Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost...
Transcript of Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost...
![Page 1: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/1.jpg)
1
Stochastic optimization method for analytic continuation
A. S. Mishchenko RIKEN (Institute of Physical and Chemical Research), Japan
RRC “Kurchatov Institute”, Russia
![Page 2: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/2.jpg)
2
Stochastic optimization method for analytic continuation
1. Quantities one can get by QMC – correlation functions. 2. Examples of useful correlation functions. 3. Extracting physical information: necessity of analytic continuation. 4. General problem to solve: Fredholm integral equation of kind I. 5. Where similar problems are encountered? 6. Why the naïve methods fail? 7. Tikhonov-Phillips regularization – first successful approach. 8. More sophisticated methods: MaxEnt and Stochastic sampling. 9. Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles.
A. S. Mishchenko RIKEN (Institute of Physical and Chemical Research), Japan
RRC “Kurchatov Institute”, Russia
![Page 3: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/3.jpg)
3
Diagrammatic Monte Carlo and new method of analytic continuation
A. S. Mishchenko RIKEN (Institute of Physical and Chemical Research), Japan
![Page 4: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/4.jpg)
4
Diagrammatic Monte Carlo and new method of analytic continuation
A. S. Mishchenko RIKEN (Institute of Physical and Chemical Research), Japan
To Tokyo 500 m
From Tokyo
![Page 5: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/5.jpg)
5
Simple but not the simplest example: polaron
Examples of problems where one can get an important correlation function
![Page 6: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/6.jpg)
6
Simple but not the simplest example: polaron
Examples of problems where one can get an important correlation function
Green function by QMC
![Page 7: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/7.jpg)
7
Simple but not the simplest example: polaron
Examples of problems where one can get an important correlation function
Green function by QMC
No simple connection to measurable properties
![Page 8: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/8.jpg)
8
Physical properties under interest: Lehman function
Lehmann spectral function (LSF)
LSF has poles (sharp peaks) at the energies of stable (metastable) states. It is a measurable (in ARPES) quantity.
Noninteracting one is simple:
![Page 9: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/9.jpg)
9
Physical properties under interest: Lehmann function.
Lehmann spectral function (LSF)
LSF has poles (sharp peaks) at the energies of stable (metastable) states. It is a measurable (in ARPES) quantity.
LSF of one particle at T=0 can be determined from equation:
Fredholm first kind.
![Page 10: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/10.jpg)
10
Physical properties under interest: Z-factor and energy
Lehmann spectral function (LSF)
If the state with the lowest energy in the sector of given momentum is stable
The asymptotic behavior is
![Page 11: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/11.jpg)
11
Physical properties under interest: Z-factor and energy
The asymptotic behavior is
![Page 12: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/12.jpg)
12
Physical properties under interest: Lehmann function.
Lehmann spectral function (LSF)
LSF of one particle at T=0 can be determined from equation:
Solving of this equation is a notoriously difficult problem
![Page 13: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/13.jpg)
13
Physical properties under interest: Lehmann function.
Lehmann spectral function (LSF)
LSF of one particle at T=0 can be determined from equation:
Solving of this equation is a notoriously difficult problem
![Page 14: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/14.jpg)
14
Solution of integral equation
![Page 15: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/15.jpg)
15
Exciton
μ E
k
Coulomb attraction
Examples of problems where one can get an important correlation function
![Page 16: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/16.jpg)
16
Exciton
Exciton- polaron
More realistic
Infinite system
Examples of problems where one can get an important correlation function
![Page 17: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/17.jpg)
17
Exciton- polaron
μ E
k
Phonon exchange
+ Hel-ph
Coulomb attraction
Examples of problems where one can get an important correlation function
![Page 18: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/18.jpg)
18
μ E
k
Phonon exchange
Coulomb attraction
Exciton-polaron: two-particle Green function
Optical absorption
Examples of problems where one can get an important correlation function
![Page 19: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/19.jpg)
19
μ E
k
Phonon exchange
Coulomb attraction
Optical absorption
Examples of problems where one can get an important correlation function
Also Fredholm integral equation of the first kind
![Page 20: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/20.jpg)
Exact solution for optical spectra of exciton-polaron A. S. Mishchenko and N. Nagaosa, CMRG, RIKEN ASI
μ E
k
Diagrammatic Monte Carlo Exciton-polaron
Coulomb attraction No particle-phonon coupling
No Coulomb attraction Particle-phonon coupling
![Page 21: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/21.jpg)
There are a lot of problems where one has to solve Fredholm integral equation of the first kind
![Page 22: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/22.jpg)
Many-particle Fermi/Boson system in imaginary times representation
![Page 23: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/23.jpg)
Many-particle Fermi/Boson system in Matsubara representation
![Page 24: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/24.jpg)
Optical conductivity at finite T in imaginary times representation
![Page 25: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/25.jpg)
Image deblurring with e.g. known 2D noise K(m,ω)
K(m,ω) is a 2D x 2D noise distributon function
m and ω are 2D vectors
![Page 26: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/26.jpg)
Tomography image reconstruction (CT scan)
K(m,ω) is a 2D x 2D distribution function
m and ω are 2D vectors
![Page 27: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/27.jpg)
Aircraft stability
Nuclear reactor
operation
Image deblurring
A lot of other…
What is dramatic in the problem?
