State and parameter estimation for a neural model of local field potentials Authors: Daniele Avitabile, Gabriel J Lord, Khadija Meddouni Published: 2026 ABSTRACT AbstractThe study of cortical dynamics during different states such as decision making, sleep and movement, is an important topic in Neuroscience. Modelling efforts aim to relate the neural rhythms present in cortical recordings to the underlying dynamics responsible for their emergence. We present an effort to capture certain features of the neural activity from the cortex of a mouse during natural sleep, obtained through local field potential measurements. Our approach relies on using a discretized Wilson–Cowan Amari neural field model for neural activity, along with a data assimilation method that allows the Bayesian joint estimation of the state and parameters. We demonstrate the feasibility of our approach on synthetic measurements before applying it to a dataset available in literature. Our findings suggest the potential of our approach to characterize the parameters of the external forcing in the neural activity model, while simultaneously inferring a state that aligns with the observed signal.Keywords: Wilson–Cowan–Amari model, Bayesian inference, Particle filters, Local field potential, Parameter estimation Introduction Data assimilation is a form of inverse problem that enhances model prediction using partial and noisy observations, and in which the unknown quantity is a dynamical system’s state, its control parameters, or a combination of these quantities. This approach to model forecast, parameter inference, and uncertainty quantification has been the subject of an intense work over the past decade, ranging from theoretical foundations to domain-specific studies in engineering, physics, and biology (Smith 2014; Law et al. 2015; Nakamura and Potthast 2015; Reich and Cotter 2015; Sanz-Alonso et al. 2023). The present paper tests and showcases the data assimilation methodology in the context of brain dynamics. Our long-term motivation is to develop a numerical tool to gain a better understanding of how cortical and limbic brain regions interact during sleep, when memories are thought to consolidate (Sanda et al. 2020). This is a compelling application for data assimilation, because experimental observations of brain activity are available for the cortex (Pedrosa et al. 2022), and one is interested in describing the coupling with hidden limbic regions. Thus it is interesting to model such interactions, even in a simple form, and test them against the partial observations of the system, in the upper (cortical) layers. Deterministic inverse problems in which the unknown quantity is a resting state or a static parameter typically suffer from ill-posedness (Curtis 2002; Kirsch 2011), which can be resolved by appealing to regularization techniques (see for example (Willoughby 1979; Kirsch 2011)) or to a Bayesian formulation as described in Joel and Franklin (1970); Stuart (2010); Kaipio and Somersalo (2005); Dashti and Stuart (2017). Within the data assimilation framework, in which data is time-dependent, we distinguish between two categories: methods such as 4D-Var where the state is estimated over a whole time window of observations, and sequential methods such as Kalman or particle filters which incorporate the observations sequentially. The latter is the approach we follow in this paper. Our aim is to examine Bayesian joint estimation of parameters and prediction of state from the local field potential measurements. Our approach relies on the use of the nested particle filter from Crisan and Míguez (2013); Miguez et al. (2023) on a nonlinear state-space system originating from the discretized deterministic Wilson–Cowan–Amari neural field originally proposed in Wilson and Cowan (1973); Shun-Ichi Amari (1977). This equation models heuristically spatially-dependent brain activity in a cortical domain using coarse activity variables, their nonlocal synaptic interaction, and the presence of an external forcing, which may come from other (cortical or non-cortical) regions, from an external stimulation, or compensate for missing biological terms. We henceforth refer to any of these sources as external forcing, and we stress that the model employed here is not sufficiently granular to distinguish between them. The present paper focuses on the numerical aspects of the inversion process, and our model choice has been dictated by mathematical structure and wide applicability, rather than biological accuracy. Although the dynamics of this model have been studied extensively (Ermentrout and Cowan 1979; Paul and