Introduction to computational quantum mechanics Lecture 5: The time independent Schrödinger equation and Ritz-Galerkin approach Simen Kvaal simen.kvaal@cma.uio.no Centre of Mathematics for Applications University of Oslo Seminar series in quantum mechanics at CMA Fall 2009 Outline Setting and formal results Ritz-Galerkin/FCI FCI: Ritz-Galerkin with Slater determinants References Outline Setting and formal results Ritz-Galerkin/FCI FCI: Ritz-Galerkin with Slater determinants References Hamiltonian, Hilbert-space and basis In this lecture, we will use the following conventions: I We consider N particles in d dimensions, i.e.: H = Π− L2 (Rd × S)⊗N We usually ignore spin S; it does not modify results. I The Hamiltonian H is given by N H= ∑ H0,k + ∑ k=1 I Vij (ij), i6=j The most general H0 is given by 1 H0 = − ∇2 + W(x), 2 ignoring irrelevant constants. W(x) smooth The time-independent Schrödinger equation I Suppose the spectral decomposition of H is available: Z ∞ E dP(E) = H= −∞ I ∑ Ek |ψk i hψk | + Z Ek ∈σd (H) σc (H) E |χ(E)i hχ(E)| dE Then, the time evolution is trivial to compute: Z ∞ U(t) = exp(−itH) = e−itE dP(E). −∞ I This – among other things – motivates the study of the eigenvalue problem for H, i.e., finding σ(E), ψk and χ(E), formally: Hψ = Eψ. The spectrum of H: HVZ theorem What can be said about σ(H) a priori? The famous HVZ theorem (after Hunziker, van Winter and Zhislin who first proved it) says a lot: Theorem Suppose Vij = V(xi − xj ), as operator on L2 (Rd ), is such that Vij (−∇2 + 1)−1 is compact Then the essential/continuous spectrum is σc (H) = [Σ, ∞). Let a be a partition of {1, . . . , N} into disjoint subsets. The number Σ is given by: Σ = min Σ(A) , #a≥2 where Σ(a) = inf σ(H(a)), with H(a) = H − ∑ (ij)6⊂a Vij . Illustration Figure: Illustration of geometric idea behind HVZ theorem. As clusters of particles move far away, the interactions between the remaining ones vanish. σc (H) “should” contain points corresponding to clusters moving “slowly away” from the origin The spectrum of H: discrete spectrum What can we say about σd (H)? Several possibilities: I Infinitely many isolated eigenvalues < Σ, accumulating at Σ (e.g., if Vij = 1/kxi − xj k) I Finitely many isolated eigenvalues < Σ (e.g., if Vij is smooth with compact support) I No eigenvalues at all, e.g., if Vij ≥ 0 (repulsive interactions) I Eigenvalues embedded in σc (H); “resonances” I No eigenvalues of infinite multiplicity (in this Hamiltonian!) The spectrum of H: purely discrete spectrum The Hamiltonian may also have a purely discrete spectrum: Theorem Let V : RNd → R be smooth and unbounded as kxk → ∞. Then the Hamiltonian 1 H = − ∇2 + V(x) 2 has a purely discrete spectrum on the form: σ(H) = σd (H) = {Ek }∞ k=1 , (Note that ∑k ∇2k = ∇2 .) E1 ≤ E2 ≤ · · · % +∞ Outline Setting and formal results Ritz-Galerkin/FCI FCI: Ritz-Galerkin with Slater determinants References Pet child has many names . . . We will describe an approximation method for σ(H) called . . . I Ritz-Galerkin/Rayleigh-Ritz(-Galerkin); especially in finite element (FE) community; continuum mechanics and studies of structural vibrations I Full configuration interaction (FCI): chemistry, nuclear physics I Exact diagonalization: quantum dots, solid state physics I Projection method, finite section method: Mathematical spectral approximation theory Variational characterization of σ(H) Ritz-Galerkin/FCI is based on the following well-known variational characterization of the eigenvalues below σc (H): Theorem Suppose H = H ∗ is bounded from below. Define a sequence {(φk , µk )}, k = 1, · · · by: µk = min hψ|Hψi =: hφk |Hφk i , ψ∈Mk where Mk = {ψ ∈ D(H) : kψk = 1, hψ|φj i = 0, for j < k.} Then for every k exactly one of the following holds: 1. There is j > k with µj > µk , and µk ∈ σd (H). Also, φk is an eigenvector with eigenvalue µk . 2. µj = µk for all j > k, and µk = inf σc (H) (the bottom of the continuous spectrum.) There are then only finitely many eigenvalues below σc (H). Ritz-Galerkin approach I I Suppose {ej } is an orthonormal basis for H , such that ej ∈ D(H) for all j. Consider the space Hn = Pn H , with n Pn := ∑ |ej i hej | (orth. projection) j=1 I Consider the compression of H onto Hn , i.e., n Hn := Pn H|Hn = ∑ n j |Hek i hek | ∑ |ej i he | {z } j=1 k=1 mat. elms. I The spectrum σ(Hn ) = σ(H) with Hjk = hej |Hek i ; I n × n matrix. One hopes that, in some sense, σ(Hn ) −→ σ(H), as n → ∞ Convergence of Ritz-Galerkin Recall the variational characterization: µk = min hψ|Hψi =: hφk |Hφk i , ψ∈Mk where Mk = {ψ ∈ D(H) : kψk = 1, hψ|φj i = 0, for j < k.