betaFeedback

WorksMathematics

High-order long-time stable methods for the forced incompressible Navier–Stokes equations

Abstract

We present in this paper two scalar auxiliary variable (SAV) correction steps for constructing long-time stable schemes for the forced incompressible Navier–Stokes equations with no-slip boundary conditions. A main feature is that the same correction steps apply to any linear predictor satisfying the order and spatial regularity requirements in Assumption~P. Their construction and stability analysis are independent of the particular form of the predictor and require no energy estimate for it. The first correction uses a damped, shifted kinetic-energy variable to obtain a uniform-in-time L2L^2 bound in two and three dimensions, together with a weighted dissipation estimate. In two dimensions, the second correction combines a higher-energy auxiliary variable with a scalar relaxation to yield a uniform-in-time H1H^1 bound. Under the stated assumptions, these estimates hold on arbitrary positive time grids, without restrictions on step sizes or adjacent step ratios. We also derive eventual absorbing bounds; for the two-SAV scheme, these apply to the admissible initial class associated with fixed parameters. The generalized BDF schemes require one constant-coefficient Helmholtz solve per velocity component and one pressure Poisson solve per step; the second correction adds no field solve. Numerical experiments exhibit second-, third-, and fourth-order convergence for the tested BDF schemes and illustrate the long-time stability and efficiency of adaptive time stepping.

Introduction

We consider in this paper the construction and long-time stability analysis of efficient numerical schemes for the forced incompressible Navier–Stokes equations

ut−νΔu+(u⋅∇)u+∇p=f,in Ω×(0,∞),(1)\mathbf{u}_t-\nu\Delta\mathbf{u}+(\mathbf{u}\cdot\nabla)\mathbf{u}+\nabla p=\mathbf{f}, \qquad\text{in }\Omega\times(0,\infty), \tag*{(1)}
∇⋅u=0,in Ω×(0,∞).(2)\nabla\cdot\mathbf{u}=0, \qquad\text{in }\Omega\times(0,\infty). \tag*{(2)}

Here u\mathbf{u} and pp denote the velocity and pressure, respectively, ν>0\nu>0 is the kinematic viscosity, and f\mathbf{f} is an external force. We prescribe an initial velocity on a bounded connected domain Ω⊂Rd\Omega\subset\mathbb{R}^{d}, d=2,3d=2,3, with a sufficiently smooth boundary. For definiteness, we consider the no-slip boundary condition u=0\mathbf{u}=0 on ∂Ω\partial\Omega. The stability and absorption results also extend to the mean-zero periodic setting described in § 2. Throughout the paper, we assume

F∞:=sup⁡t≥0∥f(⋅,t)∥L2(Ω)<∞.(3)F_{\infty}:=\sup_{t\geq0}\|\mathbf{f}(\cdot,t)\|_{L^{2}(\Omega)}<\infty. \tag*{(3)}

For long-time simulations, it is desirable that a numerical scheme preserve the dissipative properties of the underlying equations. In the presence of a persistent force, the kinetic energy need not decrease, but the balance between forcing and viscous dissipation still yields an L2L^2 absorbing bound in two and three dimensions and an H1H^1 absorbing bound in two dimensions [3, 29, 22]. Unlike an all-time bound, which may depend on the initial state, an absorbing bound gives a ball that eventually contains the solution and whose radius is independent of that state [13]. Our aim is to construct linear, fully decoupled schemes that retain these properties and can be combined with high-order time discretizations.

Long-time stability of Navier–Stokes discretizations has been studied extensively. Early results include the convergence analysis of a linearized backward Euler scheme for the unforced two-dimensional problem [4], and uniform bounds and discrete-attractor convergence for a forced linear backward Euler scheme [15]. Other developments include implicit Euler and Crank–Nicolson methods [31, 30], direct and fractional-step algorithms [28], and nonlinear Galerkin and Runge–Kutta approximations [14, 24, 21].

For linearly extrapolated BDF2 schemes, Akbas, Kaya, and Rebholz [1] established all-time L2L^2 estimates, and Shiue [27] obtained a refined unconditional long-time L2L^2 bound for the two-dimensional problem. Rebholz and Tone [23] proved long-time H1H^1 stability in the velocity–pressure formulation under a mild time-step restriction. Unconditional long-time bounds were also obtained for first-order and IMEX–BDF2 finite-element schemes in the velocity–vorticity formulation [7]. For the periodic vorticity equation, semi-implicit multistep methods of order up to four allow fully explicit convection and constant-coefficient solves, but their long-time estimates require data-dependent time-step restrictions [5, 34, 2].

The scalar auxiliary variable (SAV) approach, introduced for gradient flows [25, 26], provides an effective way to treat nonlinear terms explicitly while retaining discrete energy control. SAV methods for incompressible flows include the auxiliary-energy-variable method of Lin, Yang, and Dong [18], which requires a nonlinear scalar solve, and the linear, decoupled pressure-correction methods of Li, Shen, and Liu [16]. High-order SAV schemes have been developed for periodic Navier–Stokes equations [9], nonperiodic problems, and rotating flows [35]. However, unconditional stability on a fixed time interval does not imply uniform-in-time boundedness for a forced problem. For example, the weak stability estimate in [10 Theorem 6] and the higher-order error estimates in [12 Theorem 4.1] involve constants that depend on the final time.

Recent SAV methods address the forced problem by introducing auxiliary damping. Han and Wang [6] constructed a SAV–BDF2 scheme with uniform-in-time velocity L2L^2 bounds in two and three dimensions and higher vorticity estimates in a two-dimensional free-slip setting. Lin and Shiue [19] refined the L2L^2 analysis to obtain a time-step-independent energy bound and to allow nonsmooth initial data through regularization. For the two-dimensional periodic vorticity equation, Wang, Wang, and Zhang [33, 32] developed second-order variable-step exponential time-differencing methods with a mean-reverting SAV. These methods require either a scalar cubic solve or a linear auxiliary recovery by numerical inverse Laplace transformation. We seek to combine auxiliary damping with high-order predictors and decoupled constant-coefficient solves, while imposing no time-step restriction in the long-time velocity estimates.

There are two main difficulties. The first is to obtain an H1H^1 bound for the no-slip problem. The provisional velocity is not exactly divergence free, and the pressure and convection terms must be controlled without the cancellations available in the periodic vorticity formulation. The second is to establish a uniform estimate on arbitrary positive time grids. On a nonuniform grid, shifting an index in a time average changes its weights, while large individual steps prevent the direct use of fixed-length discrete windows.

A main contribution of this paper is the construction of two scalar correction steps whose formulation and long-time stability analysis are independent of the particular linear predictor. The same corrections can be combined with any linear predictor satisfying Assumption P. Thus, the generalized BDF schemes presented below are particular realizations of a construction that also applies to other multistep and Runge–Kutta predictors satisfying this assumption.

We first introduce a damped, shifted kinetic-energy variable. Its scalar recursion gives an L2L^2 bound and a weighted dissipation estimate without using an energy estimate for the predictor. In two dimensions, we use the Stokes-pressure estimate of Liu, Liu, and Pego [20] and interpolation inequalities to construct a second scalar equation with a nonnegative residual. A scalar relaxation transfers the first dissipation estimate to the second auxiliary variable. We then apply a uniform Gronwall argument with right-endpoint weights and treat large steps directly. This yields a uniform H1H^1 bound without an upper time-step bound or an adjacent step-ratio condition.

The main contributions of this paper are as follows.

  1. We construct two SAV correction steps that can be combined with any linear predictor satisfying Assumption P. Under the stated hypotheses, their stability estimates hold on arbitrary positive time grids and require no energy estimate for the predictor. The scalar analysis applies to any fixed correction degree.

  1. For the correction step with one-SAV, we prove uniform-in-time velocity L2L^2 bounds in two and three dimensions. The estimate separates the decaying initial-data contribution from the forcing-dependent bound and yields an eventual L2L^2 absorbing radius (Theorem 3.3).

  1. For the correction step with two-SAV, we prove a uniform-in-time velocity H1H^1 bound in two dimensions. We also establish eventual absorbing bounds for fixed algorithmic parameters and their admissible initial class (Theorem 4.5 and Corollary 4.7).

The schemes are easy to implement. The generalized BDF schemes require one constant-coefficient Helmholtz solve for each velocity component and one pressure Poisson solve per step. The two-SAV extension adds only scalar operations and a velocity rescaling. Numerical tests exhibit high-order accuracy and illustrate the effect of the corrections on long-time stability and the efficiency of adaptive time stepping. The convergence order of each coupled scheme requires a separate analysis and depends on the predictor, the starting values, and the consistency of the SAV ratios.

The rest of this paper is organized as follows. Section 2 introduces the notation, energy balances, and linear predictor. We present and analyze the one-SAV and two-SAV schemes in Sections 3 and 4, respectively. Section 5 discusses implementation, temporal accuracy, and related methods. Numerical experiments are presented in Section 6, followed by concluding remarks in Section 7.

Preliminaries and a high-order linear predictor

We introduce some notation and recall the energy balances. We then present a generalized BDF predictor on a nonuniform time grid.

Notation and continuous energy balances

We denote by (⋅,⋅)(\cdot,\cdot) and ∥⋅∥\|\cdot\| the inner product and norm in L2(Ω)L^2(\Omega), and use the same notation for scalar and vector function spaces. Set

V={v∈H01(Ω):∇⋅v=0}.V=\{v\in H^1_0(\Omega):\nabla\cdot v=0\}.

There is a Poincaré constant cP>0c_P>0 such that

∥∇v∥2≥cP∥v∥2,v∈H01(Ω).(4)\|\nabla v\|^2\ge c_P\|v\|^2,\qquad v\in H^1_0(\Omega). \tag*{(4)}

For periodic boundary conditions, we assume that the forcing and initial velocity histories have zero mean, apply

Π0v:=v−∣Ω∣−1∫Ωv dx\Pi_0v:=v-|\Omega|^{-1}\int_{\Omega}v\,dx

to the right-hand side of each velocity predictor, and take the pressure in Hper1(Ω)/RH^1_{\mathrm{per}}(\Omega)/\mathbb{R}. The scalar corrections preserve zero mean, so (2.1) holds for the periodic iterates.

Taking the inner product of (1.1a) with uu gives

ddtE(u)+ν∥∇u∥2=(f,u),E(v):=12∥v∥2.(5)\frac{d}{dt}E(u)+\nu\|\nabla u\|^2=(f,u),\qquad E(v):=\frac{1}{2}\|v\|^2. \tag*{(5)}

In two dimensions, taking the inner product with −Δu-\Delta u gives

ddtK(u)=−D2(u,p,f),K(v):=12∥∇v∥2,(6)\frac{d}{dt}K(u)=-D_2(u,p,f),\qquad K(v):=\frac{1}{2}\|\nabla v\|^2, \tag*{(6)}

where

D2(v,q,f):=ν∥Δv∥2+(f,Δv)−(∇q,Δv)−((v⋅∇)v,Δv).(7)D_2(v,q,f):=\nu\|\Delta v\|^2+(f,\Delta v)-(\nabla q,\Delta v)-((v\cdot\nabla)v,\Delta v). \tag*{(7)}

We introduce positive shifts CE,CK>0C_E,C_K>0 and define

E(v):=E(v)+CE,K(v):=K(v)+CK.(8)\mathcal{E}(v):=E(v)+C_E,\qquad\mathcal{K}(v):=K(v)+C_K. \tag*{(8)}

The shifts keep the energy ratios below well defined without changing the continuous energy balances. Unless otherwise stated, the stability constants may depend on Ω\Omega, ν\nu, the correction degree kk, the forcing bound, and the fixed algorithmic parameters, but not on the final time or the time grid. Dependence on the initial data is stated separately.

A flexible high-order predictor step

Let

0=t0<t1<⋯ ,τn+1=tn+1−tn>0.(9)0=t_0<t_1<\cdots,\qquad\tau_{n+1}=t_{n+1}-t_n>0. \tag*{(9)}

For a fixed T>0T>0, take tNT=Tt_{N_T}=T; for absorption, assume tn→∞t_n\to\infty. On a constant grid, τn+1=δt\tau_{n+1}=\delta t. The scalar estimates in late sections apply to any well-defined linear predictor with newly computed provisional velocity. We only need to make the following assumption on the linear predictor.

