PHYSICAL REVIEW B

VOLUME 58, NUMBER 23

15 DECEMBER 1998-I

Energy renormalization-group method for electronic structure of large systems

Roi Baer and Martin Head-Gordon
Department of Chemistry, University of California, Berkeley, California 94720 and Chemical Sciences Division, Lawrence Berkeley National Laboratory, Berkeley, California 94720 Received 12 June 1998 A renormalization-group method for electronic structure of large systems within a tight-binding framework is presented. A telescopic series of nested Hilbert spaces is constructed, having exponentially decreasing dimensions and electrons. The Hamiltonian matrices have exponentially converging energy ranges focusing to the Fermi level. The computational effort scales near linearly with system size even when the density matrix is highly nonlocal. This is illustrated by calculations on a model metal and a metallic nite carbon nanotube, for which standard linear scaling methods are inapplicable. S0163-1829 98 02644-7

New linear scaling methods for electronic structure of very large systems114 exploit the sparsity of the one electron density matrix DM in an R-space-localized basis. However, this is also their Achilles heel, since in many systems of biological, chemical, or physical importance the DM has long-range correlations. Metallic compounds at zero temperature exhibit no DM sparsity and small gap systems like semiconductors have very large DM correlation lengths.1518 We present a method that, unlike previous linear scaling techniques, enables near linear scaling calculations even when the DM is not sparse in R space. Our method is based on a renormalization-group RG point of view, where longranged correlations are described in varying length and energy scales.19 The numerical RG approach presented here, called the energy renormalization group ERG , is radically different from conventional real-space RG in which a super block is constructed from a given number of blocks in real space.20 In ERG, it is the energy that is rescaled by a constant factor at each step. This overcomes the problems of block boundary conditions noted by White and co-worker.21,22 We use Chebyshev expansion methods23 for calculating the FermiDirac FD matrix7,11,18 and our method for performing a renormalization step is inspired by the lter diagonalization of Neuhauser and co-worker.24,25 The electronic structure problem we concentrate on is a tight-binding TB calculation of ground-state GS expectation values, i.e., the energy or forces. We assume an R-space basis of localized functions a(r). The sparse overlap matrix S ab a b denes orthonormal basis functions

and is represented by the DM ab . The Fermi level a b in Eq. 1 is determined by the condition tr N e , where 2N e is the number of electrons. The GS energy is E tr H . 2

The FD density operator approximates : F F H 1 e 1

H

. Unlike , which is nonanalytical, F can by lim F be represented by a uniformly convergent Chebyshev series7,11 of length P:
P 1

F H
n 0

an

Tn H ,

where T n ( ) are polynomials and a n expansion coefcients. For precision 10 D , the expansion length is17 P
2 3

D 1

E,

where E (E max Emin)/2 is the spectral range of the Hamiltonian matrix (E min and E max are the smallest and largest eigenvalues of the Hamiltonian matrix . When satises /2 D ln 10 , 5

where is the HOMO-LUMO gap, the deviation of the expansion from the Heaviside function is conned to the gap, so the DM and FD-DM are identical to precision 10 D . 1 At high-enough temperatures ( ), the range of the ED-DM depends on the temperature, approximately as17 R F . 6

a r
b

1/2

ab b

with which the Hamiltonian H is represented by the sparse . Using the Heaviside function matrix H ab H b a , the one-electron density operator is a function of the Hamiltonian: H

1
PRB 58

Equations 4 and 5 are the source of trouble for current linear scaling methods. As explained in Ref. 17, the numerical work scales as N e ( E/ ) d/2 1 for the Chebyshev method and N e ( E/ ) d for the minimization methods,4,5 where d is the dimensionality of the nuclear lattice. These methods become computationally prohibitive as the gap diminishes. The approach we present is much less sensitive to E/ .
15 296 1998 The American Physical Society

0163-1829/98/58 23 /15296 4 /$15.00

PRB 58

BRIEF REPORTS

15 297

As mentioned above, the DM is the zero temperature limit of the FD-DM. As the temperature in the FD-DM is lowered, the range of matrix correlations increases see Eq. 6 . The ERG method overcomes this difculty by writing the DM as a sum of terms instead of a limit. We show that each term in the series represents an increasingly lower temperature correction in an increasingly coarse-grained space spanned by a smaller number of basis functions. Each term is sparse even though it describes longer ranged correlations. Since the systems we treat are nite, the nal term in the series is a nonsparse contribution, which however can be controlled by the choice of initial temperature and is independent of the original system size. We start by writing the density operator as a telescopic series with temperature decreasing by a factor q 1: F F
0