![Page 28: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/28.jpg)
Aircraft stability
Nuclear reactor
operation
Image deblurring
A lot of other…
What is dramatic in the problem?
Ill-posed!
![Page 29: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/29.jpg)
Ill-posed!
We cannot obtain an exact solution not because of some approximations of our approaches. Instead, we have to admit that the exact solution does not exist at all!
![Page 30: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/30.jpg)
Ill-posed!
1. No unique solution in mathematical sense
![Page 31: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/31.jpg)
Ill-posed!
1. No unique solution in mathematical sense
2. Some additional information is required which
specifies which kind of solution is expected
![Page 32: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/32.jpg)
Ill-posed!
1. No unique solution in mathematical sense
No function A to satisfy the equation
2. Some additional information is required which
specifies which kind of solution is expected. In order
to chose among many approximate solutions.
![Page 33: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/33.jpg)
Ill-posed!
Physics department: Max Ent.
How to solve?
![Page 34: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/34.jpg)
Ill-posed!
Physics department: Max Ent.
How to solve?
Engineering department: Tikhonov Regularization
![Page 35: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/35.jpg)
Ill-posed!
Physics department: Max Ent.
How to solve?
Engineering department: Tikhonov Regularization
Statistical department: ridge regression
![Page 36: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/36.jpg)
Ill-posed!
Physics department: Max Ent.
Engineering department: Tikhonov Regularization
Statistical department: ridge regression
Next player: stochastic methods
![Page 37: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/37.jpg)
Not settled!
• Still highly competitive field • Many approaches developed, some
specific ones are better for some specific cases
• Different approaches speak different languages, need some unified analysis
• Comparison of different methods, not just self-advertising, is needed
37
![Page 38: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/38.jpg)
Ill-posed!
Physics department: Max Ent. (Mark Jarrel)
Historically first: 1943: Tikhonov Regularization
Next player: stochastic Methods Since 1998
![Page 39: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/39.jpg)
Ill-posed!
The easiest way to explain the problem is to turn to a discrete form of the Fredholm equation
![Page 40: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/40.jpg)
Ill-posed!
The easiest way to explain the problem is to turn to a discrete form of the Fredholm equation
Because of noise present in the input data G(m) there is no unique A(ωn)=A(n) which exactly satisfies the equation.
![Page 41: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/41.jpg)
Ill-posed!
Because of noise present in the input data G(m) there is no unique A(ωn)=A(n) which exactly satisfies the equation. Hence, one can search for the least-square fitted solution A(n) which minimizes:
![Page 42: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/42.jpg)
Ill-posed!
Choosing the Euclidean norm one admits the absence of unique solution because there is an infinite number of deviation norms.
![Page 43: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/43.jpg)
Ill-posed!
Unique solution for the least-square fit through singular values decomposition of the matrix K
![Page 44: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/44.jpg)
Ill-posed!
Unique solution for the least-square fit through singular values decomposition of the matrix K
Explicit expression:
Typical singular values:
![Page 45: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/45.jpg)
Ill-posed!
Explicit expression:
Saw tooth noise instability due to small singular values.
![Page 46: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/46.jpg)
Ill-posed!
Saw tooth noise instability due to small singular values.
Truncating small singuar values (from 1 to 11)
![Page 47: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/47.jpg)
Ill-posed!
Tikhonov regularization to fight with the saw tooth noise instability.
![Page 48: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/48.jpg)
Ill-posed!
