Calibrated Learning and Correlated Equilibrium

Size: px
Start display at page:

Download "Calibrated Learning and Correlated Equilibrium"

Transcription

1 Unversty of Pennsylvana ScholarlyCommons Statstcs Papers Wharton Faculty Research Calbrated Learnng and Correlated Equlbrum Dean P. Foster Unversty of Pennsylvana Rakesh V. Vohra Follow ths and addtonal works at: Part of the Behavoral Economcs Commons, and the Statstcs and Probablty Commons Recommended Ctaton Foster, D. P., & Vohra, R. V. (1997). Calbrated Learnng and Correlated Equlbrum. Games and Economc Behavor, 21 (1-2), Ths paper s posted at ScholarlyCommons. For more nformaton, please contact repostory@pobox.upenn.edu.

2 Calbrated Learnng and Correlated Equlbrum Abstract Suppose two players repeatedly meet each other to play a game where: 1. each uses a learnng rule wth the property that t s a calbrated forecast of the other's plays, and 2. each plays a myopc best response to ths forecast dstrbuton. Then, the lmt ponts of the sequence of plays are correlated equlbra. In fact, for each correlated equlbrum there s some calbrated learnng rule that the players can use whch results n ther playng ths correlated equlbrum n the lmt. Thus, the statstcal concept of a calbraton s strongly related to the game theoretc concept of correlated equlbrum. Dscplnes Behavoral Economcs Statstcs and Probablty Ths journal artcle s avalable at ScholarlyCommons:

3 Calbrated Learnng and Correlated Equlbrum Dean P. Foster Unversty of Pennsylvana Phladelpha, PA Rakesh V. Vohra Oho State Unversty Columbus OH y Frst draft: May 1992, Revsed: June 1993, Ths verson: October 17, 1996 Emal:foster@hellspark.wharton.upenn.edu y Emal:vohra.1@osu.edu 1

4 Abstract Suppose two players meet each other n a repeated game where: 1. each uses a learnng rule wth the property that t s a calbrated forecast of the others plays, and 2. each plays a best response to ths forecast dstrbuton. Then, the lmt pont of the sequence of plays are Correlated Equlbra. In fact, for each Correlated equlbrum there s some calbrated learnng rule that the players can use whch result n ther playng ths correlated equlbrum n the lmt. Thus, the statstcal concept of calbraton s strongly related to the game theoretc concept of correlated equlbrum. 2

5 1 Introducton The concept of a Nash Equlbrum (NE) s so mportant to game theory that an extensve lterature devoted to ts defense and advancement exsts. Even so, there are aspects of the Nash equlbrum concept that are puzzlng. One s why any player should assume that the other wll play ther Nash equlbrum strategy? Aumann (1987) says: \Ths s partcularly perplexng when, as often happens, there are multple equlbra; but t has consderable force even when the equlbrum s unque." One resoluton s to argue that the assumpton about an opponent's plays are the outcome of some learnng process (see for example Chapter 6 of Kreps (1991a)). Learnng s modeled as recurrent updatng. Players choose a best reply on the bass of ther forecasts of ther opponents future choces. Forecasts are descrbed as a functon of prevous plays n the repeated game. Much attenton has focused on developng forecast rules by whch a Nash equlbrum (or ts renements) may be learned. Many rules have been proposed and convergence to Nash equlbrum has been establshed under certan condtons (see Skyrms 1990). For example, Fudenberg and Kreps (1991) ntroduce the class of rules satsfyng a property called `asymptotc myopc bayes.' They prove that f convergence takes place, t does so to a NE. Notce that convergence s not guaranteed. In summarzng other approaches, Kreps (1991b) ponts out, \n general convergence s not assured." Ths lack of convergence serves to lessen the mportance of NE and ts renements. On the postve sde Mlgrom and Roberts (1991) have shown that any learnng rule that requres the player to make approxmately best responses consstent wth ther expectatons, play tends towards the serally undomnated set of strateges. They call such learnng rules adaptve and prove that 3

6 f the sequence of plays converges to a NE (or correlated equlbrum) then each players play s consstent wth adaptve learnng. Learnng, as we have descrbed t, takes place at the level of the ndvdual. An mportant class of learnng models nvolve learnng at the level of populatons (evolutonary models). Here the derent strateges are represented by ndvduals n the populaton. In partcular a mxed strategy would be represented by assgnng an approprate fracton of the populaton to each strategy. A par of ndvduals s selected at random to play the game. Indvduals do not update ther strateges but ther numbers wax and wane accordng to ther average (sutably dened) payo. Even n ths envronment convergence to a NE s not guaranteed. On the postve sde, results analogous to Mlgrom and Roberts have been obtaned by Samuelson and Zheng (1992). A second objecton to NE s that t s nconsstent wth the Bayesan perspectve. A Bayesan player starts wth a pror over what ther opponent wll select and chooses a best response to that. To argue that Bayesans should play the NE of the game s to nsst that they each choose a partcular pror. Aumann (1987) has gone further and argued that the soluton concept consstent wth the Bayesan perspectve s not NE but Correlated Equlbrum (CE). Support for such a vew can be found n Nau and McCardle (1990) who characterze CE n terms of the no arbtrage condton so beloved by Bayesans. Also, Kala and Lehrer (1994) show that Bayesan players wth uncontradcted belefs learn a correlated equlbrum. In ths note, we provde a drect lnk between the Bayesan belefs of players to the concluson that they wll play a CE. We do ths by showng that a CE can be `learned'. We do not partcular a specc learnng rule, rather, we restrct our attenton to learnng rules that possess an asymptotc property 4

7 called calbraton. The key result s that f players use any forecastng rule wth the property of beng calbrated, then, n repeated plays of the game, the lmt ponts of the sequence of plays are correlated equlbra. The game theoretc mportance of calbraton follows from a theorem of Dawd (1992). Gven the Bayesans pror look at the forecasts generated by the posteror. The sequences of future events on whch ths forecast wll not be calbrated, have measure zero. That s the Bayesan's pror assgns probablty zero to such outcomes. Thus, under the common pror assumpton, a bayesan would expect all the other players to be usng ther posteror, and hence to be calbrated. Now usng our result that calbraton mples correlated equlbra, and the common pror assumpton shows that bayesans expect that n the lmt, they wll be playng a correlated equlbrum. Ths provdes an alternatve prove to Aumann's proof that the common pror assumpton and ratonalty mples a correlated equlbrum. If the common pror assumpton holds then t s common knowledge that all players are calbrated. If the players use a Bayesan forecastng scheme that s calbrated, then, by the above, n repeated plays of the game, the lmt ponts of the sequence of plays are correlated equlbra. In the next secton of ths paper we ntroduce notaton and provde a rgorous denton of some of the terms used n the ntroducton. Subsequently we state and prove the man result of our paper. For ease of exposton we consder only the 2-person case. However, our results generalze easly to the n-person case. 1 1 See dscusson after Theorem 3. 5

8 2 Notaton and Dentons For = 1; 2, denote by S() the nte set of pure strateges of player and by u (x; y) 2 < the payo to player where x 2 S(1) and y 2 S(2). Let m = js(1)j and n = js(2)j. A correlated strategy s a functon h from a nte probablty space? nto S(1) S(2),.e., h = (h 1 ; h 2 ) s a random varable whose values are pars of strateges, one from S(1) and the other from S(2). Note that f h s a correlated strategy, then u (h 1 ; h 2 ), s a real valued random varable. So as to understand the denton of a correlated equlbrum, magne an umpre who announces to both players what? and h are. Chance chooses an element g 2? and hands t to the umpre who computes h(g). The umpre then reveals h (g) to player only and nothng more. Denton: A correlated strategy h s called a correlated equlbrum f: E ( u 1 (h 1 ; h 2 ) ) E ( u 1 ((h 1 ); h 2 ) ) for all : S(1)! S(1); and, E ( u 2 (h 1 ; h 2 ) ) E ( u 2 (h 1 ; (h 2 )) ) for all : S(2)! S(2); Thus, a CE s acheved when no player can gan by devatng from the umpre's recommendaton, assumng the other player wll not devate ether. The devatons, are restrcted to be functons of h because player knows only h (g). For more on CE see Aumann (1974) and Aumann (1987). We turn now to the noton of calbraton. Ths s one of a number of crtera used to evaluate the relablty of a probablty forecast. It has been argued by a number of wrters (see Dawd (1982)) that calbraton s an 6

