mirror of
				https://github.com/saitohirga/WSJT-X.git
				synced 2025-10-30 12:30:23 -04:00 
			
		
		
		
	
		
			
	
	
		
			596 lines
		
	
	
		
			25 KiB
		
	
	
	
		
			Plaintext
		
	
	
	
	
	
		
		
			
		
	
	
			596 lines
		
	
	
		
			25 KiB
		
	
	
	
		
			Plaintext
		
	
	
	
	
	
|  | [section:root_finding_examples Examples of Root-Finding (with and without derivatives)] | ||
|  | 
 | ||
|  | [import ../../example/root_finding_example.cpp] | ||
|  | [import ../../example/root_finding_n_example.cpp] | ||
|  | [import ../../example/root_finding_multiprecision_example.cpp] | ||
|  | 
 | ||
|  | The examples demonstrate how to use the various tools for | ||
|  | [@http://en.wikipedia.org/wiki/Root-finding_algorithm root finding]. | ||
|  | 
 | ||
|  | We start with the simple cube root function `cbrt` ( C++ standard function name | ||
|  | [@http://en.cppreference.com/w/cpp/numeric/math/cbrt cbrt]) | ||
|  | showing root finding __cbrt_no_derivatives. | ||
|  | 
 | ||
|  | We then show how use of derivatives can improve the speed of convergence. | ||
|  | 
 | ||
|  | (But these examples are only a demonstration and do not try to make | ||
|  | the ultimate improvements of an 'industrial-strength' | ||
|  | implementation, for example, of `boost::math::cbrt`, mainly by using a better computed initial 'guess' | ||
|  | at [@boost:/libs/math/include/boost/math/special_functions/cbrt.hpp  cbrt.hpp]). | ||
|  | 
 | ||
|  | Then we show how a higher root (__fifth_root) [super 5][radic] can be computed, | ||
|  | and in | ||
|  | [@../../example/root_finding_n_example.cpp root_finding_n_example.cpp] | ||
|  | a generic method for the __nth_root that constructs the derivatives at compile-time. | ||
|  | 
 | ||
|  | These methods should be applicable to other functions that can be differentiated easily. | ||
|  | 
 | ||
|  | [section:cbrt_eg Finding the Cubed Root With and Without Derivatives] | ||
|  | 
 | ||
|  | First some `#includes` that will be needed. | ||
|  | 
 | ||
|  | [root_finding_include_1] | ||
|  | 
 | ||
|  | [tip For clarity, `using` statements are provided to list what functions are being used in this example: | ||
|  | you can, of course, partly or fully qualify the names in other ways. | ||
|  | (For your application, you may wish to extract some parts into header files, | ||
|  | but you should never use `using` statements globally in header files).] | ||
|  | 
 | ||
|  | Let's suppose we want to find the root of a number ['a], and to start, compute the cube root. | ||
|  | 
 | ||
|  | So the equation we want to solve is: | ||
|  | 
 | ||
|  | __spaces ['f(x) = x[cubed] -a] | ||
|  | 
 | ||
|  | We will first solve this without using any information | ||
|  | about the slope or curvature of the cube root function. | ||
|  | 
 | ||
|  | Fortunately, the cube-root function is 'Really Well Behaved' in that it is monotonic | ||
|  | and has only one root (we leave negative values 'as an exercise for the student'). | ||
|  | 
 | ||
|  | We then show how adding what we can know about this function, first just the slope | ||
|  | or 1st derivative ['f'(x)], will speed homing in on the solution. | ||
|  | 
 | ||
|  | Lastly, we show how adding the curvature ['f''(x)] too will speed convergence even more. | ||
|  | 
 | ||
|  | [h3:cbrt_no_derivatives Cube root function without derivatives] | ||
|  | 
 | ||
|  | First we define a function object (functor): | ||
|  | 
 | ||
|  | [root_finding_noderiv_1] | ||
|  | 
 | ||
|  | Implementing the cube-root function itself is fairly trivial now: | ||
|  | the hardest part is finding a good approximation to begin with. | ||
|  | In this case we'll just divide the exponent by three. | ||
|  | (There are better but more complex guess algorithms used in 'real life'.) | ||
|  | 
 | ||
|  | [root_finding_noderiv_2] | ||
|  | 
 | ||
|  | This snippet from `main()` in [@../../example/root_finding_example.cpp root_finding_example.cpp] | ||
|  | shows how it can be used. | ||
|  | 
 | ||
