Plug-and-Play Dual-Tree Algorithm Runtime Analysis

Size: px
Start display at page:

Download "Plug-and-Play Dual-Tree Algorithm Runtime Analysis"

Transcription

1 Jounal of Machine Leaning Reseach 16 (2015) Submitted 1/15; Published 12/15 Plug-and-Play Dual-Tee Algoithm Runtime Analysis Ryan R. Cutin School of Computational Science and Engineeing Geogia Institute of Technology Atlanta, GA , USA Dongyeol Lee Yahoo Labs Sunnyvale, CA William B. Mach Institute fo Computational Engineeing and Sciences Univesity of Texas, Austin Austin, TX Paikshit Ram Skytee, Inc. Atlanta, GA Edito: Nando de Feitas Abstact Numeous machine leaning algoithms contain paiwise statistical poblems at thei coe that is, tasks that equie computations ove all pais of input points if implemented naively. Often, tee stuctues ae used to solve these poblems efficiently. Dual-tee algoithms can efficiently solve o appoximate many of these poblems. Using cove tees, igoous wostcase untime guaantees have been poven fo some of these algoithms. In this pape, we pesent a poblem-independent untime guaantee fo any dual-tee algoithm using the cove tee, sepaating out the poblem-dependent and the poblem-independent elements. This allows us to just plug in bounds fo the poblem-dependent elements to get untime guaantees fo dual-tee algoithms fo any paiwise statistical poblem without e-deiving the entie poof. We demonstate this plug-and-play pocedue fo neaest-neighbo seach and appoximate kenel density estimation to get impoved untime guaantees. Unde mild assumptions, we also pesent the fist linea untime guaantee fo dual-tee based ange seach. Keywods: dual-tee algoithms, adaptive untime analysis, cove tee, expansion constant, neaest neighbo seach, kenel density estimation, ange seach 1. Dual-tee Algoithms A supising numbe of machine leaning algoithms have computational bottlenecks that can be expessed as paiwise statistical poblems. By this, we mean computational tasks that can be evaluated diectly by iteating ove all pais of input points. Neaest neighbo seach is one such poblem, since fo evey quey point, we can evaluate its distance to evey efeence point and keep the closest one. This naively equies O(N) time to answe a single quey in a efeence set of size N; answeing O(N) queies subsequently equies c 2015 Ryan R. Cutin, Dongyeol Lee, William B. Mach, Paikshit Ram.

2 Cutin, Lee, Mach, and Ram pohibitive O(N 2 ) time. Kenel density estimation is also a paiwise statistical poblem, since we compute a sum ove all efeence points fo each quey point. This again equies O(N 2 ) time to answe O(N) queies if done diectly. The efeence set is typically indexed with spatial data stuctues to acceleate this type of computation (Finkel and Bentley, 1974; Beygelzime et al., 2006); these esult in O(log N) untime pe quey unde favoable conditions. Building upon this intuition, Gay and Mooe (2001) genealized the fast multipole method fom computational physics to obtain dual-tee algoithms. These ae extemely useful when thee ae lage quey sets, not just a few quey points. Instead of building a tee on the efeence set and seaching with each quey point sepaately, Gay and Mooe suggest also building a quey tee and tavesing both the quey and efeence tees simultaneously (a dual-tee tavesal, fom which the class of algoithms takes its name). Dual-tee algoithms can be easily undestood though the ecent famewok of Cutin et al. (2013b): two tees (a quey tee and a efeence tee) ae tavesed by a puning dualtee tavesal. This tavesal visits combinations of nodes fom the tees in some sequence (each combination consisting of a quey node and a efeence node), calling a poblemspecific Scoe() function to detemine if the node combination can be puned. If not, then a poblem-specific BaseCase() function is called fo each combination of points held in the quey node and efeence node. This has significant similaity to the moe common single-tee banch-and-bound algoithms, except that the algoithm must ecuse into child nodes of both the quey tee and efeence tee. Thee exist numeous dual-tee algoithms fo poblems as divese as kenel density estimation (Gay and Mooe, 2003), mean shift (Wang et al., 2007), minimum spanning tee calculation (Mach et al., 2010), n-point coelation function estimation (Mach et al., 2012), max-kenel seach (Cutin et al., 2013c), paticle smoothing (Klaas et al., 2006), vaiational infeence (Amizadeh et al., 2012), ange seach (Gay and Mooe, 2001), and embedding techniques (Van De Maaten, 2014), to name a few. Some of these algoithms ae deived using the cove tee (Beygelzime et al., 2006), a data stuctue with compelling theoetical qualities. When cove tees ae used, dual-tee all-neaest-neighbo seach and appoximate kenel density estimation have O(N) untime guaantees fo O(N) queies (Ram et al., 2009a); minimum spanning tee calculation scales as O(N log N) (Mach et al., 2010). Othe poblems have simila wost-case guaantees (Cutin and Ram, 2014; Mach, 2013). In this wok we combine the genealization of Cutin et al. (2013b) with the theoetical esults of Beygelzime et al. (2006) and othes in ode to develop a wost-case untime bound fo any dual-tee algoithm when the cove tee is used. Section 2 lays out the equied backgound, notation, and intoduces the cove tee and its associated theoetical popeties. Reades familia with the cove tee liteatue and dual-tee algoithms (especially Cutin et al., 2013b) may find that section to be eview. Following that, we intoduce an intuitive measue of cove tee imbalance, an impotant popety fo undestanding the untime of dual-tee algoithms, in Section 3. This measue of imbalance is then used to pove the main esult of the pape in Section 4, which is a wost-case untime bound fo genealized dual-tee algoithms. We apply this esult to thee specific poblems: neaest neighbo seach (Section 5), appoximate kenel density estimation (Section 6), and ange seach / ange count (Section 7), showing linea untime 3270

3 Plug-and-Play Dual-Tee Algoithm Runtime Analysis Symbol N Desciption A tee node Set of child nodes of N i Set of points held in N i Set of descendant nodes of N i Set of points contained in N i and Di n µ i Cente of N i λ i Futhest descendant distance fom µ i C i P i Di n D p i Table 1: Notation fo tees. See Cutin et al. (2013b) fo details. bounds fo each of those algoithms. Each of these bounds is an impovement on the stateof-the-at, and in the case of ange seach, is the fist such bound. Despite the intuition this povides fo the scaling popeties of all dual-tee algoithms 1, it must be kept in mind that these wost-case bounds only apply to dual-tee algoithms that use the cove tee and the standad cove tee tavesal. 2. Peliminaies Fo simplicity, the algoithms consideed in this pape will be pesented in a tee-independent context, as in Cutin et al. (2013b), but the only type of tee we will conside is the cove tee (Beygelzime et al., 2006), and the only type of tavesal we will conside is the cove tee puning dual-tee tavesal, which we will descibe late. As we will be making heavy use of tees, we must establish notation (taken fom Cutin et al., 2013b). The notation we will be using is defined in Table The Cove Tee The cove tee is a leveled hieachical data stuctue oiginally poposed fo the task of neaest neighbo seach by Beygelzime et al. (2006). Each node N i in the cove tee is associated with a single point p i. An adequate desciption is given in thei wok (we have adapted notation slightly): A cove tee T on a dataset S is a leveled tee whee each level is a cove fo the level beneath it. Each level is indexed by an intege scale s i which deceases as the tee is descended. Evey node in the tee is associated with a point in S. Each point in S may be associated with multiple nodes in the tee; howeve, we equie that any point appeas at most once in evey level. Let C si denote the set of points in S associated with the nodes at level s i. The cove tee obeys the following invaiants fo all s i : 1. Dual-tee algoithms using kd-tees and othe types of tees have been obseved to empiically scale linealy fo tasks that take quadatic time without the use of tees; see the empiical esults of Gay and Mooe (2001); Mach et al. (2010); Vladymyov and Caeia-Pepinán (2014); Klaas et al. (2006); Gay and Mooe (2003). 3271

4 Cutin, Lee, Mach, and Ram (Nesting). C si C si 1. This implies that once a point p S appeas in C si then evey lowe level in the tee has a node associated with p. (Coveing tee). Fo evey p i C si 1, thee exists a p j C si such that d(p i, p j ) < 2 s i and the node in level s i associated with p j is a paent of the node in level s i 1 associated with p i. (Sepaation). Fo all distinct p i, p j C si, d(p i, p j ) > 2 s i. As a consequence of this definition, if thee exists a node N i, containing the point p i at some scale s i, then thee will also exist a self-child node N ic containing the point p i at scale s i 1 which is a child of N i. In addition, evey descendant point of the node N i is contained within a ball of adius 2 s i+1 centeed at the point p i ; theefoe, λ i = 2 s i+1 and µ i = p i (Table 1). Note that the cove tee may be intepeted as an infinite-leveled tee, with C containing only the oot point, C = S, and all levels between defined as above. Beygelzime et al. (2006) find this epesentation (which they call the implicit epesentation) easie fo desciption of thei algoithms and some of thei poofs. But clealy, this is not suitable fo implementation; hence, thee is an explicit epesentation in which all nodes that have only a self-child ae coalesced upwads (that is, the node s self-child is emoved, and the childen of that self-child ae taken to be the childen of the node). Figue 1 shows each of the levels of an example cove tee (in the explicit epesentation) on a simple six-point dataset. In this wok, we conside only the explicit epesentation of a cove tee, and do not concen ouselves with the details of tee constuction Expansion Constant The explicit epesentation of a cove tee has a numbe of useful theoetical popeties based on the expansion constant (Kage and Ruhl, 2002); we estate its definition below. Definition 1 Let B S (p, ) be the set of points in S within a closed ball of adius aound some p S with espect to a metic d: B S (p, ) = { S : d(p, ) }. Then, the expansion constant of S with espect to the metic d is the smallest c 2 such that B S (p, 2 ) c B S (p, ) p S, > 0. (1) The expansion constant is used heavily in the cove tee liteatue. It is, in some sense, a notion of instinic dimensionality, most useful in scenaios whee c is independent of the numbe of points in the dataset (Kage and Ruhl, 2002; Beygelzime et al., 2006; Kauthgame and Lee, 2004; Ram et al., 2009a). Note also that if points in S H ae being dawn accoding to a stationay distibution f(x), then c will convege to some finite value c f as S. To see this, define c f as a genealization of the expansion constant fo distibutions. c f 2 is the smallest value such that f(x)dx c f f(x)dx (2) B H (p,2 ) B H (p, ) 2. A batch constuction algoithm is given by Beygelzime et al. (2006), called Constuct. 3272

