Electronics Guide

Normal (Gaussian) Distribution

The normal distribution is the most commonly encountered distribution in EMC work. Many physical quantities that result from the sum of numerous independent factors tend toward normal distributions by the Central Limit Theorem. EMC emissions and immunity levels, when expressed in logarithmic units (dB), often follow approximately normal distributions.

The normal distribution is characterized by two parameters: the mean μ, which locates the center of the distribution, and the standard deviation σ, which sets its spread. For a normally distributed variable:

• Approximately 68% of values fall within one standard deviation of the mean
• Approximately 95% of values fall within two standard deviations of the mean
• Approximately 99.7% of values fall within three standard deviations of the mean

When analyzing EMC data, it is important to verify that the assumption of normality is reasonable. Graphical methods such as normal probability plots (Q-Q plots) or formal statistical tests like the Shapiro-Wilk test can assess normality. Data that appear non-normal in linear units may become approximately normal when transformed to logarithmic (dB) scale.

Log-Normal Distribution

When the logarithm of a variable follows a normal distribution, the variable itself follows a log-normal distribution. This is common in EMC because many electromagnetic quantities multiply together (coupling factors, attenuation, gain) rather than adding. When multiplied quantities are converted to decibels, they add, often producing normally distributed dB values and hence log-normally distributed linear values.

Log-normal distributions are inherently asymmetric with a long tail toward higher values. This characteristic is physically meaningful in EMC: while emissions levels cluster around typical values, occasional units may exhibit significantly higher emissions due to resonances, manufacturing variations, or other factors that multiplicatively combine.

When working with log-normal data, analysis is often performed in the logarithmic domain where normal distribution methods apply, then results are transformed back to linear units if needed.

Rayleigh and Rice Distributions

The Rayleigh distribution arises in EMC when measuring the magnitude of a signal composed of many random phasors, such as the field strength in a reverberation chamber or multipath propagation environment. When there is no dominant component and all phasors have equal mean amplitude, the magnitude follows a Rayleigh distribution.

The Rayleigh magnitude carries a companion result that is easier to work with. If the magnitude of one rectangular field component is Rayleigh distributed, its square, the received power in that component, is exponentially distributed, which is the chi-square distribution with two degrees of freedom. The two degrees of freedom are the real and imaginary parts of that single component, each an independent zero-mean Gaussian of equal variance. Summing all three rectangular components adds six such Gaussians, so the total field power follows a chi-square distribution with six degrees of freedom and the total field magnitude follows a chi distribution with six degrees of freedom. Note that the largest of the three components, the statistic several field-uniformity procedures actually record, is the maximum of three Rayleigh variables and is not itself chi distributed; it must be treated as an extreme value.

The Rice (Rician) distribution generalizes the Rayleigh distribution to include a dominant component. In EMC, Rice distributions apply when measuring fields that include both a direct coupling path and scattered components. The ratio of the dominant component power to the scattered component power, the K-factor, sets the shape: at K = 0 the distribution reduces exactly to Rayleigh, and as K grows large the distribution approaches a normal distribution centered on the amplitude of the dominant component.

These distributions are central to reverberation chamber theory, where the ideal well-stirred chamber is defined by exactly the Rayleigh and chi-square behavior described above, and departures from it signal insufficient stirring or a residual direct path. They also describe field uniformity in anechoic chambers whose absorber performance falls short at the low end of the frequency range.

Uniform and Triangular Distributions

Uniform distributions apply when a quantity is equally likely to take any value within a defined range. In EMC, uniform distributions often represent manufacturing tolerances when only the tolerance limits are known without additional information about the distribution within those limits.

Triangular distributions provide a simple model when the most likely value and the range are known. They are often used in uncertainty analysis when more detailed distribution information is unavailable but it is reasonable to assume values near the center are more likely than values near the extremes.

Both distributions earn their place in EMC through the Type B evaluation of uncertainty, where a component is characterized from information other than repeated measurement. Assigning a shape converts a stated bound into a standard uncertainty. For a rectangular distribution of half-width a the standard uncertainty is a/√3, and for a symmetric triangular distribution of the same half-width it is a/√6. A quoted component tolerance of ±5% therefore contributes a standard uncertainty of about 2.9% if nothing more is known, or about 2.0% if the manufacturer's process makes central values more likely. Uncertainty Analysis develops this treatment in full.