Assumption P. For any n>0n>0, the linear predictor is of order kk for some k≥1k\ge1 and produces provisional velocity u‾n+1∈H2(Ω)\overline{u}^{n+1}\in H^2(\Omega).

Examples include linearly implicit multistep and Runge–Kutta methods. As a particular example, we use the generalized BDF formulas of [11] with extrapolated nonlinear term and pressure as the predictor. More precisely, for an order k≥2k\ge2 and a shift β≥1\beta\ge1, define

Ak,nβ(ϕn+1)=∑q=0kak,qn(β)ϕn+1−k+q,A_{k,n}^{\beta}(\phi^{n+1})=\sum_{q=0}^{k}a_{k,q}^{n}(\beta)\phi^{n+1-k+q},
Bk,nβ(ϕn+1)=∑q=0k−1bk,qn(β)ϕn+2−k+q,B_{k,n}^{\beta}(\phi^{n+1})=\sum_{q=0}^{k-1}b_{k,q}^{n}(\beta)\phi^{n+2-k+q},
Ck,nβ(ϕn)=∑q=0k−1ck,qn(β)ϕn+1−k+q.C_{k,n}^{\beta}(\phi^{n})=\sum_{q=0}^{k-1}c_{k,q}^{n}(\beta)\phi^{n+1-k+q}.

Polynomial interpolation at the time nodes determines the coefficients. The operator Ak,nβ/τn+1A_{k,n}^{\beta}/\tau_{n+1} approximates the derivative at tn+βτn+1t_n+\beta\tau_{n+1}, while Bk,nβB_{k,n}^{\beta} and Ck,nβC_{k,n}^{\beta} approximate the solution there to order kk. On a constant grid these reduce to the formulas in [11]; β=1\beta=1 gives classical BDF. With fn+β:=f(tn+βτn+1)f^{n+\beta}:=f(t_n+\beta\tau_{n+1}), set

Ak,nβ(u‾n+1)τn+1−νΔBk,nβ(u‾n+1)+(Ck,nβ(un)⋅∇)Ck,nβ(un)+∇Ck,nβ(pn)=fn+β.(10)\frac{A_{k,n}^{\beta}(\overline{u}^{n+1})}{\tau_{n+1}}-\nu\Delta B_{k,n}^{\beta}(\overline{u}^{n+1})+\left(C_{k,n}^{\beta}(u^n)\cdot\nabla\right)C_{k,n}^{\beta}(u^n)+\nabla C_{k,n}^{\beta}(p^n)=f^{n+\beta}. \tag*{(10)}

Only u‾n+1\overline{u}^{n+1} is unknown. To verify solvability, set t∗:=tn+βτn+1t_*:=t_n+\beta\tau_{n+1}. The leading Lagrange basis functions for the derivative and diffusion interpolants are

LA(t)=∏i=n+1−knt−titn+1−ti,LB(t)=∏i=n+2−knt−titn+1−ti.L_A(t)=\prod_{i=n+1-k}^{n}\frac{t-t_i}{t_{n+1}-t_i},\qquad L_B(t)=\prod_{i=n+2-k}^{n}\frac{t-t_i}{t_{n+1}-t_i}.

Since t∗≥tn+1>tit_*\ge t_{n+1}>t_i, all factors are positive and

ak,kn(β)=τn+1LA(t∗)∑i=n+1−kn1t∗−ti>0,bk,k−1n(β)=LB(t∗)>0.a_{k,k}^{n}(\beta)=\tau_{n+1}L_A(t_*)\sum_{i=n+1-k}^{n}\frac{1}{t_*-t_i}>0,\qquad b_{k,k-1}^{n}(\beta)=L_B(t_*)>0.

Hence (2.7) has the form (aI−bΔ)u‾n+1=g(aI-b\Delta)\overline{u}^{n+1}=g with a,b>0a,b>0, and is uniquely solvable. For an L2L^2 right-hand side, elliptic regularity gives u‾n+1∈H2(Ω)\overline{u}^{n+1}\in H^2(\Omega) under the stated domain assumptions.

We use a well-defined linear starting procedure and apply the SAV updates at every starting step. The derivative and diffusion terms use the provisional-velocity histories, while the explicit convection uses the corrected histories. High-order convergence requires order-compatible starting values and the stability conditions of the chosen predictor.

A one-SAV scheme with a uniform L2L^2 bound

We construct a one-SAV scheme and prove a uniform L2L^2 bound and a weighted H1H^1 dissipation estimate for the corrected velocity.

Shifted L2L^2 auxiliary variable

Set r(t)=E(u(t))r(t)=\mathcal{E}(u(t)). For parameters α1,β1>0\alpha_1,\beta_1>0, define

Q1(v,g):=ν∥∇v∥2−(g,v)−β1E(v)+α1∥g∥2.(11)Q_1(\mathbf{v},g):=\nu\lVert\nabla\mathbf{v}\rVert^2-(g,\mathbf{v})-\beta_1\mathcal{E}(\mathbf{v})+\alpha_1\lVert g\rVert^2. \tag*{(11)}

The scalar equation

rt+β1r=−rE(u)Q1(u,f)+α1∥f∥2+β1CE(12)r_t+\beta_1r=-\frac{r}{\mathcal{E}(u)}Q_1(u,\mathbf{f})+\alpha_1\lVert\mathbf{f}\rVert^2+\beta_1C_E \tag*{(12)}

is equivalent to (5) when r=E(u)r=\mathcal{E}(u). We shall choose α1\alpha_1 and β1\beta_1 so that Q1Q_1 is nonnegative. We propose the following scheme.

Algorithm 1. Given unu^n, pnp^n, rnr^n and the required history, perform the following steps.

Step 1: velocity predictor. Compute u‾n+1\overline{u}^{n+1} from

Ak,nβ(u‾n+1)τn+1−νΔBk,nβ(u‾n+1)+(Ck,nβ(un)⋅∇)Ck,nβ(un)+∇Ck,nβ(pn)=fn+β,(13)\frac{A_{k,n}^{\beta}(\overline{u}^{n+1})}{\tau_{n+1}}-\nu\Delta B_{k,n}^{\beta}(\overline{u}^{n+1})+\left(C_{k,n}^{\beta}(u^n)\cdot\nabla\right)C_{k,n}^{\beta}(u^n)+\nabla C_{k,n}^{\beta}(p^n)=f^{n+\beta}, \tag*{(13)}

with the prescribed boundary condition. Other linear predictors from Section 2.2 may replace (13).

Step 2: shifted SAV correction. Evaluate

Q1n+1:=ν∥∇u‾n+1∥2−(fn+1,u‾n+1)−β1E(u‾n+1)+α1∥fn+1∥2,(14)Q_1^{n+1}:=\nu\lVert\nabla\overline{u}^{n+1}\rVert^2-(f^{n+1},\overline{u}^{n+1})-\beta_1\mathcal{E}(\overline{u}^{n+1})+\alpha_1\lVert f^{n+1}\rVert^2, \tag*{(14)}

and compute rn+1r^{n+1} from

rn+1−rnτn+1+β1rn+1=−rn+1E(u‾n+1)Q1n+1+α1∥fn+1∥2+β1CE,(15)\frac{r^{n+1}-r^n}{\tau_{n+1}}+\beta_1r^{n+1}=-\frac{r^{n+1}}{\mathcal{E}(\overline{u}^{n+1})}Q_1^{n+1}+\alpha_1\lVert f^{n+1}\rVert^2+\beta_1C_E, \tag*{(15)}

or, equivalently,

rn+1=rn+τn+1(α1∥fn+1∥2+β1CE)1+τn+1β1+τn+1Q1n+1/E(u‾n+1).(16)r^{n+1}=\frac{r^n+\tau_{n+1}\left(\alpha_1\lVert f^{n+1}\rVert^2+\beta_1C_E\right)}{1+\tau_{n+1}\beta_1+\tau_{n+1}Q_1^{n+1}/\mathcal{E}(\overline{u}^{n+1})}. \tag*{(16)}

Define

ξn+1:=rn+1E(u‾n+1),ηkn+1:=1−(1−ξn+1)k,u~n+1:=ηkn+1u‾n+1.(17)\xi^{n+1}:=\frac{r^{n+1}}{\mathcal{E}(\overline{u}^{n+1})},\qquad\eta_k^{n+1}:=1-(1-\xi^{n+1})^k,\qquad\widetilde{u}^{n+1}:=\eta_k^{n+1}\overline{u}^{n+1}. \tag*{(17)}

Step 3: pressure recovery. Find pn+1∈H1(Ω)/Rp^{n+1}\in H^1(\Omega)/\mathbb{R} such that

(∇pn+1,∇q)=(fn+1−(u~n+1⋅∇)u~n+1−ν∇×∇×u~n+1,∇q),∀q∈H1(Ω)/R.(18)(\nabla p^{n+1},\nabla q)=(f^{n+1}-(\widetilde{u}^{n+1}\cdot\nabla)\widetilde{u}^{n+1}-\nu\nabla\times\nabla\times\widetilde{u}^{n+1},\nabla q),\qquad\forall q\in H^1(\Omega)/\mathbb{R}. \tag*{(18)}

Equation (18) is a constant-coefficient Poisson problem. Set un+1=u~n+1u^{n+1}=\widetilde{u}^{n+1}.

Remark 3.1 (Order-preserving correction). On a fixed interval [0,T][0,T], let hmax⁡=max⁡1≤j≤NTτjh_{\max}=\max_{1\le j\le N_T}\tau_j. The polynomial ηk(x)=1−(1−x)k\eta_k(x)=1-(1-x)^k satisfies ηk(1)=1\eta_k(1)=1 and ηk(1+O(hmax⁡))=1+O(hmax⁡k)\eta_k(1+O(h_{\max}))=1+O(h_{\max}^k). Thus the correction is compatible with an order-kk predictor, provided that ξn+1=1+O(hmax⁡)\xi^{n+1}=1+O(h_{\max}).

Uniform L2L^2 stability

Assume

α1≥2νCP,0<β1≤νCP4.(19)\alpha_1\ge\frac{2}{\nu C_P},\qquad0<\beta_1\le\frac{\nu C_P}{4}. \tag*{(19)}

We shall use the following damped scalar recursion (cf. [24, Lemma 1]).

Lemma 3.2 (Damped scalar recursion). Let τj>0\tau_j>0, λ>0\lambda>0, and let aj,bja^j,b^j be nonnegative with bj≤bb^j\le b. If

aj+1−ajτj+1+λaj+1≤bj,(20)\frac{a^{j+1}-a^j}{\tau_{j+1}}+\lambda a^{j+1}\le b^j, \tag*{(20)}

then, with Θn(λ):=∏j=1n(1+λτj)−1\Theta_n(\lambda):=\prod_{j=1}^{n}(1+\lambda\tau_j)^{-1} and Θ0(λ)=1\Theta_0(\lambda)=1,

an≤Θn(λ)a0+bλ(1−Θn(λ)).(21)a^n\le\Theta_n(\lambda)a^0+\frac{b}{\lambda}(1-\Theta_n(\lambda)). \tag*{(21)}

Moreover, Θn(λ)≤(1+λtn)−1\Theta_n(\lambda) \le(1+\lambda t_n)^{-1}.

Proof. Set qj=(1+λτj)−1q_j=(1+\lambda\tau_j)^{-1}. Then aj+1≤qj+1aj+(1−qj+1)b/λa^{j+1}\le q_{j+1}a^j+(1-q_{j+1})b/\lambda. Subtract b/λb/\lambda and iterate to obtain (3.11). The last assertion follows from ∏j=1n(1+λτj)≥1+λ∑j=1nτj\prod_{j=1}^n(1+\lambda\tau_j)\ge1+\lambda\sum_{j=1}^n\tau_j. □

Define

B1:=α1F∞2+β1CE,R∞:=B1/β1,CR:=max⁡{r0,R∞}.(22)B_1:=\alpha_1F_\infty^2+\beta_1C_E,\qquad R_\infty:=B_1/\beta_1,\qquad C_R:=\max\{r^0,R_\infty\}. \tag*{(22)}

For R>0R>0, define the nondecreasing function

Cη,k(R):=max⁡0≤x≤R/CEx∣Pk−1(x)∣2,Pk−1(x):=∑ℓ=0k−1(1−x)ℓ.(23)C_{\eta,k}(R):=\max_{0\le x\le R/C_E}x|P_{k-1}(x)|^2,\qquad P_{k-1}(x):=\sum_{\ell=0}^{k-1}(1-x)^\ell. \tag*{(23)}

The following estimates also hold for the first correction in Algorithm 2.

