betaFeedback

WorksMathematics

Parameter Recovery for the Modified Leray-alpha Model via Continuous Data Assimilation

Abstract

In this study, we extend the recursive length-scale recovery method for the simplified Bardina model to the three-dimensional modified Leray-alpha model. The main difficulty is that the unknown parameter also occurs in the nonlinear convection term, so estimating the error in the parameter update requires control of second spatial derivatives of the velocity error. We combine continuous data assimilation with a recursive update of the squared length scale, using observed low Fourier modes and computed high Fourier modes. Under uniform regularity bounds on the reference solution and suitable conditions on the observations and the recovery procedure, we prove geometric convergence of both the parameter and velocity errors at the update times. When the update-window lengths are uniformly bounded, this convergence is exponential in time at the update times.

Introduction

The Navier–Stokes equations describe the motion of viscous incompressible fluids. In turbulent regimes, however, the range of dynamically relevant scales makes direct numerical simulation expensive. Regularized fluid models provide a way to describe the large-scale motion without resolving every small scale. Beginning with Leray’s regularization [23], this approach has led to several families of turbulence models, including the Navier–Stokes-α\alpha, Leray-α\alpha, modified Leray-α\alpha, Clark-α\alpha, and simplified Bardina models [7, 9, 10, 14, 18, 20, 22]. The parameter α\alpha specifies a filtering length, and its choice influences the relation between the resolved velocity and the underlying regularized flow. Recovering this parameter from observations is therefore a natural problem in model calibration.

In this paper, we consider the three-dimensional modified Leray-α\alpha model introduced in [20]. On the periodic box Ω=[0,LΩ]3\Omega= [0,L_\Omega]^3, its equations are

{vt−νΔv+(v⋅∇)u+∇p=f,∇⋅u=∇⋅v=0,v=u−α2Δu.(1)\left\{\begin{aligned} v_t - \nu\Delta v + (v \cdot\nabla)u + \nabla p &= f,\\ \nabla\cdot u = \nabla\cdot v &= 0,\\ v &= u - \alpha^2\Delta u. \end{aligned}\right. \tag*{(1)}

Here uu is the filtered velocity, vv is the associated unfiltered velocity, pp is the pressure, ν>0\nu> 0 is the viscosity, and ff is the external force. We impose periodic boundary conditions and assume

u(⋅,0)=u0,∇⋅f=0,∫Ωu0(x) dx=∫Ωf(x,t) dx=0,t≥0.u(\cdot,0)=u_0,\qquad\nabla\cdot f=0,\qquad\int_{\Omega}u_0(x)\,dx=\int_{\Omega}f(x,t)\,dx=0,\qquad t\geq0.

The force and viscosity are known, whereas the initial state and the parameter α\alpha are unknown. Only a finite number of low Fourier modes of uu are observed. We assume that known constants α0\alpha_0 and α1\alpha_1 satisfy

0<α0≤α≤α1.0<\alpha_0\leq\alpha\leq\alpha_1.

The theory of determining modes and nodes shows that sufficiently resolved partial observations determine the asymptotic behavior of the full solution under model-specific hypotheses [16, 17, 21]. Continuous data assimilation incorporates these observations into an evolution equation through a feedback term. We use the framework introduced by Azouani, Olson, and Titi [6], whose applications to turbulence models include [1, 2]. When some parameters are also unknown, the feedback system can be coupled with a parameter-recovery procedure. Such methods have been studied for viscosity recovery and sensitivity analysis [8, 11, 12, 24], identification of forcing [15], and simultaneous recovery of multiple parameters [13, 28]. Related developments include on-the-fly model discovery and relaxation-based estimation [25, 27], viscosity recovery for the steady Navier–Stokes equations [29], recovery for the primitive equations with a magnetic field [32], and data-driven operator inference [31].

For α\alpha-regularization models, the effect of a mismatch in the filtering length has been investigated in [3, 4]. A recursive recovery procedure for the simplified Bardina model was subsequently developed in [5]. We adapt that procedure to recover α\alpha and approximate the reference solution of the modified Leray-α\alpha equation. Unlike the simplified Bardina model, whose convection term is (u⋅∇)u(u\cdot\nabla)u, (1) contains (v⋅∇)u(v\cdot\nabla)u. When the system is written in terms of uu, the unknown parameter α\alpha therefore occurs not only in the time-derivative and viscous terms, but also in the convection term. To construct the approximate solution, let PNP_N denote the orthogonal projection onto the spatial Fourier modes indexed by k∈Z3k\in\mathbb{Z}^3 with 0<∣k∣≤N0<\lvert k\rvert\leq N. For a fixed approximate parameter β∈[α0,α1]\beta\in[\alpha_0,\alpha_1], we use the same feedback system as in [4] with Fourier-mode observations:

zt−νΔz+(z⋅∇)w+∇q=f−η(I−β2Δ)PN(w−u),∇⋅w=∇⋅z=0,z=w−β2Δw.(2)\begin{aligned} z_t-\nu\Delta z+(z\cdot\nabla)w+\nabla q &= f-\eta(I-\beta^2\Delta)P_N(w-u),\\ \nabla\cdot w&=\nabla\cdot z=0,\\ z&=w-\beta^2\Delta w. \tag*{(2)} \end{aligned}

The fields ww and zz satisfy the same periodic and zero-mean conditions as uu and vv; qq is the corresponding pressure, II is the identity, and η>0\eta>0 is the nudging strength. The initial value of ww is arbitrary in the energy space defined below.

The parameter update depends on the unknown velocity, whereas only its low Fourier modes are observed. Following [5], we replace the unknown velocity with a field formed from the observed low Fourier modes and the high Fourier modes of the approximate solution. For the modified Leray-α\alpha model, this replacement introduces an error involving second spatial derivatives, which cannot be controlled by a basic velocity error estimate alone. We combine a synchronization estimate for the velocity error with a time-integrated higher-order dissipation estimate to control the second spatial derivatives appearing in the parameter update. These estimates are then used to construct a recursive parameter update. Before each update window, we keep the parameter fixed for a relaxation period to allow the assimilated solution to approach the reference solution. Under suitable regularity assumptions on the reference solution and sufficient observational signal relative to the approximation error, we prove that both the estimated parameter and the approximate velocity converge geometrically to their corresponding true values at the update times. Furthermore, if the update intervals are uniformly bounded in length, this convergence is exponential in time at the update times.

The rest of the paper is organized as follows. § 2 presents the functional setting. § 3 introduces the parameter recovery procedure. § 4 establishes the estimates for the reference and data-assimilated systems. § 5 presents and proves the main theorem. § 6 concludes the paper.

Functional analytic framework

For Ω=[0,LΩ]3\Omega= [0,L_{\Omega}]^3, we denote by Lp(Ω)L^p(\Omega) the usual Lebesgue spaces of three-dimensional vector fields. For each s∈Rs \in\mathbb{R}, we define the Hilbert space

V˙s={ϕ=∑k∈Z3∖{0}ϕ^ke2πik⋅x/LΩ:ϕ^k∈C3, ϕ^−k=ϕ^k‾, k⋅ϕ^k=0, ∑k∈Z3∖{0}(2π∣k∣LΩ)2s∣ϕ^k∣2<∞}.\dot{V}_s = \left\{ \phi= \sum_{k\in\mathbb{Z}^3\setminus\{0\}} \widehat{\phi}_k e^{2\pi i k\cdot x/L_{\Omega}} : \widehat{\phi}_k \in\mathbb{C}^3,\ \widehat{\phi}_{-k}=\overline{\widehat{\phi}_k},\ k\cdot\widehat{\phi}_k=0,\ \sum_{k\in\mathbb{Z}^3\setminus\{0\}}\left(\frac{2\pi|k|}{L_{\Omega}}\right)^{2s}|\widehat{\phi}_k|^2<\infty\right\}.

where ϕ^k\widehat{\phi}_k are the Fourier coefficients of ϕ\phi, and the series is understood in the sense of distributions. We endow this space with the norm

∥ϕ∥V˙s2=∣Ω∣∑k≠0(2π∣k∣LΩ)2s∣ϕ^k∣2.\|\phi\|_{\dot{V}_s}^2 = |\Omega|\sum_{k\ne0}\left(\frac{2\pi|k|}{L_{\Omega}}\right)^{2s}|\widehat{\phi}_k|^2.

So ∥ϕ∥V˙0=∥ϕ∥L2\|\phi\|_{\dot{V}_0}=\|\phi\|_{L^2}. We write

H=V˙0,V=V˙1,V′=V˙−1,D′=V˙−2,V˙−s=(V˙s)′(s≥0).H=\dot{V}_0,\qquad V=\dot{V}_1,\qquad V'=\dot{V}_{-1},\qquad D'=\dot{V}_{-2},\qquad\dot{V}_{-s}=(\dot{V}_s)' \quad(s\ge0).

We denote by P:L2(Ω)→H\mathcal{P}:L^2(\Omega)\to H the Helmholtz–Leray orthogonal projection, given by

Pϕ=∑k∈Z3∖{0}(ϕ^k−k(k⋅ϕ^k)∣k∣2)e2πik⋅x/LΩ.\mathcal{P}\phi=\sum_{k\in\mathbb{Z}^3\setminus\{0\}}\left(\widehat{\phi}_k-\frac{k(k\cdot\widehat{\phi}_k)}{|k|^2}\right)e^{2\pi i k\cdot x/L_{\Omega}}.