Extreme Value Distributions

Extreme value distributions describe the distribution of the maximum (or minimum) of many samples. In EMC, these distributions apply when determining the expected maximum emissions from a large production population or the minimum immunity level across a fleet of products.

The Gumbel distribution (Type I extreme value) is the limiting form for data drawn from a normal or log-normal parent. The statement requires care: the maximum of n normal observations does not follow a Gumbel distribution exactly at any finite n. Rather, once the maximum is centered and scaled by sequences that depend on n, its distribution converges to the Gumbel form as n grows without bound. For the normal parent that convergence is notoriously slow, of order 1 divided by the logarithm of n, so a Gumbel fit to the maxima of small EMC samples is an approximation whose error deserves checking rather than an identity.

Two consequences matter for EMC. First, the expected maximum of a sample keeps rising as the sample grows, so the worst unit observed in a batch of ten is a systematically optimistic estimate of the worst unit in a production run of ten thousand. Second, the familiar 1 divided by the square root of n improvement in precision applies to estimates of a mean, not to estimates of a maximum; an extreme has no fixed population value to converge to, and the uncertainty in an extrapolated extreme falls far more slowly with sample size. Relating a small-sample test result to the behavior of the production population therefore calls for a tolerance limit or an explicit extreme-value model, not for the arithmetic of confidence intervals on the mean.

Confidence Intervals for the Mean

When estimating the mean of a normally distributed quantity from a sample, the confidence interval depends on the sample mean, sample standard deviation, sample size, and the desired confidence level.

For a sample of size n with mean x̄ and standard deviation s, the 100(1 − α)% confidence interval for the population mean is:

x̄ ± t(α/2, n − 1) s / √n

Here α is the significance level and t(α/2, n − 1) is the critical value of the t distribution with n − 1 degrees of freedom that leaves an area α/2 in the upper tail. For 95% confidence (α = 0.05) that critical value is 12.71 at n = 2, 2.26 at n = 10, 2.05 at n = 30, and it approaches 1.96 as n grows without bound. The interval assumes either a normally distributed population or a sample large enough for the Central Limit Theorem to apply to the mean.

Two factors shrink the interval as the sample grows: the explicit 1/√n, and the t critical value itself, which falls steeply over the first several observations. Once n reaches a few tens the t value has all but settled, and from there the 1/√n term governs alone, so halving the half-width requires roughly four times the sample. At very small n the gain from one more unit is much larger than 1/√n suggests, which is why the step from three units to six buys more than the step from thirty to sixty. Either way the message for EMC test planning is the same: tight confidence intervals are not available from small samples.

Confidence Intervals for Percentiles

EMC specifications often concern percentiles rather than means. For example, a requirement might state that 95% of production units must comply with an emissions limit. A confidence interval on the mean says nothing useful about such a requirement, and the two are easy to confuse. A confidence interval bounds where a single population parameter lies; a tolerance interval bounds where a stated proportion of the individual units lies. Compliance limits are statements about units, so they call for tolerance intervals.

For normally distributed data, the one-sided upper tolerance limit takes the form:

x̄ + k s

The 95/95 limit, meaning 95% confidence that at least 95% of the population falls below it, is the common choice. The factor k is not a t quantile or a z quantile. It is a one-sided normal tolerance factor, obtained from the non-central t distribution with n − 1 degrees of freedom and non-centrality parameter z_p√n, where z_p is the standard normal quantile for the coverage proportion p. For 95/95 coverage the factor is 4.20 at n = 5, 2.91 at n = 10, 2.22 at n = 30, and 2.07 at n = 50, falling toward 1.645, the 95th percentile of the standard normal distribution, only in the limit of an infinite sample. The excess over 1.645 is the price of estimating σ from the sample. The corresponding two-sided factors, which bound both tails at once, are larger still: about 3.38 at n = 10 and 2.55 at n = 30.

These factors are what make small-sample compliance arguments expensive. A ten-unit test demands that the sample mean sit nearly three sample standard deviations below the limit; a thirty-unit test relaxes that to about 2.2. Any of this rests on the assumption of normality in the dB domain, so a probability plot of the sample is a necessary companion to the calculation. Where normality cannot be defended, a distribution-free tolerance bound built from order statistics is available instead, at the cost of a much larger sample.

