Density-Matrix: Algorithms For Quantum Renormalization Groups
Density-Matrix: Algorithms For Quantum Renormalization Groups
Density-Matrix: Algorithms For Quantum Renormalization Groups
A formulation of numerical real space renormalization groups for quantum many-body problems is
presented and several algorithms utilizing this formulation are outlined. The methods are presented
and demonstrated using S = 1/2 and S = 1 Heisenberg chains as test cases. The key idea of the
formulation is that rather than keep the lowest-lying eigenstates of the Hamiltonian in forming a
new effective Hamiltonian of a block of sites, one should keep the most significant eigenstates of the
block density matrix, obtained from diagonalizing the Hamiltonian of a larger section of the lattice
which includes the block. This approach is much more accurate than the standard approach; for
example, energies for the S = 1 Heisenberg chain can be obtained to an accuracy of at least 10
The method can be applied to almost any one-dimensional quantum lattice system, and can provide
a wide variety of static properties.
low-lying excited states. when one puts four blocks together, as we do below), and
One begins by breaking the 1D chain into finite identi- we have not used S to further reduce the dimension of
cal blocks. It is usually convenient to start at the first it- H~~. (The value of S for a state can easily be inferred
eration with blocks consisting of just one site. We will la- by degeneracies for diB'erent values of S, . )
bel the blocks B and the block Hamiltonian H~. H~ con-
tains all terms of H involving only sites contained in B.
The lowest-lying eigenstates u,
are the states used to describe B' (BB B') Th. e new
;, ~
o. = 1, . . . m, of H~~
For example, for the Hubbard model at the first iteration, block Hamiltonian matrix H~ is diagonal. However, in
where B consists of one site, H~ = Un, gn, g p(n— ;g+n;g) the more general case where the states kept, the are u,
For the Heisenberg model at the first iteration, H~ ——0. not eigenstates of H~~ we can write
Rather than describe B and H~ in the usual way by —OH~~Ot,
listing the sites of B and using second-quantized oper-
H~
ator expressions for H~, we describe B by a list of the
many-body states on the block, and by quantum numbers
where the m x m matrix 0, , „,
= u', , ; i.e. , the rows
of 0 are the states kept. If 0 were square, this would
and matrix elements between these states. We store the be a unitary transformation. Since 0 is not square, the
number of states m, and for each state we list all quan- transformation truncates away (integrates out) the high
tum numbers which are to be used, such as S and S for energy states.
a spin system, or N~, %~, and S for an electron system. In order to obtain new matrices for S&, S„, etc. , it is
H~ is represented as an m x m matrix. In order to recon- necessary to use 0 again. First, one must construct the
struct H, additional information is needed besides H~. operators for S&, S„, etc. , for BB, which we denote by
The additional information describes the interactions be- Sg ) S ) etc. For example
tween blocks. For a Heisenberg system with interaction
1 S = [Sg]. .. b;„ (4)
S; S,+g S;S,+, + —(S+S,2+1
+, + S, S++~),
i
2122)2g22i
~ ~ ~
—— . 2
sociated with a single interval (an "onion layer" ) were IV. DENSITY-MATRIX APPROACH
added to the system at each iteration. The analogous
procedure for a 1D system would be to add a single site The fundamental difFiculty in the standard approach
to a block at each iteration. From a computational point discussed in Sec. II lies in choosing the eigenstates of
of view, this has a d.istinct advantage in that many more H~~ to be the states kept. Since H~~ contains no con-
states can be kept (m can be made larger) since at each nections to the rest of the lattice, its eigenstates have
iteration a system with nm states, as opposed to m, inappropriate features at the block ends. This is clearly
must be diagonalized, where n is the number of states on illustrated in the work of White and Noack, who sug-
a single site (n = 4 for Hubbard models, n = 2S + 1 for gested two alternatives to the standard approach. These
spin models). methods shared a common feature: The states that were
The most important difference between the Kondo sys- kept were not the eigenstates of H~~. They differed
tem and a 1D system is that the couplings between ad- in how the states to be kept were chosen. In the first
jacent layers or "sites" decreases exponentially in the method, the combination of boundary conditions (CBC)
Kondo system, whereas it remains constant for a 1D sys- approach, the lowest-lying eigenstates of several different
tem. This exponential decrease is the key to the success block Hamiltonians were kept. The several block Hamil-
of the method for the Kondo system and. related impurity tonian differed only in the boundary condition applied
systems. More discussion concerning how the detailed to a block; e.g. , one Hamiltonian might have periodic
form of the Hamiltonian makes the numerical approach boundary conditions applied and another antiperiodic.
