Estimate a time series Signal to Noise Ratio (SNR) given total and noise variance components and (optionally) a background noise model.

Given time series variance components (total variance, raw noise variance), compute the final noise, signal, and SNR. For noise estimation use one of:

A raw noise variance e.g the result of estimateNoise
A background noise model based on linear predictors
A flooring rule: final noise = max(raw, background)

The background noise model is a general linear noise baseline: baseline = sum_i beta[i] * feat[i].

Usage

estimateSNR(totalVar, noiseRaw, eps, betas, features, noise_mode = "floor")

Arguments

totalVar: Total time series variance. Typically calculated with estimateNoise
noiseRaw: noise variance estimate. Typically calculated with estimateNoise
eps: Numeric. Lower bound applied to both noise and signal.
betas, features: Coefficients (\(betas\)) and predictors (\(features\)) for the background noise model. Ignored when noise_mode = "raw". Both vectors must have the same length when used. The predictors form a design vector: include a constant 1 if the model has an intercept. For example, for \(baseline = b0 + b1 * mean\), use betas = c(b0, b1), features = c(1, mean).
noise_mode: Character scalar: one of "raw", "model", "floor". - "raw" -> use noise_raw - "model" -> use baseline = sum(betas * features) - "floor" -> use max(noise_raw, baseline)

Value

Named numeric vector with elements: - snr (log2 scale) - signal - noise final estimate used for signal and snr calculation - noise baseline (sum(betas * features); NA if not applied) - noise raw

Examples

estimateSNR(totalVar = 1.0, noiseRaw = 1.0, eps = 0.01,
            betas = c(0.5, 2.0), features = c(2.0, 4.5),
            noise_mode = "model")
#>       snr    signal     noise  baseline       raw 
#> -9.965784  0.010000 10.000000 10.000000  1.000000