Theorem 3.3 (All-time L2L^2 bounds and velocity absorption). Let d=2d=2 or 33 and assume (1.2), (3.9), and the predictor satisfies Assumption P. Initialize r0=E(u0)>0r^0=\mathcal{E}(u^0)>0. Then, on any positive time grid, Algorithm 1 satisfies

0<rn≤R∞+(r0−R∞)+Θn(β1)≤CR,n≥0,(24)0<r^n\le R_\infty+(r^0-R_\infty)_+\Theta_n(\beta_1)\le C_R,\qquad n\ge0, \tag*{(24)}

where x+:=max⁡{x,0}x_+:=\max\{x,0\}. Furthermore,

sup⁡n≥0∥un∥2≤max⁡{∥u0∥2,[2Cη,k(CR)CR]1/2}=:ML2.(25)\sup_{n\ge0}\|\mathbf{u}^n\|^2\le\max\{\|\mathbf{u}^0\|^2,[2C_{\eta,k}(C_R)C_R]^{1/2}\}=:M_{L^2}. \tag*{(25)}

If rj≤Rr^j\le R for j≥mj\ge m, then for every n>mn>m,

∑j=mn−1τj+1K(u~j+1)≤Cη,k(R)ν{rm+B1(tn−tm)}≤AR(tn−tm),(26)\sum_{j=m}^{n-1}\tau_{j+1}K(\widetilde{u}^{j+1})\le\frac{C_{\eta,k}(R)}{\nu}\{r^m+B_1(t_n-t_m)\}\le A_R(t_n-t_m), \tag*{(26)}
AR(L):=Cη,k(R)ν(R+B1L).A_R(L):=\frac{C_{\eta,k}(R)}{\nu}(R+B_1L).

If tn→∞t_n\to\infty, then for each ε>0\varepsilon>0 the velocity eventually lies in the L2L^2 ball with radius

ML2,abs(ε):=[2Cη,k(R∞+ε)(R∞+ε)]1/2.(27)M_{L^2,\mathrm{abs}}(\varepsilon):=[2C_{\eta,k}(R_\infty+\varepsilon)(R_\infty+\varepsilon)]^{1/2}. \tag*{(27)}

The radius is independent of the initial data and time grid. Entry is uniform over bounded sets of initial energies.

Proof. Young’s inequality gives

∣(fn+1,u‾n+1)∣≤νCP8∥u‾n+1∥2+2νCP∥fn+1∥2.|(f^{n+1},\overline{u}^{n+1})|\le\frac{\nu C_P}{8}\|\overline{u}^{n+1}\|^2+\frac{2}{\nu C_P}\|f^{n+1}\|^2.

Using (3.9) and the Poincaré inequality yields

Q1n+1≥ν2∥∇u‾n+1∥2+νCP4∥u‾n+1∥2≥0.(28)Q_1^{n+1}\ge\frac{\nu}{2}\|\nabla\overline{u}^{n+1}\|^2+\frac{\nu C_P}{4}\|\overline{u}^{n+1}\|^2\ge0. \tag*{(28)}

The denominator in (3.6) is at least 1+β1τn+11+\beta_1\tau_{n+1}, so r0>0r^0>0 implies rn>0r^n>0 by induction. It follows from (3.5) that

rn+1−rnτn+1+β1rn+1≤α1∥fn+1∥2+β1CE≤B1.(29)\frac{r^{n+1}-r^n}{\tau_{n+1}}+\beta_1r^{n+1}\le\alpha_1\|f^{n+1}\|^2+\beta_1C_E\le B_1. \tag*{(29)}

Lemma 3.2 proves (3.14).

To estimate the corrected velocity, we use the identity

ηk(x)=1−(1−x)k=xPk−1(x)(30)\eta_k(x)=1-(1-x)^k=xP_{k-1}(x) \tag*{(30)}

which implies ∣ηk(x)∣2≤Cη,k(R)x|\eta_k(x)|^2\le C_{\eta,k}(R)x for 0≤x≤R/CE0\le x\le R/C_E. Whenever rn+1≤Rr^{n+1}\le R,

∥u~n+1∥2≤Cη,k(R)rn+1E(u‾n+1)∥u‾n+1∥2≤2Cη,k(R)rn+1.(31)\|\widetilde{u}^{n+1}\|^2\le C_{\eta,k}(R)\frac{r^{n+1}}{\mathcal{E}(\overline{u}^{n+1})}\|\overline{u}^{n+1}\|^2\le2C_{\eta,k}(R)r^{n+1}. \tag*{(31)}

Taking R=CRR=C_R proves (25). For the pressure estimates below, set

M0:=max⁡{∥u0∥,[2Cη,k(CR)CR]1/2}.M_0 := \max\left\{\lVert u^0\rVert,\left[2C_{\eta,k}(C_R)C_R\right]^{1/2}\right\}.

Multiplying (15) at step jj by τj+1\tau_{j+1}, summing over j=m,…,n−1j=m,\ldots,n-1, and using (28), we find

ν2∑j=mn−1τj+1ξj+1∥∇u~j+1∥2≤rm+B1(tn−tm).\frac{\nu}{2}\sum_{j=m}^{n-1}\tau_{j+1}\xi^{j+1}\lVert\nabla\widetilde{u}^{j+1}\rVert^2\le r^m+B_1(t_n-t_m).

Since

K(u~j+1)≤Cη,k(R)2ξj+1∥∇u~j+1∥2,(32)K(\widetilde{u}^{j+1})\le\frac{C_{\eta,k}(R)}{2}\xi^{j+1}\lVert\nabla\widetilde{u}^{j+1}\rVert^2, \tag*{(32)}

we obtain (26).

For the absorbing bound, let Ra=R∞+εR_a=R_\infty+\varepsilon. Lemma 3.2 gives

rn≤R∞+(r0−R∞)+1+β1tn.(33)r^n\le R_\infty+\frac{(r^0-R_\infty)_+}{1+\beta_1t_n}. \tag*{(33)}

Choose an index nan_a such that

tna≥1β1((r0−R∞)+ε−1)+.(34)t_{n_a}\ge\frac{1}{\beta_1}\left(\frac{(r^0-R_\infty)_+}{\varepsilon}-1\right)_+. \tag*{(34)}

Then rn≤Rar^n\le R_a for every n≥nan\ge n_a, and (31) with R=RaR=R_a proves the eventual bound for n≥max⁡{1,na}n\ge\max\{1,n_a\}. Replacing r0r^0 in (34) by a bound over an initial set makes the entry criterion uniform over that set. □

Remark 3.4 (Transient bounds and absorbing radii). The all-time bound ML2M_{L^2} depends on the initial energy through CRC_R; the absorbing radius (27) is independent of it (cf. [13, 14]). If τj≤τmax⁡<∞\tau_j\le\tau_{\max}<\infty, then log⁡(1+x)≥x/(1+x)\log(1+x)\ge x/(1+x) gives

Θn(β1)≤exp⁡(−β1tn1+β1τmax⁡).\Theta_n(\beta_1)\le\exp\left(-\frac{\beta_1t_n}{1+\beta_1\tau_{\max}}\right).

Thus the initial contribution decays exponentially when the steps are bounded, while (33) holds on arbitrary positive grids.

A two-SAV scheme with a uniform H1H^1 bound in two dimensions

In this section, we assume that d=2d=2 and the domain has a C3C^3 boundary. We introduce a second SAV and establish a uniform-in-time H1H^1 bound by combining the weighted dissipation estimate for the first SAV with the Stokes-pressure estimate of [20], two-dimensional interpolation inequalities, and a discrete uniform Gronwall argument. The scheme uses the velocity and pressure computed in Algorithm 1 and requires no additional Helmholtz or Poisson solve.

Shifted H1H^1 auxiliary variable

Let s(t)=K(u(t))s(t)=K(u(t)) and choose α2,β2,γ2>0\alpha_2,\beta_2,\gamma_2>0. Define

G2(v):=2β2+2γ2∥∇v∥2,H2(f):=α2∥f∥2,(35)G_2(v):=2\beta_2+2\gamma_2\lVert\nabla v\rVert^2,\qquad H_2(f):=\alpha_2\lVert f\rVert^2, \tag*{(35)}

so that

G2(v)K(v)=β2∥∇v∥2+γ2∥∇v∥4.G_2(v)K(v)=\beta_2\lVert\nabla v\rVert^2+\gamma_2\lVert\nabla v\rVert^4.

Using (7), set

Q2(v,q,f):=D2(v,q,f)+G2(v)K(v)+H2(f),(36)Q_2(v,q,f):=D_2(v,q,f)+G_2(v)K(v)+H_2(f), \tag*{(36)}
Q^2(v,q,f):=Q2(v,q,f)+CKG2(v).(37)\widehat{Q}_2(v,q,f):=Q_2(v,q,f)+C_KG_2(v). \tag*{(37)}

Then

st=−sK(u)Q^2(u,p,f)+G2(u)s+H2(f)(38)s_t=-\frac{s}{K(u)}\widehat{Q}_2(u,p,f)+G_2(u)s+H_2(f) \tag*{(38)}

reduces to (6) when s=K(u)s=K(u), since the additional terms cancel and the right-hand side equals −D2(u,p,f)-D_2(u,p,f). Sufficient parameter conditions for a nonnegative discrete residual will be given below.

Algorithm 2: two SAVs

Given un\mathbf{u}^{n}, pnp^{n}, rnr^{n}, sns^{n} and the required histories, perform the following steps.

Steps 1–3. Compute the velocity predictor, first SAV correction, and pressure as in Algorithm 1.

Step 4: recursive higher-energy update. Define

KCn+1:=K(u~n+1)=12∥∇u~n+1∥2+CK,(39)\mathcal{K}_{C}^{n+1}:=\mathcal{K}(\widetilde{\mathbf{u}}^{n+1})=\frac{1}{2}\lVert\nabla\widetilde{\mathbf{u}}^{n+1}\rVert^{2}+C_{K}, \tag*{(39)}
G2n+1:=2β2+2γ2∥∇u~n+1∥2,(40)G_{2}^{n+1}:=2\beta_{2}+2\gamma_{2}\lVert\nabla\widetilde{\mathbf{u}}^{n+1}\rVert^{2}, \tag*{(40)}
H2n+1:=α2∥fn+1∥2,(41)H_{2}^{n+1}:=\alpha_{2}\lVert\mathbf{f}^{n+1}\rVert^{2}, \tag*{(41)}

and

Q2n+1:=ν∥Δu~n+1∥2+(fn+1,Δu~n+1)−(∇pn+1,Δu~n+1)−(u~n+1⋅∇)u~n+1,Δu~n+1)+β2∥∇u~n+1∥2+γ2∥∇u~n+1∥4+α2∥fn+1∥2,(42)Q_{2}^{n+1}:=\nu\lVert\Delta\widetilde{\mathbf{u}}^{n+1}\rVert^{2}+(\mathbf{f}^{n+1},\Delta\widetilde{\mathbf{u}}^{n+1})-(\nabla p^{n+1},\Delta\widetilde{\mathbf{u}}^{n+1})-(\widetilde{\mathbf{u}}^{n+1}\cdot\nabla)\widetilde{\mathbf{u}}^{n+1},\Delta\widetilde{\mathbf{u}}^{n+1})+\beta_{2}\lVert\nabla\widetilde{\mathbf{u}}^{n+1}\rVert^{2}+\gamma_{2}\lVert\nabla\widetilde{\mathbf{u}}^{n+1}\rVert^{4}+\alpha_{2}\lVert\mathbf{f}^{n+1}\rVert^{2}, \tag*{(42)}
Q^2n+1:=Q2n+1+CKG2n+1.(43)\widehat{Q}_{2}^{n+1}:=Q_{2}^{n+1}+C_{K}G_{2}^{n+1}. \tag*{(43)}