5 Plug-and-Play Dual-Tee Algoithm Runtime Analysis N a N b p 5 p 2 p 0 p 4 4 p 5 p 2 p 0 p 4 2 N c p 1 p 3 p 1 p 3 2 (a) Root node at scale 1. (b) Nodes at scale 0. N a scale 1 scale 0 1 p 5 p 2 1 N e N b N c scale 1 p 0 p 4 N d N e N d p 1 p 3 N 0 N 4 N 2 N 5 N 1 N 3 scale (c) Nodes at scale 1. (d) Abstact epesentation. Figue 1: Example cove tee on six points in R 2. (a) N a is centeed at p 0 with scale 1. (b) N b and N c ae centeed at p 0 and p 1, espectively, and have scale 0. (c) N d and N e ae centeed at p 0 and p 2, espectively, and have scale 1. The leaves, N 0 though N 6, ae centeed at each of the six points, with scale (and theefoe adius 0). Note that although node N b in subfigue (b) ovelaps node N c, point p 1 only belongs to N c, not N b. Note also that this is only one valid cove tee that could be built on the data; othe configuations ae possible; fo instance, selecting a diffeent oot point gives diffeent valid cove tees. 3273

6 Cutin, Lee, Mach, and Ram fo all p H and > 0 such that B H (p, ) f(x)dx > 0, and with B H(p, ) defined as the closed ball of adius in the space H. As a simple example, take f(x) as a unifom spheical distibution in R d : fo any x 1, f(x) is a constant; fo x > 1, f(x) = 0. It is easy to see that c f in this situation is 2 d, and thus fo some dataset S, c must convege to that value as moe and moe points ae added to S. Closed-fom solutions fo c f fo moe complex distibutions ae less easy to deive; howeve, empiical speedup esults fom Beygelzime et al. (2006) suggest the existence of datasets whee c is not stongly dependent on d. Fo instance, the covtype dataset has 54 dimensions but the expansion constant is much smalle than othe, lowe-dimensional datasets. Thee ae some othe impotant obsevations about the behavio of c. Adding a single point to S may incease c abitaily: conside a set S distibuted entiely on the suface of a unit hypesphee. If one adds a single point at the oigin, poducing the set S, then c explodes to S wheeas befoe it may have been much smalle than S. Adding a single point may also decease c significantly. Suppose one adds a point abitaily close to the oigin to S ; now, the expansion constant will be S /2. Both of these situations ae degeneate cases not commonly encounteed in eal-wold behavio; we discuss them in ode to point out that although we can bound the behavio of c as S fo S fom a stationay distibution, we ae not able to easily say much about its convegence behavio. The expansion constant can be used to show a few useful bounds on vaious popeties of the cove tee; we estate these esults below, given some cove tee built on a dataset S with expansion constant c and S = N: Width bound: no cove tee node has moe than c 4 childen (Lemma 4.1, Beygelzime et al., 2006). Depth bound: the maximum depth of any node is O(c 2 log N) (Lemma 4.3, Beygelzime et al., 2006). Space bound: a cove tee has O(N) nodes (Theoem 1, Beygelzime et al., 2006). Lastly, we intoduce a convenience lemma of ou own which is a genealization of the packing aguments used by Beygelzime et al. (2006). This is a moe flexible vesion of thei agument. Lemma 1 Conside a dataset S with expansion constant c and a subset C S such that evey two distinct points in C ae sepaated by at least δ. Then, fo any point p (which may o may not be in S), and any adius ρδ > 0: B S (p, ρδ) C c 2+ log 2 ρ. (3) Poof The poof is based on the packing agument fom Lemma 4.1 in Beygelzime et al. (2006). Conside two cases: fist, let d(p, p i ) > ρδ fo any p i S. In this case, B S (p, ρδ) = and the lemma holds tivially. Othewise, let p i S be a point such that d(p, p i ) ρδ. Obseve that B S (p, ρδ) B S (p i, 2ρδ). Also, B S (p i, 2ρδ) c 2+ log 2 ρ B S (p i, δ/2) by the 3274

7 Plug-and-Play Dual-Tee Algoithm Runtime Analysis definition of the expansion constant. Because each point in C is sepaated by δ, the numbe of points in B S (p, ρδ) C is bounded by the numbe of disjoint balls of adius δ/2 that can be packed into B S (p, ρδ). In the wost case, this packing is pefect, and B S (p, ρδ) B S(p i, 2ρδ) B S (p i, δ/2) c2+ log 2 ρ. (4) 3. Tee Imbalance It is well-known that imbalance in tees leads to degadation in pefomance; fo instance, a kd-tee node with evey descendant in its left child except one is effectively useless. A kd-tee full of nodes like this will pefom abysmally fo neaest neighbo seach, and it is not had to geneate a pathological dataset that will cause a kd-tee of this sot. This sot of imbalance applies to all types of tees, not just kd-tees. In ou situation, we ae inteested in a bette undestanding of this imbalance fo cove tees, and thus endeavo to intoduce a moe fomal measue of imbalance which is coelated with tee pefomance. Numeous measues of tee imbalance have aleady been established; one example is that poposed by Colless (1982), and anothe is Sackin s index (Sackin, 1972), but we aim to captue a diffeent measue of imbalance that uses the leveled stuctue of the cove tee. We aleady know each node in a cove tee is indexed with an intege level (o scale). In the explicit epesentation of the cove tee, each non-leaf node has childen at a lowe level. But these childen need not be stictly one level lowe; see Figue 2. In Figue 2a, each cove tee node has childen that ae stictly one level lowe; we will efe to this as a pefectly balanced cove tee. Figue 2b, on the othe hand, contains the node N m which has two childen with scale two less than s m. We will efe to this as an imbalanced cove tee. Note that in ou definition, the balance of a cove tee has nothing to do with diffeing numbe of descendants in each child banch but instead only missing levels. An imbalanced cove tee can happen in pactice, and in the wost cases, the imbalance may be fa wose than the simple gaphs of Figue 2. Conside a dataset with a single N a s a N m s m s a 1 s m 1 N b N c N d N n N e N f N g N h N i N j s a 2 N k N p N q N N s s m 2 N t (a) Balanced cove tee. (b) Imbalanced cove tee. Figue 2: Balanced and imbalanced cove tees. 3275

8 Cutin, Lee, Mach, and Ram outlie Figue 3: Single-outlie cove tee. Figue 4: A multiple-outlie cove tee. outlie which is vey fa away fom all of the othe points 3. Figue 3 shows what happens in this situation: the oot node has two childen; one of these childen has only the outlie as a descendant, and the othe child has the est of the points in the dataset as a descendant. In fact, it is easy to find datasets with a handful of outlies that give ise to a chain-like stuctue at the top of the tee: see Figue 4 fo an illustation 4. A tee that has this chain-like stuctue all the way down, which is simila to the kd-tee example at the beginning of this section, is going to pefom hoendously; motivated by this obsevation, we define a measue of tee imbalance. Definition 2 The cove node imbalance I n (N i ) fo a cove tee node N i with scale s i in the cove tee T is defined as the cumulative numbe of missing levels between the node and its paent N p (which has scale s p ). If the node is a leaf (that is, s i = ), then the numbe of missing levels is defined as the diffeence between s p and s min 1 whee s min is the smallest scale of a non-leaf node in T. If N i is the oot of the tee, then the cove node imbalance is 0. Explicitly witten, this calculation is s p s i 1 if N i is not a leaf and not the oot node I n (N i ) = max(s p s min 1, 0) if N i is a leaf 0 if N i is the oot node. 3. Note also that fo an outlie sufficiently fa away, the expansion constant is N 1, so we should expect poo pefomance with the cove tee anyway. 4. As a side note, this behavio is not limited to cove tees, and can happen to mean-split kd-tees too, especially in highe dimensions. In addition, fo this scenaio to aise with cove tees, it must be tue that c O(N). (5) 3276

9 Plug-and-Play Dual-Tee Algoithm Runtime Analysis Imbalance Dataset d N = 5k N = 50k N = 500k lcdm sdss powe susy andu higgs covetype mnist Table 2: Empiically calculated tee imbalances, nomalized by N. This simple definition of cove node imbalance is easy to calculate, and using it, we can genealize to a measue of imbalance fo the full tee. Definition 3 The cove tee imbalance I t (T ) fo a cove tee T is defined as the cumulative numbe of missing levels in the tee. This can be expessed as a function of cove node imbalances easily: I t (T ) = N i T I n (N i ). (6) A pefectly balanced cove tee T b with no missing levels has imbalance I t (T b ) = 0 (fo instance, Figue 2a). A wost-case cove tee T w which is entiely a chain-like stuctue with maximum scale s max and minimum scale s min will have imbalance I t (T w ) N(s max s min ). Because of this chain-like stuctue, each level has only one node and thus thee ae at least N levels; o, s max s min N, meaning that in the wost case the imbalance is quadatic in N. 5 Howeve, fo most eal-wold datasets with the cove tee implementation in mlpack (Cutin et al., 2013a) and the efeence implementation (Beygelzime et al., 2006), the tee imbalance is nea-linea with the numbe of points. We have constucted cove tees on N unifomly subsampled points fom a vaiety of datasets and calculated the imbalance; see Table 2 fo the esults. Ten tials wee pefomed fo each dataset and each N, and the mean imbalance is given. These esults ae nomalized with espect to N, fo which the values of 5000, 50000, and wee chosen. The powe, susy, higgs, and covetype datasets ae found in the UCI Machine Leaning Repositoy (Bache and Lichman, 2013), the mnist dataset is fom LeCun et al. (2000), the lcdm and sdss datasets ae Sloan Digital Sky Suvey data (Adelman-McCathy et al., 2008), and the andu dataset is andomlygeneated unifomly-distibuted data in 10 dimensions. The imbalances on each of these datasets tend to be nea-linea. Cuently, no cove tee constuction algoithm specifically aims to minimize imbalance. 5. Note that in this situation, c N also. 3277

10 Cutin, Lee, Mach, and Ram Algoithm 1 The standad puning dual-tee tavesal fo cove tees. 1: Input: quey node N q, set of efeence nodes R 2: Output: none max N R s 4: if (s q < s max ) then 5: {Pefom a efeence ecusion.} 6: fo each N R do 7: BaseCase(p q, p ) 8: end fo 9: R {N R : s = s max } 10: R 1 {C (N ) : N R } (R \ R ) 11: R 1 {N R 1 : Scoe(N q,n ) } 3: s max 12: ecuse with N q and R 1 13: else 14: {Pefom a quey ecusion.} 15: fo each N qc C (N q ) do 16: R {N R : Scoe(N qc,n ) } 17: ecuse with N qc and R 18: end fo 19: end if 4. Geneal Runtime Bound Pehaps moe inteesting than measues of tee imbalance is the way cove tees ae actually used in dual-tee algoithms. Although cove tees wee oiginally intended fo neaest neighbo seach (See Algoithm Find-All-Neaest, Beygelzime et al., 2006), they can be adapted to a wide vaiety of poblems: minimum spanning tee calculation (Mach et al., 2010), appoximate neaest neighbo seach (Ram et al., 2009b), Gaussian pocesses posteio calculation (Mooe and Russell, 2014), and max-kenel seach (Cutin and Ram, 2014) ae some examples. Futhe, though the tee-independent dual-tee algoithm abstaction of Cutin et al. (2013b), othe existing dual-tee algoithms can easily be adapted fo use with cove tees. In the famewok of tee-independent dual-tee algoithms, all that is necessay to descibe a dual-tee algoithm is a point-to-point base case function (BaseCase()) and a node-to-node puning ule (Scoe()). These functions, which ae often vey staightfowad, ae then paied with a type of tee and a puning dual-tee tavesal to poduce a woking algoithm. In late sections, we will conside specific examples. When using cove tees, the typical puning dual-tee tavesal is an adapted fom of the oiginal neaest neighbo seach algoithm (see Find-All-Neaest, Beygelzime et al., 2006); this tavesal is implemented in both the cove tee efeence implementation and in the moe flexible mlpack libay (Cutin et al., 2013a). The poblem-independent tavesal is given in Algoithm 1 and was oiginally pesented by Cutin and Ram (2014). Initially, it is called with the oot of the quey tee and a efeence set R containing only the oot of the efeence tee. 3278

