Download Hartree-Fock Method: Self-Consistent Fields and Orbital Optimization and more Lecture notes Chemistry in PDF only on Docsity!
MODERN ELECTRONIC STRUCTURE THEORY:
Electron Correlation
In the previous lecture, we covered all the ingredients necessary to choose a
good atomic orbital basis set. In the present lecture, we will discuss the
other half of accurate electronic structure calculations: how we compute the
energy. For simple MO theory, we used the non-interacting (NI) electron
model for the energy:
E
NI
= E
μ
μ= 1
N
∑
= ψ μ
(^1 )
H
∫^0
μ= 1
N
∑
ψ μ
(^1 ) d^ τ
Where, on the right hand side we have noted that we can write the NI
energy as a sum of integrals involving the orbitals. We already know from
looking at atoms that this isn’t going to be good enough to get us really
accurate answers; the electron-electron interaction is just too important.
In real calculations, one must choose a method for computing the energy
from among several choices, and the accuracy of each method basically boils
down to how accurately it treats electron correlation.
Self Consistent Fields
The Hartree Fock (HF) Approximation
The Hartree-Fock method uses the IPM energy expression we’ve already
encountered:
E
IPM
= E
μ
μ= 1
N
∑
J
μν
K
μν
μ< ν
N
∑
E
μ
= ψ μ
(^1 )
H
0 ∫
ψ μ
(^1 ) d^ τ
J
μν
μ
( )
ν
( )
r
1
− r
2
μ
( )
ν
( )
d r
1
d r
2
d σ
1
d σ
∫∫ 2
K
μν
μ
(^1 )^ ψ ν
( 2 )
r
1
− r
2
μ
( 2 )^ ψ ν
(^1 ) d r 1
d r
2
d σ
1
d σ
∫∫ 2
Since the energy contains the average repulsion, we expect our results will
be more accurate. However, there is an ambiguity in this expression. The
IPM energy above is correct for a determinant constructed out of any set of
orbitals { (^) μ}
y and the energy will be different depending on the orbitals we
choose. For example, we could have chosen a different set of orbitals, { ' } μ
y
, and gotten a different energy:
E '
NI
= E '
i
μ= 1
N
∑
J '
μν
K '
μν
μν
N
∑
How do we choose the best set of orbitals then? Hartree-Fock uses the
variational principle to determine the optimal orbitals. That is, in HF we find
the set of orbitals that minimize the independent particle energy. These
orbitals will be different from the non-interacting orbitals because they will
take into account the average electron-electron repulsion terms in the
Hamiltonian. Thus, effects like shielding that we have discussed
qualitatively will be incorporated into the shapes of the orbitals. This will
tend to lead to slightly more spread out orbitals and will also occasionally
change the ordering of different orbitals (e.g. s might shift below p once
interactions are included).
Now, the molecular orbitals (and hence the energy) are determined by their
coefficients. Finding the best orbitals is thus equivalent to finding the best
coefficients. Mathematically, then, we want to find the coefficients that
make the derivative of the IPM energy zero:
∂ E
IPM
∂ c
μ
∂ c
μ
E
μ
μ= 1
N
∑
J
μν
K
μν
μν
N
∑
After some algebra, this condition can be re-written to look like an
eigenvalue problem:
[ ] E
μ μ
μ
F c c = S c Eq. 1
The matrix F is called the Fock matrix and it depends on the MO
coefficients:
F
ij
≡ φ i
AO ˆ H 0
φ j
AO
d τ ∫
AO
( 1 ) ψ μ
( 2 )
r 12
φ j
AO
( 1 )ψ μ
( 2 ) d τ 1
d τ ∫ 2
μ= 1
N
∑
φ i
AO
( 1 ) ψ μ
( 2 )
r 12
φ j
AO
( 2 ) ψ μ
( 1 ) d τ 1
d τ ∫^2
The first term is just the independent particle Hamiltonian that we’ve been
dealing with in cruder forms of MO theory. The second and third terms are
new contributions that arise from the average electron-electron repulsion
present in our system. These terms reflect the fact that, on average, each
electron will see not only the nuclear attraction (first term) but also the
average Coulomb (second) and Exchange (third) interactions with all the
remaining electrons. For this reason, HF is also known as “mean field”
theory, because it includes both the interaction of each electron with the
lowest possible energy, because then 0
dE
d
c
. At this point, the Hamiltonian
will be self-consistent: the orbital coefficients we used to build F and define
H
eff
will be the same as the ones we get from the eigenvectors of H eff
. For
this reason, these iterations are called self-consistent field (SCF)
iterations. Physically, these iterations locate a set of MOs that are
eigenstates of the averaged electrostatic repulsion that they, themselves,
generate.
