Hoare Logic for Numerical Analysis
by Alex Nelson, 28 April 2024
Broadly speaking, design-by-contract implements Hoare logic in an actual programming language, and I want to consider whether it is plausible to implement numerical analysis code using design-by-contract techniques.
Case Study: Squareroot
I will be using C for my programming language and ACSL for my assertion language. The toy problem we will study will be the square root of a number, computing $\sqrt{x}$ for $x\in\mathbb{R}$ non-negative $x\geq0$.
We can use Newton’s method (a.k.a., the Babylonian method). If we know that $x_{0}^{2}\leq x\leq (x_{0}+1)^{2}$, then $x_{0}$ is our initial guess for the squareroot. Then we compute $(x_{0}+\varepsilon)^{2}\approx x_{0}^{2}+2x_{0}\varepsilon=x$, which we can solve for $\varepsilon=(x-x_{0}^{2})/(2x_{0})$ and this gives us $x_{1}=x_{0}+\varepsilon=(x+x_{0}^{2})/(2x_{0})$ (the next iteration). This gives us a sequence of real numbers $x_{0}$, $x_{1}$, $x_{2}$, …, which converges to $\sqrt{x}$.
We can use the
frexp
function
to extract the exponent of x=m2^{n}
where $1\leq m<2$ and
$n\in\mathbb{Z}$ then use this to approximate $x_{0}=2^{\lfloor n/2\rfloor}$
(or if $n$ is odd $x_{0}=2^{\lfloor n/2\rfloor}\sqrt{2}$ with $\sqrt{2}\approx1.4142135623730950488016887242096980785696$
precomputed and stored as a constant)
to get the correct order-of-magnitude estimate for the squareroot. The
contract for frexp
may be given as:
/*@
requires \is_finite(x) && \valid(exp);
ensures \is_finite(\result);
ensures 0.5 <= \result < 1.0;
ensures x == \round_double(\NearestEven, \result * \pow(2, exp));
assigns \result, *exp \from x;
*/
double frexp(double x, int *exp);
We could have used the ldexp
function to specify how to relate the
constituent parts returned by frexp
to the given double x
.
Instead, we just used the ACSL built-in \pow(a, b)
to give us $a^{b}$
as an exact real number.
We can write up the code for the Babylonian method, taking advantage of the fact that we need only $\log_{2}(b)$ steps to get $b$ bits of accuracy (and double precision floats have $b=53\approx64=2^{6}$).
/*@
requires \isfinite(x);
requires x >= 0.0;
*/
double sqrt_attempt(double x) {
int exp = 0;
const int MAX_ITER = 8;
double mantissa = frexp(x, &exp);
/*@assert x == \round_double(\NearestEven, mantissa * \pow(2, exp)); */
double sqrt = 1.0;
if (1 == exp % 2) {
sqrt = 1.4142135623730950488016887242096980785696;
}
double TOLERANCE = ldexp(x, -53);
/*@assert TOLERANCE = \round_double(\NearestEven, x * \pow(2,-53)); */
sqrt = ldexp(sqrt, exp/2); // initial guess
/*@ ghost double xs[MAX_ITER + 1];
xs[0] = sqrt; */
/*@assert sqrt*sqrt <= x & x < \pow(2, exp+1); */
for (int n=0; n < MAX_ITER; n++) {
if (fabs((sqrt*sqrt) - x) < TOLERANCE) break;
/*@ ghost xs[n+1] = 0.5*((x/xs[n]) + xs[n]); */
sqrt = 0.5*((x/sqrt) + sqrt);
/*@ assert sqrt == xs[n+1]; */
}
/* Assert somehow that xs is first MAX_ITER + 1 terms of Newton's method,
and therefore must converge to sqrt(x), so given the initial guess
of xs[0], the error must be within [error tolerance range].
