Frequency-time analysis, low-rank reconstruction and denoising of turbulent flows using SPOD

Four different applications of spectral proper orthogonal decomposition (SPOD): low-rank reconstruction, denoising, frequency-time analysis, and prewhitening are demonstrated on large-eddy simulation data of a turbulent jet. SPOD-based low-rank reconstruction can be performed by direct inversion of a truncated SPOD. This spectral inversion problem, however, is ambiguous since SPOD relies on spectral estimation. We demonstrate SPOD-based flow field reconstruction using direct inversion of the SPOD algorithm (frequency-domain approach) and propose an alternative approach based on projection of the time series data onto the modes (time-domain approach). While the SPOD optimally represents the flow in a statistical sense, the time-domain approach seeks an optimal reconstruction of each instantaneous flow field. We further propose a SPOD-based denoising strategy that is based on hard-thresholding of the SPOD eigenvalues. The proposed strategy achieves significant noise reduction while facilitating drastic data compression. In contrast to standard methods of frequency-time analysis such as wavelet transform, a proposed SPOD-based approach yields a spectrogram that characterizes the temporal evolution of spatially coherent flow structures. In the frequency-domain, time-varying expansion coefficients can be obtained by basing the SPOD on a sliding window. This approach, however, is computationally intractable, and an alternative strategy based on convolution in the time-domain is presented. When applied to the turbulent jet data, SPOD-based frequency-time analysis reveals that the intermittent occurrence of large-scale coherent structures is directly associated with high-energy events. This work suggests that the time-domain approach is preferable for low-rank reconstruction of individual snapshots, and the frequency-domain approach for denoising and frequency-time analysis.


Introduction
The curse of dimensionality (see, e.g., Meneveau et al. 1992) in the analysis of large turbulent flow data has led to the development of a number of modal decomposition techniques (Holmes et al. 2012). The primary utilities of these techniques are to extract the essential flow features and to provide a low-dimensional representation of the data. Most of these techniques seek modes that lie in the span of the snapshots that constitute the time-resolved data, and adhere to certain mathematical properties that define the decomposition. The arguably most widely used technique is the proper orthogonal decomposition (POD), introduced by Lumley (1967Lumley ( , 1970. A specific flavor of POD, the computationally inexpensive method of snapshots (Sirovich 1987;Aubry 1991), decompose the flow field into spatial modes and temporal coefficients. Its modes optimally represent the data in terms of its variance, or energy, and are coherent in space and at zero time lag. Another popular method is the dynamic mode decomposition (DMD, Schmid 2010), which is rooted in Koopman theory (Rowley et al. 2009) and assumes an evolution operator that maps the flow field from one snapshot to its next. The DMD modes are characterized by a single frequency and linear amplification rate. Refer to the reviews by Taira et al. (2017) and  for summaries of various modal techniques.
Spectral proper orthogonal decomposition (SPOD) is the frequency-domain variant of POD and computes modes as estimates of the eigenvectors of the cross-spectral density (CSD) matrix. At each frequency, SPOD yields a set of orthogonal modes, ranked by energy. The mathematical framework underlying SPOD was first outlined by Lumley (1967Lumley ( , 1970. Early implementations of SPOD include Glauser et al. (1987); Glauser & George (1992); Delville (1994); Arndt et al. (1997); Picard & Delville (2000); Citriniti & George (2000) and Gordeyev & Thomas (2000). For statistically stationary flows, Towne et al. (2018) have demonstrated that the SPOD combines the advantages of POD, namely optimality and orthogonality, and DMD, namely temporal monochromaticity. In this work, we demonstrate how these properties can be leveraged for different applications.
In the past, SPOD has been used to analyze a number of turbulent flows, including jets (Arndt et al. 1997;Gordeyev & Thomas 2000, 2002Gamard et al. 2002Gamard et al. , 2004Jung et al. 2004;Iqbal & Thomas 2007;Tinney et al. 2008a,b;Schmidt et al. 2018;Pickering et al. 2019;Nekkanti & Schmidt 2020), the wake behind a disk (Johansson et al. 2002;Johansson & George 2006a,b;Tutkun et al. 2008;Ghate et al. 2020;Nidhan et al. 2020), pipe flows (Hellström & Smits 2014;Hellström et al. 2015Hellström et al. , 2016Hellström & Smits 2017), and channel flows (Muralidhar et al. 2019). Several studies have shown that a significant amount of energy is captured by the first few modes at each frequency. For jets and disk wakes, Glauser et al. (1987), Citriniti & George (2000), Jung et al. (2004), Johansson & George (2006b), Tinney et al. (2008b) have shown that the leading mode and first three modes capture at least 40%, and 80% of the total energy, respectively. Partial reconstructions of the flow field from SPOD were previously shown by Citriniti & George (2000), Tinney et al. (2008b), and more recently, by Ghate et al. (2020). We demonstrate low-rank reconstruction using two approaches. One is by inverting the SPOD algorithm, which was previously employed by Citriniti & George (2000) in a similar manner, and for which we present an alternative means of computation based on convolution in the time domain. The other one is by taking an oblique projection of the data on the SPOD modes. The advantages and disadvantages of both approaches for different applications are discussed.
