Normal CDF Φ(z) Approximation

Normal CDF: cumulative probability under the bell curve

For Z ~ N(0, 1), the cumulative distribution function of the standard normal gives you Φ(z) = P(Z ≤ z). There's no elementary closed form for it, which is why everyone falls back on numerical approximations. The usual options are the Abramowitz & Stegun (1964) rational forms (error below 10⁻⁷), Hastings polynomial approximations, or the error-function identity Φ(z) = (1 + erf(z/√2)) / 2. A few values worth memorizing: Φ(0) = 0.5 (the median), Φ(1) ≈ 0.8413, Φ(1.96) ≈ 0.975 (95% one-tailed), Φ(2.576) ≈ 0.995 (99% one-tailed). The curve is monotonic and S-shaped, running from 0 at −∞ up to 1 at +∞, and by symmetry Φ(−z) = 1 − Φ(z).

Applications

It shows up in hypothesis testing (p-values for z-tests), confidence intervals, Monte Carlo simulation, financial risk modelling (VaR), machine learning (Bayesian inference, Gaussian processes), and quality control (Six Sigma yield calculations). Basically, wherever you need to convert a z-score into a probability — or go the other way with the inverse CDF, the quantile function.

FAQ

Why is there no exact formula? Because the integral of e^(−t²/2) simply has no antiderivative among the elementary functions. So we fall back on series, rational approximations, or numerical integration of erf.

How accurate is the approximation? The Abramowitz & Stegun rational forms keep absolute error under 7.5·10⁻⁸, which is plenty for ordinary statistics. Financial pricing engines sometimes swap in higher-precision variants when they need them.

Two-tailed vs one-tailed? For one-tailed work you read Φ(z) straight off. For the two-tailed p-value of |Z| ≥ |z|, use 2·(1 − Φ(|z|)) instead.

How do I get z from a probability? That calls for the inverse CDF Φ⁻¹(p), also known as the probit or quantile function. It's usually computed with Beasley-Springer-Moro or rational approximations like Wichura's AS241.