# Gradient descent vs coordinate descent

When it comes to function minimisation, it’s time to open a book of optimisation and linear algebra. I am currently working on variable selection and lasso-based solutions in genetics. What lasso does is basically minimising the loss function and an l1−norm

penalty in order to set to zero some regression coefficients and select only those covariates that are really associated with the response. Pheew, the shortest summary of lasso ever!

We all know that, provided the function to be minimised is convex, a good direction to follow, in order to find a local minimum, is towards the negative gradient of the function. Now, my question is how good or bad is following the negative gradient with respect to a coordinate descent approach that loops across all dimensions and minimises along each?

There is no better way to try this with real code and start measuring. Hence, I wrote some code that implements both gradient descent and coordinate descent.

The comparison might not be completely fair because the learning rate in the gradient descent procedure is fixed at 0.1 (which in some cases might be slower indeed). But even with some tuning (maybe with some line search) or adaptive learning rates, it’s quite common to see that coordinate descent overcomes its brother gradient descent.

This occurs much more often when the number of covariates becomes very high, as in many computational biology problems. In the figure below, I plot the analytical solution in red, the gradient descent minimisation in blue and the coordinate descent in green, across a number of iterations.

A small explanation is probably needed to read the function that performs coordinate descent.

Given the regression model y=Xθ, the function to minimise for a least squares regression is f(θ)=|yXθ|2. Minimising over θ is achieved by 0=Xi(Xθy)=Xi(Xiθi+Xiθiy). Therefore θi=Xi(yXiθi)XiXi

it’s quite common to see that coordinate descent overcomes its brother gradient descent

Coordinate descent will update each θi

in a Round Robin fashion. Despite the learning rate of the gradient descent procedure (which could indeed speed up convergence), the comparison between the two is fair at least in terms of complexity.

Coordinate descent needs to perform O(n) operations for each coordinate update (n operations to compute residuals r=(yXiθi) and n to compute Xir. Each cycle of this type is performed p times, where p is the number of covariates. Gradient descent performs the same number of operations O(np)

. The R code that performs this comparison and generates the plot is given below

``````    library(lassoshooting)
# make data
rm(list = ls(all = TRUE)) # make sure previous work is clear
ls()
p = 70; n = 300
x <- matrix(rnorm(n*p), nrow=n, ncol=p)
# create the y-matrix of dependent variables
y <- matrix(rnorm(n), nrow=n)
m <- nrow(y)
# analytical results with matrix algebra
analytical <- solve(t(x)%*%x)%*%t(x)%*%y
# w/o feature scaling
grad <- function(x, y, theta) {
gradient <- (1/m)* (t(x) %*% ((x %*% t(theta)) - y))
}

# define gradient descent update algorithm
thetas = matrix(0,ncol=dim(x)[2], nrow=maxit)
theta <- matrix(0, nrow=1, ncol=dim(x)[2])
numcoord = dim(x)[2]

alpha = .1 # set learning rate

for (i in 1:maxit) {
theta <- theta - alpha*grad(x, y, theta)
thetas[i, ] = theta
}
return(thetas)
}
coord.descent <- function(x, maxit, lambda=0.1){
p = dim(x)[2]
theta <- matrix(0, ncol=p, nrow=maxit) # Init parameters

for(k in 1:maxit){
for(i in 1:p) {
iter = ifelse(k>1, yes=k-1, 1)
theta[k,i] = t(x[,i])%*% (y - x[,-i]%*%theta[iter,-i])/ (t(x[,i])%*%x[,i])
theta[k,i] = softthresh(theta[k,i], lambda)
}
}
return(theta)
}

#compute
maxiter = 20