Computing Graphics

A joint newsletter of the Statistical Computing & Statistical Graphics Sections of the American Statistical Association.

April 1993 Vol.4 No.1 COMPUTING GRAPHICS

A WORD FROM OUR CHAIRS FEATURE ARTICLE Statistical Computing Saxpy, gaxpy, LAPACK, OneoftheperksoftheChairoftheStatisticalCom- puting Section is writing a column for this newsletter. and BLAS Imagine: I can write for several thousand members of Colin Goodall The Pennsylvania State University my profession, without the bene®t of refereeing. I get to tell you what I think, rather than what I know. Measuring Performance Maybe I'm a bit odd, but what I've been thinking about One of the best understood computational tasks is lin- is copyrighting and patenting of statistical software and ear algebra. Considerable effort has gone into fast and of computer material in general. My interest was piqued accurate code for these manipulations, e.g. LINPACK by an advertisement I received for a computer pack- (Dongarra et al. 1979), EISPACK, and most recently age that would display a high dimensional plot using a LAPACK (Anderson et al. 1992). The speed of these patented algorithm. The name of the method was jar- computations is measured in mega ¯ops (MFLOPS), or gon, so I could tell nothing from the advertisement about millions of ¯oating point instructions per second. Each the program and what it did. Mainly, what the adver- ¯oating point instruction is a single arithmetic opera- tisement did was make me wonder just what it meant for tion, e.g. a multiplication, a divide, an addition or sub- an algorithm to be patented. Was I supposed to believe traction, performed in full ¯oating point precision arith- that patenting was a substitute for peer review? Did it metic, usually 64 bits double precision. Two common imply certi®cation? Did it mean that I could not com- benchmarks for comparing performance are Dhrystone, pute this graph, whatever it is, with my own computer which measures speed of integer arithmetic computa- code unless I paid a royalty? tions, and Whetstone for ¯oating point computations. Assuming that most of us non-lawyers are as ignorant These measure not just the speed of the CPU but also of copyrighting and patenting as I was, I thought a compiler performance. Dhrystone performance, mea- short summary of what I've learned might be of in- sured in millions of instructions per second, can equal terest. Finding the information was easy: one call to and possibly exceed the clock speed of the CPU. my university's patent of®ce brought a few relevant pa- A more exacting benchmark involves a practical lin- pers, particularly [1], and a copy of the law. A review of

ear algebra task, speci®cally the 100 100 LINPACK my university library's on-line card catalog gave nearly benchmark (Dongarra (1993)). This measures the speed 200 references, including a journal, Software Protec- achieved, in MFLOPS, in computing the Cholesky de- tion, which has been published since 1982.

composition of an arbitrary 100 100 symmetric ma- Copyright and patent are very different. The basis trix, using a ®xed set of FORTRAN code. The numbers for copyright is contained in the U. S. Constitution, obtained from this benchmark are surprising: For exam- which gives to Congress the authority ª[t]o promote the ple a SparcStation II, with a processor rated at around

Progress of Science and the useful Arts by securing for 25 MFLOPS, has a 100 100 LINPACK speed of 4.0

Limited Times to Authors and Inventors the exclusive MFLOPS (Dongarra (1993)). Put another way, the theo-

n Right to their respective Writings and Discoveriesº [2]. retical requirement for an n Cholesky decomposition

3 2

= + n n = Exactly what can be covered by copyright was spelled is n 3 2 ¯ops, or 353,333 ¯ops when 100. CONTINUED ON PAGE ?? CONTINUED ON PAGE ?? EDITORIAL In preparing this issue we have realized yet again the herculian efforts of the newsletter's founding editors, ªThree in 93 and four in 94º was proclaimed as the Sallie Keller-McNulty and Dan Carr. We hope we can rallying call of the new editors of the joint Statistical live up to the standards they have set for us. Computing and Statistical Graphics newsletter. If we have succeeded then this ®rst issue of 1993 should be in James L. Rosenberger your hands before the April Interface meeting. We are Editor Statistical Computing Section planning to mail the second issue before the Joint Statis- [email protected] tical Meetings in August, and the third sometime around Mike Meyer Thanksgiving. More issues mean the newsletter can be Editor, Statistical Graphics Section more timely, provide announcements and encourage di- [email protected]

alogue. Please use your newsletter to communicate with the membership of the two largest sections of the ASA. Our deadlines for the remaining issues in 1993 are the SECOND FEATURE last day of June and October. Many regular columns will continue, but we solicit your Production of help with new ideas and offers to write columns or once- Stereoscopic Displays only pieces. Please keep those e-cards and e-letters coming! for Data Analysis This issue has two feature articles. The ®rst feature de- Daniel B. Carr George Mason University scribes the availability and discusses the design of public domain matrix and linear algebra routines. Many aca- Dedicated to David L. Hall, colleague and friend. demic computer installations can make available high quality subroutines of the standard algorithms without Stereoscopic displays help the analyst escape from the the need to purchase or lease commercial software pack- limited domain of 2-D visualization into the natural do- ages, e.g. IMSL or NAG. Algorithms are available, as main of 3-D visualization. The goal of producing 3-D described by Colin Goodall, to anyone with FTP soft- scatterplots motivates much of the following discussion. ware and access to the internet. The goal has strong implications in terms of selecting a stereo projection. In the every day world, familiarity The second feature, about graphics and stereoscopic dis- with objects and many depth cues facilitates fusion of plays, is like the Sunday night mini-movie on network left and right retinal images into a stereo image. Monoc- television. It looks like a feature article, but is really the ular depth cues include linear perspective (objects of ®rst episode of a new column. Dan Carr has stepped equal size transect areas inversely proportional to their down as editor but couldn't resist the challenge of a distance), interposition or occlusion (when one object regular column. The graphic images which accompany is in front of another it obscures the more distant ob- this article should also be seen as the beginning of more ject), shadows (we generally assume light comes from innovative graphical material which we would like to above), detail perspective (no ®ne detail appears in dis- print. tant objects due to limited visual acuity), and aerial perspective (greater optical depth through the air leads The Newsletter is now being set/typeset in LATEX. Af- ter three years of colorful newsletters we have tried to to a blue shift). For most elementary 3-D scatterplots restrain ourselves to typographic spice and a change in prior knowledge about the form to be perceived is lim- format. Neither of us has any sense of color, so we ited and monocular cues are restricted, so care must be will take the conservative (and probably boring) tack taken in the selection of a stereo projection. and use black text on a neutral background. We want to Two In®nite Families of Reasonable Stereo credit Kevin Fox, the design artist from Penn State, for Projections the new masthead and for keeping our link to the past Different geometric models lead to different stereo pro- with the intersecting circles punctuating the articles. jections. (Geometric models are idealized in that each Submissions should be sent by email to either of the eye has a blind spot, a region of high resolution, and editors. If you can prepare your article in TEXorLATEX various imperfections.) A simple model (Newman and that will make our lives just a little easier. Otherwise Sproull 1979) presumes that the eyes converge on a sin- plain old ASCII format is ®ne. gle focal point and constructs left and right images by

2 Statistical Computing and Statistical Graphics Newsletter April 1993 projecting onto left and right projection planes. The 0). For the right eye this yields

projection planes contain the focal point and are orthog-

s [x; y ; z e= ; ;d] + e= ; ;d=x ;y ; r 2 0 2 0 r 0

onal to the respective lines of site. This ®xed-focal- 1 point model is appropriate for advanced dynamic sys- Solving both right-eye and left-eye equations for s based tems that update immediately as the eyes change their

on the z coordinate and substituting yields:

focal point. However, the ®xed-focal-point projection

x = dx ez = =d z

does not correspond to the data analyst's typical stereo- r 2

= dx + ez = =d z viewing scenario. Valyus (1962) states, ªit has been x

l 2 (2)

shown experimentally that eye movements performed

= dy =d z when stereoscopic pictures are viewed are similar to y those performed in observing a real object. As the The fact the both left and right images have the same gaze is transferred from one object to another the eyes y coordinate is evident from geometric considerations. perform conjugate movements directed to the subjec- Consider a data point appearing behind the projection tively most important regions, and at the same time co- plane (or viewing screen). The two eyes and this data ordinated convergence movements take place.º While point form a triangle that intersects the screen. If the the ®xed-focal-point projection has proved passable for frontal view of the face is parallel to the screen and the showing non-updated images of familiar scenes, image eyes are level, then the twin projected points must also fusion problems result when looking at points in the be level. corners of the plot. Thus ®xed-focal-point projection is The second class of multiple-focal-point stereo projec- inadequate for 3-D scatterplots. tions may be called a depth-cued orthogonal projection. When the eyes have multiple focal points within the In this projection, the LCOP and RCOP are shifted for

same stereo image, a reasonable compromise uses a sin- each data point so that the midpoint between the eyes y gle common projection plane that is parallel to a frontal has the same x and coordinates as the display-scaled view of the face. In a multiple-focal-point model, a data data point to be projected. The depth-cued points are point projected into the projection plane has the same y then

coordinate for both left and right images. This is a fun-

x = x ez =[ d z ]

r 2

damental requirement for 3-D scatterplot projections.

x = x + ez =[ d z ]

l 2 (3)

y = y The two projection classes have continuous projection (x,y,z) +Y · and viewing parameters. Projection parameters include

(xr,yb,0) +X d · eye separation ( e), projection distance ( ) and the size of (xl,yb,0) RCOP · (e/2,0,d) · the viewing cube into which we translate and scale the -Z data. For convenience consider the workstation screen · LCOP as a viewing cube that is 20 centimeters on a side with (-e/2,0,d) its front face centered and aligned with the screen surface. Viewing conditions may differ from the projection

model and that introduces magni®cation (m) and view- 0

ing distance (d ) parameters. These parameters can be Figure 1 varied over an interval and still lead to comfortable im- Two classes of projections satisfy the multiple-focal- age fusion of a 3-D point cloud. Consequently both point (single projective plane) constraint. The ®rst class stereo projection classes are in®nite. uses standard projective methods with separate centers Projection Parameter Bounds and Parallax of projection for the left and right eyes, denoted LCOP and RCOP respectively. Assume a right handed coordi- While an in®nite number of projections are effective, nate system with positive z toward the viewer. Then the some parameter bounds should not be violated and some LCOP and RCOP coordinates relative to the center of projections are more desirable than others. The follow-

ing provides some background concerning parameter

e= ; ;d e= ; ;d the workstation screen are 2 0 and 2 0 constraints. The curious reader is referred to Valyus where e is the eye separation (see Figure 1). The projected coordinates can then be found by scaling the vec- (1962) for additional detail.

tor from the eye to the data point by a constant, s,sothat For most people, eye separation falls in the interval be-

: : the scaled vector touches the screen (the z coordinate is tween 5 2and74 centimeters. Fortunately, using the

April 1993 Statistical Computing and Statistical Graphics Newsletter 3 exact eye separation for each individual is not crucial magni®cation may cause problems. Magnifying the im-

as evidenced by stereo publications that are enjoyed by age magni®es the parallax. A Taylor series expansion =

diverse audiences. of the depth equation about p 0 yields 0

0 2 3

= d p=e +p=e +p=e Parallax is a key concept for understanding the pro- z 6

jection parameter bounds. The horizontal parallax, p, of a point refers to the distance between the projected When the parallax is small relative to eye separation,

magnifying the image increases the apparent depth al-

x x r coordinates on the screen, and l . Then for our multiple-focal-point stereo projection most linearly. As the parallax to eye-separation ra-

tio approaches one-half, the nonlinear terms contribute

p = x x = ez =d z r l 4 equally and magni®cation distorts the image. As magni-

A point appearing in front of the screen will have a ®ed parallax approaches eye separation, the depth of im-

z < d positive z , , and the parallax will be negative. age theoretically approaches minus in®nity. This causes Similarly, a point appearing behind the screen will have both depth distortion and fusion problems before the ap- positive parallax. Parallax must be limited to provide parent depth reaches in®nity. acceptable stereo fusion. Maintaining image focus con- strains the amount of acceptable parallax. Suppose the Control of Perspective Differences eyes converge on the twin images of a point as if the Under natural conditions the left and right eyes have

point were real. The apparent location of the point is the different views of the world. The ®eld of view for a focal point. Eye convergence is usually coupled with single human eye is about 150 horizontally and 135

accommodation (lens focusing) so that a region in front vertically. In binocular vision the ®eld of view covered and in back of the focal point is in focus. If this region

by both eyes is 120 so 30 or 1=5 of each eye's ®eld includes the workstation screen the image of the point of view is unique to the eye. For 3-D scatterplots all will be clear. While those experienced in stereo viewing points must be seen unless hidden by occlusion, so im- often learn to decouple the convergence and accommo- ages must be restricted to the shared ®eld of view. If dation of their eyes, a mismatch can lead to either fusion the viewing cube is suf®ciently smaller than available or focus problems. Constraining the parallax to keep the display space on the workstation screen this is not a perceived image depth close to the screen avoids such problem. problems. The horizontal parallax is related to angular parallax on the retina. Studies (Valyus 1962, Yeh Hodges (1992) provides an astute comment on stereo and Silverstein 1990) have related the speed of image production for workstations. He notes that some hard- fusion to angular parallax and provide guidelines. In ware systems provide only a single center of projection. short, restricting the viewing cube to 23 centimeters on Then the standard trick for producing stereo projections a side or smaller will generally satisfy the less restrictive is to use the ªoff-axisº projection. For each eye's view (but slower fusion) Valyus bound. Lipton (1982) rec- this shifts the data toward midpoint between the eyes, ommends the equivalent of centering the viewing cube projects from the midpoint and then shifts the result depthwise on the screen to control the parallax. In this back. While the projections produce coordinates iden- case the viewing cube can be made substantially larger tical to the LCOP and RCOP projections, the ®elds of before the parallax becomes excessive. view differ. This provides another reason to constrain Magni®cation and Viewing Distance the size of the viewing cube. For scatterplots all corre-

sponding left and right points must appear on the screen. 0

View related parameters include the viewing distance d Perspective differences within the shared ®eld of view and the image magni®cation m. The equation of paral- can cause fusion and interpretation problems. Perspec-

lax can be solved for the apparent distance from a point 0 tive induced fusion problems may be evident when

to the screen, z , in terms of the actual viewing distance, 0 showing the viewing-cube frame in small side-by-side

d .

