Studying propagating turbulent structures in the near wake of a sphere using Hilbert proper orthogonal decomposition Shaun Davey, Callum Atkinson and Julio Soria Abstract Turbulent flows, despite their apparent randomness, exhibit coherent structures that underpin their dynamics. Proper orthogonal decomposition (POD) has been widely used to extract these structures from experimental data. While periodic features like vortex shedding can be identified using POD mode pairs when periodicity dominates the flow, detecting such structures in complex flows is more challenging. The Hilbert proper orthogonal decomposition (HPOD) addresses this by applying POD to the analytic signal of the turbulent fluctuations, yielding complex modes with a 90 ∘ 90^{\circ} phase shift between the real and imaginary components. These modes capture propagating structures effectively but introduce filtering artefacts from the Hilbert transform that is used to derive the analytic signal. The current work investigates the relationship between the modes of the POD and those of the HPOD on the velocity fluctuations in the wake of a sphere. By comparing their outputs, POD mode pairs that correspond to the same propagating structures revealed by HPOD are identified. Furthermore, this study explored whether computing the analytic signal of the POD modes can replicate the HPOD modes, offering a more computationally efficient method for determining the pairs of POD modes that represent propagating structures. The results show that the pairs of POD modes identified by the HPOD can be more efficiently determined using the Hilbert transform directly on the POD modes. This method enhances the interpretive power of POD, enabling more detailed analysis of turbulent dynamics without introducing the filtering from the Hilbert transform. 1 Introduction While turbulence often appears random, many turbulent flows contain underlying structures, such as eddies and vortices, the scale and nature of which vary depending on the conditions of the flow. Identifying these structures allows for a more detailed description of the flow, as well as the isolation of the dominant structures in the turbulence based on their contribution to the quantity or phenomenon of interest, such as TKE ( TKE ) or mixing. Ranking these structures based on their significance to a given application or study allows the construction of parsimonious reduced-order models, which can be used to reduce the computational intensity of analyses or to reduce noise from experimental data. The complex flow phenomena present in the flow over bluff bodies, even at low Reynolds numbers, develop as the Reynolds number increases, and are a subject of ongoing study. In this respect, the flow over spheres serves as a canonical model for flows over axisymmetric bluff bodies, and is of particular interest for developing a broad understanding of these flow phenomena. The phenomena observed in the wake of a sphere include transition from axisymmetry to planar symmetry to asymmetry, the formation and shedding of vortices, the development of shear layer instabilities, and flow separation. As the flow over a sphere varies significantly with Reynolds number, experimental and numerical investigations have covered a wide range of Reynolds number regimes. In the context of experimental studies of the flow over spheres, one of the main challenges is fixing the sphere in a wind or water tunnel test section with minimal disturbance of the flow, since more often than not, the mounting/support structure of the sphere can have a significant influence on the flow. The supports chosen for an experiment should simultaneously provide rigid and stable positioning of the sphere, and minimise the disturbance to the flow, while also facilitating the placement of probes without obstructing optical measurement access. Some of the supports used in these studies include rigid downstream supports [ 1 ] , honeycomb structures with rigid supports upstream of the sphere [ 2 , 3 ] , and wires placed azimuthally [ 4 , 5 ] . The drag on the spheres has been measured using hot-wire probes [ 1 , 6 ] , with flow visualisations performed using aluminium dust [ 7 ] , smoke [ 8 , 5 ] , and dyes [ 9 , 4 , 10 ] to gain a qualitative understanding of the recirculation region, flow separation, and vortex shedding in the wake. Various forms of PIV ( PIV ), including planar [ 5 ] , stereographic [ 3 ] and tomographic [ 11 ] , have been used to quantitatively measure the flow around spheres. While mounting the sphere is not a concern in numerical studies, simulations are subject to their own set of challenges. The geometry of the problem and the range of scales involved require careful consideration of the mesh and solver to resolve the flow at high Reynolds numbers. Because of this complexity, LES ( LES ) [ 12 , 13 , 14 , 15 ] and DES ( DES ) [ 14 , 16 ] methods have been favoured over the more computationally expensive DNS ( DNS ) [ 12 , 15 , 17 , 18 ] to resolve the main features of the flow at high Reynolds numbers. In conjunction with the advancement of computational capabilities, these models have extended the feasibility of numerical investigations of the flow over spheres to large Reynolds numbers. Experimental and numerical studies have examined the development of the flow over spheres over a wide range Reynolds number regimes. At low Reynolds number, Re_{D}<20 , the flow in the wake is laminar and axisymmetric, and there is no recirculation [ 7 , 13 ] . A small vortex ring forms near the rear stagnation point around Re_{D}\approx 25 [ 7 ] , and elongates in the streamwise direction with increasing Reynolds number. This vortex ring begins to oscillate around Re_{D}\approx 130 with a long period, which becomes more prominent at higher Reynolds numbers. The recirculation region formed behind the sphere, which exhibits reverse flow in the mean, extends further downstream of the sphere as Reynolds number increases [ 9 ] . The axisymmetry in the wake at low Reynolds number is lost due to the formation of a double-threaded wake around Re_{D}\approx 210 [ 9 , 13 ] , which consists of two streamwise vortices of opposing sign that can be seen as dye threads in the wake of a free-falling sphere [ 19 ] . Single-frequency vortex shedding begins around Re_{D}\approx 300 [ 7 , 4 ] , and hairpin-shaped vortices form at Re_{D}\approx 400 [ 10 ] . The generation of an unsteady random side force, which has been observed in both experimental [ 8 ] and numerical [ 20 ] studies results in the loss of planar symmetry in the wake in the range of 400$0\\ \end{cases} (17) is the sign of the instantaneous frequency f of each term in the Fourier series given by \mathcal{F}\big[\boldsymbol{f}(\boldsymbol{x},t)\big] . For time-resolved data, the Hilbert transform can be performed in time to determine the analytic signal \boldsymbol{f}^{a}(\boldsymbol{x},t) . HPOD is performed on time-resolved data by replacing \boldsymbol{f}(\boldsymbol{x},t) with its analytic signal in equations ( 4 ) to ( 6 ). However, where the coherent structures propagate downstream, the Hilbert transform can be performed in the streamwise direction, as propagation in this direction serves as an analogy for a change in time, to apply this method to data that is not time-resolved [ 33 , 34 , 35 ] \mathcal{H}_{x}\big[\boldsymbol{f}(\boldsymbol{x},t)\big]=\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{\boldsymbol{f}\big((\xi,\ldots),t\big)}{x-\xi}\,d\xi. (18) In this case, the analytic signal of \boldsymbol{f}(\boldsymbol{x},t) is defined using \mathcal{H}_{x} \boldsymbol{f}^{a}_{x}(\boldsymbol{x},t)=\boldsymbol{f}(\boldsymbol{x},t)+i\mathcal{H}_{x}\big[\boldsymbol{f}(\boldsymbol{x},t)\big], (19) and substituted for \boldsymbol{f}(\boldsymbol{x},t) in equations ( 7 ) to ( 9 ). The resulting eigenvalues \lambda^{a}_{k} are the contribution of the k^{th} mode to the total TKE of the analytic signal, and are real-valued. The