The first of these applications is denoising. Most experimental flow field data exhibit measurement noise that hampers the physical analysis. The computation of spatial derivatives required for quantities like the vorticity or strain rate, for example, leads to particularly large errors. Another difficulty is that physically relevant small-scale structures may be concealed by the noise. The most wide-spread experimental technique for multi-dimensional flow field measurement is particle image velocimetry (PIV). Common techniques to remove noise from PIV data include spatial filtering (Discetti et al. 2013), temporal filtering based on Fourier truncation, and POD-based techniques (Raiola et al. 2015;. Spatial filters are typically based on Gaussian smoothing (Discetti et al. 2013). Different temporal filters such as median filters (Son & Kihm 2001), Hampel filters (Fore et al. 2005), Wiener filters (Vétel et al. 2011), and band-pass filters (Sciacchitano & Scarano 2014) have been used for denoising PIV data. A comparison of different spatial and temporal filters is presented, for example, in Vétel et al. (2011). As an alternative to these standard techniques, POD reconstruction has been used as a means of denoising through mode truncation. Low-dimensional reconstructions from standard POD have been applied for this purpose to flows past a backward-facing step (Kostas et al. 2005), arterial flows (Charonko et al. 2010;, turbulent wakes (Raiola et al. 2015), and vortex rings (Stewart & Vlachos 2012;. In this contribution, we demonstrate the use of SPOD for denoising on surrogate data obtained by imposing high levels of additive Gaussian noise on simulation data. We demonstrate that SPOD-based denoising combines certain advantages of temporal filters and standard POD-based denoising. Due to their chaotic nature, turbulent flows are characterized by high levels of intermittency. A common tool for the analysis of intermittent behaviour is frequency-time analysis, that is, the representation of the frequency content of a time signal as a function of time. This representation is particularly well-suited to identify events such as shorttime interval of high or low energy, and identify their wave characteristics, i.e., frequencies or wavenumbers. Frequency-time analysis can be performed using several different signalprocessing tools such as wavelet transforms (WT) (Farge 1992), the short-time Fourier transform (STFT) (Cohen 1995), the S-transform (Stockwell et al. 1996), the Hilbert-Huang transform (Huang et al. 1998), and the Wigner-Ville distribution (Boashash 1988). Short-time Fourier and wavelet transforms are arguably the most widely used techniques in fluid mechanics. Frequency-time diagrams obtained from these methods are generally referred to as spectrograms, or as scalograms for wavelet transforms. STFT performs Fourier transforms on consecutive short segments of a time signal. It has been used, for example, in the analysis of blood flows (Izatt et al. 1997;Zhang et al. 2003), magnetohydrodynamics (Bale et al. 2005), aerodynamics (Samimy et al. 2007), and physical oceanography (Brown et al. 1989). The wavelet transform is based on the convolution of the time signal with a compact waveform, the so-called mother wavelet, that is scaled to represent different frequencies. Typical applications of the WT are found in atmospheric science (Gu & Philander 1995), oceanography (Meyers et al. 1993), and, most importantly for the present work, turbulence research (Farge 1992). In the latter context, they have been used to extract coherent structures (Farge et al. 1999(Farge et al. , 2001, and analyze their intermittency (Camussi & Guj 1997;Onorato et al. 2000;Camussi 2002). All methods mentioned above are signal-processing techniques that are applied to one-dimensional data. Here, we expand on the ideas of Schmidt et al. (2017a) and Towne & Liu (2019), and analyse the intermittency of the entire flow field. The underlying idea is that the global dynamics of the entire flow field can be described in terms of a limited set of statistically prevalent, most energetic coherent flow structures. For statistically stationary flows, such structures are distilled by the SPOD, and their temporal dynamics are described by the SPOD expansion coefficients. Since SPOD is a frequency-domain technique, this idea leverages the fact that each SPOD mode is associated with a single frequency. Based on the two reconstruction techniques mentioned above, we apply analyze and compare two variants of SPOD-based frequencytime analysis. The frequency-domain approach relies on direct inversion of the SPOD algorithm and was previously demonstrated by Towne & Liu (2019). This approach, however, becomes computationally intractable even for moderate-sized, two-dimensional data. We show that this problem can be avoided by the convolution-based approach introduced herein.
Prewhitening is post-processing technique for trend detection that is commonly used in the atmospheric and geophysical sciences. It was first proposed by von Storch (1995). Prewhitening was used, for example, for the detection of trends in temperature and precipitation data (Zhang et al. 2001), rainfall (Lacombe et al. 2012), teleconnections (Rodionov 2006), and hydrological flows (Khaliq et al. 2009;Serinaldi & Kilsby 2015). Technically, prewhitening is achieved by a filtering operation that results in a flat power spectrum to remove serial correlations. We show two different SPOD-based ways to achieve this goal in the frequency domain. The remainder of this paper is organized as follows. § 2 describes the two techniques for SPOD-based flow field reconstruction. In § 3, we demonstrate these techniques on numerical data of a turbulent jet. The four different applications, SPOD-based lowdimensional reconstruction, denoising, frequency-time analysis, and prewhitenening are demonstrated in § 3.1, § 3.2, § 3.3, and § 3.4, respectively. § 4 summarizes this work.

Spectral proper orthogonal decomposition (SPOD)
In the following, we provide an outline of a specific procedure of computing the SPOD based on Welch's method (Welch 1967) and emphasize aspects that are important in the context of data reconstruction. Refer to the work of Towne et al. (2018) for details of the derivation and mathematical properties and Schmidt & Colonius (2020) for a practical introduction to the method.
Given a fluctuating flow field q i " qpt i q, where i " 1,¨¨¨, n t , that is obtained by subtracting the temporal meanq from each snapshot of the data, we start by constructing a snapshot matrix Q " rq 1 , q 2 ,¨¨¨, q nt s. (2.1) Note that multi-dimensional data is cast into the form of a vector q i of length n, corresponding to the number of variables times the number of grid points. The instantaneous energy of each time instant, or snapshot, is expressed in terms of a spatial inner product where W is a positive-definite Hermitian matrix that accounts for the component-wise weights, Ω the spatial domain of interest, and p.q˚denotes the complex conjugate. The common form of space-only POD is obtained as the eigendecomposition of QQ˚W and yields modes that are optimal in terms of equation (2.2). SPOD, however, specializes POD to statistically stationary processes and seeks modes that are optimal in terms of the space-time inner product For statistically stationary data, it is natural to proceed in the frequency domain and solve the POD eigenvalue problem for Fourier transformed two-point space-time correlation matrix, that is, the cross-spectral density matrix. To estimate the CSD, the data is segmented into n blk overlapping blocks with n fft snapshots in each of them.
If the blocks overlap by n ovlp snapshots, the j-th column in the k-th block is given by Each block is considered as a statistically independent realization of the flow under the ergodic hypothesis. The motive behind the segmentation of data is to increase the number of ensemble members. In practice, a windowing function is applied to each block to reduce spectral leakage. In this study, we use the symmetric hamming window wpi`1q " 0.54´0.46 cosˆ2 πi n fft´1˙f or i " 0, 1,¨¨¨, n fft´1 .
(2.6) Following best practices established by Harris (1978), we only apply windowing for overlapping blocks to avoid excessive loss of information at the boundaries, i.e, if n ovlp ‰ 0. Subsequently, the weighted temporal discrete Fourier transform, is performed on each windowed block to obtain the Fourier-transformed data matrix, denotes the k-th Fourier realization at the i-th discrete frequency. Next, we reorganize the data by frequency. The matrix containing all realizations of the Fourier transform at the l-th frequency readŝ From this form, the SPOD modes, Φ, and associated energies, λ, can be computed as the eigenvectors and eigenvalues of the CSD matrix S l "Q lQl . In practice, the number of spatial degrees of freedom, n, is often much larger than number of realizations. In that case, it is more economical to solve the analogous eigenvalue problem 1 n blkQl WQ l Ψ l " Ψ l Λ l (2.10) for the coefficients ψ that expand the SPOD modes in terms of the Fourier realizations. In terms of the column matrix Ψ l " rψ p1q l , ψ p2q l ,¨¨¨ψ pn blk q l s, the SPOD modes at the l-th frequency are recovered as (2.11) The matrices Λ l " diagpλ s contain the SPOD energies and modes, respectively. By construction, the SPOD modes are orthogonal in the space-time inner product, equation (2.3). At any given frequency, the modes are also orthogonal in the spatial inner product, equation (2.2).

Data reconstruction
In §2.2.1, we show how the data can be reconstructed in the frequency-domain, that is, the inversion of the SPOD. An alternative approach based on (oblique) projection in the time domain is presented in §2.2.2. Whether one approach or the other is preferred depends on the specific application. Detailed discussions for each application under consideration in this work can be found in §3. We will use the term frequency-domain if the expansion coefficients are computed using the inversion of the SPOD problem, equation (2.13), and time-domain if oblique projection, equation (2.19), is used.

