PHYSICAL REVIEW B VOLUME 40, NUMBER 6 15 AUGUST 1989-II

High-precision sampling for Brillouin-zone integration in metals

M. Methfessel* and A. T. Paxton

Max Pla-nck Inst-itut fii r Festko rper'forschung, Postfach 80-06-65, 7000 Stuttgart 80, Federal Repubit'c of Germany
(Received 15 February 1989)
We present a sampling method for Brillouin-zone integration in metals which converges exponen-
tially with the number of sampling points, without the loss of precision of normal broadening tech-
niques. The scheme is based on smooth approximants to the 5 and step functions which are con-
structed to give the exact result when integrating polynomials of a prescribed degree. In applica-
tions to the simple-cubic tight-binding band as well as to band structures of simple and transition
metals, we demonstrate significant improvement over existing methods. The method promises gen-
eral applicability in the fields of total-energy calculations and many-body physics.

I. INTRODUCTION expense of using a prohibitively fine k mesh. Thus for

each choice of 8' the broadening method will converge to
In solid-state one-electron and many-body calculations, a particular result; in the unattainable limit as 8' — +0 so
integrals of periodic functions over the Brillouin zone would each of these results converge absolutely to the
(BZ) are routinely made in the evaluation of densities of precise answer.
states, single-particle sums, charge densities, matrix ele- It is our intention to show here that a more sophisticat-
ments, response functions, and so on. Precision in energy ed choice of the broadening function can provide the
on the order of l mRy is common; in more demanding necessary freedom with which to control both the k con-
cases, such as the calculation of phonon frequencies, one vergence and the absolute convergence independently. In
requires l or 2 orders of magnitude better convergence. practice, this means that a broadening can be chosen for
In the case of insulators and semiconductors, the func- which k convergence is reached for a relatively coarse
tion to be integrated has the property that it is infinitely mesh and that the corresponding limit is close to the true
many times differentiable, due to the separation of occu- value. The method we propose revolves around a hierar-
pied and empty bands by a gap. Under these cir- chy of smooth approximations to the step and 6 func-
cumstances, it is well known' that integrals converge ex- tions, which we derive and motivate in the following two
ponentially in h, where h is the spacing between divisions sections. These are then applied to BZ integrals for the
made along the primitive vectors of the reciprocal lattice simple-cubic tight-binding band in Sec. IV and to actual
in constructing a uniform mesh over the BZ. (By ex- band structures of metals in Sec. V. The most important
ponential convergence, we mean convergence that is points are summarized in Sec. VI.
more rapid than any power in h. ) Investigations into the
properties of metals, on the other hand, are hampered by II. SMOOTH APPROXIMATIONS TO THE STEP
the fact that one must integrate a function over the BZ AND 5 FUNCTIONS
that is discontinuous due to the partial filling of the ener-
gy bands, and hence not even once differentiable. This We wish to evaluate
leads to very slow convergence when sampling over a uni-
form mesh in the BZ, or at best an h convergence when
I= fBZS(E(k) —E, )f(k)dk= f Qo
S(E EF)F(s)dE, —
using the linear tetrahedron method. Since in many ap-
where
plications the time-limiting procedure is the solution to
an eigenvalue problem at each sampling k point, it is im- F(E)= f f(k)6(E —E(k))dk .
portant that the required precision be obtained without a BZ

prohibitively fine k mesh. E(k) represents an energy band as a function of wave

We will implicitly deal with two types of convergence vector and EF is the Fermi energy. The function to be in-
in what follows, which may be best illustrated by refer- tegrated, f, is multiplied by the Fermi cutoff or step func-
ence to a very commonly used procedure for BZ integra- tion: S(x)=1 — e(x)=1 for x (0, and 0 otherwise. F(E)
tion in metals: one may immediately obtain improved k I
might be, for example, the density of states g (E), if is to
convergence by a broadening or "smearing" of contribu- be the total charge within the Fermi surface; or
tions from all k points into Gaussian, Lorentzian, or I
F (e) =g (e)e if is to be the band energy. It is easy to see
similar smooth functions with characteristic linewidth 8'- that conventional smearing, which essentially replaces
(thorough discussions can be found in Refs. 4 and 5). S(x) by a Fermi-Dirac-like distribution, leads to an error
This method ensures exponential convergence of in- in the integration which is shown schematically in Fig. l.
tegrals, albeit to the wrong answer. The only justification Note that in general this will occur unless F(e) is a con-
for this ad hoc procedure is that in the limit as 8' — +0 one stant near EF.
would recover the "absolutely" converged result —
at the The more sophisticated approach we adopt amounts to

