# Computation of MEWMA ARLs (multivariate mean monitoring), returns function mewma.arl.f <- function(l, cE, p, delta=0, r=20, ntype=NULL, qm0=20, qm1=qm0) { if ( l<=0 | l>1 ) stop("l has to be between 0 and 1") if ( cE<=0 ) stop("threshold c has to be positive") if ( p<1 ) stop("wrong dimension parameter") if ( delta<0 ) stop("wrong magnitude value") if ( r<4 ) stop("resolution too small") if ( qm0<5 ) stop("more quadrature nodes needed") if ( qm1<5 ) stop("more quadrature nodes needed") if ( is.null(ntype) ) { if ( delta <1e-10 ) { ntype <- "gl2" } else { #if ( p %in% c(2,4) ) { if ( p==2 ) { ntype <- "gl3" } else { ntype <- "gl5" } } } # collocation basis of Chebshev polynomials Tn <- Vectorize(function(z, n) { if ( n==0 ) result <- 1 if ( n==1 ) result <- z if ( n==2 ) result <- 2*z^2 - 1 if ( n==3 ) result <- 4*z^3 - 3*z if ( n==4 ) result <- 8*z^4 - 8*z^2 + 1 if ( n==5 ) result <- 16*z^5 - 20*z^3 + 5*z if ( n>5 ) result <- cos( n*acos(z) ) result }) qtyp <- pmatch(tolower(ntype), c("gl", "co", "ra", "cc", "mc", "sr", "co2", "gl2", "gl3", "gl4", "gl5", "co3", "co4", "ngl1", "ngl2", "ngl3", "ngl4", "ngl5")) - 1 if ( is.na(qtyp) ) stop("invalid type of numerical algorithm") if ( abs(delta) < 1e-10 ) { # in-control LENGTH <- 3*r zeug <- .C("mewma_arl_f", as.double(l), as.double(cE), as.integer(p), as.double(delta), as.integer(r), as.integer(qtyp), as.integer(qm0), as.integer(qm1), ans=double(length=LENGTH), PACKAGE="spc")$ans g <- zeug[1:r] w <- zeug[1:r + r] z <- zeug[1:r + 2*r] # helper functions cE <- cE * l/(2-l) l2 <- ( (1-l)/l )^2 fchi <- function(a, u) dchisq( u/l^2, p, ncp=l2*a ) / l^2 FCHI <- function(a, u) pchisq( u/l^2, p, ncp=l2*a ) if ( qtyp %in% c(0,2,5) ) arl <- Vectorize(function(a) 1 + sum( w * fchi(a, z) * g ), "a") # ordinary GL or Radau or Simpson rule Nystroem if ( qtyp==7 ) arl <- Vectorize(function(a) 1 + sum( w * fchi(a, z^2) * g * 2*z ), "a") # GL Nystroem with ()^2 substitution if ( qtyp==1 ) arl <- Vectorize(function(a) sum( Tn( (2*a-cE)/cE, 0:(r-1) ) * g ), "a") # collocation if ( qtyp==3 ) arl <- Vectorize(function(a) 1 + sum( w * fchi(a, z^2) * g ) * cE/2 , "a") # Clenshaw-Curtis if ( qtyp==4 ) arl <- Vectorize(function(a) 1 + sum( ( FCHI(a, z^2) - FCHI(a, c(0, z[-length(z)]^2)) ) * g), "a") # Markov chain (Runger/Prabhu) } else { # out-of-control if ( qtyp==4 ) { cE_ <- sqrt( cE * l/(2-l) ) w <- 2*cE_/( 2*r + 1 ) ii <- function(ix, iy) (ix-r)^2 + iy^2 < cE_^2/w^2 CIRC <- as.vector( t(outer( 0:(2*r), 0:r, ii)) ) dQ <- sum(CIRC) LENGTH <- dQ } else { r2 <- r^2 LENGTH <- r2 + 4*r } zeug <- .C("mewma_arl_f", as.double(l), as.double(cE), as.integer(p), as.double(delta), as.integer(r), as.integer(qtyp), as.integer(qm0), as.integer(qm1), ans=double(length=LENGTH), PACKAGE="spc")$ans if ( qtyp!