5.3. Iteration-dependent analyses

Here, we focus on finite-horizon analyses, which apply to algorithms that run indefinitely, as well as algorithms with finite iteration budgets matching or exceeding the horizon. For notational simplicity, we assume that

\[K\in\mathbb{N}\]

in (3.1), representing the horizon.

Following [TB19], we adopt an ansatz consisting of a sequence of iteration-dependent and quadratic Lyapunov functions, also often referred to as potential functions [BG19].

Definition 5.3.1 (Chained Lyapunov inequalities).

Suppose that Assumption 3.1 (Well-posedness) holds. Let

\[\bx_{0}\in\calH^{n}, \qquad \p{ ( \bx^{k}, \bu^{k}, \by^{k}, \bFcn^{k} ) }_{k=0}^{K},\]

be an initial point and a sequence of iterates satisfying (3.1), and let

\[\p{y^{\star}, \hat{\bu}^{\star}, \bFcn^{\star}}\]

be a point satisfying (5.1). Also let

\[c_K\in\reals_{+}.\]

Define

(5.49)\[\begin{split}\begin{aligned} \p{\forall k \in \llbracket0,K\rrbracket} \quad \left[ \begin{aligned} &\mathcal{V}(Q_k,q_k,k) {}={} \quadform{Q_{k}}{\p{\bx^{k},\bu^{k},\hat{\bu}^{\star},y^{\star}}} + q_{k}^{\top}\p{\bFcn^{k},\bFcn^{\star}}, \\ &Q_{k} \in \sym^{n+\NumEval+m}, \\ &q_{k} \in \reals^{\NumEvalFunc + \NumFunc}. \end{aligned} \right. \end{aligned}\end{split}\]

We say that \((\p{Q_{k},q_{k}})_{k=0}^{K}\) and \(c_K\) satisfy a length \(K\) sequence of chained Lyapunov inequalities for algorithm (3.1) over the problem class defined by \((\mathcal{F}_i)_{i\in\IndexFunc}\) and \((\mathcal{G}_i)_{i\in\IndexOp}\) if

(5.50)\[\begin{split}\begin{aligned} \mathcal{V}\p{Q_K,q_K,K} &\leq \mathcal{V}\p{Q_{K-1},q_{K-1},K-1} \\ &\leq \ldots \\ &\leq \mathcal{V}\p{Q_1,q_1,1} \\ &\leq c_K \mathcal{V}\p{Q_0,q_0,0} \end{aligned}\end{split}\]

holds for each

\[\bx_{0}, \qquad \p{ ( \bx^{k}, \bu^{k}, \by^{k}, \bFcn^{k} ) }_{k=0}^{K}, \qquad \p{y^{\star}, \hat{\bu}^{\star},\bFcn^{\star}},\]

and for each

\[\p{f_{i}}_{i\in\IndexFunc} \in \prod_{i\in\IndexFunc} \mathcal{F}_{i}, \qquad \p{G_i}_{i\in\IndexOp} \in \prod_{i\in\IndexOp}\mathcal{G}_i.\]

In the proposed methodology, the user specifies the boundary parameters \(\p{Q_{0},q_{0},Q_{K},q_{K}}\), which define the initial and final Lyapunov expressions \(\mathcal{V}\p{Q_0,q_0,0}\) and \(\mathcal{V}\p{Q_K,q_K,K}\). The SDP then searches for \((\p{Q_{k},q_{k}})_{k=1}^{K-1}\) and the smallest \(c_K\) such that (5.50) holds. When feasible, this yields

\[\begin{aligned} \mathcal{V}\p{Q_K,q_K,K} \leq c_K \mathcal{V}\p{Q_0,q_0,0}.\end{aligned}\]

For built-in choices of these boundary terms:

The corresponding API entry point is search_lyapunov().

The SDP characterization is given in Theorem 5.3.2 (Chained Lyapunov inequalities via SDP). Computationally, each inequality in (5.50) is enforced by a one-step analysis with positive semidefinite constraints of constant dimension. As a result, the number of variables and constraints grows linearly with \(K\).

Theorem 5.3.2 (Chained Lyapunov inequalities via SDP).

