TL;DR

Here we are with TASK #1 from The Weekly Challenge #174. Enjoy!

# The challenge

Write a script to generate first 19 Disarium Numbers.

A disarium number is an integer where the sum of each digit raised to the power of its position in the number, is equal to the number.

For example,

518 is a disarium number as (5 ** 1) + (1 ** 2) + (8 ** 3) => 5 + 1 + 512 => 518


# The questions

The only nitpicking would be about the base we should consider for doing the calculation, but I guess base-10 is safe to assume nowadays.

Then, of course, there might be the why 19? Probably 18 is too little of a challenge, and 20 is way too much.

# The solution

The nice thing about this challenge is that there is a finite number of Disarium Numbers but we’re not asked for all of them. Most of all: 19 out of 20. This is because the 20th element is… quite big.

It’s not difficult to see that there must be a finite number of these numbers: just think that every time we add a new digit, our number grows a factor of 10 but the Disarium Sum (if you know what I mean) can only grow with the sum of $9^k$, where $k$ is the number of digits. It’s a sum against a multiplication.

So we can calculate the surely not beyond this number that is allowed with $k$ digits as:

$S(k) = \sum_{i = 1}^{k} 9^i$

This allows us to also calculate the number of digits of this surely not beyond this number as:

$D(k) = \lceil log_{10}(S(k)) \rceil$

As long as $D(k)$ is equal to $k$, theoretically we can have a solution. But when it drops below, we cannot possibly have a solution any more and things will never get better:

$D(1) = 1 \\ D(2) = 2 \\ ... \\ D(21) = 21 \\ D(22) = 22 \\ D(23) = 22 < 23$

So, Disarium Numbers cannot possibly have 23 digits or more. It turns out that the biggest element of the family has 20 digits.

As I said, Disarium Numbers: brute force won’t cut it. Well, for my definition of brute force, at least. I started coding in Raku but it was taking a bit too much time that I stopped it - it’s all in the other post.

It turns out that brute force can actually cut it in about 8/9 seconds in my machine with Perl:

my $n = shift // 19; my @disariums = disariums($n, \&is_disarium_espresso);
say join ', ', @disariums;

my $n = shift // 19; my @disariums = disariums($n, \&is_disarium_espresso);
say join ', ', @disariums;

sub disariums ($n,$tester) {
my @disariums;
my $candidate = 0; while (@disariums <$n) {
push @disariums, $candidate if$tester->($candidate); ++$candidate;
}
return @disariums;
}
sub is_disarium_espresso ($n) { my$exp = 0;
$n == sum map {$_ ** ++$exp } split m{}mxs,$n;
}


So well, OK, I was wrongish for the challenge at stake (find 19 members and ignore the 20th), but way to go for finding them all.

Anyway, I did the Raku solution first. adopting a partially optimized algorithm that can solve the challenge in slightly more than one second and can theoretically find the 20th member too in about… 3 months or so 🙄

BUT: the algorihtm is highly parallelizable, so with the right adaptations and the right machine(s) it can be probably brought down to hours/minutes even in Raku (which is -Ofun as of today).

It’s not for the faint of heart and involves starting from the last digit and going leftwards, finding the maximum as above, the minimum in a similar way and recurse if their difference is too big for iteration.

#!/usr/bin/env raku
use v6;
sub MAIN (Int:D $n where * > 0 = 19) { my @disarium-numbers; my$length = 0;
my $last = time; while @disarium-numbers <$n {
++$length; put "checking length$length...";
@disarium-numbers.push: disariums-long-suffix($length).Slip; my$now = time;
put "   {$now -$last}s {@disarium-numbers.elems} found so far ({@disarium-numbers.join(', ')})";
$last =$now;
}
say @disarium-numbers;
exit 0;
}

sub disarium-calc ($n) {$n.comb.kv.map(-> $x,$y { $y ** ($x + 1) }).sum }

multi sub disariums-long-suffix (1) { return [0 .. 9] }

multi sub disariums-long-suffix ($length,$suffix = '') {
my @suffix = $suffix.comb.reverse; my$slen = $suffix.chars; my$residual = $length -$slen;

