On a class of generalized Takagi functions with linear pathwise quadratic variation Alexander Schied∗ Department of Mathematics University of Mannheim 5 68131 Mannheim, Germany 1 0 2 First version: January 15, 2015 g This version: August 4, 2015 u A 3 Abstract 1 ] We consider a class X of continuous functions on [0,1] that is of interest from two R differentperspectives. First,itiscloselyrelatedtosetsoffunctionsthathavebeenstudied P as generalizations of the Takagi function. Second, each function in X admits a linear . h pathwise quadratic variation and can thus serve as an integrator in F¨ollmer’s pathwise It¯o t a calculus. We derive several uniform properties of the class X. For instance, we compute m the overall pointwise maximum, the uniform maximal oscillation, and the exact uniform [ modulus of continuity for all functions in X. Furthermore, we give an example of a pair 5 v x,y ∈ X for which the quadratic variation of the sum x+y does not exist. 7 3 Mathematics Subject Classification 2010: 26A30, 26A15, 60H05, 26A45 8 0 0 Key words: Generalized Takagi function, Takagi class, uniform modulus of continuity, pathwise . 1 quadratic variation, pathwise covariation, pathwise Ito¯ calculus, Fo¨llmer integral 0 5 1 : 1 Introduction v i X r In this note, we study a class X of continuous functions on [0,1] that is of interest from several a different perspectives. On the one hand, just as typical Brownian sample paths, each function x ∈ X admits the linear pathwise quadratic variation, (cid:104)x(cid:105) = t, in the sense of Fo¨llmer [13] t and therefore can serve as an integrator in F¨ollmer’s pathwise Ito¯ calculus. On the other hand, X is a subset, or has a nonempty intersection, with classes of functions that have been studied ∗E-mail: [email protected] TheauthorgratefullyacknowledgessupportbyDeutscheForschungsgemeinschaftthroughtheResearchTraining Group RTG 1953 1 as generalizations of Takagi’s celebrated example [29] of a nowhere differentiable continuous function. We will now explain the connections of our results with these two separate strands of literature. 1.1 Contributions to Fo¨llmer’s pathwise Ito¯ calculus In 1981, Fo¨llmer [13] proposed a pathwise version of It¯o’s formula, which, as a consequence, yields a strictly pathwise definition of the It¯o integral as a limit of Riemann sums. Some recent developments have led to a renewed interest in this pathwise approach. Among these is the conceptionoffunctional pathwiseIto¯calculusbyDupire [10] andCont andFourni´e [6,7], which for instance is crucial in defining partial differential equations on path space [11]. Another source for the renewed interest in pathwise It¯o calculus stems from the growing awareness of model ambiguity in mathematical finance and the resulting desire to reduce the reliance on probabilisticmodels; see,e.g.,[15]forarecentsurveyand[3,4,8,14,26,27]forcasestudieswith successfulapplicationsofpathwiseIt¯ocalculustofinancialproblems. Asystematicintroduction to pathwise It¯o calculus, including an English translation of [13], is provided in [28]. A function x ∈ C[0,1] can serve as an integrator in F¨ollmer’s pathwise It¯o calculus if it admits a continuous pathwise quadratic variation t (cid:55)→ (cid:104)x(cid:105) along a given refining sequence of t partitions of [0,1]. This condition is satisfied whenever x is a sample path of a continuous semimartingale, such as Brownian motion, and does not belong to a certain nullset. This nullset, however, is generally not known explicitly, and so it is not possible to tell whether a specific realization x of Brownian motion does indeed admit a continuous pathwise quadratic variation. The first purpose of this note is to provide a rich class X of continuous functions that can be constructed in a straightforward manner and that do admit the nontrivial pathwise quadratic variation (cid:104)x(cid:105) = t for all x ∈ X . The functions in X can thus be used as a class t of test integrators in pathwise Ito¯ calculus. Our corresponding result, Proposition 2.6, slightly extends a previous result by Gantert [18, 19], from which it follows that (cid:104)x(cid:105) = 1 for all x ∈ X . 