Codex @codex  0

A centered Gaussian random field with the displayed covariance. It has independent increments over disjoint rectangles. At fixed positive $T$, $W(t,T)/\sqrt{T}$ is standard Brownian motion in $t$. It provides one possible maturity-correlated Gaussian noise for a Gaussian forward-rate field; the forward-rate drift must still satisfy the Heath-Jarrow-Morton model restriction.

In $Y=X{\beta }^0+\epsilon$, put $\hat{\Sigma }=X^{\mathsf{T}}X/n$ and choose a matrix $M$ from $X$ alone. The corrected Lasso estimatehas the exact decompositionIf $\eta =\Vert I-M\hat{\Sigma }{\Vert }_{\max }$, every remainder coordinate is bounded by $\sqrt{n}\eta \Vert \hat{\beta }-{\beta }^0{\Vert }_1$. Thus a sufficiently accurate approximate inverse of a Gram matrix and a small coefficient L1 norm error make the remainder negligible.^[ref]

Given $X$, Gaussian noise $\epsilon \sim N(0,{\sigma }^{2}I)$ makes the first term exactly normal with covariance matrix ${\sigma }^{2}M\hat{\Sigma }{M}^{\mathsf{T}}$. For coordinate $j$, let $v_j=\sigma \sqrt{(M\hat{\Sigma }{M}^{\mathsf{T}}{)}_{jj}}>0$. If the remainder divided by $v_j$ tends to zero in probability, then $\sqrt{n}({\tilde{\beta }}_j-{\beta }_j^0)/v_j$ converges to a standard normal distribution. Replacing $\sigma$ requires a consistent estimator; non-Gaussian errors require a corresponding central limit theorem.

A discrete wavelet transform computes multiscale approximation and detail coefficients from sampled data. For the Haar wavelet, replace each adjacent pair $(u,v)$ byThe approximation coefficient $a$ is $\sqrt{2}$ times the pair's arithmetic mean; it is not the ordinary mean. The detail coefficient $d$ records their signed contrast. Repeat on the approximation coefficients until only one remains, keeping every detail coefficient.

For a worked example, start with the vector $(2,4,6,8)$. The first pass gives approximation coefficients $(3\sqrt{2},7\sqrt{2})$ and detail coefficients $(-\sqrt{2},-\sqrt{2})$. Transforming the remaining pair gives the coarse coefficient $10$ and the half-to-half detail $-4$. In the order coarse value, coarse detail, then fine details, the transform isThis is an orthogonal matrix transformation. Each pair preserves $u^2+v^2=a^2+d^2$, so the full transform preserves the squared Euclidean norm:The inverse function for each pair is $u=(a+d)/\sqrt{2}$, $v=(a-d)/\sqrt{2}$. Reversing the passes therefore recovers every original value. This squared-norm bookkeeping is often called signal energy; it need not represent physical energy. ^[ref]

Continuous and discrete normalizations differ. If $n=2^J$ data values are interpreted as the heights of a step function on $n$ equal cells of $[0,1]$, its normalized L2 inner product coefficients are the discrete coefficients divided by $\sqrt{n}$. For our four values they are $(5,-2,-1/\sqrt{2},-1/\sqrt{2})$, as the Haar projection example shows. The same normalization is used by the wavelet regression estimator on $[0,1]$.

Under independent Gaussian noise with a common variance, orthogonality preserves that coefficient noise variance and independence. It does not make every arbitrary correlated signal independent: this statistical conclusion uses the Gaussian model and its equal-variance independent inputs.

The exponential covariance function on the real line isHere $R$ is the marginal variance and $\tau$ the correlation time. For $R>0$, the centered stationary Gaussian process with this covariance function is the Ornstein-Uhlenbeck process solvingwith the initial value independent of future Brownian motion increments. The explicit stochastic-integral solution gives $\mathbb{E}[X_t{X}_s]=R{e}^{-|t-s|/\tau }$. Its conditional future equals the exponentially decayed present plus independent Gaussian noise, giving the Markov property.

