[Home]
[
News]
[ Events]
[ People]
[ Research]
[ Education]
[Visitor Info]
[UCSD Only]
[Admin]
On the pervasiveness of difference-convexity in optimization and statistics
Jong-Shi Pang
University of Southern California
Abstract:
With the increasing interest in applying the methodology of difference-of-convex (dc) optimization to diverse problems in engineering and statistics, we show that many well-known functions arising therein can be represented as the difference of two convex functions. These include a univariate folded concave function commonly employed in statistical learning, the value function of a copositive recourse function in two-stage stochastic programming, and many composite statistical functions in risk analysis, such as the value-at-risk (VaR), conditional value-at-risk (CVaR), expectation-based, VaR-based, and CVaR-based random deviation functionals. We also discuss decomposition methods for computing directional stationary points of a class of nonsmooth, nonconvex dc programs that combined the Gauss-Seidel idea, the alternating direction method of multipliers, and a special technique to handle the negative of a pointwise max function.
Thursday, April 20, 2017
4:00PM AP&M 6402
News]
[ Events]
[ People]
[ Research]
[ Education]
[Visitor Info]
[UCSD Only]
[Admin]
Directors:
Randolph E. Bank
Philip E. Gill
Michael Holst
Administrative Contact:
[ Click to Send Email ]
Randolph E. Bank
Philip E. Gill
Michael Holst
Administrative Contact:
[ Click to Send Email ]
Jong-Shi Pang
University of Southern California
Abstract:
With the increasing interest in applying the methodology of difference-of-convex (dc) optimization to diverse problems in engineering and statistics, we show that many well-known functions arising therein can be represented as the difference of two convex functions. These include a univariate folded concave function commonly employed in statistical learning, the value function of a copositive recourse function in two-stage stochastic programming, and many composite statistical functions in risk analysis, such as the value-at-risk (VaR), conditional value-at-risk (CVaR), expectation-based, VaR-based, and CVaR-based random deviation functionals. We also discuss decomposition methods for computing directional stationary points of a class of nonsmooth, nonconvex dc programs that combined the Gauss-Seidel idea, the alternating direction method of multipliers, and a special technique to handle the negative of a pointwise max function.
Thursday, April 20, 2017
4:00PM AP&M 6402