The background ladder: which stage removes what

Several different controls across Asymmetry carry the word “background”, and they act at different points in the pipeline on different representations of the data. They are not alternatives to one another and they do not stack into a double subtraction — each removes a different manifestation of the same underlying quantity, at the stage where that manifestation appears. This page is the map: what each stage removes, where it sits in the chain, and when to turn it on. For the mechanics of any one stage, follow the cross-reference.

The physical quantity underneath most of these is a single thing: the steady, uncorrelated count rate — particles that reach a detector without being correlated with a muon decaying in the sample (decay positrons from muons stopped upstream, room background, detector noise that survives the coincidence logic). At a continuous source the rate is significant; at a pulsed source the beam duty factor suppresses it to nearly nothing (Blundell, De Renzi, Lancaster & Pratt, Muon Spectroscopy: An Introduction). One flat rate in the raw histograms shows up as a baseline in the counts, a zero-frequency feature in the FFT, a central peak in a zero-field MaxEnt distribution, and a pedestal under the spectrum — so it is handled, where it matters, at each of those stages.

The ladder at a glance

Stage	Operates on	Removes	When
Grouping background mode (backgrounds)	raw counts, pre-asymmetry / pre-FFT	the steady flat rate before the asymmetry ratio (one of four mutually exclusive modes, applied once)	enable per dataset at reduction time; feeds every downstream analysis
Count-fit background nuisance (Count-domain fitting (α calibration & single histograms))	raw counts, in the fit	a flat rate fitted jointly with the count model — independent of the grouping correction	when fitting raw counts and the grouping rate is not already trustworthy
σ-clip baseline (Robust baseline offset)	spectrum, post-FFT, display	the spectral noise floor (a positive pedestal of redistributed counting noise)	when measuring peak heights or areas above a true zero
MaxEnt SpecBG (Fourier analysis)	spectrum, display, ZF/LF	the zero-frequency central peak of a ZF/LF field distribution	ZF/LF field distributions with weak satellite structure buried under the central peak

Stage 1 — the grouping background mode (pre-asymmetry, pre-FFT)

The grouping dialog’s Background selector subtracts the steady flat rate from the grouped forward/backward counts before the asymmetry ratio is formed. It offers four modes — Fixed values, Range average (pre-t0), Tail fit (late-time), and Background run — and they are mutually exclusive: exactly one applies to a given dataset, through a single correction chokepoint, and it is applied once. See Background correction for how to choose among the four and for the mathematics of each.

This is the only background stage that reaches every downstream analysis. The asymmetry curve, the α estimators, the per-group time-domain fits and the Fourier input all consume the same corrected group sums. In particular, the FFT input is rebuilt from the current grouping, so the grouping background mode applies to the transform too — there is no separate FFT-only background control, and no second subtraction (see the FFT background discussion in Fourier analysis and the Fourier panel’s Background status line). A flat rate matters most here, because the grouped FFT signal is lifetime-corrected by \(e^{t/\tau_\mu}\) before the transform: an unsubtracted constant becomes a growing ramp that dumps spurious power into the low-frequency bins.

When to enable. Set a grouping background mode whenever the uncorrelated rate is non-negligible — routinely for continuous-source data (Range average), and for pulsed data analysed to long times (Tail fit), where even the small residual rate biases weak late-time relaxation.

Stage 2 — the count-fit background nuisance (in the fit, on raw counts)

Count-domain fitting (Count-domain fitting (α calibration & single histograms)) fits a flat background term bg jointly with the count model \(N_0 e^{-t/\tau_\mu}[1 + s A P(t)] + \mathrm{bg}\). This term is fitted against the raw counts.

Warning

The grouping background correction never reaches the count fit. The count fit always consumes raw histograms with deadtime applied — not the background-corrected group sums that the asymmetry and Fourier paths use. A user who believes the grouping correction has already removed the flat rate, and therefore fixes bg = 0, will bias \(N_0\) and α: the raw counts still contain the full background. Let bg float (or seed it from the grouping value) — do not assume Stage 1 has done the job for the count fit.

This is not a double subtraction; it is the opposite trap — a stage that looks downstream of the grouping correction but is not. The two are independent measurements of the same flat rate: Stage 1 stores it in the grouping and applies it to the reduced data; Stage 2 measures it from the raw counts the fit actually sees.

When to enable. Whenever you fit raw counts and want the background determined from the same data and weighting as the rest of the count model — the statistically clean route when a separate grouping estimate is unavailable or untrusted. The fitted value is a legitimate measurement of the steady rate and can be promoted into the grouping as a Fixed value (see Count-domain fitting (α calibration & single histograms)).

Stage 3 — the σ-clip baseline (post-FFT, display)

A power or magnitude spectrum sits on a positive pedestal from the redistributed counting noise — distinct from the flat time-domain rate of Stages 1–2. The Fourier panel’s robust σ-clip baseline estimates that pedestal and subtracts it from the displayed spectrum, leaving sharp peaks intact (they are rejected as outliers during the estimate). It is a display-channel operation: it never alters the underlying transform, and its converged width doubles as the noise estimate behind the signal-to-noise readout. See Robust baseline offset in Frequency-domain conditioning for the σ-clip mechanism and its pitfalls.

When to enable. Whenever peak heights or areas must be measured above a true zero — comparing intensities across runs, integrating a line. It is orthogonal to Stages 1–2: subtracting the time-domain rate flattens the low-frequency ramp, while the σ-clip baseline removes the spectral noise floor that remains.

Stage 4 — MaxEnt SpecBG (post-reconstruction, display, ZF/LF)

In zero or longitudinal field the MaxEnt reconstruction returns a field distribution \(p(B)\) dominated by a strong zero-frequency central peak, which can bury weak satellite structure. The Zero-frequency background (SpecBG) control (ZF/LF mode only) subtracts a zero-centred pseudo-Voigt model of that central peak from the displayed spectrum. Like the σ-clip baseline it is display-only and never alters the reconstructed spectrum; unlike it, it targets one specific feature — the ZF/LF central peak — rather than a broadband pedestal. See Fourier analysis for the SpecBG controls.

When to enable. Only for ZF/LF MaxEnt field distributions where a central peak is hiding the structure you care about. It is unrelated to the transverse-field precession lines that the other stages serve.

Why they coexist

Because each stage acts on a different representation — raw counts, the fitted count model, the spectral display channel, the reconstructed distribution — turning on more than one does not subtract the same rate twice. The worst case is mild redundancy: a pre-FFT-subtracted flat rate leaves a smaller zero-frequency feature for SpecBG to model, not a negative artefact. The one genuine trap is the count-fit asymmetry above — the grouping correction does not propagate into the count fit, so its background must be handled there in its own right.