One can go on to define atomic charges, bond orders and other qualitative
metrics within the HF approximation. Before talking about the practical
accuracy of HF, though, we first discuss a closely related method.
Density Functional Theory (DFT)
Here, we still use a Slater determinant to describe the electrons. Hence,
the things we want to optimize are still the MO coefficients c
μ
. However, we
use a different prescription for the energy – one that is entirely based on
the electron density. For a single determinant, the electron density, r( r ) is
just the probability of finding an electron at the point r. In terms of the
occupied orbitals, the electron density for a Slater Determinant is:
( ) ( )
2
1
N
μ
μ
r y
=
å
r r Eq.^2
This has a nice interpretation: (^) ( )
2
μ
y r is the probability of finding an
electron in orbital μ at a point r. So the formula above tells us that for a
determinant the probability of finding an electron at a point r is just the
sum of the probabilities of finding it in one of the orbitals at that point.
There is a deep theorem (the Hohenberg-Kohn Theorem) that states:
There exists a functional E v
[ r ] such that, given the ground state
density, r 0 , E v
[ r 0 ]=E 0
where E 0
is the exact ground state energy.
Further, for any density, r ’ , that is not the ground state density,
E
v
[ r ’]>E 0
This result is rather remarkable. While solving the Schrödinger Equation
required a very complicated 3N dimensional wave function Yel( r 1 , r 2 ,… r N), this
theorem tells us we only need to know the density - which is a 3D function! –
and we can get the exact ground state energy. Further, if we don’t know the
density, the second part of this theorem gives us a simple way to find it:
just look for the density that minimizes the functional E v
The unfortunate point is that we don’t know the form of the functional E v
We can prove it exists, but we can’t construct it. However, from a
pragmatic point of view, we do have very good approximations to E v
, and the
basic idea is to choose an approximate (but perhaps very, very good) form
for E v
and then minimize the energy as a function of the density. That is, we
look for the point where 0
v
dE
d r
=. Based on Eq. 2 above, we see that r just
depends on the MOs and hence on the MO coefficients, so once again we are
looking for the set of MO coefficients such that (^0)
v
dE
d
c
. Given the
similarity between DFT and HF, it is not surprising that DFT is also solved
by self consistent field iterations. In fact, in a standard electronic
structure code, DFT and HF are performed in exactly the same manner (see
flow chart above). The only change is the way one computes the energy and
dE
d c
. Note that the condition (^0)
v
dE
d
c
can also be cast as a generalized
eigenvalue-like equation:
[ ] E
μ μ
μ
KS
H c c S c
The effective Hamiltonian appearing in this equation is usually called the
Kohn-Sham Hamiltonian.
Now, as alluded to above, there exist good approximation s (note the plural)
to E v
. Just as was the case with approximate AO basis sets, these
approximate energy expressions have strange abbreviations. We won’t go
into the fine differences between different DFT energy expressions here.
I’ll simply note that roughly, the quality of the different functionals is
expected to follow:
LSDA < PBE ≈ BLYP < PBE0 ≈ B3LYP
Thus, LSDA is typically the worst DFT approximation and PBE0 and B3LYP
are typically among the best. I should mention that this is just a rule of
thumb; unlike the case for basis sets where we were approaching a well-
defined limit, here we are trying various uncontrolled approximations to an
out of time. In the first case, the primary challenge is finding enough
memory for the (10A)x(10A)=100A
2
numbers that comprise the Fock matrix.
