Title: Zero Temperature QMC Outline of lecture
1Zero Temperature QMC Outline of lecture
- Variational Monte Carlo
- Diffusion Monte Carlo
- Fermion systems
- Coupled Electron-Ion Monte Carlo
- Applications
- Electron gas
- High pressure hydrogen
2Variational Monte Carlo
- Historically first quantum simulation method
- Slater-Jastrow trial function
- Examples liquid helium and electron gas.
- Wavefunctions for Quantum solids
- WaveFunctions beyond Slater-Jastrow back flow
and 3-body - Twist Averaged Boundary Conditions
3The Variational Theorem
- Assume is a trial function
where R are the quantum degrees of freedom
(positions, spin) a are variational
parameters. - Conditions matrix elements exist, symmetries and
boundary conditions are correct.
4- Expand trial function in terms of the exact
eigenfunctions - Energy and variance are second order in
(1-overlap). - Other properties are first order.
- Temple lower bound
?
E0 E
5Linear Basis approach
- Assume trial function is a linear combination of
known functions a basis fn(R).
Problem that MC solves Unless we use 1-particle
basis, integrals are too slow to perform.
6Properties of solution to GEP
E
- For a basis of size m, there exist m
eigenvalues and orthonormal eigenfunctions - McDonalds theorem the nth eigenvalue in a basis
is an upper bound to the nth exact eigenvalue. - We can always lower all the energies by
augumenting the basis - When basis is complete, we get exact answers!
7Symmetry reduces complexity
- If an operator P commutes with H P,H0 we can
reduce complexity by working in a basis with that
symmetry. - rotational symmetry use Ylm.
- Translation symmetry use plane waves.
- Inversion symmetry even/odd functions
- Matrix elements are non-zero only for states
within the same sector. - By reordering we can block diagonalize
- Reduces complexity from M3 to k(M/k)3M3/k2.
- McDonalds theorem applies to each sector
individually.
8Notation
- Individual coordinate of a particle ri
- All 3N coordinates R (r1,r2, . rN)
- Total potential energy V(R)
- Kinetic energy
- Hamiltonian
9Liquid heliumthe prototypic quantum fluid
- Interatomic potential is known more accurately
than any other atom because electronic
excitations are so high. - A helium atom is an elementary particle. A weakly
interacting hard sphere.
- Two isotopes
- 3He (fermion antisymmetric trial function, spin
1/2) - 4He(boson symmetric trial function, spin zero)
10Helium interaction
- Repulsion at short distances because of overlap
of atomic cores. - Attraction at long distance because of the
dipole-induced-dipole force. Dispersion
interaction is - c6r-6 c8 r-8 .
- He-He interaction is the most accurate. Use all
available low density data (virial coefficients,
quantum chemistry calculations, transport
coefficients, .) Good to better than 0.1K (work
of Aziz over last 20 years). - Three body interactions are small but not zero.
11Helium phase diagram
- Because interaction is so weak helium does not
crystallize at low temperatures. Quantum exchange
effects are important - Both isotopes are quantum fluids and become
superfluids below a critical temperature. - One of the goals of computer simulation is to
understand these states, and see how they differ
from classical liquids starting from
non-relativistic Hamiltonian
12Variational Monte Carlo (VMC)
- Variational Principle. Given an appropriate trial
function - Continuous
- Proper symmetry
- Normalizable
- Finite variance
- Quantum chemistry uses a product of single
particle functions - With MC we can use any computable function.
- Sample R from ?2 using MCMC.
- Take average of local energy
- Optimize ? to get the best upper bound
- Error in energy is 2nd order
- Better wavefunction, lower variance! Zero
variance principle. (non-classical)
13First Major QMC Calculation
- PhD thesis of W. McMillan (1964) University of
Illinois. - VMC calculation of ground state of liquid helium
4. - Applied MC techniques from classical liquid
theory. - Ceperley, Chester and Kalos (1976) generalized to
fermions.
- Zero temperature (single state) method
- Can be generalized to finite temperature by using
trial density matrix instead of trial
wavefunction.
14Trial function for helium 4Jastrow or pair
product
- We want finite variance of the local energy.
- Whenever 2 atoms get close together wavefunction
should vanish. - The pseudopotential u(r) is similar to classical
potential - Local energy has the form
- G is the pseudoforce
- If v(r) diverges as ?r-n how should u(r) diverge?
Assume - u(r)?r-m
- Keep N-1 atoms fixed and let 1 atom approach
another and analyze the singular parts of the
local energy. - Gives a condition on u at small r.
- For Lennard-Jones 6-12 potential, Jastrow goes as
m5
15Periodic boundary conditions
- Minimum Image Conventiontake the closest
distance - rM min ( rnL)
- Potential is cutoff so that V(r)0 for rgtL/2
since force needs to be continuous. Remember
perturbation theory. - Image potential
- VI ? v(ri-rjnL)
- For long range potential this leads to the Ewald
image potential. You need a back ground and
convergence method.
16(No Transcript)
17Optimization of trial function
- Try to optimize u(r) using reweighting
(correlated sampling) - Sample R using P(R)??2(R,a0)
- Now find minima of the analytic function Ev(a)
- Or minimize the variance (more stable but
wavefunctions less accurate). - Statistical accuracy declines away from a0.
18Quantum Crystal Trial Function
- Jastrow trial function does not freeze at
appropriate density. - Solution is to break spatial symmetry by hand.
- Introduce a bcc lattice
- Zi
- bcc has the lowest Madelung energy, but others
may have lower zero point energy. - Introduce localized one-body terms (Wannier
functions). - Make a Slater determinant possibly with spin
ordering. - More complicated trial functions and methods are
also possible.
- C is a variational parameter to be optimized.
19Scalar Properties, Static Correlations and Order
Parameters
- What do we get out of a simulation? Energy by
itself doesnt tell you very much. - Other properties
- do NOT have an upper bound property
- Only first order in accuracy
- EXAMPLES
- Static properties pressure, specific heat etc.
