All articles

Non-uniqueness for Navier–Stokes in three dimensions

Leray solutions, self-similar variables and the construction of an unstable profile for the forced equation.

Abstract

In this article, we review the conditional uniqueness property for Navier–Stokes equation, and follow the paper [ABC22] to develop a non-uniqueness theorem of 3-dimensional case with source term. We study the linearized operator of the self-similar equation, then reduce it gradually to a 2-dimensional Euler equation. The construction of the stationary unstable velocity field relies fundamentally on Vishik’s unstable profile in [Vis18a] and [Vis18b], which we then lift into 3-dimensional axisymmetric coordinate.

The main improve compared to the original paper is that we composed the proofs in an uniform way, viewing the linearized operators of perturbations of skew-adjoint operators. The case without exterior force is still an open problem.

Transcription of the manuscript dated 15 February 2025. Historical statements and notation follow that version. A transcription note at the end of the article distinguishes several source typos from the conventions used in the numerical experiment.

1. Introduction

NOTE: Due to the length request of the Memoire, this article is a short version which only contains sketch of proof. If you are interested in the full-length proof, please refer to the longer version.

In this article, we are interested in the following Navier–Stokes Equation System with source term in dimension \(N=2,3\), which describes the evolution of local velocity with respect to time of a viscous non-compressible fluid.

\[\begin{cases}\partial_tu+(u\cdot\nabla)u-\nu\Delta u=f-\nabla p,&(t,x)\in\mathbb R_+\times\mathbb R^N,\\\nabla\cdot u=0,&(t,x)\in\mathbb R_+\times\mathbb R^N,\\u|_{t=0}=u_0,&x\in\mathbb R^N.\end{cases}\tag{N-S}\]

where \(u:\mathbb R_{\geq0}\times\mathbb R^N\to\mathbb R^N\) is a \(N\)-dimensional vector field, \(p:\mathbb R_{\geq0}\times\mathbb R^N\to\mathbb R\) is the pressure function, \(f:\mathbb R_{\geq0}\times\mathbb R^N\to\mathbb R^N\) is the given external force applied to the fluid, and \(u_0:\mathbb R^N\to\mathbb R^N\) is the given initial velocity field at time \(t=0\). The parameter \(\nu>0\) is called the viscosity factor, illustrates how much resistance the fluid produces when being deformed at a given rate, the term \(-\nu\Delta\) gives the regularity of the system. When there is no viscosity, i.e. \(\nu=0\), then we obtain the Euler equation with source term:

\[\begin{cases}\partial_tu+(u\cdot\nabla)u=f-\nabla p,&(t,x)\in\mathbb R_+\times\mathbb R^N,\\\nabla\cdot u=0,&(t,x)\in\mathbb R_+\times\mathbb R^N,\\u|_{t=0}=u_0,&x\in\mathbb R^N.\end{cases}\tag{Euler}\]

In the light of the proper rescaling below, we assume that \(\nu\equiv1\):

\[u'(t,x):=\nu u(\nu t,x),\qquad p'(t,x):=\nu^2p(\nu t,x),\qquad f'(t,x):=\nu^2f(\nu t,x).\]

Since \(u\) is divergence free, we apply the divergence operator to the first equation of system (N-S) and obtain the dependence of \(p\) on \(u\) and \(f\):

\[\Delta p=\nabla\cdot f-\nabla\cdot((u\cdot\nabla)u).\tag{1.1}\]

Here we have to introduce some notations, using row vectors, we could write the velocity and divergence operator as \(u=(u_1,\ldots,u_N)\), \(\nabla=(\partial_1,\ldots,\partial_N)\), we then define the operations below:

  1. For two vectors \(u^{(\alpha)}=(u^{(\alpha)}_1,u^{(\alpha)}_2,\ldots,u^{(\alpha)}_N)\in\mathbb R^N\), \(\alpha=1,2\), we define the tensor product of these two vectors as a \(N\times N\) matrix: \[u^{(1)}\otimes u^{(2)}:=(u^{(1)})^T(u^{(2)})=(u_i^{(1)}u_j^{(2)})_{1\leq i,j\leq N}\in\mathbb R^{N\times N}.\]
  2. For two \(N\times N\) matrices \(A^{(\alpha)}=(A^{(\alpha)}_{ij})_{1\leq i,j\leq N}\in\mathbb R^{N\times N}\), \(\alpha=1,2\), we define the contraction product of them: \[A^{(1)}:A^{(1)}:=\sum_{1\leq i,j\leq N}A^{(1)}_{ij}A^{(2)}_{ij}.\]
  3. For a matrix \(A=(A_{ij})_{1\leq i,j\leq N}\in\mathbb R^{N\times N}\), we define the divergence of \(A\): \[\nabla\cdot A=(\partial_1,\ldots,\partial_N)A^T=\left(\sum_{j=1}^N\partial_jA_{ij}\right)_{1\leq i\leq N}.\]

Then for all \(u,v\in C^\infty(\mathbb R^N)^N\) such that \(\nabla\cdot v=0\), we have:

\[\nabla\cdot(u\otimes v)=(u\cdot\nabla)v+u(\nabla\cdot v)=(u\cdot\nabla)v.\tag{1.2}\]

For \(u\in C_w([0,T],L^2(\mathbb R^N))\cap L^2([0,T],\dot H^1(\mathbb R^N))\), where \(C_w\) refers to continuous with respect to the weak topology, define the kinetic energy:

\[E_u(t):=\tfrac12\|u(t)\|_{L^2(\mathbb R^N)}^2+\nu\|u\|_{L^2([0,t],\dot H^1(\mathbb R^N))}^2,\qquad\forall t\in[0,T].\tag{1.3}\]

For \(T>0\), we define the proper space \(G_T\) of the external force:

\[G_T:=L^1([0,T],L^2(\mathbb R^N)^N)+L^2([0,T],\dot H^{-1}(\mathbb R^N)^N).\]

We follow the approach of Jean Leray [Ler34], and define the solution:

Definition 1.1 (Turbulent Solution). For all times \(T>0\), let \(u_0\in L^2(\mathbb R^N)^N\), such that \(\nabla\cdot u_0=0\), \(f\in G_T\), we say that \(u\in\mathcal S'([0,T]\times\mathbb R^N)^N\) is a turbulent solution of equation (N-S) on the time interval \([0,T]\), if:

\[u\in C_w([0,T],L^2(\mathbb R^N)^N)\cap L^2([0,T],\dot H^1(\mathbb R^N)^N).\]

For all function \(\psi\in\mathcal D([0,T]\times\mathbb R^N)\), for all \(t\in[0,T]\), we have:

\[\int_{\mathbb R^N}\nabla\psi(t,x)\cdot u(t,x)\,dx=0.\]

And for all functions \(\phi\in C^1([0,T],H^1(\mathbb R^N)^N)\), such that \(\nabla\cdot\phi=0\), we have:

\[\begin{aligned}&\int_{\mathbb R^N}u(T,x)\cdot\phi(T,x)\,dx-\int_0^T\int_{\mathbb R^N}u(s,x)\cdot\partial_t\phi(s,x)\,dx\,ds\\ &=-\nu\int_0^T\int_{\mathbb R^N}\nabla u(s,x):\nabla\phi(s,x)\,dx\,ds+\int_{\mathbb R^N}u_0(x)\cdot\phi(0,x)\,dx\\ &\quad+\int_0^T\int_{\mathbb R^N}u\otimes u(s,x):\nabla\phi(s,x)\,dx\,ds+\int_0^T\int_{\mathbb R^N}f(s,x)\cdot\phi(s,x)\,dx\,ds.\end{aligned}\tag{1.4}\]

Finally, the solution satisfies the following Energy Inequality:

\[E_u(t)\leq\tfrac12\|u_0\|_{L^2(\mathbb R^N)}^2+\int_0^t\int_{\mathbb R^N}f\cdot u\,dx\,ds,\qquad\forall t\in[0,T].\tag{1.5}\]

[Ler34] has provided the existence of turbulent solution:

Proposition 1.1. Let \(N=2,3\), given initial data \(u_0\in L^2(\mathbb R^N)^N\) divergence free, then for all \(T>0\), if \(f\in G_T\) there exists a turbulent solution \(u\) of the system (N-S) on time interval \([0,T]\).

For the uniqueness of the turbulent solution, [LP59] has proved it for dimension \(N=2\), which essentially relies on the fact that proper rescaling does not change the corresponding \(L_t^\infty L_x^2\) and \(L_t^2\dot H_x^1\) norm of the solution.