Suppose that Assumption 3.1 (Well-posedness) and Assumption 4.1 (Interpolation conditions) hold, and

\[K\in\mathbb{N}, \qquad Q_{0},Q_{K} \in \sym^{n+\NumEval+m}, \qquad q_{0},q_{K} \in \reals^{\NumEvalFunc + \NumFunc}.\]

Then a sufficient condition for there to exist

\[\p{ (Q_{k},q_{k}) }_{k=1}^{K-1}, \qquad c_K,\]

such that \(\p{ (Q_{k},q_{k}) }_{k=0}^{K}\) and \(c_K\) satisfy a length \(K\) sequence of chained Lyapunov inequalities for algorithm (3.1) over the problem class defined by \(\p{\mathcal{F}_i}_{i\in\IndexFunc}\) and \(\p{\mathcal{G}_i}_{i\in\IndexOp}\) is that the following system of constraints, which reuses the lifted matrices/vectors from 5.1. Performance estimation via SDPs, including the lifted state matrices (5.3), and the interpolation terms (5.12), (5.13), (5.14), (5.15), and (5.16),

\begin{align} &\textbf{for each}\ k \in \llbracket 0, K-1 \rrbracket \notag \\ &\quad \left( \begin{array}{@{}c@{}} \forall i \in\IndexFunc \\ \forall o \in \mathcal{O}^{\textup{func-ineq}}_{i} \\ \forall \p{ \p{j_{1},k_{1}}, \ldots, \p{j_{n_{i,o}},k_{n_{i,o}} }}\in\mathcal{J}_{i,o}^{k,k+1} \end{array} \right) \quad \lambda_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{\textup{func-ineq},\,k} \geq 0, \tag{5.51a}\\ &\quad \left( \begin{array}{@{}c@{}} \forall i \in\IndexFunc \\ \forall o \in \mathcal{O}^{\textup{func-eq}}_{i} \\ \forall \p{ \p{j_{1},k_{1}}, \ldots, \p{j_{n_{i,o}},k_{n_{i,o}} }}\in\mathcal{J}_{i,o}^{k,k+1} \end{array} \right) \quad \nu_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{\textup{func-eq},\,k} \in \reals, \tag{5.51b}\\ &\quad \left( \begin{array}{@{}c@{}} \forall i \in\IndexFunc \\ \forall o \in \mathcal{O}^{\textup{op}}_{i} \\ \forall \p{ \p{j_{1},k_{1}}, \ldots, \p{j_{n_{i,o}},k_{n_{i,o}} }}\in\mathcal{J}_{i,o}^{k,k+1} \end{array} \right) \quad \lambda_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{\textup{op},\,k} \geq 0, \tag{5.51c}\\ &\quad - W_{k} + \sum_{\substack{i\in\IndexFunc \\ o \in \mathcal{O}^{\textup{func-ineq}}_{i} \\ \p{ \p{j_{1},k_{1}}, \ldots, \p{j_{n_{i,o}},k_{n_{i,o}} }}\in\mathcal{J}_{i,o}^{k,k+1} }} \lambda_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{\textup{func-ineq},\,k} W_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{k,\,k+1,\,\textup{func-ineq}} \notag \\ &\quad \quad + \sum_{\substack{i\in\IndexFunc \\ o \in \mathcal{O}^{\textup{func-eq}}_{i} \\ \p{ \p{j_{1},k_{1}}, \ldots, \p{j_{n_{i,o}},k_{n_{i,o}} }}\in\mathcal{J}_{i,o}^{k,k+1} }} \nu_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{\textup{func-eq},\,k} W_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{k,\,k+1,\,\textup{func-eq}} \notag \\ &\quad \quad + \sum_{\substack{i\in\IndexOp \\ o \in \mathcal{O}^{\textup{op}}_{i} \\ \p{ \p{j_{1},k_{1}}, \ldots, \p{j_{n_{i,o}},k_{n_{i,o}} }}\in\mathcal{J}_{i,o}^{k,k+1} }} \lambda_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{\textup{op},\,k} W_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{k,\,k+1,\,\textup{op}} \succeq 0, \tag{5.51d}\\ &\quad - w_{k} + \sum_{\substack{i\in\IndexFunc \\ o \in \mathcal{O}^{\textup{func-ineq}}_{i} \\ \p{ \p{j_{1},k_{1}}, \ldots, \p{j_{n_{i,o}},k_{n_{i,o}} }}\in\mathcal{J}_{i,o}^{k,k+1} }} \lambda_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{\textup{func-ineq},\,k} F_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{k,\,k+1,\,\textup{func-ineq}} \notag \\ &\quad \quad + \sum_{\substack{i\in\IndexFunc \\ o \in \mathcal{O}^{\textup{func-eq}}_{i} \\ \p{ \p{j_{1},k_{1}}, \ldots, \p{j_{n_{i,o}},k_{n_{i,o}} }}\in\mathcal{J}_{i,o}^{k,k+1} }} \nu_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{\textup{func-eq},\,k} F_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{k,\,k+1,\,\textup{func-eq}} = 0, \tag{5.51e}\\ &\mathbf{end} \notag \\ & \p{\forall k\in \llbracket1,K-1\rrbracket} \quad Q_{k} \in \sym^{n+\NumEval+m}, \tag{5.51f}\\ & \p{\forall k\in \llbracket1,K-1\rrbracket} \quad q_{k} \in \mathbb{R}^{\NumEvalFunc + \NumFunc}, \tag{5.51g}\\ & c_K \in \reals_{+}. \tag{5.51h} \end{align}