Bressloff 2011; Coombes 2005; Kilpatrick and Bressloff 2010), there are far fewer works that tackle the inverse problem of estimating state or model parameters (Steven et al. 2007; Roland Potthast and Peter beim Graben 2009; Alswaihli et al. 2018). In the context of joint parameter and state estimation, related works include Kulikova and Kulikov (2023), in which the authors derived a continuous-discrete extended Kalman filter for simultaneous state and parameter estimation for stochastic dynamic neural field models and demonstrated the robustness of their approach on synthetic measurements. The authors of Steven et al. (2007) used the unscented Kalman filter to estimate both state and parameters of a spatiotemporal excitable cortical model and showed its ability to track spiral waves. In a subsequent work, Sauer and Schiff (2009) introduced the consensus set method with an ensemble Kalman filter applied to a homogeneous Wilson-Cowan model to effectively track the state and infer parameters in both synthetic and experimental measurements. Recently, state and parameter estimation for the model has been studied from a theoretical viewpoint and with synthetic data in Pieper-Sethmacher et al. (2025), in which data assimilation is introduced with a form of nudging, called guided process. The present paper differs from existing literature as it uses experimental observations with a nonlinear filter, which does not resort to linearised dynamics. Further, we test the performance of the spatio-temporal reconstruction against the dataset, and use synthetic data to show how the methodology accounts for epistemic uncertainty (model imperfections). While this paper is concerned mostly with the numerics of data assimilation, our results show the potential of this framework to quantify the influence of external forcing on the modelled cortical activity, while matching faithfully its experimental observations. The paper is organized as follows. In Sect. 2 we present the setting for the estimation problem of interest including a description of the cortical model and a background on Bayesian estimation. Section 3 provides the results of this approach on synthetic data. We present the experimental cortical data in Sect. 4 and the estimation results in Sect. 5. We conclude in Sect. 6 with a summary of this work, and we detail some additional computational experiments in the Appendices. Problem setting and background The neural field model We study the data assimilation problem on a discrete version of a one-population Amari neural field model with initial condition u(r,0)=u0(r), posed on a one-dimensional, bounded cortical domain D⊂R. The model describes the evolution of the population’s neural activity u at cortical position r and time t. The population has nonlinear firing rate σ, and typical response time τ, while w(r, s) models synaptic connections from point s to point r in the cortex. Further, the population is subject to a spatio-temporal external forcing Ξ, which depends upon a set of control parameters θ∈Rdθ that we wish to infer. The forcing may also possibly depend on other parameters a that we keep fixed. We note that the functions w and σ also depend on the parameters, but we omit this dependence for the sake of conciseness. We consider synaptic connections described by a Mexican hat kernel, which can model short-range excitation and long-range inhibition, and a sigmoidal firing rate function with threshold μ and slope λ In this paper, our aim is to characterize the external forcing Ξ by employing a sequential data assimilation method. To this end, we fix τ and all the parameters appearing in w and σ to the values given in Table 1; we are interested in inferring parameters for the forcing Ξ and simultaneously estimating a discrete approximation of the state variable u(·,t) from noisy observations. Table 1Model parameters used throughout this paper unless stated otherwiseParameterDescriptionValue(unit)τmembrane time constant of neurons0.03(s)ℓexspatial spread of of excitatory connections (short-range)0.5(mm)ℓinspatial spread of of inhibitory connections (long-range)1(mm)Aexexcitatory connections strength10(mV)Aininhibitory connections strength6(mV)λsteepness of the sigmoidal, controls how sensitive the firing rate is5(mV)-1μactivity threshold for neural firing0.5(mV) Model parameters used throughout this paper unless stated otherwise Strictly speaking, the model we use is a spatio-temporal discretisation of the continuum model (1), for specific choices of the spatio-temporal stimulus Ξ which are described later, in Sects. 