The words of this tutorial, step by step. The text lives in src/smappy/tutorial/topics/; the directions it follows are in src/smappy/tutorial/STYLE.md.
1. card: Measuring: statistics, precision and line profiles
How bright are the molecules, and how long do they stay on? How precisely were they placed, and what can the picture resolve? What shape is under a line?
One plugin for each, on the simulated data.
spoken: Three questions, and a plugin for each: how the molecules behave, how precise the data is, and what shape a structure has.
2. Localization Statistics draws each distribution with the law it should follow, so a glance says whether the data behaves.
3. Photons fall off exponentially above the detection threshold. The fitted constant is the mean photon count of a blink.
4. The precision follows from the photons. Its maximum and rising edge are landmarks you can read off the histogram.
5. The on-time, from the grouping: how many frames a molecule stays on. A straight line on this log scale is a constant switching rate.
6. card: Three measures of precision
CRLB: what the fit believes, from the photons and the background.
Pair displacement (NeNA): a molecule seen in two consecutive frames moved only by the error, so the data measures itself.
FRC: what the picture resolves, labelling and drift included.
spoken: Three measures: what the fit believes, what the data shows, and what the picture resolves. They are meant to disagree.
7. This data is so precise that FRC needs a finer pixel than it picks by itself: set it under more.
8. The pair displacement agrees with the CRLB, as it should for simulated data. In a real experiment it is larger: drift, or a PSF that is not the model.
9. FRC splits the data in two and asks up to which detail the halves agree: the resolution of the picture.
10. The CRLB histogram, with the same law as in Statistics fitted to it.
11. When the shape under a line is not known, fit all five models: a Gaussian, two Gaussians, a step, a disk and a ring.
12. Each is fitted to the localizations themselves, and compared by AIC: the lowest wins, and extra parameters have to earn their place.
13. With few localizations, bootstrap gives confidence intervals that do not rely on the fit's own error bars.
14. card: What to remember
- Statistics: photons, precision and on-time, each against its law.
- Precision: CRLB, pair displacement and FRC; where they disagree says what limits the data.
- Line Profile: compare models by AIC when the shape is not known.
spoken: That is measuring. Next: correcting drift, before any of these numbers are trusted.