Refinements

Intro from ECO 7404 Notes

Refinements in Strategic Form:
Eliminate Weakly Dominated Strategies - throwing out strictly dominated strategies (even
iteratively) never eliminates a Nash equilibrium (because Nash equilibrium never has
player using a strictly dominated strategy); goal is to eliminate weakly dominated
strategies to get a better defined equilibrium
Trembling Hand Perfect - consider possibility that players make random errors; look at limit
of probabilities of errors to get subset of original Nash equilibria
Theorem - finite, 2 player game, trembling hand perfect and eliminating weakly
dominated strategies are the same
Refinements in Extensive Form:
Tree Rules - a little more in depth on the details behind extensive form
Successor - nodes that can be reached from a given node by following arrows
Immediate Successor - node that's at end of any arrow leading away from a given node
Predecessor - analogous to successors except we trace backward through the tree;
also applies to immediate predecessor
Path - sequence of nodes that (1) starts with the initial node, (2) ends with a terminal
node, and (3) has the property that successive nodes in the sequence are immediate
successors of each other
Rule 1 - every node is a successor of the initial node, and the initial node is the only
one with this property
Rule 2 - each node except the initial node has exactly one immediate predecessor;
the initial node has no predecessors
Rule 3 - multiple branches extending from the same node have different action
labels
Rule 4 - each information set contains decision nodes for only one of the players; if
this weren't the case, at some point in the game, the players won't know who is
to make a decision
Rule 5 - all nodes in a given information set must have the same number of
immediate successors and they must have the same set of action labels on the
branches leading to these successors; if this weren't the case, the player would
be able to distinguish between the nodes
Equilibrium Path - what we observe if we watch the game
Subgame - can't break up an information set and has to start from a single node
Subgame Perfect - if game comes to initial node of a subgame, the game will progress as if
it were the subgame so we should find Nash equilibria to all subgames (don't believe
people will do irrational things); closely related to throwing out weakly dominated
strategies Equilibrium
Sequential Rationality - players ought to demonstrate rationality whenever they path
are called on to make decisions; optimal strategy for a player should maximize Belief
his or her expected payoff, conditional on every information set at which this needed
player has the move; closely related to trembling hand

New notes follow on next page

1 of 8
 A A prefers D,
but only if B
Sequential Equilibrium - builds on subgame perfection U D
plays L
Nash Equilibrium - only addresses the equilibrium path
Nash eq. says
Subgame Perfection - says all subgames have equilbrium B doesn't care B prefers L
B
play about play in L R when A plays
Kreps & Wilson - apply equilibrium play to all other areas this subgame D
(including those where there are no subgames);
conditional on getting to any point in the tree, players A
procede with optimizing behavior U D
Actions & Beliefs - have to specify actions and beliefs B No
L R subgames
at each node; actions are optimizing with respect to
beleifs; beleifs must be consistent with structure of
the game

Finding Nash Equilibria - easier to do in strategic form (for finite games); it's easy to overlook
a Nash equilibrium in extensive form
1 Player 2
L R
Example 1 - Consider this simple, 2-player game: L R

Player 1
2
Put it in strategic form and we see there are 2 L R (1,3) L 2, 2 0, 0
Nash equilibria: LL and RR R 1, 3 1, 3
Which is better? (2,2) (0,0)
In strategic form, player 2 has R weakly dominated by L so we toss the RR equilibrium
In extensive form, RR is not subgame perfect (if player 2 finds himself at the decision after
player 1 chooses L, player 2 would be irrational to pick R over L)
Player 2
1 1
L R L R L R
Player 1

L 2, 2 0, 0 2
L R (1,3) (2,2) (1,3)
R 1, 3 1, 3
(2,2) (0,0) Subgame Perfection
Weakly Dominated