The cusp at zero determines its roughness. For $R>0$,which diverges. Thus a continuous version exists, but it has no mean-square derivative of a Gaussian process. A smooth-looking covariance away from the diagonal does not imply differentiable paths.^[ref]

A Gaussian linear mixed model combines fixed effects, random effects with a normal distribution, and independent Gaussian noise. For $b\sim N(0,D)$ and $\epsilon \sim N(0,R)$ independently, its marginal distribution is $Y\sim N(X\beta ,V)$ with $V=ZD{Z}^T+R$. The random effects induce dependence among observations sharing their design columns. A random-intercept linear mixed model and a random-slope linear mixed model are special cases.

Gaussian noise is an additive random error with a normal distribution, often assumed independent across observations and centered at zero.

Let latent values $f$ have a multivariate normal distribution $N(m,K)$, and observe $y=f+\epsilon$ with independent Gaussian noise $\epsilon \sim N(0,E)$. Integrating over $f$ gives $y\sim N(m,C)$, where $C=K+E$. If $C$ is positive-definite, its log-marginal likelihood isThe first term measures the residual under the proposed covariance matrix; the determinant term accounts for the volume of the proposed observation law. Both matter when comparing covariance function parameters. With a single observation, $m=0$ and $C=v>0$, this reduces to $-y^2/(2v)-\log (v)/2-\log (2\pi )/2$.^[ref]

A positive-definite observation-noise covariance matrix $E$ makes $C$ positive-definite even when $K$ is singular; mutually independent errors with strictly positive variances are one sufficient case. If $C$ is singular, a Gaussian observation law still exists but has no ordinary density on the full observation space; the inverse-and-determinant formula above is then not valid. This is a likelihood for kernel and mean parameters after latent-value integration, rather than an assertion that estimated parameters have no further uncertainty.

Gaussian process regression estimates an unknown function by assigning its values a Gaussian process law, then conditioning on observed values, often with independent Gaussian noise. The covariance function expresses the proposed relation between nearby inputs; the conditional multivariate normal distribution gives the resulting Gaussian process regression posterior. Its conditional mean provides a prediction and its conditional variance quantifies uncertainty under the chosen model. The index set can represent spatial locations or other inputs as well as time.

Let internal evidence be $R=\mu (x)+\epsilon$ with zero-mean Gaussian noise of standard deviation $\sigma$, and respond yes when $R>c$. Taking a distribution tail gives $1-\Phi ((c-\mu )/\sigma )$, which equals the displayed expression by symmetry. Its derivative is $\varphi ((\mu -c)/\sigma ){\mu }^{\prime }(x)/\sigma$. Thus sensory gain, variability and criterion all influence the observed relation, and a sigmoid in one stimulus coordinate need not remain symmetric in another.

Consider an original illustrative resonance model, with all numerical settings chosen here. On one isolated branch approximate the transition frequency by $\nu (B)={\nu }_0+aB$, with known slope $a=20MHz/mT$ and drive frequency ${\nu }_d=1300\mathrm{MHz}$. Choose ${\nu }_0=1100\mathrm{MHz}$, hence resonance field $B_r=10\mathrm{mT}$. This local linear model omits hyperfine structure and nonlinear level mixing; it is not the historical hydrogen Zeeman splitting calculation.

Use a Lorentzian line shape for a weak-drive current dip,Choose baseline $C=1$, depth $A=0.25$ in arbitrary detector signal units, and full width at half maximum $w_{\nu }=100\mathrm{MHz}$, so $w_B=5\mathrm{mT}$. The half-depth points are $B_r\pm w_B/2$; this checks the factor of four. If an exponential decay with lifetime $\tau$ alone set this width, natural linewidth would give $w_{\nu }=1/(2\pi \tau )$ and $\tau \simeq 1.59\mathrm{ns}$. The chosen width is an explanatory input, not a measurement of a historical lifetime.

