Datasets:
ccdv
/
arxiv-classification

Dataset card Data Studio Files Community
4
text
stringlengths
2.85k
2.55M
label
class label
11 classes
SIR Coverage Analysis in Cellular Networks with Temporal Traffic: A Stochastic Geometry Approach arXiv:1801.09888v1 [cs.IT] 30 Jan 2018 Howard H. Yang, Member, IEEE, and Tony Q. S. Quek, Fellow, IEEE Abstract—The bloom in mobile applications not just bring in enjoyment to daily life, but also imposes more complicated traffic situation on wireless network. A complete understanding of the impact from traffic profile is thus essential for network operators to respond adequately to the surge in data traffic. In this paper, based on stochastic geometry and queuing theory, we develop a mathematical framework that captures the interplay between the spatial location of base stations (BSs), which determines the magnitude of mutual interference, and their temporal traffic dynamic. We derive a tractable expression for the SIR distribution, and verify its accuracy via simulations. Based on our analysis, we find that i) under the same configuration, when traffic condition changes from light to heavy, the corresponding SIR requirement can differ by more than 10 dB for the network to maintain coverage, ii) the SIR coverage probability varies largely with traffic fluctuation in the sub-medium load regime, whereas in scenario with very light traffic load, the SIR outage probability increases linearly with the packet arrival rate, iii) the mean delay, as well as coverage probability of cell edge user equipments (UEs) are vulnerable to the traffic fluctuation, thus confirms its appeal for traffic-aware communication technology. Index Terms—Poisson point process, cellular networks, random packet arrival, interacting queues, stochastic geometry, mean delay. I. I NTRODUCTION The rapid evolution of mobile applications imposes more complicated traffic condition on wireless networks, where not only the data demand grows exponentially [1], but more importantly, the the content is largely changing from mobile voice to multimedia [2]. To give an adequate response to the surge in mobile data traffic, network operators need a complete understanding on the impact of temporal traffic. In this article, we aim to evaluate how the traffic statistic affects the wireless networks, and to find those aspects that are most vulnerable. A. Background and Related Work Due to the broadcast nature of wireless channel, transmitters in space sharing a common spectrum will interact with each other through the interference they cause. To characterize the performance of such networks, stochastic geometry has been recently introduced as a way to assess performance of wireless links in large-scale networks [3]–[8]. The intrinsic elegance in modeling and analysis has popularized its application in evaluating performance among various wireless systems, including ad-hoc networks [3], cellular networks [9], or more advanced heterogeneous networks [10], even with H. H. Yang and T. Q. S. Quek are with the Information Systems Technology and Design Pillar, Singapore University of Technology and Design, Singapore (e-mail: howard [email protected], [email protected]). device-to-device (D2D) communication [11] and multipleinput multiple-output (MIMO) technology [12], [13]. However, the main drawback of these models is that they heavily rely on the full buffer assumption, i.e., every link always has a packet to transmit, and do not allow one to represent random traffic. While the additional dimension of randomness in temporal domain increases the complexity in analysis, it is nevertheless a crucial factor in understanding network performance, especially for the next generation wireless system that faces more voliated traffic conditions [14], [15]. The main difficulty with queuing in wireless network comes from the interdependency among the evolution of different queues, which is usually referred to as interacting queues [16]. Because of interference, the queue status of one transmitter can affect, and also be affected by, the queue status of its neighbors, hence making the analysis very difficult. Conventionally, the queuing interaction through wireless medium is studied using simple collision models [16]–[19]. In such models, discrete time ALOHA protocol is usually employed, where each of the N terminals initiates a transmission attempt at every slot: If more than two terminals transmit simultaneously, a collision occurs and all the terminals retransmit their packets in next slot with the same probability [17]. Analytical results about system stability can be obtained via exact form in scenarios with few (two or three) transmitters [16], or through approximations in asymptotic regime with infinitely many transmitters [20]. However, these models over simplify the wireless channel and lack the ability to tract the interference, which differs according to distance as well as channel gains, thus do not capture the information-theoretic interactions precisely. Recent attempts to address this issue are made in [21]–[25], where queuing theory is combined with stochastic geometry to model the dynamic from both temporal and spatial domains. The results provide the necessary and sufficient conditions for network to be stable [24], and different performance metrics, including success transmission probability [22], delay [21], and packet throughput [23] have been subsequently derived. While giving more refined analysis, these results either provide only bounds that are not necessary tight [21], [24], or are only appliable to networks with light traffic [23]. The most related work is from [22], where the authors applied Geo/PH/1 queuing model to account for the interference-based queues among UEs and analyzed the SIR performance under three different transmission schemes. However, the requirement for full channel inversion limits its generalization, and the restricted stable region constrains its application to relatively low traffic condition with small SIR detection threshold, and thus prevents one to take a complete treatment on traffic statistic. To 2 this end, a mathematical framework that captures the spatial– temporal dynamic of the network, and adapts to scenarios with different traffic conditions is of necessity to be explored. TABLE I N OTATION S UMMARY Notation B. Approach and Summary of Contributions In this paper, we model the BS deployment and UE locations as independent Poisson point processes (PPPs), where each BS maintains an infinite capacity buffer to store the incoming packets. The queuing dynamic is modelled via a descrete time system, where we consider the arrival of packets at each BS to be independent Bernoulli process. By combining stochastic geometry with queuing theory, we obtain a tractable expression for the SIR coverage probability. With the developed framework, we can explicitly characterize the SIR variation due to change of traffic condition, and its consequential impact on system stability, as well as delay distribution. Our main contributions are summarized below. • We develope a mathematical framework that captures the interplay between the spatial geometry of wireless links and their temporal traffic dynamic. Our analysis is tractable, and takes into account all the key features of a cellular network, including traffic profile, smallscale fading and path loss, random network topology, and queuing interaction. • Unlike [22], our result not only provides the standard SIR coverage probability, but also gives a more precise description about the fraction of UEs achieving SIR at different levels. For instance, the SIR coverage probability of cell-edge UEs can be easily derived via our result. • We discuss the sufficient and necessary conditions for the network to be stable, and provide an approximation for the stable region. We also derive the mean delay distribution, by accounting for both queuing and transmission delay. • Using the developed analysis, we find that under the same network configuration, there is more than 10 dB SIR difference between light and heavy traffic conditions. Moreover, in the very light traffic regime, the network SIR outage probability is shown to increase linearly with packet arrival rate. The result also reveals that the mean delay, as well as cell-edge UE rate, are vulnerable to the variation of traffic condition, hence urging advanced solution to adapt with traffic profile. The remainder of the paper is organized as follows. We introduce the system model in Section II. In Section III, we detail the analysis of SIR distribution in cellular networks with temporal traffic. We show the simulation and numerical results in Section IV, that confirm the accuracy of our analysis, and provide insights about the impact of traffic profile on network performance. We conclude the paper in Section V. II. S YSTEM M ODEL In this section, we provide a general introduction to the network topology, the traffic profile, as well as the propagation and failure retransmission model. The main notations used throughout the paper are summarized in Table I. Φb ; λb Φu ; λu Pmt ; α Sε ξ; ξc,ε γx0 ,t ; θ µΦ x0 ,t Definition PPP modeling the location of base stations (BSs); BS deployment density PPP modeling the location of UEs; UE deployment density BS transmit power; path loss exponent ε-stable region of wireless network, under which the fraction of unstable queues is less than ε Packet arrival rate; critical arrival rate for εstability Received SIR of typical UE at time slot t; SIR decoding threshold Conditional SIR coverage probability at time slot t A. Network Topology and Traffic Model We consider the downlink of a cellular network, as depicted in Fig. 1, that consists of randomly deployed BSs whose spatial locations follow independent Poisson point process (PPP) Φb with spatial densities λb .1 The location of UEs is modelled as another independent PPP Φu with spatial density λu , where each UE associates with its closest BS for transmission. We assume the UE density is high enough that every BS has at least one UE associates with it.2 In this network, all BSs and UEs are assumed to be equipped with single antenna, and each BS transmits with constant power Pmt . In light of its spectral efficiency, we employ universal frequency reuse throughout the network, i.e., every BS transmits in the same spectrum. We use a discrete time queuing system to model the random traffic profile. In particular, the time axis is segmented into a sequence of equal time intervals, referred to as time slots. We further assume all queuing activities, i.e., arrivals and departures, take place around the slot boundaries. Specifically, at the mth time slot, a potential packet departure may occur in the interval (m− , m), and a potential packet arrival can happen in the interval (m, m+ ). In other words, departures occur at the moment immediately before the slot boundaries while arrivals occur at the moment immediately after the slot boundaries. For a generic UE located in the cell of xi ∈ Φb , we model its packet arrival as a Bernoulli process with rate ξ ∈ [0, 1], which represents the probability of a new arrival occurs in a slot. We further assume that each BS accumulates all the incoming packets in an infinite-size buffer for further transmission purpose. In order to investigate the evolution of queuing dynamic, we limit the mobility of transceivers by considering a static network, i.e., the locations of the BSs and UEs are generated once at (0− , 0), and remained unchanged in all the following 1 PPPs serve as a good model for the planned deployment of macro cell BSs, as verified by both empirical evidence [8], [26] and theoretical analysis [27]. 2 The one UE per cell set up is mainly for the sake of notational simplicity. By adopting a similar approach as [23], the analysis in this paper can be easily extended to consider more realistic scenario where multiple UEs in each cell share the same wireless channel. 3 Transmission link Interference link BS 1 UE 1 BS 2 UE 2 BS 1 UE 1 BS 2 UE 2 Fig. 1. A snapshot of the Poisson cellular network with temperal traffic. The coverage area of different BSs form the Voronoi cells, whereas each BS may be active or inactive, depending on its buffer status. Fig. 2. Example of interacting queues with two BSs sharing the wireless channel. Due to impact from mutual interference, the service rates of BS 1 and BS 2 are different, and dynamically change with their queuing status. time slots.3 the origin, with its tagged BS located at x0 . Given the UE is receiving data at time slot t, the received SIR can be written as B. Propagation Channel and Failure Retransmission In this network, we adopt a block-fading propagation model, where the channels between any pair of antennas are assumed independent and identically distributed (i.i.d.) and quasi-static, i.e., the channel is constant during one transmission slot, and varies independently from slot to slot. We consider all propagation channels are narrowband and affected by two attenuation components, namely small-scale Rayleigh fading with unit mean power, and large-scale path loss that follows power law.4 Affected by the random channel fading and aggregated interference, the process of packet departure does not possess a constant rate and can lead to failure packet deliverty. Retransmission is thus necessary to guarantee the packet can be correctly received. By enabling retransmission, the transmission model at each BS becomes: During each time slot, every node with a non-empty buffer sends out a packet from the head of its queue. If the received SIR exceeds a predefined threshold, the transmission is successful and the packet can be removed from the queue; otherwise, the transmission fails and the packet remains in the buffer. We assume the feedback of each transmission, either success or fail, can be instantaneously awared by the BSs such that they are able to schedule transmission at next time slot. Moreover, for the BSs with empty buffer, they mute the transmissions to reduce power consumption and inter-cell interference. C. Signal-to-interference ratio (SIR) By applying Slivnyak’s theorem [5] to the stationary PPP of BS, it is sufficient to focus on the SIR of a typical UE at 3 Note that most of the pratical networks can be approximately regarded as static, since the locations of any end device cannot change drastically in a relatively short period [24]. 4 The analysis in this paper is not necessary constrained to simple propagation model, it can be further extended to incorporate more realistic setups that include multi-slope path loss [28], [29] and complicated fading environment [30], [31]. γx0 ,t = P h kx k−α P mt x0 0 , Pmt ζx,t hx kxk−α (1) x∈Φb \x0 where hx ∼ exp(1) denotes the small scale Rayleigh fading from BS x to the origin, k·k is the Euclidean distance, α stands for the path loss exponent, and ζx,t ∈ {0, 1} is an indicator showing whether a node located at x ∈ Φb is transmitting at time slot t (ζx,t = 1) or not (ζx,t = 0). It is important to note that since the spectrum is shared among BSs, the queuing status of each BS is coupled with other transmitters and hence results in interacting queue. As such, the BS active state, ζx,t , is both spatial and temperal dependent, since the location affects the pathloss and further the aggregated interference, and the time changes the queue length at each node. To better illustrate this concept, Fig. 2 gives a simple example of queuing interaction between two BSs. Note that compared to UE 1, UE 2 has an advantage location and hence enjoys better path loss and fewer interference. Consequently, BS 2 can quickly empty its queue, and the disparity between their communication conditions results in BS 2 activates less frequently than BS 1. Moreover, depending on whether packets appear at both BSs or not, the corresponding active durations are also different: If both transmitters have packets to send, the mutual interference will reduce the service rate and prolong the active duration of each BS individually. On the other hand, when one transmitter becomes silent, the other one can benefit from the reduced cross-talk and speed up its queue flushing process, hence also decreases the active period. Extending this concept to a large-scale network, we find that as the realization of PPP is irregular, there are always some BSs experience poor transmission environment, e.g., their UEs are located at the cell edge, and some others having good communication condition, e.g., their UEs are around the cell centers. In this regard, even the packet arrival rate is the same for all transmitters, the queuing status and active state 4 F Φ (u) = 1 1 − 2 π Z ∞ 0   1 Im u−jω  ω " #−1    Z 1 k ∞   X jω ξ Φ F (dt) Z(k, δ, θ) dω 1+δ F Φ (ξ) + k  k ξ t can varies largely from BS to BS, and the characterization of SIR in such network is very challenging. III. A NALYSIS We now present the main technical results of the paper. In particular, we first detail the definition of conditional SIR coverage probability, and the analysis on its distribution. Then we discuss the conditions for the queuing network to be stable. After that, we give moments as well as a computationally efficient approximation for the conditional SIR coverage probability. Finally, we derive the distribution of mean delay, which involves both queuing and transmission delay. A. Conditional SIR Coverage Probability Since both the received signal strength and the interference at a given UE are governed by a number of stochastic processes, e.g., random spatial distribution of transmitting/receiving nodes, random packet arrivals, and queuing dynamics, the SIR in (1) is a random variable and can only be characterized via distribution. In this regard, conditioning on the realization of the point process Φ , Φb ∪ Φu , we define the conditional SIR coverage probability as follows [24], [32]. Definition 1: Given the typical UE is receiving data at time slot t,, its conditional SIR coverage probability is defined as µΦ x0 ,t = P (γx0 ,t ≥ θ|Φ)   = Eexp−θkx0 k α X   hx   Φ . ζx,t kxkα x∈Φb \x0 (4) k=1 be negligible on the employed system model [19].5 Motivated by this fact, we make the following assumption. Assumption 1: The temporal interference correlation has a negligible effect on the transmission SIR coverage probability. Hence, we assume the typical UE sees almost independent interference at each time slot. It is now safe to assume that all UEs experience i.i.d. steady state queue distributions, and each BS activates independently, whereas the active probability depends on the specific service rate, or equivalently, the SIR coverage probability. To faciliate analysis in the following, we introduce a simple result from queuing theory as a preliminary, which describes the active probability of a transmitter under fixed arrival and departure rates. Lemma 1: Given the arrival rate being ξ, the service rate µ, the active probability at a generic BS is  1, if µ ≤ ξ, a η = (3) ξ/µ, if µ > ξ. Proof: The result is a standard conclusion from Geo/Geo/1 queue, which can be found in [36]. After all the above preparason, we are now ready to derive our main result of this paper, i.e., distribution of the conditional SIR coverage probability. Theorem 1: The cumulative distribution function of the conditional SIR coverage probability is given by the fixedpoint equation (4) on top of this page, which can be iterative solved as follows F Φ (u) = lim FnΦ (u) (2) Note that the conditional SIR coverage probability µΦ x0 ,t is still a random variable (as we condition on the realization of Φ), which contains all the information about the UE SIR (and therefore achievable rate) distribution across the network. Moreover, the interaction of queues is also captured by (2) via the accumulated interference. In order to analyze the distribution of µΦ x0 ,t , we need to address two issues: i) due to random packet arrival and retransmission of failed deliveries, the active state, i.e., ζx,t = 1, at each transmitter varies over time, and ii) there may exist common interferencing BSs seen by the same UE from one time slot to another, which introduces temporal correlation for the SIR coverage probability [33]–[35]. The dynamically changing active state of BSs, together with the temperal correlation, involve memory to the queues and highly complicates the analysis. Fortunately, when the number of transmitters asymptotically approaches infinity, a mean field property starts to emerge in the evolution of queues, i.e., the interaction between queues become “weak” and “global”, and the impact from aforementioned temperal and spatial correlation tends to n→∞ (5) where FnΦ (u) is given by (  Z ∞  h X jω (k) 1 1 ∞1 Φ ηn−1 Im x−jω 1 + δ Fn (x) = − 2 π 0 ω k k=1 i−1  × Z(k, δ, θ) dω (6) whereas δ = 2/α, Im{·} denotes the imaginary part of a complex number, and Z(k, δ, θ) has the form as (−1)k+1 θk 2 F1 (k, k − δ; k − δ + 1, −θ), (7) k−δ with 2 F1 (a, b; c, d) being the hypergeometry function [37], and (k) ηn−1 is given as Z 1 k ξ Φ (k) Φ F (dt). (8) ηn−1 = Fn−1 (ξ) + k n−1 ξ t Z(k, δ, θ) = (k) In particular, when n = 1, we have η0 = ξ k , ∀k ∈ N. 5 The mean field effect appears because of the cellular infrastruture, where interference is bounded away from the tagged transmitter. In Poisson bipolar networks, where the interaction is “strong” and “local”, such approximation may not hold. 5 1 Average active probability, η(ξ) 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 θ = 0 dB θ = 5 dB θ = 10 dB 0.1 0 0 0.2 0.4 0.6 0.8 1 Packet arrival rate, ξ Fig. 3. Average active probability versus packet arrival rate, under different SIR detection thresholds. for the trivial case of ξ = 0). Recall our discussion in Section II, the random locations of both BS and UE always result in some UEs located at poor coverage area, e.g., the cell edge, and having unbounded queue length. In this regard, instead of requiring all the queues to be stable, a more meaningful alternative will be to maintain the fraction of unstable queues to be below certain level. To this end, we introduce the εstable region [24], which gives the conditions for a network to be operated with less than ε portion of saturated queues. Following [24], a formal definition is in the sequel. Definition 2: For any ε ∈ [0, 1], the ε - stable region Sε and the critical arrival rate ξc are defined respectively as ) ) ( ( T X 1 ≤ε (10) µΦ Sε = ξ ∈ R+ : P lim xo ≤ ξ T →∞ T t=1 and ξc = sup Sε . Proof: See Appendix A. The expression in (4) not only quantifies how all the key features of a cellular network, i.e., deployment strategy, interference, and traffic profile, affect the distribution of SIR, but also illustrates how the interacting queues are affecting the SIR coverage via a fixed-point functional equation. The function F Φ (u) can be interpreted from two aspects: If we regard a typical cell as a queuing system, equation (4) describes the distribution of the random service rate; if we look at (4) in the view of network performance, then the CCDF, i.e., 1−F Φ (u), gives the level of certainty that at least u fraction of UEs in the network can attain SIR threshold θ. Several numerical results based on (4) will be shown in Section IV to provide more practical insights. In the following, we discuss the stable region, moments of the conditional SIR coverage probability, and approximation for the CDF. Remark 1: We introduce an auxiliary function as Z 1 ξ Φ Φ F (dt) (9) η(ξ) = F (ξ) + ξ t which can be regarded as “average” active probability. Note that as ξ grows from 0 to 1, while the value of η(ξ) also variates accordingly, the trend is very different. As depicted in Fig. 3, we can see that, with failed packet retransmissions, the average active probability goes up on a fast-and-then-slow basis with respect to the packet arrival rate ξ, and this effect is especially significant in networks with high SIR detection threshold. B. Stable Region The primal consideration in queuing systems is about stability, i.e., the critical conditions under which all the queues can remain finite length and do not explode. For an isolated system, even with random arrival and departure process, the stable region can be explicitly determined to be when average service rate is larger than the average arrival rate (Loynes theorem [38]). However, such condition cannot be directly generalized to large-scale queuing networks, as the strict stability, i.e., all queues are finite-length, is not achievable (except (11) The network is ε-stable if and only if ξ ≤ ξc . The critical arrival rate gives an explicit stable boundary, beyond which there are more queues transfering from finite size to infinite length and the requirement of less than ε portion of unstable queues cannot be guaranteed. While an exact expression for the critical arrival rate is not available, we can nevertheless obtain explicit conditions for the network to be ε-stable. Theorem 2: The sufficient condition for the network to be ε-stable is ( (  Z ∞  h X jω 1 ∞1 + 1 Sc Im x−jω 1+δ ξ ≤ ξε = sup ξ ∈ R : − 2 π 0 ω k  k=1 i−1  × Z(k, δ, θ) dω ≤ ε , (12) and the necessary condition for the network to be ε-stable is ( (  Z ∞  h X 1 ∞1 jω Nc + 1 ξ ≤ ξε = sup ξ ∈ R : − Im x−jω 1+δ 2 π 0 ω k  k=1 i−1  × ξ k Z(k, δ, θ) dω ≤ ε . (13) Proof: For the sufficient condition, we consider a dominant system, where all the nodes keep transmitting irrespect of their buffer status (for a node with empty buffer, it transmits “dummy” packets). As such, the interference in dominant system is larger than the actual ones, and the result can be attained via having ξ = 1 in (4). For the necessary condition, we consider a favorable system, where each node transmits without retransmission, i.e., even there is transmission failure, the transmitter simplily ignores the failure and withdraw the failed packets. In this scenario, the interference is smaller than that of the actual one, whereas we can obtain the result as making ηk = ξ k . From the above results, we note that the actual value of the critical arrival rate falls in the interval of [ξ Sc , ξ Nc ]. Using the ergodic property of PPP, an approximation for the ε-stable region is given in the sequel. 6 Corollary 1: The ε-stable region of the network is approximated as follows  A ξ ≤ ξc,ε = sup ξ ∈ R+ : F Φ (ξ) ≤ ε . (14) Proof: Note that due to stationary and ergodicity, the ensemble average obtained by averaging over the point process equals the spatial averages obtained by averaging an arbitrary realization of PPP over a large region, i.e., ! T 1X Φ x0 lim P µx0 ,t ≤ ξ = F Φ (ξ), (15) T →∞ T t=1 and the result follows according to the definition. Remark 2: With the notion of stability, the function η(ξ) defined in (9) can be interpreted as: On average, there are F Φ (ξ) portion of UEs in the network with infinity queue size and keep transmitting, while the rest 1 − F Φ (ξ) portion of UEs maintain R 1 stable queues, and active independently with probability ξ ξ/tF Φ (dt). C. Moments Based on the CDF of the conditional SIR coverage probability, in this part we give the corresponding moments, which can faciliate the assessment of network performance. Theorem 3: The m-th moment of the conditional SIR coverage probability is given by 1  Pm , (16) Mm = (k) 1 + δ k=1 m k η Z(k, δ, θ) where η (k) is given as η (k) = F Φ (ξ) + Z ξ 1  k ξ F Φ (dt) t P(γx0 > θ) = Proof: See Appendix B. Because of its important role in network performance assessment, we further provide bounds as well as an approximation for the SIR coverage probability to attain better insights. Corollary 2: The SIR coverage probability can be respectively bounded by the following 1 1 < P(γx0 > θ) < (19) 1 + δZ(1, δ, θ) 1 + δξZ(1, δ, θ) and when ξ ≪ 1, the SIR coverage probability can be approximated as follows (20) Proof: The upper and lower bounds can be obtained via similar approach as in the proof of Theorem 2, i.e., by considering a favorable system without retransmission and a dominant system that keeps transmitting, respectively. For the approximation, we notice that F Φ (ξ) → 0 as ξ → 0. Hence, we assume all the queues are stable in the regime with 1 1+ δξ P(γx0 >θ) Z(1, δ, θ) . (21) Remark 3: Note that the gap between the upper and lower bounds decreases with the increment of the packet arrival rate ξ, which is consistent with the conclusion in [21]. Furthermore, in the light traffic regime, the outage probability, 1 − P(γx0 > θ), increases linearly with respect to the packet arrival rate ξ, which demonstrates the significant impact of traffic profile on the system performance. D. Approximation Motivated by the fact that F Φ (u) is supported on the interval [0, 1], we content ourselves in this part by approximating the function F Φ (u) via a Beta distribution to reduce the computational complexity [32]. A formal operation is stated in the sequel. Corollary 3: The probability distribution function (pdf) of FnΦ (x) in Theorem 1 can be tightly approximated via the following µn (βn +1)−1 x 1−µn (1 − x)βn −1 fXn (x) = B(µn βn /(1 − µn ), βn ) (22) where B(a, b) denotes the Beta function, µn and βn are respectively given as (17) with F Φ (u) giving in (4). In particular, when m = 1, we have the standard SIR coverage probability given as 1 P (γx0 > θ) = . (18) 1 + δηZ(1, δ, θ) P(γx0 > θ) ≈ 1 − ξδZ(1, δ, θ) √ √ (α=4) = 1 − ξδ θ arctan( θ). very light traffic load, and approximate the active probability using the mean, i.e., ξ/P(γx0 > θ) [23]. The result then follows from solving the following equation (n) µn = M 1 , βn = (23) (n) (µn − M2 )(1 − (n) M2 − µ2n µn ) (24) (n) where Mm can be written as (n) Mm = 1+δ (k) with η̂n−1 being (k) η̂n−1 = Z Pm m k=1 k ξ fXn−1 (t)dt + 1 ,  (k) η̂n−1Z(k, δ, θ) Z ξ 0 (25) 1 k ξ fX (t)dt. tk n−1 (26) (k) When n = 1, we have η̂0 = ξ k , ∀k ∈ N. Proof: At each step of the iteration in Theorem 1, the function FnΦ (u) is supported on [0, 1]. As such, by respectively matching the mean and variance to a Beta distribution B(an , bn ), it yields an (n) = M1 , (27) an + b n i h 2 an b n (n) (n) (28) = M2 − M1 2 (an + bn ) (an + bn + 1) and the result follows by solving the above system equations. The accuracy of Corollary 3 will be verified in Fig. 6. 7 Packet arrival Packet departure 12 3 4 5 2 0 4 8 6 D1=2 D2=4 D3=3 ... ... Fig. 4. Illustration of delay in queuing system. The number of required slots to deliver each packet varies due to different queuing and retransmission results. E. Mean Delay and Analysis In the context of wireless networks, delay is an important factor that determines the Quality of Service (QoS) [14], [21]. As shown in Fig. 4, the random arrival of packets and interference-limited channel will inevitably incur waiting and retransmission in the packet delivery process, thus induce multiple slots for one successful delivery. Besides, since the number of required time slots differs among each packet, we can only quantify the delay in average manner. In its accordance, the following definition formalizes the notion of mean day. Definition 3: Let Ax (T ) be the number of packets arrived at a typical transmitter x within period [0, T ], and Di,x be the number of time slots between the arrival of the i-th packet and its successful delivery. The mean delay is defined as PAx (T ) Di,x i=1 DΦ , lim . (29) x T →∞ Ax (T ) Note that Di,x in (29) represents the number of time slots required to successfully deliver the i-th packet, and its value is affected by: (i) queueing delay, caused by other accumulated unsent packets, and (ii) transmission, or equivalently local, delay, due to link failure and retransmission [39]–[41]. By averaging over all time slots, (29) provides information on the average number of slots to successfully deliver a packet. Furthermore, using results from queuing theory and stochastic geometry, we can characterize its distribution with the following expression. Theorem 4: The CDF of mean delay at a typical UE is given as    1−ξ Φ + ξ ≤ T = 1 − F P DΦ x0 T µn (βn +1)−1 Z 1 x 1−µn (1 − x)βn −1 ≈ lim dx n→∞ ξ+ 1−ξ B(µn βn /(1 − µn ), βn ) T (30) where F Φ (u) is given by (4), and µn and βn are respectively given in (23) and (24). Proof: Without loss of generality, we focus on the typical UE. Given the service rate, i.e., the conditional success probability, µΦ x0 , we have the conditional mean delay being [21]  1−ξ Φ   µΦx −ξ , if µx0 > ξ, 0 Φ Dx0 = (31)   +∞, Φ if µx0 ≤ ξ. The distribution of mean delay can then be computed as  (a)   Φ Φ Φ P DΦ x0 ≤ T = P Dx0 ≤ T |µx > ξ P µx > ξ   Φ Φ + P DΦ x0 ≤ T |µx ≤ ξ P µx ≤ ξ  (b) Φ = P DΦ x0 ≤ T, µx > ξ   1−ξ Φ (32) + ξ ≤ µx =P T where (a) is by the law of total probability, and (b) follows by Φ noticing P(DΦ x0 ≤ T |µx ≤ ξ) = 0. We then obtain the result by using (4) and Corollary 3 to the above. The accuracy of Theorem 4 will be verified in Fig. 10. Moreover, two observations immediately follows from (30). Observation 1: When T = 1, we have P DΦ x0 ≤ T = 0, which states that the UEs who are able to success without retransmission form a probabilistic null set in large queuing networks. Observation 2: As T ≫ 1, the CCDF of mean delay Φ can be approximated as, P(DΦ x0 > T ) ≈ F (ξ) + 1/T , which implies the distribution is heavy tail. In addition, when T → ∞, we have the delay outage, i.e., 1 − P(DΦ x0 ≤ T ) converges to F Φ (ξ), which represents the fraction of UEs that are experiencing unstable queues, and it is in line with our discussion in Section-III-B. IV. N UMERICAL R ESULTS AND S IMULATIONS In this section, we validate the accuracy of our analysis through simulations, and explore the impact of traffic condition on network performance from several aspects. During each simulation run, the BSs and UEs are realized over a 100 km2 area via independent PPPs. Packets arrive at each node according to independent Bernoulli process. We average over 10,000 realizations and collect the statistic from each cell to finally calculate the SIR coverage probability. Unless differently specified, we use the following parameters for path loss exponent, BS density, and packet arrival rate, respectively: α = 3.8, λb = 10−4 BS/km2 , and ξ = 0.3 packet/slot. In Fig. 5, we compare the simulated CDF of conditional SIR coverage probability to the analysis proposed in Theorem 1, for various values of SIR detection threshold θ and packet arrival rate ξ. First, the results show a close match for all values of θ and ξ, which validate the our mathematical framework. Next, note that from Fig. 5(a) we can quantify the level of confidence about how much fraction of UEs are able to achieve the targeted SIR at different thresholds, it is also coined as the reliability of the network [32]. For instance, when ξ = 0.3, the probability that 80% of UEs in the network can attain -5 dB SIR is 0.85, showing that with high possibility the majority of UEs can have their packets correctly decoded 8 1 Simulations Analysis 0.9 Cumulative distribution function, FΦ(u) Cumulative Distribution Function, FΦ(u) 1 0.8 0.7 0.6 0.5 0.4 θ = 0 dB 0.3 0.2 0.1 0 0 0.2 0.4 0.6 Conditional SIR coverage probability, 0.8 1 Simulations Analysis 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 ξ = 0.10 0.1 0 0 0.2 Φ µx 0.4 0.6 0.8 1 Φ Conditonal SIR coverage probability, µx (a) (b) Fig. 5. Simulation versus analysis: CDF of conditional success probability. In Fig. (a), we fix the packet arrival rate to be ξ = 0.3, and varies the SIR threshold as θ = −10, −5, 0, 5, 10, 15 dB. In Fig. (b), we fix the SIR threshold to be θ = 0 dB, and varies the arrival rate as ξ = 0.05, 0.1, 0.3, 0.5, 0.8. Taking a closer look between Fig. 5(a) and Fig. 5(b), we notice that under the same cellular configuration, the conditional SIR coverage probability with θ = 0 dB is larger than that with θ = 5 dB, even their respective packet arrival rates are ξ = 0.8 and ξ = 0.3. This observation indicates that packets with smaller detection threshold, i.e., short packets, are more preferable to the network. The reason comes from the fact that higher SIR threshold not merely increases the possibility of failure in decoding process, but triggers more retransmissions thus prolong the active period of interferers, which further reduces the coverage probability. Fig. 6 compares the CDF of conditional SIR coverage 1 Cumulative distribution function, FΦ(u) without retransmission. Such probability can drop to 0.08 if the SIR detection threshold increases to 5 dB, indicating that in large-scale queuing networks, it is very unlikely to maintain the majority of UEs at high SIR. On the other hand, from Fig. 5(b) we observe that the conditional SIR coverage probability monotonically decreases with the growth of packet arrival rate, due to the fact that more and more BSs are activated because of their non-empty buffers and thus impose additional inter-cell interference. It is also worthwhile to note that the increment of traffic load defects the network SIR coverage in a non-linear manner, whereas the degradation of SIR coverage is more severe as traffic load goes from light (ξ = 0.05) to medium (ξ = 0.3), and the decreasing trend slows down as the network load further increases to heavy traffic regime (ξ = 0.8). The reason comes from the composite effect of temporal traffic. In light traffic condition, as packet arrival rate goes up, the increased traffic load not only wakes up more BSs, but also brings in more accumulated packets at the buffer. Together with the reduced service rate, the active duration of transmitters is extended, which in turns defect the SIR across the network. In the heavy traffic regime, as most of the queues are already saturated, the additional active cells cannot largely change the interference, and thus the SIR coverage probability descent is leveled off. Such explanation also goes to the observation in Remark 1. Approximation: Corollary 3 Analysis: Theorem 1 0.9 0.8 0.7 α=3 0.6 0.5 0.4 0.3 α=4 0.2 α = 3.5 0.1 0 0 0.2 0.4 0.6 0.8 1 Conditional SIR coverage probability, µΦ x Fig. 6. CDF of conditional SIR coverage probability (Theorem 1) and the approximation (Corollary 3). Parameters are set as: SIR detection threshold θ = 0 dB, and packet arrival rate ξ = 0.3. probability to its approximation in Corollary 3. We can see that the approximation matches well with the analysis under different path loss exponents, thus confirms its accuracy. We further observe that smaller path loss exponent reduces the SIR coverage, which is due to the fact that while smaller α enhances the received signal power, it nevertheless also results in higher inter-cell interference, and the latter subpass the power gain and hence defacts the SIR. Fig. 7 plots the maximum arrival rates per sufficient and necessary conditions as functions of ε. The figure shows a large difference between the necessary and sufficient conditions, while the approximation locates in between, hence confirms the necessity for better SIR characterization in traffic network. It also reveals that a slight change of arrival rate can largely impact the stable region, i.e., the network stability is vulnerable to traffic condition. Fig. 8 depicts the standard SIR coverage probability as a 9 0.6 Cell-dege coverage probability, 1-F Φ(0.95) 0.5 Maximum arrival rate, ξ 1 Necessary condition Approximation Sufficient condition 0.4 0.3 0.2 0.1 0 0 0.05 0.1 0.15 0.2 0.25 0.3 ǫ 0.8 0 SIR coverage probability, P(γ x > θ) 0.9 ξ = 0.50 0.7 ξ = 0.05 0.6 0.5 0.4 0.3 0 -15 Upper bounds Lower bounds Simulations Analysis -10 -5 0 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.2 0.4 0.6 0.8 1 Fig. 9. 95%-likely rate versus packet arrival rate ξ, for various SIR thresholds θ = −10, −5, 0, 5, 10 dB. 1 0.1 0.8 Packet arrival rate, ξ [Packets/slot] Fig. 7. Comparison of necessary and sufficient conditions. 0.2 0.9 5 10 15 SIR threshold, θ [dB] Fig. 8. Success probability versus SIR threshold θ. function of SIR detection threshold θ, under different packet arrival rates. We first note the close match between the analysis and the simulation, which validates the accuracy of Theorem 3. More importantly, the figure confirms that the traffic profile plays a crucial role in the SIR coverage probability. For instance, to maintain a 90% coverage probability, the required SIR detection threshold can differ more than 10 dB in light traffic (ξ = 0.05) and heavy traffic (ξ = 0.5) regimes, and this gap will be larger as the packet arrival rate keeps increasing. Fig. 8 also reveals that the upper and lower bounds are not tight, whereas even in heavy traffic regime the difference between the two bounds is around 5 dB, and this gap increases dramatically in light traffic condition. Moreover, the bounds fail to capture the trend of coverage probability variates according to the detection threshold θ. As what we have discussed in Remark 1, both ξ and θ significantly affect the active probability, hence, simply approximating the transmitting BSs to a thinned point process cannot provide desire results and hence validate the importance of the trafficaware analysis. In Fig. 9 we plot the cell-edge coverage probability, i.e., 1 − F Φ (0.95), as a function of packet arrival rate ξ. This quantity represents the performance of the “5% UEs”, i.e., the UEs in the bottom 5th percentile in terms of performance, and is particularly interested to operators [15], [32]. We can see that the cell-edge coverage probability is vulnerable to traffic fluctuation. Even in the case with very small SIR detection threshold, e.g., θ = −10 dB, the cell-edge coverage probability can drop by half as the packet arrival rate changes from low to high, and this defection is more severe in scenarios with high SIR thresholds. As such, it is critical to employ more advanced technology to boost up the SIR performance at the cell edge [13]. Fig. 10 compares the CDF of mean delay given in Theorem 4 to the values obtained from simulations. The figure shows that simulation results agree well with the analytical values. Moreover, Fig. 10 confirms the observation in Section III by showing a heavy tail behavior in the distribution of mean delay. Fig. 10 also reveals that the mean delay is very sensitive to the variation of traffic load, e.g., the delay outage is round 0.01 when ξ = 0.2, but this value climbs to around 0.25 when the packet arrival rate doubles, i.e., ξ = 0.4. It is becuase higher the traffic load, on one hand incurs more retransmissions via increased interference, on the other, also prolongs waiting time of each packet since the additional incomings quickly occupy all available buffers. This composite effect significantly affects the mean delay. V. C ONCLUSION In this paper, we introduced an analytical toolset to evaluate the impact of temporal traffic on the performance of cellular networks. We used a general model that accounts for key features from both spatial and temporal domain, including the channel fading, path loss, network topology, traffic profile, and queuing evolution. By exploiting queuing theory and stochastic geometry, we obtain the SIR distribution through a fixed-point functional equation, and validated its accuracy by simulation. Our results confirmed that temporal traffic profile can largely affect the network SIR performance. In particular, 10 following 1 0.9 x CDF of mean delay, P(D Φ < T) ξ = 0.2 ξ = 0.3 0.8 0.7 ξ = 0.4 0.6 Simulations Analysis 0.5 2 4 6 8 10 12 14 16 18 20 Φ Delay, D x [Time slots] Fig. 10. Simulation versus analysis: Mean delay distribution. it showed that under the same configuration, when traffic condition changes from light to heavy, the corresponding SIR requirement can differ by more than 10 dB for the network to maintain coverage. Moreover, the SIR coverage probability varies largely with traffic fluctuation in the submedium load regime, whereas in scenario with very light traffic load, the SIR outage probability increases linearly with the packet arrival rate. In addition, the mean delay, as well as coverage probability of cell-edge UEs are vulnerable to the traffic fluctuation, thus confirms its appeal for traffic-aware communication technologies. The derivation of SIR distribution as a tractable form of system parameters opens various areas to gain further design insights. On one hand, the framework can be extended to adopt more sophisticated point process, e.g., the Poisson cluster point process [42], or determinantal point processes [43]. On the other hand, the analysis can be applied to investigate the design of different wireless technologies. For example, the traffic scheduling problem in large-scale wireless networks [21], or the resource allocation problem in Dynamic TDD system [44]. Analyzing impact of temporal traffic on the performance of Massive-MIMO system is also a concrete direction to investigate in the future. A PPENDIX A. Proof of Theorem 1   Φ Φ (s) = E exp sY MYx,0 x,0   s Y  ξ +1−ξ  = E 1 + θkxkα /kzkα z∈Φ\x s  !   Z ∞ ξ (a) +1−ξ dx = exp −λb 1− 1+θkx0kα /kxkα B c (0,kxk) Z ∞  (b) = 2πλb r exp −λb πr2 0 k ! Z ∞ ∞   X s ξ 2 k+1 × exp −2πλb r (−1) vdv dr k 1+v α/2 /θ 1 k=1  ∞  h i−1 X jω k (c) = 1+δ ξ Z(k, δ, θ) (33) k k=1 where (a) follows from the probability generating functional of PPP, (b) comes from polar coordinate transform and deconditioning kxk with its pdf, which follows a Rayleigh distribution 2 as fR (r) = 2πλb re−λb πr , and (c) is by algebraic operation. By the Gil-Pelaez theorem [45], we have the CDF of µΦ x,0 given as  Φ F0Φ (u) = P (P (γx,0 > θ|Φ) < u) = P Yx,0 < ln u Z n o 1 ∞ 1 1 Im u−jω MYx,0 Φ (jω) dω. (34) = − 2 π 0 ω Next, consider all the queues have evolved to time t = n, Φ and we have obtained the CDF of µΦ x,n−1 as P(µx,n−1 < Φ Φ u) = Fn−1 (u), the moment generating function of Yx,n can be calculated as s MYx,n Φ (s) = E [P (γx,n > θ|Φ) ]    = E E P (γx,n > θ|Φ)s Fn−1    s Y qz,n Fn−1  +1−qz,n = EE 1+θkxkα/kzkα z∈Φ\x ∞   X  s 2πλb r exp −λb πr2 exp −2πλb r2 (−1)k+1 k 0 k=1 "Z  #! k ∞ qx,n ×E dr vdv Fn−1 α/2 /θ 1+v 1 ∞   i−1 h X   k s E qx,n |Fn−1 Z(k, δ, θ) = 1+δ . (35) k = Z ∞ k=1 To faciliate the presentation, we denote Fn as the σ-algebra that contains all the information about queuing status of every Φ node x ∈ Φb up to time t = n. We further introduce Yx,n and Φ Φ qx,n to denote Yx,n = ln P(γx,n > θ|Φ) and qx,n = P(ζx,n = 1), respectively. As such, the σ-algebra {Fn }∞ n=0 forms a filtration, where Fn−1 ⊂ Fn . At slot t = 0, every transmitter has one packet arrival with probability ξ, and actives with probability ξ, i.e., qx,0 = ξ, ∀x ∈ Φb . We can thus compute the moment Φ generation function of Yx,0 at a generic BS x ∈ Φb , via the Leveraging Lemma 1, we have # " k   k  ξ χ{µΦx,n >ξ} |Fn−1 E qx,n |Fn−1 = E χ{µΦx,n ≤ξ} + µΦ x,n Z 1 k ξ Φ Φ F (dt) (36) = Fn−1 (ξ) + k n−1 ξ t where χ{·} is the indicator function. By substituting (36) back into (35), and using the the Gil-Pelaez theorem for another 11 time, we have the CDF of µΦ x,n given as Φ Fx,n (u) Φ Yx,n = P (P (γx,n > θ|Φ) < u) = P < ln u Z ∞ n o 1 1 1 = − Im u−jω MYx,n Φ (jω) dω. 2 π 0 ω  (37) Noet that FnΦ (u) appears on the left hand side of (37), Φ and Fn−1 (u) appears on the right hand side. As FnΦ (u) ≤ Φ F0 (u), ∀u ∈ [0, 1], n ∈ N, by the Dominant Convergence Theorem [46], we have FnΦ (u) → F Φ (u), as n → ∞, and the result follows. B. Proof of Theorem 3 We denote F = ∪n Fn to be the σ-algebra of all the queuing status at the steady state, and qx the corresponding active probability at a generic node x. As such, the m-moment of the conditional success probability can be computed as i h m E lim P (γx,n > θ|Φ) n→∞   m Y E[qz |Fn ] +1−E[qz |Fn ]  = E lim n→∞ 1+θkxkα /kzkα z∈Φ\x m   h i−1 X   m = 1+δ E qxk |F Z(k, δ, θ) , (38) k k=1 and the result follows by using Lemma 1 and (4) to the above equation. R EFERENCES [1] D. López-Pérez, M. Ding, H. Claussen, and A. H. Jafari, “Towards 1 Gbps/UE in cellular systems: Understanding ultra-dense small cell deployments,” IEEE Commun. Surveys & Tuts., vol. 17, no. 4, pp. 2078– 2101, 4th Quart., 2015. [2] Z. Niu, “TANGO: Traffic-aware network planning and green operation,” ”IEEE Wireless Commun.”, vol. 18, no. 5, Oct. 2011. [3] M. Haenggi, J. G. Andrews, F. Baccelli, O. Dousse, and M. Franceschetti, “Stochastic geometry and random graphs for the analysis and design of wireless networks,” IEEE J. Sel. Areas Commun., vol. 27, no. 7, pp. 1029–1046, Sep. 2009. [4] M. Haenggi, Stochastic geometry for wireless networks. Cambridge University Press, 2012. [5] F. Baccelli and B. Blaszczyszyn, Stochastic Geometry and Wireless Networks. Volumn I: Theory. Now Publishers, 2009. [6] H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic geometry for modeling, analysis, and design of multi-tier and cognitive cellular wireless networks: A survey,” IEEE Commun. Surveys & Tuts., vol. 15, no. 3, pp. 996–1019, Jun. 2013. [7] H. ElSawy, A. Sultan-Salem, M.-S. Alouini, and M. Z. Win, “Modeling and analysis of cellular networks using stochastic geometry: A tutorial,” IEEE Commun. Surveys & Tuts., vol. 19, no. 1, pp. 167–203, 1st Quart., 2017. [8] M. Di Renzo, W. Lu, and P. Guan, “The intensity matching approach: A tractable stochastic geometry approximation to system-level analysis of cellular networks,” IEEE Trans. Wireless Commun., vol. 15, no. 9, pp. 5963–5983, Sep. 2016. [9] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to coverage and rate in cellular networks,” IEEE Trans. Commun., vol. 59, no. 11, pp. 3122–3134, Nov. 2011. [10] H. S. Dhillon, R. K. Ganti, F. Baccelli, and J. G. Andrews, “Modeling and analysis of k-tier downlink heterogeneous cellular networks,” IEEE J. Sel. Areas Commun., vol. 30, no. 3, pp. 550–560, Apr. 2012. [11] H. H. Yang, J. Lee, and T. Q. S. Quek, “Heterogeneous cellular network with energy harvesting-based D2D communication,” IEEE Trans. Wireless Commun., vol. 15, no. 2, pp. 1406–1419, Feb. 2016. [12] H. H. Yang, G. Geraci, and T. Q. S. Quek, “Energy-efficient design of MIMO heterogeneous networks with wireless backhaul,” IEEE Trans. Wireless Commun., vol. 5, no. 7, pp. 4914–4927, July 2016. [13] H. H. Yang, G. Geraci, T. Q. S. Quek, and J. G. Andrews, “Cell-edgeaware precoding for downlink massive MIMO cellular networks,” IEEE Trans. Signal Process., vol. 65, no. 13, pp. 3344–3358, Jul. 2017. [14] Y. Zhong, M. Haenggi, F.-C. Zheng, W. Zhang, T. Q. S. Quek, and W. Nie, “Towards a tractable delay analysis in large wireless networks,” arXiv preprint arXiv:1612.01276, 2016. [15] J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K. Soong, and J. C. Zhang, “What will 5G be?” IEEE J. Sel. Areas Commun., vol. 32, no. 6, pp. 1065–1082, Jun. 2014. [16] R. R. Rao and A. Ephremides, “On the stability of interacting queues in a multiple-access system,” IEEE Trans. Inf. Theory, vol. 34, no. 5, pp. 918–930, Sep. 1988. [17] D. P. Bertsekas, R. G. Gallager, and P. Humblet, Data networks. Prentice-Hall International New Jersey, 1992, vol. 2. [18] W. Luo and A. Ephremides, “Stability of N interacting queues in random-access systems,” IEEE Trans. Inf. Theory, vol. 45, no. 5, pp. 1579–1587, Jul. 1999. [19] C. Bordenave, D. McDonald, and A. Proutiere, “Asymptotic stability region of slotted aloha,” IEEE Trans. Inf. Theory, vol. 58, no. 9, pp. 5841–5855, Sep. 2012. [20] ——, “Asymptotic stability region of slotted aloha,” IEEE Trans. Inf. Theory, vol. 58, no. 9, pp. 5841–5855, Sep. 2012. [21] Y. Zhong, T. Q. S. Quek, and X. Ge, “Heterogeneous cellular networks with spatio-temporal traffic: Delay analysis and scheduling,” IEEE J. Sel. Areas Commun., vol. 35, no. 6, pp. 1373–1386, Jun. 2017. [22] M. Gharbieh, H. ElSawy, A. Bader, and M.-S. Alouini, “Spatiotemporal stochastic modeling of IoT enabled cellular networks: Scalability and stability analysis,” IEEE Trans. Commun., vol. 65, no. 9, pp. 3585–3600, Aug. 2017. [23] H. H. Yang, G. Geraci, Y. Zhong, and T. Q. S. Quek, “Packet throughput analysis of static and dynamic TDD in small cell networks,” IEEE Wireless Commun. Lett., vol. 6, no. 6, pp. 742–745, Dec. 2017. [24] Y. Zhong, M. Haenggi, T. Q. Quek, and W. Zhang, “On the stability of static poisson networks under random access,” IEEE Trans. Commun., vol. 64, no. 7, pp. 2985–2998, Jul. 2016. [25] A. George, T. Spyropoulos, and F. Kaltenberger, “An analytical model for flow-level performance in heterogeneous wireless networks,” IEEE Trans. Wireless Commun., 2017. [26] D. B. Taylor, H. S. Dhillon, T. D. Novlan, and J. G. Andrews, “Pairwise interaction processes for modeling cellular network topology,” in Proc. IEEE Global Telecomm. Conf., Anaheim, CA, Dec. 2012, pp. 4524– 4529. [27] B. Blaszczyszyn, M. K. Karray, and H. P. Keeler, “Using Poisson processes to model lattice cellular networks,” in Proc. IEEE Conf. on Computer Commun. IEEE, Apr. 2013, pp. 773–781. [28] M. Ding, P. Wang, D. López-Pérez, G. Mao, and Z. Lin, “Performance impact of LoS and NLoS transmissions in dense cellular networks,” IEEE Trans. Wireless Commun., vol. 15, no. 3, pp. 2365–2380, Mar. 2016. [29] J. G. Andrews, X. Zhang, G. D. Durgin, and A. K. Gupta, “Are we approaching the fundamental limits of wireless network densification?” IEEE Commun. Mag., vol. 54, no. 10, pp. 184–190, Oct. 2016. [30] Y. J. Chun, S. L. Cotton, H. S. Dhillon, F. J. Lopez-Martinez, J. F. Paris, and S. K. Yoo, “A comprehensive analysis of 5G heterogeneous cellular systems operating over κ–µ shadowed fading channels,” IEEE Trans. Wireless Commun., vol. 16, no. 11, pp. 6995–7010, Nov. 2017. [31] I. Trigui, S. Affes, and B. Liang, “Unified stochastic geometry modeling and analysis of cellular networks in LOS/NLOS and shadowed fading,” IEEE Transactions on Communications, vol. 65, no. 12, pp. 5470–5486, Dec. 2017. [32] M. Haenggi, “The meta distribution of the SIR in poisson bipolar and cellular networks,” IEEE Trans. Wireless Commun., vol. 15, no. 4, pp. 2577–2589, Apr. 2016. [33] U. Schilcher, C. Bettstetter, and G. Brandner, “Temporal correlation of interference in wireless networks with rayleigh block fading,” IEEE Trans. Mobile Comput., vol. 11, no. 12, pp. 2109–2120, Dec. 2012. [34] Y. Zhong, W. Zhang, and M. Haenggi, “Managing interference correlation through random medium access,” IEEE Trans. Wireless Commun., vol. 13, no. 2, pp. 928–941, Feb. 2014. [35] R. K. Ganti and M. Haenggi, “Spatial and temporal correlation of the interference in aloha ad hoc networks,” IEEE Wireless Commun. Lett., vol. 13, no. 9, Sep. 2009. [36] I. Atencia and P. Moreno, “A discrete-time Geo/G/1 retrial queue with general retrial times,” Queueing Systems, vol. 48, no. 1-2, pp. 5–21, 2004. [37] G. E. Andrews, R. Askey, and R. Roy, Special functions. Cambridge University Press, Cambridge, 2000. 12 [38] R. M. Loynes, “The stability of a queue with non-independent interarrival and service times,” in Math. Proc. Cambridge Philos. Soc., vol. 58, no. 3. Cambridge University Press, 1962, pp. 497–520. [39] M. Haenggi, “The local delay in poisson networks,” IEEE Trans. Inf. Theory, vol. 59, no. 3, pp. 1788–1802, Mar. 2013. [40] G. Zhang, T. Q. S. Quek, A. Huang, and H. Shan, “Delay and reliability tradeoffs in heterogeneous cellular networks,” IEEE Trans. Wireless Commun., vol. 15, no. 2, pp. 1101–1113, Feb. 2016. [41] G. Zhang, T. Q. S. Quek, M. Kountouris, A. Huang, and H. Shan, “Fundamentals of heterogeneous backhaul designanalysis and optimization,” IEEE Trans. Commun., vol. 64, no. 2, pp. 876–889, Feb. 2016. [42] R. K. Ganti and M. Haenggi, “Interference and outage in clustered wireless ad hoc networks,” IEEE Trans. Inf. Theory, vol. 55, no. 9, pp. 4067–4086, Sep. 2009. [43] Y. Li, F. Baccelli, H. S. Dhillon, and J. G. Andrews, “Statistical modeling and probabilistic analysis of cellular networks with determinantal point processes,” IEEE Trans. Commun., vol. 63, no. 9, pp. 3405–3422, Sep. 2015. [44] Z. Shen, A. Khoryaev, E. Eriksson, and X. Pan, “Dynamic uplinkdownlink configuration and interference management in TD-LTE,” IEEE Commun. Mag., vol. 50, no. 11, pp. 51–59, Nov. 2012. [45] J. Gil-Pelaez, “Note on the inversion theorem,” Biometrika, vol. 38, no. 3-4, pp. 481–482, Dec. 1951. [46] P. Billingsley, Probability and measure. John Wiley & Sons, 2008.
7cs.IT
MAXIMALLY STRETCHED LAMINATIONS ON GEOMETRICALLY FINITE HYPERBOLIC MANIFOLDS arXiv:1307.0250v2 [math.GT] 11 Jan 2017 FRANÇOIS GUÉRITAUD AND FANNY KASSEL Abstract. Let Γ0 be a discrete group. For a pair (j, ρ) of representations of Γ0 into PO(n, 1) = Isom(Hn ) with j geometrically finite, we study the set of (j, ρ)-equivariant Lipschitz maps from the real hyperbolic space Hn to itself that have minimal Lipschitz constant. Our main result is the existence of a geodesic lamination that is “maximally stretched” by all such maps when the minimal constant is at least 1. As an application, we generalize two-dimensional results and constructions of Thurston and extend his asymmetric metric on Teichmüller space to a geometrically finite setting and to higher dimension. Another application is to actions of discrete subgroups Γ of PO(n, 1) × PO(n, 1) on PO(n, 1) by right and left multiplication: we give a double properness criterion for such actions, and prove that for a large class of groups Γ the action remains properly discontinuous after any small deformation of Γ inside PO(n, 1) × PO(n, 1). Contents 1. Introduction 2 2. Preliminary results 10 3. An equivariant Kirszbraun–Valentine theorem for amenable groups 18 4. The relative stretch locus 23 5. An optimized, equivariant Kirszbraun–Valentine theorem 40 6. Continuity of the minimal Lipschitz constant 52 7. Application to properly discontinuous actions on G = PO(n, 1) 66 8. A generalization of the Thurston metric on Teichmüller space 76 9. The stretch locus in dimension 2 80 10. Examples and counterexamples 90 Appendix A. Some hyperbolic trigonometry 104 Appendix B. Converging fundamental domains 108 Appendix C. Open questions 117 References 118 The authors were partially supported by the Agence Nationale de la Recherche under the grants ANR-11-BS01-013: DiscGroup (“Facettes des groupes discrets”) and ANR-09BLAN-0116-01: ETTT (“Extension des théories de Teichmüller–Thurston”), and through the Labex CEMPI (ANR-11-LABX-0007-01). 1 2 FRANÇOIS GUÉRITAUD AND FANNY KASSEL 1. Introduction For n ≥ 2, let G be the group PO(n, 1) = O(n, 1)/{±1} of isometries of the real hyperbolic space Hn . In this paper we consider pairs (j, ρ) of representations of a discrete group Γ0 into G with j injective, discrete, and j(Γ0 )\Hn geometrically finite, and we investigate the set of (j, ρ)-equivariant Lipschitz maps Hn → Hn with minimal Lipschitz constant. We develop applications, both to properly discontinuous actions on G and to the geometry of some generalized Teichmüller spaces (via a generalization of Thurston’s asymmetric metric). Some of our main results, in particular Theorems 1.8 and 1.11, Corollary 1.12, and Theorem 7.1, were initially obtained in [Ka1, Chap. 5] in the case n = 2 with j convex cocompact. 1.1. Equivariant maps of Hn with minimal Lipschitz constant. Let Γ0 be a discrete group. We say that a representation j ∈ Hom(Γ0 , G) of Γ0 in G = PO(n, 1) is convex cocompact (resp. geometrically finite) if it is injective with a discrete image j(Γ0 ) ⊂ G and if the convex core of the hyperbolic orbifold M := j(Γ0 )\Hn is compact (resp. has finite m-volume, where m ≤ n is its dimension). In this case, the group Γ0 identifies with the (orbifold) fundamental group of M . Parabolic elements in j(Γ0 ) correspond to cusps in M ; they do not exist if j is convex cocompact. We refer to Section 2.1 for full definitions. Let j ∈ Hom(Γ0 , G) be geometrically finite and let ρ ∈ Hom(Γ0 , G) be another representation, not necessarily injective or discrete. In this paper we examine (j, ρ)-equivariant Lipschitz maps of Hn , i.e. Lipschitz maps f : Hn → Hn such that Hn . f (j(γ) · x) = ρ(γ) · f (x) for all γ ∈ Γ0 and x ∈ A constant that naturally appears is the infimum of all possible Lipschitz constants of such maps:  (1.1) C(j, ρ) := inf Lip(f ) | f : Hn → Hn (j, ρ)-equivariant . A basic fact (Section 4.2) is that C(j, ρ) < +∞ unless there is an obvious obstruction, namely an element γ ∈ Γ0 with j(γ) parabolic and ρ(γ) hyperbolic. Here we use the usual terminology: a nontrivial element of G is elliptic if it fixes a point in Hn , parabolic if it fixes exactly one point on the boundary at infinity of Hn , and hyperbolic otherwise (in which case it preserves a unique geodesic line in Hn ). To make the statements of our theorems simpler, we include the identity element of G among the elliptic elements. We shall always assume C(j, ρ) < +∞. Then there exists a (j, ρ)-equivariant map f : Hn → Hn with minimal constant C(j, ρ), except possibly if the group ρ(Γ0 ) has a unique fixed point on the boundary at infinity ∂∞ Hn of Hn (see Section 4.4, as well as Sections 10.2 and 10.3 for examples). We fix once and for all a geometrically finite representation j0 ∈ Hom(Γ0 , G). Dealing with cusps is a substantial aspect of the paper; we make the following definitions, which are relevant only when j is not convex cocompact. Definition 1.1. We say that j ∈ Hom(Γ0 , G) has the cusp type of j0 if for any γ ∈ Γ0 , the element j(γ) is parabolic if and only if j0 (γ) is parabolic. We say that ρ ∈ Hom(Γ0 , G) is cusp-deteriorating with respect to j (or that MAXIMALLY STRETCHED LAMINATIONS 3 the pair (j, ρ) is cusp-deteriorating) if for any γ ∈ Γ0 with j(γ) parabolic, the element ρ(γ) is elliptic. In the sequel, we will always assume that j has the cusp type of the fixed representation j0 . Therefore, we will often just use the phrase “ρ cuspdeteriorating”, leaving j implied. Of course, this is an empty condition when j is convex cocompact. 1.2. The stretch locus. The main point of the paper is to initiate a systematic study of the stretch locus of equivariant maps of Hn with minimal Lipschitz constant. Definition 1.2. Let f : Hn → Hn be a (j, ρ)-equivariant map realizing the minimal Lipschitz constant C(j, ρ). The stretch locus Ef of f is the (j(Γ0 )-invariant) set of points x ∈ Hn such that the restriction of f to any neighborhood of x in Hn has Lipschitz constant C(j, ρ) (and no smaller). It follows from our study that the geometry of the stretch locus depends on the value of C(j, ρ). We prove the following. Theorem 1.3. Let (j, ρ) ∈ Hom(Γ0 , G)2 be a pair of representations with j geometrically finite and C(j, ρ) < +∞. Assume that there exists a (j, ρ)equivariant map f : Hn → Hn with minimal Lipschitz constant C(j, ρ), and let E(j, ρ) be the intersection of the stretch loci of all such maps. Then • E(j, ρ) is nonempty, except possibly if C(j, ρ) = 1 and ρ is not cuspdeteriorating (see Section 10.8 for an example); • there exists an “optimal” (j, ρ)-equivariant, C(j, ρ)-Lipschitz map f0 : Hn → Hn whose stretch locus is exactly E(j, ρ); • if C(j, ρ) > 1 (resp. if C(j, ρ) = 1 and ρ is cusp-deteriorating), then E(j, ρ) is a geodesic lamination (resp. contains a k-dimensional geodesic lamination for some k ≥ 1) with the following properties: – the lamination is “maximally stretched” by any (j, ρ)-equivariant map f : Hn → Hn with minimal Lipschitz constant C(j, ρ), in the sense that f multiplies all distances by C(j, ρ) on every leaf of the lamination; – the projection to j(Γ0 )\Hn of the lamination is compact and contained in the convex core. By a geodesic lamination (resp. a k-dimensional geodesic lamination) of Hn we mean a nonempty disjoint union L of geodesic lines (resp. k-planes) of Hn , called leaves, that is closed in the space of geodesic lines (resp. kplanes) of Hn . The image in j(Γ0 )\Hn of a j(Γ0 )-invariant geodesic lamination of Hn is a geodesic lamination in the usual sense. We note that an “optimal” map f0 is usually not unique since it can be slightly perturbed outside of the stretch locus E(j, ρ). In Section 9.1 we explain how, in the case that n = 2 and that j and ρ are both injective and discrete with finite covolume, Theorem 1.3 follows from Thurston’s theory [T2] of the asymmetric metric on Teichmüller space. More precise results in the case C(j, ρ) = 1 are given (for arbitrary n) in Section 5, leading to a reasonable understanding of the stretch locus when C(j, ρ) ≥ 1. On the other hand, for C(j, ρ) < 1 the stretch locus is more mysterious; we make the following conjecture. 4 FRANÇOIS GUÉRITAUD AND FANNY KASSEL Conjecture 1.4. For n = 2, let (j, ρ) ∈ Hom(Γ0 , G)2 be a pair of representations with j geometrically finite and let E(j, ρ) be the intersection of the stretch loci of all (j, ρ)-equivariant maps with minimal Lipschitz constant C(j, ρ) ∈ (0, 1). Then E(j, ρ) is the lift to H2 of a gramination (contraction of “graph” and “lamination”) of M := j(Γ0 )\H2 , by which we mean the union of a finite set F and of a lamination in M r F with finitely many leaves terminating on F . We discuss this conjecture and provide evidence for it in Section 9.4. We also examine the behavior of the minimal Lipschitz constant C(j, ρ) and of the stretch locus E(j, ρ) under small deformations of j and ρ. We prove that the constant C(j, ρ) behaves well for convex cocompact j. Proposition 1.5. The map (j, ρ) 7→ C(j, ρ) is continuous on the set of pairs (j, ρ) ∈ Hom(Γ0 , G)2 with j convex cocompact. Here Hom(Γ0 , G) is endowed with the natural topology (see Section 6). For j geometrically finite but not convex cocompact, the constant C(j, ρ) behaves in a more chaotic way. For n ≤ 3, we prove that continuity holds when C(j, ρ) > 1 and that the condition C(j, ρ) < 1 is open on the set of pairs (j, ρ) with j geometrically finite of fixed cusp type and ρ cuspdeteriorating (Proposition 6.1). However, semicontinuity (both upper and lower) fails when C(j, ρ) ≤ 1 (see Sections 10.6 and 10.7). For n ≥ 4, the condition C(j, ρ) < 1 is not open and upper semicontinuity fails for any value of C(j, ρ) (see Sections 10.10 and 10.11). This is related to the fact that geometrical finiteness itself is not an open condition when n ≥ 4, even under fixed cusp type. We refer to Section 6 for a more thorough discussion. It is natural to hope that when the function (j, ρ) 7→ C(j, ρ) is continuous the map (j, ρ) 7→ E(j, ρ) should be at least upper semicontinuous with respect to the Hausdorff topology. We prove this semicontinuity in dimension n = 2 when C(j, ρ) > 1 and ρ(Γ0 ) does not have a unique fixed point at infinity (Proposition 9.5), generalizing a result of Thurston [T2]. 1.3. Extension of Lipschitz maps in Hn . In order to prove Theorem 1.3, following the approach of [Ka1], we develop the extension theory of Lipschitz maps in Hn and, more precisely, refine an old theorem of Kirszbraun [Kir] and Valentine [V], which states that any Lipschitz map from a compact subset of Hn to Hn with Lipschitz constant ≥ 1 can be extended to a map from Hn to itself with the same Lipschitz constant. We prove the following. Theorem 1.6. Let Γ0 be a discrete group and (j, ρ) ∈ Hom(Γ0 , G)2 a pair of representations of Γ0 in G with j geometrically finite. (1) For any j(Γ0 )-invariant subset K 6= ∅ of Hn and any (j, ρ)-equivariant map ϕ : K → Hn with Lipschitz constant C0 ≥ 1, there exists a (j, ρ)equivariant extension f : Hn → Hn of ϕ with Lipschitz constant C0 . (2) For any j(Γ0 )-invariant subset K 6= ∅ of Hn whose image in j(Γ0 )\Hn is bounded and for any (j, ρ)-equivariant map ϕ : K → Hn with Lipschitz constant C0 < 1, there exists a (j, ρ)-equivariant extension f : Hn → Hn of ϕ with Lipschitz constant < 1. The point of Theorem 1.6 is that we can extend ϕ in an equivariant way, without increasing the Lipschitz constant C0 if it is ≥ 1, and still keeping it MAXIMALLY STRETCHED LAMINATIONS 5 < 1 if it was originally < 1. Moreover, we control the local Lipschitz constant when C0 ≥ 1 (Theorem 5.1). Intuitively (at least when C0 > 1), the idea is that one should be able to choose an f whose stretch locus consists of stretch segments with endpoints in K, moved apart by a factor C0 under ϕ. We believe (see Appendix C.1) that in Theorem 1.6.(2) the best Lipschitz constant of an equivariant extension f could be bounded away from 1 in terms of C0 alone. This would allow to remove the assumption that K has a bounded image in j(Γ0 )\Hn , using the Arzelà–Ascoli theorem (see Section 5.4). In Theorem 5.1 below we refine Theorem 1.6 and actually allow K to be the empty set. (In this case we define C0 to be the supremum of ratios λ(ρ(γ))/λ(j(γ)) for γ ∈ Γ0 with j(γ) hyperbolic, where (1.2) λ(g) := infn d(x, g · x) x∈H Hn is the translation length of g in if g ∈ G is hyperbolic, and 0 if g is parabolic or elliptic.) Theorem 1.3 is equivalent to the case K = ∅ in Theorem 5.1. Theorem 1.6 and its refinements such as Theorem 5.1 should be compared to a number of recent results in the theory of extension of Lipschitz maps: see Lang–Schröder [LS], Lang–Pavlović–Schröder [LPS], Buyalo–Schröder [BuS], Lee–Naor [LN], etc. We also point to [DGK] for an infinitesimal version. 1.4. An application to the study of complete manifolds locally modeled on G = PO(n, 1). One important motivation for examining equivariant Lipschitz maps of minimal Lipschitz constant is the link with certain manifolds locally modeled on G, namely quotients of G by discrete subgroups of G × G acting properly discontinuously and freely on G by right and left multiplication: (g1 , g2 ) · g = g2 gg1−1 . This link was first noticed in [Sa], then developed in [Ka1]. For n = 2, the manifolds locally modeled on PO(2, 1)0 ∼ = PSL2 (R) are the anti-de Sitter 3-manifolds, or Lorentzian 3-manifolds of constant negative curvature, which are Lorentzian analogues of the hyperbolic 3-manifolds. For n = 3, the manifolds locally modeled on PO(3, 1)0 ∼ = PSL2 (C) are the 3-dimensional complex holomorphic-Riemannian manifolds of constant nonzero curvature, which can be considered as complex analogues of the hyperbolic 3-manifolds (see [DZ] for details). For n = 2, all compact manifolds locally modeled on G are quotients of G by discrete subgroups of G × G, up to a finite covering [Kl, KR]; for n = 3, a similar property has been conjectured in [DZ] (see Section 7.8). Recall that the quotient of G by a discrete group Γ is Hausdorff (resp. is a manifold) if and only if the action of Γ on G is properly discontinuous (resp. properly discontinuous and free). Let Γ be a discrete subgroup of G × G acting properly discontinuously on G by right and left multiplication. The key point here is that if Γ is torsion-free, then it is a graph of the form (1.3) Γj,ρ 0 = {(j(γ), ρ(γ)) | γ ∈ Γ0 } where Γ0 is a discrete group and j, ρ ∈ Hom(Γ0 , G) are representations with j injective and discrete (up to switching the two factors): this was proved in [KR] for n = 2, and in [Ka2] (strengthening partial results of [Ko2]) for general rank-one groups G. The group Γ is thus isomorphic to the fundamental 6 FRANÇOIS GUÉRITAUD AND FANNY KASSEL group of the hyperbolic n-manifold M := j(Γ0 )\Hn , and the quotient of G by Γ = Γj,ρ 0 is compact if and only if M is compact (by a classical cohomological argument, see Section 7.7). In general, if Γ is finitely generated, the Selberg lemma [Se, Lem. 8] ensures the existence of a finite-index subgroup ρ,j of Γ that is torsion-free, hence of the form Γj,ρ 0 or Γ0 as above. As before, we set λ(g) := inf x∈Hn d(x, g · x) for any g ∈ G. The following terminology is partly adopted from Salein [Sa]. Definition 1.7. A pair (j, ρ) ∈ Hom(Γ0 , G)2 is called admissible if the action of Γj,ρ 0 on G by right and left multiplication is properly discontinuous. It is called left (resp. right ) admissible if, in addition, there exists γ ∈ Γ0 such that λ(j(γ)) > λ(ρ(γ)) (resp. λ(j(γ)) < λ(ρ(γ))). By [Sa] (for n = 2) and [Ka2] (for general n), an admissible pair (j, ρ) is either left admissible or right admissible; it cannot be both. Without loss of generality, we may restrict to left admissible pairs. For a pair (j, ρ) ∈ Hom(Γ0 , G)2 with j injective and discrete, we set (1.4) C ′ (j, ρ) := sup γ∈Γ0 with j(γ) hyperbolic λ(ρ(γ)) λ(j(γ)) if the group j(Γ0 ) contains hyperbolic elements, and C ′ (j, ρ) := C(j, ρ) otherwise (case of an elementary group fixing a point in Hn or a unique point in ∂∞ Hn ). With this notation, a consequence of Theorem 1.3 is the following (double) left admissibility criterion, which was first established in [Ka1, Chap. 5] for n = 2 and convex cocompact j. Theorem 1.8. Let Γ0 be a discrete group. A pair (j, ρ) ∈ Hom(Γ0 , G)2 with j geometrically finite is left admissible if and only if (1) the infimum C(j, ρ) of Lipschitz constants of (j, ρ)-equivariant Lipschitz maps f : Hn → Hn is < 1. This is equivalent to the condition that (2) the supremum C ′ (j, ρ) of ratios of translation lengths λ(ρ(γ))/λ(j(γ)) for γ ∈ Γ0 with j(γ) hyperbolic is < 1, except possibly in the degenerate case where ρ(Γ0 ) has a unique fixed point in ∂∞ Hn and ρ is not cusp-deteriorating. In particular, left admissibility is always equivalent to (1) and to (2) if j is convex cocompact. In other words, Theorem 1.8 states that (j, ρ) is left admissible if and only if “ρ is uniformly contracting with respect to j”; this uniform contraction can be expressed in two equivalent ways: in terms of Lipschitz maps (condition (1)) and in terms of ratios of lengths (condition (2)). Note that the inequality C ′ (j, ρ) ≤ C(j, ρ) is always true (see (4.1)). It can occur quite generically that C ′ (j, ρ) < C(j, ρ) below 1, even when j and ρ are both convex cocompact (see Sections 10.4 and 10.5). In the degenerate case where ρ(Γ0 ) has a unique fixed point in ∂∞ Hn and ρ is not cusp-deteriorating, it can also happen that C ′ (j, ρ) < C(j, ρ) = 1 (see Section 10.9). However, when we are not in this degenerate case, it follows from Theorem 1.3 that C(j, ρ) ≥ 1 implies C ′ (j, ρ) = C(j, ρ) (Corollary 1.12); in particular, C ′ (j, ρ) < 1 implies C(j, ρ) < 1. MAXIMALLY STRETCHED LAMINATIONS 7 In Theorem 1.8, the fact that if C(j, ρ) < 1 then (j, ρ) is left admissible easily follows from the general properness criterion of Benoist [Be1] and Kobayashi [Ko3] (see Section 7.3); this was first observed by Salein [Sa]. Conversely, suppose that (j, ρ) is left admissible. Then C ′ (j, ρ) ≤ 1 (because (j, ρ) cannot be simultaneously left and right admissible, as mentioned above); the point is to prove that C ′ (j, ρ) = 1 is impossible. This is done in Section 7.5, following the strategy of [Ka1]: we use Theorem 1.3 to establish that C ′ (j, ρ) = 1 implies, not only that C(j, ρ) = 1 (Corollary 1.12), but also that the stretch locus E(j, ρ) contains a geodesic line of Hn ; it is then easy to find a sequence of elements of Γ0 contradicting proper discontinuity by following this geodesic line. We note that in Theorem 1.8 it is necessary for Γ0 to be finitely generated: indeed, for infinitely generated Γ0 there exist left admissible pairs (j, ρ) ∈ Hom(Γ0 , G)2 of injective and discrete representations that satisfy C(j, ρ) = C ′ (j, ρ) = 1 (see Section 10.1). It would be interesting to know whether Theorem 1.8 still holds for finitely generated but geometrically infinite j (Appendix C.2). Here is a consequence of Proposition 1.5 and Theorem 1.8. Theorem 1.9. Let G = PO(n, 1) and let Γ be a discrete subgroup of G × G acting properly discontinuously, freely, and cocompactly on G by right and left multiplication. There is a neighborhood U ⊂ Hom(Γ, G × G) of the natural inclusion such that for any ϕ ∈ U , the group ϕ(Γ) is discrete in G × G and acts properly discontinuously, freely, and cocompactly on G. A particular case of Theorem 1.9 was proved by Kobayashi [Ko4], namely the so-called “special standard” case (terminology of [Z]) where Γ is contained in G × {1}; for n = 3, this was initially proved by Ghys [Gh]. The general case for n = 2 follows from the completeness of compact anti-de Sitter manifolds, due to Klingler [Kl], and from the Ehresmann–Thurston principle on the deformation of holonomies of (G, X)-structures on compact manifolds. An interpretation of Theorem 1.9 in terms of (G, X)-structures is given in Section 7.8. We extend Theorem 1.9 to proper actions on G that are not necessarily cocompact, using the following terminology. Definition 1.10. We say that a quotient of G by a discrete subgroup Γ of G × G is convex cocompact (resp. geometrically finite) if, up to switching the two factors and to passing to a finite-index subgroup, Γ is of the form Γj,ρ 0 as in (1.3) with j convex cocompact (resp. geometrically finite) and (j, ρ) left admissible. This terminology is justified by the fact that convex cocompact (resp. geometrically finite) quotients of PO(n, 1) fiber, with compact fiber O(n), over convex cocompact (resp. geometrically finite) hyperbolic manifolds, up to a finite covering (see Proposition 7.2 or [DGK, Th. 1.2]). We prove the following extension of Theorem 1.9 (see also [Ka3] for a p-adic analogue). Theorem 1.11. Let G = PO(n, 1) and let Γ be a discrete subgroup of G × G acting properly discontinuously on G, with a convex cocompact quotient. There is a neighborhood U ⊂ Hom(Γ, G × G) of the natural inclusion 8 FRANÇOIS GUÉRITAUD AND FANNY KASSEL such that for any ϕ ∈ U , the group ϕ(Γ) is discrete in G × G and acts properly discontinuously on G, with a convex cocompact quotient. Moreover, if the quotient of G by Γ is compact, then so is the quotient of G by ϕ(Γ) for ϕ ∈ U . If the quotient of G by Γ is a manifold, then so is the quotient of G by ϕ(Γ) for ϕ ∈ U close enough to the natural inclusion. Note that Theorem 1.11 is not true if we replace “convex cocompact” with “geometrically finite”: for a given j with cusps, the constant representation ρ = 1 may have small deformations ρ′ that are not cusp-deteriorating, hence for which (j, ρ′ ) cannot be admissible. However, we prove that Theorem 1.11 is true in dimension n = 2 or 3 if we restrict to deformations of Γ of the form Γj,ρ 0 with geometrically finite j and cusp-deteriorating ρ (Theorem 7.7); a similar statement is not true for n > 3 (see Section 10.11). Theorem 1.8 implies that any geometrically finite quotient of G is sharp in the sense of [KK]; moreover, by Theorem 1.11, if the quotient is convex cocompact, then it remains sharp after any small deformation of the discrete group Γ inside G × G (see Section 7.7). This implies the existence of an infinite discrete spectrum for the (pseudo-Riemannian) Laplacian on any geometrically finite quotient of G: see [KK]. 1.5. A generalization of Thurston’s asymmetric metric on Teichmüller space. Let S be a hyperbolic surface of finite volume. The Teichmüller space T (S) of S can be defined as the space of conjugacy classes of geometrically finite representations of Γ0 := π1 (S) into PO(2, 1) ∼ = PGL2 (R) corresponding to finite-volume hyperbolic surfaces homeomorphic to S. Thurston [T2] proved that C(j, ρ) = C ′ (j, ρ) ≥ 1 for all j, ρ ∈ T (S); the function dTh := log C = log C ′ : T (S) × T (S) −→ R+ is the Thurston metric on T (S), which was introduced and extensively studied in [T2]. It is an “asymmetric metric”, in the sense that dTh (j, ρ) ≥ 0 for all j, ρ ∈ T (S), that dTh (j, ρ) = 0 if and only if j = ρ in T (S), that dTh (j1 , j3 ) ≤ dTh (j1 , j2 ) + dTh (j2 , j3 ) for all ji ∈ T (S), but that in general dTh (j, ρ) 6= dTh (ρ, j). We generalize Thurston’s result that C(j, ρ) = C ′ (j, ρ) to any dimension n ≥ 2, to geometrically finite representations j that are not necessarily of finite covolume, and to representations ρ that are not necessarily injective or discrete: as a consequence of Theorem 1.3, we obtain the following. Corollary 1.12. For G = PO(n, 1), let (j, ρ) ∈ Hom(Γ0 , G)2 be a pair of representations with j geometrically finite. If C(j, ρ) ≥ 1, then (1.5) C(j, ρ) = C ′ (j, ρ), except possibly in the degenerate case where C(j, ρ) = 1, where ρ(Γ0 ) has a unique fixed point in ∂∞ Hn , and where ρ is not cusp-deteriorating. In particular, C(j, ρ) ≥ 1 always implies (1.5) if j is convex cocompact, and C ′ (j, ρ) ≥ 1 always implies (1.5) without any assumption on j. In order to generalize the Thurston metric, we consider a nonelementary hyperbolic manifold M of any dimension n ≥ 2 and let T (M ) be the set of conjugacy classes of geometrically finite representations of Γ0 := π1 (M ) into G = PO(n, 1) with the homeomorphism type and cusp type of M . We allow MAXIMALLY STRETCHED LAMINATIONS 9 M to have infinite volume, otherwise T (M ) is trivial for n > 2 by Mostow rigidity. We set   δ(ρ) (1.6) dTh (j, ρ) := log C(j, ρ) δ(j) for all j, ρ ∈ T (M ), where δ : T (M ) → (0, n − 1] is the critical exponent (see (8.1)) giving the exponential growth rate of orbits in Hn or, equivalently in this setting, the Hausdorff dimension of the limit set [S1, S2]. It easily follows from the definition of δ that dTh (j, ρ) ≥ 0 for all j, ρ ∈ T (M ) (Lemma 8.3). Let T (M )Zs be the subset of T (M ) consisting of elements j such that the Zariski closure of j(Γ0 ) in G is simple (for instance equal to G itself). We prove the following. Proposition 1.13. The function dTh defines an asymmetric metric on T (M )Zs . The point of Proposition 1.13 is that dTh (j, ρ) = 0 implies j = ρ on T (M )Zs . On the other hand, for convex cocompact M , it follows from work of Burger [Bu, Th. 1] (see also [BCLS]) that   δ(ρ) ′ ′ dTh (j, ρ) := log C (j, ρ) δ(j) defines an asymmetric metric on T (M )Zs . Kim [Kim, Cor. 3] also proved that if log C ′ (j, ρ) = 0 and δ(j) = δ(ρ), then j = ρ in T (M )Zs . By Corollary 1.12, the asymmetric metrics dTh and d′Th are equal on the set   (j, ρ) ∈ T (M )2Zs | C(j, ρ) ≥ 1 = (j, ρ) ∈ T (M )2Zs | C ′ (j, ρ) ≥ 1 . However, they differ in general (see Sections 10.4 and 10.5). It would be interesting to compare them. In dimension n ≤ 3 the asymmetric metric dTh is always continuous, and in dimension n ≥ 4 it is continuous when M is convex cocompact (Lemma 8.1). 1.6. Organization of the paper. Section 2 contains reminders and basic facts on geometrical finiteness, Lipschitz maps, and convex interpolation in the real hyperbolic space Hn . In Section 3 we recall the classical Kirszbraun– Valentine theorem and establish an equivariant version of it for amenable groups. We then derive general properties of the stretch locus in Section 4. In Section 5 we prove an optimized, equivariant Kirszbraun–Valentine theorem for geometrically finite representations of discrete groups; this yields in particular Theorems 1.3 and 1.6, as well as Corollary 1.12. In Section 6 we examine the continuity properties of the minimal Lipschitz constant C(j, ρ); in particular, we prove Proposition 1.5. In Section 7 we apply the theory to properly discontinuous actions on G = PO(n, 1) (proving Theorems 1.8, 1.9, and 1.11), and in Section 8 we generalize Thurston’s asymmetric metric on Teichmüller space (proving Proposition 1.13). In Section 9 we focus on the case n = 2: we recover and extend results of Thurston for C(j, ρ) > 1, and discuss the nature of the stretch locus for C(j, ρ) < 1. Finally, in Section 10 we give a number of examples and counterexamples designed to make the theory more concrete while pointing out some subtleties. We collect useful 10 FRANÇOIS GUÉRITAUD AND FANNY KASSEL formulas in Appendix A, technical facts on geometrically finite representations in Appendix B, and open questions in Appendix C. Note. We have tried, inside each section, to clearly separate the arguments needed for the convex cocompact case from the ones specific to the cusps. Skipping the latter should decrease the length of the paper substantially. Acknowledgements. We are grateful to Maxime Wolff for his comments on a preliminary version of this paper, to Jeff Danciger for numerous discussions on related topics, and to Samuel Tapie for his indications on the Bowen–Margulis–Sullivan measure. We are indebted to Marc Burger for the idea of introducing the correcting factor δ(ρ)/δ(j) in the definition of dTh (in an earlier version of this paper we had to restrict to the case δ(j) = δ(ρ)). We would like to thank him for interesting discussions on the two possible generalizations dTh and d′Th of the Thurston metric. We are grateful to an anonymous referee for carefully reading the manuscript and making many valuable suggestions. Finally, we thank the University of Chicago for its support and the Institut CNRS-Pauli (UMI 2842) in Vienna for its hospitality. 2. Preliminary results In this section we recall a few well-known facts and definitions on geometrically finite hyperbolic orbifolds, on Lipschitz constants, and on barycenters in the real hyperbolic space Hn . We also expand on the notion of cuspdeterioration introduced in Definition 1.1. In the whole paper, G is the full group PO(n, 1) = O(n, 1)/{±1} of isometries of Hn . If n is even, then G identifies with SO(n, 1). 2.1. Geometrical finiteness. Let j ∈ Hom(Γ0 , G) be an injective representation of a discrete group Γ0 , with j(Γ0 ) discrete. The quotient M := j(Γ0 )\Hn is a smooth, n-dimensional orbifold; it is a manifold if and only if Γ0 is torsion-free. The convex core of M is the smallest closed convex subset of M containing all closed geodesics; its preimage in Hn is the convex hull of the limit set Λj(Γ0 ) ⊂ ∂∞ Hn of j(Γ0 ). (The convex hull is empty only in the degenerate case where the group j(Γ0 ) has a fixed point in Hn or a unique fixed point in ∂∞ Hn ; we do not exclude this case.) Following [Bo1], we will say that the injective and discrete representation j is geometrically finite if Γ0 is finitely generated and if for any ε > 0, the ε-neighborhood of the convex core of M has finite volume. In dimension n = 2, any injective and discrete representation in G of a finitely generated group is geometrically finite. In general, j is geometrically finite if and only if the convex core of M is contained in the union of a compact set and of finitely many disjoint cusps, whose boundaries have compact intersection with the convex core. We now explain what we mean by cusp, following [Bo1]. Let B be a horoball of Hn , centered at a point ξ ∈ ∂∞ Hn , and let S ⊂ Γ0 be the stabilizer of B under j. The group j(S) is discrete (possibly trivial) and consists of nonhyperbolic elements. It preserves the horosphere ∂B ≃ Rn−1 and acts on it by affine Euclidean isometries. By the first Bieberbach theorem (see [Bo1, Th. 2.2.5]), there is a finite-index normal subgroup S ′ of S that is isomorphic to Zm for some 0 ≤ m < n, and whose index in S is bounded by some ν(n) ∈ N depending only on the dimension n; we have m ≥ 1 if MAXIMALLY STRETCHED LAMINATIONS 11 and only if S contains a parabolic element. The group j(S) preserves and acts cocompactly on some m-dimensional affine subspace V of ∂B ≃ Rn−1 , unique up to translation; the subgroup j(S ′ ) acts on V by translation. Let HV be the closed (m + 1)-dimensional hyperbolic subspace of Hn containing ξ in its boundary such that HV ∩ ∂B = V, and let π : Hn → HV be the closest-point projection (see Figure 1). The group j(S) preserves the convex set C := π −1 (HV ∩ B) ⊂ Hn . Following [Bo1], we say that the image of C in M is a cusp if m ≥ 1 and C ∩ j(γ) · C = ∅ for all γ ∈ Γ0 r S. The cusp is then isometric to j(S)\C; its intersection with the convex core of M is contained in j(S)\B ′ for some horoball B ′ ⊃ B. The integer m is called the rank of the cusp. (ξ = ∞) j(S ′ ) HV B ∂B V C ∂∞ H3 Figure 1. A rank-one cusp centered at ξ = ∞ in the upper half-space model of H3 . The limit set is contained in {ξ} ∪ ∂∞ (H3 r C). When the convex core of M is nonempty, we may assume that it contains the image of V, after possibly replacing B by some smaller horoball and V by some translate. We shall use the following description. Fact 2.1. If j is geometrically finite, then M = j(Γ0 )\Hn is the union of a closed subset M ′ and of finitely many disjoint quotients j(Si )\Bi , where Bi is a horoball of Hn and j(Si ) a discrete group of isometries of Bi containing a parabolic element, such that • the intersection of M ′ with the convex core of M is compact; • for any i we have M ′ ∩ (j(Si )\Bi ) = j(Si )\∂Bi ; in particular, the intersection of j(Si )\∂Bi with the convex core of M is compact; • for any i the intersection in Hn of Bi with the preimage N of the convex core of M is the convex hull of ∂Bi ∩ N , and j(Si )\(∂Bi ∩ N ) is compact. Definition 2.2. We shall call the intersections of the sets j(Si )\Bi with the convex core of M standard cusp regions. 12 FRANÇOIS GUÉRITAUD AND FANNY KASSEL If j is geometrically finite, then the complement of the convex core of M has finitely many connected components, called the funnels of M . By definition, j is convex cocompact if it is geometrically finite with no cusp; when Γ0 is infinite, this is equivalent to the convex core being nonempty and compact. The set of convex cocompact representations is open in Hom(Γ0 , G) (see [Bo2, Prop. 4.1] or Proposition B.1). In Sections 4 and 5, we shall consider a j(Γ0 )-invariant subset K of Hn whose image in M is compact. We shall then use the following notation. Notation 2.3. In the rest of the paper, Conv(K) ⊂ Hn denotes: • the convex hull of K if K is nonempty, • the preimage in Hn of the convex core of M = j(Γ0 )\Hn if K is empty and the convex core is nonempty (leaving j implicit), • any nonempty j(Γ0 )-invariant convex subset of Hn if K and the convex core of M = j(Γ0 )\Hn are both empty (case when j(Γ0 ) is an elementary group fixing a point in Hn or a unique point in ∂∞ Hn ). In all three cases the set Conv(K) is nonempty and contains the preimage in Hn of the convex core of M . In Fact 2.1, we can take M ′ and the Bi with the following properties: • M ′ contains the (compact) image of K in M ; • the intersection of M ′ with the image of Conv(K) in M is compact; • for any i the set Bi ∩ Conv(K) is the convex hull of ∂Bi ∩ Conv(K), and j(Si )\(∂Bi ∩ Conv(K)) is compact. 2.2. Cusp deterioration. Let j ∈ Hom(Γ0 , G) be a geometrically finite representation and let B1 , . . . , Bc be horoballs of Hn whose projections j(Si )\Bi to M = j(Γ0 )\Hn are disjoint and intersect the convex core in standard cusp regions representing all the cusps, as in Section 2.1. Consider ρ ∈ Hom(Γ0 , G). Definition 2.4. For 1 ≤ i ≤ c, we say that ρ is deteriorating in Bi if ρ(Si ) contains only elliptic elements. Thus ρ is cusp-deteriorating in the sense of Definition 1.1 if and only if it is deteriorating in Bi for all 1 ≤ i ≤ c. Depending on whether ρ is deteriorating in Bi or not, we shall use the following classical fact with Γ′ = ρ(Si ). Fact 2.5 (see [Par, Th. III.3.1]). Let Γ′ be a finitely generated subgroup of G. (1) If all elements of Γ′ are elliptic, then Γ′ has a fixed point in Hn . (2) If all elements of Γ′ are elliptic or parabolic and if Γ′ contains at least one parabolic element, then Γ′ has a unique fixed point in the boundary at infinity ∂∞ Hn of Hn . Lemma 2.6. Let Γ′ be as in Fact 2.5.(2) and let wl : Γ′ → N be the word length function with respect to some fixed finite generating subset F ′ of Γ′ . Fix p ∈ Hn . • There exists R > 0 such that for all γ ′ ∈ Γ′ ,  d(p, γ ′ · p) ≤ 2 log 1 + wl(γ ′ ) + R. • If Γ′ is discrete in G, then there exists R′ > 0 such that for all γ ′ ∈ Γ′ ,  d(p, γ ′ · p) ≥ 2 log 1 + wl(γ ′ ) − R′ . MAXIMALLY STRETCHED LAMINATIONS 13 Proof. Let ξ ∈ ∂∞ Hn be the fixed point of Γ′ and let ∂B be the horosphere through p centered at ξ. For any q, q ′ ∈ ∂B, let d∂B (q, q ′ ) be the length of the shortest path from q to q ′ that is contained in ∂B. Then d∂B is a Euclidean metric on ∂B ≃ Rn−1 and   d∂B (q, q ′ ) ′ (2.1) d(q, q ) = 2 arcsinh 2 for all q, q ′ ∈ ∂B (see (A.3)). In particular, |d − 2 log(1 + d∂B )| is bounded on ∂B × ∂B. By the triangle inequality,   ′ d (p, f · p) · wl(γ ′ ) d∂B (p, γ ′ · p) ≤ max ∂B ′ ′ f ∈F γ′ Γ′ , for all ∈ which implies the first statement of the lemma. If Γ′ is discrete in G, then it acts properly discontinuously on ∂B and has a finite-index subgroup isomorphic to Zm (for some 0 < m < n), acting as a lattice of translations on some m-dimensional affine subspace V of the Euclidean space ∂B ≃ Rn−1 (see Section 2.1). In a Euclidean lattice, the norm of a vector is estimated, up to a bounded multiplicative factor, by its word length in any given finite generating set: therefore there exist c, Q > 0 such that d∂B (p, γ ′ · p) ≥ c wl(γ ′ ) − Q for all γ ′ ∈ Γ′ . The second statement of the lemma follows by using (2.1) and the properness of the function wl on Γ′ .  Here is a consequence of Lemma 2.6, explaining why the notion of cuspdeterioration naturally appears in our setting. Lemma 2.7. Let ρ ∈ Hom(Γ0 , G). If there exists a (j, ρ)-equivariant map f : Hn → Hn with Lipschitz constant < 1, then ρ is cusp-deteriorating with respect to j. Proof. Let f : Hn → Hn be a (j, ρ)-equivariant map. Suppose that ρ is not cusp-deteriorating. Then there is an element γ ∈ Γ0 such that j(γ) is parabolic and ρ(γ) is either parabolic or hyperbolic. Fix a point p ∈ Hn . By Lemma 2.6, we have d(p, j(γ k ) · p) ∼ 2 log k as k → +∞. If ρ(γ) is parabolic, then similarly d(f (p), ρ(γ k ) · f (p)) ∼ 2 log k, and if ρ(γ) is hyperbolic, then |d(f (p), ρ(γ k ) · f (p)) − k λ(ρ(γ))| is uniformly bounded (for instance by twice the distance from f (p) to the translation axis of ρ(γ) in Hn ). In both cases, we see that  d f (p), ρ(γ k ) · f (p) ≥ 1, lim sup d(p, j(γ k ) · p) k→+∞ hence the (j, ρ)-equivariant map f cannot have Lipschitz constant < 1.  2.3. Lipschitz constants. For any subset X of Hn and any map f from X to some metric space (Z, dZ ) (in practice, Hn or R), we denote by Lip(f ) = sup x,x′ ∈X, x6=x′ dZ (f (x), f (x′ )) d(x, x′ ) the Lipschitz constant of f . For any Y ⊂ X and any x ∈ X, we set LipY (f ) = Lip(f |Y ), Lipx (f ) = inf LipBx (r) (f ), r>0 14 FRANÇOIS GUÉRITAUD AND FANNY KASSEL where Bx (r) is the closed ball of radius r centered at x in Hn . We call Lipx (f ) the local Lipschitz constant of f at x. Remarks 2.8. (1) Let f be a C-Lipschitz map from a geodesic segment [x, x′ ] of Hn to Hn . If d(f (x), f (x′ )) = Cd(x, x′ ), then f “stretches maximally” [x, x′ ], in the sense that d(f (y), f (y ′ )) = Cd(y, y ′ ) for all y, y ′ ∈ [x, x′ ]. (2) Let X be a convex subset of Hn , covered by a collection of open sets Ut , t ∈ T . For any map f : X → Hn , Lip(f ) = sup LipX∩Ut (f ). t∈T (3) For any rectifiable path C in some subset X of Hn and for any map f : X → Hn , length(f (C )) ≤ sup Lipx (f ) · length(C ). x∈C Indeed, (1) follows from the fact that if the points x, y, y ′ , x′ lie in this order, then d(x, x′ ) = d(x, y) + d(y, y ′ ) + d(y ′ , x′ ) while d(f (x), f (x′ )) ≤ d(f (x), f (y)) + d(f (y), f (y ′ )) + d(f (y ′ ), f (x′ )) by the triangle inequality. To prove (2), we just need to check that the right-hand side is an upper bound for Lip(f ) (it is also clearly a lower bound). Any geodesic segment [p, q] ⊂ X can be divided into finitely many subsegments, each contained in one of the open sets Ut ; we use again the additivity of distances at the source and the subadditivity of distances at the target. Finally, (3) follows from the definition of the length of a path (obtained by summing up the distances between points of smaller and smaller subdivisions and taking a limit) and from the definition of the local Lipschitz constant. Lemma 2.9. The local Lipschitz constant function x 7→ Lipx (f ) is upper semicontinuous: for any converging sequence xk → x, Lipx (f ) ≥ lim sup Lipxk (f ). k→+∞ In particular, for any compact subset K of X, the supremum of Lipx (f ) for x ∈ K is achieved on some nonempty closed subset of K. Moreover, if X is convex, then (2.2) Lip(f ) = sup Lipx (f ). x∈X Proof. Upper semicontinuity follows from an easy diagonal extraction argument. The inequality Lip(f ) ≥ supx∈X Lipx (f ) is clear. The converse inequality for convex X follows from Remark 2.8.(3) where C is any geodesic segment [p, q] ⊂ X.  Note that the convexity of X is required for (2.2) to hold: for instance, an arclength-preserving map taking a horocycle X to a straight line is not even Lipschitz, although its local Lipschitz constant is everywhere 1. As a consequence of Lemma 2.9, the stretch locus of any Lipschitz map f : X → Hn is closed in X for the induced topology. Here we use the following terminology, which agrees with Definition 1.2. Definition 2.10. For any subset X of Hn and any Lipschitz map f : X → Hn , the stretch locus Ef of f is the set of points x ∈ X such that MAXIMALLY STRETCHED LAMINATIONS 15 ef of f is the union of Lipx (f ) = Lip(f ). The enhanced stretch locus E {(x, x) ∈ X × X | x ∈ Ef } and  (x, x′ ) ∈ X ×X | x 6= x′ and d(f (x), f (x′ )) = Lip(f ) d(x, x′ ) . ef to X are equal to Ef , but By Remark 2.8.(1), both projections of E ef records a little extra information, namely the positions of the maximally E stretched segments between points of the stretch locus Ef . 2.4. Barycenters in Hn . For any index set I equal to {1, 2, . . . , k} for some k ≥ 1 or to N∗ , and for any tuple α = (αi )i∈I of nonnegative reals summing up to 1, we set   X n I 2 n I (H )α := (pi ) ∈ (H ) αi d(p1 , pi ) < +∞ . i∈I This set contains at least all bounded sequences (pi ) ∈ (Hn )I , and it is just the direct product (Hn )k if k < +∞. The following result is classical, and actually holds in any CAT(0) space. Lemma 2.11. For any index set I equal to {1, 2, . . . , k} for some k ≥ 1 or to N∗ and for any tuple α = (αi )i∈I of nonnegative reals summing up to 1, the map mα : (Hn )Iα −→ Hn P taking (pi )i∈I to the minimizer of i∈I αi d( · , pi )2 is well defined and αi Lipschitz in its i-th entry: for any (pi ), (qi ) ∈ (Hn )Iα ,  X αi d(pi , qi ). (2.3) d mα (p1 , p2 , . . . ), mα (q1 , q2 , . . . ) ≤ i∈I Proof. Fix I and α = (αi )i∈I , and consider an element (pi ) ∈ (Hn )Iα . For any x ∈ Hn , X X 2 Φ(x) := αi d(x, pi )2 ≤ αi d(x, p1 ) + d(p1 , pi ) i∈I i∈I ≤ 2 Hn X i∈I  αi d(x, p1 )2 + d(p1 , pi )2 < +∞. The function Φ : → R thus defined is proper on Hn since it is bounded from below by any proper function αi d(·, pi )2 with αi > 0, and it achieves its minimum on the convex hull of the pi . Moreover, Φ is analytic: to see this on any ball B of Hn , note that the unweighted summands d(·, pi )2 for pi in a 1-neighborhood of B are analytic with derivatives (of any nonnegative order) bounded independently of i, while the other summands can be written φ2i + 2φi d(p1 , pi )+d(p1 , pi )2 , where φi := d(·, pi )−d(p1 , pi ) again is analytic on B, and φi and φ2i have their derivatives (of any nonnegative order) bounded independently of i. d2 2 On any unit-speed geodesic (xt )t∈R of Hn , we have dt 2 t=0 d(xt , pi ) ≥ 2. Indeed, let logx0 : Hn → Tx0 Hn be the inverse of the exponential map at x0 . Standard CAT(0) comparison inequalities with the Euclidean metric dEucl yield dEucl (logx0 (xt ), log x0 (pi ))2 ≤ d(xt , pi )2 16 FRANÇOIS GUÉRITAUD AND FANNY KASSEL for all t ∈ R; both sides are equal at t = 0, the first derivatives are equal at t = 0, and the left-hand side has second P derivative ≡ 2. It follows that t 7→ Φ(xt ) has second derivative at least 2 I αi = 2 everywhere. While mα (p1 , p2 , . . . ) is the minimizer of Φ, the point mα (q1 , q2 , . . . ) is the minimizer of Φ + Ψ, where X  −d(x, pi )2 + d(x, qi )2 αi . Ψ(x) := i∈I We claim that ψi : x 7→ −d(x, pi )2 + d(x, qi )2 is 2d(pi , qi )-Lipschitz: indeed, with (xt )t∈R as above, d dt t=0 ψi (xt ) 2d(x0 , pi ) cos p\ i x0 x1 − 2d(x0 , qi ) cos q\ i x0 x1  = 2dEucl πℓ (logx0 pi ), πℓ (logx0 qi ) ≤ 2d(pi , qi ), = where ℓ ⊂ Tx0 Hn is the tangent line to (xt )t∈R at t = 0, and πℓ : Tx0 Hn → ℓ isP the closest-point projection. Therefore, Ψ is Lipschitz with constant L := 2 i∈I αi d(pi , qi ). Thus, for any unit-speed geodesic ray (xt )t≥0 starting d Φ(xt ) > L, hence from x0 = mα (p1 , p2 , . . . ), as soon as t > L/2 we have dt d dt (Φ + Ψ)(xt ) > 0. The minimizer of Φ + Ψ is within L/2 from x0 , as promised.  Note that the map mα is G-equivariant: (2.4) mα (g · p1 , g · p2 , . . . ) = g · mα (p1 , p2 , . . . ) for all g ∈ G and (pi ) ∈ (Hn )Iα . It is also diagonal: (2.5) mα (p, p, . . . ) = p for all p ∈ Hn . If σ is a permutation of I, then (2.6) m(ασ(1) ,ασ(2) ,... ) (pσ(1) , pσ(2) , . . . ) = m(α1 ,α2 ,... ) (p1 , p2 , . . . ) 1 1 for all (pi ) ∈ (Hn )Iα ; in particular, mk := m( k ,..., k ) is symmetric in its k entries. Unlike barycenters in vector spaces however, m has only weak associativity properties: the best one can get is associativity over equal entries, i.e. if p1 = · · · = pk = p then m(α1 ,...,αk+1 ,... ) (p1 , . . . , pk+1 , . . . ) = m(α1 +···+αk ,αk+1 ,... ) (p, pk+1 , . . . ) . P We will often write i∈I αi pi for mα (p1 , p2 , . . . ). While (2.3) controls the displacement of a barycenter under a change of points, the following lemma deals with a change of weights. Lemma 2.12. Let I = {1, 2, . . . , k} for some k ≥ 1 or N∗ and let α = (αi )i∈I and β = (βi )i∈I be two nonnegative sequences, each summing up to 1. Consider points (pi )i∈I of Hn , all within distance R of some p ∈ Hn (in particular (pi ) ∈ (Hn )Iα ∩ (Hn )Iβ ). Then X  d mα (p1 , p2 , . . . ), mβ (p1 , p2 , . . . ) ≤ R |αi − βi |. i∈I Proof. For any i ∈ I, we set δi := αi − βi . The basic observation is that if for example δ1 > 0, then we can transfer δ1 units of weight from p1 to p, at the moderate cost of moving the barycenter by ≤ Rδ1 : by Lemma 2.11, the MAXIMALLY STRETCHED LAMINATIONS 17 point m := m(α1 ,α2 ,α3 ,... ) (p1 , p2 , p3 , . . . ) = m(δ1 ,β1,α2 ,α3 ,... ) (p1 , p1 , p2 , p3 , . . . ) lies at distance ≤ Rδ1 from m(δ1 ,β1 ,α2 ,α3 ,... ) (p, p1 , p2 , p3 , . . . ). Repeating this procedure for all indices i ≥ 1 such that δi > 0, we find that m lies at distance ≤ Rδ from m(δ,min{α1 ,β1 },min{α2 ,β2},min{α3 ,β3 },... ) (p, p1 , p2 , p3 , . . . ), P where we set δ := δi >0 δi . This expression is symmetric in α and β, so m P lies at distance ≤ 2Rδ = R |αi − βi | from m(β1 ,β2,... ) (p1 , p2 , . . . ).  2.5. Barycenters of Lipschitz maps and partitions of unity. Here is an easy consequence of Lemma 2.11. Lemma 2.13. Let I = {1, 2, . . . , k} for some k ≥ 1 or N∗ and let α = (αi )i∈I be a nonnegative sequence summing up to 1. Consider p ∈ X ⊂ Hn and a sequence of Lipschitz maps fi : X → Hn with (fi (p)) ∈ (Hn )Iα and with (Lip(fi ))i∈I bounded. Then the map X f= αi fi : x 7−→ mα (f1 (x), f2 (x), . . . ) i∈I is well defined on X and satisfies X Lipx (f ) ≤ αi Lipx (fi ) and i∈I LipY (f ) ≤ X αi LipY (fi ) i∈I for all x ∈ Y ⊂ X. In particular, if Lip(fi ) = C = Lip(f ) for all i ∈ I, then the (enhanced) stretch locus of f (Definition 2.10) is contained in the intersection of the (enhanced) stretch loci of the maps fi . Proof. We first note that (fi (x)) ∈ (Hn )Iα for any x ∈ Hn . Indeed, using the triangle inequality and the general inequality (a + b + c)2 ≤ 3 (a2 + b2 + c2 ) for a, b, c ≥ 0, we have X 2 αi d f1 (x), fi (x) i∈I ≤ 3 X i∈I  2 2 2  , αi d f1 (x), f1 (p) + d f1 (p), fi (p) + d fi (p), fi (x) which is finite since (fi (p)) ∈ (Hn )Iα and (Lip(fi ))i∈I is bounded. By Lemma 2.11, the map f is well defined and for any x, y ∈ Hn ,  X  d f (x), f (y) ≤ αi d fi (x), fi (y) , i∈I which implies Lemma 2.13.  We also consider barycenters of maps with variable coefficients. The following result, which combines Lemmas 2.12 and 2.13 in an equivariant setting, is one of our main technical tools; it will be used extensively throughout Sections 4 and 6. Lemma 2.14. Let Γ0 be a discrete group, (j, ρ) ∈ Hom(Γ0 , G)2 a pair of representations with j injective and discrete, and B1 , . . . , Br open subsets of Hn . For 1 ≤ i ≤ r, let fi : j(Γ0 ) · Bi → Hn be a (j, ρ)-equivariant map 18 FRANÇOIS GUÉRITAUD AND FANNY KASSEL that is Lipschitz on Bi . For p ∈ Hn , let Ip denote the set of indices 1 ≤ i ≤ r such that p ∈ j(Γ0 ) · Bi , and define Rp := diam{fi (p) | i ∈ Ip } < +∞. For 1 ≤ i ≤ r, let also ψi : Hn → [0, 1] be a j(Γ0 )-invariant Lipschitz map supported in j(Γ0 ) · Bi . Assume that ψS1 , . . . , ψr induce a partition of unity on a j(Γ0 )-invariant open subset B of ri=1 j(Γ0 ) · Bi . Then the map P f = i∈I ψi fi : B −→ P Hn p 7−→ i∈Ip ψi (p)fi (p) is (j, ρ)-equivariant and for any p ∈ B, the following “Leibniz rule” holds: X  (2.7) Lipp (f ) ≤ Lipp (ψi ) Rp + ψi (p) Lipp (fi ) . i∈Ip Proof. The map f is (j, ρ)-equivariant because the barycentric construction is, see (2.4). Fix p ∈ B and ε > 0. By definition of Ip , continuity of ψi and fi , and upper semicontinuity of the local Lipschitz constant (Lemma 2.9), there is a neighborhood U of p in B such that for all x ∈ U , • • • • • ψi |U = 0 for all i ∈ / Ip , ψi (x) ≤ ψi (p) + ε for all i ∈ Ip , Rx ≤ Rp + ε, LipU (ψi ) ≤ Lipp (ψi ) + ε for all i ∈ Ip , LipU (fi ) ≤ Lipp (fi ) + ε for all i ∈ Ip . By the triangle inequality, for any x, y ∈ U ,  X X ψi (y)fi (y) ψi (x)fi (x), d(f (x), f (y)) = d i∈Ip i∈Ip i∈Ip i∈Ip  X X ψi (y)fi (x) ψi (x)fi (x), ≤ d  X X ψi (y)fi (y) . ψi (y) fi (x), +d i∈Ip i∈Ip Using Lemma 2.12, we see that the first term of the right-hand side is bounded by  X X  (Lipp (ψi ) + ε) (Rp + ε) ; LipU (ψi ) d(x, y) Rx ≤ d(x, y) i∈Ip i∈Ip and using Lemma 2.13, that the second term is bounded by X X (ψi (p) + ε)(Lipp (fi ) + ε) . ψi (y)LipU (fi ) ≤ d(x, y) d(x, y) i∈Ip i∈Ip The bound (2.7) follows by letting ε go to 0.  3. An equivariant Kirszbraun–Valentine theorem for amenable groups One of the goals of this paper is to refine the classical Kirszbraun–Valentine theorem [Kir, V], which states that any Lipschitz map from a compact subset of Hn to Hn with Lipschitz constant ≥ 1 can be extended to a map from MAXIMALLY STRETCHED LAMINATIONS 19 Hn to itself with the same Lipschitz constant. We shall in particular extend this theorem to an equivariant setting, for two actions j, ρ ∈ Hom(Γ0 , G) of a discrete group Γ0 on Hn , with j geometrically finite (Theorem 1.6). Before we prove Theorem 1.6, we shall • reprove the classical Kirszbraun–Valentine theorem (Section 3.1), both for the reader’s convenience and because the main technical step (Lemma 3.2) will be useful later to control the local Lipschitz constant; • examine the case when the Lipschitz constant is < 1 (Section 3.2); • extend the classical Kirszbraun–Valentine theorem to an equivariant setting for two actions j, ρ ∈ Hom(S, G) of an amenable group S (Section 3.3). We shall use this as a technical tool to extend maps in cusps when dealing with geometrically finite representations j ∈ Hom(Γ0 , G) that are not convex cocompact. 3.1. The classical Kirszbraun–Valentine theorem. We first give a proof of the classical Kirszbraun–Valentine theorem [Kir, V]. Proposition 3.1. Let K 6= ∅ be a compact subset of Hn . Any Lipschitz map ϕ : K → Hn with Lip(ϕ) ≥ 1 admits an extension f : Hn → Hn with the same Lipschitz constant. The following is an important technical step in the proof of Proposition 3.1. It will also be used in the proofs of Lemmas 3.8, 5.2, and 5.4 below. Lemma 3.2. Let K 6= ∅ be a compact subset of Hn and ϕ : K → Hn a nonconstant Lipschitz map. For any p ∈ Hn r K, the function Hn −→ R+ d(q ′ , ϕ(k)) k∈K d(p, k) n admits a minimum Cq at a point q ∈ H , and q belongs to the convex hull of ϕ(K′ ) where   d(q, ϕ(k)) = Cq . K′ := k ∈ K d(p, k) Moreover, \ • either there exist k1 , k2 ∈ K′ such that 0 ≤ k\ 1 pk2 < ϕ(k1 )qϕ(k2 ) ≤ π, ′ ′ \ • or k\ 1 pk2 = ϕ(k1 )qϕ(k2 ) for (ν × ν)-almost all (k1 , k2 ) ∈ K × K , ′ where ν is some probability measure on K such that q belongs to the convex hull of the support of ϕ∗ ν. q′ 7−→ Cq′ := max c ∈ [0, π] the angle at b between three points a, b, c ∈ Hn . Here we denote by abc Proof of Lemma 3.2. The function q ′ 7→ Cq′ is proper and convex, hence admits a minimum Cq at some point q ∈ Hn . We have Cq > 0 since ϕ is nonconstant. Suppose by contradiction that q does not belong to the convex hull of ϕ(K′ ), and let q ′ be the projection of q to this convex hull. If q ′′ is a ′′ ,ϕ(k)) is point of the geodesic segment [q, q ′ ], close enough to q, then d(qd(p,k) uniformly < Cq for k ∈ K: indeed, for k ∈ K in a small neighborhood of K′ this follows from the inequality d(q ′′ , ϕ(k)) < d(q, ϕ(k)); for k ∈ K away 20 FRANÇOIS GUÉRITAUD AND FANNY KASSEL from K′ it follows from the fact that d(q,ϕ(k)) d(p,k) is itself bounded away from Cq by continuity and compactness of K. Thus Cq′′ < Cq , a contradiction. It follows that q belongs to the convex hull of ϕ(K′ ). Let K′log ⊂ Tp Hn (resp. Llog ⊂ Tq Hn ) be the (compact) set of vectors whose image by the exponential map expp : Tp Hn → Hn (resp. expq : Tq Hn → Hn ) lies in K′ (resp. in ϕ(K′ )). Let ′ ϕlog := exp−1 q ◦ ϕ ◦ expp : Klog −→ Llog be the map induced by ϕ. The fact that q belongs to the convex hull of ϕ(K′ ) implies that 0 belongs to the convex hull of Llog = ϕlog (K′log ). Therefore, there exists a probability measure νlog on K′log such that Z ϕlog (x) dνlog (x) = 0 ∈ Tq Hn . ′ kϕ (x)k Klog log (The division is legitimate since kϕlog (x)k = d(q, ϕ(expq (x))) ≥ Cq d(p, K) > 0 for all x ∈ K′log .) We set ν := (expp )∗ νlog , so that q belongs to the convex hull of the support of ϕ∗ ν. We then have ZZ   \ cos k\ 1 pk2 − cos ϕ(k1 )qϕ(k2 ) d(ν × ν)(k1 , k2 ) ′ ′ Z ZK ×K   ϕlog (x2 ) ϕlog (x1 ) x1 x2 = d(νlog × νlog )(x1 , x2 ) − kx1 k kx2 k kϕ (x1 )k kϕ (x2 )k log K′log ×K′log = Z K′log 2 x dνlog (x) − kxk Z K′log log ϕlog (x) dνlog (x) kϕlog (x)k 2 ≥ 0. \ If the function (k1 , k2 ) 7→ cos k\ 1 pk2 − cos ϕ(k1 )qϕ(k2 ) takes a positive value \ at some pair (k1 , k2 ) ∈ K′ × K′ , then we have 0 ≤ k\ 1 pk2 < ϕ(k1 )qϕ(k2 ) ≤ π. Otherwise, the function takes only nonpositive values, hence is zero (ν × ν)almost everywhere since its integral is nonnegative; in other words, k\ 1 pk2 = ′ ′ \ ϕ(k1 )qϕ(k2 ) for (ν × ν)-almost all (k1 , k2 ) ∈ K × K .  The other main ingredient in the proof of Proposition 3.1 is the following consequence of Toponogov’s theorem, a comparison theorem expressing the divergence of geodesics in negative curvature (see [BH, Lem. II.1.13]). Lemma 3.3. In the setting of Lemma 3.2, we have Cq ≤ max(Lip(ϕ), 1). Proof. We may assume Cq ≥ 1, otherwise there is nothing to prove. By \ Lemma 3.2, there exist k1 , k2 ∈ K′ such that k\ 1 pk2 ≤ ϕ(k1 )qϕ(k2 ) 6= 0. Since d(q, ϕ(k1 )) d(q, ϕ(k2 )) = = Cq ≥ 1, d(p, k1 ) d(p, k2 ) Toponogov’s theorem implies d(ϕ(k1 ), ϕ(k2 )) ≥ Cq d(k1 , k2 ). On the other hand, we have d(ϕ(k1 ), ϕ(k2 )) ≤ Lip(ϕ) d(k1 , k2 ) by definition of Lip(ϕ), hence Cq ≤ Lip(ϕ).  Proof of Proposition 3.1. It is enough to prove that for any point p ∈ Hn rK we can extend ϕ to K ∪ {p} keeping the same Lipschitz constant C0 := Lip(ϕ). Indeed, if this is proved, then we can consider a dense sequence MAXIMALLY STRETCHED LAMINATIONS 21 (pi )i∈N of points of Hn r K, construct by induction a C0 -Lipschitz extension of ϕ to K ∪ {pi | i ∈ N}, and finally extend it to Hn by continuity. Let p ∈ Hn r K. If ϕ is constant, then the constant extension of ϕ to K ∪ {p} still has the same Lipschitz constant. Otherwise we apply Lemmas 3.2 and 3.3 with (K, ϕ) := (K, ϕ).  Remark 3.4. The proof actually shows that for C ≥ 1, any map ϕ : K → Hn with Lip(ϕ) ≤ C admits an extension f : Hn → Hn with Lip(f ) ≤ C. Remark 3.5. The same proof shows that if K is a nonempty compact subset of Rn , then any Lipschitz map ϕ : K → Rn admits an extension f : Rn → Rn with the same Lipschitz constant. There is no constraint on the Lipschitz constant for Rn since the Euclidean analogue of Toponogov’s theorem holds for any C ≥ 0. This Euclidean extension result is the one originally proved by Kirszbraun [Kir], by a different approach based on Helly’s theorem. The hyperbolic version is due to Valentine [V]. Remark 3.6. Proposition 3.1 actually holds for any subset K of Hn , not necessarily compact. Indeed, we can always extend ϕ to the closure K of K by continuity, with the same Lipschitz constant, and view K as an increasing union of compact sets Ki , i ∈ N. Proposition 3.1 gives extensions fi : Hn → Hn of ϕ|Ki with Lip(fi ) ≤ Lip(ϕ), and by the Arzelà–Ascoli theorem we can extract a pointwise limit f from the fi , extending ϕ with Lip(f ) = Lip(ϕ). 3.2. A weaker version when the Lipschitz constant is < 1. Proposition 3.1 does not hold when the Lipschitz constant is < 1: see Example 9.6. However, we prove the following strengthening of Remark 3.4 with C = 1. Proposition 3.7. Let K 6= ∅ be a compact subset of Hn . Any Lipschitz map ϕ : K → Hn with Lip(ϕ) < 1 admits an extension f : Hn → Hn with Lip(f ) < 1. It is not clear whether the analogue of Remark 3.6 holds for Lip(ϕ) < 1: see Appendix C.1. Here is the main technical step in the proof of Proposition 3.7. Lemma 3.8. Let K 6= ∅ be a compact subset of Hn with convex hull Conv(K) in Hn , and let ϕ : K → Hn be a Lipschitz map with Lip(ϕ) < 1. For any p ∈ Conv(K), there is a neighborhood Up of p in Hn and a 1-Lipschitz extension fp : K ∪ Up → Hn of ϕ such that LipUp (fp ) < 1. Proof. We first extend ϕ to a map h : K ∪ {p} → Hn with Lip(h) < 1. For this we may assume p ∈ / K. By Lemma 3.2 with (K, ϕ) := (K, ϕ), we can find points q ∈ Hn and k1 , k2 ∈ K such that Cq := maxk∈K d(q, ϕ(k))/d(p, k) is minimal and such that d(q, ϕ(ki )) = Cq d(p, ki ) for i ∈ {1, 2} and k\ 1 pk2 ≤ \ ϕ(k1 )qϕ(k2 ) 6= 0. We cannot have Cq = 1, otherwise we would have d(ϕ(k1 ), ϕ(k2 )) ≥ d(k1 , k2 ) by basic trigonometry, contradicting Lip(ϕ) < 1. Therefore Cq < 1 by Lemma 3.3. We can then take h : K ∪ {p} → Hn to be the extension of ϕ sending p to q. Next, choose a small ε > 0 such that Lip(h)(1+ ε)4 < 1. Let us prove that there is a ball B of radius r > 0 centered at p and an extension h′ : B → Hn 22 FRANÇOIS GUÉRITAUD AND FANNY KASSEL of h|K∩B such that LipB (h′ ) ≤ Lip(h)(1 + ε)4 < 1. If p ∈ / K, we just ′ take B to be disjoint from K and h (B) = {q}. If p ∈ K, we remark that there is a constant r > 0 such that for any x ∈ Hn , the exponential map expx : Tx Hn → Hn and its inverse logx : Hn → Tx Hn are both (1 + ε)Lipschitz when restricted to the ball Bx (r) ⊂ Hn of radius r centered at x and to its image logx Bx (r) ⊂ Tx Hn . We set B := Bp (r). Consider the map logq ◦ h ◦ expp : logp (K) −→ Tq Hn . Its restriction to logp (K ∩ B) ⊂ Tp Hn is Lip(h)(1 + ε)2 -Lipschitz. By Remark 3.5, this restriction admits an extension ψ : logp B → Tq Hn with the same Lipschitz constant. Then h′ := expq ◦ ψ ◦ logp : B −→ Hn is an extension of h|K∩B with LipB (h′ ) ≤ Lip(h)(1 + ε)4 < 1. Let B ′ ⊂ B be another ball centered at p, of radius r ′ > 0 small enough so that Lip(h)r + Lip(h′ )r ′ < r − r ′ . We claim that we may take Up := B ′ and fp : K ∪ Up → Hn to be the map that coincides with ϕ on K and with h′ on Up . Indeed, for any distinct points (x, y) ∈ K × B ′ , if x ∈ B then Lip{x,y} (fp ) ≤ Lip(h′ ) < 1, and otherwise (3.1) d(fp (x), fp (y)) d(x, y) ≤ ≤ d(fp (x), fp (p)) + d(fp (p), fp (y)) d(x, p) − d(p, y) Lip(h) d(x, p) + Lip(h′ ) r ′ < 1, d(x, p) − r ′ where the last inequality uses the fact that d(x, p) ≥ r and the monotonicity of real Möbius maps t 7→ (t + a)/(t − b) for a, b ≥ 0.  Proof of Proposition 3.7. It is sufficient to prove that any Lipschitz map ϕ : K → Hn with C0 := Lip(ϕ) < 1 admits an extension f : Conv(K) → Hn with Lip(f ) < 1, because we can always precompose with the closest-point projection π : Hn → Conv(K), which is 1-Lipschitz. By Lemma 3.8, for any p ∈ Conv(K) we can find a neighborhood Up of p in Hn and a 1-Lipschitz extension fp : K ∪ Up → Hn of ϕ such that LipUp (fp ) < 1. By compactness of Conv(K), we can find finitely many points S p1 , . . . , pm ∈ Conv(K) such that Conv(K) ⊂ m i=1 Upi . For any 1 ≤ i ≤ m, using Remark 3.4, we extend fpi to a 1-Lipschitz map on Conv(K) ∪ Upi , still denoted by fpi . By (2.2) and Lemma 2.13, the symmetric barycenter f := m X 1 fp | : Conv(K) −→ Hn , m i Conv(K) i=1 which extends ϕ, satisfies Lip(f ) ≤ max 1≤i≤m LipUp (fpi ) + (m − 1) i m < 1.  3.3. An equivariant Kirszbraun–Valentine theorem for amenable groups. We now extend Proposition 3.1 to an equivariant setting with respect to two actions of an amenable group. Recall that a discrete group S is said to be amenable if there exists a sequence (Fi )i∈N of finite subsets of S MAXIMALLY STRETCHED LAMINATIONS 23 (called a Følner sequence) such that for any g ∈ S, #(gFi △ Fi ) −→ 0, i→+∞ #Fi where △ denotes the symmetric difference. For instance, any group which is abelian or solvable up to finite index is amenable. The following proposition will be used throughout Section 4 to extend Lipschitz maps in horoballs of Hn corresponding to cusps of the geometrically finite manifold j(Γ0 )\Hn , taking S to be a cusp stabilizer. Proposition 3.9. Let S be an amenable discrete group, (j, ρ) ⊂ Hom(S, G)2 a pair of representations with j injective and j(S) discrete in G, and K 6= ∅ a j(S)-invariant subset of Hn whose image in j(S)\Hn is compact. Any (j, ρ)-equivariant Lipschitz map ϕ : K → Hn with Lip(ϕ) ≥ 1 admits a (j, ρ)-equivariant extension f : Hn → Hn with the same Lipschitz constant. Proof. Set C0 := Lip(ϕ) ≥ 1. By Proposition 3.1 and Remark 3.6, we can find an extension f ′ : Hn → Hn of ϕ with Lip(f ′ ) = C0 , but f ′ is not equivariant a priori. We shall modify it into a (j, ρ)-equivariant map. For any γ ∈ S, the C0 -Lipschitz map fγ := ρ(γ) ◦ f ′ ◦ j(γ)−1 : Hn −→ Hn extends ϕ. For all γ, γ ′ ∈ S and all p ∈ Hn , since fγ and fγ ′ agree on K, the triangle inequality gives  (3.2) d fγ (p), fγ ′ (p) ≤ 2C0 · d(p, K). Fix a finite generating subset A of S. Using a Følner sequence of S, we see that for any ε > 0 there is a finite subset F of S such that #(γF △ F ) ≤ ε #F for all γ ∈ A. Write F = {γ1 , . . . , γk } where k = #F , and set (3.3) f ε (p) := mk (fγ1 (p), . . . , fγk (p)) 1 1 for all p ∈ Hn , where mk = m( k ,..., k ) is the averaging map of Lemma 2.11. By (2.5), the map f ε still coincides with ϕ on K. Moreover, as a barycenter of C0 -Lipschitz maps, f ε is C0 -Lipschitz (Lemma 2.13). Note, using (2.4), that (3.4) ρ(γ) ◦ f ε ◦ j(γ)−1 (p) = mk (fγγ1 (p), . . . , fγγk (p)) for all γ ∈ S and p ∈ Hn . Since #(γF △ F ) ≤ ε #F for all γ ∈ A, all but ≤ εk of the k entries of mk in (3.4) are the same as in (3.3) up to order, hence  d ρ(γ) ◦ f ε ◦ j(γ)−1 (p), f ε (p) ≤ 2C0 · d(p, K) · ε for all p ∈ Hn by (2.6), Lemma 2.11, and (3.2). We conclude by letting ε go to 0 and extracting a pointwise limit f from the f ε : such a map f : Hn → Hn is C0 -Lipschitz, extends ϕ, and is equivariant under the action of any element γ of A, hence of S.  4. The relative stretch locus We now fix a discrete group Γ0 , a pair (j, ρ) ∈ Hom(Γ0 , G)2 of representations of Γ0 in G with j geometrically finite, a j(Γ0 )-invariant subset K of Hn whose image in M := j(Γ0 )\Hn is compact (possibly empty), and a 24 FRANÇOIS GUÉRITAUD AND FANNY KASSEL (j, ρ)-equivariant Lipschitz map ϕ : K → Hn . We shall use the following terminology and notation. Definition 4.1. • The relative minimal Lipschitz constant CK,ϕ (j, ρ) is the infimum of Lipschitz constants Lip(f ) of (j, ρ)-equivariant maps f : Hn → Hn with f |K = ϕ. j,ρ the set of (j, ρ)-equivariant maps f : Hn → Hn • We denote by FK,ϕ with f |K = ϕ that have minimal Lipschitz constant CK,ϕ (j, ρ). j,ρ • If FK,ϕ 6= ∅, the relative stretch locus EK,ϕ (j, ρ) ⊂ Hn is the inter- j,ρ section of the stretch loci Ef (Definition 2.10) of all maps f ∈ FK,ϕ . n 2 e • Similarly, the enhanced relative stretch locus EK,ϕ (j, ρ) ⊂ (H ) is ef (Definition 2.10) of the intersection of the enhanced stretch loci E j,ρ . all maps f ∈ FK,ϕ Note that EK,ϕ (j, ρ) is always j(Γ0 )-invariant and closed in Hn , because j,ρ eK,ϕ (j, ρ) is always j(Γ0 )Ef is for each f ∈ FK,ϕ (Lemma 2.9). Similarly, E n invariant (for the diagonal action of j(Γ0 ) on (H )2 ) and closed in (Hn )2 . If K is empty, then CK,ϕ (j, ρ) is the minimal Lipschitz constant C(j, ρ) of (1.1) and EK,ϕ (j, ρ) is the intersection of stretch loci E(j, ρ) of Theorem 1.3, which we shall simply call the stretch locus of (j, ρ). For empty K we shall j,ρ sometimes write F j,ρ instead of FK,ϕ . 4.1. Elementary properties of the (relative) minimal Lipschitz constant and the (relative) stretch locus. We start with an easy observation for empty K. Remark 4.2. If all elements of ρ(Γ0 ) are elliptic, then C(j, ρ) = 0 and F j,ρ is the set of constant maps with image a fixed point of ρ(Γ0 ) in Hn (such a fixed point exists by Fact 2.5); in particular, E(j, ρ) = Hn . Here are now some elementary properties of CK,ϕ (j, ρ) and EK,ϕ (j, ρ) for general K. Remark 4.3. Conjugating by elements of G leaves the relative minimal Lipj,ρ schitz constant invariant and modifies the relative stretch locus (if FK,ϕ 6= ∅) by a translation: for any j, ρ ∈ Hom(Γ0 , G) and g, h ∈ G, we have  Cg·K, h◦ϕ◦g−1 (j g , ρh ) = CK,ϕ (j, ρ), Eg·K, h◦ϕ◦g−1 (j g , ρh ) = g · EK,ϕ (j, ρ), where j g := gj(·)g −1 and ρh = hρ(·)h−1 . Indeed, for any (j, ρ)-equivariant Lipschitz map f : Hn → Hn extending ϕ, the map h ◦ f ◦ g−1 : Hn → Hn extends h ◦ ϕ ◦ g−1 , is (j g , ρh )-equivariant with the same Lipschitz constant, and Lipg·p (h ◦ f ◦ g−1 ) = Lipp (f ) for all p ∈ Hn . Lemma 4.4. For any finite-index subgroup Γ′0 of Γ0 , if we set j ′ := j|Γ′0 and ρ′ := ρ|Γ′0 , then • CK,ϕ (j, ρ) = CK,ϕ (j ′ , ρ′ ); j,ρ j ′ ,ρ′ j,ρ j ′ ,ρ′ 6= ∅; , and FK,ϕ 6= ∅ if and only if FK,ϕ • FK,ϕ ⊂ FK,ϕ MAXIMALLY STRETCHED LAMINATIONS 25 • in this case, EK,ϕ (j, ρ) = EK,ϕ (j ′ , ρ′ ). By Lemma 4.4, we may always assume that • the finitely generated group Γ0 is torsion-free (using the Selberg lemma [Se, Lem. 8]); • j and ρ take values in the group G0 = PO(n, 1)0 ≃ SO(n, 1)0 of orientation-preserving isometries of Hn . This will sometimes be used in the proofs without further notice. Proof of Lemma 4.4. The inequality CK,ϕ (j ′ , ρ′ ) ≤ CK,ϕ (j, ρ) holds because any (j, ρ)-equivariant map is (j ′ , ρ′ )-equivariant. We now prove the converse inequality. Write Γ0 as a disjoint union of cosets α1 Γ′0 , . . . , αr Γ′0 where αi ∈ Γ0 . Let f ′ be a (j ′ , ρ′ )-equivariant Lipschitz extension of ϕ. For γ ∈ Γ0 , the map fγ := ρ(γ) ◦ f ′ ◦ j(γ)−1 only on the coset γΓ′0 . By (2.6), Pdepends r 1 the symmetric barycenter f := i=1 r fαi satisfies, for any γ ∈ Γ0 , −1 ρ(γ) ◦ f ◦ j(γ) γαi Γ′0 αi Γ′0 r X 1 fγαi = f, = r i=1 because the cosets are the up to order. This means that f is (j, ρ)equivariant. By Lemma 2.13, we have Lip(f ) ≤ Lip(f ′ ), hence CK,ϕ (j, ρ) ≤ CK,ϕ (j ′ , ρ′ ) by minimizing Lip(f ′ ). j,ρ Since CK,ϕ (j, ρ) = CK,ϕ (j ′ , ρ′ ), it follows from the definitions that FK,ϕ ⊂ ′ ′ j ,ρ and that, if these are nonempty, then EK,ϕ (j, ρ) ⊃ EK,ϕ (j ′ , ρ′ ). In fact, FK,ϕ ′ ′ j,ρ j ,ρ is nonempty, then so is FK,ϕ , and EK,ϕ (j, ρ) = EK,ϕ (j ′ , ρ′ ). Indeed, if FK,ϕ Pr 1 ′ ′ j ,ρ , the symmetric barycenter f = for any f ′ ∈ FK,ϕ i=1 r fαi introduced j,ρ above belongs to FK,ϕ , and the stretch locus of f is contained in that of f ′ by Lemma 2.13.  Lemma 4.5. The inequalities (4.1) C ′ (j, ρ) ≤ C(j, ρ) ≤ CK,ϕ (j, ρ), always hold, where C ′ (j, ρ) is given by (1.4). Proof. The right-hand inequality follows from the definitions. For the lefthand inequality, we observe that for any γ ∈ Γ0 with j(γ) hyperbolic and any p ∈ Hn on the translation axis of j(γ), if f : Hn → Hn is (j, ρ)-equivariant and Lipschitz, then   λ(ρ(γ)) ≤ d f (p), ρ(γ) · f (p) = d f (p), f (j(γ) · p) ≤ Lip(f ) d(p, j(γ) · p) = Lip(f ) λ(j(γ)), and we conclude by letting Lip(f ) tend to C(j, ρ).  Here is a sufficient condition for the left inequality of (4.1) to be an equality. We shall see in Theorem 5.1 that for C(j, ρ) ≥ 1 this sufficient condition j,ρ is also necessary, at least when FK,ϕ and EK,ϕ (j, ρ) are nonempty. Lemma 4.6. Let ℓ be a geodesic ray in Hn whose image in j(Γ0 )\Hn is bounded. If ℓ is maximally stretched by some (j, ρ)-equivariant Lipschitz map f : Hn → Hn , in the sense that f multiplies all distances on ℓ by Lip(f ), 26 then FRANÇOIS GUÉRITAUD AND FANNY KASSEL C ′ (j, ρ) = C(j, ρ) = Lip(f ). Proof. By (4.1) it is enough to prove that C ′ (j, ρ) ≥ Lip(f ). Parametrize ℓ by arc length as (pt )t≥0 . Since the image of ℓ in j(Γ0 )\Hn is bounded, for any ε > 0 we can find γ ∈ Γ0 and 0 < s < t with t − s ≥ 1 such that the oriented segments j(γ) · [ps , ps+1 ] and [pt , pt+1 ] of Hn are ε-close in the C 1 sense. By the closing lemma (Lemma A.1), this implies λ(j(γ)) − (t − s) ≤ 2ε. The images under f of the unit segments above are also εLip(f )-close geodesic segments. By the closing lemma again and the (j, ρ)-equivariance of f , λ(ρ(γ)) − (t − s)Lip(f ) ≤ 2ε Lip(f ). Taking ε very small, we see that λ(ρ(γ))/λ(j(γ)) takes values arbitrarily close to Lip(f ) for γ ∈ Γ0 with j(γ) hyperbolic, hence C ′ (j, ρ) ≥ Lip(f ).  4.2. Finiteness of the (relative) minimal Lipschitz constant. Lemma 4.7. (1) If j is convex cocompact, then CK,ϕ (j, ρ) < +∞. (2) In general, if j is geometrically finite, then CK,ϕ (j, ρ) < +∞ unless there exists an element γ ∈ Γ0 such that j(γ) is parabolic and ρ(γ) hyperbolic. The following proof uses Lemma 2.14 applied to an appropriate partition of unity. A similar proof scheme will be used again throughout Section 6. Proof of Lemma 4.7.(1) (Convex cocompact case). Recall Notation 2.3 for Conv(K). If j is convex cocompact, then Conv(K) is compact modulo j(Γ0 ), hence we can find open balls B1 , . . . , Br of Hn , projecting injectively to j(Γ0 )\Hn , such that Conv(K) is contained in the union of the j(Γ0 ) · Bi . For any i, let fi : Bi → Hn be a Lipschitz extension of ϕ|Bi ∩K (such an extension exists by Proposition 3.1). We extend fi to j(Γ0 ) · Bi in a (j, ρ)equivariant way (with no control on the global Lipschitz constant a priori). The function  p 7−→ Rp := diam fi (p) | 1 ≤ i ≤ r and p ∈ j(Γ0 ) · Bi is locally bounded above and j(Γ0 )-invariant, hence uniformly bounded on S i j(Γ0 )·Bi . Let (ψi )1≤i≤r be a partition of unity on Conv(K), subordinated to the covering (j(Γ0 ) · Bi )1≤i≤r , with ψi Lipschitz and j(Γ0 )-invariant for all i. Lemma 2.14 gives a (j, ρ)-equivariant map r X ψi fi : Conv(K) −→ Hn f := i=1 with Lipp (f ) bounded by some constant L independent of p ∈ Conv(K). Then LipConv(K) (f ) ≤ L by (2.2). By precomposing f with the closest-point projection π : Hn → Conv(K), which is 1-Lipschitz and (j, j)-equivariant, we obtain a (j, ρ)-equivariant Lipschitz extension of ϕ to Hn .  Proof of Lemma 4.7.(2) (General geometrically finite case). Suppose that for any γ ∈ Γ0 with j(γ) parabolic, the element ρ(γ) is not hyperbolic. The idea is the same as in the convex cocompact case, but we need to deal with the MAXIMALLY STRETCHED LAMINATIONS 27 presence of cusps, which make Conv(K) noncompact modulo j(Γ0 ). We shall apply Proposition 3.9 (the equivariant version of Proposition 3.1 for amenable groups) to the stabilizers of the cusps. Let B1 , . . . , Bc be open horoballs of Hn , disjoint from K, whose images in j(Γ0 )\Hn are disjoint and intersect the convex core in standard cusp regions (Definition 2.2), representing all the cusps. Let Bc+1 , . . . , Br be open balls of Hn that project injectively to j(Γ0 )\Hn , such that the union of the j(Γ0 ) · Bi for 1 ≤ i ≤ r covers Conv(K). For c + 1 ≤ i ≤ r, we construct a (j, ρ)-equivariant Lipschitz map fi : j(Γ0 ) · Bi → Hn as in the convex cocompact case. For 1 ≤ i ≤ c, we now explain how to construct a (j, ρ)-equivariant Lipschitz map fi : j(Γ0 ) · Bi → Hn . Let Si be the stabilizer of Bi in Γ0 for the j-action. We claim that there exists a (j|Si , ρ|Si )-equivariant Lipschitz map fi : Bi → Hn . Indeed, choose p ∈ Bi , not fixed by any element of j(Si ), and q ∈ Hn . Set fi (j(γ) · p) := ρ(γ) · q for all γ ∈ Si . Let wl : Si → N be the word length with respect to some fixed finite generating set of Si . By Lemma 2.6, there exists R′ > 0 such that  d(p, j(γ) · p) ≥ 2 log 1 + wl(γ) − R′ for all γ ∈ Si . On the other hand, there exists R > 0 such that  d(q, ρ(γ) · q) ≤ 2 log 1 + wl(γ) + R for all γ ∈ Si : if ρ(γ) is elliptic for all γ ∈ Si , this follows from the fact that the group ρ(Si ) admits a fixed point in Hn (Fact 2.5), hence d(q, ρ(γ) · q) is bounded for γ ∈ Si ; otherwise this follows from Lemma 2.6. Since the function wl is proper, we see that lim supγ∈Si d(q,ρ(γ)·q) d(p,j(γ)·p) ≤ 1, hence d(q, ρ(γ) · q) < +∞. γ∈Si r{1} d(p, j(γ) · p) sup In other words, fi is Lipschitz on j(Si ) · p. We then use Proposition 3.9 to extend fi to a (j|Si , ρ|Si )-equivariant Lipschitz map fi : Bi → Hn . Let us extend fi to j(Γ0 ) · Bi in a (j, ρ)-equivariant way (with no control on the global Lipschitz constant a priori). We claim that Rp := diam{fi (p) | 1 ≤ i ≤ r and p ∈ j(Γ0 ) · Bi } is uniformly bounded on Conv(K). Indeed, j(Γ0 ) · Bi ∩ j(Γ0 ) · Bk = ∅ for all 1 ≤ i 6= k ≤ c by definition of standard cusp regions. Therefore, if p ∈ Hn belongs to j(Γ0 )·Bi for 1 ≤ i ≤ r, then it belongs to the  one index S S more than “thick” part j(Γ0 ) · c<i≤r Bi is bounded and p 7→ Rp is c<i≤r Bi . But locally bounded above and j(Γ0 )-invariant, hence Rp is uniformly bounded on Conv(K). We conclude as in the convex cocompact case.  The converse to Lemma 4.7 is clear: if there exists an element γ ∈ Γ0 such that j(γ) is parabolic and ρ(γ) hyperbolic, then CK,ϕ (j, ρ) = +∞. Indeed, for any p, q ∈ Hn , the distance d(p, j(γ k ) · p) grows logarithmically in k (Lemma 2.6) whereas d(q, ρ(γ k ) · q) grows linearly. 28 FRANÇOIS GUÉRITAUD AND FANNY KASSEL In the rest of the paper, we shall assume CK,ϕ (j, ρ) < +∞ whenever we discuss a fixed pair (j, ρ). We shall allow C(j, ρ) = +∞ only in Section 6 and Proposition 9.5, where we discuss semicontinuity properties of the maps (j, ρ) 7→ C(j, ρ) and (j, ρ) 7→ E(j, ρ), for empty K. 4.3. Projecting onto the convex core. In the proof of Lemma 4.7, we used the closest-point projection π : Hn → Conv(K), which is 1-Lipschitz and (j, j)-equivariant. This projection will be used many times in Sections 4, 5, and 6, with the following more precise properties. j,ρ Lemma 4.8. Suppose FK,ϕ 6= ∅ and let π : Hn → Conv(K) be the closestj,ρ point projection. For any f ∈ FK,ϕ , j,ρ (1) Lip(f ◦ π) = Lip(f |Conv(K) ) = Lip(f ) = CK,ϕ (j, ρ) and f ◦ π ∈ FK,ϕ ; (2) if CK,ϕ (j, ρ) > 0, then the stretch loci and enhanced stretch loci (Definition 2.10) satisfy ( Ef ◦π = Ef |Conv(K) ⊂ Ef ∩ Conv(K),  ef ◦π = E ef | ef ∩ Conv(K) × Conv(K) . E ⊂E Conv(K) In particular, if CK,ϕ (j, ρ) > 0, then the relative stretch locus EK,ϕ (j, ρ) is always contained in Conv(K). (This is not true if CK,ϕ (j, ρ) = 0: see Remark 4.2.) j,ρ Proof. For any f ∈ FK,ϕ we have Lip(f ◦ π) = Lip(f |Conv(K) ) ≤ Lip(f ) since j,ρ π : Hn → Conv(K) is 1-Lipschitz. Equality holds and f ◦ π ∈ FK,ϕ since π is (j, j)-equivariant and Lip(f ) = CK,ϕ (j, ρ) is minimal. This proves (1). e ⊂ Hn × Hn , For (2), it is enough to consider the enhanced stretch locus E e to either of the Hn factors. since the stretch locus E is the projection of E Note that ef | ef ◦π ⊂ Conv(K) × Conv(K), E ⊂E Conv(K) because π is the identity on Conv(K) and is contracting outside Conv(K). ef ◦π ⊂ E ef | ef ◦π . By defLet us prove that E . Consider a pair (x, x′ ) ∈ E Conv(K) inition, there are sequences (xk )k∈N converging to x and (x′k )k∈N converging to x′ such that xk 6= x′k and d(f ◦ π(xk ), f ◦ π(x′k )) −→ CK,ϕ (j, ρ). k→+∞ d(xk , x′k ) By continuity of π we have π(xk ) → π(x) = x and π(x′k ) → π(x′ ) = x′ . Since d(f ◦ π(xk ), f ◦ π(x′k )) d(f ◦ π(xk ), f ◦ π(x′k )) ≤ ≤ CK,ϕ (j, ρ), d(xk , x′k ) d(π(xk ), π(x′k )) the middle term also tends to CK,ϕ (j, ρ), which shows that (π(x), π(x′ )) = ef | ef ◦π = E ef | ef . (x, x′ ) belongs to E . Thus E ⊂E  Conv(K) Conv(K) 4.4. Equivariant extensions with minimal Lipschitz constant. We shall use the following terminology. MAXIMALLY STRETCHED LAMINATIONS 29 Definition 4.9. A representation ρ ∈ Hom(Γ0 , G) is reductive if the Zariski closure of ρ(Γ0 ) in G is reductive, or equivalently if the number of fixed points of the group ρ(Γ0 ) in the boundary at infinity ∂∞ Hn of Hn is different from 1. j,ρ Lemma 4.10. The set FK,ϕ of Definition 4.1 is nonempty as soon as either K 6= ∅ or ρ is reductive. When K = ∅ and ρ is nonreductive, there may or may not exist a (j, ρ)equivariant map f : Hn → Hn with minimal constant C(j, ρ) = CK,ϕ (j, ρ): see examples in Sections 10.2 and 10.3. Proof of Lemma 4.10. The idea is to apply the Arzelà–Ascoli theorem. Set C := CK,ϕ (j, ρ) and let (fk )k∈N be a sequence of (j, ρ)-equivariant Lipschitz maps with fk |K = ϕ and C + 1 ≥ Lip(fk ) → C. The sequence (fk ) is equicontinuous. We first assume that K 6= ∅, and fix q ∈ K. For any k ∈ N and any p ∈ Hn ,  (4.2) d fk (p), ϕ(q) ≤ (C + 1) d(p, q). Therefore, for any compact subset C of Hn , the sets fk (C ) for k ∈ N all lie in some common compact subset of Hn . The Arzelà–Ascoli theorem applies, yielding a subsequence with a C-Lipschitz limit; this limit necessarily belongs j,ρ to FK,ϕ . We now assume that K = ∅ and ρ is reductive. H2 Aρ(γ1 ) Aρ(γ2 ) Figure 2. Uniform neighborhoods of lines in Hn with disjoint endpoints have a compact intersection. • If the group ρ(Γ0 ) has no fixed point in Hn and does not preserve any geodesic line of Hn (this is the generic case), then ρ(Γ0 ) contains two hyperbolic elements ρ(γ1 ), ρ(γ2 ) whose translation axes have no common endpoint in ∂∞ Hn . Fix a point p ∈ Hn . For any k ∈ N and i ∈ {1, 2},  d fk (p), ρ(γi ) · fk (p) ≤ (C + 1) d(p, j(γi ) · p). 