9 appealng mnmal condton that any respectable probablty forecast should satsfy. Dawd oers the followng ntutve denton: Suppose that, n a long (conceptually nnte) sequence of weather forecasts, we look at all those days for whch the forecast probablty of precptaton was, say, close to some gven value p and (assumng these form an nnte sequence) determne the long run proporton of such days on whch the forecast event (ran) n fact occurred. The plot of aganst p s termed the forecaster's emprcal calbraton curve. If the curve s the dagonal = p, the forecaster may be termed well calbrated. 2 To gve the noton a formal denton, suppose that player 1 s usng a forecastng scheme f. The output of f n round t of play s an n-tuple f(t) = fp 1 (t); : : : ; p n (t)g where p j (t) s the forecasted probablty that player 2 wll play strategy j 2 S(2) at tme t. Let (j; t) = 1 f player 2 plays ther j-th strategy n round t and zero otherwse. Denote by N(p; t) the number of rounds up to the t-th round that f generated a vector of forecasts equal to p. Let (p; j; t) be the fracton of these rounds for whch player 2 plays j,.e., (p; j; t) = f N(p; t) > 0 and zero otherwse. t s=1 I f (s)=p (j; s) ; N(p; t) The forecast f s sad to calbrated wth respect to the sequences of plays made by player 2 f: lm t!1 p N(p; t) j(p; j; t)? p j j = 0 t 2 Dawd (1982) page 605. Hs notaton has been changed to match ours. 7

10 for all j 2 S(2). Notce that takng 0=0 = 0 s now seen not to matter snce the only tme t wll occur s f N(p; t) = 0, and thus t would be multpled by zero anyway. Roughly, calbraton says that the emprcal frequences condtoned on the assessments converge to the assessments. Ths s to be contrasted wth the asymptotc myopc bayes condton of Fudenberg and Kreps whch says that the emprcal frequences n round t converge together wth the assessments n round t. 3 Calbraton and Correlated Equlbrum It s clear from the denton of correlated strateges that a CE s smply a jont dstrbuton over S(1) S(2) wth a partcular property. Hence, we focus on D t (x; y), the fracton of tmes up to tme t that player 1 plays x and player 2 plays y. Ths s the emprcal jont dstrbuton. We assume that when players select ther best response (for a gven forecast) they use a a statonary and determnstc te breakng rule; say the lowest ndexed strategy. Theorem 1 Let be the set of all correlated equlbra. If each player uses a forecast that s calbrated aganst the others sequence of plays, and then makes a best response to ths forecast, then, mn D2 max x2s(1);y2s(2) jd t (x; y)? D(x; y)j! 0 as t, the number of rounds of play, tends to nnty. PROOF: Observe rst that the nm-tuple each of whose components s of the form D t (x; y) les n the nm? 1 dmensonal unt smplex. By the 8

11 Bolzano-Werstrass theorem any bounded sequence n t contans a convergent subsequence. Thus, for any subsequence fd t (x; y)g and D(x; y) such that x2s(1) y2s(2) we need to show that D s a CE. jd t (x; y)? D(x; y)j! 0; For each x 2 S(1) let M b (x) be the set of mxtures over S(2) for whch x s a best response. M b (x) s a closed convex subset of the n? 1 dmensonal smplex. Let M p (x) be the set of mxtures where player 1 actually plays x gven that the forecast s n M p (x). By the assumpton that players choose best responses, M p (x) M b (x). Further, fm p (x) : x 2 S(1)g forms a partton of the smplex. The emprcal condtonal dstrbuton of y 2 S(2) gven that player 1 played x s P c2s(2) D(x;c) D t (x;y) Pc2S(2) Dt (x;c). Ths converges to D(x;y) as long as P c2s(2) D t (x; c) does not converge to zero. If t dd, t would mean that the proporton of tmes that x s played tends to zero. Hence, n the lmt, player 1 never plays x, so t can be gnored. To complete the proof t suces to show that the n-tuple whose y-th component s contaned n M b (x). Observe that: D t (x; y) = t?1 = t?1 = t?1 = t?1 + t?1 rt :f (r)2m p(x) p2m p(x) p2m p(x) p2m p(x) p2m p(x) (y; r) rt :f (r)=p (y; r) (p; y; t )N(p; t ) p y N(p; t ) + ((p; y; t )? p y )N(p; t ) D(x;y) s D(x;c) Pc2S(2) 9

12 Snce the forecasts beng used are calbrated, the second term n the last expresson goes to zero as t tends to nnty. Note: p2m p(x) N(p; t ) p y P N(q; t ) 2 M b(x) q2m p(x) because t s a convex combnaton of vectors n M b (x) f recall, M p M b g, and M b (x) s convex. Therefore D(x; y) Pc2S(2) D(x; c) = lm t!1 p2m p(x) N(p; t) p y N(p; t) p2m p(x) whch s then the y th component of a vector n M b (x) also. We have shown that any sequence fd t (x; y)g contans a convergent subsequence whose lmt s a CE. The theorem now follows. 2 In some sense the result above s not surprsng. We know from Mlgrom and Roberts (1991) f players use best responses they elmnate domnated strateges. Secondly, the calbraton requrement forces lmt ponts to satsfy an addtonal equlbrum requrement. Correlaton arses because players are able to condton on prevous plays. It s natural to ask f Theorem 1 would hold wth a non-statonary tebreakng rule. The followng verson of matchng pennes shows that ths s not possble. In each round the row player wll forecast that there s a 50% Matchng Pennes h t H 1n-1-1n1 T -1n1 1n-1 chance that column wll play heads and a 50% chance that column wll play 10

13 tals,.e., (0.5, 0.5) s the forecast. The column player wll do lkewse. Gven these forecasts there s a te for the best reply. Consder the followng te breakng rule: on even numbered rounds play heads and tals on the other rounds. Notce that the resultng sequence of plays wll be: Tt, Hh, Tt, Hh, : : :. Clearly the forecasts of each player are calbrated, but the dstrbuton of plays does not converge to a CE. Theorem 1 rases the queston of how a calbrated forecast s to be produced. Oakes (1985), has shown that there s no determnstc forecast that s calbrated for all possble sequences of outcomes. Our requrements are more modest. Gven a game, and a correlated equlbrum of ths game, s there a sequence of plays and a determnstc forecastng rule dependng only on observed hstores that s calbrated? The next theorem provdes a postve answer to ths queston. Denton: Call a pont of the dstrbuton D(x; y) a lmt pont of calbrated forecasts f there exst determnstc best reply functons R () and calbrated forecastng rules p such that f each player, plays R (p ), then the lmtng jont dstrbuton wll be D(x; y). Denton: calbrated forecasts. Let be the set of all dstrbutons whch are lmt ponts of Usng ths notaton we can restate Theorem 1 as sayng that. We can represent every game by a vector n < 2mn, where each component corresponds to a players payo. A set of games s of measure zero f the correspondng set of ponts n < 2mn has Lebesgue measure zero. 11

14 Theorem 2 For almost every game (G) = (G). In other words, for almost every game, the set of dstrbutons whch calbrated learnng rules can converge to s dentcal to the set of correlated equlbrums. Proof: Because of Theorem 1 we need only prove that. Let (x j ; y j ) be a determnstc computable sequence such that the lmtng jont dstrbuton s D(x; y). At tme j, have player 1 forecast and player two forecast p 1;j () = D(x j ; ) p 2;j () = D(; y j ), y2s(2), x2s(1) D(x j ; y) D(x j ; y) : By the assumpton that the jont dstrbuton converges to D(x; y), t s clear that both of these forecasts are calbrated. Further, x j s n fact a best response to the forecast p 1;j (), and y j to p 2;j (). So, dene R 1 (p) such that for all j, x j = R 1 (p 1;j ) and smlarly for R 2 (p). These forecasts and these best reply functons are the key dea of the proof. In fact, n the stuaton where R 1 () and R 2 () are both well dened we have completed the proof. But, R 1 () and R 2 () mght not be well dened. In other words, there mght be two derent strateges x 0 and x 00 such that x j 0 = x 0 and x j 00 = x 00, then p 1;j 0() = p 1;j 00() = p. Ths s where the \almost every game" condton comes nto play. Almost every game has the property that all the sets M b (x) have nonempty nteror. To see why ths s the case, observe that M b (x) s formed by the ntersecton of half-spaces. Start wth a closed convex set wth non-empty nteror, C, say and add these half-spaces one at a tme. We can choose C 12

