Interactive demo · Probability and Statistical Inference · Chapter 14

Conditioning a normal vector: keep the points in a strip

Conditioning on an event of probability zero, such as X1 = 4 or X1 + X2 = 2, can be pictured by simulation: draw many points and keep those within h of the line aTx = c. Left: the cloud, the strip and its points, the line of conditional means E(X | aTX = w) as w varies (for a coordinate, the regression line), and when conditioning on a coordinate also the other regression line and the major axis (dotted). Right: the histogram of the other coordinate for the points in the strip, against the exact conditional normal law (gold) and the marginal law (grey). For a normal vector the conditional law is again normal, its mean moves linearly with c, and its variance does not depend on c at all.

Condition on

Value c

Strip half-width h

Correlation ρ

Standard deviation of X2

Number of points

Try these

Points are X1 = μ1 + σ1Z1, X2 = μ2 + σ2(ρZ1 + √(1 − ρ²)Z2) from a seeded generator. With W = aTX and Y the displayed coordinate, the exact conditional law is Y | W = c ~ N(μY + β(c − μW), σY² − β²σW²) with β = Cov(Y, W)/Var(W); the "exact for this strip" values add the effect of the strip's width, using the truncated normal law of W given |W − c| ≤ h.

In the lecture notes