11 Plug-and-Play Dual-Tee Algoithm Runtime Analysis This dual-tee ecusion is a depth-fist ecusion in the quey tee and a beadth-fist ecusion in the efeence tee; to this end, the ecusion maintains one quey node N q and a efeence set R. The set R may contain efeence nodes with many diffeent scales; the maximum scale in the efeence set is s max (line 3). Each single ecusion will descend eithe the quey tee o the efeence tee, not both; the conditional in line 4, which detemines whethe the quey o efeence tee will be ecused, is aimed at keeping the elative scales of quey nodes and efeence nodes close. Keeping the quey and efeence scales close is both beneficial fo the late theoy and intuitively easonable: ecusing too quickly in the eithe the quey o efeence node will unnecessaily duplicate wok. Suppose we ecuse many levels down the quey tee befoe ecusing down the efeence tee, giving us a set of quey nodes we ae consideing. Fo each of these quey nodes, we will then need to descend the efeence tee. Because these quey nodes ae close togethe (with espect to the efeence nodes we ae consideing, which ae of lage scale and thus futhe apat), the puning decisions at each level of ecusion ae likely to be the same fo each quey node. Theefoe, ecusing too fa in the quey tee may cause a lage amount of duplicated wok. The symmetic agument applies fo ecusing too fa in the efeence tee befoe ecusing in the quey tee. This justifies the appoach of keeping the quey and efeence scales appoximately equal. A quey ecusion (lines 13 18) is staightfowad: fo each child N qc of N q, the node combinations (N qc, N ) ae scoed fo each N in the efeence set R. If possible, these combinations ae puned to fom the set R (line 17) by checking the output of the Scoe() function, and then the algoithm ecuses with N qc and R. A efeence ecusion (lines 4 12) is simila to a quey ecusion, but the puning stategy is significantly moe complicated. Given R, we calculate R, which is the set of nodes in R that have scale s max. We expand each node in R to constuct R 1 : this is the set of childen of all nodes in R. This set is then combined with R \ R (that is, the set of efeences nodes not at scale s max ) to poduce R 1. Each node in R 1 is then scoed and puned if possible, esulting in the puned efeence set R 1. The algoithm then ecuses with N q and R 1. The efeence ecusion only ecuses into the top-level subset of the efeence nodes in ode to peseve the sepaation invaiant. It is easy to show that evey pai of points held in nodes in R is sepaated by at least 2 smax : Lemma 2 Fo all distinct nodes N i, N j R (in the context of Algoithm 1) which contain points p i and p j, espectively, d(p i, p j ) > 2 smax, with s max defined as in line 3. Poof This poof is by induction. If R = 1, such as duing the fist efeence ecusion, the esult obviously holds. Now conside any efeence set R and assume the statement of the lemma holds fo this set R, and define s max as the maximum scale of any node in R. Constuct the set R 1 as in line 10 of Algoithm 1; if R 1 1, then R 1 satisfies the desied popety. Othewise, take any N i, N j in R 1, with points p i and p j, espectively, and scales s i and s j, espectively. Clealy, if s i = s j = s max 1, then by the sepaation invaiant d(p i, p j ) > 2 smax 1. Now suppose that s i < s max 1. This implies that thee exists some implicit cove tee node with point p i and scale s max 1 (as well as an implicit child of this node p i with scale 3279

12 Cutin, Lee, Mach, and Ram 2 and so foth until one of these implicit nodes has child p i with scale s i ). Because the sepaation invaiant applies to both implicit and explicit epesentations of the tee, we conclude that d(p i, p j ) > 2 smax 1. The same agument may be made fo the case whee s j < s max 1, with the same conclusion. We may theefoe conclude that each point of each node in R 1 is sepaated by 2 smax 1. Note that R 1 R 1 and that R \ R 1 R in ode to see that this condition holds fo all nodes in R 1. Because we have shown that the condition holds fo the initial efeence set and fo any efeence set poduced by a efeence ecusion (which will be R at some othe level of ecusion), we have shown that the statement of the lemma is tue. Note that in this poof, we have consideed the child efeence set R 1, not the oiginal s max efeence set R, and shown that with espect to s max as defined by R (not R 1 ), all nodes ae sepaated by 2 smax 1. Then, in the fame of the next ecusion whee R R 1, the lemma will hold, as s max will then be the maximum scale pesent in R. This obsevation means that the set of points P held by all nodes in R is always a subset of C s max. This fact will be useful in ou late untime poofs. Next, we develop notions with which to undestand the behavio of the cove tee dualtee tavesal when the datasets ae of significantly diffeent scale distibutions. If the datasets ae simila in scale distibution (that is, inte-point distances tend to follow the same distibution), then the ecusion will altenate between quey ecusions and efeence ecusions. But if the quey set contains points which ae, in geneal, much fathe apat than the efeence set, then the ecusion will stat with many quey ecusions befoe eaching a efeence ecusion. The convese case also holds. We ae inteested in fomalizing this notion of scale distibution; theefoe, define the following dataset-dependent constants fo the quey set S q and the efeence set S : η q : the lagest paiwise distance in S q δ q : the smallest nonzeo paiwise distance in S q η : the lagest paiwise distance in S δ : the smallest nonzeo paiwise distance in S These constants ae diectly elated to the aspect atio of the datasets; indeed, η q /δ q is exactly the aspect atio of S q. Futhe, let us define and bound the top and bottom levels of each tee: The top scale s T q of the quey tee T q is such that as log 2 (η q ) 1 s T q log 2 (η q ). The minimum scale of the quey tee T q is defined as s min q = log 2 (δ q ). The top scale s T of the efeence tee T is such that as log 2 (η ) 1 s T log 2 (η ). The minimum scale of the efeence tee T is defined as s min = log 2 (δ ). 3280

13 Plug-and-Play Dual-Tee Algoithm Runtime Analysis Note that the minimum scale is not the minimum scale of any cove tee node (that would be ), but the minimum scale of any non-leaf node in the tee. Suppose that ou datasets ae of a simila scale distibution: s T q = s T, and s min q = s min. In this setting we will have altenating quey and efeence ecusions. But if this is not the case, then we have exta efeence ecusions befoe the fist quey ecusion o afte the last quey ecusion (situations whee both these cases happen ae possible). Motivated by this obsevation, let us quantify these exta efeence ecusions: Lemma 3 Fo a dual-tee algoithm with S q S O(N) using cove tees and the tavesal given in Algoithm 1, the numbe of exta efeence ecusions that happen befoe the fist quey ecusion is bounded by min (O(N), log 2 (η /η q ) 1). (7) Poof The fist quey ecusion happens once s q s max. The numbe of efeence ecusions befoe the fist quey ecusion is then bounded as the numbe of levels in the efeence tee between s T and s T q that have at least one explicit node. Because thee ae O(N) nodes in the efeence tee, the numbe of levels cannot be geate than O(N) and thus the esult holds. The second bound holds by applying the definitions of s T and s T q to the expession s T s T q 1: s T s T q 1 log 2 (η ) ( log 2 (η q ) 1) 1 (8) log 2 (η ) + 1 log 2 (η q ) (9) which gives the statement of the lemma afte applying logaithmic identities. Note that the O(N) bound may be somewhat loose, but it suffices fo ou late puposes. Now let us conside the othe case: Lemma 4 Fo a dual-tee algoithm with S q S O(N) using cove tees and the tavesal given in Algoithm 1, the numbe of exta efeence ecusions that happen afte the last quey ecusion is bounded by max ( min ( O(N log 2 (δ q /δ )), O(N 2 ) ), 0 ). (10) Fo convenience, we define a tem that encapsulates this bound. Definition 4 Define θ as a bound on the numbe of exta efeence ecusions that happen afte the last quey ecusion. Then, θ = max { min ( O(N log 2 (δ q /δ )), O(N 2 ) ), 0 }. (11) 3281

14 Cutin, Lee, Mach, and Ram Poof Ou goal hee is to count the numbe of efeence ecusions afte the final quey ecusion at level s min q ; the fist of these efeence ecusions is at scale s max = s min q. Because quey nodes ae not puned in this tavesal, each efeence ecusion we ae counting will be duplicated ove the whole set of O(N) quey nodes. The fist pat of the bound follows by obseving that s min q s min log 2 (δ q ) log 2 (δ ) 1 log 2 (δ q /δ ). The second pat follows by simply obseving that thee ae O(N) efeence nodes. These two pevious lemmas allow us a bette undestanding of what happens as the efeence set and quey set become diffeent. Lemma 3 shows that the numbe of exta ecusions caused by a efeence set with lage paiwise distances than the quey set (η lage than η q ) is modest; on the othe hand, Lemma 4 shows that fo each exta level in the efeence tee below s min q, O(N) exta ecusions ae equied. Using these lemmas and this intuition, we will pove geneal untime bounds fo the cove tee tavesal. Theoem 1 Given a efeence set S of size O(N) with an expansion constant c and a set of queies S q of size O(N), a standad cove tee based dual-tee algoithm (Algoithm 1) takes O ( c 4 R χψ(n + I t (T q ) + θ) ) (12) time, whee R is the maximum size of the efeence set R (line 1) duing the dual-tee ecusion, χ is the maximum possible untime of BaseCase(), ψ is the maximum possible untime of Scoe(), and θ is defined as in Lemma 4. Poof Fist, split the algoithm into two pats: efeence ecusions (lines 4 12) and quey ecusions (lines 13 18). The untime of the algoithm is the untime of a efeence ecusion times the total numbe of efeence ecusions plus the total untime of all quey ecusions. Conside a efeence ecusion (lines 4 12). Define R to be the lagest set R fo any and any quey node N q duing the couse of the algoithm; then, it is tue that R R. The wok done in the base case loop fom lines 6 8 is thus O(χ R ) O(χ R ). Then, lines 10 and 11 take O(c 4 ψ R ) O(c 4 ψ R ) time, because each efeence node has up to c 4 childen. So, one full efeence ecusion takes O(c 4 ψχ R ) time. scale s max Now, note that thee ae O(N) nodes in T q. Thus, line 17 is visited O(N) times. The amount of wok in line 16, like in the efeence ecusion, is bounded as O(c 4 ψ R ). Theefoe, the total untime of all quey ecusions is O(c 4 ψ R N). Lastly, we must bound the total numbe of efeence ecusions. Refeence ecusions happen in thee cases: (1) s max is geate than the scale of the oot of the quey tee (no quey ecusions have happened yet); (2) s max is less than o equal to the scale of the oot of the quey tee, but is geate than the minimum scale of the quey tee that is not ; (3) s max is less than the minimum scale of the quey tee that is not. Fist, conside case (1). Lemma 3 shows that the numbe of efeence ecusions of this type is bounded by O(N). Although thee is also a bound that depends on the sizes of the datasets, we only aim to show a linea untime bound, so the O(N) bound is sufficient hee. Next, conside case (2). In this situation, each quey ecusion implies at least one efeence ecusion befoe anothe quey ecusion. Fo some quey node N q, the exact numbe of efeence ecusions befoe the childen of N q ae ecused into is bounded above 3282