15 to be the smplex of all mxed strateges. Consder a half-space H, chosen at random such that the coecents that dene H are contnuous wth respect to lebesgue measure. We clam that the ntersecton of C and H s ether the empty set, or a set wth an open nteror. Pck a pont p n the nteror of C. Let q be the pont n the boundary of H whch s closest to p. Let v be the ray from p to q and d ts length. Both v and d have contnuous dstrbutons snce they are a contnuous transformaton of the half-space H. Now consder dstrbuton of d condtonal on v. Gven v there s a unque d such that H wll be tangent to C and not contan C. The condtonal probablty of d takng ths value s 0. Hence the uncondtonal probablty s zero also. The nterors of the sets of the form M b (x) are dsjont. 3 Thus, near the pont p there are ponts p x0 and p x00 such that the unque best response to p x0 s x 0 and the unque best response to p x00 s x 00. Forecastng p x0 or p x00 nstead of p makes the reply functon well dened. Unfortunately, when the forecast of p x0 s made, the actual frequency wll turn out to be p. Thus, the calbraton score wll be o by at most jp x0? p j. If we can choose p x0 be convergent to p solves ths last problem and our proof s complete. Dene a sequence p x0 p and for all, p x0 = (1? 1=)p + (1=)p x0. Then p x0 to converges to has x 0 as ts unque best reply. For each forecast p x0 sucently many tmes to ensure that there s a hgh probablty that the emprcal dstrbuton s wthn 1= of p. Wth hgh probablty the emprcal frequency condtonal on forecast p x0 wll be wthn 2= of p x0 and hence the calbraton score wll converge to zero. 2 3 The nterors and the unon of the boundares would form a partton. 13

16 To see why theorem 2 only holds for almost every game and not every game, consder the followng game: Example of 6= A 2n2 0n3 0n1 B 2n2 0n1 0n3 C 2n0 1n1 1n0 If ROW randomzes between A and B (wth equal probablty) and COL plays 1, then ths s a Correlated Equlbrum wth a payo of (2,2). But, the only pont n s the dstrbuton whch puts all ts weght on pont (C,2) whch yelds a payo of (1,1). Ths s because: M b (A) = M b (B) = f(1; 0; 0)g and M b (c) s the entre smplex. So, f R ROW ((1; 0; 0)) = A, then ROW wll never play strategy B, and lkewse f R ROW ((1; 0; 0)) = B, then ROW wll never play A. So, a mxture of A and B s mpossble and thus the payo (2,2) s mpossble. Thus, 6=. Can Theorem 1 be strengthened such that convergence to Nash Equlbrum s assured nstead of to a CE? The prevous theorem shows f one assumes only calbraton, one gets any CE n. So, wthout further assumptons on the forecastng rule, convergence to Nash cannot be assured. In partcular by addng an assumpton that the lmt exsts does not rene the equlbrum attaned (n contrast wth Fudenberg and Kreps who show that f a lmt exsts, t must be Nash). Ths s because Theorem 2 does not just nd an accumulaton pont t nds a drect lmt. Is t easy to construct a forecast that s calbrated? Gven the mpossblty theorem of Oakes (1985) the exstence of a determnstc scheme that s calbrated for all sequences s ruled out. However, a randomzed forecastng 14

17 scheme s possble. Theorem 3 ( Foster and Vohra 1991) There exsts a randomzed forecast that player 1 can use such that no matter what learnng rule player 2 uses, player 1 wll be calbrated. That s to say, player 1 0 s calbraton score C t p j2s(2) N(p; t) j(p; j; t)? p j j t (1) converges to zero n probablty. lm t!1 P (C t < ) = 1: In other words, for all 00 we have that Proof: See the appendx. The mportant thng to notce about Theorem 3 s that each player can ndvdually choose to be calbrated. The other player can not fol ths choce. Player 1 does not have to assume that player 2 s usng an exchangeable sequence, nor that the player 2 s ratonal. Player 1 s stll calbrated f player 2 plays any arbtrary sequence. Secondly, the proof s constructve,.e., there s an explct algorthm for producng such a forecast. 4 To extend ths result to the n-person case the forecastng rules must predct the jont dstrbuton of what everyone else wll play. If n Theorem 1 we requre only that the players use a forecastng rule that s close to calbrated n the sense of Theorem 3, we obtan: Corollary There exsts a randomzed forecastng scheme, such that f both player 1 and player 2 follow ths scheme, then FOR ANY normal form matrx game and for all > 0, there exsts a t 0 > 0, such that for all t > t 0, P (mn max jd t(x; y)? D(x; y)j < ) > 1? : D2 x;y 4 The most nvolved step s nvertng a matrx. 15

18 In other words, D t converges n probablty to the set under the Hausdor topology. 4 The Shapley Game and Fcttous Play The most famous of learnng rules for games s called Fcttous Play (FP), rst conceved n 1949 by George Brown. In a two person game t goes as follows: Denton: Denton of Fcttous Play: Row computes the proporton of tmes up to the present that Column has played each of hs/her strateges. Then, Row treats these proportons as the probabltes that Column wll select from among hs/her strateges. Row then selects the strategy that s hs/her best response. Column does lkewse. In 1951 Jula Robnson proved that FP converges to a NE n 2 person zero sum games. After the Robnson paper, nterest naturally turned to tryng to generalze Robnson's theorem to non-zero sum games. In 1961, K. Myasawa proved that FP converges to a NE n 2-person non-zero sum games where each player has at most two strateges. 5 However, n 1964 Lloyd Shapley dashed hopes of a generalzaton by descrbng a non-zero sum game consstng of three strateges for each player n whch FP dd not converge to a NE. In ths secton we show that FP doesn't converge to a Correlated Equlbrum. We use Shapley's orgnal example: 5 See Monderer and Shapley (1993) for other stuatons n whch FP converges. 16

19 Payo Matrx for Shapley Game n0 0n1 0n0 2 0n0 1n0 0n1 3 0n1 0n0 1n0 As observed by Shapley, FP n ths game wll oscllate between 6 states, (1,1) then (1,2), then (2; 2); (2; 3); (3; 3); (3; 1), then repeat. Fcttous play stays longer and longer n each state, so the perods of oscllaton get larger and larger. There s only one Correlated equlbrum wth support on these sx states. 6 It assgns probablty 1=6 to each state. Fcttous play s never close to ths dstrbuton. 7 Thus, t does not converge to a CE. 6 Usng Nau and McCardle (1990) the followng lnear program produces all the CE. p 11 p 12 p 22 p 23 p 33 p 31 p 11 ; p 13 p 11 ; p 13 p 23 ; p 21 p 22 ; p 21 p 31 ; p 32 p 33 ; p 32 p 12 : Whch s equvalent to the LP : p 11 = p 12 = p 22 = p 23 = p 33 = p 31 = p 11 ; p 13 p 11 ; p 21 p 11 ; p 32 p 11. Addng the constrant that p 13 = p 21 = p 32 = 0, ths LP has a unque soluton of p 11 = p 12 = p 22 = p 23 = p 33 = p 31 = 1=6. 7 Ths can be see ether by drect calculaton, or by the followng trck. If Fcttous play was ever close to ths CE, then the margnals would have to be close to (1=3; 1=3; 1=3). But, these margnals correspond to the Nash Equlbrum. Shapley created ths example precsely to show that the margnals ddn't converge to the margnals of the Nash equlbrum, n fact the margnals are bounded away from the (1=3; 1=3; 1=3) pont. Thus the Nash equlbrum s not an accumulaton pont of the sequence of plays. Thus, we know that the margnals are never close to beng correct, and thus the jont dstrbuton s also never close. 17

20 The Shapley game s nterestng because t has a CE whch s not a mxture of Nash Equlbrums. 8 Theorem 3 tells us that there are calbrated learnng rules whch wll then converge to ths CE. The expected payo s (1=2; 1=2) whch Pareto domnates the Nash payo of (1=3; 1=3). Postscrpt Earler versons of ths paper as well as presentatons of the results at varous conferences have generated a deal of follow on papers on calbraton and ts connectons to game theory. In ths secton, we gve a bref descrpton of some of ths work. Theorem 3, whch establshes the exstence of randomzed forecastng scheme that s calbrated has prompted a number of alternatve proofs. The rst of these was due to Sergu Hart (personal communcaton) and s partcularly smple and short. It makes use of the mn-max theorem. The draw back s that the scheme mpled by the method s mpractcal to mplement. Independently, Fudenberg and Levne (1995) also gave a proof usng the mnmax theorem. The approach s more elaborate than Hart's but produces a forecastng scheme that s practcal to mplement. In a follow up paper Fudenberg and Levne [1996] consder a renement of the calbraton dea that nvolves the classcaton of observatons nto varous categores. For ths renement they derve a procedure that yelds almost as hgh a tme-average payo as could be obtaned f the player chooses knowng the condtonal dstrbuton of actons gven categores. If players use such a procedure, long run the tme average play resembeles a correlated equlbrum. 8 The unque Nash Equlbrum for ths game s (1=3; 1=3; 1=3) vs (1=3; 1=3; 1=3). So, any CE whch sn't Nash, s also not a mxture of Nash Equlbrums. 18

