What do you subtract from what?

Okay, you’ve calculated the pie correctly and even split it between the two parties. But how do
you go from there to figuring out what each side should pay? This last step seems to be a
stumbling block for some of you. As one of my students put it: What do you subtract from what?

The solution is to go back to what each person would do if they don’t reach an agreement and
then make them half a pie better off.

In the case of Aegean and Baltic that means if the party was already doing the project on its own,
then it just pays half a pie less than that amount. If the party wouldn’t do the project on its own,
then it pays half a pie less than its gross benefit, leaving it half a pie better off than doing nothing.

Let’s review the table below. In the first row, when the cost is 50, both Aegean and Baltic would
do the project on their own. Since the pie is 50, each pays 25 less than what they would pay on
their own, or 25.*

In the second row, at a cost of 150 Baltic would do the project on its own, but Aegean wouldn’t.
The pie is 100, so half the pie is 50. Baltic pays 50 less than doing the project on its own, or 100,
and Aegean pays 50 less than its gross benefit, or 50. In that case, Aegean also ends up 50
ahead when compared to doing nothing.

In the last row, when the cost is 250, neither party would do the project on its own. Here the pie
is 50, and half the pie is 25. So, Aegean pays 25 less than its gross benefit, 100 – 25 = 75, and
Baltic pays less 25 than its gross benefit, 200 – 25 =175.

Cost Aegean pays Baltic pays

(100 benefit) (200 benefit)

50 25 25

150 50 100

250 75 175

If you find mathematical formulas helpful, here’s another way to think about it. Let’s call the
amount Aegean pays when they work together CostShare _AA and what Baltic pays
CostShare_BB.

When they do the project together, Aegean’s net benefit is

Aegean’s gross benefit – CostShare _AA.

When they don’t work together, Aegean gets

Aegean’s gross benefit – Cost = Aegean’s net benefit on its own.

As we wrote above:

The solution is to go back to what each person would do if they don’t reach an agreement and
then make them half a pie better off.

In equation terms,

Aegean’s net benefit together = Aegean’s net benefit on its own + Half the pie

which is equal to

Aegean’s gross benefit – CostShare _AA = Aegean’s net benefit on its own + Half the pie.

By rearranging terms, we get

CostShare_AA = Aegean’s gross benefit – (Aegean’s net benefit on its own + Half the pie).

And likewise for Baltic

CostShare_BB = Baltic’s gross benefit – (Baltic’s net benefit on its own + Half the pie).

Let’s try it out. In the case where the software costs 50, Aegean’s gross benefit is 100. On its
own, Aegean’s net benefit would be 50, and half the pie is 25. Therefore, the equation for
Aegean’s cost share becomes:

CostShare_AA = 100 – (50 + 25) = 25.

In the case where the software costs 150, Aegean’s gross benefit is still 100. On its own,
Aegean’s net benefit would be 0, and half the pie is 50. Therefore, the equation for Aegean’s
cost share becomes:

CostShare_AA = 100 – (0 + 50) = 50.

The situation would be a little bit different if Aegean and Baltic’s gross benefits changed based
on whether or not they worked together. For instance, imagine they’re negotiating over a lease
on some office space. In this case, Aegean and Baltic would each gain more from having the
office space all to itself.
Say, Aegean’s benefit from leasing the offices on its own is 100 while its benefit from sharing the
offices with Baltic is only 85.

Similarly, Baltic’s benefit from leasing the offices on its own is 200 while its benefit from sharing
the offices is 175.

How then do we calculate how much each party should pay when they work together? Let’s go
through the example when the cost is 100.

We first want to calculate the pie:

Pie = Net benefit together – Σ Net benefits on their own.

To calculate their net benefit together:

Net benefit together = (Aegean’s gross benefit together + Baltic’s gross benefit together) – Cost

= (85 + 175) – 100 = 160.

