On Well-posed Boundary Conditions for the Linear Non-homogeneous Moment Equations in Half-space

Around the equilibrium states, the rarefied gas can be described by the linearized Boltzmann equation (LBE). For single-species monatomic molecules, we consider the LBE with the Maxwell boundary condition

	$$\displaystyle\dfrac{\partial{f}}{\partial{t}}+\xi_{d}\dfrac{\partial{f}}{\partial{x_{d}}}=\mathcal{L}[f],\quad f=f(t,\boldsymbol{x},\boldsymbol{\xi}),\ \boldsymbol{x}\in\Omega\subset\mathbb{R}^{3},\ \boldsymbol{\xi}\in\mathbb{R}^{3},$$		(2.1a)
	$$\displaystyle f(t,\boldsymbol{x},\boldsymbol{\xi})=\chi f^{w}(t,\boldsymbol{x},\boldsymbol{\xi})+(1-\chi)f(t,\boldsymbol{x},\boldsymbol{\xi}^{*}),\quad\boldsymbol{x}\in\partial\Omega,\ (\boldsymbol{\xi}-\boldsymbol{u}^{w})\cdot\boldsymbol{n}<0,$$		(2.1b)

where $f$ is the velocity distribution function, $t$ denoting the time, $\boldsymbol{x}=(x_{1},x_{2},x_{3})$ the spatial coordinates and $\boldsymbol{\xi}=(\xi_{1},\xi_{2},\xi_{3})$ the microscopic velocity. Here $\mathcal{L}$ is a linear operator depicting the collision between gas molecules.

The Maxwell model assumes that the boundary is an impermeable wall with a unit normal vector $\boldsymbol{n}$ exiting the region, a given velocity $\boldsymbol{u}^{w}$, and a given temperature $\theta^{w}$. To avoid cumbersome details about the rotation invariance, we assume

$$\Omega=\{\boldsymbol{x}\in\mathbb{R}^{3}:\ x_{2}\geq 0\},$$

with a fixed $\boldsymbol{n}=(0,-1,0)$. The reflection at the wall is divided into a sum of $\chi$ portion of the specular reflection and $1-\chi$ portion of the diffuse reflection, where $\chi\in[0,1]$ is the tangential momentum accommodation coefficient. To consider the steady layer equations, we further assume

$$\boldsymbol{u}^{w}\cdot\boldsymbol{n}=0,$$

then the velocity from specular reflection is

$$\boldsymbol{\xi}^{*}=\boldsymbol{\xi}-2[(\boldsymbol{\xi}-\boldsymbol{u}^{w})\cdot\boldsymbol{n}]\boldsymbol{n}=(\xi_{1},-\xi_{2},\xi_{3}),$$

and

$$f^{w}(t,\boldsymbol{x},\boldsymbol{\xi})=\mathcal{M}(\boldsymbol{\xi})\left(\rho^{w}+\boldsymbol{u}^{w}\cdot\boldsymbol{\xi}+\theta^{w}\frac{|\boldsymbol{\xi}|^{2}-3}{2}\right).$$

(2.2)

The reference distribution $\mathcal{M}$ is a Maxwellian at rest

$$\mathcal{M}(\boldsymbol{\xi})=\frac{1}{(2\pi)^{3/2}}\exp\left(-\frac{|\boldsymbol{\xi}|^{2}}{2}\right),$$

where $\rho^{w}$ is determined by the no mass flow condition at the wall

$$\int_{\mathbb{R}^{3}}\!\!(\boldsymbol{\xi}-\boldsymbol{u}^{w})\cdot\boldsymbol{n}f(t,\boldsymbol{x},\boldsymbol{\xi})\,\mathrm{d}\boldsymbol{\xi}=0,\quad\boldsymbol{x}\in\partial\Omega.$$

The Grad moment equations with linearized ansatz assume that the distribution function has a Hermite expansion, for a chosen integer $M\geq 2$, as

$$f(t,\boldsymbol{x},\boldsymbol{\xi})=\mathcal{M}(\boldsymbol{\xi})\sum_{|\boldsymbol{\alpha}|\leq M}w_{\boldsymbol{\alpha}}(t,\boldsymbol{x})\phi_{\boldsymbol{\alpha}}(\boldsymbol{\xi}),$$

