An addition-subtraction chain, a generalization of addition chains to include subtraction, is a sequence a0, a1, a2, a3, ... that satisfies

${\displaystyle a_{0}=1,\,}$
${\displaystyle {\text{for }}k>0,\ a_{k}=a_{i}\pm a_{j}{\text{ for some }}0\leq i,j

An addition-subtraction chain for n, of length L, is an addition-subtraction chain such that ${\displaystyle a_{L}=n}$. That is, one can thereby compute n by L additions and/or subtractions. (Note that n need not be positive. In this case, one may also include a-1=0 in the sequence, so that n=-1 can be obtained by a chain of length 1.)

By definition, every addition chain is also an addition-subtraction chain, but not vice-versa. Therefore, the length of the shortest addition-subtraction chain for n is bounded above by the length of the shortest addition chain for n. In general, however, the determination of a minimal addition-subtraction chain (like the problem of determining a minimum addition chain) is a difficult problem for which no efficient algorithms are currently known. The related problem of finding an optimal addition sequence is NP-complete (Downey et al., 1981), but it is not known for certain whether finding optimal addition or addition-subtraction chains is NP-hard.

For example, one addition-subtraction chain is: ${\displaystyle a_{0}=1}$, ${\displaystyle a_{1}=2=1+1}$, ${\displaystyle a_{2}=4=2+2}$, ${\displaystyle a_{3}=3=4-1}$. This is not a minimal addition-subtraction chain for n=3, however, because we could instead have chosen ${\displaystyle a_{2}=3=2+1}$. The smallest n for which an addition-subtraction chain is shorter than the minimal addition chain is n=31, which can be computed in only 6 additions (rather than 7 for the minimal addition chain):

${\displaystyle a_{0}=1,\ a_{1}=2=1+1,\ a_{2}=4=2+2,\ a_{3}=8=4+4,\ a_{4}=16=8+8,\ a_{5}=32=16+16,\ a_{6}=31=32-1.}$

Like an addition chain, an addition-subtraction chain can be used for addition-chain exponentiation: given the addition-subtraction chain of length L for n, the power ${\displaystyle x^{n}}$ can be computed by multiplying or dividing by x L times, where the subtractions correspond to divisions. This is potentially efficient in problems where division is an inexpensive operation, most notably for exponentiation on elliptic curves where division corresponds to a mere sign change (as proposed by Morain and Olivos, 1990).

Some hardware multipliers multiply by n using an addition chain described by n in binary:

    n = 31 = 0  0  0  1   1  1  1  1 (binary).


Other hardware multipliers multiply by n using an addition-subtraction chain described by n in Booth encoding:

    n = 31 = 0  0  1  0   0  0  0 -1 (Booth encoding).