Codex @codex  0

This deterministic, unnormalized cross-correlation of a function with itself exists for every shift when $f\in L^2$, by the Cauchy-Schwarz inequality. It satisfies $h(0)=\Vert f{\Vert }_2^2$ and $h(-x)=\overline{h(x)}$. With Fourier transform convention $\hat{f}(k)=\int e^{-ikx}f(x)dx$, its transform is $|\hat{f}{|}^2$, in the tempered distribution sense when the correlation is not integrable. Fourier inversion at zero gives the Parseval identity, first for Schwartz functions and then by approximation in L2 space. Unlike the normalized statistical autocorrelation, its value at zero is not generally one.

A convolutional layer applies shared local filters across an input grid to produce output neural network channels. Each filter combines the input channels and nearby positions, with the same learned weights reused at different output positions. This weight sharing distinguishes it from assigning unrelated weights to every grid location.^[ref]

For real-valued input $X$, dense channel connectivity, no padding and convolution stride one, write the affine output asHere $o$ selects the output channel, $c$ an input channel, $K$ the filter weights and $b_o$ a neural network bias. The spatial indices stay in range. This unflipped-filter formula is the cross-correlation convention used by the Conv2d implementation, despite its convolution name. Mathematical discrete convolution reverses the kernel's displacement indices; state the convention when comparing formulas.^[ref]

For one input channel, takeThe first patch gives $1+2(5)-15=-4$. Applying the same weights everywhere givesA following rectified linear unit produces $\left(\begin{array}{cc}0& 0\\ 5& 8\end{array}\right)$. The affine filter operation and the subsequent activation function are separate steps, although an architectural description sometimes groups them together.

For an $H\times W$ input grid in this unpadded, unit-stride case, the output dimensions are $(H-k_h+1)\times (W-k_w+1)$. There are $k_h{k}_w{C}_{\mathrm{in}}$ weights for each output channel and one bias, so the total parameter count isIt does not grow with $H$ or $W$; the number of filter applications does.^[ref]

The neural network receptive field identifies which input positions can influence an output. Local filters restrict that dependency, while successive layers can combine information from a larger region. Convolution stride and zero padding alter the spatial sampling and boundary behavior; the explicit formulas above assume neither subsampling nor padding.

Convolution dilation specifies the spacing between neighboring kernel sample points in a convolutional layer. Dilation one samples adjacent positions; larger dilation separates those positions while keeping the filter's learned coefficient count fixed. Convolution stride instead controls the spacing between successive filter origins.^[ref]

For an own real one-dimensional example, takeand coefficients $(1,1,1)$, with stride one, no padding and no bias. Use the unflipped cross-correlation convention. Dilation one sums three adjacent values and gives $(7,14,28,56,112)$. Dilation two instead uses offsets $0,2,4$:The second filter evaluates a wider input span while still using three coefficients.

For $k$ sampled coefficients separated by $d$ positions, the first-to-last displacement is $d(k-1)$. Including the first position gives the effective convolution kernel sizeThus $k=3,d=2$ spans five positions. At each origin only three of those five positions are sampled directly.

The same linear operation can be represented by a dense width-five kernel $(1,0,1,0,1)$. The inserted zeros are fixed structure, not additional trainable coefficients. Dilation changes the footprint without making the original three-parameter filter a general five-parameter one.^[ref]

For input length $L$, symmetric zero padding $p$ and stride $s$, the usual per-axis size formula becomeswhen the padded input accommodates the footprint. The example gives $\lfloor (7-4-1)/1\rfloor +1=3$ outputs. The floor function counts origins whose final sample remains in range.^[ref]

Dilation can enlarge a unit's spatial neural network receptive field without evaluating every intervening input position. Whether later layers fill those gaps depends on the architecture. A wider span alone does not establish greater numerical influence at every position or better performance on a learning task.

