The document describes applications of non-negative polynomial conic optimization (NPCO) and moment conic optimization (MCO). It discusses how problems in statistics, finance, and approximation can be formulated as NPCO and MCO problems. Examples are given such as finding the best fitting positive polynomial to data points, modeling functions that must be increasing and convex, and determining an envelope polynomial that approximates other polynomials. The applications demonstrate how shape constraints on functions can be modeled using conic constraints in NPCO and MCO.
The document discusses Chebyshev polynomials. It defines Chebyshev polynomials as orthogonal polynomials related to de Moivre's formula that can be defined recursively. It provides key properties of Chebyshev polynomials including that they are the extremal polynomials for many properties and important in approximation theory. The document also provides formulas for generating Chebyshev polynomials, their orthogonality properties, and their use in representing functions through orthogonal series expansions.
The document discusses accelerated computing through Quasi-Monte Carlo (QMC) constructions for numerical integration. It covers the theoretical background of Monte Carlo (MC), Quasi-Monte Carlo (QMC), and Randomized QMC (RQMC) methods. It also provides examples of using QMC and polynomial lattice rules for applications like option pricing in finance. The CEO presents on developing efficient software for QMC using a C++ implementation.
This document provides information about getting fully solved assignments from an assignment help service. It lists an email address and phone number to contact for assistance with assignments. It also provides details about the available programs, subjects, semesters, credits, and other assignment details like word count requirements. Students are advised to mail their request with details of their semester and specialization to get solved assignments. Calling is listed as an emergency option.
This document discusses iterative Monte Carlo methods for solving linear systems. It begins by motivating the use of stochastic approaches for solving highly dimensional systems due to the occurrence of faults in parallel computing environments. It then provides the mathematical setting for standard Monte Carlo linear solvers using both forward and adjoint methods. It discusses how the solution can be written as a power series and evaluated using random walks. It also discusses the necessary statistical constraints for the estimators to converge, such as the spectral radii of the preconditioned matrix being less than one. Finally, it addresses the choice of transition probabilities for the random walks to guarantee the constraints.
Slides of pattern recognition Course of Professor Zohreh Azimifar at Shiraz University.
اسلاید های درس شناسایی آماری الگو استاد زهره عظیمی فر در دانشگاه شیراز.
This document discusses forecast combination techniques in R using the ForecastCombinations package. It begins with an introduction to forecast combination, explaining that combining multiple forecasts can improve accuracy since different models perform better under different conditions. It then describes various combination schemes like regression-based, accuracy-based, and selecting the best individual forecast. Practical examples on topics like PPP estimation and GDP measurement are provided. Potential issues with interpretation and when combination may not be useful are discussed. The document concludes with references for further research.
The document discusses Chebyshev polynomials. It defines Chebyshev polynomials as orthogonal polynomials related to de Moivre's formula that can be defined recursively. It provides key properties of Chebyshev polynomials including that they are the extremal polynomials for many properties and important in approximation theory. The document also provides formulas for generating Chebyshev polynomials, their orthogonality properties, and their use in representing functions through orthogonal series expansions.
The document discusses accelerated computing through Quasi-Monte Carlo (QMC) constructions for numerical integration. It covers the theoretical background of Monte Carlo (MC), Quasi-Monte Carlo (QMC), and Randomized QMC (RQMC) methods. It also provides examples of using QMC and polynomial lattice rules for applications like option pricing in finance. The CEO presents on developing efficient software for QMC using a C++ implementation.
This document provides information about getting fully solved assignments from an assignment help service. It lists an email address and phone number to contact for assistance with assignments. It also provides details about the available programs, subjects, semesters, credits, and other assignment details like word count requirements. Students are advised to mail their request with details of their semester and specialization to get solved assignments. Calling is listed as an emergency option.
This document discusses iterative Monte Carlo methods for solving linear systems. It begins by motivating the use of stochastic approaches for solving highly dimensional systems due to the occurrence of faults in parallel computing environments. It then provides the mathematical setting for standard Monte Carlo linear solvers using both forward and adjoint methods. It discusses how the solution can be written as a power series and evaluated using random walks. It also discusses the necessary statistical constraints for the estimators to converge, such as the spectral radii of the preconditioned matrix being less than one. Finally, it addresses the choice of transition probabilities for the random walks to guarantee the constraints.
Slides of pattern recognition Course of Professor Zohreh Azimifar at Shiraz University.
اسلاید های درس شناسایی آماری الگو استاد زهره عظیمی فر در دانشگاه شیراز.
This document discusses forecast combination techniques in R using the ForecastCombinations package. It begins with an introduction to forecast combination, explaining that combining multiple forecasts can improve accuracy since different models perform better under different conditions. It then describes various combination schemes like regression-based, accuracy-based, and selecting the best individual forecast. Practical examples on topics like PPP estimation and GDP measurement are provided. Potential issues with interpretation and when combination may not be useful are discussed. The document concludes with references for further research.
This document provides an overview of optimization techniques. It begins with an introduction to optimization and examples of optimization problems. It then discusses the historical development of optimization methods. The document categorizes optimization problems into convex, concave, and subclasses like linear programming, quadratic programming, and semidefinite programming. It also covers advanced optimization methods like interior point methods. Finally, the document discusses applications of convex optimization techniques in fields such as engineering, finance, and wireless communications.
Monte Carlo Simulations in Ad Lift Measurement using SparkPrasad Chalasani
The document discusses methods for measuring the impact of advertising using randomized experiments and observational studies. It describes estimating response rates from experimental data and calculating Bayesian confidence intervals around those rates by sampling from the posterior distribution. It also explains how to estimate advertising lift and confidence bounds around lift by sampling from the posterior distributions of both the control and test response rates. The key ideas are using Bayesian methods and Gibbs sampling to account for uncertainty in estimates of response rates and lift from experimental data.
Monte Carlo Simulations in Ad-Lift Measurement Using Spark by Prasad Chalasan...Spark Summit
I listen to approximately 100 billion ad opportunities daily and respond with optimal bids within milliseconds. Predicting user response to ads is a machine learning problem, but quantifying the impact of ad exposure is a measurement problem. Key conceptual takeaways include issues in ad lift measurement, proper definition, confidence bounds, Bayesian methods for ad lift confidence bounds using Gibbs sampling and Markov chain Monte Carlo, and using Spark for Monte Carlo sampling and simulations.
