A point in a function where the slope is zero but it is not a local extremum.A saddle point is a point on a surface or in a function where the gradient (slope) is zero, but it is not a local maximum or minimum. In two dimensions, it looks like a mountain pass: the surface rises in one direction and falls in another, resembling a horse's saddle. In higher dimensions, saddle points are more common than extrema and play a crucial role in optimization, physics, and game theory.
The concept was formalized in the 19th century by mathematicians like Joseph-Louis Lagrange and Carl Friedrich Gauss, who studied surfaces and critical points. The term 'saddle point' became standard in the early 20th century.
Mathematically, at a saddle point, all partial derivatives are zero, but the second derivative test is inconclusive: the Hessian matrix has both positive and negative eigenvalues. This means that moving along one direction increases the function, while moving along another decreases it. In optimization algorithms, saddle points are problematic because gradient descent can get stuck there, mistaking them for minima. In physics, saddle points appear in potential energy landscapes, where they correspond to unstable equilibrium states, such as the top of a barrier between two stable configurations.
Saddle points are central to machine learning, where training neural networks involves navigating high-dimensional loss landscapes filled with saddle points. In physics, they are used in transition state theory to calculate reaction rates.
In a game of chess, the minimax theorem guarantees that every finite two-player zero-sum game has a saddle point in mixed strategies, meaning there is a value that both players can achieve with optimal play.