Next we calculate what each would do on its own. Note that each party has a greater benefit
when they don’t have to share office space (but that greater benefit doesn’t justify the extra cost).
In particular, the cost is high enough that Aegean is just on the fence.

Aegean’s net benefit on its own = 100 (Gross benefit on its own) – 100 (Cost) = 0

Baltic’s net benefit on its own = 200 – 100 = 100

Note that, on its own, Aegean gets 0 whether it signs the lease or doesn’t.

That means the pie is 160 – (0 + 100) = 60, which we split evenly between Aegean and Baltic.

How much then does each party pay? Each should be made half a pie better off by doing the
deal than they would be on their own.

Here we need to specify that the gross benefit we’re talking about is Aegean’s gross benefit from
working together with Baltic to lease the offices.

CostShare_AA = Aegean’s gross benefit together – (Aegean’s net benefit on its own + Half the
pie)

= 85 – (0 + 30) = 55.

Thus Aegean’s share of the cost is 55. In the same way, Baltic’s cost share is 45.

CostShare_BB = Baltic’s gross benefit together– (Baltic’s net benefit on its own + Half the pie)
= 175 – (100 + 30) = 45.

_________________________________________

* For your reference, here’s the calculation of the pie:

Cos Net benefit if Net benefits

Pie
t work together on their own

50 + 150 =
50 300 – 50 = 250 50
200

300 – 150 =
150 0 + 50 = 50 100
150

250 300 – 250 = 50 0+0=0 50

Further Discussion on the Shapley Value Axioms (Optional)

If you are curious, here I explain a bit more about what I meant by adding two problems together.
Let's first look at our taxi problem. We'll represent the cost of someone getting home by the
function Cost( ). Thus Cost(A) is how much it costs for A to get home. In our problem:

Cost(A) = 6

Cost(B) = 11

Cost(AB) = 12

According to the linearity (or adding) axiom, the above game has the exact same answer as a
modified game in which A pays 6 and B pays 11 right off the bat, and we then subtract those two
amounts from the joint cost of getting A and B home. In this modified game:

Cost(A) = 6 – 6 = 0

Cost(B) = 11 – 11 = 0

Cost(AB) = 12 – (6 + 11) = –5.

Everything is perfectly symmetric, so it is obvious we should split the $5 savings: A gets back
$2.50 and thus pays $6 – $2.50 = $3.50, while B pays $11 – $2.50 = $8.50. And this is exactly
the split-the-pie solution.

Now let's apply this same logic to our three-person problem.

Cost(A) = 6

Cost(B) = 11

Cost(C) = 15

Cost(AB) = 12

Cost(AC) = 17

Cost(BC) = 17

Cost(ABC) = 18

First we have A pay 6, B pay 11, and C pay 15. Then we subtract those amounts from any ride in
which A, B, and C appear. This leads to a modified game, where:

Cost(A) = 6 – 6 = 0

Cost(B) = 11 – 11 = 0

Cost(C) = 15 – 15 = 0

Cost(AB) = 12 – (6 + 11) = –5

Cost(AC) = 17 – (6 + 15) = –4

Cost(BC) = 17 – (11 + 15) = –9

Cost(ABC) = 18 – (6 + 11 + 15) = –14

Next we split the pairwise savings evenly. AB creates a savings of 5 that we split between A and
B, AC creates a savings of 4 that we split between A and C, and BC creates a savings of 9 that
we split between B and C. Adding up the changes, A paid $6 in the initial modification and then
gets back $4.50 in the second. For A, B, and C that means:

A pays 6 – (1/2)(5 + 4) = $1.50

B pays 11 – (1/2)(5 + 9) = $4.00

C pays 15 – (1/2)(4 + 9) = $8.50

But we're not quite done. Since we've allocated the savings from AB, AC, and BC to each of the
players, we have to subtract those savings out of the combined savings from ABC. The modified
cost of the triangle route is:

Cost(ABC) = –14 + (5 + 4 + 9) = 4

In other words, we had allocated 5 + 4 + 9 =18 of cost savings, but there was really only 14 to go
around. Thus we have to take back 4.

