Remove an invalid fuzz compiler arg

[gh-bc.git] / manuals / algorithms.md

# Algorithms

This `bc` uses the math algorithms below:

### Addition

This `bc` uses brute force addition, which is linear (`O(n)`) in the number of
digits.

### Subtraction

This `bc` uses brute force subtraction, which is linear (`O(n)`) in the number
of digits.

### Multiplication

This `bc` uses two algorithms: [Karatsuba][1] and brute force.

Karatsuba is used for "large" numbers. ("Large" numbers are defined as any
number with `BC_NUM_KARATSUBA_LEN` digits or larger. `BC_NUM_KARATSUBA_LEN` has
a sane default, but may be configured by the user.) Karatsuba, as implemented in
this `bc`, is superlinear but subpolynomial (bounded by `O(n^log_2(3))`).

Brute force multiplication is used below `BC_NUM_KARATSUBA_LEN` digits. It is
polynomial (`O(n^2)`), but since Karatsuba requires both more intermediate
values (which translate to memory allocations) and a few more additions, there
is a "break even" point in the number of digits where brute force multiplication
is faster than Karatsuba. There is a script (`$ROOT/scripts/karatsuba.py`) that
will find the break even point on a particular machine.

***WARNING: The Karatsuba script requires Python 3.***

### Division

This `bc` uses Algorithm D ([long division][2]). Long division is polynomial
(`O(n^2)`), but unlike Karatsuba, any division "divide and conquer" algorithm
reaches its "break even" point with significantly larger numbers. "Fast"
algorithms become less attractive with division as this operation typically
reduces the problem size.

While the implementation of long division may appear to use the subtractive
chunking method, it only uses subtraction to find a quotient digit. It avoids
unnecessary work by aligning digits prior to performing subtraction and finding
a starting guess for the quotient.

Subtraction was used instead of multiplication for two reasons:

1.      Division and subtraction can share code (one of the less important goals of
        this `bc` is small code).
2.      It minimizes algorithmic complexity.

Using multiplication would make division have the even worse algorithmic
complexity of `O(n^(2*log_2(3)))` (best case) and `O(n^3)` (worst case).

### Power

This `bc` implements [Exponentiation by Squaring][3], which (via Karatsuba) has
a complexity of `O((n*log(n))^log_2(3))` which is favorable to the
`O((n*log(n))^2)` without Karatsuba.

### Square Root

This `bc` implements the fast algorithm [Newton's Method][4] (also known as the
Newton-Raphson Method, or the [Babylonian Method][5]) to perform the square root
operation.

Its complexity is `O(log(n)*n^2)` as it requires one division per iteration, and
it doubles the amount of correct digits per iteration.

### Sine and Cosine (`bc` Math Library Only)

This `bc` uses the series

```
x - x^3/3! + x^5/5! - x^7/7! + ...
```

to calculate `sin(x)` and `cos(x)`. It also uses the relation

```
cos(x) = sin(x + pi/2)
```

to calculate `cos(x)`. It has a complexity of `O(n^3)`.

**Note**: this series has a tendency to *occasionally* produce an error of 1
[ULP][6]. (It is an unfortunate side effect of the algorithm, and there isn't
any way around it; [this article][7] explains why calculating sine and cosine,
and the other transcendental functions below, within less than 1 ULP is nearly
impossible and unnecessary.) Therefore, I recommend that users do their
calculations with the precision (`scale`) set to at least 1 greater than is
needed.

### Exponentiation (`bc` Math Library Only)

This `bc` uses the series

```
1 + x + x^2/2! + x^3/3! + ...
```

to calculate `e^x`. Since this only works when `x` is small, it uses

```
e^x = (e^(x/2))^2
```

to reduce `x`.

It has a complexity of `O(n^3)`.

**Note**: this series can also produce errors of 1 ULP, so I recommend users do
their calculations with the precision (`scale`) set to at least 1 greater than
is needed.

### Natural Logarithm (`bc` Math Library Only)

This `bc` uses the series

```
a + a^3/3 + a^5/5 + ...
```

(where `a` is equal to `(x - 1)/(x + 1)`) to calculate `ln(x)` when `x` is small
and uses the relation

```
ln(x^2) = 2 * ln(x)
```

to sufficiently reduce `x`.

It has a complexity of `O(n^3)`.

