The Matrix-Valued Systems Toolbox (MVST) is the companion MATLAB toolbox for the monograph Linear Hermitian Matrix-Valued Dynamical Systems by Corentin Briat. It is a reusable solver library for the analysis and synthesis of linear matrix-valued (impulsive and stochastic) dynamical systems, covering the book's chapters end to end:
-
stability and stabilization (exact mean-square analysis, state-feedback synthesis, dwell-time-constrained design),
-
linear-cost optimal control over the three horizons (finite, infinite, long-term average),
-
finite-time optimal steering of the state covariance,
-
dissipativity (supply rates, KYP, L2 gain) and dissipativity-based synthesis,
-
zero-sum differential games (H-infinity level),
-
further performance measures (H2, energy-to-peak, peak-to-peak),
-
linear positive systems (copositive-LP analysis and synthesis),
-
higher-order systems.
-
routed H-infinity norm and passivity (
S.hinfnorm(),S.passivity()): uniform result structs with certificates and indices; exact imaginary-axis Hamiltonian tests for CT and DT LTI systems, the DT and DT-MJLS classes, a converged (adaptive) sampled-data gain, and the book's single input-strict passivity index convention (supplyHe(w'z) - nu|w|^2), -
no SDP solver required for the SOS engines:
mvst.Options('solver','nosdp')solves the SOS programs through SOSNOSDP (see Requirements; only the legacyopts.engine = 'yalmip'paths still need an SDP solver).
The system classes handled are the continuous-time diffusion, the Markov-jump linear system (MJLS), the sampled-data / impulsive system, the Poissonian-jump diffusion, and the dwell-time switched system.
The release-facing surface is the +mvst package (mvst.System,
mvst.Options, mvst.solvercheck); every legacy engine remains available as a
mvst_* function in lib/.
The full user manual (author-ordered, one chapter per problem, with runnable
examples) is in docs/manual/index.md. A typeset PDF of
the same manual is at docs/latex/mvst_manual.pdf.
From the matlab/ directory, put the library on the path with the setup
helper (it adds lib/, bookutil/, tests/, the +mvst package parent, and the bundled
solvers under thirdparty/, and prints the solver doctor report):
run('mvst_setup.m') % or simply: mvst_setupmvst_setup accepts 'examples', true to also add examples/, and
'quiet', true to skip the doctor report. The toolbox version is available as
mvst.version.
Then build a continuous-time system and compute its L2 gain and stability:
A = [-1 0.5; 0 -2];
S = mvst.System.ct(A, zeros(2,0), 'Bw', [1;0], 'C', [1 0]); % dx = A x + Bw w, z = C x
g = S.l2gain().gamma; % exact (mean-square) L2 gain
ok = S.stability().stable; % exact mean-square stability
fprintf('L2 gain = %.4f, stable = %d\n', g, ok); % -> L2 gain = 1.0000, stable = 1
h = S.hinfnorm(); % uniform struct: gamma, certificate, class, method, diagnostics
P = mvst.System.ct([0 1;-1 -1], zeros(2,0), 'Bw', [0;1], 'C', [1 0], 'Dw', 0.5);
p = P.passivity(); % exact test: p.passive, p.nu_isp, p.rho_osp, p.methodThe main analysis and synthesis routines are exposed as one-line methods of
mvst.System (l2gain, steer, gamegamma, ocfinite, ocinfinite,
ocaverage, stability, stabilize, dissyn, pos, perf, hinfnorm,
passivity, kyp, dissipativity), each delegating to the matching backend
of the resolved system class; every opts argument takes a plain struct or an
mvst.Options. S.dissipativity(supply) takes 'l2gain', 'passivity' or the
quadratic supply struct('type','qsr','Q',Q,'S',S,'R',R), meaning
[z;w]'*[Q S;S' R]*[z;w]. The remaining engines (the SOS programs mvst_*_sos, the game variants,
dwell-time and positive-systems backends, ...) are called directly as mvst_*
functions; Contents.m (help on the toolbox folder) lists all of them.
-
MATLAB R2020a or later. The code uses
argumentsvalidation blocks (R2019b) andexportgraphics(R2020a, figure export inmvst_savepdf). It is developed and tested on R2025a; older releases are not tested. GNU Octave is not supported (theargumentsblocks and the+mvstclasses are MATLAB-only). -
Base MATLAB suffices for the Riccati-based core: stability, optimal control, covariance steering, games, and the L2 gain.
