From release 8.1.0.56 we initialize support for .NET Core, the cross-platform implementation of .NET.

MOSEK for .NET Core is distributed as a platform-independent NuGet package, which can be downloaded directly from our website. See the documentation for .NET APIs for installation instructions.

## Friday, June 29, 2018

## Friday, June 22, 2018

### MOSEK at ISMP 2018

Here is a complete schedule of our talks at ISMP 2018 in chronological order:

Check the program for any last-minute changes.

The session on Mixed-Integer Conic Optimization (Mon, 5:00PM) is organized by Sven Wiese. The full program of that session is:

Speaker | Title | Session | Time | Room |
---|---|---|---|---|

Sven Wiese | The Mixed-integer Conic Optimizer in MOSEK | Mixed-Integer Conic Optimization | Mon, 5:00PM | A. DURKHEIM, Bld. A |

Henrik Friberg | Projection and presolve in MOSEK: exponential and power cones | Theory and algorithms in conic linear programming 1 | Tue, 8:30AM | S. LC5, Bld. L |

Joachim Dahl | Extending MOSEK with exponential cones | Theory and algorithms in conic linear programming 2 | Wed, 8:30AM | S. 20, Bld. G |

Erling Andersen | MOSEK version 9 | Progress in Conic and MIP Solvers | Wed, 3:15PM | A. PITRES, Bld. O |

MichaĆ Adamaszek | Exponential cone in MOSEK: overview and applications | Relative Entropy Optimization I | Fri, 3:15PM | S. LC5, Bld. L |

Check the program for any last-minute changes.

The session on Mixed-Integer Conic Optimization (Mon, 5:00PM) is organized by Sven Wiese. The full program of that session is:

- Lucas Letocart, Exact methods based on SDP for the k-item quadratic knapsack problem
- Tristan Gally, Knapsack Constraints over the Positive Semidefinite Cone
- Sven Wiese, The Mixed-integer Conic Optimizer in MOSEK

## Tuesday, June 12, 2018

### Elementary intro to infeasibility certificates

We occasionally receive support questions asking about the meaning and practical use of infeasibility certificates, usually when the user expected the problem to be feasible. While an infeasibility certificate is a well-defined mathematical object based on Farkas' lemma, it is not intuitive for everyone how to use it to address the basic question of "What is wrong with my problem formulation?". We will try to explain this on a very simple example.

In this post we don't even consider optimization, but restrict to elementary linear algebra. Suppose we want to find a solution to a system of linear equations:

$$

\begin{array}{llllllllllll}

& 2x_1 & - & x_2 & + & x_3 & - & 3x_4 & + & x_5 & = & 1, \\

& & & x_2 & -& 2x_3 & & & +& x_5 & = & -1,\\

& 3x_1 & & & - & x_3 & - & x_4 & & & = & 2, \\

& 2x_1 & + & x_2 & - & 3x_3 & - & 3x_4 & + & 3x_5 & = & -0.5.

\end{array}

$$

This system of linear equations is in fact infeasible (has no solution). One way to see it is to multiply the equations by coefficients given on the right-hand side and add them up:

$$

\begin{array}{lllllllllllll}

& 2x_1 & - & x_2 & + & x_3 & - & 3x_4 & + & x_5 & = & 1 & / \cdot (-1) \\

& & & x_2 & -& 2x_3 & & & +& x_5 & = & -1 & / \cdot (-2) \\

& 3x_1 & & & - & x_3 & - & x_4 & & & = & 2 & / \cdot 0 \\

& 2x_1 & + & x_2 & - & 3x_3 & - & 3x_4 & + & 3x_5 & = & -0.5 & / \cdot 1 \\

\hline

& & & & & & & & & 0 & = & 0.5. &

\end{array}

$$

In this post we don't even consider optimization, but restrict to elementary linear algebra. Suppose we want to find a solution to a system of linear equations:

$$

\begin{array}{llllllllllll}

& 2x_1 & - & x_2 & + & x_3 & - & 3x_4 & + & x_5 & = & 1, \\

& & & x_2 & -& 2x_3 & & & +& x_5 & = & -1,\\

& 3x_1 & & & - & x_3 & - & x_4 & & & = & 2, \\

& 2x_1 & + & x_2 & - & 3x_3 & - & 3x_4 & + & 3x_5 & = & -0.5.