21 Our own proof of Theorem 3 (whch s descrbed n the appendx) s based on establshng the exstence of a forecastng scheme that has a property called no-regret. An proof along the lnes of Theorem 1 shows that a noregret procedure would also lead to a correlated equlbrum. Hart and Mas- Collel (1996) have extended ths dea n many ways. Frst they proved a very elegant proof of no-regret based on Blackwell's (195?) vector mn-max theorem. Second they modfy ths scheme whch requres a matrx nverson to one that nvolves regret-matchng. Ths greatly reduces the computatons requred to mplement the procedure. The smpled procedure no longer has the no-regret property but t wll converge to a correlated equlbrum. Ther theorem s much harder to prove snce they can't smply appeal to a no-regret/calbraton property as we have done. Kala, Lehrer and Smorodnsky (1996) have recently shown that the noton of calbraton s mathematcally equvalent to that of mergng. Ths allows one to establsh relatonshps between convergence results based on mergng and those based on calbraton and so derve some new convergence results. Appendx Ths appendx provdes a telegraphc proof of Theorem 3. For more detals see Foster and Vohra (1991). We wll rst prove a property called \no-regret." Consder k forecasts each wth a loss or penalty at tme t of 0 L t 1 for = 1; : : : ; k. Now consder a randomzed forecast whch pcks forecast at tme t wth probablty w t. We dene the loss from usng the combned forecast to be the weghted sum of the losses of each forecast, namely, P k =1 w tl t. 19

22 Denton: s R!j T The regret generated by changng all forecasts to j forecasts maxf0; S j T g = S j T I S j T >0 where I x>0 s the ndcator functon and S j T T t=1 We choose the probablty vector w t conservaton equatons: (8) k w t j=1 w t(l t? L j t): so that t satses the followng ow R!j t?1 = k j=1 w j tr j! t?1 : The dualty theorem of lnear programmng can be used to establsh the exstence of a non-negatve soluton w t to ths system such that P k =1 w t = 1. Lemma 1 (No-regret) For all and j the regret grows as the squareroot of T. In partcular, R!j T p 2kT. Proof: Let G (x) x2 2 I x>0. Snce x G (x) we see that R!j T G (S j T ) j G (S j T ) Now G 0 (x) = xi x>0 and so j (S j t? S j t?1)g 0 (S j t?1) = j = = 0 w t(l t? L j t)(s j t?1i S j L t w t R!j t?1? j t?1 >0) w j t R j! t?1 j {z } = 0 by ow conservaton Expandng G (S j t ) as a two term Taylor seres around S j t?1 shows j G (S j t ) j G (St?1) j + (S j t? St?1)G j 0 (Sj t?1) + (w t )2 (L? t Lj t) 2 j j 20

23 j j G (S j t?1) + k G (S j t?1) + k: (w t )2 Computng the recursve sum we see that P j G (S j T ) T k and so R!j T 1 + T k. Pckng = 1=p 2kT shows R!j 2 T p 2T k. 2 We wll now show that for a sutable loss functon, a randomzed forecast that has no regret must also be calbrated. Frst, our forecastng scheme wll choose n each round a probablty vector from the set fp j = 0; 1; : : : ; kg whch s chosen so that any probablty dstrbuton over S(2) (the opponents strateges) s wthn of one of these ponts. We denote the move made by player n 2 by the vector t = [ t;1 ; t;2 ; t;3 ; : : :] where t;j = 1 f strategy j 2 S(2) was chosen and zero otherwse. Notce that t wll be a 0-1 vector wth exactly on non-zero component. Next, the loss ncurred n round t from forecastng p wll be L t = j t? p j 2 = P j2s(2) j t;j? p j j2. The probablty of forecastng p at tme t wll be w t. We would lke to choose the w t 's so that L-2 calbraton C 2 (t) goes to zero n probablty as t gets large, where C 2 (t) = p The expected value of C 2 (t) s gven by: E(C 2 (t)) = t k ((p; j; t)? p j ) 2N(p; t) t s=1 =1 j2s(2) 21 w t( t (p ; j; s)? p j) 2 =s:

24 Smple algebra yelds max j R!j t =t E(C 2 (t)) + max R!j t =t j If the probabltes w t 's are chosen to satsfy the ow conservaton equatons dsplayed earler, we deduce that E(C 2 (t)) + O( k p t ): Thus f we let k grow slowly and go slowly to zero, we see that C 2 (t)! 0 n expectaton whch mples C 2 (t)! 0 n probablty by Jensen's nequalty. The L-1 calbraton denton of equaton (1) follows from the fact that t s smaller than the square root of the L-2 calbraton. Thus we have proved Theorem 3. REFERENCES Aumann, R. J., \Subjectvty and Correlaton n Randomzed Strateges", Journal of Mathematcal Economcs, 1, 67-96, Aumann, R. J., \Correlated Equlbrum as an Expresson of Bayes Ratonalty", Econometrca, 55, #1, 1-18, Blackwell, D., \A vector valued analog of the mn-max theorem," Pacc J. Math. 6, 1-8, Dawd, A. P., \The Well Calbrated Bayesan", Journal of the Amercan Statstcal Assocaton, 77, #379, , Foster, D. P., and R. Vohra \Asymptotc Calbraton", unpublshed manuscrpt, Foster, D. P., and H. P. Young \Stochastc Evolutonary Game Dynamcs," Theoretcal Populaton Bology, 38, ,

25 Fudenberg, D. and D. Kreps, \Expermentaton, Learnng, and Equlbrum n Extensve Form Games", unpublshed notes, Fudenberg, D. and D. Levne, \An easer way to Calbrate", unpublshed manuscrpt, Fudenberg, D. and D. Levne, \Condtonal Unversal Consstency", unpublshed manuscrpt, Hart, S. and Mas-Collel, A., \A Smple Adaptve Procedure Leadng to Correlated Equlbrum," unpublshed manuscrpt, Kala, E. and E. Lehrer, \Subjectve Games and Equlbra", unpublshed manuscrpt, Kala, E., E. Lehrer and R. Smorodnsky, \Calbrated Forecastng and Mergng", manuscrpt, Kreps, D. \Game Theory and Economc Modelng," Oxford Unversty Press, Oxford, 1991a. Kreps, D., Semnar on Learnng n Games, Unversty of Chcago, 1991b. Mlgrom, P. and J. Roberts \Adaptve and Sophstcated Learnng n Normal Form Games", Games and Economc Behavor, 3, , Monderer, D. and L. Shapley \Fcttous Play Property for games wth Identcal Interests," workng paper, Department of Economcs, UCLA, Myasawa, K., \On the Convergence of the Learnng Process n a 2 2 Non-zero- sum Two-person Game," Economc Research Program, Prnceton Unversty, Research Memorandum #33 (1961). Nau, R. F. and K. McCardle, \Coherent Behavor n Non-cooperatve Games," Journal of Economc Theory, 50, #2, Oakes, D. \Self-Calbratng Prors Do Not Exst", Journal of the Amercan Statstcal Assocaton, 80, 339, Robnson, J. \An Iteratve Method of Solvng a Game" Annals of Math- 23

26 ematcs, 54, , Samuelson, L. and J. Zhang, \Evolutonary Stablty n Asymmetrc Games", Journal of Economc Theory, 57, , Shapley, L. \Some Topcs n Two-Person Games," Advances n Game Theory, Prnceton Unversty Press, Prnceton Skyrms, B. The Dynamcs of Ratonal Delberaton, Harvard Unversty Press,

The Second Anti-Mathima on Game Theory

The Second Anti-Mathima on Game Theory The Second Ant-Mathma on Game Theory Ath. Kehagas December 1 2006 1 Introducton In ths note we wll examne the noton of game equlbrum for three types of games 1. 2-player 2-acton zero-sum games 2. 2-player

More information

COS 521: Advanced Algorithms Game Theory and Linear Programming

COS 521: Advanced Algorithms Game Theory and Linear Programming COS 521: Advanced Algorthms Game Theory and Lnear Programmng Moses Charkar February 27, 2013 In these notes, we ntroduce some basc concepts n game theory and lnear programmng (LP). We show a connecton

More information

CS286r Assign One. Answer Key

CS286r Assign One. Answer Key CS286r Assgn One Answer Key 1 Game theory 1.1 1.1.1 Let off-equlbrum strateges also be that people contnue to play n Nash equlbrum. Devatng from any Nash equlbrum s a weakly domnated strategy. That s,

More information

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X Statstcs 1: Probablty Theory II 37 3 EPECTATION OF SEVERAL RANDOM VARIABLES As n Probablty Theory I, the nterest n most stuatons les not on the actual dstrbuton of a random vector, but rather on a number

More information

Endogenous timing in a mixed oligopoly consisting of a single public firm and foreign competitors. Abstract

Endogenous timing in a mixed oligopoly consisting of a single public firm and foreign competitors. Abstract Endogenous tmng n a mxed olgopoly consstng o a sngle publc rm and oregn compettors Yuanzhu Lu Chna Economcs and Management Academy, Central Unversty o Fnance and Economcs Abstract We nvestgate endogenous