30 FRANÇOIS GUÉRITAUD AND FANNY KASSEL Therefore, the points fk (p) for k ∈ N belong to some uniform neighborhood of the translation axis Aρ(γi ) of ρ(γi ) for i ∈ {1, 2}. Since Aρ(γ1 ) and Aρ(γ2 ) have no common endpoint at infinity, the points fk (p) belong to some compact subset of Hn (see Figure 2). Since Lip(fk ) stays bounded, we obtain that for any compact subset C of Hn , the sets fk (C ) for k ∈ N all lie inside some common compact subset of Hn , and we conclude as above using the Arzelà–Ascoli theorem. • If the group ρ(Γ0 ) preserves a geodesic line A of Hn , then it commutes with any hyperbolic element of G acting as a pure translation along A. For any k ∈ N and any such hyperbolic element gk , the map gk ◦ fk is still (j, ρ)-equivariant, with Lip(gk ◦ fk ) = Lip(fk ). Fix p ∈ Hn . By the previous paragraph, the points fk (p) for k ∈ N belong to some uniform neighborhood of A. Therefore, after replacing (fk )k∈N by (gk ◦ fk )k∈N for some appropriate sequence (gk )k∈N , we may assume that the points fk (p) for k ∈ N all belong to some compact subset of Hn , and we conclude as above. • If the group ρ(Γ0 ) has a fixed point in Hn , we use Remark 4.2.  4.5. The stretch locus of an equivariant extension with minimal Lipschitz constant. Lemma 4.11. (1) If j is convex cocompact, then the stretch locus Ef of j,ρ any f ∈ FK,ϕ is nonempty. j,ρ (2) In general, the stretch locus of any f ∈ FK,ϕ is nonempty except possibly if CK,ϕ (j, ρ) = 1 and ρ is not cusp-deteriorating. Recall from Definition 1.1 that “ρ is not cusp-deteriorating” means there is an element γ ∈ Γ0 such that j(γ) and ρ(γ) are both parabolic. When CK,ϕ (j, ρ) = 1, there exist examples of pairs (j, ρ) with ρ non-cusp-deterioraj,ρ ting such that the stretch locus Ef is empty for some maps f ∈ FK,ϕ (see Sections 10.8 and 10.9, as well as Corollary 4.18 for elementary Γ0 ). Proof of Lemma 4.11.(1) (Convex cocompact case). By Lemma 4.8, it is sufficient to prove that the stretch locus of f |Conv(K) is nonempty. The function x 7→ Lipx (f |Conv(K) ) is upper semicontinuous (Lemma 2.9) and j(Γ0 )invariant. If j is convex cocompact, then Conv(K) is compact modulo j(Γ0 ), and so x 7→ Lipx (f |Conv(K) ) achieves its maximum on Conv(K), at a point that belongs to the stretch locus of f |Conv(K) .  Proof of Lemma 4.11.(2) (General geometrically finite case). Assume either that C 6= 1, or that C = 1 and ρ is cusp-deteriorating, where we set j,ρ C := CK,ϕ (j, ρ). Consider f ∈ FK,ϕ . As in the convex cocompact case, it is sufficient to prove that the stretch locus of f |Conv(K) is nonempty. Suppose by contradiction that it is empty: this means (Lemma 2.9) that LipK ′ (f ) < C for any compact subset K ′ of Conv(K), or equivalently that the j(Γ0 )-invariant function x 7→ Lipx (f ) only approaches C asymptotically (from below) in some cusps. Let B1 , . . . , Bc be open horoballs of Hn , disjoint from K, whose images in j(Γ0 )\Hn are disjoint and intersect the convex core in standard cusp regions (Definition 2.2), representing all the cusps. Our MAXIMALLY STRETCHED LAMINATIONS 31 strategy is, for each Bi on which x 7→ Lipx (f ) approaches C asymptotically, to modify f |Conv(K) on Conv(K) ∩ j(Γ0 ) · Bi in a (j, ρ)-equivariant way so as to decrease the Lipschitz constant on Conv(K) ∩ Bi . By (2.2), this will yield a new (j, ρ)-equivariant extension of ϕ to Conv(K) with a smaller Lipschitz constant than f |Conv(K) , which will contradict the minimality of Lip(f |Conv(K) ) = Lip(f ). Let us now explain the details. Let B = Bi be an open horoball as above, on which x 7→ Lipx (f ) approaches C asymptotically, and let S be the stabilizer of B in Γ0 for the j-action. The group j(S) is discrete and contains only parabolic and elliptic elements. Since CK,ϕ (j, ρ) < +∞ by assumption, the group ρ(S) also contains only parabolic and elliptic elements (Lemma 4.7). First we assume that ρ(S) contains a parabolic element, i.e. ρ is not deteriorating in B (Definition 2.4). In particular, ρ is not cusp-deteriorating, hence C ≥ 1 by Lemma 2.7 and so C > 1 by the assumption made at the beginning of the proof. Since S is amenable, in order to decrease the Lipschitz constant on Conv(K) ∩ B it is enough to prove that LipConv(K)∩∂B (f ) < C, because we can then apply Proposition 3.9. By geometrical finiteness and the assumption that the image of K in j(Γ0 )\Hn is compact (see Fact 2.1 and the remarks after Notation 2.3), we can find a compact fundamental domain D of Conv(K) ∩ j(Γ0 ) · ∂B for the action of j(Γ0 ). Fix p ∈ D. By Lemma 2.6, there exist R, R′ > 0 such that (4.3) and (4.4) d(p, j(γ) · p) ≥ 2 log(1 + wl(γ)) − R′   d f (p), f (j(γ) · p) = d f (p), ρ(γ) · f (p) ≤ 2 log(1 + wl(γ)) + R for all γ ∈ S, where wl : S → N denotes the word length with respect to some fixed finite generating set of S. Consider q, q ′ ∈ Conv(K) ∩ ∂B with q ∈ D; there is an element γ ∈ Γ0 such that d(j(γ) · p, q ′ ) ≤ ∆, where ∆ > 0 is the diameter of D. By the triangle inequality, (4.3), and (4.4), we have d(q, q ′ ) ≥ d(p, j(γ) · p) − d(p, q) − d(j(γ) · p, q ′ ) ≥ 2 log(1 + wl(γ)) − (R′ + 2∆) and, using Lip(f ) = C,  d(f (q), f (q ′ )) ≤ d f (p), f (j(γ) · p) + d(f (p), f (q)) + d(f (j(γ) · p), f (q ′ )) ≤ 2 log(1 + wl(γ)) + (R + 2C∆). Since C > 1, this implies 1+C d(f (q), f (q ′ )) ≤ <C ′ d(q, q ) 2 as soon as wl(γ) is large enough, or equivalently as soon as d(q, q ′ ) is large enough. However, this ratio is also bounded away from C when d(q, q ′ ) is bounded, because the segment [q, q ′ ] then stays in a compact part of j(Γ0 )\Conv(K). Therefore there is a constant C ′′ < C such that ′ ′′ d(f (q), f (q )) ≤ C d(q, q ′ ) for all q, q ′ ∈ Conv(K) ∩ ∂B with q ∈ D, hence LipConv(K)∩∂B (f ) ≤ C ′′ < C by equivariance. By Proposition 3.9, we can redefine f inside Conv(K)∩B so that LipConv(K)∩B (f ) < C. We then extend f to Conv(K) ∩ j(Γ0 ) · B in a (j, ρ)-equivariant way. 32 FRANÇOIS GUÉRITAUD AND FANNY KASSEL We now assume that ρ(S) consists entirely of elliptic elements, i.e. ρ is deteriorating in B (Definition 2.4). Then ρ(S) admits a fixed point q in Hn by Fact 2.5. Let f1 : j(Γ0 ) · B → Hn be the (j, ρ)-equivariant map that is constant equal to q on B, and let ψ1 : Hn → [0, 1] be the j(Γ0 )-invariant function supported on j(Γ0 ) · B given by  ψ1 (p) = ε ψ d(p, ∂B) for all p ∈ B, where ψ : R+ → [0, 1] is the 3-Lipschitz function with ψ|[0,1/3] = 0 and ψ|[2/3,+∞) = 1, and ε > 0 is a small parameter to be adjusted later. Let f2 := f , and let ψ2 := 1 − ψ1 . The (j, ρ)-equivariant map f0 := ψ1 f1 + ψ2 f2 : Hn −→ Hn coincides with f on Conv(K) ∩ ∂B. Let us prove that if ε is small enough, then Lipp (f0 ) is bounded by some uniform constant < C for p ∈ Conv(K)∩B. Let p ∈ Conv(K) ∩ B. Since Lipp (f1 ) = 0, since f1 (p) = q, and since f2 = f , Lemma 2.14 yields  Lipp (f0 ) ≤ Lipp (ψ1 ) + Lipp (ψ2 ) d(q, f (p)) + ψ2 (p) Lipp (f ). Let B ′ be a horoball contained in B, at distance 1 from ∂B. If p ∈ Conv(K)∩B ′ , then Lipp (ψ1 ) = Lipp (ψ2 ) = 0 and ψ2 (p) = 1 − ε, hence Lipp (f0 ) ≤ (1 − ε) Lipp (f ) ≤ (1 − ε) C. If p ∈ Conv(K) ∩ (B r B ′ ), then Lipp (ψ1 ), Lipp (ψ2 ) ≤ 3ε and ψ2 (p) ≤ 1, hence Lipp (f0 ) ≤ 6ε d(q, f (p)) + sup Lipx (f ) . x∈Conv(K)∩(BrB ′ ) Conv(K)∩(B rB ′ ) Note that the set is compact modulo j(S), which implies, on the one hand that the j(S)-invariant, continuous function p 7→ d(q, f (p)) is bounded on Conv(K) ∩ (B r B ′ ), on the other hand that the j(S)invariant, upper semicontinuous function x 7→ Lipx (f ) is bounded away from C on Conv(K) ∩ (B r B ′ ) (recall that the stretch locus of f |Conv(K) is empty by assumption). Therefore, if ε is small enough, then Lipp (f0 ) is bounded by some uniform constant < C for p ∈ Conv(K) ∩ B, which implies LipConv(K)∩B (f0 ) < C by (2.2). We can redefine f to be f0 on Conv(K) ∩ B. We then extend f to Conv(K) ∩ j(Γ0 ) · B in a (j, ρ)-equivariant way. After redefining f as above in each cusp where the local Lipschitz constant x 7→ Lipx (f ) approaches C asymptotically, we obtain a (j, ρ)-equivariant map on Conv(K) with Lipschitz constant < C, which contradicts the minimality of C.  4.6. Optimal extensions with minimal Lipschitz constant. j,ρ Definition 4.12. An element f0 ∈ FK,ϕ (Definition 4.1) is called optimal if e its enhanced stretch locus Ef0 (Definition 2.10) is minimal, equal to \ eK,ϕ (j, ρ) = ef . E E j,ρ f ∈FK,ϕ This means that T the ordinary stretch locus Ef0 of f0 is minimal, equal to EK,ϕ (j, ρ) = f ∈F j,ρ Ef , and that the set of maximally stretched segK,ϕ ments of f0 is minimal (using Remark 2.8.(1)). This last condition will be MAXIMALLY STRETCHED LAMINATIONS 33 relevant only when CK,ϕ (j, ρ) = 1, in the proof of Lemma 5.4: indeed, when j,ρ CK,ϕ (j, ρ) > 1, Theorem 5.1 shows that f0 ∈ FK,ϕ is optimal if and only if its ordinary stretch locus Ef0 is minimal. As mentioned in the introduction, in general an optimal map f0 is by no means unique, since it may be perturbed away from EK,ϕ (j, ρ). j,ρ j,ρ Lemma 4.13. If FK,ϕ 6= ∅, then there exists an optimal element f0 ∈ FK,ϕ . j,ρ ef is closed in Hn × Hn , the enhanced stretch locus E Proof. For any f ∈ FK,ϕ (Lemma 2.9 and Remark 2.8.(1)) and j(Γ0 )-invariant for the diagonal action. eK,ϕ (j, ρ) is also closed and j(Γ0 )-invariant. By definition, for any Therefore E eK,ϕ (j, ρ) (possibly with p = q), we can find a x = (p, q) ∈ (Hn × Hn ) r E j,ρ n , and neighborhood Ux of x in H × Hn , a (j, ρ)-equivariant map fx ∈ FK,ϕ a constant δx > 0 such that d(fx (p′ ), fx (q ′ )) = CK,ϕ (j, ρ) − δx < CK,ϕ (j, ρ). sup d(p′ , q ′ ) (p′ ,q ′ )∈Ux p′ 6=q ′ eK,ϕ (j, ρ) is exhausted by countably many compact sets, Since (Hn × Hn ) r E we can write +∞ [ n n e Ux i (H × H ) r EK,ϕ (j, ρ) = i=1 Hn ) eK,ϕ (j, ρ). Choose a for some sequence (xi )i≥1 of points of × rE n point p ∈ H and let α = (αi )i≥1 be a sequence of positive reals summing up to 1 and decreasing fast enough so that (Hn +∞ X αi d(fx1 (p), fxi (p))2 < +∞ . i=1 By Lemma 2.13, the map f0 := sup (p,q)∈Uxi p6=q P∞ i=1 αi fxi is well defined and satisfies d(f0 (p), f0 (q)) ≤ CK,ϕ (j, ρ) − αi δxi < CK,ϕ (j, ρ) d(p, q) ef ∩ Ux = ∅, which means that E ef = E eK,ϕ (j, ρ). for all i, hence E 0 i 0  Here is an immediate consequence of Lemmas 4.11 and 4.13. j,ρ Corollary 4.14. If FK,ϕ 6= ∅, then the relative stretch locus EK,ϕ (j, ρ) • is nonempty for convex cocompact j; • is nonempty for geometrically finite j in general, except possibly if CK,ϕ (j, ρ) = 1 and ρ is not cusp-deteriorating. In fact, the following holds. j,ρ Lemma 4.15. If FK,ϕ 6= ∅, then for any p ∈ Hn r (EK,ϕ (j, ρ) ∪ K) there j,ρ exists an optimal element f0 ∈ FK,ϕ that is constant on a neighborhood of p. j,ρ j,ρ Proof. Assume that FK,ϕ 6= ∅ and let f ∈ FK,ϕ be optimal (given by Lemma 4.13). Fix p ∈ Hn r (EK,ϕ (j, ρ) ∪ K). Let B ⊂ Hn be a closed 34 FRANÇOIS GUÉRITAUD AND FANNY KASSEL ball centered at p, with small radius r > 0, such that B does not meet K ∪ EK,ϕ (j, ρ) and projects injectively to j(Γ0 )\Hn . By Lemma 2.9, C ∗ := LipB (f ) < C := CK,ϕ (j, ρ). For any small enough ball B ′ ⊂ B of radius r ′ centered at p, the map ιp : ∂B ∪ B ′ −→ Hn that coincides with the identity on ∂B and is constant with image {p} on B ′ , r ∗ satisfies 1 < Lip(ιp ) = r−r ′ < C/C . Proposition 3.1 enables us to extend ιp to a map ι′p : B → Hn fixing ∂B pointwise with Lip(ι′p ) < C/C ∗ . We may moreover assume ι′p (B) ⊂ B up to postcomposing with the closestpoint projection onto B. The (j, j)-equivariant map Jp : Hn → Hn that coincides with ι′p on B and with the identity on Hn r j(Γ0 ) · B satisfies Lipx (Jp ) ≤ Lip(ι′p ) < C/C ∗ if x ∈ j(Γ0 ) · B and Lipx (Jp ) = 1 otherwise. Thus, by (2.2), we see that the (j, ρ)-equivariant map f0 := f ◦ Jp satisfies Lipx (f0 ) ≤ C ∗ Lip(ι′p ) < C if x ∈ j(Γ0 )·B and Lipx (f0 ) = Lipx (f ) otherwise. In particular, f0 is C-Lipschitz, constant on B ′ , extends ϕ, and its (enhanced) stretch locus is contained in that of the optimal map f . Therefore f0 is optimal.  4.7. Behavior in the cusps for (almost) optimal Lipschitz maps. In this section we consider representations j that are geometrically finite but j,ρ not convex cocompact. We show that when FK,ϕ is nonempty, we can find j,ρ optimal maps f0 ∈ FK,ϕ (in the sense of Definition 4.12) that “show no bad behavior” in the cusps. To express this, we consider open horoballs B1 , . . . , Bc of Hn whose images in M := j(Γ0 )\Hn are disjoint and intersect the convex core in standard cusp regions (Definition 2.2), representing all the cusps. We take them small enough so that K ∩ j(Γ0 ) · Bi = ∅ for all i. Then the following holds. Proposition 4.16. Consider C ∗ < +∞ such that there exists a C ∗ -Lipschitz, (j, ρ)-equivariant extension f : Hn → Hn of ϕ. (1) If C ∗ ≥ 1, then we can find a C ∗ -Lipschitz, (j, ρ)-equivariant extension f0 : Hn → Hn of ϕ and horoballs Bi′ ⊂ Bi such that LipBi′ (f0 ) = 0 for all deteriorating Bi and LipBi′ (f0 ) = 1 for all non-deteriorating Bi . (2) If C ∗ < 1, then we can find a C ∗ -Lipschitz, (j, ρ)-equivariant extension f0 : Hn → Hn of ϕ that converges to a point pi in any Bi (i.e. the sets f0 (Bi′ ) converge to {pi } for smaller and smaller horoballs Bi′ ⊂ Bi ). (3) If C ∗ < 1, then for any ε > 0 we can find a (C ∗ + ε)-Lipschitz, (j, ρ)equivariant extension f0 : Hn → Hn of ϕ and horoballs Bi′ ⊂ Bi such that LipBi′ (f0 ) = 0 for all i. Moreover, if C ∗ = CK,ϕ (j, ρ), then in (1) and (2) we can choose f0 such that its enhanced stretch locus is contained in that of f . In particular, f0 is optimal if f is. By “Bi deteriorating” we mean that ρ is deteriorating in Bi in the sense of Definition 2.4. Recall that all Bi are deteriorating when C ∗ < 1 (Lemma 2.7). MAXIMALLY STRETCHED LAMINATIONS 35 If Bi is not deteriorating, then any (j, ρ)-equivariant map has Lipschitz constant ≥ 1 in Bi (see Lemma 2.6), hence the property LipBi′ (f0 ) = 1 in (1) cannot be improved. We believe that the condition C ∗ ≥ 1 could be dropped in (1), which would then supersede both (2) and (3) (see Appendix C.4). Note that if f0 converges to a point pi in Bi , then pi must be a fixed point of the group ρ(Si ), where Si ⊂ Γ0 is the stabilizer of Bi under j. Here is an immediate consequence of Proposition 4.16.(1), of Lemma 4.8, and of the fact that the complement of the cusp regions in Conv(K) is j,ρ compact (Fact 2.1). Recall that FK,ϕ is nonempty as soon as K 6= ∅ or ρ is reductive (Lemma 4.10). j,ρ 6= ∅. If Corollary 4.17. Suppose FK,ϕ • CK,ϕ (j, ρ) > 1, or • CK,ϕ (j, ρ) = 1 and ρ is cusp-deteriorating, then the image of the relative stretch locus EK,ϕ (j, ρ) in j(Γ0 )\Hn is compact. Here is another consequence of Proposition 4.16 and Lemma 4.8, in the case when the group j(Γ0 ) is virtually Zm for some m < n. Corollary 4.18. If the groups j(Γ0 ) and ρ(Γ0 ) both have a unique fixed point in ∂∞ Hn , then C(j, ρ) = 1 and F j,ρ 6= ∅ and E(j, ρ) = ∅. Proof of Corollary 4.18. If j(Γ0 ) and ρ(Γ0 ) both have a unique fixed point in ∂∞ Hn , then ρ is not cusp-deteriorating with respect to j, and so C(j, ρ) ≥ 1 by Lemma 2.7. By Proposition 4.16.(1) we can find a (j, ρ)-equivariant map f : Hn → Hn and a j(Γ0 )-invariant horoball B of Hn such that LipB (f ) = 1. If we denote by πB : Hn → B the closest-point projection, then f ◦πB : Hn → Hn is (j, ρ)-equivariant and 1-Lipschitz. Thus C(j, ρ) = 1 and f ∈ F j,ρ . Lemma 4.8 shows that Ef is contained in any j(Γ0 )-invariant horoball B ′ ⊂ B, hence it is empty. In particular, E(j, ρ) = ∅.  Proof of Proposition 4.16. For any 1 ≤ i ≤ c we explain how f |Conv(K) can be modified on j(Γ0 ) · Bi ∩ Conv(K) to obtain a new (j, ρ)-equivariant Lipschitz extension f0 : Conv(K) → Hn of ϕ such that f0 (precomposed as per Lemma 4.8 with the closest-point projection πConv(K) onto Conv(K)) has the desired properties, namely (A)–(B)–(C)–(D) below. More precisely, the implications will be (A) ⇒ (2), (B) ⇒ (3), and (C)–(D) ⇒ (1). We denote by Si the stabilizer of Bi in Γ0 under j. • (A) Convergence in deteriorating cusps. We first consider the case where Bi is deteriorating and prove that there is a C ∗ -Lipschitz, (j, ρ)-equivariant extension f0 : Conv(K) → Hn of ϕ such that f0 converges to a point on Bi ∩ Conv(K), agrees with f on Conv(K) r j(Γ0 ) · Bi , and satisfies d(f0 (p), f0 (q)) ≤ d(f (p), f (q)) for all p, q ∈ Conv(K). If C ∗ = CK,ϕ (j, ρ), then this last condition implies that the enhanced stretch locus of f0 is contained in that of f . It is sufficient to prove that for any δ > 0 there is a C ∗ -Lipschitz, (j, ρ)equivariant extension fδ : Conv(K) → Hn of ϕ such that fδ agrees with f on Conv(K) r j(Γ0 ) · Bi , satisfies d(fδ (p), fδ (q)) ≤ d(f (p), f (q)) for all p, q ∈ Bi ∩Conv(K), and for some horoball Bi′ ⊂ Bi , the set fδ (Bi′ ∩Conv(K)) is contained in the intersection of the convex hull of f (Bi′ ∩ Conv(K)) with a 36 FRANÇOIS GUÉRITAUD AND FANNY KASSEL ball of radius δ. Indeed, if this is proved, then we can apply the process to f and δ = 1 to construct a map f(1) , and then inductively to f(i) and δ = 1/2i for any i ≥ 1 to construct a map f(i+1) ; extracting a pointwise limit, we obtain a map f0 satisfying the required properties. Fix δ > 0 and let us construct fδ as above. Choose a generating subset {s1 , . . . , sm } of Si , a compact fundamental domain D of ∂Bi ∩ Conv(K) for the action of j(Si ) (use Fact 2.1), and a point p ∈ D. For t ≥ 0, the closestpoint projection πt from Bi onto the closed horoball at distance t of ∂Bi inside Bi commutes with the action of j(Si ). Set pt := πt (p); by (A.5), the number max1≤k≤m d(pt , j(sk ) · pt ) goes to 0 as t → +∞. We can also find fundamental domains Dt of πt (∂Bi ) ∩ Conv(K), containing pt , whose diameters go to 0 as t → +∞. Since f is Lipschitz and (j, ρ)-equivariant, the diameter of f (Dt ) and the function t 7→ max1≤k≤m d(f (pt ), ρ(sk ) · f (pt )) also tend to 0 as t → +∞. Let Fi ⊂ Hn be the fixed set of ρ(Si ) (a single point or a copy of Hd , for some d ≤ n). There exists η > 0 such that for any x ∈ Hn , if max1≤k≤m d(x, ρ(sk ) · x) < η, then d(x, Fi ) < δ/2. Applying this to x = f (pt ), we see that for large enough t there is a point qt ∈ Fi such that d(f (pt ), qt ) < δ/2 and the diameter of f (Dt ) is < δ/2, which implies that the ρ(Si )-invariant set f (Conv(K) ∩ πt (∂Bi )) is contained in the ball Ω := Bqt (δ) of radius δ centered at qt . Let πΩ : Hn → Ω be the closest-point projection onto Ω (see Figure 3). The (j, ρ)-equivariant map fδ : Conv(K) → Hn that agrees with f on Conv(K) r j(Γ0 ) · πt (Bi ) and with πΩ ◦ f on Conv(K) ∩ πt (Bi ) satisfies the required properties. • (B) Constant maps with a slightly larger Lipschitz constant in deteriorating cusps. We still consider the case when Bi is deteriorating. For ε > 0, we prove that there is a (C ∗ + ε)-Lipschitz, (j, ρ)-equivariant extension f0 : Hn → Hn of ϕ that is constant on some horoball Bi′ ⊂ Bi and that agrees with f on Hn r j(Γ0 ) · Bi . Fix ε > 0. By (A), we may assume that f converges to a point pi on Bi , hence there is a horoball Bi′′ ⊂ Bi such that f (Bi′′ ) is contained in the ball of diameter ε/2 centered at pi . Let fi : j(Γ0 )·Bi′′ → Hn be the (j, ρ)-equivariant map that extends the constant map Bi′′ → {pi }, and let ψ : Hn → [0, 1] be a j(Γ0 )-invariant, 1-Lipschitz function equal to 1 on a neighborhood of Hn r j(Γ0 ) · Bi′′ and vanishing far inside Bi′′ . The map f0 := ψf + (1 − ψ)fi is a (j, ρ)-equivariant extension of ϕ that is constant on some horoball Bi′ ⊂ Bi and that agrees with f on Hn r j(Γ0 ) · Bi . By Lemma 2.14, Lipp (f0 ) ≤ Lipp (f ) ≤ C ∗ for all p ∈ Hn r j(Γ0 ) · Bi′′ , and Lipp (f0 ) ≤ Lipp (f ) + ε ≤ C ∗ + ε for all p ∈ j(Γ0 ) · Bi′′ , hence f0 is (C ∗ + ε)-Lipschitz by (2.2). • (C) Constant maps in deteriorating cusps when C ∗ ≥ 1. We now consider the case when Bi is deteriorating and C ∗ ≥ 1. We construct a C ∗ -Lipschitz, (j, ρ)-equivariant extension f0 : Conv(K) → Hn of ϕ that is constant on Bi′ ∩ Conv(K) for some horoball Bi′ ⊂ Bi and agrees with f on Conv(K) r j(Γ0 ) · Bi . We also prove that if C ∗ = CK,ϕ (j, ρ), then the MAXIMALLY STRETCHED LAMINATIONS Hn Bi 37 πt (Bi ) πt (D) D f πΩ Bqt (δ) = Ω f (D) Bqt (δ) = Ω f (D) Fi = Fix (ρ(Si )) Fi = Fix (ρ(Si )) Figure 3. Step (A): Postcomposition with the closest-point projection onto the small, ρ(Si )-invariant ball Ω. enhanced stretch locus of f0 (hence of f0 ◦πConv(K) by Lemma 4.8) is included in that of f . By (A), we may assume that f converges to a point pi on Bi . Let Bi′ be a horoball strictly contained in Bi . Since the set ∂Bi ∩ Conv(K) is compact modulo j(Si ) (Fact 2.1), its image under f lies within bounded distance from pi . Therefore, if Bi′ is far enough from ∂Bi , then the map from (Conv(K) r j(Γ0 ) · Bi ) ∪ (Bi′ ∩ Conv(K)) to Hn that agrees with f on Conv(K) r j(Γ0 ) · Bi and that is constant equal to pi on Bi′ ∩ Conv(K) is C ∗ Lipschitz. By Proposition 3.9, we can extend it to a C ∗ -Lipschitz, (j|Si , ρ|Si )equivariant map from (Conv(K)rj(Γ0 )·Bi )∪(Bi ∩Conv(K)) to Hn . Finally we extend this map to a (j, ρ)-equivariant map f (1) : Conv(K) → Hn . Then f (1) is C ∗ -Lipschitz, agrees with f on Conv(K) r j(Γ0 ) · Bi , and is constant on Bi′ ∩ Conv(K). Suppose that C ∗ = CK,ϕ (j, ρ). Then Lip(f (1) ) = C ∗ (and no smaller). The stretch locus (and maximally stretched segments) of f (1) are included in those of f , except possibly between ∂Bi and ∂Bi′ . To deal with this issue, we consider two horoballs Bi′′′ ( Bi′′ strictly contained in Bi′ and, similarly, construct a C ∗ -Lipschitz, (j, ρ)-equivariant map f (2) : Conv(K) → Hn that agrees with f on Conv(K) r j(Γ0 ) · Bi′′ and is constant on Bi′′′ ∩ Conv(K). 38 FRANÇOIS GUÉRITAUD AND FANNY KASSEL The (j, ρ)-equivariant map 1 1 f0 := f (1) + f (2) 2 2 still agrees with f on Conv(K)rj(Γ0 )·Bi and is constant on Bi′′′ ∩Conv(K). By Lemma 2.13, its (enhanced) stretch locus is included in that of f . • (D) Lipschitz constant 1 in non-deteriorating cusps. We now consider the case when Bi is not deteriorating; in particular, C ∗ ≥ 1 by Lemma 2.7. We construct a C ∗ -Lipschitz, (j, ρ)-equivariant extension f0 : Conv(K) → Hn of ϕ such that LipBi′ ∩Conv(K) (f0 ) = 1 for some horoball Bi′ ⊂ Bi and f0 agrees with f on Conv(K) r j(Γ0 ) · Bi . We also prove that if C ∗ = CK,ϕ (j, ρ) then the enhanced stretch locus of f0 (hence of f0 ◦πConv(K) ) is included in that of f . We assume C ∗ > 1 (otherwise we may take f0 = f ). It is sufficient to construct a 1-Lipschitz, (j|Si , ρ|Si )-equivariant map fi : Bi′ ∩Conv(K) → Hn , for some horoball Bi′ ⊂ Bi , such that the (j|Si , ρ|Si )-equivariant map  f (1) : Conv(K) r j(Γ0 ) · Bi ∪ (Bi′ ∩ Conv(K)) −→ Hn that agrees with f on Conv(K) r j(Γ0 ) · Bi and with fi on Bi′ ∩ Conv(K) satisfies Lip(f (1) ) ≤ C ∗ . Indeed, we can then extend f (1) to a C ∗ -Lipschitz, (j, ρ)-equivariant map Conv(K) → Hn using Proposition 3.9, as in step (C). Proceeding with two other horoballs Bi′′′ ( Bi′′ to get a map f (2) and averaging as in step (C), we obtain a map f0 with the required properties. To construct fi , we use explicit coordinates: in the upper half-space model Rn−1 × R∗+ of Hn , we may assume (using Remark 4.3) that j(Si ) and ρ(Si ) both fix the point at infinity, that the horosphere ∂Bi is Rn−1 ×{1}, and that f fixes the point (0, 1) ∈ Rn−1 × R∗+ . Let Wi be the orthogonal projection to Rn−1 of Conv(K) ⊂ Rn ; the group j(Si ) preserves and acts cocompactly on any set Wi × {b} with b ∈ R∗+ (use Fact 2.1). The restriction of f to Wi × {1} may be written as  f (a, 1) = f ′ (a), f ′′ (a) for all a ∈ Wi , where f ′ : Wi → Rn−1 and f ′′ : Wi → R∗+ . Let  L := max 1, Lip(f ′ ) , where Lip(f ′ ) is measured with respect to the Euclidean metric dsRn−1 of Rn−1 , and let Bi′ ⊂ Bi be a horoball Rn−1 × [b0 , +∞), with large b0 > L to be adjusted later. The map fi : Wi × [b0 , +∞) → Hn given by  fi (a, b) := f ′ (a), Lb is (j|Si , ρ|Si )-equivariant, since f is and the groups j(Si ) and ρ(Si ) both preserve the horospheres Rn−1 × {b} (see Figure 4). Moreover, fi is 1-Lipschitz, since by construction it preserves the directions of Rn−1 (horizontal) and R∗+ (vertical) and it stretches by a factor ≤ 1 in the Rn−1 -direction and 1 in the R∗+ -direction, for the hyperbolic metric ds2Rn−1 + db2 . b2 Let Di ⊂ Wi × {1} be a compact fundamental domain for the action of j(Γ0 ) on ∂Bi ∩ Conv(K), and let R := maxx∈Di d((0, 1), x) > 0. ds2 = MAXIMALLY STRETCHED LAMINATIONS 39 fi (x′ ) Bi′ x′ = (a, b) Lb0 fi b0 f (x) (0, 1) x (0, 1) ∂Bi f (∂Bi ) ∂∞ H n ∂∞ H n Figure 4. Definition of a 1-Lipschitz extension fi in the cusp in Step (D). Recall (see (A.2)) that for any (a, b) ∈ Rn−1 × R∗+ ,    kak2 + b2 + 1 d (0, 1), (a, b) = arccosh . 2b In particular,    kak2 +b ≤1 d (0, 1), (a, b) − log b as soon as b exceeds some constant, which we shall assume from now on. Therefore, for any x ∈ Di and x′ = (a, b) ∈ Bi′ ∩ Conv(K),   kak2 ′ + b − 1 − R, (4.5) d(x, x ) ≥ log b and (using the expression of fi , and the fact that f fixes (0, 1) and f ′ is L-Lipschitz)    d f (x), fi (x′ ) ≤ d f (x), f (0, 1) + d f (0, 1), fi (x′ ) (4.6)   kLak2 + Lb + 1 + C ∗ R ≤ log Lb   kak2 + b + log(L) + 1 + C ∗ R. = log b In particular, if Bi′ is far enough from ∂Bi (i.e. b0 > 0 is large enough), then the log term dominates in (4.5) and (4.6) (where b ≥ b0 ), and so d(f (x), fi (x′ )) ≤ C ∗ d(x, x′ ) 40 FRANÇOIS GUÉRITAUD AND FANNY KASSEL for all x ∈ Di and x′ ∈ Bi′ ∩ Conv(K) (recall C ∗ > 1). Therefore, the (j|Si , ρ|Si )-equivariant map  f (1) : Conv(K) r j(Γ0 ) · Bi ∪ (Bi′ ∩ Conv(K)) −→ Hn that agrees with f on Conv(K) r j(Γ0 ) · Bi and with fi on Bi′ ∩ Conv(K) satisfies Lip(f (1) ) ≤ C ∗ . This completes the proof of (D), hence of Proposition 4.16.  5. An optimized, equivariant Kirszbraun–Valentine theorem The goal of this section is to prove the following analogue and extension of Proposition 3.9. We refer to Definitions 2.10 and 4.1 for the notion of stretch locus. We denote by Λj(Γ0 ) ⊂ ∂∞ Hn the limit set of j(Γ0 ). Recall j,ρ that for geometrically finite j, the sets FK,ϕ and EK,ϕ (j, ρ) of Definition 4.1 are nonempty as soon as K is nonempty or ρ is reductive, except possibly if CK,ϕ (j, ρ) = 1 and ρ is not cusp-deteriorating (Lemma 4.10 and Corollary 4.14). Theorem 5.1. Let Γ0 be a discrete group, (j, ρ) ∈ Hom(Γ0 , G)2 a pair of representations of Γ0 in G with j geometrically finite, K a j(Γ0 )-invariant subset of Hn whose image in j(Γ0 )\Hn is compact, and ϕ : K → Hn a (j, ρ)j,ρ equivariant Lipschitz map. Suppose that FK,ϕ and EK,ϕ (j, ρ) are nonempty. Set  Lip(ϕ) if K 6= ∅, C0 := C ′ (j, ρ) if K = ∅, where C ′ (j, ρ) is given by (1.4). • If C0 ≥ 1, then there exists a (j, ρ)-equivariant extension f : Hn → Hn of ϕ with Lipschitz constant C0 , optimal in the sense of Definition 4.12, whose stretch locus is the union of the stretch locus Eϕ of ϕ (defined to be empty if K = ∅) and of a closed set E ′ such that: – if C0 > 1, then E ′ is equal to the closure of a geodesic lamination of Hn r K that is maximally stretched by f , and j(Γ0 )\E ′ is compact; – if C0 = 1, then E ′ is a union of convex sets, each isometrically preserved by f , with extremal points only in the union of K and of the limit set Λj(Γ0 ) ⊂ ∂∞ Hn ; moreover, j(Γ0 )\E ′ is compact provided that ρ is cusp-deteriorating. In particular, in these two cases CK,ϕ (j, ρ) = C0 and EK,ϕ (j, ρ) = Eϕ ∪ E ′ . • If C0 < 1 then CK,ϕ (j, ρ) < 1. By a geodesic lamination of Hn rK we mean a nonempty disjoint union L of geodesic intervals of Hn r K (called leaves), with no endpoint in Hn r K, such that L is closed for the C 1 topology (i.e. any Hausdorff limit of segments of leaves of L is a segment of a leaf of L ). By “maximally stretched by f ” we mean that f multiplies all distances by C0 on any leaf of the lamination. For Γ0 = {1} and K 6= ∅, Theorem 5.1 improves the classical Kirszbraun– Valentine theorem (Proposition 3.1) by adding a control on the local Lipschitz constant of the extension (through a description of its stretch locus). We shall give a proof of Theorem 5.1 in Sections 5.1 to 5.3, and then a proof of Theorem 1.6, as well as Corollary 1.12 under the extra assumption MAXIMALLY STRETCHED LAMINATIONS 41 E(j, ρ) 6= ∅, in Section 5.4 (this extra assumption will be removed in Section 7.5). For K = ∅, we shall finally examine how far the stretch locus E(j, ρ) goes in the cusps in Section 5.6. 5.1. The stretch locus when CK,ϕ (j, ρ) > 1. We now fix (j, ρ) and (K, ϕ) as in Theorem 5.1. To simplify notation, we set C := CK,ϕ (j, ρ) (5.1) E := EK,ϕ (j, ρ) e := E eK,ϕ (j, ρ) E ≥ C0 , ⊂ Hn , ⊂ Hn × Hn e ⊂ Conv(K)×Conv(K) (see Definition 4.1). Recall that E ⊂ Conv(K) and E as soon as C > 0 (Lemma 4.8). In order to prove Theorem 5.1, we first establish the following. Lemma 5.2. In the setting of Theorem 5.1, if C > 1, then E r K is a j,ρ geodesic lamination of Hn r K, and any f ∈ FK,ϕ multiplies arc length by C on the leaves of this lamination. Note that the projection of E to j(Γ0 )\Hn is compact (even in the presence of cusps) by Corollary 4.17. The proof is a refinement of the classical Kirszbraun–Valentine theorem (Proposition 3.1). j,ρ Proof of Lemma 5.2. By Lemma 4.13, there exists an optimal f ∈ FK,ϕ , whose stretch locus is exactly E. Fix p ∈ E r K and consider a small closed ball B ⊂ Hn r K, of radius r > 0, centered at p, which projects injectively to j(Γ0 )\Hn . By Lemma 3.2 with (K, ϕ) := (∂B, f |∂B ), we can find points q ∈ Hn and k1 , k2 ∈ ∂B such that Cq := maxk∈∂B d(q, f (k))/d(p, k) is minimal and such that d(q, f (ki )) = Cq d(p, ki ) for i ∈ {1, 2} and k\ 1 pk2 ≤ \ f (k1 )qf (k2 ) 6= 0. By minimality of Cq , we have Cq ≤ Cf (p) ≤ Lip(f ) = C. We claim that Cq = C. Indeed, suppose by contradiction that Cq < C. Let B ′ ⊂ B be another ball centered at p, of radius r ′ ∈ (0, r) small enough so that Cq r/(r −r ′ ) ≤ C. Then, as in the proof of Lemma 4.15, the extension of f |∂B to ∂B ∪ B ′ which is constant equal to q on B ′ is still C-Lipschitz. Using Proposition 3.1 (see also Remark 3.4), we extend it to a C-Lipschitz j,ρ map f ′ : B → Hn . Working by equivariance, we obtain an element of FK,ϕ , n agreeing with f on H r j(Γ0 ) · B, which is constant on a neighborhood of p, contradicting p ∈ E. Thus Cq = C. We claim that k1 , k2 ∈ ∂B are diametrically opposite, that the geodesic segment [k1 , k2 ] is maximally stretched by f , that f (p) = q, and that the set K′ of points k ∈ ∂B such that d(q, f (k)) = C d(p, k) is reduced to {k1 , k2 }. Indeed, since f is C-Lipschitz, we have d(q, f (k1 )) d(q, f (k2 )) d(f (k1 ), f (k2 )) ≤C= = , d(k1 , k2 ) d(p, k1 ) d(p, k2 ) and since C > 1, Toponogov’s theorem [BH, Lem. II.1.13] implies that \ d(f (k1 ), f (k2 )) = C d(k1 , k2 ) and k\ 1 pk2 = π (the case k1 pk2 = 0 is ruled out since k1 , k2 ∈ ∂B are distinct). In particular, the geodesic segment [k1 , k2 ] has midpoint p and is maximally stretched by f , and f (p) = q 42 FRANÇOIS GUÉRITAUD AND FANNY KASSEL \ by Remark 2.8.(1). For any k ∈ K′ we have k[ 1 pk ≤ f (k1 )qf (k) 6= 0 or \ \ \ [2 ≤ f (k)qf [ kpk (k2 ) 6= 0 (since k[ 1 pk + kpk2 = f (k1 )qf (k) + f (k)qf (k2 ) = π), and so the above reasoning shows that k ∈ {k1 , k2 }. Taking B arbitrarily small, we see that there are exactly two germs of geodesic rays through p in E that are maximally stretched by f , and they are diametrically opposite. Let ℓ be the largest geodesic interval in E r K through p that is maximally stretched by f , corresponding to these two germs. We claim that ℓ terminates on the union of K and of the limit set Λj(Γ0 ) . Indeed, any infinite end of ℓ terminates on Λj(Γ0 ) since ℓ ⊂ Eϕ,K (j, ρ) ⊂ Conv(K) by Lemma 4.8; moreover, ℓ cannot terminate on a point p′ ∈ ErK since there are exactly two germs of geodesic rays through p′ in E that are maximally stretched by f and they are diametrically opposite. This implies that E r K is a geodesic lamination of Hn r K, maximally stretched by f .  It is possible for a point p ∈ K to belong to the stretch locus E without being an endpoint of a leaf of E, or even without belonging to any closed C-stretched segment of f at all (for instance if x 7→ Lipx (ϕ) immediately drops away from p). However, the following holds. Lemma 5.3. In the setting of Theorem 5.1, if C > 1, then any p ∈ E ∩ K lies either in the closure of E r K or in Eϕ′ := {k ∈ K | Lipk (ϕ) = C}. (Note that we have not yet proved that the inequality C0 := Lip(ϕ) ≤ C is an equality; this will be done in Proposition 5.8, and will imply by definition that Eϕ′ is the stretch locus Eϕ of ϕ.) Proof. Suppose C > 1 and consider p ∈ E ∩ K. Assuming p ∈ / Eϕ′ , we shall prove that p lies in the closure of E rK. Since p ∈ / Eϕ′ , there is a small closed ball B of radius r > 0 centered at p, projecting injectively to j(Γ0 )\Hn , such that LipB (ϕ) < C (Lemma 2.9). By Proposition 3.1 and Remark 3.4, the map ϕ|B∩K admits an extension ϕ to B with LipB (ϕ) < C. Consider an j,ρ optimal f ∈ FK,ϕ , whose stretch locus is exactly E (Lemma 4.13), and let C ∗ := sup q∈(K∩B)∪∂B C∗ d(ϕ(p), f (q)) ≤C. d(p, q) We claim that = C. Indeed, suppose by contradiction that C ∗ < C. ′ For any ball B ⊂ B centered at p, with radius r ′ > 0 small enough, the map f ′ : K ∪ ∂B ∪ B ′ → Hn that coincides with f on K ∪ ∂B and with ϕ on B ′ is still C-Lipschitz. Indeed, for any x ∈ K ∪ ∂B and y ∈ B ′ , if x lies in the interior of B then Lip{x,y} (f ′ ) = Lip{x,y} (ϕ) < C, and otherwise the triangle inequality gives C ∗ r + Lip(ϕ)r ′ d(f ′ (x), f ′ (y)) ≤ d(x, y) r − r′ ′ as in (3.1), which is ≤ C if r is small enough. Therefore f ′ admits a CLipschitz extension to B by Remark 3.4; working by equivariance, we obtain j,ρ an element f ′ ∈ FK,ϕ agreeing with f on Hn rj(Γ0 )·B, such that Lipp (f ′ ) ≤ Lip(ϕ) < C, contradicting p ∈ E. Thus C ∗ = C. If the upper bound C ∗ is approached by a sequence (qi )i∈N of (K ∩B)∪∂B with qi → p, then qi ∈ K for all large enough i and p ∈ Eϕ′ , contradicting MAXIMALLY STRETCHED LAMINATIONS 43 the assumption. Therefore (qi )i∈N has an accumulation point q 6= p. The geodesic segment [p, q] is maximally stretched by f (Remark 2.8.(1)), hence [p, q] r K ⊂ E r K and any accumulation point of [p, q] ∩ K lies in Eϕ′ . Since  p∈ / Eϕ′ , we obtain that p lies in the closure of E r K. e as in 5.2. The stretch locus when CK,ϕ (j, ρ) = 1. We define C, E, E (5.1). When C = 1, the stretch locus E may contain pieces larger than j,ρ lines that are isometrically preserved by all elements of FK,ϕ . Here is the counterpart of Lemma 5.2 in this case. Lemma 5.4. In the setting of Theorem 5.1, if C = 1, then there is a canonical family (Ωp )p∈E of closed convex subsets of Hn , of varying dimensions, with the following properties: (i) any p ∈ E lies in the interior of the corresponding Ωp (where we see Ωp as a subset of its own affine span — in particular, a point is equal to its own interior); (ii) the interiors of Ωp and Ωq are either equal or disjoint for p, q ∈ E; j,ρ (iii) the restriction to Ωp of any f ∈ FK,ϕ is an isometry; n (iv) whenever two points x 6= y in H satisfy d(f (x), f (y)) = d(x, y) for j,ρ some (hence any) optimal f ∈ FK,ϕ (Definition 4.12), the geodesic segment [x, y] (called a 1-stretched segment) is contained in some Ωp ; (v) all extremal points of Ωp lie in the union of K and of the limit set Λj(Γ0 ) of j(Γ0 ); (vi) the intersection of Ωp with any supporting hyperplane is an Ωq ; S (vii) E = p∈ErK Ωp ∪ Eϕ′ where Eϕ′ = {k ∈ K | Lipk (ϕ) = 1}. Properties (i)–(vii) are reminiscent of the stratification of the boundary of a convex object, with 1-stretched segments of E replacing segments contained in the boundary of the convex object; we shall call the interiors of the sets Ωp strata of E, and the sets Ωp closed strata. Remark 5.5. In dimension n ≥ 3, the connected components of E = EK,ϕ (j, ρ) can be nonconvex. Indeed, take n = 3. Let Γ0 be the fundamental group of a closed surface, let j ∈ Hom(Γ0 , G) be geometrically finite, obtained by bending slightly a geodesic copy of H2 inside H3 along some geodesic lamination L , and let ρ ∈ Hom(Γ0 , G) be obtained by bending even a little more along the same lamination L . Then E is the first bent copy of H2 , which can be nonconvex (though connected). j,ρ Proof of Lemma 5.4. Consider an optimal f ∈ FK,ϕ (Lemma 4.13): by definition, the stretch locus Ef = E is minimal, and so is the enhanced stretch ef = E. e locus E For p ∈ E, let Wp ⊂ P(Tp Hn ) be the set of directions of 1-stretched segments containing p in their interior. This set is independent of f because ef = E. e (It is for this independence property that we use the enhanced E e here.) Since f is 1-Lipschitz, the convex hull of any two stretch locus E such 1-stretched segments is isometrically preserved by f . Therefore the set Wp is a full projective subspace (possibly empty), equal to the projectivization of a vector subspace Vp ⊂ Tp Hn . Moreover, there is a neighborhood of p in expp (Vp ) on which f coincides with an isometric embedding 44 FRANÇOIS GUÉRITAUD AND FANNY KASSEL ψp : expp (Vp ) → Hn , and the closed set  Ωp := x ∈ expp (Vp ) | f (x) = ψp (x) ⊂ E is convex and contains p in its interior. The isometric embedding ψp may depend on f , but the set Ωp depends only on the data (j, ρ, K, ϕ), because so e We shall denote by dp ≥ 0 the dimension does the enhanced stretch locus E. of Vp . Conditions (i) and (iii) are satisfied by construction, and so is (iv) by taking p in the interior of the given 1-stretched segment [x, y]. (Note that Ωp may contain points of K in its interior, even when p ∈ / K.) For any x belonging to the interior of Ωp in expp (Vp ), we have Vx = Tx Ωp . Indeed, Vx ⊃ Tx Ωp is clear since f |Int(Ωp ) is an isometry; and if x were in the interior of any 1-stretched segment s not contained in expp (Vp ), then f would be isometric on the (dp +1)-dimensional convex hull of Ωp ∪s, which contains p in its interior: this would violate the definition of Vp . From Vx = Tx Ωp we deduce in particular ψx = ψp and Ωx = Ωp . It follows that given q ∈ E, if the interiors of Ωp and Ωq intersect at a point x, then ψp = ψx = ψq and Ωp = Ωq : thus (ii) holds. Any 1-stretched segment s = [x, y] with an interior point q in Ωp is cond′ and yqp d′ for p′ ∈ Ωp , tained in Ωp . Indeed, f must preserve all angles xqp hence f is an isometry on the convex hull of s ∪ Ωp , which contains p in its interior: therefore s ⊂ expp (Vp ) by definition of dp and s ⊂ Ωp by definition of Ωp . In expp (Vp ), the intersection of Ωp with any supporting hyperplane Π at a point of ∂Ωp is the closure of an open convex subset Q of some Hd , where 0 ≤ d < dp (with H0 being a point). Consider a point q ∈ Q. By the previous paragraph, any open 1-stretched segment through q is in Ωp , hence in Π, hence in Q (see Figure 5). Therefore, d = dq and ψq = ψp |expq (Vq ) . It follows that Ωq is the closure of Q. This gives (vi). Ωp p Π q Q Figure 5. A 3-dimensional convex stratum Ωp with a supporting plane Π. By (vi), extremal points q ∈ Ωp inHn satisfy Ωq = {q}, hence q ∈ K by Lemma 5.6 below; this gives (v). MAXIMALLY STRETCHED LAMINATIONS 45 We obtain (vii) by replacing C by 1 and the call to Proposition 3.1 by Proposition 3.7 in the proof of Lemma 5.3.  Lemma 5.6. Any p ∈ E r K is contained in the interior of a 1-stretched segment. j,ρ Proof. Consider an optimal f ∈ FK,ϕ (Lemma 4.13), whose stretch locus is exactly E. Fix p ∈ E r K and consider a small closed ball B ⊂ Hn rK, centered at p, which projects injectively to j(Γ0 )\Hn . By Lemma 3.2 with (K, ϕ) := (∂B, f |∂B ), we can find a point q ∈ Hn such that Cq := maxk∈∂B d(q, f (k))/d(p, k) is minimal. In particular, Cq ≤ Cf (p) ≤ Lip(f ) = 1. In fact Cq = 1: otherwise, using Proposition 3.1 as in the proofs of Lemmas j,ρ 4.15 and 5.2, we could construct an element of FK,ϕ that would be locally constant near p, contradicting the fact that p ∈ E. In particular, there cannot exist k1 , k2 ∈ ∂B such that d(q, f (ki )) = \ Cq d(p, ki ) for i ∈ {1, 2} and 0 ≤ k\ 1 pk2 < f (k1 )qf (k2 ) ≤ π: otherwise we would have d(f (k1 ), f (k2 )) > d(k1 , k2 ) by basic trigonometry, contradicting Lip(f ) = 1. Therefore, Lemma 3.2 implies the existence of a probability measure ν on K′ := {k ∈ ∂B | d(q, f (k)) = d(p, k)} such that q belongs to the convex hull \ of the support of f∗ ν and such that k\ 1 pk2 = f (k1 )qf (k2 ) for (ν × ν)-almost ′ ′ all (k1 , k2 ) ∈ K × K . This means that the continuous map f : Hn → Hn is an isometry on the support of ν (which is contained in ∂B), hence has a unique 1-Lipschitz extension (the isometric one, with which f must agree) to the convex hull X of this support. Since f (X) contains q, we have f (p) = q and there is a 1-stretched segment through p. This completes the proof of Lemma 5.6, hence of Lemma 5.4.  Remark 5.7. When K = ∅, the closed strata Ωp of the lowest dimension (say k ≥ 1) are always complete copies of Hk : otherwise, they would by Lemma 5.4.(vi) admit supporting planes giving rise to closed strata of lower dimension. In particular, the union of these closed strata is a k-dimensional geodesic lamination in the sense of Section 1.2. In dimension n = 2, we must have k = 1 (unless j and ρ are conjugate); this implies that the stretch locus is the union of a geodesic lamination and (possibly) certain connected components of its complement. 5.3. Proof of Theorem 5.1. Theorem 5.1 is an immediate consequence of Corollary 4.17, of Lemmas 5.2 and 5.4, and of the following proposition. Proposition 5.8. In the setting of Theorem 5.1, if CK,ϕ (j, ρ) ≥ 1, then CK,ϕ (j, ρ) = C0 . In particular, C0 < 1 implies CK,ϕ (j, ρ) < 1. Proof. Recall that C0 ≤ C := CK,ϕ (j, ρ) (see (5.1)). Let us prove the converse inequality when C ≥ 1. In the particular situation where C := CK,ϕ (j, ρ) = 1 and ρ is not cuspdeteriorating, we have C0 ≥ 1 by Lemma 2.7. We now assume that we are not in this particular situation. By Corollary 4.14, the set E := EK,ϕ (j, ρ) is nonempty. Suppose E ⊂ K. By Lemma 5.3 or Lemma 5.4.(vii), for any p ∈ E ⊂ K we have Lipp (ϕ) = C, and so C0 = Lip(ϕ) ≥ C. 46 FRANÇOIS GUÉRITAUD AND FANNY KASSEL Suppose E 6⊂ K. By Lemma 5.2 or Lemma 5.4.(iii)–(v), any point p ∈ E r K belongs, either to the convex hull of a subset of K on which f multiplies all distances by C, or to a maximally stretched geodesic ray with an endpoint in Λj(Γ0 ) whose image in the quotient j(Γ0 )\Hn is bounded (Proposition 4.16.(1)). In the first case, C0 = Lip(ϕ) ≥ C since f coincides with ϕ on K. In the second case, Lemma 4.6 yields C0 ≥ C.  Theorem 1.3 is contained in Lemmas 4.10 and 4.13, Corollary 4.14, Theorem 5.1, and Remark 5.7. 5.4. Some easy consequences of Theorem 5.1. We first prove Theorem 1.6, which concerns the case where K is nonempty and possibly noncompact modulo j(Γ0 ). Proof of Theorem 1.6. Let K 6= ∅ be a j(Γ0 )-invariant subset of Hn . We can always extend ϕ to the closure K of K by continuity, with the same Lipschitz constant C0 . Suppose the image of K in j(Γ0 )\Hn is compact. If C0 ≥ 1, then Theorem 1.6 is contained in Theorem 5.1. If C0 < 1, then C(j, ρ) < 1 by Theorem 5.1, which implies Theorem 1.6 since F j,ρ 6= ∅ (Lemma 4.10). Now, for C0 ≥ 1, consider the general case where the image of K in j(Γ0 )\Hn is not necessarily compact. Let (Ck )k∈N be a sequence of j(Γ0 )-invariant subsets of Hn whose images in j(Γ0 )\Hn are compact, S with Ck ⊂ Ck+1 and k∈N Ck = Hn . For any k, Theorem 5.1 gives a (j, ρ)equivariant extension fk : Hn → Hn of ϕ|K∩Ck with Lip(fk ) = C0 , and we conclude using the Arzelà–Ascoli theorem as in Remark 3.6.  We then turn to Corollary 1.12, for which K is empty. Corollary 1.12 for E(j, ρ) 6= ∅ is an immediate consequence of Lemma 4.6 and Theorem 5.1. For E(j, ρ) = ∅, we shall prove Corollary 1.12 in Section 7.5 (Lemma 7.4), using a Cartan projection µ of G. Recall however from Corollary 4.14 that E(j, ρ) = ∅ may only happen if C(j, ρ) = 1 and ρ is not cusp-deteriorating; in that case, a direct proof of Corollary 1.12 could also be obtained by considering a sequence of closed geodesics of j(Γ0 )\Hn that spend more and more time in a cusp whose stabilizer contains an element γ ∈ Γ0 with both j(γ) and ρ(γ) parabolic. In dimension n = 2, for torsion-free Γ0 , let Cs′ (j, ρ) be the supremum of λ(ρ(γ))/λ(j(γ)) over all elements γ ∈ Γ0 corresponding to simple closed curves in the hyperbolic surface j(Γ0 )\H2 . (As for C ′ (j, ρ), we define Cs′ (j, ρ) to be C(j, ρ) in the degenerate case when j(Γ0 )\H2 has no essential closed curve.) Then Cs′ (j, ρ) ≤ C ′ (j, ρ) ≤ C(j, ρ) (see (4.1)). Here is another consequence of Theorem 5.1. Lemma 5.9 (n = 2, torsion-free Γ0 ). Suppose E(j, ρ) 6= ∅. If Cs′ (j, ρ) < 1, then C ′ (j, ρ) < 1. Proof. If C(j, ρ) ≥ 1, then (5.2) Cs′ (j, ρ) = C ′ (j, ρ) = C(j, ρ). Indeed, by Theorem 5.1 and Remark 5.7, the image of E(j, ρ) in j(Γ0 )\Hn contains a nonempty geodesic lamination L with compact image. If L contains a simple closed curve, then (5.2) is clear; otherwise we can argue as MAXIMALLY STRETCHED LAMINATIONS 47 in the proof of Lemma 4.6, but with the axis of j(γ) projecting to a simple closed geodesic nearly carried by L .  Note that if E(j, ρ) = ∅, then it is possible to have Cs′ (j, ρ) < 1 = = C(j, ρ): see Section 10.8. C ′ (j, ρ) 5.5. The recurrent set of maximally stretched laminations. For empty K, Theorem 5.1 states that the stretch locus E(j, ρ) contains a maximally stretched k-dimensional geodesic lamination with compact (nonempty) imj,ρ and EK,ϕ (j, ρ) are nonempty and C ′ (j, ρ) ≥ 1. age in j(Γ0 )\Hn as soon as FK,ϕ Conversely, we make the following observation. Lemma 5.10. Let Γ0 be a discrete group and (j, ρ) ∈ Hom(Γ0 , G)2 a pair of representations with j geometrically finite. Let L be a j(Γ0 )-invariant kdimensional geodesic lamination of Hn with a compact image in j(Γ0 )\Hn . If L is maximally stretched by some (j, ρ)-equivariant Lipschitz map f : Hn → Hn , then the recurrent set of L is contained in the stretch locus E(j, ρ). By recurrent set of L , we mean the projection to j(Γ0 )\Hn of the recurrent set of the geodesic flow (Φt )t∈R restricted to vectors tangent to L . By compactness, this recurrent set is nonempty. Recall that in this setting we have C ′ (j, ρ) = C(j, ρ) = Lip(f ) by Lemma 4.6. Proof. Set C := C(j, ρ) = Lip(f ). In order to prove that the recurrent set of L is contained in E(j, ρ), it is sufficient to prove that for any geodesic line (pt )t∈R of Hn contained in L and projecting to a geodesic which is recurrent in j(Γ0 )\Hn , we have d(f ′ (p0 ), f ′ (p1 )) = C = C d(p0 , p1 ) for all f ′ ∈ F j,ρ (hence [p0 , p1 ] ⊂ E(j, ρ)). Fix ε > 0. By recurrence, we can find γ ∈ Γ0 and t > 1 with [j(γ) · p0 , j(γ) · p1 ] arbitrarily close to [pt , pt+1 ], so that by the closing lemma (Lemma A.1) the translation axis of j(γ) passes within ε of the four points p0 , p1 , pt , pt+1 and the axis of ρ(γ) within Cε of their four images under f . Choose q0 , q1 ∈ Hn within ε of p0 , p1 , respectively, on the axis of j(γ). For any f ′ ∈ F j,ρ ,   d(f ′ (q0 ), f ′ (q1 )) ≥ d f ′ (q0 ), ρ(γ) · f ′ (q0 ) − d ρ(γ) · f ′ (q0 ), f ′ (q1 )  ≥ λ(ρ(γ)) − d f ′ (j(γ) · q0 ), f ′ (q1 )  ≥ C · (λ(j(γ)) − 4ε) − Lip(f ′ ) · λ(j(γ)) − d(q0 , q1 ) = C · (d(q0 , q1 ) − 4ε) since Lip(f ′ ) = C. But p0 , p1 are ε-close to q0 , q1 ; therefore d(f ′ (p0 ), f ′ (p1 )) ≥ ≥ ≥ ≥ d(f ′ (q0 ), f ′ (q1 )) − d(f ′ (p0 ), f ′ (q0 )) − d(f ′ (p1 ), f ′ (q1 )) C · (d(q0 , q1 ) − 4ε) − 2 Lip(f ′ ) ε C · (d(p0 , p1 ) − d(p0 , q0 ) − d(p1 , q1 ) − 4ε) − 2Cε C · (1 − 8ε). This holds for any ε > 0, hence d(f ′ (p0 ), f ′ (p1 )) = C.  48 FRANÇOIS GUÉRITAUD AND FANNY KASSEL 5.6. How far the stretch locus goes into the cusps. Suppose that j is geometrically finite but not convex cocompact. For empty K, we can control how far the stretch locus EK,ϕ (j, ρ) = E(j, ρ) goes into the cusps. Proposition 5.11. There is a nondecreasing function Ψ : (1, +∞) → R∗+ such that for any discrete group Γ0 , any pair (j, ρ) ∈ Hom(Γ0 , G)2 with j geometrically finite and C(j, ρ) > 1, and any x ∈ E(j, ρ) whose image in j(Γ0 )\Hn belongs to a standard cusp region of the convex core (Definition 2.2), the cusp thickness at x is ≥ Ψ(C(j, ρ)). Here we use the following terminology, where N ⊂ Hn is the preimage of the convex core of j(Γ0 )\Hn . Definition 5.12. Let B be a horoball of Hn such that B ∩ N projects to a standard cusp region of j(Γ0 )\Hn . Given a point x ∈ B, let ∂Bx denote the horosphere concentric to B running through x. The cusp thickness of j(Γ0 )\Hn at x is the Euclidean diameter in j(Stab(B))\∂Bx of the orthogonal projection of N to ∂Bx . By Euclidean diameter we mean the diameter for the metric induced by the intrinsic, Euclidean metric of ∂Bx ; it varies exponentially with the depth of x in the cusp region (see (A.3) for conversion to a hyperbolic distance). Note that the orthogonal projection of N is convex inside the Euclidean space ∂Bx ≃ Rn−1 . We believe that an analogue of Proposition 5.11 should also hold for C(j, ρ) < 1, see Appendix C.4. It is false for C(j, ρ) = 1 (take j = ρ). Proposition 5.11 will be a consequence of the following lemma, which applies to C = C(j, ρ) and to leaves ℓ0 , ℓ1 of the geodesic lamination E(j, ρ). It implies that any two leaves of E(j, ρ) coming close to each other must be nearly parallel. This is always the behavior of simple closed curves and geodesic laminations in dimension n = 2. Lemma 5.13. For any C > 1, there exists δ0 > 0 with the following property. Let ℓ0 , ℓ1 be disjoint geodesic lines of Hn . Suppose there exists a C-Lipschitz map f : ℓ0 ∪ ℓ1 → Hn multiplying all distances by C on ℓ0 and on ℓ1 . If ℓ0 and ℓ1 pass within δ ≤ δ0 of each other near some point x ∈ Hn , then they stay within distance 1 of each other on a length ≥ | log δ| − 10 before and after x. (The constant 10 is of course far from optimal.) Proof. We can restrict to dimension n = 3 because the geodesic span of two lines has dimension at most 3. Fix C > 1 and let ℓ0 , ℓ1 , and f be as above. The images ℓ′0 := f (ℓ0 ) and ℓ′1 := f (ℓ1 ) are geodesic lines of H3 . Fix orientations on ℓ0 , ℓ1 , ℓ′0 , ℓ′1 so that f is orientation-preserving. For i ∈ {0, 1}, let xi be a point of ℓi closest to ℓ1−i , so that the geodesic segment [x0 , x1 ] is orthogonal to both ℓ0 and ℓ1 ; let σ be the rotational symmetry of H3 around the line (x0 , x1 ). Similarly, let x′i ∈ ℓ′i be closest to ℓ′1−i , so that the segment [x′0 , x′1 ] is orthogonal to ℓ′0 and ℓ′1 ; let σ ′ be the rotational symmetry of H3 around (x′0 , x′1 ). Up to replacing f by 1 1 f + σ ′ ◦ f ◦ σ, 2 2 MAXIMALLY STRETCHED LAMINATIONS 49 which is still C-Lipschitz (Lemma 2.13), which preserves the orientations of ℓ0 , ℓ1 , ℓ′0 , ℓ′1 , and which multiplies all distances by C on ℓ0 and on ℓ1 , we may assume that f (x0 ) = x′0 and f (x1 ) = x′1 . Let η (resp. η ′ ) be the length of [x0 , x1 ] (resp. of [x′0 , x′1 ]), and θ (resp. θ ′ ) the angle between the positive directions of ℓ0 and ℓ1 (resp. of ℓ′0 and ℓ′1 ), measured by projecting orthogonally to a plane perpendicular to [x0 , x1 ] (resp. to [x′0 , x′1 ] if η ′ > 0). We claim that (∗) there exists ∆0 > 0, depending only on C, such that if η ≤ ∆0 , then min{θ, π − θ} ≤ 1.005 η. Indeed, for i ∈ {0, 1}, let ti > 0 be the linear coordinate of a point pi ∈ ℓi , measured from xi with the chosen orientation. By (A.7), (5.3) cosh d(p0 , p1 ) = cosh η · cosh t0 cosh t1 − cos θ · sinh t0 sinh t1 . Therefore, using cosh t ∼ et /2, we obtain that for t0 , t1 → +∞,   cosh η − cos θ d(p0 , p1 ) = t0 + t1 + log + o(1). 2 Similarly, since f stretches ℓ0 and ℓ1 by a factor of C and f (xi ) = x′i ,    cosh η ′ − cos θ ′ + o(1). d f (p0 ), f (p1 ) = Ct0 + Ct1 + log 2 Since f is C-Lipschitz, we must have     cosh η ′ − cos θ ′ cosh η − cos θ log ≤ C log . 2 2 Note that this must also hold if we replace θ, θ ′ with their complements to π, because we can reverse the orientations of ℓ1 and ℓ′1 . We thus obtain   cosh η ± cos θ C cosh η ′ ± cos θ ′ ≤ . 2 2 Since cosh η ′ ≥ 1, adding the two inequalities yields     cosh η + cos θ C cosh η − cos θ C (5.4) + ≥ 1. 2 2 Inequality (5.4) means that (cos θ, cosh η) lies in R2 outside of a π4 -rotated √ and 2-scaled copy of the unit ball of R2 for the LC -norm. (See also (A.10) for an interpretation of (5.4) in terms of the cross-ratio of the endpoints of ℓ0 , ℓ1 .) Since cos θ ∈ [−1, 1] and cosh η ≥ 1, we obtain that (cos θ, cosh η) lies above some concave curve through the points (−1, 1) and (1, 1), with respective slopes 1 and −1 at these points (recall that C > 1). In particular, if cosh η is very close to 1, then | cos θ| must be about as close (or closer) to 1 (see Figure 6). We obtain (∗) by using the Taylor expansions of cosh and cos (of course 1.005 can be replaced by any number > 1). To deduce the lemma from (∗), we can minimize (5.3) in t1 alone to find sinh2 d(p0 , ℓ1 ) = sinh2 η + (sinh2 η + sin2 θ) sinh2 t0 . By (∗), for small enough η we have η 2 ≤ sinh2 η + sin2 θ ≤ 2.004 η 2 , 50 FRANÇOIS GUÉRITAUD AND FANNY KASSEL p1 ℓ1 x1 η x0 ℓ0 θ p0 Figure 6. At distance d(p0 , x0 ) = t0 < | log η| from a point of closest approach of the two lines ℓ0 , ℓ1 , their angular drift ≈ θet0 cannot much exceed their height drift ≈ ηet0 . hence (5.5) sinh2 η + η 2 sinh2 t0 ≤ sinh2 d(p0 , ℓ1 ) ≤ sinh2 η + 2.004 η 2 sinh2 t0 . If t0 ∈ [0, | log η|] (for small η), we have sinh2 η sinh2 t0 2.004 + 2.004 ≤ 1.005 + = 1.506, 2 2t 2t 0 0 η e e 4 hence, on the upper side of (5.5), √  1.506 ηet0 (5.6) sinh2 η + 2.004 η 2 sinh2 t0 ≤ 1.506 η 2 e2t0 ≤ sinh2 2 2t0 and using the inequality x ≤ sinh x for x ∈ R . by multiplying + √ by η e Note that 1.506 ≤ 1.23. On the other hand, using again sinh x ≥ x, sinh2 η sinh2 t0 cosh2 t0 1 + ≥ ≥ , 2 2t 2t 2t 0 0 0 η e e e 4 hence, on the lower side of (5.5), (5.7) sinh2 η + η 2 sinh2 t0 ≥ (ηet0 sinh 0.48)2 ≥ sinh2 (0.48 η et0 ) by multiplying by η 2 e2t0 and using the inequality sinh2 0.48 < 1/4 and the convexity of sinh (recall ηet0 ≤ 1). From (5.5), (5.6), and (5.7), it follows that for η smaller than some δ0 ∈ (0, 1) (depending only on C), 0.48 η e|t0 | ≤ d(p0 , ℓ1 ) ≤ 1.23 η e|t0 | as soon as |t0 | ≤ | log η|. This two-sided exponential bound means that p0 7→ log d(p0 , ℓ1 ) is essentially a 1-Lipschitz function of p0 (plus a bounded correction), which easily implies the lemma.  Proof of Proposition 5.11. Let Γ0 be a discrete group, (j, ρ) ∈ Hom(Γ0 , G)2 a pair of representations with j geometrically finite, and B an open horoball of Hn whose image in j(Γ0 )\Hn intersects the convex core in a standard cusp region. The stabilizer S ⊂ Γ0 of B under j has a normal subgroup S ′ isomorphic to Zm for some 0 < m < n, and of index ≤ ν(n) in S, where ν(n) < +∞ depends only on n (see Section 2.1). In the upper half-space model Rn−1 × R∗+ of Hn , where ∂∞ Hn identifies with Rn−1 ∪ {∞}, we may assume that B is centered at ∞. Let Ω be the convex hull of Λj(Γ0 ) r {∞} in Rn−1 , where Λj(Γ0 ) is the limit set of j(Γ0 ). The ratio of the Euclidean diameter of j(S ′ )\Ω to that of j(S)\Ω is bounded by 2ν(n). We renormalize the metric on Rn−1 so that j(S ′ )\Ω has Euclidean diameter 1: then, by MAXIMALLY STRETCHED LAMINATIONS 51 definition of cusp thickness, it is sufficient to prove that the height of points of E(j, ρ) in Rn−1 × R∗+ is bounded in terms of C(j, ρ) alone. There is an m-dimensional affine subspace V ⊂ Ω of Rn−1 which is preserved by j(S ′ ) and on which j(S ′ ) acts as a lattice of translations (see Section 2.1). Any point of Ω lies within distance 1 of V . If C(j, ρ) > 1, then by Theorem 5.1 the stretch locus E(j, ρ) is a disjoint union of geodesic lines of Hn . Let ℓ ⊂ E(j, ρ) be such a line, reaching a height h in the upper half-space model. We must bound h. The endpoints ξ, η ∈ Ω of ℓ are 2h apart in Rn−1 . Let ξ ′ , η ′ ∈ V be within distance 1 from ′ ′ ξ, η respectively. There exists γ ∈ S ′ such that dRn−1 (j(γ) · ξ ′ , ξ +η 2 ) ≤ 1. Since ξ ′ , η ′ and their images under j(γ) form a parallelogram, we also have ′ ′ ′ dRn−1 (j(γ) · ξ +η 2 , η ) ≤ 1. By the triangle inequality,     ξ+η ξ +η , η ≤ 3. dRn−1 j(γ) · ξ , ≤ 3 and dRn−1 j(γ) · 2 2 and j(γ) · 3ξ+η are at Euclidean Adding up, it follows that the points ξ+3η 4 4 distance ≤ 3 from √ each other. But the leaves ℓ and j(γ) · ℓ of E(j, ρ) contain √ points at height h 3/2 above these two points, and are therefore ≤ 2 3/h apart in the hyperbolic metric. However, ℓ and j(γ) · ℓ form an angle close to π/3 (see Figure 7): by Lemma 5.13 (or (∗) in its proof), this places an upper bound on h (depending only on C(j, ρ)).  j(γ) · ℓ ℓ h ξ j(γ) · ξ ′ Cusp group j(S ) η V Ω j(γ) · η ∂∞ Hn Figure 7. Two leaves ℓ and j(γ) · ℓ which nearly intersect, at an angle close to π/3 (in the upper half-space model of Hn ). In Section 6.4, in order to prove the upper semicontinuity of (j, ρ) 7→ C(j, ρ) where C ≥ 1 when all the cusps of j have rank ≥ n − 2, we shall need the following consequence of Proposition 5.11. Corollary 5.14. Let Γ0 be a discrete group and (jk , ρk )k∈N∗ a sequence of elements of Hom(Γ0 , G)2 converging to some (j, ρ) ∈ Hom(Γ0 , G)2 , where • j and the jk are all geometrically finite, of the same cusp type, with all cusps of rank ≥ n − 2, • there exists C ∗ > 1 such that C(jk , ρk ) ≥ C ∗ for all k ∈ N∗ , • the stretch loci E(jk , ρk ) are nonempty (this is the case for instance if ρk is reductive, see Lemma 4.10 and Corollary 4.14). Then for any k ∈ N∗ we can find a fundamental domain Ek of E(jk , ρk ) for the action of jk (Γ0 ) so that all the Ek are contained in some compact subset of Hn independent of k. 52 FRANÇOIS GUÉRITAUD AND FANNY KASSEL Proof. By Proposition B.3, there exist a compact set C ⊂ Hn and, for any large enough k ∈ N∗ , horoballs H1k , . . . , Hck of Hn , such that the union G of all geodesic rays from C to the centers of H1k , . . . , Hck contains a fundamental n domain of the convex core of S jk (Γ0 )\Hk . In particular, the cusp thickness of n jk (Γ0 )\H at any point of 1≤i≤c ∂Hi is uniformly bounded from above by some constant independent of k. On the other hand, by Proposition 5.11, the cusp thickness of jk (Γ0 )\Hn at any point of E(jk , ρk ) is uniformly bounded from below by some constant independent of k. Since cusp thickness decreases uniformly to 0 in all cusps (at exponential rate), this means that E(jk , ρk ) ∩ G (which contains a fundamental domain of E(jk , ρk ) for the action of jk (Γ0 )) remains in some compact subset of Hn independent of k.  6. Continuity of the minimal Lipschitz constant In this section we examine the continuity of the function (j, ρ) 7→ C(j, ρ) for geometrically finite j (the set K of Sections 4 and 5 is empty). We endow Hom(Γ0 , G) with its natural topology: a sequence (jk , ρk ) converges to (j, ρ) if and only if jk (γ) → j(γ) and ρk (γ) → ρ(γ) for all γ in some (hence any) finite generating subset of Γ0 . We first prove Proposition 1.5, which states the continuity of (j, ρ) 7→ C(j, ρ) for convex cocompact j. When j is not convex cocompact, continuity, and even semicontinuity, fail in any dimension n ≥ 2: see Sections 10.6 and 10.7 for counterexamples. However, we will prove the following. Proposition 6.1. Let Γ0 be a discrete group and j0 ∈ Hom(Γ0 , G) a geometrically finite representation. If all the cusps of j0 have rank ≥ n − 2 (for instance if we are in dimension n ≤ 3), then (1) the set of pairs (j, ρ) with C(j, ρ) < 1 is open in Homj0 (Γ0 , G) × Homj0 -det (Γ0 , G), (2) the set of pairs (j, ρ) with 1 < C(j, ρ) is open in Homj0 (Γ0 , G) × Hom(Γ0 , G), (3) the map (j, ρ) 7→ C(j, ρ) is continuous on the set of pairs (j, ρ) ∈ Homj0 (Γ0 , G) × Hom(Γ0 , G) with 1 ≤ C(j, ρ) < +∞. If the cusps of j0 have arbitrary ranks, then condition (2) holds, as well as: (1’) the set of ρ with C(j0 , ρ) < 1 is open in Homj0 -det (Γ0 , G), (3’) the map (j, ρ) 7→ C(j, ρ) is lower semicontinuous on the set of pairs (j, ρ) ∈ Homj0 (Γ0 , G) × Hom(Γ0 , G) with 1 ≤ C(j, ρ): C(j, ρ) ≤ lim inf C(jk , ρk ) k for any sequence (jk , ρk ) of such pairs converging to such a pair (j, ρ), (3”) the map ρ 7→ C(j0 , ρ) is upper semicontinuous on the set of representations ρ ∈ Hom(Γ0 , G) with 1 ≤ C(j0 , ρ) < +∞: C(j0 , ρ) ≥ lim sup C(j0 , ρk ) k for any sequence (ρk ) of such representations converging to such a representation ρ. Here we denote by MAXIMALLY STRETCHED LAMINATIONS 53 • Homj0 (Γ0 , G) the space of geometrically finite representations of Γ0 in G with the same cusp type as the fixed representation j0 ; • Homj0 -det (Γ0 , G) the space of representations that are cusp-deteriorating with respect to j0 , in the sense of Definition 1.1. These two sets are endowed with the induced topology from Hom(Γ0 , G). In (3)–(3’)–(3”), we endow the set of pairs (j, ρ) satisfying 1 ≤ C(j, ρ) < +∞ or 1 ≤ C(j, ρ) with the induced topology from Hom(Γ0 , G)2 . Note that Homj0 -det (Γ0 , G) is a semi-algebraic subset of Hom(Γ0 , G); it is equal to Hom(Γ0 , G) if and only if j0 is convex cocompact. When j0 is not convex cocompact, the condition C(j, ρ) < 1 is not open in Homj0 (Γ0 , G)×Hom(Γ0 , G) or even in {j0 }×Hom(Γ0 , G), since the constant representation ρ (for which C(j, ρ) = 0) may be approached by non-cuspdeteriorating representations ρ (for which C(j, ρ) ≥ 1); see also Section 10.6 for a related example. This is why we need to restrict to cusp-deteriorating ρ in Proposition 6.1.(1). In dimension n ≥ 4, when j0 has cusps of rank < n − 2, conditions (1) and (3) of Proposition 6.1 do not hold: see Sections 10.10 and 10.11 for counterexamples. The reason, in a sense, is that the convex core of a small geometrically finite deformation of j can be “much larger” than that of j, due to the presence of parabolic elements that are not unipotent. (Such discontinuous behavior of the convex core also explains why being geometrically finite is not an open condition in the presence of cusps of rank < n − 2, even among representations of a given cusp type [Bo2, § 5].) Note finally that C(j, ρ) = +∞ must be ruled out in (3) and (3”) due to Lemma 4.7: parabolic elements can be approached by hyperbolic ones. Proposition 1.5 is proved in Section 6.1 using a partition-of-unity argument based on Lemma 2.14, together with a control on fundamental domains for converging convex cocompact representations (see Appendix B). Proposition 6.1.(1)–(1’) is proved in Section 6.2 following the same approach but using also a comparison between distances in horospheres and spheres of Hn (Lemma 6.4). Proposition 6.1.(2) and (3)–(3’)–(3”) are proved in Section 6.4; for reductive ρ, they are a consequence of the existence of a maximally stretched lamination when C(j, ρ) ≥ 1 (Theorem 1.3). The case of nonreductive ρ follows from the reductive case by using again a partition-of-unity argument, as we explain in Section 6.3. 6.1. Continuity in the convex cocompact case. In this section we prove Proposition 1.5. We fix a pair of representations (j, ρ) ∈ Hom(Γ0 , G)2 with j convex cocompact and a sequence (jk , ρk )k∈N∗ of elements of Hom(Γ0 , G)2 converging to (j, ρ). We may and shall assume that Γ0 is torsion-free, using Lemma 4.4 and the Selberg lemma [Se, Lem. 8]. 6.1.1. Upper semicontinuity. We first prove that C(j, ρ) ≥ lim sup C(jk , ρk ). k→+∞ Hn Hn Fix ε > 0 and let f : → be a (j, ρ)-equivariant, (C(j, ρ) + ε)-Lipschitz map. We explain how for any large enough k we can modify f into a (jk , ρk )equivariant map fk with Lip(fk ) ≤ Lip(f ) + ε. By Lemma 4.8, we only need 54 FRANÇOIS GUÉRITAUD AND FANNY KASSEL to define fk on the preimage Nk ⊂ Hn of the convex core of jk (Γ0 )\Hn . In order to build fk , we will paste together shifted “pieces” of f using Lemma 2.14. Let N ⊂ Hn be the preimage of the convex core of j(Γ0 )\Hn . By Proposition B.1, there exists a compact set C ⊂ Hn such that N ⊂ j(Γ0 ) · C and Nk ⊂ jk (Γ0 ) · C for all large enough k ∈ N∗ , and the injectivity radius of j(Γ0 )\Hn and jk (Γ0 )\Hn is bounded from below by some constant δ > 0 independent of k. Let B1 , . . . , Br be open balls of Hn covering C, of radius < δ. For any 1 ≤ i ≤ r, let ψi : Hn → [0, 1] be a Lipschitz, j(Γ0 )-equivariant function supported on j(Γ0 ) · Bi , such that (ψi )1≤i≤r restricts to a partition of unity on j(Γ0 )·C, subordinated to the covering (j(Γ0 )·(Bi ∩C))1≤i≤r . For 1 ≤ i ≤ r and k ∈ N∗ , let Ψi,k ψi,k := Pr , i′ =1 Ψi′ ,k where Ψi,k : Hn → [0, 1] is the jk (Γ0 )-invariant function supported on jk (Γ0 ) · Bi that coincides with ψi on Bi . Then, for k ∈ N∗ large enough, (ψi,k )1≤i≤r induces a jk (Γ0 )-equivariant partition of unity on jk (Γ0 ) · C, subordinated to the covering (jk (Γ0 ) · (Bi ∩ C))1≤i≤r . Note that there is a constant L > 0 such that ψi,k is L-Lipschitz on jk (Γ0 )P · C for all 1 ≤ i ≤ r and large k ∈ N∗ ;P indeed, the jk (Γ0 )-invariant function i′ Ψi′ ,k is Lipschitz with constant ≤ i′ Lip(ψi′ ) and it converges uniformly to 1 on each Bi ∩ C as k → +∞, by compactness. For 1 ≤ i ≤ r and k ∈ N∗ , let fi,k : jk (Γ0 ) · Bi −→ Hn be the (jk , ρk )-equivariant map that coincides with f on Bi . For k ∈ N∗ and p ∈ jk (Γ0 )·C , let Ip,k be the set of indices 1 ≤ i ≤ r such that p ∈ jk (Γ0 )·Bi . The function p 7−→ Rp,k := diam{fi,k (p) | i ∈ Ip,k }, defined on jk (Γ0 ) · C, is jk (Γ0 )-invariant and converges uniformly to 0 on C as k → +∞. By Lemma 2.14, the (jk , ρk )-equivariant map r X ψi,k fi,k : jk (Γ0 ) · C −→ Hn fk := i=1 satisfies Lipp (fk ) ≤ r X i=1  LRp,k + ψi,k (p) Lipp (fi,k )   ≤ rL sup Rp′ ,k + Lip(f ) p′ ∈C for all p ∈ C, hence for all p ∈ Nk ⊂ jk (Γ0 ) · C by equivariance. We have seen that supp′ ∈C Rp′ ,k → 0 as k → +∞. Therefore, for large enough k, the (jk , ρk )-equivariant map Hn → Hn obtained by precomposing fk with the closest-point projection onto Nk has Lipschitz constant ≤ supp∈Nk Lipp (fk ) ≤ Lip(f ) + ε by Lemma 2.9. This shows that C(jk , ρk ) ≤ C(j, ρ) + 2ε, and we conclude by taking the lim sup over k and letting ε tend to 0. MAXIMALLY STRETCHED LAMINATIONS 55 6.1.2. Lower semicontinuity. Let us now prove that C(j, ρ) ≤ lim inf C(jk , ρk ). k→+∞ If ρ(Γ0 ) has a fixed point p in Hn , then C(j, ρ) = 0 (Remark 4.2) and there is nothing to prove. We thus assume that ρ(Γ0 ) has no fixed point in Hn . • Generic case. Consider the case where ρ(Γ0 ) has no fixed point in ∂∞ Hn and does not preserve any geodesic line of Hn . Then ρ(Γ0 ) contains two hyperbolic elements ρ(γ1 ), ρ(γ2 ) whose translation axes have no common endpoint in ∂∞ Hn . For large enough k, the elements ρk (γ1 ), ρk (γ2 ) ∈ ρk (Γ0 ) are hyperbolic too and their translation axes converge to the respective axes of ρ(γ1 ), ρ(γ2 ). For any k ∈ N∗ , let fk : Hn → Hn be a (jk , ρk )-equivariant, (C(jk , ρk ) + 2−k )-Lipschitz map. The same argument as in the proof of Lemma 4.10 shows that for any compact subset C of Hn , the sets fk (C ) all lie inside some common compact subset of Hn . By the Arzelà–Ascoli theorem, some subsequence of (fk )k∈N∗ converges to a (j, ρ)-equivariant map f : Hn → Hn . (Here we use that (C(jk , ρk ))k∈N∗ is bounded, a consequence of the upper semicontinuity proved in Section 6.1.1.) This implies C(j, ρ) ≤ lim inf k C(jk , ρk ). • Degenerate reductive case. Consider the case where ρ(Γ0 ) preserves a geodesic line A of Hn . The following observation is interesting in its own right. Lemma 6.2. If the group ρ(Γ0 ) preserves a geodesic line A ⊂ Hn without fixing any point in Hn , then the stretch locus E(j, ρ) is a geodesic lamination whose projection to j(Γ0 )\Hn is compact, contained in the convex core, and whose leaves are maximally stretched. Proof. After passing to a subgroup of index two (which does not change the stretch locus by Lemma 4.4), we may assume that ρ(Γ0 ) fixes both endpoints of A in ∂∞ Hn : in other words, ρ(Γ0 ) is contained in M A, where M is the subgroup of G that (pointwise) fixes A and A is the group of pure translations along A. The groups M and A commute and have a trivial intersection; let π A : M A → A be the natural projection. We claim that ρA := π A ◦ρ satisfies C(j, ρA ) = C(j, ρ) and E(j, ρA ) = E(j, ρ). remains in F j,ρ (resp. in F j,ρA ) Indeed, any element of (resp. of after postcomposing with the closest-point projection onto A, and for a map Hn → A it is equivalent to be (j, ρ)-equivariant or (j, ρA )-equivariant. Since ρA (Γ0 ) ⊂ A is commutative, for any m ∈ Z we can consider the representam tion ρm A : γ 7→ ρA (γ) . We claim that for m ≥ 1, F j,ρ C(j, ρm A ) = m C(j, ρA ) F j,ρA ) and E(j, ρm A ) = E(j, ρA ). Indeed, let hm be an orientation-preserving homeomorphism of A ≃ R such that d(hm (p), hm (q)) = m d(p, q) for all p, q ∈ A; for any C > 0, the postcomposition with hm realizes a bijection between the (j, ρA )-equivariant, C-Lipschitz maps and the (j, ρm A )equivariant, mC-Lipschitz maps from Hn to A, which preserves the stretch locus. Since C(j, ρA ) > 0 (because ρ(Γ0 ) has no fixed point in Hn ), we have 56 FRANÇOIS GUÉRITAUD AND FANNY KASSEL C(j, ρm A ) > 1 for large enough m, hence we can apply Theorem 1.3 to the stretch locus E(j, ρm  A ) = E(j, ρA ). Suppose the group ρ(Γ0 ) preserves a geodesic line of Hn without fixing any point in Hn . By Lemmas 4.6 and 6.2, for any ε > 0 there exists γ ∈ Γ0 with j(γ) hyperbolic such that (6.1) λ(ρ(γ)) ≥ C(j, ρ) − ε. λ(j(γ)) It follows that jk (γ) is hyperbolic and λ(ρk (γ)) ≥ C(j, ρ) − 2ε λ(jk (γ)) for all large enough k. We conclude by taking the lim inf over k and letting ε tend to 0. • Nonreductive case. Finally, we consider the case where the group ρ(Γ0 ) has a unique fixed point ξ in ∂∞ Hn , i.e. ρ is nonreductive (Definition 4.9). Choose an oriented geodesic line A of Hn with endpoint ξ. For any γ ∈ Γ0 we can write in a unique way ρ(γ) = gu where g ∈ G preserves A (i.e. belongs to M A with the notation above) and u is unipotent; setting ρred (γ) := g defines a representation ρred ∈ Hom(Γ0 , G) which is reductive (with image in M A). Note that changing the line A only modifies ρred by conjugating it; this does not change the constant C(j, ρred ) by Remark 4.3. When ρ is reductive, we set ρred := ρ. Then the following holds. (6.2) C(jk , ρk ) ≥ Lemma 6.3. For any pair of representations (j, ρ) ∈ Hom(Γ0 , G)2 with j convex cocompact, C(j, ρred ) = C(j, ρ). Proof. We may assume that ρ is nonreductive. Let ξ and A be as above and let pr : Hn → A be the “horocyclic projection” collapsing each horosphere centered at ξ to its intersection point with A; it is 1-Lipschitz. For any (j, ρ)equivariant Lipschitz map f : Hn → Hn , the map pr◦f is (j, ρred )-equivariant with Lip(pr ◦ f ) ≤ Lip(f ), hence C(j, ρred ) ≤ C(j, ρ). Let a ∈ G be a hyperbolic element acting as a pure translation along A, with repelling fixed point ξ at infinity. Then ρ(i) := ai ρ(·)a−i → ρred as i → +∞. By Remark 4.3, we have C(j, ρ(i) ) = C(j, ρ) for all i ∈ N. By upper semicontinuity (proved in Section 6.1.1), C(j, ρred ) ≥ lim sup C(j, ρ(i) ) = C(j, ρ). i→+∞  We now go back to our sequence (jk , ρk )k∈N∗ converging to (j, ρ). Since ρk → ρ and ρ has conjugates converging to ρred (see above), a diagonal argument shows that there are conjugates ρ′k of ρk such that ρ′k → ρred . By the reductive case above, lim inf k C(jk , ρ′k ) ≥ C(j, ρred ), and we conclude using Remark 4.3 and Lemma 6.3. This completes the proof of Proposition 1.5. 6.2. Openness of the condition C < 1 on cusp-deteriorating pairs. In this section we prove Proposition 6.1.(1)–(1’). The strategy is analogous MAXIMALLY STRETCHED LAMINATIONS 57 to the proof of upper semicontinuity in Section 6.1.1. The partition-of-unity argument in that proof fails in the presence of cusps, since the convex core (when nonempty) is not compact anymore. However, we shall see that the argument can be adapted as long as convex cores deform continuously. Such continuous behavior is ensured under the assumptions of Prop. 6.1.(1’) (constant convex core) or Prop. 6.1.(1) (all cusps of rank ≥ n − 2, see Proposition B.3). Consider a pair (j, ρ) ∈ Homj0 (Γ0 , G)×Homj0 -det (Γ0 , G) with C(j, ρ) < 1, and a sequence (jk , ρk )k∈N∗ of elements of Homj0 (Γ0 , G) × Homj0 -det (Γ0 , G) converging to (j, ρ). If j0 has a cusp of rank < n − 2, we assume that jk = j for all k ∈ N∗ . We shall prove that C(jk , ρk ) < 1 for all large enough k. We can and shall assume that Γ0 is torsion-free (using Lemma 4.4 and the Selberg lemma [Se, Lem. 8]). We can also always assume that the convex core of jk (Γ0 )\Hn is nonempty: otherwise the group jk (Γ0 ) is elementary with a fixed point in Hn or a unique fixed point in ∂∞ Hn , and C(jk , ρk ) = 0 by Remark 4.2. Therefore the convex core of M := j(Γ0 )\Hn is nonempty too (because j and the jk have the same cusp type). Let f : Hn → Hn be a (j, ρ)-equivariant map with 0 < Lip(f ) < 1. We shall modify f into a (jk , ρk )-equivariant map fk with Lip(fk ) < 1 for all large enough k. As usual, by Lemma 4.8 we only need to define fk on the preimage Nk ⊂ Hn of the convex core of jk (Γ0 )\Hn . In order to build fk , we shall proceed as in Section 6.1.1 and paste together shifted “pieces” of f using Lemma 2.14. By Proposition 4.16.(3) we may assume that f is constant on neighborhoods of some horoballs H1 , . . . , Hc of Hn whose images in M = j(Γ0 )\Hn are disjoint and intersect the convex core of M in standard cusp regions (Definition 2.2), representing all the cusps. For 1 ≤ ℓ ≤ c, let Sℓ ⊂ Γ0 be the stabilizer of Hℓ under the j-action. Let N ⊂ Hn be the preimage of the convex core of j(Γ0 )\Hn . By Proposition B.3, if the horoballs H1 , . . . , Hc are small enough, then there exist a compact set C ⊂ Hn and, for any k ∈ N∗ , horoballs H1k , . . . , Hck of Hn , such that • the images of H1k , . . . , Hck in jk (Γ0 )\Hn are disjoint and intersect the convex core in standard cusp regions, for all large enough k ∈ N∗ ; • the stabilizer in Γ0 of Hℓk under jk is Sℓ ; • the horoballs Hℓk S converge to Hℓ for all 1 ≤ ℓ ≤ c; • N ⊂ j(Γ0 ) · (C ∪ 1≤ℓ≤c Hℓ ) and, for all large enough k ∈ N∗ ,   [ k (6.3) Nk ⊂ jk (Γ0 ) · C ∪ Hℓ ; 1≤ℓ≤c • the cusp thickness (Definition 5.12) of jk (Γ0 )\Hn at any point of ∂Hℓk is uniformly bounded by some constant Θ > 0 independent of k;  S • the injectivity radius of jk (Γ0 )\ Hn r kℓ=1 jk (Γ0 ) · Hℓk is bounded from below by some constant δ > 0 independent of k. (If j0 has a cusp of rank < n − 2, then jk = j and we take Hℓk = Hℓ for all k ∈ N∗ .) For any 1 ≤ ℓ ≤ c, by convergence of the horoballs Hℓk , the map f is constant on some neighborhood of ∂Hℓk ∩ C for large enough k, which 58 FRANÇOIS GUÉRITAUD AND FANNY KASSEL implies (6.4) sup p∈∂Hℓk ∩C Lipp (f ) = 0. Let B1 , . . . , Br be open balls of Hn covering C, of radius < δ. For any 1 ≤ i ≤ r, let ψi : Hn → [0, 1] be a Lipschitz, j(Γ0 )-equivariant function supported on j(Γ0 ) · Bi , such that (ψi )1≤i≤r restricts to a partition of unity on j(Γ0 ) · C, subordinated to the covering (j(Γ0 ) · (Bi ∩ C))1≤i≤r . As in Section 6.1.1, for large enough k we can perturb the ψi to a jk (Γ0 )equivariant partition of unity (ψi,k )1≤i≤r of jk (Γ0 ) · C, subordinated to the covering (jk (Γ0 ) · Bi )1≤i≤r , such that all the functions ψi,k are L-Lipschitz for some constant L > 0 independent of i and k. For 1 ≤ i ≤ r and k ∈ N∗ , let fi,k : jk (Γ0 ) · Bi −→ Hn be the (jk , ρk )-equivariant map that coincides with f on Bi . As in Section 6.1.1, it follows from Lemma 2.14 that the (jk , ρk )-equivariant map r X ψi,k fi,k : jk (Γ0 ) · C −→ Hn fk′ := i=1 satisfies (6.5) Lipp (fk′ ) ≤ rLRp,k + Lipp (f ) for all p ∈ jk (Γ0 )·C, where p 7→ Rp,k is a jk (Γ0 )-invariant function converging uniformly to 0 on C as k → +∞. By equivariance, lim sup sup k→+∞ p∈jk (Γ0 )·C Lipp (fk′ ) ≤ Lip(f ) < 1. It only remains to prove that for any 1 ≤ ℓ ≤ c we can extend fk′ |∂H k ∩Nk ℓ to Hℓk ∩ Nk in a (jk |Sℓ , ρk |Sℓ )-equivariant way with Lipschitz constant < 1. Indeed, then we can extend fk′ to the orbit jk (Γ0 ) · (Hℓk ∩ Nk ) in a (jk , ρk )equivariant way; piecing together these maps for varying ℓ, and taking fk′ on S the complement of cℓ=1 jk (Γ0 ) · Hℓk in Nk (which is contained in jk (Γ0 ) · C), we will obtain a (jk , ρk )-equivariant map fk : Nk → Hn with Lip(fk ) < 1 for all large enough k, which will complete the proof. Fix 1 ≤ ℓ ≤ c. By Theorem 1.6, in order to prove that fk′ |∂H k ∩Nk extends ℓ to Hℓk ∩ Nk in a (jk |Sℓ , ρk |Sℓ )-equivariant way with Lipschitz constant < 1, it is sufficient to prove that Lip∂H k ∩Nk (fk′ ) < 1. By (6.4) and (6.5), for any ℓ ε > 0 we have (6.6) sup p∈∂Hℓk ∩Nk Lipp (fk′ ) ≤ ε for all large enough k, since ∂Hℓk ∩ Nk ⊂ jk (Γ0 ) · C and the jk (Γ0 )-invariant functions p 7→ Rp,k converge uniformly to 0 on C as k → +∞. Note that (6.6) does not immediately give a bound on the global constant Lip∂H k ∩Nk (fk′ ), ℓ since the subset of horosphere ∂Hℓk ∩ Nk is not convex for the hyperbolic metric. However, such a bound follows from Lemmas 6.4 and 6.5 below, which are based on a comparison between the intrinsic metrics of horospheres and spheres in Hn . MAXIMALLY STRETCHED LAMINATIONS 59 The idea of Lemma 6.4 is that (6.6) controls the Lipschitz constant at short range, while the fixed point of ρk (Sℓ ) implies control at long range. The difficulty is that there can be an arbitrarily large “medium range” to handle inbetween, since the fixed point of ρk (Sℓ ) can lie arbitrarily far out. In dimension n ≥ 4 this is compounded by the fact that generally ∂Hℓk ∩Nk is not even convex for the Euclidean metric of the horosphere ∂Hℓk ; Lemma 6.5 deals with that issue. For t ≥ 1, we say that a subset X of a Euclidean space is t-subconvex if for any x, y ∈ X there exists a path from x to y in X whose length is at most t times the Euclidean distance from x to y. Lemma 6.4. Let S be a discrete group. For any R > 0, there exists ε > 0 with the following property: if (j, ρ) ∈ Hom(S, G)2 is a pair of representations with j injective such that • the group j(S) is discrete and preserves a horoball H of Hn , • the group ρ(S) has a fixed point in Hn , • there exists a closed, j(S)-invariant, 2-subconvex set N ⊂ ∂H such that the quotient j(S)\N has Euclidean diameter ≤ R, then any (j, ρ)-equivariant map f ′ : N → Hn satisfying Lipp (f ′ ) ≤ ε for all p ∈ N satisfies Lip(f ′ ) < 1. Lemma 6.5. In our setting, up to enlarging the compact set C ⊂ Hn and replacing the horoballs H1 , . . . , Hc and H1k , . . . , Hck with smaller horoballs, still satisfying the list of six properties (6.3), we may assume that ∂Hℓk ∩ Nk is 2-subconvex in ∂Hℓk ≃ Rn−1 for all 1 ≤ ℓ ≤ c and large enough k ∈ N∗ . Here Lemma 6.4 applies to N := Nk ∩ ∂Hℓk (which is 2-subconvex by Lemma 6.5) and to f ′ := fk′ |N (which satisfies (6.6)). Note that ρk (Sℓ ) has a fixed point in Hn by Fact 2.5, since ρk is cusp-deteriorating with respect to jk , and that the Euclidean diameter of jk (Sℓ )\(Nk ∩ ∂Hℓk ) is uniformly bounded for k ∈ N∗ , by the uniform bound Θ on cusp thickness. Therefore it is sufficient to prove Lemmas 6.4 and 6.5 to complete the proof of Proposition 6.1.(1)–(1’). Proof of Lemma 6.4. Fix R > 0 and let j, ρ, H, N be as in the statement. Consider a (j, ρ)-equivariant map f ′ : N → Hn such that Lipp (f ′ ) ≤ ε for all p ∈ N , for some ε > 0. Let us show that if ε is smaller than some constant independent of f ′ , then Lip(f ′ ) < 1. Let d∂H be the natural Euclidean metric on ∂H. By (A.3), for any p, q ∈ N ,   d∂H (p, q) . (6.7) d(p, q) = 2 arcsinh 2 If d(p, q) ≤ 1, then d(p, q) ≥ κ d∂H (p, q) for some universal κ > 0 (specifically, κ = (2 sinh(1/2))−1 by concavity of arcsinh). On the other hand, d(f ′ (p), f ′ (q)) ≤ 2ε d∂H (p, q) by Remark 2.8.(3) and 2-subconvexity, hence d(f ′ (p), f ′ (q)) 2ε 2 ≤ < d(p, q) κ 3 for all p, q ∈ N with 0 < d(p, q) ≤ 1 as soon as ε < κ/3. We now assume that this is satisfied and consider pairs of points p, q ∈ N with d(p, q) ≥ 1. 60 FRANÇOIS GUÉRITAUD AND FANNY KASSEL Let ∂B be a sphere of Hn centered at a fixed point of ρ(S), and containing f ′ (p0 ) for some p0 ∈ N (see Figure 8). Then f ′ (j(γ) · p0 ) ∈ ∂B for all γ ∈ S, since f ′ is (j, ρ)-equivariant and ρ(S) preserves ∂B. By Remark 2.8.(3) and 2-subconvexity, the set f ′ (N ) is contained in the 2εR-neighborhood of ∂B. 1 we have If the radius of ∂B is ≤ 1/3, then as soon as ε < 24R 2 5 d(f ′ (p), f ′ (q)) ≤ + 4εR ≤ d(p, q) 3 6 ′ for all p, q ∈ N with d(p, q) ≥ 1, hence Lip(f ) ≤ 5/6 < 1. We now assume that the radius r of ∂B is > 1/3 (possibly very large!). There exists a universal constant η > 0 such that the closest-point projection onto any sphere of Hn of radius > 1/3 is 2-Lipschitz on the η-neighborhood (inner η , which we shall assume and outer) of this sphere. In particular, if ε ≤ 2R from now on, then the projection onto ∂B is 2-Lipschitz on the set f ′ (N ). p0 Hn Hn H j(S) f′ f ′ (p0 ) ρ(S) B ′ f (∂H) Figure 8. An equivariant map f ′ , contracting at small scale, taking a horosphere to (or near) a sphere, is contracting at all scales. Let x, y ∈ ∂B be the respective projections of f ′ (p), f ′ (q); the distances d(x, f ′ (p)) and d(y, f ′ (q)) are bounded from above by 2εR. Let d∂B (x, y) be the length of the shortest path from x to y that is contained in the sphere ∂B. The formulas (A.8) and (A.14) yield    d∂B (x, y) (6.8) d(x, y) = 2 arcsinh sinh(r) · sin . 2 sinh(r) On the other hand, by 2-subconvexity, we can find a path ω from p to q in N of length at most 2 d∂H (p, q). Then d∂B (x, y) is bounded from above by the length of the projection of the path f ′ (ω) to ∂B, hence, by Remark 2.8.(3), (6.9) d∂B (x, y) ≤ 4ε d∂H (p, q). Using sin(t) ≤ min{1, t} for t ≥ 0, it follows from (6.8) and (6.9) that d(f ′ (p), f ′ (q)) ≤ d(f ′ (p), x) + d(x, y) + d(y, f ′ (q))  ≤ min 2r, 2 arcsinh(d∂B (x, y)/2) + 4εR   ≤ min 2r, 2 arcsinh 2ε d∂H (p, q) + 4εR. MAXIMALLY STRETCHED LAMINATIONS 61 Comparing with (6.7), we see that if ε is smaller than some constant depending only on R, then d(f ′ (p), f ′ (q)) < d(p, q) for all p, q ∈ N with d(p, q) ≥ 1. Since d(f ′ (p), f ′ (q)) is bounded independently of p and q, the ratio d(f ′ (p), f ′ (q))/d(p, q) is uniformly bounded away from 1 by compactness of N modulo j(S). This proves that LipN (f ′ ) < 1.  Proof of Lemma 6.5. Fix 1 ≤ ℓ ≤ c, where c is still the number of cusps. • Subconvexity for ∂Hℓ ∩ N . We first prove that, up to replacing Hℓ with some smaller, concentric horoball, the set ∂Hℓ ∩ N is 2-subconvex in ∂Hℓ . This will prove Lemma 6.5 when j0 has a cusp of rank < n − 2, since in that case jk = j and Hℓk = Hℓ for all k by assumption. The stabilizer Sℓ ⊂ Γ0 of Hℓ under j has a finite-index normal subgroup S ′ isomorphic to Zm for some 0 < m < n (see Section 2.1). Consider the upper half-space model Rn−1 ×R∗+ of Hn , so that ∂∞ Hn identifies with Rn−1 ∪{∞}. We may assume that Hℓ is centered at infinity, so that ∂Hℓ = Rn−1 × {b} for some b > 0. Let Ω be the convex hull of Λj(Γ0 ) r{∞} in Rn−1 , where Λj(Γ0 ) is the limit set of j(Γ0 ). The group j(S ′ ) acts on Rn−1 by Euclidean isometries and there exists an m-dimensional affine subspace V ⊂ Ω, preserved by j(S ′ ), on which j(S ′ ) acts as a lattice of translations (see Section 2.1). We claim that N contains V × [b0 , +∞) for some b0 > 0. Indeed, since V ⊂ Ω, some point p0 ∈ V × R∗+ ⊂ Hn belongs to N . The convex hull in Hn of the orbit j(S ′ ) · p0 is also contained in N . This convex hull contains all the j(S ′ )-translates of the (compact) convex hull of  ε1 εm ) · p0 | (ε1 , . . . , εm ) ∈ {0, 1}m , j(γ1 . . . γm where (γ1 , . . . , γm ) is a generating subset of S ′ ; the union X of these j(S ′ )translates projects vertically onto the whole of V and has bounded height since j(S ′ ) preserves the horospheres centered at ∞. Then N contains V × [b0 , +∞) where b0 > 0 is the maximal height of X. Up to replacing Hℓ with some smaller, concentric horoball, we may assume that b ≥ max{b0 , 7δ}, where δ > 0 is the Euclidean diameter of j(S ′ )\Ω. Let us show that ∂Hℓ ∩ N is then 2-subconvex. Consider p, q ∈ ∂Hℓ ∩ N , with respective orthogonal projections ζp , ζq to Rn−1 . We have d∂Hℓ (p, q) = dRn−1 (ζp , ζq )/b. Suppose dRn−1 (ζp , ζq ) ≤ 6δ. By definition of δ, we can find a point ζ ∈ Λj(Γ0 ) r {∞} ⊂ Rn−1 with dRn−1 (ζ, ζp ) ≤ δ. The hyperbolic triangle (p, q, ζ) is contained in N . Since b ≥ 7δ, both edges (p, ζ] and (q, ζ] lie outside Hℓ = Rn−1 × [b, +∞). It follows that the intersection of this triangle (p, q, ζ) with ∂Hℓ is an arc of Euclidean circle from p to q, of angular measure ≤ π, and hence has Euclidean length at most π2 d∂Hℓ (p, q) ≤ 2 d∂Hℓ (p, q). Suppose dRn−1 (ζp , ζq ) ≥ 6δ. Since ζp , ζq ∈ Ω, by definition of δ we can find points p′ , q ′ in N ∩ (V × {b}) whose orthogonal projections ζp′ , ζq′ to Rn−1 satisfy and dRn−1 (ζq , ζq′ ) ≤ δ. dRn−1 (ζp , ζp′ ) ≤ δ ′ Then d∂Hℓ (p, p ) = dRn−1 (ζp , ζp′ )/b ≤ δ/b, and similarly d∂Hℓ (q, q ′ ) ≤ δ/b. As above, there is an arc of Euclidean circle from p to p′ in ∂Hℓ ∩ N , of 62 FRANÇOIS GUÉRITAUD AND FANNY KASSEL length at most 2 d∂H (p, p′ ) ≤ 2δ/b. Similarly, there is an arc of Euclidean circle from q ′ to q in ∂Hℓ ∩ N , of Euclidean length ≤ 2δ/b. Concatenating these arcs with the Euclidean segment [p′ , q ′ ] ⊂ V × {b}, which is contained in ∂Hℓ ∩ N and has Euclidean length b−1 dRn−1 (ζp′ , ζq′ ), we find a path from p to q in ∂Hℓ ∩ N of Euclidean length at most dRn−1 (ζp′ , ζq′ ) + 4δ d n−1 (ζp , ζq ) + 6δ ≤ R ≤ 2 d∂Hℓ (p, q). b b This proves that ∂Hℓ ∩ N is 2-subconvex in ∂Hℓ . • Convexity for ∂Hℓk ∩ Nk in the case of cusps of rank ≥ n − 2. Finally, we suppose that all cusps of j0 have rank ≥ n − 2, in which case the representation jk is allowed to vary with k. Recall that the cusp thickness of jk (Γ0 )\Hn at ∂Hkℓ is bounded by some constant Θ > 0 independent of ℓ and k. If we replace every horoball Hkℓ with the smaller, concentric horoball at distance log(3Θ) from ∂Hkℓ (and correspondingly enlarge the compact set C ⊂ Hn ), we obtain new horoballs Hkℓ still satisfying the list of six properties (6.3), such that the cusp thickness of jk (Γ0 )\Hn at ∂Hkℓ is ≤ 1/3 for all ℓ and k. Then ∂Hℓk ∩ Nk is convex in ∂Hℓk by Lemma B.4, hence in particular 2-subconvex.  6.3. The constant C(j, ρ) for nonreductive ρ. In order to prove conditions (2), (3), (3)’, (3)” of Proposition 6.1 (in Section 6.4), we shall rely on the existence of a maximally stretched lamination for C(j, ρ) ≥ 1, given by Theorem 1.3. However, Theorem 1.3 assumes that the space F j,ρ of equivariant maps realizing the best Lipschitz constant C(j, ρ) is nonempty: this holds for reductive ρ (Lemma 4.10), but may fail otherwise (see Section 10.3). In order to deal with nonreductive ρ, we first establish the following lemma, which extends Lemma 6.3. Lemma 6.6. For any pair of representations (j, ρ) ∈ Hom(Γ0 , G)2 with j geometrically finite, C(j, ρ) = C(j, ρred ), unless the representation ρ is not cusp-deteriorating and C(j, ρred ) < 1, in which case C(j, ρ) = 1. Here ρred ∈ Hom(Γ0 , G) is the “reductive part” of ρ, defined in Section 6.1.2: if ρ is nonreductive, then the group ρred (Γ0 ) preserves some geodesic line of Hn with an endpoint in ∂∞ Hn equal to the fixed point of ρ(Γ0 ). Since ρred is well defined up to conjugation, the constant C(j, ρred ) is well defined by Remark 4.3. If ρ is reductive, then ρred := ρ. Proof of Lemma 6.6. We may assume that ρ is nonreductive, with fixed point ξ ∈ ∂∞ Hn . Then ρred is cusp-deteriorating and preserves an oriented geodesic line A of Hn with endpoint ξ. If the group j(Γ0 ) is elementary and fixes a unique point in ∂∞ Hn , then C(j, ρ) = 1 by Corollary 4.18 and C(j, ρred ) = 0 by Remark 4.2. We now assume that we are not in this case, which means that the convex core of M := j(Γ0 )\Hn is nonempty. As in the proof of Lemma 6.3, by using the projection onto A along concentric horocycles we see that C(j, ρred ) ≤ C(j, ρ), MAXIMALLY STRETCHED LAMINATIONS 63 and there is a sequence (ak )k∈N∗ of pure translations along A, with repelling fixed point ξ, such that the conjugates ρk := ak ρ(·)a−1 k (which still fix ξ) converge to ρred as k → +∞. By invariance of C(j, ρ) under conjugation (Remark 4.3), it is sufficient to prove that  C(j, ρred )  if ρ is cusp-deteriorating, lim sup C(j, ρk ) ≤ red ) max 1, C(j, ρ otherwise. k→+∞ To prove this, we use a partition-of-unity argument as in Sections 6.1.1 and 6.2. Fix ε > 0. By using Proposition 4.16 and postcomposing with the closest-point projection onto A, we can find a (j, ρred )-equivariant map f : Hn → A with Lip(f ) ≤ C(j, ρred )+ε/2 that is constant on neighborhoods of some horoballs B1 , . . . , Bc of Hn whose images in M = j(Γ0 )\Hn are disjoint and intersect the convex core in standard cusp regions (Definition 2.2), representing all the cusps. We shall use f to build (j, ρk )-equivariant maps fk with Lip(fk ) bounded from above by Lip(f ) + ε or 1 + ε, as the case may be, for all large enough k. Let S1 , . . . , Sc ⊂ Γ0 be the respective stabilizers of B1 , . . . , Bc under j; the singleton f (Bi ) is fixed by ρ(Si ) for all 1 ≤ i ≤ c. Let also Bc+1 , . . . , BrSbe open balls of Hn , each projecting injectively to j(Γ0 )\Hn , such that ri=1 j(Γ0 ) · Bi contains the preimage N ⊂ Hn of the convex core of M . For c < i ≤ r, let fi,k : j(Γ0 ) · Bi → Hn be the (j, ρk )equivariant map that coincides with f on Bi . We first assume that ρ is cusp-deteriorating. For 1 ≤ i ≤ c, all the elements of ρ(Si ) are elliptic, hence ρ(Si ) fixes a point in Hn (Fact 2.5). Since it also fixes ξ ∈ ∂∞ Hn , it fixes pointwise a full line A′ with endpoint ξ. ′ Then ρk (Si ) = ak ρ(Si )a−1 k fixes pointwise the line ak · A , which converges to A as k → +∞. In particular, we can find a sequence (pi,k )k∈N∗ that converges to the singleton f (Bi ) ∈ A as k → +∞, with pi,k fixed by ρk (Si ) for all k. For 1 ≤ i ≤ c and k ∈ N∗ , let fi,k : j(Γ0 ) · Bi −→ Hn be the (j, ρk )-equivariant map that is constant equal to pi,k on the horoball Bi . Let (ψi )1≤i≤r be a Lipschitz partition of unity subordinated to the covering (j(Γ0 ) · Bi )1≤i≤r of N , and let L := max1≤i≤r Lip(ψi ). By Lemma 2.14, the (j, ρk )-equivariant map r X ψi fi,k : N −→ Hn fk := i=1 satisfies Lipp (fk ) ≤ rLRp,k + Lipp (f ) for all p ∈ N , where the j(Γ0 )-invariant function  d fi,k (p), fi′ ,k (p) p 7−→ Rp,k := max ′ i,i Sr converges uniformly to 0 for p ∈ i=1 Bi , as k → +∞. For large enough k this yields LipN (fk ) ≤ Lip(f ) + ε/2 by (2.2), hence C(j, ρk ) ≤ C(j, ρred ) + ε by Lemma 4.8. Letting ε go to 0, we obtain lim supk C(j, ρk ) ≤ C(j, ρred ) as desired. 64 FRANÇOIS GUÉRITAUD AND FANNY KASSEL Suppose now that ρ is not cusp-deteriorating. WeSproceed as in the cuspdeteriorating case, but work with the union of balls c<i≤r j(Γ0 ) · Bi instead S of the union of balls and horoballs 1≤i≤r j(Γ0 ) · Bi . Let (ψi )c<i≤r be a S Lipschitz partition of unity of N ′ := N r 1≤ℓ≤c j(Γ0 ) · Bℓ subordinated to the covering (j(Γ0 ) · Bi )c<i≤r , and let L := maxc<i≤r Lip(ψi ). As in the cusp-deteriorating case, by Lemma 2.14, the (j, ρk )-equivariant map X fk′ := ψi fi,k : N ′ −→ Hn c<i≤r satisfies Lipp (fk′ ) ≤ Lipp (f ) + ε/2 for all p ∈ N ′ when k is large enough. In particular, for 1 ≤ ℓ ≤ c, since f is constant on a neighborhood of the horoball Bℓ , we obtain Lipp (fk′ ) ≤ ε/2 for all p ∈ N ∩ ∂Bℓ . It is sufficient to prove that (6.10) LipN ∩∂Bℓ (fk′ ) ≤ 1 for all 1 ≤ ℓ ≤ c, since Theorem 1.6 (or Proposition 3.9) then lets us extend fk′ |N ∩∂Bℓ to a 1-Lipschitz, (j|Sℓ , ρk |Sℓ )-equivariant map (Bℓ ∪∂Bℓ )∩N → Hn . We can then extend fk′ to the orbit j(Γ0 ) · (Bℓ ∪ ∂Bℓ ) ∩ N in a (j, ρk )equivariant way. Piecing together these maps for varying 1 ≤ ℓ ≤ c, and taking fk′ on N ′ , we then obtain a (j, ρk )-equivariant map fk : Hn → Hn with Lip(fk ) ≤ max(1, Lip(f ) + ε/2) for all large enough k (using (2.2)). Letting ε go to 0, we obtain lim supk C(j, ρk ) ≤ max(1, C(j, ρred )), as desired. To prove (6.10), it is sufficient to establish the following analogue of Lemma 6.4, which together with Lemma 6.5 completes the proof of Lemma 6.6.  Lemma 6.7. Let S be a discrete group. For any R > 0, there exists ε > 0 with the following property: if (j, ρ) ∈ Hom(S, G)2 is a pair of representations with j injective such that • the group j(S) is discrete and preserves a horoball H of Hn , • the group ρ(S) has a fixed point in ∂∞ Hn , • there exists a closed, j(S)-invariant, 2-subconvex set N ⊂ ∂H such that the quotient j(S)\N has (Euclidean) diameter ≤ R, then any (j, ρ)-equivariant map f ′ : N → Hn satisfying Lipp (f ′ ) ≤ ε for all p ∈ N satisfies Lip(f ′ ) ≤ 1. Proof. We proceed as in the proof of Lemma 6.4, but the sphere ∂B will now be a horosphere. Fix R > 0 and let j, ρ, H, N be as in the statement. Consider a (j, ρ)-equivariant map f ′ : N → Hn such that Lipp (f ′ ) ≤ ε for all p ∈ N , for some ε > 0. Let us show that if ε is smaller than some constant independent of f ′ , then Lip(f ′ ) ≤ 1. As in the proof of Lemma 6.4, if ε is smaller than some universal constant, then d(f ′ (p), f ′ (q)) ≤ d(p, q) for all p, q ∈ N with d(p, q) ≤ 1. We now consider p, q ∈ N with d(p, q) ≥ 1. Let ∂B be a horosphere centered at the fixed point of ρ(S) in ∂∞ Hn , containing f ′ (p0 ) for some p0 ∈ N . As in the proof of Lemma 6.4, the set f ′ (N ) is contained in the 2εR-neighborhood of ∂B. We now use the existence of a universal constant η > 0 such that the closest-point projection onto any horosphere of Hn is 2-Lipschitz on the η-neighborhood (inner and outer) of MAXIMALLY STRETCHED LAMINATIONS 65 η this horosphere. In particular, if ε ≤ 2R , which we shall assume from now on, then the projection onto ∂B is 2-Lipschitz on the set f ′ (N ). Denoting by x, y ∈ ∂B the projections of f ′ (p), f ′ (q), the (in)equalities (6.7) and (6.9) still hold, but (6.8) becomes   d∂B (x, y) d(x, y) = 2 arcsinh , 2 where d∂B is the natural Euclidean metric on ∂B. We obtain d(f ′ (p), f ′ (q)) ≤ d(f ′ (p), x) + d(x, y) + d(y, f ′ (q))  ≤ 2 arcsinh 2ε d∂H (p, q) + 2εR. Comparing with (6.7) we see that if ε is small enough then d(f ′ (p), f ′ (q)) ≤ d(p, q) for all p, q ∈ ∂H ∩ N with d(p, q) ≥ 1. Hence, Lip(f ′ ) ≤ 1.  6.4. Semicontinuity for C(j, ρ) ≥ 1 in the general geometrically finite case. We now complete the proof of Proposition 6.1. Condition (1) when all the cusps of j0 have rank ≥ n − 2 and condition (1’) in general have already been proved in Section 6.2. We now show that for pairs (j, ρ) of representations with j geometrically finite representations of fixed cusp type, (2) the condition 1 < C is open, (3’) the function (j, ρ) 7→ C(j, ρ) is lower semicontinuous on the set of pairs where 1 ≤ C, (3”) it is upper semicontinuous on the set of pairs where 1 ≤ C < +∞ when either all the cusps of j0 have rank ≥ n − 2 (for instance n ≤ 3) or the representation j is constant. Upper semicontinuity on the set of pairs where 1 ≤ C < +∞ does not hold in general in dimension n ≥ 4: see Section 10.10. The condition C(j, ρ) = +∞ is open in Hom(Γ0 , G)j0 × Hom(Γ0 , G) by Lemma 4.7, hence in (2) and (3’) we may actually restrict to 1 ≤ C < +∞. Let (jk , ρk )k∈N∗ be a sequence of elements of Hom(Γ0 , G)j0 × Hom(Γ0 , G) converging to an element (j, ρ) ∈ Hom(Γ0 , G)j0 × Hom(Γ0 , G) with C(j, ρ) < +∞. It is sufficient to prove the following two statements: (A) if C(j, ρ) > 1, then C(j, ρ) ≤ lim inf k C(jk , ρk ), (B) if C ∗ := lim supk C(jk , ρk ) > 1 and if either all the cusps of j0 have rank ≥ n − 2 or jk = j for all k ∈ N∗ , then C(j, ρ) ≥ C ∗ . If ρ is reductive, then (A) is an easy consequence of Corollary 1.12 (here E(j, ρ) 6= ∅ by Corollary 4.14, in which case Corollary 1.12 has been proved in Section 5.4): namely, for any ε > 0 there is an element γ ∈ Γ0 with j(γ) hyperbolic such that λ(ρ(γ)) ≥ C(j, ρ) − ε. λ(j(γ)) If k is large enough, then λ(jk (γ)) is hyperbolic and λ(ρk (γ))/λ(jk (γ)) ≥ C(j, ρ) − 2ε by continuity of λ, hence C(jk , ρk ) ≥ C(j, ρ) − 2ε by (4.1). We conclude by letting ε tend to 0. If ρ is nonreductive, then C(j, ρ) > 1 entails C(j, ρred ) = C(j, ρ) by Lemma 6.6, and the ρk have conjugates converging to ρred (see the end of Section 6.1.2), so we just apply the reductive case to obtain (A). 66 FRANÇOIS GUÉRITAUD AND FANNY KASSEL To prove (B), suppose that C ∗ > 1 and that either all the cusps of j0 have rank ≥ n − 2 or jk = j for all k ∈ N∗ . Up to passing to a subsequence, we may assume C(jk , ρk ) > 1 for all k ∈ N∗ and C(jk , ρk ) → C ∗ . Then C(jk , ρred k ) = C(jk , ρk ) for all k ∈ N∗ by Lemma 6.6. We now use Theorem 1.3, and either Proposition 5.11 (if jk = j) or Corollary 5.14 (if all the cusps of j0 have rank ≥ n−2): in either case we obtain that the stretch locus E(jk , ρred k ) is a (nonempty) geodesic lamination admitting a fundamental domain that remains in some compact subset of Hn , independent of k. This implies, up to passing to a subsequence, that E(jk , ρred k ) converges to some (nonempty) j(Γ0 )-invariant geodesic lamination L , with a compact image in j(Γ0 )\Hn . For any ε > 0, a closed curve ε-nearly carried by L is also nearly carried by E(jk , ρred k ) and will give (as in the proof of Lemma 4.6) an element γ ∈ Γ0 such that jk (γ) is hyperbolic and λ(ρred λ(ρk (γ)) k (γ)) = ≥ C(jk , ρred k )−ε λ(jk (γ)) λ(jk (γ)) for all large enough k. The right hand side converges to C ∗ − ε, hence by continuity of λ, λ(ρ(γ)) ≥ C ∗ − ε, λ(j(γ)) therefore C(j, ρ) ≥ C ∗ − ε by (4.1) (in particular, C ∗ < +∞). We conclude by letting ε tend to 0. This completes the proof of Proposition 6.1. 7. Application to properly discontinuous actions on G = PO(n, 1) In this section we prove the results of Section 1.4 on the geometrically finite quotients of G := PO(n, 1), namely Theorem 1.8 (properness criterion) and Theorems 1.9 and 1.11 (deformation). We adopt the notation and terminology of Section 1.4, and proceed as in [Ka1]. Note that all the results remain true if G is replaced by O(n, 1), SO(n, 1), or SO(n, 1)0 . In Section 7.1 we start by introducing a constant Cµ (j, ρ), which we use in Section 7.2 to state a refinement of Theorem 1.8. This refinement is proved in Sections 7.5 and 7.6. Before that, in Section 7.3 we discuss the connection with the general theory of properly discontinuous actions on reductive homogeneous spaces, and in Section 7.4 we make two side remarks. Section 7.7 is devoted to the proof of Theorems 1.9 and 1.11, and Section 7.8 to their interpretation in terms of completeness of geometric structures. 7.1. The constant Cµ (j, ρ). We shall refine Theorem 1.8 by characterizing properness, not only in terms of the constants C(j, ρ) of (1.1) and C ′ (j, ρ) of (1.4), but also in terms of a third constant Cµ (j, ρ). We start by introducing this constant. Fix a basepoint p0 ∈ Hn and let µ : G → R+ be the displacement function relative to p0 : (7.1) µ(g) := d(p0 , g · p0 ) MAXIMALLY STRETCHED LAMINATIONS 67 for all g ∈ G. This function is continuous, proper, and surjective; we shall see in Section 7.3 that it corresponds to a Cartan projection of G. Note that µ(g−1 ) = µ(g) for all g ∈ G because G acts on Hn by isometries. By the triangle inequality, µ(gg ′ ) ≤ µ(g) + µ(g ′ ), (7.2) and (7.3) g, g ′ λ(g) ≤ µ(g) for all ∈ G, where λ : G → R+ is the translation length function of (1.2). For hyperbolic g, the function k 7→ µ(g k ) grows linearly because µ(gk ) − kλ(g) is bounded (for instance by twice the distance from g · p0 to the translation axis Ag of g). For parabolic g, the function k 7→ µ(gk ) grows logarithmically (Lemma 2.6), while for elliptic g it is bounded. Therefore, 1 (7.4) λ(g) = lim µ(gk ) k→+∞ k for all g ∈ G. For any discrete group Γ0 and any pair (j, ρ) ∈ Hom(Γ0 , G)2 of representations, we denote by Cµ (j, ρ) the infimum of constants t ≥ 0 for which the set {µ(ρ(γ)) − t µ(j(γ)) | γ ∈ Γ0 } is bounded from above, i.e. for which µ(ρ(γ)) ≤ t µ(j(γ)) + O(1) as γ ranges over Γ0 . Note that (7.5) C ′ (j, ρ) ≤ Cµ (j, ρ) ≤ C(j, ρ). Indeed, the left-hand inequality follows from (7.4). The right-hand inequality follows from the fact that for any (j, ρ)-equivariant map f : Hn → Hn and any γ ∈ Γ0 , µ(ρ(γ)) = d(p0 , ρ(γ) · p0 )   ≤ d f (p0 ), ρ(γ) · f (p0 ) + 2 d p0 , f (p0 )   = d f (p0 ), f (j(γ) · p0 ) + 2 d p0 , f (p0 )  ≤ Lip(f ) d(p0 , j(γ) · p0 ) + 2 d p0 , f (p0 )  = Lip(f ) µ(j(γ)) + 2 d p0 , f (p0 ) . 7.2. A refinement of Theorem 1.8. Let Γ0 be a discrete group. In Sections 7.5 and 7.6, we shall refine Theorem 1.8 by establishing the following implications for any pair (j, ρ) ∈ Hom(Γ0 , G)2 with j geometrically finite. We refer to Definitions 1.1 and 1.7 for the notions of cusp-deterioration and left admissibility; recall that any ρ is cusp-deteriorating if j is convex cocompact. 68 FRANÇOIS GUÉRITAUD AND FANNY KASSEL if 5 holds px 1 C(j, ρ) < 1 KS +3 C(j, ρred ) < 1 KS   2 (j, ρ) left admissible KS +3 (j, ρred ) left admissible KS Cµ (j, ρ) < 1 KS +3 Cµ (j, ρred ) < 1 KS C ′ (j, ρ) < 1 and ρ cusp-deteriorating +3 C ′ (j, ρred ) < 1 ρ cusp-deteriorating iv   5 iii   4 +3 ρred cusp-deteriorating We define the “reductive part” ρred of ρ as in Section 6.1.2: in the generic case when ρ is reductive (Definition 4.9), we set ρred := ρ. In the degenerate case when ρ is nonreductive, we fix a Levi factor M A of the stabilizer P in G of the fixed point at infinity of ρ(Γ0 ) (see Section 7.6), denote by π : P → M A the natural projection, and set ρred := π ◦ ρ, so that ρred is reductive and preserves an oriented geodesic line A ⊂ Hn , depending only on M A. The implications 1 ⇒ 3 and i ⇒ iii ⇒ iv are immediate consequences of (7.5), while 3 ⇒ 4 ⇒ 5 and iii ⇒ v follow from (7.5) and from the estimate µ(g k ) = 2 log k + O(1) for parabolic g (Lemma 2.6); these implications do not require any geometrical finiteness assumption on j. The implications 1 ⇒ i and i ⇒ 1 , the latter assuming 5 , are immediate consequences of Lemma 6.6. We shall explain: • • • 3 iv 2 ⇒ ⇒ ⇒ ii   3 i and iii ⇒ ii and 2 ⇒ 5 in Section 7.3, and ii ⇒ i in Section 7.5, ii , 3 ⇒ iii , 4 ⇒ iv , and 5 ⇒ v in Section 7.6. 2 iii 7.3. General theory of properly discontinuous actions and sharpness. Before proving the implications above, we discuss the connection with the general theory of properly discontinuous actions on reductive homogeneous spaces. The group G endowed with the transitive action of G × G by right and left multiplication identifies with the homogeneous space (G × G)/Diag(G), where Diag(G) is the diagonal of G × G. Let K be the stabilizer in G of the basepoint p0 of (7.1): it is a maximal compact subgroup of G = PO(n, 1), isomorphic to O(n). Let A be a one-parameter subgroup of G whose nontrivial elements are hyperbolic, all pure translations, with a translation axis A v MAXIMALLY STRETCHED LAMINATIONS 69 passing through p0 . Choose an endpoint ξ ∈ ∂∞ Hn of A and let A+ be the subsemigroup of A sending p0 into the geodesic ray [p0 , ξ). Then the Cartan decomposition G = KA+ K holds: any element g ∈ G may be written as g = kak ′ for some k, k′ ∈ K and a unique a ∈ A+ (see [H, Th. IX.1.1]). The Cartan projection µ of (7.1) is the projection onto A+ composed with an appropriate identification of A+ with R+ (namely the restriction of λ to A+ ). Likewise, the group G × G admits the Cartan decomposition G × G = (K × K)(A+ × A+ )(K × K) , with Cartan projection µ• = µ × µ : G × G −→ R+ × R+ . The general properness criterion of Benoist [Be1] and Kobayashi [Ko3] states, in this context, that a closed subgroup Γ of G × G acts properly on G by right and left multiplication if and only if the set µ• (Γ) “drifts away from the diagonal at infinity”, in the sense that for any R > 0, there is a compact subset of R+ × R+ outside of which any point of µ• (Γ) lies at distance > R from the diagonal of R+ × R+ . Consider a group Γj,ρ 0 as in (1.3), with j injective and discrete. Then the properness criterion states that Γj,ρ 0 acts properly discontinuously on G (i.e. (j, ρ) is admissible in the sense of Definition 1.7) if and only if (7.6) for any R ≥ 0, |µ(j(γ)) − µ(ρ(γ))| > R for almost all γ ∈ Γ0 µ(ρ(γ)) < µ(j(γ)) − R for almost all γ ∈ Γ0 , (i.e. for all γ ∈ Γ0 but finitely many exceptions). In particular, this gives the implications 3 ⇒ 2 and iii ⇒ ii of Section 7.2 above. It also gives 2 ⇒ 5 by the contrapositive: if ρ is not cusp-deteriorating, then there exists an element γ ∈ Γ0 with j(γ), ρ(γ) both parabolic, hence j(γ k ) = 2 log k + O(1) and ρ(γ k ) = 2 log k + O(1) as k → +∞, violating (7.6). (Note that we needed no geometrical finiteness assumption on j so far.) By [Ka2, Th. 1.3], if Γ0 is residually finite (for instance finitely generated) j,ρ and Γj,ρ 0 acts properly discontinuously on G, then the set µ• (Γ0 ) lies on one side only of the diagonal of R+ × R+ , up to a finite number of points. This means, up to switching j and ρ, that condition (7.6) is in fact equivalent to the following stronger condition: (7.7) for any R ≥ 0, and that properness implies λ(ρ(γ)) < λ(j(γ)) for all γ ∈ Γ0 (using (7.4)). Condition (7.7) is a necessary and sufficient condition for left admissibility in the sense of Definition 1.7; right admissibility is obtained by switching j and ρ. The implication 2 ⇒ 3 of Section 7.2 for geometrically finite j (which will be proved in Sections 7.5 and 7.6 below) can be interpreted as follows. Theorem 7.1. Let Γ be a discrete subgroup of G × G such that the set µ• (Γ) lies below the diagonal of R+ × R+ (up to a finite number of points) and such that the projection of Γ to the first factor of G × G is geometrically finite. Then Γ acts properly discontinuously on G by right and left multiplication if and only if there are constants C < 1 and D ∈ R such that µ(γ2 ) ≤ C µ(γ1 ) + D 70 FRANÇOIS GUÉRITAUD AND FANNY KASSEL for all γ = (γ1 , γ2 ) ∈ Γ. The point of Theorem 7.1 is that if Γ acts properly discontinuously on G, then the set µ• (Γ) “drifts away from the diagonal at infinity” linearly; in other words, Γ is sharp in the sense of [KK, Def. 4.2]. In particular, Theorem 7.1 corroborates the conjecture [KK, Conj. 4.10] that any discrete group acting properly discontinuously and cocompactly on a reductive homogeneous space should be sharp. Sharpness has analytic consequences on the discrete spectrum of the (nonelliptic) Laplacian defined by the natural pseudoRiemannian structure of signature (n, n(n − 1)/2) on the quotients of G: see [KK]. 7.4. Properness and the topology of the quotients of G = PO(n, 1). Let us make two side remarks. • First, here is for convenience a short proof of the properness criterion (7.6) of Benoist and Kobayashi in our setting. Note that there is no geometrical finiteness assumption here. Proof of the properness criterion of Benoist and Kobayashi. Suppose that condition (7.6) holds. Let C be a compact subset of G and let R := max µ(g). g∈C By the subadditivity (7.2) of µ, for any g ∈ C and γ ∈ Γ0 ,  µ ρ(γ)gj(γ)−1 ≥ |µ(j(γ)) − µ(ρ(γ))| − µ(g). By (7.6), the right-hand side is > R for almost all γ ∈ Γ0 , hence C ∩ ρ(γ)C j(γ)−1 = ∅ for almost all γ ∈ Γ0 . Thus the action of Γj,ρ 0 on G is properly discontinuous. Conversely, suppose that (7.6) does not hold, i.e. there exists R > 0 and a sequence (γm )m∈N of pairwise distinct elements of Γ0 such that |µ(j(γm )−1 ) − µ(ρ(γm )−1 )| ≤ R for all m ∈ N. By definition (7.1) of µ, this means that for any m ∈ N there is an element km ∈ K such that d(ρ(γm )−1 · p0 , km j(γm )−1 · p0 ) ≤ R. Since ρ(γm ) acts on Hn by an isometry, we obtain µ(ρ(γm )km j(γm )−1 ) = d(p0 , ρ(γm )km j(γm )−1 · p0 ) ≤ R. Therefore C ∩ ρ(γm )C j(γm )−1 6= ∅, where C is the compact subset of G consisting of the elements g with µ(g) ≤ R. This shows that the action of Γj,ρ  0 on G is not properly discontinuous. • Second, still without any geometrical finiteness assumption, here is a topological consequence of the inequality C(j, ρ) < 1; we refer to [DGK] for further developments and applications. Proposition 7.2. Let Γ0 be a discrete group and (j, ρ) ∈ Hom(Γ0 , G)2 a pair of representations with j injective and discrete. If C(j, ρ) < 1, then the group Γj,ρ 0 = {(j(γ), ρ(γ)) | γ ∈ Γ0 } acts properly discontinuously on G by right and left multiplication and the quotient is homeomorphic to a K-bundle over M := j(Γ0 )\Hn , where K ∼ = O(n) is a maximal compact subgroup of G = PO(n, 1). MAXIMALLY STRETCHED LAMINATIONS 71 Proof. The group K is the stabilizer in G of some point of Hn . Choose a (j, ρ)-equivariant map f : Hn → Hn with Lip(f ) < 1. For any p ∈ Hn , Lp := {g ∈ G | g · p = f (p)} is a right-and-left coset of K. An element g ∈ G belongs to Lp if and only if p is a fixed point of g −1 ◦ f ; since Lip(g −1 ◦ f ) = Lip(f ) < 1, such a fixed point exists and is unique, which shows that g belongs to exactly one set Lp . We denote this p by Π(g). The fibration Π : G → Hn is continuous: if h ∈ G is close enough to g so that d(Π(g), h−1 ◦f ◦Π(g)) ≤ (1−Lip(f )) ε, then h−1 ◦f takes the ε-ball centered at Π(g) to itself, hence Π(h) is within ε from Π(g). Moreover, Π : G → Hn is by construction (Γj,ρ 0 , j(Γ0 ))-equivariant: ρ(γ) Lp j(γ)−1 = Lj(γ)·p for all γ ∈ Γ0 and p ∈ Hn . Since the action of j(Γ0 ) on Hn is properly discontinuous, the action of Γj,ρ 0 on G is properly discontinuous. The fibration Π descends to a topological fibration of the quotient of G by Γj,ρ 0 , with base n M = j(Γ0 )\H and fiber K. Note that for constant ρ, i.e. ρ(Γ0 ) = {1}, this fibration naturally identifies with the orthonormal frame bundle of M .  7.5. Proof of the implications of Section 7.2 when ρ is reductive. We first consider the generic case where ρ is reductive (Definition 4.9), i.e. ρred = ρ. We have already explained the easy implications i ⇒ iii ⇒ iv and iii ⇒ v (Section 7.2), as well as iii ⇒ ii which is an immediate consequence of the properness criterion of Benoist and Kobayashi (Section 7.3). We now explain iv ⇒ iii and ii ⇒ i . The implication iv ⇒ iii is an immediate consequence of the following equality. Lemma 7.3. Let Γ0 be a discrete group and (j, ρ) ∈ Hom(Γ0 , G)2 a pair of representations. If ρ is reductive, then C ′ (j, ρ) = Cµ (j, ρ). Proof. By (7.5), we always have C ′ (j, ρ) ≤ Cµ (j, ρ). Let us prove the converse inequality. If ρ is reductive, then by [AMS, Th. 4.1] and [Be2, Lem. 2.2.1] there are a finite subset F of Γ0 and a constant D ≥ 0 with the following property: for any γ ∈ Γ0 there is an element f ∈ F such that |µ(ρ(γf )) − λ(ρ(γf ))| ≤ D (the element γf is proximal — see [Be2]). Then (7.2) and (7.3) imply where we set µ(ρ(γ)) ≤ ≤ ≤ ≤ ≤ µ(ρ(γf )) + µ(ρ(f )) λ(ρ(γf )) + D + µ(ρ(f )) C ′ (j, ρ) λ(j(γf )) + D + µ(ρ(f )) C ′ (j, ρ) µ(j(γf )) + D + µ(ρ(f )) C ′ (j, ρ) µ(j(γ)) + c,  c := D + max C ′ (j, ρ) µ(j(f )) + µ(ρ(f )) < ∞. f ∈F Thus Cµ (j, ρ) ≤ C ′ (j, ρ), which completes the proof.  72 FRANÇOIS GUÉRITAUD AND FANNY KASSEL The implication ii ⇒ i (or its contrapositive) for geometrically finite j is a consequence of the existence of a maximally stretched lamination when C(j, ρ) ≥ 1 (Theorem 1.3). We first establish the following. Lemma 7.4. Let Γ0 be a discrete group and (j, ρ) ∈ Hom(Γ0 , G)2 a pair of representations with j geometrically finite. If ρ is reductive and C(j, ρ) ≥ 1, then there is a sequence (γk )k∈N of pairwise distinct elements of Γ0 such that µ(ρ(γk )) − C(j, ρ) µ(j(γk )) is uniformly bounded from below; in particular (using Lemma 7.3 and (7.5)), (7.8) C ′ (j, ρ) = Cµ (j, ρ) = C(j, ρ). The equality C ′ (j, ρ) = C(j, ρ) is Corollary 1.12, which has already been proved in Section 5.4 when the stretch locus E(j, ρ) is nonempty. Here we do not make any assumption on E(j, ρ). Proof of Lemma 7.4. If C(j, ρ) = 1 and ρ is not cusp-deteriorating, then there exists γ ∈ Γ0 with j(γ) and ρ(γ) both parabolic. Since µ(ρ(γ k )) and µ(j(γ k )) are both equal to 2 log(k) + O(1) as k → +∞ (Lemma 2.6), the sequence (µ(ρ(γk )) − C(j, ρ) µ(j(γk )))k∈N is uniformly bounded from below. If C(j, ρ) > 1 or if C(j, ρ) = 1 and ρ is cusp-deteriorating, then by Theorem 1.3 there is a (j, ρ)-equivariant map f : Hn → Hn with minimal Lipschitz constant C(j, ρ) that stretches maximally some geodesic line ℓ of Hn whose image in j(Γ0 )\Hn lies in a compact part of the convex core. Consider a sequence (γk )k∈N of pairwise distinct elements of Γ0 such that d(j(γk ) · p0 , ℓ) is uniformly bounded by some constant R > 0. For k ∈ N, let yk be the closest-point projection to ℓ of j(γk ) · p0 . If p0 ∈ Hn is the basepoint defining µ in (7.1) and if we set ∆ := d(p0 , f (y0 )) + d(p0 , f (p0 )), then the triangle inequality implies µ(ρ(γk )) = d(p0 , ρ(γk ) · p0 )  ≥ d f (y0 ), ρ(γk ) · f (p0 ) − ∆  = d f (y0 ), f (j(γk ) · p0 ) − ∆   ≥ d f (y0 ), f (yk ) − d f (yk ), f (j(γk ) · p0 ) − ∆ ≥ C(j, ρ) d(y0 , yk ) − C(j, ρ) R − ∆  ≥ C(j, ρ) d(p0 , j(γk ) · p0 ) − 2R − C(j, ρ) R − ∆ = C(j, ρ) µ(j(γk )) − 3 C(j, ρ) R − ∆. Thus the sequence (µ(ρ(γk )) − C(j, ρ) µ(j(γk )))k∈N is uniformly bounded from below.  We can now prove the implication ii ⇒ i . Corollary 7.5. Let Γ0 be a discrete group and (j, ρ) ∈ Hom(Γ0 , G)2 a pair of representations with j geometrically finite. If ρ is reductive and (j, ρ) left admissible, then C(j, ρ) < 1. Proof. Assume that ρ is reductive and (j, ρ) left admissible. By [Ka2, Th. 1.3], condition (7.7) is satisfied, hence Cµ (j, ρ) ≤ 1. Suppose by contradiction that C(j, ρ) ≥ 1. By Lemma 7.4, we have Cµ (j, ρ) = C(j, ρ) = 1 and there is a sequence (γk ) ∈ (Γ0 )N of pairwise distinct elements with |µ(ρ(γk )) − µ(j(γk ))| MAXIMALLY STRETCHED LAMINATIONS 73 uniformly bounded. This contradicts the properness criterion (7.6) of Benoist and Kobayashi.  7.6. Proof of the implications of Section 7.2 when ρ is nonreductive. We now prove the implications of Section 7.2 for geometrically finite j when ρ is nonreductive (Definition 4.9). We have already explained the easy implications 1 ⇒ 3 ⇒ 4 ⇒ 5 (Section 7.2) and 3 ⇒ 2 ⇒ 5 (Section 7.3), as well as 1 ⇒ i and i ⇒ 1 under the cusp-deterioration assumption 5 (Section 7.2). Moreover, in Section 7.5 we have established the implications i ⇔ ii ⇔ iii ⇔ iv ⇒ v for the “reductive part” ρred of ρ. Therefore, we only need to explain the “horizontal” implications 2 ⇒ ii , 3 ⇒ iii , 4 ⇒ iv , and 5 ⇒ v Let ρ ∈ Hom(Γ0 , G) be nonreductive. The group ρ(Γ0 ) has a unique fixed point ξ in the boundary at infinity ∂∞ Hn of Hn . Let P be the stabilizer of ξ in G: it is a proper parabolic subgroup of G. Choose a Levi decomposition P = (M A) ⋉ N , where A ∼ = R∗+ is a Cartan subgroup of G (i.e. a one-parameter subgroup of purely translational, commuting hyperbolic elements), M ∼ = O(n − 1) is a compact subgroup of G such that M A is the centralizer of A in G, and N ∼ = Rn−1 is the unipotent radical of P . For instance, for n = 2 (resp. n = 3), the identity component of the group G identifies with PSL2 (R) (resp. with PSL2 (C)), and we can take A to be the projectivized real diagonal matrices, N the projectivized upper triangular unipotent matrices, and the identity component of M to be the projectivized diagonal matrices with entries of module 1. We set ρred := π ◦ ρ, where π : P → M A is the natural projection. The implications 2 ⇒ ii , 3 ⇒ iii , 4 ⇒ iv , and 5 ⇒ v of Section 7.2 are consequences of the following easy observation. Lemma 7.6. After possibly changing the basepoint p0 ∈ Hn of (7.1) (which modifies µ only by a bounded additive amount, by the triangle inequality), we have λ(g) = λ(π(g)) = µ(π(g)) ≤ µ(g) for all g ∈ P . Proof. Take the basepoint p0 ∈ Hn on the geodesic line A preserved by M A, which is pointwise fixed by M and on which the elements of A act by translation. Then λ(π(g)) = µ(π(g)) for all g ∈ P . The projection onto A along horospheres centered at ξ is (P, π(P ))-equivariant, and restricts to an isometry on any line ending at ξ: therefore, if g is hyperbolic, preserving such a line, then π(g) translates along A by λ(g) units of length, yielding λ(g) = λ(π(g)). If g is parabolic or elliptic, then λ(g) = 0 and g preserves  each horosphere centered at ξ, hence λ(π(g)) = 0. Proof of 3 ⇒ iii and 4 ⇒ iv . Lemma 7.6 and (7.5) imply C ′ (j, ρ) = C ′ (j, ρred ) = Cµ (j, ρred ) ≤ Cµ (j, ρ), which immediately yields the implications.  Proof of 2 ⇒ ii . Assume that (j, ρ) is left admissible. By [Ka2], the pair (j, ρ) is not right admissible and the stronger form (7.7) of the properness 74 FRANÇOIS GUÉRITAUD AND FANNY KASSEL criterion of Benoist and Kobayashi holds. Since µ ◦ π ≤ µ by Lemma 7.6, the condition (7.7) also holds for ρred = π◦ρ, hence (j, ρred ) is left admissible.  Proof of 5 ⇒ v . Assume that ρ is cusp-deteriorating. For any γ ∈ Γ0 with j(γ) parabolic, ρred (γ) is not hyperbolic, otherwise ρ(γ) would be hyperbolic too by Lemma 7.6, and it is not parabolic since ρred takes values in the group M A which has no parabolic element.  7.7. Deformation of properly discontinuous actions. Theorems 1.9 and 1.11 follow from Theorem 1.8 (properness criterion) and Proposition 1.5 (continuity of (j, ρ) 7→ C(j, ρ) for convex cocompact j), together with a classical cohomological argument for cocompactness. Proof of Theorems 1.9 and 1.11. Let Γ be a finitely generated discrete subgroup of G × G acting properly discontinuously on G by right and left multiplication. By the Selberg lemma [Se, Lem. 8], there is a finite-index subgroup Γ′ of Γ that is torsion-free. By [KR] and [Sa] (case n = 2) and [Ka2] (general case), up to switching the two factors of G × G, the group Γ′ is of the form Γj,ρ 0 as in (1.3), where Γ0 is a torsion-free discrete subgroup of G and j, ρ ∈ Hom(Γ0 , G) are two representations of Γ0 in G with j injective and discrete, and (j, ρ) is left admissible in the sense of Definition 1.7. Assume that j is convex cocompact. Then C(j, ρ) < 1 by Theorem 1.8. By Proposition 1.5 and the fact that being convex cocompact is an open condition (see [Bo2, Prop. 4.1] or Proposition B.1), there is a neighborhood U ⊂ Hom(Γ, G × G) of the natural inclusion such that for any ϕ ∈ U , the group ϕ(Γ′ ) is of the ′ ′ form Γj0 ,ρ for some (j ′ , ρ′ ) ∈ Hom(Γ0 , G)2 with j ′ convex cocompact and C(j ′ , ρ′ ) < 1. In particular, ϕ(Γ′ ) is discrete in G × G and acts properly discontinuously on G by Theorem 1.8, and the same conclusion holds for ϕ(Γ). Assume that the action of Γ on G is cocompact. We claim that the action of ϕ(Γ) on G is cocompact for all ϕ ∈ U . Indeed, let Γ′ = Γj,ρ 0 be a finite-index subgroup of Γ as above; it is sufficient to prove that the ′ ′ action of ϕ(Γ′ ) is cocompact for all ϕ ∈ U . Since ϕ(Γ′ ) is of the form Γj0 ,ρ with j ′ injective, the group ϕ(Γ′ ) has the same cohomological dimension as Γ′ . We then use the fact that when a torsion-free discrete subgroup of G × G acts properly discontinuously on G, it acts cocompactly on G if and only if its cohomological dimension is equal to the dimension of the Riemannian symmetric space of G, namely n in our case (see [Ko1, Cor. 5.5]). ′ ′ (Alternatively, the cocompactness of the action of Γj0 ,ρ on G also follows from Proposition 7.2 and from the cocompactness of j ′ (Γ0 ).) Finally, assume that the action of Γ on G is free. This means that for any γ ∈ Γ r {1}, the elements pr1 (γ) and pr2 (γ) are not conjugate in G, where pri : G × G → G denotes the i-th projection. In fact, since the action of Γ on G is properly discontinuous, pr1 (γ) and pr2 (γ) can never be conjugate in G when γ is of infinite order. Therefore freeness is seen exclusively on torsion elements. We claim that Γ has only finitely many conjugacy classes of torsion elements. Indeed, Γ has a finite-index subgroup of the form Γj,ρ 0 with j injective and convex cocompact (up to switching the two factors of G × G), and a convex cocompact subgroup of G (or more generally a geometrically finite subgroup) has only finitely many conjugacy classes of torsion elements MAXIMALLY STRETCHED LAMINATIONS 75 (see [Bo1]). For any nontrivial torsion element γ ∈ Γ, there is a neighborhood Uγ ⊂ Hom(Γ, G × G) such that for all ϕ ∈ Uγ , the elements pr1 (ϕ(γ)) and pr2 (ϕ(γ)) are not conjugate in G; then pr1 (ϕ(γ ′ )) and pr2 (ϕ(γ ′ )) are also not conjugate for any Γ-conjugate γ ′ of γ.  The same argument, replacing Proposition 1.5 (continuity of (j, ρ) 7→ C(j, ρ) for convex cocompact j) by Proposition 6.1.(1) (openness of the condition C(j, ρ) < 1 for geometrically finite j and cusp-deteriorating ρ in dimension n ≤ 3), yields the following. Theorem 7.7. For G = PO(2, 1) or PO(3, 1) (i.e. PSL2 (R) or PSL2 (C) up to index two), let Γ be a discrete subgroup of G × G acting properly discontinuously on G, with a geometrically finite quotient (Definition 1.10). There is a neighborhood U ⊂ Homdet (Γ, G × G) of the natural inclusion such that for any ϕ ∈ U , the group ϕ(Γ) is discrete in G × G and acts properly discontinuously on G, with a geometrically finite quotient; moreover, this quotient is compact (resp. is convex cocompact, resp. is a manifold) if the initial quotient of G by Γ was. The set Homdet (Γ, G × G) is defined as follows. We have seen that the ρ,j group Γ has a finite-index subgroup Γ′ of the form Γj,ρ 0 or Γ0 , where Γ0 is a discrete subgroup of G and j, ρ ∈ Hom(Γ0 , G) are two representations of Γ0 in G with j injective and geometrically finite and (j, ρ) left admissible in the sense of Definition 1.7. By Lemma 2.7, the representation ρ is cuspdeteriorating with respect to j in the sense of Definition 1.1. We define Homdet (Γ, G × G) to be the set of group homomorphisms from Γ to G × G j,ρ whose restriction to Γ′ is of the form (j ′ , ρ′ ) (if Γ′ ∼ = Γ0 ) or (ρ′ , j ′ ) (if ρ,j Γ′ ∼ = Γ0 ) with j ′ injective and geometrically finite, of the cusp type of j, ′ and ρ cusp-deteriorating with respect to j. It is necessary to restrict to Homdet (Γ, G × G) in Theorem 7.7, for the following reasons: • as mentioned in the introduction, for a given j with cusps, the constant representation ρ = 1 can have small, non-cusp-deteriorating deformations ρ′ , for which (j, ρ′ ) is nonadmissible; • if we allow for small deformations j ′ of j with a different cusp type than j (fewer cusps), then the pair (j, ρ) can have small, nonadmissible deformations (j ′ , ρ′ ) with ρ′ cusp-deteriorating with respect to j ′ : this shows that we must fix the cusp type. Note that properly discontinuous actions on G = PO(3, 1) of finitely generated groups Γ = Γj,ρ 0 with j geometrically infinite do not deform into properly discontinuous actions in general, for the group j(Γ0 ) (e.g. the fiber group of a hyperbolic surface bundle over the circle) may have small deformations j ′ (Γ0 ) that are not even discrete (e.g. small perturbations of a nearby cusp group in the sense of [M1]). It would be interesting to know whether in this case one can define an analogue of Homdet (Γ, G×G) in which admissibility becomes again an open condition; see also Section C.2. 7.8. Interpretation of Theorem 1.9 in terms of (G, X)-structures. We can translate Theorem 1.9 in terms of geometric structures, in the sense of Ehresmann and Thurston, as follows. We set X = G = PO(n, 1) and 76 FRANÇOIS GUÉRITAUD AND FANNY KASSEL G = G × G, where G × G acts on G by right and left multiplication. Let e . Recall that a (G, X)-structure N be a manifold with universal covering N on N is a (maximal) atlas of charts on N with values in X such that the transition maps are given by elements of G. Such a structure is equivalent to a pair (h, D) where h : π1 (N ) → G is a group homomorphism called the e → X an h-equivariant local diffeomorphism called the holonomy and D : N developing map; the pair (h, D) is unique modulo the natural action of G by  g · (h, D) = gh(·)g−1 , gD . A (G, X)-structure on N is said to be complete if the developing map is a covering; this is equivalent to a notion of geodesic completeness for the natural pseudo-Riemannian structure induced by the Killing form of the Lie algebra of G (see [Go]). For n > 2, the fundamental group of G0 = PO(n, 1)0 is finite, hence completeness is equivalent to the fact that the (G, X)-structure identifies N with the quotient of X by some discrete subgroup Γ of G acting properly discontinuously and freely on X, up to a finite covering. For n = 2, this characterization of completeness still holds for compact manifolds N [KR, Th. 7.2]. Therefore, Theorem 1.9 can be restated as follows. Corollary 7.8. Let X = G = PO(n, 1) and G = G×G, acting on X by right and left multiplication. The set of holonomies of complete (G, X)-structures on any compact manifold N is open in Hom(π1 (N ), G). We note that for a compact manifold N , the so-called Ehresmann–Thurston principle asserts that the set of holonomies of all (not necessarily complete) (G, X)-structures on N is open in Hom(π1 (N ), G) (see [T1]). For n = 2, Klingler [Kl] proved that all (G, X)-structures on N are complete, which implies Corollary 7.8. For n > 2, it is not known whether all (G, X)-structures on N are complete; it has been conjectured to be true at least for n = 3 [DZ]. The question is nontrivial since the Hopf–Rinow theorem does not hold for non-Riemannian manifolds. 8. A generalization of the Thurston metric on Teichmüller space In this section we prove Proposition 1.13, which generalizes the Thurston metric on Teichmüller space to higher dimension, in a geometrically finite setting. (Corollary 1.12 has already been proved as part of Lemma 7.4; see also Section 5.4.) 8.1. An asymmetric metric on the space of geometrically finite structures. Let M be a hyperbolic n-manifold and T (M ) the space of conjugacy classes of geometrically finite representations of Γ0 := π1 (M ) into G = PO(n, 1) with the homeomorphism type and cusp type of M . In this section we assume that M contains more than one essential closed curve. We define a function dTh : T (M ) × T (M ) → R as in (1.6): for any j, ρ ∈ T (M ),   δ(ρ) , dTh (j, ρ) := log C(j, ρ) δ(j) MAXIMALLY STRETCHED LAMINATIONS 77 where δ(j) is the critical exponent of j. Recall that, by definition,  1 (8.1) δ(j) := lim sup log # j(Γ0 ) · p ∩ Bp (R) , R→+∞ R where p is any point of Hn and Bp (R) denotes the ball of radius R centered at p in Hn . We have δ(j) ∈ (0, n − 1] [Bea, S1] and the limsup is in fact P −s d(p,j(γ)·p) converges a limit [Pat, S1, Ro1]. The Poincaré series γ∈Γ0 e for s > δ(j) and diverges for s ≤ δ(j). Equivalently, δ(j) is the Hausdorff dimension of the limit set of j(Γ0 ) [S1, S2]. If M is a surface of finite volume, then T (M ) is the Teichmüller space of M , we have δ ≡ 1 on T (M ), and dTh is the Thurston metric on Teichmüller space, which was introduced in [T2] (see Section 1.5). This metric is asymmetric: in general dTh (j, ρ) 6= dTh (ρ, j), see [T2, § 2]. Lemma 8.1. The function dTh : T (M ) × T (M ) → R is continuous as soon as M is convex cocompact or all the cusps have rank ≥ n − 2. In particular, it is always continuous if n ≤ 3. Proof. If M is convex cocompact or all the cusps have rank ≥ n − 2, then the function (j, ρ) 7→ C(j, ρ) is continuous on T (M ) × T (M ) by Proposition 1.5, Lemma 2.7, and Proposition 6.1.(3). Moreover, convergence in T (M ) implies geometric convergence (see Proposition B.3), and so δ is continuous on T (M ) by [M2, Th. 7.3].  The following remark justifies the introduction of the correcting factor δ(ρ)/δ(j) in the definition (1.6) of dTh . Remark 8.2. If M has infinite volume, then log C(j, ρ) can take negative values, and log C(j, ρ) = 0 does not imply j = ρ. Proof. The following example is taken from [T2, proof of Lemma 3.4]; see also [PT]. Let M be a pair of pants, i.e. a hyperbolic surface of genus 0 with three funnels. Let α be an infinite embedded geodesic of M whose two ends go out to infinity in the same funnel, and let α′ be another nearby geodesic (see Figure 9). Cutting out the strip between α and α′ and gluing back so that the endpoints of the common perpendicular to α, α′ are identified yields a new hyperbolic surface M ′ such that two boundary components of the convex core of M ′ have the same lengths as in M , and the third one is shorter. There is a 1-Lipschitz map between M and M ′ , and the corresponding holonomies j 6= ρ satisfy log C(j, ρ) = 0. In fact, it is easy to see that after repeating the process with all three funnels we obtain an element ρ′ ∈ T (M ) with log C(j, ρ′ ) < 0.  With the correcting factor δ(ρ)/δ(j), the following holds. Lemma 8.3. For any j, ρ, j1 , j2 , j3 ∈ T (M ), (1) dTh (j, ρ) ≥ 0, (2) dTh (j1 , j3 ) ≤ dTh (j1 , j2 ) + dTh (j2 , j3 ). Proof. Let f : Hn → Hn be a (j, ρ)-equivariant Lipschitz map. For any p ∈ Hn , γ ∈ Γ0 , and R > 0, if j(γ) · p ∈ Bp (R/Lip(f )), then  ρ(γ) · f (p) = f j(γ) · p ∈ Bf (p) (R), 78 FRANÇOIS GUÉRITAUD AND FANNY KASSEL α α′ Figure 9. The strip between the geodesics α and α′ can be collapsed to create a new hyperbolic metric with shorter curves. In general, the closed geodesic at the bottom (met by α, α′ ) will not collapse to a closed geodesic of the new metric. hence δ(ρ) ≥ δ(j)/Lip(f ) by definition (8.1) of δ. Then (1) follows by letting Lip(f ) tend to C(j, ρ) and taking the logarithm. The triangle inequality (2) follows from the general inequality Lip(f1 ◦ f2 ) ≤ Lip(f1 ) Lip(f2 ).  As in Section 1.5, let T (M )Zs be the subset of T (M ) consisting of elements j such that the Zariski-closure of j(Γ0 ) in G is simple (e.g. equal to G). In order to prove Proposition 1.13, it remains to prove the following. Proposition 8.4. If j, ρ ∈ T (M )Zs are distinct, then dTh (j, ρ) > 0. Note that Proposition 8.4 is not true when j or ρ does not belong to T (M )Zs : for instance, if j ∈ T (M )Zs takes values in PO(2, 1) ⊂ G = PO(n, 1), then we may multiply it by a homomorphism with values in the centralizer O(n − 2) of PO(2, 1) in G without changing the set {µ(j(γ)) | γ ∈ Γ0 }; this yields a homomorphism ρ in T (M ) r T (M )Zs with dTh (j, ρ) = dTh (ρ, j) = 0. 8.2. Proof of Proposition 8.4. In our setting, the rigidity of the marked length spectrum {λ(j(γ)) | γ ∈ Γ0 } is well known: see [Kim, Th. 2]. We shall use a slightly stronger property, namely the rigidity of the projectivized marked length spectrum. Lemma 8.5. Let j, ρ ∈ T (M )Zs and C > 0. If λ(ρ(γ)) = C λ(j(γ)) for all γ ∈ Γ0 , then j = ρ in T (M )Zs . Proof. For i ∈ {1, 2}, let pri : G × G → G be the i-th projection. Let H1 (resp. H2 ) be the identity component (for the real topology) of the Zariski closure of j(Γ0 ) (resp. ρ(Γ0 )) in G, and let H be the identity component of the Zariski closure of (j, ρ)(Γ0 ) in G × G. Since H1 is simple, the kernel of pr2 |H is either {1} or H1 . It cannot be H1 , otherwise H = H1 × H2 would be semisimple of real rank 2, and so by [Be2] the cone spanned by {(λ(ρ(γ)), λ(j(γ))) | γ ∈ Γ0 } would have nonempty interior, contradicting the fact that λ(ρ(γ)) = C λ(j(γ)) for all γ ∈ Γ0 . Therefore pr2 |H is injective, and MAXIMALLY STRETCHED LAMINATIONS 79 similarly pr1 |H is injective, and so there is a (j, ρ)-equivariant isomorphism H1 → H2 . Note that any noncompact simple subgroup Hi of G = PO(n, 1) is conjugate to some PO(k, 1) embedded in the standard way, for k ≥ 2, up to finite index. Indeed, any Hi -orbit in Hn is a totally geodesic subspace of Hn [H, Chap. IV, Th. 7.2], hence is a copy of Hk for some k ≥ 2, and there is only one way to embed Hk into Hn up to isometry (this is clear in the hyperboloid model). Therefore the (j, ρ)-equivariant isomorphism H1 → H2 above is given by conjugation by some element of G, and j and ρ are conjugate under G.  Proof of Proposition 8.4. Assume that dTh (j, ρ) = 0 for some j 6= ρ in T (M )Zs , and let C := C(j, ρ) = δ(j)/δ(ρ). By Lemma 8.5, there exists γ0 ∈ Γ0 such that λ(ρ(γ0 )) < C λ(j(γ0 )). Let f : Hn → Hn be a (j, ρ)equivariant, C-Lipschitz map (Lemma 4.10). The translation axis A ⊂ Hn of j(γ0 ) cannot be C-stretched by f since λ(ρ(γ0 )) < Cλ(j(γ0 )). Therefore we can find p, q ∈ A and ∆ > 0 such that d(f (p), f (q)) ≤ C d(p, q)−(2+C)∆. Let Bp (resp. Bq ) be the ball of diameter ∆ centered at p (resp. q), so that d(f (p′ ), f (q ′ )) ≤ C d(p′ , q ′ ) − ∆ for all p′ ∈ Bp and q ′ ∈ Bq . We can assume moreover that p, q are close enough in the sense that no segment [p′ , q ′ ] with p′ ∈ Bp and q ′ ∈ Bq intersects any ball j(γ)·Bp or j(γ)·Bq with γ ∈ Γ0 r{1}. → Let Ue be the open set of all vectors (x, − v ) in the unit tangent bundle T 1 Hn → − such that x ∈ Bp and expx (R+ v ) intersects Bq . Let X := j(Γ0 )\T 1 Hn be the unit tangent bundle of the quotient manifold j(Γ0 )\Hn , and U ⊂ X the projection of Ue. For γ ∈ Γ0 with j(γ) hyperbolic, let Nγ be the number of times that the axis of j(γ) traverses U in X (see Figure 10); the triangle inequality yields λ(ρ(γ)) ≤ C λ(j(γ)) − Nγ ∆. (8.2) Let ν be the Bowen–Margulis–Sullivan probability measure on X (see [Ro2, Bp p Nγ copies of the axis Aj(γ) q Bq f f (p) f (q) Figure 10. Illustration of the proof of Proposition 8.4 when Nγ = 2. § 1.C]). We have ν(U ) > 0 since U intersects the projection of the axis of j(γ0 ); therefore we can find a continuous function ψ : X →R [0, 1] with compact support contained in U such that kψk∞ = 1 and ε := X ψ dν > 0. For any γ ∈ Γ0 with j(γ) hyperbolic and primitive (i.e. not a power of any other j(γ ′ )), we denote by νγ the uniform probability measure on X 80 FRANÇOIS GUÉRITAUD AND FANNY KASSEL supported on the axis of j(γ). Since the support of ψ is contained in Bp , Z Nγ diam(Bp ) =∆ . ψ dνγ ≤ kψk∞ · Nγ (8.3) λ(j(γ)) λ(j(γ)) X For R > 0, let ΓR,j be the set of elements γ ∈ Γ0 such that j(γ) is primitive 0 hyperbolic and λ(j(γ)) ≤ R. By [Ro2, Th. 5.1.1] (see also [L, DP] for special cases), Z X Z ψ dν = ε. (8.4) δ(j)R e−δ(j)R ψ dνγ −→ γ∈ΓR,j 0 R→+∞ X X Moreover, this convergence is still true if we replace ψ with the constant function equal to 1 on X [Ro2, Cor. 5.3], yielding (8.5) #(Γj,R 0 ) ∼ R→+∞ eδ(j)R . δ(j)R Combined, formulas (8.3), (8.4), (8.5) imply that the average value of Nγ /λ(j(γ)), ε for γ ranging over ΓR,j 0 , is ≥ 2∆ for all large enough R. Since Nγ 1 1 ≤ ≤ λ(j(γ)) d(p, q) − ∆ 2∆ ε for all γ ∈ ΓR,j 0 , this classically implies that a proportion ≥ 2 of elements ε , which by (8.2) entails γ ∈ ΓR,j satisfy Nγ /λ(j(γ)) ≥ 4∆ 0   ε ε λ(j(γ)) ≤ C − R. λ(ρ(γ)) ≤ Cλ(j(γ)) − Nγ ∆ ≤ C − 4 4 Thus  (C− ε )R,ρ  ε # Γ0 4 ≥ #(ΓR,j 0 ) 2 for all large enough R. Then (8.5) yields (C− 4ε ) δ(ρ) ≥ δ(j), hence C δ(ρ) δ(j) > 1.  