accurate are given by Wilson. The rationale for this was that quantum fluctuations
Bray and Chui applied the approach described in the in the rest of the system effectively apply a variety of
previous section to the 1D Hubbard model. The results boundary conditions to the block. States from any sin-
were quite discouraging. Even when a large number of gle boundary condition cannot respond properly to these
states were kept (m = 1000), results for the energies of fluctuations. By applying a representative set of bound-
the lowest few levels were off' by 5 —10%%up for 16 site chains. ary cond. itions, which is in some sense "complete" enough
Results for 32 sites or larger were not considered reliable for the problem at hand, one obtains a set of states which
and were not presented. The origin of the difIiculties was are able to respond to these fluctuations. This approach
not clearly understood. proved very effective for the simple single-particle prob-
Other uses of this general approach since then have lems studied by White and Noack, as well as for Anderson
included the work of Pan and Chen and Kovarik. Very localization models.
recently Xiang and Gehring applied a slight variation of The CBC approach appears to be ill suited to inter-
the method of the previous section to the 1D Heisenberg acting systems. It is useful to consider a noninteracting
model. Their method added a single site to a block at many-particle system, such as the Hubbard model with
each iteration, rather than doubling the block size each U = 0. An arbitrary state of this system can be described
time. (The algorithms we present below also add a single in terms of the single-particle wave functions of each of
site to a block. ) This improvement gave results rather its particles. Some of these single-particle wave functions
more encouraging than those of Bray and Chui; keeping may have nodes at the ends of a block, and some may
about 200 states they obtained an error of about 0. 5%%uoin have antinodes.
the ground state energy. This can be compared to our It is easy to choose boundary conditions with gener-
results below, where keeping fewer states we obtain the ate block states where every particle on the block has a
ground state energy to at least nine digits. node or every particle has an antinode, but it is difFicult
There have also been a number of analytical real-space to get different boundary behavior for different particles.
RGs applied to the 1D Hubbard and related models. In order to properly represent the block, the states kept
These methods keep only a handful of states, and one not only need to allow for different end behavior for dif-
generally only expects to obtain qualitative features. ferent particles; they must represent a complete range of
These will not be discussed further here. boundary behavior.
Lee used a numerical RG method to study 2D Ander- This general line of reasoning is supported by numer-
son localization. Since the model studied was noninter- ical tests on Heisenberg chains. We have tried to find a
acting, a two-dimensional calculation was feasible. It was simple set of boundary conditions which can be used to
clear from an examination of the density of states that treat the S = 2 Heisenberg chain. We tried combina-
the algorithm was not especially accurate. However, Lee tions of periodic and antiperiodic couplings between the
argued that the errors would simply result in the system ends of the block, as well as varying the magnitude of the
being in another realization of the random potential. Lee coupling between the ends of a block. We were unable
concluded that there was a critical amplitude for the ran- to find any set of boundary conditions which was at all
dom potential which induced localization. However, not satis factory.
long after Lee's work, Lee and Fisher found that the The other approach suggested by White and Noack,
2D model is logarithmically localized even for arbitrarily the superblock method, forms the basis for the density
small randomness using a different approach. This result matrix approach. In the superblock method, one diag-
is now generally accepted. onalizes a larger system (the "superblock"; the name is
In summary, the standard. numerical RG approach gen- analogous to "supercell, " as used in electronic structure
erally performs poorly. The exceptions to this rule in- calculations) composed of three or more blocks which in-
volve Hamiltonians which can be put into a special form. cludes the two blocks B'B which are used to form B'. The
10 348 STEVEN R. WHITE 48
state when one wishes to obtain several of the lowest- S, = 0 states). An eigenstate of the superblock Hamil-
lying states: If we put the lattice with equal probability tonian is called a target state if it is used in forming the
into each of several states, then the block states obtained block density matrix. The most efFicient algorithms use
from the density matrix will equally well represent each only a single target state (usually the ground state) in
of these lattice states. We represent the mixed case by constructing the density matrix. By targeting only one
saying that the lattice has probability TVA. to be in state state, the block states are more specialized for represent-
~Q"). If the system is at a finite temperature, then the ing that state, and fewer are needed for a given accuracy.