my $baseline = 0; # only from suffix my$exp = $length; for @suffix ->$base { $baseline +=$base ** $exp-- } my$max_increment = 0;
while $exp > 0 {$max_increment += 9 ** $exp-- } # optimize probably? my$min_baseline = $baseline + 1; # 1000...NNN$min_baseline = (('1' x $residual) ~$suffix).Int
if $min_baseline.chars <$length;
my $min_baseline_prefix =$min_baseline.substr(0, $residual).Int; my$min_baseline_suffix = $min_baseline.substr(*-$slen, $slen).Int; ++$min_baseline_prefix if $suffix <$min_baseline_suffix;
$min_baseline = ($min_baseline_prefix ~ $suffix).Int; my$max_baseline = $baseline +$max_increment;
my $n_digits =$max_baseline.chars;
return [] if $n_digits <$length;
my $max_baseline_prefix =$max_baseline.substr(0, $residual).Int; my$max_baseline_suffix = $max_baseline.substr(*-$slen, $slen).Int;$max_baseline_prefix-- if $suffix >$max_baseline_suffix;
$max_baseline = ($max_baseline_prefix ~ $suffix).Int; my @collected; my$delta = $max_baseline_prefix -$min_baseline_prefix;
if $delta > 10 { for (0 .. 9).reverse ->$digit {
my $new_suffix =$digit ~ $suffix; my @children = disariums-long-suffix($length, $new_suffix); @collected.push: |@children; } } elsif$delta >= 0 {
for $min_baseline_prefix ..$max_baseline_prefix -> $prefix { my$candidate = ($prefix ~$suffix).Int;
my $check =$baseline + disarium-calc($prefix); @collected.push:$candidate if $candidate ==$check;
}
}
return @collected;
}


The trivial case for 1-digit numbers is taken care by the multi mechanism. We then interate over the lenghts to find all elements of that length, stopping when we have enough numbers. For the challenge, this means stopping at 7 digits which is fair.

I know the implementation sucks and there’s no comment to understand what’s going on, but I’ll call this a week and move on.

On the Perl side, I toyed with the idea of optimizing the check for disarium numbers, pre-computing the exponentiation inside a AoA and inside a linear array, so I ended up with this:

#!/usr/bin/env perl
use v5.24;
use warnings;
use experimental 'signatures';
no warnings 'experimental::signatures';
use List::Util 'sum';
use Benchmark 'timethese';

$|++; my$n = shift // 19;
my @disariums = disariums($n, \&is_disarium_espresso); say join ', ', @disariums; timethese( 5, { espresso => sub { disariums($n, \&is_disarium_espresso)  },
precached => sub { disariums($n, \&is_disarium_precached) }, llcached => sub { disariums($n, \&is_disarium_llcached)  },
},
);

sub disariums ($n,$tester) {
my @disariums;
my $candidate = 0; while (@disariums <$n) {
push @disariums, $candidate if$tester->($candidate); ++$candidate;
}
return @disariums;
}

sub is_disarium_precached ($n) { state$pow = [];
while ((my $k =$pow->@*) < length($n)) { ++$k;
push $pow->@*, [ 0, 1, map {$_ ** $k } 2 .. 9 ]; } my$exp = 0;
$n == sum map {$pow->[$exp++][$_] } split m{}mxs, $n; } sub is_disarium_llcached ($n) {
state $pow = []; state$kpow = 0;
while ($kpow < length($n)) {
++$kpow; push$pow->@*, 0, 1, map { $_ **$kpow } 2 .. 9;
}
my $exp = 0;$n == sum map { $pow->[$_ + 10 * $exp++] } split m{}mxs,$n;
}

sub is_disarium_espresso ($n) { my$exp = 0;
$n == sum map {$_ ** ++$exp } split m{}mxs,$n;
}


My gut feeling was that the linear cache (llcached) would score better than the AoA cache (precached), which would be better than the basic approach of doing all exponentiations all the times (espresso).

I got back yet another demonstration that benchmarks are a thing and my gut feelings are disastrously not:

\$ time perl perl/ch-1.pl 19
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 89, 135, 175, 518, 598, 1306, 1676, 2427, 2646798
Benchmark: timing 5 iterations of espresso, llcached, precached...
espresso: 42 wallclock secs (42.21 usr +  0.05 sys = 42.26 CPU) @  0.12/s (n=5)
llcached: 48 wallclock secs (46.75 usr +  0.09 sys = 46.84 CPU) @  0.11/s (n=5)
precached: 47 wallclock secs (47.10 usr +  0.06 sys = 47.16 CPU) @  0.11/s (n=5)


Go figure.

Stay safe!

Comments? Octodon, , GitHub, Reddit, or drop me a line!