Performance specifications

Performance specifications describe constraints on the behavior of the algorithm. Analysis attempts to certify that an algorithm obeys the specification, and synthesis tries to form an algorithm that obeys the specifications.

Every specification is a condition on the performance channel \((w_p, z_p)\) of the generalized plant.

All specifications have the fields

Field	Description
iwp	Indices of performance input \(w_p\)
izp	Indices of performance output \(z_p\)
target	Should this specification be minimized

Most specifications have an additional field rho discount rate \(\rho >0\) as an argument. Choosing \(\rho < 1\) imposes that the property holds at an exponential rate.

Refer to the Performance specification documentation for information about the specifications and their interfaces.

Linear convergence

This imposes exponential stability of the iterates at rate \(\rho\). For every initial condition \(x_0\) there is a fixed point \((x^*(x_0), w^*(x_0), z^*(x_0))\) with

\[\begin{split}\begin{align*} \mav{c}{x_k - x^*(x_0) \\ w_k - w^*(x_0) \\ z_k - z^*(x_0)}_2 \leq \gamma_0\, \rho^{k} \norm{x_0 - x^*(x_0)}_2 = 0, & & \forall k \in \N. \end{align*}\end{split}\]

Linear convergence holds if \(\rho < 1\).

This is the most common specification and needs no performance channel (by default iwp=[], izp=[]).

It is specified with

perf = spec_stability(rho);

Note

The default specification used if none are supplied (specs = []) is {spec_stability(1)} with target=true.

Infinite-horizon exponential stability is that for \(x_0\),

\[\begin{split}\begin{align*} \lim_{k \rightarrow \infty}\rho^{-k}\mav{c}{x_k - x^*(x_0) \\ w_k - w^*(x_0) \\ z_k - z^*(x_0)}_2 = 0 \end{align*}\end{split}\]

Quadratic performance

Several specifications are special cases of general quadratic supply-rate conditions on \((w_p, z_p)\). These conditions must be respected for all performance input sequence \((w_{p, k})_{k \in \N}\) with a finite \(\rho\)-weighted \(\ell_2\)-norm \((\sum_{k=0}^\infty \rho^{-2k} \norm{w_{p, k}}_2^2 < \infty)\).

A quadratic supply rate condition with respect to matrices \((Q, S, R)\) is the existence of an \(\epsilon > 0\) such that

\[\begin{split}\begin{align*} \sum_{k=0}^{T} \rho^{-2k} \mat{c}{w_{p,k} \\ z_{p,k}}^\top \mat{cc}{Q & S \\ S^\top & R} \mat{c}{w_{p,k} \\ z_{p,k}} \; \leq \; \sum_{k=0}^T -\epsilon \rho^{-2k} \norm{w_{p, k}}_2^2 \end{align*}\end{split}\]

for all time horizons \(T > 0\).

Infinite-horizon quadratic performance is that

\[\begin{split}\begin{align*} \limsup_{T \rightarrow \infty} \sum_{k=0}^{T} \rho^{-2k} \mat{c}{w_{p,k} \\ z_{p,k}}^\top \mat{cc}{Q & S \\ S^\top & R} \mat{c}{w_{p,k} \\ z_{p,k}} \; + \; \epsilon \rho^{-2k} \norm{w_{p, k}}_2^2 \leq 0. \end{align*}\end{split}\]

Quadratic performance specifications are imposed by

M = [Q, S; S', R];
perf = spec_quad(M, iwp, izp);

The synthesis procedure requires that \(Q = Q^\top\) and \(R \succ 0\).

The following specifications are all specific instances of quadratic performance.

\(\ell_2\)-stability

Stability and ISS. Certifies that the performance input has a bounded effect on the state,

\[\begin{align*} \sum_{k=0}^T \rho^{-2k} \norm{x_k - x^*(x_0)}_2^2 \leq \gamma\norm{x_0 - x^*(x_0)}_2^2 + \gamma \sum_{k=0}^T \rho^{-2k} \norm{w_{p, k}}_2^2, & & \forall k \in \N. \end{align*}\]

If \(\rho \in (0, 1)\), then \(\ell_2\)-stability implies an input-to-state stability property [1], [2]

\[\begin{align*} \norm{x_k - x^*(x_0)}_2^2 \leq \gamma\, \rho^{2k} \norm{x_0 - x^*(x_0)}_2^2 + \frac{\gamma \rho^2}{1-\rho^2} \max_{t \in 0, \ldots, k} \norm{w_{p, t}}_2^2, & & \forall k \in \N. \end{align*}\]

Linear convergence is recovered from input-to-state stability when the disturbance vanishes (\(w_p= 0\)).

