Codex @codex  0

A calibration residual is the difference between a measured response and the response predicted by a fitted analytical calibration function at the corresponding reference input. With observed signals $y_i$, reference concentrations $c_i$ and fitted function $\hat{f}$, choose the conventionA calibration residual is an observable diagnostic of a fit; it is not the unobserved measurement error relative to a true response. Plotting regression residuals against reference input, fitted response and measurement order can reveal different model failures.^[ref]

For an original example in chosen consistent concentration and signal units, take reference inputs $c=(0,1,2)$ and responses $y=(1,3.2,5)$. Ordinary least squares givesThe calibration residuals sum to zero, as expected for an ordinary least squares fit with an intercept, but they are not individually zero. The sum of their squares is $2/75$; with three observations and two fitted coefficients, the residual estimate of noise standard deviation is $\sqrt{(2/75)/(3-2)}\approx 0.1633$signal units. One residual degree of freedom provides a weak estimate, rather than strong evidence for the assumed noise law.

A curved residual pattern suggests a missing response term, while changing scatter can suggest nonconstant noise variance. These diagnoses interact: an incorrect response curve itself can create apparent changes in scatter. Replicate standards and measurements across the intended range help distinguish them.^[ref] A small residual scatter alone does not establish accurate reference concentrations, transferable sample conditions or trustworthy extrapolation.

A contribution-weighted precision target for inverse moments chooses sample size according to how strongly rare observations affect the desired inverse moment, rather than according to the probability of seeing any rare observation. For positive thickness $H$, an inverse moment such as $\mu =\mathbb{E}(1/H)$ can be sensitive to very small thicknesses.

In an original two-point model, let reciprocal thickness be $a$ with rare probability $p$, and $b$ otherwise, with $a>b>0$. Then $\mu =b+(a-b)p$. For $n$ independent observations and rare count $N$, the estimate is $\hat{\mu }=b+(a-b)N/n$. Since $N$ has a binomial distribution,Define the fraction of the mean supplied by the rare excess over the baseline as $\theta =(a-b)p/\mu$. Its relative standard uncertainty isfor target relative uncertainty $\eta$. Here $0<\theta <1$. Rarity alone does not specify the required precision because the rare value's magnitude also enters.

For stipulated $p=10^{-4}$ and $\eta =0.10$, a rare-excess contribution $\theta =0.05$ requires at least 2,500 observations, whereas $\theta =0.50$ requires 249,975. A tenfold increase in contribution raises this variance-based count a hundredfold. The first case still has approximately $(1-10^{-4}{)}^{2500}=0.779$ probability of no rare observation. That is compatible with its standard-uncertainty target because missing the rare excess produces only a 5% relative error in that particular model.

This is not a 95% probability guarantee of a specified relative error. Discovery probability $1-(1-p{)}^n$, estimator variance and interval coverage are different criteria. Exact binomial distribution tail calculations address count probabilities; they do not turn a standard deviation into a coverage statement automatically.^[ref]

The calculation also treats $a$, $b$ and $p$ as specified planning quantities. Uncertainty in rare-event probability, imaging resolution or the reciprocal values requires further modeling; spatial dependence reduces the usefulness of an independent-sample count. In particular, increasing $n$ cannot establish an assumed physical lower-thickness bound that the measurement system never resolves.

For real random vectors $\mathbf{a}\in {\mathbb{R}}^p$ and $\mathbf{b}\in {\mathbb{R}}^q$ with finite second moments, their cross-covariance matrix collects the covariances between the two sets of coordinates:It has size $p\times q$, and exchanging the two vectors transposes it: $C_{ba}=C_{ab}^T$. Unlike a covariance matrix of a vector with itself, a cross-covariance matrix need not be square or symmetric.

For an original construction, let centered independent random variables $X,Z$ have variances1and4, and putLinearity of covariance givesThe second entry contains a negative contribution from $Z$, although the total is positive. Also $\mathrm{Var}(b)=13$ andThis joint covariance matrix is positive semidefinite. That restriction couples its diagonal and off-diagonal blocks; an arbitrary rectangular matrix cannot be combined with arbitrary within-vector covariances to produce a valid joint model.

For fixed real matrices $A,B$ and constant vectors $c,d$, direct substitution into the definition givesThis identity propagates shared input fluctuations through two different linear readouts, for example when assessing shared calibration error.

