We are saddened to share that Cleve Moler passed away on May 20, 2026, at the age of 86 at his home surrounded by his family. Cleve was chief mathematician and cofounder of MathWorks and the author of the first version of MATLAB.
In his early years, he was a professor of math and computer science for almost 20 years at the University of Michigan, Stanford University, and the University of New Mexico. During this time he was known for being one of the authors of LINPACK and EISPACK, two foundational Fortran libraries for numerical computing. One popular paper of his is “Nineteen Dubious Ways to Compute the Exponential of a Matrix.”
Cleve made extraordinary contributions to the field of numerical computing. His work had a profound impact on how mathematical algorithms are developed, analyzed, and applied across science and engineering. From his early contributions to matrix computations and numerical linear algebra to the creation of MATLAB, Cleve helped make advanced computational methods more accessible, reliable, and widely used.
Moler’s creation, Matlab, revolutionized the teaching of numerical methods to two generations of college students. Moler had major roles in the creation of BLAS and Lapack, as well. No matter which programming language we use, we all rely on BLAS and Lapack.
Cleve had been interviewed a couple of years ago for the inControl podcast. If you haven’t listened to it (and if you’re interested in control theory in general), I strongly invite you to do so !
I only met him once at a conference, probably around 2010 or so, we had a very brief conversation during a discussion at a talk, I think he was asking how many integrals SymPy can do from some test suite (I forgot which one), but I remember it could do about half back then and he said that’s very good. He was very nice.
I think he created MATLAB while he was at UNM here in Albuquerque, NM in the 1970s. We have a thread about how to compile the original version (in Fortran!) here: Compiling the original Matlab.
I first met Cleve Moler when I was a grad student at a workshop in 1978. I was a chemist, but this workshop was about using algebraic methods in chemistry. He was one of the co-organizers of the meeting, which was at Santa Cruz California. The speakers at the workshop were a mixture of chemists and numerical methods experts. Even now, some 48 years later, I can remember talks about eigenvalue methods, linear equations, Cholesky decomposition, tensor transformations, and the difference between Gram-Schmidt and modified Gram-Schmidt orthonormalization. I met him again a year later at another joint numerical methods and chemistry workshop at Salt Lake City where he helped me modify a LINPACK routine in order to solve singular linear equation systems. I later became a chemistry staff member at Argonne National Laboratory, and Cleve Moler was a frequent speaker there as he visited Jack Dongarra, Danny Sorensen, Jorge More, and others in the applied math division there. I was never a user of MATLAB, but that and his other contributions to numerical methods impacted my work and that of many others over that period of time.
edit: I found this workshop report for that 1978 meeting (edited by Cleve Moler) https://www.osti.gov/servlets/purl/6169634, and this report for the 1979 meeting (Cleve Moler was a participant). https://www.osti.gov/servlets/purl/5491306. It is odd that I can’t remember to stop for milk on the way home, but I can remember some of the talks from these meetings from almost 50 years ago.
I never had the pleasure of meeting Moler, but I came across a paper of his (Moler C. and Morrison D, “Replacing square roots by Pythagorean sums.” Dept. of Computer Science, Univ. of New Mexico, Albuquerque, NM, 1981) while I was writing the book "FORTRAN Optimization" (Academic Press, 1982). In that paper, the authors describe an iterative algorithm that is highly stable for computing the square root of a value that is the sum of two squares (it is reproduced on p. 50 of the book), something that was of some importance back in the days of sometimes limited range and precision.
I think that algorithm is the basis for adding the two fortran intrinsics hypot() and norm2() to the language.
I think the “limited range” you are talking about was the VAX D float format. It had only an 8-bit exponent field (and a longer 55-bit fraction), so the square of the machine epsilon was close to the underflow limit, which did make things like hypot() and norm2() tricky when using the straightforward expressions. I remember using the VAX cabs() function (which also used that algorithm I think) sometimes to compute hypot() for that reason.