Resolvent-Based Framework Overview

• Resolvent-based frameworks are methods that leverage resolvent operators to map inputs to outputs, supporting modal decomposition, estimation, and control in complex systems.
• They enable low-rank modeling and efficient estimation by combining linearized analysis with data-driven techniques to capture dominant physical mechanisms.
• The approach unifies applications from turbulent flow analysis to operator splitting in optimization, providing actionable insights into stability and system dynamics.

A resolvent-based framework is a family of analytical and computational methodologies organized around a resolvent operator. In one major usage, prevalent in fluid mechanics and dynamical-systems analysis, the resolvent is the input–output map obtained after linearizing governing equations about a mean or periodic base state, so that forcing is mapped to response through an operator such as

$$\hat{\boldsymbol q} = \mathbf R(\omega)\hat{\boldsymbol f}, \qquad \mathbf R(\omega) = (i\omega \mathbf I - \mathbf A)^{-1}.$$

In another major usage, prevalent in monotone operator theory and optimization, the resolvent is the proximal-type map

$J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$

Across recent work, these formulations support modal decomposition, low-rank statistical modeling, estimation and control, variational optimization, decentralized splitting algorithms, and contour-integral solution theories for fractional evolution equations (Liu et al., 2020, Tam, 2022, Wakrim, 6 Jan 2026).

1. Foundational formulations

In fluid and flow-control settings, the framework typically begins by linearizing the Navier–Stokes equations about a time-averaged or otherwise prescribed base flow and treating the remaining nonlinear terms as forcing. A representative formulation is

$\frac{\partial q'}{\partial t}=L(\bar q)q'+f,$

followed by a harmonic and spanwise-periodic ansatz,

$$q'(\boldsymbol x,t)=\hat q(x,y)e^{i(\beta z+\omega t)}, \qquad f(\boldsymbol x,t)=\hat f(x,y)e^{i(\beta z+\omega t)},$$

which yields the resolvent operator

$$\hat q_{\omega,\beta}=\left[i\omega I-L(\bar q,\beta)\right]^{-1}\hat f_{\omega,\beta} =H(\bar q;\omega,\beta)\hat f_{\omega,\beta}.$$

When the time-averaged base flow is unstable, discounted resolvent analysis introduces

$$\mathcal H(\bar q;\kappa,\omega,\beta)=[(\kappa+i\omega)I-L(\bar q;\beta)]^{-1},$$

with discount rate $\kappa$ set above the maximum linear growth rate (Liu et al., 2020).

In monotone operator theory, the resolvent is defined for a possibly set-valued operator $A:H\rightrightarrows H$ on a Hilbert space by

$J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$

For maximally monotone $A$, $J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$0 is single-valued, firmly nonexpansive, and has full domain. This definition is the primitive building block for proximal, Douglas–Rachford, forward–backward, and related splitting schemes (Artacho et al., 2020).

The same term also appears in spectral and contour-integral settings. In strongly correlated many-body systems, the starting point is the pole expansion of the Green-function resolvent

$J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$1

or its projected form

$J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$2

used to capture the global analytic structure rather than a local Taylor expansion (Huang et al., 1 Apr 2026). In rational-kernel fractional evolution equations with almost sectorial operators, the framework constructs contour-defined resolvent families

$J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$3

which support existence, uniqueness, and smoothing estimates beyond standard analytic semigroup theory (Wakrim, 6 Jan 2026).

2. Input–output analysis, modal structure, and physical mechanisms

The most characteristic operation in fluid-dynamical resolvent analysis is a singular value decomposition of the weighted resolvent. In compressible settings this is written as

$J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$4

where $J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$5 contains response modes, $J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$6 forcing modes, and $J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$7 the gains. The leading singular value $J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$8 quantifies the maximum linear amplification for a given $J_{\gamma A}:=(\mathrm{Id}+\gamma A)^{-1}.$9, and the associated low-rank structure is often interpreted physically in terms of dominant coherent mechanisms (Liu et al., 2020).