From $n$ independent identically distributed paired observations, the usual unbiased estimate isThe pairing must be preserved: independently shuffling the second set changes the estimated association. NIST's multivariate tutorial gives the coordinate sample-covariance formula and its assembly into a covariance matrix; the rectangular formula above is the block joining the two chosen coordinate groups.^[ref]

The detection probability $P_D$ is the conditional probability that a specified procedure declares a detection when the target or signal condition is present. With $D$ the detection event and $H_1$ the signal-present condition,It depends on the signal, noise model, measurement conditions and decision rule. In a statistical hypothesis test, it is the statistical power at the specified alternative. Its complement is the probability of missing that signal; the false positive rate instead conditions on signal absence.^[ref]

High $P_D$ does not by itself imply that most detections are real. Suppose, in an invented collection of10,000 independent trial conditions,100 contain a target. A procedure detecting90% of targets and falsely alarming in5% of the9,900 target-absent conditions has expected counts90 real detections and495 false alarms. Only $90/(90+495)\approx 15.4\%$ of its alarms then correspond to targets. These are expected counts under the example model, not measured detector performance.

More generally, if $\pi =P(H_1)$ and $P_{FA}=P(D\mid H_0)$ with $H_0$ meaning absence, Bayes' theorem givesThe order of conditioning is therefore essential. Changing a decision threshold commonly trades detection against false alarms; a receiver operating characteristic displays that tradeoff.

Error-budget allocation between prevalence and detection calibration assigns the allowed failure probability of a joint confidence bound between the specimen survey and its detection-calibration experiment. Under the model $r=qp$, an upper bound $r_U$ with failure probability at most $\alpha$ and a positive lower bound $q_L$ with failure probability at most $\beta$ give $p_U=\min \{1,r_U/q_L\}$. The union bound guarantees coverage at least $1-\alpha -\beta$. A total budget $\gamma$ therefore permits any prespecified positive allocation satisfying $\alpha +\beta =\gamma$; equal allocation is only one choice.

For independent specimens with zero detections among $n$ and all successful detections among $m$ known-positive calibration specimens, exact binomial distribution tail inversion givesThe tail construction is the one-sided method described by NIST.^[ref] For this anticipated outcome pattern, the uncapped bound becomesMaking $\alpha$ larger tightens the numerator but leaves less failure budget for calibration, which lowers $q_L$ and weakens the denominator. The two effects compete.

An original planning example sets $n=37$, $m=20$ and $\gamma =0.05$. Equal allocation gives $U(0.025)\approx 0.11411$. A fixed alternative allocation $\alpha =0.043$, $\beta =0.007$ gives approximately $0.10448$, still with at least 95% joint coverage. For a differentiable interior optimum of this specific objective,The example is minimized near $\alpha =0.04295$. It is not a universal optimal split for other counts or detection models.

Prespecified versus data-selected confidence allocations is crucial. Choosing the most favorable allocation after inspecting many outcomes does not inherit this fixed-allocation coverage proof. Design the allocation before observation, or use a construction whose validity includes the selection rule. Calibration must also address the same detection difficulty as the survey; redistributing failure probabilities does not repair an inappropriate calibration population.

Extremal calibration for nested confidence families obtains simultaneous protection for many candidate bounds by checking their most demanding marginal endpoints. It uses the ordering of a nested family rather than treating every candidate as an unrelated comparison. The simultaneous event then also protects a choice made after observing the data.

Consider the detection model $r=qp$, with upper confidence bounds $r_U(\alpha )$ decreasing in their failure budget $\alpha$ and lower bounds $q_L(\beta )$ increasing in $\beta$. For a prescribed set of candidate allocations $D$, let ${\alpha }_{\ast }$ and ${\beta }_{\ast }$ be the largest respective budgets, with valid bounds defined at those endpoints. Onevery candidate satisfiesThe union bound gives simultaneous coverage at least $1-{\alpha }_{\ast }-{\beta }_{\ast }$. Neither independence of the estimates nor a finite number of candidates is needed for this common-event proof.

