I stand corrected. At some point, I probably read the linked list thing that LanX mentioned and now misremembered it.

To convince myself, I threw together:

#!/usr/bin/env perl    

use strict;
use warnings;
use 5.010;

use Memory::Usage;

my @array1;
my @array2;

my $mu = Memory::Usage->new();
$mu->record('ready to go');

$array1[5268] = 1;
$mu->record('array1 has an element');

$array2[8675309] = 1;
$mu->record('array2 has an element');

$mu->dump();
[download]

Running this on a Debian 8.11 machine with perl 5.20.2, I get the result:

  time    vsz (  diff)    rss (  diff) shared (  diff)   code (  diff)
+   data (  diff)
     0  20824 ( 20824)   2568 (  2568)   1916 (  1916)      8 (     8)
+    920 (   920) ready to go
     0  20824 (     0)   2568 (     0)   1916 (     0)      8 (     0)
+    920 (     0) array1 has an element
     0  88600 ( 67776)  70416 ( 67848)   2048 (   132)      8 (     0)
+  68696 ( 67776) array2 has an element
[download]

The array index 5268 that I used for array1 is a magic number, apparently corresponding to the minimum size that my perl allocates for an array when it's initially declared. If I increase the index to 5269, it shows an additional 132k (all the numbers are in kilobytes) allocated when array1 is assigned to.