Confidence Intervals for Standard Deviation

Quantifying variability is often as important as quantifying central tendency in EMC. For a sample drawn from a normal population, the quantity (n − 1)s²/σ² follows a chi-square distribution with n − 1 degrees of freedom, and inverting that statement gives the interval.

For a sample of size n with standard deviation s, the 100(1 − α)% confidence interval for the population standard deviation σ runs from:

√[(n − 1)s² / χ²(α/2, n − 1)]

up to:

√[(n − 1)s² / χ²(1 − α/2, n − 1)]

The tail convention has to be stated, because the formula is correct under one reading and inverted under the other. Here χ²(α/2, n − 1) denotes the critical value that leaves an area α/2 in the upper tail, matching the convention used for the t critical value above. That is the larger of the two critical values, and because it sits in the denominator it produces the smaller bound. Texts that index the chi-square distribution by lower-tail area write the same interval with the two subscripts exchanged. Reading the two bounds off a table and checking that the lower one is indeed below s is the quickest guard against the swap.

The interval is asymmetric, extending further above s than below it. That asymmetry is physically meaningful: one can be fairly confident that the true variability is not much less than observed, while it could be substantially more. It also carries a warning. The interval depends on normality far more critically than the interval for the mean does, because no Central Limit Theorem rescues it. A sample mean from a non-normal population is still approximately normal, but a sample variance from a heavy-tailed population is not approximately chi-square, and the resulting interval can be badly wrong. Verify normality before quoting it.

Confidence intervals for the standard deviation require even larger samples than those for the mean. At n = 10 and 95% confidence, with 9 degrees of freedom, the critical values are 19.02 and 2.70, giving an interval from 0.69s to 1.83s, a factor of 2.65 between its limits. Narrowing it to 0.80s to 1.34s requires n = 30. An EMC variability figure quoted from ten units is therefore uncertain by better than a factor of two, which is worth remembering before treating one production line's spread as different from another's.

Fundamentals of Hypothesis Testing

A hypothesis test begins with a null hypothesis (H₀) representing the default assumption, and an alternative hypothesis (H₁) representing the effect the investigator seeks to demonstrate. The test calculates the probability of observing data at least as extreme as the actual observations if the null hypothesis were true. This probability, called the p-value, is compared against a significance level α to decide whether to reject the null hypothesis.

Two types of errors can occur:

• Type I error (false positive): Rejecting the null hypothesis when it is true. The probability of Type I error is controlled by the significance level α, typically set at 0.05.
• Type II error (false negative): Failing to reject the null hypothesis when it is false. The probability of Type II error is denoted β, and 1 − β is called the power of the test.

In EMC, these errors have practical interpretations. A Type I error might mean incorrectly concluding a product fails when it actually complies. A Type II error might mean approving a non-compliant product. The relative consequences of these errors should inform the choice of significance level and required sample sizes.

Two misreadings of these quantities are common enough in EMC reports to be worth naming. The first treats the p-value as the probability that the null hypothesis is true. It is not; it is a probability computed assuming the null hypothesis is true, and converting it into a probability about the hypothesis requires a prior, which is a Bayesian calculation. The second treats a large p-value as proof of the null hypothesis. Failing to reject a hypothesis is not the same as establishing it, and with the small samples typical of EMC work a test may fail to reject simply because it had too little power to detect anything. When the question is genuinely whether two conditions are equivalent, an equivalence test, which places the burden of proof on equivalence by testing against a stated tolerance, answers it directly; an ordinary significance test does not.

One-Sample Tests

One-sample tests compare sample data against a specified value. The one-sample t-test determines whether a sample mean differs significantly from a hypothesized value, such as an EMC limit.

For compliance testing, the hypothesis structure is typically:

• H₀: The true mean equals or exceeds the limit
• H₁: The true mean is below the limit

This formulation places the burden of proof on demonstrating compliance. The sample must show emissions sufficiently below the limit to provide statistical confidence that the population mean is indeed below the limit.

A test on the mean is nonetheless the wrong instrument for most emissions limits, and the distinction is the single most common statistical error in EMC compliance work. A limit applies to each unit shipped, not to the average of the population. A population whose mean sits comfortably below the limit can still place a substantial fraction of its units above it whenever the spread is wide. Demonstrating that units comply requires a tolerance limit or an explicit proportion-based criterion of the kind described later in this article, not a confidence interval on the mean.

