►
From YouTube: VASP Workshop at NERSC: Dielectric Response
Description
Presented by Martijn Marsman, University of Vienna
Published on December 18, 2016
Slides are available here http://www.nersc.gov/assets/Uploads/VASP-lecture-Dielectric.pdf
Presented at the 3-day VASP workshop at NERSC, November 9-11, 2016
A
So
why
would
we
want
to
compute
this
well
in
itself?
Obviously,
the
frequency
and
Static
dependent,
static
and
frequency,
dependent
dielectric
functions
correspond
to
two
certain
things
that
we
measure
right.
So
the
option
reflectance
and
things
like
this
and
in
the
long
wavelength
limit
of
the
frequency,
dependent,
polarizability
and
dielectric
matrices.
We
have
things
that
determine
the
optical
properties
in
the
regimes
that
are
accessible
to
optical
and
electronic
probes.
A
So
what
kind
of
quantities
we
calculate,
what
frequency
dependent
quantities
will
calculate
the
microscopic
dielectric
matrix
either
in
the
random
phase?
Approximation
or
Beyond,
including
change,
including
changes
in
the
in
the
DFT
exchange
correlation
potential.
I'll
show
you
later
on
what
that
means,
and
in
for
the
frequency
dependent,
macroscopic,
dielectric
tensor,
and
we
compute
imaginary
real
parts
of
the
dielectric
function
in
or
excluding
local
field
effects,
either
in
the
RPI
or
including
changes
in
if
the
exchange
correlation
potential
as
well.
A
So
anesthetic
properties
include
things
like
static,
dielectric,
tensor,
borne
effective
charges
so
born
effective
charges
are
other
things
out.
There
often
called
dynamical
charges
as
well.
So
if
I
move
an
atom,
how
much
of
charge
moves
with
my
atom
that's
more
an
effective
charge
and
as
we
do
in
our
excluding
local
field
effects
and
what
local
field
effects
are
L
I'll
mention
as
well,
and
this
is
either
done
from
density,
perturbation
theory
or
from
the
self-consistent
response
to
a
finite
electric
field.
A
Right
and
so
there's
there's
reasons
why
these
two
methods
that
essentially
yield
the
same
information
problem
is
the
density,
functional
perturbation,
Theory
only
works
for
density
function,
also
not
for
hybrid
functionals,
and
this
beat
method
that
allows
us
to
compute
the
self-consistent
response
to
a
finite
electric
field
that
works
for
for
hybrid
functions
as
well,
but
only
for
isolating
systems
or
not.
For
metallic
systems,
so
there
are
some
limits
to
do
these
different
methods.
So
if
you
look
at
dielectric
properties,
what
are
we
talking
about?
So
in
a
macroscopic
sense?
A
Are
we
talking
about
this
dielectric
tensor
here,
for
instance,
that
that
that
connects
an
externally
applied
field
to
the
fields
in
a
material,
sorry,
dielectric,
tensor
and
in
a
longitudinal
field,
and
which
is
a
field
caused
by
stationary
charges?
You
can
formulate
this
as
well
an
externally
applied
potential
connected
to
the
total
potential
in
a
system
and
where
our
total
potential
is
is
externally
applied
potential
and
the
induced
potential
tensor.
A
This
is
the
dielectric
tensor
some
terminology
around
this
and
induced
potentials
generated
by
an
induced
charge,
density,
one
that
makes
sense
in
a
sort
of
in
a
way
right
and
in
the
linear
response
regime
and
which
means
for
suffice
sufficiently
weak
fields.
We
have
a
relationship
by
D
between
the
externally
applied
potential
and
induced
charge
density,
and
that
is
through
the
reducible
polarizability
and
and
a
relationship
between
the
total
potential
yeah,
the
total
potential
and
the
induced
charge
through
the
irreducible,
polarizability
and
then
there's
a
bunch
of
relationships
between
them.
A
A
So
there
are
a
whole
bunch
of
illnesses
come
into
play.
So
we
have
a
field
and
a
dependency
on
on
the
positions
externally
applied
field
and
a
frequency,
and
then
we
have
a
tensor
of
this
kind.
That
depends
on
the
distance
between
two
points
right,
so
we
write
it
as
yes
and
if
we
write
it
like
this
so
force
efficiently-
and
this
is
the
microscope-
the
macroscopic
tensor
depends
only
on
the
distance.
A
If
we
have,
if
we
have
only
did
sorry
a
difference
here
and
not
two
coordinates,
we
can
fully
I
transform
this
into
momentum
space
using
only
one
coordinate
and
end
up
with
a
relationship
like
this
for
macroscopic
quantities
and
a
microscopic
dielectric
function
then
enters
in
such
a
way
right
and
in
momentum
space.
That
means
I
saw
in
a
macroscopic
way.
It's
only
the
distance
between
so
in
a
macroscopic
descriptions.
Only
the
distance
between
two
points
that
is
of
that
is
of
importance.
A
So
in
a
microscopic
electric
field
and
that
that
is
induced
by
an
externally
applied
field.
We
have
a
more
complicated
a
dielectric
function
where
we
don't
only
have
the
distance
between
R
and
R
Prime,
but
it
depends
actually
where
I
am.