*/
return sqrt;
}
Note that $x_{n+1}=(x+x_{n}^{2})/(2x_{n})$ but also that
$x_{n+1}=((x/x_{n})+x_{n})/2$. These produce distinct (but related)
numerical methods. For example, division by 2 (or multiplying by 0.5
)
may be done exactly in floating-point arithmetic. If we denote by
$\mathtt{fl}(x)$ the floating-point number for $x$ rounded to nearest
with ties towards even, then we can use the standard model
of floating-point gives us different truncation errors for these terms:
- $\mathtt{fl}((x+x_{n}^{2})/(2x_{n}))\approx ((x+x_{n}^{2})/(2x_{n}))(1+\varepsilon)^{3}$
- $\mathtt{fl}(x_{n}^{2})=x_{n}^{2}(1+\varepsilon_{1})$
- $\mathtt{fl}(x+x_{n}^{2}(1+\varepsilon_{1}))=(x+x_{n}^{2}(1+\varepsilon_{1}))(1+\varepsilon_{2})$
- $\mathtt{fl}((x+x_{n}^{2}(1+\varepsilon_{1}))(1+\varepsilon_{2})/(2x_{n}))=((x+x_{n}^{2}(1+\varepsilon_{1}))/(2x_{n}))(1+\varepsilon_{2})(1-\varepsilon_{3})$
- $\mathtt{fl}(((x/x_{n})+x_{n})/2)\approx\frac{1}{2}((x/x_{n})+x_{n})(1+\varepsilon)^{2}$
- $\mathtt{fl}(x/x_{n})=(x/x_{n})(1-\varepsilon_{1})$
- $\mathtt{fl}((x/x_{n})(1-\varepsilon_{1})+x_{n})=((x/x_{n})(1-\varepsilon_{1})+x_{n})(1+\varepsilon_{2})$
- division by two is exact, so it doesn’t affect the relative truncation error.
In general, fewer truncation errors are preferable.
We can improve performance by:
- Moving the
ldexp(...)
function call to the penultimate step of the function, so we compute the squareroot of the mantissa. - Using the initial guess based on the Taylor series expansion for $\sqrt{\frac{1}{2}+x}\approx(1 + x - x^{2}/2)/\sqrt{2}$, which would have an absolute error of approximately 0.05.
Let’s refactor out this logic into its own subroutine for determining the initial guess for the squareroot of $x$:
/*@
requires 0.5 <= mantissa && mantissa < 1;
ensures fabs(\result*\result - mantissa) < 0.05;
*/
static double sqrt_init_guess(double mantissa) {
double x = mantissa - 0.5;
double invSqrt2 = 0.707106781186547524400844362104849039;
/* @ghost double QuadraticTaylorPolynomialApproximation(double y) {
return invSqrt2*(1 + y*(1 - 0.5*y));
}
*/
/* @fact [?] exists xi st 0 <= xi < 0.5
&& TaylorPolynomial(sqrt(0.5 + x), x, 0, 3)
== QuadraticTaylorPolynomialApproximation(x) + xi*xi*xi/(2*sqrt(2))
&& xi*xi*xi/(2*sqrt(2)) <= pow(0.5,4)/sqrt(2) < 0.442;
*/
/* @assert \result == QuadraticTaylorPolynomialApproximation(x); */
return invSqrt2*(1 + x*(1 - 0.5*x));
}
Exercise: Can we use the linear approximation to the squareroot instead of the quadratic approximation? Would this cost us an extra iteration of Newton’s method? (End of exercise)
We can then rewrite the squareroot algorithm to be:
/*@
requires \isfinite(x);
requires x >= 0.0;
ensures \abs(\result - \sqrt(x)) <= \sqrt(x)*\pow(2,-53);
*/
double sqrt(double x) {
int exp = 0;
const int MAX_ITER = 8;
double mantissa = frexp(x, &exp);
/*@assert 0.5 <= mantissa && mantissa < 1; */
/*@assert x == \round_double(\NearestEven, mantissa * \pow(2, exp)); */
double sqrt_m = sqrt_init_guess(mantissa);
/*@assert \abs(sqrt_m*sqrt_m - mantissa) < 0.05; */
sqrt_m = sqrt_iter(mantissa, sqrt_m); // refactor out the iterative loop to
// its own function
/*@assert \abs(sqrt_m - \sqrt(mantissa)) < \pow(2,-53); */
double sqrt = 1.0;
if (1 == exp % 2) {
sqrt = 1.4142135623730950488016887242096980785696;
/*@assert \abs(sqrt - \sqrt(2)) <= \sqrt(2)*\pow(2,-53); */
}
/*@assert \abs(\result - \sqrt(x)) <= \sqrt(x)*\pow(2,-53);
return ldexp(sqrt*sqrt_m, exp/2);
}
Exercise: When we have $x=m2^{2e+1}$ we compute sqrt_m
to
approximate $\sqrt{m}$ and store $\sqrt{2}$ as the initial value of
sqrt
, then compute sqrt*sqrt_m
as the mantissa for the
solution. What are the error bounds for the result? [Hint: use triangle inequality.]
