$\newcommand{\bR}{\mathbb{R}}\newcommand{\pa}{\partial}$This questions has some nebulous roots in Morse theory. The most general version goes as follows. Fix an integer $n\geq 2$. Suppose that we have a smooth function $f$ defined on an open neighborhood $U$ of $0$ in $\bR^n$. To a smooth Riemannian metric $g$ defined on $U$ we associated the gradient $\nabla^g f$.

For which smooth functions $f$ is the correspondence

$$g\mapsto \nabla^g f $$

injective, i.e., from the knowledge that $\nabla^{g_0} f=\nabla^{g_1} f$ on $U$ we can conclude that $g_0=g_1$ on some neighborhood $V$ of $0$ in $\bR^n$, $V\subset U$.

Clearly, the above correspondence is not injective if $f\equiv const$, or $f(x)$ is linear in $x$.

The next interesting case is when $q$ is quadratic. Suppose that

$$ f(x)=q(x):=\frac{1}{2}(x_1^2+\cdots +x_n^2). $$

Denote by $g_0$ the canonical Euclidean metric on $\bR^n$ so that

$$\nabla^{g_0}q =\sum_i x_i\pa_{x_i}. $$

The special case of the question goes as follows.

If $g$ is a real analytic Riemann metric defined on a neighborhood $U$ of $0$ and $\nabla^g q(x)= \nabla^{g_0} q(x)$, $\forall x\in U$, can we conclude that $g=g_0$ in a neighborhood of $0$?

I was not able to track results of this kind in the literature, and I would appreciate any pointers.


If $V$ is a vector field such that $Vf=0$ at every point, then the 2-tensor $$ h^{ij}=g^{ij}+V^iV^j $$ is also an inverse metric and satisfies $\nabla^gf=\nabla^hf$. In a coordinate chart this amounts to finding a vector field orthogonal (in the Euclidean sense on the chart) pointwise orthogonal to the gradient of $f$. In the case of your second example such a vector field has to vanish at the origin but need not vanish elsewhere.

Examples of such vector fields are not hard to find. Just take $V=(\partial_2f,-\partial_1f,0,\dots,0)$ and change the indices if needed. If this simple idea doesn't give a nonzero vector field, then $\nabla f\equiv0$ and any $V$ will do. In any neighborhood one can find a nonzero vector field and therefore a distinct metric with the same gradient of $f$.

Therefore one function $f$ is never enough. If you have a collection of functions, you need any vector field orthogonal to every one of them to vanish identically. You need at least $n$ functions in $n$ dimensions.

If you take a non-analytic function $f$ and ask the question for analytic metrics, the answer may be different, but that's another story…

  • 2
    $\begingroup$ It's probably worth pointing out, though, that there are constraints: In $n$ dimensions, if you specify $n$ functions $f^1,\ldots, f^n$ and $n$ vector fields $X^1,\ldots, X^n$ and you want to know whether there is a metric $g$ such that $X^i=\nabla^g f^i$, then you will need it to be true that the functions $h^{ij} = \mathrm{d}f^i(X^j)$ satisfy $h^{ij} = h^{ji}$, since, by definition, $$h^{ij} = \mathrm{d}f^i(X^j) = g(\nabla^gf^i,\nabla^gf^j).$$ Of course, you also need the matrix $h = (h^{ij})$ to be positive definite as well as symmetric. With these assumptions, $g$ exists and is unique. $\endgroup$ – Robert Bryant Apr 26 '15 at 10:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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