1 Stillwithinthis context, asecondpurposeof thisnoteisto investigatewhetherthe existence of (cid:104)x(cid:105) and (cid:104)y(cid:105) implies the existence of (cid:104)x+y(cid:105) (or, equivalently, the existence of the pathwise quadratic covariation (cid:104)x,y(cid:105)). For typical sample paths of a continuous semimartingale, this implication is always true, but the corresponding nullset will depend on both x and y. In the literature on pathwise It¯o calculus, however, it has been taken for granted that the existence of (cid:104)x+y(cid:105) cannot be deduced from the existence of (cid:104)x(cid:105) and (cid:104)y(cid:105). In Proposition 2.7 we will now give an example of two functions x,y ∈ X for which (cid:104)x + y(cid:105) does indeed not exist. To the knowledge of the author, such an example has so far been missing from the literature. 1.2 Contributions to the theory of generalized Takagi functions In 1903, Takagi [29] proposed an example of a continuous function on [0,1] that is nowhere differentiable. This function has since been rediscovered several times and its properties have been studied extensively; see the recent surveys by Allaart and Kawamura [2] and Lagarias [24]. 2 While the original Takagi function itself does not belong to our class X , there are at least two classes of functions whose study was motivated by the Takagi function and that are intimately connected with X . One family of functions is the “Takagi class” introduced in 1984 by Hata and Yamaguti [20]. Similar but more restrictive function classes were introduced earlier by Faber [12] or Kahane [21]. The Takagi class has a nonempty intersection with X but neither one is included in the other. More recently, Allaart [1] extended the Takagi class to a more flexible class of functions. This family now contains X . By extending arguments given by Koˆno [23] for the Takagi class, Allaart [1] studies in particular the moduli of continuity of certain functions in his class. In contrast to these previous studies, the focus of this paper is not so much on the individual features of functions x ∈ X but rather on uniform properties of the entire class X . Here we compute the overall pointwise maximum, the uniform maximal oscillation, and the exact uniform modulus of continuity for all functions in X . In these computations, we cannot use previous methods that were conceived for the analysis of the Takagi functions and its generalizations. For instance, neither the result and arguments from Koˆno [23] nor the ones from Allaart [1] apply to the modulus of continuity of functions in X , and a suitable extension of the previous approaches must be developed. This new extension exploits the self-similar structure of X and its members. A special role in our analysis will be played by the function x, defined in (2.2) below. It (cid:98) has previously appeared in the work of Ledrappier [25], who studied the Hausdorff dimension of its graph, and in Gantert [18, 19]. Here we will determine its global maximum and its exact modulus of continuity. In particular the results on the global maximum of x will be needed in (cid:98) our analysis of the uniform properties of X , but these results are also interesting in their own right. This paper is organized as follows. In the subsequent Section 2 we first introduce our class X and then discuss its uniform properties in Theorems 2.2 and 2.3 and Corollary 2.5. We then recall Fo¨llmer’s [13] notions of pathwise quadratic variation and covariation and state our corresponding results. All proofs are given in Section 3. 2 Statement of results Recall that the Faber–Schauder functions are defined as e (t) := t, e (t) := (min{t,1−t})+, e (t) := 2−m/2e (2mt−k) ∅ 0,0 m,k 0,0 for t ∈ R, m = 1,2,..., and k ∈ Z. The graph of e looks like a wedge with height 2−m+2, m,k 2 width 2−m, and center at t = (k+1)2−m. In particular, the functions e have disjoint support 2 m,k for distinct k and fixed m. Now let coefficients θ ∈ {−1,+1} be given and define for n ∈ N m,k the continuous functions n−1 2m−1 (cid:88) (cid:88) xn(t) := θ e (t), 0 ≤ t ≤ 1. (2.1) m,k m,k m=0 k=0 3 It is well known (see, e.g., [1]) and easy to see that, due the uniform boundedness of the coefficients θ , the functions xn(t) converge uniformly in t to a continuous function x(t) as m,k n ↑ ∞. Let us denote by (cid:110) (cid:12) (cid:88)∞ 