Analyze second-order structure across two, three, or four variables. Detect saddle regions and local extrema. Clean outputs support deeper multivariable optimization and workflow reviews.
Enter a multivariable function, choose two to four variables, and evaluate the Hessian, curvature classification, principal minors, and eigenvalue-based behavior at a selected point.
| Function | Point | Expected Hessian | Typical classification |
|---|---|---|---|
x^2 + y^2 |
(0, 0) | [[2, 0], [0, 2]] | Positive definite |
x^2 - y^2 |
(0, 0) | [[2, 0], [0, -2]] | Indefinite |
x^2 + 2*x*y + 5*y^2 |
(1, 2) | [[2, 2], [2, 10]] | Positive definite |
x^2 + y^2 + z^2 |
(0, 0, 0) | 3x3 identity scaled by 2 | Positive definite |
The Hessian matrix stores second-order partial derivatives of a scalar function. For a function of n variables, the Hessian is the n x n matrix of entries Hij = ∂2f / (∂xi∂xj).
Diagonal term: d2f/dx_i^2 ≈ [f(x+h e_i) - 2f(x) + f(x-h e_i)] / h^2 Mixed term: d2f/(dx_i dx_j) ≈ [f(x+h e_i+h e_j) - f(x+h e_i-h e_j) - f(x-h e_i+h e_j) + f(x-h e_i-h e_j)] / (4h^2) Gradient term: df/dx_i ≈ [f(x+h e_i) - f(x-h e_i)] / (2h)After computing the Hessian, the calculator estimates eigenvalues, determinant, trace, Frobenius norm, symmetry deviation, and leading principal minors. These quantities help classify curvature and support second-derivative tests near stationary points.
x, y, z, and w as needed.0.0001 for smooth functions.The Hessian matrix converts local curvature into a structured object. In optimization, a gradient can indicate direction, but the Hessian measures how quickly that direction changes. For quadratic models, the matrix is constant and gives exact curvature. For nonlinear surfaces, it provides a local approximation that helps distinguish valleys, basins, and saddle regions around the point.
Diagonal entries report pure curvature for each variable, while off-diagonal entries measure interaction strength between variable pairs. If a two-variable function produces a Hessian [[2, 0], [0, 2]], curvature is balanced and uncoupled. If the result becomes [[2, 2], [2, 10]], cross-effects are significant, showing that changing x also alters curvature with respect to y.
Eigenvalues compress Hessian behavior into principal curvature directions. When all eigenvalues are positive, the point behaves like a bowl and supports minimum-type behavior at stationary points. When all are negative, the surface is dome-shaped. Mixed signs identify saddle structure. Small eigenvalues near zero warn that the surface may be nearly flat, ill-conditioned, or numerically sensitive.
The determinant summarizes curvature scaling, while leading principal minors support definiteness tests through Sylvester-style logic. In two dimensions, a positive first minor with a positive determinant usually indicates positive definiteness. A negative determinant signals indefinite curvature immediately. For higher dimensions, the pattern of successive minors becomes more informative, especially when eigenvalues are close to zero and classification needs context.
This calculator uses centered finite differences, so step size directly influences error. A very large step can smear local curvature, while a very small step can amplify rounding noise. Smooth polynomial functions usually behave well around 0.0001, but exponential, logarithmic, or sharply scaled models may need adjustment. Comparing repeated runs with nearby step values is a practical method for checking stability.
Hessian analysis supports multivariable calculus teaching, optimization studies, machine learning loss inspection, and sensitivity review in engineering models. Analysts often compare trace, determinant, Frobenius norm, and eigenvalues to judge curvature strength rather than relying on one metric alone. Exportable tables make reporting easier, because the matrix, derivative estimates, and classification transfer directly into technical documentation.
It is numerical. The page estimates second derivatives with centered finite differences, so results are approximate and depend on smoothness, scaling, and the selected step size.
An indefinite Hessian has both positive and negative eigenvalues. Around a stationary point, that pattern usually indicates a saddle point rather than a local minimum or maximum.
Mixed terms show variable interaction. Large off-diagonal values mean curvature in one direction changes when another variable moves, which matters in coupled models and optimization surfaces.
Start with 0.0001 for smooth functions. Then compare results with nearby values. Stable outputs across small adjustments usually suggest a reasonable finite-difference step.
Yes. The calculator supports two, three, and four active variables, then builds the matching Hessian matrix, eigenvalue set, and principal-minor summary automatically.
Export when you need documentation, classroom reporting, model comparison, or audit trails. CSV is convenient for spreadsheets, while PDF is better for fixed-format sharing.
Important Note: All the Calculators listed in this site are for educational purpose only and we do not guarentee the accuracy of results. Please do consult with other sources as well.