40 3616 1989 The American Physical Society

 HIGH-PRECISION SAMPLING FOR BRILLOUIN-ZONE. .. 3617

S~(x) =1 — f D~(t)dt,
and making use of the identity (d/dx)[H„(x)e "X ]
H—„+i(x)e and the definition of the error func-
tion, ' we obtain the central result of the present paper:
Sp(x) = —,'(1 —erfx)

S~(x) =So(x)+ g 3 H2 i (x)e

X

n=1
in which the zero-order approximation So corresponds to
EF} simple Fermi-Dirac-like smearing, and higher-order
terms serve to correct the errors inherent in such a pro-
cedure.
Sz has a similar property as Dz of integrating polyno-
mials exactly, since by partial integration of
FIG. 1. A schematic density of states g (c) and the same func-
tion multiplied by a Fermi-Dirac-like smoothing. An error 0= f [Dx(x) —5(x)]P(x)dx,
arises when the charge below FF is calculated as
Jg(s}S&(E F~}— ds since the two hatched areas are not equal, one has
unless the density is constant near EF. dI'
0= f [S~(x) —S(x)] dx dx .
oo

oo

Therefore, S~(x) can be substituted for S(x) in the in-

a search for successive approximants, Sz, to the step tegral jS(x)F(x)dx without incurring significant error
function which are smooth and for which the error is if F(x) can be represented as a polynomial of degree 2N
zero if F(s) is a polynomial of some order within the en- or less in the range where S~(x) — S(x) (or, equivalently,
ergy range determined by the choice of broadening. If we '
e X ) is appreciably nonzero; whereas Sp will serve only
use a characteristic broadening W, we can work with a
in cases where F(x) is constant in this interval. This is
dimensionless energy variable, x =(s EF )/W, and — look
our desired property.
for a function Sz(x) to approximate S(x). Our approach
In Fig. 2 we show successive approximations Sz(x)
is to write down a suitable representation of the 5 func-
and D~(x) to the step function and 5 function, respec-
tion 5(x), which we will integrate to obtain successive ap-
tively. Note that both oscillate close to x =0 as might be
proximants to S(x). One may expand 5(x) in a complete
set of functions, and since it is even we may write,
expected from the Gibbs phenomenon of Fourier
00
5(x)= g A„Hz„(x)e
n=0
X

where H„ is the Hermite polynomial' of degree n. Us- DN(x}

ing the orthogonality of the Hermite polynomials with N
respect to Gaussian weights:

f H„(x)H (x)e dx =n!2"&m5„

we obtain for the coeScients A„,

Ki„(0) (
—1)~
(2n }!4"&rr n!4"&m.

Now consider the finite sum

SN(x)
Dz(x) = g A„H2„(x)e X

which we introduce here as an approximation to the 5

function. By construction,

f D~(x)P(x)dx = f 5(x}P(x)dx =P(0)

if P(x) is a polynomial of degree 2K+ I or less; since FIG. 2. Successive approximants to the 5 function, Dz, and
then P(x) may be expanded in the Hermite polynomials to the step function Sz. The order of the approximant N is indi-
up to degree 2%+1. To obtain an approximation to the cated on each curve. Note that the zero-order approximants are
step function S(x), we integrate D~: a Gaussian and complementary error function, respectively.
3618 M. METHFESSEL AND A. T. PAXTON

analysis. The figure shows that for practical purposes, simulated with zero-order functions.
D&(x)=0 and S~(x) =S(x) if jx~ is greater than 2 —3. Consider the simplest case for which S, could give
We remark that the method is trivially easy to include better results than So: as above, we want to integrate
in a standard BZ sampling or special points program.
The approximants S~(x) may be obtained using the re-
f (k) over the occupied part of the band, and we suppose
that F(e) is linear in E near the Fermi energy EF. The er-
currence for the Hermite polynomials: ror in the integral obtained at k convergence can be cal-
culated by a standard integration, giving
Ho(x) =1, H, (x) =2x,
H„+,(x) =2xH„(x) —2nH„, (x) . S(c—
, E~—
) F(e)de
Substituting Sz(x) for S (x) in our original integral I, and
sampling over a uniform mesh which is reduced by sym- 'F'(EF) W for N =0
—,
metry to a set of points [k; J with weight factors a; in the
irreducible wedge of the BZ, we obtain the quadrature 0 for X&0 .