9. The stretch locus in dimension 2 We now focus on results specific to dimension n = 2. We first consider the case C(j, ρ) > 1, for which we recover and extend two aspects of the classical theory [T2] of the Thurston metric on Teichmüller space. The first aspect is the chain recurrence of the lamination E(j, ρ), which we prove in Section 9.2. Building on chain recurrence, the second aspect is the upper semicontinuity of E(j, ρ) for the Hausdorff topology, namely E(j, ρ) ⊃ lim sup E(jk , ρk ) k→+∞ for any (jk , ρk ) → (j, ρ) with ρ and ρk reductive, which we prove in Section 9.3. We also consider the case C(j, ρ) < 1 and provide some evidence for Conjecture 1.4 (describing the stretch locus E(j, ρ)) in Section 9.4. In fact, we believe that chain recurrence (suitably defined) should probably also hold in higher dimension for C(j, ρ) > 1, but we shall use the classification of geodesic laminations on surfaces to prove it here. Semicontinuity should also hold in higher dimension, not only for C(j, ρ) > 1 but also MAXIMALLY STRETCHED LAMINATIONS 81 in some form for C(j, ρ) ≤ 1: this is natural to expect in view of Propositions 1.5 and 6.1.(3) (if the minimal Lipschitz constant varies continuously, so should the stretch locus). However, our proof hinges on chain recurrence and on the fact that f multiplies arc length along the leaves of the stretch locus: this property does not obviously have a counterpart when C(j, ρ) < 1 (the stretch locus being no longer a lamination in general), and is at any rate harder to prove (the Kirszbraun–Valentine theorem no longer applies as in Lemma 5.2). 9.1. Chain recurrence in the classical setting. We first recall the notion of chain recurrence and, for readers interested in the more technical aspects of [T2], we make the link between the “maximal, ratio-maximizing, chain recurrent lamination” µ(j, ρ) introduced by Thurston in the latter paper, and the stretch locus E(j, ρ) introduced in the present paper. On a hyperbolic surface S, a geodesic lamination is called recurrent if every half-leaf returns arbitrarily often, arbitrarily close to its starting point. In [T2], Thurston introduced the weaker notion of chain recurrence. Definition 9.1. A geodesic lamination L˙ on S is called chain recurrent if for every ṗ ∈ L˙ and ε > 0, there exists a simple closed geodesic G passing within ε of ṗ and staying ε-close to L˙ in the C 1 sense. By “ε-close in the C 1 sense” we mean that any unit-length segment of G lies ε-close to a segment of L˙ (for the Hausdorff metric). In particular, any recurrent lamination is chain recurrent. The following is well known. Fact 9.2. Any geodesic lamination on S consists of finitely many disjoint recurrent components, together with finitely many isolated leaves spiraling from one recurrent component to another (possibly the same). The total number of recurrent components and of isolated leaves can be bounded by an integer depending only on the topology of S. By Fact 9.2, chain recurrence implies that for any ε > 0, any ṗ ∈ L˙ , and any direction of travel along L˙ from ṗ, one can return to ṗ (with the same direction of travel) by following leaves of L˙ and occasionally jumping to nearby leaves within distance ε. (For example, if L˙ has an isolated leaf spiraling to a simple closed curve and no leaf spiraling out, then L˙ is not chain recurrent.) By Fact 9.2, the number of necessary ε-jumps can be bounded by a number m depending only on the topology of the surface, and the distances in-between the jumps can be taken arbitrarily large. In the sequel, we shall call a sequence of leaf segments, separated by a number ≤ m of ε-jumps, an ε-quasi-leaf of L˙ . The closing lemma (Lemma A.1) implies that conversely any ε-quasi-leaf can be ε-approximated, in the C 1 sense, by a simple closed geodesic. We now briefly discuss the relation to [T2] (this will not be needed from Lemma 9.3 onwards). Let S be a hyperbolic surface of finite volume. In [T2], Thurston associated to any pair (j, ρ) of distinct elements of the Teichmüller space T (S) of S (i.e. type-preserving, geometrically finite representations of Γ0 := π1 (S) into PO(2, 1) ∼ = PGL2 (R), of finite covolume, up to conjugation) a subset µ(j, ρ) of S, defined as the union of all chain recurrent laminations 82 FRANÇOIS GUÉRITAUD AND FANNY KASSEL L˙ that are ratio-maximizing, in the sense that there exists a locally C(j, ρ)Lipschitz map from a neighborhood of L˙ in (S, j) to a neighborhood of L˙ in (S, ρ), in the correct homotopy class, that multiplies arc length by C(j, ρ) on each leaf of L˙ . He proved that µ(j, ρ) is a lamination [T2, Th. 8.2], necessarily chain recurrent, and that this lamination is C(j, ρ)-stretched by some C(j, ρ)-Lipschitz homeomorphism (S, j) → (S, ρ), in the correct homotopy class, whose local Lipschitz constant is < C(j, ρ) everywhere outside of µ(j, ρ). Indeed, this last property follows from the existence of a concatenation of “stretch paths” going from j to ρ in T (S) [T2, Th. 8.5] and from the definition of stretch paths in terms of explicit homeomorphisms of minimal Lipschitz constant [T2, § 4]. Therefore, the preimage µ e(j, ρ) ⊂ H2 of Thurston’s chain recurrent lamination µ(j, ρ) ⊂ S ≃ j(Γ0 )\H2 contains the stretch locus E(j, ρ) that we have introduced in this paper. In fact, this inclusion is an equality, as the following variant of Lemma 5.10 shows (with Γ0 = π1 (S) and j, ρ ∈ T (S)). Lemma 9.3. (in dimension n = 2) Let Γ0 be a torsion-free discrete group and (j, ρ) ∈ Hom(Γ0 , G)2 a pair of representations with j geometrically finite. Let µ e ⊂ H2 be the preimage of some chain recurrent lamination on S. If µ e is maximally stretched by some (j, ρ)-equivariant Lipschitz map f : H2 → H2 , then µ e is contained in the stretch locus E(j, ρ). Proof. We proceed as in the proof of Lemma 5.10, but using closed quasileaves instead of recurrent leaves. Set C := C(j, ρ). Consider a geodesic segment [x, y] contained in µ. By chain recurrence and by the closing lemma (Lemma A.1), for any ε > 0 there is a simple closed geodesic G on (S, j) that passes within ε of x and is ε-close to an ε-quasi-leaf L of µ. We may assume that L consists of m or fewer leaf segments, of which one contains [x, y]. Let γ ∈ Γ0 correspond to the closed geodesic G. Then λ(j(γ)) = length(G) ≤ length(L) + mε, and since each leaf segment of L is C-stretched by f we see, using the closing lemma again, that   λ(ρ(γ)) ≥ C · length(L) − 3mε ≥ C · λ(j(γ)) − 4mε . By considering p, q, p′ , q ′ ∈ H2 such that p, q project to x, y ∈ L and p′ , q ′ to points within ε from x, y in G, we obtain, exactly as in the proof of Lemma 5.10, that for any f ′ ∈ F j,ρ , d(f ′ (p), f ′ (q)) ≥ C · d(p, q) − (4m + 4)Cε. This holds for any ε > 0, hence d(f ′ (p), f ′ (q)) = Cd(p, q) and p belongs to the stretch locus of f ′ .  9.2. Chain recurrence for C(j, ρ) > 1 in general. We now prove that the stretch locus E(j, ρ) is chain recurrent in a much wider setting than [T2], allowing j(Γ0 ) to have infinite covolume in G and ρ to be any representation of Γ0 in G with C(j, ρ) > 1 (not necessarily injective or discrete). Proposition 9.4. (in dimension n = 2) Let Γ0 be a torsion-free discrete group and (j, ρ) ∈ Hom(Γ0 , G)2 a pair of representations with j geometrically finite and ρ reductive (Definition 4.9). If C(j, ρ) > 1, then the image in S := j(Γ0 )\H2 of the stretch locus E(j, ρ) is a (nonempty) chain recurrent lamination. MAXIMALLY STRETCHED LAMINATIONS 83 Proof. Let f0 ∈ F j,ρ be optimal (Definition 4.12), with stretch locus E := E(j, ρ). By Theorem 1.3 and Lemma 4.10, we know that E is a nonempty, j(Γ0 )-invariant geodesic lamination. Suppose by contradiction that its image Ė in S = j(Γ0 )\H2 is not chain recurrent. We shall “improve” f0 by decreasing its stretch locus, which will be absurd. Given ṗ ∈ Ė and a direction (“forward”) of travel from ṗ, define the forward chain closure Ėp of ṗ in Ė as the subset of Ė that can be reached from ṗ, starting forward, by following ε-quasi-leaves of Ė for positive time, for any ε > 0. Clearly, Ėp is the union of a closed sublamination of Ė and of an open half-leaf issued from ṗ. If Ėp contains ṗ for all ṗ ∈ Ė and choices of forward direction, then for any ε > 0 we can find a closed ε-quasi-leaf of Ė through ṗ. Since Ė is not chain recurrent by assumption, this is not the case: we can therefore choose a point ṗ ∈ Ė and a direction of travel such that Ėp does not contain ṗ. Then Ėp is orientable: otherwise for any ε > 0 we could find an ε-quasileaf of Ė through ṗ by following a quasi-leaf from ṗ, “jumping” onto another (quasi)-leaf with the reverse orientation, and getting back to ṗ, which would contradict the fact that Ėp does not contain ṗ. Let Υ̇ be the lamination of S obtained by removing from Ėp the (isolated) half-leaf issued from ṗ. Then Υ̇ inherits an orientation from the “forward” orientation of Ėp . No leaf of Ė r Υ̇ can be outgoing from Υ̇, otherwise it would automatically belong to Υ̇. But at least one leaf of Ė r Υ̇ is incoming towards Υ̇: namely, the leaf ℓ̇ containing ṗ. The geodesic lamination Υ̇ fills some subsurface Σ ⊂ S with geodesic boundary (possibly reduced to a single closed geodesic). Let U̇ ⊂ S be a uniform neighborhood of Σ, with the same topological type as Σ, such that U̇ ∩ Ė is the union of the oriented lamination Υ̇ and of some (at least one) incoming half-leaves. Up to shifting the point ṗ along its leaf ℓ̇, we may assume that ṗ ∈ ∂ U̇ . Let U and Υ be the (full) preimages of U̇ and Υ̇ in H2 . To reach a contradiction, we shall modify f0 on U . The modification on Υ itself is simply to replace f |Υ with fε := f0 ◦ Φ−ε , where (Φt )t∈R is the flow on the oriented lamination Υ and ε > 0 is small enough. We make the following two claims, for C := C(j, ρ): (i) the map fε is still C-Lipschitz on Υ, for all small enough ε > 0; (ii) the map fε extends to a C-Lipschitz, (j, ρ)-equivariant map f on U , that agrees with f0 on ∂U , for all small enough ε > 0. This will prove that the leaf ℓ of E containing a lift p of ṗ did not have to be maximally stretched after all, a contradiction: indeed, consider q ∈ ℓ far enough from p, at distance < ε/4 from some point q ′ ∈ Υ, such that Φ−ε (q ′ ) is still within < ε/4 from the point of ℓ at distance ε from q (see Figure 11). Then d(p, Φ−ε (q ′ )) ≤ d(p, q) − ε/2, which implies d(fε (p), fε (q)) ≤ d(fε (p), fε (q ′ )) + d(fε (q ′ ), fε (q))  ≤ C d(p, Φ−ε (q ′ )) + ε/4 ≤ C (d(p, q) − ε/4) < C d(p, q). 84 FRANÇOIS GUÉRITAUD AND FANNY KASSEL p ∂U ℓ Υ ′ q Φ−ε (q ) q′ Figure 11. Flowing back by Φ−ε brings q ′ closer to p. • Proof of (ii) assuming (i). By Remark 2.8.(2), it is sufficient to consider one connected component A of U r Υ in H2 and, assuming (i), to prove that for any small enough ε > 0 the map fε extends to a CLipschitz, (j, ρ)-equivariant map f on A, that agrees with f0 on ∂A. Fix such a connected component A; its image in S is an annulus. By Theorem 1.6, it is sufficient to prove that d(fε (x), f0 (y)) ≤ Cd(x, y) for any geodesic segment [x, y] across A with x ∈ Υ and y ∈ ∂A. Note that the length of such segments is uniformly bounded from below, by d(Υ, ∂A). ∂U y A Eδ ℓ′ Υ x Figure 12. A segment [x, y] across the lifted annulus A. For any 0 < δ < d(Υ, ∂A), let Eδ be the δ-neighborhood of the lamination E in the lifted annulus A (see Figure 12). By Lemma 2.9, since f0 is optimal, (9.1) sup Lipx (f0 ) < C. x∈ArEδ If no leaf of E entering Υ meets A, then all geodesic segments [x, y] as above spend a definite amount of length (at least d(Υ̇, ∂A) − δ) in A r Eδ , and so (9.1) implies d(f0 (x), f0 (y)) ≤ C (d(x, y) − ε0 ) for some ε0 > 0 independent of [x, y]. Therefore d(fε (x), f0 (y)) ≤ d(fε (x), f0 (x)) + d(f0 (x), f0 (y)) ≤ C d(x, y) MAXIMALLY STRETCHED LAMINATIONS 85 for all 0 < ε < ε0 . Now suppose that there are leaves of E entering Υ that meet A. The collection of such leaves is finite modulo the stabilizer of A. There exists δ > 0 such that if [x, y] is contained in the δ-neighborhood of some leaf ℓ′ of E entering A, then the function t 7→ d(Φ−t (x), y) is decreasing for t ∈ [0, 1], because the direction of the flow Φ at x is essentially the same as the direction of ℓ′ ; in particular, d(fε (x), f0 (y)) ≤ C d(Φ−ε (x), y) ≤ C d(x, y) for all ε ∈ (0, 1]. There also exists δ′ ∈ (0, δ) such that if a geodesic segment [x, y] as above is not contained in the δ-neighborhood of one of the finitely many leaves entering Υ, then it meets A r Eδ′ ; in particular, it spends a definite amount of length (at least δ′ /2) in A r Eδ′ /2 , and we conclude as above, using (9.1) with δ′ /2 instead of δ. • Proof of (i). By Remark 2.8.(2), it is enough to consider one connected component A of U rΥ in H2 and prove that d(fε (x), fε (y)) ≤ Cd(x, y) for all x, y ∈ Υ∩∂A. If Υ∩∂A is a geodesic line (corresponding to a closed geodesic of Υ̇), then (Φ−ε )|Υ∩∂A is an isometry and so LipΥ∩∂A (fε ) ≤ LipΥ∩∂A (f ) = C (in fact Υ ∩ ∂A is C-stretched by fε ). Otherwise, Υ ∩ ∂A is a countable union of geodesic lines Di , i ∈ Z, with Di and Di+1 asymptotic to each other, both oriented in the direction of the ideal spike they bound if i is odd, and both oriented in the reverse direction if i is even; the leaves of E entering Υ do so in the spikes. Suppose by contradiction that there is a sequence (εk ) ∈ (R∗+ )N tending to 0 and, for every k ∈ N, a pair (xk , yk ) of points of Υ ∪ ∂A such that (9.2) d(fεk (xk ), fεk (yk )) > C · d(xk , yk ). Note that the Hausdorff distance from [xk , yk ] to the nearest leaf segment of Υ tends to zero as k → +∞. Indeed, as above, for any δ > 0, if a geodesic segment [x, y] is not contained in the δ-neighborhood Eδ of E in A, then it spends a definite amount of length (at least δ/2) in A r Eδ/2 , and (9.1) with δ/2 instead of δ forces d(fε (x), fε (y)) ≤ Cd(x, y) for small enough ε. This proves that the Hausdorff distance from [xk , yk ] to the nearest segment of E tends to zero as k → +∞, and this segment actually lies in Υ because xk and yk both belong to Υ and there are locally only finitely many leaves of E entering Υ. Up to replacing xk and yk by j(Γ0 )-translates and passing to a subsequence, we can in fact suppose that there exists i ∈ Z such that both d(xk , Di ) and d(yk , Di ) tend to zero as k → +∞; indeed, the set of lines Di is finite modulo the stabilizer of A. Up to switching xk and yk and passing to a subsequence, we can suppose that either (xk , yk ) ∈ Di × Di+1 for all k, or (xk , yk ) ∈ Di−1 × Di+1 for all k; the case (xk , yk ) ∈ Di × Di is excluded by the assumption (9.2) since Di is C-stretched under fεk . Let yk′ be the point of Di on the same horocycle as yk in the ideal spike of A bounded by Di and Di+1 , and let ηk ≥ 0 be the length of the piece of horocycle from yk to yk′ . If xk ∈ Di−1 , define similarly an arc of horocycle from xk to x′k ∈ Di , of length ξk ; otherwise, set (x′k , ξk ) = (xk , 0). Since d(xk , Di ) and d(yk , Di ) tend to zero as k → +∞, so do ξk and ηk . We claim that, up to passing to a subsequence and replacing xk and yk by other points on the same leaves, still subject to (9.2), we can assume 86 FRANÇOIS GUÉRITAUD AND FANNY KASSEL yk ηk yk′ xk x∗ Di Di+1 yk xk Di−1 ηk ξk x′k Di Di+1 yk′ Figure 13. Distance estimates between points of Di−1 , Di , Di+1 . that d(xk , yk ) → +∞ as k → +∞. Indeed, this is already the case if (xk , yk ) ∈ Di−1 ×Di+1 for all k, because ξk , ηk → 0. If (xk , yk ) ∈ Di ×Di+1 for all k, note that the C-Lipschitz map f stretches Di and Di+1 maximally and sends them to two geodesic lines of H2 , necessarily asymptotic. Moreover, f (yk ) and f (yk′ ) lie at the same depth in the spike bounded by f (Di ) and f (Di+1 ): indeed, the distance between the horocycles through f (yk ) and through f (yk′ ) is independent of k; if it were nonzero, then for large enough k we would obtain a contradiction with the fact that f is Lipschitz (recall ηk → 0). Using (A.4), we see that there exists Q ≥ 0 such that for any integer k, the piece of horocycle connecting f (yk ) to f (yk′ ) has length QηkC , and the piece of horocycle connecting fεk (yk ) to fεk (yk′ ) has length Qe±εk ηkC , which is ≪ ηk since C > 1. In particular, xk 6= yk′ for all large enough k (since xk and yk satisfy (9.2)); in other words, xk and yk lie at distinct depths inside the spike of A bounded by Di and Di+1 (see Figure 13, top). If yk lies deeper than xk (which we can assume by symmetry), then for any x∗ ∈ Di less deep than xk , π ∗ x y < f (x∗ )f\ < x\ εk εk (xk )fεk (yk ) < π. k k 2 Moreover, d(fεk (x∗ ), fεk (xk )) = Cd(x∗ , xk ) and (9.2) holds, hence d(fεk (x∗ ), fεk (yk )) > Cd(x∗ , yk ) by Toponogov’s theorem [BH, Lem. II.1.13]. Thus, up to replacing xk by some fixed x∗ , we may assume that d(xk , yk ) → +∞. Using (A.6), we see that d(xk , yk ) = d(x′k , yk′ ) + (ξk2 + ηk2 )(1 + o(1)) (see Figure 13, bottom). Similarly, given that the length of the piece of horocycle from f (xk ) to f (x′k ) in the spike bounded by Di−1 and Di is QξkC for some Q ≥ 0 independent of k (see above) and that the length of the piece of horocycle from f (yk ) to f (yk′ ) in the spike bounded by Di and Di+1 is MAXIMALLY STRETCHED LAMINATIONS Q′ ηkC for some Q′ ≥ 0 independent of k, we obtain 87 2 d(fεk (xk ), fεk (yk )) ≤ d(fεk (x′k ), fεk (yk′ )) + (Q2 ξk2C + Q′ ηk2C )(1 + o(1)). Since d(fεk (x′k ), fεk (yk′ )) = Cd(x′k , yk′ ) and since ξkC = o(ξk ) and ηkC = o(ηk ), we find that d(fεk (xk ), fεk (yk )) ≤ Cd(xk , yk ) for all large enough k, contradicting (9.2). This completes the proof of (i).  9.3. Semicontinuity for C(j, ρ) > 1. The notion of chain recurrence (Definition 9.1) is closed for the Hausdorff topology: any compactly-supported lamination which is a Hausdorff limit of chain recurrent laminations is chain recurrent [T2, Prop. 6.1]. It is therefore relevant to consider (semi)continuity issues. In the classical setting, Thurston [T2, Th. 8.4] proved that his maximal ratio-maximizing chain recurrent lamination µ(j, ρ) varies in an upper semicontinuous way as j and ρ vary over the Teichmüller space T (S) of S. In other words, by Lemma 9.3, the stretch locus E(j, ρ) varies in an upper semicontinuous way over T (S): for any sequence (jk , ρk )k∈N of elements of T (S)2 converging to (j, ρ), E(j, ρ) ⊃ lim sup E(jk , ρk ), k→+∞ where the limsup is defined with respect to the Hausdorff topology. We now work in a more general setting and show how the chain recurrence of the stretch locus E(j, ρ) (Proposition 9.4) implies upper semicontinuity. Proposition 9.5. In dimension n = 2, the stretch locus E(j, ρ) is upper semicontinuous on the open subset of Homj0 (Γ0 , G) × Hom(Γ0 , G)red where C(j, ρ) > 1. Here we denote by Homj0 (Γ0 , G) the space of geometrically finite representations of Γ0 in G with the same cusp type as the fixed representation j0 (as in Section 6) and by Hom(Γ0 , G)red the space of reductive representations ρ ∈ Hom(Γ0 , G) (Definition 4.9). These two sets are endowed with the induced topology from Hom(Γ0 , G). The condition C(j, ρ) > 1 is open by Proposition 6.1.(2). Proof of Proposition 9.5. By Lemma 4.4, we may assume that Γ0 is torsionfree. Let (jk , ρk )k∈N be a sequence of elements of Homj0 (Γ0 , G)×Hom(Γ0 , G)red with C(jk , ρk ) > 1 converging to some (j, ρ) ∈ Homj0 (Γ0 , G)×Hom(Γ0 , G)red with C(j, ρ) > 1. Since the stretch locus E(j, ρ) is empty with our definitions when C(j, ρ) = +∞, and the condition C(j, ρ) = +∞ is open by Lemma 4.7, we may assume that the C(jk , ρk ) and C(j, ρ) are all finite. Recall from Section 6.4 the proof of the fact (labelled (B) there) that lim sup C(jk , ρk ), if greater than 1, gives a lower bound for C(j, ρ). By the same argument as in that proof, up to passing to a subsequence, the stretch loci E(jk , ρk ) are jk (Γ0 )-invariant geodesic laminations that converge to some j(Γ0 )-invariant geodesic lamination L , compact in j(Γ0 )\Hn . Moreover, the image of L in j(Γ0 )\Hn nearly carries simple closed curves corresponding to elements γ ∈ Γ0 with λ(ρ(γ))/λ(j(γ)) arbitrarily close to C(j, ρ). However, this does not immediately imply that L is contained in E(j, ρ): we need to improve the “multiplicative error” to an “additive error”. The idea is similar to Lemmas 5.10 and 9.3, but with varying j, ρ. 88 FRANÇOIS GUÉRITAUD AND FANNY KASSEL Suppose by contradiction that L contains a point p ∈ / E(j, ρ). According j,ρ to Lemma 4.15, there is an element f ∈ F that is constant on some ball centered at p, with radius δ > 0. Since the E(jk , ρk ) are chain recurrent (Proposition 9.4), so is L . Let G be a simple closed geodesic in j(Γ0 )\H2 δ passing within δ/2 of p and approached by a 16mC -quasi-leaf of L (in the sense of Section 9.1), made of at most m leaf segments of L , where m is the integer (depending only on the topology of S = j(Γ0 )\H2 ) defined after Fact 9.2. Let γ ∈ Γ0 correspond to G. By Hausdorff convergence, for large enough k, the geodesic representative of G in jk (Γ0 )\H2 is approached by a δ 8mC -quasi-leaf of E(jk , ρk ), made of at most m leaf segments. Since E(jk , ρk ) is maximally stretched by a factor Ck := C(jk , ρk ) by any element of F jk ,ρk , it follows, as in the proof of Lemma 9.3, that δ . λ(ρk (γ)) − Ck λ(jk (γ)) ≤ 4mCk · 8mC Since Ck tends to C by Proposition 6.1.(2)–(3) and since λ is continuous, the left-hand side converges to |λ(ρ(γ)) − C λ(j(γ))| as k → +∞, while the right-hand side converges to δ/2. However, this left-hand limit is ≥ Cδ since f is constant on the ball of radius δ centered at p, which contains a segment of G of length δ. This is absurd, hence L ⊂ E(j, ρ).  9.4. The stretch locus for C(j, ρ) < 1. Still in dimension n = 2, let Γ0 , (j, ρ), K ⊂ Hn compact, and ϕ : K → Hn be as in Section 4 (with K possibly empty). The relative stretch locus EK,ϕ (j, ρ) behaves very differently depending on whether CK,ϕ (j, ρ) is smaller than, equal to, or larger than 1. Let us give a simple example to illustrate the contrast. Example 9.6. We take Γ0 to be trivial. Fix o ∈ H2 and let (as )s≥0 , (bs )s≥0 , and (cs )s≥0 be three geodesic rays issued from o, parametrized at unit speed, forming angles of 2π/3 at o. Let K = {at , bt , ct } for some t > 0 and let ϕ : K → H2 be given by ϕ(at ) = aT , ϕ(bt ) = bT , and ϕ(ct ) = cT for some T > 0. Then Lip(ϕ) = g(T )/g(t), where g : R+ → R+ is given by g(s) = d(as , bs ). By convexity√of the distance function, the function g is strictly convex, asymptotic pto 3s for s → 0 and to 2s for s → +∞. (Explicitly, g(s) = 2 arcsinh( 3/4 sinh s) by (A.14).) • If t < T , then Lip(ϕ) > T /t > 1 by strict convexity of g. By Theorem 5.1, the map ϕ extends to H2 with the same Lipschitz constant and with stretch locus the perimeter of the triangle at bt ct ; this stretch locus is the smallest possible by Remark 2.8.(1). • If t = T , then ϕ is 1-Lipschitz and has a unique 1-Lipschitz extension to the (filled) triangle at bt ct , namely the identity map. An optimal extension to H2 is obtained by precomposing with the closest-point projection onto the triangle at bt ct ; the stretch locus is this triangle. • If t > T , then Lip(ϕ) < T /t < 1 by strict convexity of g. However, the optimal Lipschitz constant of an extension of ϕ to H2 cannot be less than T /t: indeed, such an extension may be assumed to fix o by T) symmetry, and d(o,a d(o,at ) = T /t. It follows from the construction used in Section 10.4 below that a (T /t)-Lipschitz extension of ϕ to H2 MAXIMALLY STRETCHED LAMINATIONS 89 does indeed exist, and the stretch locus is equal to the union of the geodesic segments [o, at ], [o, bt ], [o, ct ]. Although the stretch locus may vary abruptly in the above, note that this variation is upper semicontinuous in (t, T ) for the Hausdorff topology, in agreement with a potential generalization of Proposition 9.5 to C ≤ 1. We now consider the case when K is empty. Here is some evidence in favor of Conjecture 1.4, which claims that for C(j, ρ) < 1 the stretch locus E(j, ρ) should be what we call a gramination: • In Section 10.4, we give a construction, for certain Coxeter groups Γ0 , of pairs (j, ρ) with j convex cocompact, j(Γ0 )\H2 compact, and C(j, ρ) < 1, for which the stretch locus E(j, ρ) is a trivalent graph. • Consider the examples constructed in [Sa, § 4.4]: for any compact hyperbolic surface S of genus g and any integer k with |k| ≤ 2g − 2, Salein constructed highly symmetric pairs (j, ρ) ∈ Hom(π1 (S), G)2 with j Fuchsian such that ρ has Euler number k; a construction similar to Section 10.4 shows that the stretch locus of such a pair (j, ρ) is a regular graph of degree 4g. • In Section 10.5, we give a construction of pairs (j, ρ) with j convex cocompact, j(Γ0 )\H2 noncompact, and C(j, ρ) < 1, for which the stretch locus E(j, ρ) is a trivalent graph. It is actually possible to generalize this construction and obtain, for any given convex cocompact hyperbolic surface S of infinite volume and any given trivalent graph G retract of S, an open set of pairs (j, ρ) ∈ Hom(Γ0 , G)2 with j convex cocompact for which the stretch locus E(j, ρ) is a trivalent graph of H2 , with geodesic edges, whose projection to j(Γ0 )\H2 is a graph isotopic to G (see Remarks 10.3). • It is also possible to construct examples of pairs (j, ρ) ∈ Hom(Γ0 , G)2 with j geometrically finite and C(j, ρ) < 1 for which the stretch locus E(j, ρ) is a geodesic lamination: see Sections 10.2 and 10.3 for instance. Here is perhaps a first step towards proving Conjecture 1.4. Lemma 9.7. In dimension n = 2, let (j, ρ) ∈ Hom(Γ0 , G)2 be a pair of representations with j geometrically finite and F j,ρ 6= ∅. Each connected component of the complement of the stretch locus E(j, ρ) ⊂ H2 is convex. Proof. Suppose by contradiction that the open set H2 r E(j, ρ) has a nonconvex component U . There exists a smooth arc A0 ⊂ H2 r E(j, ρ) whose endpoints x0 , y0 are connected by a segment intersecting E(j, ρ). We can assume that A0 ∪ [x0 , y0 ] is a Jordan curve, and is arbitrarily small: this can be seen by moving the segment [x0 , y0 ] until it stops intersecting E(j, ρ), and then doing a small perturbative argument. In particular, we may assume that A0 ∪ [x0 , y0 ] embeds into j(Γ0 )\H2 under the quotient map. We can perturb the Jordan curve A0 ∪ [x0 , y0 ] to a Jordan curve A ∪ [x, y] whose inner (open) disk D intersects E(j, ρ) in some point z, with A ⊂ U (see Figure 14). Let B ⊂ U be a compact neighborhood of A. Choose an optimal, CLipschitz equivariant map f (Lemma 4.13): by construction, LipB (f ) =: 90 FRANÇOIS GUÉRITAUD AND FANNY KASSEL A B′ ′ A D′ V x z V′ y E(j, ρ) Figure 14. Improving the local Lipschitz constant at z. C ∗ < C. We define a (j, ρ)-equivariant map gε := f ◦ Jε : H2 → H2 , where Jε is the small deformation of the identity map idH2 given as follows: • on H2 r j(Γ0 ) · D, take Jε to be the identity map; • on D ′ := D r B, take Jε := ε · π[x,y] + (1 − ε) · idD′ where π[x,y] denotes the closest-point projection onto [x, y]; extend (j, ρ)-equivariantly to j(Γ0 ) · D ′ ; • on B ′ := D ∩ B, note that Jε is already defined on ∂B ′ and use Proposition 3.1 to find an optimal extension to B ′ ; extend (j, ρ)equivariantly to j(Γ0 ) · B ′ . We have LipD′ (Jε ) ≤ 1 because Lip(π[x,y] ) ≤ 1 (use Lemma 2.13). Also, we claim that Lip∂B ′ (Jε ) ≤ C/C ∗ for ε small enough. This is true because ∂B ′ is the union of two subsegments V, V ′ of [x, y] and two disjoint arcs, namely A and another, nearly parallel arc A′ : the only pairs of points (ξ, ξ ′ ) ∈ (∂B ′ )2 that Jε |∂B ′ can move apart are in A × A′ (up to order), but d(ξ, ξ ′ ) is then bounded from below by a positive constant d(A, A′ ). Thus, Lip∂B ′ (Jε ) (and hence Lip(Jε )) goes to 1 as ε goes to 0, which yields Lip(gε ) ≤ C as soon as Lip(Jε ) ≤ C/C ∗ . However, since π[x,y] is contracting near z ∈ D ′ , we have Lipz (gε ) < C, hence z ∈ / Egε . This contradicts the optimality of f , as z ∈ E(j, ρ).  10. Examples and counterexamples All examples below are in dimension n = 2, except the last two. For n = 2, we use the upper half-plane model of H2 and identify G = PO(2, 1) with PGL2 (R) and its identity component G0 with PSL2 (R). Example 10.1 deals with infinitely generated Γ0 . Examples 10.2 to 10.5 concern convex cocompact j, while Examples 10.6 to 10.11 illustrate phenomena that arise only in the presence of cusps. 10.1. A left admissible pair (j, ρ) with C(j, ρ) = 1, for infinitely generated Γ0 . In this section, we give an example of an infinitely generated discrete subgroup Γ of G × G that acts properly discontinuously on G but that does not satisfy the conclusion of Theorem 7.1; in other words, Γ is not sharp in the sense of [KK, Def. 4.2]. MAXIMALLY STRETCHED LAMINATIONS 91 In the upper half-plane model of H2 , let Ak (resp. Bk ) be the half-circle of radius 1 (resp. log k) centered at k2 , oriented clockwise. Let A′k (resp. Bk′ ) be the half-circle of radius 1 (resp. log k) centered at k2 + k, oriented counterclockwise (see Figure 15). Let αk ∈ G0 (resp. βk ∈ G0 ) be the shortest hyperbolic translation identifying the geodesic represented by Ak with A′k (resp. Bk with Bk′ ); its axis is orthogonal to Ak and A′k (resp. to Bk and Bk′ ), hence its translation length λ(αk ) (resp. λ(βk )) is equal to the distance between Ak and A′k (resp. Bk and Bk′ ). An elementary computation (see (A.12) below) shows that (10.1) λ(αk ) = 2 arccosh(k/2) = 2 log k + o(1), k ) = 2 log k − 2 log log k + o(1) . λ(βk ) = 2 arccosh( 2 log k Consider the free group ΓN = hγk ik≥N and its injective and discrete representations j, ρ given by j(γk ) = αk and ρ(γk ) = βk . Since λ(βk )/λ(αk ) goes to 1, we have C(j, ρ) ≥ 1 and C(ρ, j) ≥ 1. However, we claim that for N large enough, the group Γj,ρ N = {(j(γ), ρ(γ)) | γ ∈ ΓN } is left admissible (Definition 1.7), acting properly discontinuously on G. √ Indeed, fix the basepoint p0 = −1 ∈ H2 and consider a reduced word m γ = γkε11 . . . γkεm ∈ ΓN , where εi = ±1. Let DA ⊂ H2 be the fundamental 2 domain of H for the action of j(ΓN ) that is bounded by the geodesics Ak , A′k for k ≥ N . Let DB be the fundamental domain for the action of ρ(ΓN ) that is bounded by the geodesics Bk , Bk′ for k ≥ N . The geodesic segment from p0 to j(γ) · p0 projects in the fundamental domain DA to a union of m + 1 geodesic segments I0 , . . . , Im : namely, Ii connects the half-circle Aki or A′ki (depending on εi ) to the half-circle Aki+1 or A′ki+1 (depending on εi+1 ), unless i = 0 or m, in which case one of the endpoints is p0 (see Figure 15). Moreover, the geodesic line carrying Ii hits ∂∞ H2 near the centers of these half-circles, since all half-circles Ak , A′k are far from one another and from p0 (compared to their radii). Therefore, the ends of Ii are nearly orthogonal to the Ak , A′k and the length of Ii can be approximated by the distance from some side of DA to another (or to p0 ). The error is o(1) for each segment Ii , uniformly as N → +∞. I0 √ p0 = −1 I1 Bn1 An1 Bn′ 1 A′n1 Bn2 An2 Figure 15. For infinitely generated Γ0 , construction of an admissible pair (j, ρ) with C(j, ρ) = 1. The distance from p0 to ρ(γ) · p0 is likewise a sum of lengths of segments J0 , . . . , Jm between boundary components of DB : the segment Ji meets Bki (resp. Bk′ i , Bki+1 , Bk′ i+1 ) exactly when Ii meets Aki (resp. A′ki , Aki+1 , A′ki+1 ). Therefore, Ii is longer than Ji by roughly the sum of d(Aki , Bki ) = d(A′ki , Bk′ i ) and d(Aki+1 , Bki+1 ) = d(A′ki+1 , Bk′ i+1 ). Using (10.1), we obtain length(Ii ) − length(Ji ) = log log ki + log log ki+1 + o(1) 92 FRANÇOIS GUÉRITAUD AND FANNY KASSEL (with one term stricken out for i = 0 or m), with uniform error as N → +∞. In particular, for N large enough the left-hand side is always ≥ 1. Finally, µ(j(γ)) − µ(ρ(γ)) = d(p0 , j(γ) · p0 ) − d(p0 , ρ(γ) · p0 ) m X  length(Ii ) − length(Ji ) = i=0 o n  ≥ max m, log log max ki , 1≤i≤m γkε11 m . . . γkεm which clearly diverges to +∞ as γ = exhausts the countable j,ρ group ΓN . Therefore the group ΓN acts properly discontinuously on G by the properness criterion of Benoist and Kobayashi (Section 7.3). 10.2. A nonreductive ρ with F j,ρ 6= ∅. Let Γ0 be a free group on two generators α, β and let j ∈ Hom(Γ0 , G) be the holonomy representation of a hyperbolic one-holed torus S of infinite volume, such that the translation axes Aj(α) and Aj(β) of α and β meet at a right angle at a point p ∈ H2 (see Figure 16). H2 Aj(α) U t f (s′ ) s′ Aj(β) p D s U t ′ f f (p) f (t) f (t′ ) f (s) Figure 16. A nonreductive representation ρ such that the stretch locus E(j, ρ) is the j(Γ0 )-orbit of the axis Aj(α) . We first consider the representation ρ0 ∈ Hom(Γ0 , G) given by ρ0 (α) = j(α)2 and ρ0 (β) = 1. It is reductive with two fixed points in ∂∞ H2 . We claim that C(j, ρ0 ) = 2 = λ(ρ0 (α))/λ(j(α)) and that the image of the stretch locus E(j, ρ) in j(Γ0 )\H2 is the closed geodesic corresponding to α. Indeed, consider the Dirichlet fundamental domain D of the convex core centered at p, for the action of j(Γ0 ). It is bounded by four segments of the boundary of the convex core, and by four other segments s, s′ , t, t′ such that j(α) maps s to s′ and j(β) maps t to t′ . Let πAj(α) be the closest-point projection onto Aj(α) and h the orientation-preserving homeomorphism of Aj(α) such that d(p, h(q)) = 2 d(p, q) for all q ∈ Aj(α) . The map h ◦ πAj(α) : D → H2 is 2-Lipschitz and extends to a 2-Lipschitz, (j, ρ0 )-equivariant map f0 : H2 → Aj(α) whose stretch locus is exactly j(Γ0 ) · Aj(α) . Consider a small, nonreductive deformation ρ ∈ Hom(Γ0 , G) of ρ0 such that ρ(α) = ρ0 (α) = j(α)2 and such that ρ(β) has a parabolic fixed point in MAXIMALLY STRETCHED LAMINATIONS 93 ∂∞ H2 common with j(α). Then C(j, ρ) = C(j, ρ0 ) = 2 by Lemma 6.3. We claim that F j,ρ is nonempty if ρ(β) is close enough to IdH2 . Indeed, let us construct a (j, ρ)-equivariant deformation f of f0 which is still 2-Lipschitz. By Lemma 2.9, we have LipU (f0 ) < C for some neighborhood U of t ∪ t′ . Therefore, the map f defined on s ∪ s′ ∪ t ∪ t′ by f |t∪s∪s′ = f0 |t∪s∪s′ and f |t′ = ρ(β) ◦ f0 |t′ is still 2-Lipschitz if ρ(β) is close enough to IdH2 . This map f extends, with the same Lipschitz constant 2, to all of D (by the Kirszbraun–Valentine theorem, Proposition 3.1), hence (j, ρ)-equivariantly to H2 . This construction can be adapted to any hyperbolic surface S of infinite volume when the stretch locus E(j, ρ0 ) is a multicurve. 10.3. A nonreductive ρ with F j,ρ = ∅. Let again Γ0 be a free group on two generators α, β and let j ∈ Hom(Γ0 , G) be the holonomy representation of a hyperbolic one-holed torus S of infinite volume. Let L be the preimage in H2 of an oriented irrational measured lamination of S. We first construct a reductive representation ρ0 ∈ Hom(Γ0 , G) with two fixed points in ∂∞ H2 such that E(j, ρ0 ) = L and C(j, ρ0 ) < 1. It is sufficient to construct a differential 1-form ω of class L∞ on S with the following properties: (1) ω R is locally the differential of some 1-Lipschitz function ϕ, (2) I ω = length(I) for any segment of leaf I of (the image in S of) L , oriented positively. Indeed, if such an ω exists, then for any geodesic line A of H2 , any isometric identification A ≃ R, and any C ∈ (0, 1), we can define a representation ρ0 ∈ Hom(Γ0 , G) as follows: if γ ∈ Γ0 r {1} corresponds to a loop Gγ on S, then R ρ0 (γ) is the hyperbolic element of G translating along A with length C Gγ ω ∈ R. Such a representation ρ0 satisfies E(j, ρ0 ) = L and C(j, ρ0 ) = C because for any basepoint p ∈ H2 , the map Z 2 ω e ∈R≃A f0 : q ∈ H 7−→ C [p,q] (where ω e is the j(Γ0 )-invariant 1-form on H2 lifting ω) is (j, ρ0 )-equivariant, has Lipschitz constant exactly C, and stretches L maximally, and we can use Lemma 5.10. Let us construct a 1-form ω as above. The idea is similar to the “stretch maps” of [T2]. The complement of the image of L in the convex core of S is a one-holed biinfinite bigon B; each of its two spikes can be foliated by pieces of horocycles (see Figure 17). Let H and H ′ be horoball neighborhoods of the two spikes, tangent in two points of L (one for each side of B). We take ω = dϕ where  on B r (H ∪ H ′ ),  0 d(·, ∂H) on H, ϕ :=  −d(·, ∂H ′ ) on H ′ . This form ω extends continuously to all of L . Let now ρ ∈ Hom(Γ0 , G) be a nonreductive representation such that the fixed point ξ of ρ(Γ0 ) in ∂∞ H2 is one of the two fixed points of ρ0 (Γ0 ). Then C(j, ρ) = C := C(j, ρ0 ) by Lemma 6.3. We claim that F j,ρ = ∅, i.e. there 94 FRANÇOIS GUÉRITAUD AND FANNY KASSEL L {ξ} B f H L ∂ H′ ∝ e−t t J Figure 17. Left: the one-holed bigon B bounded by the irrational lamination L . The symbol ∂ denotes the boundary of the convex core. The function ϕ is constant on the shaded area; elsewhere its level curves are pieces of horocycles. Right : the Lipschitz map f must collapse all lines of L if C < 1, because e−Ct ≫ e−t for large t. exists no C-Lipschitz, (j, ρ)-equivariant map f : H2 → H2 . Indeed, suppose by contradiction that such an f exists. We first note that f stretches maximally every leaf of L . Indeed, the “horocyclic projection” taking any p ∈ H2 to the intersection of the translation axis A of ρ0 with the horocycle through p centered at ξ is 1-Lipschitz. After postcomposing f with this horocyclic projection, we obtain a C-Lipschitz map f1 which is (j, ρ0 )-equivariant, hence has to stretch maximally every leaf of E(j, ρ0 ) = L (Theorem 1.3). Then f also stretches maximally every leaf of L . In fact, this argument shows that on any leaf of L , the map f coincides with f1 postcomposed with some parabolic (or trivial) isometry of H2 fixing ξ, depending on the leaf; the leaf endpoint in ∂∞ H2 which is sent to ξ by f1 is also sent to ξ by f . (Actually, by density of leaves the (j, ρ0 )equivariant restrictions f0 |L and f1 |L differ only by a translation along the axis A of ρ0 , but we will not need this.) Let us now prove that f maps all the leaves of L to the same geodesic line of H2 . This will provide a contradiction since f (L ) is ρ(Γ0 )-invariant and ρ(Γ0 ) has only one fixed point in ∂∞ H2 (namely ξ). Let J be a short geodesic segment of H2 transverse to the lamination L , such that J r L is the union of countably many open subintervals Jk . Then each Jk intercepts an ideal sector bounded by two leaves of L that are asymptotic to each other on one of the two sides, left or right, of J. Orient J so that all the halfleaves on the left of J are mapped under f1 to geodesic rays with endpoint ξ. Then any two leaves asymptotic on the right of J have the same image under f : indeed, the right parts of the image leaves are asymptotic because f is Lipschitz, and the left parts are asymptotic because ρ sends all the left endpoints to ξ. Consider two leaves ℓ, ℓ′ of L that are asymptotic on the left of J, bounding together an infinite spike. At depth t ≫ 1 inside the spike, ℓ and ℓ′ approach each other at rate e−t (see (A.5)), and their images under f , if distinct, approach each other at the slower exponential rate e−Ct (recall that C < 1); since f is Lipschitz, this forces f (ℓ) = f (ℓ′ ). Therefore, all the MAXIMALLY STRETCHED LAMINATIONS 95 sectors intercepted by the Jk are collapsed by f . Since by [BiS] the length of J is the sum of the lengths of the Jk , passing to the limit we see that all the leaves of L meeting J have the same image under f . We conclude by observing that the projection of J ∪ L to S carries the full fundamental group of S. This proves that F j,ρ = ∅. It is not clear whether the same can happen when C(j, ρ) ≥ 1, but the natural conjecture would be that it does not. 10.4. A pair (j, ρ) with C ′ (j, ρ) < C(j, ρ) < 1 and j(Γ0 )\Hn compact. While the constants C(j, ρ) and C ′ (j, ρ) are equal above 1 (Corollary 1.12), they can differ below 1. To prove this, we only need to exhibit a pair (j, ρ) such that any closed geodesic of j(Γ0 )\H2 spends a definite (nonzero) proportion of its length in a compact set V disjoint from the stretch locus (on V the local Lipschitz constant of an optimal Lipschitz equivariant map stays bounded away from C(j, ρ), see Lemma 2.9). In H2 , consider a positively oriented hyperbolic triangle ABC with angles b = π, C b= π b = π, B A 3 2 14 ′ ′ ′ and another, smaller triangle A B C with the same orientation and with angles c′ = π , B c′ = π , C c′ = π . A 3 2 7 The edge [A′ ,B ′ ] is shorter than [A,B]; let ϕ : [A,B] → [A′ ,B ′ ] be the uniform parametrization, with ϕ(A) = A′ and ϕ(B) = B ′ , so that C0 := Lip(ϕ) = d(A′ , B ′ ) < 1. d(A, B) Claim 10.1. The map ϕ admits a C0 -Lipschitz extension f to the filled triangle ABC, taking the geodesic segment [A,C] (resp. [C,B]) to the geodesic segment [A′ ,C ′ ] (resp. [C ′ ,B ′ ]), and with stretch locus the segment [A,B]. Proof. Let ℓ be the geodesic line of H2 containing [A,B], oriented from A to B. Any point p ∈ H2 may be reached in a unique way from A by first applying a translation of length v(p) ∈ R along the geodesic line orthogonal to ℓ at A, positively oriented with respect to ℓ (“vertical direction”), then a translation of length h(p) ∈ R along ℓ itself (“horizontal direction”). The real numbers h(p) and v(p) are called the Fermi coordinates of p with respect to (ℓ, A). Similarly, let h′ and v ′ be the Fermi coordinates with respect to (ℓ′ , A′ ), where ℓ′ is the geodesic line containing [A′ ,B ′ ], oriented from A′ to B ′ (see Figure 18). Let Ψ : R+ → R+ be a diffeomorphism whose derivative Ψ′ is everywhere < C0 on R∗+ and let f ∗ : H2 → H2 be given, in Fermi coordinates, by h′ (f ∗ (p)) = C0 h(p) and v ′ (f ∗ (p)) = Ψ(v(p)). Then Lipp (f ∗ ) < C0 for all p ∈ / ℓ. Indeed, the differential of f ∗ at p ∈ / ℓ ′ has principal value Ψ (v(p)) < C0 in the vertical direction and, by (A.9), Ψ(v(p)) principal value C0 cosh cosh v(p) < C0 in the horizontal direction. We shall take f := πA′B′C ′ ◦ f ∗ for a suitable choice of Ψ, where πA′B′C ′ is the closest-point projection onto the filled triangle A′ B ′ C ′ . Since we wish f to map [A,C] to [A′ ,C ′ ], we need to choose Ψ so that for any p ∈ [A,C] the 96 FRANÇOIS GUÉRITAUD AND FANNY KASSEL C f ∗ (C) C′ f∗ p v(p) A h(p) B A′ B′ Figure 18. Defining a contracting map between rightangled triangles. point f ∗ (p) lies above (or on) the edge [A′ ,C ′ ]. By (A.13), [= tan pAB tanh v(p) sinh h(p) and tanh Ψ(v(p)) tanh v ′ (f ∗ (p)) = . sinh h′ (f ∗ (p)) sinh(C0 v(p)) Note that tanh(C0 t) > C0 tanh(t) and sinh(C0 t) < C0 sinh(t) for all t > 0, by strict concavity of tanh and convexity of sinh (recall that 0 < C0 < 1). Therefore the function Ψ : t 7→ C0 t yields a map f ∗ with f ∗ ([A,C]) above [A′ ,C ′ ]. We can decrease this function slightly to obtain Ψ with Ψ′ (t) < C0 for all t > 0 while keeping f ∗ ([A,C]) above [A′ ,C ′ ]. In fact, by the above formulas, we can also ensure f ∗ ([A,C]) = [A′ ,C ′ ] directly, by taking Ψ = ΨAb where   tanh v −1 : (10.2) ΨAb(v) = σ (C0 σ(v)) with σ(v) = arcsinh b tan A then Ψ′b < C0 (on R∗+ ) easily follows from C0 < 1 and from the concavity A of σ.  \′ B ′ = tan f ∗ (p)A Let Γ0 be the group generated by the orthogonal reflections in the sides of ABC and j its natural inclusion in G = PGL2 (R). Let ρ ∈ Hom(Γ0 , G) be the representation taking the reflections in [A,B], [B,C], [C,A] to the reflections in [A′ ,B ′ ], [B ′ ,C ′ ], [C ′ ,A′ ] respectively; it is well defined (relations are preserved) because π/7 is a multiple of π/14. The group Γ0 has a finiteindex normal subgroup Γ0 which is torsion-free and such that j(Γ0 ) and ρ(Γ0 ) are orientation-preserving, i.e. with values in G0 = PSL2 (R). Let j, ρ ∈ Hom(Γ0 , G) be the corresponding representations. The map f given by Claim 10.1 extends, by reflections in the sides of ABC, to a C0 -Lipschitz, (j, ρ)-equivariant map on H2 . Its stretch locus is the Γ0 -orbit of the segment [A,B], which is the 1-skeleton of a tiling of H2 by regular 14-gons meeting 3 at each vertex. MAXIMALLY STRETCHED LAMINATIONS 97 We claim that C(j, ρ) = C0 and that f is an optimal element of F j,ρ , in the sense of Definition 4.12. Indeed Lemma 4.4, applied to Γ0 and its finite-index subgroup Γ0 , shows that there exists an element f ∈ F j,ρ which is optimal and (j, ρ)-equivariant. In particular, if p ∈ H2 is fixed by some j(γ) ∈ j(Γ0 ), then f (p) is fixed by ρ(γ). Applying this to the three sides of the triangle ABC, we see that f sends A, B, C to A′ , B ′ , C ′ respectively. In particular, d(A′ , B ′ ) C(j, ρ) ≥ = C0 . d(A, B) Since f is C0 -Lipschitz with stretch locus the Γ0 -orbit of the segment [A,B], this shows that C(j, ρ) = C0 and that E(j, ρ) is the stretch locus of f ; in other words, f is an optimal element of F j,ρ . It is easy to see that no geodesic of H2 can spend more than a bounded proportion of its length near the regular trivalent graph E(j, ρ), which implies that C ′ (j, ρ) < C(j, ρ). 10.5. A pair (j, ρ) with C ′ (j, ρ) < C(j, ρ) < 1 and j(Γ0 )\Hn noncompact. Let Γ0 be a free group on two generators and j ∈ Hom(Γ0 , G) the holonomy representation of a hyperbolic three-holed sphere S with three funnels. Let G be a geodesic trivalent graph on S, with two vertices v, w and three edges, such that the natural symmetry of S switches v and w and preserves each edge. Let ℓ1 , ℓ2 , ℓ3 > 0 be the lengths of the three edges and θ1 , θ2 , θ3 ∈ (0, π) the angles between consecutive edges at both vertices, so that θ1 + θ2 + θ3 = 2π. The preimage of G in H2 is an embedded trivalent tree T . For any C0 ∈ (0, 1), there exists an immersed trivalent tree T ′ with the same combinatorics as T , with all (oriented) angles between adjacent edges of T ′ the same as in T , but with all edges of length ℓi in T replaced by edges of length C0 ℓi in T ′ . The natural map ϕ : T → T ′ , multiplying all lengths along edges by C0 , is (j, ρ)-equivariant for a unique ρ ∈ Hom(Γ0 , G). If C0 < 1 is large enough, then the immersed tree T ′ is in fact embedded, and ρ is convex cocompact. Claim 10.2. Suppose ℓ1 , ℓ2 , ℓ3 are large enough so that the bisecting rays at the vertices of T meet either outside of the preimage N ⊂ H2 of the convex core of S, or not at all. Then the map ϕ extends to a (j, ρ)-equivariant map f : H2 → H2 with Lipschitz constant C0 and stretch locus T . Proof. Let e be an edge of T and e′ , e′′ two of its neighbors, so that e′ , e, e′′ are consecutive edges of some complementary component of T . By symmetry of the pair of pants S, the edge e forms the same angle θi with e′ and with e′′ . Let β ′ and β ′′ be the corresponding bisecting rays, issued from the endpoints of e. Let Q ⊂ H2 be the compact quadrilateral bounded by e, β, β ′ , and a segment of the boundary of N . There is a similarly defined quadrilateral on each side of each edge of T , and their union is N : therefore, it is sufficient to define the map f on Q in a way that is consistent (along the bisecting rays β ′ , β ′′ ) for Q and neighboring quadrilaterals Q′ , Q′′ (see Figure 19). The construction is similar to Claim 10.1, whose notation we borrow: let (h, v) : Q → R×R+ be the Fermi coordinates with respect to the edge e, and (h′ , v ′ ) the Fermi coordinates with respect to ϕ(e). Define f |Q by h′ (f (p)) = C0 h(p) and v ′ (f (p)) = Ψθi /2 (v(p)) for all p ∈ Q, where Ψθi /2 is given by 98 FRANÇOIS GUÉRITAUD AND FANNY KASSEL T e′ Q′ ′ β θi /2 Q′′ ∂N Q β ′′ e′′ e Figure 19. Defining a contracting map on the convex hull N of a tree T , one quadrilateral Q at a time. (10.2). Since the quadrilaterals Q, Q′ , Q′′ have all their angles along T equal to θi /2, the map f just defined takes the bisecting rays β, β ′ to the bisecting rays of the corresponding angles of T ′ , in a well-defined manner. The proof that f is C0 -Lipschitz on N is the same as in Claim 10.1.  We claim that C(j, ρ) = C0 and that f is an optimal element of F j,ρ , in the sense of Definition 4.12. Indeed, since G is invariant under the natural symmetry of S, the group Γ0 is contained, with index two, in a discrete subgroup Γ0 of G = PGL2 (R). Let j ∈ Hom(Γ0 , G) be the natural inclusion and let ρ ∈ Hom(Γ0 , G) be the natural extension of ρ. All reflections in perpendicular bisectors of edges of T (resp. T ′ ) belong to j(Γ0 ) (resp. ρ(Γ0 )). Lemma 4.4, applied to (Γ0 , Γ0 ), shows that there exists an element f ∈ F j,ρ which is optimal and (j, ρ)-equivariant. Let us show that f agrees with ϕ on T . Let v be a vertex of T and let e1 , e2 , e3 be the three incident edges of T , connecting v to v1 , v2 , v3 , with perpendicular bisectors M1 , M2 , M3 . For 1 ≤ i ≤ 3, by (j, ρ)-equivariance, f (vi ) is the symmetric image of f (v) with respect to the perpendicular bisector M′i of ϕ(ei ). In particular, d(f (v), f (vi )) = 2 d(f (v), M′i ). Note that the convex function q 7−→ max 1≤i≤3 H2 , d(q, M′i ) d(v, Mi ) is always ≥ C0 on with equality if and only if q = ϕ(v), in which case all three ratios are equal to C0 . Therefore C(j, ρ) = Lip(f ) ≥ C0 and the constant C0 is achieved, if at all, only by maps that agree with ϕ on the vertices of the tree T . Since the map f of Claim 10.2 is C0 -Lipschitz with stretch locus T , this shows that C(j, ρ) = C0 and that E(j, ρ) = T ; in other words, f is an optimal element of F j,ρ . As in Section 10.4, it is easy to see that no closed geodesic can spend more than a bounded proportion of its length near the trivalent graph G, which implies C ′ (j, ρ) < C(j, ρ) = C0 . Remarks 10.3. • This construction actually gives an open set of pairs (j, ρ) ∈ Hom(Γ0 , G)2 with j convex cocompact and C ′ (j, ρ) < C(j, ρ) < 1. MAXIMALLY STRETCHED LAMINATIONS 99 Indeed, Hom(Γ0 , G)2 has dimension 12 and we have 12 independent parameters, namely ℓ1 , ℓ2 , ℓ3 , θ1 , θ2 , C0 , and a choice of a unit tangent vector in H2 for T and for T ′ (i.e. conjugation of j and ρ). The map from this parameter space to Hom(Γ0 , G)2 is injective since different parameters give rise to different stretch loci; therefore its image is open by Brouwer’s invariance of domain theorem. • There is no constraint on C0 ∈ (0, 1): in particular, ρ could be noninjective or nondiscrete. • A similar construction works for any trivalent topological graph that is a retract of a convex cocompact hyperbolic surface. The invariance of the geodesic realizations under the natural symmetry of the threeholed sphere is replaced by a minimization property for the sum of weighted edge-lengths of the graph. All remaining examples show phenomena specific to the presence of cusps. 10.6. The function (j, ρ) 7→ C(j, ρ) is not upper semicontinuous when j(Γ0 ) has parabolic elements. The following example shows that Proposition 1.5 fails in the presence of cusps, even if we restrict to C < 1. (It certainly fails for larger C since the constant representation ρ can have noncusp-deteriorating deformations, for which C ≥ 1.) Let Γ0 be a free group on two generators α, β and let j ∈ Hom(Γ0 , G) be given by     1 0 1 3 . and j(β) = j(α) = −3 1 0 1 The quotient j(Γ0 )\H2 is homeomorphic to a sphere with three holes, two of which are cusps (corresponding to the orbits of 0 and ∞ in ∂∞ H2 ). Let ρ ∈ Hom(Γ0 , G) be the constant representation, so that C(j, ρ) = 0. We shall exhibit a sequence ρk → ρ with C(j, ρk ) < 1 for all k and C(j, ρk ) √ → 1. k −1 Define ρk (α) (resp. ρ (β)) to be the rotation centered at A := 2 k k √ (resp. Bk := 2−k −1), of angle 2π/(2k k). Note that a circle of radius r in H2 has circumference 2π sinh(r) (see (A.8)), which is equivalent to πer as r → +∞. Therefore, √  √ π −→ 0, ∼ d −1, ρk (α) · −1 k→+∞ k→+∞ k hence (ρk )k∈N converges to the constant representation ρ. By construction, k−1 k−1 ρk (α2 k ) is a rotation of angle π centered at Ak , and ρk (β 2 k ) a rotation k−1 k−1 of angle π centered at Bk . Therefore if ωk = α2 k β 2 k then ρk (ωk ) is a translation of length 2d(Ak , Bk ) = 4k log 2. On the other hand, one can compute explicitly |Tr (j(ωk ))| = (3 · 2k−1 k)2 − 2 which shows that j(ωk ) is a translation of length 4(k log 2 + log k) + O(1). It follows that 1 log k , +O C(j, ρk ) ≥ 1 − k log 2 k which goes to 1 as k → +∞. See Figure 20 for an interpretation of ρk as the holonomy of a singular hyperbolic metric. However, we have C(j, ρk ) < 1 for all k: otherwise, by Corollary 4.14 and Lemmas 4.10, 5.2, and 5.4, the stretch locus E(j, ρk ) would contain a maximally stretched geodesic lamination Lk with compact image L˙k in 100 FRANÇOIS GUÉRITAUD AND FANNY KASSEL H2 mod j(Γ0 ) 2π 2k k f mod Γ0 2π 2k k Figure 20. The representation ρk can be seen as the holonomy of a singular hyperbolic metric on a sphere with three 2π cone points of angle 22π The angle k k , 2k k , and close to 2π. at the third cone point determines the distance between the other two, and is adjusted so that no equivariant map f can be better than (1 − o(1))-Lipschitz, as k → +∞. j(Γ0 )\H2 . Necessarily any recurrent component of L˙k would be a geodesic boundary component of the convex core (a three-holed sphere carries no other recurrent geodesic laminations!), corresponding to αβ ∈ Γ0 . Therefore we would have λ(ρk (αβ)) = C(j, ρ) λ(j(αβ)) ≥ λ(j(αβ)) > 0. This is impossible since ρk tends to the constant √ representation and λ is continuous. ±k −1 for different values of t in (1, 2] Note that by placing Ak , Bk at t (without changing the rotation angle of ρk (α) and ρk (β)), we could also have forced C(j, ρk ) to converge to any value in (0, 1]. 10.7. The function (j, ρ) 7→ C(j, ρ) is not lower semicontinuous when j(Γ0 ) has parabolic elements. Let Γ0 be a free group on two generators α, β and j ∈ Hom(Γ0 , G) the holonomy representation of a hyperbolic metric on a once-punctured torus, with j(αβα−1 β −1 ) parabolic. We assume that j(Γ0 ) admits an ideal square Q of H2 as a fundamental domain, with the axes of j(α) and j(β) crossing the sides of Q orthogonally. Fix two points p, q ∈ H2 distance 1 apart. For each k ≥ 1, let rk ∈ H2 be the point at distance k from p and q, so that pqrk is counterclockwise oriented. Fix a small number δ > 0 and let ρk be the representation of Γ0 taking α (resp. β) to the translation of length δ along the oriented geodesic line (p, rk ) (resp. (q, rk )) — see Figure 21. As k → +∞, the representations ρk converge to a representation ρ fixing exactly one point at infinity (the limit of (rk )k≥1 ), and ρ(αβα−1 β −1 ) is parabolic: hence C(j, ρ) ≥ 1. However, C(j, ρk ) is bounded away from 1 from above. To see this, observe that the fixed points of ρk (αβα−1 β −1 ), ρk (βα−1 β −1 α), ρk (α−1 β −1 αβ), ρk (β −1 αβα−1 ) are the vertices of a quadrilateral Qk with four sides of equal length, centered at rk , of size roughly 2δ. The maps ρk (α), ρk (β) identify pairs of opposite sides of Qk . Taking δ very small, it is not difficult to construct maps Q → Qk (taking whole neighborhoods of the ideal vertices of Q to the vertices of Qk ) that are equivariant with very small Lipschitz constant. MAXIMALLY STRETCHED LAMINATIONS Qk H2 Q H2 j(β) j(α) f 101 rk ρk (α) p ρk (β) q Figure 21. If λ(ρk (α)) and λ(ρk (β)) are small enough, then C(j, ρk ) stays small and bounded away from 1. Note however that the inequality C(j, ρ) ≤ lim inf k C(jk , ρk ) of lower semicontinuity holds as soon as the Arzelà–Ascoli theorem applies for maps fk ∈ F jk ,ρk , i.e. as soon as the sequence (fk (p))k≥1 does not escape to infinity in H2 : this fails only when ρ fixes exactly one point at infinity. 10.8. A reductive, non-cusp-deteriorating ρ with E(j, ρ) = ∅. Let S be a hyperbolic surface of infinite volume with at least one cusp and j ∈ Hom(Γ0 , G) its holonomy representation, where Γ0 := π1 (S). Consider a collection of disjoint geodesics α1 , . . . , αm of S with both ends going out in the funnels, subdividing the convex core of S into contractible polygons and polygons with one puncture (cusp). We apply Thurston’s construction from the proof of Remark 8.2: for each αi we consider another geodesic α′i very close to, but disjoint from αi , and construct the holonomy ρ of a new hyperbolic metric by cutting out the strips bounded by αi ∪ α′i and gluing back the boundaries, identifying the endpoints of the common perpendicular to αi and α′i (see Figure 22). It is easy to check that the (j, ρ)-equivariant map f : H2 → H2 defined by this “cut and paste” procedure is 1-Lipschitz, hence C(j, ρ) ≤ 1. In fact, C(j, ρ) = 1 since ρ is not cusp-deteriorating (Lemma 2.7). However, E(j, ρ) = ∅: otherwise, (5.2) would imply Cs′ (j, ρ) = 1, where Cs′ (j, ρ) is the supremum of λ(ρ(γ))/λ(j(γ)) over all elements γ ∈ Γ0 corresponding to simple closed curves G in S. To see that this is impossible, first note that any such G intersects the arcs αi nontrivially, yielding λ(ρ(γ)) < λ(j(γ)). In fact, G stays in the complement of some cusp neighborhoods, which is compact: this means that G intersects the αi a number of times roughly proportional to the length of G. Moreover, each of these intersections is responsible for a definite (additive) drop in length between λ(j(γ)) and λ(ρ(γ)): this simply follows from the fact that αi is a definite distance away from α′i , and forms with G an angle which can be bounded away from 0 (again by compactness: αi exits the convex core and G must not). This implies Cs′ (j, ρ) < 1. Therefore E(j, ρ) = ∅. A similar argument can be found in [PT]. (This is an example where Cs′ (j, ρ) < 1 = C(j, ρ) = C ′ (j, ρ), the last equality coming from Lemma 7.4.) 102 FRANÇOIS GUÉRITAUD AND FANNY KASSEL α′1 α1 j(Γ0 )\H2 α′2 α2 f mod Γ0 ρ(Γ0 )\H2 Figure 22. In the second surface (with strips removed), simple closed curves are uniformly shorter than in the first. 10.9. A nonreductive, non-cusp-deteriorating ρ with C ′ (j, ρ) < 1 = C(j, ρ) (and E(j, ρ) = ∅). Let Γ0 be a free group on two generators α, β and j ∈ Hom(Γ0 , G) the holonomy representation of a hyperbolic three-holed sphere with one cusp and two funnels, such that j(α) is hyperbolic and j(β) parabolic. For any nonreductive ρ ∈ Hom(Γ0 , G), if ρ(α) and ρ(β) are not hyperbolic (for instance if ρ(Γ0 ) is unipotent), then C ′ (j, ρ) = 0; if moreover ρ(β) is parabolic, then ρ is not cusp-deteriorating and so C(j, ρ) ≥ 1 by Lemma 2.7, which implies E(j, ρ) = ∅ and C(j, ρ) = 1 by Theorem 1.3. Here is another example with C ′ (j, ρ) > 0. Let ρ ∈ Hom(Γ0 , G) be a nonreductive representation with ρ(α) hyperbolic and ρ(β) parabolic; set ε := λ(ρ(α)) > 0. There exists L > 0 with the following property (see [DOP, p. 122], together with Lemma 2.6): for any nontrivial cyclically reduced word γ = αm1 β m2 αm3 β m4 . . . in Γ0 , with m2 · · · ms 6= 0 where ms is the last exponent,   X X |mi | + (1 + log |mi |) . λ(j(γ)) ≥ L  i∈[1,s] odd i∈[1,s] even MAXIMALLY STRETCHED LAMINATIONS 103 On the other hand, for such a γ, λ(ρ(γ)) = ε X mi , i∈[1,s] odd hence λ(ρ(γ))/λ(j(γ)) ≤ ε/L. This shows that C ′ (j, ρ) ≤ ε/L, which is < 1 for ε small enough. However, since ρ(β) is parabolic we have C(j, ρ) ≥ 1 as above, which implies E(j, ρ) = ∅ and C(j, ρ) = 1 by Theorem 1.3. 10.10. In dimension n ≥ 4, the function (j, ρ) 7→ C(j, ρ) is not upper semicontinuous even above 1. When n ≥ 4, the existence of nonunipotent parabolic elements, coming from cusps of rank < n − 2, destroys certain semicontinuity properties of C. We first give an example, in dimension n = 4, where 1 ≤ C(j, ρ) < lim inf C(jk , ρk ) k for some (jk , ρk ) → (j, ρ) with j, jk geometrically finite of the same cusp type, with a cusp of rank 1. This shows that condition (3) of Proposition 6.1 is not satisfied in general for n ≥ 4. Identify ∂∞ H4 with R3 ∪{∞} and let G := PO(4, 1). For ξ ∈ R3 , we denote by Pξ ⊂ H4 the copy of H3 bordered by the unit sphere of R3 centered at ξ. Let Γ0 be a free group on two generators α and β, and let j ∈ Hom(Γ0 , G) be the representation such that • j(α) is the unipotent isometry of H4 fixing ∞ and acting on R3 by translation along the vector (2π, 0, 0); • j(β) is the pure translation (hyperbolic element) taking ξ := (3, 0, 0) to ∞, and ∞ to η := (0, 0, 0), and Pξ to Pη . It is a standard argument (sometimes called “ping pong”) that j(α) and j(β) generate a free discrete group in G; the representation j is geometrically finite and the quotient manifold j(Γ0 )\H4 has one cusp, with stabilizer hαi ⊂ Γ0 . Take ρ = ρk = j, so that C(j, ρ) = 1. Choose an integer p ≥ 2 and, for k ≥ 1, let jk ∈ Hom(Γ0 , G) be the representation such that • jk (α) is the parabolic element of G fixing ∞ and acting on R3 as the corkscrew motion preserving the line ℓk := √ {0} × R × {k}, with rotation angle 2π/k around ℓk and progression p k/k along ℓk ; • jk (β) = j(β). It is an easy exercise to check that jk → j as k → +∞. Moreover, jk is geometrically finite with the same cusp type as j for large k, by a standard ping pong argument. (Note however that fundamental domains for the action of jk (Γ0 ) do not converge to a fundamental domain for the action of j(Γ0 ), but to a smaller set.) The element ρk (αk β) = j(αk β) takes ξ to ∞, and ∞ to j(αk )(η) = (2kπ, 0, 0), and Pξ to Pj(αk )(η) ; by (A.11),  (10.3) λ j(αk β) ≥ 2 log 2πk − R for some R > 0 independent of k. On the other hand, jk (αk β) takes ξ to ∞, √ p k and ∞ to jk (α )(η) = (0, k, 0), and Pξ to Pjk (αk )(η) , hence √  p λ jk (αk β) ≤ 2 log k + R + 1 104 FRANÇOIS GUÉRITAUD AND FANNY KASSEL by (A.11). It follows, by (4.1), that λ(ρk (αk β)) 2 log 2πk − R √ , ≥ λ(jk (αk β)) 2 log p k + R + 1 which accumulates only to values ≥ p as k → +∞. Since p was arbitrary, we see that (j ′ , ρ′ ) 7→ C(j ′ , ρ′ ) is not even bounded near (j, ρ). C(jk , ρk ) ≥ 10.11. The condition C(j, ρ) < 1 is not open in dimension n ≥ 4. We finally give an example, in dimension n = 4, where C(j, ρ) < 1 < lim inf C(jk , ρk ) k for (jk , ρk ) → (j, ρ) with j, jk geometrically finite of the same cusp type, with a cusp of rank 1, and with ρk (and ρ) cusp-deteriorating. This proves that condition (1) of Proposition 6.1 need not be satisfied for n ≥ 4 when there is a cusp of rank < n − 2. Let Γ0 be a free group on two generators α and β, and let j and jk be as in Section 10.10. We take a representation ρ ∈ Hom(Γ0 , G) such that • ρ(α) = 1 ∈ G, • ρ(β) is a pure translation along some line ℓ of H4 . Since ρ(Γ0 ) is contained in the stabilizer of ℓ, multiplying the translation length of ρ(β) by some constant ε > 0 multiplies the translation length of all elements ρ(γ) by ε. Therefore, up to taking λ(ρ(β)) small enough, we may assume C(j, ρ) < 1. Up to conjugating ρ, we can furthermore assume that apart by (A.12)) such that ρ(β) there exist ξ, η ∈ R3 (distance 2 cosh λ(ρ(β)) 2 takes ξ to ∞, and ∞ to η, and Pξ to Pη . We still normalize to η = (0, 0, 0) for convenience. We then take ρk ∈ Hom(Γ0 , G) such that • ρk (α) is an elliptic transformation fixing pointwise the hyperbolic √ ′ 2-plane bordered by the line ℓk := {0} × R × { k} of R3 (compactified at ∞), and acting as a rotation of angle π/k in the orthogonal direction, • ρk (β) = ρ(β). Clearly ρk (α) → ρ(α) as k → +∞, since this holds in restriction to any horosphere centered at ∞ (such a horosphere is stable under √ ρk (α)). This time, ρk (αk β) takes ξ to ∞, and ∞ to ρk (αk )(η) = (0, 0, 2 k), and Pξ to Pρk (αk )(η) , hence √  λ ρk (αk β) ≥ 2 log 2 k − R by (A.11). Using (10.3), we obtain √ λ(ρk (ak b)) 2 log 2 k − R √ , C(jk , ρk ) ≥ ≥ λ(jk (ak b)) 2 log p k + R + 1 which accumulates only to values ≥ p/2 as k → +∞. Since p was arbitrary, we see that (j ′ , ρ′ ) 7→ C(j ′ , ρ′ ) is not bounded near (j, ρ), even in restriction to cusp-deteriorating ρ′ . Appendix A. Some hyperbolic trigonometry We collect a few well-known formulas in hyperbolic trigonometry, from which we derive several formulas used at various places in the paper. MAXIMALLY STRETCHED LAMINATIONS 105 A.1. Distances in H2 and H3 . Let n = 2 or 3. We use the upper halfspace model of Hn : if n = 2, then Hn ≃ {z ∈ C : Im(z) > 0}, the hyperbolic metric is given by d|z|2 , ds2 = Im(z)2 the isometry group G of Hn identifies with PGL2 (R) acting by Möbius transformations, and ∂∞ Hn ≃ R ∪ {∞}. If n = 3, then Hn ≃ C × R∗+ , the hyperbolic metric is given by d|a|2 + db2 b2 ∗ for (a, b) ∈ C × R+ , the identity component G0 of G identifies with PSL2 (C), which acts on the boundary ∂∞ Hn ≃ C ∪ {∞} by Möbius transformations, and this action extends in a natural way to Hn . The matrix  η/2  e 0 Tη := ∈ G0 0 e−η/2 ds2 = defines a translation of (complex) √ length η along the geodesic line with endpoints 0, ∞ ∈ ∂∞ Hn . Set p0 := −1 ∈ Hn if n = 2, and p0 := (0, 1) ∈ Hn if n = 3. Then   cos(θ/2) sin(θ/2) Rθ := ∈ G0 − sin(θ/2) cos(θ/2) defines a rotation of angle θ around p√0 if n = 2, and around the geodesic line (containing p0 ) with endpoints ± −1 ∈ ∂∞ Hn if n = 3. The stabilizer of p0in G0 is K = PSO(2) if n = 2, and K = PSU(2) if n = 3. For any a b g= ∈ G0 , c d   2 a b (A.1) 2 cosh d(p0 , g · p0 ) = := |a|2 + |b|2 + |c|2 + |d|2 . c d Indeed this holds for g = Tη and the right-hand side is invariant under multiplication of g by elements of K on either side (recall the Cartan deA := {Tη | η ∈R}, see Section 7.3). Suppose composition G0 = KAK for  v 1/2 uv −1/2 n = 2; applying (A.1) to g = , we find in particular that for 0 v −1/2 any u, v ∈ R with v > 0,   2 √ √ u + v2 + 1 . (A.2) d −1, u + −1v) = arccosh 2v   1 L A.1.1. Horospherical distances. Applying (A.1) to g = , we see that 0 1 for any points p, q on a common horosphere ∂H, the distance d(p, q) from p to q in Hn and the distance L = d∂H (p, q) of the shortest path from p to q contained in the horosphere ∂H (“horocyclic distance”) satisfy  d (p, q)   d∂H (p, q)2  ∂H = 2 arcsinh . (A.3) d(p, q) = arccosh 1 + 2 2 106 FRANÇOIS GUÉRITAUD AND FANNY KASSEL Let t 7→ pt and t 7→ qt be the geodesic rays from p and q to the center ξ ∈ ∂∞ Hn of the horosphere ∂H, parametrized by arc length. Then (A.4) d∂Ht (pt , qt ) = e−t d∂H (p, q) for all t ≥ 0, where ∂Ht is the horocycle through pt and qt centered at ξ. Using (A.3) and the concavity of arcsinh, we find that there exists D > 1 such that (A.5) e−t d(p, q) ≤ d(pt , qt ) ≤ D e−t d(p, q) for all t ≥ 0; moreover, an upper bound on d(p, q) yields one on D. A.1.2. Distances in two ideal spikes of H2 . The following situation is considered in the proof of Proposition 9.4. Let ζ1 6= ζ2 6= ζ3 6= ζ4 be points of ∂∞ H2 , not necessarily all distinct. Let Di−1 , Di , Di+1 be the geodesic lines of H2 running from ζ1 to ζ2 , from ζ2 to ζ3 , and from ζ3 to ζ4 respectively. Consider two points x ∈ Di−1 and x′ ∈ Di on a common horocycle centered at ζ2 and let ξ ≥ 0 be their horocyclic distance. Similarly, consider two points y ∈ Di+1 and y ′ ∈ Di on a common horocycle centered at ζ3 and let η ≥ 0 be their horocyclic distance. Setting L := d(x′ , y ′ ), we have (A.6) d(x, y) = L + ξ 2 + η 2 + o(ξ 2 + η 2 ) as ξ 2 + η 2 + e−L → 0. Indeed, by (A.1),     1 1 0 1 −η T cosh d(x, y) = ξ 1 L 0 1 2 2 = cosh L + (sinh L) · (ξ 2 + η 2 )(1 + o(1)) and we conclude using the degree-1 Taylor series of cosh at L. A.1.3. Distances in a prism in H3 . The following situation is considered in the proof of Lemma 5.13. Consider a geodesic segment I of H3 , of length η ≥ 0, together with two oriented geodesic lines ℓ, ℓ′ of H3 meeting I orthogonally at its endpoints, and forming an angle θ ∈ [0, π] with each other. Note that I is the shortest geodesic segment between ℓ and ℓ′ ; the complex number L := η + iθ ∈ C is called the complex distance between ℓ and ℓ′ and will be expressed in terms of cross-ratios in Section A.1.4. For now, let us compute the distance between points p ∈ ℓ and q ∈ ℓ′ , at respective signed distances s and t from I. Note that Ts′ := Rπ/2 Ts R−π/2 ∈ G0 is a translation of length s along the geodesic line from −1 ∈ ∂∞ H3 to 1 ∈ ∂∞ H3 , which intersects the translation axis of TL (with endpoints 0, ∞ ∈ ∂∞ H3 ) perpendicularly at the basepoint p0 = (0, 1) ∈ H3 . Define ′ T T ′ . Without loss of generality, we may assume that p = p and g := T−s L t 0 ′ · p , T ′ T · p ]. Using (A.1) and the identities q = g · p, and that I = [T−s 0 −s L 0 2| cosh L2 |2 = cosh η + cos θ and 2| sinh L2 |2 = cosh η − cos θ, this gives (A.7) cosh d(p, q) = cosh η cosh s cosh t − cos θ sinh s sinh t . When η = 0 and θ = 0 or π we recover the formulas for cosh(s ± t). When η = 0 and s = t, we find that points p, q on a circle of radius s, forming an angle θ from the center, are a distance ∼ θ sinh(s) apart when θ is small. This estimate is needed in the proof of Lemma 6.4: approaching the arc of circle C from p to q with a union of short geodesic segments, we MAXIMALLY STRETCHED LAMINATIONS 107 find in the limit (A.8) length(C ) = θ sinh r. When θ = 0 and s = t, we find that points p, q at (signed) distance s from a straight line A of H2 , whose projections to A are distance η apart, satisfy d(p, q) ∼ η cosh s (A.9) when η is small. (This situation is considered in the proof of Claim 10.1.) ′ , ξ ′ ∈ ∂ H3 , the complex A.1.4. Line-to-line distances. For any ξ− , ξ+ , ξ− ∞ + ′ , ξ ′ ) satisfies distance L = η + iθ between the oriented lines (ξ− , ξ+ ) and (ξ− + (A.10) cosh η = 1 + |c| |1 − c| and cos θ = 1 − |c| , |1 − c| ′ : ξ ′ ] is the cross-ratio of ξ , ξ , ξ ′ , ξ ′ , defined so where c := [ξ− : ξ+ : ξ− − + − + + that [∞ : 0 : 1 : ξ] = ξ. Indeed, by invariance under the action of G0 , it is sufficient to check (A.10) when (ξ− , ξ+ ) = (−1, 1) and the shortest geodesic ′ , ξ ′ ) is contained in (0, ∞). In this case, segment between (ξ− , ξ+ ) and (ξ− + ′ ′ (ξ− , ξ+ ) is of the form (−z, z) for some z ∈ C∗ and we have   cosh η = (|z| + |z|−1 )/2, cos θ = Re(z/|z|),  c = (1 − z)2 /(1 + z)2 ; it is then elementary to check that (A.10) holds. Now let n ≥ 2 be arbitrary and consider as above the upper half space model of Hn , so that ∂∞ Hn identifies with Rn−1 ∪ {∞}. In Sections 10.10 and 10.11 we use the following consequence of (A.10): there exists R > 0 such that for any D ≥ 2, any ξ, η ∈ Rn−1 ⊂ ∂∞ Hn distance D apart for the Euclidean metric, and any g ∈ Isom(Hn ), if g(ξ) = ∞ = g−1 (η) and if g maps the unit hemisphere (geodesic hyperplane) Pξ centered at ξ to the unit hemisphere Pη centered at η, then g is hyperbolic and its translation length λ(g) satisfies (A.11) |λ(g) − 2 log D| ≤ R. Indeed, by (A.10) the hyperbolic distance between the closest points of Pξ and Pη is (A.12) 2 arccosh(D/2), and the intersection point of Pξ with the line (ξ, ∞) is at hyperbolic distance 2 arcsinh(D/2) from Pη ∩ (η, ∞) by (A.3). The translation length λ(g) is bounded in-between these two values, which are both 2 log D + O(1). A.2. Relations in a right-angled hyperbolic triangle. Consider a trianb B, b C b and opposite edge lengths a, b, c. Suppose gle ABC in H2 with angles A, b = π/2. Then B b = tanh c , and sin A b = sinh a . b = tanh a , cos A (A.13) tan A sinh c tanh b sinh b 108 FRANÇOIS GUÉRITAUD AND FANNY KASSEL b b Indeed, let (α, β, γ) := (ea/2 , eb/2 , ec/2 ) and (X, Y ) := (cos A2 , sin A2 ): following the perimeter of the triangle in the order C, A, B, C shows that ! γ α 1 X αγ + Y −X + Y β βγ αβ αβγ T−b RAb Tc R−π/2 Ta = αβγ β X αγ + Y βγ X αβ γ −Y 1 α must be (projectively) a rotation matrix, namely R−Cb . After multiplying all entries by αβγ, this means α2 (Xγ 2 + Y ) = β 2 (X + Y γ 2 ) and α2 β 2 (X − Y γ 2 ) = Xγ 2 − Y. It follows that 2 tanh a = = and γ 2 − γ −2 2XY b = sinh c tan A 2 X2 − Y 2 2 tanh b = 2 X+Y γ 2 X−Y γ β 2 Xγ 2 +Y − β Xγ 2 −Y α2 − α−2 = 2 X+Y γ 2 2 X−Y γ α2 + α−2 β 2 Xγ 2 +Y + β Xγ 2 −Y 2 +Y 2 X−Y γ α2 Xγ β 2 − β −2 X+Y γ 2 − α Xγ 2 −Y = 2 +Y X−Y γ 2 β 2 + β −2 + α2 Xγ α2 Xγ 2 −Y X+Y γ 2 γ 2 − γ −2 X 2 + Y 2 tanh c = . b γ 2 + γ −2 X 2 − Y 2 cos A The last identity in (A.13) follows from the first two and from the Pythagorean identity cosh b = cosh a cosh c, which is just (A.7) for (η, θ) = (0, π/2). As a consequence of the last identity in (A.13), if x, y are two points on a circle of radius r in H2 , forming an angle θ from the center, then sinh(d(x, y)/2) θ . (A.14) sin = 2 sinh r A.3. The closing lemma. Finally, we recall the following classical statement; see [BBS, Th. 4.5.15] for a proof. = Lemma A.1. For any δ > 0 and D > 0, there exists ε > 0 with the following property: given any broken line L = p0 · · · pk+1 in Hn , if d(pi , pi+1 ) ≥ D for all 1 ≤ i < k and if the angle pi−1 \ pi pi+1 is ≥ π − ε for all 1 ≤ i ≤ k, then L stays within distance δ from the segment [p0 , pk+1 ], and has total length at most d(p0 , pk+1 ) + kδ. Moreover, when δ is fixed, ε = δ will do for all large enough D. Taking limits as k → +∞, this implies in particular that for a broken line (pi )i∈Z invariant under a hyperbolic element g ∈ G taking each pi to pi+m , under the same assumptions on length and angle we have m X d(pi , pi+1 ) ≤ mδ. λ(g) − i=1 Appendix B. Converging fundamental domains Let Γ0 be a discrete group. It is well known that, in any dimension n ≥ 2, the set of convex cocompact representations of Γ0 into G = PO(n, 1) = Isom(Hn ) is open in Hom(Γ0 , G) (see [Bo2, Prop. 4.1] for instance). The set MAXIMALLY STRETCHED LAMINATIONS 109 of geometrically finite representations is open in the set of representations Γ0 → G of fixed cusp type if n ≤ 3 [Ma], or if all cusps have rank ≥ n − 2 [Bo2, Prop. 1.8], but not in general for n ≥ 4 [Bo2, § 5]. In Sections 6.1.1 and 6.2 of the paper, where we examine the continuity properties of the function (j, ρ) 7→ C(j, ρ), we need, not only this openness, but also a control on fundamental domains in Hn for converging sequences of geometrically finite representations. Propositions B.1 and B.3 below are certainly well known to experts, but we could not find a proof in the literature. Note that they easily imply the Hausdorff convergence of the limit sets, but are a priori slightly stronger. B.1. The convex cocompact case. Proposition B.1. Let Γ0 be a discrete group and (jk )k∈N∗ a sequence of elements of Hom(Γ0 , G) converging to a convex cocompact representation j ∈ Hom(Γ0 , G). Then for any large enough k ∈ N∗ the representation jk is convex cocompact. Moreover, there exists a compact set C ⊂ Hn that contains fundamental domains of the convex cores of j(Γ0 )\Hn and jk (Γ0 )\Hn for all large enough k ∈ N∗ . If Γ0 is torsion-free, then the injectivity radius of jk (Γ0 )\Hn is bounded away from 0 as k → +∞. Proposition B.1 for torsion-free Γ0 implies the general case, due to the Selberg lemma [Se, Lem. 8]. We henceforth assume Γ0 to be torsion-free. Proof. We build fundamental domains as finite unions of simplices coming from j(Γ0 )-invariant triangulations: the main step is the following. Claim B.2. There exists a j(Γ0 )-invariant geodesic triangulation ∆ of a nonempty convex subset of Hn which is finite modulo j(Γ0 ) and induces dihedral angles < π on the boundary. Let us prove Claim B.2 (note that the projection of ∆ to M := j(Γ0 )\Hn will automatically contain the convex core). The idea is to use a classical construction, the hyperbolic Delaunay decomposition (an analogue of the Euclidean Delaunay decomposition of [D]), and make sure that it is finite modulo j(Γ0 ). Let N ⊂ Hn be the preimage of the convex core of M = j(Γ0 )\Hn and let N be the uniform 1-neighborhood of N . For R ≥ 0, we call R-hyperball of Hn any convex region of Hn bounded by a connected hypersurface at constant distance R from a hyperplane. Since N is the intersection of all half-spaces containing N , we see that N is the intersection of all 1-hyperballs containing N . By the strict convexity of the distance function in Hn , there exists α > 0 such that whenever points p, q of a 1hyperball are distance ≥ 1 apart, the ball of radius α centered at the midpoint of [p, q] is also contained in the 1-hyperball. Therefore, whenever p, q ∈ N are distance ≥ 1 apart, the ball of radius α centered at the midpoint of [p, q] is also contained in N . Let X be a j(Γ0 )-invariant subset of N that is finite modulo j(Γ0 ) and intersects every ball of radius ≥ α/2 centered at a point of N . We view X as a subset of Rn+1 via the embedding of Hn as the upper hyperboloid sheet  H := x ∈ Rn+1 : x21 + · · · + x2n − x2n+1 = −1 and xn+1 > 0 . 110 FRANÇOIS GUÉRITAUD AND FANNY KASSEL b of X in Rn+1 . There is a natural bijection Consider the convex hull X between the following two sets: b separating X from 0 ∈ Rn+1 , • the set of supporting hyperplanes of X • the set of open balls, horoballs, or hyperballs of Hn that are disjoint from X but whose boundary intersects X. Namely, the bijection is given by taking any supporting hyperplane to the set of points of H that it separates from X (see Figure 23). This set is a ball (resp. a horoball, resp. a hyperball) if the intersection of H with the supporting hyperplane is an ellipsoid (resp. a paraboloid, resp. a hyperboloid). The degenerate case of a supporting hyperplane tangent to H corresponds to an open ball of radius 0 (the empty set!) centered at a point of X; the limit case of a supporting hyperplane containing 0 ∈ Rn+1 corresponds to a 0-hyperball, i.e. a half-space of Hn . H H2 Figure 23. Balls, horoballs, and hyperballs of Hn are intersections of the hyperboloid sheet H with affine half-spaces of Rn+1 containing the origin. b the corresponding open ball, horoball, For any supporting hyperplane of X, n or hyperball B ⊂ H intersects N in a region of diameter ≤ 1. Indeed, if p, q ∈ B ∩ N were distance > 1 apart, then the ball B ′ of radius α centered at the midpoint of [p, q] would be contained in N , by choice of α. But at least one hemisphere of B ′ (the hemisphere closest to the center or to the defining hyperplane of B, depending on whether B is a (horo)ball or a hyperball) would also be contained in B. A ball of radius α/2 contained in this hemisphere would intersect X (by assumption on X), while being contained in B: impossible. Thus ∂B ∩ N has diameter ≤ 1. In particular, ∂B ∩ X has diameter ≤ 1. In particular, ∂B ∩ X is finite. b be the union of all points that belong to supporting hyperLet Y ⊂ ∂ X planes separating X from 0 ∈ Rn+1 . (In other words, Y is the portion of b that is “visible from the origin”. There can also be an “invisible” por∂X tion, corresponding to hyperballs whose complement is disjoint from X.) By the previous paragraph, Y has the structure of a locally finite polyhedral hypersurface in Rn+1 , with vertex set X. Projecting each polyhedron of Y MAXIMALLY STRETCHED LAMINATIONS 111 to the hyperboloid H ≃ Hn (along the rays through the origin 0 ∈ Rn+1 ), we obtain a cellulation ∆ of the convex hull Conv(X) of X in Hn , called the Delaunay cellulation of Conv(X) relative to X. It is characterized by the fact that any cell of ∆ is inscribed in a hypersurface of Hn bounding some open ball, horoball, or hyperball disjoint from X. The cellulation ∆ is j(Γ0 )-invariant and finite modulo j(Γ0 ). Since j(Γ0 ) is torsion-free, up to taking the points of X in general position we may assume that ∆ is a b This triangulation and induces dihedral angles < π on the boundary of X. completes the proof of Claim B.2. Proposition B.1 easily follows from Claim B.2. Indeed, let F ⊂ Hn be a finite set such that X = j(Γ0 ) · F . The vertices of a d-dimensional simplex of the triangulation ∆ can be listed in the form j(γ0 ) · p0 , . . . , j(γd ) · pd , where p0 , . . . , pd ∈ F and γ0 , . . . , γd ∈ Γ0 . By finiteness of the triangulation, when jk is close enough to j the points jk (γ0 ) · p0 , . . . , jk (γd ) · pd still span a simplex and these simplices (obtained by following the combinatorics of ∆) still triangulate a region of Hn that is locally convex, hence globally convex. In particular, this region of Hn contains the preimage of the convex core of jk (Γ0 )\Hn . Thus jk (Γ0 ) is still convex cocompact for large k ∈ N∗ . Any compact neighborhood C of a union U of representatives of simplex orbits of ∆ under j(Γ0 ) contains a fundamental domain of the convex core of jk (Γ0 )\Hn for all large enough k. To bound the injectivity radius of jk (Γ0 )\Hn away from 0, we argue as follows. For any p ∈ F , let Up be the union of all simplices of ∆ containing p. Then p is an interior point of Up . Provided X is dense enough in N , each Up projects injectively to M = j(Γ0 )\Hn . For ε > 0, let Upε be the complement in Up of the ε-neighborhood of ∂Up . If ε is small enough, then any point of ∆ has a translate belonging to some Upε with p ∈ F , whose ε-neighborhood therefore projects injectively to M . This property remains true as ∆ (hence the finitely many sets Up ) are deformed slightly, up to taking a smaller ε. This completes the proof of Proposition B.1.  Note that in the above proof, for torsion-free Γ0 , the hyperbolic manifolds jk (Γ0 )\Hn and j(Γ0 )\Hn are in fact homeomorphic since their convex cores admit combinatorially identical triangulations. B.2. The geometrically finite case when all cusps have rank ≥ n − 2. Here is an analogue of Proposition B.1 for geometrically finite representations of fixed cusp type with all cusps of rank ≥ n − 2. Note that cusps always have rank ≥ n − 2 in dimension n ≤ 3. Proposition B.3. Let Γ0 be a discrete group, j ∈ Hom(Γ0 , G) a geometrically finite representation with all cusps of rank ≥ n − 2, and (jk )k∈N∗ a sequence of elements of Hom(Γ0 , G) converging to j, all of the same cusp type as j (Definition 1.1). Then for any large enough k ∈ N∗ the representation jk is geometrically finite and (jk )k∈N converges geometrically to j. More precisely, if H1 , . . . , Hc are horoballs of Hn whose images in j(Γ0 )\Hn are disjoint, small enough, and intersect the convex core in standard cusp regions (Definition 2.2), representing all the cusps, then there exist a compact set C ⊂ Hn and, for any large enough k ∈ N∗ , horoballs H1k , . . . , Hck of Hn , such that 112 FRANÇOIS GUÉRITAUD AND FANNY KASSEL • the images of H1k , . . . , Hck in jk (Γ0 )\Hn are disjoint and intersect the convex core in standard cusp regions; • the stabilizer in Γ0 of Hik under jk is the stabilizer in Γ0 of Hi under j for all 1 ≤ i ≤ c; • the horoballs Hik converge to Hi for all 1 ≤ i ≤ c, as k → +∞; • the union of C and of H1 ∪ · · · ∪ Hc (resp. H1k ∪ · · · ∪ Hck ) contains a fundamental domain of the convex core of j(Γ0 )\Hn (resp. jk (Γ0 )\Hn ); • the union of all geodesic rays from C to the centers of H1 , . . . , Hc (resp. H1k , . . . , Hck ) contains a fundamental domain of the convex core of j(Γ0 )\Hn (resp. jk (Γ0 )\Hn ); in particular, the S cusp thickness (Definition 5.12) of jk (Γ0 )\Hn at any point of 1≤i≤c ∂Hik is uniformly bounded by some constant independent of k. Moreover, if j(Γ0 ) is torsion-free, then the infimum of injectivity radii of jk (Γ0 )\Hn at projections of points of C is bounded away from 0 as k → +∞. Again, we can and will assume that Γ0 is torsion-free. Proposition B.3 fails in dimension n ≥ 4 when j has a cusp of rank < n − 2, as can be seen from [Bo2, § 5] or by adapting the examples of geometrically finite representations jk from Sections 10.10 and 10.11. In order to prove Proposition B.3, we need the following lemma, which is also used in Section 6.2. Lemma B.4. Let j ∈ Hom(Γ0 , G) be a geometrically finite representation with all cusps of rank ≥ n − 2, and let N ⊂ Hn be a uniform neighborhood of the preimage N of the convex core of j(Γ0 )\Hn . For any horoball H of Hn such that H ∩ N projects to a standard cusp region and such that the cusp thickness (Definition 5.12) of j(Γ0 )\Hn at any point of H is ≤ 1/3, the set ∂H ∩ N is convex in ∂H ≃ Rn−1 , equal to • the full Euclidean space ∂H if the cusp has rank n − 1; • the region contained between two parallel (possibly equal) Euclidean hyperplanes of ∂H if the cusp has rank n − 2. Proof of Lemma B.4. The stabilizer S ⊂ Γ0 of H under j has a finite-index normal subgroup S ′ isomorphic to Zm , where m ∈ {n − 1, n − 2} is the rank of the cusp (see Section 2.1). In the upper half-space model Rn−1 × R∗+ of Hn , in which ∂∞ Hn identifies with Rn−1 ∪ {∞}, we may assume that H is centered at infinity, so that ∂H = Rn−1 × {b} for some b > 0. Let Ω be the convex hull of Λj(Γ0 ) r {∞} in Rn−1 , where Λj(Γ0 ) is the limit set of j(Γ0 ). The group j(S) acts on Rn−1 by Euclidean isometries and there is an m-dimensional affine subspace V ⊂ Ω, preserved by j(S), on which j(S ′ ) acts as a lattice of translations (see Section 2.1). This implies that Ω is either Rn−1 (if m = n−1) or the region contained between two parallel hyperplanes of Rn−1 (if m = n − 2). Let δ > 0 be the Euclidean diameter of j(S)\Ω. Then δ is the cusp thickness of j(Γ0 )\Hn at Rn−1 × {1}, or alternatively the cusp thickness of j(Γ0 )\Hn is ≤ 1/3 exactly on Rn−1 × [3δ, +∞). Every S-orbit in Ω is δ-dense in Ω. Lemma B.4 reduces to:  (B.1) N ∩ Rn−1 × [3δ, +∞) = Ω × [3δ, +∞). MAXIMALLY STRETCHED LAMINATIONS 113 The left-hand side is always contained in the right-hand side since N is the convex hull in Hn of the limit set Λj(Γ0 ) . For the converse, suppose a point p ∈ Ω × R∗+ lies outside N . Then p belongs to a closed half-space D of Hn disjoint from N , whose ideal boundary ∂∞ D is a Euclidean (n − 1)dimensional ball disjoint from Λj(Γ0 ) . If ∂∞ D is centered inside Ω (e.g. if m = n − 1), then ∂∞ D has radius < δ because Λj(Γ0 ) is δ-dense in Ω. This yields p ∈ Ω × (0, δ) ⊂ Ω × (0, 3δ), as desired. Now suppose ∂∞ D is centered outside Ω; this may only happen if m = n−2. Let P be the connected component of ∂Ω closest to the center of ∂∞ D, and P ′ the other component: P and P ′ are parallel hyperplanes of Rn−1 both intersecting Λj(Γ0 ) (and possibly equal, if j(Γ0 ) preserves a copy of Hn−1 ). We claim that P ∩ Λj(Γ0 ) is 3δ-dense in P : indeed, fix ξ ∈ P ∩ Λj(Γ0 ) . If j(S) · ξ ⊂ P (in particular if P = P ′ ) we are done, as j(S) · ξ is already δdense in Ω. If not, since j(S) · ξ ⊂ P ∪ P ′ we can choose ξ ′ ∈ P ′ ∩ j(S) · ξ. Let S ′ ⊂ S be the stabilizer of P , which has index two in S, so that j(S ′ )\Ω has twice the Euclidean volume of j(S)\Ω. Let U and U ′ be (closed) Dirichlet fundamental domains of j(S)\Ω for the Euclidean metric, centered at ξ and ξ ′ respectively. Any lifts of U and U ′ to j(S ′ )\Ω must overlap, because of their total volume. Since U and U ′ are contained in δ-balls centered at ξ and ξ ′ , this shows that ξ and ξ ′ are at most 2δ apart in the quotient j(S ′ )\Ω. Since j(S) · ξ = j(S ′ ) · {ξ, ξ ′ } is δ-dense in Ω, we get that j(S ′ )ξ ⊂ P is 3δ-dense (and contained in P ∩ Λj(Γ0 ) ). The intersection ∂∞ D ∩ P , being disjoint from Λj(Γ0 ) , is therefore an (n − 2)-dimensional Euclidean ball of radius < 3δ. Since the hemisphere D is centered outside Ω, this shows that D does not achieve any height ≥ 3δ inside Ω × R∗+ . In particular, p ∈ D ∩ (Ω × R∗+ ) ⊂ Ω × (0, 3δ). This proves (B.1).  Proof of Proposition B.3. We proceed as in the proof of Proposition B.1 and first establish the following analogue of Claim B.2. Claim B.5. There exists a j(Γ0 )-invariant geodesic triangulation ∆ of a nonempty convex subset of Hn with the following properties: • ∆ is finite modulo j(Γ0 ), with vertices lying both in Hn and in ∂∞ Hn ; • the vertices in ∂∞ Hn are exactly the parabolic fixed points of j(Γ0 ); • no edge of ∆ connects two such (ideal) vertices; • in a neighborhood of a parabolic fixed point ξ of rank n−2, the boundary of ∆ consists of two totally geodesic hyperplanes of Hn meeting only at ξ. Let N ⊂ Hn be the uniform 1-neighborhood of the preimage N of the convex core of M = j(Γ0 )\Hn . Let X be a j(Γ0 )-invariant subset of N that is locally finite modulo j(Γ0 ) and intersects every ball of diameter ≥ α centered at a point of N , where α > 0 is chosen as in the proof of Claim B.2: whenever points p, q of a 1-hyperball of Hn are distance ≥ 1 apart, the ball of radius α centered at the midpoint of [p, q] is also contained in the 1-hyperball. By a similar argument to the proof of Claim B.2, the Delaunay cellulation ∆ of Conv(X) with respect to X is locally finite, with all cells equal to compact polyhedra of diameter ≤ 1. It remains to make ∆ finite modulo j(Γ0 ) by modifying it inside each cusp. For this purpose, we choose X carefully. 114 FRANÇOIS GUÉRITAUD AND FANNY KASSEL Let H1 , . . . , Hc be open horoballs of Hn , centered at points ξ1 , . . . , ξc ∈ ∂∞ Hn , whose images in j(Γ0 )\Hn are disjoint and intersect the convex core in standard cusp regions, representing all the cusps. We take these horoballs at distance > 2 from each other in j(Γ0 )\Hn , and small enough so that the conclusions of Lemma B.4 are satisfied. Choose the j(Γ0 )-invariant, locally finite set X in general position subject to the following constraints: S (∗) X r ci=1 ∂Hi stays at distance ≥ α′ from ∂N and from each ∂Hi , for some α′ ∈ (0, α) to be determined below; (∗∗) for any 1 ≤ i ≤ c, the set X ∩ ∂Hi intersects any ball of Hn of radius α′ /2 centered at a point of ∂Hi ∩ N ; (∗∗∗) if the stabilizer of Hi has rank n − 2, then X intersects any Euclidean ball of radius α′ /8 in the boundary of ∂Hi ∩ N in ∂Hi ≃ Rn−1 , while all other points of X in ∂Hi ∩ N are at distance ≥ α′ /4 from the boundary of ∂Hi ∩ N (which by Lemma B.4 consists of two parallel (n − 2)-dimensional Euclidean hyperplanes of ∂Hi ). Qualitatively, this implies X is especially concentrated on the horospheres ∂Hi and even more on ∂N ∩ ∂Hi , but with a little buffer around these high-concentration regions. Consider the Delaunay cellulation ∆ of Conv(X) with respect to such a set X. Suppose two vertices x, y of a given cell of ∆ (inscribed in a hypersurface bounding an open ball, horoball, or hyperball B of Hn disjoint from X) lie on opposite sides of one of the horospheres ∂Hi . By (∗), the points x and y lie at distance ≥ α′ from ∂N , hence so does the intersection point {z} = [x, y]∩∂Hi . But at least one half of the ball of radius α′ centered at z is contained in B, hence B ∩ X 6= ∅ by (∗∗): impossible. Therefore the given cell of ∆ either has all its vertices in the closure of Hi , or has all its vertices in Hn r Hi . We can thus partition the cells of ∆ (of any dimension) into • interface cells, with all their vertices in some j(γ) · ∂Hi ; • thin-part cells, with all their vertices in the closure of some j(γ) · Hi (not all in the horosphere j(γ) · ∂Hi ); S • thick-part cells, with all their vertices in Hn r j(Γ0 ) · ci=1 Hi (not all in the horospheres j(γ) · ∂Hi ). Consider the Euclidean Delaunay cellulation ∆∂Hi of the Euclidean convex hull of X ∩ ∂Hi in ∂Hi , with respect to X ∩ ∂Hi , in the classical sense (see [D]): by definition, any cell of ∆∂Hi is inscribed in some Euclidean sphere bounding an open Euclidean ball of ∂Hi disjoint from X ∩ ∂Hi . Claim B.6. The geodesic straightenings of the Euclidean triangulations ∆∂Hi give exactly the interface cells. Proof. For any interface cell W of ∆, the projection of W to ∂Hi is a cell of ∆∂Hi . Indeed, if W is inscribed in an open ball, horoball, or hyperball B of Hn disjoint from X, then the projection of W is inscribed in B ∩ ∂Hi , which is a Euclidean ball (or half-plane) of ∂Hi disjoint from X. Conversely, for any cell WE of ∆∂Hi , the geodesic straightening of WE is contained in ∆ as an interface cell. Indeed, suppose WE is inscribed in an open Euclidean ball BE of ∂Hi , disjoint from X ∩ ∂Hi , and centered in ∂Hi ∩ N . By (∗∗), the hyperbolic ball B concentric to BE such that MAXIMALLY STRETCHED LAMINATIONS 115 B ∩ ∂Hi = BE has radius ≤ α′ /2, hence is disjoint from X by (∗), which means that the geodesic straightening of WE is contained in ∆. Therefore, we just need to see that WE is always inscribed in such a ball BE . If Hi has rank n − 1, this follows from the fact that ∂Hi ∩ N = ∂Hi by Lemma B.4. If Hi has rank n − 2, this follows from (∗∗∗): if WE is inscribed ′ of ∂H , disjoint from X ∩ ∂H , and centered in a Euclidean open ball BE i i ′ is contained in a boundary component P outside ∂Hi ∩ N , then X ∩ ∂BE of N ∩ ∂Hi (an (n − 2)-dimensional Euclidean hyperplane by Lemma B.4) and WE is inscribed in another ball BE of ∂Hi , still disjoint from X, but centered at the projection of p to P .  In fact, (∗∗∗) also implies that the Euclidean Delaunay cellulation ∆P of P with respect to X ∩ P is contained in ∆∂Hi . Up to taking the points of X in generic position in P , in ∂Hi , and in Hn , we can make sure that all three Delaunay cellulations ∆P ⊂ ∆∂Hi ⊂ ∆ (where the last inclusion holds up to geodesic straigthening) are in fact triangulations. It follows from the comparison between hyperbolic and Euclidean Delaunay cellulations above that any geodesic ray escaping to the point at infinity ξi ∈ ∂∞ Hn of the cusp crosses the interface cells at most once. Therefore the thin-part cells form a star-shaped domain relative to ξi . We now modify ∆ by removing all thin-part simplices and coning the interface simplices of ∆∂Hi off to ξi . We repeat for each cusp (these operations do not interfere, since the distance between two horoballs Hi is larger than twice the diameter of any cell), and still denote by ∆ the resulting complex (see Figure 24): it is now finite modulo j(Γ0 ). To complete the proof of Claim B.5, we must check that the new complex ∆ is still convex. This is clear at the cusps of rank n − 1, since the corresponding Hi satisfy N ∩ Hi = Hi by Lemma B.4. At a cusp of rank n − 2, above the interface ∆∂Hi (which is convex in ∂Hi by the above discussion), the boundary of ∆ consists of two geodesic hyperplanes tangent at infinity (by Lemma B.4), and is therefore convex. At the boundary of ∆∂Hi , dihedral angles are convex because they already were before removal of the thin simplices. This completes the proof of Claim B.5. We now deduce Proposition B.3 from Claim B.5. As above, let H1 , . . . , Hc be horoballs of Hn , centered at points ξ1 , . . . , ξc ∈ ∂∞ Hn , whose images in j(Γ0 )\Hn are disjoint and intersect the convex core in standard cusp regions, representing all the cusps. Let p1 , . . . , pr ∈ Hn be orbit representatives of the vertices of ∆ lying in Hn . For 1 ≤ i ≤ c and k ∈ N∗ , let ξik ∈ ∂∞ Hn be the fixed point of jk (Si ), where Si is the stabilizer in Γ0 of ξi under j. Since converging parabolic elements have converging fixed points, (ξik )k∈N∗ converges to ξi for all 1 ≤ i ≤ c. We can thus find horoballs Hik centered at ξik that converge to Hi . Whenever the corresponding cusp has rank n − 2, the direction of the jk (Si )-invariant (n − 2)-planes in ∂Hik converges to the direction of the j(Si )-invariant (n − 2)-planes in ∂Hi . For 1 ≤ i ≤ r, we also choose a sequence (pki )k∈N∗ of points of Hn , converging to pi , such that if [j(γ) · pi , j(γ ′ ) · pi′ ] is a boundary edge of ∆∂Hi (such as [x0 , x1 ] in Figure 24), then jk (γ) · pki and jk (γ ′ ) · pki′ belong to a horocycle of ∂Hik contained in some jk (Si )-stable (n − 2)-plane of ∂Hik . (Inside each boundary component 116 FRANÇOIS GUÉRITAUD AND FANNY KASSEL ∂N x2 y1 x1 x0 y0 ∂Hi j(S ′ ) ∂∞ H3 Figure 24. The triangulation ∆ in a rank-1 cusp bounded by a horosphere ∂Hi centered at ∞ in the upper half-space model of H3 . At great height, the uniform neighborhood N of the convex core is bounded by just two oblique Euclidean planes. To simplify the picture, we have chosen X to intersect each boundary line P of N ∩ ∂Hi in only one j(S ′ )-orbit, (xs )s∈Z or (ys )s∈Z . Since j(S ′ ) is unipotent, the triangles (∞, xs , xs+1 ) are coplanar. In the center we showed a thickpart tetrahedron and a thin-part tetrahedron (after coning off) which share an interface triangle. of j(Si )\∆∂Hi , it is enough to enforce this condition over boundary edges that form a spanning tree in the quotient.) The simplices spanned by the jk (Γ0 ) · pki and jk (Γ0 ) · ξik , following the combinatorics of ∆, still locally form a triangulation for large k, because there are only finitely many orbits of simplices to check. It remains to check that the jk (Γ0 )-invariant collection ∆k of such simplices triangulates a convex region. This can be ensured locally, at every codimension-2 face W contained in the boundary of ∆. If W is compact, then the dihedral angle of ∆k at W goes to that of ∆, which is strictly convex. If W has an ideal vertex ξi , then ∂∆ is flat at W by Claim B.5, and ∂∆k is flat by choice of the pki . Therefore ∆k triangulates a convex region, which necessarily contains the convex core of jk (Γ0 )\Hn . In particular, jk (Γ0 ) is still geometrically finite for large k (in the absence of torsion, the quotient hyperbolic manifolds are in fact homeomorphic since their convex cores admit combinatorially identical triangulations). For the compact set C of Proposition B.3, we can take a neighborhood of a union of orbit representatives of the compact simplices of ∆. To bound injectivity radii away from 0, we argue as in the convex cocompact case, but in restriction to thick-part simplices only.  MAXIMALLY STRETCHED LAMINATIONS 117 Appendix C. Open questions Here we collect a few open questions, organized by themes; some of them were already raised in the core of the paper. C.1. General theory of extension of Lipschitz maps in H2 . Does there exist a function F : (0, 1) → (0, 1) such that for any compact subset K of H2 and any Lipschitz map ϕ : K → H2 with Lip(ϕ) < 1, there is an extension f : H2 → H2 of ϕ with Lip(f ) ≤ F (Lip(ϕ))? By controlling the sizes of the neigborhoods Up in the proofs of Proposition 3.7 and Lemma 3.8, it is possible to deal with the case where a bound on the diameter of K has been fixed a priori. An encouraging sign for unbounded K is that in Example 9.6, where K consists of three equidistant points, CK,ϕ (j, ρ) = Lip(ϕ) + o(1) as the diameter of K goes to infinity with Lip(ϕ) ∈ (0, 1) fixed. Fix a compact subset K of H2 and a Lipschitz map ϕ : K → H2 . Is it possible to find an extension f : H2 → H2 of ϕ with minimal Lipschitz constant CK,ϕ (j, ρ), which is optimal in the sense of Definition 4.12 and satisfies Lipp (f ) = Lipp (ϕ) for all points p ∈ K outside the relative stretch locus EK,ϕ (j, ρ)? Under the same assumptions, if C := CK,ϕ (j, ρ) < 1, is it true that for any p ∈ EK,ϕ (j, ρ) r K, there exists a point q 6= p such that [p, q] is C-stretched, i.e. d(f (p), f (q)) = Cd(p, q)? By definition of the relative stretch locus, some segments near p are nearly C-stretched, but it is not clear whether we can take p as an endpoint. C.2. Geometrically infinite representations j in dimension n = 3. Does Theorem 1.8 hold for finitely generated Γ0 but geometrically infinite j? To prove this in dimension 3, using the Ending Lamination Classification [BCM], one avenue would be to extend Theorem 1.3 in a way that somehow allows the stretch locus E(j, ρ) to be an ending lamination. One would also need to prove a good quantitative rigidity statement for infinite ends: at least, that if two geometrically infinite manifolds j(Γ0 )\H3 and j ′ (Γ0 )\H3 have a common ending lamination, then |µ(j(γk )) − µ(j ′ (γk ))| is bounded for some appropriate sequence (γk )k∈N of elements of Γ0 whose associated loops go deeper and deeper into the common end. (Here µ : G → R+ is the Cartan projection of (7.1).) C.3. Nonreductive representations ρ. For (j, ρ) ∈ Hom(Γ0 , G)2 with j geometrically finite and ρ reductive, we know (Lemma 4.10) that the infimum C(j, ρ) of Lipschitz constants for (j, ρ)-equivariant maps Hn → Hn is always achieved (i.e. F j,ρ 6= ∅). Is it still always achieved for nonreductive ρ when C(j, ρ) ≥ 1? When C(j, ρ) < 1, we know that it may or may not be achieved: see the examples in Sections 10.2 and 10.3. C.4. Behavior in the cusps for equivariant maps with minimal Lipschitz constant. Is there always a (j, ρ)-equivariant, C(j, ρ)-Lipschitz map that is constant in each deteriorating cusp? The answer is yes for C(j, ρ) ≥ 1 (Proposition 4.16), but for C(j, ρ) < 1 we do not even know if the stretch locus E(j, ρ) has a compact image in j(Γ0 )\Hn . If it does, then one might ask for a uniform bound: do Proposition 5.11 and Corollary 5.14 extend to C(j, ρ) < 1? 118 FRANÇOIS GUÉRITAUD AND FANNY KASSEL Suppose that C(j, ρ) = 1 and that ρ is not cusp-deteriorating. If the stretch locus E(j, ρ) is nonempty, does it contain a geodesic lamination whose image in j(Γ0 )\Hn is compact? C.5. Generalizing the Thurston metric. To what extent can the 2dimensional theory of the Thurston (asymmetric) metric on Teichmüller space be transposed to higher dimension? In particular, how do the two asymmetric metrics dTh and d′Th of Section 1.5 compare on the deformation space T (M ) of a geometrically finite hyperbolic manifold M ? The topology and geometry of T (M ), or of level sets of the critical exponent δ : T (M ) → (0, n − 1], seem difficult but interesting to study. Are any two points connected by a dTh -geodesic? Is there an analogue of stretch paths (particular geodesics introduced in [T2])? Is it possible to relate infinitesimal dTh -balls to the space of projective measured laminations as in [T2]? C.6. Chain recurrence of the stretch locus. In dimension n ≥ 3, when C(j, ρ) > 1, does the stretch locus E(j, ρ) have a chain-recurrence property as in Proposition 9.4, in the sense that any point in the geodesic lamination j(Γ0 )\E(j, ρ) sits on a closed quasi-leaf? Since there is no classification of geodesic laminations (Fact 9.2) available in higher dimension, quasi-leaves can be generalized in at least two ways: either with a bound ε → 0 on the total size of all jumps from one leaf to the next, or (weaker) on the size of each jump separately. It is not clear whether the two definitions coincide, even under constraints such as the conclusion of Lemma 5.13. In dimension n ≥ 2, does chain recurrence, suitably defined, extend to the convex strata of Lemma 5.4 when C(j, ρ) = 1? C.7. Semicontinuity of the stretch locus. Is the stretch locus map (j, ρ) 7→ E(j, ρ) upper semicontinuous for the Hausdorff topology when C(j, ρ) is arbitrary, in arbitrary dimension n? Proposition 9.5 answers this question affirmatively in dimension n = 2 for C(j, ρ) > 1; the case C(j, ρ) = 1 might allow for a proof along the same lines, using chain recurrence (suitably generalized as in Section C.6 above). C.8. Graminations. If C(j, ρ) < 1 and F j,ρ 6= ∅, is the stretch locus E(j, ρ) generically a trivalent geodesic tree (as in the example of Section 10.5)? Is it, in full generality, what in Conjecture 1.4 we called a gramination, namely the union of a closed discrete set F and of a lamination in the complement of F (with leaves possibly terminating on F )? References [AMS] [BBS] [Bea] [Be1] H. Abels, G. A. Margulis, G. A. Soifer, Semigroups containing proximal linear maps, Israel J. Math. 91 (1995), p. 1–30. W. Ballmann, M. Brin, R. Spatzier, Structure of manifolds of nonpositive curvature II, Ann. of Math. 122 (1985), p. 205–235. A. F. Beardon, The Hausdorff dimension of singular sets of properly discontinuous groups, Amer. J. Math. 88 (1966), p. 722–736. Y. Benoist, Actions propres sur les espaces homogènes réductifs, Ann. of Math. 144 (1996), p. 315–347. MAXIMALLY STRETCHED LAMINATIONS [Be2] [BiS] [BJ] [Bo1] [Bo2] [BH] [BCLS] [BCM] [Bu] [BuS] [DP] [DOP] [DGK] [D] [DZ] [Gh] [Go] [H] [Ka1] [Ka2] [Ka3] [Ka4] [KK] [Kim] 119 Y. Benoist, Propriétés asymptotiques des groupes linéaires, Geom. Funct. Anal. 7 (1997), p. 1–47. J. S. Birman, C. Series, Geodesics with bounded intersection number on surfaces are sparsely distributed, Topology 24 (1985), p. 217–225. C. J. Bishop, P. W. Jones, Hausdorff dimension and Kleinian groups, Acta Math. 179 (1997), p. 1–39. B. H. Bowditch, Geometrical finiteness for hyperbolic groups, J. Funct. Anal. 113 (1993), p. 245–317. B. H. Bowditch, Spaces of geometrically finite representations, Ann. Acad. Sci. Fenn. Math. 23 (1998), p. 389–414. M. R. Bridson, A. Haefliger, Metric spaces of non-positive curvature, Grundlehren der mathematischen Wissenschaften 319, Springer-Verlag, Berlin, 1999. M. Bridgeman, R. D. Canary, F. Labourie, A. Sambarino, The Pressure metric for Anosov representations, Geom. Funct. Anal. 25 (2015), p. 1089–1179. J. Brock, R. D. Canary, Y. N. Minsky, The classification of Kleinian surface groups, II: The Ending Lamination Conjecture, Ann. of Math. 176 (2012), p. 1–149. M. Burger, Intersection, the Manhattan curve, and Patterson–Sullivan theory in rank 2, Int. Math. Res. Not. 7 (1993), p. 217–225. S. Buyalo, V. Schröder, Extension of Lipschitz maps into 3-manifolds, Asian J. Math. 5 (2001), p. 685–704. F. Dal’Bo, M. Peigné, Groupes du ping-pong et géodésiques fermées en courbure −1, Ann. Inst. Fourier 46 (1996), p. 755–799. F. Dal’Bo, J.-P. Otal, M. Peigné, Séries de Poincaré des groupes géométriquement finis, Israel J. Math. 118 (2000), p. 109–124. J. Danciger, F. Guéritaud, F. Kassel, Geometry and topology of complete Lorentz spacetimes of constant curvature, Ann. Éc. Norm. Sup. 49 (2016), p. 1– 57. B. Delaunay, Sur la sphère vide, Izvestia Akademii Nauk SSSR, Otdelenie Matematicheskikh i Estestvennykh Nauk (1934), p. 793–800. S. Dumitrescu, A. Zeghib, Global rigidity of holomorphic Riemannian metrics on compact complex 3-manifolds, Math. Ann. 345 (2009), p. 53–81. É. Ghys, Déformations des structures complexes sur les espaces homogènes de SL2 (C), J. Reine Angew. Math. 468 (1995), p. 113–138. W. M. Goldman, Nonstandard Lorentz space forms, J. Differential Geom. 21 (1985), p. 301–308. S. Helgason, Differential geometry, Lie groups, and symmetric spaces, corrected reprint of the 1978 original, Graduate Studies in Mathematics 34, American Mathematical Society, Providence, RI, 2001. F. Kassel, Quotients compacts d’espaces homogènes réels ou p-adiques, PhD thesis, Université Paris-Sud 11, November 2009, available at http://math.univ-lille1.fr/∼kassel/. F. Kassel, Proper actions on corank-one reductive homogeneous spaces, J. Lie Theory 18 (2008), p. 961–978. F. Kassel, Quotients compacts des groupes ultramétriques de rang un, Ann. Inst. Fourier 60 (2010), p. 1741–1786. F. Kassel, Deformation of proper actions on reductive homogeneous spaces, Math. Ann. 353 (2012), p. 599–632. F. Kassel, T. Kobayashi, Poincaré series for non-Riemannian locally symmetric spaces, Adv. Math. 287 (2016), p. 123–236. I. Kim, Ergodic theory and rigidity on the symmetric spaces of non-compact type, Ergodic Theory Dynam. Systems 21 (2001), p. 93–114. 120 [Kir] [Kl] [Ko1] [Ko2] [Ko3] [Ko4] [KR] [L] [LPS] [LS] [LN] [Ma] [M1] [M2] [PT] [Par] [Pat] [Ro1] [Ro2] [Sa] [Se] [S1] [S2] [T1] [T2] FRANÇOIS GUÉRITAUD AND FANNY KASSEL M. D. Kirszbraun, Über die zusammenziehende und Lipschitzsche Transformationen, Fund. Math. 22 (1934), p. 77–108. B. Klingler, Complétude des variétés lorentziennes à courbure constante, Math. Ann. 306 (1996), p. 353–370. T. Kobayashi, Proper action on a homogeneous space of reductive type, Math. Ann. 285 (1989), p. 249–263. T. Kobayashi, On discontinuous groups acting on homogeneous spaces with noncompact isotropy subgroups, J. Geom. Phys. 12 (1993), p. 133–144. T. Kobayashi, Criterion for proper actions on homogeneous spaces of reductive groups, J. Lie Theory 6 (1996), p. 147–163. T. Kobayashi, Deformation of compact Clifford–Klein forms of indefiniteRiemannian homogeneous manifolds, Math. Ann. 310 (1998), p. 394–408. R. S. Kulkarni, F. Raymond, 3-dimensional Lorentz space-forms and Seifert fiber spaces, J. Differential Geom. 21 (1985), p. 231–268. S. P. Lalley, Renewal theorems in symbolic dynamics, with applications to geodesic flows, non-Euclidean tessellations and their fractal limits, Acta Math. 163 (1989), p. 1–55. U. Lang, B. Pavlović, V. Schröder, Extensions of Lipschitz maps into Hadamard spaces, Geom. Funct. Anal. 10 (2000), p. 1527–1553. U. Lang, V. Schröder, Kirszbraun’s Theorem and metric spaces of bounded curvature, Geom. Funct. Anal. 7 (1997), p. 535–560. J. R. Lee, A. Naor, Extending Lipschitz functions via random partitions, Invent. Math. 160 (2005), p. 59–95. A. Marden, The geometry of finitely generated kleinian groups, Ann. of Math. 99 (1974), p. 383–462. C. T. McMullen, Cusps are dense, Ann. of Math. 133 (1991), p. 217–247. C. T. McMullen, Hausdorff dimension and conformal dynamics. I. Strong convergence of Kleinian groups, J. Differential Geom. 51 (1999), p. 471–515. A. Papadopoulos, G. Théret, Shortening all the simple closed geodesics on surfaces with boundary, Proc. Amer. Math. Soc. 138 (2010), p. 1775–1784. A. Parreau, Dégénérescences de sous-groupes discrets de groupes de Lie semisimples et actions de groupes sur des immeubles affines, PhD thesis, Université Paris-Sud 11, January 2000, see http://www-fourier.ujf-grenoble.fr/∼parreau/travaux/these.pdf. S. J. Patterson, The limit set of a Fuchsian group, Acta Math. 136 (1976), p. 241–273. T. Roblin, Sur la fonction orbitale des groupes discrets en courbure négative, Ann. Inst. Fourier 52 (2002), p. 145–151. T. Roblin, Ergodicité et équidistribution en courbure négative, Mém. Soc. Math. Fr. 95 (2003). F. Salein, Variétés anti-de Sitter de dimension 3 exotiques, Ann. Inst. Fourier 50 (2000), p. 257–284. A. Selberg, On discontinuous groups in higher-dimensional symmetric spaces (1960), in “Collected papers”, vol. 1, p. 475–492, Springer-Verlag, Berlin, 1989. D. Sullivan, The density at infinity of a discrete group of hyperbolic motions, Publ. Math. Inst. Hautes Études Sci. 50 (1979), p. 419–450. D. Sullivan, Entropy, Hausdorff measures old and new, and limit sets of geometrically finite Kleinian groups, Acta Math. 153 (1984), p. 259–277. W. P. Thurston, The geometry and topology of 3-manifolds, Princeton University mimeographed notes, 1977, see http://library.msri.org/books/gt3m/. W. P. Thurston, Minimal stretch maps between hyperbolic surfaces, preprint (1986), arXiv:9801039. MAXIMALLY STRETCHED LAMINATIONS [V] [Z] 121 F. A. Valentine, Contractions in non-Euclidean spaces, Bull. Amer. Math. Soc. 50 (1944), p. 710–713. A. Zeghib, On closed anti-de Sitter spacetimes, Math. Ann. 310 (1998), p. 695– 716. CNRS and Université Lille 1, Laboratoire Paul Painlevé, 59655 Villeneuve d’Ascq Cedex, France E-mail address: [email protected] [email protected]
4math.GR
Under review as a conference paper at ICLR 2017 L EARNING P YTHON C ODE S UGGESTION S PARSE P OINTER N ETWORK WITH A arXiv:1611.08307v1 [cs.NE] 24 Nov 2016 Avishkar Bhoopchand, Tim Rocktäschel, Earl Barr & Sebastian Riedel Department of Computer Science University College London [email protected], {t.rocktaschel,e.barr,s.riedel}@cs.ucl.ac.uk A BSTRACT To enhance developer productivity, all modern integrated development environments (IDEs) include code suggestion functionality that proposes likely next tokens at the cursor. While current IDEs work well for statically-typed languages, their reliance on type annotations means that they do not provide the same level of support for dynamic programming languages as for statically-typed languages. Moreover, suggestion engines in modern IDEs do not propose expressions or multi-statement idiomatic code. Recent work has shown that language models can improve code suggestion systems by learning from software repositories. This paper introduces a neural language model with a sparse pointer network aimed at capturing very longrange dependencies. We release a large-scale code suggestion corpus of 41M lines of Python code crawled from GitHub. On this corpus, we found standard neural language models to perform well at suggesting local phenomena, but struggle to refer to identifiers that are introduced many tokens in the past. By augmenting a neural language model with a pointer network specialized in referring to predefined classes of identifiers, we obtain a much lower perplexity and a 5 percentage points increase in accuracy for code suggestion compared to an LSTM baseline. In fact, this increase in code suggestion accuracy is due to a 13 times more accurate prediction of identifiers. Furthermore, a qualitative analysis shows this model indeed captures interesting long-range dependencies, like referring to a class member defined over 60 tokens in the past. 1 I NTRODUCTION Integrated development environments (IDEs) are essential tools for programmers. Especially when a developer is new to a codebase, one of their most useful features is code suggestion: given a piece of code as context, suggest a likely sequence of next tokens. Typically, the IDE suggests an identifier or a function call, including API calls. While extensive support exists for statically-typed languages such as Java, code suggestion for dynamic languages like Python is harder and less well supported because of the lack of type annotations. Moreover, suggestion engines in modern IDEs do not propose expressions or multi-statement idiomatic code. Recently, methods from statistical natural language processing (NLP) have been used to train code suggestion systems from code usage in large code repositories (Hindle et al., 2012; Allamanis & Sutton, 2013; Tu et al., 2014). To this end, usually an n-gram language model is trained to score possible completions. Neural language models for code suggestion (White et al., 2015; Das & Shah, 2015) have extended this line of work to capture more long-range dependencies. Yet, these standard neural language models are limited by the so-called hidden state bottleneck, i.e., all context information has to be stored in a fixed-dimensional internal vector representation. This limitation restricts such models to local phenomena and does not capture very long-range semantic relationships like suggesting calling a function that has been defined many tokens before. To address these issues, we create a large corpus of 41M lines of Python code by using a heuristic for crawling high-quality code repositories from GitHub. We investigate, for the first time, the use of attention (Bahdanau et al., 2014) for code suggestion and find that, despite a substantial improvement 1 Under review as a conference paper at ICLR 2017 in accuracy, it still makes avoidable mistakes. Hence, we introduce a model that leverages long-range Python dependencies by selectively attending over the introduction of identifiers as determined by examining the Abstract Syntax Tree. The model is a form of pointer network (Vinyals et al., 2015a), and learns to dynamically choose between syntax-aware pointing for modeling long-range dependencies and free form generation to deal with local phenomena, based on the current context. Our contributions are threefold: (i) We release a code suggestion corpus of 41M lines of Python code crawled from GitHub, (ii) We introduce a sparse attention mechanism that captures very long-range dependencies for code suggestion of this dynamic programming language efficiently, and (iii) We provide a qualitative analysis demonstrating that this model is indeed able to learn such long-range dependencies. 2 M ETHODS We first revisit neural language models, before briefly describing how to extend such a language model with an attention mechanism. Then we introduce a sparse attention mechanism for a pointer network that can exploit the Python abstract syntax tree of the current context for code suggestion. 2.1 N EURAL L ANGUAGE M ODEL Code suggestion can be approached by a language model that measures the probability of observing a sequence of tokens in a Python program. For example, for the sequence S = a1 , . . . , aN , the joint probability of S factorizes according to Pθ (S) = Pθ (a1 ) · N Y Pθ (at | at−1 , . . . , a1 ) (1) t=2 where the parameters θ are estimated from a training corpus. Given a sequence of Python tokens, we seek to predict the next M tokens at+1 , . . . , at+M that maximize Equation 1 arg max Pθ (a1 , . . . , at , at+1 , . . . , at+M ). (2) at+1 , ..., at+M In this work, we build upon neural language models using Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM, Hochreiter & Schmidhuber, 1997). This neural language model estimates the probabilities in Equation 1 using the output vector of an LSTM at time step t (denoted ht here) according to exp (vτT ht + bτ ) T 0 τ 0 exp (vτ 0 ht + bτ ) Pθ (at = τ | at−1 , . . . , a1 ) = P (3) where vτ is a parameter vector associated with token τ in the vocabulary. Neural language models can, in theory, capture long-term dependencies in token sequences through their internal memory. However, as this internal memory has fixed dimension and can be updated at every time step, such models often only capture local phenomena. In contrast, we are interested in very long-range dependencies like referring to a function identifier introduced many tokens in the past. For example, a function identifier may be introduced at the top of a file and only used near the bottom. In the following, we investigate various external memory architectures for neural code suggestion. 2.2 ATTENTION A straight-forward approach to capturing long-range dependencies is to use a neural attention mechanism (Bahdanau et al., 2014) on the previous K output vectors of the language model. Attention mechanisms have been successfully applied to sequence-to-sequence tasks such as machine translation (Bahdanau et al., 2014), question-answering (Hermann et al., 2015), syntactic parsing (Vinyals et al., 2015b), as well as dual-sequence modeling like recognizing textual entailment (Rocktäschel et al., 2016). The idea is to overcome the hidden-state bottleneck by allowing referral back to previous output vectors. Recently, these mechanisms were applied to language modelling by Cheng et al. (2016) and Tran et al. (2016). 2 Under review as a conference paper at ICLR 2017 Formally, an attention mechanism with a fixed memory Mt ∈ Rk×K of K vectors mi ∈ Rk for i ∈ [1, K], produces an attention distribution αt ∈ RK and context vector ct ∈ Rk at each time step t according to Equations 4 to 7. Furthermore, W M , W h ∈ Rk×k and w ∈ Rk are trainable parameters. Finally, note that 1K represents a K-dimensional vector of ones. Mt = [m1 . . . mK ] Gt = tanh(W M Mt + 1TK (W h ht )) T αt = softmax(w Gt ) ct = Mt αTt ∈ Rk×K (4) ∈R k×K (5) ∈R 1×K (6) ∈R k (7) For language modeling, we populate Mt with a fixed window of the previous K LSTM output vectors. To obtain a distribution over the next token we combine the context vector ct of the attention mechanism with the output vector ht of the LSTM using a trainable projection matrix W A ∈ Rk×2k . The resulting final output vector nt ∈ Rk encodes the next-word distribution and is projected to the size of the vocabulary |V |. Subsequently, we apply a softmax to arrive at a probability distribution yt ∈ R|V | over the next token. This process is presented in Equation 9 where WV ∈ R|V |×k and bV ∈ R|V | are trainable parameters.    h nt = tanh W A t ∈ Rk (8) ct yt = softmax(W V nt + bV ) ∈ R|V | (9) The problem of the attention mechanism above is that it quickly becomes computationally expensive for large K. Moreover, attending over many memories can make training hard as a lot of noise is introduced in early stages of optimization where the LSTM outputs (and thus the memory Mt ) are more or less random. To alleviate these problems we now turn to pointer networks and a simple heuristic for populating Mt that permits the efficient retrieval of identifiers in a large history of Python code. 2.3 S PARSE P OINTER N ETWORK We develop an attention mechanism that provides a filtered view of a large history of Python tokens. At any given time step, the memory consists of context representations of the previous K identifiers introduced in the history. This allows us to model long-range dependencies found in identifier usage. For instance, a class identifier may be declared hundreds of lines of code before it is used. Given a history of Python tokens, we obtain a next-word distribution from a weighed average of the sparse pointer network for identifier reference and a standard neural language model. The weighting of the two is determined by a controller. Formally, at time-step t, the sparse pointer network operates on a memory Mt ∈ Rk×K of only the K previous identifier representations (e.g. function identifiers, class identifiers and so on). In addition, we maintain a vector mt = [id1 , . . . , idK ] ∈ NK of symbol ids for these identifier representations (i.e. pointers into the large global vocabulary). As before, we calculate a context vector ct using the attention mechanism (Equation 7), but on a memory Mt only containing representations of identifiers that were declared in the history. Next, we obtain a pseudo-sparse distribution over the global vocabulary from  αt [j] if mt [j] = i st [i] = (10) −C otherwise ∈ R|V | it = softmax(st ) (11) where −C is a large negative constant (e.g. −1000). In addition, we calculate a next-word distribution from a standard neural language model yt = softmax(W V ht + bV ) 3 ∈ R|V | (12) Under review as a conference paper at ICLR 2017 Figure 1: Sparse pointer network for code suggestion on a Python code snippet, showing the nextword distributions of the language model and identifier attention and their weighted combination through λ and we use a controller to calculate a distribution λt ∈ R2 over the language model and pointer network for the final weighted next-word distribution yt∗ via " # ht hλt = xt ∈ R3k (13) ct λt = softmax(W λ hλt + bλ ) yt∗ ∈ R2 ∈R = [yt it ] λt |V | (14) (15) Here, xt is the representation of the input token, and W λ ∈ R2×3k and bλ ∈ R2 a trainable weight matrix and bias respectively. This controller is conditioned on the input, output and context representations. This means for deciding whether to refer to an identifier or generate from the global vocabulary, the controller has access to information from the encoded next-word distribution ht of the standard neural language model, as well as the attention-weighted identifier representations ct from the current history. Figure 1 overviews this process. In it, the identifier base_path appears twice, once as an argument to a function and once as a member of a class (denoted by *). Each appearance has a different id in the vocabulary and obtains a different probability from the model. In the example, the model correctly chooses to refer to the member of the class instead of the out-of-scope function argument, although, from a user point-of-view, the suggestion would be the same in both cases. 3 L ARGE - SCALE P YTHON C ORPUS Previous work on code suggestion either focused on statically-typed languages (particularly Java) or trained on very small corpora. Thus, we decided to collect a new large-scale corpus of the dynamic programming language Python. According to the programming language popularity website Pypl (Carbonnelle, 2016), Python is the second most popular language after Java. It is also the 3rd most common language in terms of number of repositories on the open-source code repository GitHub, after JavaScript and Java (Zapponi, 2016). We collected a corpus of 41M lines of Python code from GitHub projects. Ideally, we would like this corpus to only contain high-quality Python code, as our language model learns to suggest code from how users write code. However, it is difficult to automatically assess what constitutes high-quality code. Thus, we resort to the heuristic that popular code projects tend to be of good quality, There are 4 Under review as a conference paper at ICLR 2017 Table 1: Python corpus statistics. Dataset #Projects #Files #Lines #Tokens Vocabulary Size Train Dev Test 489 179 281 118 298 26 466 43 062 26 868 583 5 804 826 8 398 100 88 935 698 18 147 341 30 178 356 2 323 819 Total 949 187 826 41 071 509 137 261 395 Figure 2: Example of the Python code normalization. Original file on the left and normalized version on the right. two metrics on GitHub that we can use for this purpose, namely stars (similar to bookmarks) and forks (copies of a repository that allow users to freely experiment with changes without affecting the original repository). Similar to Allamanis & Sutton (2013) and Allamanis et al. (2014), we select Python projects with more than 100 stars, sort by the number of forks descending, and take the top 1000 projects. We then removed projects that did not compile with Python3, leaving us with 949 projects. We split the corpus on the project level into train, dev, and test. Table 1 presents the corpus statistics. 3.1 N ORMALIZATION OF I DENTIFIERS Unsurprisingly, the long tail of words in the vocabulary consists of rare identifiers. To improve generalization, we normalize identifiers before feeding the resulting token stream to our models. That is, we replace every identifier name with an anonymous identifier indicating the identifier group (class, variable, argument, attribute or function) concatenated with a random number that makes the identifier unique in its scope. Note that we only replace novel identifiers defined within a file. Identifier references to external APIs and libraries are left untouched. Consistent with previous corpus creation for code suggestion (e.g. Khanh Dam et al., 2016; White et al., 2015), we replace numerical constant tokens with $NUM$, remove comments, reformat the code, and replace tokens appearing less than five times with an $OOV$ (out of vocabulary) token. 4 E XPERIMENTS Although previous work by White et al. (2015) already established that a simple neural language model outperforms an n-gram model for code suggestion, we include a number of n-gram baselines to confirm this observation. Specifically, we use n-gram models for n ∈ {3, 4, 5, 6} with Modified Kneser-Ney smoothing (Kneser & Ney, 1995) from the Kyoto Language Modelling Toolkit (Neubig, 2012). We train the sparse pointer network using mini-batch SGD with a batch size of 30 and truncated backpropagation through time (Werbos, 1990) with a history of 20 identifier representations. We use 5 Under review as a conference paper at ICLR 2017 Table 2: Perplexity (PP), Accuracy (Acc) and Accuarcy among top 5 predictions (Acc@5). Model Train PP 3-gram 4-gram 5-gram 6-gram Dev PP Test PP Acc [%] Acc@5 [%] All IDs Other All IDs Other 12.90 7.60 4.52 3.37 24.19 21.07 19.33 18.73 26.90 23.85 21.22 20.17 13.19 13.68 13.90 14.51 – – – – – – – – 50.81 51.26 51.49 51.76 – – – – – – – – LSTM LSTM w/ Attention 20 LSTM w/ Attention 50 9.29 7.30 7.09 13.08 11.07 9.83 14.01 11.74 10.05 57.91 61.30 63.21 2.1 21.4 30.2 62.8 64.8 65.3 76.30 79.32 81.69 4.5 29.9 41.3 82.6 83.7 84.1 Sparse Pointer Network 6.41 9.40 9.18 62.97 27.3 64.9 82.62 43.6 84.5 an initial learning rate of 0.7 and decay it by 0.9 after every epoch. As additional baselines, we test a neural language model with LSTM units with and without attention. For the attention language models, we experiment with a fixed-window attention memory of the previous 20 and 50 tokens respectively, and a batch size of 75. All neural language models were developed in TensorFlow (Abadi et al., 2016) and trained using cross-entropy loss. While processing a Python source code file, the last recurrent state of the RNN is fed as the initial state of the subsequent sequence of the same file and reset between files. All models use an input and hidden size of 200, an LSTM forget gate bias of 1 (Jozefowicz et al., 2015), gradient norm clipping of 5 (Pascanu et al., 2013), and randomly initialized parameters in the interval (−0.05, 0.05). As regularizer, we use a dropout of 0.1 on the input representations. Furthermore, we use a sampled softmax (Jean et al., 2015) with a log-uniform sampling distribution and a sample size of 1000. 5 R ESULTS We evaluate all models using perplexity (PP), as well as accuracy of the top prediction (Acc) and the top five predictions (Acc@5). The results are summarized in Table 2. We can confirm that for code suggestion neural models outperform n-gram language models by a large margin. Furthermore, adding attention improves the results substantially (2.3 lower perplexity and 3.4 percentage points increased accuracy). Interestingly, this increase can be attributed to a superior prediction of identifiers, which increased from an accuracy of 2.1% to 21.4%. An LSTM with an attention window of 50 gives us the best accuracy for the top prediction. We achieve further improvements for perplexity and accuracy of the top five predictions by using a sparse pointer network that uses a smaller memory of the past 20 identifier representations. 5.1 Q UALITATIVE A NALYSIS Figures 3a-d show a code suggestion example involving an identifier usage. While the LSTM baseline is uncertain about the next token, we get a sensible prediction by using attention or the sparse pointer network. The sparse pointer network provides more reasonable alternative suggestions beyond the correct top suggestion. Figures 3e-h show the use-case referring to a class attribute declared 67 tokens in the past. Only the Sparse Pointer Network makes a good suggestion. Furthermore, the attention weights in 3i demonstrate that this model distinguished attributes from other groups of identifiers. We give a full example of a token-by-token suggestion of the Sparse Pointer Network in Figure 4 in the Appendix. 6 R ELATED W ORK Previous code suggestion work using methods from statistical NLP has mostly focused on n-gram models. Much of this work is inspired by Hindle et al. (2012) who argued that real programs fall 6 Under review as a conference paper at ICLR 2017 (a) Code snippet for referencing variable. (b) LSTM Model. (c) LSTM w/ Attention (d) Sparse Pointer Net50. work. (e) Code snippet for referencing class member. (f) LSTM Model. (g) LSTM w/ Attention (h) Sparse Pointer Net50. work. (i) Sparse Pointer Network attention over memory of identifier representations. Figure 3: Code suggestion example involving a reference to a variable (a-d), a long-range dependency (e-h), and the attention weights of the Sparse Pointer Network (i). into a much smaller space than the flexibility of programming languages allows. They were able to capture the repetitiveness and predictable statistical properties of real programs using language models. Subsequently, Tu et al. (2014) improved upon Hindle et al.’s work by adding a cache mechanism that allowed them to exploit locality stemming from the specialisation and decoupling of program modules. Tu et al.’s idea of adding a cache mechanism to the language model is specifically designed to exploit the properties of source code, and thus follows the same aim as the sparse attention mechanism introduced in this paper. While the majority of preceding work trained on small corpora, Allamanis & Sutton (2013) created a corpus of 352M lines of Java code which they analysed with n-gram language models. The size of the corpus allowed them to train a single language model that was effective across multiple different project domains. White et al. (2015) later demonstrated that neural language models outperform n-gram models for code suggestion. They compared various n-gram models (up to nine grams), including Tu et al.’s cache model, with a basic RNN neural language model. Khanh Dam et al. (2016) compared White et al.’s basic RNN with LSTMs and found that the latter are better at code suggestion due to their improved ability to learn long-range dependencies found in source code. Our paper extends this line of work by introducing a sparse attention model that captures even longer dependencies. The combination of lagged attention mechanisms with language modelling is inspired by Cheng et al. (2016) who equipped LSTM cells with a fixed-length memory tape rather than a single memory cell. They achieved promising results on the standard Penn Treebank benchmark corpus (Marcus et al., 1993). Similarly, Tran et al. added a memory block to LSTMs for language modelling of English, German and Italian and outperformed both n-gram and neural language models. Their memory encompasses representations of all possible words in the vocabulary rather than providing a sparse view as we do. An alternative to our purely lexical approach to code suggestion involves the use of probabilistic context-free grammars (PCFGs) which exploit the formal grammar specifications and well-defined, deterministic parsers available for source code. These were used by Allamanis & Sutton (2014) to extract idiomatic patterns from source code. A weakness of PCFGs is their inability to model context-dependent rules of programming languages such as that variables need to be declared before 7 Under review as a conference paper at ICLR 2017 being used. Maddison & Tarlow (2014) added context-aware variables to their PCFG model in order to capture such rules. Ling et al. (2016) recently used a pointer network to generate code from natural language descriptions. Our use of a controller for deciding whether to generate from a language model or copy an identifier using a sparse pointer network is inspired by their latent code predictor. However, their inputs (textual descriptions) are short whereas code suggestion requires capturing very long-range dependencies that we addressed by a filtered view on the memory of previous identifier representations. 7 C ONCLUSIONS AND F UTURE W ORK In this paper, we investigated neural language models for code suggestion of the dynamically-typed programming language Python. We released a corpus of 41M lines of Python crawled from GitHub and compared n-gram, standard neural language models, and attention. By using attention, we observed an order of magnitude more accurate prediction of identifiers. Furthermore, we proposed a sparse pointer network that can efficiently capture long-range dependencies by only operating on a filtered view of a memory of previous identifier representations. This model achieves the lowest perplexity and best accuracy among the top five predictions. The Python corpus and code for replicating our experiment is released at https://github.com/uclmr/pycodesuggest. The presented methods were only tested for code suggestion within the same Python file. We are interested in scaling the approach to the level of entire code projects and collections thereof, as well as integrating a trained code suggestion model into an existing IDE. Furthermore, we plan to work on code completion, i.e., models that provide a likely continuation of a partial token, using character language models (Graves, 2013). ACKNOWLEDGMENTS This work was supported by Microsoft Research through its PhD Scholarship Programme, an Allen Distinguished Investigator Award, and a Marie Curie Career Integration Award. R EFERENCES Martı́n Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving, Michael Isard, Manjunath Kudlur, Josh Levenberg, Rajat Monga, Sherry Moore, Derek Gordon Murray, Benoit Steiner, Paul A. Tucker, Vijay Vasudevan, Pete Warden, Martin Wicke, Yuan Yu, and Xiaoqiang Zhang. Tensorflow: A system for large-scale machine learning. CoRR, abs/1605.08695, 2016. URL http://arxiv.org/abs/1605. 08695. Miltiadis Allamanis and Charles Sutton. Mining idioms from source code. In Proceedings of the 22Nd ACM SIGSOFT International Symposium on Foundations of Software Engineering, FSE 2014, pp. 472–483, New York, NY, USA, 2014. ACM. ISBN 978-1-4503-3056-5. doi: 10.1145/2635868.2635901. URL http://doi.acm.org/10.1145/2635868.2635901. Miltiadis Allamanis and Charles A. Sutton. Mining source code repositories at massive scale using language modeling. In Thomas Zimmermann, Massimiliano Di Penta, and Sunghun Kim (eds.), MSR, pp. 207–216. IEEE Computer Society, 2013. ISBN 978-1-4673-2936-1. URL http://dblp.uni-trier.de/db/conf/msr/msr2013.html#AllamanisS13a. Miltiadis Allamanis, Earl T. Barr, Christian Bird, and Charles Sutton. Learning natural coding conventions. In Proceedings of the 22Nd ACM SIGSOFT International Symposium on Foundations of Software Engineering, FSE 2014, pp. 281–293, New York, NY, USA, 2014. ACM. ISBN 9781-4503-3056-5. doi: 10.1145/2635868.2635883. URL http://doi.acm.org/10.1145/ 2635868.2635883. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. CoRR, abs/1409.0473, 2014. URL http://arxiv.org/abs/ 1409.0473. 8 Under review as a conference paper at ICLR 2017 Pierre Carbonnelle. Pypl popularity of programming language. http://pypl.github.io/ PYPL.html, 2016. URL http://pypl.github.io/PYPL.html. [Online; accessed 30August-2016]. Jianpeng Cheng, Li Dong, and Mirella Lapata. Long short-term memory-networks for machine reading. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. 551–561. Association for Computational Linguistics, 2016. URL http:// aclweb.org/anthology/D16-1053. Subhasis Das and Chinmayee Shah. Contextual code completion using machine learning. 2015. Alex Graves. Generating sequences with recurrent neural networks. CoRR, abs/1308.0850, 2013. URL http://arxiv.org/abs/1308.0850. Karl Moritz Hermann, Tomás Kociský, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pp. 1693–1701, 2015. URL http://papers.nips.cc/paper/ 5945-teaching-machines-to-read-and-comprehend. Abram Hindle, Earl T. Barr, Zhendong Su, Mark Gabel, and Premkumar Devanbu. On the naturalness of software. In Proceedings of the 34th International Conference on Software Engineering, ICSE ’12, pp. 837–847, Piscataway, NJ, USA, 2012. IEEE Press. ISBN 978-1-4673-1067-3. URL http://dl.acm.org/citation.cfm?id=2337223.2337322. Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735– 1780, November 1997. ISSN 0899-7667. doi: 10.1162/neco.1997.9.8.1735. URL http://dx. doi.org/10.1162/neco.1997.9.8.1735. Sébastien Jean, Kyunghyun Cho, Roland Memisevic, and Yoshua Bengio. On using very large target vocabulary for neural machine translation. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp. 1–10, Beijing, China, July 2015. Association for Computational Linguistics. URL http://www.aclweb.org/anthology/P15-1001. Rafal Jozefowicz, Wojciech Zaremba, and Ilya Sutskever. An empirical exploration of recurrent network architectures. In David Blei and Francis Bach (eds.), Proceedings of the 32nd International Conference on Machine Learning (ICML-15), pp. 2342–2350. JMLR Workshop and Conference Proceedings, 2015. URL http://jmlr.org/proceedings/papers/v37/ jozefowicz15.pdf. H. Khanh Dam, T. Tran, and T. Pham. A deep language model for software code. ArXiv e-prints, August 2016. R. Kneser and H. Ney. Improved backing-off for m-gram language modeling. In Acoustics, Speech, and Signal Processing, 1995. ICASSP-95., 1995 International Conference on, volume 1, pp. 181–184 vol.1, May 1995. doi: 10.1109/ICASSP.1995.479394. Wang Ling, Edward Grefenstette, Karl Moritz Hermann, Tomas Kocisky, Andrew Senior, Fumin Wang, and Phil Blunsom. Latent predictor networks for code generation. arXiv preprint arXiv:1603.06744, 2016. Chris J Maddison and Daniel Tarlow. Structured generative models of natural source code. In International Conference on Machine Learning, 2014. Mitchell P. Marcus, Beatrice Santorini, and Mary Ann Marcinkiewicz. Building a large annotated corpus of english: The penn treebank. COMPUTATIONAL LINGUISTICS, 19(2):313–330, 1993. Graham Neubig. Kylm - the kyoto language modeling toolkit. http://www.phontron.com/ kylm/, 2012. URL http://www.phontron.com/kylm/. [Online; accessed 23-July2016]. 9 Under review as a conference paper at ICLR 2017 Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. On the difficulty of training recurrent neural networks. In Proceedings of the 30th International Conference on Machine Learning, ICML 2013, Atlanta, GA, USA, 16-21 June 2013, pp. 1310–1318, 2013. URL http://jmlr.org/ proceedings/papers/v28/pascanu13.html. Tim Rocktäschel, Edward Grefenstette, Karl Moritz Hermann, Tomas Kocisky, and Phil Blunsom. Reasoning about entailment with neural attention. In ICLR, 2016. Ke M. Tran, Arianna Bisazza, and Christof Monz. Recurrent memory networks for language modeling. In NAACL HLT 2016, The 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, San Diego California, USA, June 12-17, 2016, pp. 321–331, 2016. URL http://aclweb.org/anthology/N/ N16/N16-1036.pdf. Zhaopeng Tu, Zhendong Su, and Premkumar Devanbu. On the localness of software. In Proceedings of the 22Nd ACM SIGSOFT International Symposium on Foundations of Software Engineering, FSE 2014, pp. 269–280, New York, NY, USA, 2014. ACM. ISBN 978-1-4503-3056-5. doi: 10.1145/2635868.2635875. URL http://doi.acm.org/10.1145/2635868.2635875. Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. Pointer networks. In Advances in Neural Information Processing Systems, pp. 2692–2700, 2015a. Oriol Vinyals, Lukasz Kaiser, Terry Koo, Slav Petrov, Ilya Sutskever, and Geoffrey E. Hinton. Grammar as a foreign language. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pp. 2773–2781, 2015b. URL http://papers.nips.cc/paper/ 5635-grammar-as-a-foreign-language. Paul J Werbos. Backpropagation through time: what it does and how to do it. Proceedings of the IEEE, 78(10):1550–1560, 1990. Martin White, Christopher Vendome, Mario Linares-Vásquez, and Denys Poshyvanyk. Toward deep learning software repositories. In Proceedings of the 12th Working Conference on Mining Software Repositories, MSR ’15, pp. 334–345, Piscataway, NJ, USA, 2015. IEEE Press. URL http://dl.acm.org/citation.cfm?id=2820518.2820559. Carlo Zapponi. Githut - programming languages and github. http://githut.info/, 2016. URL http://githut.info/. [Online; accessed 19-August-2016]. 10 Under review as a conference paper at ICLR 2017 A PPENDIX Figure 4: Full example of code suggestion with a Sparse Pointer Network. Boldface tokens on the left show the first declaration of an identifier. The middle part visualizes the memory of representations of these identifiers. The right part visualizes the output λ of the controller, which is used for interpolating between the language model (LM) and the attention of the pointer network (Att). 11
