18.155 LECTURE 15, 2015 RICHARD MELROSE Abstract. Notes before and after lecture – if you have questions, ask! Read: Direct proofs of the spectral theorem for compact operators abound – you can find one in my undergraduate Functional Analysis notes for instance. Before lecture • For a compact self-adjoint operator the orthocomplement of the null space has an orthonormal basis of eigenfunctions. Proof: Choose a cut-off χ ∈ Cc∞ (R) with support in [−2, 2] which is equal to 1 in [−, ]. Then 1 = χ− + χ + χ+ and using the functional calculus, A = A− + A0 + A+ for any self-adjoint operator A ∈ B(H). Now A commutes with A+ and so acts on Ran(A+ ) and hence on H+ = Ran(A). The restriction of A+ to H+ is compact and satisfies (Av, v) ≥ kvk2 so is invertible, and hence H+ is finite-dimensional. It follows that H+ has an orthormal basis of eigenvectors for A – look at the critical points of F (u) = (Au, u) which is a real quadratic form (start from the maximum, check it occurs at an eigenvector and pass to its orthocomplement on which A acts). The same is true for the analogous H− so A has a finite-dimensional space spanned by eigenvectors on the orthocomplement of which it has norm at most 2. Now work with = 1/n, if A is not of finite rank then this produces an orthonormal sequence of eigenvectors and on the orthocomplement of the span of these A vanishes. • Conversely, if an operator is such that there is an orthonormal basis of eigenvectors and the sequence of eigenvalues tends to 0 then the operator is compact. (I mean real and self-adjoint, but in general this corresponds to a compact normal operator). • If A = A∗ ∈ B(H) then it has closed range precisely if it is of finite rank. We can define a ‘generalized inverse’ of A, as a linear map B : D −→ H, from D = Nul(A)⊥ ⊕ Ran(A) to Nul(A)⊥ . You can check that this is an example of an unbounded self-adjoint operator. That is, D is dense, (Bu, v) = (u, Bv) for all u, v ∈ D, and (1) D∗ = {w ∈ H; D 3 u −→ (Bu, w) extends by continuity to H} = D. This is what we will find for the Laplacian on the sphere, where D = H 2 (Sn−1 ) and H = L2 (Sn−1 ). s • We identify H s (Sn−1 ) with the subspace of Hloc (Rn \ {0}) which consists of functions (or distributions) which are homogeneous of degree 0; this also works for C ∞ (Sn−1 ), being s = ∞. It also works for s < 0 if we want. • We can replace 0 by any z ∈ C and get the same space. 1 2 RICHARD MELROSE • Check that projective local coordinates identify H s (Sn−1 ) with functions ± s n−1 such that u = (Fj± )∗ u± ) for each j and sign. j , uj ∈ Hloc (R s • The flat Laplacian is homogeneous of degree −2, and maps Hloc (Rn \ {0}) s−2 n to Hloc (R \ {0}) for each s, so if we define 2 ∞ n−1 g ∆ ) S u = |x| (∆ũ), u ∈ C (S (2) then we get a map ∆S : H s (Sn−1 ) −→ H s−2 (Sn−1 ) for each s. • This is a differential operator, in the sense that it is local and is a differential operator in local coordinates. • Suppose we consider uz = rz ũ where ũ ∈ C ∞ (Rn−1 \ {0}) is homogeneous of degree 0. So uz is homogeneous of degree z and ∆rz ũ = rz−2 (−z(z + n − 2) + ∆)) ũ. (3) (4) • Suppose 0 ≤ µ ∈ Cc∞ (Rn \ {0}) is of the form µ = ν(|x|) satisfies Then define Z hu, vi = ũṽµdx, u, v ∈ L2 (Sn−1 ). R µ = 1. As part of the homework for next week check that this is given by an integral Z h(Fj± )∗ u0 , (Fj± )∗ v 0 i = u0 v 0 δdx0 Rn−1 0 0 L2c (Rn−1 ) whenever u , v ∈ depending on j and ±). • If u, v ∈ C ∞ (Sn−1 ) then for a fixed positive smooth function δ (not even h∆S u, vi = hu, ∆S vi is non-negative when u = v. • Suppose u ∈ H s (Sn−1 ) is an eigenfunction of ∆S , so ∆S u = λu for some λ ∈ C, then u ∈ C ∞ (Sn−1 ), λ = k(k + n − 2) for some k ∈ N0 and rk ũ is an harmonic polynomial. After lecture I did the proof of (3) above in lecture; it and the symmetry of ∆S are based on the formula xi 0 f (|x|) in |x| > 0 ∂xi f (|x|) = |x| for f ∈ C ∞ (0, ∞). The proof of the last statement above uses (1) Elliptic regularity for ∆Rn : If u ∈ H 2 (Sn−1 ) satisfies ∆S u = λu then its homogeneous extension satisfies 2 |x|2 ∆Rn ũ = λũ, u ∈ Hloc (Rn ). 2 This can be written ∆Rn ũ = |x|−2 ũ with the right side in Hloc (Rn \ {0}), 4 n so elliptic regularity shows that ũ ∈ Hloc (R \ {0}). Then the left side is in 4 Hloc (Rn \{0}) and proceeding by induction it follows that ũ ∈ C ∞ (Rn \{0}) so u ∈ C ∞ (Sn−1 ). From symmetry (∆S u, u) = (u, ∆S u) so λkuk2 = λ̄kuk2 and if u is nontrivial then λ ∈ R. I should have written out the proof that (∆S u, u) ≥ 0 L15 (5) 3 which implies that λ ≥ 0. In fact this follows by the same sort of integration by parts argument. Namely Z XZ (∆S u, u) = |x|2 ∆Rn ũũµ(|x|)dx = |x|2 ∂xi ũ∂xi ũµ(|x|)dx ≥ 0 Rn (6) i where the missing term is Z XZ 2 (∂xi |x| µ)(∂xi ũ)ũdx = i Rn Rn g(|x|)(x · ∂x ũ)ũdx = 0 Rn by the homogeneity of ũ. Now, once we know that λ ≥ 0 for an eigenvalue of ∆S we can see what these eigenfunctions must be. Namely, if λ ≥ 0 then there is a solution of z(z + n − 2) = λ with z ≥ 0, r n−2 2 n−2 n−2 2 n−2 2 ) =λ+( ) , so z = − + λ+( ) . (z + 2 2 2 2 Now consider |x|z ũ, for an eigenfunction with z above. From (3) it follows that ∆Rn (|x|z ũ) = |x|z−2 (−z(z + n − 2) + λ)ũ = 0. Since z ≥ 0, |x|z ũ is locally integrable across 0 and the equation (1) must actually hold across 0, since the error would have to be homogeneous of degree z −2 and supported at the origin (think more carefully about n = 2). So in fact |x|z ũ must be a polynomial and so z = k is an integer and λ = k(k + n − 2) is also an integer. Thus we have seen that the only eigenfunctions of ∆S are the ‘spherical harmonics’ the restrictions to the sphere of homogeneous harmonic polynomials. Department of Mathematics, Massachusetts Institute of Technology E-mail address: rbm@math.mit.edu