J. Fluid Mech. (2016), vol. 809, pp. 843–872.

c Cambridge University Press 2016 843

doi:10.1017/jfm.2016.678

Recursive dynamic mode decomposition of

transient and post-transient wake flows
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

Bernd R. Noack1,2, †, Witold Stankiewicz3 , Marek Morzyński3

and Peter J. Schmid4
1 Laboratoire d’Informatique pour la Mécanique et les Sciences de l’Ingénieur, LIMSI-CNRS,
Rue John von Neumann, Campus Universitaire d’Orsay, Bât 508, F-91403 Orsay, France
2 Institut für Strömungsmechanik, Technische Universität Berlin, Hermann-Blenk-Straße 37,
D-38108 Braunschweig, Germany
3 Chair of Virtual Engineering, Poznań University of Technology, ul. Jana Pawła II 24,
60-965 Poznań, Poland
4 Department of Mathematics, Imperial College London, South Kensington Campus,
London SW7 2AZ, UK

(Received 21 November 2015; revised 31 July 2016; accepted 19 October 2016;

first published online 21 November 2016)

A novel data-driven modal decomposition of fluid flow is proposed, comprising key

features of proper orthogonal decomposition (POD) and dynamic mode decomposition
(DMD). The first mode is the normalized real or imaginary part of the DMD mode
that minimizes the time-averaged residual. The Nth mode is defined recursively in an
analogous manner based on the residual of an expansion using the first N − 1 modes.
The resulting recursive DMD (RDMD) modes are orthogonal by construction, retain
pure frequency content and aim at low residual. Recursive DMD is applied to transient
cylinder wake data and is benchmarked against POD and optimized DMD (Chen
et al., J. Nonlinear Sci., vol. 22, 2012, pp. 887–915) for the same snapshot sequence.
Unlike POD modes, RDMD structures are shown to have purer frequency content
while retaining a residual of comparable order to POD. In contrast to DMD, with
exponentially growing or decaying oscillatory amplitudes, RDMD clearly identifies
initial, maximum and final fluctuation levels. Intriguingly, RDMD outperforms both
POD and DMD in the limit-cycle resolution from the same snapshots. Robustness of
these observations is demonstrated for other parameters of the cylinder wake and for
a more complex wake behind three rotating cylinders. Recursive DMD is proposed
as an attractive alternative to POD and DMD for empirical Galerkin models, in
particular for nonlinear transient dynamics.
Key words: low-dimensional models, wakes
https://doi.org/10.1017/jfm.2016.678

1. Introduction
This study proposes a novel flow field expansion tailored to the construction of low-
dimensional empirical Galerkin models. Such reduced-order models (1) help in data
compression, (2) allow quick visualizations and kinematic mixing studies (Rom-Kedar,

† Email address for correspondence: [email protected]

 844 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid
Leonard & Wiggins 1990), (3) provide a testbed for physical understanding (Lorenz
1963), (4) serve as computationally inexpensive surrogate models for optimization
(Han, Stefan & Zimmermann 2013) or (5) may be used as a plant for control design
(Gerhard et al. 2003; Bergmann & Cordier 2008).
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

In 1858, Helmholtz laid the foundation for the first low-dimensional dynamical
models in fluid mechanics with his famous theorems on vortices (see, e.g., Lugt
1995). Subsequently, a rich set of vortex models has been developed for vortex pairs,
for the recirculation zone (Föppl 1913; Suh 1993), for the vortex street (von Kármán
& Rubach 1912) and for numerous combustion related problems (Coats 1997), to
name just a few. For non-periodic open flows, low-dimensional vortex models come
at the expense of a hybrid state-space structure: new degrees of freedom (vortices) are
created at the body, merged or removed. Most forms of applications, like dynamical
systems analyses or control design, are not suitable for hybrid models but assume a
continuous evolution in a fixed finite-dimensional state space. Hence, most currently
developed reduced-order models are formulated by the Galerkin method and based
on modal expansions (Fletcher 1984).
In the last 100 years, Galerkin (1915)’s method of solving partial differential
equations has enjoyed ample generalizations and applications, from high-dimensional
grid-based computational methods to low-dimensional models. This study focuses on
low-dimensional flow representations by an expansion in global modal structures.
In principle, any space of square-integrable velocity fields has a complete set
of orthonormal modes: any flow field can be arbitrarily closely approximated
by a finite-dimensional expansion. In practice, however, the construction of such
mathematical bases is restricted to simple geometries and the use of Fourier
expansions or Chebyshev polynomials (Orszag 1971). Stability modes based on a
linearization of the Navier–Stokes equations tend to be more efficient in terms of
resolution for a given number N of modes. However, these physical modes generally
lack a proof of completeness, and are afflicted by a reduced dynamic bandwidth
and an O(N 3 ) operation count for the quadratic terms, as opposed to O(N log N)
operations for Fourier or Chebyshev modes. The most efficient representations of
a Navier–Stokes solution are obtained from empirical expansions based on flow
snapshots. These data-driven Galerkin expansions are confined to a subspace spanned
by the snapshots, i.e. have a narrow dynamic bandwidth defined by the training set.
This study focuses on data-driven expansions. A Galerkin expansion with
guaranteed minimal residual over the training snapshots was first pioneered by Lorenz
(1956) as principal axis modes and later popularized in fluid mechanics as proper
orthogonal decomposition (POD) by Lumley (1967). Proper orthogonal decomposition
guarantees an optimal data reconstruction in a well-defined sense (see, e.g., Holmes,
Lumley & Berkooz 1998). Apart from data compression applications, Lumley devised
POD as a mathematical tool for distilling coherent structures from data. Yet, only in
rare cases have POD modes been found to resemble physically meaningful structures,
like stability modes or base-flow deformations (Oberleithner et al. 2011). In particular,
they tend to mix spatial and temporal frequencies in most modes, which complicates
a physical interpretation. As a remedy for this shortcoming, the dynamic mode
decomposition (DMD) – also referred to as Koopman analysis – was pioneered by
https://doi.org/10.1017/jfm.2016.678

Rowley et al. (2009) and Schmid (2010). Dynamic mode decomposition is able to
distil stability eigenmodes from transient snapshot data and produce temporal Fourier
modes for post-transient data. The downside of these DMD features – compared with
POD – is non-orthogonality of the extracted modes, suboptimal convergence, the
need for time-resolved snapshots and numerical challenges when filtering is omitted.
Mezić (2013) provides an excellent review of recent DMD developments.
 Recursive DMD 845
In the past, numerous Galerkin models based on POD and DMD have been
constructed. In addition, numerous generalizations have been proposed to address,
among other topics, multi-operating conditions (Jørgensen, Sørensen & Brøns 2003),
changes during transients (Siegel et al. 2008), optimal correlations to observables
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

(Hoarau et al. 2006; Schlegel et al. 2012) and control design (Brunton & Noack
2015).
This study proposes a novel data-driven expansion preserving key features of POD,
such as orthonormality of the computed modes and a low residual, and of DMD,
such as distilling the dominant frequencies and their associated structures contained
in the data. The paper is organized as follows. Section 2 describes the cylinder
wake configuration and gives details on the direct numerical simulation employed as
well as the snapshots extracted. Section 3 outlines the computation of the proposed
expansion, called ‘recursive DMD (RDMD)’ in what follows. Recursive DMD is then
applied to a transient cylinder wake, demonstrating its advantages (§ 4) over previous
decompositions. In § 5, a similar investigation is performed for the non-periodic wake
behind three rotating cylinders. Our results are summarized in § 6.

