ࡱ>  bjbjߍ xH.$8"$F$1jj' ' ' 61818181818181,47Zd1Q'  "' ' ' d1U'1U'U'U'' J61U'' 61U'U'rb0T0 6Ҽq%0*"110107U'70U'04' ' ' d1d1U'' ' ' 17' ' ' ' ' ' ' ' '  J:   Searching the Heavens: Astronomy, Computation, Statistics, Data Mining and Philosophy Clark Glymour һ And Florida Institute for Human and Machine Cognition Abstract Data mining, statistics, computation and philosophy figure in an historical exchange of mutual succor and mutual disdain. Data mining was the tool that created astronomy; the necessity of computation in astronomy was the mother of statistical inference; philosophy was the stimulus to Bayesian inference, and computation was the Achilles heel of Bayesian statistics. Professionalized in the first half of the 20th century, statisticians invented philosophical dogmas to justify computational limitations, ignored scientific historyespecially astronomy--and denounced the data mining approaches that technological developments were making ever more valuable. The development of the digital computer at once promoted Bayesian statistics from toy to tool, and opened the path for data mining by machine. Initially nave, machine learning methods soon adapted the array of tools statistics had developed, and deployed them in defiance of statistical shibboleths. We may be coming to a time when statistics has preserved its mathematical depths but shed its bad philosophy, when machine learning has embraced both the tools and the rigor of mathematical statistics, and when astronomy and other sciences will fully benefit from the merger they themselves are leading. The mutual defenses are falling. The Conflicts Astronomy. Our first and purest science, the mother of scientific methods, sustained by sheer curiosity, searching the heavens we cannot manipulate. From the beginning, astronomy has combined mathematical idealization, technological ingenuity, and indefatigable data collection with procedures to search through assembled data for the processes that govern the cosmos. Astronomers are, and ever have been, data miners, and for that reason astronomical methods (but not astronomical discoveries) have often been despised by statisticians and philosophers. Epithets laced the statistical literature: Ransacking! Data dredging! Double Counting! Statistical disdain was usually directed at social scientists and biologists, rarely if ever at astronomers, but the methodological attitudes and goals that many 20th century philosophers and statisticians rejected were creations of the astronomical tradition. The philosophical criticisms were earlier and more direct. In the shadow (or in Alexander Popes phrasing, the light) cast on nature in the 18th century by the Newtonian triumph, David Hume revived arguments from the ancient Greeks to challenge the very possibility of coming to know what causes what. His conclusion was endorsed in the 20th century by many philosophers who found talk of causation unnecessary or unacceptably metaphysical, and absorbed by many statisticians as a general suspicion of causal claims, except possibly when they are founded on experimental manipulation. And yet in the hands of a mathematician, Thomas Bayes, and another mathematician and philosopher, Richard Price, Humes essays prompted the development of a new kind of statistics, the kind we now call Bayesian. The computer and new data acquisition methods have begun to dissolve the antipathy between astronomy, philosophy and statistics. But the resolution is practical, without much reflection on the arguments or the course of events. So, I offer a largely unoriginal history, substituting rather dry commentary on method for the fuller, livelier history of astronomers' ambitions, politics and passions.  Planetary Theory from Data Mining: Ptolemy, Copernicus, Kepler, Newton By data mining I mean inferences about scientific hypotheses based on historical data that was collected for some other purpose or previously used, or inferences to hypotheses prompted by aspects of the very data under scrutiny. Ptolemy worked from his own observations and historical data, apparently taken from Hipparchus. Unelaborated, the familiar epicycle on deferent device and the geocentric hypothesis left unspecified the order and distances of the planets from the Earth, the radii of the deferents, and radii, centers, and numbers of the epicycles, and the angular velocities of deferents and epicyclesthat is, almost everything. Many of these quantities remained indeterminate in Ptolemy's theory, which needed only various ratios to account for planetary motions. The order of the planets remained formally indeterminate as well, although the periods between oppositions, periods of revolutions of longitude, and comparative brightnesses gave grounds for Ptolemys ordering of the superior planets. He ordered the inferior planets incorrectly, placing Mercury closer to Earth and farther from the Sun than Venus. In addition to the epicycle on deferent, Ptolemy used an eccentric (moving, for the superior planets), the off-Earth center of the deferents, and an equant, the off-Earth point with respect to which the deferents have constant angular velocities. In principle, we now know thanks to Harold Bohr, Ptolemy could have made do instead with more epicycles, as did some Arab astronomers. All of the planetary, solar and lunar parameters came from mining the data.  Figure 1: Ptolemy's Solar system Although it is not fashionable in our mathematical times to think of it this way, the astronomers task is not only to collect data but also to find an explanation of a body of phenomena, an explanation that is consonant with physical knowledge, and, in the best case, at least approximately true. Explanation is a vague notion for which there is no formal theory. Ptolemy is a case in point. Over a long interval of time, each superior planet approximately satisfies a regularity: the number of solar years in that interval is equal to the number of oppositions of the planet in that interval plus the number of revolutions of longitude of the planet (with respect to the fixed stars) in the interval. To accommodate this regularity, Ptolemy required that the line from each superior planet to its epicyclic center on the deferent always remain parallel to the Earth-Sun line. He observed that he had no explanation for this constraint, and thereby expressed a sensedifficult to make precisethat constraints among parameters in a theory should emerge from the essential structure of the theory. That is exactly what heliocentric theory does in this case: the approximate regularity reduces to a necessary truth on the heliocentric model, and that was one of Keplers arguments for a sun-centered planetary system. The epitome of data mining was Keplers exploitation of the data left by Tycho Brahe. We know Kepler built his laws from working over Brahes data with enormous computational skill and ingenuity, one hypothesis after another. (Kepler gave voice to another idea in advance of his time. Making part of his living as an astrologer, when one of Keplers forecasts--of weather as I recollect, but dont trust me--went wrong, he complained of his critics that they did not understand the role of chance variations.) Descriptions of natures regularities are an efficient form of data compression. When laws such as Keplers are used to infer or support other generalizations, the data are being mined indirectly. Newton was many things, among them a data miner. The argument for universal gravitation in Book III of the Principia uses received data from astronomers on the orbits of the satellites of Jupiter and Saturn and of the moon, Keplers laws, and simple experiments with pendula. Data often require a massage to reveal the processes that have generated them. Newton did not explicitly use Keplers first law, although he of course knew it, (perhaps because he wished to claim he had proved it, whereas Kepler had merely guessed it). Instead, in his argument for the law of universal gravitation, Newton assumed the planets and satellites (except the moon) move in circular orbits, and then used the force law to show that the orbits are ellipses. Computation and Estimation: From Boscovitch to Legendre The value of data mining depends on the quality of the data and the methods applied to them. Astronomy created the study of both. Near the end of the 18th century Nevil Maskelyne, the fifth Astronomer Royal, dismissed his assistant observer for errors in his timing of transits across an observing wire. Friedrich Bessel soon found by experiment that individuals differed, some estimating transits early, some late. Formulas were