The Last Round of Betting in Poker Author(s): Jack Cassidy Source: The American Mathematical Monthly, Vol.

105, No. 9 (Nov., 1998), pp. 825-831 Published by: Mathematical Association of America Stable URL: http://www.jstor.org/stable/2589211 . Accessed: 22/12/2013 17:37
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp

.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

Mathematical Association of America is collaborating with JSTOR to digitize, preserve and extend access to The American Mathematical Monthly.

http://www.jstor.org

This content downloaded from 69.123.206.101 on Sun, 22 Dec 2013 17:37:12 PM All use subject to JSTOR Terms and Conditions

The Last Round ofBetting in Poker

JackCassidy
1. INTRODUCTION. Most mathematical analysesof poker make the simplifying assumption thatall playershave equal hands goingintothe finaldraw,or thatbet size equals pot size. We show that one can reach a more general resultwithout such assumptions. The method of "crossingpoints" is explained and systematically applied, to make the calculationsless cumbersome. A closed solutionto one special case is given,as well as a real-world poker example. The a-uthor was drivento thisworkby years of readingpoker books that,with the exception of [2],[4], and [5],were fullof advice thatwas eithertoo vague or too anecdotal. The techniquesused in this paper are applicable to any numberof players, but forthe sake of brevity, we'll limitthis discussionto a game withtwo playersleftin the finalround. The finalroundof betting in poker startsas Player1 (he) has the optionto bet or check (pass withoutconcedingthe pot). If he checks,Player 2 (she) gets the option to bet or check. If neitherplayerbets, the pot goes to the best hand-a showdown. If one playerbets, the othercan call, leading to a showdown, or can drop,concedingthe pot to the bettor.Raises are discussedlater in thispaper. It can be shownthatan optimalsimplestrategy existsforeach player.Assuming a continuous space of hand values,withzero representing an unbeatablehand, the strategiesare defined by three values-b, c, and d (mnemonically Bet, Call, Deceive). When a playerhas an option to bet, he or she bets withhands better than or equal to b (real bet) or worse than d (a bluff). A playercalls withhands betterthan or equal to c. Player i has a distribution functionrepresentedby Fi(x), defined as the probability that Player i has a hand that is betterthan or equal to hand x. The function is illustrated translation betweenprobabilities and the distribution by the thatPlayer1 bets: probability P(x < b1) + P(x ? d1) = F1(bl) + (1
-

F1(dj))

Player 2's callinghand, c2, is greaterthan or equal to (worse than) Player l's hand d1. This is bettinghand b1 and less than or equal to Player l's bluffing because, if c2 < b1,Player1 has a betterstrategy (bI, cl, d'), in whichbI = c2 and is dl = strategy F1(d1)= F1(d1)- (F1(b1)- F1(b'1)). If c2 > dl, a dominating and F1(bl) = F1(b1) + F1(dl) - F1(dj). Likewise,b2 ? c1 < d2. The functionGi(x) representsa player's distribution after the player has checked.
(x < bi = 0

Gi(x) = Fi(x1playeri checked) =bi

< x < di = F'(d)

x > di = 1

-F(b

1998]

THE LAST ROUND OF BETTING IN POKER

825

This content downloaded from 69.123.206.101 on Sun, 22 Dec 2013 17:37:12 PM All use subject to JSTOR Terms and Conditions

The moneyin the pot beforethe finalroundof betting beginsis denoted by a, and bet size is denoted by /3. The easiest way to calculate optimalstrategies is to use the methodof crossing points,suggestedas being "intuitively reasonable"by Zadeh [2, pp. 201-202]. This says thatthe optimalvalues for bi, ci, and di are the values at whichthe player's expectationEi(x) is equal foreitherof the possible actions.The graphin Figure 1 the situation illustrates forPlayer2.

Points Crossing
E2(bet)

E2(check)

best
b2

I
Cl

I worst
d2

HandValues 1. Player 2'sworst Figure betting handandbestbluffing handarefound where thegraphs cross.

