Maximum Entropy and Bayesian Methods : by G. J. Erickson, J. T. Rychert and C. R. Smith. Kluwer...

4
Book Reviews Maximum Entropy and Bayesian Methods by G. J. Erickson, J. T. Rychert and C. R. Smith. Kluwer Academic Publishers, Boise, Idaho, U.S.A. 1997. ix 297 pp. £84·00. ISBN 07923 5047 2. This book is the latest volume of proceedings arising from a series of conferences on the use of maximum entropy and Bayesian methods in science, engineering, medicine and economics. For the past 18 years the conference series has followed a self-consciously independent path under the intellectual leadership of Edwin Jaynes, a Professor of Physics at Washington University, St Louis, Missouri, and the ‘inventor’ of the maximum entropy method. It is therefore fitting that this book is dedicated to the memory of Ed Jaynes, who died on 30 April 1998 (biographical details of Ed Jaynes and copies of many of his scientific papers can be found at http://bayes.wustl.edu/). In common with other volumes in this series the pro- ceedings are a mixed bag of theoretical, historical, philo- sophical and applied papers, all more or less imbued with a Bayesian approach to inference. The Preface draws attention to the breadth of the volume by listing some of the keywords used. These include deconvolution, inverse problems, point- spread function, model comparison, multisensor data fusion, image processing, tomography, reconstruction, deformable models, pattern recognition, classification and group analy- sis, segmentation, complexity, algorithms, Ockham’s razor and symmetry. In my opinion, the first paper in the volume is the only really essential one. Here, John Skilling provides a concise overview of the relationship between the quantified maxi- mum entropy (QME) and massive inference techniques for spectral reconstruction and deconvolution. Until recently the QME approach, epitomised in Skilling’s MEMSYS5 computer code, was arguably the best spectral deconvolution method available. However, despite its considerable success, the QME approach had a number of inherent and fundamental difficulties. These included a problem of pixelation and the use of a multidimensional Gaussian approximation in calcu- lating error bars. Recently, a completely new Bayesian tech- nique has been developed that outclasses QME in a number of important theoretical and practical ways. The new tech- nique, developed primarily by Sibusiso Sibisi in collabora- tion with John Skilling, has been christened ‘massive inference’ (MI); readers should see Sibisi & Skilling (1997) for details. In short, the MI method is a general Bayesian prior for positive, distributed quantities, such as light intensity across an image or flux of energy along a spectrum. The method solves some of the theoretical problems inherent in QME and also produces visually and quantitatively superior reconstructions. The novelty of the MI method is both in its generality and its computability. When the MI prior and a Markov chain Monte-Carlo (MCMC) method are combined, multidimensional inference problems previously thought intractable can be performed robustly and quickly. An outstanding example of the MI method in action is the work being carried out by Ian Archibald and co-workers at the Shell Research laboratory near Chester. This group uses MI to make three-dimen- sional reconstructions of the location and intensity of leaking gas sources on large scale industrial plant from open-path (i.e. line integral) concentration measurements. Other recent applications of MI include non-parametric density estimation (Sibisi & Skilling, 1997) and the analysis of inductively coupled plasma-mass spectrometry data (Sharp et al., 1999). A general, non-mathematical, intro- duction to massive inference has recently been published by Skilling (1998). The remainder of the book is to my mind of lesser importance than the first paper. Both the quality and relevance of the papers to applied workers is also quite variable. I personally found several papers very interesting. These include a paper by Rodriguez on ‘Non-parametric cross validated analysis using MCMC’, a paper describing the use of MaxEnt as a way of deciding which computer algorithms are feasible, by Cooke, Kreinovich and Longpre ´, and a num- ber of papers by Anthony Garrett on the historical and logical roots of probability theory including novel derivations of the well known sum and product rules. More applied papers include the analysis of energy confinement data from large fusion experiments, by Preuss, Dose and von der Linden and a Bayesian approach to determining charge density from elastic electron scattering data, by Mohammad-Djafari and Miller. A final example, also indicating the breadth of the volume, is provided by Osegueda et al., who report on their use of MaxEnt and symmetry arguments to design an optimal network of sensors for non-destructive testing in the aerospace industry. I found the index rather skimpy considering the breadth of the topics addressed in the component papers. Unfortu- nately there are very few common references that could be used to gain an overall background to the material in the book and a reader interested in a specific subject would need to delve into the reference list of a particular paper to Journal of Microscopy, Vol. 196, Pt 3, December 1999, pp. 352–355. q 1999 The Royal Microscopical Society 352

Transcript of Maximum Entropy and Bayesian Methods : by G. J. Erickson, J. T. Rychert and C. R. Smith. Kluwer...