This document discusses smart buildings and precincts, including building energy management. It introduces model predictive control (MPC) as a technique to optimize building energy usage. MPC uses mathematical models to predict a building's thermal response and compute optimal control inputs over a future horizon to minimize energy costs while maintaining comfort levels. The document provides an example thermal model of a building and shows experimental data matching the model's predictions. It also discusses using low-cost IoT sensors and controllers throughout a building to monitor conditions and implement optimized control strategies from an MPC system.
The document summarizes research comparing the Particle Swarm Optimization (PSO) and Differential Evolution (DE) algorithms for optimizing power consumption using smart energy meter data. Both algorithms were implemented in MATLAB and tested on 15 days of meter data from a university lab in India. PSO achieved an 11.5% reduction in power consumption while DE achieved a 9.4% reduction. PSO outperformed DE for this application, showing it is an effective technique for optimizing energy use and reducing electricity costs for consumers. Future work could integrate the models with real smart meters and controllers to achieve automated scheduling and greater savings.
Building Institutional Capacity in Thailand to Design and Implement Climate P...UNDP Climate
23-25 November 2016, Thailand - A centerpiece of the Integrating Agriculture in National Adaptation Plans Programme (NAP-Ag) in Thailand is its support to develop a new five-year Strategy on Climate Change in Agriculture (2017-2021). This is spearheaded by the Ministry of Agriculture and Cooperatives (MOAC) and its Office of Agriculture Economics (OAE). The strategy was unveiled after a series of meetings by a Technical Working Group at a three-day workshop held on 23-25 November 2016 in Bangkok, organized by UNDP. Over 60 participants from each MOAC line department and 10 participants from academia and civil society were briefed by the Office of the Natural Resources and Environmental Policy and Planning (ONEP) and GIZ on the status of the National Adaption Plan (NAP) and learned how NAP-Ag programme efforts could support a broader NAP process and align with the Sector Plan. The new strategy focuses on improving evidence and data for informing policy choices, building the capacity of farmers and agri-businesses to adapt, promoting low-carbon development and productivity growth in the sector, and building institutional and managerial capacities to cope with climate change impacts.
The document discusses Boolean algebra and its applications in combinational logic circuit design. It covers topics like Boolean expressions, standard forms (sum of products and product of sums), converting between forms, truth tables, and determining logic expressions from truth tables. Standard forms allow for simplification using techniques like Karnaugh maps. Boolean algebra is used to analyze and design basic combinational circuits like encoders, decoders, and adders.
For some management programming problems, multiple objectives to be optimized rather than a single
objective, and objectives can be expressed with ratio equations such as return/investment, operating
profit/net-sales, profit/manufacturing cost, etc. In this paper, we proposed the transformation
characteristics to solve the multi objective linear fractional programming (MOLFP) problems. If a MOLFP
problem with both the numerators and the denominators of the objectives are linear functions and some
technical linear restrictions are satisfied, then it is defined as a multi objective linear fractional
programming problem MOLFPP in this research. The transformation characteristics are illustrated and the
solution procedure and numerical example are presented.
This document summarizes a presentation on learning to hash for large scale image retrieval. It discusses locality sensitive hashing (LSH), which maps similar data points to the same "buckets" with high probability to enable efficient nearest neighbor search. The presentation evaluates different approaches to learn the quantization thresholds and hashing hypersurfaces in LSH, including learning multiple thresholds to better capture data distributions. It proposes a semi-supervised objective function to jointly optimize the thresholds based on supervised neighborhood information and unsupervised projections between points. The goal is to improve retrieval effectiveness by making the hashing more influenced by data distributions.
For some management programming problems, multiple objectives to be optimized rather than a single objective, and objectives can be expressed with ratio equations such as return/investment, operating profit/net-sales, profit/manufacturing cost, etc. In this paper, we proposed the transformation characteristics to solve the multi objective linear fractional programming (MOLFP) problems. If a MOLFP problem with both the numerators and the denominators of the objectives are linear functions and some technical linear restrictions are satisfied, then it is defined as a multi objective linear fractional programming problem MOLFPP in this research. The transformation characteristics are illustrated and the solution procedure and numerical example are presented.
APPLYING TRANSFORMATION CHARACTERISTICS TO SOLVE THE MULTI OBJECTIVE LINEAR F...ijcsit
For some management programming problems, multiple objectives to be optimized rather than a single objective, and objectives can be expressed with ratio equations such as return/investment, operating
profit/net-sales, profit/manufacturing cost, etc. In this paper, we proposed the transformation characteristics to solve the multi objective linear fractional programming (MOLFP) problems. If a MOLFP problem with both the numerators and the denominators of the objectives are linear functions and some
technical linear restrictions are satisfied, then it is defined as a multi objective linear fractional programming problem MOLFPP in this research. The transformation characteristics are illustrated and the solution procedure and numerical example are presented.
This document provides an overview of optimization techniques. It begins with an introduction to optimization and examples of optimization problems. It then discusses the historical development of optimization methods. The document categorizes optimization problems into convex, concave, and subclasses like linear programming, quadratic programming, and semidefinite programming. It also covers advanced optimization methods like interior point methods. Finally, the document discusses applications of convex optimization techniques in fields such as engineering, finance, and wireless communications.
Monte Carlo Simulations in Ad Lift Measurement using SparkPrasad Chalasani
The document discusses methods for measuring the impact of advertising using randomized experiments and observational studies. It describes estimating response rates from experimental data and calculating Bayesian confidence intervals around those rates by sampling from the posterior distribution. It also explains how to estimate advertising lift and confidence bounds around lift by sampling from the posterior distributions of both the control and test response rates. The key ideas are using Bayesian methods and Gibbs sampling to account for uncertainty in estimates of response rates and lift from experimental data.
Monte Carlo Simulations in Ad-Lift Measurement Using Spark by Prasad Chalasan...Spark Summit
I listen to approximately 100 billion ad opportunities daily and respond with optimal bids within milliseconds. Predicting user response to ads is a machine learning problem, but quantifying the impact of ad exposure is a measurement problem. Key conceptual takeaways include issues in ad lift measurement, proper definition, confidence bounds, Bayesian methods for ad lift confidence bounds using Gibbs sampling and Markov chain Monte Carlo, and using Spark for Monte Carlo sampling and simulations.
This document discusses smart buildings and precincts, including building energy management. It introduces model predictive control (MPC) as a technique to optimize building energy usage. MPC uses mathematical models to predict a building's thermal response and compute optimal control inputs over a future horizon to minimize energy costs while maintaining comfort levels. The document provides an example thermal model of a building and shows experimental data matching the model's predictions. It also discusses using low-cost IoT sensors and controllers throughout a building to monitor conditions and implement optimized control strategies from an MPC system.
