Vidvuds Ozolins Department of Materials Science & Engineering
University of California, Los Angeles
UCLA Laboratory for the Quantum Prediction of Advanced Materials
Supported by:
Sparse physics via compressive sensing
DMR-1106024
Gus Hart (BYU Provo)
Lance Nelson (BYU Idaho)
Fei Zhou (LLNL) Weston Nielson
(UCLA)
Rogjie Lai (UC Irvine)
Russ Caflisch (UCLA)
Stan Osher (UCLA)
Co-conspirators
Outline
Premise: • Most physics models are approximately sparse
(i.e., have a few terms) in some basis
Questions: • How do we pick the basis? • How do we determine which terms to pick? • How should we generate fitting data? • How do we systematically improve accuracy?
High-rate charge storage in Nb2O5
V. Augustyn, J. Come, M. A. Lowe, J. W. Kim, P.-L. Taberna, S. H. Tolbert, H. D. Abruña, P. Simon, B. Dunn, Nature Materials 12, 518–522 (2013)
Structure of orthorhombic Nb2O5
Acta Cryst. B31, 673 (1975)
Layered structure (stacked along c) 001 layer with octahedra and pentagonal bipyramids
Atomic disorder in Nb2O5
• Nb (8i): 16 Nb off-center in the c direction
Nb layer
• Nb (4g) partial occupancy: 0.8 Nb between Nb layers
• O: distorted close packing within {001} layers
Lithium energy landscape in the (001) plane
Oxygen
C.-P. Liu, F. Zhou, and V. Ozolins (2013)
Li intralayer diffusion map
50 meV
> 700 meV
C.-P. Liu, F. Zhou, and V. Ozolins (2013)
Configurational problem for Nb2O5
Occupation Si: Vacancy, Li, or Nb
C.-P. Liu, F. Zhou, and V. Ozolins (2013)
Configurational Ordering in Alloys
Cu3Au CuAu CuAu3
All these structures are based on the face-centered cubic (FCC) lattice.
Cluster Expansion
Rewrite as an expansion in clusters of lattice sites:
= + +
+ + ...
+
ADVANTAGES: ü Very fast – can be used in Monte Carlo simulations ü Works for any structure based on a given lattice
E = E0 + J1c + J f Sii∈f∏
f
Pairs, Triplets, ...
∑
Cluster selection is difficult
0.5 1.0 1.5 2.0 2.5 3.01.
10.
100
1000
10 000
cluster radius
pairs
triplets
4-bodies5-bodies
Compressive sensing CE
=
?0
BBBBBBBBBB@
JnnJnnnJtrip.....
1
CCCCCCCCCCA
. . . .0
BB@
�1 1 . . . . . .0 �1 . . . . . .. . . . . . . .. . . . . . . .
1
CCA
0
BBBBBB@
E( )E( )E( )
.
.
.
1
CCCCCCA
JCS = argminJ
J 1 : ΠJ − E 2 ≤ ε{ }
Clusters
Stru
ctur
es Si
i∈f∏
L. J. Nelson, F. Zhou, G.L.W. Hart, and V. Ozolins, Phys. Rev. B 87, 035125 (2013)
Why compressive sensing works
x* = argmin x: Ax=b x 1
x 1 = xii=1
n
∑
x1
x2
a11x1 + a12x2 = b1
Why Euclidean distance is worse
x* = argmin x: Ax=b x 2
x1
x2
a11x1 + a12x2 = b1
x 2 = xi2
i=1
n
∑
Theorem
E. Candès, J. Romberg, and T. Tao, Comm. Pure Appl. Math., vol. 59, no. 8, pp. 1207–1223 (2006).
M ≥C ⋅µ2 ⋅S ⋅ log N δ( )Coherence
1≤µ≤√N Sparsity Size of basis set
P >1−δ
Probability to find the correct S-sparse solution is
if the number of data points satisfies
Ag-Pt cluster expansion
Pairs Triplets Quadruplets
Cluster radius
L. J. Nelson, F. Zhou, G.L.W. Hart, and V. Ozolins, Phys. Rev. B 87, 035125 (2013).
Ag-Pt cluster expansion
Pairs Triplets Quadruplets
Cluster radius
L. J. Nelson, F. Zhou, G.L.W. Hart, and V. Ozolins, Phys. Rev. B 87, 035125 (2013).
Performance of CSCE
Training set size
RMSError(meV)(solidline)
kJfitk
1(D
ashed
Line)
Compressive sensing
Discrete Opt.