More information

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4)

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4) I. Classcal Assumptons Econ7 Appled Econometrcs Topc 3: Classcal Model (Studenmund, Chapter 4) We have defned OLS and studed some algebrac propertes of OLS. In ths topc we wll study statstcal propertes

More information

Subjective Uncertainty Over Behavior Strategies: A Correction

Subjective Uncertainty Over Behavior Strategies: A Correction Subjectve Uncertanty Over Behavor Strateges: A Correcton The Harvard communty has made ths artcle openly avalable. Please share how ths access benefts you. Your story matters. Ctaton Publshed Verson Accessed

More information

Computing Correlated Equilibria in Multi-Player Games

Computing Correlated Equilibria in Multi-Player Games Computng Correlated Equlbra n Mult-Player Games Chrstos H. Papadmtrou Presented by Zhanxang Huang December 7th, 2005 1 The Author Dr. Chrstos H. Papadmtrou CS professor at UC Berkley (taught at Harvard,

More information

APPENDIX A Some Linear Algebra

APPENDIX A Some Linear Algebra APPENDIX A Some Lnear Algebra The collecton of m, n matrces A.1 Matrces a 1,1,..., a 1,n A = a m,1,..., a m,n wth real elements a,j s denoted by R m,n. If n = 1 then A s called a column vector. Smlarly,

More information

Linear Approximation with Regularization and Moving Least Squares

Linear Approximation with Regularization and Moving Least Squares Lnear Approxmaton wth Regularzaton and Movng Least Squares Igor Grešovn May 007 Revson 4.6 (Revson : March 004). 5 4 3 0.5 3 3.5 4 Contents: Lnear Fttng...4. Weghted Least Squares n Functon Approxmaton...

More information

= z 20 z n. (k 20) + 4 z k = 4

= z 20 z n. (k 20) + 4 z k = 4 Problem Set #7 solutons 7.2.. (a Fnd the coeffcent of z k n (z + z 5 + z 6 + z 7 + 5, k 20. We use the known seres expanson ( n+l ( z l l z n below: (z + z 5 + z 6 + z 7 + 5 (z 5 ( + z + z 2 + z + 5 5

More information

Chapter 9: Statistical Inference and the Relationship between Two Variables

Chapter 9: Statistical Inference and the Relationship between Two Variables Chapter 9: Statstcal Inference and the Relatonshp between Two Varables Key Words The Regresson Model The Sample Regresson Equaton The Pearson Correlaton Coeffcent Learnng Outcomes After studyng ths chapter,

More information

Computation of Higher Order Moments from Two Multinomial Overdispersion Likelihood Models

Computation of Higher Order Moments from Two Multinomial Overdispersion Likelihood Models Computaton of Hgher Order Moments from Two Multnomal Overdsperson Lkelhood Models BY J. T. NEWCOMER, N. K. NEERCHAL Department of Mathematcs and Statstcs, Unversty of Maryland, Baltmore County, Baltmore,

More information

a b a In case b 0, a being divisible by b is the same as to say that

a b a In case b 0, a being divisible by b is the same as to say that Secton 6.2 Dvsblty among the ntegers An nteger a ε s dvsble by b ε f there s an nteger c ε such that a = bc. Note that s dvsble by any nteger b, snce = b. On the other hand, a s dvsble by only f a = :

More information

Numerical Heat and Mass Transfer

Numerical Heat and Mass Transfer Master degree n Mechancal Engneerng Numercal Heat and Mass Transfer 06-Fnte-Dfference Method (One-dmensonal, steady state heat conducton) Fausto Arpno f.arpno@uncas.t Introducton Why we use models and

More information

Foundations of Arithmetic

Foundations of Arithmetic Foundatons of Arthmetc Notaton We shall denote the sum and product of numbers n the usual notaton as a 2 + a 2 + a 3 + + a = a, a 1 a 2 a 3 a = a The notaton a b means a dvdes b,.e. ac = b where c s an

More information

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity LINEAR REGRESSION ANALYSIS MODULE IX Lecture - 30 Multcollnearty Dr. Shalabh Department of Mathematcs and Statstcs Indan Insttute of Technology Kanpur 2 Remedes for multcollnearty Varous technques have

More information

Perfect Competition and the Nash Bargaining Solution

Perfect Competition and the Nash Bargaining Solution Perfect Competton and the Nash Barganng Soluton Renhard John Department of Economcs Unversty of Bonn Adenauerallee 24-42 53113 Bonn, Germany emal: rohn@un-bonn.de May 2005 Abstract For a lnear exchange

More information

MATH 829: Introduction to Data Mining and Analysis The EM algorithm (part 2)

MATH 829: Introduction to Data Mining and Analysis The EM algorithm (part 2) 1/16 MATH 829: Introducton to Data Mnng and Analyss The EM algorthm (part 2) Domnque Gullot Departments of Mathematcal Scences Unversty of Delaware Aprl 20, 2016 Recall 2/16 We are gven ndependent observatons

More information

1 The Mistake Bound Model

1 The Mistake Bound Model 5-850: Advanced Algorthms CMU, Sprng 07 Lecture #: Onlne Learnng and Multplcatve Weghts February 7, 07 Lecturer: Anupam Gupta Scrbe: Bryan Lee,Albert Gu, Eugene Cho he Mstake Bound Model Suppose there

More information

Lecture 4. Instructor: Haipeng Luo

Lecture 4. Instructor: Haipeng Luo Lecture 4 Instructor: Hapeng Luo In the followng lectures, we focus on the expert problem and study more adaptve algorthms. Although Hedge s proven to be worst-case optmal, one may wonder how well t would

More information

Feature Selection: Part 1

Feature Selection: Part 1 CSE 546: Machne Learnng Lecture 5 Feature Selecton: Part 1 Instructor: Sham Kakade 1 Regresson n the hgh dmensonal settng How do we learn when the number of features d s greater than the sample sze n?

More information

Games of Threats. Elon Kohlberg Abraham Neyman. Working Paper

Games of Threats. Elon Kohlberg Abraham Neyman. Working Paper Games of Threats Elon Kohlberg Abraham Neyman Workng Paper 18-023 Games of Threats Elon Kohlberg Harvard Busness School Abraham Neyman The Hebrew Unversty of Jerusalem Workng Paper 18-023 Copyrght 2017

More information

Bayesian predictive Configural Frequency Analysis

Bayesian predictive Configural Frequency Analysis Psychologcal Test and Assessment Modelng, Volume 54, 2012 (3), 285-292 Bayesan predctve Confgural Frequency Analyss Eduardo Gutérrez-Peña 1 Abstract Confgural Frequency Analyss s a method for cell-wse

More information

Difference Equations

Difference Equations Dfference Equatons c Jan Vrbk 1 Bascs Suppose a sequence of numbers, say a 0,a 1,a,a 3,... s defned by a certan general relatonshp between, say, three consecutve values of the sequence, e.g. a + +3a +1

More information

THE CHINESE REMAINDER THEOREM. We should thank the Chinese for their wonderful remainder theorem. Glenn Stevens

THE CHINESE REMAINDER THEOREM. We should thank the Chinese for their wonderful remainder theorem. Glenn Stevens THE CHINESE REMAINDER THEOREM KEITH CONRAD We should thank the Chnese for ther wonderful remander theorem. Glenn Stevens 1. Introducton The Chnese remander theorem says we can unquely solve any par of

More information

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal Inner Product Defnton 1 () A Eucldean space s a fnte-dmensonal vector space over the reals R, wth an nner product,. Defnton 2 (Inner Product) An nner product, on a real vector space X s a symmetrc, blnear,

More information

j) = 1 (note sigma notation) ii. Continuous random variable (e.g. Normal distribution) 1. density function: f ( x) 0 and f ( x) dx = 1

j) = 1 (note sigma notation) ii. Continuous random variable (e.g. Normal distribution) 1. density function: f ( x) 0 and f ( x) dx = 1 Random varables Measure of central tendences and varablty (means and varances) Jont densty functons and ndependence Measures of assocaton (covarance and correlaton) Interestng result Condtonal dstrbutons

More information

CS : Algorithms and Uncertainty Lecture 17 Date: October 26, 2016

CS : Algorithms and Uncertainty Lecture 17 Date: October 26, 2016 CS 29-128: Algorthms and Uncertanty Lecture 17 Date: October 26, 2016 Instructor: Nkhl Bansal Scrbe: Mchael Denns 1 Introducton In ths lecture we wll be lookng nto the secretary problem, and an nterestng

More information

CSci 6974 and ECSE 6966 Math. Tech. for Vision, Graphics and Robotics Lecture 21, April 17, 2006 Estimating A Plane Homography