3 and 5, as they change depending on the setup of the problem and numerical experiments. We highlight, however, that we do not quantify how spatial and temporal discretizations affect the data assimilation process, but we rather take the discrete model as the starting point for our investigation. The rest of this section is devoted to describing this discrete model. We take D=[0,L] and discretize space using J evenly spaced nodes r0,…,rJ-1∈D, with rj=jΔr for j=0,…,J-1. Further, we approximate the integral using the standard composite trapezium rule for the quadrature with nodes {rj} and weights {bj} given by b0=bJ-1=Δr/2, and bj=Δr otherwise. Let U(t)∈RJ be the vector approximating (u(r0,t),…,u(rJ-1,t))T, W∈RJ×J the matrix with components Wij=w(ri,rj)bj, S:RJ↦RJ is the multidimensional analogous of the firing rate σ, S(U(t)) and I(t;θ,a) are the vectors in RJ, such that S(U(t)) approximates (σ(u(r0,t)),…,σ(u(rJ-1)))T and I(t;θ,a)=(Ξ(r0,t;θ,a),…,Ξ(rJ-1,t;θ,a))T. This leads to the semi-discrete model As common in sequential data assimilation, we work with a discrete-time approximation to the dynamical system of Eq. 4. Here we integrate Eq. 4 in time using a standard explicit Euler method on a uniform grid of the temporal domain [0, T], hence for a given K∈N we let tk=kΔt with k=0,1,…,K and Δt=T/K, and consider the approximating sequence Uk≈U(tk) given by which we take as our model of cortical dynamics. Note that Ψk also depends on other non-inferred parameters such as a and the parameters appearing in the functions w and σ. For clarity of notation, we omit this dependence when no ambiguity arises. Background on particle filters and Bayesian estimation The data assimilation framework Data assimilation combines uncertain estimations from a model with noisy (and sometimes partial) observations to provide an estimate of an unobserved (hidden) dynamical state. Consider two stochastic processes {Xk}k=0∞⊂Rdx and {Yk}k=1∞⊂Rdy, with realisations xk, yk, which describe the hidden state and measurements, respectively. We adopt the standard notation xk:l=xk,xk+1,…,xl and similarly for yk:l. At time step k≥0, Xk∈Rdx represents the hidden state to be estimated and for k≥1, Yk∈Rdy represents the measurements. They are assumed to satisfy The state Eq. (6) describes how the hidden state Xk evolves over time and the observation Eq. (7) links the hidden state to the measurement Yk. The model function Fk:Rdx×Rdθ×Rdx→Rdx and observation function Gk:Rdx×Rdy→Rdy are assumed to be known. The process noise, ηk∈Rdx, represents the model uncertainty and is assumed to have a known probability distribution. Similarly γk∈Rdy is the observation noise, also with a known distribution. We assume that for k≠l, the process (observation) noise vectors ηk, ηl (γk, γl) are mutually independent and are mutually independent of the initial state X0. Further, for all k, l, ηk and γl are mutually independent. Equations (6) and (7) form a state-space system. While the function Fk depends on the control parameters θ, we may omit this dependence wherever possible from now on. In this paper, we assume additive noise in the state and observation equations, and use a linear observation operator for the state variables of Eq. (5), that is, we consider for k=1,…,K, where dx=dy=J, G∈RJ×J is an observation matrix that is specified later for each experiment, γk,ηk∼N(0,IdJ) are independent and identically distributed normal random variables with values in RJ, and Q,R∈R+ are the noise intensities. With a slight abuse of notation we use π(x) and π(y) to denote the probability densities of X and Y respectively, even though they are different functions (the argument is used to distinguish them) and assume that the probability distributions of Xk and Yk are absolutely continuous with respect to the Lebesgue measure (μXk(dxk)=π(xk)dxk and similarly for Yk). It is standard to assume {Xk}k=0∞ is a Markovian process and depends on the past observations y1:k only through its own history Similarly {Yk}k=1∞ is assumed to be a Markovian process with respect to the history of {Xk}k=0∞, such that Bayesian filtering In the Bayesian setting, the sequential filtering problem is formulated as determining π(xk|y1:k) the posterior probability density of the state Xk at time step k. Bayesian filtering methods operate sequentially by incorporating each new observation using Bayes’ formula. At timestep k+1, the posterior density of the state from the previous timestep π(xk|y1:k) is evolved in time through the state Eq. (6), leading to the prediction density π(xk+1|y1:k), so This prediction is then taken as a prior in the Bayes’ formula and is updated or corrected with the new observation yk+1 using the likelihood