spatial modes \psi^{a}_{k}(\boldsymbol{x}) are complex, with the imaginary part of \psi^{a}_{k}(\boldsymbol{x}) representing a phase shift of its real part by \mp 90^{\circ} . The instantaneous amplitude and phase angle of the k^{th} mode are defined by the real \Re and imaginary \Im parts of a^{a}_{k}(t) as |a^{a}_{k}(t)|=\sqrt{\Re{\big(a^{a}_{k}(t)\big)}^{2}+\Im{\big(a^{a}_{k}(t)\big)}^{2}}, (20) and \phi^{a}_{k}(t)=\tan^{-1}\left(\mp\frac{\Im\big(a^{a}_{k}(t)\big)}{\Re\big(a^{a}_{k}(t)\big)}\right), (21) respectively. The instantaneous amplitude denotes the strength of the mode at each instant in time, and the instantaneous phase angle denotes the mode shape, which oscillates between the real and imaginary parts of \psi_{k}^{a}(\boldsymbol{x}) at \phi_{k}^{a}(t) equal to 0^{\circ} and \mp 90^{\circ} , respectively. The analytic signal can be reconstructed using \hat{K} modes by \hat{\boldsymbol{f}^{a}}(\boldsymbol{x},t)=\sum_{k=1}^{\hat{K}}a^{a}_{k}(t)\psi^{a}_{k}(\boldsymbol{x}), (22) the real part of which is the reduced-order model of \boldsymbol{f}(\boldsymbol{x},t) including the modes which contribute the most TKE to the analytic signal. This reconstruction favours the propagating components of \boldsymbol{f}(\boldsymbol{x},t) , but introduces artefacts due to the Hilbert transform. 2.3 Hilbert Transform of POD Modes Although the POD modes do not contain temporal information, the streamwise Hilbert transform can be performed directly on each of the modes. Thus, the analytic signal of the POD mode \psi_{k}(\boldsymbol{x}) can be defined as (\psi_{k})^{a}(\boldsymbol{x})=\psi_{k}(\boldsymbol{x})+i\mathcal{H}_{x}\big[\psi_{k}(\boldsymbol{x})\big]. (23) Substituting the analytic signal of the modes for the POD modes in equation ( 1 ) yields \sum_{k=1}^{K}a_{k}(t)\Big((\psi_{k})^{a}(\boldsymbol{x})\Big)=\sum_{k=1}^{K}a_{k}(t)\Big(\psi_{k}(\boldsymbol{x})+i\mathcal{H}_{x}\big[\psi_{k}(\boldsymbol{x})\big]\Big), (24) which can be expanded to \sum_{k=1}^{K}a_{k}(t)\Big((\psi_{k})^{a}(\boldsymbol{x})\Big)=\sum_{k=1}^{K}a_{k}(t)\psi_{k}(\boldsymbol{x})+a_{k}(t)i\mathcal{H}_{x}\big[\psi_{k}(\boldsymbol{x})\big]. (25) As the Hilbert transform is a linear operator, this can be rewritten as \begin{aligned} \sum_{k=1}^{K}a_{k}(t)\Big((\psi_{k})^{a}(\boldsymbol{x})\Big)&=\sum_{k=1}^{K}a_{k}(t)\psi_{k}(\boldsymbol{x})+i\mathcal{H}_{x}\Big[\sum_{k=1}^{K}a_{k}(t)\psi_{k}(\boldsymbol{x})\Big]\\ &=\boldsymbol{f}(\boldsymbol{x},t)+i\mathcal{H}_{x}[\boldsymbol{f}(\boldsymbol{x},t)]\\ &=\boldsymbol{f}^{a}_{x}(\boldsymbol{x},t)\\ \end{aligned}, (26) which is the analytic signal of \boldsymbol{f}(\boldsymbol{x},t) . Therefore, applying the transform to each of the POD modes of \boldsymbol{f}(\boldsymbol{x},t) individually and using the resulting analytic signals in place of the original modes allows for the reconstruction of \boldsymbol{f}^{a}(\boldsymbol{x},t) . However, the analytic signal of the POD modes are not equal to the HPOD modes due to the filtering of the Hilbert transform being introduced after the decomposition of \boldsymbol{f}(\boldsymbol{x},t) . While this method is not equivalent to HPOD , where the propagation of structures is embedded in the modes, it is a computationally efficient alternative for matching pairs of POD modes which represent the same propagating structure. 3 Methodology 3.1 Experimental Facility Experiments were performed in a vertical water tunnel, the test section of which is 1.5 m long with a cross-section of 0.25 m \times 0.25 m, shown in figure 1 a [ 37 , 38 , 39 ] . The settling chamber is connected to the top of the test section by a 16:1 contraction, which includes a honeycomb and screens to reduce the turbulence level and scale in the tunnel test section. Water is pumped from the plenum chamber below the test section up to the settling chamber such that the flow through the test section is downwards. The walls of the test section are made of 15 mm thick clear acrylic. A removable panel is located on one wall for access to the test section, and is set flush to the internal side of the wall to minimise asymmetry in the tunnel. Seeding and filtration of the tunnel is performed via an auxiliary circuit, which is isolated during experiments. A sting of diameter 9.3 mm is mounted to a crossbeam of thin aerofoil cross-sections above the test section, and tapers in diameter over the last 100 mm down to 4.65 mm in diameter and ends in a threaded rod which is used to mount a sphere in the test section. A close-up view of the sphere mounting structure is shown in figure 1 b. A sphere with a diameter of D=40 mm was manufactured using a Phrozen Sonic Mighty 8k 3-D printer, which has a 35 µm layer height and a planar resolution of 22 µm \times 22 µm, as shown in figure 1 c. The sphere has a shell thickness of 2.5 mm and an external boss at the top for mounting to a threaded rod on the end of the sting. The boss begins parallel to the sting and smoothly transitions to the sphere surface. The sphere is supported internally by a conical structure, which features slits to prevent uncured resin from being trapped in the cavities during printing. Figure 1: (a) Vertical water tunnel facility, and (b) close-up of the mounting structure and sphere, and (c) 3-D printed sphere design [ 39 ] . 3.2 Experimental Method Measurements of the flow in the wake of the sphere were taken with a freestream velocity of U_{\infty}=200 m/s, resulting in a Reynolds number of Re_{D}=7780 . Under these conditions, the background turbulence is 2%, and the mean convective time of the flow t_{c}=\frac{D}{U_{\infty}}, (27) is 0.2 seconds [ 39 ] . The flow was seeded with 11 µm glass spheres, and the single-exposed particle image pairs were recorded using a PCO Panda 26 DS camera with a 5120 \times 5120 px array of 2.5 µm \times 2.5 µm pixels. A Zeiss lens with a fixed 50 mm focal length was used to capture a 190 mm \times 190 mm field of view, resulting in a spatial resolution is 37 µm/px. The field of view began just behind the trailing edge of the sphere, to avoid reflections off the surface, and extended to 5.2 diameters downstream of the sphere. The field of view was aligned with the centreline of the sphere, and extended to 2.3 diameters either side of the centreline in the transverse direction. The f-stop was set to 2.8, resulting in a 5.0 mm depth of field. The diffraction limited minimum image diameter and expected particle image diameter for an 11 µm particle at these setting are 3.9 µm and 1.6 px, respectively. The particle images were illuminated using 7 ns pulses from a pair of 400mJ Spectra Physics Nd:YAG lasers. The beams of these lasers were aligned co-linearly before being shaped into a sheet aligned to the centre of the tunnel and the corresponding meridian of the sphere. The thickness of the laser sheet throughout the field of view was 300 µm. Pairs of single-exposed images where captured with a separation of \Delta t=1 ms at an acquisition rate of 1 Hz ( \Delta T=1 s), using a BeagleBone Black pulse generator [ 40 ] to control the timing and synchronisation of the camera and laser pulses. 