The Hessian test is powerful because a definite quadratic form dominates a smaller-order error. When the quadratic form vanishes in some nonzero direction, that argument has a gap: the higher-order terms may decide the sign. The gap does not mean there is no answer. It means we must use more information about the particular function.
What a second-order expansion actually says
Definition 4.6.1
(Definite, semidefinite, indefinite and saddle)
A symmetric matrix is positive definite if for all , and positive semidefinite if for all ; negative definite and semidefinite are defined with the inequalities reversed. It is indefinite if takes both signs. A critical point (where ) is a saddle if every neighborhood of contains points with and points with .
Suppose is near a critical point , so . Taylor's theorem gives
Theorem 4.6.2
(Second-derivative test at a critical point)
Let be near a critical point . If is positive definite, is a strict local minimum. If it is negative definite, is a strict local maximum. If it is indefinite, is a saddle.
Proof
(Quadratic part against the remainder)
If is positive definite, there is a constant such that . For sufficiently small nonzero , the remainder is smaller in magnitude than , and . Thus is a strict local minimum. A negative definite Hessian similarly gives a strict maximum. If the Hessian is indefinite, directions with positive and negative quadratic values give points of both signs sufficiently close to , so it is a saddle.
Proof X-ray
Where definiteness enters the proof
The proof compares two quantities of the same order. Definiteness is used exactly once, to produce a constant with ; the remainder is then made smaller than by taking small, and the quadratic part wins. In the indefinite case the same comparison is run along one fixed direction with , whose sign survives in . If the Hessian is only semidefinite, may be zero in the null direction, so the remainder is no longer dominated and the argument has no conclusion.
When the Hessian is merely positive semidefinite, a direction may satisfy . The quadratic lower bound by is no longer available. The error is small compared with , but may still be positive, negative, or zero on a quadratic-null direction. A zero determinant therefore makes the two-variable determinant test inconclusive; it does not imply a saddle or a minimum.
Complete a square before guessing a path
Worked example 4.6.3
(Completing a square at a semidefinite point)
Classify the critical point at the origin of
At the origin the gradient vanishes and is positive semidefinite. Instead of applying a test whose hypotheses fail, rewrite the entire expression:
Both terms are nonnegative. Their sum is zero only when and then . Thus the origin is a strict global minimum. This proof controls every path at once, including curves such as that are invisible to a test restricted to straight lines. Along that curve the first square vanishes, but the remaining is positive for .
Worked example 4.6.4
(Opposite paths at the same Hessian)
Now take the function
It has the same Hessian at the origin. Along it is for , whereas along it is . Both paths approach the origin, so the origin is a saddle. These two functions show why the Hessian alone cannot decide the semidefinite case.
The explorer below keeps the Hessian fixed while changes, probes the origin along lines and parabolas, and compares the result with the completed square.
Paths can disprove a claim; a bound proves one
To prove a saddle it is enough to find two sequences or curves approaching the point on which the function difference has opposite signs. To prove a local minimum, checking finitely many paths cannot rule out a negative value on another curve. A factorization, an inequality, or a carefully controlled remainder is stronger because it applies to an entire neighborhood.
Counterexample mode
Every line gives a minimum, yet the origin is a saddle
Checking finitely many straight paths can mislead. Let . Its Hessian at the origin is , which is semidefinite. On the line with we get for small ; on , ; on , . So along every line near the origin. Yet along the parabola , , while along , . Both signs occur arbitrarily near the origin, so it is a saddle. Lines see only straight directions; a curve tangent to the null direction can still change the sign.
Worked example 4.6.5
(A minimum that is not strict)
Classify the origin for
The origin is a local and global minimum, but not a strict one: every point has value zero. Equality conditions matter just as much as the inequality if the question asks for strictness or uniqueness.
For constrained extrema, the feasible set changes the available directions. A gradient need not vanish at a regular constrained extremum, and a test in the ambient plane need not classify the function restricted to the constraint. First determine existence, then check interior candidates and every boundary component; use multipliers only where their regularity hypotheses hold.
Common mistake
(Reading a zero determinant as a verdict)
For and the Hessian at the origin is the same singular matrix . The first has a strict minimum and the second a saddle. A zero determinant says only that the test is silent.
Common mistake
(Taking a minimum to be strict)
The function has a global minimum at the origin, but on the whole curve , so no neighborhood of the origin has off the origin. Strictness needs its own argument.
Checklist before classifying
- Verify the point is in the domain; an interior local extremum where is differentiable must be stationary.
- State the regularity needed for the expansion and calculate the Hessian at the point.
- If definite or indefinite, justify the classification using a valid test. If semidefinite, do not infer a classification from the determinant alone.
- Seek a global identity or local bound for a minimum/maximum; use opposing paths to establish a saddle.
- State whether the extremum is strict and whether the conclusion is local or global.
Quick checks
Quick check
(Classify the origin for .)
The Hessian is .
Solution · Answer
Both terms are nonnegative and vanish together only at , so elsewhere: a strict global minimum, although the Hessian test is inconclusive.
Quick check
(Classify the origin for .)
Compare the paths for and .
Solution · Answer
The Hessian is . Along , is positive for and negative for , so the origin is a saddle.
Quick check
(Classify the critical points of .)
Where does vanish?
Solution · Answer
, with equality on the whole line , and every point of that line is critical. Each is a global minimum, none strict. The Hessian is semidefinite and singular everywhere.