With
respect
to
my
atomic
nucleus
or
something
like
this
and
these
microscopic
aspects,
and
if
we
Fourier
transform
this,
we
have
no
longer
only
one
Fourier
component
but
three
positions.
So
we
end
up
with
a
function
that
depends
on
G
and
G
Prime.
A
So,
and
that
already
gives
you
some
inkling
of
of
what
is
going
to
be
costly,
for
instance,
because
this
is
going
to
be
costly
to
store,
for
instance,
because
it
doesn't
depend
only
on
one
coordinate
or
on
one
for
VA
component,
but
already
on
to
so,
and
both
of
them
obviously
are
linked
in
the
usual
manner
by
some
some
matter
of
averaging
right,
some
averaging
over
a
certain
volume.
All
that
that's
not
something
that
we
are
going
to
be
using.
But
yes,
that
is
the
link
between
microscopic
and
macroscopic
worlds.
A
We
have
a
more
a
simpler
connection
between
the
externally
applied
field
and
the
resulting
field,
and
that's
given
by
the
head
of
this
inverse
dielectric
matrix,
and
then
we
have
a
very,
very
simple
connection
between
a
macroscopic,
dielectric,
constant
and
some
aspect,
some
quantity
that
we
that
we
in
fact
can
calculate
a
microscopic
quantity
so
that
so
for
materials
that
are
homogeneous
on
on
a
microscopic
scale.
The
off
diagonal
of
diagonal
parts
of
this
matrix
are
of
no
consequence.
An
air
of
Dera,
zero.
A
Sorry
so
for
G
is
not
G
prime
they're
0,
and
then
it
reduces
even
to
this
particular
relationship
so
being
able
to
compute
these
microscopic
quantities.
Under
the
assumptions
that
we
have
an
external
field
that
varies
on
a
length
scale,
that's
much
larger
than
atomic
distances
and
materials.
There
are
homogeneous
on
a
macroscopic
scale.
A
On
the
recent
sanctions,
we
have
a
direct
access
to
a
quantity
to
quantity
in
the
macroscopic
world,
and
this
is
what
we
call
for
instance,
so
the
fact
that
we
assume
that
the
materials
homogeneous
on
a
microscopic
scale-
that
is
what
we
call
a
neglected
local
field
effects,
I'm
just
sort
of
throwing
this
in,
because
you
will
see
this
in
in
in
literature.
It's
not
I
just
want
to
give
you
a
feeling
for
what
what
these
terms,
if
you
encounter
them
in
literature,
what
these
terms
mean?
A
Okay,
so
these
functions
that
we
saw
before
is
these
microscopic
dielectric
functions,
I've
written
out
here,
more
explicitly
in
terms
of
well
functions
and
quantities
in
reciprocal
space,
and
so
here
we
see
this,
which
was
called
the
reducible
polarizability,
which
is
the
change
in
density.
So
the
induced
density
when
we
change
the
external
potential,
so
I've
changed
in
the
induced
density
as
a
response
to
change
in
the
external
potential,
well
at
a
certain
at
a
certain
Ruby
I
component
at
a
certain
frequency
of
our
field,
and
so
this
is
spatial
variation
of
the
density.
A
This
would
be
with
respect
to
time,
but
then
transfer
it
into
the
frequency
domain
same
thing
for
the
external
potential
that
we
use.
This
is
what
I
said.
This
is
Coulomb
kernel,
so
this
this
represents
the
Coulomb.
Interaction
in
reciprocal
space
does
look
a
bit
different
than
what
we
saw
before
because
it
has
been
symmetrized,
so
it's
symmetric
with
respect
to
G
and
G
prime.
You
are
allowed
to
do
this
because
the
the
spectrum
of
such
an
operator
is
exactly
the
same.
A
You're
always
allowed
to
either
use
the
Coulomb
operator,
which
would
be
1
over
G
squared,
for
instance,
or
1
over
Q
plus
G
squared
or
it's
symmetric
counterpart,
which
is
Q
plus
G
times
Q
plus
G
Prime.
Okay.
So-
and
this
is
the
inverse
dielectric
function
same
thing
for
the
dielectric
function,
as
it
depends
on
the
irreducible
polarizability
right,
so
our
irreducible
one,
the
response
in
the
induced
charge
due
to
change
in
the
total
potential.
A
Well,
these
are
these
quantities
that
we
saw
before
reducible
polarizability
irreducible
one
symmetric,
Coulomb,
kernel
and
dyson
equation
that
relates
the
reducible
to
the
irreducible
polarizability.
So
that
looks
all
well.
It
looks
horrible,
but
this
is
all
sort
of
clean
and
and
and
we
can
have-
we
can
cast
our
problem
in
quantities
that
that
we
that
we
can
write
down
in
formulas,
but
we
don't
know
them
right.
So
how
do
we
compute
P
and
how
do
we
compute
the
Chi?
A
So
this
is
a
quantity
that
we
can
actually
compute
in
our
in
our
calculations
right.
So
an
other
other
and
wiser
derived
expressions
for
this,
which
are
completely
well.
The
only
ingredients
that
go
in
are
things
that
we
that
we
actually
know
from
our
computations.
