iluvjava / Proximal-Gradient

The proximal gradient methods in julia.

Proximal-Gradient

The proximal gradient methods in julia. Credit to <A Fast Iterative-Thresholding Algorithm for Linear Inverse Problem> by Amir Beck and Marc Teboule. The implementations is based on their paper and the book by Amir <First-Order method for Optimization> is an essential reference for the theoretical background of this repo.

---

How to Run

Run setup.jl with julia. And then you can start running the examples we have.

---

The Proximal Gradient

Smooth and nonsmooth function with gradient and proximal oracles are defined as types. The Proximal Gradient algorithm with adaptive stepsize and Nesterov Accelerations is impelmented in one function in this file.

---

Inverse Linear

We consider the LASSO Problem applied to image deblurring. The LASSO problem is simply:

$$ \arg\min_{x}\left\lbrace \Vert Ax - b\Vert^2_2 + \lambda \Vert x\Vert_1 \right\rbrace, $$

in which we solve it using the FISTA algorithm with $A$ being the discretized Guassian Blurr matrix. The example is in here. The sparse matrix multiplication is slow in Julia, I don't have mental strength to optimize the speed yet. It's proved in Beck book that the convergence is first order for both the accelerated and unaccelerated case.

---

Math

Read this for more math. We did:

1. Coding up numerical experiment.
2. Proved and explained everything.
3. Showed where the Nesterov Momentum came from.