The Kirsch equations describe the stress tensor around a traction-free circular hole in an infinite linear elastic plate under uniform remote tension. For remote stress ${\sigma }_{\infty }$ along $x$, hole radius $a$ and polar coordinates $(r,\theta )$ measured from that direction, put $q=a^2/r^2$. In the classical isotropic plane stress model,The fields satisfy equilibrium and the prescribed boundary tractions. They can be obtained using an Airy stress function satisfying the biharmonic compatibility equation.^[ref]

At $r=a$, $q=1$, so ${\sigma }_{rr}={\sigma }_{r\theta }=0$ whileThe tangential stress is $3{\sigma }_{\infty }$ at the top and bottom of the hole and $-{\sigma }_{\infty }$ at its sides. A traction-free boundary does not require every stress component to vanish: the outward boundary normal is radial, so radial normal and shear tractions must vanish.

For an original distance comparison along $\theta =\pi /2$, the tangential stress is ${\sigma }_{\infty }(1+q/2+3q^2/2)$. At $r=2a$ it is $1.21875{\sigma }_{\infty }$, already far below the edge value three. At large distance the polar components approach those of the original uniform Cartesian tension.

A finite-width plate, internal hole pressure, plasticity or out-of-plane bending changes the boundary-value problem. The stress concentration factor three is therefore specific to the stated remote-tension model. The MIT laboratory uses local strain measurements and photoelasticity to examine how actual geometries compare with ideal concentration fields.^[ref]

A mini-batch gradient averages gradients of several sampled components of a finite-sum objective. For $F(x)=n^{-1}{\sum }_{i=1}^n{f}_i(x)$ and batch indices $I_1,\dots ,I_b$,With independent uniform draws at fixed $x$, it is an unbiased stochastic gradient: $\mathbb{E}{G}_b(x)=\nabla F(x)$. Its covariance matrix is $\Sigma (x)/b$, where $\Sigma (x)$ is the covariance of one sampled gradient. The factor follows because centered independent cross terms have expectation zero.

If the indices are sampled uniformly without replacement, the covariance is instead $(n-b)\Sigma (x)/(b(n-1))$ for $n>1$, when $\Sigma$ uses the uniform-population normalization. At $b=n$ this gives the exact full gradient. Larger batches reduce sampling variance but require more component evaluations per update; biased sampling or data dependence requires its own analysis.

A mini-batch is a group of training examples processed together in one iteration of a learning algorithm. Its batch size counts examples, while a training epoch usually counts one pass over the available training data. In stochastic gradient descent, a mini-batch gradient combines the component gradients from the selected examples; the group of examples and the resulting numerical estimate are distinct objects.^[ref]

For an original counting example, partition eight examples into mini-batches with nominal batch size three and retain the final partial group. The sizes are $3,3,2$, giving three updates in one training epoch if each group produces one update. Discarding the final partial group changes which observations participate.

Unequal group sizes also matter when reporting a loss function. At fixed model parameters, suppose the three mini-batch mean losses are $1,1,10$. Their unweighted mean is $4$, but the mean over all eight examples isThe smaller last group must not receive the same weight as a larger group when the intended statistic is a per-example sample mean. This bookkeeping issue is separate from the randomness used to select a mini-batch.

SGDR means stochastic gradient descent with warm restarts. Its learning rate schedule repeatedly lowers the learning rate along a cosine curve, then raises it to begin another cycle while retaining the current parameters. Ilya Loshchilov and Frank Hutter's method uses this warm restart to continue an existing optimization trajectory.^[ref]

In cycle $i$, let $s$ be elapsed training time since the last restart and $T_i>0$ the cycle duration, measured in the same units. The cosine learning rate schedule isIt starts at ${\eta }_{\max ,i}$, reaches the midpoint of the two bounds at $s=T_i/2$, and reaches ${\eta }_{\min ,i}$ at the cycle endpoint. The original implementation updates the elapsed training epoch count within each mini-batch, allowing fractional epoch progress.^[ref]

For an original schedule, choose bounds 0 and 0.1, initial duration $T_0=4$ epochs and duration multiplier 2. The successive cycle lengths are 4, 8 and 16 epochs, with restart boundaries at total elapsed times 4, 12 and 28. In the first continuous cycle, the rates at $s=0,2,4$ are respectively $0.1,0.05,0$.