Tikhonov regularization to fight with the saw tooth noise instability.
Filter factors
![Page 49: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/49.jpg)
Ill-posed!
Tikhonov functional to minimize (Г is unitary matrix):
![Page 50: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/50.jpg)
Ill-posed!
General formulation of methods to
solve ill-posed problems in terms of
Bayesian statistical inference.
![Page 51: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/51.jpg)
Bayes theorem: P[A|G] P[G] = P[G|A] P[A] P[A|G] – conditional probability that the spectral function is A provided the correlation function is G
![Page 52: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/52.jpg)
Bayes theorem: P[A|G] P[G] = P[G|A] P[A] P[A|G] – conditional probability that the spectral function is A provided the correlation function is G
To find it is just the analytic continuation
![Page 53: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/53.jpg)
P[A|G] ~ P[G|A] P[A] P[G|A] is easier problem of finding G given A: likelihood function P[A] is prior knowledge about A:
Analytic continuation
![Page 54: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/54.jpg)
P[A|G] ~ P[G|A] P[A] P[G|A] is easier problem of finding G given A: likelihood function P[A] is prior knowledge about A:
All methods to solve the above problem can be formulated in terms of this relation
![Page 55: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/55.jpg)
Historically first method to solve the problem of Fredholm kind I integral equation. Tikhonov regularization method (1943) A.N.Tikhonov, Dokladyu Akdemii Nauk SSSR (1943) A.N.Tikhonov, Dokladyu Akdemii Nauk SSSR (1963) (Soviet mathematics)
Tikhonov & Arsenin, Solution of Ill-posed problems, (Washington, 1977).
![Page 56: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/56.jpg)
Historically first method to solve the problem of Fredholm kind I integral equation. A.N.Tikhonov, Dokladyu Akdemii Nauk SSSR (1943) Tikhonov & Arsenin, Solution of Ill-posed problems, (Washington, 1977).
The regularization method was developed not by …… in 1977, it was rediscovered….
![Page 57: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/57.jpg)
Historically first method to solve the problem of Fredholm kind I integral equation. Tikhonov regularization method (1943)
![Page 58: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/58.jpg)
Tikhonov regularization method (1943)
If Г is unit matrix:
![Page 59: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/59.jpg)
Ill-posed!
Filter factors
![Page 60: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/60.jpg)
Ill-posed!
Tikhonov regularization to fight with the saw tooth noise instability.
![Page 61: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/61.jpg)
Ill-posed!
Tikhonov regularization to fight with the saw tooth noise instability.
Is it not too smooth???
![Page 62: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/62.jpg)
P[A|G] ~ P[G|A] P[A] Maximum entropy method
Likelihood (objective) function
Prior knowledge function
![Page 63: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/63.jpg)
P[A|G] ~ P[G|A] P[A] Maximum entropy method
Prior knowledge function
D(ω) is default model
![Page 64: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/64.jpg)
P[A|G] ~ P[G|A] P[A] Maximum entropy method
Prior knowledge function
1. One has escaped extra smoothening.
2. But one has got default model as an extra price.
![Page 65: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/65.jpg)
P[A|G] ~ P[G|A] P[A]
Prior knowledge function
1. We want to avoid extra smoothening.
2. We want to avoid default model as an extra price.
Maximum entropy method
![Page 66: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/66.jpg)
P[A|G] ~ P[G|A] P[A] Maximum entropy method
Prior knowledge function
Both items (extra smoothening and arbitrary default model) can be somehow circumvented by the group of stochastic methods.
![Page 67: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/67.jpg)
P[A|G] ~ P[G|A] P[A] Stochastic methods
Both items (extra smoothening and arbitrary default model) can be somehow circumvented by the group of stochastic methods.
![Page 68: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/68.jpg)
P[A|G] ~ P[G|A] P[A] Stochastic methods
The main idea of the stochastic methods is: 1. Restrict the prior knowledge to the minimal possible level (positive, normalized, etc…). 2. Change the likelihood function to the likelihood functional.
![Page 69: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/69.jpg)
P[A|G] ~ P[G|A] P[A] Stochastic methods
The main idea of the stochastic methods is: 1. Restrict the prior knowledge to the minimal possible level (positive, normalized, etc…). Avoids default model. 2. Change the likelihood function to the likelihood functional. Avoids saw-tooth noise.
