Professional Documents
Culture Documents
Contents
1 Introduction
2.1
Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2
2.3
15
3.1
3.2
3.3
18
4.1
4.2
21
5.1
5.2
Laplace transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.3
Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
25
28
7.1
7.2
Recursive methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
8 Summary
36
2
9 Appendix
37
Introduction
At the dawn of quantum mechanics physicists realized that not every partial
differential equation nature has to offer can be solved analytically. Even the
most fundamental equation of quantum mechanics, the Schrodinger equation,
can be solved analytically only for very few systems. As Dirac put it in 1929,
the underlying physical laws necessary for the mathematical theory of a
large part of physics and the whole of chemistry are thus completely known,
and the difficulty is only that the exact application of these laws leads to
equations much too complicated to be soluble. Thus, to apply quantum
mechanical concepts to generic molecular systems one has to find some other
way of solving the complicated equations.
One approach is to solve differential equation numerically, via introduction
of a finite grid. This reduces the problem to having to solve a (large) system
of linear equations, a problem well suited for solving on digital computers.
However, high accuracy via this approach can be achieved only with non
uniform grids, with high density of points around the nuclei.
An alternative approach, more intuitively obvious for chemists familiar with
the concept of atomic orbitals (physicists familiar with Bloch functions), is
to expand the solution in terms of functions that behave like the solution,
atomic orbitals in the chemists case (plane waves in the physicists). Again,
this simplifies the problem down to solving matrix equations, with one complication being that the matrix elements become more or less nontrivial to
compute. The matrix elements of the Hamiltonian operator consist of several
types of molecular integrals, and the evaluation of these integrals is the main
focus of this document.
Of course, one has to anticipate what the exact solution looks like, but it turns
out that the fundamental properties of atomic and molecular wavefunctions,
such as nuclear and electronelectron cusps, exponential decay at infinity,
etc. are known and it is relatively easy to find trial functions to satisfy
4
them. Atom-centered basis sets help to describe the critical region around
the nuclei efficiently, hence these have been used in a majority of quantum
chemical studies of molecules. (Homework problem find a study that didnt
utilize atom-centered functions...)
The choice of a particular functional form for the expansion is more difficult.
If one had to solve the Schrodinger equation for the hydrogen atom in finite
basis, the natural choice would be to choose functions of the form
= rL er YLM (, )
(1.1)
that resemble the wellknown discretespectrum solutions found in any textbook on quantum mechanics. Introduced by Slater, this type of basis functions is particularly well suited for atomic calculations. However, the applicability Slatertype functions to other than trivial molecular problems is
hindered by the enormous computational complexity of the resulting expressions for matrix elements of the Hamiltonian.
Boys found a more tractable choice in 19501 as he introduced Gaussiantype
functions of the form
2
= xl y m z n er
(1.2)
and they have been with us ever since. Much simpler expressions for the
matrix elements more than compensate for improper behavior of Gaussians
at the origin and infinity. Indeed, an s-type Gaussian (a Gaussian functions
with l + m + n equal 0) is smooth at the origin, whereas an s-type Slater-type
function has a cusp at the origin (non-zero derivative with respect to r). Also,
Gaussian-type functions decay with r much faster that Slater-type functions.
However, a fact that a given Slater-type functions can be well represented
as a linear combination of only few Gaussians with different exponents was
noticed early on, thus STOnG basis sets were introduced, in which a single
Slater-type function is represented as a linear combination of n Gaussians.
One important note to make before we move on to mathematical details is
1
the value of efficient algorithms for computing integrals. Normally, in a typical high accuracy calculation only a small portion of CPU time is spent in
computing molecular integrals, and the major part is spent in computing
wavefunction parameters. However, the situation has changed dramatically
since the late 1980s. Direct methods which do not store the integrals but recompute them as needed put the integrals evaluation process into a spotlight
again. OK, enough said, on to the integrals!
2
2.1
Normalization
is called a contracted Cartesian Gaussian function. Standard notation. A corresponding primitive spherical harmonic, or pure angular momentum, Gaussian is simply
2
(r) = rL er YLM (, )
(2.3)
Note that we use capital letters for the angular momentum number and
the projection of angular momentum on z-axis to distinguish those from
2
Angular momentum is a slippery term here, some authors just use orbital quantum number
(2.8)
Full derivation of these integrals (Ix , etc) is left to the reader (otherwise this
would not be a learning experience). The result is
Z
(2l
1)!!