I~= ga;f(k;)S~(x;), x;=

E(k; ) EF— In this, and in what follows, primes denote differentiation
with respect to c.. One sees that the zero-order result
8' converges to the correct value as 8'~0, but only as
The sum converges exponentially in k space because of slowly as 8'. For the case under consideration, all
the smoothness of S~. This procedure is as simple as higher-order approximants already give the exact result.
conventional broadening methods for metals. In fact The tests in Secs. IV and V will show that k convergence
such a method is equivalent to our zero-order approxima- is only marginally slower for S, compared to So. Thus,
tion with 8' equal to the width of broadening used. If there is an immediate gain in using S& instead of So
Dz/W is substituted for Sz in the above equation, one whenever F(e) has a nonzero linear term near EF.
obtains a completely analogous method for evaluating in- Next we determine the analogous relation for the case
tegrals over the Fermi surface. We remark in passing of an integral over the Fermi surface, i.e.,
that Sz and Dz could be used in other contexts, for in-
tegrating or interpolating a function of a single variable J= F c D
tabulated on a regular mesh.
To conclude this section, we note for completeness that
the special integrative property of D& and S&, namely
which converges to F(EF) as W — +0 or asX~ ca. Since
Do integrates polynomials of linear order exactly, F(E)
that they integrate polynomials up to a certain degree ex- must have a nonvanishing quadratic term near EF if D, is
actly, is not sufhcient to define these functions uniquely. to give a better result. If F(s) is quadratic in e, the error
A similar formalism would result for another choice of in Jz at k convergence is
weight function instead of the Gaussian. However, it is
clear that for the application of integrating bands of arbi- 4F"(Ez) W— for X =0
trary width with a step or 5 function at F+, the weight
function must be smooth over the whole real axis and
0 for X&0.
strongly localized near x=0. It is convenient that the Again, we have found that k convergence of integrals
simplest choice, a Gaussian, is also the weight function made using D, is only slightly slower than those using
for a convenient and well-known orthogonal-polynomial Do. The criterion for making Fermi-surface integrals is,
sequence. therefore, use of D, in place of Do gives an immediate
improvement if F(e) has a substantial quadratic term at
III. THE BENEFIT FROM USING EF. As a practical example, note that an approximation
to the density of states (DOS) for a given set of bands
HIGHER-ORDER APPROXIMANTS
E„(k) can be accumulated by adding Dz(x„; )/W,
For the k-space integration procedure which we are x„;=[E—E„(k;)]/W, with the proper weight into the
presenting here, it is a complication that two parameters DOS for each band n and each mesh point k, . This is
8'and X must be chosen suitably. We would like to ac- equivalent to the case under discussion if one considers
quire a feeling for their significance. By construction, use E~ as a variable. From the expression for the error hJ~,
of S~ (or D~) leads to a negligible error if F(E) is it follows that the use of Do, i.e. , making the DOS by
representable as a polynomial of degree 2X (or 2N + 1) or conventional Gaussian smearing, has the effect of reduc-
less in an interval of-5 W around the Fermi energy (see ing peaks and filling in valleys, and that this undesirable
Fig. 2). The k-space converged result can therefore be phenomenon vanishes only as 8' . On the other hand,
made to approach the true value either by increasing X the DOS made using D, will have the proper heights and
or by reducing JK Obviously, our approach would be depths if 8'is no larger than about one-tenth of the dis-
superAuous if it were always possible to reproduce the re- tance between successive peaks. For a larger broadening,
sult for some set 8' and X, with equally fast k conver- the same improvement can be obtained through the use
gence, by simply using the zero-order function at a small- of functions of higher order than X =1. Tests for energy
er broadening. In this section, we show that the use of bands of transition metals have confirmed this.
orders X) 0 brings definite advantages which cannot be The connection between the two cases (integrals over
 HIGH-PRECISION SAMPLING FOR BRILX.OUIN-ZONE 3619

the occupied states and over the Fermi surface) can be

made in the following way. It is easy to show that

EF
f f F
C.
F(s)S~ de= F(e}dE, ia

where F(e) is the result of smoothing F(E) by convoluting

it with D~/W:

F(e)= f F(t) D~

It follows that F(e} made using D0 is equal to F(s) at

those energies for which F(e) is linear in E on the scale
given by W. On the other hand, the integral up to EF
made using SQ is already in error if F(e) has a linear
term, i.e., unless F(s) is approximately constant near EF.
The reason for this seeming discrepancy is that the in-
tegral has summed the errors in F(e) for all energies up to
EF. The preceding equations show an interesting fact,
namely that this accumulated error made by smoothing
F(e) into F(e) depends only on the behavior of F(e) near ENERGY
EF. More precisely, the error estimate above gives this FIG. 3. Density of states of the simple cubic tight-binding
sum (to leading order) as proportional to the slope of
band (Ref. 8) (dashed line) and successive approximations ob-
F(E) at EF. tained from its convolution with approximants D& to the 5
function. Only the negative energy range is shown. The energy
broadening 8'is 0. 10.
IV. THE SIMPLE-CUBIC TIGHT-BINDING BAND
The tight-binding band (TB) on a simple-cubic lattice,
E ( k ) = —' ( cosr k„+cosmk + cosmk,
—, }, Characteristic for the curves made by sampling using the
Sz is that they become fiat very rapidly after a threshold
provides us with an ideal DOS g Tn(e) with which to both for the fineness of the mesh is reached. For larger X the
illustrate and test the principles of the present integration threshold is reached later. A more exact analysis shows
method. Near the minimum of the band at —1,
i= that each curve converges to its limit as exp( — C/h )
gTa(E) shows free-electron-like behavior; near the center whereas the linear method converges as h . In view of
at c. =0, it is practically constant, and there are Van Hove
singularities at c, =+ —, and +1. These features are seen in
'

Fig. 3, where we also show the e6'ect of convoluting

EF = -0.1
gTB(E) with 5-function approximants of various orders.
As expected, the convoluted DOS approaches the true
DOS as N increases. Near the singularity at e. = ——,', the tO

convoluted DOS has structure similar to the correspond- O

ing step-function approximant. The benefit gained O

0)
through use of the higher-order functions is evident when ED 2, 3,4, 5
g c

compared with the zero-order approximation both where o I

OP
the DOS deviates from linearity and near the singularity. J3
E
In the regions where the DOS is almost linear, all orders
work equally well. The smoothing near the singularity is
reduced if a smaller broadening is used; then Sz(x) varies
more strongly in k space and a finer k mesh is needed. 10 20 30 40 10 20 30 40 10 20 30 40
Thus the choice of broadening is determined by the dis- Nk Ng Nk
tance of EF from the Van Hove singularities and by the
amount of variation of the DOS near EF. The order X is FICi. 4. The k convergence of the charge within the Fermi
surface for the simple-cubic tight-binding band for three choices
determined by the required precision: small X requires
of E+ as indicated. Nz is the number of divisions along each re-
fewer k points at the expense of cog.vergence to a less-
ciprocal lattice vector. %&=10 and 40 correspond to 56 and
precise result, higher orders guarantee convergence to the 1771 irreducible k points, respectively; LTM denotes linear
correct result but require more k points. tetrahedron method; other curves result from successive ap-
For three choices of EF, Fig. 4 illustrates the k conver- proximations SN. Each division on the ordinate represents
gence of the charge within the Fermi surface using the 0.001 electrons (the band is normalized to one electron). The
present method and the linear tetrahedron method. energy broadening is 0. 10.
3620 M. METHFESSEL AND A. T. PAXTON