is feasible for the scalars

\[\begin{split}\begin{aligned} \lambda_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{\textup{func-ineq},\,k}, \\ \nu_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{\textup{func-eq},\,k}, \\ \lambda_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{\textup{op},\,k}, \end{aligned}\end{split}\]

the matrices and vectors \(\p{\p{Q_{k},q_{k}}}_{k=1}^{K-1}\), and the scalar \(c_K\), where we have introduced

(5.52)\[\begin{aligned} W_{0} &= \p{\Theta_{1}^{\p{0}}}^{\top}Q_{1}\Theta_{1}^{\p{0}} - c_K \Theta_{0}^{\top}Q_{0}\Theta_{0}. \end{aligned}\]
(5.53)\[\begin{aligned} w_{0} &= \theta_{1}^{\top}q_{1} - c_K\theta_{0}^{\top}q_{0}. \end{aligned}\]
(5.54)\[\begin{aligned} \p{\forall k\in\llbracket1,K-1\rrbracket}\quad W_{k} &= \p{\Theta_{1}^{\p{k}}}^{\top}Q_{k+1}\Theta_{1}^{\p{k}} - \Theta_{0}^{\top}Q_{k}\Theta_{0}. \end{aligned}\]
(5.55)\[\begin{aligned} \p{\forall k\in\llbracket1,K-1\rrbracket}\quad w_{k} &= \theta_{1}^{\top}q_{k+1} - \theta_{0}^{\top}q_{k}. \end{aligned}\]

and

(5.56)\[\begin{split}\begin{aligned} \Theta_{0} &= \underbracket{ \begin{bmatrix} I_{n+\NumEval } & 0_{\p{n+\NumEval}\times\NumEval } & 0_{\p{n+\NumEval}\times m} \\ 0_{m \times \p{n+\NumEval}} & 0_{m \times\NumEval } & I_{m} \end{bmatrix} }_{ \in \reals^{ \p{n + \NumEval + m} \times \p{n + 2\NumEval + m} } }. \end{aligned}\end{split}\]
(5.57)\[\begin{split}\begin{aligned} \theta_{0} &= \underbracket{ \begin{bmatrix} I_{\NumEvalFunc} & 0_{\NumEvalFunc \times \NumEvalFunc } & 0_{\NumEvalFunc \times \NumFunc } \\ 0_{\NumFunc\times \NumEvalFunc} &0_{\NumFunc\times \NumEvalFunc} & I_{\NumFunc} \end{bmatrix} }_{ \in \reals^{\p{\NumEvalFunc + \NumFunc}\times\p{2\NumEvalFunc + \NumFunc}} }. \end{aligned}\end{split}\]
(5.58)\[\begin{split}\begin{aligned} \p{\forall k\in\llbracket0,K-1\rrbracket}\quad \Theta_{1}^{\p{k}} &= \underbracket{ \begin{bmatrix} X_{k+1}^{k,k+1} \\ 0_{\p{\NumEval + m} \times \p{n + \NumEval } } \quad I_{\NumEval + m} \end{bmatrix} }_{ \in \reals^{ \p{n + \NumEval + m} \times \p{n + 2\NumEval + m} } }. \end{aligned}\end{split}\]
(5.59)\[\begin{aligned} \theta_{1} &= \underbracket{ \begin{bmatrix} 0_{\p{\NumEvalFunc + \NumFunc}\times\NumEvalFunc} & I_{\NumEvalFunc + \NumFunc} \end{bmatrix} }_{ \in \reals^{\p{\NumEvalFunc + \NumFunc}\times\p{2\NumEvalFunc + \NumFunc}} }. \end{aligned}\]

