Zoutendijk Theorem in Line Search

• Zoutendijk Theorem is a convergence result in line-search methods that guarantees the summability of squared directional cosine-weighted gradient norms under Wolfe conditions.
• It provides quantitative decrease bounds using Armijo and Wolfe conditions, which form a rigorous basis for establishing global convergence in gradient-based algorithms.
• The theorem extends to nonmonotone line searches and set-valued optimization, unifying descent strategies across various nonlinear optimization frameworks.

Searching arXiv for the cited papers and closely related work on Zoutendijk theorem and line search methods. The Zoutendijk theorem is a key convergence result in line-search theory for gradient-based iterative methods. In the formulation formalized in Lean 4, it concerns a sequence generated by

$x_{k+1}=x_k+\alpha_k d_k$

for a continuously differentiable objective $f:\mathbb{R}^n\to\mathbb{R}^n$ or, in the scalar case treated explicitly, $f:\mathbb{R}^n\to\mathbb{R}$, with descent directions and step-sizes satisfying Wolfe conditions. Its conclusion is the summability of a squared directional cosine weighted by the squared gradient norm, namely

$$\sum_{k=0}^\infty \cos^2\theta_k\,\|\nabla f(x_k)\|^2<\infty,$$

which underpins global convergence arguments for first-order methods with inexact line search (Zhang et al., 24 Jun 2026). A closely analogous “Zoutendijk-like condition” appears for set-valued optimization, where the scalar directional derivative is replaced by a scalarization based on the Gerstewitz function and the conclusion becomes

$$\sum_{k=0}^\infty \frac{\bigl(\mathcal{F}^k(x_k,d_k)\bigr)^2}{\|d_k\|^2}<+\infty,$$

again serving as the central summability estimate behind global convergence of nonlinear conjugate gradient schemes (Ghosh et al., 2024).

1. Analytical setting and standing assumptions

In the scalar formulation, the ambient space is $\mathbb{R}^n$ equipped with the Euclidean inner product

$$\langle x,y\rangle=\sum_{i=1}^n x_i y_i,\qquad \|x\|=\sqrt{\langle x,x\rangle}.$$

The objective $f:\mathbb{R}^n\to\mathbb{R}$ is continuously differentiable, with gradient $\nabla f:\mathbb{R}^n\to\mathbb{R}^n$. The standing assumptions are:

$$\|\nabla f(x)-\nabla f(y)\|\le L\|x-y\|\quad \forall x,y\in\mathbb{R}^n$$

for some $f:\mathbb{R}^n\to\mathbb{R}^n$0, or equivalently $f:\mathbb{R}^n\to\mathbb{R}^n$1-Lipschitz continuity of $f:\mathbb{R}^n\to\mathbb{R}^n$2 on the level set $f:\mathbb{R}^n\to\mathbb{R}^n$3;

$f:\mathbb{R}^n\to\mathbb{R}^n$4

for some $f:\mathbb{R}^n\to\mathbb{R}^n$5, so $f:\mathbb{R}^n\to\mathbb{R}^n$6 is bounded below;

the search directions satisfy the descent condition

$f:\mathbb{R}^n\to\mathbb{R}^n$7

the directions are gradient-related in the sense that there exists $f:\mathbb{R}^n\to\mathbb{R}^n$8 such that

$f:\mathbb{R}^n\to\mathbb{R}^n$9

The step-size $f:\mathbb{R}^n\to\mathbb{R}$0 is required to satisfy the Wolfe conditions with constants $f:\mathbb{R}^n\to\mathbb{R}$1: $f:\mathbb{R}^n\to\mathbb{R}$2 and

$f:\mathbb{R}^n\to\mathbb{R}$3

These hypotheses isolate the standard ingredients of inexact line search: a smooth objective, lower boundedness, descent directions, and sufficient decrease plus curvature control. In the Lean development, these assumptions are presented as the formal basis for the mechanized proof of the theorem and for subsequent corollaries on global convergence (Zhang et al., 24 Jun 2026).