2(cid:88)m−1 (cid:111) X := x ∈ C[0,1](cid:12)x = θ e for coefficients θ ∈ {−1,+1} (cid:12) m,k m,k m,k m=0 k=0 the class of limiting functions arising in this way. A function x ∈ X belongs to the “Takagi class” introduced by Hata and Yamaguti [20] if and only if the coefficients θ in (2.1) are m,k independent of k. Moreover, X is a subset of the more flexible class of generalized Takagi functions studied by Allaart [1]. The original Takagi function, however, is obtained by taking θ = 2−m/2 and therefore does not belong to X . m,k Remark 2.1 (On similarities with Brownian sample paths). The functions in X can exhibit interesting fractal structures; see Figure 1. Figure 2, on the other hand, displays some similarities with the sample paths of a Brownian bridge. This similarity is not surprising since the well-known L´evy–Ciesielski construction of the Brownian bridge consists in replacing the coefficients θ ∈ {−1,+1} with independent standard normal random variables (see, m,k e.g., [22]). As a matter of fact, using arguments of de Rham [9] and Billingsley [5], it was shown in [1, Theorem 3.1 (iii)] that functions in X share with Brownian sample paths the property of being nowhere differentiable. Moreover, Ledrappier [25] showed that the Hausdorff dimension of the graph of the function ∞ 2m−1 (cid:88) (cid:88) x := e (2.2) (cid:98) m,k m=0 k=0 isthesameasthatofthegraphsoftypicalBrowniantrajectories,namely3/2. InProposition2.6 we will see, moreover, that the functions in X have the same pathwise quadratic variation as Brownian sample paths. Our first result is concerned with (uniform) maxima and oscillations of the functions in X . It will also be concerned with the function x defined in (2.2), a function that will play a special (cid:98) role throughout our analysis. The maximum of the original Takagi function was computed by Kahane [21], but his method does not apply in our case, and more complex arguments are needed here. Theorem 2.2 (Uniform maximum and oscillations). The class X has the following uni- form properties. (a) The uniform maximum of functions in X is attained by x and given by (cid:98) 1 √ max max |x(t)| = max x(t) = (2+ 2). (cid:98) x∈X t∈[0,1] t∈[0,1] 3 Moreover, the maximum of x(t) is attained at t = 1 and t = 2. (cid:98) 3 3 4 √ 1(2+ 2) 3 1 0.5 0.4 0.3 1/2 0.2 0.1 0.2 0.4 0.6 0.8 1.0 -0.1 1/3 1/2 2/3 1 1.0 1.0 0.8 0.6 0.5 0.4 0.2 0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0 -0.5 Figure 1: Plots of functions in X for various choices of θ . The upper left-hand panel shows m,k thefunctionx, whichisdefinedthroughθ = 1, togetherwithitsglobalmaximum. Theupper (cid:98) m,k right-hand panel corresponds to θ = (−1)m, the lower left-hand panel to θ = (−1)m+k, m,k m,k and the lower right-hand panel to θ = (−1)(cid:98)m/5(cid:99). m,k (b) The maximal uniform oscillation of functions in X is 1 √ max max |x(t)−x(s)| = (5+4 2), x∈X s,t∈[0,1] 6 where the respective maxima are attained at s = 1/3, t = 5/6, and ∞ (cid:18)2m−1−1 2m−1 (cid:19) (cid:88) (cid:88) (cid:88) x∗ := e + e − e ; (2.3) 0,0 m,k m,(cid:96) m=1 k=0 (cid:96)=2m−1 We refer to Figure 6 for a plot of the function x∗. In our next result, we will investigate the modulus of continuity of x and the uniform (cid:98) modulus of continuity of the class X . Kˆono [23] analyzed the moduli of continuity for some functions in the Takagi class of Hata and Yamaguti [20], and Allaart [1] later extended this result. However, neither the result and arguments from [23] nor the ones from [1] apply to the functions in X , because the sequence a := 2m/2 is not bounded. To state our results, let us m 5 1.0 0.5 0.8 0.6 0.2 0.4 0.6 0.8 1.0 0.4 0.2 -0.5 0.2 0.4 0.6 0.8 1.0 (cid:45)0.2 -1.0 Figure 2: Plots of x ∈ X when the coefficients θ form an independent and identically m,k distributed {−1,+1}-valued random sequence such that θ = +1 with probability 1 (left) m,k 2 and 1 (right). The dashed line corresponds to the approximation (cid:104)x(cid:105)8 of the quadratic variation 4 along the 8th dyadic partition T . 8 denote by ν(h) := (cid:98)−log h(cid:99), h > 0, (2.4) 2 the integer part of log 1, and define 2 h (cid:16) 1 (cid:17) 1 √ ω(h) := 1+ √ h2ν(h)/2 + ( 8+2)2−ν(h)/2. 