Ix =
dxx2l e2x =
.
(2.9)
(4)l 2
Spherical harmonic Gaussians arent as nice in this respect, thats why normally all integrals are computed in
the Cartesian basis and then transformed into the spherical harmonic basis.
1)!!(2m
1)!!(2n
1)!!
SO = N 2
=1
(4)(l+m+n) (2)3/2
(2.10)
N=
(2.11)
(2.12)
Since the bracketed term contains a product of two polynomials, only two
types of terms can result; the square of each uncontracted function and the
biproducts of different uncontracted functions. Take an example where n is
three:
2
In this case, as in all others, there are only two types of terms of which the
integral needs to be taken. They may be written and evaluated as
3/2
Z
2 2i r2
2
1.
ai e
dr = ai
(2.15)
2i
3/2
Z
2
.
(2.16)
2.
2ai aj e(i +j )r dr = 2ai aj
i + j
It is realized that 1. above can be obtained by setting i = j in 2., and
henceforth only the general case needs to be considered. Generalizing to arbitrary n is straightforward, and so the normalization of contracted Gaussian
functions can proceed as
Z
a21
21 2
2 3/2
(r)(r)dr = N
+ +
+
(21 )3/2
(1 + 2 )3/2
n
n X
X
ai aj
2 3/2
= N
= 1,
(2.17)
3/2
(
+
)
i
j
j
i
thus the normalization constant for the entire contraction will be
" n
#1/2
X
ai aj
N = 3/4
(i + j )3/2
i,j
(2.18)
h
i
l m n 1 r2
l m n 2 r2
(r) = N a1 x y z e
+ a2 x y z e
+
l m n
= Nx y z
n
X
ai ei r
(2.19)
(r)(r)dr = N 2
"
x2l y 2m z 2n
n
X
ai ei r
(r)(r)dr = N 2
n
n X
X
n
X
#
aj ej r
dr(2.20)
Z h
i
2l 2m 2n i r2 j r2
ai aj
x y z e
e
dr (2.21)
i=0 j=0
.
(r)(r)dr =
l+m+n+3/2
2l+m+n
(
+
)
i
j
i,j
(2.23)
Calling l + m + n = L, the angular momentum of the shell, and solving for
N,
Z
n
N 2 3/2 (2l 1)!!(2m 1)!!(2n 1)!! X
ai aj
=
=1
L+3/2
2L
(
+
)
i
j
i,j
(2.24)
"
n
X
i,j
ai aj
(i + j )L+3/2
#1/2
(2.25)
Before we approach one-electron integrals, we need to consider one very important result.
10
2.3
The Gaussian Product Theorem states that the product of two arbitrary
angular momentum Gaussian functions on centers A and B can be written
as
G1 G2 = G1 (r, 1 , A, l1 , m1 , n1 )G2 (r, 2 , B, l2 , m2 , n2 )
= exp[1 2 (AB)2 /]
"l +l
#
1
2
X
2
fi (l1 , l2 , PAx , PBx )xiP exP
"mi=0+m
1
2
X
#
2
j=0
"n +n
1
2
X
#
2
(2.26)
k=0
2
2 rB
(2.27)
.
(2.28)
G1 = e1 rA ;
G2 = e2 rB .
(2.29)
e1 rA e2 rB = KerP .
11
(2.30)
e1 rA 2 rB = exp[(1 + 2 )r r + 2(1 A + 2 B) r
1 A A 2 B B]
= K exp[(r r r P + P P)]
(2.31)
(2.32)
Comparing terms,
= 1 + 2
P = (1 A + 2 B), thus
P=
1 A + 2 B
(2.33)
(2.34)
(2.35)
From Eqn (2.33), we expand P P and use that to get a final expression for
K,
P P = 1 12 A A + 21 2 A B + 22 B B
(2.36)
K = exp 1 A A 2 B B + (12 A A + 21 2 A B + 22 B B)/
= exp (12 A A 1 2 A A 1 2 B B 22 B B
+12 A A + 21 2 A B + 22 B B) 1
2
= e
[1 2 (AB )/]
(2.37)
(2.38)
For more general Cartesian Gaussians, ones with arbitrary angular momentum,
(AB)2 /) rP2
G1 G2 = xlA1 xlB2 yAm1 yBm2 zAn1 zBn2 |e(1 2{z
(2.39)
}e
K
12
where weve used Eqn (2.38) to take care of the product of the exponentials.