2. Configuration and direct numerical simulation

As a test case, a two-dimensional incompressible viscous flow past a circular
cylinder has been chosen. The flow is described in a Cartesian coordinate system
where the x-axis is aligned with the streamwise direction and the y-axis is transverse
and orthogonal to the cylinder axis. The origin of the coordinate system coincides
with the cylinder axis. The location vector is denoted by x = (x, y). Analogously, the
velocity is represented by u = (u, v), where u and v are the x- and y-components of
the velocity. The time is represented by t. The Newtonian fluid is characterized by
the density ρ and dynamic viscosity µ. In the following, all variables are assumed to
be non-dimensionalized by the cylinder diameter D, the oncoming velocity U and the
fluid density ρ. The resulting Reynolds number Re = UDρ/µ is set to 100, i.e. well
above the onset of vortex shedding (Zebib 1987; Schumm, Berger & Monkewitz
1994), but well below the onset of three-dimensional instabilities (Zhang et al. 1995;
Barkley & Henderson 1996; Williamson 1996).
A direct numerical simulation of the Navier–Stokes equations has been performed
using an in-house solver based on a second-order finite-element discretization with
Taylor–Hood elements (Taylor & Hood 1973) in primitive variables. The time stepping
is performed using a third-order-accurate scheme and a time step equal to 0.1.
Following Noack et al. (2003), the computational domain extends from x = −5 to
x = 25 and y = −5 to y = 5 and is discretized with 56 272 finite elements. The initial
condition for the transient from t = 0 to t = 150 is a small perturbation of the steady
Navier–Stokes solution us and reads u(x, 0) = us (x) + 0.01f 1 (x), where f 1 is the real
part of the unstable complex eigenvector normalized to unit norm.
The simulation provides M = 450 equidistantly sampled velocity snapshots um (x) =
u(x, tm ), m = 1, . . . , M, covering the entire unforced transient phase, from the steady
solution to the fully developed von Kármán vortex street. The snapshot times are
https://doi.org/10.1017/jfm.2016.678

tm = 30 + m1t based on a sampling interval of 1t = 0.2 and cover the time interval
t ∈ [30, 120]. The initial and final base flows are depicted in figure 1 together with the
energy norm of the fluctuations v = u − us around the steady solution us . The transition
from the unstable fixed point to the stable limit-cycle oscillation is characterized by
the growing amplitude of the fluctuations accompanied by an increase in the Strouhal
number (Zebib 1987; Schumm et al. 1994).
 846 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid

(a) (b)
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

(c)
Selected range of data

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150

t

F IGURE 1. (Colour online) Incompressible two-dimensional flow around a circular

cylinder at Re = 100. The steady (a) and time-averaged (b) solutions are illustrated with
streamlines. The total fluctuation level K normalized with its asymptotic value K∞ is
shown in (c).

Within the first 30 convective time units, the flow is governed by linear dynamics
spanned by the steady solution and the unstable eigenmode. In the intermediate
phase, approximately t ∈ [60, 90], the vortex shedding undergoes significant changes
and moves farther upstream towards the cylinder. In the final 30 convective time units,
the flow has converged to a limit-cycle dynamics exhibiting a fully developed von
Kármán vortex street. These three characteristic stages of the transient evolution are
depicted in figure 2. For the construction of the Galerkin expansions, we ignore the
phase t ∈ [0, 30] as there is hardly any noticeable fluctuation and the corresponding
snapshots effectively duplicate the steady Navier–Stokes solution. We also ignore the
converged limit-cycle data at t > 120 as this would also give emphasis to redundant
information.
All modal decompositions are based on these fluctuations around the steady solution.
The mean flow is discarded as a base flow because it is only well defined for this
particular initial condition and the chosen time interval. Our sampling Strouhal
frequency of 10 is approximately 30 times larger than the shedding frequency – a
value that can be considered adequate for a Fourier transformation while avoiding
excessive redundancy for the statistical POD.

3. Modal decomposition
In this section, a new snapshot-based modal decomposition is proposed. This
decomposition comprises properties of the POD (Lumley 1967; Sirovich 1987) and the
DMD (Rowley et al. 2009; Schmid 2010). First (§§ 3.1 and 3.2), the snapshot-based
POD and DMD are briefly recapitulated. In § 3.3, the RDMD is proposed as an
https://doi.org/10.1017/jfm.2016.678

appealing compromise inheriting the orthonormality and low truncation error of POD
and the oscillatory representation of the flow behaviour of DMD.

3.1. Proper orthogonal decomposition

We analyse a time-dependent velocity field u(x, t) in a steady domain x ∈ Ω and
sample M flow snapshots with constant sampling frequency corresponding to a time
 Recursive DMD 847

(a) (b) (c)

Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

F IGURE 2. (Colour online) Transient evolution of the cylinder wake illustrated with
snapshots of the vorticity field at an initial (a, t = 50), an intermediate (b, t = 70) and
a final state (c, t = 90).

step 1t. The velocity field at instant tm = m1t, with m = 1, . . . , M, is denoted by

um := u(x, tm ).
Proper orthogonal decomposition requires the definition of an inner product and a
time average. We assume the standard inner product of two square-integrable velocity
fields u1 , u2 ∈ L2 (Ω), given as
Z
(u1 , u2 )Ω := dx u1 (x) · u2 (x). (3.1)
Ω
√
The corresponding norm reads kukΩ := (u, u)Ω . The snapshot-based time average of
a function F is defined in a canonical manner,
M
1 X
hF(u)iM := F(um ). (3.2)
M m=1

In the following, the snapshot-based POD (Sirovich 1987) is applied to the

fluctuations
v m (x) := u(x, tm ) − us (x) (3.3)
around the steady solution us for the reasons mentioned in § 2. First, the correlation
matrix C = (C mn ) ∈ RM×M of the fluctuations is determined,
1 m n
C mn := (v , v )Ω . (3.4)
M
Second, a spectral analysis of this matrix is performed. It should be noted that C is
a symmetric positive semi-definite Gramian matrix. Hence, its eigenvalues λi are real
and non-negative and can be assumed to be sorted according to λ1 > λ2 > . . . > λM > 0.
The corresponding eigenvectors ei = [e1i , . . . , eM T
i ] satisfy

C ei = λi ei , i = 1, . . . , M, (3.5)
and can – without loss of generality – be assumed to be orthonormalized, satisfying
ei · ej = δij , with δij as the Kronecker symbol. Third, each POD mode is expressed as
a linear combination of the snapshot fluctuations,
M
1 X m m
e v ,
https://doi.org/10.1017/jfm.2016.678

ui := √ i = 1, . . . , N. (3.6)
Mλi m=1 i

It follows that the POD modes are orthonormal, with (ui , uj )Ω = δij , ∀i, j ∈
{1, . . . , M}. Finally, the mode amplitudes read

ami := λi Meim , i = 1, . . . , N.
p
(3.7)
 848 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid
These amplitudes are uncorrelated (orthogonal in time), or in mathematical terms

hai aj iM = λi δij , i, j ∈ {1, . . . , N}. (3.8)

Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

It should be noted that the first moments hai iM do not need to vanish as the
fluctuations are based on the steady solution and not on the mean flow of this transient.
The POD defines a second-order statistics providing the two-point autocorrelation
function
N
X
v(x, t)v(y, t) = λi ui (x)ui (y). (3.9)
i=1

Hence, a minimum requirement imposed on the snapshot ensemble is the accuracy

of the extracted mean flow and the second moments of the flow. This accuracy of
the statistics for a given number of snapshots is increased by processing uncorrelated
snapshots, as required in the original paper on the snapshot POD method (Sirovich
1987).
The resulting expansion exactly reproduces the snapshots for N = M modes; we have
N
u(x, tm ) = us (x) + ai (tm )ui (x).
X
(3.10)
i=1

For N < M, the truncated expansion (3.10) has a non-vanishing residual rm (x) :=
r(x, tm ). The corresponding time-averaged truncation error

χ 2 := hkrm k2Ω iM (3.11)

can be shown to be minimal; in other words, no other Galerkin expansion with the
same number of modes will have a smaller error (Holmes et al. 1998). This optimality
property makes POD an attractive data compression technique.
For later reference, the instantaneous truncation error χ 2 (t) := krm (·, t)k2Ω is
introduced. The size of this error may be compared with the corresponding fluctuation
level on the limit cycle 2K∞ = kvk2 , where K∞ denotes the turbulent kinetic energy
(TKE) and the overbar represents averaging over the post-transient phase. We also
introduce K as the corresponding instantaneous quantity.
As a motivation for the proposed new decomposition, we recall that POD can also
be defined in a recursive manner, following Courant & Hilbert (1989) on the spectral
analysis of positive definite symmetric matrices. Taking u1 as the first expansion mode,
the resulting one-mode expansion reads

v m = am1 u1 + rm1 , m = 1, . . . , M. (3.12)