(2.3)

where $\boldsymbol{\alpha}=(\alpha_{1},\alpha_{2},\alpha_{3})\in\mathbb{N}^{3}$ is the multi-index, $|\boldsymbol{\alpha}|=\alpha_{1}+\alpha_{2}+\alpha_{3}$. The orthonormal Hermite polynomial $\phi_{\boldsymbol{\alpha}}=\phi_{\boldsymbol{\alpha}}(\boldsymbol{\xi})$ is defined [13] by ensuring

$$\left\langle{\mathcal{M}\phi_{\boldsymbol{\alpha}}\phi_{\boldsymbol{\beta}}}\right\rangle=\delta_{\boldsymbol{\alpha},\boldsymbol{\beta}},\quad\phi_{\boldsymbol{0}}=1,\ \phi_{\boldsymbol{e}_{i}}=\xi_{i},\quad\left\langle{\cdot}\right\rangle:=\int_{\mathbb{R}^{3}}\!\!\cdot\,\mathrm{d}\boldsymbol{\xi},$$

where $\boldsymbol{e}_{i}\in\mathbb{N}^{3}$ only has the $i$-th component being one. So the ansatz (2.3) gives

$$w_{\boldsymbol{\alpha}}=w_{\boldsymbol{\alpha}}(t,\boldsymbol{x})=\left\langle{f\phi_{\boldsymbol{\alpha}}}\right\rangle.$$

Substituting the ansatz into the linearized Boltzmann equation and matching the coefficients before basis functions, we have the $M$-th order moment equations

$$\dfrac{\partial{w_{\boldsymbol{\alpha}}}}{\partial{t}}+\left\langle{\mathcal{M}\xi_{d}\phi_{\boldsymbol{\alpha}}\phi_{\boldsymbol{\beta}}}\right\rangle\dfrac{\partial{w_{\boldsymbol{\beta}}}}{\partial{x_{d}}}=\left\langle{\mathcal{L}[\mathcal{M}\phi_{\boldsymbol{\beta}}]\phi_{\boldsymbol{\alpha}}}\right\rangle w_{\boldsymbol{\beta}},\quad|\boldsymbol{\alpha}|\leq M,$$

(2.4)

where Einstein’s convention is used to omit the summation notation about $|\boldsymbol{\beta}|\leq M.$

Intuitively, the moment variables $w_{\boldsymbol{\alpha}}$ can be related to the macroscopic variables such as the density, the macroscopic velocity, and the temperature. So (2.4) would contain the linearized Euler equations (around the equilibrium state given by $\mathcal{M}$). We can regard (2.4) as an extension of hydrodynamic equations [14].

It’s not difficult to check [22] that (2.4) is symmetric hyperbolic. More precisely, the coefficient matrices are all symmetric. To ensure the correct number of boundary conditions for the hyperbolic system [17], Grad [14] suggested testing the Maxwell boundary condition (2.1b) with odd polynomials (about the argument $(\boldsymbol{\xi}-\boldsymbol{u}^{w})\cdot\boldsymbol{n}=-\xi_{2}$) no higher than the $M$-th degree.

The Grad boundary condition has some equivalent representations [14, 30, 8, 12]. To exhibit the continuity of fluxes at the boundary, we extract the factor $\xi_{2}$ of these odd polynomials and write the Grad boundary condition in an equivalent form:

	$\displaystyle\int_{\mathbb{R}^{2}}\!\!\int_{0}^{+\infty}\!\!\xi_{2}\phi_{\boldsymbol{\alpha}}f(t,\boldsymbol{x},\boldsymbol{\xi})\,\mathrm{d}\boldsymbol{\xi}$	$\displaystyle=$	$\displaystyle\chi\int_{\mathbb{R}^{2}}\!\!\int_{0}^{+\infty}\!\!\xi_{2}\phi_{\boldsymbol{\alpha}}f^{w}(t,\boldsymbol{x},\boldsymbol{\xi})\,\mathrm{d}\boldsymbol{\xi}+$
			$\displaystyle(1-\chi)\int_{\mathbb{R}^{2}}\!\!\int_{0}^{+\infty}\!\!\xi_{2}\phi_{\boldsymbol{\alpha}}f(t,\boldsymbol{x},\boldsymbol{\xi}^{*})\,\mathrm{d}\boldsymbol{\xi},$

where $\alpha_{2}$ is even and $|\boldsymbol{\alpha}|\leq M-1.$ Plugging the ansatz (2.3) into the above formula, we can utilize the even-odd parity of Hermite polynomials [23] to get

$$\left(1-\frac{\chi}{2}\right)\sum_{\beta_{2}\text{\ odd}}\left\langle{\xi_{2}\mathcal{M}\phi_{\boldsymbol{\alpha}}\phi_{\boldsymbol{\beta}}}\right\rangle w_{\boldsymbol{\beta}}=-\frac{\chi}{2}\sum_{\beta_{2}\text{\ even}}\left\langle{|\xi_{2}|\mathcal{M}\phi_{\boldsymbol{\alpha}}\phi_{\boldsymbol{\beta}}}\right\rangle(w_{\boldsymbol{\beta}}-b_{\boldsymbol{\beta}}),$$