So
if
you
look
here,
don't
don't
spend
too
much
on
trying
to
understand
this
equation
per
se,
but
just
see
that
what
it
contains
our
block
functions,
so
eigen
function
solutions
of
this
konchem
equations
that
we
calculate
and
eigenenergies
sorts
of
things
that
we
calculate
here.
A
We
see
these
are
these
occupational
numbers
of
our
of
our
block
states.
So
we
see
here
that
only
if
one
of
those
states
is
occupied
and
the
other
one
is
unoccupied,
it
actually
contributes
sums
that
we
here
see
r
/,
/
k,
+,
/,
sorry,
Alfre,
offer
and
and
and
prime
to
offer
to
banned
indices
where
one
is
an
occupied
band
and
the
other
one
an
unoccupied
band.
So
we
have
two
sums
of
abandoned.
A
It
says,
but
not
only
after
occupied
states,
so
that
is
that
is
one
aspect
of
these
computations
that
now
comes
along
for
the
first
time.
In
all
the
other
things
that
we
have
been
looking
at,
it
was
always
occupied
States
occupied
States
and
give
us
the
density
occupied
states
that
that
participate
in
the
folk
potential.
A
So
it's
already
expensive
in
in
itself,
and
so
this
evaluating
these
scales
is
n
to
the
power
4
with
respect
to
system
size,
but
not
only
that
one
of
these
indices
runs
over
unoccupied
states,
which
of
which
they
are
essentially
an
infinite
number.
So
obviously
we
will
not
be
using
an
infinite
number,
but
yes,
the
computational
effort
will
will
grow.
A
So
when
we
notice
this,
the
irreducible
polarizability
in
the
independent
particle
picture
is
kind,
not
the
other
ones,
the
ones
that
we
need
to
add
reducible
and
irreducible
polarizability
that
we
need
to
compute
our
microscopic
dielectric
function.
They
are
related
that
it
can
be
computed
out
of
this
out
of
this
irreducible
polarizability
in
the
N
independent
particle
picture,
so
the
one
that
we
compute
from
our
conscience-
and
that
is
written
here.
So
those
are
dyes
and
like
equations
again,
where
we
have
kind
note.
Well,
we
can
compute,
we
see
our
Coulomb
kernel
here.
A
We
see
something
that
depends
on
the
density
functional,
so
this
is
f.
X
C
have
existed
change
in
our
exchange
correlation
potential
with
respect
to
the
dense,
it's
the
quantity
that
we
can
actually
compute
and
are
these
relationships
that
hold,
and
these
are
the
ones
that
we
saw
before
that
give
us
access
to
these
dielectric
functions
that
were,
after
so
in
the
random
face
approximation,
and
that's
one
thing
that
that
that
we
are
going
to
use
widely
because
it's
because
it's
computationally
attractive
in
the
random
face
approximation.
A
We
say
that
the
irreducible
polarizability
is
simply
equal
to
the
irreducible
polarizability
in
the
independent
particle
picture,
which
essentially
means
that
we
set
this
FX
C
to
0
and
that's
the
essence
of
the
of
the
of
the
random
face
approximation
so,
and
that
gives
us
this
particular
relationship
between
a
dielectric
function,
a
microscopic
dielectric
function
and
this
polarizability
that
we
compute
from
our
concern
system.
So
we
can
go
beyond
and
that
that
was
mentioned
on
one
of
the
first
slides.
A
So
we
can
go
beyond
to
random,
face
approximation
and
include
changes
in
the
exchange
correlation
potential.
So
not
set
this
thing
to
zero
and
and
we
get
get
such
an
equation,
and
that
is
what
we
commonly
call
local
field
effects
in
in
density.
Functional
theory.
So-
and
this
is
local
field
effects
in
the
random
face
approximation.
B
A
So,
as
we
saw
before
in
the
long
wavelength
limit,
so
so
so
now
we
try
to
connect
to
experiment
to
experiment.
So
in
the
long
wavelength
limits
4q
goes
to
zero
of
the
dielectrics
matrix.
We,
the
dielectric
matrix,
gives
us
access
to
two
optical
properties
in
regimes
that
are
that
are
probed
by
optical
measurements
and
a
microscopic
dielectric
tensor
epsilon
infinity,
for
instance,
of
Omega,
is
given
by
this
particular
relationship
right.
This
quantity
we
now
can
can
calculate,
for
instance,
in
this
random
phase
approximation
from
from
our
cone
shampoo,
Laura's
ability.
We
can
calculate
this.
A
This
inverse
dialect
matrix
in
the
random
face
approximation,
for
instance,
and
get
access
to
2
epsilon
infinity.
So
we
can
do
this
at
various
levels
of
approximation
like
I
mentioned
before
we
can
stay
within
the
independent
particle
picture.
That
is
what
we
saw
before.
That
was
this.
This
is
truly
truly
trivial
relationship,
but
we
can
also
include
local
field
effects,
as
I
mentioned
before,
in
the
random
face
approximation
or
beyond,
adding
them
adding
changes
in
the
DFT
exchange
correlation
potential
as
well.
So
these
are
tags
that
are
related
to
this,
so
you
switch
on.