|  | [root_finding_main_1] | ||
|  | 
 | ||
|  | [pre | ||
|  |   cbrt_noderiv(27) = 3 | ||
|  |   cbrt_noderiv(28) = 3.0365889718756618 | ||
|  | ] | ||
|  | 
 | ||
|  | The result of `bracket_and_solve_root` is a [@http://www.cplusplus.com/reference/utility/pair/ pair] | ||
|  | of values that could be displayed. | ||
|  | 
 | ||
|  | The number of bits separating them can be found using `float_distance(r.first, r.second)`. | ||
|  | The distance is zero (closest representable) for 3[super 3] = 27 | ||
|  | but `float_distance(r.first, r.second) = 3` for cube root of 28 with this function. | ||
|  | The result (avoiding overflow) is midway between these two values. | ||
|  | 
 | ||
|  | [h3:cbrt_1st_derivative Cube root function with 1st derivative (slope)] | ||
|  | 
 | ||
|  | We now solve the same problem, but using more information about the function, | ||
|  | to show how this can speed up finding the best estimate of the root. | ||
|  | 
 | ||
|  | For the root function, the 1st differential (the slope of the tangent to a curve at any point) is known. | ||
|  | 
 | ||
|  | This algorithm is similar to this [@http://en.wikipedia.org/wiki/Nth_root_algorithm nth root algorithm]. | ||
|  | 
 | ||
|  | If you need some reminders, then | ||
|  | [@http://en.wikipedia.org/wiki/Derivative#Derivatives_of_elementary_functions derivatives of elementary functions] | ||
|  | may help. | ||
|  | 
 | ||
|  | Using the rule that the derivative of ['x[super n]] for positive n (actually all nonzero n) is ['n x[super n-1]], | ||
|  | allows us to get the 1st differential as ['3x[super 2]]. | ||
|  | 
 | ||
|  | To see how this extra information is used to find a root, view | ||
|  | [@http://en.wikipedia.org/wiki/Newton%27s_method Newton-Raphson iterations] | ||
|  | and the [@http://en.wikipedia.org/wiki/Newton%27s_method#mediaviewer/File:NewtonIteration_Ani.gif animation]. | ||
|  | 
 | ||
|  | We define a better functor `cbrt_functor_deriv` that returns | ||
|  | both the evaluation of the function to solve, along with its first derivative: | ||
|  | 
 | ||
|  | To '['return]' two values, we use a [@http://en.cppreference.com/w/cpp/utility/pair std::pair] | ||
|  | of floating-point values. | ||
|  | 
 | ||
|  | [root_finding_1_deriv_1] | ||
|  | 
 | ||
|  | The result of [@boost:/libs/math/include/boost/math/tools/roots.hpp `newton_raphson_iterate`] | ||
|  | function is a single value. | ||
|  | 
 | ||
|  | [tip There is a compromise between accuracy and speed when chosing the value of `digits`. | ||
|  | It is tempting to simply chose `std::numeric_limits<T>::digits`, | ||
|  | but this may mean some inefficient and unnecessary iterations as the function thrashes around | ||
|  | trying to locate the last bit.  In theory, since the precision doubles with each step | ||
|  | it is sufficient to stop when half the bits are correct: as the last step will have doubled | ||
|  | that to full precision.  Of course the function has no way to tell if that is actually the case | ||
|  | unless it does one more step to be sure.  In practice setting the precision to slightly more | ||
|  | than `std::numeric_limits<T>::digits / 2` is a good choice.] | ||
|  | 
 | ||
|  | Note that it is up to the caller of the function to check the iteration count | ||
|  | after the call to see if iteration stoped as a result of running out of iterations | ||
|  | rather than meeting the required precision. | ||
|  | 
 | ||
|  | Using the test data in [@../../test/test_cbrt.cpp  /test/test_cbrt.cpp] this found the cube root | ||
|  | exact to the last digit in every case, and in no more than 6 iterations at double | ||
|  | precision.  However, you will note that a high precision was used in this | ||
|  | example, exactly what was warned against earlier on in these docs!  In this | ||
|  | particular case it is possible to compute ['f(x)] exactly and without undue | ||
|  | cancellation error, so a high limit is not too much of an issue. | ||
|  | 
 | ||
|  | However, reducing the limit to `std::numeric_limits<T>::digits * 2 / 3` gave full | ||
|  | precision in all but one of the test cases (and that one was out by just one bit). | ||
|  | The maximum number of iterations remained 6, but in most cases was reduced by one. | ||
|  | 
 | ||
|  | Note also that the above code omits a probable optimization by computing z[sup2] | ||
|  | and reusing it, omits error handling, and does not handle | ||
|  | negative values of z correctly. (These are left as the customary exercise for the reader!) | ||
|  | 
 | ||
|  | The `boost::math::cbrt` function also includes these and other improvements: | ||
|  | most importantly it uses a much better initial guess which reduces the iteration count to | ||
|  | just 1 in almost all cases. | ||
|  | 
 | ||
|  | [h3:cbrt_2_derivatives Cube root with 1st & 2nd derivative (slope & curvature)] | ||
|  | 
 | ||
|  | Next we define yet another even better functor `cbrt_functor_2deriv` that returns | ||
|  | both the evaluation of the function to solve, | ||
|  | along with its first [*and second] derivative: | ||
|  | 
 | ||
|  | __spaces['f''(x) = 6x] | ||
|  | 
 | ||
|  | using information about both slope and curvature to speed convergence. | ||
|  | 
 | ||
|  | To [''return'] three values, we use a `tuple` of three floating-point values: | ||
|  | [root_finding_2deriv_1] | ||
|  | 
 | ||
|  | The function `halley_iterate` also returns a single value, | ||
|  | and the number of iterations will reveal if it met the convergence criterion set by `get_digits`. | ||
|  | 
 | ||
|  | The no-derivative method gives a result of | ||
|  | 
 | ||
|  |   cbrt_noderiv(28) = 3.0365889718756618 | ||
|  | 
 | ||
|  | with a 3 bits distance between the bracketed values, whereas the derivative methods both converge to a single value | ||
|  | 
 | ||
|  |   cbrt_2deriv(28) = 3.0365889718756627 | ||
|  | 
 | ||
|  | which we can compare with the [@boost:/libs/math/doc/html/math_toolkit/powers/cbrt.html boost::math::cbrt] | ||
|  | 
 | ||
|  |   cbrt(28)              = 3.0365889718756627 | ||
|  | 
 | ||
|  | Note that the iterations are set to stop at just one-half of full precision, | ||
|  | and yet, even so, not one of the test cases had a single bit wrong. | ||
|  | What's more, the maximum number of iterations was now just 4. | ||
|  | 
 | ||
|  | Just to complete the picture, we could have called | ||
|  | [link math_toolkit.roots.roots_deriv.schroder `schroder_iterate`] in the last | ||
|  | example: and in fact it makes no difference to the accuracy or number of iterations | ||
|  | in this particular case.  However, the relative performance of these two methods | ||
|  | may vary depending upon the nature of ['f(x)], and the accuracy to which the initial | ||
|  | guess can be computed.  There appear to be no generalisations that can be made | ||
|  | except "try them and see". | ||
|  | 
 | ||
|  | Finally, had we called `cbrt` with [@http://shoup.net/ntl/doc/RR.txt NTL::RR] | ||
|  | set to 1000 bit precision (about 300 decimal digits), | ||
|  | then full precision can be obtained with just 7 iterations. | ||
|  | To put that in perspective, | ||
|  | an increase in precision by a factor of 20, has less than doubled the number of | ||
|  | iterations.  That just goes to emphasise that most of the iterations are used | ||
|  | up getting the first few digits correct: after that these methods can churn out | ||
|  | further digits with remarkable efficiency. | ||
|  | 
 | ||
|  | Or to put it another way: ['nothing beats a really good initial guess!] | ||
|  | 
 | ||
|  | Full code of this example is at | ||
|  | [@../../example/root_finding_example.cpp  root_finding_example.cpp], | ||
|  | 
 | ||
|  | [endsect] | ||
|  | 
 | ||
|  | [section:lambda Using C++11 Lambda's] | ||
|  | 
 | ||
|  | Since all the root finding functions accept a function-object, they can be made to | ||
|  | work (often in a lot less code) with C++11 lambda's.  Here's the much reduced code for our "toy" cube root function: | ||
|  | 
 | ||
|  | [root_finding_2deriv_lambda] | ||
|  | 
 | ||
|  | Full code of this example is at | ||
|  | [@../../example/root_finding_example.cpp  root_finding_example.cpp], | ||
|  | 
 | ||
|  | [endsect] | ||
|  | 
 | ||
|  | [section:5th_root_eg Computing the Fifth Root] | ||
|  | 
 | ||
|  | Let's now suppose we want to find the [*fifth root] of a number ['a]. | ||
|  | 
 | ||
|  | The equation we want to solve is : | ||
|  | 
 | ||
|  | __spaces['f](x) = ['x[super 5] -a] | ||
|  | 
 | ||
|  | If your differentiation is a little rusty | ||
|  | (or you are faced with an function whose complexity makes differentiation daunting), | ||
|  | then you can get help, for example, from the invaluable | ||
|  | [@http://www.wolframalpha.com/ WolframAlpha site.] | ||
|  | 
 | ||
|  | For example, entering the commmand: `differentiate x ^ 5` | ||
|  | 
 | ||
|  | or the Wolfram Language command: ` D[x ^ 5, x]` | ||
|  | 
 | ||
|  | gives the output: `d/dx(x ^ 5) = 5 x ^ 4` | ||
|  | 
 | ||
|  | and to get the second differential, enter: `second differentiate x ^ 5` | ||
|  | 
 | ||
|  | or the Wolfram Language command: `D[x ^ 5, { x, 2 }]` | ||
|  | 
 | ||
|  | to get the output: `d ^ 2 / dx ^ 2(x ^ 5) = 20 x ^ 3` | ||
|  | 
 | ||
|  | To get a reference value, we can enter: [^fifth root 3126] | ||
|  | 
 | ||
|  | or: `N[3126 ^ (1 / 5), 50]` | ||
|  | 
 | ||
|  | to get a result with a precision of 50 decimal digits: | ||
|  | 
 | ||
|  | 5.0003199590478625588206333405631053401128722314376 | ||
|  | 
 | ||
|  | (We could also get a reference value using __multiprecision_root). | ||
|  | 
 | ||
|  | The 1st and 2nd derivatives of x[super 5] are: | ||
|  | 
 | ||
|  | __spaces['f]\'(x) = 5x[super 4] | ||
|  | 
 | ||
|  | __spaces['f]\'\'(x) = 20x[super 3] | ||
|  | 
 | ||
|  | [root_finding_fifth_functor_2deriv] | ||
|  | [root_finding_fifth_2deriv] | ||
|  | 
 | ||
|  | Full code of this example is at | ||
|  | [@../../example/root_finding_example.cpp  root_finding_example.cpp] and | ||
|  | [@../../example/root_finding_n_example.cpp  root_finding_n_example.cpp]. | ||
|  | 
 | ||
|  | [endsect] | ||
|  | 
 | ||
|  | [section:multiprecision_root  Root-finding using Boost.Multiprecision] | ||
|  | 
 | ||
|  | The apocryphally astute reader might, by now, be asking "How do we know if this computes the 'right' answer?". | ||
|  | 
 | ||
|  | For most values, there is, sadly, no 'right' answer. | ||
|  | This is because values can only rarely be ['exactly represented] by C++ floating-point types. | ||
|  | What we do want is the 'best' representation - one that is the nearest __representable value. | ||
|  | (For more about how numbers are represented see __floating_point). | ||
|  | 
 | ||
|  | Of course, we might start with finding an external reference source like | ||
|  | __WolframAlpha, as above, but this is not always possible. | ||
|  | 
 | ||
|  | Another way to reassure is to compute 'reference' values at higher precision | ||
|  | with which to compare the results of our iterative computations using built-in like `double`. | ||
|  | They should agree within the tolerance that was set. | ||
|  | 
 | ||
|  | The result  of `static_cast`ing to `double` from a higher-precision type like `cpp_bin_float_50` is guaranteed | ||
|  | to be the [*nearest representable] `double` value. | ||
|  | 
 | ||
|  | For example, the cube root functions in our example for `cbrt(28.)` compute | ||
|  | 
 | ||
|  | `std::cbrt<double>(28.)  = 3.0365889718756627` | ||
|  | 
 | ||
|  | WolframAlpha says  `3.036588971875662519420809578505669635581453977248111123242141...` | ||
|  | 
 | ||
|  | `static_cast<double>(3.03658897187566251942080957850) = 3.0365889718756627` | ||
|  | 
 | ||
|  | This example `cbrt(28.)   =  3.0365889718756627` | ||
|  | 
 | ||
|  | [tip To ensure that all potentially significant decimal digits are displayed use `std::numeric_limits<T>::max_digits10` | ||
|  | (or if not available on older platforms or compilers use `2+std::numeric_limits<double>::digits*3010/10000`).[br] | ||
|  | 
 | ||
|  | Ideally, values should agree to `std::numeric-limits<T>::digits10` decimal digits. | ||
|  | 
 | ||
|  | This also means that a 'reference' value to be [*input] or `static_cast` should have | ||
|  | at least `max_digits10` decimal digits (17 for 64-bit `double`). | ||
|  | ] | ||
|  | 
 | ||
|  | If we wish to compute [*higher-precision values] then, on some platforms, we may be able to use `long double` | ||
|  | with a higher precision than `double` to compare with the very common `double` | ||
|  | and/or a more efficient built-in quad floating-point type like `__float128`. | ||
|  | 
 | ||
|  | Almost all platforms can easily use __multiprecision, | ||
|  | for example, __cpp_dec_float or a binary type __cpp_bin_float types, | ||
|  | to compute values at very much higher precision. | ||
|  | 
 | ||
|  | [note With multiprecision types, it is debatable whether to use the type `T` for computing the initial guesses. | ||
|  | Type `double` is like to be accurate enough for the method used in these examples. | ||
|  | This would limit the exponent range of possible values to that of `double`. | ||
|  | There is also the cost of conversion to and from type `T` to consider. | ||
|  | In these examples, `double` is used via `typedef double guess_type`.] | ||
|  | 
 | ||
|  | Since the functors and functions used above are templated on the value type, | ||
|  | we can very simply use them with any of the __multiprecision types.  As a reminder, | ||
|  | here's our toy cube root function using 2 derivatives and C++11 lambda functions to find the root: | ||
|  | 
 | ||
|  | [root_finding_2deriv_lambda] | ||
|  | 
 | ||
|  | Some examples below are 50 decimal digit decimal and binary types | ||
|  | (and on some platforms a much faster `float128` or `quad_float` type ) | ||
|  | that we can use with these includes: | ||
|  | 
 | ||
|  | [root_finding_multiprecision_include_1] | ||
|  | 
 | ||
|  | Some using statements simplify their use: | ||
|  | 
 | ||
|  | [root_finding_multiprecision_example_1] | ||
|  | 
 | ||
|  | They can be used thus: | ||
|  | 
 | ||
|  | [root_finding_multiprecision_example_2] | ||
|  | 
 | ||
|  | A reference value computed by __WolframAlpha is | ||
|  | 
 | ||
|  |    N[2^(1/3), 50]  1.2599210498948731647672106072782283505702514647015 | ||
|  | 
 | ||
|  | which agrees exactly. | ||
|  | 
 | ||
|  | To [*show] values to their full precision, it is necessary to adjust the `std::ostream` `precision` to suit the type, for example: | ||
|  | 
 | ||
|  | [root_finding_multiprecision_show_1] | ||
|  | 
 | ||
|  | [root_finding_multiprecision_example_3] | ||
|  | 
 | ||
|  | which outputs: | ||
|  | 
 | ||
|  | [pre | ||
|  | cbrt(2) = 1.2599210498948731647672106072782283505702514647015 | ||
|  | 
 | ||
|  | value = 2, cube root =1.25992104989487 | ||
|  | value = 2, cube root =1.25992104989487 | ||
|  | value = 2, cube root =1.2599210498948731647672106072782283505702514647015 | ||
|  | ] | ||
|  | 
 | ||
|  | [tip Be [*very careful] about the floating-point type `T` that is passed to the root-finding function. | ||
|  | Carelessly passing a integer by writing | ||
|  | `cpp_dec_float_50  r = cbrt_2deriv(2);` or `show_cube_root(2);` | ||
|  | will provoke many warnings and compile errors. | ||
|  | 
 | ||
|  | Even `show_cube_root(2.F);` will produce warnings because `typedef double guess_type` defines the type | ||
|  | used to compute the guess and bracket values as `double`. | ||
|  | 
 | ||
|  | Even more treacherous is passing a `double` as in `cpp_dec_float_50  r = cbrt_2deriv(2.);` | ||
|  | which silently gives the 'wrong' result, computing a `double` result and [*then] converting to `cpp_dec_float_50`! | ||
|  | All digits beyond `max_digits10` will be incorrect. | ||
|  | Making the `cbrt` type explicit with `cbrt_2deriv<cpp_dec_float_50>(2.);` will give you the desired 50 decimal digit precision result. | ||
|  | ] [/tip] | ||
|  | 
 | ||
|  | Full code of this example is at | ||
|  | [@../../example/root_finding_multiprecision_example.cpp  root_finding_multiprecision_example.cpp]. | ||
|  | 
 | ||
|  | [endsect] | ||
|  | 
 | ||
|  | [section:nth_root Generalizing to Compute the nth root] | ||
|  | 
 | ||
|  | If desired, we can now further generalize to compute the ['n]th root by computing the derivatives [*at compile-time] | ||
|  | using the rules for differentiation and `boost::math::pow<N>` | ||
|  | where template parameter `N` is an integer and a compile time constant.  Our functor and function now have an additional template parameter `N`, | ||
|  | for the root required. | ||
|  | 
 | ||
|  | [note Since the powers and derivatives are fixed at compile time, the resulting code is as efficient as as if hand-coded as the cube and fifth-root examples above. | ||
|  | A good compiler should also optimise any repeated multiplications.] | ||
|  | 
 | ||
|  | Our ['n]th root functor is | ||
|  | 
 | ||
|  | [root_finding_nth_functor_2deriv] | ||
|  | 
 | ||
|  | and our ['n]th root  function is | ||
|  | 
 | ||
|  | [root_finding_nth_function_2deriv] | ||
|  | 
 | ||
|  | [root_finding_n_example_2] | ||
|  | 
 | ||
|  | produces an output similar to this | ||
|  | 
 | ||
|  | [root_finding_example_output_1] | ||
|  | 
 | ||
|  | [tip Take care with the type passed to the function. It is best to pass a `double` or greater-precision floating-point type. | ||
|  | 
 | ||
|  | Passing an integer value, for example, `nth_2deriv<5>(2)` will be rejected, while `nth_2deriv<5, double>(2)` converts the integer to `double`. | ||
|  | 
 | ||
|  | Avoid passing a `float` value that will provoke warnings (actually spurious) from the compiler about potential loss of data, | ||
|  | as noted above.] | ||
|  | 
 | ||
|  | [warning Asking for unreasonable roots, for example, `show_nth_root<1000000>(2.);` may lead to | ||
|  | [@http://en.wikipedia.org/wiki/Loss_of_significance Loss of significance] like | ||
|  | `Type double value = 2, 1000000th root = 1.00000069314783`. | ||
|  | Use of the the `pow` function is more sensible for this unusual need. | ||
|  | ] | ||
|  | 
 | ||
|  | Full code of this example is at | ||
|  | [@../../example/root_finding_n_example.cpp  root_finding_n_example.cpp]. | ||
|  | 
 | ||
|  | [endsect] | ||
|  | 
 | ||
|  | [section:elliptic_eg A More complex example - Inverting the Elliptic Integrals] | ||
|  | 
 | ||
|  | The arc length of an ellipse with radii ['a] and ['b] is given by: | ||
|  | 
 | ||
|  | [pre L(a, b) = 4aE(k)] | ||
|  | 
 | ||
|  | with: | ||
|  | 
 | ||
|  | [pre k = [sqrt](1 - b[super 2]/a[super 2])] | ||
|  | 
 | ||
|  | where ['E(k)] is the complete elliptic integral of the second kind - see __ellint_2. | ||
|  | 
 | ||
|  | Let's suppose we know the arc length and one radii, we can then calculate the other | ||
|  | radius by inverting the formula above.  We'll begin by encoding the above formula | ||
|  | into a functor that our root-finding algorithms can call. | ||
|  | 
 | ||
|  | Note that while not | ||
|  | completely obvious from the formula above, the function is completely symmetrical | ||
|  | in the two radii - which can be interchanged at will - in this case we need to | ||
|  | make sure that `a >= b` so that we don't accidentally take the square root of a negative number: | ||
|  | 
 | ||
|  | [import ../../example/root_elliptic_finding.cpp] | ||
|  | 
 | ||
|  | [elliptic_noderv_func] | ||
|  | 
 | ||
|  | We'll also need a decent estimate to start searching from, the approximation: | ||
|  | 
 | ||
|  | [pre L(a, b) [approx] 4[sqrt](a[super 2] + b[super 2])] | ||
|  | 
 | ||
|  | Is easily inverted to give us what we need, which using derivative-free root | ||
|  | finding leads to the algorithm: | ||
|  | 
 | ||
|  | [elliptic_root_noderiv] | ||
|  | 
 | ||
|  | This function generally finds the root within 8-10 iterations, so given that the runtime | ||
|  | is completely dominated by the cost of calling the ellliptic integral it would be nice to | ||
|  | reduce that count somewhat. We'll try to do that by using a derivative-based method; | ||
|  | the derivatives of this function are rather hard to work out by hand, but fortunately | ||
|  | [@http://www.wolframalpha.com/input/?i=d%2Fda+\[4+*+a+*+EllipticE%281+-+b^2%2Fa^2%29\] | ||
|  | Wolfram Alpha] can do the grunt work for us to give: | ||
|  | 
 | ||
|  | [pre d/da L(a, b) = 4(a[super 2]E(k) - b[super 2]K(k)) / (a[super 2] - b[super 2])] | ||
|  | 
 | ||
|  | Note that now we have [*two] elliptic integral calls to get the derivative, so our | ||
|  | functor will be at least twice as expensive to call as the derivative-free one above: | ||
|  | we'll have to reduce the iteration count quite substantially to make a difference! | ||
|  | 
 | ||
|  | Here's the revised functor: | ||
|  | 
 | ||
|  | [elliptic_1deriv_func] | ||
|  | 
 | ||
|  | The root-finding code is now almost the same as before, but we'll make use of | ||
|  | Newton-iteration to get the result: | ||
|  | 
 | ||
|  | [elliptic_1deriv] | ||
|  | 
 | ||
|  | The number of iterations required for `double` precision is now usually around 4 - | ||
|  | so we've slightly more than halved the number of iterations, but made the | ||
|  | functor twice as expensive to call! | ||
|  | 
 | ||
|  | Interestingly though, the second derivative requires no more expensive | ||
|  | elliptic integral calls than the first does, in other words it comes | ||
|  | essentially "for free", in which case we might as well make use of it | ||
|  | and use Halley-iteration.  This is quite a typical situation when | ||
|  | inverting special-functions.  Here's the revised functor: | ||
|  | 
 | ||
|  | [elliptic_2deriv_func] | ||
|  | 
 | ||
|  | The actual root-finding code is almost the same as before, except we can | ||
|  | use Halley, rather than Newton iteration: | ||
|  | 
 | ||
|  | [elliptic_2deriv] | ||
|  | 
 | ||
|  | While this function uses only slightly fewer iterations (typically around 3) | ||
|  | to find the root, compared to the original derivative-free method, we've moved from | ||
|  | 8-10 elliptic integral calls to 6. | ||
|  | 
 | ||
|  | Full code of this example is at | ||
|  | [@../../example/root_elliptic_finding.cpp  root_elliptic_finding.cpp]. | ||
|  | 
 | ||
|  | [endsect] | ||
|  | 
 | ||
|  | 
 | ||
|  | [endsect] [/section:root_examples Examples of Root Finding (with and without derivatives)] | ||
|  | 
 | ||
|  | [section:bad_guess The Effect of a Poor Initial Guess] | ||
|  | 
 | ||
|  | It's instructive to take our "toy" example algorithms, and use deliberately bad initial guesses to see how the | ||
|  | various root finding algorithms fair.  We'll start with the cubed root, and using the cube root of 500 as the test case: | ||
|  | 
 | ||
|  | [table | ||
|  | [[Initial Guess=][-500% ([approx]1.323)][-100% ([approx]3.97)][-50% ([approx]3.96)][-20% ([approx]6.35)][-10% ([approx]7.14)][-5% ([approx]7.54)][5% ([approx]8.33)][10% ([approx]8.73)][20% ([approx]9.52)][50% ([approx]11.91)][100% ([approx]15.87)][500 ([approx]47.6)]] | ||
|  | [[bracket_and_solve_root][12][8][8][10][11][11][11][11][11][11][7][13]] | ||
|  | [[newton_iterate][12][7][7][5][5][4][4][5][5][6][7][9]] | ||
|  | [[halley_iterate][7][4][4][3][3][3][3][3][3][4][4][6]] | ||
|  | [[schroder_iterate][11][6][6][4][3][3][3][3][4][5][5][8]] | ||
|  | ] | ||
|  | 
 | ||
|  | As you can see `bracket_and_solve_root` is relatively insensitive to starting location - as long as you don't start many orders of magnitude away from the root it will | ||
|  | take roughly the same number of steps to bracket the root and solve it.  On the other hand the derivative-based methods are slow to start, but once they have some digits | ||
|  | correct they increase precision exceptionally fast: they are therefore quite sensitive to the initial starting location. | ||
|  | 
 | ||
|  | The next table shows the number of iterations required to find the second radius of an ellipse with first radius 50 and arc-length 500: | ||
|  | 
 | ||
|  | [table | ||
|  | [[Initial Guess=][-500% ([approx]20.6)][-100% ([approx]61.81)][-50% ([approx]61.81)][-20% ([approx]98.9)][-10% ([approx]111.3)][-5% ([approx]117.4)][5% ([approx]129.8)][10% ([approx]136)][20% ([approx]148.3)][50% ([approx]185.4)][100% ([approx]247.2)][500 ([approx]741.7)]] | ||
|  | [[bracket_and_solve_root][11][5][5][8][8][7][7][8][9][8][6][10]] | ||
|  | [[newton_iterate][4][4][4][3][3][3][3][3][3][4][4][4]] | ||
|  | [[halley_iterate][4][3][3][3][3][2][2][3][3][3][3][3]] | ||
|  | [[schroder_iterate][4][3][3][3][3][2][2][3][3][3][3][3]] | ||
|  | ] | ||
|  | 
 | ||
|  | Interestingly this function is much more resistant to a poor initial guess when using derivatives. | ||
|  | 
 | ||
|  | [endsect] | ||
|  | 
 | ||
|  | [section:bad_roots Examples Where Root Finding Goes Wrong] | ||
|  | 
 | ||
|  | There are many reasons why root root finding can fail, here are just a few of the more common examples: | ||
|  | 
 | ||
|  | [h3 Local Minima] | ||
|  | 
 | ||
|  | If you start in the wrong place, such as z[sub 0] here: | ||
|  | 
 | ||
|  | [$../roots/bad_root_1.svg] | ||
|  | 
 | ||
|  | Then almost any root-finding algorithm will descend into a local minima rather than find the root. | ||
|  | 
 | ||
|  | [h3 Flatlining] | ||
|  | 
 | ||
|  | In this example, we're starting from a location (z[sub 0]) where the first derivative is essentially zero: | ||
|  | 
 | ||
|  | [$../roots/bad_root_2.svg] | ||
|  | 
 | ||
|  | In this situation the next iteration will shoot off to infinity (assuming we're using derivatives that is).  Our | ||
|  | code guards against this by insisting that the root is always bracketed, and then never stepping outside those bounds. | ||
|  | In a case like this, no root finding algorithm can do better than bisecting until the root is found. | ||
|  | 
 | ||
|  | Note that there is no scale on the graph, we have seen examples of this situation occur in practice ['even when | ||
|  | several decimal places of the initial guess z[sub 0] are correct.] | ||
|  | 
 | ||
|  | This is really a special case of a more common situation where root finding with derivatives is ['divergent].  Consider | ||
|  | starting at z[sub 0] in this case: | ||
|  | 
 | ||
|  | [$../roots/bad_root_4.svg] | ||
|  | 
 | ||
|  | An initial Newton step would take you further from the root than you started, as will all subsequent steps. | ||
|  | 
 | ||
|  | [h3 Micro-stepping / Non-convergence] | ||
|  | 
 | ||
|  | Consider starting at z[sub 0] in this situation: | ||
|  | 
 | ||
|  | [$../roots/bad_root_3.svg] | ||
|  | 
 | ||
|  | The first derivative is essentially infinite, and the second close to zero (and so offers no correction if we use it), | ||
|  | as a result we take a very small first step.  In the worst case situation, the first step is so small | ||
|  | - perhaps even so small that subtracting from z[sub 0] has no effect at the current working precision - that our algorithm | ||
|  | will assume we are at the root already and terminate.  Otherwise we will take lot's of very small steps which never converge | ||
|  | on the root: our algorithms will protect against that by reverting to bisection. | ||
|  | 
 | ||
|  | An example of this situation would be trying to find the root of e[super -1/z[super 2]] - this function has a single | ||
|  | root at ['z = 0], but for ['z[sub 0] < 0] neither Newton nor Halley steps will ever converge on the root, and for ['z[sub 0] > 0] | ||
|  | the steps are actually divergent. | ||
|  | 
 | ||
|  | [endsect] | ||
|  | 
 | ||
|  | [/ | ||
|  |   Copyright 2015 John Maddock and Paul A. Bristow. | ||
|  |   Distributed under the Boost Software License, Version 1.0. | ||
|  |   (See accompanying file LICENSE_1_0.txt or copy at | ||
|  |   http://www.boost.org/LICENSE_1_0.txt). | ||
|  | ] | ||
|  | 
 |