![Page 70: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/70.jpg)
P[A|G] ~ P[G|A] P[A] Stochastic methods
Change the likelihood function to the likelihood functional. Avoids sawtooth noise.
![Page 71: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/71.jpg)
Stochastic methods Likelihood functional. Avoids sawtooth noise.
Sandvik, Phys. Rev. B 1998, is the first practical attempt to think stochastically.
![Page 72: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/72.jpg)
Stochastic methods Likelihood functional. Avoids sawtooth noise.
Sandvik, Phys. Rev. B 1998, is the first practical attempt to think stochastically.
SOM was suggested in 2000. Mishchenko et al, Appendix B in Phys. Rev. B.
![Page 73: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/73.jpg)
73
Some applications of SOM: Phys. Rev. Lett., vol. 86, 4624 (2001) Phys. Rev. Lett., vol. 87, 186402 (2001) Phys. Rev. Lett., vol. 91, 236401 (2003) Phys. Rev. Lett., vol. 93, 036402 (2004) Phys. Rev. Lett., vol. 96, 136405 (2006) Phys. Rev. Lett., vol. 99, 226402 (2007) Phys. Rev. Lett., vol. 100, 166401 (2008) Phys. Rev. Lett., vol. 101, 116403 (2008) Phys. Rev. Lett., vol. 104, 056602 (2010) Phys. Rev. Lett., vol. 107, 076403 (2011)
![Page 74: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/74.jpg)
Stochastic methods Likelihood functional. Avoids sawtooth noise.
Sandvik, Phys. Rev. B 1998, is the first practical attempt to think stochastically.
SOM was suggested in 2000. Mishchenko et al, Appendix B in Phys. Rev. B.
![Page 75: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/75.jpg)
Stochastic methods Likelihood functional. Avoids sawtooth noise.
Sandvik, Phys. Rev. B 1998, is the first practical attempt to think stochastically.
SOM was suggested in 2000. Mishchenko et al, Appendix B in Phys. Rev. B.
Beach, 2004, cond-mat
![Page 76: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/76.jpg)
Stochastic methods Likelihood functional. Avoids sawtooth noise.
Sandvik, Phys. Rev. B 1998, is the first practical attempt to think stochastically.
SOM was suggested in 2000. Mishchenko et al, Appendix B in Phys. Rev. B.
Beach, 2004, cond-mat
Other variants after 2004
![Page 77: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/77.jpg)
Stochastic methods Likelihood functional. Avoids sawtooth noise.
What is the special need for the stochastic sampling methods?
![Page 78: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/78.jpg)
Stochastic methods
What is the special need for the stochastic sampling methods?
1. Avoid saw-tooth noise. 2. Avoid over-smoothing of the δ-function
Typical spectrum of QP at T=0.
![Page 79: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/79.jpg)
Stochastic methods
What is the special need for the stochastic sampling methods?
1. Avoid saw-tooth noise. 2. Avoid over-smoothing of the δ-function Tikhonow regularization over-smoothes the δ-function. MaxEnt – default model has to fix δ-function in advance.
Typical spectrum of QP at T=0.
![Page 80: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/80.jpg)
Stochastic methods
What is the special need for the stochastic sampling methods?
1. Avoid saw-tooth noise. 2. Avoid over-smoothing of the δ-function Stochastic methods is a way to circumvent these problems.
Typical spectrum of QP at T=0.
![Page 81: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/81.jpg)
Stochastic methods
Back to Sandvik approach
![Page 82: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/82.jpg)
Stochastic methods
One does not search for a single solution but samples through difference “configurations” (spectral functions A) Using the likelihood function P which is characterized by fictitious “temperature” T and fictitious “energy” χ2.
![Page 83: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/83.jpg)
Stochastic methods
One does not search for a single solution but samples through difference “configurations” (spectral functions A) Using the likelihood function P which is characterized by fictitious “temperature” T and fictitious “energy” χ2.
One interprets χ2 as an “energy” of fictitious Hamiltonian and T as a fictitious “temperature”. Hence, one involves the Metropolis algorithm for Monte Carlo to sample through configurations A.
![Page 84: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/84.jpg)
Stochastic methods
One interprets χ2 as an “energy” of fictitious Hamiltonian and T as a fictitious “temperature”. Hence, one involves the Metropolis algorithm for Monte Carlo to sample through configurations. 1. T is not too high. Otherwise A is far from spectra which fit well the correlation function G. 2. T is not too small otherwise we are back again to the sawtooth noise problem. Ower-fitting of the noise.