The mode amplitude am1 := (v m , u1 )Ω minimizes the residual krm1 kΩ for a given
u1 . The first POD mode can be shown to minimize the averaged energy of the
residual hkrm1 k2Ω iM . Furthermore, the residual rm1 , m = 1, . . . , M is orthogonal to u1
https://doi.org/10.1017/jfm.2016.678

by construction. The second step then searches for a mode u2 that best resolves the
residual r1 ,
rm1 = am2 u2 + rm2 , m = 1, . . . , M, (3.13)
i.e. that minimizes hkrm2 k2Ω iM . The other remaining modes are computed in a similar
manner.
 Recursive DMD 849
3.2. Dynamic mode decomposition
Dynamic mode decomposition (Rowley et al. 2009; Schmid 2010) is another
data-driven modal expansion which can approximate stability eigenmodes from
transient data or Fourier modes from post-transient data. The time step 1t needs to
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

be sufficiently small for a meaningful Fourier analysis, but sufficiently large so that
the changes in the flow state exceed the noise level.
We consider the fluctuation snapshots of § 3.1, v m , m = 1, . . . , M − 1, as linearly
independent modes of an expansion and write
M−1
bi v i .
X
v= (3.14)
i=1

Evidently, the modes are generally not orthogonal. With this basis, the mode amplitude
vector of the mth snapshot becomes a unit vector,

bm = [bm1 , . . . , bmM−1 ]T , where bmi = δim . (3.15)

Dynamic mode decomposition assumes a linear relationship between the (m + 1)th and
mth snapshot,
bm+1 = Abm , (3.16)
where A ∈ R(M−1)×(M−1) is a square matrix which is generally identified from the data.
For m < M − 1, (3.16) acts as a shift map: the mth unit vector is mapped on the
(m + 1)th unit vector, leading to unity in the first subdiagonal of A and zeros in
the remaining column. The Mth snapshot has no successor in the time series and is
expanded in terms of the previous snapshots,
M−1
vM = ci v i + r.
X
(3.17)
i=1

The coefficients ci are chosen to minimize the residual norm krkΩ .

From (3.16) and (3.17), the matrix is easily seen to be of companion type,

0 0 ... 0 0 c1
 
 1 0 ... 0 0 c2 
 . . .
A= . . ... .. (3.18)

 .. ..
.
. 
 0 0 ... 1 0 cM−2 
0 0 ... 0 1 cM−1

In what follows, we depart from the classical DMD literature and propose a simpler
derivation of the DMD modes. Let P(s) = c1 + c2 s + · · · + cM−1 sM−2 + sM−1 be a
polynomial in s and let P(s) = (s − λ1 )(s − λ2 ) · · · (s − λM−1 ) be its factorization with
distinct eigenvalues λi . We introduce
https://doi.org/10.1017/jfm.2016.678

1 λ1 . . . λM−2
 
1
 1 λ2 . . . λM−2 2

V = . (3.19)
 
. .
. .
. .
. 
 . . . . 
1 λM−1 . . . λM−2
M−1
 850 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid
as the corresponding Vandermonde matrix. It can be easily verified that the
Vandermonde matrix V diagonalizes the companion matrix A. We obtain
V AV −1 = diag(λ1 , . . . , λM−1 ), (3.20)
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

where the right-hand side is a diagonal matrix with the eigenvalues as its elements.
We introduce new variables a defined by
a = V b. (3.21)
With these definitions, the evolution equation (3.16) can be cast into eigenform
according to
am+1 = V CV −1 am = D am . (3.22)
Here, D denotes the diagonal matrix
λ1 0 . . . 0
 
 0 λ2 . . . 0 
D = diag(λ1 , . . . , λM−1 ) =  .
 .. . . ..  , (3.23)
.. .. . 
0 0 . . . λM−1
T
where the ith eigenvalue is λi , with corresponding eigenvector ai = δ1i , . . . , δ(M−1)i .

Equations (3.15), (3.16) and (3.21) imply that the snapshots can be expressed as
M−1
vm = λm−1
X
i Φi , (3.24)
i=1

where the λi are referred to as DMD eigenvalues and the Φi as (complex) DMD
modes. It should be noted that the derivation of (3.24) rests on the diagonalization
of the companion matrix (3.20) and does not require the Koopman operator.
The choice of N DMD modes for the Galerkin expansion (3.10) minimizes the
truncation error (3.11) following Chen, Tu & Rowley (2012) and consistent with the
optimal property of POD.

3.3. Recursive DMD

Recursive DMD serves a multi-objective task: extracting oscillatory modes from the
snapshot sequence (like DMD) while ensuring orthogonality of the modes and a low
truncation error (like POD).
The initialization step prepares the residual to be processed. We take
rm0 := v m , m = 1, . . . , M. (3.25)
During the ith step (0 < i 6 N), the ith mode is determined from a DMD
M−1
rmi−1 = λm−1 Φji ,
X
(3.26)
https://doi.org/10.1017/jfm.2016.678

j
j=1

where Φji represents the jth DMD mode of the ith step. The candidate modes to be
considered are
<{Φki }
<{Φki } , k = 1, . . . , M − 1,
u?k = (3.27)

Ω
 Recursive DMD 851
and each mode reduces the residual rmi,k according to

rmi−1 = a?m
k uk + ri,k ,
? m
(3.28a)
a?mk = (rmi−1 · u?k )Ω . (3.28b)
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

It should be noted that the residual rmi,k depends on the index of iteration i, the
snapshot index m and the trial mode index k. The truncation error of the ith candidate
mode u?k for all snapshots is given by
2
D E
2
χi,k := rmi,k Ω . (3.29)
M

As mode i we select the trial mode k with the lowest averaged error, i.e.
2
ui := u?k so that ∀l ∈ {1, . . . , M − 1} : χi,k 6 χi,l2 . (3.30)

The resulting expansion reads after the ith step

i
v m := aj uj + rmi .
X
(3.31)
j=1

The (i + 1)th mode is computed following the same steps. The iteration terminates
when the desired number of modes is reached, i = N, or in the unlikely case that all
residuals vanish, rmi ≡ 0, m = 1, . . . , M – whichever criterion is satisfied first. We
emphasize that the candidate modes are normalized (see (3.27)) and the (i + 1)th mode
is constructed from the orthogonal residual of the Galerkin expansion with the first i
modes (see (3.28)). Hence, the modes form an orthonormal system.

4. Modal decomposition of the transient cylinder wake

The transient cylinder wake is analysed with POD (§ 4.1), DMD (§ 4.2) and
the proposed new decomposition (§ 4.3) discussed in § 3. In § 4.4, all modal
decompositions are subjected to a comparison with respect to the instantaneous
residual, the averaged residual and the convergence with increasing number of modes.
The transient wake snapshots are the same for all decompositions and have been
described in § 2.

4.1. Proper orthogonal decomposition

The snapshots of the cylinder wake simulations described in § 2 are subjected to
a snapshot POD outlined in § 3.1. The POD modes of the post-transient periodic
cylinder wake mimic a Fourier decomposition. They arise in pairs representing the
two phases at the first and higher harmonic frequencies (Deane et al. 1991; Noack
et al. 2003). The energy level of each pair rapidly decreases with the order of
the harmonics. The transient, however, exhibits a gradual change from the initial
https://doi.org/10.1017/jfm.2016.678

stability modes (with a fluctuation maximum far from the cylinder) to von Kármán
vortex shedding (with fluctuations peaking near the cylinder). The other change is
a base flow with a recirculation region that decreases in streamwise extent from
approximately 7 diameters to approximately 1 diameter length. The resulting POD
modes and associated mode amplitudes are depicted in figures 3 and 4 and appear
to be different from the post-transient analogues. Proper orthogonal decomposition
 852 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid

(a) (b)
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

(c) (d)

(e) (f)

(g) (h)

(i) ( j)

F IGURE 3. (Colour online) The first 10 POD modes of the transient phase of the flow,
with t ∈ [30, 120] and Re = 100. (a,c,e,g,i) Modes 1–5, (b,d,f,h,j) modes 6–10. The vorticity
of the mode is visualized in colour (green, zero; red, above a positive threshold; blue,
below a negative threshold).

