mathematics//calculus

Calculus is the branch of mathematics that studies rates of change and accumulation, and it is the language in which almost every model of a machine is written: a drone, a tank or a motor arrives as a rule for how fast its state changes, and calculus is what turns that rule into a trajectory, a sensitivity or a linear model a controller can be designed on. Physics states how a thing changes and leaves its position to be worked out, so reading it takes four ideas, and an engineer meets each of them in a different part of the loop.

Calculus is the branch of mathematics that studies rates of change and accumulation, and it is the language in which almost every model of a machine is written: a drone, a tank or a motor arrives as a rule for how fast its state changes, and calculus is what turns that rule into a trajectory, a sensitivity or a linear model a controller can be designed on. Physics states how a thing changes and leaves its position to be worked out, so reading it takes four ideas, and an engineer meets each of them in a different part of the loop.

The derivative is the rate of change at an instant: velocity is the derivative of position, and an ODE is nothing more than a rule for derivatives. It is also the question how much does the output move if I nudge this input, which is how a sensitivity is asked. The integral runs the other way, accumulating a rate into a state: position from velocity, volume from flow, angle from a gyroscope's rate. Every simulator, every dead reckoning estimate and the integral term of every PID compute one, by summing small pieces. The Taylor polynomial says that close up every smooth function looks like a line; it is the root of linearization, the step that lets a nonlinear plant borrow the whole linear toolkit, and its most used case is the small-angle approximation of a drone near hover. The chain rule multiplies rates along a chain of dependencies, and it reappears as backpropagation when the chain is a neural network.

With several variables the same ideas take a matrix form. A partial derivative moves one input and holds the rest. The Jacobian collects them for a vector function (one row per output), and it is what linearization, the extended Kalman filter and an error budget actually compute. The gradient, one derivative per input of a scalar function, points uphill and is what gradient descent follows (gradient, divergence and curl). The Hessian holds the second derivatives, the curvature that decides how hard an optimization is. Vector calculus extends all this to fields (temperature along a bar, the velocity of a fluid), and the Laplace transform turns a linear ODE into algebra, which is where transfer functions come from.

A computer does none of this symbolically inside the loop. An autopilot integrates by adding IMU samples a thousand times a second, a derivative of a measured signal is a difference of two samples (filtered, because differencing amplifies noise), and a training framework gets its gradients by program transformation. How each is done, and how it fails, is the subject of numerical methods: numerical integration, finite differences and automatic differentiation.

The depth an engineer needs is reading: knowing what a derivative in an equation describes and which way a sign pushes, more than computing one by hand. The rare step that needs a derivation is linked from the note that uses it.