A
L
optics
is
true.
You
get
this
particular
this
particular
quantity
written
out.
If
you
want
to
go
beyond
you
can
use,
algo
is
chi.
It
uses
a
lot
of
the
machinery
that
that
will
be
used
for
4gw
as
well,
so
it's
quite
a
bit
more
expensive
and
then
you
can
get
a
better
approximation
or
a
more
fancy
approximation
to
your
dielectric
screening
right.
A
A
A
So
how
do
we
get
this?
We
compute
the
imaginary
part
of
the
office
of
this
tensor
directly
from
from
the
cell
periodic
parts
of
the
of
the
column
orbitals
and
then
do
a
grommet
corner
gram,
grammars
clonic
transformation
to
get
at
at
the
real
part
of
this
particular
quantity.
So
the
computational
effort
in
this
case
lies
in
computing.
These
functions,
so
these
ones
are,
are
easily
easily
are
we
can
get
at
easily
and
those
are
our
block
states,
but
actually,
in
this,
in
this
queue
to
go
to
zero
going
to
zero
limit.
A
So
what
we
actually
need
to
compute
is
the
first
order,
change
in
the
cell
barrier,
part
of
our
of
our
block
functions
and
that
we
do
in
in
perturbation
theory,
so
that
is
that
is
written
here
so
and
so
an
orbital
at
a
cell
burial
periodic
part
of
a
cone.
Some
function,
sorry
of
a
block
function
slightly
away
from
K,
can
be
written
in
this
way.
So
it's
a
Taylor
expansion.
A
A
Well,
we
can
cost
it
that
way,
at
least
that
one
runs
over
occupied
States
and
one
runs
over
unoccupied
states
and
yes,
all
kinds
of
things
that
we
can
evaluate.
But
again
the
need
to
be
here,
because
these
such
an
expansion
in
perturbation
Theory
essentially
runs
over
all
states
all
states
of
your
Hamiltonian
of
the
unperturbed
Hamiltonian.
So
for
these
kind
of
things
again,
we
have
to
include
many
more
states
than
we
would
need
before
for
this
ground
state
calculation,
so
ground
state
calculations
are
only
interested
in
occupied
states.
A
So
does
it
work?
Yes,
it
does
work
so
there's
paper
describing
the
way
that
we
have
implemented
this,
and
there
are
some
test
cases
where
we
complain
where
we
are
where
we
compared
to
a
PW
plus
local
orbital.
So
this
is
the
live
into
K
code
and,
if
you're
interested
so
there's
examples
where
we,
where
we
run
exactly
such
a
calculation
in
hands
on
section
and
the
quantities
that
that
you're
after
well
I've
put
in
the
strings
that
you
can
search
for
in
the
outcome
file.
A
C
A
So
yes,
well,
this
I
think
this
is
only
a
repeat
of
what
we
already
seen
before
sorry
yeah.
So
this
is
sort
of
a
recap
right,
so
the
quantity
that
we
are,
that
we
use
the
quantity
that
we
have
access
to
this
irreducible
polarizability
in
the
independent
partial
picture-
and
this
is
this
other
and
wiser
expression
that
gives
us
access
to
this
particular
quantity
and
that
scales
evaluating
this
scale
says
n
to
the
power
4.
So
why
is
it
n
to
the
power
4
so
for
a
large
system
where
we
don't
have
K
points?
A
A
Actually
you
heard
this
random
face
approximation
that
will
that
will
pop
up
in
GW
and
and
another
methods
later
on,
so
in
the
random
face
approximation,
the
dielectric
screening
is
given
by
such
an
expansion,
and
so
our
screened
potential
in
the
random
face
approximation
is
given
by
Weldon
the
baculum,
the
bear,
coulumb
interaction
and
then
the
electronic
environment.
So
the
second
term
is
the
electronic
environment
that
that
reacts
to
the
field
generated
by
a
particle.
A
And
so
electrons
react
to
the
induced
change
in
the
potential
and
that
will
give
an
additional
change
in
the
charge
density
and
a
corresponding
change
in
the
Hartree
potential
and
so
on
and
so
on.
Advocacy's.
These
are
all
changes
in
the
Hartree
potential
through
his
school
on
Cornell
in
the
random
face
approximation,
and
this
this
can
be
written
as
a
geometrical
series,
and
so
this
whole
collection
of
terms
up
to
up
to
infinitely
long
terms.
A
You
can
write
us
into
your
metrical
series
given
by
the
naked
Coulomb
interaction
times
the
inverse
of
this,
and
this
is
exactly
the
inverse
of
this
dielectric
function
and
right
so
in
the
random
face
approximation.
But
do
we
actually
compute,
we
compute
a
screened,
Coulomb
interaction
where
all
the
older,
all
the
physical
processes
involve
involve
the
Hartree
interaction
right,
involve
changes
in
the
Hartree
potential.
B
A
So
you
can
include
that,
as
we
saw
before
in
using
this
elbow
is
Chi.
You
can
include
all
these
all
these
processes,
this
random
face
approximation
in
into
your
computation
of
the
dielectric
properties,
and
you
will
get
them
in
written
into
the
are
out
car
file
right
so
per
default.