2 3 √ Note that ω(h) is of the order O( h) as h ↓ 0. More precisely, (cid:114) ω(h) 4 √ ω(h) 1 √ liminf √ = 2 + 2, limsup √ = (11+7 2). h↓0 h 3 h↓0 h 6 These exact limits will however not be needed in the remainder of the paper. Theorem 2.3 (Moduli of continuity). (a) The function x has ω as its modulus of continuity. More precisely, (cid:98) |x(t+h)−x(t)| (cid:98) (cid:98) limsup max = 1. h↓0 0≤t≤1−h ω(h) √ (b) An exact uniform modulus of continuity for functions in X is given by 2ω. That is, |x(t+h)−x(t)| √ limsup sup max = 2. h↓0 x∈X 0≤t≤1−h ω(h) Moreover, the above supremum over functions x ∈ X is attained by the function x∗ defined in (2.3) in the sense that |x∗(t+h)−x∗(t)| √ limsup max = 2. h↓0 0≤t≤1−h ω(h) 6 Remark 2.4. In the proof of Theorem 2.3, we will actually show the following upper bounds that are stronger than the corresponding statements in the theorem: √ |x(t+h)−x(t)| ≤ ω(h) and sup |x(t+h)−x(t)| ≤ 2ω(h) (cid:98) (cid:98) x∈X for all h ∈ [0,1) and t ∈ [0,1−h]. Corollary 2.5. X is a compact subset of C[0,1] with respect to the topology of uniform con- vergence. Theorem 2.3 (b) implies moreover that each x ∈ X is Ho¨lder continuous with exponent 1 2 and hence admits a finite 2-variation in the sense that (cid:88) sup (x(t(cid:48))−x(t))2 < ∞, (2.5) T t∈T where the supremum is taken over all partitions T of [0,1] and t(cid:48) denotes the successor of t in T, i.e., (cid:40) min{u ∈ T|u > t} if t < 1, t(cid:48) = 1 if t = 1. Each x ∈ X can therefore serve as an integrator in the pathwise integration theory of rough paths; see, e.g., Friz and Hairer [17]. A different pathwise integration theory was proposed earlier by F¨ollmer [13]. It is based on the following notion of pathwise quadratic variation. Instead of considering the supremum over all partitions as in (2.5), one fixes an increasing sequence of partitions T ⊂ T ⊂ ··· of [0,1] such that the mesh of T tends to zero; such 1 2 n a sequence (T ) will be called a refining sequence of partitions. For x ∈ C[0,1] one then n n∈N defines the sequence (cid:88) (cid:104)x(cid:105)n := (x(s(cid:48))−x(s))2. (2.6) t s∈Tn,s≤t The function x ∈ C[0,1] is said to admit the continuous quadratic variation (cid:104)x(cid:105) along the sequence (T ) if for all t ∈ [0,1] the limit n (cid:104)x(cid:105) := lim(cid:104)x(cid:105)n (2.7) t t n↑∞ exists, and if t (cid:55)→ (cid:104)x(cid:105) is a continuous function. Fo¨llmer’s pathwise Ito¯ calculus uses this1 class t of functions x as integrators. For given x ∈ C[0,1], the approximations (cid:104)x(cid:105)n are typically not monotone in n, and so it is t not clear a priori whether the limit in (2.7) exists. Moreover, even if the limit exists, it may 1Pathwise It¯o calculus also works for c`adl`ag functions x, but this requires that the continuous part of x admits a continuous quadratic variation along (T ); see [13] and [6]. For this reason we will concentrate here n on the case of continuous functions x. 7 depend strongly on the particular choice of the underlying sequence of partitions. For instance, it is known that for any x ∈ C[0,1] there exists a refining sequence (T ) of partitions such n n∈N thatthequadraticvariationofxalong(T ) vanishesidentically; see[16, p. 47]. Itisalsonot n n∈N difficult to construct x ∈ C[0,1] for which the limit in (2.7) exists but satisfies (cid:104)x(cid:105) = 1 (t) t ]1/2,1] and is hence discontinuous. On the other hand, it is easy to see that the quadratic variation of a continuous function with bounded variation exists and vanishes along every refining sequence of partitions. Functions that do admit a nontrivial quadratic variation for some refining sequence of partitions must hence be of infinite total variation. Sothefirstquestionthatarisesinthiscontextishowonecanobtainfunctionsthatdoadmit a continuous quadratic variation along a given refining sequence of partitions, (T ) ? Of n n∈N course one can take all sample paths of a Brownian motion (or, more generally, of a continuous semimartingale) that are not contained in a certain null set A. But A is generally not given explicitly, and so it is not possible to tell whether a specific realization x of Brownian motion does indeed admit the quadratic variation (cid:104)x(cid:105) = t along (T ) . Moreover, this selection t n n∈N principle for functions x