the considerations of Secs. II and III, the value to which resolve it properly. The DOS of hcp Zr is very steep at
the Nth curve converges should depend on the shape of EF; correspondingly, a much smaller broadening of 15
the DOS near the Fermi energy. In the first panel, EF mRy was found suitable. Use of S& with N=2, 3, or 4
lies in the free-electron part of the band where the DOS then gives quite good convergence for a reasonably small
has a large variation. Consequently, there is a marked number of k points. The improvement over the zero-
improvement using S~ with N larger than zero. The order function is pronounced. The comparison to the
DOS is very fIat near EI; for the third panel, so that here linear method shows that the two schemes work about
So is already adequate. Finally, the second panel shows equally we11 here. Similar conclusions apply to the case
the result when EF is chosen exactly on a Van Hove of Nb. The band structure at the Fermi energy is still
singularity. While convergence with increasing N is now more complicated. Therefore neither method is com-
slower, the advantage of using a higher-order Sz is obvi- pletely converged for the largest number of k points we
ous and the present approach still compares favorably have used. However, in applications such as calculations
with the linear method. of frozen-phonon energies it is probably more important
to work on a fIat part of the curve than to have reached
absolute convergence. Then the result is insensitive to
V. APPLICATION TO REALISTIC BANDS
the position of the mesh points relative to the Fermi sur-
To test the scheme for realistic band structures, we face apd consequently no artificial discontinuities are in-
have applied it to a number of the metallic elements. The troduced when the Fermi surface passes over mesh points
aim was also to give rules for choosing 8'and N in prac- as the crystal is perturbed. Use of higher step-function
tical applications. Our examples are fcc Al, hcp Zr, and approximants makes it possible to reach this Aat part and
bcc Nb. The energy bands were generated using the stan- obtain results which are substantially better than with the
dard self-consistent I.MTO method. For each case, we zero-order function. For the linear method, this kind of
have calculated the charge within the Fermi surface and convergence cannot be achieved in practive. We have
the band energy as a function of the fineness of the uni-
form k mesh. The potential and the Fermi energy were
kept fixed at the values given by the standard program.
Again, the linear tetrahedron method was used for com- Al (W=65 re Ry ) Zr(V/=15mRy) Nb (W=35rnRy)
parison. For each metal, various choices of broadening N= 0..

were tried and one was selected which was characteristic

LTM
of the range which seemed most suited. One can only 0
give rules of thumb for this choice, since it depends on 40
O
«1.
2
the shape of the bands in question. The most difFicult sit- O
3'
uations are those in which both steep free-electron-like tD

and Oat tight-binding-like bands cross the Fermi surface. O

5
0 —1"
For each metal considered, we have chosen a broadening L
OP —
—2-
8' such that the first 2 —3 approximants are well con- I 3. .

verged with relatively coarse meshes to their own values,

while the next few higher approximants converge abso- N=O-
lutely (i.e., to the true value) with the use of a fine mesh.
One then has the freedom to choose an N that will simul-
N=O
taneously optimize the mesh size and the absolute con-
vergence required for the application at hand. For the
higher orders (say, N) 5) we have found that similar con-
10
vergence properties can be obtained by taking the next-
lower order and a slightly smaller broadening. For the
low orders, this is not true: use of Sz with some low N -2
——1
gives a gain in precision together with fast k convergence
which cannot be obtained by simply making the zero- 5,10,
order function narrower (as anticipated in Sec. III).
The results obtained with our chosen broadenings are
shown in Fig. 5. For the free-electron metal Al, behavior
was found which is similar to the low-energy range in the 10 20 30 40 10 20 30 40 10 20 30 40
Nk Nk Na
simple-cubic tight-binding band. Because of the wide
bands and smooth DOS, a large broadening of 65 mRy
could be used. As before, a marked improvement is FIG. 5. The k convergence of the charge Q and band energy
found if higher-order step functions are used in place of U at fixed EF for the elemental metals Al, Zr, and Nb. Abscissa
So. For this case, the present scheme is superior to the and labeling of curves as in Fig. 4. Nk = 10 corresponds to 47 ir-
tetrahedron method. reducible k points in Al and Nb and to S4 in Zr, while Nk =44
The situation is qualitatively different for transition corresponds to 2168 and 4232, respectively. Spacings between
metals because of the narrow d bands. The Fermi surface markers on the ordinates are 0.01 electrons for the Q curves and
is more complicated and more k points are needed to l mRy for the U curves. Note the expanded scale for the Al g.
 HIGH-PRECISION SAMPLING FOR BRILLOUIN-ZONE. . . 3621

