Say you have a matrix A which is of size P × P and a number Q < P can be used to
take a partition of said matrix, where:
A1 is the upper-left sub matrix, with dimension Q × Q,
A2 is the upper-right sub matrix, with dimension Q × (P-Q),
A3 is the lower-left sub matrix, with dimension (P-Q) × Q,
A4 is the lower-rightsub matrix, with dimension (P-Q) × (P-Q).
Which looks like this:
A1 | A2
A = ---+----
A3 | A4
How can you calculate the matrix:
Where 0q is a Q × Q matrix with zero elements.
I'm learning from a book called "Discovering Statistics using R" and although it discusses partitioned matrices, it doesn't show how to calculate one like the one given above and unfortunately I'm having no luck on the programming or maths based searches...
Any help, either mathematically and/or example R code would be great. Thanks in advance.
R has various ways to grab blocks from matrices. For instance, you can use a vector of indexes to reference a set of rows or set of columns using the extract function [, as shown in this example. (The option drop=FALSE is needed if you must handle the case p=1 or q=1 so that R continues to treat the results as matrices and not just vectors.)
#
# Create a symmetric p-d matrix of size p+q.
#
p <- 2; q <- 3 # Both must be 1 or greater
x <- matrix(rnorm((p+q)^2 * 2), ncol=p+q)
a <- cov(x)
#
# Compute b.
#
i <- 1:p; j <- 1:q + p # Indexes of the blocks
b <- a[i,i, drop=FALSE] -
a[i,j, drop=FALSE] %*% solve(a[j,j, drop=FALSE], a[j,i, drop=FALSE])
Matrix inversion is implemented by solve (which is numerically more stable and efficient than computing the inverse of a[j,j] and multiplying that by a[j,i]), the single remaining multiplication is carried out by %*%, and - subtracts one matrix from another component by component. In this fashion the code closely parallels the mathematical expression in the question.
Related
I have an expression that contains several parts. However, for simplicity, consider only the following part as MWE:
Let's assume we have the inverse of a matrix Y that I want to differentiate w.r.t. x.
Y is given as I - (x * b * t(b)), where I is the identity matrix, x is a scalar, and b is a vector.
According to The Matrix Cookbook Equ. 59, the partial derivative of an inverse is:
Normally I would use the function D from the package stats to calculate the derivatives. But that is not possible in this case, because e.g. solve to specify Y as inverse and t() is not in the table of derivatives.
What is the best workaround to circumvent this problem? Are there any other recommended packages that can handle such input?
Example that doesn't work:
f0 <- expression(solve(I - (x * b %*% t(b))))
D(f0, "x")
Example that works:
f0 <- expression(x^3)
D(f0, "x")
3 * x^2
I assume that the question is how to get an explicit expression for the derivative of the inverse of Y with respect to x. In the first section we compute it and in the second section we double check it by computing it numerically and show that the two approaches give the same result.
b and the null space of b are both eigenspaces of Y which we can readily verify by noting that Yb = (1-(b'b)x)b and if z belongs to the nullspace of b then Yz = z. This also shows that the corresponding eigenvalues are 1 - x(b'b) with multiplicity 1 and 1 with multiplicity n-1 (since the nullspace of b has that dimension).
As a result of the fact that we can expand such a matrix into the sum of each eigenvalue times the projection onto its eigenspace we can express Y as the following where bb'/b'b is the projection onto the eigenspace spanned by b and the part pre-multiplying it is the eigenvalue. The remaining terms do not involve x because they involve an eigenvalue of 1 independently of x and the nullspace of b is independent of x as well.
Y = (1-x(b'b))(bb')/(b'b) + terms not involving x
The inverse of Y is formed by taking the reciprocals of the eigenvalues so:
Yinv = 1/(1-x(b'b)) * (bb')/(b'b) + terms not involving x
and the derivative of that wrt x is:
(b'b) / (1 - x(b'b))^2 * (bb')/(b'b)
Cancelling the b'b and writing the derivative in terms of R code:
1/(1 - x*sum(b*b))^2*outer(b, b)
Double check
Using specific values for b and x we can verify it against the numeric derivative as follows:
library(numDeriv)
x <- 1
b <- 1:3
# Y inverse as a function of x
Yinv <- function(x) solve(diag(3) - x * outer(b, b))
all.equal(matrix(jacobian(Yinv, x = 1), 3),
1/(1 - x*sum(b*b))^2*outer(b, b))
## [1] TRUE
I'm currently trying to recreate this Matlab function in R:
function X = uniform_sphere_points(n,d)
% X = uniform_sphere_points(n,d)
%
%function generates n points unformly within the unit sphere in d dimensions
z= randn(n,d);
r1 = sqrt(sum(z.^2,2));
X=z./repmat(r1,1,d);
r=rand(n,1).^(1/d);
X = X.*repmat(r,1,d);
Regarding the the right matrix division I installed the pracma package. My R code right now is:
uniform_sphere_points <- function(n,d){
# function generates n points uniformly within the unit sphere in d dimensions
z = rnorm(n, d)
r1 = sqrt(sum(z^2,2))
X = mrdivide(z, repmat(r1,1,d))
r = rnorm(1)^(1/d)
X = X * matrix(r,1,d)
return(X)
}
But it is not really working since I always end with a non-conformable arrays error in R.
