You are on page 1of 4

Eric Lee 21-759

Connections between the graph Laplacian and the manifold Laplace-Beltrami operator Two well-known objects in graph theory and analysis respectively are the Laplacian dened on a graph and the Laplace-Beltrami operator dened on a differentiable manifold. Both of these objects measures the pull of a function away from a point. Let (M, g) be a Riemannian manifold embedded in Rn . It is equipped with a natural volume element, which we use to sample some points uniformly randomly. Between each pair of points we put an edge with a weight, which is large if the points are close together and vanishes quickly if they are not. The sequence of weighted Laplacians of this sequence of graphs, considered as operators on C (M), converges to the Laplace-Beltrami operator on the manifold (scaled appropriately). The Laplace-Beltrami operator on Riemannian manifolds is dened by (1)
M( f )

= div( grad( f ))

In Rn this is
n (2) i=1 2 f 2 dxi

As mentioned before, this intuitively measures how quickly the function f pulls away from each point - the gradient is a vector eld which intuitively tells us how quickly f is changing at each point, and the divergence of a vector eld tells us how quickly a vector eld is pulling each point out of a small sphere centered at that point. The Laplacian of a weighted graph G = (V, E) with weight function w : E R is dened by (3) L( G )ij = n=1 w(k, k ) for i = j, w(i, j) for i = j. k This matrix may be viewed as an operator on functions V R (which may themselves be viewed as column vectors ( f (v1 ), f (v2 ), ...) T . If the vertices of the graph come from a manifold and the weight function wt is given as a (radially symmetric w.r.t. xing one argument and positive) continuous real function on M, we may extend it naturally to an operator on functions in C (M) in the following way. (4) Lt [ f ( x )] = G
1 n n i=1 ( f ( x ) f ( xi )wt ( x, xi )

This operator is the one which we would like to see converge to the Laplace-Beltrami operator on the manifold. In order to prove this, we will need to dene a mediator - a functional approximation to the Laplace-Beltrami operator. (5) Lt [ f ( x )] =
M ( f (x)

f (y))wt ( x, y)d(y)
|| x y||2

From now on we will let wt ( x, y) = e 4t , which should be familiar as the Gaussian in Rn . Note that it makes sense to write || x y|| as we have demanded that M be embedded in Rn . We can now state the main theorem: Let x1 , x2 , ..., xn be sampled from a uniform distribution on a manifold M R. Let tn = n k+2+a where a > 0 and let f C (M). Then the following holds: (6) limn
k tn (4tn ) 2 1

L tn f ( x ) = n

1 vol (M)

f (x)

where the limit is taken in probability and vol (M) is the volume of the manifold with respect to the volume element. Roughly speaking, tn is a measure of how concentrated the weight function is, so sending n to innity, L becomes more and more localized, which is what we want. Various other versions of this theorem exist: for example, given a family of functions F which are uniformly bounded up to the third derivative, we may nd a sequence {tn }nN which forces each Ltn [ f ( x )] f F to converge to M simultaneously. Another generalizaG tion allows for arbitrary probability distributions to be used for sampling points. These proofs are not far off from the proof of the main theorem here, and they can be found in [1]. We will now proceed to the main proof: The strategy here will be as follows. We will rst show that in probability Lt [ f ( x )] G converges to Lt [ f ( x )] up to scaling. Then we will deal with the integral in Lt - the idea is that we would like to change it to an integral in Rn . To do this, we would like to use the exponential map, which is somehow a very natural way to change coordinates. However the exponential map is not generally injective on M, so in order to use it we must reduce the domain of integration to a small open set about x. Having changed to an integral in Rn , we bound terms of the integral until we have the desired result. The expectation of Lt [ f ( x )] is Lt [ f ( x )]. We apply a concentration inequality (HoeffdG T ings inequality) in order to bound the difference between LG and its expectation: we have (7) P[
1

k t(4t) 2

| Lt [ f ( x )] ELt [ f ( x )]| > ] 2e 2 G G


1

1 2 nt(4t)k/2

Choosing t = tn = n k+2+a for a > 0 we have that for any xed

>0

(8) P[ .

k t(4t) 2

| Lt [ f ( x )] ELt [ f ( x )]| > ] = 0 G G

It remains to deal with Lt . We will use three lemmas: Lemma 1. Given B M, p B, and f L (M), for any positive a as t 0 we have (9) |
Be
|| py||2 4t

f (y)d(y)

Me

|| py||2 4t

f (y)d(y)| = o (t a )

This is easy to see from the boundedness of f and the rate at which the Gaussian vanishes. Lemma 2.
M f (x)

Rn

f(0) where f = f exp x

This was given as an exercise. Lemma 3. For any two points x, y M, y = exp x ( p), the Euclidean and geodesic distances are related by (10) || p||2 k ||y x ||2 k = O(|| p||4 k uniformly. R R R The rst lemma tells us that we lose a very small amount from reducing the integral into a neighborhood of x; the second gives us the Laplacian, and the third allows us to dispense with certain terms in the integral in Rn . Let B be an open set about x which is contained in the radius of injectivity of exp x ; we are now dealing with the integral (11)
1 t(4t)k/2

B ( f ( x ) f ( y )) e

|| x y||2 4t

d(y)

Changing coordinates by the exponential map, we have (12)


1 tvol (M)(4t)k/2
B( f(0) f( p))e ||exp x ( p) x ||2 4t

det( gij dp

The metric tensor has the expansion in exponential coordinates (13) det( gij ) = 1 1 x T Rx + O(|||3 ) 6 where R is the Ricci curvature tensor. Since M is smooth and compact, the elements of R are bounded and we have det( gij ) = 1 + O(|| x ||2 ). Together with Lemma 3 we have 1 (14) tvol (M)(4t)k/2 || x ||2 g( x )
B( f(0) f( p))e
4t

(1+O(|| x ||2 )dp

Expanding f into its Taylor polynomial f( x ) f(0) = x we have 1 (15) tvol (M)(4t)k/2 || x ||2 g( x )
B( x f + 1 x T ( Hess[ f ]) x +O(|| x ||3 )e 2

4t

1 f + 2 x T ( Hess[ f ]) x + O(|| x ||3 )

(1+O(|| x ||2 )dp

We throw away the g in the exponential, since its very small near x, and we also throw away the O(|| x ||2 ). Since the Gaussian is radially symmetric we have 3
B

xi e

|| x ||2 4t

= 0 and

for i = j, B xi x j e 4t = 0. What remains is the diagonal terms in the Hessian, which yield exactly the Laplace-Beltrami operator, and the argument is nished. The reader interested in the other versions of this theorem mentioned earlier may read [1], Towards a Theoretical Foundation for Laplacian-Based Manifold Methods; an application for manifold learning may be found in [2]. [1] Belkin, M., Niyogi, P. Towards a Theoretical Foundation for Manifold-Based Learning Techniques, Journal of Computer and System Sciences, Volume 74 , Issue 8, pp. 12891308, 2008. [2] Belkin, M., Niyogi, P. Laplacian Eigenmaps for Dimensionality Reduction and Data Representation, Neural Computation, June 2003; 15 (6):1373-1396.

|| x ||2

You might also like