Compute a provisional scalar S^n+1\widehat{S}^{n+1} from

S^n+1−snτn+1=−S^n+1KCn+1Q^2n+1+G2n+1sn+H2n+1,(44)\frac{\widehat{S}^{n+1}-s^{n}}{\tau_{n+1}}=-\frac{\widehat{S}^{n+1}}{\mathcal{K}_{C}^{n+1}}\widehat{Q}_{2}^{n+1}+G_{2}^{n+1}s^{n}+H_{2}^{n+1}, \tag*{(44)}

or, explicitly,

S^n+1=(1+τn+1G2n+1)sn+τn+1H2n+11+τn+1Q^2n+1/KCn+1.(45)\widehat{S}^{n+1}=\frac{(1+\tau_{n+1}G_{2}^{n+1})s^{n}+\tau_{n+1}H_{2}^{n+1}}{1+\tau_{n+1}\widehat{Q}_{2}^{n+1}/\mathcal{K}_{C}^{n+1}}. \tag*{(45)}

Step 5: scalar relaxation and final correction. Set

sn+1:=min⁡{S^n+1,KCn+1},(46)s^{n+1}:=\min\{\widehat{S}^{n+1},\mathcal{K}_{C}^{n+1}\}, \tag*{(46)}

and define

ζn+1:=sn+1KCn+1,λkn+1:=1−(1−ζn+1)k,un+1:=λkn+1u~n+1.(47)\zeta^{n+1}:=\frac{s^{n+1}}{\mathcal{K}_{C}^{n+1}},\qquad\lambda_{k}^{n+1}:=1-(1-\zeta^{n+1})^{k},\qquad\mathbf{u}^{n+1}:=\lambda_{k}^{n+1}\widetilde{\mathbf{u}}^{n+1}. \tag*{(47)}

The relaxation (46) bounds the new scalar by the shifted energy of u~n+1\widetilde{\mathbf{u}}^{n+1} and thus transfers the time-averaged estimate from the first SAV to sns^{n}. Together with the recursion (44), this allows us to apply a uniform Gronwall argument. The pressure is recovered before the second rescaling so that the residual (42) uses the pair (u~n+1,pn+1)(\widetilde{\mathbf{u}}^{n+1},p^{n+1}).

Positivity of the higher-energy residual

Lemma 4.1 (Two-dimensional pressure and convection estimates). Let v∈H2(Ω)∩H01(Ω)\mathbf{v}\in H^{2}(\Omega)\cap H_{0}^{1}(\Omega) and let q∈H1(Ω)/Rq\in H^{1}(\Omega)/\mathbb{R} satisfy

(∇q,∇φ)=(f−(v⋅∇)v−ν∇×∇×v,∇φ),∀φ∈H01(Ω).(48)(\nabla q,\nabla\varphi)=(\mathbf{f}-(\mathbf{v}\cdot\nabla)\mathbf{v}-\nu\nabla\times\nabla\times\mathbf{v},\nabla\varphi),\qquad\forall\varphi\in H_{0}^{1}(\Omega). \tag*{(48)}

Assume ∥v∥≤M0\lVert\mathbf{v}\rVert\le M_{0}. For every sufficiently small ε>0\varepsilon>0, there exist constants Cf,C1,C2,C3,C4>0C_{f},C_{1},C_{2},C_{3},C_{4}>0, independent of v\mathbf{v}, f\mathbf{f}, and the time grid, such that

∣(∇q,Δv)∣≤ν(34+32ε)∥Δv∥2+Cf∥f∥2+C1∥∇v∥2+C2M02∥∇v∥4,(49)\lvert(\nabla q,\Delta\mathbf{v})\rvert\le\nu\left(\frac{3}{4}+\frac{3}{2}\varepsilon\right)\lVert\Delta\mathbf{v}\rVert^{2}+C_{f}\lVert\mathbf{f}\rVert^{2}+C_{1}\lVert\nabla\mathbf{v}\rVert^{2}+C_{2}M_{0}^{2}\lVert\nabla\mathbf{v}\rVert^{4}, \tag*{(49)}
∣((v⋅∇)v,Δv)∣≤νε∥Δv∥2+C4M02∥∇v∥4,(50)\lvert((\mathbf{v}\cdot\nabla)\mathbf{v},\Delta\mathbf{v})\rvert\le\nu\varepsilon\lVert\Delta\mathbf{v}\rVert^{2}+C_{4}M_{0}^{2}\lVert\nabla\mathbf{v}\rVert^{4}, \tag*{(50)}
∣(f,Δv)∣≤νε∥Δv∥2+C3∥f∥2.(51)\lvert(\mathbf{f},\Delta\mathbf{v})\rvert\le\nu\varepsilon\lVert\Delta\mathbf{v}\rVert^{2}+C_{3}\lVert\mathbf{f}\rVert^{2}. \tag*{(51)}

The constants may depend on Ω\Omega, ν\nu, and ε\varepsilon.

Proof. Let P\mathbb{P} be the L2L^2 Leray projection and set Q=I−P\mathbb{Q}=I-\mathbb{P} and N(v)=(v⋅∇)vN(\boldsymbol{v})=(\boldsymbol{v}\cdot\nabla)\boldsymbol{v}. It follows from (48) that

∇q=Q(f−N(v))+ν∇pS,∇pS=Q(Δv−∇∇⋅v).\nabla q=\mathbb{Q}(\boldsymbol{f}-N(\boldsymbol{v}))+\nu\nabla p_S,\qquad\nabla p_S=\mathbb{Q}(\Delta\boldsymbol{v}-\nabla\nabla\cdot\boldsymbol{v}).

The Stokes-pressure estimate in [20] gives

∥∇pS∥2≤(1/2+ε)∥Δv∥2+Cε∥∇v∥2.\|\nabla p_S\|^2\leq(1/2+\varepsilon)\|\Delta\boldsymbol{v}\|^2+C_\varepsilon\|\nabla\boldsymbol{v}\|^2.

This estimate holds for v∈H2∩H01\boldsymbol{v}\in H^2\cap H_0^1 without a divergence-free constraint. Hence,

ν∣(∇pS,Δv)∣≤ν(3/4+ε/2)∥Δv∥2+C1∥∇v∥2.\nu|(\nabla p_S,\Delta\boldsymbol{v})|\leq\nu(3/4+\varepsilon/2)\|\Delta\boldsymbol{v}\|^2+C_1\|\nabla\boldsymbol{v}\|^2.

By the two-dimensional interpolation inequality and elliptic regularity, we have

∥N(v)∥≤C∥v∥1/2∥∇v∥∥Δv∥1/2.\|N(\boldsymbol{v})\|\leq C\|\boldsymbol{v}\|^{1/2}\|\nabla\boldsymbol{v}\|\|\Delta\boldsymbol{v}\|^{1/2}.

Applying Young’s inequality, we obtain, for every θ>0\theta>0,

∣(N(v),Δv)∣≤νθ∥Δv∥2+Cν,θM02∥∇v∥4.|(N(\boldsymbol{v}),\Delta\boldsymbol{v})|\leq\nu\theta\|\Delta\boldsymbol{v}\|^2+C_{\nu,\theta}M_0^2\|\nabla\boldsymbol{v}\|^4.

Since Q\mathbb{Q} is an L2L^2 contraction, the same bound holds for QN(v)\mathbb{Q}N(\boldsymbol{v}). Taking θ=ε/2\theta=\varepsilon/2 and using ∣(Qf,Δv)∣≤(νε/2)∥Δv∥2+Cf∥f∥2|(\mathbb{Q}\boldsymbol{f},\Delta\boldsymbol{v})|\leq(\nu\varepsilon/2)\|\Delta\boldsymbol{v}\|^2+C_f\|\boldsymbol{f}\|^2, we obtain (49). Taking θ=ε\theta=\varepsilon gives (50), while (51) follows from Young’s inequality.

Lemma 4.1 yields the following sufficient conditions for positivity of the higher-energy residual.

Proposition 4.2 (Positivity of the second residual). Let 0<ε<1/140<\varepsilon<1/14 and fix δ>0\delta>0. With the constants from Lemma 4.1, assume

α2≥Cf+C3+δ,β2≥C1+δ,γ2≥(C2+C4)M02+δ.(52)\alpha_2\geq C_f+C_3+\delta,\qquad\beta_2\geq C_1+\delta,\qquad\gamma_2\geq(C_2+C_4)M_0^2+\delta. \tag*{(52)}

Then

Q^2n+1≥cΔ∥Δu^n+1∥2+δ∥∇u^n+1∥2+δ∥∇u^n+1∥4+δ∥fn+1∥2≥0,(53)\widehat{Q}_2^{n+1}\geq c_\Delta\|\Delta\widehat{\boldsymbol{u}}^{n+1}\|^2+\delta\|\nabla\widehat{\boldsymbol{u}}^{n+1}\|^2+\delta\|\nabla\widehat{\boldsymbol{u}}^{n+1}\|^4+\delta\|\boldsymbol{f}^{n+1}\|^2\geq0, \tag*{(53)}

where cΔ=ν(1/4−7ε/2)>0c_\Delta=\nu(1/4-7\varepsilon/2)>0. In particular, Q^2n+1≥0\widehat{Q}_2^{n+1}\geq0.

Proof. The estimates (49)–(51) with v=u~n+1\boldsymbol{v}=\widetilde{\boldsymbol{u}}^{n+1}, q=pn+1q=p^{n+1}, and f=fn+1\boldsymbol{f}=\boldsymbol{f}^{n+1} use the fraction 3/4+3ε/2+ε+ε=3/4+7ε/23/4+3\varepsilon/2+\varepsilon+\varepsilon=3/4+7\varepsilon/2 of the viscous term. Together with (52), this gives (53). Finally, Q^2n+1=Q2n+1+CKG22≥0\widehat{Q}_2^{n+1}=Q_2^{n+1}+C_KG_2^2\geq0.

Weighted uniform Gronwall argument and the H1H^1 bound

We shall use the following variable-step version of the discrete uniform Gronwall lemma [24 Lemma 3]. Its right-endpoint weights agree with those in the dissipation estimate for the first SAV.

Lemma 4.3 (Variable-step uniform Gronwall inequality). Let dj,gj,hjd^j,g^j,h^j be nonnegative and τj=tj−tj−1>0\tau_j=t_j-t_{j-1}>0. Suppose

dj+1≤(1+τj+1gj+1)dj+τj+1hj+1.(54)d^{j+1}\leq(1+\tau_{j+1}g^{j+1})d^j+\tau_{j+1}h^{j+1}. \tag*{(54)}

Fix integers 0≤m<n0\leq m<n, set L=tn−tmL=t_n-t_m, and assume

∑j=mn−1τj+1dj+1≤Ad,∑j=mn−1τj+1gj+1≤Ag,∑j=mn−1τj+1hj+1≤Ah.(55)\sum_{j=m}^{n-1}\tau_{j+1}d^{j+1}\leq A_d,\qquad\sum_{j=m}^{n-1}\tau_{j+1}g^{j+1}\leq A_g,\qquad\sum_{j=m}^{n-1}\tau_{j+1}h^{j+1}\leq A_h. \tag*{(55)}

Then

dn≤(AdL+Ah)eAg,(56)d^n\leq\left(\frac{A_d}{L}+A_h\right)e^{A_g}, \tag*{(56)}

without a restriction on adjacent step ratios. If AdA_d, AgA_g, AhA_h are uniformly bounded over a family of intervals with L≥L0>0L\geq L_0>0, then (56) gives a uniform bound at their endpoints.

Proof. Iterating (54) from ii to nn, we obtain

dn≤di∏ℓ=i+1n(1+τℓgℓ)+∑ℓ=i+1nτℓhℓ∏r=ℓ+1n(1+τrgr).d^n \le d^i \prod_{\ell=i+1}^{n}(1+\tau_\ell g^\ell)+\sum_{\ell=i+1}^{n}\tau_\ell h^\ell\prod_{r=\ell+1}^{n}(1+\tau_r g^r).