The document summarizes research comparing the Particle Swarm Optimization (PSO) and Differential Evolution (DE) algorithms for optimizing power consumption using smart energy meter data. Both algorithms were implemented in MATLAB and tested on 15 days of meter data from a university lab in India. PSO achieved an 11.5% reduction in power consumption while DE achieved a 9.4% reduction. PSO outperformed DE for this application, showing it is an effective technique for optimizing energy use and reducing electricity costs for consumers. Future work could integrate the models with real smart meters and controllers to achieve automated scheduling and greater savings.
Building Institutional Capacity in Thailand to Design and Implement Climate P...UNDP Climate
23-25 November 2016, Thailand - A centerpiece of the Integrating Agriculture in National Adaptation Plans Programme (NAP-Ag) in Thailand is its support to develop a new five-year Strategy on Climate Change in Agriculture (2017-2021). This is spearheaded by the Ministry of Agriculture and Cooperatives (MOAC) and its Office of Agriculture Economics (OAE). The strategy was unveiled after a series of meetings by a Technical Working Group at a three-day workshop held on 23-25 November 2016 in Bangkok, organized by UNDP. Over 60 participants from each MOAC line department and 10 participants from academia and civil society were briefed by the Office of the Natural Resources and Environmental Policy and Planning (ONEP) and GIZ on the status of the National Adaption Plan (NAP) and learned how NAP-Ag programme efforts could support a broader NAP process and align with the Sector Plan. The new strategy focuses on improving evidence and data for informing policy choices, building the capacity of farmers and agri-businesses to adapt, promoting low-carbon development and productivity growth in the sector, and building institutional and managerial capacities to cope with climate change impacts.
The document discusses Boolean algebra and its applications in combinational logic circuit design. It covers topics like Boolean expressions, standard forms (sum of products and product of sums), converting between forms, truth tables, and determining logic expressions from truth tables. Standard forms allow for simplification using techniques like Karnaugh maps. Boolean algebra is used to analyze and design basic combinational circuits like encoders, decoders, and adders.
For some management programming problems, multiple objectives to be optimized rather than a single
objective, and objectives can be expressed with ratio equations such as return/investment, operating
profit/net-sales, profit/manufacturing cost, etc. In this paper, we proposed the transformation
characteristics to solve the multi objective linear fractional programming (MOLFP) problems. If a MOLFP
problem with both the numerators and the denominators of the objectives are linear functions and some
technical linear restrictions are satisfied, then it is defined as a multi objective linear fractional
programming problem MOLFPP in this research. The transformation characteristics are illustrated and the
solution procedure and numerical example are presented.
This document summarizes a presentation on learning to hash for large scale image retrieval. It discusses locality sensitive hashing (LSH), which maps similar data points to the same "buckets" with high probability to enable efficient nearest neighbor search. The presentation evaluates different approaches to learn the quantization thresholds and hashing hypersurfaces in LSH, including learning multiple thresholds to better capture data distributions. It proposes a semi-supervised objective function to jointly optimize the thresholds based on supervised neighborhood information and unsupervised projections between points. The goal is to improve retrieval effectiveness by making the hashing more influenced by data distributions.
For some management programming problems, multiple objectives to be optimized rather than a single objective, and objectives can be expressed with ratio equations such as return/investment, operating profit/net-sales, profit/manufacturing cost, etc. In this paper, we proposed the transformation characteristics to solve the multi objective linear fractional programming (MOLFP) problems. If a MOLFP problem with both the numerators and the denominators of the objectives are linear functions and some technical linear restrictions are satisfied, then it is defined as a multi objective linear fractional programming problem MOLFPP in this research. The transformation characteristics are illustrated and the solution procedure and numerical example are presented.
APPLYING TRANSFORMATION CHARACTERISTICS TO SOLVE THE MULTI OBJECTIVE LINEAR F...ijcsit
For some management programming problems, multiple objectives to be optimized rather than a single objective, and objectives can be expressed with ratio equations such as return/investment, operating
profit/net-sales, profit/manufacturing cost, etc. In this paper, we proposed the transformation characteristics to solve the multi objective linear fractional programming (MOLFP) problems. If a MOLFP problem with both the numerators and the denominators of the objectives are linear functions and some
technical linear restrictions are satisfied, then it is defined as a multi objective linear fractional programming problem MOLFPP in this research. The transformation characteristics are illustrated and the solution procedure and numerical example are presented.
APPLYING TRANSFORMATION CHARACTERISTICS TO SOLVE THE MULTI OBJECTIVE LINEAR F...
Thesis_Proposal2
1. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Polynomial and Moment Conic Optimizations:
Theory and Applications
Mohammad M. Ranjbar,
Advisor: Farid Alizadeh
Rutgers, The State University Of New Jersey
Dec/01/2016
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
2. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Outline
1 Definitions
2 Applications and Motivation
3 SDP Representation of NPCO and MCO
4 Drawbacks and Remedies
5 Chebyshev Change of Basis
6 Non-Symmetric Interior Point Method
7 Homogeneous Self-Dual Embedding
8 Numerical Results
9 Discretization
10 Conclusion and Future Works
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
3. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Non-Negative Polynomial Cone:
Definition: Standard Basis
un
t = (1, t, t2
, . . . , tn
)T
,
Definition: Non-Negative (Positive) Polynomial Cone in standard
basis on interval [a, b]
Pn+
[a,b] = {p ∈ Rn+1
: p, un
t 0, ∀t ∈ [a, b]}.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
4. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Non-Negative Polynomial Cone vs Semidefinite Positive
Cone:
Non-Negative Polynomial Cone Semidefinite Positive Cone
{p :
n
i=0
pi ti ≥ 0, ∀t ∈ [a, b]} {X : yT Xy ≥ 0, ∀y}
p =
n
i=1
λi conv(pi , pi ) X =
n
i=1
λi qi qT
i
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
5. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Non-Negative Polynomial Conic Optimization:
Non-Negative Polynomial Conic Optimization (NPCO):
(NPCO) min cT
1 s1 + · · · + cT
k sk
s.t. A1s1 + · · · + Aksk = b,
si
P
n+
i
[a,b]
0, i = 1, ..., k,
where ci ∈ Rni +1, Ai ∈ Rm×ni +1, b ∈ Rm and si ∈ Rni +1.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
6. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Moment Cone
Definition: Moment Cone in standard basis on interval [a, b]
Mn
[a,b] = cl(cone{un
t | ∀t ∈ [a, b]}).