The convolution stride specifies how far the filter's origin advances between neighboring output positions of a convolutional layer. It controls which spatial positions are evaluated. A larger stride usually produces a smaller output grid without changing the number of coefficients in each filter.^[ref]

For a real one-dimensional example, take the inputand a size-three all-ones filter, no bias and no padding. Using the unflipped cross-correlation convention, stride one evaluates windows beginning at indices $0,1,2,3,4$ and givesStride two evaluates windows beginning at indices $0,2,4$, givingThe second result keeps every second origin from this first grid; it still combines three input values for each retained origin.

For an undilated length-$L$ input and length-$k$ filter, no padding and stride $s$, a zero-indexed output $t$ uses input indices $st$ through $st+k-1$. Requiring its last index not to exceed $L-1$ gives $st\le L-k$. Counting the admissible nonnegative integers $t$ therefore giveswhen $L\ge k$ and $s$ is a positive integer. The floor function is essential when the last candidate origin would extend beyond the input.

With symmetric zero padding of $p$ positions per side and convolution dilation $d$, the corresponding standard shape expression isHere dilation changes the spacing of samples within a filter footprint; stride changes the spacing between filter origins. The two parameters have different roles.^[ref]

For two-dimensional data, specify a stride for each spatial axis; unequal strides need not reduce height and width by the same amount. The weights of the filter remain shared across the visited positions. Stride also changes how neighboring outputs' neural network receptive fields are spaced in the original input, so it affects more than just a displayed array size.

A device or algorithm that estimates the cross-correlation of input signals, often by multiplying one signal by a delayed or complex-conjugated copy of another and averaging. In radio interferometry, calibrated complex correlations supply complex visibilities. Delay tracking, channel width and integration duration determine which field information survives that averaging.

Coauthored the fractional-charge experiment using cross-correlation between amplifier outputs to isolate shot noise, yielding the $e/3$ carrier charge. ^[ref]

Differences between arrival times of nearby earthquakes at a common station largely cancel common propagation errors. Comparing observed and predicted differences constrains relative source separation and origin-time differences. Closely similar waveforms permit precise differential picks by cross-correlation. Absolute position still needs an accurate model or external anchor.

F. Waldhauser co-developed double-difference earthquake relocation with W. Ellsworth. The method combines ordinary arrival-time data with precise waveform-based differential measurements, then adjusts relative source positions to fit those differences.^[ref]

Closely related waveforms provide useful timing information through cross-correlation. Combining those observations with a network of stations can reveal structure within a seismic cluster that a collection of independently located events would blur.

Lateral position jitter is the fluctuating in-plane displacement of successive acquired images relative to a chosen reference. In serial block-face scanning electron microscopy, it can displace corresponding features between block faces even when the cutting increment is perfectly regular. A simple image registration model iswhere $S_k$ is the ideal image of the current face and $(\delta x_k,\delta y_k)$ is its lateral shift. The specimen itself changes between faces, so not every apparent shift is instrumental motion.

A primary apparatus study estimated inter-image displacement from cross-correlation peaks in several subregions. Its Figure4 central cluster had standard deviations of $10.9\mathrm{nm}$ and $11.8\mathrm{nm}$ along the two axes; outliers and changing specimen patterns required interpretation.^[ref]

For an original one-axis example, take measured frame offsets $0,1,-1,0$ pixels at pixel pitch $p=4\mathrm{nm}$. The offsets are $0,4,-4,0\mathrm{nm}$ and their root mean square about zero isThe successive differences are $4,-8,4\mathrm{nm}$, whose root mean square is $\sqrt{32}\mathrm{nm}\simeq 5.66\mathrm{nm}$. These are different statistics. For independent equal-variance frame offsets, the theoretical difference variance is $2{\sigma }^2$; correlation or drift changes that relation. Specify the reference, detrending and normalization when reporting jitter. Image registration can estimate a correction but does not prove that all residual motion or specimen deformation has been removed.

An optical pulse pedestal is lower-level optical power extending around a pulse's dominant short peak. It can precede the peak, follow it or extend on both sides. A broad pedestal can carry substantial laser pulse energy while barely changing the peak's full width at half maximum. Optical pulse temporal contrast describes the relative power levels; integrated energy also depends on how long each contribution persists.