Page 1: Maximum Entropy and Bayesian Methods : by G. J. Erickson, J. T. Rychert and C. R. Smith. Kluwer Academic Publishers, Boise, Idaho, U.S.A. 1997. ix + 297 pp. £84.00. ISBN 07923

Book Reviews

Maximum Entropy and Bayesian Methods by G. J. Erickson,

J. T. Rychert and C. R. Smith. Kluwer Academic Publishers,

Boise, Idaho, U.S.A. 1997. ix�297 pp. £84´00. ISBN 07923

5047 2.

This book is the latest volume of proceedings arising from a

series of conferences on the use of maximum entropy and

Bayesian methods in science, engineering, medicine and

economics. For the past 18 years the conference series has

followed a self-consciously independent path under the

intellectual leadership of Edwin Jaynes, a Professor of Physics

at Washington University, St Louis, Missouri, and the

`inventor' of the maximum entropy method. It is therefore

®tting that this book is dedicated to the memory of Ed

Jaynes, who died on 30 April 1998 (biographical details of

Ed Jaynes and copies of many of his scienti®c papers can be

found at http://bayes.wustl.edu/).

In common with other volumes in this series the pro-

ceedings are a mixed bag of theoretical, historical, philo-

sophical and applied papers, all more or less imbued with a

Bayesian approach to inference. The Preface draws attention

to the breadth of the volume by listing some of the keywords

used. These include deconvolution, inverse problems, point-

spread function, model comparison, multisensor data fusion,

image processing, tomography, reconstruction, deformable

models, pattern recognition, classi®cation and group analy-

sis, segmentation, complexity, algorithms, Ockham's razor

and symmetry.

In my opinion, the ®rst paper in the volume is the only

really essential one. Here, John Skilling provides a concise

overview of the relationship between the quanti®ed maxi-

mum entropy (QME) and massive inference techniques for

spectral reconstruction and deconvolution. Until recently the

QME approach, epitomised in Skilling's MEMSYS5 computer

code, was arguably the best spectral deconvolution method

available. However, despite its considerable success, the QME

approach had a number of inherent and fundamental

dif®culties. These included a problem of pixelation and the

use of a multidimensional Gaussian approximation in calcu-

lating error bars. Recently, a completely new Bayesian tech-

nique has been developed that outclasses QME in a number

of important theoretical and practical ways. The new tech-

nique, developed primarily by Sibusiso Sibisi in collabora-

tion with John Skilling, has been christened `massive

inference' (MI); readers should see Sibisi & Skilling (1997)

for details. In short, the MI method is a general Bayesian

prior for positive, distributed quantities, such as light

intensity across an image or ¯ux of energy along a

spectrum. The method solves some of the theoretical

problems inherent in QME and also produces visually and

quantitatively superior reconstructions. The novelty of the

MI method is both in its generality and its computability.

When the MI prior and a Markov chain Monte-Carlo

(MCMC) method are combined, multidimensional inference

problems previously thought intractable can be performed

robustly and quickly. An outstanding example of the MI

method in action is the work being carried out by Ian

Archibald and co-workers at the Shell Research laboratory

near Chester. This group uses MI to make three-dimen-

sional reconstructions of the location and intensity of

leaking gas sources on large scale industrial plant from

open-path (i.e. line integral) concentration measurements.

Other recent applications of MI include non-parametric

density estimation (Sibisi & Skilling, 1997) and the analysis

of inductively coupled plasma-mass spectrometry data

(Sharp et al., 1999). A general, non-mathematical, intro-

duction to massive inference has recently been published by

Skilling (1998).

The remainder of the book is to my mind of lesser

importance than the ®rst paper. Both the quality and

relevance of the papers to applied workers is also quite

variable. I personally found several papers very interesting.

These include a paper by Rodriguez on `Non-parametric cross

validated analysis using MCMC', a paper describing the use of

MaxEnt as a way of deciding which computer algorithms

are feasible, by Cooke, Kreinovich and LongpreÂ, and a num-

ber of papers by Anthony Garrett on the historical and

logical roots of probability theory including novel derivations

of the well known sum and product rules. More applied

papers include the analysis of energy con®nement data from

large fusion experiments, by Preuss, Dose and von der Linden

and a Bayesian approach to determining charge density

from elastic electron scattering data, by Mohammad-Djafari

and Miller. A ®nal example, also indicating the breadth of

the volume, is provided by Osegueda et al., who report on

their use of MaxEnt and symmetry arguments to design an

optimal network of sensors for non-destructive testing in

the aerospace industry.

I found the index rather skimpy considering the breadth

of the topics addressed in the component papers. Unfortu-

nately there are very few common references that could be

used to gain an overall background to the material in the

book and a reader interested in a speci®c subject would

need to delve into the reference list of a particular paper to

Journal of Microscopy, Vol. 196, Pt 3, December 1999, pp. 352±355.

q 1999 The Royal Microscopical Society352

Page 2: Maximum Entropy and Bayesian Methods : by G. J. Erickson, J. T. Rychert and C. R. Smith. Kluwer Academic Publishers, Boise, Idaho, U.S.A. 1997. ix + 297 pp. £84.00. ISBN 07923

proceed further. For these reasons scientists and statisticians

seeking an introductory text on the MaxEnt method would

get far more from the collection edited by Buck & Macaulay

(1991). In my opinion it is also questionable whether the

volume is of general enough interest to be recommended for

a library purchase. Despite these reservations, those trying

to solve otherwise dif®cult problems involving the recon-

struction of a distributed quantity or those with an ongoing

interest in the latest developments in the MaxEnt, massive

inference and Bayesian community will ®nd the volume

essential reading.

M AT T R E E D

References

Buck, B. & Macaulay, V.A. (1991) Maximum Entropy in Action.

Clarendon Press, Oxford.

Sharp, B.L., Batey, J., Begley, I.S., Gregson, D., Skilling, S., Sulaiman,

A.B. & Verbogt, G. (1999) Information retrieval from the

inductively coupled plasma. J. Anal. At. Spectrom. 14, 99±108.

Sibisi, S. & Skilling, J. (1997) Prior distributions on measure space.

J. R. Statist. Soc. B, 59, 217±235.

Skilling, J. (1998) Probabalistic data analysis: an introductory

guide. J. Microsc. 190, 28±36.

Biological Specimen Preparation for Transmission Electron

Microscopy by Audrey M. Glauert & Peter R. Lewis.

Practical Methods in Electron Microscopy Vol. 17. Series

Editor Audrey Glauert. Portland Press, 1998. Softback,

£39.50, ISBN 185578 060 7, hardback, £110, ISBN

185578 061 5. 326 pages. Co-published by Princeton

University Press, http://pup.princeton.edu

Audrey Glauert has contributed greatly to the ®eld of

electron microscopy with a distinguished series of books she

has either written or edited. These books have, for the most

part, rightly earned their place in EM laboratories and

libraries all over the world. I must confess some disappoint-

ment when I read this book, although many will doubtless

®nd it useful in some respects. Even upon browsing through

its contents I was shocked to see that the words `cryo

electron microscopy' are nowhere to be seen! As much of

today's state of the art EM involves rapid freezing and

vitri®cation (a term incorrectly referred to and only in

passing as `microcrystalline ice'), this is quite an omission,

given the title. The authors actually start the preface by

stating that `most biological material can be examined in

the transmission electron microscope only in the form of

ultra thin resin sections'. This statement ignores all the

work on isolated particles that is going on in

many laboratories around the world. The authors

should, in retrospect, have entitled this book `Preparation

methods for resin embedding' or words to that effect. It is

therefore only subjects covered by such a title that I shall try

to give my opinion on as I discuss the contents of this

handbook.

The book deals with plastic resin methods that can be used

for preparing thin sections for EM and there is a wealth of

information on this topic. A lot, perhaps most, of this

information is already available in the earlier books in this

series and many parts, such as that on sectioning, simply refer

to one of those. There is a detailed coverage of the Lowicryl

and other resins which were not covered previously in this

series. Newcomers to EM and even experienced technicians

would do well to read the extensive parts of this book

covering laboratory safety, a daily concern in all EM labs.