- Density
- Pair correlation in real space and fourier space.
- Order parameters and broken symmetry How to tell
a liquid from a solid - Specifically quantum the momentum distribution
20Other quantum properties
- Kinetic energy
- Potential energy
- Pair correlation function
- Structure function
- Pressure (virial relation)
- Momentum distribution
- Non-classical showing effects of bose or fermi
statistics - Fourier transform is the single particle
off-diagonal density matrix - Compute with McMillan Method.
- Condensate fraction 10
Like properties from classical simulations No
upper bound property Only first order in accuracy
21Order parameters
- A system has certain symmetries e.g.translation
invariance. - At high temperatures one expects the system to
have those same symmetries at the microscopic
scale. (e.g. a gas) - BUT as the system cools those symmetries can be
broken. (a gas condenses or freezes). - At a liquid gas-transition the density is no
longer fills the volume droplets form. The
density is the order parameter. - At a liquid-solid transition, both rotational
symmetry and translational symmetry are broken - The best way to monitor the transition is to look
how the order parameter changes during the
simulation.
22Microscopic Density
- ?(r) lt ?i ?(r-r i) gt
- Or you can take its Fourier Transform
- ? k lt ?i exp(ikri) gt
- (This is a good way to smooth the density.)
- A solid has broken symmetry (order parameter).
The density is not constant. - At a liquid-gas transition the density is also
inhomogeneous. - In periodic boundary conditions the k-vectors are
on a grid k2?/L (nx,ny,nz) Long wave length
modes are absent. - In a solid Lindemanns ratio gives a rough idea
of melting - u2 lt(ri-zi)2gt/d2
23Snapshots of densities
- Liquid or crystal or glass? Blue spots are
defects
24Pair Correlation Function, g(r)
- Primary quantity in a liquid is the probability
distribution of pairs of particles. Given a
particle at the origin what is the density of
surrounding particles - g(r) lt ?iltj ? (ri-rj-r)gt (2 ?/N2)
- Density-density correlation function
- From g(r) you can calculate all pair quantities
- (potential, pressure, )
- A function gives more information than a number!
25g(r) in liquid and solid helium
- First peak is at inter-particle spacing. (shell
around the particle) - goes out to rltL/2 in periodic boundary
conditions.
26(The static) structure factor S(k)
- The Fourier transform of the pair correlation
function is the structure factor -
- problem with (2) is to extend g(r) to infinity
- S(k) is measured in neutron and X-Ray scattering
experiments. - Can provide a direct test of the assumed
potential. - Used to see the state of a system
- liquid, solid, glass, gas? (much better than g(r)
) - Order parameter in solid is ?G where G is a
particular wavevector (reciprocal lattice vector).
27Bragg peak
- In a perfect lattice S(k) will be non-zero only
on the reciprocal lattice vectors G S(G) N - At non-zero temperature (or for a quantum system)
this structure factor is reduced by the
Debye-Waller factor - S(G) 1 (N-1)exp(-G2u2/3)
- To tell a liquid from a crystal we see how S(G)
scales as the system is enlarged. In a solid,
S(k) will have peaks that scale with the number
of atoms.
- The compressibility is given by
- We can use this is detect the liquid-gas
transition since the compressibility should
diverge as k approaches 0. (order parameter is
density)
28Crystal liquid
29Momentum Distribution
- Momentum distribution
- Classically momentum distribution is always a
Gaussian - Non-classical showing effects of bose or fermi
statistics - Fourier transform is the single particle
off-diagonal density matrix - Compute with McMillan Method.
- For fermions we need to use the determinant
update formulas to find the effect of the
movement of 1 electron.
30Derivation of momentum formula
- Suppose we want the probability nk that a given
atom has momentum hk. - Find wavefunction in momentum space by FT wrt all
the coordinates and integrating out all but one
electron - Expanding out the square and performing the
integrals we get. - Where
- (states occupied with the Boltzmann
distribution.) - For a homogeneous system, n(r,s)n(r-s)
31Calculation of n(r)
- Naïve procedure
- Generate sample with VMC
- Take one particle at random and displace by a
distance r. - Find ratio of trial function for old and new
position. - O(1) work/1 distance
- McMillan procedure
- Generate sample with VMC
- Determine change of trial function if each
particle is destroyed. ak - Insert a new particle at a random position in the
box and determine trial function for the
insertion b - For each k perform the average b/ak and add to
n(r-rk). - Repeat for O(N) insertions.
- O(N) work/O(N2) distances
32The electron gasD. M. Ceperley, Phys. Rev. B 18,
3126 (1978)
- Standard model for electrons in metals
- Basis of DFT.
- Characterized by 2 dimensionless parameters
- Density
- Temperature
- What is energy?
- When does it freeze?
- What is spin polarization?
- What are properties?
33Charged systems
- How can we handle charged systems?
- Just treat like short-ranged potential cutoff
potential at rgtL/2. Problems - Effect of discontinuity never disappears (1/r)
(r2) gets bigger. - Will violate Stillinger-Lovett conditions because
Poisson equation is not satisfied - Image potential solves this
- VI ? v(ri-rjnL)
- But summation diverges. We need to resum. This
gives the ewald image potential. - For one component system we have to add a
background to make it neutral. - Even the trial function is long ranged and needs
to be resummed.
34Ewald summation method
- Key idea is to split potential into k-space part
and real-space part. We can do since FT is
linear. - Bare potential converges slowly at large r (in
r-space) and at large k (in k-space)
35Classic Ewald
- Split up using Gaussian charge distribution
- If we make it large enough we can use the minimum
image potential in r-space. - Extra term for insulators
36How to do it
O(N3/2) O(N) O(N3/2) O(N3/2) O(N1/2) O(N3/2) O(
1)
- r-space part same as short-ranged potential
- k-space part
- Compute exp(ik0xi)(cos (ik0xi), sin (ik0xi)),
k02?/L ? i - Compute powers exp(i2k0xi) exp(ik0xi
)exp(ik0xi) etc. This way we get all values of
exp(ik . ri) with just multiplications. - Sum over particles to get ?k all k.
