2.1. Force estimation

Defining the instantaneous error, $\boldsymbol{e}_f(t)$, between the force, $\boldsymbol {f}(t)$, and its estimation, $\tilde {\boldsymbol {f}}(t)$, as

(2.4)\begin{equation} \boldsymbol{e}_f(t) = \boldsymbol{f}(t)-\tilde{\boldsymbol{f}}(t), \end{equation}

we seek an optimal estimation such that a stationary point of the time-averaged error correlation matrix

(2.5)\begin{equation} \langle\boldsymbol{e}_f \boldsymbol{e}_f^{\dagger}\rangle=\int_{-\infty}^{\infty} \langle \boldsymbol{e}_f(t) \boldsymbol{e}_f^{\dagger}(t)\rangle\, \textrm{d}t = \dfrac{1}{2\pi}\int_{-\infty}^{\infty} \langle\hat{\boldsymbol{e}}_f(\omega)\hat{\boldsymbol{e}}_f^{\dagger} (\omega)\rangle\,\textrm{d}\omega \end{equation}

is found. Note that this is a generalization of root-mean-square errors, which is given by the trace of $\langle \boldsymbol {e}_f \boldsymbol {e}_f^{\dagger }\rangle$. The integration limits reflect the estimation objective: to estimate $\boldsymbol {f}(t)$ for all times. Here $\langle \cdot \rangle$ represents an ensemble average, $\hat{\boldsymbol{e}}_{f}(\omega)$ is the Fourier transform of $\boldsymbol{e}_{f}(t)$, defined by

(2.6)\begin{equation} \hat{\boldsymbol{e}}_f (\omega) = \int_{-\infty}^{\infty} \boldsymbol{e}_f(t) \mathrm{e}^{ {\mathrm{i}} \omega t}\,\textrm{d}t, \end{equation}

and the equivalence of the time and frequency domains in (2.5) is given by Parseval's theorem (Arfken, Weber & Harris Reference Arfken, Weber and Harris2013, p. 595). No assumption is made regarding the underlying probability functions of forces, responses and errors. As discussed by Kalman (Reference Kalman1960), this provides the optimal estimation if forces are Gaussian distributed. For other force distributions, optimal estimation is nonlinear, the method above providing the optimal linear estimation. Note that this is analogous to a least-squares method for fitting curves: if errors have a Gaussian distribution, the method is equivalent to a maximum-likelihood method; however, the method is still effective for other distributions.

Formal solutions of (2.1) in time and frequency domains are obtained as

(2.7a,b)\begin{equation}\boldsymbol{u}(t) = \int_{-\infty}^{t} e^{\boldsymbol{\mathsf{A}} (t-\tau)}\boldsymbol{\mathsf{B}}\, \boldsymbol{f}(\tau)\,\textrm{d}\tau,\quad \hat{\boldsymbol{u}} (\omega) = \boldsymbol{\mathsf{R}} (\omega) \boldsymbol{\mathsf{B}}\,\hat{f}(\omega), \end{equation}

where $\boldsymbol{\mathsf{R}} = (-\boldsymbol{\mathsf{A}} -\mathrm {i}\omega \boldsymbol{\mathsf{I}})^{-1}$. We assume a time evolution from $-\infty$ to $\infty$, where an initial condition, $\boldsymbol {u}_0$, at time, $t_0$, can be represented via a forcing, $\boldsymbol{\mathsf{B}} \,\boldsymbol {f}(t)=\boldsymbol {u}_0\delta (t-t_0)$.

Considering the state $\tilde {\boldsymbol {u}}$ obtained by integration of an estimated force $\tilde {\boldsymbol {f}}$,

(2.8a,b)\begin{equation} \tilde{\boldsymbol{u}}(t) = \int_{-\infty}^{t} e^{\boldsymbol{\mathsf{A}} (t-\tau)}\boldsymbol{\mathsf{B}} \,\tilde{\boldsymbol{f}}(\tau)\, \textrm{d}\tau,\quad \hat{\tilde{\boldsymbol{u}}} (\omega) = \boldsymbol{\mathsf{R}} (\omega) \boldsymbol{\mathsf{B}}\, \hat{\tilde{\boldsymbol{f}}}(\omega), \end{equation}

we seek to obtain $\tilde {\boldsymbol {f}}$ as a linear function of the readings

(2.9a,b)\begin{equation}\tilde{\boldsymbol{f}}(t) = \int_{-\infty}^{\infty} \boldsymbol{\mathsf{T}}_{\boldsymbol{f}} (t-\tau) \boldsymbol{y}(\tau)\, \textrm{d}\tau,\quad \hat{\tilde{\boldsymbol{f}}}(\omega) = \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} (\omega) \hat{\boldsymbol{y}}(\omega), \end{equation}

such that the estimation problem involves finding an estimation function, $\skew3\hat {\boldsymbol{\mathsf{T}}}_{\boldsymbol {f}}(\omega )$, that would minimise the norm of $\langle \hat {\boldsymbol {e}}_f \hat {\boldsymbol {e}}_f^{\dagger } \rangle$. Note that the estimation function in (2.9) is generally non-causal.

The error correlation matrix is rewritten, with frequency dependencies omitted for clarity, as

(2.10)\begin{align} \left\langle\hat{\boldsymbol{e}}_f \hat{\boldsymbol{e}}_f^{\dagger}\right\rangle & = \left\langle \left(\skew2\hat{\boldsymbol{f}} - \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \left(\boldsymbol{\mathsf{R}}_{\boldsymbol{y}}\skew2\hat{\boldsymbol{f}}+\hat{\boldsymbol{n}} \right)\right) \left(\skew2\hat{\boldsymbol{f}} - \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \left(\boldsymbol{\mathsf{R}}_{\boldsymbol{y}}\, \skew2\hat{\boldsymbol{f}}+\hat{\boldsymbol{n}} \right) \right)^{\dagger} \right\rangle, \nonumber\\ &= \left(\boldsymbol{\mathsf{I}} -\skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \right) \skew4\hat{\boldsymbol{\mathsf{F}}} \left(\boldsymbol{\mathsf{I}} -\skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \right)^{\dagger} + \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \skew3\hat{\boldsymbol{\mathsf{N}}} \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}}^{\,\dagger}. \end{align}

The operator $\boldsymbol{\mathsf{R}}_{\boldsymbol {y}} = \boldsymbol{\mathsf{C}}\boldsymbol{\mathsf{R}} \boldsymbol{\mathsf{B}}$ relates forces to sensor readings, and the matrices $\skew4\hat {\boldsymbol{\mathsf{F}}}(\omega )$ and $\skew4\hat {\boldsymbol{\mathsf{N}}}(\omega )$ are Fourier transforms of $\boldsymbol{\mathsf{F}}(t)$ and $\boldsymbol{\mathsf{N}}(t)$, defined in (2.3). Typically, the forcing rank is much larger than the number of sensors, $n_y \ll n_b$, and, thus, it is natural that the forcing cannot be fully estimated. The estimation is limited to the subspace of observable forces (Towne etal. Reference Towne, Lozano-Durán and Yang2020). By definition, any force that generates a non-zero sensor reading has a non-zero projection in this subspace.

A connection can be made with the observability Gramian (Bagheri etal. Reference Bagheri, Henningson, Hœpffner and Schmid2009), which is obtained through a time integral

(2.11)\begin{equation} \mathcal{G} = \int_{0}^{\infty} \mathrm{e}^{\boldsymbol{\mathsf{A}}^{\dagger}\tau}\boldsymbol{\mathsf{C}}^{\,\dagger} \boldsymbol{\mathsf{C}} \mathrm{e}^{\boldsymbol{\mathsf{A}}\tau}\,\textrm{d}\tau. \end{equation}

Alternatively, the observability Gramian can be written using Parseval's theorem (Zhou, Salomon & Wu Reference Zhou, Salomon and Wu1999; Dergham etal. Reference Dergham, Sipp, Robinet and Barbagallo2011)

(2.12)\begin{equation} \mathcal{G} = \frac{1}{2\pi} \int_{-\infty}^{\infty} \boldsymbol{\mathsf{R}}^{\,\dagger}\boldsymbol{\mathsf{C}}^{\,\dagger} \boldsymbol{\mathsf{C}} \boldsymbol{\mathsf{R}}\,\textrm{d}\omega. \end{equation}

At a given frequency, the observable space is spanned by the eigenvectors of $\boldsymbol{\mathsf{R}}^{\,\dagger }\boldsymbol{\mathsf{C}}^{\,\dagger } \boldsymbol{\mathsf{C}} \boldsymbol{\mathsf{R}}$ associated with non-zero eigenvalues. When viewed in the time domain through the observability Gramian, each force component is weighted by its overall observability at all frequencies. A low-rank truncation of this Gramian can be used to obtain a reduced basis for the construction of time-domain ROMs, implicitly favouring some frequencies over others. A similar reasoning applies to the construction of balanced modes (Bagheri etal. Reference Bagheri, Henningson, Hœpffner and Schmid2009). In the approach presented here, we use the full observable space at each frequency, without the aforementioned truncation.

As $\langle \hat{\boldsymbol{e}}_f \hat{\boldsymbol {e}}_f^{\dagger}\rangle$ is a matrix, it is typically necessary to specify a minimisation criterion: minimisation of the trace, for instance, as is done in the design of Kalman filters, or of the determinant. We will show, however, that a stationary point can be obtained simultaneously for all matrix terms, illustrating a certain robustness of the method. We impose ${\textrm {d} {\langle \hat{\boldsymbol {e}}_f \hat {\boldsymbol {e}}_f^{\dagger }\rangle }}/{\textrm {d} {\skew4\hat {\boldsymbol{\mathsf{T}}}}_{\boldsymbol {f}}^{\,\dagger }}= 0$ and ${\textrm {d} {\langle \hat {\boldsymbol {e}}_f \hat {\boldsymbol {e}}_f^{\dagger }\rangle }}/{\textrm {d} {\skew4\hat {\boldsymbol{\mathsf{T}}}_{\boldsymbol {f}}}}= 0$, where ${\skew4\hat {\boldsymbol{\mathsf{T}}}}_{\boldsymbol {f}}^{\,\dagger }$ and ${\skew4\hat {\boldsymbol{\mathsf{T}}}_{\boldsymbol {f}}}$ are treated as independent variables (Ahlfors Reference Ahlfors1979, p. 79). It can be shown that both expressions lead to the same equation, we thus focus only on the first. As both $\langle \hat {\boldsymbol {e}}_f \hat {\boldsymbol {e}}_f^{\dagger }\rangle$ and ${\skew4\hat {\boldsymbol{\mathsf{T}}}}_{\boldsymbol {f}}^{\,\dagger }$ are matrices, the derivative is a fourth-order tensor, it is thus simpler to take the derivative using Einstein's summation convention. Equation (2.10) has the form

(2.13)\begin{equation} \left\langle\hat{\boldsymbol{e}}_f \hat{\boldsymbol{e}}_f^{\dagger}\right\rangle_{il} = \Gamma_{ij} {T_f}^{\dagger}_{jk} + \Lambda_{ik}, \end{equation}

with

(2.14)\begin{gather} \boldsymbol{\Gamma} = \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \left( \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \skew4\hat{\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger} + \skew3\hat{\boldsymbol{\mathsf{N}}} \right) - \skew4\hat{\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger}, \end{gather}

(2.15)\begin{gather} \boldsymbol{\Lambda} = \left(\boldsymbol{\mathsf{I}}-\skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}}\right) \skew4\hat{\boldsymbol{\mathsf{F}}}. \end{gather}

The derivative is given by

(2.16)\begin{equation} \dfrac{\textrm{d} \langle\hat{\boldsymbol{e}}_f \hat{\boldsymbol{e}}_f^{\dagger}\rangle_{il}}{{\textrm{d} {\hat{T}_f}^{\dagger}}_{mn}} = \Gamma_{ij} \delta_{jm}\delta_{kn} = \Gamma_{im} \delta_{kn}, \end{equation}

where $\delta _{ij}$ is the Kronecker delta. The expression is a tensor product between two matrices, and it is zero only if one of these is zero. As only $\boldsymbol{\Gamma}$ is a function of ${\boldsymbol{\mathsf{T}}}_{\boldsymbol {f}}$, the stationary point is found for

(2.17)\begin{equation} \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \left( \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \skew4\hat{\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger} + \skew2\hat{\boldsymbol{\mathsf{N}}} \right) = \skew4\hat{\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger}. \end{equation}

As $\boldsymbol{\mathsf{R}}_{\boldsymbol {y}} \skew3\hat {\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{R}}_{\boldsymbol {y}} ^{\,\dagger }$ is semi-positive definite, $\boldsymbol{\mathsf{R}}_{\boldsymbol {y}} \skew3\hat {\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{R}}_{\boldsymbol {y}} ^{\,\dagger } + \skew3\hat {\boldsymbol{\mathsf{N}}}$ is always invertible, and, thus,

(2.18)\begin{equation} \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} = \skew4\hat{\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger} \left( \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \skew4\hat{\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger} + \skew2\hat{\boldsymbol{\mathsf{N}}} \right)^{-1}. \end{equation}

Note that sensor noise has a similar role of Tikhonov regularization parameter, but instead of being a regularization which is imposed on the system, it arises naturally from the system considered.

2.2. Response estimation

Here we follow a procedure similar to that developed in the previous subsection, but with the objective of obtaining an optimal estimation of the system response. Defining the error as

(2.19)\begin{equation} \hat{\boldsymbol{e}}_u = \hat{\boldsymbol{u}} - \hat{\tilde{\boldsymbol{u}}} = \boldsymbol{\mathsf{R}} \boldsymbol{\mathsf{B}} \left(\skew3\hat{\boldsymbol{f}} - \hat{\tilde{\boldsymbol{f}}} \right), \end{equation}

the error cross-correlation is given by

(2.20)\begin{align} \left\langle \hat{\boldsymbol{e}}_u \hat{\boldsymbol{e}}_u^{\dagger}\right\rangle &= \left\langle \boldsymbol{\mathsf{R}} \boldsymbol{\mathsf{B}} \left(\skew2\hat{\boldsymbol{f}} - \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \left(\boldsymbol{\mathsf{R}}_{\boldsymbol{y}}\,\skew2\hat{\boldsymbol{f}}+\hat{\boldsymbol{n}} \right)\right) \left(\skew2\hat{\boldsymbol{f}} - \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \left(\boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \,\skew2\hat{\boldsymbol{f}}+\hat{\boldsymbol{n}} \right) \right)^{\dagger} \boldsymbol{\mathsf{B}}^{\dagger} \boldsymbol{\mathsf{R}} ^{\dagger} \right\rangle\nonumber\\ & = \boldsymbol{\mathsf{R}} \boldsymbol{\mathsf{B}} \left( \left(\boldsymbol{\mathsf{I}} - \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \right) \skew4\hat{\boldsymbol{\mathsf{F}}} \left(\boldsymbol{\mathsf{I}}- \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \right)^{\dagger} + {\boldsymbol{\mathsf{T}}_{\boldsymbol{f}}} \skew2\hat{\boldsymbol{\mathsf{N}}} {\boldsymbol{\mathsf{T}}_{\boldsymbol{f}}}^{\dagger} \right) \boldsymbol{\mathsf{B}}^{\dagger} \boldsymbol{\mathsf{R}}^{\,\dagger}. \end{align}

The stationary point is given by

(2.21)\begin{equation} \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} = \skew4\hat{\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger} \left( \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \skew4\hat{\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger} + \skew2\hat{\boldsymbol{\mathsf{N}}} \right)^{-1}, \end{equation}

which is the same as that obtained for optimal force estimation (2.18).

The equivalence between optimal force and response estimation motivates use of the same nomenclature for $\skew4\hat {\boldsymbol{\mathsf{T}}}_{\boldsymbol {f}}$. This equivalence between force and response estimation is expected in the Kalman framework: in optimal estimation, only components correlated with the sensor readings are estimated. As responses are correlated with their driving forces, estimation of the former is synonymous with estimation of the latter.

State estimation is thus obtained using (2.8) and (2.9) as

(2.22)\begin{equation} \hat{\tilde{\boldsymbol{u}}} = \boldsymbol{\mathsf{R}} \boldsymbol{\mathsf{B}} \boldsymbol{\mathsf{T}}_{\boldsymbol{f}} \hat{\boldsymbol{y}} = \boldsymbol{\mathsf{T}}_{\boldsymbol{u}} \hat{\boldsymbol{y}}. \end{equation}

The expressions (2.18) and (2.22) are related to Wiener filter estimation, commonly used in its scalar version (Meditch Reference Meditch1973), but also defined in vector form (Martinelli Reference Martinelli2009). The Wiener filter is given by a transfer function obtained from cross-spectral densities as $\boldsymbol{\mathsf{S}}_{uy}\boldsymbol{\mathsf{S}}^{-1}_{yy}$, where $\boldsymbol{\mathsf{S}}_{uy}$ is the cross-spectrum between state and measurement and $\boldsymbol{\mathsf{S}}_{yy}$ is the measurement cross-spectral density (CSD). In the method proposed here, CSDs are computed a priori, assumptions being made regarding the system model and the forces. For force and sensor noise CSDs respectively given by $\skew3\hat {\boldsymbol{\mathsf{F}}}$ and $\skew3\hat {\boldsymbol{\mathsf{N}}}$, it is straightforward to show that $\boldsymbol{\mathsf{S}}_{uy} = \boldsymbol{\mathsf{F}} \boldsymbol{\mathsf{R}}_{\boldsymbol {y}}^{\dagger }$ and $\boldsymbol{\mathsf{S}}_{yy} = \boldsymbol{\mathsf{R}}_{\boldsymbol {y}}\skew3\hat {\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{R}}_{\boldsymbol {y}}^{\dagger } +\skew3\hat {\boldsymbol{\mathsf{N}}}$.

As previously mentioned, classical derivations minimise the trace of the error CSD. This suggests that optimal estimation may involve trading accuracy in one region in favour of another, so as to obtain a global minimum. If this were the case, one could localise the region where estimation is desired in order to improve it. Our derivation shows that estimation is optimal everywhere.

2.4. Matrix-free approach

The simplest method for obtaining the resolvent-based estimator described above involves matrix inversion, followed by direct application of (2.18) and (2.22). The approach thus becomes prohibitively expensive for large matrices. Matrix inversion can be avoided by solution of the linear system

(2.31)\begin{equation} ( -\mathrm{i} \omega \boldsymbol{\mathsf{I}} - \boldsymbol{\mathsf{A}}) \hat{\boldsymbol{u}} = \boldsymbol{\mathsf{B}} \skew2\hat{\boldsymbol{f}}, \end{equation}

as done by Schmidt etal. (Reference Schmidt, Towne, Rigas, Colonius and Brès2018). Although less demanding, computational cost typically limits this approach to two-dimensional problems. Time-marching schemes have been used by Tam & Pastouchenko (Reference Tam and Pastouchenko2002), in which individual rows of $\boldsymbol{\mathsf{R}}_{\boldsymbol {y}}$ are obtained via integration of harmonically forced adjoint equations; the approach is applicable for much larger systems. Repeating the procedure for different frequencies and sensors, an estimation based on (2.9), (2.21) and (2.22) can be obtained. Here we propose a method that significantly further reduces computational cost, providing solutions for all frequencies with a single time-march. This is achieved by integration of the direct and adjoint equations, similar to what is done in other matrix-free approaches (Semeraro etal. Reference Semeraro, Pralits, Rowley and Henningson2013).

For simplicity, we assume that $\boldsymbol{\mathsf{F}}(t)=\boldsymbol{\mathsf{I}}\delta (t)$, $\boldsymbol{\mathsf{B}}=\boldsymbol{\mathsf{I}}$ and $\boldsymbol{\mathsf{N}}(t)=\epsilon \boldsymbol{\mathsf{I}}\delta (t)$.

Consider the system

(2.32)\begin{equation} -\dfrac{\textrm{d}\boldsymbol{w}_i}{\textrm{d}t} = \boldsymbol{\mathsf{A}}^{\dagger} \boldsymbol{w}_i + \boldsymbol{\mathsf{C}}_i^{\dagger} \delta(t), \end{equation}

with a null terminal condition, and where $\boldsymbol{\mathsf{A}}^{\dagger }$ corresponds to the adjoint linearised Navier–Stokes operator and $\boldsymbol{\mathsf{C}}_i^{\dagger }$ is the $i$th column of the adjoint of $\boldsymbol{\mathsf{C}}$. The impulse response of (2.32) can be replaced by a terminal condition, $\boldsymbol {w}_i(0) = \boldsymbol{\mathsf{C}}_i^{\dagger }$. Taking the Fourier transform of (2.32) leads to

(2.33a)\begin{equation} ({\rm i}\omega\boldsymbol{\mathsf{I}} - \boldsymbol{\mathsf{A}}^{\dagger} ) \hat{\boldsymbol{w}}_i = \boldsymbol{\mathsf{C}}_i^{\dagger} \end{equation}

and, thus,

(2.33b)\begin{equation} \hat{\boldsymbol{w}}_i = \boldsymbol{\mathsf{R}}^{\,\dagger} \boldsymbol{\mathsf{C}}_i^{\dagger}. \end{equation}

The $i$th component of $\boldsymbol{\mathsf{R}}_{\boldsymbol {y}}^{\dagger }$, given by $\boldsymbol{\mathsf{R}}^{\,\dagger }\boldsymbol{\mathsf{C}}_i^{\dagger }$, is the Fourier transform of the response of (2.32). The resolvent can thus be constructed row-by-row as

(2.34)\begin{equation} \hat{\boldsymbol{\mathsf{R}}_{\boldsymbol{y}}} (\omega) = \left[ \begin{matrix} - & \hat{\boldsymbol{w}}_1^{\dagger}(\omega) & - \\ - & \hat{\boldsymbol{w}}_2^{\dagger}(\omega) & - \\ & \vdots & \end{matrix} \right]. \end{equation}

The values of $\boldsymbol {w}_i(-t)$ are the sensitivities of the $i$th measurement at time $t=0$ to forces at the instant $-t$; in the frequency domain the same information is contained in $\hat {\boldsymbol {w}}_i(\omega )$.

From solutions of the equation

(2.35)\begin{equation} \dfrac{\textrm{d}\boldsymbol{q}_i}{\textrm{d}t} = \boldsymbol{\mathsf{A}} \boldsymbol{q}_i + \boldsymbol{w}_i(t), \end{equation}

which has frequency domain representation

(2.36a)\begin{equation} (-\textrm{i}\omega\boldsymbol{\mathsf{I}} - \boldsymbol{\mathsf{A}}) \hat{\boldsymbol{q}}_i = \hat{\boldsymbol{w}}_i \end{equation}

and, thus,

(2.36b)\begin{equation} \hat{\boldsymbol{q}}_i = \boldsymbol{\mathsf{R}} \hat{\boldsymbol{w}}_i, \end{equation}

the operator $\boldsymbol{\mathsf{R}}\boldsymbol{\mathsf{R}}_{\boldsymbol {y}}^{\dagger }$ can be constructed from Fourier transforms of the solutions as

(2.37)\begin{equation} \boldsymbol{\mathsf{R}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}}^{\dagger} (\omega) = \boldsymbol{\mathsf{R}} \left[ \begin{matrix} | & |& \\ \hat{\boldsymbol{w}}_1(\omega) & \hat{\boldsymbol{w}}_2(\omega) & \cdots \\ | & |& \end{matrix} \right] = \left[ \begin{matrix} | & |& \\ \hat{\boldsymbol{q}}_1(\omega) & \hat{\boldsymbol{q}}_2(\omega) & \cdots \\ | & |& \end{matrix} \right]. \end{equation}

As

(2.38)\begin{equation} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}}^{\dagger}= \boldsymbol{\mathsf{C}}\boldsymbol{\mathsf{R}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}}^{\dagger}, \end{equation}

this term can be obtained from observations of the system $\hat {\boldsymbol {y}}_i = \boldsymbol{\mathsf{C}} \hat {\boldsymbol {q}}_i$. State and force estimation functions,

(2.39a,b)\begin{equation} \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{u}} = \boldsymbol{\mathsf{R}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger} \left( \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger} + \epsilon \boldsymbol{\mathsf{I}} \right)^{-1}, \quad \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}} = \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger} \left( \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}} ^{\,\dagger} + \epsilon \boldsymbol{\mathsf{I}} \right)^{-1}, \end{equation}

can then be constructed.

The procedure can be summarised in the following steps.

1. (i) Adjoint run: (2.32) is integrated for each sensor. Snapshots are saved on disk for later use.

2. (ii) Direct run: (2.35) is integrated for each sensor, with snapshots from the adjoint run loaded and interpolated at each time step, to provide the force term. The readings at the sensors and some other points of interest are calculated at each time step and saved separately.

Although the term $\boldsymbol{\mathsf{R}}_{\boldsymbol {y}}\boldsymbol{\mathsf{R}}_{\boldsymbol {y}}^{\dagger }$ from (2.37) and (2.38) can be computed from flow snapshots of the direct run, saving sensor readings at each simulation time step is computationally cheap, and provides extra accuracy with negligible extra cost. Generalisation for any $\boldsymbol{\mathsf{B}}$ and $\boldsymbol{\mathsf{F}}$ is obtained by multiplying $\hat {\boldsymbol {w}}_i$ by $\boldsymbol{\mathsf{B}} \skew3\hat {\boldsymbol{\mathsf{F}}} \boldsymbol{\mathsf{B}}^{\dagger }$ prior to the integration of (2.35).

3.1. Model problem

We compare resolvent-based estimation (non-causal) and Kalman-filter estimation (causal) on a complex-valued linearised Ginzburg–Landau (GL) model, which is frequently used as a simple model that qualitatively mimics the behaviour of complex flows (Chomaz, Huerre & Redekopp Reference Chomaz, Huerre and Redekopp1991; Couairon & Chomaz Reference Couairon and Chomaz1999; Bagheri etal. Reference Bagheri, Henningson, Hœpffner and Schmid2009; Cavalieri etal. Reference Cavalieri, Jordan and Lesshafft2019; Towne etal. Reference Towne, Lozano-Durán and Yang2020), and, thus, constitutes a convenient benchmark. The comparison provides insights on advantages of using non-causal estimation tools when the necessary information for such is available.

The model takes the form

(3.1a,b)\begin{equation} \frac{\partial \boldsymbol{u}(x,t)}{\partial t} = \boldsymbol{\mathsf{A}} \boldsymbol{u}(x,t) + \boldsymbol{f}(x,t), \quad \boldsymbol{\mathsf{A}} = - U \frac{\partial }{\partial x} + \mu(x) + \gamma \frac{\partial^{2} }{\partial x^{2}}, \end{equation}

and we use the parameters: $U=6$, $\gamma =1-\mathrm {i}$ and $\mu (x)=\beta \mu _c(1-x/20)$, where $\mu _c=U^{2}\, {\rm Re} (\gamma)/|\gamma |^{2}$ is the critical value for onset of absolute instability (Bagheri etal. Reference Bagheri, Henningson, Hœpffner and Schmid2009). The parameters are similar to those used by Lesshafft (Reference Lesshafft2018). The terms in $\boldsymbol{\mathsf{A}}$ correspond to advection, growth/decay and diffusion, respectively. Dirichlet boundary conditions are considered at $x=0$ and $40$, $\boldsymbol {u}(0,t)=\boldsymbol {u}(40,t)=0$, and the initial condition $\boldsymbol {u}(x,0)=0$ is used. We consider a system with $\beta = 0.1$, leading to a moderate gain separation between optimal and suboptimal modes.

System observations are given by

(3.2)\begin{equation} \boldsymbol{y}(t) = \boldsymbol{\mathsf{C}} \boldsymbol{u}(x,t), \end{equation}

where the operator $\boldsymbol{\mathsf{C}}$ imposes Gaussian-shaped sensors: $\boldsymbol{\mathsf{C}}$ is defined such that the $i$th entry of $\boldsymbol {y}(t)$ is

(3.3)\begin{equation} \boldsymbol{y}_i(t) = \frac{1} {\sqrt{2\pi\sigma_c^{2}}} \int \textrm{exp}{({-(x-x_i)^{2}}/{2\sigma_{c}^{2}})} \boldsymbol{u}(x,t)\,\textrm{d}x, \end{equation}

with $\sigma _c=0.5$. The number and positions of sensors ($x_i$) will be indicated for each case we consider.

The spatial domain is discretised using a second-order upwind differentiation scheme at points evenly distributed between $x=0$ and $40$, with ${\rm \Delta} x=0.1$, and the system is time integrated from $t=0$ to $t=500$ with a Crank–Nicolson scheme with time step, ${\rm \Delta} t = 10^{-2}$.

For application to a turbulent flow, the force term, $\boldsymbol {f}(x,t)$, would represent nonlinear interactions contained in the nonlinear advection term of the Navier–Stokes equations (McKeon & Sharma Reference McKeon and Sharma2010). Various assumptions will be made regarding the statistics of this term. We first consider $\boldsymbol {f}(x,t)$ to be spatially and temporally white. With the parameters cited above, $\boldsymbol{\mathsf{A}}$ is globally stable; it is locally convectively unstable in the first part of the domain ($x<20$), and locally convectively stable in the second part ($x>20$). A realisation of the system response to spatiotemporally white forcing is shown in figure 2.

Figure 2.

Sample of the response of the GL model under white noise forcing. Colour scale corresponds to responses real part. The growth of perturbations for $x<20$, their decay for $x>20$, and the convection behaviour of the model can be observed.

System gains are shown in figure 3. Optimal force and response modes for the full-rank system are shown in figure 4. The force modes observable by two sensors, and their corresponding responses, are shown in figure 5. We note that forces observable by different sensors, although linearly independent, can be quite similar: this is typically the case when the system has large gain separations. In this scenario, the extra information that can be obtained by adding a given sensor is more clearly visualised by plotting an orthogonal basis for the observable space: i.e. the component of the newly added observable force that is orthogonal to the previous observable space.

Figure 3.

Gains on the GL model.

Figure 4.

Ginzburg–Landau response ($a$) and force ($b$) modes. From top to bottom are the optimal and first two suboptimal modes. Solid lines correspond to the amplitude envelope and dashed lines show the real part.

Figure 5.

Observable forces ($\boldsymbol{\mathsf{R}}_{\boldsymbol {y}}^{\dagger } \boldsymbol{\mathsf{C}}^{\,\dagger }$ and their corresponding excited responses $\boldsymbol{\mathsf{R}}\boldsymbol{\mathsf{R}}_{\boldsymbol {y}}^{\dagger } \boldsymbol{\mathsf{C}}^{\,\dagger }$) for two Gaussian sensors at $x=10$ (blue) and $30$ (red). Solid and dotted lines correspond to the amplitude envelope and real part, the dashed back lines indicate sensor positions. On the right, observable forces and responses are made orthogonal.

Figure 6 compares Kalman-filter and resolvent-based estimations for $\boldsymbol {u}(10,t)$ and $\boldsymbol {u}(30,t)$ using one sensor at $x=20$. Force estimation from the Kalman filter were obtained using the estimated state in (2.1). Alternatively, they can be estimated using the frequency-domain counterpart of (2.1); in that case, it is necessary to account for windowing effects. This can be accomplished by multiplying (2.1) by a window function $w(t)$ and manipulation of the terms, leading to

(3.4)\begin{equation} \dfrac{\textrm{d}(w\boldsymbol{u})}{\textrm{d}t}(t) = \boldsymbol{\mathsf{A}} (w\boldsymbol{u})(t) + \boldsymbol{\mathsf{B}} (w\boldsymbol{f})(t) + \dfrac{\textrm{d}w}{\textrm{d}t}(t) \boldsymbol{u}(t), \end{equation}

with frequency-domain representation given by

(3.5)\begin{equation} \overline{\boldsymbol{u}}(\omega) = \boldsymbol{\mathsf{R}} \left(\boldsymbol{\mathsf{B}} \overline{\boldsymbol{f}}(\omega) + \overline{\boldsymbol{u}}'(\omega) \right), \end{equation}

where

(3.6a,b)\begin{equation} \overline{\boldsymbol{u}} (\omega) = \int_{-\infty}^{\infty} w(t)\boldsymbol{u}(t) \mathrm{e}^{ {\mathrm{i}} \omega t}\,\textrm{d}t, \quad \overline{\boldsymbol{u}}' (\omega) = \int_{-\infty}^{\infty} \frac{\textrm{d}w}{\textrm{d}t}(t)\boldsymbol{u}(t) \mathrm{e}^{ {\mathrm{i}} \omega t}\,\textrm{d}t \end{equation}

are respectively the estimation of $\hat {u}(\omega )$ using a window $w(t)$, and an associated force correction term that must be included to account for the window function. The method is described in detail by Martini etal. (Reference Martini, Cavalieri, Jordan and Lesshafft2019), with a discussion of the impact of window choice on the accuracy for a given sampling rate.

Figure 6.

Comparison between resolvent and Kalman-filter methods for state estimation with one sensor.

The Kalman-filter and resolvent-based estimations provide identical results for the downstream position, but only the resolvent-based estimation can estimate the upstream position, $x=10$. Figure 7 compares root-mean-square (r.m.s.) errors of both methods for different sensor configurations. Both methods show the same error downstream of the last sensor, with the resolvent method consistently showing smaller errors in other regions.

Figure 7.

Comparison between resolvent and Kalman-filter methods for states estimation with one, three and five sensors, indicated by vertical dotted lines. The dashed black line represents state r.m.s. on each position. Root-mean-square errors on each point are normalized by the global error r.m.s.

Two-point state correlations were constructed from the original and estimated systems. As seen in figure 8, the resolvent-based estimation requires a smaller number of sensors for an accurate estimation of two-point state statistics. Being smoother, responses are easier to estimate than forces, which, being white in space, are difficult to represent with a small number of force modes.

Figure 8.

Response and force two-point correlations for zero time-lag, obtained using the raw data (left most), Kalman-filter and resolvent-based estimations. Colour scale indicates the absolute level, normalized by the maximum value found on the raw data. White markers indicate the sensor position. ($a$) Response two-point correlations. ($b$) Force two-point correlations.

Cross-spectral density estimates were obtained for both forcing and state via the Welch (Reference Welch1967) method, with a window length of 40 % and 80 % overlap. Comparison of CSDs obtained with the original signal and results from Kalman-filter and resolvent-based estimations are shown in figure 9. Results again show that the resolvent-based approach leads to a more accurate CSD estimation for a given number of sensors. For cases with small numbers of sensors, it becomes clear that measurements may be used to estimate upstream information, which is not possible using a Kalman filter. Furthermore, the resolvent-based approach provides a faster convergence of both force and response estimations.

Figure 9.

Same as figure 8 for CSD at $\omega =1$. ($a$) Response CSD. ($b$) Force CSD.

3.2. Estimation of a system driven by coloured forcing

To assess the robustness of resolvent-based estimation when exact knowledge of force statistics is not available a priori, we construct resolvent-based and Kalman-filter estimators based on an assumption of spatiotemporally white forcing, and use these to estimate a system driven by coloured forces. We consider a forcing cross-spectral density model given by

(3.7)\begin{equation} \left\langle\,\skew2\hat{\boldsymbol{f}}(x_1,\omega) \skew2\hat{\boldsymbol{f}} (x_2,\omega)\right\rangle = {\rm e}^{{\rm i}k_h(x_2-x_1)}{\rm e}^{-({(x_1-x_2)^{2}}/{L_c^{2}})}{\rm e}^{-({(x_1-x_c(\omega))^{2}}/{L^{2}})} {\rm e}^{-({(x_2-x_c(\omega))^{2}}/{L^{2}})} + {\rm c.c.}, \end{equation}

where $\langle \cdot \rangle$ represents an ensemble average. The first term creates wave-like behaviour, the second imposes a coherence length and the two final terms define an amplitude envelope (Cavalieri etal. Reference Cavalieri, Jordan and Lesshafft2019). The following parameters were used: $k_h = 0.42$, $L=4$, $L_c=5$ and $x_c(\omega ) = 25-5|\omega |$.

Estimation is performed under the assumption that $\boldsymbol{\mathsf{F}}(t)=\boldsymbol{\mathsf{I}}\delta (t)$. Results for vanishing sensor noise are presented in figures 10 and 11. The method is capable of distinguishing the CSD for different frequencies, despite the underlying assumption of white forcing.

Figure 10.

Same as figure 8 for frequency-dependent CSDs. Estimation for $\omega =1$. ($a$)Response CSD. ($b$) Force CSD.

Figure 11.

Same as figure 8 for frequency-dependent CSDs. Estimation for $\omega =3$. ($a$)Response CSD. ($b$) Force CSD.

If, on the other hand, information is available concerning the force statistics, this can be used to improve the estimation performance. We illustrate this case using a rank-2 force CSD, constructed as

(3.8)\begin{gather} \langle\,\skew2\hat{\boldsymbol{f}}(x_1,\omega) \skew2\hat{\boldsymbol{f}} (x_2,\omega)\rangle = \boldsymbol{f}_1\ \boldsymbol{f}_1^{\dagger} + \boldsymbol{f}_2\ \boldsymbol{f}_2^{\dagger}, \end{gather}

(3.9)\begin{gather} \boldsymbol{f}_1(x) = e^{ -(x-5)^{2}}+e^{ -(x-35)^{2}}, \end{gather}

(3.10)\begin{gather} \boldsymbol{f}_2(x) = e^{ -(x-15)^{2}}+e^{ -(x-25)^{2}}. \end{gather}

This produces forces at $x=5$ that are perfectly correlated with forces at $x=35$, and likewise for positions $15$ and $25$. Figure 12 compares resolvent-based estimations obtained using the white force assumption to those obtained using the real force CSD. Use of the correct force CSD leads to a substantial improvement in the estimation of both response and force when a single sensor is used, and an exact estimation when two sensors are used, as a result of the very low rank of the force considered.

Figure 12.

Same as figure 8, comparing estimation with the assumption of white-noise forcing, and using the correct underlining force CSDs. Estimation for $\omega =3$. ($a$) Response CSD. ($b$) Force CSD.

Although observable forces for this system are always upstream of the sensors, downstream forces are correlated with observable forces: these forces, and their responses are thus correctly estimated.

3.3. Resolvent-based estimation as an alternative to the Kalman filter

Towne etal. (Reference Towne, Lozano-Durán and Yang2020) suggested that resolvent-based estimation can provide a departure point for the control of complex turbulent flows. But real-time control requires causal estimation. We therefore consider a truncation of the kernel of the resolvent-based estimator to is causal component (Sasaki etal. Reference Sasaki, Tissot, Cavalieri, Silvestre, Jordan and Biau2016), and we compare this to kernels obtained using the standard Kalman filter.

The comparisons of figure 6 suggest that Kalman-filter and resolvent-based estimators may be equivalent at positions downstream of the sensor. This is an interesting possibility, as the Kalman filter is a central feature of LQG control methods. If the methods are indeed equivalent when the resolvent-based estimator is truncated to its causal component, then its low cost would enable extension of LQG methods to large systems. We proceed by comparing the kernels obtained for the two methods.

The resolvent-based kernel is obtained by converting the resolvent-based estimation function to the time domain by inverse Fourier transform of (2.22). A state estimation is then obtained via convolution of the kernel, ${\boldsymbol{\mathsf{T}}}_{\boldsymbol {u}}$, and readings, $\boldsymbol {y}(t)$, i.e.

(3.11)\begin{equation} \boldsymbol{u}(t) = \int_{-\infty}^{\infty} \boldsymbol{\mathsf{T}}_{\boldsymbol{u}} (\tau) \boldsymbol{y}(t-\tau) \,\textrm{d} \tau, \end{equation}

where ${\boldsymbol{\mathsf{T}}}_{\boldsymbol {u}} (t) = \mathcal {F}^{-1}\left ({ \skew4\hat {\boldsymbol{\mathsf{T}}}_{\boldsymbol {u}}} (\omega ) \right )$ (Sasaki etal. Reference Sasaki, Tissot, Cavalieri, Silvestre, Jordan and Biau2016). The Kalman-filter estimation is obtained via integration of

(3.12)\begin{equation} \dfrac{\textrm{d}\tilde{\boldsymbol{u}}}{\textrm{d}t} = \boldsymbol{\mathsf{A}} \tilde{\boldsymbol{u}} + \boldsymbol{\mathsf{L}}(\boldsymbol{y}-\boldsymbol{\mathsf{C}} \tilde{\boldsymbol{u}}), \end{equation}

where $\boldsymbol{\mathsf{L}}$ is the Kalman gain. A formal solution is

(3.13)\begin{equation} \tilde{\boldsymbol{u}}(t) = \int_{-\infty}^{t} \mathrm{e}^{\left(\boldsymbol{\mathsf{A}}-\boldsymbol{\mathsf{L}}\boldsymbol{\mathsf{C}}\right)(t-\tau)} \boldsymbol{\mathsf{L}} \boldsymbol{y}(\tau) \,\textrm{d} \tau = \int_{-\infty}^{\infty} \boldsymbol{\mathsf{T}}_{\boldsymbol{u}}^{kal}(\tau) \boldsymbol{y}(t-\tau) \,\textrm{d} \tau, \end{equation}

where

(3.14)\begin{equation} \boldsymbol{\mathsf{T}}_{\boldsymbol{u}}^{kal}(t) = \begin{cases} \mathrm{e}^{\left(\boldsymbol{\mathsf{A}}-\boldsymbol{\mathsf{L}}\boldsymbol{\mathsf{C}}\right)t} \boldsymbol{\mathsf{L}}, & t \ge 0, \\ 0, & t < 0, \end{cases} \end{equation}

is the Kalman-filter kernel.