Commitment - if player 2 can figure out a way to commit to R, he can force player 1 to choose
R (instead of L) so player 2 would get a payoff of 3 instead of 2; views on commitment:
(a) some say it's independent of the game tree
(b) commitment must be built into the game tree... we'll use this
2
Example 2 - modify example 1 by allowing player 2 to post a bond for C N
commitment prior to player 1's decision (i.e., player 2 decides 1 1
L R L R
whether or not to post a bond; if he does (C), he pays a small fee (ε) 2 2
and incurs a large penalty for choosing L... effectively, the bond L R (1,3-ε) L R (1,3)
allows player 2 to change his payoffs for his treat of playing R is
credible) (2,-1) (0,-ε) (2,2) (0,0)
Put this tree in strategic form to find all the equilibria
Note 1: player 1's strategies are listed as XY, where X Player 2
is the strategy played if player 2 chooses C and Y NL NR CL CR
is the strategy played if player 2 chooses N LL 2, 2 0, 0 2, -1 0, 0
Player 1

Note 2: we're ignoring ε in the payoffs to save space LR 2, 2 0, 0 1, 3 1, 3

There are 5 of Nash equilibria
RL 1, 3 1, 3 2, -1 0, 0
RR 1, 3 1, 3 1, 3 1, 3

2 of 8
"Best Equilibrium" (i.e., the most likely or most logical, not necessarily the Pareto optimal)
In strategic form, we can use iterated weak dominance to find that (LR,CR) is the best
Player 2
NL NR CL CR Weak Dominance
NL over NR
LL 2, 2 0, 0 2, -1 0, 0
Player 1
CR over CL
LR 2, 2 0, 0 1, 3 1, 3
Iterated Weak Dominance
RL 1, 3 1, 3 2, -1 0, 0 LR over RR (with NR gone)
RR 1, 3 1, 3 1, 3 1, 3 LR over LL (with CL gone)

2 2
Extensive form again uses subgame perfection to C N C N
get the same result 1 1
L R L R (1,3) (2,2)
Complication - if 2 can commit, can 1 also commit? 2 2
L R (1,3-ε) L R (1,3) Subgame Perfection
If so, game tree should show it; we can also
have multiple levels of commitment so this (2,-1) (0,-ε) (2,2) (0,0)
method can get very complicated
Commitment in Game Tree - once the game tree is written, don't talk about actions that aren't
reflected in the tree (i.e., if a player doesn't have commitment devices shown in the tree,
don't talk about the player using commitment to get to his preferred equilibrium)

Problem with Subgame Perfection - (we're leading up to sequential equilibrium), subgame

perfect doesn't apply if there are no proper subgames as in example 3 1
b1
a1
Example 3 - there are no subgames so every Nash equilibrium is subgame perfect 2
a2 b2
Strategic form - Slutsky did this on the fly and admitted that his choice was not
3
the best to work with (i.e., making player 3 the row player, 2 the column a3 b3 a3 b3
player and 1 the matrix player); there are two equilibria: 1
(a1 , b2 , a3 ) and (b1 , b2 , b3 ) 1
3 0 4 0
1
2 0 4 0
Player 2 Player 2
2 0 0 1
a2 b2 a2 b2
Player 3

Player 3

a3 3,2,2 3,2,2 a3 4,4,0 1,1,1

b3 0,0,0 0,0,0 b3 0,0,1 1,1,1

a1 b1
Matrix (plr 1) payoff
Column (plr 2) payoff Player 1
Row (plr 3) payoff

Kreps & Wilson - argue that not all Nash equilibria are equally compelling; look for players
doing things off the equilibrium path that are irrational
(a1, b2, a3) Equilibrium - player 2 isn't on the equilibrium path, so check if he's rational
playing b2 given his belief that player 1 plays a1 and player 3 plays a3 ; the only way
player 2 gets to make a decision is if player 1 deviates (i.e., plays b1 instead of a1 );
there's no reason to think player 3 changes (he doesn't know player 1 deviated), so
given player 2 gets his choice and he expects player 3 to play a3 , player 2 is best off
playing a 2 (payoff of 4 instead of 1)