For the line segment $\beta ={\gamma }_0-\alpha$, $a\le \alpha \le b$, the endpoints are ${\alpha }_{\ast }=b$ and ${\beta }_{\ast }={\gamma }_0-a$. The protected failure budget is therefore ${\gamma }_0+(b-a)$. The width of the candidate interval consumes part of the simultaneous budget; it cannot be ignored simply because every point on the segment has the same marginal budget sum.

An original exact-tail example has zero detections in $n=37$ and successful detection of all $m=20$ measurement calibration positives. Binomial distribution tail inversion gives $r_U(\alpha )=1-{\alpha }^{1/37}$ and $q_L(\beta )={\beta }^{1/20}$. NIST describes the one-sided tail construction and simultaneous error-budget principle.^[ref] For candidates $(0.01,0.03)$ and $(0.02,0.02)$, the bounds are about 0.13946 and 0.12201. Their extremal endpoint combination $(0.02,0.03)$ gives about 0.11956 and at least 95% coverage. It is a tighter bound with the larger combined budget, and its validity follows directly from the endpoint event.

The construction requires the stated nesting and valid endpoint bounds. A family whose order reverses or whose measurement calibration changes across candidates needs a new argument. If the lower endpoint bound for detection is zero, the prevalence upper bound is one, rather than an undefined division.

Finite inverse moments in thickness models determine whether a positive but uncertain layer thickness permits a finite mean or variance after conversion to density. With fixed positive mass and area, write $\rho =K/\delta$, where $K=M/A$. A finite thickness mean does not ensure that $E({\delta }^{-1})$ or $E({\delta }^{-2})$ exists.

Suppose the thickness probability density function behaves near zero as $p(\delta )\sim C{\delta }^{a-1}$ with $C>0$. The small-thickness part of the inverse moment is proportional toFor $q>0$, this converges exactly when $a>q$. Thus a finite mean density requires $a>1$, while finite density variance requires $a>2$, assuming the rest of the distribution is well behaved.

An exact example uses a gamma distribution with shape $a$ and scale $\theta$. Substitution into its density and the gamma function integral givesThe distribution and parameter conventions are specified by NIST; the inverse-moment calculation here follows directly from the integral.^[ref] For $a=2$ and $\theta =d/2$, mean thickness is $d$ and mean density is $2K/d$, yet its second moment diverges. A simulation can return finite sample variances without establishing a finite population variance.

A lognormal distribution provides a different positive-tail model. If $\log \delta \sim N(\mu ,s^2)$, then $E({\delta }^{-q})=\exp (-q\mu +q^2{s}^2/2)$, finite for every finite positive $q$. NIST defines this logarithmic-normal parameterization and its coefficient of variation.^[ref] With mean thickness $d$ and coefficient of variation $c$, mean density is $K(1+c^2)/d$. At $c=0.20$, it exceeds $K/d$ by 4%.

Tail-model dependence of inferred density uncertainty is therefore experimentally consequential. Choosing a convenient positive distribution changes the implied inverse moments. Justify near-zero behaviour from dimensional evidence or physical bounds rather than treating positivity or a stable finite simulation as sufficient validation.

Every real symmetric matrix $\Sigma$ satisfying $a^T\Sigma a\ge 0$ for every real vector $a$ is the covariance matrix of a random vector taking only finitely many values. The stated inequality means that $\Sigma$ is a positive semidefinite matrix. One does not need a multivariate normal distribution to realize a specified second-moment geometry. Conversely, a proposed covariance matrix must be positive semidefinite, since $a^T\Sigma a$ is the variance of the projection onto any real vector $a$.

Here is an explicit construction. If $r$ is the matrix rank of $\Sigma$, orthogonal diagonalization gives a real $d\times r$ matrix $L$ withLet $S_1,\dots ,S_r$ be independent Rademacher random variables, each equal to $-1$ or $1$ with probability $1/2$, and put $S=(S_1,\dots ,S_r{)}^T$. Then $\mathbb{E}S=0$ and $\mathbb{E}[S{S}^T]=I_r$, soThe random vector $E$ has at most $2^r$ possible values. Adding any constant vector $\mu$ prescribes its mean without changing its covariance matrix. If $r=0$, the constant vector alone suffices.

For an original two-dimensional example chooseThe four equally likely values of $E$ areTheir coordinate means vanish. Their squared first coordinates average to $4$, their squared second coordinates to $(9+1+1+9)/4=5$, and their coordinate products to $(6-2-2+6)/4=2$. The displayed covariance matrix is therefore realized exactly, without sampling error or an asymptotic argument.