L. J. Nelson, F. Zhou, G.L.W. Hart, and V. Ozolins, Phys. Rev. B 87, 035125 (2013).
Next step – Bayesian CSCE
L. J. Nelson, G. L. W. Hart, S. Reese, F. Zhou, and V. Ozolins, Physical Review B 88, 155105 (2013).
CSCE results for zinc-finger protein
-30 -20 -10 0 10
-30
-20
-10
0
10
Direct energy HkcalêmolL
PredictedenergyHkca
lêmolL
L. J. Nelson, F. Zhou, G.L.W. Hart, and V. Ozolins, Phys. Rev. B 87, 035125 (2013).
Direct conversion of heat to electricity
Up to 30% conversion efficiency with right materials
Thermopower S = ΔV/ΔT
hot cold
ZT =σ ⋅ S2
κ total
•T
σ ⋅ S 2Power factor
Total thermal conductivity
electrical conductivity thermopower
ZT =σ ⋅ S2
κ total
•T
σ ⋅ S 2Power factor
Total thermal conductivity
electrical conductivity thermopower
PbTe: Standard of excellence
From Wei & Zunger, Physical Review B 55, 13605–13610 (1997)
(direct gaps at L point) result from the occurrence of the Pb s band below the top of the valence band, setting up coupling and level repulsion at the L point.
Image from Pei et al. Nature 473, 66–69 (2011)
Giant anharmonicity of PbTe TO mode
Chemical compound space
PbTe IV-VI
Group I + V Group VI
I-V-VI2
AgSbTe2, NaSbSe2, …
M. Nielsen, V. Ozolins, and J. P. Heremans, Energy & Environ. Sci. 6, 570-578 (2013).
New materials: Search space
ABX2 in cubic D4 (AF-IIb) for A=Cu, Ag, Au ABX2 in rhombohedral R-3m (AF-II) for A=Na, K, Rb, Cs, Tl Screened a total of 8 × 3 × 3 = 72 compounds
M. Nielsen, V. Ozolins, and J. P. Heremans, Energy & Environ. Sci. 6, 570-578 (2013).
4
3
2
1
0
Grun
eise
n pa
ram
eter
a
AuAs
Te2
CuBi
Se2
AuBi
Te2
CuAs
Te2
AgAs
Te2
AgSb
Se2
CuBi
Te2
AuSb
Te2
NaSb
S2KA
sSe2
AgBi
S2Cu
SbTe
2Ag
BiSe
2Ag
BiTe
2Ag
SbTe
2Na
AsSe
2Cs
SbSe
2Rb
SbSe
2Na
SbSe
2KS
bSe2
CsBi
S2Rb
BiS2
NaAs
Te2
NaBi
S2KB
iS2
NaSb
Te2
KAsT
e2Na
BiSe
2Rb
AsTe
2Cs
AsTe
2Rb
BiSe
2Cs
BiSe
2Na
BiTe
2KB
iSe2
RbSb
Te2
KSbT
e2Cs
BiTe
2Rb
BiTe
2KB
iTe2
CsSb
Te2
Calculated Gruneisen parameters
M. Nielsen, V. Ozolins, and J. P. Heremans, Energy & Environ. Sci. 6, 570-578 (2013).
Experimental results from OSU
M. Nielsen, V. Ozolins, and J. P. Heremans, Energy & Environ. Sci. 6, 570-578 (2013).
Cu12Sb4S13: Thermoelectric mineral with ZT=1
Cu
S
Sb
X. Lu, D. T. Morelli, Y. Xia, F. Zhou, V. Ozolins, H. Chi, and C. Uher. Advanced Energy Materials (2013)
Natural tetrahedrite mineral
!!!!!!!!!!!!!!!!!!!!!!!!!! !
!!!1 cm
Figure of Merit
0
0.2
0.4
0.6
0.8
1
1.2
300 400 500 600 700 800
Temperature (K)
Figu
re o
f Mer
it zT
a)
!
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
Brillouin Zone Occupation Fraction
Figu
re o
f Mer
it zT
b)
f =1 for Zn (2+) f =2 for Fe (3+)
Lu, Morelli, Xia, Zhou, Ozolins, Chi, Zhou, and Uher, Advanced Energy Materials (2013).
