--- title: "Introduction to the Successive Projection Algorithm" author: "Mark van der Loo" date: "`r Sys.Date()`" output: rmarkdown::html_vignette: toc: true vignette: > %\VignetteIndexEntry{Introduction to the Successive Projection Algorithm} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ## Introduction The successive projection algorithm (SPA) solves quadratic optimization problems under linear equality and inequality restrictions. That is, given a vector $\boldsymbol{x}$, find the vector $\boldsymbol{x}^*$ that minimizes the weighted Euclidian distance $$ (\boldsymbol{x}-\boldsymbol{x}^*)^T\boldsymbol{W}(\boldsymbol{x}-\boldsymbol{x}^*), $$ subject to $$ \boldsymbol{Ax}^*\leq \boldsymbol{b}. $$ Here, $\boldsymbol{W}$ is a diagonal matrix with positive weights. The system of restrictions can contain equality and/or inequality restrictions. ### Example Suppose we have the vecor $(x,y)=(0.8,-0.2)$, depicted by the black dot in the figure below. Furthermore, we have the demands that $$ \begin{array}{lcl} y &\geq& x\\ x &\geq& 1-y \end{array} $$ The regions where $y\geq x$ or $x\geq 1-y$ are indicated by the single-shaded regions in the figure. The area where both demands are satisfied is indicated by the doubly-shaded region. ```{r,echo=FALSE,fig.width=5,fig.height=5} x <- c(-0.5,2) y1 <- x y2 <- 1-x plot( x,y1,'l',xlim=c(-0.5,1.5),ylim=c(-0.5,1.5),xlab='x',ylab='y',lwd=1.8) lines(x,y2,lwd=2) polygon( x = c(-0.5, 1.5,-.5) , y = c(-0.5, 1.5,1.5) , density = 5 , border=NA , angle=-45 ) polygon( x = c(-0.5,1.5,1.5) ,y =c(1.5,-0.5,1.5) ,density=5 ,border=NA ) abline(h=0,v=0) points(0.8,-0.2,pch=16,cex=1.8) arrows(x0=0.8,y=-0.2,x1=1,y1=0,length=0.1,lwd=2) points(0.5,0.5,pch=16,col='grey',cex=1.8) arrows(x0=1,y0=0,x1=0.5,y1=0.5,length=0.1,lwd=2) text(-0.15,-0.3,labels="y = x") text(1.05,-0.3,labels="y = 1 - x") ``` To find a solution, the successive projection algorithm projects the start vector iterativelty on the borders of the convex region that is defined by the linear inequalities. In the figure this is indicated by the arrows. The solution is a point on or numerically very near the border of the allowed region. When all weights on the diagonal of $\boldsymbol{W}$ are equal, projections are orthogonal, as shown in the figure. If the weights differ, the direction of projections will be scaled accordingly. ### Optimization in R In the `lintools` package, all inequalities must be written in the $\leq$-form. So first note that the above constraints can be written as $$ \left(\begin{array}{cc} 1 & -1\\ 1 & 1\\ \end{array}\right) \left(\begin{array}{c} x\\y \end{array}\right) \leq \left(\begin{array}{c} 0\\ 1 \end{array}\right) $$ So we formulate the problem with the `lintools` package as follows. ```{r} library(lintools) x <- c(0.8,-0.2) A <- matrix(c(1,-1,-1,-1), byrow=TRUE, nrow=2) b <- c(0,-1) ``` The function `project` solves the problem for us. By passing `neq=0` we tell `project` that every restriction is an inequality (setting `neq>0` means that the first `neq` restrictions are equalities). ```{r} project(x=x,A=A,b=b,neq=0) ``` The result is a list with the following elements. - `x` : the optimized vector - `status`: A status indicator: 0 means that the algorithm converged to a solution. (1= not enough memory, 2= divergence detected, 3 = maximum nr of iterations exceeded) - `tol` A measure of how far the final vector lies from the borders defined by the constaint (it is the $L_\infty$ distance to the valid region). - `iterations` The number of iterations performed. - `duration` The amount of time elapsed during optimization. - `objective` This is the weighted distance between the start vector and the solution. ## Sparse problems For problems where a great many coeffiecients need to be optimized under a large number of restrictions, it is possible to forumate the restrictions in sparse format. In the `lintools` package, the row-column-coefficient format is used. That is, in stead of defining the full matrix $\boldsymbol{A}$ as in the previous example, we set up a `data.frame` with columns - `row` : the row number - `column`: the column number - `coef` : the coefficient Of course, only non-zero coefficients need to be listed. As a -rather simple- example, we define the same problem as above, but now in a sparse manner. ```{r} A <- data.frame( row = c(1,1,2,2) ,col = c(1,2,1,1) ,coef = c(1,-1,-1,-1) ) b <- c(0,-1) x <- c(0.8,-0.2) ``` Solving is done with the `sparse_project` function. ```{r} sparse_project(x, A=A, b=b, neq=0) ``` We have been able to solve problems with up to $\sim$ 6 milion variables and hundreds of thousands of linear (in)equality restrictions with the algorithm as implemented in this package. ### Reusing sparsely defined restrictions The `sparse_project` function performs the following steps: 1. It creates a particular sparse representation of the restrictions 2. It solves the minimization problem 3. It gathers results and returns them to the user. Step 1 takes a little bit of time (not much) but if you need to do a lot of optimizations it may pay to do it once and reuse the representation. This can be done as follows, using the same definition of sparse constraints as in the previous subsection. First, create an object of class `sparse_constraints`. ```{r} sc <- sparse_constraints(A,b,neq=0) ``` Now, using its `project` method, we can optimize from multiple starting points, for example: ```{r} sc$project(x=c(0.8,-0.2)) # the same problem, but with differing weights sc$project(x=c(0.8,-0.2),w=c(1,10)) ``` ## References The algorithm implemented here may have been invented a number of times. The earliest reference known to this author is - Hildreth, C. (1957), A quadratic programming procedure. _Naval research logistics quarterly_ 4, pp 79-85. The method was more recently discussed in the context of restricted imputation methodology by - Pannekoek, J. and Zhang, L.-C. (2012), Optimal adjustments for inconsistency in imputed data. _[Discussion Paper 201219](https://www.cbs.nl/-/media/imported/documents/2012/38/2012-19-x10-pub.pdf) Statistics Netherlands. ## Some words on the rspa package Users of the `rspa` package will no doubt recognize the algorithm and the `sparse_constraints` object. We chose to separate the functionality from `rspa` to be able to reuse the successive projection algorithm for multiple purposes, without depending on the `editrules` package. In the future, `rspa` will depend on `lintools` with guaranteed backward compatibility.