A Student t random-effect model assigns Student's t-distributions to exchangeable study effects, allowing heavier tails than a normal distribution hierarchy. An equivalent Gaussian scale mixture is , , with independent latent draws. For the variance is , so is a scale rather than a standard deviation. Small latent precisions weaken shrinkage for atypical studies while retaining partial pooling for the rest.
Sub-Gaussian distribution Created 2026-10-13 Updated 2026-11-05
A real centered random variable is sub-Gaussian random variable with parameter if its moment-generating function satisfies
The parameter is an upper bound on fluctuation size, not necessarily the standard deviation. For and , exponential Markov inequality gives . Choosing and applying the same argument to yields the Chernoff bound
A centered normal distribution of variance attains equality in the moment-generating-function condition. A Rademacher distribution also satisfies it with , since . Products of moment-generating functions show that a sum of independent centered sub-Gaussian random variables with parameters is sub-Gaussian with parameter . Hence the sample mean of independent variables with a common parameter has tail bound . This gives normal-like concentration without assuming a normal distribution. For an uncentered variable, the convention is to apply the condition to .
Symmetric probability distribution Created 2026-09-24 Updated 2026-11-05
A real probability distribution is symmetric about when and have the same distribution. A centered normal distribution is symmetric about zero.
A broad component plus a narrow central component can have one local maximum while still displaying two distinct width scales. This is what “bimodal” means in this imaging context; it does not require two separated peaks. To isolate the population effect, construct an original normalized line profile
Both terms are normal distributions; their integrals are one, so the component areas are and . With , the narrow component has central height , while the broad component contributes . A tall central feature therefore need not contain most of the particles. Holding these widths fixed deliberately isolates changing weights; it does not model how temperature, interactions or camera resolution determine actual widths.
Figure 1.
Ideal harmonic-trap condensate fraction and synthetic profiles
. Original calculated condensate fraction at left; normalized two-width normal distribution mixtures at right, with weights from the cubic law. Chosen widths and parameters, no measured data. A profile is a line marginal, not an unintegrated camera pixel cut. Details.
For absorption imaging in the weak-probe regime with negligible saturation, the optical depth is , where is the reference light intensity without atoms, the transmitted light intensity, an effective absorption cross section, and the column density. Integrating over both image coordinates gives the particle number; a central pixel height alone does not. Finite resolution can obscure a narrow component, and saturation invalidates this simple logarithmic calibration.
The tail assumptions needed for conservative Wald coverage concern the distribution of the standardized estimation error and the realized scale of its interval. An expected-value bound on a conservative variance estimator does not establish either property. The discrete and rare-large-variance counterexamples in conservative variance estimation does not guarantee Wald interval coverage isolate those failures. NIST similarly makes a normal-model qualification before interpreting an expanded uncertainty as approximately 95 percent coverage.[ref]
A precise sufficient asymptotic argument is the following. Let , take , and suppose
The first condition includes centring at the true parameter; asymptotically material bias would change it. The second condition is convergence in probability of a variance ratio, substantially more than an expectation inequality. By Slutsky's theorem, the Wald confidence interval with normal critical value then has limiting coverage probability
The limiting normal distribution has no mass on the interval boundary, which justifies passing to that probability. This is asymptotic conservative Wald coverage under variance-ratio convergence, not a finite-sample guarantee.
For an original exactly normal comparison, let a sample mean have true variance and use the deterministic scale . A nominal 95 percent Wald confidence interval uses a radius times the usual one and has exact coverage probability . The calculation uses an actual normal sampling law and a fixed ratio; it cannot be inferred from alone. The normal cdf and inverse-cdf implementations provide the numerical evaluation.[ref]
A samplewise variance upper bound supplies a separate finite-sample route when the centre is exactly normal, because each reported interval contains the known-variance normal interval. Without that shape assumption, use a justified tail inequality as in confidence bounds from a samplewise variance bound.
Thermal-product dependence beyond covariance is the influence of higher joint moments on uncertainty in a nonlinear thermal expansion correction. Two input models can have identical means, variances and covariance yet predict different uncertainty in a product. A thermal-product input-distribution check must therefore examine how the temperature offset and expansion-coefficient error occur together.
Let and , where the centered variables have expected values zero, variances and covariance . Direct expansion gives
The third joint moments and the mixed fourth moment of thermal inputs are extra information. For centered inputs, only the last two terms remain. A first-order uncertainty propagation rule at zero input means would miss the product entirely; the GUM explicitly requires higher-order treatment when nonlinearity matters.[ref]
An original counterexample keeps even the marginal shapes fixed. Let have a standard normal distribution and let independently take values and with equal probabilities. Define and . Both have centered normal distributions, the quoted variances and zero covariance. They are not independent: their absolute standardized values are identical.
