Tragic Loss or Good Riddance?

The Impending Demise of Traditional Scholarly Journals
FULL VERSION

Andrew M. Odlyzko
AT&T Bell Laboratories
amo@research.att.com

October 24, 1994

To be published in Intern. J. Human-Computer Studies (formerly Intern. J. Man-Machine Studies) and reprinted in Electronic Publishing Confronts Academia: The Agenda for the Year 2000, Robin P. Peek and Gregory B. Newby, eds., MIT Press/ASIS monograph, MIT Press, 1995. Condensed version to appear in Notices Amer. Math. Soc., Jan. 1995.

Page 3

1 Introduction

Traditional printed journals are a familiar and comfortable aspect of scholarly work. They have been the primary means of communicating research results, and as such have performed an invaluable service. However, they are an awkward artifact, although a highly developed one, of the print technology that was the only means available over the last few centuries for large-scale communication. The growth of the scholarly literature, together with the rapidly increasing power and availability of electronic technology, are creating tremendous pressures for change. The purpose of this article is to give a broad picture of these pressures and to argue that the coming changes may be abrupt.

It is often thought that changes will be incremental, with perhaps a few electronic journals appearing and further use of email, ftp, etc. My guess is that change will be far more drastic. Traditional scholarly journals will likely disappear within 10 to 20 years. and the electronic alteratives will be different from current periodicals, even though they may carry the same titles. There are obvious dangers in discontinuous change away from a system that has served the scholarly community well [Quinn]. However, I am convinced that future systems of communication will be much better than the traditional journals. Although the transition may be painful, there is the promise of a substantial increase in the effectiveness of scholarly work. Publications delays will disappear, and reliability of the literature will increase with opportunities to add comments to papers and attach references to later works that cite them. This promise of improved communication is especially likely to be realized if we are aware of the issues, and plan the evolution away from the present system as early as possible. In any event, we do not have much choice since drastic change is inevitable no matter what our preferences are.

Predictions and comments in this article apply to most scholarly disciplines. However, I will write primarily about mathematics, since I am most familiar with that field and the data that I have is clearest for it. Different areas have different needs and cultures and are likely to follow somewhat different paths in the evolution of their communications.

The rest of this section is devoted to an outline of the paper. The impending changes in scholarly publications are caused by the confluence of two trends. One is the growth in the size of the scholarly literature, the other is the growth of electronic technology.

[Section 2] discusses the first factor that is forcing a change in traditional publishing. The number of scientific papers published annually has been doubling every 10-15 years for the last two centuries. Growth has stopped recently, but this is likely to be a temporary pause of the kind that have occurred before.

The exponential growth in scholarly publishing has interesting implications. When the number of articles per year doubles in 10 years, as it did for several decades after World War 2 in many fields, about half of the entire literature will have been produced during the preceding 10 years. Even if growth then stops suddenly, the total number of articles will double in another 20 years. Since library costs are already high, there is bound to be pressure to change the system of scholarly publications.

[Section 3] describes one of the electronic information dissemination systems that are becoming widespread. Their costs are negligigble compared to those of traditional print journals, and

Page 4

[Section 4] and [Section 5] discuss the technological trends that are making new methods of publication possible. Processor and transmission speeds are increasing at rates far higher than the growth rates of scholarly literature. This will make a change to electronic publishing feasible in the next few years.

[Section 6] describes some of the recently established electronic journals. Most are operated by scholars and do not charge fees for access, and thus solve the cost and availability problem of print journals.

[Section 7] argues that although there have been many previous predictions of dramatic changes in scholarly publishing, this time it will really be different, and the change is imminent. Technology is about to provide tools for a far superior information dissemination system, and the problems with the existing one are severe enough that a change will be unavoidable.

[Section 8] is devoted to a description of future electronic publications. It is not simply that the escalating costs of print journals will force scholars to move away from paper publications. There are also novel features that can be implemented on the networks but not in print that are already pulling scholars towards electronic communication. First some of the present network communication methods are surveyed, and then projections are presented of what future systems will be like.

[Section 9] examines in detail the costs of the current system of print journals, how it might change, and how this change will affect publishers and librarians. There are many uncertainties about the precise shape of future systems, but costs will have to be reduced dramatically. This will force publishers to shrink. Libraries will also be affected. It is possible that in most fields, such as mathematics, a few dozen scholars and librarians at a central organization might be able to provide electronically all the services that a thousand reference librarians and their assistants now do.

2 Growth of literature

The impending changes in scholarly publications are caused by the confluence of two trends. Both trends have exponential growth rates (where exponential refers to the mathematical meaning of this word, not the current journalistic usage). One trend is in the size of the scholarly literature, which is causing a crisis in the traditional approach to dissemination of research results. The other trend is the increasing power and availability of electronic technology. In this section we look at the growth in publications.

Not all scholars perceive that there is a genuine crisis in the present system of journal publications. However, librarians and publishers are well aware of it. New journals are springing up all the time, yet libraries are not buying them, and are even dropping subscriptions to old journals. The blame for this crisis is usually ascribed either to greedy publishers or short-sighted administrators. The basic underlying problem, though, is the exponential growth in the scholarly literature. The number of scientific papers published annually has been doubling every 10-15 years for the last two centuries [Price]. In many fields, there was an acceleration in the rate of growth after World War 2 to a doubling every 10 years or less. For example,

Page 5

Mathematics is an old discipline, with an extensive literature accumulated over the centuries. However, what is seldom realized is that, just as with other areas, the bulk of this literature is young. In 1870 there were only about 840 papers published in mathematics [GroetschelLS]. Today, about 50,000 papers are published annually. (If you ever wonder why we no longer have mathematicians like Hilbert and Poincare, who had a comprehensive understanding of mathematics, these figures are surely a large part of the story.) The increase from 840 to 50,000 over 124 years corresponds to a doubling in the rate of publication about every 20 years. However, this growth was not even, and a more careful look at the statistics shows that during the post-World War 2 era, the number of papers published has been doubling about every 10 years [MR]. Growth has stopped recently. According to Jane Kister of Mathematical Reviews ( Math. Rev.), the number of items entered into the Math. Rev. data base reached about 57,000 per year in 1990, and has remained at that level since then (private communication). Math. Rev. has stayed at approximately 47,000 reviews per year only by limiting its coverage.

The exponential growth in mathematical publishing has interesting implications. Adding up the numbers in [MR] or simply extrapolating from the current figure of about 50,000 papers per year and a doubling every 10 years, we come to the conclusion that about 1,000,000 mathematical papers have ever been published. What is much more surprising to most people (but is a simple consequence of the geometric growth rate) is that almost half of them have been published in the last 10 years. Even if the rate of publication were to stay at 50,000 papers per year, the size of the mathematical literature would double in another 20 years.

Similar exponential growth rates can be seen in other indicators of the size of the mathematical research enterprise. The American Mathematical Society (AMS) had 1,926 members in 1930, 6,725 in 1960, and 25,623 in 1990, a doubling roughly every 16 years. (The number of papers published has been doubling faster, almost every 10 years, and the share of the papers written by mathematicians in North America has not dropped too much recently, and went up substantially in the '30's and '40's. Does this mean that mathematicians have become more productive, possibly because their jobs involve less teaching and more research, or is it just that they publish more, or are they publishing shorter papers? These are intriguing questions that invite further study.) The Division of Mathematical Sciences of the NSF has seen its budget grow from about $13 M in 1971 to about $73 M in 1991, which is about a doubling in inflation-adjusted dollars. Awards (which are roughly proportional to the number of researchers supported) went from 537 to 1424, an increase by a factor of 2.65. The complaints one hears about reduced NSF support, just like those about reduced library budgets, are not so much about reductions in absolute or even inflation-adjusted budgets, but about their growth rates not keeping up with the number and output of researchers.

Exponential growth of the scholarly community cannot continue for long. The data from Chem. Abs. and Math. Rev. cited above show that there has already been substantial slowdown in the growth of publications. There are clear signs (evident in the current job market and

Page 6

Scholarly publishing has some features that sharply differentiate it from the popular fiction or biography markets, and make rapid growth difficult to cope with. Research papers are written by specialists for specialists. Various estimates have been made of how many readers a scholarly paper usually attracts. These estimates are not too trustworthy, since there is difficulty in collecting data, and also in defining what it means to read a paper. Typical estimates for the number of serious readers in technical fields are under 20. (One version of an old joke has an editor tell a referee, ``You are probably the second and last person to read this paper,'' to which the referee replies, ``Second? Are you sure the author has read it?'') To check whether this number is reasonable, ask yourself how many papers you have read in the last year. Great value is also derived by scholars from less thorough reading of articles, but even there, the number of times that an article is scanned is not very high. (See [Section 9.4] and [KingMR] for some estimates of this number.) Whatever the correct average is for either careful study or scanning of papers, it is small, and is unlikely to grow. This is a consequence of a simple arithmetic relationship. If a scholar reads x papers and writes y papers per year, and those numbers do not change much, then the average paper will be read by x/y scholars, no matter how large the total community is. (This argument is not rigorous, since it assumes a steady state. With growth in the ranks of scholars, many scholars who do read papers but do not write any, and papers being read many years after they are published the numbers can be tricky to evaluate, but the general conclusion seems right. All the numbers in this article are back-of-the-envelope type estimates. Given the exponential growth rates we are dealing with, more precise figures would not change the implications of the message.) This would change if we could attract more readers from outside our areas. However, given the increasing specialization that applies in all fields, this is unlikely. Interdisciplinary interactions are occurring with increasing frequency, but they almost always involve specialists, and this does not affect the trend towards a narrower focus of research, which precludes attracting a general lay audience. The trade press is different. If the potential audience doubles (either because of natural increase in population, or because readers in other countries start buying translations), a popular writer like Le Carre can expect to sell twice as many copies of his books. (There will be further discussion on the differences between scholarly and trade presses later. See also [Harnad2].)

Scholarly publishing, because of its nature, cannot benefit from the economies of scale that the trade press enjoys. As our numbers grow, we tend to work in narrower specialties, so that the audience for our results stays constant. Further, the centers in which we work (typically university departments) do not grow much, so we get increasingly dispersed. On the other

Page 7

Scholarly publishing would be facing a minor inconvenience and not a crisis if the scale of this enterprise were small enough. If a university department were paying $5,000 per year for journals, it could deal with several decades of doubling in size and cost of the subscriptions before anything drastic had to be done. However, good mathematics libraries spend well over $100,000 per year just for journal subscriptions, and the best ones spend close to $200,000 (with those at the top of the range obtaining also some of the most relevant physics, computer science, and engineering literature). To obtain the true cost of the library, we have to include the cost of staff and space. According to data for 1993 compiled by the Association of Research Libraries (ARL), and available on-line on the World Wide Web at URL http://arl.cni.org, costs of all acquisitions are usually about one third of the total cost of a research library. Hence a mathematics library that spends $150,000 on books and journals per year probably costs close to $500,000 per year to run. Totals for all departments at a university are often staggering. According to figures from ARL, Harvard University spends $58 million per year on all its libraries. Harvard stands out for the size and expense of its libraries, but other institutions also have large budgets, with (to cite just a few examples) Stanford at $36 M, University of Michigan at $28 M, Princeton at $21 M, MIT at $12 M, and Dartmouth at $10 M. Budgets that large are not easy to increase, and are likely to be scrutinized for possible reductions.

3 A brave new world

Modern technology is making possible methods of information dissemination that are dramatically cheaper than traditional journals. An example can be seen in the system that my colleagues in the Mathematical Sciences Research Center of Bell Labs and I have started to use recently. My recent preprints (including this essay) can be accessed through Mosaic at URL ftp://netlib.att.com/netlib/att/math/odlyzko/index.html.Z. (Preprints of some older, already published papers are also available there, but may have to be removed if publishers complain.) For those without access to Mosaic, ftp access is available on machine netlib.att.com. After logging in as ``anonymous'' and giving the full email address as password, all the user has to do is give the commands

to obtain a copy of the (compressed) index file, which describes what preprints are available. Alternatively, those preferring email or simply not having ftp available can send the message

send index from att/math/odlyzko

Page 8

to netlib@research.att.com, and the index file will arrive via return mail, with instructions for retrieving individual papers. (For papers of my colleague Neil Sloane, use the same commands as above, but with ``odlyzko'' replaced by ``sloane,'' and so on.)

The system described above provides access for all the 20 million users of the Net (as the Internet and various other networks are called) to the preprints that my colleagues and I write. This access is free and available around the clock. Further, this access is easy, and can accommodate users with various levels of connectivity and terminal equipment. What is most remarkable about it, though, is that it places a minimal burden on the author. All I need to do (once a paper has been typeset in , say) is to give the commands

and edit the file /usr/math/odlyzko/index by adding to it the lines

Everything else is done automatically by this public domain system, which was written by my colleague Eric Grosse. (In practice there is a bit more work, since I also make the source files available in the src directory.) The plummeting costs of storage mean that I do not have to worry about sizes of data sets. I therefore make available both PostScript files, for portability, and the source files, to make text searches easier.

The only time-consuming part in using Grosse's system is the typesetting of the paper, but that would be done in any case. The extra effort needed to make the preprint available is a matter of a minute or two. This is a dramatic change compared to the situation of even a few years ago, and certainly to that of a few decades ago, when the only way for a scholar to communicate with a wide audience was to go through the slow and expensive process of publishing in a conventional journal. Now it is possible to reach a much broader audience with just a few keystrokes.

Similar systems are becoming widely used, and are likely to spread. There is a compelling logic to them, as there is to the uniform email addressing convention that is becoming common, in which if Gauss were active at Celestial University, he might be reachable as

gauss@math.celestial.edu. Mathematics departments are beginning to set up preprint archives at addresses of the form ftp.math.celestial.edu. Existence of such publicly accessible preprint archives is a great boon to scholars, but it is extremely subversive of journal publications. If I can get a preprint of a published paper for free, why should I (or my library) pay for the journal?

Preprints are treated differently from refereed journal publications. However, the system described above can be used just as easily to run an electronic journal. If I am an editor, all I need to do after receipt of a revised paper that has been refereed is to follow a procedure analogous

Page 9

The facilities described in this section are examples of what is becoming common. The next two sections consider in detail developments in modern technology that are making this possible.

4 Hardware improvements

A doubling of scholarly papers published each decade corresponds to an exponential growth rate of about 7% per year. This is fast, but nowhere near as fast as the rate of growth in information processing and transmission. Microprocessors are currently doubling in speed every 18 months, corresponding to a growth rate of 60% per year. Similarly dramatic growth figures are valid for information storage and transmission. For example, the costs of the NSF-supported backbone of the Internet increased by 68% during the period 1988-91, but the packets transmitted went up by a factor of 128 [MacKieV]. The point of citing these figures and those below is that advances in technology have made it possible to transform scholarly publishing in ways that were impossible even a couple of years ago.

Recall that about 50,000 mathematical papers are published each year. If they were all typeset in , then at a rough average of 50,000 bytes per paper, they would require 2.5 GB of storage. If we include the fewer than 1,000 mathematics books (not counting college textbooks, say) that are published each year, we find that the annual output of the mathematical community requires under 4 GB of storage. (This assumes everything is in . If we use dvi files, the requirement about doubles, while with PostScript it about quintuples. On the other hand, compression can reduce storage requirements by a factor of 2 or 3 for source code, and 5 for PostScript. Given the rapid growth in the availability and cost of storage system, factors of 2 or 5 mean only a difference of a couple of years.) These numbers, which looked daunting a few years ago, are now trivial.

For comparison, the EDGAR database of financial reports receives about 30 GB of data per year. (It is available through several commercial vendors and is now being made available for free on the Internet by the New York University School of Business Administration with the support of a grant from NSF.) The Canadian Meteorological Centre receives 3 GB of data per day.

We can now buy a 9 GB magnetic disk for about $3,000. For archival storage of papers, though, we can use other technologies, such as optical disks. A disk with a 7 GB capacity that can be written once costs $200-300. (The equipment for writing data on it is still expensive, and costs $20,000 - 40,000, but it can be shared by many individuals and even departments). Digital magnetic tape is even cheaper. The standard CD-ROM disks, with about 0.7 GB of storage capacity, cost a few dollars each to produce (with the information written on it) in runs of a few thousand. Digital tapes with 250 GB capacities are expected to become available soon. Thus the electronic storage capacity needed for dissemination of research results in mathematics is

Page 10

We conclude that is it already possible to store all the current mathematical publications at an annual cost much less than that of the subscription to a single journal. What about the papers published over the preceding centuries? Since there are 1,000,000 of them, it would require about 50 GB to store them if they were all in . It appears unlikely (but not impossible) that anyone will undertake the project of converting them all into (or any other modern electronic format). What is more likely is that optical character recognition will eventually be applied to the texts, since this will enable rapid computerized text searches, while the equations will be stored as bitmaps. To provide reliable access to the text, whole papers will have to be available as full bitmaps. This dramatically increases the storage requirements. However, even then, they are not prohibitive. With current fax standards (which produce copies that are not pleasant to read, but are usable), a page of mathematical text requires 30-50 KB. Therefore all the 1,000,000 mathematical papers can be stored in less than 1,000 GB. This is large, but it is still less than 150 of the current large optical disks. For comparison, Wal-Mart has a database of over 1,000 GB that is stored on magnetic disks, and is processed intensively all the time. The credit card transaction records maintained by American Express come to 220 GB. The DIALOG set of databases contains about 3,500 GB of data.

The storage requirements of scholarly literature are likely to become even more ordinary in the near future. Cable TV and phone companies are getting ready to deliver movie-on-demand to the home. This will require tremendous storage capacity, and is bound to stimulate development of much denser and cheaper media. There is not much progress in optical disks, but magnetic hard disks are becoming both larger and cheaper, and there are many promising new technologies on the horizon, such as optical tape that might be able to store over 1,000 GB on a single unit. We can start thinking of 1,000 GB storage devices for personal computers becoming available in a decade or so. This means that any one will be able to have all the mathematical literature available right on the desktop. Librarians and scholars often express fears about availability and durability of data in the electronic world. However, once we can put all the mathematical papers on a single $50 tape, every professional mathematician, every graduate student, and every local public library will be able to have one. Instead of 1,000 libraries in the world having a copy of a particular journal, we will have 100,000 libraries and individuals owning it.

Even before systems able to store all the mathematical literature become available to individuals at low prices, though, larger institutions, such as university mathematics departments, will be able to make such systems available to their members. This ability will mean a dramatic change in the way we operate. For example, if you can call up any paper on your screen, and after deciding that it looks interesting, print it out on the laser printer on your desktop, will you need your university's library?

Although the arguments used here are repetitive, it is important to stress that technological progress has pushed the state of what is available with routine off-the-shelf systems far ahead of what is required for scholarly publishing. This makes many arguments against electronics invalid. For example, there is the famous story of warehouses full of tapes of US census data that have deteriorated so that they cannot be read. The argument is that if census data cannot be preserved in digital format, how can one possibly expect to preserve scholarly literature? However, the difficulty with the census tapes was that they were the product of an earlier low-density storage technology that required warehouses for what could today be accommodated

Page 11

The argument about scale of scholarly publishing suggests that some technologies will not suffice. CD-ROMs are rapidly growing in popularity, and seem set to become the preferred medium of electronic publishing for general literature. For example, in 1993, more CD-ROM encyclopedias were sold in the US than paper ones. However, CD-ROMs are not ideal for scholarly publishing because of their size. At 0.7 GB capacity, they are not capable of storing even the complete set of reviews of the mathematical literature. It is true that a CD-ROM might hold almost all the papers that an individual mathematician will work with, but only if one can decide beforehand what those papers are going to be. There is simply not enough capacity to put in all that might possibly be needed, as could be done if the capacity were 70 GB, and not 0.7 GB. (Multiple CD-ROMs can be an answer, but not a convenient one.) Therefore I suspect that mathematicians and other scholars will depend on network connections more than on CD-ROMs for access to information, even though this technology will be useful for distribution of some data. CD-ROMs may even act as a distraction, with great effort devoted to trying to squeeze data onto them. As in the case of fax vs. electronic mail, this may develop into a case where the ``good enough'' may delay the arrival of ``best.''

Optical disks can be shipped by parcel post at low cost. However, it is desirable to have faster communication links. Information transmission is a barrier right now. Anyone who has to download files at 2,400 baud (roughly bits per second, bps) can attest how annoyingly slow that is. However, communication speeds are increasing rapidly. Most departments have their machines on Ethernet networks, which operate at almost 10 Mbs (millions of bits per second). Further, almost all universities now have access to the Internet, which was not the case even a couple of years ago. The Internet backbone operates at 45 Mbs, and prototypes of much faster systems are already in operations. Movies-on-demand will mean wide availability of networks with speed in the hundreds of megabits per second. If your local suppliers can get you the movie of your choice at the time of your choice for under $10 (as they will have to, in order for the system to be economic), then sending over the 50 MB of research papers in your specialty for the last year will cost pennies. Scholars might not like to depend on systems that owe their existence to the demand for X-rated movies, but they will use them when they become available.

Page 12

There is concern in the scientific community that with the withdrawal of NSF support for the Internet, the electronic data highway will charge tolls that will be prohibitive for university researchers. That concern is misplaced. In the first place, the NSF subsidy is only $20 M per year [MacKieV], and covers around 10% of the cost of running the Internet. Compared to the 20 M users of the Internet, that is not much. Further, this concern is based on a static view of the world and does not take into account the increasing power and decreasing cost of technology. Yes, the commercial data highway will have tolls. Yes, the networks that the cable TV companies build may be structured primarily for the broadcast delivery of movies to the home, and may not have the full communication capabilities that scientists require. The point is that these networks are going to be large enough to justify tremendous development efforts that will drive down the costs of all communication technologies. The tolls on the commercial data highway will have to be low enough to allow transmission of movies. Therefore the cost of transmitting mathematics will be trivial. Even if the commercial data highway is not structured in an ideal way for science, or is a commercial flop, universities will be able to use the technology that is developed and build their own cheap networks. Just as the recent progress in computers is caused much more by demand from businesses running spreadsheets on PCs than from scientists modeling global warming on supercomputers, so the progress in communications will come primarily from commercial uses of the data highway. Some scholars (such as those getting atmospheric data from satellites) will have to worry about the costs of communications, since they will be transmitting giga- and tera-bytes of data. The rest of us can sit back and go along for the ride provided almost for free by commercial demand.

Stevan Harnad (private communication) notes that even now, non-scientific and non-educational uses of the Internet are a significant fraction of the traffic. Some 1993 statistics for netnews (the decentralized system of discussion groups) collected by Brian Reid of DEC show that the largest volume was generated by alt.binaries.pictures.erotica, which has an estimated 240,000 readers, and generated 31.4 MB of postings during one sampling period, even though it is received by only about half the sites. The second largest load was 16.3 MB from alt.binaries.pictures.misc, which has 150,000 readers. The third highest was 14.1 MB from bionet.molbio.genbank.updates, which has 24,000 readers. Sci.math, which has an estimated 120,000 readers (the highest of any of the sci.* groups, but only 39-th highest among all the groups) generated only 3.6 MB. Even alt.politics.clinton generated 7.1 MB during the sampling period! This only reinforces the conclusion drawn earlier that most of scientific communication can be accommodated in a tiny fraction of the capacity of future networks. (The gene bank data, like the atmospheric data mentioned before, is a special case.)

The arguments above do not mean that the existing networks are already sufficient for scholarly communication. Even though the traffic over the Internet backbone has been doubling every year for the last three years, it still comes to only between 1 and 3 MB per person per month (depending on how many users there really are). If email and ftp of papers were all that was going to be used, then 3 MB per month would be plenty. With the growth of multimedia traffic, though, we might soon run into a crisis situation that will require a substantial increase of the network, and might require imposition of tolls. Still, this is likely to be only a momentary problem, since technology is progressing rapidly, and faster networks are being developed.

Page 13

5 Software improvements

Not only have information storage and transmission capacities grown, but the software has become much easier to use. Computerized typesetting systems have become so common that it is rare to encounter a manuscript typed on an ordinary typewriter. Moreover, scholars are increasingly doing their own typesetting. This trend is partially due to cutbacks in secretarial support, but is caused primarily by scholars preferring the greater control and faster execution that they can obtain by doing their own typesetting.

Two centuries ago there was a huge gap between what a scholar could do and what the publishers provided. A printed paper was far superior in legibility to hand-written copies of the preprint, and it was cheaper to produce than hiring scribes to make hundreds of copies. Today the cost advantage of publishers is gone, as it is far cheaper to send out electronic versions of a paper than to have it printed in a journal. The quality advantage of journals still exists, but it is rapidly eroding. Systems such as , widely used by scholars, are being increasingly adopted by publishers in response to the economic pressure to lower costs. (It was the switch to a poor electronic typesetting system by Don Knuth's publisher that prompted him to invent in the first place.) Furthermore, these systems, because of their wide use, are improving so that they are almost as good as the best conventional typesetting systems. The software improvement is aided by increasing quality of laser printers (with 600 bpi printers replacing 300 bpi ones, and 1200 bpi printers likely to become common soon). Authors of papers and books with numerous figures are frequently told by publishers that it would not have been economically feasible to publish their manuscripts with old systems. Most journal papers still have the advantage of better editing, but even that advantage is eroding with the development of better software and better training of scholars. Further, many authors complain that the advantages of professional manuscript preparation are often vitiated by the mistakes that this process introduces.

The progress in communications is even more striking than in preparation of manuscripts. At the beginning of the electronic era most of us had email. Then came anonymous ftp. Together with the network connections that are now almost universal, they provide powerful tools for communication. However, this is only the beginning. We now have a plethora of new tools, such as Archie, Gopher, and WAIS, which allow easy access to the huge stores of information on the Internet without having to learn arcane details of how it functions. In the next few years these tools are likely to evolve much further, and become widely used in the day-to-day life by most scholars. The one software tool that seems likely to have the greatest impact is Mosaic. It is a browser, a tool for reading documents in the World Wide Web, a system of distributed hypermedia databases. The usage of Mosaic is growing explosively, as it seems to have finally reached the level of user-friendliness that the wide public can live with. (This seems to be another case of the discontinuous change that is common with new technologies. As some threshold of either price or convenience is passed, there is a snowball effect. We have seen this with fax machines and CDs, and I argue that we are likely to see this with electronic publishing of scholarly information.) Mosaic is often touted as the ``killer app'' for the Internet. With its ``point-and-click'' feature, it enables users to move from file to file, and from machine to machine. It's as if a visitor to a library, on encountering a reference in a book, could press a button and be magically transported to the shelf containing the referenced journal article. Many of the existing and planned electronic journals are adapting to Mosaic as their main access tool. I will not try to describe it, as that has been done extremely well in many places.

Page 14

While Mosaic is a great tool, it does have its limitations. After a prolonged period of ``surfing the Infobahn,'' users begin to express their dissatisfaction with the lack of structure in the databases and difficulty in quickly reaching what is needed. Mosaic may eventually be replaced by a better tool, just as VisiCalc was the first spreadsheet and helped launch the PC revolution, yet was completely displaced within a few years by better products such as Lotus 1-2-3. It might also be that Mosaic will evolve. Many software tools can be integrated easily with Mosaic. For example, RightPages(TM) is an experimental Bell Labs system that gives the user access to about 60 technical journals from about 10 publishers [Story]. After selecting an issue of a journal involved in the experiment, the user is shown the table of contents, and by clicking on an item, he or she is shown the first page of the article. If it looks interesting, a click on the appropriate icon orders a paper copy for delivery to the user. (At the moment only the first page is scanned, but this is not a fundamental limitation, and is caused by the lack of resources in a small-scale experiment to do more.) The RightPages system is used inside Bell Labs, and also at the University of California in San Francisco, in the Red Sage (TM) project. Its advantage is that it is better adapted for presenting journal papers than is raw Mosaic.

In addition to improvements to or on Mosaic, we will soon have automated personalized ``agents'' that will perform data searches automatically. Eventually such agents will have the ability to interact among themselves, specialize and become a secondary level of information.

Systems like RightPages are meant to work with the current paper journals and memoranda. They are supposed to compensate for some of deficiencies of the present system, by allowing users to sift through the increasing amounts of information that are becoming available, and also by giving them access to information they cannot get in their local libraries. However, while they do ameliorate the crisis in scientific publishing, they also contribute to it. I can now use the Internet to find out what books and journals the Harvard libraries have. Suppose I could also look on my screen at the articles in the latest issues of the journals that Harvard has received recently, and order photocopies (or even electronically digitized scans that can be reproduced right away on my laser printer) of the articles that interest me. Would there be any incentive to pressure my local library to order that journal? Would there be any reason to have more than one copy of the journal in the country (or the world)? If we do have only one copy, how is it going to be paid for? Thus the arrival of technological solutions to the current crisis in scholarly publishing also threatens to aggravate that crisis.

6 Electronic journals

What I am predicting is that scholarly publishing will soon move to almost exclusively electronic means of information dissemination. This will be caused by the economic push of having to cope with increasing costs of the present system and the attractive pull of the new features that electronic publishing offers. The costs of conventional journals have been mentioned in [Section 2] and will be discussed in much greater detail in [Section 9]. Here we discuss briefly electronic journals.

It is estimated there are currently around 500 regular electronic journals. At most 100 of these

Page 15

Electronic journals do not have to be inexpensive, and inexpensive journals do not have to be electronic. However, electronic publication does offer the best opportunity of reducing costs. The economic argument will probably all by itself force a move to electronic journals. There are additional reasons for preferring electronic journals. Some of the most compelling ones involve the interactive potential of the Net, and are discussed in [Section 8]. They involve novel features that are likely to be of great value to scholars, yet are impossible to implement in print journals. Even if electronic journals did not include any of those features and restricted themselves to the familiar format of print ones, they would still have many advantages. One is easy and general access. Most mathematical journals are available at about 1,000 research libraries around the world. Even for the scholars at those institutions, access to journals requires a physical trip, often to another building, and is restricted to certain hours. Electronic journals will make access available around the clock from the convenience of the scholar's study. It will also make literature searches much easier. For journals without subscription fees, access will be available from anywhere in the world. Electronic publishing will also have a healthy influence on preprint distribution. One point that is made about Ginsparg's theoretical physics preprint service, both by Ginsparg and other users [Ginsparg], [Sci1], is that it has made the latest results much more widely available, and diminished the importance of various small ``in'' groups.

Page 16

The advantages of easy and almost universal access to electronic publications have a measurable impact on scholars' activities. For example, it is noted in [Ginsparg] that many physicists obtain copies of their colleagues' papers (from the same institution) through Ginsparg's preprint server instead of directly from authors. Similarly, numerical analysts often use the Dongarra-Grosse netlib system to repeatedly obtain the same basic subroutines, since it is easier to obtain them from a centralized and well-organized source than to remember where the previous copy was placed.

Concern is often expressed that electronic publishing will deprive poorer institutions, especially those in the less developed countries, of access to the scholarly literature. The opposite is bound to be true. Few institutions can afford the $21 M per year that Princeton University spends on its libraries. Yet a T1 connection to the Internet (of 1.5 Mbps capacity) costs $20,000-50,000 per year in the US, and would suffice to bring in all the scholarly information that is generated in the world, if only that information were electronic. In other countries connections are more expensive, but even so, less than 1% of what Princeton spends will pay for a satellite earth station of high capacity. Further, while the cost of print journals is going up, that of electronics is going down. Therefore electronic publication is the most promising route for scholars in less developed countries to become full participants in intellectual life.

Once many journals become available electronically, paper copies are likely to disappear. It will be a case of positive feedback operating in a destructive mode. Necessary information will be available electronically, and most of it will have to be accessed electronically, since the local libraries will not be able to provide copies of all relevant journals. Therefore researchers will have less incentive to press for paper journal subscriptions to be maintained, which will lead to diminished circulation, and therefore to higher prices and more pressure from libraries to cut back on subscriptions.

7 Will it really be different this time?

Anyone venturing to predict dramatic changes in scholarly publications has to be mindful of the long history of failed forecasts in this area. Back in 1913 Thomas Edison predicted that movies were going to replace books. A much more carefully considered forecast that also predicted the demise of books was made by Vannevar Bush in 1945 in the article ``As we may think,'' which presented the idea of Memex, a personal data storage device that would contain massive amounts of information. This influential essay is usually regarded as the progenitor of modern hypertext. Since Bush was one of the creators of digital computers, and the article was published in 1945, it is usually thought that it was stimulated by developments in computing. However, the studies in the interesting collection [NyceK] show that Bush started writing early drafts of his article during 1930s, stimulated by the possibilities of dense storage of information on microfilm. These possibilities were also fascinating such thinkers as H. G. Wells, who published the book World Brain, motivated by the novel ways to combine information. There were predictions that being able to put entire libraries in cabinet-sized devices in every person's home would lead to a general uplift in the intellectual level of the population. However, microfilm has played a limited role (and one that is usually cordially disliked by scholars).

In more recent times there have been other predictions of how electronic publications were going

Page 17

A skeptical look at the predictions of a rapid switch to electronic journals has been presented by Schaffner [Schaf1], [Schaf2]. She notes various practical obstacles, such as lack of standards for presenting scholarly data in digital formats. She also emphasizes the slow evolution of the print journal, which adapted it well to serve scholarly needs. This evolution owed little to technological developments, and was driven by developments in scholarly culture. Also, while scholars may be intellectually adventurous, they tend to be conservative in their work habits. This conservatism is often reinforced by experiences with new technologies. Much is made of the interactive potential of the Net, and [Section 8] of this article will be devoted to this topic. However, Don Knuth, the creator of and an eminent computer scientist, has stopped using email, as he felt it consumed too much of his time. Thus not all the features that are beckoning us to the electronic world are as great as they seem.

Although the arguments against rapid change in scholarly publishing do have some merit, they do not take into account the drastic recent changes in available technology. The traditional journals are refined products of a long evolution. However, the environment they operate in is changing drastically. Many of the features of traditional journals might hinder their survival in the new world. They will not vanish instantly, of course, and can persist in their mode of operation for a few more years, but eventually they will have to either change or die. They still have some time left, since technology is still not quite able to handle scholarly communication needs, and traditions take time to overcome. The dreams of the visionaries of several decades ago who foresaw the dramatic effect that electronics could have on scholarly communication are still not fully realized. The main reason is that it took a long time for technology to provide the tools that made those futuristic dreams possible. Even 20 years ago, computing was largely done in batch mode on mainframes, and the few fortunate enough to have access to time-sharing systems had to content themselves with printing terminals communicating at 300 bits per second. In that environment, electronic publishing and intensive collaborations across oceans were not feasible. With the rapid advance in technology, though, we are just about at the stage where the needs of an electronic publishing system for scholars can be met easily. Moreover, the early predictions (such as those of capacity of storage systems in [Licklider]) often anticipated correctly that it would only be around now that the necessary capability would become available. Thus it is not surprising that no revolution in scholarly publishing has taken place yet.

A major reason for expecting a revolution soon is that we are not dealing with some special system developed just for scholars, as Bush's Memex might have been. Instead, scholarly publishing will be swept along in the general move to the electronic world. We can see this already in the rapid growth in manuscripts that are typeset on computers and in the increasing use of email. We can also see it in some abrupt changes that have taken place in the scholarly

Page 18

Even if we accept that change to electronic journals is bound to take place, it is conceivable that it might be on a gradual evolutionary path, with electronic journals slowly gaining on paper ones. The high energy theoretical physicists who now distribute their preprints almost exclusively through Ginsparg's system are still submitting them for publication in traditional journals. Numerical analysts, who have been relying on the Dongarra-Grosse netlib system for distribution of software, are also still publishing their papers conventionally. However, I feel that slow evolution is unlikely. The problem is that the natural development of present preprint distribution systems, described in the next section, is going to make scholarly papers freely available on the Net, so that scholars will be relying on their libraries less and less. They will therefore have less and less incentive to press for paper journal subscriptions to be maintained, which will lead to diminished circulation, and therefore to higher prices and more pressure from libraries to cut back on subscriptions. (Circulations of many mathematics journals have been declining at about 4% per year recently, which has contributed to the price increases.) This phenomenon all by itself can lead to catastrophic declines over a period of a couple of years in print journal circulation. There could also be even more sudden changes at individual libraries. The costs of the traditional system are so high (and are still growing), that they are bound to attract attention of administrators. A library that costs $400,000 per year to maintain, after all, takes the resources that would provide several faculty positions. If nothing is done, then in a decade or so, during the next financial squeeze at a university, a dean might come to a mathematics department and offer a deal: ``Either you give up paper journal subscriptions, or you give up one position.'' Once the hurdle of canceling journal subscriptions is overcome, the next offer is likely be ``Either you give up maintaining your old bound journal collection, or you give up a position.''

While the transition to electronic publication does appear inevitable, print journals are unlikely to suffer catastrophic declines in their circulation for a few more years. Even though some fields might switch to almost complete reliance on electronic information distribution soon, and do it suddenly (as high energy theoretical physics has done recently), the inertia of a decentralized and heterogeneous scholarly publishing business means that change will take time. Even if the number of electronic journals were doubling every year, it would be over 7 years before it would equal the number of print journals. Thus it seems unlikely that major changes in scholarly publishing will be visible within the next 5 years, at least when measured in the revenues of print publishers. Most papers will continue to be published in the traditional way. However, subscriptions will continue to drop, prices will continue to increase, and the system will be showing more and more signs of stress. At the same time, electronic publications will be developing rapidly, and eventually they will become dominant, most likely between 2000 and 2010.

Page 19

8 The interactive potential of the Net

Because conventional print journals have been an integral part of scholarly life for so long, their inflexibility is often not appreciated. As one example, consider surveys or bibliographies. They are invariably obsolete even before they are printed, and there are few options for updating them. In contrast, an electronic version can be continually updated by the author. This is only a small example of the novel tools that electronic communication offers to scholars.

[Section 9] will discuss the economics of publishing and what role various institutions are likely to have. In this section I describe my vision of what future scholarly publications will be like, with special emphasis on mathematical publications. I start by disputing Frank Quinn's vision [Quinn] of what the present system is and ought to be, and then discuss some novel information dissemination systems that have certain features I expect to find in future systems. Finally I present the basics of the system I expect to emerge over the next decade.

Scholarly journals have evolved during the last three centuries, in the world shaped by Gutenberg's invention of movable type. This invention made possible wide dissemination of scholarly publications. However, because printing, although much cheaper than hand copying, was still expensive, journals were constrained into a format that emphasized brevity. Further, the standards have promoted correctness. Since it took a long time to print and distribute journal issues, and corrections likewise required a long time to disseminate, it was natural to develop a rigorous refereeing standard. (This was not the only reason, of course, but I believe it was an important one in the development of our current editorial practices.) As a result, mathematical literature at least has become reliable, in that mathematicians feel free to use results published in reputable journals in their work, without necessarily verifying the correctness of the proofs on their own. (This correctness of the mathematical literature has increased substantially over the last two centuries. A perusal of Dickson's History of the Theory of Numbers shows, for example, that old papers seemed to contain serious mistakes with distressing frequency.) Frank Quinn [Quinn] argues that this feature justifies extreme caution in moving away from paper journals, especially in mathematics, lest we be tempted into ``blackboard-style'' publishing practices that are common in some fields. In particular, he advocates keeping a strong distinction between informal preprint distribution and the formal refereed publications, even in an electronic format. I agree that mathematicians should strive to preserve and enhance the reliability of mathematical literature. However, I feel that Quinn's concerns are largely misplaced, and might serve to keep mathematicians and other scholars from developing better methods for information dissemination.

The first point that should be made is that electronic publication does not in any way prevent the maintenance of present publishing standards. Electronic journals can follow exactly the same policies (and might even have the same names) as the current paper journals. On the other hand, paper journals are no guarantee against unreliable results, since the practices Quinn deplores are common in some fields, and have been present for a long time. Thus the reliability of literature in any field depends primarily on the publishing norms in that field, and not on the medium. Quinn is right that electronic publication does present increased temptations to careless communication. Computers do promote a less thoughtful style of correspondence. However, that can also be said of the telephone, or even a good postal service. Just read the letters that people used to write in the 18th century. By today's standards, they tended to be literary masterpieces. The difference was that letters took a long time to deliver, and were

Page 20

Paper journals serve several purposes in addition to that of providing reliable results. By having a hierarchy of journals of various prestige levels, the present system serves a filtering role, alerting readers to what the most important recent results are, as well as a recognition role, helping in grant and tenure decisions. Here again there is no reason that electronic journals could not provide the same services.

A more serious objection to Quinn's article [Quinn] (and to a large extent also to [JaffeQ]) is that its picture of mathematicians whose main goal is to produce formally correct proofs is unrealistic. (See [JaffeQ2] for some additional comments.) I agree completely with Quinn that it is desirable to have correct proofs. However, it's a mistake to insist on rigor all the time, as this can distract from the main task at hand, which is to produce mathematical understanding. There are many areas of mathematics today that are not completely rigorous (at least so far), with the classification of finite simple groups just one example. This has been true at various times in the past as well. After all, Bishop Berkeley through his ``ghosts of departed quantities'' jibe about infinitesimals and related arguments did show that 18-th century calculus was not necessarily any more rigorous than theology. On the other, from a historical perspective he lost the argument, since the necessary rigor was eventually supplied. In a similar vein, we speak of the Riemann Mapping Theorem, even though experts agree that Riemann did not have a rigorous proof of it, and that a proper proof was only supplied later. Standards have not improved all that much in the intervening years. The present system does not do a good job of providing the expected reliability, even if by reliability we mean the standards accepted in a given field, and not formal correctness. Even conscientious referees often miss important points. Furthermore, many referees are not all that conscientious. Once a paper appears, there are some additional controls. Sometimes a careful reviewer for Math. Rev. will catch a mistake, but that does not happen often. More typically, an error will be pointed out by someone, and then, depending on who and what is involved, there may be an erratum or retraction published, or else a note will be inserted into another paper, or else the mistake will only be known to the experts in the field. If the topic is important, eventually some survey will point out the deficiencies of the paper, but most papers do not get this treatment. Thus we already have situations where published work has to be treated with caution. It is more reliable than that of just about any other field, but it is not as reliable as Quinn's article might lead us to believe.

Lack of complete reliability is only one defect of the current paper journal system. Delays in publication are the one that is best known and most disliked. This is the one area where electronic publication offers a clear advantage. (There is an interesting parallel here to the rise of scholarly journals. They originated in the 17-th century in an attempt to improve on the system of personal letters through which discoveries were being communicated, which in turn developed because the traditional method of communicating through book publications was too slow.) There are others as well. A major one is caused by the emphasis on brevity that was encouraged by an expensive system of limited capacity. Although this is seldom said explicitly, the standard for presentation in a research paper is roughly at the advanced graduate student level. If you write to be understandable to undergraduates, referees and editors will complain

Page 21

Bill Thurston [Thurston], in an article that is largely a rejoinder to that of Jaffe and Quinn [JaffeQ], argues convincingly that formal proofs are fundamental to the correctness of mathematics, but they are a terrible way to convey mathematical understanding. Although this is seldom stated explicitly, implicitly it seems to be well understood by everyone. After all, we do have advanced seminars instead of handing out copies of papers and telling everybody to read them. The reason is that the speaker is expected, by neglecting details, by intonation, and by body language, to convey an impression of the real essence of the argument, which is hard to acquire from the paper. The medium of paper journals and the standards enforced by editors and referees limit what can be done.

While Thurston argues for a more intuitive exposition of mathematical results, Lamport [Lamport] advocates just the opposite. Both Lamport and Thurston feel that the usual standards of mathematical presentation fall far short of the rigor of formal proofs. Lamport feels that our literature would be far more reliable if proofs were more formal. One problem with this suggestion is that it would make proofs even harder to understand. Lamport's solution is to have a hierarchy of proofs, of increasing levels of rigor. However, the current system cannot accommodate this, given the premium placed on brevity.

The ideal system would, as Lamport suggests, have multiple presentations of the results, starting possibly with a video of the author lecturing about them, and going down to a detailed formal proof. Such a system is possible with electronic publishing, given the availability of almost unlimited resources (although it will be a while before video presentations can be included routinely), but cannot be accommodated with our paper journals.

Electronic publishing and electronic communication in general are likely to have a profound influence on how scholarly work is performed, beyond destroying paper journals. It is likely to promote a much more collaborative mode of research. One mode of research that is highly prized in our mythology is that of the individual who goes into a study and emerges a few months or years later with a great result. However, that is not the only mode of operation. In general, team efforts have been increasing, and rapid electronic communication via email and fax has been instrumental in this. Inspection of Math. Rev. shows that the proportion of coauthored papers has increased substantially over the last few decades, and this has also been noted in other disciplines. Given the increasing specialization of researchers, such developments are only natural. Further, collaboration is a congenial mode of operation for many. Laszlo Babai wrote a marvelous article [Babai] that I highly recommend. Its title is a play on words. It is an account of the proof of an important recent result in computational complexity, on the power of interactive proofs. At the same time this article is a description of how the proof was developed through exchanges of files of manuscripts and email messages among a group of about two dozen researchers. There are many such interactions going on, and the electronic superhighway will make them easier and more popular. (It will also create new problems, such as that of assigning proper credit for a work resulting from interaction of dozens of researchers. That is another story, however.)

The following two subsections examine some of the existing methods for information dissemination and their relevance for scholarly publications. All have serious deficiencies, but all have

Page 22

8.1 Netnews

The Net provides several interesting examples of information dissemination systems. Netnews, the decentralized system of discussion groups, is extremely popular. However, not many serious scholars participate. As has been noted by many, unmoderated discussion groups, such as sci.math, are at the opposite end of the spectrum from the traditional scholarly journals. They have been called a ``global graffiti board for trivial pursuit'' [Harnad1]. They are full of arrant nonsense, and uninformed commentary of the ``As I recall, my high school teacher said that ...'' variety. They are also beginning to attract cranks. (I was amazed that sci.math survived for many years without any serious problems with crackpots, although that is unfortunately changing.) Most mathematicians who try sci.math give up in disgust after a few days, since only a tiny percentage of the postings (of which there have been over 80,000 so far) have any real information or interest. I have continued reading it sporadically, more from a sociological interest than anything else. What I have found fascinating is that although there are now cranks posting to it (and, what is worse, generating long discussions), and there is plenty of ``flaming,'' as well as the nonsense alluded to above, there are occasional nuggets of information that show up. Sometimes a well-known researcher like Noam Elkies or Philippe Flajolet will provide a sophisticated solution to a problem that has been posted. What is perhaps even more interesting is that every once in a while some totally unknown person from a medical or agricultural school, say, will post an erudite message, giving a proof, a set of references, or a history of a problem. These are not the people I would think of when choosing referees, yet they clearly have expert knowledge of at least the topic at hand.

Reading sci.math also provides a strong demonstration of the self-correcting nature of science and mathematics. The opposite of Gresham's law operates, in that good proofs tend to drive out the bad ones. For example, every few months, the Monty Hall paradox (with a contestant given a choice of three doors, etc.) crops up again, as a new reader brings it up. There is typically a flurry of a few hundred messages (this is netnews, after all, and there is no centralized control and no synchronization) but after a week or so the discussion dies down, and everyone (or almost everyone, since there are always a few crackpots) is convinced of what the right answer is.

(In these days when we hear constant complaints about lack of public interest in science and mathematics, it is interesting to note that sci.math has an estimated 120,000 readers world wide. This is a large group of people who do have at least some interest in serious mathematics.)

Although the netnews model does have some redeeming features, it is not a solution to the scholarly publishing problem. The information content is far too low. There are real gems out there, such as ``This week's finds in mathematical physics'' series that John Baez posts, but they tend to get lost in mounds of nonsense. Methods are evolving for screening out unwanted messages, such as the ``kill'' files that keep messages posted by certain individuals or those responding to postings from such people, from ever being seen by the reader. There are also the thread-following readers that enable users to screen out quickly discussions on uninteresting topics. Even that is not sufficient, though. Even the specialized groups, such as sci.math.num-analysis,

Page 23

There are other kinds of moderated discussion groups that engage in a form of publication that is clearly useful, but would not fit into the traditional paper journal mode. For example, F. Bookstein (posting to vpiej-l, Pub-EJournals, December 2, 1993) oversees a morphometrics bulletin board. Its main function is to provide technical answers to questions from readers. They are seldom novel, as one goal is to avoid mathematical novelty, and provide references to existing results, or combinations of known results. However, they serve an important function, saving researchers immense efforts by providing needed technical advice.

There are many mailing lists that provide some of the services that Bookstein's bulletin board does. Many mathematicians are familiar with the Maple and Mathematica user group mailing lists. They consist of email messages from users with questions and complaints, and responses to those messages from other users and the paid support staff for those symbolic algebra systems. They do not qualify for traditional paper journal publications. Too many are basic questions about simple features of the system. Most are about bugs in the current release, or system incompatibilities, and nobody would want to make that part of the archival record. However, they are extremely useful, largely because with electronic storage, it is possible to search the great mess of them for the tidbits relevant to one's current needs. Further, they sometimes do veer into deep questions, as when a simple query about solving a system of polynomial equations evolves into a discussion of what is effectively computable with Groebner bases.

All the discussion group formats mentioned above fall easily into the informal communication category. However, they already begin to blur the line that Quinn [Quinn] wishes to preserve between the informal and the formal, reviewed publication. Scholars are increasingly relying on these informal methods. The demarcation line is blurred even further by preprints, discussed below. Quinn regards such blurring as pernicious. However, where Quinn sees danger, I see opportunity.

8.2 Preprint servers and directories

Clear evidence that the present scholarly publication system is not satisfactory is shown by the popularity of preprints. Half a century ago, there were practically no preprints, since there were no technical means for producing them. Journals were the primary means for information dissemination. Today, with xerox machines widely available, preprints are regarded

Page 24

Preprint distribution can be done, and increasingly is done, via email. It is much easier to write a shell script or create an alias to send 50 copies electronically than it is to make 50 xerox copies, stick them in envelopes, and address them. Still, this does lead to a messy decentralized system where each author has to decide who is to receive the preprints. There are two possible enhancements to it. Either one of them would be a major advance, and either one could become the main method of disseminating scholarly information in the space of a year or so. Either one would be extremely subversive to the present journal system, and could lead to its demise.

One enhancement to the present preprint distribution system is to use anonymous ftp directories, with possible email and WWW enhancements, of the sort described in [Section 3]. The other enhancement is to have an automated preprint server. There are several of them already operational in mathematics, such as the one at Duke that covers algebraic geometry. They have not had much influence yet. However, this could change suddenly. An instructive example is provided by the system that was set up by Paul Ginsparg at Los Alamos [Ginsparg], [Sci1], [Sci2]. In only one year, starting about two years ago, the high energy theoretical physics community switched almost completely to a uniform system of preprint distribution that is run from Ginsparg's workstation (and now also from some other sites that copy the material from his). Apparently nobody in theoretical physics can afford to stay out of this system, as it has become the primary means of information dissemination. Even brief interruptions of service [Sci2] bring immediate heated complaints. This system has already been extended to several other fields, aside from high energy theoretical physics, and the requests to help in setting up the system are a major chore for Ginsparg (private communication). The main point about this system is that it is cheap. The software is available for free, and not much maintenance is required. Physicists submit preprints electronically (in a prescribed version of , and in prescribed format), the system automatically files them and sends out abstracts to lists of subscribers (which is maintained automatically), and then the entire papers can be retrieved through email requests. Recently the system was enhanced by putting it in WWW, so preprints can be browsed using Mosaic.

An important observation about Ginsparg's system is that the transition to it was sudden, at least by the standards of the publishing world. It took under a year from the time Ginsparg wrote his program to the time it became the standard and almost exclusive method for distributing new results in high energy theoretical physics. Ginsparg [Ginsparg] attributes the

Page 25

Can Ginsparg's system coexist with paper journals? Theoretical high energy physicists are still submitting their papers to the traditional journals. However, it is not clear how long that will continue. If I have heard of an interesting result that has been published in some journal, and if I can order via email a preprint of the same version (except possibly for the formatting) from an automated preprint server, why should I bother to go to the library to look for the journal? If I don't look at the journal, why should I mind if my library cancels its subscription? If there is anything that is certain, it is that my library will keep coming back each year with lists of proposed journal cancellations, so the pressure to give up more and more subscriptions will continue. One solution would be for publishers to require that preprint copies be deleted from preprint archives once a paper is published. I doubt if this approach can work. It is technologically hard to enforce. How can anyone keep track of all the copies that might have been squirreled away on private disks? Even more important, would scholars tolerate such requirements?

Given the evolution of software, the distinction between a centralized preprint server and a decentralized database will soon be immaterial. As long as preprints are available in electronic form and their authors are interested in disseminating them widely, it will soon be easy to locate and obtain them for anyone on the Net with minimal effort by the authors.

Wide distribution of preprints is supported even by Quinn [Quinn]. However, preprints in most areas have a different status from refereed papers, and Quinn argues strongly for maintaining this distinction. My feeling is that there is no way to do this effectively. The temptation to use preprints stems from the desire for faster dissemination of information. If preprints are to be widely distributed, though, we need to adapt the peer review system to accommodate them. At the moment this is not done in a satisfactory way.

It is possible for a field to rely just on preprints. Ginsparg [Ginsparg] writes that in theoretical high energy physics, they have long been the primary means of communicating research results. Papers are still published in conventional journals, but do not play much role in the development of the field, since they are typically obsolete by the time they appear. Peer review operates, but in an informal way, as several physicists in that area have told me. Rob Pike, a colleague working in operating systems, reports (private communication) that in his area journals have also become irrelevant. Communication is via email and electronic exchange of preprints, typically through anonymous ftp. A recent announcement of reports about a new operating system resulted in over a thousand copies being made via anonymous ftp in just the first week. Peer review operates in this area also, but again journal publications do not play much of a role in it (and probably cannot, since the main product is not embodied in an article, but in the software).

It is easy to argue that the experience of operating systems is not relevant to mathematics. However, the same problems are arising in various areas of mathematics. There have been complaints by some mathematicians that their efforts were not getting proper recognition,

Page 26

There are even fields close to mathematics that do not follow the usual procedures for mathematical publications. Theoretical computer science has standards of rigor that are comparable to those in mathematics. There are over a dozen annual meetings, with the original STOC and FOCS conferences the most general and prestigious. Papers (or, to be precise, extended abstracts of about 10 pages) to these conferences are submitted about six months in advance. A program committee then selects around 80 out of 200-300 submissions. Expanded abstracts (which are often full papers) are then published in the proceedings, which are given out at registration to all participants, and can be purchased by anyone from the publishers, ACM and IEEE. The official policy is that since the committee is making decisions on the basis of extended abstracts only, and in any case does not have time to do a thorough refereeing job, the proceedings publications should be treated as preliminary. The final revised versions are supposed to be submitted to refereed journals. However, there have been perennial complaints that this was not being done, and that researchers were leaving the proceedings drafts as the only ones on record. Recently, however, David S. Johnson has compiled a bibliography of STOC and FOCS papers, and it appears that around 60% of them do get published in a polished form in a journal. (Moreover, the 40% that do not get published are not by any means the bottom 40%, but include some of the most influential results. It's just that some authors are negligent and do not carry out their duties to their field.) Although terrible, this 60% figure is considerably better than the folklore would lead one to believe. Still, for working researchers, practically all the information dissemination takes place through early preprint exchanges, and interactions at these conferences. In the words of Joan Feigenbaum, in theoretical computer science ``if it didn't happen at a conference, it didn't happen.'' Moreover, acceptance at these conferences is regarded as more prestigious than journal publications, as one can see from letters of recommendation. Thus here also it is possible to have a healthy field that does not depend on journal publications for information dissemination. The reason computer scientists seem to tolerate the present lack of reliability is that the conference system provides them with rapid evaluation and feedback, and also opportunities for extensive personal interaction. However, their situation is not ideal. There are frequent complaints about errors in the proceedings papers, and there have been a few notorious cases where important claimed results turned out to be wrong. What is lacking here, as with other informal preprint systems mentioned above, is a better peer review system, to provide better assurance of reliability. What we see is the seductive influence of rapid communication through conference proceedings that has led theoretical computer science into the errors that Quinn [Quinn] warns against.

8.3 The publication and reviewing continuum

The popularity of preprints, of netnews groups, and of mailing lists shows that they do fill an important role for scholars. We cannot turn back the tide. On the other hand, there is also a need for reliability in the literature, to enable scholars to build on the accumulated knowledge. One way to resolve the conflict is to follow Quinn's advice [Quinn] and rigidly separate distribution of information, such as preprints, from publication in journals. The former would be done rapidly and informally, while the latter would follow the conventional

Page 27

I will describe the system I envisage as if it were operating on a single centralized database machine. However, this is for convenience only, and any working system would almost certainly involve duplicated or different but coordinated systems. I will not deal with the software aspects of this system, which will surely involve hypertext links, so that a click on a reference or comment would instantly bring up a window with that paper or comment in it. My main concern will be with how scholars would contribute to developments of their fields. My basic assumption is that in most scholarly areas, articles will remain the main method of communicating specialized information, although they may be enhanced with multimedia features.

At the bottom level of future systems, anyone could submit a preprint to the system. There would have to be some control on submissions (after all, computers are great at generating garbage, and therefore malicious users could easily exceed the capacity of any storage system), but it could probably be minor. Standards similar to those at the Abstracts of the AMS might be appropriate, so that proofs that the Earth is flat, or that special relativity is a Zionist conspiracy, would be kept out. Discussions of whether Bacon wrote Shakespeare's plays might be accepted (since there are interesting statistical approaches to this question). There would also be digital signatures and digital timestamping, to provide authentication. The precise rules for how the system would function would have to be decided by experimentation. For example, one feature of the system might be that nothing that is ever submitted could be withdrawn. A similar policy is already part of Ginsparg's system, so that papers can be withdrawn, but the withdrawal is noted in the record. This helps enforce quality, since posters submitting poorly prepared papers risk having their errors exposed and publicized for ever. On the other hand, such a rule might be felt to be too inhibiting, and so might not be imposed.

Once a preprint was accepted, it would be available to anyone. Depending on subject classification or keywords, notification of its arrival would be sent to those subscribing to alerting services in the appropriate areas. Comments would be solicited from anyone (subject again to some minor limitations), and would be appended to the original paper. There could be provisions for anonymous comments as well as signed ones. The author would have the opportunity to submit revised versions of the paper in response to the comments (or his/her own further work). All the versions of the papers, as well as all the comments, would remain part of the record. This process could continue indefinitely, even a hundred years after the initial submission. Author , writing a paper that improves an earlier result of author , would be encouraged to submit a comment to to that effect. Even authors who just reference would be encouraged to note that in comments on . (Software would do much of this automatically.) This way a research paper would be a living document, evolving as new comments and revisions were added. This process by itself would go a long way towards providing trustworthy results. Most important, it would provide immediate feedback to scholars. While the unsolicited comments would require evaluation to be truly useful, and

Page 28

Grafted on top of this almost totally uncoordinated and uncontrolled system there would be an editorial and refereeing structure. This would be absolutely necessary to deal with many submissions. While unsolicited comments are likely to be helpful in deciding on the novelty and correctness of many papers, they are unlikely to be sufficient in most cases. What can one do about a poorly written 100--page manuscript, for example? There is need to assure that all the literature that scholars might rely on is subject to a uniform standard of refereeing (at least as far as correctness is concerned), and at the same time control the load on reviewers by minimizing duplicate work. Both tasks are hard to achieve with an uncoordinated randomized system of commentary. A formal review process will be indispensable. That is also Harnad's conclusion [Harnad1], [Harnad2]. There would have to be editors who would then arrange for proper peer review. The editors could be appointed by learned societies, or even be self-appointed. (The self-correcting nature of science would take care of the poor ones, I expect. We do have vanity presses even now, and they have not done appreciable damage.) These editors could then use the comments that have accumulated to help them assess the correctness and importance of the results in a submission and to select official referees. (After all, who is better qualified to referee a paper than somebody who had enough interest to look at it and comment knowledgeably on it? It is usually easy to judge someone's knowledge of a subject and thoroughness of reading a manuscript from their comments.) The referee reports and evaluations could be added as comments to the paper, but would be marked as such. That way someone looking for information in homological algebra, say, and who is not familiar with the subject, could set his or her programs to search the database only for papers that have been reviewed by an acknowledged expert or a trusted editorial board. Just as today, there would be survey and expository papers, which could be treated just like all the other ones. (Continuous updating would have obvious advantages for surveys and bibliographies.) As new information accumulated with time, additional reviews of old papers might be solicited as needed, to settle disputes. All the advantages that Quinn claims for our present system in providing reliable literature could be provided by the new one.

Harnad [Harnad1], [Harnad2] advocates a hierarchy of groups, with individuals required to pass scrutiny before being allowed to participate in discussions at higher levels. I feel this will be neither necessary nor desirable, especially in mathematics. The best structure might vary from field to field. Harnad is a psychologist, and he may be reacting to the fact that the proverbial people in the street all fancy themselves experts in psychology, and qualified to lecture on the subject to anybody. On the other hand, most people disclaim any expertise in mathematics. Therefore I do not expect that systems in mathematics will be flooded by crank contributions. After all, how many cranks will be interested in discussions of crystalline cohomology, or pre-homogeneous vector spaces? There is a need to provide protection against malicious abusers of the system, but beyond that restrictions to specific topics might suffice. In a few cases stronger controls might be needed. For example, any attempted proof of Fermat's Last Theorem might attract many crank or simply uninformed comments that would be just a distraction. That is not likely to be a problem for the bulk of mathematical papers.

Page 29

Will there be any place for paper in a system such as that sketched above? I suspect it will be limited, if there will be any. In paper publishing nothing can be changed once it has gone to the printer. Therefore this system cannot be adapted to provide the opportunities for continuous updating and correcting that is available with electronic publishing. We rely on this system because it was the only one that was feasible in the past. However, we now have better alternatives, and I expect them to dominate. Perhaps the exceptionally important papers will be collected periodically (after selection by a panel of experts) and printed as a mark of distinction. If that is done, then the proposed system will in many ways resemble the one advocated by Quinn. Quinn suggested a minimum six month delay between submission of a paper and publication. The scenario I am sketching would provide a 10 or 20 year delay, and even greater assurance of reliability! For all practical purposes, though, it is likely that traditional paper journals will cease to matter to scholars.

8.4 A possible scenario

Let me show in detail how I would like the future system to function by using an example of a recent mathematical discovery. This will also offer a way to explore the inadequacies of the present system, and what the difficulties of the current preprint and electronic announcement systems are. One may well object that generalizing from pet cases is not reliable. However, it is often preferable to deal with concrete cases, as opposed to talking in vague terms that are hard to interpret.

The example I will cite is not typical. No example could be typical in the immense area of scholarly inquiry, but this one has several unusual features. I use it precisely for that purpose, because it allows me to specify what I would like to happen in some extreme cases that are not likely to apply most of the time.

On March 19, 1993, Brendan McKay and Stanislaw Radziszowski posted an announcement on sci.math.research that they had proved that . One reason for choosing this example is that their result can be explained even to a lay person. What McKay and Radziszowski showed is that if there are 25 people in a room, either there is a subset of 4 of them so that any two know each other, or else there is a subset of 5 of them so that no two know each other. Previously it was known only that this was true if there were at least 28 people in the group. This result falls in the area of Ramsey theory, and a basic theorem in that subject is that for all natural numbers m and n, the Ramsey number exists, so that in any group of at least people, either there will be m mutual acquaintances or else n mutual strangers. What Ramsey theory says is that you cannot have complete disorder, that no matter how you arrange relationships in a group of people, there will be systematic patterns. (There is much more to Ramsey theory than this, as one might expect, but the above is a basic and classical result.)

So much for a description of the result. One reason it is unusual is that it attracted attention in the popular press. This then led a lay reader to write a letter to Ann Landers, which she published in her widely syndicated column, asking whether taxpayers' money was not being wasted on such endeavors. This led to a flurry of messages on sci.math, discussing the best ways to respond to the letter and to Ann Landers' request for an explanation of the value of research of the kind McKay and Radziszowski performed. (Unfortunately, while various letters

Page 30

How significant is the McKay-Radziszowski result? There is a small part of Ramsey theory devoted to computing values of various Ramsey numbers. However, it is not among the most significant areas of combinatorics. The methods are a combination of mathematics (since raw enumeration of the various possibilities is beyond the ability of any computer that could be constructed in the known universe using the constraints of currently known physical laws) and heavy computation. Judging from the announcement, it appeared that McKay and Radziszowski had extended the state of the art in both mathematics and computing. However, that is about it, and there are no current applications for this result that I know of.

We now take a detour and ask what is worth publishing in mathematics. A certain eminent mathematician said recently that < The only thing I care about are the 100 really interesting, and 400 interesting, papers in mathematics published each year. The rest is trash good for the wastebasket, on a par with junk mail.>

The McKay-Radziszowski paper (of whose existence I learned by contacting the authors, since none of the announcements on the Net that I had seen mentioned it) does not qualify by most experts' standards for the most interesting 500 papers written in mathematics in 1993. However, I feel that this eminent mathematician is wrong, and that papers such as that of McKay and Radziszowski should be published (as this one will be, in J. Graph Theory). The reason is that I, along with many other mathematicians, scientists, and engineers, often find myself looking for some specific result, such as the value of . I don't particularly care how hard it was to derive the answer. If it's not a triviality that I should be able to see for myself in a few minutes, I am glad to find it in print, even if I could have done the work better and faster than the author (something that is surely not true here). The point is that I want the answer for some specific reason, and don't want to take a day or a week on a detour from the project I am working on to figure out the answer for myself. (A major strength of the mathematical literature, which Quinn [Quinn] stresses, is precisely that by providing reliable results, it makes this mode of operation possible.) What I want from the Net is the tools that will let me locate such answers quickly, and the editorial/refereeing system to provide me with the assurance that I can trust what I find.

I should stress that I do not have an application for the McKay-Radziszowski result right now. However, I can cite numerous examples where I or my colleagues have used similar results found in the literature. In those cases we did rely heavily on the reliability of the published papers. If I design a code that is supposed to correct certain types of errors, I need to be able to guarantee that it will do so. If disaster strikes, and some product fails because some of the errors that were supposed to be corrected slip through, it won't do to say ``Well, I saw this claim on sci.math, by some chap I have never heard of, that ....''

How could the McKay-Radziszowski claim be disproved? One way would be to exhibit a counterexample. If you find a collection of 25 people such that no 4 are mutual acquaintances, and no 5 are mutual strangers, then this will be it. To exhibit such a counterexample, you only need to specify for each of the pairs whether they know each other or not. That is a small amount of data. For any specification, there are efficient algorithms that will tell whether your collection has the desired property. How you find the counterexample is

Page 31

The reader might well ask why am I taking so much time to explain things about a paper that is of little general interest. I do this to be able to discuss the question of reliability that is attainable with the present system. A hundred years ago, McKay and Radziszowski would have submitted their paper to a journal, and after a peer review a revised version would have appeared in print. For almost everybody, the printed version would have been the first one they would have seen. What assurance would those readers have gotten? If it was a prestigious journal, they would have assumed that at least one competent expert had looked at it and pronounced himself satisfied that it was correct. That is also what readers of current journals get. However, both then and now, there would be no indication of the care taken in the refereeing process. For McKay and Radziszowski, there is both the mathematics and the computing to consider. If one were to see it in print, it would be easy to imagine that the referee or referees checked the graph theoretic arguments that reduce the problem to manageable size. But how carefully was the mathematics checked? There would be no indication of that. What about the computational part? Programs to check problems of this size often take thousands of lines of code. Did the referees check those? It is uncommon for authors to submit program listings with papers. Even if a program listing is available (and according to the McKay-Radziszowski announcement, each author wrote an independent program, and both programs were run with the same outcomes, which I find commendable), how carefully was it scrutinized? Was the program available electronically, so the referees could run it themselves? Did the referees run it? (Seems unlikely, since the announcement mentioned a cluster of workstations that were used, with total run time of over 3 years for a single workstation.)

The aim of the above discussion is to point out that conventional peer review is not too reliable. Moreover, it is not serving the needs of the scholarly community in a timely fashion. In the discussion above, I said that 100 years ago, the first time that almost all people would have seen the paper would have been in the journal. However, we cannot roll back the clock. Researchers will not be denied the interactive potential of the Net. Scholars such as McKay and Radziszowski will keep announcing their results on the Net. However, even without the Net, we would have had problems, caused by the interactive potential of the mail and the telephone. Ever since the invention of xerography, preprints have been proliferating, and have become the dominant method of disseminating information among experts. Now what is the status of a preprint? Just how much can it be trusted? One answer, which seems to be the one advocated by Quinn, is that we should not trust a preprint at all, and should regard it as being in a form of purgatory until it undergoes proper refereeing. However, that is not what is happening. Let me stress that most of what I am saying about publications is descriptive, not prescriptive. I am saying < ``Here is how these fields are developing, and these are the reasons that drive these developments. Let us try to anticipate where this is likely to lead, and whether we can introduce some small modifications on the natural evolution of the system to make it better for scholars.''

Page 32