3.3 Data Processing The mean background images of the first and second exposures were subtracted from the corresponding single-exposed particle image pairs. The instantaneous streamwise and transverse velocities, u(x,y,t) and v(x,y,t) , were determined from the particle image pairs using 2C-2D ( 2C-2D ) MCCD ( MCCD ) PIV [ 41 , 42 ] . A final interrogation window of 64 px \times 32 px on a grid with spacing of 32 px \times 16 px, in x and y , respectively, was used. This aspect ratio was chosen to capture the higher streamwise velocity component and the higher velocity gradients in the transverse direction, while still being sufficiently large for the interrogation window to contain approximately 10 particles common to both images of each pair. The quality of the cross-correlation was increased through the application of an 8-point Hart filter. A maximum displacement limit of 20 px and normalised local median filter [ 43 ] , with a threshold of 2 standard errors, was applied to the raw displacement vectors. The validation criteria resulted in a rejection of 2.3% of vectors, which were replaced by an interpolation of a 2nd-order polynomial fit of their 13 nearest neighbours [ 42 ] . A total of N_{T}=12,600 two-component velocity vector fields, each containing 50,000+ vectors, were determined. The instantaneous velocity fluctuations u^{\prime}(x,y,t) and v^{\prime}(x,y,t) were calculated by subtracting the mean velocity field, \overline{u}(x,y) and \overline{v}(x,y) , from the instantaneous velocity field, u(x,y,t) and v(x,y,t) . The 95% confidence interval of the 2C-2D MCCD PIV algorithm used is 0.06 pixels [ 42 ] , which corresponds to a standard error of \sigma_{\varepsilon_{u}}=0.03 pixels. The standard error of the instantaneous velocities is thus 1.1\times 10^{-3} m/s. The uncertainty of the mean velocity arising from the random nature of the turbulent fluctuations can be estimated by including the mean streamwise Reynolds stress [ 44 , 45 ] \sigma_{\overline{u}}=\sqrt{\frac{\overline{u^{\prime}u^{\prime}}+\sigma^{2}_{\varepsilon_{u}}}{N_{T}}}, (28) where \sigma_{\varepsilon_{u}} is the standard error of the instantaneous streamwise velocity measurements. As the maximum of both \overline{u^{\prime}u^{\prime}} and \overline{v^{\prime}v^{\prime}} are of the order of 0.05U_{\infty}^{2} [ 39 ] , a conservative estimate of the uncertainty in the mean velocity components is 4.0\times 10^{-4} m/s. The uncertainty in the fluctuations is thus \sigma_{u^{\prime}}=\sqrt{\sigma^{2}_{\overline{u}}+\sigma^{2}_{\varepsilon_{u}}}, (29) which is equal to 1.2\times 10^{-3} m/s. The key experimental and processing parameters are summarised in table 1 . Parameter Symbol Units Relative Units Freestream velocity U_{\infty} 200 mm/s - Sphere diameter D 40 mm - Reynolds number Re_{D} 7780 - Convective time t_{c} 0.2 s - Time between laser pulses \Delta t 1 ms 0.005t_{c} Time between velocity fields \Delta T 1 s 5t_{c} Depth of field - 5.0 mm 0.125D Laser sheet thickness - 0.3 mm 0.0075D Diffraction limited particle diameter - 3.9 µm Spatial resolution SR 37 µm - Grid spacing - 32 px \times 16 px 0.03D\times 0.015D Interrogation window - 64 px \times 32 px 0.06D\times 0.06D Field of view - 190 mm \times 190 mm 4.7D\times 4.7D Number of 2C-2D velocity fields N_{T} 12,600 - Velocity vectors per field N_{S} 50,000+ - Velocity vector validation rate - 97.7% Instantaneous velocity fluctuation uncertainty \sigma_{u^{\prime}} 1.2\times 10^{-3} m/s 0.006U_{\infty} Table 1: Summary of experimental and processing parameters. 4 Analysis and Results 4.1 POD and HPOD Modes POD was performed by arranging the turbulent fluctuations into the matrix \boldsymbol{X}=\begin{bmatrix}u^{\prime}(\boldsymbol{x}_{1},t_{1})&\cdots&u^{\prime}(\boldsymbol{x}_{N_{S}},t_{1})&v^{\prime}(\boldsymbol{x}_{1},t_{1})&\cdots&v^{\prime}(\boldsymbol{x}_{N_{S}},t_{1})\\ \vdots&u^{\prime}(\boldsymbol{x}_{n},t_{m})&\vdots&\vdots&v^{\prime}(\boldsymbol{x}_{n},t_{m})&\vdots\\ u^{\prime}(\boldsymbol{x}_{1},t_{N_{T}})&\cdots&u^{\prime}(\boldsymbol{x}_{N_{S}},t_{N_{T}})&v^{\prime}(\boldsymbol{x}_{1},t_{N_{T}})&\cdots&v^{\prime}(\boldsymbol{x}_{N_{S}},t_{N_{T}})\\ \end{bmatrix} (30) where N_{S} is the number of velocity vectors per velocity field and \boldsymbol{x}_{n}={(x,y)}_{n} is the spatial coordinates of the n^{th} vector, which has fluctuating components denoted by u^{\prime} and v^{\prime} in the streamwise and transverse directions, respectively. As N_{S}>N_{T} , the method of snapshots was used to perform the POD , with the two-point correlation of \boldsymbol{X} in time given by the matrix form of equation ( 7 ) \boldsymbol{C}=\boldsymbol{X}\boldsymbol{X}^{T}, (31) the eigendecomposition of which \boldsymbol{C}\boldsymbol{\Phi}=\lambda\boldsymbol{\Phi}, (32) yielded the eigenvectors \boldsymbol{\Phi} and eigenvalues \lambda . The temporal coefficients were determined from \boldsymbol{\Phi} a_{k}(t_{m})=\boldsymbol{\Phi}_{m,k}, (33) and the spatial modes were given by the projection of \boldsymbol{X} onto \boldsymbol{\Phi} \psi_{k}(\boldsymbol{x}_{n})=(\boldsymbol{\Phi}^{T}\boldsymbol{X})_{k,n}. (34) The analytic signal of \boldsymbol{X} \boldsymbol{X}^{a}=\boldsymbol{X}+i\mathcal{H}_{x}[\boldsymbol{X}], (35) was determined using equations ( 18 ) and ( 19 ), and HPOD was performed by substituting \boldsymbol{X}^{a} for \boldsymbol{X} in equations ( 31 ) to ( 34 ). The resulting eigenvalues, temporal coefficients and spatial modes are denoted by \lambda^{a}_{k} , a_{k}^{a}(t) and \psi_{k}^{a}(\boldsymbol{x}) , respectively. The contribution of the leading HPOD modes to the TKE of \boldsymbol{X}^{a} is significantly larger than the contribution of the leading POD modes to the TKE of \boldsymbol{X} . The first, second and third modes of each decomposition contribute 7.70%, 4.53% and 3.70%, and 4.8%, 3.39% and 2.55%, respectively. However, the difference in the contribution of the higher order modes to the TKE is less significant, as shown in figure 2 a. The result is that the cumulative contribution to TKE of a given number of the HPOD modes is larger than that of the equivalent number of POD modes, as shown in figure 2 b and table 2 . However, this is not a one-to-one comparison, as the total TKE of the analytic signal \boldsymbol{X}^{a} is larger than that of the data \boldsymbol{X} , at 571,871 m 2 /s 2 and 300,542 m 2 /s 2 , respectively. This is due to the imaginary component of \boldsymbol{X}^{a} , which is not present in \boldsymbol{X} . Figure 2: (a) Individual and (b) cumulative TKE contribution of the POD and HPOD modes. TKE [%] POD Modes HPOD Modes Ratio 25 12 7 0.58 50 65 38 0.58 75 311 180 0.58 90 1027 593 0.58 95 1973 1172 0.59 99 5904 4148 0.70 Table 2: Number of POD modes and HPOD modes required to meet 25% to 99% of the TKE of the data \boldsymbol{X} and its analytic signal \boldsymbol{X}^{a} , respectively, and the ratio of HPOD modes to POD modes required to meet each proportion of TKE . Figure 3: (i) Streamwise and (ii) transverse components of the first ten POD modes. The first POD mode consists of strong fluctuations directly behind the sphere x/D<2.5 , with substantially weaker fluctuations present around x/D=3.5 . The streamwise fluctuations, shown in figure 3 a(i), are opposite in sign across the centreline, at which they are zero, while the transverse fluctuations, shown in figure 3 a(ii), alternate in sign in the streamwise direction and are maximum at the centreline. The second POD mode, the streamwise and transverse components of which are shown in figures 3 b(i) and 3 b(ii), respectively, appears similar to the first mode, with the structures propagated downstream. The structures present at x/D<2.5 in the first mode are now located at 1.52.5 , which increase in wavelength further downstream. These modes differ in shape from the lower-order modes upstream of x/D=2 , where the structures have an aspect ratio closer to unity than those in the higher-order modes. The transverse components of the ninth and tenth modes, shown in figures 3 i(ii) and 3 j(ii), respectively, resemble those of the fourth and fifth modes with a short wavelength. The wavelength of these modes also increase in the downstream direction. Figure 4: (i) Streamwise and (ii) transverse components of the first four HPOD modes. The first HPOD mode, the real and imaginary parts of which are shown in figure 4 a and 4 b, respectively, is qualitatively similar in structure to the first two POD modes, shown in figure 3 a and 3 b, respectively. The streamwise components of the first HPOD mode, shown in figures 4 a(i) and 4 b(i) for the real and imaginary parts, respectively, consist of alternating positive and negative regions in the streamwise direction and is antisymmetric about the streamwise axis. The transverse components, shown in figures 4 a(ii) and 4 b(ii) for the real and imaginary parts, respectively, consist of symmetric regions of alternating sign. This structure is very similar to that of the first two POD modes, with the exception of the ends of the measurement domain. This difference is due to the filtering of the Hilbert transform, which places more emphasis on the components of the structure which propagate in the streamwise direction. The effect of this filtering is seen most clearly comparing figures 4 b(i) and 3 b(i) around x/D=5 . The second HPOD mode features the same V-shaped structure in the streamwise components, the real and imaginary parts of which are shown in figures 4 c(i) and 4 d(i), respectively, present in the fourth and fifth POD modes, shown in figures 3 d(i) and 3 e(i), respectively. The transverse components of the real and imaginary parts of the second HPOD mode are antisymmetric with regions of alternating sign in the streamwise direction, as shown in figures 4 c(ii) and 4 d(ii), respectively. While similar to transverse components of the fourth and fifth POD modes, shown in figures 3 d(ii) and 3 e(ii), respectively, the HPOD modes are significantly more symmetrical/anti-symmetrical than the POD modes, due to the filtering effect of the Hilbert transform. The third HPOD mode, the real and imaginary components of which are shown in figures 4 e and 4 f, have a similar structure to the sixth and eight POD modes, shown in figures 3 f and 3 h, respectively. Namely, antisymmetric streamwise velocity regions and symmetric transverse velocity regions of alternating sign in the streamwise direction, with a shorter wavelength than the first HPOD mode and the first two POD modes. The fourth HPOD mode, the real and imaginary components of which are shown in figures 4 g and 4 h, have a similar structure to the ninth and tenth POD modes, shown in figures 3 i and 3 j, respectively. The streamwise components of these modes feature symmetrical ovular regions of alternating sign in the streamwise direction in the near wake becoming V-shaped structures around x/D=2 , while the transverse components feature antisymmetric regions of alternating sign in the streamwise direction. 4.2 Pairing POD Modes using HPOD Modes As \boldsymbol{f}(\boldsymbol{x},t) is given by the real part of \boldsymbol{f}^{a}(\boldsymbol{x},t) , it can be expressed in terms of the HPOD modes as \boldsymbol{f}(\boldsymbol{x},t)=\Re\big(\boldsymbol{f}^{a}(\boldsymbol{x},t)\big)=\sum_{k=1}^{K}\Re\big(a_{k}^{a}(t)\psi^{a}_{k}(\boldsymbol{x})\big), (36) which can be expanded to \boldsymbol{f}(\boldsymbol{x},t)=\Re\big(\boldsymbol{f}^{a}(\boldsymbol{x},t)\big)=\sum_{k=1}^{K}\Re\big(a_{k}^{a}(t)\big)\Re\big(\psi^{a}_{k}(\boldsymbol{x})\big)-\Im\big(a_{k}^{a}(t)\big)\Im\big(\psi_{k}^{a}(\boldsymbol{x})\big). (37) As the real and imaginary parts of a_{k}^{a}(t) can be expressed as |a^{a}_{k}(t)|\cos(\phi^{a}_{k}(t)) and -|a^{a}_{k}(t)|\sin(\phi^{a}_{k}(t)) , respectively, the shape of the k^{th} HPOD mode at the phase angle \phi_{k}^{a}(t) is given by \frac{\psi_{k}^{a}(\boldsymbol{x},\phi_{k}^{a}(t))}{|a_{k}^{a}(t)|}=\cos\big(\phi_{k}^{a}(t)\big)\Re\big(\psi_{k}^{a}(\boldsymbol{x})\big)+\sin\big(\phi_{k}^{a}(t)\big)\Im\big(\psi_{k}^{a}(\boldsymbol{x})\big). (38) The mode shape of the k^{th} HPOD mode matches the j^{th} POD mode when \psi_{j}(\boldsymbol{x})=\begin{bmatrix}\cos(\Delta\phi^{a}_{k,j})&\sin(\Delta\phi^{a}_{k,j})\\ \end{bmatrix}\begin{bmatrix}\Re\big(\psi^{a}_{k}(\boldsymbol{x})\big)\\ -\Im\big(\psi^{a}_{k}(\boldsymbol{x})\big)\\ \end{bmatrix}, (39) where \Delta\phi_{k,j}^{a} is the phase offset of the k^{th} HPOD mode which results in the best match of the j^{th} POD mode. The negative of the imaginary part is used to account for the negative introduced by the Hilbert transform for positive instantaneous frequencies, given in equation ( 16 ), which results in the phase of the HPOD modes being reversed relative to the motion of the structure. The phase angles at which the k^{th} HPOD mode matches each of the POD modes can be determined simultaneously by \begin{bmatrix}\cos(\Delta\phi^{a}_{k,1})&\sin(\Delta\phi^{a}_{k,1})\\ \vdots&\vdots\\ \cos(\Delta\phi^{a}_{k,j})&\sin(\Delta\phi^{a}_{k,j})\\ \vdots&\vdots\\ \cos(\Delta\phi^{a}_{k,K})&\sin(\Delta\phi^{a}_{k,K})\\ \end{bmatrix}^{T}={\left(\begin{bmatrix}\Re\big(\psi^{a}_{k}(\boldsymbol{x})\big)\\ -\Im\big(\psi^{a}_{k}(\boldsymbol{x})\big)\\ \end{bmatrix}\begin{bmatrix}\Re\big(\psi^{a}_{k}(\boldsymbol{x})\big)\\ -\Im\big(\psi^{a}_{k}(\boldsymbol{x})\big)\\ \end{bmatrix}^{T}\right)}^{-1}\begin{bmatrix}\Re\big(\psi^{a}_{k}(\boldsymbol{x})\big)\\ -\Im\big(\psi^{a}_{k}(\boldsymbol{x})\big)\\ \end{bmatrix}\begin{bmatrix}\psi_{1}(\boldsymbol{x})\\ \vdots\\ \psi_{j}(\boldsymbol{x})\\ \vdots\\ \psi_{K}(\boldsymbol{x})\\ \end{bmatrix}^{T}, (40) and \Delta\phi^{a}_{k,j}=\tan^{-1}\left(\frac{\sin(\Delta\phi^{a}_{k,j})}{\cos(\Delta\phi^{a}_{k,j})}\right). (41) The correlation coefficient of the k^{th} HPOD mode at \Delta\phi^{a}_{k,j} and the j^{th} POD mode is \mathcal{R}_{k,j}=\frac{\operatorname{cov}\big(\psi_{k}^{a}(\boldsymbol{x},\Delta\phi^{a}_{k,j}),\psi_{j}(\boldsymbol{x})\big)}{\sqrt{\operatorname{var}\big(\psi_{k}^{a}(\boldsymbol{x},\Delta\phi^{a}_{k,j})\big)\operatorname{var}\big(\psi_{j}(\boldsymbol{x},)\big)}}, (42) and the corresponding R^{2} value is R^{2}_{k,j}=1-\frac{\sum\big(\psi_{k}^{a}(\boldsymbol{x},\Delta\phi^{a}_{k,j})-\psi_{j}(\boldsymbol{x})\big)^{2}}{\sum\big(\psi_{j}(\boldsymbol{x})\big)^{2}}. (43) The HPOD modes which best match each of the first ten POD modes, and the corresponding phase offset, correlation coefficient, and R^{2} value, are shown in table 3 . POD Mode, j HPOD Mode, k \Delta\phi_{k,j}^{a} \mathcal{R}_{k,j} R^{2}_{k,j} 1 1 1.71\pi 0.89 0.78 2 1 1.18\pi 0.97 0.93 3 3 0.18\pi 0.62 0.26 4 2 1.70\pi 0.84 0.65 5 2 1.21\pi 0.96 0.91 6 3 1.52\pi 0.82 0.65 7 4 1.70\pi 0.63 0.29 8 3 0.88\pi 0.76 0.50 9 4 1.10\pi 0.85 0.70 10 4 0.50\pi 0.80 0.49 Table 3: HPOD modes and phase shifts for best match of the first ten POD modes, and the corresponding \mathcal{R}_{k,j} values. The first and second POD modes are the best matches to the first HPOD mode, with correlation coefficients of 0.89 and 0.97, and R^{2} values of 0.78 and 0.93, respectively. The corresponding phase angles are 1.71\pi and 1.18\pi , which are approximately 90^{\circ} apart. This is consistent with the separation of the real and imaginary parts of the HPOD modes, suggesting that the relationship between these POD modes is similar to that of the real and imaginary parts of the first HPOD mode. Similarly, the fourth and fifth POD modes are the best matches to the second HPOD mode, with correlation coefficients of 0.84 and 0.96, and R^{2} values of 0.65 and 0.91, respectively. The corresponding phase angles are 1.7\pi and 1.21\pi , which are also separated by approximately 90^{\circ} . The sixth and eighth POD modes are the best match to the third HPOD mode, with correlation coefficients of 0.82 and 0.76, and R^{2} values of 0.65 and 0.50, respectively. The separation of the corresponding phase angles, 1.52\pi and 0.88\pi , respectively, is greater than 90^{\circ} . This is also the case for the ninth and tenth POD modes, which are the best matches to the fourth HPOD mode at phase angles of 1.10\pi and 0.50\pi , respectively. The corresponding correlation coefficients and R^{2} values of these POD modes are 0.85 and 0.80, and 0.70 and 0.49, respectively. The phase average of the paired POD modes were compared to the corresponding HPOD modes at phase angles from 0 to 2\pi in increments of 0.01\pi , accounting for the phase angle shift in the first POD mode in each pair, \Delta\phi_{k,i} . The R^{2} values of each phase average R_{k,(i,j)}^{2}=1-\frac{\sum{\Big(\psi^{a}_{k}(\boldsymbol{x},\phi+\Delta\phi^{a}_{k,i})-\widetilde{\psi_{i,j}}(\boldsymbol{x},\phi)\Big)}^{2}}{\sum{\big(\widetilde{\psi_{i,j}}(\boldsymbol{x},\phi)\big)}^{2}}, (44) was calculated for each case. The joint probability density functions of the phase angle of the HPOD modes and the phase angle of the corresponding pairs of POD modes, adjusted for the phase angle shift between the HPOD mode and the first POD mode of each pair. HPOD Mode k POD Modes i,j R_{k,(i,j)}^{2} 1 1,2 0.81 2 4,5 0.76 3 6,8 0.57 4 9,10 0.61 Table 4: R_{k,(i,j)}^{2} values of the phase-averaged HPOD modes and the corresponding POD modes. Figure 5: Phase-averaged (a) streamwise and (c) transverse components of the first and second POD modes, and the phase-averaged (b) streamwise and (d) transverse components of the first HPOD mode, adjusted for the phase angle offset \Delta\phi_{1,1} , at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . The streamwise component of the phase-averaged first and second POD modes, shown on figure 5 a, is good agreement with that of the phase average of the first HPOD mode at the adjusted phase angle, shown in figure 5 b, with the main differences in the structures being a result of the filtering from the Hilbert transform. Namely, the intensity of the structures in the POD mode is higher directly behind the sphere than the HPOD modes, while the intensity in the structures further downstream is higher in the HPOD modes. This difference is also seen in the transverse comments, shown in figures 5 c and 5 d for the POD modes and HPOD modes, respectively. Additionally, the structures in the near wake of the sphere in the POD modes, shown in figure 5 a(iv), are filtered out of the HPOD mode, shown in figure 5 b(iv), as they do not propagate downstream. Despite these differences, it is clear that the first and second POD modes and the first HPOD mode represent the same propagating structure consisting of planar antisymmetric streamwise fluctuations and alternating positive and negative transverse fluctuations. This structure is consistent with flapping of the wake, and the similarity of the third POD mode, shown in figure 3 c, to the first and second suggests that the flapping also has an out-of-plane component. Figure 6: Phase-averaged (a) streamwise and (c) transverse components of the fourth and fifth POD modes, and the phase-averaged (b) streamwise and (d) transverse components of the second HPOD mode, adjusted for the phase angle offset \Delta\phi_{2,4} , at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . While the streamwise component of the phase-averaged fourth and fifth POD modes, shown in figure 6 a, is similar to that of the second HPOD mode, shown in figure 6 b, the latter demonstrates significantly stronger symmetry, due to the inherent filtering of propagating features by HPOD . This is most clearly seen when comparing figures 6 a(ii) and 6 a(vi) to figures 6 b(ii) and 6 b(vi), respectively, where the V-shaped structures in the HPOD case remain symmetric downstream, while they become asymmetric in the case of the POD modes. The transverse component of the second HPOD mode, shown in figure 6 d, shows a stronger antisymmetry than the transverse components of the phase average of the fourth and fifth POD modes, shown in figure 6 c. The variation between the phase average of the fourth and fifth POD modes and the second HPOD mode highlights the filtering effect of HPOD , which prioritises the structures that are periodic in the direction of the Hilbert transform. While the fourth and fifth POD modes are highly asymmetric, particularly after x/D>3 , the second HPOD mode is symmetric in the streamwise component and antisymmetric in the transverse direction. The asymmetry in the POD modes is likely due to the three-dimensionality in the flow being more significant in the structure represented by the fourth and fifth POD modes than that represented by the first and second POD modes. In this case, the use of HPOD is beneficial for isolating the propagating structures in the flow in the presence of asymmetry in the POD modes. The alternating positive and negative streamwise fluctuations in these modes suggests that they represent a pulsating motion in the wake of the sphere. Figure 7: Phase-averaged (a) streamwise and (c) transverse components of the sixth and eighth POD modes, and the phase-averaged (b) streamwise and (d) transverse components of the third HPOD mode, adjusted for the phase angle offset \Delta\phi_{3,6} , at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . The streamwise component of the phase average of the sixth and eight POD modes, shown in figure 7 a, is qualitatively similar to that of the third HPOD mode, shown in figure 7 b. However, the alternating positive and negative structures are better defined in the HPOD mode than the phase average of the POD modes. This is most evident when comparing figures 7 a(i) and 7 a(ii) to figures 7 b(i) and 7 b(ii), in which the positive region in the HPOD modes around x/D=3 and y/D>0 is not present in the phase-averaged POD modes. This is most likely due to the structure moving in the out-of-plane direction as it propagates, which is filtered out by the Hilbert transform. Additionally, the intensity of the structures is generally lower in the HPOD modes than those in the POD modes due to the filtering of the non-propagating components. The transverse component of the phase average of the sixth and eighth POD modes, shown in figure 7 c, shows a transverse shifting of the structures of alternating sign, particularly in figures 7 c(i)-(iii). This shifting is not present in the transverse component of the third HPOD mode, shown in figure 7 d. The shifting of the structures in the transverse velocity of the POD modes is consistent with the the three-dimensional motion of the wake, and coupled with the behaviour of the streamwise components of the POD modes. The structure represented by these modes appears to be a flapping in the wake with a shorter wavelength than that represented by the first and second POD modes and the first HPOD mode. Figure 8: Phase-averaged (a) streamwise and (c) transverse components of the ninth and tenth POD modes, and the phase-averaged (b) streamwise and (d) transverse components of the fourth HPOD mode, adjusted for the phase angle offset \Delta\phi_{4,9} , at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . The streamwise component of the phase-averaged ninth and tenth POD modes, shown in figure 8 a, is qualitatively similar to the phase average of the fourth HPOD mode, shown in figure 8 b, for x/D>2.5 . In this domain, the streamwise components of both the POD modes and HPOD mode exhibit V-shaped structures similar to those in the fourth and fifth POD modes and the second HPOD mode. For x/D<2.5 , there is more detail in the structure of the POD modes, which is most noticeable in figures 8 a(iii) and 8 a(iv) compared to figures 8 b(iii) and 8 b(iv). As the V-shaped structures in the streamwise components are more prominent in the phase-averaged ninth and tenth POD modes than the fourth HPOD mode, it can be inferred that they do not propagate in the streamwise direction. The transverse components of phase average of the ninth and tenth POD modes and the phase average of the fourth HPOD modes, shown in figures 8 c and 8 d, respectively, are also qualitatively similar. As is the case for the streamwise components, the transverse component of the phase-averaged POD modes contains features not present in the HPOD mode in the very near wake, due to the filtering effect of the Hilbert transform. This can be seen clearly in figures 8 c(ii) and 8 c(iii) compared to figures 8 d(ii) and 8 d(iii). Figure 9: Joint probability density function of the phase angles of (a) the first HPOD mode and the first and second POD modes, (b) the second HPOD mode and the fourth and fifth POD modes, (c) the third HPOD mode and the sixth and eighth POD modes, and (d) the fourth HPOD mode and the ninth and tenth POD modes. Phase angles are relative to the phase angle of each mode in the first instantaneous velocity field. The joint probability density functions of the HPOD modes and the corresponding POD mode pairs, shown in figure 9 , are all the greatest around \phi^{a}_{k}=\phi_{i,j} . The phase angles of both have been adjusted by the phase angle of the first instantaneous velocity field to account for the phase shift between the POD modes and the HPOD modes and to ensure all phase angles are in the domain \phi\in[0,2\pi] . 4.3 Pairing POD Modes using the Hilbert Transform While HPOD modes can be used to determine matched pairs of POD modes, this process requires performing both decompositions. As the Hilbert transform can be applied in the streamwise direction, it can be applied directly to the POD modes despite their lack of temporal information. The result is a shift of the fundamental wavenumbers of each POD mode by \mp 90^{\circ} , which is analogous to a propagation of the mode downstream. Thus, the streamwise Hilbert transform of a POD mode is analogous to a POD mode which represents a \mp 90^{\circ} phase shift of the same streamwise periodic structure. In order to determine matched pairs of POD modes using the Hilbert transform, the Hilbert transform of each POD mode \psi_{k}(\boldsymbol{x}) in the streamwise direction \mathcal{H}_{x}[\psi_{k}(\boldsymbol{x})] is used in place of the imaginary part of the HPOD modes to define the analytic signal of \psi_{k}(\boldsymbol{x}) {(\psi_{k})}^{a}(\boldsymbol{x})=\psi_{k}(\boldsymbol{x})+i\mathcal{H}_{x}\big[\psi_{k}(\boldsymbol{x})\big]. (45) The shape of the analytic signal of the k^{th} POD mode at the phase angle \phi_{k} is given by \big(\psi_{k}\big)^{a}(\boldsymbol{x},\phi_{k})=\cos(\phi_{k})\psi_{k}(\boldsymbol{x})+\sin(\phi_{k})\mathcal{H}_{x}[\psi_{k}(\boldsymbol{x})], (46) and the phase angle at which the k^{th} POD mode best matches the j^{th} POD mode is given by \begin{bmatrix}\cos(\Delta\phi_{k,1})&\sin(\Delta\phi_{k,1})\\ \vdots&\vdots\\ \cos(\Delta\phi_{k,j})&\sin(\Delta\phi_{k,j})\\ \vdots&\vdots\\ \cos(\Delta\phi_{k,K})&\sin(\Delta\phi_{k,K})\\ \end{bmatrix}^{T}=\left(\begin{bmatrix}\psi_{k}(\boldsymbol{x})\\ \mathcal{H}_{x}[\psi_{k}(\boldsymbol{x})]\\ \end{bmatrix}\begin{bmatrix}\psi_{k}(\boldsymbol{x})\\ \mathcal{H}_{x}[\psi_{k}(\boldsymbol{x})]\\ \end{bmatrix}^{T}\right)^{-1}\begin{bmatrix}\psi_{k}(\boldsymbol{x})\\ \mathcal{H}_{x}[\psi_{k}(\boldsymbol{x})]\end{bmatrix}\begin{bmatrix}\psi_{1}(\boldsymbol{x})\\ \vdots\\ \psi_{j}(\boldsymbol{x})\\ \vdots\\ \psi_{K}(\boldsymbol{x})\\ \end{bmatrix}^{T}, (47) and \Delta\phi_{k,j}=\tan^{-1}\frac{\sin(\Delta\phi_{k,j})}{\cos(\Delta\phi_{k,j})}. (48) The corresponding correlation coefficient is \mathcal{R}_{k^{a},j}=\frac{\operatorname{cov}\big((\psi_{k})^{a}(\boldsymbol{x},\Delta\phi_{k,j}),\psi_{j}(\boldsymbol{x})\big)}{\sqrt{\operatorname{var}\big((\psi_{k})^{a}(\boldsymbol{x},\Delta\phi_{k,j})\big)\operatorname{var}\big(\psi_{j}(\boldsymbol{x})\big)}}. (49) The R^{2} value of the phase average of the k^{th} and j^{th} POD modes compared to the phase average of the analytic signal of the k^{th} POD mode is given by R^{2}_{k^{a},\widetilde{(k,j)}}=1-\frac{\sum\big((\psi_{k})^{a}(\boldsymbol{x},\phi)-\widetilde{\psi}_{k,j}(\boldsymbol{x},\phi)\big)^{2}}{\sum\big(\widetilde{\psi}_{k,j}(\boldsymbol{x},\phi)\big)^{2}}, (50) where (\psi_{k})^{a}(\boldsymbol{x},\phi) is the analytic signal of the k^{th} POD mode and \widetilde{\psi}_{k,j}(\boldsymbol{x},\phi) is the phase average of the k^{th} and j^{th} POD modes at phase angle \phi , which was determined using phase angles \phi=[0,2\pi] in increments of 0.01\pi . The POD mode which best matches the analytic signal of the each of the first ten POD modes, the corresponding correlation coefficient, and the R^{2} of the phase averages are given in table 5 . POD Mode k Paired POD Mode j \Delta\phi_{k,j} \mathcal{R}_{k^{a},j} R^{2}_{k^{a},\widetilde{(k,j)}} 1 2 -0.5\pi 0.88 0.89 2 1 0.5\pi 0.76 0.76 3 2 0.5\pi 0.70 0.74 4 5 -0.5\pi 0.79 0.81 5 4 0.5\pi 0.72 0.72 6 8 -0.5\pi 0.80 0.82 7 9 -0.5\pi 0.60 0.67 8 6 0.5\pi 0.76 0.77 9 10 -0.5\pi 0.64 0.65 10 9 0.5\pi 0.63 0.65 Table 5: Paired POD modes using the analytic signal of the k^{th} POD mode, and the corresponding phase angle, correlation coefficient and R^{2} of the phase averages. For each of the first ten POD modes, the paired POD mode corresponds to the analytic signal at a phase angle of \mp 0.5\pi . Thus, the j^{th} POD mode best matches the Hilbert transform of the corresponding k^{th} POD mode. A pair of POD modes representing a structure propagating in the streamwise direction therefore requires that the POD modes are a mutual best match with phase angle shifts of 0.5\pi with opposite signs. Thus, the POD mode pairs identified using the Hilbert transform are the first and second POD modes, the fourth and fifth POD modes, the sixth and eighth POD modes, and the ninth and tenth POD modes. These are the same pairings as those identified using the HPOD modes. Figure 10: Phase-averaged (a) streamwise and (c) transverse components of the first and second POD modes and the phase-averaged (b) streamwise and (d) transverse components of the phase average of the analytic signal of the first POD mode at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . The streamwise component of the analytic signal of the first POD mode, shown in figure 10 b, is a close match to the phase average of the first two POD modes, shown in figure 10 a. Both show the same long regions of alternating sign propagating downstream from the sphere, with the main variation between them being for x/D<2 , where the non-propagating features are located, as shown in figure 5 a(iv). This is an artefact of applying the Hilbert transform to the POD mode, due to these structures being present in the second POD mode but not the first POD mode, on which the Hilbert transform was performed. This also explains the variation in the R^{2} of the analytic signal of the first POD mode and the second POD mode and that of the analytic signal of the second POD mode and the first POD mode, being 0.89 and 0.76, respectively. This difference is due to the non-propagating components of these modes being absent in the other mode, but propagated by the Hilbert transform. The variation between the transverse components of the phase-averaged first and second POD modes, shown in figure 10 c, and the analytic signal of the first POD mode, shown in figure 10 d, is significantly less than in the streamwise components. Both exhibit the same symmetric regions of alternating sign, which increase in length and decrease in intensity as they travel downstream. This is consistent with the vertical fluctuations from the flow coming around the sphere decaying once they are sufficiently far into the wake. Figure 11: Phase-averaged (a) streamwise and (c) transverse components of the fourth and fifth POD modes and the phase-averaged (b) streamwise and (d) transverse components of the phase average of the analytic signal of the fourth POD mode at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . The streamwise component of the analytic signal of the fourth POD mode, shown in figure 11 b, retains more of the asymmetry of the phase-averaged fourth and fifth POD modes, shown in figure 11 a, than the second HPOD mode, shown in figure 6 b. This is also true for the transverse component of the analytic signal of the fourth POD mode, shown in figure 11 d. The transverse component of the analytic signal of the fourth POD mode, shown in figure 11 d, also retains the asymmetry seen in the phase average of the fourth and fifth POD modes, shown in figure 11 c. Although the filtering of the Hilbert transform is less prominent in the analytic signal of the fourth POD mode than in the second HPOD mode, shown in figures 6 b and 6 d, it is still present, most noticeably figures 11 b(iii) and 11 d(iv). Figure 12: Phase-averaged (a) streamwise and (c) transverse components of the sixth and eighth POD modes and the phase-averaged (b) streamwise and (d) transverse