The same formula defines its extension to periodic distributions. For each s∈Rs\in\mathbb{R}, the Stokes operator A:V˙s→V˙s−2A:\dot{V}_s\to\dot{V}_{s-2} is given by

Aϕ=∑k∈Z3∖{0}4π2∣k∣2LΩ2ϕ^ke2πik⋅x/LΩ.A\phi=\sum_{k\in\mathbb{Z}^3\setminus\{0\}}\frac{4\pi^2|k|^2}{L_{\Omega}^2}\widehat{\phi}_k e^{2\pi i k\cdot x/L_{\Omega}}.

In particular,

Aϕ=−PΔϕ=−Δϕ(ϕ∈D(A)=V˙2),λ1=4π2LΩ2.A\phi=-\mathcal{P}\Delta\phi=-\Delta\phi\quad(\phi\in D(A)=\dot{V}_2),\qquad\lambda_1=\frac{4\pi^2}{L_{\Omega}^2}.

For the inner product and norms, we use the notation

(ϕ,ψ)=(ϕ,ψ)L2,∥ϕ∥=∥ϕ∥L2,∥ϕ∥V=∥∇ϕ∥.(\phi,\psi)=(\phi,\psi)_{L^2},\qquad\|\phi\|=\|\phi\|_{L^2},\qquad\|\phi\|_V=\|\nabla\phi\|.

Thus round parentheses denote the L2L^2 inner product, whereas ⟨⋅,⋅⟩\langle\cdot,\cdot\rangle denotes the appropriate Sobolev duality pairing, which extends this inner product. The zero-mean condition gives the Poincaré inequalities

∥u∥2≤λ1−1∥∇u∥2(u∈V),∥∇u∥2≤λ1−1∥Au∥2(u∈D(A)).(3)\|u\|^2\le\lambda_1^{-1}\|\nabla u\|^2\quad(u\in V),\qquad\|\nabla u\|^2\le\lambda_1^{-1}\|Au\|^2\quad(u\in D(A)). \tag*{(3)}

We recall the following particular cases of the Gagliardo–Nirenberg inequalities [19, 30]:

∥g∥L3≤C∥g∥1/2∥∇g∥1/2,∥g∥L4≤C∥g∥1/4∥∇g∥3/4,∥g∥L6≤C∥∇g∥.\|g\|_{L^3} \le C\|g\|^{1/2}\|\nabla g\|^{1/2}, \qquad\|g\|_{L^4} \le C\|g\|^{1/4}\|\nabla g\|^{3/4}, \qquad\|g\|_{L^6} \le C\|\nabla g\|.

These inequalities hold for every g∈Vg \in V. We shall also use the Sobolev embeddings V˙2↪L∞\dot{V}_2 \hookrightarrow L^\infty and V˙3↪W1,∞\dot{V}_3 \hookrightarrow W^{1,\infty}.

For each ϑ>0\vartheta> 0, we denote the Helmholtz operator by Tϑ=I+ϑ2AT_\vartheta= I + \vartheta^2 A. Its inverse is the Fourier multiplier operator with multiplier (1+ϑ2λ1∣k∣2)−1(1+\vartheta^2\lambda_1|k|^2)^{-1}, and the following estimates hold:

∥Tϑ−1ϕ∥≤∥ϕ∥(ϕ∈H),∥Tϑ−1F∥D′≤ϑ−2∥F∥D′(F∈D′).(4)\|T_\vartheta^{-1}\phi\| \le\|\phi\| \quad(\phi\in H), \qquad\|T_\vartheta^{-1}F\|_{D'} \le\vartheta^{-2}\|F\|_{D'} \quad(F \in D'). \tag*{(4)}

The observations are represented by the low-mode projection introduced in the Introduction. For N∈NN \in\mathbb{N}, this projection and its complement are given by

PNϕ=∑0<∣k∣≤Nϕ^ke2πik⋅x/LΩ,QN=I−PN.P_N\phi= \sum_{0<|k|\le N}\widehat{\phi}_k e^{2\pi i k\cdot x/L_\Omega}, \qquad Q_N = I-P_N.

By orthogonality in each space V˙s\dot{V}_s, these projections satisfy the following relations:

∥QNϕ∥2≤1λ1N2∥∇ϕ∥2,∥PNϕ∥≤∥ϕ∥,∥∇PNϕ∥≤∥∇ϕ∥,PNA=APN.(5)\|Q_N\phi\|^2 \le\frac{1}{\lambda_1N^2}\|\nabla\phi\|^2, \qquad\|P_N\phi\| \le\|\phi\|, \qquad\|\nabla P_N\phi\| \le\|\nabla\phi\|, \qquad P_NA = AP_N. \tag*{(5)}

We define the bilinear operator B:V×V→V′B: V \times V \to V' by

B(u,v)=P[(u⋅∇)v].B(u,v)=\mathcal{P}[(u\cdot\nabla)v].

It extends continuously to B:H×V→D′B:H\times V\to D', with

∥B(u,v)∥D′≤C∥u∥∥∇v∥(u∈H, v∈V).(6)\|B(u,v)\|_{D'} \le C\|u\|\|\nabla v\| \qquad(u\in H,\ v\in V). \tag*{(6)}

For u,v,w∈Vu,v,w\in V, and also for u∈Hu\in H and v,w∈D(A)v,w\in D(A), we have

⟨B(u,v),v⟩=0,⟨B(u,v),w⟩=−⟨B(u,w),v⟩.(7)\langle B(u,v),v\rangle=0,\qquad\langle B(u,v),w\rangle=-\langle B(u,w),v\rangle. \tag*{(7)}

Throughout the paper, CC denotes a positive constant that may change from one occurrence to another, and may depend on the domain.

Parameter recovery algorithm

Throughout the analysis, we assume that f∈L∞(0,∞;H)f\in L^\infty(0,\infty;H) and write F=∥f∥L∞(0,∞;H)F=\|f\|_{L^\infty(0,\infty;H)}. Applying P\mathcal{P} to (1) gives the functional formulation

{dvdt+νAv+B(v,u)=f,v=Tαu=(I+α2A)u,u(0)=u0.(8)\begin{cases} \dfrac{dv}{dt}+\nu Av+B(v,u)=f,\\ v=T_\alpha u=(I+\alpha^2A)u,\\ u(0)=u_0. \tag*{(8)} \end{cases}

Similarly, for fixed β∈[α0,α1]\beta\in[\alpha_0,\alpha_1], the assimilation system (2) becomes

{dzdt+νAz+B(z,w)=f−ηTβPN(w−u),z=Tβw=(I+β2A)w.(9)\begin{cases} \dfrac{dz}{dt}+\nu Az+B(z,w)=f-\eta T_\beta P_N(w-u),\\ z=T_\beta w=(I+\beta^2A)w. \tag*{(9)} \end{cases}

Here the feedback field PN(w−u)P_N(w-u) is divergence free.

For sufficiently regular ϕ\phi, we introduce two operators that separate the terms with and without the squared filtering length:

R(ϕ)=ϕt+νAϕ+B(ϕ,ϕ),R(\phi)=\phi_t+\nu A\phi+B(\phi,\phi),
S(ϕ)=Aϕt+νA2ϕ+B(Aϕ,ϕ).S(\phi)=A\phi_t+\nu A^2\phi+B(A\phi,\phi).

With this notation, systems (8) and (9) can be written in the form

R(u)+α2S(u)=f,R(w)+β2S(w)=f−ηTβPN(w−u).(10)R(u)+\alpha^2S(u)=f,\qquad R(w)+\beta^2S(w)=f-\eta T_\beta P_N(w-u). \tag*{(10)}

Estimate (6) yields

ϕ∈L∞(V)∩L2(D(A)),ϕt∈L2(H)⟹Aϕt,A2ϕ,B(Aϕ,ϕ)∈L2(D′).\phi\in L^\infty(V)\cap L^2(D(A)),\qquad\phi_t\in L^2(H)\quad\Longrightarrow\quad A\phi_t,A^2\phi,B(A\phi,\phi)\in L^2(D').

Fixing t1≥0t_1\geq0, a regular solution satisfies

u∈C(J;V)∩L2(J;D(A)),ut∈L2(J;H),u\in C(\mathcal{J};V)\cap L^2(\mathcal{J};D(A)),\qquad u_t\in L^2(\mathcal{J};H),

on a bounded interval J=[r,r+ℓ]\mathcal{J}=[r,r+\ell], where r≥t1r\geq t_1 and 0<ℓ<∞0<\ell<\infty.

For ϑ>0\vartheta>0 and ϕ∈V\phi\in V, we define the corresponding energy by

Eϑ(ϕ)=∥ϕ∥2+ϑ2∥∇ϕ∥2.E_\vartheta(\phi)=\|\phi\|^2+\vartheta^2\|\nabla\phi\|^2.