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
7. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Moment Conic Optimization:
Moment Conic Optimization (MCO):
(MCO) min cT
1 x1 + · · · + cT
k xk
s.t. A1x1 + · · · + Akxk = b,
xi M
ni
[a,b]
0, i = 1, ..., k,
where ci ∈ Rni +1, Ai ∈ Rm×ni +1, b ∈ Rm and xi ∈ Rni +1.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
8. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Notations: For simplicity we will show
c = (c1; · · · ; ck),
A = [A1, · · · , Ak],
x = (x1; · · · ; xk),
s = (s1; · · · ; sk),
M[a,b] = Mn1
[a,b] ⊗ · · · ⊗ Mnk
[a,b],
P+
[a,b] = P
n+
1
[a,b] ⊗ · · · ⊗ P
n+
k
[a,b].
Conic Optimization:
min cT
x
s.t. A x = b,
x K 0,
where K is either M[a,b] or P+
[a,b].
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
9. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Definition: Hankel Operator:
H(.) : R2n+1
→ R(n+1)×(n+1)
H(x) =
x0 x1 . . . xn
x1 ...
...
xn+1
... ...
... ...
xn xn+1 . . . x2n
Definition: Dehankel Operator:
H∗
(.) : R(n+1)×(n+1)
→ R2n+1
[H∗
(Y )]2n
i=0 =
y0,0
y0,1 + y1,0
...
yn,n
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
11. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Applications:
Applications of NPCO and MCO
1 Non-Negative Polynomial Conic Optimization
Statistics: Nonparametric Estimation With Shape Constraint
Finance: Nonparametric Cost Function, Production Function,
Utility Function and Option Pricing Under Shape Restriction
Approximation
Time Variant Network Flows: Maximum Flows Problem
with Time Variant Capacities, Min-cost Flows Problem with
Time Variant Costs
Engineering: Envelope Signal
2 Moment Cone Optimization
Statistics: Different orders of moment of a distribution with
desired properties.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
12. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Statistics:
Consider a set of points in two dimension, (xi , yi ) for i = 1, ..., p.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.2
0.4
0.6
0.8
1
1.2
Data Points
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
13. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Statistics:
What is the Best Fitting Polynomial with the given degree?
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
data1
Best Fitting Poly(deg=40)
Best Fitting Poly (deg=100)
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
14. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Statistics:
Now what if the Best Fitting Polynomial has to be Positive
Polynomial?
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
-1
-0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
Best Fitting Poly(deg=40)
15. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Statistics:
Now what if the Best Fitting Polynomial has to be Positive
Polynomial?
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
-1
-0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
Best Fitting Poly(deg=40)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
-1
-0.5
0
0.5
1
1.5
2
2.5
3
3.5
4
Data Points
Best Fitting Poly(deg=40)
Best Fitting Positive Poly(deg=40)
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
16. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Statistics:
What is the Best Envelope Convex Polynomial with the given
degree Below all data points?
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
2
4
6
8
10
12
Data Points
Best Envelope Convex Poly(deg=60)
How about the Best Fitting Convex Polynomial with the given
degree?
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
17. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Statistics: Nonparametric Estimation under Shape
Constraints I
In general, all of these Nonparametric Estimation under shape
constraints can be formulated as:
min
s
p
i=1
(s(ti ) − yi )2
s.t s(t) ≥ 0, ∀ t ∈ [a, b],
s (t) ≥ 0, ∀ t ∈ [a, b], ←− increasing or decreasing
s (t) ≥ 0, ∀t ∈ [a, b], ←− convexity or concavity
where s and s are the first and the second derivative of s(t)
respectively.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
18. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Statistics: Nonparametric Estimation under Shape
Constraints II
This problem can be cast as a NPCO and SOCP.
min z
s.t
z
Vs − y socp 0,
s Pn+
[a,b]
0,
s
P
(n−1)+
[a,b]
0,
s
P
(n−2)+
[a,b]
0,
where V is the Vandermonde matrix (Vi,j = tj
i ).
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
19. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Finance:
Usually, in finance and economy:
Cost function must be Increasing and Convex
Production function must be Increasing and Concave
Utility function must be Increasing and Concave
Call option pricing function must be Decreasing and Convex
All of these cases can be formulated as:
min T(s)
s.t. A s = b,
s(t) Pn+
[a,b]
0,
s (t) Pn−1+
[a,b]
0,
s (t) Pn−2+
[a,b]
0.
where T(.) is a linear operator.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
20. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Approximation or Envelope Polynomial:
What is the Best Polynomial (coefficients of the polynomial) that
approximates all other polynomials and is below (or above) them?
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
2.5
3
3.5
4
4.5
5
5.5
21. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Approximation or Envelope Polynomial:
What is the Best Polynomial (coefficients of the polynomial) that
approximates all other polynomials and is below (or above) them?
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
2.5
3
3.5
4
4.5
5
5.5
-0.7 -0.65 -0.6 -0.55 -0.5 -0.45 -0.4 -0.35 -0.3 -0.25
3
3.1
3.2
3.3
3.4
3.5
3.6
Approx Poly(deg=80)
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
22. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Increase Degree of Polynomial
As the degree of polynomial increases, we get better optimal
solution.
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
2.5
3
3.5
4
4.5
5
Degree, n=20,
Degree, n=200,
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
23. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Approximation:
This class of problem can be formulated as:
max
b
a
x(t)dt = u[a,b], x
s.t. x(t) Pn+
[a,b]
pi (t), i = 1, ..., k,
x(t) Pn+
[a,b]
0.
What if we have arbitrary functions instead of polynomials?
max
b
a
x(t)dt = u[a,b], x
s.t. x(t) ≤ fi (t), i = 1, ..., k, ∀t ∈ [a, b],
x(t) Pn+
[a,b]
0.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
24. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Time Dependent Network Flows: I
Conventional Network Flows problems: Network flows with
Constant Input Data.
For example, in Maximum Flows problem “One seeks to push
maximum flows from source to sink with respect to the
Constant Capacities of the edges”.
Time Dependent Network Flows Problems: Network flows with
Time-Variant Input Data.
For example in Time Variant Maximum Flows, “One seeks
to push maximum flows from source to sink with respect to
the Time Dependent Capacities of the edges”.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
25. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Time Dependent Network Flows: II
1Source
2
3
i
j
k
r
m Sink
p1,2(t)
p1,3(t)
pk,m(t)
pr,m(t)
Time Dependent Maximum Flows problem can be formulated as:
max
i|(s,i)∈E
b
a
Xs,i (t) dt
s.t.
j|(j,i)∈E
Xj,i (t) −
j|(i,j)∈E
Xi,j (t) = 0, ∀i, ∀t ∈ [a, b],
Xi,j (t) Pn+
[a,b]
Pi,j (t), (i, j) ∈ E, ∀t ∈ [a, b],
Xi,j (t) Pn+
[a,b]
0, (i, j) ∈ E, ∀t ∈ [a, b].