One-sample tests for percentiles are more complex. Non-parametric approaches using order statistics can test whether a specified proportion of the population meets a criterion without distributional assumptions, but require larger sample sizes.

Two-Sample Tests

Two-sample tests compare two groups to determine whether they differ significantly. Applications in EMC include comparing before/after performance of a design change, comparing two measurement methods, or comparing products from different manufacturing lines.

The two-sample t-test compares means from two groups. When sample sizes are equal and variances can be assumed equal, the test is straightforward. When variances may differ, Welch's t-test provides a more robust approach that does not assume equal variances.

The F-test compares variances between two groups, useful when determining whether a design change affected emissions variability as well as emissions level. However, the F-test is sensitive to non-normality and should be supplemented with graphical assessment of the data.

Multiple Comparisons

When performing multiple hypothesis tests simultaneously, the probability of at least one false positive increases. Suppose 20 independent tests are performed at α = 0.05 and every one of the 20 null hypotheses is in fact true. Each test then avoids a false positive with probability 0.95, and all 20 do so with probability 0.95²⁰ = 0.358, leaving a 64.2% chance of at least one false positive, far above the nominal 5% level. The calculation depends on both stated conditions: the tests must be independent, and the family-wise rate applies to the case where no real effect is present.

Correction methods address this multiple comparison problem:

• Bonferroni correction: Test each comparison at α divided by the number of comparisons. For the 20 tests above that means judging each at 0.0025, which holds the family-wise error rate at or below 0.05 whatever the dependence between tests. Simple, valid in general, and conservative: the power to detect any one real effect falls accordingly.
• Holm-Bonferroni method: A step-down procedure that is less conservative than Bonferroni while still controlling the family-wise error rate.
• False Discovery Rate (FDR): Controls the expected proportion of false positives among rejected hypotheses, appropriate when some false positives are acceptable.

In EMC, multiple comparisons arise when testing at multiple frequencies, multiple configurations, or multiple units. The choice of correction method depends on the consequences of false positives and false negatives in the specific application.

Linear Regression

Simple linear regression fits a straight line to data relating a response variable y to a predictor variable x:

y = a + bx + ε

The coefficients a (intercept) and b (slope) are estimated by least squares, minimizing the sum of squared residuals, and ε is the random error term, assumed independent across observations and of constant variance. Key outputs include:

• Coefficient estimates: The fitted values of a and b, with their standard errors and confidence intervals
• R-squared: The proportion of variance in y explained by x, ranging from 0 (no linear relationship) to 1 (perfect linear relationship)
• Residual analysis: Examination of the differences between observed and fitted values to verify model assumptions

In EMC, regressing a level in dB against the logarithm of frequency often reveals the coupling mechanism, because each mechanism has a characteristic slope. A slope of −20 dB per decade is the signature of a single-pole low-pass roll-off, and −40 dB per decade of a two-pole one. A positive slope points the other way, toward a mechanism whose coupling strengthens with frequency; capacitive coupling behaves this way, because the coupling impedance 1/(2πfC) falls as frequency rises. Reading a rising mechanism as a falling one is an easy slip, so fix the sign of the slope before attaching a physical story to it.

Multiple Regression

Multiple regression extends linear regression to multiple predictor variables:

y = a + b₁x₁ + b₂x₂ + ⋯ + b_kx_k + ε

This enables modeling EMC behavior as a function of multiple factors simultaneously. For example, emissions might depend on clock frequency, supply voltage, load current, and temperature. Multiple regression quantifies the individual contribution of each factor while accounting for the others.

Key considerations in multiple regression include:

• Multicollinearity: When predictor variables are correlated with each other, coefficient estimates become unstable. Variance Inflation Factors (VIF) quantify this problem.
• Variable selection: Including unnecessary variables reduces model precision. Stepwise regression, information criteria (AIC, BIC), or regularization methods (LASSO, ridge regression) guide variable selection.
• Interaction effects: The effect of one variable may depend on another. Including a product term x₁x₂ captures the dependency.