![Page 85: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/85.jpg)
Stochastic methods
One interprets χ2 as an “energy” of fictitious Hamiltonian and T as a fictitious “temperature”. Hence, one involves the Metropolis algorithm for Monte Carlo to sample through configurations.
Simple rule T = M
1. T is not too high. Otherwise A is far from spectra which fit well the correlation function G. 2. T is not too small otherwise we are back again to the sawtooth noise problem. Ower-fitting of the noise.
![Page 86: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/86.jpg)
Stochastic methods
One interprets χ2 as an “energy” of fictitious Hamiltonian and T as a fictitious “temperature”. Hence, one involves the Metropolis algorithm for Monte Carlo to sample through configurations.
Which features of Sandvik method are artificial? 1. There is no real Hamiltonian and T and, hence, one has no requirement to sample through Metropolis 2. Algorithm is not effective at low T and use the tempering procedures with sampling at different Ts.
![Page 87: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/87.jpg)
Stochastic optimization method.
1. One has to sample through solutions A(ω) which fit the correlation function G well.
2. One has to make some weighted sum of these well solutions A(ω).
![Page 88: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/88.jpg)
SOM is very similar to Sandvik method but circumvent its artificial features and, as a result, turns out more effective
88
![Page 89: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/89.jpg)
Stochastic optimization method.
One collects and averages large amount of “well” solutions and take an average.
1. What is the likelihood function (deviation measure of fit quality? 2. How the spectrum is parameterized 3. How to find one “well” solution? 4. When the number of solutions is enough? 5. Tests.
![Page 90: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/90.jpg)
Stochastic optimization method.
1. What is the likelihood function (deviation measure of fit quality?
![Page 91: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/91.jpg)
Stochastic optimization method. Parameterization of the particular solution:
![Page 92: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/92.jpg)
Stochastic optimization method. Parameterization of the particular solution:
No predefined mesh for the energy (ω) space.
![Page 93: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/93.jpg)
Stochastic optimization method. Contribution of rectangle to
If no analytic expression.
One tabulates:
Contribution:
![Page 94: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/94.jpg)
Stochastic optimization method. Contribution of rectangle to
Particular cases. Imaginary time, T=0:
K(m,ω) = exp(-iτmω)
Kernel is
Contribution:
![Page 95: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/95.jpg)
Stochastic optimization method. Contribution of rectangle to
Particular cases. Matsubara, any T:
Kernel is
Contribution:
![Page 96: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/96.jpg)
96
Stochastic Optimization method.
Particular solution L(i)(ω) for LSF is presented as a sum of a number K of rectangles with some width, height and center.
L
L
ω
ω
![Page 97: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/97.jpg)
97
How to find
one of solutions?
![Page 98: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/98.jpg)
98
Stochastic Optimization method.
Particular solution L(i)(ω) for LSF is presented as a sum of a number K of rectangles with some width, height and center. Initial configuration of rectangles is created by random number generator (i.e. number K and all parameters of of rectangles are randomly generated).
L
ω
![Page 99: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/99.jpg)
99
Stochastic Optimization method.
Particular solution L(i)(ω) for LSF is presented as a sum of a number K of rectangles with some width, height and center. Initial configuration of rectangles is created by random number generator (i.e. number N and all parameters of of rectangles are randomly generated). Each particular solution L(i)(ω) is obtained by a naïve method without regularization (though, varying number K).
L ~ ~
Deviation measure for configuration:
![Page 100: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/100.jpg)
100
How to minimize
the deviation?
Which updates?
L ~ ~
![Page 101: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/101.jpg)
101
Stochastic Optimization method: update procedures.
L
ω
Shift of rectangular.
L
ω
![Page 102: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/102.jpg)
102
Stochastic Optimization method: update procedures.
L
ω
Change of height of rectangular with fixed Z-factor.
L
ω
![Page 103: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/103.jpg)
103
Stochastic Optimization method: update procedures.
L
ω
Split of rectangular.
L
ω
![Page 104: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/104.jpg)
104
Stochastic Optimization method: update procedures.
L
ω
Glue two rectangles.
