Does zero gradient always mean a minimum?

No. It can also indicate a maximum, saddle, or other stationary behavior.

Does convexity guarantee a unique minimizer?

Not by itself. Strict convexity gives at most one minimizer; existence also matters.

Can a function have a minimum where the gradient is not zero?

Yes, at a constrained boundary; a nonsmooth minimum may not have an ordinary gradient at all.

AI Glossary / Mathematics and optimization / Intermediate

Zero-gradient condition

Definition

The zero-gradient condition states that every first partial derivative of a differentiable objective is zero at a candidate point. Written as gradient f(x)=0, it identifies a stationary point. For a differentiable convex function optimized without constraints over all real vectors, this condition is necessary and sufficient for global optimality. Without those assumptions, a zero gradient may identify a maximum or saddle point instead of a minimum.

What it means

The gradient records how a function changes to first order in different coordinate directions. If it is zero, no direction has a nonzero first-order directional change at that point. Higher-order changes can still make nearby values smaller or larger.
This is why the condition has different force in different settings. Convexity connects local first-order information to the whole objective. In a general nonconvex problem, the condition narrows the candidate set without deciding which candidates minimize the function.
The zero-gradient condition also concerns the mathematical objective being examined. It does not establish that the objective represents the right task, that the data are correct, or that a trained model generalizes.

How it works

Differentiate the objective with respect to every variable and solve the resulting simultaneous equations. Then check the problem's assumptions and classify the candidates using suitable reasoning.
For differentiable convex f, the first-order inequality says f(y) is at least f(x)+gradient f(x) dotted with (y−x). If the gradient is zero, this becomes f(y)≥f(x) for every y, establishing a global minimum.
In unconstrained twice-differentiable problems, curvature information can help distinguish candidates. A positive-definite Hessian at a stationary point is sufficient for a strict local minimum, while indefinite curvature indicates a saddle. A semidefinite Hessian alone may leave the classification unresolved. Constraints and nonsmooth objectives require appropriate alternative optimality conditions.

Three practical examples

A linear least-squares derivation can set the objective gradient to zero and obtain the normal equations.
An optimizer can use a small gradient norm as one stopping diagnostic for a smooth unconstrained problem, interpreted with scale and numerical tolerance in mind.
A researcher can check a claimed optimum analytically and discover that it is actually a stationary saddle or maximum.

Think of it this way

Think of standing where the ground has no immediate slope under your feet. You might be at the bottom of a bowl, the top of a hill, or the center of a saddle-shaped surface. The analogy explains why flat first-order behavior alone is ambiguous, but convexity is an additional global property that makes the bowl-like conclusion rigorous.

Worked example

Consider f(x,y)=2(x−1)^2+(y+2)^2. Its gradient is [4(x−1),2(y+2)]. Setting both components to zero gives x=1 and y=−2. Each squared term is nonnegative and both vanish there, proving the minimum value is zero. The positive coefficients also make this minimizer unique.
Now change the objective to g(x,y)=(x+y−4)^2. Its gradient is [2(x+y−4),2(x+y−4)], which vanishes at every point on the line x+y=4. All those points achieve value zero. The condition still identifies global minima, but it does not establish uniqueness.
For a nonconvex contrast, take h(x)=x^4−2x^2. Its derivative is 4x(x^2−1), giving stationary points 0, 1, and −1. At zero the value is 0, while at either 1 or −1 it is −1. Near zero, small nonzero x gives a negative value, so zero is a local maximum rather than a minimum.
These examples use exact derivatives. They show what additional structure contributes beyond merely solving the gradient equations.

Illustrative case study

In a fictional model-development project, a team stops an optimization run because the reported gradient norm is tiny. The validation result remains poor, and the team initially assumes the best possible model has been reached.
A review separates several possibilities: the routine may be near a stationary point, the objective may be poorly scaled, or the model and loss may be unsuitable for the task. The team checks the mathematical formulation and compares meaningful alternatives rather than treating a stopping message as a business conclusion.
The lesson is that an optimization diagnostic answers a narrower question than model adequacy. The fictional case does not assert which explanation caused the poor result.

Common misconceptions and mistakes

Do not omit the unconstrained assumption. Minimizing x^2 subject to x≥1 gives the solution x=1, where the derivative is 2 rather than zero. The feasible boundary changes the optimality condition.
Do not require an ordinary gradient where the function is not differentiable. The function |x| is minimized at zero despite having no ordinary derivative there.
Do not confuse a zero gradient with a zero objective value. The function (x−2)^2+7 has zero derivative at its minimum, but the minimum value is 7.
Finally, a small numerical gradient is not exactly zero and can depend on units, scaling, derivative accuracy, and numerical precision.

Business use

For a practical optimization report, state whether the objective is smooth, convex, and constrained before interpreting first-order conditions. Keep the achieved objective, gradient or residual measure, and stopping reason together.
When the task is model training, also report appropriate evaluation outside the fitted objective. A mathematically optimal training solution may still be inaccurate or unsuitable for the intended decision. The zero-gradient condition helps establish a specific optimization property; it does not replace the rest of the model assessment.

Check your understanding

For f(x)=(x−5)^2+3 over all real x, where is the gradient zero and what is the minimum value?

Answer: The derivative 2(x−5) vanishes at x=5. The minimum value is 3, not zero.

Related terms

Gradient
Hessian
Global minimum
Constrained optimization
Subgradient
stationary point
convexity
local minimum
saddle point
optimality condition

Learn next

Differentiate the three worked objectives and verify their stationary sets. Then compare a constrained boundary optimum with an unconstrained optimum. Study first-order convexity inequalities and KKT conditions to understand which assumptions support each optimality claim.

Sources

Aalto Dictionary of Machine Learning, zero-gradient condition: necessary and sufficient condition for smooth convex unconstrained minimization. https://aaltodictionaryofml.github.io/ADictML_Main.pdf Boyd and Vandenberghe, Convex Optimization lecture slides: first-order convexity and optimality conditions. https://web.stanford.edu/~boyd/cvxbook/bv_cvxslides.pdf

Previous: Zero degrees of freedom Next: Zero-inflated Poisson regression