done similar tests for bcc Mo and paramagnetic bcc Fe. zone precisely, using straightforward sampling tech-
The DOS for Mo has a wide smooth valley near the Fer- niques such as the special-points method. Comparison to
mi energy and a choice of 8'=65 mRy gives good con- results obtained using standard Gaussian smearing show
vergence similar to the Zr results. The case of Fe is simi- the superiority of the present approach. Tests for realis-
lar to Nb and 8' near 20 mRy is suitable; both the tic metal band structures indicate that the scheme is
present approach and the linear tetrahedron method need better suited than the linear tetrahedron method for
a very fine mesh to attain absolute convergence here. free-electron metals and about equally successful for tran-
The preceding results indicate that two kinds of appli- sition metals. Our method has advantages common to all
cation are conceivable. For high-precision calculations, sampling schemes, namely (a) all k points are treated in-
it is necessary first to optimize the values of 8' and X, dependently so that it is extremely easy to implement and
once, for a typical band structure. For this purpose we no large disk files with eigenvalues and eigenvectors are
have found it instructive to make plots of the integrals needed, and (b) no errors are introduced by band cross-
against the order 1V for diA'erent finenesses of the mesh ings. The scheme is equally appropriate for k-dependent
and with various fixed broadenings O'. On the other matrix elements as for densities of states and single-
hand, it suKces to be aware in many instances that particle sums.
without any exhaustive testing, an extension of the zero- It is envisaged that the method may be employed on
order smearing method by employing the first- or two levels of complexity. (i) An immediate improvement
second-order approximant to 5(x) or S(x) will give a over standard smearing methods may be obtained, simply
significant enhancement in precision at the expense of by use of the first- or second-order approximants to the
minimal extra e6'ort. step or 5 function. (2) By careful choice of the parame-
ters W and X very high precision in BZ integrations may
VI. SUMMARY AND CONCLUSIONS be achieved which was hitherto unobtainable.
We have introduced a systematic set of smooth approx- ACKNO%'LEDGMKNTS
imants to the step and 5 functions. When convoluted
with a given function, these approximants give no error if We thank O. K. Andersen for helpful discussions.
the function satisfies a well-defined smoothness property. A. T.P. acknowledges the financial support of the Science
This makes it possible to evaluate Fermi-surface integrals and Engineering Research Council of the United King-
as well as integrals over the occupied part of the Brillouin dom.

'Present address: Fritz-Haber-Institut, Faradayweg 4-6, 1000 4K.-M. Ho, C.-L. Fu, B. N. Harmon, W. Weber, and D. R.
Berlin 33, Federal Republic of Germany. Hamann, Phys. Rev. Lett. 49, 673 (1982); C.-L. Fu and K.-M.
'M. Abratnowitz and I. A. Stegun, Handbook of Mathematical Ho, Phys. Rev. 8 28, 5480 (1983).
Functions (Dover, New York, 1972). 5R. J. Needs, R. M. Martin, and O. H. Nielsen, Phys. Rev. B 33,
2M. Methfessel, M. H. Boon, and F. M. Mueller, J. Phys. C 16, 3778 (1986).
L949 (1983). T. S. Chihara, An Introduction to Orthogonal Polynomials
O. Jepsen and O. K. Andersen, Solid State Commun. 9, 1763 (Gordon and Breach, New York, 1978}.
(1971); G. Lehmann and M. Taut, Phys. Status Solidi 8 54, 7D. J. Chadi and M. L. Cohen, Phys. Rev. B 8, 5747 (1973).
469 (1972); It has recently been shown that convergence like 8T. Morita and T. Horiguchi, J. Math. Phys. 12, 981 (1971).
h can be obtained in the linear tetrahedron method for cer- O. K. Andersen, Phys. Rev. 8 12, 3060 (1975); in The Electron-
tain functions F(E;) (P. E. Blochl, Ph. D. thesis, University of ic Structure of Complex Systems, edited by P. Pharisean and
Stuttgart, 1989). W. M. Temmerman (Plenum, New York, 1984},p. 11.