15 Plug-and-Play Dual-Tee Algoithm Runtime Analysis by I n (N q ) + 1: if N q has imbalance 0, then it is exactly one level below its paent, and thus thee is only one efeence ecusion. On the othe hand, if N q is many levels below its paent, then it is possible that a efeence ecusion may occu fo each level in between; this is a maximum of I n (N q ) + 1. Because each quey node in T q is ecused into once, the total numbe of efeence ecusions befoe each quey ecusion is N q T q I n (N q ) + 1 = I t (T q ) + O(N) (13) since thee ae O(N) nodes in the quey tee. Lastly, fo case (3), we may efe to Lemma 4, giving a bound of θ efeence ecusions in this case. We may now combine these esults fo the untime of a quey ecusions with the total numbe of efeence ecusions in ode to give the esult of the theoem: O ( c 4 R ψχ (N + I t (T q ) + θ) ) + O ( c 4 R ψn ) O ( c 4 R ψχ (N + I t (T q ) + θ) ). (14) When we conside the monochomatic case (whee S q = S ), the esults tivially simplify. Coollay 1 Given the situation of Theoem 1 but with S q = S = S so that c q = c = c and T q = T = T, a dual-tee algoithm using the standad cove tee tavesal (Algoithm 1) takes O ( c 4 R χψ (N + I t (T )) ) (15) time, whee R is the maximum size of the efeence set R (line 1) duing the dualtee ecusion, χ is the maximum possible untime of BaseCase(), and ψ is the maximum possible untime of Scoe(). An intuitive undestanding of these bounds is best achieved by fist consideing the monochomatic case (this case aises, fo instance, in all-neaest-neighbo seach). The linea dependence on N aises fom the fact that all quey nodes must be visited. The dependence on the efeence tee, howeve, is encapsulated by the tem c 4 R, with R being the maximum size of the efeence set R; this value must be deived fo each specific poblem. The poo pefomance of tees on datasets with lage c (o, in the wost case, c N) is then captued in both of those tems. These datasets fo which tees pefom pooly may also have a high cove tee imbalance I t (T ); the linea dependence of untime on imbalance is thus sensible fo datasets whee tees pefom well. The bichomatic case (S q S ) is a slightly moe complex esult which deseves a bit moe attention. The intuition fo all tems except θ emain vitually the same. The tem θ captues the effect of quey and efeence datasets with diffeent widths, and has one unfotunate cone case: when δ q > η, then the quey tee must be entiely descended befoe any efeence ecusion. This esults in a bound of the fom O(N log(η /δ )), 3283

16 Cutin, Lee, Mach, and Ram o O(N 2 ) (see Lemma 4). This is because the efeence tee must be descended sepaately fo each quey point. The quantity R bounds the amount of wok that needs to be done fo each ecusion. In the wost case, R can be N. Howeve, dual-tee algoithms ely on banch-and-bound techniques to pune away wok (lines 11 and 16 in Algoithm 1). A small value of R will imply that the algoithm is extemely successful in puning away wok. An (uppe) bound on R (and the algoithm s success in puning wok) will depend on the poblem and the data. As we will show, bounding R is often possible. Fo many dual-tee algoithms, χ ψ O(1); often, cached sufficient statistics (Mooe, 2000) can enable O(1) untime implementations of BaseCase() and Scoe(). These esults hold fo any dual-tee algoithm egadless of the poblem. Hence, the untime of any dual-tee algoithm can be bounded no moe tightly than O(N) with ou bound, which matches the intuition that answeing O(N) queies will take at least O(N) time. Fo a paticula poblem and data, if c, R, χ, and ψ ae bounded by constants independent of N and θ is no moe than linea in N (fo lage enough N), then the dual-tee algoithm fo that poblem has a untime linea in N. Ou theoetical esult sepaates out the poblem-dependent and the poblem-independent elements of the untime bound, which allows us to simply plug in the poblem-dependent bounds in ode to get untime bounds fo any dual-tee algoithm without equiing an analysis fom scatch. Ou esults ae simila to that of Ram et al. (2009a), but those esults depend on a quantity called the constant of bichomaticity, denoted κ, which has unclea elation to cove tee imbalance. The dependence on κ is given as c 4κ q, which is not a good bound, especially because κ may be much geate than 1 in the bichomatic case (whee S q S ). The moe ecent esults of Cutin and Ram (2014) ae moe elated to these esults, but they depend on the invese constant of bichomaticity ν which suffes fom the same poblem as κ. Although the dependence on ν is linea (that is, O(νN)), bounding ν is difficult and it is not tue that ν = 1 in the monochomatic case. The quantity ν coesponds to the maximum numbe of efeence ecusions between a single quey ecusion, and κ coesponds to the maximum numbe of quey ecusions between a single efeence ecusion. The espective poofs that use these constants then apply them as a wost-case measue fo the whole algoithm: when using κ, Ram et al. (2009a) assume that evey efeence ecusion may be followed by κ quey ecusions; similaly, Cutin and Ram (2014) assume that evey quey ecusion may be followed by ν efeence ecusions. Hee, we have simply used I t (T q ) and θ as an exact summation of the total exta efeence ecusions, which gives us a much tighte bound than ν o κ on the unning time of the whole algoithm. Futhe, both ν and κ ae difficult to empiically calculate and equie an entie un of the dual-tee algoithm. On the othe hand, bounding I t (T q ) (and θ) can be done in one pass of the tee (assuming the tee is aleady built). Thus, not only is ou bound tighte when the cove tee imbalance is sublinea in N, it moe closely eflects the actual behavio of dual-tee algoithms, and the constants which it depends upon ae staightfowad to calculate. In the following sections, we will apply ou esults to specific poblems and show the utility of ou bound in simplifying untime poofs fo dual-tee algoithms. 3284

17 Plug-and-Play Dual-Tee Algoithm Runtime Analysis Algoithm 2 Neaest neighbo seach BaseCase() Input: quey point p q, efeence point p, list of candidate neighbos N and distances D Output: distance d between p q and p if d(p q, p ) < D[p q ] and BaseCase(p q, p ) not yet called then D[p q ] d(p q, p ), and N[p q ] p end if etun d(p q, p ) Algoithm 3 Neaest neighbo seach Scoe() Input: quey node N q, efeence node N Output: a scoe fo the node combination (N q, N ), o if the combination should be puned if d min (N q, N ) < B(N q ) then etun d min (N q, N ) end if etun 5. Neaest Neighbo Seach The standad task of neaest neighbo seach can be simply descibed: given a quey set S q and a efeence set S, fo each quey point p q S q, find the neaest neighbo p in the efeence set S. The task is well-studied and well-known, and thee exist numeous appoaches fo both exact and appoximate neaest neighbo seach, including the cove tee neaest neighbo seach algoithm due to Beygelzime et al. (2006). We will conside that algoithm, but in a tee-independent sense as given by Cutin et al. (2013b); this means that to descibe the algoithm, we equie only a BaseCase() and Scoe() function; these ae given in Algoithms 2 and 3, espectively. The point-to-point BaseCase() function compaes a quey point p q and a efeence point p, updating the list of candidate neighbos fo p q if necessay. The node-to-node Scoe() function detemines if the entie subtee of nodes unde the efeence node N can impove the candidate neighbos fo all descendant points of the quey node N q ; if not, the node combination is puned. The Scoe() function depends on the function d min (, ), which epesents the minimum possible distance between any two descendants of two nodes. Its definition fo cove tee nodes is d min (N q, N ) = d(p q, p ) 2 sq+1 2 s+1. (16) Given a type of tee and tavesal, these two functions stoe the cuent neaest neighbo candidates in the aay N and thei distances in the aay D. (See Cutin et al., 2013b, fo a moe complete discussion of how this algoithm woks and a poof of coectness.) The Scoe() function depends on a bound function B(N q ) which epesents the maximum distance that could possibly impove a neaest neighbo candidate fo any descendant point 3285

18 Cutin, Lee, Mach, and Ram of the quey node N q. The standad bound function B(N q ) used fo cove tees is adapted fom Beygelzime et al. (2006): B(N q ) := D[p q ] + 2 sq+1 (17) In this fomulation, the quey node N q holds the the quey point p q, the quantity D[p q ] is the cuent neaest neighbo candidate distance fo the quey point p q, and 2 sq+1 coesponds to the futhest descendant distance of N q. Fo notational convenience in the following poof, take c q = max((max pq Sq c ), c ), whee c is the expansion constant of the set S {p q }. Theoem 2 Using cove tees, the standad cove tee puning dual-tee tavesal, and the neaest neighbo seach BaseCase() and Scoe() as given in Algoithms 2 and 3, espectively, and also given a efeence set S of size O(N) with expansion constant c, and a quey set S q of size O(N), the unning time of the algoithm is bounded by O(c 4 c 5 q(n +I t (T q )+θ)) with I t (T q ) and θ defined as in Definition 3 and Lemma 4, espectively. Poof The unning time of BaseCase() and Scoe() ae clealy O(1). Due to Theoem 1, we theefoe know that the untime of the algoithm is bounded by O(c 4 R (N +I t (T q )+θ)). Thus, the only thing that emains is to bound the maximum size of the efeence set, R. Assume that when R is encounteed, the maximum efeence scale is s max and the quey node is N q. Evey node N R satisfies the popety enfoced in line 11 that d min (N q, N ) B(N q ). Using the definition of d min (, ) and B( ), we expand the equation. Note that p q is the point held in N q and p is the point held in N. Also, take ˆp to be the cuent neaest neighbo candidate fo p q ; that is, D[p q ] = d(p q, ˆp ) and N[p q ] = ˆp. Then, d min (N q, N ) B(N q ) (18) d(p q, p ) d(p q, ˆp ) + 2 sq s sq+1 (19) whee the last step follows because s q + 1 s max d(p q, ˆp ) + 2(2 smax +1 ) (20) and s s max. Define the set of points P as the points held in each node in R (that is, P = {p P(N ) : N R }). Then, we can wite P B S (p q, d(p q, ˆp ) + 2(2 smax +1 )). (21) Suppose that the tue neaest neighbo is p and d(p q, p ) > 2 smax +1. Then, p must be held as a descendant point of some node in R which holds some point p. Using the tiangle inequality, d(p q, ˆp ) d(p q, p ) d(p q, p ) + d( p, p ) d(p q, p ) + 2 smax +1. (22) This gives that P B S {pq}(p q, d(p q, p ) + 3(2 smax +1 )). The pevious step is necessay: to apply the definition of the expansion constant, the ball must be centeed at a point in the set; now, the cente (p q ) is pat of the set. 3286