With empty products equal to one and empty sums equal to zero, this also holds for i=ni=n. Since the coefficients are nonnegative and 1+x≤ex1+x\le e^x, we have

dn≤eAg(di+Ah),m≤i≤n.d^n\le e^{A_g}(d^i+A_h),\qquad m\le i\le n.

Multiplying by τi\tau_i, summing over i=m+1,…,ni=m+1,\ldots,n, and dividing by ∑i=m+1nτi=L\sum_{i=m+1}^{n}\tau_i=L, we obtain (56) from (55). The last assertion follows directly. □

For Algorithm 2, the first SAV controls the right-endpoint average ∑τj+1sj+1\sum\tau_{j+1}s^{j+1}. We next establish the recursion required by Lemma 4.3.

Lemma 4.4 (Recursion and upper energy bound). Assume the conditions of Proposition 4.2 and s0>0s^0>0. Then

0<sn+1≤KCn+1,(57)0<s^{n+1}\le\mathcal{K}_{C}^{n+1}, \tag*{(57)}

and

sn+1≤(1+τn+1G2n+1)sn+τn+1H2n+1.(58)s^{n+1}\le(1+\tau_{n+1}G_2^{n+1})s^n+\tau_{n+1}H_2^{n+1}. \tag*{(58)}

Proof. By Proposition 4.2, the denominator in (45) is at least one, and the numerator is positive whenever sn>0s^n>0. Thus s^n+1>0\widehat{s}^{n+1}>0, and (46) gives (57) by induction. Dropping the denominator in (45) and using sn+1≤s^n+1s^{n+1}\le\widehat{s}^{n+1} gives (58). □

We now prove the uniform H1H^1 bound.

Theorem 4.5 (All-time H1H^1 stability on arbitrary positive grids). Let T>0T>0, 0=t0<⋯<tNT=T0=t_0<\cdots<t_{N_T}=T, and d=2d=2. Assume (1.2), (3.9), (4.18), and the predictor satisfies Assumption P. For some fixed σ∈(0,1]\sigma\in(0,1], suppose that

u0=uˉ0∈H1+σ(Ω),u^0=\bar{u}^0\in H^{1+\sigma}(\Omega),

and that the starting histories satisfy

uj,uˉj∈H2(Ω),j=1,…,k−1,pj∈H1(Ω)/R,j=0,…,k−1.u^j,\bar{u}^j\in H^2(\Omega),\qquad j=1,\ldots,k-1,\qquad p^j\in H^1(\Omega)/\mathbb{R},\qquad j=0,\ldots,k-1.

Require all velocity histories to satisfy the prescribed homogeneous Dirichlet conditions. Initialize

r0=E(u0),s0=K(u0),r^0=\mathcal{E}(u^0),\qquad s^0=\mathcal{K}(u^0),

and apply both SAV updates and the scalar relaxation at every starting step. Prescribe the initial pressure p0p^0 in H1(Ω)/RH^1(\Omega)/\mathbb{R} and apply pressure recovery to each newly computed positive-time velocity. Let the subsequent values be generated by Algorithm 2. Then

∥un∥≤ML2,∥∇un∥≤MH1,0≤n≤NT.(59)\lVert u^n\rVert\le M_{L^2},\qquad\lVert\nabla u^n\rVert\le M_{H^1},\qquad0\le n\le N_T. \tag*{(59)}

The constants may depend on the initial data and fixed algorithmic parameters, but are independent of TT, NTN_T, and all step sizes. The bounds require neither an upper step-size bound nor a restriction on adjacent step ratios, and hold for all nn on an infinite positive grid.

Proof. By Theorem 3.3, the first correction satisfies ∥u~n∥≤M0\lVert\widetilde{u}^n\rVert\le M_0 and (3.16). For R>0R>0, set

DR:=AR(2)+2CK,ΓR:=4β2+4γ2AR(2),H∗:=2α2F∞2.(60)D_R:=A_R(2)+2C K,\qquad\Gamma_R:=4\beta_2+4\gamma_2 A_R(2),\qquad H_*:=2\alpha_2 F_\infty^2. \tag*{(60)}

Fix the reference window length at 11, independently of the grid. For any interval [tm,tn][t_m,t_n] of length at most 22 with rj≤Rr^j\le R for j≥mj\ge m, (3.16) and (4.23) imply

∑j=mn−1τj+1sj+1≤DR,∑j=mn−1τj+1(G2j+1)2≤ΓR,∑j=mn−1τj+1(H2j+1)2≤H∗.(61)\sum_{j=m}^{n-1}\tau_{j+1}s^{j+1}\le D_R,\qquad \sum_{j=m}^{n-1}\tau_{j+1}(G_2^{j+1})^2\le\Gamma_R,\qquad \sum_{j=m}^{n-1}\tau_{j+1}(H_2^{j+1})^2\le H_*. \tag*{(61)}

Here G2j+1=2β2+4γ2K(u~j+1)G_2^{j+1}=2\beta_2+4\gamma_2\mathcal{K}(\widetilde{u}^{j+1}).

Taking R=CRR=C_R, we shall prove

sn≤MS:=[max⁡{s0,DCR}+H∗]eΓCRfor every n.(62)s^n \le M_S := [\max\{s^0,D_{C_R}\}+H_*]e^{\Gamma_{C_R}} \quad\text{for every }n. \tag*{(62)}

First, for any step τℓ≥1\tau_\ell\ge1, the one-step estimate (3.16) gives

K(u~ℓ)≤Cη,k(CR)ν(CRτℓ+B1)≤ACR(1).K(\widetilde{u}^{\ell}) \le\frac{C_{\eta,k}(C_R)}{\nu}\left(\frac{C_R}{\tau_\ell}+B_1\right) \le A_{C_R}(1).

Thus sℓ≤ACR(1)+CK≤DCRs^\ell\le A_{C_R}(1)+C_K \le D_{C_R}.

For a fixed endpoint nn, let ℓ\ell be the index of the last step of size at least 1 among steps 1,…,n1,\ldots,n, and set ℓ=0\ell=0 if there is no such step. All steps after ℓ\ell are less than 1. If tn−tℓ<1t_n-t_\ell<1, iterating (4.24) and using (4.27), we obtain

sn≤eΓCR(sℓ+H∗)≤eΓCR(max⁡{s0,DCR}+H∗).s^n \le e^{\Gamma_{C_R}}(s^\ell+H_*)\le e^{\Gamma_{C_R}}(\max\{s^0,D_{C_R}\}+H_*).

The case ℓ=n\ell=n is included. If tn−tℓ≥1t_n-t_\ell\ge1, choose the largest m≥ℓm\ge\ell such that tn−tm≥1t_n-t_m\ge1. Since τm+1<1\tau_{m+1}<1, we have 1≤tn−tm<21\le t_n-t_m<2. Lemma 4.3, with dj=sjd^j=s^j, gj=G2jg^j=G_2^j, hj=H2jh^j=H_2^j, and (4.27), yields

sn≤(DCR+H∗)eΓCR≤MS.s^n\le(D_{C_R}+H_*)e^{\Gamma_{C_R}}\le M_S.

This proves (4.28).

To estimate the velocity, note that 0<ζn+1≤10<\zeta^{n+1}\le1 and

0≤λk(z)=1−(1−z)k≤min⁡{1,kz},0≤z≤1.0\le\lambda_k(z)=1-(1-z)^k\le\min\{1,kz\},\qquad0\le z\le1.

It follows that λk(z)2≤λk(z)≤kz\lambda_k(z)^2\le\lambda_k(z)\le kz, and hence

K(un+1)≤kζn+1K(u~n+1)≤ksn+1≤kMS.(63)K(u^{n+1})\le k\zeta^{n+1}K(\widetilde{u}^{n+1})\le ks^{n+1}\le kM_S. \tag*{(63)}

Also ∥un+1∥≤∥u~n+1∥≤M0\lVert u^{n+1}\rVert\le\lVert\widetilde{u}^{n+1}\rVert\le M_0. Thus (4.25) holds with ML2=M0M_{L^2}=M_0 and MH1=(2kMS)1/2M_{H^1}=(2kM_S)^{1/2}, including n=0n=0 because MS≥s0≥K(u0)M_S\ge s^0\ge K(u^0).

Remark 4.6 (Regularity of the starting values). At the first full multistep update, u0u^0 enters the derivative and explicit convection, but not the diffusion interpolation. In two dimensions, we have

∥(w⋅∇)v∥L2≤∥w∥L∞∥∇v∥L2≤Cσ∥w∥H1+σ∥v∥H1+σ,σ>0.\lVert(w\cdot\nabla)v\rVert_{L^2}\le\lVert w\rVert_{L^\infty}\lVert\nabla v\rVert_{L^2}\le C_\sigma\lVert w\rVert_{H^{1+\sigma}}\lVert v\rVert_{H^{1+\sigma}},\qquad\sigma>0.

Thus its convection contribution belongs to L2L^2 for H1+σH^{1+\sigma} initial data. The estimate fails when both factors are only in H1H^1, so the argument does not cover σ=0\sigma=0.

For the generalized diffusion interpolation, each independent old provisional velocity with a nonzero diffusion coefficient must satisfy Δu‾j∈L2\Delta\overline{u}^j\in L^2, or equivalently u‾j∈H2\overline{u}^j\in H^2 under the stated boundary and domain assumptions. Together with pj∈H1/Rp^j\in H^1/\mathbb{R}, this gives an L2L^2 right-hand side for the Helmholtz equation. Elliptic regularity yields u‾k∈H2\overline{u}^k\in H^2, scalar rescaling preserves this regularity, and pressure recovery gives pk∈H1/Rp^k\in H^1/\mathbb{R}. Repeating the argument gives the same regularity at all subsequent steps.

These sufficient regularity conditions can be weakened if a diffusion coefficient vanishes or a history combination has additional regularity. For classical BDF, diffusion is evaluated only at the new time level, so H1+σH^{1+\sigma} starting velocity histories suffice for the first full multistep predictor. The H2H^2 regularity justifies each solve and residual evaluation; no uniform-in-time H2H^2 bound is assumed or obtained.

Applying the preceding argument after the first scalar enters its absorbing interval gives the following result.

Corollary 4.7 (Eventual H1H^1 velocity bounds). Fix the algorithmic parameters and consider initialized histories satisfying the hypotheses of Theorem 4.5 with these parameters. Let tn→∞t_n\to\infty. For ϵ>0\epsilon>0, set Ra=R∞+ϵR_a=R_\infty+\epsilon and choose nan_a by (3.24). Then

sn≤(DRa+H∗)eΓRa=:MS,abs(ϵ)whenever tn−tna≥1.(64)s^n\le(D_{R_a}+H_*)e^{\Gamma_{R_a}}=:M_{S,\mathrm{abs}}(\epsilon)\qquad\text{whenever }t_n-t_{n_a}\ge1. \tag*{(64)}

Consequently, at all such endpoints,

∥un∥H12≤2Cη,k(Ra)Ra+2kMS,abs(ϵ).(65)\lVert u^n\rVert_{H^1}^2\le2C_{\eta,k}(R_a)R_a+2kM_{S,\mathrm{abs}}(\epsilon). \tag*{(65)}

Thus (65) defines an absorbing ball for the admissible initial class, with radius independent of the particular history and the time grid. The entry criterion is uniform over bounded sets of admissible initial energies.

Proof. By Theorem 3.3, rj≤Rar^{j}\le R_{a} for j≥naj\ge n_{a}. Hence (26) and (61) hold with R=RaR=R_{a} on every interval starting at or after nan_{a}. We apply the proof of Theorem 4.5 starting at nan_{a}. If a step of size at least 1 occurs among steps na+1,…,nn_{a}+1,\ldots,n, let ℓ\ell be the index of the last such step. Then sℓ≤DRas^{\ell}\le D R_{a}. For tn−tℓ<1t_{n}-t_{\ell}<1, direct iteration gives the bound (DRa+H∗)eΓRa(D R_{a}+H_{*})e^{\Gamma R_{a}}. Otherwise, the remaining steps are less than 1, and we apply Lemma 4.3 on an interval [tm,tn][t_{m},t_{n}] with 1≤tn−tm<21\le t_{n}-t_{m}<2. If no such large step occurs after nan_{a}, the condition tn−tna≥1t_{n}-t_{n_{a}}\ge1 gives the same backward interval. This proves (64) without a bound on snas^{n_{a}}. Finally, (31) and (63) yield (65).