Frustrated Cu(1) bonding environment
Cu(1) (3-fold)
Lu, Morelli, Xia, Zhou, Ozolins, Chi, Zhou, and Uher, Advanced Energy Materials (2013).
Anharmonic optical mode
-0.4 -0.2 0.0 0.2 0.4
-10
0
10
20
30
40
50
uCu HfiL
DEHme
VL
Lu, Morelli, Xia, Zhou, Ozolins, Chi, Zhou, and Uher, Advanced Energy Materials (2013).
CS lattice dynamics
Taylor expansion of the total energy in terms of the atomic displacements ua=Ra-R0
a:
� = �0 + �aua +1
2�abuaub +
1
6�abcuaubuc + · · ·
Fa = �@�/@a = �(�a + �abub +1
2�abcubuc + · · · )
Forces:
Use for: Thermal transport, free energies, phase transformations, etc.
Expansion in cluster series
Φaua, Φaaua2, Φaaaua
3, …
Φabuaub, Φaabua2ub, …
Φabcuaubuc , Φaabcua2ubuc , …
Φabcduaubucud, …
…
=
…
How to calculate FCT’s from DFT?
Calculate forces Fa Displace atoms ua
Fa = �@�/@a = �(�a + �abub +1
2�abcubuc + · · · )Fa = �@�/@a = �(�a + �abub +
1
2�abcubuc + · · · )
Compressive sensing
ΦCS = argminΦ
µ Φ 1 +12
ϒΦ + F 22
1 ub1 ub
1uc1
1 ub2 ub
2uc2
⎛
⎝
⎜⎜⎜⎜
⎞
⎠
⎟⎟⎟⎟
Φa
Φab
Φabc
⎛
⎝
⎜⎜⎜⎜
⎞
⎠
⎟⎟⎟⎟=
−Fa1
−Fa2
⎛
⎝
⎜⎜⎜⎜
⎞
⎠
⎟⎟⎟⎟
We need to solve an underdetermined linear system of equations:
• Commutativity of derivatives
• Translational invariance (Noether’s theorem)
Symmetry constraints
Dynamical properties of solids, eds. G.K. Horton & A.A. Maradudin (1974)
�abc = �acb = . . .
X
a
�I({a, b, c · · · }) = 0
�(aa) = �X
b 6=a
�(ab) Pair ASR
Multibodies
• Space group symmetry
• Transformation matrix:
• Reduces ## of FCT’s by a factor of 10 in cibic crystals
Symmetry constraints
Dynamical properties of solids, eds. G.K. Horton & A.A. Maradudin (1974)
F2 F3
u1
�I(s↵) = �IJ(s)�J(↵)
�IJ(s) = �i1j⇡(1)
· · · �inj⇡(n)
γ =cosα −sinα 0sinα cosα 00 0 1
⎛
⎝
⎜⎜⎜
⎞
⎠
⎟⎟⎟
The final CSLD L1 problem
ΦCS = argminΦS
µ Φ 1 +12
ϒΦ + F 22⎛
⎝⎜⎞⎠⎟
DFT data constraints are imposed via a regular least-squares norm:
where Φ = BTTrans BS
Rot BP
PermΦS.
are symmetry-distinct FCT components. ΦS
100 200 300 400 500 600
-0.10
-0.05
0.05
0.10
0.15
F
• Solution is sparse • Force prediction error = 3%
2 3 4 5 6
DFT tests: NaCl
Independent FC’s by order, then by distance
-15 -10 -5 0 5 10 15 20-15
-10
-5
0
5
10
15
20
FDFT @eVêfiD
F CSLD@eVêfiD
Δu/r0=25%
Applications
Performance of CSLD
Applications
Sparsity for PDEs
Classically, sparsity is limited to discrete objects. [Questions] • What is an analog of the “sparsity” for continuous
functions? • How to design a tractable method to create
“sparse” continuous functions?
L1 regularization of DFT
E = minψ i{ }i=1
Nψ j H ψ j
j=1
N
∑ s.t. ψ i ψ j = δ ij
Conventional Kohn-Sham problem:
This method converges to spatially localized wave f-ns with compact support. O(N) becomes possible.
