# confidence interval in quadratic programming

This is a quadratic programming problem coming from constrained linear regression

$A b + \varepsilon = y$ and $B^T b >0$

then minimize $\varepsilon^T \varepsilon=(A b - y)^T (A b - y)$ which is equivalent to $min(-d^T b + 1/2 b^T D b)$ where $D=2A^TA$, and $2y^TA=d^T$

I would like to get confidence interval in a quadratic programming problem. I am using R. For the curious I provide the below code where I would like to get the confidence interval of the estimated coefficients. I have tried bootstrapping but I am having issues with non-positive definite matrix, I am hoping another approach might be possible. Maybe just shock the estimates and calculate $min(-d^T b + 1/2 b^T D b)$ which is the cost function for quadratic program.

Amat = matrix(c(1,1,1,1,1,1,1,3,6,9,12,15,1,9,36,81,144,225,1,27,216,729,1728,3375), nrow=6)

Dmat = matrix( c(12,92,992,12152,92,992,12152,158600,992,12152,158600,2150552,12152,158600,2150552,29910872), nrow=4)
dvec = matrix(c(12,96.1128386,1086.2395754,13835.6202146), ncol=1)
Bmat = matrix( c(0,1,3,7,0,1,5,19,0,1,7,37,0,1,9,61,0,1,11,91,0,1,13,127,0,1,15,169,0,1,17,217,0,1,19,271,0,1,21,331,0,1,23,397,0,1,25,469,0,1,27,547,0,1,29,631,0,1,31,721,0,1,33,817,0,1,35,919,0,1,37,1027,0,1,39,1141,0,1,41,1261), nrow=4)
bvec = matrix(c(0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0), nrow=1)

bEst <- solve.QP(Dmat, dvec, Bmat, bvec)

bEst$solution [1] 0.879426809 0.026242400 -0.005319406 0.00035462  • You have nominally stated a deterministic quadratic programming problem, which is specified by the inputs Dmat, dvec, Amat, bvec. This problem has a deterministic solution. Where does the randomness come in? You need to provide a context. Nov 20, 2015 at 18:27 • Thanks, I added further information. Let me know if it is more clear now – adam Nov 23, 2015 at 11:15 • I believe the difficulty here is that for components of$B^\prime b$that are close to zero, the CI needs to terminate at zero itself. Related to this--and even thornier--is the problem of determining a CI for the components that have been forced to zero by the restriction. No simple analytical procedure is going to work and developing the asymptotics looks difficult. Consider a bootstrapped CI or something generic like that, assuming you have a much larger dataset. – whuber Dec 1, 2015 at 16:39 • this is my dataset. 6 points, yo = c(0.9978865,1.0054508,0.8994358,0.820207,0.9638653,1.3131546), b=c(1,3,6,9,12,15) – adam Dec 1, 2015 at 16:47 • In general one thing you could do is to run Monte Carlo simulation on a sub sample; iff you don't have time series data. In any case you have too few data. Sep 5, 2016 at 15:39 ## 1 Answer yo = c(0.9978865,1.0054508,0.8994358,0.820207,0.9638653,1.3131546) e <- yo-Amat%*%bEst$solution
R2 <- 1-crossprod(e,e)/crossprod(yo-mean(yo),yo-mean(yo))

degreesOfFreedom <- (nRow - polyOrder - 1)
paramVar <- diag(solve(Dmat*0.5))*crossprod(e, e)/degreesOfFreedom
bEst$paramVar <- paramVar bEst$tValue <- bEst$solution/sqrt(paramVar) bEst$pValue <- (1- pt(abs(bEst$tValue), df=degreesOfFreedom ))*2  I propose to use a solution similar to linear regression such as$(A^TA)^{-1}\sigma$where$\sigma$is error standard deviation. page 18 of nonlinear regression notes has similar formula • This answer looks incorrect: a student$t\$ distribution is not the right one to use when any of the constraints holds (or is close to holding). Could you explain why you think this is correct?
– whuber
Oct 8, 2019 at 13:06