19 Plug-and-Play Dual-Tee Algoithm Runtime Analysis B S {p q}(p q, d(p q, p ) + 3(2 smax +1 )) B S {p q}(p q, 4d(p q, p )) (23) c 3 q B S {p q}(p q, d(p q, p )/2) (24) which follows because the expansion constant of the set S {p q } is bounded above by c q. Next, we know that p is the closest point to p q in S {p q }; thus, thee cannot exist a point p p q S {p q } such that p B Sq (p q, d(p q, p )/2) because that would imply that d(p q, p ) < d(p q, p ), which is a contadiction. Thus, the only point in the ball is p q, and we have that B S {pq}(p q, d(p q, p )/2) = 1, giving the esult that R c 3 q in this case. The othe case is when d(p q, p ) 2 smax, and theefoe that P C s max +1, which means that d(p q, ˆp ) 2 smax +2. Note P B S (p q, d(p q, p ) + 3(2 smax +1 )) C s max (25) B S (p q, 4(2 smax +1 )) C s max. (26) Evey point in C s max is sepaated by at least 2 smax. Using Lemma 1 with δ = 2 smax ρ = 8 yields that P c 5. This gives the esult, because c 5 c 5 q. and In the monochomatic case whee S q = S 6, the bound is O(c 9 (N + I t (T )) because c = c = c q and θ = 0. Fo well-behaved tees whee I t (T q ) is linea o sublinea in N, this epesents the cuent tightest wost-case untime bound fo neaest neighbo seach. 6. Appoximate Kenel Density Estimation Ram et al. (2009a) pesent a cleve technique fo bounding the unning time of appoximate kenel density estimation based on the popeties of the kenel, when the kenel is shiftinvaiant and satisfies a few assumptions. We will estate these assumptions and povide an adapted poof using Theoem 1, which gives a tighte bound. Appoximate kenel density estimation is a common application of dual-tee algoithms (Gay and Mooe, 2003, 2001). Given a quey set S q, a efeence set S of size N, and a kenel function K(, ), the tue kenel density estimate fo a quey point p q is given as f (p q ) = p S K(p q, p ). (27) In the case of an infinite-tailed kenel K(, ), the exact computation cannot be acceleated; thus, attention has tuned towads tactable appoximation schemes. Two simple schemes fo the appoximation of f (p q ) ae well-known: absolute value appoximation and elative value appoximation. Absolute value appoximation equies that each density estimate f(p q ) is within ɛ of the tue estimate f (p q ): f(p q ) f (p q ) < ɛ p q S q. (28) 6. In the monochomatic case, we do not take a point as its own neaest neighbo, so slight modification of BaseCase() is necessay. The untime bound esult emains unchanged. 3287

20 Cutin, Lee, Mach, and Ram Relative value appoximation is a moe flexible appoximation scheme; given some paamete ɛ, the equiement is that each density estimate is within a elative toleance of f (p q ) : f(p q ) f (p q ) f (p q ) < ɛ p q S q. (29) Kenel density estimation is elated to the well-studied poblem of kenel summation, which can also be solved with dual-tee algoithms (Lee and Gay, 2006, 2009). In both of those poblems, egadless of the appoximation scheme, simple geometic obsevations can be made to acceleate computation: when K(, ) is shift-invaiant, faaway points have vey small kenel evaluations. Thus, tees can be built on S q and S, and node combinations can be puned when the nodes ae fa apat while still obeying the eo bounds. In the following two subsections, we will sepaately conside both the absolute value appoximation scheme and the elative value appoximation scheme, unde the assumption of a shift-invaiant kenel K(p q, p ) = K( p q p ) which is monotonically deceasing and non-negative. In addition, we assume that thee exists some bandwidth h such that K(d) must be concave fo d [0, h] and convex fo d [h, ). This assumption implies that the magnitude of the deivative K (d) is maximized at d = h. These ae not estictive assumptions; most standad kenels fall into this class, including the Gaussian, exponential, and Epanechnikov kenels. 6.1 Absolute Value Appoximation A tee-independent algoithm fo solving appoximate kenel density estimation with absolute value appoximation unde the pevious assumptions on the kenel is given as a BaseCase() function in Algoithm 4 and a Scoe() function in Algoithm 5 (a coectness poof can be found in Cutin et al., 2013b). The list f p holds patial kenel density estimates fo each quey point, and the list f n holds patial kenel density estimates fo each quey node. At the beginning of the dual-tee tavesal, the lists f p and f n, which ae both of size O(N), ae each initialized to 0. As the tavesal poceeds, node combinations ae puned if the diffeence between the maximum kenel value K(d min (N q, N )) and the minimum kenel value K(d max (N q, N )) is sufficiently small (line 3). If the node combination can be puned, then the patial node estimate is updated (line 4). When node combinations cannot be puned, BaseCase() may be called, which simply updates the patial point estimate with the exact kenel evaluation (line 3). Afte the dual-tee tavesal, the actual kenel density estimates f must be extacted. This can be done by tavesing the quey tee and calculating f(p q ) = f p (p q )+ N i T f n(n i ), whee T is the set of nodes in T q that have p q as a descendant. Each quey node needs to be visited only once to pefom this calculation; it may theefoe be accomplished in O(N) time. Note that this vesion is fa simple than othe dual-tee algoithms that have been poposed fo appoximate kenel density estimation (see, fo instance, Gay and Mooe, 2003); howeve, this vesion is sufficient fo ou untime analysis. Real-wold implementations, such as the one found in mlpack (Cutin et al., 2013a), tend to be fa moe complex. 3288

A Bijective Approach to the Permutational Power of a Priority Queue

A Bijective Approach to the Permutational Power of a Priority Queue A Bijective Appoach to the Pemutational Powe of a Pioity Queue Ia M. Gessel Kuang-Yeh Wang Depatment of Mathematics Bandeis Univesity Waltham, MA 02254-9110 Abstact A pioity queue tansfoms an input pemutation

More information

Math 301: The Erdős-Stone-Simonovitz Theorem and Extremal Numbers for Bipartite Graphs

Math 301: The Erdős-Stone-Simonovitz Theorem and Extremal Numbers for Bipartite Graphs Math 30: The Edős-Stone-Simonovitz Theoem and Extemal Numbes fo Bipatite Gaphs May Radcliffe The Edős-Stone-Simonovitz Theoem Recall, in class we poved Tuán s Gaph Theoem, namely Theoem Tuán s Theoem Let

More information

ON INDEPENDENT SETS IN PURELY ATOMIC PROBABILITY SPACES WITH GEOMETRIC DISTRIBUTION. 1. Introduction. 1 r r. r k for every set E A, E \ {0},

ON INDEPENDENT SETS IN PURELY ATOMIC PROBABILITY SPACES WITH GEOMETRIC DISTRIBUTION. 1. Introduction. 1 r r. r k for every set E A, E \ {0}, ON INDEPENDENT SETS IN PURELY ATOMIC PROBABILITY SPACES WITH GEOMETRIC DISTRIBUTION E. J. IONASCU and A. A. STANCU Abstact. We ae inteested in constucting concete independent events in puely atomic pobability

More information

Solution to HW 3, Ma 1a Fall 2016

Solution to HW 3, Ma 1a Fall 2016 Solution to HW 3, Ma a Fall 206 Section 2. Execise 2: Let C be a subset of the eal numbes consisting of those eal numbes x having the popety that evey digit in the decimal expansion of x is, 3, 5, o 7.

More information

Surveillance Points in High Dimensional Spaces

Surveillance Points in High Dimensional Spaces Société de Calcul Mathématique SA Tools fo decision help since 995 Suveillance Points in High Dimensional Spaces by Benad Beauzamy Januay 06 Abstact Let us conside any compute softwae, elying upon a lage

More information

Encapsulation theory: radial encapsulation. Edmund Kirwan *

Encapsulation theory: radial encapsulation. Edmund Kirwan * Encapsulation theoy: adial encapsulation. Edmund Kiwan * www.edmundkiwan.com Abstact This pape intoduces the concept of adial encapsulation, wheeby dependencies ae constained to act fom subsets towads

More information

Lecture 8 - Gauss s Law

Lecture 8 - Gauss s Law Lectue 8 - Gauss s Law A Puzzle... Example Calculate the potential enegy, pe ion, fo an infinite 1D ionic cystal with sepaation a; that is, a ow of equally spaced chages of magnitude e and altenating sign.

More information

The Substring Search Problem

The Substring Search Problem The Substing Seach Poblem One algoithm which is used in a vaiety of applications is the family of substing seach algoithms. These algoithms allow a use to detemine if, given two chaacte stings, one is

More information

4/18/2005. Statistical Learning Theory

4/18/2005. Statistical Learning Theory Statistical Leaning Theoy Statistical Leaning Theoy A model of supevised leaning consists of: a Envionment - Supplying a vecto x with a fixed but unknown pdf F x (x b Teache. It povides a desied esponse

More information

Pearson s Chi-Square Test Modifications for Comparison of Unweighted and Weighted Histograms and Two Weighted Histograms

Pearson s Chi-Square Test Modifications for Comparison of Unweighted and Weighted Histograms and Two Weighted Histograms Peason s Chi-Squae Test Modifications fo Compaison of Unweighted and Weighted Histogams and Two Weighted Histogams Univesity of Akueyi, Bogi, v/noduslód, IS-6 Akueyi, Iceland E-mail: nikolai@unak.is Two

More information

Unobserved Correlation in Ascending Auctions: Example And Extensions

Unobserved Correlation in Ascending Auctions: Example And Extensions Unobseved Coelation in Ascending Auctions: Example And Extensions Daniel Quint Univesity of Wisconsin Novembe 2009 Intoduction In pivate-value ascending auctions, the winning bidde s willingness to pay

More information

Relating Branching Program Size and. Formula Size over the Full Binary Basis. FB Informatik, LS II, Univ. Dortmund, Dortmund, Germany

Relating Branching Program Size and. Formula Size over the Full Binary Basis. FB Informatik, LS II, Univ. Dortmund, Dortmund, Germany Relating Banching Pogam Size and omula Size ove the ull Binay Basis Matin Saueho y Ingo Wegene y Ralph Wechne z y B Infomatik, LS II, Univ. Dotmund, 44 Dotmund, Gemany z ankfut, Gemany sauehof/wegene@ls.cs.uni-dotmund.de

More information

COMP Parallel Computing SMM (3) OpenMP Case Study: The Barnes-Hut N-body Algorithm