As a rule of thumb, each number takes up 8 bytes so that the Fock matrix
requires 800A
2
bytes of storage. A typical computer these days has on the
order of 2 Gigabytes (2x
9
bytes) of storage, which will get maxed out if
800A
2
=2x
9
, which happens when A=2500 or so. So, based on storage, we
conclude HF calculations are possible on systems up to a couple thousand
atoms in size. Looking at the time the calculation takes, the key step is
finding the eigenvectors of the Fock matrix. As a rule of thumb, finding the
eigenvectors of an N-by-N matrix takes about 5N
3
operations (add,
subtract, multiply, divide). Thus, finding the eigenvectors of F will require
about 5x(10A)
3
=5000A
3
operations. A modern CPU can do about 3 billion
operations in a second (the clock speed on processors is on the order of 3
GHz). Thus, in an hour a good CPU can do (3600 s/hr)x(3x
9
ops/s)≈1x
13
operations. Thus, the calculation will take an hour when 5000A
3
≈1x
13
,
which again occurs when A≈1200. This estimate is a little bit less accurate
than the storage estimate because we’ve neglected all the other steps in the
SCF procedure, and we’ve also ignored the fact that in the SCF iterations
will require us to find the eigenvectors of F many, many times. Thus, a 1200
atom calculation will actually take quite a bit more than an hour, but that’s
OK because in practice even waiting 1-2 days for an answer isn’t so bad,
because the computer can work while we’re off doing something else.
In any case, whether we look at time or storage, we arrive at the same
conclusion: SCF calculations can be run on molecules with up to a couple
thousand atoms. This opens up a huge variety of chemical systems to ab
initio computation and has led to MO theory calculations becoming par for
the course in many research publications. The limiting factor in expanding
the size of systems one can study with HF or DFT is that the cost grows
non-linearly with the size of the system. If your present computer can
handle A atoms, then it has something like A
2
memory and A
3
CPU speed. In
order to handle 2A atoms, you’ll need (2A)
2
=4A
2
memory and (2A)
3
=8A
3
CPU.
Thus, the computational demands grow faster than the system size – if you
double the system size, you need four times the memory and 8 times the
CPU power. This makes the calculation increasingly difficult for large
systems and one inevitably hits a wall beyond which the calculation is just
too difficult. To circumvent this, many groups are investing time in
developing linear scaling algorithms for MO simulations that hinge on two
points: First, for very large systems, many of the terms in the RH equations
are very, very small (e.g. the interaction between two very distant atoms).
Avoiding computing these irrelevant terms can save an immense amount of
time and storage. Second, by reformulating the SCF iterations, one can
avoid computationally intensive steps like finding eigenvectors in favor of
many less expensive operations (like matrix-vector multiplication). At the
moment, these algorithms are still work in progress, but they will be
required if one wants to be able to simulate truly large molecules and
aggregates in the 100 kDa range.
In any case, HF and DFT are both limited by the same computational
bottlenecks and thus, in practice, the two methods have very similar
computational cost. The same can not be said about the methods’ accuracy,
however, as we now discuss.
Accuracy of SCF Calculations
Even though Hartree Fock and DFT are much more accurate than qualitative
MO theory, they are still an approximation. The accuracy varies depending
on the property one is interested in, and the reliability of HF/DFT for
various predictions has been extensively tested empirically. We now
summarize a few key rules of thumb that come out of those computer
experiments. Note that, as discussed above, the quality of DFT results
depends on the functional used. The figures given below are characteristic
of the B3LYP functional (one of the better forms available). In the future,
as new and better functionals are developed, these error bars may improve.
Ionization Potentials and Electron Affinities: Just as for qualitative MO
theory, we can use SCF calculations to compute the IP and EA of a molecule,
M. There are two ways we can do this, which will give somewhat different
answers. The first is to use Koopman’s theorem. Here we do just one
calculation on the neutral molecule and then estimate the IP (EA) from e HOMO
(e LUMO
). In the second route, we perform two separate calculations on the
neutral and cation (or anion) to compute the IP (or EA) directly from
IP=E[M
]-E[M] (or EA=E[M]-E[M
]). In practice the first route is somewhat
faster, while the second is somewhat more accurate. Taking the more
accurate route, HF predicts IPs and EAs to within about 0.5 eV. This isn’t
perfect, but is close enough to sometimes be a useful guide. DFT on the
inherently unstable. However, once these points have been found, one can
predict the barrier height by computing energy difference between the
minimum and the saddle point. HF is not terribly good at predicting these
barrier heights, typically overestimating them by 30-50%. Unfortunately,
current density functionals are not very good at this, either, typically
underestimating barrier heights by ~25%. The challenge of improving
barrier heights in DFT (without making other properties worse!) is an active
area of research.
Binding Energies One of the most important properties we can predict is the
binding energy of a complex. We can obtain the binding energy of AX by
doing three separate calculations (on AX, A and X). The binding energy is
then ∆E=E(AX)-E(A)-E(X). Unfortunately, HF is quite bad at predicting
binding energies in this fashion, underestimating them by 50% in practice.
