F5SLD's Free HAM Magazine 100% HAM RADIO ANTENNAS FOR LIMITED SPACE
Kohns Ham
-
Upload
ying-hsuan-lin -
Category
Documents
-
view
219 -
download
0
Transcript of Kohns Ham
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 1/93
Introduction to Kohn-Sham Density Functional
Theory: Analysis and Algorithms
Weinan E 1 and Jianfeng Lu 2
1Princeton University
2Courant Institute of Mathematical SciencesNew York University
Collaborators: Roberto Car, Carlos Garcıa-Cervera, Weiguo Gao,Lin Lin, Juan Meza, Chao Yang, Xu Yang, Lexing Ying.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 2/93
Fundamentals
Macroscopic limit of density functional theoryDerivation of nonlinear elasticity and macroscopic electrostaticequation from Kohn-Sham DFT
Derivation of macroscopic Maxwell equation from TDDFT
AlgorithmsIntroductionDiscretization of the Kohn-Sham HamiltonianRepresentation of the Fermi OperatorDensity evaluation
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 3/93
Quantum many-body problem
The (non-relativistic) ground state electronic structure of a system
is determined by the lowest eigenvalue and eigenfunction of themany-body time independent Schrodinger operator (omitting spin):
H Ψ =
i −1
2 ∆x i +i < j 1
|x i
−x j
| −i ,α
Z α
|x i
−X α
|Ψ = E Ψ,
within the Born-Oppenheimer approximation.
E (X α) = inf Ψ=1
Ψ|H |Ψ.
Here the many-body wave function Ψ(x 1, x 2, . . . , x N ) is anantisymmetric function of N variables, according to Pauli’sexclusion principle.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 4/93
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 5/93
Tight-binding model
Consider one-body effective Hamiltonian, and assume one-bodywave functions take the form
ψ(x ) =α,i
c i ,αϕi (x − X α),
where ϕi are atomic orbitals. Similar to a generalized finiteelement discretization.
Hence, the Hamiltonian becomes a matrix acting on c i ,α.
Nonlinear tight-binding model is also used sometimes. Theeffective Hamiltonian depends on the density (like a discreteversion of density functional theory)
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 6/93
Hartree-Fock theory
Assumes the many-body wave function is a single Slater
determinant
Ψ = 1√
N !det
ψ1(x 1) ψ2(x 1) · · · ψN (x 1)ψ1(x 2) ψ2(x 2) · · · ψN (x 2)
... ...
. . . ...
ψ1(x N ) ψ2(x N ) · · · ψN (x N )
with ψi a set of orthonormal functions.
For example, in this approximation, the kinetic energy reduces to R3N
Ψ∗(−i
12 ∆x i )Ψ =
i
1
2
R3
|∇ψi (x )|2.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 7/93
Hartree-Fock theory (cont’d)
Hartree-Fock equation
−1
2∆ + V ext +
R3
ρ(y )
|x − y | dy − Kψi = λi ψi .
K is the exchange operator:
Kψi = ψ∗ j (y )ψi (y )
|x − y | dy ψ j (x ).
As the energy is minimized over a smaller space,
E ≤ E HF = E + E c .
The error made E c is called correlation energy, as a result of ignoring many-body interactions (besides Coulomb and exchange).
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 8/93
Density functional theory
Brief history: Thomas-Fermi: Concepts of using solely the density to
describe quantum mechanics (simple empirical models).
Hohenberg-Kohn: Proves that for ground state in quantum
mechanics is indeed only a function of the electron density. Kohn-Sham: Mean field theory for non-interacting electrons in
an effective potential.
Levy-Lieb variational principle: Mathematical rigorousfoundation.
Development of exchange-correlation functional (Becke,Burke, Ernzerhof, Parr, Perdew, Yang, ...)
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 9/93
Density functional theory [Hohenberg-Kohn 1964]: The many-bodyvariational problem can also be reformulated using electron densityas variable:
E = inf ρ≥0,√ ρ∈H 1R ρ=N
F [ρ].
The basic variable is the electron density:
ρ(x ) = N |Ψ(x , x 2, . . . , x N )|2
dx 2 · · · dx N .
Levy-Lieb formulation:
E = inf ρ≥0,√ ρ∈H 1R
ρ=N
inf Ψ∈H 1,Ψ→ρ
Ψ antisymmetricΨ
|T + V ee
|Ψ
+ V extρ
≡ inf ρ≥0,
√ ρ∈H 1R
ρ=N
F KS[ρ].
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 10/93
Kohn-Sham density functional theory
The energy functional depends on N electron orbitals
ψi
:
F KS(ψi ) =i
1
2
|∇ψi |2
+ 1
2 (ρ − m)(x )(ρ − m)(y )
|x − y | + E xc[ρ].
The orbitals are orthonormal, ρ is the electron density given byρ(x ) =
i |ψi (x )|2. m is the background charge distribution.
Formally exact. All errors are encoded in the last term, theexchange-correlation energy, which contains chemistry, as itmodels the quantum correlation of electrons. However, the explicitform is unknown and needs approximation.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 11/93
Input to the model
Background charge: m(y ) =
y j ∈Ω ma j (y − y j )
y j = positions of the nuclei (ions). ma
j = ionic potential describing the atoms in the system.
All electron model: ma
j = delta function Valence electron model (view core electrons as part of thenuclei): ma
j is (local) pseudopotential.
Chemistry: y j ,ma j describes a set of molecules.
Materials: y j ,m
a
j describes a deformed lattice with defects.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 12/93
Local density approximation
In principle E xc[ρ] is a nonlocal functional depending on ρ. In thelocal density approximation, it is assumed that E xc[ρ] is a local
functional:E xc[ρ] =
xc(ρ(x )) dx .