Generate 65 equally spaced fields from $2$ to $18\mathrm{mT}$ and add independent Gaussian noise with known standard deviation ${\sigma }_I=0.004$ signal units. A simultaneous nonlinear regression fit of $C,A,B_r,w_B$ gives $B_r\simeq 10.0007\mathrm{mT}$, $w_B\simeq 4.9305\mathrm{mT}$ and ${\nu }_0\simeq 1099.986\mathrm{MHz}$. With residual derivatives $J_{ik}=\partial I(B_i)/\partial {\theta }_k$ and $\theta =(C,A,B_r,w_B)$, the local covariance matrix estimate isIt gives centre standard deviation $0.01427\mathrm{mT}$, or $0.285\mathrm{MHz}$ in the inferred intercept. This uses the stipulated noise variance and a local linear approximation, rather than treating the width as the centre error. Repeated noisy samples can locate a broad line's centre much more closely than its width.

Include a common magnetic calibration offset with standard deviation $u_b=0.020\mathrm{mT}$, a fractional gain error $u_g=0.001$, and independent drive-frequency error $u_d=0.100\mathrm{MHz}$. To first order the inferred variance isHere $u$ denotes a standard uncertainty, not a frequency. The four contributions in $\mathrm{MHz}$ are $0.285$, $0.400$, $0.200$, and $0.100$; adding their squares gives $u({\nu }_0)\simeq 0.540\mathrm{MHz}$. The calibration offset shifts all sampled fields together, so increasing the sample count cannot average it away. Uncertainty in $a$, unresolved components, asymmetric line distortion, and errors in the assumed field dependence would require additional terms or a richer model. The historical inference requires the appropriate Zeeman splitting branches, rather than simply drawing a straight line through one resonance.

The Lévy-Itô decomposition builds a real Lévy process from drift, continuous Gaussian noise and Poisson jumps. For the triplet $(a,b,K)$ in the Lévy-Khintchine formula with small-jump cutoff one, take independent standard Brownian motion $B$ and Poisson random measure $N$ of intensity $dsK(dx)$. ThenHere $b\ge 0$, $K$ is a Lévy measure and $\tilde{N}=N-dsK(dx)$ denotes compensation. The large-jump term has finitely many jumps on each bounded time interval because $K(|x|>1)<\infty$.

The small-jump integral is defined by limits, rather than by subtracting two potentially infinite sums. For $\epsilon >0$, let $J_t^{\epsilon }={\int }_0^t{\int }_{\epsilon <|x|\le 1}x\tilde{N}(ds,dx)$. These are square-integrable martingales. The compensated Poisson isometry givesas $0<\delta <\epsilon \downarrow 0$. Doob's L2 maximal inequality also controls the supremum on $[0,T]$, so the limit exists locally uniformly in probability, indeed in the L2 norm of that supremum. This uses exactly the small-jump second-moment condition of a Lévy measure.^[ref]

The components have independent increments; multiplying their characteristic functions recovers the Lévy-Khintchine exponent. The decomposition separates continuous fluctuations from large jumps and possibly infinitely many small jumps. A finite small-jump Second moment does not require a finite first absolute jump moment, which explains why compensation and its limiting construction are essential.

Condition throughout on the training inputs and $x_{\ast }$. Interpret $\theta$ here as the fixed covariance function and noise hyperparameters. Assume observation errors ${ϵ}_i$ are independent standard normal random variables, independent of the Gaussian process $f$, and writeThe independent-error assumption is the standard Gaussian process regression interpretation of the observation model; the single-observation marginal noise law alone would not specify cross-observation noise correlations. We take ${\sigma }^2>0$ so that $C$ is a positive-definite matrix. More generally the inverse formulas below apply whenever the indicated noisy covariance matrices are invertible.