The E2(check) line shows that if Player 2 checks everyhand, her expectation decreases linearly froma max of a forher best possible hand down to 0 forher worst.The E2(bet) line startshigher,because she wins more by bettingthan by a greathand. For intermediate hands she does worse by betting. checking Betting hand worse thanPlayerl's callingvalue c1 givesa constantexpectation any equal to a*(1 - G1(cl)) - j8*Gj(cj). is to bet in the partsof the graphwherebetting has a Player2's optimalstrategy thanchecking, and to checkwhere checking is better.In other higherexpectation words,she should locate b2 at the crossing pointon the left,and locate d2 at the crossing pointon the right. 2. THE SIX EQUATIONS. To findthe crossing pointford2, we need the pointat whichthe players'expectations are the same whetherPlayer2 bets or checksher hand. We'll startby using Player l's expectation, thoughwe could use either.If he has a betterhand. Player2 checks d2 (a bad hand), Player 1 wins a whenever If she bets d2, Player1 wins a + /3whenever he calls. E1(P2 checksd2) = E1(P2 bets d2)
a*Gl(d2) + 0*(1G(d2))
=

(a + j8)*Gj(cj)

+ 0*(1

Gl(cl))

G1(c1)

G1(d2)

(1: P1 call ratio)

826

THE LAST ROUND OF BETTING IN POKER

[November

This content downloaded from 69.123.206.101 on Sun, 22 Dec 2013 17:37:12 PM All use subject to JSTOR Terms and Conditions

When G1(d2) = 1, thisbecomes Zadeh's FundamentalCallingTheorem [3, Theorem 1]. The crossing between c1 and b2. pointfor b2 leads to a relationship E1(P2 checks b2) = E1(P2 bets b2) a*Gl(b2) + 0*(1
-

G1(b2)) = (a + 13)*Gl(b2)
1*(Gj(cj)
-

G1(b2))

+ 0*(1

G1(c1))

G1(c1) = 2G1(b2)

(2: winhalf)

hand b2, halfof Player l's calls This says thatwhen Player2 bets withborderline are betterand halfare worse. The c1 crossingpoint uses the fact that afterPlayer2 bets, we know that her hand was eitherbetterthan b2 or worse than d2. This calculationleads to Zadeh's Theorem[3, Theorem2]. FundamentalBluffing E1(P1 calls c1) = E1(P1 drops c1)

-/3(F2(b2)
1 - F2(d2)

+ (1 - F(d))
= F2(b2) a + /

F2(b2)

3(F(b)

1F2(d2)) + (1 -F2(d

)=

(3: P2 bluff ratio)

of a and /3. to real bets is a function Equation (3) says the optimalratio of bluffs solved a Von Neumannand Morgenstern [1, pp. 186-219] simplified, asymmetand solved for of poker.In our terminology, ric form theyset G1 = F2, set a = /3, c1, b2, and d2. Makingthese changesand solvingequations(1), (2), and (3) yields F(b2) = 2/9, F(d2) = 8/9, F(cl) = 4/9, whichis the solutionin [1]. To derivethe otherthreeequations needed forthe two person game,we'll use she Player2's expectation. Facing a bet of d1,she winsthe pot plus a bet whenever calls. Facing a check of d1, she wins the pot unless she checkswitha worse hand. between d1 and d2: There are two cases, dependingon the relationship E2(P1 bets d1) = E2(P1 checksd1) E2(P1 bets d1) = (a + /)F2(c2)
+ O*(1 - F2(C2))

-F2(b2)) +ba(F2(dl) (d, < d2 = a*F2(b) a + 0*(F2(d2) -F2(dl)) + a(1 - F2(d2 ) ) E2(P1 checksd1) = - F2(b2)) d, 2 d2= a*F2(+b2)+a(F2(d2) +a(1 -F2(d2))

( Jd1<
F2(C2)
=

d2

d a/ = 2 d2 ld, a+ 8

/ a + (1-

~~a

(F2(d2) - F2(dl)))

(4: P2 callratio)