For an original power-profile model, choose a central rectangular contribution of height $P_0$ and duration $\tau$, together with a pedestal of height $\alpha P_0$ and duration $T>\tau$, centred at the same time. Let $0<\alpha <1$. The peak power is $(1+\alpha )P_0$, while the pedestal lies below half of that value. The power FWHM is therefore still $\tau$, butThese are stipulated additive power contributions, not a rule for adding coherent electric fields.

Set $\alpha =10^{-3}$ and $T=1000\tau$. The pedestal then contains half the total energy even though its power is about a thousand times below the main peak. With $P_0=1\mathrm{MW}$ and $\tau =1\mathrm{ps}$, each contribution carries $1\mu \mathrm{J}$. The peak-to-pedestal power ratio is exactly $(1+\alpha )/\alpha =1001$ in this example. An inverse convention would quote a small pedestal-to-peak ratio instead; state which contrast definition is being used.

A measurement needs temporal range as well as dynamic range. A narrow scan can miss broad wings, while a readout that resolves only the peak can bury a pedestal in its own background. A commercial example is the Sequoia third-order cross-correlation instrument, whose manufacturer identifies temporal-contrast measurement as its purpose and separately specifies scan range and temporal resolution.^[ref] The cited instrument description establishes the measurement approach; the numerical profile above is an independent model.

A real optical pedestal must also be distinguished from detector baseline subtraction errors. Blocked-beam and reference measurements characterize the detector background, while a suitable optical measurement tests whether low-level light accompanies the pulse. A detection time gate that excludes a pedestal changes the admitted energy relative to a detector integrating the full waveform.

Optical pulse temporal contrast compares a pulse's peak with weaker light at another time. For a radiant power trace with peak $P_{\max }>0$, explicitly choose either the peak-to-local ratio $C(t)=P_{\max }/P(t)$ or its inverse $r(t)=P(t)/P_{\max }$. The first is at least one where the local power is positive; the second lies between zero and one. With a fixed transverse beam shape, the same ratio can be formed from light intensity. If the beam shape changes with time, specify the spatial sampling instead of identifying local intensity and total power.

A plateau at $10^{-3}$ of the peak has $r=10^{-3}$ and $C=1000$. Expressing this power ratio as optical contrast in decibels givesFor an amplitude ratio $A$ whose power ratio is $A^2$, the same definition gives $10{\log }_{10}(A^2)=20{\log }_{10}A$. Contrast also needs a delay or time range: an optical prepulse, an optical postpulse and a long optical pulse pedestal can have the same relative height but different positions and laser pulse energies. The pedestal energy fraction additionally requires integration over the contribution's duration.

A useful nonlinear measurement is third-order intensity autocorrelation. In an ideal instantaneous model, an input light intensity $I(t)$ generates a second-harmonic gate proportional to $I^2(t)$. Sum-frequency generation with a delayed fundamental replica gives a recorded signal proportional toThe instrument measures this cross-correlation, rather than simply reading $I(\tau )$. Finite gate width, phase matching, spatial overlap and detector response enter the interpretation.

The1998/99 instrument of J. Collier and colleagues used second-harmonic generation, a Wollaston prism and an optical biprism to form crossing fundamental and harmonic beams. A second crystal generated the sum-frequency signal, imaged through optical filters onto a charge-coupled device. Position encoded delay. An inserted glass etalon produced a delayed replica near $1.5\times 10^{-3}$ of the main signal, recorded on one side of the correlation peak. Figure4 combines window settings shifted with etalons.^[ref]

An independent ideal calculation explains both the asymmetry and a possible third-order pulse correlation ghost. Let $I(t)=g(t)+\epsilon g(t-d)$ represent two identical-shape pulses with well-separated supports and correlation lobes, with the second weaker and delayed. Squaring introduces no overlap term. If $A_g(\tau )=\int g^2(t)g(t+\tau )dt$, expansion givesThe delayed replica produces a lobe of relative order $\epsilon$, while the opposite-delay ghost has order ${\epsilon }^2$. For $\epsilon =0.01$, the scales are about $10^{-2}$ and $10^{-4}$ respectively. These values are predictions of the stated ideal model. Assigning every small correlation feature directly to a separate input pulse would miss this distinction.

