3 dimensionnal matching to partition transformation

We want to transform $3DM$ to $PARTITION$, I am reading Garey and Johnson book and I really don't understand how they do the transformation, I know how they create elements $a_i$ from triples of set $M = W \times X \times Y$, but I don't know why they add two other elements $b_1$ and $b_2$ with definition
$$
s(b_1) = 2 (\sum_{i=1}^{k} s(a_i)) - B \\
s(b_2) = (\sum_{i=1}^{k} s(a_i)) + B
$$

with $B$ being:
$$
B = \sum_{j = 0}^{3q - 1} 2 ^ { pj}
$$
and $p = [log_2(k)] + 1$ and $q$ is the number of elements in $X$ (Y, W also have the same number of elements) and also $k$ is the number of triples in $M$.

It would be very nice if anyone could clarify this transformation.

What they actually do is a composition of two reductions (and some optimization). The first is from 3DM to SUBSET-SUM, and the second is from SUBSET-SUM to PARTITION. For the first part, I have it written down here, and the second reduction is quite well known, e.g. here.

Note that there is a small caveat in the first reduction, where we assume that the numbers are encoded in base $|X|+1$, while SUBSET-SUM requires the numbers to be in binary. The reduction you mention does the conversion explicitly, which accounts for the strange $p$ you encountered (which is just the base-conversion constant).

StackExchange Computer Science Q#33871, answer score: 3