Prediction Intervals

Regression models enable prediction of the response for new values of the predictors. Prediction intervals quantify the uncertainty in these predictions, accounting for both the uncertainty in the fitted model and the inherent variability of individual observations.

Prediction intervals are always wider than confidence intervals for the mean response at the same predictor values. The confidence interval addresses uncertainty in the mean; the prediction interval addresses uncertainty in a single observation.

In EMC, prediction intervals are essential when using regression models for compliance assessment. A prediction interval for emissions at an untested frequency must be narrow enough to provide confidence that the limit will not be exceeded, accounting for both model uncertainty and unit-to-unit variability.

One-Way ANOVA

One-way ANOVA compares means across multiple groups defined by a single factor. For example, comparing emissions from three different production lines or four different component suppliers.

The analysis decomposes the total sum of squares into between-group and within-group components. The F-statistic is the ratio of the between-group mean square to the within-group mean square. If the group means were identical, both mean squares would estimate the same error variance and F would sit near 1; a value far above 1, judged against the F distribution with the appropriate pair of degrees of freedom, is evidence that the factor affects the response. A significant F says only that the group means are not all equal; it does not say which differ, nor by how much, and the size of the effect in decibels matters more to an EMC decision than its significance.

Following a significant ANOVA result, post-hoc tests (such as Tukey's HSD or Dunnett's test) identify which specific groups differ from each other, with appropriate adjustment for multiple comparisons.

Two-Way ANOVA

Two-way ANOVA examines two factors simultaneously, along with their potential interaction. For example, analyzing how both operating mode and temperature affect immunity levels.

The analysis partitions variance into:

• Main effect of factor A
• Main effect of factor B
• Interaction between A and B
• Residual (unexplained) variance

An interaction effect means that the effect of one factor depends on the level of the other factor. In EMC, interactions are common: the effect of shielding may depend on frequency, or the effect of grounding scheme may depend on operating current. Identifying and quantifying interactions is essential for developing robust designs.

Random and Mixed Effects

ANOVA factors can be fixed or random. Fixed effects have specific levels of interest (e.g., three specific operating modes). Random effects represent samples from a larger population of levels (e.g., units randomly selected from production).

Mixed models include both fixed and random effects. In EMC, mixed models are valuable when analyzing data with hierarchical structure, such as multiple measurements on each of several units, where unit is a random effect and operating condition is a fixed effect.

Variance components analysis extracts the variance attributable to each random factor, providing direct insight into sources of variability. This supports decisions about where to focus design improvements: reducing between-unit variability requires different strategies than reducing within-unit variability.

Factorial Designs

Full factorial designs test all combinations of factor levels. A 2^k factorial tests k factors at two levels each, requiring 2^k experimental runs. These designs estimate all main effects and all interactions without confounding.

For example, a 2³ factorial with factors A (clock frequency: low/high), B (supply voltage: low/high), and C (load: light/heavy) requires 8 runs and estimates:

• Three main effects (A, B, C)
• Three two-factor interactions (AB, AC, BC)
• One three-factor interaction (ABC)

Fractional factorial designs reduce the number of runs by confounding higher-order interactions with main effects or lower-order interactions. A 2^k-p fractional factorial requires only 2^k-p runs. The resolution of the design indicates the level of confounding: Resolution III designs confound main effects with two-factor interactions; Resolution IV designs keep main effects clear of two-factor interactions but confound those interactions with each other; Resolution V designs keep both main effects and two-factor interactions clear of one another, confounding them only with three-factor and higher interactions.

Screening Designs

When many factors may potentially affect EMC performance, screening designs efficiently identify the most important factors. Plackett-Burman designs can screen up to k-1 factors in k runs (where k is a multiple of 4), estimating main effects under the assumption that interactions are negligible.

Definitive screening designs offer a more robust alternative, accommodating some two-factor interactions and quadratic effects while remaining economical. These designs are particularly useful in early-stage EMC investigations when the factor-response relationships are unknown.

Response Surface Methodology

Response Surface Methodology (RSM) extends DoE to characterize the relationship between factors and response over a continuous region, not just at discrete levels. RSM designs support fitting polynomial models that describe how the response varies across the factor space.

