1 Definition and Basic Examples
1.1 Critical points for functions of one variable
1.1.1 Derivative-based criterion (f′(x)=0)
For a real-valued function \(f:\mathbb{R}\to\mathbb{R}\), a point \(x_0\) in the domain is called a critical point if the derivative at that point vanishes, i.e. \(f'(x_0)=0\). Intuitively, \(f'(x_0)=0\) means the function has no instantaneous tendency to increase or decrease at \(x_0\); in smooth settings this corresponds to a horizontal tangent line.
A related but closely tied notion appears when the derivative does not exist at \(x_0\). In many standard calculus courses, such points are excluded unless the underlying assumptions about regularity are addressed. In broader analysis, “critical” can be defined to include certain derivative failures, provided the point still behaves in a controlled way (see below).
1.1.2 Handling non-differentiable points (left/right derivatives, conventions)
At a point \(x_0\) where \(f\) is not differentiable, one common approach is to treat \(x_0\) as critical if appropriate one-sided derivatives exist and “balance” in the sense that the slope changes direction without a clear tangent. For example, if left and right derivatives exist with opposite signs, the point can act like a peak, trough, or corner where local extremum candidates arise.
In encyclopedia-style usage, it is typical to distinguish:
- Smooth critical points, where \(f'(x_0)=0\).
- Nonsmooth critical candidates, where differentiability fails but generalized or one-sided derivative information can still justify treating the point as stationary in an extended sense.
The exact convention varies by context, especially when generalized derivatives are introduced.
1.2 Critical points for multivariable functions
1.2.1 Gradient criterion (∇f(x)=0)
For a differentiable function \(f:\mathbb{R}^n\to\mathbb{R}\), a point \(x_0\) is a critical point if the gradient vanishes: \[ \nabla f(x_0)=0. \] This condition generalizes “horizontal tangent” to higher dimensions. Since directional derivatives are given by \(\mathrm{D}f(x_0)[v]=\nabla f(x_0)\cdot v\), the zero gradient means the first-order change is zero along every direction, so no direction produces a linear increase or decrease.
1.2.2 Jacobian-rank criterion (general differentiable mappings)
When studying mappings other than scalar functions—e.g. \(F:\mathbb{R}^n\to\mathbb{R}^m\)—one often uses a Jacobian rank criterion. A point is considered critical for such a map when the Jacobian matrix loses full rank. In scalar-to-vector contexts, this relates to whether the mapping is locally “regular” or whether it collapses dimensions in a way that affects the geometry of level sets and images.
In optimization and calculus of variations, the most common scalar case uses the gradient criterion, while the Jacobian rank viewpoint becomes important for constraint elimination and implicit-function-style reasoning.
1.3 Geometric interpretation
1.3.1 Tangent slope and level-set behavior
In one variable, \(f'(x_0)=0\) means the tangent line at \(x_0\) is horizontal. For multivariable \(f\), the condition \(\nabla f(x_0)=0\) implies that nearby level sets \(f(x)=c\) have no first-order tilt at \(x_0\). Geometrically, the gradient is normal to the level surface; when it vanishes, this normal direction becomes indeterminate and the surface exhibits local “flatness” to first order.
1.3.2 Stationary points and local “flatness”
A critical point is a place where the first-order approximation offers no guidance toward increasing or decreasing values. Higher-order terms, if present, determine whether the point is a maximum, minimum, saddle, or a more degenerate stationary configuration.
In practice, this means that classification often hinges on what the Taylor expansion looks like beyond the leading nonzero term.
2 Necessary Conditions and Local Behavior
2.1 Stationary points vs. extrema
2.1.1 Why critical points are necessary for interior extrema
For differentiable functions on open intervals, a standard result asserts: if \(f\) attains a local maximum or local minimum at an interior point \(x_0\), then \(f'(x_0)=0\). This reflects the idea that at an extremum, infinitesimal moves in either direction cannot improve the value, forcing the first-order slope to vanish.
The reasoning is often expressed using the mean value theorem: if \(f'(x_0)\neq 0\), then for sufficiently small steps the function moves consistently upward or downward, contradicting local maximality/minimality.
