CS 364A Algorithmic Game Theory Stanford University

CS364A Exercise Set 2

Austin Liu SUNet ID: yliu2
April 19, 2019

AGT - Algorithmic Game Theory by Nisam, Roughgarden, Tardos, Vazirani

Quick summary:

Single parameter environment has n bidders, each with private valuation vi (“value per unit of stuff”),
and a feasible set X. Each element of X is an n-vector (x1 , x2 , . . . , xn ), xi is the “amount of stuff” given to
bidder i.
A sealed bid auction has steps:
• Collect bids b

• Choose allocation rule x(b) ∈ X ⊆ Rn

• Choose payment rule p(b) ∈ Rn
Use quasilinear utility model, so, in auction with allocation rule x, payment rule p, bidder i has utility

ui (b) = vi xi (b) − pi (b)

Focus on payment rules satisfying:

pi (b) ∈ [0, bi xi (b)]

where pi ≥ 0 prevents seller from paying bidders and pi ≤ bi xi ensures that truthtelling bidder receives
nonnegative utility.
Allocation rule x for single-parameter environment is implementable if there is payment rule p such that
sealed bid auction (x, p) is DSIC
An allocation rule x for a single-parameter environment is monotone if for every bidder i and bids b−i
by the other bidders, the allocation xi (z, b−i ) to i is nondecreasing in its bid z.
Theorem 1 (Myerson’s lemma). Fix a single-parameter environment.

(a) An allocation rule x is implementable if and only if it is monotone.

(b) If x is monotone, then there is a unique payment rule such that the sealed-bid mechanism (x, p) is
DSIC [assuming the normalization that bi = 0 implies pi (b) = 0].
(c) The payment rule in (b) is given by an explicit formula:
Z bi
dxi (z, b−i )
pi (bi , b−i ) = z dz
0 dz

allocation rule xG bids b feasible set S with total size

P
i∈S wi ≤ W (W is capacity)

Austin Liu: [email protected] 1

CS 364A Algorithmic Game Theory Stanford University

Greedy Knapsack algorithm

A. Index bidders such that b1 /w1 ≥ b2 /w2 ≥ · · · ≥ bn /wn
B. Take items until one doesn’t fit, then halt
C. Return step-2 solution or highest bidder, whichever creates more surplus
Theorem 2 (Theorem 2.1 in notes). Assuming truthful bids, the surplus of the greedy allocation rule is at
least 50% of the maximum-posible surplus.
Proof: Consider truthful bids v1 , . . . vn , known sizes w1 , . . . , wn , capacity W , allow a bidder to be
chosen fractionally. Solutions look like z ∗ = (1, . . . , 1, α, 0, . . . , 0) α is at the break item k and is α =
W −w1 −...wk
wk+1 (fill the remaining with the fraction of the item) Let optimum solution for Knapsack be x∗ ,
x = (1, . . . , 1, 0, 0, . . . , 0) y = (0, . . . , 0, 1, 0, . . . , 0)

val(x∗ ) ≤ val(z ∗ ) = val(x) + αvk+1 ≤ val(x) + val(y) ≤ 2 max{val(x), val(y)}

(the better solution will be at least half the size of the optimal fractional solution, else we get a contra-
diction)

Let k is first item not accepted by Greedy Knapsack If wi ≤ αW for every bidder i with α ∈ (0, 1/2]

For1 ≤ i ≤ k, vi /wi ≥ vk /wk

=⇒ v1 + v2 + · · · + vk ≥ (w1 + w2 + . . . wk )vk /wk
=⇒ vk ≤ wk (v1 + v2 + · · · + vk )/W (w1 + · · · + wk > W )
=⇒ vk ≤ α(v1 + v2 + · · · + vk ) (since wk /W ≤ α)
≤ α(v1 + v2 + · · · + vk−1 )/(1 − α)
Fourth line is because v1 + v2 + · · · + vk ≥ (w1 + w2 + . . . wk )vk /wk ≥ W vk /wk ≥ vk /α
so α(v1 + v2 + · · · + vk ) ≥ vk = v1 + v2 + · · · + vk − v1 + v2 + · · · + vk−1
so (v1 + v2 + · · · + vk−1 )/(1 − α) ≥ v1 + v2 + · · · + vk ≥ OPT
where OPT is the value of the optimal solution.
Hence v1 + . . . vk−1 ≥ (1 − α)OPT

Exercise 9
Want to show that Vickery auction is the unique single-item auction that is DSIC, always awards the good
to the highest bidder, and charges losers 0, using Myerson’s lemma.
By Myerson’s lemma (a) - auction is DSIC iff it is monotone, which is equivalent to awarding the good
to the highest bidder.
So the function xi (z, b−i ) is 0 up to B = maxj6=i bj , 1 thereafter. So the payment is either 0 for bi < B
or B for bi > B, losers are charged 0.

Exercise 10
Consider payment difference sandwich (4) in lecture 3 notes for 0 ≤ y < z: (where inequalities are from
invoking DSIC constraint - if x, p are DSIC, truthful bidding incentivized if zx(z) − p(z) ≥ zx(y) − p(y) if
private valuation is z, and yx(y) − p(y) ≥ yx(z) − p(z)
z[x(y) − x(z)] ≤ p(y) − p(z) ≤ y[x(y) − x(z)]
If allocation rule is not monotone, we can have x(y) > x(z). But z > y, so inequality is impossible to satisfy,
so p is not implementable.
Hence allocation rule implementable implies it is also monotone.