It is accepted at the outset that most EM labs around the

world spend the predominant part of their time preparing

plastic sections. However, it is inconceivable to me that the

authors hardly mention the negative staining approaches!

This is a great pity because, in contrast to cryo EM, this is a

technique whose essence can be learnt in minutes and

would be used much more extensively by many cell biology

labs, especially for immunolabelling, if it was advertised

more widely.

It must be admitted that the processes starting with

chemical ®xation of a living cell and culminating in the

visualization of a stained thin section in the EM are

exceedingly complex. It is virtually impossible at present

to know what happens at the cellular, organellar and

molecular levels when a ®xative enters cells; it is even more

of a problem to imagine what happens to these structures

when they are subsequently osmicated, dehydrated, em-

bedded, sectioned and exposed to the electron beam.

Consequently, we are all at the mercy of empiricism: if an

EM technique works adequately for our purposes we use it,

even if we do not understand how it works. However, when

a technique has been widely used for many years we can

learn something from surveying the literature and trying to

come up with some consensus about the best approach, and

perhaps even about how or why the method works. My

main criticism of this book is that that there are far too

many statements that are not substantiated by references to

publications that support those statements. I would be

happy to read that `in our experience such-and-such an

approach works well', even if the theoretical basis was

unclear. However, statements are made throughout the

book that make me ask, how do they know? In the chapter

on ®xation, for example, there are a number of unsub-

stantiated statements with which I tend to disagree but

which I could have checked by reading the published

papers, had they been cited. For example:

1 `glutaraldehyde . . . stabilizes structures by cross-linking

before there is any opportunity for extraction by the buffer

to occur'.

2 `formaldehyde, a mono-aldehyde, is less likely than

glutaraldehyde to introduce extra free aldehyde groups

into the tissue'.

BOOK REVIEWS 353

q 1999 The Royal Microscopical Society, Journal of Microscopy, 196, 352±355

Page 3: Maximum Entropy and Bayesian Methods : by G. J. Erickson, J. T. Rychert and C. R. Smith. Kluwer Academic Publishers, Boise, Idaho, U.S.A. 1997. ix + 297 pp. £84.00. ISBN 07923

3 Acrolein is a lipid ®xative but glutaraldehyde and

formaldehyde are not'.

4 `the polar Lowicryl K4M always extracts more lipid than

the non-polar HM2O'

This book has plenty of recipes and many are well-tried

and tested. One has to work hard, however, to separate the

`wheat' from the `chaff ', because well-known recipes (such

as that for phosphate buffer) are mixed with obscure

techniques, in many cases long relegated to obscurity. I

would have appreciated more of the authors' opinions on

some of these methods, some guidance concerning their

ef®cacy. For example, the authors cite Idelman (1964) and

Stein & Stein (1971), who developed the approach of going

directly from 70% ethanol to the pure resin without using

the 100% step in order to extract less lipids. But on page

134 the authors warn us of the danger of incomplete

dehydration, a danger of which most experienced people in

EM labs are probably aware. This apparent contradiction is

not commented upon.

A lack of critical evaluation can also be seen in the

discussion of the various recipes, although the overall

description of the recipes is usually adequate for a novice to

be able to follow them. My gripe here again concerns

empiricism, combined with lack of substantiation by

references. I am sceptical, for example, of the authors'

repeated claim throughout book that calcium should be

added to (essentially) all ®xatives. If adding calcium is truly

an advantage they should at least cite papers supporting

that claim. I can intuitively accept that for some situations

adding extracellular (mM) levels of calcium might be

bene®cial: it would be more convincing to ®nd some

published quantitative evaluations that give credence to

this idea. Lacking such data, I emphasize that I would also

®nd it perfectly acceptable for the authors to admit

something like `in our experience, as well as that of many

other laboratories, the addition of 3 mM calcium to the

primary ®xative qualitatively improves the ultrastructural

appearance of the tissue'.

While I am critical of the theoretical discussion of many

of the topics covered, the `hands on' practical descriptions

are mostly well done and easy to follow. There is a wealth of

information in these practical details which will surely help

people who want to apply one of these methods for the ®rst

time. In contrast to the theoretical parts, the practical

descriptions of the diverse methods are (mostly) well

supported by satisfactory references to the original papers.