DFT on the other hand is very successful at predicting binding energies, with
a typical error of about 0.1 eV
Putting all this data together, we can make a table of expected accuracy for
HF and DFT:
Property HF Accuracy DFT Accuracy
IPs and EAs ±0.5 eV ±0.2 eV
Bond Lengths - 1% ± 1 pm
Vibrational
Frequencies
Barrier
Heights
Bond Energies - 50% ±0.1 eV
Because DFT is essentially universally more accurate than HF but comes at
the same computational cost, DFT calculations have taken over the quatum
chemistry literature.
Correlated Wave Function Approximations
While DFT gives us terrific bang for our buck, the one thing it does not
provide is a convergent hierarchy for doing calculations. We have some
empirical rules of thumb that one functional should be better than another,
but no guarantees. And further, if the best functional (e.g. B3LYP) gives the
wrong answer, we have no avenue by which we can improve it. To obtain
something that is more controlled, we use methods that are commonly
referred to as correlated wave function methods. Here, the idea is to
employ wave functions that are more flexible than a Slater determinant.
This can be done by adding up various combinations of Slater determinants,
by adding terms that explicitly correlate pairs of electrons (e.g. functions
that depend on r 1
and r 2
simultaneously) and a variety of other creative
techniques. These approaches are all aimed at incorporating the correlation
between electrons – i.e. the fact that electrons tend to spend more time far
apart from one another as opposed to close together. This correlation
reduces the average repulsion employed in HF and brings us closer to the
true ground state energy.
First, we must properly frame the task. The best ground state wave
function we have obtained so far comes from HF (in technical terms, DFT
only gives us a good density, not a good wave function). Of course, the HF
wave function is not the true ground state wave function and the HF energy
is not the true ground state energy
Ψ HF
≠ Ψ exact
E HF
≠ E exact
In particular, we know from the variational principle that the HF energy is
above the ground state, so we define the correlation energy as
E corr
= E exact
− E HF
This energy captures the energy difference between our HF estimate and
the correct ground state energy. It is negative (stabilizing) by definition
and physically arises from the fact that electrons are not independent (as in
HF) but move in a correlated fashion so that they avoid one another and thus
minimize the electron-electron repulsion. As a rule of thumb, E HF
is about
99% of the total energy, so that E corr
is only about 1% of the total energy.
However, this 1% determines a whole range of chemical properties (recall
how bad HF is for predicting, say, bond energies) and so it is crucial that we
describe it accurately.
Hence, our goal is to derive the wave function corrections necessary to turn
E
HF
into E corr
- or at least get a better approximation. There are generally two
routes to this: the first treats the missing correlation as a perturbation, the
second seeks to directly expand the wave function and use the tools of
matrix mechanics.
Property HF DFT MP
IPs and
EAs
±0.5 eV ±0.2 eV ±0.2 eV
Bond
Lengths
Vibrational
Frequencies
Barrier
Heights
Bond
Energies
kcal/mol
In practice, one very rarely sees MP3 or MP4 calculations in the literature
for two reasons. The first is cost. If we compile the computational cost of
MP2 as compared to HF or DFT we find a somewhat sobering situation:
Resource HF/DFT MP
Storage
O(N
2
) O(N
3
CPU Time O(N
3
) O(N
4 - 5
Maximum
Feasible N
Maximum
Feasible
Atoms
Here we define the number of basis functions to be N , and keep track of the
highest power of N required by the method with the notation O(N
x
). Thus,
MP2 requires exponentially more time and space than the corresponding SCF
calculation. This is compounded by the fact that, as discussed at the end of
the basis set discussion, an accurate MP2 calculation requires a more
accurate AO basis (e.g. QZ2P) to get the type of precision that HF or DFT
can get in a smaller basis (e.g. DZP). Thus, with an MP2 calculation, one is
limited to much smaller molecules than with DFT or HF. Going to MP3 or
even MP4 requires more storage and more CPU time, making them almost
prohibitively expensive.
The other problem is that the perturbative expansion will only converge if
the perturbation is, in some sense, “small”. A precise characterization of
when the MP series does and does not converge is difficult to come by. By
its nature peturbation theory is what is called an asymptotic expansion, and
it is very difficult to guarantee convergence of these types of expansions.