Taking LDA, the energy functional becomes
F KS(ψi ) =i
1
2 |∇ψi |2
+ 1
2 (ρ − m)(x )(ρ − m)(y )
|x
−y
|
+
xc(ρ(x )).
The function xc is obtained from calculations of Jellium system.
xc(ρ) = −ρ4/3 + ...
.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 13/93
Developments about exchange-correlation functional
Theoretically, it is a universal functional of the density field.
In practice, it is obtained by physics intuition and argument,plus fitting from quantum Monte Carlo calculations.
Examples of functional forms: Becke88, LYP, PBE, so on; Jacob’s ladder of exchange-correlation energy: meta-GGA
(TPSS), hybrid functionals (B3LYP), etc.
Question: Mathematical derivations? Which asymptotic regime?
(Burke, Friesecke, Solovej, ...)
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 14/93
Alternative formulations
Several alternative formulations for the Kohn-Sham densityfunctional theory:
orbitals or wavefunctions density matrix or projection operator, or subspace formulation
density (in terms of the Kohn-Sham map)
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 15/93
Subspace problem
The energy functional F KS(ψ j ) is invariant under rotations of
the wave functions. More generally, define the non-orthogonalenergy functional as
F KS(ψi ) =
ij 1
2 ∇ψ∗i S −1
ij ∇ψ∗ j
+ 1
2
(ρ − m)(x )(ρ − m)(y )
|x − y | + xc(ρ(x )).
with S ij = ψi , ψ j and ρ(x ) =
ij S −1ij ψ
∗i (x )ψ j (x ). The energy
functional is invariant under general non-degenerate lineartransformation.
The Kohn-Sham density functional theory is a minimization overoccupied subspace. The particular basis ψ j is not relevant.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 16/93
Euler-Lagrange equation
The Euler-Lagrange equation of E KS with respect to ψi gives theeigen-equations
H [ρ]ψi =−1
2 ∆ +
(ρ − m)(y )
|x − y | + V xc(ρ)ψi = i ψi
H [ρ] is the effective Hamiltonian, depends on ρ and hence on ψi .This is a self-consistent equation (nonlinear eigenvalue problem).
The effective potential (depending on ρ) consists of two parts:Coulomb and exchange-correlation.
The Kohn-Sham equation is local, unlike Hartree-Fock equation,which contains the nonlocal exchange operator.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 17/93
Kohn-Sham map
Given an effective Hamiltonian H [ρ], the density corresponding tothe occupied states can be written as
ρ(x ) = φ0
FD(H [ρ] − µ)(x , x )
where φ0
FD
(x ) = χx ≤0
is the Heaviside function, Fermi-Diracdistribution at zero temperature. µ is the chemical potential,suitably chosen so that
ρ = N .
The map from ρ to
ρ is called Kohn-Sham map F .
This is extended to the finite temperature case by takingFermi-Dirac distribution function instead of the Heaviside functionin the above (corresponds to Mermin functional).
O f f
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 18/93
Orbital free density functional theory
The orbital-free density functional theory is a further simplificationof the Kohn-Sham DFT so that the functional only involves thedensity.
In particular, the kinetic energy functional is replaced by afunctional depends on ρ only, approximates
T [ρ] = inf ψ j →ρ
j
R3
|∇ψ j |2 dx .
Thomas-Fermi approximation: R3 ρ5/3
dx .
Thomas-Fermi-von Weiszacker approximation: R3 ρ5/3 dx + R3|∇√
ρ|2dx based on gradient expansion.
Wang-Teter and Wang-Govind-Carter functionals based onlinear response.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 19/93
On the mathematical level, the nature of orbital-based andorbital-free DFTs are quite different.
The orbital-free DFT is more of a conventional variational problemin applied mathematics, like Landau-Lifschitz, Ginzburg-Landau,
liquid crystals, nonlinear elasticity and so on.
We will focus mainly on orbital-based Kohn-Sham densityfunctional theory from now on.
I f K h Sh d i f i l h
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 20/93
Issues of Kohn-Sham density functional theory
On the analysis side:
Existence is not trivial due to possible loss of compactness; Uniqueness is not always expected as the functionals are
non-convex;
The property and structure of the solutions are not easy to
investigate.On the numerics side:
Conventional cubic scaling algorithms is too expensive.Calls for fast algorithms and efficient parallelization to address
large systems. Choices of discretization to achieve the balance between
accuracy and efficiency;
Issue of numerical analysis: accuracy, convergence, so on.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 21/93
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 22/93
Fundamentals
Macroscopic limit of density functional theoryDerivation of nonlinear elasticity and macroscopic electrostaticequation from Kohn-Sham DFTDerivation of macroscopic Maxwell equation from TDDFT
AlgorithmsIntroductionDiscretization of the Kohn-Sham Hamiltonian
Representation of the Fermi OperatorDensity evaluation
Moti atio
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 23/93
Motivation
Physical systems (solids, materials) can be modeled at differentscales:
Quantum mechanics: Many-body Schrodinger equation,electronic structure models, lattice models, ...
Atomistic models: Molecular statics and dynamics withempirical potentials;
Continuum theories: Elasticity, dielectricity, micromagnetism,phase field models, ...
Multiscale modeling and analysis: Understanding the connectionsand coupling between models on different scales.
Continuum theories
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 24/93
Continuum theories
We will focus on in this talk two representatives of continuummodels:
Nonlinear elasticity:
inf u
W (∇u (x )) − f (x )u (x ) dx .
Macroscopic Maxwell equation:
∇ · D = ρf ,
∇ · B = 0,
∇ ×E =
−∂ t B ,
∇ × H = J f + ∂ t D ;
The continuum theories are obtained by physical principle(minimum action principle, conservation laws, ...) plus empirical
constitutive relations.
Macroscopic limit
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 25/93
Macroscopic limit
Want to understand the following questions for the connectionsbetween micro and macro models:
How can we obtain the constitutive relations of themacroscopic models from the microscopic ones?