L
ω
![Page 105: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/105.jpg)
105
Parameters for changing are obtained by optimizing some continuous parameter making quadratic (intra)extrapolation. For example: Measure of deviation for the shift of rectangle is calculated for distances x, 2x, and 3x and then 3 points D(x) is reproduced by parabola. Variable x can be any other continuous parameter of the update.
![Page 106: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/106.jpg)
106
Stochastic Optimization method: update procedures.
CONFIGURATON
Deviation measure.
Accept only updates which decrease the deviation measure
WRONG STRATEGY
![Page 107: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/107.jpg)
107
Stochastic Optimization method: update procedures.
CONFIGURATON
Deviation measure.
Always accept with some probability some updates which decrease the deviation measure WRONG STRATEGY: Sandvik 1998, Beach 2004
![Page 108: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/108.jpg)
108
Stochastic Optimization method: update procedures.
CONFIGURATON
Deviation measure.
Shake-off two-step strategy: Step 1: Increase of deviation measure is allowed during M steps with high probability Step 2: Only decrease of deviation measure is allowed during last K steps.
![Page 109: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/109.jpg)
109
Stochastic Optimization method: update procedures.
CONFIGURATON
Deviation measure.
Shake-off two-step strategy: Step 1: Increase of deviation measure is allowed during M steps with high probability Step 2: Only decrease of deviation measure is allowed during last K steps.
K+M chain is rejected if final D is larger than initial
![Page 110: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/110.jpg)
110
Stochastic Optimization method: update procedures.
CONFIGURATON
Deviation measure.
Shake-off two-step strategy: Step 1: Increase of deviation measure is allowed during M steps with high probability Step 2: Only decrease of deviation measure is allowed during last K steps.
K+M chain is
accepted if final D
is smaller than initial
![Page 111: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/111.jpg)
111
How to judge that one of solutions
is “GOOD”
![Page 112: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/112.jpg)
How to judge that one of solutions is “GOOD”
Κ>1/4 (Ideal limit κ=1/2)
![Page 113: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/113.jpg)
113
Stochastic Optimization method.
Particular solution L(i)(ω) for LSF is presented as a sum of a number K of rectangles with some width, height and center. Initial configuration of rectangles is created by random number generator (i.e. number K and all parameters of of rectangles are randomly generated). Each particular solution L(i)(ω) is obtained by a naïve method without regularization (though, varying number K). Final solution is obtained after M steps of such procedure L(ω) = M-1 ∑i
L(i)(ω) Each particular solution has saw tooth noise Final averaged solution L(ω) has no saw tooth noise though not regularized with sharp peaks/edges!!!!
![Page 114: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/114.jpg)
114
We can find many particular solutions each of which
fits the input data reasonably.
![Page 115: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/115.jpg)
115
We can find many particular solutions each of which
fits the input data reasonably.
Which particular solutions one has to
take into account?
![Page 116: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/116.jpg)
116
Self-averaging of the saw-tooth noise.
![Page 117: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/117.jpg)
117
Self-averaging of the saw-tooth noise.
![Page 118: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/118.jpg)
118
Self-averaging of the saw-tooth noise.
![Page 119: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/119.jpg)
119
Which particular solutions one has to take into account?
![Page 120: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/120.jpg)
120
Which particular solutions one has to take into account?
One has to include solution with deviation measure D[A] which is less that twice of minimal MIN{D[A]} D[A] < 2 MIN{D[A]}
![Page 121: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/121.jpg)
Sandvik method
1. T is not too high. Otherwise A is far from spectra which fit well the correlation function G. 2. T is not too small otherwise we are back again to the sawtooth noise problem. Ower-fitting of the noise.
Simple rule T = M
![Page 122: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/122.jpg)
Tikhonov functional: similar strategy for choice of λ
![Page 123: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/123.jpg)
Max Ent
123
Similar strategy everywhere: equate noise contribution
with regularization contribution
Avoid over-fitting
![Page 124: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/124.jpg)
124
Similar strategy everywhere: equate noise contribution
with regularization contribution
Tikhonov & Arsenin, Solution of Ill-posed problems, (Washington, 1977). Arsenin (1986): the art of finding solution for ill posed problem lies in an intuition which tells us when to stop improve the deviation before the noise of input data overruns the information contained in the input data.
![Page 125: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/125.jpg)
125
Which particular solutions one has to take into account?