It's
it's
all
done
at
the
level
of
our
PA,
but
you
can.
You
can
include
changes
in
the
dance
in
the
in
the
exchange
correlation
potential
as
well.
A
So
yes,
so
there's
a
few
practical
points
and
that
will
that
will
be
of
importance
to
a
computation
of
dielectric
properties,
but
it's
also
of
importance
to
well
very
much
of
importance
for
GW
calculation,
where
I
will
use
this
dielectric
properties
I've
been
talking
about
the
fact
that
we
are
going
to
be
using
these
these
unoccupied
States,
and
the
thing
is
that
that
we
we
use
iterative
matrix
diagonalization
to
refine
our
states
right.
We
talked
about
it
yesterday
and
I
said
okay.
A
This
is
very
nice
because
we
can
get
at
the
so
many
lowest
states
in
our
eigen
value
spectrum
of
our
Hamiltonian
matrix,
and
the
thing
is
that
is
iterative
techniques.
They
converge
most
rapidly
for
the
deepest
lying
States.
As
you
go
up
into
the
eigen
eigen
value
spectrum,
the
states
converge
less
and
less
rapidly,
so
in
for
a
ground
set
calculation,
we
look
to
the
convergence
of
the
occupied
states
right
because
they
give
us
the
total
energy
and
our
convergence
criterion
only
looks
at
total
energy.
A
It
says:
okay,
if
my
total
energy
doesn't
change
by
soul
and
so
much
anymore,
I
say:
I'm
converged
right,
but
now
we
will
use
the
virtual
States
or
the
empty
states
in
the
spectrum
as
well,
and
they
don't
contribute
to
the
total
energy.
So
the
quality
of
those
states
and
whether
or
not
they
are
converged
doesn't
express
itself
in
the
total
energy.
A
So
you
could,
you
can't
simply
do
do
a
calculation
that
includes
simply
more
bands
and
then
say:
okay
now
I'm
converts,
because
then
the
virtual
states
are
well
less
converged
than
the
than
the
occupied
states,
so
the
quality
of
your
virtual
states
of
your
unoccupied
bands
is
not
guaranteed.
So
we
do
a
trick
to
to
get
around
this.
We
do
a
ground
set
calculation,
a
simple
one
where
we
we
simply
use
the
the
number
of
bands
that
that
we
would
use
for
any
normal
ground
state
calculation
and
then
to
obtain
virtual
orbitals
of
sufficient
quality.
A
We
do
exact,
diagnose
Asian
of
the
Hamiltonian,
so
that's
the
thing
that
I
said
yesterday
well
that
that
is
the
thing
that
we
want
to
avoid
it,
because
why
would
I
do
an
exact
diagonalization
and
compute
50,000
50,000
eigenstates
of
my
Hamiltonian,
when
I'm
only
interested
in
four
well
here
we'll
do
one
step
of
this
on
top
of
such
ground
set
calculation.
So
essentially
our
occupied
states
have
converged.
Our
Hamiltonian
is
correct
in
that
sense,
and
then
we
build
up
this
mo
Tony,
so
we
restart
the
calculation
reading.
A
Those
wave
functions
build
up
the
Hamiltonian
that
is
defined
by
the
occupied
States
diagonalize
at
one
time,
and
then
we
have
all
unoccupied
states
correct
as
well,
so
the
quality
of
those
states
will
be
as
good
as
the
ones
that
we
have
optimized
for
occupy
part
of
the
spectrum.
So
that's
an
additional
step
that
that
you
would
need
to
do
to
get
to
get
high-quality
virtual
orbitals.
B
A
That
is,
that
is
something
that
that
that
we
will
have
to
implement
very
soon.
It
would
be
a
better
way
to
do
it,
but
at
the
moment
that's
not
present
yet
so
one
can
obviously
say
I,
don't
look
at
the
at
the,
so
you
wouldn't
have
to
look
per
se
at
the
total
energy
to
judge
whether
your
converts,
you
can
look
at
the
eigen
energies
of
of
any
number
of
states
and
as
they
stop
changing
or
are
converts
with,
with
a
certain
degree,
you
could
say:
okay,
I'm
done.
A
That
would
allow
you
to
keep
using
these
iterative
matrix
diagonal
asian
techniques.
Yes,
that's
definitely
true,
and
we
should
change
this,
but
at
the
moment
that's
not
possible.
Yet
it's
quite
unfortunate,
so
we
are
forced
to
do
this
to
do
this
three
steps
and
and
you're
right.
This
is
a
very
costly
step
for
large
systems.
It's
even
so,
and
that
is
why
we'll
definitely
will
definitely
change
it
pretty
soon.
A
So
this
this
will
be
a
valuable
thing,
even
in
the
future,
but
for
very
large
systems
and
for
GW.
This
is
something
that
that
might
kill
you
at
the
moment.
That
is
true.
So
these
are
the
typical
steps
that
you
would
do
and
well
in
this
case
it's
only
a
calculation
of
the
dielectric
properties
per
se.
So
the
three
steps
is
a
ground
set
calculation.