Taking into account the payments already assigned, our new game is:

Cost(A) = 0

Cost(B) = 0

Cost(C) = 0

Cost(AB) = 0

Cost(AC) = 0

Cost(BC) = 0

Cost(ABC) = 4

Everything here is perfectly symmetric, so we simply divide Cost(ABC) three ways. Thus the final
payments are:

A pays 1.5 + 4/3 = $2.83

B pays 4.0 + 4/3 = $5.33

C pays 8.5 + 4/3 = $9.83

This is exactly the same as what we found using the random order approach. What I especially
like about this way of calculating the Shapley Value is how it emphasizes the splitting-the-pie
perspective. Everyone gets an equal share of the cost savings they create. Thus A shares in the
cost savings he creates with B and C, while B shares in the cost savings he creates with A and
C. And to the extent that we double count—and the allocated savings add up to more than what
exists—we take back some of those savings, splitting the take-back equally among the parties.

If you would like to download this discussion as a PDF, see the Downloads section below.
Nucleolus (advanced and very much optional)
I want to be careful that I don’t leave you with the impression that the Shapley Value is the only
fair way to divide the pie. I think it has many desirable properties, but it isn’t the only option. Here
I’ll explain one other approach, called the Nucleolus (like the largest structure inside the nucleus
of a cell), which was developed by David Schmeidler in 1969.

To explain how the Nucleolus works, let’s take a slightly different version of the Runway Problem.
We’ll have three airlines as before.

A needs a runway of length 12

B needs a runway of length 18

C needs a runway of length 18

Under the Shapley Value approach, we’d say the three airlines would split the cost of the first
length three ways and then B and C, as the only users of the next half length, would share that
cost equally. Thus,

A pays 4

B pays 4 + 3 = 7

C pays 4 + 3 = 7

Before turning to the Nucleolus, we can determine a few properties that any fair solution should
obey.

For starters, since A is using less of the runway than B or C, whatever amount A pays, it should
be no more than what B or C pays.

From this, it follows that B and C should pay the exact same amount, as B uses no more than C
and C uses no more than B.

Combining these two properties, we can conclude the most A should ever pay is 6. Since B and
C must pay at least as much as A, once A pays 6 so must B and C, and together that covers the
full cost of the runway.

The most A should pay is 6. What is the least? Well, it wouldn’t be fair for A to pay less than 4, as
that is its equal share of the cost of the runway it uses. Therefore, A should pay something
between 4 and 6. The Shapley Value has A paying 4. The Nucleolus makes the argument for
why A should pay 6. It has to do with how much each side gains when one individual (or group)
joins another.
Imagine B and C have already formed a partnership. In that case, the two of them have already
saved 18 by coming together. If A joins them, the three-way partnership will create another 12 of
savings. Under the Shapley Value, A pays 4 and thus saves 8, while the (B, C) partnership only
saves 4. Thus Airline A gains double what the (B, C) partnership gets. This seems unfair. A
needs the (B, C) group just as much as the (B, C) pair needs A. The Nucleolus proposes this
gain be split evenly, namely 6 and 6, which requires A to pay 6.

You might be wondering why I picked the combination of A joining (B, C). Why not look at B
joining an (A, C) partnership? Indeed, the Nucleolus looks at all the possible combinations. When
B joins (A, C), there is a gain of 18. Using the Shapley Value approach, B would pay 7 and save
11, while (A, C) would only save 7. This isn’t equal, either. To make this equal, we’d have to have
B pay 9. By symmetry, C would have to pay 9. But that creates a bigger issue as then A would
pay nothing. We’ve made the asymmetry even worse than in the case where A joins (B, C). If A
pays nothing, then all of the gain goes to A and none to (B, C).

