17

Let $(M,g)$ be a compact, boundaryless Riemannian $n$-manifold. Write $\Delta$ for the Laplace–Beltrami operator on $M$, and let $\{\phi_j\}_{j\ge 0}$ be an $L^2(M)$-orthonormal basis of smooth eigenfunctions, $\Delta \phi_j=\lambda_j \phi_j$ with $0=\lambda_0<\lambda_1\le\lambda_2\le\cdots$.

Is the following statement true?

If $\phi_j(x)=\phi_j(y)$ for all $j$, then $x=y$.

I would be interested in more general principles, concerning the seperation of points, that imply it.

What I think might work (and where I’m unsure):

  • Pick $f\in C^\infty(M)$ with $f(x)\neq f(y)$. Expand in the $L^2$-eigenbasis:

    $$ f=\sum_{j=0}^\infty c_j\,\phi_j\quad\text{in }L^2(M). $$

    If $\phi_j(x)=\phi_j(y)$ for all $j$, then formally evaluating termwise would give $f(x)=f(y)$, a contradiction. The gap is that this is only an $L^2$-expansion, so termwise evaluation at points needs justification, that might involve properties such as elliptic regularity.

Mathbds
  • 485
  • Doesn't this already follow from having an L^2 basis of continuous function, without invoking any of the Laplace eigenfunction properties? Assume $x\neq y$, pick a function $f$ such that $f(x)=0$ and $f(y)=1$, look at is representation in the basis. Contradiction. – quarague Aug 31 '25 at 02:35
  • 1
    @quarague please elaborate on your final step of “look at its representation in the basis” I really doubt it’s that easy – peek-a-boo Aug 31 '25 at 02:58
  • @peek-a-boo It's been a long time since I looked at this stuff but note that $M$ is compact so we can pick both the basis and $f$ to be absolutely continuous, so it seems to me we can just evaluate termwise and get absolutely converging series at $x$ and $y$, so there is a contradiction. – quarague Aug 31 '25 at 06:12
  • 3
    @quarague we can even take the basis and $f$ to be $C^{\infty}$ functions, that’s not the part I’m worried about. It’s the mode of convergence of the series which is the issue. By definition the basis only gives convergence in $L^2$-norm, but you’re using pointwise convergence everywhere. This is the unjustified step (and is why I had to really use elliptic PDE facts to justify the convergence also takes place in $C^{\ell}$-norms). Note by basic measure theory, $L^2$ convergence would give, despite smoothness, a subsequence which merely converge pointwise a.e, but this isn’t enough. – peek-a-boo Aug 31 '25 at 13:36
  • 2
    The part about absolute continuity and absolutely converging series is probably some intuition/well-known facts regarding Fourier series on the circle, but I guess the point is we have to generalize that to higher dimensions (and arbitrary Riemmannian metrics) as well. – peek-a-boo Aug 31 '25 at 13:40
  • 1
    I think you can mimic the Fourier-theoretic argument for the circle (or more generally the flat $n$-torus), but you'd need a) the identification of the Fréchet space $C^\infty(M)$ as the projective limit of all the $L^2$ Sobolev spaces on $M$, and b) spectral asymptotics à la Weyl for the Laplace--Beltrami operator. However, you can't obviously do any of that without some combination of elliptic regularity and heat kernel machinery, so you'd might as well stick with the two answers already given. – Branimir Ćaćić Aug 31 '25 at 19:03

2 Answers2

16

This is provable using the heat kernel of the Laplace-Beltrami operator. I'll use as a reference the book Topics in Spectral Geometry by Mangoubi, Levitin and Polterovich, but you can also check Doubt on the definition of heat kernel and the second reference within, or other places.

There exists a unique function $e = e(t,x,y)$ named heat kernel or fundamental solution of the heat equation on $M$ which is $C^\infty$ on $(0, +\infty) \times M \times M$ and satisfies the following assertion:
For all $f \in C^0(M)$ and all $x \in M$, we have: $$\lim_{t \downarrow 0} \int_{M} e(t,x,y)f(y) \mathrm{d}y = f(x)$$ Moreover, and more importantly for this problem, we know that it has the following series expression: $$e(t,x,y) = \sum_{j = 0}^\infty e^{-\lambda_j t}\phi_j(x)\phi_j(y)$$ where the equality holds pointwise.

Therefore, if we assume that $\phi_j(x) = \phi_j(x')$ (just changing the notation to have something convenient to write) for all $j$, then we'd have $e(t,x,y) = e(t,x',y)$ for all $t$ and all $y$, looking at the series expansion.
But this implies that, for all $f \in C^0(M)$, we have: $$f(x) = \lim_{t \downarrow 0} \int_{M} e(t,x,y)f(y) \mathrm{d}y = \lim_{t \downarrow 0} \int_{M} e(t,x',y)f(y) \mathrm{d}y = f(x')$$

By a partition of unity argument for example though it's not the only way, this indeed implies that $x = x'$, and we are done.