For u0∈Vu_0\in V, the reference system admits a unique global regular solution: see [20] for autonomous forcing and [4] for the time-dependent forcing considered here. Taking the duality pairing of (8) with uu and using (7) and Young’s inequality together with (3), we obtain

ddtEα(u)+νλ1Eα(u)≤F2νλ1.\frac{d}{dt}E_\alpha(u)+\nu\lambda_1E_\alpha(u)\leq\frac{F^2}{\nu\lambda_1}.

Applying Gronwall’s inequality, we conclude that, for t≥t1t\geq t_1,

Eα(u(t))≤e−νλ1(t−t1)Eα(u(t1))+F2ν2λ12(1−e−νλ1(t−t1)).(11)E_\alpha(u(t))\leq e^{-\nu\lambda_1(t-t_1)}E_\alpha(u(t_1))+\frac{F^2}{\nu^2\lambda_1^2}\left(1-e^{-\nu\lambda_1(t-t_1)}\right). \tag*{(11)}

Assume that a known, parameter-independent bound Eα1(u(t1))1/2≤RrefE_{\alpha_1}(u(t_1))^{1/2}\leq R_{\mathrm{ref}} is available. Then (11) gives

Eα(u(t))1/2≤Mref:=max⁡{Rref,Fνλ1},t≥t1.(12)E_\alpha(u(t))^{1/2}\leq M_{\mathrm{ref}}:=\max\left\{R_{\mathrm{ref}},\frac{F}{\nu\lambda_1}\right\},\qquad t\geq t_1. \tag*{(12)}

For the parameter recovery estimates, we impose the following additional regularity assumptions on the reference solution.

Assumption 1. There exist constants M3,M4>0M_3,M_4>0, depending only on the reference solution, such that

(i)

sup⁡t≥t1∥u(t)∥V3≤M3.(13)\sup_{t\geq t_1}\|u(t)\|_{V_3}\leq M_3. \tag*{(13)}

(ii)

sup⁡r≥t1∫rr+1(∥u(s)∥V42+∥ut(s)∥V22) ds≤M42.\sup_{r\geq t_1}\int_r^{r+1}\left(\|u(s)\|_{V_4}^2+\|u_t(s)\|_{V_2}^2\right)\,ds\leq M_4^2.

For every T>t1T>t_1, condition (ii) and the embedding L2(t1,T;V˙4)∩H1(t1,T;V˙2)↪C([t1,T];V˙3)L^2(t_1,T;\dot{V}_4)\cap H^1(t_1,T;\dot{V}_2)\hookrightarrow C([t_1,T];\dot{V}_3) give a continuous representative of uu, which we use throughout. Thus the supremum in (13) agrees with the essential supremum.

As in [5], we keep NN and η\eta fixed throughout the recovery iteration. Let t1<t2<⋯t_1<t_2<\cdots denote the successive parameter-update times, with βn\beta_n held fixed on [tn,tn+1][t_n,t_{n+1}]. On each interval [tn,tn+1][t_n,t_{n+1}], we use the approximate parameter βn∈[α0,α1]\beta_n\in[\alpha_0,\alpha_1] and solve

{(zn)t+νAzn+B(zn,wn)=f−ηTβnPN(wn−u),zn=Tβnwn,wn(tn)=wn−1(tn)(n≥2),w1(t1)=w0.(14)\begin{cases} (z_n)_t+\nu Az_n+B(z_n,w_n)=f-\eta T_{\beta_n}P_N(w_n-u), \quad z_n=T_{\beta_n}w_n,\\ w_n(t_n)=w_{n-1}(t_n)\quad(n\ge2), \quad w_1(t_1)=w_0. \tag*{(14)} \end{cases}

The initial state w0∈Vw_0\in V is arbitrary.

Let gn=wn−ug_n=w_n-u, so that PNgn=PNwn−PNuP_Ng_n=P_Nw_n-P_Nu. Subtracting the two equations in (10), we obtain

(α2−βn2)S(u)=R(wn)−R(u)+βn2(S(wn)−S(u))+η(I+βn2A)PNgn.(15)(\alpha^2-\beta_n^2)S(u)=R(w_n)-R(u)+\beta_n^2(S(w_n)-S(u))+\eta(I+\beta_n^2A)P_Ng_n. \tag*{(15)}

Following [5], we approximate the reference solution by combining the observed low modes of uu with the computed high modes of wnw_n:

uN∗,n=wn−PNgn=QNwn+PNu,u−uN∗,n=−QNgn.(16)u_N^{*,n}=w_n-P_Ng_n=Q_Nw_n+P_Nu,\qquad u-u_N^{*,n}=-Q_Ng_n. \tag*{(16)}

Define

DN(n)(t):=A(PNu)t+νA2PNu+PNB(AuN∗,n,uN∗,n).(17)D_N^{(n)}(t):=A(P_Nu)_t+\nu A^2P_Nu+P_NB(Au_N^{*,n},u_N^{*,n}). \tag*{(17)}

Using the same replacement in the projected error equation, we define

YN(n):=Tβn(PN(gn)t+νAPNgn+ηPNgn)+PN{B(Tβnwn,wn)−B(TβnuN∗,n,uN∗,n)}.(18)\begin{aligned} Y_N^{(n)}:=T_{\beta_n}\bigl(P_N(g_n)_t+\nu AP_Ng_n+\eta P_Ng_n\bigr)\\ {}+P_N\bigl\{B(T_{\beta_n}w_n,w_n)-B(T_{\beta_n}u_N^{*,n},u_N^{*,n})\bigr\}. \tag*{(18)} \end{aligned}

Since PNA=APNP_NA=AP_N, (17) and (18) can be computed from wnw_n, PNuP_Nu, and (PNu)t(P_Nu)_t. Moreover,

DN(n),YN(n)∈L2(tn,tn+1;PNH).D_N^{(n)},Y_N^{(n)}\in L^2(t_n,t_{n+1};P_NH).

We use exact observations in continuous time.

Definition 1. For each n≥1n\ge1, we choose an intermediate time t^n\widehat{t}_n and denote the corresponding update interval and its length by

tn<t^n<tn+1,In=[t^n,tn+1],τn=∣In∣.t_n<\widehat{t}_n<t_{n+1},\qquad I_n=[\widehat{t}_n,t_{n+1}],\qquad\tau_n=|I_n|.

The interval [tn,t^n][t_n,\widehat{t}_n] is used for relaxation, and InI_n is used to compute the parameter update. We require the update signal to satisfy

δn:=∫tntn+1∥DN(n)(s)∥2 ds>0.(19)\delta_n:=\int_{t_n}^{t_{n+1}}\left\|D_N^{(n)}(s)\right\|^2\,ds>0. \tag*{(19)}

Fix ζ>0\zeta>0. An update interval InI_n is accepted if δn/τn≥ζ\delta_n/\tau_n\ge\zeta.

On each update interval, we first form a candidate value for the squared parameter:

γn+1:=βn2+1δn∫tntn+1(YN(n)(s),DN(n)(s)) ds,(20)\gamma_{n+1}:=\beta_n^2+\frac{1}{\delta_n}\int_{t_n}^{t_{n+1}}\left(Y_N^{(n)}(s),D_N^{(n)}(s)\right)\,ds, \tag*{(20)}

then project this value onto [α02,α12][\alpha_0^2,\alpha_1^2], obtaining

βn+1=min⁡{α12,max⁡{α02,γn+1}}.(21)\beta_{n+1}=\sqrt{\min\left\{\alpha_1^2,\max\left\{\alpha_0^2,\gamma_{n+1}\right\}\right\}}. \tag*{(21)}

We propose the following parameter updating algorithm.

Algorithm 1.

Initialization. Given the observations PNu(t)P_Nu(t) on [t1,T][t_1,T], where T>t1T>t_1, choose w0∈Vw_0\in V and β1∈[α0,α1]\beta_1\in[\alpha_0,\alpha_1]. Let L>0L>0 be the fixed relaxation time before each update window. Keep N∈NN\in\mathbb{N}, η>0\eta>0, and the threshold ζ>0\zeta>0 fixed, and choose a finite list of trial lengths satisfying

0<τ0=ℓ1<⋯<ℓJ≤τmax⁡,J≥1.0<\tau_0=\ell_1<\cdots<\ell_J\le\tau_{\max},\qquad J\ge1.

Require the reference estimates from t1t_1 onward; set w1(t1)=w0w_1(t_1)=w_0, n=1n=1.

Iteration. Given tn,βn,wn(tn)t_n,\beta_n,w_n(t_n):

  1. If tn+L+τ0>Tt_n+L+\tau_0>T, stop. Otherwise set t^n=tn+L\widehat{t}_n=t_n+L.

  1. Try the candidate window lengths ℓ1,…,ℓJ\ell_1,\ldots,\ell_J in increasing order, restricted to t^n+ℓj≤T\widehat{t}_n+\ell_j\le T. Set

τn=ℓj,tn+1=t^n+τn.\tau_n=\ell_j,\qquad t_{n+1}=\widehat{t}_n+\tau_n.

On [tn,tn+1][t_n,t_{n+1}], solve (14) with the same βn,wn(tn)\beta_n,w_n(t_n) for every trial. Compute (16)–(19); accept the first trial with δn/τn≥ζ\delta_n/\tau_n\ge\zeta.