When are the macroscopic models valid characterizations of
the system?
How does the failure of the macroscopic models happen?What is the onset of breaking down?
For example, elasticity theory
→ plasticity, fracture.
These questions can be addressed by studying macroscopic limit of microscopic models.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 26/93
Fundamentals
Macroscopic limit of density functional theoryDerivation of nonlinear elasticity and macroscopic electrostaticequation from Kohn-Sham DFTDerivation of macroscopic Maxwell equation from TDDFT
AlgorithmsIntroductionDiscretization of the Kohn-Sham Hamiltonian
Representation of the Fermi OperatorDensity evaluation
Related works
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 27/93
Related works
For perfect crystal, the thermodynamic limit was studied for KSDFT model without exchange-correlation
[Catto-Le Bris-Lions 2001].
For perfect crystal with local defects, the macroscopic limit wasstudied
for KSDFT model without exchange-correlation[Cances-Deleurence-Lewin 2009, Cances-Lewin 2010].
For elastically deformed crystal, the macroscopic limit was studied for KSDFT model [E-Lu preprint].
Derivation of nonlinear elasticity and macroscopic
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 28/93
Derivation of nonlinear elasticity and macroscopicelectrostatic equation from DFT
For Kohn-Sham DFT, under sharp stability conditions The equilibrium system is insulating;
The equilibrium system is stable with respect to plasmons;
The effective dielectric constant for the equilibrium system is
positive definite,the electronic structure for the elastically deformed system ischaracterized by the Cauchy-Born rule (the electron density andlocal energy density is determined by the local deformationgradient).
The dielectric response of the system couples with the elasticdeformation (sort of piezoelectric effects), and is characterized byan effective Poisson equation.
Cauchy-Born rule for electronic structure
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 29/93
Cauchy Born rule for electronic structureCauchy-Born rule is a recipe links together micro- and macro-scalemodels.
Main idea: On the microscopic scale, the smooth macroscopicelastic displacement is effectively linear. Hence, at each point x ,the energy density and also other physical quantities should begiven by those of a system with homogeneous deformation withdeformation gradient
∇u (x ) (Independent of u (x ) due to
symmetry invariance).
Smoothly deformed crystal
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 30/93
Smoothly deformed crystal
Assumption: Atoms follow the prescribed smooth displacement
field u .
At equilibrium, the atoms form a crystal εL, with underlyingBravais lattice εL and unit cell εΓ. Take a smooth Γ-periodicdisplacement field u (x ), so that under deformation the atoms arelocated at Y εi = X εi + u (X εi ), where X εi ∈ εL.
Remarks about periodicity of u :
The period gives the characteristic length of u : O(1);
With the PBC assumption, only the bulk behavior is present.The interesting surface phenomenon (e.g., surface plasmon) isruled out.
Continuum limit
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 31/93
Continuum limit
The small parameter ε is understood as the ratio between the
lattice constant (atomic length scale) and the characteristic lengthscale of the elastic deformation.
The continuum limit ε → 0 will be considered. Physically, theatomic spacing is tiny compared to the macroscopic deformation.
Questions:
Derivation of nonlinear elasticity from KSDFT;
Characterization of the electronic structure for deformed
crystals;
Identification of the onset of failure of nonlinear elasticitymodel (sharp stability conditions).
Electronic structure for perfect crystal
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 32/93
Electronic structure for perfect crystalAt equilibrium, we assume there exists a Γ-periodic electron densityρe , such that
ρe (x ) = F (ρe )(x ) = 12πi
C
1λ − H e [ρe ]
dλ(x , x ).
The effective potential V e is also Γ-periodic and one can apply theBloch-Floquet theory
H e =
Γ∗H ξ dξ, H ξ =
n
E n(ξ )|ψn,ξψn,ξ|.
We assume that the system under consideration is an insulator:
dist(σZ , σ(H e )\σZ ) = E g > 0
where σZ =
n≤Z E n(Γ∗). Therefore, the Kohn-Sham map is welldefined (through Bloch-Floquet theory) for a compact contour C
encloses the occupied spectrum.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 33/93
Effective Hamiltonian operator
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 34/93
pHamiltonian (in the Lagrangian coordinates):
H ετ =
−ε2J 1/2∆τ J −1/2 + V ετ (x )
with ∆τ = τ ∗∆τ ∗ and J (x ) = det(∇τ (x )).
Potential (Coulomb + exchange-correlation):
V ετ [ρ](x ) = φε
τ [ρ](x ) + η(J (x )
−1ε3ρ(x ));
− ∆τ φετ [ρ] = 4πεJ −1(ρ − mετ );
Γφετ [ρ] dx = 0.
The deformation u enters through the background chargedistribution:
mετ (x ) = J (x )
X εα∈εL1
ε3ma((τ (x ) − τ (X εα))/ε)
Kohn-Sham equation
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 35/93
q
Look for fixed point of the Kohn-Sham map for the deformedsystem:
ρ(x ) = F ετ (ρ)(x ) = 1
2πi
C
1
λ − H ετ [ρ] dλ(x , x ).
Here C is a fixed contour in the resolvent set of the Hamiltonianenclosing the occupied spectrum.
The right hand side is the diagonal of kernel of a operator (not intrace class), which is not a priori well defined. Nevertheless, it can
be proved to be well defined for the cases we will consider byelliptic regularity results.
Cauchy-Born rule for electron density
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 36/93
y y
For the system deformed homogeneously with deformation gradient
A = ∇u (x 0), we still have a periodic problem.
By implicit function theorem + stability condition, when |A| issufficiently small, there exists a solution to the Kohn-Shamequation, denoted as ρCB(x ; A). Also, ρCB(
·; A) is Γ-periodic (as
defined in Lagrangian coordinates).