Reconstruction in the Frequency Domain
The common numerator of the different applications of the SPOD considered in this study is that they require truncation or re-weighting of the SPOD basis. In practice, this is achieved by modifying the expansion coefficients. The original realizations of the Fourier transform at each frequency can be reconstructed aŝ where A l is the matrix of expansion coefficients, (2.13) Equation (2.13) shows that the expansion coefficients can either be saved during the computation of the SPOD or be recovered later by projecting the Fourier realizations onto the modes. In the following, we omit the frequency index l with the understanding that the SPOD eigenvalue problem is solved at each frequency separately. From equation (2.12), it can be inferred that each column of the matrix a n blk 1 a n blk 2¨¨¨an blk n blk fi ffi ffi ffi fl , (2.14) contains the expansion coefficients that allow for the reconstruction of a specific Fourier realization from the SPOD modes. Vice versa, the coefficients contained in each row of A can be used to expand a specific SPOD mode in terms of the Fourier realizations. This can most easily be seen by rewriting equation (2.12) as Φ l " 1 n blkQ l Al Λ´1 l . The Fourier-transformed data of the k-th block can be reconstructed aŝ . (2.15) The original data in the k-th blocks Q pkq can now be recovered using the inverse (weighted) Fourier transform, . (2.16) Reconstructing the time series from the reconstructed data segments concludes the inversion of the SPOD. A schematic of the windowing and blocking strategy is shown in figure 1. The use of overlapping blocks leads to an ambiguity in the reconstruction as the i-th snapshot can either be obtained from the k-th block as q pkq j , or from the pk`1q-th block as q pk`1q j`n ovlp´nfft . Different possibilities to remove this ambiguity are described in appendix A. Based on this discussion, snapshots are reconstructed as averages of two reconstructions from overlapping blocks, weighted by the relative value of their windowing function. Partial reconstructions in the frequency-domain are readily achieved by zeroing the expansion coefficients of specific modes prior to applying the inverse Fourier transform.

Reconstruction in the Time Domain
As an alternative to the reconstruction in the frequency domain, we present in the following a projection-based approach in the time domain. This approach is computationally efficient, and has the advantage that it can be applied to new data that was Figure 1: Schematic representation of overlapping blocks, hamming window and the SPOD parameters such as n t , n fft , n ovlp and n blk . The left, and right blocks are denoted using blue, and red lines, respectively. The vertical line corresponds to the snapshot number 210, and the blue and red circles indicate the corresponding value of the window function. not used to compute the SPOD and to individual snapshots. However, the time-domain reconstruction does not leverage the orthogonality of the SPOD modes in the space-time inner product. Instead, it is based on an oblique projection of the data onto the modal basis. We start by representing the data as a linear combination of the SPOD modes as Q «ΦÃ. (2.17) The matrixΦ contains the basis of SPOD modes at all frequencies. Arranging the basis vectors by frequency first, we writẽ Assuming thatΦ has full column rank, the matrix of expansion coefficient is obtained from the weighted oblique projectioñ The oblique projection is required as SPOD modes at different frequencies are not orthogonal in the purely spatial inner product, x¨,¨y x , defined in equation (2.2). We furthermore use a weighted oblique projection based on the weight matrix W to guarantee compatibility with this inner product. Using the oblique projection, a single snapshot q " qpx, tq is represented in the SPOD basis as 20) and the entire data is recovered asQ "ΦÃ.
(2.21) In the case whereΦ is rank deficient or ill-conditioned,Φ˚WΦ is not invertible. For a generally rank-r matrix, we perform the symmetric eigenvalue decomposition, where D 1 " diagpd 1 , d 2 ,¨¨¨, d r q with d 1 ě d 2 ě¨¨¨ě d r ą 0 is the diagonal matrix of (numerically) non-zero eigenvalues and U 1 is the corresponding matrix of orthonormal eigenvectors. In some applications, we desire a more aggressive truncation to rank k ă r.
For full-rank reconstructions, we use a truncation threshold of d k d1 " 10´6 in this work. Denoted byÃ is the rank-k approximation ofÃ, where D tku " diagpd 1 , d 2 ,¨¨¨, d k q and U tku " ru 1 , u 2 ,¨¨¨, u k s. In the truncated basis, U tku D´1 tku Ut ku approximates´Φ˚WΦ¯´1. In the following, we demonstrate how the above and other truncation and partial reconstruction strategies can be used to achieve a number of objectives in the processing of flow data.
3. Applications of SPOD: Low-rank reconstruction, denoising, frequency-time analysis, and prewhitening In this section, four different uses of SPOD-based applications are introduced and demonstrated on the example of a turbulent jet. The theoretical background of each applications is presented in the context of the jet. In particular, low-dimensional reconstruction is discussed in §3.1, denoising in §3.2, frequency-time analysis in §3.3, and prewhitening in §3.4 respectively.
We consider the LES data of an isothermal subsonic turbulent jet, the Reynolds number, Mach number and temperature ratio are defined as Re " ρ j U j D{µ j " 0.45ˆ10 6 , M j " U j {c j " 0.4 and T j {T 8 " 1.0, respectively, where ρ is the density, U velocity, D nozzle diameter, µ dynamic viscosity, c speed of sound, and T temperature. The subscripts j and 8 refer to the jet inlet and free-stream conditions, respectively. We use the large-eddy simulation data computed by Brès & Lele (2019). The simulation was performed using the compressible flow solver 'Charles' ) on an unstructured grid using a finite-volume method. The reader is referred to Brès et al. (2018; for further details on the numerical method. The LES database consists of 10,000 snapshots sampled at an interval of ∆tc 8 {D " 0.2 acoustic time units. Data interpolated on a cylindrical grid spanning x, r P r0, 30sˆr0, 6s is used in this analysis. The flow is non-dimensionalized by the nozzle exit values, namely, velocity by U j , pressure by ρ j U 2 j , length by the nozzle diameter D, and time by D{U j . Frequencies are reported in terms of the Strouhal number St " f D{U j . For simplicity, and without loss of generality, we perform our analysis only on the pressure field in what follows. Refer to Freund & Colonius (2009) for analysis on different energy norms and POD modes of a jet. We further exploit the rotational symmetry of the jet and consider individual azimuthal Fourier components. The helical (m " 1) component provides an example of complex data.
To determine the spectral estimation parameters for the SPOD, we follow the guidelines provided in Schmidt & Colonius (2020) and Schmidt et al. (2018), which, in turn, follow standard practice in spectral estimation. The SPOD is computed for blocks containing n fft " 256 snapshots with 50% overlap, resulting in a total number of n blk " 77 blocks. A 50% overlap is used to minimize the variance of the spectral estimate (Welch 1967).