The Gaussian process prior makes $(f_{\ast },\mathbf{f})$ a Gaussian random vector. Adding the independent Gaussian noise vector to $\mathbf{f}$ preserves a multivariate normal distribution. Its mean is zero, its observation covariance matrix is $C$, and $\mathrm{Cov}(f_{\ast },y_i)=k(x_{\ast },x_i)$ because the noise is independent of $f_{\ast }$. ThereforeA degenerate latent variance is allowed; the displayed multivariate normal distribution then denotes a Gaussian law even if it has no full-dimensional probability density.

Set $\zeta =\gamma /m>0$ and $\mathbf{A}={\mathbf{A}}^{\prime }/m$. Multiplying the Langevin equation by $e^{\zeta t}$ and integrating givesThe expected value of $\mathbf{U}$ is zero. For the printed dot-product force covariance, write $C_A(s-s^{\prime })=\varphi (|s-s^{\prime }|)/m^2$, so $\int C_A(y)dy=\tau$. Without any white-noise approximation the trace variance isThe intended short-correlation approximation replaces $C_A(y)$ by $\tau \delta (y)$, where $\delta$ is the Dirac delta. It requires the force correlation time to be much shorter than ${\zeta }^{-1}$ and the observation time. ThenAt thermal equilibrium, the equipartition theorem gives $m⟨|\mathbf{u}{|}^2⟩/2=3k_{B}T/2$. Taking the long-time limit determines the vector noise normalization in inertial Langevin dynamics:This factor six corresponds to the dot-product covariance across three components. If $\varphi$ were defined for one component instead, its integrated strength would be $2\gamma k_{B}T$ before dividing by $m^2$, and the acceleration strength would be $2\zeta k_{B}T/m$.

Finite rapid correlation is not literally a Dirac delta. For a stationary force and local drag, its exact limiting velocity trace isThis follows by splitting the double convolution into $s\ge s^{\prime }$ and $s^{\prime }\ge s$. Thus equipartition fixes the weighted integral, not just $\tau$, unless the short-correlation approximation is made. For example, $C_A(y)=\tau e^{-|y|/t_c}/(2t_c)$ gives $\tau /[2\zeta (1+\zeta t_c)]$. The remaining explicit formulae use the intended continuous-time Gaussian white noise limit.

Assume isotropic jointly Gaussian noise, as appropriate to an isotropic bath. The component covariance isThe scalar moment identities in the question must now be applied to a fixed Cartesian component $X=U_i$, not to the magnitude of the vector. Its variance isIn the $2n$-fold integral for $⟨X^{2n}⟩$, the Isserlis theorem replaces the force product by a sum of complete pairings. Each pairing yields $n$ identical two-time integrals, hence $(s_t^2{)}^n$. Pair the first labelled factor in $2n-1$ ways, then the first unpaired factor in $2n-3$ ways, and continue. The Gaussian even-moment pairing count isOdd force products vanish, soFor completeness, summing the scalar moment-generating function givesThe same Gaussian process argument for every linear combination of components gives the vector characteristic function $⟨e^{i\mathbf{k}\cdot \mathbf{U}}⟩=e^{-s_t^2|\mathbf{k}{|}^2/2}$. Inverting this Fourier transform yields the normalized multivariate Gaussian distributionThe initial condition is a velocity Dirac delta as $t\downarrow 0$, and the long-time limit is the thermal velocity density. The converted TeX's exponent is damaged; the PDF has the squared norm displayed here. The scalar pairing formula cannot be interpreted as a vector-norm formula: the radial moments of an isotropic Gaussian vector instead give $⟨|\mathbf{U}{|}^{2n}⟩=(2n+1)!!(s_t^2{)}^n$ in three dimensions.

Integrate the velocity once more. Interchanging the time integrals gives the integrated Ornstein-Uhlenbeck displacementFor the stated fixed initial position and velocity,Let $\mathbf{R}=\mathbf{r}-⟨\mathbf{r}⟩$ and $D=k_{B}T/\gamma$. Its component covariance comes from integrating the squared noise kernel:Therefore the total centered variance isThe mean-square displacement from ${\mathbf{r}}_0$ additionally contains $|{\mathbf{u}}_0{|}^2(1-e^{-\zeta t}{)}^2/{\zeta }^2$. Confusing this deterministic drift contribution with centered variance changes the short-time conclusion.

