Measurements in biology rarely have the same noise at every level: counts have a variance that grows with their mean, and microarray intensities have a constant noise floor plus a noise proportional to the signal. If Var(X) = v(μ), the delta method says that sd(g(X)) ≈ |g′(μ)|√v(μ), so choosing g′ ∝ 1/√v makes the spread the same everywhere. The plot shows the standard deviation of g(X) against the mean on a log axis: exactly (solid), by the delta method (dashed) and from 200 simulated replicates at each mean (dots).
Variance functions: Poisson v(μ) = μ; negative binomial (gamma-Poisson) v(μ) = μ + αμ2; microarray model X ~ N(μ, a2 + b2μ2). Their variance-stabilising transformations are 2√x, (2/√α) asinh(√(αx)) and asinh(bx/a)/b. Exact values: sums over the probability mass function, or Simpson's rule for the normal model (the logarithm is taken only on the positive part, which is reported). Simulations use a seeded generator.