Furthermore, if the interpolation conditions for \(\p{\mathcal{F}_{i}}_{i\in\IndexFunc}\) and \(\p{\mathcal{G}_{i}}_{i\in\IndexOp}\) are tight, \(\dim \calH \geq n + 2\NumEval + m\), and there exists

\[\begin{split}\begin{aligned} \p{\forall k\in \llbracket0,K-1\rrbracket} \quad \left[ \begin{aligned} &G_{k}\in\sym_{++}^{n + 2\NumEval + m}, \\ &\bchi_{k}\in\reals^{2\NumEvalFunc + \NumFunc}, \end{aligned} \right. \end{aligned}\end{split}\]

such that

\[\begin{split}\begin{aligned} & \textbf{for each}\ k\in \llbracket0,K-1\rrbracket \\ & \quad \left( \begin{array}{@{}c@{}} \forall i \in\IndexFunc \\ \forall o \in \mathcal{O}^{\textup{func-ineq}}_{i} \\ \forall \p{ \p{j_{1},k_{1}}, \ldots, \p{j_{n_{i,o}},k_{n_{i,o}} }}\in\mathcal{J}_{i,o}^{k,k+1} \end{array} \right) \quad \begin{aligned} & \bm{\chi}^{\top}_{k} F_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{k,\,k+1,\,\textup{func-ineq}} \\ & + \trace\p{W_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{k,\,k+1,\,\textup{func-ineq}}G_{k}} \leq 0, \end{aligned} \\ & \quad \left( \begin{array}{@{}c@{}} \forall i \in\IndexFunc \\ \forall o \in \mathcal{O}^{\textup{func-eq}}_{i} \\ \forall \p{ \p{j_{1},k_{1}}, \ldots, \p{j_{n_{i,o}},k_{n_{i,o}} }}\in\mathcal{J}_{i,o}^{k,k+1} \end{array} \right) \quad \begin{aligned} & \bm{\chi}^{\top}_{k} F_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{k,\,k+1,\,\textup{func-eq}} \\ & + \trace\p{W_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{k,\,k+1,\,\textup{func-eq}}G_{k}} = 0, \end{aligned} \\ & \quad \left( \begin{array}{@{}c@{}} \forall i \in\IndexFunc \\ \forall o \in \mathcal{O}^{\textup{op}}_{i} \\ \forall \p{ \p{j_{1},k_{1}}, \ldots, \p{j_{n_{i,o}},k_{n_{i,o}} }}\in\mathcal{J}_{i,o}^{k,k+1} \end{array} \right) \quad \trace\p{W_{\p{i,j_{1},k_{1},\ldots,j_{n_{i,o}},k_{n_{i,o}},o}}^{k,\,k+1,\,\textup{op}}G_{k}} \leq 0, \\ & \mathbf{end} \end{aligned}\end{split}\]

then (5.51) is also a necessary condition.

Proof. Note that

\[\begin{split}\begin{aligned} \p{\bx^{k},\bu^{k},\hat{\bu}^{\star},y^{\star}} &= \p{\Theta_{0} \kron \Id }\p{\bx^{k},\bu^{k},\bu^{k+1},\hat{\bu}^{\star},y^{\star}}, \\ \p{\bFcn^{k},\bFcn^{\star}} &= \theta_{0}\p{\bFcn^{k},\bFcn^{k+1},\bFcn^{\star}}, \\ \p{\bx^{k+1},\bu^{k+1},\hat{\bu}^{\star},y^{\star}} &= \p{ \Theta_{1}^{\p{k}} \kron \Id }\p{\bx^{k},\bu^{k},\bu^{k+1},\hat{\bu}^{\star},y^{\star}}, \\ \p{\bFcn^{k+1},\bFcn^{\star}} &= \theta_{1}\p{\bFcn^{k},\bFcn^{k+1},\bFcn^{\star}}, \end{aligned}\end{split}\]

where we have used in (5.56), (5.57), (5.58), and (5.59), respectively.