A discrete convention must specify which update uses the endpoint. For example, sampling phases $s=0,1,2,3$ in the first cycle and restarting before the next step gives rates approximately $0.1,0.08536,0.05,0.01464$, followed by $0.1$. The zero endpoint in the continuous curve is then a limit, rather than an additional zero-rate optimizer step.

Retaining parameters also has a concrete meaning. In an original deterministic gradient descent example with $f(\theta )={\theta }^2/2$ and current $\theta =0.5$, resetting only the schedule leaves $\theta =0.5$. The next step at learning rate0.1 gives0.45; a step at0.01 would give0.495. The restart changes the next movement rather than drawing a new initial parameter. Handling a momentum buffer is a further state convention; restarting the schedule and discarding all optimizer state are different operations. Record cycle phase, durations and rate bounds with an optimization checkpoint.

Clustering groups observations according to a specified notion of similarity or shared structure. Unlike supervised learning, it need not begin with known class labels. The representation, distance and objective determine what counts as a group: nearby numerical values, a connected dense region and membership in a probabilistic component are different criteria.

In K-means clustering, $K$ centers $m_j$ minimize the within-group squared-distance objectiveFor fixed groups the minimizing center is their mean; for fixed centers each point is assigned to a nearest center. These elementary minimizations explain the alternating updates. Stuart P. Lloyd's original quantization paper develops the corresponding nearest-representative and mean conditions for scalar signal reconstruction.^[ref]

For an independent one-dimensional example, the data $0,2,8,10$ can be grouped as $\{0,2\}$ and $\{8,10\}$, with means $1,9$ and objective $4$. A single center at $5$ gives objective $68$. Reducing this objective by allowing more groups does not itself tell us the scientifically appropriate $K$: with one center per observation it becomes zero.

Clustering describes the chosen representation, not an automatic discovery of natural kinds. Changing units can change distances; elongated or nested groups can challenge a squared-distance criterion; different initial centers can produce different fitted solutions. Stability, the scientific question and knowledge of how observations were collected help determine whether a partition is informative.

A training epoch is conventionally one pass through the training data used by an iterative learning procedure. Several parameter updates can occur within that pass. A mini-batch supplies a group of examples for an update, while the epoch counts progress through the data collection rather than counting individual updates.^[ref]

Suppose there are $N=10$ training examples. Visit each exactly once per pass, form mini-batches of at most $B=4$ examples and retain the smaller final group. Their sizes are $4,4,2$. With one update for every mini-batch, an epoch hasupdates. Five epochs therefore make fifteen updates and fifty example visits. Fifty visits do not mean fifty distinct examples: the same ten examples are reused.

Changing the mini-batch size changes the update count. Taking all ten examples together gives one update per epoch under the same rule, while groups of two give five. An epoch count alone consequently does not specify the number of stochastic gradient descent steps, the amount of computation or the final fitted model.

The finite-pass assumption matters. Dropping an incomplete final group in this example gives two updates and only eight visits during that iterator traversal. Reordering the examples on later passes can change which examples are omitted. Sampling with replacement can repeat some examples before visiting others, so a chosen number of sampled updates need not be a literal once-per-example pass. Report those conventions rather than assuming that every loop named epoch guarantees full coverage.

The learning rate schedule needs a declared clock too: changing the learning rate once per epoch differs from changing it once per mini-batch. The optimization tutorial separates the outer loop from the operations that clear gradients, perform backpropagation and update parameters inside the training loop.^[ref] Keeping data passes and update steps distinct makes training runs easier to compare.

A directional tipping margin is the distance an interior projection of the center of mass can move along one specified direction before reaching a boundary of its support polygon. It differs from the nearest-edge static tipping margin, which protects against movement in every horizontal direction. For a bounded polygon $P=\{x:a_i\cdot x\le b_i\}$, interior point $c$, and unit direction $u$, the ray $c+su$ stays inside exactly while all edge inequalities hold. HenceEdges with nonpositive denominator do not limit positive motion along that ray. The underlying gravity-only feasibility criterion is the nonnegative-reaction convex-hull condition in support polygon.^[ref]

For an original triangular footprint, take $x\ge 0$, $y\ge 0$, $x+y\le 1$ in metres and $c=(0.20,0.30)\mathrm{m}$. Moving right gives $s_{\max }=0.50\mathrm{m}$; moving downward gives $0.30\mathrm{m}$; moving along $u=(1,1)/\sqrt{2}$ gives $0.50/\sqrt{2}\approx 0.3536\mathrm{m}$. The all-direction margin is only $0.20\mathrm{m}$, set by the left edge. A displacement allowance restricted to the right can therefore be larger than the isotropic allowance without contradiction.