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
26. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Time Dependent Network Flows: III
How about when the input data are arbitrary functions?
max
i|(s,i)∈E
b
a
Xs,i (t) dt
s.t.
j|(j,i)∈E
Xj,i (t) −
j|(i,j)∈E
Xi,j (t) = 0, ∀i, ∀t ∈ [a, b],
Xi,j (t) Pn+
[a,b]
fi,j (t), (i, j) ∈ E, ∀t ∈ [a, b],
Xi,j (t) Pn+
[a,b]
0, (i, j) ∈ E, ∀t ∈ [a, b].
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
27. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Engineering:
Envelope Signal: In Electrical Engineering and Signal Processing
we need to have the Envelope Signal of given signal or signals
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
28. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
NPCO+SOCP+LP:
All of preceding problems can be formulated as Non-Negative
Polynomial Conic Optimization with the combination of the
other well-known cones such as LP, SOCP.
min cL
xL
+
i
cS
j xS
j +
l
cP+
l xP+
l
s.t. AL
xL
+
i
AS
j xS
j +
l
AP+
l xP+
l = b,
xL
≥ 0, xS
j SOCP 0, xP+
l Pn+
[a,b]
0.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
29. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
SDP Representation of NPC:
Theorem (Nesterov 97)
a) p ∈ P2n+
R if and only if ∃Y 0 s.t. p = H∗(Y ),
b) p ∈ P2n+
[a,b] if and only if
∃Y1, Y2 0, s.t. p = H∗
(Y1) + AH∗
(Y2),
c) p ∈ P
(2n+1)+
[a,b] if and only if
∃Y1, Y2 0, s.t. p = BH∗
(Y1) + CH∗
(Y2),
where A, B, C are three matrices dependent on the interval shift
the vectors H∗(Y1), H∗(Y2).
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
30. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
SDP Representation of NPCO:
Using this theorem, NPCO can be cast as SDP. For instance, when
K = P2n+
R :
min cT
x =⇒ min H(c)•Y
s.t. Ax = b, s.t. H(ai )•Y = bi , i = 1, .., k,
x P2n+
R
0, Y 0,
where s = H∗(Y ).
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
31. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
SDP Representation of MC:
Theorem (Karlin-Studden 66, Nesterov 97)
a) x ∈ M2n
R ⇔ H(x) 0.
b) x ∈ M2n
[a,b] ⇔
H(x) 0,
H((−abxi + (b + a)xi+1 − xi+2)2n−2
i=0 ) 0
c) x ∈ M2n+1
[a,b] ⇔
H((−axi + xi+1)2n
i=0) 0,
H(bxi − xi+1)2n
i=0) 0.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
32. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
SDP Representation of MCO:
Using this theorem, MCO can be cast as SDP. For instance, when
K = M2n
R :
min cT
x =⇒ min cT
x
s.t. Ax = b, s.t. Ax = bi , i = 1, .., k,
x M2n
R
0, H(x) 0.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
33. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Drawbacks and Remedies:
However, casting these conic optimizations as SDP have two major
drawbacks:
(a) It is extremely ill-conditioned.
(b) It quadratically increases the dimension of the problem.
We have proposed two remedies for these drawbacks, this is:
(a) Orthogonal change of basis.
(b) Non-symmetric interior point methods.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
34. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Chebyshev Change of Basis
Chebyshev Basis
T0(t) = 1,
T1(t) = t,
Tn+1(t) = 2tTn(t) − Tn−1(t).
Definition: Moment and Non-negative polynomial cone in
Chebyshev basis on [−1, 1] are defined as:
Mn
Ch = cl(Cone{(T0(t), T1(t), ..., Tn(t))T
| ∀t ∈ [−1, 1]}),
Pn+
Ch = {s ∈ Rn+1
|
n
j=0
sj Tj (t) 0, ∀t ∈ [−1, 1]}.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
35. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Representation of Moment Cone in Chebyshev Basis
Theorem (Papp 2011, Ranjbar 2016)
a) x ∈ M2n
Ch over R if and only if H(x) + T(x) 0.
b) x ∈ M2n
Ch over [−1, 1] if and only if
H(x) + T(x) 0,
H([xi − 1
2xi+2]2n−2
i=0 − 1
2[x2; x1; [xi ]2n−4
i=0 ])
+T([xi − 1
2xi+2]n−1
i=0 ] − 1
2[x2; x1; [xi ]n−3
i=0 ]) 0.
c) x ∈ M2n+1
Ch over [−1, 1] if and only if
H([xi + xi+1]2n
i=0) + T([xi + xi+1]n
i=0) 0,
H([xi − xi+1]2n
i=0) + T([xi − xi+1]n
i=0) 0.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
36. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Change of Basis Result
By using Chebyshev change of basis, SDP representation of MCO
and NPCO of much higher sizes can be solved.
Our experiences showed problem with block dimension of 3000 can
be solved.
But casting MCO or NPCO as SDP requires squaring the number of
decision variables and increases the number of constraints.
For example, if NPCO has 103
blocks and the degree of each block
is of size 103
, then it needs O(109
) floating point memory storage
to only save the hessian.
The situation is even worse in term of running time, for example, a
problem of 200 blocks and each block of dimension 1200 needs
more than 4 days to be solved by a server with 32 GB RAM and
3GHZ processor.
The so-called non-symmetric interior point method will be the remedy.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
37. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Primal-Dual Path-Following Algorithm (IPM): I
To use Primal-Dual Path-Following IPM, we need to have
followings:
An efficiently computable LHSCB function for the cones.
The gradient and Hessian of barrier function should be
efficiently computable
The pair of primal-dual problem.
An efficient algorithm that in the case of feasible problem
converges to optimal solution and in the case of infeasibility
detects it.
The algorithm should not require any information about the
dual barrier function.
Following has been offered
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
38. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Theorem (3) gives a suitable barrier function for moment cone
in Chebyshev basis.
It has been shown that the gradient and the hessian of the
barrier are efficiently computable.
Karlin-Studden solved the primal-dual setting.
Homogeneous Self-Dual embedding (HSD) is used which
either converges to optimal solution or detects infeasibility.