According the Cauchy-Born philosophy, one expects
ρε(x )
∼ε−3ρCB(x /ε;
∇u (x )).
In other words, electron density is given locally by that of thehomogeneous deformed system.
Locality of quantum system
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 37/93
Recall that the quantum mechanical model is a rather nonlocalfrom the first sight.
Coulomb interaction is nonlocal;
The Pauli exclusion principle (orthogonal constraint) isnonlocal;
The Schrodinger eigenvalue problem is nonlocal.
Nevertheless, the Cauchy-Born rule states that the electronicstructure at a point only depends on the local surrounding
environment. This is related with the property of “near-sightedness” in the physics literature: for insulators, thephysics is essentially local [Kohn 1996, Prodan-Kohn 2005].
Stability conditions
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 38/93
The Cauchy-Born rule is valid under sharp stability conditions. Inother words, the electron density is well approximated by theCauchy-Born guess, provided that the equilibrium system is stable :
Stability of band gap: The equilibrium system is a band
insulator. Stability wrt charge density wave: The linearized Kohn-Sham
operator is invertible in suitable spaces.
Stability of dielectric response: The permittivity tensor is
positive definite (The effective Poisson equation is elliptic).
Stability conditions
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 39/93
We will assume that the electronic structure is stable, in the senseof the following two assumptions.
Assumption (Stability of charge density wave)
For every n ∈ N, I − Le as an operator on H −1n ∩ H 2n is invertible,
and the norm of its inverse is bounded independent of n:
( I − Le )−1
L ( H −1n ∩H 2n) 1.
Assumption (Stability of dielectric response)The effective permittivity tensor E is positive definite.
Stability of dielectric response
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 40/93
Recall E = 12 (Ae + A∗
e ) + 14π I is the macroscopic permittivity
tensor for the undeformed crystal.
The matrix Ae = (Ae ,αβ ) for α, β = 1, 2, 3 is given by
Ae ,αβ = − 2
n≤Z m>Z
Γ∗
dξ
E n(ξ ) − E m(ξ )
× u m,ξ, ∂ ξβu n,ξu m,ξ, ∂ ξαu n,ξ− g e ,α, δ ρe V e ( I − Le )
−1g e ,β ,
and
g e ,α(z ) = 2n≤Z
m>Z
Γ∗
dξ
E n(ξ ) − E m(ξ )
× u ∗n,ξ(z )u m,ξ(z )
u m,ξ, i ∂ ξαu n,ξ
;
Linearized Kohn-Sham map
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 41/93
Consider the linearization of the Kohn-Sham map
F e at the
equilibrium density ρe . Formally, Le w → Le (w ), where
Le (w ) = 1
2πi
C
1
λ − He (ρe )δ ρe V e (w )
1
λ − He (ρe ) dλ(x , x ).
Here δ ρe V e is the linearized operator of V e (ρ) at ρe , given by (fornΓ-periodic function w )
δ ρe V e (w )(x ) = δφe (w )(x ) + η(ρe )w (x );
−∆δφe (w )(x ) = 4πw ,
with periodic boundary condition on nΓ and nΓ δφe = 0 to fix the
arbitrary constant.
Spaces of periodic functions
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 42/93
For a given n, define the periodic Sobolev space
W m,p n (R3) =
f
∈S
(R3)
|τ R f = f ,
∀R
∈nL; f
∈W m,p (nΓ)
with its natural norm f W m,p n (R3) = f W m,p (nΓ). We will also
write H mn for W m,2n . Here, (τ R f )(x ) = f (x − R ).
Moreover, define the periodic Coulomb space (homogeneousSobolev space with index
−1) H −1
n (R3) as
H −1n (R3) = f ∈ S
(R3) | τ R f = f , ∀R ∈ nL;
k ∈L∗/n1
|k |2|
f (k )|2 < ∞.
Here, f (k ) denotes the Fourier coefficients of the nΓ-periodicfunction f . Also the higher order spaces H m+1
n , defined by thenorm
f
˙H
m+1
n (R3
)
=∇
f H m
n (R3)3 .
Analysis of the linearized Kohn-Sham map
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 43/93
Theorem (Uniform boundedness of Le )The operator Le is bounded on spaces H −1
n ∩ H 2n uniformly in n.
Write Le w = χe δ ρe V e w with the polarizability operator
(χe V )(x ) = 1
2πi
C
1
λ − HV 1
λ − H dλ(x , x ).
The proof consists of showing
δ ρe V e : ˙H −
1
n ∩ H
2
n → H
2
n + ˙H
4
n is uniformly bounded; χe : H 2n + H 4n → H −1
n ∩ H 2n is uniformly bounded.
Polarizability operator
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 44/93
One may represent χe in more explicit terms:
χe V = 2n≤Z
m>Z
(Γ∗)2
dξ dζ
ψn,ξ,V ψm,ζ
E n(ξ ) − E m(ζ ) ψn,ξψ∗m,ζ .
For example, for Jellium model (H = −∆), we have
χe V (k ) = m(k ) V (k ),
where
m(k ) = 1
8βπ2k ∞
0
d
ln
1 + e −β (( − k /2)2 − µ)
1 + e −β (( + k /2)2
−µ) .
We remark that the behavior the linearized Kohn-Sham map israther different for metal and insulator. In particular, theboundedness results only hold for insulators (due to a finite bandgap).
Ingredients of the estimate for χe :
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 45/93
L1n → L1
n: Show the operator is in trace class, using
f (x )g (
−i
∇ξ)
Sp (L2
n)
f
Lp n
g
l p (L∗/n),
L∞n → L∞n : Based on Agmon type’s argument of regularity of kernel of elliptic operators.