What the Nucleolus does is finds the division that maximizes the smallest gain. And once that is
done, it maximizes the next smallest gain subject to not lowering the smallest gain. It isn’t always
possible to make all the gains equal and when it isn’t, the Nucleolus comes as close as possible.

Here’s what I mean. Recall that when A joins (B, C), there is only 12 to go around. If we split this
evenly, A saves 6 which implies it pays 6. And since B and C must not pay less than A, they pay
6 as well. So, we have a cost division where A, B, and C all pay 6.

Let’s see how the gains are split in each combination of an individual joining a pair. We know that
the gains from A joining (B, C) are split evenly. When B joins (A, C), since B pays 6, it gains 12
while (A, C) gains 6. And the same is true for C when it joins (A, B). We might like to increase
how much (A, C) or (A, B) gains as the divisions are lopsided toward the other party, but to do so
would require that B pays more and, by symmetry, C pays more. If B and C were to each pay
more than 6, then the (B, C) pair would gain less than 6 when A joins them. Thus we can’t make
the result for when B joins (A, C) more fair without making the A joins (B, C) scenario even less
fair.

John Rawls argued that society should work to maximize the welfare of its worst-off members. In
a similar vein, the Nucleolus looks to find the cost division that maximizes the gain to the group
that is getting the least from coming together. And it keeps on doing that to the extent possible.

Let me provide a few more examples.

A needs a runway of length 12

B needs a runway of length 24

C needs a runway of length 24

Under the Shapley Value approach, we’d say that all three airlines would split the cost of the first
length three ways and then B and C, as the only users of the next length, would share that cost
equally. Therefore,

A pays 4

B pays 4 + 6 = 10

C pays 4 + 6 = 10

But under this cost division, when A joins (B, C), A gains 8 and (B, C) only gains 4. To equalize
this, we should have A pay 6. Thus the Nucleolus solution is:

A pays 6

B pays 9

C pays 9

Unlike our first example, it is not the case here that all three parties split the cost evenly. Indeed,
as the runway needed by Airlines B and C increases, they pay all of the additional costs. A never
pays more than half the cost of the first length.

And to the extent that the runway needs of B and C decrease (from their starting point of 18), the
three airlines will continue to split the full cost three ways. For example, with the numbers below,
A, B, and C would each pay 5 under the Nucleolus.

A needs a runway of length 12

B needs a runway of length 15

C needs a runway of length 15

Still wondering why it’s fair for A to be paying an equal share of the full runway cost? The reason
is B and C can together create a large amount of savings without A. Basically, A is very lucky to
be joining (B, C) and should be happy to split the surplus created when it joins this pair.

This becomes even clearer if we add additional airlines that need the longer runway. Consider
this case:

A needs a runway of length 12

B needs a runway of length 16

C needs a runway of length 16

D needs a runway of length 16

Under the Shapley Value, A pays 3, while under the Nucleolus, A pays 4. Either way, A gets a
great deal in joining the (B, C, D) trio. A gains either 9 or 8 while the trio gains 3 or 4. To make
the gains as equal as possible, A should pay 4. (A can’t pay more than 4 as that would mean it
pays more than B, C, and D.)

For our final example, we’ll forgo airport runways for estate settlements. The Babylonian Talmud
provides an illustration for how to divide up an estate in the face of competing claims.
Surprisingly (for a 2,000 year old book), the unusual solution proposed is the same as the
Nucleolus.

Here's the set up. A person dies owing money to three creditors. They are owed 100, 200, and
300 respectively. These debts are larger than the assets of the estate, so they can't all be paid
off. The table below shows how the assets should be allocated, according to the Talmud. In the
event the estate only has 100 to pay out, each of the three creditors is paid 33 1/3. If the estate
has 200, A is paid 50, while B and C receive 75. If the estate has 300, A is paid 50, B is paid 100,
and C is paid 150.

It seems like there are three different approaches being taken. The first row suggests equal
treatment, the third row suggests proportional treatment, and the middle row just looks weird.