Skajaa-Ye (2015) proposed a HSD primal-dual
predictor-corrector IPM which does not need any information
of the dual barrier.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
39. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Barrier Function for Moment Cone in Chebyshev Basis
logarithmic homogeneous self-concordant barrier (LHSCB)
function for moment cone in Chebyshev basis are defined:
F(x) = − ln det((H + T)(x)), ∀x ∈ M2n
ChR
,
F(x) = − ln det((H + T)(x))−
ln det((H + T)([xi −
1
2
xi+2]2n−2
i=0 −
1
2
[x2; x1; [xi ]2n−4
i=0 ])),
∀x ∈ M2n
Ch,
F(x) = − ln det((H + T)([xi + xi+1]2n
i=0))−
ln det((H + T)([xi − xi+1]2n
i=0)), ∀x ∈ M2n+1
Ch ,
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
40. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Gradient of Barrier Function
x (F(x)) = −(H∗
+ T∗
)((H + T)−1
(x)), ∀x ∈ int(M2n
ChR
),
x (F(x)) = −(H∗
+ T∗
)((H + T)−1
(x))−
(H∗
+ T∗
)((H + T)−1
([xi −
1
2
xi+2]2n−2
i=0 −
1
2
[x2; x1; [xi ]2n−4
i=0 ])),
∀x ∈ int(M2n
Ch),
x (F(x)) = −(H∗
+ T∗
)((H + T)−1
([xi + xi+1]2n
i=0))−
(H∗
+ T∗
)((H + T)−1
([xi − xi+1]2n
i=0)), ∀x ∈ int(M2n+1
Ch ).
Let us show gradient by gx .
gx ∈ R2n+1
in contrast to (n + 1) × (n + 1) matrix in SDP
formulation.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
41. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Hessian of Barrier Function
2
x (F(x)) =conv2T((H + T)−1
(x)), ∀x ∈ int(M2n
ChR
),
2
x (F(x)) =conv2T((H + T)−1
(x))+
conve2T((H + T)−1
([xi −
1
2
xi+2]2n−2
i=0 −
1
2
[x2; x1; [xi ]2n−4
i=0 ])),
∀x ∈ int(M2n
Ch),
2
x (F(x)) =conv2T((H + T)−1
([xi + xi+1]2n
i=0))+
conv2T((H + T)−1
([xi − xi+1]2n
i=0)), ∀x ∈ int(M2n+1
Ch ).
conv2T(.) is the convolution of two bivariate polynomials in
Chebyshev basis.
Let us show hessian by Hx .
Hessian is a (2n + 1) × (2n + 1) matrix in contrast to
(n + 1)2
× (n + 1)2
in SDP formulation.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
42. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Computation of Gradient and Hessian I
Solving a linear systems of equations with (hankel+toeplitz)
coefficient matrices can be done in O(rn2
) and O(rn) memory.
In implementation, we have used “drsolve” package written by Arico
and Rodriguez [2] to computing the inverse of (hankel+toeplitz)
matrix.
The convolution of two polynomials of degree n in Chebyshev basis
can be computed, in O((12n + 3) log 2n − 14n + 15) via Discrete
Cosine Transform (DCT) .
The convolution of two bivariate polynomial in Chebyshev basis of
degree n × n can be done by convolution of two single variable
polynomials of degree n2
.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
43. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Computation of Gradient and Hessian II
In implementation, we have used a version of DCT method based
on the work of Baszenski and Tasche [3]. This is
conv2T(A, B) =
4
N2
CN ((CN ACT
N ) ◦ (CN BCT
N ))CT
N
where A and B ∈ R(N+1)×(N+1)
, ◦ is the Hadamard product, and
CN ∈ R(N+1)×(N+1)
is
CN = (εN,j cos
ijπ
N
)N
i,j=0,
with εN,0 = 1
2 and εN,j = 1, j = 1, ..., N − 1.
The computation work of this method is O(n2
log n).
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
44. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Barrier Function for Non-Negative Polynomial Cone I
We have investigated two LHSCB function for non-negative
polynomial cone.
Faybusovich Barrier Function [4]: Faybusovich has
proposed following LHSCB function for P2n+
R
F∗
(s) =
1
2
ln det D(s),
Di,j (s) =
1
−1
ui (t)uj (t) − uj (t)ui (t)
s2(t)
dt, i, j = 0, ..., n.
This barrier is not efficiently and practically computable.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
45. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Barrier Function for Non-Negative Polynomial Cone II
Convex conjugate of the moment barrier function: Using
the definition of convex conjugate function the LHSCB
function of s ∈ P2n+
R can be defined as:
F∗
(s) = − inf
x∈int(M2n
R )
{ x, s − ln det H(x)},
This is a convex optimization which does not have any closed
form solution.
Therefore, even finding the value of the barrier function of
non-negative polynomial cone, is a challenge.
Hence, we need an algorithm which does not need the information
of the dual barrier function.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
46. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Duality:
Theorem (Karlin-Studden 66)
Moment cone and Non-Negative Polynomial cone are dual.
(M[a,b])∗
= P+
[a,b].
Using this duality, MCO and NPCO can be put in a non-symmetric
primal-dual setting.
(MCO) min cT
x
dual
⇐=⇒ (NPCO)max bT
y
s.t. A x = b, s.t. AT
y+s = c
x M[a,b]
0, s P+
[a,b]
0.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
47. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Homogeneous Self-Dual (HSD) Embedding
min 0
s.t. Ax−bτ = 0
AT
y +cτ−s = 0
bT
y−cT
x −κ= 0
(x, τ) ∈ M[a,b] × R+
, y ∈ Rm
, (s, κ) ∈ P+
[a,b] × R+
Path-following algorithm on HSD model:
does not need initial feasible solution.
does not increase the size of the problem.
In the case of feasibility gives the optimal solution.
In the case of infeasibility detects the source.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
48. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
HSD Solution
Theorem (Ye-Todd-Mizuno)
Assume (x∗, τ∗, y∗, s∗, κ∗) solves HSD model. Then
a) (x∗, τ∗, s∗, κ∗) is complementary, this is, x∗T s∗ + τ∗κ∗ = 0.
b) If τ∗ > 0 then (x∗, y∗, s∗)/τ∗ is optimal for HSD.
c) If κ∗ > 0 then either primal or dual is infeasible.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
49. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
More Notations:
¯x =
x
τ
¯s =
s
κ
¯F(¯x) = F(x) − log τ, ¯F∗
(¯s) = F∗
(s) − log κ,
¯K = K × R+
, ¯K∗
= K+
× R+
,
¯ν = ν + 1
µ(z) = ¯xT
¯s/¯v,
ψ(¯x, ¯s, µ(z)) = ¯s − µ(z)g¯x .
where z = (¯x, y, ¯s)
G =
0 A −b
−AT
0 c
bT
−cT
0
,
rp
rd
rc
= G
y
¯x
−
0
¯s
.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
50. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Path Following IPM:
The idea of Path Following IPM is to Approximately Follow
the Central Path and Stay Close to it.