First, suppose that the parameters \(\p{\p{Q_{k},q_{k}}}_{k=1}^{K-1}\) and \(c_K\) are fixed and consider \(\p{\mathcal{V}\p{Q_k,q_k,k}}_{k=0}^{K}\) in (5.49). Note that

(5.60)\[\begin{split}\begin{aligned} &\mathcal{V}\p{Q_1,q_1,1} - c_K\mathcal{V}\p{Q_0,q_0,0} \notag\\ & = \quadform{\p{\Theta_{1}^{\p{0}}}^{\top}Q_{1}\Theta_{1}^{\p{0}} - c_K\Theta_{0}^{\top}Q_{0}\Theta_{0}}{\p{\bx^{0},\bu^{0},\bu^{1},\hat{\bu}^{\star},y^{\star}}} \notag \\ & \quad + \p{q_{1}^{\top}\theta_{1} - c_K q_{0}^{\top}\theta_{0} } \p{\bFcn^{0},\bFcn^{1},\bFcn^{\star}} \notag \\ &= \quadform{W_{0}}{\p{\bx^{0},\bu^{0},\bu^{1},\hat{\bu}^{\star},y^{\star}}} + w_{0}^{\top} \p{\bFcn^{0},\bFcn^{1},\bFcn^{\star}} \end{aligned}\end{split}\]

where (5.52) and (5.53) is used in the last equality. Therefore, using the (5.60) as the objective function in (PEP), Theorem 5.1.1 (Performance estimation via SDP) gives that (5.51), with \(k=0\), is a sufficient condition for the inequality \(\mathcal{V}\p{Q_1,q_1,1} \leq c_K\mathcal{V}\p{Q_0,q_0,0}\).

For \(k\in \llbracket1,K-1\rrbracket\), note that

(5.61)\[\begin{split}\begin{aligned} &\mathcal{V}\p{Q_{k+1},q_{k+1},k+1} - \mathcal{V}\p{Q_k,q_k,k} \notag \\ & =\quadform{Q_{k+1}}{\p{\bx^{k+1},\bu^{k+1},\hat{\bu}^{\star},y^{\star}}} + q_{k+1}^{\top}\p{\bFcn^{k+1},\bFcn^{\star}} \notag\\ &\quad -\quadform{Q_{k}}{\p{\bx^{k},\bu^{k},\hat{\bu}^{\star},y^{\star}}} - q_{k}^{\top} \p{\bFcn^{k},\bFcn^{\star}} \notag \\ & = \quadform{\p{\Theta_{1}^{\p{k}}}^{\top}Q_{k+1}\Theta_{1}^{\p{k}} - \Theta_{0}^{\top}Q_{k}\Theta_{0}}{\p{\bx^{k},\bu^{k},\bu^{k+1},\hat{\bu}^{\star},y^{\star}}} \notag\\ &\quad + \p{q_{k+1}^{\top}\theta_{1} - q_{k}^{\top}\theta_{0} } \p{\bFcn^{k},\bFcn^{k+1},\bFcn^{\star}} \notag \\ & = \quadform{W_k}{\p{\bx^{k},\bu^{k},\bu^{k+1},\hat{\bu}^{\star},y^{\star}}} + w_k^{\top}\p{\bFcn^{k},\bFcn^{k+1},\bFcn^{\star}} \end{aligned}\end{split}\]

where (5.54) and (5.55) is used in the last equality. Therefore, using the (5.61) as the objective function in (PEP), Theorem 5.1.1 (Performance estimation via SDP) gives that (5.51), constrained to this particular \(k\), is a sufficient condition for the inequality \(\mathcal{V}\p{Q_{k+1},q_{k+1},k+1} \leq \mathcal{V}\p{Q_k,q_k,k}\).

Second, note that the proof is complete if we let the parameters \(\p{\p{Q_{k},q_{k}}}_{k=1}^{K-1}\) and \(c_K\) free, as in (5.51f), (5.51g), and (5.51h). \(\square\)

References

[BG19]

Nikhil Bansal and Anupam Gupta. Potential-function proofs for gradient methods. Theory of Computing, 15(1):1–32, 2019. doi:10.4086/toc.2019.v015a004.

[TB19]

Adrien Taylor and Francis Bach. Stochastic first-order methods: Non-asymptotic and computer-aided analyses via potential functions. In Conference on Learning Theory (COLT). 2019.