Wk are normalized Boltzmann weights. In this case the Probably the most important characteristic of a density-
appropriate definition for the error in the representation matrix algorithm is the rate at which the accuracy in-
is creases with the number of states m. We have found that
the accuracy of the representation of the target states in-
(14) creases roughly exponentially with rn, , at least for open
boundary conditions. The coefIicient governing the in-
crease of accuracy with m, is largest with a single target
Note that we are interested in determining a single set state.
of optimal u,
whereas we allow the rest of the universe Several considerations enter in the construction of the
additional freedom to choose a difI'erent v for each state superblock to be used in an algorithm. Generally it is
k. Minimizing over the v"', and a", we find
u, more eKcient to enlarge the block by adding a single
pu = Q)~u )
site, rather than doubling a block, for the same reasons
as in the standard RG approach (see Sec. II). A block
with can be represented more accurately with a given m, if it
(16) connects to the rest of the chain only on one end, rather
than both ends. A block can connect to the rest of the
chain by only one end if it is on an end of a chain with
open boundary conditions. If periodic boundary condi-
= )k
Wk(a") tions are used, it is not possible for a block to connect
on only one end. Below we give results of calculations
This equation for p is the definition of the reduced density which show that the open boundary condition case per-
matrix when the universe is in a mixed state, and the u forms much better than the periodic boundary condition
are the eigenstates of p. case. Our intuitive picture for this numerical result is the
Thus the conclusion when the universe is in a mixed following: Roughly speaking, each eigenstate of the block
state is identical to the result for a pure state: The op- density matrix represents the response of the block to a
timal states to keep are the eigenvectors of the reduced particular quantum fluctuation in the rest of the chain.
density matrix with the largest eigenvalues. A block with two ends which connect to the rest of the
system must respond to nearly independent fluctuations
near each of the ends. In the case of a long block, where
V. DEN SIT Y-MATRIX ALGORITHMS the ends are nearly independent, if m states are required
to accurately describe a single end to a given accuracy,
Incorporating the result of the previous section in a nu- then approximately m states would be required to accu-
merical renormalization group algorithm involves a fun- rately represent both ends. In other words, if for a given
damental change in the way the calculation is carried accuracy open boundary condition require m states, pe-
out. In the standard RG approach, to find the states riodic boundary conditions require roughly m states.
to be kept, one diagonalizes only the system BB, which Figure 1 shows the superblock configuration used for
becomes B'. In the density-matrix approach, in order most of the calculations reported here. We adopt the
to obtain any reasonable approximation to the density notation Bg ~ B&, for this configuration, where Bg repre-
matrix, it is necessary to diagonalize the Hamiltonian sents a block composed of / sites, B& is a reflected block
of a larger system which includes BB, namely, some (right interchanged with left) of length P, ~ represents
sort of superblock, and then use the eigenstates of the a single site, and the total length of the superblock is
superblock to determine the density matrix. The den- I = E+ E'+ 2. Here B' = By+i is formed from the left
sity matrix is then diagonalized, and its most significan block plus a single site, i.e. , By+i —Bg. Open boundary
eigenstates are the states kept. The number of eigen- conditions are used. The right-hand block and site B&,
states of the Hamiltonian of the superblock used to pro-
duce the density matrix can be as small as 1; this single
state produces a density matrix for BB which has many B,
eigenstates to be used as block states to be kept.
A density-matrix algorithm is defined mainly by the Bl+I
form of the superblock and the manner in which the
blocks are enlarged (such as by doubling the block, FIG. 1. The configuration of blocks used for the density
B' = BB, or by adding a single site, B' = B + site), matrix calculations considered here. The rectangles represent
and by the choice of superblock eigenstates used in con- blocks containing 3 and l' sites, and the solid circles represent
structing the density matrix (e.g. , the two lowest-lying single sites.