We're alwaysrelievedwhen a case turnsout to have d1 ? d2, so we can use the simplerexpression. 1998]
THE LAST ROUND OF BETTING IN POKER

827

This content downloaded from 69.123.206.101 on Sun, 22 Dec 2013 17:37:12 PM All use subject to JSTOR Terms and Conditions

In the bet vs. check equation for b1, we can assume b2 ? b1, because if not, Player2 can improveher strategy more. by betting E2(P1 bets b1) = E2(P1 checks b1)
(a + f)F2(bl)
-

(F2(C2)

F2(bl))

+ 0*(
-

F2(c2)) -(1

(a + f)F2(bl) +o*(F2(d2)
fd F(c)
=

f(F2(b2)-

F2(bl))
-

F2(b2))

F2(d2))

-(F2(b2)

1 -F2(d2))

F2(C2)

F2(b2)

+ (1

- F2(d2))

(5: bets equal calls)

Note that b1 drops out of the equation, leavingthe resultthat Player 2's calling percentageexactly equals the sum of her bet and bluff percentages. The finalequation givesPlayer l's bluff ratio:

E2(P2 drops c2) = E2(P2 calls c2)

Fl( bl)
+ (1- Fl(dl)) + ,

p (Fl(bl) 1 - Fl(d1)
=

a+

1 - Fl(dl)
-

F1(b1) + (1

Fl(dl) )

Fl(b1)

(6: P1 bluff ratio)

This is Playerl's analogue to (3). withthese equations,thateach playeris trying to One gets the feeling, working protectthe otherfrom If Player2, forexample,uses the optimal makingmistakes. bluff between b2 and f2 without ratio,thenPlayer 1 can set c1 anywhere hurting This is because a minimax his results. strives to minimize one's own losses strategy the opponent'soptions.The optimalstrategy is a defensiveone, to be by limiting adjustedonce you see the weaknessesof youropponent. A Closed Solution. Equations (1)-(6) allow calculation of Equal Distributions: forthe finalround in any situation are where the distributions optimalstrategies In known.They are easily solved for some artificial distributions. mathematical the case of twoplayersstarting withequal distributions particular, (F1 = F2), leads to a closed solution: F(bl)= F(cl) (1 -F(dl)) F(b2)
F(C2) a2

(a+/3)(a+2/23)
=

(Playeribet) (Player 1 call)

a a 2/ a 2,8
(a
R O ) 2(a+2

2 )

(Player 1 bluff) 2bet)

a a+/3(Player

=
=

2 call) (Player
(lyrbuf

(1 -F(d2))

a/3(lae (a +/3)(a +2/3)

buf

This content downloaded from 69.123.206.101 on Sun, 22 Dec 2013 17:37:12 PM All use subject to JSTOR Terms and Conditions

We use these equations for Loball (five card draw,worst hand wins) as an approximation whenbothplayersdrawthe same numberof cards.If the bet size is about half the pot size, F(bj) = 0.33, F(cl) = 0.67, F(dj) = 0.89, F(b2) = 0.50, F(c2) = 0.67, and F(d2) = 0.83. Raising. Raising and reraising can introduceenough new variablesand distributionsto drivea mathematician to distraction. In [6], Cutlersolves the special case of unlimited raises when distributions are equal and pot size equals bet size. If we use generaldistributions, but simplify byallowing onlyone raise (byPlayer 2), the situationis manageable. To go withthe previously definedborderline values bi, ci, and di, we'll define 2 raiseswithe2 or better and bluff e2, f2, and g1.Player raises with f2 or worse. Player1 calls a raise withg1 or better.Note that e2 < g1 < b1 and f2 > d2. H1 is Playerl's distribution afterhe has bet. In our game,the bet size is b, so a raise costs 2,3.

raisef2) = E1(P2 dropf2) E1(P2bluff

(a + 3/)Hl(gl) = a + a
a

H1(g1) =
-

(7)

E1(P1 call g1) = E1(Pj drop g1)

f8*F2(e2) + (a + 3,3)(1 -F2(f2))
=

1 - F2(f2)
- /3(Hj(gj) - Hl(e2))