F 1

2 ,

where 0 is an initial inverse temperature parameter and n n q 0 . From Eq. 6 , F ( 0 ) is strongly localized if 0 is chosen small. Long-range DM correlations that are quenched by 0 are then introduced by cooling the system successively by the renormalization factor q. This yields DM corrections n (n 1,2, . . . ), which are increasingly delocalized. Notice however that n ( ) F n ( ) F n 1 ( ) is only nonzero within an energy of order n 1 1 around the Fermi level. Hence, the n corrections can be represented in Hilbert spaces of decreasing dimensionality. Thus, the cooling by a factor q also allows us to renormalize the energy scale downwards by the same factor. The GS energy of Eq. 2 , is now E tr H F tr H 1 tr H 2 . 8

The matrix G (n) (n) G (n) (n) where the N (n) basis a b a,b functions of the nth space are (n) (r) r (n) can be cala a culated column by column using a length P Chebyshev expansion in the same way the FD-DM columns are computed but with different expansion coefcients . G (n) is highly singular, and its rank is the dimension of the (n 1)th space N (n 1) . The DOS at the Fermi level tr G (n) serves to estimate N (n 1) as M n 1 tr G (n) . A precise method for determining N (n 1) will be given below. Note an important difference between the ERG method and traditional lattice RG. The dimension of the nested space is not imposed, it is determined by the DOS of the system. Since G (n) is obtained by a nite number of Hamiltonian operations, every column of G (n) represents a localized funca,b tion b G (n) (n) (r). Thus, to create a basis for the (n b ab 1)th Hilbert space, we select N (n 1) columns of G (n) and a,b arrange them in a rectangular N (n) N (n 1) matrix, designated G (n) . On the average, one column is selected out of ba every N (n) /N (n 1) columns representing spatially close functions according to an algorithm we describe below. The (n 1)th space is spanned by the functions a
n 1 b

G ba , n

and the new Hamiltonian matrix is H

n 1

n 1

1/2

n 1

1/2

, 10

where S (n 1) G (n) T G (n) is the new overlap matrix. H (n 1) can be calculated without explicitly constructing S (n 1) 1/2 which is less sparse by the following iterations:26 Bk Ak with
(n) T 1 1

B k 3 B 2 /4, k 11

Crucially, the increasingly nonlocal contributions due to n can be evaluated in increasingly small, coarse-grained subspaces, spanned by increasingly less localized basis functions. Furthermore, because of the reciprocal changes in increasing by q and the energy scale decreasing by q on each step, it follows from Eq. 4 that each n can be calculated by a Chebyshev expansion of the same length P. The procedure terminates with a nal explicit diagonalization when the subspace size is small enough to permit this. The crucial step of the ERG approach is the construction of reduced-matrix representations for the (n 1)th Hilbert space out of those known in nth space. The basic idea is much inuenced by lter diagonalization.24 Dene the Fermi-level density of states DOS (n 0,1, . . . ): G
n

Ak

A k B k B k A k /2 B k A k B k /4

e
n

n H n H

A0 k 0,1, . . . , B 0 ( S (n 1) 1), (n) (n) 1 G H G , and N (n 1) maxi,j S(n 1) i,j . The matrices of Eq. 11 converge quadratically B k 0, A k H (n 1) within order 10 iterations depending on the condition number of S (n 1) . We now describe how G (n) is constructed from G (n) . Viewing G (n) as a set of N (n) columns, G (n) is a subset of size N (n 1) . Our algorithm determines the subset and its size in a manner that yields Cond(G (n) ) 10D/2 the condition number of G (n) , where 10 D is the precision. This ensures Cond(S (n 1) ) 10D , which is the minimal value needed to represent the nested space to precision 10 D . The algorithm, based on Ref. 27, and of numerical complexity O(N (n) ) is as follows.
(n)

1 e

When G (n) operates on any vector in the nth space, it zeros components of energy outside the interval M n 1, 1 M n M is a number of order 10 . Thus, G (n) is a pseudoprojector onto the (n 1)th space.

1 For all k(k 1, . . . ,N (n) ), calculate the kth column of G denoted v k . If k 1 the rst column add it to G (n) ; if k 1 perform steps 26 to update G (n) . 2 Identify all the columns that are currently in G (n) and that have nonzero overlap with v k . Assume L 1 such columns exist L is independent of system size, related only to the range of the basis functions .