Remark 4.8 (Parameter dependence and the meaning of absorption). To impose (52) uniformly on an initial class whose all-time L2L^{2} bound satisfies M0≤M∗M_{0}\le M_{*}, fix γ2≥(C2+C4)M∗2+δ\gamma_{2}\ge(C_{2}+C_{4})M_{*}^{2}+\delta. Corollary 4.7 then gives a common absorbing radius for this class. The radius depends on the fixed γ2\gamma_{2}, but not on the particular initial history or its H1H^{1} norm. Enlarging the class may require a different γ2\gamma_{2} and hence a different radius. The all-time bound retains its dependence on r0r^{0} and s0s^{0} through CRC_{R} and MSM_{S}.

Implementation and comparison

In this section, we discuss the implementation and temporal accuracy of the two schemes, and compare their long-time stability properties and computational costs with those of existing methods.

Implementation and computational cost

After spatial discretization, the generalized BDF predictor takes the form

(ak,kn(β)τn+1M+νbk,k−1n(β)A)U~n+1=Rn,(66)\left(\frac{a_{k,k}^{n}(\beta)}{\tau_{n+1}}M+\nu b_{k,k-1}^{n}(\beta)A\right)\widetilde{U}^{n+1}=R^{n}, \tag*{(66)}

where MM and AA are the fixed mass and Laplace matrices, respectively, and RnR^{n} contains only known quantities. The pressure is recovered by solving the Poisson equation

Appn+1=G(U~n+1,Fn+1).(67)A_{p}p^{n+1}=G(\widetilde{U}^{n+1},F^{n+1}). \tag*{(67)}

For a fixed spatial mesh and a constant time step, we can factorize the matrices in (66)–(67) once and reuse the factorizations at subsequent steps. For variable time steps, the spatial operators remain unchanged, but the scalar weights in the Helmholtz matrix may require a new factorization. Finite-element, finite-difference, and spectral discretizations on nonperiodic domains preserve this constant-coefficient structure in space.

In addition to the field solves and the evaluation of residuals and norms, Algorithm 1 requires the scalar division (16) and one velocity rescaling. Algorithm 2 adds the division (45), the minimum in (46), and a second rescaling. Hence, the stronger H1H^{1} bound requires no additional Helmholtz or Poisson solve. The coefficients of the linear systems are independent of the numerical velocity, and no nonlinear iteration is needed. These costs are for the multistep schemes presented above; the cost of a Runge–Kutta predictor depends on its stages.

The stability results are established for the time-discrete schemes. To obtain mesh-uniform estimates for a fully discrete scheme, one needs discrete counterparts of the Poincaré, pressure, and interpolation estimates used in the analysis. In particular, for standard C0C^{0} finite elements, replacing the continuous Laplacian in the higher-energy residual by an elementwise Laplacian requires further analysis.

Temporal accuracy

The correction polynomials convert a first-order error in the SAV ratios into higher-order perturbations of the velocity. On a fixed time interval [0,T][0,T], let hmax⁡=max⁡1≤j≤NTτjh_{\max}=\max_{1\le j\le N_{T}}\tau_{j}. Then

ξn=1+O(hmax⁡) ⟹ ηkn=1+O(hmax⁡k),ζn=1+O(hmax⁡) ⟹ λkn=1+O(hmax⁡k).(68)\xi^{n}=1+O(h_{\max})\ \Longrightarrow\ \eta_{k}^{n}=1+O(h_{\max}^{k}),\qquad\zeta^{n}=1+O(h_{\max})\ \Longrightarrow\ \lambda_{k}^{n}=1+O(h_{\max}^{k}). \tag*{(68)}

For the minimum relaxation, we have

1−ζn+1=(1−S^n+1KCn+1)+,1-\zeta^{n+1}=\left(1-\frac{\widehat{S}^{n+1}}{K_C^{n+1}}\right)_{+},

so the minimum does not increase the provisional ratio error relative to the current shifted energy. The ratio estimates in (68) require an error analysis of the coupled schemes; they do not follow from uniform boundedness alone.

For the BDF scheme, we retain the provisional histories in the derivative and diffusion terms and use corrected velocities in the explicit terms, as in the degree-kk correction of [9]. For a one-step method restarted from the corrected velocity, a perturbation of order hkh^k at each step can accumulate to order hk−1h^{k-1} on a constant grid. In this case, a degree-(k+1)(k+1) correction is one possible choice, provided that the corresponding ratio-consistency estimates hold. The scalar stability analysis applies to both choices and, more generally, to any fixed correction degree.

High-order convergence further requires sufficiently smooth solutions and forcing, starting values of the appropriate order, and the stability and time-grid conditions of the predictor. Thus, boundedness on arbitrary positive grids does not imply accuracy on such grids or zero-stability of classical BDF formulas at all orders. A finite-time error analysis of the coupled schemes remains to be developed.

Remark 5.1 (Relaxation variants). Any alternative relaxation satisfying

0<sn+1≤KCn+1,sn+1≤S^n+1,0<s^{n+1}\leq K_C^{n+1},\qquad s^{n+1}\leq\widehat{S}^{n+1},

preserves the two inequalities used in the stability proof. Accuracy also requires consistency of the scalar ratio. For example, one half of the minimum satisfies both inequalities but need not converge to the continuous energy. Hence, these stability conditions alone do not ensure the accuracy of an alternative relaxation.

Comparison with existing long-time stable methods

In Table 1, we compare methods for which uniform-in-time bounds have been proved. The third column gives the order and time-step conditions of the cited long-time analysis. For the present schemes, we list the correction degree, since the stability analysis for an arbitrary degree does not establish the corresponding convergence order.

Method and referenceSetting and uniform boundOrder and long-time conditionPrincipal solves per step
Fully implicit Euler and Crank–Nicolson [31, 30]No-slip, 2D; velocity L2L^2 and H1H^1Orders 1 and 2; the H1H^1 analysis requires a time-step condition (CFL type for Crank–Nicolson)Nonlinear velocity–pressure problem
Linear backward Euler [15]Primitive variables, 2D; no-slip or periodic; bounds in VV and D(A)D(A), where AA is the Stokes operatorOrder 1; bounds uniform with respect to δt\delta tLinear Oseen/Stokes problem with a convection coefficient depending on the solution
Linearly extrapolated BDF2 [1, 23, 27]Primitive variables, no-slip; velocity L2L^2 in 2D/3D and H1H^1 in 2DOrder 2; L2L^2 unconditional; H1H^1 requires a mild time-step restriction in the cited analysisLinear Oseen/Stokes problem with a convection coefficient depending on the solution
Velocity–vorticity IMEX finite elements [7]Periodic, 2D; velocity and vorticity in both L2L^2 and H1H^1BDF2 and a first-order variant; unconditionalTwo sequential linear Oseen-type field solves; the vorticity solve is decoupled
Explicit-convection periodic vorticity schemes [5, 34, 2]Periodic, 2D; vorticity L2L^2 and H1H^1, with higher Sobolev bounds in the later schemesOrders 1–4; time-step restrictionOne or two constant-coefficient Poisson/heat solves, depending on the scheme
Earlier high-order SAV–IMEX scheme [9]Periodic, 2D/3D; velocity H1H^1 in 2D and L2L^2 in 3D; zero external force in the analyzed modelBDF orders 1–5; unconditionalOne constant-coefficient Poisson-type velocity solve and an explicit scalar update
Forced SAV–BDF2 scheme [6, 19]Periodic, no-slip, or free-slip; velocity L2L^2 in 2D/3D; vorticity and H1H^1 in the 2D free-slip settingOrder 2; unconditional; the later analysis allows nonsmooth initial data and gives a δt\delta t-independent energy boundTwo constant-coefficient Brinkman/Stokes solves and one linear scalar update
Variable-step ETD–mr-SAV schemes [33, 32]Periodic, 2D; vorticity L2L^2Variable-step order 2 (with an embedded first-order companion); unconditional[33]: two heat solves and a scalar cubic; [32]: two heat solves and a linear scalar recovery by Laplace–Talbot inversion
Present one-SAV schemeVelocity L2L^2 for no-slip problems in 2D/3DAny correction degree kk; arbitrary positive variable stepsOne vector Helmholtz solve, one Poisson solve, and one scalar division
Present two-SAV schemeNo-slip, 2D; velocity H1H^1Any correction degree kk; arbitrary positive variable stepsThe same field solves, plus one scalar division and a scalar relaxation

Table 1. Comparison of numerical schemes with proved uniform-in-time bounds for the Navier–Stokes equations. Here “unconditional” means that the cited long-time estimate requires no CFL-type or data-dependent time-step restriction.

The methods in Table 1 differ mainly in the systems to be solved at each step. Fully implicit methods require nonlinear solves. Linearly extrapolated BDF2 and velocity–vorticity methods lead to linear Oseen-type systems whose coefficients depend on the numerical solution. The periodic vorticity methods with explicit convection use constant-coefficient operators, but their cited long-time estimates require time-step restrictions.

Among the SAV methods, the high-order scheme in [9] treats the unforced periodic problem. The forced SAV–BDF2 schemes in [6, 19] provide uniform-in-time bounds and require constant-coefficient Brinkman/Stokes solves. The variable-step ETD methods in [33, 32] treat the two-dimensional periodic vorticity equation and require either a scalar cubic solve or a linear auxiliary recovery by numerical Laplace inversion.

Our schemes use the velocity–pressure formulation and require only decoupled Helmholtz/Poisson solves and explicit scalar updates. The second SAV strengthens the velocity bound in two dimensions without additional field solves. Moreover, the scalar estimates are valid for the admissible predictors and arbitrary positive time grids.

Numerical examples

In this section, we present three groups of numerical tests in two dimensions with no-slip boundary conditions. We first verify the temporal convergence of the two-SAV scheme. We then compare the BDF3 predictor with and without SAV corrections, and finally examine the efficiency of adaptive time stepping for a problem with a short initial transient and persistent long-time motion. We use constant time steps in the first two tests and variable time steps in the third. The algorithmic parameters are fixed within each refinement study.

Convergence test

We take Ω=(−1,1)2\Omega= (-1,1)^2, ν=1\nu= 1, and T=1T = 1, and choose the forcing such that the exact velocity and pressure are given by

u(x,y,t)=π(sin⁡(2πy)sin⁡2(πx)sin⁡(t−12)−sin⁡(2πx)sin⁡2(πy)sin⁡(t−12)),(69)u(x,y,t) = \pi\begin{pmatrix} \sin(2\pi y)\sin^2(\pi x)\sin\left(t-\frac{1}{2}\right) \\ -\sin(2\pi x)\sin^2(\pi y)\sin\left(t-\frac{1}{2}\right) \end{pmatrix}, \tag*{(69)}
p(x,y,t)=sin⁡(t−12)cos⁡(πx)sin⁡(πy).p(x,y,t) = \sin\left(t-\frac{1}{2}\right)\cos(\pi x)\sin(\pi y).

We use the Legendre–Galerkin method in space with polynomial degree Nx=Ny=32N_x = N_y = 32. In each coordinate, the velocity basis consists of Lj−Lj+2L_j - L_{j+2}, j=0,…,Nx−2j = 0,\ldots,N_x - 2, where LjL_j is the Legendre polynomial of degree jj. For the pressure, we use all polynomials up to degree 32 and fix the constant mode by the zero-mean condition. We test the generalized BDF schemes of orders k=2,3,4k=2,3,4 with the shifts and time steps

(k,β)=(2,2), (3,3), (4,8),δt=180,1160,1320,1640.(70)(k,\beta)=(2,2),\ (3,3),\ (4,8),\qquad\delta t=\frac{1}{80},\frac{1}{160},\frac{1}{320},\frac{1}{640}. \tag*{(70)}

For Algorithm 2, we set

CE=CK=100,α1=4π2,β1=π28,(71)C_E=C_K=100,\qquad\alpha_1=\frac{4}{\pi^2},\qquad\beta_1=\frac{\pi^2}{8}, \tag*{(71)}
α2=5,β2=γ2=10−2.\alpha_2=5,\qquad\beta_2=\gamma_2=10^{-2}.