developed for pairwise comparisons of differences of observers in timing a transit, but experimental work soon enough developed an equationthe personal equation-- for absolute errors. So began psychophysics and the study of human factors. Among the first problems in astronomical data analysis are to find a correct way to estimate a relationship among variables from a data sample, and to find a method to test a hypothesis with free parameters against sample data. Eighteenth century astronomers sometimes used a non-probabilistic method to test hypotheses against inconsistent observations, the method now sometimes called "uncertain but bounded error" in engineering texts. The idea is simple: in using a set of measurements either to test hypotheses or to estimate parameters, the investigator believes that the errors in the measurements are within some specific bound. That is, when a value x of variable X is recorded in measurement, the true value of X on that occasion is within e of x where e is some positive quantity assumed known. The measured values of variables then determine bounds on the values of parameters, and may exclude a hypothesis when no values of its parameters exist that are consistent with both the measurements and the error bounds assumed. In 1755, Father Boscovitch analyzed five data points on the length of meridian arc at various latitudes, taken for the purpose of testing the Newtonian hypothesis of an ellipsoidal Earth. Boscovitch had five data points and a linear equation (see Stigler, p. 42) in two unknown, allowing 10 determinations of the unknown parameters. He computed all 10 values, and also computed the average value of one of the parameters, the ellipticity, and argued that the difference between the individual values of the elipticity and the average value was too large to be due to measurement error. He concluded that the elliptical hypothesis must be rejected. Boscovitch's argument is valid if we accept his bounds on the errors of measurement. Stigler claims that Boscovitch's error bounds were unreasonable at the time, and using a more reasonable error bound (100 toisesabout 100 feet) he finds that the least squares line is within the error interval for all five observations. But that is only to deny Boscovitch's premise, not the validity of Boscovitch's argument. Boscovitch's procedure is sensible, simple in description, informative about the truth or falsity of hypotheses of interest, and requires only an elementary kind of prior belief that could readily be elicited from scientific practitioners. It corresponds to an interval estimation procedure that is, if the assumptions about error bounds are correct, logically guaranteed to give an interval estimate containing the true value if one exists. Boscovitch himself later abandoned the method for a procedure that minimized the sum of absolute values of differences between estimate and observations, the L1 norm as it is now called. Uncertain but bounded errors have advantages. If the error bounds are correct, and errors approaching the bound will occur on both sides of a quantity, then the method converges to the true value. But the method had the overwhelming disadvantage for scientists before the late 20th century that, by hand calculation, uncertain but bounded error estimation is intractable except in the simplest cases. Early in the 19th century, uncertain but bounded error methods and L1 norm methods, were rapidly displaced by least squares, a criterion that had been proposed and investigated in the middle of the 18th century but was not publicly solved (Carl Gauss, as usual, claimed priority) until the appendix of Adrien Legendres essay on the orbits of comets in 1805. Part of the reason for the success of least squares was that Gauss and Pierre Laplace gave least squares what remains the standard justification for its use: the expected value of least-squares estimates is the mean for normally distributed variables, and least squares minimizes the expected squared error of the estimate. The central-limit theorem justified the assumption of a Normal distribution of measurement errors as the limit of the binomial distribution, or more substantively, the Normal results in the limit from summing appropriately small, unrelated causes. But, as Abraham DeMoivre emphasized in The Doctrine of Chances, probability had no logical connection with truth. Jacob Bemoulli's theorem, for example, did not say (contrary to proposals of some 20th century frequentists) that a binomial probability could be defined as the limiting frequency of a sequence of trials; for that purpose, the theorem would have been twice circular, requiring that the trials be independent and giving convergence only in probability. In contrast, under assumptions investigators were disposed to make, the method of uncertain but bounded error afforded finite sample guarantees about where the truth lay, assuming only what astronomers thought they knew. But, with Legendre's method, least squares solutions could be computed by hand, and provided a uniform assumption not dependent on individual judgements of measurement accuracy. For those benefits, astronomers were willing to change what they assumed. From Philosophical Skepticism to Bayesian Inference: Hume, Bayes and Price Plato's Meno is the ancient source of skepticism about the very possibility of using experience to learn general truths about the world. After a sequence of conjectures and counterexamples to definitions of virtue, Meno asks Socrates how they would know they had the right answer if they were to conjecture it. In the second century, A.D., Sextus Empiricus revived Meno's query in a more direct form: the absence of an available counterexample to a generalization does not entail that there is no counterexample yet to be found; hence general knowledge cannot be founded on experience. David Hume received a classical 18th century British education, and he surely knew Plato's writings and probably some of those of Sextus as well. In 1740 he published A Treatise of Human Nature, which sought to trace the basis for claims of scientific knowledge to human psychology, and sought to discover the facts and regularities of psychology by informal introspection, the Experimental Method as the full title of his book claimed. No philosopher has been so much cited by scientists, and so little read. Hume distinguished between propositions about matters of fact' and propositions about relations of ideas, and he retained a traditional account of mental faculties, including Reason and Imagination. The scope of Reason is confined to necessary relations of ideas, in effect to mathematics and logic. Our senses deliver impressions and we have as well internal impressions, as of hunger and pain and pleasure and memory. The mind copies and abstracts impressions into ideas, which the mind then compounds to form more complex ideas, and associates (by similarity of content, nearness or contiguity in time, etc.) to make inferences about as yet unobserved sequences of impressions. The intelligibility of language depends on the sources of terminology in ideas or impressions of sensation. To find what an expression such as cause or probability means, one must reduce its content to a composition of claims about ideas and their relations, and hence ultimately to a source in impressions, whether internal or external. Some of these associations are so frequent that they become habits, which we make without reflection or deliberation, involuntarily as it were. We expect the sun to rise tomorrow from habit, not knowledge. The source of our notion of causality is not in any impressions of sensation, but in the habitual association of ideas by the constancy of one kind of impression succeeded by another closely in space and time. Neither do we perceive causal relations, or necessary connections between events, nor, by the arguments of Plato and Sextus (from whom Hume borrows but does not cite), can causal relations be established by Reason. So there are two threads. A psychological argument about meaning: causality refers to events of a kind that are contiguous in space, and constantly conjoined in a fixed order. (In one phrase, which never reoccurs in his writings, Hume does suggest that causal claims are about counterfactualsthe effect would not have occurred had the cause not been present.) And a logical argument: constant conjunction of observed events does not entail that as yet unobserved events will satisfy the same regularity. Although frequentist statistical reasoning is drenched in counterfactual reasoning, 20th century