Since SPOD yields one set of eigenpairs per frequency, we may investigate the contributions of different frequencies independently. The SPOD eigenvalues are represented in the form of a spectrum, reminiscent to a power spectrum, in figure 2(a). Gray lines of decreasing intensity connect eigenvalues of constant mode number and decreasing mode energy. The first, most energetic mode is shown in magenta. The red line represents the sum of all eigenvalues and corresponds to the power spectral density (PSD) integrated over the physical domain. This line of the integral total energy can be compared to truncated sums of eigenvalues, that is, the energy contained in reconstructions of different Figure 2: SPOD spectra of the turbulent jet for m " 1. All eigenvalues (gray lines) and the sum of all eigenvalues (red line), corresponding to the integral PSD, are shown. The normalized cumulative energy content and the percentage of energy accounted by each mode as a function of frequency are shown in (b) and (c), respectively. The solid and dashed white lines indicate the number of modes required to retain 90% and 50% energy at each frequency. ranks. As the eigenvalues are sorted by energy, the lines corresponding to the truncated sums of the leading 3 and 10 eigenvalues fall between the leading-eigenvalue spectrum and the total energy curve. Figure 2(b) shows the normalized cumulative energy content, independent of frequency. The first and the leading 10 modes contain 30% and 80% of the total energy, respectively. Figure 2(c) shows the percentage of energy accounted for by each mode as a function of frequency. At low frequencies, the first few modes contain a high percentage of energy, whereas the energy is more dispersed at higher frequencies.
The solid and dashed white lines indicate the number of modes required to retain 50% and 90% of the total energy at each frequency.
The first and second modes at two representative frequencies are shown in figure 3. The frequency St " 0.5 corresponds to the maximum difference between first and second eigenvalues. The leading mode at this frequency shows a Kelvin-Helmholtz wavepacket (Suzuki & Colonius 2006;Gudmundsson & Colonius 2011;Schmidt et al. 2018) in the shear layer of the jet. A similar, but a more compact wavepacket structure is observed at St " 1.0. The suboptimal mode at both frequencies exhibit a multi-lobed wavepacket structure, whose amplitude peaks in near the end of the potential core at x « 6. The reader is referred to  and (Tissot et al. 2017) for a physical discussion of this observation and the link to non-modal instability. In the present context, our preliminary interest is in the desirable mathematical property of the SPOD that guarantee that the modes optimally represent the turbulent flow field in terms of the space-time inner product, equation (2.3). In the following, different uses of lowdimensional reconstructions that use SPOD modes as basis vectors are introduced and discussed.

Low-dimensional flow field reconstruction
Since SPOD seeks an optimal series expansion for each frequency, the choice of what eigenpairs to include in a low-dimensional reconstruction is not obvious. Here we first discuss the most elementary way of truncation based on the frequency-wise optimality property, that is, a certain number of modes is retained at each frequency. We refer to this as a n modesˆnfreq -mode reconstruction, where n modes is the number of modes retained at each frequency, and n freq " n fft 2`1 is the number of positive frequencies, including zero. If all modes are linearly independent, the overall rank of the reconstructions is then given by the total number of basis vectors, n modes n freq .
Following the discussion in §2, we present two means of obtaining an SPOD-based low-dimensional reconstruction: (i) in the frequency domain (see §2.2.1) using equation ( Tinney et al. (2008a). The time-domain approach can be viewed as the most general approach that can be applied to any given modal basis. It is not specific to SPOD, but commonly used for low-order modeling.
In the following, we first compare different low-dimensional reconstructions in terms of their block-wise and snapshot-wise energy in figure 4. It follows from equation (2.3), that the energy of a single block is where ∆T " rt 1`pk´1qpn fft´novlp q , t n fft`p k´1qpn fft´novlp q s is the time interval of the k-th block. The spatial norm (2.2) measures the energy present in each snapshot. Both the block-wise and snapshot-wise pressure norms are computed for both reconstruction approaches. The evolution of the space-time norm is shown for the entire database in figure 4(a,c). For the spatial norm, we focus on the first 768 snapshots in figure 4(b,d). This segment corresponds to 5 blocks and exhibits dynamics that are representative of the rest of the data. Low-dimensional reconstructions using 1ˆ129, 3ˆ129, 10ˆ129 modes and the reconstruction using all modes are presented in figure 4. The dimension of the modal bases directly reflects their ability to capture the pressure norms of the data. The full-dimensional reconstructions in the frequency and time domain recover the data completely. Notably, the dynamics in space-time norm are accurately captured, even by the 1ˆ129-mode reconstruction. For a fixed number of modes, the time domain approach captures more energy and provides a better approximation of the data than the frequency domain approach. Take as an example the 10ˆ129 basis: the time-domain reconstruction accurately approximates for the full data (figure 4(c,d)), which is notably under-predicted by frequency-domain reconstruction ( figure 4(a,b)). This difference can be understood by considering the SPOD energy content of the reconstruction. The dashed lines in figure 4(a,b) denotes this energy, which is given by the sum of first n modes eigenvalues over all frequencies. As expected, the space-time and spatial norm of the different reconstructions fluctuate about the sum of the eigenvalues. Higher energies are obtained by the timedomain reconstruction in figure 4(c,d) as the modal expansion coefficients obtained via oblique projection are not bound to specific frequencies. In what follows, we will see again and again that this flexibility of the expansion coefficients of the time-domain approach leads to an overall better reconstruction. This additional degree of freedom can be leveraged to obtain an accurate reconstruction of the flow dynamics. It is, in fact, the optimality property of the oblique projection, equation (2.23), that guarantees that the time-domain approach yields the best possible approximation in a least-square sense. For example, the 1ˆ129 time-domain reconstruction seen in figure 4(d) is sufficient to capture the dynamics of the original data accurately. From figure 4(b), it can be seen that the frequency-domain reconstruction significantly over-predicts the pressure norm during the first 32 snapshots. This effect only occurs for the two out-most blocks, which do not posses neighbouring blocks in one direction. In appendix A (figure 20), we show by comparison with rectangular window that this error is a result of the Hamming window. The presence of the windowing-effect in the first and last blocks is equally reflected in the space-time norm, see figure 4(a). Figure 5 demonstrates that the frequency-domain approach accurately recovers the mode energies given by the SPOD eigenvalues. By adding the residual energy contained in the truncated eigenvalues, both the space-time norm (figure 5(a)) and the spatial norm ( figure 5(b)) can be collapsed to the total energy. The remaining differences are largely due to the windowing effect. Figure 6 compares a single time instant of the original data in (a) with reconstructions of increasing fidelity in the frequency (left) and the time domain (right). Both the 1ˆ129mode reconstructions shown in figure 6(b,c) capture the dominant wavepackets. However, the frequency-domain reconstruction lacks the detail of the time-domain reconstruction. The higher accuracy of the time-domain reconstruction can be explained by its less stringent nature. As the leading SPOD modes often represent a spatially highly confined structure, other structures associated with the same frequency cannot be represented by the frequency-domain reconstruction. The KH-type wavepacket seen in figure 3(c) is a good example of such a confined structure. This difference between the approaches also explains the better reconstruction of the integral energy in the time domain, as previously observed in figure 4. The higher-dimensional versions for both approaches shown in figure 6(d-g) yield increasingly more detailed and accurate reconstructions.  Both approaches yield reconstructions that are indistinguishable from the original data when all modes are used (see figure 6(h,i)).