10 350 STEVEN R. WHITE 48
are only used to help form the density matrix for Bg+q., in states, then from Eq. (12), the density matrix p for the
the construction of the density matrix, the states of B&, system Bg has at most m nonzero eigenvalues. Initially,
are traced over. This configuration can be used in two when B& (and Bg) consists of just a single site, it has only
difFerent ways: in an infinite chain method, in which the a few states. With this method, the number of states in
chain size increases by two at each step, and in a finite By+i could not be larger than the nuinber in B& (which is
chain method, in which the chain size is fixed. the same as the number in Br), unless one included states
with density-matrix eigenvalues of zero. In general, for a
A. In6nite system method robust method, the number of nonzero eigenvalues of the
density matrix should be larger than the number of states
In the first step of the infinite system method, we start kept, except in the first few steps when all states of the
with a four site chain and diagonalize the Hamiltonian of block are kept. If more than one state is targeted, then
the superblock configuration Bq ~ Bz, where Bq and the Bg B& configuration is presumably feasible, although
we have not tested it. In any case, the presence of extra
Bz both represent a single site. We use the Davidson
algorithm for the sparse matrix diagonalization, but sites in the center which are not part of the outer blocks
one could also use the more well-known Lanczos method. makes the method more accurate, since B& is represented
Using the target states calculated with this configura- by an approximate Hamiltonian, whereas the central sites
tion, we calculate a density matrix and form an effective are represented exactly. Thus extra sites in the center
Hamiltonian for B2 —Bq ~ . In the second step we diago- produce a more accurate density matrix. The penalty
nalize B2 B2, where we have formed B2 by reflecting for these extra sites is that the diagonalization of the
B2. We continue in this manner, diagonalizing the con- superblock is more work.
figuration Bg ~ B&, and setting By+i —Bg, and using The infinite system algorithm is summarized in Table
Bg+q and. its reflection in the next step of the iteration.
II. The representation of the blocks is identical to that
At each step, both blocks increase in length by one site, of the standard algorithm: We describe a block by list-
and the total length of the chain increases by 2 at each ing how many states it has and the quantum numbers
step of the iteration. The infinite chain method is usually for each state, and by storing matrices for H~, S, , etc.
used when one is interested in ground state properties of Once the matrix 0
is constructed using the most signif-
the infinite chain. Each step of the iteration pushes the icant eigenvectors of the density matrix, the change of
ends of the chain farther from the two sites in the center. basis procedure it also identical to that of the standard
After many steps, each block approximately represents algorithm. For the purposes of organizing the algorithm,
one-half of an infinite chain. In order to represent one- it is easiest to think of the two sites in the middle as
half of an infinite chain, B must not only contain many blocks which can be treated similarly to the two outer
sites itself; its effective Hamiltonian must be formed from blocks, although they contain only a few states.
a system in which the rest of the chain has many sites.
The efFective Hamiltonian formed from the left-hand side B. Finite system method
Bp ~ depends strongly on the right-hand. side B& . The
infinite chain algorithm converges in two senses simulta- The finite system algorithm is designed to calculate ac-
neously: in the length of Bg going to infinity and in the curately the properties of a finite system of size L, which
sense that Bg is adapted to respond to an infinite chain we will assume for simplicity to be even. It is summarized
connected to it on the right. in Table III. It begins with the use of the infinite system
Why not use the simpler configuration Bg ~ B&? In I
algorithm for j2 —1 steps, so that the final superblock
other words, why make the right-hand side B& rather used is of size L. In the infinite system method, there
than B&? If there is only one target state, and B& has vn is no need to store Bg once we have B~+q, we need only
Make four initial blocks, each consisting of a single site, representing the initial four site system. Set up matrices
representing the block Hamiltonian and other operators.
Form the Hamiltonian matrix (in sparse form) for the superblock.
Using the Davidson or Lanczos method, diagonalize the superblock Hamiltonian to find the target state @(ii,i 2, i3, i4).
vP is usually the ground state. Expectation values of various operators can be measured at this point using @.
Form the reduced density matrix for the two-block system 1-2, using p(ii, i2, ii, i2) = P
z3)z4 Q(ii, i2, i3, i4)g(ii, i2, i3, i4).
. .
5. ..
Diagonalize p to find a set of eigenvalues m and eigenvectors u. , . Discard all but the largest m eigenvalues
and associated eigenvectors.
6. Form matrix representations of operators (such as H) for the two-block system 1-2 from operators for each separate
block [cf. Eq. (4)].
Form a new block 1 by changing basis to the u and truncating to m, states using H = OH 0,
etc. If blocks
1 and 2 have mi and m2 states, then
Replace old block 1 with new block l.
0 ..
is an m x mim2 matrix, with matrix elements O(a; ii, i2) = u. , , a = 1, . . . , m.
I
TABLE III. Finite system density-matrix algorithm for a 1D system consisting of sites. A calculation consists of several
I
iterations, indexed by I, with each iteration consisting of —3 steps, indexed by l, where l is the size of the erst block.