H 2n → H 2n : Estimates of commutators. H 1n
→ H −1
n : Based on the generalized trace class technique
[Hainzl-Lewin-Sere 2005, Cances-Deleurence-Lewin 2008]. H 3n → H 2n : Uses the “trick of projection”
Q V = 1
2πi
C
1
λ − HV 1
λ − H dλ
= 12πi
C
P 1λ − HV 1
λ − HP ⊥ dλ
+ 1
2πi
C
P ⊥ 1
λ − HV 1
λ − HP dλ,
where P is the projection operator on the occupied space.
What happens when the stability condition is violated?
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 46/93
Instability of charge density wave (plasmon instabilities):
Wigner crystal: Crystal formation of Fermi gas at low density
Change of lattice structure: The electrons and nuclei may
have different (and possibly incommensurate) crystalstructures
Defects formation: Possibility to have electronic defects (inanalogy to crystal defects).
More on stability
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 47/93
There are three scales of instability in the system (electronic,
atomic and continuum). Due to our assumption that the atomsfollow the smooth displacement, we are only looking at thestability of electronic structure.
Without the assumption, the situation becomes more complicated.Then the system could present instability in electronic structureand also instability in atom positions (phonon analysis).
It is of interest to identify the boundary of these instabilities.Questions like whether the atomistic stability still holds while theelectronic structure become unstable.
Main results
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 48/93
Theorem (E-Lu, preprint)
Under stability assumptions, there exist constants a, ε0 and M,such that if ε ≤ ε0 and if M A = sup j ∇ j u L∞ ≤ a, then there exists ρε ∈ L∞ε with the property:
ρε is a solution to the Kohn-Sham equation:
ρε(x ) = F ετ (ρε)(x ).
ρε − ε−3ρCB(x /ε; ∇u (x ))L∞ε ≤ M ε1/2, i.e.,
ρε − ρCB(x ; ∇u (εx ))L∞ ≤ M ε1/2, where
ρε(x ) = ε3ρ(εx ),
and ρε satisfies the normalization constraint: Γρε(x ) dx = Z ε−3.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 49/93
Theorem (E-Lu, preprint)
Moreover, The macroscopic potential satisfies a second-order elliptic equation of the form:
Aαβ ∂ x α∂ x βU 0(x ) + 1
4π
L2U 0(x ) + Bα∂ x αU 0(x )
+ D − m2(x , ·) = 0
This is a piezo-electric effect: Mechanical deformationintroduces an electric potential.
Sketch of the proof
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 50/93
General strategy: Build a higher order approximated solution to the Kohn-Sham
equation ρ0 using asymptotic expansion:
ρ0 = ε−3ρCB(x /ε; ∇u (x )) + ε−2ρ1(x , x /ε) + · · · Start from the approximated solution ρ0, use Newton iteration
to find the unique fixed point nearby. The stability conditionguarantees the convergence and uniformality with ε.
The overall strategy is standard, but technically both steps arenon-conventional and require careful analysis.
Some remarks
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 51/93
The Cauchy-Born rule for the energy and also the expressionfor the stored energy density follow.
To simplify the presentation, the result is not written in itsoptimal form. In particular, the norm can be sharpened intoH −1ε ∩ H 2ε and higher order approximation can be constructed.
The effective permittivity at macroscopic level has beenstudied for undeformed crystal in reduced Hartree-Fock model
in [Cances-Lewin 2010], in the cases that the amplitude of theexternal potential is small (linear perturbation regime).
Characterization of the occupied space
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 52/93
We know that for the perfect crystal, the electronic structure can
be understood by Bloch-Floquet theory, which gives a clearpicture. In particular, the occupied space can be represented by theBloch wave functions ψn,ξ(x ).
H ξψn,ξ = E n,ξψn,ξ.
Another useful set of basis functions for the occupied space isWannier functions
W n,R (x + R ) = Γ∗ e i ξ·(R −R )ψn,ξ(x ) dξ,
The Wannier functions decays exponentially for band insulators[Kohn 1959, Nenciu 1983, Panati 2007].
Wannier functions for deformed crystals
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 53/93
How about deformed crystals that translational symmetry is
broken? The Bloch-Floquet theory no long applies. What is thegeneralization of Wannier functions?
Project Wannier functions for the equilibrium state onto theoccupied space of the deformed system.
Let W εe ,k be the set of (rescaled) Wannier functions for theequilibrium configuration. The projected Wannier functions
W ετ,k = P ετ W εe ,k
form a basis for the occupied space for H ετ provided that thedisplacement is small. The functions W ετ,k are still exponentiallylocalized.
Cauchy-Born rule for the projected Wannier functions
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 54/93
Denote the center of the Wannier function for the equilibrium
configuration W εe ,k as centered at c
εk = εc k .
Take the Wannier function associated with the homogeneouslydeformed system: W ε
A(c εk
),k with A(c εk ) = ∇u (c εk ).
Theorem (E-Lu, 2010)Under the same assumption as the previous theorem, we have
(I + ε2∆)(W ετ,k − W εA(c εk
),k )L2 ≤ C ε.
In other words, the Cauchy-Born construction gives a goodapproximation to the Wannier functions of the deformed system.The Wannier functions are stable under deformation.
Example of Cauchy-Born rule for Wannier functions
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 55/93
Figure: Example of Cauchy-Born rule construction of Wannier functions:Parameter ε = 1/32
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 56/93
Fundamentals
Macroscopic limit of density functional theoryDerivation of nonlinear elasticity and macroscopic electrostaticequation from Kohn-Sham DFTDerivation of macroscopic Maxwell equation from TDDFT
AlgorithmsIntroductionDiscretization of the Kohn-Sham Hamiltonian
Representation of the Fermi OperatorDensity evaluation
Time dependent density functional theory
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 57/93
Time dependent (current) density functional theory:
i ∂ψ
j ∂ t = 12me
−i ∇ − e c (A + Aext)2
ψ j + e (V + V ext)ψ j ,
− ∆φ = e
0(ρ − m),
1
c
∂
∂ t 1
c
∂
∂ t A +
∇φ−
∆A = e
c 0
J ,
∇ · A = 0,
V (t , x ) = φ(t , x ) + η(ρ(t , x )),
Model assumptions:
ALDA approximation for exchange-correlation scalar potential;
No exchange-correlation vector potential;
Spin degeneracy is ignored.
