Assume there are two Riemannian metrics on a manifold ( open or closed) with the same set of all geodesics. Are they proportional by a constant? If not in general, what are the affirmative results in this direction?

2$\begingroup$ A question related in spirit: "Probing a manifold with geodesics" mathoverflow.net/questions/81622/… $\endgroup$– Joseph O'RourkeMay 29, 2013 at 17:57

5$\begingroup$ The counterexamples to 'uniqueness up to constant multiples' mentioned by alvarezpaiva and Christi Stoica, notwithstanding, it is true that for the generic metric $g$, even in dimension $2$, the only metrics that have the same geodesics (up to reparametrization) as $g$ are the constant multiples of $g$. Only very special metrics (in dimension $2$, these were classified by Dini) share their geodesics with more metrics than that. $\endgroup$– Robert BryantMay 30, 2013 at 10:36

$\begingroup$ A related question: "From shortest path to manifold structure" mathoverflow.net/questions/116292/… $\endgroup$– AerykOct 11, 2013 at 1:07

1$\begingroup$ This was one of Hilbert's questions. $\endgroup$– Brian RushtonFeb 3, 2015 at 15:57
6 Answers
The short answer is no, the geodesics do not determine the metric. For example, in the CayleyKlein model of hyperbolic geometry the geodesics are straight lines. It is however rather rare for two Riemannian metrics to have the same geodesics. In two dimensions these metrics were studied by Liouville (see Livre VI of Darboux's Théorie générale des surfaces). Their geodesic flow is completely integrable and, in fact, admits an additional integral of motion quadratic in the momenta.
A basic results is that if locally the geodesics are straight lines (or can be mapped to straight lines), the metric has constant curvature (Beltrami's theorem). Other rigidity results of this kind exist.
For more on this topic you can consult the works of Topalov and Matveev.
A simple counterexample is the space $\mathbb R^n$, with a metric $g_{ab}$ independent on the point. As examples, the Euclidean space $\mathbb R^n$, but also the Minkowski spacetime (which is semiRiemannian, but has the same geodesics as $\mathbb R^4$). The geodesics are the lines in $\mathbb R^n$, no matter how we choose the constant metric $g_{ab}$. Hence, different metrics can give the same set of geodesics.
I hope that my ``answer'' will not be understood solely as a propaganda of my survey http://arxiv.org/abs/1101.2069 where I discussed
(1) how, given geodesics, to reconstruct a connection (in both cases: when geodesics are parameterised and not parameterised)
(2) how, given a class od projectively related connections, to reconstruct a metric, and what are advantages of additional curvature assumptions on the metric
(3) what is the freedom in reconstructing the metric by geodesics  in particular I proved the statement wellknown to experts and mentionen by Robert Bryant in his comment that for generic metric the geodesics, even unparameterized, determine the metric.
I attended today a talk at a conference, where Graham Hall presented some very interesting results about recovering the metric from geodesics, and mentioned some previous results too. Here are some of his papers about this:
http://www.sciencedirect.com/science/article/pii/S0393044011002269
http://link.springer.com/article/10.2478/s1153301200876
The example of Cristi Stoica is isometric to Euclidean space $\mathbb R^n$ if the $g_{ab}$ defines a Riemannian metric.

1$\begingroup$ But it is still a counterexample to the question. $\endgroup$ Jun 1, 2013 at 7:32

1$\begingroup$ What is the point in having an isometric counterexample? In that case why not taking $\mathbb R^2$ divide the distance in the $y$ direction by $2$. $\endgroup$ Jun 1, 2013 at 7:57

1$\begingroup$ I learned from Vladimir Matveev, that even isometric examples of metrics with the same geodesics can be interesting. He considered the standard metric on the projective plane, and then its pullback by a projective transformation. Athough this seems obvious it gives rise to quadratic integrals of motion and the construction of interesting metrics such as the "Poisson spheres" (that can also be constructed using the Hopf fibration as submersive metrics obtained from metrics $SO(3)$ with a left invariant metric. $\endgroup$ Jun 1, 2013 at 14:53

4$\begingroup$ That's true. But the question was whether, by assuming Riemannian metric, and by knowing the geodesics, we can recover the metric, say up to a scale factor. I just gave the simplest counterexample. Now, if you add the condition that the resulting manifolds are not isometric, we can modify the counterexample to satisfy this. A simple example is a flat torus. Flat torii have the same geodesics, but are not isometric. Moreover, in general they can't be made isometric by rescaling the metric (unless you use different scaling factor in different directions). $\endgroup$ Jun 4, 2013 at 4:17