2. Formal theorem statement and equivalent bound

The theorem introduces the angle quantity

$f:\mathbb{R}^n\to\mathbb{R}$4

Under the assumptions above, the conclusion is that the series

$f:\mathbb{R}^n\to\mathbb{R}$5

converges. Equivalently, the explicit estimate

$f:\mathbb{R}^n\to\mathbb{R}$6

holds (Zhang et al., 24 Jun 2026).

The role of $f:\mathbb{R}^n\to\mathbb{R}$7 is structural. It measures the alignment of the search direction with the negative gradient. The summand

$f:\mathbb{R}^n\to\mathbb{R}$8

can also be written as

$f:\mathbb{R}^n\to\mathbb{R}$9

which is the form naturally produced by the one-step decrease estimate.

A common misunderstanding is to treat the theorem as a statement that $$\sum_{k=0}^\infty \cos^2\theta_k\,\|\nabla f(x_k)\|^2<\infty,$$0 unconditionally. The stated theorem gives summability of the weighted quantity. The stronger conclusion $$\sum_{k=0}^\infty \cos^2\theta_k\,\|\nabla f(x_k)\|^2<\infty,$$1 requires additional information, such as directions that are truly gradient-related so that $$\sum_{k=0}^\infty \cos^2\theta_k\,\|\nabla f(x_k)\|^2<\infty,$$2 is uniformly bounded away from $$\sum_{k=0}^\infty \cos^2\theta_k\,\|\nabla f(x_k)\|^2<\infty,$$3, or the special case of steepest descent where $$\sum_{k=0}^\infty \cos^2\theta_k\,\|\nabla f(x_k)\|^2<\infty,$$4 (Zhang et al., 24 Jun 2026).

3. One-step decrease, telescoping, and the Lean proof structure

The formal proof in Lean is decomposed into two principal components. The first is the lemma zoutendijkStepCore, which establishes a lower bound on the decrease in one iteration: $$\sum_{k=0}^\infty \cos^2\theta_k\,\|\nabla f(x_k)\|^2<\infty,$$5 The derivation uses three ingredients exactly as stated in the Lean exposition:

the Armijo inequality

$$\sum_{k=0}^\infty \cos^2\theta_k\,\|\nabla f(x_k)\|^2<\infty,$$6

the Wolfe curvature condition together with Lipschitz continuity of $$\sum_{k=0}^\infty \cos^2\theta_k\,\|\nabla f(x_k)\|^2<\infty,$$7, yielding

$$\sum_{k=0}^\infty \cos^2\theta_k\,\|\nabla f(x_k)\|^2<\infty,$$8

substitution of this lower bound on $$\sum_{k=0}^\infty \cos^2\theta_k\,\|\nabla f(x_k)\|^2<\infty,$$9 into the Armijo decrease estimate.

The second component is zoutendijkTheorem, obtained by summing the per-step estimate from $$\sum_{k=0}^\infty \frac{\bigl(\mathcal{F}^k(x_k,d_k)\bigr)^2}{\|d_k\|^2}<+\infty,$$0 to $$\sum_{k=0}^\infty \frac{\bigl(\mathcal{F}^k(x_k,d_k)\bigr)^2}{\|d_k\|^2}<+\infty,$$1: $$\sum_{k=0}^\infty \frac{\bigl(\mathcal{F}^k(x_k,d_k)\bigr)^2}{\|d_k\|^2}<+\infty,$$2 Since $$\sum_{k=0}^\infty \frac{\bigl(\mathcal{F}^k(x_k,d_k)\bigr)^2}{\|d_k\|^2}<+\infty,$$3 is bounded below by $$\sum_{k=0}^\infty \frac{\bigl(\mathcal{F}^k(x_k,d_k)\bigr)^2}{\|d_k\|^2}<+\infty,$$4, the left-hand side is at most $$\sum_{k=0}^\infty \frac{\bigl(\mathcal{F}^k(x_k,d_k)\bigr)^2}{\|d_k\|^2}<+\infty,$$5. A telescoping argument then yields the finite-sum bound and hence convergence of the infinite series (Zhang et al., 24 Jun 2026).

The proof architecture is significant because it exposes the theorem as a summability statement produced by a quantified decrease estimate. In the mechanized setting, this decomposition also clarifies the dependencies among Armijo, Wolfe, Lipschitz continuity, and bounded-below assumptions.