\end{array}

$$

This system of linear equations is in fact infeasible (has no solution). One way to see it is to multiply the equations by coefficients given on the right-hand side and add them up:

$$

\begin{array}{lllllllllllll}

& 2x_1 & - & x_2 & + & x_3 & - & 3x_4 & + & x_5 & = & 1 & / \cdot (-1) \\

& & & x_2 & -& 2x_3 & & & +& x_5 & = & -1 & / \cdot (-2) \\

& 3x_1 & & & - & x_3 & - & x_4 & & & = & 2 & / \cdot 0 \\

& 2x_1 & + & x_2 & - & 3x_3 & - & 3x_4 & + & 3x_5 & = & -0.5 & / \cdot 1 \\

\hline

& & & & & & & & & 0 & = & 0.5. &

\end{array}

$$

We get an obvious contradiction, which proves the system has no solution. The vector

$$ y = [-1, -2, 0, 1] $$

of weights used above is therefore a proof (certificate) of infeasibility. MOSEK produces such a certificate automatically. Here is a simple script in MATLAB that computes precisely that vector:

As output we get:

Here are a few comments:

###

In the example above $y_3=0$ which means that the third equation does not matter: infeasibility is caused already by some combination of the 1st, 2nd and 4th equation. In many practical situations the infeasibility certificate $y$ will have very few nonzeros, and those nonzeros determine a subproblem (subset of equations) which alone cause infeasibility. The user can configure MOSEK to print an infeasibility report, which in our example will look like:

As output we get:

Here are a few comments:

- General theory guarantees that the dual variable y is a convenient place that can store the certificate.
- When a system of equations has no solution then an appropriate certificate is guaranteed to exist. This is a basic variant of Farkas' lemma, but it should be intuitively clear: your favorite method of solving linear systems (for instance Gaussian elimination) boils down to taking successive linear combinations of equations, a process which ends up either with a solution or with an "obvious" contradiction of the form $0=1$.

### Using the certificate to debug a model

This report is nothing else than a listing of equations with nonzero entries in $y$, and $y$ is the difference between Dual lower and Dual upper. Analyzing this set (which hopefully is much smaller than the full problem) can help locate a possible modeling error which makes the problem infeasible.

To conclude, let us now phrase the above discussion in matrix notation. A linear equation system

$$ Ax=b $$

is infeasible if and only if there is a vector $y$ such that

$$ A^Ty = 0 \quad \mbox{and}\quad b^Ty \neq 0. $$

The situation is analogous for linear problems with inequality constraints. We will look at an example of that in another blog post.

Labels:
infeasibility

## Thursday, May 31, 2018

### Perspective functions (and Luxemburg norms)

If $\varphi:\mathbb{R}_+\to\mathbb{R}$ is a convex function then its perspective function $\tilde{\varphi}:\mathbb{R}_+\times\mathbb{R}_+\to\mathbb{R}$ is defined as

$$\tilde{\varphi}(x,y) = y\varphi(\frac{x}{y}).$$

Moreover, under these conditions, the epigraph of the perspective function

$$\left\{(t,x,y)~:~t\geq y\varphi(\frac{x}{y})\right\}$$

(or, to be precise, its appropriate closure) is a convex cone. Here are some familiar examples:

We bring this up in connection with a series of blogposts by Dirk Lorenz here and here. For a monotone increasing, convex, nonnegative $\varphi$ with $\varphi(0)=0$ he defines a norm on $\mathbb{R}^n$ via

$$\|x\|_\varphi=\inf\left\{\lambda>0~:~\sum_i\varphi\left(\frac{|x_i|}{\lambda}\right)\leq 1\right\},$$

and we can ask when the norm bound $t\geq \|x\|_\varphi$ is conic representable. The answer is: if the epigraph of the perspective function $\tilde{\varphi}$ is representable, then so is the epigraph of $\|\cdot\|_\varphi$. The reason is that the inequality

$$\sum_i\varphi\left(\frac{|x_i|}{t}\right)\leq 1$$

is equivalent to

$$

\begin{array}{ll}

w_i\geq |x_i| & i=1,\ldots,n,\\

s_i\geq t\varphi\left(\frac{w_i}{t}\right) & i=1,\ldots,n,\\

t=\sum_i s_i. &

\end{array}

$$

That covers for example $\varphi(x)=x^2$, $\varphi(x)=x^p$ ($p>1$), $\varphi(x)=\exp(x)-1$, $\varphi(x)=x\log(1+x)$ and so on.

$$\tilde{\varphi}(x,y) = y\varphi(\frac{x}{y}).$$

Moreover, under these conditions, the epigraph of the perspective function

$$\left\{(t,x,y)~:~t\geq y\varphi(\frac{x}{y})\right\}$$

(or, to be precise, its appropriate closure) is a convex cone. Here are some familiar examples:

- $\varphi(x)=x^2$. Then $\tilde{\varphi}(x,y)=\frac{x^2}{y}$, familiar to some as quad-over-lin. The epigraph of $\tilde{\varphi}$, described by $ty\geq x^2$, is the Lorentz cone (rescaled rotated quadratic cone).
- $\varphi(x)=x^p$. Then $\tilde{\varphi}(x,y)=\frac{x^p}{y^{p-1}}$ and the epigraph of $\tilde{\varphi}$, described equivalently by $t^{1/p}y^{1-1/p}\geq |x|$ is the 3-dimensional power cone (with parameter $p$).
- $\varphi(x)=\exp(x)$. Then the epigraph $t\geq y\exp(x/y)$ is the exponential cone.
- $\varphi(x)=x\log(x)$. Then the epigraph $t\geq x\log(x/y)$ is the relative entropy cone.

We bring this up in connection with a series of blogposts by Dirk Lorenz here and here. For a monotone increasing, convex, nonnegative $\varphi$ with $\varphi(0)=0$ he defines a norm on $\mathbb{R}^n$ via

$$\|x\|_\varphi=\inf\left\{\lambda>0~:~\sum_i\varphi\left(\frac{|x_i|}{\lambda}\right)\leq 1\right\},$$

and we can ask when the norm bound $t\geq \|x\|_\varphi$ is conic representable. The answer is: if the epigraph of the perspective function $\tilde{\varphi}$ is representable, then so is the epigraph of $\|\cdot\|_\varphi$. The reason is that the inequality

$$\sum_i\varphi\left(\frac{|x_i|}{t}\right)\leq 1$$

is equivalent to

$$

\begin{array}{ll}

w_i\geq |x_i| & i=1,\ldots,n,\\

s_i\geq t\varphi\left(\frac{w_i}{t}\right) & i=1,\ldots,n,\\

t=\sum_i s_i. &

\end{array}

$$

That covers for example $\varphi(x)=x^2$, $\varphi(x)=x^p$ ($p>1$), $\varphi(x)=\exp(x)-1$, $\varphi(x)=x\log(1+x)$ and so on.

Figure: smallest $(\exp(x)-1)$-Luxemburg-norm disk containing a random set of points. |

## Wednesday, May 9, 2018

### New Modeling Cookbook

We released a new, revised version of the MOSEK Modeling Cookbook

where you can find practical modeling examples and some theory behind:

- linear optimization,
- conic quadratic optimization,
- the power cone,
- the exponential cone (relative entropy cone),
- semidefinite optimization,
- duality in linear and conic optimization,
- mixed-integer optimization,
- quadraticaly constrained optimization,
- and numerical issues.

The Modeling Cookbook is a general compendium of practical conic optimization. It is a generic mathematical publication without any code samples or references to the MOSEK API, although it covers areas of optimization that are supported by MOSEK 8 or the upcoming MOSEK 9.