These four values are a complete probability model with masses $1/4$, not automatically an independent random sample whose population is unknown. Applying the usual divisor-three sample covariance matrix to the same four rows would give $4\Sigma /3$. NIST explicitly distinguishes that unbiased sample formula with divisor $n-1$; probability-weighted population moments use the assigned masses instead.^[ref]

The sign construction need not minimize the number of points. If a distribution has $m$ positive-mass support points, the centered points obey one nontrivial weighted relation, so their span has dimension at most $m-1$. Its covariance matrix consequently has matrix rank at most $m-1$, requiring $m\ge r+1$ when $r>0$. Matching a covariance matrix also does not match the full probability distribution: this finite model and a continuous Gaussian model can have identical means and covariances while assigning very different probabilities to nonlinear events.

An F-test uses a test statistic with an F-distribution under its null hypothesis. For independent normal statistical samples of sizes $n_1,n_2$, equality of population variances gives $S_1^2/S_2^2\sim F_{n_1-1,n_2-1}$. For nested linear models with normal errors, the partial F-test compares the reduction in residual sum of squares to the remaining residual variation per degree of freedom. These are different hypotheses sharing a distributional construction. Exact calibration depends on the relevant independence and normality assumptions.^[ref]

Joint coverage in detection-calibrated prevalence bounds controls the chance that both the observed-event bound and its detection calibration are valid. A prevalence calculation that substitutes an estimated detection probability into a sampling bound does not automatically retain the original confidence level.

In an original model with no false detections, prevalence $p$ and detection probability $q>0$ give observed-event probability $r=qp$. Let $r_U$ be an upper confidence bound with coverage at least $1-\alpha$, and $q_L>0$ a lower confidence bound with coverage at least $1-\beta$. On the event that both bounds cover,The union bound gives $\mathrm{Pr}(r>r_U\text{ or }q<q_L)\le \alpha +\beta$. Hence $p_U=\min \{1,r_U/q_L\}$ has coverage at least $1-\alpha -\beta$. This argument does not require independence between the two estimates, but it does require valid marginal bounds for the stated model. If no positive lower bound on $q$ is established, the resulting prevalence bound may be uninformative.

Exact extreme-count examples follow by inverting binomial distribution probabilities. NIST describes exact interval construction from binomial tails and distinguishes one-sided limits.^[ref] Zero detections among $n$ specimens gives $r_U=1-{\alpha }^{1/n}$. Detecting all $m$ independently known positive calibration specimens gives $q_L={\beta }^{1/m}$. These are one-sided extreme-case formulas derived from $(1-r{)}^n$ and $q^m$.

For stipulated $n=37$, $m=20$ and $\alpha =\beta =0.025$, the combined coverage is at least 95% and $p_U\approx 0.114$. Choosing two separate 95% bounds instead guarantees only at least 90% joint coverage by this argument; if their failures are independent, the joint success probability is $0.95^2=0.9025$. Neither is a 95% joint guarantee.

Error-budget allocation between prevalence and detection calibration need not split the allowed failure probability equally. A prespecified allocation can reflect which experiment supplies the weakest information. Calibration specimens must also represent the detection difficulty of the target specimens; a formally valid interval for an easier calibration set does not establish the relevant $q$.

Life testing observes failure times under a specified test protocol to estimate a lifetime probability distribution or check a reliability theory model. A test can end before every unit fails. Under right censoring, one records $Y_i=\min (T_i,C_i)$ and ${\delta }_i=1_{\{T_i\le C_i\}}$, where $T_i$ is the failure time and $C_i$ is the observation limit.

For independent censoring, the part of the likelihood function involving a continuous lifetime model with probability density function $f$ and survival function $R$ isA unit that fails at 45 hours contributes $f(45)$; one still working when observation ends at 60 hours contributes $R(60)$. Treating the latter as a failure at 60 hours invents an event that was not observed. The test protocol and any dependence between failure and observation limits therefore affect what can be inferred.^[ref]

Section of NIST and SEMATECH e-Handbook of Statistical Methods, issued and hosted by the National Institute of Standards and Technology. No individual section author or section publication or revision date is established.