The same geometry can enter a planar load calculation. A horizontal force $Fu$ applied at height $h$ shifts the effective support-resultant point by $(Fh/mg)u$ under the corresponding static torque balance. The directional threshold is $F_{\mathrm{tip}}=mg{s}_{\max }/h$, subject also to friction and the other contact assumptions. With $m=10\mathrm{kg}$ and $h=0.50\mathrm{m}$, the rightward example gives $98.1\mathrm{N}$, whereas the nearest-edge worst-direction value is $39.24\mathrm{N}$.

A measured pull direction and contact layout are therefore part of the specification. Changes in contact set, applied couples, or slipping require the fuller analysis in tipping of a rigid body; this geometric directional reserve alone is not a dynamical stability certificate.

The scalaron is the additional propagating scalar degree of freedom in a generic nonlinear $f(R)$ metric theory of gravity. In the model $f(R)=R+R^2/(6m^2)$, it can be expressed in the Einstein frame as a canonically normalized scalar $\varphi =\sqrt{3/2}M\log f_R$, where $M$ is the reduced Planck mass and $f_R=df/dR>0$.

The potential is $V(\varphi )=3M^2{m}^2[1-e^{-\sqrt{2/3}\varphi /M}{]}^2/4$. Near its minimum, expanding the exponential gives $V\simeq m^2{\varphi }^2/2$, identifying $m$ as the scalaron's small-oscillation mass. At large positive field values the potential flattens, providing the slow-roll mechanism of Starobinsky inflation.^[ref]

Identifying flow-sensor gain and offset with timed transfers uses a known gas volume delivered over different durations to distinguish multiplicative response from accumulated zero error. In the model $Q_m=gQ+b$, a transfer of volume $V_0$ lasting $T_i$ producesThe observations form a linear regression with design matrix rows $(V_0,T_i)$. If $V_0\ne 0$, gain and offset are identifiable only when the durations are not all equal. Repeating one duration measures the same combination of $g$ and $b$ more precisely, without separating them.

For two noise-free transfers,Stipulated 3 L transfers lasting 1 and 6 s, measured as 3.020 and 3.120 L, give $\hat{b}=0.020\mathrm{L}{\mathrm{s}}^{-1}$ and $\hat{g}=1.000$. The different durations expose an offset that one transfer could misinterpret as a gain error.

Timing spread and gain-offset estimator covariance quantifies the design. Suppose $n$ observations have independent equal volume-error variance $s^2$, and define $S_{TT}={\sum }_i(T_i-\overline{T}{)}^2$. Ordinary least squares givesUnder the stipulated model, wider timing separation improves offset precision. For the two durations above and $s=0.010$ L, the offset's standard uncertainty is about $0.00283\mathrm{L}{\mathrm{s}}^{-1}$ and the gain's about 0.00406.

The reference syringe volume was treated as known in those formulas. If it is uncertain, repeated sensor tests share that reference uncertainty. Indeed the observations identify only the product $g{V}_0$ when both factors are unknown. Additional repetitions cannot establish an absolute gain without an independently characterized volume. The GUM treats fit uncertainties and common-reference correlations as distinct contributions.^[ref]

The primary spirometry statement's multiple-flow syringe checks test the practical response over different delivery rates.^[ref] Actual leaks, flow-dependent nonlinearities and uncertain timing require a richer model than a constant gain and offset.

Inverse spectral theory reconstructs an operator or its coefficients from spectral data such as eigenvalues, spectral weights or boundary responses. It asks what the data determine and what additional information is needed for uniqueness.