- Sum over k to get the potentials.
- Forces can also be done by taking gradients.
- Constant terms to be added.
- Checks perfect lattice V-1.4186487/a (cubic
lattice).
37Optimized EwaldJ. Comput. Physics 117, 171
(1995).
- Division into Long-range and short-ranged
function is convenient but is it optimal? No - Trial functions are also long-ranged but not
simply 1/r. We need a procedure for general
functions. - Natoli-Ceperley procedure. What division leads
to the highest accuracy for a given radius in r
and k? - Leads to a least squares problem.
- FITPN code does this division.
- Input is fourier transform of vk
- on grid appropriate to the supercell
- Output is a spline of vsr(r)
- and table of long ranged function.
38Problems with Image potential
- Introduces a lattice structure which may not be
appropriate. - Example a charge layer.
- We assume charge structure continues at large r.
- Actually nearby fluid will be anticorrelated.
- This means such structures will be penalized.
- One should always consider the effects of
boundary conditions, particularly when
electrostatic forces are around! - You need to have a continuum model to understand
the results of a microscopic simulation.
39Fermions antisymmetric trial function
- At mean field level the wavefunction is a Slater
determinant. Orbitals for homogenous systems are
a filled set of plane waves. - We can compute this energy analytically (HF).
- To include correlation we multiply by a
pseudopotential. We need MC to evaluate
properties. - New feature how to compute the derivatives of a
deteminant and sample the determinant. Use tricks
from linear algebra. - Reduces complexity to O(N2).
Slater-Jastrow trial function.
40Jastrow factor for the e-gas
- Look at local energy either in r space or
k-space - r-space as 2 electrons get close gives cusp
condition du/dr0-1 - K-space, charge-sloshing or plasmon modes.
- Can combine 2 exact properties in the Gaskell
form. Write EV in terms structure factor making
random phase approximation. (RPA). - Optimization can hardly improve this form for
the e-gas in either 2 or 3 dimensions. RPA works
better for trial function than for the energy. - NEED EWALD SUMS because potential trial function
is long range, it also decays as 1/r, but it is
not a simple power.
- Long range properties important
- Give rise to dielectric properties
- Energy is insensitive to uk at small k
- Those modes converge t1/k2
41Derivation of the e-gas Jastrow
- For simplicity, consider boson trial function
42Generalized Feynman-Kacs formula
- Lets calculate the average population resulting
from DMC starting from a single point R0 after a
time t.
43Wavefunctions beyond Jastrow
smoothing
- Use method of residuals construct a sequence of
increasingly better trial wave functions.
Justify from the Importance sampled DMC. - Zeroth order is Hartree-Fock wavefunction
- First order is Slater-Jastrow pair wavefunction
(RPA for electrons gives an analytic formula) - Second order is 3-body backflow wavefunction
- Three-body form is like a squared force. It is a
bosonic term that does not change the nodes.
44Backflow wave function
- Backflow means change the coordinates to quasi-
coordinates. - Leads to a much improved energy and to
improvement in nodal surfaces. Couples nodal
surfaces together. - Kwon PRB 58, 6800 (1998).
3DEG
45Dependence of energy on wavefunction 3d Electron
fluid at a density rs10 Kwon, Ceperley,
Martin, Phys. Rev. B58,6800, 1998
- Wavefunctions
- Slater-Jastrow (SJ)
- three-body (3)
- backflow (BF)
- fixed-node (FN)
- Energy ltf H fgt converges to ground state
- Variance ltf H-E2 fgt to zero.
- Using 3B-BF gains a factor of 4.
- Using DMC gains a factor of 4.
FN -SJ
FN-BF
46Comparison of Trial functions
- What do we choose for the trial function in VMC
and DMC? - Slater-Jastrow (SJ) with plane wave orbitals
- For higher accuracy we need to go beyond this
form. - Need correlation effects in the nodes.
- Include backflow-three body.
- Example of incorrect physics within SJ
47Analytic backflowHolzmann et al, Phys. Rev. E
68, 0467071-15(2003).
- Start with analytic Slater-Jastrow using Gaskell
trial function - Apply Bohm-Pines collective coordinate
transformation and express Hamiltonian in new
coordinates - Diagonalize resulting Hamiltonian.
- Long-range part has Harmonic oscillator form.
- Expand about k0 to get backflow and 3-body
forms. - Significant long-range component to BF
- OPTIMIZED BF
ANALYTIC BF - 3-body term is non-symmetric
rs1,5,10,20
48Results of Analytic tf
- Analytic form EVMC better for rslt20 but not for
rs?20. - Optimized variance is smaller than analytic.
- Analytic nodes always better! (as measured by
EDMC) - Form ideal for use at smaller rs since it will
minimize optimization noise and lead to more
systematic results vs N, rs and polarization. - Saves human machine optimization time.
- Also valuable for multi-component system of
metallic hydrogen.
49Twist averaged boundary conditions
- In periodic boundary conditions (? point), the
wavefunction is periodic?Large finite size
effects for metals because of shell effects. - Fermi liquid theory can be used to correct the
properties. - In twist averaged BC we use an arbitrary phase ?
as r ?rL - If one integrates over all phases the momentum
distribution changes from a lattice of k-vectors
to a fermi sea. - Smaller finite size effects
PBC TABC
50Twist averaged MC
- Make twist vector dynamical by changing during
the random walk. - Within GCE, change the number of electrons
- Within TA-VMC
- Initialize twist vector.
- Run usual VMC (with warmup)
- Resample twist angle within cube
- (iterate)
- Or do in parallel.