**Note**: this series can also produce errors of 1 ULP, so I recommend users do
their calculations with the precision (`scale`) set to at least 1 greater than
is needed.

### Arctangent (`bc` Math Library Only)

This `bc` uses the series

```
x - x^3/3 + x^5/5 - x^7/7 + ...
```

to calculate `atan(x)` for small `x` and the relation

```
atan(x) = atan(c) + atan((x - c)/(1 + x * c))
```

to reduce `x` to small enough. It has a complexity of `O(n^3)`.

**Note**: this series can also produce errors of 1 ULP, so I recommend users do
their calculations with the precision (`scale`) set to at least 1 greater than
is needed.

### Bessel (`bc` Math Library Only)

This `bc` uses the series

```
x^n/(2^n * n!) * (1 - x^2 * 2 * 1! * (n + 1)) + x^4/(2^4 * 2! * (n + 1) * (n + 2)) - ...
```

to calculate the bessel function (integer order only).

It also uses the relation

```
j(-n,x) = (-1)^n * j(n,x)
```

to calculate the bessel when `x < 0`, It has a complexity of `O(n^3)`.

**Note**: this series can also produce errors of 1 ULP, so I recommend users do
their calculations with the precision (`scale`) set to at least 1 greater than
is needed.

### Modular Exponentiation

This `dc` uses the [Memory-efficient method][8] to compute modular
exponentiation. The complexity is `O(e*n^2)`, which may initially seem
inefficient, but `n` is kept small by maintaining small numbers. In practice, it
is extremely fast.

### Non-Integer Exponentiation (`bc` Math Library 2 Only)

This is implemented in the function `p(x,y)`.

The algorithm used is to use the formula `e(y*l(x))`.

It has a complexity of `O(n^3)` because both `e()` and `l()` do.

However, there are details to this algorithm, described by the author,
TediusTimmy, in GitHub issue [#69][12].

First, check if the exponent is 0. If it is, return 1 at the appropriate
`scale`.

Next, check if the number is 0. If so, check if the exponent is greater than
zero; if it is, return 0. If the exponent is less than 0, error (with a divide
by 0) because that is undefined.

Next, check if the exponent is actually an integer, and if it is, use the
exponentiation operator.

At the `z=0` line is the start of the meat of the new code.

`z` is set to zero as a flag and as a value. What I mean by that will be clear
later.

Then we check if the number is less than 0. If it is, we negate the exponent
(and the integer version of the exponent, which we calculated earlier to check
if it was an integer). We also save the number in `z`; being non-zero is a flag
for later and a value to be used. Then we store the reciprocal of the number in
itself.

All of the above paragraph will not make sense unless you remember the
relationship `l(x) == -l(1/x)`; we negated the exponent, which is equivalent to
the negative sign in that relationship, and we took the reciprocal of the
number, which is equivalent to the reciprocal in the relationship.

But what if the number is negative? We ignore that for now because we eventually
call `l(x)`, which will raise an error if `x` is negative.

Now, we can keep going.

If at this point, the exponent is negative, we need to use the original formula
(`e(y * l(x))`) and return that result because the result will go to zero
anyway.

But if we did *not* return, we know the exponent is *not* negative, so we can
get clever.

We then compute the integral portion of the power by computing the number to
power of the integral portion of the exponent.

Then we have the most clever trick: we add the length of that integer power (and
a little extra) to the `scale`. Why? Because this will ensure that the next part
is calculated to at least as many digits as should be in the integer *plus* any
extra `scale` that was wanted.

Then we check `z`, which, if it is not zero, is the original value of the
number. If it is not zero, we need to take the take the reciprocal *again*
because now we have the correct `scale`. And we *also* have to calculate the
integer portion of the power again.

Then we need to calculate the fractional portion of the number. We do this by
using the original formula, but we instead of calculating `e(y * l(x))`, we
calculate `e((y - a) * l(x))`, where `a` is the integer portion of `y`. It's
easy to see that `y - a` will be just the fractional portion of `y` (the
exponent), so this makes sense.

But then we *multiply* it into the integer portion of the power. Why? Because
remember: we're dealing with an exponent and a power; the relationship is
`x^(y+z) == (x^y)*(x^z)`.

So we multiply it into the integer portion of the power.

Finally, we set the result to the `scale`.

### Rounding (`bc` Math Library 2 Only)

This is implemented in the function `r(x,p)`.