This operation for sampling n random points from the d-dimensional unit sphere could be stated in words as:
Construct a n x d matrix with entries drawn from the standard normal distribution
Normalize each row so it has (2-norm) magnitude 1
For each row, compute a random value by taking a draw from the uniform distribution (between 0 and 1) and raise that value to the 1/d power. Multiply all elements in the row by that value.
The following R code does these operations:
unif.samp <- function(n, d) {
z <- matrix(rnorm(n*d), nrow=n, ncol=d)
z * (runif(n)^(1/d) / sqrt(rowSums(z^2)))
}
Note that in the second line of code I have taken advantage of the fact that multiplying a n x d matrix in R by a vector of length n will multiply each row by the corresponding value in that vector. This saves us the work of using repmat to construct matrices of exactly the same size as our original matrix for these sorts of row-specific operations.
Does anyone know how to generate matrix with certain rank in R?
I ultimately want to create data matrix Y = X + E
where rank(X)=k and E~i.i.d.N(0,sigma^2).
The easiest is the identity matrix, which has always full rank. So e.g. use:
k <- 10
mymatrix <- diag(k)
Here, rows and columns are equal to the rank you specify
I suppose you want to mimic a regression model, so you might want to have more rows (meaning 'observations') than columns, (e.g. 'variables'). The following code allows you to specify both:
k <- 5 # rank of your matrix
nobs <- 10 # number of lines within X
X <- rbind(diag(k), matrix(rep(0,k*(nobs-k)), ncol=k))
y <- X + rnorm(nobs)
Note, that X - and therefore also y - now have full column rank. there is no multicollinearity in this 'model'.
I have data which I want to fit to the following equation using R:
Z(u,w)=z0*F(w)*[1-exp((-b*u)/F(w))]
where z0 and b are constants and F(w), w=0,...,9 is a decreasing step function that depends on w with F(0)=1 and u=1,...,50.
Z(u,w) is an observed set of data in the form of a 50x10 matrix (u=50,...,1 down the side of the rows and w=0,...,9 along the columns). For example as I haven't explained that great, Z(42,3) will be the element in the 9th row down and the 4th column along.
Using F(0)=1 I was able to get estimates of b and z0 using just the first column (ie w=0) with the code:
n0=nls(zuw~z0*(1-exp(-b*u)),start=list(z0=283,b=0.03),options(digits=10))
I then found F(w) for w=1,...,9 by going through each columns and using the vlaues of b and z0 I found.
However, I was wanting to find a way to estimate all the 12 parameters at once (b, z0 and the 10 values of F(w)) as b and z0 should be fitted to all the data, not just the first column.
Does anyone know of any way of doing this? All help would be greatly appreciated!
Thanks
James
This may be a case where the formula interface of the nls(...) function works against you. As an alternative, you can use nls.lm(...) in the minpack.lm package to perform non-linear regression with a programmatically defined function. To demonstrate this, first we create an artificial dataset which follows your functional form by design, with random error added (error ~ N[0,1]).
u <- 1:50
w <- 0:9
z0 <- 100
b <- 0.02
F <- 10/(10+w^2)
# matrix containing data, in OP's format: rows are u, cols are w
m <- do.call(cbind,lapply(w,function(w)
z0*F[w+1]*(1-exp(-b*u/F[w+1]))+rnorm(length(u),0,1)))
So now we have a matrix m, which is equivalent to your dataset. This matrix is in the so-called "wide" format - the response for different values of w is in different columns. We need it in "long" format: all responses in a single column, with a separate columns identifying u and w. We do this using melt(...) in the reshape2 package.
# prepend values of u
df.wide <- data.frame(u=u, m)
library(reshape2)
# reshape to long format: col1 = u, col2=w, col3=z
df <- melt(df.wide,id="u",variable.name="w", value.name="z")
df$w <- as.numeric(substr(df$w,2,4))-1
Now we have a data frame df with columns u, w, and z. The nls.lm(...) function takes (at least) 4 arguments: par is a vector of initial estimates of the parameters of the fit, fn is a function that calculates the residuals at each step, observed is the dependent variable (z), and xx is a vector or matrix containing the independent variables (u, v).