We infer from figures 4 to 6 that reconstruction in the time-domain provides a better estimate of the flow field than the frequency-domain version. To understand this observation, figure 7 reports the SPOD eigenspectra of the 1ˆ129-mode reconstructions and compares them to those of the full data. Only the leading three eigenvalues are shown for clarity. The leading eigenvalue of the frequency-domain reconstruction in figure 7(a) approximately follows the full data with some discrepancies at lower frequencies. No such discrepancies are observed for the time-domain reconstruction in figure 7(b); in fact, the leading eigenvalue spectra are indistinguishable. Contrast this observation with the expectation that a 1ˆn freq frequency-domain reconstruction should exactly reproduce the leading-mode eigenspectrum, and that all higher-eigenvalue spectra are expected to be zero. In the context of figure 21 in appendix A, we show that this is an effect of windowing that is not observed when using a rectangular window. The time-domain reconstruction, on the other hand, accurately approximates the first, and, to some degree, the leading suboptimal eigenvalue spectra. This again demonstrates the higher accuracy of the timedomain reconstruction that results from the higher flexibility of the expansion.
To quantify the accuracy of the two approaches, figure 8 compares their 2-norm errors as a function of the number of basis vectors (modes). For both the methods, the error reduces significantly as the number of modes retained per frequency increases from one to ten. For a fixed number of modes, the time-domain reconstruction is consistently more accurate. Recall that this is guaranteed by the optimality property of the oblique projection. It is, in fact, observed that the error of the time-domain reconstruction approaches machine precision for 40 or more modes per frequency. For the frequencydomain approach this only occurs for the full reconstruction using all modes.

Denoising
After using SPOD truncation for low-dimensional approximation above, we now explore its potential for denoising. We will show that additive noise is captured by certain parts of the spectrum, and that truncation of these parts leads to efficient denoising. This strategy is most efficiently implemented in the frequency domain. The local-in-time optimality of the time-domain approach is a hindrance in this context as it tends to reconstruct the noise. On the contrary, the one-to-one correspondence between modes and frequencies of the frequency domain approach leads to efficient denoising. As the arguably most common type of noise occurring in experimental environments, we demonstrate denoising on additive Gaussian white noise. In particular, we add Gaussian white noise that has a standard deviation equal to the spatial mean of the standard deviation along the lipline of the pressure data. This scenario mimics, for example, heavily contaminated particle image velocimetry data in which the variance of the noise is of the same order as the variance of the physical phenomena of interest. The SPOD eigenvalue spectrum of this noisy data is shown in figure 9(a). Most noticeably, the addition of noise has introduced a noise floor at λ « 4¨10´5, effectively cutting off the spectrum at St « 1.5. Information above this frequency lies below the noise floor and is not directly accessible. The leading eigenvalue of the original data (red dashed line) is shown for comparison. It lies well below the leading eigenvalue of the noisy data, which is elevated due to the energy contained in the added noise. To illustrate the ability of SPOD to differentiate between spatially correlated, physical structures and noise, examples of modes that are above and below the noise floor are compared in figure 9(bd). The leading mode at St " 0.5 (figure 9(b)) clearly reveals the K-H wavepacket and is indistinguishable from the corresponding mode of the original data shown in figure 3(a) above. The fifth mode at St " 0.5 (figure 9(c)) and the leading mode at St " 1.5 (figure 9(d)), on the contrary, are heavily contaminated by noise. The noise floor in the SPOD spectrum is found to be a very good indicator of this distinction. We hence propose a denoising-strategy based on hard-thresholding of the spectrum. In this example, we pick a threshold of 5ˆ10´5 (blue line), slightly above the noise floor. To address the effect of the truncation on the SPOD spectrum, we report the leading SPOD eigenvalue of the denoised data (magenta dashed line) in the same figure. It coincides with the leading eigenvalue of the noisy data up to the point where it intersects with the threshold limit, beyond which it falls off sharply, giving it the characteristics of a low-pass filter. A closer analysis of the truncated and original spectra reveals that the denoised field contains only 2.6% energy of the noisy field, but that it contains 92.7% energy of the original flow field. Another positive side effect is that only 74 out of 9933 modes (n freqˆnblk ) have been retained, resulting in a space saving of 99.26%. These significant space savings are an advantage over standard denoising strategies based on low-pass filtering. Figure 11: Comparison of the two denoising strategies: (a) signal to noise ratio (SNR) along the lipline (r " 0.5) for the noise added flow field (black line), SPOD-based denoised flow field (red lines), and the low-pass filtered flow field (blue line); (b) error of the noisy data, SPOD-based denoised data, and the low-pass filtered data. The amplitude of the additive noise was adjusted such that the average SNR along the lipline is one.
A representative instantaneous snapshots of the original data is compared to its noisy and denoised counterparts, and to the result of standard low-pass filtering in figure 10. The standard low-pass filter uses the cut-off frequency of St " 0.8 of the SPOD approach (inferred from figure 9) in the truncation of the long-time Fourier transform. A higher threshold, and its associated cut-off frequency, was found to lead to more aggressive filtering that can partially remove relevant flow structures. A threshold below the noise floor, on the other hand, leads to unsatisfactory noise rejection. In practice, a good trade-off between noise rejection and preservation of physically relevant flow structures is achieved by using the SPOD spectrum as a gauge to choose a threshold slightly above the noise floor. A comparison of the denoised data in figure 10(c,d) with the noisy data in figure 10(b) shows that significant noise reduction was achieved in all parts of the domain using both strategies. The resulting denoised flow fields clearly reveal the flow structures present in the original data. By visual inspection of the filtered pressure fields shown in figures 10(c) and 10(d), the SPOD-based strategy appears somewhat more efficient at removing the noise.
For a more quantitative assessment, we compare the denoised flow fields in terms of two quantities. First, their signal-to-noise ratio (SNR) along the lipline, and second, the relative error between the denoised snapshots and the original data. The SNR is defined as where P is power and σ standard deviation. We further define the integral (over the physical domain) error as where q andq are the original and the denoised flow fields, respectively. Figure 11(a) compares the SNR along the lipline at r " 0.5 for the noisy and the denoised flow fields. Both methods achieve to increase the signal-to-noise ratio over large parts of the domain. The low-pass filter performs better for x À 10, and the SPOD-based filter beyond that point. For x À 2.5, the SPOD-filtered pressure field exhibits a marginally lower SNR than the unfiltered data. We find that this is the results of the aggressive truncation of high-frequency components by the SPOD-based filter. A result that is almost identical to that of the low-pass filter can be achieved by lowering the λ-threshold (a similar value for both methods is used here for consistency). Figure 11(b) compares the time traces of the errors of the noisy and the two denoised flow fields. The error of the noisy data serves as a reference, and it is observed that both methods significantly reduce this error. The SPOD-based approach performs consistently better than the low-pass filter. This result is consistent with the visual observation of the denoised fields in figure 10(c,d).
Note that the threshold is an adjustable parameter in both methods. In practice, we find that by adjusting this parameter qualitatively very similar results can be obtained by both methods. This, however, leaves the SPOD-based approach with the advantage of significant data reduction.

