Differential Calculus
Letting h go to zero…
Initializing Sorting Algorithms...
Sorting
Trees
Graphs
Preparing interactive visualizations...
Differential Calculus
Letting h go to zero…
The mathematics of change. Squeeze a secant into a tangent, read a curve's shape from its derivatives, and follow slopes to roots, to maxima and down a loss surface.
Walk towards a point from the left and from the right. A removable hole, a jump and an asymptote, then an ε–δ band you can tighten.
lim x→a f(x) = LFunctions and graphsApproach from Both SidesA secant line through two points of a curve, pulled together until it becomes the tangent — with the difference quotient updating live.
f′(x) = lim h→0 (f(x + h) − f(x)) / hLimitsSecant to TangentThe power rule, the product rule as a rectangle growing on two sides, and the chain rule as rates multiplying through a composition.
(uv)′ = u′v + uv′, (f∘g)′ = f′(g)·g′The DerivativeRules from Picturesf, f′ and f″ stacked on one x-axis. Rising and falling, bending up and down, maxima, minima and inflection points, all tied together.
f′ > 0 ⇒ f rises, f″ > 0 ⇒ f bends upDerivative RulesSign AnalysisCut squares from the corners of a sheet and fold it into an open box. Which cut holds the most? Find it where the tangent goes flat.
V′(x) = 0 at the best cutGraphing with DerivativesCritical PointsFollow the tangent down to the x-axis, then do it again. Correct digits double each step near a root — until a bad start sends it cycling or flying off.
xₙ₊₁ = xₙ − f(xₙ) / f′(xₙ)The DerivativeTangent IterationAdd one term at a time and watch the polynomial hug sin x and eˣ. For ln(1 + x), see exactly where adding terms stops helping.
f(x) = ∑ f⁽ⁿ⁾(a) (x − a)ⁿ / n!Derivative RulesPolynomial ApproximationSlice a surface along x and along y to get the two partial slopes, then see the gradient on the contour map — always perpendicular to the level curves.
∇f = (∂f/∂x, ∂f/∂y)The DerivativeSlices of a SurfaceA ball stepping downhill on a loss surface in 3D and on its contour map. Slide the learning rate from crawling to converging to zigzagging to blowing up.
x ← x − η ∇f(x)Partial Derivatives and GradientsIterative OptimisationGradient descent is the derivative put to work: every neural network learns by following ∂loss/∂weight downhill, millions of parameters at a time.
Any question that asks for the largest volume, the shortest time or the lowest cost becomes "where is the derivative zero?" — and then a picture tells you which one wins.
Square roots come from Newton's method, sin and eˣ from Taylor polynomials. The functions in a maths library are calculus run a few steps at a time.