In practice, one can simply calculate very high orders of the MPn hierarchy
and look for convergence. In a very good case, we might see something like
the figure below, which shows the MPn energies for the BH molecule as a
function of n:
Here, we see that the energy converges monotonically and rapidly with n:
MP2 is within 0.012 au (0.3 eV) of the exact total energy (which is the
dotted line), MP4 is within .005 au (0.1 eV), MP5 is within .002 au (0.03 eV)
…. Meanwhile for N 2
we see
which has somewhat slower convergence as well as some oscillation. For neon
we see very fast near-convergence, followed by divergence(!):
Here, the dots just indicate that the number of electrons is arbitrary. For
any particular number of electrons, the list terminates. Thus, for four
electrons:
Ψ( 1 , 2 , 3 , 4 ) = C pqrs
p < q < r < s
∑
ψ p
( 1 ) ψ q
( 1 ) ψ r
( 1 ) ψ s
( 1 )
ψ p
( 2 ) ψ q
( 2 ) ψ r
( 2 ) ψ s
( 2 )
ψ p
( 3 ) ψ q
( 3 ) ψ r
( 3 ) ψ s
( 3 )
ψ p
( 4 ) ψ q
( 4 ) ψ r
( 4 ) ψ s
( 4 )
Using this expansion, we can represent all wave functions as vectors in this
space and all operators as matrices. Thus, we can find all the energy
eigenstates by solving an orthogonal eigenvalue problem:
This expansion is known as Full CI and it is guaranteed to give us the right
answer – unlike perturbation theory (where there are an infinite number of
terms) here the expansion is finite and there is no question of convergence.
The problem is that the dimension of this eigenvalue problem is very large.
Say we have a benzene molecule in a minimal basis, which has 42 electrons
and around 100 AOs. To enumerate the number of terms in the sum, we
recognize that there are (100 possible values of p ) times (99 possible values
of q ) times (98 possible values of r ) … for a total of unique
basis states! Thus something the size of benzene is out of the question, and
FCI will be limited to very small molecules (like diatomics).
In order to get something more tractable, we first rearrange the expansion.
We note that one of the determinants involved will be the HF determinant,
YHF. There will also be a number of determinants that differ from the HF
state by only one orbital (e.g. orbital i gets replaced by orbital a ). This
singly substituted states will be denoted Y i
a
. Next, there will be
determinants that differ by two orbitals (e.g. i and j get and replaced by a
and b , denoted Yij
ab
). Similarly, we will have triple, quadruple, quintuple …
substitutions which allows us to reorganize the expansion as
Ψ = Ψ HF
a
Ψ i
a
ia
∑
ab
Ψ ij
ab
ia
∑
This expansion will terminate only after we have performed N substitutions
(where N is the number of electrons). Because this is an exact re-writing of
the CI expansion, we have saved no effort in this step. However, this way of
writing things suggests a useful approximation. We recognize that the HF
H ⋅ C
i
= E
i
C
i
≈ 1 x 10
30
state is a pretty good approximation to the ground state. As a result,
determinants that differ by many, many orbitals from Y HF
shouldn’t be that
important in the expansion, and so it is natural to truncate the expansion at
some level. Thus, if we only include Single substitutions, we have CIS. If we
include Double substitutions, too, we have CISD. Adding Triples gives
CISDT and so on. In any of these cases, we still obtain the eigenstates by
solving the linear eigenvalue problem.
However, we have reduced the dimension of our Hilbert space (e.g. so that it
only includes HF plus single and double substitutions) so the dimension of our
matrices is much, much smaller. These forms of truncation allow CI
expansions to be routinely used in quantum chemistry.
There is one other way to truncate CI that is based more on chemical
intuition. First, we recognize that each of the terms above can be
associated with a group of stick diagrams:
H ⋅ C
i
= E
i
C
i
Where Y S
and Y D
are sums of singly and doubly excited sates, respectively.