51Grand Canonical Ensemble QMC
- GCE at T0K choose N such that E(N)-?N is
minimized. - According to Fermi liquid theory, interacting
states are related to non-interacting states and
described by k. - Instead of N, we input the fermi wavevector(s)
kF. Choose all states with k lt kF (assuming
spherical symmetry) - N will depend on the twist angle ?. number of
points inside a randomly placed sphere. - After we average over ? (TA) we get a sphere of
filled states. - Is there a problem with Ewald sums as the number
of electrons varies? No! average density is
exactly that of the background. We only work with
averaged quantities.
52Single particle size effects
- Exact single particle properties with TA within
HF - Implies momentum distribution is a continuous
curve with a sharp feature at kF. - With PBC only 5 points on curve
- rs4 N33 polarized
- No size effect within single particle theory!
- Kinetic energy will have much smaller size
effects.
53Potential energy
- Write potential as integral over structure
function - Error comes from 2 effects.
- Approximating integral by sum
- Finite size effects in S(k) at a given k.
- Within HF we get exact S(k) with TABC.
- Discretization errors come only from non-analytic
points of S(k). - the absence of the k0 term in sum. We can put
it in by hand since we know the limit S(k) at
small k (plasmon regime) - Remaining size effects are smaller, coming from
the non-analytic behavior of S(k) at 2kF.
543DEG at rs10
TABC
TABC1/N
PBC
GC-TABC1/N
We can do simulations with N42! Size effects now
go like We can cancel this term at special
values of N! N 15, 42, 92, 168, 279,
55Brief History of Ferromagnetism in electron gas
- What is polarization state of fermi liquid at low
density? - Bloch 1929 got polarization from exchange
interaction - rs gt 5.4 3D
- rs gt 2.0 2D
- Stoner 1939 include electron screening contact
interaction - Herring 1960
- Ceperley-Alder 1980 rs gt20 is partially
polarized - Young-Fisk experiment on doped CaB6 1999 rs25.
- Ortiz-Balone 1999 ferromagnetism of e gas at
rsgt20. - Zong et al Redo QMC with backflow nodes and
TABC.
56Ceperley, Alder 80
T0 calculations with FN-DMC
- 3d electron gas
- rslt20 unpolarized
- 20ltrslt100 partial
- 100ltrs Wigner crystal
Energies are very close together at low density!
More recent calculations of Ortiz, Harris and
Balone PRL 82, 5317 (99) confirm this result but
get transition to crystal at rs65.
57Polarization of 3DEG
- We see second order partially polarized
transition at rs52 - Is the Stoner model (replace interaction with a
contact potential) appropriate? Screening kills
long range interaction. - Wigner Crystal at rs105
- Twist averaging makes calculation possible--much
smaller size effects. - Jastrow wavefunctions favor the ferromagnetic
phase. - Backflow 3-body wavefunctions more paramagnetic
58Phase Diagram
- Partially polarized phase at low density.
- But at lower energy and density than before.
- As accuracy gets higher, polarized phase shrinks
- Real systems have different units.
59Recent calculations in 2D
Tanatar,Ceperley 89 Rapisarda, Senatore 95 Kwon
et al 97
T0 fixed-node calculation Also used high
quality backflow wavefunctions to compute energy
vs spin polarization. Energies of various phases
are nearly identical Attaccalite et al PRL 88,
256601 (2002)
- 2d electron gas
- rs lt25 unpolarized
- 25lt rs lt35 polarized
- rs gt35 Wigner crystal
60Polarization of 2D electron gas
- Same general trend in 2D
- Partial polarization before freezing
- Results using phase averaging and BF-3B
wavefunctions
rs20
rs10
rs30
61Linear response for the egas
- Add a small periodic potential.
- Change trial function by replacing plane waves
with solutions to the Schrodinger Eq. in an
effective potential. - Since we dont care about the strength of
potential use trial function to find the
potential for which the trial function is
optimal. - Observe change in energy since density has mixed
estimator problems.
62Fermi Liquid parameters
- Do by correlated sampling Do one long MC random
walk with a guiding function (something
overlapping with all states in question). - Generate energies of each individual excited
state by using a weight function - Optimal Guiding function is
- Determine particle hole excitation energies by
replacing columnsfewer finite size effects this
way. Replace columns in slater matrix - Case where states are orthogonal by symmetry is
easier, but non-orthogonal case can also be
treated. - Back flow needed for some excited state since
Slater Jastrow has no coupling between unlike
spins.
63Summary of Variational (VMC)
Simple trial function
Better trial function
applications
64Summary and problems with variational methods
- Powerful method since you can use any trial
function - Scaling (computational effort vs. size) is almost
classical - Learn directly about what works in wavefunctions
- No sign problem
- Optimization is time consuming
- Energy is insensitive to order parameter
- Non-energetic properties are less accurate. O(1)
vs. O(2) for energy. - Difficult to find out how accurate results are.
- Favors simple states over more complicated
states, e.g. - Solid over liquid
- Polarized over unpolarized
What goes into the trial wave function comes out!
GIGO We need a more automatic method! Projector
Monte Carlo
65Projector Monte Carlo
- Originally suggested by Fermi and implemented in
1950 by Donsker and Kac for H atom. - Practical methods and application developed by
Kalos
66Projector Monte Carlo(variants Greens function
MC, Diffusion MC, Reptation MC)
- Project single state using the Hamiltonian
- We show that this is a diffusion branching
operator. Maybe we can interpret as a
probability. But is this a probability? - Yes! for bosons since ground state can be made
real and non-negative. - But all excited states must have sign changes.
This is the sign problem. - For efficiency we do importance sampling.
- Avoid sign problem with the fixed-node method.
67Diffusion Monte Carlo
- How do we analyze this operator?
- Expand into exact eigenstates of H.
- Then the evolution is simple in this basis.
- Long time limit is lowest energy state that
overlaps with the initial state, usually the
ground state. - How to carry out on the computer?