Central Path
Approximately
Close to central path
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
51. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Central Path:
Central Path
Central Path is defined as the unique
solution of the parametrized equations.
G
y
¯x
−
0
¯s
= γ
rp
rd
rc
(CP)
¯s − γµg¯x = 0,
where 0 ≤ γ ≤ 1.
z=(x,,y,s,𝜿)
Central Path
Opt Sol
Feasibility
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
52. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Approximately:
Approximately
Approximately means to find the solution
of central path by using First Order
Newton Method, this is,
G
dy
d ¯x
−
0
d¯s
= (1 − γ)
rp
rd
rc
(Dir)
ds + µHx dx = −s − γµgx
τdκ + κdτ = −τκ + γµ.
z=(x,𝞃,y,s,𝜿)
Central Path
Opt Sol
Feasibility
dz=(dx,d𝞃,dy,ds,d𝜿)
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
53. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Close to Central Path:
Hessian Neighborhood:
N(η) = {z ∈ F :
s − µ(z)gx
κ + µ(z)/τ
∗
(x,τ)
≤ ηµ(z)}
where v ∗
x = H
−1/2
x v .
In implementation computing Hessian is expensive instead infinite
norm will be used which is related to hessian.
|| F(x)||−1
∞ (s + µ F(x))
τκ − µ ∞
≤ ηµ(z)
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
54. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Predictor-Corrector:
The idea of Predictor-Corrector IPM is to start from an initial
point near the central path and then alternate between the
Prediction Phase and the Correction Phase.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
55. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Prediction Phase (γ = 0)
Find direction that Reduces the
Residuals and Duality gap and is not
Concerned about the Centrality
G
dyp
d ¯xp
−
0
d¯sp
=
rp
rd
rc
(PD)
dsp + µHx dxp = −s
τdκp + κdτp = −τκ.
z=(x,𝞃,y,s,𝜿)
Larg 𝓝
Small 𝓝
zp=(xp,𝞃p,yp,sp,𝜿p) dzp=(dxp,d𝞃p,dyp,dsp,d𝜿p)
Central Path
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
56. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Correction Phase (γ = 1)
Find direction that Reduce the
Centrality but Keeps the Residuals
and the Duality Gap Unchanged
G
dyc
d ¯xc
−
0
d¯sc
=
0
0
0
(CD)
dsc + µHxp dxc = −sp − µpgxp
τpdκc + κpdτc = −τpκp + µp.
z=(x,𝞃,y,s,𝜿)
Larg 𝓝
Small 𝓝
zp=(xp,𝞃p,yp,sp,𝜿p) dzp=(dxp,d𝞃p,dyp,dsp,d𝜿p)
Central Path
Zc=(xc,𝞃c,yc,sc,𝜿c)
dzp=(dxp,d𝞃p,dyp,dsp,d𝜿p)
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
57. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Next Prediction and Correction Iteration:
z=(x,𝞃,y,s,𝜿)
Larg 𝓝
Small 𝓝
zp=(xp,𝞃p,yp,sp,𝜿p) dzp=(dxp,d𝞃p,dyp,dsp,d𝜿p)
Central Path
Zc=(xc,𝞃c,yc,sc,𝜿c)
dzp=(dxp,d𝞃p,dyp,dsp,d𝜿p)
58. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Next Prediction and Correction Iteration:
z=(x,𝞃,y,s,𝜿)
Larg 𝓝
Small 𝓝
zp=(xp,𝞃p,yp,sp,𝜿p) dzp=(dxp,d𝞃p,dyp,dsp,d𝜿p)
Central Path
Zc=(xc,𝞃c,yc,sc,𝜿c)
dzp=(dxp,d𝞃p,dyp,dsp,d𝜿p)
Larg 𝓝
Small 𝓝
zp=(xp,𝞃p,yp,sp,𝜿p) dzp=(dxp,d𝞃p,dyp,dsp,d𝜿p)
Central Path
Zc=(xc,𝞃c,yc,sc,𝜿c)
dzp=(dxp,d𝞃p,dyp,dsp,d𝜿p)
z=(x,𝞃,y,s,𝜿)
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
59. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Convergency:
Skajaa-Ye [7] showed if η ≤ 1/6 then by taking a fixed
step-length, αp = Ω(1/
√
ν), z+
p ∈ N(2η).
In practical implementation a larger step-length can be taken.
Skajaa-Ye (see lemma 6 in [7]) showed if z+
p ∈ N(2η) then by
applying the correction phase at most two times z+
c ∈ N(η).
They also showed that by using the specified step-length the
algorithm terminates with a -solution in no more than
O( (ν) log(1/ )).
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
60. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Symmetric vs Non-Symmetric HSD P-D P-C for LP
Non-symmetry comes from duality gap
LP has both formats
Symmetric
Prediction
Sdxp + Xdsp = −Xs
Correction
Spdxc + Xpdsc = −Xpsp + µpe
Non-Symmetric
Prediction
µX−2
dxp + dsp = −s
Correction
µpX−2
p dxc + dsc = −sp + µpX−1
p e
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
61. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Symmetric vs Non-Symmetric HSD P-D P-C for LP
Two experiments have been done:
1- Number of iterations in correction phase are fixed to 1 in
symmetric and to 2 in non-symmetric algorithm.
Algorithm m n Iter. Avg. Corr.
Symmetric 20 2e2 11 1
Non-Symmetric 20 2e2 14 2
Symmetric 200 2e4 19 1
Non-Symmetric 200 2e4 19 2
Symmetric 300 2e5 48 1
Non-Symmetric 300 2e5 45 2
Table: Symmetric vs Non-Symmetric HSD P-D P-C IPM for LP
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
62. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Symmetric vs Non-Symmetric HSD P-D P-C for LP
2- Number of iterations in correction phase is decided by
algorithm.
Algorithm m n Iter. Avg. Corr.