Their product is . Its expected value is zero, while its variance is , because . Independent centered normal inputs with the same individual variances instead give product variance . The joint model, rather than a marginal histogram or correlation coefficient alone, determines the difference.
For an illustrative length correction with zero means, , and , these models give standard deviations nm and 10 nm. This constructed dependence is a mathematical test case, not a proposed universal thermometer model.
In a physical gauge block study, track paired temperature and coefficient information, shared thermometer corrections and any procedure that estimates one input using the other. NIST's thermal correction explicitly combines block length, expansion coefficient and temperature difference.[ref] Preserve that dependence when generating joint draws. JCGM 101 distinguishes model formulation from propagation and gives a multivariate Gaussian input model as a specific choice, not a consequence of normal marginals alone.[ref] More simulation trials reduce numerical error conditional on the chosen model; they cannot identify unmeasured joint moments.
Thomas J. Rothenberg 2026-11-05
Thomas J. Rothenberg is a mathematical econometrician whose work concerns statistical inference and the identifiability of models. The American Economic Association's 2015 citation recognizes his contributions to identification and inference and identifies him as a professor emeritus at Berkeley.[ref]
His primary paper Identification in Parametric Models by Thomas J. Rothenberg, published in 1971, develops identification criteria beyond one model considered in isolation. It connects local identification with the information matrix under regularity conditions and treats identification of individual parameters separately from that of the full vector.[ref] The distinction is useful when a scientifically relevant combination is recoverable even though the entire parameterization is not; identifiability from linear observations gives an explicit example.
An original counterexample explains why the regularity qualifications matter. Consider one observation with normal distribution
The mean map is one-to-one, so distinct parameters have distinct observable distributions. Nevertheless its Fisher information is
which vanishes at zero. The information rank changes there, so a nonsingularity criterion cannot be applied without its assumptions. Nonregular identification at a flat mean map separates exact distinguishability from first-order sensitivity: a zero derivative does not by itself establish that two different parameter values produce the same law.
The example illustrates a broader contribution of identification theory. Before asking an estimator for a precise answer, determine whether the observations can distinguish the proposed quantities and whether the local approximation used for inference is valid.
Figure 1.
Thomas J. Rothenberg in the American Economic Association award portrait
. Externally hosted image identified by the society's named Distinguished Fellow citation.
Suppose conditionally independent latent values have normal distribution , with , but the observation only says which side of a known threshold contains each value. With for a value above the threshold and for a value below, the likelihood is the displayed product, where is the standard normal distribution function. Repeated thresholds give corresponding powers. These factors are tail probabilities, not probability densities evaluated at the thresholds. Multiplying normal densities at would describe a different observation process. The thresholds must be treated as fixed or otherwise modelled appropriately; informative selection of thresholds can change the likelihood.
Time series Created 2026-09-28 Updated 2026-10-13
A time series is a sequence of observations indexed by time: hourly temperatures, daily sales, or successive measurements of a fluctuating signal. A probabilistic model treats these observations as one sample path of a stochastic process . For equally spaced observations, take and measure a lag of a time series in observation intervals. The order matters: shuffling a record preserves its observed values but destroys information about how one value follows another.
Two basic ways to model temporal dependence are an autoregressive model, which feeds past observations into the next value, and a moving-average model, which combines present and past white noise. An AR(1) carries a disturbance forward with diminishing weights; an MA(1) carries it into just two observations. Their autocorrelation functions make this difference visible. A weakly stationary process supplies a common mean and a lag-dependent autocovariance with which to describe that dependence.
These models help forecast future observations and quantify the uncertainty of averages. Successive measurements need not provide as much information as the same number of independent random variables: the long-run variance of a stationary process accounts for their accumulated covariances. Trend and seasonality require attention before fitting a stationary model; differencing and a seasonal difference operator can help, but do not automatically make every process stationary. See Peter Bartlett, Introduction to Time Series Analysis, lecture 1: objectives and models.
Figure 1.
Stationary Gaussian examples: positive AR(1) persistence, negative AR(1) alternation, and MA(1) short memory
. Details.
The three illustrated sample paths have zero mean and unit variance, so the visual comparison concerns dependence rather than scale. The AR(1) paths start from their stationary normal distribution, independently of future discrete Gaussian white noise; the MA(1) path includes the noise preceding its first observation. These are finite random examples, not exact pictures of an ACF. Compare their theoretical autocorrelation functions in autocorrelation.
Historically, Yule's 1927 study investigated disturbed periodicities in sunspot observations; it should not be read as the discovery of the particular AR(1) examples here. Slutsky studied how summing random inputs could produce wave-like records, a warning against inferring a periodic cause from appearance alone. The editorial note to his 1937 English paper identifies the 1927 Russian original. See G. Udny Yule, On a method of investigating periodicities in disturbed series, with special reference to Wolfer's sunspot numbers (1927) and Eugen Slutsky, The Summation of Random Causes as the Source of Cyclic Processes (1937), including the editorial note on the 1927 original.
