Noncentral normal distribution

1 Background and Definition

1.1 Relation to the standard normal distribution

The noncentral normal distribution refers to a Gaussian random variable whose mean is shifted away from the value implied by the baseline (central) model, commonly in a way governed by an additional “noncentrality” parameter. In its most common probabilistic form, if \(X\) is modeled as normal with mean \(\mu\) and variance \(\sigma^2\), then the standardized variable \[ Z=\frac{X-\mu}{\sigma} \] is standard normal. The “noncentral” perspective instead emphasizes that the effective mean offset (relative to some baseline) is controlled by a parameter that does not vanish, even when other parameters are held fixed.

In applications, this parameter often enters through a projection, conditioning step, or latent-variable construction, so the resulting observed quantity behaves as if it were normal but with a shift determined by the underlying noncentrality.

1.2 Noncentrality parameter and its interpretation

A noncentrality parameter typically represents the magnitude of an offset relative to noise or variability. Informally, it measures “how strongly” the mean is pushed away from the central reference. Depending on the modeling route, it may correspond to:

A true mean shift in the observation equation,
A deterministic component appearing after transformation of a random vector,
A signal-to-noise effect that survives conditioning or marginalization.

Because the name “noncentral” is used broadly across related families (for example, in noncentral chi-square settings), the key idea is that the parameter quantifies an asymmetry or offset that would be absent in the central case.

1.3 Parameterizations and notation

The distribution is often written in the same parameterization as the normal family: \[ X \sim \mathcal{N}(\mu,\sigma^2). \] The noncentrality is then introduced by relating \(\mu\) to a baseline quantity and an extra parameter. A common convention is to express the shift in standardized units, for instance: \[ \mu = \sigma \delta + \mu_0, \] where \(\delta\) is a noncentrality parameter (interpretable as a mean shift measured in standard deviations) and \(\mu_0\) is the central reference mean. Different sources choose different baselines; the essential point is that the additional parameter controls the mean displacement.

1.4 Where the noncentral normal arises

Noncentral normal behavior emerges whenever a Gaussian component is observed after a step that creates a deterministic offset or induces a shifted conditional mean. Typical routes include:

Projection of a Gaussian vector onto a direction with nonzero alignment with a deterministic mean component;
Conditioning in linear-Gaussian models where the conditional expectation includes a nonzero term;
Linear transformation of a normal variable plus a deterministic shift;
Measurement models in which a latent “signal” adds to noise, and the observed variable remains Gaussian.

Although the resulting distribution is still normal, the “noncentral” label highlights that the mean offset is not assumed to be zero and is tracked explicitly.

2 Distribution Properties

2.1 Probability density function (PDF)

For \(X\sim \mathcal{N}(\mu,\sigma^2)\), the PDF is \[ f_X(x)=\frac{1}{\sqrt{2\pi\sigma^2}}\exp\!\left(-\frac{(x-\mu)^2}{2\sigma^2}\right). \] When \(\mu\) is parameterized via a noncentrality parameter \(\delta\) (for example, \(\mu=\mu_0+\sigma\delta\)), the PDF becomes a function of \(\delta\) through the shifted mean. The functional form remains Gaussian; only the mean location changes with the additional parameter.

2.2 Cumulative distribution function (CDF)

The CDF is \[ F_X(x)=\Phi\!\left(\frac{x-\mu}{\sigma}\right), \] where \(\Phi\) is the standard normal CDF. As with the PDF, if the noncentrality parameter influences \(\mu\), then it enters through the standardized argument \((x-\mu)/\sigma\).

2.3 Moment generating function (MGF) and characteristic function

The MGF exists for all real \(t\) and is \[ M_X(t)=\exp\!\left(\mu t+\frac{1}{2}\sigma^2 t^2\right). \] The characteristic function is \[ \varphi_X(t)=\exp\!\left(i\mu t-\frac{1}{2}\sigma^2 t^2\right). \] If \(\mu\) depends on a noncentrality parameter, both transforms inherit that dependence through the linear term in \(t\).

2.4 Mean, variance, and higher moments

For \(X\sim\mathcal{N}(\mu,\sigma^2)\),

Mean: \(\mathbb{E}[X]=\mu\),
Variance: \(\mathrm{Var}(X)=\sigma^2\),
Higher central moments: depend only on \(\sigma^2\), as in the standard normal family.

Noncentrality affects the mean (and thus raw moments), but central moments of order \(\ge 2\) are unchanged relative to an ordinary normal with the same \(\sigma^2\). For example, the third central moment is \(0\), and the fourth central moment is \(3\sigma^4\).

2.5 Limiting and special cases

Several special cases follow directly:

If the noncentrality parameter corresponds to \(\delta=0\), then \(\mu\) equals the baseline mean \(\mu_0\), yielding the “central” normal model.
As \(\sigma^2\to 0\), the distribution concentrates at \(x=\mu\), approaching a degenerate distribution.

- If \(	\delta	\to\infty\) while \(\sigma\) is fixed, the mean diverges; probabilities concentrate far into the tails relative to a fixed evaluation point.

2.6 Connection to standardized forms

Many calculations reduce to standard normal quantities via standardization: \[ \frac{X-\mu}{\sigma}\sim \mathcal{N}(0,1). \] In this view, noncentrality is simply a way of tracking how \(\mu\) shifts with respect to a baseline. As a result, most probabilities and expectations can be expressed using \(\Phi\), \(\varphi\) (the standard normal PDF), and algebraic substitutions.

3 Estimation and Inference

3.1 Likelihood for the noncentral normal model

Given independent observations \(x_1,\dots,x_n\) from a normal model with parameters \((\mu,\sigma^2)\) (with \(\mu\) tied to a noncentrality parameter in specific applications), the likelihood is \[ L(\mu,\sigma^2)=\prod_{i=1}^n \frac{1}{\sqrt{2\pi\sigma^2}} \exp\!\left(-\frac{(x_i-\mu)^2}{2\sigma^2}\right). \] In parameterizations where \(\mu=\mu_0+\sigma\delta\) or \(\mu=\mu(\delta)\), the likelihood becomes a function of \(\delta\) and any remaining unknown parameters.

3.2 Maximum likelihood estimation (MLE)

For the standard normal parameterization:

The MLE of the mean is the sample mean: \(\hat{\mu}=\bar{x}\).
The MLE of the variance is \(\hat{\sigma}^2=\frac{1}{n}\sum_{i=1}^n (x_i-\bar{x})^2\).

When the mean is constrained by a relationship to a noncentrality parameter (e.g., \(\mu=\mu_0+\sigma\delta\)), the MLE can be obtained by maximizing the likelihood over the reduced parameter set. In many practical setups, this reduces to solving for \(\delta\) using the fitted \(\mu\) and \(\sigma\), followed by checking how uncertainty propagates.

3.3 Method of moments

Moment estimators equate theoretical moments to sample moments. For a normal model, the sample mean estimates \(\mu\) and the sample second central moment estimates \(\sigma^2\). Under a noncentral parameterization, \(\delta\) is then inferred from the estimated mean and variance through the chosen mapping between \(\delta\) and \(\mu\).

3.4 Confidence intervals and hypothesis tests (general)

Inference typically proceeds by leveraging the exact normal-theory distribution of sample statistics:

For known \(\sigma^2\), a normal-based confidence interval for \(\mu\) (and hence for \(\delta\) after transformation) is straightforward.
For unknown \(\sigma^2\), the usual t-distribution arises for \(\mu\)-based inference.

If the noncentrality parameter is the target, confidence intervals are constructed by either:

Building an interval for \(\mu\) (or \(\mu/\sigma\) depending on the mapping) and transforming it to \(\delta\), or
Directly profiling the likelihood over \(\delta\) and using standard likelihood-ratio logic.

Hypothesis testing follows the same logic, with the null often corresponding to a “central” value of the noncentrality (commonly \(\delta=0\)), though the exact null depends on the modeling definition.

3.5 Model checking and diagnostics

Even with a noncentral mean, the Gaussian assumptions remain central. Diagnostics include:

Residual plots to assess deviations from normality and constant variance,
Normal Q-Q plots of residuals (after centering by the estimated mean),
Checks for outliers or influential points,
Verification that the assumed variance model is compatible with the data.

Because the mean shift is explicitly parameterized, mismatches in \(\sigma^2\) or non-Gaussian tails can be masked if only the mean is fitted; variance diagnostics are therefore important.

4.1 Linear transformations of noncentral normal variables

If \(X\sim\mathcal{N}(\mu,\sigma^2)\) and \(Y=aX+b\) with \(a\neq 0\), then \[ Y\sim\mathcal{N}(a\mu+b,a^2\sigma^2). \] In a noncentral parameterization, the transformation alters the mean and rescales the variance. As a result, the effective noncentrality (as measured relative to the transformed noise scale) changes predictably.

4.2 Sums and differences under model assumptions

For independent normal variables \(X_1\sim\mathcal{N}(\mu_1,\sigma_1^2)\) and \(X_2\sim\mathcal{N}(\mu_2,\sigma_2^2)\), the sum \[ S=X_1+X_2 \] is normal with mean \(\mu_1+\mu_2\) and variance \(\sigma_1^2+\sigma_2^2\). The same holds for differences \(X_1-X_2\) with variance \(\sigma_1^2+\sigma_2^2\). Noncentrality parameters combine through their effect on the means, while variances add.

4.3 Conditional distributions (general treatment)

In linear-Gaussian models, conditional distributions remain Gaussian. If a joint vector is multivariate normal with nonzero mean components, conditioning typically produces a normal distribution whose conditional mean includes deterministic terms that reflect the noncentrality. While the exact formula depends on the covariance structure, the general pattern is:

Conditional variance depends only on the covariance matrices,
Conditional mean shifts according to both the observed value and the nonzero mean component.

This is one common mechanism by which a noncentral normal emerges: the conditioning step preserves normality but transfers the underlying offset into the conditional expectation.

4.4 Truncation and censoring considerations

If a normal variable is observed only when it falls within a region (truncation) or is only partially recorded (censoring), the resulting distribution is no longer a simple normal. However, the noncentral mean shift influences the probability of falling into each observable region and therefore changes the conditional likelihood.

Inference under truncation/censoring often involves:

Writing the likelihood using CDF differences over the observed region,
Computing moments of the truncated normal, which depend on standardized truncation points that include the shifted mean,
Using numerical routines for stability when tail probabilities are small.

5 Computation and Numerical Methods

5.1 Evaluating PDF/CDF numerically

Direct evaluation of the PDF and CDF relies on stable implementations of \(\exp(\cdot)\), the standard normal PDF, and the standard normal CDF \(\Phi\). For large positive or negative arguments, care is needed to avoid underflow or loss of precision in tail regions. Many libraries provide specialized functions for log-CDFs or complementary CDFs to maintain accuracy.

5.2 Approximations and series expansions (overview)

When extremely small tail probabilities matter (for example, in inverse probability calculations), approximations can be used:

Tail approximations for the standard normal (e.g., asymptotic expansions),
Series-based approximations for related special functions when embedded in more complex expressions.

For the noncentral normal itself, computations remain relatively straightforward because the distribution is standard normal after centering and scaling; nonetheless, numerical issues can arise in derived quantities such as confidence bounds, likelihood surfaces, or transformed parameters.

5.3 Stability issues and implementation notes

Potential numerical challenges include:

Computing \(\Phi(z)\) when \(z\) is very negative or very positive,
Subtracting nearly equal CDF values in probability differences (which can cause catastrophic cancellation),
Optimizing likelihoods where the parameterization involves \(\sigma\) inside the mean shift (which can create strong curvature).

Stable strategies include using log-probabilities, complementary CDFs, and careful parameter transformations to reduce ill-conditioned optimization landscapes.

5.4 Software considerations and best practices

Typical best practices for implementations:

Use established statistical libraries for \(\Phi\) and related functions (rather than custom approximations),
Prefer parameterizations that keep optimization unconstrained when possible (e.g., working with \(\log\sigma\)),
Validate code by comparing against known closed-form results for means, variances, and basic probabilities,
Benchmark computations in extreme regimes (large noncentrality, small variance, or high censoring thresholds).

6 Applications

6.1 Measurement error and calibration contexts (general)

In calibration problems, observed values often combine a true signal with noise. When the true signal contains a deterministic component (such as a bias or reference offset), the measured output can be modeled as normal with a nonzero mean shift relative to a central target. The noncentrality parameter then summarizes how large the systematic offset is compared to the noise level.

6.2 Signal processing and detection theory (general)

In detection frameworks, a test statistic is frequently modeled as Gaussian under different hypotheses. When one hypothesis includes a deterministic signal component added to noise, the distribution of the statistic under that hypothesis becomes a noncentral normal relative to the central (noise-only) case. This underlies calculations of error rates, thresholds, and receiver operating characteristics, where the shifted mean changes separability between distributions.

6.3 Quality control and process shifts (general)

Quality control procedures often assume that measurements fluctuate normally around a nominal target, with occasional shifts due to changes in the process. If a process shift changes the mean level but does not alter the variance, the post-shift observations are well described by a noncentral normal model. Such models support reasoning about detection delays and the likelihood of alarms under shifted conditions.

6.4 Bayesian modeling perspective (high level)

In Bayesian treatments, the noncentrality parameter can be treated as an unknown quantity with a prior distribution, allowing inference about the size of the mean offset after observing data. Conjugate structure often remains available under normal likelihood assumptions, enabling analytic updates for \(\mu\) (or for transformed versions of \(\delta\)) when variance is known or modeled with standard conjugate priors. When truncation or censoring is present, Bayesian computation typically relies on numerical integration or sampling methods, but the shifted-mean interpretation stays central.