Now, xlA1 , xlB2 and the like need to be considered.
xlA1 xlB2 = (x Ax )l1 (x Bx )l2
(2.40)
(2.41)
(xP (PA)x ) =
l1
X
(xP )
(PA)lx1 i
i=0
1
X
l1 !
=
(xP )i (PA)lx1 i
i!(l1 i)!
i=0
l1
i
(2.42)
Likewise,
l2
l2
(x Bx ) = (xP (PB)x ) =
l2
X
(xP )
(PB)lx2 j
j=0
l2
j
.
(2.43)
lX
1 +l2
(2.44)
k=0
(2.45)
i=0,l1 j=0,l2
2i = k + q
2j = k q
increments of 2
13
Whence we write the full Gaussian Product Theorem as Eqn (2.26). The
derivation of Eqn (2.46) is left to the reader.
14
3
3.1
(3.1)
(3.2)
= e1 2 (AB)
Z
e
x2P
Z
dx
S12
3.2
yP2
Z
dy
ezP dz
(3.3)
3/2
1 2 (AB)2 /
= e
(3.4)
n
X
2
i rA
ai e
, 2 (r) = N2
Z
S12 =
1 (r)2 (r)dr = N1 N2
m
X
bj ej rB
(3.5)
j
n X
m
X
i
Z
ai b j
(3.6)
i j (AB)2 /ij
= e
ij
15
ei rA ej rB dr
(3.7)
(3.8)
n X
m
X
i
3.3
i A+j B
.
ij
So
ai bj ei j (AB)
/ij
ij
3/2
(3.9)
Z lX
1 +l2
(3.13)
i=0
lX
1 +l2
Z
fi (l1 , l2 , PAx , PBx )
i=0
xiP exP dx
(3.14)
Noting that
odd value of i produces a zero integral, and then using Eqn
R any
i x2P
(2.22) for xP e
dx, we finally write Ix as
(l1 +l2 )/2
Ix =
X
i=0
(2i 1)!!
f2i (l1 , l2 , PAx , P Bx )
(2)i
1/2
(3.15)
The last remaining case is the overlap of two contracted Gaussians of arbitrary angular momentum is left out of the consideration. It is not difficult
16
to write simple routines to compute overlap integrals over primitive Gaussian functions of arbitrary angular momentum using Eqns (3.12) and (3.15)
and use those routines in the evaluation of overlap integrals over contracted
functions.
17
(4.2)
Now we need to determine the action of the Lagrangian (or any piece thereof)
on a particular Gaussian function. Sequentially applying the differential operator,
l2 2 x2B
2
2
(xB e
) = l2 xlB2 1 e2 xB 22 xlB2 +1 e2 xB
(4.3)
x
l2 2 x2B
2
2
( (xB e
)) = l2 (l2 1)xlB2 2 e2 xB 22 (2l2 + 1)xlB2 e2 xB
x x
2
(4.4)
+422 xlB2 +2 e2 xB
2
l2 (l2 1) l2 2 2 x2B
1
2
(xlb2 e2 xB ) =
xB e
2
2 x
2
2
2
+2 (2l2 + 1)xlB2 e2 xB 222 xlB2 +2 e2 xB
(4.5)
Clearly, this is just a sum of three Gaussian functions related to the original
by a shift of 0, 2, or -2 in the angular momentum portion, aside from some
constants.