modes 1 and 2 effectively represent von Kármán vortex shedding growing from zero
to asymptotic values. These values are larger than the corresponding fluctuation levels
of the other modes. Modes 6 and 7 describe the second harmonics, as can be seen
from the visualization and the amplitude evolution. Modes 9 and 10 look similar
to the stability eigenmodes at slightly lower frequencies, with peak activity around
t = 55. Modes 4 and 5 also represent vortex shedding, with peak activity around t = 65,
i.e. two shedding periods later. Mode 3 depicts the shift mode (Noack et al. 2003),
i.e. it characterizes the base-flow change between steady and time-averaged periodic
solution. Mode 8 represents another base-flow correction with different topology and
is mainly active during the most rapid changes of the transient around t = 65. It has
a small second harmonic component. The base-flow modes 3 and 8 correspond to an
expansion with the steady solution as basic modes. Taking the post-transient averaged
https://doi.org/10.1017/jfm.2016.678

velocity as basic mode is numerically found to give rise to similar base-flow modes
but with different energy content and thus different indices.
All of the displayed POD modes show pronounced frequencies: six modes display
oscillations near the shedding frequency, two resolve the second harmonics and
two feature slowly varying base-flow changes. Unlike POD for periodic shedding,
the third and higher harmonics are found as modes with indices after the first 10
 Recursive DMD 853

(a) 3 (b) 0.4

2 0.2
1 0
0 –0.2
–1 –0.4
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

–2 –0.6
–3 –0.8
30 40 50 60 70 80 90 100110120 30 40 50 60 70 80 90 100110120

(c) 3 (d ) 0.6
2 0.4
1 0.2
0 0
–1 –0.2
–2 –0.4
–3 –0.6
30 40 50 60 70 80 90 100110120 30 40 50 60 70 80 90 100110120

(e) 0 (f) 1.0

–0.5 0.8
0.6
–1.0
0.4
–1.5 0.2
–2.0 0
–2.5 –0.2
30 40 50 60 70 80 90 100110120 30 40 50 60 70 80 90 100110120

(g) 1.5 (h) 0.6

1.0 0.4
0.5 0.2
0 0
–0.5 –0.2
–1.0 –0.4
–1.5 –0.6
30 40 50 60 70 80 90 100110120 30 40 50 60 70 80 90 100110120

(i) 1.5 ( j ) 0.6

1.0 0.4
0.5 0.2
0 0
–0.5 –0.2
–1.0 –0.4
–1.5 –0.6
30 40 50 60 70 80 90 100110120 30 40 50 60 70 80 90 100110120

F IGURE 4. The mode amplitudes of the POD modes of figure 3.

ones displayed. This comparatively ‘clean’ frequency content may be attributed

to the narrow-bandwidth transient dynamics, with dominant frequencies between the
eigenfrequency of the steady solution and the shedding frequency of the post-transient
wake. Moreover, the maximum of the fluctuation envelope moves upstream during
the transient. For broadband dynamics, such as a mixing layer with multiple vortex
pairings, POD modes with multiple frequency content are far more common (Noack
et al. 2004).
https://doi.org/10.1017/jfm.2016.678

4.2. Dynamic mode decomposition

The snapshots of the transient flow have been decomposed with DMD. The DMD
procedure can identify eigenmodes for the transient data and the Fourier modes for
the attractor.
 854 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid

(a) (b)
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

(c) (d )

(e) (f)

(g) (h)

(i) ( j)

F IGURE 5. (Colour online) The same as figure 3 but for the first 10 DMD modes.

The result of the procedure is a set of complex Ritz vectors and complex
eigenvalues characterizing the growth rate and the frequency of the respective mode.
The first 10 eigenmodes are depicted in figure 5. The first DMD mode corresponds
to a real eigenvalue leading to a real Ritz vector. The remaining modes represent
the real and imaginary parts of complex Ritz vectors. The phase-shifted analogue
of oscillatory mode 10 is mode 11 (not displayed). In figure 6, the corresponding
amplitudes of the real modes are displayed.
Modes 1, 6 and 7 act as shift modes and resolve slow base-flow changes. This
interpretation is corroborated by the behaviour of the mode amplitudes. The remaining
modes describe vortex shedding (i = 2, 3, 4, 5, 10) or its second harmonics (i = 8, 9).
The oscillatory mode amplitudes are slowly growing, like the first vortex shedding
pair (for i = 2, 3) and the second harmonics (for i = 8, 9), or slowly decaying (for
i = 4, 5, 10). Intriguingly, the DMD modes describing vortex shedding have nearly
identical frequencies and nearly identical shapes. This implies a redundancy which
https://doi.org/10.1017/jfm.2016.678

constitutes a challenge for reduced-order modelling.

By construction, the mode amplitudes have an exponentially growing or decaying
envelope and can thus give no indication of initial or asymptotic values or temporal
periods of maximum activity. In particular, an extrapolation beyond the sampling
interval t ∈ [30, 120] is not meaningful. An additional potential challenge for
reduced-order modelling is posed by the non-orthogonality of the DMD modes.
 Recursive DMD 855

(a) –0.270 (b) 0.40

–0.272 0.30
–0.274 0.20
0.10
–0.276 0
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

–0.278 –0.10
–0.280 –0.20
30 40 50 60 70 80 90 100 110 120 30 40 50 60 70 80 90 100110120

(c) 0.20 (d ) 0.40

0.15 0.30
0.10
0.05 0.20
0 0.10
–0.05 0
–0.10
–0.15 –0.10
–0.20 –0.20
30 40 50 60 70 80 90 100 110 120 30 40 50 60 70 80 90 100110120

(e) 0.20 (f) 0.15

0.15 0.10
0.10
0.05 0.05
0 0
–0.05 –0.05
–0.10
–0.15 –0.10
–0.20 –0.15
30 40 50 60 70 80 90 100 110 120 30 40 50 60 70 80 90 100110120

(g) 0.40 (h) 0.15

0.30 0.10
0.20
0.10 0.05
0 0
–0.10 –0.05
–0.20
–0.30 –0.10
–0.40 –0.15
30 40 50 60 70 80 90 100 110 120 30 40 50 60 70 80 90 100110120

(i) 0.40 ( j) 0.60

0.30 0.40
0.20
0.10 0.20
0 0
–0.10 –0.20
–0.20
–0.30 –0.40
–0.40 –0.60
30 40 50 60 70 80 90 100 110 120 30 40 50 60 70 80 90 100110120

F IGURE 6. The same as figure 4 but for the first 10 DMD modes.

Already in the early literature (Rowley et al. 2009; Schmid 2010; Chen et al.
2012), DMD has been shown to accurately capture the onset of fluctuations in the
linear regime or the post-transient behaviour on the attractor. The current results
indicate difficulties of the DMD concerning the modal interpretation for a complete
transient from the steady solution to the post-transient attractor. This issue has also
been pointed out and analysed by Bagheri (2013). In the next section, we present an
alternative DMD decomposition which addresses and removes the above-mentioned
https://doi.org/10.1017/jfm.2016.678

challenges.

4.3. Recursive DMD

In this section, the results of RDMD are presented. The procedure is demonstrated for
the same set of snapshots as employed previously for POD and DMD.
 856 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid

(a) (b)
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

(c) (d)

(e) (f)

(g) (h)

(i) ( j)

F IGURE 7. (Colour online) The same as figure 3 but for the first 10 RDMD modes.

The first 10 RDMD modes are depicted in figure 7. Intriguingly, RDMD resolves
the first harmonics (modes i = 1, 2), the second harmonics (i = 6, 7) and the
third harmonics (i = 8, 9). The associated mode amplitudes (see figure 8) show
corresponding oscillations starting near zero and reaching asymptotic values on the
limit cycle. Mode 3 shows a nearly pure shift mode, describing the transition of the
base flow from the steady solution to the time-averaged flow. In contrast to Noack
et al. (2003), the amplitude becomes negative, since the sign of the RDMD mode is
arbitrary. Modes 4 and 5 resolve intermediate vortex shedding patterns with maximum
activity around t = 70 and rather small residual fluctuations on the limit cycle. Modes
10 and 11 (not shown) are reminiscent of stability eigenmodes and peak near t = 55,
i.e. more than two periods earlier.
The first seven modes have significant similarities with the POD modes of figures 3
and 4. Yet, the oscillations are more symmetric (compare RDMD and POD mode 6)
and show no apparent frequency mixing, as in POD modes i = 3, 6, 8. Recursive DMD
https://doi.org/10.1017/jfm.2016.678

modes have by definition a lower averaged residual, compared with POD modes, but
they look much cleaner and even reveal the third harmonics. It appears that RDMD
modes have more in common with POD modes than with DMD modes. This should
not come as a surprise, since the primary construction principle is an orthonormal
decomposition leading to minimization of the residual, while the secondary criterion
is the emulation of single-frequency DMD-mode behaviour.
 Recursive DMD 857

(a) 3.0 (b) 0.4

2.0
1.0 0.2
0 0
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

–1.0 –0.2
–2.0 –0.4
–3.0
30 40 50 60 70 80 90 100110120 30 40 50 60 70 80 90 100110120

(c) 3.0 (d )
2.0 0.4
1.0 0.2
0 0
–1.0 –0.2
–2.0
–0.4
–3.0
30 40 50 60 70 80 90 100110120 30 40 50 60 70 80 90 100110120

(e) 0 ( f ) 0.3
–0.5 0.2
–1.0 0.1
0
–1.5 –0.1
–2.0 –0.2
–2.5 –0.3
30 40 50 60 70 80 90 100110120 30 40 50 60 70 80 90 100110120

(g) 2.0 (h) 0.3

1.5 0.2
1.0
0.5 0.1
0 0
–0.5 –0.1
–1.0
–1.5 –0.2
–2.0 –0.3
30 40 50 60 70 80 90 100110120 30 40 50 60 70 80 90 100110120

(i) 2.0 ( j) 0.6

1.5 0.4
1.0 0.2
0.5 0
0
–0.5 –0.2
–1.0 –0.4
–1.5 –0.6
–2.0 –0.8
30 40 50 60 70 80 90 100110120 30 40 50 60 70 80 90 100110120

F IGURE 8. The same as figure 4 but for the first 10 RDMD modes.