68Notation
- Individual coordinate of a particle ri
- All 3N coordinates R (r1,r2, . rN)
- R can depend on imaginary time, time slice or
Trotter index t or on iteration number Rt. - Total potential energy V(R)
- Kinetic energy
- Hamiltonian
69The Greens function
- Operator notation
- We define the coordinate greens function (or
density matrix by - Roughly the probability density of going from R0
to R in time t. (but is it a probability?) - Properties
70Froebinius Theorem
- When can we consider the wavefunction as a
probability? First how about the Greens
function? - But if we start with a non-negative function it
will stay non-negative, and can be interpreted as
a p.d.f. - Not true for all Hamiltonians (require
off-diagonal matrix elements to be non-positive.)
(not pseudopotentials, not magnetic fields.) - Only true for the bosonic ground state.
71Monte Carlo process
- Now consider the variable t as a continuous
time (it is really imaginary time). - Take derivative with respect to time to get
evolution. - This is a diffusion branching process.
- Justify in terms of Trotters theorem.
- Requires interpretation of the wavefunction as a
probability density. - But is it? Only in the boson ground state.
Otherwise there are nodes. Come back to later.
72Trotters theorem
- How do we find the solution of
- The operator solution is
- Trotters theorem (1959)
- Assumes that A,B and AB are reasonable
operators. - This means we just have to figure out what each
operator does independently and then alternate
their effect. This is rigorous in the limit as
n??. - In the DMC case A is diffusion operator, B is a
branching operator. - Just like molecular dynamics At small time we
evaluate each operator separately.
73Evaluation of kinetic density matrix
74Putting this together
- n is number of time slices.
- ? is the time-step
- V is diagonal
- Error at finite n comes from commutator is
roughly - Diffusion preserves normalization but potential
does not! -
75Basic DMC algorithm
- Construct an ensemble (population P(0)) sampled
from the trial wavefunction. R1,R2,,RP - Go through ensemble and diffuse each one
(timestep ?) - number of copies
- Trial energy ET adjusted to keep population
fixed. - Problems
- Branching is uncontrolled
- Population unstable
- What do we do about fermi statistics?
ndrn uprn floor function
76Sampling the normal distribution
- Inverse mapping is a little slow, also of
infinite range. - Trick generate 2 at a time r(x,y)
- Or sample angle using rejection technique
- Sample (x,y) in square
- Accept if x2y2 lt1
- Normalize to get the correct r.
77Code to sample normal distribution
- Normal distribution ltxgtx0 and lt(x-x0)2gt?2
1 xsprng()-0.5 ysprng()-0.5 r2xxyy
if (r2gt0.25) go to 1 radius sigmasqrt
(-2ln(sprng())/r2) xnormalx0xradius
ynormaly0yradius
- No trig functions, 1 log, 1 sqrt, 1 divide
- Mixes up regularity of random numbers
- Efficiency of angle generation is 4/?.
- Gets 2 ndrns each time.
78Population Bias
- Having the right trial energy guarantees that
population will on the average be stable, but
fluctuations will always cause the population to
either grow too large or too small. - Various ways to control the population
- Suppose P0 is the desired population and P(t) is
the current population. How much do we have to
adjust ET to make P(tT)P0? - Feedback procedure
- There will be a (small) bias in the energy caused
by a limited population.
79Importance SamplingKalos 1970, Ceperley 1979
- Why should we sample the wavefunction? The
physically correct pdf is ?2. - Importance sample (multiply) by trial wave
function. - Evolution diffusion drift
branching - Use accept/reject step for more accurate
evolution. - make acceptance ratiogt99 . Determines time
step. - We have three terms in the evolution equation.
Trotters theorem still applies.
Commute ? through H
80Brownian Dynamics
- Consider a big molecule in a solvent. In the high
viscosity limit the master equation
(Smoluchowski or Fokker-Planck eq.) is
Diffusion Quantum Monte Carlo without branching
is the same as Brownian Dynamics. Use same
techniques.
81Greens function for a gradient
- What is Greens function for the operator?
- This operator just causes probability
distribution to drift in the direction of F. - Smoluchowski equation for Brownian motion it was
the effect of gravitational field on the motion
of colloids. - In practice, we limit the gradient so the walk is
not pushed too far.
82- To the pure diffusion algorithm we have added a
drift step that pushes the random walk in
directions of increasing trial function - Branching is now controlled by the local energy
- Because of zero variance principle, fluctuations
are controlled. - Cusp condition can limit infinities coming from
singular potentials. - We still determine ET by keeping asymptotic
population stable. - Must have accurate time evolution. Adding
accept/reject step is a major improvement.
83- Importanced sampled Greens function
- Exact property of DMC Greens function
- We enforce detailed balance to decrease time step
errors. - VMC satisfies detailed balance.
- Typically we choose time step to have 99
acceptance ratio. - Method gives exact result if either time step is
zero or trial function is exact.
84Schematic of DMC
- Ensemble evolves according to
- Diffusion
- Drift
- branching
- ensemble
85(No Transcript)
86Local Markov process
- Let us make a course approximation to a molecular
dynamics of a large molecule in the presence of a
solvent that we want to eliminate. - Particle will follow a continuous trajectory
R(t). - Let us assume that the movement is Markovian and
local. - Describe in terms of the Greens function
?(r,r,t) - This is probability density of a particle at
(r,t) starting at (r,0). - The master equation describes its time evolution
- Evolution
- Initial condition
87General form of evolution
- What could we have for O?
- Not hard to prove that O must be a sum of first
and second derivative operators. The most
general form is - Clearly probability is conserved.
- Steady state solution is
88- How do we prove dynamics is unique?
- If it is a continuous walk-it has to be made of
little random steps. - Central limit theorem says that the only things
that survive adding together many little steps
are - Mean value this is the drift term
- The variance this is the diffusion term
- Hence the smart MC gaussian is the most general
Markovian form, if we allow a general diffusion
matrix. - If we choose it to be isotropic and independent
of position then D(R) is a constant. - it could vary in some media both in position and
direction - also it could couple particles far apart,
hydrodynamical effects.
