# Time complexity of an enumeration of SUBSET SUM instances

An instance of the SUBSET SUM problem (given $y$ and $A = \{x_1,...,x_n\}$ is there a non-empty subset of $A$ whose sum is $y$) can be represented on a one-tape Turing Machine with a list of comma separated numbers in binary format. If $\Sigma = \{0,1,\#\}$ a reasonable format could be:

$( 1 \; (0|1)^* \; \#)^* \#$

Where the first required argument is the value $y$ and $\#\#$ encodes the end of the input. For example:

 1  0  0  #  1  0  #  1  #  #
^^^^^^^^     ^^^^     ^
y          x1     x2
Instance: y=4, A={2,1}


I would like to enumerate the SUBSET SUM instances.

Question: What is the (best) time complexity that can be achieved by a Turing Machine $TM_{Enum}$ that on input $m$ (which can be represented on the tape with a string of size $\log m + 1$) - outputs the $m$-th SUBSET SUM instance in the above format?

EDIT:

Yuval's answer is fine, this is only a longer explanation.

Without loss of generality we set that $y > 0$ and $0 < x_1 \leq x_2 \leq ... \leq x_n$, $n \geq 0$

And we can represent an instance of subset sum using this encoding:

$y \# x_1\# d_2\# ...\# d_{n} \#\#$ where $d_i \geq 1, x_i = x_{i-1} + d_i - 1 \; , i \geq 2$

Using a binary representation for $y,x_1, d_2, d_3, ...$ we have the following representation:

$1 \; ((0|1)^* \# 1)^* \; \#\#$

Equivalent to $1 \; (0|1|\#1)^* \; \#\#$. There is always a leading 1 and a trailing ## so we can consider only the $(0|1|\#1)^*$ part.

So the decoder TM on input $m$ in binary format should:

• convert $m$ to base 3 mapping digit 2 to $\#1$
• when outputing the i-th intermediate $\#$ calculate $x_i = d_i + x_{i-1}-1$
• output the trailing $\#\#$

No duplicate instances are generated.

• What do you mean by "the enumeration"? I assume the specific enumeration is up to the answerer to fix? – Raphael Sep 29 '12 at 18:01
• @Raphael: yes, I'll edit the question – Vor Sep 29 '12 at 18:17
• May the enumeration repeat itself? – Raphael Oct 4 '12 at 10:02
• @Raphael: no it should not repeat itself (otherwise Yuval's enumeration is fine) – Vor Oct 4 '12 at 10:43
• ?? you say you want to enumerate, but this is more like outputing the n'th occurrence in an enumeration as specified by input which is not really the same thing. also you have to specify the set of numbers that you want to enumerate on the input ("A") unless its hardcoded into the TM, right? anyway if you fix that, what you seem to be asking for takes linear time on the input to only pick the 'nth' enumeration.... – vzn Oct 4 '12 at 15:17

SUBSET-SUM instances can be encoded in base 3. We have codes for $0,1,\#$. Some codings are invalid, but in that case we can just immediately output $\#\#$ (or $\#$, if we have just written $\#$). Every SUBSET-SUM problem has infinitely many encodings, I hope that's not a problem.
If the input has length $\ell$, then (assuming the tape alphabet has at least 4 symbols) we can do the conversion in time $O(\ell^2)$. I don't know whether this is the "best" time complexity achievable.
Edit: Here is a better encoding. We still have only three input codes, $0,1,\#$. The output string always starts with $1$ and ends with $\#\#$. Further, $\#$ is output as $\#1$. Now each output string is generated once, though several output strings could correspond to the same instance.
• if I understand well you propose to use the input $m$ ($\ell = \log m + 1$) as the (direct) base 3 encoding of the subset sum instance and map invalid decoded strings to a dum instance $\#\#$ (and the $O(\ell^2)$ comes from "compacting" the decoded string to the left). But do you think that is possible to avoid the invalid instances in the enumeration and decode it without counting them from 1 to $m$? – Vor Oct 4 '12 at 7:55
• thanks the new encoding is better and I think that with the same technique you can generate an enumeration without duplicated instances: just output $x_{i} = x_{i-1} + d_i - 1$ for $i > 1$ ($d_i$ is the i-th argument in your encoding)! – Vor Oct 4 '12 at 15:57