Estate Creditor A (100) Creditor B (200) Creditor C

100 33 1/3 33 1/3 33 1/3

200 50 75 75

300 50 100 150

As it turns out, all three rows are consistent with the Nucleolus.

First, look at the case where there is 300 to divide. If (B, C) get together, they can divide up 200
without any agreement from A. That is because (B, C) could write A a check for 100 and A would
have no more claim. Thus (B, C) can get 200 on their own, but they need to bring A onboard to
divide up that last 100. If they divide it evenly, then A’s gain from joining the group will be 50, the
same as the gain to (B, C). You can check that this is the smallest amount anyone or pair gets in
forming a threesome. For example, (A, B) on their own can get nothing as if they pay C off, that
leaves nothing for themselves. So if (A, B) want to get anything, they need to bring C into the
group. Similarly, C can get nothing on its own as paying off A and B would use up all the assets.
Thus bringing C together with (A, B) creates 300 to share. In the proposed solution, C gets 150
and (A, B) get 150, both of which are much bigger than 50.
The middle row looks more confusing. Here (B, C) can get 100 of the 200 without an agreement
from A (again by writing a check to A). Thus, just as in the last row, the addition of A is worth
100. If A gets 50 and (B, C) gets 50, then the two gains from joining are equal and they are also
tied for the lowest amount any party gets for joining another. For example, (A, C) can get nothing
on its own, so there is then 200 to split up when B joins (A,C). Here B gets 75 and (A, C) gets
125. To equalize this, B would have to get more and either A or C get less. But A can't get less
as it is only getting 50 when joining (B, C) and that is already the minimum. If C were to get more
than 75, then when C joins (A, B) that means B would have to get less than 75 (as A is getting
50)—but then B would be getting less for joining (A, C) than C would be getting.

If we go to the first row, no pair can get anything on their own. Even (B, C) without A gets nothing
as absent an agreement there is no ability to pay A his claim of 100 and have anything left over.
Thus if the payouts were anything but equal, then whoever got the lowest payout would get less
for joining the other pair. Only when the payouts are equalized do we maximize the smallest
payout.

If you are wondering how the Shapley Value would work in the Talmud case, just imagine that
the three creditors get in line in a random order. Each is paid off until the money runs out. Thus if
the order is (A, B, C) and the amount is 200, then A gets 100 for being first in line. There's still
100 left and B, as next in line, gets all of it leaving nothing for C. If the order is (C, A, B) then C
gets 200 while A and B each get nothing. And so on. The Shapley Value is the average of all
these possibilities.

There's one other way of explaining the Nucleolus solution. Recall our discussion of the Principle
of the Divided Cloth. The Nucleolus has the attractive feature that if we take away one of the
players and the amount that person gets then the other two players divide what's left according
to the Principle of the Divided Cloth. Consider, for example, the middle row. If we take away A
who is paid 50, then B and C have 150 to divide up. since they are claiming 200 and 300
respectively, each is claiming the entire cloth and so each is given half or 75. If we remove B with
his 75 payment, that leaves 125 for A and C to divide. A is only claiming 100 which concedes 25
to C; meanwhile C claims 300 conceding nothing to A. Thus 100 is in dispute, and this is split
50/50. So A gets 50 and C gets 50 plus the conceded 25 for 75 in total.

I don’t want to pretend this is a rigorous explanation of the Nucleolus, but I hope I've given you a
brief taste of why it is a very reasonable alternative to the Shapley Value. For more information
about the Nucleolus, have a look at the original article by David Schmeidler and an elegant
application of the Nucleolus to our airport cost-sharing problem by S.C. Littlechild. The
application of bargaining theory to problems from the Talmud starts with Barry O'Neill in his 1982
paper; the connection to the Nucleolus was developed by Nobel Prize winner Robert Aumann
and Michael Maschler.