statisticians embraced the first thread of Humes argument and overcame the second by appeal to probability. What of the probability of causal judgements? Hume has the same psychological take: probability is merely degree of opinion, and of course opinions differ. There is no objective content to causal or probabilistic claims, and they cannot be known by Reason or through experience, and hence not at all. It is not known whether Hume's writing helped to prompt Thomas Bayes to work out his account of probability and inverse inference, but Hume's work certainly influenced the publication in 1784 of Bayes' Essay towards Solving a Problem in the Doctrine of Chances. Bayes' literary executor was a fellow reverend, mathematician and Fellow of the Royal Society, Richard Price, who saw to the posthumous publication of two of Bayes essays. Price also had philosophical interests, and he found in Bayes' work a response to Hume's skepticism about the possibility of knowledge of nature, and, not incidentally, to Hume's skepticism about miracles. Price gave Bayes' essay its title, an introduction, and an appendix directed to Hume without naming him: Price used Bayes' method to calculate the probability that the sun will rise tomorrow given that it has been seen to rise on a first day, and so on. Bayes' formulated his theory of probability in terms of the betting odds one ought to hold given a prior state of knowledge. Bayes construes pure ignorance of the probability of an event type as the prior belief that there is an equal chance for any frequency of the event type in a given finite sequence of trials, and from that assumption infers that the prior probability distribution should be uniform. (George Boole later objected that if the equal chances are for each sequence of possible outcomes in n trials, for every n, then the prior must be exactly .) Hume, not Bayes, was the subjectivist about probability, and the stimulus of his skepticism was the last useful contribution that philosophy was to make to probability for more than a century. Bayes method had two difficulties, the selection of the prior probability distribution and computability. Posterior probabilities could not be computed analytically except in simple cases. Bayes' theory had a minor following in the 19th century, and was dismissed by the 20th century giants of statistics until the digital computer, and the approximation algorithms it prompted, changed Bayesian methods from an interesting mathematical/philosophical toy into a practical tool. Early Planet Hunting Observing is mining the data of our senses. The invention of the telescope soon prompted sky surveys. What was found was sometimes not what was sought. In 1781, using a seven foot reflector with a 6 and inch mirror, of his own construction, in the course of a sky survey of objects down to the 8th magnitude, William Hershel, the most astonishing amateur astronomer ever, found the object we now call Uranus (Hershel had another name in mind, for King George III), and put to rest the ancient assumption of six planets. Hershel thought for some time that he had seen a comet, but the subsequent calculation of a nearly circular orbit settled the matter. Now, besides comets, there were other solar system objects to look for. Early planet hunting was a multi-stage data mining process of anomaly detection and classification: survey the sky for an object that does not move with the fixed stars or that has a distinctive disc; estimate its orbit sufficiently to demonstrate that the object is a planet. The search was not random. As early as 1715 David Gregory had noted a progression of planetary distances, repeated in subsequent textbooks. In 1764 in a translation of Charles Bonnet's Contemplation de la Nature, Johann Titus elaborated the point, adding that there should be a planet between Jupiter and Mars and calling for a search. In 1772 Johann Bode repeated the point, initially without reference to Titus. In keeping with Stigler's law of eponymy (nothing in science is named after the person who first discovered it), the corresponding power relation has come to be known as Bode's law. Baron von Zach organized a group of astronomers to look for the missing planet, but Ceres was discovered serendipitously in 1801 by Giuseppe Piazza in collaboration with Gauss, who did not make the observations but calculated the orbit. The distance was closer to the prediction from Bodes law than was that of any of the other known extra-terrestrial planets except Jupiter. The celebrated story of the discovery of Neptune does not need repetition here, except to note that the process pursued independently by Urban LeVerrier and by John Adams was in interesting ways different from the data mining procedure that found Uranus. In this case, the anomaly was not sought, but found in the motion of Uranus. Adams assumed a single planet, initially located in accord with Bode's law (which Neptune actually substantially violates), calculated the predicted perturbations of Uranus, changed the hypothesis to better account for the residuals, and iterated until a satisfactory solution was found. Adams' solution changed several times, and his method appears to have been one of the reasons that James Challis, the director of the Cambridge Observatory, was reluctant to search for the then hypothetical planet. At the nudging of the Astronomer Royal, Challis eventually did a plodding search, and actually observed Neptune, but (like many others before) failed to recognize the planet. LeVerrier had similar problems with French astronomers, and the planet was eventually found at his suggestion by the Berlin Observatory. It is tempting to speculate (and why resist?) that the episode had a benign influence on astronomical practice. Many astronomers were subsequently eager to use their telescopes to search for effects that might explain the subtle anomalies in the orbit of Mercury. The discovery of Neptune raised a question about language and knowledge that no philosopher at the time seems to have seriously addressed. (The contemporary philosophers had chiefly become a conservative burden on scientific advances: Georg Hegel, in an incompetent dissertation on Newtonian celestial mechanics, had proposed replacing Bode's law with a philosophical series taken from Plato; William Whewell opposed atomic theory and evolution; Auguste Comte announced that the material composition of stars would never be known.) Neptune did not follow the orbits computed either by Adams or by LeVerrier. Benjamin Pierce argued that the discovery of the planet was a happy accident--the observed planet was not the computed planet. Indeed Adams and LeVerrier had the good fortune to calculate Neptune's position near conjunction with Uranus, when perturbations were maximal. John Hershel argued that the essential thing in discovering the planet was not the calculation of an orbit, but the identification of its position, which is what Adams and LeVerrier had almost given (there were errors in their distances). Herschel's point can be viewed as the primacy of ostensionpointing--over description in referring to an object. What counted most for success of reference were not the various descriptions that Adams and LeVerrier gave of the objectthe cause of the perturbations of Uranus' orbit, the body with such and such orbital elementsbut that they could say for the time being where to look to see it, and that it was a planet. In retrospect, Leverriers indignation at the happy accident account seems more appropriate than either Pierces criticism or Herschels response to it: Leverriers happy accident, like most good results of data mining, depended on a very un-accidental and thoughtful program of search and inference, Mining historical data to improve orbital elements for the planets, in 1858 LeVerrier announced an anomalous advance of 38 of arc per century in Mercury's perihelion, which he thought could be explained by several small planets orbiting between the Sun and Mercury. So began a search that lasted for the rest of the century and found lots of sunspots, but no planet. Near the end of the century, LeVerrier's calculation was corrected by Simon Newcomb using better (and, unlike LeVerrier, consistent) planetary masses--and worsening the anomaly. By 1909, when Newcomb reviewed the history, astronomers and physicists were left with no alternatives but to alter the inverse square law, as Asaph Hall had suggested, or to postulate some as yet undetectable massive miasma such as Hugo von Seeliger's zodiacal light. Bayesian Statistics, Data Mining and The Classical Tests of General Relativity In October of 1915, Einstein produced an explanation of Newcomb's 43 second per century anomalous advance of the perihelion of Mercury. The paper provoked controversy at the timeEinstein did not have the field equations he was soon to produce, and used what amounted to a special case (in 1916 Schwarzschild and Droste independently found the exact solution for a spherically symmetrical field) ; his paper was littered with typographical errors (kept in the standard English translation!) because the paper was printed by the Berlin Academy without review; his derivation was obscure enough that a central pointwhether he was using a power series approximation he did not explicitly give or merely analyzing a single term deviationare still matters of scholarly discussion. Any number of contrary hypotheses and criticisms were published, including the fact that Einstein's and Schwarzchilds analyses of perihelion advance neglected planetary perturbations of Mercurys orbit amounting to about 5 of arc per century.  