F2(e2) a + 3/3

/3

(8)

E1(P2 raise e2) = E1(P2 call e2)

+ (a + 3/3)Hl(e2) Hl(e2) = (a + 2/3)Hj(e2) =

-Hl(gl)

(9)

raises and real raises is smallerthanthe Equation (8) says the ratiobetweenbluff ratiobetweenbluff in that a and /3 bets and real bets. Equation (9) is interesting between e2 and g1. drop out, leavinga simplerelationship raises intothe picturealterstwo of the earlierequations,because it Introducing thatcan happen afterPlayer 1 bets. expandsthe set of things E2(P1 bets d1) = E2(P1 checksd1)
(a + /)(1
-F2(f2)) + (a + /)F2(c2) = a =

F2(c2) E2(P1 bets b1) = E2(P1 checksb1) (a + /)(1

-

a +

-(1

-F2(f2))

(4r)

F2(f2))

/3(F2(C2)
-

= -/3(1

F2(bl)) + (a + /)F2(bl) F2(d2)) - /3(F2(b2) - F2(bl))

-

F2(b2)

(1

- F2(d2))

+(a a
+

+ /3)F2(bj)
-

a + 2,/ a+/3

(1 (1

F2(f2))

F2(b2)

a
a

+ 2p-

- F2(f2))

(5r)

1998]

THE LAST ROUND OF BETTING IN POKER

829

This content downloaded from 69.123.206.101 on Sun, 22 Dec 2013 17:37:12 PM All use subject to JSTOR Terms and Conditions

The derivation of (4r) assumesthe simplecase, d2 < d1.The derivation of (5r) uses (4r) to substitute forF2(c2). Equations (4r) and (5r) revert to the earlierequations (4) and (5) in the case thatPlayer2 neverraises. Example: 7 Card Stud. These equations can be used forreal lifesituations, ifyou supplya reasonable approximation of the two players'distributions. As an example, we'll calculate optimalstrategy fora finalround situationin seven card stud (dealt two cards down,fourcards up, one card down). Player 1 shows K, 10,5, 3, and three of the cards are hearts. Player 2 shows 9 8, 8, Q, 3 withno hearts.There are 42 unseen cards. If Player1 has a fourflush, of thesegivehima Flush, 15 giveOne Pair, and 18 givea Bust. If he has a Pair but not a fourth heart,2 cards give Three of a Kind, 12 give Two Pair, and 28 leave himwithOne Pair. The betting fromearlierroundswould lead a reasonable observer to say there's a 50% chance thatPlayer 1 has a fourflush,a 25% chance he has a Pair, and a 25% chance he alreadyhas a Flush. Adding the probabilities togethergives the distribution F1(x) in Figure 2. The same reasonable observerwould say that Player 2 could have Two Pair (50% chance), Three of a Kind (25% chance), or One Pair (25% chance). Her probabilities, similarly calculated,yieldthe distribution F2(x). Four of a Kind Full House 0.00 0.00 0.01 0.11 Three Flush of a Kind Two Pair Pair Bust 0.36 0.37 0.44 0.79 1.00 0.11 0.31 1.00 1.00 0.83

F1(x) F2(x)

fora situation Figure2. Approximate distributions in 7 card stud.

The pot size is twicethe bet size, a = 2, /3= 1. We'll startby applying the six equations,without with(4). F2(c2) = a/(a + 8) = 2/3. Player2 raises,beginning calls witha hand of 8866xor better(where"x" represents a meaningless side card). Combining (3) and (5), we see that F2(b2) = 3/4*F2(c2), or 1/2, so Player 2 bets with a JJ88x or better.Applying (5) again, we see that (1 - F2(d2)) = 1/6. With Player 2's distribution, this means she bluffsany time she has only a pair, 88xxx (or worse). To get Player l's optimalstrategy, combine(1) and (2). Applying the relationship betweenGi(x) and Fi(x), we find 2 Fj(d2) F1(b1) 3 \F1(dj) -F1(bl)
-

