Skip to article
← All posts

Optimization

Fast gradient flows without convexity

The Polyak–Łojasiewicz condition, and a favourite banana-shaped valley.

Convexity is not the only route to fast optimization. The Polyak–Łojasiewicz (P–L) condition asks for something more directly connected to gradient descent: away from the minimum value, the gradient must remain large enough.

For a differentiable function with minimum value \(f_\star\), the condition is

\[\frac12\|\nabla f(x)\|^2\geq \mu\bigl(f(x)-f_\star\bigr), \qquad \mu>0.\]

It rules out stationary points with a nonoptimal value, without requiring a convex landscape.

Along the gradient flow \(\dot x=-\nabla f(x)\),

\[\frac{\mathrm d}{\mathrm dt}(f(x(t))-f_\star) =-\|\nabla f(x(t))\|^2 \leq -2\mu(f(x(t))-f_\star).\]

Integration gives the exponential decay

\[f(x(t))-f_\star \leq e^{-2\mu t}\bigl(f(x(0))-f_\star\bigr).\]

This is a statement about the objective value, not necessarily convergence to a unique minimizer.

A curved valley

One of my favourite illustrations is the banana-shaped gradient flow in the Mathematical Nexus. Let \(I=[0,1]\) and, for \(b\geq0\), consider

\[F_b(x,y)=\operatorname{dist}(x,I)^2+ \bigl(y+b(x-\tfrac12)^2\bigr)^2.\]

Its minimizers form the parabolic arc \(y=-b(x-\tfrac12)^2\), \(x\in I\). For \(b>0\) this set is not convex, so neither is the function. Nevertheless, its gradient flow has an exponential energy bound on each bounded sublevel set.

Explore the banana valley

Teal curve: minimizers. Orange dots: evolving particles. Pale blue curves: positive energy levels.

Change the bend, then play or scrub time to follow the gradient flow. The trajectories use fourth-order Runge–Kutta integration.

Enable JavaScript to use this figure’s animation and sliders. The equations and explanation remain available in the article.

There is a useful qualification here: this banana does not have one uniform, global P–L constant when \(b>0\). Set

\[r=x-\operatorname{proj}_I(x),\qquad z=y+b(x-\tfrac12)^2,\qquad c=2b(x-\tfrac12).\]

Then \(F_b=r^2+z^2\) and \(\nabla F_b=2(r+cz,z)\). Inverting this shear gives

\[\sqrt{r^2+z^2} \leq (1+|c|)\sqrt{(r+cz)^2+z^2}.\]
On \(F_b\leq R\), one has $$ x-\tfrac12 \leq \tfrac12+\sqrt R$$, hence
\[\frac12\|\nabla F_b\|^2 \geq \frac{2}{\bigl[1+2b(\tfrac12+\sqrt R)\bigr]^2}\,F_b.\]

The flow stays in its initial sublevel, so this is enough for exponential decay along every trajectory. The rate may depend on the starting sublevel.

Ten years of a useful perspective

The 2016 paper by Hamed Karimi, Julie Nutini, and Mark Schmidt, Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak–Łojasiewicz Condition, is an excellent entry point into this viewpoint, including its discrete and proximal variants.

Update, 15 September 2026: the paper received the ECML PKDD 2026 Test-of-Time Award, as announced by PICUS Lab, the award committee chair’s research group.

Back to the blog · Gabriel Peyré