Further Reading
Littlechild, S.C. “A Simple Expression for the Nucleolus in a Special Case.” Int. Journal of Game
Theory 3 (1974): 21–29. [LINK = ]

Schmeidler, David. “The Nucleolus of a Characteristic Function Game.” SIAM Journal on Applied

Mathematics 17 (1969): 1163–1170.

O'Neill, Barry. “A Problem of Rights Arbitration from the Talmud.” Mathematical Social

Sciences 2 (1982): 345-371. [LINK = ]

Aumann, Robert J. and Michael Maschler. “Game Theoretic Analysis of a Bankruptcy Problem
from the Talmud.” Journal of Economic Theory 36 (1985): 195-213.

Planet–Gazette Case

Merger Negotiations

The New Haven Planet and the Hartford Gazette are contemplating a merger. Roughly speaking,
newspapers are valued on a per-reader basis. The Planet has 100,000 subscribers, while
the Gazette is twice as large with 200,000 readers. Currently, the Planet has a market cap of
$10m, while the Gazette has a market cap of $22m.*

The reasons for a merger are as follows:

By combining their joint purchasing, the two papers expect to reduce paper and printing costs by
2.3%. The present discounted value of this cost savings is $2m to the Planet and $4m to the
Gazette.

The Gazette is starting from a position with lower production costs than the Planet. The cost
advantages of the Gazette can all be transferred to the Planet’s operation. The projected savings
are $125k annually to the Planet, which adds $1 million to the current value (or market cap) of
the company.

The merger will allow the two papers to cut overhead. The reduction in headcount is worth a total
of $150k annually, or $1.2m in present discounted value.

It is believed there is a possibility of expanding readership through joint subscription offers. It is

thought that 5% of Planet readers will start subscribing to the Gazette and 5% of Gazette readers
will start subscribing to the Planet. The value of the new 5,000 Gazette readers is worth $550k
(at current market prices), and the value of the new 10,000 Planet readers is worth $1m (at
current market prices). If we assume the Planet will have the same profitability as the Gazette
post-merger, then the new Planet readers will be worth $1.1m.
In addition, there is a reduced need for working capital, synergies in selling ads, and in building
the online paper. These synergies have not been quantified.

As a result of the merger discussions, both parties have all the information above. These two
papers have agreed on how the editorial side of the business will combine operations. The major
sticking point is the financial terms for the merger. The two sides agree the new combined entity
will have a market cap of $41.85 million.

Based on the relatively small size of these two papers, you should assume that there are no
other potential merger partners. No joint ventures are possible. Either the Planet and the Gazette
reach a deal or they don’t. If no deal is reached, then both sides continue with their business as
usual and none of the synergies are achieved.

IF YOUR BIRTHDAY IS BETWEEN JANUARY 1 & JUNE 30 THEN:

You are representing the Hartford Gazette in the merger. Your job is to make the best case for
paying a low price. What is the lowest purchase price you can justify, and how do you justify it?
What do you think is the fair outcome?

IF YOUR BIRTHDAY IS BETWEEN JULY 1 & DECEMBER 31 THEN:

You are representing the New Haven Planet in the merger. Your job is to make the best case for
getting a high price. What is the highest purchase price you can justify, and how do you justify it?
What do you think is the fair outcome?

________________________

* The Gazette is more profitable on a per-reader basis and thus has a higher market cap per
reader.

Once we were part of Coca-Cola, there was still a question of how to share the cost savings. The
way it is done at Coca-Cola—and I expect at most large companies—is that there is one central
group that does all the purchasing, and the operating units (like brands) are charged an annual
fee to cover the overhead costs of the purchasing group. The annual fee is typically tied to the
sales of the operating unit. This arrangement worked out particularly well for Honest Tea. The
sales were low, so Honest Tea’s share of the overhead cost was low. However, the savings were
large as Honest Tea's costs coming into Coca-Cola’s system were unusually high.

This is the flip of the first argument I mentioned in the video. Instead of dividing the benefits in
proportion to sales, we divide the costs in proportion to sales.