Then
this
exact
diagonalization
step
where
you,
where
you
ask
for
asked
for
a
large
number
of
empty
states
and
those
aren't
so
it's
done.
D
B
E
B
A
A
A
Yes,
so
I
recently
did
a
calculation
on
on
a
cell
with
256,
so
at
GW
calculation
on
a
cell
with
256
silicon
atoms
and
there
it
was
something
like
47
thousand
fans,
so
the
whole
spectrum
depends
is
really
yeah.
It's
quite
sizable,
yes
yeah,
so
it
would
pay
off
to
to
be
able
to
skip
this
step.
Yes,
definitely.
A
Well,
that
is,
that
is
the
thing.
So
so
is
there
a
rule
of
thumb?
Yes,
there
are
many
rules
of
thumb,
so
I
but
I
mean
so
for
silicon,
and
that
is
the
well.
This
is
our
favorite
test
system.
Let's
say,
and
everything
always
works
for
silicon
because
it
has
been
tested
with
silicon.
I
I
would
use
something
like
sixty
four
states
per
silicon
atom
right
so
which
means
that
in
that
particular
calculation
that
I
gave
for
256
silicon
atoms,
I
kept
16,000
bands.
Well,
a
large
search
can
be
quite
sizable.
A
I
think
it
depends
on
how
accurate
you
want
your
results
to
be.
You
could
get
away,
probably
32
bands
for
for
at
them
as
well,
go
to
another
system.
I,
don't
know
I,
don't
want
to
venture
I.
Don't
want
to
to
tell
you
yes
always
use
64
bands
per
atom
right,
so
I
I
simply
don't
know,
and
it's
one
of
the
things
white
white
I
mean
you
can
get
around
this
step
by
simply
using
all
bands
that
the
number
of
bands
you
get
is
then
simply
defined
by
your
cutoff.
A
And
then
your
results
again
are
going
to
converge
with
your
energy
cutoff,
because
the
size
of
this
matrix
is
directly
defined
by
your
basis
set
cutoff,
because
it's
the
size
of
the
f50
grid
and
that
is
defined
by
your
cutoff.
And
so
in
this
way
you
get
around
around
this
uncertainty
uncertainty
by
simply
using
all
bands
and
then
there's
clever
algorithms,
by
which
we
can
extrapolate
to
infinite
basis
set
size,
because
we
know
how
it
should
how
it
should
converge
with
respect
to
a
s.
Your
cutoff
is
high
enough.
A
We
know
the
convergence
behavior
with
respect
to
the
cutoff,
and
then
we
do
an
extrapolation,
but
that
can
only
work
if
you
use
all
bands
that
there
are,
if
you
use
all
states
of
your
Hamiltonian
yeah.
So
it
does
pay
off
to
still
do
that.
Actually,
especially
for
a
40
w
and
a
CF
DT
calculation
as
we'll
see
later
on.
A
Yes,
so
here
come
these
GW
potentials
that
maybe
people
have
have
already
seen
and
because
and
already
mentioned
this
yesterday-
that
we
now
depend
on
everything.
We
need
a
certain
amount
of
virtual
states
as
well,
and
so
we
need
potentials
that
are
that
are
sufficiently
accurate,
that
they
describe
the
scattering
properties
up
to
up
to
higher
in
the
in
the
energy
spectrum.
A
You
start
to
see
deviations,
so
this
potential
will
not
yield
very
high
quality
virtual
states
because
they
depend
on
scattering
properties
higher
up
in
the
spectrum,
and
these
two
W
potentials
have
been
constructed
to
match
the
scattering
properties
up
to
much
higher
energies,
and
that
is
the
idea,
the
essence
the
essential
step
which,
which
is
why
we
why
we
advise
to
use
these
potentials.
If
you
go
to
any
of
these
methods
that
include
sums
of
our
virtual
states.
B
A
So,
let's
turn
to
static
dielectric
response
and
so
static.
Dielectric
response,
there's
a
few
things
that
we
have
to
calculate
so
that
the
program
can
calculate
for
you.
The
iron
clamp
static,
macroscopic,
dielectric,
tensor
solecism.
This
epsilon
invent
infinity
at
Omega
is
0
at
0
frequency.
The
board
effective
charge
tensors
that
I
that
I
spoke
about.
A
So
there's
examples
in
the
hands-on
session
that
deal
with
these
aspects
right
so
for
so
using
density,
functional
perturbation
theory
so
essentially
added
the
thing
that
we
have
to
compute
here
is
again
the
derivative
of
this
of
of
are
the
cell
periodic
part
of
our
block
functions
with
respect
to
the
block
wave
vector
in
case
of
density
function
of
perturbation
theory.
We
don't
use
the
sum
over
empty
States
to
do
this.
It's
not
this.
A
This
first
order
perturbation
theory
where
we
have
a
sum
over
over
sorry
over
empty
states,
but
we
we
solve
a
linear,
Stern,
himer
equation.
That
looks
like
this
I
saw
this,
so
this
is
the
number
of
states
for
which
we
have
to
do.
This
is
of
the
order
of
the
occupied
states
right
so
that
that
is
nice.
That's
a
nice
thing
from
density,
functional
perturbation
theory,
so
actually
so
that
we
solve
for
this
particular
quantity,
and
this
particular
quantity
actually
is,
is-
is
needed
to
express
our
perturbation.
