Oral-History:Enders Robinson

About Enders Robinson

Dr. Enders Anthony Robinson was born in Boston, 18 March 1930. He received his S.B. in mathematics and statistics in 1950 from the Massachusetts Institute of Technology. In 1952, he received an S.M. in economics from MIT where he did his thesis work under Professor Paul Samuelson and Professor Robert Solow, both later winners of the Nobel Prize. From 1952 to 1954, Dr. Robinson was director of the MIT Geophysical Analysis Group and he developed the first digital signal filtering methods to process seismic records used in oil exploration. The available computer, the MIT Whirlwind digital computer, was not powerful enough to make this research work commercially feasible at the time. Professor Norbert Wiener of MIT took an active interest in this work which represented the first successful application of his recently developed theory of prediction and filtering of time series. In 1954, Dr. Robinson received his Ph.D. in geophysics from MIT. He then worked in the oil industry both as a geophysicist and as an economist In 1958, Dr. Robinson joined the mathematics faculty at the University of Wisconsin at Madison where he became acting director of the computer science department. In 1960, the University of Wisconsin granted a fellowship to Dr. Robinson to work in Uppsala, Sweden under Professor Herman Wold and Professor Harold Cramer, earlier developers of time series analysis. Dr. Robinson stayed in Sweden from 1960 to 1964 as Deputy Professor of Statistics at Uppsala University. With the advent of transistorized computers in the 1960s, digital methods became economically feasible for use in oil exploration. In 1965, six geophysicists and Dr. Robinson formed Digicon Inc. which was one of the first companies to process seismic records by computers. From 1970 to 1983, Dr. Robinson divided his time between the oil exploration industry and universities. He was a member of the Boston Statistical Consulting Cooperative, a group set up by professors at MIT and Harvard. In 1981-82, he was a visiting professor of geophysics and theoretical mechanics at Cornell University and in 1983 he became the McMan Professor of Geophysics at the University of Tulsa. He spent 1990-1992 at the Department of Earth and Planetary Sciences at Harvard. In 1969, Dr. Robinson received the Medal Award of the Society of Exploration Geophysicists in recognition of outstanding contributions to the digital processing of seismic data and the Conrad Schlumberger Award of the European Association of Exploration Geophysicists for a rationalization and formalization of the geophysicist's approach to data enhancement. In 1983, he was made an honorary member of the Society of Exploration Geophysicists, and in 1984 he received the Donald G. Fink Prize Award of the IEEE. In 1988, he was elected a member of the National Academy of Engineering for pioneering contributions to digital seismic processing. Dr. Robinson is the author of books on probability and statistics, time series analysis, and geophysics. His most recent book is Einstein's Relativity in Metaphor and Mathematics. Robinson died on December 6th, 2022.

The interview begins with a discussion of the ENIAC computer. Robinson describes the application of Wiener's theories to seismology, and explains the seismic convolutional model and how the method of deconvolution evolved. Robinson discusses the earliest applications of digital signal processing in computing by oil companies, and how the computer fit into the oil refinery business. He and his partners then created a consulting firm, called Digicon, to program the computers for oil refinery companies. Companies such as Texas Instruments led the way for the application of existing computer technology to geophysical exploration, as is demonstrated by the advent of digital filters to remove reverberations as early as 1953. Robinson also discusses the peculiar difficulties he found in working with digital filters and primitive computing equipment and gives a detailed description of the application of Wiener's mathematical theories to geophysical computing and the advent of underwater surveying. The evolution of geophysical computing is also discussed. Highlights of this evolution which Robinson points out are deconvolution, migration, common depth point method and three dimensional data collection. Robinson concludes with an in-depth look at the evolution of the computing equipment and by reminiscing about his days at the Massachusetts Institute of Technology.

About the Interview

ENDERS ROBINSON: An Interview Conducted by Andrew Goldstein, Center for the History of Electrical Engineering, 6 March 1997

Interview #326 for the Center for the History of Electrical Engineering, The Institute of Electrical and Electronics Engineers, Inc.

Copyright Statement

This manuscript is being made available for research purposes only. All literary rights in the manuscript, including the right to publish, are reserved to the IEEE History Center. No part of the manuscript may be quoted for publication without the written permission of the Director of IEEE History Center.

Request for permission to quote for publication should be addressed to the IEEE History Center Oral History Program, IEEE History Center, 445 Hoes Lane, Piscataway, NJ 08854 USA or ieee-history@ieee.org. It should include identification of the specific passages to be quoted, anticipated use of the passages, and identification of the user.

It is recommended that this oral history be cited as follows:

Enders Robinson, an oral history conducted in 1997 by Andrew Goldstein, IEEE History Center, Piscataway, NJ, USA.

Interview

Interview: Enders Robinson

Interviewer: Andrew Goldstein

Date: 6 March 1997

Place: New York, New York

Education

Goldstein:

Your story begins with the ENIAC computer and work at Aberdeen Proving Ground.

Robinson:

Yes. To go back to my student days at MIT in 1946 as a freshman, we used to have lectures in physics, chemistry and mathematics under the great dome. Close by was a room that had Vannevar Bush's differential analyzer. Every day I would walk by, but the door was always closed. I was just a freshman, but simply walking in the presence of that machine was something—it was the most advanced and best known analog computer at the time. Then in 1949, I was in the Reserve Officers Training Corps and went to Aberdeen Proving Ground in Maryland as a cadet, and got to visit the ENIAC. It was impressive just for the space it took up. The next summer I was on active duty there and actually got to know some of the programmers. At that time, programming was an art form—there were no manuals—you just learned by doing.

I came back to MIT in the fall of 1950, and I was in the Mathematics Department as a graduate research assistant. I was working under Norbert Wiener to find applications for his time-series analysis. As you know, Norbert Wiener was an eminent mathematician at MIT and he worked in generalized harmonic analysis back in 1930. He always felt that people didn't realize that was probably his most important contribution. But during World War II, he worked on prediction theories for anti-aircraft fire control. He developed a theory—classified at the time—but published by MIT Press in 1949. We used the book when I was a student there, so I was familiar with it. By 1950, we were ready to apply it. Meanwhile, Wiener had written another book called Cybernetics which was first published about 1948. That book was an instant success; he was the one that introduced the word cybernetics.

Norbert Wiener; seismic convolutional model

Robinson:

So Wiener was a celebrity by 1950, and my research assistantship depended on finding applications of his work. Another MIT professor, George Wadsworth, was working in weather prediction which he originally started during World War II; Wiener's classified book had come out, but it was too difficult mathematically to be applied as such. So Wadsworth asked Norman Levinson, who was another eminent mathematician at MIT, to take Wiener's book and to simplify it into numerical algorithms that he, Wadsworth, could use. Levinson published these papers in the Journal of Mathematics and Physics in 1947. They were added as appendices to Wiener's 1949 book, so Levinson's algorithms with Wiener's theory became the way to do this type of thing. We could then apply it to geophysics because seismic records are essentially noisy records. The idea was to take these records and apply Wiener's theory.

Goldstein:

How was data like this collected?

Robinson:

I'll show you an example. This one was from the Atlantic Refining Company, which is now called Atlantic Richfield. Essentially it's a record taken in what they call exploration seismology. Dynamite is exploded and the seismic waves travel through the earth and are reflected from various interfaces and boundaries in the subsurface and come back to the surface where they are recorded as a seismic trace. Recording instruments are spread out so there are multiple traces for each shot, then for any prominent interface, things will line up on the different recordings. This record here is called a no-good record, because things don't line up.

Goldstein:

You mean the peaks?

Robinson:

If the peaks lined up, that would show an interface. The basic problem was that the things that lined up were the primary reflections, but there was also a host of other signals coming in, multiple reflections, diffractions, surface waves, refracted waves, all of which hide the primary reflections that you want to see. The idea was to use Wiener's theory to unscramble this. Wiener was very interested in this problem because unlike other pure mathematicians, he was very interested in applications. Although he's remembered for his pure mathematics, he always had contacts with applications people.

When you read Wiener's books you get the feeling that he might be a little arrogant, but if you knew him, he was anything but arrogant. He was a very kind, gentle person, but when he wrote books he tried to write them objectively, from the point of view of a disinterested observer, not from a personal point of view. So when you read his accounts, it's not the real Wiener who is talking, it's more like an objective person talking about a situation. He didn't really put himself into the book like they would today, because at that time people weren't supposed to do that sort of thing. A lot of stories were generated at MIT. For example, he would ask people for directions to his classroom and all that. But a lot of that was part of his game, his sense of humor—to add a little fun to the somber atmosphere around MIT at that time. He certainly knew what direction he was going in and which class he was teaching. He sort of helped form that professor image of himself. Then he would go to the faculty club for lunch and quite often he would play chess. He was probably thinking about mathematics when he played, because he was never an excellent chess player. He played for fun. A lot of the old MIT professors who played Wiener could brag that they had beat him. It was a big deal to them. [laughter]

Goldstein:

The decision to try and apply his techniques to seismology—how did that come to be?

Robinson:

Professor Wadsworth was in a carpool with Professor Hurley, a geology professor, and used to discuss his weather records with him. One day, probably in 1949, Professor Hurley said, "Well, you know in geophysics, we also have curves that look like that, wiggly curves that go up and down. Perhaps Wiener's methods could be applied to them." So Professor Hurley obtained eight seismic records, like this one, so that I actually had the data in the fall of 1950. Even at that time it was a small amount of data (because all the graphs were examined by eye). Everything then was done just by observation. In a sense, today everything is still done by observation. The human being still has to make the final decision. They thought this would be a good application, and my assistantship was an MIT Mathematics Department assistantship; it wasn't sponsored by any agency in those days. The first step was for me to hand-digitize these eight records by putting a T-square down with a scale and reading off the traces point-by-point, putting them in numerical form.

Goldstein:

You probably had a pretty modest sample rate!

Robinson:

Yes, I had a pretty modest sample rate, and yet it took a couple of days to do a record. And then you'd have to check to make sure. Professor Wadsworth in weather prediction had a staff of people who used Marchant desk calculators. The machines could add, multiply, and divide, but couldn't accumulate or run a program. In other words, the person using the calculator was the programmer. That summer in 1951 we actually used Wiener's prediction theory on the digitized data.

Now, there is another little factor that comes in, because at the time I was taking courses in the Economics Department with Professors Paul A. Samuelson and Robert M. Solow, who later each won the Nobel prize, and they were very interested in time-series analysis, but their approach was more traditional statistics instead of the Wiener method. Having both these points of view was very helpful, because if you think of the time, you have Norbert Wiener, Norman Levinson, Paul Samuelson and Robert Solow, and then later even John Nash, who also won a Nobel Prize. It was quite a group of people and an exciting time. I did want to see if I could apply Wiener's theory, all right. To make a long story short, I actually ran the Wiener theory, not only with the concepts of Samuelson and Solow, but also those of John Tukey. We put it all together in the summer of 1951 and the methods worked.

Goldstein:

The summer of 1951, you say?

Robinson:

Yes, because we started in the fall of 1950, and by the spring we had tried various things. That summer I had the computing done by Virginia Woodward. It took several weeks of her time using the Marchant calculator. I had to wait for her time because she was working on other projects. Then, when she was available, she was only able to give a few weeks of her time. It became very exciting, my method worked, and then Professor Hurley took the result to the oil companies and showed it to them.

Goldstein:

When you say it worked, tell me what you mean. You were able to achieve mathematically the same results that people had achieved digitally before?

Robinson:

Audio File

MP3 Audio
(326 - robinson - clip 1.mp3)

Let me explain the seismic convolutional model. We had a seismic trace which needed a mathematical model. That's where Samuelson and Solow came in, because in economics they always want a model. In geophysics, they had a model of the earth, treated as a large set of differential equations, but the solutions were impossible. To get a simplified mathematical model, we devised what is called a convolutional model. The idea is that there are many different geologic interfaces in the ground as you go down in depth. Each interface has a reflection coefficient, and the sequence of all the reflecting coefficients is called the reflectivity series. Then a source signal, called a seismic wavelet, is initiated at the surface. It goes down and is reflected from each layer. So the actual trace that you receive at the surface from one of these seismic signals is the convolution of the wavelet, which is the reflectivity series. Having that model meant that we could now do business in signal processing. And that's where all these people helped, because I had Wiener, Samuelson, Solow, Wadsworth, and Levinson. Talking to all of them, the model finally evolved. Having a model of the trace, the next step was to deconvolve it. In other words, unconvolve the trace to get back the components.

The deconvolution was the process that worked. In order for the deconvolution to work, I had to make two assumptions. One was that the interfaces laid down in geologic time were random—there was no correlation between them—so the reflectivity series was a random series. The other assumption was that the wavelet was minimum phase. Minimum phase was a term introduced by H. W. Bode for the feedback amplifier design. He wrote the definitive work on the feedback amplifier back in the early forties. By 1951, I think, he was in charge of mathematics at Bell Labs, I believe, and later he became vice president. In the mathematics group were Claude E. Shannon and John Tukey. Tukey was a half-time professor at Princeton and half-time at Bell Labs. Shannon had obtained his degree at MIT and then went to Bell Labs, where he developed information theory. Tukey was very interested in this seismic project, so I corresponded with him starting in 1951. When I visited them at Bell Labs in 1955 and showed them all the seismic processing, it was a revelation in a sense because Shannon, who was a theoretician, actually was most interested in the practical things and the data processing side. In other words, although he was a theoretician like Wiener, he had a definite interest in digital processing. Bode was very interested in my use of minimum phase, because the one he had been using was for continuous time analog circuits, whereas mine was for digital signals.

I remember the evening I went to dinner at Bode's house where his wife was cooking. Bode was gracious, but I was very nervous, and he asked if I would like something to drink. I had really never had much to drink, ever, but in the movies they would say "Scotch and soda," so I said "Scotch and soda." It was so sour I could hardly stand it. [laughter]

At one point Bode said, "You know, Tukey doesn't really understand minimum phase. Statisticians don't really understand it. Only electrical engineers do." But now, everybody gets it. Bode was a wonderful person. Also David Slepian was there, who was a nice person.

Deconvolution and computers

Robinson:

To get back to 1951. We had a method that worked, which was called deconvolution. The next question was how to compute it. It took all summer to deconvolve a few of these traces with Virginia Woodward. At that time MIT had a computer called Whirlwind. It actually started out as an analog computer in 1946, but with the success of the ENIAC they converted the Whirlwind to digital. Jay Forrester pushed digital when he got in charge. He developed the magnetic-core memory which went into the big IBM machines at the time. Like the ENIAC, it took whole rooms, or a whole building, such as the Barta Building at MIT. The Barter Building on Massachusetts Avenue had two floors. One floor was all taken up by the Whirlwind and the other floor held the offices. This was in 1952, and the Whirlwind, like the ENIAC, was a computer designed for military use. The military took about eight hours a day, and since it was a vacuum tube machine the maintenance took about eight hours a day, and the academic people were supposed to get eight hours a day, but it never worked out that way because sometimes the maintenance took more time.

So in the spring of 1952, I went to Whirlwind with Howard Briscoe and put deconvolution on Whirlwind. However, because of the limited computer time available, as soon as the oil companies were supporting us we were able to get another computer to carry it out too. The next step was Raytheon, which in about 1952 was becoming very interested in computers. Richard F. Clippinger, who I knew in Aberdeen, had worked on the ENIAC, and Clippinger was the person who first implemented John von Neumann's suggestion. Recall that ENIAC as designed was not a stored program machine. It stored data, but the programs were wired in. So every time you ran a new program you would have to re-wire the machine. Von Neumann said you could put the program into code like data, then run the machine with the internal program. His suggestion is what made the general purpose digital computer possible; before that every digital computer was special purpose and had to be re-wired for each new program. But in 1948, von Neumann said you could actually put programs in like data and then the machine would run from memory, not the wiring.

Goldstein:

Von Neumann did that in his report on ENIAC?

Robinson:

That's right. Von Neumann is responsible for the success of the ENIAC. Richard Clippinger was head of the computing laboratory at Aberdeen and also a mathematician of note. He worked out the required program for von Neumann's suggestion and ran it on the ENIAC in September 1948. Although it slowed the machine's operation, it accelerated the programmer's task enormously. The change was so profound that the old method, rewiring, was never used again. So, Clippinger was a person who had worked with von Neumann. He left Aberdeen and became head of computing at Raytheon in Massachusetts, and I contacted him there in 1952. He was with B. Dimsdale and J. H. Levin, two other great mathematicians who took on this seismic project because they wanted to do business with the oil companies. They programmed these deconvolution codes as well. They used a machine at the University of Toronto called the Ferranti computer, which was an early British computer. By 1954, R .F. Clippinger, B. Dimsdale, and J .H. Levin of the Computing Services Section of Raytheon wrote a report to the oil companies, called "Utilization of Electronic Digital Computers in Analysis of Seismograms." In it they discussed how the computer was used, so this report on digital signal processing was one of the earlier production reports on how to do it. They talked about the different machines available, the methods used, and various costs and philosophies. They also showed that the large computer would be cheaper per record than the smaller computer, even though the large computer would have hourly computation costs much greater than those for a small computer. They were advocating the use of larger computers, and this was the beginning of the age when computers were starting to become much larger and more efficient.

The Raytheon report was in 1954. Unfortunately, the oil companies were unable commercially to do deconvolution because the computers available still weren't large enough to handle all the mountains of data they had. At that time, the oil companies were converting to the tape recording of seismic data, instead of recording the traces on photographic paper, but the tape recorded data were still analog.

Goldstein:

Is that problem addressed in this report or did the oil companies discover themselves that there was too much data for the computers?

Robinson:

It's interesting, because the geophysical people look at visual images, and the early computers were not able to generate enough of these deconvolved records for them to look at. Also, the data were only just beginning to be recorded on magnetic tape. So, you might say the geophysical people weren't ready at that time. Now it's interesting because a few years later, in 1956 and 1957, I actually worked at Standard Oil Company of New Jersey (or Esso; it is now called Exxon) in Rockefeller Center in New York. This was a job that Samuelson got for me because I wanted to work for an oil company. I was in what they called the Petroleum Economics Department; I was in the Coordination Division, which reviewed exploration and refining and transportation, so I was in with a lot of chemical engineers. The chemical engineers at Esso were very interested in computers, and actually key people were promoting the use of computers in refining analysis because linear programming was then what could be used. I have a letter here written to me by J .E. Gardner of Esso, dated 13 June 1956, in which he wrote, "During our discussion we touched on the use of computers for simulating refining operations. I thought you might be interested in the attached two papers which I did not have available at that time. They cover information on the subject." So, basically, the oil companies got into computers more through refining, and more in New Jersey, because that's where Esso had their refineries—Linden, New Jersey.

Computers and refinery

Goldstein:

I don't want to spend too much time on it, but where did the computer fit into the refining part of the business?

Robinson:

When they run a refinery, they want to know how much of the feed stock to use and how much of each refined component to produce. It's really an optimization problem, not signal processing. The computers were better for that because signal processing took tremendous amounts of data. Imagine all the data in just one of those seismic records, and there's thousands of these records. Optimization problems, on the other hand, required that you produce so many gallons of this, so many gallons of that—all the data fit on one page. The early computers, which had small memories, were much better adapted to that type of problem. Signal processing was really out of bounds for computers then. Even today, on home computers you need tremendous memory to record television and things like that. But because of the oil companies' early interest in this type of computer use for refineries, within five years they told the geophysical people, "You will use computers, too."

Goldstein:

Who are you referring to when you say geophysical people?

Robinson:

Take an oil company like Exxon; it's really a holding company controlling many subsidiary companies. Two of their subsidiaries at that time were Carter Oil Company in Oklahoma, and Humble Oil Company in Texas. These were companies that explored for oil. Then Exxon had Standard Oil, headquartered in Manhattan, that refined oil in New Jersey. So, the refining people had no contact with the exploration people, except through the parent company, Exxon. The parent company saw that the refineries were doing very well using computers, and the IBM people were saying to Exxon, “You should be using more computers." IBM was very active in those days. Exxon’s management were convinced, and so told Humble Oil, “You should use computers in oil exploration, too.” The computers were getting better, so around 1960 there was a digital revolution in the exploration industry. Computers started to be used commercially because they were recording seismic data on tape, and they could use transistorized computers. The real transition came with the transistor. As soon as they had the IBM 7090, which was a transistor computer, the geophysical people could do business, because it was reliable; the vacuum tube computers hadn't been that reliable.

Goldstein:

It was a reliability problem, not a data input problem?

Robinson:

The vacuum tube computers took a tremendous amount of maintenance. All the vacuum tubes had to be checked constantly by programs and by inspection. If one tube went out you could get machine errors, which were always a problem with early computers. Today people don't think about machine errors, but they were a major problem in the early days. As soon as the transistorized computer came, you didn't have the machine errors, you didn't have the maintenance, and by then you also had a much larger memory and higher speed. Everything was good.

Goldstein:

So it sounds like you are saying the analytic tools were in place by the early fifties.

Robinson:

But the hardware wasn't.

Analog-to-digital conversion in geophysics

Goldstein:

Another limiting factor that I would have worried about is this process of manually digitizing.

Robinson:

That went out. Once they started recording on magnetic tape they could use an analog-to-digital conversion.

Goldstein:

That happened somewhere during the fifties?

Robinson:

Analog-to-digital converters were available from early days, but the early seismic records were recorded in the field on photographic paper. But then by mid-fifties, because it was a boom time for exploration, they realized it was better to record the seismic data on large magnetic tapes—the tapes were quite large then. Then they took the tapes back at the office, where they played them out on paper.

The next logical step—it was actually in the Raytheon report in 1954—was to convert the analog signals on the tape to digital. That was an accepted, well-known method; it wasn't that complicated for the oil industry to do. Basically, there were two factors working. One factor was that the amount of data the oil companies collected went up exponentially; that was possible because they were recording on magnetic tape. The second factor was that, having this large amount of data, they wanted to do more with it. A major problem in the 1950s was that they could not explore in the ocean very well because the water reverberations hid the signals. The water layer itself is reverberating like a drum head and that hid the signals coming from the depths. Deconvolution removed those reverberations.

By the late 1950s there were lots of areas they wanted to explore offshore, like the Gulf of Mexico and the Persian Gulf, and in Venezuela they had Lake Maracaibo. They realized that the only way they could get rid of water reverberations was by deconvolution. The only way they could deconvolve was by signal processing, because analog cannot do it. So they were sort of forced into digital that way.

Goldstein:

When you say by analog, what do you mean?

Robinson:

The analog methods they used would be electrical band-pass filtering, high-pass filtering, and low-pass filtering. They could also adjust the traces in time: move one trace with respect to the other. In other words, they might think, "Well, the signal's coming in at 40 hz, so we will band-pass it close to 40 hz to find the signal." That would be analog processing because they did it by an electric circuit.

Goldstein:

And that was supposed to strip out the reverberations?

Robinson:

They could use band-pass filtering if the reverberations were at a different frequency than the deep reflections. That was the idea, but it didn't work because all these signals overlapped in frequency. The companies decided they had to get rid of those reverberations, which they could do through deconvolution. That meant going digital, so they started spending the money. That was the advantage of the oil industry: they had money. The other thing is that they had a history of spending money on computers through their refining operations. They had close contacts with IBM and other computer companies. Top management people were not unfamiliar with computers, even though many geologists and other exploration people were.

What happened about 1960 was a digital revolution in geophysics. It came in with the IBM 7090, and the Control Data Corporation 1604 (which was designed by Seymour Cray). These were the first big transistorized computers. And then there was the ILLIAC at the University of Illinois and the Stretch computer by IBM. By the '60s, everybody was getting into computers, and the oil industry became a very big spender in that area.

Goldstein:

Where did the analytic expertise come from? I can see how they could buy the hardware through the sales operation of the computer companies. But where did the programs come from? Were there software packages that would do this analysis?

Robinson:

When the oil companies started supporting it in a major way, about 1953, MIT formed the Geophysical Analysis Group (GAG), which I directed. A lot of MIT graduate students came in on the project, because they were financially supported. Even though they might have been more interested in pure physics, they came in because they could get support in geophysics. So we had about eight or ten outstanding graduate assistants at any given time at GAG. Most of these people actually went out and worked in oil exploration. One of the companies that hired three or four of these people was called Geophysical Service Incorporated (GSI).

Goldstein:

The forerunner to Texas Instruments?

Robinson:

That's right. They were one of the first geophysical companies in seismic prospecting, originally formed about 1930. In 1942, during the war, everything went down and Cecil Green was able to buy up a lot of GSI stock, and he started running the company with Ken Burg and other people. About 1950 or so, they needed instruments, and there was a little company in Houston called Texas Instruments which they bought and used to make their instruments. But Texas Instruments saw beyond seismic prospecting and soon outstripped GSI, which had been the parent company—Texas Instruments became the parent company. They were the ones that actually built their own computer, called the TIAC (Texas Instruments Automatic Computer), and they put the deconvolution process on that computer. So you had companies like Texas Instruments who were actively involved in digital processing by the early 1960s, and then the oil companies could easily buy digital processing services from them. The oil companies also bought IBM computers for their own use.

Goldstein:

Did the oil companies have in-house expertise in programming, or did the programming come from IBM?

Robinson:

It came from the MIT group, essentially. The signal processing algorithms on those early computers, small by today’s standards, took a lot of programming effort (although on today's computers they are quite simple). Ninety percent of your programming effort was spent trying to fit the operations into the shoe—the small computer. It wasn't the mathematics. The actual programs that we developed at MIT were given to all these oil companies, so they had the programs in-hand. In other words, all the programs were developed at MIT on the Whirlwind, and they were just given to the oil and geophysical companies. The problem the companies faced was adapting those programs to their own computers.

The big thing that happened then was that in 1955, IBM came out with FORTRAN. Once you had FORTRAN, it was easy to program for the 1960s-generation computers. Whenever you have a problem, you always use the equipment that you have, and then you want more. So if they had a computer this size they'd say, "Well, now we need a bigger computer, and then we need still a bigger computer...." And so on. This spurred IBM and other companies to develop bigger computers.

Goldstein:

Did the exploration companies acquire all this expensive capital equipment, or were computers provided by the parent holding company, Exxon?

Robinson:

Exxon actually owned Humble Oil Company outright. Humble really was Exxon. So Humble would just buy the computers. Actually, in those days, I think you usually rented computers, though you could buy them. The oil companies didn't have any problem buying and getting the computers.

In 1955, six people left Texas Instruments and, with me as a seventh, formed a geophysical company. We had no money at all, but these six people were very adept businessmen, although I was not. We'd offer to process an oil company's records if they'd give us a contract, then we'd go to the bank and say, "We've got a contract; we need a loan to buy a computer." We bought an SDS computer, which was a good one at that time. I also went around to some oil companies trying to sell digital seismic processing—one was Conoco, another was Chevron. Our company then was called Digital Consultants, Inc., later called Digicon (an abbreviation of Digital Consultants). There was another company formed about the same time called Seismic Computing Company. So these little companies were formed that would do the processing. Meanwhile, the oil companies were also doing it—they always liked to have two ways of doing things. Chevron, Tenneco—they would do it in-house and then they would contract it out and compare results. In that way they kept their people alert. So the '60s was an active time in digital signal processing in the oil industry.

Industrial reception of signal processing, 1950s-1960s

Goldstein:

Can you give me some sense of what the attitudes were in the early '60s? Were signal processing techniques controversial, or was there consensus that they were economical and essential to being competitive?

Robinson:

Audio File

MP3 Audio
(326 - robinson - clip 2.mp3)

That's a good question. In 1953, I spent three days at Shell Oil Company talking to research people. Francis A. Van Melle was one of their top people. Shell is a Dutch company, so most of the people in the parent company, Royal Dutch Shell, were Dutch and Shell of the United States also had a lot of nice Dutch people here. They were very good scientists, as well. I went into Van Melle’s office and the whole side of the room was covered with frequency spectra of the various analog filters they used. They'd take these seismic signals and run them through analog filters hoping to filter out the noise and preserve the signal, the trouble being that the signal and noise were overlapping in frequency. So he had this whole wall covered with the amplitude and phase characteristics of these analog filters. I said, "You know, we can do that on a computer with a digital phase filter. You give us the amplitude and the phase factor and we can design a digital filter that will do that operation. There is such a thing as digital filtering that does what an analog filter does. If you specify the amplitude and phase characteristics, you can design a filter. It won't be perfect, but we can approximate it pretty well.” He did not believe it. He said, "No. There's no way."

In other words, the oil companies had a lot of different analog filters, which were state-of-the-art in the 1950s, and they had good people designing them. They had high-cut and low-cut filters and various cut-outs and other things. Van Melle had this whole room filled with all the amplitude and phase characteristics of those filters, because he was constantly comparing them. Geophysicists always hang things up on the wall to look them, and they need a big wall space. This was in 1953, and I told him a digital filter can filter just like an analog filter. You can convert the analog signal to digital. The digital filter is nothing more than a sequence of numbers, and you can convolve it with the input and get the output. Van Melle could not believe that it could be done.

Another fascinating thing happened at MIT in 1952. Dayton H. Clewell was in charge of geophysical research for Mobil Oil Company. He said "Well, we like the idea of filters, of computers and digital signal processing, but unfortunately, we do things in real time and all our analog filters are realizable, but you're not doing things in real time and your filters are not realizable. There could be no application of your method to our analog filtering because we are in real time and you are not." Although I had devoted space in my Ph.D. thesis to clear up this misconception, his mind was made up. The idea that computers couldn't really do a physical thing was prevalent among some groups. But they were becoming the minority, because most people were very interested in computers and wanted to try to use them for signal processing. The trouble was that the hardware wasn't up to doing digital signal processing yet. Digital signal processing takes large amounts of data, but the memory and the speed of computers in the 1950s were not great enough to handle that data. The early computers were much better on problems that involved a little data and a lot of computing, instead of a lot of data and limited computing.

Goldstein:

Can you re-cap the transition from when people were skeptical to when the companies started investing in it?

Robinson:

It came when Texas Instruments Company, which was Geophysical Services Incorporated, said, "Give us your exploration records from the Gulf of Mexico and we'll remove the reverberation by deconvolution." Then they showed the results, and everybody flip-flopped.

Goldstein:

And that happened in the late 1950s?

Robinson:

Early 1960s. They were working on it in the late 1950s. The geophysical people involved at Texas Instruments earlier had been my research assistants at MIT. Mark Smith was one person who was really working on this; he became vice-president at Texas Instruments. Cecil Green was so ecstatic he gave MIT a building—the Green Building—which was his first real act of philanthropy, and from there he gave professorships and more buildings to several universities.

Goldstein:

You're saying that Texas Instruments led the way, then other exploration companies developed that ability too?

Robinson:

It is fair to say that of all the geophysical companies, Texas Instruments had the ability and motivation to carry out this program. And Cecil Green had the vision to hire people from MIT. They built the TIAC computer, which was quite a feat in itself. Meanwhile, the other oil companies had been working on the same thing, but always in secret, whereas a contractor had its work advertised. Once the contractor—Texas Instruments—advertised it, then a lot of people were saying, “Well, we can do it too!” Since it was relatively simple to do the signal processing, anybody could buy a computer and do it, and that meant a lot of small companies started up, one of which was Digicon, in 1965. As I mentioned, six out of the seven people that founded Digicon were Texas Instruments people.

Digital filtering and computer programs, 1950s

Goldstein:

I'm surprised to hear you say that digital filtering could be done as early as 1953. The impression I had was that you don't start seeing digital filters developed until the late 1950s.

Robinson:

This is my actual thesis written in 1954, and it was also reproduced as an MIT Geophysical Analysis Group report. This figure shows the actual filter coefficients, and this shows the power transfer function, which is really the square of the amplitude spectrum. And this shows the inverse of it. So, I had both the amplitude spectrum and the minimum-phase digital filter. We were actually filtering digitally then. We had the amplitude spectrum from which we computed the minimum-phase digital filter. And here is another example: This figure actually shows the linear operator, that is, the filter coefficients. This is the inverse linear operator. And this shows the amplitude spectrum and the phase spectrum. So that is the whole thing right there. And that's what I was telling Van Melle, because he had the analog filter characteristics—the impulse response and the amplitude and phase spectra. He actually had pictures like these which were done by an oscilloscope, and I was saying that he could actually do the same thing digitally.

Goldstein:

What was your motivation to work on digital filtering at that early time?

Robinson:

One thing is that in the oil companies probably half the geophysicists were electrical engineers. In other words, when you have an electrical engineering degree, you can do anything. If an oil company is doing exploration, all the equipment is electrical or electronic, like the geophones which convert ground motion into electrical signals which are then passed through analog filters. So you need electrical engineers, and a lot of their people were trained in electrical engineering.

At MIT I started working with the oil people in 1952. The first couple of years I was working with Wiener and other mathematicians who were talking pure mathematics. But once I got to the oil company people and told them that digital filters could do much more, they said they wanted to know the impulse response, amplitude and phase factor of the filter. So I computed the filter characteristics, and I did everything I could to put digital filters in terms familiar to electrical engineers.

Goldstein:

But why digitalize? What was wrong with analog filters?

Robinson:

We started out digital—I never used analog. I started with the mathematics of Wiener, and used his various approximations to switch his continuous-time formulations to digital. I had a digital computer, but I didn’t have any electrical equipment, so I was forced to do it digitally.

Goldstein:

Because your methodologies were discrete, you wanted to work in that domain entirely?

Robinson:

That is right. I never had to convert—I started out digital. A person at MIT who influenced me a great deal was Y. W. Lee in the Electrical Engineering Department; he taught a course on Wiener's theory, which I took in 1951. Subsequently he published the essence of this course in his book. He had worked out all of Wiener's analog theories, which were just beautiful. Now, Wiener and Lee were working with analog circuits, but from where I came from in the Mathematics Department we were working with Samuelson, Solow and other people from the Economics Department, and everything was digital. The economics data were digital. The economic models had discrete times—every year, every month, or every day. So, I started with digital and never started with analog.

Goldstein:

But the geophysical data does start out analog, doesn't it?

Robinson:

Yes, the geophysical people were all analog, but I couldn’t do what they did. I had none of their instruments or equipment. All I had were pencil and paper. So I would digitize the continuous-time seismic traces, and then enter the digital data into a computer. If I had had the equipment they had, I never would have done it the digital way. All I had was a digital computer, and that's the only advantage I had. Otherwise, they had everything— and plenty of money—but they didn't have the digital computer, and they didn’t really use computers. I had a digital computer, starting with programming the ENIAC in 1951, and then in 1952, I got the Whirlwind computer.

Getting back to programs, the Whirlwind was built in the Digital Computer Laboratory of the Electrical Engineering Department at MIT. Charlie Adams was in charge of that Laboratory. I thought, “Well, if I just go over there, they must have all these programs for solving linear equations, inverting matrices, and so forth.” But they didn't, because in those days, just the development of the computer itself and the programming of basic operations needed just to use the computer took all the time; the applied mathematics had not been done. So I did a lot of the applied mathematics programs myself, in 1951 and 1952. Then of course when IBM got into the picture, they did the applied mathematics.

Let me see if I can pick out of this story a couple of key things. I had the deconvolution model working in 1951, before I related it to electrical filtering. I had Wiener's prediction theory, which is today called linear prediction coding, working on Whirlwind in 1952. Then in the next couple of years, in order to justify what I was doing, I got into what was going on. I related the digital filter to analog, and I would compute the amplitude and phase.

Goldstein:

There was other work going on at MIT around that time in deconvolution and digital filtering. Al Oppenheim was using his homomorphic filtering for speech processing. Was your work connected to that, or was it independent?

Robinson:

That was a different generation, because our work was from 1950 to 1954. The only person in electrical engineering really working with the Wiener methods then was Y. W. Lee. People like Al Oppenheim, Tom Kailath, and R. E. Kalman were younger, and came after that—they came in the late 1950s and 1960s. Oppenheim and other electrical engineering people worked on digital filtering and all the new things at MIT from the late 1950s on. I did get to know those people, but I wasn't at MIT when I did.

Goldstein:

Did their work bear on the geophysics problems?

Robinson:

Yes. Oppenheim did the homomorphic work, which was excellent. All his students have done very well, and some of them went into geophysics. In the 1970s, at various conferences, when all of the scientific work was done, we'd get together at some nice restaurant. We had more of a social time then. It was very pleasant, because Oppenheim is a wonderful person. What MIT would do is develop something, and then they would also teach it and write the textbooks on it. Oppenheim took digital signal processing and gave it to the world, so to speak. He could do everything—he was in research, teaching, and education.

But my focus, in the early 1950s, was to try to satisfy the oil companies. Basically, the thing that converted them was when they saw that deconvolution worked on marine data. That took place in the early 1960s, and it was made possible by the transistorized computers.

It's interesting, I recently realized that I had never really given Samuelson and Solow credit for some of the work they did. Back in the spring of 1950, Samuelson taught a course on prediction which I took. I also took Wiener's course, so I had exposure to both approaches. Samuelson and Solow were interested in time series analysis. So, last year to Paul Samuelson and Robert Solow at MIT, and I told them I was trying to give them a little acknowledgment. I wrote, "Belatedly, I am now in the process of making an acknowledgment of how much the discipline of geophysical signal processing owes to your pioneering works and influence. Both of you are always in my mind when I look at time series and models." Bob Solow wrote back, "Many thanks for your letter, which brings back memories. Can it really have been almost fifty years ago? The answer is, yes it can. I was in my twenties, you were a mere lad, and even Herman Wold was young and vigorous. Your summary of early deconvolution will be an eye opener to today's students, who have all the computer packages available and take ideas for granted that were new then. I am glad you are making a record of those events, and proud to have played a part in it."

I think that it is hard for people to realize how difficult those early computers were to program. Now, to give you an example, when I first programmed deconvolution on the Whirlwind in 1951 and 1952, I was very fortunate because I got to know an undergraduate student, Howie Frisco, who was what today we call a “computer person”—I mean, he loved computers. He had gone over to the Digital Computer Lab at MIT and learned how to program. Because he was in geophysics, he and I worked on programming this deconvolution. That machine was very bare bones—just clear, add, subtract, and multiply—no division or floating point arithmetic. You could do conditional transfer and things like that. Howie and I worked all the spring of 1952 on that. And then Howie had to go away in the summer of 1952, and the program wasn't quite working. In those days, you gave the hand-written program to two typists, each one whom typed the program on paper tape in binary form. Then they compared the tapes. If there was a mistake then the two tapes didn't match, and they had to re-do them. That way, you got your proper binary tapes. There was only one teletypewriter in the lab, and it was on the machine, so you couldn't print out the tape; you had to enter the tape into the computer and then run the program. If the program didn't work, you could just get a machine dump that showed all the binary digits in the computer when the program crashed. Well, the deconvolution program didn’t work. I spent about a week trying to figure out the dump. What happened was, on the page where I wrote the original program in pencil there was a little mark on the paper, and BOTH typists had put a comma in where there shouldn't have been one. I told that incident to my brother, and he said, “Computers will never work.” That was July 4, 1952.

Deconvolution in geophysics

Goldstein:

Is the deconvolution problem the same in geophysical applications as it is in other early applications of signal processing, like speech or radar?

Robinson:

There are special things in geophysics. The bandwidths are different, the signals are different, the types of noise are different. But if you just define deconvolution as unconvolving something, then the basic idea is the same. For different types of applications the models are different. In speech, it's basically the same model, but there is one fundamental difference. Seismic is a reflection problem, in that you have the source and the receiver at the same point. It's like radar. Radar has a reflection problem, and so does sonar. Now, speech is a transmission problem, because the speech originates in the lungs and goes through vocal chords to the lips. It turns out that the reflection problems are easier to solve than transmission problems—and that’s why you might say seismic analysis is easier.

Gulf Oil Company; Standard Oil Company

Goldstein:

Tell more about your own research during this period in the 1950s. How did your research fit into the toolbox of the possible?

Robinson:

I graduated with my Ph.D. in 1954 from MIT, and went to Gulf Oil Company in Exploration Geophysics. They were not interested in digital computers, so I was learning conventional geophysics—"Stone-Age" geophysics.

Goldstein:

You mean that they did everything in analog?

Robinson:

Right. But it was interesting in the sense that they were exploring in Kuwait, so I learned about Kuwait when nobody had ever heard of it. But except for that, it was only knowing about the rock structure somewhere—it wasn't really mathematics. I was using geophysics to find oil, and not doing mathematics.

Basically, I've always been an applied mathematician. I preferred to do other things, like digital mathematics or digital signal processing. So I went back to MIT in the Mathematics Department and was again with Wadsworth and Wiener and John Nash and all those people. And then at MIT, I was talking to Professor Samuelson and he said, “I know the Chief Economist at Standard Oil Company in New Jersey.” He called him up and they gave me a job, so I worked in New York for a couple of years, from 1956 to 1957. That's when I got to know the computers from the other side, because they were very interested in using computers to do refining problems, but these refining problems weren't data-intensive like signal processing.

University of Wisconsin

Theory, mathematics

Robinson:

Then Sputnik went up and everybody wanted to get back into science, so I wanted to go back to a university. I went to the University of Wisconsin and I went back into digital work. I became director of what they called their Computer Science Laboratory, which was then not a department—I was in the mathematics department. So I was back in mathematics and doing signal processing, the Wiener-type thing. I was working on theory a lot, especially about minimum phase. There were certain theorems that came out of what I was working on. I was able to prove a theorem which actually got into a textbook, Fourier Series and Integrals by H. Dym and H. P. McKean. This text is used at Harvard and other universities in graduate school mathematics. Here Dym and McKean are talking about Hardy functions and filters. This is pure mathematics. It says, "Outer filters (i.e. minimum-phase filters) have the biggest power. The fact alluded to is that among all filters with the same gain, the outer filter makes the energy built-up as large as possible, and it does so for every positive time. This beautiful fact is due to Robinson in 1962. It does not seem to have found its way into the purely mathematical literature.” This sort of made me happy, because their book was on Fourier series, and I have a theorem in it. When I was an undergraduate at MIT, Professor Salem taught the course in Fourier theory, and he was an outstanding expert in Fourier that subject. I took a math course from him in my junior year in 1948 and it was analysis, which in the Mathematics Department was a key course. I usually got in the nineties on my math exams, and I got like an eighty-five or something on one exam, and I was shocked. I felt I didn't understand anything. I was depressed, and so I went to him and said, "I don't know what has happened; I had a bad day." He said, "Well, some people can do mathematics and some people can't. Maybe you should drop out of mathematics." In any case, I didn't, but worked hard and got an A. But now I feel good because my theorem got into a Fourier series textbook. I feel it would make my old professors happy, and in those days—the good old days—professors were like gods.

Goldstein:

What did your 1962 paper prove? What techniques did it bring about?

Robinson:

My thesis was called "Predictive Decomposition of Time Series with Applications to Seismic Exploration." Decomposition is deconvolution, because they used to say “composition” instead of “convolution.” It was predictive because it used Wiener's prediction ideas.

Goldstein:

How did you extend Wiener?

Robinson:

The Wiener theory was being used for predictions, and this use did not require any physical model of the signal. What I did was to form a geophysical convolutional model, where you actually identified a correspondence between things in the mathematical expression for the signal with and things in the earth. You could say that the earth produces a seismic trace which is a convolutional signal. That is, you have the seismic trace as the convolution of a wavelet and a reflectivity series. The wavelet is minimum phase, and the reflectivity series is a white noise series. So we now have linked a mathematical model of the seismic traces with the physical model of the earth. Having that physical model, you can then operate on the trace, deconvolve it by digital signal processing, and recover the two components, namely the reflectivity series and the minimum-phase wavelet. Basically what I did was to take Wiener's theory and connect it with the earth, so it could be used. Without that connection it would make no sense, because geophysicists wouldn't know what they were getting. Today the convolution model is completely accepted, and geophysicists use it all the time.

Again, my innovation was getting the model, and that's where I think Samuelson and Solow came in, because economists were always looking for a model of the real world, whereas mathematicians were looking more for mathematical relationships as such. There was no earth model in Wiener’s time series book—nothing connected to the physical earth, only very deep and difficult mathematics. Still, the convolutional would not have been possible without Wiener’s help in explaining his mathematics to me. It was good fortune to be able to know Wiener and work closely with him

Minimum phase and seismogram

Robinson:

Also very important to me was the influence of Bode and Shannon who wrote a critical paper on minimum phase. Basically, my work was taking mathematics and connecting it with physics. Once you had that connection, then you could use the digital signal processing for other things, such as removing multiple reflections and reverberations.

Goldstein:

Could you give me a quick introduction to minimum phase?

Robinson:

Bode introduced the idea. If you have a network with a feedback loop in it, your output depends on the input and the feedback of the previous outputs. This feedback process has to go at a certain speed, otherwise you'll get an unstable output. The speed at which it goes is controlled by the phase function, and it has to have a minimum phase to yield a stable feedback system. Bode actually introduced the term “minimum phase.” Now, suppose you plot the phase spectra of two systems, both are the same gain, that is, they both have the same amplitude spectrum. If one system has minimum phase and the other system doesn't, then the phase spectrum of the minimum phase system will lie under the phase spectrum of the other system. It actually is below. So for any system with that gain, it's the minimum-phase system that has the smallest phase shift for any frequency. This frequency-domain description was the way that minimum phase existed in electrical engineering. What I did was to figure out what minimum phase means in the time domain. What does the impulse response look like? In the time domain, the impulse response is made up of a sum of sinusoids of different frequencies. The amplitude of the sinusoid for a certain frequency is given by the value of the amplitude spectrum at that frequency, and the phase-shift by the value of the phase spectrum at that frequency. For a minimum phase system, the phase shifts are smaller than any other system with the same amplitude spectrum. Because the minimum phase system is actually the system with the smallest phase shifts, it follows that it has the smallest energy delay. And that smallest energy delay gives stability to the feedback loop. When driving an automobile, a person notes the discrepancy, or error, between his steering and the direction of the road.

There is always a time-delay between the instant the human operator observes an error and the instant he starts to take corrective action. The operator with the minimum-delay is the safest driver. So in other words, if you have a feedback loop, you want that loop to return energy as fast as possible, and that occurs for a feedback loop with minimum delay. If the energy is spread out in time too much, if the energy is delayed too much, then the energy is returned too late, so you get the instability. So that's the relationship. So minimum phase means minimum delay. What does that look like in the time domain? The geophysicist sees a signal as a function time, and what does minimum delay mean to him? Now suppose you have two wavelets, or two impulse responses, and one is minimum phase, and the other is not, and they both have the same amplitude spectrum. The minimum phase impulse response has the sharpest leading edge. It has the fastest growth of energy with time. It has the quickest energy build-up. So a minimum phase impulse response has most of its energy up front. In contrast, a non-minimum-phase impulse response would have most of its energy in the middle or at the end. In this picture, that's the minimum phase impulse response, the one with the big coefficients at the front. That's my theorem: a minimum phase wavelet has the fastest possible energy build-up of any wavelet with the same gain. A prime example of a minimum-delay wavelet is a curve that decreases exponentially with time; the big part is at the front. If a signal is the input to a minimum delay system, then we can recover this signal in its original form by passing the output of the minimum-delay system into its realizable inverse system. The recovery of the signal is accomplished with no overall time-delay. On the other hand, there is no realizable inverse system for a non-minimum-delay system. For some non-minimum-delay systems, the original signal can be recovered, at least approximately, but with an overall time-delay. For other non-minimum-delay systems, especially multichannel ones, the original signal cannot be recovered, even with an indefinitely long time-delay. Does that sort of explain it?

Goldstein:

Yes, that helps.

Robinson:

So in other words, there are two ways to look it. In the frequency domain, the phase curve is as low as possible, so it's minimum phase. In the time domain, the energy is up front, so the impulse response curve rises swiftly and has a big whack up front and then it tails off. And a curve that rises slowly and has its big whack in the middle or further back would not be minimum phase. You can see why in feedback you'd want the big response up front in time.

Goldstein:

Right, I can see why this is important in terms of designing a feedback system. And so where does this figure in analyzing a seismogram?

Robinson:

A seismogram. The way it figures in is that originally the source signals that were put in the ground were explosions. The energy travels down through the broken and crushed rock produced by the explosion, and then the energy reaches the unbroken rock, which is elastic, so that a wavelet is formed. Because this wavelet is caused by an explosion, the energy in the wavelet is up front, so the wavelet is minimum phase. The actual physical situation gives a minimum phase wavelet. In the application of deconvolution, it is assumed that within the time gate (1) the wavelet is unknown, but is minimum phase, and (2) the frequency content of the reflectivity is completely white. Thus the frequency content of the wavelet on the trace can be obtained by taking the autocorrelation of the trace. The reason is that the power spectrum of a white reflectivity series is completely flat, so the colored frequency content of the wavelet is the same as that of the trace.

The deconvolution filter is computed by the least-squares method, which leads to a set of normal equations, in which the known quantities are the autocorrelation coefficients of the trace. The deconvolution filter is found by solving the normal equations. The purpose of deconvolution is to remove the wavelet from the trace while leaving the reflectivity intact. The deconvolution filter is the least-squares inverse of the minimum-delay wavelet. When the deconvolution filter is convolved with the trace, the deconvolution filter is, in fact, convolved in turn with all of the wavelets that make up the trace. The operation converts each of these wavelets, as well as possible, into a spike, greatly improving seismic resolution. Then, when geophysicists started using other types of signals as sources, like chirp vibroseis signals and air gun signals, then, of course, the wavelets produced were not minimum phase, and so they had to be converted to minimum phase in order to apply deconvolution. So the way things are done now in geophysics is that they try to use minimum phase equipment and signals as much as possible, all down the line, and they convert the rest to minimum phase, in order to carry out deconvolution. At the final step they convert the deconvolved signals into nice symmetric wavelets of zero phase, so the final section looks better to the interpreters.

Computer science laboratory

Goldstein:

Let's get back to the chronology of your career.

Robinson:

We're back to the year 1958. Well, I went to the University of Wisconsin, and fortunately another professor there had worked with John von Neumann. The Wisconsin professor was Preston Hammer, and he was in the mathematics department. He was very interested in computers, but basically Wisconsin had no large computer for university use. They had an IBM 650, which was a very small drum computer. But they had lots of punch card machines because several of the economists and some other people had been using them. Then Preston Hammer went away, and I became head of the computer science laboratory which, as I have said, was part of the mathematics department. The demand for computers was increasing. All the various departments wanted more computing time, but it couldn't be done with the existing equipment. So we applied to the National Science Foundation and got a $1 million grant to buy a big computer. The Wisconsin Alumni Research Foundation (WARF) put up an equal amount of money, so we had $2 million and we were able to buy the Control Data 1604 computer, which had been designed by Seymour Cray. Because of this new computer, the computer science lab could do business. Because of that, WARF gave me a fellowship to go over and work with Herman Wold in Sweden, who was one of the pioneers in time series analysis. I stayed in Sweden a while, and when I came back it was the digital revolution in geophysics, so I got into geophysics again, first with Geoscience, Inc., and with Digicon.

Digital signal processing and oil exploration

Goldstein:

How did the digital revolution in signal processing effect the oil exploration business?

Robinson:

The thing that converted electrical engineering to digital was the fast Fourier transform. Tukey had been looking for that for many years—it wasn't just something he did instantly, he had been working on that problem. Somehow he had this vision that we needed that fast Fourier transform, and we did because convolution takes a lot of computer time in the time domain, but if you use the fast Fourier transform, it's much faster. A lot of operations were possible with the fast Fourier transform. Tukey knew that, worked on it, and published it with Cooley in 1965. I think as soon as he did that the electrical engineers went digital. That was a big transition in electrical engineering. The oil company people had switched a few years before because they could see reflections in marine records by deconvolution. They had good results on actual field records. For example, at about that time they were exploring the North Sea and they used deconvolution there and it was successful. But there was a problem because it took a lot of time to do the convolutions involved in deconvolution, and instead of doing them in the frequency domain, they had IBM develop what's called the Array Processor, which is a special purpose machine designed to do convolutions. They used that for many years. You might say the geophysical people were sort of isolated from the electrical community until the mid 1970s. The electrical engineers in the oil exploration community became geophysicists, but not the other way around.

Goldstein:

When you say that they pushed IBM to develop the Array Processor, can you tell me more about that?

Robinson:

They were doing a lot of convolutions and it took a lot of time, so they wanted a special purpose machine that would just do convolutions.

Goldstein:

Who are they?

Robinson:

The individual oil companies. The funny thing is that while the oil companies are very competitive, the scientists within the organizations know what everybody is doing. They would go to meetings and talk with each other. It became apparent that they would need a faster way to do all this, so they actually built a special purpose digital computer for that, called the Array Processor. With the introduction of more powerful computers, they do not use Array Processors anymore.

Goldstein:

You said that it was already possible to do this kind of work with deconvolution. Did the oil exploration or geophysics applications benefit from the fast Fourier transform?

Robinson:

Yes. Deconvolutions were done in the time domain. What turned out was that the geophysicists who actually did the interpretations always looked at things in the time domain, not the frequency domain. They were comfortable in the time domain, and if the computations were done in the time domain, they were comfortable with that. But computers by themselves were not really big enough or good enough to do extensive deconvolutions economically in the 1960s. With the addition of Array Processors, they were okay. To do the big jobs they want to do today it is a different matter; the old machines would be too small. But for the objectives then, the machines were good. They were pushing the limits even then, but the machines did work.

Goldstein:

Tell me how we've gotten from there to the present day. Why did the job change?

Robinson:

I think what happened was an evolution. If you go back to the 1950s, there was no such thing as digital signal processing being taught at the universities. Then in the 1960s, Alan Oppenheim, Ronald Schafer, and others did digital signal processing and wrote textbooks. So, by the 1970s there was an educated group of people hired by the oil companies. By the 1970s, everything sort of integrated—it was a natural sort of thing. The 1960s were probably the most exciting time, because that's when computers came into their own. And today is another exciting time, because you’re in a whole new level of computers, but in the 1970s Oppenheim actually went to oil companies and talked to them. So there was a general linking up at that time.

Goldstein:

What was there to add to the oil companies' technique? You said that they were able to do the analysis that they needed using the deconvolution techniques in the time domain.

Robinson:

As soon as the oil companies could process so much data, someone said, "Let's collect twice as much data and do so much signal processing.” Then someone else said, “Well, let’s collect twice as much data to get twice as good answers." The data collection became much more efficient with the new equipment. The magnetic tapes were better, and so were the recording instruments, and other things. There was quite an instrumentation development. Geophones that receive the signals on land used to be big heavy things that had to be laid out with great difficulty. As they evolved, they became very tiny and could be laid out very easily. You could even use radio waves so you wouldn't have to connect the geophones with wires. The marine cables, which held the hydrophones that received signals at sea, also evolved, so the explorationists were able to collect much more data. Collecting all this data, they wanted better and better answers. So the signal processing people were faced with huge amounts of data, and to do better processing, they developed new programs. They needed, for example, programs that would give better estimates of seismic parameters such as the wave velocities. The explorationists needed bearings as well as ranges—they needed all sorts of things! So various new types of digital processing methods were developed which were suited more to just the oil industry—the processing brought them closer to the actual physical structure of the earth. One of those techniques is migration, where you try to sort out all the signals and put them in the proper spatial position. Basically, their appetite and aspiration levels went way up. If they were satisfied with what they were doing in 1960 it would be a static field, and they wouldn’t have needed new processes—but as soon as you get people doing one thing, they want a bigger thing.

Geophysics and disciplinary divisions

Goldstein:

Right in the mid-60s the IEEE Group on Audio started getting interested in signal processing and took advantage of the excitement over things like the digital computers, digital filters and the FFT. Were geophysical engineers involved in that new group, or did they follow their own kind of activity?

Robinson:

They followed the new group, but they were also very concerned with their own special problems. Their problem was to collect data, then produce a map of where to drill the oil well. They faced problems having to do with the earth. For example, because the terrain has hills and valleys, they had to make corrections. Computers could be used for this, too, so they went into an area that's called inversion. Deconvolution is a type of inversion, but full inversion involved a much more complicated model than the simple convolutional model. The trouble is that the mathematicians run away with the problems, but in the end they always have to come back to basic signal processing techniques to solve them.

However, I think by 1970 or so there was quite a bit of mixing of groups, and geophysics wasn’t really something that was dominated by just geophysicists. It's interesting, because in every generation a new group of people comes in and forms a new batch of companies. What's happened in the last ten years is that while the old signal processing methods still work, today, with the greater amount of data available, you can produce very beautiful pictures to display, and you see so much more. The emphasis today is on developing new visualization algorithms because processing is done essentially by the old algorithms, such as deconvolution and migration, that have been around for thirty years or more. In fact, quite excellent results come from applying these old algorithms to a lot more data and then putting out beautiful pictures, which they can do with visual displays today.

Another big change is that one geophysicist now fills many roles. Traditionally, there have always been three steps—seismic acquisition of data; processing, and interpretation—and they were done by three different groups: People in the field collected the data; people in the computer lab processed the data; and interpreters looked at the processed data. Now, with the workstations and everything, they can actually put computers in the field, and it is possible do all the steps right there and see what's going on. You just pull down a menu and point and click, where in the old days you entered numbers and things. Now they have default values so if you don't know what parameters to put in, the program will get the parameters for you, and then optimize them. During the past five years there has been a transition, turning the whole of digital signal processing from an art form where people had to know the mathematics, to what it is now, where someone looks at a computer and pulls down a menu and picks something, and they see a beautiful picture. He doesn’t quite like this picture, so he clicks some other spot and brings that up. It is now at the point where you don't really have to know the theory of digital signal processing to do digital signal processing. It's a beautiful thing; the only problem is it's very expensive, because the companies that have developed these menu-driven, user friendly programs, like Landmark Graphics, will charge you a lot of money to buy their programs.

Geophysical processing tools

Goldstein:

A second ago you were giving examples of processing tools and I want to be sure to get a complete list. Thirty years ago you had deconvolution and migration; what else have you got?

Robinson:

One of the things that I didn’t mention that is very important was that in the 1950s Harry Mayne developed the common depth point method (CDP), which is based on redundancy. It goes back to Shannon's idea that if you want to receive a signal through a lot of noise, you repeat that same signal over and over and then take an average. By averaging all that redundant information you can find the signal and get rid of the noise. So instead of illuminating the subsurface from one perspective, by one set of shots, you try to illuminate every depth point from many different angles. They shoot shots from all over the place and cover every different depth point many different ways. They have what they call folds. One fold would illuminate the whole subsurface once; with twenty-four fold data, it is essentially illuminated from twenty-four perspectives—and then to forty-eight, and ninety-six, and even up to 1,000 fold. By using this redundancy you can get rid of a lot of unwanted signals. This common depth point method could be done by analog, and it was developed for analog. But once they went to digital, it became so much easier. Multifold coverage is something that is specific to geophysics. That is a basic method that is still used. From all these different illuminations they could figure out the velocity structure: how fast the wave is traveling in each strata.

Goldstein:

When you say it can be done analog, what do you mean? Are they done simultaneously?

Robinson:

Yes, by using the geophysical process known as stacking. Stacking takes the data obtained by the CDP method and then, using Shannon's idea, adds the data together with the necessary time shifts so as to eliminate the noise and preserve the signal. They call it stacking, but it's really adding. You can actually add signals by analog. In the 1950s these signals would be fed by magnetic tape into an analog computer which would just average them together with dynamic time shifts.

Goldstein:

But not in real time.

Robinson:

It's done from tapes by an analog machine. So in other words, anything that depended on adding traces together and making dynamic time shifts could be done by analog. But when you got to deconvolution, it had to be done by digital, because you were actually designing and building a digital filter. The oil industry went to digital in the early 1960s because of deconvolution, but once they had digital computers, they could do all the other things that had previously been done by analog with essentially no additional cost, because they had already bought the computer. So they became very enamored with the computer. Originally, migration used to be done by paper and pencil—the interpreter would mark the reflected events and then he would take straightedge and compass and move the events around to their correct positions in space. A lot of the interpretation then was done by paper and pencil, but once you got the computer, then stacking and migration could be done on the computer, and they found out that it made life simpler all around.

So the big processing operations in geophysics are: The common depth point method which led to the stacking process to eliminate noise; velocity analysis; deconvolution; and migration, which is really based on Huygen’s principle and the Kirchoff integral solution of the wave equation. Then you have the static corrections, which take into account the terrain differences directions. Then you have source signature corrections. If you are not using a minimum phase source signal, you have to use source signature deconvolution, which will take out the non-minimum-phase component. All these methods were in place in the 1960s, but they were very time consuming on the computer so they could only be used to a limited extent. Now they've gone to what's called three-dimensional data collection (3-D). Before 3-D, all the data were collected with sources and receivers on a seismic line, and so by adding the depth dimension you only got a two-dimensional slice of the earth. Now they are working with sources and receivers on a two-dimensional grid on the surface, so adding the depth dimension you get a volume of the earth. You get so much more information with 3-D because, before, with 2-D, if some energy was coming from the side—out of the plane—you couldn't see where it was coming from. But with 3-D you know where the seismic energy is coming from. Of course the computer time required by this three-dimensional processing goes up by a whole order of magnitude, so while 3-D is very successful, they do need the big computers.

Goldstein:

In the late 1950s, deconvolution provided the breakthrough that made underwater surveying possible. Was there anything of that order of importance between the late 1950s and now?

Robinson:

Audio File

MP3 Audio
(326 - robinson - clip 3.mp3)

Unlike the other commonly used processes, deconvolution had never done before, because it's completely digital. The migration process was done by ruler and pencil as early as 1921 by J. C. Karcher. Then J. G. Hagedoorn in 1954 wrote an article which really showed all the geometry. Migration was then converted to digital. In the conversion to digital, geophysicists used the finite-difference wave equations and the Kirchoff integral, which taught them a lot more about how the waves propagated. So it has been a learning process. Stacking using common depth point data was done first by analog computer, which involved adding with time-shifts, but then that was converted to digital as well. It could be said that everything was in place, but the conversion to digital still represented a revolution. Another revolution—the 3-D revolution—is taking place today. You do a two-dimensional survey you see certain things, but when you do a three-dimensional survey, it’s like all of a sudden you see a whole new picture that you never saw before.

The exciting thing today is in areas like the Gulf of Mexico, the North Sea, and the Persian Gulf, where deep in the rock layers under the water are big salt domes, and oil is found on the flanks of the salt domes. But the seismic method was never able to look beneath the salt domes. Now, with today's three-dimensional methods, they can lay out a large survey and look beneath the salt domes. With 3-D you can actually see the shape of the salt domes, so you can go back and look for oil deposits below them. As a result, in the last two or three years, there have been tremendous oil discoveries in the United States in the Gulf of Mexico. Looking for oil with three-dimensional seismic methods has opened up exciting new possibilities.

An interesting thing is that often the two-dimensional and three-dimensional theories are the same—you just add the third dimension. The calculations are not very different, but nature is really three-dimensional, so 3-D gives you a better picture. The revolution today is found in the beautiful pictures they get in 3-D work. They get some gorgeous pictures of the sub-surface. It takes much of the guess work out of interpretation.

Something interesting about computers is that in a way, scientists used to content to get a computer and do number crunching. But once the public got computers, and the children too, they wanted too see beautiful pictures and to manipulate them. The computer industry began developing visual displays for the general market. For years, scientists were content with a lot of printed pages of numbers. You can look back in time, and see the scientist with all those old computer print-outs they used to look at all the numbers. Then once the general public got into it, the visual thing was pushed and now the visual thing has come decisively into science. It's a two-way street. Science helps the general public, and the general public is helping scientists. It’s amazing!

Diverse applications of geophysical signal processing

Goldstein:

peaking of a two-way street, did any signal processing methods developed for geophysics find their way into broader signal processing applications?

Robinson:

I think deconvolution has done that, but a lot of other things are special to seismic. There is also radar signal processing and sonar signal processing, and they have had some contact with seismic, but basically I’d say radar and sonar people work independently of the geophysicists. But they do read each others papers. The other aspect that's interesting is medical signal processing. Even as early as 1965, Digicon had a woman whose husband was a doctor, and she was trying to get into medical imaging. Of course in those days medicine had various types of imaging, but they were looking toward imaging with ultrasonic waves. Geophysical people got into medical in a little way, but could never really make much progress because of the instrumentation and because they weren't close enough to the problem. But now, if you go to the Mayo Clinic, they have an echo-cardiogram which essentially uses reflected ultrasonic waves. It can tell you more about your heart than you want to know—what every valve is doing or if there's any leakage. The doctor can use an instrument made by Hewlett Packard to look at your heart on a screen. You're connected up with these little ultrasonic transducers, and he can see inside you without cutting you open. So these signal processing methods have found their way into the medicine, and geophysical people have tried to get into medicine, but because of the special complications, the people actually in the medical field produced the instrumentation. Overall, geophysicists have not been very successful breaking into other fields. It's interesting, because here at Columbia we have two brothers, the Chudnovsky brothers—David and Gregory. They’re world-class mathematicians and computer experts, and they are interested in geophysics. But when you talk to geophysicists, you realize that they would like the Chudnovsky brothers to go all the way and become geophysicists—to get really involved in it. They do not seem to realize that the two brothers can really work in many fields at once. Geophysicists like to be fully devoted to their field.

Geoscience Incorporated

Goldstein:

You've been mentioning your company, Digicon. What's its history?

Robinson:

In 1964, after I came back from Sweden, I became vice president of a little company called Geoscience Incorporated, which was next to MIT. It had a lot of the MIT people in it, including the president, who was an MIT professor, and it carried on the traditional MIT approach. Shell Oil came to me, and I got a big contract for Geoscience to get Shell into digital. It turned out that people like Van Melle, who had ridiculed digital processing in 1953, by 1964 were great supporters of it. In other words, Van Melle was the first to admit that digital was the way to go. Here is the Geoscience contract with Shell Oil Company, dated December of 1964, along with notes I took at the time. Shell had immediate problems like normal move-out, static corrections, stacking—these were things they wanted to improve. They had access to the IBM 7094, which was the successor to the IBM 7090, and shortly they got an IBM 360 M40. What they wanted from a long-term standpoint was improved access to the computer, both research and in Shell divisions. They wanted automation of the interpretative processes, which isn't even routinely done today, you know. They wanted better displays—ways to plug in the human more effectively. They wanted help in developing systems philosophy. They wanted an introduction to our ideas and how they might be effective, and instruction and training on specific things to develop general capability in seismic areas. They wanted to get into this, and they essentially came to the MIT group, which now offered its services through the Geoscience company.

Shell also had the problem of high-speed conversion of paper seismic records and well logs to digital tape. They had to adapt scanning devices. They had problems of channeling data. See, there's a lot to just handling data in signal processing; it's a big job. And for seismic research, they wanted high-speed correlation, methods for multi-trace data, and further development and evaluation of correlation techniques. The end objective was the reliable and automatic picking and rating of seismic events, something they can't do fully even today—the human being can still see a picture better than a machine. The machine can process the data, but when you come to the pictures, human beings can see things better. Shell wanted automatic stacking done and dynamic corrections made by computer. Shell wanted determination of move-out, specification of digital filters and parameters, evaluation of signal-to-noise conditions. How does a human interpretation form its operations? Can a machine do it as well? In other words, when they say Shell Oil was like a university, it was because they really wanted to get into computing on a deep level. They already had educational programs for their geophysicists. They wanted to offer course to exploration people in digital processing. In their research program, Shell wanted to apply the things we had, the new theories: three-dimensional; adaptive; modular computer programs not linked to specific hardware. Shell formed an information and Computer Services Section to offer services for getting data off and on the machine, and to do back-up and programming for seismic processing. Shell’s research programs included such topics as similarity from trace to trace, static corrections, velocity analysis, and wave fronts and rays. Shell had deconvolution, but because of its importance they weren’t ready to disclose that. This is despite the fact that deconvolution had been well published in my 1954 MIT thesis.

Here is our progress report, which Ralph Wiggins and I wrote on August 30, 1965. It says, "A FORTRAN program has been developed that has successfully picked the reflected events on a seismic section where the signal-to-noise ratio was high. The performance of the program as the signal-to-noise ratio deteriorates must still be evaluated. As to the migration and diffraction problems, the primary effort has been the development of basic input/output routines that are needed to interpret the stacked sections." As you see, all this involved a lot of data handling.

We set up a FORTRAN program and did a lot of data handling and sub-routine libraries. Approximately fifty routines from Geoscience's FORTRAN II program were converted to FORTRAN IV and checked out. This memo here notes that "Most of these have been transmitted to Mike Johnson at Shell Development Company. The translation of these will continue."

Then there is a section on software. Research on correlation enhancement by digital processing. "Research on better digital computer methods to enhance the correlation properties of seismic events includes progress from the following five items. Item one: The development of time-varying and space-varying filters. Item two: Work on the minimum phase property of the response of a horizontally stratified absorptive earth to acoustic waves. Item three: Development of hybrid analog/digital methods to speed up the computation of coefficients of least-squares digital filters, if possible by an order of magnitude over the Levinson single-channel methods and the Robinson method now in current use in geophysical data processing." In other words, there was an effort about that time to try to do optical processing, which is essentially analog, but which never worked out. But in time it might.

Goldstein:

What was the idea?

Robinson:

The idea was that digital was too expensive. Computers were expensive. An analog optical method that could do seismic processing physically in one step would be a very nice thing. The oil companies were always interested in building analog models of the earth. So the analog is a little closer to nature to them. Also, until the Cooley-Tukey FFT method came out in that very same year, 1965, analog was cheaper.

Goldstein:

Okay. So the optical method I guess is an analog method, but what is it?

Robinson:

The optical method would be essentially using lasers to do computing. I think that the Japanese have worked on this same thing. You can think of it this way. In seismic exploration you make use of the reflections of seismic waves. The optical method passes laser light through a transparency of a variable-density seismic record section. A lens forms a diffraction pattern of the transparency. This pattern is the Fourier transform of the section. In order to produce a filtered section, obstructions of suitable design are placed in the path of the light so as to remove events having certain frequencies or inclinations that cause interference with the desired reflection information. The modified light is passed through a second lens which converts it to the required filtered image. Optical filtering is useful not only for noise elimination, but also as a means of isolating various events bearing on the interpretation, but it could not do deconvolution. The laser had just become sort of operational around that time. Almost all the oil companies bought big laser machines, which were massive things back then. They thought, "Well maybe we can really do optical processing." Today, for example, optical telephone cables carry signals by internal reflections, so I can see their thinking. Why compute a big matrix and use the Levinson recursion if you could just do the processing in one step based on a nice analog method? It would have been nice if it had worked out. Now back to the August 30, 1965 report. Item four: "The development of suitable nonlinear iterative computational techniques to design digital filters which provide an optimum balance between feed-forward loops and feedback loops." And item five: "The setting up of an integrated course on digital filtering in the broad sense for field geophysicists and the preparation of exercises and study aids for this course." So Shell wanted to set up its own program. And it was helpful because then we were able to write a lot of papers which went into books and things.

Digicon

Robinson:

So, then Digicon was founded in 1965. Basically what happened was that there were six geophysicists at Texas Instruments, and at first three of them decided to found their own company. This is very common in geophysics, in that way a lot of small companies are founded. People leave big companies and found small companies. Three more people joined, so there were six. Digicon wanted to use deconvolution, but Texas Instruments said, “No you can’t. It's proprietary.” So that’s why they came to me and said, “Well, if you join Digicon we can use your deconvolution.”

Goldstein:

Was the hardware implementation of the mathematical theory or the analytic tool itself protected by Texas Instrument's patent?

Robinson:

TI, through its subsidiary GSI, was a geophysical service company. GSI would either collect the seismic records themselves, or take an oil company's records, process them, and supply the results. But they wouldn't tell the oil companies how they were doing it—they used their proprietary programs. When these six people left, they were not allowed to use Texas Instruments' actual software to do the deconvolution. So Digicon came to me, and I joined the company and gave them the software. All the deconvolution software used the same mathematics, because it all came from MIT to begin with, but Texas Instruments had modified it for their machines. Digicon got one of the SDS 9300 machines, which I think were built in California. The SDS company is not around anymore, but the 9300 was a good machine at the time, and they could put my software on it. Actually, all these MIT signal processing programs had been around for many years. O. M. Osman and I recently took all those original programs developed in the 1950s and republished them in a reprint book on Deconvolution, published by the Society of Exploration Geophysicists. I took all those old programs and converted them into Mathematica so students today can use them. Mathematica is a joy to use because it is so beautiful and gives beautiful pictures and graphics. Anyway, the Levinson Recursion and deconvolution were well known, but the fact that these six founders of Digicon had been Texas Instruments employees and had left the company could give an indication or appearance that they were taking along TI proprietary information. That's where I came in. In the end, it turned out deconvolution was quickly used by everybody, so no one had any proprietary claims to it.

Geophysical methods and patents

Robinson:

But other geophysical methods weren't public, like CDP and stacking; these were patented so users had to pay royalties on them.

Goldstein:

Who held the patent on that?

Robinson:

It was the Petty Geophysical Engineering Company that patented that. The company was named after a man named Petty. The CDP method was patented, and they had it. When something is patented, it can bring in a lot of money. Another geophysical method that was patented, and that made a lot of money for Conoco, was what they call vibroseis. This method uses a vibrator as the source of energy that goes into the ground. Actually, that method came right out of the first formal presentation I ever gave to the oil companies. It was at the MIT Industrial Liaison Conference on August 6, 1952. Bill Doty of Conoco was there. What Conoco did was to take a radar-type method and apply it to seismic. Conoco used a seismic swept frequency signal as a source and then correlated the source signal with the received seismic trace, and each event was collapsed into a nice sharp autocorrelation. So Conoco used a radar method, but they wrote the patent in such a way that you wouldn't know vibroseis had anything to do with radar, and so for years and years the other oil companies paid royalties to them, which amounted to a great deal of money. I knew a person who worked on the Conoco patent, one of the technical people, and he said, "Well, essentially, we patented correlation," and I said, "That's pretty good!"

But what saved the oil industry from other patents was the MIT GAG, because whenever anyone went to patent something, all the other companies remembered the MIT group and said, "Oh, this really came out of that group." As a result, many of these processes were not patented—they were opened up for everybody to develop without hindrance. I feel that really speeded things up, because patents can really slow things down.

Digicon; deconvolution and data collection

Goldstein:

What sort of things did Digicon have to do to stay competitive, what services did it begin to offer?

Robinson:

The first thing Digicon did was to offer deconvolution—that was the hot item in 1965, and there were only a few companies offering it then, even though most oil companies could do it in-house, like Shell was doing. Of the geophysical contracting companies, Digicon was one of the first. TI had been the very first to offer deconvolution a few years earlier. Then Seismic Computing Company got into it. By about 1968 everyone was offering deconvolution. It was those few years from 1965 to 1967 when deconvolution was still sort of a new thing.

Goldstein:

Started in 1965? I thought you were saying that some of the early deconvolution explorations were in '60?

Robinson:

You're right. Actually, let's see, from about 1964 to 1967 it was offered commercially by a few geophysical companies as a common thing. From about 1960 to 1964 it was done in-house by oil companies, and only offered commercially by TI. At that time there was a great deal of secrecy among oil companies, and they kept their processes secret. These major companies had the deconvolution process, and they were using it, but they would not divulge it. Then in the early 1960s, TI started advertising deconvolution and selling it. Then in 1965 Digicon started a company, and other companies started up, and they all started selling it. There were only a few years when there was a little oligopoly by those few companies that were selling it, then soon it was out and everybody had it. There were only a few years when it was “hot.” It is similar to a lot of things in the computer industry. Apple Computer had the mouse for a while, and now everybody has the mouse.

Goldstein:

After it become commonplace, say in 1967, what did Digicon start offering?

Robinson:

Digicon became a full service geophysical company and started collecting data. Some of their people were data collectors to begin with. Then there were other digital processing methods coming in, like migration and velocity analysis and the use of larger receiver arrays, so they started doing all sorts of things. But around 1970 there was a recession in engineering and science, and because Digicon had over-expanded in too many areas, it started losing money, especially in field work, which is very expensive. If you don't have all your crews hired out all the time, it could just kill you. Digicon almost went out of business, but it did survive. In the late 1970s, with the oil shortages, oil exploration boomed again, so Digicon bought this and that—tugboat companies, airplanes, everything—because business was booming.

Image processing

Robinson:

When the oil bust came in 1985 and the lean years returned, Digicon essentially went bankrupt and wiped out everybody who owned it. But the people were still there. So the creditors to whom Digicon actually owed the money took the people in the company and formed a new company, still called Digicon but with new ownership. So Digicon was like the Phoenix that rose up out of the ashes. Recently it merged with another company, and it’s now called Veritas DGC, where DGC stands for Digicon Geophysical Company. So it seems that even though the financial things might go bust, the people are still there and they go on. Digicon wasn’t the only one. There were a lot of small companies that went under. Then, in the late 1980s, a lot of new companies came in that were strong on modern computer graphics. They gave you a menu-driven operating system with beautiful displays. It’s like all computer programs sold today. If you don't have a program with beautiful displays, nobody will buy it.

Goldstein:

Do you think these things are cosmetic, or are they central to the analysis of the data?

Robinson:

Much of the signal processing today is geared towards image processing. People don't want to see just text, they want to see a picture. It might be cosmetic because you're essentially using the old signal processing algorithms but in much greater detail. But in some ways anything that gives you a better picture is not cosmetic. It's sort of like when the magnifying glass or the microscope or the telescope came in. You could see the moon before, but when Galileo saw it, it was a different moon. And that's sort of what's going on today. You could see pictures of the underground, but now you can see them almost like looking through a telescope. So it's a different ball game. So, in a sense, it's cosmetic if you're a mathematician. If you're an image processing person, then it's not.

Wavelets and geophysics

Robinson:

It is interesting to talk about wavelets, because the word wavelet goes back to Huygens in 1690. If you take an elementary physics book they talk about Huygens principle, which is: "Every point on a primary wave front serves as the source of spherical secondary wavelets. These secondary wavelets advance with a speed and frequency equal to that of the primary wave at each point in space. The primary wave front at some later time is the envelope of these secondary wavelets." But that word was never really used in physics or electrical engineering except in that sense. Then there is an old geophysicist, Norman Ricker, who is one of the great men of geophysics. He introduced the idea of a seismic wavelet. Ricker made use of a modified wave equation which was first given by George Gabriel Stokes in 1845. And essentially the wavelet that Ricker derived is the second derivative of a normal probability density. So it had a lot of nice properties, because a Gaussian distribution had so many beautiful properties. The seismic, or "Ricker" wavelet became the cornerstone of geophysics.

Goldstein:

When did that happen?

Robinson:

That happened about 1940. Ricker received his Ph.D. degree in physics from Rice University in 1920, and then worked for Western Electric. He joined Exxon in 1923 where he organized their first geophysical group in 1924. So when I wrote my thesis in 1954, I took that wavelet idea of his, but instead of an analytic wavelet, I had an empirical wavelet, which was simply a digital signal. I showed how to obtain the empirical wavelet from the seismic trace, and then how to remove the wavelet by deconvolution to give an empirical reflectivity series, which could be verified physically after the oil well was drilled. In this way I actually linked the empirical reflectivity series with the actual physical reflectivity in the ground. So, whereas his wavelet has a definite mathematical shape, my wavelet is empirical and is the basis of the convolutional model used in most seismic processing and interpretation schemes in use today. So among geophysicists, you can think of Ricker as a giant. I was like a student of Ricker's, and when I wrote my thesis I made a big deal out of the wavelets. The empirical wavelet is physically stable and minimum phase. No one believed that it was possible to pull out the individual wavelet, and that was the breakthrough. So I actually could compute a wavelet from the trace. So in a way Ricker was the one who started the wavelet, and I took it from there. Then the story switches to Pierre Goupillaud, a geophysicist who worked at Conoco, and I'm an old friend of his. Actually when I was in Digicon I taught a course for Conoco because in ways Conoco had gotten the idea of vibroseis from my MIT work, so they were grateful. When we founded Digicon, they made a contract with us, and part of the contract was to teach them about all the latest digital things. So when I was teaching that course in 1966, one student was Pierre L. Goupillaud, who was working with Conoco. He was French and a very interesting person. I just saw him a couple of months ago, and he's about seventy-eight years old now. Anyhow, in the course I got into wavelets, and got him into discussions about the wavelet idea. And then, Jean P. Morlet, who is a French geophysicist and a mutual friend, was dealing with Gabor's transform. Gabor covered the time-frequency plane with uniform cells, and associated each cell with a wavelet of invariant envelope with a carrier of variable frequency. Morlet perceived that it was the wavelet shape that must be invariant to give uniform resolution in the entire plane. To do this, Morlet adapted the sampling rate to the frequency, thereby creating a changing time scale producing a stretching of the wavelet. He called this technique the cycle-octave transform. Pierre Goupillaud, as editor of Geophysics, published Morlet's contribution in 1982. While a student at the Ecole Polytechnique, Morlet had a classmate named Roger Balian, who was a theoretical physicist. In 1982 Balian introduced Morlet to Alexandre F. Grossman, a professor of mathematical physics at Marseilles. Morlet visited Grossman in Marseilles, and Grossman developed the rigorous basis of the cycle-octave transform. In 1983, Goupillaud, Morlet, and Grossman visited me at the University of Tulsa and we had a nice meeting. Grossman and Morlet used the expression "wavelets of constant shape" in their 1984 SIAM paper so the cycle-octave transform became universally known as the wavelet transform. Even though the word wavelet had been pre-empted in geophysics, the word caught on in mathematics and is now firmly entrenched. So Goupillaud, who is a geophysicist, and Grossman, who is a theoretical physicist, gave Morlet the support needed to bring the wavelet transform to the attention of people outside of geophysics. The word wavelet comes from geophysics, and that is the origin of the name wavelet transform, which the mathematicians picked up as part of their domain.

Goldstein:

When was that?

Robinson:

This was in the mid-1980s. The mathematicians picked it up very readily because it was right down their alley. And so, in fact, this is like a little history of it. So it's kind of interesting because, in the name wavelet transform, the word wavelet is very descriptive. In fact, since it was done by a geophysicist, it's called the wavelet transform; if it had been done by a mathematician, it would have some other name, and then it would have been the such-and-such transform. It's just the word. If you have a good name for something, use it.

Goldstein:

Exactly. It has already been captured in the term there.

Robinson:

Yes. Because a name is a way of encompassing a concept, and the word wavelet was available from geophysics and it was powerful. But if somebody had used a different word, the wavelet transform would have a different name. Morlet originally called it the cycle-octave transform. It could have stayed that. It just happened that the name wavelet transform caught on. If Ricker hadn't introduced the word wavelet into geophysics, it most certainly would have been called something else.

Overview of signal processing in geophysics

Goldstein:

Okay, I think we've covered a lot of things. Is there something major in the development of signal processing and geophysics or is there a connection with other signal processing techniques that we've overlooked?

Robinson:

In summary you might say that there's never really a beginning to anything. If you go back to the 1930s and 1940s, the people at Bell Labs, MIT, and places like that, even , and before that to geophysicists like Richter and to mathematicians like Wiener, Shannon, Bode, and Tukey, you see that they all had important ideas. But these ideas could not be fully implemented until the computer matured enough to handle so much data. I think that Shannon wrote an article way back in the 1940s on whether a computer could play chess, and he believed that it could. Today it's considered nothing for a computer to play chess. And Wiener had all these ideas and would try to implement them, but things were so primitive then. Now they can be implemented. If only those pioneers could come back and see how their ideas have become all these wonderful things. It was a happy circumstance that so many good people were at these places at that time. The electrical engineering department at MIT is quite an exceptional department. They had Professor E. A. Guillemin who wrote a lot of their electrical engineering textbooks and laid the foundation, and they continue that tradition. So, from my point of view, it was good fortune to be at MIT in those years.

Radiation Lab

Goldstein:

Let me ask about that. Was the influence of the Radiation Lab important in the electrical engineering environment?

Robinson:

Yes. The Rad Lab functioned during the war, and I entered MIT in 1946, which was really the first post-war class—students who entered in 1945 were still considered a war-time class. There was a place called Building Twenty—I think it's still there—a wooden structure built for the Rad Lab. There was another building, too, which they were converting to house the Lincoln Laboratory. They rebuilt this old wooden building, and spent a lot of money on it. The Lincoln Lab went in there for a while, but then it moved to the Bedford-Lexington Line, and that’s where it is now. The radiation lab was converted from the Lincoln laboratory. All this activity had a tremendous influence on the students at the time. When we formed the Geophysical Analysis Group in 1953, we had our offices in Building Twenty, the old Rad Lab building. People wanted to rip that building down, but then they said no, it was sort of an antique at MIT. Last time I was there, Building Twenty was still there.

Geophysical Analysis Group

Goldstein:

How did the Geophysical Analysis Group get organized?

Robinson:

In 1950 we started work on the Mobil Oil Company records. It was essentially myself, but with the help of professors Wiener, Wadsworth, and Levinson. I was a research assistant in the mathematics department trying to apply their ideas. So I digitized those records, and worked out the theory of deconvolution successfully on the seismic records. Mobil Oil, which had given me the records, was then asked to support the project. Mobil Oil said it was already giving MIT $50,000 a year, and that was in 1951 or something, so that would be like $500,000 or more in today’s dollars. But the money was all being used by the Chemical Engineering Department—the math department wasn't getting any of this. So, the people at Mobil said, “Give the mathematics department some of that money!” This upset the chemical engineers because some of it would have to come from their share. So to please the chemical engineers, MIT said “We could do it for a year or two, but then the mathematics department will have to get oil companies to give new money.” So, the project went on in the math department for about two years.

And then, on August 6, 1952, the results of everything were presented at the MIT Industrial Liaison Conference that I mentioned earlier. As a result of that, fourteen companies gave new money. So in 1953, the project was moved into the geophysics department, actually the Department of Geology and Geophysics, and was called the MIT Geophysical Analysis Group—GAG—which I have told you about. I moved with it and became the head of GAG. I became a geophysicist, which was probably a mistake—I should have stayed in mathematics. With that large amount of money, GAG was able to hire many graduate students, and upon receiving their degrees, the graduate students were able to go out and populate the oil companies. Cecil Green was an MIT graduate in electrical engineering. He arranged with Robert Shrock, who was head of the Geology Department and became a very good friend of Cecil Green, an interchange of students in which MIT students would go work with Texas Instruments in the summer. As a result, many students ended up working for TI.

Mobil Oil

Goldstein:

How did you get connected with Mobil in the first place?

Robinson:

Mainly through Professor Hurley, who happened to know people at Mobil Oil.

Goldstein:

So you were looking for support for your studies?

Robinson:

No. Actually, a research assistantship had been offered to me in mathematics when I had applied to graduate school in the spring of 1950. I graduated in June 1950, and then I was called to active duty in the army and spent that summer at Aberdeen Proving Grounds, where I learned to program the ENIAC at nights. Upon graduating from the Army Ordnance School in September 1950, I was assigned to an early ready reserve unit at Watertown Arsenal in Massachusetts. By this good fortune I was able to start graduate school at MIT in October of 1950. I had a research assistantship in the mathematics department, and I went to work applying Wiener’s methods. Wadsworth said, “Well, I have some seismic records that I can get from Professor Hurley, and you can work on them.” We had no working connection with Mobil—we just had the records. But we started going, Hurley asked for more records and then the connection grew. It's interesting because in 1952, I was supposed to go down to Dallas. I still have the telegram from Cecil Green saying, “I’ve made hotel reservations for you.” Here's Cecil Green, a multi-billionaire today, saying he was making hotel reservations for me. Cecil Green was always very careful with money. I think that's why he has so much. He invested every nickel and dime wisely to make TI into a great company. It pays off.

Whirlwind

Robinson:

So, in summary, you might say that there were a lot of fortuitous things. One was that MIT had one of the first digital computers, the Whirlwind, which became the prototype for the North American Defense system. It was the machine for which the magnetic core memory was invented. Whirlwind had a very small register length of about 16 bits. Whirlwind was good for signal processing because it could handle a lot of data, but the short register length made it difficult for the inversion of matrices. But for the signal processing, for deconvolution, it was very fast and efficient. So in some ways Whirlwind was equipped for signal processing; it could handle a lot of data that the other machines of the early 1950s could not.

Goldstein:

What makes the short register length good for signal processing?

Robinson:

In those days every computer had a very small memory, which meant you could only have so many 32 bit words, or twice that number of 16 bit words. In signal processing you have long data series, and 16 bits is more than enough to hold each amplitude of the data series. So the computer can hold longer signals using 16 bit registers than 32 bit registers under the constraint that the total number of bits is fixed. But it did make the inversions more difficult—you had to use double precision for that, but it could be done. So MIT had Whirlwind, and it had Wiener. When Wiener's cybernetics book came out in 1949, he became a celebrity. He was on television. In 1953. I turned on the television and there he was, on the morning show, Norbert Wiener, and he was in the math department. Just having him there was an inspiration, and you could talk to him, or have lunch with him. The nice thing was that I was able to work with him and get his help. When he went to Mexico—he used to spend one semester in Mexico every year—the department used to put me in his office, because they had a shortage of offices, and then I could read all his books and articles. So, in having him, and in having the enthusiasm of science in those days, it was the golden age of science. And then the electrical engineering department had Y. W. Lee, who was very helpful and good. I don't think any of Oppenheim's people where there at that time, 1950 to 1954. I think they were all younger than that.

Goldstein:

Thomas Stockham's is another name I keep hearing. Did you know him at MIT?

Robinson:

Just a little bit. Tom Kailath left MIT and went to Stanford. I've kept up with him.

Goldstein:

I’m actually supposed to meet him next week.

Robinson:

He is an excellent person. In fact, I went to his sixtieth birthday party, so when you see him, give him my regards.

Goldstein:

I will, and thank you for your time.