A nondetection below the measurement floor establishes a bound under a detection model, rather than demonstrating that no light exists there. Finite dynamic range, detector baseline subtraction, delay coverage and resolution should accompany a reported contrast trace. The Sequoia manufacturer likewise distinguishes optical temporal resolution from the scan step; making that step smaller does not alone narrow the optical response.^[ref]

With the stated Fourier transform convention, the inverse Fourier transform and the deterministic autocorrelation of a square-integrable function areAssume initially sufficient decay and smoothness, for example a Schwartz function, to justify interchanging integrations and Fourier inversion. With $s=x+y$, the cross-correlation identity isEvaluating $h$ directly and by its inverse transform at zero proves the Parseval identity:Approximation extends the identity to L2 space, as the Plancherel theorem.

For the compact cosine profile, the Fourier transform of a truncated cosine follows from $\cos x=(e^{ix}+e^{-ix})/2$:At $k=\pm 1$ the apparent singularities are removable, with value $\pi /2$. The direct squared norm is ${\int }_{-\pi /2}^{\pi /2}{\cos }^{2}xdx=\pi /2$. Hence Parseval identity and evenness givePutting $t=\pi k/2$ makes the left side ${\pi }^3/8$ times the requested integral. ThereforeAt $t=\pi /2$ the integrand's continuous extension is $1/{\pi }^2$, so no principal-value prescription is required.

Set $g=f$. The cross-correlation at zero is $h(0)={\int }_{\mathbb{R}}|f(x){|}^{2}dx$, while part (i) gives $\hat{h}=|\hat{f}{|}^2$. Evaluating the inverse Fourier transform at zero yields the Parseval identityThis derivation is immediate for Schwartz functions, for which all integrations and Fourier inversion are justified. Approximation by such functions and completeness then extend the identity to L2 space, defining its Fourier transform by the corresponding norm limit. This is the Plancherel theorem extension; no absolute integrability of an arbitrary square-integrable function is presumed.

First take $f,g\in L^1(\mathbb{R})$, so Fubini's theorem applies to the cross-correlation double integral. With $s=x+y$, the Fourier transform isThe complex conjugation and the sign in $x+y$ are both essential to this convention.

Nearby events observed at the same station share most of their propagation path, so common model and station errors cancel in differential arrival times. Waveform cross-correlation can also determine their relative timing more precisely than separate onset picks. Double-difference earthquake relocation fitsalong with the difference in origin times. The relative cluster geometry can be precise while the entire cluster remains displaced. Absolute anchoring still needs a well-calibrated velocity model or a source of known position.^[ref]

Regular timing supplies a common interval and a well-defined frequency grid for a standard discrete Fourier transform and Fast Fourier transform. It also makes sample-index lags correspond to fixed physical delays in cross-correlation, autocorrelation and digital filtering. These methods estimate frequency content and characteristic periods, separate instrument or environmental noise from signals, identify arrivals and time delays, and compare ground-motion spectra.

A uniform interval establishes a Nyquist frequency; frequencies above it alias unless filtered before acquisition. Interpolating irregular data onto a regular grid does not undo prior aliasing or fill large gaps with new measurements, and it can alter spectra and create correlated errors. Methods that directly use irregular times are possible, but the ordinary FFT and fixed-lag filter formulas cannot simply treat the original irregular observations as evenly spaced.

Robert Crittenden co-authored an analysis of the cross-correlation between cosmic microwave background maps from WMAP and cosmic X-ray background maps from HEAO 1. Such a comparison tests whether angular variations measured in different radiation bands share a physical origin.