The problem for the uniqueness for general turbulent solutions in dimension \(N=3\) has remained long open. Leray’s work has proven the “strong-weak” uniqueness of the turbulent solution, i.e. if the solution is strong, which satisfies a certain regularity condition, then all turbulent solutions coincide.

In this article, we consider the equation (N-S) with source term not regular enough to admit a strong solution, then, the “strong-weak” principle cannot be applied. We follow the lines of [ABC22], to construct a source term which gives a counter-example for the uniqueness in dimension \(N=3\):

Theorem 1.1. Assume \(\nu=1\), then there exists a time \(T>0\), an external source term \(f\in L^1([0,T],L^2(\mathbb R^N)^N)\), and two distinct turbulent solutions \(u\) and \(\tilde u\) to the system (N-S) on time interval \([0,T]\) with initial condition \(u_0\equiv0\).

2. Leray’s Turbulent Solution

This section is dedicated to give an outline of the proof of Leray’s existence theorem (1.1), we consider the case \(f\in L^2_{\mathrm{loc}}(\mathbb R,\dot H^{-1}(\mathbb R^N)^N)\), the general case could be treated with minor modification.

In this section, we use the following convention:

\[H:=L^2(\mathbb R^N)^N,\qquad V:=H^1(\mathbb R^N)^N,\qquad V':=\dot H^{-1}(\mathbb R^N)^N,\]

and the space with subscript \(\sigma\) is the divergence-free subspace.

2.1. Construction of Approximate Solutions

Firstly, we want to construct a sequence of approximate solutions to the Navier–Stokes equation.

To get rid of the pressure \(p\), we apply the following Leray operator onto the equation (N-S):

\[\mathbb P:\mathcal M\longrightarrow\mathcal M,\qquad f=(f_1,\ldots,f_N)\longmapsto\left(\delta_{i,j}-\frac{D_iD_j}{|D|^2}\right)_{i,j=1}^{N}f.\tag{2.1}\]

where \(\mathcal M=H^k(\mathbb R^N)^N\) or \(\dot H^k(\mathbb R^N)^N\), \(k\in\mathbb R\). We use the Fourier multiplier convention:

\[\left(\delta_{i,j}-\frac{D_iD_j}{|D|^2}\right)_{i,j=1}^{N}f=\mathcal F^{-1}\left[\left(\delta_{i,j}-\frac{\xi_i\xi_j}{|\xi|^2}\right)_{i,j=1}^{N}\hat f\right].\]

Let \((P_k)_{k\in\mathbb N}\) be a family of spectral projections approximating \(\mathbb P\), which satisfies the following properties: \(\forall f\in V'\), then we have \(P_kf\in V_\sigma\), and:

\[\|P_kf\|_{V'_\sigma}\leq\|f\|_{V'_\sigma},\qquad\lim_{k\to+\infty}\|P_kf-\mathbb Pf\|_{V'_\sigma}=0.\]

Due to the length request, we deliberately omit the construction of \((P_k)_{k\in\mathbb N}\), denote by \(H_k\) the space \(P_kH\).

Definition 2.1. Let us define the (continuous) bilinear map

\[Q:V\times V\longrightarrow V',\qquad(u,v)\longmapsto-\nabla\cdot(u\otimes v)=-\left(\sum_{i=1}^{N}\partial_i(u_iv_j)\right)_{1\leq j\leq N}.\]

The continuity of \(Q\) from the Sobolev embedding. Moreover, using integration by parts and a smoothing argument, we have:

\[\langle Q(u,v),v\rangle=0,\qquad(u,v)\in V_\sigma\times V.\]

Let us take \((f_k)_{k\in\mathbb N}\) an approximation of \(f\) in the space \(C^1(\mathbb R_+,V_\sigma)\) such that \(f_k(t)\in H_k\) and \(\lim_{k\to\infty}\|f_k-f\|_{L^2([0,T],V'_\sigma)}=0\). Now, let us introduce the following ordinary differential equation:

\[\begin{cases}\dot u_k(t)=P_k\Delta u_k(t)+F_k(u_k(t))+f_k(t),\\u_k(0)=P_ku_0,\end{cases}\tag{2.2}\]

with \(F_k(u)=P_kQ(u,u)\).

  1. Thanks to linearity of \(P_k\Delta:H_k\to H_k\) and continuity of \(Q\), the Cauchy–Lipschitz theorem give the existence of a unique maximal solution \(u_k\) of \((NS_k)\) in \(C^\infty([0,T_k[,H_k)\) with \(T_k\in]0,+\infty]\).
  2. We have a control of \(\|\dot u_k(t)\|_{L^2}\) with the following inequality: \[\|\dot u_k(t)\|_{L^2}\leq k\|u_k(t)\|_{L^2}+Ck^{4/d}\|u_k(t)\|_{L^2}^2+\|f_k(t)\|_{L^2}.\]
  3. If \(T_k<+\infty\) and \(\|u_k(t)\|_{L^2}\) is bounded on the interval \([0,T_k[\), so does \(\|\dot u_k(t)\|_{L^2}\) and then we can extend the solution beyond \(T_k\).
  4. Actually, because each \(u_k\) verifies the energy equality, we can show that \((u_k)_{k\in\mathbb N}\) remains uniformly bounded on \(H\) for all time, hence that \(T_k=+\infty\). Moreover, \((u_k)_{k\in\mathbb N}\) is bounded in the space \(L^\infty_{\mathrm{loc}}(\mathbb R_+,H)\cap L^2_{\mathrm{loc}}(\mathbb R_+,V_\sigma)\cap L^{8/d}_{\mathrm{loc}}(L^4(\Omega))\) and \((\Delta u_k)_{k\in\mathbb N}\) is bounded in the space \(L^2_{\mathrm{loc}}(\mathbb R_+,V'_\sigma)\).

2.2. Compactness Properties

The next result will be crucial for the rest of the proof. It will give us a candidate for the global weak solution to (N-S).

Proposition 2.1. A vector field \(u\) exists in \(L^2_{\mathrm{loc}}(\mathbb R_+,V_\sigma)\) such that up to an extraction (which we omit to note) \((u_k)_{k\in\mathbb N}\) converges to \(u\) strongly in \(L^2_{\mathrm{loc}}(\mathbb R_+\times\mathbb R^N)\) and weakly in \(L^2_{\mathrm{loc}}(\mathbb R_+,V_\sigma)\). Moreover, \((u_k(t))_{k\in\mathbb N}\) converges weakly to \(u(t)\) uniformly on \([0,T]\) for functions in \(C^1(\mathbb R_+,V_\sigma)\).

Let us consider a test function \(\Psi\in C^1(\mathbb R_+,V_\sigma)\). By definition of \(u_k\), we have

\[\begin{aligned}\frac{d}{dt}\langle u_k(t),\Psi(t)\rangle&=\langle\dot u_k(t),\Psi(t)\rangle+\langle u_k(t),\dot\Psi(t)\rangle\\&=\langle P_k\Delta u_k(t),\Psi(t)\rangle+\langle P_kQ(u_k(t),u_k(t)),\Psi(t)\rangle\\&\quad+\langle f_k(t),\Psi(t)\rangle+\langle u_k(t),\dot\Psi(t)\rangle.\end{aligned}\]

After integrating by parts and integrating in time from \(0\) to \(t\), we obtain

\[\begin{aligned}&\int_\Omega u_k(t,x)\cdot\Psi(t,x)\,dx\\&\quad+\int_0^t\int_\Omega(\nabla u_k:\nabla P_k\Psi-u_k\otimes u_k:\nabla P_k\Psi-u_k\cdot\partial_t\Psi)(t',x)\,dx\,dt'\\&=\int_\Omega u_k(0,x)\cdot\Psi(0,x)\,dx+\int_0^t\langle f_k(t'),\Psi(t')\rangle\,dt'.\end{aligned}\]

By using the local strong convergence of \((u_k)_{k\in\mathbb N}\) we can pass to the limit and prove that \(u\) is a solution of (N-S) in the sense of (1.1).

2.3. End of the Proof

Thanks to Proposition 1.2, we can show by using properties of weak convergence that for all \(t\geq0\),

\[\|u(t)\|_{L^2}^2\leq\liminf_{k\to+\infty}\|u_k(t)\|_{L^2}^2,\qquad\int_0^t\|\nabla u(t')\|_{L^2}^2\,dt'\leq\liminf_{k\to+\infty}\int_0^t\|\nabla u_k(t')\|_{L^2}^2\,dt'.\]

Then we can deduce the energy inequality.