(2.5)

where the entries $b_{\boldsymbol{0}}=\rho^{w},\ b_{\boldsymbol{e}_{i}}=u_{i}^{w},\ b_{2\boldsymbol{e}_{i}}=\theta^{w}/\sqrt{2}$ and otherwise $b_{\boldsymbol{\alpha}}=0.$

In conclusion, the $M$-th order linear Grad moment equations (in the plane geometry) are (2.4) with the Grad boundary condition (2.5), where $\rho^{w}$ is determined by the no mass flow condition

$$w_{\boldsymbol{e}_{2}}-u_{2}^{w}=0,\quad\text{at}\ x_{2}=0.$$

Assume there is a boundary layer near $\partial\Omega=\{x_{2}=0\}.$ Then we may introduce the fast variables, which vary dramatically in the normal direction of the wall and vanish outside the boundary layer. According to the asymptotic analysis [29] of (2.4), these fast variables may satisfy the linear steady moment equations in half-space:

$$\left\langle{\mathcal{M}\xi_{2}\phi_{\boldsymbol{\alpha}}\phi_{\boldsymbol{\beta}}}\right\rangle\dfrac{\mathrm{d}{w_{\boldsymbol{\beta}}}}{\mathrm{d}{y}}=\left\langle{\mathcal{L}[\mathcal{M}\phi_{\boldsymbol{\beta}}]\phi_{\boldsymbol{\alpha}}}\right\rangle w_{\boldsymbol{\beta}}+h_{\boldsymbol{\alpha}},\quad|\boldsymbol{\alpha}|\leq M,$$

where $w_{\boldsymbol{\beta}}=w_{\boldsymbol{\beta}}(y),\ y\in[0,+\infty),$ represents the fast variables, with $w_{\boldsymbol{\beta}}(+\infty)=0$. The non-homogeneous term $h_{\boldsymbol{\alpha}}=h_{\boldsymbol{\alpha}}(y)$ is given, arising from the other contributions in the boundary layer.

The Grad boundary condition becomes

$$\left(1-\frac{\chi}{2}\right)\sum_{\beta_{2}\text{\ odd}}\left\langle{\xi_{2}\mathcal{M}\phi_{\boldsymbol{\alpha}}\phi_{\boldsymbol{\beta}}}\right\rangle(w_{\boldsymbol{\beta}}+\bar{w}_{\boldsymbol{\beta}})=-\frac{\chi}{2}\sum_{\beta_{2}\text{\ even}}\left\langle{|\xi_{2}|\mathcal{M}\phi_{\boldsymbol{\alpha}}\phi_{\boldsymbol{\beta}}}\right\rangle(w_{\boldsymbol{\beta}}+\bar{w}_{\boldsymbol{\beta}}-b_{\boldsymbol{\beta}}),$$

where $\alpha_{2}$ is even with $|\boldsymbol{\alpha}|\leq M-1,$ and $\bar{w}_{\boldsymbol{\beta}}$ is given by the bulk flow outside the boundary layer.

For the classical linearized Boltzmann operator [10], we shall write the Grad moment equations in half-space as an abstract form

	$$\displaystyle\boldsymbol{A}\dfrac{\mathrm{d}{W}}{\mathrm{d}{y}}=-\boldsymbol{Q}W+h,\quad W=W(y),\ y\in[0,+\infty),$$		(2.6)
	$$\displaystyle W(+\infty)=0,$$

where $\boldsymbol{A}=\boldsymbol{A}^{T}\in\mathbb{R}^{N\times N},\ \boldsymbol{Q}\geq 0$ and $W\in\mathbb{R}^{N}.$ Here $N=\#\{\boldsymbol{\alpha}\in\mathbb{N}^{3}:\ |\boldsymbol{\alpha}|\leq M\}$ and we assume

$$\mathrm{Null}(\boldsymbol{A})\cap\mathrm{Null}(\boldsymbol{Q})=\{0\}.$$

(2.7)

In this paper, we will discuss the well-posed conditions of (2.6) with $h\neq 0,$ i.e., which boundary condition should be prescribed at $y=0$ to ensure the well-posedness of (2.6). Roughly, we only need to write general solutions for the ODEs and analyze their stability.