89Moment Expansion
For the variance
90Mixed estimators
- Problem is that PMC samples the wrong
distribution. - OK for the energy
- Linear extrapolation helps correct this
systematic error - Other solutions
- Maximum overlap
- Forward walking
- Reptation/path integrals
91Forward WalkingKalos et al. 1974.
- Lets calculate the average population resulting
from DMC starting from a single point R0 after a
time t. - We can estimate the correction to the mixed
estimator by weighting with the number of
descendants of a given configuration. - Problem the fluctuations in the weights
eventually diverge. - Dont make t too large.
92Fusion sticking coefficientPhys. Rev. A 31, 1999
(1985).
- Consider the 3 body system (? d t)
- For the sticking coefficient, we need the exact
wavefunction at the point where 2 nuclei are at
the same position. (this is a singular point)
93Other projector functions can be used
- Common effect on long-time (iteration) limit.
- 3rd choice generates a Krylov sequence. Only
works for bounded spectra such as a lattice model.
G
E
94Greens Function Monte CarloKalos, Levesque,
Verlet Phys. Rev. A9, 2178 (1974).
- It is possible to make a zero time-step-error
method - Works with the integral formulation of DMC
- Sample time-step from Poisson distribution
- Express operator in a series expansion and sample
the terms stochastically. - Recent Revival Continuous time Monte Carlo for
lattice models.
95Fermions?
- How can we do fermion simulations? The initial
condition can be made real but not positive (for
more than 1 electron in the same spin state) - In transient estimate or released-node methods
one carries along the sign as a weight and
samples the modulus. - Do not forbid crossing of the nodes, but carry
along sign when walks cross. - Whats wrong with node release
- Because walks dont die at the nodes, the
computational effort increases (bosonic noise) - The signal is in the cancellation which dominates
- Monte Carlo can add but not subtract
96Transient Estimate Approach
- ?(?) converges to the exact ground state
- E is an upper bound converging to the exact
answer monotonically
97VMC
300K
DMC
timestep
bcc hydrogen, N16
98Model fermion problem Particle in a box
- Symmetric potential V(r) V(-r)
- Antisymmetric state f(r)-f(-r)
Initial (trial) state
Final (exact) state
Sign of walkers fixed by initial position. They
are allowed to diffuse freely. f(r) number of
positive-negative walkers. Node is dynamically
established by diffusion process. (cancellation
of positive and negative walkers.)
99Scaling in Released-Node
Initial distribution
Later distribution
- At any point, positive and negative walkers will
tend to cancel so the signal is drown out by the
fluctuations. - Signal/noise ratio is tprojection time
- EF and EB are Fermion, Bose energy (proportional
to N) - Converges but at a slower rate. Higher accuracy,
larger t. - For general excited states
- Exponential complexity!
- Not a fermion problem but an excited state
problem. - Cancellation is difficult in high dimensions.
100Exact fermion calculations
- Possible for the electron gas for up to 60
electrons. - 2DEG at rs1 N26
- Transient estimate calculation with SJ and BF-3B
trial functions.
101General statement of the fermion problem
- Given a system with N fermions and a known
Hamiltonian and a property O. (usually the
energy). - How much time T will it take to estimate O to an
accuracy e? How does T scale with N and e? - If you can map the quantum system onto an
equivalent problem in classical statistical
mechanics then
With 0 lta lt 4
- This would be a solved quantum problem!
- All approximations must be controlled!
- Algebraic scaling in N!
- e.g. properties of Boltzmann or Bose systems in
equilibrium.
102Solved Problems
- 1-D problem. (simply forbid exchanges)
- Bosons and Boltzmanons at any temperature
- Some lattice models Heisenberg model, 1/2 filled
Hubbard model on bipartite lattice (Hirsch) - Spin symmetric systems with purely attractive
interactions ult0 Hubbard model, nuclear Gaussian
model. - Harmonic oscillators or systems with many
symmetries. - Any problem with ltiHjgt ?0
- Fermions in special boxes
- Other lattice models
- Kalos and coworkers have invented a pairing
method but it is not clear whether it is
approximation free and stable.
103The sign problem
- The fermion problem is intellectually and
technologically very important. - Progress is possible but danger-the problem maybe
more subtle than you first might think. New ideas
are needed. - No fermion methods are perfect but QMC is
competitive with other methods and more general. - The fermion problem is one of a group of related
problems in quantum mechanics (e.g dynamics). - Feynman argues that general many-body quantum
simulation is exponentially slow on a classical
computer. - Troyer Wiese show that some quantum sign
problems are NP hard. - Maybe we have to solve quantum problems using
analog quantum computers programmable quantum
computers that can emulate any quantum system.
104Fixed-node method
- Initial distribution is a pdf.
- It comes from a VMC simulation.
- Drift term pushes walks away from the nodes.
- Impose the condition
- This is the fixed-node BC
- Will give an upper bound to the exact energy, the
best upper bound consistent with the FNBC.
- f(R,t) has a discontinuous gradient at the nodal
location. - Accurate method because Bose correlations are
done exactly. - Scales well, like the VMC method, as N3.
Classical complexity. - Can be generalized from the continuum to lattice
finite temperature, magnetic fields, - One needs trial functions with accurate nodes.
105Proof of fixed-node theorem
- Suppose we solve S.E. in a subvolume V determined
by the nodes of an antisymetric trial function.
106Nodal Properties
- If we know the sign of the exact wavefunction
(the nodes), we can solve the fermion problem
with the fixed-node method. - If f(R) is real, nodes are f(R)0 where R is the
3N dimensional vector. - Nodes are a 3N-1 dimensional surface. (Do not
confuse with single particle orbital nodes!) - Coincidence points ri rj are 3N-3 dimensional
hyper-planes - In 1 spatial dimension these points exhaust the
nodes fermion problem is easy to solve in 1D
with the no crossing rule. - Coincidence points (and other symmetries) only
constrain nodes in higher dimensions, they do not
determine them. - The nodal surfaces define nodal volumes. How many
nodal volumes are there? Conjecture there are
typically only 2 different volumes ( and -)
except in 1D. (but only demonstrated for free
particles.)
