A vector in ℝⁿ is an ordered list of numbers with both magnitude and direction. The dot product of vectors a and b is a·b = ∑ a_ib_i = |a||b|cosθ — it measures how aligned two vectors are, and is zero exactly when they are perpendicular. In ℝ³, the cross product a × b produces a new vector perpendicular to both.

For f(x, y), the partial derivative ∂f/∂x differentiates with respect to x while treating every other variable as a constant. Collecting all partials gives the gradient:

The gradient points in the direction of steepest increase of f, with magnitude equal to that steepest rate. The directional derivative along a unit vector u is D_uf = ∇f · u. Multiple integrals extend the same accumulation idea to two or three dimensions, e.g. a double integral over a region gives the volume under a surface.

Imagine standing on a hillside described by f(x, y) = elevation. A partial derivative asks: "if I only ever step east-west (holding north-south fixed), how steep is it?" That is one slice of the full picture. The gradient stitches every direction's slope together into a single vector that always points straight uphill, as steeply as possible — walk in that direction and you gain elevation fastest of all.

The dot product's role in the directional derivative is elegant: ∇f · u measures how much of the gradient's "push" lines up with the direction u you actually chose to walk. Walk along the gradient itself and you get the maximum possible rate; walk perpendicular to it and the rate is zero — you're moving along a contour line, at constant height. This machinery — the gradient — is exactly what powers gradient descent, the algorithm that trains most modern machine learning models by repeatedly stepping downhill on an error surface.