COMP Parallel Computing SMM (3) OpenMP Case Study: The Barnes-Hut N-body Algorithm COMP 633 - Paallel Computing Lectue 8 Septembe 14, 2017 SMM (3) OpenMP Case Study: The Banes-Hut N-body Algoithm Topics Case study: the Banes-Hut algoithm Study an impotant algoithm in scientific computing»

More information

Stanford University CS259Q: Quantum Computing Handout 8 Luca Trevisan October 18, 2012

Stanford University CS259Q: Quantum Computing Handout 8 Luca Trevisan October 18, 2012 Stanfod Univesity CS59Q: Quantum Computing Handout 8 Luca Tevisan Octobe 8, 0 Lectue 8 In which we use the quantum Fouie tansfom to solve the peiod-finding poblem. The Peiod Finding Poblem Let f : {0,...,

More information

B. Spherical Wave Propagation

B. Spherical Wave Propagation 11/8/007 Spheical Wave Popagation notes 1/1 B. Spheical Wave Popagation Evey antenna launches a spheical wave, thus its powe density educes as a function of 1, whee is the distance fom the antenna. We

More information

ASTR415: Problem Set #6

ASTR415: Problem Set #6 ASTR45: Poblem Set #6 Cuan D. Muhlbege Univesity of Mayland (Dated: May 7, 27) Using existing implementations of the leapfog and Runge-Kutta methods fo solving coupled odinay diffeential equations, seveal

More information

On the ratio of maximum and minimum degree in maximal intersecting families

On the ratio of maximum and minimum degree in maximal intersecting families On the atio of maximum and minimum degee in maximal intesecting families Zoltán Lóánt Nagy Lale Özkahya Balázs Patkós Máté Vize Mach 6, 013 Abstact To study how balanced o unbalanced a maximal intesecting

More information

Classical Worm algorithms (WA)

Classical Worm algorithms (WA) Classical Wom algoithms (WA) WA was oiginally intoduced fo quantum statistical models by Pokof ev, Svistunov and Tupitsyn (997), and late genealized to classical models by Pokof ev and Svistunov (200).

More information

arxiv: v1 [math.nt] 12 May 2017

arxiv: v1 [math.nt] 12 May 2017 SEQUENCES OF CONSECUTIVE HAPPY NUMBERS IN NEGATIVE BASES HELEN G. GRUNDMAN AND PAMELA E. HARRIS axiv:1705.04648v1 [math.nt] 12 May 2017 ABSTRACT. Fo b 2 and e 2, let S e,b : Z Z 0 be the function taking

More information

Supplementary information Efficient Enumeration of Monocyclic Chemical Graphs with Given Path Frequencies

Supplementary information Efficient Enumeration of Monocyclic Chemical Graphs with Given Path Frequencies Supplementay infomation Efficient Enumeation of Monocyclic Chemical Gaphs with Given Path Fequencies Masaki Suzuki, Hioshi Nagamochi Gaduate School of Infomatics, Kyoto Univesity {m suzuki,nag}@amp.i.kyoto-u.ac.jp

More information

3.1 Random variables

3.1 Random variables 3 Chapte III Random Vaiables 3 Random vaiables A sample space S may be difficult to descibe if the elements of S ae not numbes discuss how we can use a ule by which an element s of S may be associated

More information

10/04/18. P [P(x)] 1 negl(n).

10/04/18. P [P(x)] 1 negl(n). Mastemath, Sping 208 Into to Lattice lgs & Cypto Lectue 0 0/04/8 Lectues: D. Dadush, L. Ducas Scibe: K. de Boe Intoduction In this lectue, we will teat two main pats. Duing the fist pat we continue the

More information

Probablistically Checkable Proofs

Probablistically Checkable Proofs Lectue 12 Pobablistically Checkable Poofs May 13, 2004 Lectue: Paul Beame Notes: Chis Re 12.1 Pobablisitically Checkable Poofs Oveview We know that IP = PSPACE. This means thee is an inteactive potocol

More information

A Multivariate Normal Law for Turing s Formulae

A Multivariate Normal Law for Turing s Formulae A Multivaiate Nomal Law fo Tuing s Fomulae Zhiyi Zhang Depatment of Mathematics and Statistics Univesity of Noth Caolina at Chalotte Chalotte, NC 28223 Abstact This pape establishes a sufficient condition

More information

New problems in universal algebraic geometry illustrated by boolean equations

New problems in universal algebraic geometry illustrated by boolean equations New poblems in univesal algebaic geomety illustated by boolean equations axiv:1611.00152v2 [math.ra] 25 Nov 2016 Atem N. Shevlyakov Novembe 28, 2016 Abstact We discuss new poblems in univesal algebaic

More information

Chapter 3: Theory of Modular Arithmetic 38

Chapter 3: Theory of Modular Arithmetic 38 Chapte 3: Theoy of Modula Aithmetic 38 Section D Chinese Remainde Theoem By the end of this section you will be able to pove the Chinese Remainde Theoem apply this theoem to solve simultaneous linea conguences

More information

Goodness-of-fit for composite hypotheses.

Goodness-of-fit for composite hypotheses. Section 11 Goodness-of-fit fo composite hypotheses. Example. Let us conside a Matlab example. Let us geneate 50 obsevations fom N(1, 2): X=nomnd(1,2,50,1); Then, unning a chi-squaed goodness-of-fit test

More information

On the ratio of maximum and minimum degree in maximal intersecting families

On the ratio of maximum and minimum degree in maximal intersecting families On the atio of maximum and minimum degee in maximal intesecting families Zoltán Lóánt Nagy Lale Özkahya Balázs Patkós Máté Vize Septembe 5, 011 Abstact To study how balanced o unbalanced a maximal intesecting

More information

Chapter 5 Linear Equations: Basic Theory and Practice

Chapter 5 Linear Equations: Basic Theory and Practice Chapte 5 inea Equations: Basic Theoy and actice In this chapte and the next, we ae inteested in the linea algebaic equation AX = b, (5-1) whee A is an m n matix, X is an n 1 vecto to be solved fo, and

More information

I. CONSTRUCTION OF THE GREEN S FUNCTION

I. CONSTRUCTION OF THE GREEN S FUNCTION I. CONSTRUCTION OF THE GREEN S FUNCTION The Helmohltz equation in 4 dimensions is 4 + k G 4 x, x = δ 4 x x. In this equation, G is the Geen s function and 4 efes to the dimensionality. In the vey end,

More information

Temporal-Difference Learning

Temporal-Difference Learning .997 Decision-Making in Lage-Scale Systems Mach 17 MIT, Sping 004 Handout #17 Lectue Note 13 1 Tempoal-Diffeence Leaning We now conside the poblem of computing an appopiate paamete, so that, given an appoximation

More information

Geometry of the homogeneous and isotropic spaces

Geometry of the homogeneous and isotropic spaces Geomety of the homogeneous and isotopic spaces H. Sonoda Septembe 2000; last evised Octobe 2009 Abstact We summaize the aspects of the geomety of the homogeneous and isotopic spaces which ae most elevant

More information

Method for Approximating Irrational Numbers

Method for Approximating Irrational Numbers Method fo Appoximating Iational Numbes Eic Reichwein Depatment of Physics Univesity of Califonia, Santa Cuz June 6, 0 Abstact I will put foth an algoithm fo poducing inceasingly accuate ational appoximations

More information

Do Managers Do Good With Other People s Money? Online Appendix

Do Managers Do Good With Other People s Money? Online Appendix Do Manages Do Good With Othe People s Money? Online Appendix Ing-Haw Cheng Haison Hong Kelly Shue Abstact This is the Online Appendix fo Cheng, Hong and Shue 2013) containing details of the model. Datmouth

More information

Likelihood vs. Information in Aligning Biopolymer Sequences. UCSD Technical Report CS Timothy L. Bailey

Likelihood vs. Information in Aligning Biopolymer Sequences. UCSD Technical Report CS Timothy L. Bailey Likelihood vs. Infomation in Aligning Biopolyme Sequences UCSD Technical Repot CS93-318 Timothy L. Bailey Depatment of Compute Science and Engineeing Univesity of Califonia, San Diego 1 Febuay, 1993 ABSTRACT:

More information

Encapsulation theory: the transformation equations of absolute information hiding.

Encapsulation theory: the transformation equations of absolute information hiding. 1 Encapsulation theoy: the tansfomation equations of absolute infomation hiding. Edmund Kiwan * www.edmundkiwan.com Abstact This pape descibes how the potential coupling of a set vaies as the set is tansfomed,

More information

An Exact Solution of Navier Stokes Equation

An Exact Solution of Navier Stokes Equation An Exact Solution of Navie Stokes Equation A. Salih Depatment of Aeospace Engineeing Indian Institute of Space Science and Technology, Thiuvananthapuam, Keala, India. July 20 The pincipal difficulty in

More information

PROBLEM SET #1 SOLUTIONS by Robert A. DiStasio Jr.

PROBLEM SET #1 SOLUTIONS by Robert A. DiStasio Jr. POBLM S # SOLUIONS by obet A. DiStasio J. Q. he Bon-Oppenheime appoximation is the standad way of appoximating the gound state of a molecula system. Wite down the conditions that detemine the tonic and

More information

Brief summary of functional analysis APPM 5440 Fall 2014 Applied Analysis

Brief summary of functional analysis APPM 5440 Fall 2014 Applied Analysis Bief summay of functional analysis APPM 5440 Fall 014 Applied Analysis Stephen Becke, stephen.becke@coloado.edu Standad theoems. When necessay, I used Royden s and Keyzsig s books as a efeence. Vesion

More information

arxiv: v1 [math.co] 4 May 2017

arxiv: v1 [math.co] 4 May 2017 On The Numbe Of Unlabeled Bipatite Gaphs Abdullah Atmaca and A Yavuz Ouç axiv:7050800v [mathco] 4 May 207 Abstact This pape solves a poblem that was stated by M A Haison in 973 [] This poblem, that has

More information

6 PROBABILITY GENERATING FUNCTIONS

6 PROBABILITY GENERATING FUNCTIONS 6 PROBABILITY GENERATING FUNCTIONS Cetain deivations pesented in this couse have been somewhat heavy on algeba. Fo example, detemining the expectation of the Binomial distibution (page 5.1 tuned out to

More information

EM Boundary Value Problems

EM Boundary Value Problems EM Bounday Value Poblems 10/ 9 11/ By Ilekta chistidi & Lee, Seung-Hyun A. Geneal Desciption : Maxwell Equations & Loentz Foce We want to find the equations of motion of chaged paticles. The way to do

More information

Math 124B February 02, 2012

Math 124B February 02, 2012 Math 24B Febuay 02, 202 Vikto Gigoyan 8 Laplace s equation: popeties We have aleady encounteed Laplace s equation in the context of stationay heat conduction and wave phenomena. Recall that in two spatial

More information

arxiv: v1 [physics.pop-ph] 3 Jun 2013

arxiv: v1 [physics.pop-ph] 3 Jun 2013 A note on the electostatic enegy of two point chages axiv:1306.0401v1 [physics.pop-ph] 3 Jun 013 A C Tot Instituto de Física Univesidade Fedeal do io de Janeio Caixa Postal 68.58; CEP 1941-97 io de Janeio,

