1. Introduction
Near-wall turbulence is organised into streamwise rolls, and alternating low- and high-speed streamwise streaks (Klebanoff, Tidstrom & Sargent Reference Klebanoff, Tidstrom and Sargent1962; Kline et al. Reference Kline, Reynolds, Schraub and Runstadler1967; Blackwelder & Eckelmann Reference Blackwelder and Eckelmann1979; Smith & Metzler Reference Smith and Metzler1983; Johansson, Her & Haritonidis Reference Johansson, Her and Haritonidis1987). These coherent structures are well-documented and their characterisation is the subject of multiple works (Bakewell Jr. & Lumley Reference Bakewell and Lumley1967; Landahl Reference Landahl1980; Butler & Farrell Reference Butler and Farrell1993; Chernyshenko & Baig Reference Chernyshenko and Baig2005; Del Alamo & Jiménez Reference del ÁLAMO and JIMÉNEZ2006). Additionally, many studies point to a quasiperiodic cycle, wherein the streamwise streaks are amplified by streamwise vortices, meander then break down, which subsequently regenerates new quasistreamwise vortices (Kim, Kline & Reynolds Reference Kim, Kline and Reynolds1971; Robinson Reference Robinson1991; Hamilton, Kim & Waleffe Reference Hamilton, Kim and Waleffe1995; Waleffe Reference Waleffe1997; Panton Reference Panton2001; Adrian Reference Adrian2007; Smits, McKeon & Marusic Reference Smits, McKeon and Marusic2011; Jiménez Reference Jiménez2018). The cycle, also known as the self-sustaining process, can be more clearly observed in a minimal flow unit (Jiménez & Moin Reference Jiménez and Moin1991), where the domain is artificially restricted in the streamwise and spanwise directions in order to exclude the dynamics of the outer region of the channel. Recent methods based on graph-theoretic approaches (Elnahhas et al. Reference Elnahhas, Lenz, Moin, Lozano-Durán and Bae2024) have provided new evidence of the self-sustaining process and revealed consistent patterns of energy exchange between rolls and streaks in both the minimal flow unit and larger channels.
Although nonlinear mechanisms play a role in the self-sustaining process, much attention has been given to linear mechanisms and instabilities as the drivers of this process (Panton Reference Panton2001; Jiménez Reference Jiménez2013; Lozano-Durán et al. Reference Lozano-Durán, Constantinou, Nikolaidis and Karp2021). One example is the Orr mechanism (Orr Reference Orr1907; Jiménez Reference Jiménez2013), in which the mean shear near the wall tilts velocity perturbations forward in the streamwise direction and stretches vertical scales, intensifying the wall-normal velocity perturbations. Another example is lift-up (Hwang & Cossu Reference Hwang and Cossu2010), which occurs when wall-normal velocity perturbations transport slow-moving fluid near the wall away into the faster flow field farther away from the wall. Works such as Del Alamo & Jiménez (Reference del ÁLAMO and JIMÉNEZ2006) and Pujals et al. (Reference Pujals, García-Villalba, Cossu and Depardon2009) show that, even after removing the nonlinear term from the perturbation equations, linear transient growth via the mean shear generates the dominant (streaky) structures in wall-bounded turbulence. The linearised system additionally accounts for much of the energy spectra and reproduces the self-similar profile in the logarithmic region. Similarly, Lozano-Durán et al. (Reference Lozano-Durán, Constantinou, Nikolaidis and Karp2021) show through numerical experiments that the minimal flow unit can sustain turbulence without the nonlinear feedback between the velocity fluctuations and the mean velocity profile, except when the Orr mechanism or push-over (momentum transfer from the spanwise mean shear into the streamwise velocity perturbation) are suppressed. The authors thus argue for linear transient growth as a prominent mechanism for transferring energy from the mean flow to turbulent fluctuations. In addition to the traditional near-wall streaks, smaller scales in the minimal channel are also capable of significant transient growth via purely linear mechanisms (Markeviciute & Kerswell Reference Markeviciute and Kerswell2024). This is shown using a linearisation of the Navier–Stokes equations about a base flow composed of the mean turbulent channel flow profile with an added near-wall streak. The transient growth of the small scales is found to be especially driven by spanwise gradients, further underlining the ability of the linear push-over mechanism to amplify perturbations. In this work, using methods from Cho, Hwang & Choi (Reference Cho, Hwang and Choi2018), Symon, Illingworth & Marusic (Reference Symon, Illingworth and Marusic2021) and Ding, Chung & Illingworth (Reference Ding, Chung and Illingworth2025), we lend special attention to the role of the spanwise self-advection term in transferring energy across scales.
The linear amplification process linking streamwise vortices and streamwise streaks has also been fruitfully studied through the lens of resolvent analysis (McKeon & Sharma Reference McKeon and Sharma2010; Moarref et al. Reference Moarref, Sharma, Tropp and McKeon2013; McKeon Reference McKeon2017). In resolvent analysis, the Navier–Stokes equations are reframed as a linear dynamical system for the velocity fluctuations; the nonlinear terms, along with other exogenous inputs to the system, are represented by an external forcing term acting on this system. The goal is then to solve for the spatial structure of the (nonlinear) forcing that generates the response (velocity) with the largest linear energy amplification. Resolvent analysis is traditionally applied to systems exhibiting spatial and temporal homogeneity. The linearised Navier–Stokes equations are first Fourier-transformed in time and the homogeneous spatial directions prior to resolvent analysis, which thus reveals the linear amplification properties of individual wavenumbers and frequencies. In the context of wall-bounded turbulent flows, the equations are Fourier-transformed in the streamwise and spanwise directions and time, and the targeted length scales are the traditional streamwise and spanwise streak spacings in the buffer layer (McKeon & Sharma Reference McKeon and Sharma2010; Moarref et al. Reference Moarref, Sharma, Tropp and McKeon2013; McKeon Reference McKeon2017; Bae, Lozano-Durán & McKeon Reference Bae, Lozano-Durán and McKeon2021). Despite using a linearisation of the equations of motion, traditional resolvent analysis identifies streamwise rolls as the most perturbing structures and streamwise streaks as the most amplified structures. Resolvent analysis thus provides a partial explanation for why rolls and streaks are observed in the turbulent channel, though the linearly obtained modes fail to reproduce second-order turbulent statistics with quantitative accuracy, especially for higher Reynolds numbers. Indeed, resolvent analysis ignores the feedback of the output modes into the nonlinear forcing term. One way to reincorporate this effect into resolvent models is through the inclusion of a linear eddy viscosity model in the linearised equations of motion (Morra et al. Reference Morra, Semeraro, Henningson and Cossu2019, Reference Morra, Nogueira, Cavalieri and Henningson2021; Symon et al. Reference Symon, Madhusudanan, Illingworth and Marusic2023). Another useful correction includes weighting the resolvent operator with the statistical properties of the nonlinear fluctuation terms (Zare, Jovanović & Georgiou Reference Zare, Jovanović and Georgiou2017; Morra et al. Reference Morra, Semeraro, Henningson and Cossu2019, Reference Morra, Nogueira, Cavalieri and Henningson2021); the forcing terms, which are considered uncorrelated (‘white’) across space and time in the traditional formulation of resolvent analysis, are reweighted to reflect the observed coherence of the nonlinear terms in the turbulent system.
Beyond simply using resolvent analysis as a way to study the self-sustaining process, resolvent modes have been used as a tractable way of tackling the control of turbulent flows. The resolvent response modes are often used as models for the fully turbulent flow. These modes can be cheaply computed for a variety of wavenumbers and frequencies to identify the structures that undergo the largest kinetic energy amplification under the linearised dynamics and are thus expected to figure prominently in the turbulent flow. In Toedtli, Luhar & McKeon (Reference Toedtli, Luhar and McKeon2019), resolvent modes were used to model the response of turbulent channel flow at
$Re_\tau \approx 180$
to a varying-phase opposition control. Though the Reynolds stresses of the model did not accurately capture the profile from direct numerical simulations (DNS), the change in the profiles and drag were well represented by the resolvent model. This suggests that the cheap computation of resolvent modes can be used in lieu of DNS to identify forcing frequencies and length scales that can enact desirable changes in the turbulent flow. Liu et al. (Reference Liu, Sun, Yeh, Ukeiley, Cattafesta and Taira2021) have also applied this framework to reduce pressure fluctuations along a cavity in supersonic flow (
$Re = 10\,000$
, computed using the cavity depth and the free stream velocity). In Yeh & Taira (Reference Yeh and Taira2019), resolvent analysis of a flow over an aerofoil (
$Re = 23\,000$
, based on chord length) is also used to identify a forcing frequency and length scale that enhance momentum mixing and reduce flow separation. A similar framework is used in Lin, Tsai & Tsai (Reference Lin, Tsai and Tsai2023) for a plunging cylinder (
$Re = 500$
, based on cylinder diameter) to reduce lift fluctuations. The assumption underpinning the above approach is that resolvent forcing modes, though only optimal for the linearised equations, are nevertheless efficient at actuating the fully coupled system.
In Bae et al. (Reference Bae, Lozano-Durán and McKeon2021), linearly identified resolvent modes were indeed found to play a role in the transfer of energy to coherent near-wall turbulent perturbations, even within a fully nonlinear turbulent flow at
$Re_\tau = 186$
, lending credibility to the assumption underpinning the works of Yeh & Taira (Reference Yeh and Taira2019), Liu et al. (Reference Liu, Sun, Yeh, Ukeiley, Cattafesta and Taira2021) and Lin et al. (Reference Lin, Tsai and Tsai2023). Via a modified simulation in which the contribution of the leading resolvent forcing mode is subtracted from the nonlinear term at every time step, the streak-regeneration process is interrupted and buffer layer turbulence in the minimal flow unit is greatly suppressed.
Resolvent forcing and response modes have shown promise in emulating the behaviour of turbulent flows under forcing, and this is linked to the role the forcing modes themselves play in amplifying near-wall turbulence in simulated flows. In a more physical scenario, however, the flow would only be forced intermittently, either by spontaneous events arising in the unforced flow or by externally imposed actuation. It would be valuable to study how a turbulent flow instantaneously reacts to such forcing, but since the flow in Bae et al. (Reference Bae, Lozano-Durán and McKeon2021) is altered at each time step, its time-dependent reaction is difficult to measure. We thus emphasise the importance of computing optimal time-localised forcing modes and transient response mode. Traditional formulations of resolvent analysis – in which the linearised equations of motion are Fourier-transformed in time, or Laplace-transformed in time and the real part of the Laplace variable is taken to zero to allow transients to decay – are incapable of representing such scenarios. The resulting resolvent modes are Fourier modes in time that lack transient growth information. In this work, we use a version of resolvent analysis formulated in a wavelet basis (Ballouz et al. Reference Ballouz, Lopez-Doriga, Dawson and Bae2023, Reference Ballouz, Dawson and Bae2024a , Reference Ballouz, Lopez-Doriga, Dawson and Baeb ). The individual wavelets can capture information localised in a particular time interval. Additionally, in contrast to the Laplace transform which prescribes the same decay rate for the forcing and response modes, wavelet-based resolvent analysis, particularly its windowed version, allows one to compute optimal forcing and response modes that are separated in time or frequency, e.g. optimal pulse-like forcing modes and their corresponding transient response trajectories that can extend over the entire time domain. A discretised and wavelet-transform signal in time is a concatenation of increasingly subsampled signals, each discretised over time shift. By windowing for a particular element of the transform in either the response or the forcing, the user can easily select the degree to which the original signal is subsampled, which roughly determines the frequency range (scale), as well as the interval in time (shift) in which the user wishes to constrain the forcing or response. More on this topic can be found in § 2.2.
We note that maximally growing transient trajectories for the linearised Navier–Stokes equations have long been computed under the optimal transient growth framework, and used to study turbulent flows (Butler & Farrell Reference Butler and Farrell1993; Schmid, Henningson & Jankowski Reference Schmid, Henningson and Jankowski2002; Pujals et al. Reference Pujals, García-Villalba, Cossu and Depardon2009; Jiménez Reference Jiménez2013; Encinar & Jiménez Reference Encinar and Jiménez2020). This framework finds the optimal initial condition that leads to a maximally energetic state at a chosen final time. While optimal transient growth has also been successful at producing rolls and streaks as the optimal perturbation and response structures, respectively, important differences exist between this method and wavelet-based resolvent analysis. Optimal transient growth is sensitive to the choice of final time. More importantly, as is expounded in § 2.2, optimal transient growth and wavelet-based resolvent analysis use different measures of optimality: the former maximises the kinetic energy ratio between the initial condition and the solution at the chosen terminal time, while the latter maximises the integrated kinetic energy of the response over the entire time domain. Maximising the integrated kinetic energy may better capture structures that tend to persist in time, rather than spike and decay rapidly. Moreover, wavelet-based resolvent analysis can identify optimal forcing terms arising at various points of the time domain, and not just its origin. For example, when applied to turbulent oscillating Stokes flows (Ballouz et al. Reference Ballouz, Dawson and Bae2024a ), the computed forcing and response modes coincide with the times in the cycle when the streamwise root-mean-square velocity peaks. The forcing modes usually precede the response modes. Another version of this spatiotemporal resolvent analysis computes time-sparse modes and similarly yields time-localised optimal forcing modes that precede their transiently varying responses (López-Doriga et al. Reference López-Doriga, Ballouz, Bae and Dawson2023, Reference López-Doriga, Ballouz, Bae and Dawson2024). Both of these results suggest that time-localised resolvent analysis correctly extracts cause-and-effect relations between the computed modes.
The transient growth of any structure within a fully turbulent flow is modulated by the myriad nonlinear interactions not considered when using methods that rely on the linearised equations of motion. Therefore, with the objective of controlling near-wall turbulence in mind, we wish to determine whether significant transient growth can be achieved in a turbulent flow via the injection of a resolvent forcing mode into the simulation of a turbulent flow. The transient resolvent response mode corresponding to the injected forcing will allow us to measure the instantaneous discrepancy between the actuated turbulent and linearised flows, and probe the efficacy of the resolvent forcing modes at actuating nonlinear flows. We aim to identify the time scales during which the turbulent system responds similarly to the optimal linear response mode and beyond which nonlinear effects distort the effect of the resolvent forcing mode. We study the mechanisms that erode the effects of the injected mode, especially the nonlinear interactions that transfer energy from the actuated scale to secondary ones and force the turbulent flow to deviate from the optimal linear response. The system we use is the minimal flow unit at
$Re_\tau = 186$
.
This paper is organised as follows. In § 2.1, we compute the base flow for the minimal flow unit at
$Re_\tau = 186$
, which we use in § 2.2 to formulate the resolvent operator in a time-localised wavelet basis. We compute resolvent modes as in Ballouz et al. (Reference Ballouz, Lopez-Doriga, Dawson and Bae2024b
), making sure to constrain the forcing to a wavelet-shaped pulse. This yields a forcing mode in the shape of streamwise rolls that is compactly supported in time, in addition to an optimal streak-like response that grows transiently before decaying. The justification for the choice of spatial wavenumbers and wavelets is given in § 2.3. We then solve the fully nonlinear forced Navier–Stokes equations for the minimal flow unit at
$Re_\tau = 186$
, using the time-localised wavelet-based resolvent forcing mode as our forcing term. This step is detailed in § 2.4. We track the evolution of this resolvent forcing mode as it generates and amplifies streamwise streaks, and compute relevant turbulent statistics, which we present in § 3. In § 4, we focus on the nonlinear energy transfer from the induced streak to secondary modes. Concluding remarks are given in § 5.
2. Methods
In this work, we consider the flow in the minimal flow unit of size
$L_{1} \times L_{2} \times L_{3} = 1.72\delta \times 2\delta \times 0.86 \delta$
, where
$\delta$
is the channel half-height, and
$x_1$
,
$x_2$
and
$x_3$
are the streamwise, wall-normal and spanwise directions, respectively. We denote the velocity fluctuation field by
$\boldsymbol u = [u_1, u_2, u_3]^{\textrm T}$
, where
$u_1$
,
$u_2$
and
$u_3$
are the streamwise, wall-normal and spanwise components, respectively. The system is characterised by the friction Reynolds number
$Re_\tau = \delta u_\tau / \nu \approx 186$
, where
$u_\tau$
is the friction velocity, and
$\nu$
is the kinematic viscosity. The flow is periodic in the streamwise and spanwise directions, and the no-slip and no-penetration conditions hold at the walls of the channel.
2.1. Base flow
For the DNS in this work, we discretise the streamwise and spanwise directions uniformly using
$N_1 = N_3 = 32$
grid points, which results in streamwise and spanwise grid spacings of
$\Delta x_1^+ \approx 10$
and
$\Delta x_3^+ \approx 5$
. In the wall-normal direction, the grid is of size
$N_2 = 128$
and stretched according to a hyperbolic tangent distribution, which results in a wall-normal spacing of
$\min (\Delta x_2^+) \approx 0.17$
near the wall and
$\max (\Delta x_2^+) \approx 7.6$
at the centreline. Here, the superscript
$+$
denotes wall units normalised with
$u_\tau$
and
$\nu$
. We discretise the incompressible Navier–Stokes equations with a staggered, second-order accurate, central finite difference method in space (Orlandi Reference Orlandi2000), and a fractional step method is used to compute pressure (Kim & Moin Reference Kim and Moin1985). Time-advancement is performed with an explicit third-order-accurate Runge–Kutta method (Wray Reference Wray1990). The DNS code has been validated in previous studies of turbulent channel flows (Lozano-Durán & Bae Reference Lozano-Durán and Bae2016; Bae et al. Reference Bae, Lozano-Durán, Bose and Moin2018, Reference Bae, Lozano-Durán, Bose and Moin2019). Using this discretisation, we obtain a mean streamwise velocity profile
$U_1(x_2)$
by averaging DNS results of the unforced system in the homogeneous directions and time. This mean profile is used in the subsequent sections.
2.2. Spatiotemporal resolvent modes
In this section , we describe how we compute the time-localised resolvent forcing modes and their corresponding transient responses for the minimal flow unit (Ballouz et al. Reference Ballouz, Dawson and Bae2024a
,
Reference Ballouz, Lopez-Doriga, Dawson and Baeb
). We first formulate the incompressible Navier–Stokes equations for the velocity fluctuations about a mean turbulent flow field
$\boldsymbol{U} = (U_1, U_2, U_3)$
,
Here
$p$
represents the pressure fluctuation, and
$f_i$
the nonlinear term in the
$x_i$
-momentum equation. The base flow is obtained from § 2.1, and satisfies
$U_2 = U_3 = 0$
while the mean streamwise component
$U_1$
is constant in time. Equation (2.1) is Fourier-transformed in the
$x_1$
and
$x_3$
directions and discretised in
$x_2$
and time. The time domain
$[0, T]$
is periodic – i.e. when constructing the resolvent operator for the discretised system, periodic boundary conditions are used for the temporal derivative matrix; this is further discussed below – and we choose
$T = 22 \; \delta / u_\tau$
, which is long enough to allow the resolvent modes to decay to zero. The discretisation in time is uniform, with a grid size of
$N_t = 128$
corresponding to a spacing of
$\Delta\, t \approx 0.17 \; \delta / u_\tau$
. The wall-normal discretisation is the same as in the DNS (§ 2.1).
The discretised equations are further wavelet-transformed in time by premultiplying them by a discrete wavelet transform operator
$\unicode{x1D652}$
. For an arbitrary square-integrable function
$g(t)$
, the wavelet expansion on a dyadic time grid of size
$N_t$
– the grid is uniform and
$N_t$
is a power of two – is defined as
\begin{equation} g(t) = \sum _{\ell = 1}^L \sum _{m = 0}^{N_t/2^\ell -1} \frac {1}{\sqrt {2^\ell }}\tilde g^{(w)}(\ell , m) \eta \left (\frac {t}{2^\ell } - m \right ) + \sum _{m = 0}^{N_t/2^L-1} \frac {1}{\sqrt {2^L}} \tilde g^{(s)}(m) \zeta \left (\frac {t}{2^L} - m \right ), \end{equation}
where
$\eta (t)$
and
$\zeta (t)$
denote the wavelet and scaling functions, respectively (Mallat Reference Mallat1999; Najmi Reference Najmi2012), and
$L$
satisfies
$2^L \leqslant N_t$
and represents the largest scale captured by the wavelet expansion. The dilations and shifts of
$\eta (t)$
act as high-pass filters and coefficients
$\tilde g^{(w)}(\ell , m)$
capture the high-frequency information of
$g(t)$
; while
$\zeta (t)$
acts as a low-pass filter and coefficients
$\tilde g^{(s)}(m)$
capture the remaining low-frequency information of the function. The matrix
$\unicode{x1D652}$
approximates this wavelet expansion: given
$\boldsymbol{g}$
– the discretisation of
$g(t)$
in time – its wavelet transform
$\boldsymbol{\tilde g}$
is computed as
where the elements of
$\boldsymbol{\tilde g^{(s)}}$
and
$\boldsymbol{\tilde g^{(w)}}$
are, respectively,
$\tilde g^{(w)}(\ell , m)$
and
$\tilde g^{(s)}(m)$
for all
$\ell$
and
$m$
. The matrix
$\unicode{x1D652}$
is of size
$N_t \times N_t$
, and the transform
$\boldsymbol{\tilde g}$
is of size
$N_t$
. The projection onto
$\eta (t/2^\ell - m)$
or
$\zeta (t/2^L - m )$
roughly captures a portion of the frequency content of
$g$
determined by
$\ell$
, centred in a time interval determined by
$m$
. Larger
$\ell$
corresponds to a narrower band of frequencies closer to zero, while larger
$m$
corresponds to later times. Each of
$\eta$
and
$\zeta$
capture different portions of the frequency spectrum.
The choice of wavelet-scaling-function pair is not unique and determines the properties of the transform operator
$\unicode{x1D652}$
. In this work, we use a single-level Daubechies-8 wavelet transform (Daubechies Reference Daubechies1992). The Daubechies wavelets and their corresponding scaling functions are compactly supported in time and form an orthonormal basis, resulting in a sparse banded and unitary operator
$\unicode{x1D652}$
(Mallat Reference Mallat1999; Najmi Reference Najmi2012; Ballouz et al. Reference Ballouz, Lopez-Doriga, Dawson and Bae2024b
).
For a given streamwise and spanwise wavenumber pair
$(2\pi k_{1}/L_{1}, 2\pi k_{3}/L_{3})$
, where
$k_{1}, \; k_{3} \in {\mathbb{Z}}$
, we obtain the following discretised and transformed equations:
Here,
$\tilde {\unicode{x1D66A}}_i$
,
$\tilde{\unicode{x1D665}}_i$
and
$\tilde{\unicode{x1D65B}}_i$
denote the discretised and transformed velocity, pressure and forcing, respectively. These transformed quantities are functions of wall-normal position
$x_2$
, and the wavelet scale and shift parameters
$\ell$
and
$m$
, which, respectively, represent the time interval and frequency support of the wavelet mode (Ballouz et al. Reference Ballouz, Lopez-Doriga, Dawson and Bae2024b
). The transformed spatial derivative operators are defined as follows:
$\widetilde {{\unicode{x1D63F}}_1} = \text{i} \breve k_{1} \unicode{x1D644}$
, where
$\unicode{x1D644}$
is the identity matrix of size
$(N_2N_t) \times (N_2N_t)$
,
$\widetilde {{\unicode{x1D63F}}_2} = {\unicode{x1D63F}}_2$
, which denotes a block diagonal second-order-accurate central finite difference operator on the staggered
$x_2$
grid,
$\widetilde {{\unicode{x1D63F}}_3} = \text{i} \breve k_{3} \unicode{x1D644}$
and
$\widetilde {{\unicode{x1D63F}}^2_{jj}} = -\breve k_{1}^2 \unicode{x1D644} + {\unicode{x1D63F}}^2_{2} - \breve k_{3}^2 \unicode{x1D644}$
, where
${\unicode{x1D63F}}^2_{2}$
denotes a second-order-accurate second-order finite difference operator on the staggered
$x_2$
grid. The wavenumbers
$\breve k_i$
are the modified wavenumbers for the discretisation scheme used in the DNS, and are defined as
$\breve k_{1} := 2\delta \textrm{sin}( \delta \Delta x_1 \pi k_{1}/L_{1} )/\Delta x_1$
and
$\breve k_{3} := 2 \delta \textrm{sin}(\delta \Delta x_3 \pi k_{3}/L_{3})/\Delta x_3$
. Here, the use of the modified wavenumbers ensures that the modes satisfy the conservation laws – namely continuity – on the simulation grid and for the discretisation scheme used in the DNS; this is particularly important for the forcing modes, which will be injected into the DNS of the minimal flow unit.
The modified time derivative operator is defined as
$\widetilde {{\unicode{x1D63F}}_t} = \unicode{x1D652} {\unicode{x1D63F}}_t \unicode{x1D652}^{\,-1}\!$
, where
$\unicode{x1D652}$
now denotes the discrete wavelet transform in time for the
$(N_tN_2 \times N_tN_2 )$
-dimensional system (i.e. the transform defined previously for a one-dimensional discrete time signal, premultiplying the
$(N_2\times N_2)$
-dimensional identity matrix
$\unicode{x1D644}$
via a Kronecker product), and
${\unicode{x1D63F}}_t$
is a second-order-accurate central finite difference matrix in time. Though not shown, using a fourth-order-accurate finite difference operator does not strongly affect the resolvent modes. We also define the mean flow term
$\widetilde {\unicode{x1D650}}_j := \unicode{x1D652} \unicode{x1D650}_j \unicode{x1D652}^{\,-1}$
and the mean shear term
$\widetilde {{\unicode{x1D659}\unicode{x1D650}}}_{ij} := \unicode{x1D652} {\unicode{x1D659}\unicode{x1D650}}_{ij} \unicode{x1D652}^{\,-1}\!$
, where
$\unicode{x1D650}_j$
and
${\unicode{x1D659}\unicode{x1D650}}_{ij}$
are diagonal matrices with diagonal terms corresponding to
$U_j$
and
${\textrm d}U_i/{\textrm d}x_{\!j}$
at each
$x_2$
and time grid point, respectively.
We can rearrange (2.4) as
\begin{equation} \begin{bmatrix} \widetilde {\unicode{x1D66A}}_1(x_2, \ell , m) \\[3pt] \widetilde {\unicode{x1D66A}}_2(x_2, \ell , m) \\[3pt] \widetilde {\unicode{x1D66A}}_3(x_2, \ell , m) \\[3pt] \widetilde {\unicode{x1D665}}(x_2, \ell , m) \end{bmatrix} = \,\widetilde {\! \unicode{x1D643}}^{(k_{1}, k_{3})} \begin{bmatrix} \,\widetilde {\! \unicode{x1D65B}}_1(x_2, \ell , m) \\[3pt] \,\widetilde {\! \unicode{x1D65B}}_2(x_2, \ell , m) \\[3pt] \,\widetilde {\! \unicode{x1D65B}}_3(x_2, \ell , m) \\[3pt] 0 \end{bmatrix}, \end{equation}
where the linear operator
$\,\widetilde {\! \unicode{x1D643}}^{(k_{1}, k_{3})}$
is the resolvent operator in this formulation. The superscript
$(k_{1}, k_{3})$
indicates the choice of Fourier parameters for the transformation of (2.1), and the functional dependence on
$x_2$
,
$\ell$
and
$m$
denotes the discretisation of (2.1) over these quantities. Since we use the same
$x_2$
grid as the DNS that produced the mean flow and shear profiles, and taking
$N_t$
as the temporal resolution, we note that
$\,\widetilde {\! \unicode{x1D643}}^{(k_{1}, k_{3})}$
is a
$(4N_2 N_t) \times (4 N_2 N_t)$
matrix. This formulation of the resolvent operator targets all the Fourier frequencies resolved by the temporal grid, rather than one as in traditional resolvent analysis.
We introduce the additional step of constraining the forcing along a wavelet-shaped pulse of any desired scale
$\ell$
and shift
$m$
using a temporal windowing matrix
$\unicode{x1D63D}$
(Jeun, Nichols & Jovanović Reference Jeun, Nichols and Jovanović2016; Kojima et al. Reference Kojima, Yeh, Taira and Kameda2020; Ballouz et al. Reference Ballouz, Lopez-Doriga, Dawson and Bae2023, Reference Ballouz, Dawson and Bae2024a
). Typically, the unconstrained resolvent modes for channel flow are Fourier modes in time centred at a chosen critical layer (McKeon & Sharma Reference McKeon and Sharma2010; Bae et al. Reference Bae, Lozano-Durán and McKeon2021; Ballouz et al. Reference Ballouz, Lopez-Doriga, Dawson and Bae2024b
). In order to compute an optimal transiently growing response trajectory, we must constrain the forcing in time. The windowing matrix
$\unicode{x1D63D}$
takes the form
so that we are restricting the forcing to a wavelet or scaling function, centred at a desired time and frequency determined by
$\ell _d$
and
$m_d$
, respectively. For a discussion on the choice of
$\ell _d$
and
$m_d$
, see § 2.3. We then take the singular value decomposition of the combined operator
where
$(\cdot )^H$
denotes the conjugate transpose. We index the singular values
$\{\sigma _i\}_{i=1}^{\infty }$
such that
$\sigma _1 \geqslant \sigma _2 \geqslant \ldots \geqslant 0$
. The right and left singular vectors
$\{\boldsymbol{\tilde \phi _j}\}_{j=1}^{\infty }$
and
$\{\boldsymbol{\tilde \psi _j}\}_{j=1}^{\infty }$
, respectively, define orthonormal bases for the spaces containing the nonlinear term (forcing) and the velocity and pressure fluctuations (response). For the singular value decomposition, we choose the inner product to be the seminorm representing action – or time-integrated energy – which we define for an arbitrary vector
$\boldsymbol{\tilde b} = [\tilde b_1, \tilde b_2, \tilde b_3, \tilde b_p]^{\textrm T}$
to be
where
$\boldsymbol{b} = [b_1, b_2, b_3, b_p]^{\textrm T}$
is the inverse wavelet- and Fourier-transform of
$\boldsymbol{\tilde b}$
(Ballouz et al. Reference Ballouz, Lopez-Doriga, Dawson and Bae2024b
). The kinetic energy amplification factor is given by the square of the singular values. The forcing modes are therefore ordered decreasingly according to the integrated kinetic energy amplification they undergo when acted on by
$\,\widetilde {\! \unicode{x1D643}}^{(k_{1}, k_{3})} \unicode{x1D63D}$
, and the response modes are the corresponding amplified coherent structures arising from this action. Thus,
$\boldsymbol{\tilde \phi _1}= [\tilde \phi _{1, 1}, \; \tilde \phi _{1, 2}, \; \tilde \phi _{1, 3}, \;0]^{\textrm T}$
generates the largest linear energy amplification via the windowed resolvent operator, and
$\sigma _1 \boldsymbol{\tilde \psi _1} = \sigma _1 [ \tilde \psi _{1, 1}, \; \tilde \psi _{1, 2}, \; \tilde \psi _{1, 3}, \; \tilde \psi _{1, p}]^{\textrm T}$
is the optimally amplified transient velocity and pressure fluctuation resulting from the application of
$\boldsymbol{\tilde \phi _1}$
on the linearised system. Here,
$ \tilde \psi _{j, i}$
and
$ \tilde \phi _{j, i}$
correspond to the
$i$
th velocity component, and
$ \tilde \psi _{j, p}$
and
$ \tilde \phi _{j, p}$
refer to the pressure component. For all
$j$
,
$\boldsymbol{\phi _j}$
, the inverse transform of
$\boldsymbol{\tilde \phi _j}$
, is shaped in time according to the wavelet or scaling function chosen by
$\unicode{x1D63D}$
. As described in Moarref et al. (Reference Moarref, Sharma, Tropp and McKeon2013), the resolvent modes for certain Fourier parameters occur in equivalent pairs of equal singular values due to the symmetry of the channel geometry. The pairs of singular modes each form a singular plane, and for the numerical experiments in § 2.4, we linearly combine the two equivalent forcing modes (e.g.
$\boldsymbol{\tilde \phi _1}$
and
$\boldsymbol{\tilde \phi _2}$
) to form a forcing vector in the singular plane that acts primarily on the bottom-half of the channel. Thus, upon injecting this mode into the DNS of the forced Navier–Stokes equations, only the bottom of half of the channel is subject to the forcing, allowing us to use the unforced top-half as a benchmark for comparison (Bae et al. Reference Bae, Lozano-Durán and McKeon2021). Henceforth,
$\boldsymbol{\tilde \phi _1}$
refers to the linear combination of the first two equivalent forcing modes that concentrates the forcing in the bottom-half of the channel, renormalised to satisfy
$\|\boldsymbol{\tilde \phi _1} \| = 1$
, and
$\boldsymbol{\tilde \psi _1}$
denotes the corresponding response mode. The modes
$\boldsymbol{\tilde \phi _3}$
and
$\boldsymbol{\tilde \psi _3}$
are defined similarly with regards to the third and fourth resolvent forcing and response modes. We use
$\boldsymbol{\breve \psi _i}$
and
$\boldsymbol{\breve \phi _i}$
to denote the inverse wavelet transforms in time of
$\boldsymbol{\tilde \psi _i}$
and
$\boldsymbol{\tilde \phi _i}$
, or equivalently, the spatial Fourier transforms of
$\boldsymbol{ \psi _i}$
and
$\boldsymbol{ \phi _i}$
, respectively.
2.3. Choice of spatial and temporal scales
The minimal flow unit allows us to isolate one buffer layer streak. This streak appears to stretch the entire streamwise length of the unit, which is only large enough to contain one low- and high-speed streak pair in the spanwise direction. We thus choose to target the Fourier mode given by the streamwise and spanwise wavenumbers of
$k_{1} = 0$
and
$k_{3} = 1$
, respectively. These length scales also correspond to a peak in the spectral energy content for the minimal flow unit (Bae et al. Reference Bae, Lozano-Durán and McKeon2021).
Traditional resolvent analysis in which the Navier–Stokes equations are Fourier-transformed in time reveals that a temporal frequency of
$\omega = 0$
produces the modes with the largest kinetic energy amplification for the minimal flow unit at
$Re_\tau = 186$
(Bae et al. Reference Bae, Lozano-Durán and McKeon2021). To target this frequency, we constrain the forcing term to a Daubechies-8 scaling function of arbitrary shift
$m$
by using
$\unicode{x1D63D}$
to select the corresponding elements of
$\,\widetilde {\! \unicode{x1D65B}}$
. The scaling function is shown in figure 1(a), and its Fourier spectrum, shown in figure 1(b), indeed encompasses the target frequency. Since the scaling function is compactly supported in time, wavelet-based resolvent analysis will not be able to target
$\omega = 0$
uniquely but will capture a wide range of frequencies: a trade-off exists between time and frequency localisation, and the more precision we require in one domain, the less we preserve in the other (Mallat Reference Mallat1999; Najmi Reference Najmi2012). We note that the simulations detailed in § 2.4 resolve temporal wavenumbers up to
$\omega \delta / u_\tau \approx 55\,000$
. The scaling function satisfies
$ ( \int _{-\infty }^{+\infty } |\zeta (t)|^2 \textrm{d}t ) \; u_\tau / \delta = 1$
.
The obtained resolvent modes are shown in figure 2. Notably, the response modes exhibit transient energy growth and decay as seen in figure 2(a,b). The inverse transforms of the modes are shown in figure 2(c,d). The modes share many similarities with the Fourier-based modes computed for
$\omega = 0$
in Bae et al. (Reference Bae, Lozano-Durán and McKeon2021): the optimal transient nonlinear forcing mode appears in the shape of streamwise rolls, and the optimal velocity fluctuation response appears as predominantly streamwise streaks with alternating signs of the same magnitude. This supports the extensively examined claim that streamwise streaks can be linearly generated by a linear lift-up mechanism, whereby slower moving fluid close to the wall is swept upwards into the faster moving mean flow farther away from the wall. The streak-shaped response mode grows in intensity before fading, showcasing the transient growth that is characteristic of non-normal systems.
Daubechies-8 scaling function
$\zeta$
in time (a) and frequency (b) domain.

(a) Magnitudes of the (i) wall-normal component of the principal forcing mode and the (ii) streamwise component of the principal response mode. (b) Integrated energy of the principal forcing (red) and response (black) modes. (c) Principal forcing mode shown at peak amplitude (
$t u_\tau /\delta \approx 0.40)$
. (d) Principal response modes at peak amplitude (
$t u_\tau /\delta \approx 2.41)$
. In (c) and (d), the contours represent the streamwise magnitude; the arrows, which show the direction of the cross-flow components, are coloured according to their magnitudes
$\sqrt {| \phi _{1,2}|^2 + | \phi _{1,3}|^2}$
in (c), or
$\sqrt {| \psi _{1,2}|^2 + |\psi _{1,3}|^2}$
in (d).

2.4. Forced DNS
The time-localised resolvent forcing mode obtained in the previous section maximises the action for the minimal flow unit under the linearised Navier–Stokes equations. The aim of the work is to study the effect of the forcing mode on the fully nonlinear minimal flow unit. To that end, we aim to inject the time-localised resolvent forcing mode into a DNS of the minimal flow unit. The forcing mode will be introduced into an already turbulent state of the minimal flow unit. The turbulent initial condition will contain energy at varying scales, triggering cross-scale interactions that will interfere with the linearly driven transient growth of the injected mode. We inject the forcing at different magnitudes, all small relative to the energy of the initial condition. We expect the produced energy amplification to vary nonlinearly across the different cases. To characterise how the turbulent system reacts to actuation of varying strength, we aim to track the ensemble-averaged transient response, and measure the energy growth of the targeted mode along with the nonlinear energy transfer to secondary scales. The optimal resolvent response mode will serve as a benchmark for the achievable energy amplification caused by the injected mode. Below we present the numerical details of the DNS.
To initialise the ensemble of the forced simulations, we first perform a DNS of the non-actuated minimal flow unit, fixing the mean profile to be
$\boldsymbol{U} = (U_1(x_2), 0, 0)$
(§ 2.1), which is used to calculate the resolvent modes. Snapshots from this simulation will serve as the initial conditions to the forced simulations. Freezing the mean profile ensures that the DNS mean profile matches the one used to compute the resolvent modes for all time. We do this by initialising the flow to have the desired mean streamwise profile of
$U_1$
, then by removing the steady-state contribution of the right-hand side of the Navier–Stokes equations. The initial snapshots from the fixed-mean simulation are separated by
$1 \leqslant \Delta\, t u_\tau / \delta \leqslant 5.2$
, which reduces the correlation between the snapshots. The ensemble sizes range from 1000 to 4000, in order to ensure statistical convergence. Doubling the time between initial snapshots – equivalently, averaging over half the available time series data – led to streamwise energy fields that only differed by at most
$0.5$
–
$1.3\, \, \%$
at any given wall-normal location and time. For each initial condition, we also obtain a corresponding unforced fixed-mean time series. We denote the velocity fluctuations for the unforced simulations by
$\boldsymbol{u}_0 (x_1, x_2, x_3, t)$
.
Before injecting the forcing mode into the DNS of the minimal flow unit, the mode is normalised so that
$\| \boldsymbol{\tilde {\phi _1}}\|^2 = 1$
and scaled by a complex constant
$\kappa$
with magnitude
\begin{equation} |\kappa | := \gamma \left (\frac {\delta }{u_\tau ^3} \frac {1}{L_{1} (2\delta )L_{3}}\int _{-\infty }^{+\infty }\int _0^{L_{1}}\int _0^{2\delta }\int _0^{L_{3}} \left \vert \frac {\partial \boldsymbol{{u}_0}}{\partial t}\right \vert ^2_{t=0} \zeta (t)^2\textrm{d}x_3 \, \textrm{d}x_2 \, \textrm{d}x_1\, \textrm{d}t \right )^{1/2}, \end{equation}
where
$\gamma \in \{ 1\, \, \%, 2\, \, \%, 5\, \, \%, 10\, \, \%\}$
such that the resolvent forcing mode is increasing the initial energy of the right-hand side by
$\gamma \, \%$
. Thus,
$|\kappa |^2$
determines the integrated energy injected into the system by the forcing. Using
$\angle \cdot$
to denote the phase, we choose
so that the forcing mode is in phase with the right-hand side of the unforced flow field. In the limiting case where the nonlinear interactions are negligible, this provision ensures that the added forcing maximally increases the energy of the target
$(0, 1)$
-mode.
The forced DNS is the solution to the full incompressible Navier–Stokes equations with the additional right-hand side forcing terms as follows:
\begin{eqnarray} && \frac {\partial {u_{i}}}{\partial t} + u_{\!j} \frac {\partial u_{i}}{\partial x_{\!j}} + U_{\!j}\frac {\partial {u_{i}}}{\partial {x_{\!j}}} + u_{\!j}\frac {\partial {U_{i}}}{\partial {x_{\!j}}} = -\frac {\partial {p}}{\partial {x_{i}}} + \frac {1}{Re} \frac {\partial ^{2}u_{i}}{\partial {x_{\!j}}\partial {x_{\!j}}} \nonumber\\[8pt]&&\quad +\, 2 {\textrm {Re}} \left (\kappa \breve \phi _{1, i}(x_{2}, t) e^{\textrm{i} \frac {2\pi x_{3}}{L_{3}}} \right ) + {\mathcal{F}}_i, \end{eqnarray}
The term
$\boldsymbol{\mathcal{F}} = ({\mathcal{F}}_1, 0, 0)$
enforces the condition that the mean streamwise profile of the forced simulation stay fixed and equal to
$U_1$
(§ 2.1) as in the unforced simulation, i.e.
$\boldsymbol{\mathcal{F}}$
removes the steady-state (
$k_{1} = 0$
,
$k_{3} = 0$
) contribution of the right-hand side of the Navier–Stokes equations. We remind the reader that
$\boldsymbol{u} (x_1, x_2, x_3, 0 ) = \boldsymbol{u}_0 (x_1, x_2, x_3, 0 )$
for each ensemble member, where
$ \boldsymbol{u}_0$
is obtained from the unforced DNS. We run the forced DNS for a total time of
$T = 5.69 \; \delta / u_\tau$
. To test the optimality of
$\boldsymbol{\phi _1}$
at forcing the turbulent channel, we repeat the case for
$\gamma = 5\, \%$
using
$\boldsymbol{\phi _3}$
and a forcing term with a randomly generated spatial component
$\boldsymbol{\phi }_{\boldsymbol{rand}} = \hat {\boldsymbol{\phi }}_{\boldsymbol{rand}}(x_2) \zeta (t)$
, which we normalise and scale the same way (
$\Vert {\boldsymbol{\phi }}_{\boldsymbol{rand}} \Vert ^2 = 1$
).
3. Results and discussion
In this section, the notation
$\widehat {(\boldsymbol{\cdot})}^{(k_{1}, k_{3})}$
denotes the Fourier transform in the streamwise and spanwise directions corresponding to the streamwise and spanwise wavenumbers
$2\pi k_{1}/L_{1}$
and
$2\pi k_{3}/L_{3}$
, respectively. We define the deviation operator
$\varDelta$
as the difference between the forced and unforced simulations, e.g.
$\Delta \hat {u}_1^{(0,1)} = \hat u_1^{(0,1)} - \hat u_{0,1}^{(0,1)}$
. We denote the ensemble average by
$\overline { (\boldsymbol{\cdot})}$
.
3.1. Transient energy growth and decay of streaks in the forced DNS
We define the instantaneous streak energy as
\begin{align} \hat E_1^{(0,1)}(t) = \left [ \frac {|\hat u_1^{(0,1)}|^2}{2} \right ], \end{align}
where
$[\cdot ] := \int _0^{2\delta } (\boldsymbol{\cdot} ) {\textrm d}x_2 / (2\delta )$
denotes the wall-normal average. Figure 3(a) shows the streak energy contained in the
$(0, 1)$
-mode as a function of time, for different resolvent-forcing amplitudes. For all cases, the energies grow and peak before decaying and reverting back to non-forced levels. We observe that the peak energy deviation
$\Delta \hat E_1^{(0,1)}$
scales subquadratically with the forcing amplitudes and is proportional to
$ | \gamma | ^{1.44}$
(figure 3
b). For a linear system, the energy peaks would scale quadratically, which indicates that the nonlinearities cause the energy to start decaying before it can reach the peak allowed by optimal linear growth. The stronger the forcing, the faster the streak energy’s growth rate and the faster its decay. The peak times,
$t_{{peak}}$
, defined as the times at which the energies reach their maxima, decrease slightly with forcing amplitude, but are relatively constant compared with the decay times,
$\Delta\, t_{\textit{decay}}$
, which we define as the time it takes for the energy to reduce from the peak to
$10\, \%$
of its peak (figure 3
c,d). Indeed, the differences in decay rate are more dramatic across the different forcing amplitude cases, and scale as
$\Delta\, t_{\textit{decay}} \sim |\gamma |^{-0.65}$
. We note that all fully coupled simulations decay significantly faster (
$\Delta\, t_{\textit{decay}} \approx 1 \delta / u_\tau$
) than the linear response (
$\Delta\, t_{\textit{decay}}\approx 15 \delta / u_\tau$
).
(a) Average streak energy as a function of time; the cases plotted are
$\gamma = 1\, \%$
(black
$\circ$
),
$\gamma = 2\, \%$
(cyan
$\bullet$
),
$\gamma = 5\, \%$
(purple
$\times$
) and
$\gamma = 10\, \%$
(red
$*$
). (b) Streak energy peaks, (c) peak times and (d) decay times as a function of
$\gamma$
. The dashed line represents the trends (b)
$80.48 |\gamma |^{1.44}$
, and (d)
$|\gamma |^{-0.65}$
.

To measure the proportion of the linearly amplified energy captured by the forced simulations, we compute an ensemble-averaged forcing efficiency, or effective amplification, which we define to be
\begin{equation} \sigma _{\textit{eff}} = \left (\, \frac {1}{u_\tau \delta } \int _{0}^{T} \overline {\frac { \Delta \hat E^{(0,1)}(t)}{|\kappa |^2/2} } \textrm{d}t \,\right )^{{1}/{2}}. \end{equation}
This is analogous to
$\sigma _1 = \max _{\tilde {\unicode{x1D65B}}} \|\tilde {\unicode{x1D643}}^{(0, 1)} \unicode{x1D63D} \tilde {\unicode{x1D65B}}\, \|_2 / \|\tilde {\unicode{x1D65B}}\,\|_2$
, where the numerator reflects the energy contained in the velocity perturbation field, and the denominator corresponds to the forcing amplitude. The computed
$\sigma _{\textit{eff}}$
is shown in figure 4. The forcing efficiencies decrease with the intensity of the forcing, and all effective amplifications are lower than
$\sigma _1 = 11.54$
, indicating that, across forcing amplitudes, the resolvent response mode overpredicts the response of the system to the injected forcing, even for the low-amplitude cases. We expect the agreement to be improved when (2.1) are augmented by an eddy viscosity model (Zare et al. Reference Zare, Jovanović and Georgiou2017; Symon et al. Reference Symon, Madhusudanan, Illingworth and Marusic2023). Regarding the trend of
$\sigma _1$
with forcing amplitude, we see that for smaller resolvent forcing amplitudes, more of the forcing energy is linearly converted into streak energy. For higher values of
$\gamma$
, nonlinear interactions that scale superlinearly curtail the growth of the response mode – the integrated action of which scales linearly with
$\gamma$
in the linearised setting – and hinder the effectiveness of the resolvent forcing mode.
Effective amplification
$\sigma _{\textit{eff}}$
(solid) and
$\sigma _1$
(dashed).

3.2. Comparison of velocity deviations with linear response
To visualise the alignment of the velocity fluctuation fields with the linear response mode across all wall normal heights, we plot the contours of
$|\overline {\Delta \hat {u}^{(0,1)}/\kappa }|/u_\tau$
, i.e. the magnitude of the
$\hat u_1^{(0,1)}$
deviations normalised by the forcing coefficient
$\kappa$
, along with the contours of the linear resolvent response modes (figure 5). We note that
$\Delta \hat {u}^{(0,1)}$
is divided by the complex value of
$\kappa$
prior to ensemble averaging so as to align the phases of the forcing (§ 2.4) across the ensemble members. At earlier times (
$t \lt 0.7 \delta / u_\tau$
), the responses for both the
$\gamma = 1\, \%$
and
$\gamma = 10\, \%$
cases are very similar to the linear mode. The strongly forced case, however, quickly reverts to the unforced channel flow statistics beyond an eddy turnover time unit. In contrast, the lightly forced case of
$\gamma = 1\, \%$
exhibits a longer-lasting velocity deviation, especially in the near wall region (
$x_2^+ \lt 15$
). To more closely investigate how the agreement of the forced simulations and the optimal linear response varies with
$x_2^+$
, we show
$\Delta \hat {u}_1^{(0,1)}$
at two wall-normal heights, along with the linear response at those heights (figure 6). At both wall-normal locations, the initial growth rates are similar to the linear case for all
$\gamma$
, and we obtain good collapse prior to
$t \approx 0.7 \delta /u_\tau$
. However, increasing the forcing amplitude causes the
$\kappa$
-normalised velocity deviation to decay earlier and at a smaller amplitude for both wall-normal heights shown in figure 6. For a fixed forcing amplitude, we note that
$\Delta \hat {u}^{(0,1)}$
diverges from the optimal linear response around the same time at both wall-normal locations plotted in figure 6, but the agreement between the forced simulations and the optimal linear response turns out better in the near-wall region. As
$x_2^+$
moves closer to the wall, the growth rate due to linear mechanisms increases, and
$\Delta \hat {u}_1^{(0,1)}$
manages to recover more of the linearly amplified energy before decaying due to nonlinear effects.
Average deviation in the streamwise velocity of the
$(0,1)$
-Fourier mode,
$|\overline {\Delta \hat {u}^{(0,1)}/\kappa }|/u_\tau$
, for the (a)
$\gamma = 1\, \%$
and (b)
$\gamma = 10\, \%$
cases. The contours correspond to
$7\, \%$
,
$15\, \%$
,
$25\, \%$
,
$75\, \%$
and
$90\, \%$
of the maximum value of
$\sigma _1 \boldsymbol{\psi _1}$
. The lines represent the forced DNS case (blue), and the resolvent response (red).

Average deviation in the streamwise velocity of the
$(0,1)$
-Fourier mode at
$x_2^+ \approx 16$
(a) and
$x_2^+ \approx 39$
(b). The cases plotted are
$\gamma = 1\, \%$
(black
$\circ$
),
$\gamma = 2\, \%$
(cyan
$\bullet$
),
$\gamma = 5\, \%$
(purple
$\times$
),
$\gamma = 10\, \%$
(red
$*$
), and the linear response mode
$\sigma _1 \boldsymbol{\psi _1}$
(black, dashed).

Average streak energy for cases forced by
$\boldsymbol{\phi _1}$
(red
$*$
),
$\boldsymbol{\phi _3}$
(cyan
$\bullet$
) and
$\boldsymbol{\phi }_{\boldsymbol{rand}}$
(dashed black). The unforced case is shown in black.

Integrated streamwise spectral energy content
$\overline {[|\hat {u_1}^{(k_{1},k_{3})}|^2]}/(2 u_\tau ^2)$
for
$\gamma = 2\, \%$
(a–c) and
$\gamma = 10\, \%$
(d–f). The spectra shown are at times
$t = 0$
(a,d),
$t = 0.6 \delta /u_\tau$
(b,e) and
$t = 1.2 \delta /u_\tau$
(c,f). The energy contained in the
$(0,0)$
,
$(0,1)$
and
$(0,-1)$
-modes (black) are excluded for clarity.

3.3. Optimality of resolvent forcing
We compare the results of forcing using the principal resolvent forcing mode
$\boldsymbol{\phi _1}$
, the first suboptimal mode
$\boldsymbol{\phi _3}$
and the random mode
$\boldsymbol{\phi }_{\boldsymbol{rand}}$
, using
$\gamma = 5\, \%$
. We see that the streak energy grows to higher peak when the minimal flow unit is forced by
$\boldsymbol{\phi _1}$
than when forced by
$\boldsymbol{\phi _3}$
(figure 7). In both cases, the minimal flow unit is much more responsive compared with the case with random forcing. This suggests that resolvent analysis identifies a forcing structure to which the minimal flow unit is indeed sensitive, even when governed by the fully nonlinear Navier–Stokes equations. The advantage of the optimal forcing mode, however, is significantly reduced. The effective amplifications,
$\sigma _{\textit{eff}}$
in DNS forced by
$\boldsymbol{\phi _1}$
and
$\boldsymbol{\phi _3}$
differ only by a factor of
$1.03$
, whereas
$\sigma _1/\sigma _3 = 2.16$
. Nonlinear effects quickly clip the transient growth of the induced streak and drain energy from the actuated mode before the differences between the responses to
$\boldsymbol{\phi _1}$
and
$\boldsymbol{\phi _3}$
can deepen.
We remind the reader that the mean profile is fixed by an additional forcing
$\boldsymbol{\mathcal{F}}$
which removes the
$(0, 0)$
contribution to
$\partial {\boldsymbol u}/\partial t$
at every time step. Freezing the mean replenishes the energy in the
$(k_{1}, k_{3}) = (0, 0)$
mode and preserves the energy transfer mechanism to the resolvent response mode. In other words, allowing the mean profile to vary may reduce the effectiveness of the forcing mode, since the mode is only optimal for the mean profile without forcing. This is observed in Bae et al. (Reference Bae, Lozano-Durán and McKeon2021): subtracting the contribution of the resolvent forcing mode from the nonlinear term succeeds at reducing the turbulent kinetic energy initially, but this effect fades after
$t u_\tau / \delta \approx 12$
as the mean flow is modified. However, our injected forcing has a short extent in time (
$t u_\tau / \delta \approx 1.5$
), and we do not expect the initial linearly driven growth of the induced response mode to change much whether we fix the mean profile or not. Though not shown, the streak energy profile obtained for
$\gamma = 5\, \%$
and an unfixed mean velocity is indeed very close to the what is shown in figure 3(a).
3.4. Spectra
We compute the
$x_2$
-integrated streamwise and spanwise spectra for
$\gamma = 2\, \%$
and
$\gamma = 10\, \%$
, at the initial time, right before the peak in streak energy (
$t = 0.6 \delta / u_\tau$
) and during the streak energy decay (
$t = 1.2 \, \delta / u_\tau$
). The results are shown figure 8. Though not plotted, we note that the
$(0, \pm 1)$
-mode is the most dominant one across all times, accounting for
$30 {-}70\, \%$
of the total turbulent energy over the entire simulated time horizon. For all forcing amplitudes – though only the
$\gamma = 2\, \%$
and
$\gamma = 10\, \%$
cases are shown, the behaviour of the non-actuated modes is similar for all
$\gamma$
: the energy of the non-actuated modes grows during the decay of the
$(0, 1)$
-mode, with the
$(0, \pm 2)$
and
$(\pm 1, \pm 1)$
-modes growing most significantly, which highlights their key role in exchanging energy with the actuated
$(0, 1)$
-mode. The share of total turbulent energy accounted for by the
$(0, \pm 2)$
and
$(\pm 1, \pm 1)$
-modes is roughly constant across all forcing amplitudes, amounting to approximately
$9 {-}10\, \%$
and
$7 {-}8\, \%$
at their peaks, respectively. Some differences do exist across forcing amplitudes: the transient behaviour of the
$(0, \pm 2)$
-mode appears more sensitive to the value
$\gamma$
. For the lightly forced case, the energy of the
$(0, \pm 2)$
-mode continues to grow beyond
$t_{{peak}}$
, while for the strongly forced case, the energy of the mode peaks soon after the peak in streak energy before decaying rapidly and ceding to the
$(\pm 1, \pm 1)$
-mode. This suggests that the
$(0, \pm 2)$
-mode tends to grow faster and peak earlier as the forcing amplitude increases. The particular sensitivity of the integrated energy of the
$(0, \pm 2)$
-mode to forcing amplitude can be partially explained by the fact that it is fed by the dyadic interaction involving the self-interaction of the actuated
$(0, \pm 1)$
-mode. The nonlinear energy transfer from the
$(0, \pm 1)$
-mode to secondary scales is discussed in more detail in the next section.
The instantaneous growth and decay of the two preferred secondary modes, the
$(0, \pm 2)$
- and
$(\pm 1, \pm 1)$
-modes, at different wall-normal locations can be seen in figure 9(a,b). Closer to the wall, the energy of the
$(0, \pm 2)$
-mode first decreases during the growth of the
$(0, \pm 1)$
-mode, then exhibits two peaks for the larger amplitude cases, one coinciding with the peak of the
$(0, \pm 1)$
-mode, and one much later at around
$t u_\tau / \delta \approx 3$
, as the energy of the mode reverts back to its initial state. The first peak, which is only visible for the higher amplitude cases, occurs earlier as
$\gamma$
increases, while the later peak occurs roughly at the same time across forcing amplitudes. The behaviour of the
$(\pm 1, \pm 1)$
-mode is simpler and its energy peaks earlier as the forcing amplitude increases. Farther away from the wall, the energy peaks of both the
$(0, \pm 2)$
- and
$(\pm 1, \pm 1)$
-modes depend little on
$\gamma$
and occur roughly at the same time for all forcing amplitudes, at
$t u_\tau / \delta \approx 0.8$
and
$t u_\tau / \delta \approx 1$
, respectively. Thus, the energy of the
$(0, \pm 2)$
- and
$(\pm 1, \pm 1)$
-modes is similarly influenced by
$x_2$
and
$\gamma$
: while the forcing magnitude
$\gamma$
affects the growth time scale of the secondary modes, causing their energy to peak earlier as it increases, the wall-normal location modulates the sensitivity of these modes to
$\gamma$
.
Streamwise spectral energy content at (a)
$x_2^+ \approx 16$
and (b)
$x_2^+ \approx 39$
for the
$(0,\pm 2)$
–Fourier mode (dotted lines) and the
$(\pm 1,\pm 1)$
–Fourier mode (dashed lines). (c) Energy peak times for the
$(0, \pm 1)$
-mode (solid) and the smaller scales (dashed) cases. (d) Time scale for the cross-scale energy transport. Colours indicate
$\gamma = 2\, \%$
(blue),
$\gamma \approx 5\, \%$
(purple) and
$\gamma = 10\, \%$
(red). The vertical lines in (a) and (b) represent the local peaks of
$|\hat u_1^{(0,1)}|$
. The horizontal black lines in (c) and (d) represent
$x_2^+ = 15$
and
$x_2^+ = 40$
, which delineate the buffer layer.

To better visualise the cross-scale energy transport time scales, we plot the streamwise energy peak times for the
$(0, \pm 1)$
-mode and the smaller scales as a function of
$x_2^+$
and for different forcing amplitudes (figure 9
c). Across all wall-normal heights, increasing the forcing amplitude causes the energy of both the
$(0, \pm 1)$
-mode and the smaller scales to peak earlier. For all forcing amplitudes, the peak times for the
$(0, \pm 1)$
-mode and the smaller scales vary inversely with
$x_2^+$
. For the
$(0, \pm 1)$
-mode, they are roughly constant with
$x_2^+$
in the near-wall region, increase slightly in the buffer layer as we move farther away from the wall, and plateau in the outer region of the flow. For the non-actuated modes, the peak times are also constant in the near-wall region, but decrease dramatically within the buffer layer as
$x_2^+$
increases, before levelling off in the outer region. The growth of the smaller scales is thus more sensitive to both forcing amplitude and distance to the wall than the actuated
$(0, \pm 1)$
-mode.
We define the time scale for cross-scale energy transport,
$\Delta\, t_{\textit{trans}}$
, as the time delay between the energy peaks of the
$(0, \pm 1)$
-modes and the smaller scales; its dependence on
$x_2^+$
is mostly determined by the energy peak time for the smaller scales (figure 9
d). While the results are generally consistent with the established understanding of the energy cascade from larger to smaller scales, the trend of
$\Delta\, t_{\textit{trans}}$
probes additional aspects of the energy transfer, mainly how the rate of the energy transfer to the secondary scales varies with the forcing amplitude and wall-normal height. Interestingly, though a larger forcing amplitude accelerates the energy transport from the
$(0, \pm 1)$
-mode to smaller scales for all wall normal heights, the sensitivity of
$\Delta\, t_{\textit{trans}}$
to forcing amplitude decreases as we move farther away from the wall. Indeed, in the outer regions of the flow,
$\Delta\, t_{\textit{trans}}$
converges to a value of approximately
$0.18 \, \delta / u_\tau$
for high forcing amplitude. The plots in figure 9(c,d) indicate two cross-scale energy transfer mechanisms: one for the near-wall region (
$x_2^+ \leq 25$
) which is highly dependent on
$\gamma$
, and one for the outer (
$x_2^+ \geq 25$
) region which is less dependent on
$\gamma$
. The outer region is already highly nonlinear, which allows for a rapid energy cascade from the actuated mode to smaller scales. Perturbing this region seems to have little effect on the time scale of this cross-scale energy transfer. Near the wall, linear mechanisms dominate; the
$(0, \pm 1)$
-mode, growing significantly under the action of the resolvent forcing term, greatly enhances the nonlinear interactions involving the mode and markedly changes the underlying energy transfer to secondary scales, which may explain the heightened sensitivity of
$\Delta\, t_{\textit{trans}}$
to
$\gamma$
in the near-wall region.
4. Nonlinear energy transfer
The energy content of the secondary modes (figures 8 and 9) are the result of nonlinear interactions amongst all length scales. We wish to disentangle these interactions and focus on the nonlinear energy transfer that specifically drain energy from the actuated
$(0, 1)$
-mode. As in Symon et al. (Reference Symon, Illingworth and Marusic2021) and Ding et al. (Reference Ding, Chung and Illingworth2025), we represent the nonlinear energy transfer from a mode
$(k_{1}, k_{3})$
using the following term:
\begin{equation} \hat N(k_{1}, k_{3}) = -\hat u_i^{(-k_{1}, -k_{3})} \widehat {\frac {\partial u_i u_{\!j}}{\partial x_{\!j}}}^{(k_{1}, k_{3})}. \end{equation}
We note that
$(\boldsymbol{\cdot} )^{(-k_{1}, -k_{3})}$
refers to the complex conjugate of
$(\boldsymbol{\cdot} )^{(k_{1}, k_{3})}$
. The term
$\hat N(k_{1}, k_{3})$
satisfies
\begin{align} \int _0^{2\delta } \sum \limits _{k_{1}} \sum \limits _{k_{3}} \hat N(k_{1}, k_{3}) {\textrm d}y &= -\int _0^{2\delta } \widehat {u_i \frac {\partial u_i u_{\!j}}{\partial x_{\!j}}}^{(0, 0)} \nonumber\\[6pt]&= \displaystyle\frac {1}{L_{1} L_{3}} \int _0^{2\delta } \int _0^{L_{3}} \int _0^{L_{1}} u_i \displaystyle\frac {\partial u_i u_{\!j}}{\partial x_{\!j}} {\textrm d}x_1 {\textrm d}x_3 {\textrm d}x_2 \nonumber \\[10pt] &= 0, \end{align}
due to continuity and the no-penetration boundary conditions at the walls. This reflects the fact that the nonlinear transfer (NLT) in channel flow does not add or remove energy from the system but simply redistributes it between scales (Tennekes & Lumley Reference Tennekes and Lumley1972; Pope Reference Pope2000). We can express
$\hat N (k_{1}, k_{3})$
as a sum of contributions from interacting scales,
\begin{eqnarray} \hat N(k_{1}, k_{3}) &= - \displaystyle \sum \limits _{s_1} \sum \limits _{s_3} \hat u_i^{(-k_{1}, -k_{3})} \widehat {\displaystyle\frac {\partial u_i}{\partial x_{\!j}}}^{(s_1, s_3)} \hat u_{\!j} ^{(k_{1} - s_1, k_{3} - s_3)}. \end{eqnarray}
We refer to an individual contribution to the sum as
$\hat M^{(k_{1}, k_{3})} (s_1, s_3)$
, defined as follows:
\begin{eqnarray} \hat M^{(k_{1}, k_{3})}(s_1, s_3) = - 2 {\textrm{Re}} \left \{ \hat u_i^{(-k_{1}, -k_{3})} \widehat {\displaystyle\frac {\partial u_i}{\partial x_{\!j}}}^{(s_1, s_3)} \hat u_{\!j} ^{(k_{1} - s_1, k_{3} - s_3)} \right \} . \end{eqnarray}
The term
$\hat M^{(k_{1}, k_{3})}(s_1, s_3)$
represents the energy transfer from the
$(k_{1}, k_{3})$
-mode to the
$(s_1, s_3)$
-mode and satisfies the following properties:
These properties are also a consequence of the incompressibility condition and the boundary conditions. For example, to obtain (4.5), consider
$\hat u_i^{(-k_{1}, -k_{3})} \widehat {\partial u_i/\partial x_{\!j}}^{(s_1, s_3)} \hat u_{\!j} ^{(k_{1} - s_1, k_{3} - s_3)}$
and
$\hat u_i^{(s_1, s_3)} \widehat {\partial u_i/\partial x_{\!j}}^{(-k_{1}, -k_{3})} \hat u_{\!j} ^{(k_{1} - s_1, k_{3} - s_3)}$
which, along with their conjugates, constitute the quantities
$\hat M^{(k_{1}, k_{3})}(s_1, s_3)$
and
$\hat M^{(s_1, s_3)}(k_{1}, k_{3})$
, respectively. These terms satisfy the following:
\begin{align} & \int _0^{2\delta } \left ( \hat u_i^{(-k_{1}, -k_{3})} \widehat {\frac {\partial u_i}{\partial x_{\!j}}}^{(s_1, s_3)} \hat u_{\!j} ^{(k_{1} - s_1, k_{3} - s_3)} + \hat u_i^{(s_1, s_3)} \widehat {\frac {\partial u_i}{\partial x_{\!j}}}^{(-k_{1}, -k_{3})} \hat u_{\!j} ^{(k_{1} - s_1, k_{3} - s_3)} \right ) {\textrm d}x_2 \nonumber\\[6pt]&\qquad = \int _0^{2\delta } \textrm{i} (s_1-k_{1}) \hat u_i^{(-k_{1}, -k_{3})} \hat u_i ^{(s_1, s_3)} \hat u_1 ^{(k_{1} - s_1, k_{3} - s_3)} \nonumber\\[6pt]& \qquad \quad + \frac {\partial }{\partial x_2} \left ( \hat u_i^{(-k_{1}, -k_{3})} \hat u_i ^{(s_1, s_3)} \right )\hat u_2 ^{(k_{1} - s_1, k_{3} - s_3)} \nonumber\\[6pt]& \qquad \quad + \textrm{i} (s_3 - k_{3}) \hat u_i^{(-k_{1}, -k_{3})} \hat u_i ^{( s_1, s_3)} \hat u_3 ^{(k_{1} - s_1, k_{3} - s_3)} {\textrm d}x_2 \nonumber\\[6pt]&\qquad = - \int _0^{2\delta } \hat u_i^{(-k_{1}, -k_{3})} \hat u_i ^{(s_1, s_3)} \left ( \textrm{i}(k_{1}-s_1) \hat u_1 ^{(k_{1} - s_1, k_{3} - s_3)} \right ) \nonumber\\[6pt] & \qquad \quad + \hat u_i^{(-k_{1}, -k_{3})} \hat u_i ^{(s_1, s_3)} \left ( \frac {\partial }{\partial x_2} \hat u_2 ^{(k_{1} - s_1, k_{3} - s_3)} \right ) \nonumber\\[6pt] & \qquad \quad + \hat u_i^{(-k_{1}, -k_{3})} \hat u_i ^{(s_1, s_3)} \left ( \textrm{i} (k_{3} - s_3) \hat u_3 ^{(k_{1} - s_1, k_{3} - s_3)} \right ) {\textrm d}x_2 \nonumber\\[6pt] &\qquad = 0. \end{align}
Note,
$\textrm{i} = \sqrt {-1}$
and
$(\boldsymbol{\cdot} )_i$
denotes a spatial direction. To obtain the wall-normal derivative term in the second equality, we integrate by parts and apply the no-penetration boundary condition; we enforce continuity for
$\hat u_{\!j}^{(k_{1} -s_1, k_{3}- s_3)}$
to obtain the final result. Equation (4.5) implies that the scale-to-scale energy transfer is conservative; (4.6), a consequence of (4.5), implies that it purely captures the energy transferred from one scale to a different scale and excludes self-interactions. We finally define
$ \Delta \hat M^{(k_{1}, k_{3})}$
as
$\hat M^{(k_{1}, k_{3})} - \hat M_0^{(k_{1}, k_{3})}$
, where
$\hat M_0^{(k_{1}, k_{3})}$
is defined as (4.4) using the unforced velocity field
$\boldsymbol{u_0}$
.
4.1. Interacting modes
Figure 10 shows the total NLT from the
$(k_{1}, k_{3}) = (0, 1)$
mode integrated over the wall-normal direction. The integrated transfer is negative and confirms that nonlinear effects are indeed draining energy from the actuated scale. The trends in the plot echo the results in figure 3(a): the larger the forcing term, the larger the integrated transfer to unforced scales. The total energy transfer tends to peak during the decay of the streak, approximately
$0.06{-}0.09 \; \delta / u_\tau$
later than the streak energy. Considering the total energy transfer normalised by forcing magnitude
$|\kappa |^2$
, we see that the plots collapse for all values of
$\gamma$
, for
$t u_\tau / \delta \lt 0.8$
. The initial growth rate is the same across all forcing intensities and its magnitude scales with
$|\gamma |^2$
. The energy transfer mechanism at early times thus seems to be similar for all cases, regardless of forcing magnitude. The larger the forcing, however, the earlier the energy transfer reaches its maximum and starts its decay. At later times, additional multiscale effects, more prominent for the strongly forced cases, likely take over and help drain energy more efficiently from the
$(0, 1)$
-mode.
(a) Integrated nonlinear energy transfer from the
$(0,1)$
-mode to the non-forced modes; (b) integrated nonlinear energy transport from the
$(0,1)$
-mode normalised by the forcing magnitude
$|\kappa |^2$
. The cases plotted are
$\gamma = 1\, \%$
(black
$\circ$
),
$2\, \%$
(cyan
$\bullet$
),
$5 \, \%$
(purple
$\times$
) and
$10 \, \%$
(red
$*$
).

Nonlinear energy transfer
$\Delta \hat M^{(0, 1)} (s_1, s_3)$
for
$\gamma = 2\, \%$
(a–c) and
$\gamma = 10\, \%$
(d–f), at
$tu_\tau / \delta = 0.35$
(a,d),
$0.74$
(b,e),
$2$
(c,f).

In figure 11, we plot the streamwise–spanwise spectra for
$\hat M^{(0, 1)}( s_1, s_3)$
at different times. Only wavenumbers corresponding to
$ s_1, s_3 \in [-4, 4]$
are shown for clarity; the remaining wavenumbers interact negligibly with the
$(0,1)$
-mode. We observe that
$\hat M^{(0, 1)} (s_1, s_3) \approx \hat M^{(0, 1)} (-s_1, s_3)$
. Since the target mode extends the entire length of the channel, the streamwise phase of the interacting modes matters little once the results are averaged over the ensemble of simulations. The figure reveals that the nonlinear energy transfer is dominated by energy transfer to the
$(0, 2)$
- and
$(1, 1)$
-modes. This pattern is consistent across forcing intensities, though only the cases for
$\gamma = 2\, \%$
and
$\gamma = 10\, \%$
are shown. The qualitatively similar results across forcing magnitudes highlight the privileged role of the
$(0, 2)$
- and
$(1, 1)$
-modes at exchanging energy with the actuated mode. Since the transfer to the
$(0, 2)$
-mode results from the self-interaction of the actuated mode, the importance of this particular scale is not surprising. Figure 12 shows the
$x_2$
-integrated energy transferred to the
$(1, 1)$
- and
$(0, 2)$
-modes only. As in figure 10(b), the initial growth rate of the magnitude-normalised quantities across forcing amplitudes, with the plot for
$\gamma = 10\, \%$
peaking and starting its decay earlier than for
$\gamma = 2\, \%$
. Another notable observation is that, across forcing amplitudes, the energy transfer to the
$(0, 2)$
-mode occurs at a faster time scale than the transfer to the
$(1, 1)$
-mode.
(a) Integrated nonlinear energy transfer from the
$(0,1)$
-mode to the
$(0, 2)$
-mode (
${\cdots}$
) and the
$(1,1)$
-mode (
$-\,-$
); (b) integrated nonlinear energy transfer from the
$(0,1)$
-mode normalised by the forcing magnitude
$|\kappa |^2$
. The cases plotted are
$2\, \%$
(cyan) and
$10 \, \%$
(red).

(a) Total nonlinear energy transfer from the
$(0,1)$
-mode, i.e.
$ \sum _{s_1} \sum _{s_3} \overline {\Delta \hat M^{(0, 1)}(s_1, s_3)} \delta / u_\tau ^3$
. The cases plotted are for (a)
$\gamma = 2\, \%$
and (b)
$\gamma = 10 \, \%$
. The black lines are contours of
$\overline {|\hat u_1^{(0, 1)}|^2/2}$
and represent
$10\, \%, \; 25 \, \%, 50\, \%, 75\, \%$
and
$90\, \%$
of the maximum value across
$x_2$
and
$t$
.

In figure 13, we study the spatial distribution of the nonlinear energy transfer term
$\hat M^{(0, 1)}\!$
. We observe that, for
$\gamma = 2\, \%$
and
$\gamma = 10\, \%$
, the additional NLT to other scales due to the induced streak is centred on two wall-normal locations,
$x_2^+ \approx 16$
located in the buffer layer, and
$x_2^+ \approx 75$
located in the outer region. This is consistent across all forcing magnitudes. The presence of two hubs of energy transfer echo the results in figure 9(d), which similarly shows two distinct regions each characterised by a time scale of energy transfer from the
$(0, 1)$
-mode to smaller scales.
(a,b)
$\overline {\Delta \hat M^{(0,1)}(0, 2)} \delta / u_\tau ^3$
and (c,d)
$\overline {\Delta \hat M^{(0,1)}(1, 1)} \delta / u_\tau ^3$
in the
$t-x_2$
plane. The cases plotted are for (a,c)
$\gamma = 2\, \%$
and (b,d)
$\gamma = 10 \, \%$
. The black lines are contours of
$\overline {|\hat u_1^{(0, 1)}|^2/2}$
and represent
$10\, \%, \; 25 \, \%, 50\, \%, 75\, \%$
and
$90\, \%$
of the maximum value across
$x_2$
and
$t$
.

In figure 14, we plot the nonlinear energy transfer to the
$(0, 2)$
- and
$(1, 1)$
-modes, for
$\gamma = 2\, \%$
and
$\gamma = 10\, \%$
. The figure reveals that the transfer to the
$(0, 2)$
-mode accounts for the transfer at
$x_2^+ \leq 25$
, while the transfer to the
$(1, 1)$
-mode accounts for the transfer at
$x_2^+ \gt 25$
. In the buffer layer, the forced
$(0, 1)$
-mode thus tends to transfer its energy to a mode that is twice periodic in the spanwise direction; the induced streak splits into two branches, which can be seen in figure 15(a,b). In the outer region, the nonlinear energy transfer favours the
$(1, 1)$
-mode, suggesting that the induced streak breaks up along the streamwise direction. This is seen in figure 15(c,d), which show the streak meandering but not splitting into two branches as for
$x_2^+ = 16$
. We note that the transfer from the
$(0, 1)$
-mode to the
$(0, 2)$
-mode is due to the self-interaction of the actuated mode, and the term
$\hat u^{(0, -1)}$
is squared in the expression for
$\hat M ^{(0, 1)}(0, 2)$
(4.4). This along with the fact that the energy transfer to the
$(0, 2)$
-mode is predominant in the near-wall region explains the increased sensitivity of the energy transfer time scale
$\Delta\, t_{\textit{trans}}$
to
$\gamma$
for
$x_2^+ \leq 25$
(figure 9
d). More specifically, a higher value of
$\gamma$
disproportionately intensifies the transfer of energy to the
$(0, 2)$
-mode, causing the NLT of energy near the wall to occur earlier – closer to the focus of
$\hat M^{(0, 1)}(0, 2)$
– and decreasing
$\Delta\, t_{\textit{trans}}$
.
Ensemble-averaged streamwise velocity deviation
$\overline {\Delta u_1}$
for (a,c)
$\gamma = 2\, \%$
and (b,d)
$\gamma = 10 \, \%$
. The results are for
$x_1 = 0$
and two wall-normal heights: (a,b)
$x_2^+ = 16$
and (c,d)
$x_2^+ = 75$
. Before ensemble averaging, the
$(0, 1)$
-mode for each ensemble member is multiplied by
$\textrm{e}^{-i\angle \kappa }$
to ensure their phase alignment.

4.2. Mechanism of nonlinear energy transfer
To shed light on the mechanism of this nonlinear energy transfer, we split the NLT
$\hat M^{(0, 1)}(s_1, s_3)$
into the terms that, respectively, reflect the contributions of the streamwise, spanwise and wall-normal nonlinear advection. Specifically, in order to study the nonlinear energy transfer from a mode
$(k_{1}, k_{3})$
due to streamwise nonlinear advection – henceforth referred to as NLT
$_1$
(
$x_1$
), we define
$\hat M_{1}^{(k_{1}, k_{3})}(s_1, s_3)$
as
\begin{equation} \hat M_{1} ^{(k_{1}{,} k_{3})}(s_1, s_3) := - 2 {\textrm{Re}} \left \{ \hat u_i^{(-k_{1}, -k_{3})} \widehat {\displaystyle\frac {\partial u_i}{\partial x_1}}^{(s_1, s_3)} \hat u_1 ^{(k_{1} - s_1, k_{3} - s_3)} \right \}. \end{equation}
We similarly define
$\hat M_{2}$
and
$\hat M_{3}$
to study NLT
$_2$
and NLT
$_3$
, the nonlinear energy transfer due to wall-normal and spanwise gradients self-advection, respectively.
Nonlinear energy transfer from the
$(0,1)$
-mode to unforced modes, broken down by streamwise (blue
$-\bullet$
), wall-normal (green
$-\blacktriangle$
) and spanwise (red
$-\blacksquare$
) contributions. The cases plotted are (a,c)
$\gamma = 2\, \%$
and (b,d)
$\gamma = 10\, \%$
. Panels (a) and (b) represent the contributions to the total nonlinear energy transport from the
$(0,1)$
-mode, with the black line representing the sum of the contributions; panels (c) and (d) represent the transfer to the
$(0, 2)$
- and
$(1, 1)$
-modes, denoted by
$(\cdots )$
and (
$-\,-$
), respectively.

Nonlinear energy transfer from the
$(0,1)$
-mode split by NLT
$_2$
and NLT
$_3$
contributions: (a,b)
$\overline {\Delta \hat M^{(0,1)}_2} \delta / u_\tau ^3$
and (c,d)
$\overline {\Delta \hat M^{(0,1)}_3} \delta / u_\tau ^3$
. The cases shown are for (a,c)
$\gamma = 2\, \%$
and (b,d)
$\gamma = 10 \, \%$
. The black lines are contours of
$\overline {|\hat u_1^{(0, 1)}|^2/2}$
and represent
$10\, \%, \; 25 \, \%, 50\, \%, 75\, \%$
and
$90\, \%$
of the maximum value across
$x_2$
and
$t$
.

Figure 16 shows that the energy transfer due to the coupling with the spanwise velocity component dominates the total energy transfer to other scales, even before the growth of the injected mode and after the mode completely decays for
$t u_\tau / \delta \gt 2$
. During the growth of the injected mode, the NLT
$_3$
term grows significantly more than the transfer due to self-advection in the streamwise and wall-normal directions. This demonstrates the unique role of NLT
$_3$
in transferring energy to secondary scales and restoring the system to its unforced state. This behaviour is consistent across forcing amplitudes, though only
$\gamma = 2\, \%$
and
$\gamma = 10\, \%$
are shown. The dominance of the NLT
$_3$
contributions also holds for the
$(0, 2)$
- and
$(1, 1)$
-modes individually (figure 16
c,d). We do note that the NLT
$_2$
is relatively important for the
$(0, 2)$
-mode. The induced streak thus tends to shed its energy to secondary modes via spanwise self-advection. We note that these results match those found in Markeviciute & Kerswell (Reference Markeviciute and Kerswell2024), which computes a streak structure via optimal transient growth analysis then studies the growth of secondary instabilities added to the streak. Authors find that suppressing the pushover mechanism, i.e. the spanwise advection of the secondary perturbation due to the streak, prevents the growth of these secondary instabilities. Thus, both this work and Markeviciute & Kerswell (Reference Markeviciute and Kerswell2024) agree that coupling via spanwise gradients dominate the interaction between the streak and secondary modes.
Additionally, by considering the wall-normal variations (figure 17), we see that spanwise self-advection is the primary pathway across all wall-normal heights by which the forced
$(0, 1)$
-mode sheds its energy. This is consistent across all forcing magnitudes, though only the cases for
$\gamma =2\, \%$
and
$\gamma = 10\, \%$
are shown. For the lightly forced case
$\gamma = 2\, \%$
, we do note a significant NLT
$_2$
contribution at
$x_2^+ \approx 18$
(figure 17
a), but as
$\gamma$
increases, the NLT
$_2$
contribution is quickly overtaken by the NLT
$_3$
. The NLT
$_3$
term thus has a stronger dependence on
$\gamma$
than the NLT
$_2$
term, especially close to the wall; this suggests its key role in the nonlinear interactions which grow rapidly with
$\gamma$
. The growth of secondary instabilities due to push-over occurs locally at both foci of nonlinear energy transfer, and both the
$(1, 1)$
- and
$(0, 2)$
-modes benefit from this energy transfer mechanism.
4.3. Quasilinear approximation
In this section, we seek to determine whether the instantaneous behaviours of the nonlinear energy transfer can be simply modelled by the interactions of the principal response mode with the ‘background’ turbulence – that is, the turbulence frozen in its state at
$t = 0$
. By neglecting subsequent nonlinear feedback, the velocity field after the injection of the mode can be represented as
$\boldsymbol{u}_\psi := \boldsymbol{u_0} + \kappa \sigma _1 \boldsymbol{\psi _1} \textrm{exp}(\textrm i 2 \pi / L_{3}) + \kappa ^* \sigma _1 \boldsymbol{\psi _1}^* \textrm{exp}(-\textrm i 2 \pi / L_{3})$
. Using this field, we define the following nonlinear energy transfer term for the quasilinear field analogously to (4.4):
\begin{align} \widetilde { M} ^{(k_{1}, k_{3})}(s_1, s_3) := - 2 {\textrm{Re}} \left \{ \hat u_{\psi , i}^{(-k_{1}, -k_{3})} \widehat {\displaystyle\frac {\partial u_{\psi , i}^{(s_1, s_3)}}{\partial x_{\!j}}} \hat u_{\psi , j} ^{(k_{1} - s_1, k_{3} - s_3)} \right \}, \end{align}
where
$u_{\psi , i}$
is the
$i$
th component of
$\boldsymbol u_\psi$
. This term captures the nonlinear energy transfer from the
$(k_{1}, k_{3})$
-mode to the
$(s_1, s_3)$
-mode due to interactions of the injected mode with the background turbulence and ignoring the self-interactions of
$\boldsymbol u_\psi$
at every time step. We can thus view the term
$\widetilde {M}^{(k_{1} , k_{3})}$
as a quasilinear model of the nonlinear energy transfer. We define the quantities representing the NLT
$_1$
, NLT
$_2$
and NLT
$_3$
contributions for the
$\boldsymbol u_\psi$
field as
\begin{align} \widetilde { M}_{1} ^{(k_{1}, k_{3})}(s_1, s_3) := - 2 {\textrm{Re}} \left \{\hat u_{\psi , i}^{(-k_{1}, -k_{3})} \widehat {\displaystyle\frac {\partial u_{\psi , i}}{\partial x_1}}^{(s_1, s_3)} \hat u_{\psi , 1} ^{(k_{1} - s_{1}, k_{3} - s_{3})} \right \}. \end{align}
Quasilinear approximation
$\widetilde {M}^{(0,1)}$
for the nonlinear energy transfer from the
$(0,1)$
-mode, normalised by the forcing magnitude
$|\kappa |^2$
. Panels (a) and (b) correspond to the transfer to the
$(0,2)$
-mode, and panels (c) and (d) correspond to the transfer to the
$(1,1)$
-mode. The cases plotted are for (a,c)
$\gamma = 2\, \%$
and (b,d)
$\gamma = 10\, \%$
. The black lines represent contours of
$\hat { M}^{(0,1)}$
and correspond to
$10\, \%$
,
$25\, \%$
,
$50\, \%$
,
$75\, \%$
and
$90\, \%$
of the minimum (most negative) value (–), or
$10\, \%$
,
$25\, \%$
,
$50\, \%$
,
$75\, \%$
and
$90\, \%$
of the maximum (most positive) value (
$-\,-$
).

