---
title: Causal Resolvent Estimation Kernels
url: https://www.emergentmind.com/topics/causal-resolvent-based-estimation-kernels
type: topic
---

# Causal Resolvent Estimation Kernels

Searching arXiv for the cited papers and closely related work to ground the article.
Causal resolvent-based estimation kernels are linear impulse-response operators that map sensor measurements to estimates of target quantities under an explicit real-time constraint: the estimate may depend only on present and past data. In the flow-estimation literature, the construction is resolvent-based because it is derived from the linear input-output map of a linearized dynamical system together with forcing and noise cross-spectral densities, and it is causal because the kernel is required to vanish for negative time. This formulation is developed for globally stable linear systems in "Resolvent-based tools for optimal estimation and control via the Wiener-Hopf formalism" [2201.00361] and extended to a globally unstable turbulent compressible wake in "Resolvent-based estimation of a turbulent wake" [2507.18837]. Related work uses causal kernels and resolvent regularization in adjacent senses, including Gaussian-process connectivity kernels in neuroscience [1705.05603], kernel instrumental variable regression [2511.21603], and proximal causal learning with proxy variables [2105.04544].

## 1. Formal definition and operator setting

In the canonical formulation, the underlying system is linear time-invariant with stochastic forcing, noisy measurements, and a target quantity to be estimated. One form used in the literature is
$$
\frac{du}{dt}(t)=Au(t)+f(t)+a(t),\qquad y(t)=u(t)+n(t),\qquad z(t)=u(t),
$$
with spatial supports handled by input-output operators and with forcing and noise specified through cross-spectral densities \(F(\omega)\) and \(N(\omega)\). In Fourier space, the linear dynamics are represented by the resolvent
$$
R(\omega)=(-i\omega I-A)^{-1}.
$$
The non-causal estimator is written as a convolution over all times,
$$
\tilde z(t)=\int_{-\infty}^{\infty}T_{z,nc}(\tau)\,y(t-\tau)\,d\tau,
$$
whereas the causal estimator is written as
$$
\tilde z(t)=\int_{0}^{\infty}T_{z,c}(\tau)\,y(t-\tau)\,d\tau.
$$
The latter is the defining support constraint: \(T_{z,c}(t<0)=0\) [2201.00361].

The estimator is optimal with respect to a quadratic mean-square criterion. In the non-causal case, minimization produces a frequency-domain optimality condition expressed through sensor auto-spectra and sensor-target cross-spectra. In the stable-system formulation, the quantities entering this condition are
$$
\hat G = RFR^\dagger + N,\qquad \hat G = RFR^\dagger,
$$
and the estimator is the linear map from measurements to targets that minimizes the estimation cost under those second-order statistics. This establishes the basic meaning of a causal resolvent-based estimation kernel: it is the causal version of the optimal linear estimator induced by the resolvent and the spectral statistics of disturbances and noise [2201.00361].

## 2. Spectral construction in turbulent and unstable flows

The 2025 turbulent-wake extension applies the same conceptual framework to the wake of a spanwise-periodic NACA0012 airfoil at \(M_\infty=0.3\), \(Re_{L_c}=23{,}000\), and \(\alpha=6^\circ\). The flow contains a laminar separation bubble on the suction side, transition to turbulence around \(x/L_c\approx 0.6\), a turbulent wake downstream, and a shedding band around \(St_\alpha\approx 0.15\)–\(0.2\). The dataset contains \(n_t=75{,}000\) snapshots over \(tU_\infty/L_c\in[0,324]\) with sampling \(\Delta t U_\infty/L_c=0.00432\), corresponding to about \(466\) shedding cycles. Three representations are considered: spanwise-averaged flow, spanwise-Fourier-transformed flow, and the mid-span plane [2507.18837].

The technical obstacle in that setting is that the linearized Navier-Stokes operator about the mean flow is not globally stable. The classical resolvent
$$
\mathsfbi{R}(\omega)=(i\omega \mathsfbi{I}-\mathsfbi{A})^{-1}
$$
is therefore not well-posed in the usual sense when \(\mathsfbi{A}\) has unstable eigenvalues. The paper circumvents explicit unstable-resolvent construction by replacing the needed quantities with empirical cross-spectral densities from large-eddy simulation. With measured outputs \(\boldsymbol{y}\), target outputs \(\boldsymbol{z}\), forcing \(\boldsymbol{f}\), and sensor noise \(\boldsymbol{n}\), the noncausal estimator is written directly as
$$
\hat{\mathsfbi{T}_{nc}}=\mathsfbi{S}_{zy}(\mathsfbi{S}_{yy}+\hat{\mathsfbi{N}})^{-1}.
$$
This makes the method data-driven and preserves the role of colored nonlinear forcing statistics through \(\hat{\mathsfbi{F}}\), which the paper identifies as essential for turbulence; white-noise forcing would generally be too crude [2507.18837].

