Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

It happened so I need to work with an algorithm for solving QP problems. The main issue here is that I can't find any references to this algorithms (at least no references in sources in english). There is something on Wikipedia, but rather obscure. I was able to find a couple of books where this algorithm is refered to as 'Sigular Points Method' (exact translation from Russian to English), but I want to know if there is globally accepted name for this algorithm, and maybe existing implementations in Matlab, Maple or SciPy.

The algorithm looks like some kind of 'Simplex Method', but for QP problems, and it always converges in finite number of steps.

Brief algorithm description:

We have the following QP problem:

\begin{equation} f(x) \to min, \quad x \in D, \tag{1} \end{equation}

\begin{equation} D = \{ x \in R^n \, | \, Ax \leq b \}, \tag{2} \end{equation}

where $$ f: R^n \to R, \quad f(x) = \frac{1}{2} <Cx, x> + <c, x> $$ $C \in R(n, n)$ is symmetric positive semidefinite matrix, $c \in R^n$, $A \in R(m, n)$, $b \in R^m$.

The notation $<a, b>$ stands for dot-product, $R(m, n)$ is a set of all $(m \times n)$-dimensional matrices.

Next, let's define an analog of simplex vertices we iterate over in 'Simplex Method'.

Definition 1. Point $\hat x$ is called singular for the problem (1), (2), if exists a set $I \subset \{ 1, \ldots, m \} $, such as $\hat x$ is a solution for the following problem:

\begin{equation} f(x) \to min, \quad x \in D_I, \tag{3} \end{equation}

\begin{equation} D_I = \{ x \in R^n \, | \, <a_i, x> = b_i, \, i \in I \} \tag{4} \end{equation}

(cases when $I = \emptyset$ and $I = \{ 1, \ldots, m \}$ are both included).

Ok, now let's formulate a theorem that will shed some light on the actual algorithm.

Theorem 1. For every symmetric positive semidefinite matrix $ C \in R(n, n) $, and every $c \in R^n$, $A \in R(m, n)$ and $b \in R^m$ there is only a finite set of singular points for problem (1), (2), and solution of this problem is its singular point.

At this step it is clear that we could iterate over singular points, and pick the one where the value of the objective function is minimal, and that one singular point will be the solution.

Of course the iteration process is rather complex. It builds a sequence of singular points $x^1, x^2, \ldots, x^k$ such that $ f( x^{i+1} ) < f(x^i) $. Conjugate gradien method with some modifications for semidefinite matrices could be used to solve the problem (3), (4) (for quadratic functions it converges in finite number of steps).

I could provide even more details, but I'm pretty sure that at this point the algorithm could be identified.


The main idea of algorithm:

Suppose, we have some point $x_0 \in D$.

1) Select the constraints from (2) that is satisfied as an equations for $x_0$. This constraints define some face $D_0$ of a polyhedron described by (2).

2) Find $\hat x_0 = min f(x), \, x \in D_0$.

$\hat x_0$ is either the solution for the problem (1), (2) or it is possible to switch to another face, and start with step 1).

By the way, there is no non-negative constraints on problem variables

share|cite|improve this question

I'm not sure if this is what you ate thinking of, but there is Wolfe's method, which is an extended simplex method. A description can be found here and in Winston, Operations Research: Applications and Algorithms.

share|cite|improve this answer
I looked over the details on Wolfe's method, and I don't think it is it. Wolfe's method is only applied to QPP with non-negative problem variables, while there is no such restrictions in a problem described above. And in general, Wolfe's method and algorithm described above doesn't seem similar to me. – Artem Pyanykh Nov 29 '12 at 10:20
@Deshene I wasn't sure, but I do know this is suitable. Your problem can be rewritten in this form by setting $x=x_+ - x_-$, where $x_+$ and $x_-$ are both positive. All the best for finding the algorithm. – Daryl Nov 29 '12 at 10:27
@Deshene I just read your edit. Again, I'm not certain of the implementation, but the quadprog function in the MATLAB optimisation toolbox may be helpful to you. – Daryl Nov 29 '12 at 10:33
about $x = x_+ - x_-$ - yes, this is a common technique, but it increases problem dimensionality twice, which is not that good. Quadprog, though, could be really helpful, thanks! – Artem Pyanykh Nov 29 '12 at 11:29
up vote 0 down vote accepted

It seems like I have successfully identified the method described above. If I got it right, it is called an Active Set method for quadratic programming. In particular, as @Daryl menthioned, there is a 'quadprog' function in Matlab, which solves QPP, and supports 'active set' algorithm.

share|cite|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.