Next we define a function, f(par, xx), where par is an 11 element vector. The first two elements contain estimates of z0 and b. The next 9 contain estimates of F(w), w=1:9. This is because you state that F(0) is known to be 1. xx is a matrix with two columns: the values for u and w respectively. f(par,xx) then calculates estimate of the response z for all values of u and w, for the given parameter estimates.
library(minpack.lm)
# model function
f <- function(pars, xx) {
z0 <- pars[1]
b <- pars[2]
F <- c(1,pars[3:11])
u <- xx[,1]
w <- xx[,2]
z <- z0*F[w+1]*(1-exp(-b*u/F[w+1]))
return(z)
}
# residual function
resids <- function(p, observed, xx) {observed - f(p,xx)}
Next we perform the regression using nls.lm(...), which uses a highly robust fitting algorithm (Levenberg-Marquardt). Consequently, we can set the par argument (containing the initial estimates of z0, b, and F) to all 1's, which is fairly distant from the values used in creating the dataset (the "actual" values). nls.lm(...) returns a list with several components (see the documentation). The par component contains the final estimates of the fit parameters.
# initial parameter estimates; all 1's
par.start <- c(z0=1, b=1, rep(1,9))
# fit using Levenberg-Marquardt algorithm
nls.out <- nls.lm(par=par.start,
fn = resids, observed = df$z, xx = df[,c("u","w")],
control=nls.lm.control(maxiter=10000, ftol=1e-6, maxfev=1e6))
par.final <- nls.out$par
results <- rbind(predicted=c(par.final[1:2],1,par.final[3:11]),actual=c(z0,b,F))
print(results,digits=5)
# z0 b
# predicted 102.71 0.019337 1 0.90456 0.70788 0.51893 0.37804 0.27789 0.21204 0.16199 0.13131 0.10657
# actual 100.00 0.020000 1 0.90909 0.71429 0.52632 0.38462 0.28571 0.21739 0.16949 0.13514 0.10989
So the regression has done an excellent job at recovering the "actual" parameter values. Finally, we plot the results using ggplot just to make sure this is all correct. I can't overwmphasize how important it is to plot the final results.
df$pred <- f(par.final,df[,c("u","w")])
library(ggplot2)
ggplot(df,aes(x=u, color=factor(w)))+
geom_point(aes(y=z))+ geom_line(aes(y=pred))
I have a matrix and I would like to know if it is diagonalizable. How do I do this in the R programming language?
If you have a given matrix, m, then one way is the take the eigen vectors times the diagonal of the eigen values times the inverse of the original matrix. That should give us back the original matrix. In R that looks like:
m <- matrix( c(1:16), nrow = 4)
p <- eigen(m)$vectors
d <- diag(eigen(m)$values)
p %*% d %*% solve(p)
m
so in that example p %*% d %*% solve(p) should be the same as m
You can implement the full algorithm to check if the matrix reduces to a Jordan form or a diagonal one (see e.g., this document). Or you can take the quick and dirty way: for an n-dimensional square matrix, use eigen(M)$values and check that they are n distinct values. For random matrices, this always suffices: degeneracy has prob.0.
P.S.: based on a simple observation by JD Long below, I recalled that a necessary and sufficient condition for diagonalizability is that the eigenvectors span the original space. To check this, just see that eigenvector matrix has full rank (no zero eigenvalue). So here is the code:
diagflag = function(m,tol=1e-10){
x = eigen(m)$vectors
y = min(abs(eigen(x)$values))
return(y>tol)
}
# nondiagonalizable matrix
m1 = matrix(c(1,1,0,1),nrow=2)
# diagonalizable matrix
m2 = matrix(c(-1,1,0,1),nrow=2)
> m1
[,1] [,2]
[1,] 1 0
[2,] 1 1
> diagflag(m1)
[1] FALSE
> m2
[,1] [,2]
[1,] -1 0
[2,] 1 1
> diagflag(m2)
[1] TRUE
You might want to check out this page for some basic discussion and code. You'll need to search for "diagonalized" which is where the relevant portion begins.
All symmetric matrices across the diagonal are diagonalizable by orthogonal matrices. In fact if you want diagonalizability only by orthogonal matrix conjugation, i.e. D= P AP' where P' just stands for transpose then symmetry across the diagonal, i.e. A_{ij}=A_{ji}, is exactly equivalent to diagonalizability.
If the matrix is not symmetric, then diagonalizability means not D= PAP' but merely D=PAP^{-1} and we do not necessarily have P'=P^{-1} which is the condition of orthogonality.
you need to do something more substantial and there is probably a better way but you could just compute the eigenvectors and check rank equal to total dimension.
See this discussion for a more detailed explanation.