2.1. General nonlinear system

We consider the nonlinear autonomous system

(2.1)\begin{equation} \frac{\textrm{d}}{\textrm{d}t}{\boldsymbol{q}}(t) = {\boldsymbol{f}}\left({\boldsymbol{q}}(t)\right) \end{equation}

with state ${\boldsymbol {q}}(t)$. In three-dimensional incompressible fluid flows, the state is the three-dimensional vector velocity field along with the scalar pressure.

In the harmonic resolvent framework, we are interested in studying the amplification mechanisms of small perturbations ${\boldsymbol {q}}'(t)$ about a time-varying base flow that is periodic with period $T$, given by

(2.2)\begin{equation} {\boldsymbol{Q}}(t) = \sum_{\omega \in \Omega_b} \hat{\boldsymbol{Q}}_{\omega}\,\textrm{e}^{\textrm{i}\,\omega t}, \end{equation}

with $\Omega _b\subset ({2{\rm \pi} }/{T})\mathbb {Z}$. The base flow does not need to satisfy the governing equations and $\Omega _b$ usually contains a small subset of frequencies that approximate the dynamics of the large coherent structures present in the flow. We proceed by seeking perturbations of the form

(2.3)\begin{equation} {\boldsymbol{q}}'(t) = \sum_{\omega \in \Omega}\hat{\boldsymbol{q}}'_{\omega}\,\textrm{e}^{\textrm{i}\,\omega t}, \end{equation}

with $\Omega \subset ({2{\rm \pi} }/{T})\mathbb {Z}$. Usually $\Omega _b \subset \Omega$, as $\Omega$ is the set of temporal frequencies associated with the flow structures that one wishes to resolve. (In general, the set $\Omega$ need not be restricted to multiples of the fundamental frequency $2{\rm \pi} /T$, although we do make this assumption in this paper.) Upon substituting the decomposition ${\boldsymbol {q}}(t)={\boldsymbol {Q}}(t)+{\boldsymbol {q}}'(t)$ in (2.1) we obtain

(2.4)\begin{equation} \frac{\textrm{d}}{\textrm{d}t}{\boldsymbol{q}}'(t) = \underbrace{\mathcal{D}_{{\boldsymbol{q}}}{\,\boldsymbol{f}}\left({\boldsymbol{Q}}(t)\right)}_{{\boldsymbol{\mathsf{A}}}(t)}{\boldsymbol{q}}'(t) + {\boldsymbol{h}}'(t), \end{equation}

where

(2.5)\begin{equation} {\boldsymbol{h}}'(t) = \left[-\frac{\textrm{d}}{\textrm{d}t}{\boldsymbol{Q}}(t) + {\boldsymbol{f}}\left({\boldsymbol{Q}}(t)\right)\right] + o\left(\|{\boldsymbol{q}}'(t)\|\right). \end{equation}

The first term in (2.5) is the error associated with the base flow not satisfying the governing equations, while the second represents the higher-order terms in the dynamics. Before proceeding further, we observe that ${\boldsymbol{\mathsf{A}}}(t)$ is periodic with period $T$ (since ${\boldsymbol {Q}}$ is periodic), and it can therefore be represented in terms of a Fourier series, analogous to (2.2). We then obtain the following expression for the perturbation at frequency $\omega$:

(2.6)\begin{equation} \textrm{i}\,\omega \hat{\boldsymbol{q}}'_{\omega} = \sum_{\alpha\in \Omega} \hat{\boldsymbol{\mathsf{A}}}_{\omega-\alpha} \hat{\boldsymbol{q}}'_{\alpha} + \hat{\boldsymbol{h}}'_{\omega} \quad \forall \omega \in \Omega. \end{equation}

We neglect frequencies $\omega$ that are not in $\Omega$. For ease of notation, let $\hat {\boldsymbol {q}}'$ be the vector of $\hat {\boldsymbol {q}}'_\omega$ for all frequencies $\omega \in \Omega$, and let $\hat {\boldsymbol {h}}'$ be defined similarly. We then define the operator ${\boldsymbol{\mathsf{T}}}$ by

(2.7)\begin{equation} \big[{\boldsymbol{\mathsf{T}}}\hat {\boldsymbol{q}}'\big]_\omega = \textrm{i}\,\omega \hat {\boldsymbol{q}}_\omega' - \sum_{\alpha\in\Omega} \hat {\boldsymbol{\mathsf{A}}}_{\omega-\alpha}\hat {\boldsymbol{q}}_\alpha'. \end{equation}

Before introducing the formal definition of the harmonic resolvent, it is necessary to address a subtlety. If the base flow ${\boldsymbol {Q}}(t)$ is an exact solution of (2.1), then ${\boldsymbol{\mathsf{T}}}$ is singular and its null space contains a vector associated with the neutrally stable direction of phase shift about the base flow given by the Fourier coefficients of $({\textrm {d}}/{\textrm {d}t}) {\boldsymbol {Q}}(t)$. Similarly, if ${\boldsymbol {Q}}(t)$ is an approximate solution of (2.1), then ${\boldsymbol{\mathsf{T}}}$ is nearly singular along the direction of phase shift. In either case, effects associated with shifting the original base flow in time can dominate and obscure the genuine amplification mechanisms of interest. We define the harmonic resolvent in a way that removes the phase shift and preserves the desired steady-state periodic response up to a constant shift in time. In particular, we restrict ${\boldsymbol{\mathsf{T}}}$ to a subspace $\Sigma$ that is orthogonal to the direction of phase shift given by $\widehat{\textrm{d}{\boldsymbol{Q}}/{\textrm{d}t}} $. This is analogous to constructing a Poincaré map by reducing the dynamics onto a codimension-1 surface pierced by the limit cycle (Guckenheimer & Holmes Reference Guckenheimer and Holmes2002). We notice that when we take $\hat {\boldsymbol {q}}'\in \Sigma$, (2.6) ensures that ${\boldsymbol {h}}'$ must be in the range of the restricted operator ${\boldsymbol{\mathsf{T}}} \vert _{\Sigma }$. Therefore, we define the harmonic resolvent operator on $W_{\Sigma } = Range ({\boldsymbol{\mathsf{T}}} \vert _{\Sigma } )$ as

(2.8)\begin{equation} {\boldsymbol{\mathsf{H}}} = \left( \left. {\boldsymbol{\mathsf{T}}} \right\vert_{\Sigma} \right)^{{-}1}. \end{equation}

Details on efficient computation with this operator can be found in appendix A.

Finally, after removing the phase shift, (2.6) may be written as

(2.9)\begin{equation} \hat{\boldsymbol{q}}' = {\boldsymbol{\mathsf{H}}}\hat{\boldsymbol{h}}', \end{equation}

where $\hat {\boldsymbol {q}}'\in \Sigma$, and ${\boldsymbol {h}}'\in W_{\Sigma }$ now live in compatible codimension-1 subspaces. Therefore, the harmonic resolvent describes the steady-state dynamics of small periodic perturbations $\hat {\boldsymbol {q}}'$ about a periodic base flow in response to a periodic input forcing $\hat {\boldsymbol {h}}'$, up to a constant shift in time. Note that, if external inputs are present (such as a control input, or noise forcing the system), these enter into the formulation in the same way that ${\boldsymbol {h}}'$ does, and this leads to the harmonic transfer function of Wereley (Reference Wereley1991). A more thorough discussion of the characteristics of the harmonic resolvent and its connection with the usual resolvent operator is given at the end of § 2.2.

2.2. Bilinear system: incompressible fluid flow

We now consider an incompressible fluid flow governed by the Navier–Stokes equations, given by

(2.10)\begin{equation} \begin{array}{c} \dfrac{\partial}{\partial t}{\boldsymbol{u}} + {\boldsymbol{u}}\boldsymbol{\cdot} \boldsymbol{\nabla} {\boldsymbol{u}} ={-}\boldsymbol{\nabla} p + Re^{{-}1}\nabla^2 {\boldsymbol{u}} ,\\ \boldsymbol{\nabla} \boldsymbol{\cdot} {\boldsymbol{u}} = 0. \end{array} \end{equation}

Here, ${\boldsymbol {u}}({\boldsymbol {x}},t)$ and $p({\boldsymbol {x}},t)$ are the velocity and pressure, respectively, over the spatial domain $\mathcal {X}\subseteq \mathbb {R}^3$. For ease of notation, we will drop the explicit dependence on ${\boldsymbol {x}}$ from here on. Equation (2.10) can be written compactly as

(2.11)\begin{equation} \frac{\partial}{\partial t} \underbrace{\begin{bmatrix} {\boldsymbol{\mathsf{I}}} & 0 \\ 0 & 0 \end{bmatrix}}_{{\boldsymbol{\mathsf{M}}}} \begin{bmatrix} {\boldsymbol{u}}(t)\\ p(t)\end{bmatrix} = \underbrace{\begin{bmatrix} Re^{{-}1}\nabla^2 & -\boldsymbol{\nabla} \\ -\boldsymbol{\nabla} \boldsymbol{\cdot} & 0 \end{bmatrix}}_{{\boldsymbol{\mathsf{L}}}} \begin{bmatrix}{\boldsymbol{u}}(t)\\ p(t)\end{bmatrix} + \underbrace{\begin{bmatrix} -{\boldsymbol{u}}(t)\boldsymbol{\cdot} \boldsymbol{\nabla} {\boldsymbol{u}}(t) \\ 0 \end{bmatrix}}_{{\boldsymbol{g}}\left({\boldsymbol{u}}(t),{\boldsymbol{u}}(t)\right)}. \end{equation}

We denote the state vector by ${\boldsymbol {q}}=({\boldsymbol {u}},p)$, and consider perturbations about a periodic base flow, as in § 2.1,

(2.12)\begin{equation} {\boldsymbol{q}}(t) = {\boldsymbol{Q}}(t)+{\boldsymbol{q}}'(t) = \sum_{\omega \in \Omega_b}\hat{\boldsymbol{Q}}_{\omega}\,\textrm{e}^{\textrm{i}\,\omega t} + \sum_{\omega \in \Omega}\hat{\boldsymbol{q}}'_{\omega}\,\textrm{e}^{\textrm{i}\,\omega t}, \end{equation}

where $\Omega _b \subseteq \Omega \subset ({2{\rm \pi} }/{T})\mathbb {Z}$. As before, we seek an input–output representation for the perturbations ${\boldsymbol {q}}'$. Substituting (2.12) in (2.11), and neglecting frequencies $\omega \notin \Omega$, we obtain

(2.13)\begin{equation} \textrm{i}\,\omega {\boldsymbol{\mathsf{M}}} \hat{\boldsymbol{q}}'_{\omega} = {\boldsymbol{\mathsf{L}}} \hat{\boldsymbol{q}}'_{\omega} + \sum_{\alpha \in \Omega}\big[{\boldsymbol{g}}(\hat{\boldsymbol{Q}}_{\omega-\alpha},\hat{\boldsymbol{q}}'_{\alpha}) + {\boldsymbol{g}}(\hat{\boldsymbol{q}}'_{\alpha},\hat{\boldsymbol{Q}}_{\omega-\alpha})\big] + \hat{\boldsymbol{h}}'_{\omega},\quad \forall \omega \in \Omega , \end{equation}

where, as in the previous section, $\hat {\boldsymbol {h}}'_{\omega }$ is the Fourier mode of the base flow error along with the terms that are nonlinear in $\hat {\boldsymbol {q}}'_\omega$ (see (2.5)). We again let $\hat {\boldsymbol {q}}'$ denote the vector of $\hat {\boldsymbol {q}}_\omega '$ for all frequencies $\omega \in \Omega$, and define the operator ${\boldsymbol{\mathsf{T}}}$ by

(2.14)\begin{equation} \big[{\boldsymbol{\mathsf{T}}}\hat{\boldsymbol{q}}'\big]_\omega = (\textrm{i}\,\omega {\boldsymbol{\mathsf{M}}} - {\boldsymbol{\mathsf{L}}})\hat{\boldsymbol{q}}'_\omega -\sum_{\alpha\in\Omega} \big[{\boldsymbol{g}}(\hat{\boldsymbol{Q}}_{\omega-\alpha},\hat{\boldsymbol{q}}'_{\alpha}) + {\boldsymbol{g}}(\hat{\boldsymbol{q}}'_{\alpha},\hat{\boldsymbol{Q}}_{\omega-\alpha}) \big]. \end{equation}

Finally, after removing the phase shift, as per the discussion at the end of § 2.1, (2.13) may be written compactly as

(2.15)\begin{equation} \hat{\boldsymbol{q}}' = {\boldsymbol{\mathsf{H}}}\hat{\boldsymbol{h}}'. \end{equation}

Also, as specified at the end of § 2.1, inputs such as a control signal or an external disturbance enter in the system in the same way as ${\boldsymbol {h}}'$.

At this point, a few comments on the structure of the harmonic resolvent operator are in order. First, note that the number of frequencies in the set $\Omega$ may be infinite (e.g. $\Omega =({2{\rm \pi} }/{T})\mathbb {Z}$), in which case the harmonic resolvent is an infinite-dimensional operator. However, in practice, one truncates its dimensionality by selecting a finite number of frequencies $\omega \in \Omega$ that are considered to be of interest. The dimension of ${\boldsymbol{\mathsf{H}}}$ is therefore proportional to the number of frequencies in $\Omega$. As mentioned previously, we usually consider the frequencies $\Omega _b$ in the base flow to be a subset of the frequencies $\Omega$ of the perturbations. That is, one usually wishes to study the dynamics of perturbations at multiple frequencies, about a filtered representation of the large-scale structures that are observed in the flow. The number of frequencies $\omega \in \Omega _b$ affects the accuracy of the linear operator in representing the nonlinear dynamics of the flow. This becomes clear once we observe from (2.13) that perturbations at different temporal frequencies are linearly coupled to one another via the base flow. More precisely, structures at frequency $\omega$ are coupled to structures at frequency $\alpha$ through the base flow at the frequency difference $\omega - \alpha$. Of course, $\omega -\alpha$ needs to be in $\Omega _b$ if one wishes to capture the aforementioned interaction. For instance, if the base flow has frequencies $\Omega _b = \{-\omega,0,\omega \}$ and we want to study the dynamics of perturbations over the set of frequencies $\Omega = \{-2\omega,-\omega,0,\omega,2\omega \}$, then

\[ {\boldsymbol{\mathsf{T}}} = \begin{bmatrix} {\boldsymbol{\mathsf{R}}}^{{-}1}_{{-}2\omega} & {\boldsymbol{\mathsf{G}}}_{-\omega} & 0 & 0 & 0\\ {\boldsymbol{\mathsf{G}}}_\omega & {\boldsymbol{\mathsf{R}}}^{{-}1}_{-\omega} & {\boldsymbol{\mathsf{G}}}_{-\omega} & 0 & 0\\ 0 & {\boldsymbol{\mathsf{G}}}_\omega & {\boldsymbol{\mathsf{R}}}^{{-}1}_{0} & {\boldsymbol{\mathsf{G}}}_{-\omega} & 0\\ 0 & 0 & {\boldsymbol{\mathsf{G}}}_\omega & {\boldsymbol{\mathsf{R}}}^{{-}1}_\omega & {\boldsymbol{\mathsf{G}}}_{-\omega}\\ 0 & 0 & 0 & {\boldsymbol{\mathsf{G}}}_\omega & {\boldsymbol{\mathsf{R}}}^{{-}1}_{2\omega}\\ \end{bmatrix}, \]

where

\begin{gather*} {\boldsymbol{\mathsf{G}}}_\omega {\boldsymbol{q}} ={-}{\boldsymbol{g}}(\hat{\boldsymbol{Q}}_\omega,{\boldsymbol{q}}) - {\boldsymbol{g}}({\boldsymbol{q}},\hat{\boldsymbol{Q}}_\omega),\\ {\boldsymbol{\mathsf{R}}}^{{-}1}_\omega = \textrm{i}\,\omega {\boldsymbol{\mathsf{M}}} - {\boldsymbol{\mathsf{L}}} + {\boldsymbol{\mathsf{G}}}_0. \end{gather*}

Note that ${\boldsymbol{\mathsf{R}}}_\omega$ is the usual resolvent operator at frequency $\omega$ (i.e. the resolvent of the operator linearized about the constant base flow $\hat {\boldsymbol {Q}}_0$). In fact, in the special case that the base flow is constant (i.e. $\Omega _b=\{0\}$), the harmonic resolvent becomes block diagonal, and perturbations at different frequencies are decoupled.

Finally, we wish to make a remark about the resolvent formalism. Whether the base flow is time varying or time invariant, (2.13) is exactly equivalent to the Navier–Stokes equations if $\hat {{\boldsymbol {h}}}'_{\omega }$ is known. This introduces a convenient splitting of the dynamics into linear and nonlinear parts, where the linear part is amenable to classical input–output analyses; this is the point of view taken by McKeon & Sharma (Reference McKeon and Sharma2010), for instance. If $\hat {{\boldsymbol {h}}}'_{\omega }$ is unknown, but can be modelled as random noise (as in McKeon & Sharma (Reference McKeon and Sharma2010)), the linear part evaluated about the temporal mean provides useful insight into the dominant amplification mechanisms. If, on the other hand, $\hat {{\boldsymbol {h}}}'$ has some structure (e.g. large-scale coherent structures at some dominant frequency), then the linearization about the temporal mean may not be useful, if the deviations are large. In the harmonic resolvent framework, we extend the input–output analysis of the linearized Navier–Stokes operator to flows that exhibit such large periodic deviations from the mean.

2.3. Global amplification mechanisms from the harmonic resolvent

The mathematical formulation presented in the previous sections leads to a linear time-periodic input–output system in Fourier space, represented by

(2.16)\begin{equation} \hat{\boldsymbol{q}}' = {\boldsymbol{\mathsf{H}}}\hat{\boldsymbol{w}}', \end{equation}

where the harmonic resolvent ${\boldsymbol{\mathsf{H}}}$ governs the dynamics of perturbations about a periodic base flow $\hat {\boldsymbol {Q}}$ in response to some periodic forcing $\hat {\boldsymbol {w}}'$. There are several ways to view the perturbation $\hat {\boldsymbol {w}}'$. From the point of view of control theory, $\hat {\boldsymbol {w}}'$ can be interpreted as an external input, which might be chosen to achieve some control objective. In a more physics-driven approach, $\hat {\boldsymbol {w}}'$ can be understood as the frequency-domain representation of the nonlinearities that feed back into the linear harmonic resolvent. Alternatively, $\hat {\boldsymbol {w}}'$ can be viewed as an external disturbance that perturbs the system around a known periodic orbit.

In any of these circumstances, one may want to understand the dominant mechanisms by which space–time inputs are amplified through ${\boldsymbol{\mathsf{H}}}$. One way to do so is by seeking a unit-norm space–time input $\hat {\boldsymbol {w}}'$ that leads to the most amplified space–time response $\hat {\boldsymbol {q}}'$. Formally,

(2.17)\begin{equation} \begin{array}{c} \displaystyle\max_{\hat{\boldsymbol{w}}'} \quad \langle{\boldsymbol{\mathsf{H}}}\hat{\boldsymbol{w}}', {\boldsymbol{\mathsf{H}}}\hat{\boldsymbol{w}}'\rangle \\ \text{subject to } \quad \langle \hat{\boldsymbol{w}}',\hat{\boldsymbol{w}}'\rangle = 1. \\ \end{array} \end{equation}

It can be shown that (2.17) leads to the eigenvalue problem

(2.18)\begin{equation} {\boldsymbol{\mathsf{H}}}^{*}{\boldsymbol{\mathsf{H}}}\hat{\boldsymbol{\psi}} = \sigma \hat{\boldsymbol{\psi}}, \end{equation}

where the optimal unit-norm forcing $\hat {\boldsymbol {\psi }}$ is the first right singular vector of ${\boldsymbol{\mathsf{H}}}$, and $\sigma$ is the largest singular value. If we left-multiply (2.18) by ${\boldsymbol{\mathsf{H}}}$ and define $\hat {\boldsymbol {\phi }} = {\boldsymbol{\mathsf{H}}}\hat {\boldsymbol {\psi }}$ we obtain

(2.19)\begin{equation} {\boldsymbol{\mathsf{H}}}{\boldsymbol{\mathsf{H}}}^*\hat{\boldsymbol{\phi}} = \sigma \hat{\boldsymbol{\phi}} , \end{equation}

from which we can conclude that the optimal (most amplified) response, $\hat {\boldsymbol {\phi }}$, is the corresponding left singular vector of ${\boldsymbol{\mathsf{H}}}$. We will refer to the right singular vectors of ${\boldsymbol{\mathsf{H}}}$ as input modes and we will refer to the left singular vectors as output modes.

Proceeding further, the response of the linear time-periodic system to an arbitrary input $\hat {\boldsymbol {w}}'$ can be expressed as a linear combination of input and output modes of the harmonic resolvent, as follows:

(2.20)\begin{equation} \hat{\boldsymbol{q}}' = {\boldsymbol{\mathsf{H}}}\hat{\boldsymbol{w}}'=\sum_{j = 1}^{N-1}\sigma_j\hat{\boldsymbol{\phi}}_j\underbrace{\hat{\boldsymbol{\psi}}_j^*\hat{\boldsymbol{w}}'}_{\langle\hat{\boldsymbol{w}}', \hat{\boldsymbol{\psi}}_j \rangle}, \end{equation}

where $\hat {{\boldsymbol {\phi }}}_j, \hat {{\boldsymbol {\psi }}}_j \in \mathbb {C}^N$. Observe that we sum to $N-1$ since, as per the discussion at the end of § 2.1, we have constrained the range of ${\boldsymbol{\mathsf{H}}}$ to an $(N-1)$-dimensional subspace orthogonal to $\widehat{\textrm{d}{\boldsymbol{Q}}/{\textrm{d}t}}$. Equation (2.20) sheds some light on the information contained within the output and input modes. In particular, the output modes $\hat {\boldsymbol {\phi }}_j$ form an orthonormal basis for the range of ${\boldsymbol{\mathsf{H}}}$ and identify the spatio-temporal structures that are preferentially excited in response to some external input. The input modes $\hat {\boldsymbol {\psi }}_j$ form an orthonormal basis for the domain of ${\boldsymbol{\mathsf{H}}}$ and identify the spatio-temporal structures that are most effective at exciting an energetic response. That is, the input modes relate to the spatio-temporal sensitivity of the flow to external inputs. This concept can be easily understood in terms of the inner product in the underbrace of (2.20). As an example, let us consider a rank-1 approximation of ${\boldsymbol{\mathsf{H}}}$ and assume that $\sigma _1 \approx 1$. If the external input $\hat {\boldsymbol {w}}'$ aligns poorly with the input mode $\hat {\boldsymbol {\psi }}_1$, then $\langle \hat {\boldsymbol {w}}', \hat {\boldsymbol {\psi }}_1 \rangle \ll 1$ and consequently $\|\hat {\boldsymbol {q}}'\|\ll 1$, meaning that $\hat {\boldsymbol {w}}'$ is not effective at exciting a very energetic response through the harmonic resolvent. On the other hand, if the external input aligns well with $\hat {\boldsymbol {\psi }}_1$, then $\langle \hat {\boldsymbol {w}}', \hat {\boldsymbol {\psi }}_1 \rangle \approx \|\hat {\boldsymbol {w}}'\|$. Consequently $\hat {\boldsymbol {q}}' \approx \hat {\boldsymbol {\phi }}_1 \|\hat {\boldsymbol {w}}'\|$ and $\|\hat {\boldsymbol {q}}'\|\approx \|\hat {\boldsymbol {w}}'\|$. In this case $\hat {\boldsymbol {w}}'$ is (close to) optimal and it excites the (close to) optimal most energetic response. Understanding the sensitivity information contained within the input modes is especially important if one is interested in controlling the flow. For instance, if $\hat {\boldsymbol {w}}'$ is a chosen control input, it is advisable to design it in such a way that $\langle \hat {\boldsymbol {w}}',\hat {\boldsymbol {\psi }}_j\rangle$ is maximized.

The singular values $\sigma _j$ can be understood as the gains on the input–output pairs $\hat {\boldsymbol {\psi }}_j$, $\hat {\boldsymbol {\phi }}_j$ and they provide information about the rank of the harmonic resolvent. For instance, if $\sigma _j$ is very small, then the corresponding modes have little effect on the input–output response and can be neglected. Often the effective rank $r$ of ${\boldsymbol{\mathsf{H}}}$ (i.e. the number of singular values that exceed some threshold) is such that $r \ll N-1$, and ${\boldsymbol{\mathsf{H}}}$ may be approximated as

(2.21)\begin{equation} {\boldsymbol{\mathsf{H}}} \approx\sum_{j = 1}^{r}\sigma_j\hat{\boldsymbol{\phi}}_j\hat{\boldsymbol{\psi}}_j^*. \end{equation}

We would like to add that the low rank property of the harmonic resolvent could likely be exploited to develop reduced-order models for the nonlinear dynamics of the flow about the periodic orbit under consideration. For instance the dimension of the nonlinear algebraic system (2.6) in the frequency domain could be reduced by projecting onto the leading input and output modes of the harmonic resolvent. However, investigating this idea further is beyond the scope of the present work.

2.4. Cross-frequency amplification mechanisms from the harmonic resolvent

The global analysis that was carried out in the previous section can be easily extended to selected frequency pairs or selected subsets of $\Omega$. Since the harmonic resolvent accounts for the coupling between different frequencies, we may ask which cross-frequency interactions are most significant. More precisely, for given $\alpha ,\omega \in \Omega$, we can seek the unit-norm forcing at frequency $\alpha$ that triggers the most amplified response at frequency $\omega$. This corresponds to the following optimization problem:

(2.22)\begin{equation} \begin{array}{c} \displaystyle \max_{\hat{\boldsymbol{w}}'_{\alpha}} \quad \langle \hat{\boldsymbol{q}}'_{\omega},\hat{\boldsymbol{q}}'_{\omega} \rangle \\ \text{subject to } \quad \langle \hat{\boldsymbol{w}}'_{\alpha},\hat{\boldsymbol{w}}'_{\alpha} \rangle = 1, \end{array} \end{equation}

where

(2.23)\begin{equation} \hat{\boldsymbol{q}}'_{\omega} = {\boldsymbol{\mathsf{H}}}_{\omega,\alpha}\hat{\boldsymbol{w}}'_{\alpha}, \end{equation}

and where ${\boldsymbol{\mathsf{H}}}_{\omega ,\alpha }$ is the block of ${\boldsymbol{\mathsf{H}}}$ that couples structures at frequency $\omega$ with structures at frequency $\alpha$. It can be shown that the optimal input $\hat {\boldsymbol {w}}'_{\alpha }$ and the optimal output $\hat {\boldsymbol {q}}'_{\alpha }$ are, respectively, the first right singular vector and the first left singular vector of ${\boldsymbol{\mathsf{H}}}_{\omega ,\alpha }$. The corresponding singular value, $\sigma _{\omega ,\alpha }$, is the gain on the $\omega ,\alpha$ cross-frequency pair. For different values of $\omega ,\alpha \in \Omega$, the magnitudes of $\sigma _{\omega ,\alpha }$ provide a measure to determine which cross-frequency couplings are responsible for the development of the structures that are observed in the full nonlinear flow.

3.1. Comparison between the harmonic resolvent framework and resolvent analysis

In this section we compare the effectiveness of different linearizations in predicting the response of the nonlinear system to some external periodic forcing. For this purpose, we introduce forcing to (3.1) with parameters $\mu = \alpha =\beta = 1/5$

(3.5)\begin{equation} \left.\begin{array}{c} \dot x = \mu x - y - \alpha x z - \beta x y + w_1' ,\\ \dot y = x + \mu y - \alpha y z + \beta x^2 + w_2' ,\\ \dot z ={-}\alpha z + \alpha \left(x^2 + y^2 \right) \end{array}\right\} \end{equation}

and define ${\boldsymbol {w}}'(t)=(\varepsilon \sin (\omega t), \varepsilon \cos (\omega t))$ as our external periodic forcing at the fundamental frequency $\omega = 0.9798$ (from (3.4)) and with $\varepsilon = 0.1$. The forcing ${\boldsymbol {w}}'(t)$ can be expanded in a Fourier series

(3.6)\begin{equation} {\boldsymbol{w}}'(t) = \sum_{k \in \{{-}1,1\}} \hat{\boldsymbol{w}}'_{k}{\rm e}^{\textrm{i}\, k\omega t}, \end{equation}

and the dynamics of perturbations ${\boldsymbol {q}}'(t)$ about a given base flow in response to ${\boldsymbol {w}}'(t)$ can be written as

(3.7)\begin{equation} \hat{\boldsymbol{q}}' = {\boldsymbol{\mathsf{H}}}{\boldsymbol{\mathsf{B}}}\hat{\boldsymbol{w}}', \end{equation}

where ${\boldsymbol{\mathsf{H}}}$ is the harmonic resolvent evaluated about the chosen base flow and ${\boldsymbol{\mathsf{B}}}$ is a block-diagonal operator through which the input $\hat {\boldsymbol {w}}'$ enters the system. Throughout this section we consider perturbations with spectral energy content up to the seventh harmonic of the fundamental frequency

(3.8)\begin{equation} {\boldsymbol{q}}'(t) = \sum_{k ={-}7}^{7}\hat{\boldsymbol{q}}_{k}{\rm e}^{\textrm{i}\,k\omega t}, \quad \Omega = \{{-}7\omega,-6\omega, \ldots ,7\omega \}, \end{equation}

and we compare the predictions obtained by linearizing about the temporal mean $\Omega _b = \{0\}$ (see the marker in figure 2b) to predictions obtained by linearizing about the exact limit cycle $\Omega _b = \{-3\omega ,\ldots ,3\omega \}$ (see figure 2a). (Recall that linearizing about the mean is equivalent to performing resolvent analysis.) The results are compared against a ground truth computed by numerical integration of (3.5). In recalling from § 2.1 that the range of the harmonic resolvent is orthogonal to any phase shift about the chosen base flow, we warn the reader that in order to draw a direct comparison between the harmonic resolvent results and the nonlinear simulation, it may be necessary to phase match the forced solution ${\boldsymbol {q}}(t)$ from (3.5) to the base flow ${\boldsymbol {Q}}(t)$. Specifically, we advance (3.5) for a few forcing periods until transients have decayed, and then we phase match in the least squares sense

(3.9)\begin{equation} \min_{\phi \in \mathbb{R}} \quad \sum_{\omega \in \Omega_b} \big\vert \big\vert \hat{{\boldsymbol{q}}}_{\omega}\,\textrm{e}^{\textrm{i}\,\omega \phi } - \hat{{\boldsymbol{Q}}}_{\omega}\big\vert \big\vert_2^2. \end{equation}

Finally, the aforementioned ground truth perturbation is given by ${\boldsymbol {q}}'(t) = {\boldsymbol {q}}(t) - {\boldsymbol {Q}}(t)$.

Figure 3 shows the energy spectrum of the post-transient perturbations $\hat {\boldsymbol {q}}'$ in response to the periodic input $\hat {\boldsymbol {w}}'$ as well as the state evolution ${\boldsymbol {q}}'(t)$ over one fundamental period. We observe from the nonlinear simulation that forcing at frequency $\omega$ leads to a response with energy content also at the zeroth, second and third harmonics. We observe also that the prediction obtained by linearizing about the exact limit cycle accurately matches the ground truth. This is because the time-varying base flow about which we evaluate the harmonic resolvent couples structures at different frequencies and we are therefore able to predict (to first order) the frequency off-scatter that is observed in the nonlinear system. The extent to which we are able to capture cross-frequency interactions is given by the block singular values of ${\boldsymbol{\mathsf{H}}}$ shown in figure 4(b). We colour code the cross-frequency blocks according to the fractional variance

(3.10)\begin{equation} E_{j,k} = \frac{\displaystyle\sum_{m = 1}^3 \sigma^2_{m,\left(\,j,k\right)}}{\displaystyle\sum_{n = 1}^{N-1}\sigma_n^2}, \end{equation}

where $\sigma _{m,(\,j,k)}$ is the $m$th singular value of the $(\,j\omega ,k\omega )$ block of ${\boldsymbol{\mathsf{H}}}$ and $\sigma _n$ is the $n$th singular value of ${\boldsymbol{\mathsf{H}}}$. The normalization is such that $\sum _{j,k} E_{j,k} = 1$. We observe from the $1\omega$-column in figure 4(b) that forcing at the fundamental frequency may trigger a response with spectral energy content up to the third harmonic, and that is precisely what the energy spectrum in figure 3(a) confirms.

Figure 3.

Response of the system (3.5) to periodic forcing at the fundamental frequency $\omega = 0.9798$. (a) Normalized spectrum of the perturbation $\boldsymbol {q'}(t)$ and (b) state evolution over one period $T = 2{\rm \pi} /\omega$. Symbols: $\ast$, Ground truth computed by numerical integration of (3.5); $\square$, harmonic resolvent prediction with $\Omega _b = \{-3\omega ,\ldots ,3\omega \}$; $\diamond$, harmonic resolvent prediction with $\Omega _b = \{0\}$ (equivalent to resolvent analysis).

Figure 4.

Fractional variance (3.10) contained within each block of the harmonic resolvent for the toy model (3.1) (a) evaluated about the temporal mean $\Omega _b = \{0\}$ and (b) evaluated about the exact periodic solution $\Omega _b = \{k\omega \}_{k\in \{-3,\ldots ,3\}}$.

Linearizing about the temporal mean, however, does not provide an accurate representation of the response of the nonlinear system to the given periodic forcing at frequency $\omega$. First, observe that through this linearization we overestimate the spectral energy at the fundamental frequency. Second, it is clear from figures 3(a) and 3(b) that the prediction is monochromatic at frequency $\omega$. This is because ${\boldsymbol{\mathsf{H}}}$ is block diagonal, as mentioned at the end of § 2.2 and illustrated in figure 4(a). Therefore, no cross-frequency interaction can be accounted for through the base flow, and forcing at frequency $\omega$ will only produce a response at the same frequency.

4.1. Amplification mechanisms about a time-varying base flow

We linearize the dynamics in (4.1) about a time-periodic base flow over the set of frequencies $\Omega _b =\{-3\omega ,\ldots ,3\omega \}$ with $\omega = 2.40$ as in figure 5(a), and we compute the singular value decomposition of the harmonic resolvent.

First, note from figure 6(a) that there is more than an order of magnitude separation between the first and the second singular values of the harmonic resolvent and we can therefore argue that the harmonic resolvent has low-rank structure.

Figure 6.

Singular values of the harmonic resolvent for flow past an airfoil with $\Omega _b = \{-3\omega ,\ldots ,3\omega \}$, $\Omega = \{-7\omega ,\ldots ,7\omega \}$ and $\omega = 2.40$, showing (a) singular values of ${\boldsymbol{\mathsf{H}}}$ and (b) blockwise fractional variance $E_{j,k}$ defined by an expression similar to (3.10).

Second, figure 6(b) shows that the nonlinear flow is very susceptible to perturbations at the fundamental frequency, since the block singular values of ${\boldsymbol{\mathsf{H}}}$ suggest that introducing forcing at $\omega$ has an effect on flow structures up to the fourth harmonic. Likewise, we can conclude that the flow is less sensitive to perturbations at higher harmonics of the fundamental frequency as we observe that the singular values of the blocks governing those dynamics are one (or more) orders of magnitude less than those in the $1\omega$-column.

We may also draw conclusions about the sensitivity of the flow from the $k\omega$-entries of the first input mode of ${\boldsymbol{\mathsf{H}}}$, shown in figure 7. Recall from the previous sections that the input modes describe the spatio-temporal structures that are most effective at exciting a response, while the output modes describe the spatio-temporal structures that are preferentially excited by these inputs. Specifically, we learn from the magnitude of the entries of the input mode that the flow is most sensitive to perturbations at frequency $\omega$ and it is the least sensitive to perturbations at frequency $3\omega$. Moreover, the mode shapes of the entries of the input mode suggest that the flow is very sensitive to perturbations that are spatially localized around the body, while the output mode entries in figure 7 illustrate the spatial structures that should arise in the flow in response to a disturbance or control input that aligns well with the input mode. In order to verify this statement, we introduce a small amplitude forcing in the flow by sinusoidally moving the airfoil in the vertical direction, with velocity

(4.3)\begin{equation} v = \varepsilon\omega\cos\omega t. \end{equation}

The forcing frequency is taken to be the fundamental frequency of vortex shedding, $\omega = 2.40$, while $\varepsilon = 0.01$.

Figure 7.

Real part of the vorticity field computed from the input mode and the output mode associated with $\sigma _1$ in figure 6(a).

As per the discussion in § 3.1, in order to draw a direct comparison with the output modes of the harmonic resolvent operator we proceed as follows. We let the flow reach the limit cycle ${\boldsymbol {Q}}(t)$ whose spectrum is shown in figure 5(a) and then we introduce the sinusoidal motion described in (4.3). We let the flow evolve for a few periods until transients have decayed and then we phase match the forced solution ${\boldsymbol {q}}(t)$ to the base flow ${\boldsymbol {Q}}(t)$ according to (3.9). Finally, the perturbation introduced by the sinusoidal motion is ${\boldsymbol {q}}'(t) = {\boldsymbol {q}}(t) - {\boldsymbol {Q}}(t)$.

Figure 8 shows the vorticity of the first few Fourier modes computed from ${\boldsymbol {q}}'(t)$. These highlight the vortical structures that result when the flow is forced sinusoidally according to (4.3). Remarkably, the first output mode of the harmonic resolvent, shown in figure 7, provides a surprisingly accurate prediction of the structures, at all four frequencies shown in figure 8. This close agreement is presumably a consequence of the low-rank structure of the harmonic resolvent (see figure 6): regardless of the type of forcing, the resulting flow perturbations will resemble the output modes shown in figure 7. The fact that the input modes shown in figure 7 are supported near the airfoil suggests that the flow is sensitive to perturbations near the airfoil. While this result is not surprising, the simulations with sinusoidal motion of the airfoil confirm this behaviour.

Figure 8.

Real part of the Fourier modes of the vorticity perturbations for the airfoil with sinusoidal motion (4.3) for frequencies $0{-}3\omega$.

4.2. Amplification mechanisms about the temporal mean

We now evaluate the harmonic resolvent about the temporal mean. We recall that in this case, with $\Omega _b = \{0\}$, the harmonic resolvent becomes block diagonal, where each diagonal block is the usual resolvent operator at its corresponding frequency. Once again, we consider perturbations over the set of frequencies $\Omega = \{-7\omega ,\ldots ,7\omega \}$. The first five singular values of the harmonic resolvent ${\boldsymbol{\mathsf{H}}}$ are shown in figure 9(a), which shows that the effective rank of ${\boldsymbol{\mathsf{H}}}$ is 2. Here, $\sigma _1 = \sigma _2 \approx 6\times 10^3$, where $\sigma _1$ is the singular value associated with the block at frequency $\omega$, and $\sigma _2$ corresponds to the block at frequency $-\omega$. In figure 10 we show the first input and output pairs computed from the singular value decompositions of the resolvent operators evaluated at the frequencies of interest.

Figure 9.

Singular values for the harmonic resolvent of flow past an airfoil, with $\Omega _b = \{0\}$, $\Omega = \{-7\omega ,\ldots ,7\omega \}$ and $\omega = 2.40$, showing (a) singular values of ${\boldsymbol{\mathsf{H}}}$ and (b) blockwise fractional variance $E_{j,k}$ defined by an expression similar to (3.10).

Figure 10.

Real part of the vorticity field computed from the first input and output pairs of the resolvent operators at the given frequencies. The corresponding leading singular values are $\sigma \approx 10$, $\sigma \approx 6\times 10^3$ and $\sigma \approx 5$ for frequencies $0$, $\omega$ and $2\omega$, respectively.

It appears that meaningful information is obtained only at the fundamental frequency $\omega$ in figure 10. Specifically, the $\omega$-component of the input mode provides accurate information on the sensitivity of the flow, since it is analogous to the $\omega$-component of the input mode of the harmonic resolvent operator evaluated about the time-varying base flow in figure 7(c). The corresponding output mode in figure 10, however, does not capture the qualitative behaviour that has been observed in the forced nonlinear simulation, shown in figure 8.

Finally, it appears that no meaningful information is provided by the resolvent operators at the zero frequency or at higher harmonics of the fundamental frequency $\omega$. The reason behind this can be understood by looking at the diagonal blocks of figures 9(b) and 6(b). Both, in fact, suggest that the temporal mean (block-diagonal entries of the harmonic resolvent) does not amplify disturbance at higher frequencies, meaning that we cannot expect the resolvent operators at frequencies $k\omega$ with $k\neq 1$ to provide any meaningful information about the flow structures at those frequencies. Furthermore, figure 6 suggests that the presence of higher harmonics is exclusively due to perturbations at the fundamental frequency $\omega$ that scatter off the base flow to excite a response at higher harmonics. It appears that nonlinear mechanisms dominate the dynamics of this flow, and it is therefore necessary to perform a linearization about a time-varying base flow in order to study the amplification mechanisms.

The brief quantitative analysis that was conducted in this section attempts to show how the information contained within the harmonic resolvent can be used to draw conclusions about the physics of the flow. In turn, the understanding of the flow physics can be leveraged for control purposes. For instance, the sensitivity of the flow with respect to some external forcing at different frequencies is of utmost importance in actuator design and placement. Moreover, the output modes at different frequencies provide insight into the first-order response of the flow to the control input. In turbulence applications and in the problem of reattachment of separated flows (Deem et al. Reference Deem, Cattafesta, Yao, Hemati, Zhang and Rowley2018), for example, one may be interested in understanding how the zero frequency (temporal mean) is affected by a harmonic input. Finally, the cross-frequency information contained within the harmonic resolvent (see figure 6b, for example) can provide useful insight into the energy transfer mechanisms between different time scales in turbulent flows.