2.1.2 Boundary considerations in constrained/unconstrained problems
At endpoints of a closed interval, local extrema may occur even when derivatives do not vanish. For example, if the function is differentiable inside \((a,b)\) but the best value is at \(a\) or \(b\), then there is no requirement that the derivative be zero because one cannot take steps beyond the boundary.
In constrained optimization, similar boundary-like effects appear: the admissible set restricts movement, so “flatness” in all ambient directions is not required—only stationarity along feasible directions.
2.2 Fermat’s theorem (standard version)
Fermat’s theorem (in its standard calculus form) states that if \(f\) is differentiable near \(x_0\) and \(x_0\) is a point of local extremum, then \(f'(x_0)=0\) (for interior points). The theorem provides a necessary condition used both in theoretical proofs and algorithmic design: candidate extrema are found among critical points and boundary candidates.
2.3 Relationship to monotonicity and sign changes of derivatives
In one dimension, the derivative’s sign often controls monotonicity. When \(f'(x)\) changes sign at \(x_0\), one typically sees \(x_0\) behave like a turning point. However, a vanishing derivative does not automatically imply a sign change: the derivative might touch zero and remain the same sign on both sides, corresponding to a flat inflection-like behavior.
Thus:
- \(f'(x_0)=0\) is a necessary condition for an extremum (interior, smooth case),
- sign changes in \(f'\) are diagnostic but not equivalent to extremality,
- higher-order derivatives clarify cases where \(f'\) merely “hugs” the axis.
3 Classification of Critical Points
3.1 Second-derivative test in one dimension
3.1.1 f″(x0)>0, f″(x0)<0, and the degenerate case
Suppose \(f\) is twice differentiable near \(x_0\) and \(f'(x_0)=0\). Then:
- If \(f''(x_0)>0\), \(x_0\) is a local minimum.
- If \(f''(x_0)<0\), \(x_0\) is a local maximum.
When \(f''(x_0)=0\), the test is inconclusive. Such points are often called degenerate with respect to the second-order criterion, because the first nonzero term in the Taylor expansion may occur at order three or higher.
3.1.2 Higher-order derivatives and order of vanishing
If the first nonzero derivative beyond \(f'\) occurs at order \(k\), the parity of \(k\) influences the classification. For instance, if the leading nonzero term is even order with a positive coefficient, the graph curves upward near the point, suggesting a local minimum. Odd order leading terms typically produce saddle-like “through” behavior in one dimension (often realized as an inflection point).
In short, the order of vanishing of derivatives at \(x_0\) governs the local shape, and the second-derivative test only covers the simplest nondegenerate case.
3.2 Hessian-based tests in multiple dimensions
3.2.1 Positive/negative definiteness and local extrema
For \(f:\mathbb{R}^n\to\mathbb{R}\) twice differentiable and with \(\nabla f(x_0)=0\), consider the Hessian matrix \(H=\nabla^2 f(x_0)\). The second-order Taylor approximation involves the quadratic form \(v^T H v\).
- If \(H\) is positive definite, then \(x_0\) is a strict local minimum.
- If \(H\) is negative definite, then \(x_0\) is a strict local maximum.
Positive/negative definiteness means the quadratic form is uniformly positive/negative on all nonzero directions, preventing any direction from improving the value in the opposite sense.
3.2.2 Indefinite Hessian and saddle points
If the Hessian is indefinite (the quadratic form takes both positive and negative values depending on the direction), then \(x_0\) is a saddle point. Along directions corresponding to positive curvature, \(f\) increases; along directions corresponding to negative curvature, \(f\) decreases. The critical point itself is stationary to first order, but the second-order landscape has competing “uphill” and “downhill” directions.
3.3 Degenerate critical points
3.3.1 Failure of the basic second-order criteria
A critical point is degenerate when the Hessian is singular (neither positive definite nor negative definite). In such cases, some directions have zero quadratic curvature, so the second-order test cannot determine the local type.
Degeneracy also arises more subtly when the first nonzero term in a Taylor expansion has order greater than two, or when symmetry forces cancellations.