In turbulent jets, resolvent modes have been connected to modal shear instability and the nonmodal Orr mechanism. In “Resolvent-based modelling of coherent wavepackets in a turbulent jet,” the optimal forcing–response pair is associated with Kelvin–Helmholtz-type amplification, while suboptimal pairs separate out the Orr mechanism; a low-rank cross-spectral-density model based on resolvent modes showed close agreement with experimentally derived SPOD wavepackets around preferred-mode Strouhal numbers near $\frac{\partial q'}{\partial t}=L(\bar q)q'+f,$0 (Lesshafft et al., 2018).

The same framework has been adapted to modified operators. For turbulent flow over anisotropic permeable substrates, the effect of permeability enters through the Volume-Averaged Navier–Stokes equations and a generalized form of Darcy’s law, introducing direction-dependent damping through a permeability tensor. High streamwise permeability together with low spanwise permeability suppresses the gain of the resolvent mode used as a surrogate for the energetic near-wall cycle, while increasing wall-normal permeability predicts the emergence of spanwise-coherent rollers resembling Kelvin–Helmholtz vortices (Chavarin et al., 2020).

In fully developed Langmuir turbulence, scale-dependent resolvent analyses of the Craik–Leibovich equations with LES-derived mean state and eddy viscosity identify two-dimensional circulating rolls and three-dimensional turbulent coherent vortices as linearly amplified responses to sustained harmonic forcing. The integrated energy spectra predicted by the principal resolvent modes under broadband forcing capture dominant spanwise length scales consistent with LES data (Xuan et al., 20 Aug 2025).

Periodic base states require a different construction. The time-spectral resolvent method uses Fourier collocation in time and replaces harmonic-balance mappings between truncated Fourier coefficients by an operator acting directly on forcing and response envelopes sampled on a temporal grid,

$\frac{\partial q'}{\partial t}=L(\bar q)q'+f,$1

This formulation achieves spectral convergence and avoids computing Fourier coefficients of the base-flow Jacobian (Howell et al., 16 Feb 2026).

3. Estimation, sensing, and control

Resolvent-based estimation treats the linearized dynamics as a stochastic input–output system and uses forcing statistics to construct optimal linear estimators from limited measurements. In turbulent channel flow, the state-space model is written as

$\frac{\partial q'}{\partial t}=L(\bar q)q'+f,$2

and the optimal linear estimator for the forcing in frequency space is

$\frac{\partial q'}{\partial t}=L(\bar q)q'+f,$3

When the true space-time forcing statistics are used, the estimator is optimal in the Wiener filtering sense and provides an upper bound for the accuracy of any linear estimator from wall data. The same study showed that all models lead to accurate results up to the buffer layer, but only the true forcing statistics allow accurate estimation of large-scale log-layer structures throughout the channel (Amaral et al., 2020).

For turbulent airfoil wakes, causal resolvent-based estimation kernels can be constructed directly from cross-spectral densities: $\frac{\partial q'}{\partial t}=L(\bar q)q'+f,$4 and then causalized by a Wiener–Hopf decomposition. In the wake of a spanwise-periodic NACA0012 airfoil at Mach $\frac{\partial q'}{\partial t}=L(\bar q)q'+f,$5, Reynolds number $\frac{\partial q'}{\partial t}=L(\bar q)q'+f,$6, and angle of attack $\frac{\partial q'}{\partial t}=L(\bar q)q'+f,$7, the framework addresses global instability, multi-scale turbulence, and high dimensionality by using LES-derived statistics, causal kernels, and parallel algorithms. Using limited shear-stress measurements on the airfoil surface, the method accurately estimates streamwise velocity in spanwise-averaged, spanwise-Fourier-transformed, and mid-span representations; for the spanwise-averaged wake, the reported error is $\frac{\partial q'}{\partial t}=L(\bar q)q'+f,$8 using four sensors (Jung et al., 24 Jul 2025).

