To propagate an extremum without a weight, we write the linearized operator in real coordinates. Its symbol is continuous and semipositive. At a point where a ball outside the maximum set touches it, the normal direction must lie in the kernel of the symbol; this forces the maximum set to be invariant under the flows of vector fields with values in the range of the symbol.
In a holomorphic chart $z_ j=x_ j+\mathrm{i} y_ j$, let $C_{\rm cf}$ be the Hermitian matrix characterized by
04DC9E828E52
14.22.9
\begin{equation} \frac{\bigl (\mathrm{i}\sum H_{j\bar k}\mathrm{d}z_ j\wedge \mathrm{d}\bar z_ k\bigr )\wedge T}{\mathrm{d}V} =\operatorname {Tr}(C_{\rm cf}H) \end{equation}
for every Hermitian matrix $H=(H_{j\bar k})$. Positivity of mixed products makes $C_{\rm cf}$ positive semidefinite. Write $C_{\rm cf}=P_{\rm cf}+\mathrm{i} Q_{\rm cf}$, where $P_{\rm cf}$ is real symmetric and $Q_{\rm cf}$ is real skew-symmetric. In the real coordinate order $(x_1,\ldots ,x_ n,y_1,\ldots ,y_ n)$,
7A7369E544A8
14.22.10
\begin{equation} L\varphi :=\frac{\mathrm{dd}^{\mathrm{c}}\varphi \wedge T}{\mathrm{d}V} =\operatorname {Tr}(aD^2\varphi ),\qquad a=\frac1{8\pi } \begin{pmatrix} P_{\rm cf}
& Q_{\rm cf}
\\ -Q_{\rm cf}
& P_{\rm cf}
\end{pmatrix}. \end{equation}
Indeed $\varphi _{j\bar k}=(\varphi _{x_ jx_ k}+\varphi _{y_ jy_ k} +\mathrm{i}\varphi _{x_ jy_ k}-\mathrm{i}\varphi _{y_ jx_ k})/4$; inserting this expression in (14.22.9) proves (14.22.10). For real coordinate vectors $\xi ,\eta \in \mathbb R^ n$, moreover,
\[ (\xi ,\eta )^ Ta(\xi ,\eta ) =\frac1{8\pi }(\xi -\mathrm{i}\eta )^*C_{\rm cf}(\xi -\mathrm{i}\eta )\geq 0. \]
Thus $a$ is a continuous positive semidefinite principal symbol, the operator $L$ has no first-order term in holomorphic coordinates, and the range of $a$ is that of $Q^\sharp $.
An exterior-ball argument propagates the maximum for this operator. Let $W\in C^2$ satisfy $LW\geq 0$ in a chart, suppose $W\leq M$, and put $K=\{ W=M\} $. If $z\in K$ is a point of $K$ nearest to a point $y\notin K$ of the chart, then
2692BD3333CA
14.22.11
\begin{equation} (y-z)^ Ta(z)(y-z)=0. \end{equation}
To prove this, suppose the left side is positive, set $r_0=|y-z|$, and shrink the tangent ball toward $z$: $y'=z+\theta (y-z)$, $R=\theta r_0$, with $0<\theta <1$ small. The closed ball $\overline{B(y',R)}$ lies in the chart and meets $K$ only at $z$. This follows from the triangle inequality; equality in $|k-y|\leq |k-y'|+|y'-y|\leq r_0$, for $k\in K$, forces $k=z$. For large $\alpha >0$, the function
\[ \psi (x)=\mathrm{e}^{-\alpha R^2}-\mathrm{e}^{-\alpha |x-y'|^2} \]
satisfies
\[ L\psi =2\alpha \mathrm{e}^{-\alpha |x-y'|^2} \bigl (\operatorname {tr}a-2\alpha (x-y')^ Ta(x)(x-y')\bigr )<0 \]
in a neighborhood of $z$, because $z-y'$ is a nonzero multiple of $y-z$, so that $(z-y')^ Ta(z)(z-y')>0$ by assumption. Choose a sufficiently small $r<R$, and consider the lens $D=B(z,r)\cap B(y',R)$. On the part of its boundary lying on the sphere $\partial B(y',R)$, we have $\psi =0$ and $W-M\leq 0$. On the remaining compact boundary portion, $W-M<0$, since the tangent ball meets $K$ only at $z$. Hence, for sufficiently small $\varepsilon >0$, $W-M\leq \varepsilon \psi $ on $\partial D$. At a positive interior maximum of $W-M-\varepsilon \psi $, the Hessian would be nonpositive, whereas $L(W-M-\varepsilon \psi )>0$; this is impossible since $a\geq 0$. Hence the inequality $W-M\leq \varepsilon \psi $ holds in $D$. Outside the ball $B(y',R)$ we have $\psi \geq 0$, so this inequality holds in a full neighborhood of $z$. Since equality holds at $z$, the function $W-M-\varepsilon \psi $ has a local maximum at $z$, and therefore
\[ 0\leq LW(z)\leq \varepsilon L\psi (z)<0, \]
proving (14.22.11).