′E = minψ i{ }i=1
Nψ j H ψ j
j=1
N
∑ + 1µ
ψ j 1j=1
N
∑⎛⎝⎜
⎞⎠⎟
s.t. ψ i ψ j = δ ij,
“Sparse” = localized or short-ranged
ψ 1 ≡ ψ (x) dx, x ∈Rd∫
V. Ozolins, R. Lai, R. E. Caflisch, and S. Osher, “Compressed modes … ,“ PNAS (2013)
Results
Results
V. Ozolins, R. Lai, R. E. Caflisch, and S. Osher, “Compressed modes … ,“ PNAS (2013)
Dependence on µ
0 5 10 15 20 25 30 35 40 45 50
0
0.2
0.4
0.6
!1
!2
!3
!4
!5
µ = 300 5 10 15 20 25 30 35 40 45 50
0
0.2
0.4
0.6
!1
!2
!3
!4
!5
µ = 50
0 5 10 15 20 25 30 35 40 45 50
0
0.2
0.4
0.6
!1
!2
!3
!4
!5
µ = 500 5 10 15 20 25 30 35 40 45 50
0
0.2
0.4
0.6
!1
!2
!3
!4
!5
µ = 300
0 5 10 15 20 25 30 35 40 45 50
0
0.2
0.4
0.6
!1
!2
!3
!4
!5
µ = 5000 5 10 15 20 25 30 35 40 45 50
0
0.2
0.4
0.6
!1
!2
!3
!4
!5
µ = 5000
Fig. 4. Computation results of CMs with di↵erent values of µ. The first column: the first 5 CMs of the 1D free-electron model. The second column: the first 5 CMs of the1D Kronig-Penny model.
electrons in a one-dimensional crystal, where the potentialfunction V (x) consists of a periodic array of rectangular po-tential wells. For simplicity, in our experiments we replace therectangular wells with inverted gaussians so that the poten-
tial is given by V (x) = �V0P
Nelj=1 exp
h� (x�xj)
2
2�2
i. We choose
Nel = 5, V0 = 1, � = 3 and xj
= 10j in our discussion be-low and, in spite of the di↵erent potential, continue to referto this case as the 1D KP model. This model exhibits twolow-energy bands separated by finite gaps from the rest of the(continuous) eigenvalue spectrum, and the Wannier functionscorresponding to these bands are exponentially localized.
In our experiments, we choose ⌦ = [0, 50] and discretize⌦ with 128 equally spaced nodes. The proposed variationalmodel Eq. [4] is solved using algorithm 1, where parametersare chosen as � = µN/20 and r = µN/5. We report the com-putational results of the first 5 CMs of the 1D free-electronmodel (the first column) and the 1D KP model (the secondcolumn) in Figure 4, where we use 5 di↵erent colors to dif-ferentiate these CMs. To compare all results more clearly, weuse the same initial input for di↵erent values of µ in the free-electron model and the 1D KP model. We flip the CMs ifnecessary such that most values of CMs on their support arepositive, since sign ambiguities do not a↵ect minimal valuesof the objective function in Eq. [4]. For comparison, Figure 3plots the first 5 eigenfunctions of the Schrodinger operatorused in the free-electron model and KP model. It is clearthat all these eigenfunctions are spatially extended withoutany compact support. However, as we can observe from Fig-ure 4, the proposed variational model does provide a series ofcompactly supported functions. Furthermore, all numericalresults in Figure 4 clearly show the dependence of the size of
compact support on µ, as suggested by general considerationsbased on the variational formula Eq. [4]. In other words, themodel with smaller µ will create CMs with smaller compactsupport, and the model with larger µ will create CMs withlarger compact support. In addition, we find that the result-ing compressed modes are not interacting for small µ (the firstrow of Figure 4). By increasing µ to a moderate value, themodes start to interact each other via a small amount of over-lap (the second row of Figure 4). Significant overlap can beobserved using a big value of µ (the third row of Figure 4).
0 5 10 15 20 25 30 35 40 45 50
!0.2
!0.1
0
0.1
0.2
0.3
!1
!2
!3
!4
!5
0 5 10 15 20 25 30 35 40 45 50
!0.2
!0.1
0
0.1
0.2
0.3
!1
!2
!3
!4
!5
Fig. 3. The first 5 egienfunctions of the Schrodinger operator H in the free-electron model (top) and the KP model (bottom).