3 of 8
Trembling Hand - the strategic form version of what Kreps & Wilson argued (sort of)
(a1, b2, a3) Equilibrium - we should be able to eliminate this; define:
ε = Pr[player 1 deviates] (i.e., plays b1 instead of a1 )
δ = Pr[player 3 deviates] (i.e., plays b3 instead of a3 )
Assume ε and δ are small
1
Consider gain to player 2 by playing deviating (i.e., plays a 2 instead of b2 ): b1
a1
(Note: "deviation" for player's 1 & 3 are random errors, but for player 2 2
a2 b2
it's a choice) 3
Condition Prob ∆Payoff a3 b3 a3 b3 1
1 & 3 don't deviate (1 − ε )(1 − δ ) 2−2=0
1
1 deviates (3 doesn't) ε (1 − δ ) 4−1=3 3 0 4 0
1
2 0 4 0
3 deviates (1 doesn't) (1 − ε )δ 0−0=0
2 0 0 1
1 & 3 deviate εδ 0 − 1 = −1
Expected Gain ε (1 − δ )3 − εδ = 3ε − 3εδ − εδ = 3ε − 4εδ = ε (3 − 4δ )
∴ player 2 gains if 3 − 4δ > 0 δ < 3/ 4
We assumed ε and δ are small so player 2 will prefer to deviate and (a1 , b2 , a3 )
is not a trembling hand perfect equilibrium
Note: 1 & 3 deviate is "second order small" ( εδ ) so that loss isn't a big deal
(b1, b2, b3) Equilibrium - define:
ε = Pr[player 1 deviates] (i.e., plays a1 instead of b1 )
δ = Pr[player 3 deviates] (i.e., plays a3 instead of b3 )
Assume ε and δ are small
Consider gain to player 2 by playing deviating (i.e., plays a 2 instead of b2 ):
Condition Prob ∆Payoff
1 & 3 don't deviate (1 − ε )(1 − δ ) 0 − 1 = −1
1 deviates (3 doesn't) ε (1 − δ ) 0−0=0
3 deviates (1 doesn't) (1 − ε )δ 4−1=3
1 & 3 deviate εδ 2−2=0
Expected Gain − (1 − ε )(1 − δ ) + (1 − ε )δ 3 = (1 − ε )[− 1 + δ + 3δ ] = (1 − ε )(−1 + 4δ )
∴ player 2 gains if − 1 + 4δ > 0 δ > 1/ 4
We assumed ε and δ are small so player 2 will not prefer to deviate
Note: we only have to find one player that wants to deviate so for the (a1 , b2 , a3 )
equilibrium we were able to stop; for this equilibrium we still have to check if
player 1 or player 3 want to deviate
Check player 3; define:
ε = Pr[player 1 deviates] (i.e., plays a1 instead of b1 )
δ = Pr[player 2 deviates] (i.e., plays a 2 instead of b2 )
Assume ε and δ are small
Consider gain to player 3 by playing deviating (i.e., plays a3 instead of b3 ):

4 of 8
 1
b1
a1
2
Condition Prob ∆Payoff a2 b2
1 & 2 don't deviate (1 − ε )(1 − δ ) 1−1=0 3
a3 b3 a3 b3
1 deviates (2 doesn't) ε (1 − δ ) 2−0=2 1

(1 − ε )δ
1
2 deviates (1 doesn't) 0 − 1 = −1 3 0 4 0
1
1 & 2 deviate εδ 2−0=2 2 0 4 0

Expected Gain ε (1 − δ ) 2 − (1 − ε )δ + εδ 2 = 2ε − 2εδ − δ + εδ + εδ 2 = 2 0 0 1

2ε − δ + εδ
∴ player 3 gains if 2ε − δ > 0 ε >δ /2
So if player 2 is more than twice as likely to make an error as player 1 then
player 3 will gain by deviating... not exactly concrete; need to know player 3's
beliefs about ε and δ