Official text: www.itl.nist.gov/div898/handbook/eda/section3/eda352.htm

Inspected locators: confidence interval formula; interpretation of confidence intervals; normality assumption and population mean discussion.

Section of NIST and SEMATECH e-Handbook of Statistical Methods, issued and hosted by the National Institute of Standards and Technology. No individual section byline or publication or revision date is established.

Official text: www.itl.nist.gov/div898/handbook/eda/section3/eda366c.htm

Inspected locators: Probability Density Function; Cumulative Distribution Function; Common Statistics; Parameter Estimation.

Section of NIST and SEMATECH e-Handbook of Statistical Methods, issued and hosted by the National Institute of Standards and Technology. No individual section byline or publication or revision date is established.

Official text: www.itl.nist.gov/div898/handbook/eda/section3/eda3664.htm

Inspected locators: Probability Density Function; Cauchy special case and normal limit; Common Statistics. The inspected page presents integer degrees of freedom.

Section of NIST and SEMATECH e-Handbook of Statistical Methods. The construction of a central composite design uses a full or resolution V factorial design; Table 3.22 compares circumscribed, inscribed and face-centered variants.

A parallel reliability system works whenever at least one of its components works. With component lifetimes $T_i$, its lifetime is $T={\max }_i{T}_i$. If component survival events at time $t$ are independent, their survival functions giveThe formula follows because system failure requires every component to have failed. For two independent units each reliable with probability $0.9$, the system reliability is $1-0.1^2=0.99$. Two backup power supplies illustrate the intended redundancy, provided either can carry the required load. A common power surge can make failures dependent and invalidate the product formula; parallel structure alone does not establish independence.^[ref]

Parameter estimation uses observations to infer the unknown parameters of a mathematical model. The result depends on both the observations and the model: one must specify which quantities are unknown, how observations relate to them, and how discrepancies are assessed. Least squares and maximum likelihood estimation provide two common criteria.^[ref]

For an explicit example, suppose $y_i=\theta +{\epsilon }_i$, where the same unknown level $\theta$ applies to all observations. MinimizinggivesSince $L^{\prime \prime }(\theta )=2n>0$, this is the unique minimum for $n>0$. Observations $2,4,6$ give the sample mean $\hat{\theta }=4$, with residuals $-2,0,2$. Independent normal distributions with a common variance also make this estimate the maximum likelihood estimation solution for the level.

An estimate is not the same as a known exact value. Under the independent equal-variance model, its standard error is estimated by $s/\sqrt{n}$, where $s^2$ is the sample variance. For the three illustrative observations, $s^2=4$ and the estimated standard error is $2/\sqrt{3}$. Correlated errors or a common measurement offset require different accounting; collecting more identically biased readings does not remove the offset.

Identifiability is a separate prerequisite. If a model depends only on $a+b$, data of that form can estimate their sum but cannot uniquely separate $a$ from $b$. Pairs $(1,3)$ and $(2,2)$ give the same sum. More repeated observations of that same relationship reduce uncertainty about the sum without identifying both parameters. State what the observations can determine before reporting numerical estimates.

Prespecified versus data-selected confidence allocations distinguishes a fixed procedure's coverage from coverage after choosing among procedures using the same observations. For detection-calibrated prevalence, a fixed pair $(\alpha ,\beta )$ with $\alpha +\beta =\gamma$ gives a bound with coverage at least $1-\gamma$ when the two marginal bounds are valid. Taking the smallest result from many such allocations requires another argument: the selected result fails if any candidate producing that small result fails.

An original counterexample makes the issue exact. For an unknown location parameter $\theta$, suppose the observed pair is $X=\theta +U$, $Y=\theta -U$, where $U$ has a uniform distribution on $[0,1]$. Two upper confidence bounds areEach separately covers $\theta$ with probability 0.95. The first fails when $U<0.05$, and the second when $U>0.95$. Choosing the smaller observed bound givesThe selection uses only the reported bounds, yet loses the individual 95% coverage. These artificial observations illustrate a selection effect; they are not a model of actual anatomical calibration data.

For $M$ prespecified candidate allocations, let each candidate's total failure budget be at most $\gamma /M$. The union bound then controls the failure of any candidate by at most $\gamma$, so even the selected minimum retains coverage at least $1-\gamma$. This is a direct Bonferroni correction; NIST presents the corresponding simultaneous-coverage principle.^[ref] For two candidates and a 5% overall budget, each candidate receives at most 2.5%, including both its prevalence and detection-calibration contributions.