CSci 6974 and ECSE 6966 Math. Tech. for Vision, Graphics and Robotics Lecture 21, April 17, 2006 Estimating A Plane Homography CSc 6974 and ECSE 6966 Math. Tech. for Vson, Graphcs and Robotcs Lecture 21, Aprl 17, 2006 Estmatng A Plane Homography Overvew We contnue wth a dscusson of the major ssues, usng estmaton of plane projectve

More information

Maximizing the number of nonnegative subsets

Maximizing the number of nonnegative subsets Maxmzng the number of nonnegatve subsets Noga Alon Hao Huang December 1, 213 Abstract Gven a set of n real numbers, f the sum of elements of every subset of sze larger than k s negatve, what s the maxmum

More information

ANSWERS. Problem 1. and the moment generating function (mgf) by. defined for any real t. Use this to show that E( U) var( U)

ANSWERS. Problem 1. and the moment generating function (mgf) by. defined for any real t. Use this to show that E( U) var( U) Econ 413 Exam 13 H ANSWERS Settet er nndelt 9 deloppgaver, A,B,C, som alle anbefales å telle lkt for å gøre det ltt lettere å stå. Svar er gtt . Unfortunately, there s a prntng error n the hnt of

More information

Outline. Bayesian Networks: Maximum Likelihood Estimation and Tree Structure Learning. Our Model and Data. Outline

Outline. Bayesian Networks: Maximum Likelihood Estimation and Tree Structure Learning. Our Model and Data. Outline Outlne Bayesan Networks: Maxmum Lkelhood Estmaton and Tree Structure Learnng Huzhen Yu janey.yu@cs.helsnk.f Dept. Computer Scence, Unv. of Helsnk Probablstc Models, Sprng, 200 Notces: I corrected a number

More information

Game Theory. Lecture Notes By Y. Narahari. Department of Computer Science and Automation Indian Institute of Science Bangalore, India February 2008

Game Theory. Lecture Notes By Y. Narahari. Department of Computer Science and Automation Indian Institute of Science Bangalore, India February 2008 Game Theory Lecture Notes By Y. Narahar Department of Computer Scence and Automaton Indan Insttute of Scence Bangalore, Inda February 2008 Chapter 10: Two Person Zero Sum Games Note: Ths s a only a draft

More information

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 12 10/21/2013. Martingale Concentration Inequalities and Applications

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 12 10/21/2013. Martingale Concentration Inequalities and Applications MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.65/15.070J Fall 013 Lecture 1 10/1/013 Martngale Concentraton Inequaltes and Applcatons Content. 1. Exponental concentraton for martngales wth bounded ncrements.

More information

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011 Stanford Unversty CS359G: Graph Parttonng and Expanders Handout 4 Luca Trevsan January 3, 0 Lecture 4 In whch we prove the dffcult drecton of Cheeger s nequalty. As n the past lectures, consder an undrected

More information

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009 College of Computer & Informaton Scence Fall 2009 Northeastern Unversty 20 October 2009 CS7880: Algorthmc Power Tools Scrbe: Jan Wen and Laura Poplawsk Lecture Outlne: Prmal-dual schema Network Desgn:

More information

Axiomatizations of Pareto Equilibria in Multicriteria Games

Axiomatizations of Pareto Equilibria in Multicriteria Games ames and Economc Behavor 28, 146154 1999. Artcle ID game.1998.0680, avalable onlne at http:www.dealbrary.com on Axomatzatons of Pareto Equlbra n Multcrtera ames Mark Voorneveld,* Dres Vermeulen, and Peter

More information

Lecture 20: Lift and Project, SDP Duality. Today we will study the Lift and Project method. Then we will prove the SDP duality theorem.

Lecture 20: Lift and Project, SDP Duality. Today we will study the Lift and Project method. Then we will prove the SDP duality theorem. prnceton u. sp 02 cos 598B: algorthms and complexty Lecture 20: Lft and Project, SDP Dualty Lecturer: Sanjeev Arora Scrbe:Yury Makarychev Today we wll study the Lft and Project method. Then we wll prove

More information

More metrics on cartesian products

More metrics on cartesian products More metrcs on cartesan products If (X, d ) are metrc spaces for 1 n, then n Secton II4 of the lecture notes we defned three metrcs on X whose underlyng topologes are the product topology The purpose of

More information

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification

2E Pattern Recognition Solutions to Introduction to Pattern Recognition, Chapter 2: Bayesian pattern classification E395 - Pattern Recognton Solutons to Introducton to Pattern Recognton, Chapter : Bayesan pattern classfcaton Preface Ths document s a soluton manual for selected exercses from Introducton to Pattern Recognton

More information

Here is the rationale: If X and y have a strong positive relationship to one another, then ( x x) will tend to be positive when ( y y)

Here is the rationale: If X and y have a strong positive relationship to one another, then ( x x) will tend to be positive when ( y y) Secton 1.5 Correlaton In the prevous sectons, we looked at regresson and the value r was a measurement of how much of the varaton n y can be attrbuted to the lnear relatonshp between y and x. In ths secton,

More information

Lecture 12: Discrete Laplacian

Lecture 12: Discrete Laplacian Lecture 12: Dscrete Laplacan Scrbe: Tanye Lu Our goal s to come up wth a dscrete verson of Laplacan operator for trangulated surfaces, so that we can use t n practce to solve related problems We are mostly

More information

Complete subgraphs in multipartite graphs

Complete subgraphs in multipartite graphs Complete subgraphs n multpartte graphs FLORIAN PFENDER Unverstät Rostock, Insttut für Mathematk D-18057 Rostock, Germany Floran.Pfender@un-rostock.de Abstract Turán s Theorem states that every graph G

More information

20. Mon, Oct. 13 What we have done so far corresponds roughly to Chapters 2 & 3 of Lee. Now we turn to Chapter 4. The first idea is connectedness.

20. Mon, Oct. 13 What we have done so far corresponds roughly to Chapters 2 & 3 of Lee. Now we turn to Chapter 4. The first idea is connectedness. 20. Mon, Oct. 13 What we have done so far corresponds roughly to Chapters 2 & 3 of Lee. Now we turn to Chapter 4. The frst dea s connectedness. Essentally, we want to say that a space cannot be decomposed

More information

A note on almost sure behavior of randomly weighted sums of φ-mixing random variables with φ-mixing weights

A note on almost sure behavior of randomly weighted sums of φ-mixing random variables with φ-mixing weights ACTA ET COMMENTATIONES UNIVERSITATIS TARTUENSIS DE MATHEMATICA Volume 7, Number 2, December 203 Avalable onlne at http://acutm.math.ut.ee A note on almost sure behavor of randomly weghted sums of φ-mxng

More information

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix Lectures - Week 4 Matrx norms, Condtonng, Vector Spaces, Lnear Independence, Spannng sets and Bass, Null space and Range of a Matrx Matrx Norms Now we turn to assocatng a number to each matrx. We could

More information

LECTURE 9 CANONICAL CORRELATION ANALYSIS

LECTURE 9 CANONICAL CORRELATION ANALYSIS LECURE 9 CANONICAL CORRELAION ANALYSIS Introducton he concept of canoncal correlaton arses when we want to quantfy the assocatons between two sets of varables. For example, suppose that the frst set of

More information

Computational Biology Lecture 8: Substitution matrices Saad Mneimneh

Computational Biology Lecture 8: Substitution matrices Saad Mneimneh Computatonal Bology Lecture 8: Substtuton matrces Saad Mnemneh As we have ntroduced last tme, smple scorng schemes lke + or a match, - or a msmatch and -2 or a gap are not justable bologcally, especally

More information

MMA and GCMMA two methods for nonlinear optimization

MMA and GCMMA two methods for nonlinear optimization MMA and GCMMA two methods for nonlnear optmzaton Krster Svanberg Optmzaton and Systems Theory, KTH, Stockholm, Sweden. krlle@math.kth.se Ths note descrbes the algorthms used n the author s 2007 mplementatons

More information

Expected Value and Variance

Expected Value and Variance MATH 38 Expected Value and Varance Dr. Neal, WKU We now shall dscuss how to fnd the average and standard devaton of a random varable X. Expected Value Defnton. The expected value (or average value, or

More information

Department of Computer Science Artificial Intelligence Research Laboratory. Iowa State University MACHINE LEARNING

Department of Computer Science Artificial Intelligence Research Laboratory. Iowa State University MACHINE LEARNING MACHINE LEANING Vasant Honavar Bonformatcs and Computatonal Bology rogram Center for Computatonal Intellgence, Learnng, & Dscovery Iowa State Unversty honavar@cs.astate.edu www.cs.astate.edu/~honavar/

More information

Hidden Markov Models & The Multivariate Gaussian (10/26/04)

