Ok so I have today found an expression that wolfram alpha breaks on. I think so anyhow.

I would like to evaluate (very roughly)


Wolfram Alpha answers that this expression is $1$ which is wrong and is the first expressions I've seen it stumble on (quite badly too).

The reason why it's wrong is that unless I made a mistake somewhere this is (roughly) the chance of getting a collision when choosing $6*10^{10}$ primes of length $2^{1024}$ uniformly at random. Since you should get a probability of one half after choosing roughly $2^{510}$ primes you should be getting something very close to zero after choosing "only" 60 billion. I have a few ideas how to approximate the result but the factorials are giving me trouble.

Any ideas on how to do this will be welcomed. Obviously any answer that can bound the number within a couple orders of magnitude would be great.

  • $\begingroup$ As an aside: how do you know that this is the first expression that Wolfram Alpha has miscalculated for you? $\endgroup$ – Rob Arthan Aug 10 '16 at 20:26
  • $\begingroup$ @RobArthan Well to be perfectly honest I don't, I don't tend to check every digit when I compute things. On the other hand I do tend to sanity check answers so I'm fairly confident it is the first one. $\endgroup$ – DRF Aug 10 '16 at 20:30
  • $\begingroup$ Thanks for the info. It's interesting to hear opinions about the quality of results from computer algebra systems. $\endgroup$ – Rob Arthan Aug 10 '16 at 20:55
  • $\begingroup$ You could write a Java program to calculate this for you using the BigDecimal class. $\endgroup$ – kamoroso94 Aug 10 '16 at 21:19
  • 1
    $\begingroup$ @RobArthan: Check out this amusing example of how and why WA gets limits wrong. The reason is that WA is basically a product of programmers who do not really know the mathematics behind algorithms and heuristics that they implement, so they always get it wrong. $\endgroup$ – user21820 Mar 10 '17 at 10:48

Using $n! \approx \sqrt{2\pi n}\left(\frac{n}{e}\right)^n$, $\ln n! \approx n\ln(n) - n + \tfrac12\ln(2\pi n)$ and $\ln(1+x) \approx x-\tfrac12x^2$ for small $x$, and some liberal use of these and other approximations, we might get something like

$\dfrac{m!}{(m-k)!} $ $\approx \exp\left(m\ln(m) - m + \tfrac12\ln(2\pi m) - (m-k)\ln(m-k) + (m-k) - \tfrac12\ln(2\pi (m-k)) \right)$ $=\exp\left(k\ln(m)+\left(m-k+\tfrac12\right)\ln\left(1+\tfrac{k}{m-k} \right) -k\right)$ $\approx m^k\exp\left(\left(m-k+\tfrac12\right)\left(\tfrac{k}{m-k}-\tfrac12\left(\tfrac{k}{m-k}\right)^2 \right) -k \right) \qquad$ $= m^k\exp\left( -\tfrac12 \tfrac{k^2}{m-k} +\tfrac12 \tfrac{k}{m-k}-\tfrac14\tfrac{k^2}{(m-k)^2} \right) \approx m^k\exp\left( -\tfrac12 \tfrac{k^2}{m-k} \right)$

so $1-m^{-k}\tfrac{m!}{(m-k)!} \approx \tfrac12 \tfrac{k^2}{m-k} $ when $k^2 \ll m$, in which case we may as well use $$1-m^{-k}\dfrac{m!}{(m-k)!} \approx \dfrac{k^2}{2m}.$$

As an example with $m=10^6$ and $k=200$, we would have $\tfrac{k^2}{2m} =0.02$ while in fact $1-m^{-k}\tfrac{m!}{(m-k)!} \approx 0.0197046$.

In your case with setting $m=2^{1021}$ and $k=6\times 10^{10}$, we would have $\tfrac{k^2}{2m} \approx 8 \times 10^{-287}$ which, as you suspected, is close to $0$.


We want $1-f(n,m)$, where $f(n,m) = \frac{n!}{(n-m)!} n^{-m}$, for $n = 2^{1021},\ \ m=6\cdot10^{10}$.

Using Sage to first compute $\log(f(n,m))$:

R = RealField(prec=200)
def log_factorial(n):
    x = n + 1
    return (x - 1/2)*log(x) - x + log(2*pi)/2 + 1/(12*x)
def log_f(n,m): return R(log_factorial(n) - log_factorial(n-m) - m*log(n))
n = 2^1021
m = 6*10^10
print log_f(n,m)
print 1-e^log_f(n,m)

with output

  • $\begingroup$ So... very very very close to one? $\endgroup$ – Pedro Tamaroff Aug 10 '16 at 21:35
  • $\begingroup$ I like the try and +1 to you but I feel we should be getting a bound in at least double digits of 0's beyond the decimal point. Still good try. $\endgroup$ – DRF Aug 10 '16 at 21:41
  • $\begingroup$ @PedroTamaroff - No, the final line of output is $\quad 1-\exp(\log(f(n,m)))=1-f(n,m)\approx 0$. $\endgroup$ – r.e.s. Aug 10 '16 at 21:43
  • $\begingroup$ @DRF - I've edited to show the results of increasing the precision (to 200 bits) for the real number representations in Sage. I think the number of terms used in the computation of the log-factorial function is adequate for this. $\endgroup$ – r.e.s. Aug 10 '16 at 21:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.