More information

Absorption Rate into a Small Sphere for a Diffusing Particle Confined in a Large Sphere

Absorption Rate into a Small Sphere for a Diffusing Particle Confined in a Large Sphere Applied Mathematics, 06, 7, 709-70 Published Online Apil 06 in SciRes. http://www.scip.og/jounal/am http://dx.doi.og/0.46/am.06.77065 Absoption Rate into a Small Sphee fo a Diffusing Paticle Confined in

More information

working pages for Paul Richards class notes; do not copy or circulate without permission from PGR 2004/11/3 10:50

working pages for Paul Richards class notes; do not copy or circulate without permission from PGR 2004/11/3 10:50 woking pages fo Paul Richads class notes; do not copy o ciculate without pemission fom PGR 2004/11/3 10:50 CHAPTER7 Solid angle, 3D integals, Gauss s Theoem, and a Delta Function We define the solid angle,

More information

Fractional Zero Forcing via Three-color Forcing Games

Fractional Zero Forcing via Three-color Forcing Games Factional Zeo Focing via Thee-colo Focing Games Leslie Hogben Kevin F. Palmowski David E. Robeson Michael Young May 13, 2015 Abstact An -fold analogue of the positive semidefinite zeo focing pocess that

More information

Exploration of the three-person duel

Exploration of the three-person duel Exploation of the thee-peson duel Andy Paish 15 August 2006 1 The duel Pictue a duel: two shootes facing one anothe, taking tuns fiing at one anothe, each with a fixed pobability of hitting his opponent.

More information

Question 1: The dipole

Question 1: The dipole Septembe, 08 Conell Univesity, Depatment of Physics PHYS 337, Advance E&M, HW #, due: 9/5/08, :5 AM Question : The dipole Conside a system as discussed in class and shown in Fig.. in Heald & Maion.. Wite

More information

Contact impedance of grounded and capacitive electrodes

Contact impedance of grounded and capacitive electrodes Abstact Contact impedance of gounded and capacitive electodes Andeas Hödt Institut fü Geophysik und extateestische Physik, TU Baunschweig The contact impedance of electodes detemines how much cuent can

More information

COLLAPSING WALLS THEOREM

COLLAPSING WALLS THEOREM COLLAPSING WALLS THEOREM IGOR PAK AND ROM PINCHASI Abstact. Let P R 3 be a pyamid with the base a convex polygon Q. We show that when othe faces ae collapsed (otated aound the edges onto the plane spanned

More information

THE JEU DE TAQUIN ON THE SHIFTED RIM HOOK TABLEAUX. Jaejin Lee

THE JEU DE TAQUIN ON THE SHIFTED RIM HOOK TABLEAUX. Jaejin Lee Koean J. Math. 23 (2015), No. 3, pp. 427 438 http://dx.doi.og/10.11568/kjm.2015.23.3.427 THE JEU DE TAQUIN ON THE SHIFTED RIM HOOK TABLEAUX Jaejin Lee Abstact. The Schensted algoithm fist descibed by Robinson

More information

Compactly Supported Radial Basis Functions

Compactly Supported Radial Basis Functions Chapte 4 Compactly Suppoted Radial Basis Functions As we saw ealie, compactly suppoted functions Φ that ae tuly stictly conditionally positive definite of ode m > do not exist The compact suppot automatically

More information

Stress Intensity Factor

Stress Intensity Factor S 47 Factue Mechanics http://imechanicaog/node/7448 Zhigang Suo Stess Intensity Facto We have modeled a body by using the linea elastic theoy We have modeled a cack in the body by a flat plane, and the

More information

Fresnel Diffraction. monchromatic light source

Fresnel Diffraction. monchromatic light source Fesnel Diffaction Equipment Helium-Neon lase (632.8 nm) on 2 axis tanslation stage, Concave lens (focal length 3.80 cm) mounted on slide holde, iis mounted on slide holde, m optical bench, micoscope slide

More information

ANA BERRIZBEITIA, LUIS A. MEDINA, ALEXANDER C. MOLL, VICTOR H. MOLL, AND LAINE NOBLE

ANA BERRIZBEITIA, LUIS A. MEDINA, ALEXANDER C. MOLL, VICTOR H. MOLL, AND LAINE NOBLE THE p-adic VALUATION OF STIRLING NUMBERS ANA BERRIZBEITIA, LUIS A. MEDINA, ALEXANDER C. MOLL, VICTOR H. MOLL, AND LAINE NOBLE Abstact. Let p > 2 be a pime. The p-adic valuation of Stiling numbes of the

More information

Many Electron Atoms. Electrons can be put into approximate orbitals and the properties of the many electron systems can be catalogued

Many Electron Atoms. Electrons can be put into approximate orbitals and the properties of the many electron systems can be catalogued Many Electon Atoms The many body poblem cannot be solved analytically. We content ouselves with developing appoximate methods that can yield quite accuate esults (but usually equie a compute). The electons

More information

Math 2263 Solutions for Spring 2003 Final Exam

Math 2263 Solutions for Spring 2003 Final Exam Math 6 Solutions fo Sping Final Exam ) A staightfowad appoach to finding the tangent plane to a suface at a point ( x, y, z ) would be to expess the cuve as an explicit function z = f ( x, y ), calculate

More information

On the Structure of Linear Programs with Overlapping Cardinality Constraints

On the Structure of Linear Programs with Overlapping Cardinality Constraints On the Stuctue of Linea Pogams with Ovelapping Cadinality Constaints Tobias Fische and Mac E. Pfetsch Depatment of Mathematics, TU Damstadt, Gemany tfische,pfetsch}@mathematik.tu-damstadt.de Januay 25,

More information

To Feel a Force Chapter 7 Static equilibrium - torque and friction

To Feel a Force Chapter 7 Static equilibrium - torque and friction To eel a oce Chapte 7 Chapte 7: Static fiction, toque and static equilibium A. Review of foce vectos Between the eath and a small mass, gavitational foces of equal magnitude and opposite diection act on

More information

High Dimensional Clustering with r-nets

High Dimensional Clustering with r-nets High Dimensional Clusteing with -nets Geogia Avaikioti, Alain Ryse, Yuyi Wang, Roge Wattenhofe ETH Zuich, Switzeland {zetava,ayse,yuwang,wattenhofe}@ethz.ch Abstact Clusteing, a fundamental task in data

More information

Deterministic vs Non-deterministic Graph Property Testing

Deterministic vs Non-deterministic Graph Property Testing Deteministic vs Non-deteministic Gaph Popety Testing Lio Gishboline Asaf Shapia Abstact A gaph popety P is said to be testable if one can check whethe a gaph is close o fa fom satisfying P using few andom

More information

Central Coverage Bayes Prediction Intervals for the Generalized Pareto Distribution

Central Coverage Bayes Prediction Intervals for the Generalized Pareto Distribution Statistics Reseach Lettes Vol. Iss., Novembe Cental Coveage Bayes Pediction Intevals fo the Genealized Paeto Distibution Gyan Pakash Depatment of Community Medicine S. N. Medical College, Aga, U. P., India

More information

Convergence Dynamics of Resource-Homogeneous Congestion Games: Technical Report

Convergence Dynamics of Resource-Homogeneous Congestion Games: Technical Report 1 Convegence Dynamics of Resouce-Homogeneous Congestion Games: Technical Repot Richad Southwell and Jianwei Huang Abstact Many esouce shaing scenaios can be modeled using congestion games A nice popety

More information

OSCILLATIONS AND GRAVITATION

OSCILLATIONS AND GRAVITATION 1. SIMPLE HARMONIC MOTION Simple hamonic motion is any motion that is equivalent to a single component of unifom cicula motion. In this situation the velocity is always geatest in the middle of the motion,

More information

Duality between Statical and Kinematical Engineering Systems

Duality between Statical and Kinematical Engineering Systems Pape 00, Civil-Comp Ltd., Stiling, Scotland Poceedings of the Sixth Intenational Confeence on Computational Stuctues Technology, B.H.V. Topping and Z. Bittna (Editos), Civil-Comp Pess, Stiling, Scotland.

More information

Secret Exponent Attacks on RSA-type Schemes with Moduli N = p r q

Secret Exponent Attacks on RSA-type Schemes with Moduli N = p r q Secet Exponent Attacks on RSA-type Schemes with Moduli N = p q Alexande May Faculty of Compute Science, Electical Engineeing and Mathematics Univesity of Padebon 33102 Padebon, Gemany alexx@uni-padebon.de

More information

MATH 415, WEEK 3: Parameter-Dependence and Bifurcations

MATH 415, WEEK 3: Parameter-Dependence and Bifurcations MATH 415, WEEK 3: Paamete-Dependence and Bifucations 1 A Note on Paamete Dependence We should pause to make a bief note about the ole played in the study of dynamical systems by the system s paametes.

More information

KOEBE DOMAINS FOR THE CLASSES OF FUNCTIONS WITH RANGES INCLUDED IN GIVEN SETS

KOEBE DOMAINS FOR THE CLASSES OF FUNCTIONS WITH RANGES INCLUDED IN GIVEN SETS Jounal of Applied Analysis Vol. 14, No. 1 2008), pp. 43 52 KOEBE DOMAINS FOR THE CLASSES OF FUNCTIONS WITH RANGES INCLUDED IN GIVEN SETS L. KOCZAN and P. ZAPRAWA Received Mach 12, 2007 and, in evised fom,

More information

AST 121S: The origin and evolution of the Universe. Introduction to Mathematical Handout 1

AST 121S: The origin and evolution of the Universe. Introduction to Mathematical Handout 1 Please ead this fist... AST S: The oigin and evolution of the Univese Intoduction to Mathematical Handout This is an unusually long hand-out and one which uses in places mathematics that you may not be

More information

YINGHUA DONG AND GENNADY SAMORODNITSKY

YINGHUA DONG AND GENNADY SAMORODNITSKY CONTACT DISTRIBUTION IN A THINNED BOOLEAN MODEL WITH POWER LAW RADII YINGHUA DONG AND GENNADY SAMORODNITSKY Abstact. We conside a weighted stationay spheical Boolean model in R d. Assuming that the adii

More information

The Millikan Experiment: Determining the Elementary Charge

The Millikan Experiment: Determining the Elementary Charge LAB EXERCISE 7.5.1 7.5 The Elementay Chage (p. 374) Can you think of a method that could be used to suggest that an elementay chage exists? Figue 1 Robet Millikan (1868 1953) m + q V b The Millikan Expeiment:

More information

Physics 2A Chapter 10 - Moment of Inertia Fall 2018

Physics 2A Chapter 10 - Moment of Inertia Fall 2018 Physics Chapte 0 - oment of netia Fall 08 The moment of inetia of a otating object is a measue of its otational inetia in the same way that the mass of an object is a measue of its inetia fo linea motion.