4. Consequences for global convergence

The theorem is described as the cornerstone of the global-convergence theory for any gradient-based method with an inexact line search (Zhang et al., 24 Jun 2026). Once

$$\sum_{k=0}^\infty \frac{\bigl(\mathcal{F}^k(x_k,d_k)\bigr)^2}{\|d_k\|^2}<+\infty,$$6

several consequences follow.

If the directions are truly gradient-related, so that $$\sum_{k=0}^\infty \frac{\bigl(\mathcal{F}^k(x_k,d_k)\bigr)^2}{\|d_k\|^2}<+\infty,$$7 is uniformly bounded away from $$\sum_{k=0}^\infty \frac{\bigl(\mathcal{F}^k(x_k,d_k)\bigr)^2}{\|d_k\|^2}<+\infty,$$8, then

$$\sum_{k=0}^\infty \frac{\bigl(\mathcal{F}^k(x_k,d_k)\bigr)^2}{\|d_k\|^2}<+\infty,$$9

and therefore

$\mathbb{R}^n$0

In the special case of steepest descent,

$\mathbb{R}^n$1

one has

$\mathbb{R}^n$2

so the theorem immediately yields

$\mathbb{R}^n$3

Even without a uniform lower bound on $\mathbb{R}^n$4, the finiteness of the series implies

$\mathbb{R}^n$5

Accordingly, the algorithm cannot “stall” at a non-stationary point: either the angle goes to $\mathbb{R}^n$6, meaning the directions become nearly orthogonal to the gradient, or the gradient norm goes to zero (Zhang et al., 24 Jun 2026).

The Lean library records these implications via the lemmas cos_theta_neg_grad_sq_eq_one and gradient_norm_inf_zero, culminating in globalConvergence, which asserts

$\mathbb{R}^n$7

This distinction matters: the theorem itself is a summability theorem, while the full convergence conclusion depends on the geometry of the directions and, in some cases, on the algorithmic specialization.

5. Nonmonotone line searches and set-valued generalization

Beyond classical Armijo/Wolfe backtracking, the Lean formalization also covers two non-monotone line-search conditions, Grippo and Zhang–Hager. The same stepwise-decrease argument is reported to remain valid if one replaces the reference value $\mathbb{R}^n$8 by a sliding maximum in the Grippo case or by the $\mathbb{R}^n$9-sequence in the Zhang–Hager case. The resulting analogue of zoutendijkStepCore yields exactly the same summability of

$$\langle x,y\rangle=\sum_{i=1}^n x_i y_i,\qquad \|x\|=\sqrt{\langle x,x\rangle}.$$0

so the Zoutendijk theorem and its corollary $$\langle x,y\rangle=\sum_{i=1}^n x_i y_i,\qquad \|x\|=\sqrt{\langle x,x\rangle}.$$1 extend verbatim to these non-monotone line searches (Zhang et al., 24 Jun 2026).

A broader generalization appears in optimization of set-valued mappings of finite cardinality. There the problem is to minimize

$$\langle x,y\rangle=\sum_{i=1}^n x_i y_i,\qquad \|x\|=\sqrt{\langle x,x\rangle}.$$2

with respect to the lower-set order $$\langle x,y\rangle=\sum_{i=1}^n x_i y_i,\qquad \|x\|=\sqrt{\langle x,x\rangle}.$$3, where the image space $$\langle x,y\rangle=\sum_{i=1}^n x_i y_i,\qquad \|x\|=\sqrt{\langle x,x\rangle}.$$4 is ordered by a closed, convex, pointed, solid cone $$\langle x,y\rangle=\sum_{i=1}^n x_i y_i,\qquad \|x\|=\sqrt{\langle x,x\rangle}.$$5. Descent is tested through the Gerstewitz scalarizing function

$$\langle x,y\rangle=\sum_{i=1}^n x_i y_i,\qquad \|x\|=\sqrt{\langle x,x\rangle}.$$6

which satisfies

$$\langle x,y\rangle=\sum_{i=1}^n x_i y_i,\qquad \|x\|=\sqrt{\langle x,x\rangle}.$$7

and is Lipschitz continuous (Ghosh et al., 2024).

