Singular Value Decomposition: Reduced Density Matrix
We will introduce the singular value decomposition of a matrix in the context of the reduced density matrix of a quantum system connected to an environment. REDUCED DENSITY MATRIX1 Let us consider a quantum system (block) B, which is spanned by the N-dimensional orthonormal basis set { i | i = 1,...,N}, surrounded by an environment E, which is spanned by the M-dimensional orthonormal basis set { j | j = 1,...,M} (see the figure below).
The ground state of the total (= block + environment) system can be represented as
N M
= ij i j .
i=1 j=1
(1)
Now consider the expectation value of an arbitrary operator, A, which acts only within the block:
* A = ij j i A i j i j i j i * i j ij i j j
=
i j
i A i j j
, (2)
=
i i
* i jij i j i
i A i
i i Aii = trB ( A)
where the reduced density matrix is defined as
* i i i jij , j
(3)
and the matrix element of the operator is Aii i A i . SINGULAR VALUE DECOMPOSITION (SVD) Problem: What is the optimal reduced density matrix of rank-m (<< N)? Solution: Singular value decomposition (SVD) of RNRM. (Theorem) An NM matrix (assume N M) can be decomposed as
or
1
d1 dM
VT
(4)
R. P. Feynman, Statistical Mechanics (Benjamin/Cummings, Reading, MA, 1972) Chap. 2.
= UDV T ,
where U = Ui = ui( ) RNRM is column orthogonal, i.e.,
(5)
ui( ) ui( ) = ,
i=1
(6)
or
U T U = IM , RMRM is column orthogonal, i.e.,
(7)
and V = Vi = vi( )
vi( )vi( ) = ,
i=1
(8)
or (9) V T V = IM . The columns of U, whose same-numbered elements d are nonzero, are an orthonormal set of basis vectors that span the range (see Appendix for the range); the columns of V, whose same-numbered elements d are zero, are an orthornormal basis for the nullspace that is mapped to zero, i.e., the subspace of x RM, where x = 0. The program, singular.c, in the source code directory of the class home page demonstrates this automatic construction of orthonormal bases for the range and the nullspace. TRUNCATED SVD AS OPTIMAL APPROXIMATION (Theorem) Let = UDV T be the SVD of with the diagonal elements in descending order d1 d2 ... dM and let
(m ) u ( ) d v ( )T ,
=1
(10)
be the rank-m truncation of the SVD. Then
min A rank(A)=m
2
= (m)
= dm +1 ,
2
(11)
where the matrix 2-norm is defined in terms of the vector 2-norm as A
Therefore, (m) is the optimal rank-m approximation to .
x R
min
M
) 2 =1
Ax R N
) 2.
Equation (10) shows that SVD is a representation of a matrix as a sum of outer products of two vectors, just as a density matrix is. LOW-RANK APPROXIMATION TO THE REDUCED DENSITY MATRIX Substituting the rank-m approximation (10) in the definition of the reduced density matrix, Eq. (3),
= T
= u ( ) d v ( )T v ( ) d u ( )T = u ( ) d ( ) d u ( )T
=1 =1 m 2 = u ( ) d u ( )T =1 =1 =1 m m
m m
(12)
(Summary) The rank-m truncation of the SVD of the global (= block + environment) ground state wave function, m (m ) = u ( ) d v ( )T , (13)
=1
or
(m ij ) = ui ( ) d v j ( ) ,
(14)
=1
produces the rank-m approximation to the reduced density matrix,
or
(m ) = u ( ) w u ( )T ,
=1
(15)
i(m ) = ui ( ) w ui ( ) , i
=1
(m)
(16)
where w = d . The rank-m approximation is optimal in the least square sense. DENSITY MATRIX RENORMALIZATION GROUP The density matrix renormalization group (DMRG) algorithm by Steven White2 is a systematic procedure to accurately obtain a quantum ground state with a modest computational cost. The DMRG incrementally add environments to the block, solve the global (= block + environment) ground state, and construct a low-rank block density matrix to represent the block with a reduced degrees of freedom. APPENDIXRANK OF A MATRIX For an NM matrix A, consider the mapping,
x R M A b = Ax R N .
(A1)
The range of matrix A is the vector space spanned by all linearly independent vectors {b}, which are mapped from some x. The rank of matrix A is the size (i.e., the number of linearly independent vectors) of its range.
S. R. White, Density-matrix algorithms for quantum renormalization groups, Physical Review B 48, 10345 (1993).