I
(First half of = 1.) Use the infinite system algorithm for L/2 —1 steps to build up the lattice to L sites.
At each iteration store the block Hamiltonian and end operator matrices for block 1. Label the blocks by their size,
Bg, E = 1, . . . , L/2
2. I
(Start of second half of = 1.) Set / = L/2. Use Bg as block 1, and the re8ection of BI, g 2 as block 4.
3. Steps 2 —8 of Table II.
4 Store the new block 1 as Bg+q, replacing the old Bg+q.
Replace block 4 with the reflection of Bl. g q, obtained from the erst half of this iteration.
I
If E & —3, set E = E+ 1 and go to step 3.
7. (Start of iteration I,I) 2. ) Make four initial blocks, the first three consisting of a single site, and the fourth
consisting of the reflection of B'L, 3 from the previous iteration. Set Z = 1.
8. Steps 2 —8 of Table II.
9. Store the new block 1 as Bg+q, replacing the old Bg+~.
10. Replace block 4 with the reflection of Bl. q 2, obtained from the previous iteration (if / ( L/2 —1) or the first
)
half of this iteration (if E I/2 —1).
( I
If f. L —3, set E = I. + 1 and go to step 8. If 8 = —3, start a new iteration by going to step 7. (Stop after 2
or 3 iterations. )
the Hamiltonian matrix is proportional to m, whereas It is usually more convenient to choose points i and j
on
if block 2 is a single site, the number is proportional to diferent blocks, rather than keep track of complicated
m. In the case of periodic boundary conditions, these operators such as [S'S&];„~. A convenient way to calcu-
considerations are relevant because block 4 is connected late a correlation function such as this as a function of
to block 1, and thus one of these should be a single site. j j
—k is to always put and k at the same distance (within
Otherwise, the periodic case is nearly identical to the a lattice spacing) from the center of the chain. As —k j
open boundary case discussed above. is increased, both points move outwards symmetrically
towards the ends of the chain.
VI. MEASUREMENTS
VII. RESULTS
Properties of the I-site system can be obtained from
the wave functions of any of the L —3 superblock config- In this section we give a brief survey of results, with
urations, although we find that the symmetric configura- the intention of illustrating features of the algorithms
tion (with both the left and right blocks of size L/2 —1) rather than giving a study of the model systems. Figure
usually gives the most accurate results. The procedure 2 shows the density-matrix eigenvalues m for a 32 site
is to use the wave function ~Q) resulting from the diag- system for both open and periodic boundary conditions,
onalization of the L-site system to evaluate expectation and for both S = 1/2 and S = 1, targeting one state.
values of the form (g~A~g). Rather complicated opera- The figure shows the eigenvalues only for one particular
tors can be evaluated fairly easily, but dynamical infor- size of block 1, but similar results are obtained at other
mation is not easily obtained. In order to measure A ) one steps in the iteration. The falloK of m is most rapid
must have kept operator matrices for the components of for the open S = 1/2 case. The figure shows that it
A. For example, to measure the on-site spin density S'- is possible to obtain an accuracy (truncation error) of
for all sites j, one must keep track of matrices ~Sg J't1 Z1 ) ~ better than 10 in this case, keeping only m = 20 states.
for all sites in block 1, and similarly for blocks 2, 3, and 4 The S = 1 open case shows a slightly slower fallofF, with
(regarding the single sites in the middle as blocks 2 and an accuracy of 10 obtained with I, = 48. The periodic
3). At each step of each iteration, these operators for cases fall oà more slowly than either of the open cases.
blocks 1 and 4 must be updated using analogues of Eqs. Keeping m = 50 states yields an accuracy of roughly
— —
(4—6). One then obtains the expectation value using 10 5 —10 6 . Even more accurate treatment of periodic
an
): ~:,......[S]... ~„„....
X1'C2'L3'L42, 1
, ,
systems is quite feasible; we have been able to keep as
many as m = 200 states in a calculation of a 60 site
system. Degeneracies in the m are clearly visible, which
etc. This procedure gives exact evaluations of (g~A~g) come from spin symmetries. One must adjust I,
so as not
for the approximate eigenstate ~g).
to split a degenerate set of levels, since this will destroy
For a correlation function such as (@ S'S& ~g), the eval- spin symmetries that would otherwise be preserved.