Hidden Markov Models & The Multivariate Gaussian (10/26/04) CS281A/Stat241A: Statstcal Learnng Theory Hdden Markov Models & The Multvarate Gaussan (10/26/04) Lecturer: Mchael I. Jordan Scrbes: Jonathan W. Hu 1 Hdden Markov Models As a bref revew, hdden Markov models

More information

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction ECONOMICS 5* -- NOTE (Summary) ECON 5* -- NOTE The Multple Classcal Lnear Regresson Model (CLRM): Specfcaton and Assumptons. Introducton CLRM stands for the Classcal Lnear Regresson Model. The CLRM s also

More information

Affine transformations and convexity

Affine transformations and convexity Affne transformatons and convexty The purpose of ths document s to prove some basc propertes of affne transformatons nvolvng convex sets. Here are a few onlne references for background nformaton: http://math.ucr.edu/

More information

Time-Varying Systems and Computations Lecture 6

Time-Varying Systems and Computations Lecture 6 Tme-Varyng Systems and Computatons Lecture 6 Klaus Depold 14. Januar 2014 The Kalman Flter The Kalman estmaton flter attempts to estmate the actual state of an unknown dscrete dynamcal system, gven nosy

More information

Errors for Linear Systems

Errors for Linear Systems Errors for Lnear Systems When we solve a lnear system Ax b we often do not know A and b exactly, but have only approxmatons  and ˆb avalable. Then the best thng we can do s to solve ˆx ˆb exactly whch

More information

THE SUMMATION NOTATION Ʃ

THE SUMMATION NOTATION Ʃ Sngle Subscrpt otaton THE SUMMATIO OTATIO Ʃ Most of the calculatons we perform n statstcs are repettve operatons on lsts of numbers. For example, we compute the sum of a set of numbers, or the sum of the

More information

Lecture 17 : Stochastic Processes II

Lecture 17 : Stochastic Processes II : Stochastc Processes II 1 Contnuous-tme stochastc process So far we have studed dscrete-tme stochastc processes. We studed the concept of Makov chans and martngales, tme seres analyss, and regresson analyss

More information

Understanding Reasoning Using Utility Proportional Beliefs

Understanding Reasoning Using Utility Proportional Beliefs Understandng Reasonng Usng Utlty Proportonal Belefs Chrstan Nauerz EpCenter, Maastrcht Unversty c.nauerz@maastrchtunversty.nl Abstract. Tradtonally very lttle attenton has been pad to the reasonng process

More information

Lecture 3. Ax x i a i. i i

Lecture 3. Ax x i a i. i i 18.409 The Behavor of Algorthms n Practce 2/14/2 Lecturer: Dan Spelman Lecture 3 Scrbe: Arvnd Sankar 1 Largest sngular value In order to bound the condton number, we need an upper bound on the largest

More information

Online Appendix. t=1 (p t w)q t. Then the first order condition shows that

Online Appendix. t=1 (p t w)q t. Then the first order condition shows that Artcle forthcomng to ; manuscrpt no (Please, provde the manuscrpt number!) 1 Onlne Appendx Appendx E: Proofs Proof of Proposton 1 Frst we derve the equlbrum when the manufacturer does not vertcally ntegrate

More information

Generalized Linear Methods

Generalized Linear Methods Generalzed Lnear Methods 1 Introducton In the Ensemble Methods the general dea s that usng a combnaton of several weak learner one could make a better learner. More formally, assume that we have a set

More information

Genericity of Critical Types

Genericity of Critical Types Genercty of Crtcal Types Y-Chun Chen Alfredo D Tllo Eduardo Fangold Syang Xong September 2008 Abstract Ely and Pesk 2008 offers an nsghtful characterzaton of crtcal types: a type s crtcal f and only f

More information

Linear Correlation. Many research issues are pursued with nonexperimental studies that seek to establish relationships among 2 or more variables

Linear Correlation. Many research issues are pursued with nonexperimental studies that seek to establish relationships among 2 or more variables Lnear Correlaton Many research ssues are pursued wth nonexpermental studes that seek to establsh relatonshps among or more varables E.g., correlates of ntellgence; relaton between SAT and GPA; relaton

More information

Module 9. Lecture 6. Duality in Assignment Problems

Module 9. Lecture 6. Duality in Assignment Problems Module 9 1 Lecture 6 Dualty n Assgnment Problems In ths lecture we attempt to answer few other mportant questons posed n earler lecture for (AP) and see how some of them can be explaned through the concept

More information

Finding Dense Subgraphs in G(n, 1/2)

Finding Dense Subgraphs in G(n, 1/2) Fndng Dense Subgraphs n Gn, 1/ Atsh Das Sarma 1, Amt Deshpande, and Rav Kannan 1 Georga Insttute of Technology,atsh@cc.gatech.edu Mcrosoft Research-Bangalore,amtdesh,annan@mcrosoft.com Abstract. Fndng

More information

PROCEEDINGS OF THE AMERICAN MATHEMATICAL SOCIETY Volume 125, Number 7, July 1997, Pages 2119{2125 S (97) THE STRONG OPEN SET CONDITION

PROCEEDINGS OF THE AMERICAN MATHEMATICAL SOCIETY Volume 125, Number 7, July 1997, Pages 2119{2125 S (97) THE STRONG OPEN SET CONDITION PROCDINGS OF TH AMRICAN MATHMATICAL SOCITY Volume 125, Number 7, July 1997, Pages 2119{2125 S 0002-9939(97)03816-1 TH STRONG OPN ST CONDITION IN TH RANDOM CAS NORBRT PATZSCHK (Communcated by Palle. T.

More information

Parametric fractional imputation for missing data analysis. Jae Kwang Kim Survey Working Group Seminar March 29, 2010

Parametric fractional imputation for missing data analysis. Jae Kwang Kim Survey Working Group Seminar March 29, 2010 Parametrc fractonal mputaton for mssng data analyss Jae Kwang Km Survey Workng Group Semnar March 29, 2010 1 Outlne Introducton Proposed method Fractonal mputaton Approxmaton Varance estmaton Multple mputaton

More information

Edge Isoperimetric Inequalities

Edge Isoperimetric Inequalities November 7, 2005 Ross M. Rchardson Edge Isopermetrc Inequaltes 1 Four Questons Recall that n the last lecture we looked at the problem of sopermetrc nequaltes n the hypercube, Q n. Our noton of boundary

More information

n α j x j = 0 j=1 has a nontrivial solution. Here A is the n k matrix whose jth column is the vector for all t j=0

n α j x j = 0 j=1 has a nontrivial solution. Here A is the n k matrix whose jth column is the vector for all t j=0 MODULE 2 Topcs: Lnear ndependence, bass and dmenson We have seen that f n a set of vectors one vector s a lnear combnaton of the remanng vectors n the set then the span of the set s unchanged f that vector

More information

Comparison of the Population Variance Estimators. of 2-Parameter Exponential Distribution Based on. Multiple Criteria Decision Making Method

Comparison of the Population Variance Estimators. of 2-Parameter Exponential Distribution Based on. Multiple Criteria Decision Making Method Appled Mathematcal Scences, Vol. 7, 0, no. 47, 07-0 HIARI Ltd, www.m-hkar.com Comparson of the Populaton Varance Estmators of -Parameter Exponental Dstrbuton Based on Multple Crtera Decson Makng Method

More information

Hopfield networks and Boltzmann machines. Geoffrey Hinton et al. Presented by Tambet Matiisen

Hopfield networks and Boltzmann machines. Geoffrey Hinton et al. Presented by Tambet Matiisen Hopfeld networks and Boltzmann machnes Geoffrey Hnton et al. Presented by Tambet Matsen 18.11.2014 Hopfeld network Bnary unts Symmetrcal connectons http://www.nnwj.de/hopfeld-net.html Energy functon The

More information

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg prnceton unv. F 17 cos 521: Advanced Algorthm Desgn Lecture 7: LP Dualty Lecturer: Matt Wenberg Scrbe: LP Dualty s an extremely useful tool for analyzng structural propertes of lnear programs. Whle there

More information

Random Walks on Digraphs

Random Walks on Digraphs Random Walks on Dgraphs J. J. P. Veerman October 23, 27 Introducton Let V = {, n} be a vertex set and S a non-negatve row-stochastc matrx (.e. rows sum to ). V and S defne a dgraph G = G(V, S) and a drected

More information

The equation of motion of a dynamical system is given by a set of differential equations. That is (1)

The equation of motion of a dynamical system is given by a set of differential equations. That is (1) Dynamcal Systems Many engneerng and natural systems are dynamcal systems. For example a pendulum s a dynamcal system. State l The state of the dynamcal system specfes t condtons. For a pendulum n the absence

More information

Excess Error, Approximation Error, and Estimation Error

Excess Error, Approximation Error, and Estimation Error E0 370 Statstcal Learnng Theory Lecture 10 Sep 15, 011 Excess Error, Approxaton Error, and Estaton Error Lecturer: Shvan Agarwal Scrbe: Shvan Agarwal 1 Introducton So far, we have consdered the fnte saple

More information

NUMERICAL DIFFERENTIATION

NUMERICAL DIFFERENTIATION NUMERICAL DIFFERENTIATION 1 Introducton Dfferentaton s a method to compute the rate at whch a dependent output y changes wth respect to the change n the ndependent nput x. Ths rate of change s called the

More information

A note on the one-deviation property in extensive form games

A note on the one-deviation property in extensive form games Games and Economc Behavor 40 (2002) 322 338 www.academcpress.com Note A note on the one-devaton property n extensve form games Andrés Perea Departamento de Economía, Unversdad Carlos III de Madrd, Calle

More information

Linear, affine, and convex sets and hulls In the sequel, unless otherwise specified, X will denote a real vector space.

