A Matlab implementation of the optimal quadratic averaging algorithm. The accompanying paper can be found at http://arxiv.org/abs/1604.06543.
The file optave.m contains a function that implements the optimal quadratic averaging algorithm. A detailed description of optave.m is given in the comments in the function itself.
The function optave.m expects the objective in a structured form (see the comments in optave.m for more information). The function in make_optave_objective.m is called by optave.m to process and check the structure of the objective passed into optave.m.
The file bracket_minimizer.m contains a method for bracketing the minimizer of a 1-dimensional convex function. The function bracket_minimizer.m and the Matlab function fminbnd are used by optave.m to perform exact line search.
The file main.m shows examples of using optave.m to minimize the Rosenbrock function, the logistic loss function, and the log-sum-exp function.
The functions log_sum_exp.m, mean_log_one_plus_exp.m, and c_half_square_norm.m are component functions used to make the Rosenbrock function, the logistic loss function, and the log-sum-exp function.
The functions in make_rosenbrock.m, make_logistic_loss.m, and make_logsumexp.m put together the component functions and return structured objectives that can be passed into optave.m.
Running the file memory_comparison_plot.m produces a plot that compares minimizing the logistic loss using optave.m for various memory sizes.
The function callback.m is passed into optave.m to collect plotting data.
The files a1a.txt and colon-cancer.txt are LIBSVM data sets (for use in the logistic regression examples).