Frequency-time analysis
Intermittency, that is the occurrence of flow events at irregular intervals, is an inherent feature of any turbulent flow. A common approach for the characterization of intermittent behaviour is frequency-time analysis. The arguably most wide-spread tools of frequencytime analysis are wavelet transforms and short-time Fourier transform. Their outcomes are scalograms and spectrograms, respectively, that indicate the presence of certain scales (WT), or frequency components (STFT), at certain times. Both methods are signalprocessing techniques that are applied to 1-D time series, and therefore only quantify intermittency locally. As an alternative to this local perspective, we demonstrate how SPOD expansion coefficients can be used to study the intermittency of the spatially coherent flow structures represented by the modes. Below, frequency-time analyses based on both, time-domain and frequency-domain reconstructions, are introduced and compared.

Time-domain approach
We first consider the time-domain approach, in which the expansion coefficients obtained via oblique projection readily describe the temporal behaviour of each mode. The amplitudes of the expansion coefficients computed from equation (2.19), | ř n modes j"1ã pjq pf l , tq|, hence yields the desired frequency-time representation for the leading n modes . The expansion coefficients are calculated using the full basis, i.e,φ in equation (2.18) consists of all modes at all frequencies. Subsequently, we only consider the expansion coefficients of the leading n modes modes at each frequency. An alternative  approach is to perform the oblique projection using a reduced basis that consists of only the leading n modes modes at each frequency. We find that the first approach is preferable in the context of frequency-time analysis and is explained in the appendix B (see figure  22). The frequency-time diagrams for n modes " 10, and 1, are shown in figure 12(a), and (b), respectively. The leading 10 modes correspond approximately to 80% of the total energy as shown in figure 2(b). Most of the energy is concentrated at low frequencies, St À 0.2, as expected from the eigenvalue spectrum in figure 2. The eigenvalue spectrum provides a statistical representation of the structures that are coherent in space and time, whereas the frequency-time diagrams provide a temporal information of these structures. Bright yellow spots indicate high similarity of the instantaneous flow field with the leading mode in (b). These regions also correspond to high energy events as we will show in figure 14.

Convolution-based approach
A direct way of using SPOD for frequency-time analyses is in terms of the SPOD expansion coefficients. Since each block is associated with a finite time interval, this approach requires the computation of the SPOD using an overlap of n ovlp " n fft´1 to obtain time-resolved coefficients (Towne & Liu 2019). This approach assumes that the value of the expansion coefficient obtained from a finite time segment (block) represents the instantaneous frequency content at the center of the time segment. A limitation of this approach is its high memory requirement (3.6 TB for the present example). As an alternative, we propose a computationally tractable way of calculating time-continuous expansion coefficients based on the convolution theorem. Applying the convolution theorem to the inverse SPOD problem yields a piq l ptq "´pφ piq l e´i 2πf l t q f q¯ptq " ż ∆T ż Ω´φ piq l pxq¯˚Wpxqqpx, t`τ qwpτ qe´i 2πf l τ dxdτ, (3.4) where f indicates the convolution between the time evolving SPOD mode and the data, that takes into the account the windowing function, wpτ q, and the weight matrix W. In practise this convolution is computed by expanding the SPOD mode in time as φ piq l e´i 2πf l t and convolving it over the data one snapshot at a time. In this step, we leverage the orthogonality property of the SPOD mode in the space-time inner product, which allows us to compute the expansion coefficient one at a time. If the SPOD was computed using an overlap of n fft -1, then the expansion coefficients, a piq l , obtained from equation (3.4) and equation (2.13) are mathematically identical. Here, the underlying idea is to apply the continuously-discrete convolution integral to the SPOD mode computed with a significantly lower overlap to make it computationally feasible. We confirmed that the frequency-time diagrams of the expansion coefficients obtained using equation (3.4) for an overlap of 50% are virtually indistinguishable to those obtained from equation (2.13) for an n ovlp " n fft´1 (shown in the appendix C, figure 24). This is to be expected since the convergence of the SPOD modes does not improve significantly for overlap over 50%. In practice, the convolution integral in equation (3.4) is most efficiently computed using the FFT.
Frequency-time diagrams of the expansion coefficients for the convolution approach are shown in figure 13. Figure 13(a), and (b) show the contribution of the leading 10 modes, and the leading mode, respectively. These frequency-time diagrams appear less detailed compared to the frequency-time diagrams of the time-domain approach. In the context of figure 15, however, we will show that figures 12 and 13 basically contain the same information. For now, it is sufficient to note that the convolution and timedomain approaches detect the same trends. Take as an example, the high energy events occurring at low frequency in the time ranges, 550 À t À 590 and 1370 À t À 1400, in the spectrograms of figure 12(b) and 13(b).

Comparison of methods and interpretation of results
Next, we investigate if the high similarity between the instantaneous flow field and the modes indicates high energy. Figure 14 shows the temporal evolution of the pressure 2-norm for the original data, low-dimensional 1ˆ129-mode reconstructions using the time-domain, frequency-domain and the convolution approaches. The pressure 2-norm of the data is highly underpredicted by the time-domain approach that uses a full basis (n blkˆ1 29), but is able to capture the major trends of the original data. The time-domain reconstruction performed using a modal basis of 1ˆ129 is also shown for comparison. This accurately follows the spatial norm of the original data, except for an offset, similar to figure 4(d). The frequency-domain curve also shows a similar trend. In addition, the spatial norm of the 1ˆ129-mode reconstruction using the convolution approach is shown. It follows the trend of the original data and attains its global maximum at the same time instant. All curves in figure 14 peak at the time of maximum instantaneous energy, previously indicated in figure 13(b). This indicates that a high similarity between the instantaneous flow field and the leading mode implies high overall energy. We highlight Figure 14: Temporal evolution of the pressure 2-norm in the vicinity of its global maximum at t " 1380. The convolution approach, based on equation (3.4), is compared to the time-and frequency-domain approaches, previously shown in figure 4(b,d). The 1ˆ129-mode reconstructions are shown. that this finding is not self-evident as the leading SPOD mode represents the most energetic flow structure in a purely statistical sense. The important physical insight hence is that the intermittent occurrence of large-scale coherent structures is directly associated with high-energy events.
To understand the qualitative differences of the frequency-time diagrams in figure 12 and 13, we now look at the expansion coefficients of the two approaches. Figure 15 compares the expansion coefficient of the leading mode, obtained by the time-domain and convolution approaches. As an example, the expansion coefficient at St « 0.50 is shown in figure 15(a). It is centered around its global maximum, in the time interval 1000 ď t ď 1300. We observe that the time traces of the expansion coefficients obtained from the two approaches show similar trends. In particular, the local peaks occur at similar locations, with both curves exhibiting the global maximum at t " 1150 (black dashed line). Compared to the time-domain approach, the convolution curve is smoother and resembles a moving average of the time-domain curve (black dash-doted line). For optimal comparison with the convolution approach, the moving average is computed by weighting the time-domain curve by the Hamming window in equation (2.6) and averaging it 256 points. To further quantify the relation between the expansion coefficients of the two approaches, we show the cross-correlation coefficient in figure 15(b). The crosscorrelation coefficient confirms the observation that the expansion coefficients obtained from the two approaches are similar, by demonstrating a cross-correlation coefficient of 0.77 at 0-time lag (τ " 0). We have confirmed that this correspondence holds in general. In figure 15(c) and 15(d), for example, the same trends are observed for the expansion coefficients at the lower frequency of St " 0.05 over the time interval previously shown in figure 14 above. From figure 15, we infer that the intermittency of the coherent structures can be captured using both approaches.