π(yk+1|xk+1) derived from the observation Eq. (7), that is where When the state-space model is linear and Gaussian, the densities in (12) and (13) can be computed exactly leading to the Kalman filter (Kalman 1960). In the presence of non-linearities, several approximations exist such as the extended Kalman filter (Jazwinski 1970; Gelb 1974) or the unscented Kalman filter (Wan and Van Der Merwe 2000). Particle filters (Neil 1993; Doucet et al. 2001) are an alternative approach and are a class of filters that do not require any additional assumptions of linearity or Gaussianity of the state-space model. They rely only on Monte Carlo methods and importance sampling to approximate the posterior density through an ensemble of random samples, referred to as particles. In practice, particle filters implement the Bayesian filtering Eqs. (12) and (13): they rely on having an ensemble of M particles {xk1,⋯,xkM} (positions in the state space at time k) along with their weights {ωk1,⋯,ωkM} (the probability that a particle represents the true position) as an approximation to the density π(xk|y1:k). The aim is to first produce an ensemble of particles to approximate the prediction measure π(xk+1|y1:k) and then another ensemble for the posterior π(xk+1|y1:k+1) with updated weights that reflect how the particles match the observations. The first step uses the state Eq. (6) to approximate Eq. (12), while the second makes use of the prior distribution as an importance sampling distribution in order to make the update formula of the weights in Eq. (13) sequential. A direct implementation will run into a problem of degeneracy: as time progresses, only few particles will have significant weight while the majority of the particles end up with extremely low weights. This implies that only a few particles contribute significantly to the approximation of the posterior distribution, effectively reducing the ensemble diversity and the quality of the estimation (since not all regions of the posterior are sampled). In fact, at time step k the posterior distribution is approximated through a weighted empirical measure where δ is the Dirac delta. From which we have the estimate E[g(x)|y1:k]≈∑m=1Mωkmg(xkm), for any integrable function of interest g (e.g. the posterior mean or variance). The quality of such estimates relies on the particles covering the relevant regions of the posterior and on their weights corresponding to their likelihood. This fails when the ensemble is composed of few particles with non-negligible weights. Indeed, in the worst case, the ensemble may collapse to a single particle with non-negligible weight. To counter this, a resampling step is normally added, so that particles that do not match the observation (and thus have extremely low weight) are thrown away and replaced with duplicates of the particles with higher probability. In practice, we resample our ensemble of particles each time the effective number of particles Meff, approximated by Meff:=1/∑m=1M(ωkm)2, falls below a threshold of M/2. Meff can be interpreted as the number of independent, equally-weighted samples that would give the same quality approximation. A high Meff indicates a diverse particle ensemble, for example, when all the particles are uniformly weighted with 1/M, Meff=M. On the other hand, when all weights are zero except for a single particle with weight one, Meff=1 indicating that the entire posterior is represented by a single particle, regardless of how many particles were nominally propagated. The resampling step consists in drawing the particles with replacement according to their normalized weight. Based on the findings of Douc and Cappe (2005); Kuptametee and Aunsri (2022), we use stratified resampling in our numerical simulation as given in Kitagawa (1996). Algorithm 1 presents a straightforward implementation of this instance of particle filters, known as the sequential importance resampling (SIR) particle filter. SIR Nested particle filters In the previous sections, it was implicitly assumed that the model function Fk is fully known and depends only on the state itself. However, in most applications, the exact values of the model parameters θ are not known and need to be estimated along with the state. One approach to perform a joint state and parameter estimation is the nested particle filter presented in Crisan and Míguez (2013); Miguez et al. (2023), which we adopt here. This nested filter is a recursive algorithm that can be described as two layers of particle filters. In the outer layer, N particles for the parameters are generated, and for each particle, an inner filter from Algorithm 