Concept
Calculus on functions of several variables. Partial derivatives, gradients, Jacobians, multiple integrals — the language of ML loss, field theory, utility functions.
Understand it in one breath
Vary one variable while holding the others fixed. On a mountain you can measure the east–west slope and the north–south slope separately; combine them and you get the direction of steepest ascent — that's the gradient.
At a glance
Operation | Input → output | Intuition | Example |
|---|---|---|---|
∂f/∂x | Scalar field → scalar | Hold y fixed; vary only x | East-to-west slope of a temperature field |
∇f (gradient) | Scalar field → vector field | Direction of steepest ascent | Direction of decreasing loss in machine learning |
Jacobian Jᵢⱼ = ∂fᵢ/∂xⱼ | Vector field → matrix | Derivative of a multivariable function | Core of neural-network backpropagation |
∬f dA (double integral) | Function → scalar | Accumulate area × function value | Integral of a probability density |
∇²f (Laplacian) | Scalar → scalar | Difference from the mean | Heat and electrostatic equations |
Gradient descent moves against the local gradient of a loss. It is common in neural-network training, while nondifferentiable models, closed-form methods, and tree-based procedures use other learning algorithms.
Key formula
Modern applications
Neural-network backpropagation, physical simulation, optimization, and marginal analysis in economics.
Beyond MathVoyage
Loading…