You already know: distributions, expectation, variance, Bayes' rule and confidence intervals (Probability); eigenvectors and covariance (Linear systems, SVD and PCA).
Why this matters for Atlas. Real data has many variables at once, so we need joint distributions: how variables vary together (covariance, correlation), what one tells us about another (conditionals), and when they're unrelated (independence). The multivariate Gaussian is the workhorse model for all of it. And the central limit theorem explains why averages (eval accuracy, A/B lifts, mini-batch gradients) are approximately Gaussian with error shrinking like 1/√n, which is the justification behind nearly every error bar in this course.
1. Joint, marginal and conditional distributions
The question. How do you describe two variables together?
Start from nothing. A joint distribution p(x, y) gives probabilities for pairs. Summing (or integrating) out one variable gives a marginal: p(x) = Σ_y p(x, y). Dividing by a marginal gives a conditional: p(y | x) = p(x, y) / p(x). Variables are independent if p(x, y) = p(x)p(y), i.e., knowing x tells you nothing about y. Covariance Cov(X, Y) = E[(X − μx)(Y − μy)] measures linear co-movement; correlation rescales it to [−1, 1]. Zero correlation doesn't imply independence (a relationship can be nonlinear).