Remark: This is precisely at this moment that we lose the energy equality. Actually, we can show that \(u_k\) verifies also the energy equality, but the weak convergence of \((u_k)_k\) does not allow \(u\) to verify the equality, reason why we have only an inequality and then we cannot maintain the uniqueness of the solutions.

Finally, let’s prove the time continuity of \(u\): for any \(\Psi\in V_\sigma\), because \(u\) satisfies the weak formulation,

\[\begin{aligned}\langle u(t),\Psi\rangle-\langle u(t'),\Psi\rangle&=\int_t^{t'}\langle f(s),\Psi\rangle\,ds\\&\quad+\int_t^{t'}\int_\Omega(\nabla u(s):\nabla\Psi-u\otimes u(s):\nabla\Psi)\,dx\,ds.\end{aligned}\]

The inequality \(\|u(t)\|_{L^4}\leq\|u(t)\|_{L^2}^{1-d/4}\|\nabla u(t)\|_{L^2}^{d/4}\) implies \(u\in L^{8/d}([0,T],L^4(\mathbb R^N)^N)\). Then we deduce that

\[\begin{aligned}|\langle u(t),\Psi\rangle-\langle u(t'),\Psi\rangle|&\leq |t-t'|^{1-d/4}\|u\|_{L^{8/d}([0,T],L^4(\Omega))}^2\|\Psi\|_{V_\sigma}\\&\quad+|t-t'|^{1/2}\big(\|f\|_{L^2([0,T],V'_\sigma)}+\|\nabla u\|_{L^2([0,T],L^2)}\big)\|\Psi\|_{V_\sigma}.\end{aligned}\]

which concludes the proof of (1.1).

3. Previous Results: A Revisit

In this section, we review the uniqueness properties of the turbulent solutions in dimension \(N=2,3\).

3.1. Case 2D: Unconditional Uniqueness

Proposition 3.1. Given initial data \(u_0\in L^2(\mathbb R^2)^2\), \(f\in L^1([0,T],\dot H^{-1}(\mathbb R^2)^2)\), then there exists a unique turbulent solution \(u\) of the system (N-S) defined on time interval \([0,T]\).

Sketch of Proof. The proof of uniqueness essentially relies on Sobolev embedding and Grönwall’s inequality. Assume that there exists two turbulent solution \((u,p_u)\) and \((v,p_v)\), where \(p\) denotes the associated pressure function.

Apply the Leray projector (2.1) onto the equation, we obtain:

\[\begin{cases}\partial_tw-\Delta w=-\mathbb P(\nabla\cdot(u\otimes u-v\otimes v)),\\\nabla\cdot w=0,\\w|_{t=0}\equiv0.\end{cases}\tag{3.1}\]

Apply Sobolev inequality \(H^{1/2}(\mathbb R^2)\hookrightarrow L^4(\mathbb R^2)\),

\[\|\mathbb P(\nabla\cdot(u\otimes u-v\otimes v))\|_{L_t^2\dot H_x^{-1}}\lesssim\big(\|u\|_{L_t^4\dot H_x^{1/2}}+\|v\|_{L_t^4\dot H_x^{1/2}}\big)\|w\|_{L_t^4\dot H_x^{1/2}}.\]

And by interpolation inequality,

\[\|f\|_{L_t^4\dot H_x^{1/2}}\lesssim\|f\|_{L^\infty([0,T],L^2(\mathbb R^2)^2)}^{1/2}\|f\|_{L^2([0,T],\dot H^1(\mathbb R^2)^2)}^{1/2}.\]

Thus, the source term is in the classical energy space, apply standard energy estimate for heat equation, combine all estimates above:

\[\begin{aligned}&\tfrac12\|w(t)\|_{L^2(\mathbb R^2)^2}^2+\int_0^t\|w(s)\|_{\dot H^1(\mathbb R^2)^2}^2\,ds\\&\lesssim\int_0^t\big(\|u(s)\|_{\dot H^{1/2}(\mathbb R^2)^2}+\|v(s)\|_{\dot H^{1/2}(\mathbb R^2)^2}\big)\|w(s)\|_{L^2(\mathbb R^2)^2}^{1/2}\|w(s)\|_{\dot H^1(\mathbb R^2)^2}^{3/2}\,ds\\&\leq\int_0^t\left[c\big(\|u(s)\|_{\dot H^{1/2}(\mathbb R^2)^2}+\|v(s)\|_{\dot H^{1/2}(\mathbb R^2)^2}\big)^4\|w(s)\|_{L^2(\mathbb R^2)^2}^2+\tfrac12\|w(s)\|_{\dot H^1(\mathbb R^2)^2}^2\right]ds.\end{aligned}\]

where in the last line we applied the multi-variable case for arithmetic-geometric inequality. Absorb \(H^1\) term in the right, and the conclusion follows from Grönwall’s inequality.

3.2. Case 3D: Strong-Weak Principle

Proposition 3.2. Given initial data \(u_0\in L^2(\mathbb R^3)^3\), \(f\in L^2([0,T],\dot H^{-1}(\mathbb R^3)^3)\), then if the system (N-S) admits a turbulent solution \(u\) defined on \([0,T]\), furthermore satisfies the extra regularity below:

\[u\in L^4([0,T],\dot H^1(\mathbb R^3)^3),\]

Then, every turbulent solution \(v\) defined on \([0,T]\) coincide with \(u\).

Sketch of Proof. Similarly, we consider the difference of two solutions:

\[w:=u-v\in C_w([0,T],L^2(\mathbb R^3)^3)\cap L^2([0,T],\dot H^1(\mathbb R^3)^3).\]

Then, we have the following estimate of energy comes from the energy inequality:

\[\begin{aligned}E_w(t)&\leq\|u_0\|_{L^2(\mathbb R^3)^3}+\int_0^t\int_{\mathbb R^N}f\cdot(u+v)\,dx\,ds\\&\quad-\langle u(t),v(t)\rangle_{L^2(\mathbb R^3)^3}-2\int_0^t\langle\nabla u(s),\nabla v(s)\rangle_{L^2(\mathbb R^3)^9}\,ds.\end{aligned}\tag{3.2}\]

The extra regularity of \(u\) guarantees that we could couple the (N-S) of \(u\) with \(v\) and vice-versa:

\[\int_0^t\int_{\mathbb R^3}\partial_tu\cdot v+(\nabla u-u\otimes u):\nabla v\,dx\,ds=\int_0^t\int_{\mathbb R^3}f\cdot v\,dx\,ds,\tag{3.3}\] \[\begin{aligned}&\langle v(t),u(t)\rangle_{L^2(\mathbb R^3)^3}-\int_0^t\int_{\mathbb R^3}v\cdot\partial_tu+(v\otimes v-\nabla v):\nabla u\,dx\,ds\\&=\|u_0\|_{L^2(\mathbb R^3)^3}^2+\int_0^t\int_{\mathbb R^3}f\cdot u\,dx\,ds.\end{aligned}\tag{3.4}\]

Combining (3.2), (3.3) and (3.4), we have:

\[E_w(t)\leq\int_0^t\int_{\mathbb R^3}u\otimes u:\nabla v+v\otimes v:\nabla u\,dx\,ds=\int_0^t\int_{\mathbb R^3}(w\cdot\nabla)u\cdot w\,dx\,ds.\]

The last equality is by divergence free property. Apply Sobolev embedding:

\[\int_{\mathbb R^3}(w\cdot\nabla)u\cdot w\,dx\lesssim\|w\|_{L^2(\mathbb R^3)^3}^{1/2}\|w\|_{\dot H^1(\mathbb R^3)^3}^{3/2}\|u\|_{\dot H^1(\mathbb R^3)^3}.\]

Similarly absorb \(\dot H^1\) term, and the conclusion follows from Grönwall’s inequality.

4. Non-Uniqueness: Review and Foreseen

In this section, we are going to briefly review the history of non-uniqueness problem of Navier–Stokes equation, and give a self-similar refined version of the main theorem (1.1).

4.1. A Brief History

Non-Uniqueness of Navier–Stokes equation has been a long question in different contexts, less regular the solution is, the more non-uniqueness it may holds.

In the weakest sense, Buckmaster and Vicol [BV19] have proven that in the class of finite kinetic energy distributions, for any given initial datum, the solution is always non-unique. However, their solution lies in the finite energy class \(L_t^\infty L_x^2\) but is not regular enough to be a Leray solution.