Use this when you need the iterates to stay bounded and convergent under noise but do not need a specific gain; for a gain bound, use \(\ell_2\)-gain instead.

perf = spec_l2(iwp);        % no performance output; optional bound: spec_l2(iwp, MU)

Infinite-horizon \(\ell_2\)-stability is

\[\begin{align*} \limsup_{T \to \infty} \frac{\sum_{k=0}^{T} \rho^{-2k} \norm{x_k - x^*(0)}_2^2}{ \sum_{k=0}^{T} \rho^{-2k} \norm{w_{p,k}}^2_2} \; < \; \gamma^2. \end{align*}\]

\(\ell_2\)-gain

Energy-to-energy gain. Bounds the induced \(\ell_2\)-gain \(\gamma\) from the performance input to the performance output,

\[\begin{align*} \limsup_{T \to \infty} \frac{\sum_{k=0}^{T} \rho^{-2k} \norm{z_{p,k}}_2^2}{ \sum_{k=0}^{T} \rho^{-2k} \norm{w_{p,k}}^2_2} \; < \; \gamma^2. \end{align*}\]

For linear systems this is the \(\rho\)-weighted \(H_\infty\)-gain. The \(\ell_2\)-gain is an infinite-horizon penalty.

Use it to quantify how strongly a disturbance at \(w_p\) (for example, noise in the gradient evaluations) is amplified at a tracked output \(z_p\).

perf = spec_e2e(GAIN, iwp, izp);   % GAIN is the initial bound
perf.target = false; %enforce the gain bound

perf.target = true; %minimize the gain 

Passivity

Imposes a passivity relation between the performance input and output. Passivity may optionally include an input passivity index \(\nu_w\) and an output passivity index \(\nu_z\):

Passivity is obeyed if for all time horizons \(T\) with with \(x_0 = 0, x^*(x_0) = 0\), it holds that

\[\begin{align*} \sum_{k=0}^{T} z_{p,k}^\top w_{p,k} \; \geq \; \sum_{k=0}^{T} \big( \nu_w \norm{w_{p,k}}_2^2 + \nu_z \norm{z_{p,k}}_2^2 \big). \end{align*}\]

Setting both indices to zero corresponds to classical passivity. Positive indices correspond to the stronger input or output strict passivity property, respectively. The performance input and output channels must have the same length.

% ind_w = nu_w, ind_z = nu_z
perf = spec_passivity(ind_w, ind_z, iwp, izp);   

Stochastic sensitivity

Stochastic sensitivity imposes a mean-square boundedness requirement on the performance output [3]. The performance input sequence \(\{w_{p, k}\}\) is a sequence of i.i.d. random variables. These inputs are zero-mean and bounded: there exists a known \(\Omega \succ 0\) such that \(\E[w_{p, k}] = 0\) and \(\E[w_{p, k} w_{p, k}^\top] \preceq \Omega\) at all \(k \in \N\). The algorithm achieves stochastic sensitivity with gain \(\gamma \geq 0\) if for all initial conditions \(x_0\) and performance inputs \(w_p\), it holds that

\[ \limsup_{T \rightarrow \infty} \frac{1}{T} \sum_{k=0}^{T} \mathbb{E}[|| z_{p,k}||^2_2] \leq \gamma^2 \text{Trace}(\Omega).\]

Stochastic sensitivity is invoked by the command

perf = spec_h2(GAIN, Omega, iwp, izp);   %GAIN = gamma

Stochastic sensitivity is only certified if \(\rho = 1\) and the oracle input \(z\) is independent of \(w_p\). In contrast, the \(\ell_2\)-gain is usable if these conditions are violated.

Performance for time-varying dynamical systems

The specific performance constraints imposed by specs may vary for systems with time-variations.

All specifications on this page are used as presented for LTI, periodic, and periodic-orbit systems.

For switched systems, a performance specification in specs imposes a worst-case bound over all possible switching sequences.

More specifications

We plan to implement further performance criteria for both analysis and synthesis.

---

[1]

Eduardo D Sontag. Smooth stabilization implies coprime factorization. IEEE transactions on automatic control, 34(4):435–443, 1989.

[2]

Lukas Schwenkel, Johannes Köhler, Matthias A Müller, Carsten W Scherer, and Frank Allgöwer. Multi-objective robust controller synthesis with integral quadratic constraints in discrete-time. International Journal of Robust and Nonlinear Control, 36(3):935–954, 2026.

[3]

Bryan Van Scoy and Laurent Lessard. The speed-robustness trade-off for first-order methods with additive gradient noise. arXiv preprint arXiv:2109.05059, 2021.