Next, consider a pair of M molecules very far apart. At the same level of
approximation, the correct dimer wave function will be a product
Ψ M ... M
= Ψ M
= Ψ HF
( ) Ψ HF
( )
= Ψ HF
Ψ HF
Ψ S
Ψ HF
( ) + Ψ HF
Ψ D
Ψ HF
Ψ S
( )
Ψ D
Ψ S
( ) +^ Ψ D
Ψ D
In the last equality, we have grouped the terms according to the number of
substitutions in the dimer: the first term has no substitutions (since it is
just the product of the Hartree-Fock determinants for the two monomers),
the second has singles, the third doubles, the fourth triples and the last
quadruples [Note that the exictation levels add here. Hence, if we have two
orbitals substituted on the left and two substituted on the right, the total
wave function has four substitutions.] Thus, we see that if we include only
double substitutions for the monomer, that implies we need quadruple
substitutions for the dimer (and six-fold substitutions for the trimer, eight-
fold for the tetramer…) Hence, any fixed level of truncation (e.g. to doubles
or triples or quadruples) will fail to describe non-interacting dimer wave
functions correctly. This failure is called the size consistency problem. It
manifests itself in a number of ways. One prominent mistake here is that,
since CISD wave functions never decompose as products, the energies do not
add properly:
In fact, one can even prove the more stringent fact that for very large
systems, the correlation energy per particle predicted by CISD goes to
zero! Thus, one finds strange results like the following: if you compute the
CISD energy of an infinite number copies of molecule A that are all far from
one another, the total energy per molecule will actually come out the same as
the HF energy for A:
This error is called the size extensivity error. It reflects the fact that,
while we know the energy is an extensive quantity, in CISD the correlation
energy is not extensive (i.e. it does not grow linearly with N). Thus, when
the system is large enough, the CISD correlation per particle is negligible.
Coupled Cluster (CC)
E
CISD
( A ... X ) ≠ E CISD
( A ) + E
CISD
( X )
lim
n → ∞
E
CISD
A
n
( )
n
= E
HF
( A )
Because of the problems with size extensivity and size consistency, it is not
meaningful to discuss the quality of results one expects with CI. The rules
would depend strongly on the size of the system and even on the presence or
absence of spectator molecules. Instead, we need to modify the CI
expansion. Toward this end, we define an operator, , that generates all
the substituted configurations in the CI expansion:
ˆ C Ψ HF
≡ C i
a
Ψ i
a
ia
∑
ab
Ψ ij
ab
ia
∑
We can use this operator to generate the entire CISD wave function:
Ψ CISD
= 1 +
ˆ
( C )Ψ HF
≡ Ψ HF
a
Ψ i
a
ia
∑
ab
Ψ ij
ab
ia
∑
As we saw, this CI expansion led to difficulties with size extensivity. Rather
than this linear prescription, Coupled Cluster postulates an exponential form
for the wave function. Thus, the CCSD wave function is
Ψ CCSD
= e
ˆ C
Ψ HF
= 1 +
ˆ C +
1
2
ˆ C
2
( )
Ψ HF
= Ψ HF
ˆ C Ψ HF
1
2
ˆ C
2
Ψ HF
In the last expression, the first term has zero substitutions, the second has
up to double substitutions, the third has up to quadruple substitutions,….
While this exponential form may seem artificial at first, it is easily
confirmed that this format alleviates the size consistency problem. Suppose
molecules A and B are well represented by CCSD
Ψ CCSD
A
= e
ˆ T A Ψ HF
A
Ψ CCSD
B
= e
ˆ T B Ψ HF
B
Then the noninteracting A…B heterodimer will be well-represented by a
product:
Ψ CCSD
A .... B
= Ψ CCSD
A
Ψ CCSD
B
= e
T ˆ A Ψ HF
A
e
T ˆ B Ψ HF
B
= e
T ˆ A e
T ˆ B Ψ HF
A
Ψ HF
B
= e
ˆ TA +
ˆ TB
Ψ HF
A
Ψ HF
B
= e
ˆ TAB
Ψ HF
A ... B
Where on the last line we have made the definitions
ˆ T AB
≡
ˆ T A
ˆ T B
and
Ψ HF
A ... B
≡ Ψ HF
A
Ψ HF
B
. Now, if
ˆ T A
and
ˆ T B
involve at most double substitutions, then
ˆ T AB
≡
ˆ T A
ˆ T B
will also involve at most doubles. Thus, the product of two
noninteracting CCSD wave functions is another CCSD wave function. We
conclude that a truncated CC prescription is capable of describing
noninteracting fragments correctly. As a result, the CCSD energy is
properly size consistent and size extensive and consequently preferable to
CISD for looking at large systems.
There are truncated CC cousins for every CI approximation (e.g. CCS, CCSD,
CCSDT, CCSDTQ…). There is one approximation that has gained widespread
C