If we can solve the generalized eigenvalue problem of $(\boldsymbol{A},\boldsymbol{Q})$, we may deal with the characteristic equations

$$\lambda_{i}\dfrac{\mathrm{d}{v_{i}}}{\mathrm{d}{y}}=-v_{i}+h_{i},\quad v_{i}(+\infty)=0,$$

where $\lambda_{i}\in\mathbb{R}\cup\{\infty\}$ and $h_{i}=h_{i}(y)$ is a given 1D function.

If $\lambda_{i}=0,$ we have $v_{i}=h_{i}$ and $v_{i}(0)=h_{i}(0)$, which means that there is no need to prescribe extra boundary conditions at $y=0.$ If $\lambda<0$, we have

$$v_{i}(y)=e^{-\lambda_{i}^{-1}y}v_{i}(0)+\lambda_{i}^{-1}\int_{0}^{y}\!\!e^{-\lambda_{i}^{-1}(y-s)}h_{i}(s)\,\mathrm{d}s.$$

So to make sure $v_{i}(+\infty)=0,$ there must be

$$v_{i}(0)=-\lambda_{i}^{-1}\int_{0}^{+\infty}\!\!e^{\lambda_{i}^{-1}s}h_{i}(s)\,\mathrm{d}s,$$

if the integral exists. If $\lambda_{i}>0,$ from the above formula we can see that $v_{i}(0)$ can be arbitrarily given when

$$\lim_{y\rightarrow+\infty}\int_{0}^{y}\!\!e^{-\lambda_{i}^{-1}(y-s)}h_{i}(s)\,\mathrm{d}s=0.$$

Finally, if $\lambda_{i}=\infty$, formally $v_{i}$ should be a constant and must be zero.

Our results are based on these simple observations. Due to the structure of coefficient matrices, the role of the generalized eigenvalue decomposition can be replaced by an interpretable simultaneous transformation of the matrices $\boldsymbol{A}$ and $\boldsymbol{Q}$. We note that a similar transformation appears in [3, 5] in the language of projection operators.

For this purpose, we assume that $\boldsymbol{G}\in\mathbb{R}^{N\times p}$ is the orthonormal basis matrix of $\mathrm{Null}(\boldsymbol{Q}),$ and $\boldsymbol{X}\in\mathbb{R}^{p\times r}$ is the orthonormal basis matrix of $\mathrm{Null}(\boldsymbol{G}^{T}\boldsymbol{A}\boldsymbol{G}),$ where

$$p=\mathrm{dim}\ \mathrm{Null}(\boldsymbol{Q}),\quad r=\mathrm{dim}\ \mathrm{Null}(\boldsymbol{G}^{T}\boldsymbol{A}\boldsymbol{G}).$$

Since $\mathrm{Null}(\boldsymbol{A})\cap\mathrm{Null}(\boldsymbol{Q})=\{0\}$, we have

$$\mathrm{rank}(\boldsymbol{A}\boldsymbol{G})=\mathrm{rank}(\boldsymbol{G})=p.$$

So we can let $\boldsymbol{V}_{1}=\boldsymbol{G}\boldsymbol{X}\in\mathbb{R}^{N\times r}$ and construct $\boldsymbol{V}_{2}\in\mathbb{R}^{N\times p}$ by a Gram-Schmidt orthogonalization of $\boldsymbol{A}\boldsymbol{G}$. Then

$$\mathrm{span}\{\boldsymbol{V}_{2}\}=\mathrm{span}\{\boldsymbol{A}\boldsymbol{G}\}$$

and $\boldsymbol{V}_{2}^{T}\boldsymbol{V}_{1}=\boldsymbol{0}.$ Let $\boldsymbol{V}_{3}$ be the orthogonal complement of $[\boldsymbol{V}_{1},\boldsymbol{V}_{2}]$. So $\boldsymbol{V}=[\boldsymbol{V}_{1},\boldsymbol{V}_{2},\boldsymbol{V}_{3}]$ is orthogonal.

For technical reasons, we further define $\boldsymbol{U}=[\boldsymbol{U}_{1},\boldsymbol{U}_{2},\boldsymbol{U}_{3}]$, where

$$\boldsymbol{U}_{1}=\boldsymbol{G}\in\mathbb{R}^{N\times p},\ \boldsymbol{U}_{2}=\boldsymbol{A}\boldsymbol{G}\boldsymbol{X}\in\mathbb{R}^{N\times r},\ \boldsymbol{U}_{3}=\boldsymbol{V}_{3}.$$

We claim that $\boldsymbol{U}$ is invertible. Since $\boldsymbol{V}_{3}^{T}\boldsymbol{V}_{2}=\boldsymbol{0},$ we have $\boldsymbol{V}_{3}^{T}\boldsymbol{U}_{2}=\boldsymbol{0}$. To make $\boldsymbol{U}$ invertible, it’s enough to show that

$$\mathrm{span}\{\boldsymbol{V}_{3}\}\cap\mathrm{span}\{\boldsymbol{G}\}=\{0\}.$$

(2.8)

In fact, if $\boldsymbol{G}c_{1}=\boldsymbol{V}_{3}c_{2}$ for some $c_{1}\in\mathbb{R}^{p}$ and $c_{2}\in\mathbb{R}^{(N-p-r)}$, then $(\boldsymbol{G}c_{1})^{T}\boldsymbol{A}\boldsymbol{G}=0$ since $\boldsymbol{V}_{3}^{T}\boldsymbol{V}_{2}=\boldsymbol{0}$. So $c_{1}\in\mathrm{span}\{\boldsymbol{X}\}$ and $\boldsymbol{V}_{3}c_{2}\in\mathrm{span}\{\boldsymbol{G}\boldsymbol{X}\}=\mathrm{span}\{\boldsymbol{V}_{1}\}$. Thus, from $\boldsymbol{V}_{3}^{T}\boldsymbol{V}_{1}=\boldsymbol{0}$ we have $c_{2}=0$, which implies (2.8).

Thanks to the above transformation, we may consider the generalized eigenvalue problem of $(\boldsymbol{A}_{33},\boldsymbol{Q}_{33})$ rather than $(\boldsymbol{A},\boldsymbol{Q}),$ where $\boldsymbol{Q}_{33}$ is symmetric positive definite while $\boldsymbol{Q}$ is symmetric positive semi-definite. Here $\boldsymbol{A}_{33}$ and $\boldsymbol{A}$ are both symmetric matrices which may have zero eigenvalues. The choices of $\boldsymbol{V}_{2}$ and $\boldsymbol{V}_{3}$ are not unique, but they do not affect the results of Lemma 2.1.

The general solutions of ODEs should be written with the aid of the eigenvalue decomposition. According to Sylvester’s law of inertia, $\boldsymbol{Q}_{33}^{-1}\boldsymbol{A}_{33}$ should have the same number of positive, negative and zero eigenvalues as $\boldsymbol{A}_{33}$. Assume the Cholesky decomposition

$$\boldsymbol{Q}_{33}=\boldsymbol{L}\boldsymbol{L}^{T},$$

then we must have an orthogonal eigenvalue decomposition of the real symmetric matrix:

$$\boldsymbol{L}^{-1}\boldsymbol{A}_{33}\boldsymbol{L}^{-T}\boldsymbol{R}=\boldsymbol{R}\boldsymbol{\Lambda}.$$

(2.9)

We assume the diagonal matrix

$$\boldsymbol{\Lambda}=\begin{bmatrix}\boldsymbol{\Lambda}_{+}&&\\ &\boldsymbol{0}&\\ &&\boldsymbol{\Lambda}_{-}\end{bmatrix},$$

where $\boldsymbol{\Lambda}_{+}\in\mathbb{R}^{n_{+}\times n_{+}}$ has postive entries and $\boldsymbol{\Lambda}_{-}\in\mathbb{R}^{n_{-}\times n_{-}}$ has negative entries. The matrix $\boldsymbol{R}=[\boldsymbol{R}_{+},\boldsymbol{R}_{0},\boldsymbol{R}_{-}]$ is assumed orthogonal, where the columns of $\boldsymbol{R}_{+},\boldsymbol{R}_{0},\boldsymbol{R}_{-}$ are individually the number of postive, zero and negative eigenvalues of $\boldsymbol{A}_{33}$, i.e., $n_{+},n_{0},n_{-}$. Suppose

$$\boldsymbol{T}=\boldsymbol{L}^{-T}\boldsymbol{R}=[\boldsymbol{T}_{+},\boldsymbol{T}_{0},\boldsymbol{T}_{-}],$$

where the columns of $\boldsymbol{T}_{+},\boldsymbol{T}_{0},\boldsymbol{T}_{-}$ are individually $n_{+},n_{0},n_{-}$. Then $\boldsymbol{T}$ is invertible and we have

$$\boldsymbol{Q}_{33}^{-1}\boldsymbol{A}_{33}[\boldsymbol{T}_{+},\boldsymbol{T}_{0},\boldsymbol{T}_{-}]=[\boldsymbol{T}_{+},\boldsymbol{T}_{0},\boldsymbol{T}_{-}]\begin{bmatrix}\boldsymbol{\Lambda}_{+}&&\\ &\boldsymbol{0}&\\ &&\boldsymbol{\Lambda}_{-}\end{bmatrix}.$$

(2.10)

For a positive number $a$, we define the norm in $L^{2}(\mathbb{R}_{+},L^{2}(\mathbb{R}^{N}))$ as

$$\|h\|_{a}=\left(\int_{0}^{+\infty}\!\!e^{2ay}h^{T}h(y)\,\mathrm{d}y\right)^{1/2},$$

and the vector norm as

$$\|g\|=\sqrt{g^{T}g}.$$

Then we have the following well-posed theorem, whose complete proof is put in Section 4.

For initial-boundary value problems of the linear hyperbolic system, Majda and Osher [26] emphasize that the linear space determined by the boundary condition should contain the null space of the boundary matrix. Otherwise, the boundary condition may be unstable. Based on the similar observation, [28] points out that the linear Grad boundary condition is unstable for the initial-boundary value problem.

For the half-space problem, we will use a simple example to illustrate the instability of the Grad boundary condition. From the example, we can see that the instability not only comes from the half-space problem.

If we consider Kramers’ problem with the BGK collision term [12], the simplest moment system when $M=3$ can read as

	$$\displaystyle\dfrac{\mathrm{d}{\sigma_{12}}}{\mathrm{d}{y}}=0,$$
	$$\displaystyle\dfrac{\mathrm{d}{u_{1}}}{\mathrm{d}{y}}+\sqrt{2}\dfrac{\mathrm{d}{f_{3}}}{\mathrm{d}{y}}=-\nu\sigma_{12},$$		(3.1)
	$$\displaystyle\sqrt{2}\dfrac{\mathrm{d}{\sigma_{12}}}{\mathrm{d}{y}}=-\nu f_{3}+h_{3},$$

where $\nu>0$ is a constant and we write the moment variables as $\sigma_{12},u_{1},f_{3}$. The term $h_{3}=h_{3}(y)$ is the given non-homogeneous term. The Grad boundary condition reads as

$$\bar{\sigma}+\sigma_{12}+\hat{\chi}\left(u_{1}+\bar{u}+\frac{\sqrt{2}}{2}f_{3}\right)=0,$$

(3.2)

where $\hat{\chi}=\displaystyle\frac{2\chi}{2-\chi}\frac{1}{\sqrt{2\pi}}$ and $\bar{\sigma},\ \bar{u}$ are given by the far field. Since the moment variables vanish when $y=+\infty$, we can solve from (3.1) that

$$\sigma_{12}=0,\ u_{1}=-\sqrt{2}f_{3},\ f_{3}=h_{3}/\nu.$$

(3.3)

So the Grad boundary condition gives

$$\bar{u}=-\frac{1}{\hat{\chi}}\bar{\sigma}+\frac{\sqrt{2}}{2}\frac{h_{3}(0)}{\nu}.$$

Only when $\bar{u}$ and $\bar{\sigma}$ satisfy the above relation can the half-space problem has a unique solution. However, the term $h_{3}(0)$ can not be controlled by the weighted $L^{2}$ norm $\|h_{3}\|_{a}$. In this sense, we claim that $\bar{u}$ is unstable because $\|\bar{u}\|$ can not be controlled by $\|h_{3}\|_{a}$.

Using the notations in Theorem 2.1, in this example, we have

$$\boldsymbol{A}=\begin{bmatrix}0&0&1\\ 0&0&\sqrt{2}\\ 1&\sqrt{2}&0\end{bmatrix},\ \boldsymbol{Q}=\begin{bmatrix}0&&\\ &\nu&\\ &&\nu\end{bmatrix},\ h=\begin{bmatrix}0\\ h_{3}\\ 0\end{bmatrix},\ W=\begin{bmatrix}u_{1}\\ f_{3}\\ \sigma_{12}\end{bmatrix}.$$

The previous procedure will give

$$\boldsymbol{V}_{1}=\begin{bmatrix}1\\ 0\\ 0\end{bmatrix},\ \boldsymbol{V}_{3}=\begin{bmatrix}0\\ 1\\ 0\end{bmatrix},\ n_{+}=0,\ n_{0}=1.$$

Since $n_{+}=0,$ Theorem 2.1 tells that the half-space problem does not need boundary conditions at $y=0$, where the moment system directly gives the solution (3.3). But the Grad boundary condition gives one more condition. So this condition asks for relations of the given values, e.g., $\bar{u}$ and $\bar{\sigma}$. We should also consider the stability of these additional solutions.

When $M=5$, we can check that $n_{+}>0$ and the Grad boundary condition does not satisfy

$$\boldsymbol{B}\boldsymbol{T}_{0}=\boldsymbol{0},$$

which shows the instability of the Grad boundary condition from Theorem 2.1. An analogous example is given in [28].

To overcome this drawback, one may change (3.2) as

$$c(\bar{\sigma}+\sigma_{12})+\hat{\chi}\left(u_{1}+\bar{u}+\sqrt{2}f_{3}\right)=0,$$

where $c>0$ is an arbitrary positive constant. Then the modified boundary condition will give

$$\bar{u}=-\frac{1}{c\hat{\chi}}\bar{\sigma},$$

which is stable about $h_{3}$. The boundary conditions involving higher-order moment variables can be modified in the same way. Authors of [28] suggest choosing $c=1$, such that the modified boundary condition differs from the Grad boundary condition only by the coefficients before the highest order moment variables, i.e., $f_{3}$ in this example. However, it’s not clear whether the modification is optimal in the sense of providing the most accurate solutions.

We systematically introduce a class of modified boundary conditions and prove their well-posedness. Assume

$$\mathbb{I}_{e}=\{\boldsymbol{\alpha}\in\mathbb{N}^{3}:\ \alpha_{2}\text{\ even},\ |\boldsymbol{\alpha}|\leq M\},\quad\mathbb{I}_{o}=\{\boldsymbol{\alpha}\in\mathbb{N}^{3}:\ \alpha_{2}\text{\ odd},\ |\boldsymbol{\alpha}|\leq M\},$$

and $m=\#\mathbb{I}_{e},\ n=\#\mathbb{I}_{o}.$ So we have $m+n=N$ and $m\geq n.$ Suppose the multi-indices with the even second component are always ordered before the ones with the odd second component, e.g., $(a_{1},0,a_{3})$ is ordered before $(b_{1},1,b_{3})$ for any $a_{1},a_{3},b_{1},b_{3}$. Then we can write the Grad boundary condition (2.5) as

$$\boldsymbol{E}\boldsymbol{M}(W_{o}-b_{o})+\hat{\chi}\boldsymbol{E}\boldsymbol{S}(W_{e}-b_{e})=0,$$

(3.4)

where $\hat{\chi}=\displaystyle\frac{2\chi}{2-\chi}\frac{1}{\sqrt{2\pi}}$. The matrix $\boldsymbol{M}\in\mathbb{R}^{m\times n}$ has entries $\left\langle{\mathcal{M}\xi_{2}\phi_{\boldsymbol{\alpha}}\phi_{\boldsymbol{\beta}}}\right\rangle$ for $\boldsymbol{\alpha}\in\mathbb{I}_{e}$ and $\boldsymbol{\beta}\in\mathbb{I}_{o}.$ And the matrix $\boldsymbol{S}\in\mathbb{R}^{m\times m}$ has entries $\left\langle{\mathcal{M}|\xi_{2}|\phi_{\boldsymbol{\alpha}}\phi_{\boldsymbol{\beta}}}\right\rangle$ for $\boldsymbol{\alpha},\ \boldsymbol{\beta}\in\mathbb{I}_{e}$. The variables are divided as

$$W=\begin{bmatrix}W_{e}\\ W_{o}\end{bmatrix},\ b=\begin{bmatrix}b_{e}\\ b_{o}\end{bmatrix},$$

where the elements of $W_{e}\in\mathbb{R}^{m}$ are $w_{\boldsymbol{\alpha}},\ \boldsymbol{\alpha}\in\mathbb{I}_{e}$ and the elements of $W_{o}\in\mathbb{R}^{n}$ are $w_{\boldsymbol{\alpha}},\ \boldsymbol{\alpha}\in\mathbb{I}_{o}.$ Every row of $\boldsymbol{E}\in\mathbb{R}^{n\times m}$ is a unit vector with only one component being one, such that the entries of $\boldsymbol{E}\boldsymbol{M}$ are $\left\langle{\mathcal{M}\xi_{2}\phi_{\boldsymbol{\alpha}}\phi_{\boldsymbol{\beta}}}\right\rangle$ $,\boldsymbol{\alpha}\in\mathbb{I}_{e},\ |\boldsymbol{\alpha}|\leq M-1,$ and $\boldsymbol{\beta}\in\mathbb{I}_{o}.$

Due to the recursion relation and orthogonality of Hermite polynomials [6], we can write

$$\boldsymbol{A}=\begin{bmatrix}\boldsymbol{0}&\boldsymbol{M}\\ \boldsymbol{M}^{T}&\boldsymbol{0}\end{bmatrix},\quad\boldsymbol{Q}=\begin{bmatrix}\boldsymbol{Q}_{e}&\boldsymbol{0}\\ \boldsymbol{0}&\boldsymbol{Q}_{o}\end{bmatrix},$$

where $\boldsymbol{Q}_{e}\in\mathbb{R}^{m\times m}$ and $\boldsymbol{Q}_{o}\in\mathbb{R}^{n\times n}.$ By definition, we also know that $\boldsymbol{M}$ is of full column rank and $\boldsymbol{S}$ is symmetric positive definite [23].

The modified boundary condition for the initial-boundary value problem is

$$\boldsymbol{H}(W_{o}-b_{o})+\hat{\chi}\boldsymbol{M}^{T}(W_{e}-b_{e})=0,$$

(3.5)

where $\boldsymbol{H}\in\mathbb{R}^{n\times n}$ is an arbitrary symmetric positive definite matrix. For the half-space problem, the boundary condition should write as

$$\boldsymbol{H}(W_{o}+\bar{W}_{o}-b_{o})+\hat{\chi}\boldsymbol{M}^{T}(W_{e}+\bar{W}_{e}-b_{e})=0,$$

(3.6)

where $\bar{W}_{e}$ and $\bar{W}_{o}$ are given by the flow outside the boundary layer.

Inspired by the illustrative examples, we can first find a matrix $\boldsymbol{C}\in\mathbb{R}^{n\times n_{+}}$ and multiply (3.6) left by $\boldsymbol{C}^{T}$. Then, we can use Theorem 2.1 to check the well-posedness of the half-space problem. Finally, we may solve the remaining part of (3.6) to obtain relations between the given vectors, e.g., $\bar{W}_{o}$ and $\bar{W}_{e}$.

Following this way, we first calculate the value of $n_{+}$ by the special block structure of $\boldsymbol{A}$ and $\boldsymbol{Q}$. We may as well write

$$\boldsymbol{G}=\begin{bmatrix}\boldsymbol{G}_{e}&\\ &\boldsymbol{G}_{o}\end{bmatrix},\ \boldsymbol{X}=\begin{bmatrix}\boldsymbol{X}_{e}&\\ &\boldsymbol{X}_{o}\end{bmatrix},\$$

where $\boldsymbol{G}_{e}\in\mathbb{R}^{m\times p_{1}}$, $\boldsymbol{G}_{o}\in\mathbb{R}^{n\times p_{2}}$, $\boldsymbol{X}_{e}\in\mathbb{R}^{m\times r_{1}}$, $\boldsymbol{X}_{o}\in\mathbb{R}^{n\times r_{2}}$. And

$$\boldsymbol{V}_{1}=\begin{bmatrix}\boldsymbol{Y}_{1}&\\ &\boldsymbol{Z}_{1}\end{bmatrix},\ \boldsymbol{V}_{2}=\begin{bmatrix}\boldsymbol{Y}_{2}&\\ &\boldsymbol{Z}_{2}\end{bmatrix},\ \boldsymbol{V}_{3}=\begin{bmatrix}\boldsymbol{Y}_{3}&\\ &\boldsymbol{Z}_{3}\end{bmatrix},$$

where $\boldsymbol{Y}_{1}=\boldsymbol{G}_{e}\boldsymbol{X}_{e},\ \boldsymbol{Z}_{1}=\boldsymbol{G}_{o}\boldsymbol{X}_{o}$,

$$\mathrm{span}\{\boldsymbol{Y}_{2}\}=\mathrm{span}\{\boldsymbol{M}\boldsymbol{G}_{o}\},\ \mathrm{span}\{\boldsymbol{Z}_{2}\}=\mathrm{span}\{\boldsymbol{M}^{T}\boldsymbol{G}_{e}\},$$

$\boldsymbol{Y}_{3}\in\mathbb{R}^{m\times(m-r_{1}-p_{2})}$ and $\boldsymbol{Z}_{3}\in\mathbb{R}^{n\times(n-r_{2}-p_{1})}$. Under these assumptions, we claim that

Incidentally, applying the above result in Lemma 3.1 to $\boldsymbol{G}^{T}\boldsymbol{A}\boldsymbol{G}$, we have $r_{1}+p_{2}=r_{2}+p_{1}.$ Then the required matrix $\boldsymbol{C}\in\mathbb{R}^{n\times n_{+}}$ can be found.