In words
What it is, why it matters, and what it is like.
Why am I learning this?
The Kalman filter is the standard way to track something moving when your measurements are noisy. It powers GPS in your phone, keeps self-driving cars aware of where they are, and is used in robotics, finance, and even to smooth sensor data in AI systems. Learning it here gives you a foundation for understanding how AI systems handle uncertainty — a key skill for working with real-world data.
The idea, in plain terms
Imagine you're trying to find your friend in a crowded market. Your friend sends you a text saying they're near the chai stall, but they might be off by a bit. You also have a rough guess from the last time you saw them. You don't just trust the text — you combine it with what you already believed, weighting each by how reliable you think it is. That's the Kalman filter: it's a way to combine a prediction (where you think the system should be) with a noisy measurement (what a sensor tells you) to get a best estimate. It does this again and again, each time using the new estimate to make the next prediction. It's recursive because it uses the previous estimate as a starting point, and it's optimal under certain conditions — meaning it gives the best possible estimate given the information you have.
An analogy
Think of tracking a bird flying across a field. You have two sources of information: your own guess from watching it move (the prediction) and a camera that occasionally snaps a picture (the measurement). The camera is good but not perfect — sometimes the bird is a blur. The Kalman filter is like having a wise friend who watches the bird with you. Every moment, your friend makes a prediction: 'The bird should be about here, because it was there and moving this fast.' Then, when a new photo arrives, your friend doesn't just replace the prediction with the photo. Instead, they blend the two: if the camera is usually very accurate (low noise), they trust the photo more; if the bird is moving erratically (high process noise), they trust the prediction more. The result is a smooth, accurate track that never jumps around wildly. The analogy breaks down when the bird does something truly unpredictable, like suddenly flying behind a tree and changing direction — the filter assumes the movement follows a pattern (its dynamic model), so it can be caught off guard by sharp, unexpected changes.
Definition
A Kalman filter is a recursive algorithm that estimates the state of a system by alternating between a prediction step (using a model of how the system moves) and an update step (incorporating a noisy measurement), producing an estimate with minimal error when the system is linear and the noise is Gaussian.
Where this sits
This concept belongs to Numerical Methods, which is about computing mathematical answers with finite precision and finite time. The Kalman filter is a numerical technique for solving problems in tracking and estimation. It builds on Linear Algebra (you'll need vectors and matrices to represent state, covariance, and the equations), and it connects to Probability (the filter treats noise as random variables with a mean and variance). You have notes on Least Squares Regression, which is a cousin — both find a 'best' estimate by minimising error. The Kalman filter extends this to a time-varying system, updating the estimate as new data arrives.
