Skip to content


Subversion checkout URL

You can clone with HTTPS or Subversion.

Download ZIP
automatic differentiation library for common lisp
Common Lisp
branch: master

Fetching latest commit…

Cannot retrieve the latest commit at this time

Failed to load latest commit information.


Written by Mason Smith

  1. What is cl-autodiff?
  2. Installation
  3. Usage
  4. Limitations
  5. TODO

1. What is cl-autodiff?

cl-autodiff is a library for automatic differentiation of mathematical functions. Automatic Differentiation is a technique whereby derivatives of function are computed to machine precision, without the need for programming the derivative explicitly and without resorting to numerical approximation.

cl-autodiff, specifically, is an implentation of forward accumulation automatic differentiation. This has some performance implications, noted in Limitations.

It is licensed using the LLGPL, which should be included with a copy of this distribution. If you find a use for cl-autodiff, be sure to let me know.

2. Installation

cl-autodiff is installable via ASDF. Simply make a symbolic link to cl-autodiff.asd in your ASDF directory, and use (asdf:oos 'asdf:load-op :cl-autodiff) to compile and load the program. Alternatively, if you use clbuild, you can add cl-autodiff to your wnpp-projects file. The repo is located at:

3. Usage

Simply put, use DEFINE-WITH-DERIVATIVES as you would DEFUN to define mathematical functions.

For instance, p(x) = 3*x^2 - 2*x + 1.

CL-USER> (define-with-derivatives p (x) (+ (* 3 (expt x 2)) (* -2 x) 1))
>> P

When you evaluate P at point now, P returns both the value and the derivative at that point.

CL-USER> (p 3)
>> 22

Note that the second value is a vector. DEFINE-WITH-DERIVATIVES computes the gradient, so your function can have any number of inputs. For instance, take q(x, y, z) = cos(xy) - ysin(z)

CL-USER> (define-with-derivatives q (x y z) (- (cos (* x y)) (* y (sin z))))
>> Q

CL-USER> (q pi 1 (/ pi 2))
>> -2
 #(0 -1 0)

DEFINE-WITH-DERIVATIVES works with virtually all analytic* Common Lisp math functions. It also works seamlessly with complex numbers.

CL-USER> (define-with-derivatives r (z) (* z (sqrt z)))
>> R

CL-USER> (r #C(0.0 1.0))
>> #C(-0.70710677 0.70710677)
#(#C(1.0606601 1.0606601))

To a reasonable extent, it understands let, labels, etc. as well. One can even define recursive functions, assuming the underlying function is differentiable.

CL-USER> (define-with-derivatives recur-cube (x)
           (labels ((f (x n) 
           (if (> n 0) (* x (f x (1- n))) 1)))
             (f x 3)))

CL-USER> (recur-cube 2)

Note that you can also use your user-defined functions in other functions.

CL-USER> (define-with-derivatives sixth-power (x) 
       (let ((y (recur-cube x)))
         (* y y)))



LAMBDA-AD also does what you would expect.

CL-USER> (funcall (lambda-ad (x y) (/ (expt x 3) (expt y 3))) 2 2)
>> 1
#(1.5 -1.5)

4. Limitations

cl-autodiff is still in, at best, a pre-alpha stage. As far as I have tested it (which has been almost not at all), it seems to work. However, there are a number of known limitations (and some unknown ones, I'm sure).

  • The only way of defining functions with automatic derivatives is with the DEFINE-WITH-DERIVATIVES macro. As of yet, there is no equivalent of LABELS or LAMBDA for it.

  • LAMBDA-AD doesn't have the syntactic leeway that LAMBDA does. In particular, neither ((lambda-ad (x) x) 1) nor (funcall #'(lambda (x) x) 1) will work.

  • `DEFINE-WITH-DERIVATIVES` does not handle some special forms very well. Theoretically it can handle do loops, for instance, but I haven't tested it. The problem is that `DEFINE-WITH-DERIVATIVES` does nothing to a form it doesn't recognize. An alternate (and probably more effective) approach would be to assume that, for any unrecognized form, the subforms should be parsed. I'll probably change this in a few versions.
  • DEFINE-WITH-DERIVATIVES should be able to handle all macros, special forms, etc., but the functionality has not been tested at all. In fact, testing overall is rather limited.

  • DEFINE-WITH-DERIVATIVES can handle lambdas, but it CANNOT handle #'. So, for instance, (funcall (lambda (x) (+ x 3)) y) should differentiate fine (w/rt y, of course), but (funcall #'+ y 3) won't work at all.

  • Partial derivatives are implemented by passing them around throughout the function evaluation. A better approach would be to use reverse accumulation, rather than forward accumulation.

In general, if you stick with the built in mathematical operators, labels, flets, let*, lets, and lambdas, you should be fine. As far as I can tell, if you can compile/evaluate a form and get an answer at all, it will probably be correct.


  • [labels/flet]-with-derivatives

  • Thorough unit testing

  • Experiment with an operator overloading version

    • would require non-standard functions (ad-+, ad-sin, etc.)
Something went wrong with that request. Please try again.