For each iterate $$\langle x,y\rangle=\sum_{i=1}^n x_i y_i,\qquad \|x\|=\sqrt{\langle x,x\rangle}.$$8, the scalar-directional measure is defined by

$$\langle x,y\rangle=\sum_{i=1}^n x_i y_i,\qquad \|x\|=\sqrt{\langle x,x\rangle}.$$9

and a direction $f:\mathbb{R}^n\to\mathbb{R}$0 is a $f:\mathbb{R}^n\to\mathbb{R}$1-descent direction if

$f:\mathbb{R}^n\to\mathbb{R}$2

Under boundedness of the level set, Lipschitz continuity of the gradients $f:\mathbb{R}^n\to\mathbb{R}$3 on an open neighborhood of that level set, and an additional boundedness hypothesis for nonincreasing sequences $f:\mathbb{R}^n\to\mathbb{R}$4, the standard Wolfe conditions for set-valued maps imply the Zoutendijk-like summability property

$f:\mathbb{R}^n\to\mathbb{R}$5

for iterates satisfying

$f:\mathbb{R}^n\to\mathbb{R}$6

with $f:\mathbb{R}^n\to\mathbb{R}$7-descent directions and Wolfe step-lengths (Ghosh et al., 2024).

The relation to the classical theorem is exact at the level of structure. In scalar unconstrained minimization one has

$f:\mathbb{R}^n\to\mathbb{R}$8

In the set-valued theorem, the scalar directional derivative $f:\mathbb{R}^n\to\mathbb{R}$9 is replaced by the scalarization $\nabla f:\mathbb{R}^n\to\mathbb{R}^n$0, and the Armijo decrease is replaced by a set-order inequality. Apart from these differences, the proof pattern and the final “derivative$\nabla f:\mathbb{R}^n\to\mathbb{R}^n$1/$\nabla f:\mathbb{R}^n\to\mathbb{R}^n$2” summability property are stated to be in exact analogy (Ghosh et al., 2024).

6. Algorithmic role and verified nonlinear optimization

In the set-valued nonlinear conjugate gradient framework, the Zoutendijk-like condition is used together with a sufficient-descent condition

$\nabla f:\mathbb{R}^n\to\mathbb{R}^n$3

and the nondegeneracy assumption

$\nabla f:\mathbb{R}^n\to\mathbb{R}^n$4

The contradiction argument reported in the paper shows that if $\nabla f:\mathbb{R}^n\to\mathbb{R}^n$5 stays bounded away from zero, then $\nabla f:\mathbb{R}^n\to\mathbb{R}^n$6 is bounded below by a positive constant; sufficient descent then forces

$\nabla f:\mathbb{R}^n\to\mathbb{R}^n$7

contradicting the Zoutendijk-like condition. Hence

$\nabla f:\mathbb{R}^n\to\mathbb{R}^n$8

Since $\nabla f:\mathbb{R}^n\to\mathbb{R}^n$9 characterizes stationarity, every limit point is stationary, which establishes global convergence of the method (Ghosh et al., 2024).

In the scalar setting, the Lean formalization presents the theorem as part of a broader program of machine verification of nonlinear optimization theory. It formalizes fundamental notions related to gradient descent and descent directions, adaptive step-size selection via backtracking line search, and the Armijo, Goldstein, and Wolfe conditions, as well as nonmonotone variants. Within that development, the Zoutendijk theorem functions as the formal backbone for global convergence of first-order methods in the Lean optimization ecosystem (Zhang et al., 24 Jun 2026).

This suggests a unifying interpretation of the theorem across settings: whether expressed through $$\|\nabla f(x)-\nabla f(y)\|\le L\|x-y\|\quad \forall x,y\in\mathbb{R}^n$$0 in scalar optimization or through $$\|\nabla f(x)-\nabla f(y)\|\le L\|x-y\|\quad \forall x,y\in\mathbb{R}^n$$1 in set-valued optimization, the theorem is fundamentally a summability principle derived from sufficient decrease, curvature control, and regularity of the derivative map. Its primary significance lies not in guaranteeing convergence of iterates by itself, but in providing the quantitative estimate from which global convergence arguments are built.