x 0

0 plots. If, for example, the left cube face has the

= d = e=p

z 1 5 e= coordinate at 2, the left cube face will project as If the parallax is ®xed then doubling the viewing dis- a line in the left eye image but as a trapezoid in the tance doubles the apparent distance of the point to the right eye image. This radical perspective-induced dis- screen. Thus the viewing cube frame can be made to crepancy complicates image fusion. Since the viewing

appear squashed or elongated by selecting a different cube frame is simple, avoiding radical differences in

0 d viewing distance, d , than actual projection distance, . perspective is straightforward. For example rotating Often a stereo image is shown to a large audience on a the viewing cube often suf®ces. In fact some analysts big screen. If the image is not designed for the room, prefer perspective views of the cube that have two or

4 Statistical Computing and Statistical Graphics Newsletter April 1993 Figure 2 - Random Points on a Mobius Strip. The structure is apparent with any 3-D viewing approach. In ¯atland recognizing the structure is much harder. Piecing scatterplot matrix views together is not so easy without brushing or conditioning. Conditioning or slicing is helpful in moving to higher dimensions. For real data, overplotting is more of a problem and density- based presentations become advantageous.

Figure 3 - Selected Stereo Contours for a 3-D Density Estimate.

Figure 4 - Steepest Ascent Ridge Traces for a 2-D Density Surface. Ridges and their projections in domain space provide another way of viewing density estimates. This example is provided courtesy of Qiang Luo.

Figure 5 - Random Dot Stereogram. This contour completion illusion at different depths takes a while to fuse. This image is provided courtesy of Nathan Carr, who used it in a high school science project that repeated experiments pioneered by Bela Julesz (1971).

April 1993 Statistical Computing and Statistical Graphics Newsletter 5 three in®nity points rather than the common head-on than data analysis. The legends provide a brief descrip- view that has only one in®nity point. Another trick tion of the simple examples. bases the projection on a distance substantially larger The examples are designed for parallel fusion. To fuse than the actual viewing conditions. This reduces per- the images look at the left image with the left eye and spective differences. the right image with the right eye from a distance of The foreshortening (interpoint distances appear smaller about a 50 centimeters. It may be helpful to separate as a function of depth) of perspective views compli- the images with a card or to use an inexpensive magni- cates the reading of coordinates based on given axes. fying stereopticon. These parallel fusion ®gures were Equations (3) provide an orthographic stereo projection produced using an S function (stereo.pairs) which is that removes all perspective differences and allows the available via anonymous ftp from galaxy.gmu.edu x and y coordinates to be interpreted directly. Carr (in subdirectory /submissions/eda). and Little®eld (1983) describe a simple implementation Future Articles that exploits the scaling in standard statistical graphics

packages. Plot the left eye and right eye coordinates of a Separate articles will give background for more ad-

X X ;Y X + X ;Y

vanced examples and provide some discussion of statis- p point using p and respectively

where the values are in data units. The expression for tical interpretation. The next article will focus on alpha

X X blending and with luck will include a translucent stereo the half parallax, p ,in data units, is

rendering of overlapping contours of a density estimate

X = k X Z Z = Z p range min range 7 for 3-D data. While many practicing statisticians do not yet have the requisite hardware, computing environ- The constant k is chosen as a conveniently small value ments are changing very rapidly. Soon a high percent- such as :026. age of statisticians will be able to study 3-D structure Little®eld provided a color-anaglyph implementation in in the calm of non-rotating images. Future articles will a statistical package over a decade ago by modifying discuss enhancement tools for the representation and Minitab (remember when source code was available). study of data. This modi®cation allowed drawing of arbitrary glyphs and included color table control to handle light mix- Additional References ing. Red and green points overplotted as yellow as they Many additional details are available in the literature, should and color control allowed subtle depth-based concerning stereoscopic resolution, hypersteroscopy, shading. Color polaroids demonstating that a third vari- and enhancements for statistical images, common pro- able helps little in group discrimination were shown at duction dif®culties, etc. Some useful starting references the ®rst ASA Graphics Exposition in 1982 and addi- are Wegman and Carr (1992), Hodges (1992), Carr and tional examples were published (Carr, Nicholson, Lit- Nicholson (1988), Papathomas and Julesz (1988), Hu- tle®eld, and Hall 1986). While color anaglyph stereo is ber (1987) and Lipton (1982). Valyus (1962) provides not the most desirable form of stereo, the anaglyph work one of the most detailed expositions and the fascinating demonstrated the simplicity of stereo plot production as work of Julesz (1971) produces considerable insight into long as statistical packages provide for control of color visual processing via study of optical illusions encoded mixing (hint). in random dot stereograms. Side by Side Stereo Examples References An article on graphics should have some graphics. Carr. D. B. and W. L. Nicholson. (1988), ªEX- Stereo production methods are incredibly diverse. For PLOR4: A Program for Exploring Four-dimensional workstations, time-multiplexed methods that alternately Data.ºDynamic Graphics for Statistics,eds.W.S. route images to the left and right eyes are gaining in Cleveland and M. E. McGill, pp. 309-329. Belmont, popularity. Side-by-side methods are common in non- California: Wadsworth. electronic publications and so will be used here. Note Carr, D. B., W. L. Nicholson, R. J. Little®eld, and D. L. that other disciplines often publish stereo images in Hall. (1986), ªInteractive Color Display Methods for color (Editors' note: Maybe someday the newsletter will Multivariate Data.º Statistical Image Processing and be able to afford to do this). For example see the ray- Graphics, eds. E. J. Wegman and D. J. Depriest, pp. traced image in Hodges (1992). The additional color- 215-250, New York: Marcel Decker. based depth cues are especially helpful when showing Carr, D. B. and R. J. Little®eld. (1983), ªColor surfaces. The images here are monochrome point and Anaglyph Stereo Scatterplots ± Construction De- line drawings and for brevity focus on geometry rather tails.º Computer Science and Statistics: Proceedings

6 Statistical Computing and Statistical Graphics Newsletter April 1993 of the 15th Symposium on the Interface.NewYork: computers) UNIX-based environment in academia, the North Holland Publishing Company. issues surrounding departmental computing appear in Hodges, L. R. (1992), ªTutorial: Time-Multiplexed business and government and across platforms. My de- Stereoscopic Computer Graphics.º IEEE Computer partment has networked PCs, Macs and multiple Vaxes Graphics & Applications, pp. 20-30. in addition to the UNIX network. I'll try to keep the Huber, P. J. (1987), ªExperiences With Three- presentations as independent as I can of any particular Dimensional Scatterplots.º Journal of the American computing platform. We'll discuss platform speci®cs Statistical Association, 82(398), 448-453. when necessary for examples of general principles. Julesz, B. (1971), The Foundations of Cyclopean Per- I confess to having several basic premises that guide my ception. Chicago: University of Chicago Press. thinking about departmental systems. Lipton, L. (1982), Foundations of the Stereo-Scopic Cinema, A Study in Depth. New York: Van Nostrand Departmental computing is the right level for sup- Reinhold. port of statistical activities. Reliance on Univer- Newman, W. M. and R. F. Sproull. (1979), Principles sity or corporate level computing systems natu- of Interactive Computer Graphics, Second Edition. rally leads to inadequate resources. New York: McGraw-Hill. Papathomas, T. V. and B. Julesz. (1988), ªThe Applica- A coordinated system in which users can send and tion of Depth Separation to the Display of Large Data receive mail, share ®les, share printers and other Sets.º Dynamic Graphics for Statistics,eds.W.S. peripherals, share software installations, share Cleveland and M. E. McGill, pp. 353-377. Belmont, documentation and share user support is a rea- California: Wadsworth. sonable goal.

Valyus, N. A. (1962), Stereoscopy. New York: The Planning is possible. Despite budget uncertain- Focal Press. ties and power politics, planning can and should Wegman, E. J. and D. B. Carr. (1992), ªStatistical be done. Graphics and Visualization.º Technical Report No. Diversity happens. We can't stop it and we have 84. Center for Computational Statistics, George Ma- to work to accept it. The best system for a ®scal son University, Fairfax, VA 22030. person may not be the best system for a statistical Yeh, Y. Y. and Silverstein, L. D. (1990), ªLimits of scientist. Should all the statistical scientists in a Fusion and Depth Judgement in Stereoscopic Color department have the same platform? Let's dis- Displays.º Human Factors, 32, 45-60. cuss this. Let me know what you think. See the end of this column for my e-mail address. *Minitab is a trademark of Minitab,Inc.

S is a trademark of AT&T. A network connection to the Internet is essential for proper access to the information needed to do Daniel B. Carr one's job. Again, let me know what you think.

George Mason University Larger systems require full-time systems manage-

[email protected] ment. We'll start here (see the next section).

In future columns I hope to cover issues that hinder ac- DEPARTMENTAL COMPUTING ceptance and development of departmental computing resources. Here are some of the issues I see as important Not just hardware and (in no particular order):

Acquisition of systems. Choosing platform(s), software fund raising, planning, purchasing. Breaking A computer system has four major componentsÐ inÐhow does one get started? What about solo hardware, software, communications and people. The machines? hardware and software angles of computing get dis-

Administration. System management, user sup- cussed at length. In this column I will try to focus port. on issues in computing that have to do with getting a coordinated computing system going and maintained. Maintenance (hardware, software, network gear) The issues cut across types of hardware and software, of modern departmental computing networks.

operating systems and organizations. While my per- Resource sharing. Coordinating funds as well as sonal experience has been in developing a large (100+ equipment.

April 1993 Statistical Computing and Statistical Graphics Newsletter 7 Speci®cations. What should the system be able monitor system usage, provide and maintain basic secu- to do. Performance, software, licensing, vendors. rity for data, interact with the institution's networking organizations, respond to hardware, software and net- Diversity and innovation. Supporting diversity, planning and implementing innovation. work failures, design and implement solutions to work- group problems, train users in the operation of the sys-

The computer life-cycle. Trickle-down computing and opportunities for old equipment. tem, write documentation regarding system procedures, use of software, and local conditions, maintain the var-

Migration to new systems and new ways of work- ious operating system software subsystems, respond to ing. Acceptance and training. changing external network conditions, simplify system

Institutional barriers to progress. Technical, use by providing new tools, negotiate with vendors, and physical, ®nancial and bureaucratic impediments. generally make sure that everyone can use the comput- As a ®rst topic, let's consider the often neglected re- ers to do their jobs. quirements for system management. Why would anyone think this is not a full-time, dif®cult System Management job? It reminds me of the problems we have convincing some non-statisticians that statisticians are necessary. Providing system management for a new or evolving ªWhy can't I do it myself with a computer programº, system can be a real ªcatch-22º. How can you justify they ask. And we stand ready to tell them. We must system management if you don't yet have a complex be equally clear that system management should not be running system? How can you build a complex running done as a hobby. system without system management? In planning for system management or making the Most departments recruit a knowledgeable faculty case for system management there are additional is- member into the role of system administrator or man- sues. Who will provide backup for the system manager ager. If things go well, the system grows and the faculty so that there is someone to turn to when the system member gets burned out trying to do system manage- manager is on vacation and things break? The obvious ment in his or her ªspare timeº. Graduate students may choice is the faculty member who recently retired from be called in to assist or a scienti®c programmer may creating the system in the ®rst place. A second issue become involved. If things go very well, the system involves hiring and retention. System managers can be becomes complex and the case is made to hire full-time hard to ®nd. Senior system managers have large salaries help. The faculty member is relieved of system man- and may have tired of the daily grind of responding to agement duties. The start-up phase from no system to a problems. Junior people may lack either the technical system requiring a system manager may last years. knowledge necessary to do a thorough job, or lack the Just as a department needs secretarial help to support maturity and responsibility that comes with experience. the work of its faculty, a department with a complex In any case, system managers are in very great demand, modern computing system will need a full-time system and rightly so. One needs to expect turnover in the manager to support the systems and its use by students, system manager's position. staff and faculty. Twenty systems is more than a part- What's Next? time job and 100 systems is more than a full-time job. Check with other departments at your institution to see I would very much like to hear from the readers. We'll what staf®ng commitments they have made to system all get bored hearing what I did, what I'm doing, and management. what I think is important. What issues have you faced in the construction of your departmental system? How Centralized support organizations may be available at were problems solved? What problems remain? Which your institution to relieve some system management issues that I plan to discuss are the most important to duties, but these rarely compensate for having someone you? Which are irrelevant? What should I discuss that locally. I haven't mentioned? System managers are responsible for the day to day op- The best way for me to communicate with you is via eration of the system. They perform backups, install electronic mail. My address is at the end of the column. software, install revisions, install new hardware, send I will collect comments, questions, criticisms and anec- hardware out for repair, make minor repairs in-house, dotes and use them in future columns (with your prior prepare strategic plans, keep up-to-date on hardware permission, of course). and software trends, manage the department's software licenses, track down bugs in hardware and software, At the ASA meeting in San Francisco this Au-

8 Statistical Computing and Statistical Graphics Newsletter April 1993 gust, I will be hosting a roundtable luncheon on the tion such as a cassette tape. Many manufacturers can topic of Departmental Computing. Please join us! make cassette tapes, and they would pay a royalty to the patent holder. Michael Conlon University of Florida Under U.S. patent law, it appears that a mathematical [email protected] algorithm, de®ned as a ªprocedure for solving a given

type of mathematical problem,º [5] is not patentable,

while a particular application of an algorithm may be

FROM OUR CHAIRS ::: patentable. For example, a patent would not be issued for a factor analysis, but a patent could be issued for a CONTINUED FROM PAGE 1 speci®c implementation in computer code that does fac- out in the 1976 copyright law: protection is available tor analysis. Different computer code using the same for ªoriginal works of authorship:::®xed in any tangible computational algorithm may not be in violation of the medium of expression, now known or later developed, patent. from which they can be perceived, reproduced, or other- According to Greg Aharonian, over 9000 software wise communicated, either directly or with the aid of a patents have been issued, with over 1300 issued in 1992 machine or deviceº [3]. Originality does not imply that alone, mostly to a few large companies. Many of the the work must be novel or unique, only that it originate patents are issued for code that serves a speci®c purpose. with the author. Thus, for example, a computer program For example, IBM was issued patents for a ªComputer written as a class assignment is presumably a candidate user interface with window title bar iconsº and for a for copyright even though many other students would ªMethod for deleting a marked portion of a structured produce similar programs to perform the same task. document.º Again according to Aharonian, patent in- Some things are speci®cally excluded from copyright, fringement lawsuits are rare. He believes that most such as ªany idea, procedure, process, system, method patents are obtained as a defensive measure to protect of operation, concept, principle, or discovery, regard- against future law suits, not to prosecute others. less of the form in which it is described, explained, il- After this bit of research, I guess I now understand a lustrated, or embodied :::º [4]. I would interpret this to program that uses a patented algorithm for computing a mean that one could copyright a program that computes graph: the statement is advertising hyperbole. Patent- a singular value decomposition, for example, but one ing is no substitute for peer review, and if the graph is could not copyright the singular value decomposition of use, and I understand it, I am free to write my own itself. code to compute it. The owner of a copyright has authority to sell or other- References wise distribute copies of the program. If I buy a copy of Systat, for example, I have whatever rights are laid out [1] Wadley, James B. (1986), ªAn Introduction to copy- in the agreement between me and the copyright owner. right protection of computer programs and the semi- The license for Systat allows me to use the program on conductor chip protection act of 1984," Journal of the one CPU for its intended purpose of statistical analysis. Kansas Bar Association, July, 185-190. I do not have the right to reverse engineer the software [2] United States Constitution, Article 1, Section 8. to ®nd out how it works. I can sell my copy, but the [3] 17 U.S.C. 102(a). license agreement will apply to the buyer as well. If [4] 17 U.S.C. 102(b). IweretouseSystat with my own data, output will be [5] 1106 OG 6 produced in a format that may be unique to Systat;for Sanford Weisberg example, I might get a graph that could not be produced Chair, Statistical Computing Section in any other way. Copyright law speci®cally excludes [email protected] forms such as bank checks or date books from copyright, so presumably the graph produced in Systat would be owned be me, not by Systat, Inc. Statistical Graphics Whereas a copyright gives its holder exclusive use of a Hello. I know you're busy, so I won't run on too long creation, a patent gives its holder the right to claim roy- here at the keyboard. I just wanted to let you know alties to an invention. Patents are generally not intended that the Statistical Graphics Section has many activities to limit use of an invention, only to require payment for planned for 1993 and that you are invited to take part. use. For example, a patent could be issued for an inven- The easiest way to particiate is to come to the Joint Sta-

April 1993 Statistical Computing and Statistical Graphics Newsletter 9 tistical Meetings in San Francisco; it will be a prime but in actuality only 11 per second are possible at 4.0 focus for graphics events. I'd like to tell you about a MFLOPS. The performance of other machines is given few of the things we have planned. in Table 1, selected from 671 entries in Dongarra (1993).

Research contributions are always an important compo- The 100 100 LINPACK benchmark uses standard nent of the meetings, and David Scott has put together a compiler options. For example, the -dalign options strong set of invited paper sessionsÐthe list appears on on SUNs, which forces the alignment of each double page 21. If you would enjoy stimulating conversation precision datum on an 8-byte boundary, can yield faster on graphical topics, try one of the luncheon roundta- speeds. Signi®cantly faster speeds can be achieved if bles. Sally Morton has been working to catalog, dis- the code itself is modi®ed. These attempts are recorded

tribute, and display our growing collection of graphics against the 1000 1000 LINPACK benchmark in Table video tapes. The newest and best of these will be shown 1 (col.2). Considerable ingenuity has gone into opti-

in a screening room at the meetings. Another feature mizing 1000 1000 LINPACK performance, and it is of this year's graphics program is a continuing educa- apparent from Table 1 that a two to three-fold increase tion course that we are sponsoring: Bill Cleveland on in performance or better is common, which then ap- ªVisualizing Dataº. proaches nominal performance (col.3).

Every other year the graphics section sponsors a poster Table 1: LINPACK (Cholesky decomposition)

session where any interested person or group is invited performance in MFLOPS for to analyze a common dataset. As a participant, it's al- 100 100 and 1000 1000 matrices ways stimulating to be able to discuss a project with Computer 100 1000 CPU others who have looked carefully at the same data; as CRAY Y-MP C90 (16 proc.) 479 9715 15238 a spectator, it is interesting to compare and contrast the CRAY Y-MP C90 (1 proc.) 387 874 952 different approaches to the data. This year, David Cole- NEC SX-3/14R (1 proc.) 368 5199 6400 man has provided two datasets for the session: one on DEC 10000-610 (200 MHz) 43 112 200 nutritional content of breakfast cereals and another one HP 9000/735 (99 MHz) 41 107 198 a time-series with surprises. Make sure you show up to IBM 6000-560 (50 MHz) 31 84 100 see the posters this year, and make a note to do the anal- DEC 3000-500 (150 MHz) 30 80 150 ysis and contribute a poster of your own for the 1995 SGI R4000 (50 MHz) 16 32 50 IBM 6000-530 (25MHz) 15 42 50 meetings. SUN SPARC 10/30 (36MHz) 9.3 Keep on the lookout for the student poster competition SUN SPARC 2 4.0 that we sponsor with the Statistical Education section. Apple MAC Quadra 700 1.4 The work of these school children is always refreshing NeXTCube 1.4 and often displays data in a very creative way. Compaq Deskpro 486 1.3 nCube 2 (1024 proc.) 258 2409 Under the new ASA constitution, the sections have nCube 2 (32 proc.) 46 75 much more autonomy to provide services for their nCube 2 (1 proc.) 0.78 2.02 2.35 members. The statistical graphics section can ac- Apple Mac IIfx 0.37 complish a lot, but we need to know what activi- Apple Mac IIsi 0.19 ties would be of most interest to you. Let me or Compaq 386/20 w/387 0.16 any of the other section of®cers know if you have a microVAX II 0.12 good idea that the section can pursue. Also, please IBM AT w/80287 0.012 get involvedÐwe are always looking for volunteers The initial goal of this article is to elaborate further on who would like to play a role in section activies. the gap between nominal and actual performance. The Rick Becker key issue is memory management, discussed in the next Chair, Statistical Graphics Section section. The ef®cient use of memory is illustrated for [email protected] the apparently straightforward task of matrix multipli-

cationintheMatrix Multiplication section. The details

are illuminating, and include description of the ªele-

SAXPY, GAXPY, ::: mentary operationsº of vector and matrix arithmetic, saxpy and gaxpy. Two other, very effective tech- CONTINUED FROM PAGE 1 niques for speeding computations are block algorithms

Nominally, around 71 100 100 Cholesky decompo- and loop unrolling, discussed in the Performance Issues sitions can be calculated each second at 25 MFLOPS, section.

10 Statistical Computing and Statistical Graphics Newsletter April 1993 In programming the LINPACK subroutines for linear benchmark which utilizes IBM's ESSL library imple- algebra, Dongarra et al. (1979) noted that ef®cient code mentation of BLAS. On the other hand, on a molecular could be achieved through a modular approach. That is, dynamics benchmark, the price/performance ratio of a an ef®cient and numerically accurate algorithm for, say, DEC, SGI, or HP workstation can be better than that of the QR decomposition or singular value decomposition, the IBM. is implemented through a series of calls to a standard Memory Management library of basic subroutines for vector and matrix manipulation, the BLAS (Basic Linear Algebra Subroutines). The fast cycle time of a CPU can only be utilized if The same is true for LAPACK (Demmel et al. (1991)), the data and instructions can be placed in CPU quickly but in a more sophisticated sense, re¯ecting the memory enough. The ®rst concern is to con®gure the system management issues discussed in this article. The BLAS with suf®cient RAM, so that paging between RAM and encompass many operations, from the most simple such disk is eliminated. As 16MB of RAM on a workstation as vector copy or vector inner product (BLAS level 1, is now standard, and 32MB is common, this will be used in LINPACK), to matrix multiplication and the so- taken for granted. The second concern is data transfer lution of triangular systems of linear equations (BLAS between RAM and CPU. Whether the memory and CPU level 3, used in LAPACK). boards connect by the system bus or a dedicated bus, the transfer rate is slow compared to the CPU speed. Thus The portable FORTRAN public-domain implementa- some memory, so-called primary cache memory, is typ- tion of BLAS has been improved by several hardware ically placed on the CPU board and integrated tightly vendors, including IBM and Silicon Graphics Inc., and with the CPU. can include some assembly coding. Dyad Software markets an assembly code version of BLAS for DOS. The BLAS routines are useful very generally: While it is quite easy to code a routine to handle such tasks as copying an array to another array, computing the inner product of two vectors, or matrix multiplication, BLAS provides these routines also. A tabular summary of the BLAS routines is given on page ??. They are carefully Figure 1 Typical Architecture coded for numerical accuracy and speed, and can often be called from C. LINPACK, LAPACK, and the portable implementation of BLAS can be obtained by ªmail orderº from netlib, using e-mail, ftp, or the X-based graphical interface program xnetlib. The xnetlib software is particularly sophisticated, as it includes the library contents hier- archically by name and by subject classi®cation, in- Figure 2 Simpli®ed Architecture cludes various search options, and follows dependen- Typical cache memory may comprise 8KB for data and cies to facilitate downloading all relevant ®les. Calls 8KB for instructions. (A greater amount of secondary to the BLAS library can then be substituted for the cache memory, 1 MB say, may be available, but access BLAS routines distributed with individual LINPACK speeds to it are not comparable to those for primary routines; LAPACK distribution keeps the two separate. cache access.) Figure 1 shows a typical con®guration; The xnetlib software can be obtained by ftp from netlib. for our purposes the simpli®ed version, Figure 2, suf- ®ces. The size of the primary cache (from now on, This section concludes with a caveat. In a user appli- simply `cache') is small. For example, the data for the cation, such as graphics, combinatorial optimization, or

100 100 LINPACK benchmark greatly exceeds the data base manipulation, even if the code is built on the

1024 double precision number (32 32 matrix) capacity BLAS library, the usage of memory may be substan- of an 8KB cache. Larger primary caches are included in tially more non-local, and include much more indirect some recently introduced workstations, but the amount addressing, than is found in typical linear algebra ap- is still nowhere the size of a modest application. plications. Then the gap between nominal and actual performance is wider still. For example, the IBM RISC An ef®cient program is one that minimizes the transfer 6000 architecture provides exceptional performance on of data between RAM and cache, through maximizing

the LINPACK benchmarks, especially the 1000 1000 the number of ¯oating point operations for each copy

April 1993 Statistical Computing and Statistical Graphics Newsletter 11 between RAM and cache. Vector-vector and matrix- generalized saxpy,

vector routines in levels 1 and 2 BLAS respectively

x + y 7! y :

A 15

n O n

perform O operations on an amount of data, A (In the saxpy the vector x is a column of , in the gaxpy

but the matrix-matrix routines in level 3 BLAS perform B

3 2 the vector x is the ®rst column of , which contains

n O n O ¯oating point operations on an amount of

the scalars .) In their Second Edition (1989), Golub data. and van Loan strongly emphasize saxpy and gaxpy, and Matrix Multiplication; saxpy and gaxpy write of saxpy as a ®fth elementary operation, along

with addition, subtraction, multiplication, and division. Consider the 3 3by3 3 matrix multiplication

They point to the increasing role that saxpy and gaxpy

= C

AB (8) has played in numerical linear algebra in the past 10-15

0 1 0 1

: :

123 :1 2 3 years, beginning with the use of saxpy in LINPACK

@ A @ A

: : : C

456 4 5 6 = (9) (Dongarra et al. 1979).

: : 789 :7 8 9 A third approach to multiplying two matrices is by an

The conventional approach is to compute the elements outer product, building the matrix in three steps,

t t t

b + a b + a b = C; of one by one, by the dot (inner) products a

1 1 2 2 3 3 16

a j A b j

a b : c =

where j is the th column of and is the th row

ij kj

ik 10 B

k of . The ®rst outer product, which can be viewed as 3

saxpy's, is depicted

c =

In more detail, ®rst initialize ij 0 and then sequen-

1 1 0 1 0 0

a b c k = ;::: ;

ij kj

tially add ik to the current ,for 1 3.

A A @ A @ @

: =

For c11 this may be depicted

1 1 0 1 0 0

A A @ A @ @

: =

The dot product, saxpy-gaxpy, and outer product algo-

= 11 rithms are the principal alternatives from among 3! 6 arrangements of the computation, eq. (??), as three

Another approach is, after initializing the elements of

j k

nested loops indexed by the subscripts i, ,and in

A b C to 0, to multiply the ®rst column of by and add 11 every order. Each saxpy in eq. (??)andeq.(??)is the result to the current ®rst column of C , then multiply

acolumn saxpy, because the x is a column vec- b the second column of A by and add the result to 21 tor. The 6 algorithms comprise 2 algorithms each based the current ®rst column of C , then multiply the third

on dot products, 2 on column saxpy's, and 2 on row b column of A by and add the result to the current ®rst

31 saxpy's. C column of C to obtain the ®nal ®rst column of .The

®rst step is depicted Performance Issues

0 1 0 1 0 1

Although each of the six algorithms for matrix multipli-

3 2

@ A @ A @ A

;

= cation involves precisely 3 multiplications and 2 3

additions, they differ markedly in the manner in which

12 data is transferred to and from memory. For the ef- C where indicates that the element of is partially com- ®cient use of cache and also in vector processing, we

puted after this step, instead of fully computed (). The prefer algorithms that both minimize data transfer and

three steps may be depicted operate on vectors of elements that are contiguous in

1 0 1 0 1

0 memory. Given that matrices are stored by column,

each dot product accesses a vector of contiguous entries

@ A @ A @ A

? ? :

= and a vector of non-contiguous entries. Column saxpy's

? y

access two vectors of contiguous entries (x and ,anda 13 scalar), and are preferable to row saxpy's, which access

Each of the three steps is called a saxpy, for ªscaler two vectors of non-contiguous entries. When three col- y

times (vector) x plus .º Algebraically, a saxpy opera- y umn saxpy's are arranged as a gaxpy, the x differ but

tion is does not. When three column saxpy's are arranged as

x + y 7! y :

14 x

an outer product, the y differ, but does not. The for- x The ®rst column of C is constructed from three saxpy mer (gaxpy) is preferable, as the need only be loaded operations. Together they are known as a gaxpy,or from memory, whereas in computing the outer product

12 Statistical Computing and Statistical Graphics Newsletter April 1993 the y must be loaded from memory, incremented, and often 32, but can be 64, for example for operations on replaced in memory. triangular matrices. These memory management issues are known as unit As is the case for ordinary matrix multiplication, block stride and vector touches. Two additional topics, block matrix multiplication can be arranged in 3! = 6 ways. algorithms and loop unrolling, are discussed shortly. In The block gaxpy approach is

summary:

C + A B 7! C

IK KJ IJ IJ 19

Unit stride. The stride of a vector is the distance be-

K J where the subscript I varies fastest, second, and

tween successive elements of the vector in mem-

A B KJ slowest. When each matrix multiplication IK is ory. Unit stride is preferred. A column saxpy has computed using a gaxpy, there are six subscripts, which

unit stride, a row saxpy does not.

k j I K J vary in the order i (fastest), , , , ,to (slowest).

Vector touches. A vector touch is the loading of a vec- J The cache hit rate is particularly high if, for each I , ,

tor of data, to or from RAM, cache or a vector

K A B C

KJ IJ and , the entire matrices IK ,,and are con- processor. A gaxpy requires approximately half tained in cache. If each matrix is 16 16, 6144 bytes

the vector touches of an outer product. of cache are used. In fact, though, a block size of 32

J K Block algorithms. When a computation is divided into is feasible with an 8KB cache: For given I , ,and ,

blocks, a greater fraction of the computation can A

and using gaxpy multiplication, the matrix IK is used

be performed using only data transfer between B

repeatedly, once for each column of KJ . The matrix

cache and the CPU, without the use of main mem- C

IJ is used repeatedly, but one column at a time, and

ory. B

the elements of KJ are referenced once only. There- A

Loop unrolling. Utilization of data currently in mem- fore cache need contain only the entire matrix IK and C

ory is enhanced when several successive itera- a column of IJ . tions, 4 say, of a loop are written out explicitly Loop unrolling (The number of iterations of the loop is decreased by the same factor.) Panzeira (1992) describes the use of loop unrolling to further reduce the ratio of memory accesses to ¯oating Interestingly, very similar issues arise in the design of point operations, Consider the gaxpy

algorithms for parallel processing. Much of the excite-

= ; ; k = ; ; ment surrounding coding for parallel processing can be do j 1 32 1;do 132 1

found in ef®cient coding for single processors. It is

t = b

widely known that it is essential to structure code care- kj

= ; ; fully if it is to be vectorized for a multiple processor, or do i 1 32 1

vector, computer. But similar concerns apply to single

c = c + a t

ij ij processor architecture, as vector operations may be used ik to move data into and out of cache. end do Block algorithms end do; end do

3 The inner loop is unrolled four times as follows:

In general the full O operations to multiply two

n n

= ; ; k = ; ;

matrices cannot be accomplished with all three do j 1 32 1;do 132 1

B C

matrices, A, ,and entirely in cache, at least for large

t = b

matrices. However, it is possible to substantially reduce kj

= ; ; the number of memory loads and stores through the use do i 1 29 4

of operations on submatrices, that is, block algorithms.

c = c + a t

i+ j i+ j

+ k

The ¯oating point operations are unchanged in number, 0 0 i 0

c = c + a t

i+ j i+ j

+ k

but their order is rearranged. 1 1 i 1

c = c + a t

A B C IK KJ

i+ j i+ j

i+ k

KJ IJ

Let IK ,and denote the th, th and 2 2 2

A B C

IJ th block of , and respectively. Then

c = c + a t

i+ j i+ j

3 3 3 k

C = A B :

IK KJ

IJ 18 end do K end do; end do

Apart from the last row and column of submatrices, each

A A

submatrix IK of , say, will have the same number of Inner loop unrolling reduces the looping overhead but rows and of columns, the block size. The block size is not memory accesses. Thus the performance gains are

April 1993 Statistical Computing and Statistical Graphics Newsletter 13 not signi®cant. Panzeira reports that for 50 50 ma- solvers. Electronic mail requests can be sent to netlib trices actual gaxpy performance is 31% of theoretical [email protected], with a message such as (CPU) speed. With inner loop unrolling ef®ciency is 33% of theoretical speed. For matrix multiplication send index with matrices of 250 or more rows and columns, ef®- send index from LAPACK ciency increases from 24% to 25%. With blocking, the send index from BLAS increase in ef®ciency is from 31% to 33% for matrices send index from benchmark of all sizes. Alternatively software can be downloaded interactively Middle and outer loop unrolling is much more success- using ftp to research.att.com (login as user ftp), ful. Unrolling the inner loop twice, the middle loop four or the utility xnetlib. times, and the outer loop twice, the code becomes:

BLAS routines are in sets of 2 or 4, with suf®ces S and D

= ; ; k = ; ;

do j 1 31 2;do 129 4 for single and double precision real data, and C and Z for

= b t = b

t ; single and double precision complex data. Tables 2-4

+ j k + j 00 k 0 10 1

show the contents of BLAS at level 1 (vector-vector op-

= b t = b

t ;

+ j k + j

20 k 2 30 3 erations), level 2 (matrix-vector operations), and level

= b t = b

t ;

+ j + k + j + 01 k 0 1 11 1 1 3 (matrix-matrix operations) respectively. These sum-

maries are intended to show what routine programming

= b t = b

t ;

+ j + k + j + 21 k 2 1 31 3 1

tasks can be circumvented using the BLAS library.

= ; ;

do i 1 31 2

= c + a t + a t + a t +

c Table 2: BLAS LEVEL 1,

ij ij

ik+ ik+ ik 00 110 220

Vector-Vector Operations

a t +

ik 330

7! y

copy scopy x

c = c +a t + a t + a t +

x $ y

ij+ ij+

ik+ ik+

11ik 01 111 221 swap sswap

x 7! x

scale sscal

a t +

ik 331

x + y 7! x

saxpy saxpy

t c = c + a + a t +

x y

i+ j i+ j

k k + + i+

1 1 i 1 00 1 1 10 dot product sdot

t x

Euclidean norm snrm2 x

a t + a t

k + k + + i+

i 1 2 20 1 3 30

L jx j

1 norm sasum i

t c = c + a + a t +

j + j + i+ i+

k k + + i+ 1 1 1 1 i 1 01 1 1 11 maximum index isamax index of max

planar rotations drot planar rotations

a t + a t

k + k + + i+ i 1 2 21 1 3 31 end do Table 3: BLAS LEVEL 2

end do; end do Matrix-Vector Operations

+ 7! multiplication (gaxpy) dgemv Ax y y

Ef®ciency is now around 70% and 51% for the 50 50 7!

triangular matrices dtrmv T x x

and 250+ 250 matrices respectively, and 69% and = back substitution dtrsv solve T x b

65% respectively with blocking. Thus using loop un-

A + A rank 1 matrix update dger xy 7!

rolling, actual speeds can be at least doubled, e.g. from t

+ 7! A symmetric rank 1 update dsyr A xx

31% to 65% ef®ciency. Automatic innermost loop un- t

+ + symmetric rank 2 update dsyr2 A xy

rolling has been implemented in compilers, e.g. Sil- t

7! A yx icon Graphics FORTRAN compiler (IRIX release 3.3 and above). More specialized compiler technology is Table 4: BLAS LEVEL 3, needed to implement other types of automatic loop un- Matrix-Matrix Operations rolling. For example, the SGI Power FORTRAN Ana- Matrix Multiplication

lyzer implements automatic outermost loop unrolling.

+ C 7! C

general matrices dgemm AB

+ C 7! C

Software for Linear Algebra: An Overview one symmetric matrix dsymm AB

+ C 7! C one triangular matrix dtrmm T B Software for linear algebra includes libraries of high Solving Triangular Systems

level routines, such as LINPACK and LAPACK, and

= B back substitution dtrsm solve TX

also lower level routines, the BLAS library. CORE- k

Rank k and Rank 2 Matrix Updates

C + AA 7! C

MATH and netlib are repositories for sets of routines, rank k matrix update dsyrk

+ AB +

such as EISPACK, LINPACK, LAPACK, the generic symmetric rank dsyr2k C

B A 7! C version of BLAS, optimizers, and nonlinear equation 2 k update

14 Statistical Computing and Statistical Graphics Newsletter April 1993 The BLAS routines are carefully coded, sometimes in LAPACK, and Olivier Schreiber at Silicon Graphics assembly code, and, particularly for BLAS level 3, may Inc. for information on their work with the LINPACK utilize loop unrolling, blocking, and parallel processing benchmarks and compilers. This work is based in part Ð as well as saxpy's and gaxpy's Ð to take advantage on Goodall (1993), and was supported by NSF Grant of modern machine architecture. (On rare occasions the DMS-9208656.

use of higher level BLAS routines is not optimal, e.g. to References

p invert a p triangular matrix, dtrsm from BLAS is Anderson, E., Bai, Z., Bischof, C., Demmel, J., Don- inferior to dtrtri from LAPACK.) Original sources garra, J., Du Croz, J., Greenbaum, A., Hammarling, for BLAS are Lawson et al. (1979) and Dongarra et S., McKenney, A., Ostrouchov, S., and Sorensen, D. al. (1988, 1989). Coleman and van Loan (1988) give (1992). LAPACK Users' Guide. Philadelphia, PA: examples of the use of the BLAS, emphasizing level 1. SIAM. Anderson et al. (1992, Appendix C) give a summary of Coleman, T.F. and Van Loan, C. (1988). Handbook for the BLAS. Other summaries can be found in UNIX man Matrix Computations. Philadelphia, PA: SIAM. pages distributed with the software. The summary given Demmel, J.W., Dongarra, J.J., and Kahan, W. (1991). here includes only some of the routine names, but other ªOn designing portable high performance numerical routines are closely related and their names can be found libraries. LAPACK working notes no. 39. Univer- from the references, or the man pages (which may com- sity of Tennessee. (to netlib: send lawn39 from bine descriptions of several routines on a single page). lapack) LINPACK uses only BLAS level 1 routines, but can be Dongarra, J.J. (1993). ªPerformance of various comput- re-coded quite easily to take advantage of BLAS level ers using standard linear equations software.º Tech- 2. However, performance can then diminish as the size nical report CS-89-85, University of Tennessee and of the problem increases. LAPACK uses the more ef- Oak Ridge National Laboratory. 1 March 1993. (to ®cient BLAS level 3 routines, although some tuning netlib: send performance from benchmark) of LAPACK is possible for optimal performance. The Dongarra, J.J, Moler, C.B., Bunch, J.R., and Stewart, LAPACK auxiliary enquiry function ILAENV, provides G.W. (1979). LINPACK Users' Guide. Philadelphia, a lookup table of parameters, such as block size, for PA: SIAM. each primary function. The generic ILAENV in the Dongarra, J.J., DuCroz, J., Hammarling, S., and Han- LAPACK distribution may be modi®ed for better person, R.J. (1988). ªAn extended set of Fortran basic formance on individual systems, but, as LAPACK itself linear algebra subprograms.º ACM Trans. on Math. is new, departures from the generic version are uncom- Software, 14, 1-17. (See also pp. 18-32.) mon. This is perhaps an opportunity for an excellent Dongarra, J.J., DuCroz, J., Duff, I.S., and Hammarling, and worthwhile class exercise in experimental design. S. (1989). ªA set of level 3 basic algebra subpro- See Demmel et al. (1991), Section 4. grams.º ACM Trans. on Math. Software Golub, G.H. and van Loan, C.F. (1983, 1989). Ma- The necessary BLAS routines are included when LIN- trix Computations. Baltimore, MD: Johns Hopkins PACK routines are sent from netlib; LAPACK requires University Press. Second edition, 1989. that the BLAS routines have been installed separately. Goodall, C.R. (1993), ªComputation Using the QR De- LAPACK is intended to be a replacement to LINPACK. composition.º In: Handbook in Statistics, Volume 9: LAPACK, like BLAS and unlike LINPACK and EIS- Statistical Computing (C.R.Rao,ed.).Amsterdam, PACK, consistently includes routines for all four data NL: Elsevier/North-Holland. types. A little of LINPACK's functionality is absent Lawson, C.L., Hanson, R.J., Kincaid, D., and Krough, F. from LAPACK, for example, the LINPACK routines (1979) ªBasic linear algebra subprograms for Fortran DCHUD and DCHDD to update the Cholesky factor of usage.º ACM Trans. Math. Software, 5, 308-325. a symmetric matrix when a rank 1 matrix is added or Panzeira, (1992). ªNested loops optimization.º SGI subtracted; a more general alternative is TOMS algo- technical report. rithm 686, Reichel and Gragg (1990). The LAPACK Reichel, L. and Gragg, W.B. (1990). ªAlgorithm 686: release notes (send release notes from lapack) Fortran subroutines for updating the QR decomposi- include practical guidance on using LAPACK on a num- tion,º ACM Trans. Math. Software, 16, 369-377. ber of machines, e.g. by suggesting compiler options. Acknowledgements Colin Goodall I wish to thank Jim Rosenberger for suggesting I write The Pennsylvania State University this article, Jack Dongarra for advice on the use of [email protected]

April 1993 Statistical Computing and Statistical Graphics Newsletter 15 COMPUTER COMMUNICATION AND NET connected to the Internet. SNOOPING The collection of all information and services available through gopher is called gopherspace. Gopherspace Gopher and other is still growing but is already vast with hundreds of gopher-servers worldwide. Once you are connected to resource discovery tools one gopher-server you can reach other gopher-servers by selecting the menu item ªOther Gopher and Infor- The tremendous growth of services and information mation Serversº and then selecting the server you want. available over the Internet has prompted the develop- Most gopher-servers are very general and have similar ment of several tools that ease the search for these re- menu entries, while others can be focused on particular sources. Recently, I described archie in this column. information available at a given site. Archie is a tool that maintains and searches an inventory of information available at FTP sites. One disadvantage of archie is that one has to know at least a substring of Gopher is a menu driven interface the ®le name or directory in which the information is that allows browsing through Internet located to ®nd it. Recently, a reader of this column resources ::: brought gopher to my attention. I used archie to ®nd Gopher is great for browsing through information on gopher, installed a gopher-client on my machine (xgo- the Internet, especially if you have some idea which pher in my case), and then I used gopher to ®nd and gopher-servers have information that interests you. Be- retrieve information about itself and other related tools. cause gopherspace is now so vast, browsing is not an I found that most of the Internet services that I de- ef®cient way to look for information. A recent answer to scribed in previous issues of this newsletter are avail- this problem, available through gopher on most gopher- able through gopher. Some of these are: the statlib servers, is veronica (very easy rodent-oriented net-wide archive, the netlib archive, whois servers, archie,and index to computerized archives). It was developed at many more. Some new resources that I discovered with the University of Nevada and provides a keyword search gopher include: other white page servers, databases of most gopher-server menus in the entire gopherspace. (such as census data, weather data, and geographical Searching the menus with veronica is not a replacement data), library catalogs of many universities around the for browsing because the menus are not always very de- world, electronic books (such as the King James Bible, scriptive of the contents and represent only one of many Moby Dick, and Peter Pan), a wavelet archive, the AMS interpretations of the contents. For example, the same combined membership list, and many more. information could be listed as ªgeological databases,º ªrock data,º or even ªJack's favorite dataº because Jack Gopher is really more dif®cult to describe than it is to is a geologist. use, but I will try anyway. Gopher is a menu driven interface that allows browsing through Internet resources re- WAIS (Wide Area Information Server) is another tool gardless of their type and without having to worry about for ®nding resources on the Internet. Unlike guided Internet addresses. By selecting menu items one can ob- browsing through Internet resources with gopher or tain programs, documents, pictures, and even sounds, veronica searches of gopher menus, WAIS provides look up addresses, do keyword searches, and use many a facility for keyword search of documents over the other quite different Internet resources. Gopher origi- Internet. If the Internet is a collection of libraries of nated as a distributed campus information service at the documents, WAIS is an attempt to automate the interac- University of Minnesota, the ªGolden Gophers,º and tion between a library patron and a reference librarian. because it evolved to ªgo ferº things on the Internet, WAIS does not search through the actual documents, the name was coined. A recent book that contains a rather it searches through indexes built from the docu- whole chapter on gopher as well as a wealth of other ments. Only those libraries that have a WAIS index can information about the Internet is Kroll (1992). The best be searched. The uniqueness and strength of WAIS is way to learn about gopher is to ®nd a gopher-client and that it can rate the relevance of documents in terms of start using it. Telnet to consultant.micro.umn.edu,lo- similarity to other user speci®ed documents. You can gin as gopher and ªgo ferº it! This telnet-accessible get to a WAIS server with gopher, but there are some anonymous gopher-client can help you obtain the soft- advantages to running your own WAIS client program. ware required to install your own gopher-client. There Many are available from think.com, because much of the are now versions available for just about any hardware development of WAIS is done at Thinking Machines. Of and operating system combination. You only need to be course, you can also use gopher or archie to ®nd them.

16 Statistical Computing and Statistical Graphics Newsletter April 1993 Reference re-distribution point for the list. I strongly suggest that you ®nd a local feed for the S-news mailing list, or set Kroll, Ed (1992), The Whole Internet User's Guide and one up yourself. To subscribe to STAT-L, send a one Catalog, O'Reilley & Associates. line e-mail message containing just the line

The Last Word::: SUBSCRIBE STAT-L Let me end this column by saying that it is time for to [email protected]. My fourth statistical me to sign off and let someone else continue in my related bboard is the xlisp-stat-news mailing list. There place. While writing this column for the past three is also a ¯edgling mailing list for those interested in years, I have learned much about computer commu- Bayesian statistics. If you are interested in that, contact nication and various Internet services. Please contact me via e-mail. me [email protected] or the computing ed- Finding your way around netnews and the mailing lists itor [email protected] of this newsletter if you are can be daunting both for novices and experienced users interested in writing or editing a similar column or sim- alike. One way of making quick headway is to read the ply contributing to it. Also, please send any comments FAQs (Frequently Asked Questions) postings. or suggestions regarding this column to me and I will forward them to the next column editor. Many bboards have a periodic posting (often monthly) containing the FAQs. These posts are often long mes- George Ostrouchov sages containing answers to many questions which Oak Ridge National Laboratory are frequently discussed on the bboard. For example [email protected] there are several forums for discussing the TEX type- setting language. One of the more common ones is Getting your Facts from netnews.comp.text.tex. The FAQ posting contains information about FTP sites that carry collections FAQs of TEX macros, various public domain implementa- tions of TEX for different platforms (e.g., EmTeX for This is the ®rst of an occasional series about navigating DOS/Windows, OzTeX for the Mac) and so on. For the Internet. Since there are entire books devoted to new users the FAQ posting will often contain answers this subject (see above), I won't even try to be com- to many of the simple (and not-so-simple) questions. prehensive. Instead, I'll provide some clues on Internet For a®cionados the FAQ ®le often contains pointers to snooping for the statistician. sources of detailed information. If you are considering The Internet and bitnet mailing lists and netnews sys- posting a question to a bulletin board or newsgroup (par- tem is an amazing resource. It contains technical ma- ticularly the more technical newsgroups) it is always a terial and discussions about all sorts of other things. good idea to read the FAQ ®le ®rst. At the least it saves At my university the campus electronic bulletin board you getting a number indignant ªread the FAQº replies system contains a wide collection of the Internet bit- and in the best of circumstances it gets you an answer net mailing lists (including things like the s-news mail- in minutes rather than hours or days. ing list), an electronic newswire, netnews, and a gag- One way of making quick headway is to gle of local discussion boards. I ®nd it very useful read the FAQs (Frequently Asked Ques- to be able to watch netnews and the Internet mailing lists without getting my personal mailbox even more tions) postings. clogged than it already is. I'll use the generic term Not every bulletin board or newsgroup contains a FAQ ªbboardº to refer to both netnews and the various In- post and, even worse, many campus bulletin board sys- ternet mailing lists. I regularly read a large number tem generally don't keep a months worth of back posts. of bboards (I won't admit to how many). Four of So what should you do when you can't ®nd a FAQ them are devoted to statistics. The three high vol- post? Suppose you are interested in the LISP program- ume bboards are netnews.sci.math.stat,the ming language and you can't ®nd the FAQ ®le. (There Internet S-news mailing list, and the Bitnet STAT-L is one, it is posted around the middle of every month, mailing list. Many readers will already know about in about 6 parts, with a total of over 200Kb of text, to these two mailing lists, but if you do not, here is how several lisp-related forums). A good strategy is to start to subscribe. To subscribe to S-news send a message out by subscribing to the appropriate newsgroup (in this to [email protected]. A hu- case netnews.comp.lang.lisp is a good place to start) and man will either add you to the list or suggest a local become a passive reader/scanner of the messages. If the

April 1993 Statistical Computing and Statistical Graphics Newsletter 17 FAQ ®le is posted every month, and it is not in the old Uncertainty Analysis for Computationally- messages available on your bulletin board system, then Demanding System Codes it is likely that you will see the post in the next week or two. While you are waiting you can sometimes learn a Part one of a two-part series lot by just reading the other discussions. Some degree of anthropogenic environmental change is A ®nal tidbit. The USENET community does quite a an inevitable consequence of global-scale human activ- good job of collecting FAQ ®les in one place. There ity. Both policy makers and the public are extremely are several USENET newsgroups that consist solely of interested in the possible consequences of environmen- FAQs. Some that I know of are netnews.comp.answers tal changes. The Department of Energy is particularly (many FAQs for newsgroups in the netnews comp hier- interested in the effects on energy consumption, both archy), netnews.misc.answers, netnews.news.answers, of naturally occurring environmental changes and those netnews.rec.answers, and netnews.sci.answers. resulting from human activity. Large-scale modeling efforts are currently underway to assess the possible Of the statistics related bboards, I'm only aware of a effects of potential environmental change scenarios. FAQ for the S-news mailing list. This is archived in Global estimates of energy consumption, for example, StatLib. To get a copy, send the message are obtained by integrating the results from several regional models. The regional models incorporate factors send faq from S such as regional climatology and hydrology, and include models of terrestrial ecosystems and human activities (agriculture, energy use, adaptation, etc.). Boundary to [email protected], or use gopher or conditions for regional models are derived from general FTP to access StatLib. circulation models, which are run for the climate-change scenarios of interest. (We can debate the usefulness of By adroitly using your bulletin board system you can models, but such models are in fact being used to make quickly learn a lot about many technical areas. Of substantive policy decisions and their use for this pur- course, the trick is to limit the amount of time you pose will only increase with time.) waste reading pointless material. Again, this is where it is useful to read the FAQs. In a few minutes one can of- One important use of models such as those described ten ®nd answers to most of the interesting and common above is uncertainty analysis, which refers to methods questions. So go forth and explore the bulletin boards for estimating the probability distribution of model re- and be a smart reader by sticking to the FAQs. sponse resulting from variability in its input variables. However, traditional methods of uncertainty analysis Mike Meyer (Monte Carlo simulation) fail for system models such Carnegie Mellon University as these because each component may in its own right [email protected] be a complex code with heavy computing demands. A

new approach to uncertainty analysis is required. To

address this problem, McKay et al. (1979) developed BITS FROM THE PITS a sampling method (Latin hypercube sampling) that is more ef®cient for simulation than simple random sampling. Taking another approach, Downing et al. (1985) Statistical Computing and used response surface approximations to reduce com- Graphics in Science and puting requirements. In Liebetrau and Scott (1991) and Liebetrau et al. (1993), we propose a strategy for com- Industry plex system codes that employs both response-surface approximations and ef®cient sampling strategies. The This column features statistical computing and statisti- basic idea is to develop a simpli®ed analog to the com- cal graphics activities in science and industry. I invite plex system code, which we shall call a performance your comments and suggestions for future columns. assessment (PA) code, by approximating its computa- Please send comments, inquiries, and suggestions to tionally demanding components. The PA code is ab- Albert M. Liebetrau, Analytic Sciences Department, stracted from the underlying code so as to preserve the Battelle-Northwest, MS K7-34, P.O. Box 999, Rich- essential features of component processes and the in- land, WA 99352, AM [email protected], teractions among them. The PA code can be used for 509-375-2694. uncertainty analysis because it requires less computing

18 Statistical Computing and Statistical Graphics Newsletter April 1993 resources than its underlying analog. interest. Focus groups might appear inappropriately Our strategy for developing PA codes has three basic distant a topic for a column on geographic information elements. These are (i) ef®cient selection of an initial systems (GIS). But anyone who has tried to use a GIS set of model inputs (realizations), (ii) development of intended for more than pedagogic purposes will appre- an approximation to the response surface of the under- ciate its relevance. Developers of GIS software have lying codes, and (iii) an updating algorithm that uses paid scant attention to both interface design and the ex- existing information to determine the locations (inputs) periences and opinions of actual and potential users, and for additional runs. The overall idea is to develop a their products re¯ect an appalling ignorance of human two-tiered system model that consists of a performance factors equaled only, one might argue, by the develop- assessment model that ªsits atopº an underlying model ers of videocassette recorders. That GIS developers still which is made up of the detailed component models. rely almost exclusively on after-the-fact feedback from After response surfaces for the component models have users and disgruntled letter-writers re¯ects an immature been approximated, most computations for uncertainty technology inadequately exploited by specialized devel- analysis are done at the PA level. When it is necessary opers facing limited competition in a fragmented mar- to drop down to the lower level, results of these new ketplace. That people struggle to use their products is runs are used systematically to improve approximations convincing evidence of human persistence in exploiting at the upper level. a promising technologyÐnot to mention the pro®tabil- ity of vendor-run workshops and short courses. I had no In previous columns, I have described some of the tools delusions that the comparatively infantile graphic nar- available to implement the approximating strategy out- ratives discussed in my last column were any better than lined here. In the second part of this article, I will the typical GIS. But group interviews seemed a useful describe our experiences in attempting to implement way to solicit critical advice in a synergetic setting. this strategy for two real-world examples. References Focus Groups Downing, D. J., R. H. Gardner and F. O. Hoff- Although the focus group is not a full-¯edged human- man. (1985). ªAn Examination of Response-Surface factors technique, human-factors experts use group in- Methodologies for Uncertainty Analysis in Assess- terviews (as they are also called) to pretest product de- ment Models.º Technometrics 27(2), 151-163. signs as well as to identify issues worth addressing with Liebetrau, A. M. and M. J. Scott. (1991). ªStrategies for more exacting testing procedures. Employed princi- Modeling the Uncertain Impacts of Climate Change.º pally in marketing and media research, the group inter- Journal of Policy Modeling 13(2), 185-204. view is a systematic method for identifying the range Liebetrau, A. M., P. D. Whitney, D. W. Engel and C. of attitudes and preferences that buyers, readers, view- A. LoPresti. (1993). ªComputational Analogues to ers, and voters have about a product, publication, TV Complex Computer-Based Codes.º Technical Re- program, or political candidate. Social scientists have port. little use for focus groups, which are wholly unreliable McKay, M. D., R. J. Beckman, and W. J. Conover. 1979. for estimating means, variances, covariances and other ªA Comparison of Three Methods for Selecting Val- fodder for the variation-explaining paradigm. But mar- ues of Input Variables in the Analysis of Output from keting experts and product designers have found group a Computer Code.º Technometrics 21(2), 239-245. interviews an effective strategy for identifying ¯aws and gaining insight about user needs. A synergy develops Albert M. Liebetrau in which participants stimulate each other to recognize Battelle Paci®c Northwest Laboratories problems and propose alternatives. It seems obvious AM [email protected] but I'll say it anyway: to generate meaningful results,

focus-group participants must re¯ect the target popula-

tion. GEOGRAPHIC INFORMATION SYSTEMS To gather useful opinions about my prototype Atlas Touring scripts, I needed groups of more or less typ- Designing the GIS ical users of geographic data or graphic software. So instead of coercing undergraduates or bribing church Interface groups, I wrote some letters, made some calls, and ar- In my last column, I promised to describe the use of fo- ranged for sessions with groups of six to ten profes- cus groups if a signi®cant number of readers expressed sionals at four sites: Syracuse University's School of

April 1993 Statistical Computing and Statistical Graphics Newsletter 19 Information Studies (faculty and doctoral students only, namic spatial-temporal maps from the ªhistorical scriptº please), the Cartographic Division of the National Ge- and a new, interactive version of a key introductory ographical Society (in Washington, D.C.), the Of®ce of graphic phrase from the correlation script. Myke nar- Geographic and Cartographic Research in the U.S. Geo- rated each demonstrationÐwe rehearsed this carefully, logical Survey's National Mapping Division (in Reston, for consistencyÐand then presented a number of is- VA), and the GIS applications unit at IBM Corporation's sues for discussion. After the ®rst demonstration, the research and development center at Kingston, NY. The participants were asked to discuss the issues of infor- managers with whom I dealt were cooperative and en- mativeness and coherence as well as the script's good thusiastic, and graciously recruited staff with a range points and bad points, and after the second demo they of experiences in geographic analysis, map use, and addressed issues of user interaction and customization. interface design. They bought the argument that their For each issue the protocol included a sequence of dis- commitment of valuable staff time might be treated as a cussion questions. When addressing the issue of coher- colloquium or training seminar. At IBM I even sweet- ence, for instance, Myke asked questions such as: Did ened the offer by lecturing later in the day to a somewhat the progression of graphic phrases seem logical? Did larger group on the theme of How to Lie with Maps the maps and graphs ®t together? Were the text and (University of Chicago Press, 1991)Ðan opportunity to labels useful or helpful? Our protocol also included plug my book proved dif®cult to resist, as it does now. several ªprobesº with which Myke coaxed further criticisms or suggestions from the participants. The probes Developers of GIS software have paid included: Can you give me a speci®c example? Why scant attention to both interface design do you feel that way? Does anyone feel differently? and the experiences and opinions of ac- How do you see it? Myke taped the entire session with two recorders, both in full view. As part of the uni- tual and potential users, ::: versity's human-subjects review process, we prepared I didn't conduct the focus groups myself; to do so would a statement, read to participants at the outset, agreeing have tainted the results. I wanted honest, frank opin- to hide their identity and to destroy the tapes when ®n- ions about the prototype graphic narratives, not nice, ished with the analysis. Our statement also noted that polite comments inhibited by the designer's presence. they should feel free to leave at any time for any rea- Although I could have tried to avoid defensive reponses son. Fortunately for us no one did. Then came the hard to criticism, I ®nd body languageÐthe crossed arms, the part. Interpreting the results required hours of listening shocked look, the odd twitch, the higher pitch, the long to tapes and taking and comparing notes. We found it pauseÐnearly impossible to control. So early in the helpful to listen to the tapes separately, record our own study I engaged an experienced focus-group ªfacilita- observations, exchange notes, listen to the tapes again, tor,º Myke Gluck, then a doctoral student in information and discuss the results. One of us then wrote an initial studies. (Myke defended his dissertation on December, draft, which the other modi®ed extensively. and moved on to a faculty position in the School of Findings Library and Information Studies at Florida State Uni- versity, in Tallahassee.) I accompanied Myke only on Although communications and marketing researchers the visit to IBM, and while he ran the session, I hid out in often carry out a more systematic content analysisÐ an of®ce. Myke and I collaborated in developing a par- after having their tapes transcribed and coded by spe- ticipant questionnaire and a focus-group protocol. The cially trained assistantsÐthat level of rigor is more questionnaire helped both to verify that the participants appropriate to studies comparing two or more designs did in fact re¯ect our target population and to reveal or products than to exploratory studies concerned, like backgrounds and viewpoints useful in interpreting in- ours, with identifying issues and pinpointing ¯aws. Our dividual responses and group dynamics. The protocol intermediate goal was two lists, one registering widely promoted uniformity among the four groups and guar- shared complaints or suggestions to be addressed in the anteed ef®cient use of the hour or so participants and next design cycle and the other inventorying less fre- their supervisors could commit to the session. quent yet nonetheless worthwhile hints and options. In Group Interviews addition to suggestions explicitly stated by group participants, this second list typically includes the investi- We divided the group interview into two parts, the ®rst to gators' own ideas spontaneously triggered by the taped address a demonstration of the full ªcorrelation scriptº discussion. More important, though, are shortcomings (about 11 minutes) and the second to discuss a mixed and insights registered by two or more groups: in ad- demonstration (about 10 minutes) consisting of two dy- dition to blatantly obvious problems, the primary list

20 Statistical Computing and Statistical Graphics Newsletter April 1993 commonly includes design de®ciencies and strategies buy. So, I am looking for some contributing assistance, affecting only a minority of potential users. How ef- correspondents who can report on reviews of statistical fectively a software designer appreciates and addresses computing and graphics books that are reviewed in other these not-widely-shared concerns can be crucial in de- journals. Particularly I am interested in the proliferating veloping an ef®cient, effective, and broadly inclusive crop of statistical computing journals, none of which I system. Findings based on focus groups are much closer will see any more, plus Biometrics, Applied Statistics, to the soft end of the measurement-rigor-reliability con- and Shorter Book Reviews. tinuum, and the risk of bias seems markedly greater than with the more tightly controlled research strate- There were two other problems that I encountered in gies of con®rmatory analysis. Whoever reads our re- preparing this column. First, there have been fewer port (still under editorial review) might well question reviews of statistical computing and graphics books re- whether Myke and I have carefully listened to partici- cently. Second, some of the ones that have appeared pants' complaints about the demonstration software and have been written by me in the report section of Tech- openly reported their doubts and reservations about nar- nometrics Book Reviews. It seems rather ridiculous to rative graphics. Conscious or unconscious bias is al- quote myself, so I shall just remark brie¯y on those re- ways a problem when researchers rely heavily on inter- ports. In November I commented on new editions of two views, and our study is no exception. But the possibility books on statistical analysis using SAS, Sas System for of bias is cause only for caution and skepticism, not out- Regression, Freund and Littell, and Sas System for Lin- right rejection. The value of our study can be judged ear Models, Littell, Freund, and Spector. I found each to by the apparent reasonableness and thoroughness of the be signi®cantly improved over its previous edition and collective and constructive criticisms we report. an essential personal library item for any statistician who uses SAS for doing statistical analysis. In Febru- What did we learn? LotsÐfar too much to describe ary I commented on a newcomer to this collection, Sas in the current column. Would we do use focus groups System for Statistical Graphics,M.Friendly.Forme again? You betÐdespite the dif®culties of pleading this much larger book was equally valuable, especially with managers, coping with D.C. traf®c to meet a tight for its providing of extensive SAS code and macros, schedule, and taking our own Macintosh system into also for its comprehensive coverage of graphics in all and out of an IBM facility. The exercise yielded valu- types of statistical analyses. able insights, and I can't wait for the opportunity this summerÐNSF willingÐto plunge in, implement and Other statistical packages were included as central fea- re®ne some new strategies, and generate vastly im- tures of books, too. In JASA for December is a review proved demos for another round of focus-group eval- for the 2nd Edition of Data Analysis for Managers with uation. The next iteration might not yield a satisfactory Minitab, H. Roberts, Scienti®c Press. John McKen- solution, but I surely have a clearer sense of where the zie praises the book's use of real data, effort at making Atlas Touring project is headed. statistics interesting, inclusion of statistics as part of business, and use of software to analyze data, stating that ªthe author has written a book that should be exam- Mark Monmonier ined by anyone dealing with elementary business statis- Syracuse University ticsº. In comments in the Technometrics Editor Reports

[email protected] in February, I opine that the book is better suited for

teaching statisticians about business applications than BOOK REVIEW BEAT managers about statistics. GENSTAT also gets some textbook support with Applied Statistics: A Handbook of Genstat Analyses, E. Snell and H. Simpson (Chap- This column returns, with the same proprietorship, be- man and Hall). Mike Driscoll, in Technometrics for cause the new editors invited it, and because I was will- November, notes that the ªhandbook details the use of ing to continue. I had reservations, however. The previ- GENSTAT to analyze the data used for the examples ous editors always asked for the next column, but there in Snell's book, Applied Statistics: Principles and Ex- was never any feedback from anyone about whether the amplesº. For that book, he ®nds this one a ªuseful column was good or bad, useful or wasted space. I will supplementº with ªlimited value otherwiseº. need to have positive feedback before I will feel that it is worth continuing. Actually, I will need more than There are some reviews of books not about statistical that. Cost cutbacks in my corporation have taken away packages. In JASA for December, Martin David reviews from me all the statistics journals, except the ones that I a collection of papers, Statistical and Scienti®c Data

April 1993 Statistical Computing and Statistical Graphics Newsletter 21 Bases, Z. Michalewicz, editor (Ellis Horwood). Noting ªGraphical Methods for Finding Structure in Multi- the proliferation of persons who are embedding scien- variate Data,º Qiang Luo, George Mason Univer- ti®c data in data bases, he says that ªthis volume will sity, [email protected] leave them inadequately informed on the conceptual structure to use in their work and on the strategic trade- Section: Statistical GraphicsÐco-sponsor Stat Ed, offs that are possible with powerful off-the-shelf Re- Teaching Health Stat lational Data Base Management Systems (RDBMS)º. Session Title: Statistical Graphics and Animation For In Technometrics for February, Minoo Niknian reviews Instruction/Classroom Randomization and Monte Carlo Methods in Biology, Session Time: Wednesday, August 11, 10:30 - 12:20 B. Manly (Chapman and Hall), and ®nds the book ªap- Organizer: Joseph Newton propriate for master's level students of statistics and Session Chair: Joseph Newton practicing statisticians, as well as subject matter inves- Speakers: tigators with good backgrounds in statistics and expe- ªIntegrating Dynamic Graphics into a Lin- rience in computingº. He calls the book ªwelcome in ear Regression Course,º Sandy Weis- applied statisticsº. In Biometrics for September, Jim berg/Dennis Cook, University of Minnesota, Gentle reviews the same book, saying that ªthe book is [email protected] generally well written, and the examples are interesting ªExploratory Dynamic Graphics for Survival and easy to understandº. He also notes that ªthe word Data,º Neely Atkinson, Texas Medical Center, `biology' in the title should not limit the readershipº. [email protected] ªMethods for the Analysis of Data From Designed The Editors or this columnist need some feedback to Experiments,º Wei-Yin Loh, University of Wis- determine whether a look at reviews of statistical com- consin, [email protected] puting and graphics books is useful, or whether some other format should be used to deal with the books. If Section: ACM SIGGRAPH [subj to ®nal signing]Ð the overview of reviews is to remain reasonably com- Statistical Graphics, Stat Ed prehensive, some contributors for the aforementioned Session Title: Multimedia: Past, Present, and Future journals are needed. Contact Eric Ziegel or the editors. Session Time: Tuesday, August 10, 2:00 - 3:50 Organizer: Jan Pedersen, Xerox Parc Eric Ziegel Chair: Jan Pedersen Technometrics Book Reviews Editor TBA [Current Multimedia Environment], Dan Rus-

[email protected] sell, Xerox PARC, [email protected]

ªIntelligent Agents as a User-Interface NEWS CLIPPINGS Metaphor,º Tim Oren, Kaleida Labs, Inc., [email protected] ªMultimedia Futures,º Enrique Godreau, Aldus Cor- Joint Statistical Meetings poration, enrique [email protected] Discussant: William Cleveland, AT&T Bell Labs, Statistical Graphics Invited Program [email protected] Section: Statistical Graphics Session Title: Young Researchers In Statistical Graph- Section: Statistical Graphics ics Session Title: Novel Applications of Multivariate Sta- Session Time: Tuesday, August 10, 8:30 - 10:20 tistical Graphics Organizer: David W. Scott Session Time: Tuesday, August 10, 4:00 - 5:50 Session Chair: David Scott Organizer: David Scott Speakers: Chair: TO BE NAMED ªGraphics KeysÐMisha's Resource Database Ap- ªVisualizing Speech Data and Hidden Markov proach to Extensible Graphics,ºJames Hardin, Bat- Models,º Jim Schimert [speaker], Andreas telle Lab, [email protected] Buja, Werner Stuetzle, Statistical Sciences, Inc., ªThe Mode Tree: Nonparametric Visualization of [email protected] Density Features,º Michael Minnotte, Utah State ªVisualizing Multivariate Rank Data,º G. Thompson University, [email protected] and K. Baggerly, Southern Methodist University ªAssisting Inductive Modeling With Visualiza- ªVariable Resolution Bivariate Plots,º Chisheng tion,º John Elder, University of Virginia, Huang, Univ. of Washington, [email protected] [email protected]

22 Statistical Computing and Statistical Graphics Newsletter April 1993 Statistical Graphics Roundtable Luncheon organized by Janis Hardwick, University of Michigan, Discussions who will present ªComputational Analyses of Sequen- tial Allocation Problems.º Phil Spector, University of 1. ªSoftware for statistical graphics,º James California, Berkeley, will present ªParallelizing CARTº E. Gentle, George Mason University, and George Ostrouchov, Oak Ridge National Labora- [email protected] tories,will present ªStatistical Modeling Applications 2. ªSoftware for Multivariate Functional Estimation on Parallel Computers.º Quentin Stout, University of and Visualization,º David W. Scott, Rice Univer- Michigan, will be the discussant. sity, [email protected] 3. ªIssues in implementing graphical func- Sallie Keller-McNulty, Kansas State University, orga- tions,º Linda Clark, AT&T Bell Labs, nized a session that looks at nontraditional models enti- [email protected] tled ªStatistics Driven by Non-Classical Assumptions: Spatial Sampling, Estimation, and Modeling.º She will 4. ªMeaningful Statistical Graphics for Manufactur- present ªMean Field Estimation for Probability Distri- ing,º Karen Kafadar, National Cancer Institute, bution Function Methods of Turbulent Reactive Flows.º [email protected] Katherine Campbell, Los Alamos National Laborato- 5. ªSingle Frame Video Recording: The Wave of ries, will discuss spatial considerations in environmen- the Future?,º William F. Eddy, Carnegie Mellon tal sampling in ªSampling the Continuum: The Spa- University, [email protected] tial Support of Environmental Data.º Douglas Nychka, 6. ªGraphics in Bayesian Statistical Com- North Carolina State University, will discuss ªStretch- puting,º William DuMouchel, New Eng- ing and Translating Surfaces to Model the Drain Current land Biomedical Research Foundation, in a Semiconductor.º [email protected] G. Pliego of ITESL, Mexico, has organized a session Statistical Computing Invited Program ªWavelet Software and Algorithms in Statistics.º that includes an introduction to wavelets. ªComputational The Statistical Computing Section has an exciting col- Techniques for Wavelet Applicationsº is the title of Wim lection of eight invited and special contributed sessions Sweldens, University of South Carolina and Katholieke for the annual ASA meeting in San Francisco. Universiteit Leuven, Belgium. Z.Wang, Purdue Uni- The session that received the top number of votes in versity, will present ªSoftware Examples for Statistical the ASA competition for extra sessions was organized Wavelet Analysis.º Carolyn Carroll, IBM, will discuss by Ed Wegman of George Mason University and titled ªParallelizing Wavelets.º ªVirtual Reality for Exploratory Analysis.º The ses- John Maryak, The Johns Hopkins University, organized sion is a tutorial on using virtual reality to explore high ªRecursive Methods for Optimization in Time Seriesº dimensional data. Rotating a point cloud on a screen with Larry Goldstein, University of Southern Califor- to ªseeº in three dimensions is an example of the kind nia, as discussant. T.L.Lai, Stanford University, will of advances in graphical data analysis we saw in the present ªParallel Recursive Algorithms in Time Series 1980's. Virtual reality appears to offer the next step for Estimation and Control.º James Spall, The Johns Hop- the 1990's: it can be used to ªimmerseº the statistician kins University, discusses ªAccelerated Stochastic Opti- in the data world. mization in Multivariate Problems,º and John Monahan, A second session entitled ªSoftware Statisticsº orga- North Carolina State University, presents ªApplications nized by David James, AT&T Bell Laboratories, ex- of Stationary Stochastic Approximation.º amines the role of statistics in the relatively new ®eld The session ªSmoothing in Data Analysis: Tutorial of software engineering. Scott Vander Wiel, AT&T on Methods and Applicationsº is organized by Karen Bell Laboratories, will discuss ªCapture-Recapture Kafadar, National Cancer Institute, who will present and Other Statistical Methods for Software Inspection ªChoosing among Two-dimensional Smoothers in Prac- Data,º while Wendell Jones, Bell Northern Research, tice.º Trevor Hastie, AT&T Bell Laboratories, will will discuss ªModeling for Software Systems.º Also present ªWhat Smoothies Doº and Colin Goodall, Penn Bill Curtis, Carnegie-Mellon University, will discuss State University, will present ªNonlinear Smoothers for ªConundrums in Applying Statistics to Software Engi- One-Dimensional Data.º Katherine Hansen, Sandia neering.º National Laboratories, will present ªSmoothing Mul- Getting up to speed on high performance computers is tidimensional Data,º and Owen Devine, Center for Dis- the topic of the session ªHigh Performance Computingº ease Control, will act as discussant.

April 1993 Statistical Computing and Statistical Graphics Newsletter 23 Tim Hesterberg, Franklin and Marshall College, orga- 4. ªDeveloping Departmental Computingº nized the session ªWeighted Samples in Simulation,º Mike Conlon from University of Florida and he will present ªImportance Sampling and Control ([email protected]) will discuss the de- Variates for the Bootstrap.º George Easton, Univer- velopment, administration and maintenance of sity of Chicago, will discuss ªCon®gural Polysampling modern departmental computing networks. Is- and Some Related Researchº while Peter Glynn, Stan- sues of resource sharing (coordinating funds as ford University, will discuss ªRare Event Simulation for well as equipment), speci®cations (what should Queues.º the system be able to do), performance, software, Mary Ellen Bock, Program Chairman 1993 vendors, migration (acceptance and training), in- Statistical Computing Section stitutional barriers (technical and bureaucratic) and more will be discussed. Mike will consult on how to get from various starting points (stand Statistical Computing Roundtable Luncheon alone PCs, mainframes) to operational networked Discussions systems. Five interesting Luncheon Discussions are planned in 5. ªFuture Directions in Statistical Computingº San Francisco. The topics and a brief description of John Chambers from AT&T Bell Laboratories each are given below. Feel free to contact the Discus- ([email protected]) will lead this dis- sion Leaders for more information. To participate in cussion. This luncheon parallels an Invited these discussions, you need to sign-up for the luncheon Speaker Session with John, Wayne Oldford, when you register for the conference. All of these lun- Werner Stuetzle, and Dave Andrews on the same cheons will be scheduled for the same day topic. Lively discussion on our future is expected 1. ªImpact of Computing on Industrial Experimental to be found here!