One has to include solution with deviation measure D[A] which is less that twice of minimal MIN{D[A]} D[A] < 2 MIN{D[A]}
![Page 126: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/126.jpg)
126
Some tests
![Page 127: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/127.jpg)
127
Some tests
B=10-4
![Page 128: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/128.jpg)
128
Some tests
Particular cases.
Imaginary time, T=0:
K(m,ω) = exp(-τmω)
Kernel is
![Page 129: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/129.jpg)
129
Some tests
Particular cases.
Imaginary time, finite T, fermions
Kernel is
![Page 130: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/130.jpg)
130
Some tests
Particular cases.
Imaginary time, finite T, optical conductivity
Kernel is
![Page 131: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/131.jpg)
131
Some tests
Particular cases.
Matsubara frequencies, finite T, fermions
Kernel is
![Page 132: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/132.jpg)
132
Particular cases.
Imaginary time, finite T, optical conductivity
Kernel is
Back to optical conductivity. Let us compare MaxEnt and Stochastic.
![Page 133: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/133.jpg)
133
![Page 134: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/134.jpg)
134
Stochastic Optimization method.
Particular solution L(i)(ω) for LSF is presented as a sum of a number K of rectangles with some width, height and center. Initial configuration of rectangles is created by random number generator (i.e. number K and all parameters of of rectangles are randomly generated). Each particular solution L(i)(ω) is obtained by a naïve method without regularization (though, varying number K). Final solution is obtained after M steps of such procedure L(ω) = M-1 ∑i
L(i)(ω) Each particular solution has saw tooth noise Final averaged solution L(ω) has no saw tooth noise though not regularized with sharp peaks/edges!!!!
![Page 135: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/135.jpg)
135
Conclusions: 1. Analytic continuation is ill posed problem. 2. Similar Fredholm I integral equatuion problem in many
applications. 3. Long history of the methods: Tikhonov -> MaxEnt ->
stochastic. 4. All methods bear similar strategy of regularization: not
to over-fit the noise 5. Each method is the best in each particular case. There
is no universal method which is “the best” for all cases. 6. We are still on the way to improve the analytic
continuation. 7. Combinations of methods might help.
![Page 136: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/136.jpg)
136
Questions?
![Page 137: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/137.jpg)
New Method for Low Temperature analysis of the ESR spectra
137
Andrey Mishchenko CMRG, RIKEN
Collaborations:
Tatsuo Hasegawa (AIST) Hiroyuki Matsui (AIST)
Phys Rev. Lett. 104, 056602 (2010)
![Page 138: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/138.jpg)
New Method for Low Temperature analysis of the ESR spectra
138
1. Nature of the inhomogeneous ESR lineshape and line narrowing 3. Analysis of the lineshape of an electron trapped by an impurity 4. Analysis of the fine structure of the ESR line can give a complete information on the distribution of the traps versus localization parameters
![Page 139: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/139.jpg)
139
Basics of ESR Transition between Zeeman split levels under the influence of the electromagnetic field. For example, the frequency is fixed and magnetic field B is varied. Then, the intensity of signal I(B) is I(B) ~ δ(B-B0)
![Page 140: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/140.jpg)
140
Basics of ESR Transition between Zeeman split levels under the influence of the electromagnetic field. For example, the frequency is fixed and magnetic field B is varied. Then, the intensity of signal I(B) is I(B) ~ δ(B-B0)
Hyperfine splitting
![Page 141: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/141.jpg)
141
Basics of ESR Transition between Zeeman split levels under the influence of the electromagnetic field. For example, the frequency is fixed and magnetic field B is varied. Then, the intensity of signal I(B) is I(B) ~ δ(B-B0)
Hyperfine splitting
In complex system at low temperatures the lineshape is set by the sum of random contributions coming from
hyperfine and superhyperfine interactions.