107Nodal Picture 2d slice thru 322d space
- Free electron
- Other electrons
- Nodes pass thru their positions
- Divides space into 2 regions
- Wavelength given by interparticle spacing
108SPIN?
- How do we treat spin in QMC?
- For extended systems we use the Sz
representation. - We have a fixed number of up and down electrons
and we antisymmetrize among electrons with the
same spin. - This leads to 2 Slater determinants.
- For a given trial function, its real part is also
a trial function (but it may have different
symmetries), for example momentum - For the ground state, without magnetic fields,
spin-orbit interaction we can always work with
real functions. - However, in some cases it may be better to work
with complex functions.
109Fixed-Phase methodOrtiz, Martin, DMC 1993
- Generalize the FN method to complex trial
functions - Since the Hamiltonian is Hermitian, the
variational energy is real - We see only one place where the energy depends on
the phase of the wavefunction. - If we fix the phase, then we add this term to the
potential energy. In a magnetic field we get also
the vector potential. - We can now do VMC or DMC and get upper bounds as
before. - The imaginary part of the local energy will not
be zero unless the right phase is used. - Used for twisted boundary conditions, magnetic
fields, vortices, phonons, spin states,
110Problem with core electrons
- Bad scaling in both VMC and DMC
- In VMC, energy fluctuations from core dominate
the calculation - In DMC, time step will be controlled by core
dynamics - Solution is to eliminate core states by a
pseudopotential - Conventional solution semi-local form
- Ensures that valence electrons go into well
defined valence states with the wavefunction and
energy for each angular momentum state
prescribed. - PP is non-local OK for VMC. Leads to an extra MC
integral. But DMC uses a locality approximation
and good trial functions. Extra approximation.
111Solid state / chemical applications
- Use LDA derived pseudopotentials
- Take orbitals from other methods
- Gaussian-xx give orbitals for molecules
- DFT-PW codes give orbitals for extended systems
- HF is slightly better because of self-interaction
effects within DFT - Multiply by a Jastrow function (electron gas or
otherwise). Can include higher order e-e-n terms - Must add a compensating e-n term in order to
cancel out purely repulsive character of e-e
correlation. - Assuming LDA density is correct, this can be done
by making sure VMC electron densityLDA electron
density. (Fahy correction).
112Summary of T0 methodsVariational(VMC),
Fixed-node(FN), Released-node(RN)
113Problems with projector methods
- Fixed-node is a super-variational method
- DMC dynamics is determined by Hamiltonian
- Zero-variance principle allows very accurate
calculation of ground state energy if trial
function is good. - Projector methods need a trial wavefunction for
accuracy. They are essentially methods that
perturb from the trial function to the exact
function. (Note if you dont use a trial
function, you are perturbing from the ideal gas) - Difficulty calculating properties other than
energy. We must use extrapolated estimators or
forward walking. - Bad for phase transitions and finite temperature,
complex systems. - Path Integral MC solves some of these problems.
114Reptation Methods
115Reptation Monte Carlo (RQMC)
- Similar technique to Diffusion MC
- Instead of imaginary timecomputer time, keep
entire path in memory - Update with a Metropolis based method instead of
branching diffusing random walks - Get exact properties no extrapolation or mixed
estimators - Good for energy differences.
- How to move the particles? Reptation means move
like a snake. This is how polymers can move.
116Reptation Monte Carlo
- ?(?) converges to the exact ground state as a
function of imaginary time. More accurate than
VMC - E is an upper bound converging to the exact
answer monotonically - Do Trotter break-up into a path of p steps a la
PIMC. - Bosonic action for the links
- Trial function at the end points.
- For fixed-phase add a potential to avoid the
sign problem. Exact answer if potential is
correct.
117Reptation moves
- Let d be the direction of the move
- -1 tail move
- 1 head move
- Standard method.
- Choose d at random
- Acceptance probability is
- Takes O(p2) steps to decorrelate.
- One way reptation gives the wrong answers.
- Bounce method
- add d to the state.
- Change d only on rejections.
- Use same acceptance formula!!
- Does not satisfy detailed balance but still gives
correct answer since it is an eigenfunction of T. - Moves are 1/(rejection rate) times more
effective!
118Link action
- PIMC method pair potential plus nodal action
- DMC method. Use importanced-sampled evolution to
suggest action (good for accurate trial
functions) - Symmetrize with respect to R and R to get higher
accuracy.
119VMC
300K
DMC
timestep
Tests on bcc hydrogen, N16 Good results in a
few slices p20.
120Coupled Ionic-Electronic Simulations Carlo
Pierleoni, Kris Delaney and DMCIllinois ,
LAquila
As computers get more powerful, we can use more
accurate intermolecular potentials
121Quantum Monte Carlo
- We need to use simulation techniques to treat
many-body quantum problems, just as in classical
liquids, lattice spin models - Various QMC methods for electronic structure
problem - Variational Monte Carlo
- Projector Monte Carlo e.g. Diffusion MC,
Reptation MC - Path Integral Monte Carlo
- Let us assume that QMC gives most accurate
energies for general quantum many-body systems?
we want to use it for the energy surface of an
ionic system. - We want to use the Born-Oppenheimer
approximation a big separation in electron and
ionic energies. - Try out CEIMC method on a simple but difficult
many-body system warm dense hydrogen.
122Coupled electron ion MC
- To calculate thermodynamic properties, we sample
the classical (or quantum) Boltzmann
distribution. - The electronic energy E(S) is obtained by solving
the electronic Schroedinger equation for a given
position, S, of the ions. - MC is simpler and more rigorous than MD
- No ergodic problems.
- Flexibility in choosing transition moves.
- Possibility of canceling noise. (how do we do
this in MD?)