We use correction degree kk and the minimum relaxation (4.12). The derivative and diffusion terms use the provisional velocity histories, while the explicit convection uses the final corrected velocities. The pressure is recovered from the first corrected velocity before applying the second SAV correction.

For q=u,pq=u,p, we define the errors at the final time by

eq,X(δt)=∥qNT−q(T)∥X,X=L2,H1.(72)e_{q,X}(\delta t)=\lVert q^{N_T}-q(T)\rVert_X,\qquad X=L^2,H^1. \tag*{(72)}

Here ∥v∥H12=∥v∥2+∥∇v∥2\lVert v\rVert_{H^1}^2=\lVert v\rVert^2+\lVert\nabla v\rVert^2, and the numerical and exact pressures have zero mean. From Figure 1, we observe the expected orders 2, 3, and 4 for both velocity and pressure in the two norms over the tested range of time steps.

Temporal convergence plots for velocity and pressure in the $L^2$ and $H^1$ norms

Figure 1. Temporal convergence of Algorithm 2 for (6.1) at T=1T=1. Solid lines with filled markers show the results for k=2,3,4k=2,3,4, with different colors and marker shapes. Dotted segments of the same colors indicate slopes 2, 3, and 4, respectively.

Next, let u∗u_* denote the exact velocity in (6.1), and define the scalar and ratio errors by

er=∣rNT−E(u∗(T))∣,es=∣sNT−K(u∗(T))∣,e_r=\left|r^{N_T}-\mathcal{E}(u_*(T))\right|,\qquad e_s=\left|s^{N_T}-\mathcal{K}(u_*(T))\right|,
eξ=∣ξNT−1∣,eζ=∣ζNT−1∣.(73)e_\xi=\left|\xi^{N_T}-1\right|,\qquad e_\zeta=\left|\zeta^{N_T}-1\right|. \tag*{(73)}

In Figure 2, we observe approximately first-order convergence of these errors. Thus, in this test, first-order scalar errors coexist with higher-order velocity and pressure convergence, consistent with (68). The minimum relaxation is active in these computations. All computed SAV residuals are nonnegative, and all update denominators are positive.

Scalar and ratio errors for Algorithm 2 in the convergence test

Figure 2. Scalar and ratio errors for Algorithm 2 in the convergence test. Dotted lines indicate slope one. The scalar errors are first order, while the velocity and pressure errors have the higher orders shown in Figure 1.

Effect of the SAV corrections on stability

We compare the BDF3 predictor with and without the two SAV corrections on the unit disk D={(x,y):x2+y2<1}D = \{(x,y): x^2 + y^2 < 1\} and the square S=(−1,1)2S = (-1,1)^2, both with no-slip boundary conditions. In each comparison, the predictor and spatial discretization are unchanged, so the difference is due to the scalar updates and rescalings.

We first consider the unit disk. Set ν=10−2\nu= 10^{-2}, A=1.05A = 1.05, σ=x2+y2\sigma= x^2 + y^2, and prescribe

Ψ=(1−σ)2[A+10−4e−tRe⁡[(x+iy)7]],u∗=(∂yΨ,−∂xΨ),\Psi= (1-\sigma)^2\left[A + 10^{-4}e^{-t}\operatorname{Re}\left[(x+\mathrm{i}y)^7\right]\right], \qquad\boldsymbol{u}_* = (\partial_y\Psi,-\partial_x\Psi),
p∗=A2[8(σ−σ2+σ33)−2].(74)p_* = A^2\left[8\left(\sigma-\sigma^2+\frac{\sigma^3}{3}\right)-2\right]. \tag*{(74)}

The velocity is smooth and divergence free. We evaluate the bounded force analytically from

f=∂tu∗−νΔu∗+(u∗⋅∇)u∗+∇p∗,\boldsymbol{f} = \partial_t\boldsymbol{u}_*-\nu\Delta\boldsymbol{u}_*+(\boldsymbol{u}_*\cdot\nabla)\boldsymbol{u}_*+\nabla p_*,

independently of the numerical solution.

We use the Fourier–Jacobi velocity space (1−σ)PN−2(D)2(1-\sigma)P_{N-2}(D)^2, with N=14N = 14 and 19×6519 \times65 quadrature points. The BDF3 formula uses the provisional velocity histories, while convection uses 3un−3un−1+un−23\boldsymbol{u}^n-3\boldsymbol{u}^{n-1}+\boldsymbol{u}^{n-2}. We use the scalar updates, pressure recovery, and minimum relaxation of Algorithm 2 with cubic correction polynomials. We take δt=0.04\delta t = 0.04, T=120T = 120, and

CE=CK=1,α1=4νπ2,β1=νπ28,α2=50,β2=γ2=10−2.C_E = C_K = 1,\qquad\alpha_1 = \frac{4}{\nu\pi^2},\qquad\beta_1 = \frac{\nu\pi^2}{8},\qquad\alpha_2 = 50,\qquad\beta_2 = \gamma_2 = 10^{-2}.

Both computations use the same forcing, spatial operators, and RK4 startup with 20 substeps on each of the first two intervals. The computation without SAV omits only the scalar updates and rescalings. Neither computation uses filtering or resets to the exact solution.

In Figure 3, we observe that the uncorrected predictor becomes unstable near t=99t = 99, while the two-SAV solution remains bounded through all 3000 steps. In Figure 4, we compare the numerical solutions with the exact solution. For the two-SAV scheme, the maximum relative errors over the full interval are 2.0118% and 7.1616% for velocity in L2L^2 and the full H1H^1 norm, respectively, and 0.7921% for pressure in L2L^2.

Velocity norms normalized by their initial values

Figure 3. Velocity norms normalized by their initial values (N=14N = 14, δt=0.04\delta t = 0.04). Top: results with and without SAV on logarithmic scales. Bottom: two-SAV results on linear scales.

Relative errors

Figure 4. Relative errors. Left: two-SAV velocity and pressure errors in percent. Right: velocity L2L^2 errors with and without SAV on a logarithmic scale.

This example also shows that boundedness does not ensure accuracy. For N=18N = 18 and δt=0.04\delta t = 0.04, the two-SAV solution remains bounded, but its maximum relative velocity L2L^2 error is about 21%. With δt=0.02\delta t = 0.02, both methods give accurate results for N=14,18N = 14, 18. Thus, although the scalar corrections control the velocity norms, the time step must still be chosen to resolve the solution.

The H1H^{1} analysis in Section 4 uses a Stokes-pressure estimate for domains with C3C^{3} boundaries. We next consider a square to test the scheme on a polygonal domain, which is outside this assumption. We consider S=(−1,1)2S=(-1,1)^{2} with homogeneous no-slip boundary conditions. Let D={x2+y2<1}D=\{x^{2}+y^{2}<1\} and retain the disk data

ψD=(1−σ)2[A+εe−tRe⁡[(x+iy)7]],uD=(∂yψD,−∂xψD),σ=x2+y2,\psi_D=(1-\sigma)^2[A+\varepsilon e^{-t}\operatorname{Re}[(x+iy)^7]],\qquad\boldsymbol{u}_D=(\partial_y\psi_D,-\partial_x\psi_D),\qquad\sigma=x^2+y^2,

with A=1.05A=1.05, ε=10−4\varepsilon=10^{-4}, and pD=A2[8(σ−σ2+σ3/3)−2]p_D=A^2[8(\sigma-\sigma^2+\sigma^3/3)-2]. Writing fD=∂tuD−νΔuD+(uD⋅∇)uD+∇pD\boldsymbol{f}_D=\partial_t\boldsymbol{u}_D-\nu\Delta\boldsymbol{u}_D+(\boldsymbol{u}_D\cdot\nabla)\boldsymbol{u}_D+\nabla p_D, we prescribe

u0=1DuD(⋅,0),f(⋅,t)=1DfD(⋅,t)on S.(75)\boldsymbol{u}_0=\mathbf{1}_D\boldsymbol{u}_D(\cdot,0),\qquad\boldsymbol{f}(\cdot,t)=\mathbf{1}_D\boldsymbol{f}_D(\cdot,t)\quad\text{on }S. \tag*{(75)}

Since uD=0\boldsymbol{u}_D=0 on ∂D\partial D, the extended initial velocity belongs to H01(S)2H^1_0(S)^2 and is weakly divergence free. We do not assume that the extended disk field is an exact time-dependent solution on the square.

We use a Legendre–Galerkin BDF3 predictor in [(1−x2)(1−y2)Q12,12]2[(1-x^2)(1-y^2)Q_{12,12}]^2, with ν=10−2\nu=10^{-2}, δt=0.04\delta t=0.04, and T=120T=120. Both computations use the same projected initial data and RK4 starting histories, with 20 substeps per initial interval. For the two-SAV scheme, we use cubic corrections and the minimum relaxation, with

CE=CK=1,α1=4νπ2,β1=νπ28,α2=50,β2=γ2=10−2.C_E=C_K=1,\qquad\alpha_1=\frac{4}{\nu\pi^2},\qquad\beta_1=\frac{\nu\pi^2}{8},\qquad\alpha_2=50,\qquad\beta_2=\gamma_2=10^{-2}.

The computation without SAV omits only the scalar updates and rescalings. We evaluate the initial-data and forcing integrals on their disk support, and the velocity integrals on the entire square.

In Figure 5, the computation without SAV becomes unstable: ∥un∥/∥u0∥\lVert\boldsymbol{u}^{n}\rVert/\lVert\boldsymbol{u}_0\rVert exceeds 106010^{60} at t=6.20t=6.20. The two-SAV computation reaches T=120T=120, with maximum normalized L2L^2 and full H1H^1 norms of 1.0873 and 1.0476, respectively. All computed SAV residuals and update denominators are positive. We observe the same qualitative stability behavior as on the disk. Extending the stability analysis to polygonal domains remains an open problem.

Stability comparison on the square

Figure 5. Stability comparison on the square at δt=0.04\delta t=0.04. Velocity norms are normalized by the corresponding continuous initial norms. The main axes show all recorded values; the computation without SAV terminates at the runaway threshold before T=120T=120. Insets show enlarged views of the two-SAV norm histories.

Adaptive long-time integration

We now examine the efficiency of adaptive time stepping. We combine the two-SAV corrections with the two-stage IMEX–SDIRK2 predictor and embedded step-selection strategy of [17], replacing its SAV equation by our scalar equations and minimum relaxation. Each step starts from the corrected velocity. For this second-order predictor, we use the cubic rescalings η3=1−(1−ξ)3\eta_3=1-(1-\xi)^3 and λ3=1−(1−ζ)3\lambda_3=1-(1-\zeta)^3, as discussed in Section 5. We set the embedded relative velocity tolerance to 5×10−35 \times10^{-3} and the tolerance for the scalar indicator max⁡{∣1−ξ∣,∣1−ζ∣}\max\{|1-\xi|,|1-\zeta|\} to 0.5. The safety factor is 0.9, the initial step is 10−310^{-3}, and the step bounds are [10−7,0.2][10^{-7},0.2]. If a step is rejected, all trial velocity, pressure, and scalar values are discarded.

On the no-slip unit disk, let σ=x2+y2\sigma=x^2+y^2, A=1.05A=1.05, and prescribe

a(t)=1+e−100t+0.2sin⁡(0.1t),a(t)=1+e^{-100t}+0.2\sin(0.1t),
Ψ(t)=(1−σ)2[Aa(t)+10−4e−tRe⁡[(x+iy)7]],u∗=(∂yΨ,−∂xΨ),(76)\Psi(t)=(1-\sigma)^2\left[Aa(t)+10^{-4}e^{-t}\operatorname{Re}[(x+iy)^7]\right],\qquad\boldsymbol{u}_*=(\partial_y\Psi,-\partial_x\Psi), \tag*{(76)}
p∗(t)=a(t)2A2[8(σ−σ2+σ33)−2].p_*(t)=a(t)^2A^2\left[8\left(\sigma-\sigma^2+\frac{\sigma^3}{3}\right)-2\right].