18
4.1
Simply applying the results shown in Eqn (4.5) within Eqn (4.1) gives a form
of Tij which appears as a sum of three overlaptype integrals with various
multiplicative constants. To display the particular overlap integrals involved
in that sum we will use a particular notation derived from the bra and ket
notation common in physics. Let hn| denote a Gaussian where the angular
momentum has been increased or decreased by n in the coordinate. In other
words,
2
h+2|x = xl+2 y m z n er
(4.6)
Thus, given that the overlap between two Gaussians G1 and G2 is
Z
G1 G2 = h0|0i,
(4.7)
l2 (l2 1)
h0| 2ix
2
(4.8)
1 (r) 2 2 (r)dxdydz
(4.9)
Ix =
2
x
and integrating by parts in x,
"Z Z
#
+
ZZZ
1
2 (r)
1 (r) 2 (r)
Ix =
1 (r)
dydz
dxdydz
2
x
x
x
(4.10)
19
The first term is of course zero because both 1 (r) and 2 (r)/x go to zero
as x . So
ZZZ
1
1 2
Ix =
dxdydz
(4.11)
2
x x
Recalling Eqn (4.3),
ZZZ h
i
1
2
l1 1
l1 +1
yAm1 zAn1 e1 rA
l2 xA 21 xA
Ix =
2h
i
2
l2 1
l2 +1
l2 xB 22 xB
yBm2 zBn2 e2 rB dxdydz.
(4.12)
Thus we can reduce this cumbersome notation to something a little more
convenient to code up.
Ix =
1
l1 l2 h1| 1ix + 21 2 h+1| + 1ix
2
1 l2 h+1| 1ix 2 l1 h1| + 1ix
(4.13)
20
5.1
Since the potential energy is due to Coulomb interaction of the nuclei with
the electron in question, the operator to deal with is r1C , where rC = |r C|.
Thus the integral we need to evaluate is
Z
1
C
Vij =
i (r) j (r)dr
rC
Z
2 1
2
=
xlA1 yAm1 zAn1 e1 rA xlB2 yBm2 zBn2 e2 rB dr
(5.1)
rC
Since the operator does not affect the operand (), we can combine the two
Gaussian functions via the Gaussian product theorem, and make the final
statement
XXX
VijC = K
fl (l1 , l2 , PAx , PBx )fm (m1 , m2 , PAy , PBy )
l
Z
fn (n1 , n2 , PAz , PBz )
1
dr
rC
(5.2)
21
and result in resolution of the variables. There are many possibilities the
most commonly used are Laplace or Fourier transforms. Lets look at the
Laplace transform solution in detail.
5.2
Laplace transform
1 Z
2
= ( )
esr s/21 ds,
2
(5.4)
where (x) is the standard gamma function. You can just evaluate the RHS
to confirm this. We want the instance where = 1, thus
1
= 1/2
rC
(5.5)
What occurs when we use this in the context of a potential energy integral
involving only sfunctions?
Z
2
2 1
V =
e1 rA e2 rB dr
rC
Z
2 1
= K erP dr
rC
Z
Z
2
2
= K 1/2 erP esrC s1/2 dsdr
(5.6)
0
Conveniently, the Laplace transform takes rC1 into a function with the appearance of an s-type Gaussian of orbital exponent s centered at C. A second
application of the GPT and we can switch the order of integration, evaluating
22
dss1/2
erP esrC dr
= K 1/2
dss1/2 esPC
/(+s)
dre(+s)rD
(5.7)
Z
= K
/(+s)
(5.8)
0
s
, ds = 2 s1/2 ( + s)3/2 dt amazingly
Now making the substitution t2 = (+s)
cancels just about everything leaving
Z1
2 2
2K
V =
ePC t dt
(5.9)
0
3/2
V =
K
1/2
PC), P 6= C
erf(
3/2 PC
(5.11)
or as
V =
2K
, P=C
(5.12)
23
Note that
1/2
erf( T ), T > 0
2 T
F0 (0) = 1
F0 (T ) =
5.3
(5.14)
(5.15)
Fourier transform
The Fourier transform approach is essentially very similar, the only difference
is that r1C is represented as a three-dimensional integral:
ZZZ
1
1
= 2
k 2 eikrC dk
(5.16)
rC
2
A substitution of Eqn (5.16) into Eqn (5.2) and a reversal of the order of integration allow the integral over r to be separated into three one-dimensional
integrals. The resulting derivation is a little bit longer than in the case of
Laplace transform, and the final expression looks a bit different, but both
formulae give the same values, and thats what counts.