Trembling Hand Perfect - basic idea is introducing a little noise to the game; different ways to
do it (strategic vs. extensive form); players have beliefs in what opponents will do based on
Nash equilibrium, but they also consider the chance of mistakes
Error - define error as a mixed strategy E i where some positive probability is assigned to
every strategy of player i ; by assumption this error does not depend on the player's
equilibrium strategy; examples where errors are dependent: 20
1
Football - can choose strategy to pass or run; errors are different (incomplete pass or
interception if player passes; fumble if player runs)
Darts - aim at 20 and probability of hitting 1 (error) is more than if player aims at 19 19
Chance of Error - with some small probability ε , player i makes an error (i.e., plays the
mixed strategy E i instead of his Nash equilibrium strategy X i )
Sequence - look at error as a sequence ε n with lim ε n = 0
n →∞
Not Serious - not taking errors seriously; not modeling it and don't care what causes it
because ε is small
Overall Mixed Strategy - pick an ε (or some n in the sequence); combining player's
equilibrium strategy and error results is the mixed strategy (1 − ε n ) X i + ε n E i (note limit
as n → ∞ is equal to X i )
Example - assume player 1 has three strategies x1 , x 2 , x3 and player has pure strategy
equilibrium x1 ; overall mixed strategy is:
1 E11 1 − ε + εE11
(1 − ε ) 0 + ε E31 = εE31
0 E31 εE31
New Equilibrium - this overall mixed strategy game that results from incorporating errors
satisfies the assumptions of the Nash theorem so we can find a new Nash equilibrium
using errors for all the players NE ( n , E) , where:
n
is matrix of all error probability sequences for each player

E is matrix of all mixed strategy vectors for each player

General Case - this formulation allows errors to affect everyone; Zelton's original work
had player's own errors affect rivals only (not himself)

5 of 8
 Trembling Hand Perfect Equilibria - for some n in the sequence lim NE ( n
, E) ⊆ NE(0)
n →∞

∴ set of trembling hand perfect equilibria is ∪ lim NE ( n

, E) ⊆ NE(0)
E n →∞
Subgame Perfect - since everything has some positive probability in the error, trembling
hand perfect equilibria looks at player decisions at every node; that means these
equilibria are subgame perfect

Sequential Equilibrium - similar to trembling hand, but not exactly the same; start with Nash
equilibrium and use equilibrium strategies to define beliefs at all nodes (including those not
on the equilibrium path and those with zero probability of occurring); there are lots of rules
for how to do this
Changing Beliefs - once deviation is observed (i.e., end up on non-equilibrium path)
consider how a player's beliefs change
Consistency - beliefs must be consistent with equilibrium strategies, including subgame
perfect equilibrium strategies (if they get you to that information set); if player is on a
node that's not part of any equilibrium path, then beliefs can be arbitrary
1 1 1

2 2 2 2 2 2

3 3 3

Beliefs (0,1,0) Beliefs (1/3,1/3,1/3) Beliefs (1,0)

Set of consistent beliefs: ∪ lim Beliefs( n

, E)
E n →∞
Mixed Strategies - if a player is using a mixed strategy in equilibrium, beliefs must be
consistent with that strategy

1 Player 2 must have beliefs consistent with

0.9 0.1
player 1's strategy ∴ beliefs are (0.9,0.1)
2
L R L R Given player 2's dominant strategy, player 3 would never
3 expect to be here during the game (zero probability event);
still has to be consistent with player 2's information set (add
errors for player 2 and take limit of error going to zero)
∴ Beliefs (0.9,0.1)

Difference -
Trembling hand puts in errors, finds NE, then takes limit of equilibria
Sequential puts in errors, puts in beliefs, takes limit of beliefs, the find equilibrium
Their both very similar, but examples can be constructed where they're different
Trembling hand imposes more restrictions so TH ⊂ Seq ⊂ NE
Benefit of Seq - emphasizes using actions and beliefs (TH does too, but not explicitly); this
is a good bridge to incomplete information
Example - shows the difference between sequential and trembling hand
Trembling Hand - effectively generates a new payoff 1 Player 2
matrix based on what strategies the players intend to U D
L R
Player 1

play (in this case we have to check (D,L) and (U,R)) 2

L R L R U 1, 1 1, 1
D 2, 0 -1, -3
1, 1, 2, -1,
6 of 8 1 1 0 -1
 Expected Payoff - define (x, y ) as the expected payoff vector when player 1 plays
mixed strategy x and player 2 plays mixed strategy vector y
Player 2
Aim at L Aim at R
Player 1 Aim at U ((1 − δ )U + δD, (1 − ε )L + εR ) ((1 − δ )U + δD, (1 − ε )R + εL)

