Eds.), Turbulent Shear Flows 5, Springer Berlin Heidelberg, Berlin, Heidelberg, 1987, pp. 134–145.DOI
  34. [34]G. Strang, T. Nguyen, Wavelets and filter banks, rev. ed. ed., Wellesley-Cambridge Press, Wellesley, Mass., 1997. Literaturverz. S. [475] - 486.
  35. [35]T. Hlavatý, M. Isoz, M. Belda, V. Uruba, P. Procházka, Is the proper orthogonal decomposition suitable to validate simulation of turbulent wake?, Journal of Wind Engineering and Industrial Aerodynamics 255 (2024) 105953.DOI

Paper details

vlnc:W_jtfrd3qq
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents