---
title: Hilbert Proper Orthogonal Decomposition (HPOD)
url: https://www.emergentmind.com/topics/hilbert-proper-orthogonal-decomposition-hpod
type: topic
---

# Hilbert Proper Orthogonal Decomposition (HPOD)

Searching arXiv for the cited HPOD papers and closely related POD/Hilbert-space references.
Hilbert Proper Orthogonal Decomposition (HPOD) is a complex-valued extension of proper orthogonal decomposition (POD) in which the dataset is first converted into an analytic signal by a Hilbert transform and is then decomposed by POD. In the fluid-mechanical usage established in recent work, HPOD is designed to make travelling, advecting coherent structures explicit, especially when standard real-valued POD distributes one physical propagating structure across two or more modes in phase quadrature. The method has been developed for advection-dominated flows and applied to bluff-body wakes, turbulent jets, and temporally under-resolved particle-image-velocimetry datasets; its central outputs are complex modes and coefficients that encode propagation, local amplitude, and local phase in a single modal object [2507.02487].

## 1. Definition and motivation

Standard POD decomposes a fluctuating field into orthonormal spatial modes and temporal coefficients,
\[
\mathbf{f}(\mathbf{x},t)=\sum_{k=1}^{K} a_k(t)\,\psi_k(\mathbf{x}),
\]
with modes ranked by energy. In turbulent-flow analysis, the eigenvalues quantify each mode’s turbulent kinetic energy contribution. POD is attractive because it does not require time-resolved measurements and often reveals coherent structures efficiently. Its limitation, for the present topic, is that a propagating or oscillating structure is usually represented by two or more real-valued modes corresponding to different phases of the same physical motion, while POD itself does not identify which modes belong together [2507.19167].

That limitation is mild in simple strongly periodic flows, where mode pairs can often be recognized from approximately circular coefficient portraits or from an external phase reference. It becomes more serious in complex turbulent flows containing multiple interacting structures, asymmetry, and three-dimensional effects. The wake of a sphere at \(Re_D=7780\) is presented as precisely such a case: a propagating motion may be split across POD modes without an explicit mathematical linkage, so pairing by inspection becomes difficult [2507.19167].

The broader physical target of HPOD is the advective wavepacket. The canonical representation used in the foundational paper is
\[
w_j(x,t)=A_j(x,t)\cos(k_j x-\omega_j t)
       =A_j(x,t)\cos\!\big(k_j(x-c_j t)\big),
\]
where \(A_j(x,t)\) is a slowly varying amplitude envelope, \(k_j\) is local wavenumber, \(\omega_j\) is local frequency, and \(c_j=\omega_j/k_j\) is phase velocity. This representation is intended to retain propagation, local amplification and decay, amplitude modulation, frequency modulation, and intermittency, which are all difficult to encode compactly in standard real-valued POD [2507.02487].

## 2. Mathematical construction

The defining step of HPOD is the formation of an analytic signal. For a real scalar signal \(s(t)\), the Hilbert transform is
\[
\mathcal{H}_t[s(t)] = \lim_{\epsilon\to 0} \int_{|\tau-t|>\epsilon}\frac{s(\tau)}{\pi (t-\tau)}\,d\tau,
\]
and the analytic signal is
\[
\tilde{s}(t)=s(t)+i\,\mathcal{H}_t[s(t)].
\]
In Fourier space,
\[
\mathcal{F}\{\mathcal{H}[s]\}(f) = -i\,\operatorname{sign}(f)\,\mathcal{F}[s](f),
\]
so the transform acts as a \(\pm \pi/2\) phase shift, while the analytic signal keeps positive frequencies, doubles them, and removes negative frequencies [2507.02487].

For a flow field, the temporal version is written
\[
\tilde{\mathbf{u}}(\mathbf{x},t)=\mathbf{u}(\mathbf{x},t)+i\,\mathcal{H}_t[\mathbf{u}(\mathbf{x},t)].
\]
The space-only formulation replaces the temporal transform by a transform along the advection direction \(x\),
\[
\tilde{\mathbf{u}}(\mathbf{x},t)=\mathbf{u}(\mathbf{x},t)+i\,\mathcal{H}_x[\mathbf{u}(\mathbf{x},t)],
\]
with
\[
\mathcal{H}_x[\mathbf{u}(x,y,z,t)] = \lim_{\epsilon\to 0} \frac{1}{\pi} \int_{|\xi-x|>\epsilon} \frac{\mathbf{u}(\xi,y,z,t)}{x-\xi}\,d\xi.
\]
HPOD is then simply POD applied to the complexified field,
\[
\tilde{\mathbf{u}}(\mathbf{x},t) = \sum_j \psi_j(t)\,\sigma_j\,\boldsymbol{\phi}_j(\mathbf{x}),
\]
where both \(\boldsymbol{\phi}_j(\mathbf{x})\) and \(\psi_j(t)\) are generally complex-valued [2507.02487].