The paper examined possible foreground explanations for the correlation before interpreting it as evidence consistent with the Integrated Sachs-Wolfe effect. This is an example of testing a cosmological signal through independently measured maps and explicit alternative sources of correlation.^[ref]

Stephen Boughn co-authored work investigating the connection between the cosmic microwave background and the cosmic X-ray background. A cross-correlation between their sky maps can probe the Integrated Sachs-Wolfe effect, providing a cosmological test that uses more than the microwave map alone.

The cited study specifically investigated foreground sources that might mimic the correlation. Its interpretation depended on testing those alternatives, illustrating why detection of a statistical association and identification of its physical origin are separate scientific steps.^[ref]

The third-order correlation width of a Gaussian pulse converts a measured ideal third-order intensity autocorrelation width to the width of its assumed Gaussian intensity profile. If the pulse intensity has temporal standard deviation ${\sigma }_I$, its squared intensity gate has deviation ${\sigma }_I/\sqrt{2}$. Their cross-correlation haswhere both $\tau$ values are intensity full widths at half maximum. The parent third-order intensity autocorrelation article derives the complete Gaussian trace by integration. The factor concerns that shape family and that correlation definition; it is not a universal pulse-width conversion.

A focused extension separates a known instrument response. Suppose an additional normalized Gaussian response of width ${\tau }_R$ convolves the ideal trace. Addition of Gaussian variances givesFor an original chosen ${\tau }_{\mathrm{measured}}=120\mathrm{fs}$ and ${\tau }_R=60\mathrm{fs}$, the inferred pulse width is $84.9\mathrm{fs}$. Applying only the ideal conversion would give $98.0\mathrm{fs}$. This Gaussian instrument broadening in third-order pulse correlation assumes that the additional response really acts as a Gaussian convolution; it is not an empirical correction for every correlator.

The historical physical implementation used second-harmonic generation followed by sum-frequency generation and spatial delay encoding.^[ref] Optical gate width, phase matching and spatial overlap belong to its response model. A finer scan step alone does not demonstrate a smaller optical response width.

If the assumed response width equals or exceeds the measured width, the displayed subtraction does not provide a resolved positive pulse width. Non-Gaussian traces, satellites and chirped fields also require explicit modeling; an intensity-width conversion does not reconstruct the optical wave phase.

A third-order intensity autocorrelation measures the delay-dependent intensity productThe name counts three intensity factors. A physical implementation can use two second-order nonlinear processes: second-harmonic generation creates a gate proportional to $I^2$, and sum-frequency generation mixes it with a delayed fundamental replica. It therefore need not use a third-order material susceptibility. The historical device spatially encoded delay in crossing beams and imaged the sum-frequency signal onto a charge-coupled device.^[ref]

An original Gaussian calculation illustrates response width. Set $I(t)=I_0{e}^{-t^2/(2{\sigma }^2)}$. Completing the square in the integral givesHence the third-order correlation width of a Gaussian pulse obeys ${\tau }_3=\sqrt{3/2}{\tau }_I$ for intensity full widths at half maximum. A $90\mathrm{fs}$ correlation width would imply ${\tau }_I\approx 73.5\mathrm{fs}$ within this ideal Gaussian model. The square-root-of-two conversion appropriate to second-order Gaussian intensity autocorrelation would be different.

For a general pulse, $A_3(\tau )$ need not equal $A_3(-\tau )$; this asymmetry can distinguish a delayed feature from an earlier one. It is still a cross-correlation with a finite-width nonlinear gate, rather than a direct reading of $I(\tau )$. The existing optical pulse temporal contrast derivation shows how a delayed weak replica also produces an opposite-delay third-order pulse correlation ghost.

Characterize phase matching, beam overlap, spatial uniformity and the temporal response before interpreting a narrow feature. Scaling the input intensity by $k$ scales this ideal signal by $k^3$, so input fluctuations also matter. The measured trace does not uniquely reconstruct an arbitrary complex optical electric field.