## Monday, March 26, 2018

### Progress on the pooling problem and new reformulation techniques

On the 8th of March 2018, an interesting paper on the pooling problem landed on arxiv. The authors take the pq-formulation, known to have the strongest continuous relaxation, and strengthen it further it by adding convex hull-describing valid inequalities for subproblems -- one for each (product, attibute, pool)-combination. Two of the inequalities are linear, one is conic quadratic, and the final is general convex. We took up the challenge and also cast the latter on conic quadratic form. Here is what we learned.

The inequality is this:

$$

\label{eq:cut}

\overline{\beta}(\overline{\gamma} x - u) + h(y, v) \leq \overline{\beta}(\overline{\gamma} - t),

$$

valid on $y > 0$ when

$$

h(y,v) := A y + B g(y,v),\quad g(y,v) := \frac{yv}{y+v},\quad v \geq 0,

$$

where $A = \overline{\gamma}-\underline{\gamma} > 0$ and $B = \underline{\gamma} < 0$ are coefficients. The challenge is first to put this on conic quadratic form, and then to extend its validity to all of $y \in \mathbb{R}$, while staying tight on $y > 0$. On the first challenge of conic representability, our analysis led to a small surprise, namely, the possibility to represent harmonic means (see this blog post). The inequality $\eqref{eq:cut}$ is thus representable on $y > 0$ by $\overline{\beta}(\overline{\gamma} x - u) + h \leq \overline{\beta}(\overline{\gamma} - t)$ and

$$

\label{eq:repr:ypos}

h = A y + B g,\quad \left(\begin{matrix}0.5 (y-g)\\ y+v\\ y\end{matrix}\right) \in Q_r^3,\quad v \geq 0.

$$

We stress that the conic form is a proof of convexity. Next is the challenge of extending this representation to be valid on all of $y \in \mathbb{R}$, while staying tight on $y > 0$. That is, as proven in the paper, we need to extend the representation such that $h \leq 0$ on $y \leq 0$. Notably, however, the current representation in $\eqref{eq:repr:ypos}$ fails this criterion as shown by the graph of $h(y, 5)$ where $A=1$ and $B=-3$.

For this task we deploy another reformulation technique which, in lack of better names, may be denoted as

$$

\min_{s \in \mathbb{R}} h(y+s,v) = \begin{cases}

h(y+\overline{s},v), & {\rm if\ } y_{min} \geq y+\overline{s}, \\

h(y_{min},v), & {\rm if\ } y+\underline{s} \leq y_{min} \leq y+\overline{s},\\

h(y+\underline{s},v), & {\rm if\ } y+\underline{s} \geq y_{min}.

\end{cases}

$$

The visual interpretation of this is that we cut the graph of the function in two parts at $y_{min}$, move the two parts away from each other along the y-axis by amounts determined by $\underline{s}$ and $\overline{s}$, and finally fill in the gap with the extremum value $h(y_{min},v)$. In case of $\eqref{eq:repr:ypos}$ we use $\underline{s} = 0$ and $\overline{s} = \infty$ to allow $h$ to take smaller values of $h(y,v)$ at any higher value of $y$. In the graphed example of $h(y, 5)$ above, we get that the lower bound on $h$ is relaxed to the left of the extremum.

As $h(0,v) = 0$ for all $v \geq 0$, we get the desired property that $h \leq 0$ is possible on $y \leq 0$. In conclusion, the globally valid conic quadratic representation of the convex inequality we started out from, is given by $\overline{\beta}(\overline{\gamma} x - u) + h \leq \overline{\beta}(\overline{\gamma} - t)$ and

$$

h = A \hat{y} + B g,\quad \left(\begin{matrix}0.5 (\hat{y}-g)\\ \hat{y}+v\\ \hat{y}\end{matrix}\right) \in Q_r^3,\quad v \geq 0, \quad \hat{y} \geq y.

