bm_logsumexp: Mapper for log-sum-exp aggregation

In inlabru: Bayesian Latent Gaussian Modelling using INLA and Extensions

Mapper for log-sum-exp aggregation

Description

Constructs a mapper that aggregates elements of exp(state), with optional non-negative weighting, and then takes the log(), so it can be used e.g. for v_k=\log[\sum_{i\in I_k} w_i \exp(u_i)] and v_k=\log[\sum_{i\in I_k} w_i \exp(u_i) / \sum_{i\in I_k} w_i] calculations. Relies on the input handling methods for bm_aggregate, but also allows the weights to be supplied on a logarithmic scale as log_weights. To avoid numerical overflow, it uses the common method of internally shifting the state blockwise; v_k=s_k+\log[\sum_{i\in I_k} \exp(u_i + \log(w_i)- s_k)] , where s_k=\max_{i\in I_k} u_i + \log(w_i) is the shift for block k.

Usage

bm_logsumexp(rescale = FALSE, n_block = NULL)

bru_mapper_logsumexp(...)

## S3 method for class 'bm_logsumexp'
ibm_jacobian(mapper, input, state = NULL, ...)

## S3 method for class 'bm_logsumexp'
ibm_eval(mapper, input, state = NULL, log = TRUE, ..., sub_lin = NULL)

Arguments

rescale	logical; For bm_aggregate and bm_logsumexp, specifies if the blockwise sums should be normalised by the blockwise weight sums or not: FALSE: (default) Straight weighted sum, no rescaling. TRUE: Divide by the sum of the weight values within each block. This is useful for integration averages, when the given weights are plain integration weights. If the weights are NULL or all ones, this is the same as dividing by the number of entries in each block.
n_block	Predetermined number of output blocks. If NULL, overrides the maximum block index in the inputs. The priority order is input$n_block, the mapper definition n_block, then max(input$block).
...	Arguments passed on to other methods
mapper	A mapper S3 object, inheriting from bru_mapper.
input	Data input for the mapper.
state	A vector of latent state values for the mapping, of length ibm_n(mapper, inla_f = FALSE)
log	logical; control log output. Default TRUE, see the ibm_eval() details for logsumexp mappers.
sub_lin	Internal, optional pre-computed sub-mapper information

Methods (by generic)

• ibm_jacobian(bm_logsumexp): input should be a list with elements block and weights. block should be a vector of the same length as the state, or NULL, with NULL equivalent to all-1. If weights is NULL, it's interpreted as all-1.

• ibm_eval(bm_logsumexp): When log is TRUE (default), ibm_eval() for logsumexp returns the log-sum-weight-exp value. If FALSE, the sum-weight-exp value is returned.

See Also

bru_mapper, bru_mapper_generics

Other mappers: bm_aggregate(), bm_collect(), bm_const(), bm_factor(), bm_fmesher(), bm_harmonics(), bm_index(), bm_linear(), bm_marginal(), bm_matrix(), bm_mesh_B(), bm_multi(), bm_pipe(), bm_repeat(), bm_scale(), bm_shift(), bm_sum(), bm_taylor(), bru_get_mapper(), bru_mapper(), bru_mapper.fm_mesh_1d(), bru_mapper.fm_mesh_2d(), bru_mapper_generics

Examples

m <- bm_logsumexp()
ibm_eval2(m, list(block = c(1, 2, 1, 2), weights = 1:4), 11:14)
ibm_eval2(m, list(block = c(1, 2, 1, 2), weights = 1:4, n_block = 3), 11:14)

inlabru documentation built on Aug. 9, 2025, 1:08 a.m.