Design,º Perry D. Haaland from Becton Dicken-

son Research Center ([email protected]) will be leading a discussion on a variety of issues including new results in optimal design, inference Interface `93 Meeting based on robust methods, graphical analysis, new results on dispersion modeling, Taguchi or ªclas- Highlights sicalº approaches, industrial training, and practical experiences. Perry also plans to leave some 25th Symposium on the Interface: Computing Sci- time for the sharing of ªwar stories.º ence and Statistics 2. ªUsing the Statistical Computing and Statistical Graphics Newsletter as a Vehicle to Keep Up with Statistical Applications and Expanding Computer Recent Developments in Computingº Newsletter Capabilities Editor Jim Rosenberger from Penn State Univer- April 14-17, 1993 sity ([email protected]) will lead this dis- San Diego, California cussion. This roundtable will be the perfect vehi- Pan Paci®c Hotel cle for those interested in providing input to our Keynote Speaker: David Brillinger "Statistics and Com- ever growing and evolving Section Joint Newslet- puting in Science" ter. Sponsor: Interface Foundation of North America 3. ªThe Design and Analysis of Computer-Based Experimentsº Albert M. Liebetrau from Battelle- Hosts: Precision Data Group and University of Califor- Northwest (AM [email protected]) nia, Berkeley plans to cover this topic by encompassing the Cooperating Societies and Institutions: work of the ªSandiaº School (Iman and Conover American Statistical Association (ASA); Institute of latin hypercube sampling approach), the ªOak Mathematical Statistics (IMS); Society for Industrial Ridgeº School (Worley, Oblow, Pin and oth- & Applied Mathematics (SIAM); Operations Research ers), and the ªSecond Oak Ridgeº School (Sacks, Society of America (ORSA); The Biometrics Soci- Welsh, Morris, and Mitchell). Discussion will ety (WNAR); University of California, Berkeley; San also cover the recent theoretical work by Art Diego State University, San Diego; Northern and South- Owen on latin hypercube designs and orthogonal ern California Chapters of the American Statistical arrays. Association.

24 Statistical Computing and Statistical Graphics Newsletter April 1993 Invited Sessions include: charged a late fee of $20. The registration fee covers the Data Compression; Computing with Environmental reception, coffee breaks, banquet, and the proceedings. Data; Biopharmaceutical Maps and Graphics; Clini- Please make checks payable to Interface '93. cal Trials; Protein Structure; Digital Networks; User Interfaces; Geosciences; Software Engineering and GENERAL INFORMATION Statistical Methods; The Interface at 25; Likelihood Accommodations: All meetings will be held at The Applications; Library Systems; Medical Applications; Pan Paci®c Hotel in downtown San Diego, California. Multivariate Function Estimation; Networked Informa- Conference room rates of $99 per room have been ar- tion Systems;Supercomputers; Time Series Analysis; ranged for all attendees. Early registration is strongly Wavelets; Computers and Statistics in Drug Discovery; recommended to assure a room. Please contact the hotel Quality Data Bases. directly for reservations. Inquiries should be sent to: THE PAN PACIFIC HOTEL, 402 West Broadway, San Interface '93 Diego, CA 92102-3580. Telephone (619)239-4500 or Michael E. Tarter, Program Chair (800)626-3988. 140 Warren Hall Conference rate for Interface '93 participants honored University of California, Berkeley until March 14, 1993. Rate can be extended 3 days prior Berkeley, CA 94720 and/or following the conference. Rates are $99 for ei- (510) 642-4601 ther single or double room. Reservations held until 4 [email protected] p.m. without deposit or accepted credit card. Conference proceedings of invited and contributed pa- Air Fare: At the present time, no airlines are offer- pers will be published. Camera-ready copy will be due ing discount association fares. Vineyard Travel in San on June 1, 1993. Diego will be happy to assist you with your travel plans. Their telephone number is (619) 741-6669. Rosemary Conference Schedule: The conference begins on Nigro is our Corporate Travel Consultant. Wednesday evening, April 14, with a get-acquainted reception. Technical sessions will be held Thursday San Diego International Airport: San Diego Interna- and Friday with a banquet Thursday evening. There tional Airport is served by most major airlines. Shuttle will be ®nal technical sessions Saturday morning. bus and limousine service are available to the Pan Pa- ci®c Hotel in downtown San Diego. Short Courses: Program Committee: Lynne Billard, Mary Ellen Bock,

SYSTAT: Data Analysis ± 2 days prior to the con- Noel Cressie, Arnold Goodman, Sam Greenhouse, Jon ference, Monday, April 12 and Tuesday, April Kettenring, Diane Lampert, Michael Lock, Bob New- 13. SYSTAT produces statistical software de- comb, Joseph Newton, John Rice, Ernest Scheuer, Bob signed for data analysis. Registration for this Shumway, Michael Tarter, Grace Wahba, Edward Weg- short course will be handled separately. man

BMDP: A New Windows-Based Statistical Anal- Interface Foundation: The Interface Symposium is an ysis Environment will be presented Wednesday activity of the Interface Foundation of North Amer- afternoon, April 14. There is no fee for this ica, a nonpro®t, educational corporation founded in Au- course, however, please note on the registration gust of 1987. The aim of the IFNA is to promote the card if you plan to attend. Interface Symposium and related activities at the in- Randomization Tests: Jenny Baglivo, Marcello terface of computing science and statistics. The 25th Pagano, and Cathie Spino will offer a short course Symposium is the sixth held under the auspices of the in Randomization Tests: Theory and Practice. Interface Foundation. The next symposium will be There is no fee for this course. Please note on the hosted by the SAS institute in North Carolina in 1994.

registration card if you plan to attend.

Registration: The registration fee is $165 for members of the cooperating societies, ASA, IMS, SIAM, ORSA, the Biometrics Society (ENAR and WNAR), and for persons af®liated with University of California, Berke- ley. The fee is $55 for students; for others it is $185. Registrations received after February 15, 1993 will be

April 1993 Statistical Computing and Statistical Graphics Newsletter 25 Lieutenant General William E. Odom, announced the News from NSF expansion and redirection of Program OCREAE, NSA's February 16, 1993 grants program for research in Cryptology and related Dear Colleague: areas. This effort is currently being vigorously sup- Announcement of Proposal Target Dates ported by the Agency. In order to improve the Division's proposal manage- The grant proposals submitted to the program are re- ment, and possibly to employ disciplinary panels in the viewed by the NSA Mathematics Review Panel which merit review of proposals, the Division of Mathemati- is appointed and administered by the American Math- cal Sciences plans to introduce target dates for proposal ematical Society. Under the guidance of this panel, submission for disciplinary research activities for FY the program has been particularly interested in support- 1994 NSF funds. ing promising young investigators with small summer salary grants. The program has also directed a sig- Beginning in the fall of 1993 the Division will introduce ni®cant portion of its funds to senior investigators for two target dates for proposals submitted to the follow- the support of their graduate students and to univer- ing programs: sity departments for the support of special conferences Oct. 22, 1993 Algebra and Number Theory Program and workshops. In an attempt to formalize these ob- Classical Analysis Program jectives, the MSP program now offers funding in four Modern Analysis Program distinct categories; the Young Investigators Grant,the Topology and Foundations Program Standard Grant,theSenior Investigators Grant,and Nov. 19, 1993 Applied Mathematics Program the Conferences, Workshops, and Special Situations Computational Mathematics Program Grants. Geometric Analysis Program Firm deadline of October 15, 1993, for all proposals Statistics and Probability Program except conferences. Funding as soon as possible after These dates do not overlap substantially with other October 1, 1994. known Foundation target dates, mesh reasonably well Conference proposals accepted any time. Allow 8 with academic calendars, and cluster the programs so months for funding. as to provide a balance with respect to both overlapping scienti®c content and anticipated program proposal Further information on grants is available by calling loads. (301) 688-0400 or writing: Proposals which miss the target dates will be handled Dr. Charles F. Osgood, Director as time permits. Priority will be given to proposals NSA Mathematical Sciences Program arriving on or before the above target dates. National Security Agency Attention: R51A The above dates do not apply to the activities of the Ft. George G. Meade, MD 20755-6000 Division's Of®ce of Special Projects. These activities [email protected]

already have speci®ed target or deadline dates.

Sincerely, Frederic Y. M. Wan Electronic White House Division Director Division of Mathematical Sciences There has been a lot of recent information and disinfor-

mation about the Clinton-Gore electronic White House.

We have seen various postulated e-mail addresses for contacting the new president of the United States, but News from NSA none of the addresses is completely satisfactory. One ANNOUNCEMENT common outcome of e-mail to the President is a reply, via regular paper mailÐnot quite what one would ex- The NSA Mathematical Sciences Program continues its

pect! The most common explanation? ::: lack of funds efforts at funding high quality mathematical research in to buy computers and connect to the outside world. With the areas of Algebra, Number Theory, Discrete Math- that cleared up, it is still worth reporting some of the ematics, Probability, Statistics and Cryptology. The ways of contacting the white cottage. program, in its present form, had its beginning in 1987 when the then director of the National Security Agency, The new Clinton-Gore Administration has several elec-

26 Statistical Computing and Statistical Graphics Newsletter April 1993 tronic mail addresses. The MCI Mail box address (see cussion/®le areas related to the new administration: below) and bulletin board have received the most pub- licity. Compuserve: 75300,3115 (e-mail); GO: WHITE HOUSE (White House forum) GEnie has a PF (Public Forum) section which carries

America OnLine: clinton pz (e-mail); KEY- MIT-generated ®les from White House press brie®ngs WORD: WHITEHOUSE (White House area) and speeches, but that forum offers no e-mail feedback to the Clinton Administration. It serves the public in- MCI; WHITE HOUSE (e-mail); VIEW WHITE terest by making important text ®les and other public HOUSE (views bulletin boards)

information widely available. Internet e-mail address: What if you really want to get a message through to [email protected]; Washington? In the past this meant letters or telegrams, [email protected]; but now you can save paper and attempt direct electronic [email protected] contact in several ways. We haven't personally tried any of these addresses or The following mailbox addresses are reported to work discussion forums. Perhaps our readers could give us for sending e-mail to the White House and entering dis- some feedback!

1994 Joint Statistical Meetings in Toronto Canada As Statistical Computing Program Chair for 1994, I am soliciting your help. Now is the time to give me suggestions for topics, speakers, and organizers for our invited sessions. If you have any ideas along these lines please contact me before or at this year's Joint Meetings in San Francisco.

Sallie Keller-McNulty, Kansas State University [email protected] 913-532-6883

April 1993 Statistical Computing and Statistical Graphics Newsletter 27 The Statistical Computing and Statistical Graphics Newsletter is a publication of the Statistical Computing and Statistical Graphics Sections of the ASA. All communications regarding this publication should be addressed to:

James L. Rosenberger Michael M. Meyer Editor, Statistical Computing Section Editor, Statistical Graphics Section Department of Statistics Department of Statistics The Pennsylvania State University Carnegie Mellon University University Park, PA 16802-2111 Pittsburgh, PA 15213-1380 (814) 865-1348 (412) 268-3108 [email protected] [email protected]

All communications regarding membership in the ASA and the Statistical Computing or Statistical Graphics Sections, including change of address, should be sent to: American Statistical Association 1429 Duke Street, Alexandria, VA 22314-3402 (703) 684-1221

Where to ®nd it

: : : : : : : : : : : : : : : : : : : : :

A WORD FROM OUR CHAIRS : 1 Getting your Facts from FAQs 17

: : : : : : : : : : : : : : :

Statistical Computing 1 BITS FROM THE PITS : 18

: : : : : : : : : : : : : : : :

Statistical Graphics : 9 Statistical Computing and Graphics in Science and

: : : : : : : : : : : : : : : : :

: : : : : : : : : : : : : : : : : :

FEATURE ARTICLE 1 Industry : 18

: : : : : : :

Saxpy, gaxpy, LAPACK, and BLAS : 1

: : : :

GEOGRAPHIC INFORMATION SYSTEMS : 19

: : : : : : : : : : : : : : : : : : : :

EDITORIAL : 2

: : : : : : : : : : :

Designing the GIS Interface : 19

: : : : : : : : : : : : : : : :

SECOND FEATURE : 2

: : : : : : : : : : : : : : : BOOK REVIEW BEAT : 21

Production of Stereoscopic Displays for Data Anal-

: : : : : : : : : : : : : : : : :

NEWS CLIPPINGS : 22

: : : : : : : : : : : : : : : : : : : : :

ysis : 2

: : : : : : : : : : : : :

Joint Statistical Meetings : 22

: : : : : : : : :

DEPARTMENTAL COMPUTING : 7

: : : : : : : :

Interface 93 ± Meeting Highlights : 24

: : : : : : : : :

Not just hardware and software : 7

: : : : : : : : : : : : : : : : :

COMPUTER COMMUNICATION AND NET News from NSF : 26

: : : : : : : : : : : : : : : : : :

SNOOPING : 16 News from NSA 26

: : : : : : : : : : : : : : : : : : Gopher and other resource discovery tools : 16 Electronic White House 26