But in the context of Leray’s turbulent solution, the problem remains open. The method adopted by [ABC22], which is presented in this paper, is partially inspired by Vishik’s study of non-uniqueness of Euler’s equation with external force in [Vis18a] and [Vis18b]. The main idea is to consider a unstable steady state, and apply an external force as perturbation in \(L_t^1L_x^p\). We will utilize his method of constructing an unstable vortex solution.

How about removing the external force? It’s still an open problem for the uniqueness of (N-S) without external force. But in the context of Euler equation, Bressan, Murray, and Shen have proposed a potential initial data which potentially leads to non-unique solutions without external force in [BM20] and [BS21]. We still need further exploration for similar conclusions on Navier–Stokes equation.

4.2. Strategy of Refining

The scheme is described below: The term unstable refers to the fact that the linearized operator around a given stationary profile in some equivalent coordinate has an unstable eigenvalue, and the corresponding perturbation term turns to \(0\) while \(t\to-\infty\). It’s hard to treat the question on positive time \(\mathbb R_+\), thus, we introduce the similarity variables:

\[\xi:=\frac{x}{\sqrt t},\qquad\tau:=\log t\in\mathbb R,\tag{4.1}\] \[u(t,x)=:\frac1{\sqrt t}U(\tau,\xi),\qquad f(t,x)=:\frac1{t^{3/2}}F(\tau,\xi).\tag{4.2}\]

Here the lower case letters denote the functions in physical space, and the upper case letters denote functions with similarity variables. The Navier–Stokes equation becomes:

\[\begin{cases}\partial_\tau U-\tfrac12(1+\xi\cdot\nabla_\xi)U-\Delta U+U\cdot\nabla U=F-\nabla P,&(t,x)\in\mathbb R\times\mathbb R^3,\\\nabla\cdot U=0,&(\tau,\xi)\in\mathbb R\times\mathbb R^3.\end{cases}\tag{4.3}\]

We call a solution \(\tilde U\) to (4.3) linear unstable, if the linearized operator \(L_{ss}\) at \(\tilde U\) defined below has eigenvalue with positive real part.

\[L_{ss}(U):=\tfrac12(1+\xi\cdot\nabla_\xi)U+\Delta U+\mathbb P(\tilde U\cdot\nabla U+U\cdot\nabla\tilde U).\tag{4.4}\]

where \(\mathbb P\) is the Leray projector defined in (2.1). We will study the unstable manifold associated with the most unstable eigenvalue \(\lambda_0\), i.e. the eigenvalue with the largest real part which we denote as \(a=\Re(\lambda_0)\in\mathbb R\). Then, the solutions on this manifold satisfy the following asymptotic behavior:

\[U=\tilde U+U^{\mathrm{lin}}+O(e^{2\tau a}),\qquad\tau\to-\infty,\]

where \(U^{\mathrm{lin}}\) is a non-trivial solution to the linearized equation \(\partial_tU^{\mathrm{lin}}=L_{ss}U^{\mathrm{lin}}\) on \(\mathbb R\times\mathbb R^3\) corresponding to an eigenfunction of \(\lambda_0\). Since \(U^{\mathrm{lin}}\) decays at the rate \(e^{\tau a}\) as \(\tau\to-\infty\), thus,

\[\lim_{\tau\to-\infty}U=\tilde U.\]

Since we can view \(\tau=-\infty\) as \(t=0\), and by the formula (4.2), the corresponding velocity field in physical space satisfies:

\[\lim_{t\to0+}\|u(t)\|_{L^2(\mathbb R^3)^3}=\lim_{\tau\to-\infty}e^{\tau/2}\|U(\tau)\|_{L^2(\mathbb R^3)^3}=0.\tag{4.5}\]

The same holds true for \(\tilde u\), this corresponds to the non-uniqueness of \(t=0\) in time variable with initial value \(u_0=0\). We obtain a refined version of (1.1):

Theorem 4.1. There exists a smooth, compactly supported stationary real-valued velocity field \(\tilde U\) which does not depend on the self-similar time \(\tau\), and a smooth, compactly supported source term

\[\tilde F:=-\tfrac12(1+\xi\cdot\nabla_\xi)\tilde U-\Delta\tilde U+\tilde U\cdot\nabla\tilde U,\]

satisfying the following properties:

  1. The linearized operator \(L_{ss}\) defined by (4.4) has an unstable eigenvalue \(\lambda\) with a non-zero eigenfunction \(\eta\in H^k(\mathbb R^3)\) for all \(k\geq0\). Denote one of the solutions to the linearized equation \(\partial_tU^{\mathrm{lin}}=L_{ss}U^{\mathrm{lin}}\) as below: \[U^{\mathrm{lin}}(\tau):=\Re(e^{\lambda\tau}\eta).\]
  2. There exists a time-like upper bound \(T\in\mathbb R\) and a velocity field \(U^{\mathrm{per}}\), satisfying the following smallness condition. For all \(k\in\mathbb N\), we have: \[\|U^{\mathrm{per}}(\tau)\|_{H^k(\mathbb R^3)}\lesssim_k e^{2\tau a},\qquad\forall\tau\in(-\infty,T].\]

    And in view of (4.5), \(\tilde u\) and \(\tilde u+u^{\mathrm{lin}}+u^{\mathrm{per}}\) are two distinct turbulent solutions to the system (N-S) with initial value \(u_0\equiv0\) and external force \(\tilde f\).

Furthermore, we can make the following observations:

  • The external force gives us a large range of choice. And by simply modifying the amplitude, we can view nonlinear convection term as the leading term, and the other terms, including viscosity, act as small perturbations.

Thus, the problem is reduced to finding a smooth unstable steady state for the 3-dimensional Euler equation with external force. It’s natural to impose extra symmetry and lower the dimension to 2D case, a frequently considered family of steady state functions are so-called vortices, which write in polar coordinates as:

\[\bar u(r,\theta)=\bar u^\theta(r)e_\theta,\qquad(r,\theta)\in\mathbb R_+\times(\mathbb R\setminus2\pi\mathbb Z),\]

where \(e_\theta\) denotes the clockwise tangent vector at \((1,\theta)\) of the unit circle. Vishik has constructed an unstable vortex with power-law decay as \(r\to+\infty\). Our method is lifting Vishik’s vortex in a axisymmetric without swirl way into 3 dimensions, of the form

\[u(r,\theta,z)=u^r(r,z)e_r+u^z(r,z)e_z,\]

where \(e_r,e_z\) are respectively the unit vector in the \(r\) and \(z\)-direction. The corresponding vorticity and stream function are of following form:

\[\omega:=\nabla\times u=-\omega^\theta(r,z)e_\theta,\qquad\phi=-\Delta^{-1}\omega=\phi^\theta(r,z)e_\theta.\tag{4.6}\]

Then, with these variables, applying the curl operator on the Euler equation (Euler) with \(f=0\), together with the definition of \(\omega,\phi\), leads to the following system:

\[\begin{cases}\partial_t\omega^\theta+u\cdot\nabla\omega^\theta-\dfrac{u^r}{r}\omega^\theta=0,\\\omega^\theta=\left(\partial_r^2+\dfrac1r\partial_r-\dfrac1{r^2}\right)\phi^\theta,\\u=-\partial_z\phi^\theta e_r+\left(\partial_r+\dfrac1r\right)\phi^\theta e_z.\end{cases}\tag{4.7}\]

As \(r\to+\infty\), all the terms of order \(r^{-1}\) and \(r^{-2}\) can be neglected, so we reduce to the following 2-dimensional Euler-like system:

\[\begin{cases}\partial_t\omega+u\cdot\nabla\omega=0,\\\omega=\Delta\phi,\qquad u=\nabla^\perp\phi.\end{cases}\tag{4.8}\]

Hence, we can lift Vishik’s unstable 2D vortex into a 3D hollow vortex ring with large diameter in order to ease the influence of those higher order terms. This suggests us a possible way to construct the unstable steady state.

5. Instability: From 2D to Axisymmetric Case

In this section, we study the instability property of the 3-dimensional axisymmetrical Euler system (4.8). It relies fundamentally on the construction of Vishik’s unstable vortex for 2-dimensional Euler’s equation in [Vis18a] and [Vis18b].

5.1. 2D Instability: Vishik’s Vortex

This section is devoted to stating a version of unstable vortex constructed by Vishik in [Vis18b].