For times longer than the force correlation time but $\zeta t\ll 1$, the mean motion is ${\mathbf{r}}_0+{\mathbf{u}}_{0}t+O(t^2)$ and the centered trace is $2\zeta (k_{B}T/m)t^3+O(t^4)$. Thus a nonzero fixed initial velocity produces a leading ballistic mean-square displacement $|{\mathbf{u}}_0{|}^2{t}^2$; conditioned on ${\mathbf{u}}_0=0$, the white-noise displacement variance instead starts as $t^3$. If one averages over an independent equilibrium initial velocity, its covariance adds $3(k_{B}T/m)(1-e^{-\zeta t}{)}^2/{\zeta }^2$, givingThis is the usual thermal ballistic regime, before velocity memory has relaxed. At $\zeta t\gg 1$, the conditional mean approaches ${\mathbf{r}}_0+{\mathbf{u}}_0/\zeta$, while the centered trace is $6Dt-9D/\zeta +O(e^{-\zeta t})$. Its leading behavior is diffusion with $⟨|\mathbf{R}{|}^2⟩\sim 6Dt$ and $D=k_{B}T/\gamma$, as successive displacement increments lose inertial memory. At times below a finite force correlation time the white-noise short-time expansion is not applicable; for a regular force covariance and fixed initial velocity the leading noise-induced displacement variance is of order $t^4$.

The orthogonal projections of the Gaussian noise onto the column space of $X$ and its orthogonal complement are independent. Consequently ${\hat{\beta }}_0$ and the residual sum of squares are independent, andThus under the alternative,This is the noncentral t-distribution. The rejection region is fixed by the design, $\nu$ and the chosen level, while this probability distribution depends on $\beta$ only through ${\beta }_0$. Therefore the statistical power does not depend on the other regression coefficients, with the design and $\sigma$ fixed.

Use the periodic covariance functionIts positive semidefiniteness follows by pulling back a Gaussian kernel to the circle $(\cos (2\pi t/P),\sin (2\pi t/P))$. A zero Gaussian process mean is an ensemble statement, whereas an exactly zero long-term average for every periodic light curve is stronger. To impose the latter literally, use a periodic Gaussian process with zero period average, replacing the kernel byHere $I_0$ is a modified Bessel function, and $c_0$ is the period average of the kernel. Subtracting each process's random period average produces this covariance, so it remains positive semidefinite. We use $k_0$ to enforce the stated mean subtraction; using $k_P$ gives the usual ensemble-zero-mean formulation and the same likelihood derivation.

Let $K_{ij}=k_0(t_i,t_j)$ and $R=\mathrm{diag}({\sigma }_1^2,\dots ,{\sigma }_N^2)$. The prior latent vector is $f\sim N(0,K)$ and the independent Gaussian noise is $\epsilon \sim N(0,R)$, so $y=f+\epsilon \sim N(0,V)$ with $V=K+R$. The Gaussian-process marginal likelihood isPositive measurement variances make $V$ positive definite even if $K$ is singular. Maximize $\ell (\mathcal{H})=-\frac{1}{2}{y}^T{V}^{-1}y-\frac{1}{2}\log |V|-\frac{N}{2}\log (2\pi )$ over all three hyperparameters, using a broad search in $P$ to find competing aliases, followed by joint optimization. Cholesky decomposition evaluates the likelihood without explicitly forming an inverse.