  1. If all eligible trials fail, stop without changing βn\beta_n or wn(tn)w_n(t_n).

  1. For an accepted window, apply (20)–(21), set wn+1(tn+1)=wn(tn+1)w_{n+1}(t_{n+1})=w_n(t_{n+1}), increment nn, and repeat.

Output. Return βn,wn(tn)\beta_n,w_n(t_n), or β1,w0\beta_1,w_0 if no update was accepted.

Estimates for the reference and assimilation systems

For each κ>0\kappa>0, we introduce the quantities

KS(κ)=C((1+ν2)M421−e−κ+M34κ),MS2=C((1+ν2)M42(1+τ0−1)+M34),(22)\begin{aligned} K_S(\kappa)&=C\left(\frac{(1+\nu^2)M_4^2}{1-e^{-\kappa}}+\frac{M_3^4}{\kappa}\right),\\ M_S^2&=C\left((1+\nu^2)M_4^2(1+\tau_0^{-1})+M_3^4\right), \tag*{(22)} \end{aligned}

where τ0\tau_0 is the shortest trial length and CC is a sufficiently large constant depending only on the domain.

Lemma 1. Under Assumption 1, for J=[r,r+ℓ]⊂[t1,∞)\mathcal{J}=[r,r+\ell]\subset[t_1,\infty), with 0<ℓ<∞0<\ell<\infty, we have S(u)∈L2(J;H)S(u)\in L^2(\mathcal{J};H) and

∫J∥S(u(t))∥2 dt≤C((1+ν2)M42(ℓ+1)+M34ℓ)≤MS2ℓ(ℓ≥τ0>0).(23)\int_{\mathcal{J}}\|S(u(t))\|^2\,dt\le C\left((1+\nu^2)M_4^2(\ell+1)+M_3^4\ell\right)\le M_S^2\ell\qquad(\ell\ge\tau_0>0). \tag*{(23)}

Moreover, for every κ>0\kappa>0,

∫rte−κ(t−s)∥S(u(s))∥2 ds≤KS(κ),t≥r≥t1.(24)\int_r^t e^{-\kappa(t-s)}\|S(u(s))\|^2\,ds\le K_S(\kappa),\qquad t\ge r\ge t_1. \tag*{(24)}

Proof. Applying Assumption 1 on successive unit intervals starting at rr, we obtain

∫J(∥ut∥V2′2+∥u∥V42) ds≤M42(ℓ+1).(25)\int_{\mathcal{J}}\left(\|u_t\|_{V_2'}^2+\|u\|_{V_4}^2\right)\,ds\le M_4^2(\ell+1). \tag*{(25)}

For the nonlinear term, Hölder’s inequality and the Sobolev embeddings give, for a.e. s≥t1s\ge t_1,

∥B(Au(s),u(s))∥≤C∥Au(s)∥L3∥∇u(s)∥L6≤C∥u(s)∥V32≤CM32.\|B(Au(s),u(s))\|\le C\|Au(s)\|_{L^3}\|\nabla u(s)\|_{L^6}\le C\|u(s)\|_{V_3}^2\le CM_3^2.

Since S(u)=Aut+νA2u+B(Au,u)S(u)=Au_t+\nu A^2u+B(Au,u), we have

∥S(u)∥2≤3∥Aut∥2+3ν2∥A2u∥2+3∥B(Au,u)∥2≤C(∥ut∥V22+ν2∥u∥V42+M34).\|S(u)\|^2 \le3\|Au_t\|^2+3\nu^2\|A^2u\|^2+3\|B(Au,u)\|^2 \le C\left(\|u_t\|_{V_2}^2+\nu^2\|u\|_{V_4}^2+M_3^4\right).

Integrating over J\mathcal{J} and applying (25), we obtain

∫J∥S(u(s))∥2 ds≤C((1+ν2)M42(ℓ+1)+M34ℓ).\int_{\mathcal{J}}\|S(u(s))\|^2\,ds \le C\left((1+\nu^2)M_4^2(\ell+1)+M_3^4\ell\right).

For ℓ≥τ0>0\ell\ge\tau_0>0, we have ℓ+1≤(1+τ0−1)ℓ\ell+1\le(1+\tau_0^{-1})\ell. Hence

∫J∥S(u(s))∥2 ds≤C((1+ν2)M42(1+τ0−1)+M34)ℓ=M22ℓ,\int_{\mathcal{J}}\|S(u(s))\|^2\,ds \le C\left((1+\nu^2)M_4^2(1+\tau_0^{-1})+M_3^4\right)\ell=M_2^2\ell,

where the constant in (22) is chosen sufficiently large. This proves (23).

For t≥r≥t1t\ge r\ge t_1, set Ij=[t−j−1,t−j]∩[r,t]I_j=[t-j-1,t-j]\cap[r,t] for j≥0j\ge0. Each nonempty IjI_j is contained in a unit interval starting at or after t1t_1. Applying Assumption 1 on these intervals, we obtain

∫rte−κ(t−s)∥S(u(s))∥2 ds≤C∑j=0∞e−κj∫Ij(∥Aut∥2+ν2∥A2u∥2) ds+CM43κ≤C(1+ν2)M421−e−κ+CM34κ.\begin{aligned} \int_r^t e^{-\kappa(t-s)}\|S(u(s))\|^2\,ds \le C\sum_{j=0}^{\infty}e^{-\kappa j}\int_{I_j}\left(\|Au_t\|^2+\nu^2\|A^2u\|^2\right)\,ds+\frac{CM_4^3}{\kappa} \\ &\le\frac{C(1+\nu^2)M_4^2}{1-e^{-\kappa}}+\frac{CM_3^4}{\kappa}. \end{aligned}

We have the following well-posedness result.

Proposition 1. Let βn∈[α0,α1]\beta_n\in[\alpha_0,\alpha_1], η>0\eta>0, wn(tn)∈Vw_n(t_n)\in V, f∈L∞(0,∞;H)f\in L^\infty(0,\infty;H), and let uu be the reference solution. With βn\beta_n fixed, (14) has a unique global solution satisfying, for every T>tnT>t_n,

wn∈C([tn,T];V)∩L2(tn,T;D(A)),(wn)t∈L2(tn,T;H).w_n\in C([t_n,T];V)\cap L^2(t_n,T;D(A)),\qquad(w_n)_t\in L^2(t_n,T;H).

Proof. Fix nn and T>tnT>t_n, and write β=βn\beta=\beta_n. We use the Galerkin construction in [4 Theorem 2]. For m≥1m\ge1, let Πm\Pi_m be the orthogonal projections onto the spaces in an increasing sequence of finite-dimensional Stokes eigenspaces such that PNH⊂ΠmHP_NH\subset\Pi_mH and Πm→I\Pi_m\to I strongly in HH and VV. Let wmw_m denote the corresponding Galerkin approximation, with wm(tn)=Πmwn(tn)w_m(t_n)=\Pi_mw_n(t_n). Since PNP_N is an orthogonal projection commuting with AA,

(TβPNwm,wm)=Eβ(PNwm)≥0.(T_\beta P_Nw_m,w_m)=E_\beta(P_Nw_m)\ge0.

Testing the Galerkin equation with wmw_m, using (7) and Young’s inequality, gives

ddtEβ(wm)+ν(∥∇wm∥2+β2∥Awm∥2)+ηEβ(PNwm)≤F2νλ1+ηEβ(PNu).\frac{d}{dt}E_\beta(w_m)+\nu\left(\|\nabla w_m\|^2+\beta^2\|Aw_m\|^2\right)+\eta E_\beta(P_Nw_m)\le\frac{F^2}{\nu\lambda_1}+\eta E_\beta(P_Nu).

Thus, for every η>0\eta>0, the approximations exist on [tn,T][t_n,T] and are uniformly bounded in L∞(tn,T;V)∩L2(tn,T;D(A))L^\infty(t_n,T;V)\cap L^2(t_n,T;D(A)). Applying Tβ−1T_\beta^{-1} to their equations and using (4) and (6), we obtain

∥(wm)t∥≤ν∥Awm∥+Cα0−2∥Tβwm∥∥∇wm∥+F+η∥PN(wm−u)∥.\|(w_m)_t\|\le\nu\|Aw_m\|+C\alpha_0^{-2}\|T_\beta w_m\|\|\nabla w_m\|+F+\eta\|P_N(w_m-u)\|.

Thus (wm)t(w_m)_t is bounded in L2(tn,T;H)L^2(t_n,T;H). By the Aubin–Lions theorem, after extracting a subsequence,

wm→wnin L2(tn,T;V),Tβwm⇀Tβwnin L2(tn,T;H).w_m\to w_n\quad\text{in }L^2(t_n,T;V),\qquad T_\beta w_m\rightharpoonup T_\beta w_n\quad\text{in }L^2(t_n,T;H).

The limit solves (14) with the given initial data. The embedding L2(tn,T;D(A))∩H1(tn,T;H)↪C([tn,T];V)L^2(t_n,T;D(A))\cap H^1(t_n,T;H)\hookrightarrow C([t_n,T];V) yields the continuity.

