3. DMRG algorithm

The density matrix renormalization group (DMRG) was first used for ab initio quantum chemistry in 1999 [DMRG1]. The method variationally optimizes a low-rank tensor approximation of the full configuration interaction (FCI) solution. Suppose we have \(L\) spatial orbitals. The FCI solution can in general be written as

\[\begin{split}\left|\Psi\right\rangle & = & \sum\limits_{\{ n_{i\sigma} \}} C^{n_{1\uparrow} n_{1\downarrow} n_{2\uparrow} n_{2\downarrow} n_{3\uparrow} ... n_{L\uparrow} n_{L\downarrow} } \left( \hat{a}^{\dagger}_{1\uparrow} \right)^{n_{1\uparrow}} \left( \hat{a}^{\dagger}_{1\downarrow} \right)^{n_{1\downarrow}} \left( \hat{a}^{\dagger}_{2\uparrow} \right)^{n_{2\uparrow}} ... \left( \hat{a}^{\dagger}_{L\uparrow} \right)^{n_{L\uparrow}} \left( \hat{a}^{\dagger}_{L\downarrow} \right)^{n_{L\downarrow}} \left|-\right\rangle \nonumber \\ & = & \sum\limits_{\{ n_{i\sigma} \}} C^{n_{1\uparrow} n_{1\downarrow} n_{2\uparrow} n_{2\downarrow} n_{3\uparrow} ... n_{L\uparrow} n_{L\downarrow} } \left| n_{1\uparrow} n_{1\downarrow} n_{2\uparrow} n_{2\downarrow} n_{3\uparrow} ... n_{L\uparrow} n_{L\downarrow} \right\rangle.\end{split}\]

With successive singular value decompositions, the FCI \(C\)-tensor can be composed into a matrix product state (MPS):

\[\begin{split}C^{ n_{1\uparrow} n_{1\downarrow} n_{2\uparrow} n_{2\downarrow} n_{3\uparrow} ... n_{L\uparrow} n_{L\downarrow} } & = & \sum\limits_{\alpha_1, \alpha_2, ..., \alpha_L} A[1]^{ n_{1\uparrow} n_{1\downarrow} }_{ \alpha_1 } A[2]^{n_{2\uparrow} n_{2\downarrow}}_{ \alpha_1 ; \alpha_2 } A[3]^{n_{3\uparrow} n_{3\downarrow}}_{ \alpha_2 ; \alpha_3 } ... A[L-1]^{n_{L-1\uparrow} n_{L-1\downarrow}}_{ \alpha_{L-2} ; \alpha_{L-1} } A[L]^{n_{L\uparrow} n_{L\downarrow}}_{ \alpha_{L-1} } \nonumber \\ & = & \mathbf{A}[1]^{ n_{1\uparrow} n_{1\downarrow} } \mathbf{A}[2]^{n_{2\uparrow} n_{2\downarrow}} \mathbf{A}[3]^{n_{3\uparrow} n_{3\downarrow}} ... \mathbf{A}[L-1]^{n_{L-1\uparrow} n_{L-1\downarrow}} \mathbf{A}[L]^{n_{L\uparrow} n_{L\downarrow}},\end{split}\]

where \(dim(\alpha_i) = min(4^i,4^{L-i})\). To make the method of polynomial complexity, the rank of the decomposition is truncated to a fixed dimension \(D\):

\[dim(\alpha_i) = min(4^i,4^{L-i},D).\]

The integer \(D\) is called the bond, virtual, or auxiliary dimension. The DMRG algorithm consists of consecutive sweeps over the chain of orbitals, during which two neighbouring MPS tensors are variatonally optimized. Thereto they are combined into a two-orbital tensor:

\[\mathbf{B}[i]^{n_{i\uparrow} n_{i\downarrow} n_{i+1\uparrow} n_{i+1\downarrow}} = \mathbf{A}[i]^{n_{i\uparrow} n_{i\downarrow}} \mathbf{A}[i+1]^{n_{i+1\uparrow} n_{i+1\downarrow}}.\]

The Lagrangian

\[\mathcal{L} = \left\langle \Psi( \mathbf{B}[i] ) \mid \hat{H} \mid \Psi( \mathbf{B}[i] ) \right\rangle - E \left\langle \Psi( \mathbf{B}[i] ) \mid \Psi( \mathbf{B}[i] ) \right\rangle\]

is varied with respect to \(\mathbf{B}[i]\) to yield an effective Hamiltonian eigenvalue equation. By exploiting the gauge freedom in the MPS, this eigenvalue equation can always be turned into a numerically stable standard eigenvalue equation for each local optimization step:

\[\mathbf{H}^{\text{effective}}[i] \times \mathbf{B}[i] = E \mathbf{B}[i].\]

Once \(\mathbf{B}[i]\) is found, it is decomposed with a singular value decomposition:

\[B[i]^{n_{i\uparrow} n_{i\downarrow} n_{i+1\uparrow} n_{i+1\downarrow}}_{\alpha;\beta} = M_{(\alpha n_{i\uparrow} n_{i\downarrow});(n_{i+1\uparrow} n_{i+1\downarrow} \beta)} = \sum\limits_{\kappa} U_{(\alpha n_{i\uparrow} n_{i\downarrow});\kappa} \lambda_{\kappa} V^{\dagger}_{\kappa;(n_{i+1\uparrow} n_{i+1\downarrow} \beta)} = \sum\limits_{\kappa} A[i]^{n_{i\uparrow} n_{i\downarrow}}_{\alpha;\kappa} \lambda_{\kappa} A[i+1]^{n_{i+1\uparrow} n_{i+1\downarrow}}_{\kappa;\beta}.\]

For a normalized wavefunction \(\sum\limits_{\kappa} \lambda_{\kappa}^2 = 1\). If \(dim(\alpha) = dim(\beta) = D\) then \(dim(\kappa) = 4D\). In order to keep the virtual dimension fixed to \(D\), the summation over \(\kappa\) is truncated to the \(D\) largest values \(\lambda_{\kappa}\). The discarded weight \(w_D[i] = \sum\limits_{\kappa > D} \lambda_{\kappa}^2\) is a measure for the information loss.

For more information on the DMRG method, please read Ref. [DMRG2].

DMRG1

S.R. White and R.L. Martin, Journal of Chemical Physics 110, 4127 (1999), doi: 10.1063/1.478295

DMRG2

19. Wouters and D. Van Neck, European Physical Journal D 68, 272 (2014), doi: 10.1140/epjd/e2014-50500-1