Linear, affine, and convex sets and hulls In the sequel, unless otherwise specified, X will denote a real vector space. Lnear, affne, and convex sets and hulls In the sequel, unless otherwse specfed, X wll denote a real vector space. Lnes and segments. Gven two ponts x, y X, we defne xy = {x + t(y x) : t R} = {(1 t)x +

More information

5 The Rational Canonical Form

5 The Rational Canonical Form 5 The Ratonal Canoncal Form Here p s a monc rreducble factor of the mnmum polynomal m T and s not necessarly of degree one Let F p denote the feld constructed earler n the course, consstng of all matrces

More information

PHYS 705: Classical Mechanics. Calculus of Variations II

PHYS 705: Classical Mechanics. Calculus of Variations II 1 PHYS 705: Classcal Mechancs Calculus of Varatons II 2 Calculus of Varatons: Generalzaton (no constrant yet) Suppose now that F depends on several dependent varables : We need to fnd such that has a statonary

More information

Basically, if you have a dummy dependent variable you will be estimating a probability.

Basically, if you have a dummy dependent variable you will be estimating a probability. ECON 497: Lecture Notes 13 Page 1 of 1 Metropoltan State Unversty ECON 497: Research and Forecastng Lecture Notes 13 Dummy Dependent Varable Technques Studenmund Chapter 13 Bascally, f you have a dummy

More information

1 Convex Optimization

1 Convex Optimization Convex Optmzaton We wll consder convex optmzaton problems. Namely, mnmzaton problems where the objectve s convex (we assume no constrants for now). Such problems often arse n machne learnng. For example,

More information

Chapter 11: Simple Linear Regression and Correlation

Chapter 11: Simple Linear Regression and Correlation Chapter 11: Smple Lnear Regresson and Correlaton 11-1 Emprcal Models 11-2 Smple Lnear Regresson 11-3 Propertes of the Least Squares Estmators 11-4 Hypothess Test n Smple Lnear Regresson 11-4.1 Use of t-tests

More information

Lecture 10 Support Vector Machines II

Lecture 10 Support Vector Machines II Lecture 10 Support Vector Machnes II 22 February 2016 Taylor B. Arnold Yale Statstcs STAT 365/665 1/28 Notes: Problem 3 s posted and due ths upcomng Frday There was an early bug n the fake-test data; fxed

More information

Tail Dependence Comparison of Survival Marshall-Olkin Copulas

Tail Dependence Comparison of Survival Marshall-Olkin Copulas Tal Dependence Comparson of Survval Marshall-Olkn Copulas Hajun L Department of Mathematcs and Department of Statstcs Washngton State Unversty Pullman, WA 99164, U.S.A. lh@math.wsu.edu January 2006 Abstract

More information

MIMA Group. Chapter 2 Bayesian Decision Theory. School of Computer Science and Technology, Shandong University. Xin-Shun SDU

MIMA Group. Chapter 2 Bayesian Decision Theory. School of Computer Science and Technology, Shandong University. Xin-Shun SDU Group M D L M Chapter Bayesan Decson heory Xn-Shun Xu @ SDU School of Computer Scence and echnology, Shandong Unversty Bayesan Decson heory Bayesan decson theory s a statstcal approach to data mnng/pattern

More information

Welfare Properties of General Equilibrium. What can be said about optimality properties of resource allocation implied by general equilibrium?

Welfare Properties of General Equilibrium. What can be said about optimality properties of resource allocation implied by general equilibrium? APPLIED WELFARE ECONOMICS AND POLICY ANALYSIS Welfare Propertes of General Equlbrum What can be sad about optmalty propertes of resource allocaton mpled by general equlbrum? Any crteron used to compare

More information

Lecture 14: Bandits with Budget Constraints

Lecture 14: Bandits with Budget Constraints IEOR 8100-001: Learnng and Optmzaton for Sequental Decson Makng 03/07/16 Lecture 14: andts wth udget Constrants Instructor: Shpra Agrawal Scrbed by: Zhpeng Lu 1 Problem defnton In the regular Mult-armed

More information

Problem Set 9 Solutions

Problem Set 9 Solutions Desgn and Analyss of Algorthms May 4, 2015 Massachusetts Insttute of Technology 6.046J/18.410J Profs. Erk Demane, Srn Devadas, and Nancy Lynch Problem Set 9 Solutons Problem Set 9 Solutons Ths problem

More information

Introduction to Vapor/Liquid Equilibrium, part 2. Raoult s Law:

Introduction to Vapor/Liquid Equilibrium, part 2. Raoult s Law: CE304, Sprng 2004 Lecture 4 Introducton to Vapor/Lqud Equlbrum, part 2 Raoult s Law: The smplest model that allows us do VLE calculatons s obtaned when we assume that the vapor phase s an deal gas, and

More information

A New Refinement of Jacobi Method for Solution of Linear System Equations AX=b

A New Refinement of Jacobi Method for Solution of Linear System Equations AX=b Int J Contemp Math Scences, Vol 3, 28, no 17, 819-827 A New Refnement of Jacob Method for Soluton of Lnear System Equatons AX=b F Naem Dafchah Department of Mathematcs, Faculty of Scences Unversty of Gulan,

More information

Bézier curves. Michael S. Floater. September 10, These notes provide an introduction to Bézier curves. i=0

Bézier curves. Michael S. Floater. September 10, These notes provide an introduction to Bézier curves. i=0 Bézer curves Mchael S. Floater September 1, 215 These notes provde an ntroducton to Bézer curves. 1 Bernsten polynomals Recall that a real polynomal of a real varable x R, wth degree n, s a functon of

More information

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016 U.C. Berkeley CS94: Spectral Methods and Expanders Handout 8 Luca Trevsan February 7, 06 Lecture 8: Spectral Algorthms Wrap-up In whch we talk about even more generalzatons of Cheeger s nequaltes, and

More information

Ph 219a/CS 219a. Exercises Due: Wednesday 23 October 2013

Ph 219a/CS 219a. Exercises Due: Wednesday 23 October 2013 1 Ph 219a/CS 219a Exercses Due: Wednesday 23 October 2013 1.1 How far apart are two quantum states? Consder two quantum states descrbed by densty operators ρ and ρ n an N-dmensonal Hlbert space, and consder

More information

Homework Assignment 3 Due in class, Thursday October 15

Homework Assignment 3 Due in class, Thursday October 15 Homework Assgnment 3 Due n class, Thursday October 15 SDS 383C Statstcal Modelng I 1 Rdge regresson and Lasso 1. Get the Prostrate cancer data from http://statweb.stanford.edu/~tbs/elemstatlearn/ datasets/prostate.data.

More information

Additional Codes using Finite Difference Method. 1 HJB Equation for Consumption-Saving Problem Without Uncertainty

Additional Codes using Finite Difference Method. 1 HJB Equation for Consumption-Saving Problem Without Uncertainty Addtonal Codes usng Fnte Dfference Method Benamn Moll 1 HJB Equaton for Consumpton-Savng Problem Wthout Uncertanty Before consderng the case wth stochastc ncome n http://www.prnceton.edu/~moll/ HACTproect/HACT_Numercal_Appendx.pdf,

More information

Engineering Risk Benefit Analysis

Engineering Risk Benefit Analysis Engneerng Rsk Beneft Analyss.55, 2.943, 3.577, 6.938, 0.86, 3.62, 6.862, 22.82, ESD.72, ESD.72 RPRA 2. Elements of Probablty Theory George E. Apostolaks Massachusetts Insttute of Technology Sprng 2007

More information

Bezier curves. Michael S. Floater. August 25, These notes provide an introduction to Bezier curves. i=0

Bezier curves. Michael S. Floater. August 25, These notes provide an introduction to Bezier curves. i=0 Bezer curves Mchael S. Floater August 25, 211 These notes provde an ntroducton to Bezer curves. 1 Bernsten polynomals Recall that a real polynomal of a real varable x R, wth degree n, s a functon of the

More information