24
i (r1 )
j (r1 )
k (r2 )
l (r2 )
=
=
=
=
(6.2)
np ,nq 1 ,2 t000
where
= lp + lq + np + nq + mp + mq 2u1 2u2 2v1 2v2
2w1 2w2 t0 t00 t000
2
K1 = exp(1 2 AB /p )
2
K2 = exp(3 4 CD /q )
0
G(x) = (1)lp +t flp (la , lb , P Ax , P Bx )flq (lc , ld , QCx , QDx )
p q lp +lq 2u1 2u2 t0
0
)
(P Qx )lp +lq 2u1 2u2 2t (
p + q
0
(p )u1 lp (q )u2 lq lp !lq !(lp + lq 2u1 2u2 )!4u1 u2 t
u1 !u2 !(lp 2u1 )!(lq 2u2 )!(lp + lq 2u1 2u2 2t0 )!(t0 )!
Z1
Fm (T ) =
u2m exp(T u2 )du
0
0 lp la + lb
0 u1 lp /2
0 lq lc + ld
0 u2 lq /2
2 5/2 K1 K2
2
F
(PQ
p q /(p + q ))
0
p q (p + q )1/2
(6.5)
Computing ERI of any but (ss|ss) type in closed form is not very practical. Coding it up is relatively easy, but the efficiency is very poor,
especially if you need to compute more than just one integral.
Eqn (6.4) may be rewritten as
I=
L
X
Cm Fm (T )
(6.6)
m=0
2
27
Up to this point weve used a notation system that we feel is the most suitable
for manipulating closed form expressions for one-electron integrals. To move
on further we have to enhance our notation to bring it in accordance with
the one used in the literature. Following Obara and Saika 6 , we write the
unnormalized Cartesian Gaussian function centered at R as
(r; , n, R) = (x Rx )nx (y Ry )ny (z Rz )nz
exp[(r R)2 ] ,
(7.1)
28
7.1
N
X
(7.2)
Of course, we want to compute a six dimensional integral, and our approach will have to be a little bit different (why cannot we just use a formula
analogous to Eqn (7.2)? Read the answer at the bottom of this page).7 The
trick is to represent an ERI as a one-dimensional integral over some function.
Lets start with Eqn (6.6). After plugging the definition of the incomplete
gamma function in and moving the summation sign inside the integral we
obtain
Z1 X
L
I =
Cm t2m exp(T t2 )dt
0
Z1
=
m=0
(7.3)
Kl (x)w(x)dx
(7.4)
where Kl (x) is a polynomial of degree l, w(x) is positive on the interval [a, b].
The theory of orthogonal polynomials offers a way of computing this type of
7
29
integrals exactly. Let a set of polynomials {Sn (x)} be defined on [a; b] and
orthogonal in the sense that
Zb
Sn (x)Sm (x)w(x)dx = mn ,
(7.5)
then integral (7.4) can be evaluated exactly by an n-point numerical quadrature formula:
n
X
I=
Kl (ti )Wi ,
(7.6)
i=1
Li (x)w(x)dx ,
(7.7)
(7.8)
All the glorious details with the proofs can be found in any textbook on
numerical methods.
Fortunately, there exists a set of polynomials {Ri (t, T )}, known as Rys polynomials, that satisfy all necessary conditions:
Rn (t, T ) is an even polynomial of degree 2n in the variable t;
For any real T there exists an infinite set of such entities orthogonal to
each other:
Z1
Rn (t, T )Rm (t, T ) exp(T t2 )dt = mn
(7.9)
0
30
n
X
PL (ti )Wi ,
(7.10)
i=1
where n is an integer greater than L/2, ti is a positive zero of the n-th Rys
polynomial
Rn (ti , T ) = 0 ,
(7.11)
and Wi is the corresponding weight which depends on T . Since Wi depends
on T only, and every integral in a given class of integrals will share the
same set of Wi , it is beneficial to use Eqn (7.10) to compute whole classes of
integrals rather than a single integral. In fact, all practical strategies of ERI
evaluation employ this so-called shell structure of integrals.