15 298

BRIEF REPORTS TABLE I. Renormalized parameters at the different iterations for a 16 000-site 1D metal. Iteration Space dimensions No. of electrons Energy range Inv. temperature Hamiltonian range FD-DM range Energy contribution n N (n) N e (n) En
n

PRB 58

0 16 000 16 000 2.00 15 3 132 5884.48

1 1692 1660.6 0.31 94 73 165 32.75

2 298 265.5 0.06 531 55 120 1.23

Diag. 89 36.3 0.01

R n (H) R n (F) tr H ( n ) ( n )

36.42

3 Arrange the columns and v k in a matrix C of dimensions N (n) L and calculate the L L matrix S C T C. 4 Determine the singular value decomposition of S: S Uw v T , where v and U are orthogonal matrices and w is diagonal with elements w 1 w 2 w L 0. Locate the largest integer r the rank with w 1 /w r 10D . 5 Zero the last L-r rows of the matrix v T , yielding a new matrix v and perform a QR factorization with pivoting: QR, where is a permutation of columns. v 6 Replace the L-l columns of G (n) from which C was built by the rst r columns of C . The dimension N (n 1) of the nested space is the nal number of columns in G (n) . The number of electrons in this space is N en
1

iteration. The method is most effective when high DOS peaks are situated away from the Fermi level. The numerical work as a function of system size N scales near linearly as N ln N, the factor ln N reects the increase in number of iterations. ERG was applied to a half-lled one-dimensional 1D single-band metal, whose Hamiltonian is H 2
i

ai ai

ai a j

i, j nearest neighbors

N en

tr F q n .

12

N (n) may be nonintegral if is inaccurately known. e This concludes the denition of one renormalization step. With each iteration the dimension of the matrices decreases and the energy range collapses to the Fermi level . For a nite system the computation terminates after f iterations, when the dimension N ( f ) is small enough so the eigenvalues (f ) of H ( f ) can be determined by numerical diagonalizai ab tion. The series of Eq. 8 is terminated after f steps, where the last term is dened by other terms were dened above
N f

tr H f

i 1

F ,

f i

f
f 1,

f i

13

In this expression, is the initial guess for the Fermi level, which was used throughout the calculation. The quantity is a corrected Fermi level adjusted, together with so the following relation is maintained:
N f i 1 N f

F ,

f i

Ne

f i 1

f
f, 0

14

is taken much larger than the inverse highest occupied molecular orbitallowest unoccupied molecular orbital HOMO-LUMO gap known after the diagonalization but otherwise arbitrary. Since the range of the FD-DM is roughly proportional to 1/2 , the numerical work in the ERG method is proportional to P d/2 1 see Ref. 17 , where d is the number of extended spatial dimensions of the nuclear lattice. P should be chosen small, yet it must be large enough to achieve a signicant reduction in the number of Hilbert-space dimensions at each

with N e N electrons. We treat several system sizes N. In all calculations the initial guess for the Fermi energy is 2.01 and diagonalization is applied once N (n) 100. The Chebyshev expansion length is P 80 terms implemented to a precision D 5 using a tree code.18 Values of parameters for different levels of the computation are shown for the case of N 16 000 in Table I. The energy renormalization factor is q 6. The length renormalization factor is 10 in the rst iteration dropping to 6 and 3 in following iterations. This is due to higher DOS far from the Fermi level. The range R in Table I is dened as the number of elements with magnitude greater than 10 D . R(F) is the range of the FD-DM in renormalized length units and R(H) is the range of interactions. After the rst iteration both ranges are roughly constant. Notice in Table I the large correction to the energy in the nal level, due to inaccurate initial guess of the Fermi level. The energy per site is determined to accuracy of 2 10 5 . Performance of the ERG method as a function of system size is shown in Fig. 1. CPU time scales quasilinearly, approximately as N(ln N)2 the additional factor of ln N is due to the tree-code implementation18 . Crossover is at N 4000 for this single processor Cray C90 256MW calculation. 60% of CPU time was consumed by the rst iteration, sufcient to converge insulators with a gap 1 Eq. 5 with 15 . As a more realistic example, the ERG method was applied to 3,0 carbon nanotubes of varying lengths, predicted to be metals,28 using a TB Hamiltonian.29 Nuclear position was not optimized C-C distance was set to 1.55 and localcharge neutrality was not imposed in these calculations thus, these calculations constitute a rst iteration in the selfconsistent-eld procedure . The Chebyshev length used was P 45. Results are shown in Fig. 2 for D 4 accuracy. The ERG method becomes more efcient than diagonalization at 320 carbon atoms for these calculations on a 175-MHz DEC alpha workstation. Again, the calculation complexity scales approximately as N(ln N)2.