1 is used to estimate the state with M particles for each of the N parameter particles. For the sake of completeness, we present the method in Algorithm 2. The jittering or rejuvenation step of the algorithm consists in drawing new parameter samples from a kernel to avoid the collapse of the ensemble of particles for the parameters. The new sample θ¯ki must be close enough to the previous sample θk-1i, so that the filter approximation of xk-1 computed for θk-1i can be used as a particle approximation of the filter for the new sample θ¯ki. Following Crisan and Míguez (2013); Miguez et al. (2023), we use the following kernel where ϵN=1N is the probability of jittering a particle, TN is the truncated normal distribution with mean θ, variance σ2θN∝N-3/2 and a support [Aθ,Bθ] that varies for each component of the parameter vector θ. We consider the empirical posterior means as our estimates at time step k of the parameters and state respectively. Nested Particle Filter Nested particle filter applied to synthetic data Before using the particle filter on cortical data, we test the data-assimilation algorithm on synthetic data. We consider the states {Uk∗}k=0∗⊂R∗ generated by model (5), in which the mapping Ψk∗(U,θ∗) is specified via a matrix W∗∈RJ∗×J∗ and input I∗(t;θ∗,a∗)∈R∗. The input features 3 inferable parameters θ∗=(A∗,ν∗,f∗) as well as an additional parameter a∗, and has components Ξ(rj,t,θ∗,a∗)j=0J∗-1 at time t and spatial nodes rj=jΔr where The external input models a chirp travelling wave with amplitude A∗ (mV), spatial frequency ν∗ (mm-1), and time-dependent frequency f(t)=(f∗+a∗t), hence f∗ (Hz) is the initial temporal frequency and a∗ (Hz/s) is the chirp rate. The data-assimilation task aims to reconstruct {Uk∗} and infer θ∗ under noisy observations, as well as (epistemic) model uncertainty. To this end, we setup system (8) as follows: Noisy observations. The observations {Yk}k=1K⊂RJ are obtained by sampling the original signal {Uk∗}k=0∗⊂R∗ on a coarse spatio-temporal grid (K0.9}; k90 number of modes accounting for 90% of energy, calculated as k90=min{k:∑i=1kσi2∑i=1Jσi2>0.9}; R1 roughness of the first spatial mode, given by R1=‖∇Φ1‖22; R1 roughness of the first spatial mode, given by R1=‖∇Φ1‖22; T1 period of the first temporal mode, given by the inverse of the dominant frequency T1=1/argmaxf|Ψ^1|2. T1 period of the first temporal mode, given by the inverse of the dominant frequency T1=1/argmaxf|Ψ^1|2. We use these different quantities of interest (rather than a single one) to examine different aspects of the model dynamics (energy, frequencies, ...), which yields a global picture of the model’s sensitivity structure. Third stage, we compute the first order S1 and total order ST Sobol indices of the parameters A, ν and f. S1 and ST measure the individual and total contribution of each input (parameter) to the output (QOI) variance respectively. We report their values in Table 3. Table 3First order (S1) and total order (ST) Sobol indices for parameters A, ν and f drawn from [0,10]×[0,1]×[0,1], and for scalar quantities of interest η1, k90, R1 and T1η1k90R1T1STS1STS1STS1STS1A0.7840.4050.8050.10.3680.0760.7480.355ν0.1610.0630.3030.0630.9030.4930.2980.082f0.4430.0650.8230.0610.2820.0280.4790.127 First order (S1) and total order (ST) Sobol indices for parameters A, ν and f drawn from [0,10]×[0,1]×[0,1], and for scalar quantities of interest η1, k90, R1 and T1 The three parameters exhibit moderate to high total indices ST and gap ST-S1 across all QOIs, which indicates that each parameter is influential mainly through its interactions with other parameters and not on its own. A has large ST and moderate S1 for all QOIs except R1, which indicates that this parameter is very influential, mainly through its interactions with the other parameters. ν primarily controls R1 both directly and through interactions, and influences the other QOIs mainly through interactions as its S1 is negligible. f appears to have almost no independent effect as its S1 is negligible compared to ST for all QOIs. For the parameter inference task, this suggests that the observations can constrain combinations of the parameters, not each separately. However considering a smaller range for parameters, around the true values from the experiments of Sect. 3 (A=1,ν=0.1,f∈[0,0.5]), paints a different picture for the parameter f in particular. We calculate total and first order Sobol indices for parameters in the range θ=(A,ν,f)∈[0.5,1.5]×[0.05,0.5]×[0,1], while keeping the same setting as before. The values are reported in Table 4. Table 4First order (S1) and total order (ST) Sobol indices for parameters