So
we
solve
our
perturbation.
A
Is
a
it's
a
it's
a
bit
of
a
strange
thing
in
this
in
these
electro
static
electric
fields,
and
so
this
is
a
static.
Electric
field
at
at
Q
is
zero.
So
that
means
it's
a
constant
field
right
over
our
material,
and
that
poses
a
problem
for
for
the
kind
of
methods
that
we
use,
because
the
constant
electric
field
is
a
that
is
a
constantly
dropping
potential
out
for
a
system
with
periodic
boundary
conditions.
So
it's
not
it's
not
a
perturbation
that
we
can
express
under
periodic
boundary
conditions.
Right
is
this
potential
that
we
add.
A
Our
equation
gives
us
this
gives
us
actually
a
representation
of
our
electric
field
of
our
perturbation,
and
then
we
solve
a
second
second
mr.
himer
equation
to
get
at
the
response
in
our
orbitals.
Due
to
this
due
to
this
perturbation-
and
this
can
include
all
kinds
of
local
field
effects
so
which
means
that
we
can
include
the
response
in
our
self
consistent
hamiltonian
due
to
this
perturbation
right
so
which
is
commonly
called
these
local
field
effects
and
they
may
be
included
at
the
RTL
level
again
or
or
beyond.
A
A
Only
so
there's
no
sum
over
unoccupied
states
anymore,
and
we
can
compute
a
host
of
other
quantities
from
this
like
born
effective
charges
and
and
these
beautiful
electric
tensors
by
simply
looking
at
changes
in
the
helm
environment
forces
as
we
change
our
orbitals
with
respect
to
this
to
the
to
their
response
to
the
perturbation.
And
we
do
this
in
finite
differences.
F
B
F
A
B
A
B
A
A
Electric
field
is
not
something
that
we
can
represent
under
periodic
boundary
conditions,
there's
a
way
to
actually
include
the
effects
for
I
for
insulating
systems,
and
that
that
is
based
on
this.
What
is
called
the
modern
theory
of
polymerization?
Well,
it's
always
very
bad
to
call
anything
modern,
because
it
is
modern
at
some
point-
and
this
is
already
some
years
old,
so
right
so
in
100
years.
It's
definitely
not
going
to
be
modern
anymore,
but
it's
still
called
this
so,
and
that
says
we
can
change.
A
So
it's
one
of
the
things
that
there
is
maybe
a
bit
counter
intuitive
the
fact
that
we
can
compute
the
polarization
under
periodic
boundary
conditions
is
not
a
trivial
statement
in
itself,
and
that
is
maybe
I
can
can
sketch
again
from
this
nice
wall
and
that
is
under
periodic
boundary
conditions.
We
have
two
following
problems,
so
let's
say:
if
I
have,
if
I
have
a
dipole
in
a
system,
it's
easy
to
compute.
A
The
other
polarization
because
of
this
is
no
there's
no
essential
problem,
but
if
I,
if
I,
not
yes,
I
cannot
put
repeated
dipoles,
which
we
have
on
the
periodic
boundary
conditions.
It
becomes
a
non
trivial
thing
to
compute
this,
because
I
can
choose
my
cell
to
be
this
right
and
then
integrate
out
for
the
charge.
It
will
give
me
one
answer,
but
it
can
also
choose
and
that's
completely
equivalent.
A
It
is
easy
to
define
a
dipole
moment
in
in
a
system,
and
then
you
can,
then
you
go
through
a
whole
bunch
of
mathematical,
manipulations
and
end
up
with
an
equation
that
tells
you
that
your
macroscopic
polarization
depends
on,
and
that
is
what
we
saw
before
right.
I
already
mentioned
it,
the
derivative
of
the
self
periodic
part
of
our
block
functions
with
respect
to
the
block
wave
vector.
A
Okay
for
the
interest
of
time,
we'll
we'll
simply
believe
this
and-
and
we
say
ok,
we
have
a
way
to
compute
the
polarization
of
our
system,
and
this
opens
now
a
route
to
include
the
effects
of
a
finite
electric
field
as
well,
because,
we'll
simply
add
a
term.
So
this
is
our
normal
ground
state,
DFT,
energy,
right
with
wave
functions
and
now
we'll
add
a
term
which
is
essentially
the
in
product
between
the
finite
electric
field
that
we're
considering
and
the
macroscopic
polarization
of
our
system.
A
Sorry
Susan
gives
us
a
new
total
energy
that
includes
the
effect
of
a
finite
electric
field
on
our
system,
so
an
additional
term
in
our
total
energy.
Add
a
defense
of
this
expression
that
we
saw
before
that
comes
out
of
the
modern
theory
of
polarization.
An
additional
term
in
our
total
energy
will
add
and
corresponding
term
to
the
Hamiltonian,
and
that
is
well
this
one
and
so
the
difference.
A
The
variation
of
our
polarization
with
respect
to
an
orbital
is
added
to
our
Hamiltonian,
and
then
we
simply
do
a
self-consistent
calculation
for
this
particular
Hamiltonian
and
minimize
this
total
energy.