Time dependent density functional theory (cont’d)
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 58/93
The electron density and current are given by
ρ(t , x ) =N
j =1
|ψ j (t , x )|2,
J (t , x ) =
me
N
j =1
(ψ∗ j (t , x )∇ψ j (t , x )) − e
me c ρ(t , x )A(t , x ).
The system describes the quantum dynamics of electrons under theeffect of external potentials Aext and V ext.
The system consists of nonlinear Schrodinger equations (manyelectrons) coupled with microscopic (vacuum) Maxwell equations.
Nondimensionalization and macroscopic scalingWe consider the physical situation that the external fields V ext and
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 59/93
p yAext are slowly varying in space. Denotes ε as the small parametercharacterizing the ratio between the lattice parameter and the
external fields.
i ∂ψ j ∂ t
= 1
2 (−i ε∇ − ε(A + Aext))2 ψ j + (V + V ext)ψ j ,
−∆φ = ε(ρ
−m),
∂ 2
∂ t 2A − ∆A +
∂ ∂ t
∇φ = ε2J ,
∇ · A = 0,
V (t , x ) = φ(t , x ) + η(ε3ρ(t , x )).
ρ(t , x ) =
|ψ j (t , x )|2,
J (t , x ) = ε ψ∗ j (t , x )∇ψ j (t , x ) − ερ(t , x )A(t , x ).
Macroscopic limit
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 60/93
Assume that without external fields, the ground state of the
system forms a perfect lattice and is an insulator (finite gap in thespectrum).
Questions:
Derivation of macroscopic Maxwell equation, in particular, the
constitutive relations; Identification of sharp stability conditions (future work);
Dynamic coupling of elastic deformation and electromagneticfields (future work).
[E, Lu and Yang, in press]
Macroscopic Maxwell equation
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 61/93
By asymptotic analysis, the following macroscopic Maxwellequation is obtained in the macroscopic limit:
∇ · E (ω) E (ω, x ) = ρext(ω, x ),
∇ · B (ω, x ) = 0,
∇ × E (ω, x ) = i ω
B (ω, x ),
∇ × B (ω, x ) = −i ωE (ω) E (ω, x ) + J ext(ω, x ),
with ρext, J ext given by V ext and Aext.
Here the effective dynamic permittivity E is given by the electronic
structure (at equilibrium). Note that the permeability tensor is thesame as in the vacuum (comments ...)
Note that the above system is written in Fourier space and isdispersive.
Dynamic permittivity tensor
W h E ( ) δ + A ( )
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 62/93
We have E αβ (ω) = δ αβ + Ae ,αβ (ω).
Ae ,αβ (ω) =n≤Z
m>Z
Γ∗
dξ ω + ωmn(ξ )
u n,ξ, ∂ ξαu m,ξu n,ξ, ∂ ξβu m,ξ
−n≤Z m>Z Γ∗
dξ
ω
−ωmn(ξ )
u n,ξ, ∂ ξαu m,ξu n,ξ, ∂ ξβu m,ξ
− 2i
ωn≤Z
m>Z
Γ∗
u n,ξ, ∂ ξαu m,ξu n,ξ, ∂ ξβu m,ξ dξ
−g e ,ω,α, δ ρe V e (
I −χe ,ωδ ρe V e )
−1g e ,ω,β ,where we have used the shorthand ωmn(ξ ) = E m(ξ ) − E n(ξ ).Recall that δ ρe V e is the linearized effective potential operator atequilibrium.
Dynamic permittivity tensor (cont’d)
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 63/93
The dynamic polarizability operator χe ,ω and the vector valued
functions g e ,ω are given by
χe ,ωV = −n≤Z
m>Z
Γ∗
dξ
ω + ωmn(ξ )u n,ξu ∗m,ξu n,ξ,Vu m,ξ
+ n≤Z
m>Z
Γ∗ dξ
ω − ωmn(ξ ) u ∗n,ξu m,ξu m,ξ,Vu n,ξ,
g e ,ω = −n≤Z
m>Z
Γ∗
dξ
ω + ωmn(ξ )u n,ξu ∗m,ξu n,ξ, i ∇ξu m,ξ
+ n≤Z
m>Z
Γ∗
dξ ω − ωmn(ξ ) u ∗n,ξu m,ξi ∇ξu m,ξ, u n,ξ.
F d l
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 64/93
Fundamentals
Macroscopic limit of density functional theoryDerivation of nonlinear elasticity and macroscopic electrostaticequation from Kohn-Sham DFTDerivation of macroscopic Maxwell equation from TDDFT
AlgorithmsIntroductionDiscretization of the Kohn-Sham Hamiltonian
Representation of the Fermi OperatorDensity evaluation
F d l
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 65/93
Fundamentals
Macroscopic limit of density functional theoryDerivation of nonlinear elasticity and macroscopic electrostaticequation from Kohn-Sham DFTDerivation of macroscopic Maxwell equation from TDDFT
AlgorithmsIntroductionDiscretization of the Kohn-Sham Hamiltonian
Representation of the Fermi OperatorDensity evaluation
Self consistent iteration and direct minimization
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 66/93
Self consistent iterationLook for a fixed point of the Kohn-Sham map
ρ = F KS(ρ).
Direct minimization
Minimize the Kohn-Sham energy functional directly
E KS(ψi ) = 1
2
N
i =1
|∇ψi (x )|2
dx +
V ext(x )ρ(x ) dx + E xc[ρ]
+
γ
N i =1
|b , ψi |2 + 1
2
ρ(x )ρ(y )
|x − y | dx dy .
Flowchart in a general scope
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 67/93
Kohn-Sham map at finite temperature
ρ = F KS (ρ) = diag f (H [ρ] − µ) = diag 2
1 + e β (H [ρ]−µ).
Mixing scheme for self consistent iteration
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 68/93
Mixing schemes can be understood in the context of solving the
nonlinear equation ρ = F KS(ρ): Simple mixing (linear mixing);
Pulay mixing [Pulay 1980];
Quasi-Newton method: Anderson mixing [Anderson 1965],
Broyden mixing [Johnson 1988] and multisecant versions; Newton method [Gao, Yang and Meza preprint];
Preconditioners: ...
Our focus in the rest of the talk will be the evaluation of Kohn-Sham map: To obtain F KS(ρ).
General behavior of metallic and insulating system
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 69/93
Insulator Metal
Insulator Metal
Band gap finite 0Density matrix
decay
exponential algebraic
General scheme for both metallic and insulating systems;
Special O(N ) techniques for insulating systems.
Difference between metal and insulator
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 70/93
Differences can be understood from several perspective: Localization (Linear scaling algorithms);
Behavior of linear response (Self-consistent iteration);
Charge screening (Multiscale methods).
Representative examples:
Metal: Jellium model, H = −∆;
Insulator: Array of deep narrow potentials.Analytical solvable example?
Discretization
Conventional basis sets:
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 71/93
Fourier space (Plane-wave);
Real space (Finite difference, finite element); Wavelet basis (multiscale).
Atomic orbital as basis functions:
Tight binding (parametrized atomic orbital);
Gaussian type orbitals (GTO); Numerical atomic orbital (NAO) [Blum et al 2009].
Mixed basis functions with atomic orbital:
Augmented plane-wave (APW) [Slater 1937];
Linear augmented-plane-wave (LAPW) [Andersen 1975]; Projector augmented-wave (PAW) [Blochl 1994];
Enriched finite element [Sukumar and Pask 2009].
Discontinuous basis functions
Representation
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 72/93
Spectral decomposition:
(diag f (H [ρ]))(x ) =n
f (E n)|Ψn(x )|2.
Fermi Operator expansion:
diag f (H [ρ]) ≈P i =1
diag f i (H [ρ]),
where f i (H [ρ]) are simple (polynomials, rational functions)that the matrix function can be evaluated directly.
Evaluation
Diagonalization:
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 73/93
g
Jacobi-Davidson Diagonalization.
Chebyshev filtering [Zhou, Saad, Tiago et al 2006]
Fermi operator expansion:
Polynomial expansion [Goedecker and Colombo 1994]
Rational expansion [Baroni and Giannozzi 1992] [Lin, Lu, Ying
et al 2009]
Special techniques for insulators
Divide and Conquer [Yang 1991]
Orbital minimization [Mauri, Galli and Car 1993]
Density matrix minimization [Li, Nunes and Vanderbilt 1993]
Localized subspace iteration [Garcıa-Cervera, Lu and E 2007]
Orbital minimization with localization [E and Gao 2010]
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 74/93
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 75/93
Existing discretization methods
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 76/93
Uniform grid (Fourier, real space)
Large number of basis functions per atom (500 ∼ 5000).
Atomic orbitals and mixed basis functions:
Fine tunning of parameters.
Different parameters for different exchange-correlationfunctional.
Overcomplete and incomplete basis sets.
Interstitial region.
Basis function with large support: metallic system.
Discontinuous Galerkin framework with locally adaptivebasis
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 77/93
Goal: Construct local basis functions on the fly by solving a small
part of the system.
Local solve in a buffer to obtain basis functions adapted tothe environment;
Discontinuous Galerkin framework to discretize the systemusing these (discontinuous) basis functions.
See Lin Lin’s talk (Thursday) for more details
DG Formulation
Minimize the DG discrete effective energy functional to get the
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 78/93
Minimize the DG discrete effective energy functional to get thedensity (for given effective potential):
E DG(ψi ) = 1
2
N i =1
∇ψi ,∇ψi T +V eff , ρT +
γ
N i =1
|b , ψi T |2
−N i =1
∇ψi , ψi S + αh
N i =1
ψi , ψi S .
u = u 1n1 + u 2n2 on S .q
= 12 (q 1 + q 2) on S .
Constructing adaptive local basis function
Buffer region associated with Ek : Qk ⊃ Ek .
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 79/93
Buffer region associated with E k : Q k ⊃ E k .
Restrict the effective Hamiltonian on Q k by assuming the
periodic boundary condition on ∂ Q k and obtain H eff ,Q k .
Take the first several eigenfunctions of H eff ,Q k called ϕk , j , j = 1, · · · , J k and restrict them on E k .
Red: E k ; Red+Blue: Q k
Numerical results: Accuracy
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 80/93
(a) Quasi-1D Na(1 × 1 × 4 Cells)
3 50
0.5
1
1.5
2x 10
−3
DOF/Atom
E r r o r / A
t o m ( a
u )
(b) Quasi-2D Na
(1 × 4 × 4 Cells)
5 10 150
1
2
3
x 10−3
DOF/Atom
E r r o r / A
t o m ( a
u )
(c) Bulk 3D Na(4 × 4 × 4 Cells)
15 20 25 30 350
2
4
6x 10
−3
DOF/Atom
E r r o r / A
t o m ( a
u )
Buffer length per dimension: 2 unit cells (Red); 3 unit cells (Blue)
Numerical results: Efficiency
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 81/93
128 432 1024 2000 3456
100
101
102
103
104
x2
Atom #
W a l l C l o c k T i m e ( s e c )
Buffer SolveDG Eigensolver
DG OverheadComputational time perprocessor comparison:
Atom# Proc# Global DG
time time128 64 35 s 4 s432 216 248 s 35 s
Fundamentals
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 82/93
Macroscopic limit of density functional theoryDerivation of nonlinear elasticity and macroscopic electrostaticequation from Kohn-Sham DFTDerivation of macroscopic Maxwell equation from TDDFT
AlgorithmsIntroductionDiscretization of the Kohn-Sham HamiltonianRepresentation of the Fermi OperatorDensity evaluation
Fermi Operator Expansion
2
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 83/93
2
1 + e β (H
−µI )
≈ P
i =1
c i
H −µI
∆E
i
+Q i =1
ωi
(z i I − (H − µI )/∆E )q i
β inverse temperature; ∆E width of the spectrum of thediscretized Hamiltonian;
Want P ,Q , q i be as small as possible given β and ∆E .
Past work
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 84/93
P : Number of polynomials. Q : Number of rational functions. q i :
the order of each rational function. P ∼ O(β ∆E ),Q = 0 [Goedecker and Colombo 1994];
P = 0, Q ∼ O(β ∆E ), q i ≡ 1 [Baroni and Giannozzi 1992];
P = 0, Q ∼ O(β ∆E )1/2, q i ≡ 1 [Ozaki 2007];
P ∼ C , Q ∼ O(β ∆E )1/2, q i ≡ 1[Ceriotti, Kuhne and Parrinello 2008];
Multipole expansion: P ∼ C , Q ∼ O(log(β ∆E )), q i ∼ C [Lin, Lu, Car and E 2009];
Pole expansion: P = 0, Q ∼ O(log(β ∆E )), q i ≡ 1[Lin, Lu, Ying and E 2009].
Pole expansion
2 1 f ( ) 1P
f ( )
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 85/93
f (x ) = 2
1 + e β ∆Ex
= 1
2πi Γf (z )
z − x dz
≈ 1
2πi
P
i =1
f (z i )w i
z i − x .
Optimal choice of the contour Γ, integration points z i ∈ C andintegration weights w i ∈ C → Number of discretization points
∼ O(log(β ∆E )).
non-analytic
spectrum
1-1
Geometric convergence with small pre-constant
2D discretized Laplacian with small perturbation: energy gap6
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 86/93
around 10−6au.
10 20 30 40 50 60 70
10−8
10−6
10−4
10−2
N Pole
L 1
e r r o r
p e r
e l e c t r o n
104
105
106
60
65
70
75
80
85
90
β ∆E
N P o l e
Fundamentals
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 87/93
Macroscopic limit of density functional theoryDerivation of nonlinear elasticity and macroscopic electrostaticequation from Kohn-Sham DFTDerivation of macroscopic Maxwell equation from TDDFT
AlgorithmsIntroductionDiscretization of the Kohn-Sham HamiltonianRepresentation of the Fermi OperatorDensity evaluation
Selected inversion
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 88/93
Evaluate density based on the Fermi operator expansion:
ρ ≈ diag
P i =1
ωi
z i I − H .
Naive approach: Invert z i I − H first and then take thediagonal: cubic scaling.
Fast diagonal extraction? Use sparse matrix algebra.Selected inversion [Lin, Lu, Ying et al 2009]
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 89/93
SelInv: Selected Inversion for Sparse Symmetric Matrix
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 90/93
SelInv is a selected inversion algorithm for general sparsesymmetric matrix written in Fortran[Lin, Yang, Meza, et al, in press]
Symbolic Analysis: matrix reordering LDLT factorization
Selected inversion
Numerical results: SelInv
Problems from Harwell-Boeing Test Collection and the Universityof Florida Matrix Collection.
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 91/93
of Florida Matrix Collection.
problem n selected inversion direct inversion speeduptime time
bcsstk14 1,806 0.01 sec 0.13 sec 13bcsstk24 3,562 0.02 sec 0.58 sec 29bcsstk28 4,410 0.02 sec 0.88 sec 44bcsstk18 11,948 0.24 sec 5.73 sec 24
bodyy6 19,366 0.09 sec 5.37 sec 60crystm03 24,696 0.78 sec 26.89 sec 34
wathen120 36,441 0.34 sec 48.34 sec 142thermal1 82,654 0.44 sec 95.06 sec 216shipsec1 140,874 17.66 sec 3346 sec 192
pwtk 217,918 14.55 sec 5135 sec 353
parabolic fem 525,825 20.06 sec 7054 sec 352tmt sym 726,713 13.98 sec > 3 hours > 772ecology2 999,999 16.04 sec > 3 hours > 673
G3 circuit 1,585,478 218.7 sec > 3 hours > 49
Parallel implementation5-pt discretization of 2D Laplacian operator:
103 PSelInv
104 PSelInv
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 92/93
100
101
102
103
102
10
number of processors
w a l l c l o c k t i m e ( s e c )
ideal
100
101
102
103
101
102
103
10
number of processors
G f l o p s
ideal
Figure: Log-log plot of total wall clock time and total Gflops with respectto number of processors, compared with ideal scaling. The largest matrixsolved has (4.3 billion)2 degrees of freedom.[Lin, Yang, Lu, et al, preprint]
Conclusion and Outlook
Summary
8/10/2019 Kohns Ham
http://slidepdf.com/reader/full/kohns-ham 93/93
Understanding the macroscopic limit of the Kohn-Sham
density functional theory, based on sharp stability conditions: Nonlinear elasticity; Macroscopic Maxwell equations;
Algorithmic development for metallic systems Discontinuous Galerkin framework with locally adapted basis; Fermi operator expansion with optimal scaling; Selected inversion for sparse discrete Hamiltonian matrix;
(not mentioned) Linear and sublinear scaling algorithms fordensity functional theory.
Density functional theory is a challenge and opportunity for appliedmathematicians. Many open questions remain on both analyticaland numerical sides.