4 www.pnas.org/ Footline Author
V. Ozolins, R. Lai, R. E. Caflisch, and S. Osher, “Compressed modes … ,“ PNAS (2013)
Total energy convergence
µ = 10,M = 50, N = 50 µ = 10,M = 50, N = 60 µ = 10,M = 50, N = 128
0 10 20 30 40 50!0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Eigs of !T H !
Eigs of H
0 10 20 30 40 50!0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Eigs of !T H !
Eigs of H
0 10 20 30 40 50!0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Eigs of !T H !
Eigs of H
0 10 20 30 40 50!1
!0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
Eigs of !T H !
Eigs of H
0 10 20 30 40 50!1
!0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
Eigs of !T H !
Eigs of H
0 10 20 30 40 50!1
!0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
Eigs of !T H !
Eigs of H
Fig. 5. Comparisons of the first 50 eigenvalues of the 1D free electron model (the first row) and the 1D Kronig-Penney model (the second row).
10 50 100 200 300 400 5000
0.02
0.04
0.06
0.08
0.1
µ
Re
lative
Err
or
M = N = 50
Potential free
Kronig!Penney
50 60 70 80 90 100 1280
0.02
0.04
0.06
0.08
0.1
N
Rela
tive E
rror
M = 50, µ = 10
Potential free
Kronig!Penney
Fig. 6. Relative eigenvalue error of the 1D free-electron model (red dots) and 1DKP model (blue circles). Top: relation of the relative error via di↵erent values of µfor fixed M = N = 50. Bottom: relation of the relative error via di↵erent valuesof N for fixed µ = 10 and M = 50.
We further test Conjecture 1 numerically, i.e., uni-tary transformation of the derived compactly supported
compressed modes can represented the eigenmodes of theSchrodinger operator. We compare the first M eigenvalues(�1, · · · ,�M
) of the matrix h T
N
H N
i obtained by the 1DKP model and 1D free-electron model with the first M eigen-values (�1, · · · ,�M
) of the corresponding Schrodinger oper-ators. Figure 5 illustrates the comparisons with a relativesmall value µ = 10, when the CMs are highly localized. Wecan clearly see that {�
i
} gradually converges to {�i
} withincreasing number N of CMs. In addition, we also plot therelative error E =
PM
i=1(�i
��i
)2/P
M
i=1(�i
)2 in Figure 4. Aswe speculated in conjecture 1, the relative error will convergeto zero as µ ! 1 for fixed M = N = 50, which is illustratedin the top panel of Figure 4. The relative error will also con-verge to zero as N ! 1 for fixed µ = 10 and M = 50, whichis illustrated in the bottom panel of Figure 4.
Moreover, the proposed model and numerical algorithmalso work on domains in high dimensional space. As an ex-ample, Figure. 7 shows computational results of the first 25CMs of the free-electron case on a 2D domain [0, 10]2 withµ = 30. All the above discussions of 1D model are also truefor 2D cases. In addition, our approach can also be naturallyextended to irregular domains, manifolds as well as graphs,which will be investigated in our future work.
In conclusion, the above numerical experiments validatethe conjecture that the proposed CMs provide a series of com-pactly supported orthonormal functions, which approximatelyspan the low-energy eigenspace of the Schrodinger operator(i.e., the space of linear combinations of its first few lowesteigenmodes).
Footline Author PNAS Issue Date Volume Issue Number 5
V. Ozolins, R. Lai, R. E. Caflisch, and S. Osher, “Compressed modes … ,“ PNAS (2013)
Higher dimensions, etc.
Discussions and ConclusionsIn conclusion, we have presented a method for producing com-pressed modes (i.e., modes that are sparse and spatially local-ized with a compact support) for the Laplace operator plus apotential V , using a variational principle with an L1 penal-ization term that promotes sparsity. The trade-o↵ betweenthe degree of localization and the accuracy of the variationalenergy is controlled by one numerical parameter, µ, withoutthe need for physical intuition-informed spatial cuto↵s. TheSOC algorithm of Ref. [13] has been used to numerically con-struct these modes. Our tests indicate that the CMs can beused as an e�cient, systematically improvable orthonormalbasis to represent the low-energy eigenfunctions, energy spec-trum of the Schrodinger operator. Due to the fact the CMsare compactly supported, the computational e↵ort of total en-ergy calculations increases linearly with the number of modesN , overcoming the O(N3) orthogonalization bottleneck lim-iting the performance of methods that work by finding theeigenfunctions of the Schrdinger operator.