Austin Liu: [email protected] 2

CS 364A Algorithmic Game Theory Stanford University

Exercise 11
Used “proof by picture” to show that coupling monotone, piecewise allocation rule x with the payment
formula:
Z bi Z bi
pi (bi , b−i ) = z d[xi (z, b−i )] = bi xi (b) − xi (z, b−i )dz
0 0

where x is a distribution (not necessarily a function), yields a DSIC mechanism.

Proof-by-picture breaks down x is not monotone in the following way:
Let breakpoints in x be x1 , x2 , x3 , those in z be z1 , z2 , z3 in Fig. 2(b,e,h) with b > v. Suppose x drops to
the first step, x1 at v < z < z3 , then the z > b case has surplus vx1 , more payment z1 x1 − (v − z2 )(x2 − x1 ),
utility is surplus minus payment.
Honest case has utility b = v, utility is vx2 − z1 x1 − (z2 − z1 )(x2 − x1 ).
So if x2 is large enough, can have more utility in the non-monotone case (imagine a negative payment,
for instance).

Exercise 12
Proof that coupling a monotone and piecewise constant allocation rule x with the payment rule (1) yields a
DSIC mechanism.
Surplus is given by vi xi (b)
Z bi Z bi
pi (bi , b−i ) = z d[xi (z, b−i )] = bi xi (b) − xi (z, b−i )dz
0 0

First term is the rectangle of height xi (b), z goes from 0 to bi . Utility is surplus less price.
(Basically the pictures at https://www.cs.mcgill.ca/~cai/COMP_MATH_553/LectureNotes04.pdf)
If bi > vi , utility up to the truthful bid is the same, but always have utility loss (from increase in price)
Rb
of (bi − vi )xi (b) − vii xi (z, b−i )dz which is strictly positive if xi (z, b−i ) is increasing on [v, bi ]. (else the first
term may be smaller than the second). Rv
If bi < vi , always have utility change of of bi xi (b) − bi xi (z, b−i )dz which is strictly negative if xi (z, b−i )
is increasing on [bi , v].

Exercise 13
Each bidder i now has a publicly known quality βi (in addition to a private valuation vi per click). As usual,
each slot j has a CTR αj , and kslots have α1 ≥ α2 · · · ≥ αk . We assume that if bidder i (n bidders total)
is placed in slot j, its probability of a click is βi αj - thus, bidder i derives value vi βi αj from this outcome.
As before, order by bi βi , x(b) is the allocation rule assigning jth highest bidder to the jth highest slot,
for j = 1, 2, . . . k.
The allocation xi (z, b−i ) ranges from 0 to α1 as z ranges from 0 to b1 , with a jump of αj − αj+1 at the
point where z becomes the jth highest bid in the profile (z, b−i ), namely bj+1 βj+1 .

X
pi (b) = βj+1 bj+1 (αj − αj+1 )
j=i

Exercise 14
Consider an arbitrary single parameter
Pn environment, with feasible set X. The surplus-maximizing allocation
rule is x(b) = arg max(x1 ,...xn )∈X i=1 bi xi (arg max - arguments of the maxima, points of domain of function
at which function values are maximized).

Austin Liu: [email protected] 3

CS 364A Algorithmic Game Theory Stanford University

WLOG assume values are sorted x1 ≥ x2 · · · ≥ xn , b1 ≥ b2 · · · ≥ bn . Then the rule returns (x1 , x2 , . . . , xn )
in sorted order (greedy algorithm in Exercise 8).
WLOG consider xi (z, b−i ), WTS that this is non-decreasing in z, i.e. monotone.
Suppose there exists a better valuation with i, j swapped, xi ≥ xj , vi ≥ vj . Change in valuation is:

xj vi + xi vj − xi vi − xj vj = (xj − xi )(vi − vj )

Above quantity is negative, contradicting claim that social surpus can increase in this ordering. Hence initial
ordering maximizes social surplus.

Exercise 15
Consider Ex. 14, restrict to sets with only 0-1 vectors (each bidder either wins or loses). Feasible outcome
identified with “feasible set” of bidders (winners in that outcome). Assume further that the environment
is “downward closed”, meaning that subsets of a feasible set are again feasible. Recall from lecture that
Myerson’s payment formula dictates that a winning bidder pays its “critical bid” - the lowest bid at which
it would continue to win.
Prove that, when S ∗ is the set of winning bidders and i ∈ S ∗ , i’s critical bid equals the difference between
(i) the maximum surplus of a feasible set that excludes i (you should assume there is at least one such set);
and (ii) the surplus j∈S ∗ \{i} vj of the bidders other than i in the chosen outcome S ∗ . Also, is this difference
P
always nonnegative?
Order so x1 ≥ x2 · · · ≥ xn , b1 ≥ b2 · · · ≥ bn . DSIC so bids equal valuations.
(i) is maximum surplus
P of S ∗ \ {i}, which is v1 + v2 + · · · + vn−1
(ii) is surplus in j∈S ∗ \{i} vj = v1 + v2 + · · · + vn − vi
Difference is vi − vn . In the above sense, a winning bidder pays its “externality” - the surplus loss it
imposes on others.

Austin Liu: [email protected] 4