The same input–output logic extends naturally to control. For a laminar NACA 0012 airfoil wake at chord-based Reynolds number $\frac{\partial q'}{\partial t}=L(\bar q)q'+f,$9, Mach number $$q'(\boldsymbol x,t)=\hat q(x,y)e^{i(\beta z+\omega t)}, \qquad f(\boldsymbol x,t)=\hat f(x,y)e^{i(\beta z+\omega t)},$$0, and angle of attack $$q'(\boldsymbol x,t)=\hat q(x,y)e^{i(\beta z+\omega t)}, \qquad f(\boldsymbol x,t)=\hat f(x,y)e^{i(\beta z+\omega t)},$$1, the linearized system is augmented with actuator inputs and target outputs. Under equivalent assumptions, the resolvent-based estimator and controller reproduce the Kalman filter and LQG controller, respectively, but at substantially lower computational cost, and they can incorporate colored-in-time forcing statistics directly. Using four shear-stress sensors, the estimator predicts downstream targets with approximately $$q'(\boldsymbol x,t)=\hat q(x,y)e^{i(\beta z+\omega t)}, \qquad f(\boldsymbol x,t)=\hat f(x,y)e^{i(\beta z+\omega t)},$$2 error in clean freestream conditions and $$q'(\boldsymbol x,t)=\hat q(x,y)e^{i(\beta z+\omega t)}, \qquad f(\boldsymbol x,t)=\hat f(x,y)e^{i(\beta z+\omega t)},$$3 error in noisy freestream conditions; with four surface actuators, the controller reduces turbulent kinetic energy in the wake by $$q'(\boldsymbol x,t)=\hat q(x,y)e^{i(\beta z+\omega t)}, \qquad f(\boldsymbol x,t)=\hat f(x,y)e^{i(\beta z+\omega t)},$$4 in the noisy case (Jung et al., 2024).

A related development for incompressible turbulent channel flow is the Resolvent-informed White-noise-based Estimation method. Rather than using Resolvent-Based Estimation to infer absolute energy directly from distant measurements, it uses RBE to estimate the relative energy distribution near the wall, then modifies the white-noise forcing distribution by minimizing the norm of the CSD tensor of the near-wall error. The reported result is low sensitivity to $$q'(\boldsymbol x,t)=\hat q(x,y)e^{i(\beta z+\omega t)}, \qquad f(\boldsymbol x,t)=\hat f(x,y)e^{i(\beta z+\omega t)},$$5 and to measurement locations ranging from the near-wall region to the upper bound of the logarithmic region, together with high accuracy in predicting energy spectra (Ying et al., 2023).

4. Low-rank modeling, data-driven identification, and machine-learning hybrids

A central attraction of the framework is that many physically important systems are low-rank in the resolvent sense. In turbulent jet acoustics, LES realizations were projected onto a truncated acoustic resolvent basis, yielding a reduced cross-spectral-density matrix for the expansion coefficients. For round isothermal jets at Mach numbers $$q'(\boldsymbol x,t)=\hat q(x,y)e^{i(\beta z+\omega t)}, \qquad f(\boldsymbol x,t)=\hat f(x,y)e^{i(\beta z+\omega t)},$$6 and $$q'(\boldsymbol x,t)=\hat q(x,y)e^{i(\beta z+\omega t)}, \qquad f(\boldsymbol x,t)=\hat f(x,y)e^{i(\beta z+\omega t)},$$7, a single resolvent mode reconstructs the most energetic regions of the far-field acoustic field across $$q'(\boldsymbol x,t)=\hat q(x,y)e^{i(\beta z+\omega t)}, \qquad f(\boldsymbol x,t)=\hat f(x,y)e^{i(\beta z+\omega t)},$$8 and azimuthal wavenumbers $$q'(\boldsymbol x,t)=\hat q(x,y)e^{i(\beta z+\omega t)}, \qquad f(\boldsymbol x,t)=\hat f(x,y)e^{i(\beta z+\omega t)},$$9, and a simple rank-1 model agrees within $$\hat q_{\omega,\beta}=\left[i\omega I-L(\bar q,\beta)\right]^{-1}\hat f_{\omega,\beta} =H(\bar q;\omega,\beta)\hat f_{\omega,\beta}.$$0 dB of the peak noise for both jets (Pickering et al., 2021).