Transience of Brownian motion in dimension at least three Created 2026-09-24 Updated 2026-11-05
Standard Brownian motion in is transient for : almost surely. In dimension three, start away from zero and use the positive local martingale . It is a supermartingale, hence converges to a finite limit. Brownian scaling and the normal distribution of show that in probability, so its almost-sure limit must be zero. Thus .
For a start at zero, apply the same argument from any positive deterministic time; the position there is nonzero almost surely. In higher dimensions, project onto three coordinates: the full radius dominates a radius already tending to infinity. This differs from recurrence of planar Brownian motion, where every neighbourhood is visited repeatedly.
Truncated lognormal moment Created 2026-10-06 Updated 2026-11-05
If with , and having the standard normal distribution, then for every real ,
Here is the standard normal distribution function. Multiplying the normal distribution density by and completing the square shifts its mean by and supplies the factor ; this is a Gaussian exponential tilt. Integrating above or below the shifted threshold proves both formulas. They apply to noninteger and negative powers, including the inverse-price moment . The lower-tail formula gives the European power put formula after a binomial theorem expansion.
Truncated normal distribution Created 2026-09-28 Updated 2026-11-05
A truncated normal distribution is a normal distribution conditioned on belonging to a retained measurable set of positive probability. If , , and , its density is
where is the standard normal density. The indicator is essential: the density is zero outside . For an interval , possibly with infinite endpoints,
The same expression holds for any endpoint inclusion convention because the original normal distribution has no atoms.
For a standard normal variable restricted to positive values, and the density is for . Its mean is : integrate using . Symmetry gives second moment one, hence variance . Truncation therefore changes the mean and variance; they are not generally the original parameters .
sign-truncated normal sampling uses inverse transform sampling to generate this conditional law directly. Rejection from the original normal is also exact but may be slow when the retained set has very small probability.
Stipulate an original Gaussian mixture distribution with mixture weights , , component means , , and component variances both equal to one. These are chosen statistical parameters, with no observed sample or fitting. If is the standard normal distribution probability density function, then
In the figure, the two colored curves are weighted probability density functions, with areas and . Adding their heights at every gives the black curve. Neither colored curve stops at a dividing line.
Figure 1.
Stipulated two-Gaussian mixture: weighted component densities and their sum
. Original analytic curves for component means and , component variances one, and mixture weights and . No observed points, histogram, or fitted parameters are used. Details.
The law of total expectation and law of total variance give the total mean and variance:
All of the extra variance comes from variation between component means. Equivalently, the second moment is , so the variance is .
A mixture distribution is different from a deterministic linear combination of random variables. If independent and are drawn and , then has a normal distribution with the same mean but variance , rather than . Every draw of uses both random variables. A draw from the mixture distribution instead uses , where the mixture allocation indicators are independent of and select exactly one of them. The expression for adds densities, not simultaneous random outcomes.
Underdamped Langevin dynamics Created 2026-09-28 Updated 2026-11-05
Underdamped Langevin dynamics retains both position and momentum . For mass , temperature , friction rate and smooth potential energy , one common normalization is
Here is vector Brownian motion. The noise amplitude is tied to friction by the fluctuation-dissipation relation for a Langevin particle. If the density is normalizable and the dynamics is nonexplosive, its invariant distribution is the Boltzmann distribution
The Hamiltonian flow preserves the total energy and phase-space volume; the momentum Ornstein-Uhlenbeck process preserves a centered normal distribution with coordinate variance . These two facts verify the invariant density. In dimensionless units with and momentum variance , it is , with the corresponding choice .
Retaining inertia resolves the short-time ballistic-to-diffusive Brownian displacement. The overdamped Langevin dynamics approximation removes momentum relaxation when only much longer times are relevant.
Uniform-alternative binomial Bayes factor Created 2026-10-14 Updated 2026-10-25
For a binomial distribution count, compare a point null with the uniform alternative prior distribution on . The alternative Bayesian model evidence is
so the exact Bayes factor favouring the point null is . With standardized deviation , the local normal distribution approximation is . At fixed the frequentist p-value remains approximately constant, while the factor grows like . This demonstrates the prior-width mechanism behind the Jeffreys-Lindley paradox for a Gaussian point null in a discrete sampling model.
For a separate original example, change the point null to and take to have a uniform prior , fixed before the ten Bernoulli trials. Here is drawn once under and shared by all ten trials: they are independent conditional on . For one particular ordered sequence with eight heads and two tails, the Euler beta function gives
Thus the Bayes factor for the uniform alternative against the point null is
The binomial distribution count formulation gives and . Their ratio is again ; these count masses must not be combined with the ordered-sequence masses above. The general formula at the start of this article puts the point null in the numerator, so it equals here.
In contrast, maximum likelihood estimation gives : differentiating the log-likelihood yields , and the endpoints have zero likelihood. Its peak-to-null likelihood ratio is
This peak is chosen after seeing the data, so it is not the Bayes factor of the prespecified . The Bayesian model evidence for averages over its entire uniform prior; it is smaller than the peak. It is also smaller than . Against the same null, a point alternative fixed at would give a Bayes factor about , greater than . An integrated alternative need not outperform a point alternative.
Figure 1.
A uniform-prior average of the likelihood differs from its maximum
. Original calculation for a stipulated ordered sequence with eight heads and two tails. Left: the likelihood function and its uniform prior average , with the point null at and peak at . Right: the average, fixed value, and peak divided by the same null value . Only the first two comparisons use prespecified alternatives; the peak uses a maximum-likelihood estimate. These are calculated model quantities, not measured frequencies. Details.
For a general proper beta distribution prior distribution with , the same ordered-sequence marginal likelihood is . This follows by substituting the normalized prior density into the integral, and shows exactly where the parameter prior distribution changes the Bayes factor. Holding the parameter prior distributions fixed while changing model prior odds leaves that Bayes factor unchanged.
For a normal distribution with unrestricted and unknown , an integrable data-only unbiased estimator of the mean-vector prediction risk of a fixed rank- orthogonal projection matrix exists exactly when . To prove necessity, randomize the mean by independent Gaussian variance : conditioning adds to the squared bias term, whereas the marginal variance identity would add . When , the residual sum of squares itself is unbiased.
Variance normalization Created 2026-09-24 Updated 2026-11-05
Variance normalization rescales a centered sum to have variance one. Let and assume . Then
For independent variables with finite variances, . With dependence it also contains . For example, if every equals the same random variable of variance , the sum's variance is , not .
This is the natural scale for central limit theorems, but variance one alone does not imply a Normal distribution or convergence to one. Appropriate assumptions, such as a Lindeberg condition for a suitable independent triangular array, control the shape of the normalized sum. Centering, choosing the correct variance and proving a distributional limit are distinct steps.
Consider the explicitly specified objective
and estimate its derivative by a symmetric finite difference with step . The choice of shared or independent noise changes the estimator's variance even though both use correct normal distributions at the two parameter values.
With common random numbers, the same appears on both sides:
Thus and , independent of . Cancellation removes the noisy square term. For this quadratic, the symmetric difference has no truncation bias in exact arithmetic.
For independent standard normals , instead use
If , expand . The standard-normal moments , and give
Adding the two independent square variances and dividing by yields
Both estimators have the correct expected value , but the independent-noise variance diverges as tends to zero. At and , the variances are and . An average of independent replicate estimators divides each variance by .
This is a complete calculation for the stated quadratic, not a universal dominance theorem. At , and , the independent variance is , below the shared value . The covariance principle in common random numbers explains why coupling needs a specified objective and comparison. Very small steps in floating-point evaluation also introduce cancellation concerns that this exact-arithmetic model does not include.
Variance-stabilizing transformation Created 2026-10-14 Updated 2026-10-25
If a response has mean and variance , the delta method gives . Choosing proportional to makes this first-order variance independent of . For on positive means, the resulting transformation is the natural logarithm. This is a local approximation and requires the transformation to be defined on the observations; it does not make arbitrary normal responses positive or preserve an exact normal distribution.
The variational approximation gap is the part of the evidence lower bound deficit that remains even after optimizing separately for an observation within a specified variational family. With the generative model fixed and finite , define
The equality uses the usual evidence lower bound identity for variational inference. A supremum allows for a best value that is not attained. The amortization gap separately measures failure of the shared inference rule to reach this value.[ref]
An original example shows what independence can cost. Stipulate a centred two-dimensional multivariate normal distribution as the posterior distribution, with covariance matrix
Restrict the variational family to centred independent normal distributions, with . This is a simple mean-field variational inference restriction. Writing , the Kullback-Leibler divergence between normal distributions becomes
Each variance contributes a term , minimized at . Consequently
For , the best independent variances are both , and the unavoidable deficit is about nat. Matching both marginal variances to one instead would give about nat. The reverse Kullback-Leibler divergence therefore chooses narrower independent marginals here, while still failing to represent the correlation.
Enlarging the family to include the full covariance matrix permits and reduces this example's gap to zero. More generally, inclusion cannot increase the optimum approximation gap at a fixed model. Merely training the same shared rule longer cannot remove a strictly positive family restriction. Numerical improvements are not proofs that the infimum has been reached.