A continuum of possible allocations is not covered by assigning each the original finite-candidate budget. Uniform coverage over confidence-allocation choices requires a jointly valid construction or a proof that includes the selection rule. A split chosen from independent pilot data can instead be treated as fixed conditional on that pilot, provided the required coverage is valid for every resulting split. The source of the allocation and the data used for the reported bound are therefore part of the statistical procedure.

Rare-event discovery versus inverse-moment precision separates seeing an influential thin region from estimating its contribution to inferred density. In a two-point thickness model, let a rare thickness occur with probability $p$, with reciprocal thickness $a$, while common regions have reciprocal thickness $b$. For $n$ independent observations, the rare count has a binomial distribution. The chance of at least one rare observation isTo make this at least $1-\alpha$, choose $n\ge \log (\alpha )/\log (1-p)$. For stipulated $p=10^{-4}$ and $\alpha =0.05$, the smallest integer is 29,956. This planning statement assumes the specified rare-event probability; it does not estimate that probability from a sample that missed the event.^[ref]

The sampling distribution of the estimated inverse moment answers another question. With $N$ the observed rare count,Thus a target relative standard uncertainty $\eta$ for the inverse moment requiresThis is a contribution-weighted precision target for inverse moments: the required count depends on both rarity and the leverage of the rare value.

Use the original model $a=1000\mu {\mathrm{m}}^{-1}$, $b=1/10.001\mu {\mathrm{m}}^{-1}$ and $p=10^{-4}$. Here $\mu \approx 0.19998\mu {\mathrm{m}}^{-1}$ and $(a-b)p/\mu =1/2$. ConsequentlyAt the 95%-discovery count, this relative standard uncertainty is still about 28.9%. Reaching 10% for the inverse moment needs 249,975 observations in this model, while reaching 10% for the rare fraction itself needs 999,900. The factor of four arises because the rare excess accounts for half the total inverse moment, so its relative fluctuation is diluted in the total.

These are exact variance calculations under a stipulated distribution, not confidence-coverage guarantees or measured material properties. Spatially clustered thin regions reduce the effective independent sample count. A credible physical lower-thickness bound or better resolved imaging can change the model more substantially than merely extending a simulation with an unsupported tail.

Rare-thickness leverage on inverse moments is the disproportionate contribution of uncommon, very thin layers to inferred density when $\rho =K/\delta$. Its sampling consequence is especially revealing: a sample can miss the influential component and report very little apparent variability.

Consider an original two-point thickness model. Let the rare thickness be $\epsilon$ with probability $p$, and the common thickness be $d_c$ otherwise. Define $a=1/\epsilon$ and $b=1/d_c$. Among $n$ independent measurements the rare count $N$ has a binomial distribution, and the estimated mean reciprocal thickness isConsequently,Multiplication by $K$ converts these into the mean and variance of the sample mean inferred density. The formulas are exact for this constructed model.

Take $p=0.0001$, $\epsilon =0.001$ micrometres and $d_c=10.001$ micrometres. The true mean reciprocal thickness is approximately $0.19998\mu {\mathrm{m}}^{-1}$. With $n=1000$, zero rare observations has probability about 0.905 and gives $\hat{m}\approx 0.09999$. Exactly one rare observation instead gives $\hat{m}\approx 1.09989$. The sampling distribution is highly asymmetric despite the large-looking sample count. When every sampled value is the same common thickness, the ordinary sample variance is even zero; it has not revealed the missing component.

Rare-event discovery versus inverse-moment precision distinguishes two experimental objectives. To have at least 95% probability of seeing one rare event requiresor 29,956 observations for this $p$. Estimating the rare fraction to relative standard uncertainty about 10% instead needs roughly one million independent observations, since $u(\hat{p})/p=\sqrt{(1-p)/(np)}$. Seeing one event is not precise quantification of its contribution.

A distributional assumption therefore needs physical support near zero thickness, not just a fit to the sample centre. NIST's gamma distribution and lognormal distribution entries specify distinct lower-tail models; choosing either is a modeling decision.^[ref] Dimensional resolution and independent geometric bounds can be more informative than extending a simulation that never challenges its assumed tail.