=4 ) { g <- zeug[1:r2] w0 <- zeug[1:r + r2] z0 <- zeug[1:r + r2 + r] w1 <- zeug[1:r + r2 + 2*r] z1 <- zeug[1:r + r2 + 3*r] } else { g <- zeug[1:dQ] } # helpers l2 <- ( (1-l)/l )^2 h <- cE * l/(2-l) rdc <- l * sqrt( delta/h ) sig <- l / sqrt( h ) lsd <- l * sqrt( delta ) if ( qtyp %in% c(0, 2, 3, 5) ) arl <- Vectorize(function(a, b) { # ordinary GL or Radau or CC or Simpson rule Nystroem if ( abs(h-a) < 1e-10 ) a_ <- 1 else a_ <- ( a - b^2/delta ) / ( h - b^2/delta ) b_ <- b / sqrt( delta * h ) m <- rdc + (1-l)*b_ eta <- l2 * h * (1-b_^2) * a_ if ( eta < 1e-10 ) eta <- 0 result <- 1 for ( i in 1:r ) { korr <- h * (1-z1[i]^2) / l^2 outer <- korr * w1[i] * dnorm( z1[i], mean=m, sd=sig) inner <- sum( w0 * dchisq( korr*z0, p-1, eta) * g[ (i-1)*r + 1:r ] ) result <- result + inner * outer } result }) if ( qtyp==7 ) arl <- Vectorize(function(a, b) { # GL Nystroem with ()^2 substitution if ( abs(h-a) < 1e-10 ) a_ <- 1 else a_ <- ( a - b^2/delta ) / ( h - b^2/delta ) b_ <- b / sqrt( delta * h ) m <- rdc + (1-l)*b_ eta <- l2 * h * (1-b_^2) * a_ if ( eta < 1e-10 ) eta <- 0 result <- 1 for ( i in 1:r ) { korr <- h * (1-z1[i]^2) / l^2 outer <- korr * w1[i] * dnorm( z1[i], mean=m, sd=sig) inner <- sum( w0 * dchisq( korr*z0^2, p-1, eta) * 2*z0 * g[ (i-1)*r + 1:r ] ) result <- result + inner * outer } result }) if ( qtyp==8 ) arl <- Vectorize(function(a, b) { # GL Nystroem with ()^2 plus sin() substitution if ( abs(h-a) < 1e-10 ) a_ <- 1 else a_ <- ( a - b^2/delta ) / ( h - b^2/delta ) b_ <- b / sqrt( delta * h ) m <- rdc + (1-l)*b_ eta <- l2 * h * (1-b_^2) * a_ if ( eta < 1e-10 ) eta <- 0 result <- 1 for ( i in 1:r ) { korr <- h * ( 1 - sin(z1[i])^2 ) / l^2 outer <- korr * w1[i] * dnorm( sin(z1[i]), mean=m, sd=sig) * cos(z1[i]) inner <- sum( w0 * dchisq( korr*z0^2, p-1, eta) * 2*z0 * g[ (i-1)*r + 1:r ] ) result <- result + inner * outer } result }) if ( qtyp==9 ) arl <- Vectorize(function(a, b) { # GL Nystroem with ()^2 plus tan() substitution if ( abs(h-a) < 1e-10 ) a_ <- 1 else a_ <- ( a - b^2/delta ) / ( h - b^2/delta ) b_ <- b / sqrt( delta * h ) m <- rdc + (1-l)*b_ eta <- l2 * h * (1-b_^2) * a_ if ( eta < 1e-10 ) eta <- 0 result <- 1 for ( i in 1:r ) { korr <- h * ( 1 - tan(z1[i])^2 ) / l^2 outer <- korr * w1[i] * dnorm( tan(z1[i]), mean=m, sd=sig) / cos(z1[i])^2 inner <- sum( w0 * dchisq( korr*z0^2, p-1, eta) * 2*z0 * g[ (i-1)*r + 1:r ] ) result <- result + inner * outer } result }) if ( qtyp==10 ) arl <- Vectorize(function(a, b) { # GL Nystroem with ()^2 plus sinh() substitution norm <- sinh(1) if ( abs(h-a) < 1e-10 ) a_ <- 1 else a_ <- ( a - b^2/delta ) / ( h - b^2/delta ) b_ <- b / sqrt( delta * h ) m <- rdc + (1-l)*b_ eta <- l2 * h * (1-b_^2) * a_ if ( eta < 1e-10 ) eta <- 0 result <- 1 for ( i in 1:r ) { korr <- h * ( 1 - (sinh(z1[i])/norm)^2 ) / l^2 outer <- korr * w1[i] * dnorm( sinh(z1[i])/norm, mean=m, sd=sig) * cosh(z1[i])/norm inner <- sum( w0 * dchisq( korr*z0^2, p-1, eta) * 2*z0 * g[ (i-1)*r + 1:r ] ) result <- result + inner * outer } result }) if ( qtyp %in% c(1, 6, 11, 12) ) arl <- Vectorize(function(a, b) { # collocation if ( abs(h-a) < 1e-10 ) a_ <- 1 else a_ <- ( a - b^2/delta ) / ( h - b^2/delta ) b_ <- b / sqrt( delta * h ) result <- 0 for ( i in 1:r ) { outer <- Tn( 2*a_-1, i-1 ) inner <- sum( Tn( b_, 0:(r-1)) * g[ (i-1)*r + 1:r ] ) result <- result + inner * outer } result }) if ( qtyp==4 ) arl <- Vectorize(function(a, b) { # Markov chain (Runger/Prabhu) #a <- sqrt(a) cE_ <- sqrt( cE * l/(2-l) ) w <- 2*cE_/( 2*r + 1 ) wl <- w^2 / l^2 ii <- function(ix, iy) (ix-r)^2 + iy^2 < cE_^2/w^2 Vf <- function(iy,jy) pchisq( (jy+.5)^2*wl, p-1, ncp=(iy*w)^2*l2) - as.numeric(jy>0)*pchisq( (jy-.5)^2*wl, p-1, ncp=(iy*w)^2*l2) Hf <- function(ix,jx) pnorm( (-cE_+(jx+1)*w-(1-l)*(-cE_+(ix+.5)*w) )/l, mean=delta) - pnorm( (-cE_+jx*w-(1-l)*(-cE_+(ix+.5)*w) )/l, mean=delta) CIRC <- as.vector( t(outer( 0:(2*r), 0:r, ii)) ) Vv <- Vf( a/w, 0:r ) Hv <- Hf( (b+cE_)/w-.5, 0:(2*r) ) dQ <- sum(CIRC) Qv <- as.vector( Vv %o% Hv ) Qv_ <- Qv[ CIRC ] result <- 1 + sum( Qv_ * g ) }) if ( qtyp %in% c(13, 14, 15, 16, 17) ) arl <- Vectorize(function(a, b) { # new GL designs gam <- 0 #if ( a > 0 ) gam <- b / sqrt( a * delta ) if ( a > 0 ) gam <- b / sqrt( a ) mij <- lsd + (1 - l ) * sqrt(a) * gam ncpij <- l2 * a * ( 1 - gam^2 ) norm <- sinh(1) result <- 1 for ( i in 1:r ) { if ( qtyp == 13 ) korr <- w0[i] * sqrt( z0[i] ) / l^2 if ( qtyp > 13 ) korr <- 2 * w0[i] * z0[i]^2 / l^2 wl <- z1 korr2 <- rep(1, r) if ( qtyp == 15 ) { wl <- sin( z1 ); korr2 <- cos( z1 ) } if ( qtyp == 16 ) { wl <- tan( z1 ); korr2 <- 1 / ( cos( z1 )^2 ) } if ( qtyp == 17 ) { wl <- sinh( z1 ) / norm; korr2 <- cosh( z1 ) / norm } if ( qtyp == 13 ) term <- korr * sum( w1 * dnorm( (sqrt(z0[i])*wl - mij)/l ) / l * dchisq( z0[i]*( 1 - wl^2 ) / l^2, p-1, ncpij ) * g[ (i-1)*r + 1:r ] * korr2 ) if ( qtyp > 13 ) term <- korr * sum( w1 * dnorm( (z0[i]*wl - mij)/l ) / l * dchisq( z0[i]^2*( 1 - wl^2 ) / l^2, p-1, ncpij ) * g[ (i-1)*r + 1:r ] * korr2 ) result <- result + term } result }) } arl }