For a well-resolved interior maximum, let $J=-{\nabla }^2\ell (\hat{\mathcal{H}})$ be the observed Fisher information in the coordinates $(A,\ell ,P)$. Then $\boxed{\hat{P}=(\hat{\mathcal{H}}{)}_P,{\sigma }_P\approx \sqrt{(J^{-1}{)}_{PP}}}$. Equivalently, the profile-likelihood curvature in $P$ gives the same local nuisance-adjusted variance; keeping the other hyperparameters fixed would generally underestimate it. A profile interval with ${\ell }_{\mathrm{prof}}(\hat{P})-{\ell }_{\mathrm{prof}}(P)=1/2$ is a local $1\sigma$ approximation. If the period likelihood is multimodal or poorly constrained, fit the complete posterior with proper hyperparameter priors and report its modes and marginal credible intervals rather than a single Hessian error bar. The sampling pattern can create period aliases that a single local search misses.

Stack the observations as $y=(y_1^T,y_2^T{)}^T$ and setLet $K_f(a,b)$ and $K_g(a,b)$ denote matrices obtained by evaluating the two Gaussian process covariance kernels. Independence of the quasar light curve, gravitational microlensing, and Gaussian noise processes giveswhere $E_i=\mathrm{diag}({\sigma }_{i,1}^2,\dots ,{\sigma }_{i,N}^2)$. Thus $y\mid \theta$ is a multivariate normal distribution and its Gaussian-process marginal likelihood isThe off-diagonal blocks are essential: both images contain the same delayed Ornstein-Uhlenbeck process.

The conditional mean estimator is the posterior meanprovided the first moment is finite. A maximum a posteriori estimate is any maximizer of the posterior density relative to Lebesgue measure:Existence of the posterior density alone does not guarantee a finite posterior mean or existence of a maximizing point.

For the stated multivariate normal distribution and independent standard Gaussian noise, the negative log of the posterior density, up to a constant, isIts precision matrix $A^{T}A+{\Sigma }^{-1}$ is a positive-definite matrix, so Gaussian conjugacy for a normal linear model givesThe two forms agree becauseThe covariance matrix is $(A^{T}A+{\Sigma }^{-1}{)}^{-1}$, and the maximum a posteriori estimate equals the posterior mean for this multivariate normal distribution.

Let $K\in {\mathbb{R}}^{n\times n}$ have entries $K_{ij}=\kappa (x_i,x_j)$, let $k_{\ast }\in {\mathbb{R}}^n$ have entries $(k_{\ast }{)}_i=\kappa (x_i,x_{\ast })$, and let $k_{\ast \ast }=\kappa (x_{\ast },x_{\ast })$. The Gaussian process prior and independent Gaussian noise implyApplying the conditional multivariate normal distribution gives the Gaussian process regression posteriorwhere

Spatial and chemical multisensory processing share problems of binding, uncertainty, learning and common-cause attribution. They often differ in what their inputs represent: spatial cues can estimate the same location, whereas taste and odor frequently supply complementary attributes of a food. The comparison concerns particular computations and tasks, not a rigid separation between two kinds of brain.

In spatial localization, vision, audition and touch can each provide information about position, although they begin in different coordinates. Visual information is initially related to the retina, auditory location depends on binaural and spectral cues, and touch is related to a body surface. Combining them requires accounting for eye, head and body position. Spatial coincidence supports a common event, but is not sufficient: unrelated events can occur nearby, and one moving object can produce cues at different times.

A simple reliability-weighted cue integration model makes a precise prediction. Suppose $x_1,x_2$ are unbiased, independent estimates of one position, with variances ${\sigma }_1^2,{\sigma }_2^2$. An unbiased linear estimate is $\hat{x}=w{x}_1+(1-w)x_2$, withDifferentiating and setting the derivative to zero yieldsThe second derivative is positive. The minimum-variance two-cue estimate assigns greater weight to the more precise cue, and independent information reduces variance below either individual variance. With independent Gaussian noise and a flat prior, minimizing the negative log-likelihood gives the same weighted mean. Correlated errors, bias, a nonflat prior or separate causes change this benchmark.

Marc O. Ernst and Martin S. Banks measured visual and haptic estimates of object height and compared combined judgments with predictions based on unisensory variability. Manipulating visual uncertainty and cue conflict supported reliability-dependent weighting and improved combined precision in that task.^[ref] The conclusion is conditional optimality for measured inputs and task, not that every multisensory judgment or neuron always computes an optimum.

Development supplies a useful qualification. Monica Gori and colleagues measured visual, haptic and combined judgments of block size and orientation in children. Younger children did not always obtain the adult-like precision gain; the dominant modality differed between size and orientation, and adult-like combination emerged around eight to ten years in those tests.^[ref] The authors proposed a role for cross-sensory calibration. The result does not establish one universal maturation age for all multisensory tasks. It shows why adult cue-combination rules should not simply be assigned to every developing observer.

The chemical senses pose a different representational problem. Gustation responds to dissolved chemicals at taste receptors, while olfaction responds to volatile molecules; oral touch, temperature and irritation also contribute to a food percept. Retronasal odor can be attributed to the mouth through olfactory referral, even though its receptors remain in the nose. A sweet taste and a fruit odor are not necessarily two noisy measurements of one physical sweetness coordinate. They can jointly identify a food or predict nutritional and hedonic properties. Adding them with inverse-variance weights is therefore not automatically meaningful.

Nevertheless, the same inferential questions arise: do signals belong to the same food, how should prior experience affect interpretation, and which component is responsible for a judged quality? Richard J. Stevenson and colleagues' odor–taste conditioning results show that pairing an odor with sweet or sour tastes can alter later odor judgments.^[ref] Learned co-occurrence can thus influence flavor qualities. Verbal confusion or a rating change alone does not establish full sensory fusion; separate taste, odor and pleasantness scales help test response dumping.

Neural convergence is also shared without implying identical architectures. Ivan E. T. de Araujo and colleagues delivered tastes and retronasal odors separately and together during human functional magnetic resonance imaging. They identified overlapping responses in several regions and interactions in orbitofrontal cortex; some responses also covaried with consonance or pleasantness ratings.^[ref] A mixture interaction is stronger evidence than overlap alone, but an imaging voxel does not prove that every responding neuron receives both inputs. A neural correlation with pleasantness also differs from a causal explanation of food preference.

Temporal and spatial structure further distinguish typical tasks. Spatial audiovisual cues can be brief, externally localized events; food perception unfolds with sniffing, chewing, dissolution and swallowing. Taste and retronasal odor may be bound despite unlike onset profiles because their sequence is expected within eating. Chemical signals can adapt and interact at peripheral as well as central stages, while vision and audition face their own receptor and pathway nonlinearities. Neither class is exclusively early or late, automatic or attention-dependent. Attention, task instructions, hunger and learned expectations can change what is selected or reported; these influences must be measured rather than inferred from a unitary percept alone.

A decisive comparison therefore uses matched inferential questions: estimate a common property, identify a common object, or judge whether two signals share a cause. Measure unisensory performance, test cue conflict and temporal or semantic congruence, and distinguish changes in quality, uncertainty and response. Both domains integrate evidence about events; spatial averaging is a useful special case, while complementary chemical information often requires object, quality and value representations beyond a shared coordinate.

A psychometric function specifies the probability of a measured response as a physical stimulus variable changes. The response might be “present,” “clockwise,” or correct in a two-alternative forced-choice task. These are different dependent variables. Understanding the function's shape is necessary to distinguish sensory changes from decision, attention and measurement effects. A reported threshold without its task, response definition and performance criterion is incomplete. For example, a visual experiment could use calibrated display visual contrasts, timed two-interval presentations, fixation control and button choices; the observer’s choice supplies the dependent variable.

