tests/maternGmrfPrec.R

In geostatsp: Geostatistical Modelling with Likelihood and Bayes

library('geostatsp')
matrix(NNmat(7, 7)[,25], 7, 7)

myr = squareRaster(ext(0,600,0,300), 60)
theNN = NNmat(myr)


params=c(range = 6*xres(myr),
		cellSize=xres(myr),
		shape=2,
		variance=1600)


# precision matrix without adjusting for edge effects
system.time({precMat = maternGmrfPrec(N=theNN, param=params,
					adjustEdges=FALSE)})


system.time({theNNadj = NNmat(N=myr, nearest=params['shape']+1, adjustEdges=TRUE)})
# and with the adjustment
system.time({precMatCorr =maternGmrfPrec(N=theNNadj, param=params, 
					adjustEdges=TRUE)}) 

# find better parameters using a numerical optimizer
system.time({precMatAdj =maternGmrfPrec(theNNadj, param=params, 
					adjustEdges='optimalShape') })

attributes(precMat)$info$adjustEdges
attributes(precMatAdj)$info$adjustEdges
attributes(precMatCorr)$info$adjustEdges

Nx = attributes(theNN)$Nx
Ny = attributes(theNN)$Ny
midcell = Nx*round(Ny/2) + round(Nx/2) # the middle cell
edgecell = Nx*5 + 5 # cell near corner


# show precision of mid cell
precMid=matrix(precMat[,midcell], Ny, Nx, byrow=TRUE)
precMid[round(Ny/2)+seq(-3, 3), round(Nx/2)+seq(-3, +3)]

if(Sys.info()['user'] =='patrick') {
	
# invert to get variance matrices
	precMatAdjInv = solve(precMatAdj)
	precMatCorrInv = solve(precMatCorr)
	precMatInv = solve(precMat)
	
# check range of diagonal entries, should be
	params['variance']
	range(diag(precMatInv))
	range(diag(precMatCorrInv))
	range(diag(precMatAdjInv))
	
# map marginal variances
	tempA=tempC=tempU = attributes(precMat)$raster
	values(tempC) = diag(precMatCorrInv)
	values(tempU) = diag(precMatInv)
	values(tempA) = diag(precMatAdjInv)
	
	if(!interactive()){
		pdf("maternGmrfMarginalVarRasters.pdf",height=10,width=5)
	}
	par(mfrow=c(3,1))
	plot(tempU)
	plot(tempC)
	plot(tempA)
	if(!interactive()){
		dev.off()
	}
	
	
	
# variance matrices
	Ncell = Nx*Ny
	midVec = sparseMatrix(midcell,1,x=1,dims=c(Ncell,1))
	edgeVec = sparseMatrix(edgecell,1,x=1,dims=c(Ncell,1))
	
	therast = attributes(precMat)$raster
	values(therast)=NA
	varRast = c(therast, therast,therast,therast,therast,therast)
	names(varRast) = c('mid','edge','midCor','edgeCor','midAdj','edgeAdj')
	values(varRast[['mid']]) = as.vector(Matrix::solve(precMat, midVec))
	values(varRast[['edge']]) = as.vector(Matrix::solve(precMat, edgeVec))
	values(varRast[['midCor']]) =  as.vector(Matrix::solve(precMatCorr, midVec))
	values(varRast[['edgeCor']]) = as.vector(Matrix::solve(precMatCorr, edgeVec))
	values(varRast[['midAdj']]) =  as.vector(Matrix::solve(precMatAdj, midVec))
	values(varRast[['edgeAdj']]) = as.vector(Matrix::solve(precMatAdj, edgeVec))
	
	
	if(!interactive()){
		pdf("maternGmrfPredRasters.pdf",height=11,width=4)
	}
	par(mfrow=c(7,2))
	
	mMid = matern(varRast[[1]],y=xyFromCell(varRast,midcell),param=params)
	mEdge = matern(varRast[[1]],y=xyFromCell(varRast,edgecell),param=params)
	plot(mMid)
	plot(mEdge)
	
	
	plot(varRast[['mid']])
	plot(varRast[['edge']])
	
	plot(varRast[['mid']]-mMid)
	plot(varRast[['edge']]-mEdge)
	
	plot(varRast[['midCor']])
	plot(varRast[['edgeCor']])
	plot(varRast[['midCor']]-mMid)
	plot(varRast[['edgeCor']]-mEdge)
	
	plot(varRast[['midAdj']])
	plot(varRast[['edgeAdj']])
	plot(varRast[['midAdj']]-mMid)
	plot(varRast[['edgeAdj']]-mEdge)
	
	
	if(!interactive()){
		dev.off()
	}
	
	
# compare covariance matrix to the matern
	
	xseqFull = seq(-1.5, 1.5, len=401)*params["range"]
	
	diagdist = sqrt(2)
	
	xymid = xyFromCell(varRast, midcell)
	
	xseq = seq(xmin(varRast)+xres(varRast)/2,xmax(varRast), by=xres(varRast))
	yseq = seq(ymin(varRast)+yres(varRast)/2,ymax(varRast), by=yres(varRast))
	
	vx= extract(varRast, 
			vect(cbind(xseq,
							rep(xymid[2],length(xseq)))
			))
	vy = extract(varRast, 
			vect(cbind(
							rep(xymid[1],length(yseq)),
							yseq)
			)
	)
	
	aseq = seq(0,100,by=xres(varRast))
	aseq = sort(unique(c(aseq, -aseq)))
	
	lur = vect(cbind(xymid[1]+aseq,xymid[2]+aseq))
	lll = vect(cbind(xymid[1]-aseq,xymid[2]+aseq))
	
	vur = extract(varRast, 
			lur		)
	
	
	vll= extract(varRast, 
			lll		)
	xyedg = xyFromCell(varRast, edgecell)
	
	xseqe = seq(xmin(varRast)+xres(varRast)/2,xmax(varRast), by=xres(varRast))
	yseqe = seq(ymin(varRast)+yres(varRast)/2,ymax(varRast), by=yres(varRast))
	
	vxe= extract(varRast, 
			vect(cbind(xseqe,
							rep(xyedg[2],length(xseqe)))
			))
	vye = extract(varRast, 
			vect(cbind(
							rep(xyedg[1],length(yseqe)),
							yseqe)
			)
	)
	
	aseqe = seq(0,100,by=xres(varRast))
	aseqe = sort(unique(c(aseqe, -aseqe)))
	
	lure = vect(cbind(xyedg[1]+aseqe,xyedg[2]+aseqe))
	llle = vect(cbind(xyedg[1]-aseqe,xyedg[2]+aseqe))
	
	vure = extract(varRast, 
			lure		)
	
	vlle= extract(varRast, 
			llle		)
	
	
	
	
	theedg = c('edge','edgeAdj','edgeCor')
	
	
	
	themid = c('mid','midAdj','midCor')
	
	
	thecol =c(mid='red',edge='red',midCor='green',edgeCor='green',
			midAdj='blue',edgeAdj='blue')
	thelwd = c(mid=3,midCor=1,midAdj=1,edge=3,edgeCor=1,edgeAdj=1)
	
	if(!interactive()){
		pdf("maternGmrfPred.pdf",height=5,width=10)
	}
	par(mfrow=c(2,2),mar=c(3,2,0,0))
# middle cell, full plot
	plot(xseqFull, matern(xseqFull, param=params),
			type = 'l',ylab='cov', xlab='dist',
			main="matern v gmrf",
			ylim=c(0,params['variance']*1.1),col='grey',lwd=9)
	
	matlines(xseq-xymid[1],vx[,themid],col=thecol[themid],lty=1,
			type='o',lwd=thelwd[themid])
	matlines(yseq-xymid[2],vy[,themid],col=thecol[themid],lty=1,
			type='o',lwd=thelwd[themid])
	
	
	matlines(aseq*diagdist, 
			vur[,themid],col=thecol[themid],lty=2,type='o',
			lwd=thelwd[themid])
	matlines(aseq*diagdist, 
			vll[,themid],col=thecol[themid],lty=2,type='o',lwd=thelwd[themid])
	
	
	
	
	
	
	legend("topright", lty=1, 
			col=thecol[c('edge','edgeCor','edgeAdj')],
			legend=c('vanilla','edge','adj')
	)
	
	
# edge cell, full plot
	plot(xseqFull, matern(xseqFull, param=params),
			type = 'l',ylab='cov', xlab='dist',
			main="matern v gmrf",
			ylim=c(0,params['variance']*1.1))
	
	
	
	matlines(xseqe-xyedg[1],vxe[,theedg],col=thecol[theedg],lty=1,
			lwd=thelwd[theedg])
	matlines(yseqe-xyedg[2],vye[,theedg],col=thecol[theedg],lty=1,
			lwd=thelwd[theedg])
	
	
	matlines(aseqe*diagdist, vure[,theedg],col=thecol[theedg],lty=2,
			lwd=thelwd[themid])
	
	matlines(aseqe*diagdist, vlle[,theedg],col=thecol[theedg],lty=2,
			lwd=thelwd[theedg])
	
	# middle cell detail plot
	
	plot(xseqFull, matern(xseqFull, param=params),
			type = 'l',ylab='cov', xlab='dist',
			main="matern v gmrf",
			ylim=params['variance']*c(0.5,1.05),
			xlim=c(-0.5, 0.5)*params['range'],col='grey',lwd=4)
	
	
	
	matlines(xseq-xymid[1],vx[,themid],col=thecol[themid],
			lty=1,
			lwd=thelwd[themid])
	matlines(yseq-xymid[2],vy[,themid],col=thecol[themid],lty=1,
			lwd=thelwd[themid])
	
	
	matlines(aseq*diagdist, vur[,themid],col=thecol[themid],lty=2,
			lwd=thelwd[themid])
	
	matlines(aseq*diagdist, vll[,themid],col=thecol[themid],lty=2,
			lwd=thelwd[themid])
	
	
	
	# edge cells
	
	plot(xseqFull, matern(xseqFull, param=params),
			type = 'l',ylab='cov', xlab='dist',
			main="matern v gmrf",
			ylim=params['variance']*c(0.5,1.05),
			xlim=c(-0.5, 0.5)*params['range'])
	
	
	matlines(xseqe-xyedg[1],vxe[,theedg],col=thecol[theedg],
			lty=1,
			lwd=thelwd[theedg])
	matlines(yseqe-xyedg[2],vye[,theedg],col=thecol[theedg],
			lty=1,
			lwd=thelwd[theedg])
	
	
	matlines(aseqe*diagdist, vure[,theedg],col=thecol[theedg],
			lty=2,
			lwd=thelwd[themid])
	
	matlines(aseqe*diagdist, vlle[,theedg],col=thecol[theedg],
			lty=2,
			lwd=thelwd[theedg])
	
	
	
	if(!interactive()){
		dev.off()	
	}
	
# matern variance matrix
	covMatMatern = 
			matern(attributes(precMat)$raster, 
					param=params[c('range','shape','variance')])
	
	covMatMaternAdj = 
			matern(attributes(precMat)$raster, 
					param=attributes(precMatAdj)$info$optimal[
							c('range','shape','variance')])
	
	
	
	prodUncor = covMatMatern %*% (precMat)
	prodCor = covMatMatern %*% (precMatCorr)
	prodAdj = covMatMatern %*% (precMatAdj)
	prodAdjOpt = covMatMaternAdj %*% (precMatAdj)
	
	
	
	thebreaks = c(-100, -0.25, 0.25, 0.75, 1.5, 5, 10, 100)
	
	thecol = terrain.colors(length(thebreaks)-1)
	
	if(!interactive()){
		pdf("maternGmrfDiagRast.pdf",height=8,width=5)
	}
	par(mfrow=c(3,1))
	temp = attributes(precMat)$raster
	values(temp) = diag(prodCor)
	plot(temp,breaks=thebreaks,col=thecol,legend=F)
#	mapmisc::legendBreaks('right',breaks=thebreaks, col=thecol)
	
	temp = attributes(precMat)$raster
	values(temp) = diag(prodUncor)
	plot(temp,breaks=thebreaks,col=thecol,legend=FALSE)
	
	temp = attributes(precMat)$raster
	values(temp) = diag(prodAdj)
	plot(temp,breaks=thebreaks,col=thecol,legend=FALSE)
	
	if(!interactive()){
		dev.off()
	}
	
	if(!interactive()){
		pdf("maternGmrfHist.pdf",height=9,width=7)
	}
	par(mfrow=c(4,2),mar=c(4,4,0,0))	
	
	hist(Matrix::diag(prodUncor),breaks=60,ylab='vanilla')
	abline(v=1)
	hist(prodUncor[lower.tri(prodUncor,diag=FALSE)],breaks=60)	
	abline(v=0)
	
	thebreaks = c(-10,seq(0.4, 1.1, len=61),10)
	
	if(interactive()){
		hist(Matrix::diag(prodCor),breaks=thebreaks,ylab='edge')
	} else {
		hist(Matrix::diag(prodCor),breaks=60,ylab='edge')
	}
	abline(v=1)
	hist(prodCor[lower.tri(prodCor,diag=FALSE)],breaks=60)	
	abline(v=0)
	
	if(interactive()){
		hist(Matrix::diag(prodAdj),breaks=thebreaks,ylab='adj')
	} else {
		hist(Matrix::diag(prodAdj),breaks=60,ylab='edge')
	}
	abline(v=1)
	hist(prodAdj[lower.tri(prodAdj,diag=FALSE)],breaks=60)	
	abline(v=0)
	
	if(interactive()){
		hist(Matrix::diag(prodAdjOpt),breaks=thebreaks,ylab='adj optimal')
	} else {
		hist(Matrix::diag(prodAdjOpt),breaks=60,ylab='edge')
	}
	abline(v=1)
	hist(prodAdjOpt[lower.tri(prodAdjOpt,diag=FALSE)],breaks=60)	
	abline(v=0)
	
	
	if(!interactive()){
		dev.off()	
	}
	
	
	# testing marginal variance
	if(FALSE){
		
# order one
		Sa = seq(from=4.01,to=5,len=10)
		maxvar=NULL
		for(a in Sa) {
			myp = theNN
			myp@x = c(4+a^2,-2*a,2,1,0,0,0,0)[myp@x]
			myv = solve(myp)
			values(myr) = diag(myv)
			maxvar = c(maxvar,max(myr[15,]))
			
		}
		
		phi = 4/Sa
		scale = sqrt(Sa-4)
		range = sqrt(8)/scale
		
		margvar = (4/pi)^(scale^0.5/2)/(4*pi*(Sa-4))
		margvar = (4/pi)^(scale^2/2)/(4*pi*(Sa-4))
		
		margvarOld = 1/(4*pi*(Sa-4))
		
		
		par(mfrow=c(4,1),mar=c(2,4,0,0))
		they= c(-0.1, 0.1)
		plot(range, maxvar/margvar-1,log='x',ylab='range',ylim=they)
		lines(range, maxvar/margvarOld-1,col='grey')
		abline(h=0)
		
		plot(1-phi, maxvar/margvar-1,ylab='phi',ylim=they)
		lines(1-phi, maxvar/margvarOld-1,col='grey')
		abline(h=0)
		plot(scale, maxvar/margvar-1,ylab='scale',ylim=they)
		lines(scale, maxvar/margvarOld-1,col='grey')
		abline(h=0)
		plot(Sa-4, maxvar/margvar-1,log='x',ylab='Sa-4',ylim=they)
		lines(Sa-4, maxvar/margvarOld-1,col='grey')
		abline(h=0)
		max(maxvar/margvar)
		them=which.max(maxvar/margvar)
		c(range[them],phi[them],scale[them],Sa[them])
		
		
# order two
		
		Sa2 = seq(from=4.01,to=5,len=40)
		maxvar2=NULL
		for(a in Sa2) {
			myp = theNN
			myp@x = c("1" = a*(a*a+12),
					"2" = -3*(a*a+3),
					"3" = 6*a,
					"4" = 3*a, 
					"5" =  -3,
					"6" = -1)[myp@x]
			myv = solve(myp)
			values(myr) = diag(myv)
			maxvar2 = c(maxvar2,max(myr[15,]))
			
		}
		
		phi2 = 4/Sa2
		scale2 = sqrt(Sa2-4)
		range2 = sqrt(8*2)/scale2
		
		
		margvar = (4/pi)^(scale^2/2)/(4*pi*2*(Sa2-4)^2)
		margvarOld = 1/(4*pi*2*(Sa2-4)^2)
		
		
		par(mfrow=c(4,1),mar=c(2,4,0,0))
		they= c(-0.1, 0.1)
		plot(range, maxvar2/margvar-1,log='x',ylab='range',ylim=they)
		lines(range, maxvar2/margvarOld-1,col='grey')
		abline(h=0)
		
		plot(1-phi, maxvar2/margvar-1,ylab='phi',ylim=they)
		lines(1-phi, maxvar2/margvarOld-1,col='grey')
		abline(h=0)
		plot(scale, maxvar2/margvar-1,ylab='scale',ylim=they)
		lines(scale, maxvar2/margvarOld-1,col='grey')
		abline(h=0)
		plot(Sa-4, maxvar2/margvar-1,log='x',ylab='Sa-4',ylim=they)
		lines(Sa-4, maxvar2/margvarOld-1,col='grey')
		abline(h=0)
		max(maxvar2/margvar)
		them=which.max(maxvar2/margvar)
		c(range2[them],phi2[them],scale2[them],Sa2[them])
		
		
		
		plot(xseqFull, matern(xseqFull, param=params),
				type = 'l',ylab='cov', xlab='dist',
				main="matern v gmrf",
				ylim=params['variance']*c(0.25,1.05),
				xlim=c(-0.75, 0.75)*params['range'])
		
		# middle cell
		themid='mid'
		
		plot(xseq-xymid[1],
				vx[,themid] - matern(xseq-xymid[1], param=params),
				col='red',lty=thelty[themid],
				lwd=thelwd[themid],type='o',pch=1,ylim=c(-55,12))
		matlines(yseq-xymid[2],
				vy[,themid]- matern(yseq-xymid[2], param=params),
				col='orange',lty=thelty[themid],
				lwd=thelwd[themid],type='o',pch=2)
		
		
		matlines(aseq*diagdist, 
				vur[,themid]- matern(aseq*diagdist, param=params),
				col='yellow',lty=thelty[themid],
				lwd=thelwd[themid],type='o',pch=3)
		
		matlines(aseq*diagdist, vll[,themid]- matern(aseq*diagdist, param=params),
				col='pink',lty=thelty[themid],
				lwd=thelwd[themid],type='o',pch=4)
		
		xseqFull = seq(-400,400,len=1000)
		
		mbase =  matern(xseqFull, param=
						c(params[c('variance','range')],
								params['shape'])
		)
		lines(xseqFull, 
				matern(xseqFull, 
						param=
								c(params['variance'],
										params['range']/1.025,
										params['shape']/1.23)
				)-mbase,
				col='blue'
		)
		lines(xseqFull, 
				matern(xseqFull, 
						param=
								c(params['variance'],
										params['range']/1.0225,
										params['shape']/1.22)
				)-mbase,
				col='grey'
		)
	}
	
# test providing ar parameter
	
	paramsAR = c(shape=1,oneminusar=0.1)
#	N=theNN;param=paramsAR;adjustEdges=TRUE;adjustParam=FALSE 
	
	precMatAR =maternGmrfPrec(theNN, param=paramsAR,
			adjustEdges=FALSE,
			adjustParam=FALSE) 
	
	
# find better parameters using a numerical optimizer
	precMatAdjAR =maternGmrfPrec(theNN, param=paramsAR, 
			adjustEdges=TRUE,adjustParam=TRUE) 
	
	
# and with the adjustment
	precMatCorrAR =maternGmrfPrec(theNN, param=paramsAR, 
			adjustEdges=TRUE,adjustParam=FALSE) 
}