![Page 142: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/142.jpg)
142
Basics of ESR
Hyperfine splitting
In complex system at low temperatures the lineshape is set by the sum of random contributions coming from
hyperfine and superhyperfine interactions. The distribution of the sum of random variable is Gaussian:
S(B) = dG(B) / dB
![Page 143: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/143.jpg)
143
Basics of ESR The distribution of the sum of random variable is Gaussian:
If the electron is spread over N molecules, one has a distribution over sum of random variables. Then, according to the Central Limit Theorem the distribution is Gaussian with more narrow dispersion σ:
![Page 144: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/144.jpg)
144
Basics of ESR
N molecules, Ai=Ai/N, ni=ni/N
![Page 145: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/145.jpg)
145
If the electron is spread over N molecules, one has a distribution over sum jf random variables. Then, according to the Central Limit Theorem the distribution is Gaussian with more narrow dispersion σ:
Above knowledge is from the theory of inhomogeneous lineshape in molecules. In solids? When an electron is localized on the trap, there is a charge distribution f(i) and one needs to look at the distribution of different variable <B>:
![Page 146: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/146.jpg)
146
If the electron is spread over N molecules, one has a distribution over sum jf random variables. Then, according to the Central Limit Theorem the distribution is Gaussian with more narrow dispersion σ:
Above knowledge is from the theory of inhomogeneous lineshape in molecules. In solids? When an electron is localized on the trap, there is a charge distribution f(i) and one needs to look at the distribution of different variable <B>:
Experiment reveals non-Gaussian signal. Maybe this is the reason.
![Page 147: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/147.jpg)
147
If the electron is spread over N molecules, one has a distribution over sum of random variables. Then, according to the Central Limit Theorem the distribution is Gaussian with more narrow dispersion σ:
Above knowledge is from the theory of inhomogeneous lineshape in molecules. In solids? When an electron is localized on the trap, there is a charge distribution f(i) and one needs to look at the distribution of variables <B>:
![Page 148: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/148.jpg)
148
If the electron is spread over N molecules, one has a distribution over sum jf random variables. Then, according to the Central Limit Theorem the distribution is Gaussian with more narrow dispersion σ:
Above knowledge is from the theory of inhomogeneous lineshape in molecules. In solids? When an electron is localized on the trap, there is a charge distribution f(i) and one needs to look at the distribution of variables <B>:
Numerical simulations show that (although CLT
does not work in this case) the distribution is still
Gaussian.
![Page 149: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/149.jpg)
149
CLT for non-uniformly distributed variables: nontrivial!
The uniform distribution is the best case for narrowing
In extreme limit of localized case Neff 1.
![Page 150: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/150.jpg)
150
CLT for non-uniformly distributed variables: nontrivial!
Neff = [ Σi p(i)2 ]-1/2
![Page 151: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/151.jpg)
151
However, experimental signal is not Gaussian which means that there is no traps with some given value of Neff which dominate.
Experimental ESR in pentacene
For T<50K the wave saturation experiment shows that all carriers are localized and no broadening except nonhomogeneous one is expected!
![Page 152: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/152.jpg)
152
However, experimental signal is not Gaussian which means that there is no traps with some given value of Neff which dominate.
Experimental ESR in pentacene
Two kinds of traps:
I(B) = α G(N1eff ,B-B0) + β G(N2
eff ,B-B0)
![Page 153: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/153.jpg)
153
However, experimental signal is not Gaussian which means that there is no traps with some given value of Neff which dominate.
Experimental ESR in pentacene
Two kinds of traps:
I(B) = α G(N1eff ,B-B0) + β G(N2
eff ,B-B0)
Three kinds of traps:
I(B) = α G(N1eff ,B-B0) + β G(N2
eff ,B-B0) + γ G(N3eff ,B-B0)
![Page 154: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/154.jpg)
154
Experimental ESR in pentacene Broader
view: distribution
of traps
![Page 155: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/155.jpg)
155
Experimental ESR in pentacene
Note, such interpretation
requires that all molecules of molecular crystal are
equally oriented with respect to surface.
Broader view:
distribution of
traps
![Page 156: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/156.jpg)
156
Fredholm integral equation of the 1-st kind: so called ill posed problem
Previously (A.S. Mishchenko et al, PRB v. 62, 6317 (2000)) a method more flexible and less capricious than MEM was developed for solving the analytic continuation problem -
Stochastic Optimization
method
![Page 157: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/157.jpg)
157
![Page 158: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/158.jpg)
158
![Page 159: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/159.jpg)
159
There are 100-s methods to fit the signal. We say for the 1-st time that we do not need any fit!!!
ESR spectrum of organic FET
![Page 160: Stochastic optimization method for analytic …Stochastic optimization method (SOM) as the utmost accomplishment of stochastic methods principles. A. S. Mishchenko RIKEN (Institute](https://reader034.fdocuments.us/reader034/viewer/2022050102/5f418adfb17f6462287bd063/html5/thumbnails/160.jpg)
160
ESR spectrum of organic FET Reliability of result: