Recently I revisited two’s complement arithmetic. It’s one of the coolest tricks I’ve seen.

Computers internally represent integers (whole numbers like -1, 0, 42) with a a fixed number of bits. Consider a tiny computer that allows only 3-bit integers. So if we are talking only non-negative integers, 0 could be coded as the bit vector 000, 1 as 001, 2 as 010, … 7 as 111. This is just one possible coding, probably the most obvious one.

But say we also want to represent negative integers. A simple coding there could be to treat the most significant (leftmost) bit as the sign bit, and the remaining two bits for absolute value of the integer being represented. Here’s a possible mapping:

000 --> +0
001 --> +1
010 --> +2
011 --> +3
100 --> -0
101 --> -1
110 --> -2
111 --> -3

Note two points:

  1. There are two zeros, which means we waste one combination of bits (000 and 100 both are technically the number 0; we could have used one of them for an extra integer, either in the negative or the positive half)
  2. We need to build circuitry to be aware of negative numbers for operations like addition (adding a positive and a negative should produce a sign bit equal to the larger of the two numbers, and should have the value equal to the bitwise subtractions of the respective values).

Another coding of negative numbers, called the two’s complement encoding, sidesteps both of these issues. Here’s how the coding looks:

000 --> +0
001 --> +1
010 --> +2
011 --> +3
100 --> -4
101 --> -3
110 --> -2
111 --> -1

In this approach, the core idea is this: When we add (using the usual bitwise addition rules) together the bitvector representing an integer N and its negative -N, the result should be a bit vector 1 followed by as many zeros as our representation’s bit width. So adding +1 (001) and -1 (111) gives 1000. So does adding +2 (010) and -2 (110). Technically, these pairs are called two’s complements of each other – Any two numbers representable with <= $W$ bits that add up together to give exactly $2^W$ are two’s complements of each other. In our example, $W=3$, so $2^W = 8$, which is written 1000 in binary.

One exception here is -4, which does not have its corresponding positive under this coding. That’s because its two’s complement for a 3-bit representation is itself! (100 + 100 = 1000). In general, in the 2c representation, the negative half is always exactly one element larger than the positive half, owing to one element in the positive half being allotted to 0.

Here’s the cool bit: Say we want to subtract a number from itself. One way to do this is to add the number to its negative. As shown above, in the two’s complement representation, this operation always produces 1000. That’s 4 bits, but we only have 3 bits. What happens in this case? The left most bit is simply discarded (it “overflows”), and we are left with 000, which is 0, which should be the result of subtracting a number from itself! Now this works no matter what representible numbers we pick: Consider subtracting 1 from 3. That’s the same as adding -1 and 3:

   111      (-1)
   011      (+3)
 1 010
 ^ ^^^------(+2)

Note that one “feature” of this representation is that incrementing the highest positive value “wraps” the result to the most negative value (incrementing 011, which is +3 gives 100, which is -4).

One way to easily compute the two’s complement (for bit width W):

  1. First write down the number in binary, with necessary padding on the left with 0’s to fill all of W bits.
  2. Flip the bits (0s become 1s, 1s become 0s).
  3. Add 1 to the result

BTW the result of step (2) above is called the one’s complement of the original number: When you add a number with its one’s complement, you get a bitvector of all 1’s, W-bits wide (e.g., one’s complement, for a 3-bit representation, of 001 is 110, because adding them together gives 111).

Today’s computers do not have 3 bit integers: an int64 is quite standard, and it stands for 64 bit wide bit-vector, internally represented exactly as above: The positive half goes from $0$ to $2^{63} - 1$, and the negative half from $-2^{63}$ to $-1$.

See this page and the Wikipedia entry for more discussion.