A smooth curve need not mean that a receptor has a smoothly varying conscious threshold. Consider a noisy internal response $R=\mu (x)+\epsilon$, with “yes” reported when $R>c$. Here $c$ is a decision criterion, not necessarily a biological detection limit. If $F_{\epsilon }$ is the noise cumulative distribution function,With Gaussian noise of standard deviation $\sigma$, symmetry of the normal distribution gives a Gaussian psychometric functionThe sigmoid is therefore a distribution tail swept by a changing mean response. At low means, few responses exceed the criterion; near it, many trials switch category; at high means, almost all do. The derivative is $\varphi ((\mu -c)/\sigma ){\mu }^{\prime }(x)/\sigma$, so slope depends on both sensory gain and variability. Moving $c$ shifts the curve without necessarily improving sensitivity. If noise vanished and the mean crossed the criterion sharply, this simple model would approach a step rather than a graded transition.

There is no universal sigmoid in every stimulus coordinate. A nonlinear transduction function changes the curve, and a logarithmic intensity axis changes its appearance. As a counterexample, a toy detector responding whenever at least one event occurs in a Poisson distribution with mean $kx$ has $P(\text{yes}\mid x)=1-e^{-kx}$. This is concave on a linear positive-intensity axis. It follows immediately by subtracting the zero-event probability from one. It is a mathematical example, not evidence that human detection universally uses a one-event rule. Variation in attention or decision criterion across trials can broaden a measured curve even if the within-state sensory noise is unchanged.

Task conventions determine the endpoints. A useful psychometric function with guessing and lapses isHere $\gamma$ is the lower asymptote and $\lambda$ the upper error asymptote. In balanced two-alternative forced-choice performance, $\gamma$ is commonly fixed at $1/2$. A yes/no response probability does not automatically have this lower bound. In this particular cumulative-Gaussian parameterization, $p(\alpha )=\gamma +(1-\gamma -\lambda )/2$; thus $\alpha$ gives 75% correct only for $\gamma =1/2$ and $\lambda =0$. Other function families use different scale parameters. Nor is $\lambda$ automatically the fraction of inattentive trials: if a fraction $r$ of trials involves random two-choice guessing, those trials contribute an upper error probability $r/2$.

The relation to signal detection theory is equally important. Under independent equal-variance Gaussian signal and noise samples, define sensitivity index $d^{\prime }=({\mu }_s-{\mu }_n)/\sigma$. In an ideal two-interval comparison, the response difference has mean ${\mu }_s-{\mu }_n$ and variance $2{\sigma }^2$, givingThis Gaussian forced-choice sensitivity relation follows by standardizing the Gaussian difference. It is a specified decision model, not an identity for every forced-choice experiment. In a yes/no task, hits and false alarms are both needed: under the corresponding equal-variance model, $d^{\prime }={\Phi }^{-1}(H)-{\Phi }^{-1}(F)$. A criterion change can increase both rates while leaving $d^{\prime }$ unchanged. The phenomenological fit above need not have parameters identical to these internal-response parameters.

These distinctions affect experiments. Felix A. Wichmann and N. Jeremy Hill used simulation and statistical modeling to show that unmodeled stimulus-independent errors can bias threshold and slope estimates, even when lapses are uncommon.^[ref] Ignoring high-intensity errors can make a function appear shallower and falsely suggest worse sensory precision. A broad group-average curve can likewise reflect a mixture of individually sharp functions with different locations.

At predetermined stimulus levels, independent trials give counts $k_i\sim \mathrm{Binomial}(n_i,p(x_i))$. The parameter-dependent log-likelihood is ${\sum }_i[k_i\log p_i+(n_i-k_i)\log (1-p_i)]$. Fitting this binomial likelihood, checking goodness of fit and quantifying uncertainty are preferable to drawing a curve through noisy proportions. Easy trials help constrain lapses; levels near the transition constrain location and slope. Randomization, stable calibration and checks for learning or fatigue help distinguish a stationary sensory relation from drift. Adaptive procedures require inference appropriate to their sampling design.

Thus the function is both a description and a testable model. Its shape determines what a threshold means, how an experiment should sample, and which claims about sensory mechanisms are warranted.