The low-rank interpretation of experimental statistics appears even more directly in SPOD-based comparisons. In the turbulent-jet wavepacket study, the response CSD was modeled as

$$\hat q_{\omega,\beta}=\left[i\omega I-L(\bar q,\beta)\right]^{-1}\hat f_{\omega,\beta} =H(\bar q;\omega,\beta)\hat f_{\omega,\beta}.$$1

using only a truncated set of resolvent response modes, and the leading SPOD mode extracted from two-point velocity measurements was found to agree closely with the leading coherent structure of the resolvent-based model in the preferred-mode range (Lesshafft et al., 2018).

Equation-free variants replace the explicit operator by a learned invariant subspace. “Data-driven resolvent analysis” uses exact DMD on transient snapshots to approximate eigenvalues and eigenvectors of the propagator, builds a projected resolvent

$$\hat q_{\omega,\beta}=\left[i\omega I-L(\bar q,\beta)\right]^{-1}\hat f_{\omega,\beta} =H(\bar q;\omega,\beta)\hat f_{\omega,\beta}.$$2

and obtains physical-space forcing and response modes by weighted SVD within the learned DMD basis. The method is equation-free and adjoint-free, but its recovery of resolvent structure depends on the richness of the transients: only dynamics that are sufficiently excited or observable in the data can be learned (Herrmann et al., 2020).

Hybrid physics–ML formulations keep the resolvent modes but learn the modal weights. In incompressible channel flow at friction Reynolds number $$\hat q_{\omega,\beta}=\left[i\omega I-L(\bar q,\beta)\right]^{-1}\hat f_{\omega,\beta} =H(\bar q;\omega,\beta)\hat f_{\omega,\beta}.$$3, the optimal resolvent response modes serve as basis functions while a convolutional neural network maps the mean streamwise velocity profile $$\hat q_{\omega,\beta}=\left[i\omega I-L(\bar q,\beta)\right]^{-1}\hat f_{\omega,\beta} =H(\bar q;\omega,\beta)\hat f_{\omega,\beta}.$$4 to the projection coefficients $$\hat q_{\omega,\beta}=\left[i\omega I-L(\bar q,\beta)\right]^{-1}\hat f_{\omega,\beta} =H(\bar q;\omega,\beta)\hat f_{\omega,\beta}.$$5. The reconstructed turbulent-energy distributions show close agreement with DNS; in the testing phase, the reported streamwise-energy error is $$\hat q_{\omega,\beta}=\left[i\omega I-L(\bar q,\beta)\right]^{-1}\hat f_{\omega,\beta} =H(\bar q;\omega,\beta)\hat f_{\omega,\beta}.$$6 when the loss fits only $$\hat q_{\omega,\beta}=\left[i\omega I-L(\bar q,\beta)\right]^{-1}\hat f_{\omega,\beta} =H(\bar q;\omega,\beta)\hat f_{\omega,\beta}.$$7, while simultaneous fitting of $$\hat q_{\omega,\beta}=\left[i\omega I-L(\bar q,\beta)\right]^{-1}\hat f_{\omega,\beta} =H(\bar q;\omega,\beta)\hat f_{\omega,\beta}.$$8 yields $$\hat q_{\omega,\beta}=\left[i\omega I-L(\bar q,\beta)\right]^{-1}\hat f_{\omega,\beta} =H(\bar q;\omega,\beta)\hat f_{\omega,\beta}.$$9, $$\mathcal H(\bar q;\kappa,\omega,\beta)=[(\kappa+i\omega)I-L(\bar q;\beta)]^{-1},$$0, $$\mathcal H(\bar q;\kappa,\omega,\beta)=[(\kappa+i\omega)I-L(\bar q;\beta)]^{-1},$$1, and $$\mathcal H(\bar q;\kappa,\omega,\beta)=[(\kappa+i\omega)I-L(\bar q;\beta)]^{-1},$$2 (Fan et al., 2024).

5. Variational optimization and statistical approximation

Resolvent modes have also been used as trial spaces in variational formulations. In wall-bounded invariant-flow computations, the Navier–Stokes equations are recast as minimization of the global residual

$$\mathcal H(\bar q;\kappa,\omega,\beta)=[(\kappa+i\omega)I-L(\bar q;\beta)]^{-1},$$3

over periodic or equilibrium fields. The state and residual are expanded in divergence-free, no-slip, orthonormal response modes obtained from resolvent analysis, which eliminates pressure-gradient terms after projection and provides a reduced-order representation. In rotating plane Couette flow, this construction recovers exact equilibrium and periodic solutions consistent with DNS, and the Hessian eigenvalues are directly linked to inverse-squared resolvent singular values,

$$\mathcal H(\bar q;\kappa,\omega,\beta)=[(\kappa+i\omega)I-L(\bar q;\beta)]^{-1},$$4

so truncation of low-gain modes functions as an optimal preconditioning strategy (Burton et al., 2 Sep 2025).

A related but statistically oriented use appears in chaotic dynamics. For the Lorenz 1963 system, a resolvent basis is used to reduce the search for long periodic trajectories to a low-dimensional optimization over modal coefficients minimizing the projected residual. The resulting quasi-trajectories need not be exact solutions of the governing equations, yet the paper reports that key observables, probability distributions, and spectra converge rapidly to values from long chaotic simulations. The authors explicitly argue that exact periodic solutions may not be necessary for approximating the system’s statistical behavior, because partially optimized trajectories can provide a sufficient “sketch” of the attractor (Burton et al., 2024).

In strongly correlated many-body systems, the framework is reorganized around the projected resolvent and an ETH-inspired statistical closure of cross terms. The projected resolvent equation

$$\mathcal H(\bar q;\kappa,\omega,\beta)=[(\kappa+i\omega)I-L(\bar q;\beta)]^{-1},$$5

is paired with an exact recursive re-expansion of cross-correlated terms beyond mean field, generating higher-order corrections that control distribution tails, branch splitting, and fluctuations. Practical closure is achieved through Lorentzian, Gaussian, and hybrid ansatzes for the overlap distribution, with the Lorentzian describing the bulk, the Gaussian the tails, and the hybrid form the full distribution (Huang et al., 1 Apr 2026).

6. Resolvent splitting, approximate resolvents, and abstract operator frameworks

In optimization and monotone inclusion theory, the phrase refers to frameworks that compute or exploit $$\mathcal H(\bar q;\kappa,\omega,\beta)=[(\kappa+i\omega)I-L(\bar q;\beta)]^{-1},$$6 without forming the resolvent of a sum explicitly. A foundational abstract line is the theory of approximate resolvents and Fejér-convergent algorithms. For a maximal monotone operator $$\mathcal H(\bar q;\kappa,\omega,\beta)=[(\kappa+i\omega)I-L(\bar q;\beta)]^{-1},$$7, the $$\mathcal H(\bar q;\kappa,\omega,\beta)=[(\kappa+i\omega)I-L(\bar q;\beta)]^{-1},$$8-approximate resolvent $$\mathcal H(\bar q;\kappa,\omega,\beta)=[(\kappa+i\omega)I-L(\bar q;\beta)]^{-1},$$9 is defined using the $\kappa$0-enlargement $\kappa$1 and a relative error criterion, and methods based on such approximate resolvents fall within a general class of Fejér-convergent schemes. This construction unifies Forward–Backward splitting, Tseng’s modified Forward–Backward method, Korpelevich’s method, and the Hybrid Proximal-Extragradient method (Svaiter, 2012).