The kernels, ${\boldsymbol{\mathsf{T}}}_{\boldsymbol {u}}$ and ${\boldsymbol{\mathsf{T}}}_{\boldsymbol {u}}^{kal}$, are compared using different sensor configurations. Three probes, at locations $x=5,20$ and $35$, referred to as $z_1$, $z_2$ and $z_3$, are used. Figure 13 shows kernels constructed using one sensor. The Kalman-filter kernel for probe $z_3$ exactly reproduces the resolvent-based kernel for all values of sensor noise, and the causal part of the resolvent-based kernel closely approximates that of the Kalman filter for the probe $z_2$. While the Kalman filter makes no estimation on the probe $z_1$, the resolvent-based estimation provides a non-causal estimation.

Figure 13.

Comparison between resolvent estimation (solid lines) and Kalman-filter estimation (dashed lines) kernels using one sensor for different values of sensor noise $\epsilon$. Blue, red and yellow lines indicate kernels for probes $z_1$, $z_2$ and $z_3$, respectively. Sensor and probes locations are shown on the bottom. In the legend, $\mathcal {T}_{u,y_i,z_j}$, corresponds to the transfer function term to be convoluted with the $i$th sensor for obtaining estimates at the $j$th probe.

The match between the causal component of the resolvent-based kernel and that of the Kalman filter reinforces the idea that the latter can be obtained via a truncation of the resolvent-based kernel. This is in accordance with the results obtained by Fraser & Potter (Reference Fraser and Potter1969): an optimal non-causal (smoothing) estimation is equivalent to the combination of a pair of Kalman filters, one moving forward and the other backward in time; in systems where no information can be gained from the backward-directed filter, the optimal causal and non-causal estimates coincide.

Figure 14 shows the kernels obtained with inclusion of an additional upstream sensor. The kernels are similar for probes $z_2$ and $z_3$. However, the resolvent-based kernel for probe $z_1$ has causal contributions from sensor $y_1$ and non-causal contributions from sensor $y_2$; and, as suggested by the foregoing argument, resolvent-based and Kalman-filter kernels do not match. Addition of a further downstream sensor leads to kernels that differ for all probe locations, as shown in figure 15.

Figure 14.

Same as figure 13 using two sensors, with sensor noise of $\epsilon =10^{-2}$.

Figure 15.

Same as figure 13 using three sensors, with sensor noise of $\epsilon =10^{-2}$.

When the methods are not equivalent, we note that the resolvent-based kernel is consistently lower than those of the Kalman filter. Truncation of the former to its causal components will thus lead to an underestimation. This can also be interpreted in terms of the results of Fraser & Potter (Reference Fraser and Potter1969): as the non-causal estimation is the average of the forward- and backward-moving Kalman filters, truncating the resolvent-based kernel to its causal part corresponds to a zeroing of the backward-directed filter, leading to an underestimation.

The dominance of the downstream-travelling mode in the present model makes it easy to identify sensor locations which will produce predominantly causal resolvent-based kernels, and from which Kalman-filter estimation can be obtained. Many fluid flows of interest are dominated by downstream-travelling modes (incompressible boundary layers and jets, see Beneddine etal. (Reference Beneddine, Yegavian, Sipp and Leclaire2017); Sasaki etal. (Reference Sasaki, Piantanida, Cavalieri and Jordan2017b) for examples) and may therefore be amenable to the obtention of a Kalman-filter estimator, using the resolvent-based approach we describe, at substantially lower computational cost and without requiring the construction of a reduced-order model. In more complex systems, where upstream- and downstream-travelling modes may be relevant, the procedure is not necessarily simple, or possible. In such cases optimal non-causal kernels can be converted to optimal causal kernels by solving associated Wiener–Hopf equations (Martinelli Reference Martinelli2009). Such equations are, however, not easily solved if more than one sensor is used.

3.4. Comparison with truncated response-mode estimation

The recognition that large resolvent gain separation implies that the response of a system may be relatively insensitive to the specific details of its underlying driving forces (McKeon & Sharma Reference McKeon and Sharma2010) has motivated several studies (Gomez Carrasco etal. Reference Gomez Carrasco, Blackburn, Rudman, Sharma and McKeon2014; Beneddine etal. Reference Beneddine, Yegavian, Sipp and Leclaire2017; Symon etal. Reference Symon, Dovetta, McKeon, Sipp and Schmid2017) in which the system response is estimated using the dominant response modes. We refer to such approaches as truncated response-mode estimation (TRME). Response-mode amplitudes and phases are obtained by fitting them to available flow data. Parabolized stability equations (PSE) have also been used with the same underlying idea (Sasaki etal. Reference Sasaki, Piantanida, Cavalieri and Jordan2017b): these track the fastest growing perturbation, which often approximates the optimal resolvent response mode (Towne, Rigas & Colonius Reference Towne, Rigas and Colonius2019), and can therefore be understood as a rank-1 truncated response-mode estimation.

Beneddine etal. (Reference Beneddine, Yegavian, Sipp and Leclaire2017) showed that sensors for a jet need to be located in regions where the optimal modes have significant amplitude; with this choice of sensor location, flow quantities downstream were estimated. Although successful, the approach lacks a rigorous justification: it is not expected that such a basis will be optimal for any choice of sensors. The authors argue that sensors should be located around the peak of the response mode, which should lead to better signal-to-noise ratios. Our results suggest that such locations are indeed desirable, but with a slightly different interpretation: the observable force subspace for sensors at these locations has a larger projection of the optimal force mode. A rank-1 TRME was also used to improve a data assimilation of the flow around an airfoil by Symon (Reference Symon2018), but to the best of our knowledge higher-rank estimations were not attempted. In what follows we show that higher-order TRMEs are not suitable for estimation, and we provide an explanation as to why this is the case.

The orthogonality of the response-mode basis suggests that this constitutes an efficient basis if the projection is made using an $L^{2}$ inner product. This norm can be emulated using many sensors uniformly distributed throughout the flow. Figure 16 compares errors associated with resolvent-based and truncated response-mode estimations when uniformly spaced sensors are used. For the latter, the number of response modes corresponds to the number of sensors. Response-truncation errors are always larger than the resolvent-based estimation, but the results are similar.

Figure 16.

Comparisons between r.m.s. errors of resolvent-based estimations and TRME. Allcurves are normalised by the global r.m.s. error.

A different picture emerges when sensors are not uniformly spaced, but concentrated in the upstream region of the domain. Figure 17 shows that TRME errors are orders of magnitude larger than the proposed method. This is a consequence of an ill-posed fit of the optimal response mode basis to the collocation points used. From figure 4 we see that only the optimal force mode is biased towards the upstream flow region, suboptimal force modes being spread throughout the domain. As a sensor in a convection-dominated system can only observe forces upstream of its location, the use of upstream sensors to obtain forces spread throughout the domain leads to an ill-posed fitting. The resolvent-based estimation, on the other hand, focuses on the observable force subspace, thus providing a naturally stable method.

Figure 17.

Comparisons between r.m.s. errors of resolvent-based and TRME methods with sensors located at an upstream region. All curves are normalised by the global r.m.s. error.

We illustrate the origin of the TRME ill-posed fitting with a simpler problem: curve fitting using polynomials, as shown in figure 18. Fittings were obtained using uniformly spaced collocation points, Gauss–Lobatto collocation points and collocation points located on a small domain. Even though all fits use the same basis, fit performances vary considerably: Gauss–Lobatto points provide uniform, exponential convergence (Trefethen Reference Trefethen2000); equispaced points provide non-uniform convergence; and localised collocation points lead to large errors, which increase with the polynomial order. This situation can be understood as analogous to that which underpins the results of figure 17, once it is realised that the procedure is implicitly trying to represent the systems forces with a particular choice of basis: force modes in the case of TRME and observable forces in the case of the resolvent-based method.

Figure 18.

Fitting of a Gaussian curve using polynomials of orders 4 and 15, markers indicate the collocation points used. Collocation points are ($a$) uniformly spaced on the first half of the domain, ($b$) uniformly spaced throughout the domain, ($c$) on Gauss–Lobatto points.

For estimation purposes, the instability of TRME may be worked around by using more sensors than response modes, leading to a least-square minimisation of the sensor errors, as was done by Gomez Carrasco etal. (Reference Gomez Carrasco, Blackburn, Rudman, Sharma and McKeon2014). The sensors should be weighted in order to reproduce an inner product for which the response modes are orthogonal. This approach, however, discards sensor information, which could be used for higher accuracy. It can be justified if resolvent modes are readily available, but if they need to be constructed by traditional methods, this would be more demanding than using the method we propose.

4.1. Transitional flat-plate boundary layer

We perform state-estimation in a two-dimensional spatially evolving transitional boundary layer, with Reynolds number $Re=1000$ based on the displacement thickness at the beginning of the domain. The free stream velocity is taken as the reference value for non-dimensional quantities. The base flow, obtained from a Blasius boundary layer solution, is shown together with sensors and probes used in figure 19. The linearised Navier–Stokes equations are solved using the spectral-element code Nek5000 (Fischer& Patera Reference Fischer, Patera, Kane, Carlson and Cox1989; Fischer Reference Fischer1998), which uses $n$th-order Lagrangian interpolants within each element to solve a weak formulation of the incompressible Navier–Stokes equations.

Figure 19.

Boundary layer base flow. Colour contours indicate the streamwise velocity component, dotted and dashed lines indicate the boundary layer thickness ($\delta _{0.99}$) and displacement thickness ($\delta ^{*}$). White markers correspond to the locations of perturbation sensors used for estimation and green markers show the control positions used in figures 22 and 23.

The computational domain is a box of size $1000\times 20$, discretised with $250 \times 25$ elements using sixth-order polynomials, ninth-order polynomials being used for de-aliasing. Time integration was performed using a time step of $10^{-1}$. Convergence with mesh and time step were checked. Dirichlet boundary conditions for velocity fluctuations were used on all boundaries. At positions $x=0$ and $x=1000$ the boundary conditions create a viscous boundary which is not well resolved by the mesh. Outflow conditions can alternatively be obtained using extended domains together with sponge regions, as described by Bodony (Reference Bodony2006), but such treatment was not necessary here, as the regions around the boundary are stable with the numerical methods used here, and associated errors are both small and localised.

The flow is forced with a spatiotemporally white noise for $y<5$. The distributed forcing excites, in addition to unstable Tollmien–Schlichting waves, many other stable modes, leading to a more challenging case for estimation in comparison to forcing in a limited upstream region, which is often considered in flow-control problems (Bagheri etal. Reference Bagheri, Henningson, Hœpffner and Schmid2009; Belson etal. Reference Belson, Semeraro, Rowley and Henningson2013; Sasaki etal. Reference Sasaki, Morra, Fabbiane, Cavalieri, Hanifi and Henningson2018).

Readings from control sensors at $x=150,500$ and $850$ were saved during the direct run, defined in § 2.4. Estimation for these positions is obtained via the transfer function

(4.1)\begin{equation} \hat{\boldsymbol{\mathsf{T}}_{ \boldsymbol{z}}} = \boldsymbol{\mathsf{C}}_{ \boldsymbol{z}} \hat{\boldsymbol{\mathsf{T}}_{\boldsymbol{u}}} = \boldsymbol{\mathsf{C}}_{ \boldsymbol{z}}\boldsymbol{\mathsf{R}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}}^{\dagger} \left(\boldsymbol{\mathsf{R}}_{\boldsymbol{y}} \boldsymbol{\mathsf{R}}_{\boldsymbol{y}}^{\dagger} +\epsilon \boldsymbol{\mathsf{I}}\right), \end{equation}

which can be directly evaluated from the available direct-run readings. The observable force and response modes for each sensor are shown in figure 20. In figure 21 the modes observable by the second sensor are made orthogonal to those observable by the first, as was previously done for the Ginzburg–Landau model (cf. figure 5). Figure 22 shows estimations at these points using only one sensor at $x=300$, and using sensors at $x=300$ and $700$. The estimates are especially accurate for the two downstream positions, and inclusion of the second sensor reduces the error for both. For the upstream point, $x=150$, a reasonable accuracy is obtained. The kernels are shown in figure 23. As in the Ginzburg–Landau model, estimation is causal for positions downstream of the last sensor; this suggests that resolvent-based estimation is equivalent to Kalman-filter estimation for these locations.

Figure 20.

Observable forces ($a,b$) and responses ($c,d$) for the sensors located at $x=300$ ($a,c$) and $x=700$ ($b,d$) for $f=0.004$. Dotted and dashed lines indicate the boundary layer thickness ($\delta _{0.99}$) and displacement thickness ($\delta ^{*}$).

Figure 21.

Same as figure 20, with forces and responses on the right made orthogonal to the ones on the left.

Figure 22.

Boundary layer time series estimation at $x=150$ ($a$), 500 ($b$) and 850 ($c$). Reconstructions using one (red line) and two (blue line) sensors are shown along with the simulation signal at these position (dashed line).

Figure 23.

Resolvent-based kernels for probes $z_1,\ z_2 \ {\rm and}\ z_3$, at three positions ($x=150,500$ and $850$). Blue and red lines correspond to the term to be convoluted with $y_1$ and $y_2$, respectively. ($a$)Kernel using only the upstream sensor, $y_1$. ($b$) Kernel using both sensors, $y_1$ and $y_2$.

Appendix B shows how to estimate a flow snapshot in a memory efficient manner. Snapshots of the flow and its estimation, and their error, are seen in figure 24; the fields are virtually indistinguishable at downstream locations, slight differences being observed upstream: this region is not dominated by Tollmien–Schlichting waves, which here have low amplitudes; the flow here thus has higher rank, and the presence of multiple modes makes this region harder to estimate using downstream sensors. Nonetheless, the method provides a good overall estimation given that the upstream region contains less energy. The ratio between error energy, defined as

(4.2)\begin{equation} e_{err} = \dfrac{1}{2}\int \left( (u(x,y) - \tilde{u}(x,y))^{2} + (v(x,y) - \tilde{v}(x,y))^{2} \right)\,\textrm{d}v \end{equation}

and the flow energy is $2.3\,\%$.

Figure 24.

Snapshot of the $u$ component of the boundary layer. ($a$) Real flow, ($b$) two-sensor estimation and ($c$) error. The plots are normalized by the maximum snapshot streamwise flow speed. Note the lower range of the colour scale of the error plot.

4.2. Turbulent channel flow

We now consider a turbulent channel flow with friction Reynolds number $Re_{\tau }=180$. We use data obtained via direct numerical simulation (DNS) to compute the flow and nonlinear forcing statistic, both of which were previously documented by Morra (Reference Morra2020) and Morra etal. (Reference Morra, Nogueira, Cavalieri and Henningson2020), to compare with our estimation results. Inner (viscous) units can be defined in terms of a friction velocity ${ u_{\tau } = \sqrt {{\tau _w} / {\rho }}}$, where $\tau _w$ is the mean wall shear and $\rho$ the fluid density. These are denoted using the superscript ‘$+$’.

A direct numerical solution of the nonlinear Navier–Stokes equations was obtained using the channelflow code (Gibson, Halcrow & Cvitanović Reference Gibson, Halcrow and Cvitanović2008; Gibson Reference Gibson2012) version 2.0. Fourier bases are used for streamwise ($x$) and spanwise ($z$) directions, Chebyshev polynomials being used in the wall-normal ($y$) direction. De-aliasing is performed in the $x$ and $z$ directions. A box of size $4\pi \times 2 \times 2\pi$ with $288 \times 129 \times 288$ grid points was used. The simulation was started with a random initial condition and run until transient effects vanished and stationary turbulence was obtained; only data after this time was used. A snapshot of the flow is shown in figure 25, and flow statistics (mean flow and r.m.s. profiles) are presented in figure 26. Comparison with previous DNS results from Del Alamo & Jiménez (Reference Del Alamo and Jiménez2003) validates the database.

Figure 25.

Snapshot of the streamwise velocity component of the turbulent channel flow.

Figure 26.

Turbulent-channel mean flow ($a$), and perturbation root-mean-square values ($b$). Lines indicate results from the current study, while squares results from Del Alamo & Jiménez (Reference Del Alamo and Jiménez2003).

The nonlinear terms, considered as external forces of the linear operator as discussed earlier, were computed using Fourier and Chebyshev differentiation matrices (Weideman& Reddy Reference Weideman and Reddy2000) saved along with the responses. A total number of 2000 snapshots were saved every ${\rm \Delta} t = 0.5$ in outer units (${\rm \Delta} t^{+} = 5.7$). Force and response statistics were computed, CSDs being obtained using the Welch method with a Hanning window, 256 samples per block, and with a 75 % overlap. Force CSDs are denoted by $\skew4\hat {\boldsymbol{\mathsf{F}}}^{\,\prime} (\omega )$, while the symbol $\skew4\hat {\boldsymbol{\mathsf{F}}}(\omega )$ is reserved for the CSD model used to construct the estimator. Windowing effects were corrected using the method developed by Martini etal. (Reference Martini, Cavalieri, Jordan and Lesshafft2019) and described in § 3.1; Nogueira etal. (Reference Nogueira, Morra, Martini, Cavalieri and Henningson2020) showed that the correction is fundamental for the correspondence of force and response modes in turbulent flows.

The Navier–Stokes equations, linearised around the mean, are homogeneous in the $x$ and $z$ directions. Applying a Fourier transform in the homogeneous directions, the streamwise perturbation velocity component, $u(x,y,z,t)$, can be written as $u'(\alpha,y,\beta,t)$. For brevity, the wavenumbers $\alpha$ and $\beta$ will be omitted in what follows.

It is well known that near-wall structures are dominated by $({2\pi }/{\alpha _+},\ {2\pi }/{\beta _+})=(\lambda _{x}^{+},\lambda _{z}^{+})=(1000,100)$ (Del Alamo & Jiménez Reference Del Alamo and Jiménez2003), and we thus focus on this wavenumber. Two Fourier-mode sensors are used, each measuring streamwise and spamwise components of the wall shear stress. For these wavenumbers, the linearised Navier–Stokes operator depends only on $y$ and can thus be explicitly constructed and inverted to obtain the resolvent operator. Here we consider the standard linearised Navier–Stokes operator with terms nonlinear in fluctuations considered as forcing (McKeon & Sharma Reference McKeon and Sharma2010).

Non-causal resolvent-based estimation was performed using (2.18) and (2.22), using the following four different models for the force CSD.

1. (i) White. Spatiotemporally white forcing: $\skew3\hat {\boldsymbol{\mathsf{F}}}(\omega ) = \boldsymbol{\mathsf{I}}$.

2. (ii) White in time. White-noise in time, with spatial correlation given by the two-point correlation of the DNS: $\skew3\hat {\boldsymbol{\mathsf{F}}}(\omega ) = \boldsymbol{\mathsf{F}}'(0)$.

3. (iii) Estimated colour. Force statistics estimated from the CSD of streamwise and wall-normal velocity components.

4. (iv) True colour. Force statistics calculated from the DNS: $\skew3\hat {\boldsymbol{\mathsf{F}}}(\omega ) = \skew4\hat{\boldsymbol{\mathsf{F}}}^{\,\prime}(\omega )$.

Force-statistic estimation (iii) was performed using the CSD of an auxiliary set of sensors, referred to as $y'$, that consisted of streamwise and wall-normal velocity perturbations at $y^{+}=5,10,15,20,35$ and $40$. The force CSD was estimated as

(4.3)\begin{equation} \skew4\hat{\boldsymbol{\mathsf{F}}} = \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}}^{\ \prime} \skew4\hat{\boldsymbol{\mathsf{Y}}}^{\ \prime} \skew4\hat{\boldsymbol{\mathsf{T}}}_{\boldsymbol{f}}^{\ \prime}, \end{equation}

where $\skew4\hat{\boldsymbol{\mathsf{Y}}}^{\,\prime}$ is the CSD of the auxiliary sensor set: $\skew4\hat {\boldsymbol{\mathsf{Y}}}^{\ \prime} =\langle \hat {\boldsymbol {y}}' \hat {\boldsymbol {y}}'^{\dagger } \rangle$, and $\skew3\hat {\boldsymbol{\mathsf{T}}}_{\boldsymbol {f}}^{\ \prime}$ is the equivalent of (2.18) constructed for the auxiliary set of sensors $\boldsymbol {y}'$. This is similar to the procedure in Towne etal. (Reference Towne, Lozano-Durán and Yang2020), with the difference that here the estimated force CSD is used to improve estimation using a different set of measurements. Note that $\skew4\hat {\boldsymbol{\mathsf{Y}}}^{\,\prime}$ can be constructed in experiments using only two sensors. The auxiliary sensors are not used in the flow estimation described next. The choice of sensors was motivated by Nogueira etal. (Reference Nogueira, Morra, Martini, Cavalieri and Henningson2020), where correlations between wall-normal velocities and streamwise forces is shown to be important for the flow dynamics. Force CSDs for $\lambda _t^{+}=100$, where $\lambda _t^{+} = 2\pi /\omega ^{+}$ denotes the period in inner units, obtained directly from the DNS and estimated using the auxiliary sensors are shown in figure 27. The estimated colour has considerable differences when compared to the true colour. Nevertheless, as will be shown next, it contains the necessary information for an accurate state estimation.

Figure 27.

Force CSDs ($\skew3\hat {\boldsymbol{\mathsf{F}}}\,$) for $\lambda _t^{+} =100$ obtained directly form DNS data (Morra Reference Morra2020) ($a$) and estimated using the extra sensor set ($b$). Colour scale is saturated for the $\skew3\hat {\boldsymbol{\mathsf{F}}}_{f_x,f_x}$ component, which is an order of magnitude greater than the others as to make cross-correlations clearer.

A comparison of power spectral densities (PSD) obtained from the DNS and estimation at $y^{+}=12$, where the PSD of the streamwise velocity component peaks, is presented in figure 28. The PSD for $\lambda _t^{+} = 100$ at different wall-normal positions is show in figure 29, and a time-series comparison is provided in figure 30. The largest velocity fluctuations are in the streamwise direction, and it is therefore the component for which estimation performance is best.

Figure 28.

Turbulent-channel power-spectral densities for $\lambda _{t}^{+} = 100$ of $u$, $v$ and $w$ from DNS and estimation using a different assumption of force statistics.

Figure 29.

Turbulent-channel power-spectral densities of $u$ at $y^{+} =12$ obtained from the DNS and estimation using a different assumption of force statistics.

Figure 30.

Turbulent-channel streamwise velocity perturbation time series obtained from the DNS and estimation using a different assumption of force statistics. Estimation at $y^{+}=12$.

The assumption of white-noise force in time and space does not provide a satisfactory flow estimation, as previously observed by Chevalier etal. (Reference Chevalier, Hœpffner, Bewley and Henningson2006). Studying the structure of the nonlinear force terms, Nogueira etal. (Reference Nogueira, Morra, Martini, Cavalieri and Henningson2020) showed that streaks generated by the lift-up mechanism have their amplitudes reduced by streamwise forces. This effect is not present if forces are assumed to be spatiotemporally white, leading to the overshoot observed in figure 28.

Chevalier etal. (Reference Chevalier, Hœpffner, Bewley and Henningson2006) explored spatial colouring of the forces, which were nonetheless simplified as white noise in time, probably due to the complexity of including full colour information in a Kalman-filter estimation. To the best of the authors’ knowledge, no previous study has included full force colour. With the resolvent-based estimation procedure we propose, estimation of relevant force components and their use for response estimation is straightforward.

The use of spatiotemporal forcing colour provides significant improvements in estimation performance. Low-rank forcing models can easily be obtained from experimental set-ups from sensor CSDs and the resolvent operator, and can thus be a useful tool to estimate full flow states in experimental set-ups, where one is typically limited to low-rank observations of the system.