William Jeffereys and James Berger, have argued that Einsteins explanation of the anomalous advance was a triumph of subjective Bayesian reasoning. A subjective view of probability holds, with Hume, that probability is opinion, but opinion with degrees constrained by the axioms of probability, and that, for rationality, on coming to know or accept some datum, the probability of any claim (including that datum) must be altered to its conditional probability on the datum. The subjective Bayesian interpretation of probability was first given a justification in terms of betting odds by Frank Ramsey in 1926. Since digital computing made computation of posterior probabilities feasible, a considerable number of astronomers have claimed to be Bayesian in this sense. Useful as Bayesian statistics now is for astronomy and many other sciences, it does not provide a plausible account of the reasoning in the classical tests of general relativity; instead, the circumstances of the classical tests illustrate fundamental problems for subjective probability as an account of an important aspect of actual scientific reasoning. Jeffereys and Berger compute posterior comparisons, conditional on the perihelion advance value, for general relativity and for alternative explanations of the perihelion advance at the time which, unlike the relativistic prediction, had free parameters that must be adjusted to produce the advance. As usual with posterior ratios for alternative hypotheses on the same data, the prior probability of the data in the usual expression for conditional probability can be canceled. And that hides a problem. Bayesians abhor double counting: starting from a prior opinion or subjective probability distribution D1, considering a datum (or body of data) and in light of it altering the prior opinion to D2, and then computing and announcing a posterior probability distribution D3 by conditioning the altered distribution, D2, on the datum. The procedure is regarded as especially vicious when D2 is announced as the prior. The exact logical error in double counting from a subjectivist viewpoint is never (in my reading anyway) made explicit. Were the final posterior distribution D3 equal to D1 conditional on the datum, ignoring D2, there would be no problem in having peeked at the datum. If the initial probability D1 is consistently altered to obtain D2 by conditioning on the datum, then subsequent conditioning again on the datum is harmless and does not change any degrees of belief: in that case D3 must equal D2 because in D2 the datum (conditioned on itself) has probability 1, and conditioning (to form D3) on a proposition (the datum) that already has probability 1 does not change the probabilities of any other propositions. Double counting can be objectionable from a subjective Bayesian point of view only if the intermediate distribution D2, is incoherent (i.e., does not satisfy the axioms of probability) or disingenuous, and in particular if in D2 the datum is not given probability 1 although in D2 hypotheses of interest are given their conditional probability (in D1) on the datum. In devising the general theory, Einstein peeked at the perihelion data. In fact, Einstein developed a series of relativistic gravitational theories between 1907 and 1915, and repeatedly used the perihelion advance to test and reject hypotheses. So, from a subjective Bayesian point of view, in regarding the derivation of the perihelion advance from the general theory as a confirmation of the general theory, Einstein must either have been (Bayesian) incoherent, or else have been mistaken that the perihelion advance increased the probability of the general theory of relativity.  Even that is not quite right, because at the time of his perihelion advance paper, general relativity was not in the algebra of propositions to which Einstein could have given degrees of belief: neither he, nor anyone else, had yet conceived the theory. More generally, as a Bayesian economist, Edward Leamer, seems first to have made vivid, subjective probability gives no account of the rational transition of subjective degrees of belief upon the introduction of novel hypotheses, and of how the logical relations of novel hypotheses with previously accepted data are to be weighted in that transition. That is exactly what was going on in the case of general relativity and the anomalous advance of Mercury's perihelion. Photography changed astronomy. Photography made possible data mining that was not dependent on whatever the observer happened to record in the telescope's field of view. And it did much more. It made possible long exposures, increasing the resolving power of the instrument; it allowed the recording of spectral lines; and it allowed repeated examination and measurement of images. Only digitalization, which came a century later, was comparably important to astronomical methodology. By the end of the 19th century astrophotographers had available a gelatin coated dry plate process that replaced the elaborate collodion process that had made astrophotography feasible by the 1850s but required preparation of a glass plate immediately before exposure and development immediately after. Dry plates enormously facilitated photography during eclipse expeditions. And so it was that when Einstein predicted a gravitational deflection of light, astronomers ventured to solar eclipse sites to prove him right or wrong. Karl Popper regarded Einstein's prediction of the gravitational deflection of light as the paradigm of scientific method: conjecture, predict, test, and revise if refuted. The historical sequence was rather different: conjecture and predict, conjecture and predict differently, test and refute, retain conjecture and prediction, test and refute and test and not refute, test and not refute. Einstein produced a relativistic theory of gravity in 1911 which yielded a Newtonian deflection of 0.87 at the limb of the sun. An eclipse expedition from the Argentine National observatory attempted to measure the deflection in 1912 but was rained out. Expeditions from Argentina, Germany, and the Lick Observatory tried again in 1914 but the German expedition became Russian prisoners of war before reaching the observation site and the other two groups of observers were rained out. In 1915, using his new theory, Einstein predicted a deflection of 1.74 of arc. An expedition from the Lick Observatory took eclipse photographs at Goldendale, Washington in 1918. Heber Doust Curtis, the astronomer in charge of reducing the observations, and W. W. Campbell, the Director of the observatory, on separate occasions initially announced that the observations were inconsistent with Einstein's second prediction. For reasons to be discussed later, Campbell refused to publish a final account of the 1918 measurements. Late in the summer of 1919 the results of three sets of measurements with three telescopes from two British expeditions, one to Sobral, Brazil, the other to Principe, off the coast of Africa, were announced. The British observations and data analyses have recently been discussed several times by historians, but the first blush results were these: One instrument in Brazil gave fuzzy images and some rotational distortion but showed a lot of stars; measurement of the glass plates and reduction by least squares gave the Newtonian deflection with a substantial standard deviation (0.48); a second instrument give precise images, a small standard deviation (0.178), and a deflection (1.98) substantially above the prediction of general relativity, which would