The algorithm is a simple method to check if rounding away from zero is
necessary, and if so, adds `1e10^p`.

It has a complexity of `O(n)` because of add.

### Ceiling (`bc` Math Library 2 Only)

This is implemented in the function `ceil(x,p)`.

The algorithm is a simple add of one less decimal place than `p`.

It has a complexity of `O(n)` because of add.

### Factorial (`bc` Math Library 2 Only)

This is implemented in the function `f(n)`.

The algorithm is a simple multiplication loop.

It has a complexity of `O(n^3)` because of linear amount of `O(n^2)`
multiplications.

### Permutations (`bc` Math Library 2 Only)

This is implemented in the function `perm(n,k)`.

The algorithm is to use the formula `n!/(n-k)!`.

It has a complexity of `O(n^3)` because of the division and factorials.

### Combinations (`bc` Math Library 2 Only)

This is implemented in the function `comb(n,r)`.

The algorithm is to use the formula `n!/r!*(n-r)!`.

It has a complexity of `O(n^3)` because of the division and factorials.

### Logarithm of Any Base (`bc` Math Library 2 Only)

This is implemented in the function `log(x,b)`.

The algorithm is to use the formula `l(x)/l(b)` with double the `scale` because
there is no good way of knowing how many digits of precision are needed when
switching bases.

It has a complexity of `O(n^3)` because of the division and `l()`.

### Logarithm of Base 2 (`bc` Math Library 2 Only)

This is implemented in the function `l2(x)`.

This is a convenience wrapper around `log(x,2)`.

### Logarithm of Base 10 (`bc` Math Library 2 Only)

This is implemented in the function `l10(x)`.

This is a convenience wrapper around `log(x,10)`.

### Root (`bc` Math Library 2 Only)

This is implemented in the function `root(x,n)`.

The algorithm is [Newton's method][9]. The initial guess is calculated as
`10^ceil(length(x)/n)`.

Like square root, its complexity is `O(log(n)*n^2)` as it requires one division
per iteration, and it doubles the amount of correct digits per iteration.

### Cube Root (`bc` Math Library 2 Only)

This is implemented in the function `cbrt(x)`.

This is a convenience wrapper around `root(x,3)`.

### Greatest Common Divisor (`bc` Math Library 2 Only)

This is implemented in the function `gcd(a,b)`.

The algorithm is an iterative version of the [Euclidean Algorithm][10].

It has a complexity of `O(n^4)` because it has a linear number of divisions.

This function ensures that `a` is always bigger than `b` before starting the
algorithm.

### Least Common Multiple (`bc` Math Library 2 Only)

This is implemented in the function `lcm(a,b)`.

The algorithm uses the formula `a*b/gcd(a,b)`.

It has a complexity of `O(n^4)` because of `gcd()`.

### Pi (`bc` Math Library 2 Only)

This is implemented in the function `pi(s)`.

The algorithm uses the formula `4*a(1)`.

It has a complexity of `O(n^3)` because of arctangent.

### Tangent (`bc` Math Library 2 Only)

This is implemented in the function `t(x)`.

The algorithm uses the formula `s(x)/c(x)`.

It has a complexity of `O(n^3)` because of sine, cosine, and division.

### Atan2 (`bc` Math Library 2 Only)

This is implemented in the function `a2(y,x)`.

The algorithm uses the [standard formulas][11].

It has a complexity of `O(n^3)` because of arctangent.

[1]: https://en.wikipedia.org/wiki/Karatsuba_algorithm
[2]: https://en.wikipedia.org/wiki/Long_division
[3]: https://en.wikipedia.org/wiki/Exponentiation_by_squaring
[4]: https://en.wikipedia.org/wiki/Newton%27s_method#Square_root_of_a_number
[5]: https://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Babylonian_method
[6]: https://en.wikipedia.org/wiki/Unit_in_the_last_place
[7]: https://people.eecs.berkeley.edu/~wkahan/LOG10HAF.TXT
[8]: https://en.wikipedia.org/wiki/Modular_exponentiation#Memory-efficient_method
[9]: https://en.wikipedia.org/wiki/Root-finding_algorithms#Newton's_method_(and_similar_derivative-based_methods)
[10]: https://en.wikipedia.org/wiki/Euclidean_algorithm
[11]: https://en.wikipedia.org/wiki/Atan2#Definition_and_computation
[12]: https://github.com/gavinhoward/bc/issues/69