Symmetric 20 2e2 11 1.54
Non-Symmetric 20 2e2 11 1.81
Symmetric 200 2e4 20 1.75
Non-Symmetric 200 2e4 22 1.82
Symmetric 300 2e5 44 2.34
Non-Symmetric 300 2e5 45 1.80
Table: Symmetric vs Non-Symmetric HSD P-D P-C IPM for LP
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
63. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Non-Symmetric HSD P-D P-C for MCO-NPCO
To show the numerical results of non-symmetric HSD P-D P-C for
MCO-NPCO, we have considered the approximation problem.
Algorithm blk m n Iter. Avg. Corr.
MCO-NPCO 3 41 123 16 1.5
SDP 3 121 1362 12 1
MCO-NPCO 3 201 603 20 1.5
SDP 3 601 30802 22 1
MCO-NPCO 3 401 1203 23 1.65
SDP 3 1201 1216022 20 1
Table: Symmetric vs Non-Symmetric HSD P-D P-C IPM for MCO-NPCO
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
64. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Discretization of MCO and NPCO I
Considered a refined grid of points in [a, b].
S[a b] := {a ≤ t1 < t2 < ... < tp ≤ b}
Therefore S[a,b] is going to be an approximation of [a, b] when
p is large enough.
Then, the discrete moment and non-negative polynomial
cones over S[a,b] are defined:
Mn
S[a,b]
= cl(Cone{un
tj
, j = 1, .., p})
Pn+
S[a,b]
= {s ∈ Rn+1
: s, un
tj
≥ 0, j = 1, ..., p}
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
65. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Discretization of MCO and NPCO II
It has been shown that
Mn
S[a,b]
⊂ Mn
[a,b],
Pn+
[a,b] ⊂ Pn+
S[a,b]
.
As p −→ ∞,
Mn
S[a,b]
−→ Mn
[a,b],
Pn+
S[a,b]
−→ Pn+
[a,b].
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
66. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Discretization of MCO and NPCO III
Using this approximation, discrete MCO and NPCO can be
reformulated as a linear programming. This is:
min cT
x min cT
s
s.t. Ax = b, s.t. As = b,
x =
j=1
p
αj un
tj
, un
tj
,s ≥ 0, j = 1, ..., p,
αj ≥ 0, j = 1, ..., p,
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
67. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
LP vs SDP vs Non-Symmetric for NPCO:
Consider NPCO:
min cT
s
s.t. A s = b (NPCO)
s Pn+
[a,b]
0.
The number of variables and constraints in LP, SDP and
NPCO formulations:
Problem # of Const. # of Var.
LP p n
SDP kn O(kn2)
Non-Symmetric n kn
Table: Number of Const. & Var. in NPCO with k Blocks
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
68. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Numerical Results: SDP vs Discrete (LP)
Approximation Problem:
max
b
a
s(t)dt = e[a,b], s
s.t. s(t) Pn+
[a,b]
pi (t), i = 1, ..., m,
s(t) Pn+
[a,b]
0.
Time and Feasibility of SDP and LP:
Form m n p time(Sec.) feasibility optval
SDP 26 100 −− 3.66E + 01 Yes 6.0000
LP 26 100 102
3.33E + 00 infeas. prob. −−
LP 26 100 103
3.49E + 00 infeas. sol. 6.0026
LP 26 100 104
5.84E + 00 Yes 6.0015
Table: NPCO: SDP vs LP as p Increases
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
69. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Figure: SDP vs Discrete (LP)
Time and Feasibility:
0.97 0.975 0.98 0.985 0.99 0.995 1
2.985
2.99
2.995
3
3.005
3.01
3.015
3.02
max flow LP 26-100-1000
max flow LP 26-100-10000
max flow SDP 26-100
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
70. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Conclusion: I
Non-negative polynomial and Moment conic optimizations
have been defined.
Different applications in statistics, finance, network flows etc
have been mentioned.
SDP representation of NPCO and MCO have been shown.
The drawbacks of SDP have been mentioned.
The remedies have been proposed.
Orthogonal change of basis has been shown.
A tailor-made non-symmetric HSD P-D P-C based on
Skajaa-Ye algorithm has been proposed.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
71. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Conclusion: II
Numerical results for non-symmetric vs symmetric HSD for LP
have been shown.
Numerical results of implementation of non-symmetric HSD
P-D P-C have been shown.
Discretization of NPCO and MCO have been shown.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
72. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Future: I
Many real world problems can be formulated as an optimization
problem over the combination of different cones (LP,SOCP, NPCO)
Write a software package that can solve following problem
min cL
xL
+
i
cS
j xS
j +
l
cP+
l xP+
l
s.t. AL
xL
+
i
AS
j xS
j +
l
AP+
l xP+
l = b,
xL
≥ 0, xS
j SOCP 0, xP+
l Pn+
[a,b]
0.
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
73. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Future: II
Write a software package that can solve following problem
min f ( xL
, xS
j , xP+
l )
s.t. AL
xL
+
i
AS
j xS
j +
l
AP+
l xP+
l = b,
xL
≥ 0, xS
j SOCP 0, xP+
l Pn+
[a,b]
0.
where f (x) is a well-known convex function, like entropy
f (x) =
n
i=1
xi log xi .
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
74. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Question!?
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
75. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Akle Serrano S., Algorithms for unsymmetric cone
optimization and an implementation for problems with the
exponential cone, PhD Thesis, Stanford university (2015)
Arico, A., Rodriguez, G., A fast solver for linear systems with
displacement structure, NUMERICAL ALGORITHMS, 55,
529-556 (2010)
Baszenski, G., Tasche, M., Fast Polynomial Multiplication and
Convolution Related to the Discrete Cosine Transform, Linear
Algebra and Its Applications, 252, 1-25 (1997)
Faybusovich, L., Self-Concordant Barriers for Cones Generated
by Chebyshev Systems, SIAM J. Optim., 12(3), 770-781
(2002)
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications
76. Definitions Applications and Motivation SDP Representation of NPCO and MCO Drawbacks and Remedies Chebyshev Change of
Nesterov, Y., Squared Functional Systems and Optimization
Problems, High Performance Optimization, 33, 405-440,
Kluwer Academic Press, (2000)
Nesterov, Y. E., Towards Nonsymmetric Conic Optimization.
Optim. Method. Softw., 27, 893-917 (2012)
Skajaa A., Ye Y., A homogeneous interior-point algorithm for
nonsymmetric convex conic optimization, Math. Program.,
Ser. A, 150, 391-422 (2015)
Mohammad M. Ranjbar, Advisor: Farid Alizadeh Rutgers
Polynomial and Moment Conic Optimizations: Theory and Applications