have been rejected at about the 0.1 significance level (no such test was reported at the time and no variances calculatedthe standard at the time was probable error.) A third instrument, at Principe, supervised by Arthur Eddington, showed only five stars, not very clearly. By aid of some legerdemain, Eddington calculated a deflection of about 1.65 with a standard deviation (0.44) about equal to that of the first Sobral instrument. Setting up the problem as a forced choice between the Newtonian and the general relativistic values, at a joint meeting of the Royal Astronomical Society of London in the summer of 1919, Frank Dyson, the Astronomer Royal, announced that Einstein's theory had been confirmed. No explicit Bayesian analysis was feasible at the time; modern methods for combining estimates from multiple data sets, or for testing hypotheseslet alone from multiple data setswere not available. Useful statistical methods were limited to least squares and probable errors. (I have never seen a modern analysis taking account of the three data sets, their varying numbers of stars, and the three distinct group variances. My vague guess is that a reasonable hypothesis test would reject both the 0.87 and 1.74 hypotheses, and a Bayesian comparison of the two with equal priors would favor the larger deflection. (There were any number of possible deflection values in between, but none were considered at the meeting.) The conflicting data from Brazil were explained away post hoc, and Eddington, in his Mathematical Theory of Relativitythe first and for a good while standard reference on the theory in Englishpretended the expedition to Brazil had never occurred. Needless to say, a lot of astronomers, including Campbell, were unconvinced. Under Campbell's direction, a Lick Observatory eclipse expedition to Australia in 1922 produced the first unequivocal confirmation of the general relativistic prediction. At the joint meeting in the summer of 1919, after the presentation of the results of the expeditions and Dyson's verdict, the acerbic and ambitious Ludwig Silberstein, who was understandably not fond of English weather, objected that solar spectra could even in this country, be observed many times a year. Silberstein advocated a theory he called Relativity without the Equivalence Principle. Silberstein's point was that measurements of solar spectra to test the gravitational red shift were not going Einstein's way. Even Eddingtons endorsement of the theory was initially qualified because of the solar spectra results. The case of the gravitational red shift presents an example of what might be called reverse data mining: using a theory to alter the interpretation of previous data that did not support it. Einstein had given heuristic arguments for a gravitational red shift in 1907 and 1911, and in 1916 he gave another derivation from general relativity assuming a spherically symmetric and static gravitational field. The argument was heuristic and theorists had a devilish time keeping straight clock measurement, coordinate time, and proper time. Even the great Max von Laue derived a blue shift! Work in the 1880s had found that solar lines shifted most often towards the red compared with terrestrial sources, varied with the element considered, with the line intensity, and were not proportional to the wavelength. A Doppler shift could not explain them. In the first decade of the 20th century solar shift measurement were made by a variety of observers, of course without reference to relativity or the equivalence principle. Between 1915 and 1919 repeated measurements were made to test the gravitational red shift with various lines, notably and persistently by Charles St. John at the Mt. Wilson observatory. It was not until after the announcement of the British eclipse expeditions that consistent confirmations of the predicted red shift appeared. Some of the confirming measurements were new, but many were by reverse data mining. Old measurements conflicting with the data were reviewed and reanalyzed in accord with the relativistic hypothesis, for example by claiming that solar lines had previously been mismatched with terrestrial lines. It is hard not to think that the reanalyses were driven by the aim of reconciling the historical measurements with the new theory. (The reworking of old numbers to accord with a new theory was not novel in science. In 1819, for example, Dulong and Petit proposed that specific heats of elements are proportional to their atomic weights. They had measured specific heats calorimetrically, but the only atomic weights available were the tables produced by Berzelius, largely by guesswork, assumptions about molecular formulas, and chemical analogy. Dulong and Petit simply replaced the entries in the Berzelius tables with values that agreed with their hypothesis. Good methods may more often find the truth than bad methods, but truth doesn't care about how it is found.) Statisticians Become Philosophers By the early 20th century astronomy and physics had established the causes of the motions of the solar system bodies, and were beginning to unravel the cause of the sun's heat, the composition of stars, classifications and laws that allowed distance measurements to stars, the resolution of nebulae into galaxies, a relation between red shifts and distancesHubble's lawthat gave insight into cosmological structure, and much more. The laws, mechanisms, and natural kinds (e.g., star types) of the universe were being revealed by physical theory and observation, connected by serendipity and data mining. Meanwhile, statistics was undergoing a revolution limited by two factors, computation and the suspicion of causality, a suspicion sustained by philosophy. In the 18th and 19th centuries, in the hands of Laplace, Bayes, Price, George Boole, George Udny Yule and many others, probability was the mathematical tool for the discovery of causes. There was no other mathematics of causal relations, no formal representation of causal structures as mathematical objects other than those implied by concrete theories, whether atomic theory or Newtonian celestial dynamics. That division continued with the remarkable development of experimental design by Ronald Fisher, whose analyses always took known potential causal relations as providing the structure within which to apply his computationally tractable statistical methods for their estimation and testing. Among philosophers and many statisticians, however, the 20th century saw a remarkable change in outlook, signaled by Karl Pearson's announcement in 1911 in The Grammar of Science that causation is nothing but correlation and by Bertrand Russell's argument two years later that the notion of causation was a harmful relic of a bygone age. Russell's argument focused on the reversibility of Newtonian equations, Pearson's on a kind of subjectivism in the spirit of George Berkeley, the 18th century idealist philosopher. (Pearson, possibly the worst philosopher ever to write in English, claimed that material objects do not exist: they are merely illusions produced by (a material object) the brain! Pearson seems not to have noticed the contradiction.) Positivist philosophers of the time such as Rudolf Carnap and Carl Hempel avoided the notion of causation altogether and sought to replace it with deductive relations among sentences. Just as in the 17th century Gottfried Leibniz adopted a bizarre metaphysics apparently because the formal logic he knew could not accommodate relations, the logically minded philosophers of the early 20th century seemed to have been suspicious of causality because causal phrasings were not easily accommodated by the novel and impressive developments of formal logic that had begun in the 1870s. Fisher's attitude seems to have been that talk of causation makes sense in experiments, but not otherwise. (He seems to have had a similar view about talk of probability.) A lot of astronomy was about finding the causes of observations, but Fisher, whose Bachelors degree was in astronomy and who had worked initially on the theory of errors for astronomical observations, did not appear to see any tension between that and his commitment to experiment. A general sense that causation was an objectionably metaphysical notion became pervasive; probability was acceptable, even though no one could provide a tenable explanation of what the term referred to. The chief difference was that probability had a developed, and developing, mathematical theory, and causation did not. A second difference was that with observational data, associations could, it seemed, always be explained by postulating further common causes, whether historical or contemporary with the associated events, and so no direct causal relationships could be established. Fisher used the second argument to attack the conclusion that smoking causes cancer, and even as late as 1998, Judea Pearl, the computer scientist who was later to become the most prominent contemporary champion of causation, used it to reject causal interpretations of non-experimental data. Causal claims based on observing might make sense, but they were always unwarranted. The counterexamples provided by astronomy and evolution were seldom noted. Along with an aversion to causation, prompted by the wide influence of statistical hypothesis testing 20th century statistics developed an aversion to search The principal argument, repeated in scores, perhaps hundreds of sources throughout the century, was by worst case. Given a sample of N independent variables, and, say, a hundred hypotheses of correlation tested on the sample, the chance that at least one null hypothesis of zero correlation would be rejected at a .05 alpha level can be much higher than .05, depending on the sample size. Statistical methodology produced two responses: the Bonferroni adjustment, which reduced the rejection region by the reciprocal of the number of tests; and the methodological dogma that a hypothesis to be tested must be specified before the data used in testing it is examined. It seemed not to matter (and, so far as I have read, almost never noted) that the same worst case argument would apply if 100 independent scientists each pre-specified one of the 100 hypotheses and tested it on the sample, or even if each of the 100 independent scientists each had their own samples from the same probability distribution. Neither did it matter that worst and expected cases might differ. The logical conclusion of the conception of science as implicit or explicit hypothesis testing seemed to be that subsequent testing of other hypotheses on samples from the same distribution, or on the same sample, should reduce the confidence in whatever hypotheses passed earlier tests. The unavoidable implication was that almost the whole of the history of science, and especially astronomy, was a statistical mistake. In an unpublished paper, one prominent economist, Hal White, drew that very conclusion. By the 1960s, conventional statistics was fundamentally in conflict with science, and Bayesian remedies were unconventional and in practice unavailable. Then both science and statistics were both changed by a profound technological revolution: the computer and the digitalization of data. 8. The Digital Revolution The 20th century saw an expansion of sky surveys with new telescopes, and near mid-century the practical use in astronomy of sky surveys outside the visible spectrum using radio, infrared and, later, X-ray and g-ray detection. Those remarkable developments were complimented by the development of the digital computer and, around 1970, of the charge coupled device (CCD) camera and other solid state digital detectors. Early radio telescope instrumentation produced huge amounts of data, printed on page after page. Magnetic recording in the 1960s, followed by digitalization and automated computation of correlations allowed real time reduction of raw data and the production of images for scientific analysis. Once data were digitalized, all of the convenience, speed and potential analytical power of the computer could be put to work on the flood of data new instrumentation was producing. It may be that the digital revolution had a social effect on data intensive sciences such as astronomy and psychology. While particular female astronomers from the 19th century such as Caroline Herschel are historical quasi-celebrities, women were rare in astronomy until past the middle of the 20th century when their numbers began to grow. Before the digital revolution it was common for technically able, educated women to work as computers in observatories. For the most part the careers of these women are unknown, but there are exceptions. Annie Jump Cannon was such a computer at the Harvard College Observatory, hired to help complete a collection of optical spectra of stars and to classify them. She developed the standard classification of star types, O , B, F, A , G, K, M (Oh Be A Fine Girl and Kiss Me) Although eventually appointed an astronomer at Harvard, initially she, like other women assistants, was paid less than the secretaries. The American Astonomical Society awards a prize in her name. Henrietta Leavitt was another computer at the same observatory at about the same time. She discovered the correlation of luminosity periods and distances of Cepheid variable stars, an essential basis for her supervisors, Edwin Hubble, famous law relating distances of stars to their red-shifts. Leavitt was belatedly promoted the year of her death, and still more belatedly, posthumously nominated for a Nobel Prize. Less recognized, Adelaide Hobe was an assistant at the Lick Observatory in the same period, the first quarter of the 20th century. She conducted the re-measurement and recalculation of the gravitational deflection from the plates of the Observatorys 1918 eclipse expedition. Her work helped convince Campbell that the photographs were not decisive against Einstein, and may have helped to prompt the Lick Observatory expedition to Australia in 1922. Curtis, the astronomer who had done the original measurements and calculations, and who by 1920 had moved to become director of an observatorythe Allegheny, in Pittsburghturned out to have been arithmetically challenged. The more capable Hobe suffered the indignities of her time. Graduating from the University of California in physics (although her transcript says Home Economics) and unwelcome in graduate school because of her sex, she had taken work at the Lick Observatory as a human computer. One astronomer noted that it was a pity she was a woman. The University of California, and other universities, did come to admit women to graduate study in astronomy, but it is difficult to believe that the coincidence of growth of women in the subject and the emergence of technology that removed the need for human computers was entirely a matter of chance. 9. Statistics and Machine Learning Meet One Another The computer prompted new statistics and rejuvenated old, and in consequence the logical divide between scientific practice and statistical orthodoxy began to be bridged. Bayesian statistics suddenly became a useful tool. The calculation of posterior probabilities became possible through fast converging asymptotic approximations, notably Schwartz's, and through Monte Carlo methods, notably Gibb's sampler and Metropolis algorithms. Bayesian methods developed for analyzing time series and their breaks, and for fusing disparate data sets. In astronomy, cosmology and elsewhere Bayesian statistics succeeded by treating each new problem with model driven likelihoods and priors generally given by mathematical convenience, dispensing in practice if not always in rhetoric with any pretense that the probabilities represented anyone's, or any ideal agent's, degrees of belief, but yielding scientifically valuable comparisons of hypotheses nonetheless. Bayesian methods made Issues of multiple hypothesis testing, for example in searches for extra-solar planets, largely irrelevant. The cornucopia of new digital data in genomics, Earth science and astronomy begged for automated search methods, methods of handling hitherto unthinkable numbers of variables, and data sets varying from small in number of units to, well, astronomical. They were not long in coming. The major developments occurred first in classification, regression and clustering. Flavors of neural net classifiers abounded, followed by probabilistic decision trees, support vector machines, Markov Blanket variable selection algorithms, and more. In many cases successful classification or regression became possible even when the number of variables in a dataset is two orders of magnitude greater than the sample size. New source separation methods developed, notably varieties of independent components algorithms. Statistical and model based data correction, as in outlier rejection, was supplemented with classifier based data correction, as with boosting and data polishing. Problems of overfitting, which traditionally would have been addressed only by testing out-of-sample predictions (no peeking!), were addressed through resampling methods. The worst case Bonferroni correction is coming to be replaced by the more sensible false discovery rate. As the contributions to this volume show, developments such as these are becoming central to the analysis of astronomical data. In many cases these new developments were heuristic, but in other casesMarkov blanket variable selection for example--they came with well developed asymptotic theories in place of more statistically standard but heuristic procedures, and in other cases--support vector machines for example--with demonstrable finite sample error properties. 