For uniqueness, let e=w(1)−w(2)e=w^{(1)}-w^{(2)} for two solutions with identical data. Applying the interpolation estimates and Young’s inequality, we have

∣⟨B(Tβe,w(1)),e⟩∣≤νβ22∥Ae∥2+C(1+∥∇w(1)∥∥Aw(1)∥)Eβ(e).\left|\left\langle B(T_\beta e,w^{(1)}),e\right\rangle\right|\leq\frac{\nu\beta^2}{2}\|Ae\|^2+C\left(1+\|\nabla w^{(1)}\|\|Aw^{(1)}\|\right)E_\beta(e).

Testing the difference equation with ee yields

ddtEβ(e)≤C(1+∥∇w(1)∥∥Aw(1)∥)Eβ(e).\frac{d}{dt}E_\beta(e)\leq C\left(1+\|\nabla w^{(1)}\|\|Aw^{(1)}\|\right)E_\beta(e).

Gronwall’s inequality implies e=0e=0. Since T>tnT>t_n is arbitrary, the solution is global.

For a sufficiently large domain-dependent constant CC, we define

C0=C(1+ν−3)(Mref4α0−4+α12(1+M34)).(26)C_0=C\left(1+\nu^{-3}\right)\left(M_{\mathrm{ref}}^4\alpha_0^{-4}+\alpha_1^2(1+M_3^4)\right). \tag*{(26)}

We require the observation resolution and nudging strength to satisfy

2C0λ1N2≤ν8,η>C0.(27)\frac{2C_0}{\lambda_1N^2}\leq\frac{\nu}{8},\qquad\eta>C_0. \tag*{(27)}

Using the quantities in (22), we also fix

κ=νλ18,CS∗=8νλ1,K0=CS∗KS(κ),KA=4CS∗MS2να02.(28)\kappa=\frac{\nu\lambda_1}{8},\qquad C_S^*=\frac{8}{\nu\lambda_1},\qquad K_0=C_S^*K_S(\kappa),\qquad K_A=\frac{4C_S^*M_S^2}{\nu\alpha_0^2}. \tag*{(28)}

The following estimate refines [4] by retaining the higher-order dissipation and keeping the constants uniform over all assimilation intervals. This improvement is useful in the parameter update.

Proposition 2. Let uu be a solution of (8) satisfying Assumption 1 and Eα(u)≤Mref2E_\alpha(u)\leq M_{\mathrm{ref}}^2 on [tn,tn+1][t_n,t_{n+1}]. Let ww be a solution of (14) with fixed β∈[α0,α1]\beta\in[\alpha_0,\alpha_1], and set g=w−ug=w-u. Under (27), with constants (26) and (28), we have

ddtEβ(g)+κEβ(g)+ν4∥∇g∥2+νβ24∥Ag∥2≤CS∗∥S(u(t))∥2∣β2−α2∣2(29)\frac{d}{dt}E_\beta(g)+\kappa E_\beta(g)+\frac{\nu}{4}\|\nabla g\|^2+\frac{\nu\beta^2}{4}\|Ag\|^2\leq C_S^*\|S(u(t))\|^2|\beta^2-\alpha^2|^2 \tag*{(29)}

for a.e. t∈[tn,tn+1]t\in[t_n,t_{n+1}]. Moreover, for t∈[tn,tn+1]t\in[t_n,t_{n+1}] and J=[r,r+ℓ]⊂[tn,tn+1]\mathcal{J}=[r,r+\ell]\subset[t_n,t_{n+1}] with ℓ≥τ0>0\ell\geq\tau_0>0,

Eβ(g(t))≤e−κ(t−tn)Eβ(g(tn))+K0∣β2−α2∣2,E_\beta(g(t))\leq e^{-\kappa(t-t_n)}E_\beta(g(t_n))+K_0|\beta^2-\alpha^2|^2,
∫J∥Ag(s)∥2 ds≤4νβ2Eβ(g(r))+KA∣β2−α2∣2ℓ.(30)\int_{\mathcal{J}}\|Ag(s)\|^2\,ds\leq\frac{4}{\nu\beta^2}E_\beta(g(r))+K_A|\beta^2-\alpha^2|^2\ell. \tag*{(30)}

The constants κ\kappa, CS∗C_S^*, K0K_0, and KAK_A are independent of NN and η\eta once (27) holds.

Proof. Subtracting the reference equation from (9), we obtain

Tβgt+νATβg+B(Tβg,u)+B(Tβw,g)+ηTβPNg=(α2−β2)S(u).T_\beta g_t+\nu AT_\beta g+B(T_\beta g,u)+B(T_\beta w,g)+\eta T_\beta P_Ng=(\alpha^2-\beta^2)S(u).

Here we have used the identity B(Tβw,w)−B(Tβu,u)=B(Tβg,u)+B(Tβw,g)B(T_\beta w,w)-B(T_\beta u,u)=B(T_\beta g,u)+B(T_\beta w,g). Pairing with gg and using (7), we have

12ddtEβ(g)+ν(∥∇g∥2+β2∥Ag∥2)+η(∥PNg∥2+β2∥∇PNg∥2)≤∣⟨B(g,u),g⟩∣+β2∣⟨B(Ag,u),g⟩∣+∣β2−α2∣∣(S(u),g)∣.\frac{1}{2}\frac{d}{dt}E_\beta(g)+\nu\left(\|\nabla g\|^2+\beta^2\|Ag\|^2\right)+\eta\left(\|P_Ng\|^2+\beta^2\|\nabla P_Ng\|^2\right) \leq\left|\langle B(g,u),g\rangle\right|+\beta^2\left|\langle B(Ag,u),g\rangle\right|+|\beta^2-\alpha^2|\left|(S(u),g)\right|.

Using the interpolation inequalities, Young’s inequality, and ∥∇u∥≤Mref/α0\|\nabla u\| \le M_{\mathrm{ref}}/\alpha_0, ∥∇u∥L∞≤CM3\|\nabla u\|_{L^\infty} \le CM_3, and β≤α1\beta\le\alpha_1, we have

∣⟨B(g,u),g⟩∣≤ν16∥∇g∥2+Cν−3∥∇u∥4∥g∥2.|\langle B(g,u),g\rangle| \le\frac{\nu}{16}\|\nabla g\|^2 + C\nu^{-3}\|\nabla u\|^4\|g\|^2.
β2∣⟨B(Ag,u),g⟩∣≤νβ216∥Ag∥2+Cν−1β2M32∥g∥2.\beta^2|\langle B(Ag,u),g\rangle| \le\frac{\nu\beta^2}{16}\|Ag\|^2 + C\nu^{-1}\beta^2M_3^2\|g\|^2.

The Poincaré and Young inequalities give

∣β2−α2∣∣⟨S(u),g⟩∣≤λ1−1/2∣β2−α2∣∥S(u(t))∥∥∇g∥≤ν16∥∇g∥2+4νλ1∥S(u(t))∥2∣β2−α2∣2.|\beta^2-\alpha^2||\langle S(u),g\rangle| \le\lambda_1^{-1/2}|\beta^2-\alpha^2|\|S(u(t))\|\|\nabla g\| \le\frac{\nu}{16}\|\nabla g\|^2 + \frac{4}{\nu\lambda_1}\|S(u(t))\|^2|\beta^2-\alpha^2|^2.

Choosing C0C_0 large enough, we then obtain

ddtEβ(g)+ν2∥∇g∥2+νβ22∥Ag∥2+2η(∥PNg∥2+β2∥∇PNg∥2)≤C0∥g∥2+CS∗∥S(u(t))∥2∣β2−α2∣2.\frac{d}{dt}E_\beta(g) + \frac{\nu}{2}\|\nabla g\|^2 + \frac{\nu\beta^2}{2}\|Ag\|^2 + 2\eta(\|P_Ng\|^2+\beta^2\|\nabla P_Ng\|^2) \le C_0\|g\|^2+C_S^*\|S(u(t))\|^2|\beta^2-\alpha^2|^2.

Using the orthogonality of the low and high Fourier modes together with (5), we have

C0∥g∥2≤2C0∥PNg∥2+2C0λ1N2∥∇g∥2.C_0\|g\|^2 \le2C_0\|P_Ng\|^2+\frac{2C_0}{\lambda_1N^2}\|\nabla g\|^2.

By (27), the low-mode contribution is absorbed by the feedback term, and the high-mode contribution is absorbed by part of the viscous dissipation. Consequently,

ddtEβ(g)+3ν8∥∇g∥2+νβ22∥Ag∥2≤CS∗∥S(u(t))∥2∣β2−α2∣2.\frac{d}{dt}E_\beta(g)+\frac{3\nu}{8}\|\nabla g\|^2+\frac{\nu\beta^2}{2}\|Ag\|^2\le C_S^*\|S(u(t))\|^2|\beta^2-\alpha^2|^2.

From (3), we have

ν8∥∇g∥2+νβ24∥Ag∥2≥νλ18(∥g∥2+β2∥∇g∥2).\frac{\nu}{8}\|\nabla g\|^2+\frac{\nu\beta^2}{4}\|Ag\|^2\ge\frac{\nu\lambda_1}{8}(\|g\|^2+\beta^2\|\nabla g\|^2).