Since $a\geq 0$, its range is the orthogonal complement of its kernel, and (14.22.11) implies that $y-z$ lies in this kernel. Consequently $(y-z)\cdot Z(z)=0$ for every vector field $Z$ with values in the range of $a$. If $Z$ is locally Lipschitz and $\gamma $ is one of its flow curves starting in $K$, let $d(t)$ denote the squared distance from $\gamma (t)$ to $K$ in a sufficiently small compact chart neighborhood. Shortening the time interval ensures that nearest points $z_ t$ lie in the chart interior. Write $D^+d(t)=\limsup _{h\downarrow 0}(d(t+h)-d(t))/h$ for the upper right derivative. Holding such a nearest point fixed when differentiating the distance from above gives
\begin{align*} D^+d(t)& \leq 2(\gamma (t)-z_ t)\cdot Z(\gamma (t))\\ & =2(\gamma (t)-z_ t)\cdot \bigl (Z(\gamma (t))-Z(z_ t)\bigr ) \leq 2\operatorname {Lip}(Z)d(t). \end{align*}
As $d(0)=0$, Gronwall's inequality gives $d(t)=0$. The same argument for $-Z$, followed by continuation, proves invariance under both signs of time and under finite concatenations of such flows.
of 14.22.1(2).
Apply the preceding argument to $W=w$ and to $W=-w$ in each chart; both satisfy $LW=0$ by (14.20.1). The maximum set of $w$ is nonempty, closed, and contains every control orbit starting in it. If every orbit is dense, this maximum set is all of $X$, proving constancy. The argument for the minimum set is identical.
It remains to prove the bracket criterion. Suppose that a family of smooth vector fields with values in $\operatorname {im}Q^\sharp $ is bracket generating. At $p$, let $\mathcal D_ p$ be the span of the values of all original fields pushed forward by finite compositions of their local flows and inverse flows. This distribution contains the original fields and is invariant under their flows. Differentiation of pushed-forward fields along a flow shows that it contains their nested Lie brackets; finite-dimensional subspaces are closed, so the derivative remains in $\mathcal D_ p$. Induction and the Jacobi identity give every iterated bracket. The assumed rank condition therefore gives $\mathcal D_ p=T_ pX$.
Choose $2n$ such pushed-forward fields independent at $p$. The flow of each is a conjugate of an original flow and hence preserves its orbit. The map obtained by successively flowing for $2n$ small times has invertible derivative at zero. The inverse function theorem puts a neighborhood of $p$ in its orbit. All orbits are open, and their partition of connected $X$ has only one member. This proves the Hörmander-type sufficient condition; it requires neither a constant-rank hypothesis nor a smooth matrix square root of $a$. In particular, the columns of a smooth symbol $a$ may be used as the original vector fields.
$\square$
For $C^{2,1}$ potentials, the columns of $a$ are locally Lipschitz, so the orbit criterion still applies to them. The bracket criterion requires smooth vector fields.
Comments
Corrections, questions, and mathematical discussion are welcome. Sign in with GitHub to comment. Comments are public.