MathOverflow is a question and answer site for professional mathematicians. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

For a smooth function $f:M\to \mathbb{R}$ one usually defines the degeneracy and index of a critical point $p\in M$ in terms of the eigenvalues of the Hessian matrix $(\partial^2 f/\partial x_i\partial x_j)$.

On the other hand, if we have a Riemannian metric $g$ we can define the Hessian tensor $H(f,g)=\nabla df$. From what I understand, one can recover the Hessian matrix from this tensor and define the index and degeneracy of a critical point as above.

My question is: can we cut out the middle man, i.e., is there a natural way to define the index and degeneracy of a critical point from the Hessian tensor without using the Hessian matrix?

share|cite|improve this question
A symmetric bilinear function has an index, so the answer would be yes. – Ryan Budney Aug 6 '10 at 22:27
up vote 9 down vote accepted

The Hessian tensor (defined at a critical point of a function) is a symmetric bilinear form on the tangent space of the manifold at that point. For a symmetric bilinear form on a finite-dimensional vector space (in characteristic different from $2$) there is always a basis such that the corresponding symmetric matrix is diagonal. The diagonal entries are not well-defined independent of diagonalizing basis, but if your field is $\mathbb R$ then the number of positive (or negative) entries is well-defined. You can intrinsically define the index of a nondegenerate critical point as the maximal dimension of a tangent vector subspace on which the Hessian is negative definite.

share|cite|improve this answer

Let $f : M\to \mathbb R$ be a smooth function, where $M$ is a smooth (infinite-dimensional, if you like) manifold. Pick a point $m\in M$. Then the value $f(m) = f^{(0)} \in \mathbb R$ is well-defined, of course, and also $f^{(1)}(m) = {\rm d}f_m$ is well-defined as a linear map ${\rm T}_mM \to \mathbb R$, i.e. as an element in ${\rm T}^*_m M$. In general, the higher derivatives are not well-defined, even if you choose a metric, although they are if you choose coordinates. What is well-defined is the "$k$-jet" of the function for each $k$; the $k$-jet of $f$ takes values in a vector bundle over $M$ that's locally isomorphic to $\bigoplus_{j=0}^k (\operatorname{Sym}^j( {\rm T}_m M))^*$, but choosing such an isomorphism essentially requires choosing local coordinates.

(Actually, more is true: the $k$-jet bundle surjects onto the $(k-1)$-jet bundle, so that there is a well-defined $\infty$-jet bundle; these surjections are compatible with the local trivializations, so that a trivialization identifies the $\infty$-jet bundle with $(\operatorname{Sym}^\bullet({\rm T}_mM))^*$. Naively you would define the Taylor series as living here. The system of surjections defines a filtration on the $\infty$-jet bundle, and $(\operatorname{Sym}^\bullet({\rm T}_mM))^*$ is the corresponding associated graded bundle.)

But there is one situation in which you can define a symmetric tensor $f^{(k)}(m) : ({\rm T}_mM)^{\otimes k} \to \mathbb R$. Namely, if all $f^{(1)}(m), \dots, f^{(k-1)}(m)$ vanish, then $f^{(k)}(m)$ is well-defined. This is because under changes of coordinates, the coordinate-dependent term $f^{(k)}(m) \in (\operatorname{Sym}^k({\rm T}_m M))^*$ changes by some term depending linearly on $\{f^{(1)}(m), \dots, f^{(k-1)}(m)\}$ (and polynomially on the derivatives of the change-of-coordinates function).

This is the situation for the Hessian. If $m$ is a critical point for $f$, then $f^{(1)}(m) = {\rm d}f_m = 0$, and so the second derivative is well-defined as a symmetric map $H = f^{(2)}(m) : ({\rm T}_mM)^{\otimes 2} \to \mathbb R$.

In any case, if $V$ is any vector space, and $\phi: V^{\otimes 2} \to \mathbb R$ any symmetric linear map, then the Morse index of $\phi$ is defined as the dimension (if it is finite, or $\infty$ otherwise) of any maximal subspace $W\subseteq V$ so that $\phi\rvert_{W\otimes W}$ is negative-definite. It is a straightforward lemma that this number (or $\infty$) does not depend on which maximal subspace you use. Of course, you can also define a "Morse coindex" as the Morse index of $-\phi$, and finally one can write down the degeneracy of $f$ at $m$ as the dimensional of the kernel of $\phi$ thought of as a map $\phi : V \to V^*$.

Point being, there are no other choices necessary to go: smooth function $f: M\to \mathbb R$ and choice $m\in M$ of critical point $\Rightarrow$ Hessian of $f^{(2)}(m) : ({\rm T}_mM)^{\otimes 2} \to \mathbb R$ $\Rightarrow$ Morse index of $f^{(2)}(m)$. At non-critical points, you cannot do the first step.

Finally, let me comment on the use of a metric. For applications of Morse theory to the topology of manifolds, picking a metric is important, as it lets you convert the differential one-form ${\rm d}f$ into a vector field and thus consider the corresponding flow lines, etc. From a topological point of view, what particular metric you use doesn't matter much: the space of metrics is contractible. But it is important; just not for the definition of Morse index.

Recall that a metric $g$ determines for each $m\in M$ a linear map $\operatorname{Sym}^2 {\rm T}_m M \to \mathbb R$, and hence a linear map $g_m: {\rm T}_m M \to {\rm T}^*_m M$. Being a metric means that this linear map is an injection (and hence an iso in finite-dimensions). So you should try to write down an inverse $g^{-1}_m: {\rm T}_m^* M \to {\rm T}_m M$. You can do this in finite-dimensions; in infinite dimensions the inverse is not everywhere-defined. Then you can define $\nabla f_m = {\rm d}f_m \, g^{-1}_m$ and, if you want, the Hessian to be the matrix $f^{(2)}_m g^{-1}_m$. Since $g$ is positive-definite, so is $g^{-1}$, and so the dimensions of the kernel and positive and negative eigenspaces of $f^{(2)}_m g^{-1}_m$ are the same as for $f^{(2)}$.

share|cite|improve this answer

What you call the Hessian matrix is defined using a choice of local co-ordinates near the critical point. The Hessian matrix at the critical point consists of the components of the Hessian tensor written with respect to the local-ordinates. In other words, at any critical point the two don't just have the same index; they are the same thing.

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.