20th century suspicions about causalityespecially about causal conclusions drawn by computational procedures applied to observational data--still remain, despite dramatic research developments in the last 20 years in machine learning of graphical causal models. This research fractured model specification searches into a variety of cases of varying informativeness. Correlation, for example, is not causation, but correlation is a second moment quantity: in linear systems with non-Gaussian additive, independent disturbances, higher moments determine direction of influence. Consistent, computable search procedures, generalized to tolerate unknown latent variables and sample selection bias, provided set valued model estimates for any distribution family for which conditional independence tests are available; algorithms for linear and non-linear models with additive errors provided finer, in some cases unique, estimates of causal structure; correct algorithms for estimating relations among indirectly measured variables appeared for linear, and some non-linear systems. The new methods were adapted to time series. Most recently, work on machine learning of graphical models has extended traditional factorial designs to reveal enormous possible efficiencies in experimental design and has provided guides to the integration of fragmentary causal models. Some of this work has been put to use in genomics, brain imaging, biology and economics but as yet has had little if any impact in astronomy. 10. A Brief Reflection The essays in this volume illustrate some of the many applications of machine learning and novel statistics in astronomy, which need no review here. They illustrate a wave of new inference methodology that new data acquisition methods are forcing. From the seemingly routine task of matching sky catalogs to the excitement of the search for new planetary systems, machine learning, statistics and astronomy are happily merging. Bad arguments for old shibboleths are ignored, real concerns for reliable inference retained. Philosophers are ignored. All good.  My accounts of various episodes in the astronomical tradition are taken from standard sources, especially Otto Neugebauer, The Exact Sciences in Antiquity. Princeton: Princeton University Press, 1952; Richard Baum and William Sheehan, In Search of Planet Vulcan, Basic Books, 1997; Jeffrey Crelensten, Einsteins Jury, Princeton University Press, 2006; and Stephen Stigler, The Measurement of Uncertainty, Harvard University Press, 1990. Methodological commentary is mine, not that of these sources.  See J. Earman and M. Janssen, Einsteins Explanation of the Motion of Mercurys Perihelion, in J. Earman, M. Janssen and John Norton, eds. The Attraction of Gravitation: New Studies in the History of General Relativity. Birkhauster, Boston, 1993.  James Berger and William Jeffereys, The Application of Robust Bayesian Analysis to Hypothesis Testing and Occams Razor. Statistical Methods and Applications, 1, 17-32.  . Berger (private correspondence) has acknowledged that Einstein's use of the perihelion advance vitiates the subjective Bayesian account of Einstein's reasoning.  Edward Leamer, Specification Searches, Wiley, 1978. A similar argument is given with respect to the perihelion advance in Clark Glymour, Theory and Evidence, Princeton University Press, 1980  See J. Earman and C. Glymour, Relativity and Eclipses: The British Expeditions of 1919 and Their Predecessors. Historical Studies in the Physical Sciences,W^ ` =h !i[f''p(r(y*....456677[8]899<.<<<<<3=8=?ȸȸȦȒ~~~Ȓ~ȒȒ&hk+B*CJH*OJQJ^JaJph&hk+6B*CJOJQJ^JaJph#hYB*CJOJQJ^JaJphhk+B*CJOJQJaJph#hk+B*CJOJQJ^JaJphjhk+Ujhk+0JU hk+6 hk+H*hk+ hk+51We "iQ!-dh^- & Fdh dh]^dhhdh^h$dha$!#%''y*069??4BMN3R.U W!WY^]cfdkn&o- & Fdh^`-dh^/dh- & Fdhdh????VBXBBB M MMMVV?VAV W!WJXLX[[n&orr's(sKsLs||~~?@#GI֮خݵߵj02,bh]hk+OJQJ hk+6jhk+0JUhY hk+5\ hk+H* hk+6]hk+ hk+5M&o)swyl܎ujA%'dh- & Fdh^`-dh^-dh^bj,:; "#$'(/Xgno12M_`컹𵭵jhUhU hk+6 hk+6]jhk+0JUjhY0J&U hk+5 hk+5\ hk+H*hk+hYhMh]hSCD'n1_dh55$a$ dgdk+-^3 11, 1980, 49-85. A defense of Eddingtons and Dysons interpretation of the measurements is given by B. Almass, in Trust in Expert Testimony: Eddingtons 1919 Eclipse Expedition and the British Response to General Relativity. Studies In History and Philosophy of Science Part B: Studies In History and Philosophy of Modern Physics 13 ,2009, 57-67.  I thank Jeff Scargle and Michael Way for helpful comments.      PAGE 2 h DmHnHuhk+jhk+UhjhU ,1h/ =!"#$% n~B<89"VPNG  IHDRO4IDATx]w\G>ߐaA`vQc16(y-D%vPĂ( `{AEP," yYw=~wv;3ϟ?+ ܨH" _mB  ӧ? c…ǏZ؆(0I|W|) 0Z\"!2 / JjDih}@rJx*Uddd+WUsC )l+9R+To'*Uo]gj{p4 fffKAHMvZf7V\yٲeŋ׫cƌA`ii)!F Ԍ ~PbŊW\;w#/ȹdR3~7Rʛ5kֺuk777Io vjnnZT)m1>Ȥf4^ ԩ#!`bbݻm;>ڷo/-Ԍ۷o[nNNNѢEŶC,Y2))oC7omA&5$ϭ[^(^׏?&-Ƈٳgjժ2TrAmᅉ'nذ!==] @NΜ9#_I&>]ضdR32#~;;;mKX;y$D͛Ŷ [}؆zAAA#LjFH>}/_^ls򇕕֭[w)1~xpG`` v\&5M!A7nprr:qضߑ#Gb+Vܵkג%KjԨ"9>~W_2%4i7\~Yf/Æ C|d$";v;wf7o4,ȤfX\2J<NlsA%!ٞ59h aIMSG\v֬Y|Zk%EjpHL,Ȥf(^իWm۶HmBtt955J*bۢ$ݻw wuu2 Lj-[ܵk sm^^^Vw߉mȿP;4LjnjԨѤI=m kcƌ}:8tPIMlD3Ӹqcm1ȤV8y$PmO?t5mQ 33D$^z-_\b :wܵkWooo۝qM~m?#޽_uѢE 1ʕ+ׯ_oțrf͚[nI'dR+ rwwߴid{???p ٢E }}|P-?|@!u:uK.-J >yDa[G5.]aLj//vppJ2?}.p˖-񙝝uw}իW\5KwEl+2LL:u˖-*U2׬YӢE3fpw~zǎ… >/*Wl٬wJgl*OW򔎂ŋ9rd޼y~~~o||WJJ M厃6776l؞={tI1c޽{&50)11Q^c<^z7o^~欁pӴ4yt) ڵkITi۶tH 5jTxxLjܐς`C.6;p *dʕ<}۷.]Zؽ]HAa;}tP؆H2d 4Y}vN UGǻ,Y׷}::t(QbܸqKCr"""ĶBI ?azڵkSM277͟?r:$`<Ћ/Jm_<92px{{[XXlܸQZL΀֭u1.\cRuB%;{>DV〳ի €.w}70ժk K&~ҩ)fff&+Y`bbbgg{%/^6ŋǎKj7Lj*APhQ.`Ofi߾}||V0M!+†+VڵkB%X|y(#AxwsIKK={סFѼys(Mx^fہ)T"}rc3رcxx8b}cȧ|(`iݺÅUV}(C HR(JԫWT/ ۷6idȑ]t)W\Ju1bj׮Mj03.gѣG:tH[l L /SLy 5#+%dZhO?@qqqqw c,,,J(qe{{{\{-0Pmv*U91Sti(5θ`$†#G貍SQ,]K_z̙5k2 +D :uꄢ ;vi˗/iڴ%sTt>v,0B}ƌ}#VnA9;;sRDA?ҥKSRR/^|ΝgddR)(?|@H'.xH3gʔ)zjP oKS nWX188^Sw@ =zxY[tFիW'W2:ܸq#""B%զMݓzLѢEݻgnn4x`:C)2|98."5*BM6i K7o޼h"F~Ug/͚5eKX}tgee%ݥ YFz(QwN7>}jMj&M:x ^lY!EjժctuppB,5kÇԩ{RŊ;};u{jD &풂yڵJ*o*AJ UDɥ_+XR 1},6lt c e_-6%Kܱcǖ-[}ˊQ}Ԥ|xrƿ7|!͆ )ի׸qiӦ 3~W^^^qqq=5 55Uӽ/]F'V`]FԔfX -R+ wd 3`C>}rfիWwhDj?~l׮ŋu57w=z.ʱ\rb[!-NԊ-w&MnӧwDJN4R0Zq;~%~ 6̛7OJz<ʀ~^899 4H- d^zuA3f߇{ܹZN%0T4111E:t[4b€Ne*2Xɀsf4h˸jݺus͛7[hssq\q~dBڵoܸ!҄ eVP|y;;o/^\d/^XD 8Z& EؘF%i25^m}j2NU1ݻ'Nd`ȑ:hBɒ%gϒyv?taP+GX"J T ҄ܧVѵkcǎ x*U?ޣG &S6mܜ>}QUv3LY ft2eʈma]{H/3! "BBBq*TZpVhh(c\/77>}ZȠ[tDO! N`y)]#YuԁWbE|Bd۷oʕ eJKKx񢣣er֨Qrj PGpwA+$@j{&MH8fH &Z8ڌ p,WbŦNzԩ%KoM61?@yAE&%%A߿ĉc@)0p>K.}ʕ+$ЧfbbҾ}{>l\u!x:CMLL4Z^zO};u"C_~aDD~֯_ԨQZ[m&B /@B9rD"hI]6r T+-6JE LFaehvݺuYf#G7%>}X[zڴi0F ѓ&}ʸ/z]P6S:@VwPWhd?(uV"h%ԝ "D u=dP0qoĞ:Ď=%?iӦuImggoJm*ݻwo:5puwiBiGf !5Rkܸ1O] )5/^uwiBM uY2ibV5jHlCTڐǏǎ3fX(Q"::ZKL,PԤaD}/"U&5Qp9_}P ׯ_h!R](IaѫWWvMlC@ҥ޽+*ЩSlH$)E>{Ll+!I MbԥL+Vl^}qq>[O4"4lĉb[H-==]o͛7Zo',m@= ^v||ӧ322RSS_z'M6͛7777/S A#O8Ql+055> 5yH"Gʹjg?UR ('?|-Z(22rҤI۷/_<( ?///--mϞ=h۶ɓBQ<|).!DDDʏbl{' 6~jQ6!lPkN,ċtڵk8dT%u[ZZf͚ EPBoo9s$$$(U5M_)4kx,IRis3]ᰀAdgg uxUίĤQFz|c2'l')33S.yBق.ۤ$TYeDB9GܧرcE%ܾ$aڵAmz w8\Q DϿեtjt#7$R.)x9&|dm͝ZU}jZ86m֭ݻI[J p3F3fo,C or.o8VV!q3?J<#=PaFӂR͛7gjMD%׷F666[E P҇mnn"$[]˖- "N"D`|%ZݳOIq\Jި 6L>].jٸqco QzMP8bdFG?AR1ӧO;uTV-]+rF5UV;v:)B9KKK|իÇ_`A9%tIY[M`ܓ%e΢ーf ׭[ʕ+K~9yPP/rRjC+%>(!z,?˗ԩhD^|YD R4?KpǏ/c(Ip$tr- M6>>>=zrQ^@XJ*%TjB%2MP.M" IWMQ](зP'PZhrFEPJvw:7+WLf~Lӧ q۷o>x;GbTJ/2h*@2'ZC"&+x|xk$wfi FΡP]"#.LP<*|JsQ"q9s #|M_I aQJ޾}ۼyׯ_CS^k,S YI lVV͛7믵k^pFB vqפ`i/ ?KL?E'OHR*H l*T"B;AJPm ӧO@@@pppNƌSFzxB%}bO4 @$A?~֭[/{߿ΎF2ŋ5j ?k׮Gh4_Y*EUkṅ;ZOA%t.ʕamr2䂦53?*"gH<MP#ণ7n ///[[&MP@ U[r{쉍mٲ%jժYjկk/_ ֆ'L⒖=EMu8=`OOOpܒ%K9|5#)|Հa (Z[nU8Cw'''ˣtP>|a >lѢur^P"Z8q"bŊV/E:;;?\>cXbŊ:u긻S'[իNa},52d(Ç1鼆;†ڵk*Hsss1iР75T 8P_dRÄT%ïa fMB‰JիG) pKe=z5k֤\vԩSz2W{ 6l…v+_(%N ׮ZjÆ h(akkիK.n߿g>zBaNNNN螸Pk۷olAn0NP1*O#?SSS0t2e6 (5333$''>Uҵ?D*^ ju# j#}iӦn"35}x1,X@?d+Wܯ_>}DDD[nӦMGܹ!C֭+x? \?xVVV<9rsjԨ!e!!!cƌ|x H>H)UwDQo|qttD5֭[Z cЮRv1rl̙ZU꦳p͛7Dϟ?3\/Uhmm 'ݽ{wc4KK#Gԯ_̙3z|{xwp${RS{ݼyСC-O8]FVxEA@pSN Fm0S~; ߇ަMr耽(@Wq…ƍI/]v eWh-%ݻZ~n |Z7Mp{a d`g|AdRv ҥK_rXI$5|jM25cƌxs2ם\jUH6|4׮]SzuhV0F0}d&&&ǎCV"ALU?@uZt)|h8%K4CBPUxpԓ/Fo?Hhʕ+ǧ=V3ŗu pPĞ>}zժU(DM AEԩdǒ%K@ϗX>oVGII1dzm !B_yTB -^ HV/4:?ѣ&00/... h7o߾}ǎ 8{p:WܡPnC70r HJȮ}8HNz7on߾ d۷o) *^I ³ٽ{7F!˗/C򋞁 } Q(W^ݹs'UO)F/.\v}&T ܹs;w ;Tvvv`SSSu (lN?22{!?n[[[6l^v]8O#HDFHJN2.3336߿? !e77͛7}7okd'O<<<ρjFPf e Y6{l)> M&ϔ$[~4URZI]V'9;;O8W^ W!Ax v{- vE r q3G]>r#,@)G" PcJuse 0,>>4ٟÞH<7n\ φ݅"p4.;TF+quu}J:|䣕ߩ|C%gmh8nq!XҵkW2w >ǰ2":.<%)Nx`Њ'=EpȑxT֭[Tt߯^<ϷƘT9Xz]ؙFi!C R=b2p,N>$ۇНŗ8Zݸnռk#Ҏp̡џ)kС`H#.