A, ν and f drawn from [0.5,1.5]×[0.05,0.5]×[0,1], and for scalar quantities of interest η1, k90, R1 and T1η1k90R1T1STS1STS1STS1STS1A0.1820.0660.3090.0180.3440.0180.3580.060ν0.0440.01450.1620.0270.5740.0360.7120.365f0.9360.8241.0380.5950.9970.2390.6280.260 First order (S1) and total order (ST) Sobol indices for parameters A, ν and f drawn from [0.5,1.5]×[0.05,0.5]×[0,1], and for scalar quantities of interest η1, k90, R1 and T1 In this restricted range, the three parameters exhibit different behaviours: A has negligible first order indices S1 across all quantities of interest, but moderate total indices ST. ν has negligible S1 and ST except for R1 where it influences mainly through interactions, and T1 for which it has high S1 and ST; finally, f seems to be dominant for all four quantities of interest through its interactions as well as directly. This exploration of the sensitivity near the true values of the parameters suggests that once the particles approach this restricted region in the parameters space, inference of the parameters should be easier, since the influence of the parameters is more direct. Appendix B Robustness B.1 Discretization parameters While an extensive study of the effect of discretization schemes on inference is beyond the scope of this paper, we present some evidence that our results are robust to changes in the temporal and spatial discretization parameters Δt=0.01 and J=30 used in Sect. 3. We only require that these steps permit us to resolve the frequencies present in the observations. For this purpose, we conduct the same experiment as described in Sect. 3 while only changing Δt and J. In Fig. 14, we show the effect of using a Δt that is double the one used in Sect. 3. We observe that the posterior mean estimates of A and ν still converge to the true value, while f exhibits a similar behaviour to Fig. 1c for the first 60s, after which its deviations from the truth are higher than in Fig. 1c. In Fig. 15, where Δt is half the one used in Sect. 3, the estimates of A and ν behave the same as in Fig. 1a and b. However, the estimate of f for this smaller Δt seems to track more closely the temporal frequency variations exhibited by the state in Fig. 2c. Figure 16, which corresponds to a smaller state dimension J (a bigger spatial discretisation step Δr), shows slower convergence for A and ν with more fluctuations around their true values, while the estimate for f manages to remain relatively close to the truth until around 60s after which it diverges. In Fig. 17, increasing J (smaller Δr) yields posterior mean estimates of A and ν that converge almost instantly with small and stable variance, and a behaviour similar to Fig. 1c for f. Across these tests, the posterior mean estimates of A and ν remain stable across varying and reasonable discretisation resolutions. As for f, the discretisation steps must be sufficiently small to ensure that the numerical integration of the evolution model captures the observed dynamics of this parameter. J=30,Δt=0.02 J=30,Δt=0.005 J=15,Δt=0.01 J=45,Δt=0.01 B.2 Noise intensities The performance of particle filters is highly sensitive to process and observation noise intensities, though the two play different roles. In fact, Q governs the spread of the ensemble of particles in the prediction step. If Q is too small, the particles remain tightly clustered and fail to cover the region of the space where the true state is, which leads to an impoverished sample after resampling. If Q is too large, the ensemble will be overly spread and therefore the filter estimates lose precision. R controls the sharpness of the likelihood and therefore the weighting of the particles. If R is too small, non-negligible weights will be assigned to a handful of particles which will then be duplicated while discarding the rest of the ensemble through resampling, leading to a loss of diversity. On the other hand, a very large R will result in all particles receiving similar weights regardless of their proximity to the observation, making the later non informative. This misspecification of the noise intensities for the inner layer of the nested particle filter can directly distort the outer weights of the parameters, since these are calculated from the state weights (Algorithm 2, line 13). Practically, this can manifest in pathological behaviours such as weight collapse or an invariant particle ensemble that does not update with new observations. To illustrate this, we run the same experiment as Sect. 3 for different Q and R pairs while eliminating the model mismatch aspect. That is, we set a∗ to 0 for the process generating observations. Figures 