If it is less than desired precision, how can we recover the desired precision? (End of exercise)
The iterative loop is then its own function, which we can write as:
/*@
requires \abs(init_guess*init_guess - x) < 0.05
requires 0.5 <= x && x < 1;
requires 0.5*sqrt(2) <= init_guess && init_guess < 1;
ensures \abs(\result - \sqrt(x)) < \pow(2,-53);
*/
static double sqrt_iter(double init_guess, double x) {
int MAX_ITER = 3;
/*@ghost double s[MAX_ITER + 1];
real error[MAX_ITER + 1];*/
/*@assert \abs(init_guess/\sqrt(x) - 1) < 0.05/\sqrt(x) < 0.025*\sqrt(2); */
/*@assert \abs(init_guess/\sqrt(x) - 1) < \pow(2,-4); */
double sqrt_m = 0.5*((x/init_guess) + init_guess);
/*@ghost s[0] = 0.5*((x/init_guess) + init_guess);
error[0] = s[0]/\sqrt(x) - 1; */
/*@assert 0 <= error[0] && error[0] < \pow(\pow(2,-4), 2)/2; */
/*@assert error[0] < \pow(2,-9); */
/*@assert 0.5*sqrt(2) <= sqrt_m && sqrt_m < 1; */
int i;
for(i = 0; i < MAX_ITER; i++) {
/*@ghost s[i+1] = 0.5*((x/s[i]) + s[i]);
error[i+1] = s[i+1]/\sqrt(x) - 1; */
/*@assert error[i+1] == error[i]*error[i]/(2*(1 + error[i])); */
/*@assert error[i+1] <= error[i]*error[i]/2; */
/*@assert error[i+1] < \pow(2, 1-5*\pow(2, i+1)); */
sqrt_m = 0.5*((x/sqrt_m) + sqrt_m);
/*@assert \exists real eps;
\abs(eps) < \pow(2, -53)
&& sqrt_m == s[i+1]*(1 + i*eps); */
/*@assert \abs(sqrt_m - \sqrt(x)) <= error[i+1]*\sqrt(x); */
/*@assert \abs(sqrt_m - \sqrt(x)) < error[i+1]; */
}
/*@assert \abs(error[i+1]) < \pow(2, 1-5*\pow(2, i+1)); */
/*@assert \abs(sqrt_m - \sqrt(x)) < \pow(2, -64); */
return sqrt_m;
}
Concluding Remarks
Alright, so this…actually works adequately as pidgin code. I’m not certain the Frama-C tool will accept this input, but it communicates the expectations at each step of the calculation.
However, this particular case study (calculating the squareroot of a non-negative real number) is a rather simple task in numerical analysis. We just wanted to consider a small “proof of concept” to see if Hoare logic could ostensibly express the “mathematical state” to guarantee correctness of the implementation.
Puzzle: Can we use Hoare logic to prove the correctness of numerical linear algebra implementations? What about numerical differential equations? (End of Puzzle)
The difficulty with numerical linear algebra is that we would need to use separation logic, since we are working with arrays allocated on the heap, and we mutate them quite readily. Numerical partial differential equations is even harder, since the proof assistant needs to know quite a bit of analysis.
Exercise: Use separation logic or ACSL to prove the correctness for some implementations of the BLAS routines. (End of exercise)
Perhaps someone has worked on this silently over the years, I could not find anything in the literature. If you are aware of anyone who has formalized Numerical Linear Algebra using separation logic, then please email me with the details!