Eigenvalues alone generally lose information. The matrices $\mathrm{diag}(1,3)$ and $\left(\begin{array}{cc}2& 1\\ 1& 2\end{array}\right)$ have the same eigenvalues but different entries in the fixed coordinate system. Orthogonal conjugation changes eigenvectors while preserving the spectrum. An inverse problem must therefore specify whether recovery is sought in fixed coordinates, up to a change of basis, or within a more restricted class. Spectral weights can retain information about how selected measurements couple to eigenvectors.

Correlated spectral endpoint error means that the errors in two estimated frequencies have nonzero covariance. Their effect on a frequency difference depends on that covariance as well as on the uncertainty of each endpoint. Shared errors can cancel in a difference without making the absolute frequencies accurately known.

Let endpoint errors be ${ϵ}_0,{ϵ}_T$, with standard uncertainties $u_0,u_T$ and correlation coefficient $\rho$. The error in the increment is ${ϵ}_T-{ϵ}_0$, soFor equal uncertainties $u$, the result is $2u^2(1-\rho )$. Independent errors give $2u^2$, perfectly shared errors give zero, and perfect anticorrelation gives $4u^2$. This is correlated-input uncertainty propagation applied to a difference.^[ref]

An original common-offset model writes ${ϵ}_0=b+{\eta }_0$ and ${ϵ}_T=b+{\eta }_T$. Suppose $b$ has variance $s_b^2$ and the two independent residual errors each have variance $s^2$, all mutually independent. Each endpoint then has variance $s_b^2+s^2$ and their covariance is $s_b^2$, butWith $s_b=0.3$ and $s=0.1$ in compatible frequency units, each endpoint has standard uncertainty about $0.316$, while the increment error has standard uncertainty about $0.141$. This spectral common-mode offset rejection follows from subtraction, rather than from knowing the common offset's value.

A shared gain error behaves differently. If $Y_t=(1+g){\omega }_t+b$, the constant offset cancels but $Y_T-Y_0=(1+g)({\omega }_T-{\omega }_0)$. A shared $1\%$ scale error therefore remains a $1\%$ increment error. Calling an error common does not establish which part cancels.

To estimate covariance, repeated measurements should preserve the reference and data-processing relationships relevant to the two endpoints. Filtering or overlapping fitting windows can introduce correlations. The spectral frequency-increment measurement noise contribution also depends on covariance between physical fluctuations and estimation errors; the endpoint formula alone does not justify dropping that separate term. A two-dimensional infrared spectroscopy statistic should be interpreted through its actual estimator and response model, not by automatically adding independent noise to two abstract frequency labels.

Area weighting of radiant exposure combines patch-average exposures in proportion to the receiving areas they represent. For disjoint patches with areas $A_j$ and mean radiant exposures $H_j$, the coefficients ${\alpha }_j=A_j/{\sum }_k{A}_k$ satisfy ${\sum }_j{\alpha }_j=1$, and the combined spatially averaged radiant exposure is ${\sum }_j{\alpha }_j{H}_j$. This follows by adding the received energies $A_j{H}_j$, rather than averaging a list of values without its geometry.^[ref]

An original coordinate example explains a calibration issue in image-based measurements. Let a rectangular receiving strip have length $\ell$ and constant width. Stipulate $H(x)=H_{0}x/\ell$ for $0\le x\le \ell$. Suppose a calibrated image coordinate $u\in [0,1]$ maps to physical position $x=\ell u^2$. Uniform sampling in $u$ is therefore nonuniform in physical area.

The physical area average isAn unweighted average over increasingly dense, uniformly spaced $u$ samples would instead converge to ${\int }_0^1{H}_0{u}^{2}du=H_0/3$. The factor $2u$ is the change of variables formula's area correction for the chosen mapping; the constant strip width and $\ell$ cancel in the normalized average.

Thus a Jacobian area correction for radiant-exposure maps can matter even if every local exposure value is correct. An optical camera measurement needs its spatial calibration and the receiving-surface map, not just pixel values. Overlapping patches must not be counted as disjoint areas, and local point samples need an explicit quadrature model before they can stand for patch averages.

A box averaging filter gives equal weight to all positions within a specified window and zero weight outside it. In one spatial dimension, its normalized kernel and action areThis is a spatial averaging filter expressed as a convolution. In several dimensions, averaging over an axis-aligned rectangular box uses the product of the one-dimensional kernels.