More information

Circular Orbits. and g =

Circular Orbits. and g = using analyse planetay and satellite motion modelled as unifom cicula motion in a univesal gavitation field, a = v = 4π and g = T GM1 GM and F = 1M SATELLITES IN OBIT A satellite is any object that is

More information

DOING PHYSICS WITH MATLAB COMPUTATIONAL OPTICS

DOING PHYSICS WITH MATLAB COMPUTATIONAL OPTICS DOING PHYIC WITH MTLB COMPUTTIONL OPTIC FOUNDTION OF CLR DIFFRCTION THEORY Ian Coope chool of Physics, Univesity of ydney ian.coope@sydney.edu.au DOWNLOD DIRECTORY FOR MTLB CRIPT View document: Numeical

More information

LINEAR AND NONLINEAR ANALYSES OF A WIND-TUNNEL BALANCE

LINEAR AND NONLINEAR ANALYSES OF A WIND-TUNNEL BALANCE LINEAR AND NONLINEAR ANALYSES O A WIND-TUNNEL INTRODUCTION BALANCE R. Kakehabadi and R. D. Rhew NASA LaRC, Hampton, VA The NASA Langley Reseach Cente (LaRC) has been designing stain-gauge balances fo utilization

More information

Chem 453/544 Fall /08/03. Exam #1 Solutions

Chem 453/544 Fall /08/03. Exam #1 Solutions Chem 453/544 Fall 3 /8/3 Exam # Solutions. ( points) Use the genealized compessibility diagam povided on the last page to estimate ove what ange of pessues A at oom tempeatue confoms to the ideal gas law

More information

On a quantity that is analogous to potential and a theorem that relates to it

On a quantity that is analogous to potential and a theorem that relates to it Su une quantité analogue au potential et su un théoème y elatif C R Acad Sci 7 (87) 34-39 On a quantity that is analogous to potential and a theoem that elates to it By R CLAUSIUS Tanslated by D H Delphenich

More information

On the integration of the equations of hydrodynamics

On the integration of the equations of hydrodynamics Uebe die Integation de hydodynamischen Gleichungen J f eine u angew Math 56 (859) -0 On the integation of the equations of hydodynamics (By A Clebsch at Calsuhe) Tanslated by D H Delphenich In a pevious

More information

Functions Defined on Fuzzy Real Numbers According to Zadeh s Extension

Functions Defined on Fuzzy Real Numbers According to Zadeh s Extension Intenational Mathematical Foum, 3, 2008, no. 16, 763-776 Functions Defined on Fuzzy Real Numbes Accoding to Zadeh s Extension Oma A. AbuAaqob, Nabil T. Shawagfeh and Oma A. AbuGhneim 1 Mathematics Depatment,

More information

Analytical Solutions for Confined Aquifers with non constant Pumping using Computer Algebra

Analytical Solutions for Confined Aquifers with non constant Pumping using Computer Algebra Poceedings of the 006 IASME/SEAS Int. Conf. on ate Resouces, Hydaulics & Hydology, Chalkida, Geece, May -3, 006 (pp7-) Analytical Solutions fo Confined Aquifes with non constant Pumping using Compute Algeba

More information

Information Retrieval Advanced IR models. Luca Bondi

Information Retrieval Advanced IR models. Luca Bondi Advanced IR models Luca Bondi Advanced IR models 2 (LSI) Pobabilistic Latent Semantic Analysis (plsa) Vecto Space Model 3 Stating point: Vecto Space Model Documents and queies epesented as vectos in the

More information

Light Time Delay and Apparent Position

Light Time Delay and Apparent Position Light Time Delay and ppaent Position nalytical Gaphics, Inc. www.agi.com info@agi.com 610.981.8000 800.220.4785 Contents Intoduction... 3 Computing Light Time Delay... 3 Tansmission fom to... 4 Reception

More information

Analysis of simple branching trees with TI-92

Analysis of simple branching trees with TI-92 Analysis of simple banching tees with TI-9 Dušan Pagon, Univesity of Maibo, Slovenia Abstact. In the complex plane we stat at the cente of the coodinate system with a vetical segment of the length one

More information

Syntactical content of nite approximations of partial algebras 1 Wiktor Bartol Inst. Matematyki, Uniw. Warszawski, Warszawa (Poland)

Syntactical content of nite approximations of partial algebras 1 Wiktor Bartol Inst. Matematyki, Uniw. Warszawski, Warszawa (Poland) Syntactical content of nite appoximations of patial algebas 1 Wikto Batol Inst. Matematyki, Uniw. Waszawski, 02-097 Waszawa (Poland) batol@mimuw.edu.pl Xavie Caicedo Dep. Matematicas, Univ. de los Andes,

More information

16 Modeling a Language by a Markov Process

16 Modeling a Language by a Markov Process K. Pommeening, Language Statistics 80 16 Modeling a Language by a Makov Pocess Fo deiving theoetical esults a common model of language is the intepetation of texts as esults of Makov pocesses. This model

More information

Multiple Criteria Secretary Problem: A New Approach

Multiple Criteria Secretary Problem: A New Approach J. Stat. Appl. Po. 3, o., 9-38 (04 9 Jounal of Statistics Applications & Pobability An Intenational Jounal http://dx.doi.og/0.785/jsap/0303 Multiple Citeia Secetay Poblem: A ew Appoach Alaka Padhye, and

More information

DIMENSIONALITY LOSS IN MIMO COMMUNICATION SYSTEMS

DIMENSIONALITY LOSS IN MIMO COMMUNICATION SYSTEMS DIMENSIONALITY LOSS IN MIMO COMMUNICATION SYSTEMS Segey Loya, Amma Koui School of Infomation Technology and Engineeing (SITE) Univesity of Ottawa, 6 Louis Pasteu, Ottawa, Ontaio, Canada, KN 6N5 Email:

More information

The geometric construction of Ewald sphere and Bragg condition:

The geometric construction of Ewald sphere and Bragg condition: The geometic constuction of Ewald sphee and Bagg condition: The constuction of Ewald sphee must be done such that the Bagg condition is satisfied. This can be done as follows: i) Daw a wave vecto k in

More information

AQI: Advanced Quantum Information Lecture 2 (Module 4): Order finding and factoring algorithms February 20, 2013

AQI: Advanced Quantum Information Lecture 2 (Module 4): Order finding and factoring algorithms February 20, 2013 AQI: Advanced Quantum Infomation Lectue 2 (Module 4): Ode finding and factoing algoithms Febuay 20, 203 Lectue: D. Mak Tame (email: m.tame@impeial.ac.uk) Intoduction In the last lectue we looked at the

More information

( ) [ ] [ ] [ ] δf φ = F φ+δφ F. xdx.

( ) [ ] [ ] [ ] δf φ = F φ+δφ F. xdx. 9. LAGRANGIAN OF THE ELECTROMAGNETIC FIELD In the pevious section the Lagangian and Hamiltonian of an ensemble of point paticles was developed. This appoach is based on a qt. This discete fomulation can

More information

Journal of Inequalities in Pure and Applied Mathematics

Journal of Inequalities in Pure and Applied Mathematics Jounal of Inequalities in Pue and Applied Mathematics COEFFICIENT INEQUALITY FOR A FUNCTION WHOSE DERIVATIVE HAS A POSITIVE REAL PART S. ABRAMOVICH, M. KLARIČIĆ BAKULA AND S. BANIĆ Depatment of Mathematics

More information

ON THE INVERSE SIGNED TOTAL DOMINATION NUMBER IN GRAPHS. D.A. Mojdeh and B. Samadi

ON THE INVERSE SIGNED TOTAL DOMINATION NUMBER IN GRAPHS. D.A. Mojdeh and B. Samadi Opuscula Math. 37, no. 3 (017), 447 456 http://dx.doi.og/10.7494/opmath.017.37.3.447 Opuscula Mathematica ON THE INVERSE SIGNED TOTAL DOMINATION NUMBER IN GRAPHS D.A. Mojdeh and B. Samadi Communicated

More information

7.2. Coulomb s Law. The Electric Force

7.2. Coulomb s Law. The Electric Force Coulomb s aw Recall that chaged objects attact some objects and epel othes at a distance, without making any contact with those objects Electic foce,, o the foce acting between two chaged objects, is somewhat

More information

MAGNETIC FIELD AROUND TWO SEPARATED MAGNETIZING COILS

MAGNETIC FIELD AROUND TWO SEPARATED MAGNETIZING COILS The 8 th Intenational Confeence of the Slovenian Society fo Non-Destuctive Testing»pplication of Contempoay Non-Destuctive Testing in Engineeing«Septembe 1-3, 5, Potoož, Slovenia, pp. 17-1 MGNETIC FIELD

More information

CSCE 478/878 Lecture 4: Experimental Design and Analysis. Stephen Scott. 3 Building a tree on the training set Introduction. Outline.

CSCE 478/878 Lecture 4: Experimental Design and Analysis. Stephen Scott. 3 Building a tree on the training set Introduction. Outline. In Homewok, you ae (supposedly) Choosing a data set 2 Extacting a test set of size > 3 3 Building a tee on the taining set 4 Testing on the test set 5 Repoting the accuacy (Adapted fom Ethem Alpaydin and

More information

Matrix Colorings of P 4 -sparse Graphs

Matrix Colorings of P 4 -sparse Graphs Diplomabeit Matix Coloings of P 4 -spase Gaphs Chistoph Hannnebaue Januay 23, 2010 Beteue: Pof. D. Winfied Hochstättle FenUnivesität in Hagen Fakultät fü Mathematik und Infomatik Contents Intoduction iii

More information

q i i=1 p i ln p i Another measure, which proves a useful benchmark in our analysis, is the chi squared divergence of p, q, which is defined by

q i i=1 p i ln p i Another measure, which proves a useful benchmark in our analysis, is the chi squared divergence of p, q, which is defined by CSISZÁR f DIVERGENCE, OSTROWSKI S INEQUALITY AND MUTUAL INFORMATION S. S. DRAGOMIR, V. GLUŠČEVIĆ, AND C. E. M. PEARCE Abstact. The Ostowski integal inequality fo an absolutely continuous function is used

More information

MASSACHUSETTS INSTITUTE OF TECHNOLOGY Physics Department. Problem Set 10 Solutions. r s

MASSACHUSETTS INSTITUTE OF TECHNOLOGY Physics Department. Problem Set 10 Solutions. r s MASSACHUSETTS INSTITUTE OF TECHNOLOGY Physics Depatment Physics 8.033 Decembe 5, 003 Poblem Set 10 Solutions Poblem 1 M s y x test paticle The figue above depicts the geomety of the poblem. The position

More information

Auchmuty High School Mathematics Department Advanced Higher Notes Teacher Version

Auchmuty High School Mathematics Department Advanced Higher Notes Teacher Version The Binomial Theoem Factoials Auchmuty High School Mathematics Depatment The calculations,, 6 etc. often appea in mathematics. They ae called factoials and have been given the notation n!. e.g. 6! 6!!!!!

More information