This proves (29). Multiplying by the exponential integrating factor and applying (24), we obtain the first bound in (30). Integrating the differential inequality over J\mathcal{J} and applying (23), we obtain

νβ24∫J∥Ag(s)∥2 ds≤Eβ(g(r))+CS∗MS2∣β2−α2∣2ℓ.\frac{\nu\beta^2}{4}\int_{\mathcal{J}}\|Ag(s)\|^2\,ds\le E_\beta(g(r))+C_S^*M_S^2|\beta^2-\alpha^2|^2\ell.

Since β≥α0\beta\ge\alpha_0, we obtain the second bound in (30). □

Remark 1. For the synchronization estimates above, once NN satisfies (27), any fixed η>C0\eta>C_0 is allowed. Unlike condition (20) in [5], these estimates impose no upper bound on η\eta.

Synchronization estimates

We use the synchronization and dissipation estimates to control the error introduced by replacing the unobserved part of the reference solution. The following identity relates this error to the parameter update.

Lemma 2. Under Assumption 1, let wnw_n be given by Proposition 1. For a.e. t∈Int \in I_n, there is a residual RN(n)\mathcal{R}^{(n)}_N satisfying

YN(n)(t)=(α2−βn2)DN(n)(t)+RN(n)(t),Y^{(n)}_N(t) = (\alpha^2-\beta_n^2)D^{(n)}_N(t)+\mathcal{R}^{(n)}_N(t),
∥RN(n)(t)∥≤CR(1+Eβn(gn(t))1/2)(Eβn(gn(t))1/2+∥AQNgn(t)∥).(31)\left\|\mathcal{R}^{(n)}_N(t)\right\| \le C_R(1+E_{\beta_n}(g_n(t))^{1/2})(E_{\beta_n}(g_n(t))^{1/2}+\|AQ_Ng_n(t)\|). \tag*{(31)}

where CRC_R depends on ν,α0,α1,M3,N\nu,\alpha_0,\alpha_1,M_3,N, but is independent of nn.

Proof. Applying the low-mode projection to (15) and using the definitions in (17) and (18), we obtain the identity in (31), with residual

RN(n)=PN{B(Tαu,u)−B(TαuN∗,n,uN∗,n)}.\mathcal{R}^{(n)}_N=P_N\{B(T_\alpha u,u)-B(T_\alpha u_N^{*,n},u_N^{*,n})\}.

To estimate this residual, we write h=QNgnh=Q_Ng_n so that uN∗,n=u+hu_N^{*,n}=u+h. For ϕ∈PNH\phi\in P_NH with ∥ϕ∥=1\|\phi\|=1, we have ∥∇ϕ∥L∞≤cN\|\nabla\phi\|_{L^\infty}\le c_N, where cNc_N depends only on NN and Ω\Omega. By bilinearity and integration by parts, using ∇⋅u=∇⋅h=0\nabla\cdot u=\nabla\cdot h=0, we obtain

∣⟨B(u,u)−B(u+h,u+h),ϕ⟩∣≤[cN(∥u∥+∥h∥)+∥∇u∥L∞]∥h∥,|\langle B(u,u)-B(u+h,u+h),\phi\rangle| \le[c_N(\|u\|+\|h\|)+\|\nabla u\|_{L^\infty}]\|h\|,
∣⟨B(Au,u)−B(A(u+h),u+h),ϕ⟩∣≤cN(∥Au∥+∥Ah∥)∥h∥+∥∇u∥L∞∥Ah∥.|\langle B(Au,u)-B(A(u+h),u+h),\phi\rangle| \le c_N(\|Au\|+\|Ah\|)\|h\|+\|\nabla u\|_{L^\infty}\|Ah\|.

Combining these estimates and taking the supremum over unit test fields ϕ∈PNH\phi\in P_NH, we obtain

∥RN(n)∥≤C(N,α1,Ω)(∥u∥+∥Au∥+∥∇u∥L∞+∥h∥)(∥h∥+∥Ah∥).\|\mathcal{R}^{(n)}_N\|\le C(N,\alpha_1,\Omega)(\|u\|+\|Au\|+\|\nabla u\|_{L^\infty}+\|h\|)(\|h\|+\|Ah\|).

Finally, we use (13), the embedding V˙3↪W1,∞\dot{V}_3\hookrightarrow W^{1,\infty}, the inequality ∥h∥≤Eβn(gn)1/2\|h\|\le E_{\beta_n}(g_n)^{1/2}, and the identity ∥Ah∥=∥AQNgn∥\|Ah\|=\|AQ_Ng_n\|. These bounds yield (31), with a constant uniform in n,Ln,L.

For δn>0\delta_n>0, substituting (31) into (20) gives

γn+1−α2=δn−1∫In⟨RN(n),DN(n)⟩ ds.\gamma_{n+1}-\alpha^2=\delta_n^{-1}\int_{I_n}\langle\mathcal{R}^{(n)}_N,D^{(n)}_N\rangle\,ds.

Applying the Cauchy–Schwarz inequality, we obtain

∣βn+12−α2∣≤1δn∫In∥RN(n)∥∥DN(n)∥ ds≤∥RN(n)∥L2(In;H)δn1/2=τn−1/2∥RN(n)∥L2(In;H)(δn/τn)1/2.(32)\begin{aligned} |\beta_{n+1}^2-\alpha^2| \le\frac{1}{\delta_n}\int_{I_n}\|\mathcal{R}^{(n)}_N\|\|D^{(n)}_N\|\,ds \\ &\le\frac{\|\mathcal{R}^{(n)}_N\|_{L^2(I_n;H)}}{\delta_n^{1/2}} =\frac{\tau_n^{-1/2}\|\mathcal{R}^{(n)}_N\|_{L^2(I_n;H)}}{(\delta_n/\tau_n)^{1/2}}. \tag*{(32)} \end{aligned}

With CRC_R as in Lemma 2, we fix the following constants before choosing the relaxation length LL:

P∗=α12−α02,E∗=1+K0P∗2,CA=2α0τ0.P_*=\alpha_1^2-\alpha_0^2,\qquad E_*=1+K_0P_*^2,\qquad C_A=\frac{2}{\alpha_0\sqrt{\tau_0}}.

We then define

Crec=CR(1+E∗)max⁡{1+CA,(1+CA)K0+KA}.(33)C_{\mathrm{rec}}=C_R(1+\sqrt{E_*})\max\{1+C_A,(1+C_A)\sqrt{K_0}+\sqrt{K_A}\}. \tag*{(33)}

The constants K0K_0, KAK_A, E∗E_*, CrecC_{\mathrm{rec}} are independent of LL and w0w_0. The remaining assumptions require successful update windows and a sufficiently strong signal relative to the bound on RN(n)\mathcal{R}^{(n)}_N.

Assumption 2. (H1) α∈[α0,α1]\alpha\in[\alpha_0,\alpha_1], with α0>0\alpha_0 > 0. The reference solution satisfies Assumption 1 and the initial energy bound used in (12).

(H2) NN and η\eta remain fixed and satisfy (27), with C0C_0 given by (26).

(H3) Fix the trial lengths 0<ℓ1<⋯<ℓJ≤τmax⁡<∞0 < \ell_1 < \cdots< \ell_J \le\tau_{\max} < \infty and threshold ζ>0\zeta> 0 independently of LL. There exists Lsig>0L_{\mathrm{sig}} > 0 such that, for every L≥LsigL \ge L_{\mathrm{sig}}, Algorithm 1, run without a terminal time, accepts a trial at every iteration. Hence τn∈{ℓ1,…,ℓJ}\tau_n \in\{\ell_1,\ldots,\ell_J\} and

δn/τn≥ζ>0for every n.(34)\delta_n/\tau_n \ge\zeta> 0 \qquad\text{for every } n. \tag*{(34)}

(H4) With CrecC_{\mathrm{rec}} from (33),

Crecζ−1/2<1.(35)C_{\mathrm{rec}}\zeta^{-1/2} < 1. \tag*{(35)}

Remark 2. Conditions (H3)–(H4) are additional assumptions: (H3) requires a successful update window at every iteration, while (H4) requires ζ>Crec2\zeta> C_{\mathrm{rec}}^2. Neither follows from Assumption 1. Unlike [5], our bound on RN(n)R^{(n)}_N has no explicit decay in NN; the constant CRC_R contains cNc_N, which may grow with NN. Thus increasing NN does not necessarily improve (H4).

Theorem 1. Under Assumption 2, generate wn,βnw_n,\beta_n by (14), (20) and (21), from w0∈Vw_0 \in V, β1∈[α0,α1]\beta_1 \in[\alpha_0,\alpha_1], with t^n=tn+L\widehat{t}_n=t_n+L. There exists L0≥LsigL_0 \ge L_{\mathrm{sig}}, which may depend on w0w_0, such that every fixed L≥L0L \ge L_0 admits CL>0C_L > 0 and θL∈(0,1)\theta_L \in(0,1) satisfying

∣βn2−α2∣≤CLθLn−1,Eβn(wn(tn)−u(tn))1/2≤CLθLn−1.\lvert\beta_n^2-\alpha^2\rvert\le C_L\theta_L^{n-1}, \qquad E_{\beta_n}\left(w_n(t_n)-u(t_n)\right)^{1/2} \le C_L\theta_L^{n-1}.

