Calculus III: Second Derivative Test for Multivariable Functions

In engineering and physics, optimizing a design or understanding a system's stable states often depends on finding the high and low points of a complex, multidimensional landscape. While finding where the terrain is flat is straightforward, determining whether that flat spot is a peak, a valley, or a saddle pass requires more sophisticated tools. The Second Derivative Test for multivariable functions provides this essential classification system, extending the intuition from single-variable calculus to the richer, more complex world of surfaces defined by functions of two or more variables. This test is a cornerstone for solving unconstrained optimization problems in engineering design, machine learning, and economic modeling.

Locating Critical Points: Where the Gradient is Zero

The first step in analyzing a function's landscape is to find all the points where interesting things might happen—the critical points. For a function $f(x,y)$ of two variables, a critical point occurs where its instantaneous rate of change in all directions is zero. This condition is captured by the gradient vector, $\nabla f$.

The gradient is defined as: $$\nabla f(x,y)=⟨f_x(x,y),f_y(x,y)⟩$$ where $f_x$ and $f_y$ are the first-order partial derivatives with respect to $x$ and $y$, respectively. A point $(a,b)$ is a critical point if and only if: $$\nabla f(a,b)=⟨0,0⟩\text{or equivalently}{f}_x(a,b)=0\text{and}{f}_y(a,b)=0.$$

For example, consider the function $f(x,y)=x^3+y^3-3xy$. We find the partial derivatives: $f_x=3x^2-3y$ and $f_y=3y^2-3x$. Setting them equal to zero gives the system: $3x^2-3y=0$ and $3y^2-3x=0$. Solving this system yields the critical points $(0,0)$ and $(1,1)$. At these points, the tangent plane to the surface is horizontal.

Constructing the Hessian Matrix: A Summary of Curvature

Once you have a critical point, you need to examine the local curvature of the surface around it. In one variable, curvature is summarized by the second derivative $f^{\prime \prime }(x)$. For two variables, we need a matrix of all possible second-order partial derivatives, called the Hessian matrix $H$.

For $f(x,y)$, the Hessian is a 2x2 symmetric matrix: $$H(x,y)=\left[\begin{array}{cc}{f}_{xx}(x,y)& f_{xy}(x,y)\\ f_{yx}(x,y)& f_{yy}(x,y)\end{array}\right]$$ By Clairaut's Theorem, for most well-behaved functions encountered in engineering, the mixed partials are equal: $f_{xy}=f_{yx}$. The entries $f_{xx}$ and $f_{yy}$ tell us about the concavity in the pure $x$ and $y$ directions, while $f_{xy}$ captures the "twisting" or interaction between the two directions. This matrix encodes all the second-order information needed to classify the critical point.

The Discriminant (D) and the Second Derivative Test

The classification hinges on the determinant of the Hessian matrix evaluated at the critical point $(a,b)$. This determinant is called the discriminant, often denoted by $D$. $$D(a,b)=\det (H(a,b))=f_{xx}(a,b)f_{yy}(a,b)-[f_{xy}(a,b){]}^2$$

The Second Derivative Test for a function of two variables states:

1. Local Minimum: If $D>0$ and $f_{xx}(a,b)>0$, then $f$ has a local minimum at $(a,b)$. Think of a bowl-shaped surface opening upward.
2. Local Maximum: If $D>0$ and $f_{xx}(a,b)<0$, then $f$ has a local maximum at $(a,b)$. Think of a bowl-shaped surface opening downward (an upside-down bowl).
3. Saddle Point: If $D<0$, then $f$ has a saddle point at $(a,b)$. The surface curves upward in one direction and downward in another, like a mountain pass or a Pringles chip.
4. Inconclusive (Degenerate Case): If $D=0$, the test provides no information. The point could be a local minimum, maximum, saddle point, or none of these. Further analysis is required.

Applying the Test to Classify Critical Points

Let's classify the critical points we found earlier for $f(x,y)=x^3+y^3-3xy$.

First, compute the second partial derivatives: $f_{xx}=6x$, $f_{yy}=6y$, and $f_{xy}=-3$.

Now, evaluate at each critical point.

• At $(0,0)$:

$f_{xx}(0,0)=0$, $f_{yy}(0,0)=0$, $f_{xy}(0,0)=-3$. The discriminant is $D=(0)(0)-(-3{)}^2=-9$. Since $D<0$, the point $(0,0)$ is a saddle point.

• At $(1,1)$:

$f_{xx}(1,1)=6$, $f_{yy}(1,1)=6$, $f_{xy}(1,1)=-3$. The discriminant is $D=(6)(6)-(-3{)}^2=36-9=27$. Since $D>0$ and $f_{xx}(1,1)=6>0$, the point $(1,1)$ is a local minimum.

This step-by-step application—find gradient, solve for critical points, compute second derivatives, build Hessian, calculate $D$, and interpret—is the standard workflow for the test.

Handling Degenerate Cases and Advanced Interpretation

The case where $D=0$ is a degenerate case and signals that the second-derivative information is insufficient. This often happens when the critical point has higher-order flatness. For instance, for $f(x,y)=x^4+y^4$, the point $(0,0)$ is a minimum, but $f_{xx}(0,0)=0$, $f_{yy}(0,0)=0$, and $f_{xy}(0,0)=0$, giving $D=0$. Conversely, for $g(x,y)=x^4-y^4$, $(0,0)$ is a saddle point, but the test is again inconclusive because $D=0$.

To analyze degenerate cases, you must examine the function's behavior more closely. Techniques include:

• Analyzing cross-sections or contours of the function.
• Using a higher-order Taylor series expansion.
• Substituting a path (e.g., $y=mx$) into the function to see if it changes sign around the point.

In engineering contexts, a degenerate result can indicate a design with a broad, flat optimum region or a structural instability that requires more nuanced analysis than local derivatives can provide.

Common Pitfalls

1. Miscomputing Partial Derivatives: A single error in a first or second partial derivative invalidates all subsequent steps. Always check your work, especially for mixed partials. For $f(x,y)=x{e}^{xy}$, remember to use the product rule: $f_{xy}=e^{xy}+xy{e}^{xy}$.

1. Misinterpreting the Sign of $f_{xx}$: The sign of $f_{xx}$ only matters when $D>0$. If $D<0$, the point is a saddle point regardless of what $f_{xx}$ is. A common mistake is to see $f_{xx}>0$ and incorrectly conclude "minimum" without first checking that $D$ is positive.

1. Forgetting to Check All Critical Points: Systems from $\nabla f=0$ can have multiple solutions. You must find and test all critical points. Overlooking one could mean missing the global optimum of your design function.

1. Applying the Test at Non-Critical Points: The Second Derivative Test is defined only at critical points where $\nabla f=0$. Evaluating $D$ at a random point has no standard interpretation for classification.

Summary

• The Second Derivative Test classifies critical points (where $\nabla f=0$) of a two-variable function as local maxima, minima, or saddle points.
• The test relies on the Hessian matrix $H$ of second partial derivatives and its determinant, the discriminant $D=f_{xx}{f}_{yy}-(f_{xy}{)}^2$.
• $D>0$ and $f_{xx}>0$ indicates a local minimum. $D>0$ and $f_{xx}<0$ indicates a local maximum.
• $D<0$