In addition, note that the discretized variational principlein Eq. [13] is related to sparse principal component analysis(SPCA) (Ref. [18, 19]). SPCA, however, does not involvean underlying continuum variational principle and the sparseprincipal components are not localized, since the componentnumber does not correspond to a continuum variable.
These results are only the beginning. We expect that CM-related techniques will be useful in a variety of applicationsin solid state physics, chemistry, materials science, and otherfields. Future studies could explore the following directions:
1. Use CMs to develop spatially localized basis sets that spanthe eigenspace of a di↵erential operator, for instance, theLaplace operator, generalizing the concept of plane wavesto an orthogonal real-space basis with multi-resolution ca-pabilities. More details will be discussed in Ref. [20].
2. Use the CMs to construct an accelerated (i.e., O(N)) sim-ulation method for density-functional theory (DFT) elec-tronic structure calculations.
3. Construct CMs for a variety of potentials and develop CMsas the modes for a Galerkin method for PDEs, such asMaxwell’s equations.
4. Generalize CMs for use in PDEs (such as heat type equa-tions) that come from the gradient descent of a variationalprinciple.
5. Extend CMs to higher dimensions and di↵erent geometries,including the Laplace-Beltrami equation on a manifold anda discrete Laplacian on a network.
Finally, we plan to perform an investigation of the formalproperties of CMs to rigorously analyze their existence andcompleteness, including the conjecture that was hypothesizedand numerically tested here.
Fig. 7. The first 25 CMs of freel-electron case on a 2D domain [0, 10]2 withµ = 30. Each CM is color-coded by its height function.
ACKNOWLEDGMENTS. V.O. gratefully acknowledges financial support from theNational Science Foundation under Award Number DMR-1106024 and use of com-puting resources at the National Energy Research Scientific Computing Center, whichis supported by the US DOE under Contract No. DE-AC02-05CH11231. The re-search of R.C. is partially supported by the US DOE under Contract No. DE-FG02-05ER25710. The research of S.O. was supported by the O�ce of Naval Research(Grant N00014-11-1-719).
1. D.L. Donoho. Compressed sensing. Information Theory, IEEE Transactions on,52(4):1289–1306, 2006.
2. E. J. Candes, J. Romberg, and T. Tao. Robust uncertainty principles: Exact signal re-construction from highly incomplete frequency information. Information Theory, IEEETransactions on, 52(2):489–509, 2006.
3. B. Recht, M. Fazel, and P. Parrilo. Guaranteed minimum-rank solutions of linearmatrix equations via nuclear norm minimization. SIAM Review, 52(3):471–501, 2010.
4. E. J. Candes, Y.C. Eldar, T. Strohmer, and V. Voroninski. Phase retrieval via matrixcompletion. SIAM Journal on Imaging Sciences, 6(1):199–225, 2013.
5. E. J. Candes, X. Li, Y. Ma, and J. Wright. Robust principal component analysis?Journal of the ACM (JACM), 58(3):11, 2011.
6. H. Schae↵er, R. Caflisch, C. D. Hauck, and S. Osher. Sparse dynamics for partial dif-ferential equations. Proceedings of the National Academy of Sciences, 110(17):6634–6639, 2013.
7. L.J. Nelson, G. Hart, F. Zhou, and V. Ozolins. Compressive sensing as a paradigm forbuilding physics models. Physical Review B, 87(3):035125, 2013.
8. E. Prodan and W. Kohn. Nearsightedness of Electronic Matter. Proceedings of theNational Academy of Sciences, 102(33):1635–11638, 1996.
9. G. H. Wannier. The structure of electronic excitation levels in insulating crystals.Physical Review, 52(3):0191–0197, August 1937.
10. N. Marzari and D. Vanderbilt. Maximally localized generalized Wannier functions forcomposite energy bands. Physical Review B, 56(20):12847–12865, 1997.
11. W. E, T. Li, and J. Lu. Localized bases of eigensubspaces and operator compression.Proceedings of the National Academy of Sciences, 107(1273–1278), 2010.
12. D. L. Donoho and M. Elad. Optimally sparse representation in general (nonorthogo-nal) dictionaries via 1 minimization. Proceedings of the National Academy of Sciences,100(5):2197–2202, 2003.