$\begingroup$ You may be interested in contributing to a proposed Spanish language version of math stackexchange; it could use some input from fluent professors and students: area51.stackexchange.com/proposals/64529/… $\endgroup$ Feb 2, 2014 at 20:51
For the sake of completeness, here is the local story, assuming that geodesics are understood as maps from intervals to the manifold and "the same" is understood literally (rather than "up to reparameterization"). Equivalently, two metrics have the same exponential map.
Results along these lines are wellknow but hard to find (with usual references pointing to old papers by Eisenhart which I find unreadable), while this should be standard textbook material...
I will say that two Riemannian metrics with the same LeviCivita connection form an LC pair. I will say that an LC pair is trivial if the metrics are homothetic to each other, i.e. are scalar multiples of each other.
The first thing to observe is that two Riemannian metrics on the given manifold have the same exponential map if and only if they form an LC pair. One direction is immediate, as geodesics on a Riemannian manifold are curves defined by the differential equation $\nabla_{c'} c'=0$. The opposite direction is harder, it is proven (for instance) in:
M.Spivak, "Comprehensive Introduction to Differential Geometry" (Publish Or Perish, 2000), volume 2, chapter 6, Appendix 1.
Theorem. Suppose that $(M,g_1)$ is a Riemannian manifold such that $M$ admits another Riemannian metric $g_2$ such that $g_1, g_2$ form a nontrivial LC pair. Then $(M,g_1)$ is locally isometric to a product of Riemannian manifolds.
Proof. I will need several ingredients. Recall that the holonomy group $Hol_{p,g}$ of the Riemannian metric $g$ (rel. basepoint $p\in M$) depends only on the Riemannian affine connection $\nabla$, $Hol_{p,g}=Hol_{p,\nabla}$. Our discussion is local, it suffices to restrict to loops contained in a totally convex neighborhood of $p$. Since $\nabla$ is a Riemannian connection (for a metric $g$) then $G=Hol_{p,\nabla}$ will preserve the quadratic form $g_p$ on $T_pM$. Thus, $G$ is a relatively compact subgroup of $O(n)=Aut(T_pM, g_p)$. This group is essentially independent of the basepoint: Holonomy groups at different basepoints $p_1, p_2$ in $M$ are "conjugate" via the parallel transport along a path from $p_1$ to $p_2$.
Next, comes a fact from elementary representation theory. Suppose that $V$ is a finitedimensional real vector space and $G< GL(V)$ is a (relatively) compact subgroup whose action on $V$ is irreducible, i.e. $V$ does not admit a nontrivial $G$invariant direct sum decomposition $V=V_1\oplus V_2$. Then any two $G$invariant quadratic forms $q_1, q_2$ on $V$, are scalar multiples of each other.
Applying this result to the holonomy groups of Riemannian metrics $g_1, g_2$ on a connected manifold $M$ we obtain:
If $g_1, g_2$ have the same LeviCivita connection $\nabla$ then either the holonomy groups $Hol_{p,\nabla}$ are reducible for some (equivalently, every) $p\in M$ or the metrics $g_1, g_2$ are conformal to each other: $$ g_2= e^{2f}g_1, f\in C^\infty(M). $$
Another ingredient that we need is deRham's theorem:
If $g$ is a Riemannian metric on $M$ whose holonomy is reducible, then $M$ locally splits as a Riemannian direct product.
See e.g. Theorem 3.1 on p. 228 in Petersen's book "Riemannian Geometry."
Applying this to an LC pair $g_1, g_2$, we conclude:
Either (1) $(M,g_1)$ locally splits as a Riemannian direct product or (2) the metrics $g_1, g_2$ are conformal to each other.
Consider the case (2):
$$
g_2= e^{2f}g_1, f\in C^\infty(M).
$$
I will prove that the function $f$ is constant, i.e. $g_1, g_2$ form a trivial LC pair. Let $grad(f)$ denote the gradient field of $f$ with respect to $g_1$. Then, since $g_1, g_2$ form an LC pair, one obtains:
$$
X(f)Y + Y(f)X  g_1(X,Y) grad(f)=0,
$$
for any two vector fields $X, Y$ on $M$. I claim that $grad(f)$ is identically zero on $M$, i.e. $f$ is constant.
Indeed, consider vector fields $X=Y= grad(f)$. Then, the equation becomes
$$
2 grad(f)^2 grad(f)  grad(f)^2 grad(f) = grad(f)^2 grad(f)=0,
$$
i.e. $grad(f)$ is identically zero. Thus, the metric $g_2$ is a constant multiple of $g_1$. qed
Clearly, the "converse" to this theorem holds as well assuming that $(M,g_1)$ has a global nontrivial deRham decomposition.
Applying the theorem inductively, one obtains the following corollary:
Corollary. Let $(M,g)$ be a Riemannian manifold which has local deRham decomposition $$ M=M_0\times M_1\times ... \times M_m, g= g_0 \oplus g_1\oplus ... \oplus g_m $$ where $g_0=\delta_{ij}$ is the standard flat metric on ${\mathbb R}^k$ (and the rest of the factors are nonflat and do not split any further). Suppose that $g, h$ is an LC pair on $M$. Then (locally) $h$ has the form $$ h= h_0 \oplus a_1 g_1\oplus ... \oplus a_m g_m $$ where $h_0$ is a constant metric tensor on ${\mathbb R}^k$ and $a_1,...,a_m$ are certain constants.