## Comparison of CHSH and J for recent Bell experiments, 
## together with optimally noise-reduced versions of both.
## Theory: https://pub.math.leidenuniv.nl/~gillrd/Peking/Peking_4.pdf

## In short: assume four multinomial samples, 
## estimate covariance matrix of estimated relative frequencies, 
## use sample deviations from no-signalling to optimally reduce 
## the noise in the estimate of Bell's S or Eberhard's J

## AKA: generalized least squares

############# MUNICH #############

## The basic data, four 2x2 tables

table11 <- matrix(c(16, 4, 3, 13), 
    2, 2, byrow = TRUE,
    dimnames = list(Alice = c("+", "-"), Bob = c("+", "-")))
table12 <- matrix(c(11, 4, 2, 17), 
    2, 2, byrow = TRUE,
    dimnames = list(Alice = c("+", "-"), Bob = c("+", "-")))
table21 <- matrix(c(19, 4, 3, 16), 
    2, 2, byrow = TRUE,
    dimnames = list(Alice = c("+", "-"), Bob = c("+", "-")))    
table22 <- matrix(c(4, 22, 10, 2), 
    2, 2, byrow = TRUE,
    dimnames = list(Alice = c("+", "-"), Bob = c("+", "-")))

table11
##      Bob
## Alice  +  -
##     + 16  4
##     -  3 13
table12
##      Bob
## Alice  +  -
##     + 11  4
##     -  2 17
table21
##      Bob
## Alice  +  -
##     + 19  4
##     -  3 16
table22
##      Bob
## Alice  +  -
##     +  4 22
##     - 10  2
## Check of the total number of trials

# "The number of valid trials is N = 150"
sum(table11) + sum(table12) + sum(table21) + sum(table22)
## [1] 150
## The same data now in one 4x4 table

tables <- cbind(as.vector(t(table11)), as.vector(t(table12)), as.vector(t(table21)), as.vector(t(table22)))
dimnames(tables) = list(outcomes = c("++", "+-", "-+", "--"), 
                      settings = c(11, 12, 21, 22))
tables
##         settings
## outcomes 11 12 21 22
##       ++ 16 11 19  4
##       +-  4  4  4 22
##       -+  3  2  3 10
##       -- 13 17 16  2
## The total number of trials for each setting pair

Ns <- apply(tables, 2, sum)
Ns
## 11 12 21 22 
## 36 34 42 38
## observed relative frequencies, one 4x4 matrix

rawProbsMat <- tables / outer(rep(1,4), Ns)
rawProbsMat
##         settings
## outcomes         11         12         21         22
##       ++ 0.44444444 0.32352941 0.45238095 0.10526316
##       +- 0.11111111 0.11764706 0.09523810 0.57894737
##       -+ 0.08333333 0.05882353 0.07142857 0.26315789
##       -- 0.36111111 0.50000000 0.38095238 0.05263158
## Convert the relative frequencies to one vector of length 16

VecNames <- as.vector(t(outer(colnames(rawProbsMat), rownames(rawProbsMat), paste, sep = "")))
rawProbsVec <- as.vector(rawProbsMat)
names(rawProbsVec) <- VecNames

VecNames
##  [1] "11++" "11+-" "11-+" "11--" "12++" "12+-" "12-+" "12--" "21++" "21+-"
## [11] "21-+" "21--" "22++" "22+-" "22-+" "22--"
rawProbsVec
##       11++       11+-       11-+       11--       12++       12+- 
## 0.44444444 0.11111111 0.08333333 0.36111111 0.32352941 0.11764706 
##       12-+       12--       21++       21+-       21-+       21-- 
## 0.05882353 0.50000000 0.45238095 0.09523810 0.07142857 0.38095238 
##       22++       22+-       22-+       22-- 
## 0.10526316 0.57894737 0.26315789 0.05263158
## Building up the 4 no-signalling constraints, combined in one 16 x 4 matrix "NS"

Aplus <- c(1, 1, 0, 0)
Aminus <- - Aplus
Bplus <- c(1, 0, 1, 0)
Bminus <- - Bplus
zero <- c(0, 0, 0, 0)
NSa1 <- c(Aplus, Aminus, zero, zero)
NSa2 <- c(zero, zero, Aplus, Aminus)
NSb1 <- c(Bplus, zero, Bminus, zero)
NSb2 <- c(zero, Bplus, zero, Bminus)
NS <- cbind(NSa1 = NSa1, NSa2 = NSa2, NSb1 = NSb1, NSb2 = NSb2)
rownames(NS) <- VecNames
NS
##      NSa1 NSa2 NSb1 NSb2
## 11++    1    0    1    0
## 11+-    1    0    0    0
## 11-+    0    0    1    0
## 11--    0    0    0    0
## 12++   -1    0    0    1
## 12+-   -1    0    0    0
## 12-+    0    0    0    1
## 12--    0    0    0    0
## 21++    0    1   -1    0
## 21+-    0    1    0    0
## 21-+    0    0   -1    0
## 21--    0    0    0    0
## 22++    0   -1    0   -1
## 22+-    0   -1    0    0
## 22-+    0    0    0   -1
## 22--    0    0    0    0
## Build the 16x16 estimated covariance matrix of the 16 observed relative frequencies

cov11 <- diag(rawProbsMat[ , "11"]) - outer(rawProbsMat[ , "11"], rawProbsMat[ , "11"])
cov12 <- diag(rawProbsMat[ , "12"]) - outer(rawProbsMat[ , "12"], rawProbsMat[ , "12"])
cov21 <- diag(rawProbsMat[ , "21"]) - outer(rawProbsMat[ , "21"], rawProbsMat[ , "21"])
cov22 <- diag(rawProbsMat[ , "22"]) - outer(rawProbsMat[ , "22"], rawProbsMat[ , "22"])