Aim at D

((1 − δ )D + δU, (1 − ε )L + εR )

((1 − δ )D + δU, (1 − ε )R + εL )
In this case, we don't really need to solve all the boxes (it gets very tedious in larger
problems); we can just examine each Nash equilibrium to see if it's still an
equilibrium
(U,R) - determine if player 2 really prefers playing R:
EV L = (1 − δ )(1) + δ (0) = 1 − δ
EV R = (1 − δ )(1) + δ (−1) = 1 − 2δ
EV L > EV R so player 2 actually prefers L... (U,R) is not trembling hand perfect
(D,L) - determine if player 2 really prefers playing L:
EV L = δ (1) + (1 − δ )(0) = δ
EV R = δ (1) + (1 − δ )(−1) = 2δ − 1
EV L > EV R (recall δ ∈ [0,1] because it's a probability, but it's assumed to be small)
so player 2 does prefer L; now check for row player:
EV U = (1 − ε )(1) + ε (1) = 1
EV D = (1 − ε )(2) + ε (−1) = 2 − 3ε
EV D > EV U (as long as ε < 1 / 3 ) so player 1 does prefer D
∴ (D,L) is trembling hand perfect
Limit of Equilibria - effectively took limit (wrt error probabilities) of equilibria
Gets Hard - with more than two strategies, need to look at union of all distributions of
errors
Sequential - takes limit of beliefs (usually easier than TH)
Belief - let π n ( x) be sequence of beliefs of column player about row player choosing
strategy x , and θ n ( y ) be sequence of beliefs of row player about column player
choosing strategy y
Consistent, Not Equilibrium - beliefs to do not have to be equilibrium beliefs, but
have to be consistent in the limit (i.e., as n → ∞ , the beliefs have to agree with
equilibrium)
For this case, consider:
π n ( U) = 1 − 1 / n θ n (R ) = 1 − 1 / n
π n ( D) = 1 / n θ n ( L) = 1 / n
Limits as n → ∞ :
π ∞ ( U ) = 1 , π ∞ ( D) = 0 θ ∞ ( R ) = 1 , θ ∞ ( L) = 0
This is consistent because (U,R) is a Nash equilibrium
Sequential Equilibrium - specifies two things: actions and beliefs:
SE = (( U, R ); π ( U) = 1,θ (L) = 0 ) ... action: (U.R); beliefs π ∞ ( U) = 1 & θ ∞ (L) = 0
Rationality - given belief (taken at the limit), is action a Nash equilibrium
Consistency - limit of beliefs could result from fully mixed errors

7 of 8
 In this case, sequential does nothing because (D,L) and (U,R) are sequent equilibria
Kreps & Wilson Theorem - paraphrased by Slutsky: typically, sequential and trembling hand
equilibria are the same; for small sets of games ("measure zero") they're not the same; in
those cases, trembling hand eliminates more than sequential (TH ⊂ Seq ⊂ Nash);
theorem is very hard to prove

Summary -
For 2 players: IWD ⊂ TH = WD ⊂ Seq ⊂ Nash
- TH is exactly equivalent to single round of eliminating weakly dominated strategies
- Iteratively eliminating weakly dominated strategies eliminates more than TH
More than 2 players: TH ⊂ WD ⊂ Nash; TH ⊂ Seq ⊂ Nash; can't compare WD & Seq or
anything with IWD
- TH eliminates more equilibria than a single round of eliminating weakly dominated
strategies

-------------
Aside - classic game theory movies:
Dr. Strangelove - credibility
The Maltese Falcon - the book by Dashiell Hammett has a sub-plot about husband
disappearing (not in the movie) that deals with trembling hand (error doesn't mean game
is changed)

8 of 8