3.3.2 Using higher-order Taylor terms (overview)
To classify degenerate points, one typically examines higher-order terms in the Taylor expansion or uses refined geometric criteria. A common approach is to restrict attention to directions in which the quadratic form vanishes and analyze the leading behavior there. This can reveal whether the point is:
- a flat minimum,
- a flat saddle,
- or another non-generic stationary configuration.
4 Taylor Expansions and Higher-Order Analysis
4.1 Taylor’s theorem near a critical point
4.1.1 Reducing behavior to leading nonzero terms
Near a critical point \(x_0\), Taylor’s theorem writes \(f(x_0+h)\) as a sum of terms of increasing order in \(h\). Because the first-order term vanishes at a critical point, the dominant behavior is dictated by the first nonzero derivative term after the gradient (or by the relevant generalized leading term in nonsmooth settings).
This reduction—discarding higher-order remainder terms for sufficiently small \(h\)—turns local classification into an examination of a polynomial-like expression.
4.1.2 Order of contact and directional effects
In multivariable settings, the leading term might depend strongly on direction. For example, a point may appear flat in certain directions due to the vanishing of low-order derivatives along those directions, while in other directions curvature or even sign changes appear earlier. Such directional order of contact is central to understanding why degenerate critical points can exhibit complex shapes.
4.2 Directional derivatives and local models
4.2.1 Examining along rays or lines through the point
A practical method for analysis is to study the restriction \(g(t)=f(x_0+t v)\) along a line through \(x_0\) in direction \(v\). The criticality of \(x_0\) ensures \(g'(0)=0\), and then the one-dimensional higher-order behavior of \(g\) can inform classification in that direction.
By scanning multiple directions (or systematically using symmetry), one can reconstruct a local geometric picture.
4.2.2 Connection to saddle-point geometry
Saddle behavior can be understood as having mixed signs in the leading term’s directional model. If the first nonzero homogeneous polynomial term is indefinite, then different rays exhibit different growth or decay, producing the hallmark saddle structure. When the leading term is positive in all directions of some order, the point resembles a minimum; when it changes sign, it behaves like a saddle at that scale.
5 Methods for Locating and Studying Critical Points
5.1 Solving critical-point equations
5.1.1 Systems of equations from ∇f=0
For a differentiable multivariable function \(f\), critical points are solutions of the system \(\nabla f(x)=0\). This typically forms a set of nonlinear equations, one for each partial derivative.
In theory, solutions can be:
- isolated points (nondegenerate or isolated stationary states),
- or continuous sets (when degeneracies or symmetries produce flat directions).
5.1.2 Practical considerations: scaling and normalization
In real calculations, one often rescales variables, removes dimensions, or normalizes parameters to simplify the algebra. Such steps do not change the conceptual definition of critical points but can improve conditioning of numerical methods or reduce symbolic complexity.
When analytic solving is difficult, one may rely on approximate methods, but the conceptual workflow remains: derive stationarity conditions, then classify candidates using local expansions or second-order information.
5.2 Stability and perturbation ideas (qualitative)
5.2.1 Structural sensitivity near degenerate points
Nondegenerate critical points (where the Hessian is definite) tend to be stable under small perturbations: the qualitative classification persists, with the critical point shifting slightly. Degenerate points are more delicate, because small changes can split one stationary configuration into several.
This sensitivity is visible in optimization landscapes, where flat ridges or plateaus can suddenly develop distinct minima or saddle points when the objective function is modified.
5.2.2 Bifurcation intuition (high-level, non-controversial)
Bifurcation intuition describes how the number or type of critical points can change when a parameter varies. For example, varying a coefficient in a polynomial objective can transform a single degenerate stationary point into multiple nondegenerate ones, or vice versa. While detailed bifurcation analysis belongs to more specialized texts, the key message is that critical points organize local behavior and their evolution reveals underlying structure.
6 Special Topics: Constraints and Generalizations
6.1 Constrained critical points (Lagrange multipliers)
6.1.1 Setup with equality constraints
Consider an objective \(f(x)\) subject to an equality constraint \(g(x)=0\), where \(f,g\) are differentiable and \(x\) lies in \(\mathbb{R}^n\). A constrained critical point is commonly characterized by the existence of a scalar \(\lambda\) such that: \[ \nabla f(x_0)=\lambda \nabla g(x_0). \] This expresses that the gradient of the objective is parallel to the gradient of the constraint surface at the optimal candidate.