lets a probabilistic model enter through the backdoor, although the initial purpose of pathwise It¯o calculus was to get rid of probabilistic models altogether. In the following proposition, we show that each x ∈ X admits the linear pathwise quadratic variation (cid:104)x(cid:105) = t for t ∈ [0,1] along the sequence of dyadic partitions: t T := {k2−n|k = 0,...,2n}, n = 1,2,... (2.8) n This slightly extends a result by Gantert [18, 19], from which it follows that (cid:104)x(cid:105) = 1 for all 1 x ∈ X . Our proposition implies that each x ∈ X can serve as an integrator in Fo¨llmer’s pathwise Ito¯ calculus. Proposition 2.6. Every x ∈ X admits the quadratic variation (cid:104)x(cid:105) = t along the sequence t (T ) from (2.8). n The second question that we will address in this context is concerned with a standard assumption that is made in pathwise Ito¯ calculus whenever the covariation of two functions x,y ∈ C[0,1] is needed. Let (cid:88) (cid:104)x,y(cid:105)n := (x(s(cid:48))−x(s))(y(s(cid:48))−y(s)) (2.9) t s∈Tn,s≤t and observe that 1(cid:16) (cid:17) (cid:104)x,y(cid:105)n = (cid:104)x+y(cid:105)n −(cid:104)x(cid:105)n −(cid:104)y(cid:105)n . (2.10) t 2 t t t If x and y admit the continuous quadratic variations (cid:104)x(cid:105) and (cid:104)y(cid:105) along (T ) , then it follows n n∈N from (2.10) that the covariation of x and y, (cid:104)x,y(cid:105) := lim(cid:104)x,y(cid:105)n, (2.11) t t n↑∞ 8 exists along (T ) and is continuous in t if and only if x+y admits a continuous quadratic n n∈N variationalong(T ) . Whenxandy aresamplepathsofBrownianmotionor, moregenerally, n n∈N of a continuous semimartingale, the quadratic variation (cid:104)x+y(cid:105), and hence (cid:104)x,y(cid:105), will always exist almost surely. But for arbitrary functions x,y ∈ C[0,1] it has so far not been possible to reducetheexistenceofthelimitin(2.11)totheexistenceof(cid:104)x(cid:105)and(cid:104)y(cid:105). Inournextproposition we will provide an example of two functions x,y ∈ X for which the limit in (2.11) does not exist, even though (cid:104)x(cid:105) = t = (cid:104)y(cid:105) . This shows that the existence of (cid:104)x,y(cid:105) and (cid:104)x+y(cid:105) is not t t implied by the existence of (cid:104)x(cid:105) and (cid:104)y(cid:105). It follows in particular that the class of functions that admit a continuous quadratic variation along (T ) is not a vector space. To the knowledge n n∈N of the author, a corresponding example has so far been missing from the literature. Proposition 2.7. Consider the sequence (T ) of dyadic partitions (2.8) and the functions n n∈N ∞ 2m−1 ∞ 2m−1 (cid:88) (cid:88) (cid:88) (cid:88) x = e and y = (−1)me , (cid:98) m,k m,k m=0 k=0 m=0 k=0 which belong to X and hence admit the quadratic variation (cid:104)x(cid:105) = t = (cid:104)y(cid:105) along (T ). Then (cid:98) t t n 4 8 lim(cid:104)x+y(cid:105)2n = t and lim(cid:104)x+y(cid:105)2n+1 = t (2.12) n↑∞ (cid:98) t 3 n↑∞ (cid:98) t 3 and 1 1 lim(cid:104)x,y(cid:105)2n = − t and lim(cid:104)x,y(cid:105)2n+1 = t. (2.13) n↑∞ (cid:98) t 3 n↑∞ (cid:98) t 3 In particular, for t > 0, the limits of (cid:104)x + y(cid:105)n and (cid:104)x,y(cid:105)n do not exist as n ↑ ∞, but x + y (cid:98) t (cid:98) t (cid:98) admits different continuous quadratic variations along the two refining sequences (T ) and 2n n∈N (T ) . 2n+1 n∈N See the two upper panels in Figure 1 for plots of the two functions x and y occurring in (cid:98) Proposition 2.7. See Figure 3 for a plot of ∞ 22m−1 (cid:88) (cid:88) x+y = 2e . (2.14) (cid:98) 2m,k m=0 k=0 3 Proofs 3.1 Proof of Theorem 2.2 We start with the following lemma, which computes maxima and maximizers of the functions n−1 2m−1 (cid:88) (cid:88) xn(t) := e (t), t ∈ [0,1] and n = 1,2,...; (cid:98) m,k m=0 k=0 see Figure 4 for an illustration. 9 2.5 2.0 1.5 1.0 0.5 0.2 0.4 0.6 0.8 1.0 Figure 3: Plot of the function x+y defined in (2.14). The dotted line is (cid:104)x+y(cid:105)7, the dashed (cid:98) (cid:98) line is (cid:104)x+y(cid:105)8. (cid:98) M5 M4 M3 n=5 M2 n=4 M1 n=3 n=2 n=1 t2 t4 t5 t3 t1= 12 Figure 4: Illustration of Lemma 3.1 and its proof. The functions xn are plotted over the interval (cid:98) [0,1/2] for various values of n, together with the corresponding values for t and M . n n 10