C++ MEX plugins to compute error functions (erf, erfc, erfi, erfcx, Faddeeva, ...) of complex args
C++ source code for compiled plugins (MEX files) to compute various error functions for complex arguments:
** Faddeeva_erf(z)  the error function
** Faddeeva_erfc(z) = 1  erf(z)  complementary error function
** Faddeeva_erfi(z) = i erf(iz)  imaginary error function
** Faddeeva_erfcx(z) = exp(z^2) erfc(z)  scaled complementary error function
** Faddeeva_w(z) = exp(z^2) erfc(iz)  Faddeeva function
** Faddeeva_Dawson(z) = 0.5 sqrt(pi) exp(z^2) erfi(z)  Dawson function
From e.g. the Faddeeva function, one can also obtain the Voigt functions and other related functions.
Assuming you have a C++ compiler (and have configured it in MATLAB with mex setup), compile by running the included Faddeeva_build.m script in MATLAB:
Faddeeva_build
All of the functions have usage of the form:
w = Faddeeva_w(z)
or optionally Faddeeva_w(z, relerr), where relerr is a desired relative error (default: machine precision). z may be an array or matrix of complex or real numbers.
This code may also be downloaded from
http://abinitio.mit.edu/Faddeeva
along with documentation and other versions. As described in the source code, this implementation uses a combination of algorithms for the Faddeeva function: a continuedfraction expansion for large z [similar to G. P. M. Poppe and C. M. J. Wijers, "More efficient computation of the complex error function," ACM Trans. Math. Soft. 16 (1), pp. 38–46 (1990)], and a completely different algorithm for smaller z [Mofreh R. Zaghloul and Ahmed N. Ali, "Algorithm 916: Computing the Faddeyeva and Voigt Functions," ACM Trans. Math. Soft. 38 (2), 15 (2011).]. Given the Faddeeva function, we can then compute the other error functions, although we must switch to Taylor expansions and use other tricks in certain regions of the complex plane to avoid cancellation errors or other floatingpoint problems.
1.5  portability fixes, slight accuracy improvements 

1.4  Now includes separate plugins for all of the error functions. 

1.3  note how to compute erfi using Faddeeva function 

1.1  Improve accuracy in Re[w(z)] taken by itself. 
Janos (view profile)
This is a fantastic implementation. This code works about 2000x faster for me (when tested with large multidimensional arrays) than the builtin Matlab erfi function.
Karan Gill (view profile)
Alternatively, the Symbolic Math Toolbox provides the error and dawson functions for complex inputs.
Take the error function for example:
>> double(erf(sym(1+1i)))
ans =
1.3162 + 0.1905i
You could define an anonymous function to make it easier:
>> erfCmplx = @(x) double(erf(sym(x)))
erfCmplx =
@(x)double(erf(sym(x)))
>> erfCmplx(1+1i)
ans =
1.3162 + 0.1905i
The Symbolic Math Toolbox functions are:
erf: http://www.mathworks.com/help/symbolic/erf.html
erfc: http://www.mathworks.com/help/symbolic/erfc.html
erfcinv: http://www.mathworks.com/help/symbolic/erfcinv.html
erfi: http://www.mathworks.com/help/symbolic/erfi.html
erfinv: http://www.mathworks.com/help/symbolic/erfinv.html
dawson: http://www.mathworks.com/help/symbolic/dawson.html
Brian Hannan (view profile)
Ian (view profile)
One update I would really like to see in this package is the derivative functions, at least for W(z). As pointed out in Zaghloul and Ali, the derivative functions of W(z) (equations 2123) become numerically unstable near dV/dx = 0 (V=real(W(z)), at the peak of the Voigt function. This can cause problems when trying to compute analytical Jacobians for doing nonlinear fits of the Voigt function to optical spectra.
Since you are already using the Zaghloul and Ali algorithm in this region, it would be helpful to also use their method to output a function, say Faddeeva_dw(z) = dV/dx + i*dL/dx (L=imag(W(z)). The corresponding y derivatives can then be trivially computed.
Ian (view profile)
Works as advertised and is extremely fast.
Niklas (view profile)
Steven G. Johnson (view profile)
It won't work with lcc, since that is a C compiler, not a C++ compiler. The failure to compile with Visual C++ is a bug, which is fixed in the latest release on my web site (and which should appear on Matlab Central shortly).
alan (view profile)
Hello, Love this program. I have it up and running on my mac to simulate voigt broadening. Im trying to add this functionality to a lab computer running win32 matlab 7.12.0 (R2011a), but I cannot successfully compile with mex.
With Lccwin32 C 2.4.1 in C:\PROGRA~1\MATLAB\R2011a\sys\lcc: lcc preprocessor error: .\Faddeeva.hh:30 .\Faddeeva_mex.cc:35 Faddeeva_w_mex.cc:3 Could not find include file complex
full verbose: http://pastebin.com/YyLYYe8C
With Microsoft Visual C++ 2010 Express in C:\Program Files\Microsoft Visual Studio 10.0: Faddeeva.cc Faddeeva.cc(184) : error C2124: divide or mod by zero Faddeeva.cc(822) : error C3861: 'copysign': identifier not found C:\PROGRA~1\MATLAB\R2011A\BIN\MEX.PL: Error: Compile of 'Faddeeva.cc' failed.
full verbose: http://pastebin.com/K4LPd9FN
The only instances of complex.h are in a pythonwx folder not related to matlab. I see that it was released for R2012a. Does the package need a specific C++ compiler?