Bruno B
  • 8,475
  • Nobody told me that I forgot to say that the heat kernel is unique not just for the assertion we used but also because it solves the heat equation... It's not needed here so I won't bump the question with a useless edit though. – Bruno B Sep 15 '25 at 16:24
13

See Uniform convergence of Eigenfunction decomposition on Riemannian sphere? on MO. Unless I’m tripping, I feel like the following is a “softer” argument which only uses Sobolev embedding and elliptic estimates, rather than an explicit Weyl-law type of eigenvalue growth estimate.


Throughout, let $(M,g)$ be a smooth compact boundaryless Riemannian manifold of dimension $n$, and let $A$ be a self-adjoint second-order elliptic differential operator with smooth coefficients (like the Laplacian). Due to compactness of $M$, there are several equivalent ways of describing the function spaces below (e.g via local charts and partitions of unity, or by using the Levi-Civita connection to define these spaces). Now, we record the following estimates.

  • (Sobolev Embedding) For all $\ell\in \Bbb{N}_0$ and all integer $m>\ell +\frac{n}{2}$, there is a $B_{\ell,m,n}>0$ such that for all $f\in C^{\infty}(M)$, we have \begin{align} \|f\|_{C^{\ell}}\leq B_{\ell,m,n}\cdot \|f\|_{H^m}. \end{align} Roughly speaking, we can quantiatively trade-in some super high Sobolev regularity for some uniform pointwise regularity.
  • (Higher Order Elliptic Estimate for $A$) For all $m\in\Bbb{N}_0$, there is a constant $B_{m,n,A}>0$ such that for all $f\in C^{\infty}(M)$, we have \begin{align} \|f\|_{H^{m+2}}\leq B_{m,n,A}\left(\|Af\|_{H^{m}}+ \|f\|_{L^2}\right). \end{align}

Before coming to the main result, let’s introduce some notation. Suppose $\{\phi_j\}_{j=0}^{\infty}$ is an orthonormal basis of $L^2(M)$ consisting of smooth functions which are eigenfunctions of $A$. For each $N\in\Bbb{N}_0$, let $P_N$ be the orthogonal projection of $L^2(M)$ onto the subspace spanned by $\{\phi_0,\dots, \phi_N\}$, i.e for all $f\in L^2(M)$, \begin{align} P_Nf:=\sum_{j=0}^N\langle f,\phi_j\rangle_{L^2}\,\phi_j. \end{align} We then have that for each $f\in L^2(M)$, $\|f-P_Nf\|_{L^2}\to 0$ because the $\phi_j$’s form an orthonormal basis. If $f$ is smooth, then we observe also that $Af$ makes sense, so we can commute $A(P_Nf)=P_N(Af)$.

Theorem.

With notation as above, for each $\ell\in\Bbb{N}_0$ and $f\in C^{\infty}(M)$, we have $\|f-P_Nf\|_{C^{\ell}}\to 0$ as $N\to\infty$.

To prove this, fix $\ell\in\Bbb{N}_0$, and an integer $m$ so large that $2m>\ell+ \frac{n}{2}$. Then, by the Sobolev embedding, and applying the elliptic estimate $m$ times, we see that for all $f\in C^{\infty}(M)$, \begin{align} \|f\|_{C^{\ell}}&\leq \widetilde{B}_{\ell,m,n}\cdot \|f\|_{H^{2m}} \leq B_{\ell,m,n,A}\sum_{j=0}^m\|A^jf\|_{L^2}. \end{align}

Now, apply this same estimate with $f-P_Nf$ instead (which is still a smooth function since the eigenfunctions are smooth), then commute with the $A$’s to get that for all $f\in C^{\infty}(M)$, \begin{align} \|f-P_Nf\|_{C^{\ell}}&\leq B_{\ell,m,n,A}\sum_{j=0}^m\|A^j(f-P_Nf)\|_{L^2}\\ &=B_{\ell,m,n,A}\sum_{j=0}^m\|(I-P_N)(A^jf)\|_{L^2}. \end{align} Since $f$ is smooth, we have that $f,Af,\dots, A^mf$ all lie in $L^2(M)$, so as $N\to\infty$, the RHS vanishes. This shows $P_Nf\to f$ in the $C^{\ell}$-norm, hence completes the proof.


In particular, this shows that for smooth $f$, the eigenfunction expansion is uniform, so if $x,y\in M$ are such that for all $j$, $\phi_j(x)=\phi_j(y)$, then we can indeed perform pointwise evaluations to get for all smooth $f$ that $f(x)=f(y)$. Hence, we deduce $x=y$ (we use bump functions to prove that smooth functions separate points on any manifold, compact or not).

peek-a-boo
  • 69,172