So
there
are
some
considerations.
You
cannot
make
this
field
too
big.
Now,
because
at
some
point
you
will,
you
will
start
to
close
your
gap,
so
that
is
that
is
sketched
on
your
slide,
and
that
is
argued
in
this
particular
paper.
If
you're
interested
in
this,
so
there's
a
criterion
implemented
how
strong
you
can
make
your
field
before
you
before
you
close
your
gap.
A
You
must
have
forget,
so
the
system
must
be
insulating,
but
then
you
can.
Then
you
can
optimize
your
wave
functions
with
respect
to
this
particular
energy,
functional
and
use
these
use
these
solutions
to
compute
the
quantities
that
we
have
seen
before
right,
because
we
can,
we
can
compute
the
change
in
the
polarization
now
with
respect
to
and
a
finite
electric
field,
and
we
saw
that
this
change
in
the
polarization
with
respect
to
the
field
gives
us
our
static,
macroscopic,
dielectric
matrix
and
the
born
effective
charges
is
associated
with
this.
A
We
can
again
like
before
compute
from
and
we
apply
an
electric
field.
We
have
a
slight
change
in
the
wave
functions
and
through
this
it's
changed
in
the
helm,
environment
forces
and
the
change
in
these
forces
with
respect
to
the
field
give
us
the
born
effective
charges,
so
the
essentially
quantities
that
we
get
from
density
function
of
perturbations.
Here
we
can
get
from
a
response
to
to
this
finite
fields
as
well
and,
of
course,
all
local
field
effects
are
included
in
a
very
natural
manner,
because
we
do
a
self-consistent
calculation.
A
You
apply
a
field
and
then
you
do
a
self-consistent
calculation.
So
all
variations.
All
changes
in
potentials
to
to
the
field
are
naturally
included
in
this
right.
You
through
the
self-consistency,
so
that's
very
nice,
and
we
can
do
this
for
any
Hamiltonian,
so
density,
functional
perturbation
theory.
You
can
do
for
density
functionals,
but
not
for
hybrid
functions,
but
this
works
for
hybrid
functions.
Well,
okay,
so
there's
some
numbers
here
that
we
did
we
gotten
with
this
and
yeah,
so
I
think
we'll
skip
this.
These
are
essentially
only
details
on
on
this
particular
terms.
B
E
E
A
E
A
So
the
biggest
system
that
I
have
been
able
to
do
is
not
very
big,
because
we
are
poor
austrians
and
we
don't
have
computers,
but
now
I
know
people
that
do
that.
Do
you
see
calculations
and
systems
with
thousands
of
atoms
completely
beyond
anything
that
I
have
experience
with
I,
don't
know
how
painful
it
is
and
how
big
the
machine
has
to
be,
but
I
think
what
is
you
have?
Maybe
an
idea
of
what
what
a
common
size
at
Newark
is
for.
D
A
I
got
some
some
benchmarks
from
from
a
guy
on
and
he
was
using
a
crane
machine
and
he
could
go
up
to
thousands
of
course,
so
he
had
access
to
a
fair
amount,
of
course
and
I
think
he
did
systems
of
some
thousands
of
atom
or
a
thousand
of
atoms.
Things
like
this,
so
it
is
possible,
it's
surely
painful,
I'm,
absolutely
sure
this.
It
hurts
to
do
it.
Yeah,
yes,
sometimes.
F
B
F
F
A
Although,
although-
and
that
is
something
that
will
come
up,
although
that
then
you're
already
so,
if
give
T
gives
you
such
a
bad
answer
for
that
for
the
ground
state,
then
I
would
try
with
a
hybrid
functional,
but
it's
something
that
that
I'll
discuss
tomorrow.
What
what
I
mean?
There's
there's
this
there's
a
compensation,
there's
a
compensation
of
errors
involved
in
using
the
random
face
approximation
and
DFT.
So
you
do
get
really
good
screening
properties
if
you
use
two
random
face
approximation
in
connection
with
DFT,
orbitals
and
eigenenergies.
A
A
A
F
F
A
Perfect
so
yeah,
so
how
do
you
apply?
The
sister
of
yeah
I
will
have
to
look.
Look
it
up.
I
said
I
mean
it's
in
one.
It's
used
in
one
of
the
examples:
okay,
but
I
will
have
to
make
sure
that
it
that
it
will
work
in
the
way
that
that
you,
that
you
need
yeah
and
it's
something
I
mean
it.
It
is
so
simple
that
if
there's
something
still
missing
there,
we
can
easily
quickly
fix
this
yeah,
yes
to
make
it
work
in
a
in
a
manner
that
is
convenient
to
you.
A
So
would
you
get
the
dielectric
properties
of
isolated
molecules?
I,
don't
see
a
reason
why
why
not
I,
don't
know
how
good
DFT
or
or
these
methods
would
or
the
random
phase
approximation
would
be
for
isolated
molecules,
though
or
so
so.
I
have
no
experience
doing
this
for
for
isolated
molecules,
but
there's
no
there's
no
essential,
there's
nothing
in
any
approximation.
That
would
would
invalidate
the
procedure
for
isolated
molecules
right,
yeah,.