Central Composite Designs (CCD) and Box-Behnken designs are common RSM approaches. They enable fitting quadratic response surfaces:

y = b₀ + ∑ b_ix_i + ∑ b_iix_i² + ∑ b_ijx_ix_j + ε

The first sum runs over the k factors and collects the linear effects, the second collects the pure quadratic effects that give the surface its curvature, and the third runs over distinct pairs i < j and collects the two-factor interactions.

RSM is valuable in EMC for optimization problems: finding the combination of design parameters that minimizes emissions or maximizes immunity margin. The fitted response surface identifies both the optimal settings and the sensitivity of performance to deviations from optimal. That second output is often the more useful one, because a shallow optimum that tolerates ordinary manufacturing spread is worth more in production than a sharp optimum that does not.

Basic Monte Carlo Method

The basic Monte Carlo approach involves:

1. Define the model relating inputs to outputs
2. Specify probability distributions for each input variable
3. Generate random samples from each input distribution
4. Evaluate the model for each sample set
5. Analyze the resulting output distribution

For EMC problems, the model might be a circuit simulation, an analytical coupling model, or a system-level interference calculation. Inputs include component values, geometry, material properties, and environmental factors. Outputs include emissions levels, immunity margins, or interference probabilities.

The number of samples N determines the precision of the results, but the way precision improves with N depends entirely on what is being estimated, and the familiar 1/√N rule covers only the easiest case.

For a mean, the standard error of the estimate is σ/√N, so precision improves as 1/√N and the rule applies as usually stated. Four times the samples halve the error, and the result does not depend on the shape of the output distribution.

For a tail probability p, such as the probability of exceeding an emissions limit, the estimate is a proportion and its standard error is √[p(1 − p)/N]. The absolute error still falls as 1/√N, but the quantity of interest is small, so what matters is the relative error, √[(1 − p)/(pN)]. That grows as the event becomes rarer. Achieving a 10% relative standard error needs roughly 100/p samples: about 1,000 runs at p = 0.1, about 10,000 at p = 0.01, and about 100,000 at p = 0.001. Rare-event estimation is therefore not merely more expensive than mean estimation, it is more expensive in proportion to the rarity, which is exactly why importance sampling earns its place.

For a percentile, precision depends on the density of the output distribution at that percentile, which is low in the tails. Estimating a 99th percentile well takes far more samples than estimating a median from the same runs.

For an extreme value, the sample maximum, the 1/√N rule does not apply at all. The maximum has no fixed value to converge to: as N grows the observed maximum drifts upward without limit for an unbounded output distribution, and the spread of that maximum narrows only as fast as the extreme-value theory of the parent distribution allows, which for a normal parent is of order 1 divided by the logarithm of N. Reporting the largest of N Monte Carlo runs as though it were an estimate with 1/√N precision is a serious error. Extreme quantities call for fitting an extreme-value model to block maxima or threshold exceedances and quoting the uncertainty from that fit.

Variance Reduction Techniques

Advanced Monte Carlo techniques improve efficiency by reducing the number of samples required for a given precision:

• Latin Hypercube Sampling: Ensures samples are spread evenly across the input space, improving coverage compared to purely random sampling.
• Importance Sampling: Concentrates samples in regions that contribute most to the quantity of interest, particularly useful for rare-event probability estimation.
• Control Variates: Uses correlated auxiliary quantities with known expectations to reduce variance in the quantity of interest.
• Antithetic Variates: Pairs negatively correlated samples to reduce variance.

In EMC, importance sampling is particularly valuable when estimating the probability of exceeding emissions limits, which may occur in only a small fraction of the parameter space.

Sensitivity Analysis

Monte Carlo results support sensitivity analysis identifying which input variables most affect output variability. Several approaches exist:

• Correlation analysis: Examines correlations between inputs and outputs across the Monte Carlo samples.
• Regression analysis: Fits a regression model to the Monte Carlo data, with standardized coefficients indicating relative sensitivity.
• Variance-based sensitivity (Sobol indices): Decomposes output variance into contributions from individual inputs and their interactions.

Sensitivity analysis guides design decisions by identifying which parameters most need to be controlled or optimized. In EMC, this might reveal that emissions are sensitive to PCB trace length but insensitive to component tolerance, directing attention to geometry control rather than component selection.

Bayes' Theorem

Bayes' theorem relates the posterior probability (belief after seeing data) to the prior probability (belief before data) and the likelihood (probability of the data given the parameter):

P(θ | data) ∝ P(data | θ) P(θ)

In words: the posterior is proportional to the likelihood times the prior. What the proportionality omits is the denominator of the full theorem, P(data) = ∫ P(data | θ) P(θ) dθ, known as the marginal likelihood or the evidence. It does not depend on θ, so it changes no shape and can be ignored whenever the question is the relative plausibility of different parameter values, which is what sampling methods work with. It cannot be ignored in two situations: whenever a properly normalized probability is needed, since the posterior must integrate to one, and whenever models rather than parameters are being compared, since the Bayes factor between two models is precisely the ratio of their evidences.

The prior distribution encodes existing knowledge before testing. In EMC, prior knowledge might come from simulation results, similar products, physics-based bounds, or expert judgment. As test data accumulate, the posterior distribution narrows, reflecting increased knowledge about the parameter.

Prior Selection

Choosing appropriate priors requires balancing informativeness with objectivity:

• Uninformative priors: Express minimal prior knowledge, letting data dominate the analysis. Uniform distributions or Jeffreys priors are common choices.
• Weakly informative priors: Provide soft constraints based on physical reasonableness without strongly influencing conclusions.
• Informative priors: Incorporate specific prior knowledge, such as distributions derived from historical data or physics-based models.

In EMC, physics provides natural constraints: emissions cannot be negative, shielding effectiveness cannot exceed theoretical limits, and coupling cannot occur without a physical path. Priors that encode these constraints improve inference efficiency without introducing bias.

Bayesian Inference in EMC

Bayesian methods offer several advantages for EMC applications:

• Sequential updating: As each test unit is measured, the posterior distribution updates, providing a continuously refined estimate. Testing can stop when the posterior provides sufficient certainty.
• Incorporating prior information: Simulation results, supplier data, or historical measurements can formally contribute to the analysis.
• Direct probability statements: A Bayesian credible interval carries the interpretation most engineers already assume, namely that the parameter lies within it with the stated probability. A frequentist confidence interval does not: its coverage is a property of the procedure over repeated sampling, not a probability attached to the one interval in hand.
• Decision integration: Bayesian decision theory provides a coherent framework for making optimal decisions under uncertainty.

For compliance assessment, Bayesian methods can calculate the probability that a product meets its specification, incorporating all available evidence. That is the quantity a release decision actually needs. A p-value cannot supply it, because a p-value is computed under an assumed hypothesis and says nothing on its own about how probable that hypothesis is. The cost of the Bayesian answer is that it depends on the prior, so the prior must be stated, justified, and tested for influence before the result is used in a regulatory argument.

Computational Methods

Modern Bayesian analysis relies on computational methods to handle complex models:

• Markov Chain Monte Carlo (MCMC): Generates samples from the posterior distribution through a random walk that converges to the target distribution. Allows Bayesian inference for arbitrarily complex models.
• Hamiltonian Monte Carlo: An advanced MCMC method that uses gradient information for more efficient sampling in high dimensions.
• Variational inference: Approximates the posterior with a simpler distribution, trading accuracy for speed in large-scale problems.

Software packages such as Stan, PyMC, and JAGS make Bayesian analysis accessible for practical EMC applications. The choice of method depends on model complexity, data volume, and computational resources.

The CISPR 80/80 Rule

CISPR does not leave the statistics of series production to the practitioner. CISPR 16-4-3, Statistical considerations in the determination of EMC compliance of mass-produced products, sets the criterion, and product standards such as CISPR 11 restate it in their own clauses on the statistical assessment of series-produced equipment.

The criterion is the 80/80 rule, and its two figures do not mean what the common paraphrase suggests. It does not say that 80% of units must pass. It says that compliance shall be demonstrated such that at least 80% of the mass-produced items comply with the limit, with at least 80% confidence. The first 80 is the covered proportion of the population; the second is the confidence attached to that statement given a finite sample. It is, in the vocabulary used earlier in this article, an 80/80 one-sided tolerance limit rather than a pass rate.

The principal test in CISPR 16-4-3 is based on the non-central t distribution and takes the familiar tolerance-limit form. Compliance is judged from:

x̄_n + k S_n ≤ L

where x̄_n is the arithmetic mean of the levels measured on the n items, S_n is the sample standard deviation computed with the divisor n − 1, L is the permissible limit, and all four quantities are expressed logarithmically in dB(µV), dB(µV/m), or dB(pW). The factor k is taken from tables of the non-central t distribution for 80% confidence that 80% of the type falls below the limit, and it depends only on the sample size:

The test should be performed on a sample of not fewer than five items; three are permitted only when five are not available. Two further points are worth carrying away. First, because the highest emission of one unit rarely falls at the same frequency as the highest emission of the next, the standard divides the spectrum into frequency sub-ranges and applies the rule to the highest level within each sub-range, rather than to a single peak per unit. Second, the 80/80 rule protects the consumer; it says little about the probability that a conforming batch will actually be accepted, which is the manufacturer's risk and is treated separately.

These k values are markedly smaller than the 95/95 factors quoted earlier, which is simply what the relaxation from 95/95 to 80/80 costs in protection. Comparing the two is a useful calibration: a criterion is only as strong as the confidence and coverage attached to it, and neither number is visible in a bare pass or fail.

Sample Size Determination

Determining appropriate sample sizes is crucial for both testing efficiency and result reliability. Sample size depends on:

• Desired confidence level and precision
• Expected variability in the measured quantity
• Effect size to be detected (for hypothesis tests)
• Cost and time constraints

Power analysis calculates the sample size needed to detect a specified effect with specified power. For comparing two groups with expected difference d and common standard deviation s, a widely used approximation is:

n ≈ 16(s/d)² per group

The constant 16 is not arbitrary. The exact normal-theory expression is 2(z_α/2 + z_β)²(s/d)², and for a two-sided test at α = 0.05 with 80% power the bracket is (1.96 + 0.84)² = 7.85, giving 15.7, which rounds to 16. Raising the power to 90% replaces 16 with about 21. The approximation assumes a two-sided test, two groups of equal size, equal variances, and a standard deviation known in advance; because it substitutes normal quantiles for t quantiles it understates the requirement slightly, by one or two units per group at the sample sizes typical of EMC work. Treat it as a planning figure, then confirm it with an exact power calculation.

For proportion-based requirements, which is what EMC compliance usually reduces to, larger samples are required. If n units are tested and every one passes, the confidence that at least a proportion p of the population passes follows from the binomial distribution alone, with no assumption about the shape of the underlying distribution:

n ≥ ln(1 − C) / ln(p)

where C is the required confidence and p the required coverage. For 95% confidence that at least 95% of the population complies, this gives ln(0.05)/ln(0.95) = 58.4, so 59 units must be tested with no failures. The equivalent 90/90 statement needs 22 units. The contrast with the tolerance-limit approach is instructive: a normal tolerance limit reaches the same 95/95 statement from ten units, but only because it buys the difference with an assumption of normality that the distribution-free calculation does not make.

Software Tools

Statistical analysis for EMC applications can be performed using various software platforms:

• General-purpose statistics packages: R, Python (SciPy, StatsModels), MATLAB Statistics Toolbox, Minitab, JMP
• Monte Carlo simulation: Crystal Ball, @Risk, Python (NumPy, SciPy), MATLAB
• Bayesian analysis: Stan, PyMC, JAGS
• Design of experiments: JMP, Minitab, Design-Expert

Many EMC test equipment manufacturers also provide built-in statistical analysis capabilities. However, understanding the underlying statistical methods remains essential for proper application and interpretation.

Documentation and Reporting

Statistical EMC analyses should be thoroughly documented, including:

• Clear statement of the question addressed and hypotheses tested
• Description of the data, including sample size, measurement conditions, and any data transformations
• Justification for statistical methods chosen
• Verification that method assumptions are satisfied
• Complete results including point estimates, confidence intervals, and p-values where applicable
• Practical interpretation in EMC terms, not just statistical significance

Three documents divide the ground between them. ISO/IEC Guide 98-3, the Guide to the Expression of Uncertainty in Measurement (GUM), sets the general framework for evaluating and stating measurement uncertainty. CISPR 16-4-2 applies that framework to the uncertainty of EMC measurement instrumentation. CISPR 16-4-3 covers the separate question of statistical compliance for mass-produced products. Measurement uncertainty and production spread are different quantities with different treatments, and conflating them is a recurring source of confusion in test reports; citing the right document for each keeps them apart.