The conceptual gain is that a single complex mode contains two quadrature components of one propagating structure. In the sphere-wake formulation, the complex HPOD mode is written
\[
\psi_k^a(\mathbf{x}) = \Re(\psi_k^a(\mathbf{x})) + i\,\Im(\psi_k^a(\mathbf{x})),
\]
with the imaginary part interpreted as the quadrature counterpart of the real part, approximately a \(\mp 90^\circ\) phase-shifted version of the same propagating structure. The complex temporal coefficient similarly yields an instantaneous amplitude and phase,
\[
|a_k^a(t)|=\sqrt{\Re(a_k^a(t))^2+\Im(a_k^a(t))^2},
\]
\[
\phi_k^a(t)=\tan^{-1}\!\left(\mp \frac{\Im(a_k^a(t))}{\Re(a_k^a(t))}\right).
\]
A phase-conditioned mode shape then interpolates continuously between real and imaginary parts as phase advances, embedding propagation directly into the modal representation [2507.19167].

A central analytical result is that HPOD eigenfunctions inherit the analytic-signal property: time-analytic HPOD yields temporal modes with only positive frequencies, while space-analytic HPOD yields spatial modes with only positive wavenumbers. This one-sidedness is the formal basis for interpreting HPOD modes as complex wavepackets rather than merely complexified energetic structures [2507.02487].

## 3. Conventional and space-only HPOD

Two versions of HPOD are treated explicitly. Conventional HPOD computes the analytic signal in time and therefore requires time-resolved data. Space-only HPOD computes the analytic signal along the advection direction and therefore requires adequate spatial resolution along that direction but not temporal resolution. The latter was introduced specifically to address snapshot datasets in which temporal sampling is too slow for time-based modal methods [2507.02487].

The physical intuition comes from the travelling-wave factor \(e^{i(kx-\omega t)}\). For such a structure, a \(\pi/2\) phase shift can be generated either temporally or spatially,
\[
e^{i(kx-\omega t)} = f(x,t)+i\,f\!\left(x,t+\frac{\pi}{2\omega}\right)
\]
or
\[
e^{i(kx-\omega t)} = f(x,t)+i\,f\!\left(x+\frac{\pi}{2k},t\right).
\]
Accordingly, the Hilbert transform in time and the Hilbert transform in space both recover the missing quadrature component of the same travelling structure [2507.02487].

The equivalence argument is given under an idealized stationary and homogeneous setting with a phase-velocity relation \(c=\omega/k\). Under that assumption, the space-only and conventional formulations recover the same physical structures, up to conjugation or phase convention. The equivalence is exact for wave-like advecting structures satisfying a meaningful phase-velocity relation and adequate resolution in the transformed dimension. It is not expected for non-advective phenomena or structures without such a relation, and the authors regard that selectivity as a feature because it emphasizes advective content [2507.02487].

This distinction is practically decisive for experimental work. Conventional HPOD degenerates toward standard POD when time resolution is inadequate, whereas space-only HPOD can remain informative when spatial sampling resolves the advective pattern but temporal sampling does not. The intended application domain is therefore not only time-resolved DNS or LES, but also snapshot PIV and related under-resolved measurements [2507.02487].

## 4. Interpretation of modes and comparison with related decompositions

An HPOD structure consists of a complex spatial mode, a complex temporal coefficient, and a singular value. The real and imaginary parts of the spatial mode are two quadrature components of one propagating pattern. Propagation is encoded in the phase: a monotonic phase progression in the advection direction indicates a travelling wave, and a \(\pi/2\)-type quadrature relation between real and imaginary parts distinguishes propagation from a standing pattern. The modulus of the mode acts as an envelope, revealing amplification and decay, while the modulus of the temporal coefficient reveals temporal modulation and active or inactive intervals [2507.02487].

Because the outputs are analytic signals, HPOD permits local amplitude, local phase, and instantaneous frequency in time,
\[
f(t)=\frac{1}{2\pi}\frac{d\varphi(t)}{dt},
\]
as well as the corresponding spatial interpretation in terms of local wavenumber from spatial phase variation. In the published examples, these quantities are examined through time-frequency spectrograms of temporal coefficients and space-wavenumber spectrograms of spatial modes. The extracted structures are described not as spectrally pure, but as broadband, often band-limited or narrowband around a dominant frequency or wavenumber, and capable of exhibiting modulation and intermittency [2507.02487].

This places HPOD between standard POD and SPOD. Standard POD is energy-optimal on real-valued data but does not encode propagation phase in a single mode; travelling structures are therefore often split into mode pairs. SPOD isolates orthogonal modes at each temporal frequency using the cross-spectral density and is powerful for statistically stationary flows and harmonic coherent structures, but it imposes spectral purity, effectively infinite temporal support, and frequency-by-frequency decomposition. HPOD is presented as an alternative when local or instantaneous wave characteristics, amplitude modulation, frequency modulation, and intermittency are of primary interest [2507.02487].

The contrast with DMD is also explicit. DMD seeks modes tied to linear dynamics with exponential time behavior and is often interpreted through Koopman theory; HPOD does not attempt to fit a linear propagator, but instead extracts energy-ranked complex wavepacket structures from analytic data. The literature discussion also places HPOD alongside shifted POD, centering or template-based symmetry reduction, characteristic DMD, permuted POD, and spectral or multi-scale POD variants, while stressing that HPOD handles translation symmetry through analytic-signal complexification rather than explicit snapshot shifting [2507.02487].

## 5. Empirical validation on DNS, LES, and PIV

The first validation dataset is a 2D DNS of a laminar bluff-body wake at \(Re=100\), with \(10{,}800\) time-resolved snapshots on a cropped domain \(10.6D\times4.8D\) and \(91\times201\) grid points. In this case, both conventional HPOD and space-only HPOD extract the von Kármán vortex street as a single complex travelling structure, whereas standard POD splits it into the expected first pair of real modes. The two HPOD variants produce almost identical spatial modes and temporal phase portraits, differing mainly by a phase sign or conjugation convention. The same study also reports FFT-induced edge effects and shows that removing about 15% at each end cleans the modes considerably [2507.02487].

A deliberately shuffled version of the DNS dataset is used to mimic non-time-resolved measurements. Under temporal shuffling, space-only HPOD is essentially unchanged, while conventional HPOD fails to complexify correctly and reverts toward POD-like behavior. This is one of the clearest demonstrations of why the space-only variant matters for snapshot-style experiments [2507.02487].

The second validation dataset is a high-fidelity LES of a turbulent jet at \(M=0.9\) and \(Re_j\approx 10^6\), with \(10{,}000\) time-resolved snapshots. Here the target structures are broadband, modulated, intermittent wavepackets in the shear layer rather than nearly harmonic shedding. Both HPOD variants extract leading complex modes corresponding to jet wavepackets. The first three modes capture about \(8.1\%\), \(7.2\%\), and \(6.0\%\) of the turbulent kinetic energy in conventional HPOD, and about \(8.4\%\), \(7.5\%\), and \(6.2\%\) in space-only HPOD. The mode magnitudes show streamwise amplification and decay, with mode 1 peaking roughly between \(x/D=10\) and \(20\), while time-frequency and space-wavenumber spectrograms reveal active and inactive periods, varying but band-limited peak frequencies, one-sided spatial spectra, and increasing dominant wavenumber with mode number [2507.02487].

The third validation dataset is a planar snapshot PIV of a subsonic jet at \(Re=33{,}000\), comprising 2600 velocity fields sampled at 10 Hz with spatial vector spacing \(0.029D\). Only space-only HPOD is meaningful in this case because the temporal resolution is effectively not useful for the dynamics. The first three modes capture approximately \(8.8\%\), \(6.7\%\), and \(5.4\%\) of turbulent kinetic energy. The extracted complex spatial modes closely resemble those of the LES: real and imaginary parts appear in spatial quadrature, the structures display wavepacket-like spatial oscillations with streamwise growth and decay envelopes, and wavenumber increases with mode rank. Temporal cyclograms are essentially random or fuzzy, consistent with temporal undersampling, while the spatial modes remain physically interpretable as advecting structures [2507.02487].

## 6. Sphere-wake application, POD pairing, and computational shortcuts

A complementary study examined HPOD on 12,600 planar 2C-2D PIV velocity fields in the near wake of a \(D=40\) mm sphere at \(Re_D=7780\). The field of view extended from just downstream of the sphere to about \(5.2D\) downstream and \(\pm 2.3D\) transversely. Since the acquisition rate was only 1 Hz while the convective time was \(t_c=0.2\) s, the dataset was not time-resolved. The authors therefore used the streamwise spatial analogue of the Hilbert transform and applied HPOD to the complex matrix
\[
X^a = X + i\,H_x[X],
\]
constructed from the two-component velocity fluctuations \(u'(x,y,t)\) and \(v'(x,y,t)\) [2507.19167].

The leading HPOD modes were found to correspond closely to pairs of standard POD modes. The first HPOD mode matched POD modes 1 and 2, the second matched POD modes 4 and 5, the third matched POD modes 6 and 8, and the fourth matched POD modes 9 and 10. The clearest pairings, \((1,2)\) and \((4,5)\), had phase offsets relative to the associated HPOD mode that were approximately \(90^\circ\) apart. Physically, the pair \((1,2)\) corresponded to a flapping wake motion; \((4,5)\) to a pulsating streamwise fluctuation pattern; \((6,8)\) to a shorter-wavelength flapping; and \((9,10)\) to a higher-order pulsating structure with small-scale near-wake features and a V-shaped downstream structure [2507.19167].

The same study emphasizes that HPOD is not neutral because the Hilbert transform acts as a directional filter. Applied directly to the fluctuation field, it accentuates structures periodic in the chosen direction and may suppress or distort structures that do not satisfy the streamwise-propagation assumption. In the reported comparisons, HPOD modes showed stronger symmetry or antisymmetry, cleaner downstream propagation, reduced near-wake non-propagating features, and modified intensity distributions. For inherently three-dimensional flow measured only in a two-dimensional plane, this introduces an interpretive caution: a feature may be physically real but appear non-propagating or asymmetric in the plane, and HPOD may suppress it [2507.19167].

To address that issue, the paper proposes a cheaper alternative: apply the streamwise Hilbert transform to the POD modes themselves,
\[
(\psi_k)^a(\mathbf{x})=\psi_k(\mathbf{x})+i\,H_x[\psi_k(\mathbf{x})],
\]
rather than to the full field prior to decomposition. Because the Hilbert transform is linear, summing these analytic POD modes with the original coefficients reconstructs the analytic signal of the original field, but the transformed POD modes are not identical to HPOD modes because transform and decomposition are performed in opposite order. Their practical use is pair identification. By correlating each mode with the Hilbert-shifted version of another, the method identified exactly the same POD pairs as HPOD—\((1,2)\), \((4,5)\), \((6,8)\), and \((9,10)\)—with mutual matches occurring at phase shifts of approximately \(\pm 0.5\pi\). The study concludes that this alternative is much more computationally efficient and preserves the interpretive power of POD while avoiding the need for a full complex decomposition of the analytic signal [2507.19167].

## 7. Terminology, limitations, and adjacent meanings

In the recent fluid-mechanics literature, HPOD denotes the Hilbert-transform-based complex POD used to extract propagating coherent structures or wavepackets from flow-field data [2507.02487]. A potential misconception is that the acronym has a single universally fixed meaning across the broader POD literature. Other arXiv papers use closely related language to denote different, though conceptually adjacent, ideas.

One adjacent usage is Hilbert-space POD in the operator-theoretic sense. “Hierarchical Approximate Proper Orthogonal Decomposition” formulates POD on a Hilbert space \(V\) with inner product \((\cdot,\cdot)\), defines the snapshot map \(\mathcal{S}\), the Gramian \(G=\mathcal{S}^*\mathcal{S}\), and the optimality property
\[
\sum_{s\in S}\|s-P_{V_N}(s)\|^2
=
\min_{\dim X=N}\sum_{s\in S}\|s-P_X(s)\|^2,
\]
before extending it to the tree-based HAPOD algorithm [1607.05210]. Another is the generalized POD theory for PDE data in two Hilbert spaces \(X\) and \(Y\) linked by an operator \(L\), with exact error identities for seminorm-type quantities such as \(\|L(w-\Pi_r w)\|_Y\) and for non-orthogonal projections [1910.08174]. A further related direction is dissipation-based POD, formulated in a gradient-based \(H^1\)-type metric for turbulent Rayleigh–Bénard convection, where the chosen Hilbert metric is tailored to viscous and thermal dissipation rather than to \(L^2\) energy [2311.11807]. Meshless POD from scattered data likewise restores the continuous \(L^2(\Omega)\)-type inner product through radial-basis-function regression and quadrature rather than through grid-based dot products [2407.03173]. This suggests an acronym overlap rather than a single universally fixed usage.

For the Hilbert-transform-based HPOD itself, the documented limitations are specific. The transformed direction must be adequately resolved: temporal resolution for conventional HPOD, spatial resolution along the advection direction for space-only HPOD. FFT-based analytic-signal computation causes boundary corruption, so trimming, padding, or windowing are relevant remedies; the demonstrations use trimming. The equivalence of the time and space formulations holds for advective structures with a phase-velocity relation, not for arbitrary non-advective variability. HPOD is broadband rather than spectrally pure, which is advantageous for modulation analysis but a drawback when strict frequency separation is desired. Finally, the directional Hilbert transform biases the analysis toward streamwise-propagating content, so HPOD should be interpreted as a diagnostic tool that makes propagation explicit, not as a neutral representation of all flow physics [2507.02487].

Source: https://www.emergentmind.com/topics/hilbert-proper-orthogonal-decomposition-hpod