Clarabel Solver for Java

Setup

To use clarabel4j, the latest (at the time of writing) JDK 23 is needed as clarabel4j depends on Java’s FFM API.

The library is available at Maven Central. For Maven add the latest version to your pom.xml

<dependency>
    <groupId>com.ustermetrics</groupId>
    <artifactId>clarabel4j</artifactId>
    <version>x.y.z</version>
</dependency>

or using Gradle

implementation 'com.ustermetrics:clarabel4j:x.y.z'

For Windows, Linux, or MacOS users, there is a separate library clarabel4j-native that wraps the native shared library binaries. Using Maven add the latest version from Maven Central to your pom.xml

<dependency>
    <groupId>com.ustermetrics</groupId>
    <artifactId>clarabel4j-native</artifactId>
    <version>x.y.z</version>
    <scope>runtime</scope>
</dependency>

or using Gradle

runtimeOnly 'com.ustermetrics:clarabel4j-native:x.y.z'

Alternatively, install the native Clarabel solver on the machine and add the location to the java.library.path. clarabel4j dynamically loads the native solver.

Since clarabel4j invokes some restricted methods of the FFM API, use the --enable-native-access=com.ustermetrics.clarabel4j or --enable-native-access=ALL-UNNAMED (if you are not using the Java Platform Module System) VM options when running the Java code to avoid warnings from the Java run-time.

Using clarabel4j

Several examples from the Clarabel website can be found in the Model class unit tests of clarabel4j.

For example, for the following quadratic program

\[\begin{align*} & \text{minimize} & & 3 x_1^2 + 2 x_2^2 - x_1 - 4 x_2\\ & \text{subject to} & & -1 \leq x \leq 1, \, x_1 = 2 x_2 \end{align*}\]

the problem data for the objective is

\[ P = 2 \begin{bmatrix} 3 & 0 \\ 0 & 2 \end{bmatrix}, \; q = \begin{bmatrix} -1 \\ -4 \end{bmatrix} \] and for the constraints

\[ A = \left[\begin{array} {rr} 1 & -2 \\ 1 & 0 \\ 0 & 1 \\ -1 & 0 \\ 0 & -1 \end{array}\right], \; b = \begin{bmatrix} 0 \\ 1 \\ 1 \\ 1 \\ 1 \end{bmatrix}, \; \mathcal{K} = \{0\}^1 \times \mathbf{R}^4_{\geq 0} \]

In Java the data can be setup as follows

val p = new Matrix(2, 2, new long[]{0, 1, 2}, new long[]{0, 1}, new double[]{6., 4.});
val q = new double[]{-1., -4.};
val a = new Matrix(5, 2, new long[]{0, 3, 6}, new long[]{0, 1, 3, 0, 2, 4},
        new double[]{1., 1., -1., -2., 1., -1.});
val b = new double[]{0., 1., 1., 1., 1.};
val cones = List.of(new ZeroCone(1), new NonnegativeCone(4));

Note that clarabel4j expects the matrices \(P\) and \(A\) in Compressed Sparse Column (CSC) format.

Then, the quadratic program can be solved and the solution can be checked

try (val model = new Model()) {
    val parameters = Parameters.builder()
            .verbose(false)
            .build();
    model.setParameters(parameters);
    model.setup(p, q, a, b, cones);

    val status = model.optimize();
    
    assertEquals(SOLVED, status);
    assertArrayEquals(new double[]{0.4285714282, 0.2142857141}, model.x(), TOLERANCE);
    assertArrayEquals(new double[]{-1.5714285714, 0., 0., 0., 0.}, model.z(), TOLERANCE);
    assertArrayEquals(new double[]{0., 0.5714285718, 0.7857142859, 1.4285714282, 1.2142857141}, model.s(),
            TOLERANCE);
}

Portfolio optmimization with clarabel4j

We updated the portfolio optimization example from ecos4j with an implementation for clarabel4j. Furthermore, we also implemented the Kotlin Notebook portfolio optimization example with clarabel4j. The notebook can be found here.