components of the phase average of the analytic signal of the sixth POD mode at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . While there is some filtering in the analytic signal of the sixth POD mode, the asymmetry seen in the phase average of the sixth and eighth POD modes is preserved where the HPOD modes were planar symmetric. This can be seen in the positive region introduced to the streamwise component of the third HPOD mode at x/D=2.5 , shown in figure 7 b(i), which is not present in the phase average of the sixth and eight POD modes, shown in figure 12 a(i), or the analytic signal of the sixth POD mode, shown in figure 12 b(i). It is also shown in the asymmetric region at x/D=3.5 in the transverse component of the sixth and eighth POD modes, shown in figure 12 c(i), which is also off-centre in the analytic signal of the sixth POD mode, shown in figure 12 d(i), but located at the centreline in the third HPOD mode, shown in figure 7 d(i). Additionally, the intensity of the structures further downstream, such as in figure 12 a(i), is clearly preserved in the analytic signal of the sixth POD mode, as shown in figure 12 b(i), where the intensity of the same region is significantly reduced in the third HPOD mode, shown in figure 7 b(i). This intensity in the phase average of the sixth and eighth POD modes around x/D=4 , shown in figure 12 a(v), is propagated downstream in the analytic signal of the sixth mode, as shown in figure 12 b(v), whereas the intensity is lower in the same region in the third HPOD mode, as shown in figure 7 b(v). This shows that the filtering which occurs from the Hilbert transform also affects the analytic signal of the POD modes. Figure 13: Phase-averaged (a) streamwise and (c) transverse components of the ninth and tenth POD modes and the phase-averaged (b) streamwise and (d) transverse components of the phase average of the analytic signal of the ninth POD mode at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . While there is a significant difference between the streamwise components of the analytic signal of the ninth POD mode, shown in figure 13 b, and that of the phase average of the ninth and tenth POD modes, shown in figure 13 a, for x/D<2 , they are more similar further downstream of the sphere. The structures in the phase average of the ninth and tenth modes are smaller scale for x/D<2 than they are for x/D>2 , which is consistent with smaller scale fluctuations in the near wake breaking down as they travel downstream while the larger scales remain. As a result, these scales are not propagated by the Hilbert transform and are thus filtered in the analytic signal. The differences between the transverse components of the analytic signal of the ninth POD mode, shown in figure 13 d, and the phase average of the ninth and tenth POD modes, shown in figure 13 c, are significantly less prominent. However, there is still variation in the structures upstream of x/D=1.5 , as shown in figure 13 c(ii) and 13 d(ii), and figures 13 c(iii) and 13 d(iii). 4.4 Propagating Structures in the Wake of a Sphere The contribution of the pairs of POD modes identified to the total TKE of \boldsymbol{X} is given in table 6 . HPOD Mode TKE [%] POD Modes TKE [%] 1 7.70 1, 2 8.24 2 4.53 4, 5 4.75 3 3.70 6, 8 3.20 4 2.70 9,10 2.60 Table 6: Turbulent kinetic energy contributions of the HPOD modes to \boldsymbol{X}^{a} and of the POD mode pairs to \boldsymbol{X} . In order to determine the nature and effects of the propagating structures identified from the POD modes, the paired POD modes were used to phase average the instantaneous velocity fluctuations \begin{bmatrix}\widetilde{u^{\prime}}(\boldsymbol{x},\phi)\\ \widetilde{v^{\prime}}(\boldsymbol{x},\phi)\\ \end{bmatrix}=\frac{1}{N_{\phi}}\sum_{k=1}^{K}a_{k}(t)\psi_{k}(\boldsymbol{x})\,|\,\phi_{i,j}=\phi, (51) where i and j are paired POD modes, and N_{\phi} is the number of instantaneous velocity fields included for the phase angle \phi . The mean velocity was added to the phase-averaged velocity fluctuations to determine the phase-averaged velocity \begin{bmatrix}\widetilde{u}(\boldsymbol{x},\phi)\\ \widetilde{v}(\boldsymbol{x},\phi)\\ \end{bmatrix}=\begin{bmatrix}\widetilde{u^{\prime}}(\boldsymbol{x},\phi)\\ \widetilde{v^{\prime}}(\boldsymbol{x},\phi)\\ \end{bmatrix}+\begin{bmatrix}\overline{u}(\boldsymbol{x})\\ \overline{v}(\boldsymbol{x})\\ \end{bmatrix}, (52) from which the phase-averaged vorticity is defined as \widetilde{\omega}(\boldsymbol{x},\phi)=\frac{\partial\widetilde{v}(\boldsymbol{x},\phi)}{\partial x}-\frac{\partial\widetilde{u}(\boldsymbol{x},\phi)}{\partial y}. (53) The phase-averaged planar Reynolds stress is given by the phase average of the product of the streamwise and transverse components of the instantaneous velocity fluctuations. The TKE of the phase-averaged velocity is given by \widetilde{\operatorname{TKE}}(\boldsymbol{x},\phi)=\big(\widetilde{u^{\prime}}(\boldsymbol{x},\phi)\big)^{2}+\big(\widetilde{v^{\prime}}(\boldsymbol{x},\phi)\big)^{2}. (54) Figure 14: Phase-averaged (a) velocity, (b) vorticity, (c) planar Reynolds stress, and (d) TKE using the first and second POD modes at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . Figure 15: (a) Velocity, (b) vorticity, (c) planar Reynolds stress, and (d) TKE of the first HPOD mode at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . The motion represented by the first and second POD modes, shown in figure 14 , and the first HPOD mode, shown in figure 15 , is a flapping of the wake. The phase-averaged velocity from using the first and second POD modes and the first HPOD mode, shown in figure 14 a and 15 a, respectively, are qualitatively indistinguishable, as is the phase-averaged vorticity, shown in figure 14 b and 15 b, respectively. The structures in the planar Reynolds stresses of the phase-average of POD modes, shown in figure 14 c, are longer and more intense for x/D<3 than those in the first HPOD mode, shown in figure 15 c, and disappear after x/D=3 . The structures in the planar Reynolds stress of the HPOD mode are shorter and more consistent in size across the measured domain. The intensity of the structures is lower than those in the POD modes, but decays more slowly, with structure visible past x/D=3 . The phase-averaged TKE of the POD modes, shown in figure 14 d, is concentrated in periodically shed packets which initially increase in intensity until x/D=2 , and then decay by x/D=3 . The TKE of the HPOD mode, shown in figure 15 d, reaches exhibits the same periodic structures as the POD modes, however, the intensity is increased. This is consistent with the TKE of the analytic signal of the turbulent fluctuations being significantly higher than that of the actual turbulent fluctuations. Figure 16: Phase-averaged (a) velocity, (b) vorticity, (c) planar Reynolds stress, and (d) TKE using the fourth and fifth POD modes at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . Figure 17: (a) Velocity, (b) vorticity, (c) planar Reynolds stress, and (d) TKE of the second HPOD mode at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . The structure represented by the fourth and fifth POD modes and the second HPOD mode is a pulsating in the streamwise fluctuations, as can be seen from the deformation of the low velocity region behind the sphere, shown in figure 16 a and 17 a, respectively. The streamwise pulsation can also be seen in the phase-averaged out-of-plane vorticity, with the high-vorticity regions either side of the sphere extend further toward the centreline close to the sphere in figures 16 b(ii) and 17 b(ii). This extension toward the centreline occurs around x/D=2 in figures 16 b(v) and 17 b(v). While the structures in the planar Reynolds stress are highly variable in shape and size in the phase-averaged POD modes, shown in figure 16 c, they are consistent in size and shape in the HPOD mode, shown in figure 17 c. The size and shape of the phase-averaged TKE