The book is also nicely illustrated with pictures of the

important machines and gadgets, as well as examples of

micrographs showing the ®nal appearance of cells after

particular specimen preparation schedules. The book covers

a vast array of methods for plastic embedding and gives, for

the most part, suf®cient information on `how to do it'; the

reader with an open mind can learn a lot from these pages.

However, the reader who needs some guidance as to which

specimen preparation to use for a particular purpose will

probably be disappointed, as will those who may want to

use some of these resins for immunocytochemical purposes.

Indeed, the latter probably represents the major reason why

anyone might want to experiment with the (non-epoxy)

resins in the ®rst place. In this book immunocytochemistry

is mentioned only in passing.

G A R E T H G R I F F I T H S

Topics in Electron Diffraction and Microscopy of Materials.

Edited by P. B. Hirsch. Part of the Microscopy in Materials

Science series. Institute of Physics Press, 1999. Hardback,

196 pp £75, $120. ISBN 07503 0538 X.

Where has electron microscopy got to? Is it a standard

laboratory tool which has reached maturity and is bought

off-the-shelf to solve speci®c problems? How has it responded

to the challenge of the new scanned probe microscopies?

Has confocal light microscopy stolen a substantial part of its

market? Is it still an exciting ®eld to be working in because

the brightest scientists of their day are pushing forward its

frontiers? I have a view on these questions, and you probably

do too. In this book you can read the views of 13 inter-

nationally renowned microscopists; you will not ®nd a

direct answer to any of these particular questions but you

might be able to draw some good inferences.

This book emerged from a meeting to celebrate the

outstanding work of Mike Whelan. To those of us content to

reveal our age by admitting that we were brought up on

Hirsch, Howie and Whelan, this volume might be called `the

return of HHW' since Hirsch and Howie are contributors

and the spirit of Whelan pervades the whole book. It opens

with a short historical survey by Professor Hirsch, which is

followed by a wonderful analysis by Alec Moodie which

demonstrates just how full of insight were the early papers

of HHW. The crucial thing they got right, time and time

again, was what approximations to make. Two absolutely

critical decisions were to use the two-beam approximation

and the column approximation. Not only did they work, but

they simpli®ed the concepts suf®ciently to enable journey-

men like myself (and thousands of others) to understand

what was going on. Or maybe tricked us into thinking we

understood, since every time I have taught the subject I

have discovered new features which I clearly did not

understand properly at the time.

The remaining seven chapters of the book, each of which

is succinct enough to be readable at one sitting, either deal

with techniques which are well established but still of great

importance or attempt to show directions in which modern

electron microscopy is moving. In the ®rst category are

chapters by David Cockayne (who better?) on weak beam

methods, by S. L. Dudarev on using backscattered electrons

for diffraction and Lian-Mao Peng on the use of RHEED for

354 BOOK REVIEW S

q 1999 The Royal Microscopical Society, Journal of Microscopy, 196, 352±355

Page 4: Maximum Entropy and Bayesian Methods : by G. J. Erickson, J. T. Rychert and C. R. Smith. Kluwer Academic Publishers, Boise, Idaho, U.S.A. 1997. ix + 297 pp. £84.00. ISBN 07923

monitoring single crystal growth. In the second category are

chapters by H. Hashimoto on the reduction of aberrations,

by Archie Howie on valence loss spectroscopy, by Colin

Humphreys and Gianluigi Botton on probing bonding in

solids using density functional theory with EELS and by John

Spence's group on the possibility of molecular imaging.

Who would bene®t from reading this fascinating collec-

tion? The primary readership must be practising research

electron microscopists with a physical bent. It might serve

both as a guide to the content of the book and as an indicator

of the way electron microscopy is going (to partially answer

my questions at the beginning of this review) if I reveal that

the ratios micrographs : simulated micrographs : diagrams :

equations are 1 : 1 : 2 : 6. However, before letting that

frighten you off, let me say that John Spence actually

devotes quite a lot of space to the effect of the beam on the

specimen itself (not a simulation of it!). Overall, this book is

a wonderful tribute to Mike Whelan's extraordinary con-

tributions over many years and is worth reading just to get

a feel for the breadth as well as the depth of his vision. It is

also written in a scholarly and well-referenced style which

should ensure that it retains its usefulness for many years. I

recommend it.P E T E R G O O D H E W

BOOK REVIEWS 355

q 1999 The Royal Microscopical Society, Journal of Microscopy, 196, 352±355