In Fig. 3 we show the relative accuracy of the ground
j
~
''
10o
and [S&];„~,and one has
10 i
c, S=1
(41S;Sil&) = W 2y2g1gl4 L g $211y I. &]1414 P Ky22 7324 —2 , S=1/Z
b1 L2 43 L4 X1 Z4
'I 'I 10 =1
(20) 3
=1/~
10
If j and k are on the same block, one should not use 10-4
41 22 'C3 '5421 Z1
10 6
10 7
This expression does not evaluate the correlation function
10 30 40 50
exactly within the approximate state ~g). The sum over
I
should run over a complete set of states, but does
not, whereas the sums over the other variables need run FIG. 2. Density-matrix eigenvalues to vs eigenvalue in-
only over those states needed to represent ~g), since they dex n for a 32 site system for both periodic and open bound-
appear as a subscript in either the ~@) on the left or on ary conditions, S = 1/2 and S = 1. The eigenvalues were
the right. To evaluate this type of correlation function, obtained from the BJ5 ~ BJ5 system. The steplike struc-
one needs to have kept track of (S S&],„rthroughout the. ture comes
. from the presence of spin degeneracies.
tional near degeneracies in the 5 = 1 cases come from the
Addi-
calculation. One then evaluates
presence of nearly free S = 1/2 effective spins on the ends of
(@IS;S~l&) = ).
t 12 g 'C
3 24 'E 1
(22) S = 1 blocks. The errors in the calculation is determined by
, which can be estimated from the figure.
DENSITY-MATRIX ALGORITHMS FOR QUANTUM. . . 10 353
FIG. 3. Relative error in the ground state energy AE/E Haldane's conjecture that integral-spin Heisenberg chains
for a 28 site S = 1/2 system as a function of the number are gapped. The gap in the S = 1/2 system tends to 0
of states kept m, for periodic and open boundary conditions. as L M oo, in agreement with the Bethe ansatz exact
The exact energy was obtained from Ref. 16 for the periodic solution of the model. These results predict an infinite
case, and from the renormalization group calculation itself system gap for the S = 1 chain of 0.411(1). Using open
with m = 58 for the open case. boundary conditions (of the "soft" type discussed below)
we have obtained what we believe is the most accurate
value of A to date: A = 0.4105(l).
the number of states kept. The error for the periodic case It is clearly more accurate to use open boundary condi-
tions than periodic in these calculations, but a bulk sys-
was obtained from comparison with an exact diagonaliza-
for the open case, we used the renor-
tem is generally better described using periodic bound-
tion calculation;
malization group calculation itself with m = 60, which ary conditions. It is important to understand. the effect
we believe reproduces the ground state energy to better
of an open boundary on these systems. The efFect of an
than 10 . (For smaller systems, we have verified that open boundary is most easily seen for an S = 1/2 sys-
tem by measuring the local bond strength (S~ S~+i) as
the open case really is this accurate by comparison with
a function of site index i. In Fig. 6(a) we show the local
our own exact diagonalization calculations. ) Again, we
bond strength for a 60 site system. The open boundaries
see that the open boundary condition case is extremely
cause a strong alternation in the bond strength which
accurate even for a relatively small m. The relative er-
ror in the energy is somewhat larger than the truncation decays very slowly. (We emphasize that this is the effect
error indicated by the density-matrix eigenvalues. The
of the boundaries; errors in our calculation are negligi-
ble. ) This effect is easily understood using the valence
fallofF in the error in the energy is much slower for the
bond picture of the ground state of the S = 1/2 system.
periodic case; it is dificult to achieve an accuracy of bet-
ter than 10 in this case. Figure 4 shows similar results In the valence bond picture, the ground state is viewed
as a resonance between two different states, one having
strong singlet bonds on the even numbered links of the
10 i chain, and no bonds on the odd links, and the other with
strong bonds on the odd links, and none on the even.
10 ~
10 3 0.5
1O-4
0.4
&105
1O-6 0.3
0.2
0. 1
50 100 125
0.0
FIG. 4. Relative error in the ground state energy E for a
16 site S = 1 system, as a function of the number of states
0.00 0.02 0.04 0.06 0.08 0. 10
kept m, for periodic and open boundary conditions. The exact
energy was obtained from Ref. 17 for the periodic case, and FIG. 5. Gaps Ag between the ground and first excited
from the renormalization group calculation itself with m = states for S = 1/2 and S = 1 systems with periodic boundary
140 for the open case. conditions versus the inverse length of the chain 1/L
10 354 STEVEN R. WHITE
Simple variational estimates show that this type of state can participate in a bond.
is greatly preferred over a Neel ordered state. This pic- If the system has an odd number of sites, one end favors
ture strongly suggests that the system can be perturbed the strong bonds on the even links, and the other favors
very easily into one of these two dimerized states. The them on the odd links; thus the system is frustrated. In
open boundary conditions favor a strong bond on the Fig. 6(b) we show the local bond strength on a 61 site
outermost links, since that is the only way the end sites system. The system shows the odd links with stronger
bonds on the left, and the even links stronger on the right.