We evaluate the bounded force analytically from f=∂tu∗−νΔu∗+(u∗⋅∇)u∗+∇p∗f=\partial_t\boldsymbol{u}_*-\nu\Delta\boldsymbol{u}_*+(\boldsymbol{u}_*\cdot\nabla)\boldsymbol{u}_*+\nabla p_*. We use ν=10−2\nu=10^{-2}, N=14N=14, 19 ×\times 65 quadrature points, and the SAV parameters of the preceding disk example. The solution has an initial layer on the time scale 0.01, followed by persistent motion with period 20π20\pi. All computations use the same spatial operators, forcing, and initial data, with final time T=1000T=1000 and no solution resets.

We first test several constant time steps up to T=200T=200. Among these, τ=0.005\tau=0.005 is the largest for which the maximum relative velocity error is at most 5% in L2L^2. Both the fixed-step and adaptive two-SAV schemes satisfy this requirement over [0,1000][0,1000], with maximum errors 2.444% and 0.949%, respectively. In Figure 6, we observe that the adaptive scheme uses small steps to resolve the initial layer and larger steps thereafter. The number of accepted steps decreases from 200000 to 14083, with two rejected trials, and the reported solver wall time decreases from 591.92 to 42.01 seconds, giving a speedup of 14.09. Both methods meet the same accuracy requirement, but their errors are not matched. In particular, the fixed-step solution is more accurate after the initial transient (Figure 7(b)).

Efficiency plots for the adaptive two-SAV scheme

Figure 6. Efficiency of the adaptive two-SAV scheme. Top: accepted step sizes over the full interval and during the initial transient. Bottom: cumulative step counts and solver wall times compared with the fixed-step two-SAV scheme.

Velocity $L^2$ norms and relative velocity $L^2$ errors

Figure 7. (a) Velocity L2L^2 norms with and without SAV on the same adaptive time grid; the computation without SAV terminates near t=77t=77. (b) Relative velocity L2L^2 errors of the adaptive and fixed-step two-SAV schemes, with a logarithmic time axis.

Finally, we repeat the computation on the accepted adaptive grid without SAV corrections. The uncorrected solution becomes unstable near t=77t=77, while the two-SAV solution remains bounded to T=1000T=1000 (Figure 7(a)), with a maximum relative velocity H1H^1 error of 2.640%. A computation without SAV that selects its own adaptive steps also reaches T=1000T=1000, with maximum relative velocity L2L^2 error 0.737%. Thus, the comparison on the same grid shows the stabilizing effect of the SAV corrections, while a suitable adaptive grid also allows the uncorrected method to remain stable in this example. The reported speedup compares adaptive and fixed steps within the two-SAV method.

Concluding remarks

We constructed in this paper two SAV correction steps for the forced incompressible Navier–Stokes equations. A main feature is that their formulation and long-time stability analysis are independent of the particular linear predictor. The same correction steps apply to any linear predictor satisfying Assumption P. The first correction uses a damped, shifted kinetic-energy variable to obtain uniform-in-time velocity L2L^2 bounds in two and three dimensions. In two dimensions, the second correction combines a higher-energy scalar equation with a relaxation to yield a uniform-in-time H1H^1 bound. Under the stated assumptions, the estimates hold on arbitrary positive time grids. We also established eventual absorbing bounds; for the two-SAV scheme, these apply to the admissible initial class associated with fixed algorithmic parameters.

The analysis combines the weighted dissipation estimate from the first SAV with the Stokes-pressure estimate and a discrete uniform Gronwall argument. By retaining right-endpoint weights and treating large steps directly, we avoid restrictions on step sizes and adjacent step ratios. The scalar estimates do not require an energy identity for the predictor. Moreover, the second SAV requires no additional Helmholtz or Poisson solve.

The numerical results exhibit second-, third-, and fourth-order convergence for the tested BDF schemes and illustrate the effect of the scalar corrections on long-time stability. In the adaptive example, variable steps reduced the reported computational time by about a factor of fourteen relative to the fixed-step two-SAV computation, while both satisfied the same accuracy requirement. This result demonstrates the benefit of adaptivity for the tested implementation.

The uniform-in-time bounds established here do not by themselves imply uniform-in-time accuracy. Solutions of the forced Navier–Stokes equations need not be stable with respect to perturbations of the initial data. Consequently, small discretization errors may be amplified over long times even when both the exact and numerical solutions remain bounded. Error estimates of the expected convergence order with constants independent of the final time generally require additional stability assumptions on the exact solution, such as uniform exponential stability [8]. A finite-time error analysis of the coupled schemes and the approximation of long-time statistical quantities remain to be studied.

References

  1. [1]M. Akbas, S. Kaya, and L. G. Rebholz. On the stability at all times of linearly extrapolated BDF2 timestepping for multiphysics incompressible flow problems. Numer. Methods Partial Differ. Equ., 33(4):999–1017, 2017.DOI
  2. [2]K. Cheng and C. Wang. Long time stability of high order multistep numerical schemes for two-dimensional incompressible Navier–Stokes equations. SIAM J. Numer. Anal., 54(5):3123–3144, 2016.DOI
  3. [3]P. Constantin and C. Foias. Navier–Stokes Equations. University of Chicago Press, Chicago, 1988.DOI
  4. [4]T. Geveci. On the convergence of a time discretization scheme for the Navier–Stokes equations. Math. Comput., 53(187):43–53, 1989.DOI
  5. [5]S. Gottlieb, F. Tone, C. Wang, X. Wang, and D. Wirosoetisno. Long time stability of a classical efficient scheme for two-dimensional Navier–Stokes equations. SIAM J. Numer. Anal., 50(1):126–150, 2012.arxiv.org/abs/1105.4349
  6. [6]D. Han and X. Wang. Long-time stable SAV-BDF2 numerical schemes for the forced Navier–Stokes equations. arXiv preprint arXiv:2410.06362, 2024.arxiv.org/abs/2410.06362
  7. [7]T. Heister, M. A. Olshanskii, and L. G. Rebholz. Unconditional long-time stability of a velocity–vorticity method for the 2D Navier–Stokes equations. Numer. Math., 135(1):143–167, 2017.arxiv.org/abs/1511.08072
  8. [8]John G Heywood and Rolf Rannacher. Finite element approximation of the nonstationary Navier–Stokes problem, part II: stability of solutions and error estimates uniform in time. SIAM journal on numerical analysis, 23(4):750–777, 1986.
  9. [9]F. Huang and J. Shen. Stability and error analysis of a class of high-order IMEX schemes for Navier–Stokes equations with periodic boundary conditions. SIAM J. Numer. Anal., 59(6):2926–2954, 2021.arxiv.org/abs/2103.11025
  10. [10]F. Huang and J. Shen. Stability and error analysis of a second-order consistent splitting scheme for the Navier–Stokes equations. SIAM J. Numer. Anal., 61(5):2408–2433, 2023.DOI
  11. [11]F. Huang and J. Shen. On a new class of BDF and IMEX schemes for parabolic type equations. SIAM J. Numer. Anal., 62(4):1609–1637, 2024.arxiv.org/abs/2405.00300
  12. [12]F. Huang and J. Shen. Stability and error analysis of a new class of higher-order consistent splitting schemes for the Navier–Stokes equations. Math. Comput., 95(362):2789–2816, 2026.arxiv.org/abs/2507.01296
  13. [13]A. R. Humphries and A. M. Stuart. Runge–Kutta methods for dissipative and gradient dynamical systems. SIAM J. Numer. Anal., 31(5):1452–1485, 1994.DOI
  14. [14]F. Jauberteau, C. Rosier, and R. Temam. A nonlinear Galerkin method for the Navier–Stokes equations. Comput. Methods Appl. Mech. Eng., 80(1-3):245–260, 1990.
  15. [15]N. Ju. On the global stability of a temporal discretization scheme for the Navier–Stokes equations. IMA J. Numer. Anal., 22(4):577–597, 2002.
  16. [16]X. Li, J. Shen, and Z. Liu. New SAV-pressure correction methods for the Navier–Stokes equations: Stability and error analysis. Math. Comput., 91(333):141–167, 2022.arxiv.org/abs/2002.09090
  17. [17]H. Liao, H. Wang, and X. Wang. An efficient IMEX-SDIRK2 mr-ccSAV scheme for the forced Navier–Stokes equations with uniform-in-time enstrophy bounds. arXiv preprint arXiv:2608.22445, 2026.arxiv.org/abs/2608.22445
  18. [18]L. Lin, Z. Yang, and S. Dong. Numerical approximation of incompressible Navier–Stokes equations based on an auxiliary energy variable. J. Comput. Phys., 388:1–22, 2019.DOI
  19. [19]Q.-Y. Lin and M.-C. Shiue. On the unconditional long-time L² stability of the SAV–BDF2 time-stepping schemes for the forced Navier–Stokes equations with nonsmooth initial data. Appl. Math. Lett., 171:109665, 2025.DOI
  20. [20]J.-G. Liu, J. Liu, and R. L. Pego. Stability and convergence of efficient Navier–Stokes solvers via a commutator estimate. Commun. Pure Appl. Math., 60(10):1443–1487, 2007.DOI
  21. [21]C. Lubich and A. Ostermann. Runge–Kutta time discretization of reaction–diffusion and Navier–Stokes equations: Nonsmooth-data error estimates and applications to long-time behaviour. Appl. Numer. Math., 22(1–3):279–292, 1996.DOI
  22. [22]A. J. Majda and X. Wang. Nonlinear Dynamics and Statistical Theories for Basic Geophysical Flows. Cambridge University Press, Cambridge, 2006.DOI
  23. [23]L. G. Rebholz and F. Tone. Long-time H¹-stability of BDF2 time stepping for 2D Navier–Stokes equations. Appl. Math. Lett., 141:108624, 2023.DOI
  24. [24]J. Shen. Long time stability and convergence for fully discrete nonlinear Galerkin methods. Appl. Anal., 38(4):201–229, 1990.DOI
  25. [25]J. Shen, J. Xu, and J. Yang. The scalar auxiliary variable (SAV) approach for gradient flows. J. Comput. Phys., 353:407–416, 2018.DOI
  26. [26]J. Shen, J. Xu, and J. Yang. A new class of efficient and robust energy stable schemes for gradient flows. SIAM J. Numer. Anal., 57(3):1647–506, 2019.arxiv.org/abs/1710.01331
  27. [27]M.-C. Shiue. On the unconditional long-time L²-stability of the BDF2 time stepping scheme for the two-dimensional Navier–Stokes equations. Appl. Numer. Math., 214:104–109, 2025.DOI
  28. [28]J. C. Simo and F. Armero. Unconditional stability and long-term behavior of transient algorithms for the incompressible Navier–Stokes and Euler equations. Comput. Methods Appl. Mech. Eng., 111(1–2):111–154, 1994.DOI
  29. [29]R. Temam. Infinite-Dimensional Dynamical Systems in Mechanics and Physics, volume 68 of Applied Mathematical Sciences. Springer, New York, 1988.
  30. [30]F. Tone. On the long-time stability of the Crank–Nicolson scheme for the 2D Navier–Stokes equations. Numer. Methods Partial Differ. Equ., 23(5):1235–1248, 2007.DOI
  31. [31]F. Tone and D. Wirosoetisno. On the long-time stability of the implicit Euler scheme for the two-dimensional Navier–Stokes equations. SIAM J. Numer. Anal., 44(1):29–40, 2006.DOI
  32. [32]H. Wang, X. Wang, and M. Zhang. A linear variable-step embedded ETD scheme with uniform-in-time stability for the 2D Navier–Stokes equations. Numer. Methods Partial Differ. Equ., 42(4):e70132, 2026.
  33. [33]H. Wang, X. Wang, and M. Zhang. Unconditionally long-time stable variable-step second-order ETD schemes for the 2D periodic incompressible NSE. SIAM J. Sci. Comput., 2026. Accepted for publication; available at arXiv:2602.10268.arxiv.org/abs/2602.10268
  34. [34]X. Wang. An efficient second order in time scheme for approximating long time statistical properties of the two dimensional Navier–Stokes equations. Numer. Math., 121(4):753–779, 2012.arxiv.org/abs/1108.5409
  35. [35]K. Wu, F. Huang, and J. Shen. A new class of higher-order decoupled schemes for the incompressible Navier–Stokes equations and applications to rotating dynamics. J. Comput. Phys., 458:111097, 2022.DOI

Paper details

vlnc:W_fcfykbdg
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