13. R. Lai and S. Osher. A splitting method for orthogonality constrained problems. Jouralof Scientific Computing, to appear, 2013.
14. S. Osher, M. Burger, D. Goldfarb, J. Xu, and W. Yin. An iterative regularizatoinmethod for total variation-based image restoration. Multiscale Model. Simul., 4:460–489, 2005.
15. W. Yin, S. Osher, D. Goldfarb, and J. Darbon. Bregman iterative algorithms forl1-minimization with applications to compressed sensing. SIAM Journal on ImagingSciences, 1(1):143–168, 2008.
16. T. Goldstein and S. Osher. The split bregman method for `1-regularized problems.SIAM Journal on Imaging Sciences, 2(2):323–343, 2009.
17. R. de L. Kronig and W. G. Penney. Quantum mechanics of electrons in crystal lattices.Proceedings of the Royal Society of London. Series A, 130(814):499–513, 1931.
18. A. d’Aspremont, L. El Ghaoui, M.I. Jordan, and G. Lanckriet. A direct formulationfor sparse PCA using semidefinite programmingpca using semidefinite programming.SIAM Review, 49(3):434–448, 2007.
19. X. Qi, R. Luo, and H. Zhao. Sparse principal component analysis by choice of norm.Journal of multivariate analysis, 114:127–160, 2013.
20. V. Ozolins, R. Lai, R. Calflish, and S. Osher. Compressed plane waves - compactlysupported multiresolution basis for the Laplace operator. In preparation, 2013.
6 www.pnas.org/ Footline Author
V. Ozolins, R. Lai, R. E. Caflisch, and S. Osher, “Compressed modes … ,“ PNAS (2013)
Compressed plane waves
ψ CPW (x) = argminψ
− 12ψ Δψ + 1
µψ 1
⎛⎝⎜
⎞⎠⎟
s.t. ψ (x)ψ (x − jw) = δ0 j
Set V=0 to derive general compactly supported basis for the Laplacian:
V. Ozolins, R. Lai, R. E. Caflisch, and S. Osher, “Compressed plane waves … ,“ UCLA CAM Report (2013)
Compressed plane waves
V. Ozolins, R. Lai, R. E. Caflisch, and S. Osher, “Compressed plane waves … ,“ UCLA CAM Report (2013)
Introduction Compressed Modes (CMs) Compressed Plane Waves (CPWs)
Numerical algorithm and results
The first six basic CPWs
0 10 20 30 40 50 60 70 80 90 100!2
0
2
4
6
1
0 10 20 30 40 50 60 70 80 90 100
!5
0
5
2
0 10 20 30 40 50 60 70 80 90 100
!5
0
5
3
0 10 20 30 40 50 60 70 80 90 100
!5
0
5
4
0 10 20 30 40 50 60 70 80 90 100
!5
0
5
5
0 10 20 30 40 50 60 70 80 90 100
!5
0
5
6
Figure: From top to the bottom, the first six modes 1, 2, 3, 4, 5, 6 obtained by Eq. (10 -11)using L = 100, µ = 5, w = 5.
V. Ozolins & R. Lai & R. Caflisch & S. Osher Compressed Modes for Variational Problems in Math. & Phy.
Compressed plane waves
V. Ozolins, R. Lai, R. E. Caflisch, and S. Osher, “Compressed plane waves … ,“ UCLA CAM Report (2013)
Introduction Compressed Modes (CMs) Compressed Plane Waves (CPWs)
Numerical algorithm and results
Spectral density distribution
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0
0.2
0.4
0.6
0.8
1
Wave vector G
|!i (G
)|2
!1
!2
!3
!4
!5
!6
0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
0
0.2
0.4
0.6
0.8
1
Wave vector G
!i=
1
6|"
i (G)|
2
Figure: Spectral density distribution of CPWs. Top: the spectral density distribution of 1, 2, 3, 4, 5, 6. Bottom: The total spectral density distribution of the first four modes.
V. Ozolins & R. Lai & R. Caflisch & S. Osher Compressed Modes for Variational Problems in Math. & Phy.
Summary
✔ CS-based sparse physics is based on rigorous math, not intuition
✔ Accuracy can be improved systematically ✔ Constraints (symmetry and translational
invariance) can be straightforwardly incorporated ✔ CS automatically picks out important terms ✔ A simple prescription for gathering data