Moreover, there exist constants C~L,σL>0\widetilde{C}_L,\sigma_L > 0 such that

∣βn2−α2∣+Eβn(wn(tn)−u(tn))1/2≤C~Le−σL(tn−t1),n≥1,\lvert\beta_n^2-\alpha^2\rvert+ E_{\beta_n}\left(w_n(t_n)-u(t_n)\right)^{1/2} \le\widetilde{C}_L e^{-\sigma_L(t_n-t_1)}, \qquad n \ge1,

where σL=∣log⁡θL∣/(L+τmax⁡)\sigma_L=\lvert\log\theta_L\rvert/(L+\tau_{\max}). In particular, βn→α\beta_n \to\alpha and wn(tn)→u(tn)w_n(t_n) \to u(t_n) in VV as n→∞n \to\infty.

Proof. Set an:=∣βn2−α2∣a_n := \lvert\beta_n^2-\alpha^2\rvert, bn:=Eβn(wn(tn)−u(tn))1/2b_n := E_{\beta_n}\left(w_n(t_n)-u(t_n)\right)^{1/2}, and

B∗:=(∥w0∥2+α12∥∇w0∥2)1/2+α1α0Mref.B_* := \left(\lVert w_0\rVert^2+\alpha_1^2\lVert\nabla w_0\rVert^2\right)^{1/2}+\frac{\alpha_1}{\alpha_0}M_{\mathrm{ref}}.

Since α,β1∈[α0,α1]\alpha,\beta_1 \in[\alpha_0,\alpha_1], the triangle inequality and (12) give

b1≤Eβ1(w0)1/2+Eβ1(u(t1))1/2≤Eα1(w0)1/2+α1α0Eα(u(t1))1/2≤B∗.b_1 \le E_{\beta_1}(w_0)^{1/2}+E_{\beta_1}(u(t_1))^{1/2} \le E_{\alpha_1}(w_0)^{1/2}+\frac{\alpha_1}{\alpha_0}E_\alpha(u(t_1))^{1/2}\le B_*.

Moreover, (21) ensures βn∈[α0,α1]\beta_n \in[\alpha_0,\alpha_1], so

an=∣βn2−α2∣≤α12−α02=P∗.a_n=\lvert\beta_n^2-\alpha^2\rvert\le\alpha_1^2-\alpha_0^2=P_*.

Applying the first estimate in (30) at the end of the assimilation interval, we obtain

Eβn(gn(tn+1))1/2≤e−κ(tn+1−tn)/2bn+K01/2an.E_{\beta_n}(g_n(t_{n+1}))^{1/2}\le e^{-\kappa(t_{n+1}-t_n)/2}b_n+K_0^{1/2}a_n.

We set Ceq=α1/α0C_{\mathrm{eq}}=\alpha_1/\alpha_0 and C1=CeqK01/2C_1=C_{\mathrm{eq}}K_0^{1/2}. Using Eβn+1≤Ceq2EβnE_{\beta_{n+1}}\le C_{\mathrm{eq}}^2E_{\beta_n} and tn+1−tn≥Lt_{n+1}-t_n\ge L, we conclude that

bn+1≤qLbn+C1an,qL:=Ceqe−κL/2.b_{n+1}\le q_Lb_n+C_1a_n,\qquad q_L:=C_{\mathrm{eq}}e^{-\kappa L/2}.

Choose Lpre≥LsigL_{\mathrm{pre}} \ge L_{\mathrm{sig}} sufficiently large that qL≤1/2q_L \le1/2 for every L≥LpreL \ge L_{\mathrm{pre}}. Iterating the recurrence and using an≤P∗a_n \le P_* gives

bn≤qLn−1B∗+C1P∗(1−qLn−1)1−qL≤Bbd:=max⁡{B∗,2C1P∗}.b_n \le q_L^{n-1} B_* + \frac{C_1P_*(1-q_L^{n-1})}{1-q_L} \le B_{\mathrm{bd}} := \max\{B_*,2C_1P_*\}.

We increase LpreL_{\mathrm{pre}}, if necessary, so that e−κLBbd2≤1e^{-\kappa L}B_{\mathrm{bd}}^2 \le1. The first estimate in (30) then gives, on every update interval InI_n,

Eβn(gn(s))≤e−κLBbd2+K0P∗2≤E∗.E_{\beta_n}(g_n(s)) \le e^{-\kappa L}B_{\mathrm{bd}}^2 + K_0P_*^2 \le E_*.

This bound makes the coefficient in the residual estimate uniform over the update intervals. Denoting E=Eβn(gn)≤E∗E=E_{\beta_n}(g_n)\le E_* and using ∥AQNgn∥≤∥Agn∥\lVert AQ_Ng_n\rVert\le\lVert Ag_n\rVert, we combine (32), Lemma 2, and (34), and apply Minkowski's inequality to obtain

an+1≤CR(1+E∗)ζ−1/2[(1τn∫InE(s) ds)1/2+(1τn∫In∥Agn(s)∥2 ds)1/2].a_{n+1}\le C_R(1+\sqrt{E_*})\zeta^{-1/2}\left[\left(\frac{1}{\tau_n}\int_{I_n}E(s)\,\mathrm{d}s\right)^{1/2}+\left(\frac{1}{\tau_n}\int_{I_n}\lVert Ag_n(s)\rVert^2\,\mathrm{d}s\right)^{1/2}\right].

Applying (30) over the update interval, since τn≥τ0\tau_n\ge\tau_0 and E(tn)1/2≤e−κL/2bn+K0anE(t_n)^{1/2}\le e^{-\kappa L/2}b_n+\sqrt{K_0}a_n, we have

(1τn∫InE(s) ds)1/2≤e−κL/2bn+K0an,\left(\frac{1}{\tau_n}\int_{I_n}E(s)\,\mathrm{d}s\right)^{1/2}\le e^{-\kappa L/2}b_n+\sqrt{K_0}a_n,
(1τn∫In∥Agn(s)∥2 ds)1/2≤CAe−κL/2bn+(CAK0+KA)an.\left(\frac{1}{\tau_n}\int_{I_n}\lVert Ag_n(s)\rVert^2\,\mathrm{d}s\right)^{1/2}\le C_Ae^{-\kappa L/2}b_n+(C_A\sqrt{K_0}+\sqrt{K_A})a_n.

Substituting these bounds into the preceding estimate for an+1a_{n+1} and using (33), we obtain

an+1≤Crecζ−1/2an+Crecζ−1/2e−κL/2bn.a_{n+1}\le C_{\mathrm{rec}}\zeta^{-1/2}a_n+C_{\mathrm{rec}}\zeta^{-1/2}e^{-\kappa L/2}b_n.

Combining this estimate with bn+1≤Ceqe−κL/2bn+C1anb_{n+1}\le C_{\mathrm{eq}}e^{-\kappa L/2}b_n+C_1a_n, we obtain the recurrence

TL:=(Crecζ−1/2Crecζ−1/2e−κL/2C1Ceqe−κL/2),(an+1bn+1)≤TL(anbn).T_L:=\begin{pmatrix} C_{\mathrm{rec}}\zeta^{-1/2} & C_{\mathrm{rec}}\zeta^{-1/2}e^{-\kappa L/2}\\ C_1 & C_{\mathrm{eq}}e^{-\kappa L/2} \end{pmatrix}, \qquad \begin{pmatrix}a_{n+1}\\ b_{n+1}\end{pmatrix} \le T_L\begin{pmatrix}a_n\\ b_n\end{pmatrix}.

Let ρ(TL)\rho(T_L) denote the spectral radius of TLT_L. As L→∞L\to\infty,

TL⟶(Crecζ−1/20C10).T_L\longrightarrow\begin{pmatrix} C_{\mathrm{rec}}\zeta^{-1/2} & 0\\ C_1 & 0 \end{pmatrix}.

The limiting matrix has eigenvalues Crecζ−1/2C_{\mathrm{rec}}\zeta^{-1/2} and 00. Hence, by continuity of the spectral radius and (H4),

ρ(TL)⟶Crecζ−1/2<1.\rho(T_L)\longrightarrow C_{\mathrm{rec}}\zeta^{-1/2}<1.

Thus ρ(TL)<1\rho(T_L)<1 for all sufficiently large LL. Choose L0≥LpreL_0\ge L_{\mathrm{pre}} so that this holds for every L≥L0L\ge L_0. By (H3) and the nonnegativity of TLT_L, induction yields

(anbn)≤TLn−1(a1b1).\begin{pmatrix}a_n\\ b_n\end{pmatrix} \le T_L^{n-1}\begin{pmatrix}a_1\\ b_1\end{pmatrix}.

For ρ(TL)<θL<1\rho(T_L)<\theta_L<1, the standard matrix-power estimate gives, for some Cθ,L>0C_{\theta,L}>0,

0≤an+bn≤Cθ,LθLn−1(a1+b1).0\le a_n+b_n\le C_{\theta,L}\theta_L^{n-1}(a_1+b_1).