4.4. Comparison of POD, DMD and RDMD

In this section, results from POD, DMD and RDMD are quantitatively compared.
Figure 9 shows the truncation error of the POD, DMD and RDMD expansions as a
function of the number of modes. As expected, all errors decrease monotonically with
the number of modes, and POD outperforms DMD and RDMD. The RDMD residual,
https://doi.org/10.1017/jfm.2016.678

however, follows the POD value remarkably well and stays within similar orders of
magnitude. In contrast, DMD approaches an effective asymptote near 10 % of the final
fluctuation level.
The temporal evolution of the instantaneous truncation error is displayed in
figure 10 for POD, DMD and RDMD at N = 10. The maximum value in t ∈ [60, 70]
is lowest for POD and largest for DMD. Recursive DMD performs, as expected,
 858 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid

100
DMD
POD
RDMD
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

10–1

10–2

0 10 20 30 40 50 60 70
N

F IGURE 9. The time-averaged fluctuation level of the residual (3.11) for POD, DMD and
RDMD with increasing number of modes. The value is normalized by the corresponding
fluctuation level (2K∞ ) on the limit cycle.

0.5
10 DMD modes
10 POD modes
10 RDMD modes
0.4

0.3

0.2

0.1

0
30 40 50 60 70 80 90 100 110 120
t

F IGURE 10. The instantaneous normalized truncation error (3.11) as a function of time
for 10 POD, DMD and RDMD modes. The figure displays the sampling interval.

between the alternative expansions. A surprising feature is the asymptotes. Proper

orthogonal decomposition and DMD have similar truncation errors near 10 % of
the terminal fluctuation level, while RDMD outperforms both with a final value
https://doi.org/10.1017/jfm.2016.678

at approximately 5 %. This is no contradiction to the optimal property of POD,

as this property only guarantees a minimal time-averaged value, or, equivalently,
a minimal value of the integral over the instantaneous truncation error, which it
evidently has. The DMD-based frequency filtering included in RDMD appears to
have ‘anticipated’ the limit-cycle behaviour. One indication of this ‘anticipation’ is
the third harmonics which are featured in RDMD but absent in both POD and DMD.
 Recursive DMD 859

(a) (b)
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

(c) (d )

(e) (f)

(g) (h)

(i) ( j)

F IGURE 11. (Colour online) The same DMD modes as figure 5 but for the transient
cylinder with t ∈ [30; 120] and at Re = 120.

At this stage, RDMD appears to be more suitable for reduced-order modelling when
compared with POD or DMD. Like POD, RDMD yields orthonormal modes, but
with purer frequency content. Like DMD, RDMD extracts oscillatory modes, but
with well-defined initial and asymptotic behaviour of the mode amplitudes. In this
sense, a strength of RDMD is that its amplitudes depart from oscillatory behaviour
with exponential growth or decay of the DMD decomposition. In addition, the
cycle-to-cycle variation of amplitude and frequency of RDMD coefficients is smaller
than for similar POD mode coefficients.
The relative performance of POD, DMD and RDMD is rarely affected by a small
change of the interval time under investigation. A numerical study with numerous
time intervals indicates that the first modes keep their spatial structure but may have
different energy content. Even significant changes of the interval only give rise to
https://doi.org/10.1017/jfm.2016.678

changes in the relative order of the modes. For instance, shifting the investigation
interval to include more converged limit-cycle data decreases the energy content of
slowly varying base-flow modes while higher harmonics become more dominant. In
the extreme case of snapshots from the converged limit cycle, POD, DMD and RDMD
become very similar. An opposite observation is made for intervals focusing on the
initial transient.
 860 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid

(a) 0.240 (b) 0.20

0.15
0.238 0.10
0.236 0.05
0
0.234 –0.05
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

0.232 –0.10
–0.15
0.230 –0.20
30 40 50 60 70 80 90 100 110 120 30 40 50 60 70 80 90 100 110 120

(c) 0.20 (d) 0.20

0.15 0.15
0.10 0.10
0.05 0.05
0 0
–0.05 –0.05
–0.10 –0.10
–0.15 –0.15
–0.20 –0.20
30 40 50 60 70 80 90 100 110 120 30 40 50 60 70 80 90 100 110 120
(e) 0.20 ( f ) 0.80
0.15 0.60
0.10 0.40
0.05 0.20
0 0
–0.05 –0.20
–0.10 –0.40
–0.15 –0.60
–0.20 –0.80
30 40 50 60 70 80 90 100 110 120 30 40 50 60 70 80 90 100 110 120

(g) 1.00 (h) 0.80

0.75 0.60
0.50 0.40
0.25 0.20
0 0
–0.25 –0.20
–0.50 –0.40
–0.75 –0.60
–1.00 –0.80
30 40 50 60 70 80 90 100 110 120 30 40 50 60 70 80 90 100 110 120

(i) 1.00 ( j) 0.15

0.75 0.10
0.50
0.25 0.05
0 0
–0.25 –0.05
–0.50
–0.75 –0.10
–1.00 –0.15
30 40 50 60 70 80 90 100 110 120 30 40 50 60 70 80 90 100 110 120

F IGURE 12. The modal amplitudes of the DMD modes of figure 11.

4.5. Reconstruction of the flow at Re = 120

In the previous section (4.4), the data-driven Galerkin expansions have been evaluated
for the training data. The residuals of these modal expansions are now tested for
another data set: the transient cylinder wake at Re = 120 with the same initial
conditions during the time t ∈ [30, 120].
The change of the wake transient may be appreciated when the new DMD mode
https://doi.org/10.1017/jfm.2016.678

basis displayed in figure 11 is compared with the Re = 100 reference in figure 5.

While the first five modes have similar frequency content (see figure 12), modes 6–10
are distinctly different. For instance, the quasi-steady modes 6 and 7 of the Re =
100 reference have been replaced by the second harmonics for Re = 120. The third
harmonics do not appear in the top 10 of the reference data but are found as mode 10
for Re = 120.
 Recursive DMD 861

1.00
10 DMD
10 DMD
0.50 10 POD
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

10 RDMD

0.10

0.05
0 10 20 30 40 50 60 70
N

F IGURE 13. The time-averaged fluctuation level of the residual (3.11) for POD, DMD and
RDMD (obtained from the transient at Re = 100), and DMD at Re = 120 with increasing
number of modes. The value is normalized by the corresponding fluctuation level (2K∞ )
on the limit cycle.

As cross-validation, the POD, DMD and RDMD expansions obtained from the
transient wake at Re = 100 are employed to resolve the transient wake at Re = 120.
First, the residuals for the whole transients are plotted as a function of the number
of modes in figure 13. The residuals of POD and RDMD decrease similarly to a low
fluctuation level as the number of modes increases to 64. The observation that RDMD
outperforms POD particularly for high-order expansion is not a contradiction to the
optimality property of POD, as the residual is determined on new data. Intriguingly,
DMD shows a particularly poor performance compared with POD and RDMD. As
expected, the residual of DMD decreases when the training data (Re = 100) are
replaced by the testing data (Re = 120). However, we emphasize that RDMD based
on training data at Re = 100 performs much better for the new data at Re = 120
compared with DMD obtained and tested for these new data. In other words, RDMD
obtained for some operating condition appears to be robust against new data from
changing operating conditions.
Figure 14 shows the residual for the four modal expansions of figure 13 as a
function of time for a modal order of N = 10. The maximum instantaneous residuals
are ordered like the averaged residual values at N = 10: POD beats RDMD and
RDMD beats DMD. The DMD obtained for the new transient data shows the largest
maximum residual but is more accurate on the limit cycle.
Similar time-dependent residual behaviour is observed for expansions with more
modes, except that RDMD is more often observed to outperform POD, keeping in
mind that the expansions are obtained from Re = 100 data but tested for the Re = 120
transient.
https://doi.org/10.1017/jfm.2016.678

5. Modal decomposition of the wake past three rotating cylinders

The performance of POD, DMD and RDMD is investigated for a non-periodic flow
behind three rotating cylinders, which extends the previous cylinder wake simulations
to a more complex configuration. All cylinders have unit diameter D = 1. Their
 862 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid

0.6
10 DMD modes
10 DMD modes
0.5
10 POD modes
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

10 RDMD modes
0.4

0.3

0.2

0.1

0
30 40 50 60 70 80 90 100 110 120
t

F IGURE 14. The instantaneous normalized truncation error (3.11) as a function of time for
10 POD, DMD and RDMD modes obtained for Re = 100, and 10 DMD modes from Re =
120 transient data. The figure displays the same sampling interval as used for figure 13.

(a) (b)
15D

F IGURE 15. (Colour online) Flow past three rotating cylinders. Details of the test case
(a) and the instantaneous vorticity field (b).

centres are on the vertices of an equilateral triangle with an edge length of 1.5D.
This triangle is centred at the origin and has one vertex on the negative x-axis and two
vertices at same x-value downstream. The two rear cylinders rotate counterclockwise
with tangential √velocity vT = 1. The front cylinder rotates clockwise, and its tangential
velocity vT = ( 5 − 1)/2 ≈ 0.618 (figure 15a). Due to the rotation speeds, the wake
flow is intentionally not symmetrical. Figure 15(b) depicts the vorticity field.
In this section, we investigate the converged post-transient flow. The flow is
computed with the same Navier–Stokes solver and the same accuracy as the cylinder
wake in § 4. The computational domain is bounded by the rectangle −6 6 x 6 20
−y 6 6 6 y and is discretized by 4225 quadratic six-node elements. The Reynolds
numbers, based on the diameter of a single cylinder, is 100.
https://doi.org/10.1017/jfm.2016.678

We are working at companion experiments for this configuration. The motivation

is to have a simple geometry with three actuation inputs, the rotations of the
cylinders, which can display complex dynamics but remains easily computable and
experimentally realizable. The front cylinder controls the stagnation point while the
rear cylinders manipulate each shear-layer individually. Thus, many drag-reducing
mechanisms can be realized, like stagnation point stabilization (front cylinder),
 Recursive DMD 863

(a) (b)
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

(c) (d)

(e) ( f)

(g) (h)

(i) ( j)

F IGURE 16. (Colour online) The first 10 POD modes of the flow past three rotating
cylinders at Re = 100. (a,c,e,g,i) Modes 1–5, (b,d,f,h,j) modes 6–10. The vorticity of the
mode is visualized in colour (green, zero; red, above a positive threshold; blue, below a
negative threshold).

symmetric and antisymmetric high-frequency shear-layer excitation (rear cylinders) or

base bleeding (rear cylinders generating a jet on the x-axis), to name a few.
Following the exposition of § 4, we discuss POD (§ 5.1), DMD (§ 5.2) and RDMD
(§ 5.3), followed by a quantitative comparison of the residuals (§ 5.4).
https://doi.org/10.1017/jfm.2016.678

5.1. Proper orthogonal decomposition

The snapshots representing the attractor of the flow past three rotating cylinders
are decomposed using POD. The first 10 of them, representing more than 99 % of
the fluctuation energy, are depicted in figure 16. The first two modes describe von
Kármán vortex shedding and the next two modes resolve a low-frequency base-flow
 864 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid

(a) 0.0030 (b) 0.0080

0.0020 0.0060
0.0040
0.0010 0.0020
0 0
– 0.0020
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

– 0.0010
– 0.0040
– 0.0020 – 0.0060
– 0.0030 – 0.0080
145 150 155 160 165 170 175 178 145 150 155 160 165 170 175 178

(c) 0.0030 (d) 0.0200

0.0020 0.0150
0.0100
0.0010 0.0050
0 0
– 0.0010 – 0.0050
– 0.0100
– 0.0020 – 0.0150
– 0.0030 – 0.0200
145 150 155 160 165 170 175 178 145 150 155 160 165 170 175 178

(e) 0.0050 ( f ) 0.0200

0.0150
0.0030 0.0100
0.0010 0.0050
0
– 0.0010 – 0.0050
– 0.0030 – 0.0100
– 0.0150
– 0.0050 – 0.0200
145 150 155 160 165 170 175 178 145 150 155 160 165 170 175 178

(g) 0.0050 (h) 0.0400

0.0300
0.0030 0.0200
0.0010 0.0100
0
– 0.0010 – 0.0100
– 0.0030 – 0.0200
– 0.0300
– 0.0050 – 0.0400
145 150 155 160 165 170 175 178 145 150 155 160 165 170 175 178

(i) 0.0080 ( j) 0.0400

0.0300
0.0040 0.0200
0.0100
0 0
– 0.0100
– 0.0040 – 0.0200
– 0.0300
– 0.0080 – 0.0400
145 150 155 160 165 170 175 178 145 150 155 160 165 170 175 178

F IGURE 17. Flow past three rotating cylinders. The mode amplitudes of the POD modes
of figure 16.

modulation. Modes 5–10 form pairs describing fluctuations at higher frequencies. The
spatial mode interpretation is corroborated by the behaviour of the mode amplitudes
in figure 17.
https://doi.org/10.1017/jfm.2016.678

5.2. Dynamic mode decomposition

The same snapshots are subjected to DMD. The results, i.e. the eigenmodes and their
amplitudes, are depicted in figures 18 and 19 respectively.
The modal amplitudes show, as expected, pure harmonic behaviour, while the non-
symmetric modes arise from a non-symmetrically forced wake.
 Recursive DMD 865

(a) (b)
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

(c) (d)

(e) ( f)

(g) (h)

(i) ( j)

F IGURE 18. (Colour online) The same as figure 16 but for the first 10 DMD modes.

5.3. Recursive DMD

In this section, the results of RDMD are presented. Again, the procedure is
demonstrated for the same set of snapshots as employed previously for POD
and DMD. Figures 20 and 21 depict the modes and their amplitudes respectively.
The frequency contents of POD and RDMD are very similar, while the amplitude
modulation of POD is less pronounced in RDMD, corroborating the very purpose or
https://doi.org/10.1017/jfm.2016.678

RDMD, namely more harmonic modes.

5.4. Comparison of POD, DMD and RDMD

In this section, results from POD, DMD and RDMD are quantitatively compared.
Figure 22 shows the time-averaged residual as a function of the number of modes
 866 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid

(a) 0.10
0.08
(b) 0.10
0.08
0.06 0.06
0.04 0.04
0.02 0.02
0 0
– 0.02 – 0.02
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

– 0.04 – 0.04
– 0.06 – 0.06
– 0.08 – 0.08
– 0.10 – 0.10
145 150 155 160 165 170 175 178 145 150 155 160 165 170 175178

(c) 0.10 (d) 0.10

0.08 0.08
0.06 0.06
0.04 0.04
0.02 0.02
0 0
– 0.02 – 0.02
– 0.04 – 0.04
– 0.06 – 0.06
– 0.08 – 0.08
– 0.10 – 0.10
145 150 155 160 165 170 175 178 145 150 155 160 165 170 175178

(e) 0.10 ( f ) 0.10

0.08 0.08
0.06 0.06
0.04 0.04
0.02 0.02
0 0
– 0.02 – 0.02
– 0.04 – 0.04
– 0.06 – 0.06
– 0.08 – 0.08
– 0.10 – 0.10
145 150 155 160 165 170 175 178 145 150 155 160 165 170 175178

(g) 0.10 (h) 0.10

0.08 0.08
0.06 0.06
0.04 0.04
0.02 0.02
0 0
– 0.02 – 0.02
– 0.04 – 0.04
– 0.06 – 0.06
– 0.08 – 0.08
– 0.10 – 0.10
145 150 155 160 165 170 175 178 145 150 155 160 165 170 175178

(i) 0.10 ( j) 0.10

0.08 0.08
0.06 0.06
0.04 0.04
0.02 0.02
0 0
– 0.02 – 0.02
– 0.04 – 0.04
– 0.06 – 0.06
– 0.08 – 0.08
– 0.10 – 0.10
145 150 155 160 165 170 175 178 145 150 155 160 165 170 175178

F IGURE 19. The same as figure 17 but for the first 10 DMD modes.

for all three expansions. The performance of RDMD follows POD closely, while the
DMD residual is approximately one order of magnitude larger at N = 32.

6. Conclusions
https://doi.org/10.1017/jfm.2016.678

We propose a novel data-driven flow decomposition which combines the modal

orthonormality and low truncation error χ 2 (3.11) of POD with the frequency-distilling
features of DMD. This decomposition is recursively defined. First, the data set is
subjected to DMD, after which a normalized DMD mode is chosen which minimizes
the averaged error χ 2 . This procedure is recursively repeated in the orthogonal
subspace of computed modes. The resulting RDMD modes are orthonormal by
 Recursive DMD 867

(a) (b)
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

(c) (d)

(e) ( f)

(g) (h)

(i) ( j)

F IGURE 20. (Colour online) The same as figure 16 but for the first 10 RDMD modes.

construction and can be expected to have lower error χ 2 than DMD modes while
retaining the monochromatic features of DMD.
Proper orthogonal decomposition, DMD and RDMD have been applied to the same
snapshots of a transient cylinder wake starting near the steady solution and terminating
on the limit cycle. As expected, RDMD significantly outperforms DMD in terms of
the maximum, time-averaged and asymptotic truncation error for all considered mode
numbers by a large margin. In addition, the exponentially growing or decaying DMD
https://doi.org/10.1017/jfm.2016.678

amplitudes resemble neither initial nor asymptotic flow behaviour in a meaningful

manner, while the RDMD amplitudes clearly identify the initial, transient and post-
transient flow phases.
Also as expected, RDMD modes have far purer frequency content than POD modes
but maintain the residual at a comparable level. While the maximum and average
truncation errors of POD outperform RDMD, the latter shows a better resolution of
 868 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid

(a) (b) 2.00

4 1.50
2 1.00
0.50
0 0
– 0.50
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

–2 – 1.00
–4 – 1.50
– 2.00
145 150 155 160 165 170 175 180 145 150 155 160 165 170 175 180

(c) (d) 0.80

4 0.60
0.40
2 0.20
0 0
– 0.20
–2 – 0.40
– 0.60
–4 – 0.80
– 1.00
145 150 155 160 165 170 175 180 145 150 155 160 165 170 175 180

(e) 3.00 ( f ) 0.80

2.00 0.60
0.40
1.00 0.20
0 0
– 1.00 – 0.20
– 0.40
– 2.00 – 0.60
– 3.00 – 0.80
145 150 155 160 165 170 175 180 145 150 155 160 165 170 175 180

(g) 3.00 (h) 0.60

2.00 0.40
1.00 0.20
0 0
– 1.00 – 0.20
– 2.00 – 0.40
– 3.00 – 0.60
145 150 155 160 165 170 175 180 145 150 155 160 165 170 175 180

(i) 2.00 ( j) 0.60

1.50 0.40
1.00
0.50 0.20
0 0
– 0.50 – 0.20
– 1.00
– 1.50 – 0.40
– 2.00 – 0.60
145 150 155 160 165 170 175 180 145 150 155 160 165 170 175 180

F IGURE 21. The same as figure 17 but for the first 10 RDMD modes.

the limit cycle: the asymptotic value for N = 10 is half as large as the corresponding
POD value, and the third harmonic frequency is only captured by RDMD in the first
10 modes. Table 1 provides a brief comparison of the main characteristics and features
of POD, DMD and RDMD.
In the comparison, all modal decompositions were compared for the same data from
https://doi.org/10.1017/jfm.2016.678

which the modes were constructed. The robustness of the Galerkin expansions has
also been tested for a transient at a higher Reynolds number of 120. Here, RDMD
and POD performed very similarly, even better than DMD for the new data. A third
investigation has been performed for a non-periodic and non-symmetric wake behind
three rotating cylinders. Again, the main conclusions for RDMD, DMD and POD were
corroborated.
 Recursive DMD 869

1.000
DMD
POD
0.500 RDMD
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

0.100

0.050

0.010

0.005
0 5 10 15 20 25 30
N

F IGURE 22. The time-averaged fluctuation level of the residual (3.11) for POD, DMD
and RDMD for the flow past three rotating cylinders, with increasing number of modes.
The value is normalized by the corresponding converged fluctuation level (2K∞ ) on the
attractor.

Aspect POD DMD RDMD

Ideal snapshots Uncorrelated Time-resolved Time-resolved
Averaged truncation error Optimal Poor Good
Maximum truncation error 13 % 46 % 28 %
Asymptotic truncation error 10 % 10 % 5%
Modal frequency content Mixed Pure Almost pure
Noise sensitivity Low High Like
without filter DMD
Niche applications Statistics Stability modes Transient
Fourier modes dynamics
TABLE 1. Comparison of POD, DMD and RDMD for the transient cylinder wake.
The maximum and asymptotic truncation errors are for expansions with N = 10 modes
normalized with the post-transient fluctuation level 2K∞ .

The literature contains alternative approaches for spectrally purified POD modes.
One important recent contribution is spectral POD (Sieber, Paschereit & Oberleithner
2016), which interpolates between POD and DMD by filtering the correlation matrix.
This continuous interpolation offers an additional degree of freedom not present in
RDMD; the price is the loss of strict orthonormality of the modes for all interpolation
parameters. In a similar vein, Bourgeois, Martinuzzi & Noack (2013) construct
orthogonal POD modes with purer frequency content after Morlet-filtering the flow
https://doi.org/10.1017/jfm.2016.678

data at design frequencies. Recursive DMD can be considered as a simpler approach,

which can be expected to perform well in an unsupervised manner, but leaves little
room for tuning the modes for special purposes or applications. As a word of warning,
all data-driven Galerkin expansions are – by definition – based on modes that are
linear combinations of the snapshots, i.e. they cannot leave the space spanned by the
training data. The benefit of one reduced-order representation method over another
 870 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid
may actually be small. Noack (2016) sketches how future data-driven low-dimensional
flow representations may be generalized to cure inherent limitations of the Galerkin
expansion.
Summarizing, the current study indicates that RDMD is an attractive ‘compromise’
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

between POD and DMD. It sacrifices little of the optimal residual property of
POD while retaining the single-frequency behaviour of DMD. Its potential in
control-oriented reduced-order modelling will be explored in a future study. In
addition, a particularly attractive opportunity for RDMD is the unsupervised extraction
of generalized mean-field models with few dominant frequencies (Brunton & Noack
2015).

Acknowledgements
B.N. acknowledges the funding and excellent working conditions of the Senior
Chair of Excellence ‘Closed-loop control of turbulent shear flows using reduced-order
models’ (TUCOROM 2010–2015) supported by the French Agence Nationale de la
Recherche (ANR) and hosted by Institute PPRIME and the Collaborative Research
Centre (CRC 880) ‘Fundamentals of High Lift of Future Civil Aircraft’ supported by
the Deutsche Forschungsgemeinschaft (DFG) and hosted at the Technical University
of Braunschweig. This work has also been funded by the Polish National Centre
of Science under research grant no. DEC-2011/01/B/ST8/07264 ‘Novel Method
of Physical Modal Basis Generation for Reduced Order Flow Models’. We have
highly profited from stimulating discussions with M. Abel, S. Brunton, L. Cordier,
R. Martinuzzi, B. Protas, R. Radespiel and R. Semaan. We thank the Ambrosys
Ltd. Society for Complex Systems Management and the Bernd Noack Cybernetics
Foundation for additional support.

REFERENCES

BAGHERI , S. 2013 Koopman-mode decomposition of the cylinder wake. J. Fluid Mech. 726, 596–623.
BARKLEY, D. & H ENDERSON , R. D. 1996 Three-dimensional Floquet stability analysis of the wake
of a circular cylinder. J. Fluid Mech. 322, 215–241.
B ERGMANN , M. & C ORDIER , L. 2008 Optimal control of the cylinder wake in the laminar regime
by trust-region methods and POD reduced order models. J. Comput. Phys. 227, 7813–7840.
B OURGEOIS , J. A., M ARTINUZZI , R. J. & N OACK , B. R. 2013 Generalised phase average with
applications to sensor-based flow estimation of the wall-mounted square cylinder wake. J. Fluid
Mech. 736, 316–350.
B RUNTON , S. L. & N OACK , B. R. 2015 Closed-loop turbulence control: progress and challenges.
Appl. Mech. Rev. 67 (5), 050801.
C HEN , K. K., T U , J. H. & ROWLEY, C. W. 2012 Variants of dynamic mode decomposition: boundary
condition, Koopman, and Fourier analyses. J. Nonlinear Sci. 22 (6), 887–915.
C OATS , C. M. 1997 Coherent structures in combustion. Prog. Energy Combust. Sci. 22, 427–509.
C OURANT, R. & H ILBERT, D. 1989 Methods of Mathematical Physics, vol. 1. Wiley-VCH.
D EANE , A. E., K EVREKIDIS , I. G., K ARNIADAKIS , G. E. & O RSZAG , S. A. 1991 Low-dimensional
https://doi.org/10.1017/jfm.2016.678

models for complex geometry flows: application to grooved channels and circular cylinders.
Phys. Fluids A 3, 2337–2354.
F LETCHER , C. A. J. 1984 Computational Galerkin Methods, 1st edn. Springer.
F ÖPPL , L. 1913 Wirbelbewegung hinter einen Kreiszylinder (transl.: vortex motion behind a circular
cylinder). Sitzb. d. k. bayr. Akad. d. Wiss. 1, 1–18.
G ALERKIN , B. G. 1915 Rods and plates: series occurring in various questions regarding the elastic
equilibrium of rods and plates (translated). Vestn. Inzhen. 19, 897–908.
 Recursive DMD 871
G ERHARD , J., PASTOOR , M., K ING , R., N OACK , B. R., D ILLMANN , A., M ORZY ŃSKI , M. &
TADMOR , G. 2003 Model-based control of vortex shedding using low-dimensional Galerkin
models. In 33rd AIAA Fluids Conference and Exhibit, Orlando, Florida, USA, Paper 2003-4262.
H AN , Z.-H., S TEFAN , G. & Z IMMERMANN , R. 2013 Improving variable-fidelity surrogate modeling
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

via gradient-enhanced kriging and a generalized hybrid bridge function. Aerosp. Sci. Technol.
25 (1), 177–189.
H OARAU , C., B ORÉE , J., L AUMONIER , J. & G ERVAIS , Y. 2006 Analysis of the wall pressure trace
downstream of a separated region using extended proper orthogonal decomposition. Phys.
Fluids 18, 055107.
H OLMES , P., L UMLEY, J. L. & B ERKOOZ , G. 1998 Turbulence, Coherent Structures, Dynamical
Systems and Symmetry, 1st edn. Cambridge University Press.
J ØRGENSEN , B. H., S ØRENSEN , J. N. & B RØNS , M. 2003 Low-dimensional modeling of a driven
cavity flow with two free parameters. Theor. Comput. Fluid Dyn. 16, 299–317.
VON K ÁRMÁN , T. & RUBACH , H. 1912 Über den Mechanismus des Flüssigkeits- und Luftwiderstandes.
Phys. Zeitschr. XIII, 49–59.
L ORENZ , E. N. 1956 Empirical orthogonal functions and statistical weather prediction. Tech. Rep..
MIT, Department of Meteorology, Statistical Forecasting Project.
L ORENZ , E. N. 1963 Deterministic nonperiodic flow. J. Atmos. Sci. 20, 130–141.
L UGT, H. J. 1995 Vortex Flow in Nature and Technology. Krieger Publishing Company.
L UMLEY, J. L. 1967 The structure of inhomogeneous turbulent flows. In Atmospheric Turbulence
and Wave Propagation (ed. A. M. Yaglom & V. I. Tatarski), pp. 166–178.
M EZI Ć , I. 2013 Analysis of fluid flows via spectral properties of the Koopman operator. Annu. Rev.
Fluid Mech. 45, 367–378.
N OACK , B. R. 2016 From snapshots to modal expansions – bridging low residuals and pure
frequencies. J. Fluid Mech. 802, 1–4.
N OACK , B. R., A FANASIEV, K., M ORZY ŃSKI , M., TADMOR , G. & T HIELE , F. 2003 A hierarchy
of low-dimensional models for the transient and post-transient cylinder wake. J. Fluid Mech.
497, 335–363.
N OACK , B. R., P ELIVAN , I., TADMOR , G., M ORZY ŃSKI , M. & C OMTE , P. 2004 Robust low-
dimensional Galerkin models of natural and actuated flows. In Fourth Aeroacoustics Workshop,
pp. 0001–0012. RWTH Aachen.
O BERLEITHNER , K., S IEBER , M., NAYERI , C. N., PASCHEREIT, C. O., P ETZ , C., H EGE , H.-C.,
N OACK , B. R. & W YGNANSKI , I. 2011 Three-dimensional coherent structures in a swirling
jet undergoing vortex breakdown: stability analysis and empirical mode construction. J. Fluid
Mech. 679, 383–414.
O RSZAG , S. A. 1971 Numerical simulation of incompressible flows within simple boundaries: accuracy.
J. Fluid Mech. 49, 75–112.
ROM -K EDAR , V., L EONARD , A. & W IGGINS , S. 1990 An analytical study of transport, mixing and
chaos in unsteady vortical flow. J. Fluid Mech. 214, 347–394.
ROWLEY, C. W., M EZI Ć , I., BAGHERI , S., S CHLATTER , P. & H ENNINGSON , D. S. 2009 Spectral
analysis of nonlinear flows. J. Fluid Mech. 645, 115–127.
S CHLEGEL , M., N OACK , B. R., J ORDAN , P., D ILLMANN , A., G RÖSCHEL , E., S CHRÖDER , W., W EI ,
M., F REUND , J. B., L EHMANN , O. & TADMOR , G. 2012 On least-order flow representations
for aerodynamics and aeroacoustics. J. Fluid Mech. 697, 367–398.
S CHMID , P. J. 2010 Dynamic mode decomposition for numerical and experimental data. J. Fluid.
Mech 656, 5–28.
S CHUMM , M., B ERGER , E. & M ONKEWITZ , P. A. 1994 Self-excited oscillations in the wake of
https://doi.org/10.1017/jfm.2016.678

two-dimensional bluff bodies and their control. J. Fluid Mech. 271, 17–53.
S IEBER , M., PASCHEREIT, C. O. & O BERLEITHNER , K. 2016 Spectral proper orthogonal
decomposition. J. Fluid Mech. 792, 798–828.
S IEGEL , S. G., S EIDEL , J., FAGLEY, C., L UCHTENBURG , D. M., C OHEN , K. & M CLAUGHLIN , T.
2008 Low dimensional modelling of a transient cylinder wake using double proper orthogonal
decomposition. J. Fluid Mech. 610, 1–42.
 872 B. R. Noack, W. Stankiewicz, M. Morzyński and P. J. Schmid
S IROVICH , L. 1987 Turbulence and the dynamics of coherent structures. Part I: coherent structures.
Q. Appl. Maths XLV, 561–571.
S UH , Y. K. 1993 Periodic motion of a point vortex in a corner subject to a potential flow. J. Phys.
Soc. Japan 62, 3441–3445.
Downloaded from https:/www.cambridge.org/core. Columbia University Libraries, on 04 Jul 2017 at 10:01:40, subject to the Cambridge Core terms of use, available at https:/www.cambridge.org/core/terms.

TAYLOR , C. & H OOD , P 1973 A numerical solution of the Navier–Stokes equations using the finite
element technique. Comput. Fluids 1 (1), 73–100.
W ILLIAMSON , C. H. K. 1996 Vortex dynamics in the cylinder wake. Annu. Rev. Fluid Mech. 28,
477–539.
Z EBIB , A. 1987 Stability of viscous flow past a circular cylinder. J. Engng Maths 21, 155–165.
Z HANG , H.-Q., F EY, U., N OACK , B. R., K ÖNIG , M. & E CKELMANN , H. 1995 On the transition of
the cylinder wake. Phys. Fluids 7 (4), 779–795.
https://doi.org/10.1017/jfm.2016.678