Morse lemma via Moser's trick

969 Views Asked by At

In Abraham and Marsden's Foundations of Mechanics, they prove Morse lemma via Moser's trick. They are able to reduce the proof so that it suffices to find a smooth family of vector fields $Z_t$ such that $$\iota_{Z_t}\omega_t+(f-g)=0, \quad Z_t(0)=0,$$ where $f$ is the function under consideration in Morse lemma statement, $g(x)=\frac{1}{2}D^2f(0)(x,x)$ and $\omega_t=tdf+(1-t)dg$.

Now, the book says that it is "easy to see that $Z_t$ exists near $0$ by the nondegeneracy hypothesis", which I assume is referring to the non-degeneracy of the hessian $D^2f(0)$.

However, I can't see how it is easy to see that. For instance, $\omega_t(0)$ is $0$, so $Z_t(0)$ could be whatever we want and I see no way to easily reconcile this*. Furthermore, $\omega_t$ is a $1$-form and we want a $Z_t$ such that $\omega_t(Z_t)$ is some number. This gives a lot of redundancy for $Z_t$.

My question is: How to build such $Z_t$?

*This is quite different from the usage of Moser's trick on the symplectic situations (Darboux's theorem, say), where $\omega_t$, when constant on a region of interest (like the origin in this case), is usually non-degenerate (not to mention being a $2$-form, which is relevant for the next thing I am about to say)


For those not familiar, Morse lemma is as follows.

Let $f:M \to \mathbb{R}$ be such that $\mathrm{Hess}_p$ is non-degenerate and $f(p)=0$. Then there exists a chart $\phi$ around $p$ such that in local coordinates $$f(x)=D^2f(0)(x,x).$$

1

There are 1 best solutions below

0
On BEST ANSWER

Since $df(0) = 0$, given a vector $v$ we can write $$ df(x)(v) = \big<G(x)v,x\big>, $$ where $$ G(x) = \int_0^1 \mathrm{Hess}\;f(sx)\;ds. $$ In fact, $$ df(x)(v) = \int_0^1 \frac{d}{ds}df(sx)(v)\;ds = \int_0^1 D^2f(sx)(v,x)\;ds = \big<\big(\int_0^1\mathrm{Hess}\;f(sx)\;ds\big)v,x\big>. $$ Thus, the equation $$ i_{Z_t}\omega_t = g - f $$ becomes $$ \big<(tG(x) + (1 - t)H)Z_t(x),x\big> = \big<((1/2)H - G(x))x,x\big> $$ where $H = \mathrm{Hess}\;f(0)$. Whence, it suffices to find $Z_t$ such that $$ (tG(x) + (1 - t)H)Z_t(x) = ((1/2)H - G(x))x. $$ Since $G(0) = H$, the operator $(tG(x) + (1 - t)H)$ is invertible for $x$ near $0$. Define $$ Z_t(x) = (tG(x) + (1 - t)H)^{-1}((1/2)H - G(x))x $$ in a neighbourhood of $0$.