This proves the geometric bounds. Moreover, since 0<θL<10 < \theta_L < 1, the preceding estimate gives an→0a_n \to0 and bn→0b_n \to0. By their definitions and α,βn≥α0>0\alpha, \beta_n \ge\alpha_0 > 0, we obtain

∣βn−α∣≤an2α0→0,∥∇(wn(tn)−u(tn))∥≤bnα0→0.|\beta_n-\alpha| \le\frac{a_n}{2\alpha_0} \to0,\qquad\|\nabla(w_n(t_n)-u(t_n))\| \le\frac{b_n}{\alpha_0} \to0.

Finally, since n−1≥(tn−t1)/(L+τmax⁡)n-1 \ge(t_n-t_1)/(L+\tau_{\max}), we have

θLn−1=e(n−1)log⁡θL≤exp⁡(−∣log⁡θL∣L+τmax⁡(tn−t1)).\theta_L^{n-1}=e^{(n-1)\log\theta_L}\le\exp\left(-\frac{|\log\theta_L|}{L+\tau_{\max}}(t_n-t_1)\right).

This gives exponential decay in tn−t1t_n-t_1 at the update times. For a fixed geometric factor, increasing LL decreases this rate, so a longer relaxation interval does not by itself guarantee faster convergence.

Conclusions

In this study, we have constructed a recursive algorithm for recovering the length-scale parameter in the modified Leray-α\alpha model. The synchronization estimate controls the velocity error, while the integrated dissipation estimate supplies the higher-order control needed for the analysis. With fixed NN, η\eta, a sufficiently large fixed relaxation length, uniform reference regularity, successful update windows, and the condition Crecζ−1/2<1C_{\mathrm{rec}}\zeta^{-1/2}<1, the parameter and velocity errors converge geometrically at the update times. If the update windows have a uniform upper bound on their lengths, the errors also decay exponentially in time at the update times.

The convergence result remains conditional on the signal hypotheses. Deriving these conditions from the reference trajectory remains open. Martinez, Strikwerda, and Wan [26] study excitation and viscosity identifiability for the two-dimensional Navier–Stokes equations using spectral observations and a nudging-based sensitivity analysis. A natural direction for future work is to investigate whether a similar analysis can provide verifiable sufficient conditions for (H3)–(H4). Moreover, the present analysis assumes exact observations available continuously in time. Extensions to discrete-time observations and noisy data are left for future work.

Acknowledgments

Jing Tian’s work was partially supported by the NSF LEAPS-MPS Grant #2316894.

References

  1. [1]D. A. F. Albanez, H. J. Nussenzveig Lopes, and E. S. Titi, Continuous data assimilation for the three-dimensional Navier–Stokes-α model, Asymptot. Anal. 97 (2016), 139–164.
  2. [2]D. A. F. Albanez and M. J. Benvenutti, Continuous data assimilation algorithm for simplified Bardina model, Evol. Equ. Control Theory 7 (2018), 33–52.DOI
  3. [3]D. A. F. Albanez, M. J. Benvenutti, S. Little, and J. Tian, Parameter analysis in continuous data assimilation for various turbulence models, Commun. Nonlinear Sci. Numer. Simul. 151 (2025), 109073.
  4. [4]D. A. F. Albanez, M. J. Benvenutti, S. Little, and J. Tian, Parameter error analysis for the 3D modified Leray-α model: analytical and numerical approaches, Comput. Appl. Math. 44 (2025), 200.arxiv.org/abs/2411.16324
  5. [5]D. A. F. Albanez, M. J. Benvenutti, and J. Tian, Parameter recovery for the simplified Bardina model via continuous data assimilation, Ann. Phys. (N.Y.) 493 (2026), 170612.DOI
  6. [6]A. Azouani, E. Olson, and E. S. Titi, Continuous data assimilation using general interpolant observables, J. Nonlinear Sci. 24 (2014), 277–304.DOI
  7. [7]J. Bardina, J. H. Ferziger, and W. C. Reynolds, Improved subgrid-scale models for large-eddy simulation, AIAA Paper 80-1357, 1980.DOI
  8. [8]A. Biswas and J. Hudson, Determining the viscosity of the Navier–Stokes equations from observations of finitely many modes, Inverse Problems 39 (2023), 125012.
  9. [9]C. Cao, D. D. Holm, and E. S. Titi, On the Clark-α model of turbulence: global regularity and long-time dynamics, J. Turbul. 6 (2005), N20.DOI
  10. [10]Y. Cao, E. M. Lunasin, and E. S. Titi, Global well-posedness of the three-dimensional viscous and inviscid simplified Bardina turbulence models, Commun. Math. Sci. 4 (2006), 823–848.DOI
  11. [11]E. Carlson, J. Hudson, and A. Larios, Parameter recovery for the 2 dimensional Navier–Stokes equations via continuous data assimilation, SIAM J. Sci. Comput. 42 (2020), A250–A270.DOI
  12. [12]E. Carlson and A. Larios, Sensitivity analysis for the 2D Navier–Stokes equations with applications to continuous data assimilation, J. Nonlinear Sci. 31 (2021), 84.arxiv.org/abs/2007.01860
  13. [13]E. Carlson, J. Hudson, A. Larios, V. R. Martinez, E. Ng, and J. P. Whitehead, Dynamically learning the parameters of a chaotic system using partial observations, Discrete Contin. Dyn. Syst. 42 (2022), 3809–3839.arxiv.org/abs/2108.08354
  14. [14]A. Cheskidov, D. D. Holm, E. Olson, and E. S. Titi, On a Leray-α model of turbulence, Proc. R. Soc. A 461 (2005), 629–649.DOI
  15. [15]A. Farhat, A. Larios, V. R. Martinez, and J. P. Whitehead, Identifying the body force from partial observations of a two-dimensional incompressible velocity field, Phys. Rev. Fluids 9 (2024), 054602.DOI
  16. [16]C. Foias and G. Prodi, Sur le comportement global des solutions non-stationnaires des équations de Navier–Stokes en dimension 2, Rend. Sem. Mat. Univ. Padova 39 (1967), 1–34.
  17. [17]C. Foias and R. Temam, Determination of the solutions of the Navier–Stokes equations by a set of nodal values, Math. Comp. 43 (1984), 117–133.DOI
  18. [18]C. Foias, D. D. Holm, and E. S. Titi, The Navier–Stokes-α model of fluid turbulence, Physica D 152–153 (2001), 505–519.
  19. [19]C. Foias, O. Manley, R. Rosa, and R. Temam, Navier–Stokes Equations and Turbulence, Cambridge University Press, Cambridge, 2001.DOI
  20. [20]A. A. Ilyin, E. M. Lunasin, and E. S. Titi, A modified-Leray-α subgrid scale model of turbulence, Nonlinearity 19 (2006), 879–897.DOI
  21. [21]D. A. Jones and E. S. Titi, Upper bounds on the number of determining modes, nodes, and volume elements for the Navier–Stokes equations, Indiana Univ. Math. J. 42 (1993), 875–887.
  22. [22]W. Layton and R. Lewandowski, On a well-posed turbulence model, Discrete Contin. Dyn. Syst. Ser. B 6 (2006), 111–128.DOI
  23. [23]J. Leray, Sur le mouvement d’un liquide visqueux emplissant l’espace, Acta Math. 63 (1934), 193–248.DOI
  24. [24]V. R. Martinez, Convergence analysis of a viscosity parameter recovery algorithm for the 2D Navier–Stokes equations, Nonlinearity 35 (2022), 2241–2287.
  25. [25]V. R. Martinez, J. Murri, and J. P. Whitehead, Relaxation-based schemes for on-the-fly parameter estimation in dissipative dynamical systems, Inverse Problems 41 (2025), 055001.arxiv.org/abs/2408.14296
  26. [26]V. R. Martinez, S. Strikwerda, and X. Wan, Excitation and identifiability in the 2D Navier–Stokes equations, arXiv preprint arXiv:2609.31230 (2026).arxiv.org/abs/2609.31230
  27. [27]J. Newey, J. P. Whitehead, and E. Carlson, Model discovery on the fly using continuous data assimilation, J. Comput. Phys. 537 (2025), 114121.arxiv.org/abs/2411.13561
  28. [28]B. Pachev, J. P. Whitehead, and S. A. McQuarrie, Concurrent multiparameter learning demonstrated on the Kuramoto–Sivashinsky equation, SIAM J. Sci. Comput. 44 (2022), A2974–A2990.DOI
  29. [29]L. G. Rebholz, J. Reyes, and J. P. Whitehead, Continuous data assimilation in steady Navier–Stokes equations with unknown viscosity: robust and efficient solvers and fast parameter recovery, arXiv preprint arXiv:2609.02862 (2026).arxiv.org/abs/2609.02862
  30. [30]R. Temam, Navier–Stokes Equations and Nonlinear Functional Analysis, 2nd ed., SIAM, Philadelphia, 1995.
  31. [31]Y. Wang and L. Yan, Data-driven operator inference for parameter estimation in nonlinear partial differential equations, J. Comput. Phys. 544 (2026), 114442.DOI
  32. [32]Y. Zhao, W. Liu, J. Zhang, G. Lv, and Y. Wang, Parameter recovery for the three-dimensional primitive equations with magnetic field via continuous data assimilation, ZAMM 106 (2026), e70534.DOI

Paper details

vlnc:W_68erm7xs
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents