c++: strategies for stability of floating point arithmetic - c++

Can anyone recommend any C++ libraries/routines/packages that contain strategies for maintaining the stability of various floating point operations?
Example: suppose you would like to sum across a vector/array of one million long double in the unit interval (0,1), and that each number is of about the same order of magnitude. Naively summing for (int i=0;i<1000000;++i) sum += array[i]; is unreliable - for large enough i, sum will be of a much larger order of magnitude than array[i], and so sum += array[i] would be equivalent to sum += 0.00.
(Note: the solution to this example is a binary summing strategy.)
I deal with sums and products of thousands/millions of miniscule probabilities. I am using the arbitrary-precision library MPFRC++ with a 2048 bit significand, but the same concerns still apply.
I am chiefly concerned with:
Strategies for accurately summing many numbers (e.g. above Example).
When is multiplication and division potentially unstable? (If I want to normalize a large array of numbers, what should my normalization constant be? The smallest value? The largest? A median?)

Binary summation doesn't guarantee accurate result. The most reliable (albeit slower) method is to use Kahan summation. Boost.Accumulators has an implementation of the above and much more.
Multiplication and division stability: unless you get to denormalized floats they don't suffer from the same problems as summation and substraction. In fact the error of multiplication is at most 0.5 ulp (units last place).
... what should my normalization constant be?
What do you mean by 'normalize'? It depends on the norm you use. Possible candidates: use the maximum absolute value in the array, or any other generalized mean. (Other choices you listed do not work since they may be zero even for non-zero array.)

Related

What is the logic behind dividing arrays of double precision matrix to a large number before sending that matrix to a solver?

I am given a code which divides the members of [A] in [A]{X}=[B] to 10^4 when assembling matrix [A].
Then it also divides the results array to this number to calculate correct value.
I cannot understand why this should be done? Does double precision has limitation on numbers of integer it can hold; so the maker of this code wanted to increase number of float digits? Or maybe he misunderstood the conception of double precision?
Dividing by some constant will not improve the accuracy of the solution as floating point numbers have an exponent representation. Thus scaling by some large number will to a large degree affect the exponent and to a minor degree the significant bits.
The numerically hard part when solving is adding floating point numbers of different orders of magnitude, as this will affect the number of significant bits. However, this does not get affected by a global scaling as the difference of exponents remains the same. Similarly, division, multiplication and square root does not get affected.
What does affect the solution is using different scaling factors,e.g. enforcing the diagonal of a symmetric A to be one. However, this mostly helps with iterative solvers, while cholesky decomposition and similar will only be affected to a minor degree.
Without seeing the code it is hard to say.
Sometimes scaling was done for IO reasons rather than computational reasons.
If the '*1E-4' is single precision then it is likely done to discretize the incoming data, and the computations can be in double precision following that.
Obviously seeing code can help.

What algorithm should I use for high-performance large integer division?

I am encoding large integers into an array of size_t. I already have the other operations working (add, subtract, multiply); as well as division by a single digit. But I would like match the time complexity of my multiplication algorithms if possible (currently Toom-Cook).
I gather there are linear time algorithms for taking various notions of multiplicative inverse of my dividend. This means I could theoretically achieve division in the same time complexity as my multiplication, because the linear-time operation is "insignificant" by comparison anyway.
My question is, how do I actually do that? What type of multiplicative inverse is best in practice? Modulo 64^digitcount? When I multiply the multiplicative inverse by my divisor, can I shirk computing the part of the data that would be thrown away due to integer truncation? Can anyone provide C or C++ pseudocode or give a precise explanation of how this should be done?
Or is there a dedicated division algorithm that is even better than the inverse-based approach?
Edit: I dug up where I was getting "inverse" approach mentioned above. On page 312 of "Art of Computer Programming, Volume 2: Seminumerical Algorithms", Knuth provides "Algorithm R" which is a high-precision reciprocal. He says its time complexity is less than that of multiplication. It is, however, nontrivial to convert it to C and test it out, and unclear how much overhead memory, etc, will be consumed until I code this up, which would take a while. I'll post it if no one beats me to it.
The GMP library is usually a good reference for good algorithms. Their documented algorithms for division mainly depend on choosing a very large base, so that you're dividing a 4 digit number by a 2 digit number, and then proceed via long division.
Long division will require computing 2 digit by 1 digit quotients; this can either be done recursively, or by precomputing an inverse and estimating the quotient as you would with Barrett reduction.
When dividing a 2n-bit number by an n-bit number, the recursive version costs O(M(n) log(n)), where M(n) is the cost of multiplying n-bit numbers.
The version using Barrett reduction will cost O(M(n)) if you use Newton's algorithm to compute the inverse, but according to GMP's documentation, the hidden constant is a lot larger, so this method is only preferable for very large divisions.
In more detail, the core algorithm behind most division algorithms is an "estimated quotient with reduction" calculation, computing (q,r) so that
x = qy + r
but without the restriction that 0 <= r < y. The typical loop is
Estimate the quotient q of x/y
Compute the corresponding reduction r = x - qy
Optionally adjust the quotient so that the reduction r is in some desired interval
If r is too big, then repeat with r in place of x.
The quotient of x/y will be the sum of all the qs produced, and the final value of r will be the true remainder.
Schoolbook long division, for example, is of this form. e.g. step 3 covers those cases where the digit you guessed was too big or too small, and you adjust it to get the right value.
The divide and conquer approach estimates the quotient of x/y by computing x'/y' where x' and y' are the leading digits of x and y. There is a lot of room for optimization by adjusting their sizes, but IIRC you get best results if x' is twice as many digits of y'.
The multiply-by-inverse approach is, IMO, the simplest if you stick to integer arithmetic. The basic method is
Estimate the inverse of y with m = floor(2^k / y)
Estimate x/y with q = 2^(i+j-k) floor(floor(x / 2^i) m / 2^j)
In fact, practical implementations can tolerate additional error in m if it means you can use a faster reciprocal implementation.
The error is a pain to analyze, but if I recall the way to do it, you want to choose i and j so that x ~ 2^(i+j) due to how errors accumulate, and you want to choose x / 2^i ~ m^2 to minimize the overall work.
The ensuing reduction will have r ~ max(x/m, y), so that gives a rule of thumb for choosing k: you want the size of m to be about the number of bits of quotient you compute per iteration — or equivalently the number of bits you want to remove from x per iteration.
I do not know the multiplicative inverse algorithm but it sounds like modification of Montgomery Reduction or Barrett's Reduction.
I do bigint divisions a bit differently.
See bignum division. Especially take a look at the approximation divider and the 2 links there. One is my fixed point divider and the others are fast multiplication algos (like karatsuba,Schönhage-Strassen on NTT) with measurements, and a link to my very fast NTT implementation for 32bit Base.
I'm not sure if the inverse multiplicant is the way.
It is mostly used for modulo operation where the divider is constant. I'm afraid that for arbitrary divisions the time and operations needed to acquire bigint inverse can be bigger then the standard divisions itself, but as I am not familiar with it I could be wrong.
The most common divider in use I saw in implemetations are Newton–Raphson division which is very similar to approximation divider in the link above.
Approximation/iterative dividers usually use multiplication which define their speed.
For small enough numbers is usually long binary division and 32/64bit digit base division fast enough if not fastest: usually they have small overhead, and let n be the max value processed (not the number of digits!)
Binary division example:
Is O(log32(n).log2(n)) = O(log^2(n)).
It loops through all significant bits. In each iteration you need to compare, sub, add, bitshift. Each of those operations can be done in log32(n), and log2(n) is the number of bits.
Here example of binary division from one of my bigint templates (C++):
template <DWORD N> void uint<N>::div(uint &c,uint &d,uint a,uint b)
{
int i,j,sh;
sh=0; c=DWORD(0); d=1;
sh=a.bits()-b.bits();
if (sh<0) sh=0; else { b<<=sh; d<<=sh; }
for (;;)
{
j=geq(a,b);
if (j)
{
c+=d;
sub(a,a,b);
if (j==2) break;
}
if (!sh) break;
b>>=1; d>>=1; sh--;
}
d=a;
}
N is the number of 32 bit DWORDs used to store a bigint number.
c = a / b
d = a % b
qeq(a,b) is a comparison: a >= b greater or equal (done in log32(n)=N)
It returns 0 for a < b, 1 for a > b, 2 for a == b
sub(c,a,b) is c = a - b
The speed boost is gained from that this does not use multiplication (if you do not count the bit shift)
If you use digit with a big base like 2^32 (ALU blocks), then you can rewrite the whole in polynomial like style using 32bit build in ALU operations.
This is usually even faster then binary long division, the idea is to process each DWORD as a single digit, or recursively divide the used arithmetic by half until hit the CPU capabilities.
See division by half-bitwidth arithmetics
On top of all that while computing with bignums
If you have optimized basic operations, then the complexity can lower even further as sub-results get smaller with iterations (changing the complexity of basic operations) A nice example of that are NTT based multiplications.
The overhead can mess thing up.
Due to this the runtime sometimes does not copy the big O complexity, so you should always measure the tresholds and use faster approach for used bit-count to get the max performance and optimize what you can.

IEEE-754 floating point: Divide first or multiply first for best precision?

What's better if I want to preserve as much precision as possible in a calculation with IEEE-754 floating point values:
a = b * c / d
or
a = b / d * c
Is there a difference? If there is, does it depend on the magnitudes of the input values? And, if magnitude matters, how is the best ordering determined when general magnitudes of the values are known?
It depends on the magnitude of the values. Obviously if one divides by zero, all bets are off, but if a multiplication or division results in a denormal subsequent operations can lose precision.
You may find it useful to study Goldberg's seminal paper What Every Computer Scientist Should Know About Floating-Point Arithmetic which will explain things far better than any answer you're likely to receive here. (Goldberg was one of the original authors of IEEE-754.)
Assuming that none of the operations would yield an overflow or an underflow, and your input values have uniformly distributed significands, then this is equivalent. Well, I suppose that to have a rigorous proof, one should do an exhaustive test (probably not possible in practice for double precision since there are 2^156 inputs), but if there is a difference in the average error, then it is tiny. I could try in low precisions with Sipe.
In any case, in the absence of overflow/underflow, only the exact values of the significands matter, not the exponents.
However if the result a is added to (or subtracted from) another expression and not reused, then starting with the division may be more interesting since you can group the multiplication with the following addition by using a FMA (thus with a single rounding).

Floats vs rationals in arbitrary precision fractional arithmetic (C/C++)

Since there are two ways of implementing an AP fractional number, one is to emulate the storage and behavior of the double data type, only with more bytes, and the other is to use an existing integer APA implementation for representing a fractional number as a rational i.e. as a pair of integers, numerator and denominator, which of the two ways are more likely to deliver efficient arithmetic in terms of performance? (Memory usage is really of minor concern.)
I'm aware of the existing C/C++ libraries, some of which offer fractional APA with "floats" and other with rationals (none of them features fixed-point APA, however) and of course I could benchmark a library that relies on "float" implementation against one that makes use of rational implementation, but the results would largely depend on implementation details of those particular libraries I would have to choose randomly from the nearly ten available ones. So it's more theoretical pros and cons of the two approaches that I'm interested in (or three if take into consideration fixed-point APA).
The question is what you mean by arbitrary precision that you mention in the title. Does it mean "arbitrary, but pre-determined at compile-time and fixed at run-time"? Or does it mean "infinite, i.e. extendable at run-time to represent any rational number"?
In the former case (precision customizable at compile-time, but fixed afterwards) I'd say that one of the most efficient solutions would actually be fixed-point arithmetic (i.e. none of the two you mentioned).
Firstly, fixed-point arithmetic does not require any dedicated library for basic arithmetic operations. It is just a concept overlaid over integer arithmetic. This means that if you really need a lot of digits after the dot, you can take any big-integer library, multiply all your data, say, by 2^64 and you basically immediately get fixed-point arithmetic with 64 binary digits after the dot (at least as long as arithmetic operations are concerned, with some extra adjustments for multiplication and division). This is typically significantly more efficient than floating-point or rational representations.
Note also that in many practical applications multiplication operations are often accompanied by division operations (as in x = y * a / b) that "compensate" for each other, meaning that often it is unnecessary to perform any adjustments for such multiplications and divisions. This also contributes to efficiency of fixed-point arithmetic.
Secondly, fixed-point arithmetic provides uniform precision across the entire range. This is not true for either floating-point or rational representations, which in some applications could be a significant drawback for the latter two approaches (or a benefit, depending on what you need).
So, again, why are you considering floating-point and rational representations only. Is there something that prevents you from considering fixed-point representation?
Since no one else seemed to mention this, rationals and floats represent different sets of numbers. The value 1/3 can be represented precisely with a rational, but not a float. Even an arbitrary precision float would take infinitely many mantissa bits to represent a repeating decimal like 1/3. This is because a float is effectively like a rational but where the denominator is constrained to be a power of 2. An arbitrary precision rational can represent everything that an arbitrary precision float can and more, because the denominator can be any integer instead of just powers of 2. (That is, unless I've horribly misunderstood how arbitrary precision floats are implemented.)
This is in response to your prompt for theoretical pros and cons.
I know you didn't ask about memory usage, but here's a theoretical comparison in case anyone else is interested. Rationals, as mentioned above, specialize in numbers that can be represented simply in fractional notation, like 1/3 or 492113/203233, and floats specialize in numbers that are simple to represent in scientific notation with powers of 2, like 5*2^45 or 91537*2^203233. The amount of ascii typing needed to represent the numbers in their respective human-readable form is proportional to their memory usage.
Please correct me in the comments if I've gotten any of this wrong.
Either way, you'll need multiplication of arbitrary size integers. This will be the dominant factor in your performance since its complexity is worse than O(n*log(n)). Things like aligning operands, and adding or subtracting large integers is O(n), so we'll neglect those.
For simple addition and subtraction, you need no multiplications for floats* and 3 multiplications for rationals. Floats win hands down.
For multiplication, you need one multiplication for floats and 2 multiplications for rational numbers. Floats have the edge.
Division is a little bit more complex, and rationals might win out here, but it's by no means a certainty. I'd say it's a draw.
So overall, IMHO, the fact that addition is at least O(n*log(n)) for rationals and O(n) for floats clearly gives the win to a floating-point representation.
*It is possible that you might need one multiplication to perform addition if your exponent base and your digit base are different. Otherwise, if you use a power of 2 as your base, then aligning the operands takes a bit shift. If you don't use a power of two, then you may also have to do a multiplication by a single digit, which is also an O(n) operation.
You are effectively asking the question: "I need to participate in a race with my chosen animal. Should I choose a turtle or a snail ?".
The first proposal "emulating double" sounds like staggered precision: using an array of doubles of which the sum is the defined number. There is a paper from Douglas M. Priest "Algorithms for Arbitrary Precision Floating Point Arithmetic" which describes how to implement this arithmetic. I implemented this and my experience is very bad: The necessary overhead to make this run drops the performance 100-1000 times !
The other method of using fractionals has severe disadvantages, too: You need to implement gcd and kgv and unfortunately every prime in your numerator or denominator has a good chance to blow up your numbers and kill your performance.
So from my experience they are the worst choices one can made for performance.
I recommend the use of the MPFR library which is one of the fastest AP packages in C and C++.
Rational numbers don't give arbitrary precision, but rather the exact answer. They are, however, more expensive in terms of storage and certain operations with them become costly and some operations are not allowed at all, e.g. taking square roots, since they do not necessarily yield a rational answer.
Personally, I think in your case AP floats would be more appropriate.

Accurate evaluation of 1/1 + 1/2 + ... 1/n row

I need to evaluate the sum of the row: 1/1+1/2+1/3+...+1/n. Considering that in C++ evaluations are not complete accurate, the order of summation plays important role. 1/n+1/(n-1)+...+1/2+1/1 expression gives the more accurate result.
So I need to find out the order of summation, which provides the maximum accuracy.
I don't even know where to begin.
Preferred language of realization is C++.
Sorry for my English, if there are any mistakes.
For large n you'd better use asymptotic formulas, like the ones on http://en.wikipedia.org/wiki/Harmonic_number;
Another way is to use exp-log transformation. Basically:
H_n = 1 + 1/2 + 1/3 + ... + 1/n = log(exp(1 + 1/2 + 1/3 + ... + 1/n)) = log(exp(1) * exp(1/2) * exp(1/3) * ... * exp(1/n)).
Exponents and logarithms can be calculated pretty quickly and accuratelly by your standard library. Using multiplication you should get much more accurate results.
If this is your homework and you are required to use simple addition, you'll better add from the smallest one to the largest one, as others suggested.
The reason for the lack of accuracy is the precision of the float, double, and long double types. They only store so many "decimal" places. So adding a very small value to a large value has no effect, the small term is "lost" in the larger one.
The series you're summing has a "long tail", in the sense that the small terms should add up to a large contribution. But if you sum in descending order, then after a while each new small term will have no effect (even before that, most of its decimal places will be discarded). Once you get to that point you can add a billion more terms, and if you do them one at a time it still has no effect.
I think that summing in ascending order should give best accuracy for this kind of series, although it's possible there are some odd corner cases where errors due to rounding to powers of (1/2) might just so happen to give a closer answer for some addition orders than others. You probably can't really predict this, though.
I don't even know where to begin.
Here: What Every Computer Scientist Should Know About Floating-Point Arithmetic
Actually, if you're doing the summation for large N, adding in order from smallest to largest is not the best way -- you can still get into a situation where the numbers you're adding are too small relative to the sum to produce an accurate result.
Look at the problem this way: You have N summations, regardless of ordering, and you wish to have the least total error. Thus, you should be able to get the least total error by minimizing the error of each summation -- and you minimize the error in a summation by adding values as nearly close to each other as possible. I believe that following that chain of logic gives you a binary tree of partial sums:
Sum[0,i] = value[i]
Sum[1,i/2] = Sum[0,i] + Sum[0,i+1]
Sum[j+1,i/2] = Sum[j,i] + Sum[j,i+1]
and so on until you get to a single answer.
Of course, when N is not a power of two, you'll end up with leftovers at each stage, which you need to carry over into the summations at the next stage.
(The margins of StackOverflow are of course too small to include a proof that this is optimal. In part because I haven't taken the time to prove it. But it does work for any N, however large, as all of the additions are adding values of nearly identical magnitude. Well, all but log(N) of them in the worst not-power-of-2 case, and that's vanishingly small compared to N.)
http://en.wikipedia.org/wiki/Arbitrary-precision_arithmetic
You can find libraries with ready for use implementation for C/C++.
For example http://www.apfloat.org/apfloat/
Unless you use some accurate closed-form representation, a small-to-large ordered summation is likely to be most accurate simple solution (it's not clear to me why a log-exp would help - that's a neat trick, but you're not winning anything with it here, as far as I can tell).
You can further gain precision by realizing that after a while, the sum will become "quantized": Effectively, when you have 2 digits of precision, adding 1.3 to 41 results in 42, not 42.3 - but you achieve almost a precision doubling by maintaining an "error" term. This is called Kahan Summation. You'd compute the error term (42-41-1.3 == -0.3) and correct that in the next addition by adding 0.3 to the next term before you add it in again.
Kahan Summation in addition to a small-to-large ordering is liable to be as accurate as you'll ever need to get. I seriously doubt you'll ever need anything better for the harmonic series - after all, even after 2^45 iterations (crazy many) you'd still only be dealing with a numbers that are at least 1/2^45 large, and a sum that's on the order of 45 (<2^6), for an order of magnitude difference of 51 powers-of-two - i.e. even still representable in a double precision variable if you add in the "wrong" order.
If you go small-to-large, and use Kahan Summation, the sun's probably going to extinguish before today's processors reach a percent of error - and you'll run into other tricky accuracy issues just due to the individual term error on that scale first anyhow (being that a number of the order of 2^53 or larger cannot be represented accurately as a double at all anyhow.)
I'm not sure about the order of summation playing an important role, I havent heard that before. I guess you want to do this in floating point arithmetic so the first thing is to think more inline of (1.0/1.0 + 1.0/2.0+1.0/3.0) - otherwise the compiler will do integer division
to determine order of evaluation, maybe a for loop or brackets?
e.g.
float f = 0.0;
for (int i=n; i>0; --i)
{
f += 1.0/static_cast<float>(i);
}
oh forgot to say, compilers will normally have switches to determine floating point evaluation mode. this is maybe related to what you say on order of summation - in visual C+ these are found in code-generation compile settings, in g++ there're options -float that handle this
actually, the other guy is right - you should do summation in order of smallest component first; so
1/n + 1/(n-1) .. 1/1
this is because the precision of a floating point number is linked to the scale, if you start at 1 you'll have 23 bits of precision relative to 1.0. if you start at a smaller number the precision is relative to the smaller number, so you'll get 23 bits of precision relative to 1xe-200 or whatever. then as the number gets bigger rounding error will occur, but the overall error will be less than the other direction
As all your numbers are rationals, the easiest (and also maybe the fastest, as it will have to do less floating point operations) would be to do the computations with rationals (tuples of 2 integers p,q), and then do just one floating point division at the end.
update to use this technique effectively you will need to use bigints for p & q, as they grow quite fast...
A fast prototype in Lisp, that has built in rationals shows:
(defun sum_harmonic (n acc)
(if (= n 0) acc (sum_harmonic (- n 1) (+ acc (/ 1 n)))))
(sum_harmonic 10 0)
7381/2520
[2.9289682]
(sum_harmonic 100 0)
14466636279520351160221518043104131447711/278881500918849908658135235741249214272
[5.1873775]
(sum_harmonic 1000 0)
53362913282294785045591045624042980409652472280384260097101349248456268889497101
75750609790198503569140908873155046809837844217211788500946430234432656602250210
02784256328520814055449412104425101426727702947747127089179639677796104532246924
26866468888281582071984897105110796873249319155529397017508931564519976085734473
01418328401172441228064907430770373668317005580029365923508858936023528585280816
0759574737836655413175508131522517/712886527466509305316638415571427292066835886
18858930404520019911543240875811114994764441519138715869117178170195752565129802
64067621009251465871004305131072686268143200196609974862745937188343705015434452
52373974529896314567498212823695623282379401106880926231770886197954079124775455
80493264757378299233527517967352480424636380511370343312147817468508784534856780
21888075373249921995672056932029099390891687487672697950931603520000
[7.485471]
So, the next better option could be to mantain the list of floating points and to reduce it summing the two smallest numbers in each step...