A second line introduces “strengthening” of operators. For $\kappa$2, the $\kappa$3-strengthening is

$\kappa$4

and its resolvent is computable from the resolvent of the original operator: $\kappa$5 This enables the resolvent of a sum to be characterized through zeros of strengthened, shifted operators and yields generalized Douglas–Rachford, forward–backward–forward, and multi-operator splitting schemes (Artacho et al., 2020).

The general framework for frugal resolvent splittings pushes this further by characterizing fixed-point operators that use each resolvent once per iteration together with only vector addition and scalar multiplication. Under matrix conditions such as $\kappa$6 and $\kappa$7, the resulting map is $\kappa$8-averaged nonexpansive, which provides a unified convergence proof for Douglas–Rachford, Ryu splitting, and the Malitsky–Tam minimal lifting scheme. The same paper gives a new decentralized algorithm on regular networks, where local updates require only neighbor communications and one local resolvent evaluation per node per iteration (Tam, 2022).

Recent work has turned algorithm design itself into an optimization problem. A semidefinite-programming framework based on the Performance Estimation Problem constructs custom frugal resolvent splitting algorithms with prescribed communication structure and operator access patterns. Mixed-integer SDP variables encode the schedule of communication and computation, while continuous variables such as step size are tuned to minimize contraction factors or wall-clock convergence time (Bassett et al., 2024).

The abstract scope of the framework extends beyond splitting. In rational-kernel fractional evolution equations with almost sectorial operators, admissible kernel multipliers $\kappa$9 define contour-integral resolvent families, and the associated mild solutions satisfy

$A:H\rightrightarrows H$0

The framework unifies Atangana–Baleanu–Caputo and generalized $A:H\rightrightarrows H$1-operator dynamics, establishes existence and uniqueness of mild solutions, and proves fractional smoothing estimates

$A:H\rightrightarrows H$2

for almost sectorial generators (Wakrim, 6 Jan 2026).

7. Recurring themes, limits, and common misconceptions

A recurring misconception is that the largest local gain automatically determines the best control input. In supersonic turbulent cavity flow, the control design did not select the most amplified forcing mode at a single point; it selected the frequency–wavenumber pair that sustained the leading resolvent-based kinetic energy distribution over the entire cavity length, because sustained amplification was needed to prevent the formation of large spanwise vortices. The resulting leading-edge three-dimensional forcing achieved up to $A:H\rightrightarrows H$3 reduction in pressure root-mean-square level along the aft and bottom cavity walls relative to the baseline flow (Liu et al., 2020).

A second misconception is that white-in-time or white-in-space forcing is an adequate default closure. The channel-flow estimation study showed that true forcing statistics provide a theoretical upper bound for linear estimation accuracy and are essential for accurate recovery of large-scale log-layer structures (Amaral et al., 2020). The laminar-airfoil estimation and control study likewise emphasized that colored-in-time nonlinear forcing can be accommodated naturally and that doing so improves estimation accuracy and control efficacy relative to white-noise assumptions (Jung et al., 2024).

A third misconception is that resolvent methods are necessarily equation-based and tied to stationary base states. Data-driven resolvent analysis demonstrated a completely equation-free and adjoint-free route based on DMD and transient snapshots, while also making clear that insufficiently excited or unobservable directions cannot be reconstructed (Herrmann et al., 2020). The time-spectral resolvent method extended resolvent analysis to periodic base flows by operating directly in the time domain on a collocation grid and by handling quasi-periodic responses without assembling block-Toeplitz Fourier-domain operators (Howell et al., 16 Feb 2026).

The literature also identifies recurrent technical constraints. Data richness and observability limit equation-free reconstructions; forcing-color modeling limits linear estimators and controllers; and conditioning can dominate optimization-based variants. In the invariant-flow variational framework, convergence rates depend on the stability properties of the targeted solution, and conditioning degrades when eigenvalues of the linearized Navier–Stokes operator approach the imaginary axis. This suggests that the term “resolvent-based framework” is best understood not as a single algorithm, but as a methodological family in which the resolvent organizes amplification, statistical closure, and numerical conditioning in a common operator-theoretic language (Burton et al., 2 Sep 2025).