autodiff

c++14 compile-time symbolic differentiation with zero memory footprint

This is a proof-of-concept. Using modern c++14 techniques we are able to perform symbolic calculus and symbolic differentiation.

    using X = Var<0>;
    using Y = Var<1>;
    using Z = Var<2>;
    X x;
    Y y;
    Z z;
    auto r1 = (x*y) / (x+y);
    auto r2 = sin(x)*sin(x) + cos(x)*cos(x);
    auto r3 = (3_K / 12_K) * (x+y)*(x+y);
    cerr << r1.subs( x(2.), y(3.) ) << endl;
    cerr << r1.diff(x,x,x).subs( x(2.), y(3.) ) << endl;
    cerr << r1.diff(x,y).subs( x(2.), y(3.) ) << endl;
    //decltype(r3) = Mul<Mul<Div<Const<1>, Const<4>>, Add<Var<0>, Var<1>>>, Add<Var<0>, Var<1>>> 

Features

• Each symbol is a type.
• Symbolic expressions are types. We push decltype to its limits to deduce complex expression trees.
• Notion of symbols and Lvalues: operator(T&&) on a symbol yields a perfectly forwarded object initialization:
  • Lvalue<std::decay_t<T>,Id>{std::forward<T>(v)}
• Nothing is stored inside the type (empty types). Other approaches store the value in the symbol. We store nothing:
  • substitution subs(const Lvalue<Ts,Is>&...) will deduce an evaluation sequence at the point of call
  • as a variadic function, it works with any type
• Trivial simplifications are supported: x-x=0, x*0=0, 1*x=x, x+x=2*x etc ...
• static_assert can catch errors at compile-time like:
  • trying to substitue a variable twice or not at all
  • division by 0 or undefined quantity 0/0
• Multi-variate and multi-degree differentiation is again a variadic function.
• Fractions are simplified via their GCD at compile time.
• Suffix operator _K for constants.

Dessert

• I have used CRTP to block funny type deductions.
• I had to fiddle a little bit with SFINAE, but I tried to limit its usage (might be unavoidable since I'm exploiting overload resolution...)
• Of course everything is constexpr :)