of the POD modes, shown in figure 16 d, is more consistent than that of the planar Reynolds stresses. These structures are similar to, but less planar symmetric, than the structures in the phase-averaged TKE of the HPOD mode, which also exhibit a higher intensity with a more consistent pattern. The intensity of the TKE increases from directly behind the sphere until x/D=2 , as seen in figure 17 d(ii), before decreasing further downstream, as seen in figure 17 d(iv). Figure 18: Phase-averaged (a) velocity, (b) vorticity, (c) planar Reynolds stress, and (d) TKE using the sixth and eighth POD modes at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . Figure 19: (a) Velocity, (b) vorticity, (c) planar Reynolds stress, and (d) TKE of the third HPOD mode at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . The sixth and eighth POD modes represent a flapping motion in the wake, similar to that of the first and second POD modes but with a shorter wavelengths. Thus, the phase-averaged velocity, shown in figure 18 a, and phase-averaged vorticity, shown in figure 18 b, are similar to those of the first and second POD modes. The smaller structures are evident in the phase-averaged planar Reynolds stress, shown in figure 18 c, particularly near the sphere. Further downstream of the sphere, the structures increase in streamwise extent, most notable in figure 18 c(iv). The smaller structures in the phase-averaged TKE , shown in figure 18 d, are more definitively separated across the centreline of the sphere than in the case of the first and second POD modes, where the larger structures extended to the centreline. The structures in the phase-averaged planar Reynolds stresses of the third HPOD mode, shown in figure 19 c, are more consistent in size, shape and arrangement than those of the POD modes. There is also a region with low planar Reynolds stress around x/D=2 for all phase angles shown in figure 19 , which reflects the smaller structures in the streamwise turbulent fluctuations in this regions, shown in figure 7 c. The structures in the phase-averaged TKE are also more clearly defined for the HPOD modes, as shown in figure 19 d. Figure 20: Phase-averaged (a) velocity, (b) vorticity, (c) planar Reynolds stress, and (d) TKE using the ninth and tenth POD modes at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . Figure 21: (a) Velocity, (b) vorticity, (c) planar Reynolds stress, and (d) TKE of the fourth HPOD mode at \phi= (i) 0, (ii) \pi/6 , (iii) \pi/3 , (iv) \pi/2 , (v) 2\pi/3 , and (vi) 5\pi/6 . The structure represented by the ninth and tenth POD modes, and the fourth HPOD mode, is similar to that represented by the fourth and fifth POD modes, and the second HPOD mode, namely periodic streamwise fluctuations in the wake. The phase-averaged velocity, shown in figures 20 a and 21 a for the POD modes and HPOD mode, respectively, and the phase-averaged vorticity, shown in figures 20 b and 21 b for the POD modes and HPOD mode, respectively, are qualitatively similar to those of the lower order modes. The structures in the planar Reynolds stress phase-averaged using the POD modes, shown in figure 20 c, are small and round in the near wake of the sphere and larger and more ovular further downstream, with the latter being angles away from the centreline in accordance with the V-shaped structures in the POD modes. While the arrangement of these structures is variable, particularly close to the sphere, in the phase-average using the POD modes, the planar Reynolds stress of the HPOD mode, shown in figure 21 c, clearly shows the round structures for x/D<2 and more V-shaped structures for x/D>2 . Both the POD modes and HPOD modes show a noticeable reduction in the planar Reynolds stress around x/D=2 , which delineates the rounder structures from the V-shaped ones. However, the HPOD modes display this more prominently than the POD modes, which also exhibit additional structures in the very near wake which are filtered out in the HPOD modes. The change between round structures near the sphere to more V-shaped structures is reflected in the phase-averaged TKE , shown in figure 20 d and 21 d, for the POD modes and HPOD mode, respectively. This also shows that the structures close to the sphere are more intense and decay in intensity toward x/D=2 , with the V-shaped structures increasing in intensity around x/D=3 before decaying in intensity downstream. As in the case of the planar Reynolds stress, the structures in the TKE are more clearly defined in the HPOD mode due to the filtering of the Hilbert transform. The structures elucidated by the POD mode pairs identified from the HPOD modes differ from those in the corresponding HPOD modes. The primary difference between these methods results from the filtering inherent in the Hilbert transform, as the non-propagating structures are given less importance in the HPOD , while propagating structures are amplified. The phase-averaging of the resulting modes thus results in more consistently shaped structures making it easier to identify and interpret the propagating structures. However, the filtering out of non-propagating structures may result in a poorer representation of the flow structure where the flow is three-dimensional. While applying the Hilbert transform directly to the POD modes is more computational efficient than performing the HPOD for pairing POD modes, it also forcibly propagates structures in the flow which may not propagate in reality. Thus, both methods of pairing POD modes introduce non-physical artefacts into the analysis of the flow and this should be considered when interpreting results. 5 Concluding Remarks The velocity in the wake of a sphere at Re_{D}=7780 was measured using 2C-2D MCCD PIV , and POD and HPOD were applied to the turbulent fluctuations. The resulting POD and HPOD modes were compared to determine if the HPOD modes, which accentuate propagating structures, could be used to match POD modes which represented the same structure at different phase angles. The Hilbert transform of the POD modes was then used as a substitute for the HPOD modes in order to determine paired POD modes without performing the computationally-intensive HPOD . The POD mode pairs identified using these methods were then used to phase average the turbulent fluctuations to examine the propagating structures in the wake of the sphere. The leading HPOD modes are shown to represent the same structures as the leading POD modes, with a greater emphasis on the propagating structures in the former. While each HPOD mode consists of a real and imaginary part representing a phase-shift of 90 ∘ in the mode, this comes at the cost of filtering from the Hilbert transform. As a result, it can be difficult to differentiate between what is part of the flow structures and what is an artefact of this filtering, particularly in the case of two-dimensional measurements of a three-dimensional flow. Performing the Hilbert transform on the POD modes to create an analytic signal of the POD modes is shown to identify the same pairs of POD modes identified using the HPOD modes. This method allows the same structures identified using the HPOD modes to be identified, without introducing the filtering resulting from the HPOD . The phase average of the turbulent fluctuations using the identified POD mode pairs represented flapping and pulsating motions in wake of the sphere. As the velocity fields in the current work are not time-resolved, the Hilbert transform was performed in the streamwise direction, and thus the structures identified are those which propagate in that direction. As a result, the temporal information of the structures is not given by either decomposition. However, the paired POD modes gave insight into the dynamics of the structures in the turbulent fluctuations when the complexity of the flow made identifying related POD modes challenging. Declaration of Interests The authors report no conflict of interest.