Abstract Details
| Error Analysis in XAS: Which Method for Which Situation? | |
|---|---|
| Abstract ID | W_EXAFS-05 |
| Presenter | Emmanuel Curis |
| Presentation Type | EXAFS Workshop - SSRL |
| Full Author List | E. Curis (1) |
| Affiliations | (1) Laboratoire de Biomathématiques, EA 2498, Faculté de Pharmacie, Université Paris Descartes |
| Category | Instrumentation/Development |
| Abstract | The aim of this presentation will be to make an overview of the
various methods available to estimate the uncertainties in XAS
analysis and to give hints to select one or the other of these methods
according to the situation faced to. After a short classification of
the various experimental errors, methods for handling statistic errors
will be presented, then methods for systematic errors.
Errors in XAS experiments can come from many sources: random character of the photon absorption/emission process, departure from ideal experimental conditions (monochromatic beam, homogeneous sample of constant thickness, linear detectors,...), theoretical approximations, data analysis distortions,... These errors can be roughly classified in two great families: statistic and systematic errors, based on their time scale; handling of these errors uses completely different procedures. As far as statistic errors are considered, every single experimental point is modeled as a centered random variable added to the ideal value. The aim is to determine the variance of this random variable and to use it to obtain variances of the estimated parameters, also seen as random variables. Despite various methods exist when only a single spectrum is available, most efficient methods rely on the average of several spectra. Once variances ("error bars") are known for each point of the experimental spectrum, they are used in fitting to estimate error bars on the parameters. Mainly 3 techniques are used for that, that will be presented and compared: Hessian inversion, "brute-force" and Monte-Carlo. Dealing with systematic errors is much more difficult, since there is no universal model for that. Strong systematic errors can be directly corrected or, in the other hand, can make the data useless. However, estimate the effect of small, unknown, systematic errors on fitted parameters values is a challenge. An approach based on modeling the experimental conditions will be discussed; basically, it relies on the assumption that distortions of the experimental data observed when simulating "real", non-ideal, experimental conditions are comparable to the distortion between the experimental data and the "real" value. Finally, the link between the two approaches will be made by the mean of the various interpretations that can be made of the Monte-Carlo analysis results — and of the method itself! |
| Footnotes | |
| Funding Acknowledgement | |