123- In Coupled Electron-Ionic Monte Carlo (CEIMC )
- Do classical MC or Path Integral MC for the ions
at Tgt0. - Let electrons be at zero temperature, a
reasonable approximation for room temperature
simulations. - Do MC, not MD, for the ions.
- Assumptions are convenient, not necessary.
- Advantages to having T0 electrons
- Zero variance principle in ground state can
reduce computation time for high quality wave
functions. - Quality of zero temperature wavefunctions is
higher than for Tgt0 implies fewer time slices. - Possibility of using existing DFT/SCF/QMC
wavefunctions at T0. - Energy difference methods can be efficient.
124Plan of talk
- Describe some features of the CEIMC method
- The penalty method
- New trial functions
- Reptation MC
- Energy difference method
- Path Integrals for protons
- Twist averaged boundary conditions
- Hydrogen Phase diagram
- Crystal-liquid transition
- Molecular-atomic transition
- Advantages of Coupled-Electron Ion MC
125A single CIEMC step
- Move ions from S to S
- Pre-reject moves based on empirical potential
- Determine new trial function ?T (RS)
- Sample electron coordinates from P(RS,S)
- Determine energy difference E(S)-E(S).
- Use penalty method to accept or reject
R
electrons
ions
S ? S
126Basics of the classical random walk methodMarkov
chain with rejections
- The Metropolis, Rosenbluth, Teller (1953) method
- Move from S to S with probability
T(S?S)T(S?S) - Accept move with probability
- a(S?S) min 1 , exp ( - (E(S)-E(S))/kBT )
- Repeat many times
- Given ergodicity, the distribution of the state,
S, will be - exp(-E(S)/kBT)/Z
- E(S)energy of ionic arrangement S,
- Zpartition function.
- Only the difference in energy enters in the
acceptance probability.
127Problem QMC energy difference will be noisy
Average energy of Lennard-Jones liquid
- Ignoring noise gives a systematic increase in the
energy because high energy moves are occasionally
wrongly accepted. - Acceptance formula is non-linear min 1 , exp (
- ?E/kBT ) - But it is possible get the exact distribution,
independent of noise level!!
128Approaches to deal with this noise
- Ignore problem will smooth out distributions.
- Compute energy difference very accurately takes
a lot of computer time!! - Langevin Dynamics (Grossman and Mitas PRL, 2005)
- Solve MD with random component to forces.
- Noise must be small with respect to time step and
temperature. - Dynamics is artificial since noise doesnt
correspond to a physical process. - Linear Noise (Kennedy and Kuti, PRL 54, 2473,
1985). - Used in lattice gauge simulations where noise is
small. - Sign problem for large fluctuations
- Non-optimal it increases rejections
unnecessarily, especially for large noise. - Our approach modify the formula for accepting
moves. - Similar to fuzzy MC by Krajcí and Hafner, PRL
74, 5100 (1995) but with higher efficiency and
fewer approximations. - Recent development (more general distributions)
- R. C. Ball et al. PRL 91, 030201 (2003).
129The Penalty method DMC Dewing, J. Chem. Phys.
110, 9812(1998).
- Assume estimated energy difference ?e is normally
distributed with variance ?2 and the correct
mean. - lt ?e gt ?E
- lt ?e- ?E2 gt ?2
- OK because of central limit theorem for ?lt?
- a(?e ?) is acceptance ratio.
- average acceptance A(?E) lt a(?e) gt
- Markov chain goes to the correct distribution if
flux of transitions from S to S is symmetric
detailed balance A(?E) exp (- ?E ) A(-?E) - An exact solution is to use a modifed acceptance
formula a(x,?) min 1, exp(-x- ?2/2) - ?2/2 is penalty additional rejections caused
by the noise.
130Estimating the error
- Problem both the mean and variance are unknown
and are spatially varying. - Both energy and variance are estimated from the
data and must be included in the acceptance rate
formula a(?,v) - Sample M independent estimates of ?E ek
1?k?M. - ?? ek/M ? ?E
- v? ek - ?2/(M(M-1)) ? ?2
- 2 parameter integral equation, normal in ? and
?2(vM) in v. - Unique smooth solution is a Bessel function
JM/2(?Ms). Expand - a(?,s) exp - ? - v/2 - v2/4(M1)
-v3/3(M1)(M3) - Error of noise causes an additional penalty!
- Asymptotic series only good for vlt4M. But very
accurate for vlt0.1M.
1312 Parameter Integral Equation
The estimated variance, s, is distributed
according to a chi-squared distribution with
(N-1) degrees of freedom.
Detailed Balance Equation
to be satisfied for all values of y and ?2 .
132Optimization of Efficiency w.r.t. Noise Level
- As the noise level increases, the acceptance
ratio decreases because of the penalty. - Assume it is usual MC rule to reduce the
variance. That is - ?2t/M where M data points ? CPU
time ? (?/?)2 constant - ? error of the energy per step.
- For double well, optimal noise turns out to be
3kBT! - Large noise (order k B T) is much more efficient
than small noise. - Trying many cheap steps, rejecting most of
them, is more successful than spending much
computer time crafting a step which will be
accepted. - Nonetheless there is n upper limit to the noise
because you need to estimate the variance also.
Average time to achieve a given error in mean
energy
133Other uses of the Penalty method
- Consider a problem where there are a finite but
large number of terms in the potential. - Example long range potential (order N terms
after 1 particle is moved.) Ewald sums for
coulomb potential scale as N1/2 . - Fast multipole method O(1) is only appropriate
if all particles are moved together as in
molecular dynamics. - We must compute the neighboring term
directly--otherwise error will not be normally
distributed. Split pair potential into
short-range and long-range terms v(r)
vlr(r)vsr(r) - We can sample terms at position, r, of the
long-range part. Since it does not diverge
anywhere, its variance is finite. - The optimal sampling probability is
134Pre-rejection
- For efficiencys sake we want to make moves as
far as possible. - We use an empirical potential (Vm) to decide if a
move is at all reasonable (first level
rejection) choose Vm self-consistently. - 2-stag