-
Optimization Toolbox: the positive-systems copositive linear programs (
mvst_pos_*, vialinprog). -
SOSTOOLS and an SDP solver: the sum-of-squares engines (
mvst_*_sos, dissipativity, games, stabilization, and optimal control by SOS). SeDuMi is the default backend; SDPT3 and MOSEK are alternatives, selected throughmvst.Options('solver', ...). -
YALMIP (optional, not bundled): needed only by the legacy
opts.engine = 'yalmip'paths. The LMIs ofmvst_kyp,mvst_pos_ms,mvst_stab_delayand the decay-rate branch ofmvst_perf_dwellare SOSTOOLS programs by default (opts.engine = 'sostools', every backend including'nosdp'); their former YALMIP formulation remains available for one release asopts.engine = 'yalmip'and fails withmvst:opts:yalmipMissingwhen YALMIP is not installed. Install it yourself (https://yalmip.github.io/);mvst_setupandmvst.solvercheckdetect an installation already on the path and report it as optional. -
Control System Toolbox: not required (the Lyapunov solves in
mvst_perfandmvst_kypuse base MATLAB:sylvesterand a Kronecker solve). -
sosnosdp (optional): required for
mvst.Options('solver','nosdp'), which solves the SOS programs (themvst_*_sosengines,mvst_stabz_sd_range,mvst_kyp,mvst_isp_margin_ct,mvst_stab_delay,mvst_pos_ms, and the timer-varying and decay-rate branches ofmvst_perf_dwell) without any SDP solver, see https://github.com/cbriat/sosnosdp. SOSTOOLS is still needed; the legacy YALMIP engines reject'nosdp'(mvst:opts:nosdpYalmip).Limits of
'nosdp'(a different algorithm, not an interior-point SDP solver): accuracy is lower (bisected gains agree with SeDuMi to about 1e-3 to 1e-2 relative), quantities recovered from a non-unique optimum (for example the synthesis gains ofmvst_stabz_sd_range) can differ from the SeDuMi ones by a few percent, programs with large Gram blocks are much slower (the SDP-range synthesis suite takes minutes instead of seconds), and SOSNOSDP's certification step may print benignMatrix is close to singular or badly scaledwarnings near the feasibility boundary (the certificate is verified by explicit Gram-matrix residual and eigenvalue tests, so the warning does not affect soundness). A SOSNOSDP optimum that stays uncertified (statusuncertified_*) is returned by the synthesis engines withinfo.certified = falseand a warning (mvst:uncertified); the classical backends treat such a solve as failed.
SOSTOOLS, SeDuMi, SDPT3, and sosnosdp are shipped in thirdparty/,
each as an unmodified snapshot of the upstream release in its own subfolder
thirdparty/<solver>/, together with its own license file. mvst_setup adds
whatever is present there, so installation is one step. To use a newer version
or your own installation, replace or delete the subfolder: mvst_setup and
mvst.solvercheck discover what is available. The component list, versions,
and licenses are in THIRD-PARTY.md and
thirdparty/VERSIONS.md. SDPT3 ships without compiled
MEX files: run install_sdpt3 once (from thirdparty/sdpt3/) before selecting it.
SeDuMi runs from its MATLAB source, but fast solving needs its platform-specific
MEX files (not tracked in git): run install_sedumi once (from thirdparty/sedumi/).
mvst.solvercheck prints this hint when a solver is found but does not solve.
YALMIP is not bundled. If you use it (legacy engine paths), please cite it (J. Löfberg, "YALMIP: A toolbox for modeling and optimization in MATLAB", CACSD 2004), as its license requests.
-
Solver doctor. Run
mvst.solvercheckat any time to see which backends resolve and actually solve a trivial SDP:mvst.solvercheck
| Path | Contents |
|---|---|
+mvst/ |
Release package: System, Options, solvercheck. |
lib/ |
Reusable engine layer, the mvst_* functions (Riccati, SOS, positive-systems, simulation). |
bookutil/ |
Figure style, PDF export, and table helpers used by the examples. |
thirdparty/ |
Bundled upstream solver snapshots (see THIRD-PARTY.md). |
docs/ |
User manual (Markdown and typeset PDF). |
examples/ |
Example systems and figure scripts (one per book example). |
tests/ |
Unit-test suites (the release battery). |
mvst_setup.m |
Path setup + solver doctor (top level). |
When an example script is run, vector PDFs (white background, black text, so exports are correct even in dark
mode) are written to figures/ under the current folder (or to the folder named
by the environment variable MVST_FIGDIR, see mvst_figdir), and the LaTeX data
macros the book \inputs are written to its data/ subfolder. The SOS/SDP examples require an SDP solver on
the path; the rest need only base MATLAB.
One consolidated entry point runs every unit-test suite, prints a PASS/FAIL/SKIPPED summary, and exits nonzero only if a test fails (tests that need a solver not on the path are reported as SKIPPED):
cd tests; run_all_testsor from the shell:
matlab -batch "cd tests; run_all_tests"
The battery has 26 classdef matlab.unittest suites (about 250 tests; a full run
takes a few minutes), all auto-discovered so a new t*.m TestCase is
picked up without editing the harness. The development tree additionally
carries a golden-data regression harness tied to the book's figure data; it is
not part of the release battery. Individual suites can be run with runtests('tpos') etc.
MIT (see LICENSE). The bundled solvers in thirdparty/ carry
their own licenses (SOSTOOLS GPLv3, SeDuMi GPLv2, SDPT3
GPLv2, sosnosdp MIT); the distribution is an aggregate of separately licensed
components, listed in THIRD-PARTY.md.
See CITATION.cff. Release history: CHANGELOG.md.