Nonlinear energy transport from the
$(0,1)$
-mode to the (a,b)
$(0,2)$
-mode and the (c,d)
$(1,1)$
-mode, broken down by streamwise (blue
$-\bullet$
), wall-normal (green
$-\blacktriangle$
) and spanwise (red
$-\blacksquare$
) contributions, and normalised by forcing magnitude. The solid lines correspond to the nonlinear energy transfer in the quasilinear model
$[\overline {\widetilde { M}^{(0,1)} / |\kappa |^2}]$
, while the dashed or dotted lines correspond to
$[\overline {\hat { M}^{(0,1)} / |\kappa |^2}]$
, the results from DNS (also shown in figure 16). The cases plotted are (a,c)
$\gamma = 2\, \%$
and (b,d)
$\gamma = 10\, \%$
.

We call these terms QLT
$_1$
, QLT
$_2$
and QLT
$_3$
(‘quasilinear transfer,
$x_i$
’), respectively, to distinguish them from NLT
$_1$
, NLT
$_2$
and NLT
$_3$
. Figure 18 shows the quasilinear estimates for the nonlinear energy transfer to the
$(0,2)$
- and
$(1, 1)$
-modes, normalised by forcing amplitude. For the lightly forced case of
$\gamma =2\, \%$
, the quasilinear model predicts the regions of energy loss to the
$(0, 2)$
- and
$(1, 1)$
-modes well, especially for
$t \lt 1 \delta /u_\tau$
(figure 18
a,c). For
$\gamma =10\, \%$
, the quasilinear interactions fail to predict the region of energy transfer to the
$(0, 2)$
-mode, but perform better for the
$(1, 1)$
-mode (figure 18
b,d).
In an attempt to test whether the quasilinear model can predict the general NLT
$_i$
trends observed in the previous section – mainly the dominance of the NLT
$_3$
contributions – we plot the QLT
$_1$
, QLT
$_2$
and QLT
$_3$
quantities in figure 19. We see that the quasilinear estimates for the transfer to mode
$(0,2)$
can roughly predict a strongly negative QLT
$_3$
component for the lightly forced system at early times, but fail to do so for later times. For large forcing amplitudes, the quasilinear mode performs poorly for all times, and even reverses the observed trends for the QLT
$_2$
and QLT
$_3$
components of the NLT. The quasilinear model indirectly assumes that the growth of the secondary modes interacting with the injected streak is slow enough so that the dominant nonlinear interactions remain due to the ‘background’ unperturbed field. The failure of the quasilinear model to explain the trends of the
$(0, 2)$
-mode suggests that the assumption of slow growth cannot be applied to this mode. Thus, in the near-wall region where streak splitting occurs, we suspect that transient growth mechanisms are key to explaining the behaviour mode
$(0, 2)$
. One such mechanism is linearly driven transient algebraic growth, which is responsible for amplifying the injected resolvent mode at early times. In contrast, the quasilinear estimates for the transfer to mode
$(1,1)$
qualitatively match the behaviour of the DNS and reveal a strongly negative QLT
$_3$
contribution and a weakly positive QLT
$_2$
contribution. Outside the buffer layer, the instantaneous nonlinear interactions between the target
$(0,1)$
-mode and the background turbulence can qualitatively explain the streak breakup. For better quantitative accuracy, we suspect that transient growth effects are also important in the outer region.
Magnitude of the streamwise component of the principal resolvent response modes for (a)
$(k_{1}, k_{3}) = (0, 2)$
and (b)
$(k_{1}, k_{3}) = (1, 1)$
. The modes are normalised so that their action pseudonorm is unity. The horizontal lines represent the peak location of (a)
$\hat M^{(0, 1)}(0, 2)$
and (b)
$\hat M^{(0, 1)}(1, 1)$
for
$\gamma = 2\, \%$
(blue) and
$\gamma = 10\, \%$
(red).

In addition to studying the interaction of the transiently growing
$(0, 1)$
-mode with the background turbulence, we consider the linear optimal growth of the
$(0, 2)$
- and
$(1, 1)$
-modes. We repeat the wavelet-based resolvent analysis described in § 2.2 but with spatial parameters
$(k_{1}, k_{3}) = (0, 2)$
and
$(k_{1}, k_{3}) = (1, 1)$
. We use the same spatial grid and choose
$T = 44 \delta / u_\tau$
to allow the modes ample time to decay to zero, which occurs at
$t u_\tau /\delta \approx 20$
. The size of the temporal grid is chosen to be
$N_t = 880$
. Despite the fact that resolvent analysis captures a purely linear process, the
$x_2$
location preferred by the principal resolvent response for
$(k_{1}, k_{3}) = (0, 2)$
and
$(k_{1}, k_{3}) = (1, 1)$
coincide with the foci of nonlinear interaction between the
$(0, 1)$
-mode, and the
$(0, 2)$
- and
$(1, 1)$
-modes, respectively (figure 20). The resolvent modes, which capture the maximal linear growth of the two scales considered, accurately predict the locations of energy exchange with the actuated mode, and thus, transient growth via linear mechanisms can dictate the spatial structure of the nonlinear energy cascade. Moreover, we observe that the principal resolvent response mode for
$(k_{1}, k_{3}) = (0, 2)$
grows faster than the mode for
$(k_{1}, k_{3}) = (1, 1)$
, peaking earlier, which may explain the earlier peaks of
$\hat M^{(0,1)}(0, 2)$
compared with
$\hat M^{(0,1)}(1, 1)$
(figure 12).
The results in figure 20 support works such as Huang et al. (Reference Huang, Toedtli, Chini and McKeon2023), which studies the efficacy of individual dyadic interactions at exciting an energetic scale in the channel. The influence of individual dyadic contributions to the nonlinear term of the Navier–Stokes equations is measured via the projection of the contributions onto the principal resolvent forcing mode for the scale of interest, and dyadic interactions highly aligned with the forcing mode are deemed important contributors to turbulence. The underpinning assumption is that the response mode corresponding to the forcing mode indeed capture the behaviour of turbulence in the channel, which figure 20 suggests is true, even for secondary scales like the
$(0, 2)$
- and
$(1, 1)$
-modes. Figure 20 also helps explain the success of restricted nonlinear models at replicating turbulent statistics (Thomas et al. Reference Thomas, Lieu, Jovanović, Farrell, Ioannou and Gayme2014; Farrell, Gayme & Ioannou Reference Farrell, Gayme and Ioannou2017; Gayme & Minnick Reference Gayme and Minnick2019). In restricted nonlinear models, the mean profile is governed by a modified version of the fully nonlinear Navier–Stokes equations where only a reduced subset of fluctuation scales contributes to the nonlinear advection terms, while the fluctuations obey the linearised Navier–Stokes equations about the mean profile. The location of the modes in figure 20 indeed suggest that linear mechanisms are enough to at least predict the correct spatial distribution of the nonlinear energy transfer to the two most important secondary modes involved in the energy cascade in this experiment.
5. Conclusions
In this work, we study the growth of time-localised resolvent modes in the minimal flow unit at
$Re_\tau \approx 186$
. We formulated resolvent analysis in a wavelet-basis in time that endows the resolvent modes with transient information, and obtained a linearly optimal time-localised forcing mode and its corresponding transient response mode.
Resolvent analysis ignores the feedback between the velocity fluctuations and the nonlinear terms; we thus tested the optimality of the resolvent forcing within a DNS of a minimal flow unit by numerically injecting the principal resolvent forcing mode into the flow at varying amplitudes. This allowed us to investigate the interactions between the transiently growing linear response mode and the nonlinear effects of the turbulent flow. We compared the resulting flow with one forced by the first suboptimal forcing mode and another forced by a spatially random forcing. The principal resolvent forcing produces a larger transient energy growth than the suboptimal mode, but the energy amplification is notably lower in both cases compared with the linearised case. Both systems were significantly more amplified than the random forcing case. In all cases, despite the fact that the injected forcing term is small compared with the initial nonlinearities, the amplification of the velocity perturbation due to linear mechanisms is significant enough that the simulated fields track the optimal linear response for a short time.
The nonlinearities of turbulence interrupt the initial algebraic energy growth driven by the linear dynamics of the flow. This is seen in all cases forced by the principal resolvent mode, and the amplitude of the resolvent forcing affects how closely the turbulent trajectory behaves like the optimal resolvent response mode. Across all forcing amplitudes, the initial growth phase is similar, and the systems peak at roughly the same time, but, the higher the forcing amplitude, the faster the decay of the system back to the unforced turbulent system. The more intense turbulence in the high-amplitude-forcing cases is more effective at damping the effects of the initial forcing. We observe that the induced velocity perturbation field for the actuated Fourier mode matches the resolvent mode better, both in magnitude and spatiotemporal structure, within the near-wall region, where viscous effects are more prominent. During the decay of the streak decays, the spectral energy content of the simulations becomes increasingly multiscale due to a transfer of energy to the non-forced spatial scales. This cross-scale energy transfer is more prominent and occurs faster for the high-amplitude-forcing cases.
Nonlinear effects lead to streak breakdown, and by considering the nonlinear energy transfer from the induced streak to the non-actuated modes, two secondary modes are found to play a dominant role in draining the streak of its energy and curtailing its growth. The first mode, constant in the streamwise direction and twice periodic in the spanwise direction, corresponds to a splitting of the streak into two branches and dominates the energy transfer in the near-wall region, including the buffer layer. The second, once periodic in the streamwise and spanwise directions, corresponds to a streamwise breakup of the streak and dominates the energy transfer in the outer region. The branching mode receives energy from the actuated mode via the self-interaction of the injected streak, which explains the particular sensitivity of the near-wall region to forcing amplitude. Though a larger forcing amplitude accelerates the energy transfer from the actuated mode to smaller scales across all wall-normal heights, this effect is indeed more prominent in near the wall than in the outer region.
The streak interacts with these secondary modes mostly through an NLT
$_3$
-type energy transfer, i.e. via the spanwise nonlinear self-advection term. We model the predicted nonlinear energy transfer by computing the interaction of the injected streak with the background turbulence of the initial conditions used, and find that these one-way interactions generally predict the trends for the streak breakup that dominates in the outer region. This crude model cannot, however, predict the dominance of the NLT
$_3$
contribution that occurs during the streak splitting near the wall. We postulate that in the near-wall region and the buffer layer, transient growth phenomena are necessary to explain the behaviour of the
$(0, 2)$
-mode. We compute the transiently growing principal resolvent modes for the two preferred secondary scales, and find that they are located exactly at the foci of nonlinear energy exchange with the actuated scale. The structures found using resolvent analysis predict the
$x_2$
distribution of the nonlinear energy transfer to the
$(0, 2)$
- and
$(1, 1)$
-modes in the nonlinear DNS, further proving that resolvent modes can be informative of turbulent flows despite using the linearised equations of motion.
Testing the effectiveness of the principal resolvent forcing mode reveals valuable insights on the stability of the streak, the mechanism by which it sheds energy, and the sensitivity of this nonlinear energy transfer mechanism to forcing amplitude. To find structures more effective than the resolvent forcing mode at actuating the minimal flow unit, one could use nonlinear optimisation techniques (Kerswell Reference Kerswell2018; Heide & Hemati Reference Heide and Hemati2023). These broadly aim to maximise the growth of kinetic energy within a user-defined time window, and enforce the satisfaction of the (nonlinear) Navier–Stokes equations as an optimisation constraint. Nevertheless, an important advantage of wavelet-based resolvent analysis is its computational efficiency and tractability.
Funding
This work is supported in part by the European Research Council under the Caust grant ERC-AdG-101018287 and AFOSR grant FA9550-22-1-0109.
Declaration of interests
The authors report no conflict of interest.



















































































































































