For other system sizes the behavior is nearly identical to
that shown for I= 60 in the sense that there are never
L=60, S=i/2, Open BCs more than a few links in the center with almost no bond-
—0.3 strength alternation present, and there is approximately
linear falloff of the bond strength in the central region.
—0.4 We interpret this behavior as evidence of a mobile domain
+ wall (soliton), with a nearly uniform probability to be at
CG
—0.5 any particular site in the central region. If one averages
over the various positions of the domain wall, one obtains
results consistent with the figure.
If we desire the accuracy of open boundary conditions,
but not the bond-strength alternation which is absent
—0.7 in the bulk system, it is possible to considerably weaken
the alternation using "soft" boundary conditions. In Fig.
20 30 40 50 60
1 6(c) we show results from an especially simple version of
these boundary conditions. In this case we weakened
—0.2 the two outermost exchange constants J from their bulk
(b)
L=61, S=1/2, Open BCs
0 3 —1.2
—0.6 M —1.5
10 20 30 40 50 60
10 20 40 50 60
(c)
L=60, S=1/2, Open BCs
—0.3
&59-6O 0.6
(b)
—0.4 L=60, S=1, Open BCs
+ S tot —g
M
M —0.5
0.2
—0.7 —0.2
10 20 40 50 60
—0.4
FIG. 6. Local bond strength for (a) 60 site and (b) 61 site 10 20 30 40 50 60
S = 1/2 chains with open boundary conditions. The systems
try to have strong single bonds at the outermost links of the
chain; this induces a domain wall in the odd-length chain. In FIG. 7. Local bond strength (a) and local value of S, (b)
(c), the bond-strength alternation is diminished by weakening for a S = 1 system with 60 sites. The S = 1/2 efFective spina
the local value of the exchange constant J
on the outermost on the ends of are clearly visible in (b), which shows one of
links. the low-lying triplet states just above the ground state.
48 DENSITY-MATRIX ALGORITHMS FOR QUANTUM. . . 10 355
K.G. Wilson, Rev. Mod. Phys. 4'7, 778 (1975). M. D. Kovarik, Phys. Rev. B 41, 6889 (1990).
S.R. White and R.M. Noack, Phys. Rev. Lett. 68, 3487 T. Xiang and G.A. Gehring, 3. Magn. Magn. Mater. 104-
(1992). 107, 861 (1992); T. Xiang and G. A. Gehring (unpublished).
S.R. White, Phys. Rev. Lett. 69, 2863 (1992). See, for example, J.E. Hirsch, Phys. Rev. B 22, 5259 (1980);
J.W. Bray and S.T. Chui, Phys. Rev. B 19, 4876 (1979). C. Dasgupta and P. Pfeuty, J. Phys. C 14, 717 (1981), and
C.Y. Pan and Xiyao Chen, Phys. Rev. B $6, 8600 (1987). references therein.
10 356 STEVEN R. WHITE
P.A. Lee, Phys. Rev. Lett. 42, 1492 (1979). ing (Cambridge University Press, New York, 1986).
P. A. Lee, and D. S. Fisher, Phys. Rev. Lett. 47, 882 (1981). E.R. Davidson, 3. Compnt. Phys. 17, 87 (1975).
R.M. Noack and S.R. White, Phys. Rev. B 47 9243 (1993). S.R. White and D. Hnse, Phys. Rev. B 48, 3844 (1993).
R.P. Feynman, Statistical Mechanics: A Set of Lectures D. Medeires and G. G. Cabrera, Phys. Rev. 8 431 3703
(Benjamin, Reading, MA, 1972). (1991).
W. H. Press, B.P. Flannery, S.A. Teukolsky, and W. T Vet- "A. Moreo, Phys. Rev. B 35, 8562 (1987).
terling, Numerical Recipes: The Art of Scientific Comput S.R. White, Phys. Rev. B 45, 5752 (1992).