\ /

(F1(b2) -F1(b1)
-

\ F1(d)

-F1(bl) F1(b1))

2 -(0.615

F1(b1)) = 2(0.405
Fl(bl)
=

0.30

about 10% Player1 bets with5/6 of his Flushes. Equation (6) sayshe shouldbluff of the time overall, that is, with half of his Busts. Applying(2) again gives Fl(c1) = 0.51. Player1 should call withKKxxxor better. Adding the raise equations to the analysisyields no surprises.Player 2 raises withany fullhouse, and bluff raises with 13% of her pairs. Player 1 calls a raise withthe best 2/3 of his flushes. but b2 drops fromJacks Player2's callingvalue c2 doesn't change noticeably, and Eights to Tens and Eights. She bluffsa little less, with 88% of her pairs. 830
THE LAST ROUND OF BETTING IN POKER

[November

This content downloaded from 69.123.206.101 on Sun, 22 Dec 2013 17:37:12 PM All use subject to JSTOR Terms and Conditions

The values b1, c1, and d1 do not change in this example,because F1(b2) doesn't change much. raises and stillget the same result. For thisexample,we could allow unlimited Player l's best hand (a Flush) is not good enough to reraise afterPlayer 2 has raised,probably witha Full House. step. You can use the optimalstrategy Calculatingoptimalstrategy is onlya first in a game withunknown players.But as you play,compareyouropponents'actual too infrequently? Then call less often.Are strategy to the optimal.Do theybluff more. theytoo slow to call? Then bluff It may be difficult to keep trackof the habits of each individualplayer at a the overall table. But it's easy to keep trackof the table as a whole. By counting of calls to drops and bluffs to real bets,you usuallyfinda non-optimal proportion in the grouppercentages. Use thatto guideyourplay.It's whenyou take tendency advantageof non-optimal play by othersthatyou win consistently. CONCLUSION. The method of crossingpoints can be applied to a generalized raises, twoplayerpokerproblem.This yieldssix equationsforoptimalplaywithout and threemore equations fora game withone raise allowed. these equationsgiveprecisebetting When applied to a practicalpokerproblem, equal distribution case, theyproduce a guidelines. When applied to the theoretical closed solutioninvolving the ratiobetweenbet size and pot size. fromthe Effective use of these resultsin actual poker play requires straying to take advantageof the opponents'non-optimal tendencies. optimalstrategy
The author thanks Jim for their helpinpreparAmidei andJanice Steinberg ACKNOWLEDGMENT. ingthis paper.
REFERENCES Theoryof Games and Economic Behavior, Princeton 1. J. von Neumann and 0. Morgenstern, University Press, 1944. Prentice-Hall, Englewood Cliffs, N.J.,1974. 2. N. Zadeh, Winning PokerSystems, Oper. Res. 25 (1977) 541-562. 3. N. Zadeh, Computation of OptimalPoker Strategies, Las Vegas, NV, 1994. 4. D. Sklansky, The Theory ofPoker,Two Plus Two Publishing, With Game Theory, Basic Books, New York, N.Y., 1981. 5. N. C. Ankeny, PokerStrategy-Winning forPot LimitPoker, Amer. Math. Monthly 82 (1975) 368-376. 6. W. H. Cutler,An OptimalStrategy

Chance. 3651 Third Ave., San Diego, CA 92103 cassidy@sdd.hp.com

AppleII utility University. He haswritten a BA inmathematics from Cornell JACK CASSIDYreceived charging simulations, and spenta year spacecraft programs forBeagle Bros.,createdelectrostatic OfficeJet firmware for Hewlett-Packard California. He nowdevelops playing Loballpoker in Gardena, a published is married tomystery author Janice Steinberg. He previously Jack printer/scanner/copiers. fiction no advice or mathematical content), Winning AtPoker and GamesOf bookofshort (containing

1998]

THE LAST ROUND OF BETTING IN POKER

831

This content downloaded from 69.123.206.101 on Sun, 22 Dec 2013 17:37:12 PM All use subject to JSTOR Terms and Conditions