6.1.2 Interpreting multipliers and tangent space conditions
The equality \(\nabla f=\lambda \nabla g\) can be read geometrically: the direction along which movement is allowed (tangent to the constraint) must produce no first-order change in \(f\). Equivalently, the directional derivative of \(f\) vanishes for every feasible tangent direction at \(x_0\).
In this view, the multiplier \(\lambda\) measures the relative weighting of how the objective changes when one moves in the normal direction to the constraint—often interpreted as a sensitivity parameter in applications like mechanics and optimization.
6.2 Critical points in non-smooth settings (brief)
6.2.1 Using subgradients / generalized derivatives (conceptual)
When \(f\) fails to be differentiable, the gradient criterion cannot be applied directly. Instead, one uses generalized notions that extend the idea of “no first-order descent direction.” For convex functions, subgradients provide such a replacement: a point is critical in an optimization sense if the zero vector belongs to the subdifferential.
For non-convex nonsmooth problems, the definition is more nuanced, but the guiding principle stays: criticality corresponds to the absence (or balance) of allowable first-order changes.
6.2.2 Clarke critical points (high-level overview)
One common framework for nonsmooth nonconvex analysis uses Clarke’s generalized gradient. Clarke critical points are points where the generalized derivative contains the zero element, again capturing a “stationary” condition without requiring classical differentiability. These notions support existence results and algorithmic designs for nonsmooth optimization.
6.3 Critical points on manifolds (conceptual)
6.3.1 Intrinsic gradient and tangency conditions
On a manifold \(M\subseteq \mathbb{R}^n\), one considers functions defined on \(M\). The relevant notion of gradient becomes the intrinsic gradient on the tangent space \(T_{x_0}M\). A manifold critical point satisfies that the intrinsic gradient vanishes, or equivalently that the differential annihilates all tangent directions: \[ \mathrm{d}f(x_0)[v]=0\quad \text{for all } v\in T_{x_0}M. \] This aligns with the constrained viewpoint, where feasible directions are tangent to the constraint set.
7 Applications and Connections in Analysis
7.1 Variational perspective
7.1.1 Critical points as stationary points of functionals
In the calculus of variations, one studies functionals—mappings that assign a real number to a function (not just to a point in space). Critical points of such functionals correspond to functions that are stationary under infinitesimal perturbations. The “first variation equals zero” condition plays a role analogous to \(\nabla f=0\) in finite-dimensional optimization.
This connection emphasizes the unifying theme: criticality means no first-order change under admissible perturbations.
7.2 Optimization links
7.2.1 Local minima/maxima criteria
In optimization, classifying critical points is central. Local minima correspond to feasible stationary candidates with curvature supporting stability. For smooth problems, second-order conditions based on the Hessian (or its constrained analog) provide practical criteria.
Even when global optimality is hard, many algorithms seek points where first-order stationarity holds and then use curvature information to avoid saddles.
7.2.2 Saddle points and methods that exploit them
Saddles can be obstacles in gradient-based methods, but they also have structural value. In some settings, algorithms and analyses exploit saddle geometry—e.g. by using second-order information or modified descent directions—to escape unstable stationary points.
Moreover, saddle points often encode transitions between different local regimes of the objective function.
7.3 Dynamical systems viewpoint
7.3.1 Equilibria vs. critical points (gradient flows)
In dynamical systems, equilibria often correspond to zeros of a vector field. For gradient flows of the form \[ \dot{x}=-\nabla f(x), \] equilibria coincide with critical points of \(f\). The landscape of \(f\) therefore organizes long-term behavior: minima attract trajectories, while saddles can act as unstable sets.
This perspective reinforces the interpretation of critical points as places where the system has no instantaneous drift.
7.3.2 Lyapunov-type intuition (overview)
A function \(f\) is sometimes used as a Lyapunov function to study stability, especially in gradient-like dynamics. When trajectories follow a rule that monotonically decreases \(f\), critical points become natural candidates for asymptotic states. Under suitable conditions, definiteness properties of the Hessian near a critical point translate into local stability or instability in the dynamical system.