1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054
|
% Copyright Matthew Pulver 2018 - 2019.
% Distributed under the Boost Software License, Version 1.0.
% (See accompanying file LICENSE_1_0.txt or copy at
% https://www.boost.org/LICENSE_1_0.txt)
\documentclass{article}
\usepackage{amsmath} %\usepackage{mathtools}
\usepackage{amssymb} %\mathbb
\usepackage{array} % m{} column in tabular
\usepackage{csquotes} % displayquote
\usepackage{fancyhdr}
\usepackage{fancyvrb}
\usepackage[margin=0.75in]{geometry}
\usepackage{graphicx}
\usepackage{hyperref}
%\usepackage{listings}
\usepackage{multirow}
\usepackage[super]{nth}
\usepackage{wrapfig}
\usepackage{xcolor}
\hypersetup{%
colorlinks=false,% hyperlinks will be black
linkbordercolor=blue,% hyperlink borders will be red
urlbordercolor=blue,%
pdfborderstyle={/S/U/W 1}% border style will be underline of width 1pt
}
\pagestyle{fancyplain}
\fancyhf{}
\renewcommand{\headrulewidth}{0pt}
\cfoot[]{\thepage\\
\scriptsize\color{gray} Copyright \textcopyright\/ Matthew Pulver 2018--2019.
Distributed under the Boost Software License, Version 1.0.\\
(See accompanying file LICENSE\_1\_0.txt or copy at
\url{https://www.boost.org/LICENSE\_1\_0.txt})}
\DeclareMathOperator{\sinc}{sinc}
\begin{document}
\title{Autodiff\\
\large Automatic Differentiation C++ Library}
\author{Matthew Pulver}
\maketitle
%\date{}
%\begin{abstract}
%\end{abstract}
\tableofcontents
%\section{Synopsis}
%\begingroup
%\fontsize{10pt}{10pt}\selectfont
%\begin{verbatim}
% example/synopsis.cpp
%\end{verbatim}
%\endgroup
\newpage
\section{Description}
Autodiff is a header-only C++ library that facilitates the
\href{https://en.wikipedia.org/wiki/Automatic_differentiation}{automatic differentiation} (forward mode) of
mathematical functions of single and multiple variables.
This implementation is based upon the \href{https://en.wikipedia.org/wiki/Taylor_series}{Taylor series} expansion of
an analytic function $f$ at the point $x_0$:
\begin{align*}
f(x_0+\varepsilon) &= f(x_0) + f'(x_0)\varepsilon + \frac{f''(x_0)}{2!}\varepsilon^2 + \frac{f'''(x_0)}{3!}\varepsilon^3 + \cdots \\
&= \sum_{n=0}^N\frac{f^{(n)}(x_0)}{n!}\varepsilon^n + O\left(\varepsilon^{N+1}\right).
\end{align*}
The essential idea of autodiff is the substitution of numbers with polynomials in the evaluation of $f(x_0)$. By
substituting the number $x_0$ with the first-order polynomial $x_0+\varepsilon$, and using the same algorithm
to compute $f(x_0+\varepsilon)$, the resulting polynomial in $\varepsilon$ contains the function's derivatives
$f'(x_0)$, $f''(x_0)$, $f'''(x_0)$, ... within the coefficients. Each coefficient is equal to the derivative of
its respective order, divided by the factorial of the order.
In greater detail, assume one is interested in calculating the first $N$ derivatives of $f$ at $x_0$. Without loss
of precision to the calculation of the derivatives, all terms $O\left(\varepsilon^{N+1}\right)$ that include powers
of $\varepsilon$ greater than $N$ can be discarded. (This is due to the fact that each term in a polynomial depends
only upon equal and lower-order terms under arithmetic operations.) Under these truncation rules, $f$ provides a
polynomial-to-polynomial transformation:
\[
f \qquad : \qquad x_0+\varepsilon \qquad \mapsto \qquad
\sum_{n=0}^Ny_n\varepsilon^n=\sum_{n=0}^N\frac{f^{(n)}(x_0)}{n!}\varepsilon^n.
\]
C++'s ability to overload operators and functions allows for the creation of a class {\tt fvar}
(\underline{f}orward-mode autodiff \underline{var}iable) that represents polynomials in $\varepsilon$. Thus
the same algorithm $f$ that calculates the numeric value of $y_0=f(x_0)$, when
written to accept and return variables of a generic (template) type, is also used to calculate the polynomial
$\sum_{n=0}^Ny_n\varepsilon^n=f(x_0+\varepsilon)$. The derivatives $f^{(n)}(x_0)$ are then found from the
product of the respective factorial $n!$ and coefficient $y_n$:
\[ \frac{d^nf}{dx^n}(x_0)=n!y_n. \]
\section{Examples}
\subsection{Example 1: Single-variable derivatives}
\subsubsection{Calculate derivatives of $f(x)=x^4$ at $x=2$.}
In this example, {\tt make\_fvar<double, Order>(2.0)} instantiates the polynomial $2+\varepsilon$. The {\tt Order=5}
means that enough space is allocated (on the stack) to hold a polynomial of up to degree 5 during the proceeding
computation.
Internally, this is modeled by a {\tt std::array<double,6>} whose elements {\tt \{2, 1, 0, 0, 0, 0\}} correspond
to the 6 coefficients of the polynomial upon initialization. Its fourth power, at the end of the computation, is
a polynomial with coefficients {\tt y = \{16, 32, 24, 8, 1, 0\}}. The derivatives are obtained using the formula
$f^{(n)}(2)=n!*{\tt y[n]}$.
\begin{verbatim}
#include <boost/math/differentiation/autodiff.hpp>
#include <iostream>
template <typename T>
T fourth_power(T const& x) {
T x4 = x * x; // retval in operator*() uses x4's memory via NRVO.
x4 *= x4; // No copies of x4 are made within operator*=() even when squaring.
return x4; // x4 uses y's memory in main() via NRVO.
}
int main() {
using namespace boost::math::differentiation;
constexpr unsigned Order = 5; // Highest order derivative to be calculated.
auto const x = make_fvar<double, Order>(2.0); // Find derivatives at x=2.
auto const y = fourth_power(x);
for (unsigned i = 0; i <= Order; ++i)
std::cout << "y.derivative(" << i << ") = " << y.derivative(i) << std::endl;
return 0;
}
/*
Output:
y.derivative(0) = 16
y.derivative(1) = 32
y.derivative(2) = 48
y.derivative(3) = 48
y.derivative(4) = 24
y.derivative(5) = 0
*/
\end{verbatim}
The above calculates
\begin{alignat*}{3}
{\tt y.derivative(0)} &=& f(2) =&& \left.x^4\right|_{x=2} &= 16\\
{\tt y.derivative(1)} &=& f'(2) =&& \left.4\cdot x^3\right|_{x=2} &= 32\\
{\tt y.derivative(2)} &=& f''(2) =&& \left.4\cdot 3\cdot x^2\right|_{x=2} &= 48\\
{\tt y.derivative(3)} &=& f'''(2) =&& \left.4\cdot 3\cdot2\cdot x\right|_{x=2} &= 48\\
{\tt y.derivative(4)} &=& f^{(4)}(2) =&& 4\cdot 3\cdot2\cdot1 &= 24\\
{\tt y.derivative(5)} &=& f^{(5)}(2) =&& 0 &
\end{alignat*}
\subsection{Example 2: Multi-variable mixed partial derivatives with multi-precision data type}\label{multivar}
\subsubsection{Calculate $\frac{\partial^{12}f}{\partial w^{3}\partial x^{2}\partial y^{4}\partial z^{3}}(11,12,13,14)$
with a precision of about 50 decimal digits,\\
where $f(w,x,y,z)=\exp\left(w\sin\left(\frac{x\log(y)}{z}\right)+\sqrt{\frac{wz}{xy}}\right)+\frac{w^2}{\tan(z)}$.}
In this example, {\tt make\_ftuple<float50, Nw, Nx, Ny, Nz>(11, 12, 13, 14)} returns a {\tt std::tuple} of 4
independent {\tt fvar} variables, with values of 11, 12, 13, and 14, for which the maximum order derivative to
be calculated for each are 3, 2, 4, 3, respectively. The order of the variables is important, as it is the same
order used when calling {\tt v.derivative(Nw, Nx, Ny, Nz)} in the example below.
\begin{verbatim}
#include <boost/math/differentiation/autodiff.hpp>
#include <boost/multiprecision/cpp_bin_float.hpp>
#include <iostream>
using namespace boost::math::differentiation;
template <typename W, typename X, typename Y, typename Z>
promote<W, X, Y, Z> f(const W& w, const X& x, const Y& y, const Z& z) {
using namespace std;
return exp(w * sin(x * log(y) / z) + sqrt(w * z / (x * y))) + w * w / tan(z);
}
int main() {
using float50 = boost::multiprecision::cpp_bin_float_50;
constexpr unsigned Nw = 3; // Max order of derivative to calculate for w
constexpr unsigned Nx = 2; // Max order of derivative to calculate for x
constexpr unsigned Ny = 4; // Max order of derivative to calculate for y
constexpr unsigned Nz = 3; // Max order of derivative to calculate for z
// Declare 4 independent variables together into a std::tuple.
auto const variables = make_ftuple<float50, Nw, Nx, Ny, Nz>(11, 12, 13, 14);
auto const& w = std::get<0>(variables); // Up to Nw derivatives at w=11
auto const& x = std::get<1>(variables); // Up to Nx derivatives at x=12
auto const& y = std::get<2>(variables); // Up to Ny derivatives at y=13
auto const& z = std::get<3>(variables); // Up to Nz derivatives at z=14
auto const v = f(w, x, y, z);
// Calculated from Mathematica symbolic differentiation.
float50 const answer("1976.319600747797717779881875290418720908121189218755");
std::cout << std::setprecision(std::numeric_limits<float50>::digits10)
<< "mathematica : " << answer << '\n'
<< "autodiff : " << v.derivative(Nw, Nx, Ny, Nz) << '\n'
<< std::setprecision(3)
<< "relative error: " << (v.derivative(Nw, Nx, Ny, Nz) / answer - 1) << '\n';
return 0;
}
/*
Output:
mathematica : 1976.3196007477977177798818752904187209081211892188
autodiff : 1976.3196007477977177798818752904187209081211892188
relative error: 2.67e-50
*/
\end{verbatim}
\subsection{Example 3: Black-Scholes Option Pricing with Greeks Automatically Calculated}
\subsubsection{Calculate greeks directly from the Black-Scholes pricing function.}
Below is the standard Black-Scholes pricing function written as a function template, where the price, volatility
(sigma), time to expiration (tau) and interest rate are template parameters. This means that any Greek based on
these 4 variables can be calculated using autodiff. The below example calculates delta and gamma where the variable
of differentiation is only the price. For examples of more exotic greeks, see {\tt example/black\_scholes.cpp}.
\begin{verbatim}
#include <boost/math/differentiation/autodiff.hpp>
#include <iostream>
using namespace boost::math::constants;
using namespace boost::math::differentiation;
// Equations and function/variable names are from
// https://en.wikipedia.org/wiki/Greeks_(finance)#Formulas_for_European_option_Greeks
// Standard normal cumulative distribution function
template <typename X>
X Phi(X const& x) {
return 0.5 * erfc(-one_div_root_two<X>() * x);
}
enum class CP { call, put };
// Assume zero annual dividend yield (q=0).
template <typename Price, typename Sigma, typename Tau, typename Rate>
promote<Price, Sigma, Tau, Rate> black_scholes_option_price(CP cp,
double K,
Price const& S,
Sigma const& sigma,
Tau const& tau,
Rate const& r) {
using namespace std;
auto const d1 = (log(S / K) + (r + sigma * sigma / 2) * tau) / (sigma * sqrt(tau));
auto const d2 = (log(S / K) + (r - sigma * sigma / 2) * tau) / (sigma * sqrt(tau));
switch (cp) {
case CP::call:
return S * Phi(d1) - exp(-r * tau) * K * Phi(d2);
case CP::put:
return exp(-r * tau) * K * Phi(-d2) - S * Phi(-d1);
}
}
int main() {
double const K = 100.0; // Strike price.
auto const S = make_fvar<double, 2>(105); // Stock price.
double const sigma = 5; // Volatility.
double const tau = 30.0 / 365; // Time to expiration in years. (30 days).
double const r = 1.25 / 100; // Interest rate.
auto const call_price = black_scholes_option_price(CP::call, K, S, sigma, tau, r);
auto const put_price = black_scholes_option_price(CP::put, K, S, sigma, tau, r);
std::cout << "black-scholes call price = " << call_price.derivative(0) << '\n'
<< "black-scholes put price = " << put_price.derivative(0) << '\n'
<< "call delta = " << call_price.derivative(1) << '\n'
<< "put delta = " << put_price.derivative(1) << '\n'
<< "call gamma = " << call_price.derivative(2) << '\n'
<< "put gamma = " << put_price.derivative(2) << '\n';
return 0;
}
/*
Output:
black-scholes call price = 56.5136
black-scholes put price = 51.4109
call delta = 0.773818
put delta = -0.226182
call gamma = 0.00199852
put gamma = 0.00199852
*/
\end{verbatim}
\section{Advantages of Automatic Differentiation}
The above examples illustrate some of the advantages of using autodiff:
\begin{itemize}
\item Elimination of code redundancy. The existence of $N$ separate functions to calculate derivatives is a form
of code redundancy, with all the liabilities that come with it:
\begin{itemize}
\item Changes to one function require $N$ additional changes to other functions. In the \nth{3} example above,
consider how much larger and inter-dependent the above code base would be if a separate function were
written for \href{https://en.wikipedia.org/wiki/Greeks\_(finance)#Formulas\_for\_European\_option\_Greeks}
{each Greek} value.
\item Dependencies upon a derivative function for a different purpose will break when changes are made to
the original function. What doesn't need to exist cannot break.
\item Code bloat, reducing conceptual integrity. Control over the evolution of code is easier/safer when
the code base is smaller and able to be intuitively grasped.
\end{itemize}
\item Accuracy of derivatives over finite difference methods. Single-iteration finite difference methods always
include a $\Delta x$ free variable that must be carefully chosen for each application. If $\Delta x$ is too
small, then numerical errors become large. If $\Delta x$ is too large, then mathematical errors become large.
With autodiff, there are no free variables to set and the accuracy of the answer is generally superior to finite
difference methods even with the best choice of $\Delta x$.
\end{itemize}
\section{Mathematics}
In order for the usage of the autodiff library to make sense, a basic understanding of the mathematics will help.
\subsection{Truncated Taylor Series}
Basic calculus courses teach that a real \href{https://en.wikipedia.org/wiki/Analytic_function}{analytic function}
$f : D\rightarrow\mathbb{R}$ is one which can be expressed as a Taylor series at a point
$x_0\in D\subseteq\mathbb{R}$:
\[
f(x) = f(x_0) + f'(x_0)(x-x_0) + \frac{f''(x_0)}{2!}(x-x_0)^2 + \frac{f'''(x_0)}{3!}(x-x_0)^3 + \cdots
\]
One way of thinking about this form is that given the value of an analytic function $f(x_0)$ and its derivatives
$f'(x_0), f''(x_0), f'''(x_0), ...$ evaluated at a point $x_0$, then the value of the function
$f(x)$ can be obtained at any other point $x\in D$ using the above formula.
Let us make the substitution $x=x_0+\varepsilon$ and rewrite the above equation to get:
\[
f(x_0+\varepsilon) = f(x_0) + f'(x_0)\varepsilon + \frac{f''(x_0)}{2!}\varepsilon^2 + \frac{f'''(x_0)}{3!}\varepsilon^3 + \cdots
\]
Now consider $\varepsilon$ as {\it an abstract algebraic entity that never acquires a numeric value}, much like
one does in basic algebra with variables like $x$ or $y$. For example, we can still manipulate entities
like $xy$ and $(1+2x+3x^2)$ without having to assign specific numbers to them.
Using this formula, autodiff goes in the other direction. Given a general formula/algorithm for calculating
$f(x_0+\varepsilon)$, the derivatives are obtained from the coefficients of the powers of $\varepsilon$
in the resulting computation. The general coefficient for $\varepsilon^n$ is
\[\frac{f^{(n)}(x_0)}{n!}.\]
Thus to obtain $f^{(n)}(x_0)$, the coefficient of $\varepsilon^n$ is multiplied by $n!$.
\subsubsection{Example}
Apply the above technique to calculate the derivatives of $f(x)=x^4$ at $x_0=2$.
The first step is to evaluate $f(x_0+\varepsilon)$ and simply go through the calculation/algorithm, treating
$\varepsilon$ as an abstract algebraic entity:
\begin{align*}
f(x_0+\varepsilon) &= f(2+\varepsilon) \\
&= (2+\varepsilon)^4 \\
&= \left(4+4\varepsilon+\varepsilon^2\right)^2 \\
&= 16+32\varepsilon+24\varepsilon^2+8\varepsilon^3+\varepsilon^4.
\end{align*}
Equating the powers of $\varepsilon$ from this result with the above $\varepsilon$-taylor expansion
yields the following equalities:
\[
f(2) = 16, \qquad
f'(2) = 32, \qquad
\frac{f''(2)}{2!} = 24, \qquad
\frac{f'''(2)}{3!} = 8, \qquad
\frac{f^{(4)}(2)}{4!} = 1, \qquad
\frac{f^{(5)}(2)}{5!} = 0.
\]
Multiplying both sides by the respective factorials gives
\[
f(2) = 16, \qquad
f'(2) = 32, \qquad
f''(2) = 48, \qquad
f'''(2) = 48, \qquad
f^{(4)}(2) = 24, \qquad
f^{(5)}(2) = 0.
\]
These values can be directly confirmed by the \href{https://en.wikipedia.org/wiki/Power_rule}{power rule}
applied to $f(x)=x^4$.
\subsection{Arithmetic}
What was essentially done above was to take a formula/algorithm for calculating $f(x_0)$ from a number $x_0$,
and instead apply the same formula/algorithm to a polynomial $x_0+\varepsilon$. Intermediate steps operate on
values of the form
\[
{\bf x} = x_0 + x_1\varepsilon + x_2\varepsilon^2 +\cdots+ x_N\varepsilon^N
\]
and the final return value is of this polynomial form as well. In other words, the normal arithmetic operators
$+,-,\times,\div$ applied to numbers $x$ are instead applied to polynomials $\bf x$. Through the overloading of C++
operators and functions, floating point data types are replaced with data types that represent these polynomials. More
specifically, C++ types such as {\tt double} are replaced with {\tt std::array<double,N+1>}, which hold the above
$N+1$ coefficients $x_i$, and are wrapped in a {\tt class} that overloads all of the arithmetic operators.
The logic of these arithmetic operators simply mirror that which is applied to polynomials. We'll look at
each of the 4 arithmetic operators in detail.
\subsubsection{Addition}
The addition of polynomials $\bf x$ and $\bf y$ is done component-wise:
\begin{align*}
{\bf z} &= {\bf x} + {\bf y} \\
&= \left(\sum_{i=0}^Nx_i\varepsilon^i\right) + \left(\sum_{i=0}^Ny_i\varepsilon^i\right) \\
&= \sum_{i=0}^N(x_i+y_i)\varepsilon^i \\
z_i &= x_i + y_i \qquad \text{for}\; i\in\{0,1,2,...,N\}.
\end{align*}
\subsubsection{Subtraction}
Subtraction follows the same form as addition:
\begin{align*}
{\bf z} &= {\bf x} - {\bf y} \\
&= \left(\sum_{i=0}^Nx_i\varepsilon^i\right) - \left(\sum_{i=0}^Ny_i\varepsilon^i\right) \\
&= \sum_{i=0}^N(x_i-y_i)\varepsilon^i \\
z_i &= x_i - y_i \qquad \text{for}\; i\in\{0,1,2,...,N\}.
\end{align*}
\subsubsection{Multiplication}
Multiplication produces higher-order terms:
\begin{align*}
{\bf z} &= {\bf x} \times {\bf y} \\
&= \left(\sum_{i=0}^Nx_i\varepsilon^i\right) \left(\sum_{i=0}^Ny_i\varepsilon^i\right) \\
&= x_0y_0 + (x_0y_1+x_1y_0)\varepsilon + (x_0y_2+x_1y_1+x_2y_0)\varepsilon^2 + \cdots +
\left(\sum_{j=0}^Nx_jy_{N-j}\right)\varepsilon^N + O\left(\varepsilon^{N+1}\right) \\
&= \sum_{i=0}^N\sum_{j=0}^ix_jy_{i-j}\varepsilon^i + O\left(\varepsilon^{N+1}\right) \\
z_i &= \sum_{j=0}^ix_jy_{i-j} \qquad \text{for}\; i\in\{0,1,2,...,N\}.
\end{align*}
In the case of multiplication, terms involving powers of $\varepsilon$ greater than $N$, collectively denoted
by $O\left(\varepsilon^{N+1}\right)$, are simply discarded. Fortunately, the values of $z_i$ for $i\le N$ do not
depend on any of these discarded terms, so there is no loss of precision in the final answer. The only information
that is lost are the values of higher order derivatives, which we are not interested in anyway. If we were, then
we would have simply chosen a larger value of $N$ to begin with.
\subsubsection{Division}
Division is not directly calculated as are the others. Instead, to find the components of
${\bf z}={\bf x}\div{\bf y}$ we require that ${\bf x}={\bf y}\times{\bf z}$. This yields
a recursive formula for the components $z_i$:
\begin{align*}
x_i &= \sum_{j=0}^iy_jz_{i-j} \\
&= y_0z_i + \sum_{j=1}^iy_jz_{i-j} \\
z_i &= \frac{1}{y_0}\left(x_i - \sum_{j=1}^iy_jz_{i-j}\right) \qquad \text{for}\; i\in\{0,1,2,...,N\}.
\end{align*}
In the case of division, the values for $z_i$ must be calculated sequentially, since $z_i$
depends on the previously calculated values $z_0, z_1, ..., z_{i-1}$.
\subsection{General Functions}
Calling standard mathematical functions such as {\tt log()}, {\tt cos()}, etc. should return accurate higher
order derivatives. For example, {\tt exp(x)} may be written internally as a specific \nth{14}-degree polynomial to
approximate $e^x$ when $0<x<1$. This would mean that the \nth{15} derivative, and all higher order derivatives, would
be 0, however we know that $\frac{d^{15}}{dx^{15}}e^x=e^x$. How should such functions whose derivatives are known
be written to provide accurate higher order derivatives? The answer again comes back to the function's Taylor series.
To simplify notation, for a given polynomial ${\bf x} = x_0 + x_1\varepsilon + x_2\varepsilon^2 +\cdots+
x_N\varepsilon^N$ define
\[
{\bf x}_\varepsilon = x_1\varepsilon + x_2\varepsilon^2 +\cdots+ x_N\varepsilon^N = \sum_{i=1}^Nx_i\varepsilon^i.
\]
This allows for a concise expression of a general function $f$ of $\bf x$:
\begin{align*}
f({\bf x}) &= f(x_0 + {\bf x}_\varepsilon) \\
& = f(x_0) + f'(x_0){\bf x}_\varepsilon + \frac{f''(x_0)}{2!}{\bf x}_\varepsilon^2 + \frac{f'''(x_0)}{3!}{\bf x}_\varepsilon^3 + \cdots + \frac{f^{(N)}(x_0)}{N!}{\bf x}_\varepsilon^N + O\left(\varepsilon^{N+1}\right) \\
& = \sum_{i=0}^N\frac{f^{(i)}(x_0)}{i!}{\bf x}_\varepsilon^i + O\left(\varepsilon^{N+1}\right)
\end{align*}
where $\varepsilon$ has been substituted with ${\bf x}_\varepsilon$ in the $\varepsilon$-taylor series
for $f(x)$. This form gives a recipe for calculating $f({\bf x})$ in general from regular numeric calculations
$f(x_0)$, $f'(x_0)$, $f''(x_0)$, ... and successive powers of the epsilon terms ${\bf x}_\varepsilon$.
For an application in which we are interested in up to $N$ derivatives in $x$ the data structure to hold
this information is an $(N+1)$-element array {\tt v} whose general element is
\[ {\tt v[i]} = \frac{f^{(i)}(x_0)}{i!} \qquad \text{for}\; i\in\{0,1,2,...,N\}. \]
\subsection{Multiple Variables}
In C++, the generalization to mixed partial derivatives with multiple independent variables is conveniently achieved
with recursion. To begin to see the recursive pattern, consider a two-variable function $f(x,y)$. Since $x$
and $y$ are independent, they require their own independent epsilons $\varepsilon_x$ and $\varepsilon_y$,
respectively.
Expand $f(x,y)$ for $x=x_0+\varepsilon_x$:
\begin{align*}
f(x_0+\varepsilon_x,y) &= f(x_0,y)
+ \frac{\partial f}{\partial x}(x_0,y)\varepsilon_x
+ \frac{1}{2!}\frac{\partial^2 f}{\partial x^2}(x_0,y)\varepsilon_x^2
+ \frac{1}{3!}\frac{\partial^3 f}{\partial x^3}(x_0,y)\varepsilon_x^3
+ \cdots
+ \frac{1}{M!}\frac{\partial^M f}{\partial x^M}(x_0,y)\varepsilon_x^M
+ O\left(\varepsilon_x^{M+1}\right) \\
&= \sum_{i=0}^M\frac{1}{i!}\frac{\partial^i f}{\partial x^i}(x_0,y)\varepsilon_x^i + O\left(\varepsilon_x^{M+1}\right).
\end{align*}
Next, expand $f(x_0+\varepsilon_x,y)$ for $y=y_0+\varepsilon_y$:
\begin{align*}
f(x_0+\varepsilon_x,y_0+\varepsilon_y) &= \sum_{j=0}^N\frac{1}{j!}\frac{\partial^j}{\partial y^j}
\left(\sum_{i=0}^M\varepsilon_x^i\frac{1}{i!}\frac{\partial^if}{\partial x^i}\right)(x_0,y_0)\varepsilon_y^j
+ O\left(\varepsilon_x^{M+1}\right) + O\left(\varepsilon_y^{N+1}\right) \\
&= \sum_{i=0}^M\sum_{j=0}^N\frac{1}{i!j!}\frac{\partial^{i+j}f}{\partial x^i\partial y^j}(x_0,y_0)
\varepsilon_x^i\varepsilon_y^j + O\left(\varepsilon_x^{M+1}\right) + O\left(\varepsilon_y^{N+1}\right).
\end{align*}
Similar to the single-variable case, for an application in which we are interested in up to $M$ derivatives in
$x$ and $N$ derivatives in $y$, the data structure to hold this information is an $(M+1)\times(N+1)$
array {\tt v} whose element at $(i,j)$ is
\[
{\tt v[i][j]} = \frac{1}{i!j!}\frac{\partial^{i+j}f}{\partial x^i\partial y^j}(x_0,y_0)
\qquad \text{for}\; (i,j)\in\{0,1,2,...,M\}\times\{0,1,2,...,N\}.
\]
The generalization to additional independent variables follows the same pattern.
\subsubsection{Declaring Multiple Variables}
Internally, independent variables are represented by vectors within orthogonal vector spaces. Because of this,
one must be careful when declaring more than one independent variable so that they do not end up in
parallel vector spaces. This can easily be achieved by following one rule:
\begin{itemize}
\item When declaring more than one independent variable, call {\tt make\_ftuple<>()} once and only once.
\end{itemize}
The tuple of values returned are independent. Though it is possible to achieve the same result with multiple calls
to {\tt make\_fvar}, this is an easier and less error-prone method. See Section~\ref{multivar} for example usage.
%\section{Usage}
%
%\subsection{Single Variable}
%
%To calculate derivatives of a single variable $x$, at a particular value $x_0$, the following must be
%specified at compile-time:
%
%\begin{enumerate}
%\item The numeric data type {\tt T} of $x_0$. Examples: {\tt double},
% {\tt boost::multiprecision::cpp\_bin\_float\_50}, etc.
%\item The maximum derivative order $M$ that is to be calculated with respect to $x$.
%\end{enumerate}
%Note that both of these requirements are entirely analogous to declaring and using a {\tt std::array<T,N>}. {\tt T}
%and {\tt N} must be set as compile-time, but which elements in the array are accessed can be determined at run-time,
%just as the choice of what derivatives to query in autodiff can be made during run-time.
%
%To declare and initialize $x$:
%
%\begin{verbatim}
% using namespace boost::math::differentiation;
% autodiff_fvar<T,M> x = make_fvar<T,M>(x0);
%\end{verbatim}
%where {\tt x0} is a run-time value of type {\tt T}. Assuming {\tt 0 < M}, this represents the polynomial $x_0 +
%\varepsilon$. Internally, the member variable of type {\tt std::array<T,M>} is {\tt v = \{ x0, 1, 0, 0, ... \}},
%consistent with the above mathematical treatise.
%
%To find the derivatives $f^{(n)}(x_0)$ for $0\le n\le M$ of a function
%$f : \mathbb{R}\rightarrow\mathbb{R}$, the function can be represented as a template
%
%\begin{verbatim}
% template<typename T>
% T f(T x);
%\end{verbatim}
%Using a generic type {\tt T} allows for {\tt x} to be of a regular type such as {\tt double}, but also allows for\\
%{\tt boost::math::differentiation::autodiff\_fvar<>} types.
%
%Internal calls to mathematical functions must allow for
%\href{https://en.cppreference.com/w/cpp/language/adl}{argument-dependent lookup} (ADL). Many standard library functions
%are overloaded in the {\tt boost::math::differentiation} namespace. For example, instead of calling {\tt std::cos(x)}
%from within {\tt f}, include the line {\tt using std::cos;} and call {\tt cos(x)} without a namespace prefix.
%
%Calling $f$ and retrieving the calculated value and derivatives:
%
%\begin{verbatim}
% using namespace boost::math::differentiation;
% autodiff_fvar<T,M> x = make_fvar<T,M>(x0);
% autodiff_fvar<T,M> y = f(x);
% for (int n=0 ; n<=M ; ++n)
% std::cout << "y.derivative("<<n<<") == " << y.derivative(n) << std::endl;
%\end{verbatim}
%{\tt y.derivative(0)} returns the undifferentiated value $f(x_0)$, and {\tt y.derivative(n)} returns $f^{(n)}(x_0)$.
%Casting {\tt y} to type {\tt T} also gives the undifferentiated value. In other words, the following 3 values
%are equal:
%
%\begin{enumerate}
%\item {\tt f(x0)}
%\item {\tt y.derivative(0)}
%\item {\tt static\_cast<T>(y)}
%\end{enumerate}
%
%\subsection{Multiple Variables}
%
%Independent variables are represented in autodiff as independent dimensions within a multi-dimensional array.
%This is perhaps best illustrated with examples. The {\tt namespace boost::math::differentiation} is assumed.
%
%The following instantiates a variable of $x=13$ with up to 3 orders of derivatives:
%
%\begin{verbatim}
% autodiff_fvar<double,3> x = make_fvar<double,3>(13);
%\end{verbatim}
%This instantiates {\bf an independent} value of $y=14$ with up to 4 orders of derivatives:
%
%\begin{verbatim}
% autodiff_fvar<double,0,4> y = make_fvar<double,0,4>(14);
%\end{verbatim}
%Combining them together {\bf promotes} their data type automatically to the smallest multidimensional array that
%accommodates both.
%
%\begin{verbatim}
% // z is promoted to autodiff_fvar<double,3,4>
% auto z = 10*x*x + 50*x*y + 100*y*y;
%\end{verbatim}
%The object {\tt z} holds a 2-dimensional array, thus {\tt derivative(...)} is a 2-parameter method:
%
%\[
%{\tt z.derivative(i,j)} = \frac{\partial^{i+j}f}{\partial x^i\partial y^j}(13,14)
% \qquad \text{for}\; (i,j)\in\{0,1,2,3\}\times\{0,1,2,3,4\}.
%\]
%A few values of the result can be confirmed through inspection:
%
%\begin{verbatim}
% z.derivative(2,0) == 20
% z.derivative(1,1) == 50
% z.derivative(0,2) == 200
%\end{verbatim}
%Note how the position of the parameters in {\tt derivative(...)} match how {\tt x} and {\tt y} were declared.
%This will be clarified next.
%
%\subsubsection{Two Rules of Variable Initialization}
%
%In general, there are two rules to keep in mind when dealing with multiple variables:
%
%\begin{enumerate}
%\item Independent variables correspond to parameter position, in both the initialization {\tt make\_fvar<T,...>}
% and calls to {\tt derivative(...)}.
%\item The last template position in {\tt make\_fvar<T,...>} determines which variable a derivative will be
% taken with respect to.
%\end{enumerate}
%Both rules are illustrated with an example in which there are 3 independent variables $x,y,z$ and 1 dependent
%variable $w=f(x,y,z)$, though the following code readily generalizes to any number of independent variables, limited
%only by the C++ compiler/memory/platform. The maximum derivative order of each variable is {\tt Nx}, {\tt Ny}, and
%{\tt Nz}, respectively. Then the type for {\tt w} is {\tt boost::math::differentiation::autodiff\_fvar<T,Nx,Ny,Nz>}
%and all possible mixed partial derivatives are available via
%
%\[
%{\tt w.derivative(nx,ny,nz)} =
% \frac{\partial^{n_x+n_y+n_z}f}{\partial x^{n_x}\partial y^{n_y}\partial z^{n_z} }(x_0,y_0,z_0)
%\]
%for $(n_x,n_y,n_z)\in\{0,1,2,...,N_x\}\times\{0,1,2,...,N_y\}\times\{0,1,2,...,N_z\}$ where $x_0, y_0, z_0$ are
%the numerical values at which the function $f$ and its derivatives are evaluated.
%
%In code:
%\begin{verbatim}
% using namespace boost::math::differentiation;
%
% using var = autodiff_fvar<double,Nx,Ny,Nz>; // Nx, Ny, Nz are constexpr size_t.
%
% var x = make_fvar<double,Nx>(x0); // x0 is of type double
% var y = make_fvar<double,Nx,Ny>(y0); // y0 is of type double
% var z = make_fvar<double,Nx,Ny,Nz>(z0); // z0 is of type double
%
% var w = f(x,y,z);
%
% for (size_t nx=0 ; nx<=Nx ; ++nx)
% for (size_t ny=0 ; ny<=Ny ; ++ny)
% for (size_t nz=0 ; nz<=Nz ; ++nz)
% std::cout << "w.derivative("<<nx<<','<<ny<<','<<nz<<") == "
% << w.derivative(nx,ny,nz) << std::endl;
%\end{verbatim}
%Note how {\tt x}, {\tt y}, and {\tt z} are initialized: the last template parameter determines which variable
%$x, y,$ or $z$ a derivative is taken with respect to. In terms of the $\varepsilon$-polynomials
%above, this determines whether to add $\varepsilon_x, \varepsilon_y,$ or $\varepsilon_z$ to
%$x_0, y_0,$ or $z_0$, respectively.
%
%In contrast, the following initialization of {\tt x} would be INCORRECT:
%
%\begin{verbatim}
% var x = make_fvar<T,Nx,0>(x0); // WRONG
%\end{verbatim}
%Mathematically, this represents $x_0+\varepsilon_y$, since the last template parameter corresponds to the
%$y$ variable, and thus the resulting value will be invalid.
%
%\subsubsection{Type Promotion}
%
%The previous example can be optimized to save some unnecessary computation, by declaring smaller arrays,
%and relying on autodiff's automatic type-promotion:
%
%\begin{verbatim}
% using namespace boost::math::differentiation;
%
% autodiff_fvar<double,Nx> x = make_fvar<double,Nx>(x0);
% autodiff_fvar<double,0,Ny> y = make_fvar<double,0,Ny>(y0);
% autodiff_fvar<double,0,0,Nz> z = make_fvar<double,0,0,Nz>(z0);
%
% autodiff_fvar<double,Nx,Ny,Nz> w = f(x,y,z);
%
% for (size_t nx=0 ; nx<=Nx ; ++nx)
% for (size_t ny=0 ; ny<=Ny ; ++ny)
% for (size_t nz=0 ; nz<=Nz ; ++nz)
% std::cout << "w.derivative("<<nx<<','<<ny<<','<<nz<<") == "
% << w.derivative(nx,ny,nz) << std::endl;
%\end{verbatim}
%For example, if one of the first steps in the computation of $f$ was {\tt z*z}, then a significantly less number of
%multiplications and additions may occur if {\tt z} is declared as {\tt autodiff\_fvar<double,0,0,Nz>} as opposed to \\
%{\tt autodiff\_fvar<double,Nx,Ny,Nz>}. There is no loss of precision with the former, since the extra dimensions
%represent 0 values. Once {\tt z} is combined with {\tt x} and {\tt y} during the computation, the types will be
%promoted as necessary. This is the recommended way to initialize variables in autodiff.
\section{Writing Functions for Autodiff Compatibility}\label{compatibility}
In this section, a general procedure is given for writing new, and transforming existing, C++ mathematical
functions for compatibility with autodiff.
There are 3 categories of functions that require different strategies:
\begin{enumerate}
\item Piecewise-rational functions. These are simply piecewise quotients of polynomials. All that is needed is to
turn the function parameters and return value into generic (template) types. This will then allow the function
to accept and return autodiff's {\tt fvar} types, thereby using autodiff's overloaded arithmetic operators
which calculate the derivatives automatically.
\item Functions that call existing autodiff functions. This is the same as the previous, but may also include
calls to functions that are in the autodiff library. Examples: {\tt exp()}, {\tt log()}, {\tt tgamma()}, etc.
\item New functions for which the derivatives can be calculated. This is the most general technique, as it
allows for the development of a function which do not fall into the previous two categories.
\end{enumerate}
Functions written in any of these ways may then be added to the autodiff library.
\subsection{Piecewise-Rational Functions}
\[
f(x) = \frac{1}{1+x^2}
\]
By simply writing this as a template function, autodiff can calculate derivatives for it:
\begin{Verbatim}[xleftmargin=2em]
#include <boost/math/differentiation/autodiff.hpp>
#include <iostream>
template <typename T>
T rational(T const& x) {
return 1 / (1 + x * x);
}
int main() {
using namespace boost::math::differentiation;
auto const x = make_fvar<double, 10>(0);
auto const y = rational(x);
std::cout << std::setprecision(std::numeric_limits<double>::digits10)
<< "y.derivative(10) = " << y.derivative(10) << std::endl;
return 0;
}
/*
Output:
y.derivative(10) = -3628800
*/
\end{Verbatim}
As simple as $f(x)$ may seem, the derivatives can get increasingly complex as derivatives are taken.
For example, the \nth{10} derivative has the form
\[
f^{(10)}(x) = -3628800\frac{1 - 55x^2 + 330x^4 - 462x^6 + 165x^8 - 11x^{10}}{(1 + x^2)^{11}}.
\]
Derivatives of $f(x)$ are useful, and in fact used, in calculating higher order derivatives for $\arctan(x)$
for instance, since
\[
\arctan^{(n)}(x) = \left(\frac{d}{dx}\right)^{n-1} \frac{1}{1+x^2}\qquad\text{for}\quad 1\le n.
\]
\subsection{Functions That Call Existing Autodiff Functions}
Many of the standard library math function are overloaded in autodiff. It is recommended to use
\href{https://en.cppreference.com/w/cpp/language/adl}{argument-dependent lookup} (ADL) in order for functions to
be written in a way that is general enough to accommodate standard types ({\tt double}) as well as autodiff types
({\tt fvar}).
\\
Example:
\begin{Verbatim}[xleftmargin=2em]
#include <boost/math/constants/constants.hpp>
#include <cmath>
using namespace boost::math::constants;
// Standard normal cumulative distribution function
template <typename T>
T Phi(T const& x)
{
return 0.5 * std::erfc(-one_div_root_two<T>() * x);
}
\end{Verbatim}
Though {\tt Phi(x)} is general enough to handle the various fundamental floating point types, this will
not work if {\tt x} is an autodiff {\tt fvar} variable, since {\tt std::erfc} does not include a specialization
for {\tt fvar}. The recommended solution is to remove the namespace prefix {\tt std::} from {\tt erfc}:
\begin{Verbatim}[xleftmargin=2em]
#include <boost/math/constants/constants.hpp>
#include <boost/math/differentiation/autodiff.hpp>
#include <cmath>
using namespace boost::math::constants;
// Standard normal cumulative distribution function
template <typename T>
T Phi(T const& x)
{
using std::erfc;
return 0.5 * erfc(-one_div_root_two<T>() * x);
}
\end{Verbatim}
In this form, when {\tt x} is of type {\tt fvar}, the C++ compiler will search for and find a function {\tt erfc}
within the same namespace as {\tt fvar}, which is in the autodiff library, via ADL. Because of the using-declaration,
it will also call {\tt std::erfc} when {\tt x} is a fundamental type such as {\tt double}.
\subsection{New Functions For Which The Derivatives Can Be Calculated}\label{new_functions}
Mathematical functions which do not fall into the previous two categories can be constructed using autodiff helper
functions. This requires a separate function for calculating the derivatives. In case you are asking yourself what
good is an autodiff library if one needs to supply the derivatives, the answer is that the new function will fit
in with the rest of the autodiff library, thereby allowing for the creation of additional functions via all of
the arithmetic operators, plus function composition, which was not readily available without the library.
The example given here is for {\tt cos}:
\begin{Verbatim}[xleftmargin=2em]
template <typename RealType, size_t Order>
fvar<RealType, Order> cos(fvar<RealType, Order> const& cr) {
using std::cos;
using std::sin;
using root_type = typename fvar<RealType, Order>::root_type;
constexpr size_t order = fvar<RealType, Order>::order_sum;
root_type const d0 = cos(static_cast<root_type>(cr));
if constexpr (order == 0)
return fvar<RealType, Order>(d0);
else {
root_type const d1 = -sin(static_cast<root_type>(cr));
root_type const derivatives[4]{d0, d1, -d0, -d1};
return cr.apply_derivatives(order,
[&derivatives](size_t i) { return derivatives[i & 3]; });
}
}
\end{Verbatim}
This uses the helper function {\tt fvar::apply\_derivatives} which takes two parameters:
\begin{enumerate}
\item The highest order derivative to be calculated.
\item A function that maps derivative order to derivative value.
\end{enumerate}
The highest order derivative necessary to be calculated is generally equal to {\tt fvar::order\_sum}. In the case
of {\tt sin} and {\tt cos}, the derivatives are cyclical with period 4. Thus it is sufficient to store only these
4 values into an array, and take the derivative order modulo 4 as the index into this array.
A second helper function, not shown here, is {\tt apply\_coefficients}. This is used the same as
{\tt apply\_derivatives} except that the supplied function calculates coefficients instead of derivatives.
The relationship between a coefficient $C_n$ and derivative $D_n$ for derivative order $n$ is
\[
C_n = \frac{D_n}{n!}.
\]
Internally, {\tt fvar} holds coefficients rather than derivatives, so in case the coefficient values are more readily
available than the derivatives, it can save some unnecessary computation to use {\tt apply\_coefficients}.
See the definition of {\tt atan} for an example.
Both of these helper functions use Horner's method when calculating the resulting polynomial {\tt fvar}. This works
well when the derivatives are finite, but in cases where derivatives are infinite, this can quickly result in NaN
values as the computation progresses. In these cases, one can call non-Horner versions of both function which
better ``isolate'' infinite values so that they are less likely to evolve into NaN values.
The four helper functions available for constructing new autodiff functions from known coefficients/derivatives are:
\begin{enumerate}
\item {\tt fvar::apply\_coefficients}
\item {\tt fvar::apply\_coefficients\_nonhorner}
\item {\tt fvar::apply\_derivatives}
\item {\tt fvar::apply\_derivatives\_nonhorner}
\end{enumerate}
\section{Function Writing Guidelines}
At a high level there is one fairly simple principle, loosely and intuitively speaking, to writing functions for
which autodiff can effectively calculate derivatives: \\
{\bf Autodiff Function Principle (AFP)}
\begin{displayquote}
A function whose branches in logic correspond to piecewise analytic calculations over non-singleton intervals,
with smooth transitions between the intervals, and is free of indeterminate forms in the calculated value and
higher order derivatives, will work fine with autodiff.
\end{displayquote}
Stating this with greater mathematical rigor can be done. However what seems to be more practical, in this
case, is to give examples and categories of examples of what works, what doesn't, and how to remedy some of the
common problems that may be encountered. That is the approach taken here.
\subsection{Example 1: $f(x)=\max(0,x)$}
One potential implementation of $f(x)=\max(0,x)$ is:
\begin{verbatim}
template<typename T>
T f(const T& x)
{
return 0 < x ? x : 0;
}
\end{verbatim}
Though this is consistent with Section~\ref{compatibility}, there are two problems with it:
\begin{enumerate}
\item {\tt f(nan) = 0}. This problem is independent of autodiff, but is worth addressing anyway. If there is
an indeterminate form that arises within a calculation and is input into $f$, then it gets ``covered up'' by
this implementation leading to an unknowingly incorrect result. Better for functions in general to propagate
NaN values, so that the user knows something went wrong and doesn't rely on an incorrect result, and likewise
the developer can track down where the NaN originated from and remedy it.
\item $f'(0) = 0$ when autodiff is applied. This is because {\tt f} returns 0 as a constant when {\tt x==0}, wiping
out any of the derivatives (or sensitivities) that {\tt x} was holding as an autodiff variable. Instead, let us
apply the AFP and identify the two intervals over which $f$ is defined: $(-\infty,0]\cup(0,\infty)$.
Though the function itself is not analytic at $x=0$, we can attempt somewhat to smooth out this transition
point by averaging the calculation of $f(x)$ at $x=0$ from each interval. If $x<0$ then the result is simply
0, and if $0<x$ then the result is $x$. The average is $\frac{1}{2}(0 + x)$ which will allow autodiff to
calculate $f'(0)=\frac{1}{2}$. This is a more reasonable answer.
\end{enumerate}
A better implementation that resolves both issues is:
\begin{verbatim}
template<typename T>
T f(const T& x)
{
if (x < 0)
return 0;
else if (x == 0)
return 0.5*x;
else
return x;
}
\end{verbatim}
\subsection{Example 2: $f(x)=\sinc(x)$}
The definition of $\sinc:\mathbb{R}\rightarrow\mathbb{R}$ is
\[
\sinc(x) = \begin{cases}
1 &\text{if}\; x = 0 \\
\frac{\sin(x)}{x} &\text{otherwise.}\end{cases}
\]
A potential implementation is:
\begin{verbatim}
template<typename T>
T sinc(const T& x)
{
using std::sin;
return x == 0 ? 1 : sin(x) / x;
}
\end{verbatim}
Though this is again consistent with Section~\ref{compatibility}, and returns correct non-derivative values,
it returns a constant when {\tt x==0} thereby losing all derivative information contained in {\tt x} and
contributions from $\sinc$. For example, $\sinc''(0)=-\frac{1}{3}$, however {\tt y.derivative(2) == 0} when
{\tt y = sinc(make\_fvar<double,2>(0))} using the above incorrect implementation. Applying the AFP, the intervals
upon which separate branches of logic are applied are $(-\infty,0)\cup[0,0]\cup(0,\infty)$. The violation occurs
due to the singleton interval $[0,0]$, even though a constant function of 1 is technically analytic. The remedy
is to define a custom $\sinc$ overload and add it to the autodiff library. This has been done. Mathematically, it
is well-defined and free of indeterminate forms, as is the \nth{3} expression in the equalities
\[
\frac{1}{x}\sin(x) = \frac{1}{x}\sum_{n=0}^\infty\frac{(-1)^n}{(2n+1)!}x^{2n+1}
= \sum_{n=0}^\infty\frac{(-1)^n}{(2n+1)!}x^{2n}.
\]
The autodiff library contains helper functions to help write function overloads when the derivatives of a function
are known. This is an advanced feature and documentation for this may be added at a later time.
For now, it is worth understanding the ways in which indeterminate forms can occur within a mathematical calculation,
and avoid them when possible by rewriting the function. Table~\ref{3nans} compares 3 types of indeterminate
forms. Assume the product {\tt a*b} is a positive finite value.
\begin{table}[h]
\centering\begin{tabular}{m{7em}||c|c|c}
& $\displaystyle f(x)=\left(\frac{a}{x}\right)\times(bx^2)$
& $\displaystyle g(x)=\left(\frac{a}{x}\right)\times(bx)$
& $\displaystyle h(x)=\left(\frac{a}{x^2}\right)\times(bx)$ \\[0.618em]
\hline\hline
Mathematical\newline Limit
& $\displaystyle\lim_{x\rightarrow0}f(x) = 0$
& $\displaystyle\lim_{x\rightarrow0}g(x) = ab$
& $\displaystyle\lim_{x\rightarrow0}h(x) = \infty$ \\
\hline
Floating Point\newline Arithmetic
& {\tt f(0) = inf*0 = nan} & {\tt g(0) = inf*0 = nan} & {\tt h(0) = inf*0 = nan}
\end{tabular}
\caption{Automatic differentiation does not compute limits.
Indeterminate forms must be simplified manually. (These cases are not meant to be exhaustive.)}\label{3nans}
\end{table}
Indeterminate forms result in NaN values within a calculation. Mathematically, if they occur at locally isolated
points, then we generally prefer the mathematical limit as the result, even if it is infinite. As demonstrated in
Table~\ref{3nans}, depending upon the nature of the indeterminate form, the mathematical limit can be 0 (no matter
the values of $a$ or $b$), or $ab$, or $\infty$, but these 3 cases cannot be distinguished by the floating point
result of nan. Floating point arithmetic does not perform limits (directly), and neither does the autodiff library.
Thus it is up to the diligence of the developer to keep a watchful eye over where indeterminate forms can arise.
\subsection{Example 3: $f(x)=\sqrt x$ and $f'(0)=\infty$}
When working with functions that have infinite higher order derivatives, this can very quickly result in nans in
higher order derivatives as the computation progresses, as {\tt inf-inf}, {\tt inf/inf}, and {\tt 0*inf} result
in {\tt nan}. See Table~\ref{sqrtnan} for an example.
\begin{table}[h]
\centering\begin{tabular}{c||c|c|c|c}
$f(x)$ & $f(0)$ & $f'(0)$ & $f''(0)$ & $f'''(0)$ \\
\hline\hline
{\tt sqrt(x)} & {\tt 0} & {\tt inf} & {\tt -inf} & {\tt inf} \\
\hline
{\tt sqr(sqrt(x)+1)} & {\tt 1} & {\tt inf} & {\tt nan} & {\tt nan} \\
\hline
{\tt x+2*sqrt(x)+1} & {\tt 1} & {\tt inf} & {\tt -inf}& {\tt inf}
\end{tabular}
\caption{Indeterminate forms in higher order derivatives. {\tt sqr(x) == x*x}.}\label{sqrtnan}
\end{table}
Calling the autodiff-overloaded implementation of $f(x)=\sqrt x$ at the value {\tt x==0} results in the
\nth{1} row (after the header row) of Table~\ref{sqrtnan}, as is mathematically correct. The \nth{2} row shows
$f(x)=(\sqrt{x}+1)^2$ resulting in {\tt nan} values for $f''(0)$ and all higher order derivatives. This is due to
the internal arithmetic in which {\tt inf} is added to {\tt -inf} during the squaring, resulting in a {\tt nan}
value for $f''(0)$ and all higher orders. This is typical of {\tt inf} values in autodiff. Where they show up,
they are correct, however they can quickly introduce {\tt nan} values into the computation upon the addition of
oppositely signed {\tt inf} values, division by {\tt inf}, or multiplication by {\tt 0}. It is worth noting that
the infection of {\tt nan} only spreads upward in the order of derivatives, since lower orders do not depend upon
higher orders (which is also why dropping higher order terms in an autodiff computation does not result in any
loss of precision for lower order terms.)
The resolution in this case is to manually perform the squaring in the computation, replacing the \nth{2} row
with the \nth{3}: $f(x)=x+2\sqrt{x}+1$. Though mathematically equivalent, it allows autodiff to avoid {\tt nan}
values since $\sqrt x$ is more ``isolated'' in the computation. That is, the {\tt inf} values that unavoidably
show up in the derivatives of {\tt sqrt(x)} for {\tt x==0} do not have the chance to interact with other {\tt inf}
values as with the squaring.
\subsection{Summary}
The AFP gives a high-level unified guiding principle for writing C++ template functions that autodiff can
effectively evaluate derivatives for.
Examples have been given to illustrate some common items to avoid doing:
\begin{enumerate}
\item It is not enough for functions to be piecewise continuous. On boundary points between intervals, consider
returning the average expression of both intervals, rather than just one of them. Example: $\max(0,x)$ at $x=0$.
In cases where the limits from both sides must match, and they do not, then {\tt nan} may be a more appropriate
value depending on the application.
\item Avoid returning individual constant values (e.g. $\sinc(0)=1$.) Values must be computed uniformly along
with other values in its local interval. If that is not possible, then the function must be overloaded to
compute the derivatives precisely using the helper functions from Section~\ref{new_functions}.
\item Avoid intermediate indeterminate values in both the value ($\sinc(x)$ at $x=0$) and derivatives
($(\sqrt{x}+1)^2$ at $x=0$). Try to isolate expressions that may contain infinite values/derivatives so
that they do not introduce NaN values into the computation.
\end{enumerate}
\section{Acknowledgments}
\begin{itemize}
\item Kedar Bhat --- C++11 compatibility, Boost Special Functions compatibility testing, codecov integration,
and feedback.
\item Nick Thompson --- Initial feedback and help with Boost integration.
\item John Maddock --- Initial feedback and help with Boost integration.
\end{itemize}
\begin{thebibliography}{1}
\bibitem{ad} \url{https://en.wikipedia.org/wiki/Automatic\_differentiation}
\bibitem{ed} Andreas Griewank, Andrea Walther. \textit{Evaluating Derivatives}. SIAM, 2nd ed. 2008.
\end{thebibliography}
\end{document}
|