What are potential advantages of using Eqn (7.10) instead of Eqn (6.6)?
At first glance none. To compute PL one has to know coefficients Cm ,
computing which is the main obstacle in using Eqn (6.6). Yet it turns out
there exists a way to determine numerical values of PL (ti ) without computing
Cm . The details are too lengthy to be presented here and can be found in a
paper by Dupuis, Rys, and King 8 . The resulting expression is written as
I
N
X
(7.12)
i=1
31
7.2
Recursive methods
= 2(r, , n + 1i , R) ni (r, , n 1i , R)
(7.13)
Ri
to the expression for the (ss|ss) integral given in Eqn. (6.5) to obtain the
corresponding (ps|ss) integral. Hence, multiple applications yield higher angular momentum ERIs.
McMurchie and Davidson proposed recurrence relations for ERIs over Hermite Gaussian functions11 to compute the integrals over Cartesian Gaussians.
Their algorithm is a significant improvement over Boys method, but it requires a special transformation of the resulting integrals over Hermite Gaussians back into the Cartesian basis.
In 1986 Obara and Saika published a new recursive scheme based also on the
differential relation for Cartesian Gaussians (7.13). Their approach treats
one-, two-, and potentially n-electron integrals on equal footing. OS suggested the following recurrence relation for ERIs over primitive Cartesian
Gaussians: 12
[a + 1i , b|cd](m) = P Ai [ab|cd](m) + W Pi [ab|cd](m+1)
ai
(m)
(m+1)
+
[a 1i , b|cd]
[a 1i , b|cd]
2
+
bi
+
[a, b 1i |cd](m)
[a, b 1i |cd](m+1)
2
+
ci
di
+
[ab|c 1i , d](m+1) +
[ab|c, d 1i ](m+1) , (7.14)
2( + )
2( + )
11
Hermite Gaussians are a special form of Gaussian functions which resemble the eigenfunctions of the harmonic
oscillator. Their use is motivated by extreme simplicity of the differential relations for this type of Gaussians. For
more details see ...
12
The presented equation is only one of the four possible cases, the other three being the relations increasing
angular momentum on centers B, C, and D.
32
where
Z 2 m
2
u
(ab|cd)(m) =
du
(ab|u|cd)
(7.15)
+ u2
Z 0
2
(ab|u|cd) =
dr1 dr2 (r1 ; a , a, A)(r1 ; b , b, B) exp(u2 r12
)
(r2 ; c , c, C)(r2 ; d , d, D) ,
(7.16)
and
= a + b
= c + d
=
+
a A + b B
P =
a + b
c C + d D
.
Q =
c + d
(7.17)
(7.18)
(7.19)
(7.20)
(7.21)
(7.22)
(m)
(m+1)
+
[a 1i , 0|c0]
[a 1i , 0|c0]
2
+
ci
+
[a0|c 1i , 0](m+1) .
2( + )
(7.23)
(7.24)
The prefactors in (7.23) still depend on the exponents of the Gaussian functions, and therefore cannot be applied to ERIs over contracted functions.
Prefactors in HRR depend on the geometric variables only, which are common for all primitive functions in a contraction. Hence, HRR can be applied
to contracted ERIs. The computation strategy has to be adjusted accordingly:
1. For each combination of primitive Gaussians [ab|cd] do:
(a) Compute Fm (T );
(b) Apply VRR to build primitive [e0|f 0], where a e a + b ,
c f c + d ;
13
34
35
Summary
36
Appendix
Let us derive the explicit expression for the electron repulsion integral over
primitive Gaussians using the Fourier transform for the r112 factor:
ma na
2
a
xl1A
y1A
z1A exp(1 r1A
)
m b nb
2
b
xl1B
y1B
z1B exp(2 r1B
)
lc mc nc
2
x2C y2C z2C exp(3 r2C )
md nd
2
d
xl2D
y2D
z2D exp(4 r2D
)
Z
1
I =
i (r1 )j (r1 ) k (r2 )l (r2 )dr1 dr2
r12
ZZZ
1
1
=
k 2 eikr12 dk
2
r12
2
i (r1 )
j (r1 )
k (r2 )
l (r2 )
=
=
=
=
37
(9.25)