Cov <- matrix(0, 16, 16)
rownames(Cov) <- VecNames
colnames(Cov) <- VecNames
Cov[1:4, 1:4] <- cov11/Ns["11"]
Cov[5:8, 5:8] <- cov12/Ns["12"]
Cov[9:12, 9:12] <- cov21/Ns["21"]
Cov[13:16, 13:16] <- cov22/Ns["22"]

## The vector "S" is used to compute the CHSH statistic "CHSH"
## The sum of the first three sample correlations minus the fourth

S <- c(c(1, -1, -1 ,1), c(1, -1, -1 , 1), c(1, -1, -1, 1), - c(1, -1, -1, 1))
names(S) <- VecNames
CHSH <- sum(S * rawProbsVec)
CHSH
## [1] 2.609047
## Compute the estimated variance of the CHSH statistic, 
## its estimated covariances with the observed deviations from no-signalling,
## and the 4x4 estimated covariance matrix of those deviations.
## We'll later also need the inverse of the latter.

varS <- t(S) %*% Cov %*% S
covNN <- t(NS) %*% Cov %*% NS
covSN <- t(S) %*% Cov %*% NS
covNS <- t(covSN)

InvCovNN <- solve(covNN)

## Estimated variance of the CHSH statistic, 
## and estimated variance of the optimally "noise reduced" CHSH statistic.

varCHSH <- varS
varCHSHopt <- varS - covSN %*% InvCovNN %*% covNS

## The variance, and the improvement as ratio of standard deviations

varS
##           [,1]
## [1,] 0.0617252
sqrt(varCHSH / varCHSHopt)
##         [,1]
## [1,] 1.01325
## The coefficients of the noise reduced CHSH statistic and the resulting improved estimate

Sopt <- S - covSN %*% InvCovNN %*% t(NS)
Sopt
##           11++      11+-       11-+ 11--      12++      12+-      12-+
## [1,] 0.7656009 -1.279026 -0.9553731    1 0.9320339 -0.720974 -1.346992
##      12--      21++      21+-      21-+ 21--       22++     22+-     22-+
## [1,]    1 0.8837271 -1.071646 -1.044627    1 -0.5813619 1.071646 1.346992
##      22--
## [1,]   -1
CHSHopt <- sum(Sopt * rawProbsVec)
CHSHopt
## [1] 2.582261
## p-values assuming approximate normality for testing CHSH inequality

pnorm((CHSH - 2)/ sqrt(varCHSH), lower.tail = FALSE)
##             [,1]
## [1,] 0.007114475
pnorm((CHSHopt - 2)/ sqrt(varCHSHopt), lower.tail = FALSE)
##             [,1]
## [1,] 0.008782296
## Now we repeat for the Eberhard J statistic
## First, the coefficients in the vector "J"
## and the observed value of the statistic

J <- c(c(1, 0, 0 ,0), c(0, -1, 0 ,0), c(0, 0, -1, 0), c(-1, 0, 0, 0))
names(J) <- VecNames
sum(J * rawProbsVec)
## [1] 0.1501057
## Next, its estimated variance and resulting p-value

varJ <- t(J) %*% Cov %*% J
sum(J * rawProbsVec) / sqrt(varJ)
##          [,1]
## [1,] 1.270007
pnorm(sum(J * rawProbsVec) / sqrt(varJ), lower.tail = FALSE)
##           [,1]
## [1,] 0.1020411
## The covariances between J and the observed deviations from no-signaling
## The variance of the usual estimate of J and of the improved estimate of J
## The improvement as a ration of standard deviations

covJN <- t(J) %*% Cov %*% NS
covNJ <- t(covJN)
varJopt <- varJ - covJN %*% InvCovNN %*% covNJ

varJ
##            [,1]
## [1,] 0.01396953
sqrt(varJ / varJopt)
##         [,1]
## [1,] 1.92813
## The coefficients of an improved estimataor of Eberhard's J

Jopt <- J - covJN %*% InvCovNN %*% t(NS)
Jopt
##             11++       11+-       11-+ 11--        12++       12+-
## [1,] -0.05859978 -0.5697565 -0.4888433    0 -0.01699153 -0.4302435
##           12-+ 12--        21++       21+-       21-+ 21--      22++
## [1,] -0.586748    0 -0.02906821 -0.5179115 -0.5111567    0 0.1046595
##           22+-     22-+ 22--
## [1,] 0.5179115 0.586748    0
## Observed estimate of J, and improved estimate of J

sum(J * rawProbsVec)
## [1] 0.1501057
sum(Jopt * rawProbsVec)
## [1] 0.1455654
## p-values based on J and on improved J
## Note that the p-value based on improved J is the same as that of improved CHSH

pnorm(sum(J * rawProbsVec) / sqrt(varJ), lower.tail = FALSE)
##           [,1]
## [1,] 0.1020411
pnorm(sum(Jopt * rawProbsVec) / sqrt(varJopt), lower.tail = FALSE)
##             [,1]
## [1,] 0.008782296
## The p-value of the optimized J got much better
## The procedure for CHSH made things slightly worse
## The deviation from no-signalling is small; it is responsible for these small changes