$$

The inequality is this:

$$

\label{eq:cut}

\overline{\beta}(\overline{\gamma} x - u) + h(y, v) \leq \overline{\beta}(\overline{\gamma} - t),

$$

valid on $y > 0$ when

$$

h(y,v) := A y + B g(y,v),\quad g(y,v) := \frac{yv}{y+v},\quad v \geq 0,

$$

where $A = \overline{\gamma}-\underline{\gamma} > 0$ and $B = \underline{\gamma} < 0$ are coefficients. The challenge is first to put this on conic quadratic form, and then to extend its validity to all of $y \in \mathbb{R}$, while staying tight on $y > 0$. On the first challenge of conic representability, our analysis led to a small surprise, namely, the possibility to represent harmonic means (see this blog post). The inequality $\eqref{eq:cut}$ is thus representable on $y > 0$ by $\overline{\beta}(\overline{\gamma} x - u) + h \leq \overline{\beta}(\overline{\gamma} - t)$ and

$$

\label{eq:repr:ypos}

h = A y + B g,\quad \left(\begin{matrix}0.5 (y-g)\\ y+v\\ y\end{matrix}\right) \in Q_r^3,\quad v \geq 0.

$$

We stress that the conic form is a proof of convexity. Next is the challenge of extending this representation to be valid on all of $y \in \mathbb{R}$, while staying tight on $y > 0$. That is, as proven in the paper, we need to extend the representation such that $h \leq 0$ on $y \leq 0$. Notably, however, the current representation in $\eqref{eq:repr:ypos}$ fails this criterion as shown by the graph of $h(y, 5)$ where $A=1$ and $B=-3$.

For this task we deploy another reformulation technique which, in lack of better names, may be denoted as

*stretching the extremum*. The idea is to locally, in the representation of $h(y,v)$, exchange $y$ by $\hat{y} = y + s$ for some amount of stretch $\underline{s} \leq s \leq \overline{s}$. To see the effect of this let $y_{min} = \arg\min_{y \in \mathbb{R}} h(y,v)$, and note that$$

\min_{s \in \mathbb{R}} h(y+s,v) = \begin{cases}

h(y+\overline{s},v), & {\rm if\ } y_{min} \geq y+\overline{s}, \\

h(y_{min},v), & {\rm if\ } y+\underline{s} \leq y_{min} \leq y+\overline{s},\\

h(y+\underline{s},v), & {\rm if\ } y+\underline{s} \geq y_{min}.

\end{cases}

$$

The visual interpretation of this is that we cut the graph of the function in two parts at $y_{min}$, move the two parts away from each other along the y-axis by amounts determined by $\underline{s}$ and $\overline{s}$, and finally fill in the gap with the extremum value $h(y_{min},v)$. In case of $\eqref{eq:repr:ypos}$ we use $\underline{s} = 0$ and $\overline{s} = \infty$ to allow $h$ to take smaller values of $h(y,v)$ at any higher value of $y$. In the graphed example of $h(y, 5)$ above, we get that the lower bound on $h$ is relaxed to the left of the extremum.

As $h(0,v) = 0$ for all $v \geq 0$, we get the desired property that $h \leq 0$ is possible on $y \leq 0$. In conclusion, the globally valid conic quadratic representation of the convex inequality we started out from, is given by $\overline{\beta}(\overline{\gamma} x - u) + h \leq \overline{\beta}(\overline{\gamma} - t)$ and

$$

h = A \hat{y} + B g,\quad \left(\begin{matrix}0.5 (\hat{y}-g)\\ \hat{y}+v\\ \hat{y}\end{matrix}\right) \in Q_r^3,\quad v \geq 0, \quad \hat{y} \geq y.

$$

## Friday, March 23, 2018

### Easter holidays 2018

Support and sales will be closed for Easter holidays

We wish everyone a good Easter.

from

**Thursday, March 29th**until**Monday, April 2nd**(inclusive)**.**We wish everyone a good Easter.

*MOSEK Team*
Subscribe to:
Posts
(
Atom
)