պ4،ő޵Wv&EҳyJ3Qc۵kG"ZM"##x8q_lcw'++}htS T:*|,iTbcc|||&M끍 &i{u\dA!NA&t=H*0,X#Cú(/Ν;p< (7o6m!O9sܺukΝCnݺu̙7"Ԉu1#F.]sN:#i6l%&&S&Jy{/_W^h#EڡLCڴڦv{FFs g &,1l,Y2((>*FO+''N:)))Df2CcǎvvvPy111 &ڵ+^[ȽիWsѢETH?GX{ ZAWTFcw*i]>[HgjJ2-~ǎX4E޽{ &WBϟ?_"!%k[u:OV- Hdɒ%SN]]Sxq;Ca4iҕ+W*Vq9Jʿ(c+9Byyyŵl.-"Wܰaöm6mdkkiG!왐ФIvfo&_*cMAmr(ߓv"IP c>_k *!d=Y:Rݓܹþ9L֤!3fs1bРB+kғR7ӘK_r=:4o^sȑUV/9AqQRcV͝.Yk!>}ڡC3g@.I VA7h j0T*U仞< 1<={viPpѢEVZ,$\޽*hBp_vmx}trFDDIt7v}9L@7L>}'UbFLG,zXЍa&' RkCgƲ#Ʌdo8/"K"uayV* 35K1Kh"'}͚5)7 ܹ3u˗֫W/;QnB_D Bv2^lYdddCBB`@5]^M JR0(dLQ~;;TV!u3b)}idƍSbA>ɱ6nG~}m۶ >*M*? #[e ?}vy#PI A҆BSP=Ufƕ[l;v3U@>5o gϞ`^`w'#YW'99I6IJ(-IWY?ǣ:m۶氨k,ʚҥuO`VLj⺨4 Q(w'WH);3!5`OnAFNOOM._ -?`u뒠i|$2'IW?#`לB ^f /pf/_uV\oĀ{E4̙3hAaFڵ+vɢC pr("S,uFy| g ]pr> "(I۷I%Z%ݥ77߃N< &CB|'j?~x(3AmvKUj -j:[`5(i—$:"/_O>b`wܹpR]fLs/T]vծ]M htk+ quc6q HIEbݗTdҕE/c___vO w|y-Vgq34] ;wxzz=zǸqmذtȞjTs"dNҤF}Lc)Sڵkw-RZ:{,O ~AX(dz-,nggJOZjg%&&JK%::QhTrgUkP^˗Çg@Br/#^kP+O(֎H8BOeMʽmUI&`CyxxZ `@yjAL_ԵTDҥKUq5˗/ϙ3+&lll}777W(z7n@'5{Aznsz}(?6nܨ%gOb\Ku.yz a) YG);^Ж]4a:ro[P ?!r12D桩vqq?S*K5uInNk6OA:+qq RdV#'' OjFct5ḧG^mTB&"k ޷oݻ5|3M# @X @&3b*qb*Bޡgz.ʕ+C9rdьMIE\!0q/GRܣB:= Xںu #*4R6d[O\rǏBm׺G yE͛Z-S@Zz5C-^x˖-AL68Q|F d@Հ-:r6A,C>I0Y5jO?Q#e$]HX焓ir/֭[TA Ԏ VIMX۵k .M3L?~<|F]^$yժUNNNaaaP`7h ~:&&~ . x7ыdɠ~BCCwpA:D`q)w}~G t\˗wyȑڥ 8sLX;yu5k -]tϞ=#""@A-ZH @^:f QS@2S4jK(Yg:.%]$ 8˖-;v?-;;ϟ?;wv%&%%%88xРAC !GP&L0}y恗oںlٲ(P쭫 @۶mt0 k'|⑔3^zu֍c}& $F 3{=c ҵ2iPN:?E}]f }!@s`xZƍh0%00 q˗/?0fTч)=Z@!c/ҳgO%K̷颟Ξ_)=Gq9MG 4)Xhh]˔)ӠA͛7wE]ԴdmeeuMj-D%Jɞ]oӦMQAy-[<~x=p&kng:٭d7n(ԌE ޛnc*uclp _޽{ׯFۣBݻLmp5kVZZaoK&uxp4V&fffr 䒍Ǐ[V8"Yo>c T1zbW^#(TR[l\!C,--<1gHZsE G*d)SPFLj:իA͜9Ngr>=z_-6jNÆ R O?G$ܽ{w̘1D5p4}1͌WFaliLM=#Sԩ LBSD)q$Y"09ݻ˗/^e`C&5۷oҥAYXX@շi;إe,&&m۷oGKTvmS|w:QHΌӧt%  ֭Kw/S.͌.diqҳM49}""?2Rpj[P`ea%P`ߋ1h׮ĉ7oNb R^|X[,lذ &| CvoӯӉuf:6q,\lG/q#,VwFR ѣ'RMؤ-X޾AC, Vжz'Eȗu)v؁{APłԴ: Q322­x1 fffVlY|3f\ C;(4/a-zBz(d@zʂPi;]wΝ...d:}m N%MeƮN[~=5vRPdG!J 8ʏ/µ(TAv_ E&?[R U|k&$%ߞ 9 Ucƌw'bC8-[F(f *15q,h3M4$Ç=<<ҡwÁN<9sL9rfZd dR/K)x dW򧎦O $-ܹsff&ԉ'&>^XSRR@#Gx񢭭md1dR+8}?3"qAts֭ HȀ4dx??駟LMM"- v=qDGGG<TȤV(зo_5YCCc>|pwwS:b޽=z0Mzyӻwo8P+VTw^^^VV۷ϟZ%%%1 >ӭM5Ԩ`-"y7o"09s&A[UرcnnӧO-,,BCCkԨ&3cǎ*ϛ7ںJ*X..dR+ظq={TE43MeԽ_ #n rn2q]5*ȸTŲȄ:/Ub*ՔθҥK0yWܾ#ٰڅq"<]駟?,!/SUB&5.h3עHP718TƆ N]e0)>=))) l 7LشiI2xv>a}B⣎+,x𠗗KuF}ڵ%KjbsibiggK:zLjjr61{v(GرcXVG?hѢsCfffnW_ZN%4RI*t||eݴP  I"+++,,,55UlCZ- ::zĉtDwVRqڶm+!* Zʕ+^jff&!\FrVz%Ķ lݺɓb[?=4h !* ZAJܺmRRRlmm6m*!* H͛7ې!11ml0ȤVڸvǏKm7=w/^pttۊ/p]KKKusgϞ%;H2(\|nYAzxxH*ɓ'm={6tڪ}mjȤ% ڵk~ixAɓ'VVVҩ M6-..NlC#Gې&u$] 0EΝ[paϞ=Ŷ >}Zl*uGtttrrrb4J\NNY[:k=6 2}R|yb F)}FFMMp)B{}___  sE?Q7 oڵkRjU IIIbKjUZh>>Ԣ4IftSo` HǗGN^:**#_%YͿOW+{XC~/**_ܸqcݺu# /K%* :ZDnH=+3xСC߼y.ZJȻw{jGmggW~}fhS}WʍB&55w TZU]|ƍh$ ~6 ۷OR,pLj_h->|{bfz{{XBlC:sqqې/-WLV祖I͈@Dٛ7o!|aZE2zj777a/tǙ3g&L X:peR8ecc#|~{N 6/vZxO޿Ϙ;-~zС4M\qp… uLIMB#rrr=ZdɆ JAG!CƌØ}ο1ܹ3qDY؃S3W'I -ԈSdRMCCC&5##IqFE]\\Əo2I`իWCMAk?~\lCA^^ޥK$޽{O.5quu^IMҀ32bĈ̙3gmjې|rJ*m'NzXx>6l(! qa8۷o`mm !c7o &޼y؂e'NXjw)`VN$&&:88Ep:ܺukС}ې|HNNf/+:_^V\7L11 Ȥ&QdjٳgRc4,X&M^z%hɣGm.]  &ssVVVʕ6/dR"}իW TTJ*I}xyymxdRiii`$0 ? K9rd׮]bnݺ5|p2suQ3Zxx#٩Ox؆@ddUʖ-+!5rHb dR34<==1f›7o517Yӟ? mj|b[/]111jm ᐓs //ӧOE'N:emm-!pΝI&x̬,I>}ZbE%IM::ֵkWgϞ5T%Ç6mڸqIIS'Sٞwȑ`7*^hh(-ݓ{ FCQKOOҥ U֭%X4u&%%鞔FQcSNKK1bxXssbUYmٲ ]w߾}]ּysch(:uu]iqÇ!B=t@tLwӧUT!;;?mLk֬X4hPvv؆0j쾧B&5Aػwo|r20RRR֭{}QQQŋعs)#Cb+dR yyyaaa={,06077߰aq>yA(3'OF=I s]5 4 2pyIMQ~ ԩ#!LfIMcڤڵСC7nܨVО7o6̈ Իw獵Ϙ1C}搓eB LSȤ) ?|𠠠ÇH\vuӠQ(%Ǐ 6LJJlOZZi #R`7؀|-2ŇN:믿vԩAhʕ3%)` F3O?XB ڵK mK_vӧEGXHԤʖ?44t̙nz$b>5 /[~VVVh$8 ӧ:t ERo޼"wA{{{:tÆ %JwO@v2h㋍ vƍIII;w/,\pĉeʔ  jmmժav_tʁYBb A)L<J-ֹsg˗)_re͚5GKTg?@&KvF&$$o߾Y7dRCgB(]4\ )J*ƂԝD|()||||ڵ9čxN+Ǐ}>>>RPg爈} a-LdFP P}O\0d l޽wA? 4$9`={•p@#_׻wu|Ŕ%$//?s.]dgW:ZB!Aw/!z {cZ}ps޽+VhsСh ؈.HMM ^"*<|_~eʔ)}%)u׬Yի͚5;tP!HӮqϑ lKKKUB4 N)DF.\w ׮]@=r3f@K0䕹y\\ש#߾} Q`A.}_ϟjٲ%~:88xV$H} ߨK aΜ9#4:^|ٯ_?4ob/$L/oݺu;;\߇np^ }̙aaa;vHp5_hƨ]\\l7'6ףx\l4*lقO y<:kԨђ%Kv]PUHLLѾ;?~عsg}ֺu+WVT~ֶ\rpH2*>zhG hxΝ{q &$8?[HIIY}vvPŋM:qz;ܼyHnܸ&K.bG+ 6m$$$ Ӛ5k^Gyxx檔|ͩS6m ((^VP<(΍LjgϞ:u*FWtt4C;;;}$Yv8άN:[*]l+]˖-”`6,{9!Zj׹sݯ\bUƚ7o;} hnn>q|8aq>/7ҥKtEp[VU Ppa4GGG=+88;3FӪUÒ|1tЃJsk׮@̊m!ZAC^^իZ(_)+ۭ92++,ZÃrTo.AjhER CÆ njެ;)Ņ3%ݺuhm2¶t_p, rFӅ:oeeO|mڴG XH !C_^ 'O߫$5FQb$ 6mDԐFj+HndffhqAv?XNjTnn+ϟ3bg\JIj@yXH4KU 4P> Y NS/_n) 6 >4ڶl+ks (ѤnΝκnnn˗/ק6#6 Q)d!r޾aÆ8vx|uuu4 YӃhpڵV?zhW\Zt)8f4%qАDGƈlffF###_~ٲe ׆p jJ7o_H✞={{$HÒ%K:L¡ٯ_?nD"QEEEtt4xm~wVVVr B@ O*޽{R>}zdduo+ Y$SZ j|zGGGp U 7!!gEm֬Y X_Aؾ};h_JJW_}J~zzPa/jg,Ebb"xLLo _%[nSjIQF2rˡ{{{;::B{׶GWss3 R/lȐ!Ht) ;w@0rJqX/Q_DT wvva`GRP5gS:2vҥ B {,8$%FpTJR8G5GLB |EC m ߿?ؼ}Y%cw) CxQ :VUU :4&&:^zqm mll D[.Ϟ= $r*~}5jT~~>t6bCxȑۣ'Yp,/=ͫVȀX9q8|?li[XXL>m63--m֬Y8 94%44ի5},Rѥ\QpРwssرc\FFF =?h6m7 3UwDzFP}MM e2۷oo.u%-,Rs¨%J8}& *@ >Ń6#ܺܰڴ6Ǵݻw<r'Nv! Q~򗆆z{{[ZZB3޶m$58p+#l *,,\j D|&7n@6l֚̆nNOOoooWpHGT5@˄v۷]6!!A__]"z_|EjjFѧO L>}ʔ)p 555|ζ6(˙i#|Q:wM6}l ঁ/Qu jAވݹsg֭O}4erŠ&1.[kk={_|ɔ !p[ 9; 2QLd7YvttePeDFFfffc~Y޽?c C? H ]9R?[¾5556lhii9wE0؟ܫWSps\zGN ,ȐXTnCPOhɔ(Rۛ]g_pYmݺu12dt"5Y%.Љ N8AS$$Q8#)/4An 3CKKK'O,=ZK &7% 7~͚5666˖-{A~%"D\3|K eȑ#sssǏOIT\?=zgKǫՍQLGիWHwm›P +j?UmV [2fCSTT$>'Nj MMM /_$;ٓ1P:QM6Ȋֱ/HApLS3I@V*JOOFbMIǪ;wnFF ;9BI^}pnܸHԯVQQ ܴiSNNΓ'OhƓ&ѣǃˣ,ܽ{ e h<}eΜ9|AÖࣁ%''(":--ASR򚛛9rk[HH?#MRzfO/KYY١C޼yÒ ŋ<==4>|MD(_x%]]]Νvm6tЀo߆XZZzyyqm Qz, J`%56YNԆR._,5%S,hSSSN?>*~Oס&6݀?~\BjGGG_%3:x_tdf000]'tttΘ1C{p jfjٳg===?ѝ tpmOM@@3g<"?R?#^nN۷!!!9S/: !3Ğcǎ\v*ڿyC~ .O8^УG:#_X휜ɓ'Ϛ5 @" nV]]ˣnjÈ ;ZEg"j|A|D+Gi3vA +[7?6 gvuG[[[!v4h'S= E׀m%~N8W uC-33޾'PϗEiF}ڻw/I\UVR"h~b;IH9㰗'~""~gbPTN#^ZZZ,Yr} S2GKJJڿSSg}FuEZUV' @dG"<j%tO3B9MBm' WWW//gRWih(.0<vdC|%3C6tܻwoΜ9ΝSGn߾ caˋԘ_?ƪ6vvvVVV\B*8^FP>QPjD"QLLLpppccε9tɱ?&8uuuWXSp_xv… &pmD:*/߿_\\lllmbb 4dȐ, mJ_B:L\ٹnݺ<~xذa~akk+:kϞ=MP4hQ@Pނ'"2Xy͜9gϞ@^IJ Fµ-RP!+TJmrAS-dG0] Y qqؑk<~ٳcccΝFf---atk )NH@Q&1`UV]D6}dQL{ QLp1|FM_#kRkW{{{TTTxxxnn |m5ؘq@~~~ΞD T }>IENDB`7666666666vvvvvvvvv66666686666666666666666666666666666666666666666666666666hH6666666666666666666666666666666666666666666666666666666666666666662 0@P`p2( 0@P`p 0@P`p 0@P`p 0@P`p 0@P`p 0@P`p8XV~8XV~_HmH nH sH tH Z`Z Normald*$$CJOJQJ^J_HaJmH sH tHDA D Default Paragraph FontViV 0 Table Normal :V 44 la (k ( 0No List 2o2 WW8Num1z0OJQJDA`D Default Paragraph FontJ/J Absatz-StandardschriftartP/!P WW-Absatz-StandardschriftartR/1R WW-Absatz-Standardschriftart12oA2 WW8Num1z1OJQJ2oQ2 WW8Num1z2OJQJ.oa. WW8Num2z0^J2oq2 WW8Num2z1OJQJ2o2 WW8Num2z2OJQJ2o2 WW8Num2z3OJQJ.o. WW8Num3z0^J.o. WW8Num4z0^J.o. WW8Num5z0^JJ/J WW-Default Paragraph Font>o> Footnote Characters@&`@ Footnote ReferenceH*@o@ Endnote CharactersH*B/B WW-Endnote Characters:/!: Numbering Symbols>*`1> Endnote ReferenceH*BAB Header CharCJOJQJ^JaJBQB Footer CharCJOJQJ^JaJ@&`a@ Footnote ReferenceH*>*`q> Endnote ReferenceH*NON Heading (x$OJQJCJPJ^JaJ6B@6 Body Text )x(/@( List*^JH"@H Caption +xx $CJ6^JaJ].O. Index, $^JD@D List Paragraph-^]`FOF CM1.d1$7$OJ QJ CJaJFOF CM9/d1$7$OJ QJ CJaJFOF CM40d1$7$OJ QJ CJaJ<O< Table Contents1 $FO"F Table Heading 2$a$ $5\N@2N Footnote Text3^]` $CJaJ4B4 Header 4 H$4 @R4 Footer 5 H$^Ob^ Preformatted Text 6OJQJCJPJ^JaJPK!pO[Content_Types].xmlj0Eжr(΢]yl#!MB;.n̨̽\A1&ҫ QWKvUbOX#&1`RT9<l#$>r `С-;c=1g'}ʅ$I1Ê9cY<;*v7'aE\h>=,*8;*4?±ԉoAߤ>82*<")QHxK |]Zz)ӁMSm@\&>!7;ɱʋ3װ1OC5VD Xa?p S4[NS28;Y[꫙,T1|n;+/ʕj\\,E:! t4.T̡ e1 }; [z^pl@ok0e g@GGHPXNT,مde|*YdT\Y䀰+(T7$ow2缂#G֛ʥ?q NK-/M,WgxFV/FQⷶO&ecx\QLW@H!+{[|{!KAi `cm2iU|Y+ ި [[vxrNE3pmR =Y04,!&0+WC܃@oOS2'Sٮ05$ɤ]pm3Ft GɄ-!y"ӉV . `עv,O.%вKasSƭvMz`3{9+e@eՔLy7W_XtlPK! ѐ'theme/theme/_rels/themeManager.xml.relsM 0wooӺ&݈Э5 6?$Q ,.aic21h:qm@RN;d`o7gK(M&$R(.1r'JЊT8V"AȻHu}|$b{P8g/]QAsم(#L[PK-!pO[Content_Types].xmlPK-!֧6 -_rels/.relsPK-!kytheme/theme/themeManager.xmlPK-!!Z!theme/theme/theme1.xmlPK-! ѐ'( theme/theme/_rels/themeManager.xml.relsPK]# iixOyG  GJ  ?bsvx{!&o'tuwy !l,b$B<89"V.@(    <A#" ?0 # CNLw4Z`_h#V_hn 49JPAIMU4<{ &&A'K'/(:((( ))v))))i*u*****e,o,//g2o222(34355U6Y6'7+77777779#9@@#C.C DDSSTTUUVV;YDY\#\`\g\\\%].]]]^^`_i__```waaYbbbccccdd!e+effiiiimnIyOy| }Z}c}QZW`# ژgk5?;F.8 fqXcyy/?3?s?)@"TaT{YY`a}c-dgghhiiqkknnq/qݒޒͫ5Ks~йmp(1NX J::::::::::::::::::::::::::::::::::(J_ ^`OJQJo( 8^8`OJQJo( ^`OJQJo(o  p^ `OJQJo(  @ ^ `OJQJo( x^x`OJQJo( H^H`OJQJo(o ^`OJQJo( ^`OJQJo(^`OJQJ.^`OJQJ.pL^p`LOJQJ.@ ^@ `OJQJ.^`OJQJ.L^`LOJQJ.^`OJQJ.^`OJQJ.PL^P`LOJQJ.^`.88^8`.^`.^`.pp^p`.  ^ `.@ @ ^@ `.  ^ `.^`.P^`P@@^@`0^`0``^``^`^`^``^``00^0`WW8Num1WW8Num2SC D]k+YM@RLg%g&gg9`@`.``@``@`@Unknown G*Ax Times New Roman5Symbol3 *Cx Arial7@ CalibriU XMKLBI+ArialMSCambria? Courier New;WingdingsG=  jMS Mincho-3 fg5 Tahoma[ HKBKUS+Arial-BoldMSArialACambria Math"h G Gm p!20!e3P]! xx ,Valued Acer CustomerUser     Oh+'0   0 < H T`hpx'Valued Acer Customer Normal.dotmUser2Microsoft Macintosh Word@F#@@T}ļ@T}ļ ՜.+,0 hp|  'm!  Title  !"#$%&'()*+,-./0123456789:;<=>?@ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqrstuvwxyz{|}~Root Entry F}Ҽ1Table7WordDocument xSummaryInformation(DocumentSummaryInformation8MsoDataStoreҼ9Ҽ4DQRC30ODEC==2εҼҼItem PropertiesUCompObj `  F Microsoft Word 97-2004 DocumentNB6WWord.Document.8