This construction also clarifies the relation between “resolvent-based” and “data-driven.” The estimator remains resolvent-based at the level of interpretation, because the spectra are those induced by the input-output dynamics of the linearized system, but the computational route proceeds through empirical second-order statistics rather than through direct assembly of the unstable resolvent [2507.18837].

## 3. Causality, Wiener–Hopf factorization, and analytic structure

Causality is not imposed by truncating a non-causal kernel after it has been computed. Instead, the cited work imposes causality as a constrained optimization problem, with Lagrange multipliers supported on the negative-time side. In the stable-system formulation, the causal optimality condition becomes a matrix Wiener-Hopf equation,
$$
\hat T_{z,c}\hat G + \hat{}_- = \hat G,
$$
whose solution requires factorization into plus and minus parts that are analytic in the upper and lower half-planes, respectively. The explicit estimator is
$$
\hat T_{z,c}(\omega)=\left(\hat G(\omega)\hat G_-^{-1}(\omega)\right)_+ \hat G_+^{-1}(\omega),
$$
where \((\cdot)_+\) and \((\cdot)_-\) denote the causal and anti-causal parts in the Wiener-Hopf sense [2201.00361].

The turbulent-wake paper gives an equivalent causal formula in terms of empirical spectral matrices,
$$
\hat{\mathsfbi{T}_{c}(\omega)} = \Big[ \mathsfbi{S}_{zy}(\mathsfbi{S}_{yy}+\hat{\mathsfbi{N}})^{-1}_{-} \Big]_+ (\mathsfbi{S}_{yy}+\hat{\mathsfbi{N}})_{+}^{-1}.
$$
Its appendix states that multiplicative factorization is the hard part, and solves it numerically through a Fredholm-type integral equation together with the Hilbert transform
$$
\mathcal{H}(\hat{\boldsymbol{x}})=P.V.\frac{1}{\pi}\int_{-\infty}^{\infty}\frac{\hat{\boldsymbol{x}}(u)}{\omega-u}\,du,
$$
using GMRES [2507.18837].

A central misconception addressed by both papers is that a causal estimator can be obtained by simply discarding the negative-time part of the optimal non-causal kernel. The turbulent-wake results show why this fails: in the spanwise-averaged case, the noncausal kernels contain negative-time peaks, including \(\tau_{\text{peak}}=[-0.360,\ 0.171,\ 0.676,\ 1.25]\) for selected downstream targets, so naive truncation discards informative structure. The causal Wiener-Hopf solution instead redistributes the kernel optimally under the support constraint [2507.18837].

## 4. Numerical realization, scalability, and sensor placement

The computational appeal of the framework is that the design problem is organized around spectral objects whose size depends on sensor and actuator channels rather than on the full state dimension. In the stable-system formulation, the required Wiener-Hopf coefficients are built by matrix-free time-marching, using direct and adjoint runs to assemble quantities such as \(R\), \(R^\dagger\), \(RR^\dagger\), and \(R^\dagger R\) without forming the full resolvent explicitly. The numerical Wiener-Hopf factorization is handled with a Daniele-type Fredholm/GMRES procedure, and the practical implementation requires zero-padding because FFT-based Hilbert transforms assume periodicity; the paper reports padding about \(20\) times the time signal [2201.00361].

The turbulent-wake implementation is integrated directly into CharLES in C++, uses PETSc and FFTW, and is designed for HPC use. For high-rank targets, an online DFT is used to reduce memory usage. Cross-spectral densities are constructed from streaming spectral estimates with Hamming windows and \(50\%\) overlap. A convergence study selects the largest tested setup, with \(N_{freq}=5000\), \(N_{ovlp}=2500\), and \(N_{blks}=29\) [2507.18837].

Sensor placement is treated empirically rather than as a formal optimization problem. The paper uses single-sensor estimation error maps and coherence analysis. The error metric is
$$
E= \frac{\sum_i \int (\tilde{\boldsymbol{z}}_i(t)-\boldsymbol{z}_i(t))^2\,dt} {\sum_i \int \boldsymbol{z}_i(t)^2\,dt},
$$
and the magnitude-squared coherence is
$$
\hat{\gamma}_{yz}(\omega)= \frac{|\hat{\mathsfbi{S}}_{yz}(\omega)|} {\sqrt{\hat{\mathsfbi{S}}_{yy}(\omega)\hat{\mathsfbi{S}}_{zz}(\omega)}}.
$$
The main findings are that sensors near the trailing edge are generally best; good locations occur on the pressure side and also at some upstream suction-side locations; sensors in the laminar region may still be effective for spanwise-averaged flows but are less effective for Fourier modes and some mid-span cases; and the best locations are typically among \(y_1,y_4,y_5,y_6\), with \(y_5\) and \(y_6\) especially important [2507.18837].

## 5. Performance characteristics, validation, and limitations

The stable-system Wiener-Hopf formalism is validated on the linearized Ginzburg-Landau equation, where the causal estimator and controller produce identical gains to the classical Kalman/LQG solution when the assumptions reduce to white-noise forcing. In the backward-facing step example, the framework shows that downstream flow response is dominated by the Kelvin-Helmholtz mechanism while receptivity to incoming upstream perturbations is high-rank, so multiple sensors improve performance. The paper also notes that causal control is generally inferior to non-causal optimal control because causality forbids use of future information, and that performance degrades when disturbances become strongly nonlinear [2201.00361].

In the turbulent-wake application, accurate causal estimation of streamwise velocity is reported for spanwise-averaged, spanwise-Fourier-transformed, and mid-span flow using limited surface shear-stress measurements on the airfoil. The spanwise-averaged representation is the best estimated of the three. For the extended wake, the reported error is about \(0.254\) for the causal estimator and about \(0.352\) for the truncated noncausal estimator. The gain from causal enforcement is largest when the noncausal kernel contains important negative-time information; it diminishes when the kernel is already strongly convective and its main peak lies mostly at positive time. Accuracy decreases with increasing downstream distance, with increasing spanwise complexity, and in the presence of more random and smaller-scale fluctuations in the mid-span plane [2507.18837].

These findings delimit the method’s scope. In one branch of the literature, the formal theory assumes global stability of the linearized operator [2201.00361]. In another, empirical cross-spectral construction is used specifically to handle a not globally stable operator in a turbulent compressible wake [2507.18837]. The contrast is substantive rather than contradictory: it identifies two regimes of use, one based on direct stable resolvent manipulation and another based on data-driven replacement of unstable spectral objects.

## 6. Related uses of causal kernels and resolvent regularization

A neighboring usage appears in network neuroscience. GP CaKe models effective connectivity through causal convolution kernels \(c_{i\to j}(t)\) in the integro-differential system
$$
\mathcal{D}_j x_j(t)=\sum_{i=1}^{N}(c_{i\to j}\star x_i)(t)+w_j(t),
$$
with causality enforced by \(c_{i\to j}(t)=0\) for \(t<0\). The kernels are learned nonparametrically by Gaussian process regression in the frequency domain, using a causal covariance constructed from a spectral squared exponential kernel and its Hilbert transform,
$$
\mathfrak{K}_C(\zeta)= \mathfrak{K}_{SE}(\zeta)+ i\,\mathcal{H}\mathfrak{K}_{SE}(\zeta).
$$
The paper states that this construction yields a complex-valued GP over analytic functions, with temporal localization and causality encoded in the prior; it also notes an important limitation, namely that the additional smoothness factor slightly weakens strict causality by allowing small anti-causal leakage when \(\nu\) is not much larger than \(t_s\) [1705.05603].

A broader operator-theoretic connection appears in causal inference with RKHS methods. In kernel instrumental variable regression, the structural equation
$$
Y=h_0(X)+\varepsilon,\qquad \mathbb{E}[\varepsilon\mid Z]=0
$$
is turned into an ill-posed inverse problem involving the operator
$$
T=S^*S_z^{-1}S,
$$
with a resolvent-style regularized operator
$$
T_{\mu,\lambda}=S^*(S_z+\mu)^{-1}S+\lambda.
$$
The 2025 inference paper provides finite-sample, uniform confidence bands for KIV and a multiplier bootstrap that requires only one run of the KIV estimator [2511.21603]. In proximal causal learning with proxies, the bridge-function equation is a Fredholm integral equation of the first kind, and the PMMR estimator is written explicitly as
$$
h_\lambda=(T^*T+\lambda I)^{-1}T^*g.
$$
That paper presents both a two-stage kernel estimator and a maximum moment restriction estimator, together with consistency guarantees and convergence rates [2105.04544].

These adjacent literatures do not use “resolvent-based estimation kernels” in exactly the same sense as the flow-estimation papers. A plausible implication is that the phrase names a narrower family of causal Wiener-Hopf filters in the resolvent-analysis tradition, while also sitting within a larger class of kernelized inverse methods in which causality is encoded either by temporal support constraints or by causal structural equations, and resolvent language refers either to spectral transfer operators or to regularized inverses of compact operators.

Source: https://www.emergentmind.com/topics/causal-resolvent-based-estimation-kernels