Responses and priors¶

A kernel describes a curve; Response determines what is learned and shared. Read the illustrated kernel guide before selecting a shape. Normal and KernelPrior are R penalties; GP uses Prior inside BayesianPriors instead.

Common kernel operations¶

Finite response families inherit these operations from the kernel base class:

SampledKernel has no free scalar parameters; its supplied arrays describe a fixed curve. Response.free_parameters, parameter_bounds() and support_envelope() expose the resolved learning specification.

Response¶

from multimodalsrm import Response · Source

Response(
    kernel,
    pooling="partial",
    estimate=True,
    fixed=None,
    bounds=None,
    pooling_strength=1.0,
    lag_prior=None,
    prior=None,
)

Configure response sharing and which kernel parameters are learned.

estimate=True learns parameters not listed in fixed. For example, a Gaussian with fixed={"width": 0.4} learns only its lag, while omitting fixed learns both width and lag. estimate=False fixes the full response. pooling="shared" shares a modality's response across people.

Response.lag_only¶

Response.lag_only(
    kernel,
    *,
    pooling="partial",
    bounds=None,
    pooling_strength=1.0,
    lag_prior=None,
)

Compatibility convenience that fixes every non-lag parameter.

Prefer Response(kernel, estimate=True, fixed={...}) in new examples to make the selected parameters explicit.

Response.initial_kernel¶

Response.initial_kernel()

Response.free_parameters¶

Response.free_parameters (property)

Response.parameter_bounds¶

Response.parameter_bounds()

Response.support_envelope¶

Response.support_envelope()

Fixed union over bounded family extremes, independent of fit iterates.

Response.metadata¶

Response.metadata (property)

Identity¶

from multimodalsrm import Identity · Source

Identity()

Identity.evaluate¶

Identity.evaluate(lags)

Gaussian¶

from multimodalsrm import Gaussian · Source

Gaussian(width=1.0, lag=0.0)

Gaussian.support¶

Gaussian.support (property)

Gamma¶

from multimodalsrm import Gamma · Source

Gamma(shape=3.0, scale=1.0, lag=0.0)

Gamma.support¶

Gamma.support (property)

DoubleGamma¶

from multimodalsrm import DoubleGamma · Source

DoubleGamma(
    peak_shape=6.0,
    peak_scale=1.0,
    undershoot_shape=16.0,
    undershoot_scale=1.0,
    undershoot_ratio=1 / 6,
    lag=0.0,
)

DoubleGamma.support¶

DoubleGamma.support (property)

BatemanSCR¶

from multimodalsrm import BatemanSCR · Source

BatemanSCR(rise=0.7, decay=3.0, lag=0.0)

Causal SCR with two exponential stages and finite continuous L2 scale.

rise and decay are positive time constants in seconds; lag shifts the onset. Equal constants are supported. Exchanging constants leaves the response unchanged: use separated bounds to identify labels. The response ends 90 seconds after lag and uses continuous finite L2 normalization.

BatemanSCR.support¶

BatemanSCR.support (property)

SampledKernel¶

from multimodalsrm import SampledKernel · Source

SampledKernel(lags, values)

SampledKernel.support¶

SampledKernel.support (property)

Normal¶

from multimodalsrm import Normal · Source

Normal(mean, sd)

KernelPrior¶

from multimodalsrm import KernelPrior · Source

KernelPrior(reference, strength)

Prior¶

from multimodalsrm.bayesian import Prior · Source

Prior(family, loc=0.0, scale=1.0, lower=-np.inf, upper=np.inf)

Normal or lognormal prior, optionally truncated, or a finite uniform.

For lognormal priors, loc and scale describe log(parameter). The density itself is with respect to the physical parameter. bounded truncates and renormalizes; it does not add a second penalty.

Prior.normal¶

Prior.normal(mean=0.0, sd=1.0)

Prior.lognormal¶

Prior.lognormal(log_mean, log_sd)

Prior.uniform¶

Prior.uniform(lower, upper)

Prior.bounded¶

Prior.bounded(lower, upper)

Prior.ppf¶

Prior.ppf(probability)

Prior quantiles for independent, reproducible initial designs.

Prior.distribution¶

Prior.distribution()

NumPyro distribution; also defines the MAP log density.

BayesianPriors¶

from multimodalsrm.bayesian import BayesianPriors · Source

BayesianPriors(noise, loading_sd=1.0, offset_sd=1.0, filters=dict())

Priors in supplied data units; filter priors map modality then parameter.

noise is a prior on variance, not on standard deviation or log variance. Loading and offset scales are illustrative defaults; applications should set all scales from their measurement units and prior predictive checks.