PRB 58

BRIEF REPORTS

15 299

FIG. 1. ERG D 5, dots and standard diagonalization triangles CPU time for an energy calculation vs number of sites in a TB model of a metal. Line depicts ideal asymptotic N(ln N)2 scaling.

FIG. 2. ERG D 4, dots and diagonalization triangles CPU time for an energy calculation vs number of atoms in a metallic 3,0 carbon nanotube. Line depict the ideal asymptotic N(ln N)2 scaling.

The Chebyshev expansion length within the ERG method, of P 45 terms, should be compared to an expansion length of about P 30 000 that would have been needed to resolve the HOMO-LUMO gap using naively the conventional Chebyshev scheme.7,18 Because the DM is highly nonlocal in these metallic systems, similar difculties are also expected in the direct minimization methods.4,5 We have presented a computational method which enables near-linear scaling in previously inaccessible metallic

and small gap systems. The expectation value of any operator in the electronic GS can be written in a form analogous to Eq. 8 , and thus be computed by tracing its nested space matrices Eq. 11 with corresponding (n) matrices. This work was supported by the Laboratory Directed Research and Development Program of Lawrence Berkeley Laboratory, US DOE Contract No. DE-AC03-76SF00098. M.H.G. thanks the Packard Foundation for nancial support.

W. T. Yang, Phys. Rev. Lett. 66, 1438 1991 . G. Galli and M. Parrinello, Phys. Rev. Lett. 69, 3547 1992 . 3 F. Mauri, G. Galli, and R. Car, Phys. Rev. B 47, 9973 1993 . 4 X.-P. Li, R. W. Nunes, and D. Vanderbilt, Phys. Rev. B 47, 10 891 1993 . 5 M. S. Daw, Phys. Rev. B 47, 10 895 1993 . 6 R. W. Nunes and D. Vanderbilt, Phys. Rev. B 50, 17 611 1994 . 7 S. Goedecker and L. Colombo, Phys. Rev. Lett. 73, 122 1994 . 8 W. Kohn, Phys. Rev. Lett. 76, 3168 1996 . 9 P. Ordejon, D. A. Drabold, R. M. Martin, and M. P. Grumbach, Phys. Rev. B 51, 1456 1995 . 10 E. B. Stechel, A. R. Williams, and P. J. Feibelman, Phys. Rev. B 49, 10 088 1994 . 11 S. Goedecker and M. Teter, Phys. Rev. B 51, 9455 1995 . 12 C. A. White, B. G. Johnson, P. M. W. Gill, and M. Head-Gordon, Chem. Phys. Lett. 230, 8 1994 . 13 M. C. Strain, G. E. Scuseria, and M. J. Frisch, Science 271, 5245 1996 . 14 T. Zhu, W. Pan, and W. Yang, Phys. Rev. B 53, 12 713 1996 .
2

15 16

W. Kohn, Phys. Rev. 115, 809 1959 . A. Nenciu and G. Nenciu, Phys. Rev. B 47, 10 112 1993 . 17 R. Baer and M. Head-Gordon, Phys. Rev. Lett. 79, 3962 1997 . 18 R. Baer and M. Head-Gordon, J. Chem. Phys. 107, 10 003 1997 . 19 R. Shankar, Rev. Mod. Phys. 66, 129 1994 . 20 K. G. Wilson, Rev. Mod. Phys. 47, 773 1975 . 21 S. R. White and R. M. Noack, Phys. Rev. Lett. 68, 3487 1992 . 22 S. R. White, Phys. Rev. B 48, 10 345 1993 . 23 R. Kosloff and H. Tal-Ezer, Chem. Phys. Lett. 127, 223 1986 . 24 D. Neuhauser, J. Chem. Phys. 93, 2611 1990 . 25 M. R. Wall and D. Neuhauser, J. Chem. Phys. 102, 8011 1995 , especially Appendix E. 26 R. Baer and M. Head-Gordon, J. Chem. Phys. to be published . 27 G. H. Golub and C. F. Van Loan, Matrix Computations, 3rd ed. John Hopkins University Press, Baltimore, 1996 , p. 591. 28 X. Blase, L. X. Benedict, E. L. Shirley, and S. G. Louie, Phys. Rev. Lett. 72, 1878 1994 . 29 A. P. Horseld, P. D. Godwin, D. G. Petifor, and A. P. Sutton, Phys. Rev. B 54, 15 773 1996 .