18, 19 and 20 show some examples of the effect of misspecified Q and R on the parameter inference. For Fig. 18, after an initial period where the estimates are near the true values, the filter diverges and the variance blows up. In Fig. 19, the ensemble is not moving and particle weights are not updated after each new observation. As a result, the estimates are stuck at the mean of the prior with a variance that does not shrink as new information accumulates. As for Fig. 20, the filter settles near the wrong values with almost zero variance, indicating a possible ensemble collapse to few particles. q=0.1,R=0.1 q=0.5, R=0.1 q=1,R=0.5 B.3 Prior range The choice of the prior range for the parameters and the jittering kernel is also critical. A prior that does not include the true value, or has insufficient initial coverage near it will result in a failed recovery of the parameters. By (14), the parameter particles are constrained to vary in the range dictated by the prior ( which is also the support of the jittering kernel), they cannot escape this range regardless of what the observations may suggest. If the true value is not included in the initial range, the filter will not recover it and will produce an incorrect estimate. This failure can manifest itself as the particles, and by consequence the posterior mean, clustering near the boundary of the prior range closest to the true value. This can have further consequences for the estimation of the parameters: the filter may partially compensate for a misplaced parameter by adjusting the estimation of one or more other parameters in an effort to make the predicted state close to the observations. To illustrate this, we repeat the same experiment as Sect. 3 while changing only the prior range of A to from U([0,10]) to U([2,10]) which excludes the true value A∗=1. Figure 21 shows that while we successfully recover ν, the posterior mean of A stalls near the boundary of its prior while f varies to absorb the mismatch between the state and observations due to the error in A. A plausible choice of the prior should then be guided by physical knowledge of the parameter (known bounds, typical magnitude, etc.) and the characteristics of the observations and how they are related to the parameters. For the same numerical experiment as Fig. 1 except for changing the prior of A to U([0,2]) which does not include the true value A∗=1, we show the sequence of empirical posterior means (blue line) within 2σ band (shaded region) compared to their true values (red line) (color figure online) B.4 Tukey window The tapering fraction β of the Tukey window used as observation operator in the experiments of Sect. 5, controls how much of the signal is preserved without modification and how much is subject to decay at the edges. In order to evaluate the effect of β on the dynamics of the estimated parameters and quality of state estimation, we run experiments in the same setting described in Sect. 5 for angle α=60∘ for different values of the tapering fraction β=0.3,0.5,0.7. Figure 22a–c show the posterior means of the inferred input parameters with added/ subtracted offset to separate them, while Fig. 22d–f show them overlaid. These figures show that the dynamics of the mean posteriors are globally conserved through the change of β, with the amplitude parameter A slightly decreasing with β and f showing some slight variations between the different values of β. We report the aggregated RMSE and frequency deviation Δft for the three cases in Table 5. While Δft remains almost unchanged, the RMSE for β=0.5 and 0.7 is lower which indicates these values are more suitable for our application. Inferred parameters for α=60∘ for three tapering fractions of the Tukey window. The top row shows the trajectories with an offset Table 5RMSE and Δft for different tapering fractions β of the Tukey window for α=60∘RMSEΔftβ=0.32.25650.0150β=0.52.07120.0156β=0.72.05540.0141 RMSE and Δft for different tapering fractions β of the Tukey window for α=60∘ Data availability Imaging data used in this work have been deposited in the Donders Repository https://doi.org/10.34973/2w2s-tg07. Code availability A repository with codes is available at https://doi.org/10.5281/zenodo.21795156. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Daniele Avitabile, Gabriel J. Lord and Khadija Meddouni have contributed equally to this work. References Associated Data Data Availability Statement Imaging data used in this work have been deposited in the Donders Repository https://doi.org/10.34973/2w2s-tg07. A repository with codes is available at https://doi.org/10.5281/zenodo.21795156.