The discrete counterpart averages a fixed number of samples. On an infinite sequence that is zero except for value ten at index zero, a centred five-sample box average produces value two at indices $-2,-1,0,1,2$ and zero elsewhere. The total remains ten, while the isolated peak spreads. On a finite array, boundary extension must be specified; reflection, zero padding and wrapping can give different edge values.^[ref]

Applying the same box average twice is not the same operation as applying it once. The new kernel is the convolution of the two kernels:Indeed, the integral counts the length of overlap of two intervals, divided by ${\Delta }^2$. The resulting support has width $2\Delta$, and the weight decreases away from its centre.

For a Fourier mode $\cos (kx)$, one application multiplies the amplitude by $\sin (k\Delta /2)/(k\Delta /2)$; two applications square this multiplier. The sinc function response explains why a box average smooths a signal without implementing an ideal sharp frequency cutoff. In large eddy simulation, that distinction matters when relating a grid filter to a coarser test filter.

A grid filter defines the resolved field associated with a large eddy simulation, commonly written $\bar{u}=G_{\Delta }\ast u$. Its name reflects the connection between resolved scales and a computational grid, but the filtering operator and the mesh are different specifications. An explicit spatial averaging filter states its kernel; in other methods the retained representation and numerical operations supply an implicit scale selection.^[ref]

A characteristic width is often related to cell dimensions. One possible three-dimensional convention isCells with widths $0.01$, $0.02$ and $0.04\mathrm{m}$ then have $\Delta =0.02\mathrm{m}$. This single number preserves the volume scale but does not describe the strong directional differences of the cell. The filter shape and directional resolution still matter.

Sampling alone also cannot determine which unresolved oscillations were present. At eight equally spaced points on a periodic interval of length $L$, the modes with indices ten and two have identical sine samples:This is aliasing. Removing the high mode by a specified filter before sampling and merely sampling the unfiltered field are different operations.

The subgrid-scale stress depends on the filter actually adopted. A test filter then applies an additional operation to the already resolved data. Reporting only a grid spacing, without the filtering and discretization choices, can leave the mathematical meaning of the resolved field and its closure ambiguous.

The inhomogeneous $H^1$ norm on a domain $D\subseteq {\mathbb{R}}^n$ iswhere the derivatives are weak derivatives. It measures both the size of the function and its first-order variation. The underlying Sobolev space consists of the $L^2$ functions whose first weak derivatives also lie in $L^2$.

For $u(x)=x$ on $(0,1)$, the squared norm is $1/3+1=4/3$, so the norm is $2/\sqrt{3}$. A nonzero constant has zero derivative but positive $H^1$ norm. Thus the gradient alone defines a seminorm on the full space, not the displayed norm.

On a suitable zero-boundary space over a bounded domain, a Poincaré inequality controls the $L^2$ term by the gradient and makes the gradient norm equivalent to the full norm. The boundary or normalization condition is essential for this replacement.

In finite-population sampling, a simple random sample of fixed size $n$ without replacement is a subset $S$ of $N$ labelled units drawn so that every size-$n$ subset has probabilityThe sample is the selected subset; simple random sampling is the procedure that generates it. This definition does not require the observed values to be distinct, only the selected population labels. The without-replacement convention is important because an independent repeated draw from a population is a different design.^[ref]

One construction randomly orders all $N$ labels and takes the first $n$. Each particular subset can occupy those positions in $n!(N-n)!$ permutations, out of $N!$, proving the probability above. Equivalently, draw successively and uniformly from the remaining labels, then discard their order.

For $N=5,n=2$, there are $\left(\genfrac{}{}{0ex}{}{5}{2}\right)=10$ possible samples. A particular unit belongs to four of them, so its inclusion probability is $2/5$; a particular pair is one complete sample, so its joint inclusion probability is $1/10$. These are properties of the procedure before observing which subset was obtained.

A simple random sample need not resemble the whole population in every realization. For fixed values $(0,0,0,0,100)$, six of the ten size-two samples contain only zeros. Thus the sample mean is zero with probability $0.6$, although its expected value is the finite population mean $20$. The remaining four samples each have mean $50$, giving $0.6\times 0+0.4\times 50=20$. Unbiasedness concerns repeated selection, rather than a guarantee about any one sample. Coverage of the population labels and faithful implementation of the selection rule remain essential.