We briefly state the Biot–Savart Operator to construct a velocity profile when given a vorticity:

\[\mathrm{BS}_{2d}:L^2(\mathbb R^2)\longrightarrow\dot H^1(\mathbb R^2)^2,\qquad\omega\longmapsto\nabla\times\Delta^{-1}\omega,\tag{5.1}\]

and we can write the image above into the convolution form:

\[\nabla\times\Delta^{-1}\omega=K_{\mathrm{BS}_{2d}}*\omega,\qquad K_{\mathrm{BS}_{2d}}=\frac1{2\pi}\frac{x^\perp}{|x|^2},\tag{5.2}\]

where for \(x=(x_1,x_2)\), \(x^\perp=(-x_2,x_1)\).

Remark 5.1. Following the same procedure, we can define the Biot–Savart operator \(\mathrm{BS}_{3d}:L^2(\mathbb R^3)^3\to\dot H^1(\mathbb R^3)^3\) in the 3-dimensional case:

\[\mathrm{BS}_{3d}[\omega](x)=\int_{\mathbb R^3}K_{\mathrm{BS}_{3d}}(y)\times\omega(x-y)\,dy,\qquad K_{\mathrm{BS}_{3d}}(x):=-\frac1{4\pi}\frac{x}{|x|^3}\in L_w^{3/2}(\mathbb R^3).\tag{5.3}\]

Let us consider the behavior of the linearized operator associated with the Euler-like equation (4.8) around a steady state. Consider a smooth, real-valued, divergence-free vector field with sufficient decay at \(r\to+\infty\), under the form:

\[\bar u(r,\theta)=\zeta(r)e_\theta.\]

Set the corresponding vorticity which is a radial function:

\[\bar\omega=\nabla\times\bar u=\partial_r\zeta+\frac1r\zeta.\]

We would like to work on a subspace of \(L^2(\mathbb R^2)\) to obtain better decay properties of our eigenfunction. Let us introduce our desired weighted space, where we use the coordinate \((x,y)\) in \(\mathbb R^2\)

\[L_\gamma^2:=L^2(\mathbb R^2,\gamma\,dx\,dy),\tag{5.4}\]

where the smooth weight function \(\gamma>0\) satisfies:

\[\gamma|_{B_0(\bar R)}\equiv1,\qquad(1+x^2+y^2)^{50}\lesssim\gamma\lesssim(1+x^2+y^2)^{50}.\tag{5.5}\]

Now we consider the operator \(L_\infty\) defined below:

\[L_\infty:D(L_\infty)\longrightarrow L_\gamma^2,\qquad\omega\longmapsto-(u\cdot\nabla\bar\omega+\bar u\cdot\nabla\omega),\tag{5.6}\]

where \(u=\mathrm{BS}_{2d}[\omega]\), and the domain of definition

\[D(L_\infty)=\{\omega\in L_\gamma^2\mid\bar u\cdot\nabla\omega\in L_\gamma^2\}.\]

In fact, \(u\cdot\nabla\bar\omega\) is compact by Rellich’s embedding if given \(\nabla\tilde\omega\) with compact support, and \(\bar u\cdot\nabla\omega\) is skew-adjoint. We will use the following scheme for several times:

Remark 5.2. Consider a skew-adjoint linear operator \(L\) define on \(H\), we perturb \(L\) with a compact operator \(T\). Then, the essential spectrum satisfies:

\[\sigma_{\mathrm{ess}}(L+T)\subset\{x\in\mathbb C\mid\Re(x)\leq0\},\]

and the remaining spectrum of \(L+T\) consists of isolated eigenvalues of finite order in \(\{x\in\mathbb C\mid\Re(x)>0\}\).

We now state a modified version of Vishik’s unstable vortex:

Theorem 5.1. There exists a smooth, real, compactly supported velocity field expressed in polar coordinate \(\bar u(r,\theta)=\bar\zeta(r)e_\theta\), with its corresponding smooth vorticity \(\bar\omega(r,\theta)=\bar\omega(r)\), satisfying that the linearized operator \(L_\infty\) has an unstable eigenvalue \(\lambda\), i.e. \(\Re(\lambda)>0\).

5.2. Axisymmetric Instability

We want to lift the previous argument to the 3-dimensional axisymmetric-without-swirl case, with coordinates written as \((r',z)\in\mathbb R_+\times\mathbb R\), the function \(u\) defined on this coordinate is independent of the angle \(\theta\), and of the form:

\[u(r',z)=u^r(r',z)e_r+u^z(r',z)e_z,\qquad\omega:=\nabla\times u=-\omega^\theta(r,z)e_\theta.\]

In the discussion below, we will use 2D Euclidean coordinates to represent 3D behavior, beware that the operators are different from the original 2D operators!

To ease the influence of \(r^{-1}\) and \(r^{-2}\) term, the lifted vortex should be away from \(0\). Thus, we work on the shifted variable \(r:=r'-l\in\mathbb R_{\geq-l}\), where \(l>2\bar R\).

From now on, we use the following convention:

Convention 1. All the operators with subscript \(l\) are the 2-dimensional operators induced by the corresponding 3-dimensional operators applied to the shifted axisymmetric-without-swirl variable \((r',z)=(r+l,z)\).

Denote our 2-dimensional unstable vortex in Euclidean coordinate \((r,z)\in\mathbb R^2\) in alignment with 3D axisymmetrical case as:

\[\bar u(r,z)=\bar u^r(r,z)e_r+\bar u^z(r,z)e_z.\]

By the discussion above, it can be lifted to a 3-dimensional axisymmetric-without-swirl profile in the coordinate \((r,z)\), denoted also by \(\bar u\). Notice that \(\bar u\) is not a 3D divergence-free profile in the shifted coordinate \((r',z)\):

\[\nabla_l\cdot\bar u:=\left(\partial_r+\frac1{r+l}\right)\bar u^r+\partial_z\bar u^z=\frac1{r+l}\bar u^r,\]

where \(\nabla_l\) is the operator induced by the gradient of 3D shifted cylindrical coordinate \((r',z)\). Observe that the error term is decaying as \(l\to\infty\), we have:

Lemma 5.1. For all \(l\geq2\bar R\), there exists a real correction \(v_l\in C_0^\infty(B_0(\bar R),\mathbb R^2)\), satisfying:

\[\nabla_l\cdot(\bar u+v_l)=0,\]

and the following decay property:

\[v_l\xrightarrow[l\to\infty]{}0\quad\text{in }C^k(B_0(\bar R)),\qquad\forall k\geq0.\]

Sketch of Proof. Only need to solve the following equation:

\[\nabla_l\cdot u=\frac1{r+l}(\partial_r,\partial_z)\cdot[(r+l)u],\qquad(\partial_r,\partial_z)\cdot[(r+l)v_l]=-\bar u^r.\]

Recalling that \(\bar u^r\in C_0^\infty(B_0(\bar R))\) with \(\int_{\mathbb R^2}\bar u^r(r,z)\,dr\,dz=0\) inherited from its rotational symmetry, applying the Bogovskii’s operator in case \(N=2\), we obtain the desired conclusion.

For any \(l>2\bar R\), we can define the corrected background profile:

\[\bar u_l:=\bar u+v_l\in C_0^\infty(B_0(\bar R),\mathbb R^2),\qquad\bar\omega_l:=\nabla_l\times\bar u_l\in C_0^\infty(B_0(\bar R)).\]

As the flow is axisymmetric without swirl, we can identify the vorticity with its angular component as a scalar function:

\[\bar\omega_l=-\bar\omega_l^\theta e_\theta,\qquad\bar\omega^\theta=-\partial_z\bar u_l^r+\partial_r\bar u_l^z.\]

We consider the linearized operator of equation (4.7) around the background profile \((\bar u_l,\bar\omega_l)\). Firstly, we have to define the proper functional space we work on. As before, we expect that the functions are decaying sufficiently fast:

\[L_{\gamma,l}^2:=L^2(\mathbb R_{r\geq-l}\times\mathbb R_z,\gamma\,dr\,dz),\]

where \(\gamma\) is the same weight function defined in (5.5).

Define the linearized operator:

\[L_l:D(L_l)\longrightarrow L_{\gamma,l}^2,\qquad\omega\longmapsto L_l\omega,\tag{5.7}\]

where,

\[-L_l\omega:=\bar u_l\cdot\nabla\omega+\mathrm{BS}_l[\omega]\cdot\nabla\bar\omega_l-\frac{\bar u_l^r}{r+l}\omega-\frac{(\mathrm{BS}_l[\omega])^r}{r+l}\bar\omega_l.\tag{5.8}\]

\(\nabla=(\partial_r,\partial_z)\) is the normal gradient operator, \(\mathrm{BS}_l\) is the operator induced by 3-dimensional Biot–Savart operator \(\mathrm{BS}_{3d}\) applied to the axisymmetric coordinate \((r',z)=(r+l,z)\), satisfy the following system of equations:

\[\begin{aligned}\Delta_l\phi_l&:=\left(\partial_r^2+\frac1{r+l}\partial_r-\frac1{(r+l)^2}+\partial_z^2\right)\phi_l=\omega\quad\text{in }\mathbb R_{r\geq-l}\times\mathbb R,\\\partial_r\phi_l|_{r=-l}&=0,\\\mathrm{BS}_l[\omega]&:=-\partial_z\phi_l e_r+\left(\partial_r+\frac1{r+l}\right)\phi_l e_z.\end{aligned}\tag{5.9}\]

The domain of definition is:

\[D(L_l):=\{\omega\in L_{\gamma,l}^2\mid\bar u_l\cdot\nabla\omega\in L_{\gamma,l}^2\}.\]

We observe the following decomposition of the operator \(L_l\):

\[L_l=M_l+K_l+S_l,\]

into the main skew-symmetric term, the compact term, and the small term:

\[\begin{aligned}-M_l\omega&:=\bar u_l\cdot\nabla\omega+\tfrac12\omega\nabla\cdot v_l,\\-K_l\omega&:=\mathrm{BS}_l(\omega)\cdot\nabla\bar\omega_l,\\-S_l\omega&:=-\frac{\bar u_l^r}{r+l}\omega-\frac{(\mathrm{BS}_l(\omega))^r}{r+l}\bar\omega_l-\tfrac12\omega\nabla\cdot v_l.\end{aligned}\]

Note that the additional term \(\tfrac12\omega\nabla\cdot v_l\) is to make \(M_l\) a skew-adjoint operator.

Here are the basic properties of the compact and small operators:

Lemma 5.2. \(K_l,S_l:L_{\gamma,l}^2\to L_{\gamma,l}^2\) are two bounded operators. Moreover, \(K_l\) is a compact operator, and

\[\|S_l\|_{L_{\gamma,l}^2\to L_{\gamma,l}^2}\xrightarrow[l\to\infty]{}0.\]

Sketch of Proof. The compactness of \(K_l\) essentially relies on Rellich’s theorem, and the smallness of \(S_l\) is intuitive.

Since we are considering \(l\) large enough, as the small term converges to zero, only need to consider the formal limits of the main and compact term:

\[\begin{aligned}K_\infty:L_\gamma^2&\longrightarrow L_\gamma^2,&\omega&\longmapsto-\mathrm{BS}_{2d}[\omega]\cdot\nabla\bar\omega,\\M_\infty:D(M_\infty)\subset L_{\gamma,l}^2&\longrightarrow L_{\gamma,l}^2,&\omega&\longmapsto-\bar u\cdot\nabla\omega,\end{aligned}\]

where the domain of definition:

\[D(M_\infty)=\{\omega\in L_\gamma^2\mid\bar u\cdot\nabla\omega\in L^2(B_0(\bar R))\}.\]

We observe that the total formal limit \(M_\infty+K_\infty=L_\infty\) has an unstable eigenvalue by the discussion of the previous subsection. Thus, we expect that for \(l\) large enough, the operator \(L_l\) is unstable:

Proposition 5.1. Let \(\lambda_\infty\) be an unstable eigenvalue of \(L_\infty\), then, for all \(\epsilon\in(0,\Re(\lambda_\infty))\), there exist \(l_0=l_0(\bar u,\lambda_\infty,\epsilon)\geq2\bar R\), such that for all \(l\geq l_0\), the operator \(L_l\) has an unstable eigenvalue \(\lambda_l\) with \(|\lambda_l-\lambda_\infty|<\epsilon\).

The original problem can be transformed into the generalized lemma below, and we can justify all the conditions with the setting of proposition above.

Lemma 5.3. Let \((M_n)_{n\in\mathbb N},M_\infty:D(M_\infty)\subset H\to H\) closed normal operators, \((K_n)_{n\in\mathbb N},K_\infty:H\to H\) compact operators, \((S_n)_{n\in\mathbb N}:H\to H\) bounded operators, satisfying the following properties:

  1. There exist \((\mu_n)_{n\in\mathbb N\cup\{\infty\}}\in\mathbb R\), such that \(\lim_{n\to\infty}\mu_n=\mu_\infty\), and: \[\sigma(M_n)\subset\{z\in\mathbb C\mid\Re(z)\leq\mu_n\},\qquad\forall n\in\mathbb N\cup\{\infty\}.\]
  2. For all \(f\in H\), \[(\lambda-M_n)^{-1}f\xrightarrow[n\to\infty]{}(\lambda-M_\infty)^{-1}f\quad\text{in }H,\tag{5.10}\]

    locally uniformly on \(\lambda\in\{z\in\mathbb C\mid\Re(z)>\mu_\infty\}\).

  3. \(K_n\to K_\infty\) in operator norm.
  4. The smallness of \(\{S_n\}_{n\in\mathbb N}\): \[\lim_{n\to\infty}\|S_n\|_{H\to H}=0.\tag{5.11}\]

Let \(\lambda_\infty\) be an isolated eigenvalue of \(M_\infty+K_\infty\), such that \(\Re(\lambda_\infty)>\mu_\infty\), and let \(V\subset\{z\in\mathbb C\mid\Re(z)>\mu_\infty\}\) such that \(V\cap\sigma(M_\infty+K_\infty)=\{\lambda_\infty\}\), then, for \(n\) big enough, there exist an eigenvalue of \(M_n+K_n+S_n\) in \(V\).

Sketch of Proof. We claim that: for all \(C\subset\mathrm{res}(C_\infty)\cap\{z\in\mathbb C\mid\Re(z)>\mu_\infty\}\) compact, \(f\in H\), we have the following convergence for \(\lambda\in C\) uniformly:

\[(\lambda-(M_n+K_n+S_n))^{-1}f\xrightarrow[n\to\infty]{}(\lambda-(M_\infty+K_\infty))^{-1}f\quad\text{in }H.\tag{5.12}\]

The formula above can be proven by resolvent calculus, and the conclusion can be drawn by contour integration.

6. Instability: From Euler to Navier–Stokes

As mentioned in the strategy part, the linearized operator of Navier–Stokes equation formally approaches the linearized operator of Euler equation when the amplitude of the velocity profile is sufficiently large. In this section, we will give a rigorous statement of the idea above.

6.1. Vorticity Instability

For \(l\) sufficiently large, we fix \(\bar U\) to be the 3-dimensional axisymmetrical-no-swirl velocity profile associated with \(\bar u_l\) with respect to coordinate \((r',z)\):

\[\bar U(r',z):=\bar u_l^r(r'-l,z)e_r+\bar u_l^z(r'-l,z)e_z,\]

and \(\bar\Omega\) is its corresponding vorticity field.

Define the set of \(L^2\) integrable axisymmetrical-pure-swirl vector field on \(\mathbb R^3\):

\[L_{\mathrm{aps}}^2=\{\Omega^\theta(r',z)e_\theta\mid\Omega^\theta\in L^2(\mathbb R^3)\}.\]

We consider the base profile with relative amplitude \(\beta\), i.e. the operator is linearized around \(\beta\bar U\), take curl on the operator \(L_{ss}\) defined in (4.4) and restrict it on the set of \(L_{\mathrm{aps}}^2\). We obtain the following operator:

\[\begin{aligned}L_{\mathrm{vor}}^{(\beta)}:D(L_{\mathrm{vor}}^{(\beta)})&\longrightarrow L_{\mathrm{aps}}^2,\\\Omega&\longmapsto\tfrac12(2+\xi\cdot\nabla_\xi)\Omega+\Delta\Omega-\beta([\bar U,\Omega]+[U,\bar\Omega]),\end{aligned}\]

where the domain of definition:

\[D(L_{\mathrm{vor}}^{(\beta)}):=\{\Omega\in L_{\mathrm{aps}}^2\cap H^2(\mathbb R^3)\mid\xi\cdot\nabla_\xi\Omega\in L^2(\mathbb R^3)\},\]

the lifted velocity field \(U:=\mathrm{BS}_{3d}[\Omega]\), and the bracket \([f,g]:=(f\cdot\nabla)g-(g\cdot\nabla)f\).

We decompose the operator into the following terms and a multiplication by constant:

\[\begin{aligned}D\Omega&=\left(\tfrac34+\tfrac\xi2\cdot\nabla_\xi\right)\Omega+\Delta\Omega,&M\Omega&=-\bar U\cdot\nabla\Omega,\\S\Omega&=\bar\Omega\cdot\nabla U+\Omega\cdot\nabla\bar U,&K\Omega&=-U\cdot\nabla\bar\Omega.\end{aligned}\tag{6.1}\]

\(D\) is the diffusion term, \(M\) is the main term, \(S\) is the small term, and \(K\) is the compact term. Observe that \(D-\Delta\) is a skew-adjoint operator.

Finally, we define:

\[T_\beta:=\frac1\beta D+M+S+K=\frac1\beta\left(L_{\mathrm{vor}}^{(\beta)}-\frac14\right),\]

and we define the formal limit when \(\beta\to\infty\):

\[T_\infty:=M+S+K,\]

where the domain of definition is

\[D(T_\infty)=\{\Omega\in L_{\mathrm{aps}}^2\mid\bar U\cdot\nabla\Omega\in L^2\}.\]

We view \(T_\infty\) as a lift of \(L_l\) into 3-dimensions, and the part \(S\) is exactly a part of \(S_l\) whose smallness can be obtained by slightly modifying the proof of Lemma 5.2. Thus, we can fix \(l\) sufficiently large, such that:

\[\alpha:=\sup_{\lambda\in\sigma(T_\infty)}\Re(\lambda)>\mu:=\|S\|_{L_{\mathrm{aps}}^2\to L^2(\mathbb R^3)}.\tag{6.2}\]

From now on, we consider \(l\) and \(\bar U\) as fixed and satisfying the estimate (6.2).

Now we state the instability of self-similar Navier–Stokes equation:

Theorem 6.1. Let \(\lambda_\infty\) be an unstable eigenvalue of \(T_\infty\) such that \(\Re(\lambda_\infty)>\mu\), then for all \(\epsilon\in(0,\Re(\lambda_\infty)-\mu)\), there exists \(\beta_0>0\), such that for all \(\beta>\beta_0\), \(T_\beta\) has an unstable eigenvalue \(\lambda_\beta\) such that \(|\lambda_\beta-\lambda_\infty|<\epsilon\), and \(L_{\mathrm{vor}}^{(\beta)}\) has an unstable eigenvalue defined by

\[\tilde\lambda_\beta:=\beta\lambda_\beta+\frac14.\]

Sketch of Proof. We set ourselves in the setting of Lemma (5.3), let:

\[M_\beta=\beta^{-1}D+M+S,\qquad K_\beta=K,\qquad S_\beta=0.\]

It remains to verify the statements in the Lemma.

6.2. Return to Velocity Form

In the following discussion, it’s more convenient to work on the original linearized operator (4.4) which acts on the velocity fields, we introduce the amplified operator \(L_{ss}^{(\beta)}\):

\[\begin{aligned}L_{ss}^{(\beta)}:D(L_{ss})&\longrightarrow L_\sigma^2,\\U&\longmapsto\tfrac12(1+\xi\cdot\nabla_\xi)U+\Delta U-\beta\mathbb P(\bar U\cdot\nabla U+U\cdot\nabla\bar U),\end{aligned}\]

where we define the divergence-free subspace and the domain of definition:

\[L_\sigma^2=\{u\in L^2(\mathbb R^3)^3\mid\nabla\cdot u=0\},\qquad D(L_{ss})=\{U\in L_\sigma^2\cap H^2(\mathbb R^3)^3\mid\xi\cdot\nabla_\xiU\in L^2(\mathbb R^3)^3\}.\]

The following corollary identifies the unstable eigenvalue of \(L_{\mathrm{vor}}^{(\beta)}\) and \(L_{ss}\).

Corollary 6.1. In the setting of Theorem 6.1, for \(\beta\geq\beta_0\), the unstable eigenvalue \(\tilde\lambda_\beta\) is also an unstable eigenvalue of \(L_{ss}^{(\beta)}\).

Sketch of Proof. By choosing \(\beta_0\) sufficiently large, we can ensure that \(\Re(\tilde\lambda_\beta)>1\). Denoting the corresponding eigenfunction as \(\Omega_\beta\in D(L_{\mathrm{vor}}^{(\beta)})\), then we have:

\[(\tilde\lambda_\beta-1)\Omega_\beta-\frac{\xi\cdot\nabla_\xi\Omega_\beta}{2}-\Delta\Omega_\beta=-\beta([\bar U,\Omega_\beta]+[U_\beta,\bar\Omega]).\tag{6.3}\]

We can see that \(\mathrm{BS}_{3d}[\Omega_\beta]\) is the eigenvalue corresponding to \(\tilde\lambda_\beta\) of \(L_{ss}^{(\beta)}\) as long as \(\mathrm{BS}_{3d}[\Omega_\beta]\in D(L_{ss}^{(\beta)})\).

7. Nonlinear Instability: Conclusion

In the previous paragraphs, we have studied the instability property of the linearized operator \(L_{ss}\) defined in (4.4) with a given background real, compactly supported velocity profile \(\beta\bar U\) with \(\beta\) sufficiently large. Now, we want to apply this property into our self-similar equation (4.3), and obtain the asymptotic behavior when \(\tau\to-\infty\).

In the following sections, we will use the notion introduced in (4.1) and (4.2) to separate the physical space and self-similar space and introduce the semigroup generated by the operator \(L_{ss}\) and finally construct \(U^{\mathrm{per}}\) via fix point argument.

7.1. Study of the Linearized Operator

In this subsection, we want to consider the time evolution under the linearized operator. The spectrum of \(L_{ss}\) is bounded in the right and the norm of its resolvent is of order \(O(\Re(z)^{-1})\). Applying Hille–Yosida’s theorem, the operator generates a semigroup \(e^{\tau L_{ss}}:L_\sigma^2\to L_\sigma^2\). For a detailed proof, please refer to [JS13, Chapter 2].

We define the right spectral bound of \(L_{ss}\) as:

\[s(L_{ss}):=\sup\{\Re(\lambda)\mid\lambda\in\sigma(L_{ss})\}>0.\]

In the light of [EN01, Proposition 2.2], [Kat66, Corollary 2.11] and [JS13, Lemma 2.7], we conclude the spectral property by the proposition below:

Proposition 7.1. We have \(s(L_{ss})<\infty\), and there exists an eigenvalue \(\lambda=s(L_{ss})+ib\in\sigma(L_{ss})\) and its corresponding eigenfunction \(\eta=\eta_1+i\eta_2\), \(\eta_i\in D(L_{ss})\) such that \(L_{ss}\eta=\lambda\eta\).

Moreover, for \(\delta>0\), there exists \(M(\delta)>0\), such that:

\[\|e^{\tau L_{ss}}\|_{L_\sigma^2\to L_\sigma^2}\leq M(\delta)e^{(s(L_{ss})+\delta)\tau}.\]

We take the linear perturbation to be

\[U^{\mathrm{lin}}(\tau)=\Re(e^{\lambda\tau}\eta),\]

then it satisfies the linearized equation:

\[\partial_\tau U^{\mathrm{lin}}=L_{ss}U^{\mathrm{lin}}.\]

As there is a diffusion term in \(L_{ss}\), we expect parabolic regularity:

Proposition 7.2. For any integer \(\sigma_2\geq\sigma_1\geq0\) and \(\delta>0\), then there exists \(M=M(\sigma_1,\sigma_2,\delta)\), for all \(U_0\in L_\sigma^2\cap H^{\sigma_1}(\mathbb R^3)^3\), we have:

\[\|e^{\tau L_{ss}}U_0\|_{H^{\sigma_2}}\leq\frac{M}{\tau^{(\sigma_2-\sigma_1)/2}}e^{(s(L_{ss})+\delta)\tau}\|U_0\|_{H^{\sigma_1}}.\]

Sketch of Proof. Observe that by a change of variables back into the physical space,

\[u(t,x):=\frac1{\sqrt{t+1}}U\left(\log(t+1),\frac{x}{\sqrt{t+1}}\right),\qquad\bar u(t,x):=\frac1{\sqrt{t+1}}\bar U\left(\frac{x}{\sqrt{t+1}}\right),\]

with \(u(0,x)=U(0,x)=:U_0(x)\), then, we have the following equation:

\[\begin{cases}\partial_tu-\Delta u=-\mathbb P(\bar u\cdot\nabla u+u\cdot\nabla\bar u),&(t,x)\in[0,\infty)\times\mathbb R^3,\\u(0,x)=U_0(x).\end{cases}\]

The proof essentially relies on the parabolic regularity for the heat semigroup:

\[t^{s/2}\|e^{t\Delta}f\|_{\dot H^s}\lesssim\|f\|_{L^2},\qquad\forall s\in\mathbb N.\]

As a corollary, we have the eigenfunction \(\eta\in H^k(\mathbb R^3)^3\) for all \(k\in\mathbb N\), and,

\[\|U^{\mathrm{lin}}(\tau)\|_{H^k}=\|\Re(e^{\tau L_{ss}}\eta)\|_{H^k}=e^{s(L_{ss})\tau}\|\Re(e^{ib\tau}\eta)\|_{H^k}\sim e^{s(L_{ss})\tau},\tag{7.1}\]

where \(\|\Re(e^{ib\tau}\eta)\|_{H^k}\) is a periodic function.

7.2. Construction of the Nonlinear Profile

In this subsection, we study the equation satisfied by the perturbation term in the self-similar variables:

\[\partial_\tau U^{\mathrm{per}}+L_{ss}U^{\mathrm{per}}=-\mathbb P[((U^{\mathrm{per}}+U^{\mathrm{lin}})\cdot\nabla)(U^{\mathrm{per}}+U^{\mathrm{lin}})],\tag{7.2}\]

with formal initial value \(U^{\mathrm{per}}(-\infty)=0\).

We state the existence and decay of \(U^{\mathrm{per}}\) in the proposition below:

Proposition 7.3. Assume that \(N>\tfrac52\) is an integer, then, there exists lifespan \(T=T(\bar U,U^{\mathrm{lin}},N)\in\mathbb R\), \(\epsilon_0=\epsilon_0(s(L_{ss}))>0\), such that (7.2) admits a solution \(U^{\mathrm{per}}\in C((-\infty,T],H^N(\mathbb R^3)^3)\), and we have the following estimate:

\[\|U^{\mathrm{per}}(\tau)\|_{H^N}\leq e^{(s(L_{ss})+\epsilon_0)\tau},\qquad\forall\tau\leq T.\]

Let \(T\in\mathbb R\) and \(\epsilon_0>0\) to be two undetermined constants. We consider the Banach space equipped with norm decaying in time and the Duhamel functional:

\[X:=\left\{U\in C((-\infty,T],H^N(\mathbb R^3)^3)\ \middle|\ \|U\|_X:=\sup_{\tau\leq T}e^{-(s(L_{ss})+\epsilon_0)\tau}\|U\|_{H^N}<+\infty\right\},\] \[\mathcal TU(\tau):=-\int_{-\infty}^\tau e^{(\tau-s)L_{ss}}\mathbb P[((U+U^{\mathrm{lin}})\cdot\nabla)(U+U^{\mathrm{lin}})(s)]\,ds.\]

Proposition 7.4. There exists \(T=T(\bar U,U^{\mathrm{lin}},N)\in\mathbb R\) and \(\epsilon_0=\epsilon_0(s(L_{ss}))>0\), such that:

\[\mathcal T:\{U\in X\mid\|U\|_X\leq1\}\longrightarrow\{U\in X\mid\|U\|_X\leq1\}\]

is a contraction.

Sketch of Proof. The proposition essentially comes from the fact that for \(M>\tfrac32\), \(H^M(\mathbb R^3)\) is a Banach algebra, i.e. for \(f,g\in H^N(\mathbb R^3)\), we have:

\[\|fg\|_{H^M}\lesssim_M\|f\|_{H^M}\|g\|_{H^M}.\]

Simply choose \(\epsilon_0=\tfrac12s(L_{ss})\), \(T\) sufficient small will satisfy the proposition.

Indeed, by a bootstrap argument, we have:

\[\sup_{\tau\in(-\infty,T]}e^{-2s(L_{ss})\tau}\|\mathcal TU(\tau)\|_{H^N}<\infty.\]

Since \(U^{\mathrm{per}}\) is a fixed point of \(\mathcal T\), it satisfies the same estimate, thus the decay rate of \(\|U^{\mathrm{per}}\|_{H^N}\) is actually \(O(e^{2s(L_{ss})\tau})\).

Combining all the conclusions we have obtained, we have constructed the real, smooth, compactly supported unstable profile \(\bar U\); the linear term \(U^{\mathrm{lin}}\) associated to the most unstable eigenvalue, with decay of order \(O(e^{s(L_{ss})\tau})\); and the perturbation term \(U^{\mathrm{per}}\) with decay of order \(O(e^{2s(L_{ss})\tau})\). The proof of Theorem 4.1 is done. Finally, rewrite the equation into the physical variables, with initial value \(u|_{t=0}=0\), Theorem 1.1 is proven.

References

  1. [ABC22] Dallas Albritton, Elia Bruè, and Maria Colombo. Non-uniqueness of Leray solutions of the forced Navier–Stokes equations. Annals of Mathematics, 196, 07 2022. doi:10.4007/annals.2022.196.1.3.
  2. [BM20] Alberto Bressan and Ryan Murray. On self-similar solutions to the incompressible Euler equations. Journal of Differential Equations, 269(6):5142–5203, 2020.
  3. [BS21] Alberto Bressan and Wen Shen. A posteriori error estimates for self-similar solutions to the Euler equations. Discrete and Continuous Dynamical Systems, 41(1):113–130, 2021.
  4. [BV19] Tristan Buckmaster and Vlad Vicol. Nonuniqueness of weak solutions to the Navier–Stokes equation. Annals of Mathematics, 189(1):101–144, 2019.
  5. [EN01] Klaus-Jochen Engel and Rainer Nagel. One-parameter semigroups for linear evolution equations. Semigroup Forum, 63:278–280, 06 2001.
  6. [JS13] Hao Jia and Vladimir Sverak. Are the incompressible 3D Navier–Stokes equations locally ill-posed in the natural energy space? arXiv: Analysis of PDEs, 2013.
  7. [Kat66] Tosio Kato. Perturbation Theory for Linear Operators. Springer, 1966.
  8. [Ler34] Jean Leray. On the motion of a viscous liquid filling space. [English translation of the title.] Acta Mathematica, 63(none):193–248, 1934.
  9. [LP59] Jacques-Louis Lions and Giovanni Prodi. An existence and uniqueness theorem for the Navier–Stokes equations in dimension two. [English translation of the title.] C. R. Acad. Sci., Paris, 248:3519–3521, 1959.
  10. [Vis18a] Misha Vishik. Instability and non-uniqueness in the Cauchy problem for the Euler equations of an ideal incompressible fluid. Part I. arXiv e-prints, arXiv:1805.09426, May 2018.
  11. [Vis18b] Misha Vishik. Instability and non-uniqueness in the Cauchy problem for the Euler equations of an ideal incompressible fluid. Part II. arXiv e-prints, arXiv:1805.09440, May 2018.
Transcription notes

The seven sections, proofs and eleven references of the manuscript are reproduced above. Formulas are typeset in LaTeX on the page. Paragraph breaks and punctuation have been adapted for screen reading; the mathematical statements remain those of the text dated 2025.

Several typos in that version have been retained to distinguish the source from the implementation: (4.4) has a positive sign in front of convection, whereas the linearisation of (4.3), reproduced in Section 6.2, has a negative sign; in (4.5), the \(L^2(\mathbb R^3)\) norm is multiplied by \(e^{\tau/4}\), and its square by \(e^{\tau/2}\); in (7.2), the displayed Duhamel formula corresponds to \(\partial_\tau U^{\mathrm{per}}-L_{ss}U^{\mathrm{per}}\). The term \(\partial_z^2\phi^\theta\) is missing from the second line of (4.7), but appears in (5.9). The matrix contraction written as \(A^{(1)}:A^{(1)}\) means \(A^{(1)}:A^{(2)}\). Identity (1.2) and the divergence convention must be aligned: with the convention explicitly given, \(\operatorname{div}(u\otimes v)=(v\cdot\nabla)u+u\operatorname{div}v\). The reference to “Proposition 1.2” in Section 2.3 denotes the compactness result in Proposition 2.1.

The experiment follows the signs and physical normalisation of the numerical report: it retains every term of the difference equation, including \(\partial_z^2\), and uses \(\|u(t)\|_2=t^{1/4}\|U(\log t)\|_2\).

Back to top