} (n) I The eigenvalues µk of H are obtained by replacing D(H) with H\ . I Thus, we immediately obtain: (n) µk ≤ µk , I for all k ≤ n and n. Conclusion: The eigenvalues are approximated from above. If (n) #σd (H) = m < ∞ (at least), the n − m remaining µk converges to Σ = inf σc (H). Convergenve in Hausdorff metric But we can say more: Theorem If H = H ∗ is bounded from below and we have have σc (H) = [Σ, ∞) or empty , then σ(H) → σ(H) in Hausdorff metric. Convergence in Hausdorff metric: illustration Figure: Convergence of discrete spectrum in Hausdorff metric, meaning that the convergence is ordinary for the discrete points in σ(H), while the other discrete eigenvalues fill up the continuous spectrum, with smaller and smaller distance among them. Simple example: one-particle quantum dot I We consider a simple d = 1 problem, N = 1 particle. I Hamiltonian: H=− I 1 ∂2 + W(x); 2 ∂x2 2 /2σ2 Basis functions: ej (x) = Nn Hn (x)e−x I W(x) = −V0 e−x 2 /2 (Hermite functions) Matrix elements evaluated using Gaussian quadrature, matrix diagonalized in Matlab for an increasing number of basis functions. Simple example: plot of eigenvalues Convergence history of the approximate eigenvalues 10 8 6 µ(n) k 4 2 0 −2 −4 −6 −8 0 50 100 150 n 200 250 (n) Figure: Convergence with increasing n of eigenvalues µk towards µk . Seemingly, σc (H) = [0, ∞), which in fact can be proven. Note extremely rapid convergence in this case for the bound states, i.e., for σd (H). Outline Setting and formal results Ritz-Galerkin/FCI FCI: Ritz-Galerkin with Slater determinants References Brief outline 1. Given an N-body Hamiltonian H/Hilbert space HN 2. Use subset of Slater-determinant basis BN 3. Construct matrix 4. Diagonalize to obtain approximation to σd (H). Figure: Typical structure of an N = 3 matrix A little repetition does little harm I N-particle Hilbert space: HN = Π− H1⊗N = H1 ∧ · · · ∧ H1 {z } | N factors I Single-particle orthonormal “orbitals”: H1 = Span{φj : j = 1, 2, · · · } I Slater determinants constitute a basis for HN : ΦSD j1 ,j2 ,··· ,jN (x1 , x2 , · · · , xN ) = φj1 ∧ φj2 ∧ · · · ∧ φjN I We could also express these using creation operators: ΦSD = c†j1 · · · c†jN Φ0 , c†j : HN → HN+1 , where Φ0 is a trivial Slater determinant with zero particles. FCI Slater determinant subset I The complete basis: BN = ΦSD j1 ,··· ,jN : j1 < j2 < · · · < jN I Introduce a cut parameter L and consider: BNL := ΦSD j1 ,··· ,jN : ji ≤ L ⊂ BN I There are of course other ways to truncate the basis, but this is the most common. I Note: #BNL L = N ⇒ ”Curse of dimensionality” More repetition I The Hamiltonian can be written: H = ∑ hα,β c†α cβ + αβ I 1 ∑ vαβγδ c†α c†β cδ cγ 2 αβγδ Here, hαβ is a “one particle matrix element”: hαβ = hφα , H0 φβ i , (inner product in one-particle space) and vαβγδ is a “two-particle matrix element”. (See earlier lecture for more) Building matrix elements Recall two basic facts: Theorem Even though N-body Hilbert space consists of Nd-dimensional functions, the matrix elements can be computed solely with low-dimensional integrals. Theorem The matrix obtains a sparsity structure due to the Slater–Condon rules: The Hamiltonian only “changes two single-particle functions” when acting on ΦSD . Therefore, matrix element between Slater determinants are zero whenever they have > 2 different indices ji A convergence result for quantum dot systems We have considered the fact that FCI does converge as L increases. What about the speed of convergence? I Depends on the analytic properties of the eigenfunction ψk . Typical question: “How smooth is it?” I This in turns depends on the Hamiltonian H. I The speed of convergence with respect to L depends on the approximation properties of φj , the single-particle functions I For electronic problems, Vij ∼ 1/kxi − xj k, and typically EkL − Ek ≤ C1 L−ν1 , withν1 > 0. I In nuclear physics, Vij is, in fact, smooth, and we obtain EkL − Ek ≤ C2 e−ν2 L , withν2 > 0. Outline Setting and formal results Ritz-Galerkin/FCI FCI: Ritz-Galerkin with Slater determinants References References Simon, B. Schrödinger operators in the 20th century J. Math. Phys 41, p. 3523 2000 Hansen, A.C. On the approximation of spectra of linear operators on Hilbert spaces J. Func. Anal. 254, pp. 2092–2126 2008 Babuska, I. and Osborn, J.E. Finite Element-Galerkin Approximation of the Eigenvalues and Eigenvectors of Selfadjoint Problems Math. Comp. 52, pp. 275–297 1989 S.K. Analysis of many-body methods for quantum dots PhD thesis 2009