After establishing the correspondence between the time-domain and convolution approaches, we now compare the spectral characteristics of the two approaches. The PSDs of the expansion coefficients associated with the leading mode at St " 0.2, 0.5, and 1.0, for the time and convolution-domain approaches are shown in figure 16 (a). As expected the PSDs peak at the frequency of the corresponding mode for both approaches. The expansion coefficients computed in the convolution approach exhibit a much narrower peak than those in the time-domain approach. Note that, we cannot expect a sharp spectral peak even for the convolution approach due to spectral leakage and the modulation of the wave amplitude as seen in figure 15. Figure 16(c) compares the PSDs of the leading mode and the 10 leading modes at St " 0.5, which underlines the dominance of the leading mode at this particular frequency. The expansion coefficients are presented in terms of the spectral energy content in figure 16(b) and (d). As the expansion coefficients are uncorrelated, and their expected value is the equal to SPOD modal energy, it is expected that the convolution approach accurately approximates the eigenvalue spectrum. The blue and magenta lines are almost coincident in figure 16(b,d), thus confirming this conjecture. For the time-domain approach, the expected value for the sum of first 10 modes approximates the total integral PSD for all but the low frequencies expansion coefficients are obtained from a full basis here, but from a smaller basis in figure 4(d).
After examining the properties of the expansion coefficients, we focus on the spatial composition of the reconstructed flow field in terms of contributions from individual modes in figure 17. The time instant of high energy, previously marked in figure 15, is chosen as an example. The contribution of the leading SPOD modes at two representative frequencies, St " 0.5, and 0.05, to the original flow field is shown in figure 17. Figure  17(b) and (c) show the modes weighted by their expansion coefficient, a p1q φ p1q at the corresponding time instant and these two frequencies. Close to the nozzle exit the LES flow field clearly exhibits a KH-type instability wave. The leading mode at St " 0.5, closely resembles this structure. Similarly, the dominant wave pattern with a large wavelength (« 5) observed in the flow field, is represented by the mode at St " 0.05 in a location downstream of the potential core (12 À x À 20). The real part of the pressure field along the lipline (r " 0.5) is compared with the contributions of the leading modes at St " 0.5, and St " 0.05 in figure 17(d), and (e), respectively. It can be seen that the phases of the pressure field and the mode at St " 0.5 are aligned in the region where the mode attains its maximum. Weaker but similar kind of phase alignment is also observed (d,e) comparison of the real part of the pressure field along the lipline (r " 0.5) for St=0.5 and St=0.05, respectively. The instantaneous flow field at t "1150 (see figure 15) is shown. Contours in (a, b, c) are reported on the same color axis. for low frequency, St " 0.05, despite the disturbed nature of the wavepacket in the region, 12 À x À 20. For the flow field at the current time instant, this also explains that the contribution of the leading mode at St " 0.05 is lower than St " 0.5, where the KH wavepacket dominates the flow field. These observations confirm that the maximum in the frequency-time diagrams indicate close resemblance of the instantaneous flow field with the corresponding SPOD modes. Since the SPOD modes are the most energetic structures it is not surprising that the maxima in the frequency-time diagrams indicate the intervals of high energy. Furthermore, as the leading SPOD modes are spatially coherent and contain the most energy at each frequency, we infer that SPOD-based frequency-time analysis can be used to gauge the intermittency of large-scale coherent structures. Here, for brevity, only the convolution approach is shown, but we note that these conclusions also hold for the time-domain approach.

Prewhitening
As mentioned in §1, prewhitening is a filtering operation that results in a flat power spectrum, and is commonly used for trend detection in atmospheric and geophysical applications. The goal hence is to use SPOD to scale the data to have the same energy at all frequencies. We propose to achieve this goal by re-scaling the expansion coefficients of the reconstruction in the frequency-domain. The frequency-domain is chosen for the same reasons as for denoising in §3.2. We leverage the fact that expansion coefficients are uncorrelated, and that their expected value is the equal to SPOD modal energy, to scale the integral mode energy, that is, the sum of all eigenvalues, to one at each frequency. Both methods achieve this goal, but result in different relative scalings of individual SPOD modes. Figure 18(a) and 18(b) show the effect of the two scalings on the SPOD eigenvalue spectrum. The reconstructed flow fields obtained from the expansion coefficients scaled using equations (3.7) and (3.8) are shown in 18(c) and 18(d), respectively. Note that equation (3.7) collapses all eigenvalues in figure 18(a) to the same value. Scaling 2, on the contrary, preserves both the mode hierarchy and the relative energy content. In comparison to the original flow field shown in figure  6(a), prewhitening emphasizes high-frequency structures in the shear-layer, whereas it de-emphasizes the highly energetic large-scale structures associated with low frequencies downstream of the potential core. This portrayal of the flow field might appear unfamiliar; we emphasize that the objective of prewhitening is not physical interpretation, but pattern identification. Here, for example, the prewhitened pressure fields bring to light the trapped acoustic modes in the potential core. These modes have only recently been described in detail (Towne et al. 2017;Schmidt et al. 2017b). Previously, they remained largely unnoticed in the analysis of jet data because of their low energy content. An important difference between SPOD-based prewhitening and classical local, point-wise prewhitening techniques is that the SPOD-based approach preserves the spatial coherence of the flow structures identified by the SPOD modes.

Summary and Conclusions
Different applications of SPOD including low-rank reconstruction, denoising, prewhitening, and frequency-time analysis are demonstrated on the example of LES data of a turbulent jet. A fundamental building block for these applications is the capability to reconstruct the original data from the SPOD. In the frequency domain, this can be accomplished by inverting the SPOD problem (see, e.g., Citriniti & George 2000). We demonstrate that this inversion can be computed either directly in the frequency-domain, or using a convolution-based strategy. The latter approach becomes a necessity in the context of frequency-time analysis, where the corresponding SPOD problem becomes intractable. As an alternative to frequency-domain reconstruction, we introduce a time-domain approach that is based on the oblique projection of the data onto the SPOD modes.
First, we show the complete recovery of the data using all modes and compare to lowdimensional reconstructions. The low-dimensional reconstructions from both approaches accurately capture the integral energy of the segmented data (in the space-time norm). However, the time-varying dynamics (in the purely spatial norm) are only captured by the low-dimensional reconstructions in the time-domain. For a fixed number of modes, the time-domain approach captures more of the energy (in both norms). On the downside, the association of the SPOD modes with a single frequency is lost. Instantaneous pressure fields reconstructed in the frequency-domain, on the contrary, preserve this monochromatic property of the SPOD modes, but may lack the finer details of the flow field reconstructions in the time-domain. The main advantage of the frequencydomain approach is that it conserves the orthogonality property and the frequency-mode correspondence of the SPOD. The main advantage of the time-domain approach is its optimality in reconstructing the instantaneous flow field with the least possible number of modes.
After establishing the advantages and disadvantages of both approaches, we demonstrate SPOD-based denoising as an application of the frequency-domain approach. As expected, noise is mainly captured by higher SPOD modes at low frequencies and all modes at high frequencies. As a best practice, we propose a hard threshold above the noise floor that is identified from the SPOD eigenvalue spectrum. Significant noise reduction is achieved. At the same time, a substantial amount of energy of the original flow field is retained. Compared to a standard low-pass filter, SPOD-based denoising has the additional advantage of drastic space savings.
Finally, we demonstrate how SPOD-based frequency-time analysis can be used to analyse the intermittency of turbulent flows. Established means of frequency-time analysis such as wavelets transforms are signal-processing techniques that are applied to 1-D time signals. The alternative, SPOD-based approach demonstrated here, provides a global perspective in which spectrograms characterize the temporal evolution of the spatially coherent flow structures represented by the SPOD modes. The SPOD-based frequencytime analysis requires the computation of time-varying expansion coefficients at each time instant, and is computationally intractable in the frequency-domain. This problem is mitigated by the convolution-based strategy, which is mathematically equivalent in the limit of the intractable continuously-discrete (in time) SPOD problem. This convolutionbased approach is compared to the projection-based approach in the time domain. The expansion coefficients calculated from both methods show similar trends. We further demonstrate that a moving average of the spectrogram obtained via oblique projection resembles the spectrogram obtained from the convolution approach. For consistency, the moving time average is directly based on the SPOD windowing. The main advantage of the frequency-domain, and therefore the convolution approach, is that it retains the orthogonality property and mode-frequency correspondence of the SPOD. The frequencytime analysis of the jet data confirms the highly intermittent nature of this turbulent flow. In accordance with the SPOD eigenvalue spectrum, it is found that most of the energy is concentrated at low frequencies, St À 0.2. A comparison of the total flow energy as a function of time with the spectrograms shows that high energy events are directly linked to the presence of flow structures resembling the leading SPOD modes. We highlight that this behaviour is not necessarily expected as the SPOD modes represent the most energetic structures only in a statistical sense. From previous work , it is well-known that SPOD modes often isolate certain, prevailing physical phenomena. SPOD-based frequency-time analysis hence provides additional physical insight by indicating time intervals during which a particular mechanism is active.
Based on the results, we recommend the use of the time-domain approach for low-rank reconstruction of individual snapshots, and the frequency-domain approach for denoising and frequency-time analysis. For the latter application the proposed convolution strategy facilitates efficient computation of the time-continuous expansion coefficients. Figure 19: Errors of different ways to compute the frequency-domain reconstruction for 1ˆ129, 3ˆ129 and 10ˆ129 modes: (a) 2-norm; (b) spatial norm (equation (2.2)). Analogous to figure 5, the residual energy of the truncated modes is subtracted for comparison. particular 4(b)), it becomes apparent that these distinctions only matter for truncated series reconstructions; full-dimensional reconstructions are generally accurate.
The sudden jumps in the local energy of the reconstruction observed in figure 4 are a windowing effect. We demonstrate this by comparison with reconstruction using rectangular windows and no overlap in figure 20. The low-dimensional reconstructions using 1ˆ129, 3ˆ129, and 10ˆ129 modes, and all modes are shown. Many observations made in the context of figure 4, also hold here: the dimension of the modal bases is directly proportional to its ability to capture the pressure norm of the data, and for a fixed number of modes, the time-domain results in a better approximation of the data than the frequency-domain approach. The most notable difference can be seen between figure 20(b) and 4(b). The windowing effect in the frequency-domain reconstructions is absent if the rectangular window is used. Note in particular the difference during the first few snapshots and near the locations of switching from one block to another (vertical dotted black lines). Despite this advantage in the context of frequency-domain reconstructions for small n modes , rectangular windowing is generally not recommended because of spectral leakage (Schmidt & Colonius 2020).
Analogous to figure 7, we report in figure 21 the SPOD eigenspectra of the 1ˆ129-mode frequency and time-domain reconstructions for a rectangular window and n ovlp " 0. The SPOD eigenvalue spectra of the full data is also shown for comparison. Only the leading three eigenvalues are shown for clarity. The leading eigenvalue of the frequency-domain and time-domain reconstructions are indistinguishable from the leading eigenvalues of the full data. For the frequency-domain reconstruction, the higher eigenvalue spectra are zero to machine precision, as expected. This indicates that the windowing-effect causes the elevation of the higher eigenvalue spectra in figure 7(a). The time-domain reconstruction, on the other hand, is able predict the higher eigenvalues as explained in the context of 7(b). This implies that the time-domain reconstruction is much less sensitive to the choice of windowing function.   based on all SPOD modes, the majority of maxima is not found in the low-frequency regime, St À 0.2. To understand this difference, the PSD of the expansion coefficient associated with the leading mode at St " 0.5 is shown in figure 22(b). The expansion coefficient computed with the 1ˆ129-modal basis exhibits a much broader peak than the one computed with full basis. This behaviour indicates a loss of the mode-frequency correspondence for the heavily truncated basis. Next, both approaches are compared by taking the convolution-based expansion coefficient as the reference. The cross-correlation of the expansion coefficients from both approaches with the reference signal from the convolution approach are shown in figure 22(c). The expansion coefficient computed using the 1ˆ129-modal basis exhibits a much lower correlation with the reference. We conclude from this analysis that the time-domain approach should be conducted using all SPOD modes, as it yields a more accurate description of the intermittency of the coherent structure represented by the SPOD modes.
In figure 15, we demonstrated that the expansion coefficients computed from a moving average of the time-domain approach resemble those from the convolution approach. For further evidence, we show in figure 23 the frequency-time diagrams obtained by taking the moving mean, at each frequency, of the time-domain diagram previously shown in figure  12. The outcome should be compared to the frequency-time diagrams obtained using the convolution approach, i.e., figure 13. It is observed that the frequency-time diagrams are qualitatively very similar. The effect of taking the moving average is mainly visible at higher frequencies, where it leads to minor loss of detail. For qualitative flow analysis, we hence conclude that the moving average of the time-domain approach can well be used to approximate the computationally much more involved convolution approach.
Appendix C. Frequency-domain approach based frequency-time analysis: effect of overlap and correspondence to convolution-based approach Figure 24 demonstrates the similarity between the frequency-time diagrams obtained using the convolution method and the frequency-domain approach. For the convolution method, equation (3.4), a basis of precomputed SPOD modes with 50% overlap was used. To obtain time-continuous expansion coefficients using the frequency-domain approach, we require n ovlp " n fft´1 . As this is computationally intractable, the data (only here) was reduced to every third grid point in the streamwise and radial directions. The time traces of the individual expansion coefficients for St " 0.05 and 1.0 are reported in 24(c,d).