catch incorrect length of lengthscale vector

[goldingn]

When a kernel is specified with a single lengthscale, but evaluated on a design matrix with multiple columns, only the first column is used and no warning is issued. We should let the user know about a dimension mismatch, and either error, or in the case of a single lengthscale bing provided, just replicate it and assume they want an isotropic kernel.

See here for a reprex: https://forum.greta-stats.org/t/gaussian-process-in-greta-matern-covariance/151/17

[jdyen]

This is messy. I wanted to add a check similar to check_active_dims when setting up the kernels. See here:

greta.gp/R/kernel_constructors.R

Line 81 in 448f29c

arguments = list(active_dims = check_active_dims(columns, lengthscales)))

But the dimension of the design matrix is unknown at this stage. So possibly needs to go in the kernel call, e.g.,

greta.gp/R/greta_kernel_class.R

Line 172 in 448f29c

tf_K <- function(operation, X, X_prime, greta_kernel, out_dim) {

Then the issue is that it needs to check only when lengthscale is a valid parameter.

[jdyen]

Probably need to do the same with other parameters, e.g., variance in the rbf kernel. It's surprising that it doesn't error inside the tf calls, e.g.,

greta.gp/R/tf_kernels.R

Lines 74 to 86 in 448f29c

	tf_rbf <- function(X, X_prime, lengthscales, variance, active_dims) {
	# pull out active dimensions
	X <- tf_cols(X, active_dims)
	X_prime <- tf_cols(X_prime, active_dims)
	# calculate squared distances
	r2 <- squared_dist(X, X_prime, lengthscales)
	# construct and return RBF kernel
	variance * tf$math$exp(-r2 / fl(2))
	}

I assume the extra dims aren't dropped before here, in which case it's a result of how tf handles broadcasting.

[njtierney]

I think this might be captured here? https://github.com/greta-dev/greta.gp/blob/main/R/checkers.R

https://github.com/greta-dev/greta.gp/blob/main/R/kernel_constructors.R#L107