9cs.NE
GENE ONTOLOGY (GO) PREDICTION USING MACHINE LEARNING METHODS HAOZE WU * Department of Mathematics and Computer Science, Davidson College Box 5996, Davidson, North Carolina, 28035, USA [email protected] YANGYU ZHOU *1 Biology Department, Davidson College Box 6891, Davidson, North Carolina, 28035, USA [email protected] We applied machine learning to predict whether a gene is involved in axon regeneration. We extracted 31 features from different databases and trained five machine learning models. Our optimal model, a Random Forest Classifier with 50 submodels, yielded a test score of 85.71%, which is 4.1% higher than the baseline score. We concluded that our models have some predictive capability. Similar methodology and features could be applied to predict other Gene Ontology (GO) terms. Keywords: Gene Ontology; Machine Learning; Axon Regeneration 1. Introduction Demand for bioinformatics methods to analyze genomic information has never been greater. Researchers are particularly interested in the overall functions of genes. Despite the best curation efforts, functions of genes remain incomplete and biased toward much-studied genes and pathways. Machine learning methods have been extensively used to identify functional elements of a gene, such as transcription start sites, promoters, and transcription factor binding sites 1. However, few studies have applied machine learning to predict whether a gene has certain function, that is, whether a gene is involved in a particular biological process. * HaoZe Wu and YangYu Zhou are co-authors. As a proof-of-concept, in this paper, we aim to predict whether a gene is involved in axon regeneration in C. elegans. To establish the dataset, we find definitively positive or negative samples from literature and engineer features with the information provided with several on-line databases. We train five classifiers and assess the performance of each one of them. In the following sections, we will first briefly introduce some biology background and show how we extract features and establish the dataset. After that, we will present our methodology of training and evaluating different machine learning models. Finally, we will present and discuss our results. 2. BACKGROUND As a proof-of-concept, we chose to test the performance of our tool on axon regeneration (GO:0031103) in model organism Caenorhabditis elegans genome. We chose to analyze the genome of C. elegans, because of its well-annotated genome and vast amount of knowledge available in the scientific community. Gene ontology associates a gene and the biological function of the gene's product. There are 24500+ gene ontology terms, each term describes a biological function. The structure of GO terms resembles a directed acyclic graph (DAG), with the most general term (biological process, GO: 0008150) at the root of the graph, and the most specific terms (including the one we are concerned with, axon regeneration, GO:0031103) at the leaves. Damaged axons can sometime regenerate, restoring connections to target tissues after injury or disease. However, axon regeneration is initiated by intrinsic and extrinsic signals at the site of injury, such as neuron type, age, type of surrounding tissues and more. It is thought that the heterogeneity arises partially from regulation at the genetic level, including methylation of genome and chromatin structure. Transcription of certain key genes and synthesis of proteins are also required for successful axon regeneration. To achieve the highest predictive capabilities, eleven features were selected (Table 1). Table 1. Features, and their sources, used as input to the machine learning models. # 1 Feature Gene length 2 RNA-seq expression Mean GO distance 3 4 Minimum GO distance 5 ChIP-seq score 6 Avg. Sequence conservation Std dev Sequence conservation Ratio of positive interacting proteins Number of positive interacting proteins Phenotype association GO enriched terms 7 8 9 10 11 Description The length of a gene (in basepair) starting from its transcription start site (TSS) Gene expression level under standard conditions Database UCSC modENCODE The mean distance of all GO terms of a gene to GO:0031103 (axon regeneration) in the graph representation of all GO terms The minimum distance among the distances between a gene’s GO terms and GO: 0031103 (axon regeneration). ChIP-seq expression profile on young adult worms. Average conservation score of 15% length of the gene upstream of its TSS Standard deviation of score of 15% length of the gene upstream of its TSS AmiGo2 WormBase Among the proteins that interact with a gene, what ratio of them are products of positive genes. STRING Among the proteins that interact with a gene, how many of them are products of positive genes. STRING Whether a gene has the phenotypes that are most common in the dataset Whether a gene has the GO terms that are most common in the dataset. STRING AmiGo2 WormBase WormBase WormBase WormBase WormBase 2.1. Sequence conservation Conserved sequences are identical DNA sequences across different species. A region of high conservation score indicates that this segment of DNA is highly identical among species, suggesting that it may encode proteins of important biological functions so that it has been protected throughout evolution despite speciation. 2.2. Protein-protein interaction Proteins often interact with other proteins when performing biological functions such as signal transduction and cellular metabolism. Multiple transcription factors often act together when regulating gene expressions. 2.3. Phenotype association Phenotype is the observable characteristics of an organism, such as morphology and behavior. The phenotype largely results from the genotype, the expression of genes. In this paper, the phenotype of a gene refers to the abnormal changes when the gene is deleted. 2.4. ChIP-seq Chromatin Immunoprecipitation (ChIP) in combination with massive parallel DNA sequencing is a technique that identifies the binding sites of DNA binding proteins. Promoters of genes usually are high in ChIP-seq score due to high levels of transcription factor binding. 2.5. RNA-seq RNA sequencing quantifies the levels of transcriptions of genes genome-wide. Similar to ChIP-seq data, RNA-seq measures how active the gene is at a given time. We used RNA-seq data of young adult worms living at standard conditions. 3. Data Preprocessing Our samples are different genes, and our target variable is a boolean value that represents whether a gene is involved in axon regeneration. To determine whether a gene is involved in axon regeneration, we searched literature for definitive positive examples (genes that are for sure involved in axon regeneration) and negative examples (for sure not involved). We found three papers that describe 347 positive genes and 1401 negative genes in total2,3,4. Since most of the datasets have different formats, we wrote programs to extract useful information and compiled them into a structured format. Here we will briefly introduce the algorithms we used to extract some features and make further clarification on some features. 3.1. GO distance Each gene has several gene ontology terms. To get the distance between each of a gene's GO terms to the GO term GO:0031103 (axon regeneration), we first created a graph representation of all GO terms that captured the relationships between them, and then obtained the shortest distance between two GO terms with the Dijkstra's algorithm. 3.2. Protein interaction We first constructed a graph representation of protein interactions. Each vertex constitutes a protein, and each edge xy constitutes an interaction between protein x and y. Given a gene, we searched its neighborhood and looked for the neighbors that are protein products of the known positive examples. The data we obtained from different databases were well-established, so that we did not have a lot of missing values. We imputed missing values with mean imputation. Then, we randomly chose 20% of the dataset (350 samples) as the testing set, and used the rest 80% (1398 samples) for training and validation. From now on in this paper, for sake of convenience, we will refer to the 80% samples for training and cross-validation as the ``dataset." We will refer to the rest 20% samples as the ``testing set," as usual. Finally, we mean-normalized the dataset and the testing set. 4. Experiment We trained five different classifiers, K-nearest Neighbor (K-NN), Random Forest (RF), Logistic Regression, Support Vector Machine (SVC), and Neural Network (NN). The Neural Network model we used was a single-layer feedforward Neural Network Classifier. We used the Python Scikit-learn implementations of these models scikit-learn 5. We evaluated a model primarily based on its accuracy. In this paper, we will use the words ``score" and ``accuracy" interchangeably. We will use ``test score" to refer to the accuracy of a model on the test set; and we will use ``validation score" to refer to the mean cross-validation accuracy of a model. Besides the scores, we were also interested in the models' performance on predicting positive samples. Therefore, we recorded the precision and recall on the test set for each model. For each classifier, we tuned several hyperparameters. For each hyperparameter, we conducted a grid search for the optimal value. Table 2 shows the hyperparameters we tuned for each model, and the ranges of the grid searches. Table 2. Hyperparameters tuned for different models. Model Logistic K-NN RF SVC NN - Hyperparameters and Range of Grid Search Penalty parameter {0.001, 0.01, 0.1, 1, 10} # of neighbors (1-13 with step size 2) Distance metric {Euclidean, manhattan} Weight function {uniform, distance} # of components (10 – 100 with step size 10) Kernel {linear, rbf, sigmoid, 2nd poly} Penalty parameter {0.001, 0.01, 0.1, 1, 10} # of perceptrons (2 – 16 with step size 2) Penalty parameter {0.0001, 0.001, 0.01, 0.1, 1} Learning rate {adaptive, constant} Weight optimization algorithm {l-bfgs, sgd, adam} Besides, for K-NN, SVC, and Logistic, we tried to use Adaboost to enhance their performance. We also tried to enhance the last two classifiers with Bagging. For each model enhanced with these two meta-algorithms, we tuned the number of sub-classifiers of Adaboost or Bagging, with a grid search from 10 to 50, step size 10. We decided not to use Bagging to enhance K-NN because that would create similar effect to a weighted version of K-NN, while we would tune the weight function as a hyperparameter for the K-NN classifier. To compare models of a classifier with different values of hyperparameters, we conducted cross-validation. However, due to the small size of our dataset, we wanted to maximize the size of the training set while receiving stable validation scores. Therefore, we used leave-one-out cross-validation, that is, for each sample in the dataset, we trained a model with the rest of the samples, and predicted the target variable of that sample. In this case, the validation set is the entire dataset, and the cross-validation score would be the mean of the accuracies of all predictions (either 100% or 0%). For each optimized classifier, we reported its validation score, test score, (test) precision, and (test) recall. We also compared the validation scores of the five models horizontally, and chose the classifier with the highest score on the validation set as the model that has the highest predictive capability. To test whether our models have predictive capability, we also compared our models with the baseline model. Our training set contained 79.9% negative samples, and our test set contained 81.1% negative samples. If a model always predicted false, it would have a validation score of 79.9%, and a test score of 81.1%. Therefore, we decided that to outperform the baseline model, a model must have a validation score of more than 79.9% and a test score of more than 81.1%. 5. Results In this section, we introduce the training results of the five models in the order of their performances (from low to high). Then we compare the five models horizontally. 5.1. Logistic Regression We tuned the penalty parameter C for the Logistic Regressor and found that when C = 1, the model had the highest validation score. The model has a validation score of 79.01%, which is worse than the validation score of the baseline model. Therefore, we considered the Logistic Regressor a weak learner and used Adaboost to enhance its result. As shown by figure 1, as the number of sub-models increased, there was no significant improvement of the validation score. The model with 30 sub-models has the highest validation score. And this model has a test score of 80.57%, a precision of 42.86% and a recall of 9.09%. Though Adaboost did improve the accuracy of the Logistic Regressor, the boosted test score is still lower than the baseline. Fig. 1. Validation score vs. number of sub-models for Adaboost (Logisitic, C =1). On the other hand, Bagging led to even less improvement. Our best Bagging Logistic Regressor contains 40 sub-models, and has a validation score of 79.28%. This minimal improvement suggests that our original model did not have high variation. This is supported by figure 2 which shows that as the number of training examples increased, the training scores and the validation scores came close to each other. This means our model did not have a high variation. Fig 2. Learning Curve (Logisitic, C = 1). None of our Logistic Regression model has a test score higher than the baseline model. Therefore, we concluded that Logistic Regressor does not have predictive capability. 5.2. Neural Network (NN) We discovered that the optimal learning rate was "adaptive," and the optimal penalty parameter was 0.01. Figure 3 shows the validation scores of Neural Network with different weight optimization algorithms and numbers of perceptrons. The validation scores generally increased as the number of perceptrons increased. ``l-bfgs" and ``adam" yields comparable optimal results, while ``adam" is slightly better. Our best Neural Network model uses ``adam" as the weight optimization algorithm, and has 12 perceptrons in the hidden layer. It yielded a test score of 83.14%, a precision of 61.29%, and recall of 28.79%. Fig. 3. Cross-validating the number of perceptrons and the weight optimization algorithm for NN. 5.3. Support Vector Classifier (SVC) Figure 4 shows the validation scores of SVC models with different kernels and penalty parameters. As the penalty parameter increased, the validation scores increased for models using the ``rbf" and the ``2nd degree poly" kernels. The models using the ``sigmoid" kernel always yielded the baseline accuracy, 79.9%. The models using the ``linear" kernel yielded results worse than the baseline accuracy. Fig. 4. Cross-validating the kernel and penalty parameter for SVC. We tried to use Bagging and Adaboost to enhance the SVC, however, we received worse validation scores. We discovered that the optimal penalty parameter of SVC is 1, and the optimal kernel is ``rbf." Our optimal SVC model has a test score of 84.00%, a precision of 85.71%, and a recall of 18.18%. Fig. 5. Learning Curve (SVC, ‘rbf’ kernel, C = 1). 5.4. K-nearest Neighbors (K-NN) As shown by figure 6, models that used manhattan distance metric, and weighed the neighbors using the ``distance" weight function (weighing neighbors by the inverse of their distances to the input instance), were strictly better than the other models. On the other hand, we observed that the validation scores for manhattan distance models, and Euclidean distance models continued to increase as the number of neighbors $n$ increased, and there was no sign of convergence. Thus, we hypothesized that we had not found the optimal values of n for those two types of models and decided to cross-validate n by conducting a grid search with a wider range. Fig. 6. Validation scores of K-NN. Figure 7 supports our hypothesis. For the Euclidean distance models, the validation scores increased after n > 13 and started to decrease when n ≥ 17. For the manhattan distance models, the validation scores fluctuated and showed signs of convergence after n > 13. We discovered that the optimal K-NN classifier is a manhattan distance model that searches for 21 neighbors. This model yielded a validation score of 84.26%, a test score of 85.14%, a precision of 100%, and a recall of 21.21%. Fig. 7. Further cross-validation on the number of neighbors. We observed that for the unweighted K-NN models, the optimal value of n was only 3, which was significantly smaller than that for the weighted K-NN models. We hypothesized that weighted K-NN models correctly captured the characteristics of the dataset that predicative information was available in a fairly large neighborhood of an input instance, while the closer a neighbor was, the more important information it could provide. We tried to use Adaboost to enhance the K-NN, but there was no significant improvement. 5.5. Random Forest (RF) As shown by figure 8, the optimal number of sub-models n for Random Forest is 50. Normally, as n increases, the accuracy increases until convergence. However, this was not the case for us, the model's performance improved gradually when n ≤ 50, but when n > 50, the performance dropped and started to fluctuate. Fig. 8. Validation score vs. number of sub-models for Random Forest. Our best Random Forest model, with n = 50, yielded a test score of 85.71%, precision of 83.33%, and a recall of 30.30%. 5.6. Overall Analysis In this subsection, we first compare the performance of different classifiers; then we talk about some unrelated observations and hypotheses of our experiments, including • • The inefficiency of Adaboost; The insufficiency of training samples; • • The precision/recall trade-off; The decision boundary between positive and negative samples. Table 3. Comparison between different types of classifiers. Model Logistic NN SVC K-NN RF Validation Score 79.68% 82.19% 82.98% 84.26% 85.69% Test Score 80.57% 83.14% 84.00% 85.14% 85.71% Precision 42.86% 61.29% 85.71% 100% 83.33% Recall 9.09% 28.79% 18.18% 21.21% 30.30% Table 3 shows the validation score, test score, test precision, and test recall of the optimal model for each type of classifier. Except for the Logistic Regressor, all models outperformed the baseline model. The Random Forest Classifier had the best performance, and K-NN was the second best model. The former yielded a test score of 85.71%, and the latter 85.14%, which are both higher than the baseline test score 81.1%. The fact that most of our models outperformed the baseline model suggests that some features have certain predictive capability. However, more predicative features are needed in order to achieve a higher performance, as our best model only outperformed the baseline model by 4.6%. We observed that Adaboost generally resulted in marginal improvement or even mild decrease in the validation scores. We believed that this can be explained by the high dimensionality of the feature space, and the small size of our dataset. Adaboost is a biasincreasing method which theoretically produces weak models that barely outperform the baseline model. However, a very small dataset and a high-dimensional feature space might worsen the overfitting problem and result in sub-models with accuracies lower than the baseline. We believed that the small size of training set might also lead to sub-optimally trained models. As shown by figure 5, the validation score curve of SVC does not seem to converge when the size of the training set is about 1400. It is likely that if we acquired more training examples, we would be able to obtain models with higher scores. In terms of precision and recall, all of our models tended to have high precisions and low recalls. Intuitively, this suggests that a model would decide that a sample is positive only when the model is very ``confident" about it. We believed this high confidence threshold can be explained by the unbalanced nature of our dataset, where the models had more incentives to predict negative than predict positive. Finally, the different behaviors of different classifiers shed light on the decision boundary between the negative and positive samples in the feature space. Logistic Regression, and SVC with linear kernel yielded results worse than the baseline. On the other hand, Random Forest, K-NN, and SVC with 'rbf' kernel yielded the highest accuracy. Based on this observation, we hypothesized that the decision boundary is not linear. 6. Conclusion We started out aiming to create a classifier that predicts whether a gene is involved in axon regeneration in C. elegans. We found that a Random Forest Classifier with 50 subestimators has the highest predicative capability. It yielded a test score of 85.71%, a precision of 83.33%, and a recall of 30.30%. Since the test score is 4.6% higher than the baseline test score, we decided that our model has certain predicative capability and some of the features we extracted were effective. We also believed that similar methodologies in extracting gene features as ours could be applied to the prediction of whether a gene is involved in other biological processes, besides axon regeneration. To improve predicative capability of our models, we would 1) increase sample size through literature search and lab works, 2) extract more predicative features. References 1. 2. 3. 4. 5. Libbrecht, M. W., and Noble, W. S. 2015. Machine learning applications in genetics and genomics. Nature Reviews Genetics 16:321–332. Bejjani, R. E., and Marc, H. 2012. Neural regeneration in c. elegans. Annual Review of Genetics 46:499–513. Lizhen Chen, Zhiping Wang, A. G.-R. T. H. D. Y. S. O. R. B. B. Z. W. Y. J., and Chisholm, A. D. 2011. Axon regener- ation pathways identified by systematic genetic screening in c. elegans. Neuron 71(6):1043–1057. Paola Nix, Marc Hammarlund, L. H.-M. L. E. M. J., and Bastiani, M. 2014. Axon regeneration genes identified by RNAi screening in c. elegans. The Journal of Neuroscience 34(2):499–513. Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; Vanderplas, J.; Passos, A.; Cournapeau, D.; Brucher, M.; Perrot, M.; and Duchesnay, E. 2011. Scikit- learn: Machine learning in Python. Journal of Machine Learning Research 12:2825–2830.
5cs.CE
arXiv:1508.07592v1 [math.AC] 30 Aug 2015 THE LINEAR STRAND OF DETERMINANTAL FACET IDEALS JÜRGEN HERZOG, DARIUSH KIANI AND SARA SAEEDI MADANI Abstract. Let X be an (m × n)-matrix of indeterminates, and let J be the ideal generated by a set S of maximal minors of X. We construct the linear strand of the resolution of J. This linear strand is determined by the clique complex of the m-clutter corresponding to the set S. As a consequence one obtains explicit formulas for the graded Betti numbers βi,i+m (J) for all i ≥ 0. We also determine all sets S for which J has a linear resolution. Introduction In this paper we consider ideals generated by an arbitrary set of maximal minors of an (m × n)-matrix of indeterminates. Such ideals have first been considered when m = 2, in which case the set of minors (which is a set of binomials) is in bijection with the edges of a graph G, and therefore is called the binomial edge ideal of G. This class of ideals has first been considered in [13] and [21]. In [13] the relevance of such ideals for algebraic statistics has been stressed. In the sequel binomial edge ideals have been studied in numerous papers with an attempt to better understand their algebraic and homological properties, see for example [8], [9], [15], [24] and [25]. In some special cases, the resolution of such ideals has been determined and upper bounds for their regularity have been given, see [4], [11], [14], [16], [18], [22] and [23]. When m > 2, these ideals are called determinantal facet ideals. Here its generators are in bijection to the facets of a pure simplicial complex of dimension m − 1. These ideals were introduced and first studied in [10]. About the resolution of determinantal facet ideals is known even less than for binomial edge ideals. Apart from a very special case considered in [19], the resolution of a determinantal facet ideal is only known when the underlying simplicial complex is a simplex, in which case the Eagon-Northcott complex provides a resolution. One of the motivations for writing this paper was a conjecture made in [14] by the second and third author of this paper. Given a finite simple graph G. Let JG be its binomial edge ideal. Then the graded Betti numbers βi,i+2 (JG ) give the ranks of the corresponding free modules of the linear strand of JG . Now the conjecture made in [14] says that βi,i+2 (JG ) = (i + 1)fi+1 (∆(G)) for all i ≥ 0. Here ∆(G) is the clique complex of G and fi+1 (∆(G)) is the number of faces of ∆(G) of dimension i + 1. In this paper we will not only prove this conjecture but also prove an analogue 2010 Mathematics Subject Classification. Primary 16E05, 13C40; Secondary 05E40, 13C05. Key words and phrases. Linear strands, generalized Eagon-Northcott complex, determinantal facet ideals, binomial edge ideals. The third author was supported by the German Research Council DFG-GRK 1916. 1 statement for determinantal facet ideals, see Corollary 4.3. In fact, the linear strand of any determinantal facet ideal, which is a subcomplex of its minimal graded free resolution, is explicitly described in Theorem 4.1. In general, the linear strand of a graded minimal free resolution is a linear complex, that is, a complex whose matrices describing the differentials of the complex have linear forms as entries. Now given any finite linear complex, one may ask when such a complex is the linear strand of the graded minimal free resolution of a suitable graded module. This question, which is of interest by itself, is answered in the first section of the paper, see Theorem 1.1. A different description of the linear strand is given by Eisenbud in [7, Theorem 7.4]. Our characterization of linear strands is then used to prove that the complex constructed in Section 2 is indeed the linear strand of the corresponding determinantal facet ideal. In order to describe this complex, let ∆ be a simplicial complex on the vertex set [n] and let ϕ be a linear map between free modules of rank m and n. We define a subcomplex of the Eagon-Northcott complex C(ϕ) associated with ϕ which we denote by C(∆; ϕ). The complex C(∆; ϕ) is obtained from C(ϕ) by restriction to basis elements determined by the faces of ∆. In order to see that this is indeed a well defined subcomplex one has to describe the differentials of C(ϕ) explicitly in terms of the given natural bases. For the convenience of the reader we included this description of the Eagon-Northcott complex. We call the subcomplex C(∆; ϕ) the generalized Eagon-Northcott complex (determined by ∆ and ϕ). When m = 1, C(∆; ϕ) is a very special cellular complex. Cellular complexes were considered by Bayer, Peeva and Sturmfels ([1], [2]). Fløystad in [12] considered this cellular complex and called their homology to be the enriched homology module of ∆ with respect to K. In Theorem 3.1 it is shown that C(∆; ϕ) is the linear strand of a module with initial degree m if and only if ∆ has no minimal nonfaces of cardinality ≥ m + 2. Here and in all the following statements ϕ is defined by an (m × n)-matrix X of indeterminates. We use this result in Section 4 to obtain one of the main results of this paper: let C be an m-uniform clutter, that is a collection of subsets of [n] of cardinality m. The elements of C are called the circuits of C. The clique complex of C, which we denote by ∆(C), has as faces all subsets σ of [n] with the property that each m-subset of σ belongs to C. The clique complex ∆(C) has so minimal nonfaces of cardinality ≥ m + 2, and hence Theorem 3.1 can be applied, and it is shown in Theorem 4.1 that C(∆(C); ϕ) is the linear strand of the resolution of JC where JC is generated by all maximal minors of X whose columns are determined by the circuits of C. From this fact one deduces immediately the formula for the Betti numbers of the linear strand of JC , namely ! m+i−1 fm+i−1 (∆(C)). βi,i+m (JC ) = m−1 In Section 5 the determinantal facet ideals with linear resolution are characterized, and it is shown in Theorem 5.1 that JC has linear resolution if and only if JC is linearly presented, and that this is the case if and only if the clutter C is complete, which means that C consists of all m-subsets of a given set V ⊂ [n]. 2 1. The linear strand of a graded free resolution Let K be a field and S = K[x1 , . . . , xn ] be the polynomial ring over K in the indeterminates x1 , . . . , xn . We view S as a standard graded K-algebra by assigning to each xi the degree 1. Let M be a finitely generated graded S-module. Let d be the initial degree of M, namely, the smallest integer i such that Mi 6= 0. L Let (F, ∂) be the minimal graded free resolution of M. Then Fi = j S(−j)βi,j where βij = dimK Tori (K, M)j is the ij-th graded Betti number of M. Note that βij = 0 for all pairs (i, j) with j < i + d. Let Filin be the direct summand S(−i − d)βi,i+d of Fi . It is obvious that ∂(Filin ) ⊂ lin Fi−1 for all i > 0. Thus Flin : · · · −→ F2lin −→ F1lin −→ F0lin −→ 0 is a subcomplex of F, called the linear strand of the resolution of M. Throughout the paper, by the linear strand of a module M we mean the linear strand of the resolution of M. A graded complex G : · · · → G2 → G1 → G0 → 0 of finitely generated graded free S-modules is called a linear complex (with initial degree d) if for all i, Gi = S(−i−d)bi for suitable integers bi . Obviously, Flin is a linear complex. The question arises which linear complexes are the linear strand of a finitely generated graded module. The following result answers this question. Theorem 1.1. Let G be a finite linear complex with initial degree d. The following conditions are equivalent: (a) G is the linear strand of a finitely generated graded S-module with initial degree d. (b) Hi (G)i+d+j = 0 for all i > 0 and for j = 0, 1. Proof. We may assume that d = 0. (a) ⇒ (b): Let G = Flin where F is the minimal graded free resolution of a finitely generated graded S-module M. We denote by D the quotient complex F/Flin. From the exact sequence of complexes 0 −→ Flin −→ F −→ D −→ 0 we deduce that (1) Hi (Flin)i+j ∼ = Hi+1 (D)i+1+(j−1) for all i > 0 and all j. Since (Di )j = 0 for all i ≥ 0 and all j ≤ i, we see that Hi+1 (D)i+1+(j−1) = 0 for j = 0 and j = 1. Thus (1) yields the desired conclusion. (b) ⇒ (a): Let M be a graded module whose minimal graded free resolution F satisfies the condition that F1lin → F0lin is isomorphic to G1 → G0 . (For example we could choose M = H0 (G)). We prove by induction on i that the truncated complex Gi → Gi−1 → · · · → G1 → G0 → 0 3 is isomorphic to the truncated complex lin Filin → Fi−1 → · · · → F1lin → F0lin → 0. By the definition of M this is the case for i = 1. Now let i > 1. By induction hypothesis we have Ker(Gi → Gi−1 ) ∼ = Ker(F lin → F lin ). i i−1 Let Z = Ker(Gi → Gi−1 ) and let W be the submodule of Z which is generated by lin all elements of Z of degree i + 1. Similarly we let Z ′ = Ker(Filin → Fi−1 ) and W ′ be the submodule of Z ′ which is generated by all the elements of Z ′ of degree i + 1. Then the isomorphism Z ∼ = Z ′ induces an isomorphism W ∼ = W ′. Since by assumption Hi (G)i+1 = Hi+1 (G)i+1 = 0, it follows that Gi+1 → W is a lin minimal free presentation of W . By the same reason, Fi+1 → W ′ is a minimal free presentation of W ′ . This shows that the truncations of G and Flin are isomorphic up to homological degree i + 1.  We would like to remark that the linear strand Flin of a module M with initial degree d is only determined by the submodule N of M which is generated by the elements of degree d of M. Moreover, N = Coker(F1lin → F0lin ) if and only if N is linearly presented. Consider for example the ideal I = (x2 , xy, y 3, z 2 ) ⊂ S = K[x, y, z] with minimal graded free resolution F. Then I and J = (x2 , xy, z 2 ) have the same linear strand. We have that J 6= Coker(F1lin → F0lin ), since J is not linearly presented. Indeed, Coker(F1lin → F0lin) is isomorphic to (x2 , xy)S ⊕ S(−2). Corollary 1.2. Let G be a finite linear complex with initial degree d satisfying the conditions (b) of Theorem 1.1. Furthermore, let W be a finitely generated graded Smodule with minimal graded free resolution F satisfying the property that G1 → G0 is isomorphic to F1lin → F0lin . Then G is isomorphic to the linear strand of W . Proof. In the proof of Theorem 1.1 (b) ⇒ (a) we have seen that if G is a finite linear complex with initial degree d satisfying the conditions (b) of Theorem 1.1 and such  that G1 → G0 ∼ = Flin. = F1lin → F0lin , then G ∼ 2. The generalized Eagon-Northcott complex Let F and G be free S-modules of rank m and n, respectively, with m ≤ n, and let ϕ : G → F be an S-module homomorphism. We choose a basis f1 , . . . , fm of F and a Pm basis g1 , . . . , gn of G. Let ϕ(gj ) = i=1 αij fi for j = 1, . . . , n. The matrix α = (αij ) describing ϕ with respect to these bases is an (m × n)-matrix with entries in S. The ideal of m-minors of this matrix is independent of the choice of these bases and is denoted Im (ϕ). It is know that with suitable grade conditions on Im (ϕ), the socalled Eagon-Northcott complex provides a free resolution of Im (ϕ), see [5]. There are nice and basis free descriptions of the Eagon-Northcott complex, see for example [3] and [6]. For our purpose however a description of the Eagon-Northcott complex and its differentials in terms of natural bases is required. For the convenience of the reader and due to the lack of a suitable reference we recall this description. 4 We denote by S(F ) the symmetric algebra of F . Then S(F ) is isomorphic to the polynomial ring over S in the variables f1 , . . . , fm . The module G ⊗S S(F ) is a free S(F )-module with basis g1 ⊗ 1, . . . , gn ⊗ 1 and ϕ gives rise to the S(F )-linear map G ⊗ S(F ) → S(F ) with gj ⊗ 1 7→ ϕ(gj ) for all j, which in turn induces the Koszul complex K(ϕ) : 0 → n ^ G ⊗ S(F ) → · · · → whose differential δ is defined as follows: (2) δ(gj1 ∧gj2 ∧· · ·∧gji ⊗1) = i X 1 ^ G ⊗ S(F ) → 0 ^ G ⊗ S(F ) → 0, (−1)k+1 gj1 ∧gj2 ∧· · ·∧gjk−1 ∧gjk+1 ∧· · ·∧gji ⊗ϕ(gjk ). k=1 The Koszul complex K(ϕ) splits into graded components K(ϕ)i : 0 → i ^ G ⊗ S0 (F ) → · · · → 1 ^ G ⊗ Si−1 (F ) → 0 ^ G ⊗ Si (F ) → 0, each of which is a complex of free S-modules. We are interested in the S-dual of the (n − m)-th component of K(ϕ) which is the complex (3) 0 → ( Let 0 ^ G ⊗Sn−m (F ))∗ → · · · → ( µ: n−i ^ n−m−1 ^ G ⊗S1 (F ))∗ → ( G → HomS ( Vn−i i ^ G, n ^ n−m ^ G ⊗S0 (F ))∗ → 0. G) be the isomorphism which assigns to a ∈ G the S-linear map µa ∈ HomS ( with µa (b) = a ∧ b. Composing µ with the isomorphism i ^ n ^ Vi G, Vn i ^ G)∗ which is induced by the isomorphism Sg1 ∧ . . . ∧ gn ∼ = S with g1 ∧ . . . ∧ gn 7→ 1, we obtain the isomorphisms HomS ( n−i ^ (4) G∼ =( i ^ G, G) → ( G)∗ for i = 0, . . . , n. Furthermore, if we use that for any two finitely generated free S-modules A and B there are natural isomorphisms (A ⊗ B)∗ ∼ (5) = Hom(A, B ∗ ) ∼ = A∗ ⊗ B ∗ , the complex (3) becomes the complex C(ϕ) : 0 → n ^ G ⊗ Sn−m (F )∗ → . . . → m+1 ^ G ⊗ S1 (F )∗ → which is called the Eagon-Northcott complex. There is a natural augmentation map ε: m ^ G ⊗ S0 (F )∗ → Im (ϕ) 5 m ^ G ⊗ S0 (F )∗ → 0, G) which assigns to gi1 ∧ · · · ∧ gim ⊗ 1 the m-minor of α with the columns i1 , i2 , . . . , im . It is known that grade Im (ϕ) ≤ n − m + 1 and equality holds if and only if C(ϕ) → Im (ϕ) → 0 is exact, that is, if and only if C(ϕ) provides a free resolution of Im (ϕ), see [3]. We now describe the differential of the Eagon-Northcott complex. The map δ∗ : ( r ^ G ⊗ Sn−m−r (F ))∗ → ( r+1 ^ G ⊗ Sn−m−r−1 (F ))∗ is just the dual of the differential δ defined in (2). Thus if we set gσ = gj1 ∧ · · · ∧ gjr am for σ = {j1 < j2 < · · · < jr } and set f a = f1a1 f2a2 · · · fm with |a| = a1 + · · · + am = n − m − r, then r X δ(gσ ⊗ f ) = a (−1)k+1 gσ\{jk } ⊗ ϕ(gjk )f a k=1 r X m X = am ). (−1)k+1 αijk (gσ\{jk } ⊗ f1a1 · · · fiai +1 · · · fm k=1 i=1 Therefore, (6) δ ∗ ((gσ ⊗ f a )∗ ) = (−1)sign(σ,j) X m X am ∗ ), αij (gσ∪{j} ⊗ f1a1 · · · fiai −1 · · · fm i=1 ai >0 j∈[n]\σ where sign(σ, j) = |{i ∈ σ : i < j}|. (a ) (am ) Let f (a) = f1 1 · · · fm denote the basis element in (Sj (F ))∗ which is dual to a am f a = f1 1 · · · fm . Furthermore, we set f (a) = 0, if ai < 0 for some i. Then by using the isomorphisms (4) and (5), we see that (gσ ⊗ f a )∗ is mapped to (−1)s(σ) gσc ⊗ f (a) , where σ c = [n] \ σ and where s(σ) is defined by the equation gσ ∧ gσc = (−1)s(σ) g[n] . Thus, δ ∗ identifies with the map (7) ∂(gσ ⊗ f (a) ) = (−1)|σ c| m+i m XX (a1 ) (−1)k+1 αijk (gσ\{jk } ⊗ f1 (ai −1) · · · fi (am ) · · · fm ), k=1 i=1 which is the differential of the Eagon-Northcott complex. In this formula, the sign c (−1)|σ | which only depends on the homological degree of gσ ⊗ f (a) , may be skipped, and we will do this in the final presentation of ∂. Indeed, to see that (7) holds, we note that, due to (6), δ ∗ ((−1)s(σ) gσc ⊗ f (a) ) = X j∈σc (−1)sign(σ,j) m X (a1 ) (−1)s(σ∪{j}) αij (gσc \{j} ⊗ f1 i=1 ai >0 6 (ai −1) · · · fi (am ) · · · fm ). Thus δ ∗ (gσ ⊗ f (a) ) = m X X j∈σ (−1)s(σ c )+sign(σ c ,j)+s(σ c ∪{j}) (a1 ) αij (gσ\{j} ⊗ f1 (ai −1) · · · fl (am ) · · · fm ). i=1 ai >0 Finally (−1)s(σ c ∪{j}) g[n] = gσc ∪{j} ∧ gσ\{j} = (−1)|σ |σc |−sign(σc ,j)+sign(σ,j) = (−1) c |−sign(σ c ,j) gσ c ∧ gσ |σc |−sign(σc ,j)+sign(σ,j)+s(σc ) = (−1) |σc | It follows that (−1) (−1)sign(σ,j) = (−1) gσc ∧ gj ∧ gσ\{j} g[n] . s(σc )+sign(σc ,j)+s(σc ∪{j}) , as desired. (a) In order to simplify notation we set b(σ; a) = gσ ⊗f . Then the elements b(σ; a) with |σ| = m + i and a = (a1 , . . . , am ) such that a1 + · · · + am = i form a basis of Vm+i G ⊗ Si (F )∗ , and ∂(b(σ; a)) = m+i m XX (−1)k+1αℓjk b(σ \ {jk }; a − eℓ ). k=1 ℓ=1 Here e1 , . . . , em is the canonical basis of Zm . Corollary 2.1. Suppose that grade Im (ϕ) = n − m + 1. Then Im (ϕ) has a linear resolution and ! ! m+i−1 n for i = 0, . . . , n − m. βi (Im (ϕ)) = m−1 m+i Now let ∆ be a simplicial complex on the vertex set [n]. We set Ci (ϕ) = m+i G ⊗ Si (F )∗ and denote Ci (∆; ϕ) the free submodule of Ci (ϕ) generated by all b(σ; a) such that σ ∈ ∆ with |σ| = m+i, and a ∈ Zm ≥0 with |a| = i. Since ∂(b(σ; a)) ∈ Ci−1 (∆; ϕ) for all b(σ; a) ∈ Ci (∆; ϕ), we obtain the subcomplex V C(∆; ϕ) : 0 → Cn−m (∆; ϕ) → · · · → C1 (∆; ϕ) → C0 (∆; ϕ) → 0 of C(ϕ) which we call the generalized Eagon-Northcott complex attached to the simplicial complex ∆ and the module homomorphism ϕ : G → F . In particular, if ∆ is just a simplex on n vertices, then C(∆; ϕ) coincides with C(ϕ). Moreover, note that the generalized Eagon-Northcott complex of a simplicial complex ∆ is determined by those facets of ∆ whose dimension is at least m − 1. More precisely, if we obtain the simplicial complex ∆′ from ∆ by removing all facets of dimension less than m − 1, then the complexes C(∆; ϕ) and C(∆′ ; ϕ) are the same. 3. The generalized Eagon-Northcott complex as a linear strand of a module Let X be an (m × n)-matrix of indeterminates xij . We fix a field K and let S be the polynomial ring over K in the variables xij . Moreover, let ϕ : G → F be the S-module homomorphism of free S-modules given by the matrix X. In the rest of this paper, we fix this situation. 7 Now we give a (Zm ×Zn )-grading to the polynomial ring S, by setting mdeg(xij ) = (ei , εj ) where ei is the i-th canonical basis vector of Zm and εj is the j-th canonical basis vector of Zn . Now, let ∆ be a simplicial complex. Then the chain complex C(∆; ϕ) inherits this grading. More precisely, for each i, the degree of a basis element b(σ; a) of Ci (∆; ϕ) with σ = {j1 , . . . , jm+i } is set to be (a + 1, γ) ∈ Zm × Zn , where γ = εj1 + · · · + εjm+i , and 1 is the vector in Zm whose entries are all equal to 1. Then, one can see that all the homomorphisms in the chain complex C(∆; ϕ) is homogeneous. To characterize all simplicial complexes for which the generalized Eagon-Northcott complex C(∆; ϕ) is the linear strand of a finitely generated graded S-module, we need to introduce a concept about simplicial complexes which is crucial in the sequel. Let ∆ be a simplicial complex on the vertex set V , and let σ be a subset of V . Then recall that σ is called a minimal nonface of ∆ if σ ∈ / ∆, but all its proper subsets belong to ∆. Theorem 3.1. Let ∆ be a simplicial complex on n vertices, and let m be a positive integer. Then the following conditions are equivalent: (a) C(∆; ϕ) is the linear strand of a finitely generated graded S-module with initial degree m. (b) ∆ has no minimal nonfaces of cardinality ≥ m + 2. Proof. We may assume that m ≤ dim ∆ + 1, because otherwise C(∆; ϕ) = 0 and the assertion is trivial. By Theorem 1.1 it suffices to show that ∆ has no minimal nonfaces of cardinality ≥ m + 2 if and only if Hi (C(∆; ϕ))i+m+j = 0 for all i > 0 and for j = 0, 1. First suppose that j = 0, and let z be a nonzero cycle of degree m + i in Ci (∆; ϕ) for some i > 0. Then z is also a cycle in Ci (ϕ). Since C(ϕ) is exact, z is a boundary in Ci (ϕ). So there is a nonzero element f of degree m + i in Ci+1 (ϕ) such that ∂(f ) = z which is a contradiction, since Ci+1 (ϕ)m+i = 0. This implies that there is no nonzero cycle of degree m + i in Ci (∆; ϕ), and hence Hi (C(∆; ϕ))i+m = 0. Therefore, to prove the theorem it is enough to show that ∆ has no minimal nonfaces of cardinality ≥ m + 2 if and only if Hi (C(∆; ϕ))i+m+1 = 0 for all i > 0. Equivalently, we show that ∆ has a minimal nonface of cardinality at least m + 2 if and only if Hi (C(∆; ϕ))i+m+1 6= 0 for some i > 0. To prove this, suppose that there is a minimal nonface σ = {j1 , . . . , jm+1+i } of ∆ for some i > 0. Therefore, σ ∈ / ∆ and for each t = 1, . . . , m + 1 + i, we have σ \ {jt } ∈ ∆. Now let z = ∂(b(σ; a)) for some a ∈ Zm ≥0 with |a| = i + 1. Then z ∈ Ci (∆; ϕ), and z is a nonzero element of Ci (ϕ). Since z is a cycle in Ci (ϕ), it is also a cycle in Ci (∆; ϕ). On the other hand, the multidegree of z is (a + 1, γ) / ∆, there is no basis element of multidegree where γ = εj1 + · · · + εjm+1+i . Since σ ∈ (a + 1, γ) in Ci+1 (∆; ϕ), and hence z is not a boundary in Ci (∆; ϕ). This implies that Hi (C(∆; ϕ))i+m+1 6= 0. Conversely, suppose that Hi (C(∆; ϕ))i+m+1 6= 0 for some i > 0. Then there is a cycle z ∈ Ci (∆; ϕ) which is not a boundary. We may assume that z is multihomogeneous with mdeg(z) = (a+1, γ) with γ = εj1 +· · ·+εjm+1+i and |a| = i+1. Since z is also a cycle in Ci (ϕ) and since C(ϕ) is exact, we deduce that z is a boundary in Ci (ϕ). 8 Thus there is a nonzero multihomogeneous element f ∈ Ci+1 (ϕ) such that ∂(f ) = z. Since C(ϕ) is a multigraded resolution, it follows that mdeg(f ) = (a+1, γ). Since any two basis elements of Ci+1 (ϕ) have different multidegrees, it follows that f = λb(σ; a) for some 0 6= λ ∈ K with σ = {j1 , . . . , jm+1+i }. Because z is not a boundary in Ci (∆; ϕ), the basis element b(σ; a) does not belong to Ci+1 (∆; ϕ), and hence σ ∈ / ∆. On the other hand, we have z = ∂(f ) = λ m+1+i m X X k=1 (−1)k+1 xℓjk b(σ \ {jk }; a − eℓ ). ℓ=1 Since the basis elements of Ci+1 (∆; ϕ) form a subset of the basis elements of Ci+1 (ϕ), it follows from this presentation of z that b(σ \ {jk }; a − eℓ ) ∈ Ci (∆; ϕ) for all jk and ℓ for which b(σ \ {jk }; a − eℓ ) 6= 0. Since z is not zero, there exist jk and l with b(σ \ {jk }; a − eℓ ) 6= 0. Therefore, a − eℓ ∈ Zm ≥0 . Hence b(σ \ {jt }; a − eℓ ) 6= 0 for all t = 1, . . . , m + 1 + i. It follows that b(σ \ {jt }; a − eℓ ) ∈ Ci+1 (∆; ϕ) for all t = 1, . . . , m + 1 + i. Therefore, σ \ {jt } is a face of ∆ for all t = 1, . . . , m + 1 + i, so that σ is a minimal nonface of ∆ of cardinality m + 1 + i ≥ m + 2, as desired.  We would like to remark that a subclass of simplicial complexes satisfying condition (b) of Theorem 3.1 has been considered in [20] for a different purpose. There, minimal nonfaces are called missing faces. Now, we give some examples of other classes of simplicial complexes which satisfy condition (b) of Theorem 3.1. First recall that a flag simplicial complex is a simplicial complex whose minimal nonfaces all have cardinality equal to two. Flag complexes are exactly clique complexes of graphs, i.e. a simplicial complex whose faces are cliques of a graph. In [17], a generalization of flag complexes has been introduced which we recall it in the following. In [17] a subset T of V is called a critical clique of ∆ if each pair of vertices form a 1-dimensional face of ∆ and T \ {v} ∈ ∆ for some v ∈ T . Let dim ∆ = d − 1 and let i be an integer with 1 ≤ i ≤ d. Then ∆ is called i-banner if every critical clique T of cardinality at least i + 1 is a face of ∆. In [17, Lemma 3.3] it is shown that a simplicial complex is 1-banner if and only if it is 2-banner, and this is equivalent to being flag. Now, let 1 ≤ m ≤ d. Then an (m + 1)-banner satisfies condition (b) of Theorem 3.1, namely has no minimal nonfaces of cardinality ≥ m + 2, see also [17, Lemma 3.6]. In particular, all flag complexes satisfy condition (b) of Theorem 3.1, because i-banner implies (i + 1)-banner for all i. We also would like to remark that not all simplicial complexes with no minimal nonfaces of cardinality ≥ m + 2 are (m + 1)-banner. For example, let ∆ be the simplicial complex with the vertex set {1, 2, 3, 4} whose facets are {1, 2, 3}, {1, 3, 4} and {2, 3, 4} and let m = 2. Then ∆ has no minimal nonfaces of cardinality 4, but it is not 3-banner. Indeed, T = {1, 2, 3, 4} ∈ / ∆ but it is a critical clique of ∆. 9 4. The linear strand of determinantal facet ideals and binomial edge ideals A clutter C on the vertex set [n] is a collection of subsets of [n] such that there is no containment between its elements. An element of C is called a circuit. In this paper, we assume that each vertex of a clutter C belongs to some circuits of C. If all circuits of C have the same cardinality d, then C is said to be an d-uniform clutter. Let C be an m-uniform clutter. To each circuit τ ∈ C with τ = {j1 , . . . , jm } and 1 ≤ j1 < j2 < · · · < jm ≤ n we assign the m-minor mτ of X which is determined by the columns 1 ≤ j1 < j2 < · · · < jm ≤ n. We denote by JC the ideal in S which is generated by the minors mτ with τ ∈ C. This ideal which has first been considered in [10] is called the determinantal facet ideal of C since the circuits of C may be considered as the facets of a simplicial complex. In the case that C is a 2-uniform clutter, C may be viewed as a graph G, and in this case JC = JG , where JG is the binomial edge ideal of G, as defined in [13]. A clique of C is a subset σ of [n] such that each m-subset τ of σ is a circuit of C. We denote by ∆(C) the simplicial complex whose faces are the cliques of C which is called the clique complex of C. A clutter is called complete if its clique complex is a simplex. Note that the clique complex of an m-clutter has no minimal nonface of cardinality ≥ m + 2. The example at the end of Section 3 also shows that not all simplicial complexes which satisfy condition (b) of Theorem 3.1 are clique complexes of clutters. Here we also give an example of a 3-clutter whose clique complex is not 4-banner, although all clique complexes of 2-clutters are 3-banner, as mentioned before. Let C be the 3-clutter on the vertices {1, . . . , 6} with the circuits {1, 2, 3}, {1, 3, 4}, {2, 3, 4}, {3, 4, 5}, {3, 4, 6}, {3, 5, 6} and {4, 5, 6}. Then ∆(C) = h{1, 2, 3}, {1, 3, 4}, {2, 3, 4}, {3, 4, 5, 6}i. Thus ∆(C) is not 4-banner, since {1, 2, 3, 4} is a critical clique of ∆(C) which is not a face. The following theorem is the main result of this section. Theorem 4.1. Let C be an m-uniform clutter on the vertex set [n], and let F be the minimal graded free resolution of JC . Then Flin ∼ = C(∆(C); ϕ). Proof. To prove the theorem, by Corollary 1.2 and Theorem 3.1, it suffices to prove the following F1lin → F0lin ∼ = C1 (∆(C); ϕ) → C0 (∆(C); ϕ). First, note that JC is a homogeneous ideal of S with respect to the (Zm × Zn )multigrading introduced in Section 3. Now, we define ψ C0 (∆(C); ϕ) −→ JC with ψ(b(σ; 0)) = mσ for each σ = {j1 , . . . , jm } ∈ C, so that it is a minimal free presentation of JC . Since C(ϕ) is the resolution of Im (ϕ), the following commutative diagram implies that ψ ◦ ∂ = 0. 10 C1 (∆(C); ϕ)  ∂ _  C1 (ϕ) / C0 (∆(C); ϕ)  ψ / _ ∂   / C0 (ϕ) JC _ / Im (ϕ) Let Z = ker ψ and Z lin be the S-module generated by the elements of Z whose degree is m + 1. Since C1 (∆(C); ϕ) is generated in degree m + 1, it follows that ∂′ ∂ induces a homogeneous homomorphism C1 (∆(C); ϕ) −→ Z lin . The desired result follows once we have shown that ∂ ′ induces an isomorphism of vector spaces ∂′ C1 (∆(C); ϕ)m+1 −→ (Z lin )m+1 = Zm+1 . Let z be a nonzero element of Z lin of multidegree (es + 1, γ) for some s = 1, . . . , m where |γ| = m + 1. Let γ = εj1 + · · · + εjm+1 . Since z 6= 0, at least m vectors among εj1 , . . . , εjm+1 are distinct, say εj1 , . . . , εjm , and the basis element b(σ; 0) appears in the presentation of z as the linear combination of the basis elements of C0 (∆(C); ϕ), where σ = {j1 , . . . , jm }. First suppose that jm+1 ∈ σ. Then according to the (Zm × Zn )-multigrading, the only basis element of C0 (∆(C), ϕ) which contributes to this multidegree in z, is b(σ; 0). Then, it follows that z = λxsjm+1 b(σ; 0) for some 0 6= λ ∈ K, and hence we have ψ(z) 6= 0, a contradiction. Therefore, jm+1 ∈ / σ. Let τ = {j1 , . . . , jm+1 }, and without loss of generality assume that j1 < · · · < jm+1 . Then all possible basis elements of C0 (∆(C), ϕ) which could contribute to this multidegree in z are of the form b(τ \ {jk }; 0). By comparing the multidegrees, we observe that z = Pm+1 k=1 λk xsjk b(τ \ {jk }; 0) for some λk ∈ K. From the above diagram it follows that z is a boundary of C(ϕ). This implies that there exists a nonzero multihomogeneous element f ∈ C1 (∆; ϕ) with ∂(f ) = z. Since mdeg(f ) = mdeg(z), it follows that P k+1 f = λb(τ ; es ) for some 0 6= λ ∈ K. Therefore, z = ∂(f ) = λ m+1 xsjk b(τ \ k=1 (−1) k+1 {jk }; 0). This implies that for each k = 1, . . . , m+1, we have λk = (−1) λ and the basis element b(τ \{jk }; 0) appears in the presentation of z with a nonzero coefficient. Hence for each k = 1, . . . , m + 1, τ \ {jk } is a circuit of C which implies that τ is a clique of C. This shows that the set {rs (τ ) : 1 ≤ s ≤ m, τ ∈ ∆(C), |τ | = m+1} with P k+1 rs (τ ) = m+1 xsjk b(τ \ {jk }; 0) for each clique τ = {j1 , . . . , jm+1 } of C, is a k=1 (−1) system of generators of Z lin . Since mdeg(rs (τ )) 6= mdeg(rs′ (τ ′ )) if (s, τ ) 6= (s′ , τ ′ ), we conclude that these elements form a basis of Zm+1 . Moreover, each basis element b(τ ; es ) is mapped by ∂ ′ to rs (τ ). Thus, we get the desired result.  Recall that for a simplicial complex ∆ and an integer i, the i-th skeleton ∆(i) of ∆ is the subcomplex of ∆ whose faces are those faces of ∆ whose dimension is at most i. Corollary 4.2. Let ∆ be a simplicial complex whose facets all have dimension at least m−1, and let C be the m-clutter whose circuits are exactly the facets of ∆(m−1) . Then C(∆; ϕ) is the linear strand of JC if and only if ∆ = ∆(C). Proof. It is enough to note that by our assumptions we have ∆ ⊆ ∆(C), so that by Theorem 4.1 the conclusion follows.  11 Recall that the f -vector (f0 (∆), f1 (∆), . . .) of a simplicial complex ∆ is the integer vector such that ft (∆) = |{σ ∈ ∆ : dim σ = t}| for each t = 0, . . . , dim ∆. The following corollary is a consequence of Theorem 4.1. In particular, in the case m = 2, it proves the conjecture in [22, page 338]. Corollary 4.3. Let C be an m-uniform clutter. Then ! m+i−1 fm+i−1 (∆(C)), βi,i+m (JC ) = m−1 for all i. Another straightforward consequence of Theorem 4.1 is the following result. Corollary 4.4. Let C be an m-uniform clutter. Then the length of the linear strand of JC is equal to dim ∆(C) − m + 1. In particular, proj dim JC ≥ dim ∆(C) − m + 1. The projective dimension of the determinantal facet ideals in general is bigger than the lower bound given in Corollary 4.4. This could be the case, even for 2-uniform clutters, namely binomial edge ideals of graphs. For example, it follows from [8, Theorem 1.1] that for a block graph G with n vertices, proj dim JG = n − c − 1 where c is the number of connected components of G. One can see that if G is a non-complete block graph G over n vertices, n − c − 1 > dim ∆(G) − 1, and hence in this case the lower bound given in Corollary 4.4 is strictly less than the projective dimension. 5. Determinantal facet ideals with linear resolution In this section, we determine when the minimal graded free resolution of determinantal facet ideals coincide with their linear strand. In other words, we characterize all m-uniform clutters whose determinantal facet ideal has a linear resolution. Moreover, we show that having a linear resolution is equivalent to being linearly presented for such ideals. Indeed, applying Theorem 4.1, we show that these properties occur for JC if and only if C is a complete clutter. Recall that a graded ideal I in S which is generated in degree d is called linearly presented if β1,j (I) = 0 for all j 6= d + 1. Moreover, I is said to have a linear resolution if βi,j (I) = 0 for all i and j 6= i + d. The following theorem is the main result of this section which in particular recovers the case of binomial edge ideals from [14]. Here, we consider S as a Zn -graded ring by setting deg(xij ) = εj for all i = 1, . . . , m, where by εj we mean the j-th canonical basis vector of Zn . Theorem 5.1. Let C be an m-uniform clutter. Then the following conditions are equivalent: (a) JC has a linear resolution. (b) JC is linearly presented. (c) C is a complete clutter. To prove the above theorem, we need the following lemma. 12 Lemma 5.2. Let C be an m-uniform clutter on [n] and suppose that JC is linearly presented. Then the following conditions hold: (a) for each σ1 , σ2 ∈ ∆(C) with |σ1 ∩ σ2 | ≥ m − 1, we have σ1 ∪ σ2 ∈ ∆(C). (b) for circuits τ and σ of C, ρ ∈ ∆(C) with σ, τ ⊆ ρ, and c ∈ / ρ, it follows that τ ∪ {c} ∈ ∆(C) if and only if σ ∪ {c} ∈ ∆(C). Proof. (a) Before proving (a) we first show: (∗) for each (m + 1)-subset σ of [n] which contains at least two circuits of C, we have σ ∈ ∆(C). Suppose on the contrary this is not the case, and suppose σ = {a1 , . . . , am+1 } is an (m+1)-subset of [n] which contains at least two circuits of C, but σ ∈ / ∆(C). We may assume that τ1 = {a1 , . . . , am } and τ2 = {a2 , . . . , am+1 } are circuits of C. Let (F, δ) be the minimal graded free resolution of JC . Note that F0 = C0 (∆(C); ϕ) and F1lin = C1 (∆(C); ϕ) by Theorem 4.1. So, F1 = C1 (∆(C); ϕ) ⊕ F1′ , for some free graded Smodule F1′ . Let z = mτ1 b(τ2 ; 0) − mτ2 b(τ1 ; 0) ∈ C0 (∆(C); ϕ). Since z ∈ ker ψ, there exists a nonzero homogeneous element f = (f1 , f2 ) ∈ F1 where f1 ∈ C1 (∆(C); ϕ), f2 ∈ F1′ and δ(f ) = z. On the other hand, deg(z) = εa1 + 2εa2 + · · ·+ 2εam + εam+1 , so that deg(f1 ) = deg(f2 ) = εa1 + 2εa2 + · · · + 2εam + εam+1 , because δ is homogeneous. Any basis element of C1 (∆(C); ϕ) appearing in the presentation of f1 is of the form b(σ; et ) for some t = 1, . . . , m. Since by the assumption σ is not a clique in C, no such basis element exists in C1 (∆(C); ϕ), and hence f1 = 0. Therefore, we have f2 6= 0, since f 6= 0. This contradicts to our assumption that JC is linearly presented. Now we show that (∗) implies (a). Suppose that condition (∗) holds. Let σ1 = {v1 , . . . , vk , u1 , . . . , ur } and σ2 = {v1 , . . . , vk , w1 , . . . , ws } be two cliques of C with |σ1 ∩ σ2 | = k ≥ m − 1. We show that σ1 ∪ σ2 is a clique. For this, it is enough to show that each m-subset H = {va1 , . . . , vat , ub1 , . . . , ubp , wc1 , . . . , wcq } of σ1 ∪ σ2 in which t + p + q = m and p, q ≥ 1, is a circuit of C. We prove this by induction on p + q. First assume that p = q = 1. Then we have τ1 = {va1 , . . . , vam−1 , ub1 } ∈ C and τ2 = {va1 , . . . , vam−1 , wc1 } ∈ C, since k ≥ m − 1 and σ1 , σ2 ∈ ∆(C). Thus, {va1 , . . . , vam−1 , ub1 , wc1 } contains at least two circuits τ1 and τ2 of C, and hence by (∗) it is a clique in C on m + 1 vertices. Therefore, all its m-subsets are circuits of C, and in particular, {va1 , . . . , vam−2 , ub1 , wc1 } ∈ C, so that we are done in this case. Now suppose that p + q = ℓ > 2. Then by induction hypothesis, we have (H ∪{vat+1 })\{ubp } ∈ C and (H ∪{vat+1 })\{wcq } ∈ C for all at+1 ∈ [k]\{a1 , . . . , at }. Therefore, by (∗) H ∪ {vat+1 } ∈ ∆(C) which implies that H is a circuit of C, as desired. (b) It is enough to show that σ ∪ {c} ∈ ∆(C) if τ ∪ {c} ∈ ∆(C). If τ = σ, there is nothing to prove. Suppose τ ∪ {c} ∈ ∆(C), a ∈ σ \ τ and b ∈ τ \ σ. Then (τ \ {b}) ∪ {c} is a circuit of C. Since ρ ∈ ∆(C), we have τ ′ = (τ \ {b}) ∪ {a} is also a circuit of C. Then it follows by (a) that τ ′ ∪ {c} ∈ ∆(C). Since |σ \ τ ′ | < |σ \ τ |, induction on |σ \ τ | yields the result.  Now we are ready to prove Theorem 5.1. 13 Proof of Theorem 5.1. (a) ⇒ (b) is clear, and (c) ⇒ (a) follows since the EagonNorthcott complex is the minimal graded free resolution of the determinantal facet ideal of a complete clutter. (b) ⇒ (c): We may assume that C is a clutter on the vertex set [n]. Assume that JC is linearly presented, and let F be the minimal graded free resolution of JC . Then by Theorem 4.1 we have F1 = C1 (∆(C); ϕ). Now suppose on the contrary that C is not complete. We claim that there exist two circuits in C whose union is not a clique in C. Now we prove the claim. Since C is not complete, there is an m-subset A = {c1 , . . . , cm } of [n] which is not a circuit of C. Let k be the biggest integer such that {c1 , . . . , ck } is contained in a circuit σ1 . Then k < m, because A is not a circuit. On the other hand, ck+1 ∈ σ2 for some σ2 ∈ C, but ck+1 ∈ / σ1 and therefore σ2 6= σ1 . Moreover, σ1 ∪ σ2 is not a clique in C, since otherwise the set {c1 , . . . , ck , ck+1 } is contained in a circuit of C which is a contradiction, because k is the biggest integer with this property. This proves the claim. Let z = mσ2 b(σ1 ; 0) − mσ1 b(σ2 ; 0). Since z ∈ ker ψ, there exists a nonzero multihomogeneous element w ∈ C1 (∆(C); ϕ) of the same multidegree as z with ∂(w) = z. Let A = {a ∈ σ1 \ σ2 : σ2 ∪ {a} ∈ ∆(C)} and B = {b ∈ σ2 \ σ1 : σ1 ∪ {b} ∈ ∆(C)}. Since ∂(w) = z, it follows that A, B 6= ∅. Now let ρ1 = σ1 ∪ B and ρ2 = σ2 ∪ A. Using repeatedly part (a) of Lemma 5.2, it follows that ρ1 , ρ2 ∈ ∆(C). Since ρ1 ∪ρ2 = σ1 ∪σ2 , we deduce that ρ1 ∪ρ2 ∈ / ∆(C). Then by Lemma 5.2 part (a), we have |ρ1 ∩ ρ2 | < m − 1. So, for i = 1, 2 we choose τi such that ρ1 ∩ ρ2 ⊆ τi ⊆ ρi and |τi | = m. In particular, τ1 6= τ2 . Now consider z ′ = mτ2 b(τ1 ; 0) − mτ1 b(τ2 ; 0) which is a nonzero element of ker ψ. We claim that τ1 ∪ {c} ∈ / ∆(C) for any c ∈ τ2 \ τ1 . Then it follows that there is no nonzero multihomogeneous element in C1 (∆(C); ϕ) of the same multidegree as z ′ , so that z ′ ∈ / ∂(C1 (∆(C); ϕ)) which is a contradiction, and hence C is a complete clutter and the desired result follows. Now, we prove the claim. Suppose there exists c ∈ τ2 \ τ1 such that τ1 ∪ {c} ∈ ∆(C). We have c ∈ / ρ1 , since c ∈ / τ1 . Then by Lemma 5.2 part (b), it follows that σ1 ∪ {c} ∈ ∆(C), since σ1 ⊆ ρ1 . This is a contradiction, since c ∈ τ2 ⊆ ρ2 which implies that c ∈ σ2 because c ∈ / A ⊆ σ1 , and so c ∈ B ⊆ ρ1 .  The following is an straightforward consequence of Theorem 5.1. Corollary 5.3. Let C be an m-clutter and consider the augmentation map ψ C0 (∆(C); ϕ) −→ JC −→ 0. ψ Then the complex C(∆(C); ϕ) −→ JC −→ 0 is exact if and only if C is a complete clutter. 14 References [1] D. Bayer, B. Sturmfels, Cellular resolutions of monomial ideals, J. Reine Angew. Math. 102 (1998), 123-140. [2] D. Bayer, I. Peeva, B. Sturmfels, Monomial resolutions, Math. Res. Lett. 5 (1-2) (1998), 31-46. [3] W. Bruns, U. Vetter, Determinantal rings, Lecture Notes in Mathematics, Springer, (1988). [4] A. Dokuyucu, Extremal Betti numbers of some classes of binomial edge ideals, to appear in Math. Reports. [5] J. A. Eagon, D. G. Northcott, Ideals defined by matrices and a certain complex associated with Them, Proceedings of the Royal Society of London. Series A, Mathematical and Physical Sciences 269, (1962), no. 1337, 188-204. [6] D. Eisenbud, Commutative algebra with a view towards algebraic geometry, Springer, (1995). [7] D. Eisenbud, The geometry of syzygies (A second course in Commutative Algebra and Algebraic Geometry), Springer, (2005). [8] V. Ene, J. Herzog, T. Hibi, Cohen-Macaulay binomial edge ideals, Nagoya Math. J. 204 (2011), 57-68. [9] V. Ene, J. Herzog, T. Hibi, Koszul binomial edge ideals, Bridging Algebra, Geometry, and Topology, Volume 96 of the series Springer Proceedings in Mathematics and Statistics, (2014), 125-136. [10] V. Ene, J. Herzog, T. Hibi, F. Mohammadi, Determinantal facet ideals, Michigan Math. J. 62 (2013), 39-57. [11] V. Ene, A. Zarojanu, On the regularity of binomial edge ideals, Math. Nachr. 288, No. 1 (2015), 19-24. [12] G. Fløystad, Enriched homology and cohomology modules of simiplicial complexes, J. Algebr. Comb. 25 (2007), 285-307. [13] J. Herzog, T. Hibi, F. Hreinsdotir, T. Kahle, J. Rauh, Binomial edge ideals and conditional independence statements, Adv. Appl. Math. 45 (2010), 317-333. [14] D. Kiani, S. Saeedi Madani, Binomial edge ideals with pure resolutions, Collect. Math. 65 (2014), 331-340. [15] D. Kiani, S. Saeedi Madani, Some Cohen-Macaulay and unmixed binomial edge ideals, to appear in Comm. Algebra, (arXiv:1506.01362). [16] D. Kiani, S. Saeedi Madani, The Castelnuovo-Mumford regularity of binomial edge ideals, arXiv:1504.01403. [17] S. Klee, I. Novik, From flag complexes to banner complexes, SIAM J. Discrete Math., 27 (2013), no. 2, 1146-1158. [18] K. Matsuda, S. Murai, Regularity bounds for binomial edge ideals, Journal of Commutative Algebra. 5(1) (2013), 141-149. [19] F. Mohammadi, Prime splittings of determinantal ideals, arXiv:1208.2930. [20] E. Nevo, Remarks on missing faces and lower bounds on face numbers, Electronic J. Combin. 16(2) (the Björner Festschrift volume) (2009), R8. [21] M. Ohtani, Graphs and ideals generated by some 2-minors, Comm. Algebra. 39 (2011), 905917. [22] S. Saeedi Madani, D. Kiani, Binomial edge ideals of graphs, Electronic J. Combin. 19(2) (2012), ♯ P44. [23] S. Saeedi Madani, D. Kiani, On the binomial edge ideal of a pair of graphs, The Electronic Journal of Combinatorics. 20(1) (2013), ♯ P48. [24] P. Schenzel, S. Zafar, Algebraic properties of the binomial edge ideal of a complete bipartite graph, An. St. Univ. Ovidius Constanta, Ser. Mat. 22(2) (2014), 217-237. [25] Z. Zahid, S. Zafar, On the Betti numbers of some classes of binomial edge ideals, The Electronic Journal of Combinatorics. 20(4) (2013), ♯ P37. 15 Jürgen Herzog, Fachbereich Mathematik, Universität Duisburg-Essen, Campus Essen, 45117 Essen, Germany E-mail address: [email protected] Dariush Kiani, Department of Pure Mathematics, Faculty of Mathematics and Computer Science, Amirkabir University of Technology (Tehran Polytechnic), 424, Hafez Ave., Tehran 15914, Iran, and School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P.O. Box 19395-5746, Tehran, Iran. E-mail address: [email protected], [email protected] Sara Saeedi Madani, Universität Osnabrück, Institut für Mathematik, 49069 Osnabrück, Germany E-mail address: [email protected], [email protected] 16
0math.AC
arXiv:1710.00264v1 [cs.DS] 30 Sep 2017 Bayesian estimation from few samples: community detection and related problems Samuel B. Hopkins∗ David Steurer† October 3, 2017 Abstract We propose an efficient meta-algorithm for Bayesian estimation problems that is based on low-degree polynomials, semidefinite programming, and tensor decomposition. The algorithm is inspired by recent lower bound constructions for sum-of-squares and related to the method of moments. Our focus is on sample complexity bounds that are as tight as possible (up to additive lower-order terms) and often achieve statistical thresholds or conjectured computational thresholds. Our algorithm recovers the best known bounds for community detection in the sparse stochastic block model, a widely-studied class of estimation problems for community detection in graphs. We obtain the first recovery guarantees for the mixed-membership stochastic block model (Airoldi et el.) in constant average degree graphs—up to what we conjecture to be the computational threshold for this model. We show that our algorithm exhibits a sharp computational threshold for the stochastic block model with multiple communities beyond the Kesten–Stigum bound—giving evidence that this task may require exponential time. The basic strategy of our algorithm is strikingly simple: we compute the best-possible lowdegree approximation for the moments of the posterior distribution of the parameters and use a robust tensor decomposition algorithm to recover the parameters from these approximate posterior moments. ∗ Cornell University. Supported by an NSF graduate research fellowship, a Microsoft Research PhD fellowship, a Cornell University fellowship, and David Steurer’s NSF CAREER award. Part of this work was accomplished while this author was an intern at Microsoft Research New England. † ETH Zürich. Much of this work was done while at Cornell University and the Institute for Advanced Study. Supported by a Microsoft Research Fellowship, a Alfred P. Sloan Fellowship, NSF awards, and the Simons Collaboration for Algorithms and Geometry. Contents 1 Introduction 1.1 Meta-algorithm and meta-theorems for Bayesian estimation 1.2 Detecting overlapping communities . . . . . . . . . . . . . . 1.3 Low-correlation tensor decomposition . . . . . . . . . . . . . 1.4 Information-computation gaps and concrete lower bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 3 9 14 15 2 Techniques 2.1 Approximately pairwise-independent estimators . . . . 2.2 Low-degree estimators for higher-order moments . . . 2.3 Correlation-preserving projection . . . . . . . . . . . . . 2.4 Low-correlation tensor decomposition . . . . . . . . . . 2.5 From quasi-polynomial time to polynomial time . . . . 2.6 Illustration: push-out effect in spiked Wigner matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 20 21 22 23 23 24 . . . . . . . . . . . . . . . . . . 3 Warmup: stochastic block model with two communities 26 3.1 Low-degree estimate for posterior second moment . . . . . . . . . . . . . . . . . . . . 29 4 Matrix estimation for generalized block models 31 4.1 Matrix estimation for the mixed-membership model . . . . . . . . . . . . . . . . . . . 32 4.2 Proof of Theorem 4.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 4.3 Proofs of Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 5 Tensor estimation for mixed-membership block models 5.1 Main theorem and algorithm . . . . . . . . . . . . . . . . 5.2 Low-degree estimate for posterior third moment . . . . . 5.2.1 Details of unbiased estimator . . . . . . . . . . . . 5.2.2 Details of approximate conditional independence 5.3 Cross validation . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Producing probability vectors . . . . . . . . . . . . . . . . 5.5 Remaining lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 37 43 45 46 48 50 53 6 Lower bounds against low-degree polynomials at the Kesten-Stigum threshold 53 6.1 Low-degree Fourier spectrum of the k-community block model . . . . . . . . . . . . 53 6.2 Lower bound for estimating communities . . . . . . . . . . . . . . . . . . . . . . . . . 58 7 Tensor decomposition from constant correlation 59 7.1 Proofs of Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63 7.2 Lifting 3-tensors to 4-tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 References 67 A Toolkit and Omitted Proofs 71 A.1 Probability and linear algebra tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 A.2 Tools for symmetric and Dirichlet priors . . . . . . . . . . . . . . . . . . . . . . . . . . 72 1 Introduction Bayesian1 parameter estimation [Wik17a] is a basic task in statistics with a wide range of applications, especially for machine learning. The estimation problems we study have the following form: For a known joint probability distribution p(x, θ) over data points x and parameters θ (typically both high-dimensional objects), nature draws a parameter θ ∼ p(θ) from its marginal distribution and we observe i.i.d. samples x1 , . . . , x m ∼ p(x | θ) from the distribution conditioned on θ. The goal is to efficiently estimate the underlying parameter θ from the observed samples x1 , . . . , x m . A large number of important problems in statistics, machine learning, and average-case complexity fit this description. Some examples are principal component analysis (and its many variants), independent component analysis, latent Dirichlet allocation, stochastic block models, planted constraint satisfaction problems, and planted graph coloring problems. For example, in stochastic block models the parameter θ imposes a community structure on n nodes. In the simplest case, this structure is a partition into two communities. Richer models support more than two communities and allow nodes to participate in multiple communities. The samples x1 , . . . , x m are edges between the nodes drawn from a distribution p(x | θ) that respects the community structure θ, which typically means that the edge distribution is biased toward endpoints with the same or similar communitity memberships. Taken together the samples x1 , . . . , x m form a random graph x on n vertices that exhibits a latent community structure θ; the goal is to estimate this structure θ. This problem becomes easier the more samples (i.e., edges) we observe. The question is how many samples are required such that we can efficiently estimate the community structure θ? Phrased differently: how large an average degree of the random graph x do we require to be able to estimate θ? In this work, we develop a conceptually simple meta-algorithm for Bayesian estimation problems. We focus on the regime that samples are scarce. In this regime, the goal is to efficiently compute an estimate θ̂ that is positively correlated with the underlying parameter θ given as few samples from the distribution as possible. In particular, we want to understand whether for particular estimation problems there is a difference between the sample size required for efficient and inefficient algorithms (say, exponential vs. polynomial time). In this regime, we show that our metaalgorithm recovers the best previous bounds for stochastic block models [Mas14, MNS15a, AS16a]. Moreover, for the case of richer community structures like multiple communities and especially overlapping communities, our algorithm achieves significantly stronger recovery guarantees.2 In order to achieve these improved guarantees, our meta-algorithm draws on several ideas from previous lines of work and combines them in a novel way. Concretely, we draw on ideas from recent analyses of belief propagation and their use of non-backtracking and self-avoiding random walks [Mas14, MNS15a, AS16a]. We also use ideas from recent works based on the method of moments and tensor decomposition [AGH+ 14, AGHK14, BKS15]. Our algorithm also 1Here, “Bayesian” refers to the fact that there is a prior distribution over the parameters. 2 If we represent the community structure by k vectors y1 , . . . , y k ∈ {0, 1} n that indicate community memberships, then previous algorithms [AS16a] do not aim to recover these vectors but, roughly speaking, only a random linear combination of them. While for some settings it is in fact impossible to estimate the individual vectors, we show that in many settings it is possible to estimate them (in particular for symmetric block models). 1 employs convex-programming techniques, namely the sum-of-squares semidefinite programming hierarchy, and gives a new perspective on how these techniques can be used for estimation.3 Our meta-algorithm allows for a tuneable parameter which corresponds roughly to running time. Under mild assumptions on a Bayesian estimation problem p(x, θ) (that are in particular satisfied for discrete problems such as the stochastic block model), when this parameter is set to allow the meta-algorithm to run in exponential time, if there is any estimator θ̂ of θ obtaining correlation δ, the meta-algorithm offers one obtaining correlation at least δ O(1) . While this parameter does not correspond directly to the degree paramter used in convex hiararchies such as sum of squares, the effect is similar to the phenomenon that sum of squares convex programs of exponential size can solve any combinatorial optimization problem exactly. (Since Bayesian estimation problems do not always correspond to optimization problems, this guarantee would not be obtained by sum of squares in our settings.) For many Bayesian estimation problems there is a critical number of samples n 0 such that when the number of samples n is less than n 0 , computationally-efficient algorithms seem unable to compute good estimators for θ. This is in spite of the presence of sufficient information to identify θ (and therefore estimate it with computationally inefficient algorithms), even when n < n 0 . Providing rigorous evidence for such computational thresholds has been a long-standing challenge. One popular approach is to prove impossibility of estimating θ from n < n 0 samples using algorithms from some restricted class. Such results are most convincing the chosen class captures the lowest-sample-complexity algorithms for many Bayesian inference problems, which our meta-algorithm does.4 We prove that in the k-community block model, no algorithm captured by our meta-algorithm can tolerate smaller-degree graphs than the best known algorithms. This provides evidence for a computational phase transition at the Kesten-Stigum threshold for stochastic block models. Organization In the remainder of this introduction we discuss our results and their relation to previous work in more detail. In Section 2 (Techniques) we describe the mathematical techniques involved in our meta-algorithm and its analysis, and we illustrate how to apply the meta-algorithm to recover a famous result in the theory of spiked random matrices with a much simplified proof. In Section 3 (Warmup) we re-prove (up to some loss in the running time) the result of MosselNeeman-Sly on the two-community block model as an application of our meta-algorithm, again with very simple proofs. In Section 4 (Matrix estimation) we re-interpret the best existing results on the block model, due to Abbe and Sandon, as applications of our meta-algorithm. In Section 5 (Tensor estimation) we apply our meta-algorithm to the mixed-membership block model. Following that, in Section 6 (Lower bounds) we prove that no algorithm captured by our 3 Previously, convex-programming techniques have been used in this context only as a way to obtain efficient relaxations for maximum-likelihood estimators. In contrast, our work uses convex programming to drive the method of moments approach and decompose tensors in an entropy maximizing way. 4Recent work in this area has focused on sum of squares lower bounds [HSS15, MW15, BHK+ 16]. While the sum of squares method is algorithmically powerful, it is not designed to achieve optimal sample guarantees for Bayesian estimation. Lower bounds against our meta-algorithm therefore serve better the purpose of explaining precise computational sample thresholds. 2 meta-algorithm can recover communities in the block model past the Kesten-Stigum threshold. In Section 7 (Tensor decomposition), which can be read independently of much of the rest of the paper, we give a new algorithm for tensor decomposition and prove its correctness; this algorithm is used by our meta-algorithm as a black box. 1.1 Meta-algorithm and meta-theorems for Bayesian estimation We first consider a version of the meta-algorithm that is enough to capture the best known algorithms for the stochastic block model with k disjoint communities, which we now define. Let ε, d > 0. Draw y uniformly from [k]n . For each pair i , j, add the edge {i, j} to a graph on n vertices with probability (1 + (1 − 1k )ε) nd if yi  y j and (1 − εk ) nd otherwise. The resulting graph has expected average degree d. A series of recent works has explored the problem of estimating y in these models for the sparsest-possible graphs. The emerging picture, first conjectured via techniques from statistical physics in the work [DKMZ11], is that in the k-community block model it is possible to recover a 2 nontrivial estimate of y via a polynomial time algorithm if and only if d  (1 + δ) εk 2 for δ > Ω(1). This is called the Kesten-Stigum threshold. The algorithmic side of this conjecture was confirmed by [Mas14, MNS15a] for k  2 and [AS16a] for general k. One of the goals of our meta-algorithm is that it apply in a straightforward way even to complex Bayesian estimation problems. A more complex model (yet more realistic for real-world networks) is the mixed-membership block model [ABFX08] which we now define informally. Let α > 0 be n k(α+1) an overlap parameter. Draw y from kt , where t  k+α ≈ α + 1; that is for each of n nodes pick a set S j of roughly α + 1 communities.5 For each pair i, j, add an edge to the graph with |S i ∩S j | probability (1 + ( t 2 − 1k )ε) nd . (That is, with probability which increases as i and j participate in more communities together.) In the limit α → 0 this becomes the k-community block model. Returning to the meta-algorithm (but keeping in mind the block model), let p(x, y) be a joint probability distribution over observable variables x ∈ ’n and hidden variables y ∈ ’m . Nature draws (x, y) from the distribution p, we observe x and our goal is to provide an estimate ŷ(x) for 2 y. Often the mean square error p(x,y) ŷ(x) − y is a reasonable measure for the quality of the estimation. For this measure, the information-theoretically optimal estimate is the mean of the posterior distribution ŷ(x)  p(y |x) y. This approach has two issues that we address in the current work. The first issue is that naively computing the mean of the posterior distribution takes time Í exponential in the dimension of y. For example, if y ∈ {±1} m , then p(y |x) y  y∈{±1} m y · p(y | x); there are 2m terms in this sum. There are many well-known algorithmic approaches that aim to address this issue or related ones, for example, belief propagation [Gal62, Pea82] or expectation maximization [DLR77]. While these approaches appear to work well in practice, they are notoriously difficult to analyze. 5In actuality one draws for each node i ∈ [n] a probability vector σi ∈ ∆k−1 from the Dirichlet distribution with parameter α; we describe a nearly-equivalent model here for the sake of simplicity—see Section 1.2 for details. Our guarantees for recovery in the mixed-membership model also apply to the model here because it has the same second moments as the Dirichlet distribution. 3 In this work, we can resolve this issue in a very simple way: We analytically determine a 2 low-degree polynomial f (x) so that p(x,y) f (x) − y is as small as possible and use the fact that low-degree polynomials can be evaluated efficiently (even for high dimensions n).6 Because the maximum eigenvector of an n-dimensional linear operator with a spectral gap is an O(log n)-degree polynomial of its entries, our meta-algorithm captures spectral properties of linear operators whose entries are low-degree polynomials of observable variables x. Examples of such operators include adjacency matrices (when x is a graph), empirical covariance matrices (when x is a list of vectors), as well as more sophisticated objects such as linearized belief propagation operators (e.g., [AS15]) and the Hashimoto non-backtracking operator. The second issue is that even if we could compute the posterior mean exactly, it may not contain any information about the hidden variable y and the mean square error is not the right measure to assess the quality of the estimator. This situation typically arises if there are symmetries in the posterior distribution. For example, in the stochastic block model with two communities we have p(y |x) y  0 regardless of the observations x because p(y | x)  p(−y|x). A simple way to resolve this issue is to estimate higher-order moments of the hidden variables. For stochastic block models with disjoint communities, the second moment p(y |x) y yT would suffice. (For overlapping communities, we need third moments p(y |x) y ⊗3 due to more substantial symmetries.) For now, we think of y as an m-dimensional vector and x as an n-dimensional vector (in the blockmodel on N nodes, this would correspond to m ≈ kN and n  N 2 ). Our algorithms follow a two-step strategy: 1. Given x ∼ p(x| y), evaluate a fixed, low-degree polynomial P(x) taking values in (’m )⊗ℓ . (Usually ℓ is 2 or 3.) 2. Apply a robust eigenvector or semidefinite-programming based algorithm (if ℓ  2), or a robust tensor decomposition algorithm (if ℓ  3 or higher) to P to obtain an estimator ŷ for y. The polynomial P(x) should be an optimal low-degree approximation to y ⊗ℓ , in the following sense: if n is sufficiently large that some low-degree polynomial Q(x) has constant correlation with y ⊗ℓ hQ, y ⊗ℓ i > Ω(1) · ( kQ k 2)1/2 ( k y ⊗ℓ k 2 )1/2 , x,y x then P has this guarantee. (The inner products and norms are all Euclidean.) A prerequisite for applying our meta-algorithm to a particular inference problem p(x, y) is  ⊗ℓ  that it be possible to estimate y given y | x for some constant ℓ. For such a problem, the optimal Bayesian inference procedure (ignoring computational constraints) can be captured  ⊗ℓ  by computing F(x)  y | x , then using it to estimate y. When p(x, y) is such that it is information-theoretically possible to estimate y from x, these posterior moments will generally satisfy hF(x), y ⊗ℓ i > Ω(1) · ( kF(x)k 2 )1/2 ( k y ⊗ℓ k 2 )1/2 , for some constant ℓ. Our observation is that when F is approximately a low-degree function, this estimation procedure can be carried out via an efficient algorithm. 6Our polynomials typically have logarithmic degree and naive evaluation takes time n O(log n) . However, we show that under mild conditions it is possible to approximately evaluate these polynomials in polynomial time using the idea of color coding [AYZ95]. 4 Matrix estimation and prior results for block models In the case ℓ  2, where one uses the  T  covariance y y | x to estimate y, the preceding discussion is captured by the following theorem. Theorem 1.1 (Bayesian estimation meta-theorem—2nd moment). Let δ > 0 and p(x, y) be a distriO(1) bution over vectors x ∈ {0, 1} n and unit vectors y ∈ ’d . Assume p(x) > 2−n for all x ∈ {0, 1} n .7 Suppose there exists a matrix-valued degree-D polynomial P(x) such that T p(x,y) hP(x), y y i > δ ·  p(x) kP(x)kF2  1/2 . (1.1) Then, there exists δ′ > δ O(1) > 0 and an estimator ŷ(x) computable by a circuit of size n O(D) such that p(x,y) h ŷ(x), yi 2 > δ′ . (1.2) To apply this theorem to the previously-discussed setting of samples x1 , . . . , x N generated from p(x | y), assume the samples x1 , . . . , x N are in some fixed way packaged into a single n-length vector x. One curious aspect of the theorem statement is that it yields a nonuniform algorithm—a family of circuits—rather than a uniform algorithm. If the coefficients of the polynomial P can themselves be computed in polynomial time, then the conclusion of the algorithm is that an n O(D) -time algorithm exists with the same guarantees. As previously mentioned, the meta-algorithm has a parameter D, the degree of the polynomial P. If D  n, then whenever it is information-theoretically possible to estimate y from [y y ⊤ | x], the meta-algorithm can do so (in exponential time). This follows from the fact that every function in n Boolean variables is a polynomial of degree at most n. It is also notable that, while a degree D polynomial can be evaluated by an n O(D) -size circuit, some degree-D polynomials can be evaluated by much smaller circuits. We exploit such polynomials for the block model (computable via color coding), obtaining n O(1) -time algorithms from degree log n polynomials. By using very particular polynomials, which can be computed via powers of non-backtracking operators, previous works on the block model are able to give algorithms with near-linear running times [MNS15a, AS16a].8 Using the appropriate polynomial P, this theorem captures the best known guarantees for partial recovery in the k-community stochastic block model. Via the same polynomial, applied in the mixed-membership setting, it also yields our first nontrivial algorithm for the mixed-membership model. However, as we discuss later, the recovery guarantees are weak compared to our main theorem. 7This mild condition on the marginal distribution of x allows us to rule out pathological situations where a lowdegree polynomial in x may be hard to evaluate accurately enough because of coefficients with super-polynomial bit-complexity. 8In this work we choose to work with self-avoiding walks rather than non-backtracking ones; while the corresponding polynomials cannot to our knowledge be evaluated in near-linear time, the analysis of these polynomials is much simpler than the analysis needed to understand non-backtracking walks. This helps to make the analysis of our algorithms much simpler than what is required by previous works, at the cost of large polynomial running times. It is an interesting question to reduce the running times of our algorithm for the mixed-membership block model to near-linear via non-backtracking walks, but since our aim here is to distinguish what is computable in polynomial time versus, say, exponential time, we do not pursue that improvement here. 5 Recalling the ε, d, k block model from the previous section, let y ∈ ’n be the centered indicator vector of, say, community 1. Theorem 1.2 (Implicit in [Mas14, MNS15a, AS16a], special case of our main theorem, Theorem 1.4). def k 2 (α+1)2 Let δ  1 − ε2 d . If x is sampled according to the n-node, k-community, ε-biased, α-mixed-membership block model with average degree d and y is the centered indicator vector of community 1, there is a n×n-matrix valued polynomial P of degree O(log n)/δ O(1) such that  δ hP(x), y y i > x k(α + 1) T  O(1) ( kP(x)k 2 )1/2 ( k y yT k 2 )1/2 . Together with Theorem 1.1, up to questions of n O(log n) versus n O(1) running times, when α → 0 this captures the previous best efficient algorithms for the k-community block model. (Once one has a unit vector correlated with y, it is not hard to approximately identify the vertices in community 1.) While the previous works [Mas14, MNS15a, AS16a] did not consider the mixedmembership blockmodel, this theorem is easily obtained using techniques present in those works (as we show when we rephrase those works in our meta-algorithm, in Section 4).9 Symmetries in the posterior, tensor estimation, and improved error guarantees We turn next to our main theorem on the mixed-membership model, which offers substantial improvement on the correlation which can be obtained via Theorem 1.2. The matrix-based algorithm discussed above, Theorem 1.2, contains a curious asymmetry; namely the arbitrary choice of community 1. The block model distributions are symmetric under relabeling of the communities, which means that any estimator P(x) of y yT is also an estimator of y ′ y ′T , where y ′ is the centered indicator of community j > 1. Since one wants to estimate all the vectors y1 , . . . , yk (with yi corresponding to the i-th community), it is more natural to consider the polynomial P to be an estimator of the Í matrix M  i∈[k] yi yiT .10 Unsurprisingly, P is a better estimator of M than it is of y1 . In fact, with the same notation as in the theorems, x,y hP(x), M(y)i > δ O(1) ( kP(x)k 2 )1/2 ( kM(y)k 2)1/2 , removing the k O(1) factor in the denominator. This guarantee is stronger: now the error in the estimator depends only on the distance δ of the parameters ε, d, k, α from the critical threshold k 2 (α+1)2 ε2 d  1 rather than additionally on k. If given the matrix M exactly, one way to extract an estimator ŷi for some yi is just to sample a random unit vector in the span of the top k eigenvectors of M. Such an estimator ŷi would have 9In fact, if one is willing to lose an additional 2−k in the correlation obtained in this theorem, one can obtain a similar result for the mixed-membership model by reducing it to the disjoint-communities with K ≈ 2k communities, one for each subset of k communities. This works when each node participates in a subset of communities; if one uses the Dirichlet version of the mixed-membership model then suitable discretization would be necessary. 10In more general versions of the blockmodel studied in [AS16a], where each pair i, j of communities may have a different edge probability Q i j it is not always possible to estimate all of y1 , . . . , y k . We view it as an interesting open problem to extract as much information about y1 , . . . , y k as possible in that setting; the guarantee of [AS16a] amounts, roughly, to finidng a single vector in the linear span of y1 , . . . , y k . 6 1 h ŷi , yi i 2 > k O(1) k yi k, recovering the guarantees of the theorems above but not offering an estimator ŷi whose distance to yi depends only on the distance δ above the critical threshold. Indeed, without exploiting additional structure of the vectors yi is unclear how to remove this 1/k O(1) factor. As Í a thought experiment, if one had the matrix M ′  i 6 k a i a iT , where a1 , . . . , a k were random unit vectors, then a1 , . . . , a k would be nearly orthonormal and one could learn essentially only their linear span. (From the linear span it is only possible to find aˆi with correlation h aˆi , a i i 2 > 1/k O(1) .) In the interest of generality we would like to avoid using such additional structure: while in the disjoint-community model the vectors yi have disjoint support (after un-centering them), no such special structure is evident in the mixed-membership setting. Indeed, when α is comparable to k, the vectors yi are similar to independent random vectors of the appropriate norm. To address this issue we turn to tensor methods. To illustrate the main idea simply: if a1 , . . . , a k Í are orthonormal, then it is possible to recover a1 , . . . , a k exactly from the 3-tensor T  i 6 k a i⊗3 . More abstractly, the meta-algorithm which uses 3rd moments is able to estimate hidden variables whose posterior distributions have a high degree of symmetry, without errors which worsen as the posteriors become more symmetric. Theorem 1.3 (Bayesian estimation meta-theorem—3rd moment). Let p(x, y1 , . . . , yk ) be a joint distribution over vectors x ∈ {0, 1} n and exchangable,11 orthonormal12 vectors y1 , . . . , yk ∈ ’d . Assume the O(1) marginal distribution of x satisfies p(x) > 2−n for all x ∈ {0, 1} n .13 Suppose there exists a tensor-valued degree-D polynomial P(x) such that p(x,y1 ,...,y k ) hP(x), k Õ i1 yi⊗3 i > δ·  p(x) kP(x)k 2  1/2 · √ k. (1.3) √ (Here, k·k is the√norm induced by the inner product h·, ·i. The factor k normalizes the inequality because Í k ki1 yi⊗3 k  k by orthonormality.) Then, there exists δ′ > δ O(1) > 0 and a circuit of size n O(D) that given x ∈ {0, 1} n outputs a list of unit vectors z1 , . . . , z m with m 6 n poly(1/δ) so that max hyi , z j i 2 > δ′ . p(x,y1 ,...,y k ) i∼[k] j∈[m] (1.4) That the meta-algorithm captured by this theorem outputs a list of n 1/poly(δ) vectors rather than just k vectors is an artifact of the algorithmic difficulty of multilinear algebra as compared to linear algebra. However, in most Bayesian estimation problems it is possible by using a very small number of additional samples (amounting to a low-order additive term in the total sample complexity) to cross-validate the vectors in the list z1 , . . . , z m and throw out those which are not correlated with some y1 , . . . , yk . Our eventual algorithm for tensor decomposition (see Section 1.3 11 Here, exchangeable means that for every x ∈ {0, 1} n and every permutation π : [k] → [k], we have p(y1 , . . . , y k | x)  p(y π(1) , . . . , y π(k) | x). 12 Here, we say the vector-valued random variables y1 , . . . , y k are orthonormal if with probability 1 over the distribution p we have hy i , y j i  0 for all i , j and k y i k 2  1. 13As in the previous theorem, this mild condition on the marginal distribution of x allows us to rule out pathological situations where a low-degree polynomial in x may be hard to evaluate accurately enough because of coefficients with super-polynomial bit-complexity. 7 and Section 7) bakes this step in by assuming access to an oracle which evaluates the function Í v 7→ i∈[k] hv, yi i 4 . A key component of the algorithm underlying Theorem 1.3 is a new algorithm for very robust orthogonal tensor decomposition.14 Previous algorithms for tensor decomposition require that the input tensor is close (in an appropriate norm) to only one orthogonal tensor. By contrast, our tensor decomposition algorithm is able to operate on a tensor T which is just δ ≪ 1 correlated to the Í orthogonal tensor yi⊗3 , and in particular might also be δ-correlated with 1/δ other orthogonal tensors. If one views tensor decomposition as a decoding task, taking a tensor T and decoding it into its rank-one components, then our guarantees are analogous to list-decoding. Our algorithm in this setting involves a novel entropy-maximization program which, among other things, ensures that given a tensor T which for example is δ-correlated with two distinct orthogonal tensors A and B, the algorithm produces a list of vectors correlated with both the components of A and those of B. Applying this meta-theorem (plus a simple cross-validation scheme to prune the vectors in the 1/poly(δ) n -length list) to the mixed-membership block model (and its special case, the k-disjointcommunities block model) yields the following theorem. (See Section 1.2 for formal statements.) Theorem 1.4 (Main theorem on the mixed-membership block model, informal). Let ε, d, k, α be k 2 (α+1)2 paramters of the mixed-membership block model, and let δ  1 − ε2 d > Ω(1). Let yi be the centered indicator vector of the i-th community. There is an n 1/poly(δ) -time algorithm which, given a sample x from the ε, d, k, α block model, recovers vectors ŷ1 (x), . . . , ŷk (x) such that there is a permutation π : [k] → [k] with h ŷπ(i) , yi i 2 > δ O(1) ( k ŷπ(i) k 2 )1/2 ( k yi k 2 )1/2 . The eventual goal, as we discuss in Section 1.2, is to label each vertex by a probability vector τi which is correlated with the underlying label σi , but given the ŷ vectors from this theorem this is easily accomplished. Comparison to the sum of squares method The sum of squares method has been recently been a popular approach for designing algorithms for Bayesian estimation [BKS15, HSS15, RRS16, GM15]. The technique works best in settings where the maximum-likelihood estimator can be phrased as a polynomial optimization problem (subject to semialgebraic constraints). Then the strategy is to use the sum of squares method to obtain a strong convex relaxation of the maximum-likelihood problem, solve this relaxation, and round the result. This strategy has been quite successful, but thus far it does not seem to allow the sharp (up to low-order additive terms) sample-complexity guarantees we study here. (Indeed, for some problems, including the stochastic block model, it is not clear that maximum likelihood estimation recovers those guarantees, much less the SoS-relaxed version.) One similarity between our algorithms and these applications of sum of squares is that the rounding procedures used at the end often involve tensor decomposition, which is itself often done 14An orthogonal 3-tensor is Ím ⊗3 i1 a i , where a1 , . . . , a m are orthonormal. 8 via the sum of squares method. We do employ the SoS algorithm as a black box to solve tensor decomposition problems for versions of our algorithm which use higher moments. Recent works on SoS show that the low-degree polynomials computed by our meta-algorithm are closely connected to lower bounds for the SoS hierarchy, though this connection remains far from fully understood. The recent result [BHK+ 16] on the planted clique problem first discovered this connection. The work [HKP+ 17] (written concurrently with the present paper) shows that this connection extends far beyond the planted clique setting. Comparison to the method of moments Another approach for designing statistical estimators for provable guarantees is the method of moments. Typically one considers parameters θ (which need not have a prior distribution p(θ)) and iid samples x1 , . . . , x n ∼ p(x|θ). Generally one shows that the moments of the distribution {x|θ} are related to some function of θ: for example perhaps [xx ⊤ | θ]  f (θ). Then one uses the samples x i to estimate the moment M  [xx ⊤ | θ], and finally to estimate θ by f −1 (M). While the method of moments is quite flexible, for the high-noise problems we consider here it is not clear that it can achieve optimal sample complexity. For example, in our algorithms (and existing sample-optimal algorithms for the block model) it is important to exploit the flexibility to compute any polynomial of the samples jointly—given n samples our algorithms can evaluate a polynomial P(x1 , . . . , x n ), and P often will not be an empirical average of some simpler function like Í i 6 n q(x i ). The best algorithm for the mixed-membership block model before our work uses the method of moments and consequently requires much denser graphs than our method [AGHK14]. 1.2 Detecting overlapping communities We turn now to discuss our results for stochastic block models in more detail and compare them to the existing literature. The stochastic block model is a widely studied (family of) model(s) of random graphs containing latent community structure. It is most common to study the block model in the sparse graph setting: many large real-world networks are sparse, and the sparse graph setting is nearly always more mathematically challenging than the dense setting. A series of recent works has for the first time obtained algorithms which recover communities in block model graphs under (conjecturally) optimal sparsity conditions. For an excellent survey, see [Abb17]. Such sharp results remain limited to relatively simple versions of the block model; where, in particular, each vertex is assigned a single community in an iid fashion. A separate line of work has developed more sophisticated and realistic random graph models with latent community structure, with the goal of greater applicability to real-life networks. The mixed-membership stochastic block model [ABFX08] is one such natural extension of the stochastic block model that allows for communities to overlap, as they do in large networks found in the wild. In addition to the number of vertices n, the average degree d, the correlation parameter ε, and the number of communities k, this model has an overlap parameter α > 0 that controls how many communities a typical vertex participates in. Roughly speaking, the model generates an n-vertex 9 graph by choosing k communities as random vertex subsets of size (1 + α)n/k and choosing dn/2 random edges, favoring pairs of vertices that have many communities in common. Definition 1.5 (Mixed-membership stochastic block model). The mixed-membership stochastic block model SBM(n, d, ε, k, α) is the following distribution over n-vertex graphs G and kdimensional probability vectors σ1 , . . . , σn for the vertices: • draw σ1 , . . . , σn independently from Dir(α) the symmetric k-dimensional Dirichlet distribution with parameter α > 0,15   • for every potential edge {i, j}, add it to G with probability d n · 1 + hσi , σ j i − 1 k  ε . Due to symmetry, hσi , σ j i has expected value 1k , which means that the expected degree of every vertex in this graph is d. In the limit α → 0, the Dirichlet distribution is equivalent to the uniform distribution over coordinate vectors 11 , . . . , 1k and the model becomes SBM(n, d, ε, k), the stochastic block model with k disjoint communities. For α  k, the Dirichlet distribution is uniform over the open (k − 1)-simplex [Wik17b]. For general values of α, a probability vector from Dir(α) 1 + 1k , which means that we can think turns out to have expected collision probability (1 − 1k ) α+1 of the probability vector being concentrated on about α + 1 coordinates.16 This property of the Dirichlet distribution is what determines the threshold for our algorithm. Correspondingly, our algorithm and analysis extends to a large class of distributions over probability vectors that share this property. Measuring correlation with community structures In the constant-average-degree regime of the block model, recovering the label of every vertex correctly is information-theoretically impossible. For example, no information is present in a typical sample about the label of any isolated vertex, and in a typical sample a constant fraction of the vertices are isolated. Instead, at least in the k-disjoint-community setting, normally one looks to label vertices by labels 1, . . . , k so that (up to a global permutation), this labeling has positive correlation with the true community labels. bj |, When the communities are disjoint, one can measure such correlation using the sizes of |S j ∩ S bj is an estimated set of nodes in community where S j ⊆ [n] is the set of nodes in community j and S j. The original definition of overlap, the typical measure of labeling-accuracy in the constant-degree regime, takes this approach [DKMZ11]. For present purposes this definition must be somewhat adapted, since in the mixedmembership block model there is no longer a good notion of a discrete set of nodes S j for each community j ∈ [k]. We define a smoother notion of correlation with underlying community labels to accommodate that the labels σi are vectors in ∆k−1 . In discrete settings, for example when α → 0 (in which case one recovers the k-disjoint-community model), or more generally when each σi is the uniform distribution over some number of communities, our correlation measure recovers the usual notion of overlap. 15In the symmetric k-dimensional Dirichlet distribution with parameter α > 0, the probability of a probability vector Î σ is proportional to kt1 σ(t)α/k−1 . By passing to the limit, we define Dir(0) to be the uniform distribution over the coordinate vectors 11 , . . . , 1k . k 16When k and α are comparable in magnitude, it is important to interpret this more accurately as (α + 1) · k+α coordinates. 10 Let σ  (σ1 , . . . , σn ) and τ  (τ1 , . . . , τn ) be labelings of the vertices 1, . . . , n by by k-dimensional probability vectors. We define the correlation corr(σ, τ) as max π i∼n hσi , τπ(i) i − 1 k (1.5) where π ranges over permutations of the k underlying communities. This notion of correlation is closely related to the overlap of the distributions σi , τi . To illustrate this notion of correlation, consider the case of disjoint communities (i.e., α  0), where the ground-truth labels τi are indicator vectors in k dimensions. Then, if i hσi , τπ(i) i − 1k > δ, by looking at the large coordinates of σi it is possible to correctly identify the community memberships of a δ O(1) + 1k fraction of the vertices, which is a δ O(1) fraction more than would be identified by randomly assigning labels to the vertices without looking at the graph. When the ground truth labels τi are spread over more than one coordinate—say, for example, they are uniform over t coordinates—the best recovery algorithm cannot find σ’s with correlation better than 1 1 corr(σ, τ)  − , t k which is achieved by σ  τ. This is because in this case τ has collision probability hτ, τi  1t . Main result for mixed-membership models The following theorem gives a precise bound on the number of edges that allows us to find in polynomial time a labeling of the vertices of an n-node mixed membership block model having nontrivial correlation with the true underlying labels. Here, the parameters d, ε, k, α of the mixed-membership stochastic block model may even depend on the number of vertices n. Theorem 1.6 (Mixed-membership SBM—significant correlation). Let d, ε, k, α be such that k 6 n o(1) , k 2 (α+1)2 def > 0. (Equivalently for small δ, suppose α 6 n o(1) , and ε 2 d 6 n o(1) . Suppose δ  1 − ε2 d 2 2 2 ′ O(1) ε d > (1 + δ) · k (α + 1) .) Then, there exists δ > δ > 0 and an n 1/poly(δ) -time algorithm that given an n-vertex graph G outputs τ1 , . . . , τn ∈ ∆k−1 such that  1 1 − corr(σ, τ) > δ · t k (G,σ)∼SBM(n,d,ε,k,α) ′  (1.6) k where t  (α + 1) · k+α (samples from the α, k Dirichlet distribution are roughly uniform over t out of k coordinates). In particular, as δ → 1 we have corr(σ, τ) → 1t − 1k , while corr(σ, σ)  1t − 1k . Note that in the above theorem, the correlation δ′ that our algorithm achieves depends only on δ (the distance to the threshold) and in particular is independent of n and k (aside from, for the latter, the dependence on k via δ). For disjoint communities (α  0), our algorithm achieves constant correlation with the planted labeling if ε 2 d/k 2 is bounded away from 1 from below. We conjecture that the threshold achieved by our algorithm is best-possible for polynomial-time algorithms. Concretely, if d, ε, k, α are constants such that ε 2 d < k 2 (α + 1)2 , then we conjecture that for every polynomial-time algorithm that given a graph G outputs τ1 , . . . , τn ∈ ∆k−1 , lim n→∞ (G,σ)∼SBM(n,d,ε,k,α) 11 corr(σ, τ)  0 . (1.7) This conjecture is a natural extension of a conjecture for disjoint communities (α  0), which says that beyond the Kesten–Stigum threshold, i.e., ε 2 d < k 2 , no polynomial-time algorithm can achieve correlation bounded away from 0 with the true labeling [Moo17]. For large enough values of k, this conjecture predicts a computation-information gap because the condition ε 2 d > Ω(k log k) is enough for achieving constant correlation information-theoretically (and in fact by a simple exponential-time algorithm). We discuss these ideas further in Section 1.4. Comparison with previous matrix-based algorithms We offer a reinterpretation in our metaalgorithmic framework of the algorithms of Mossel-Neeman-Sly and Abbe-Sandon. This will permit us to compare our algorithm for the mixed-membership model with what could be achieved by the methods in these prior works, and to point out one respect in which our algorithm improves on previous ones even for the disjoint-communities block model. The result we discuss here is a slightly generalized version of Theorem 1.2. Let U be a (possibly infinite or continuous) universe of labels, and let W assign to every x, y ∈ U a nonnegative real number W(x, y)  W(y, x) > 0. Let µ be a probability distribution on U, which induces the inner product of functions f , 1 : U → ’ given by h f , 1i  x∼µ f (x)1(x). The function W can be considered as linear operator on { f : U → ’}, and under mild assumptions it has eigenvalues λ1 , λ2 , . . . with respect to the inner product h·, ·i. The pair µ, W along with an average degree parameter d induce a generalized stochastic block model, where labels for nodes are drawn from µ and an edge between a pair of nodes with labels x and y is present with probability nd · W(x, y). When U is ∆k−1 and µ is the Dirichlet distribution, this captures the mixed-membership block model. Assume λ 1  1 and that µ and W are sufficiently nice (see Section 4 for all the details). Then one can rephrase results of Abbe and Sandon in this setting as follows. Theorem 1.7 (Implicit in [AS16a]). Suppose the operator W has eigenvalues 1  λ1 > λ 2 > · · · > λ r def (each possibly with higher multiplicity) and δ  1 − dλ1 2 > 0. Let Π be the projector to the second eigenspace 2 of the operator W. For types x1 , . . . , x n ∼ U, let A ∈ ’n×n be the random matrix A i j  Π(x i , x j ), where we abuse notation and think of Π : U × U → ’. There is an algorithm with running time n poly(1/δ) which outputs an n × n matrix P such that for x, G ∼ G(n, d, W, µ), x,G Tr P · A > δ O(1) · ( x,G kAk 2)1/2 ( x,G kP k 2 )1/2 . In one way or another, existing algorithms for the block model in the constant-degree regime are all based on estimating the random matrix A from the above theorem, then extracting from an estimator for A some labeling of vertices by communities. In our mixed-membership setting, Í one may show that the matrix A is s∈[k] v s v sT , where v s ∈ ’n has entries v s (i)  σi (s) − 1k . Furthermore, as we show in Section 4, the condition dλ22 > 1 translates for the mixed-membership model to the condition ε 2 d > k(α + 1)2 , which means that under the same hypotheses as our main theorem on the mixed-membership model it is possible in polynomial time to evaluate a Í constant-correlation estimator for s∈[k] v s v sT . As we discussed in Section 1.1, however, extracting estimates for v1 , . . . , v k (or, almost equivalently, estimates for σ1 , . . . , σn ) from this matrix seems 12 to incur an inherent 1/k loss in the correlation. Thus, the final guarantee one could obtain for the mixed-membership block model using the techniques in previous work would be estimates  O(1) .17 We avoid this loss in our main theorem τ1 , . . . , τn for σ1 , . . . , σn such that corr(σ, τ) > δk via tensor methods. Although this 1/k multiplicative loss in the correlation with the underlying labeling is not inherent in the disjoint-community setting (roughly speaking this is because the matrix A is a 0/1 block-diagonal matrix), previous algorithms nonetheless incur such loss. (In part this is related to the generality of the work of Abbe and Sandon: they aim to allow W where A might only have rank one, while in our settings A always has rank k − 1. For low-rank A this 1/k loss is probably necessary for polynomial time algorithms.) Thus our main theorem on the mixed membership model offers an improvement on the guarantees in the previous literature even for the disjoint-communities setting: when W only has entries 1 − ε and ε we obtain a labeling of the vertices whose correlation with the underlying labeling depends only on δ. This allows the number k of communities to grow with n without incurring any loss in the correlation (so long as the average degree of the graph grows accordingly). For further discussion of the these results and a proof of the above theorem, see Section 4. Comparison to previous tensor algorithm for mixed-membership models Above we discussed a reinterpretation (allowing a continuous space U of labels) of existing algorithms for the constantaverage-degree block model which would give an algorithm for the mixed-membership model, and discussed the advantages of our algorithm over this one. Now we turn to algorithms in the literature which are specifically designed for stochastic block models with overlapping communities. The best such algorithm requires ε 2 d > O(log n)O(1) · k 2 (α + 1)2 [AGHK13]. Our bound saves the O(log n)O(1) factor. (This situation is analogous to the standard block model, where simpler algorithms based on eigenvectors of the adjacency matrix require the graph degree to be Í logarithmic.) Notably, like ours this algorithm is based on estimating the tensor T  s∈[k] v s⊗3 , where v s ∈ ’n has entries v s (i)  σi (s) − 1k . However, the algorithm differs from ours in two key respects. 1. The algorithm [AGHK13] estimates the tensor T using a 3-tensor analogue of a high power of the adjacency matrix of an input graph, while we use self-avoiding walks (which are rather like a tensor analogue of the nonbacktracking operator). 2. The tensor decomposition algorithm used in [AGHK13] to decompose the (estimator for the) tensor T tolerates much less error than our tensor decomposition algorithm; the result is that a higher-degree graph is needed in order to obtain a better estimator for the tensor T. The setting considered by [AGHK13] does allow a more sophisticated version of the Dirichlet distribution than we allow, in which different communities have different sizes. It is an interesting open problem to extend the guarantees of our algorithm to that setting. 17In fact, it is not clear one can obtain even this guarantee using strictly matrix methods. Strictly speaking, in estimating, say, v1 using the above described matrix method, one obtains a unit vector v such that hv, v1 i 2 > Ω(1) · kv1 k 2 . Without knowing whether v or −v is the correct vector it is not clear how to transform estimates for the v s ’s to estimates for the σ’s. However, matrix methods cannot distinguish between v s and −v s . In our main algorithm we avoid this issue Í because the 3rd moments v s⊗3 are not sign-invariant. 13 1.3 Low-correlation tensor decomposition Tensor decomposition is the following problem. For some unit vectors a1 , . . . , a m ∈ ’n and a Í ⊗k constant k (often k  3 or 4), one is given the tensor T  m i1 a i + E, where E is some error tensor. The goal is to recover vectors b 1 , . . . , b m ∈ ’n which are as close as possible to a1 , . . . , a m . Tensor decomposition has become a common primitive used by algorithms for parameter learning and estimation problems [CJ10, AGH+ 14, GHK, GVX14, BKS15, MSS16, SS17]. In the simplest examples, the hidden variables are orthogonal vectors a1 , . . . , a m and there is a simple Í function of the observed variables which estimates the tensor i 6 m a i⊗k (often an empirical k-th moment of observed variables suffices). Applying a tensor decomposition algorithm to such an estimate yields estimates of the vectors a1 , . . . , a m . We focus on the case that a1 , . . . , a m are orthonormal. Algorithms for this case are already useful for a variety of learning problems, and it is often possible to reduce more complicated problems to the orthonormal case using a small amount of side information about a1 , . . . , a m (in Í T particular their covariance m i1 a i a i ). In this setting the critical question is: how much error E (and measured in what way) can the tensor decomposition algorithm tolerate and still produce useful outputs b 1 , . . . , b m ? When we use tensor decomposition in our meta-algorithm, the error E will be incurred when Í ⊗k estimating m i1 a i from observable samples. Using more samples would decrease the magnitude Ím ⊗k of T − i1 a i , but because our goal is to obtain algorithms with optimal sample complexity we need a tensor decomposition algorithm which is robust to greater errors than those in the existing literature. Our main theorem on tensor decomposition is the following. Theorem 1.8 (Informal). For every δ > 0, there is a randomized algorithm with running time n 1/poly(δ) that given a 3-tensor T ∈ (’n )⊗3 and a parameter k outputs n poly(1/δ) unit vectors b 1 , . . . , b m with the √ Ík ⊗3 following property: if T satisfies hT, i1 a i i > δ · kT k · k for some orthonormal vectors a1 , . . . , a k , then max ha i , b j i 2 > δ O(1) . i∼[k] j∈[m] Furthermore, if the algorithm is allowed to make n 1/poly(δ) calls to an oracle O which correctly answers Í 4 O(1) ?”, then it outputs just k orthonormal queries of the form “does the unit vector v satisfy m i1 ha i , vi > δ vectors, b 1 , . . . , b k such that there is a permutation π : [k] → [k] with i∈[k] ha i , b π(i) i 2 > δ O(1) . (These guarantees hold in expectation over the randomness used in the decomposition algorithm.) (For a more formal statement, and in particular the formal requirements for the oracle O, see Section 7.) √ Í Rescaling T as necessary, one may reinterpret the condition hT, ki1 a i⊗3 i > δ · kT k · k as √ Í Í ⊗3 k/δ and k · k is the Euclidean norm. T  ki1 a i⊗3 + E, where hE, m i1 a i i  0 and kEk 6 In particular, E may have Euclidean norm which is a large constant factor 1/δ larger than the 14 Í ⊗3 Euclidean norm of the tensor m i1 a i that the algorithm is trying to decompose! (One way such error could arise is if T is actually correlated with 1/δ unrelated orthogonal tensors; our algorithm in that case ensures that the list of outputs vectors is correlated with every one of these orthogonal tensors.) In all previous algorithms of which we are aware (even for the case of orthogonal a1 , . . . , a m ), the error E must have spectral norm (after flattening to an n 2 × n 2 matrix) at most ε for ε < 21 ,18 √ or E must have Euclidean norm at most ε m [SS17]. The second requirement is strictly stronger than ours (thus our algorithm has weaker requirements and so stronger guarantees). The first, on the spectral norm of E when flattened to a matrix, is incomparable to the condition in our theorem. However, when E satisfies such a spectral bound it is possible to decompose T using (sophisticated) spectral methods [MSS16, SS17]. In our setting such methods seem unable to avoid producing only vectors b which are correlated with E but not with any a1 , . . . , a m . In other words, such methods would overfit to the error E. To avoid this, our algorithm uses a novel maximum-entropy convex program (see Section 7 for details). One a priori unusual requirement of our tensor decomposition algorithm is access to the oracle O. In any tensor decomposition setting where E satisfies kEkin j  max kxk1 hE, x ⊗3 i 6 o(1), the Í oracle O can be implemented just be evaluating hT, v ⊗3 i  ki1 ha i , vi 3 + o(1). All previous works on tensor decomposition of which we are aware either assume that the injective norm kEkin j is bounded as above, or (as in [SS17]) can accomplish this with a small amount of preprocessing on the tensor T. Our setting allows, for example, E  100 · v ⊗3 for some unit vector v, and does not appear to admit the possibility of such preprocessing, hence the need for an auxiliary implementation of O. In our learning applications we are able to implement O by straightforward holdout set/cross-validation methods. 1.4 Information-computation gaps and concrete lower bounds The meta-algorithm we offer in this paper is designed to achieve optimal sample complexity among polynomial-time algorithms for many Bayesian estimation problems. It is common, though, that computationally inefficient algorithms can obtain accurate estimates of hidden variables with fewer samples than seem to be tolerated by any polynomial-time algorithm. This appears to be true for the k-community stochastic block model we have used as our running example here: efficient algorithms seem to require graphs of average degree d > k 2 /ε 2 to estimate communities, while inefficient algorithms are known which tolerate d of order k log k [AS16b, Moo17]. Such phenomena, sometimes called information-computation gaps, appear in many other Bayesian estimation problems. For example, in the classical planted clique problem [Jer92, Kuc95], a clique of size k > (2 + ε) log n is randomly added to a sample G ∼ G(n, 12 ); the goal is to find the clique given the graph. Since the largest clique in G(n, 12 ) has expected size 2 log n, so long as k > (2 + ε) log n it is information-theoretically possible, via brute-force search, to recover the planted clique. On the other hand, despite substantial effort, no polynomial-time algorithm is √ known which can find a clique of size k 6 o( n), exponentially-larger than cliques which can be 18Or, mildly more generally, E should have SoS norm less than ε [MSS16]. 15 found by brute force. For other examples one need not look far: sparse principal component analysis, planted constraint satisfaction problems, and densest-k-subgraph are just a few more problems exhibiting information-computation gaps. This ubiquity leads to several questions: 1. What rigorous evidence can be provided for the claim: no polynomial algorithm tolerates b 1 , . . . , x n ) for particular a n < n ∗ samples and produces a constant-correlation estimator θ(x Bayesian estimation problem p(x, θ)? 2. Can information-computation gaps of different problems be explained by similar underlying phenomena? That is, is there a structural feature of a Bayesian estimation problem which determines whether it exhibits an information-computation gap, and if so, what is the critical number of samples n ∗ required by polynomial-time algorithms? 3. Are there methods to easily predict the location of a critical number n ∗ of samples, without analyzing every polynomial-time algorithm one can think of? Rigorous evidence for computational phase transitions The average-case nature of Bayesian estimation problems makes it unlikely that classical tools like (NP-hardness) reductions allow us to reason about the computational difficulty of such problems in the too-few-samples regime. Instead, to establish hardness of an estimation problem when n < n ∗ for some critical n ∗ , one proves impossibility results for restricted classes of algorithms. Popular classes of algorithms for this purpose include Markov-Chain Monte Carlo algorithms and various classes of convex programs, especially arising from convex hierarchies such as the Sherali-Adams linear programming hierarchy and the sum of squares semidefinite programming hierarchy. Results like this are meaningful only if the class of algorithms for which one proves an impossibility result captures the best known (i.e. lowest-sample-complexity) polynomial-time algorithm for the problem at hand. Better yet would be to use a class of algorithms which captures the lowest-sample-complexity polynomial-time algorithms for many Bayesian estimation problem simultaneously. In the present work we study sample complexity up to low-order additive terms in the number n of samples. For example, in the k-community α-mixed-membership stochastic block model, we provide an algorithm which estimates communities in graphs of average degree d > (1 + δ)k 2 (α + 1)2 /ε 2 , for any constant δ > 0. Such precise algorithmic guarantees suggest the pursuit of equally-precise lower bounds. Proving a lower bound against the convex-programming-based algorithms most commonly considered in previous work on lower bounds for Bayesian estimation problems does not suit this purpose. While powerful, these algorithms are generally not designed to achieve optimal sample complexity up to low-order additive terms. Indeed, there is mounting evidence in the block model setting that sum of squares semidefinite programs actually require a constant multiplicative factor more samples than our meta-algorithm [MS16, BKM17]. Another approach to providing rigorous evidence for the impossibility side of a computational threshold is statistical query complexity, used to prove lower bounds against the class of statistical query algorithms [FGR+ 13, FGV15, FPV15]. To the best of our knowledge, similar to sum of squares 16 algorithms, statistical query algorithms are not known to achieve optimal sample rates for problems such as community recovery in the block model up to low-order additive terms. Such sampleoptimal algorithms seem to intrinsically require the ability to compute a complicated function of many samples simultaneously (for example, the top eigenvector of the non-backtracking operator). But even the most powerful statistical query algorithms (using the 1-STAT oracle) can access one bit of information about each sample x (by learning the value of some function h(x) ∈ {0, 1}). This makes it unclear how the class of statistical query algorithms can capture the kinds of sampleoptimal algorithms we want to prove lower bounds against. Our meta-algorithm offers an alternative. By showing that when n is less than a critical n ∗ there are no constant-correlation low-degree estimators for a hidden random variable, one rules out any efficient algorithm captured by our meta-algorithm. Concretely, Theorems 1.1 and 1.3 show that in order for an estimation problem to be intractable it is necessary that every low-degree polynomial fails to correlate with the second or third moment of the posterior distribution (in the sense of Eqs. (1.1) and (1.3)). This kind of fact about low-degree polynomial is something we can aim to prove unconditionally as a way to give evidence for the intractability of a Bayesian estimation problem. Next we discuss our example result of this form in the block model setting. Concrete unconditional lower bound at the Kesten–Stigum threshold In this work, we show an unconditional lower bound about low-degree polynomials for the stochastic block model with k communities at the Kesten–Stigum threshold. For k > 4, this threshold is bounded away from the information-theoretic threshold [AS15]. In this way, our lower bounds gives evidence for an inherent gap between the information-theoretical and computational thresholds. For technical reasons, our lower bound is for a notion of correlation mildly different from Eqs. (1.1) and (1.3). Our goal is to compare the stochastic block model distribution SBM(n, d, ε, k) graphs to the Erdős-Rényi distribution G(n, nd ) with respect to low-degree polynomials. As before we represent graphs as adjacency matrices x ∈ {0, 1} n×n . Among all low-degree polynomials p(x), we seek one so that the typical value of p(x) for graphs x from the stochastic blocks model is as large as possible compared to its typical for Erdős-Rényi graphs. The following theorem shows that a suitable mathematical formalization of this question exhibits a sharp “phase transition” at the Kesten–Stigum threshold. Theorem 1.9. Let d, ε, k be constants. Then, max p∈’[x]6 ℓ x∼SBM(n,d,ε,k) x∼G(n,d/n) p(x)   2 1/2 p(x) ( > n Ω(1) if ε 2 d > k 2 , ℓ > O(log n) 6 n o(1) if ε 2 d < k 2 , ℓ 6 n o(1) (1.8) Let µ : {0, 1} n×n → ’ be the relative density of SBM(n, d, ε, k) with respect to G(n, nd ). Basic linear algebra shows that the left-hand side of Eq. (1.8) is equal to kµ 6 ℓ k2 , where k·k2 is the Euclidean norm with respect to the measure G(n, d/n) and µ 6 ℓ is the projection (with respect to this norm) of µ to the subspace of functions of degree at most ℓ. This is closely related to the χ2 -divergence of µ with respect to G(n, d/n), which in the present notation would be given by k(µ − 1)k2 . When the latter quantity is small, k(µ − 1)k2 6 o(1), one may conclude that the distribution µ is information17 theoretically indistinguishable from G(n, d/n). This technique is used in the best current bounds on the information-theoretic properties of the block model [MNS12, BMNN16]. The quantity in Theorem 1.9 is a low-degree analogue of the χ2 -divergence. If it were true that k(µ 6 ℓ − 1)k2 6 o(1), then by a straightforward application of Cauchy-Schwarz it would follow that no low-degree polynomial p(x) distinguishes the block model from G(n, d/n), since every such p would have (after setting G(n,d/n) p(x)  0) that SBM p(x) 6 o( G(n,d/n) p(x)2 )1/2 . This condition turns out to be quite powerful: [BHK+ 16, HKP+ 17] give evidence that for problems such as planted clique, for which distinguishing instances drawn from a null model from instances with hidden structure should be computationally intractable, the condition k(µ 6 ℓ − 1)k 6 o(1) is closely related to sum of squares lower bounds.19 The situation in the k-community block model is a bit more subtle. One has only that kµ 6 ℓ k 6 n o(1) below the Kesten-Stigum threshold because even in the latter regime it remains possible to distinguish a block model graph from G(n, d/n) via a low-degree polynomial (simply counting triangles will suffice). However, we can still hope to rule out algorithms which accurately estimate communities below the Kesten-Stigum threshold. For this we prove the following theorem. Theorem 1.10. Let d, ε, k, δ be constants such that ε 2 d < (1 − δ)k 2 . Let f : {0, 1} n×n → ’ be any function, let i, j ∈ [n] be distinct. Then if f satisfies d f (x)  0 and is correlated with the indicator x∼G(n, n ) 1σi σ j that i and j are in the same community in the following sense: x∼SBM(n,d,ε,k) ( f (x)(1σi σ j − 1k ) d x∼G(n, n ) f (x)2 )1/2 > Ω(1) then deg f > n c(d,ε,k) for some c(d, ε, k) > 0. There is one subtle difference between the polynomials ruled out by this theorem and those which could be used by our meta-algorithm. Namely, this theorem rules out any f whose correlation with the indicator 1σi σ j is large compared to f ’s standard deviation under G(n, d/n), whereas our meta-algorithm needs a polynomial f where this correlation is large compared to f ’s standard deviation under the block model. In implementing our meta-algorithm for the block model and for other problems, we have found that these two measures of standard deviation are always equal (up to low-order additive terms) for the polynomials which turn out to provide sample-optimal constant-correlation estimators of hidden variables. Interesting open problems are to prove a version of the above theorem where standard deviation is measured according to the block model and to formalize the idea that SBM f (x)2 should be related to G(n,d/n) f (x)2 for good estimators f . It would also be quite interesting to see how large the function c(d, ε, k) can be made: the above theorem shows that when d < (1 − δ)k 2 /ε 2 the degree of any good estimator of 1σi σ j must be polynomial in n—perhaps it must be linear, or even quadratic in n. 19In particular, the so-called pseudocalibration approach to sum of squares lower bounds works only when k(µ 6 ℓ −1)k 6 o(1). 18 General strategies to locate algorithmic thresholds The preceding theorems suggest a general strategy to locate critical a sample complexity n ∗ for almost any Bayesian estimation problem: compute a Fourier transform of an appropriate relative density µ and examine the 2-norm of its low-degree projection. This strategy has several merits beyond its broad applicability. One advantage is that in showing k(µ > ℓ − 1)k > Ω(1), one automatically has discovered a degree-ℓ polynomial and a proof that it distinguishes samples with hidden structure from an appropriate null model. Another is the mounting evidence (see [BHK+ 16, HKP+ 17]) that when, on the other hand k(µ 6 ℓ −1)k 6 o(1) for large-enough ℓ, even very powerful convex programs cannot distinguish these cases. A final advantage is simplicity: generally computing k(µ > ℓ − 1)k is a simple exercise in Fourier analysis. Finally, we compare this strategy to the only other one we know which shares its applicability across many Bayesian estimation problems, namely the replica and cavity methods (and their attendant algorithm, belief propagation) from statistical physics [MM09]. These methods were the first used to predict the sharp sample complexity thresholds we study here for the stochastic block model, and they have also been used to predict similar phenomena for many other hidden variable estimation problems [LBB+ 16, LBB+ 16, LBB+16]. Though remarkable, the predictions of these methods are much more difficult than ours make rigorous—in particular, it is notoriously challenging to rigorously analyze the belief propagation algorithm, and often when these predictions are made rigorous, only a modified version (“linearized BP”) can be analyzed in the end. By contrast, our methods to predict critical sample complexities, design algorithms, and prove lower bounds all study essentially the same low-degree-polynomial algorithm. We view it as a fascinating open problem to understand why predicted critical sample complexities offered by the replica and cavity methods are so often identical to the predictions of the low-degree-polynomials meta-algorithm we propose here. 2 Techniques To illustrate the idea of low-degree estimators for posterior moments, let’s first consider the most basic stochastic block model with k  2 disjoint communities (α  0). (Our discussion will be similar to the analysis in [MNS15a].) Let y ∈ {±1} n be chosen uniformly at random and let x ∈ {0, 1} n×n be the adjacency matrix of a graph such that for every pair i < j ∈ [n], we have x i j  1 with probability (1 + ε yi y j ) nd . Our goal is to find a matrix-valued low-degree polynomial P(x) that correlates with y yT . It turns out to be sufficient to construct for every pair i, j ∈ [n] a low-degree polynomial that correlates with yi y j . The linear polynomial p i j (x)  n εd  xi j − d n  is an unbiased estimator for yi y j in the sense that [p i j (x) | y]  yi y j . By itself, this estimator is not particular useful because its variance p i j (x)2 ≈ εn2 d is much larger than the quantity yi y j we are trying to estimate. However, if we let α ⊆ [n]2 be a length-ℓ path between i and j (in the complete graph), then we can combine the unbiased estimators along the path α and obtain a polynomial p α (x)  Ö ab∈α 19 p ab (x) (2.1) Î  that is still an unbiased estimator [p α (x) | yi , y j ]  ab∈α y a y b | y i , y j  y i y j . This estimator n ℓ 2 has much higher variance p α (x) ≈ (ε2 d ) . But we can hope to reduce this variance by averaging over all such paths. The number of such paths is roughly n ℓ−1 (because there are ℓ − 1 intermediate vertices to choose). Hence, if these estimators {p α (x)}α were pairwise independent, this averaging would reduce the variance by a multiplicative factor n ℓ−1 , giving us a final variance of (εn2 d )ℓ · n 1−ℓ  ( ε12 d )ℓ · n. We can see that above the Kesten–Stigum threshold, i.e., ε 2 d > 1 + δ for δ > 0, this heuristic variance bound ( ε12 d )ℓ · n 6 1 is good enough for estimating the quantity yi · y j for paths of length ℓ > log1+δ n. Two steps remain to turn this heuristic argument into a polynomial-time algorithm for estimating the matrix y yT . First, it turns out to be important to consider only paths that are self-avoiding. As we will see next, estimators from such paths are pairwise independent enough to make our heuristic variance bound go through. Second, a naive evaluation of the final polynomial takes quasipolynomial time because it has logarithmic degree (and a quasi-polynomial number of non-zero coefficients in the monomial basis). We describe the high-level ideas for avoiding quasi-polynomial running time later in this section (Section 2.5). 2.1 Approximately pairwise-independent estimators Let SAWℓ (i, j) be the set of self-avoiding walks α ⊆ [n]2 of length ℓ between i and j. Consider the Í unbiased estimator p(x)  |SAW1ℓ (i, j)| α∈SAWℓ (i, j) p α (x) for yi y j . Above the Kesten–Stigum threshold and for ℓ > O(log n), we can use the following lemma to show that p(x) has variance O(1) and achieves constant correlation with z  yi y j . We remark that the previous heuristic variance bound corresponds to the contribution of the terms with α  β in the left-hand side of Eq. (2.2). Lemma 2.1 (Constant-correlation estimator). Let (x, z) be distributed over {0, 1} n × ’. Let {p α }α∈I be a collection of real-valued n-variate polynomials with the following properties: 1. unbiased estimators: [p α (x) | z]  z for every α ∈ I 2. approximate pairwise independence: for δ > 0, Õ α,β∈I Then, the polynomial p  1 |I| Í α∈I p α (x) · p β (x) 6 p α satisfies 1 δ2 p(x) · z > δ · · |I| 2 z2 p(x)2 · (2.2) z2  1/2 . Remark 2.2. In applying the lemma we often substitute for Eq. (2.2) the equivalent condition z2 · Õ α,β∈I p α (x) · p β (x) 6 1 Õ ( p α (x)z) · ( p β (x)z) · δ2 α,β∈I which is conveniently invariant to rescaling of the p α ’s. Proof. Since the polynomial p is an unbiased estimator for z, we have p(x)z  p(x)2 6 (1/δ 2 ) · z 2 . Taken together, we obtain the desired conclusion. 20 z 2 . By Eq. (2.2),  In Section 3.1, we present the short combinatorial argument that shows that above the Kesten– Stigum bound the estimators for self-avoiding walks satisfy the conditions Eq. (2.2) of the lemma. We remark that if instead of self-avoiding walks we were to average over all length-ℓ walks between i and j, then the polynomial p(x) computes up to scaling nothing but the (i, j)-entry of the ℓ-th power of the centered adjacency x − nd 1 1 T . For ℓ ≈ log n, the ℓ-th power of this matrix converges to vvT , where v is the top eigenvector of the centered adjacency matrix. For constant degree d  O(log n), it is well-known that this eigenvector fails to provide a good approximation to the true labeling. In particular, the corresponding polynomial fails to satisfy the conditions of Lemma 2.1 close to the Kesten–Stigum threshold. 2.2 Low-degree estimators for higher-order moments Let’s turn to the general mixed-membership stochastic block model SBM(n, d, ε, k, α 0). Let (G, σ) be graph G and community structure σ  (σ1 , . . . , σn ) drawn from this model. Recall that σ1 , . . . , σn are k-dimensional probability vectors, each roughly uniform over α 0 + 1 of the coordinates. Let x ∈ {0, 1} n×n be the adjacency matrix of G and let y1 , . . . , yk ∈ ’n be centered community indicator vectors, so that (ys )i  (σi )s − 1k . It’s instructive to see that, unlike for disjoint communities, second moments are not that useful for overlapping communities. As a thought experiment suppose we are given the matrix Ík T s1 (y s )(y s ) (which we can estimate using the path polynomials described earlier). In case of disjoint communities, this matrix allows us to “read off” the community structure directly (because two vertices are in the same community if and only if the entry in the matrix is 1 − O(1/k)). For overlapping communities (say the extreme case α 0 ≫ k for simplicity), we can think of each σi as a random perturbation of the uniform distribution so that (σi )s  (1 + ξ i,s ) 1k for iid Gaussians {ξ i,s } with small variance. Then, the centered community indicator vectors y1 , . . . , yk are iid Í centered, spherical Gaussian vectors. In particular, the covariance matrix ks1 ys ysT essentially only determines the subspace spanned by the vectors y1 , . . . , yk but not the vectors themselves. (This phenomenon is sometimes called the “rotation problem” for matrix factorizations.) In contrast, classical factor analysis results show that if we were given the third moment tensor Ík ⊗3 s1 y s , we could efficiently reconstruct the vectors y1 , . . . , y k [Har70, LRA93]. This fact is the reason for aiming to estimate higher order moments in order to recover overlapping communities. In the same way that a single edge x i, j − nd gives an unbiased estimator for the (i, j)-entry of the second moment matrix, a 3-star (x i,c − nd )(x j,c − nd )(x k,c − nd ) gives an unbiased estimator for Í the (i, j, k)-entry of the third moment tensor ks1 ys⊗3 . This observation is key for the previous best algorithm for mixed-membership community detection [AGHK13]. However, even after averaging over all possible centers c, the variance of this estimator is far too large for sparse graphs. In order to decrease this variance, previous algorithms [AGHK13] project the tensor to the top eigenspace of the centered adjacency matrix of the graph. In terms of polynomial estimators this projection corresponds to averaging over all length-ℓ-armed 3-stars20 for ℓ  log n. Even for 20A length-ℓ-armed 3-star between i, j, k ∈ [n] consists of three length-ℓ walks between i, j, k and a common center c ∈ [n] 21 disjoint communities, this polynomial estimator would fail to achieve the Kesten–Stigum bound. In order to improve the quality of this polynomial estimator, informed by the shape of thresholdachieving estimator for second moments, we average only over such long-armed 3-stars that are self-avoiding. We show that the resulting estimator achieves constant correlation with the desired third moment tensor precisely up to the Kesten–Stigum bound (Section 5.2). 2.3 Correlation-preserving projection A recurring theme in our algorithms is that we can compute an approximation vector P that is correlated with some unknown ground-truth vector Y in the Euclidean sense hP, Yi > δ · kP k · kY k, where the norm k·k is induced by the inner product h·, ·i. (Typically, we obtain P by evaluating a low-degree polynomial in the observable variables and Y is the second or third moment of the hidden variables.) In this situation, we often seek to improve the quality of the approximation P—not in the sense of increasing the correlation, but in the sense of finding a new approximation Q that is “more similar” to Y while roughly preserving the correlation, so that hQ, Yi > δ O(1) · kQ k · kY k. As a concrete example, we may know that Y is a positive semidefinite matrix with all-ones on the diagonal and our goal is to take an arbitrary matrix P correlated with Y and compute a new matrix Q that is still correlated with Y but in addition is positive semidefinite and has all-ones on the diagonal. More generally, we may know that Y is contained in some convex set C and the goal is “project” P into the set C while preserving the correlation. We note that the perhaps most natural choice of Q as the vector closest to P in C does not work in general. (For example, if Y  (1, 0), C  {(a, b) | a 6 1}, and P  (δ · M, M), then the closest vector to P in C is (1, M), which has poor correlation with Y for large M.) Theorem 2.3 (Correlation-preserving projection). Let C be a convex set and Y ∈ C. Let P be a vector with hP, Yi > δ · kP k · kY k. Then, if we let Q be the vector that minimizes kQ k subject to Q ∈ C and hP, Qi > δ · kP k · kY k, we have hQ, Yi > δ/2 · kQ k · kY k . (2.3) Furthermore, Q satisfies kQ k > δkY k. Proof. By construction, Q is the Euclidean projection of 0 into the set C ′ : {Q ∈ C | hP, Qi > δkP k · kY k}. It’s a basic geometric fact (sometimes called Pythagorean inequality) that a Euclidean projection into a set decreases distances to points into the set. Therefore, kY − Q k 2 6 kY − 0k 2 (using that Y ∈ C ′). Thus, hY, Qi > kQ k 2/2. On the other hand, hP, Qi > δkP k · kY k means that kQ k > δkY k by Cauchy–Schwarz. We conclude hY, Qi > δ/2 · kY k · kQ k.  In our applications the convex set C typically consists of probability distributions or similar objects (for example, quantum analogues like density matrices or pseudo-distributions—the sumof-squares analogue of distributions). Then, the norm minimization in Theorem 2.3 can be viewed as maximizing the Rényi entropy of the distribution Q. From this perspective, maximizing the entropy within the set C ′ ensures that the correlation with Y is not lost. 22 2.4 Low-correlation tensor decomposition Earlier we described how to efficiently compute a 3-tensor P that has correlation δ > 0 with Í a 3-tensor ki1 yi⊗3 , where y1 , . . . , yk are unknown orthonormal vectors we want to estimate (Section 2.2). Here, the correlation δ depends on how far we are from the threshold and may be minuscule (say 0.001). It remains to decompose the tensor P into a short list of vectors L so as to ensure that O(1) . (Ideally of course |L|  k. In the block model context this guari∈[k] max ŷ∈L h ŷ, yi > δ antee requires a small amount of additional work to cross-validate vectors in a larger list.) To the best of our knowledge, previous tensor decomposition algorithms do not achieve this kind of Í guarantee and require that the correlation of P with the orthogonal tensor ki1 yi⊗3 is close to 1 (sometimes even within polynomial factors 1/n O(1) ). In the current work, we achieve this guarantee building on previous sum-of-squares based tensor decomposition algorithms [BKS15, MSS16]. These algorithms optimize over moments of pseudo-distributions (a generalization of probability distributions) and then apply Jennrich’s classical tensor decomposition algorithms to these “pseudo-moments”. The advantage of this approach is that it provably works even in situations where Jennrich’s algorithm fails when applied to the original tensor. As a thought experiment, suppose we are able to find pseudo-moments M that are correlated Í with the orthogonal tensor ki1 yi⊗3 . Extending previous techniques [MSS16], we show that Jennrich’s algorithm applied to M is able to recover vectors that have constant correlation with a constant fraction of the vectors y1 , . . . , yk . A priori it is not clear how to find such pseudo-moments M because we don’t know the Í orthogonal tensor ki1 yi⊗3 , we only know a 3-tensor P that is slightly correlated with it. Here, the correlation-preserving projection discussed in the previous section comes in: by Theorem 2.3 we can efficiently project P into the set of pseudo-moments in a way that preserves correlation. In this way, we obtain pseudo-moments M that are correlated with the unknown orthogonal tensor Ík ⊗3 i1 y i . When P is a 3-tensor as above, we encounter technical difficulties inherent to odd-order tensors. (This is a common phenomenon in the tensor-algorithms literature.) To avoid these difficulties we give a simple algorithm, again using the correlation-preserving projection idea, to lift a 3-tensor P which is δ-correlated with an orthogonal tensor A to a 4-tensor P ′ which is δ O(1) -correlated with an appropriate orthogonal 4-tensor. See Section 7.2. 2.5 From quasi-polynomial time to polynomial time In this section, we describe how to evaluate certain logarithmic-degree polynomials in polynomialtime (as opposed to quasi-polynomial time). The idea is to use color coding [AYZ95].21 ℓ For a coloring c : [n] → [ℓ] and a subgraph α ⊆ [n]2 on ℓ vertices, let Fc,α  ℓℓ! · 1c(α)[ℓ] be a scaled indicator variable of the event that α is colorful. 21We thank Avi Wigderson for suggesting that color coding may be helpful in this context. 23 Theorem 2.4 (Evaluating colorful-path polynomials). There exists a n O(1) · exp(ℓ)-time algorithm that given vertices i, j ∈ [n], a coloring c : [n] → [ℓ] and an adjacency matrix x ∈ {0, 1} n×n evaluates the polynomial Õ p c (x) : |SAW1ℓ (i, j)| p α (x) · Fc,a . (2.4) (Here, p α ∝ Î α∈SAW ℓ (i, j) ab∈α (x ab − nd ) is the polynomial in Eq. (2.1).) Proof. We can reduce this problem to computing the ℓ-th power of the following n · 2ℓ -by-n · 2ℓ matrix: The rows and columns are indexed by pairs (a, S) of vertices a ∈ [n] and color sets S ⊆ [ℓ]. The entry for column (a, S) and row (b, T) is equal to x ab − nd if T  S ∪ {c(a)} and 0 otherwise. If we compute the ℓ-th power of this matrix, then the entry for column (i, ∅) and row (j, [ℓ]) is the sum over all colorful ℓ-paths from i to j.  For a fixed coloring c, the polynomial p c does not provide a good approximation for the Í polynomial p(x) : |SAW1ℓ (i, j)| α∈SAWℓ (i, j) p α (x). In order to get a good approximation, we will choose random colorings and average over them. If we let c be a random coloring, then by construction c Fc,α  1 for every simple ℓ-path α. Therefore, c p c (x)  p(x) for every x ∈ {0, 1} n×n . We would like to estimate the variance of p c (x). Here, it turns out to be important to consider a typical x drawn from stochastic block model distribution SBM. x∼SBM(n,d,ε) c p c (x)2  1 |SAWℓ (i, j)| 2 6e 2ℓ · Õ Õ α,β∈SAW ℓ (i, j) 1 |SAWℓ (i, j)| c Fc,α · Fc,β · α,β∈SAW ℓ (i, j) x∼SBM | p α (x)p β (x)| . x p α (x)p β (x) (2.5) (2.6) 2 6 e 2ℓ (because ℓ ℓ /ℓ! 6 e ℓ ). For the last step, we use that c Fc,α The right-hand side of Eq. (2.6) corresponds precisely to our notion of approximate pairwise independence in Lemma 2.1. Therefore, if we are within the Kesten–Stigum bound, ε 2 d > 1 + δ, the right-hand side of Eq. (2.6) is bounded by e 2ℓ · 1/δ O(1) . We conclude that with high probability over x, the variance of p c (x) for random c is bounded by O(ℓ) e . It follows that by averaging over e O(ℓ) random colorings we obtain a low-variance estimator for p(x). 2.6 Illustration: push-out effect in spiked Wigner matrices We turn to a first demonstration of our meta-algorithm beyond the stochastic block model: deriving the critical signal-to-noise ratio for (Gaussian) Wigner matrices (i.e. symmetric matrices with iid entries) with rank-one spikes. This section demonstrates the use of Theorem 1.1; more sophisticated versions of the same ideas (for example our 3rd-moment meta-theorem, Theorem 1.3) will be used in the course of our block model algorithms. Consider the following Bayesian estimation problem: We are given a spiked Wigner matrix A  λvvT + W so that W is a random symmetric matrix with Gaussian entries Wi j ∼ N(0, n1 ) and 24 v ∼ N(0, n1 Id). The goal is to estimate v, i.e., compute a unit vector v̂ so that hv, v̂i 2 > Ω(1). Since √ √ the spectral norm of a Wigner matrix satisfies kW k  2, it follows that for λ > 2, the top eigenvector v̂ of A satisfies hv, v̂i 2 > Ω(1). However, it turns out that we can estimate the spike v even for smaller values of λ: a remarkable property of spiked Wigner matrices is that as soon as λ > 1, the top eigenvector v̂ becomes correlated with the spike v [BAP+ 05]. (This property is sometimes called the “pushout effect”.) Unfortunately known proofs of this property a quite involved. In the following, we apply Theorem 1.1 to give an alternative proof of the fact that it is possible to efficiently estimate the spike v as soon as λ > 1. Our algorithm is more involved and less efficient than computing the top eigenvector of A. The advantage is that its analysis is substantially simpler compared to previous analyses. Theorem 2.5 (implicit in [BAP+ 05]). If λ  1 + δ for some 1 > δ > 0, there is a degree δ−O(1) · log n matrix-valued polynomial f (A)  { f i j (A)}i j 6 n such that ( k ⊤ W,v Tr f (A)vv f (A)kF2 )1/2 · ( kvv ⊤ kF2 )1/2 > δ O(1) . O(1) Together with Theorem 1.1, the above theorem gives an algorithm with running time n log n/δ to find v̂ with nontrivial h v̂, vi 2 .22 The analysis of [BAP+ 05] establishes the above theorem for the polynomial f (A)  A ℓ with ℓ  δ−O(1) · log n. Our proof chooses a different polynomial, which affords a substantially simpler analysis.  Î Proof of Theorem 2.5. For α ⊆ n2 , let χα (A)  {i, j}∈α A i j . Let L  log n/δ C for C a large enough constant. For i j ∈ [n], let SAWi j (L) be the collection of all self-avoiding paths from i to j in the L−1 complete graph on n vertices. Observe that nλ L χα for α ∈ SAWi j (L) is an unbiased estimator of vi v j :   χα (A) | v i , v j  v " Ö kℓ∈α W # (Wkℓ + λv k v ℓ ) | v i , v j  λ L v i v j Ö k∈α\{i, j} v 2k  λL · vi v j . n L−1 L−1 We further claim that the collection { nλ L χα }α∈SAWi j (L) is approximately pairwise independent in the sense of Lemma 2.1. To show this we must check that n 2(L−1) Õ λ2L α,β χα χβ 6 1 |SAWi j (L)| 2 δ2 v 2i v 2j  1 1 |SAWi j (L)| 2 · 2 . 2 δ n 22While this algorithm is much slower than the eigenvector-based algorithm—even after using color coding to O(1) O(1) improve the n log n/δ running time to n 1/δ —the latter requires many sophisticated innovations and ideas from random matrix theory. This algorithm, by contrast, can be derived and analyzed with our meta-theorem, little innovation required. 25 The dominant contributers to the sum are α, β which intersect only on the vertices i and j. In that case, Ö n 2(L−1) 2(L−1) v 2k  v 2i v 2j . χ χ  n α β 2L λ k∈α∪β The only other terms which might contribute to the same order are α, β such that α ∩ β is a union of two paths, one starting at i and one at j. If the lengths of these paths are t and t ′, respectively, and t ′ + t ′ < L, then n 2(L−1) λ 2L n 2(L−1) χα χβ  2(t+t ′ ) λ where we have used that  v  Ö  Ö   n t+t ′ 2 t+t ′ 2   v v ( A ) · k ℓ   t+t ′ · (1 + O(λ /n))  kℓ W λ (k,ℓ)∈α∩β  (k,ℓ)∈α△β    1 2 2 2 A2kℓ | v k , v ℓ  n (1 + O(λ /n)) · t+t ′ vi v j . choices for such pairs α, β, so long as t + t ′ < L. If t + t ′  L, There are at most |SAWi j (L)| 2 /n then there are n times more choices than the above bound. All together, n 2(L−1) λ2L Õ α,β∈SAWi j (L) L © Õ 1 χα χβ 6 |SAWi j (L)| · ­ λt « t0 !2 + n ª ®· λL ¬ v 2i v 2j 6 1 + o(1) · |SAWi j (L)| · 1 − 1/λ v 2i v 2j where we have used that λ  1 + δ > 1 and chosen C large enough that n/λ L 6 1/n. Rewriting in  terms of δ  λ − 1 and applying Lemma 2.1 finishes the proof. 3 Warmup: stochastic block model with two communities We demonstrate our meta-algorithm by applying it to the two-community stochastic block model. The algorithm achieves here the same threshold for partial recovery as the best previous algorithms [MNS13, Mas13], which is also known to be the information-theoretic threshold [MNS15b]. While the original works involved a great deal of ingenuity, the merit of our techniques is to provide a simple and automatic way to discover and analyze an algorithm achieving the same guarantees. Definition 3.1 (Two-community stochastic block model). For parameters ε, d > 0, let SBM(n, d, ε) n be the following distribution on pairs (x, y) where x ∈ {0, 1}( 2 ) is the adjacency matrix of an n-vertex graph and y ∈ {±1} n is a labeling of the n vertices. First, sample y ∼ {±1} n uniformly. Then, independently for every pair i < j, add the edge {i, j} with probability (1 + ε) nd if yi  y j and with probability (1 − ε) nd if yi , y j . The following theorem gives the best bounds for polynomial-time algorithms for partial recovery in this model. (We remark that the algorithms in [MNS13, Mas13] actually run in time close to linear. In this work, we content ourselves with coarser running time bounds.) 26 1 ε2 d and d 6 n o(1) . Then, there exists n a randomized polynomial-time algorithm A that given a graph x ∈ {0, 1}( 2 ) outputs a labeling ỹ(x) such that for all sufficiently large n > n 0 (ε, d), Theorem 3.2 ([MNS13, Mas13]). Let ε ∈ ’, d ∈ Ž with δ ≔ 1 − (x,y)∼SBM(n,d,ε) h ỹ(x), yi 2 > δ O(1) · n 2 . Here, the factor n 2 in the conclusion of the theorem normalizes the vectors ỹ(x) and y because k ỹ(x)k 2 · k y k 2  n 2 . In the remainder of this section, we will prove the above theorem by specializing our metaalgorithm for two-community stochastic block model. For simplicity, we will here only analyze a version of algorithm that runs in quasi-polynomial time. See Section 2.5 for how to improve the running time to n 1/poly(δ) . n Algorithm 3.3. For a given n-vertex graph x ∈ {0, 1}( 2) with average degree d and some parameter δ > 0, execute the following steps:23 1. evaluate the following matrix-valued polynomial P(x)  (Pi j (x)) Pi j (x) ≔ ℓ Õ α∈SAW ℓ (i, j) p α (x) . (3.1) Here as in Section 2, SAWℓ (i, j) ⊆ n2 consists of all sets of vertex pairs that form a simple O(1) .24 The polynomial p is a (self-avoiding) path between i and j of length ℓ  Θ(log n)/δ α   Î product of centered edge indicators, so that p α (x)  ab∈α x ab − nd .25 2. compute a matrix Y with minimum Frobenius norm satisfying the constraints          1 kP(x)kF ·n diag(Y)  1    · hP(x), Yi > δ′     Y0  . (3.2) and output a vector ỹ ∈ {±1} n obtained by taking coordinate-wise signs of a centered Gaussian vector with covariance Y.26 The matrix P(x) is essentially the same as the matrix based on self-avoiding walks analyzed in [MNS13]. The main departure from previous algorithms lies in the second step of our algorithm. As stated, the first step of the algorithm takes quasi-polynomial because it involves a sum over n ℓ terms (for ℓ  Θ(log n)/δ O(1) ). In prior works this running time is improved by using non-backtracking paths instead of self-avoiding paths. Non-backtracking paths can be counted in n O(1) time using matrix multiplication, but relating the non-backtracking path polynomial to the self-avoiding path polynomial requires intensive moment-method calculations. An alternative, 23The right choice of δ′ will depend in a simple way on the parameters ε and d. 24In particular, the paths in SAWℓ (i, j) are not necessarily paths in the graph x but in the complete graph on n vertices. 25Up to scaling, this polynomial is a d/n-biased Fourier character of sparse Erdős-Rényi graph. 26In other words, we apply the hyperplane rounding algorithm of Goemans and Williamson. 27 described in Section 2.5, is to compute the self-avoiding path polynomial P using color-coding, O(1) requiring time n O(1)+1/δ , still polynomial time for any constant δ > 0. The second step of the algorithm is a convex optimization problem over an explicitly represented spectrahedron. Therefore, this step can be carried out in polynomial time. We break the analysis of the algorithm into two parts corresponding to the following lemmas. The first lemma shows that if ε 2 d > 1 then the matrix P(x) has constant correlation with y yT for (x, y) ∼ SBM(n, d, ε) and n sufficiently large. (Notice that this the main preconditon to apply meta-Theorem 1.1.) Lemma 3.4 (Low-degree estimator for posterior second moment). Let ε ∈ ’ and d ∈ Ž, and assume def d  n o(1) . If δ  1 − ε12 d > 0 and n > n 0 (ε, d, δ) is sufficiently large, then the matrix-valued polynomial P(x) in Eq. (3.1) satisfies T (x,y)∼SBM(n,d,ε) hP(x), y y i > δ O(1) ·  x∼SBM(n,d,ε) kP(x)kF2  1/2 ·n (3.3) (Here, the factor n in the conclusion normalizes the matrix y yT because k y yT kF  n.) By application of Markov’s inequality to the conclusion of this theorem one shows that with P has Ω(1)-correlation with y y ⊤ with Ω(1)-probability. As we have noted several times, the same theorem would hold if we replaced P, an average over self-avoiding walk polynomials, with an average over nonbacktracking walk polynomials. This would have the advantage that the resulting polynomial can be evaluated in n O(1) time (i.e. with running time independent of δ), rather than n O(log n)/poly(δ) for P (which can be improved to n poly(1/δ) via color coding), but at the cost of complicating the moment-method analysis. Since we are aiming for the simplest possible proofs here we use P as is. The second lemma shows that given a matrix P that has constant correlation with y yT for an unknown labeling y ∈ {±1} n , we can efficiently compute a labeling ỹ ∈ {±1} n that has constant correlation with y. We remark that for this particular situation simpler and faster algorithms work (e.g., choose a random vector in the span of the top 1/δ O(1) eigenvectors of P); these are captured by the meta-Theorem 1.1, which we could use in place of the next lemma. (We are presenting this lemma, which involves a more complex and slower algorithm, in order to have a self-contained analysis in this warmup and because it illustrates a simple form of a semidefinite programming technique that is important for our tensor decomposition algorithm, which we use for overlapping communities.) Lemma 3.5 (Partial recovery from posterior moment estimate). Let P ∈ ’n×n be a matrix and 1 hP, y yT i. Let Y be the matrix of minimum Frobenius such that y ∈ {±1} n be a vector with δ′ ≔ kPk·n 1 Y  0, diagY  1, and kPk·n hY, Pi > δ′ (i.e., the constraints Eq. (3.2)). Then, the vector ỹ obtained by taking coordinate-wise signs of a Gaussian vector with mean 0 and covariance Y satisfies h ỹ, yi 2 > Ω(δ′)2 · n 2 . (Here, the factor n 2 in the conclusion normalizes the vectors ỹ, y because k ỹk 2 · k y k 2  n 2 .) 28 Proof. By Theorem 2.3, the matrix Y satisfis hY, y yTi > (δ′/2)kY k · k y k 2 and kY k > δ · k y k 2 . In particular, hY, y yTi > δ 2 n 2 /2. The analysis of rounding algorithm for the Grothendieck problem on psd matrices [AN04], shows that h ỹ, yi 2 > π2 hY, y yTi > Ω(δ 2 ) · n 2 . (Here, we use that y yT is a psd matrix.)  Taken together, the above lemmas imply a quasi-polynomial time algorithm for partial recovery in SBM(n, d, ε) when ε 2 d > 1. Proof of Theorem 3.2 (quasi-polynomial time version). Let (x, y) ∼ SBM(n, d, ε) with δ ≔ 1−1/ε 2 d > 0. 1 Run Algorithm 3.3 on x with the parameter δ′ chosen as 10 times the correlation factor in the conclusion of Lemma 3.4. Then, by Lemma 3.4, (x,y)∼SBM(n,d,ε) hP(x), y yT i > 10δ′ · x∼SBM(n,d,ε) kP(x)k · n. By a variant of Markov inequality Theorem A.1, the matrix P(x) satisfies with constant probability hP(x), y yT i > δ′ · kP(x)k · n. In this event, by Lemma 3.5, the final labeling ỹ satisfies 2 ′ 2 2 ỹ h ỹ, yi > Ω(δ ) · n . Since this event has constant probability, the total expected correlation satisfies (x,y)∼SBM(n,d,ε) h ỹ(x), yi 2 > Ω(δ′)2 · n 2 as desired.  It remains to prove Lemma 3.4. 3.1 Low-degree estimate for posterior second moment We will apply Lemma 2.1 to prove Lemma 3.4. The next two lemmas verify that the conditions of that lemma hold; they immediately imply Lemma 3.4. Lemma 3.6 (Unbiased estimators for yi y j ). For i, j ∈ [n] distinct, let SAWℓ (i, j) be the set of all simple paths from i to j in the complete graph on n vertices of length ℓ. Let x i j be the i j-th entry of the Î adjacency matrix of G ∼ SBM(n, d, ε), and for α ∈ SAWℓ (i, j), let p α (x)  ab∈α (x ab − nd ). Then for any yi , y j ∈ {±1} and α ∈ SAWℓ (i, j),  n ℓ   p α (x) | yi y j  yi y j . εd Thus, each simple path α from i to j in the complete graph provides an unbiased estimator (n/εd)ℓ p α (x) of yi y j . It is straightforward to compute that each has variance  pairwise independent, they could be averaged to give an estimator with variance n(ε 2 d)−ℓ , n ℓ−1 n ε2 d ℓ . If they were  1 n |SAWℓ (i, j)| · ε 2 d ℓ  since there are simple paths from i to j. If ℓ is logarithmic in n, this becomes small. The estimators are not strictly pairwise independent, but they do satisfy an approximate pairwise independence property which will be enough for us. def Lemma 3.7 (Approximate conditional independence). Suppose δ  1 − ε12 d > Ω(1) and d  n o(1) . For i, j ∈ [n] distinct, let SAWℓ (i, j) be the set of all simple paths from i to j in the complete graph on n vertices of length ℓ  Θ(log n)/δ C for a large-enough constant C. Let x i j be the i j-th entry of the adjacency Î matrix of G ∼ SBM(n, d, ε). Let p α (x)  ab∈α (x ab − nd ). Then yi2 y 2j Õ α,β∈SAW ℓ (i, j) p α (x)p β (x) 6 δ −O(1) · Õ α,β∈SAW ℓ (i, j) 29 p α (x)yi y j   p β (x)yi y j . To prove the lemmas we will use the following fact; the proof is straightforward. Fact 3.8. For x, y ∼ SBM, the entries of x are all independent conditioned on y, and a, b distinct,  x ab − d n | ya , yb  εd · ya yb  n  and We can prove both of the lemmas. x ab − d n 2  | ya , yb   d 1 + ε ya yb + O(d/n) . n Proof of Lemma 3.6. We condition on y and expand the expectation.   p α (x) | yi y j  y " Ö [x ab − ab∈α d n #  εd | y]  n ℓ y " Ö ya yb ab∈α # by Fact 3.8.  ℓ Because α is a path from i to j, every index a ∈ [n] except for i and j appears exactly twice in the product. So, removing the conditioning on ya for all a , i, j, we obtain as desired.   p α (x) | yi y j  εd n · yi y j  The proof of Lemma 3.7 is the heart of the proof, and will use the crucial assumption ε 2 d > 1. Proof of Lemma 3.7. Let α, β ∈ SAWℓ (i, j), and suppose that α and β share r edges. Let α△β denote the symmetric difference of α and β. Then   x ab − | ya , yb  | ya , yb · (x ab − x x  ab∈α△β    Ö Ö   ya yb  ε 2ℓ−2r  (1 + ε ya yb + O(d/n)) · y   ab∈α△β  ab∈α∩β  Ö  p α (x)p β (x)   y  ab∈α∩β   2ℓ−r  d n  d 2 n)  Ö  d n using Fact 3.8 in the second step. Since α and β are paths, the graph α△β has all even degrees, Î so ab∈α△β ya yb  1. Furthermore, any subgraph of α ∩ β contains some odd-degree vertex. So Î r y ab∈α∩β (1 + ε y a y b + O(d/n))  (1 + O(d/n)) . All in all, we obtain p α (x)p β (x)    2ℓ−r d n ε 2ℓ−2r (1 + O(d/n))r (3.4) Suppose r < ℓ. Paths α, β sharing r edges must share at least r vertices. If they share exactly r vertices, then the shared vertices must form paths in α and β beginning at i and j. Since each path has length ℓ and therefore contains ℓ − 1 vertices in addition to i and j, there are at most r · n 2(ℓ−1)−r such pairs α, β (the multiplicative factor r comes because the shared paths starting from i and j could have lengths between 0 and r). Other pairs α, β share r edges but s vertices for some s > r. For each s and r, there are at most n 2(ℓ−1)−s ℓ O(s−r) such pairs, because the shared edges must occur as at most s − r paths. Furthermore, ℓ O(s−r) n −(s−r) 6 n −Ω(1) when s > r. Putting all of this together, Õ α,β∈SAW ℓ (i, j) p α (x)p β (x) 6 n −2 · " ℓ−1 Õ r0   d 2ℓ−r ε 2ℓ−2r (1 + O(d/n))r r + n −Ω(1) + (ε 2 d)ℓ · n 30 #  n −2 · (1 + n −Ω(1) ) · (εd)2ℓ · ℓ Õ r0 ! r · (ε 2 d)−r + (ε 2 d)−ℓ · n , The additive factor of (ε 2 d)ℓ n in the first line comes from the case r  ℓ (i.e., α  β), where there are n ℓ−1 paths. In the second line we have used the assumption that d ≪ n to simplify the expression. Í m Finally, by convergence of the series ∞ m0 m · z for |z| < 1, and the choice of ℓ logarithmic in n, this is at most ! (1 + n −Ω(1) 2ℓ ) · (εd) O(1) 1 · 1− . 1 ε2 d So, now our goal is to show that Õ α,β∈SAW ℓ (i, j) ( p α (x)yi y j )( p β (x)yi y j ) > n −2 · (1 + n −Ω(1) 2ℓ ) · (εd) · 1 1− 1 ε2 d ! O(1) . Each term in the left-hand sum is (εd/n)2ℓ (by Lemma 3.6) and there are Ω(n 2ℓ−2 ) such terms, so the left-hand side of the above is at least Ω((εd)2ℓ /n 2 ). This proves the Lemma.  4 Matrix estimation for generalized block models In this section we phrase a result essentially due to Abbe and Sandon [AS16a] (and closely related to results by Bordenave et al [BLM15]) in somewhat more general terms. This turns out to be enough to capture an algorithm to estimate a pairwise-vertex-similarity matrix in the d, k, α, ε mixed-membership block model when ε 2 d > k 2 (α + 1)2 . ∫ Let U be a universe of labels, endowed with some base measure ν, such that 1 · dν  1. Let µ be a probability distribution on U, with a density relative to ν. (We abuse notation by conflating µ and its associated density). Let W : U × U → ’+ be a bounded nonnegative function with W(x, y)  W(y, x) for every x, y ∈ U. Consider a random graph model G(n, d, W, µ) sampled as follows. For each of n vertices, draw a label x i ∼ µ independently. Then for each pair i j ∈ [n]2 , independently add the edge (i, j) to the graph with probability nd W(x i , x j ). (This captures the W-random graph models used in literature on graphons.) Let F denote the space of square-integrable functions f : U → ’, endowed with the inner product h f , 1i  x∼µ f (x)1(x). That is, f ∈ F if x∼µ f (x)2 exists. We assume throughout that 1. (Stochasticity) For every x ∈ U, the average y∼µ W(x, y)  1. Í 2. (Finite rank) W has a finite-rank decomposition W(x, y)  i 6 r λ i f i (x) f i (y) where λ i ∈ ’ and f i : U → ’. The values λ i are the eigenvalues of W with respect to the inner product generated by µ. The eigenfunctions are orthonormal with respect to the µ inner product. Notice that the assumptions on W imply that its top eigenfunction f1 (x) is the constant function, with eigenvalue λ1  1. 31 3. (Niceness I) Certain rational moments of µ−1 exist; that is x∼µ µ(x)−t exists for t  −3/2, −2. p 4. (Niceness II) W and µ are nice enough that W(x, y) 6 1/ µ(x)µ(y) and |W(x, y)| 6 p λ 2 / µ(x)µ(y) for every x, y ∈ U, where W(x, y)  W(x, y) − 1. (Notice that in the case of discrete W and µ this is always true, and for smooth enough W and µ it is true via a δ-function argument.) The function W induces a Markov operator W : F → F . If f ∈ F , then (W f )(x)  y∼µ W(x, y) f (y) . (We abuse notation by conflating the function W and the Markov operator W.) Theorem 4.1 (Implicit in [AS16a]). Suppose the operator W has eigenvalues 1  λ1 > λ 2 > · · · > λ r def (each possibly with higher multiplicity) and δ  1 − dλ1 2 > 0. Let Π be the projector to the second eigenspace 2 of the operator W. For types x1 , . . . , x n ∼ µ, let A ∈ ’n×n be the random matrix A i j  Π(x i , x j ), where we abuse notation and think of Π : U × U → ’. There is an algorithm with running time n poly(1/δ) which outputs an n × n matrix P such that for x, G ∼ G(n, d, W, µ), x,G Tr P · A > δ O(1) · ( x,G kAk 2)1/2 ( x,G kP k 2 )1/2 . When U is discrete with k elements one recovers the usual k-community stochastic block model, and the condition λ 22 > 1 matches the Kesten-Stigum condition in that setting. When λ 22 > 1 + δ, the guarantees of Abbe and Sandon can be obtained by applying the above theorem to obtain an Í estimator P for the matrix M  s∈[k] v s v ⊤ s , where v s is the centered indicator vector of community s. The estimator P will have at least δ O(1) /k correlation with M, and a random vector in the span of the top k/δ O(1) eigenvectors of M will have correlation (δ/k)O(1) with some v s . Thresholding that vector leads to the guarantees of Abbe and Sandon for the k-community block model, with one difference: Abbe and Sandon’s algorithm runs in O(n log n) time, much faster than the n poly(1/δ) running time outlined above. In essence, they achieve this by computing an estimator P ′ for M which counts only non-backtracking paths in G (the estimator P counts self-avoiding paths). In Section 4.1 we prove a corollary of Theorem 4.1. This yields the algorithm discussed Theorem 1.2 for the mixed-membership blockmodel. As discussed before, the quantitative recovery guarantees of this algorithm are weaker than those of our final algorithm, whose recovery accuracy depends only on the distance δ of the signal-to-noise ratio of the mixed-membership blockmodel to 1. In Section 4.2 we prove Theorem 4.1. 4.1 Matrix estimation for the mixed-membership model We turn to the mixed-membership model and show that Theorem 4.1 yields an algorithm for partial recovery in the mixed-membership block model. However, the correlation of the vectors output by this algorithm with the underlying community memberships depends both on the signal-to-noise ratio and the number k of communiteis. (In particular, when k is super-constant this algorithm no longer solves the partial recovery task.) 32 Definition 4.2 (Mixed-Membership Block Model). Let G(n, d, ε, α, k) be the following random Í graph ensemble. For each node i ∈ [n], sample a probability vector σi ∈ ’k>0 with t∈[k] σi (t)  1 according to the following (simplified) Dirichlet distribution. (σ) ∝ Ö t∈[k] σi (t)α/k−1 For each pair of vertices i, i ′ ∈ [n], sample communities t ∼ σi and t ′ ∼ σi′ . If t  t ′, add the edge {i, i ′ } to G with probability nd (1 + (1 − 1k )ε). If t , t ′, add the edge {i, i ′ } to G with probability ε d n (1 − k ). (For simplicity, throughout this paper we consider only the case that the communities have equal sizes and the connectivity matrix has just two unique entries.) Theorem 4.3 (Constant-degree partial recovery for mixed-membership block model, k-dependent O(1) error). For every δ > 0 and d(n), ε(n), k(n), α(n), there is an algorithm with running time n O(1)+1/δ with the following guarantees when def δ  1− k 2 (α + 1)2 >0 ε2 d k, α 6 n o(1) and ε 2 d 6 n o(1) . and Let σ, G ∼ G(n, d, ε, k, α) and for s ∈ [k] let v s ∈ ’n be given by v s (i)  σi (s) − 1k . The algorithm outputs a vector x such that hx, v1 i 2 > δ′ kx k 2 kv1 k 2 , for some δ′ > (δ/k)O(1) .27 Ideally one would prefer an algorithm which outputs τ1 , . . . , τn ∈ ∆k−1 with corr(σ, τ) > δ′/(α + 1). If one knew that hx, v1 i > δ′ kx k kv k rather than merely the guarantee on hx, v1 i 2 (which does not include a guarantee on the sign of x), then this could be accomplished by correlationpreserving projection, Theorem 2.3. The tensor methods we use in our final algorithm for the mixedmembership model are able to obtain a guarantee on hx, v1 i and hence can output probability vectors τ1 , . . . , τn .28 To prove Theorem 4.3 we will apply Theorem 4.1 and then a simple spectral rounding algorithm; the next two lemmas capture these two steps. Lemma 4.4 (Mixed-membership block model, matrix estimation). If U is the (k − 1)-simplex, µ is the α, k Dirichlet distribution, and W(σ, σ′)  1 − εk + εhσ, σ′i, then G(n, d, W, µ) is the mixed-membership block model with parameters k, d, α, ε. In this case, the second eigenvalue of W has multiplicity k − 1 and ε . has value λ 2  k(α+1) Proof. The first part of the claim follows from the definitions. For the second part, note that W has the following decomposition W(σ, τ)  1 + Õ i6k ε(σi − 1k )(τi − 1k ) . 27The requirement ε2 d 6 n o(1) is for technical convenience only; as ε2 d increases the recovery problem only becomes easier. 28Such a guarantee could be obtained here by using a cross-validation scheme on x to choose between x and −x. Since we are focused on what can be accomplished by matrix estimation methods generally we leave this to the reader. 33 The functions σ 7→ σi − 1 k are all orthogonal to the constant function σ 7→ 1 with respect to µ; i.e. σ∼µ 1 · (σi − 1k )  0 because σi  1k . It will be enough to test the above Rayleigh quotient σ∼µ f (σ) · (W f )(σ) σ∼µ f (σ)2 with any function f (σ) in the span of the functions σ 7→ σi − 1k . If we pick f (σ)  σ1 − 1k the remaining calculation is routine, using only the second moments of the Dirichlet distribution (see Fact 4.5 below).  Fact 4.5 (Special case of Fact A.3). Let σ ∈ ’k be distributed according to the α, k Dirichlet distribution. 1 Let σ̃  σ − 1k · 1 be centered. Then (σ̃)(σ̃)⊤  k(α+1) · Π where Π is the projector to the complement of the all-1s vector in ’k . We analyze a simple rounding algorithm. Í be an n × n symmetric rank-k PSD matrix. Let P ∈ ’n×n be another Lemma 4.6. Let M  ki1 v i v ⊤ i symmetric matrix such that hP, Mi > δkP k kM k (where k · k is the Frobenious norm). Then for at least one vector v among v1 , . . . , v k , a random unit vector x in the span of the top (k/δ)O(1) eigenvectors of P satisfies hx, vi 2 > (δ/k)O(1) kv k 2 . Now we can prove Theorem 4.3. Proof of Theorem 4.3. Lemma 4.4 shows that the conditions of Theorem 4.1hold, and hence (via color coding) there is an n poly(1/δ) time algorithm to compute a matrix P such that hP, Mi > δ O(1) kP k kM k Í with probability at least δ O(1) , where M  s∈[k] v s v ⊤ s . (The reader may check that the matrix A of Theorem 4.1 is in this case the matrix M described here.) Applying Lemma 4.6 shows that a random unit vector x in the span of the top (k/δ)O(1) eigenvectors of P satisfies hx, vi 2 > (δ/k)O(1) kv k 2 , where v ∈ ’n has entries v i  σi (1). (The choice of 1 is without loss of generality.)  4.2 Proof of Theorem 4.1 Definition 4.7. For a pair of functions A, B : U × U → ’, we denote by AB their product, whose entries are (AB)(x, y)  z∼µ A(x, z)B(z, y). The strategy to prove Theorem 4.1 will as usual be to apply Lemma 2.1. We check the conditions of that Lemma in the following Lemmas, deferring their proofs till the end of this section. 34 Lemma 4.8. Let G i j be the 0/1 indicator for the presence of edge i ∼ j in a graph G. As usual, let SAWℓ (i, j) be the collection of simple paths of length ℓ in the complete graph on n vertices from i to j. Î Let x, G ∼ G(n, d, W, µ). Let α ∈ SAWℓ (i, j). Let p α (G)  ab∈α (G ab − nd ). Let W(x, y)  W(x, y) − 1. Then  ℓ   ℓ−1 d W (x i , x j ) p α (G) | x i , x j  n Lemma 4.9. With the same notation as in Lemma 4.8, as long as ℓ > C log n/δ O(1) for a large-enough constant C,  n  2ℓ d Õ p α (G)p β (G) 6 δ −O(1) · |SAWℓ (i, j)| 2 · α,β∈SAW ℓ (i, j) W ℓ−1 (x i , x j )2 . (The constant C depends on W and the moments of µ.) ℓ−1 −(ℓ−1) W (x i , x j ). By Lemma 4.8, Lemma 4.9, and Lemma 2.1, Proof of Theorem 4.1. Let B i j  λ 2 there is matrix polynomial P(G), computable to 1/poly(n)-accuracy in time n poly(1/δ) by color coding, such that Tr PB T > δ O(1) ( kP k 2 )1/2 ( kB k 2)1/2 . At the same time, B − A has entries (B − A)i j  Õ  λ  ℓ−1 t 36 t 6 r λ2 Πt (x i , x j ) where the Πt projects to the t-th eigenspace of W. Since W is bounded, choosing ℓ a large enough multiple of log n ensures that kB − Ak 2 6 n −100 kB k 2 , so the theorem now follows by standard manipulations.  4.3 Proofs of Lemmas Proof of Lemma 4.8. As usual, we simply expand p, obtaining   p α (G) | x i , x j   x  " d n Ö d ab∈α ℓ · (Wx a ,x b − 1) | x i , x j n ·W ℓ−1 # (x i , x j ) .  We will need some small facts to help in proving Lemma 4.9. Fact 4.10. If ℓ − t > C log n for large enough C  C(W), then λ2t 2 x,y∼µ W ℓ−t (x, y)2 6 (1 + o(1)) · ℓ x,y∼µ W (x, y)2 . Also, for any t 6 ℓ, λ 2t 2 x,y∼µ W ℓ−t (x, y)2 6 r · where r is the rank of W. 35 ℓ x,y∼µ W (x, y)2 . Proof. Using the eigendecomposition of W, we have that similarly ℓ x,y∼µ W (x, y)2  Í x,y∼µ W ℓ−t λ 2ℓ . If i > 2, then i 26 i 6 r 2(ℓ−t) λ 2t 2 λi (x, y)2  Í 2(ℓ−t) 26 i 6 r λi and 2(ℓ−t)  λ2ℓ 6 λ2ℓ 2 (λ i /λ 2 ) 2 /n by our assumption that ℓ − t > C log n for large enough C. This finishes the proof of the first claim; the second one is similar.  Proof of Lemma 4.9. Pairs α, β which share only the vertices i, j each contribute exactly ℓ−1 W (x i , x j )2 to the left-hand side, by Lemma 4.8. Consider next the contribution of α, β whose shared edges form paths originating at i and j. Suppose there are t such shared edges. Then pα pβ     2ℓ−t Ö d n x d n (1 + O(d/n))t   2ℓ−t ab∈α△β W(x a , x b ) · W Ö ab∈α∩β 2(ℓ−t−1) (W(x a , x b ) + O(d/n)) (x, y)2 , where for the second equality we used the assumption x∼µ W(x, y)  1 for every y. If ℓ − t > C log n for the constant in Fact 4.10, then this is at most (1 + o(1))   2ℓ−t d n λ −2t 2 W 2(ℓ−1) (x, y)2 , and for every t 6 ℓ it is at most r ·   2ℓ−t d n λ −2t 2 W 2(ℓ−1) (x, y)2 . There are at most |SAWℓ (i, j)| 2 /n t · t choices for such pairs α, β, except when t  ℓ, in which case there are |SAWℓ (i, j)| 2 /n t−1 choices. So the total contribution from such α, β is at most |SAWℓ (i, j)| 2 · x,y W ℓ−1 6 δ−O(1) |SAWℓ (i, j)| 2 ©Õ (x, y)2 · ­ x,y W ℓ−1 «t 6 ℓ/2 td −t λ −2t 2 + nr · Õ ℓ > t>ℓ/2 (x, y)2 . ª td −t λ −2t 2 ® ¬ It remains to handle pairs α, β which share t vertices and s edges for t > s. If t, s 6 ℓ − 2, then there are only n 2(ℓ−1)−s ℓ O(t−s) choices for such a pair α, β. The contribution of each such pair we bound as follows pα pβ  Now,     2ℓ−s d n Ö ab∈α∩β Ö (G ab − nd )2 · W x a ,x b . ab∈α△β (G ab − nd )2 | x  nd (W(x a , x b ) + O(d/n)) by straightforward calculations, so the above is (1 + O(d/n)) s   2ℓ−s d n 6 (1 + O(d/n)) s x   2ℓ−s d n Ö ab∈α∩β λ 22ℓ−s Ö a∈α∪β 36 Ö W(x a , x b ) · ab∈α△β W(x a , x b ) − degα,β (a)/2 µ(x a ) where degα,β (a) is the degree of the vertex a in the graph α ∪ β. Any degree-2 vertices simply contribute 1 in the above, since x∼µ 1/µ(x)  1. There are at most t − s vertices of higher degree; they may have degree at most 4. They each contribute at most some number C  C(µ) by the niceness assumptions on µ. So the above is at most (1 + o(1))   2ℓ−s d n λ22ℓ−s exp{O(t − s)} . Putting things together as in Lemma 3.7 finishes the proof.  Proof of Lemma 4.6. By averaging, there is some v among v1 , . . . , v k such that hP, vv ⊤i > δ δ · kP k · kM k > · kP k · kvv ⊤ k k k where the second inequality uses M  0. Renormalizing, we may assume kP k has Frobenious norm 1 and v is a unit vector; in this case we obtain hv, Pvi > δ/k. Writing out the eigendecomposition Í of P, let P  ni1 λ i u i u ⊤ and we get i n Õ i1 By Cauchy-Schwartz, and hence inequality. Ín i1 n Õ i1 λ i hv, u i i 2 > δ/k n Õ λ i hv, u i i 2 6 λ 2i hv, u i i 2 > (δ/k)2 , while Ín i1 i1 λ 2i hv, u i i 2 ! 1/2 λ 2i  1. Now the Lemma follows from Markov’s  5 Tensor estimation for mixed-membership block models 5.1 Main theorem and algorithm Theorem 5.1 (Constant-degree partial recovery for mixed-membership block model). There is a constant C such that the following holds. Let G(n, d, ε, k, α) be the mixed-membership block model. For O(1) with the every δ ∈ (0, 1) and d(n), ε(n), k(n), α(n), there is an algorithm with running time n O(1)+1/δ following guarantees when def δ  1− k 2 (α + 1)2 >0 ε2 d and k, α 6 n o(1) and ε 2 d 6 n o(1) . k (samples from the α, k Dirichlet distribution are Let σ, G ∼ G(n, d, ε, k, α). Let t  (α + 1) · k+α approximately uniform over t coordinates). Given G, the algorithm outputs probability vectors τ1 , . . . , τn ∈ ∆k−1 such that   1 C 1 . − corr(σ, τ) > δ t k 37 (Recall the definition of correlation from (1.5).)29 Let c ∈ (0, 1) be a small-enough constant. Let C(c) ∈ Ž be a large-enough constant (different from the constant in the theorem statement above). There are three important parameter regimes: 1. Large δ, when δ ∈ [1 − c, 1). 2. Small δ, when δ ∈ (1 − c, 1/k 1/C ). This is the main regime of interest. In particular when k(n) → ∞ this contains most values of δ. 3. Tiny δ, when δ ∈ (0, 1/k 1/C ]. (This regime only makes sense when k(n) 6 O(1).) Let Ginput be an n-node graph. Algorithm 5.2 (Main algorithm for mixed-membership model). Let η > 0 be chosen so that k 2 (α+1)2 1 − ε2 d(1−η) > δ 2 and o(1) > η > n −γ for every constant γ > 0. (This guarantees that enough edges remain in the input after choosing a holdout set.) 1. Select a partition of [n] into A and A at random with |A|  ηn. Let G  A ∩ Ginput 2. If δ is large, run Algorithm 5.5 on (Ginput , G, A). 3. If δ is small, run Algorithm 5.4 on (Ginput , G, A). 4. If δ is tiny, run Algorithm 5.3 on (Ginput , G, A). Algorithm 5.3 (Tiny δ). foo 1. Run the algorithm from Theorem 4.3 on G with parameters (1 − η)d, k, ε, α to obtain a vector x ∈ ’n−ηn . (4) (3) 2. Evaluate the quantities s x  S3 (Ginput \ G, x) and s x  S4 (Ginput \ G, x), the polynomials (4) from Lemma 5.8. If s x < C(n, α, k, ε, d, η), output random labels τ1 , . . . , τn . (The scalar C(n, α, k, ε, d, η) depends in a simple way on the parameters.) (3) 3. If s x < 0, replace x by −x. 4. Run the cleanup algorithm from Lemma 5.11 on the vector x, padded with zeros to make a length n vector. Output the resulting τ1 , . . . , τn . Algorithm 5.4 (Small δ). foo 1. Using color coding, evaluate the degree-log n/poly(δ) polynomial P(G)  (Pi jk (G)) from Lemma 5.6. (This takes time n poly(1/δ) .) 2. Run the 3-tensor to 4-tensor lifting algorithm (Theorem 7.14) on P(G) to obtain a 4-tensor T. 3. Run the low-correlation tensor decomposition algorithm (Corollary 7.3) on T, implementing (4) the cross-validation oracle O as follows. For each query x ∈ ’n−ηn , compute s x  S4 (Ginput \ (4) G, x), the quantity from Lemma 5.9. If s x > C(n, d, k, ε, α, η) (distinct from the C above, again depending in a simple way on the parameters), output YES, otherwise output NO. The tensor decomposition algorithm returns unit vectors x1 , . . . , x k ∈ ’n−ηn . 29The requirement ε2 d 6 n o(1) is for technical convenience only; as ε2 d increases the recovery problem only becomes easier. 38 (3) 4. For each x1 , . . . , x k , compute s i for which (4) si (4)  S3 (Ginput \ G, x i ) and s i  S4 (Ginput \ G, x i ). For any x i < C(n, d, k, ε, α, η), replace x i with a uniformly random unit vector. For any (3) si < 0, replace x i with −x i . x i for which 5. Run the algorithm from Lemma 5.10 on (x1 , . . . , x k ) (padded with zeros to make an n × k matrix) and output the resulting τ1 , . . . , τn . Algorithm 5.5 (Large δ). foo 1. Using color coding, evaluate the degree-log n/poly(δ) polynomial P(G)  (Pi jk (G)) from Lemma 5.6. (This takes time n poly(1/δ) .) 2. Run the 3-tensor to 4-tensor lifting algorithm (Theorem 7.14) on P(G) to obtain a 4-tensor T. 3. Run the low-correlation tensor decomposition algorithm on T, obtaining unit vectors x1 , . . . , x k . (4) (4) 4. For each x i , compute the quantity s i  S4 (Ginput \ G, x i ) from Lemma 5.9. If s i < C(n, d, k, ε, α, η), replace x i with a uniformly random unit vector. (The scalar threshold C(n, d, k, ε, α, η) depends in a simple way on the parameters.) (3) (3) 5. For each x i , compute the quantity s i  S3 (Ginput \ G, x i ) from Lemma 5.9. If s i < 0, replace x i with −x i . 6. Run the algorithm from Lemma 5.10 on the matrix x  (x1 , . . . , x k ) and output the resulting τ1 , . . . , τn . We will analyze each of these algorithms separately, but we state the main lemmas together because many are shared among tiny, small, and large δ cases. Two of the algorithms use the low-correlation tensor decomposition algorithm as a black box; Corollary 7.3 in Section 7 captures the guarantees of that algorithm. The first thing we need is Theorem 4.3, which describes a second-moment based algorithm used as a subroutine by Algorithm 5.3. (This subroutine was already analyzed in Section 4.) Theorem (Restatement of Theorem 4.3). For every δ > 0 and d(n), ε(n), k(n), α(n), there is an algoO(1) rithm with running time n O(1)+1/δ with the following guarantees when def δ  1− k 2 (α + 1)2 >0 ε2 d and k, α 6 n o(1) and ε 2 d 6 n o(1) . Let σ, G ∼ G(n, d, ε, k, α) and for s ∈ [k] let v s ∈ ’n be given by v s (i)  σi (s) − 1k . The algorithm outputs a vector x such that hx, v1 i 2 > δ′ kx k 2 kv1 k 2 , for some δ′ > (δ/k)O(1) .30 The proofs of all the lemmas that follow can be found later in this section. Next, we state the tensor estimation lemma used to analyze the tensor P computed in Algorithm 5.4 and Algorithm 5.5. 30The requirement ε2 d 6 n o(1) is for technical convenience only; as ε2 d increases the recovery problem only becomes easier. 39 Lemma 5.6. Suppose def δ  1− k 2 (α + 1)2 >0 ε2 d ε 2 d 6 n 1−Ω(1) and k, α 6 n o(1) . and Í For a collection σ1 , . . . , σn of probability vectors, let V(σ)  s∈[k] v s⊗3 , where the vectors v s ∈ ’n have entries v s (i)  σi (s) − 1k . Let w s ∈ ’n have entries w s (i)  v s (i) + √1 . (Note that hw s , w t i  0 for Í k α+1 s , t.) Let W(σ)  s∈[k] w s⊗3 . If G ∼ G(n, d, ε, α, k), there is a degree O(log n/δ O(1) ) polynomial P(G) ∈ (’n )⊗3 such that  σ,G hP(G), W(σ)i  1/2  1/2  2 2 · σ,G kW(σ)k σ,G kP(G)k > δ O(1) Furthermore, P can be evaluated up to (1 + 1/poly(n)) multiplicative error (whp) in time n poly(1/δ) . Two of our algorithms use the low-correlation tensor decomposition algorithm of Corollary 7.3. That corollary describes an algorithm which recovers an underlying orthogonal tensor, but the tensor W is not quite orthogonal. The following lemma, proved via standard matrix concentration, captures the notion that W is close to orthogonal. Lemma 5.7. Let σ1 , . . . ,√σn be iid draws from the α, k Dirichlet distribution. Let w s ∈ ’n be given by w s (i)  σi (s) − 1k (1 − 1/ α + 1). Then as long as k, α 6 n o(1) , with high probability 1 Õ ws w⊤ s  (1 + n −Ω(1) ) · Id . ) · Id  2 k kw s k s1 k (1 + n −Ω(1) All of the algorithms perform some cross-validation using the holdout set A. The next two lemmas offer what we need to analyze the cross-validations. Lemma 5.8. Let n 0 , n 1 satisfy n 0 + n 1  n. Let A ⊆ [n] have size |A|  n 1 > n Ω(1) . Let k  0 k(n), d  d(n), ε  ε(n), α  α(n) > 0 and α, k, ε 2 d 6 n o(1) . Let σ ∈ ∆nk−1 . Let v s ∈ ’n0 have entries 1 v s (i)  σi (s) − k . Let τ1 , . . . , τn1 be iid from the α, k Dirichlet distribution. Let G be a random bipartite graph on vertex sets A, [n] \ A, with edges distributed according to the n, d, ε, k, α mixed-membership model with labels σ, τ. Let x ∈ ’n0 . For a ∈ A, let Pa (G, x) be the expression Õ Pa (G, x)  (G ai − nd )(G a j − nd )(G ak − nd )x i x j x k . i jk∈A distinct Let S3 (G, x) be S3 (G, x)  Õ a∈A Pa (G, x) . There is a number C  C(n, d, k, ε, α, n1) such that  G,τ        C · S3 (G, x) − Õ hv , xi 3 s s∈[k] kv s k 3 40 > n −Ω(1)        6 exp(−n Ω(1) ) . Similarly, there are scalars C(n, d, k, ε, α, n1), C ′(n, d, k, ε, α, n 1) such that the following holds. For a ∈ A, let Õ Q a (G, x)  (G ai − nd )(G a j − nd )(G ak − nd )(G aℓ − nd )x i x j x k x ℓ . i jkℓ∈A distinct Õ and let R a (G, x)  Finally let S4 (G, x)  C · Then  G,τ        S4 (G, x) − (G ai − nd )(G a j − nd )x i x j . i j∈A distinct Õ a∈A Q a (G, x) − C ′ · Õ hv , xi 4 s s∈[k] kv s k 4 >n Õ a∈A −Ω(1)        R a (G, x) !2 . 6 exp(−n Ω(1) ) . Lemma 5.9. Under the same hypotheses as Lemma 5.8, there are S3 (G, x), S4(G, x), polynomials of degree 3 and 4, respectively, in x and in the edge indicators of G, such that  G,τ and  G,τ               S4 (G, x) − Õ hw , xi 4 s s∈[k] C · S3 (G, x) − kw s k 4 >n Õ hw , xi 3 s s∈[k] kw s k 3 where w 1 , . . . , w k are the vectors w s (i)  v s (i) + −Ω(1)        > n −Ω(1) √1 . k α+1 6 exp(−n Ω(1) ) ,        6 exp(−n Ω(1) ) , Finally, all of the algorithms have a cleanup phase to transform n-length vectors to probability vectors τ1 , . . . , τn ∈ ∆k−1 . The following lemma describes the guarantees of the cleanup algorithm used by the small and large δ algorithms, which takes as input vectors x correlated with the vectors w. Lemma 5.10. Let δ ∈ (0, 1) and k  k(n) ∈ Ž and α  α(n) > 0, with α, k 6 n o(1) . Suppose δ > 1/k 1/C for a big-enough constant C. There is a poly(n)-time algorithm with the following guarantees. Let σ1 , . . . , σn be iid draws from the α, k Dirichlet distribution. Let v1 , . . . , v k ∈ ’n be the vectors given by v s (i)  σi (s) − 1k . Let w 1 , . . . , w k ∈ ’n be the vectors given by w s (i)  v s (i) + √1 , so that hw s , w t i  0 for s , t. Let M  1. M −1/2 w s − ws ( kw s k 2 )1/2 6 Í T s w s w s . Let E be the event that 1 poly n for every s ∈ [k]. 2. kw s k  (1 ± 1/poly(n))( kw s k 2 )1/2 for every s ∈ [k]. 3. kv s k  (1 ± 1/poly(n))( kv s k 2 )1/2 for every s ∈ [k]. 41 k α+1 Suppose x1 , . . . , x k ∈ ’n are unit vectors such that for at least δk vectors w 1 , . . . , w m there exists t ∈ [k] such that hw s , x t i > δkw s k. The algorithm takes input x1 , . . . , x k and when E happens returns probability vectors τ1 , . . . , τn ∈ ∆k−1 such that   k+α 1 1 O(1) 2 O(1) · − corr(σ, τ) > δ kv k  δ . α+1 k k Finally, the last lemma captures the cleanup algorithm used by the tiny-δ algorithm, which takes a single vector x correlated with v1 . Lemma 5.11. Let δ ∈ (0, 1) and k  k(n) ∈ Ž and α  α(n) > 0, with α, k 6 n o(1) . Suppose δ 6 k 1/C for any constant C. There is a poly(n)-time algorithm with the following guarantees. Let σ1 , . . . , σn be iid draws from the α, k Dirichlet distribution. Let v1 , . . . , v k ∈ ’n be the vectors given by v s (i)  σi (s) − 1k . Let x ∈ ’n be a unit vector satisfying hx, v s i > δkv s k for some s ∈ [k]. On input x, the algorithm produces τ1 , . . . , τn ∈ ∆k−1 such that   O(1) δ corr(σ, τ) > k · 2 kv k  δ so long as the event E from Lemma 5.10 occurs. O(1)   k+α 1 1 · − . α+1 k k Analysis for tiny δ (Algorithm 5.3) Proof of Theorem 5.1, tiny-δ case. Let C ∈ Ž and 1 > δ > 0 be any fixed constants. We will prove that if k 6 δ C then the output of Algorithm 5.2 satisfies the conclusion of Theorem 5.1. Let x ∈ ’(1−η)n be the output of the matrix estimation algorithm of Theorem 4.3. By Markov’s inequality, with probability (δ/k)O(1) over G and σ1 , . . . , σ(1−η)n , the vector x satisfies hv, xi 2 > (δ/k)O(1) kv k 2 kx k 2 , where v ∈ ’(1−η)n is the vector v(i)  σi (1) − 1k . By our assumption k 6 δ C , this means that with probability δ O(1) the vector x satisfies hx, vi 2 > δ O(1) kx k 2 kv k 2. Now, the labels σ(1−η)n , . . . , σn and the edges from nodes 1, . . . , (1 − η)n to nodes (1 − η)n, . . . , n are independent of everything above. So, invoking Lemma 5.8, we can assume that the quantity (4) s x computed by Algorithm 5.3 satisfies (4) sx − Õ hv , xi 4 s s∈[k] kv s k 4 6 n −Ω(1) . (4) Now, if x satisfies hv s , xi 2 > δ O(1) kv s k 2 for some v s , then also s x > δ O(1) . On the other hand, if O(1) (4) s x > δ O(1) then there is some s such that hx, v s i 2 > δ k kv s k 2 . So choosing the threshold C in Algorithm 5.3 appropriately, we have obtained that with probability δ O(1) the algorithm reaches step 3 with a vector x which satisfies hx, v s i 2 > δ O(1) kv s k 2 , and otherwise the algorithm outputs random τ1 , . . . , τn . Step 3 is designed to check the sign of hx, v s i. Call x good if there is s ∈ [k] such that (3) hx, v s i > δ O(1) kv s k. If |s x | 6 δ O(1) then x there are v s , v t such that hv s , xi > δ O(1) kv s k and (3) hv t , xi 6 −δ O(1) kv t k, so both x and −x are good If |s x | > δ O(1) then clearly step 3 outputs a good vector. Since after step 3 the vector x is good, applying Lemma 5.11 finishes the proof in the tiny δ case.  42 Analysis for small and large δ (Algorithm 5.4, Algorithm 5.5) Proof of Theorem 5.1, small δ case. Let n 0  (1 − η)n and n 1  ηn with η as in Algorithm 5.2. By Markov’s inequality applied to Lemma 5.6, with probability δ O(1) the tensor P satisfies Í hP, Wi > δ O(1) kP k kW k, where W ∈ (’n0 )⊗3 is as in Lemma 5.6. Let M  s∈[k] w s w ⊤ s , where w s is as in Lemma 5.6. The vectors M −1/2 w s are orthonormal, and Lemma 5.7 guarantees that Í ws k kw − M −1/2 w s k 6 n −Ω(1) with high probability. Let W ′  s∈[k] (M −1/2 w s )⊗3 and let W4′  sk Í Then also hP, W ′i > δ O(1) kP k kW ′ k. By the guarantees of the 3-to-4 lifting algorithm (Theorem 7.14), finally we get hT, W4′ i > δ O(1) kT k kW4′ k. In order to conclude that Algorithm 5.4 successfully runs the low-correlation tensor decomposition algorithm, we have to check correctness of its implementation of the cross-validation oracle. This follows from Lemma 5.7, Lemma 5.9, the size of η, and a union bound over the exp(k/poly(δ)) 6 exp(n o(1) ) queries made by the nonadaptive implementation of the lowcorrelation tensor decomposition algorithm, and independence of the randomness in the holdout set. We conclude that with probability at least δ O(1) , the tensor decomposition algorithm returns unit vectors x1 , . . . , x k ∈ ’n0 such that a δ O(1) fraction of w s among w 1 , . . . , w k have x t such that hw s , x t i 2 > δ O(1) kw s k 2 . By the same reasoning as in the tiny δ case, using Lemma 5.9 after the signchecking step the same guarantee holds with the strengthened conclusion hw s , x t i > δ O(1) kw s k. Finally, we apply Lemma 5.10 (along with elementary concentration arguments to show that the event E occurs with high probability) to conclude that the last   step of Algorithm 5.4 gives τ1 , . . . , τn s∈[k] (M −1/2 w s) ⊗4 . such that (in expectation) corr(σ, τ) > δ O(1) 1 α+1 · k k+α − 1 k as desired.  5.2 Low-degree estimate for posterior third moment In this section we prove Lemma 5.6. The strategy is to apply Lemma 2.1 to find an estimator for the Í 3-tensor s∈[k] v s⊗3 . With that in hand, combining with the estimators in Section 4 for the second Í moments s∈[k] v s v sT is enough to obtain an estimator for W, since Õ w s⊗3  Õ Õ s∈[k] s∈[k]  (v s + c · 1)⊗3 (5.1) v s⊗3 + c(v s ⊗ v s ⊗ 1 + v s ⊗ 1 ⊗ v s + 1 ⊗ v s ⊗ v s ) + 1⊗3 (5.2) s∈[k] √1 , k α+1 where 1 is the all-1s vector, c  n poly(1/δ) and we have used that polynomial such that hR, Õ s∈[k] v s⊗3 i > δ O(1) ( kRk 2)1/2 ( Í Í Õ s∈[k] s∈[k] v s  0. Thus if R is a degree 2 v s⊗3 )1/2 and Q is similar but estimates s∈[k] v s v sT , then R and Q can be combined according to (5.2) to obtain the polynomial P from the lemma statement. 43 j k i Figure 1: A 3-armed star with arms of length 2. We will eventually use arms of length t ≈ log n. Thus in the remainder of this section we focus on obtaining such a polynomial R; we change notation to call this polynomial P. The first step will be to define a collection of polynomials {G α }α for all distinct i, j, k ∈ [n].  Definition 5.12. Any α ⊆ n2 can be interpreted as a graph on some nodes in [n]. Such an α is a long-armed star if it consists of three self-avoiding paths, each with ℓ edges, joined at one end at a single central vertex, at the other end terminating at distinct nodes i, j, k ∈ [n]. (See figure.) Let STARℓ (i, j, k) be the set of 3-armed stars with arms of length ℓ and terminal vertices i, j, k. For any  Î α ⊆ n2 let G α  ab∈α (x ab − nd ) be the product of centered edge indicators. The next two lemmas check the conditions to apply Lemma 2.1 to the sets {G α }α∈STARℓ (i, j,k) . Lemma 5.13 (Unbiased Estimator). Let i, j, k ∈ [n] all be distinct. Let α ∈ STARℓ (i, j, k). Í For a collection of probability vectors σ1 , . . . , σk , let V(σ)  s∈[k] v s⊗3 where v s (i)  σi (s) − 1k . Let G ∼ G(n, d, ε, α 0, k).  α G | σi , σ j , σk   εd  n  3ℓ  1 k(α 0 + 1)  3(ℓ−1) · C 3 · V(σ)i jk . O(1) Here α 0 > 0 is the Dirichlet concentration paramter, unrelated to the graph α, and C 3  1/(k O(1) α 0 constant related to third moments of the Dirichlet distribution. ) is a Lemma 5.14 (Approximate conditional independence). If def δ  1− k 2 (α 0 + 1)2 >0 ε2 d k, α 0 6 n o(1) and ε 2 d 6 n o(1) . and and ℓ > C log n/δ O(1) for a large enough constant C, then for G ∼ G(n, d, ε, k, α 0), h i V(σ)2i jk · Õ α,β∈STARℓ (i, j,k) G α G β 6 1/δ O(1) · Õ α,β∈STARℓ (i, j,k) Now we can prove Lemma 5.6. 44   G α V(σ)i, j,k ·   G β V(σ)i, j,k . Proof of Lemma 5.6. As discussed at the beginning of this section, it is enough to find an estimator for the tensor V(σ). Lemma 5.13 and Lemma 5.14 show that Lemma 2.1 applies to each set of polynomials STARℓ (i, j, k). The conclusion is that for every distinct i, j, k ∈ [n] there is a degree log n poly(1/δ) polynomial P(G)i jk so that P(G)i jk V(σ)i jk ( P(G)2i jk )1/2 · ( V(σ)2i jk )1/2 > Ω(1) . One may check that the entries i, j, k for i, j, k all distinct of the tensor V(σ) comprise nearly all of its 2-norm. That is, Õ V(σ)2i, j,k > (1 − o(1)) kV(σ)k 2 . i, j,k distinct This is sufficient to conclude that the tensor-valued polynomial P(G) whose (i, j, k)-th entry is Pi, j,k (G) when i, j, k are all distinct and is 0 otherwise is a good estimator of V(σ) (see Fact A.2). Thus, σ,G hP(G), V(σ)i   1/2 > Ω(1) .  1/2   2 2 · kV(σ)k kP(G)k σ,G σ,G 5.2.1 Details of unbiased estimator We work towards proving Lemma 5.13. We will need to assemble a few facts. The first will help us control moment tensors of the Dirichlet distribution. The proof can be found in the appendix. Fact 5.15 (Special case of Fact A.3). Let σ be distributed according to the α, k Dirichlet distribution. Let σ̃  σ − 1k 1. There are numbers C 2 , C 3 depending on α, k so that for every x1 , x2 , x3 in ’k with Í s∈[k] x i (s)  0, h σ̃, x1 ih σ̃, x2 i  C 2 hx1 , x2 i σ and σ h σ̃, x1 ih σ̃, x2 ih σ̃, x3 i  C 3 Furthermore, 1 k(α + 1) Now we can prove Lemma 5.13. C2  and Õ s∈[k] C3  x1 (s)x2 (s)x3 (s) . 1 k O(1) α O(1) . Proof of Lemma 5.13. For any collection of σ’s and α ∈ STARℓ (i, j, k),  εd [G | σ]  n G α  3ℓ Ö (a,b)∈α h σ̃a , σ̃b i Let a be the central vertex of the star α. Taking expectations over all the vertices in the arms of the star,   3ℓ   3(ℓ−1)  α  εd 1 G | σi , σ j , σk  h σ̃i , σ̃a ih σ̃ j , σ̃a ih σ̃k , σ̃a i . σa n k(α 0 + 1) Finally, using the second part of Fact 5.15 completes the proof.  45 5.2.2 Details of approximate conditional independence Í We prove Lemma 5.14, first gathering some facts. In the sum α,β∈STARℓ (i, j,k) G α G β , the terms α, β which (as graphs) share only the vertices i, j, k will not cause us any trouble, because such G α and G β are independent conditioned on σi , σ j , σk . Fact 5.16. If α, β ∈ STARℓ (i, j, k) share only the vertices i, j, k, then for any collection σ of probability vectors,  α β   α   β  G G | σi , σ j , σk  G | σi , σ j , σk · G | σi , σ j , σk . Proof. To sample G α , one needs to know σa for any a ∈ [n] with nonzero degree in α, and similar for b ∈ [n] and G β . The only overlap is σi , σ j , σk .  The next fact is the key one. Pairs α, β which share vertices forming paths originating at i, j, Í and k make the next-largest contribution (after α, β sharing only i, j, k) to α,β G α G β . Fact 5.17. Let i, j, k ∈ [n] be distinct. Let V(σ)i jk be as in the Lemma 5.14. Let C 2 ∈ ’ be as in Fact 5.15. Let α, β ∈ STARℓ (i, j, k) share s vertices (in addition to i, j, k) for some s 6 2t , and suppose the shared vertices form paths in α and β starting at i, j, and k. Then V(σ)2i jk · α β G G 6ε −2s   −s d n −s (1 + O(d/n))  1 · k(α 0 + 1)  −2s  ·  G α V(σ)i jk ·   G β V(σ)i jk . Proof. Let σα∩β be the σ’s corresponding to vertices sharerd by α, β. Let i ′ , j ′ , k ′ be the last shared vertices along the paths beginning at i, j, k respectively. We expand G α G β and use conditional independence of the G e ’s given the σ’s: Gα Gβ  σ i′ , j′ ,k′     (G α∩β )2 |σi′ , σ j′ , σk′ ·  G α\β | σi′ σ j′ σk′ ·  G β\α | σi′ σ j′ σk′  . Both G α\β and G β\α are long-armed stars with terminal vertices i ′ , j ′ , k ′. The arm lengths of G α\β total 3ℓ − s. By a similar argument to Lemma 5.13, G α\β is an unbiased estimator of V(σ)i′ j′ k′ with  G α\β | σi′ , σ j′ , σk ′   εd  n  3ℓ−s  1 k(α 0 + 1)  3(ℓ−1)−s · C 3 · V(σ)i′ , j′ ,k′ and the same goes for G β\α . Furthermore,  (G α∩β 2 ) | σi′ , σ j′ , σk ′    |α∩β| d  n  Ö     ′ , σ j′ , σk ′  . (1 + εh σ̃ , σ̃ i + O(d/n)) σ a i b   (a,b)∈α∩β    By our assumption that α ∩ β consists just of paths, every subset of edges in the graph α ∩ β contains   α∩β 2 a vertex of degree 1. Hence, (G ) | σi′ , σ j′ , σk′  (1 + O(d/n))|α∩β| (d/n)|α∩β| . Putting these together, α β s 6ℓ−2s G G  (1 + O(d/n)) ε   6ℓ−s  d n 46 1 k(α 0 + 1)  6(ℓ−1)−2s C 32 V(σ)2i jk G α V(σ)i jk to obtain At the same time, one may apply Lemma 5.13 to   α  G V(σ)i jk · The lemma follows. β G V(σ)i jk   εd  n  6ℓ  1 k(α 0 + 1)  6(ℓ−1) C 32 ·  V(σ)2i jk σ i ,σ j ,σ k 2 .  The last fact will allow us to control α, β which intersect in some way other than paths starting at i, j, k. The key idea will be that such pairs α, β must share more vertices than they do edges. Fact 5.18. Let i, j, k ∈ [n] be distinct. Let V(σ)i jk be as in the Lemma 5.14. Let C 2 ∈ ’ be as in Fact 5.15. C 2  k(α10 +1) . Let α, β ∈ STARℓ (i, j, k) share s vertices (in addition to i, j, k) and r edges. Then V(σ)2i jk · α β G G 6ε −2r   −r d n  · C 2−2s · k O(s−r) (1 + α 0 )O(s−r) · Proof. Expanding as usual, α β G G    6ℓ−r d n σ Ö ab∈α△β h σ̃a , σ̃b i · Ö ab∈α∩β   G α V(σ)i jk ·  G β V(σ)i jk . (1 + εh σ̃a , σ̃b i + O(d/n)) . Any nontrivial edge-induced subgraph of α ∩ β contains a degree-1 vertex; using this to expand the second product and simplifying with σ̃a  0, the above is   6ℓ−r d n σ Ö ab∈α△β h σ̃a , σ̃b i · (1 + O(d/n))r . For every degree-2 vertex in α△β we can use Fact A.3 to take the expectation. Each such vertex contributes a factor of C 2 and there are at least 3ℓ − O(s − r) such vertices. The remaining expression will be bounded by 1. The fact follows.  Now we can prove Lemma 5.14. Proof of Lemma 5.14. Let us recall that our goal is to show h i V(σ)2i jk · Õ α,β∈STARℓ (i, j,k) k 2 (α +1)2 G α G β 6 δ O(1) ·   Õ α,β∈STARℓ (i, j,k)     G α V(σ)i jk ·  G β V(σ)i jk  where δ  1 − ε20 d . Let c  G α V(σ)i jk · G β V(σ)i jk . (Notice this number does not depend on α or β.) The right-hand side above simplifies to |STARℓ (i, j, k)| 2 · c. On the left-hand side, what is the contribution from α, β sharing s vertices? First consider what happens with s 6 t/2 and the intersecting vertices form paths in α and β starting at i, j, k. Choosing a random pair α, β from STARℓ (i, j, k), the probability that they intersect along paths of length s1 , s2 , s3 starting at i, j, k respectively is at most n −s1−s2 −s3 . There are at most (1 + s 2 ) choices 47 for nonnegative integers s1 , s2 , s3 with s1 + s2 + s3  s. By Fact 5.17, such terms therefore contribute at most  q |STARℓ (i, j, k)| 2 · ε c· n −s d n (1 + O(d/n))  −2s C 2−2s · s 2  c · |STARℓ (i, j, k)| 2 · (ε 2 dC 22 (1 + O(d/n)))−s · s 2 where C 2  k(α10 +1) . By hypothesis, δ > 0. Consider the sum of all such contributions for s 6 t/2; this is at most 2 c · |STARℓ (i, j, k)| · t/2 Õ s0 (1 + s 2 ) ·  k 2 (α 0 +1)2 ε2 d s 6 δ O(1) · c · |STARℓ (i, j, k)| 2 . Next, consider the contribution from α, β which share s vertices in some pattern other than those considered above. Unless α  β, this means α, β share at least one more vertex than the number r of edges that they share. Suppose α , β and let s − r  q. There are t O(q) patterns in which such an intersection might occur, and each occurs for a random pair α, β ∈ STARℓ (i, j, k) with probabilty n −s . So using Fact 5.18, the contribution is at most 2 c · |STARℓ (i, j, k)| · t  2 q Õ ε d n q1 · k O(q) (1 + α 0 )O(q) t O(q) By the hypotheses k, α  n o(1) and ε 2 d  n 1−Ω(1) , this is all o(c|STARℓ (i, j, k)| 2 ). Finally, consider the case α  β. Then, using Fact 5.18 again, the contribution is at most c · |STARℓ (i, j, k)| 2  ε2 d k 2 (α 0 + 1)2  −t k O(1) α O(1) which is o(c|STARℓ (i, j, k)| 2 ) because t ≫ log(n). Putting these things together gives the lemma.  5.3 Cross validation In this section we show how to use a holdout set of vertices to cross-validate candidate community membership vectors. The arguments are all standard, using straightforward concentration inequalities. At the end we prove the first part of Lemma 5.8, on the estimator S3 . The proof of the second part, on S4 is similar, using standard facts about moments of the Dirichlet distribution (see Fact A.3). The proof of Lemma 5.9 is also similar, using the discussion in Section 5.2 to turn estimators for moments of the v vectors into estimators for moments of the w vectors—we leave it to the reader. We will need a few facts to prove the lemma. Fact 5.19. Let n 0 , n 1 , A, k, d, ε, α, σ, v, τ, G, x, P be as in Lemma 5.8. Let a ∈ A. There is a number C  C(k, α) 6 poly(k, α) such that  εd Pa (G, x)  n G,τ 3 ·C· Õ Õ i jk∈A distinct s∈[k] 48 σi (s)σ j (s)σk (s)x i x j x k . Proof. Immediate from Fact 5.15.  Fact 5.20. Let n 0 , n 1 , A, k, d, ε, α, σ, v, τ, G, x, P be as in Lemma 5.8. Let a ∈ A. The following variance bound holds.  2 poly(k, α, ε, d) 2 Pa (G, x) − Pa (G, x) 6 . G,τ G,τ n3 Proof. Expanding Pa (G, x) and using that |hσ, σ′i| 6 1 for any σ, σ′ ∈ ∆k−1 we get  6 d Pa (G, x) 6 n G,τ 2 Õ x i x j x k x i′ x j′ x k ′ 6 i jk distinct i′ j′ k ′ distinct  6 d n · n 3 · kx k 12 .  Fact 5.21. Let n 0 , n 1 , A, k, d, ε, α, σ, v, τ, G, x, P be as in Lemma 5.8. Let a ∈ A. For some constant γ∗ (ε, d, k, α) and every γ∗ > γ > 0,  {|Pa (G, x)| > n γ } 6 exp(−n Ω(γ) ) G,τ Proof. The fact follows from a standard exponential tail bound on the degree of vertex a.  We can put these facts together to prove the S3 portion of Lemma 5.8 (as we discussed above, the S4 portion and Lemma 5.9 are similar). The strategy will be to use the following version of Bernstein’s inequality, applied to the random variables hG a , v ⊗3 i. The proof of the inequality is in the appendix. Proposition 5.22 (Bernstein wth tails). Let X be a random variable satisfying numbers R, δ, δ′ ∈ ’, {|X | > R} 6 δ and |X | · 1 |X|>R 6 δ′ . Let X1 , . . . , X m be independent realizations of X. Then  ( 1 m Õ Xi > t + δ i6m ′ )  X  0 and, for some  −Ω(1) · m · t 2 + mδ . 6 exp X2 + t · R Now we can prove Lemma 5.8. Proof of Lemma 5.8. We apply Proposition 5.22 to the n 1 random variables X a   εd n  −3 C −1 Pa (G, x) for a ∈ A, where C  C(k, α) is the number from Fact 5.20. (For each a ∈ A these are iid over G, τ.) ′ Take t  n 3/2−γ for a small-enough constant γ′ so that n 1 t 2 /n 3 > n γ for some constant γ, using the assumption n 1 > n Ω(1) . All together, we get     Õ Õ  1 Õ    ′ 3/2−γ′ Xa −  σs (i)σs (j)σs (k)x i x j x k > n 6 exp(n −γ )  G,τ   n 1 a∈A  s∈[k] i jk∈A distinct   49 for some constants γ, γ′ (possibly different from γ, γ′ above) and large-enough n. For any unit n0 x ∈ ’n0 and σ ∈ ∆k−1 , using that k 6 n o(1) it is not hard to show via Cauchy-Schwarz that Õ s∈[k] hv s , xi 3 − Õ Õ s∈[k] i jk∈A distinct σs (i)σs (j)σs (k)x i x j x k 6 n 1+o(1) . The lemma follows.  5.4 Producing probability vectors In this section we prove Lemma 5.10. The proof of Lemma 5.11 is very similar (in fact it is somewhat easier) so we leave it to the reader. Lemma (Restatement of Theorem 5.10). Let δ ∈ (0, 1) and k  k(n) ∈ Ž and α  α(n) > 0, with α, k 6 n o(1) . Suppose δ > 1/k 1/C for a big-enough constant C. There is a poly(n)-time algorithm with the following guarantees. Let σ1 , . . . , σn be iid draws from the α, k Dirichlet distribution. Let v1 , . . . , v k ∈ ’n be the vectors given by v s (i)  σi (s) − 1k . Let w 1 , . . . , w k ∈ ’n be the vectors given by w s (i)  v s (i) + √1 , so that hw s , w t i  0 for s , t. Let M  1. M −1/2 w s − ws ( kw s k 2 )1/2 6 Í k α+1 T s w s w s . Let E be the event that 1 poly n for every s ∈ [k]. 2. kw s k  (1 ± 1/poly(n))( kw s k 2 )1/2 for every s ∈ [k]. 3. kv s k  (1 ± 1/poly(n))( kv s k 2 )1/2 for every s ∈ [k]. Suppose x1 , . . . , x k ∈ ’n are unit vectors such that for at least δk vectors w 1 , . . . , w m there exists t ∈ [k] such that hw s , x t i > δkw s k. The algorithm takes input x1 , . . . , x k and when E happens returns probability vectors τ1 , . . . , τn ∈ ∆k−1 such that   1 k+α 1 O(1) 2 O(1) corr(σ, τ) > δ kv k  δ . · − α+1 k k First some preliminaries. Let σ1 , . . . , σn be iid from the α, k Dirichlet distribution. There are two important families of vectors in ’n . Let 1 v s (i)  σi (s) − k   1 1 w s (i)  σi (s) − 1− √ . k α+1 We will also work with a normalized version of the v vectors: vs vs  . ( kv s k 2 )1/2 Í Í By construction, kv s k 2  1. Also by definition, s v s  s v s  0. Thus Í −1 k + s ,t hv s , v t i  0 and so by symmetry hv s , v t i  k−1 . We let 1 ws  vs + √ · n 50 r 1 k−1 h Í s vs , Í s vsi  so that hw s , w t i  0 for s , t. (In the facts which follow we sometimes write v as v when both normalizations are not needed; this is always noted.) We will want the following fact; the proof is elementary. Fact 5.23. Let σ, u, v, w as above, and suppose y is an n × k matrix whose rows are in ∆k−1 − 1k (that is they are shifted probability vectors). Then τ  y + 1k is a matrix whose rows are probability vectors, and τ satisfies n hτ, σi > hy, vi + . k The following fact√will be useful when δ is small but not tiny; i.e. δ < 1 − c for some fixed constant c but δ ≫ 1/ k. Fact 5.24. Suppose that x1 , . . . , x k are unit vectors and w 1 , . . . , w k are orthonormal. Also suppose that there is 1 > δ > 0 such that for at least δk vectors w s among w 1 , . . . , w k there exists a vector x t among x1 , . . . , x k such that hw s , x t i > δ. Then there is a permutation π : [k] → [k] such that if x  (x1 , . . . , x k ) is an n × k matrix and similarly for w,  1 1 hx, π · wi > δ5 − √ 4 k 1−δ  1/2 ! kx k kw k , where x  (x1 , . . . , x k ) is an n × k matrix and similarly for w. Proof. We will think of π as a matching of w 1 , . . . , w k to x1 , . . . , x k . Call x t good for w s if hw s , x t i > δ. First of all, by orthogonality of vectors w 1 , . . . , w k , any particular vector x t is good for at most 1/δ 2 vectors w s . Hence, there is a set S of δ4 k vectors w s such that for each w s there exists a good x t and all the good x t ’s are distinct. Begin by matching each w s ∈ S to its good x t . Let π be the result of extending that matching randomly to a perfect matching of k to k. Í We need to lower bound s <S hw s , x π(s) i. Consider that for a particular t, −hx t , w π−1(t) i 6 ( hx t , w π−1(t) i 2 )1/2 . The distribution of π−1 (t) is uniform among all s < S. So  1 1 1 Õ hw s , x t i 2 6 hx t , w π−1 (t) i  k − |S| k 1 − δ4 2 since Í s <S s∈[k] hw s , xt i2 6 1. It follows that    1/2 1 1 hx t , w π−1(t) i > − √ . 4 k 1−δ  1/2 √  Í 1 Therefore, k hw , x i > − . Thus there is some choice of π such that s π(s) s <S 1−δ4   1/2 √ Í 1 . Hence for this π one gets s <S hw s , x π(s) i > − k 1−δ4 Õ s∈[k] √  1 hw s , x π(s) i > δ 5 k − k 1 − δ4  1/2 51  1 1  δ5 − √ 4 k 1−δ  1/2 ! kx k kw k .  The next fact serves the same purpose as the previous one but in the large δ case (i.e. δ close to 1). Fact 5.25. Under the same hypotheses as Fact 5.24, letting δ  1 − ε for some ε > 0, there is a permutation π : [k] → [k] such that hx, π · wi > (1 − 9ε)kx k kw k. Proof. As in the proof of Fact 5.24, we construct a matching π by first matching a set S of at least δ4 k > (1 − 4ε)k vectors w s to corresponding x t . Then we match the remaining vectors arbitrarily. For any s, t we know hw s , x t i > −1. So the result is hx, π · wi > (1 − 5ε)k − 4εk  (1 − 9ε)k  (1 − 9ε)kx k kw k .  We will also want a way to translate a matrix correlated with w to one correlated with v, so that we can apply Fact 5.23. Fact 5.26. Suppose v is an n × k matrix whose rows are centered probability vectors and w  v + c is a coordinate-wise additive shift of v. Suppose y is also an n × k matrix whose rows are centered probability vectors shifted by c in each coordinate (so y − c is a matrix of centered probability vectors). Then the shifted matrix y − c satisfies hy − c, vi > hy, wi − c 2 nk . Proof. By definition, hy − c, vi  hy, vi. Since v  w − c, we get hy − c, vi  hy, vi  hy, wi − chy, 1i  hy, wi − c 2 nk .  Proof of Lemma 5.10. First assume δ < 1 − c for any small constant c. Let π be the permutation guaranteed by Fact 5.24 applied to the vectors x1 , . . . , x k and M −1/2 w 1 , . . . , M −1/2 w k . (Without loss of generality reorder the vectors so that π is the identity permutation.) Since 1 − c > δ > 1/k 1/C for big-enough C and small-enough c (which are independent of n, k) and the guarantee of Fact 5.24, by event E we get that hx, wi > δ O(1) kx k kw k . So by taking a correlation-preserving projection of x into the set of matrices whose rows are shifted probability vectors, we get a matrix y with the guarantee hy, wi > δ O(1) k y k kw k and k y k > δ O(1) kw k . Applying Fact 5.26, we obtain hy − c, vi > hy, wi − c 2 nk  hy, wi − where c  √1 . k α+1 Putting things together and using hy − c, vi > δ O(1) So applying Fact 5.23 finishes the proof in this case. 52 kv k 2 6 kv k 2 . kw k 2 k kw k 2 and the event E, we get Now suppose δ > 1 − c for a small-enough constant c. Then using event E and Fact 5.25, there is π such that hx, wi > (1 − O(c))kx k( kw k 2) (where again we have without loss of generality reordered the vectors so that π is the identity permutation). Now taking the Euclidean projection of x· by ( kwk 2 )1/2 into the n × k matrices whose rows are centered kxk 1 √ c  , we get a matrix y which again satisfies hy, k α+1 probability vectors shifted entrywise wi > (1 − O(c))k y k kw k and k y k > (1 − O(c))kw k, so (using event E), hy, wi > (1 − O(c)) kw k 2. Removing the contribution from hy, 1i, this implies that hy − c, vi > (1− O(c)) kv k 2 . For c small enough, this is at least δ O(1) kv k 2 .  Applying Fact 5.23 finishes the proof. 5.5 Remaining lemmas We provide sketches of the proofs of Lemma 5.7 and Lemma 5.10, since the proofs of these lemmas use only standard techniques. √ Proof sketch of Lemma 5.7. For σ ∈ ’k , let σ̃  σ − (1 − 1/ α + 1)/k. Standard calculations show that 1 Id. It follows by standard if σ is drawn from the α, k Dirichlet distribution then σ̃ σ̃ ⊤  k(α+1) Í matrix concentration and the assumption k, α 6 n o(1) that the eigenvalues of n1 i 6 n σ̃i σ̃i⊤ are all 1 ± n −Ω(1) , where σ1 , . . . , σn are iid draws from the α, k Dirichlet distribution. For the second part of the Lemma, use the first part to show that kvv ss k − w ′s 6 1/poly(k). Then when k > δ−C for large-enough C, if hx, v s i 3 > δ O(1) kv s k 3 it follows that also hx, w s i > δ O(1) − 1/poly(k) > δ O(1) . The lemma follows.  Proof sketch of Lemma 5.10. If δ < 1 − Ω(1), then δ 2 /2 > δ O(1) , so the Lemma follows from standard concentration and Theorem 2.3 on correlation-preserving projection. On the other hand, if δ > ˜ k−1 )n by 1 − o(1), then kv ′ − σ̃k 6 o(1) · k σ̃k, so the same is also true for the projection of v ′ into (∆ convexity and the lemma follows.  6 Lower bounds against low-degree polynomials at the Kesten-Stigum threshold In this section we prove two lower bounds for k-community partial recovery algorithms based on low-degree polynomials. 6.1 Low-degree Fourier spectrum of the k-community block model Theorem 6.1. Let d, ε, k be constants. Let µ : {0, 1} n×n → ’ be the relative density of SBM(n, d, ε, k) with respect to G(n, nd ). Let µ 6 ℓ be the projection of µ to the degree-ℓ polynomials with respect to the norm 53 induced by G(n, nd ).31 For any constant δ > 0 and ξ > 0 (allowing ξ 6 o(1)), kµ 6 ℓ k is ( > n Ω(1) if ε 2 d > (1 + δ)k 2 , ℓ > O(log n) 6 n 2ξ if ε 2 d < (1 − δ)k 2 , ℓ < n ξ . This proves Theorem 1.9 (see discussion following statement of that theorem). To prove the theorem we need the following lemmas. Lemma 6.2. Let χα : {0, 1} n×n → ’ be the nd -biased Fourier character. If α ⊆ on n vertices, has any degree-one vertex, then G∼SBM(n,d,ε,k) n 2 , considered as a graph χα (G)  0 The proof follows from calculations very similar to those in Section 5, so we omit it. Proof of Theorem 6.1. The bound kµ 6 ℓ k > n Ω(1) when ε 2 d > (1 + δ)k 2 and ℓ ≫ log(n), follows from almost identical calculations to Section 5,32 so we omit this argument and focus on the regime ε 2 d < (1 − δ)k 2 . By definition and elementary Fourier analysis, kµ 6 ℓ k 2  Also by definition, b µ(α)  d G∼G(n, n ) Õ n 2 α⊆( ) ,|α| 6 ℓ µ(G)χα (G)  b µ(α)2 G∼SBM(n,d,ε,k) (6.1) χα where {χα } are the nd -biased Fourier characters. Thus, using Lemma 6.2 we may restrict attion to  the contribution of those α ⊆ n2 with |α| 6 ℓ and containing no degree-1 vertices. Fix such an α, and suppose it has C(α) connected components and V2 (α) vertices of degree 2 (considered again as a graph on [n]). Fact 6.3 (following this proof) together with routine computations shows that  G∼SBM(n,d,ε,k)  χα (G) 2  |α|   |α|  k −2(V(α)−C(α)) 6 1 + O( nd ) ·n −|α| ·(1−δ)|α| ·k 2(|α|−V(α)+C(α)) . 6 (1 + O( nd ))ε 2 nd  |α| Let c(α)  1 + O( nd ) · n −|α| · (1 − δ)|α| · k 2(|α|−V(α)+C(α)) be this upper bound on the contribution of α to the right-hand side of (6.1). It will be enough to bound (∗) 31That is, k f k  ( d) G∼G(n, n def  Õ c(α) n 2 α⊆( ) |α| 6 ℓ α has no degree 1 nodes f (G)2 )1/2 . 32The calculations in Section 5 are performed for long-armed stars; to prove the present result the analogous calculations should be performed for cycles of logarithmic lengh. Similar calculations also appear in many previous works. 54 Given any α as in the sum, we may partition it into two vertex-disjoint subgraphs, α 0 and α 1 , where α 0 is a union of cycles and no connected component of α 1 is a cycle, such that α  α 0 ∪ α 1 . Thus, Õ (∗) 6 α0 c(α 0 ) ! Õ α1 c(α 1 ) ! where α 0 ranges over unions of cycles with |α 0 | 6 ℓ and α 1 ranges over graphs on [n] with at most ℓ where all degrees are at least 2 and containing no connected component which is a cycle. Lemmas 6.4 and 6.5, which follow, the terms above as O(1) and n 2ξ , respectively, which finishes the proof.  Fact 6.3. Let U be a connected graph on t vertices where all degrees are at least 2. For each vertex v of U let σv ∈ ’k be a uniformly random standard basis vector. Let σ̃v  σv − 1k · 1. Then Ö (u,v)∈U h σ̃v , σ̃u i 6 tk −t+1 Proof. Consider a particular realization of σ1 , . . . , σt . Suppose all but m vertices v in U are adjacent to at least 2 vertices u 1 , u 2 such that σu1 , σv and σu2 , σv . In this case, Ö (u,v)∈U h σ̃v , σ̃u i 6 k −(t−m) . The probability of such a pattern of disagreements is at most k −m , unless m  t, in which case the probability is at most k −t+1 . The fact follows.  Lemma 6.4. For α ⊆ n 2 ,  def components in α. For constants ε, d, k, let c(α)  1 + O( nd ) ℓ 6 n 0.01 and     |α| let V(α) be the number of vertices in α, let C(α) be the number of connected · n −|α| · (1 − δ)|α| · k 2(|α|−V(α)+C(α)) Let  n : α has all degrees > 2, has no connected components which are cycles, |α| 6 ℓ . U α⊆ 2 Then Õ c(α) 6 O(1) . α∈U Proof. We will use a coding argument to bound the number of α ∈ U with V vertices, E edges, and C connected components. We claim that any such α is uniquely specified by the following encoding. To encode α, start by picking an arbitrary vertex v1 in α. List the vertices v1 , . . . , v |V | of α, each requiring log n bits, starting from v1 , using the following rules to pick v i . 1. If v i−1 has a neighbor not yet appearing in the list v1 , . . . , v i−1 , let v i be any such neighbor. 55 2. Otherwise, if v i−1 has a neighbor v j which (a) appears in the list v1 , . . . , v i−1 and (b) for which either j  1 or v j−1 is not adjacent to v j in α, and (c) for which if j , i ′ for i ′ 6 i − 1 being the minimal index such that v i′ , . . . , v i−1 is a path in α (i.e. v j , . . . , v i−1 are not a cycle in α) then reorder the list as follows. Remove vertices v j , . . . , v j′ where j ′ is the greatest index so that all edges v ℓ , v ℓ+1 exist in α for j 6 ℓ 6 j ′. Also remove vertices v i′ , . . . , v i−1 where i ′ is analogously the minimal index such that edes v ℓ , v ℓ+1 exist in α for i ′ 6 ℓ 6 i − 1. Then, append the list v j′ , v j′−1 , . . . , v j , v i−1 , . . . , v i′ . By construction, all of these vertices appear in a path in α. The new list retains the invariant that every vertex either preceeds a neighbor in α or has no neighbors in α which have not previous appeared in the list. 3. Otherwise, let v i be an arbitrary vertex in α in the same connected component as v i−1 , if some such vertices has not yet appeared in the list. 4. Otherwise, let v i be an arbitrary vertex of α not yet appearing among v1 , . . . , v i−1 . After the list of vertices, append to the encoding the following information. First, a list of the R (for removed) pairs v i , v i+1 for which there is not an edge (v i , v i+1 ) in α. This uses 2R log V bits. Last, a list of the edges in α which are not among the pairs v i , v i+1 (each edge encoded using 2 log V bits). We argue that the number R of removed pairs (and hence the length of their list in the encoding) is not too great. In particular, we claim R 6 2(E −V). In fact, this is true connected-component-wise in α. To see it, proceed as follows. Fix a connected component β of α. Let v t be the first vertex in β to appear in the list v1 , . . . , v |V | . Proceeding in increasing order down the list from v t , let (v r1 , v r1+1 ), (v r2 , v r2 +1 ), . . . be the pairs encountered (before leaving β) which do not correspond to edges in α (and hence will later appear in the list of removed pairs). Construct a sequence of subgraphs β j of β as follows. The graph β 1 is the line on vertices v t , . . . , v r1 . To construct the graph β j , start from β j−1 and add the line from v r j−1 +1 to v r j (by definition all these edges appear in β). Since v r j must have at least degree 2, it has a neighbor u j in β among the vertices v a for a < r j aside from v r j −1 . (If v r j had a neighbor not yet appearing in the list, then v r j +1 would have been that neighbor, contrary to assumption.) Choose any such neighbor and add it to β j ; this finishes construction of the graph β j . For later use, note that either adding the edge to u j turns β j \ b j−1 into a cycle or u j is not itself among the v r ’s, since otherwise in constructing the list we would have done a reordering operation. In each of the graphs β j , the number of edges is equal to the number of vertices. To obtain β, we must add E β − Vβ edges (where E β is the number of edges and β and Vβ is the number of vertices). We claim that in so doing at least one half of a distinct such edge must be added per β j ; we prove this via a charging scheme. As noted above, each graph β j \ β j−1 either contains v r j−1 as a degree-1 vertex or it forms cycle. If it contains a degree-1 vertex, by construction this vertex is not u j′ for any j ′ > j, otherwise we would have reordered. So charge β j to the edge which must be added to fix the degree-1 vertex. 56 In the cycle case, either some edge among the E β − Vβ additional edges is added incident to the cycle (in which case we charge β j to this edge), or some u j′ for j ′ > j is in β j \ β j−1 . If the latter, then β j′ \ β j′−1 contains a degree-1 vertex and β j \ β j − 1 can be charged to the edge which fixes that degree 1 vertex. Every additional edge was charged at most twice. Thus, R 6 2(E − V) It is not hard to check that α can be uniquely decoded from the encoding previously described. The final result of this encoding scheme is that each α can be encoded with at most V log n + 6(E − V) log V bits, and so there are at most n V · V 6(E−V) choices for α. The contribution of such α to Í α∈U c(α) is thus at most n −(E−V) V 6(E−V) (1 − δ/2)E k 2(E−V+C) We know that C 6 E − V. So as long as k, V 6 n 0.01 , we obtain that this contributes at most n (E−V)/2 (1 − δ/2)E . Summing across all V, E 6 n 0.01 , the lemma follows.  n 2 , Lemma 6.5. For α ⊆   |α| let V(α) be the number of vertices in α, let C(α) be the number of connected def components in α. For constants 1 > δ > 0 and k, let c(α)  1 + O( nd ) 2 Let ℓ 6 n ξ/k for some ξ > 0 (allowing ξ 6 o(1)) and    ·n −|α| ·(1−δ)|α| ·k 2(|α|−V(α)+C(α))  n U α⊆ : α is a union of cycles . 2 Õ Then c(α) 6 n 2ξ α∈U Proof. Let U t be the set of α which are unions of t-cycles (we exclude the empty α). Let c t  Then Õ Ö c(α) 6 (1 + c t ) . α∈U Í α∈U t cα. t 6ℓ Count the α ∈ U t which contain exactly p cycles of length t by first choosing a list of pt vertices— there are n pt choices. In doing so we will count each alpha p!t p times, since each of the p cycles can be rotated and the cycles can themselves be exchanged. All in all, there are at most n pt /(p!t p ) such α, and they contribute at most (1 − δ/2)pt k 2p c(α)n pt 6 6 k 2p /(p!t p ) . p!t p p!t p for large enough n. Thus, summing over all α ∈ U t , we get (1 + c t ) 6 So, Ö t 6ℓ (1 + c t ) 6 exp(k 2 ℓ Õ (1 − δ/2)p k 2p p0 ℓ Õ t1 p!t p 6 exp(k 2 /t) . 2 1/t) 6 exp(k 2 log 2ℓ) 6 (2ℓ)k 6 n 2ξ .  57 6.2 Lower bound for estimating communities Theorem 6.6. Let d, ε, k, δ be constants such that ε 2 d < (1− δ)k 2 . Let f : {0, 1} n×n → ’ be any function, let i, j ∈ [n] be distinct. Then if f satisfies d f (G)  0 and is correlated with the indicator 1 σ i σ j G∼G(n, n ) that i and j are in the same community in the following sense: G∼SBM(n,d,ε,k) ( f (G)(1σi σ j − 1k ) f (G)2 )1/2 d G∼G(n, n ) > Ω(1) then deg f > n c(d,ε,k) for some c(d, ε, k) > 0. Proof. Let 1(G)  µ(G) [1σi σ j − 1k | G], where µ(G) is the relative density of SBM(n, d, ε, k). Standard Fourier analysis shows that the optimal degree-ℓ choice for such f to maximize the above correlation is 1 6 ℓ , the orthogonal projection of 1 to the degree-ℓ polynomials with respect to the measure G(n, nd ), and the correlation is at most k1 6 ℓ k. It suffices to show that for some constant c(d, ε, k), if ℓ < n c(d,ε,k) then k1 6 ℓ k 6 o(1). For this we expand 1 in the Fourier basis, noting that where as usual σ̃i  σi − tions show that 1 k b 1 (α)  σ,G∼SBM(n,d,ε,k) h σ̃i , σ̃ j iχα (G) · 1 is the centered indicator of i’s community. By-now routine computa-  b 1 (α)2 6 (1 + O(d/n))ε 2 nd  |α| Ö © ª h σ̃i , σ̃ j i ® · ­ h σ̃i , σ̃ j i · (k,ℓ)∈α « ¬ 2 We assume that (i, j) < α; it is not hard to check that such α’s dominate the norm k1 6 ℓ k. If some vertex aside from i, j in α has degree 1 then this is zero. Similarly, if i or j does not appear in α then this is zero. Otherwise, b 1 (α)2 6 ((1 + O(d/n)))|α| n −|α| (1 − δ)|α| k 2(|α|−V(α)+C(α) where as usual V(α) is the number of vertices in α and C(α) is the number of connected components in α. Let β(α) be the connected component of α containing i and j (if they are not in the same component the arguments are mostly unchanged). Then we can bound k1 6 ℓ k 2  Õ |α| 6 ℓ b 1 (α)2 6 kµ 6 ℓ k 2 · Õ β ((1 + O(d/n)))|β| n −|β| (1 − δ)|β| k 2(|β|−V(β)+1 where β ranges over connected graphs with vertices from [n], at most ℓ edges, every vertex except i and j having degree at least 2, and containing i and j with degree at least 1. There are at most n V−2 V O(E−V) such graphs containing at V vertices aside from i and j and E edges (by an analogous argument as in Lemma 6.4). The total contribution from such β is therefore at most k 2(E−V+1) V O(E−V) n E−V+2 58 Summing over V and E, we get Õ β ((1 + O(d/n)))|β| n −|β| (1 − δ)|β| k 2(|β|−V(β)+1 6 n −Ω(1) so long as ℓ 6 n c for small enough c. Using Theorem 6.1 to bound kµ 6 ℓ k finishes the proof.  7 Tensor decomposition from constant correlation Problem 7.1 (Orthogonal n-dimensional 4-tensor decomposition from constant correlation). Let Í ⊗4 n ⊗4 satisfy hA,Bi > δ  Ω(1). a1 , . . . , a m ∈ ’n be orthonormal, and let A  m i1 a i . Let B ∈ (’ ) kAkkBk Let O be an oracle such that for any unit v ∈ ’n , O(v)  ( YES if Ím i1 ha i , vi 4 > δ O(1) NO otherwise Input: The tensor B, and if δ < 0.01, access to the oracle O. Goal: Output orthonormal vectors b 1 , . . . , b m so that there is a set S ⊆ [m] of size |S| > δ O(1) · m where for every i ∈ S there is j 6 m with hb j , a i i 2 > δ O(1) . O(1) O(1) We will give an n 1/δ -time algorithm (hence using at most n 1/δ oracle calls) for this problem based on a maximum-entropy Sum-of-Squares relaxation. The main theorem is the following; the subsequent corollary arrives at the final algorithm. Theorem 7.2. Let A, B and a1 , . . . , a m and δ 6 0.01 be as in Problem 7.1. Let v1 , . . . , v r for r 6 δ 4 m be orthonormal vectors. There is a randomized algorithm ALG with running time n O(1) which takes input B, v1 , . . . , v r and outputs a unit vector v, orthogonal to v1 , . . . , v r , with the following guarantee. There is a set S ⊆ [m] of size |S| > δ O(1) · m so that for i ∈ S,   hv, a i i 2 > δ O(1) > n −1/poly(δ) . The following corollary captures the overall algorithm for tensor decomposition, using the oracle O to filter the output of the algorithm of Theorem 7.2. Corollary 7.3. Let a1 , . . . , a n , A, B, δ be as in Theorem 7.2 and O as in Problem 7.1. There is a n poly(1/δ) time algorithm which takes the tensor B as input and returns b 1 , . . . , b m such that with high probability there is a set S ⊆ [m] of size |S| > δ O(1) m which has the guarantee that for all i ∈ S there is j 6 m with ha i , b j i 2 > δ O(1) . If δ 6 1 − Ω(1), the algorithm makes n 1/poly(δ) adaptive queries to the oracle O. The algorithm can also be implemented with nonadaptive queries as follows. Once the input B and the random coins of the algorithm are fixed, there is a list of at most n poly(k/δ) . Query the oracle O nonadaptively on all these vectors and assemble the answers into a lookup table; then the decomposition algorithm can be run using access only to the lookup table. 59 Proof of Corollary 7.3. If δ > 1 − ε ∗ for a small enough constant ε ∗ then the tensor decomposition algorithm of Schramm and Steurer has the appropriate guarantees. (See Theorem 4.4 and Lemma 4.9 in [SS17]. This algorithm has several advantages, including that it does not need to solve any semidefinite program, but it cannot handle the high-error regime we need to address here.) From here on we assume δ 6 0.01 < 1 − ε ∗ . (Otherwise, we can replace δ with δ C 6 0.01 for large enough C.) Our algorithm is as follows. Algorithm 7.4 (Constant-correlation tensor decomposition). 1. Let V be an empty set of vectors. 2. For rounds 1, . . . , T  δ O(1) m, do: (a) Use the algorithm of Theorem 7.2 on the tensor B to generate w 1 , . . . , w t , where t  O(1) n 1/δ . (b) Call O on successive vectors w 1 , . . . , w t , and let w be the first for which it outputs YES. (If no such vector exists, the algorithm halts and outputs random orthonormal vectors b 1 , . . . , b m .) (c) Add w to V. 3. Let b 1 , . . . , b m−|V | be random orthonormal vectors, orthogonal to each v ∈ V. 4. Output {b 1 , . . . , b m−|V | } ∪ V. O(1) large enough, and T  δ O(1) m small enough, by Theorem 7.2 with high Choosing t  n 1/δ probability in every round 1, . . . , T there is some w among w 1 , . . . , w t for which O outputs YES. Suppose that occurs. In this case, the algorithm outputs (along with some random vectors b i ) a set of vectors V which are orthonormal, and each v ∈ V satisfies hv, a i i > δ O(1) for some a i ; say that this a i is covered by v. Each a i can be covered at most 1/δ O(1) times, by orthonormality of the set V. So, at least δ O(1) |V |  δ O(1) m vectors are covered at least once, which proves the corollary.  We turn to the proof of Theorem 7.2. We will use the following lemmas, whose proofs are later in this section. The problem is already interesting when the list v1 , . . . , v r is empty, and we encourange the reader to understand this case first. The first lemma says that a pseudodistribution of high entropy (in the 2-norm sense33) which is correlated with the tensor B must also be nontrivially correlated with A. Lemma 7.5. Let A, B be as in Problem 7.1. Let v1 , . . . , v r ∈ ’n be orthonormal, with r 6 δ 4 m. Suppose ˜ is the degree-4 pseudodistribution solving mink ˜ x ⊗4 kF s.t. ˜ satisfies {kx k 2 6 1, hx, v1i  0, . . . , hx, v r i  0} δ h ˜ x ⊗4 , Bi > 2m ˜ xx ⊤ 6 1 m ˜ xx ⊤ ⊗ xx ⊤ 6 1 m (7.1) (7.2) (7.3) 33For a distribution µ finitely-supported on a family of orthonormal vectors, the Frobenious norm k x∼µ x ⊗k k is closely related to the collision probability of µ, itself closely related to the order-2 case of Rényi entropy. 60 Í Then ˜ i 6 m hx, a i i 4 > δ 2 /8. Furthermore, it is possible to find ˜ in polynomial time.34 The second lemma says that given a high-entropy (in the spectral sense of [MSS16]) pseudodistribution ˜ having nontrivial correlation with some a ∈ ’n , contracting ˜ with a yields a matrix whose quadratic form is large at a and which does not have too many large eigenvalues. Lemma 7.6. Let a1 , . . . , a m ∈ ’n be orthonormal. Let ˜ be a degree-4 pseudoexpectation such that 1. ˜ satisfies {kx k 2 6 1} Í 2. ˜ i 6 m hx, a i i 4 > δ. 3. k ˜ xx ⊤ ]kop , k ˜ xx ⊤ ⊗ xx ⊤ kop 6 1 m .35 Let M i ∈ ’n×n be the matrix ˜ hx, a i i 2 xx ⊤ . For every i ∈ [m], the matrix M i has at most 4/δ eigenvalues δ larger than 4m . Furthermore,  δ > 2δ .  ha i , M i a i i > 2m i∼[m] The last lemma will help show that a random contraction of a high-entropy pseudodistribution behaves like one of the contractions from Lemma 7.6, with at least inverse-polynomial probability. Lemma 7.7. Let 1 ∼ N(0, Σ) for some 0  Σ  Id and let ˜ be a degree-4 pseudoexpectation where • ˜ satisfies {kx k 2 6 1}. • ˜ xx ⊤ 6 c. • ˜ xx ⊤ ⊗ xx ⊤ 6 c Then 1 ˜ h1, xi 2 xx ⊤ 6 O(c · log n) . Now we can prove Theorem 7.2. Proof of Theorem 7.2. The algorithm is as follows: Algorithm 7.8 (Low-correlation tensor decomposition). 1. Use the first part of Lemma 7.5 to obÍ 4 2 2 ˜ tain a degree-4 pseudoexpectation with i∈[m] ha i , xi > δ /4 satisfying {kx k 6 1, hx, v 1 i  0, . . . , hx, v r i  0}. 2. Sample a random 1 ∼ N(0, Id) and compute the contraction M  ˜ h1, xi 2 xx ⊤ . eigenvectors of M. 3. Output a random unit vector b in the span of the top 32 δ2 34Up to inverse-polynomial error, which we ignore here. See [MSS16] for the ideas needed to show polynomial-time solvability. 35Recall that k·k denotes the operator norm, or maximum singular value, of a matrix. 61 First note that for any v ∈ Span{v1 , . . . , v r }, we must have hv, Mvi  ˜ h1, xi 2 hv, xi 2  0, so v lies in the kernel of M. Hence, the ouput of the algorithm will always be orthogonal to v1 , . . . , v r . Let Π32/δ2 be the projector to the top 32/δ2 eigenvectors of M. For any unit vector a with kΠ32/δ2 a k > δ O(1) , the algorithm will output b with nontrivial correlation with a. Formally, for any such a, hb, ai 2 > δ O(1) . b So, our goal is to show that for a δ O(1) -fraction of the vectors a1 , . . . , a m , {kΠ32/δ2 a i k > δ O(1) } > n −1/δ O(1) 1 . For i ∈ [m], let M i  ˜ ha i , xi 2 xx ⊤ . Let i be the index of some a i so that δ2 16m ha i , M i a i i > δ2 > 32m and rank M i 6 32 δ2 as in Lemma 7.6. (There are Ω(δ2 m) possible choices for a i , according to the Lemma.) We expand the Gaussian vector 1 from the algorithm as 1  10 · a i + 1 ′ where 10 ∼ N(0, 1) and h1 ′ , a i i  0. We note for later use that 1 ′ is a Gaussian vector independent of 10 and that (1 ′)(1 ′)⊤  Id. Using this expansion, M  102 ˜ ha i , xi 2 xx ⊤ + 2 · 10 ˜ h1 ′ , xiha i , xixx ⊤ + ˜ h1 ′ , xi 2 xx ⊤ . We will show that all but the first term have small spectral norm. Addressing the middle term first, by Cauchy-Schwarz, for any unit v ∈ ’n , ˜ h1 ′ , xiha i , xihv, xi 2 6 ˜ h1 ′ , xi 2 hx, vi 2  1/2 ˜ ha i , xi 2 hv, xi 2 where in the last step we have used that ˜ xx ⊤ ⊗ xx ⊤ 6 By Markov’s inequality and Lemma 7.7, ′ 1 n ˜ h1 ′ , xi 2 xx ⊤ > t log n m Let t be a large enough constant so that ′ 1 n ˜ h1 ′ , xi 2 xx ⊤ 6 o t log n m 1 m.  1/2 6O o 1 t  6 ˜ h1 ′, xi 2 xx ⊤ 1/2 ·  1 1/2 m , . > 0.9 . For any constant c, with probability n −1/poly(δ) , the foregoing occurs and 10 (which is independent of 1 ′) is large enough that 2 102 · cδm > δ14 M − 102 M i . Choosing c large enough, in this case def 1 M 102 M′   M i + O(δ6 /m) . 62 Hence the vector a i satisfies 1 ha i , 102 Ma i i > δ2 33m This means that the projection b of a i into the span of eigenvectors of M ′ with eigenvalue at least δ2 /60m has kb k 2 > δ O(1) . This finishes the proof.  7.1 Proofs of Lemmas These lemmas and their proofs use many ideas from [MSS16]. The main difference here is that we want to contract the tensor ˜ x ⊗4 in 2 modes, to obtain the matrix ˜ h1, xi 2 xx ⊤ . For us this is useful because ˜ h1, xi 2 xx ⊤  0. By contrast, the tools in [MSS16] would only allow us to analyze the contraction ˜ hh, x ⊗ xixx ⊤ for h ∼ N(0, Idn 2 ). We start with an elementary fact. Fact 7.9. Let a1 , . . . , a m ∈ ’n be orthonormal. Let Π be the projector to a subspace of codimension at most √ Ím ⊗4 Ím ⊗4 δm. Let A  i1 a i and ΠA  i1 (Πa i ) . Then hA, ΠAi > (1 − O( δ))kAk · kΠAk. A useful corollary of Fact 7.9 is that if T is any 4-tensor satisfying hT, ΠAi > δkT k kΠAk and Π has codimension ≪ δ 2 m, then hT, Ai > Ω(δ)kT k kAk. Proof of Fact 7.9. We expand hA, ΠAi  Õ i, j 6 m ha i , Πa j i 4 > Õ i, j 6 m kΠa i k 8 Writing Π in the a i basis, we think of kΠa i k 4  Π2ii , the square of the i-th diagonal entry of Π. Since Π has codimension at most δm, rank Π  Tr Π  Õ i6n Πii > n − δm . √ Furthermore, for each i, it must be√that 0 6 Πii 6 1. By Markov’s inequality, at√most δm diagonal Í entries of Π can be less than 1 − δ in magnitude. Hence, i 6 m Π4ii > (1 − 4 δ)m. On the other hand, kAk 2  m; this proves the fact.  Now we can prove Lemma 7.5. Proof of Lemma 7.5. We will appeal to Theorem 2.3. Let C be the convex set of all pseudo-moments ˜ x ⊗4 such that ˜ is a deg-4 pseudo-distribution that satisfies the polynomial constraints {kx k 2 6 1, hx, v i i  0} and the operator norm conditions ˜ xx ⊤ 6 1 m, ˜ xx ⊤ ⊗ xx ⊤ 6 1 m . Let Π be the projector to the orthogonal space of v1 , . . . , v r . Notice that m1 ΠA ∈ C. Furthermore, hB, ΠAi > δ/2 by Fact 7.9, the assumption that r 6 δ4 m, and the assumption δ 6 0.01. By Theorem 2.3, and Fact 7.9 again, the optimizer of the convex program in the Lemma satisfies δ2 ) and the result follows.  h ˜ x ⊗4 , m1 Ai > 8m 63 Proof of Lemma 7.6. By the assumption k ˜ xx ⊤ ⊗xx ⊤ k 6 m1 , for every a i it must be that ˜ hx, a i i 4 6 m1 . Í δ 4 4 ˜ Since ˜ m i1 hx, a i i > δ, at least δm/2 of the a i ’s must satisfy hx, a i i > 2m . Rewritten, for any δ . such a i we obtain ha i , M i a i i > 2m For any M i , Tr M i  ˜ hx, a i i 2 kx k 2  ˜ hx, a i i 2 6 m1 because k ˜ xx ⊤ k 6 δ 4m . 1 m. Also, M i  0. Hence, M i can have no more than 4 δ eigenvalues larger than  Now we turn to the proof of Lemma 7.7. We will need spectral norm bounds on certain random matrices associated to the random contraction ˜ h1, xixx ⊤. The following are closely related to Theorem 6.5 and Corollary 6.6 in [MSS16]. Lemma 7.10. Let 1 ∼ N(0, Id) and let ˜ be a degree-4 pseudoexpectation where • ˜ satisfies {kx k 2  1}. • ˜ xx ⊤ 6 c. • ˜ xx ⊤ ⊗ xx ⊤ 6 c Then ˜ h1, xi 2 xx ⊤ 6 O(c · log n) . 1 Before proving the lemma, we will need a classical decoupling inequality. Fact 7.11 (Special case of Theorem 1 in [dlPnMS94]). Let 1, h ∼ N(0, Idn ) be independent. Let M i j for i, j ∈ [n] be a family of matrices. There is a universal constant C so that Õ 1 i, j 1i 1 j · Mi j 6 C · 1,h Õ i, j 1i h j · Mi j . We will also need a theorem from [MSS16]. Fact 7.12 (Corollary 6.6 in [MSS16]). Let T ∈ ’p ⊗ ’q ⊗ ’r be an order-3 tensor. Let 1 ∼ N(0, Σ) for some 0  Σ  Idr . Then for any t > 0,  1 n (Id ⊗ Id ⊗ 1)⊤ T {1},{2} and consequently, 1 h (Id ⊗ Id ⊗ 1)⊤ T o  2 > t · max kT k{1},{2,3} , kT k{2},{1,3} 6 2(p + q) · e −t /2 , {1},{2} i  6 O(log(p + q))1/2 · max kT k{1},{2,3} , kT k{2},{1,3} 64 Proof of Lemma 7.10. We expand the matrix ˜ h1, xi 2 xx ⊤ as ˜ h1, xi 2 xx ⊤  Õ i , j∈[n] i∈[n] Addressing the first term, by standard concentration, 1 Õ 1i2 ˜ x 2 xx ⊤ 6 i i∈[n] 1  max 1i2 i∈[n] Õ 1i2 ˜ x 2i xx ⊤ + · ˜ kx k 2 xx ⊤  1i 1 j · ˜ x i x j xx ⊤ . max i∈[n] 1i2  O(log n). So,  O(log n) · ˜ xx ⊤  O(c · log n) . The second term we will decouple using Fact 7.11. 1 Õ i, j 1i 1 j · ˜ x i x j xx ⊤ 6 O(1) · 1,h Õ i, j 1i h j · ˜ x i x j xx ⊤ . We add some aditional terms to the sum; by similar reasoning to our bound on the first term they do not contribute too much to the norm. 1,h Õ i, j 1i h j · ˜ x i x j xx ⊤ 6 O(1) · Õ 1,h Õ i, j∈[n] 1i h j · ˜ x i x j xx ⊤ + O(c · log n) . We can rewrite the matrix in the first term on the right-hand side as i, j∈[n] 1i h j · ˜ x i x j xx ⊤  ˜ h1, xihh, xixx ⊤ . Now we can apply Fact 7.12 twice in a row; first to 1 and then to h, which together with our norm bound on xx ⊤ ⊗ xx ⊤ , gives 1,h ˜ h1, xihh, xixx ⊤ 6 O(c · log n) . Putting all of the above together, we get the lemma.  Next we prove Lemma 7.7 as a corollary of Lemma 7.7 which applies to random contractions which are non-spherical. The proof technique is very similar to that for Fact 7.12. Proof of Lemma 7.7. Let h ∼ N(0, Id − Σ) be independent of 1, and define 1 ′  1 + h and 1 ′′  1 − h, so that 1  21 (1 ′ + 1 ′′). It is sufficient to bound 1,h ˜ h1 ′ + 1 ′′ , xi 2 xx ⊤ . Expanding and applying triangle inequality, 1,h ˜ h1 ′ + 1 ′′ , xi 2 xx ⊤ 6 1,h ˜ h1 ′ , xi 2 xx ⊤ + 2 1,h ˜ h1 ′ , xih1 ′′, xixx ⊤ + 1,h ˜ h1 ′′ , xi 2 xx ⊤ . The first and last terms are O(c · log n) by Lemma 7.10. For the middle term, consider the quadratic form of the matrix ˜ h1 ′ , xih1 ′′, xixx ⊤ on a vector v ∈ ’n : ˜ h1 ′ , xih1 ′′, xihx, vi 2 6 ˜ h1 ′ , xi 2 hx, vi 2 + ˜ h1 ′′ , xi 2 hx, vi 2 by pseudoexpectation Cauchy-Schwarz. Thus for every 1 ′ , 1 ′′, ˜ h1 ′ , xih1 ′′, xixx ⊤ 6 ˜ h1 ′ , xi 2 xx ⊤ + ˜ h1 ′′ , xi 2 xx ⊤ . Together with Lemma 7.10 this concludes the proof. 65  7.2 Lifting 3-tensors to 4-tensors Problem 7.13 (3-to-4 lifting). Let a1 , . . . , a m ∈ ’n be orthonormal. Let A3  Ím ⊗4 n×n×n satisfy hB, A i > δ · kA k · kB k. 3 3 i1 a i . Let B ∈ ’ Input: The tensor B. Goal: Output B ′ satisfying hB ′ , A4 i > δ O(1) · kA4 k · kB ′ k. Ím i1 a i⊗3 and A4  Theorem 7.14. There is a polynomial time algorithm, using the sum of squares method, which solves the 3-to-4 lifting problem. Proof. Small δ regime: δ < 1 − Ω(1): The algorithm is to output the fourth moments of the optimizer of the following convex program. min ˜ s.t. k ˜ x ⊗3 k ˜ is degree-4 ˜ satisfies {kx k 2  1} δkB k h ˜ x ⊗3 , Bi > √ m 1 k ˜ x ⊗4 k 6 √ . m To analyze the algorithm we apply Theorem 2.3. Let C be the set of degree-4 pseudodistributions √ satisfying {kx k 2  1} and having k ˜ x ⊗4 k 6 1/ m. The uniform distribution over a1 , . . . , a m , whose third and fourth moments are m1 A3 and m1 A4 , respectively, is in C. Let ˜ be the pseudoexpectation solving the convex program. By Theorem 2.3, h ˜ x ⊗3 , 1 m A3 i > δ 1 δ2 · √ · k ˜ x ⊗3 k > 2 2m m At the same time, m 1 Õ˜ 1 h ˜ x ⊗3 , m1 A3 i  hx, a i i 3 6 m m i1 ˜ m Õ i1 hx, a i i 4 ! 1/2 by Cauchy-Schwarz. Putting these together, we obtain h ˜ x ⊗4 , A4 i  ˜ m Õ i1 hx, a i i 4 > δ 4 /4 . √ Finally, kA4 k · k ˜ x ⊗4 k 6 1 (since we constrained k ˜ x ⊗4 k 6 1/ m), which finishes the proof. Large δ regime: δ > 1 − o(1): Modify the convex program from the small-δ regime to project √ (B/kB k) · 1/ m to same convex set C. The normalization is so that √ (B/kB k) · 1/ m  m1 · A3 . The analysis is similar.  66 Acknowledgments We are indebted to Avi Wigderson who suggested color coding as a technique to evaluate the kinds of polynomials we study in this work. We thank Moses Charikar for pointing out the relationship between our SOS program for low correlation tensor decomposition and Rényi entropy. We thank Christian Borgs, Jennifer Chayes, and Yash Deshpande for helpful conversations, especially relating to Section 4. We thank anonymous reviewers for many suggested improvements to this paper. References [Abb17] Emmanuel Abbe, Community detection and stochastic block models: recent developments, CoRR abs/1703.10146 (2017). 9 [ABFX08] Edoardo M. Airoldi, David M. Blei, Stephen E. Fienberg, and Eric P. Xing, Mixed membership stochastic blockmodels, NIPS, Curran Associates, Inc., 2008, pp. 33–40. 3, 9 [AGH+ 14] Animashree Anandkumar, Rong Ge, Daniel J. Hsu, Sham M. Kakade, and Matus Telgarsky, Tensor decompositions for learning latent variable models, Journal of Machine Learning Research 15 (2014), no. 1, 2773–2832. 1, 14 [AGHK13] Animashree Anandkumar, Rong Ge, Daniel J. Hsu, and Sham Kakade, A tensor spectral approach to learning mixed membership community models, COLT 2013 - The 26th Annual Conference on Learning Theory, June 12-14, 2013, Princeton University, NJ, USA (Shai Shalev-Shwartz and Ingo Steinwart, eds.), JMLR Workshop and Conference Proceedings, vol. 30, JMLR.org, 2013, pp. 867–881. 13, 21 [AGHK14] Animashree Anandkumar, Rong Ge, Daniel J. Hsu, and Sham M. Kakade, A tensor approach to learning mixed membership community models, Journal of Machine Learning Research 15 (2014), no. 1, 2239–2312. 1, 9 [AN04] Noga Alon and Assaf Naor, Approximating the cut-norm via grothendieck’s inequality, STOC, ACM, 2004, pp. 72–80. 29 [AS15] Emmanuel Abbe and Colin Sandon, Community detection in general stochastic block models: Fundamental limits and efficient algorithms for recovery, FOCS, IEEE Computer Society, 2015, pp. 670–688. 4, 17 [AS16a] , Achieving the KS threshold in the general stochastic block model with linearized acyclic belief propagation, NIPS, 2016, pp. 1334–1342. 1, 3, 5, 6, 12, 31, 32 [AS16b] , Crossing the KS threshold in the stochastic block model with information theory, ISIT, IEEE, 2016, pp. 840–844. 15 [AYZ95] Noga Alon, Raphael Yuster, and Uri Zwick, Color-coding, J. ACM 42 (1995), no. 4, 844–856. 4, 23 67 [BAP+ 05] Jinho Baik, Gérard Ben Arous, Sandrine Péché, et al., Phase transition of the largest eigenvalue for nonnull complex sample covariance matrices, The Annals of Probability 33 (2005), no. 5, 1643–1697. 25 [BHK+ 16] Boaz Barak, Samuel B. Hopkins, Jonathan A. Kelner, Pravesh Kothari, Ankur Moitra, and Aaron Potechin, A nearly tight sum-of-squares lower bound for the planted clique problem, FOCS, IEEE Computer Society, 2016, pp. 428–437. 2, 9, 18, 19 [BKM17] Jess Banks, Robert Kleinberg, and Cristopher Moore, The lovász theta function for random regular graphs and community detection in the hard regime, APPROX-RANDOM, LIPIcs, vol. 81, Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2017, pp. 28:1–28:22. 16 [BKS15] Boaz Barak, Jonathan A. Kelner, and David Steurer, Dictionary learning and tensor decomposition via the sum-of-squares method, STOC, ACM, 2015, pp. 143–151. 1, 8, 14, 23 [BLM15] Charles Bordenave, Marc Lelarge, and Laurent Massoulié, Non-backtracking spectrum of random graphs: Community detection and non-regular ramanujan graphs, FOCS, IEEE Computer Society, 2015, pp. 1347–1357. 31 [BMNN16] Jess Banks, Cristopher Moore, Joe Neeman, and Praneeth Netrapalli, Informationtheoretic thresholds for community detection in sparse networks, COLT, JMLR Workshop and Conference Proceedings, vol. 49, JMLR.org, 2016, pp. 383–416. 18 [CJ10] Pierre Comon and Christian Jutten, Handbook of blind source separation: Independent component analysis and applications, Academic press, 2010. 14 [DKMZ11] Aurelien Decelle, Florent Krzakala, Cristopher Moore, and Lenka Zdeborová, Asymptotic analysis of the stochastic block model for modular networks and its algorithmic applications, CoRR abs/1109.3041 (2011). 3, 10 [dlPnMS94] Victor H. de la Peña and S. J. Montgomery-Smith, Bounds on the tail probability of Ustatistics and quadratic forms, Bull. Amer. Math. Soc. (N.S.) 31 (1994), no. 2, 223–227. MR 1261237 64 [DLR77] Arthur P Dempster, Nan M Laird, and Donald B Rubin, Maximum likelihood from incomplete data via the em algorithm, Journal of the royal statistical society. Series B (methodological) (1977), 1–38. 3 [DM15] Yash Deshpande and Andrea Montanari, Improved sum-of-squares lower bounds for hidden clique and hidden submatrix problems, COLT, JMLR Workshop and Conference Proceedings, vol. 40, JMLR.org, 2015, pp. 523–562. 72 [FGR+ 13] Vitaly Feldman, Elena Grigorescu, Lev Reyzin, Santosh Vempala, and Ying Xiao, Statistical algorithms and a lower bound for detecting planted cliques, STOC, ACM, 2013, pp. 655–664. 16 68 [FGV15] Vitaly Feldman, Cristobal Guzman, and Santosh Vempala, Statistical query algorithms for stochastic convex optimization, CoRR abs/1512.09170 (2015). 16 [FPV15] Vitaly Feldman, Will Perkins, and Santosh Vempala, On the complexity of random satisfiability problems with planted solutions, STOC, ACM, 2015, pp. 77–86. 16 [Gal62] Robert Gallager, Low-density parity-check codes, IRE Transactions on information theory 8 (1962), no. 1, 21–28. 3 [GHK] Rong Ge, Qingqing Huang, and Sham Kakade, Learning mixtures of gaussians in high dimensions, stoc 2015, 761–770. 14 [GM15] Rong Ge and Tengyu Ma, Decomposing overcomplete 3rd order tensors using sum-ofsquares algorithms, APPROX-RANDOM, LIPIcs, vol. 40, Schloss Dagstuhl - LeibnizZentrum fuer Informatik, 2015, pp. 829–849. 8 [GVX14] Navin Goyal, Santosh Vempala, and Ying Xiao, Fourier PCA and robust tensor decomposition, STOC, ACM, 2014, pp. 584–593. 14 [Har70] Richard A Harshman, Foundations of the parafac procedure: Models and conditions for an" explanatory" multi-modal factor analysis. 21 [HKP+ 17] Samuel B Hopkins, Pravesh Kothari, Aaron Potechin, Prasad Raghavendra, Tselil Schramm, and David Steurer, The power of sum-of-squares for detecting hidden structures, Symposium on Foundations of Computer Science (2017). 9, 18, 19 [HSS15] Samuel B. Hopkins, Jonathan Shi, and David Steurer, Tensor principal component analysis via sum-of-square proofs, COLT, JMLR Workshop and Conference Proceedings, vol. 40, JMLR.org, 2015, pp. 956–1006. 2, 8 [Jer92] Mark Jerrum, Large cliques elude the metropolis process, Random Struct. Algorithms 3 (1992), no. 4, 347–360. 15 [Kuc95] Ludek Kucera, Expected complexity of graph partitioning problems, Discrete Applied Mathematics 57 (1995), no. 2-3, 193–212. 15 [LBB+16] Thibault Lesieur, Caterina De Bacco, Jess Banks, Florent Krzakala, Cristopher Moore, and Lenka Zdeborová, Phase transitions and optimal algorithms in high-dimensional gaussian mixture clustering, Allerton, IEEE, 2016, pp. 601–608. 19 [LRA93] S. E. Leurgans, R. T. Ross, and R. B. Abel, A decomposition for three-way arrays, SIAM J. Matrix Anal. Appl. 14 (1993), no. 4, 1064–1083. MR 1238921 21 [Mas13] Laurent Massoulié, Community detection thresholds and the weak ramanujan property, CoRR abs/1311.3085 (2013). 26, 27 [Mas14] , Community detection thresholds and the weak ramanujan property, STOC, ACM, 2014, pp. 694–703. 1, 3, 6 69 [MM09] Marc Mezard and Andrea Montanari, Information, physics, and computation, Oxford University Press, 2009. 19 [MNS12] Elchanan Mossel, Joe Neeman, and Allan Sly, Stochastic block models and reconstruction, arXiv preprint arXiv:1202.1499 (2012). 18 [MNS13] Elchanan Mossel, Joe Neeman, and Allan Sly, A proof of the block model threshold conjecture, CoRR abs/1311.4115 (2013). 26, 27 [MNS15a] , Consistency thresholds for the planted bisection model, STOC, ACM, 2015, pp. 69– 75. 1, 3, 5, 6, 19 [MNS15b] Elchanan Mossel, Joe Neeman, and Allan Sly, Reconstruction and estimation in the planted partition model, Probab. Theory Related Fields 162 (2015), no. 3-4, 431–461. MR 3383334 26 [Moo17] Cristopher Moore, The computer science and physics of community detection: Landscapes, phase transitions, and hardness, CoRR abs/1702.00467 (2017). 12, 15 [MS16] Andrea Montanari and Subhabrata Sen, Semidefinite programs on sparse random graphs and their application to community detection, STOC, ACM, 2016, pp. 814–827. 16 [MSS16] Tengyu Ma, Jonathan Shi, and David Steurer, Polynomial-time tensor decompositions with sum-of-squares, FOCS, IEEE Computer Society, 2016, pp. 438–446. 14, 15, 23, 61, 63, 64 [MW15] Tengyu Ma and Avi Wigderson, Sum-of-squares lower bounds for sparse PCA, NIPS, 2015, pp. 1612–1620. 2 [Pea82] Judea Pearl, Reverend bayes on inference engines: A distributed hierarchical approach, Cognitive Systems Laboratory, School of Engineering and Applied Science, University of California, Los Angeles, 1982. 3 [RRS16] Prasad Raghavendra, Satish Rao, and Tselil Schramm, Strongly refuting random csps below the spectral threshold, CoRR abs/1605.00058 (2016). 8 [SS17] Tselil Schramm and David Steurer, Fast and robust tensor decomposition with applications to dictionary learning, COLT, Proceedings of Machine Learning Research, vol. 65, PMLR, 2017, pp. 1760–1793. 14, 15, 60 [Wik17a] Wikipedia, Bayes estimator — Wikipedia, the free encyclopedia, http://en.wikipedia.org/w/index.php?title=Bayes%20estimator&oldid=754605088, 2017, [Online; accessed 30-March-2017]. 1 [Wik17b] , Dirichlet distribution — Wikipedia, the free encyclopedia, http://en.wikipedia.org/w/index.php?title=Dirichlet%20distribution&oldid=762020989, 2017, [Online; accessed 30-March-2017]. 10 70 A Toolkit and Omitted Proofs A.1 Probability and linear algebra tools Fact A.1. Consider any inner product h·, ·i on ’n with associated norm k · k. Let X and Y be joinlydistributed ’n -valued random variables. Suppose that kX k 2 6 C kX k 2 with probability 1, and that ( Then hX, Yi > δ. kY k 2 )1/2 kX k 2 )1/2 (   hX, Yi δ2 δ . > >  kX k · kY k 2 4C 2 Proof of Fact A.1. Let 1E be the 0/1 indicator of an event E. Note that   Hence, hX, Yi 1 hX, Yi 1 δ hX,Yi 6 2 ·kXk·kY k δ hX,Yi> 2 ·kXk·kY k At the same time,  hX, Yi 1 δ hX,Yi> 2 ·kXk·kY k    6 δ 2 kX k · kY k 6 2δ ( kX k 2 )1/2 ( kY k 2 )1/2 . > δ 2 kX k · kY k 6 2δ ( kX k 2)1/2 ( kY k 2 )1/2 . 2  2 1/2 6 kX k · kY k  kX k 2 · kY k 2  1/2 ·  1 δ hX,Yi> 2 ·kXk·kY k · {hX, Yi > δ 2  1/2 · kX k · kY k} 6 C( kX k 2)1/2 ( kY k 2)1/2 · {hX, Yi > δ 2  1/2 · kX k · kY k} Putting the inequalities together and rearranging finishes the proof.  1/2 .  Proof of Proposition 5.22. We decompose X i as X i  X i 1|X i | 6R +X i 1 |X i |>R . Let Yi  X i 1 |X i | 6R . Then | Yi |  | X i 1|X i |>R | 6 δ′ Xi − and – Yi 6 Yi2 6 1 Í So we can apply Bernstein’s inequality to  ( 1 m Õ ) i6m m Yi > t + δ i6m ′ X i2 . Yi to obtain that   −Ω(1) · m · t 2 . 6 exp X2 + t · R Now, with probability at least 1 − δ we know X i  Yi , so by a union bound,  ( 1 m Õ i6m Xi > t + δ ′ )   −Ω(1) · m · t 2 + mδ . 6 exp X2 + t · R 71  Fact A.2. Let {X1 , . . . , X n , Y1 , . . . , Ym } are jointly distributed real-valued random variables. Suppose there is S ⊆ [m] with |S| > (1 − o m (1)) · m such that for each i ∈ S there a degree-D polynomials p i satisfying p i (X)Yi > δ. ( Y 2 )1/2 ( p i (X)2 )1/2 Í Í Furthermore, suppose i∈S Yi2 > (1 − o(1)) i∈[m] Yi2 . Let Y ∈ ’m be the vector-valued random variable with i-th coordinate Yi , and similarly let P(X) have i-th coordinate p i (X). Then ( hP(X), Yi > (1 − o(1)) · δ · ( kP(X)k 2 )1/2 kY k 2 )1/2 Proof. The proof is by Cauchy-Schwarz. hP(X), Yi  Õ Õ i∈S >δ i∈S >δ p i (X)Y ( p i (X)2 )1/2 ( Yi2 )1/2 Õ i∈S p i (x)2 ! 1/2 A.2 Tools for symmetric and Dirichlet priors © Õ 2ª Yi ® · (1 − o(1)) ­ i∈[m] ¬ « 1/2 .  Proof of Fact 5.15. Let X be any ’k -valued random variable which is symmetric in distribution Í with respect to permutations of coordinates and satisfies s∈[k] X(s)  0 with probability 1. (The variable σ̃ is one example.) We prove the claim about hX, xihX, yihX, zihX, wi; the other proofs are similar. Consider the matrix M  (X ⊗ X)(X ⊗ X)⊤ . Since x, y, z, w are orthogonal to the all-1’s vector, we may add 1 ⊗ v, for any v ∈ ’n , to any row or column of M without affecting the statement to be proved. Adding multiples of 1 ⊗ e i to rows and columns appropriately makes M a block diagonal matrix, with the top block indexed by coordinates (i, i) for i ∈ [k] and the bottom block indexed by pairs (i, j) for i , j. The resulting top block takes the form cId + c ′ J, where J is the all-1’s matrix. The bottom block will be a matrix from the Johnson scheme. Standard results on eigenvectors of the Johnson scheme (see e.g. [DM15] and references therein) finish the proof. The values of constants C for the Dirichlet distribution follow from the next fact.  Fact A.3. Let σ ∈ ’k be distributed according to a (symmetric) Dirichlet distribution with parameter α. Î That is, (σ) ∝ j∈[k] σ α−1 . Let γ ∈ Žk be a k-tuple, and let σ γ  Î γj j6k σ j . Let |γ|  Γ(kα) · σγ  Γ(kα + |γ|) 72 Î j6k Í j6k γ j . Then Γ(α + γ j ) Γ(α)k . Furthermore, let σ̃ ∈ ’k be given by σ̃i  σi − 1k . Then   Γ(kα) Γ(α + 2) Γ(α + 1)2 1 σ̃ σ̃  ·Π − ·Π, 2 Γ(kα + 2) Γ(α) k(kα + 1) Γ(α) ⊤ where Π ∈ ’k×k is the projector to the subspace orthogonal to the all-1s vector. Proof. We recall the density of the k-dimensional Dirichlet distribution with parameter vector α 1 , . . . , α k . Here Γ denotes the usual Gamma function. Í Γ( {σ}  Î Í In particular, Γ( Î j6k j6k α j) Γ(α j ) α j ) Ö α j −1 σj . · j 6 k Γ(α j ) · j6k j6k ∫ Ö α j −1 σj dσ  1 j6k Í where the integral is taken with respect to Lebesgue measure on {σ : j 6 k σ j  1}. Using this fact we can compute the moments of the symmetric Dirichlet distribution with parameter α. We show for example how to compute second moments; the general formula can be proved along the same lines. For s , t ∈ [k], Γ(kα) · σs σt  Γ(α)k ∫ σs σt Ö σ α−1 j j6k Γ(kα + 2) Γ(α + 1)2 Γ(kα) · ·  · 2 Γ(kα + 2) Γ(α) Γ(α)k−2 Γ(α + 1)2  ∫ Γ(α + 1)2 Γ(kα) · . Γ(kα + 2) Γ(α)2 (α+1)−1 (α+1)−1 σt σs Ö σ α−1 j j ,s,t Similarly, σs2  The formula for Γ(α + 2) Γ(kα) · . Γ(kα + 2) Γ(α) σ̃ σ̃ ⊤ follows immediately.  73
8cs.DS
1 Canonical form of linear subspaces and coding invariants: the poset metric point of view arXiv:1706.09996v1 [cs.IT] 30 Jun 2017 Jerry Anderson Pinheiro and Marcelo Firer Abstract In this work we introduce the concept of a sub-space decomposition, subject to a partition of the coordinates. Considering metrics determined by partial orders in the set of coordinates, the so called poset metrics, we show the existence of maximal decompositions according to the metric. These decompositions turns to be an important tool to obtain the canonical form for codes over any poset metrics and to obtain bounds for important invariants such as the packing radius of a linear subspace. Furthermore, using maximal decompositions, we are able to reduce and optimize the full lookup table algorithm for the syndrome decoding process. Index Terms Canonical form, poset codes, poset metrics, syndrome decoding, packing radius. MSC2010: 15A21, 94B05, 52C17 C ONTENTS I Introduction 2 II Poset Metrics and Decomposition of Subspaces 4 II-A Posets and Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 II-B Partitions and Decompositions . . . . . . . . . . . . . . . . . . . . . . . . 6 II-C Construction of Maximal P-Decompositions . . . . . . . . . . . . . . . . . 13 III Comparison of Posets and Decompositions 18 III-A Decompositions and refinements . . . . . . . . . . . . . . . . . . . . . . . 19 III-B Hierarchical Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 The authors are at Institute of Mathematics, Statistics and Scientific Computing, University of Campinas, Brazil, e-mail: [email protected], [email protected]. March 1, 2018 DRAFT 2 IV Applications to Coding Theory 21 IV-A Packing Radius Bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 IV-B Optimizing the Full Lookup Table Searching in Syndrome Decoding . . . 23 References 27 I. I NTRODUCTION Coding theory is a mathematical theory that deals with many different aspects that emerges in the engineering of communication. A significant and classical part of it, the study of block-codes, is a subject that deals with many geometric and computational aspects of linear subspaces of finite dimensional vector spaces over finite fields. The most classic metric considered in this context is a metric that counts the number of different coordinates between two vectors, and it is named after R. W. Hamming, which introduced it in a seminal work [10, 1950]. Later, in ([21], 1957) and ([13], 1958), another well-known metric was proposed, the Lee metric. Up to our knowledge, the first work considering metrics in a general approach is a short communication of S. W. Golomb ([9], 1969). In that work, Golomb described a family of additive metrics (metrics defined over an alphabet and additively extended to a set of words with a fixed length) which are still being investigated nowadays, as we can see in [20]. Recently, due to the development of new communication channels and models, new families of metrics has been studied in the context of coding theory, as can be seen, for example, in [4], [2], [8] and [5]. The family of poset metrics was defined by Brualdi, Graves and Lawrence, in ([4], 1995) as an extension of the metrics proposed by Niederreiter in [15] and [16]. This is a large family of metrics which are determined by partial orders and generalizes the classical Hamming metric. Since then, it has been studied intensively, in order to understand the behaviour of metric invariants that are relevant to coding theory. As it frequently happens with generalizations, besides the possible inherent merits, some side effects may arise, either by giving a more profound understanding about the nature of previously established knowledge or by posing some new interesting (and difficult) questions. This is the case of poset metrics. Postponing the details for the next section, a poset metric depends on the non-zero positions of a vector and also on the level of these positions in a partial order (or the level in the Hasse DRAFT March 1, 2018 3 diagram). A significant part of the knowledge available for poset metrics can be understood once we consider a decomposition of a code (a vector subspace) which depend on the levels of the poset. Considering the metrics determined by the family of the so-called hierarchical posets, one can fully grasp the role of such decompositions. Being studied since the 90’s, results previously proved1 by many researchers were restated in an alternative and simple way [14], all the proofs were derived from the existence of a canonical decomposition. Besides the canonical-systematic form introduced in [7] for hierarchical poset metrics, the only known attempt to construct standard forms for poset metrics was made in [1], where a standard form for a particular case (Niederreiter-Rosenbloom-Tsfasman (NRT), or metrics induced by orders consisting of multiple disjoint chains with same cardinality) is presented. In that work, one can see that the standard form is not unique (in any possible sense). In a matter of fact, unicity of such a pre-established form is a characteristic of hierarchical posets [14, Theorem 3]. Since unicity of the form is not available, we strives for the best possible option: canonical forms determined by maximal decompositions, whose unicity is obtained according to the degree of the decomposition. Hence, the main subject of this work is how to decompose a subspace “respecting” the metric structure. For the reader that is not particularly interested in coding theory, despite the use of a terminology from coding theory, there is essentially a very short lexicon that needs to be translated: instead of “code” and “[n, k]q ”-code, one should read “vector subspace” and “k-dimensional vectors subspace of Fnq ”, respectively; instead of “C ⊆ Fnq ”, simply writes “V ⊆ Fnq ”. This work is organized as follows. In Section II we introduce the basic concepts of this work (poset metrics and P-decomposition) and describe an algorithm for decomposing a code. In Section III we show that the decomposition of a subspace is “well behaved” in the sense that when different posets are comparable, so are the corresponding decompositions. Using this fact and the knowledge about hierarchical poset metrics, in part III-B we produce bounds for the degree of a maximal P-decomposition of a code, where P is an arbitrary poset. Finally, on Section IV we go back to the coding ambient and give some applications: the determination of the packing radius (part IV-A) and the possible simplification of syndrome decoding algorithm 1 Results concerning the MacWilliams identity [12], association schemes [17], extension of isometries [3] (MacWilliams Extension Theorem) and the relation between the packing radius and the minimum distance [7]. March 1, 2018 DRAFT 4 (part IV-B). II. P OSET M ETRICS AND D ECOMPOSITION OF S UBSPACES This section has three parts. In the first part we summarize the basic concepts about poset metrics. In the second we introduce a new concept, object of study of this work: how different poset metrics allow different decompositions of a subspace. After doing so, we introduce the maximal and the primary P-decompositions of a subspace. The third part is devoted to an algorithm to find a maximal P-decompositions. A. Posets and Metrics Let P = ([n], 6P ) be a partially ordered set (or poset), where [n] := {1, . . . , n} and 6P is an order relation. We may denote the order relation only as a 6 b if the poset is immaterial or clear in the context. A set I ⊂ [n] is an order ideal of P if for every x ∈ I and y ∈ [n] with y 6P x implies y ∈ I. Given E ⊆ [n], the smallest ideal of P containing E is denoted by hEiP and called the ideal generated by E. The maximal elements (according to P) of E will be denoted by M(E). In case E = {a} is a singleton, we denote hEiP = haiP . An order isomorphism f : P → Q between posets P and Q over [n] is a bijection f : [n] → [n] such that a 6P b if, and only, if f (a) 6Q f (b), for all a, b ∈ [n]. In the case P = Q, f is said to be a P-automorphism. The group of all automorphisms of P will be denoted by Aut(P). Given a poset P = ([n], 6P ) and a subset X ⊂ [n], considering the order on X inherited from P, we say that (X, 6P ) is a sub-poset of P and denote X ⊆ P. A chain is a poset that is totally ordered, that is, for any a, b ∈ [n], either a 6 b or b 6 a. An anti-chain is a poset with no relations excepts the trivial relations a 6 a. Given a ∈ P, the height of a according to P is the maximal length of a chain in P that has a as a maximal element, i.e., h(a) = max{|C| : C ⊆ P is a chain and b 6P a ∀ b ∈ C}, and the height of P is h(P) = max{h(a) : a ∈ [n]}. The i-th level ΓPi of P is the set of all elements with height i, i.e., ΓPi = {a ∈ P : h(a) = i}. We remark that each level ΓPi is a chain and the extremal condition of chains can be grasped from the fact that h(P) = n if, and only if, P is a chain or, equivalently, if every level satisfies DRAFT March 1, 2018 5 |ΓPi | = 1. Similarly, for an anti-chain, we have h(P) = 1 and |ΓP1 | = n. If X ⊆ P is a chain or an anti-chain, we say that X is a chain or anti-chain in P. If X is a chain (anti-chain), we call its cardinality |X | the length of the chain (anti-chain). A broader class of posets, containing both anti-chain and chain posets, is the class of hierarchical posets. We say that P = ([n], 6P ) is hierarchical if elements at different levels are always j comparable, i.e., a 6P b for every a ∈ ΓPi and b ∈ ΓP where 1 ≤ i < j ≤ h(P). This is an interesting class of posets that will play a central role in the applications to coding theory, as we shall see in Section IV. We now show how a poset determines a metric over a finite-dimensional vector space. Due to the primary concern to applications in coding theory, we will restrict ourselves to the case of finite fields, so we let Fq be the finite field with q elements and Fnq be the vector space of all n-tuples over Fq . A k-dimensional subspace C ⊂ Fnq is called an [n, k]-linear code or, simply a code when the parameters are either clear from the context or not relevant. Given a vector v = (v1, . . . , vn ) ∈ Fnq , its support is the set of all non-null coordinates: supp(v) = {i ∈ [n] : vi , 0}. Given a poset P over [n], the poset weight (or P-weight) of an element v ∈ Fnq is the cardinality of the ideal generated by its support, i.e., wP (v) = |hsupp(v)iP |. The poset distance (or P-distance) dP is the distance function over Fnq induced by the P-weight, in other words, if u, v ∈ Fnq , then dP (u, v) = wP (u − v). The poset distance was introduced by Brualdi et.al. [4] and it is immediate to prove that (Fnq, dP ) is a metric space. The space (Fnq, dP ) is called P-space and the metric dP is called P-metric. A linear P-isometry is a linear map T : Fnq → Fnq preserving the P-weight, and hence the Pdistance. Two codes C1 and C2 over Fnq are said to be P-equivalent if there is a linear isometry T such that T(C1 ) = C2 . This is clearly an equivalence relation on the set of all codes and to determine the equivalence classes we need to know what are the linear P-isometries of the space. The group of all linear P-isometries over Fnq is denoted by GLP (Fnq ), and it is described by the next proposition. March 1, 2018 DRAFT 6 Proposition 1 ([19]): Let P be a poset over [n]. Given T ∈ GLP (Fnq ), the map φT (i) = M(supp(T(ei ))) is a P-automorphism. Furthermore, T ∈ GLP (Fnq ) if, and only if, Õ T(ei ) = xi j eφT (i), i6 P j where xi j are scalars satisfying x j j , 0 for every j ∈ [n]. Let GP be the subgroup of all linear isometries T such that φT = id. Considering the subgroup of linear isometries induced by the automorphisms of P (the permutation part), witch is commonly denoted by Aut(P), in [19], it was proved that GLP (Fnq ) = GP ⋊ Aut(P). B. Partitions and Decompositions A partition of a subset J ⊆ [n] is a family of non-empty subsets {J1, . . . , Jr } such that r Ø J= Ji and Ji ∩ J j = ∅ for every i , j. i=1 We will denote such partition by J = (Ji )ri=1 . The triple J ∗ = (J; J0; Ji )ri=1 where J0 = [n] \J = {i ∈ [n] : i < J} is called pointed partition. Since J is the union of the subsets Ji , the pointed partition J ∗ is completely determined once the pair (J0 ; Ji )ri=1 is given, so we may also write J ∗ = (J0 ; Ji )ri=1 . Note that J0 = ∅ if, and only if, J = [n]. We stress that the pointer J0 has a special role, since it is the only part we allow to be empty. From now on, we consider only pointed partitions, hence we will omit the symbol ∗ and the adjective “pointed”. A partition J can be refined in two ways: either by increasing the number of parts or by enlarging the distinguished part J0 . Except for the pointer J0 , the order of the other parts is irrelevant, for example, (J0 ; {1, 2} , {3, 4, 5}) = (J0 ; {5, 4, 3} , {1, 2}) . Definition 1: Given a partition J = (J0 ; Ji )ri=1 , a split of J is a partition J ′ = (J0 ; Ji′)r+1 i=1 ′ where Jl = Jl′ ∪ Jr+1 for some l ∈ [r] and Ji = Ji′ for each i , l. This means that Jl is split into two components and the others are unchanged. An aggregate of J is a partition J ′ = (J0′ ; Ji′)ri=1 where for some l ∈ [r], Ji′ = Ji if i < {l, 0}, Jl = Jl′ ∪ Jl∗ and J0′ = J0 ∪ Jl∗ for some ∅ , Jl∗ Jl , i.e., some elements of Jl were aggregated into the distinguished part J0 . Definition 2: We say that a partition J ′ is a 1-step refinement of J if J ′ is obtained from J either by a splitting or by aggregating only one element. The partition J ′ is a refinement of DRAFT March 1, 2018 7 J if J ′ can be obtained from J by a successive number of 1-step refinements and, if that is the case, we use the notation J ≥ J ′. Example 1: The partition ([4]; ∅; {1, 2, 3, 4}) can be refined in order to get the partition ([4]; {1, 3}; {2}, {4}) by using the following 1-step refinements: (∅; {1, 2, 3, 4}) ≥ ({3} ; {1, 2, 4}) ≥ ({1, 3} ; {2, 4}) ≥ ({1, 3} ; {2} , {4}) . Set partitions and refinement operations allow us to define decompositions of codes. Definition 3: We say that C = (C; C0 ; Ci )ri=1 is a decomposition of an [n, k, δ]q code C if each Ci is a subspace of Fnq and r C with dim(C ) > 0 for every i ∈ [r]; (a) C = ⊕i=1 i i (b) C0 = {(x 1, x2, . . . , xn ) : xi = 0 if i ∈ supp (C)}; (c) (supp (C0 ) ; supp (Ci ))ri=1 is a partition over [n]. Definition 4: A split, aggregate, 1-step refinement and a refinement C ′ = (C; C0′ ; Ci′)r′ i=1 of a decomposition C = (C; C0 ; Ci )ri=1 are defined according to (supp(C0′ ); supp(Ci′))r′ i=1 being a split, aggregate, 1-step refinement or a refinement of (supp (C0 ) ; supp (Ci ))ri=1 respectively. We will also use the notation C ≥ C ′ to denote a refinement. Example 2: Let P be a poset over [4]. Consider the [4, 2, δP ]2 code C given by C = {0000, 1100, 0010, 1110}. Then, 2 (C; he4 i; C) ≥ (C; he4 i; Ci )i=1 where C1 = {0000, 1100} and C2 = {0000, 0010}. Definition 5: A decomposition C = (C; C0; Ci )ri=1 is said to be maximal if it does not admit a refinement. Until now, only the algebraic aspects of the codes were used, the metric dP has not played any role in the decomposition of a code. We introduce now a decomposition that depends of the metric and consequently of the poset P. Definition 6: A P-decomposition of C is a decomposition C = (C ′; C0′ ; Ci′)ri=1 of C ′ (as in Definition 3) where C ′ ∼P C. Each Ci′ is called a component of the decomposition. A trivial P-decomposition of C is either the decomposition (C; C0; C) or any P-decomposition with a unique factor (C ′; C0′ ; C ′) where |supp(C ′)| = |supp (C)| and |supp(C0′ )| = |supp (C0 )|. March 1, 2018 DRAFT 8 Definition 7: A code C is said to be P-irreducible if it does not admit a non-trivial Pdecomposition. Example 3: Let P be the hierarchical poset with order relations 1, 2, 3 6 4. The codes C = {0000, 1001} and C ′ = {0000, 0111} are equivalent since T(e1 ) = e1, T(e2 ) = e2, T(e3 ) = e3, T(e4 ) = e1 + e2 + e3 + e4 is an isometry satisfying T(C) = C ′. Then, C ′ = (C ′; he1 i; C ′) is a P-decomposition of C while C = (C; he2, e3 i; C) is a decomposition of C. We note that if we consider the partitions ([4]; {2, 3}; {1, 4}) and ([4]; {1}; {2, 3, 4}) induced by the P-decompositions C and C ′ constructed in the previous example, these partitions can not be obtained from each other by a sequence of aggregations and splittings. Hence, we must use the order automorphism induced by the isometries in order to compare them. We now do it explicitly. ′ Let C = (C ′′; C0′′; Ci′′)ri=1 and C ′ = (C ′; C0′ ; Ci′)ri=1 be two P-decompositions of a code C. Associated to those P-decompositions there are two partitions of [n], namely: (supp(C0′′); supp(Ci′′))ri=1 ′ and (supp(C0′ ); supp(Ci′))ri=1 . By the definition of a P-decomposition, there are isometries T1, T2 ∈ GLP (Fnq ) such that T1 (C) = C ′′ and T2 (C) = C ′. Denote T = T2 ◦ (T1 )−1 . Then T is a linear isometry and T(C ′′) = C ′. Considering the automorphism φT of Proposition 1 (the order automorphism induced by T), φT induces a map on the partition of [n] determined by the Pdecomposition C , namely, φT [(supp(C0′′); supp(Ci′′))ri=1 ] = (φT (supp (C0 )) ; φT (supp (Ci )))ri=1 . Considering the map T defined in Example 3, we may define T1 = identity and T2 = T. Then, φT [([4]; {2, 3}; {1, 4})] = ([4]; φT ({2, 3}); φT ({1, 4})) = ([4]; {2, 3}; {1, 4}) is a partition which also can not be obtained by ([4]; {1}; {2, 3, 4}) using an aggregation. Considering the isometry S(e1 ) = e3, S(e2 ) = e2, S(e3 ) = e1, S(e4 ) = e2 + e4 we still have that S(C ′′) = C ′ but now φS [([4]; {2, 3}; {1, 4})] = ([4]; φS ({2, 3}); φS ({1, 4})) = ([4]; {1, 3}; {2, 4}) DRAFT March 1, 2018 9 is an aggregation of ([4]; {1}; {2, 3, 4}). Hence, it may happen that C ′ and C ′′ may not be comparable for some U ∈ GLP (Fnq ) but comparable for some T ∈ GLP (Fnq ). Therefore, we can define the analogous of the operations on decompositions by demanding the existence of such an isometry. ′′ ′′ ′′ ′′ r′′ Definition 8: Let C ′ = (C ′; C0′ ; Ci′)r′ i=1 and C = (C ; C0 ; Ci )i=1 be two P-decompositions of C. We say that C ′ is a P-refinement of C ′′ (and write C ≥ C ′) if there is T ∈ GLP (Fnq ) such ′′ ′′ ′′ r that φT [(supp(C0′ ); supp(Ci′))r′ i=1 ] is a refinement of the partition (supp(C0 ); supp(Ci ))i=1 . Since, for every T ∈ GLP (Fnq ), the induced map φT is an automorphism, if C and C ′ are comparable by S and T respectively, the number of aggregations and splittings used to obtain C ′ from C ′′ coincides when S or T are used to compare them. Hence, splittings and aggregations are well determined once two P-decompositions are given. Definition 9: A P-decomposition C = (C; C0 ; Ci )ri=1 is said to be maximal if it can not be refined or, equivalently, if each Ci is P-irreducible for every i ∈ [r]. Let us consider the Hamming metric dH over Fn2 . It is well-known that the group of linear isometries of this metric space is isomorphic to the permutation group Sn . Given a code C, let C = (C; C0; Ci )ri=1 be a maximal decomposition of C. If a code C ′ is dH -equivalent to C, then there is T ∈ GLdH (Fnq ) ∼ Sn such that T (C) = C ′. Note that the decomposition C ′ = (C ′; T(C0); T(Ci ))ri=1 is a maximal decomposition of C ′, otherwise, C would not be a maximal decomposition of C. Therefore, when considering the Hamming metric, a maximal decomposition is also a maximal H-decomposition (H is the anti-chain poset over [n]). This is not true in general, as we can see in Example 3. Definition 10: Let C = (C ′; C0; Ci )ri=1 be a P-decomposition of an [n, k, δ]q code C. The profile of C is the array profile (C ) := [(n0, k0 ) , (n1, k1 ) , . . . , (nr , kr )] , where ni = |supp (Ci )| and ki = dim (Ci ) . Let β = {e1, . . ., en } be the canonical basis of Fnq . Given I ⊂ [n], the I-coordinate subspace VI is defined by VI = span {ei : i ∈ I} = ( Õ i∈I ) xi ei : xi ∈ Fq . Given a decomposition C = (C; C0; Ci )ri=1 of a linear code C, we say that March 1, 2018 DRAFT 10 Vi := Vsupp(Ci ) = span {ei : i ∈ supp (Ci )} is the support-space of (the component) Ci . We consider [n]C = supp (C) and [n]C = [n] \ [n]C . In the case where [n]C , ∅, we write V0 = V[n] C and denote C0 = V0 . We say that the decomposition (C; C0; Ci )ri=1 is supported by the environment decomposition (V0 ; Vi )ri=1 . In the case where [n]C = ∅, we have V0 = C0 = {0}. Given a P-decomposition C = (C ′; C0′ ; Ci′)ri=1 of C, the environment decomposition determine a partition of [n], i.e., [n] = ∪ri=0 supp (Vi ). Also, if Ci′ = {0}, then i = 0. If each Ci′ is Pirreducible, denoting ni = dim(Vi ) = |supp(Ci )| and ki = dim(Ci′), then r Õ ni = n = dim(Fnq ) and i=0 r Õ ki = k = dim(C). i=1 The following theorem states that the profile of a maximal P-decomposition C of a code C depends (essentially) exclusively on C, not on C . Theorem 1: Let C be an [n, k, δ]q code and let P be a poset over [n]. If C ′ and C ′′ are two maximal P-decompositions of C with   profile(C ′) = (n′0, k ′0 ), (n′1, k1′ ), . . ., (nr′ , kr′ ) and   profile(C ′′) = (n′′0 , k0′′), (n′′1 , k1′′), . . . , (n′′s , k s′′) . Then, r = s and, up to a permutation, profile (C ′) = profile (C ′′), i.e., there is σ ∈ Sr such that ′′ ) and (n′ , k ′ ) = (n′′, k ′′). (ni′, ki′) = (n′′σ(i), k σ(i) 0 0 0 0 s Proof: Let C ′ = (C ′; C0′ ; Ci′)ri=1 and C ′′ = (C ′′; C0′′; Ci′′)i=1 be two maximal P-decompositions of C. Suppose, without loss of generality, r < s. If T ∈ GLP (Fnq ) is an isometry satisfying T (C ′) = C ′′, there is a component Ci′ of C ′ such that T(Ci′) is not contained in any component Cj′′ of C ′′, otherwise r ≥ s. Hence, there are components Ci′0 of C ′ and Cj′′0, Cj′′1, . . ., Cj′′t of C ′′ such that T(Ci′0 ) ⊂ Cj′′0 ⊕ Cj′′1 ⊕ · · · ⊕ Cj′′t and T(Ci′0 ) ∩ Cj′′l , ∅ for any l ∈ {1, . . . , t}. Therefore, T(Ci′0 ) = t Ê T(Ci′0 ) ∩ Cj′′m m=0 is a non-trivial P-decomposition for Ci′0 , contradicting the fact that each component of a maximal P-decomposition is P-irreducible. It follows that r = s. Moreover, for every i ∈ {1, . . . , r } there DRAFT March 1, 2018 11 is ji such that T(Ci′) ⊆ Cj′′i . Hence, ni′ ≤ n′′ji and ki′ ≤ k ′′ji . Applying the same reasoning to T −1 ∈ GLP (Fnq ), we get that ni′′ ≤ n′ji and ki′′ ≤ k ′ji , hence ni′ = n′′ji and ki′ = k ′′ji , so that, up to a permutation, profile(C ′) = profile(C ′′). The next Corollary follows straight from Theorem 1. Corollary 1: Let C ′ = (C ′; C0′ ; Ci′)ri=1 and C ′′ = (C ′′; C0′′; Ci′′)ri=1 be two maximal P-decompositions of C and let T ∈ GLP (Fnq ) be a linear isometry such that T (C ′) = C ′′. Then, there is a permutation ′′ . σ ∈ Sr such that T(Ci′) = Cσ(i) To express the number of operations (splitting and aggregations) performed in a decomposition from a trivial decomposition, we first need to show that every non-trivial P-decomposition is a refinement of a trivial P-decomposition. Lemma 1: If T ∈ GLP (Fnq ) then the ideals hsupp(C)i and hsupp(T(C))i are isomorphic. Proof: Given T ∈ GLP (Fnq ), follows immediate from the definition of φT (Proposition 1) that φT (M(supp(C))) = M(supp(T(C))). Hence, hsupp(T(C))i = hM(supp(T(C)))i = hφT (M(supp(C)))i. On the other hand, since φT is an isomorphism, hsupp(C)i = hM(supp(C))i ∼ φT (hM(supp(C))i) = hφT (M(supp(C)))i. Therefore, hsupp(C)i ∼ hsupp(T(C))i. Proposition 2: If C = (C ′; C0′ ; Ci′)ri=1 is a non-trivial P-decomposition of C with |supp(C0′ )| > |supp(C0)|, then C is a refinement of a trivial P-decomposition of C. Proof: It is straightforward that C is a refinement of the one-component P-decomposition (C ′; C0′ ; C ′). To conclude, we need to prove that (C ′; C0′ ; C ′) is obtained by aggregations from a trivial P-decomposition. Since |supp(C ′)| < |supp(C)| and |M(supp(C ′))| = |M(supp(C))|, there exist i0 ∈ hsupp(C ′)i \ supp(C ′), otherwise hsupp(C)i / hsupp(C ′)i contradicting Lemma 1. Take j0 ∈ M(supp(C ′)) such that j0 > P i0 and define S ∈ GLP (Fnq ) by S(ei ) = ei ∀ i , j0 March 1, 2018 and S(e j0 ) = e j0 + ei0 . DRAFT 12 Therefore, φS = id. Denoting S(C ′) = C ′′, since supp(C ′) = supp(C ′′) ∪ {i0 } we get that (C ′; C0′ ; C ′) is a 1-step refinement of (C ′′; C0′′; C ′′). Proceeding in this form (performing aggregations), at some point we get a trivial P-decomposition having C as its refinement. Definition 11: Given a P-decomposition C = (C ′; C0′ ; Ci′)ri=1 of C, its degree according to C is defined by DP (C , C) = (r − 1) + |supp(C0′ )| − |supp(C0 )|. A P-decomposition with maximum degree is called primary P-decomposition. The degree of a decomposition is basically the number of 1-step refinements necessary to obtain the given decomposition from a trivial decomposition, clearly, if C is a trivial decomposition, DP (C , C) = 0 and if C ′ is a P-refinement of C , then DP (C ′, C) > DP (C , C). Furthermore, the degree of a decomposition is completely determined by its profile. Hence, by Theorem 1, maximal P-decompositions have the same degree. Thus, the maximum degree of a decomposition of a code C will be denoted by DP (C) instead of DP (C , C). The following proposition follows straight from the previous comments. Proposition 3: A P-decomposition of a code C is maximal if, and only if, it is a primary P-decomposition. We recall that the group of linear P-isometries is the semi-direct product GLP (Fnq ) = GP ⋊ Aut(P). It is worth remarking that the permutation part Aut(P) is irrelevant regarding maximality of P-decompositions: Proposition 4: Let C = (C ′; C0; Ci )ri=1 be a maximal P-decomposition of C. Let φ ∈ Aut (P) and Tφ ∈ Aut(P) be the isometry induced by φ. Then, C ′ = Tφ (C ′) ; Tφ (C0 ) ; Tφ (Ci ) r i=1 is also a maximal P-decomposition of C. Proof: Because φ is a permutation, for every i ∈ {0, 1, . . ., r }, j ∈ supp(Ci ) ⇐⇒ φ( j) ∈ supp(Tφ(Ci )). Hence, C ′ is a P-decomposition of C. Since its profile coincides with the profile of C , C ′ is also a maximal P-decomposition of C. DRAFT March 1, 2018 13 C. Construction of Maximal P-Decompositions In order to find maximal P-decompositions of a code C, we need to obtain the maximal decompositions of its equivalent codes. In this section, we will describe the construction of a generator matrix which provides a code equivalent to C and determines a maximal P-decomposition of C. Let P be a poset over [n] and G = (gi j ) be a k × n generator matrix of an [n, k, δ]q code C. We lose no generality by assuming that G is in a reduced row echelon form, obtained by elementary operations on rows. In order to obtain a maximal P-decomposition, we need to use a slight different definition for the classical reduced row echelon form. For each i ∈ [k], let  j (i) := max j : gi j , 0 be the right-most non-zero column of the i-th row of G. Performing elementary row operations on G, we may assume that j(1) > j(2) > · · · > j(k) and gi j(l) = 0 if i , l. (1) We say that G is in inverse reduced row echelon form if the entries of G satisfy Property (1). From now on, we assumed that generator matrices have this form. In order to make more clear the difference of the proposed reduced row echelon form and the classical one, we present the next example. Example 4: Let    1 0 1 1 0      G= 1 1 0 1 1       0 1 0 1 1    be a generator matrix of a [5, 3]2 code C (the metric here is irrelevant). Then,     G1 =      are also generator matrices     0 1 1 0 1  1 0 0 0 0       and G =  0 0 1 1 0  0 1 0 1 1  2       0 0 1 1 0   1 0 0 0 0     of C. Note that G1 is in the classical reduced row echelon form while G2 is in the proposed one. A generator matrix G of C determines a unique decomposition C in the following sense: Construction of C : March 1, 2018 DRAFT 14 Suppose β1 = {v1, . . ., vk } is the set of all rows of G and I ⊂ [n] is the index set of the null columns of G. Then, define ( C0 = v ∈ Fnq : v = Õ i∈I ) xi ei . Take w1 ∈ β1 and let γ1 = {vi1, . . ., vir } ⊂ β1 be the set of all rows of G such that if v ∈ γ1 , then supp(w1 ) ∩ supp(v) , ∅. Denote    r Õ    C1 = c ∈ C : c = xi vi j .   j=1   Take β2 = β1 \γ1 , if it is empty, the decomposition (C; C0; C1 ) is the one determined by G. If β2 , ∅, take w2 ∈ β2 and γ2 = {vi1, . . ., vis } ⊂ β2 where v ∈ γ2 if, and only if, supp(w2 )∩supp(v) , ∅. Define Proceeding in this way, βr+1    s Õ    C2 = c ∈ C : c = xi vi j .   j=1   = ∅ and βr , ∅ for some r. In this case, C = (C; C0 ; Ci )ri=1 is the decomposition determined by G. It is clear that the choice of wi does not change the degree of the decomposition, therefore, the decomposition constructed is unique, up to a permutation of its components. As we shall see in the next proposition, this decomposition is maximal. Example 5: Consider the generator matrices of Example 4. The decomposition determined by G is trivial since every two rows of G have intersection in their supports. On the other hand, the decomposition obtained by the matrix G2 is not trivial, indeed, the third row has disjoint support from the first and second rows. Therefore, if C1′ = {00000, 01011, 00110, 01101}, C2′ = {00000, 10000}, and C1 = {00000, 01101, 00110, 01011}, C2 = C2′, 2 and (C; ∅; C )2 are the decomposition obtained from G and G respectively. then (C; ∅; Ci′)i=1 i i=1 1 2 A generator matrix G is said to be in generalized reduced row echelon form if there is a permutation σ ∈ Sk such that j(σ(1)) > . . . > j(σ(k)) and gi j(l) = 0 if i , l, i.e., if a matrix in inverse reduced row echelon form may be obtained by permuting the rows of G. DRAFT March 1, 2018 15 Proposition 5: Let C be an [n, k, δ]q code. If G is a generator matrix of C in a generalized reduced row echelon form, then G determines a maximal decomposition for C. Proof: Let G be a generator matrix of C in a generalized reduced row echelon form. Without loss of generality, we will assume that the decomposition determined by G has only one component, i.e., C = (C; C0; C). By the construction of C it is straightforward that C can 2 is a splitting of C . Thus, C = C ⊕ C not be refined by aggregations. Suppose C ′ = (C; C0 ; Ci )i=1 1 2 and supp(C1 ) ∩ supp(C2) = ∅. Let {w1, . . ., wk1 } and {wk1 +1, . . ., wk } be a basis of C1 and C2 respectively. The k × n matrix G1 having {w1, . . ., wk } as its rows is the generator matrix of C whose decomposition induced is C ′. Then, each row of G1 is a linear combination of the rows of G. Let β1 be the minimum set of rows of G generating {w1, . . ., wk1 }, i.e., for every i ∈ {1, . . ., k1 }, wi = Õ xvi v v∈β1 where xvi ∈ Fq . Similarly, let β2 be the minimum set of rows of G generating {wk1 +1, . . ., wk }. Since G and G1 generates the same code, β1 ∩ β2 = ∅. Since the decomposition induced by G is trivial, supp(β1 ) ∩ supp(β2 ) , ∅. Take i0 ∈ supp(β1 ) ∩ supp(β2 ), since span{β1 } = C1 and span{β2 } = C2 , it follows that i0 ∈ supp(C1) ∩ supp(C2), a contradiction. The previous proposition ensures that each maximal decomposition is obtained by taking a generator matrix in a generalized reduced row echelon form. Till now, we have not used the group of linear isometries to construct maximal P-decompositions. In the following, we will consider the subgroup of isometries GP (as we saw, the automorphism part is not relevant when searching for maximal P-decompositions since it does not change the degree of a decomposition). By definition of GP , the following two operations over a generator matrix G will provide matrices generating equivalent codes to the one generated by G = (gi j ): (OP 1) If gi0 j0 , 0, gi j0 = 0 for every i , i0 (gi0 j0 is the only non-zero entry on the j0 -th column) and r 6P j0 (r , j0 ), we may assume gi0r = 0; This is equivalent to choosing the isometry T ∈ GP such that T(e j ) = e j for every j , j0 and T(e j0 ) = e j0 − gi0r gi−1 e. 0 j0 r (OP 2) More generally, if there is s ∈ [n] such that r 6P ji and r , ji for every i ∈ {1, . . . , s}, Ís furthermore, there are two rows of G, namely, i1, i2 ∈ [k], such that gi1r = l=1 xl gi1 jl Ís and gi2r = l=1 xl gi2 jl for some choice of x1, . . . , xl ∈ Fq , and gi jl = 0 for every i , i1, i2 , March 1, 2018 DRAFT 16 we may assume gi1r = gi2r = 0. The procedure can be performed simultaneously to many lines and all those entries may be considered to be 0. Furthermore, if the column r is a linear combination of columns j1, . . . , js , then one may exchange the column r by the null column. Let {g1, . . . , gn } be the set of columns of G and suppose s Õ gr = xi g ji . i=1 In order to exchange the r-th column of G by a null column we consider the isometry T ∈ GP defined by T(ei ) = ei for every i < { j1, . . ., js } and T(e ji ) = e ji − xi er for every i ∈ {1, . . . , s}. Definition 12: Let G be a generator matrix of an [n, k, δ]q code C. If G is in generalized reduced row echelon form and none non-null entry of G may be exchanged by zero using Operations (OP 1) or (OP 2), we say that G is in P-canonical form. Example 6: Let C be the [6, 3]2 code with generator matrix given by    0 0 1 1 0 1      G=  1 0 1 1 1 0 .      1 1 0 0 0 0    Consider the poset P1 with order relations 1 6P1 2 and 3 6P1 4. Applying operations (OP 1) and (OP 2) we get the following matrix in P1 -canonical form:    0 0 0 1 0 1      G′ =  1 0 0 1 1 0  .      0 1 0 0 0 0    Furthermore, if we consider P2 a poset such that P1 ⊂ P2 and 4 6P2 5, then performing operation (OP 1) we get   0   G′′ =  1    0  ′′ It is clear that G is in P2 -canonical form. decomposition for C.  0 0 1 0 1   0 0 0 1 0 .   1 0 0 0 0   It is also clear that it determines a maximum P2 - Theorem 2: Let G be a generator matrix of a code C. If G is in a P-canonical form then the decomposition determined by G is a maximal P-decomposition of C. DRAFT March 1, 2018 17 Proof: Assume, without loss of generality, C = (C; C0; C). We will first show that C does not admit aggregations. Suppose there is a P-decomposition C ′ = (T(C); C0′ ; T(C)) where T ∈ GP and φT (supp(C0)) ⊂ supp(C0′ ). Since T ∈ GP , the map φT coincides with the identity map, hence supp(C0 ) ⊂ supp(C0′ ). Let G2 = T(G) be the matrix whose i-th row is obtained by the action of T in the i-th row of G = (gi j ). Thus, G2 is a generator matrix of T(C). If i0 ∈ supp(C0′ ), then i0 is a null column of G2 . The characterization of GP ensures that for every row gi = (gi1, . . . , gin ) of G, the i0 -th coordinate of T(gi ), which is null and denoted by T(gi )i0 , is given by Õ Õ 0 = T(gi )i0 = x ji0 gi j = gii0 xi0 i0 + x ji0 gi j (2) for every i ∈ {1, . . ., k}. Since xi0 i0 , 0, it follows that Õ  x ji  − 0 gi j gii0 = xi0 i0 j (3) j i0 6 P j j j,i0, i0 6 P j j,i0, i0 6 P j for every i ∈ {1, . . ., k}. Since G is in P-canonical form, Equation 3 together with Operation (OP 2) ensures that the i0 -th column of G is null, so i0 ∈ supp(C0 ). Therefore, supp(C0′ ) = supp(C0 ). The decomposition C also does not admit a splitting, indeed, suppose otherwise and let 2 C ′ = (T(C); C0 ; Ci )i=1 be its refinement. Let G1 be as in Proposition 5, then G1 is a generator matrix of T(C) and, by construction, the decomposition determined by G1 is C ′. Let β1 be the minimal set of rows of G such that for every i ∈ {1, . . ., k1 }, Õ wi = xgi T(g) (4) g∈β1 where xvi ∈ Fq . Let β2 be the minimal set of rows of G such that Õ wi = xgi T(g) (5) g∈β2 for every i ∈ {k1 + 1, . . . , k}. We stress that the only difference between this equalities and the ones in Proposition 5 is the presence of the linear isometry T. As before, we also have that β1 ∩ β2 = ∅. By construction, supp(β1 ) ∩ supp(β2 ) , ∅. Let i0 be the rightmost column of G such that i0 ∈ supp(β1 ) ∩ supp(β2 ), if i0 < supp(T(β1 )) then for every gi = (gi1, . . ., gin ) ∈ β1 (which is also a row of G), the i0 -th coordinate of T(gi ) is given by Equation 2. Since i0 is the rightmost column of G such that i0 ∈ supp(β1 ) ∩ supp(β2 ). Equation 3 ensures that the i0 -th column of G is null, a contradiction. The same argument shows that i0 ∈ supp(T(β2 )). Therefore, i0 ∈ supp(T(β1 )) ∩ supp(T(β2 )). By identities 4 and 5, there exist j ∈ {1, . . ., k1 } and l ∈ {k1 + 1, . . ., k} such that i0 ∈ supp(w j ) ∩ supp(wl ), a contradiction. March 1, 2018 DRAFT 18 Each time we exchange a non-null column by a null column, we exchange a code C with Pdecomposition C = (C; C0; Ci )ri=1 by a P-equivalent code with P-decomposition C ′ = (C ′; C0′ ; Ci′)ri=1 where dim(C0′ ) = dim(C0) + 1. If the Operation (OP 2) is performed in a proper subset of k-lines of G, we do not increase dim(C0 ) but, we may split some of the Ci into Ci = Ci′1 ⊕ Ci′2 with supp(Ci′1 ) ∩ supp(Ci′2 ) = ∅. Note that the role of P in such operations rests solely on the condition j 6P j1, j2, . . . , js . For the two extremal posets, namely, the anti-chain and the chain poset, the picture is absolutely clear: If P is an anti-chain (hence we are considering the Hamming metric), no such operation may be performed (since i 6P j ⇐⇒ i = j). Hence, maximal P-decompositions coincide with maximal decompositions (see the paragraph after Definition 9) and the P-canonical form is a permutation of the classical systematic form; if P is a chain with 1 6P 2 6P · · · 6P n, then (OP 1) may be performed to every j (i) in the reduced row echelon matrix G, hence, C is equivalent to a code that has a generator matrix G = (gi j ) where gi j(i) = 1 and gi j = 0 if j , j(i) (see [18]). More general, for N RT metrics, the P-canonical form is also a permutation of the N RT-triangular form presented in [1]. We stress that, fixing the natural labeling in the posets, the permutations necessary to obtain one form from other are the ones induced by Aut(P). The other case that can be easily described is the case of hierarchical posets. The algorithm to find P-decompositions according to the levels of the poset was first proposed in [7]. The matrix obtained by this form is called canonical-systematic and it is also a permutation of the P-canonical form. Furthermore, in this case, the canonical-systematic form is a standard representation in the sense that every code has such decomposition. In [14], it was proved that for every non-hierarchical poset, it is not possible to give a standard representation like this one (whose support of each component is entirely contained in a specific level of the poset). III. C OMPARISON OF P OSETS AND D ECOMPOSITIONS As proved in [7, Corollary 1], if P is a hierarchical poset, a linear code may be canonically decomposed as T(C) = C1 ⊕ · · · ⊕ Cr where supp (Ci ) ⊂ ΓPi . As we shall explain in Section IV, this canonical decomposition permits to give explicit formulae and constructions for metric invariants. Since this decomposition (according to the levels of the poset) is a characteristic of hierarchical posets, see [14, Theorem 3], we aim to use the knowledge about hierarchical poset metrics to establish bounds in the general case. To do that, we need to show that the P-primary decomposition is “well behaved”, in the sense that the decomposition are comparable when DRAFT March 1, 2018 19 the posets are comparable. The meaning of the expression “well behaved” will be explained in Theorem 3. A. Decompositions and refinements The set of all posets over [n] is denoted by Pn . It is, on itself, a partially ordered set with the order relation given by P ≤ Q if for every a, b ∈ [n] with a 6P b implies a 6Q b. The set Pn has n! (isomorphic) chain posets and one anti-chain poset, these n! + 1 elements may be considered as extremal posets, indeed, it is easy to verify that adding a relation to a chain will give rise to a cycle a 6 b 6 a since a total order on [n] is defined once we state that σ(1) ≤ σ(2) ≤ . . . ≤ σ(n) for some permutation σ ∈ Sn , contradicting the anti-symmetry property of a partial order. On the other hand, it is also immediate to realize that no relation can be removed from an anti-chain since a 6 a for all a ∈ [n] are the unique relations on an anti-chain. Theorem 3: Let P, Q ∈ Pn with P ≤ Q. Given a code C, there is a maximal P-decomposition of C which is a Q-decomposition of C. Proof: Assume P, Q ∈ Pn and P < Q. Let C ′ = (C ′; C0 ; Ci )ri=1 be a maximal P-decomposition of C and T ∈ GLP (Fnq ) such that T (C) = C ′. By the characterization of GLP (Fnq ), T = A ◦ Tφ where A ∈ GP and Tφ ∈ Aut(P). From Proposition 4, we have that  r C ′′ = Tφ−1 (C ′) ; Tφ−1 (C0 ) ; Tφ−1 (Ci ) i=1 is also a maximal P-decomposition of C. However Tφ−1 (C ′) = Tφ−1 ◦ T(C), hence Tφ−1 (C ′) = Tφ−1 ◦ A ◦ Tφ (C). We stress that GP ⊂ GQ always that P ≤ Q. Because GP is a normal subgroup of GLP (Fnq ), it follows that Tφ−1 ◦ A ◦ Tφ ∈ GP , hence Tφ−1 ◦ A ◦ Tφ ∈ GQ . Therefore, C ′′ is a Q-decomposition of C. As a direct consequence of the previous theorem, we obtain a relation among primary decom- positions of posets and the natural order over Pn . Corollary 2: If P, Q ∈ Pn with P ≤ Q. Then, DP (C) ≤ DQ (C) for every linear code C. Concerning primary P-decompositions, there is always a code that it is differently decomposed depending on the poset, that is, the preceding inequality is strict: Proposition 6: If P, Q ∈ Pn with P < Q. Then, there is a code C such that DP (C) < DQ (C). March 1, 2018 DRAFT 20 Proof: Suppose P < Q. Hence, there are i0, j0 ∈ [n] such that i0 P j0 and i0 6Q j0 . Let C = span{ei0 + e j0 } and C ′ = span{e j0 } be two codes over Fnq . The linear map T defined by T(e j ) = e j for every j ∈ [n] \ {i0, j0 }, T(ei0 ) = ei0 and T(e j0 ) = e j0 − ei0 , is a Q-linear isometry, by the characterization of GLQ (Fnq ) given in Proposition 1. Therefore, the Q-decomposition C ′ = (C ′; C0′ ; C ′) of C, where C0′ = {v ∈ Fnq : v j0 = 0}, is a primary Q-decomposition and it follows that DQ (C) = 1. On the other hand, when considering the P-metric, the decomposition C = (C; C0 ; C), where C0 = {v ∈ Fnq : vi0 , 0 and v j0 , 0}, can not be refined since C is a one-dimensional code (hence, splittings are not allowed) whose support is composed by two maximal elements of P (hence, aggregations are not allowed). Therefore, C ′ is a primary P-decomposition for C and DP (C) = 0. We remark that Corollary 2 together with Proposition 6 implies that primary decomposition is a characterization of posets, in the sense that a given poset P may be reconstructed from the profile of codes according to P. Moreover, looking at the proof of Proposition 6, one may notice that the reconstruction can be done by considering only the n(n − 1)/2 pairs of vectors (ei, e j ). B. Hierarchical Bounds As we have mentioned, metrics induced by hierarchical posets are well understood. This knowledge was mainly obtained from the existence of the decomposition derived from the canonical-systematic form. Since the degree of a decomposition express the amount of 1step refinements performed from the trivial decomposition, we aim to use the knowledge of hierarchical posets and establish bounds for the degree considering the easy-to-compute degree of primary P-decompositions relatively to hierarchical posets. As we have mentioned, hierarchical poset metrics are well understood and, in particular, if P is a hierarchical poset, the profile (and consequently the degree) of a primary P-decomposition of a code C can be easily computed once we know the canonical-systematic form. For this reason, when considering a general poset P, we aim to establish bounds for DP (C) considering the easy-to-compute primary P-decompositions relatively to hierarchical posets. Considering the natural order ≤ on Pn , out of P we can define two hierarchical posets: DRAFT March 1, 2018 21 1 - Upper neighbor: P + = min {Q ∈ Pn : P ≤ Q and Q is hierarchical} . 2 - Lower neighbor: P − = max {Q ∈ Pn : Q ≤ P and Q is hierarchical} . The next proposition follows directly from Corollary 2. Proposition 7: For any linear code C, DP+ (C) ≤ DP (C) ≤ DP− (C) . If P is not hierarchical, both inequalities are strict for some code C. Moreover, the bounds are tight, in the sense that, given a poset P, there are codes C1 and C2 such that DP+ (C1 ) = DP (C1 ) and DP− (C2 ) = DP (C2 ) (just consider any code C with supp(C) ⊂ ΓP1 ). IV. A PPLICATIONS TO C ODING T HEORY The poset-metrics, as much as the decomposition problem, arose in the context of coding theory, what justified the use of the terminology and the notation from the area: code and codeword for vector subspace and vector, respectively. Now comes the time to justify it with some content, i.e., with applications to coding theory. We combine the knowledge available for hierarchical posets with the degree’s bound found in Section III-B to produce bounds for the most important invariant in coding theory (the packing radius) and to optimize the traditional full lookup table searching algorithm in the syndrome decoding process. A. Packing Radius Bounds If Fnq is endowed with a P-metric, an [n, k, δP ] P-code is an [n, k]-linear code C with minimum distance δP = min{wP (c) : c ∈ C} = min{dP (c, c′) : c, c′ ∈ C, c , c′ }. In the context of coding theory, the minimum distance is considered the most important metric parameter of a code. Its importance comes from the fact that it may ensures the correction of a certain amount of errors. If P is an anti-chain, the P-weight and metric are known as Hamming weight and metric, denoted by wH and δH , respectively. This is the most important setting in March 1, 2018 DRAFT 22 the context of coding theory. In this situation, let us say that a message c ∈ C is transmitted and it is received with errors as a message v = c + e, where e ∈ Fnq is the error vector. If δH is the minimum distance of C and wH (e) ≤ ⌊(δH − 1)/2⌋ (where ⌊·⌋ is the floor function), then we can ensure that the original message c can be recovered. The quantity ⌊(δH − 1)/2⌋ happens to be what is known as the packing radius of the code, defined, in general, as R P (C) = max{r ≥ 0 : BP (c, r) ∩ BP (c′, r) = ∅ ∀c, c′ ∈ C, c , c′ }, where BP (u, r) = {v ∈ Fnq : dP (u, v) ≤ r } is the P-ball with center u and radius r. When considering a general poset metric, the packing radius maintains the same property of evaluating the correction capability. However, in general, the minimum distance does not determine the packing radius, only bounds for it (see [6]). Indeed, hierarchical posets are the unique posets for which the packing radius of a code is a function of its minimum distance, as can be seen in [14]. Determining the packing radius, in general settings, is a very difficult task, it may be an NP-hard problem even considering the case of a code with only two elements (see [6]). For this reason, working with a general poset, the best we can do is to find bounds and we do it considering maximal P-decompositions of the code. We start with two simple propositions. Proposition 8: Let C = (C ′; C0; Ci )ri=1 be a P-decomposition for C. Then, R P (C) ≤ min R P (Ci ). i∈{1,...,r} Proof: Note that R P (C) = R P (C ′). Furthermore, since each Ci is a subcode of C ′, it follows that R P (C) ≤ R P (Ci ) for every i ∈ {1, . . . , r }. Proposition 9: If P ≤ Q, then R P (C) ≤ RQ (C) for every linear code C. Proof: If follows directly from the fact that BQ (r, c) ∩ BQ (r, 0) ⊂ BP (r, c) ∩ BP (r, 0) for every c ∈ C and any integer r ≥ 0. Using Proposition 9 and the upper and lower neighbours P + and P − defined in the previous section, bounds for the packing radius of codes according to hierarchical posets may be obtained. The proof of the next proposition follows straight from the previous proposition. Proposition 10: Given a maximal P-decomposition C = (C ′; C0; Ci )ri=1 for C, R P− (Ci ) ≤ R P (Ci ) ≤ R P+ (Ci ) for every i ∈ {1, . . ., r }. DRAFT March 1, 2018 23 Propositions 8 and 10 yield the following upper and lower bounds for the packing radius: R P− (C) ≤ R P (C) ≤ min R P+ (Ci ) i∈{1,...,r} (6) for every i ∈ {1, . . ., r }. Since P + and P − are hierarchical, these bounds are obtained just by finding the minimum distance of C and each code Ci according to P − and P + , respectively. Furthermore, each of these codes may be decomposed according to the lower and upper neighbors in order to simplify and optimize the search for the minimum distance. If P is hierarchical, we obtain equalities in the bounds obtained in Proposition 10 and in Inequality (6), hence, these bounds are tight. B. Optimizing the Full Lookup Table Searching in Syndrome Decoding The core of coding theory is the existence of errors: the messages to be sent are elements of a code C ⊂ Fnq , but, in the occurrence of errors, any element of Fnq may be received. Once y ∈ Fnq arrives, the receiver needs to decide what was the original message. There are essentially to types of decision criteria: probabilistic and deterministic. We are interested in the deterministic criteria determined by metrics, the well-known minimum distance decoding. Minimum distance decoding works as follows: once y ∈ Fnq is received, we look for one of the codewords that minimizes the distance to y. We are in a good situation if the minimum is unique and this is the case if the weight of the error was not greater than the packing radius. Despite the fact that the minimum distance decoding is very simple to describe as a criterion, its implementation as an algorithm is very difficult, since codes (for practical purposes) are very large and searching on very large sets is infeasible. Syndrome decoding is an algorithm to perform minimum distance decoding that works for every linear code and every metric that is determined by a weight or, equivalently, a metric that is invariant by translations. Given a code C ⊂ Fnq , for each coset u + C, consider an element with minimal weight, called coset leader. In syndrome decoding, once a message y is received, instead looking for c ∈ C that minimizes dP (y, c), one looks for a coset leader u such that T(u − y) = 0, where T : Fnq → Fn−k is a linear map that has C as its kernel (determined by q the parity check matrix, in coding jargon). The advantage of using syndrome decoding is the exchange of a lookup table of size q k by one of size q n−k , what is smaller, in general, since coding theory looks for [n, k]-codes with k/n close to 1. This quantity, q n−k , is an algebraic invariant that does not depend on the metric. March 1, 2018 DRAFT 24 We do not explain here the details of syndrome decoding (it can be found in any book on coding theory, for example [11, Section 1.11]), but it is a very established process and all we need is to state that when we use the poset metric dP , we mean that the minimality of the coset leader refers to the P-weight wP . In this section we show how we can reduce the searching process in the lookup table of syndrome decoding, by considering P-decompositions of a code. Roughly speaking, the finest the decomposition, the smaller the search table. Let us do it in details. Let C = (C; C0; Ci )ri=1 be a maximal P-decomposition of an [n, k, δ]q code C. Initially, note that in order to perform decoding, we can ignore the component V0 (recall that Vi is the supportspace of Ci ) since we know that any codeword c = (c1, . . ., cn ) ∈ C should have c j = 0 for every j ∈ [n]C . Consider the projection π : Fnq −→ Fnq1 +···+nr (recall that ni = |supp(Ci )|) defined by π(x1, . . ., xn ) = (xi1, . . ., xis ) r V , the where {i1, . . ., is } = supp(C) and we assume i1 < · · · < is . The map π1,...,r = π|⊕i=1 i r V , is a bijection. Therefore, by pushing forward the metric in restriction of π to the space ⊕i=1 i the restriction, we obtain the metric dPπ in Fnq1 +···+nr defined by −1 −1 dPπ (x, y) := dP (πi,...,r (x), πi,...,r (y)) for every x, y ∈ Fnq1 +···+nr . The metric dPπ turns the restriction map π1...r into a linear isometry. Because C is a subspace of V1 ⊕ . . . ⊕ Vr , the proof of the proposition below follows straight from these observations. Proposition 11: The metric-decoding criteria of Fnq for C is equivalent to the metric-decoding criteria of Fnq1 +···+nr for π(C), i.e., dP (c′, y) = min dP (c, y) ⇐⇒ dPπ (π(c′), π(y)) = min dPπ (c, π(y)). c∈C c∈π(C) By Proposition 11, to perform syndrome decoding, instead of using a lookup table with Î |Fnq /C| = q n−k elements, we can reduce the number of cosets to |Fnq1 +···+nr /π(C)| = ri=1 q ni −ki elements. Note that n0 q × r Ö q ni −ki = q n−k . i=1 Therefore, P-decompositions having the maximum possible number of elements in the support of C0 are the best ones in order to perform syndrome decoding. DRAFT March 1, 2018 25 Besides this possible (and a-posteriori irrelevant) gain in the cardinality of the syndrome lookup table obtained considering aggregations, there is a more significant gain that can be obtained through the splitting operation. Proposition 12: If hsupp(Ci )iP ∩ hsupp(Cj )iP = ∅ for all i , j and i, j , 0, then given y ∈ Fnq , min dP (c, y) = c∈C r Õ i=1 min dPπi (πi (c), πi (y)). c∈Ci Proof: Note that if c ∈ C, then c = c1 + · · · + cr with ci ∈ Ci and for every yi ∈ Fnqi , supp(yi − ci ) ⊂ hsupp(Ci )i = supp(Vi ). Then, d(y, c) = d(y1, c1 ) + · · · + d(yr , cr ), where y = y1 + · · · + yr and yi ∈ Fnqi for all i ∈ [r]. Due to Proposition 12, if the ideals generated by each component are disjoint, syndrome decoding can be done independently in each component Ci . Therefore, the number of cosets Í can be reduced from q n−k elements to ri=1 q ni −ki elements, where the last one is the sum of the cosets in each quotient Fnqi /πi (Ci′). More generally, if r is a disjoint union of subsets Ii , i.e., [r] = I1 ⊔ . . . ⊔ Is , and hsupp(⊕i∈I j Ci )iP ∩ hsupp(⊕i∈Il Ci )iP = ∅, for every j , l, then decoding can be separately done in each projection of ⊕i∈I j Ci into FqN Í where N = i∈I j ni . Until now, in order to obtain the reductions, it was not necessary to change the syndrome decoding algorithm; we just performed it in a different (smaller) space. The hierarchical relation among elements of the poset will provide us a “quasi-independent” syndrome decoding algorithm that is performed by choosing first coordinates that hierarchically dominate others, therefore we will call this algorithm a Leveled Syndrome Decoding. Given two subsets I, J ⊂ [n], we say that I and J are hierarchically related if every element in I is smaller than every element in J. Suppose [r] is an ordered disjoint union of sets, [r] = I1 ⊔ . . . ⊔ Is , such that supp(⊕i∈I j Ci ) is hierarchically related with supp(⊕i∈I j+1 Ci ) for every j ∈ {1, . . . , s − 1}, so if i0 ∈ supp(⊕i∈I j0 Ci ), then i0 6 i for every i ∈ supp(⊕i∈Il Ci ) with March 1, 2018 DRAFT 26 l > j0 . By using the well-known syndrome decoding algorithm, the leveled syndrome decoding algorithm is as follows: Input: y = y1 + · · · + ys ∈ Fnq where yi ∈ ⊕ j∈Ii Vj For each i ∈ {1, . . ., s} do Í Decode π j∈Ii (yi ) ∈ Fq j ∈Ii nj π j ∈Ii using syndrome (dP ) outputting ci ∈ π j∈Ii (⊕ j∈Ii Cj ) −1 Output: c = π −1 j∈I1 (c1 ) + · · · + π j∈Is (cs ). Leveled Syndrome Decoding Algorithm 1. Furthermore, performing an ordered syndrome decoding from the higher level to the smaller ones, we can proceed as follows: Input: y = y1 + · · · + ys ∈ Fnq where yi ∈ ⊕ j∈Ii Vj For i = s to 1, do If π j∈Ii (yi ) ∈ ⊕ j∈Ii Cj do ci = π j∈Ii (yi ); else do Í Decode π j∈Ii (yi ) ∈ Fq j ∈Ii nj π j ∈Ii using syndrome (dP ) outputting ci ∈ π j∈Ii (⊕ j∈Ii Cj ); Go to Output; end if; end For; −1 −1 Output: c = π −1 j∈Ii (ci ) + π j∈Ii+1 (ci+1 ) + · · · + π j∈Is (cs ). Leveled Syndrome Decoding Algorithm 2. The first algorithm was described in [7] for the hierarchical poset case. The second one is also a minimum distance algorithm according to the poset metric dP (see the next proposition), but if there is an error in a particular level, the decoder outputs the null vector in the levels covered by the one where the error happened. Proposition 13: Algorithm 2 determines a minimum distance decoder according to the metric dP . Proof: Given y ∈ Fnq and c ∈ C, then π j ∈Ii dP (y, c) = dP DRAFT (π j∈Ii (y), π j∈Ii (c)) March 1, 2018 27 where i is the largest integer such that π j∈Ii (y) , π j∈Ii (c). Therefore, if c′ ∈ C satisfies dP (y, c′) = min dP (y, c), c∈C ′ then c′′ = cs′ + cs−1 + · · · + ci′ also attains the minimum and this is the codeword returned by algorithm 2. Both the algorithms demand the storage of the lookup tables of each quotient FqN /π j∈Ii (⊕ j∈Ii Cj ) Í where j∈Ii n j . The total number of elements we need to store is s Ö Õ q n j −k j . (7) i=1 j∈Ii The difference between Algorithm 1 and 2 is that while the search in the Algorithm 1 is performed over all elements of the lookup table, the ordered procedure in Algorithm 2 stop the search in the first level where an error has occurred, i.e., if this happens in the level i0 , then the search is performed only over the lookup table of π j∈Ii0 (⊕ j∈Ii0 Cj ) which has Ö q n j −k j (8) j∈Ii0 elements. Note that if the poset has only one level, the values of Expressions (7) and (8) coincide and are equal to q n−k . However, the more the support of a P-decomposition of a code is partitioned into hierarchically related subsets, the more we may transform factors in the product (7) into parcels in expression (8). ACKNOWLEDGMENT The authors would like to thank the São Paulo Research Foundation (Fapesp) for the financial support through grants 2013/25977-7 and 2016/01551-9. The second author would like to thank also CNPq, for the support through grant 303985/2014-3. R EFERENCES [1] Marcelo Muniz S. Alves. A standard form for generator matrices with respect to the niederreiter-rosenbloom-tsfasman metric. In Information Theory Workshop (ITW), pages 486–489. IEEE, 2011. [2] Marcelo Muniz S. Alves, Luciano Panek, and Marcelo Firer. Error-block codes and poset metrics. Advances in Mathematics of Communications, 2(1):95–111, 2008. [3] Aleams Barra and Heide Gluesing-Luerssen. MacWilliams extension theorems and the local-global property for codes over Frobenius rings. Journal of Pure and Applied Algebra, 219(4):703 – 728, 2015. [4] Richard A Brualdi, Janine Smolin Graves, and K Mark Lawrence. Codes with a poset metric. Discrete Mathematics, 147(1):57–72, 1995. March 1, 2018 DRAFT 28 [5] Michel Marie Deza and Elena Deza. Encyclopedia of distances. Springer, Berlin, 2009. [6] Rafael Gregorio Lucas D’Oliveira and Marcelo Firer. The packing radius of a code and partitioning problems: The case for poset metrics on finite vector spaces. Discrete Mathematics, 338(12):2143–2167, 2015. [7] Luciano Viana Felix and Marcelo Firer. Canonical-systematic form for codes in hierarchical poset metrics. Advances in Mathematics of Communications, 6(3):315–328, 2012. [8] Ernst Gabidulin. A brief survey of metrics in coding theory. Mathematics of Distances and Applications, pages 66–84, 2012. [9] S. Golomb. A general formulation of error metrics (corresp.). Information Theory, IEEE Transactions on, 15(3):425–426, 1969. [10] Richard W. Hamming. Error detecting and error correcting codes. Bell System technical journal, 29(2):147–160, 1950. [11] W. Cary Huffman and Vera Pless. Fundamentals of error-correcting codes. Cambridge university press, 2003. [12] H.K. Kim and Dong Yeol Oh. A classification of posets admitting the MacWilliams identity. IEEE Transactions on Information Theory, 51(4):1424–1431, April 2005. [13] C. Y. Lee. Some properties of nonbinary error-correcting codes. IRE Transactions on Information Theory, 4(2):77–82, 1958. [14] R. Machado, J. Pinheiro, and M. Firer. Characterization of metrics induced by hierarchical posets. IEEE Transactions on Information Theory, PP(99):1–1, 2017. [15] Harald Niederreiter. A combinatorial problem for vector spaces over finite fields. Discrete Mathematics, 96(3):221–228, 1991. [16] Harald Niederreiter. Orthogonal arrays and other combinatorial aspects in the theory of uniform point distributions in unit cubes. Discrete mathematics, 106:361–367, 1992. [17] Dong Yeol Oh. Poset metrics admitting association schemes and a new proof of Macwilliams identity. Journal of the Korean Mathematical Society, 50(5):917–931, 2013. [18] Luciano Panek, Marcelo Firer, and Marcelo Muniz Silva Alves. Classification of niederreiter–rosenbloom–tsfasman block codes. Information Theory, IEEE Transactions on, 56(10):5207–5216, 2010. [19] Luciano Panek, Marcelo Firer, Hyun Kwang Kim, and Jong Yoon Hyun. Groups of linear isometries on poset structures. Discrete Mathematics, 308(18):4116–4123, 2008. [20] Claudio Qureshi and Sueli IR Costa. On perfect q-ary codes in the maximum metric. Information Theory and Applications, 2016. [21] Werner Ulrich. Non-binary error correction codes. Bell System Technical Journal, 46(6):1341–1388, 1957. DRAFT March 1, 2018
7cs.IT
IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 1 AffectNet: A Database for Facial Expression, Valence, and Arousal Computing in the Wild arXiv:1708.03985v4 [cs.CV] 9 Oct 2017 Ali Mollahosseini, Student Member, IEEE, Behzad Hasani, Student Member, IEEE, and Mohammad H. Mahoor, Senior Member, IEEE Abstract—Automated affective computing in the wild setting is a challenging problem in computer vision. Existing annotated databases of facial expressions in the wild are small and mostly cover discrete emotions (aka the categorical model). There are very limited annotated facial databases for affective computing in the continuous dimensional model (e.g., valence and arousal). To meet this need, we collected, annotated, and prepared for public distribution a new database of facial emotions in the wild (called AffectNet). AffectNet contains more than 1,000,000 facial images from the Internet by querying three major search engines using 1250 emotion related keywords in six different languages. About half of the retrieved images were manually annotated for the presence of seven discrete facial expressions and the intensity of valence and arousal. AffectNet is by far the largest database of facial expression, valence, and arousal in the wild enabling research in automated facial expression recognition in two different emotion models. Two baseline deep neural networks are used to classify images in the categorical model and predict the intensity of valence and arousal. Various evaluation metrics show that our deep neural network baselines can perform better than conventional machine learning methods and off-the-shelf facial expression recognition systems. Index Terms—Affective computing in the wild, facial expressions, continuous dimensional space, valence, arousal. F 1 I NTRODUCTION A FFECT is a psychological term used to describe the outward expression of emotion and feelings. Affective computing seeks to develop systems and devices that can recognize, interpret, and simulate human affects through various channels such as face, voice, and biological signals [1]. Face and facial expressions are undoubtedly one of the most important nonverbal channels used by the human being to convey internal emotion. There have been tremendous efforts to develop reliable automated Facial Expression Recognition (FER) systems for use in affect-aware machines and devices. Such systems can understand human emotion and interact with users more naturally. However, current systems have yet to reach the full emotional and social capabilities necessary for building rich and robust Human Machine Interaction (HMI). This is mainly due to the fact that HMI systems need to interact with humans in an uncontrolled environment (aka wild setting) where the scene lighting, camera view, image resolution, background, users head pose, gender, and ethnicity can vary significantly. More importantly, the data that drives the development of affective computing systems and particularly FER systems lack sufficient variations and annotated samples that can be used in building such systems. There are several models in the literature to quantify affective facial behaviors: 1) categorical model, where the emotion/affect is chosen from a list of affective-related categories such as six basic emotions defined by Ekman et al. [2], 2) dimensional model, where a value is chosen over a continuous emotional scale, such as valence and arousal [3] and 3) Facial Action Coding System (FACS) model, where • Authors are with the Department of Electrical and Computer Engineering, University of Denver, Denver, CO, 80210. E-mail: amollah, bhasani, [email protected] all possible facial actions are described in terms of Action Units (AUs) [4]. FACS model explains facial movements and does not describe the affective state directly. There are several methods to convert AUs to affect space (e.g., EMFACS [5] states that the occurrence of AU6 and AU12 is a sign of happiness). In the categorical model, mixed emotions cannot adequately be transcribed into a limited set of words. Some researchers tried to define multiple distinct compound emotion categories (e.g., happily surprised, sadly fearful) [6] to overcome this limitation. However, still the set is limited, and the intensity of the emotion cannot be defined in the categorical model. In contrast, the dimensional model of affect can distinguish between subtly different displays of affect and encode small changes in the intensity of each emotion on a continuous scale, such as valence and arousal. Valence refers to how positive or negative an event is, and arousal reflects whether an event is exciting/agitating or calm/soothing [3]. Figure 1 shows samples of facial expressions represented in the 2D space of valence and arousal. As it is shown, there are several different kinds of affect and small changes in the same emotion that cannot be easily mapped into a limited set of terms existing in the categorical model. The dimensional model of affect covers both intensity and different emotion categories in the continuous domain. Nevertheless, there are relatively fewer studies on developing automated algorithms in measuring affect using the continuous dimensional model (e.g., valence and arousal). One of the main reasons is that creating a large database to cover the entire continuous space of valence and arousal is expensive and there are very limited annotated face databases in the continuous domain. This paper contributes to the field of affective computing by providing a large annotated face database of the dimensional as well as the IEEE TRANSACTIONS ON AFFECTIVE COMPUTING Fig. 1. Sample images in Valence Arousal circumplex categorical models of affect. The majority of the techniques for automated affective computing and FER are based on supervised machine learning methodologies. These systems require annotated image samples for training. Researchers have created databases of human actors/subjects portraying basic emotions [7], [8], [9], [10], [11]. Most of these databases mainly contain posed expressions acquired in a controlled lab environment. However, studies show that posed expressions can be different from unposed expressions in configuration, intensity, and timing [12], [13]. Some researchers captured unposed facial behavior while the subject is watching a short video [14], [15], engaged in laboratory-based emotion inducing tasks [16], or interacted with a computer-mediated tutoring system [17]. Although a large number of frames can be obtained by these approaches, the diversity of these databases is limited due to the number of subjects, head position, and environmental conditions. Recently, databases of facial expression and affect in the wild received much attention. These databases are either captured from movies or the Internet, and annotated with categorical model [18], [19], [20], dimensional model [21], and FACS model [22]. However, they only cover one model of affect, have a limited number of subjects, or contain few samples of certain emotions such as disgust. Therefore, a large database, with a large amount of subject variations in the wild condition that covers multiple models of affect (especially the dimensional model) is a need. To address this need, we created a database of facial Affect from the InterNet (called AffectNet) by querying different search engines (Google, Bing, and Yahoo) using 1250 emotion related tags in six different languages (English, Spanish, Portuguese, German, Arabic, and Farsi). AffectNet contains more than one million images with faces and extracted facial landmark points. Twelve human experts manually annotated 450,000 of these images in both categorical 2 and dimensional (valence and arousal) models and tagged the images that have any occlusion on the face. Figure 1 shows sample images from AffectNet and their valence and arousal annotations. To calculate the agreement level between the human labelers, 36,000 images were annotated by two human labelers. AffectNet is by far the largest database of facial affect in still images which covers both categorical and dimensional models. The cropped region of the facial images, the facial landmark points, and the affect labels will be publicly available to the research community1 . Considering the lack of inthe-wild large facial expressions datasets and more specifically annotated face datasets in the continuous domain of valence and arousal, AffectNet is a great resource which will enable further progress in developing automated methods for facial behavior computing in both the categorical and continuous dimensional spaces. The rest of this paper is organized as follows. Section 2 reviews the existing databases and state-of-the-art methods for facial expression recognition with emphasis on the dimensional model and in the wild setting databases. Section 3 explains the process of collecting AffectNet images from the Internet and annotating the categorical and dimensional models. Section 4 presents two different baselines for automatic recognition of categorical emotions and prediction of dimensional valence and arousal in the continuous space using AffecNet images. Finally Section 5 concludes the paper. 2 2.1 R ELATED W ORK Existing databases Early databases of facial expressions such as JAFFE [7], Cohn-Kanade [8], [9], MMI [10], and MultiPie [11] were captured in a lab-controlled environment where the subjects portrayed different facial expressions. This approach resulted in a clean and high-quality database of posed facial expressions. However, posed expressions may differ from daily life unposed (aka spontaneous) facial expressions. Thus, capturing spontaneous expression became a trend in the affective computing community. Examples of these environments are recording the responses of participants’ faces while watching a stimuli (e.g., DISFA [14], AM-FED [15]) or performing laboratory-based emotion inducing tasks (e.g., Belfast [16]). These databases often capture multi-modal affects such as voice, biological signals, etc. and usually a series of frames are captured that enable researchers to work on temporal and dynamic aspects of expressions. However, the diversity of these databases is limited due to the number of subjects, head pose variation, and environmental conditions. Hence there is a demand to develop systems that are based on natural, unposed facial expressions. To address this demand, recently researchers paid attention to databases in the wild. Dhall et al. [18] released Acted Facial Expressions in the Wild (AFEW) from 54 movies by a recommender system based on subtitles. The video clips were annotated with six basic expressions plus neutral. AFEW 1. Interested researcher can download a copy of AffectNet from: http: //mohammadmahoor.com/databases-codes/ IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 3 TABLE 1 The Summary and Characteristics of Reviewed Databases in Affect Recognition Database Database information # of Subjects CK+ [9] - Frontal and 30 degree images - 123 MultiPie [11] - Around 750,000 images - Under multiple viewpoints and illuminations - 337 MMI [10] - Subjects portrayed 79 series of facial expressions - Image sequence of frontal and side view are captured - 25 DISFA [14] SALDB [23], [24] - Video of subjects while watching a four minutes video - Clip are recorded by a stereo camera - SAL - Audiovisual (facial expression,shoulder, audiocues) - 20 facial feature points, 5 shoulder points for video - 27 Condition - Controlled - Posed - Controlled - Posed - Controlled - Posed & Spontaneous - Controlled - Spontaneous -4 - Controlled - Spontaneous RELOCA [25] - Multi-modal audio, video, ECG and EDA - 46 - Controlled - Spontaneous AM-FED [15] - 242 facial videos - 242 - Spontaneous DEAP [26] - 40 one-minute long videos shown to subjects - EEG signals recorded - 32 - Controlled - Spontaneous - Videos - Images queried from web - Images queried from web - 100,000 images annotated manually - 900,000 images annotated automatically - 330 - ∼35,887 - Wild - Wild - ∼100,000 - Wild Aff-Wild [21] - 500 videos from YouTube - 500 - Wild FER-Wild [20] - 24,000 images from web - ∼24,000 - Wild AffectNet (This work) - 1,000,000 images with facial landmarks - 450,000 images annotated manually - ∼450,000 - Wild AFEW [18] FER-2013 [19] EmotioNet [22] contains 330 subjects aged 1-77 years and addresses the issue of temporal facial expressions in the wild. A static subset (SFEW [27]) is created by selecting some frames of AFEW. SFEW covers unconstrained facial expressions, different head poses, age range, occlusions, and close to real world illuminations. However, it contains only 700 images, and there are only 95 subjects in the database. The Facial Expression Recognition 2013 (FER-2013) database was introduced in the ICML 2013 Challenges in Representation Learning [19]. The database was created using the Google image search API that matched a set of 184 emotion-related keywords to capture the six basic expressions as well as the neutral expression. Images were resized to 48x48 pixels and converted to grayscale. Human labelers rejected incorrectly labeled images, corrected the cropping if necessary, and filtered out some duplicate images. The resulting database contains 35,887 images most of which are in the wild settings. FER-2013 is currently the biggest publicly available facial expression database in the wild settings, enabling many researchers to train machine learning methods such as Deep Neural Networks (DNNs) where large amounts of data are needed. In FER-2013, the faces are not registered, a small number of images portray disgust (547 images), and unfortunately most of facial landmark detectors fail to extract facial landmarks at this resolution and quality. In addition, only the categorical model of affect is provided with FER-2013. The Affectiva-MIT Facial Expression Dataset (AM-FED) database [15] contains 242 facial videos (160K frames) of people watching Super Bowl commercials using their webcam. The recording conditions were arbitrary with differ- Affect Modeling - 30 AUs - 7 emotion categories - 7 emotion categories - 31 AUs - Six basic expression - 12 AUs - Valence - Quantized [23] - Continuous [24] - Valence and arousal (continuous) - Self assessment - 14 AUs - Valence and arousal (continuous) - Self assessment - 7 emotion categories - 7 emotion categories - 12 AUs annotated - 23 emotion categories based on AUs - Valence and arousal (continuous) - 7 emotion categories - 8 emotion categories - Valence and arousal (continuous) ent illumination and contrast. The database was annotated frame-by-frame for the presence of 14 FACS action units, head movements, and automatically detected landmark points. AM-FED is a great resource to learn AUs in the wild. However, there is not a huge variance in head pose (limited profiles), and there are only a few subjects in the database. The FER-Wild [20] database contains 24,000 images that are obtained by querying emotion-related terms from three search engines. The OpenCV face recognition was used to detect faces in the images, and 66 landmark points were found using Active Appearance Model (AAM) [28] and a face alignment algorithm via regression local binary features [29], [30]. Two human labelers annotated the images into six basic expressions and neutral. Comparing with FER2013, FER-Wild images have a higher resolution with facial landmark points necessary to register the images. However, still a few samples portray some expressions such as disgust and fear and only the categorical model of affect is provided with FER-Wild. The EmotioNet [22] consists of one million images of facial expressions downloaded from the Internet by selecting all the words derived from the word “feeling” in WordNet [31]. Face detector [32] was used to detect faces in these images and the authors visually inspected the resultant images. These images were then automatically annotated with AUs and AU intensities by an approach based on Kernel Subclass Discriminant Analysis (KSDA) [33]. The KSDA-based approach was trained with Gabor features centered on facial landmark with a Radial Basis Function (RBF) kernel. Images were labeled as one of the 23 (basic or compound) emotion categories defined in [6] based on IEEE TRANSACTIONS ON AFFECTIVE COMPUTING AUs. For example, if an image has been annotated as having AUs 1, 2, 12 and 25, it is labeled as happily surprised. A total of 100,000 images (10% of the database) were manually annotated with AUs by experienced coders. The proposed AU detection approach was trained on CK+ [9], DISFA [14], and CFEE [34] databases, and the accuracy of the automated annotated AUs was reported about 80% on the manually annotated set. EmotioNet is a novel resource of FACS model in the wild with a large amount of subject variation. However, it lacks the dimensional model of affect, and the emotion categories are defined based on annotated AUs and not manually labeled. On the other hand, some researchers developed databases of the dimensional model in the continuous domain. These databases, however, are limited since the annotation of continuous dimensions is more expensive and necessitate trained annotators. Examples of these databases are Belfast [16], RECOLA [25], Affectiva-MIT Facial Expression Dataset (AM-FED) [15], and recently published Aff-Wild Database [21] which is the only database of dimensional model in the wild. The Belfast database [16] contains recordings (5s to 60s in length) of mild to moderate emotional responses of 60 participants to a series of laboratory-based emotion inducing tasks (e.g., surprise response by setting off a loud noise when the participant is asked to find something in a black box). The recordings were labeled by information on self-report of emotion, the gender of the participant/experimenter, and the valence in the continuous domain. The arousal dimension was not annotated in Belfast database. While the portrayed emotions are natural and spontaneous, the tasks have taken place in a relatively artificial setting of a laboratory where there was a control on lighting conditions, head poses, etc. The Database for Emotion Analysis using Physiological Signals (DEAP) [26] consists of spontaneous reactions of 32 participants in response to one-minute long music video clip. The EEG, peripheral physiological signals, and frontal face videos of participants were recorded, and the participants rated each video in terms of valence, arousal, like/dislike, dominance, and familiarity. Correlations between the EEG signal frequencies and the participants ratings were investigated, and three different modalities, i.e., EEG signals, peripheral physiological signals, and multimedia features on video clips (such as lighting key, color variance, etc.) were used for binary classification of low/high arousal, valence, and liking. DEAP is a great database to study the relation of biological signals and dimensional affect, however, it has only a few subjects and the videos are captured in lab controlled settings. The RECOLA benchmark [25] contains videos of 23 dyadic teams (46 participants) that participated in a video conference completing a task which required collaboration. Different multi-modal data of the first five minutes of interaction, i.e., audio, video, ECG and EDA) were recorded continuously and synchronously. Six annotators measured arousal and valence. The participants reported their arousal and valence through the Self-Assessment Manikin (SAM) [35] questionnaire before and after the task. RECOLA is a great database of the dimensional model with multiple cues and modalities, however, it contains only 46 4 subjects and the videos were captured in the lab controlled settings. Audio-Visual Emotion recognition Challenge (AVEC) series of competitions [36], [37], [38], [39], [40], [41] provided a benchmark of automatic audio, video and audiovisual emotion analysis in continuous affect recognition. AVEC 2011, 2012, 2013, and 2014 used videos from the SEMAINE [42] database videos. Each video is annotated by a single rater for every dimension using a two-axis joystick. AVEC 2015 and 2016 used the RECOLA benchmark in their competitions. Various continuous affect recognition dimensions were explored in each challenge year such as valence, arousal, expectation, power, and dominance, where the prediction of valence and arousal are studied in all challenges. The Aff-Wild Database [21] is by far the largest database for measuring continuous affect in the valence-arousal space “in-the-wild”. More than 500 videos from YouTube were collected. Subjects in the videos displayed a number of spontaneous emotions while watching a particular video, performing an activity, and reacting to a practical joke. The videos have been annotated frame-by-frame by three human raters, utilizing a joystick-based tool to rate valence and arousal. Aff-Wild is a great database of dimensional modeling in the wild that considers the temporal changes of the affect, however, it has a small subject variance, i.e., it only contains 500 subjects. Table 1 summarizes the characteristics of the reviewed databases in all three models of affect, i.e., categorical model, dimensional model, and Facial Action Coding System (FACS). 2.2 Evaluation Metrics There are various evaluation metrics in the literature to measure the reliability of annotation and automated affective computing systems. Accuracy, F1-score [49], Cohens kappa [50], Krippendorfs Alpha [51], ICC [52], area under the ROC curve (AUC), and area under Precision-Recall curve (AUC-PR) [53] are well-defined widely used metrics for evaluation of the categorical and FACS-based models. Since, the dimensional model of affect is usually evaluated in a continuous domain, different evaluation metrics are necessary. In the following, we review several metrics that are used in the literature for evaluation of dimensional model. Root Mean Square Error (RMSE) is the most common evaluation metric in a continuous domain which is defined as: v u n u1 X RM SE = t (θ̂i − θi )2 (1) n i=1 where θ̂i and θi are the prediction and the ground truth of ith sample, and n is the number of samples in the evaluation set. RMSE-based evaluation can heavily weigh the outliers [54], and it is not able to provide the covariance of prediction and ground-truth to show how they change with respect to each other. Pearsons correlation coefficient IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 5 TABLE 2 State-of-the-art Algorithms and Their Performance on the Databases Listed in Table 1. Work Mollahosseini et al. [43] Database CK+ MultiPie Shan et al. [44] MMI Zhang et al. [45] DISFA Nicolaou et al. [24] SALDB He et al. [46] RECOLA McDuff et al. [15] AM-FED Koelstra et al. [26] DEAP Fan et al. [47] AFEW Tang et al. [48] FER-2013 Benitez-Quiroz et al. [22] EmotioNet Mollahosseini et al. [20] FER-Wild Method - Inception based Convolutional Neural Network (CNN) - Subject-independent and cross-database experiments - Different SVM kernels trained with LBP features - Subject-independent and cross-database experiments - lp norm multi-task multiple kernel learning - learning shared kernels from a given set of base kernels - Bidirectional LSTM - Trained on multiple engineered features extracted from audio, facial geometry , and shoulder - Multiple stack of bidirectional LSTM (DBLSTM-RNN) - Trained on engineered features extracted from audio (LLDs), video (LPQ-TOP), 52 ECG features, and 22 EDA features - HOG features extracted - SVM with RBF kernel - Gaussian naive Bayes classifier - EEG, physiological signals, and multimedia features - Binary classification of low/high arousal, valence, and liking - Trained on both video and audio. - VGG network are followed by LSTMs and combined with 3D convolution - CNN with linear one-vs-all SVM at the top - New face feature extraction method using Gabor filters - KSDA classification - Subject-independent and cross-database experiments - Trained on AlexNet - Noise estimation methods used is therefore proposed in some literature [24], [36], [37] to overcome this limitation: E[(θ̂ − µθ̂ )(θ − µθ )] COV {θ̂, θ} = (2) σθ̂ σθ σθ̂ σθ Concordance Correlation Coefficient (CCC) is another metric [40], [41] which combines the Pearsons correlation coefficient (CC) with the square difference between the means of two compared time series: 2ρσθ̂ σθ (3) ρc = 2 2 σθ̂ + σθ + (µθ̂ − µθ )2 CC = where ρ is the Pearson correlation coefficient (CC) between two time-series (e.g., prediction and ground-truth), σθ̂2 and σθ2 are the variance of each time series, and µθ̂ and µθ are the mean value of each. Unlike CC, the predictions that are well correlated with the ground-truth but shifted in value are penalized in proportion to the deviation in CCC. The value of valence and arousal are [-1,+1] and their signs are essential in many emotion-prediction applications. For example, if the ground-truth valence is +0.3, prediction of +0.7 is far better than prediction of -0.1, since +0.7 indicates a positive emotion similar to the ground-truth (despite both predictions have the same RMSE). Sign Agreement Metric (SAGR) is another metric that is proposed in [24] to evaluate the performance of a valence and arousal prediction system. SAGR is defined as: n SAGR = 1X δ(sign(θ̂i ), sign(θi )) n i=1 where δ is the Kronecker delta function, defined as: ( 1, a = b δ(a, b) = 0, a 6= b (4) (5) Results - 93.2% accuracy on CK+ - 94.7% accuracy on MultiPie - 86.9% accuracy on MMI - 0.70 F1-score on DISFA - 0.93 recognition rate on DISFA - Leave-one-sequence-out - BLSTM-NN outperform SVR - Valence (RMSE=0.15 and CC=0.796) - Arousal (RMSE=0.21 and CC=0.642) - Winner of AVEC 2015 challenge - Valence (RMSE=0.104 and CC=0.616) - Arousal (RMSE=0.121 and CC=0.753) - AUC 0.90, 0.72 and 0.70 for smile, AU2 and AU4 respectively - 0.39 F1-score on Arousal - 0.37 F1-score on Valence - 0.40 F1-score on Liking - Winner of EmotiW 2016 challenge - 56.16% accuracy on AFEW - Winner of the FER challenge - 71.2% accuracy on test set - ∼80% AU detection on EmotioNet - 82.12% accuracy on FER-Wild The above discussed metrics are used to evaluate the categorical and dimensional baselines on AffectNet in Sec. 4. 2.3 Existing Algorithms Affective computing is now a well-established field, and there are many algorithms and databases for developing automated affect perception systems. Since it is not possible to include all those great works, we only give a brief overview and cover the state-of-the-art methods that are applied on the databases explained in Sec. 2.1. Conventional algorithms of affective computing from faces use hand-crafted features such as pixel intensities [55], Gabor filters [56], Local Binary Patterns (LBP) [44], and Histogram of Oriented Gradients (HOG) [14]. These handcrafted features often lack enough generalizability in the wild settings where there is a high variation in scene lighting, camera view, image resolution, background, subjects head pose and ethnicity. An alternative approach is to use Deep Neural Networks (DNN) to learn the most appropriate feature abstractions directly from the data and handle the limitations of handcrafted features. DNNs have been a recent successful approach in visual object recognition [57], human pose estimation [58], face verification [59] and many more. This success is mainly due to the availability of computing power and existing big databases that allow DNNs to extract highly discriminative features from the data samples. There have been enormous attempts on using DNNs in automated facial expression recognition and affective computing [20], [43], [46], [47], [48] that are especially very successful in the wild settings. Table 2 shows a list of the state-of-the-art algorithms and their performance on the databases listed in Table 1. IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 6 Fig. 2. A screen-shot of the software application used to annotate categorical and dimensional (valence and arousal) models of affect and the osculation tag if existing. Only one detected face in each image is annotated (shown in the green bounding box). As shown in the table, the majority of these approaches have used DNNs to learn a better representation of affect, especially in the wild settings. Even some of the approaches, such as the winner of the AVEC 2015 challenge [46], trained a DNN with hand-crafted features and still could improve the prediction accuracy. 3 A FFECT N ET AffectNet (Affect from the InterNet) is the largest database of the categorical and dimensional models of affect in the wild (as shown in Table 1). The database is created by querying emotion related keywords from three search engines and annotated by expert human labelers. In this section, the process of querying the Internet, processing facial images and extracting facial landmarks, and annotating facial expression, valence, and arousal of affect are discussed. 3.1 Facial Images from the Web Emotion-related keywords were combined with words related to gender, age, or ethnicity, to obtain nearly 362 strings in the English language such as “joyful girl”, “blissful Spanish man”, “furious young lady”, “astonished senior”. These keywords are then translated into five other languages: Spanish, Portuguese, German, Arabic and Farsi. The direct translation of queries in English to other languages did not accurately result in the intended emotions since each language and culture has differing words and expressions for different emotions. Therefore, the list of English queries was provided to native non-English speakers who were proficient in English, and they created a list of queries for each emotion in their native language and inspected the quality of the results visually. The criteria for highquality queries were those that returned a high percentage of human faces showing the intended queried emotions rather than drawings, graphics, or non-human objects. A total of 1250 search queries were compiled and used to crawl the search engines in our database. Since a high percentage of results returned by our query terms already contained neutral facial images, no individual query was performed to obtain additional neutral face. Three search engines (Google, Bing, and Yahoo) were queried with these 1250 emotion related tags. Other search engines such as Baidu and Yandex were considered. However, they either did not produce a large number of facial images with intended expressions or they did not have available APIs for automatically querying and pulling image URLs into the database. Additionally, queries were combined with negative terms (e.g., “drawing”, “cartoon”, “animation”, “birthday”, etc.) to avoid non-human objects as much as possible. Furthermore, since the images of stock photo websites are posed unnaturally and contain watermarks mostly, a list of popular stock photo websites was compiled and the results returned from the stock photo websites were filtered out. A total of ∼1,800,000 distinct URLs returned for each query were stored in the database. The OpenCV face recognition was used to obtain bounding boxes around each face. A face alignment algorithm via regression local binary features [29], [30] was used to extract 66 facial landmark points. The facial landmark localization technique was trained using the annotations provided from the 300W competition [60]. More than 1M images containing at least one face with extracted facial landmark points were kept for further processing. The average image resolution of faces in AffectNet are 425 × 425 with STD of 349 × 349 pixels. We used Microsoft cognitive face API to extract these facial attributes on 50,000 randomly selected images from the database. According to MS face API, 49% of the faces are men. The average estimated age of the faces is 33.01 years with the standard deviation of 16.96 years. In particular, 10.85, 3.9, 30.19, 26.86, 14.46, and 13.75 percent of the faces are in age ranges [0, 10), [10, 20), [20, 30), [30, 40), [40, 50) and [50, -), respectively. MS face API detected forehead, mouth, and eye occlusions in 4.5, 1.08, and 0.49 percent of the images, respectively. Also, 9.63% of the faces wear glasses, 51.07 and 41.4% of the faces have eye and lip make-ups, respectively. In terms of head IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 7 pose, the average estimated pitch, yaw, roll are 0.0,-0.7, and -1.19 degrees, respectively. Expression Neutral Happy Sad Surprise Fear Disgust Anger Contempt None Uncertain Non-Face Annotation Crowd-sourcing services like Amazon Mechanical Turk are fast, cheap and easy approaches for labeling large databases. The quality of labels obtained from crowd-sourcing services, however, varies considerably among the annotators. Due to these issues and the fact that annotating the valence and arousal requires a deep understanding of the concept, we avoided crowd-sourcing facilities and instead hired 12 fulltime and part-time annotators at the University of Denver to label the database. A total of 450,000 images were given to these expert annotators to label the face in the images into both discrete categorical and continuous dimensional (valence and arousal) models. Due to time and budget constraints each image was annotated by one annotator. A software application was developed to annotate the categorical and dimensional (valence and arousal) models of affect. Figure 2 shows a screen-shot of the annotation application. A comprehensive tutorial including the definition of the categorical and dimensional models of affect with some examples of each category, valence and arousal was given to the annotators. Three training sessions were provided to each annotator, in which the annotator labeled the emotion category, valence and arousal of 200 images and the results were reviewed with the annotators. Necessary feedback was given on both the categorical and dimensional labels. In addition, the annotators tagged the images that have any occlusion on the face. The occlusion criterion was defined as if any part of the face was not visible. If the person in the images wore glasses, but the eyes were visible without any shadow, it was not considered as occlusion. 3.2.1 Categorical Model Annotation Eleven discrete categories were defined in the categorical model of AffectNet as: Neutral, Happy, Sad, Surprise, Fear, Anger, Disgust, Contempt, None, Uncertain, and Non-face. The None (“None of the eight emotions”) category is the type of expression/emotions (such as sleepy, bored, tired, seducing, confuse, shame, focused, etc.) that could not be assigned by annotators to any of the six basic emotions, contempt or neutral. However, valence and arousal could be assigned to these images. The Non-face category was defined as images that: 1) Do not contain a face in the image; 2) Contain a watermark on the face; 3) The face detection algorithm fails and the bounding box is not around the face; 4) The face is a drawing, animation, or painted; and 5) The face is distorted beyond a natural or normal shape, even if an expression could be inferred. If the annotators were uncertain about any of the facial expressions, images were tagged as uncertain. When an image was annotated as Nonface or uncertain, valence and arousal were not assigned to the image. The annotators were instructed to select the proper expression category of the face, where the intensity is not important as long as the face depicts the intended emotion. Table 3 shows the number of images in each category. Table 4 indicates the percentage of annotated categories for queried emotion terms. As shown, the happy emotion Number 80,276 146,198 29,487 16,288 8,191 5,264 28,130 5,135 35,322 13,163 88,895 TABLE 4 Percentage of Annotated Categories for Queried Emotion Terms (%) Annotated Expression 3.2 TABLE 3 Number of Annotated Images in Each Category NE* HA SA SU FE DI AN CO NO UN NF HA 17.3 48.9 2.6 2.7 0.7 0.6 2.8 1.3 5.4 1.3 16.3 SA 16.3 27.2 15.7 3.1 1.2 0.7 4.5 0.9 8.7 3.1 18.6 Query Expression SU FE DI 13.9 17.8 17.8 30.4 28.6 33 4.8 5.8 4.5 16 4.4 3.6 4.2 4 1.5 0.7 0.9 2.7 3.8 5.6 6 0.4 1.1 1.1 4.8 8.1 8.8 4.3 3.1 4.1 16.7 20.6 16.9 AN 16.1 29.5 5.4 3.4 1.4 1.1 12.2 1.2 9.3 3.7 16.8 CO 20.1 30.1 4.6 4.1 1.3 1 6.1 2.4 11.2 2.7 16.3 * NE, HA, SA, SU, FE, DI, AN, CO, NO, UN , and NF stand for Neutral, Happy, Sad, Surprise, Fear, Anger, Disgust, Contempt, None, Uncertain, and Non-face categories, respectively. had the highest hit-rate (48%), and the rest of the emotions had hit-rates less than 20%. About 15% of all query results were in the No-Face category, as many images from the web contain watermarks, drawings, etc. About 15% of all queried emotions resulted in neutral faces. Among other expressions, disgust, fear, and contempt had the lowest hitrate with only 2.7%, 4%, and 2.4% hit-rates, respectively. As one can see, the majority of the returned images from the search engines were happy or neutral faces. The authors believe that this is because people tend to publish their images with positive expressions rather than negative expressions. Figure 3 shows a sample image in each category and its intended queries (in parentheses). 3.2.2 Dimensional (Valence & Arousal) Annotation The definition of valence and arousal dimensions was adapted from [3] and was given to annotators in our tutorial as: “Valence refers to how positive or negative an event is, and arousal reflects whether an event is exciting/agitating or calm/soothing”. A sample circumplex with estimated positions of several expressions, borrowed from [61], was provided in the tutorial as a reference for the annotators. The provided circumplex in the tutorial contained more than 34 complex emotions categories such as suspicious, insulted, impressed, etc., and used to train annotators. The annotators were instructed to consider the intensity of valence and arousal during the annotation. During the annotation process, the annotators were supervised closely and constant necessary feedback was provided when they were uncertain about some images. IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 8 1 10000 0.8 3162 0.6 1000 0.4 Neutral (Angry) Happy (Happy) Sad (Angry) Surprise (Fear) 0.2 316 0 100 −0.2 32 −0.4 10 −0.6 Fear (Fear) Disgust (Disgust) Angry (Angry) Contempt (Happy) 3 −0.8 −1 −1 1 −0.5 0 0.5 1 Fig. 4. Histogram (number of frames in each range/area) of valence and arousal annotations (Best viewed in color). Non-face (Surprise) Uncertain (Sad) None (Fear) None (Happy) TABLE 5 Annotators’ Agreement in Dimensional Model of Affect Fig. 3. Samples of queried images from the web and their annotated tags. The queried expression is written in parentheses. To model the dimensional affect of valence and arousal, a 2D Cartesian coordinate system was used where the x-axis and y -axis represent the valence and arousal, respectively. Similar to Russell’s circumplex space model [3], our annotation software did not allow the value of valence and arousal outside of the circumplex. This allows us to convert the Cartesian coordinates to polar coordinates with 0 ≤ r ≤ 1 and 0 ≤ θ < 360. The annotation software showed the value of valence and arousal to the annotators when they selected a point in the circumplex. This helped the annotators to pick more precise locations of valence and arousal with a higher confidence. A predefined estimated region of valence and arousal was defined for each categorical emotion in the annotation software (e.g., for happy emotion the valence is in (0.0, 1.0], and the arousal is in [-0.2, 0.5] ). If the annotators select a value of valence and arousal outside of the selected emotion’s region, the software indicates a warning message. The annotators were able to proceed, and they were instructed to do so, if they were confident about the value of valence and arousal. The images with the warning messages were marked in the database, for further review by the authors. This helped to avoid mistakes in the annotation of the dimensional model of affect. Figure 4 shows the histogram (number of samples in each range/area) of annotated images in a 2D Cartesian coordinate system. As illustrated, there are more samples in the center and the right middle (positive valence and small positive arousal) of the circumplex, which confirms the higher number of Neutral and Happy images in the database compared to other categories in the categorical model. 2 2. A numerical representation of annotated images in each range/area of valence and arousal is provided in the Appendix. RMSE CORR SAGR CCC 3.3 Same Category Valence Arousal 0.190 0.261 0.951 0.766 0.906 0.709 0.951 0.746 All Valence 0.340 0.823 0.815 0.821 Arousal 0.362 0.567 0.667 0.551 Annotation Agreement In order to measure the agreement between the annotators, 36,000 images were annotated by two annotators. The annotations were performed fully blind and independently, i.e., the annotators were not aware of the intended query or other annotator’s response. The results showed that the annotators agreed on 60.7% of the images. Table 6 shows the agreement between two annotators for different categories. As it is shown, the annotators highly agreed on the Happy and No Face categories, and the highest disagreement occurred in the None category. Visually inspecting some of the images in the None category, the authors believe that the images in this category contain very subtle emotions and they can be easily confused with other categories (the last two example of Fig. 3 show images in the None category). Table 5 shows various evaluation metrics between the two annotators in the continuous dimensional model of affect. These metrics are defined in Sec. 2.2. We calculated these metrics in two scenarios: 1) the annotators agreed on the category of the image; 2) on all images that are annotated by two annotators. As Table 5 shows, when the annotators agreed on the category of the image, the annotations have a high correlation and sign agreement (SAGR). According to Table 6, this occurred on only 60.7% images. However, there is less correlation and SAGR on overall images, since the annotators had a different perception of emotions expressed in the images. It can also be seen that the annotators agreed on valence more than arousal. The authors believe that this is because the perception of valence (how positive or negative the emotion is) is easier and less subjective than arousal (how excited or calm the subject is) especially in still images. Comparing the metrics in the existing dimensional IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 9 TABLE 6 Agreement Between Two Annotators in Categorical Model of Affect (%) Neutral Happy Sad Surprise Fear Disgust Anger Contempt None Uncertain Non-Face Neutral 50.8 6.3 11.8 2.0 3.1 1.5 8.1 10.2 22.6 13.5 3.7 Happy 7.0 79.6 0.9 3.8 1.5 0.8 1.2 7.5 12.0 12.1 3.8 Sad 9.1 0.6 69.7 1.6 3.8 3.6 7.5 2.1 14.5 7.8 1.7 Surprise 2.8 1.7 1.2 66.5 15.3 1.2 1.7 0.5 8.0 7.3 1.1 Fear 1.1 0.3 3.4 14.0 61.1 3.5 2.9 0.5 6.0 4.0 0.9 Disgust 1.0 0.4 1.3 0.8 2.5 67.6 4.4 4.4 2.3 4.5 0.4 databases (shown in Table 2) with the agreement of human labelers on AffectNet, suggest that AffectNet is a very challenging database and even human annotations have more RMSE than automated methods on existing databases. 4 BASELINE In this section, two baselines are proposed to classify images in the categorical model and predict the value of valence and arousal in the continuous domain of dimensional model. Since deep Convolutional Neural Networks (CNNs) have been a successful approach to learn appropriate feature abstractions directly from the image and there are many samples in AffectNet necessary to train CNNs, we proposed two simple CNN baselines for both categorical and dimensional models. We also compared the proposed baselines with conventional approaches (Support Vector Machines [62] and Support Vector Regressions [63]) learned from hand-crafted features (HOG). In the following sections, we first introduce our training, validation and test sets, and then show the performance of each proposed baselines. 4.1 Test, Validation, and Training Sets Test set: The subset of the annotated images that are annotated by two annotators is reserved for the test set. To determine the value of valence and arousal in the test set, since there are two responses for one image in the continuous domain, one of the annotations is picked randomly. To select the category of image in the categorical model, if there was a disagreement, a favor was given to the intended query, i.e., if one of the annotators labeled the image as the intended query, the image was labeled with the intended query in the test set. This happened in 29.5% of the images with disagreement between the annotators. On the rest of the images with disagreement, one of the annotations was assigned to the image randomly. Since the test set is a random sampling of all images, it is heavily imbalanced. In other words, there are more than 11,000 images with happy expression while it contains only 1,000 images with contemptuous expression. Validation set: Five hundred samples of each category is selected randomly as a validation set. The validation set is used for hyper-parameter tuning, and since it is balanced, there is no need for any skew normalization. Training set: The rest of images are considered as training examples. The training examples, as shown in Table 3, are heavily imbalanced. 4.2 Anger 4.8 0.5 4.0 1.9 7.2 13.1 62.3 2.1 16.9 6.2 1.7 Contempt 5.3 3.0 0.3 0.6 0.0 1.7 1.3 66.9 1.3 2.6 0.4 None 11.1 4.6 3.5 4.2 1.9 2.7 5.5 3.7 9.6 12.3 1.2 Uncertain 1.9 1.0 1.2 1.9 0.4 2.3 1.9 1.5 4.3 20.6 1.4 Non-Face 5.1 2.2 2.6 2.7 3.3 2.1 3.3 0.6 2.6 8.9 83.9 Categorical Model Baseline Facial expression data is usually highly skewed. This form of imbalance is commonly referred to as intrinsic variation, i.e., it is a direct result of the nature of expressions in the real world. This happens in both the categorical and dimensional models of affect. For instance, Caridakis et al. [64] reported that a bias toward quadrant 1 (positive arousal, positive valence) exists in the SAL database. The problem of learning from imbalanced data sets has two challenges. First, training data with an imbalanced distribution often causes learning algorithms to perform poorly on the minority class [65]. Second, the imbalance in the test/validation data distribution can affect the performance metrics dramatically. Jeni et al. [53] studied the influence of skew on imbalanced validation set. The study showed that with exception of area under the ROC curve (AUC), all other studied evaluation metrics, i.e., Accuracy, F1-score, Cohens kappa [50], Krippendorfs Alpha [51], and area under Precision-Recall curve (AUCPR) are affected by skewed distributions dramatically. While AUC is unaffected by skew, precision-recall curves suggested that AUC may mask poor performance. To avoid or minimize skew-biased estimates of performance, the study suggested to report both skew-normalized scores and the original evaluation. We used AlexNet [57] architecture as our deep CNN baseline. AlexNet consists of five convolution layers, followed by max-pooling and normalization layers, and three fully-connected layers. To train our baseline with an imbalanced training set, four approaches are studied in this paper as Imbalanced learning, Down-Sampling, Up-Sampling, and Weighted-Loss. The imbalanced learning approach was trained with the imbalanced training set without any change in the skew of the dataset. To train the down-sampling approach, we selected a maximum of 15,000 samples from each class. Since there are less than 15,000 samples for some classes such as Disgust, Contempt, and Fear, the resulting training set is semi-balanced. To train the up-sampling approach, we heavily up-sampled the under-represented classes by replicating their samples so that all classes had the same number of samples as the class with maximum samples, i.e., Happy class. The weighted-loss approach weighted the loss function for each of the classes by their relative proportion in the training dataset. In other words, the loss function heavily penalizes the networks for misclassifying examples from under-represented classes, while penalizing networks less IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 10 TABLE 7 F1-Scores of four different approaches of training AlexNet Neutral Happy Sad Surprise Fear Disgust Anger Contempt Imbalanced Top-1 Top-2 Orig* Norm* Orig Norm 0.63 0.49 0.82 0.66 0.88 0.65 0.95 0.80 0.63 0.60 0.84 0.81 0.61 0.64 0.84 0.86 0.52 0.54 0.78 0.79 0.52 0.55 0.76 0.78 0.65 0.59 0.83 0.80 0.08 0.08 0.49 0.49 Down-Sampling Top-1 Top-2 Orig Norm Orig Norm 0.58 0.49 0.78 0.70 0.85 0.68 0.92 0.85 0.64 0.60 0.81 0.78 0.53 0.63 0.75 0.83 0.54 0.57 0.80 0.82 0.53 0.64 0.74 0.81 0.62 0.60 0.79 0.78 0.22 0.32 0.60 0.70 Up-Sampling Top-1 Top-2 Orig Norm Orig Norm 0.61 0.50 0.81 0.64 0.85 0.71 0.95 0.80 0.6 0.57 0.81 0.77 0.57 0.66 0.80 0.81 0.56 0.58 0.75 0.76 0.53 0.59 0.70 0.72 0.63 0.59 0.81 0.77 0.15 0.18 0.42 0.42 Weighted-Loss Top-1 Top-2 Orig Norm Orig Norm 0.57 0.52 0.81 0.77 0.82 0.73 0.92 0.88 0.63 0.61 0.83 0.81 0.51 0.63 0.77 0.86 0.56 0.66 0.79 0.86 0.48 0.66 0.69 0.83 0.60 0.60 0.81 0.81 0.27 0.59 0.58 0.79 *Orig and Norm stand for Original and skew-Normalized, respectively. for misclassifying examples from well-represented classes. The entropy loss formulation for a training example (X, l) is defined as: E=− K X Hl,i log(pˆi ) (6) i=1 where Hl,i denotes row l penalization factor of class i, K is the number of classes, and pˆi is the predictive softmax with values [0, 1] indicating the predicted probability of each class as: exp(xi ) (7) pˆi = PK j=1 exp(xj ) Equation (6) can be re-written as: exp(xi ) Hl,i log( P ) j exp(xj ) i X X X =− Hl,i xi + Hl,i log( exp(xj )) E= − X i = log( i X exp(xj )) j (8) j X Hl,i − i X Hl,i xi i The derivate with respect to the prediction xk is: X X ∂E ∂ ∂ X = [log( exp(xj )) Hl,i ] − [ Hl,i xi ] ∂xk ∂xk ∂xk j i i 1 ∂ X exp(xj ) − Hl,k j exp(xj ) ∂xk j X =( Hl,i ) P i X exp(xk ) =( Hl,i ) P − Hl,k j exp(xj ) i X =( Hl,i )pˆk − Hl,k i (9) When H = I , the identity, the proposed weighted-loss approach gives the traditional cross-entropy loss function. We used the implemented Infogain loss in Caffe [66] for this purpose. For simplicity, we used a diagonal matrix defined as: ( fi , if i = j Hij = fmin (10) 0, otherwise where fi is the number of samples of the ith class and fmin is the number of samples in the most under-represented class, i.e., Disgust class in this situation. Before training the network, the faces were cropped and resized to 256×256 pixels. No facial registration was performed at this baseline. To augment the data, five crops of 224×224 and their horizontal flips were extracted from the four corners and the center of the image at random during the training phase. The networks were trained for 20 epochs using a batch size of 256. The base learning rate was set to 0.01, and decreased step-wise by a factor of 0.1 every 10,000 iterations. We used a momentum of 0.9. Table 7 shows the top-1 and top-2 F1-Scores for the imbalanced learning, down-sampling, up-sampling, and weighted-loss approaches on the test set. Since the test set is imbalanced, both the skew-normalized and the original scores are reported. The skew normalization is performed by random under-sampling of the classes in the test set. This process is repeated 200 times, and the skew-normalized score is the average of the score on multiple trials. As it is shown, the weighted-loss approach performed better than other approaches in the skew-normalized fashion. The improvement is significant in under-represented classes, i.e., Contempt, Fear, and Disgust. The imbalanced approach performed worst in the Contempt and Disgust categories since there were a few training samples of these classes compared with other classes. The up-sampling approach also did not classify the Contempt and Disgust categories well, since the training samples of these classes were heavily upsampled (almost 20 times), and the network was over-fitted to these samples. Hence the network lost its generalization and performed poorly on these classes of the test set. The confusion matrix of the weighted-loss approaches is shown in Table 8. The weighted-loss approach classified the samples of Contempt and Disgust categories with an acceptable accuracy but did not perform well in Happy and Neutral. This is because the network was not penalized enough for misclassifying examples from these classes. We believe that a better formulation of the weight matrix H based on the number of samples in the mini-batches or other data-driven approaches can improve the recognition of wellrepresented classes. Table 9 shows accuracy, F1-score, Cohens kappa, Krippendorfs Alpha, area under the ROC curve (AUC), and area under the Precision-Recall curve (AUC-PR) on the test sets. Except for the accuracy, all the metrics are calculated in a binary-class manner where the positive class contains the samples labeled by the given category, and the negative class contains the rest. The reported result in Table 9 is the average of these metrics over eight classes. The accuracy is defined in a multi-class manner in which the number of correct predictions is divided by the total number of samples in the test set. The skew-normalization is performed by IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 11 NE HA SA SU FE DI AN CO NE 53.3 4.5 13.0 3.4 1.5 2.0 6.2 16.2 HA 2.8 72.8 1.3 1.2 1.5 2.2 1.2 13.1 SA 9.8 1.1 61.7 1.7 4.6 5.8 5.0 3.5 Predicted SU FE 8.7 1.7 6.0 0.6 3.6 5.8 69.9 18.9 13.5 70.4 3.3 6.2 3.2 5.8 3.1 0.5 DI 2.5 1.7 4.4 1.7 4.2 68.6 11.1 4.3 AN 10.4 1.0 9.2 2.8 4.3 10.6 65.8 5.7 CO 10.9 12.2 1.2 0.5 0.2 1.3 1.9 53.8 balancing the distribution of classes in the test set using random under-sampling and averaging over 200 trials. Since the validation set is balanced, there is no need for skewnormalization. We compared the performance of CNN baseline with a Support Vector Machine (SVM) [62]. To train SVM, the faces in the images were cropped and resized to 256×256 pixels. HOG [67] features were extracted with the cell size of 8. We applied PCA retaining 95% of the variance to reduce the HOG features dimensionality from 36,864 to 6,697 features. We used a linear kernel SVM in Liblinear package [68] (which is optimized for large-scale linear classification and regression). Table 9 shows the evaluation metrics of SVM. Reported AUC and AUCPR values for SVM are calculated using the LibLinear’s resulting decision values. We calculated the scores of predictions using a posterior-probability transformation sigmoid function. Comparing the performance of SVM with the CNN baselines on AffectNet, indicates that CNN models perform better than conventional SVM and HOG features in all metrics. We also compared the baseline with an available off-theshelf expression recognition system (Microsoft Cognitive Services emotion API [69]). The MS cognitive system had an excellent performance on Neutral and Happy categories with an accuracy of 0.94 and 0.85, respectively. However, it performed poorly on other classes with an accuracy of 0.25, 0.27 and 0.04 in the Fear, Disgust and Contempt categories. Table 9 shows the evaluation metrics on the MS cognitive system. Comparing the performance of the MS cognitive with the simple baselines on AffectNet indicates that AffectNet is a challenging database and a great resource to further improve the performance of facial expression recognition systems. Figure 5 shows nine samples of randomly selected misclassified images of the weighted-loss approach and their corresponding ground-truth. As the figure shows, it is really difficult to assign some of the emotions to a single category. Some of the faces have partial similarities in facial features to the misclassified images, such as nose wrinkled in disgust, or eyebrows raised in surprise. This emphasizes the fact that classifying facial expressions in the wild is a challenging task and, as mentioned before, even human annotators agreed on only 60.7% of the images. 4.3 Dimensional Model (Valence and Arousal) Baseline Predicting dimensional model in the continuous domain is a real-valued regression problem. We used AlexNet [57] architecture as our deep CNN baseline to predict the value Angry (Disgust) Disgust (Angry) Fear (Sad) Angry (Sad) Happy (Surprise) Fear (Surprise) Surprise (Fear) Angry (Fear) Angry (Disgust) Happy (Neutral) Sad (Angry) Happy (Contempt) Fig. 5. Samples of miss-classified images. Their corresponding groundtruth is given in parentheses. 0.25 Training Validation 0.2 error Actual TABLE 8 Confusion Matrix of Weighted-Loss Approach on the Test Set 0.15 0.1 0 2000 4000 6000 8000 10000 12000 14000 16000 iteration Fig. 6. Euclidean error of training valence and arousal. of valence and arousal. Particularly, two separate AlexNets were trained where the last fully-connected layer was replaced with a linear regression layer containing only one neuron. The output of the neuron predicted the value of valence/arousal in continuous domain [-1,1]. A Euclidean (L2) loss was used to measure the distance between the predicted value (ŷn ) and actual value of valence/arousal (yn ) as: N 1 X ||ŷn − yn ||22 (11) E= 2N n=1 The faces were cropped and resized to 256×256 pixels. The base learning rate was fixed and set to 0.001 during the training process. We used a momentum of 0.9. Training was continued until a plateau was reached in the Euclidean error of the validation set (approximately 16 epochs with a minibatch size of 256). Figure 6 shows the value of training and validation losses over 16K iterations (about 16 epochs). We also compared Support Vector Regression (SVR) [63] with our DNN baseline for predicting valence and arousal in AffectNet. In our experiments, first, the faces in the images were cropped and resized to 256×256 pixels. Histogram of Oriented Gradient (HOG) [67] features were extracted with the cell size of 8. Afterward, we applied PCA retaining 95% IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 12 TABLE 9 Evaluation Metrics and Comparison of CNN baselines, SVM and MS Cognitive on Categorical Model of Affect. Accuracy F1 -Score Kappa Alpha AUC AUCPR Imbalanced Orig Norm 0.72 0.54 0.57 0.52 0.53 0.46 0.52 0.45 0.85 0.80 0.56 0.55 CNN Baselines Down-Sampling Up-Sampling Orig Norm Orig Norm 0.68 0.58 0.68 0.57 0.56 0.57 0.56 0.55 0.51 0.51 0.52 0.49 0.51 0.51 0.51 0.48 0.82 0.85 0.82 0.84 0.54 0.57 0.55 0.56 TABLE 10 Baselines’ Performances of Predicting Valence and Arousal on Test Set RMSE CORR SAGR CCC CNN (AlexNet) Valence Arousal 0.394 0.402 0.602 0.539 0.728 0.670 0.541 0.450 SVR Valence Arousal 0.494 0.400 0.429 0.360 0.619 0.748 0.340 0.199 Fig. 7. RMSE of predicted valence and arousal using AlexNet and Euclidean (L2) loss (Best viewed in color). of the variance of these features to reduce the dimensionality. Two separate SVRs were trained to predict the value of valence and arousal. Liblinear [68] package was used to implement SVR baseline. Table 10 shows the performances of the proposed baseline and SVR on the test set. As shown, the CNN baseline can predict the value of valence and arousal better than SVR. This is because the high variety of samples in AffectNet allows the CNN to extract more discriminative features than hand-crafted HOG, and therefore it learned a better representation of dimensional affect. The RMSE of CNN baseline (AlexNet) between the predicted valence and arousal and the ground-truth are shown in Fig. 7. As illustrated, the CNN baseline has a lower error rate in the center of circumplex. In particular, predicting low-valence mid-arousal and low-arousal mid-valence areas were more challenging. These areas correspond to the expressions of contempt, bored, and sleepy. It should be mentioned that predicting valence and arousal in the wild is a challenging task, and as discussed in Sec. 3.3, the disagreement between two human annotators has RMSE=0.367 and RMSE=0.481 for valence and arousal, respectively. 5 Weighted-Loss Orig Norm 0.64 0.63 0.55 0.62 0.5 0.57 0.5 0.57 0.86 0.86 0.58 0.64 SVM Orig 0.60 0.37 0.32 0.31 0.77 0.39 Norm 0.37 0.31 0.25 0.22 0.70 0.37 MS Cognitive Orig 0.68 0.51 0.46 0.46 0.83 0.52 Norm 0.48 0.45 0.40 0.37 0.77 0.50 machine learning methodologies, and their performance highly depends on the amount and diversity of annotated training samples. Recently, databases of facial expression and affect in the wild received much attention. However, existing databases of facial affect in the wild only cover one model of affect, have a limited number of subjects, or contain few samples of certain emotions. The Internet is a vast source of facial images, most of which are captured in uncontrolled conditions. These images are often taken in the wild under natural conditions. In this paper, we introduced a new publicly available database of a facial Affect from the InterNet (called AffectNet) by querying different search engines using emotion related tags in six different languages. AffectNet contains more than 1M images with faces and extracted landmark points. Twelve human experts manually annotated 450,000 of these images in both the categorical and dimensional (valence and arousal) models and tagged the images that have any occlusion on the face. The agreement level of human labelers on a subset of AffectNet showed that expression recognition and predicting valence and arousal in the wild is a challenging task. The two annotators agreed on 60.7% of the category of facial expressions, and there was a large disagreement on the value of valence and arousal (RMSE=0.34 and 0.36) between the two annotators. Two simple deep neural network baselines were examined to classify the facial expression images and predict the value of valence and arousal in the continuous domain of dimensional model. Evaluation metrics showed that simple deep neural network baselines trained on AffectNet can perform better than conventional machine learning methods and available off-the-shelf expression recognition systems. AffectNet is by far the largest database of facial expression, valence and arousal in the wild, enabling further progress in the automatic understanding of facial behavior in both categorical and continuous dimensional space. The interested investigators can study categorical and dimensional models in the same corpus, and possibly co-train them to improve the performance of their affective computing systems. It is highly anticipated that the availability of this database for the research community, along with the recent advances in deep neural networks, can improve the performance of automated affective computing systems in recognizing facial expressions and predicting valence and arousal. C ONCLUSION The analysis of human facial behavior is a very complex and challenging problem. The majority of the techniques for automated facial affect analysis are mainly based on ACKNOWLEDGMENTS This work is partially supported by the NSF grants IIS1111568 and CNS-1427872. We gratefully acknowledge the IEEE TRANSACTIONS ON AFFECTIVE COMPUTING support of NVIDIA Corporation with the donation of the Tesla K40 GPUs used for this research. R EFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] J. Tao and T. Tan, “Affective computing: A review,” in International Conference on Affective Computing and Intelligent Interaction. Springer, 2005, pp. 981–995. 1 P. Ekman and W. V. Friesen, “Constants across cultures in the face and emotion.” Journal of personality and social psychology, vol. 17, no. 2, p. 124, 1971. 1 J. A. Russell, “A circumplex model of affect,” Journal of Personality and Social Psychology, vol. 39, no. 6, pp. 1161–1178, 1980. 1, 7, 8 P. Ekman and W. V. Friesen, “Facial action coding system,” 1977. 1 W. V. Friesen and P. Ekman, “Emfacs-7: Emotional facial action coding system,” Unpublished manuscript, University of California at San Francisco, vol. 2, p. 36, 1983. 1 S. Du, Y. Tao, and A. M. Martinez, “Compound facial expressions of emotion,” Proceedings of the National Academy of Sciences, vol. 111, no. 15, pp. E1454–E1462, 2014. 1, 3 M. Lyons, S. Akamatsu, M. Kamachi, and J. Gyoba, “Coding facial expressions with gabor wavelets,” in Automatic Face and Gesture Recognition, 1998. Proceedings. Third IEEE International Conference on. IEEE, 1998, pp. 200–205. 2 Y.-I. Tian, T. Kanade, and J. F. Cohn, “Recognizing action units for facial expression analysis,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 23, no. 2, pp. 97–115, 2001. 2 P. Lucey, J. F. Cohn, T. Kanade, J. Saragih, Z. Ambadar, and I. Matthews, “The extended cohn-kanade dataset (ck+): A complete dataset for action unit and emotion-specified expression,” in Computer Vision and Pattern Recognition Workshops (CVPRW), 2010 IEEE Computer Society Conference on. IEEE, 2010, pp. 94–101. 2, 3, 4 M. Pantic, M. Valstar, R. Rademaker, and L. Maat, “Web-based database for facial expression analysis,” in Multimedia and Expo, 2005. ICME 2005. IEEE International Conference on. IEEE, 2005, pp. 5–pp. 2, 3 R. Gross, I. Matthews, J. Cohn, T. Kanade, and S. Baker, “Multipie,” Image and Vision Computing, vol. 28, no. 5, pp. 807–813, 2010. 2, 3 J. F. Cohn and K. L. Schmidt, “The timing of facial motion in posed and spontaneous smiles,” International Journal of Wavelets, Multiresolution and Information Processing, vol. 2, no. 02, pp. 121– 132, 2004. 2 M. F. Valstar, M. Pantic, Z. Ambadar, and J. F. Cohn, “Spontaneous vs. posed facial behavior: automatic analysis of brow actions,” in Proceedings of the 8th international conference on Multimodal interfaces. ACM, 2006, pp. 162–170. 2 S. M. Mavadati, M. H. Mahoor, K. Bartlett, P. Trinh, and J. F. Cohn, “Disfa: A spontaneous facial action intensity database,” Affective Computing, IEEE Transactions on, vol. 4, no. 2, pp. 151–160, 2013. 2, 3, 4, 5 D. McDuff, R. Kaliouby, T. Senechal, M. Amr, J. Cohn, and R. Picard, “Affectiva-mit facial expression dataset (am-fed): Naturalistic and spontaneous facial expressions collected,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2013, pp. 881–888. 2, 3, 4, 5 I. Sneddon, M. McRorie, G. McKeown, and J. Hanratty, “The belfast induced natural emotion database,” IEEE Transactions on Affective Computing, vol. 3, no. 1, pp. 32–41, 2012. 2, 4 J. F. Grafsgaard, J. B. Wiggins, K. E. Boyer, E. N. Wiebe, and J. C. Lester, “Automatically recognizing facial expression: Predicting engagement and frustration.” in EDM, 2013, pp. 43–50. 2 A. Dhall, R. Goecke, J. Joshi, M. Wagner, and T. Gedeon, “Emotion recognition in the wild challenge 2013,” in Proceedings of the 15th ACM on International conference on multimodal interaction. ACM, 2013, pp. 509–516. 2, 3 I. J. Goodfellow, D. Erhan, P. L. Carrier, A. Courville, M. Mirza, B. Hamner, W. Cukierski, Y. Tang, D. Thaler, D.-H. Lee et al., “Challenges in representation learning: A report on three machine learning contests,” Neural Networks, vol. 64, pp. 59–63, 2015. 2, 3 A. Mollahosseini, B. Hasani, M. J. Salvador, H. Abdollahi, D. Chan, and M. H. Mahoor, “Facial expression recognition from world wild web,” in The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2016. 2, 3, 5 13 [21] S. Zafeiriou, A. Papaioannou, I. Kotsia, M. A. Nicolaou, and G. Zhao, “Facial affect “in-the-wild”: A survey and a new database,” in International Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, Affect ”in-the-wild” Workshop, June 2016. 2, 3, 4 [22] C. F. Benitez-Quiroz, R. Srinivasan, and A. M. Martinez, “Emotionet: An accurate, real-time algorithm for the automatic annotation of a million facial expressions in the wild,” in Proceedings of IEEE International Conference on Computer Vision & Pattern Recognition (CVPR16), Las Vegas, NV, USA, 2016. 2, 3, 5 [23] M. A. Nicolaou, H. Gunes, and M. Pantic, “Audio-visual classification and fusion of spontaneous affective data in likelihood space,” in Pattern Recognition (ICPR), 2010 20th International Conference on. IEEE, 2010, pp. 3695–3699. 3 [24] ——, “Continuous prediction of spontaneous affect from multiple cues and modalities in valence-arousal space,” IEEE Transactions on Affective Computing, vol. 2, no. 2, pp. 92–105, 2011. 3, 5 [25] F. Ringeval, A. Sonderegger, J. Sauer, and D. Lalanne, “Introducing the recola multimodal corpus of remote collaborative and affective interactions,” in Automatic Face and Gesture Recognition (FG), 2013 10th IEEE International Conference and Workshops on. IEEE, 2013, pp. 1–8. 3, 4 [26] S. Koelstra, C. Muhl, M. Soleymani, J.-S. Lee, A. Yazdani, T. Ebrahimi, T. Pun, A. Nijholt, and I. Patras, “Deap: A database for emotion analysis; using physiological signals,” IEEE Transactions on Affective Computing, vol. 3, no. 1, pp. 18–31, 2012. 3, 4, 5 [27] A. Dhall, R. Goecke, S. Lucey, and T. Gedeon, “Static facial expression analysis in tough conditions: Data, evaluation protocol and benchmark,” in Computer Vision Workshops (ICCV Workshops), 2011 IEEE International Conference on. IEEE, 2011, pp. 2106–2112. 3 [28] A. Mollahosseini and M. H. Mahoor, “Bidirectional warping of active appearance model,” in Computer Vision and Pattern Recognition Workshops (CVPRW), 2013 IEEE Conference on. IEEE, 2013, pp. 875–880. 3 [29] S. Ren, X. Cao, Y. Wei, and J. Sun, “Face alignment at 3000 fps via regressing local binary features,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 1685–1692. 3, 6 [30] L. Yu, “face-alignment-in-3000fps,” https://github.com/ yulequan/face-alignment-in-3000fps, 2016. 3, 6 [31] G. A. Miller, “Wordnet: a lexical database for english,” Communications of the ACM, vol. 38, no. 11, pp. 39–41, 1995. 3 [32] P. Viola and M. J. Jones, “Robust real-time face detection,” International journal of computer vision, vol. 57, no. 2, pp. 137–154, 2004. 3 [33] D. You, O. C. Hamsici, and A. M. Martinez, “Kernel optimization in discriminant analysis,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 3, pp. 631–638, 2011. 3 [34] P. Lucey, J. F. Cohn, K. M. Prkachin, P. E. Solomon, and I. Matthews, “Painful data: The unbc-mcmaster shoulder pain expression archive database,” in Automatic Face & Gesture Recognition and Workshops (FG 2011), 2011 IEEE International Conference on. IEEE, 2011, pp. 57–64. 4 [35] M. M. Bradley and P. J. Lang, “Measuring emotion: the selfassessment manikin and the semantic differential,” Journal of behavior therapy and experimental psychiatry, vol. 25, no. 1, pp. 49–59, 1994. 4 [36] B. Schuller, M. Valstar, F. Eyben, G. McKeown, R. Cowie, and M. Pantic, “Avec 2011–the first international audio/visual emotion challenge,” in International Conference on Affective Computing and Intelligent Interaction. Springer, 2011, pp. 415–424. 4, 5 [37] B. Schuller, M. Valster, F. Eyben, R. Cowie, and M. Pantic, “Avec 2012: the continuous audio/visual emotion challenge,” in Proceedings of the 14th ACM international conference on Multimodal interaction. ACM, 2012, pp. 449–456. 4, 5 [38] M. Valstar, B. Schuller, K. Smith, F. Eyben, B. Jiang, S. Bilakhia, S. Schnieder, R. Cowie, and M. Pantic, “Avec 2013: the continuous audio/visual emotion and depression recognition challenge,” in Proceedings of the 3rd ACM international workshop on Audio/visual emotion challenge. ACM, 2013, pp. 3–10. 4 [39] M. Valstar, B. Schuller, K. Smith, T. Almaev, F. Eyben, J. Krajewski, R. Cowie, and M. Pantic, “Avec 2014: 3d dimensional affect and depression recognition challenge,” in Proceedings of the 4th International Workshop on Audio/Visual Emotion Challenge. ACM, 2014, pp. 3–10. 4 [40] F. Ringeval, B. Schuller, M. Valstar, R. Cowie, and M. Pantic, “Avec 2015: The 5th international audio/visual emotion challenge and IEEE TRANSACTIONS ON AFFECTIVE COMPUTING [41] [42] [43] [44] [45] [46] [47] [48] [49] [50] [51] [52] [53] [54] [55] [56] [57] [58] [59] [60] [61] workshop,” in Proceedings of the 23rd ACM international conference on Multimedia. ACM, 2015, pp. 1335–1336. 4, 5 M. Valstar, J. Gratch, B. Schuller, F. Ringeval, D. Lalanne, M. T. Torres, S. Scherer, G. Stratou, R. Cowie, and M. Pantic, “Avec 2016-depression, mood, and emotion recognition workshop and challenge,” arXiv preprint arXiv:1605.01600, 2016. 4, 5 G. McKeown, M. Valstar, R. Cowie, M. Pantic, and M. Schroder, “The semaine database: Annotated multimodal records of emotionally colored conversations between a person and a limited agent,” IEEE Transactions on Affective Computing, vol. 3, no. 1, pp. 5–17, 2012. 4 A. Mollahosseini, D. Chan, and M. H. Mahoor, “Going deeper in facial expression recognition using deep neural networks,” IEEE Winter Conference on Applications of Computer Vision (WACV), 2016. 5 C. Shan, S. Gong, and P. W. McOwan, “Facial expression recognition based on local binary patterns: A comprehensive study,” Image and Vision Computing, vol. 27, no. 6, pp. 803–816, 2009. 5 X. Zhang, M. H. Mahoor, and S. M. Mavadati, “Facial expression recognition using {l} {p}-norm mkl multiclass-svm,” Machine Vision and Applications, pp. 1–17, 2015. 5 L. He, D. Jiang, L. Yang, E. Pei, P. Wu, and H. Sahli, “Multimodal affective dimension prediction using deep bidirectional long shortterm memory recurrent neural networks,” in Proceedings of the 5th International Workshop on Audio/Visual Emotion Challenge. ACM, 2015, pp. 73–80. 5, 6 Y. Fan, X. Lu, D. Li, and Y. Liu, “Video-based emotion recognition using cnn-rnn and c3d hybrid networks,” in Proceedings of the 18th ACM International Conference on Multimodal Interaction. ACM, 2016, pp. 445–450. 5 Y. Tang, “Deep learning using linear support vector machines,” arXiv preprint arXiv:1306.0239, 2013. 5 M. Sokolova, N. Japkowicz, and S. Szpakowicz, “Beyond accuracy, f-score and roc: a family of discriminant measures for performance evaluation,” in Australasian Joint Conference on Artificial Intelligence. Springer, 2006, pp. 1015–1021. 4 J. Cohen, “A coefficient of agreement for nominal scales,” Educational and Psychological Measurement, vol. 20, no. 1, p. 37, 1960. 4, 9 K. Krippendorff, “Estimating the reliability, systematic error and random error of interval data,” Educational and Psychological Measurement, vol. 30, no. 1, pp. 61–70, 1970. 4, 9 P. E. Shrout and J. L. Fleiss, “Intraclass correlations: uses in assessing rater reliability.” Psychological bulletin, vol. 86, no. 2, p. 420, 1979. 4 L. A. Jeni, J. F. Cohn, and F. De La Torre, “Facing imbalanced data–recommendations for the use of performance metrics,” in Affective Computing and Intelligent Interaction (ACII), 2013 Humaine Association Conference on. IEEE, 2013, pp. 245–251. 4, 9 S. Bermejo and J. Cabestany, “Oriented principal component analysis for large margin classifiers,” Neural Networks, vol. 14, no. 10, pp. 1447–1461, 2001. 4 M. Mohammadi, E. Fatemizadeh, and M. H. Mahoor, “Pca-based dictionary building for accurate facial expression recognition via sparse representation,” Journal of Visual Communication and Image Representation, vol. 25, no. 5, pp. 1082–1092, 2014. 5 C. Liu and H. Wechsler, “Gabor feature based classification using the enhanced fisher linear discriminant model for face recognition,” IEEE Trans. Image Process., vol. 11, no. 4, pp. 467–476, 2002. 5 A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in neural information processing systems, 2012, pp. 1097–1105. 5, 9, 11 A. Toshev and C. Szegedy, “Deeppose: Human pose estimation via deep neural networks,” in Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on. IEEE, 2014, pp. 1653–1660. 5 Y. Taigman, M. Yang, M. Ranzato, and L. Wolf, “Deepface: Closing the gap to human-level performance in face verification,” in Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on. IEEE, 2014, pp. 1701–1708. 5 C. Sagonas, E. Antonakos, G. Tzimiropoulos, S. Zafeiriou, and M. Pantic, “300 faces in-the-wild challenge: Database and results,” Image and Vision Computing, vol. 47, pp. 3–18, 2016. 6 G. Paltoglou and M. Thelwall, “Seeing stars of valence and arousal in blog posts,” IEEE Transactions on Affective Computing, vol. 4, no. 1, pp. 116–123, 2013. 7 14 [62] C. Cortes and V. Vapnik, “Support-vector networks,” Machine learning, vol. 20, no. 3, pp. 273–297, 1995. 9, 11 [63] A. Smola and V. Vapnik, “Support vector regression machines,” Advances in neural information processing systems, vol. 9, pp. 155– 161, 1997. 9, 11 [64] G. Caridakis, K. Karpouzis, and S. Kollias, “User and context adaptive neural networks for emotion recognition,” Neurocomputing, vol. 71, no. 13, pp. 2553–2562, 2008. 9 [65] H. He and E. A. Garcia, “Learning from imbalanced data,” IEEE Trans. Knowl. Data Eng., vol. 21, no. 9, pp. 1263–1284, 2009. 9 [66] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell, “Caffe: Convolutional architecture for fast feature embedding,” in Proceedings of the 22nd ACM international conference on Multimedia. ACM, 2014, pp. 675–678. 10 [67] N. Dalal and B. Triggs, “Histograms of oriented gradients for human detection,” in 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR’05), vol. 1. IEEE, 2005, pp. 886–893. 11 [68] R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin, “Liblinear: A library for large linear classification,” Journal of machine learning research, vol. 9, no. Aug, pp. 1871–1874, 2008. 11, 12 [69] “Microsoft cognitive services - emotion api,” https://www. microsoft.com/cognitive-services/en-us/emotion-api, (Accessed on 12/01/2016). 11 Ali Mollahosseini received the BSc degree in computer software engineering from the Iran University of Science and Technology, Iran, in 2006, and the MSc degree in computer engineering - artificial intelligence from AmirKabir University, Iran, in 2010. He is currently working toward the Ph.D. degree and is a graduate research assistant in the Department of Electrical and Computer Engineering at the University of Denver. His research interests include deep neural networks for the analysis of facial expression, developing humanoid social robots and computer vision. Behzad Hasani received the BSc degree in computer hardware engineering from Khaje Nasir Toosi University of Technology, Tehran, Iran, in 2013, and the MSc degree in computer engineering - artificial intelligence from Iran University of Science and Technology, Tehran, Iran, in 2015. He is currently pursuing his Ph.D. degree in electrical & computer engineering and is a graduate research assistant in the Department of Electrical and Computer Engineering at the University of Denver. His research interests include Computer Vision, Machine Learning, and Deep Neural Networks, especially on facial expression analysis. Mohammad H. Mahoor received the BS degree in electronics from the Abadan Institute of Technology, Iran, in 1996, the MS degree in biomedical engineering from the Sharif University of Technology, Iran, in 1998, and the Ph.D. degree in electrical and computer engineering from the University of Miami, Florida, in 2007. He is an Associate Professor of Electrical and Computer Engineering at DU. He does research in the area of computer vision and machine learning including visual object recognition, object tracking and pose estimation, motion estimation, 3D reconstruction, and humanrobot interaction (HRI) such as humanoid social robots for interaction and intervention with children with special needs (e.g., autism) and elderly with depression and dementia. He has received over $3M of research funding from state and federal agencies including the National Science Foundation. He is a Senior Member of IEEE and has published about 100 conference and journal papers. IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 15 A PPENDIX A TABLE 11 Samples of Annotated Categories for Queried Emotion Terms Happy Annotated Expression Neutral Happy Sad Surprise Fear Disgust Anger Contempt None Uncertain Non-Face Sad Queried Expression Surprise Fear Disgust Anger Contempt IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 16 TABLE 12 Samples of Annotated Images by Two Annotators (Randomly selected) Neutral Happy Sad Surprise Fear Annotator 1 Disgust Anger Contempt None Annotator 2 Neutral Happy Sad Surprise Fear Disgust Anger Contempt None Uncertain Non-Face TABLE 13 Agreement percentage between Two Annotators in Categorical Model of Affect (%) A1 A2 A3 A4 A5 A6 A7 A8 A9 A10 A11 A12 A1* 0.0** 69 70 68 0 0 0 0 0 0 0 0 A2 69 0 64.9 68.3 0 0 0 64.7 0 0 0 0 A3 70 64.9 0 70.6 67.4 69.9 63 62.3 0 48.1 0 0 A4 68 68.3 70.6 0 70.4 70.8 64.3 67.5 0 27.5 0 0 A5 0 0 67.4 70.4 0 70.6 0 0 0 0 0 0 A6 0 0 69.9 70.8 70.6 0 0 0 0 0 0 0 A7 0 0 63 64.3 0 0 0 0 0 75.8 0 0 A8 0 64.7 62.3 67.5 0 0 0 0 51.1 0 0 0 A9 0 0 0 0 0 0 0 51.1 0 0 54.4 0 * A1 to A12 indicate Annotators 1 to 12 ** Zero means that there were no common images between the two annotators A10 0 0 48.1 27.5 0 0 75.8 0 0 87.5 0 61.9 A11 0 0 0 0 0 0 0 0 54.4 0 0 0 A12 0 0 0 0 0 0 0 0 0 61.9 0 0 Uncertain Non-Face IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 17 Arousal V: -0.2 A: 0.87 V: 0.2 A: 0.85 V: -0.6 A: 0.75 V: 0.3 A: 0.62 V: -0.4 A: 0.6 V: 0.68 A: 0.66 V: 0.1 A: 0.46 V: -0.77 A: 0.45 V: -0.35 A: 0.28 Valence V: 0.35 A: 0.3 V: -0.95 A: 0.19 V: 0.95 A: 0.17 V: 0.0 A: 0.0 V: -0.55 A: -0.1 V: 0.97 A: -0.1 V: 0.4 A: -0.41 V: -0.3 A: -0.35 V: -0.85 A: -0.38 V: -0.55 A: -0.6 V: 0.23 A: -0.65 V: 0.7 A: -0.73 V: -0.47 A: -0.73 V: -0.12 A: -0.75 V: 0.1 A: -0.85 V: 0.0 A: -0.98 Fig. 8. Sample images in Valence Arousal circumplex with their corresponding Valence and Arousal values (V: Valence, A: Arousal). Arousal TABLE 14 Number of annotated images in each range/area of valence and arousal [.8,1] [.6,.8] [.4,.6] [.2,.4] [0,.2] [-.2,0] [-.4,-.2] [-.6,-.4] [-.8,-.6] [-1,-.8] [-1,-.8] 0 0 638 6770 3331 395 787 610 0 0 [-.8,-.6] 0 74 720 9283 1286 577 1364 7800 3537 0 [-.6,-.4] 21 161 312 3884 2971 5422 3700 2645 8004 4123 [-.4,-.2] 674 561 505 2473 4854 3675 6344 3571 4374 1759 Valence [-.2,0] [0,.2] 1021 521 706 1006 2689 1905 5530 2296 14083 15300 9024 23201 2804 1745 2042 2517 5066 3379 4836 1845 [.2,.4] 60 432 1228 3506 4104 6237 821 1993 4169 1672 [.4,.6] 57 738 992 1824 9998 42219 5241 467 944 739 [.6,.8] 0 530 3891 2667 13842 23281 10619 1271 873 0 [.8,1] 0 0 957 1125 9884 21040 9934 921 0 0 IEEE TRANSACTIONS ON AFFECTIVE COMPUTING 18 TABLE 15 Evaluation Metrics and Comparison of CNN baselines, SVM and MS Cognitive on Categorical Model of Affect on the Validation Set. Accuracy F 1-Score Kappa Alpha AUCPR AUC Imbalanced 0.40 0.34 0.32 0.39 0.42 0.74 CNN Baselines Down-Sampling Up-Sampling 0.50 0.47 0.49 0.44 0.42 0.38 0.42 0.37 0.48 0.44 0.47 0.75 Weighted-Loss 0.58 0.58 0.51 0.51 0.56 0.82 SVM MS Cognitive 0.30 0.24 0.18 0.13 0.30 0.68 0.37 0.33 0.27 0.23 0.38 0.70 TABLE 16 Baselines’ Performances of Predicting Valence and Arousal on the Validation Set RMSE CORR SAGR CCC CNN (AlexNet) Valence Arousal 0.37 0.41 0.66 0.54 0.74 0.65 0.60 0.34 SVR Valence Arousal 0.55 0.42 0.35 0.31 0.57 0.68 0.30 0.18
1cs.CV
EXPLICIT BELYI MAPS OVER Q HAVING ALMOST SIMPLE PRIMITIVE MONODROMY GROUPS arXiv:1703.02848v1 [math.NT] 8 Mar 2017 DOMINIK BARTH AND ANDREAS WENZ Abstract. We present all Belyi maps f : P1 C → P1 C having almost simple primitive monodromy groups (not isomorphic to An or Sn ) containing rigid and rational generating triples of degree between 50 and 250. This also leads to new polynomials having almost simple Galois groups over Q(t). 1. Introduction A triple (x, y, z) ∈ Sn3 is called nice if it satisfies the following conditions: • xyz = 1. • hx, yi is not a full symmetric or alternating group. • hx, yi is primitive and almost simple. • (x, y, z) is a genus 0 triple. • (x, y, z) is rigid and each element is contained in a rational conjugacy class of hx, yi. Using Magma [3] we can find all nice triples of permutation degree between 50 and 250. Theorem. There are exactly 10 nice triples (x, y, z) up to simultaneous conjugation with hx, y, zi being a permutation group of degree between 50 and 250. These triples generate the following groups: group degree ] nice triples Aut(PSL(3, 3)) 52 1 PGL(2, 11) 55 2 NS56 (PSL(3, 4)) 56 1 Aut(PSU(3, 3)) 63 1 Aut(M22 ) 77 1 PSp(4, 4):2 85 1 Aut(HS) 100 2 + O (8, 2) 135 1 According to Riemann’s existence theorem and the Riemann-Hurwitz formula every nice triple (x, y, z) of the previous theorem leads to a 3-branchpoint covering f˜ : P1 C → P1 C (also called Belyi map) with monodromy group isomorphic to hx, yi. Combining the rational rigidity criterion and some additional rationality considerations shows the existence of a Möbius 2010 Mathematics Subject Classification. 12F12. Key words and phrases. Inverse Galois Problem, Belyi Maps, almost simple groups. 1 2 DOMINIK BARTH AND ANDREAS WENZ transformation µ such that f := f˜ ◦ µ ∈ Q(X), see [4] and [5]. Note that Belyi maps are unique up to inner and outer Möbius transformations. In this paper we will present Belyi maps to all nice triples of the theorem and verify their respective monodromy group. 2. Computed Results and Verification We use the following definitions and facts: Let f = p/q ∈ Q(X) be of degree n where p and q are coprime polynomials in Q[X]. The arithmetic monodromy group of f is defined as A := Gal(p(X) − tq(X) ∈ Q(t)[X]), and the geometric monodromy group as G := Gal(p(X) − tq(X) ∈ C(t)[X]). Both groups act transitively on the n roots of p(X) − tq(X) in a splitting field and we have G E A ≤ Sn . The subdegrees of A correspond to the degrees of the irreducible factors of p(X) − f (t)q(X) ∈ Q(t)[X]. In order to verify the respective geometric monodromy groups we will pursue the following idea: The first step is to check whether the computed function f = p/q = 1+r/q is indeed a Belyi map, i.e. a three branch point covering P1 C → P1 C ramified over 0, 1 and ∞. This can be done by combining the Riemann-Hurwitz genus formula with the factorization of the polynomials p, q, r ∈ Q[X]. We will omit this easy computation in the following. The next step is to obtain the subdegrees of the arithmetic monodromy group A. In most cases we can then apply the following obvious divisibility criterion: If there is no subset of the subdegrees containing 1 adding up to a nontrivial divisor of the permutation degree, then A is primitive. If this fails we can apply Ritt’s theorem which states that A is imprimitive if and only if f decomposes nontrivially into f = g ◦ h with g, h ∈ Q(X). The latter can be tested by the decomposition algorithm found in [1]. By checking the Magma database for finite primitive groups having the desired subdegrees we obtain a list of possibilities for the arithmetic monodromy group A, and for the geometric monodromy group G, since G E A. Fortunately, in most cases only one of these groups contains elements having the same cycle structure as x, y and z. This turns out to be the geometric monodromy group G. Once G has been determined we check if G contains exactly one generating permutation triple (up to simultaneous conjugation) having the desired cycle structure. In our examples this is always the case and ensures that the ramification data over 0, 1, ∞ is indeed given by {x, y, z}. Note that due to the rational rigidity of (x, y, z) the splitting field of p(X) − tq(X) ∈ Q(t)[X] is regular over Q, thus we have A = G. Aut(PSL(3,3)) of degree 52. We begin with the nice triple (x, y, (xy)−1 ) generating Aut(PSL(3, 3)) where x = (1, 41, 8, 9, 45, 32, 39, 44)(2, 13, 29, 21, 50, 26, 34, 6)(3, 35, 52, 30) (4, 7, 22, 18, 33, 43, 10, 38)(5, 37, 27, 42, 25, 15, 12, 24) (11, 51, 17, 47, 36, 31, 49, 40)(14, 20, 28, 48)(16, 19, 23, 46) BELYI MAPS HAVING ALMOST SIMPLE PRIMITIVE MONODROMY GROUPS 3 and y = (1, 20)(2, 34)(3, 7)(4, 16)(5, 17)(8, 41)(9, 13)(10, 52)(11, 40) (12, 23)(14, 29)(15, 25)(18, 33)(19, 47)(21, 35)(22, 43)(26, 42) (27, 45)(28, 36)(31, 39)(32, 49)(37, 51)(38, 48)(46, 50) of type x (xy)−1 y cycle structure 85 .43 224 .14 410 .24 .14 The corresponding Belyi map f = p/q = 1 + r/q is given by p(X) = 22 · (4X 4 − 16X 3 − 24X 2 − 8X − 1)2 · (4X 4 − 16X 3 − 18X 2 − 8X − 1) · (4X 4 + 8X 3 + 36X 2 + 28X + 5)4 · (4X 6 − 36X 5 − 24X 4 − 4X 3 + 9X 2 + 3X + 1)4 and r(X) = 33 · (X + 1)8 · (2X 2 − 8X − 1)8 · (2X 2 + 1)4 · (6X 2 + 4X + 1)8 . Verification of monodromy. The factorization p(X) − f (t)q(X) ∈ Q(t)[X] yields that A has subdegrees 1, 6, 18, 27. Using the divisibility criterion we find that A is primitive. As there is only one primitive permutation group of degree 52 with these subdegrees we obtain A = Aut(PSL(3, 3)). Since G is normal in A we have G = PSL(3, 3) or G = Aut(PSL(3, 3)). The first case can be ruled out because PSL(3, 3) does not contain elements having the same cycle structure as (xy)−1 , thus G = A = Aut(PSL(3, 3)).  PGL(2,11) of degree 55. This group has exactly two nice triples (x, y, (xy)−1 ) up to simultaneous conjugation. The first one is given by x = (1, 33, 17, 8, 41, 32)(2, 51, 45, 36, 55, 50)(3, 43, 47, 24, 44, 25) (4, 40, 49, 52, 18, 6)(5, 7, 26, 16, 37, 23)(9, 21, 22, 38, 31, 20) (10, 48, 29, 30, 35, 19)(11, 28, 54)(12, 15, 13, 34, 42, 27)(39, 46, 53) and y = (1, 26)(2, 41)(3, 51)(4, 32)(5, 29)(6, 37)(7, 20)(8, 47)(10, 44)(11, 50) (12, 52)(13, 15)(14, 35)(17, 46)(18, 23)(19, 42)(21, 48)(24, 38) (25, 45)(27, 30)(31, 39)(33, 53)(34, 36)(40, 54)(49, 55) of type x y (xy)−1 cycle structure 68 .32 .11 225 .15 412 .23 .11 4 DOMINIK BARTH AND ANDREAS WENZ This triple corresponds to the Belyi map f = p/q = 1 + r/q where q(X) = 114 · (2X + 1) · (176X 3 + 1056X 2 + 330X + 31)4 · (264X 3 + 154X 2 + 22X + 1)4 · (352X 3 + 264X 2 + 99X + 14)4 · (704X 3 + 132X 2 + 1)4 · (1408X 3 + 693X 2 + 132X + 8)2 and r(X) = − 24 · (22X 2 − 11X − 2)6 · (44X 2 + 22X + 3)6 · (88X 2 + 55X + 1)6 · (176X 2 + 44X + 5)6 · (704X 2 + 242X + 17)3 . Verification of monodromy. The polynomial p(X)−f (t)q(X) ∈ Q(t)[X] factorizes into irreducible polynomials of degree 1, 6, 12, 12, 12, 12, thus A must have the subdegrees 1, 6, 12, 12, 12, 12 and the primitivity of A follows from the divisibility criterion. The only primitive group of degree 55 having these subdegrees is PGL(2, 11). Since G is normal in PGL(2, 11) we have G = PGL(2, 11) or G = PSL(2, 11). The latter case can be ruled out because PSL(2, 11) does not contain elements with the cycle structure of y. We find G = A = PGL(2, 11).  Concerning the second nice triple (x, y, (xy)−1 ) in PGL(2, 11) we have x = (2, 54, 55, 3)(4, 42, 33, 47)(5, 46, 48, 41)(6, 27, 49, 50)(7, 17, 38, 26) (8, 37, 36, 22)(9, 30, 32, 51)(10, 21, 19, 35)(11, 44, 39, 20)(12, 45, 31, 34) (13, 52, 40, 53)(14, 23, 29, 28)(15, 16, 18, 25) and y = (1, 22)(2, 34)(3, 41)(4, 39)(5, 50)(6, 52)(8, 42)(9, 54)(10, 27)(11, 23) (12, 24)(14, 44)(15, 49)(16, 33)(17, 35)(18, 37)(19, 32)(21, 40)(25, 46) (26, 47)(28, 48)(29, 45)(31, 38)(43, 53)(51, 55) of type x y (xy)−1 cycle structure 413 .13 225 .15 68 .31 .22 The computed results for this Belyi map f = p/q = 1 + r/q are q(X) = 24 · (3X − 1)3 · (2X 2 − 5X − 1)6 · (3X 2 − 2X + 4)2 · (12X 2 − 8X + 5)6 · (12X 4 + 6X 3 + 19X 2 − 3X + 3)6 and r(X) = − 33 · 115 · (4X 3 − 4X 2 + 5X − 3)4 · (6X 3 + 5X 2 + 2X + 1)4 · (12X 3 − 56X 2 + 15X − 9) · (72X 6 − 144X 5 + 230X 4 − 134X 3 + 61X 2 − 14X + 2)4 . BELYI MAPS HAVING ALMOST SIMPLE PRIMITIVE MONODROMY GROUPS 5 Verification of monodromy. The subdegrees of A are 1, 4, 6, 8, 12, 24. By applying the decomposition algorithm we see that f is indecomposable and thus A is primitive by Ritt’s theorem. It follows G = PGL(2, 11) or G = PSL(2, 11) and since there are no elements of cycle structure 225 .15 in PSL(2, 11) we conclude again G = A = PGL(2, 11).  NS56 (PSL(3, 4)) of degree 56. The only nice permutation triple (x, y, (xy)−1 ) in this group is given by x = (1, 36, 2, 5)(3, 47)(7, 45, 33, 22)(8, 31, 55, 14)(9, 21, 50, 48)(10, 16, 40, 39) (11, 54)(12, 19, 49, 23)(13, 41, 42, 15)(17, 56, 24, 30)(18, 53, 44, 25) (20, 52, 28, 35)(26, 29, 46, 37)(27, 34, 51, 43) and y = (1, 27, 36, 32)(2, 43, 50, 14, 52, 17, 30, 54)(3, 38, 47, 29, 33, 53, 16, 46) (4, 7, 22, 26, 28, 55, 44, 45)(5, 11, 24, 56, 20, 12, 41, 34) (6, 40, 18, 19, 35, 37, 10, 39)(8, 21)(9, 48, 51, 13, 15, 49, 25, 31)(23, 42) having the following cycle structure x y cycle structure 412 .22 .14 86 .41 .22 (xy)−1 225 .16 The Belyi map f = p/q = 1 + r/q consists of p(X) = (X 2 − 6X − 1)2 · (3X 4 − 468X 3 − 258X 2 − 60X − 5) · (3X 4 + 36X 3 + 54X 2 + 60X + 19)4 · (3X 8 − 96X 7 − 12X 6 + 432X 5 + 1498X 4 − 320X 3 − 348X 2 − 80X − 5)4 and r(X) = − 22 · 55 · (X 2 + 2X + 3)2 · (3X 2 + 6X + 1)8 · (3X 4 − 12X 3 + 38X 2 + 12X + 3)8 Verification of monodromy. The subdegrees of A turn out to be 1, 10, 45, thus A is primitive. There are five finite primitive groups having these subdegrees, all between PSL(3, 4) and NS56 (PSL(3, 4)). As PSL(3, 4) is simple these five groups are also the only possibilities for the geometric monodromy group G E A. However, only NS56 (PSL(3, 4)) contains a generating triple with the desired cycle structure. It follows A = G = NS56 (PSL(3, 4)).  Aut(PSU(3,3)) of degree 63. This group only has the nice triple (x, y, (xy)−1 ) where x = (1, 43, 31, 39, 63, 35, 2)(3, 17, 24, 21, 20, 55, 53)(4, 29, 62, 11, 14, 45, 27) (5, 38, 23, 32, 48, 18, 51)(6, 13, 36, 47, 25, 8, 61)(7, 9, 15, 56, 34, 28, 42) (10, 33, 59, 60, 44, 19, 37)(12, 26, 52, 30, 54, 49, 41) (16, 40, 57, 50, 22, 46, 58) 6 DOMINIK BARTH AND ANDREAS WENZ and y = (2, 53)(3, 9)(4, 38)(5, 29)(6, 51)(8, 25)(10, 19)(11, 20)(12, 44)(13, 43) (14, 21)(16, 40)(17, 32)(18, 39)(22, 62)(23, 30)(24, 26)(27, 59)(33, 54) (34, 63)(35, 56)(36, 55)(41, 49)(42, 48)(45, 60)(46, 61)(47, 50)(57, 58) of type x y (xy)−1 cycle structure 79 228 .17 412 .26 .13 The corresponding Belyi map is given by f = p/q = 1 + r/q where r(X) = − 28 · 312 · (X 2 − X + 2)7 · (X 3 + 2X 2 − X − 1)7 · (X 3 + 9X 2 − X − 1)7 and q(X) = (X 3 + 30X 2 + 27X + 6) ·(X 6 + 18X 5 + 93X 4 + 169X 3 + 144X 2 − 75X − 62)2 ·(X 12 + 15X 11 − 15X 10 − 332X 9 − 2766X 8 + 4002X 7 + 2002X 6 − 2496X 5 − 1215X 4 + 1047X 3 + 117X 2 − 108X + 36)4 . Verification of monodromy. The subdegrees of A are 1, 6, 24, 32. Again, by applying the decomposition algorithm we see that A is primitive and only three possibilities for A remain. Since G is normal in A there are four possibilities for G. Fortunately, among these groups only Aut(PSU(3, 3)) contains elements of cycle structure of y and (xy)−1 . This yields G = A = Aut(PSU(3, 3)).  Aut(M22 ) of degree 77. The only nice triple (x, y, (xy)−1 ) is given by x = (1, 14, 3, 53, 31, 27, 71, 62, 10, 65, 61)(2, 50, 46, 29, 12, 7, 56, 19, 63, 28, 25) (4, 36, 38, 44, 17, 13, 66, 43, 39, 9, 72)(5, 49, 68, 51, 58, 59, 70, 15, 11, 23, 33) (6, 55, 42, 67, 32, 21, 45, 64, 48, 77, 57)(8,41,60,20, 26, 74, 76, 24, 69, 52, 40) (16, 22, 54, 35, 34, 37, 18, 73, 75, 30, 47) and y = (1, 54)(2, 59)(3, 48)(4, 20)(6, 32)(7, 29)(11, 38)(13, 43)(14, 51)(15, 19) (18, 37)(21, 57)(22, 46)(24, 73)(30, 44)(31, 40)(33, 45)(34, 52)(35, 71) (36, 64)(39, 75)(47, 56)(49, 77)(50, 58)(53, 60)(62, 65)(63, 70)(72, 76) of type x y (xy)−1 cycle structure 117 228 .121 416 .26 .11 The corresponding Belyi map is f = p/q = 1 + r/q with p(X) = 222 · (X 2 + X + 3)11 · (X 5 − 3X 4 − 14X 3 + 15X 2 + X − 1)11 BELYI MAPS HAVING ALMOST SIMPLE PRIMITIVE MONODROMY GROUPS 7 and r(X) = − 114 · (X 4 + 2X 3 + 7X 2 − 16X − 2)4 ·(X 6 + 14X 5 + 34X 4 + 8X 3 − 30X 2 + 60X + 16)4 ·(4X 6 + X 5 + 15X 4 + 10X 3 − 10X 2 − 2X − 2)4 ·(16X 6 − 29X 5 + 71X 4 − 136X 3 + 92X 2 − 8X − 8)2 . Verification of monodromy. The subdegrees of A are 1, 16, 60, thus A is primitive. The classification of finite primitive rank 3 groups yields A = M22 or A = Aut(M22 ). Since M22 does not contain elements with the cycle structure of y, we see G = A = Aut(M22 ).  PSp(4,4):2 of degree 85. We now consider the nice triple (x, y, (xy)−1 ) where x = (1, 85, 49, 26, 15, 39, 65, 24, 37, 4, 23, 3, 28, 19, 76) (2, 82, 64, 74, 52, 58, 20, 70, 43, 7, 68, 12, 53, 40, 16) (5, 81, 51, 67, 54, 44, 41, 77, 30, 21, 71, 63, 33, 66, 18) (6, 42, 46, 50, 60, 22, 73, 80, 47, 45, 14, 31, 13, 55, 79) (8, 62, 56, 36, 72, 69, 35, 25, 10, 84, 48, 34, 59, 27, 11) (9, 57, 83, 78, 17)(29, 38, 61, 75, 32) and y = (3, 70)(5, 29)(6, 17)(7, 68)(8, 27)(9, 74)(14, 23)(15, 39)(16, 57)(18, 33) (19, 52)(20, 58)(22, 35)(24, 31)(25, 60)(26, 65)(30, 71)(32, 81)(34, 56) (36, 51)(38, 53)(40, 63)(41, 78)(42, 84)(43, 61)(44, 48)(45, 75)(46, 50) (47, 72)(49, 55)(54, 67)(64, 82)(73, 80)(76, 79)(77, 83). It has the following cycle structure: x y (xy)−1 cycle structure 155 .52 235 .115 416 .27 .17 This leads to the Belyi map f = p/q = 1 + r/q with q(X) = X · (5X 3 + 20X 2 + 20X + 6)2 · (5X 4 + 10X 3 − 14X − 10)4 · (5X 4 + 10X 3 − 8X − 4)2 · (5X 6 + 30X 5 + 60X 4 + 8X 3 − 48X 2 − 24X − 4) · (625X 12 + 3750X 11 + 7500X 10 + 3500X 9 − 3750X 8 − 1500X 7 + 2700X 6 + 3000X 5 + 2100X 4 + 1040X 3 + 240X 2 − 8)4 and r(X) = − 224 · (5X 2 + 5X + 2)5 · (5X 4 + 10X 3 − 5X − 1)15 . Verification of monodromy. Again, A turns out to be a primitive rank 3 group with subdegrees 1, 20, 64. Thus, we either have A = PSp(4, 4) or 8 DOMINIK BARTH AND ANDREAS WENZ A = PSp(4, 4):2. As PSp(4, 4) does not contain elements with the cycle structure of y, we obtain A = G = PSp(4, 4):2.  Aut(HS) of degree 100. Both nice triples (x, y, (xy)−1 ) generating Aut(HS) have already been discussed in [2]. The first one is given by x = (1, 23, 53, 86)(2, 36, 29, 43)(3, 15, 46, 6)(4, 80, 71, 81)(5, 75, 16, 47) (7, 32, 60, 8)(9, 76, 100, 51)(10, 50, 49, 34)(11, 28, 74, 84)(12, 72, 37, 52) (13, 21, 96, 88)(14, 41, 40, 87)(17, 42, 45, 79)(18, 63, 19, 20)(22, 99, 39, 89) (24, 59, 77, 38)(25, 68, 26, 35)(27, 69, 73, 48)(30, 92, 33, 82)(31, 56, 93, 58) (44, 98, 67, 64)(54, 95, 85, 62)(55, 65, 94, 61)(57, 78, 83, 97)(66, 90, 70, 91), y = (1, 75, 5, 71, 15)(2, 43, 52, 89, 39)(3, 18, 100, 33, 35, 26, 58, 32, 53, 23) (4, 81, 47, 16, 86, 7, 42, 38, 77, 59)(6, 41, 14, 87, 82, 76, 9, 97, 19, 63) (8, 60, 93, 56, 13, 61, 36, 99, 70, 45)(10, 65, 55, 88, 12, 29, 94, 34, 49, 50) (11, 44, 64, 25, 92)(17, 72, 96, 69, 28, 30, 40, 46, 80, 24)(20, 83, 78, 57, 51) (21, 31, 68, 67, 98, 84, 74, 27, 48, 73)(22, 37, 79, 90, 66, 95, 54, 62, 85, 91) of type x y z cycle structure 425 108 .54 235 .130 The Belyi map f = p/q = 1 + r/q is given by p(X) = (7X 5 − 30X 4 + 30X 3 + 40X 2 − 95X + 50)4 · (2X 10 − 20X 9 + 90X 8 − 240X 7 + 435X 6 − 550X 5 + 425X 4 − 100X 3 − 175X 2 + 250X − 125)4 · (2X 10 + 5X 8 − 40X 6 + 50X 4 − 50X 2 + 125)4 , q(X) = 28 · (X 4 − 5)5 · (X 8 − 20X 6 + 60X 5 − 70X 4 + 100X 2 − 100X + 25)10 . Verification of monodromy. Once again, A is a primitive rank 3 group with subdegrees 1, 22, 77, therefore A = HS or A = Aut(HS) and we find G = HS or G = Aut(HS). Since HS does not contain elements having the cycle structure of (xy)−1 we get A = G = Aut(HS).  The second triple (x, y, (xy)−1 ) consists of the following permutations: x = (1, 64, 8, 54, 37)(2, 20, 81, 42, 49)(3, 98, 32, 73, 89)(4, 96, 86, 15, 79) (5, 22, 28, 78, 48)(6, 67, 97, 40, 14)(7, 58, 82, 59, 18)(9, 16, 87, 85, 60) (10, 70, 41, 56, 55)(11, 77, 36, 25, 68)(12, 17, 19, 21, 80)(13, 35, 90, 33, 91) (23, 50, 66, 84, 27)(24, 72, 95, 52, 76)(26, 99, 100, 57, 93)(29, 71, 38, 69, 65) (30, 74, 94, 53, 51)(31, 45, 47, 75, 34)(43, 63, 44, 46, 62) BELYI MAPS HAVING ALMOST SIMPLE PRIMITIVE MONODROMY GROUPS 9 and y = (1, 20)(2, 64)(3, 76)(4, 45)(5, 83)(6, 26)(7, 13)(8, 74)(9, 41)(10, 63)(11, 25) (12, 66)(14, 21)(15, 52)(16, 62)(17, 33)(18, 35)(19, 42)(22, 60)(23, 58) (24, 73)(28, 98)(29, 82)(30, 53)(31, 61)(32, 59)(34, 67)(36, 95)(37, 85) (38, 47)(39, 51)(40, 80)(43, 92)(44, 78)(46, 99)(48, 55)(49, 94)(50, 91) (54, 90)(65, 88)(69, 72)(71, 75)(77, 79)(81, 87)(84, 97)(86, 100)(93, 96) of type x y z cycle structure 519 .15 247 .16 610 .310 .25 The corresponding Belyi map f = p/q = 1 + r/q can be computed as p(X) = 33 · (X 4 − 8X 3 − 6X 2 + 8X + 1)5 · (X 5 − 5X 4 + 50X 3 + 70X 2 + 25X + 3)5 · (3X 5 − 5X 4 − 5X 3 + 35X 2 + 40X + 4)· (9X 10 − 30X 9 + 55X 8 − 200X 7 + 210X 6 + 924X 5 − 890X 4 − 360X 3 + 1925X 2 − 1070X + 291)5 and q(X) = (3X 5 − 35X 4 + 90X 3 − 50X 2 + 15X + 9)2 · (9X 10 − 120X 9 + 10X 8 − 1960X 7 − 1090X 6 + 3304X 5 − 760X 4 − 920X 3 + 145X 2 + 80X + 6)3 · (3X 10 − 10X 9 − 65X 8 + 160X 7 − 90X 6 − 932X 5 − 330X 4 + 880X 3 + 1255X 2 + 830X + 27)6 . Verification of monodromy. One can apply the exact same proof as before to show G = A = Aut(HS).  O+ (8, 2) of degree 135. In this group the nice triple (x, y, (xy)−1 ) is given by x = (1, 94, 65, 71, 134, 80, 107, 98, 4)(2, 104, 58, 121, 97, 116, 88, 8, 23) (3, 69, 36, 32, 29, 73, 102, 128, 106)(5, 14, 124, 105, 67, 18, 49, 117, 34) (6, 28, 100, 41, 135, 31, 48, 109, 17)(7, 133, 112, 53, 91, 15, 25, 122, 129) (9, 62, 99, 96, 131, 77, 10, 81, 52)(11, 56, 110, 13, 115, 111, 95, 89, 54) (12, 64, 113, 108, 20, 76, 50, 22, 55)(16, 61, 83, 118, 75, 66, 39, 35, 132) (19, 85, 68, 126, 40, 125, 74, 130, 43)(21, 47, 79, 78, 72, 84, 24, 37, 57) (26, 38, 70, 90, 92, 103, 63, 120, 44)(27, 119, 127, 42, 87, 82, 101, 93, 45) (30, 59, 86, 51, 33, 60, 123, 46, 114) 10 DOMINIK BARTH AND ANDREAS WENZ and y = (3, 118)(4, 110)(5, 132)(7, 36)(9, 33)(10, 46)(12, 112)(13, 129)(16, 65) (17, 106)(20, 113)(21, 107)(22, 55)(25, 61)(26, 27)(28, 30)(29, 37)(31, 109) (32, 98)(35, 130)(40, 42)(43, 99)(44, 125)(45, 90)(47, 49)(50, 91)(51, 93) (52, 60)(54, 56)(58, 116)(59, 128)(62, 82)(63, 73)(64, 69)(66, 96)(67, 78) (71, 117)(72, 105)(74, 124)(75, 135)(76, 83)(77, 121)(80, 134)(84, 120) (85, 87)(86, 103)(88, 104)(89, 94)(95, 122)(97, 114)(100, 131)(119, 127) and is of the following type x y (xy)−1 cycle structure 915 252 .131 430 .26 .13 The corresponding Belyi map f = p/q = 1 + r/q consists of p(X) = 222 · (3X 3 − 9X 2 − 9X − 2)9 · (3X 3 + 9X 2 + 6X + 1)9 · (27X 9 + 243X 8 + 567X 7 + 513X 6 + 162X 5 − 27X 4 + 9X 3 + 27X 2 + 9X + 1)9 and q(X) = (3X 3 − 9X − 2) · (3X 3 + 27X 2 + 27X + 7)2 · (6X 3 + 9X 2 − 1)2 · (36X 6 + 189X 5 + 189X 4 + 96X 3 + 36X 2 + 9X + 1)4 · (81X 12 + 1944X 11 + 11178X 10 + 27648X 9 + 29403X 8 − 1944X 7 − 39150X 6 − 44712X 5 − 25434X 4 − 8088X 3 − 1332X 2 − 72X + 4)4 · (648X 12 + 3888X 11 + 11907X 10 + 15120X 9 + 13365X 8 + 14580X 7 + 11772X 6 + 3240X 5 − 1782X 4 − 1632X 3 − 504X 2 − 72X − 4)4 . Verification of monodromy. We find that A is a primitive rank 3 group with subdegrees 1, 64, 70. Therefore we have A = O+ (8, 2) or A = O+ (8, 2).2 and due to normality G = O+ (8, 2) or G = O+ (8, 2).2. Note that G is generated by permutations of cycle structure 915 , 252 .131 , 430 .26 .13 and by inspecting the sizes of conjugacy classes of O+ (8, 2) and O+ (8, 2).2 we can conclude that there are no elements with these cycle structures in O+ (8, 2).2 \ O+ (8, 2). It follows G = O+ (8, 2). Because G contains only one generating triple (up to simultaneous conjugation) having the desired cycle structure the ramification data over 0, 1, ∞ is given by {x, y, (xy)−1 }. Since (x, y, (xy)−1 ) is rigid we have A = G = O+ (8, 2).  Acknowledgements We would like to thank Peter Müller for introducing us to the subject of this work and for providing us with helpful suggestions. References [1] Mohamed Ayad and Peter Fleischmann. On the decomposition of rational functions J. Symbolic Comput., 43(4):259–274, 2008. [2] Dominik Barth and Andreas Wenz. Explicit Polynomials Having the Higman-Sims Group as Galois Group over Q(t). arXiv:1611.04314, 2016 BELYI MAPS HAVING ALMOST SIMPLE PRIMITIVE MONODROMY GROUPS 11 [3] Wieb Bosma, John Cannon, and Catherine Playoust. The Magma algebra system. I. The user language. J. Symbolic Comput., 24(3-4):235–265, 1997. Computational algebra and number theory (London, 1993). [4] Gunter Malle and B. Heinrich Matzat. Inverse Galois theory. Berlin: Springer, 1999. [5] Helmut Völklein. Groups as Galois groups: an introduction. Cambridge: Cambridge Univ. Press, 1996. Institute of Mathematics, University of Würzburg, Emil-Fischer-Straße 30, 97074 Würzburg, Germany E-mail address: [email protected] E-mail address: [email protected]
4math.GR
Evolutionary Optimisation Methods for Template Based Image Registration Lukasz A Machowski, Tshilidzi Marwala School of Electrical and Information Engineering University of Witwatersrand, Johannesburg, South Africa. [email protected] [email protected] Abstract This paper investigates the use of evolutionary optimisation techniques to register a template with a scene image. An error function is created to measure the correspondence of the template to the image. The problem presented here is to optimise the horizontal, vertical and scaling parameters that register the template with the scene. The Genetic Algorithm, Simulated Annealing and Particle Swarm Optimisations are compared to a Nelder-Mead Simplex optimisation with starting points chosen in a pre-processing stage. The paper investigates the precision and accuracy of each method and shows that all four methods perform favourably for image registration. SA is the most precise, GA is the most accurate. PSO is a good mix of both and the Simplex method returns local minima the most. A pre-processing stage should be investigated for the evolutionary methods in order to improve performance. Discrete versions of the optimisation methods should be investigated to further improve computational performance. 1. Introduction Image registration has great practical application in the field of computer vision, medicine, remote sensing and image watermarking [1][2]. Being able to determine how best the template image fits into the scene poses several problems that have to be overcome. The registration process may involve shifting, scaling, rotation, perspective projection or other non-linear transformations. The shear number of possible transformations makes it difficult to automate the process and usually requires a person to verify the results manually. This paper presents findings on the use of evolutionary optimisation methods for automating the template matching of 2-dimensional intensity images. 1.1 Image Registration Image registration is the process by which a template is oriented in such a way as to match an entire, or a portion of, a given scene [1][2]. The template is transformed in such a way as to match the scene as closely as possible. There are four main steps required for registration of an image. These are feature detection, feature matching, transform model estimation and image transformation [1]. Feature based detection makes it easier to determine the orientation of the template with respect to the scene. Area based detection methods are much more computationally expensive due to the amount of data that needs to be processed. Since the area based detection methods depend on the appearance of the images, they are intolerant of changes in illumination and ambient conditions [1][2]. The feature based detection methods do not suffer from this but it is more difficult to automatically extract the features for any general image. It is common to combine the advantages from both methods to form a hybrid approach to the registration process [2]. Correlation-like methods are typically used for area-based detection methods where a correlation surface is calculated for the template and the maximum point is found and interpreted as the best fit for the template [1]. This method is adversely affected by self similarity in the image and it is characterised by high computational complexity. It also does not allow much variance in template rotation or other more complex transformations. This approach, is still however attractive for real-time object tracking [1][3]. An alternative to cross correlation is to use optimisation to find the best fit for the template in the scene [4]. The advantage of this approach is that one can apply more complex transformations to the templates, and thus make the method robust when compared to cross correlation. This method also requires less computation because the entire correlation surface does not have to be determined. In this paper, we investigate the use of a Genetic Algorithm (GA), Simulated Annealing (SA) and Particle Swarm Optimisation (PSO) to register a template with a given scene. These methods are also compared to the Nelder-Mead Simplex method. For simplicity, only three transformation parameters are defined. These are horizontal translation, vertical translation and uniform scaling. 1.2 Evolutionary Optimisation The term “evolutionary” refers to the fact that the optimum solution gradually evolves from a population of individuals that share information and have group dynamics [5]. This is in contrast to the non-evolutionary or classical optimisation methods which always try to travel in the best direction. Typically, the evolutionary concept is linked with GA alone but in this paper, we group GA, SA and PSO into the subset. All evolutionary optimisation methods have the following operations [5]: • Evaluation • Selection • Alteration An initial population of individuals is initialised, covering the parameter space and the objective function is evaluated for each individual. From this data, a subset of individuals is selected and altered to form new individuals. The degree to which each of these operations is performed in GA, SA and PSO varies from algorithm to algorithm. To find the optimal registration parameters for template matching, it is important to construct a multivariate cost function that represents how well the template matches the scene [6]. The traditional techniques for optimisation make use of the objective function value, first derivative or its second derivative [6][7]. The general approach for all non-evolutionary optimisation methods is to select an initial guess for the registration parameters and travel in a direction as to improve the objective function. Once a suitable direction is found, it is possible to make either fixed or varying successive steps towards the local optimum. Evolutionary Optimisation methods, however, do not make use of any other information but the objective function values themselves. This eliminates the evaluation of gradients which may be expensive and misleading for image registration. Evolutionary methods typically sample the search space significantly more than the non-evolutionary techniques but this improves the probability of the algorithm finding the global optimum point. The evolutionary algorithms are typically based on the processes which occur in the natural world, such as genetics, the swarming behaviour of bees and the annealing of metals. Various data structures are used to simulate these [5]. A brief review of the methods suitable for image registration is given below: 1.2.1 Simplex Method This is a non-evolutionary (classical) method. A simplex is a geometric figure that has one more vertex than the number of dimensions in the parameter space (a triangle in two dimensions, as shown in Figure 1). The objective function is sampled at each vertex and the one that has the worst value gets removed from the simplex. A new vertex is then created by reflecting the simplex about the remaining points. Depending on whether the fitness of the new point improves or not, the simplex is expanded or contracted to look for a more precise solution. In this manner, the algorithm steps its way towards the local optimum point. This method is relatively robust when used for discontinuous objective functions [8]. x2 Reflected Vertex x1 Figure 1 A simplex in two dimensions, showing a reflection. 1.2.2 Simulated Annealing Simulated Annealing is a Monte Carlo Technique [5] and is based on the analogy of metals cooling slowly to form a crystalline structure with low energy. The SA introduces a probability of acceptance of a new sample point for a hill-climbing process. The probability of accepting the new point is based on a control variable called the temperature. The higher the temperature, the more likely it is for a worse point to be accepted. This allows the algorithm to escape from local optima [5][9]. This algorithm is commonly referred to as Metropolis after its founder. The SA technique used in this paper modifies the simplex method described above to allow the simplex to accept a worse vertex with a probability distribution that is based on the temperature. If the change in energy is negative (we have a better point) then the new point will always be chosen. If the change in energy is positive (the point is worse) then the probability of accepting it is given by: p=e − ∆E k B .T (1) where ∆E is the change in energy, kB is Boltzmann’s constant and T is the current temperature. The cooling schedule (how many iterations to spend at each temperature) is an important factor in the success of the algorithm. 1.2.3 Genetic Algorithm Based on the theory of genetics, the GA encodes each individual in the population with a chromosome [5][10]. This encoding represents the parameters for the objective function being optimised. There are several different techniques for encoding parameters, performing the selection, and the alteration stages of the algorithm. The alteration stage is separated into Crossover and Mutation. The method used in this paper selects a random sample of parents from the population with a specified probability. An arithmetic crossover is then performed on these individuals which creates children based on a linear interpolation of the two parents. This is shown in Figure 2. xd (t + 1) = xd (t ) + vd (t + 1) (3) The various components that make up the velocity for each particle are shown in Figure 3. pi x2 c1 = p1.(mix) + p2.(1-mix) c2 = p1.(1-mix) + p2.(mix) x where vd(t+1) is the next velocity for particle d, α is an acceleration constant, βi is an attraction constant for the individual best position, βg is an attraction constant for the group best position, pi,d is the best individual position for particle d, pi,g is the best group position for particle d’s neighbourhood and xd(t) is the current position for particle d. The position of each particle is then calculated for each iteration as: pg p1 c1 eg: Mix = 0.6 Vd(t) x x1 c2 Figure 3 Various components making up the new velocity. p2 Figure 2 One Dimensional Arithmetic crossover operator. 2. A multi-non-uniform mutation is performed which modifies the parent parameters with a binomial distribution which narrows as the number of generations gets larger. More details of the individual techniques are given in [5]. 1.2.4 Particle Swarm Optimisation PSO is based on the swarming behaviour of bees, flocking of birds, schooling of fish and social relations of humans [11]. A population of particles is randomly initialised within the parameter space and each one is given an initial velocity. At each iteration of the algorithm, the particle position is updated and a new velocity is calculated, taking the best position for the particle and the group into account. There are different ways of grouping the particles together. The method used in this paper creates a social grouping where each particle has n logical neighbours (referenced by adjacent index numbers). This means that particles can be neighbours even though they are not close to each other spatially. This method also tends to produce better global exploration by the particles since there are many more attractors. The velocity of each particle is calculated as:   α .v(t )   vd (t + 1) =  + β i .( pi ,d − xd (t ))    + β .( p − x ( t )) g g , d d   (2) Method 2.1 Objective Function In order to perform optimisation, it is necessary to define an objective function that captures the essence of the problem at hand. In image registration, one wants to maximise correspondence between the scene image and the template posed at its current position. The correspondence can be measured as the sum-squared difference between the intensities of overlapping pixels. This can be expressed as an error function where a value of zero represents a perfect match. The parameters to be optimised in this problem are horizontal translation (x), vertical translation (y) and uniform scaling (s). The objective function used when there are overlapping pixels between the template and the scene image is given by: error = ∑ ( A − T ) 2 numel ( A) (3) where A is the scene image, T is the template, the 2 refers to element-wise squaring and the summation is over each element of the resultant matrix. This error is then normalised with the number of pixels that are overlapping between both images. It is also important to interpolate sub-pixel values for the optimisation algorithm to be able to function correctly. This allows the traditional algorithms to be run unchanged and also allows it to be compared to other general optimisation algorithms. It is necessary to penalise the error function when there are template pixels that do not lie within the image. This error component is added to the existing value calculated above. The penalised error then becomes: errorp = error + OutPixels × c (4) where error is from equation (3), OutPixels is the number of template pixels that do not lie in the image, and c is a penalisation constant which should be large (~1000). This penalisation has the effect of constraining the x and y parameters back into range when the images no longer overlap. It is necessary to hard-limit the scale parameter because the optimisation algorithms might try ridiculously high values which require extremely large amounts of memory. Very seldom does a template match a scene at very high scaling values. Similarly, if the template gets scaled to one pixel in size, then a fit can be found nearly anywhere in the scene. 2.2 Test Image The test image used is the familiar picture of Lena, which has a good mix of various image features and provides several local minima for registration. The template is a cut-out of Lena’s face and is shown in Figure 4. The image scene is 256x256 pixels and the template is 170x138 pixels taken from the coordinates (151.5, 151.5) with a scale of 2.0. This means that the global optimum for our objective function is at the coordinates (151.5, 151.5, 0.5) with an error value of 0. (b) (a) Figure 4 a) Test Image b) Template for registration. 2.3 Optimisation This section describes implementation details for each of the optimisation algorithms. All routines are implemented in Matlab R13. The implementation details for each algorithm are given below: 2.3.1 Simplex Method Matlab’s Optimisation Toolbox is used to perform the Nelder-Mead simplex optimisation [8]. The function used is fminsearch.m. A simple pre-processing algorithm is used to choose initial starting points. 2.3.2 Simulated Annealing The SA algorithm used in this paper is based on the code given in [12] and is extended to include restarts. The code is abstracted into a higher level in order to take advantage of Matlab’s matrix arithmetic capabilities. The following is a high level description of the algorithm used: x = RandomStartingSimplex(); y = EvaluateSimplex(S); for each temperature in cooling schedule: for number of iterations: yFluc = AddFluctuation(y); sort(yFluc); ReflectSimplex; If better than best then ExpandSimplex(x) Else if worse than 2nd highest: ContractSimplex(x); If still bad then: ContractOtherVerices(x); End End If SimplexIsStuck() then: % Restart, keeping best point: x = RandomStartingSimplex(); KeepBestPoint(x); End End for each iteration End for each temperature The restart allows the algorithm to oscillate at a local minimum for only a limited number of iterations. After this, it is restarted with the best point as one of the vertices. This behaviour is justified because it is recommended in [12] that the algorithm be re-run in the same fashion once a solution is found. This merely allows the simplex to further explore the parameter space. 2.3.3 Genetic Algorithm The Genetic Optimization Toolbox (GAOT) [13] is used for the implementation of the GA. This is an extensive toolbox with many functions for the encoding, selection, crossover and mutation operators. The following operators were used for the optimisation: • normGeomSelect.m, with the probability of selecting the best, set to 0.6. • arithXover.m, with 10 crossovers per generation • multiNonUnifMutation.m, with 20 mutations per generation. The particular selection operator used gives a good mix of exploration and precision. The arithmetic crossover is very useful in this problem since it assists in finding more precise parameters for the objective function. It is important to use the given mutation operator so that a sufficient amount of exploration occurs. The nature of the image registration problem creates an objective function with many local minima so it is important to mutate out of these valleys. Since the GAOT maximises an objective function, we merely multiply our original objective function by a factor of -1 to perform the minimisation. 2.3.4 Particle Swarm Optimisation After evaluating the performance of a free PSO toolbox, and getting poor results, it was decided to write a custom PSO routine based on the method described in the previous section. The high level description of the algorithm is given below: Swarm = CreateRandomSwarm(); While we have more iterations to go: EvaluateObjectiveFunction(Swarm); If Swarm.BestValue<GlobalBestValue then GlobalBestValue = Swarm.BestValue; GlobalBestPosition=Swarm.BestPosition; End UpdateIndividualAndGroupBestValues(Swarm); CalculateParticleVelocities(Swarm); UpdateParticlePositions(Swarm); End While loop Output GlobalBestPosition; The initial swarm is created with random particles between the bounds of the parameter space. A record is kept of each particle’s best value that it has sampled. Similarly, a group-best is maintained for a social neighbourhood size of 3 particles to either side of the current particle. Modulo indexing is used. The α parameter (described in the previous section) is set to 0.99 so that the swarm does not become unstable and diverge. Both β parameters are set to 0.01 in order that the particles approach the best locations gradually. This samples the objective function many times along the trajectory of the particle. 3. Results 3.1 Optimisation In order to be able to compare the algorithms, the maximum number of function evaluations for each method is set to 1000. The algorithms are stopped as close to this value as possible (since the number of function evaluations may vary from run to run). The amount of time taken by each algorithm is not a good measure of its performance in this case because the amount of processing in the objective function is highly dependent on the parameter values being sampled. All four algorithms described in this paper ultimately approach the global optimum so a suitable measure for their performance is to investigate their precision and accuracy. Each algorithm is run 50 times and an accuracy histogram is calculated for how close and how consistently the algorithm reached the global optimum of (151.5, 151.5, 0.5). The distance is measured geometrically by the following equation: di = (xG − xi )2 + ( yG − yi )2 + (sG − si )2 (5) where di is the distance for the ith run, (xG, yG, sG) are the global optimum parameters and (xi, yi, si) are the optimum parameters as calculated in the ith run. The histogram is divided into 10 equal bins so that the last bin has a distance which is 10 units away from the global optimum. The accuracy histograms for the various methods are given in Figure 5 and the results are analysed below. Higher counts towards bin zero are better. 20 40 15 30 Simplex SA 20 10 10 5 0 1 2 3 4 5 6 7 8 9 101112 30 20 1 2 3 4 5 6 7 8 9 101112 30 GA 10 0 0 20 PSO 10 1 2 3 4 5 6 7 8 9 1011 0 1 2 3 4 5 6 7 8 9 101112 Figure 5 Accuracy Histograms. 3.1.1 Simplex Method This method is especially robust for functions that have peculiar gradients or that are discontinuous, as is common in image registration. The method of travel by the simplex acts as a pseudo-gradient that plays a similar role as in the gradient methods. The success of this algorithm is attributed to the pre-processing stage of the algorithm which selects suitable starting points for the Nelder-Mead optimisation. The histogram shows that when the algorithm is near the global optimum, it is reasonably precise. It does however, find local minima quiet often. 3.1.2 Simulated Annealing It is expected that this algorithm performs better than the Simplex Method described above due to the Metropolis method that it employs. It was found that the performance of the algorithm is significantly improved by restarting when the simplex becomes stuck. Without this behaviour, the algorithm tries oscillatory values which waste precious computation time since the samples do not advance the simplex at all. The histogram shows that the algorithm is the least accurate (repeatable) algorithm of the four but is precise (finds values near to the global optimum many times). 3.1.3 Genetic Algorithm The performance of the GA can be attributed to the number of mutations that take place. This is important for image registration because the objective function has many local optima that vary in shallowness. It is more important for the algorithm to explore the parameter space than it is to improve the precision (using the arithmetic crossover). It was also observed that the mutation rate is more important than the initial population size (which randomly explores the parameter space), since sufficient mutations will explore the space more wisely. The histogram shows a clear binomial distribution near the global optimum. This shows that the GA produces reasonably accurate (consistent) results. 3.1.4 Particle Swarm Optimisation The PSO produces acceptable results because of the group dynamics in the system. The social groups promote exploration of the search space while the individual best position lets each particle improve its precision. Good parameter values were found by investigating what effect they have on the swarm behaviour and then tweaking the values to suite the problem domain. The behaviour of the swarm is predictably based on the algorithm parameters so it is relatively easy to infer good parameter values by watching how the particles swarm in the objective function. The histogram shows that the algorithm finds precise results but not always accurately. 3.2 Other Sample Data The algorithms managed to register the templates of distorted and noisy images of Lena’s face to the image. The various modifications to the templates that were made include, Gaussian blurring, the addition of Gaussian noise, vortex rotation, smudging and non-uniform stretching. The characteristics and accuracy histograms for each algorithm remain relatively consistent with the results described in the previous section. 4. Recommendations The current objective function makes use of sub-pixel sampling to obtain a continuous parameter space. This is very computationally intensive and unnecessary for certain parameters in image registration such as x and y coordinates. The algorithms presented above should be modified so that the user can specify that certain parameters may vary discretely whilst others continuously. This should reduce the amount of time taken for each function evaluation. Another improvement to the algorithms would be to select regions of interest that are likely to contain the template. This requires looking for higher level features in the image first and then creating bounds for the optimisation algorithms. 5. Conclusion This paper introduced evolutionary optimisation methods and contextualised them in the image registration field. It was found that the algorithms perform well in different aspects of the image registration process. The parameters for each method need to be tuned to suite the given image but the behaviour of the methods presented in this paper are intuitive and insight into how to modify parameters can easily be gained by simulating a few initial runs. These evolutionary optimisation methods were compared to the non-evolutionary Nelder-Mead simplex method with starting points selected by doing higher level preprocessing. SA returns the most precise results, while GA returns the most accurate results. PSO is in between these two and it is followed by the Simplex method. Preprocessing should be investigated for the evolutionary methods. 6. References [1] B. Zitova, J. Flusser. “Image registration methods: a survey”; Elsevier B.V.; Image and Vision Computing 21 (2003) 977–1000; June 2003. [2] X. Peng, M. Ding, C. Zhou, Q. Ma. “A practical twostep image registration method for two-dimensional images”; Elsevier B.V.; Information Fusion; Article in press; 2004. [3] S. Chien, S. Sung. “Adaptive window method with sizing vectors for reliable correlation-based target tracking”; Elsevier B.V.; Pattern Recognition 33 (2000) 237}249; 1999. [4] H. Lim, S. Hossein Cheraghi. “An optimization approach to shape matching and recognition”; Computers & Electrical Engineering 24 (1998) 183-200; 1998. [5] Z. Michalewicz. “Genetic Algorithms + Data Structures = Evolution Programs 3rd Ed.”; Springer; New York; 1999. [6] E. Chong, S. Zak. “An Introduction to Optimization, 2nd Ed.”; Wiley; 2001. [7] I. T. Nabney. “Netlab: Algorithms for Pattern Recognition”; Springer; UK; 2002. [8] “Optimization Toolbox Mathworks Inc.; 2002. [9] C. Skiscim. “Optimisation by Simulated Annealing”; Proceedings of the 1983 Winter Simulation Conference; pp 523-535; 1983. [10] C. Houck; J. Joines; M. Kay. "A Genetic Algorithm for Function Optimization: A Matlab Implementation"; NCSU-IE TR 95-09, 1995. [11] J. Kennedy, R. C. Eberhart. “Particle swarm optimization”, Proceedings of the 1995 IEEE International Conference on Neural Networks, vol. 4, pp. 1942–1948. IEEE Press; 1995. [12] W. Press, W. Vetterling, S. Teukolsky, B. Flannery. “Numerical Recipes in C++, 2nd Ed.”; Cambridge University Press; 2002. [13] “Genetic Optimization Toolbox”; Last Accessed: 11/05/2004. http://www.ie.ncsu.edu/mirage/GAToolBox/gaot/ User’s Guide V2”; The
5cs.CE
From Principles to Practice with Class in the First Year Sam Tobin-Hochstadt David Van Horn Northeastern University Boston, Massachusetts, USA {samth,dvanhorn}@ccs.neu.edu We propose a bridge between functional and object-oriented programming in the first-year curriculum. Traditionally, curricula that begin with functional programming transition to a professional, usually object-oriented, language in the second course. This transition poses obstacles for students, and often results in confusing the details of development environments, syntax, and libraries with the fundamentals of OO programming that the course should focus on. Instead, we propose to begin the second course with a sequence of custom teaching languages which minimize the transition from the first course, and allow students to focus on core ideas. After working through the sequence of pedagogical languages, we then transition to Java, at which point students have a strong command of the basic principles. We have 3 years of experience with this course, with notable success. 1 Introduction Many universities and colleges aim to teach their students proficiency in an industrial object-oriented programming language by the end of the students’ first year. The most common approach to achieve this aim is to teach an industrial language in the first course, often Java, starting on the first day. Other curricula take a more indirect route by teaching functional programming in the first semester, followed by a second semester in Java. The latter approach is an improvement over the first, as pointed out by numerous observers [9, 23, 7, 22], but both suffer serious flaws. As an example, Northeastern University teaches introductory programming in the first semester using How to Design Programs [8], followed by object-oriented programming using How to Design Classes [11] in the second semester. This sequence was designed to provide a smooth path for incoming students with a competence in high-school level algebra to reach proficiency in Java by the end of their first year [9]. It was a major improvement over the previous Java-first curriculum in terms of student success, attrition, and preparation for subsequent courses [21]. However, significant problems remain; in particular, the second semester course violates the designers’ own principles (as recalled in [5]): 1. introduce only those language constructs that are necessary to teach programming principles, and 2. choose a language with as few language constructs as possible, and one in which they can be introduced one at a time. The problem is that the first semester ends with an advanced pedagogical functional language and the second semester starts with Java, although it focuses on a side-effect free subset. Despite this focused subset, this transition is too abrupt to meaningfully bridge the gap between functional and object-oriented programming, because several other significant transitions happen in concert: • from a highly regular and minimal syntax to a complicated irregular syntax, • from an untyped language to a typed language, • from a pedagogical programming environment (DrRacket) to a professional programming environment (Eclipse), Philip K.F. Hölzenspies (Ed.): Trends on Functional Programming In Education (TFPIE) EPTCS 136, 2013, pp. 1–15, doi:10.4204/EPTCS.136.1 c S. Tobin-Hochstadt & D. Van Horn 2 From Principles to Practice with Class in the First Year • from a language with numeric values corresponding to mathematical objects to a language with numeric values corresponding to common machine representations,1 • from a language with image literals and graphical libraries to one in which graphical programming is tedious, • from an interaction-oriented language and tool suite to a compiled, batch-oriented language. This abrupt transition has several negative consequences which we have experienced first-hand: the principles of object-oriented programming are obscured and de-emphasized, struggling with the programming environment is frustrating and can cause potentially good students to leave the program, it favors students with prior exposure to the particular tools (a set that is demographically skewed), it inhibits students from experimenting by allowing them to rely upon past skills, and it creates the false impression that courses are discrete units of instruction that can be discarded after successful completion rather than being part of a continuous and cumulative educational experience. We contribute an alternative approach to the second semester that overcomes these problems and provides a gradual introduction to object-oriented programming. Our approach starts the second semester by introducing only the concept of programming with objects, while all other aspects of the course remain where they were left off in the previous semester. This allows other concepts to be introduced at the point at which they are relevant and motivated. Despite this more gradual approach, the course accomplishes the goal of reaching industrial language competence by the end of the semester, covering a super-set of the concepts and topics covered in the How to Design Classes-based course. Outline The remainder of this paper is organized as follows: section 2 provides background on How to Design Programs and the context and constraints involved in the first year at Northeastern. Section 3 describes our approach to the second semester, which starts with a small shift in perspective to bridge the gap between functional programming and object oriented programming. Section 4 describes the path from our pedagogical languages to an industrial object-oriented programming language. Section 5 discusses the relation to existing work and section 6 concludes. 2 Background: the context at Northeastern At Northeastern, the College of Computer & Information Science (CCIS) requires a four course introductory sequence for Computer Science majors in the first year. The first semester features both a course on discrete mathematics and an introduction to programming following the How to Design Programs curriculum. The second semester follows with a course on object-oriented programming and one featuring formal reasoning about programs, both on paper and with the ACL2 theorem prover [17]. After the first year, students take a wide variety of follow-up courses, ranging from a required course in “Object-oriented design” to architecture, operating systems, robotics, and programming languages. No standard language is used in these courses. More significantly, Northeastern distinctively emphasizes experiential education, with almost all Computer Science majors participating in a 6 month “co-op” internship after their third semester. These co-ops take place at a wide variety of companies, and while most students do some software development, there is no single platform or set of tools that commands majority use. In particular, there is wide 1 While this may seem like a minor point, details of numeric representation can crop up quickly in a Java-based course—for example, 1/3 cannot be represented by any Java numeric types. S. Tobin-Hochstadt & D. Van Horn 3 variation in the languages students use while on co-op. This combination sets the constraints under which we designed our approach. 2.1 A first course on How to Design Programs In the first semester, students are introduced to the “design recipe”, a step-by-step process for going from English descriptions of problems to working programs. The design recipe involves six steps: 1. Analyze the information involved in the problem and express how to represent it as data. 2. Write down a function signature, a summary of the purpose of the function, and a function stub. 3. Illustrate the signature and the purpose statement with some functional examples. 4. Take an inventory of the input data that can be used to compute an answer. 5. Write the code for the function. 6. Verify the behavior of the program against the functional examples given earlier. Students explore program design using this process in the context of a series of successively richer pedagogical programming language levels [9, 8] that are included in the DrRacket (formerly DrScheme) programming environment [12]. These languages are called the beginning student language (BSL), the intermediate student language (ISL), and the advance student language (ASL). The language and environment include several tools in support of the design recipe. For example, functional examples can be written as executable tests by writing check-expect expression [14]; an algebraic stepper and REPL are provided to interact with programs at each step of the design process. Finally, the first semester course makes extensive use of a library for developing interactive animations and games using functional programming and functional graphics [10, 4]. 2.2 The goal of the second course After the second course, students should both (a) be prepared for subsequent courses in the curriculum, which expect familiarity with Java and standard Java libraries, (b) be prepared for co-ops in which they will use professional-grade languages and tools which will almost certainly be object-oriented. More significantly, we aim to teach the key insights behind the object-oriented approach to program design. These constraints, while in detail specific to Northeastern and the CCIS curriculum, are broadly similar to the requirements for the first year at many universities. Our course also attends to smaller and more idiosyncratic elements of our curriculum, ranging from formal reasoning to algorithmic analysis, as described in the following sections. 3 A small shift of focus On the first day of the second semester, we introduce a single linguistic concept to an otherwise unchanged context of the previous semester: the idea of an object. An object is a new kind of value that can, as a first cut, be understood as a pairing together of two familiar concepts: data and functionality. • An object is like a structure in that it has a fixed number of fields, thus an object (again, like a structure) can represent compound data. But unlike a structure, an object contains not just data, but functionality too; • An object is like a (set of) function(s) in that it has behavior—it computes; it is not just inert data. From Principles to Practice with Class in the First Year 4 This suggests that objects are a natural fit for well-designed programs since good programs are organized around data definitions and functions that operate over such data. An object, in essence, packages these two things together into a single programming apparatus. This has two important consequences: 1. Students already know how to design programs oriented around objects. Since objects are just the combination of two familiar concepts that students already use to design programs, they already know how to design programs around objects, even if they have never heard the term “object” before. 2. Objects enable new kinds of abstraction and composition. Although the combination of data and functionality may seem simple, objects enable new forms of abstraction and composition. That is, objects open up new approaches to the construction of computations. By studying these new approaches, we can distill new design principles. Because we understand objects are just the combination of data and functionality, we can understand how all of these principles apply in the familiar context of programming with functions. 3.1 The basics of objects To begin with, we introduce the notion of a class definition, which can be thought of at first as a structure definition in that it defines a new class of compound data. A class is defined using the define-class form: (define-class posn (fields x y)) This is similar to the define-struct form of the first semester, used as follows: (define-struct posn (x y)) An object is a value that is a member of this class of data, which can be constructed with the new keyword, a class name, and the appropriate number of arguments for the fields of the object: (new posn 3 4) An object understands some set of messages. Simple structure-like objects understand messages for accessing their fields and messages are sent by using the send keyword, followed by an object, a message name, and some number of arguments: (send (new posn 3 4) x) ;=> 3 (send (new posn 3 4) y) ;=> 4 The send notation is simple, but syntactically heavy. Once students are comfortable with send, we introduce shorthand to make it more convenient, writing (x . m) for (send x m). The dot notation can be nested, so (x . m . n) is shorthand for (send (send x m) n). (The approach of introducing a simple, uniform syntax and later introducing a convenient shorthand that would have been confusing to start with follows the approach of first introducing cons and then later list and quote in the first semester.) It is possible to endow objects with functionality by defining methods, which extend the set of messages an object understands. A method definition follows the same syntax as a function definition, but is located inside of a class definition. Here is a more complete development of the posn class with two methods: S. Tobin-Hochstadt & D. Van Horn 5 ;; A Posn is a (new posn Number Number), ;; which represents a point on the Cartesian plane (define-class posn (fields x y) ;; dist : Posn -> Number ;; Distance between this (check-expect ((new posn (define (dist that) (sqrt (+ (sqr (- (this (sqr (- (this posn and that posn 0 0) . dist (new posn 3 4)) 5) . x) (that . x))) . y) (that . y)))))) ;; dist-origin : -> Number ;; Distance of this posn from the origin (check-expect ((new posn 0 0) . dist-origin) 0) (check-expect ((new posn 3 4) . dist-origin) 5) (define (dist-origin) (this . dist (new posn 0 0)))) This class definition defines a new class of values which are posn objects. Such objects are comprised of two numeric values and understand the messages x, y, dist, and dist-origin. Unit tests have been included with each method definition, following the principles of the design recipe studied in the first semester. Although check-expect forms can appear within class definitions, they are lifted to the top-level when a program is run. Methods can be defined to consume any number of arguments, but they are implicitly parameterized over this, the object that received the message. 3.2 Where did the cond go? Unions, and recursive unions in particular, are a fundamental kind of data definition that students are well-versed in from the previous semester. A fundamental early lesson is how to represent (recursive) unions using classes and how to write recursive methods. As an example, figure 1 defines binary trees of numbers (an archetypal recursive union data definition) using the Beginning Student language (BSL) as used at the start of the first semester, and also using the Class language of our course. The structure of this program is analogous to the approach of the previous semester but this example brings to light an important difference with the functional approach. The method for computing the sum of a leaf is defined in the leaf class, while the method for computing the sum of a node is in the node class. When a tree object is sent the sum method, there is no function with a conditional to determine whether the object is a leaf—instead, the object itself takes care of computing the sum based on its own sum method. This shift in perspective is at the core of object-orientation: objects contain their own behavior and the case analysis previously done in functions is eliminated. 3.3 Worlds and animations At Northeastern, Programming in the first semester is often oriented around interactive event-driven video games. The basic design of a video game involves defining a data representation for states of the game and functions for transitioning between states based on events such as clock ticks, keyboard input, or mouse events. The design of a game thus involves the design of data and operations on that data; From Principles to Practice with Class in the First Year 6 #lang bsl ;; A Tree is one of: ;; - (make-leaf Number) ;; - (make-node Tree Number Tree) (define-struct leaf (v)) (define-struct node (left v right)) ;; sum : Tree -> Number ;; sums the elements of the given tree (define (sum a-tree) (cond [(leaf? a-tree) (leaf-v a-tree)] [else (+ (sum (node-left a-tree)) (node-v a-tree) (sum (node-right a-tree)))])) (check-expect (sum (make-leaf 7)) 7) (check-expect (sum (make-node (make-leaf 1) 5 (make-node (make-leaf 0) 10 (make-leaf 0)))) 16) #lang class/1 ;; A Tree is one of: ;; - (new leaf Number) ;; - (new node Tree Number Tree) ;; and implements ;; sum : -> Number ;; sums the elements of this tree (define-class leaf (fields v) (define (sum) (this . v))) (define-class node (fields left v right) (define (sum) (+ (this . left . sum) (this . v) (this . right .sum)))) (check-expect ((new leaf 7) . sum) 7) (check-expect ((new node (new leaf 1) 5 (new node (new leaf 0) 10 (new leaf 0)))) . sum) 16) Figure 1: Binary tree sum in Beginning Student and in the Class language in other words, the game involves the design of objects. We therefore continue in the second semester with the use of programming video games but supplement the course with a library for doing so in an object-oriented style. Figure 2 gives an example written in both the functional style and object-oriented style. The key difference between these two programs is that the functional program uses the 2htdp/universe library, which provides a big-bang form that consumes the initial state of the world and has a declarative form of associating event-handler functions, while the object-oriented program uses an alternative library developed for our class: class/universe. It also provides a big-bang form but it consumes a single argument, the initial state of the world represented as an object. Event handlers are simply methods of this object; for example, clock ticks trigger the on-tick method. The program on the left is the first program of the first semester, while the one on the right is the first program of the second semester. Our approach is able to make the conceptual connection between functional and object-oriented programming quite clear while appealing to the familiar event-driven interactive programs developed throughout the year. S. Tobin-Hochstadt & D. Van Horn 7 #lang bsl (require 2htdp/image 2htdp/universe) #lang class/1 (require 2htdp/image class/universe) ;; A World is a Number ;; A World is a (new world Number) (define-class world (fields n) ;; on-tick : World -> World (define (tick w) (add1 w)) ;; to-draw : World -> Image (define (draw w) (place-image (circle 10 "solid" "red") w 200 (empty-scene 400 400))) ;; on-key : KeyEvent World -> World (define (on-key k w) 10) (big-bang 10 [to-draw draw] [on-tick tick]) ;; on-tick : -> World (define (on-tick) (new world (add1 (this . n)))) ;; to-draw : -> Image (define (to-draw) (place-image (circle 10 "solid" "red") (this . w) 200 (empty-scene 400 400))) ;; on-key : KeyEvent -> World (define (on-key k) (new world 10))) (big-bang (new world 10)) Figure 2: World programs The move to object-oriented style immediately and naturally leads to designs that are enabled by organizing programs around objects. For example, the state pattern [13] becomes useful almost immediately. The programs in figure 2 animate a rocket (rendered as a circle in this example) taking off. An illustrative follow-up exercise is to animate a rocket that lands. The natural design is to have two variants for states of the rocket: one for descending rockets and one for landed rockets (an example is given in appendix A). While in the functional approach it is easy to use the state-pattern for the data representing a rocket, it is more difficult to have states of behavior. The typical solution adds conditionals to all of the event handlers. In the object-oriented approach, states of behavior are just as natural as data. It is therefore straightforward to design programs with easy to observe invariants such as “a landed rocket never changes position.” In the functional approach, even such simple properties are more involved to establish, because all event handlers must be inspected. This approach also leads naturally to discussion of inheritance. Often programs with multiple states wish to share the implementation of some methods. We first show that this can be accomplished at the cost of minor boilerplate with delegation, and then show how inheritance allows the programmer to avoid duplication and boilerplate entirely. Once inheritance is able to group identical methods, overriding is a natural next step when some but not all of the implementations are identical across the variants. 3.4 Language levels Our introduction to object-oriented programming is built on a series of “language levels”, each of which introduces additional features, adding complexity to the programming model and expressiveness to the programs. Each language is class/N for some N, with features appearing in the following order. From Principles to Practice with Class in the First Year 8 0. Classes and Objects 1. Abbreviated notation for method calls 2. Super classes 3. Overriding 4. Constructors Several commonalities run through all of these languages. First, they are all purely functional; we do not introduce imperative I/O or side-effects until after transitioning to Java in the second half of the course. Second, they all are a super set of the Intermediate Student language from How to Design Programs, meaning that they support higher-order functional programming and lists. One key principle that we adhere to in the design of the language levels is that no features of the language are added purely to support “software engineering” concerns such as specification mechanisms. Not only does that language not support declaring types or contracts, but interfaces are described purely in comments. This is not to say that interfaces and contracts are optional; in fact, they are mandatory. But the focus of the first part of the course is on the fundamentals of object-orientation. Teaching the use of software engineering tools such as type systems, while vital, is a topic which we defer to the second half of the course when we transition to Java. We made this decision after experience in which students were confused about the relationship between explicit interface specifications, type systems, and the informal data definitions and contracts which students are required to write for all methods. After removing interfaces from the language and making them purely a specification construct, this confusion disappeared. 4 From principles to industrial languages The transition from custom teaching languages to a professional language takes place about half-way through the course. At this point, students already have experience with many of the essential concepts of object-oriented programming. In particular: objects, classes, fields and methods, dynamic dispatch, inheritance, and overriding. From this point, almost any language that students might encounter in future co-op positions, summer internships, or upper-level courses would be an appropriate follow-up. Our course transitions to Java, but C#, Python, Ruby, Eiffel, or JavaScript would all work naturally. The key lesson of the transition is that the fundamental principles underlying object-oriented programming remain the same between languages, and that learning a new language is primarily a matter of mapping these concepts to specific constructs in the new language. Of course, particular languages also use unique specific mechanisms which need to be taught to use the language effectively, but these are rarely as vital as the cross-language principles. We chose the half-way point as the time for transition based on experience with earlier versions of this course. In particular, we found that a later transition, while allowing us to present additional concepts in a controlled environment, did not give students sufficient time and experience with Java. Subsequent classes found that students were strong on fundamentals but weak on Java practice. The other alternative, transitioning earlier, would not provide sufficient time to cover the fundamental topics before the transition. S. Tobin-Hochstadt & D. Van Horn 9 4.1 Functional Java The transition begins with replicating the object-oriented style of our teaching languages in Java. In particular, we do not introduce mutation, for loops, or mutable data structures such as arrays or ArrayLists until later in the semester. Instead, students design data representations using classes, with interfaces representing unions of data. Additionally, we avoid mention of the distinction between primitive and other values in Java, which is made easier by not using standard libraries early. An example of this style of programming is presented in figure 3, repeating the binary tree sum from the previous section. Comparing this figure to the previous example illustrates a number of the differences that students are exposed to upon transition to Java. 1. Explicit representation of unions and interfaces in the language. Previously, interfaces were simply described in stylized comments, following the How to Design Programs approach. 2. Types are now specified as part of the program and are (statically) enforced. Data definitions and interfaces can be transformed from the stylized comments into interface definitions and method signatures annotated with types. Students are taught the benefits of type systems, which impose syntactic restrictions sufficient to prove objects meet (the structural aspects of) their interface definitions. Students also encounter the downside of types when they discover the type system cannot always follow valid reasoning about program invariants and may reject perfectly good programs. 3. Java syntax is substantially different and more verbose. For example, constructors must be defined explicitly. 4. The testing environment is somewhat different, and requires additional boilerplate, although we are able to use the JavaLib framework [20] to support testing with structural equality. There are other differences which cannot be seen from a code snippet. 5. Students must use a new development environment and compiler. In class, we primarily develop in a text editor and run the Java compiler at the command line. In labs and on homeworks, students typically use the Eclipse IDE. 6. Installing and configuring libraries is now required. Because we use a custom library for testing, students must cope with library installation and class paths on the first day. All but the first two of these changes are unrelated to the fundamental lessons we hope to teach—the rest merely present additional hurdles for students. However, at this point in the semester, the students are far better equipped to meet these challenges. They are already familiar with objects, classes, and the other concepts we have covered. They are also fully engaged in the class, instead of making the transition in the midst of the transition between semesters. Finally, they have now been programming for 50% longer than they had at the start of the semester. 4.2 Traditional Java Thanks to the preparation in the first half of the course, we can cover OO programming in a functional subset of Java in a just a few lectures. We then increase the subset of the language we use to encompass mutation, loops, and mutable data structures and introduce the underlying design principles. We present ArrayLists, followed briefly by arrays. Students use, and then implement, hash tables as well as other mutable and immutable data structures. Conventional input and output are treated only very briefly, as we focus instead on both fundamentals and exercises making use of real APIs such as hashing functions From Principles to Practice with Class in the First Year 10 import tester.*; interface Tree { // sums the elements of this tree Integer sum(); } class Leaf implements Tree { Integer v; Leaf(Integer v) { this.v = v; } public Integer sum() { return this.v; } } class Node implements Tree { Tree left; Integer v; Tree right; Node(Tree l, Integer v, Tree r) { this.left = l; this.v = v; this.right = r; } public Integer sum() { return this.left.sum() + this.v + this.right.sum(); } } class Examples { void test_tree(Tester t) { t.checkExpect(new Leaf(7).sum(), 7); t.checkExpect(new Node(new Leaf(1), 5, new Node(new Leaf(0), 10, new Leaf(0))).sum(), 16); } } Figure 3: Binary tree sum in the style of How to Design Classes S. Tobin-Hochstadt & D. Van Horn 11 or Twitter posting. Finally, while, for, and for-each loops are presented, following the methodology of How to Design Classes which connects loops to stylized use of recursive functions with accumulators, a technique the students now have used for two semesters. 4.3 Beyond Traditional Java Finally, at the end of the course, we are able to build on the two major segments to examine less-wellexplored topics in object-oriented programming. Typically, we cover the basics of representing objects in a functional language, advanced OO techniques such as mixins and prototypes, and a new OO language such as Ruby or JavaScript. Additionally, we emphasize the ability to embed functional programming in an OO context, using techniques such as the command pattern and the visitor patterns. Again, the key message is the transferability of concepts across languages. 5 Related work Teaching programming principles in a functional style has a long history, with Abelson and Sussman’s Structure and Interpretation of Computer Programs [1] being a prominent example. Our work follows in the tradition of the Program by Design (PbD) project2 (previously known as the TeachScheme! project), which emphasizes a systematic approach to program construction. Since the introduction of functional-first curricula, and more specifically in the Program by Design framework, numerous courses have tackled the problem of transition. Typically they, as we, transition to Java in the second course. We discuss first the approach developed by some of the principal creators of PbD, and then other approaches. 5.1 Program by Design and ProfessorJ The Program by Design project initially focused only on the first course, with the second course typically taught in Java in institution-specific ways. Subsequently, the pedagogical approach was extended to Java, but without the tool support and textbook of the first course. An example of this approach is described by (author?) [5], who presents the experience integrating these courses at Adelphi. He reports that “many of Java’s concepts could be introduced more easily in a second course than a first.” With these lessons in mind, the PbD project set out to apply the lessons of teaching languages and IDE support to Java, as well as to present the approach to object-oriented programming in textbook form. ProfessorJ [15] is the resulting system, accompanying the draft textbook How to Design Classes [11]. In parallel to our course, Northeastern teaches the remainder of its computer science majors following this approach. ProfessorJ and How to Design Classes maintain many of the ideas of the first course. In particular, ProfessorJ brings language levels to Java, in an attempt to smooth the transition for students from the first course and provide more helpful feedback. ProfessorJ is also embedded in the DrRacket IDE, increasing familiarity for the students and supporting tools such as an interactive read-eval-print loop. However, the “day 1” transition from the student languages used with How to Design Programs to ProfessorJ is too abrupt and too large. Most significantly, changing languages from the first semester immediately rather than simply adding a new concept confuses too many issues for students. On the first day of a How to Design Classes-based course, students see object-orientation, a new programming 2 http://www.programbydesign.org/ From Principles to Practice with Class in the First Year 12 paradigm; Java, a new language with new syntax, and a static type system, a crucial but orthogonal concept. In contrast, our course presents just one of these concepts on the first day, but covers all of them by the end of the semester. ProfessorJ also takes on the dual challenges of implementing Java as well as subsetting it. This ultimately resulted in both a limited Java environment as well as the eventual abandoning of the tool since it was too difficult to maintain, let alone keep up with advances in Java. Committing to Java on the first day, regardless of the environment provided to students, has significant limitations. First, the syntactic and semantic heaviness of Java is a burden for beginning students, and discourages interactive and experimental programming. The very first chapter of How to Design Classes discusses the fixed size of Java integers, a topic avoided entirely in the first course. Second, by committing to a particular industrial-strength language, it closes off possibilities in the curriculum. Third, it commits entirely to the new paradigm, making it more difficult for students to compare the approaches. Since ProfessorJ is no longer available, students are faced with an even starker change on the first day. Even with a student-oriented environment such as DrJava or BlueJ [2, 16], students must learn an entirely new tool, along with new libraries. If the course uses a typical professional development environment such as Eclipse, students must also contend with compilation, loss of interactivity, and subtle issues such as classpaths, none of which are fundamental to the concepts that the course focuses on. 5.2 Other transitions Not every curriculum that begins with How to Design Programs transitions to Java after the first course. (author?) [22] describes a second course that includes both more advanced work in Scheme beyond teaching-oriented languages as well as low-level programming in C, taught to computer science majors at University of Waterloo. Radge’s course intentionally does not use student-oriented languages, although the recently-developed C0 language [19] could provide such a language. Other discussions of functional programming in the first year [7] do not discuss the problems of transition. 5.3 Other approaches to Java The problems of teaching Java in introductory courses have been well-explored; we mention only a few related directions here. DrJava [2] and BlueJ [16, 18] are introductory environments for Java, which alleviate some but not all of the drawbacks we have outlined. For example, both of these systems immediately present students with (1) type systems and (2) Java syntax, and (3) do not support the image values and exact numeric values that we rely on in our course. Several teaching-oriented graphics libraries for Java have been proposed [6, 3], but these are significantly more complex than the graphics and interaction libraries we are able to use in the introductory language we present. 6 Experience and outlook We have now completed the third iteration of this course, teaching approximately 35 students each time. Our experience has been uniformly positive, and the students have gone on to significant success in the subsequent courses, despite the curriculum differing from what the bulk of Northeastern University computer science majors take. Anecdotally, the class has also had notable success in the recruitment and S. Tobin-Hochstadt & D. Van Horn 13 retention of female students, as compared to the other versions of the second-semester course. However, the classes are sufficiently different as to make a precise comparison impossible. The course has provided a vantage point to introduce topics that will be taken up later in the curriculum. We present types, contracts, invariants, and properties of functions, all of which tie into both the concurrent course on logic and computation, as well as later classes on formal methods. The emphasis on representation-independence and interfaces both tie into later classes on software engineering, as well as preparing students for algorithms and data structures courses. Finally, the use of interactive and distributed systems connects to later courses on operating systems and networks. Despite our success, much remains to be done. Type systems are a fundamental concept, but their introduction accompanies the rest of Java. Developing a typed version of our introductory languages would allow a smoother introduction of this idea. Our class’s use of Eclipse could also be improved by first transitioning to a pedagogically-oriented Java environment, but we have not evaluated the specific options. Alternatively, introducing Java-like syntax for the teaching languages we have developed would help tease apart the difficult transitions still present in the course. Finally, the Java portion of the class does not continue the use of “World”-style interactive graphical programming, although a version of the “World” library has been developed for Java [20]. Instead, our course focuses on coverage of standard Java libraries, as well as introductory algorithmic and data structure topics. Continuing to use World-style programming in motivating examples might be valuable for continuity between the two halves of the course. Acknowledgments Matthias Felleisen’s approach to pedagogy and passion for undergraduate teaching has inspired this work from the beginning. CCIS Dean Larry Finkelstein entrusted two postdocs with the redesign of a key undergraduate course, which made this experiment possible. Our teaching assistants, Dan Brown, Asumu Takikawa, and Nicholas Labich, as well as the tutors and graders, contributed enormously to the success of our courses. Finally, and most importantly, our students at Northeastern for the last three years have put up with a curriculum in progress, and the opportunity to teach them has been truly rewarding. References [1] Harold Abelson & Gerald J. Sussman (1996): Structure and Interpretation of Computer Programs. MIT Press. [2] Eric Allen, Robert Cartwright & Brian Stoler (2002): DrJava: a lightweight pedagogic environment for Java. SIGCSE Bull. 34(1), pp. 137–141, doi:10.1145/563517.563395. [3] Carl Alphonce & Phil Ventura (2003): Using graphics to support the teaching of fundamental object-oriented principles in CS1. In: Companion of the 18th annual ACM SIGPLAN conference on Object-oriented programming, systems, languages, and applications, OOPSLA ’03, ACM, pp. 156–161, doi:10.1145/949344.949391. [4] Ian Barland, Matthew Flatt & Robby Findler (2010): The Design of a Functional Image Library. In: Workshop on Scheme and Functional Programming (SFP). [5] Stephen A. Bloch (2000): Scheme and Java in the first year. J. Comput. Sci. Coll. 15(5), pp. 157–165. 14 From Principles to Practice with Class in the First Year [6] Kim B. Bruce, Andrea Danyluk & Thomas Murtagh (2001): A library to support a graphics-based object-first approach to CS 1. In: Proceedings of the thirty-second SIGCSE technical symposium on Computer Science Education, SIGCSE ’01 33, ACM, pp. 6–10, doi:10.1145/366413.364527. [7] Manuel M. T. Chakravarty & Gabriele Keller (2004): The risks and benefits of teaching purely functional programming in first year. J. Funct. Program. 14(1), pp. 113–123, doi:10.1017/S0956796803004805. [8] Matthias Felleisen, Robert B. Findler, Matthew Flatt & Shriram Krishnamurthi (2001): How to design programs: an introduction to programming and computing. MIT Press. [9] Matthias Felleisen, Robert B. Findler, Matthew Flatt & Shriram Krishnamurthi (2004): The structure and interpretation of the computer science curriculum. Journal of Functional Programming 14(4), pp. 365–378, doi:10.1017/s0956796804005076. [10] Matthias Felleisen, Robert B. Findler, Matthew Flatt & Shriram Krishnamurthi (2009): A functional I/O system or, fun for freshman kids. In: ICFP ’09 Proceedings of the 14th ACM SIGPLAN International Conference on Functional programming, ACM, pp. 47–58, doi:10.1145/1631687.1596561. [11] Matthias Felleisen, Matthew Flatt, Robert Bruce Findler, Kathryn E. Gray, Shriram Krishnamurthi & Viera K. Proulx (2012): How to Design Classes (Draft). Available at http://www.ccs.neu.edu/home/matthias/htdc.html. [12] Robert B. Findler, John Clements, Cormac Flanagan, Matthew Flatt, Shriram Krishnamurthi, Paul Steckler & Matthias Felleisen (2002): DrScheme: a programming environment for Scheme. JFP 12(02), pp. 159–182, doi:10.1017/s0956796801004208. [13] Erich Gamma, Richard Helm, Ralph Johnson & John Vlissides (1994): Design Patterns: Elements of Reusable Object-Oriented Software. Addison Wesley. [14] Kathryn E. Gray & Matthias Felleisen (2007): Linguistic Support for Unit Tests. Technical Report UUCS-07-013, University of Utah. [15] Kathryn E. Gray & Matthew Flatt (2003): ProfessorJ: a gradual introduction to Java through language levels. In: Companion of the 18th annual ACM SIGPLAN conference on Objectoriented programming, systems, languages, and applications, OOPSLA ’03, ACM, pp. 170–177, doi:10.1145/949344.949394. [16] James I. Hsia, Elspeth Simpson, Daniel Smith & Robert Cartwright (2005): Taming Java for the classroom. SIGCSE Bull. 37(1), pp. 327–331, doi:10.1145/1047124.1047459. [17] Matt Kaufmann, J. Strother Moore & Panagiotis Manolios (2000): Computer-Aided Reasoning: An Approach. Kluwer Academic Publishers. [18] Michael Kölling, Bruce Quig, Andrew Patterson & John Rosenberg (2003): The BlueJ system and its pedagogy. Journal of Computer Science Education 13(4). [19] Frank Pfennig (2011): C0 Reference. Available at http://c0.typesafety.net/doc/c0-reference.pdf. [20] Viera K. Proulx (2012): JavaLib. Available at http://www.ccs.neu.edu/javalib/. [21] Viera K. Proulx & Kathryn E. Gray (2006): Design of class hierarchies: an introduction to OO program design. In: Proceedings of the 37th SIGCSE technical symposium on Computer science education, SIGCSE ’06, ACM, pp. 288–292, doi:10.1145/1121341.1121431. S. Tobin-Hochstadt & D. Van Horn 15 [22] Prabhakar Ragde (2008): The chilling descent: making the transition to a conventional curriculum. In: Proceedings of the 2008 international workshop on Functional and declarative programming in education, FDPE ’08, ACM, pp. 13–20, doi:10.1145/1411260.1411263. [23] Joel Spolsky (2005): The Perils of JavaSchools. Available http://www.joelonsoftware.com/articles/ThePerilsofJavaSchools.html. A Worlds and the State pattern #lang class/1 (require 2htdp/image class/universe) ;; A World is one of ;; - (new landed-world) ;; - (new downworld Number) (define-class landed-world ;; to-draw : -> Image (define (to-draw) (place-image (circle 10 "solid" "red") 390 200 (empty-scene 400 400)))) (define-class downworld (fields n) ;; on-tick : -> World (define (on-tick) (cond [(zero? (this . n)) (new landed-world)] [else (new world (sub1 (this . n)))])) ;; to-draw : -> Image (define (to-draw) (place-image (circle 10 "solid" "red") (this . w) 200 (empty-scene 400 400))) ;; on-key : KeyEvent -> World (define (on-key k) (new world 400))) (big-bang (new downworld 400)) at
6cs.PL
Tropical Vertex-Disjoint Cycles of a Vertex-Colored Digraph: Barter Exchange with Multiple Items Per Agent arXiv:1610.05115v3 [cs.DS] 3 Aug 2017 Timothy Highley 1 Hoang Le 1 1 Department of Mathematics and Computer Science, La Salle University Abstract In a barter exchange market, agents bring items and seek to exchange their items with one another. Agents may agree to a k-way exchange involving a cycle of k agents. A barter exchange market can be represented by a digraph where the vertices represent items and the edges out of a vertex indicate the items that an agent is willing to accept in exchange for that item. It is known that the problem of finding a set of vertexdisjoint cycles with the maximum total number of vertices (MAX-SIZEEXCHANGE) can be solved in polynomial time. We consider a barter exchange where each agent may bring multiple items, and items of the same agent are represented by vertices with the same color. A set of cycles is said to be tropical if for every color there is a cycle that contains a vertex of that color. We show that the problem of determining whether there exists a tropical set of vertex-disjoint cycles in a digraph (TROPICALEXCHANGE) is NP-complete and Log-APX-hard. This is equivalent to determining whether it is possible to arrange an exchange of items among agents such that every agent trades away at least one item. TROPICALMAX-SIZE-EXCHANGE is a similar problem, where the goal is to find a set of vertex-disjoint cycles that contains the maximum number of vertices and also contains all of the colors in the graph. We show that this problem is likewise NP-complete and Log-APX-hard. For the restricted case where there are at most two vertices of each color (corresponding to a restriction that each agent may bring at most two items), both problems remain NP-hard but are in APX. Finally, we consider MAX-SIZE-TROPICALEXCHANGE, where the set of cycles must primarily include as many colors as possible and secondarily include as many vertices as possible. We show that this problem is NP-hard. E-mail addresses: [email protected] (Timothy Highley) [email protected] (Hoang Le) 2 1 Highley, Le Tropical Vertex-Disjoint Cycles Introduction Consider a barter exchange where many different agents each bring multiple items to trade. For each item at the barter exchange, the agent lists which other items may be accepted in trade for that item. Based on the lists, an algorithm determines which items are traded for which other items. Instead of restricting exchanges to one-to-one trades, k-way trades are also permitted, where k is limited only by the number of items at the swap meet. If k-way trades are permitted, the traders potentially give items to and receive items from different agents. This barter exchange can be represented as a directed graph where each item brought to the exchange is a vertex and there are edges from each item to those items that may be accepted in trade for that item. The problem of determining the maximum possible number of items that can be simultaneously exchanged is equivalent to finding a set of vertex-disjoint cycles that maximizes the total number of vertices in the cycles. (Throughout this paper, all cycles are assumed to be simple cycles.) This problem is equivalent to the Assignment Problem, which can be solved in polynomial time. It has also been called MAX-SIZE-EXCHANGE [10]. The decision and optimization versions of the problem are defined as follows. Problem 1a. MAX-SIZE-EXCHANGE (MSE-d) Input: A directed graph G Question: Is there a set of vertex-disjoint cycles that includes every vertex in the graph? Problem 1b. MAX-SIZE-EXCHANGE (MSE-o) Input: A directed graph G Output: A set of vertex-disjoint cycles that includes as many vertices as possible If there are many items, it is very likely that there are multiple solutions that tie for the maximum number of items traded. How should the tie be broken? As a secondary criterion, it may be desirable to choose, from among those itemmaximizing solutions, the solution that maximizes the number of agents who trade away at least one item. This can be accomplished by converting the previously described directed graph into a vertex-colored digraph, with each participant having a unique color. The problem, then, is equivalent to finding a set of vertex-disjoint cycles with the maximum total number of vertices in the cycles, and in the case of ties, finding a solution that secondarily contains the maximum number of colors. A subset of vertices from a vertex-colored graph is called tropical if it includes every color in the vertex-colored graph [3]. We define the decision and optimization versions of the TROPICAL-MAX-SIZEEXCHANGE as follows. 3 Problem 2a. TROPICAL-MAX-SIZE-EXCHANGE (TMaxEx-d) Input: A vertex-colored digraph G Question: Is there a set of vertex-disjoint cycles S such that no other set of vertex-disjoint cycles contains more vertices in the cycles, and for every color in the graph there is a cycle in S that contains a vertex of that color? Problem 2b. TROPICAL-MAX-SIZE-EXCHANGE (TMaxEx-o) Input: A vertex-colored digraph G Output: A set of vertex-disjoint cycles that includes as many colors as possible, subject to the restriction that there does not exist another set of vertexdisjoint cycles that has more vertices If the goal is instead to simply maximize the number of agents who trade away at least one item (with no regard for the total number of items traded away), that is equivalent to the problem of finding a set of vertex-disjoint cycles that collectively contain the maximum number of colors. We define the decision and optimization versions of TROPICAL-EXCHANGE as follows. Problem 3a. TROPICAL-EXCHANGE (TEx-d) Input: A vertex-colored digraph G Question: Is there a set of vertex-disjoint cycles such that for every color there is a cycle that contains a vertex of that color? Problem 3b. TROPICAL-EXCHANGE (TEx-o) Input: A vertex-colored digraph G. Output: A set of vertex-disjoint cycles such that the total number of vertex colors in the cycles is maximized We show that both TMaxEx-d and TEx-d are NP-complete via a reduction from CNFSAT. We show that both TMaxEx-o and TEx-o are Log-APX-hard via a reduction from MIN-DOMINATING-SET. We also consider restricted cases where each agent is permitted to bring at most j items to the barter exchange. We then maximize the number of agents who get to trade an item: Problem 4. j-PER-COLOR-TROPICAL-EXCHANGE (jPC-TEx) Input: A vertex-colored digraph G with at most j vertices of each color Output: A set of vertex-disjoint cycles such that the total number of vertex colors in the cycles is maximized 4 Highley, Le Tropical Vertex-Disjoint Cycles Problem 5. j-PER-COLOR-TROPICAL-MAX-SIZE-EXCHANGE (jPC-TMaxEx) Input: A vertex-colored digraph G with at most j vertices of each color Output: A set of vertex-disjoint cycles that includes as many colors as possible, subject to the restriction that no other set of vertex-disjoint cycles has more vertices These problems can be trivially solved in polynomial time if j = 1, but even if j = 2, a reduction from MAX-2-SAT shows that these problems are NP-hard. They are, however, in APX. Finally, we define MAX-SIZE-TROPICAL-EXCHANGE (MaxTEx), which reverses the criteria of TMaxEx-o. In TMaxEx-o, the first criterion is maximizing the total number of vertices in the cycles and the second criterion is maximizing the total number of colors. MaxTEx first maximizes the total number of colors in the cycles and secondarily maximizes the total number of vertices. We show that MaxTEx is NP-hard and Log-APX-hard. We define the optimization version of MaxTEx as follows. Problem 6. MAX-SIZE-TROPICAL-EXCHANGE (MaxTEx) Input: A vertex-colored digraph G. Output: A set of vertex-disjoint cycles S such that no other set of vertexdisjoint cycles contains more colors, and no other set with the same number of colors contains more vertices. Barter exchanges as described in this paper regularly occur online. For example, since 2006 an average of 10 barter exchanges per month have been organized at boardgamegeek.com, with a total of over 500,000 items offered for trade. When participants are informed of the results, they exchange items either in person or by mailing items to each other. Usually, a participant will give an item to one person and receive an item from a different person. Maximizing the number of items traded is the primary goal, and is accomplished using a solution to the Assignment Problem. Maximizing the number of users that trade at least one item is the secondary goal, and this is addressed by solving the Assignment Problem multiple times and after a set number of iterations selecting the solution that yields the greatest number of users trading. However, there is no guarantee that the chosen solution actually maximizes the number of users who trade at least one item. This paper demonstrates that, in general, the problem of maximizing the number of users trading cannot be solved in polynomial time unless P = N P . Other websites, such as barterquest.com, swapdom.com, netcycle.com, and rehash.com, have offered similar services in a more general context. 2 Related Work The Assignment Problem is the minimum weight bipartite perfect matching problem. It is a well-known fundamental optimization problem that can be 5 solved in polynomial time. The polynomial bound was first shown by Jacobi [21], but was not widely known until Kuhn’s publication of the Hungarian Algorithm [24]. More efficient algorithms for the Assignment Problem have been discovered since then [15] [25]. It is known that MAX-SIZE-EXCHANGE can be solved by reducing it to the Assignment Problem [1] [10]. MAX-SIZE-EXCHANGE and related problems have been studied extensively in the context of kidney exchange programs. In kidney exchange programs, it is impractical to allow cycles of unbounded length because it is important for all surgeries to happen at the same time. As a result, most work related to kidney exchanges addresses the MAX SIZE ≤ k-WAY EXCHANGE problem, which is APX-complete when k ≥ 3 [1] [10]. Work on the kidney exchange problem includes investigation of approximability [9], techniques to mitigate failure after matches have been determined [16], dynamic exchanges [29], and incentivecompatible mechansims to ensure that hospitals publish all incompatible pairs [7]. The problems discussed in this paper differ from the kidney exchange problem because we are not concerned with the length of the cycles, and because of the introduction of colored vertices. We explore problems that are similar to the Assignment Problem and kidney exchange problem but based on vertex-colored graphs. Vertex-colored graphs have also been called tropical graphs [20]. As mentioned earlier, a subset of vertices from a vertex-colored graph is called tropical if it includes every color in the vertex-colored graph. Tropical connected subgraphs, tropical dominating sets, and tropical homomorphisms have been studied [3] [4] [20]. In this paper, we are concerned with sets of vertex-disjoint cycles. We say that a set of cycles is tropical if every color in the vertex-colored graph is represented by a vertex in at least one cycle. We study the total number of vertices and colors in a set of vertex-disjoint cycles, but not the total number of cycles in the set. Cycle packing is concerned with the total number of cycles, and it has been studied extensively for both edge-disjoint and vertex-disjoint cycles [11] [23] [28]. There is other work on graphs and colors that is not directly related to our work, but we provide representative citations to help readers delineate the differences. For instance, Fellows et al. and Dondi et al. explored patternmatching in vertex-colored graphs [17] [19]. Coudert et al. explored complexity and approximability properties of edge-colored graphs [13]. Our work assumes that the color of a vertex is an inherent property of the graph. Graph-coloring is a separate topic where colors are added to a graph such that adjacent nodes have different colors. The Four-Color Theorem is a famous result from graph-coloring [5] [6]. Coloring graphs are graphs where each node represents a possible coloring of another graph [8]. More recent graph-coloring work includes the search for rainbow connections [12] [27] [26]. Colorful paths are paths that are both rainbow paths (all vertices in the path have different colors) and tropical paths (all colors in the graph are represented in the path) [2]. 6 Highley, Le Tropical Vertex-Disjoint Cycles 3 CNFSAT as a Vertex-Colored Graph Consider an instance of CNFSAT with q clauses. Let x1 , ..., xn be the n variables in the expression. For each variable, create a vertex. Give all of these vertices the same color, and refer to them as the“variable vertices.” On each vertex, create two edges that loop back to the same vertex. Label one edge TRUE and one edge FALSE on each vertex. Assign each clause of the CNFSAT instance a separate color: c1 , c2 , ..., cq . For each literal in the clause, create a vertex that has the clause’s color. Refer to these vertices as “literal vertices.” For negative literals, put the corresponding vertex on the FALSE loop of the variable vertex for that variable by removing one edge from that loop and adding the new literal vertex along with two edges so that the FALSE loop remains a single cycle from the variable vertex back to itself. Similarly, for positive literals put the literal vertex on the TRUE loop of the variable vertex for that variable. Figure 1 depicts an example graph for a given CNFSAT instance. c1 Variables … … … … … … … … c2 c3 c4 c5 TRUE w … … FALSE TRUE x … … FALSE TRUE y … … FALSE TRUE z … … FALSE 𝑥˅𝑦˅𝑧 ˄ 𝑥˅𝑦˅𝑧 ˄ 𝑤˅𝑧 ˄ 𝑤˅𝑦 ˄ 𝑤˅𝑥 Figure 1: CNFSAT instance represented as a vertex-colored directed graph. 4 CNFSAT reduces to TROPICAL-EXCHANGE Constructing a graph for a CNFSAT instance can clearly be accomplished in polynomial time. Each variable only adds one vertex and two edges to the graph, and each clause only adds a vertex and an edge for each literal in the clause. 7 Based on the way that the graph is constructed, there can be at most n vertex-disjoint cycles. Each variable vertex is part of two cycles (the TRUE loop and the FALSE loop), so a set of vertex-disjoint cycles can only contain one or the other. There are no other cycles in the graph. The cycle that is chosen for the set of vertex-disjoint cycles corresponds to the truth value of that variable in the expression. If we have a polynomial time algorithm for TROPICAL-EXCHANGE, then solving TROPICAL-EXCHANGE will also solve CNFSAT in polynomial time. This is because each clause’s color will be in a cycle if and only if a variable is set to a truth value that satisfies that clause. If all colors can be in the set of vertex-disjoint cycles, then all clauses and thus the entire expression is satisfiable. Since CNFSAT is NP-complete, TROPICALEXCHANGE is NP-hard. TROPICAL-EXCHANGE is also in NP (it is trivial to verify that all colors are in the solution), so TROPICAL-EXCHANGE is NPcomplete. 5 CNFSAT reduces to TROPICAL-MAX-SIZE-EXCHANGE MAX-SIZE-EXCHANGE can be solved in polynomial time, but TROPICALEXCHANGE is NP-complete. What about TROPICAL-MAX-SIZE-EXCHANGE (TMaxEx-d)? That is, if we restrict the search space to only those sets of cycles that maximize the total number of vertices in the cycles, then does that make it easier to determine whether it is possible for a set of vertex-disjoint cycles to cover all of the colors in the graph? The previous reduction does not work for TMaxEx-d due to the fact that solving TMaxEx-d for the constructed graph does not solve CNFSAT. Because TMaxEx-d first maximizes the total number of vertices in the cycles, it may be the case that the CNFSAT instance is satisfiable even though the solution to TMax-Ex-d for the graph does not include all of the colors. That would be the case if the only way to include all of the colors in the graph in the cycles is to accept a lower total number of vertices in the cycles. We can adjust the graph by adding “balance vertices” that all have the same color (the “balance color”). First, we identify the cycle with the most vertices. In the case of a tie, we choose any of the tied cycles. We add one balance vertex to this cycle and then add balance vertices to all other cycles so that every cycle has the same number of vertices. These balance vertices accomplish two tasks: • The number of vertices in each cycle has no impact on the solution produced by TMaxEx-d. With the addition of the balance vertices, every cycle has the same number of vertices, so every possible combination of vertex-disjoint cycles has the same total number of vertices, assuming one cycle per variable vertex. As a result, the secondary criterion (number of colors in the set of cycles) determines the solution. • The balance color has no impact. Because every cycle has a balance vertex, 8 Highley, Le Tropical Vertex-Disjoint Cycles the balance color is guaranteed to be in the set of vertex-disjoint cycles, and the balance color will not have an impact on the final solution. If all possible combinations of vertex-disjoint cycles have the same number of total vertices, then by the same logic described for the previous reduction, the solution to TMaxEx-d for the constructed graph will also solve CNFSAT. Figure 2 depicts an example graph for a given CNFSAT instance. c1 Variables … … … … … … … … c2 c3 c4 c5 Balance Vertices TRUE w … … FALSE TRUE x … … FALSE TRUE y … … FALSE TRUE z … … FALSE 𝑥˅𝑦˅𝑧 ˄ 𝑥˅𝑦˅𝑧 ˄ 𝑤˅𝑧 ˄ 𝑤˅𝑦 ˄ 𝑤˅𝑥 Figure 2: CNFSAT graph with the addition of Balance Vertices. TMaxEx-d is easily shown to be in NP. A proposed solution to an instance of TMaxEx-d can be verified by first solving MAX-SIZE-EXCHANGE for the same graph, which can be accomplished in polynomial time. The next step is verifying that the proposed solution to TMaxEx-d and the solution produced by MAXSIZE-EXCHANGE have the same number of vertices in their respective sets of vertex-disjoint cycles. Assuming they do, it is simple to verify that all colors appear in the proposed solution. Therefore, TMaxEx-d is NP-complete. 6 TROPICAL-EXCHANGE and TROPICAL-MAX-SIZEEXCHANGE are Log-APX-hard A set of vertices is a dominating set of a graph if every vertex in the graph is either in the set or adjacent to a vertex in the set. MIN-DOMINATING-SET is known to be Log-APX-complete, since there exist L-reductions in both directions between MIN-DOMINATING-SET and MIN-SET-COVER [22], and MINSET-COVER is known to be Log-APX-complete [18]. 9 MIN-DOMINATING-SET ≤strict TEx-o. To reduce to TEx-o, for each vertex in an instance of MIN-DOMINATING-SET, create a vertex in an instance of TExo. Call this vertex a mirror, and give all of these mirror vertices in TEx-o the color black. On each mirror vertex, create two cycles (initially just loops). Label one cycle IN and the other NOT-IN. IN and NOT-IN respectively represent a vertex’s inclusion in or exclusion from the dominating set. For each vertex in MIN-DOMINATING-SET, designate two unused colors. The purpose of these two colors is to indicate that the vertex is dominated. Call them the domination colors. On the IN cycle of its mirror in TEx-o, add two new vertices: one for each of its domination colors. Also, for each adjacent vertex in the MINDOMINATING-SET graph, add two vertices to the IN cycle of its mirror in TEx-o, and give the new vertices the two domination colors. Finally, add a “NOT-IN” vertex to the NOT-IN cycle of each mirror vertex. Give each NOT-IN vertex its own unique color. A solution to TEx-o will choose cycles that include as many colors as possible, so the domination colors will certainly be included. It is clearly possible to include all domination colors by simply taking the IN cycle for every mirror vertex. Black is included automatically, and thus the best solution is determined simply by including as many of the unique NOT-IN colors as possible while still including all of the domination colors. If a NOT-IN color is included, that indicates that its vertex is not in the dominating set. Obviously, by maximizing the number of vertices not in the dominating set, the number of vertices in the dominating set is minimized. Note that every correct TEx-o solution will include all of the domination colors, because if any domination color were missing, two domination colors could be added by changing the cycle for the appropriate mirror vertex from its NOT-IN cycle to its IN cycle. That would remove the unique NOT-IN color, but two domination colors would be added for a gain of one additional color in the set of cycles. Similarly, MIN-DOMINATING-SET reduces to TMaxEx-o. The only change to the previous reduction is that black balance vertices are added to the NOT-IN cycles in order to give every IN cycle and NOT-IN cycle the exact same number of vertices. To illustrate the reduction, Figure 3 presents a graph that is an instance of MIN-DOMINATING-SET, and Figure 4 presents the corresponding instance of TMaxEx-o. The number of vertices in the MIN-DOMINATING-SET solution is exactly the same as the number of colors excluded from the TEx-o or TMaxEx-o solution. If there are n vertices in the MIN-DOMINATING-SET problem, then there are 3n + 1 colors in the corresponding TEx-o or TMaxEx-o problem: black, along with two domination colors and one NOT-IN color for each vertex. Therefore, if x is the optimal solution to MIN-DOMINATING-SET and y is the number of vertices in the solution proposed by the reduction, then the performance ratio is y/x, which is greater than the performance ratio for TEx-o or TMaxEx-o: (3n + 1 − x)/(3n + 1 − y). Hence, this reduction is a strict reduction [14]. Because MIN-DOMINATING-SET is Log-APX-hard, TEx-o and TMaxEx-o are also Log-APX-hard. 10 Highley, Le Tropical Vertex-Disjoint Cycles Figure 3: An instance of MIN-DOMINATING-SET Figure 4: MIN-DOMINATING-SET instance from Figure 3 reduced to TMaxEx-o 11 7 TROPICAL-EXCHANGE with only 2 vertices per color We now consider the problems jPC-TEx and jPC-TMaxEx, which are special cases of TROPICAL-EXCHANGE and TROPICAL-MAX-SIZE-EXCHANGE, respectively. For these problems, there can be at most j vertices of each color. In terms of a barter exchange, this corresponds to an organized exchange where each person is permitted to bring at most j items. We will show that these restricted cases are in APX and, unless j = 1, they are NP-hard. There is a trivial algorithm to show that both 2PC-TEx and 2PC-TMaxEx are in APX. Simply solve MAX-SIZE-EXCHANGE for the given graph, completely ignoring vertex colors. Because there are only two vertices in the graph for each color, the solution to MAX-SIZE-EXCHANGE is guaranteed to have at least half as many colors as the maximum number of colors possible in a solution to 2PC-TEx or 2PC-TMaxEx. Therefore it is a 2-approximation. Similarly, a j-approximation can be attained for any j, so jPC-TEx and jPC-TMaxEx are in APX for all j. Consider an instance of the NP-hard problem MAX-2-SAT. Build a graph in polynomial time as described in Section 3, but instead of making all variable vertices the same color, assign a unique color to each variable vertex. This graph has at most two vertices of any given color, so it is an instance of 2PC-TEx. A solution of 2PC-TEx will include as many colors as possible. If a clause’s color is included in the solution, that corresponds to a satisfied clause, so MAX-2-SAT reduces to 2PC-TEx, which means both are NP-hard. Similarly, 2PC-TMaxEx can be shown to be NP-hard with a reduction from MAX-2-SAT. As before, create a graph to represent the MAX-2-SAT instance, giving each variable vertex a unique color. Add balance vertices and give each balance vertex its own unique color. Finally, add an additional cycle that includes a vertex for each of the balance vertex colors. Each balance vertex color appears only twice: once in a TRUE or FALSE cycle and once in the balance vertex cycle. Due to the balance vertex cycle, every balance vertex color is guaranteed to appear in the solution, so those colors do not affect any decisions regarding whether to take a TRUE cycle or a FALSE cycle. As a result, every combination of TRUE and FALSE cycles has the same number of vertices in the cycles, and a solution to 2PC-TMaxEx will simply maximize the number of colors in the cycles. Maximizing the number of colors in the solution corresponds to maximizing the number of satisfied clauses. Thus, 2PC-TMaxEx is also NP-hard. Figure 5 depicts the reduction of an instance of MAX-2-SAT to jPC-TMaxEx. Because 2PC-TEx and 2PC-TMaxEx are NP-hard, jPC-TEx and jPC-TMaxEx are also NP-hard for all j > 2. 12 Highley, Le Tropical Vertex-Disjoint Cycles Figure 5: MAX-2-SAT instance reduced to TMaxEx. 8 MAX-SIZE-TROPICAL-EXCHANGE is NP-Hard and Log-APX-hard For completeness, we consider MAX-SIZE-TROPICAL-EXCHANGE (MaxTEx), where the primary criterion is maximizing the total number of colors in the set of vertex-disjoint cycles, and the secondary criterion is maximizing the number of vertices. This simply reverses the criteria of TMaxEx-o. It is easy to see that a solution to MaxTEx simultaneously solves TEx-o. One only needs to observe the number of colors in the solution to MaxTEx. Thus, MaxTEx is NP-hard and Log-APX-hard. 9 Conclusion In this paper, we have defined and analyzed problems that have practical application in the area of algorithmically arranged barter exchanges. We have shown that TROPICAL-EXCHANGE and TROPICAL-MAX-SIZE-EXCHANGE are NPcomplete and Log-APX-hard, and that MAX-SIZE-TROPICAL-EXCHANGE is NP-hard and Log-APX-hard. When instances of TROPICAL-EXCHANGE and TROPICAL-MAX-SIZE-EXCHANGE are restricted to j vertices per color (jPCTEx and jPC-TMaxEx, respectively), the optimization problems remain NP-hard if j > 1, but are in APX. 13 References [1] D. J. Abraham, A. Blum, and T. Sandholm. Clearing algorithms for barter exchange markets: Enabling nationwide kidney exchanges. In Proceedings of the 8th ACM Conference on Electronic Commerce, EC ’07, pages 295– 304, New York, NY, USA, 2007. ACM. doi:10.1145/1250910.1250954. [2] S. Akbari, V. Liaghat, and A. Nikzad. Colorful paths in vertex coloring of graphs. Electronic Journal of Combinatorics, 18(1), 2011. [3] J.-A. Angles d’Auriac, C. Bujtás, H. El Maftouhi, N. Narayanan, L. Rosaz, J. Thapper, and Z. Tuza. Tropical dominating sets in vertex-coloured graphs. In Proceedings of WALCOM: Algorithms and Computation: 10th International Workshop. Springer, 2016. [4] J.-A. Angles d’Auriac, H. El Maftouhi, S. Legay, N. Cohen, A. Harutyunan, and Y. Manoussakis. Connected tropical subgraphs in vertexcoloured graphs. Discrete Mathematics and Theoretical Computer Science, 17(3):327–348, 2016. [5] K. Appel and W. Haken. Every planar map is four colorable. Part i: Discharging. Illinois Journal of Mathematics, 21(3):429–490, 1977. [6] K. Appel, W. Haken, and J. Koch. Every planar map is four colorable. Part ii: Reducibility. Illinois Journal of Mathematics, 21(3):491–567, 1977. [7] I. Ashlagi, F. Fischer, I. A. Kash, and A. D. Procaccia. Mix and match: A strategyproof mechanism for multi-hospital kidney exchange. Games and Economic Behavior, 91:284–296, 2015. [8] J. Beier, J. Fierson, R. Haas, H. M. Russell, and K. Shavo. Classifying coloring graphs. Discrete Mathematics, 339:2100–2012, 2016. doi:10. 1016/j.disc.2016.03.003. [9] P. Biró and K. Cechlárová. Inapproximability of the kidney exchange problem. Information Processing Letters, 101(5):199–202, 2007. [10] P. Biro, D. F. Manlove, and R. Rizzi. Maximum weight cycle packing in directed graphs, with application to kidney exchange programs. Discrete Mathematics, Algorithms and Applications, 1(04):499–517, 2009. [11] H. L. Bodlaender, S. Thomassé, and A. Yeo. Kernel bounds for disjoint cycles and disjoint paths. In European Symposium on Algorithms, pages 635–646. Springer, 2009. [12] G. Chartrand, G. L. Johns, K. A. McKeon, and P. Zhang. Rainbow connection in graphs. Mathematica Bohemica, 133(1):85–98, 2008. [13] D. Coudert, P. Datta, S. Pérennes, H. Rivano, and M.-E. Voge. Shared risk resource group complexity and approximability issues. Parallel Processing Letters, 17(02):169–184, 2007. 14 Highley, Le Tropical Vertex-Disjoint Cycles [14] P. Crescenzi. A short guide to approximation preserving reductions. In Proceedings of the 12th Annual IEEE Conference on Computational Complexity, CCC ’97, pages 262–, Washington, DC, USA, 1997. IEEE Computer Society. URL: http://dl.acm.org/citation.cfm?id=791230.792302. [15] S. Daitch and D. Spielman. Faster approximate lossy generalized flow via interior point algorithms. In Proceedings of the Fortieth Annual ACM Symposium on Theory of Computing, pages 451–460. ACM, 2008. [16] J. P. Dickerson, A. D. Procaccia, and T. Sandholm. Failure-aware kidney exchange. In Proceedings of the fourteenth ACM conference on Electronic commerce, pages 323–340. ACM, 2013. [17] R. Dondi, G. Fertin, and S. Vialette. Complexity issues in vertex-colored graph pattern matching. Journal of Discrete Algorithms, 9(1):82–99, 2011. [18] B. Escoffier and V. T. Paschos. Completeness in approximation classes beyond apx. Theoretical Computer Science, 359(1):369 – 377, 2006. URL: http://www.sciencedirect.com/science/article/pii/ S0304397506003148, doi:http://dx.doi.org/10.1016/j.tcs.2006. 05.023. [19] M. R. Fellows, G. Fertin, D. Hermelin, and S. Vialette. Upper and lower bounds for finding connected motifs in vertex-colored graphs. Journal of Computer and System Sciences, 77(4):799–811, 2011. doi:doi:10.1016/ j.jcss.2010.07.003. [20] F. Foucaud, A. Harutyunyan, P. Hell, S. Legay, Y. Manoussakis, and R. Naserasr. The complexity of tropical graph homomorphisms. CoRR, abs/1607.04777, 2016. URL: http://arxiv.org/abs/1607.04777. [21] C. Jacobi. About the research of the order of a system of arbitrary ordinary differential equations, 2016. [Online; accessed 6-October-2016]. URL: http: //www.lix.polytechnique.fr/~ollivier/JACOBI/jacobiEngl.htm. [22] V. Kann. On the approximability of NP-complete optimization problems. PhD thesis, Royal Institute of Technology Stockholm, 1992. [23] M. Krivelevich, Z. Nutov, M. R. Salavatipour, J. V. Yuster, and R. Yuster. Approximation algorithms and hardness results for cycle packing problems. ACM Transactions on Algorithms (TALG), 3(4):48, 2007. [24] H. Kuhn. The Hungarian Method for the assignment problem. Naval Research Logistics Quarterly, 2:83–97, 1955. [25] Y. Lee and A. Sidford. Path√finding methods for linear programming: Solving linear programs in O( rank) iterations and faster algorithms for maximum flow. In 2014 IEEE 55th Annual Symposium on Foundations of Computer Science (FOCS), pages 424–433. IEEE, 2014. doi:10.1109/ FOCS.2014.52. 15 [26] X. Li and Y. Shi. Rainbow connection in 3-connected graphs. Graphs and Combinatorics, 29(5):1471–1475, 2013. [27] X. Li and Y. Sun. Rainbow connections of graphs. Springer Science & Business Media, 2012. [28] J. P. Pedroso. Maximizing expectation on vertex-disjoint cycle packing. In International Conference on Computational Science and Its Applications, pages 32–46. Springer, 2014. [29] M. U. Ünver. Dynamic kidney exchange. The Review of Economic Studies, 77(1):372–414, 2010.
8cs.DS
Hypergraph Two-coloring in the Streaming Model∗ arXiv:1512.04188v1 [cs.DS] 14 Dec 2015 Jaikumar Radhakrishnan† Saswata Shannigrahi ‡ Rakesh Venkat § Abstract We consider space-efficient algorithms for two-coloring n-uniform hypergraphs H = (V, E) in the streaming model, whenpthe hyperedges arrive one at a time. It is known that any such hypergraph with at most 0.7 lnnn 2n hyperedges has a two-coloring [4], which can be found deterministically in polynomial time, if allowed full access to the input. • Let sD (v, q, n) be the minimum space used by a deterministic (one-pass) streaming algorithm p that on receiving an n-uniform hypergraph H on v vertices and q hyperedges (q ≤ 0.7 lnnn 2n ), produces a proper two-coloring of H. We show that sD (n2 , q, n) = Ω(q/n). • Let sR (v, q, n) be the minimum space used by a randomized (one-pass) streaming algorithm that on receiving an n-uniform hypergraph H on v vertices and q hyperedges with high probability p n n produces a proper two-coloring of H (or declares failure). We show that 1 sR (v, 10 ln n 2 , n) = O(v log v).  2  n2 , sR nt , 2n−2 exp( 8t ), n = O(n2 /t); in particular, • We show that for any 4 ≤ t ≤ 2n−1 this shows that every n-uniform hypergraph with at most hyperedges is two-colorable. n2 t vertices and 2n−1 exp( 8t ) The above results are inspired by the study of the number q(n), the minimum possible number of p hyperedges in a n-uniform hypergraph that is not two-colorable. It is known that q(n) = Ω( lnnn 2n ) and q(n) = O(n2 2n ). The lower bound (due to Radhakrishnan and Srinivasan [4]) has a corresponding algorithm to deterministically produce the two-coloring; the upper bound due to Erdős is obtained by picking about n2 2n hyperedges randomly from a vertex set of n2 vertices. Our first result shows that no efficient deterministic streaming algorithm can match the performance of the algorithm in [4]; the second result shows that there is, however, an efficient randomized algorithm for the task; the third result shows that that if the number of vertices is substantially smaller than n2 , then every non-two-colorable hypergraph has significantly more than n2 2n hyperedges. 1 Introduction A hypergraph H = (V, E) is a set system E defined on a finite universe V (called the vertex set). The sets in the set system are called hyperedges. We consider n-uniform hypergraphs, that is, hypergraphs whose hyperedges all have n elements. We say that a hypergraph is two-colorable, or has Property B, if there is an assignment of colors χ : V → {red, blue} to the vertex set V such that ∗ A part of this work was reported earlier in the conference paper, Streaming Algorithms for 2-Coloring Uniform Hypergraphs by Jaikumar Radhakrishnan and Saswata Shannigrahi, WADS (2011): 667-678. † Tata Institute of Fundamental Research, Mumbai. E-mail: [email protected] ‡ Indian Institute of Technology, Guwahati. E-mail: [email protected] § Tata Institute of Fundamental Research, Mumbai. E-mail: [email protected] 1 every hyperedge e ∈ E(H) has a vertex colored red and a vertex colored blue, that is, no hyperedge is monochromatic. Note that the hypergraph two-coloring problem can be viewed as a constraint satisfaction problem where the clauses are of a specific kind (the Not-All-Equal predicate). In the special case of graphs (which we may view as two-uniform hypergraphs) two-colorability is easy to characterize and establish: the graph is two-colorable if and only if it does not have an odd cycle; one can find a two-coloring in linear time and O(|V |) space with random access to the input; if the edges are streamed one at a time, a two-coloring can be constructed with O(|V | log |V |) space and O((|E| + |V |) log2 |V |) bit operations. For the case of hypergraphs, the situation is not as simple. Using the probabilistic method, Erdős [2] showed that any n-uniform hypergraph with fewer than 2n−1 hyperedges is two-colorable: a random two-coloring is valid with positive probability; furthermore, this randomized method can be derandomized using the method of conditional probabilities. Erdős [2] later showed that there are n-uniform hypergraphs with Θ(n2 ) vertices and Θ(n2 2n ) hyperedges that are not two-colorable. Both these bounds remained unchanged for some time, until it was first improved by Beck [1], and further improved p n n by Radhakrishnan and Srinivasan [4], who showed that any hypergraph with fewer than 0.7 ln n 2 hyperedges is two-colorable. They also provided a polynomial-time randomized algorithm (and its derandomization) for coloring such hypergraphs. The algorithm has one-sided error, i.e., with some small probability δ that can be made arbitrarily small, it declares failure, but otherwise it always outputs a valid two-coloring. Recently, Cherkashin and Kozik [13] showed that a remarkably simple randomized algorithm achieves this bound. Erdős and Lovász [3] have conjectured that any hypergraph with fewer than n2n hyperedges is two-colorable. In related work, Achlioptas et al. [9] studied when, in terms of the number of hyperedges, does a randomly chosen n-uniform hypergraph stop being two-colorable, and relate this to similar questions for random nSAT. The above results are formally stated in terms of bounds on the number q(n), the minimum number of hyperedges in a non-two-colorable hypergraph. In its original description, the delayed recoloring algorithm of Radhakrishnan and Srinivasan [4] assumes that the entire hypergraph is available in memory, its vertices and hyperedges and may be accessed quickly. Unfortunately, the number of hyperedges can be exponentially larger than the number of vertices. For example, the number of hyperedges can be as high as Ω(2n ) even when there are just O(n) vertices, and it may be unrealistic to expect the entire hypergraph to be available in the main memory. In this paper, we ask if the performance of the delayed recoloring algorithm can be replicated when the work space of the algorithms is limited to (say) a polynomial in the size of vertex set. To study this question, we consider the hypergraph two-coloring problem in the streaming model, where the data arrives as a stream or is stored in external memory, and an algorithm with limited local work space analyzes it by making a small number of sequentially passes over it. The resources whose use one tries to minimize in this model are: the number of passes over the data, the amount of local memory required, and the maximum processing time for any data item. On receiving the complete data set, the algorithm must decide as fast as possible either to start another pass or stop with an output. The algorithm could be either deterministic or randomized, and in the latter case, its output needs to be accurate with high probability over the algorithm’s internal coin tosses. A number of important algorithms have been developed in the streaming model of computation, e.g., estimating frequency moments [8] and heavy hitters [12] of a data stream. These algorithms find applications in the context of network routing, where large amount of data flows through a router but each of them has limited memory. 2 Further motivation for our work comes from the semi-streaming model for graphs that has recently been considered widely in the literature. Proposed by Muthukrishnan [17], this model looks at solving fundamental problems on graphs such as Max-Cut, s − t-connectivity, shortest paths, etc. (see, for e.g. the works [18, 19, 20] and references therein) when the edges are streamed in one at a time. The algorithm only has O(|V |poly log(|V |) bits of workspace, and limited number of passes over the edge-stream, which means that it cannot store the entire graph in memory. Our model for hypergraphs is an extension of this setting, when we have a stream of hyperedges coming in. The parameter of interest to us is the uniformity n of the hypergraph, and we assume the number of vertices is poly(n): this is the natural regime in which all the results known for two-coloring hypergraphs are stated. The hyperedges are made available one at a time as a list of n vertices each, where every vertex is represented using B bits. Furthermore, we assume that poly(n) bits (equivalently, poly(|V |) bits are available in the algorithm’s work space. Notice that the number of hyperedges could now be exponential in n (or subexponential in |V |), and hence we allow for a space that is poly(|V |), rather than close to linear in the number of vertices like in the semi-streaming model for graphs. 1.1 Deterministic streaming algorithms: In the first part of the paper, we investigate the space requirements of deterministic streaming algorithms for the hypergraph two-coloring problem. It seems reasonable to conjecture that any deterministic two-coloring one-pass algorithm must essentially store all the hyperedges before it can arrive at a valid coloring. Let sD (v, q, n) be the minimum space used by a deterministic (one-pass) streaming algorithm that on receiving an n-uniform hypergraph H on v vertices and q hyperedges produces a proper two-coloring of H. Theorem 1.1. sD (n2 , q, n) = Ω(q/n). p Note that when q ≤ 0.7 lnnn 2n , the graph is guaranteed to have a two-coloring. However, the above theorem shows if the number of hyperedges is large, a two-coloring (though it is guaranteed to exist) cannot be found efficiently by a deterministic streaming algoirithm. Lower bounds for space bounded computations often follow from lower bounds for associated communication complexity problems. The above result is also obtained using this strategy. However, the communication complexity problem turns out to be somewhat subtle; in particular, we are not able to directly reduce it to a well-known problem and refer to an existing lower bound. We conjecture that no deterministic algorithm can do substantially better if it is allowed only a constant number of passes over the input. Note, however, that proving this might be non-trivial, because the corresponding two-round deterministic communication problem has an efficient protocol; see Section 3.1 for details. 1.2 Randomized streaming algorithms: We show that a version of the delayed recoloring algorithm can be implemented in the streaming model, and provides essentially the same guarantees as the original algorithm. Let sR (v, q, n) be the minimum space used by a randomized (one-pass) streaming algorithm that on receiving an n-uniform hypergraph H on v vertices and q hyperdges with probability at least 34 (say) produces a proper two-coloring of H (or declares failure). We suppose that each vertex is represented using B bits, and each edge is represented as an n-tuple of vertices. 3 p n n 1 Theorem 1.2. sR (v, 10 ln n 2 , n) = O(vB). Furthermore, the corresponding randomized algorithm maintains a coloring of the vertices encountered, p n n and updates this coloring in time O(nvB) 1 per hyperedge. If the hypergraph has at most 10 ln n 2 hyperedges, then with high probability the two-coloring is valid. If the two-coloring being maintained valid, the algorithm declares failure (the algorithm never outputs an invalid coloring). 1.3 The number of vertices in non-two-colorable hypergraphs The upper bound for Property B, q(n) = O(n2 2n ), was shown by Erdős by exhibiting a hypergraph with O(n2 2n ) hyperedges that is not two-colorable; this hypergraph needed n2 vertices. We ask if there are such non-two-colorable hypergraphs with o(n2 ) vertices. 2 n Theorem 1.3. (a) Let t be such that 4 ≤ t ≤ 2n−1 . An n-uniform hypergraph with at most 2 n−1 n /t vertices and at most 2 exp(t/8) hyperedges is two-colorable. The corresponding two-   2 n n−1 R exp(t/8), n = coloring can be found using an efficient randomized streaming algorithm: s t ,2 O(n2 /t). 2 2 n tn (b) If t ≤ 2n−1 , there is a graph with n2 /t vertices and O( nt · 2n exp( n−2t )) hyperedges that is not two-colorable. Setting t = 8 ln(2n) in Theorem 1.3(a) above implies that for any n-uniform hypergraph with n2 vertices and q < n2 2n hyperedges, there is a randomized one-pass streaming algorithm v ≤ 8 ln(2n) 2 that outputs a two-coloring with high probability. This algorithm requires O( lnn n ) space at any instant and O(n) processing time after reading each hyperedge. A comparison of this to the result of Erdős [2], which gave a construction of a non-two-colorable hypergraph with the same q = n2 2n hyperedges, but on Θ(n2 ) vertices shows that the number of vertices being Ω(n2 ) was crucial, and anything significantly smaller would not have worked. Furthermore, Theorem 1.3 (b) recovers Erdős’ bound when t = O(1), and generalizes it to make explicit the dependence of the Erdős’ upper bound on q(n) on the vertex-set size. Organisation of the rest of the paper: In Section 2 we introduce the notation. In Section 3, we establish Theorem 1.1, showing the limitation of deterministic streaming algorithms. In Section 4, we prove Theorem 1.2 by showing how a version of the off-line delayed recoloring algorithm of [4] (see Section 4.1) can, in fact, be implemented efficiently in the streaming model. In Section 5, we prove the results in Theorem 1.3 regarding non-two-colorability. We conclude with some remarks and open problems. 2 Notation  Our n-uniform hypergraphs will be denoted by (V, E), where V is the set of vertices, and E ⊆ Vn is the set of hyperedges of H. For a hypergraph H = (V, E), we use v for |V |, and q for |E|. In our setting, v will typically be a small polynomial in n and q will be exponential in n. For any k ∈ N, we use the notation [k] := {1, . . . , k}. A valid two-coloring of the hypergraph H is an assignment χ : V → {Red, Blue} that leaves no edge monochromatic, i.e., ∀e ∈ E, ∃i, j ∈ e such that χ(i) 6= χ(j). A hypergraph that admits a valid two-coloring is said to be two-colorable, or equivalently, to have Property B. 4 We will be interested in space-efficient streaming algorithms for finding two-colorings of hypergraphs. Consider a hypergraph with edge set E = {e1 , . . . , eq }. The hyperedges are made available to the algorithm one at a time in some order ei1 , . . . , eiq . To keep the problem general, we will not assume that the vertex set is fixed in advance. The algorithm will become aware of the vertices as they arrive as part of the stream of hyperedges. We will assume that each vertex is encoded using B bits. The goal is to design a space-efficient algorithm (deterministic or randomized) that can output a valid coloring for the entire hypergraph once all the hyperedges have passed. We may allow multiple passes over the input data. We call the algorithm a r-pass streaming algorithm, if it outputs a valid coloring after making r passes over the input stream. By space-efficient, we mean that the algorithm uses poly(n) internal workspace. We will assume that n is large, say at least 100. Denote by sD r (v, q, n) the minimum space used by a deterministic r-pass streaming algorithm that on receiving an n-uniform hypergraph H on v vertices and q hyperedges produces a proper twocoloring of H. Similarly, sR r (v, q, n) is the minimum space used by a randomized r-pass streaming algorithm that on receiving an n-uniform hypergraph H on v vertices and q hyperedges with probability at least 43 (say) produces a proper two-coloring of H (or declares failure). When r is omitted from the subscript, it is assumed r = 1. 3 Deterministic streaming algorithms We first show lower bounds in the deterministic setting. We recall the routine translation of an efficient streaming algorithm to a communication complexity protocol [8], with a view to proving lower bounds. There are two computationally unbounded players Alice and Bob, who both know of a relation R ⊆ X × Y × Z. Alice receives an input x ∈ X and Bob gets y ∈ Y; in the beginning, neither player is aware of the other’s input. Their goal is to exchange bits according to a fixed protocol and find a z ∈ Z so that (x, y, z) ∈ R. The communication complexity of R is the minimum number of bits that Alice and Bob exchange in any valid protocol for the worst case input pair (x, y). Several generalizations of this model can be defined with k >= 3 players, we will define our specific model below. For more details on communication complexity in general, please consult the book by Kushilevitz and Nisan [7]. To show lower bounds for hypergraph two-coloring, we define the class of communication problems H(v, q, k) 1 . Definition 3.1 (Problem class H(v, q, k)). For k >= 2, an instance I ∈ H(v, q, k) has k players P1 , . . . , Pk . Each player Pi has a subset Ei ⊆ E of some hypergraph H = (V, E = E1 ∪E2 ∪· · ·∪Ek ), with |V | = v and |Ei | ≤ q. The communication is done in sequential order: starting with P1 , Pi sends a message to Pi+1 for i ∈ {1, . . . , k − 1}. This sequence of communication constitutes a round. In a multiple-round protocol, Pk may start a new round by sending back a message to P1 , who would continue communication in the above order. In a valid r-round protocol Π, some Pi in the course of the r-th round will output a coloring χ that is valid for H. Denote by Π(I, Pi , l) the communication sent by Pi in round l on instance I. We define the r-round communication complexity of H(v, q, k) as follows: CCr (H(v, q, k)) = 1 min Π: r−round valid protocol max I∈H(v,q,k), i∈[k],l∈[r] |Π(I, Pi , l)| Since all hypergraphs we consider are n-uniform, we will not explicitly state n as a parameter 5 Remark 3.2. In this definition, we consider the maximum communication by any single player instead of total communication because this quantity is related more closely to the space requirement of streaming protcols (see Proposition 3.3 below). If the maximum communication by any player is s bits, then the total communication is bounded above by kr · s bits for an r round protocol. Besides, for the values of k and r that we consider, this O(kr) blowup is immaterial. Note that when k = 2 and r = 1, we get the two-player one-round model (where we traditionally call P1 as Alice and P2 as Bob): Alice sends a message m to Bob depending on her input and Bob outputs a coloring looking at m and his input. Our lower bound for hypergraph coloring in the streaming model will rely on the following well-known connection between streaming and communication complexity introduced in [8]. Proposition 3.3 ([8]). For any k ∈ Z, k ≥ 2, we have: CCr (H(v, q, k)) ≤ sD r (v, q, n) Thus, to establish Theorem 1.1, it is enough to show an appropriate lower bound on CC1 (H(v, q, k)). We start with the two-player case, which already introduces most of the ideas. 2 Theorem 3.4. CC1 (H(n2 , q, 2)) = Ω( 2nqn4 ). Remark 3.5. Note that the above theorem gives a non-trivial lower bound only when q ≫ n2 2n/2 . Ideally, we would expect a lower bound that is linear in q, for all values of q. Proof. Consider a valid one-round protocol for the two-player problem where Alice sends a message m from a set M of possible messages. Let the input to the protocol be (HA , HB ), where HA and HB are hypergraphs, each with q hyperedges on a common vertex set [n2 ]. For every hypergraph HB that Bob receives, he must output a coloring χ = f (m, HB ) based on some deterministic function f . For m ∈ M, define L(m) = {f (m, HB ) : HB is an input for Bob}. It is easy to see that Bob may identify the message m with the list L(m); on receiving m, he must find a proper coloring for HB from L(m). Thus, for every m ∈ M, we have the following. Completeness for Bob: Every possible input hypergraph to Bob (i.e. all hypergraphs on q hyperedges) should have a valid coloring in L(m). Soundness for Alice: Let P (m) = {HA : Alice sends the message m for input HA }. Then every χ ∈ L(m) should be valid for every hypergraph HA ∈ P (m). We will show that these two conditions imply the claimed lower bound on |M|. Definition 3.6 (Shadows). Given a coloring χ, we define its shadow ∆(χ) to be the set of all possible hyperedges that are monochromatic under χ. The shadow of a list L of colorings is [ ∆(L) = ∆(χ) χ∈L 6 Note that in the above definition, the shadow ∆(χ) collects all possible monochromatic hyperedges under χ, so it depends only on the coloring χ, and not on any hypergraph. Similarly, ∆(L) is also a collection of hyperedges and does not depend on any hypergraph; in particular, if a hypergraph HA is monochromatic under every coloring in L, then none of HA ’s hyperedges can appear in ∆(L). In the following, assume that n is large. Claim 3.7. For every coloring χ, we have  2 n 1 . ∆(χ) ≥ n n 10 · 2 (3.1) Proof of Claim. : One of the two color classes χ has at least ⌈n2 /2⌉ vertices. It follows that  2   2 ⌈n /2⌉ 1 n ∆(χ) ≥ ≥ . n n n 10 · 2 We next observe that the completeness condition for Bob imposes a lower bound on ∆(L(m)). Claim 3.8. For every m ∈ M,  2 n q . |∆(L(m))| ≥ 2 n n 10n 2 Proof of Claim. : Suppose the claim does not hold, that is,  2 q n |∆(L(m))| < . 2 n n 10n 2 (3.2) (3.3) Choose a random hypergraph HB by choosing q hyperedges randomly from ∆(L(m)). We will say that the hypergraph hits χ ∈ L(m), if at least one of its hyperedges falls in ∆(χ), otherwise we say it misses χ. For each χ ∈ L(m), we have, using the bounds (3.1) and (3.3), that  q |∆(χ)| Pr[H misses χ] ≤ 1 − H |∆(L(m))| q  n2 ≤ 1− q ≤ exp(−n2 ) 2 Since there are at most 2n colorings, the union bound yields: 2 Pr[∃χ ∈ L(m) : H misses χ] ≤ 2n exp(−10n2 ) ≪ 1 H Thus there exists a hypergraph H with q hyperedges that hits every coloring χ ∈ L(m), that is every coloring in L(m) is invalid for H. This, however, violates Completeness for Bob, proving the Claim. 7 We can now complete the proof of the theorem. Consider a random hypergraph H for Alice, obtained by choosing each of its hyperedges uniformly at random from the set of all hyperedges. Since Alice sends some m ∈ M for every hypergraph, the soundness condition for Alice implies that q2 q q ) ≤ |M| exp(− ), 1 = Pr[∃m : H misses all χ ∈ L(m)] ≤ |M|(1 − H 10n2 2n 10n2 2n where we used Claim 3.8 to justify the first inequality. 2 Taking logarithms on both sides yields the desired lower bound log |M| = Ω( nq4 2n ). As remarked earlier, the above communication lower bound implies that the space required by a deterministic streaming algorithm to find a valid coloring is exponential in n even for hypergraphs that have very simple randomized coloring strategies; it does not yield any such lower bound for hypergraphs that have fewer than 2n/2 hyperedges. In order to overcome this limitation, we generalize the analysis above to the k-player hypergraph coloring problem. We will see later that for q ≤ 2n/2 there do exist efficient two-player one-round protocols, so we could not have proved our lower bounds while restricting our attention to the two-player setting. Proof of Theorem 1.1: Theorem 1.1 will follow from the following theorem: Theorem 3.9. Let k ≥ 1.  " q CC1 (H(v, q, k + 1)) = Ω q v v/2 n v n #1  k  Proof. Recall that the protocol has k + 1 players, P1 , P2 , . . . , Pk , Pk+1 . player Pi receives a hypergraph Hi with q hyperedges over the vertex set [v]. The communication starts with P1 , who sends a message m1 of length ℓ1 to player P2 ; in the i-th step, Pi sends a message of length ℓi to Pi+1 . In the end, Pk+1 produces a coloring for the hypergraph H1 ∪ H2 ∪ · · · ∪ Hk+1 . It will be convenient to view this coloring as a message sent by player Pk+1 , and set ℓk+1 = v (the number of bits needed to describe a coloring). In a (k + 1)-player protocol, after the messages sent by the first i players have been fixed, we have a list of colorings that may still be output at the end; we use the following notation to refer to this list (recall that mk+1 is a coloring): L(m1 , m2 , . . . , mi ) = {mk+1 : for some input H1 , H2 , . . . , Hk+1 the transcript is of the form (m1 , m2 , . . . , mi , . . . , mk , mk+1 )}. In particular, by considering the situation at the beginning of the protocol (when no messages have yet been generated), we let L0 = {χ : mk+1 = χ is output by the protocol on some input}. Let  s0 = |∆(L0 )|/ nv , and for i = 1, . . . , k + 1, let   v si (m1 , . . . , mi ) = |∆(L(m1 , m2 , . . . , mi ))| n 8 and si = min (m1 ,m2 ,...,mi ) si (m1 , . . . , mi ). Here the minimum is taken over all possible sequences of first i messages that arise in the protocol. In particular, s0 corresponds to the union of shadows of all colorings ever output by the protocol, and sk+1 corresponds to the shadow of the output corresponding to the transcript (that is, the shadow of the last message, which is a coloring). Claim 3.10. s0 ≤ 1 ∀i ∈ {0, . . . , k} : (3.4) si ≥ si+1 sk+1 ≥ q ℓi+1 ⌈v/2⌉ n . v (3.5) (3.6) n Proof. Inequality (3.4) is immediate from the definition of s0 . For Claim (3.5), we use ideas similar to those used in the proof of Claim 3.8 for two-player protocols. Fix m1 , m2 , . . . , mi . We will show that si (m1 , m2 , . . . , mi ) is at least the right hand side of (3.5). Pick a random hypergraph H (which we will consider as a possible input to Pi+1 ) as follows: choose q hyperedges independently and uniformly from the set ∆(L(m1 , . . . , mi )). When H is presented to Pi+1 , it must respond with a message mi+1 . None of the colorings that Pk+1 produces after that can include any edge of H in its shadow, that is, ∆(L(m1 , . . . , mi+1 )) ∩ H = ∅. For each valid choice m for mi+1 , we have ∆(L(m1 , . . . , mi , m)) ⊆ ∆(L(m1 , . . . , mi )) and   q  si+1 si+1 (m1 , . . . , mi , m) q ≤ 1− . Pr[H ∩ ∆(L(m1 , . . . , mi , m)) = ∅] ≤ 1 − si (m1 , . . . , mi )) si (m1 , . . . , mi )) Thus, q si+1 H si (m1 , . . . , mi ))   si+1 q q This yields exp(ℓi+1 − si (m1 ,...,mi ) ) ≥ 1, giving si (m1 , . . . , mi ) ≥ si+1 ℓi+1 . By minimizing over valid sequences (m1 , . . . , mi ), we justify our claim.  Claim 3.6 follows from the fact the shadow of every coloring has at least ⌈v/2⌉ hyperedges. n  1 = Pr[∃mi+1 : ∆(L(m1 , . . . , mi+1 )) ∩ H = ∅] ≤ 2ℓi+1 1 − By combining parts (3.4) and (3.5), we obtain 1≥ q k+1 sk+1 ,k+1 Y ℓi . i=1 The theorem follows from this by using (3.6), noting that ℓk+1 ≤ v, and for i = 1, 2, . . . , k: ℓi ≤ maxi∈[k] ℓi . Corollary 3.11 (Restatement of Theorem 1.1). Every one-pass deterministic streaming algorithm to two-color an n-uniform hypergraph with at most q hyperedges requires Ω( nq ) bits of space. Proof. Setting k = n in the previous result immediately yields that for hypergraphs on v = n2 vertices and at most q = (n + 1)q ′ hyperedges, the communication required is Ω(q ′ ). The lower bound for streaming algorithms then follows from Proposition 3.3. 9 3.1 Deterministic communication protocols for the two-coloring problem In the previous section, we derived our lower bound for the deterministic streaming algorithms by invoking multi-player communication complexity, because the two-player lower bound did not give us a non-trivial lower bound for hypergraphs with fewer than 2n/2 hyperedges. In this section, we first show an upper bound in the two-player setting, which shows that it was essential to consider the multi-player setting in order to get the stronger lower bound. Next, we consider two-round protocols, for they are related to two-pass streaming algorithms. We show below, perhaps surprisingly, that the two-round two-player deterministic communication complexity for the problem is poly(n). However, we do not have a streaming algorithm with a matching performance. Theorem 3.12. CC1 (H(v, 2n/2 , 2)) = poly(n) Proof. Alice and Bob will base their protocol on a special collection of lists of colorings L = {L1 , . . . , Lr }, with r = 2n . Suppose Alice’s hypergraph is HA and Bob’s hypergraph is HB . The protocol will have the following form. Alice Alice sends an index i of a list Li ∈ L such that every coloring in Li is valid for HA . Bob Bob outputs a coloring χ ∈ Li that is valid for HB . We next identify some properties on the collection of lists that easily imply that the protocol above produces a valid two-coloring. Definition 3.13 (Good Lists). (a) A collection of lists L is good for Bob, if for every hypergraph HB , in every list L in L there is a valid coloring for HB in L. (b) A collection of lists L is good for Alice, if for every hypergraph HA , there is some list L in L, such that every coloring in L is valid for HA . Note that good collections are defined differently for Alice and Bob. With this definition, the existence of a collection of lists that is good for both Alice and Bob would furnish a one-way communication protocol with ⌈log r⌉ bits of communication, and establish the theorem. Such a good collection is shown to exist in Lemma 3.14 below. Lemma 3.14. Suppose HA and HB are restricted to have at most 2n/2 hyperedges. Then, there exists a collection of L = {L1 , . . . , Lr } of r = 2n lists, each with k = ⌈62n/2 log v⌉ colorings, which is good for Alice and Bob. Proof. We pick L by picking r lists randomly: list Li will be of the form {χi1 , . . . , χik }, where each coloring is chosen independently and randomly from the set of all colorings. We will show that with positive probability L is a good collection of lists. We will separately bound the probability that L fails to satisfy conditions (a) and (b) above. We may restrict attention to hypergraphs that have exactly q = ⌊2n/2 ⌋ hyperedges. (a) We first bound the probability that L is not good for Bob. Consider one list Li ∈ L. Pr[Li is not good for Bob ] = Pr[∃HB such that none of the k colorings are valid for HB ]  v   q k n . ≤ 2n−1 q 10 Since L is a collection of r lists, we have Pr[L is not good for Bob] = Pr[∃Li ∈ L is not good for Bob]  v   q k n ≤ r 2n−1 q ≤ 2(nq log v+log r)−k(n−log q−1) . (3.7) (3.8) To make the right hand side of (3.8) small, we will later choose k such that k≫ nq log v + log r . n − log q − 1 (3.9) (b) We next bound the probability that L is not good for Alice. Fix a hypergraph HA for Alice, and consider Li ∈ L.  q k Pr[every χ ∈ Li is valid for HA ] ≥ 1 − n−1 Li 2     qk x ) . (using 1 − x ≥ exp − 1−x ≥ exp − n−1 2 −q v  Since L is a collection of r such lists chosen independently and there are (nq ) choices for HA , we have  v   r  qk n Pr[L is not good for Alice] ≤ 1 − exp − n−1 2 −q q    qk ≤ exp qn ln v − r exp − n−1 . (3.10) 2 −q To make the right hand side of (3.10) small, we will choose r such that   qk r ≫ (qn ln v) exp . 2n−1 − q (3.11) Now, suppose v = poly(n). Then one can verify that if we set k = ⌈6q log v⌉ and r = 2n , then (3.9) and (3.11) both hold for all large n. It follows that the required collection of lists exists. 3.1.1 Circuit upper bounds and a 2-round communication protocol Next we present our efficient two-player, two-round communication protocol for hypergraph twocoloring. Theorem 3.15. CC2 (H(v, 2n /8, 2)) = poly(n) We will prove Theorem 3.15 by exploiting the connection between circuit complexity and KarchmerWigderson games [14]. Definition 3.16 (Karchmer-Wigderson game). Given a monotone boolean function f : {0, 1}N → {0, 1}, the Karchmer-Wigderson communication game Gf between two players Alice and Bob is the following: Alice gets an input x ∈ f −1 (0), and Bob gets an input y ∈ f −1 (1), and they both know f . The goal is to communicate and find an index i ∈ [N ] such that xi = 0 and yi = 1 (such an index exists since f is monotone). The communication complexity of this game is denoted by CC(Gf ) (or specifically CCr (Gf ) for an optimal r-round protocol). 11 Karchmer and Wigderson make the following connection: Theorem 3.17 ([14]). If a monotone boolean function f has a depth-d, size s circuit, then CCd−1 (Gf ) = O(d log s). The proof of Theorem 3.15 will be based on the circuit complexity of Approximate Majority functions. Define the partial function ApprMajN on a subset of {0, 1}N as follows: ApprMajN (x) = 1 when |x| ≥ 2N/3, and ApprMajN (x) = 0 when |x| ≤ N/3. A function f : {0, 1}N → {0, 1} is an Approximate Majority function if it satisfies: f (x) = 1 when |x| ≥ 2N/3, and f (x) = 0 when |x| ≤ N/3. We need the following Lemma. v Lemma 3.18. For any Approximate Majority function f : {0, 1}2 → {0, 1}, we have: CCr (H(v, 2n /8, 2)) ≤ CCr (Gf ) Proof. We will prove this by reducing H(v, 2n /8, 2) to Gf , for any Approximate Majority function f . For convenience, let N := 2v . Suppose Alice receives HA = (V, EA ), and Bob receives HB = (V, EB ). Let LA ∈ {0, 1}N be an indicator vector for Alice that marks which of the N possible twocolorings of V (arranged in some canonical order) are valid for HA , and let LB be the corresponding indicator vector for Bob. Since we know that both HA , HB have q ≤ 2n /8 hyperedges, a randomly chosen coloring will color either hypergraph with probability ≥ 3/4, which implies |LA | ≥ 3N/4 and |LB | ≥ 3N/4. To find a coloring valid for both HA and HB , Alice and Bob just have to find an index i ∈ [2v ] such that LA (i) = LB (i) = 1. Alice now complements her input to consider LA . Since |LA | ≤ N/4, ApprMajN (LA ) = 0, whereas ApprMajN (LB ) = 1. Clearly, playing the Karchmer-Wigderson game for any Approximate Majority function f on N bits, LA , LB would find a index i where LA = LB = 1, and consequently yield a valid coloring for HA ∪ HB . Hence, a small depth-3 monotone circuit that computes ApprMaj2v on its domain would yield an efficient 2-round protocol for H(v, 2n /8, 2), using Theorem 3.17. Ajtai [15] showed that such a circuit exists, and Viola [16] further showed uniform constructions of such circuits. Theorem 3.19. [16] There exist monotone, uniform poly(N )-sized depth-3 circuits for ApprMajN on N input bits. Proof (of Theorem 3.15). The proof of Theorem 3.15 is now immediate from Theorems 3.17, 3.19 and Lemma 3.18, since v is polynomial in n. 4 Streaming algorithms for hypergraph two-coloring In this section, we present streaming algorithms that come close to the performance of the randomized off-line algorithm of Radhakrishnan and Srinivasan [4]. We first point out why this algorithm, as stated, cannot be implemented with limited memory. Next, we show how an alternative version can be implemented using a small amount of memory. This modified version, however, does return colorings that are not valid (though it does this with small probability); we show how by maintaining a small amount of additional information, we can derive an algorithm that returns and valid coloring with high probability or returns failure, but never returns an invalid coloring. In order to describe our algorithms, it will be convenient to use u, w to denote vertices in the hypergraph. Also, in this section, we will explicitly use |V | to denote the size of the vertex set V , 12 and |E| to denote the number of hyperedges in the hypergraph. The vertex set of the hypergraph will be identified with [|V |], and B will denote the number of bits required to represent a single vertex in V . 4.1 The delayed recoloring algorithm Radhakrishnan and Srinivasan [4] showed that by introducing delays in the recoloring step of an algorithm originally proposed by Beck [1] one can two-color hypergraphs with more hyperedges than it was possible before. p Theorem 4.1. Let H = (V, E) be an n-uniform hypergraph with at most 1/10 n/ ln n × 2n hyperedges. Then H is two-colorable; also a proper two-coloring can be found with high probability in time O(poly(|V | + |E|)). Algorithm 1 (Off-line delayed recoloring algorithm) 1: Input H = (V, E). 2: For all u ∈ V , independently set χ0 (u) to Red or Blue with probability 12 . 3: Let M0 be the set of hyperedges of H, that are monochromatic under χ0 . 4: For all u ∈ V , independently set b(u) to be 1 with probability p, and 0 with probability 1 − p. 5: Let π be one of the |V |! permutations of V , chosen uniformly at random. 6: for i = 1, 2, . . . , |V | do 7: χi is obtained from χi−1 by retaining the colors of all vertices except perhaps π(i). If b(π(i)) = 1, and some hyperedge containing π(i) was monochromatic under χ0 and all its vertices still have the same color in χi−1 , then χi−1 (π(i)) is flipped to obtain χi (π(i)). 8: end for 9: Output χf = χ|V | . The term off-line refers to the fact that we expect all hyperedges to remain accessible throughout the algorithm. The term delayed recoloring refers to the fact that vertices are considered one after another, and are recolored only if the initially monochromatic edge they belonged to has not been set right by a vertex recolored earlier. For a coloring χ, let B(χ) be the set of hyperedges that are colored entirely blue under χ, and the following claims are let R(χ) be the set of hyperedges colored entirely red under χ. In p[4], 1 n n established for this algorithm under the assumption that |E| ≤ 10 2 . In the following, the ln n parameter p denotes the probability that the bit b(v) is set to 1 in Algorithm 1. r 2 n Pr[(B(χf ) ∩ B(χ0 ) 6= ∅) ∨ (R(χf ) ∩ R(χ0 ) 6= ∅)] ≤ (1 − p)n (4.1) 10 ln n 2np ; (4.2) Pr[B(χf ) \ B(χ0 ) 6= ∅] ≤ 100 ln n 2np Pr[R(χf ) \ R(χ0 ) 6= ∅] ≤ . (4.3) 100 ln n The first inequality (4.1) helps bound the probability that an initially monochromatic hyperedge does not change our attempts at recoloring; the second inequality (4.2) considers the event where an initially non-blue hyperedge become blue because of recoloring; the third inequality (4.3) similarly refers to the event where an initially non-red hyperedge becomes red because of recoloring. These 13 events together cover all situations where χf turns out to be invalid for H. Now, set p = ln ln n) so that we have r 2 4np 1 1 11 n (1 − p)n + ≤ + ≤ . 10 ln n 100 ln n 5 50 50 1 2n (ln n − (4.4) Thus, the above algorithm produces a valid two-coloring with probability at least 12 . Furthermore, since the entire graph is available in this off-line version, we can efficiently verify that the final coloring is valid. By repeating the algorithm ⌈log(1/δ)⌉ times, we can ensure that the algorithm produces a valid coloring with probability at least 1 − δ, and declares failure otherwise (but never produces an invalid coloring). Implementation in the streaming model: Now, suppose the hyperedges arrive one at a time, each hyperedge is a sequence of n vertices, each represented using B bits. In time O(Bn) per edge, we may extract all the vertices. The corresponding colors and bits for recoloring can be generated in constant time per vertex. Generating the random permutation π is a routine matter, we maintain a random permutation of the vertices received so far by inserting each new vertex at a uniformly chosen position in the current permutation; using specialized data structure, the total time taken for generating π is O(|V ||B| log |V |). Overall, the the algorithms can be implemented in time Õ(nB|V ||E|) (the Õ hides some polylog(|V |) factors). The space required is rather large because we explicitly store the hyperedges. Observe, however, that we need store only the hyperedges that are monochromaticpafter the initial random coloring. Thus, the algorithm can be implemented using Õ(|V |B + nB lnnn ) bits of space on the average. The space and time requirements may then be considered acceptable, but this implementation does not allow us to determine if the coloring produced at the end is valid; moreover, it is not clear how we may reduce the error probability by repetition. In the next section, we give an implementation that does not suffer from this deficiency. 4.2 An efficient streaming algorithm In this section, we modify the randomized streaming algorithm of the previous section so that it uses O(B|V |) bits of space and takes time comparable to the off-line algorithm. This version maintains a running list of vertices as suggested above, but it assigns them colors immediately, and with the arrival of each edge decides if some of its vertices must be recolored. The algorithm is derived from the off-line algorithm. It maintains enough information to ensure that the actions of this streaming algorithm can somehow be placed in one-to-one correspondence with the actions of the off-line algorithm. We wish to show that Algorithm 2 succeeds in constructing a valid coloring with probability at least 12 . To justify this, we compare the actions of this algorithm and the off-line delayed recoloring algorithm stated earlier, and observe that their outputs have the same distribution. Recall that χ0 was the initial coloring generated in the Algorithm 1. Let χ̂0 be the corresponding coloring for Algorithm 2: that is, let χ̂0 (u) be the color u was first assigned when the first hyperedge containing u appeared in the input. Similarly, let b̂ be the sequence of bits generated by the above algorithm, and π̂ be the final permutation of vertices that results. Notice (χ0 , b, π) and (χ̂0 , b̂, π̂) have the same distribution. In Algorithm 1, once χ0 , b and π are fixed, the remaining actions are deterministic. That is, χf = χf (χ0 , b, π) is a function of (χ0 , b, π). Similarly, once we fix (i.e., condition on) (χ̂0 , b̂, π̂), the final coloring χ̂f = χ̂f (χ̂0 , b̂, π̂) is fixed. 14 Algorithm 2 Delayed recoloring with limited space 1: Input: The hypergraph H = (V, E) as a sequence of hyperedges h1 , h2 , . . .. 2: The algorithm will maintain for each vertex u three pieces of information: (i) its initial color χ0 (u); (ii) its current color χ(u); (iii) a bit b(u) that is set to 1 with probability p; 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: Furthermore, it maintains a random permutation π of the vertices received so far. for i = 1, 2, . . . , |E| do Read hi . for u ∈ hi do if u has not been encountered before then Set χ0 (u) = χ(u) to Red or Blue with probability 12 . Set b(u) to 1 with probability p and 0 with probability 1 − p. Insert u at a random position in π. end if end for if hi was monochromatic under χ0 and all its vertices have the same color under χ then Let u be the first vertex (according to the current π) such that b(u) = 1. if χ0 (u) = χ(u) then Flip χ(u). end if end if end for Output χf = χ|V | . 15 Lemma 4.2. For all χ, b and π, we have χf (χ0 , b, π) = χ̂f (χ0 , b, π). Proof. Suppose χf (u) 6= χ̂f (u) for some vertex u. One of them must equal χ0 (u). We have two cases. Suppose χ0 (u) = χ̂f (u) 6= χf (u): Let us also assume that h is a hyperedge that necessitated u’s recoloring in the off-line algorithm. This implies that b(u) = 1, and b(w) = 0 for the vertices w of h that appeared before u in π. Now in the streaming algorithm above, when h is considered, we would find that h is monochromatic in χ0 , and flip χ(u) unless it is already flipped. Once flipped, the color of u will not change again. Thus, χ̂f (u) 6= χ0 (u)—a contradiction. Suppose χ0 (u) = χf (u) 6= χ̂f (u): Let h be the hyperedge that necessitated the recoloring of u in the above streaming algorithm. This implies that b(u) = 1, and all vertices w of h that appear before u in permutation π have b(w) = 0. But in such a situation, the original off-line algorithm will find h monochromatic when u is considered, and will flip its color. Thus, χ0 (u) 6= χf (u)—a contradiction. We conclude that (4.4) applies to Algorithm 2 as well. Corollaryp4.3 (Compare Theorem 4.1). Let H = (V, E) be an n-uniform hypergraph with at most 1/10 n/ ln n × 2n hyperedges. Then, with probability at least 12 Algorithm 2 produces a valid two-coloring for H. Note, however, Algorithm 2 lacks the desirable property that it outputs only valid colorings. A straight-forward check does not seem possible using O(|V |B) RAM space. We show below that with carefully storing some vertices of a few hyperedges, we can in fact achieve this using only O(nB) RAM space. 4.3 An efficient algorithm that produces only valid colorings As remarked before the events considered in inequalities (4.1), (4.2) and (4.3) cover all situations when the algorithm might return an invalid coloring. So, in order never to return an invalid coloring, it suffices to guard against these events. (a) To ensure that some vertex in every initially monochromatic hyperedge h does change its color, we just need to verify that b(u) = 1 for some u ∈ h. This we can ensure by examining χ0 and b just after they are generated in line 7 of Algorithm 2. (b) To ensure that no initially non-blue hyperedge h has turned blue, we will will save the red vertices of h whenever there is potential for them all turning blue, and guard against their turning blue. Consider the hypergraph HBlue with hyperedges −1 −1 E(HBlue ) = {h ∩ χ−1 0 (Red) : h ∩ χ0 (Red) 6= ∅ and b(u) = 1 for all u ∈ h ∩ χ0 (Red)}. We will then verify that no hyperedge in HBlue is entirely blue in the end. 16 (c) To ensure that no initially non-red hyperedge h has turned red, we consider the corresponding hypergraph HRed with hyperedges −1 −1 E(HRed ) = {h ∩ χ−1 0 (Blue) : h ∩ χ0 (Blue) 6= ∅ and b(u) = 1 for all u ∈ h ∩ χ0 (Blue)}, and verify in the end that no hyperedge in HRed is entirely red in the end. It remains to show how HBlue and HRed can be stored efficiently. Each hyperedge will be stored separately by listing all its vertices. Thus, the expected sum of the sizes of the hyperedges in HBlue is n   X X n i −n i E[size(HBlue )] = 2 p i i=1 h∈E(H) −n ≤ |E| np2 n−1  X i=0 −n ≤ |E| np2  n − 1 i−1 p i−1 (1 + p)n−1 . Clearly, the same bound applies to HRed . We will choose the value of p, so that with high probability size(HBlue ) and size(HBlue ) are both at most n (if either of them exceeds n, the algorithm terminates with failure), while also ensuring that the right hand sides of (4.1), (4.2) and (4.3) stay small. As 1 (ln n − ln ln n). To the failures accounted for in (4.4), we must account for before set p = 2n p n ln n p n n 1 size(HBlue ) or size(HBlue ) exceeding n. Now, E[size(HBlue )] ≤ 10 ln n 2 ln n ≤ 20 , and by 1 1 Markov’s inequality Pr[size(HBlue ) ≥ n] ≤ 20 . Similarly, Pr[size(HRed ) ≥ n] ≤ 20 . Thus, the 1 1 11 + 20 + 20 ≤ 12 , while revised algorithm fails to deliver a valid coloring with probability at most 50 it uses at most O(|V |B + nB) bits of space. 4.4 A local version The above results deal with the cases when a upper bound on the number of hyperedges of H guarantees that it has a valid two-coloring. Using the Lovász LocalpLemma one can show that n-uniform hypergraphs where no hyperedge intersects more than 0.17 n/ ln n 2n others has a twocoloring [4]. Note that this does not require a bound on the number of hyperedges of H, but only one on the number of intersections of any one hyperedge with others. The algorithms of Moser and Tardos [10] can now be used to recover a valid two-coloring of H. As in the previous case, though the algorithm a-priori requires access to the entire hypergraph in an offline fashion, we can modify it to adapt it to the streaming setting. The details are straightforward, and we omit them here to state only the result that derives from the parallel algorithm in [10]. n−1 Proposition 4.4. If each hyperedge of a hypergraph H intersects at most (1−ǫ)2 − 1 other e hyperedges, then a two-coloring of the hypergraph can be found by a streaming algorithm that makes at most O(log |V |) passes over the input stream and uses space O(|V |B). p Compared to the Moser-Tardos algorithm, which works when the number of intersections is O( n/ ln n 2n ), this works only until intersections of up to O(2n ). Furthermore, note that it requires log |V | passes over the input stream, in contrast to previous algorithms that required just one pass. 17 5 Coloring n-uniform hypergraphs with few vertices The upper bound q(n) = O(n2 2n ) was shown by Erdős by considering a random graph on n2 vertices. We ask if hypergraphs with a comparable number of hyperedges but supported on fewer vertices (perhaps o(n)) can be non-two-colorable. We show that that the choice of n2 above is essentially optimal. Theorem 5.1 (Implies Theorem 1.3(a)). Let t ≥ 4 be a parameter, and H be an n-uniform 2 hypergraph on v vertices and q hyperedges with v ≤ nt vertices. Then, a random two-coloring that partitions the vertex set into two almost equal color classes is a valid coloring with probability at least 1 − 2−(n−1) exp(− 8t )q. The following Corollary is obtained by setting t = 8 ln 2n. Corollary 5.2. Every n-uniform hypergraph with at most edges is two-colorable. n2 8 ln 2n vertices and at most n2 2n hyper- The Corollary shows that Erdős’ construction of non-two-colorable hypergraphs having q = n2 2n hyperedges on a vertex set of size v = Θ(n2 ) is nearly optimal in terms of the number of vertices. Proof of Theorem 5.1: Suppose V (H) = [v]. Consider the coloring χ obtained by partitioning [v] into two roughly equal parts and coloring one part Red and the other part Blue. Every hyperedge has the same probability p of being monochromatic under this coloring. The theorem will follow immediately from the claim below using a union bound.   2 Claim 5.3. p ≤ 2−(n−1) exp − (n−1) . 2v Proof of claim: If v ≤ 2(n − 1), then this probability is 0. In general, the probability that a hyperedge is monochromatic is   ⌈v/2⌉  2 v/2 + ⌊v/2⌋ n n n  ≤ v (since na is convex for a ≥ n − 1). p= v n n We then have p≤ 2· v2 ( v2 −1)( v2 −2)···( v2 −n+1) v(v−1)(v−1)···(v−n+1) n−1 Y v − 2i v−i i=1 n−1 Y −(n−1) 1− = 2 = 2−(n−1) i=1 −(n−1) ≤ 2 exp − i v−i n−1 X i=1 18  i v−i ! . Now, − n−1 X i=1 i v−i = (n − 1) − v Pv−1 1 j=v−n+1 j ≤ (n − 1) − v Z v v−n+1 1 dx x  n−1 ≤ (n − 1) + v ln 1 − v   n−1 1 n−1 2 ≤ (n − 1) − v + ( ) v 2 v (n − 1)2 = − . 2v    2 This shows that p ≤ 2−(n−1) exp − (n−1) and concludes the proof of the Claim. 2v To recover the statement of Theorem 5.1, we union-bound over the set of hyperedges, and observe 2 ≥ (n − 1)2 · nt2 ≥ t/8. that (n−1) v A streaming algorithm: Theorem 5.1 (more generally, Proposition 5.5 at the end of this section for k-coloring) gives rise to a straightforward randomized streaming algorithm for two-coloring (k-coloring) assuming that the number of vertices in the hypergraph is known in advance. For 2 example, suppose we are given an n-uniform hypergraph on v = nt vertices and at most 2n−2 exp( 2t ) hyperedges. We just pick a random coloring χ that assigns colors Red and Blue to roughly the same number of vertices, and verify that no edge is monochromatic under χ. The algorithm uses O(v) bits of space, and returns a valid coloring with probability 12 and fails otherwise. The probability of failure can be reduced to δ by running the algorithm log 1δ times in parallel. We further examine the trade-off between the number of vertices and the number of hyperedges. 2 Let q(n, t) be the minimum number of hyperedges in a non-two-colorable hypergraph with nt vertices. Note that when the number of vertices v ≤ 2n − 2, every n-uniform hypergraph is trivially two-colorable by any coloring that colors ⌊v/2⌋ vertices Blue and the rest Red, and so we will take n2 t ≥ 2n − 1. The above proof (of Theorem 1.3(a)) shows:   t q(n, t) ≥ 2(n−1) exp . 8 Using arguments similar to those used by Erdős [2], we will now obtain a comparable upper bound on q(n, t). This makes the dependence of two-colorability on the vertex set-size explicit. 2 Proposition 5.4 (Restatement of Theorem 1.3(b)). Suppose N ≤ nt is an integer that is at least tn 2n − 1. Then, there is an n-uniform hypergraph with N vertices and O(N · 2n exp( n−2t )) hyperedges 2 tn ). that is not two-colorable. Thus, q(n, t) ≤ ( nt )2n exp( n−2t Proof. Our non-two-colorable n-uniform hypergraph will be obtained by picking the required number of hyperedges at random from [N ]. Fix a coloring χ. Then, a randomly drawn hyperedge is 19 monochromatic with probability at least n2 2t p= n n2 t n   n−1 1 Y n2 /t − 2i 2n n2 /t − i i=0 −n  t −n ≥ 2 1+ n − 2t   −tn ≥ 2−n exp . n − 2t = ] Let H be the hypergraph obtained by picking q hyperedges uniformly and independently from [N n . The probability that χ is valid for H (i.e., none of its hyperedges is monochromatic under χ) is at most q  −tn ) . 1 − 2−n exp( n − 2t Since there are at most 2N colorings, the probability that some coloring χ is valid for H is at most   −tn q N −n 2 1 − 2 exp( . n − 2t tn ) · ln 2 this quantity is less than 1. Our claim follows from this. If q ≥ N · 2n exp( n−2t The above results can be generalized easily to k-colorings. Let qk (n, t) be the minimum number of 2 hyperedges in a non-k-colorable hypergraph with at most nt (≥ kn − 1) vertices. We then have the following bounds (we omit the proofs).     (k − 1)t (k − 1)tn n2 Proposition 5.5. kn−1 exp − ≤ qk (n, t) ≤ ( )kn exp . 2 t n − kt 6 Conclusion The lower bound we obtain on the space requirements of one-pass streaming algorithms is optimal (up to poly(n)) factors. We present an efficient two-player two-round deterministic communication protocol for two-coloring n-uniform hypergraphs with up to 2n /8 hyperedges, but we do not know if there is a corresponding streaming algorithm. As mentioned in the introduction, it is easy to come up with a deterministic streaming algorithm that works in |V | passes using the method of conditional expectations. Even in the two-player communication setting it would be interesting p to determine if the protocol can accommodate up ω(2n ) hyperedges, perhaps even Ω( lnnn 2n ) hyperedges. 2 Our two-coloring algorithm for hypergraphs with O( nt ) vertices does not improve on the bound provided via the delayed recoloring algorithm when t is small, say, o(log n). We believe it should be possible to combine our argument and the delayed recoloring algorithm to show that p ifn the 1 n number of vertices is o(n2 ) then we can two-color hypegraphs with strictly more than 10 ln n 2 hyperedges. 20 References [1] J.Beck. On 3-chromatic hypergraphs. Discrete Mathematics, 24:127-137, 1978. [2] P. Erdős. On a Combinatorial Problem. Nordisk Mat. Tidsskr, 11:5–10, 1963. [3] P. Erdős and L. Lovász. Problems and results on 3-chromatic hypergraphs and some related questions. Colloq. Math. Soc. Jnos Bolyai, 10:609–627, 1973. [4] J. Radhakrishnan and A. Srinivasan. Improved bounds and algorithms for hypergraph 2coloring. Random Structures Algorithms, 16 (1):4–32, 2000. [5] S. Shannigrahi. Coloring, Embedding, Compression and Data Structure Problems on Uniform Hypergraphs Ph.D. Thesis, School of Technology and Computer Science, Tata Institute of Fundamental Research, 2011. [6] N. Alon and J. H. Spencer. The Probabilistic Method. Wiley-Interscience Series, John Wiley and Sons, Inc., New York, U.S.A., 1992. [7] E. Kushilevitz and N. Nisan. Communication Complexity. Book, Cambridge University Press, 1996. [8] N. Alon, Y. Matias and M. Szegedy. The space complexity of approximating the frequency moments. Proceedings of the twenty-eighth annual ACM symposium on Theory of computing (STOC ’96) Pages 20-29, 1996. [9] D. Achlioptas, J.H. Kim, M. Krivelevich, and P. Tetali. Two-coloring random hypergraphs. Random Structures & Algorithms 20(2): 249-259, 2002. [10] R. A. Moser and G. Tardos. A constructive proof of the general Lovász local lemma. Journal of the ACM 57 (2): 1-15, 2010. [11] M. Mitzenmacher and E. Upfal. Probability and Computing: Randomized Algorithms and Probabilistic Analysis. Cambridge University Press, 2005. [12] R. M. Karp, C. H. Papadimitriou and S. Shenker. A simple algorithm for finding frequent elements in streams and bags. ACM Transactions on Database Systems 28: 51-55, 2003. [13] D. D. Cherkashin and J. Kozik. A note on random greedy coloring of uniform hypergraphs. Random Structures & Algorithms 47(3): 407-413, 2015. [14] M. Karchmer and A. Wigderson. Monotone Circuits for Connectivity Require SuperLogarithmic Depth. SIAM J. Discrete Math. 3(2): 255-265 , 1990. [15] M. Ajtai. Σ11 -formulae on nite structures.Ann. Pure Appl. Logic. 24(1):148, 1983. [16] Emanuele Viola. On Approximate Majority and Probabilistic Time. Computational Complexity, v.18 n.3, p.337-375, 2009. [17] S. Muthukrishnan. Data Streams: Algorithms and Applications, Foundations and Trends in Theoretical Computer Science: Vol. 1: No. 2, pp 117-236, 2005. [18] Joan Feigenbaum, Sampath Kannan, Andrew McGregor, Siddharth Suri, Jian Zhang. On Graph Problems in a Semi-streaming Model, Proc. 31st International Colloquium on Automata, Languages and Programming, (ICALP), v. 3182: 531-543, 2004. 21 [19] J. Feigenbaum, S. Kannan, and J. Ziang. Computing diameter in the streaming and sliding window models, Algorithmica, vol. 41, no. 1, pp. 2541, 2004 [20] J. Feigenbaum, S. Kannan, A. McGregor, S. Suri, and J. Zhang. Graph distances in the streaming model: The value of space. Proc. ACM-SIAM SODA, pp. 745754, 2005. 22
8cs.DS
1 Wireless Channel Modeling Perspectives for Ultra-Reliable Low Latency Communications Patrick C. F. Eggers, Marko Angjelichinoski, and Petar Popovski arXiv:1705.01725v2 [cs.IT] 12 Jan 2018 Department of Electronic Systems, Aalborg University, Denmark Abstract Ultra-Reliable Low Latency Communication (URLLC) is one of the distinctive features of the upcoming 5G wireless communication, going down to packet error rates (PER) of 10−9 . In this paper we discuss the statistical properties of the wireless channel models that are relevant for characterization of the lower tail of the Cumulative Distribution Function (CDF). We show that, for a wide range of channel models, the outage probability at URLLC levels can be approximated by a simple power law expression, whose exponent and offset depend on the actual channel model. The main insights from the analysis can be summarized as follows: (1) the two-wave model and the impact of shadowing in combined models lead to pessimistic predictions of the fading in the URLLC region; (2) the CDFs of models that contain single cluster diffuse components have slopes that correspond to the slope of a Rayleigh fading, and (3) multi-cluster diffuse components can result in different slopes. We apply our power law approximation results to analyze the performance of receiver diversity schemes for URLLCrelevant statistics and obtain a new simplified expression for Maximum Ratio Combining (MRC) in channels with power law tail statistics. Index Terms Ultra-reliable communications, URLLC, 5G, wireless channel models, fading, diversity, probability tail approximations, rare event statistics. I. I NTRODUCTION One of the features of 5G wireless communication systems is to offer wireless communication with extremely high reliability guarantees, also known as Ultra-Reliable Low Latency Communication (URLLC) [1], [2]. The level of reliability, going sometimes down to packet error rates (PER) of 10−9 , should be sufficiently convincing in order to remove cables in an industrial setting, remote control of robots and drones that need to perform a critical function, remote surgery or 2 self-driving cars [3]. A significant number of the URLLC use cases take place indoors, which resonates with the expectation for a wireless environment that is, to some extent, controlled and offers predictable signal quality. Wireless communication performance is inherently a stochastic variable; however, URLLC refocuses our attention to the statistical characterization of the rare events and the lower tail of the channel Cumulative Distribution Function (CDF). This requires a major effort in terms of measurement campaigns, purposefully designed to capture the lower tail statistics, as well as in terms of analytical channel modeling. Nevertheless, a good starting point for characterizing wireless channels is to analyze the current channel models in regimes that are relevant for URLLC, which is the topic of this article. The related works [4], [5], [22], [23], [26] have characterized the probability density functions of wireless channel models, often leading to very complex expression and analysis dependent on multiple parameters. Our objective is to get insight into the behavior of different channel models at the region of ultra-low error rates, which we define as a PER that is  ≤ 10−5 or lower, corresponding to the reliability of “five nines”. We have selected 10−5 as this is the target error rate for URLLC selected in 3GPP for a packet of 32 bytes to be delivered within 1 ms [44]. Different mission-critical services will use different levels of ultra-reliability, such as PERs of 10−6 in smart grids and 10−9 for factory automation [3]. We focus on the packet errors that occur due to outages, induced by block fading, rather than errors caused by noise. Recent studies [6] have shown that this is a very suitable model for transmission of short packets, which are in turn expected to be prevalent in the URLLC scenarios. Throughout the paper we will use the term URLLC-relevant statistics to denote the events in signal reception that occur with probabilities on the order of 10−5 or less. Our analysis shows that, despite the complex probability density functions, the behavior of the lower tail can be significantly simplified, leading to important insights on the behaviors that occur when the outage probabilities are very low. While the tuning of the models can only be done with appropriate experimental evidence, this work provides a valuable contribution to the modeling efforts that are essential for designing ultra-reliable wireless systems. Besides the analysis of various statistical channel models, in this paper we also revisit the basic assumptions adopted in channel modeling in the light of the statistically rare events that needs to be captured in an URLLC regime. The existing channel models have been developed 3 for wireless communication systems1 that deal with bit error rates (BER) of 10−3 to 10−4 [7] and there is a lack of experimental evidence to support channel models that deal with PER at the URLLC levels. This paper extrapolates the standard channel models towards URLLC-relevant statistics and identifies their limitations when modeling rare events. One of our main result is that, for a wide range of channel models (but not all, as it will be seen in the paper), the tail approximation of the cumulative distribution function at the URLLC levels has the form:  β   PR PR ≈α F A A (1) where A is the average power, PR is the minimal required power to decode the packet correctly and α, β are parameters that depend on the actual channel model. Since the support of URLLC will be significantly dependent on high levels of diversity, we have also characterised the URLLC-relevant statistics when multiple antennas are considered. Specifically, we provide a simplified analysis of M -branch receive diversity for uncorrelated branch signals, that makes use of (1), as well as approximations for some special channels. The result provides a compact Maximum Ratio Combing (MRC) solution of form FMRC ≈ αMRC (β1 ..βM )FSC (2) that is, a scaled version of a Selection Combining (SC) solution, in which the scale parameter αMRC depends only on the branch exponents β. The paper is organized as follows. After the introduction, in Section II we provide the system model. Section III contains analysis of a wide range of channel models which exhibit power law tails at URLLC-relevant probabilities. Section IV contains analysis of two models that do not not result in power law tails. Section V contains the analysis of the receive diversity schemes in URLLC-relevant regime. Section VI concludes the paper. II. W IRELESS C HANNEL M ODELING FOR URLLC-R ELEVANT S TATISTICS 1) Preliminaries: The common approach in wireless channel modeling is to assume separability of the following effects [9]: 1 • Path loss, dependent on the actual geometric setting and operating frequency. • Long-term fading (i.e., shadowing) that captures slowly-varying macroscopic effects. For example, the first generation digital systems, such as GSM, that had an emphasis on voice communication. 4 • Short-term fading processes, relevant on a time scale of a packet (i.e., quasi-static fading) or even a symbol (fast fading), assuming stationary scattering conditions. The URLLC performance is determined by the short-term process and its (un)predictability, which ultimately determines the fate of the packet at the destination. Assuming separability, the statistics of short-term fading is described via parameters that are derived from the longterm fading and path loss effects; these parameters are assumed to be constant over a period of time. However, separability becomes problematic when URLLC-relevant statistics is considered, since the estimated long-term parameters require certain level of accuracy in order to have a valid short-term statistics of rare events. Motivated by this, we also consider combined long and short term fading models. Furthermore, in absence of dedicated URLLC channel models, we investigate the behavior of a wide palette of existing wireless channel models in URLLC regime. 2) General Model: We use combination of (a) the complex baseband model of a narrowband channel with reduced wave grouping from [4], and (b) the incoherent multi-cluster channel of [17], [19]. Let P denote the total received power; we have: P =ω µ X m=1 2/β |Vm | , Vm = ξ N X i=1 ! jφi,m ρi,m e + L Y Vdif,m,l . (3) l=1 Vm denotes the complex received voltage from the m−th cluster m = 1, . . . , µ, in which ρi,m /φi,m is the amplitude/phase of the i−th specular component, i = 1, . . . , N and Vdif,m,l is the l−th diffuse component for the m−th cluster with L denoting the number of diffuse components per cluster [8]. β caters for the modeling of a Weibull channel [16], and for all other models it is set to β = 1. The shadowing effects are represented by the random variables (RVs) ξ and ω. Here ξ is a common shadowing amplitude that affects only the specular components [26], while ω induces a shadowing effect on the total power [9], [23], see section III-G. We assume that each ρi of a specular component is constant and that φi is a uniform random variable [4]. The elementary diffuse components Vdif,m,l are treated in their simplest form, as a contribution from a large number of waves and application of the central limit theorem [4], [17], [19], which leads to Vdif = XR + jXI , where XR and XI are independent Gaussian variables, each with zero mean and variance σ 2 . A more general variant of the diffuse component follows from a multi-scatter physical setup [8], [17], [19]. This leads to the cases of Nakagami, Weibull and Cascaded Rayleigh channel, as well as compound channels, such as Suzuki and shadowed κ − µ [9], [22], [23], [26]. 5 We treat narrowband channel models with block fading, such that the power at which the packet is received remains constant and equal to P given with (3). The noise power is normalized to 1, such that P also denotes the Signal-to-Noise Ratio (SNR) at which a given packet is received. For each new packet, all RVs from (3) are independently sampled from their probability distributions2 . The average received power for the channel model (3) across many realizations is denoted by A and can be computed as: (a) A = E[P ] = ω µ X " ξ2 m=1 N X ! ρ2i,m " +E i=1 L Y ## 2 |Vdif,m,l | , (4) l=1 with E[·] denoting the expectation operator. Note that (a) is valid when we treat the reduced wave grouping model from [4]. In the subsequent analysis we assume normalized shadowing power, i.e. ω = 1 and ξ 2 = 1. The diffuse term depends on link signal correlation, while for a single link (L = 1) the average power of the elementary terms is E[|Vdif,m |2 ] = 2σ 2 . 3) Descriptive Metrics: The specular component vector balancing in the reduced wave group model of [4] is given via the peak to average ratio of the two dominant specular powers: ∆= 2ρ1 ρ2 . ρ21 + ρ22 (5) Furthermore, the power ratio of the specular components and the diffuse component per cluster, called k-factor is defined as kN = PN ρ2i , which in case of the multiple clusters gives [19]: Pµ PN 2 µ ρi,m 1X kN,m = m=1 i=1 . (6) κ= µ m=1 µ · 2σ 2 i=1 2σ 2 4) URLLC-Relevant Statistics: Let R denote the transmission rate of the packet. We assume that packet errors occur due to outage only, such that the PER  is given by:  = Pr(R < log2 (1 + P )) = Pr(P < PR ), (7) where PR = 2R − 1 is the minimal required power to receive the packet sent at rate R. Denote by  the target packet error probability (PER), also referred to as outage probability. Then, for each model the objective is to find PR , defined in (7) through the CDF F (PR ), obtained as: Z √PR  = F (PR ) = f (r)dr, (8) rmin 2 The reader may object that this assumption is not valid when long-term shadowing is treated, i.e. a sample for a given ρi is applicable to several packet transmissions. See Section IV-A for discussion about this assumption. 6 TABLE I CDF TAIL APPROXIMATIONS ˜, FOR DISTRIBUTIONS IN THE FOLLOWING SUBSECTIONS . GAIN OFFSET ( SCALE ) Channel Model TW Tail ˜ = limp→0 F (p)  1 − π1 asin 1−p 2 ∆ PR 4π∆r 3W TWDP κµ κµ/m κµ/α Suz Cas LN where r = 1 4π∆r FRayl (p(k1 + 1)) e FRice (p; k2 ) I0 (k2 ∆) (Γ(1 + 1/β)p) m β (m /Γ(m + 1)) p when ρ1 < ρ2 + ρ3 1 (k1 + 1)e −k2 1 I0 (k2 ∆) 1 β β m FNak (p; µ)FRice (1; κ)  m m FNak (p; µ)(1 + κ)µ κµ+m Γ(α+µ) (α−1)µ Γ(α) Fκµ (p; κ, µ) · 1 2 ln 10 PR 10 10 (σdB ( 20 )−µdB )   1+Γ 1+Γ −p 1−Γ ln p (1−Γ) 2 ( 1 −k1 Γ(1 + 1/β) m µ − 1 e 4 1 2 when ∆ → 1 (k2 + 1)e 1 ln(P )−aσ −µ 2 R l l 2 2σ 2 l m /Γ(m + 1) (e−κ (κ+1)µ)µ Γ(µ+1)  m µ µ µ (1+κ) Γ(µ+1) (1). Slope β 1 −k1 Wei Nak Offset α √ 2 π p Rayl Rice α AND LOG - LOG SLOPE ( SHAPE ) β = R ELATIVE POWER p = PR /A, d log(F ) FOLLOWING d log(p) m µ m κµ+m µ Γ(α+µ) (α−1)µ Γ(α) ακµ · 1 2 ln 10 10 10 (σdB ( 20 )−µdB ) µ 1 ln(p)+ln ) - 1 1+ - 10 ln 10 h a σdB −2 1+Γ (1−Γ)2 PR,dB −µdB 2 2σdB i √ P is the received envelope and rmin is the minimal value of the envelope in the support set of f (r), which is the Probability Density Function (PDF) of the specific channel model. The key to the approximations presented in this paper is the fact that, for URLLC scenarios,  is very small. III. C HANNELS WITH P OWER L AW TAIL S TATISTICS We analyze the behavior of common wireless channel models in URLLC-relevant regime and derive asymptotically tight approximations ˜ of their tail probabilities , that satisfy limPR →0 ˜ =  The common trait of all models considered in this section is that ˜ takes the form of a simple power law (1). with distribution-specific values of the parameters α, β. For the channel models that are important in practice (see subsections III-A-III-F), we also provide a simple tool to analyze the convergence of ˜ to . Specifically, we introduce the nonnegative approximation error function φ(PR ) that satisfies the inequality (see Appendix A): ˜(1 − φ(PR )) ≤  ≤ ˜(1 + φ(PR )). (9) 7 φ(PR ) increases monotonically with PR and satisfies limPR →0 φ(PR ) = 0. We say that ˜ converges   η ˜ −1 asymptotically to  in the sense that  − 1 ≤ η if PR ≤ φ for some small error 1+η tolerance η > 0. In other words, φ(PR ) can be used to compute the range of envelopes over which the relative tail approximation error is less than η. Table I summarizes the tail approximations that are derived in the sequel. A. Two-Wave Model (TW) We start with the common Two-Wave channel model [4], where µ = 1, N = 2 and L = 0, i.e., single cluster, two specular and no diffuse components. The envelope PDF is given by: 2r fTW (r) = r πATW where ATW = ρ21 +ρ22 ,  ∆2 − 1 − 2 , r2 (10) ATW i hp p ATW (1 − ∆), ATW (1 + ∆) . ∆ is given by (5) and r ∈ [rmin , rmax ] = By putting fTW (r) in (8) we obtain the CDF: 1 1  = FTW (PR ) = − asin 2 π 1− PR ATW ! ∆ . Bounding  from below leads to the tail approximation: r 2 p ? 1 PR , ˜ = π ATW where PR? = 1 P ∆ R − (11) (12) 1−∆ ATW . ∆ The approximation error function is (see Appendix A): r 4 ATW (ATW + PR? )PR? p φ(PR ) = . 3 2 (2ATW − PR? )3 (13) The upper bound on the power for error tolerance η can be evaluated numerically. Fig. 1 depicts the tail  for the TW channel. When ∆ < 1, the tail falls abruptly to zero at PR = ATW (1 − ∆). However, as it is seen from Fig. 1, the log-log slope that precedes this abrupt transition to zero is 1 2 (half a decade per 10 dB), which can be also see from (12). In the singular case ∆ = 1 (0 dB), the tail approximation is given by (12) with PR? = PR and the slope continues until −∞ dB3 . For example, if the log-log slope of 1 2 should be present at  = 10−6 , then we need to have ∆ > −4 · 10−6 dB, i.e. ρ2 very close to ρ1 , which is unlikely in practice due to the losses of the reflected wave. Hence, the two-wave model should be used with high caution when evaluating URLLC scenarios. 3 The case ∆ ≈ 1 has been referred to in the literature as hyper-Rayleigh fading [10] 8 10 0 10 -1 TW ∆T W = −3 dB 10 -2 ∆ 10 -4 10 -5 10 -6 -60 ∆T W = −0.12 dB 10 -3 dB ∆T W = −0.001 dB ǫ =0 TW Rayl -50 -40 -30 -20 -10 0 PR /A dB Fig. 1. Two-wave (11) and the classical Rayleigh (15) CDFs and their tail approximations (12), (16) (black dotted lines). For Two-wave: ρ1 = 1, while ρ2 is set to correspond to the ∆ (5) given at each curve. B. Rayleigh Channel (Rayl) This model, adopted in many wireless studies, has µ = 1, N = 0 and L = 1 (single cluster and diffuse component and no specular components) and the envelope PDF is [9]: r2 2r − ARayl e fRayl (r) = , ARayl (14) with average power ARayl = 2σ 2 . The CDF follows readily as: PR Rayl −A  = FRayl (PR ) = 1 − e , (15) which can be upper bounded by retaining only the first term in the Taylor expansion, resulting in the following simple power law approximation: ˜ = PR , ARayl (16) also known as the Rayleigh rule of thumb “10dB outage margin per decade probability” due to a log-log slope of β = 1. The approximation error function can be derived via an upper bound on the Taylor remainder, yielding the simple form (see Appendix A): φ(PR ) = PR . 2ARayl (17) 9 C. Rician Channel (Rice) This is an extension of the Rayleigh channel, featuring a specular component in addition to the diffuse one. The average received power is ARice = ρ21 + 2σ 2 = 2σ 2 (k1 + 1), where k1 = is the Rician k-factor and the PDF of the received envelope is [9]: rp  −k1 fRice (r) = fRayl (r)e I0 2k1 σ ρ21 2σ 2 (18) where I0 (·) is the modified Bessel function of 1st kind and 0th order. The tail can then be expressed in closed form in terms of the 1st order Marcum Q-function as follows: ! r p PR (k1 + 1) .  = FRice (PR ) = 1 − Q1 2k1 , 2 ARice (19) Bounding  from below via 1st polynomial expansion of Q1 , we arrive at the tail approximation: ˜ = PR (k1 + 1)e−k1 . ARice (20) The approximation error function obtains the form (see Appendix A):  P  k1 R (k +1) 1 A φ(PR ) = e 2 e Rice −1 , (21) The tail approximation of the Rician channel in URLLC-relevant regime has Rayleigh slope β = 1. However, Fig. 2 shows that before attaining the slope β = 1, the Rician CDF has a steeper slope compared to the Rayleigh one. In the context of wireless communications, this can be interpreted as an increased diversity order offered by the Ricean distribution. The lower the k1 −factor, the sooner the slope becomes identical to the Rayleigh one; in other words, as k1 increases, PR decreases for fixed error tolerance η which can be also see from (21). D. Weibull Channel (Wei) The Weibull channel is a generalization of the Rayleigh model, where the diffuse component p is given by |Vdif | = (XR2 + XI2 )1/β with β 6= 1. This model has been used in empirical studies as it offers increased freedom to fit the modeling of the diffuse part [14], [15]. As in the Rayleigh case, here we also have only a diffuse component, but the received envelope follows Weibull distribution [16]: fWei (r) = 1/β with AWei = (2σ 2 ) Γ(1 + 1/β). For β = 2βr2β−1 − r2β2 e 2σ 2σ 2 1 2 (22) we get the TW case, while β = 1 leads to the Rayleigh case. The tail is given as:  β P − Γ(1+1/β) A R  = FWei (PR ) = 1 − e W ei . (23) 10 10 0 10 -2 ǫ 10 -4 10 -6 k1 = 9 dB 10 -8 k1 = 12 dB 10 k1 = 16 dB -10 -50 -40 -30 -20 -10 0 PR /A dB Fig. 2. Rician CDF (19) and its tail approximation (20) (black dotted lines): The Rician k-factor k1 is indicated at the respective curves. Using first order Taylor expansion, we obtain the following tail approximation:  β PR ˜ = Γ(1 + 1/β) . AWei (24) Here β denotes the log-log slope and an example with β = 2 is shown in Fig. 3. The approximation error function for the Weibull channel is given by (see Appendix A): β  R Γ(1 + 1/β) APWei φ(PR ) = β .  PR 1 + Γ(1 + 1/β) AWei (25) E. Nakagami-m Channel (Nak) The envelope of this model behaves similarly to the Weibull model, although the diffuse pPm 2 2 component is modeled differently as r = |Vdif | = i=1 (XRi + XIi ), with m integer. This model can be interpreted as an incoherent sum of m i.i.d. Rayleigh-type clusters, each with mean power 2σ 2 and total power ANak = m · 2σ 2 . The PDF of the envelope r is given by [17]:  2 m 2 r2 r 2mm 2 − r2 2m−1 −m ANak 2σ , (26) fNak (r) = r e = e Γ(m)Am rΓ(m) 2σ 2 Nak where we interpret m ∈ R ≥ 1 2 for generality. For m = 1 2 we get an exponential, while for m = 1 a Rayleigh distribution. The CDF is given as:   PR γ m; m AN ak  = FNak (PR ) = Γ(m) (27) 11 where γ(a, x) is the lower incomplete gamma function. The power law tail approximation can be obtained via the upper bound γ(a; x) ≤ xa /a [18], resulting in:  m mm PR ˜ = . Γ(m + 1) AN ak (28) We see that (28) has the same flexibility and slope behavior as the Weibull model (for m ≥ 21 ), but a different offset. This can be also observed in Fig. 3 with m = 2, where the wide shoulder sends the tail (28) to lower levels compared with the Weibull case. By bounding the lower incomplete gamma function in (27) from both sides, i.e., e−x xa /a ≤ γ(a; x) ≤ xa /a [18], we derive the approximation error function: PR Nak −m A φ(PR ) = 1 − e ≤e PR Nak mA − 1. (29) F. κ − µ Channel (κµ) The κ−µ model was developed in [19] as a generalization to the Nakagami model, by considerpPµ 2 2 ing incoherent sum of µ Rician type clusters, i.e. envelope r = i=1 (XRi + pi ) + (XIi + qi ) where XRi + jXIi are complex Gaussian diffuse components (all same mean power 2σ 2 ) and pi + jqi the corresponding specular components with arbitrary power ρ2i = p2i + qi2 . Here κ is a generalized Rician type k-factor defined in (6). Consequently, the total mean power is Aκµ = µ(1 + κ) · 2σ 2 and the PDF of r is given by [19]: " r #  µ  2  2 r 2(κµ)(1−µ)/2 r r − √ √ fκµ (r) = e 2σ2 Iµ−1 2 κµ 2 , µ ≥ 0 κµ 2 2 2σ e 2σ 2σ (30) Again, for generality, we interpret µ ∈ R. The CDF in closed form is described via the generalized Marcum Q function, from [19]:   q p 2κµ, 2(1 + κ)µPR /Aκµ Fκµ (PR ) = 1 − Qµ (31) Using a first-order polynomial expansion of the generalized Marcum Q-function, we obtain the following tail approximation: µ (e−κ (κ + 1)µ) ˜ = Γ(µ + 1)  PR Aκµ µ = FN ak  PR ;µ Aκµ  FRice (1; κ)µ (32) i.e. a multiplicative form of the previous Rician and Nakagami-m tail approximation. The approximation error function obtains the simple form (see Appendix A):   P κµ (κ+1)µ A R κµ φ(PR ) = e 2 e −1 . (33) 12 10 0 10 -2 ǫ 10 -4 10 -6 Nak 10 κ−µ Wei -8 10 -10 -50 -40 -30 -20 -10 0 PR /A dB Fig. 3. Nakagami-m (27) (m = 2), Weibull (23) (β = 2) and κ − µ (31) (κ = 3.9, µ = 2) CDFs and their tail approximations (28),(24),(32) (black dotted lines). We see that both ˜ and φ(PR ) reduce to the forms derived earlier as special cases; specifically, for µ = 1 the κ − µ model reduces to a Rician situation, while for κ = 0 the Nakagami-m situation emerges. The tail (32) in a typical Rician setting (κ = 3.9, µ = 2) is seen in Fig. 3, where it can be seen that the tail is pushed to lower probabilities compared to Nakagami and Weibull models. G. Generalizations We analyze several generalizations of the channels presented in the previous subsection and derive their power law tail approximations. First, we explore the transition of the behavior from few paths to many paths. In this sense, we expand the previous TW model to cater for 3−vector components or include a diffuse part that is generalized compared to the models with diffuse part in the previous section. As it will be shown, both cases result in a tail behavior that conforms to a behavior dominated by diffuse components. In other words, three specular components can be sufficient to produce the behavior of a Rayleigh diffuse component at URLLC levels. Another important generalization is to use combined short and long term processes, particularly when such are inseparable. We consider three combined models in following sub-sections: 1) Log-normal shadowed Rayleigh fading, i.e. the Suzuki distribution typically used to model Macro-cell behavior [9]; 2) κ − µ fading with Nakagami-m shadowing; and 3) κ − µ fading with 13 inverse Gamma distributed shadowing. While 1) is a classical case, 2) and 3) have been found useful for modeling close range propagation [22], [23], [26]. We can think of the combined channel models to be applicable to the following situation. When there is only short term block fading, the outage probability can be controlled by selecting the rate R according to the known average power of the short term channel. Equally important is the specular component balancing or ratio towards the diffuse parts, captured by ∆, κ in (5), (6). However, when the sender does not have a reliable estimate of the average power (or impact of specular components), then this uncertainty can be modeled by assuming that the average power or ∆, κ are random variables. The independent sampling from the shadowing distribution is a pessimistic case that assumes sporadic transmissions, sufficiently separated in time. 1) Three-Wave Model (3W): We consider the Three-Wave generalization of the TW model. Here N = 3, Vdif = 0, received envelope r = |ρ1 + ρ2 ejφ2 + ρ3 ejφ3 | and average power A3W = P3 2 n=1 ρn . The probability density function [4] is given by:   2  √ ∆r  2√ r K ∆2r ≤ ρ1 ρ2 ρ3 r π ρ1 ρ2 ρ3 ρ1 ρ2 ρ3 r   (34) f3W (r) = 2  2r K ρ1 ρ22ρ3 r > ρ ρ ρ r ∆ 1 2 3 r π ∆r ∆ r for r ∈ [rmin , rmax ], and it is 0 otherwise, with rmin = max(2 max(ρ1 , ρ2 , ρ3 ) − ρ1 − ρ2 − ρ3 , 0), rmax = ρ1 + ρ2 + ρ3 . In (34), K(·) is an elliptic integral of the first kind4 and the quantity ∆r is defined as: ∆2r = 1 [(r + ρ1 )2 − (ρ2 − ρ3 )2 ][(ρ2 + ρ3 )2 − (r − ρ1 )2 ]. 16 (35) Without losing generality, we can take ρ1 ≥ ρ2 ≥ ρ3 and define the difference ∆ρ = ρ1 −(ρ2 +ρ3 ), such that rmin = max(∆ρ , 0). Three cases can be considered: (1) rmin = 0 when ∆ρ < 0; (2) rmin = 0 and ∆ρ = 0; and (3) rmin > 0 otherwise. Here we treat the case ∆ρ < 0, which sets the basis for the reader to treat the other two cases. The integral (8) is evaluated for values √ r ∈ [0, PR ] that are very small and taking r → 0: lim ∆2r = [ρ21 − (ρ2 − ρ3 )2 ][(ρ2 + ρ3 )2 − ρ21 ] > 0 r→0 which implies that ∆2r > ρ1 ρ2 ρ3 r holds in (34). With r → 0:   ρ1 ρ2 ρ3 r r π r r→0 f3W (r) → 2 K ≈ 2 2 π ∆r ∆r π ∆r 2 4 Convention of [4] is K(m) with argument m = k2 (instead of K(k) with modulus k). (36) (37) 14 10 0 10 -1 − k∗ 3 = 04 00 10 -3 10 -6 -60 3 1. 8 1. 7 k∗ 3 = k∗ 3 = 10 -5 dB dB 10 -4 k∗ 3 = − 0. ǫ dB 2. 82 dB 10 -2 -50 -40 -30 -20 -10 0 10 PR /A dB Fig. 4. Three Wave CDF (numerical integration of (34)) and the tail approximation (38) (black dotted line): Here ρ1 = 1 while ρ2 and ρ3 are set according to k3∗ indicated at the curves. The solid dots depicts the locations of the “breaking points” where the asymptotic slopes change (as determined by the value of |∆ρ |2 ). where we have used limx→0 K(x) = π2 . Approximating ∆2r as a constant for small values of r, we get the following tail approximation: r2 PR ˜ = = 4π∆r 4π∆r (38) such that the log-log linear slope is β ≈ 1. In the singular case rmin = 0 and ∆ρ = 0 it can be shown that β = 43 , while the case rmin > 0 has a slope of 21 , 34 or 1, before an abrupt fall to zero 2 . when PR = rmin The 3W CDF is shown in Fig. 4 for ρ1 = 1 and different variations of ρ2 and ρ3 . The curves are labeled by k3∗ = ρ21 2 (ρ2 +ρ23 ) 5 2 ρ| = −50dB . We select to represent two cases with identical 10 log |∆ A3W that are seen to diverge significantly when PR A3W < −40dB. The difference between these two cases emerges due to the different sign of ∆ρ , which in one case results in rmin > 0 (ρ2 =0.7850, ρ3 =0.2109) and in the other case rmin = 0 (ρ2 =0.7914, ρ3 =0.2126). The latter case has a log-log slope of β = 1, identical to the Rayleigh distribution. Hence, if the sum of the two smallest components can cancel and overshoot the strongest component, the URLLC-level behavior of the 3W model is practically identical to that of a Rayleigh channel in terms of a slope. 5 Note that this metric differs from the k-factor which involves diffuse parts. k3∗ is the ratio between specular component powers only 15 2) Two-Wave Diffuse Power (TWDP) Channel: In this model N = 2 and Vdif , with envelope r = |ρ1 +ρ2 +Vdif | and average received power ATWDP = ρ21 +ρ22 +2σ 2 [4]. The PDF is obtained by averaging of the Rician PDF [5]: 1 fTWDP (r) = 2π Z 2π fRice (r; k2 [1 + ∆ cos (ψ)]) dψ (39) 0 with ∆ defined in (5) and k2 in (6). The integration over ψ involves only I0 (·) and the exponential R 2π k ∆ cos ψ PR 1 terms in (18). Using I0 ≥ 1 for ATWDP e2 · 1 dψ =  4k2 (k12 +1) , this integration is ∼ 2π 0 I0 (k2 ∆), i.e. it leads to a constant with respect to r. Hence, the tail can be lower-bounded through a scaled Rician tail:  = FTWDP (PR ) ≥ FRice (PR ; k2 ) I0 (k2 ∆) = ˜ (40) and the analysis from the Rician case can be directly applied, scaled by I0 (k2 ∆). From Fig. 5 it can be seen that TWDP6 starts to differ from a Rician model (with k1 = k2 ) when ∆ is sufficiently high, such that ρ2 can be distinguished from Vdif . The second specular component ρ2 lifts-off the lower tail as ∆ → 0 dB, while preserving the Rayleigh tail slope. Note that, in order to reach the extreme slope of the singular TW model at the URLLC levels, one needs ∆ = 0dB and k2 in range 50 to 60dB, which is very unlikely to happen in practice. 3) Suzuki Channel (Suz): This is a compound channel consisting of a diffuse component only, which is a mixture between a Rayleigh envelope and a log-normal varying mean [9]. The compound envelope is r = |XR + jXI |, where XR and XI are zero-mean Gaussian variables with variance σLN = eN that has a log-normal distribution. The PDF and CDF of the Suzuki channel can be found as follows. Let us denote by A the average power used to generate Rayleigh-faded power level A. The power A is log-normal distributed, such that we can obtain its PDF from the PDF of the log-normal envelope (53) by substituting A = r2 . This leads to the following joint distribution of P and A: − 1 1 P √ e fSuz (P, A) = e− A · A 2Aσl 2π ( 12 ln A−µl ) 2 2σ 2 l For given PR , the outage probability can be calculated as follows: Z PR Z ∞ = dP fSuz (P, A)dA 0 6 (41) (42) 0 No tractable closed form of PDF or CDF exists. In [4] the PDF is approximated, while we use a complete expansion as in [13]. However, due to the numerical sensitivity at URLLC levels, it requires the use of high-precision numerical tools. 16 10 0 10 -2 dB dB ∆T W DP increases = −6 k2 = ∆T =0 ǫ 10 -4 dB 16 P WD k2 dB =9 ∆T W =− k2 = 16 dB ∆T W 10 DP DP B 6d -6 B 6d 10 -8 B 2d ∆T P WD =− 1 k2 = 10 -10 -50 -40 -30 -20 -10 0 PR /A dB Fig. 5. Two-Wave Diffuse Power CDF and its tail approximation (black dotted line): ρ = 1, while ρ2 is set to correspond to the ∆ (5) given at each curve. The corresponding k2 (6) is also given. The upper bound for (42) is obtained by noting that P A ≥ 0 and it can be removed from (41), after which we get: 2 Z PR Z ∞ − 1 √ e  ≤ dP 2A2 σl 2π 0 0 PR 2 2 = e2σl −2µl PR = · e4σl ASuz ( 12 ln A−µl ) 2σ 2 l dA (43) 2 where ASuz = e2σl +2µl = ALN [37]7 . The lower bound can be found by using the inequality e−x ≥ 1 − x which leads to: − 1 √ e fSuz (P, A) ≥ 2 2A σl 2π and results in PR 4σl2 ≥ e − ASuz  PR ASuz 2 2 ( 12 ln A−µl ) · e12σl ≈ 2σ 2 l 2 (1 − P ) A 2 2 ln(10) PR 4σdB 202 e = ˜ ASuz (44) (45) For URLLC-relevant levels it is PR  ASuz , such that the upper bound can be treated as tight. The tail has a Rayleigh-like slope of β = 1, but pushed to lower levels as seen in Fig. 6. 2 2 Note that ASuz = E[r2 ]e2σl +2µl , where E[r2 ] = 2 · σRayl is the mean power of the Rayleigh part. [37] has used unit √ 2 variance Gaussians σRayl = 1. Without loss of generality we assume σRayl = 1/ 2, such that all mean level shifts are attributed 7 solely to µl . 17 10 0 10 -2 Suzuki ǫ 10 -4 10 -6 κµ/m κµ/α -40 -30 10 -8 10 -10 -50 -20 -10 0 PR /A dB Fig. 6. Shadowed CDFs from (47) and numerical integration of (42),(49) and their tail approximations (45),(48),(51) (black dotted lines) and heuristic CDF expansion (52) (dash-dot curve): Suzuki (σdB = 6dB), κ − µ/m and κ − µ/α (κ = 3.9, µ = 2, m = 0.25, α = 1.5). 4) Nakagami-m shadowed κ − µ Channel (κµ/m): Shadowing the total signal has been investigated in [22], [23], but provides complicated PDF and no known closed-form solution for the CDF. A model that considers shadowing of only the dominant signal parts has been P developed by [26]. The instant power is p = µi=1 (XRi + ξpi )2 + (XIi + ξqi )2 , where ξ is a power normalized Nakagami-m distributed shadowing amplitude acting on specular components pi + jqi = ρi ejφi . The closed form PDF of PR (with p = PR /Aκµ/m ) is given by [26]: m    µµ m κ(1 + κ)µ2 µ−1 −(1+κ)µ·p p e p fκµ/m (PR ) = 1 F1 m; µ; Γ(µ)Aκµ/m κµ + m κµ + m (46) with 1 F1 (·) Kummer’s function of the first kind [43]. Essentially, the first part is a Nakagami-m PDF of order µ, while the latter part holds a µ−order modified Rician impact. The tail is:   µ m m µµ PR µ  = Fκµ/m (PR ) = (κ + 1) Φ2 (b1 , b2 ; c, x, y) (47) Γ(µ + 1) Aκµ/m κµ + m | {z } {z } | →e−κµ |m→∞ ≈FN ak (p;µ) µµ (κ + 1)µ ≈ Γ(µ + 1)  PR Aκµ/m µ  m κµ + m m = ˜ (48) where Φ2 (·) is Humberts function [27] with arguments (b1 = µ − m, b2 = m, c = µ + 1, x = PR m −µ(1 + κ) Aκµ/m , y = x κµ+m ). The power law tail approximation follows from [26, eq. (13)]. Since ξ → 1 when m → ∞, expression (47) should reduce to the regular κ − µ in (32); this m is indeed so, as x+m → ex and (47) is evidently in accordance with (32). Fig. 6 shows m 18 a strongly shadowed example (m = 0.25). It can be noticed that the shoulder is significantly broadened to a degree that the elevation of the shoulder visible in the regular κ − µ case, has vanished. This is expected as only the LOS part has been shadowed, thereby effectively averaging the κ − µ distributions shape (κ-factor). Consequently, the tail is being pushed to significantly lower outages. 5) Inverse Γ−shadowed κ − µ Channel (κµ/α): Shadowing the total κ − µ (30) fading envelope rκµ = |Vκµ | by an inverse Gamma (Γ−1 ) distributed varying mean power ω = Aκµ , leads to a closed form PDF but no tractable closed-form CDF solution [23]8 . The combined signal √ (r = rκµ ω) PDF is obtained [23, (6)] by averaging the conditional envelope PDF fκµ|Aκµ (30)  βα 1 · exp − ωβ , with shape α > 0 and scale over the mean power statistics fΓ−1 (ω) = Γ(α) ω α+1 β > 0 parameters. The combined signal power PDF of [23, (10)] can be written as κµ · c/Aκµ/αβ µ−1 (e−k /κµ)µ fκµ/αβ (PR ) = · x 1 F1 (α + µ; µ; x) B(α, µ) (c · p + 1)α−1 (c · p + 1)2 | {z } (49) dx/dPR with B(·, ·) the beta function and argument scaling c = power is p = r2 E[r2 ] tail levels p  1 c = PR Aκµ/αβ µ(1+κ) β c·p in x = κµ c·p+1 . The relative and Aκµ/αβ is the mean power of the combined signal. For lower or α → 1+ , (c · p + 1)α−1 → 1 and constraining this approximation to the leading term only, we essentially have a function of form f (x) = xb−1 1 F1 (a, b, x). To obtain the R Γ(b) CDF we make use of f (x)dx = xb Γ(b+1) 1 F1 (a, b + 1, x) [25]. Thus, via variable transform and reordering of terms, we arrive at the following upper bound for the CDF:  µ   e−κµ c·p c·p  = Fκµ/αβ (PR ) ≤ · 1 F1 α + µ; µ + 1; κµ µB(α, µ) c · p + 1 c·p+1 (50) which is exact in the limit α → 1+ . Furthermore, we can simplify (50) as 1 F1 (a, b, x) → 1 for x → 0 and realizing that the scale β in [23] is set arbitrarily, such that fΓ−1 is not normalized. As ω = E[ω] = β α−1 valid for α > 1 [23], normalizing shadowing by setting ω = 1, we get β = α − 1. Thus, we can represent the impact of the shadowing through a single parameter: Fκµ/α (PR ) & (µ(1 + κ)e−k )µ µ Γ(α + µ) p = ˜, Γ(1 + µ) (α − 1)µ Γ(α) | {z } (51) ≈Fκµ (p;κ;µ) 8 Very recently [24, (3.14)] has provided a closed-form CDF. However, this is provided through a complex Kampé de Fériet function, for which no readily available numerical evaluation exists in tools such as MatlabTM . Furthermore, no simple analytical approximation seems available to be used in an URLLC setting. The underlying inverse Gamma PDF in [24, (3.14)] seems normalized in a skewed manner with ω = α . α−1 Thus we make our approximation analysis based on the PDF in [23]. 19 i.e. in form of a scaled κ − µ tail, representing a lower bound. However, for α . 1 other normalization methods must be used. We can heuristically reintroduce the denominator term (c · p + 1)α−1 into the leading term of (50) for larger arguments: e−κµ Fκµ/α (PR ) & µB(α, µ)(c · p + 1)α−1 where we can redefine c = µ(1+κ) α−1  c·p c·p+1 µ   c·p · 1 F1 α + µ; µ + 1; κµ = ˜ (52) c·p+1 via the above normalization. This result provides a significantly better fit than (50) or (51), especially in a strongly shadowed Rician regime (κ/α & 1), as it can be seen in Fig. 6. Furthermore, this expanded expression seems to act as a lower bound. It is also seen that the elevated shoulder from the underlying κ − µ signal is better preserved than in the case of the previous κµ/m model, while also having strong shadowing (α = 1.5) that indicates that the complete signal has been shadowed. IV. OTHER C HANNELS In this section we analyze two special models that do not exhibit power-law tail behavior and derive their corresponding tail approximations. First, we consider the log-normal distribution [9], as a classical reference distribution for shadowing. Next, we treat cascaded channel type models that arise in NLOS propagation, backscatter communication and in ’pin hole’ channels [9]. The two models also represent two extremes, the macro scale (log-normal shadowing) and short range (e.g. device-to-device). Furthermore, these models can be used as instances to illustrate cases that do not follow the power law in the diversity analysis presented in the next section. A. Log-Normal Channel (LN) In this model there is a single specular component N = 1 and no diffuse component. The specular component is not constant, but subject to a log-normal shadowing, such that log-envelope ln(r) is modeled as Gaussian variable [9] with PDF: 2 (ln(r)−µl ) − 1 1 2σ 2 l √ e fLN (r) = Nln(r) (µl , σl ) = r rσl 2π with logarithmic mean and standard deviation µl = E[ln(r)] = µdB ln(10) and σl = 20 (53) p E[ln(r)2 ] − µ2l = 2 σdB ln(10) . The average power is ALN = e2σl +2µl [9] and the CDF 20  = FLN (r) = 1 2 + 12 erf(x(r)) (54) 20 √ with x = (ln(r) − µl )/(σl 2) and erf being the error function. Using Bürmann-type asymptotic   √ 2 2 1 −x approximation [35] leads to FLN (x) ≈ 2 1 + sgn(x) 1 − e ≈ 14 e−x , when omitting higher order terms and approximating the square root for |x|  0. A tighter approximation can be obtained if we use FLN (x) ≈ 14 e−f (x) with a polynomial fitting function f (x) [36]. Comparing 1 −x2 e 4 with (54), it appears to be shifted proportionally to σl , such that: 2  = FLN (PR ) ≈ 41 e log  ≈ log 14 − − ( 12 ln(PR )−aσl −µl ) 2σ 2 l = ˜ (55) 1 (PR,dB − aσdB − µdB )2 2 ln(10) 2σdB (56) With a = 0.223, the relative error is η . 10−1 for 10−12 ≤  ≤ 10−2 and 3 ≤ σdB ≤ 24dB. The deviation on the margin matters most for outage analysis and is here below 1 dB. 3 This accuracy is still very useful, considering the simplicity of the expression for analytical studies. Finding a root of the second order equation in ln(PR ) (55) we get √ √ −1 2[(aσl +µl )+ 2σl − ln()+ln(1/4)] PR = FLN () ≈ e For a given PR , we can find the log-log slope as β = d log(FLN ) d log(PR ) ≈ 10 ln 10 h a σdB (57) i PR,dB −µdB . − 2 2σ2 dB From Fig. 7 it is observed that, for large σdB , a log-normal channel can exhibit extreme slopes when the level PR A is in the region −10 to −30 dB, which makes it hard to distinguish from a TW or TWDP channel. However, when going towards URLLC-relevant levels, the deviation from a linear slope is noticeable. B. Cascaded Rayleigh Channel (Cas) This model also contains only a diffuse component, which is a product of the envelopes of two Rayleigh links r1 and r2 . The compound received envelope is r = r1 r2 = |XR1 + jXI1 | · |XR2 + jXI2 | with PDF equal to [30], [31]9 : fCas (r) = where rΓ = r . σ1 σ2 (1−Γ)  √  rΓ I0 rΓ Γ K0 (rΓ ) σ1 σ2 (58) Using (58) we get ACas = E[r2 ] = 4σ12 σ22 (1 + Γ) = P¯1 P¯2 (1 + Γ) with correlation coefficient Γ between powers P1 = r12 and P2 = r22 . In and Kn are the Modified Bessel functions of 1st and 2nd kind, of order n. The CDF follows as: 9 Ref. [31] uses the convention Γ = ΓV1 ,V2 , whereas we use Γ = ΓP1 ,P2 as in [30]. Note that for Rayleigh fading |ΓV1 ,V2 |2 = Γ2XR 1 ,XR2 + Γ2XR 1 ,XI2 = ΓP1 ,P2 ≈ Γr1 ,r2 and Γ2XR 1 ,XI2 ≈ 0 for random links [32]. 21  √   √  √  = FCas (PR ) = 1 − rΓ [ Γ· I1 rΓ Γ K0 (rΓ ) + I0 rΓ Γ K1 (rΓ )] (59) Approximating the Bessel functions for rΓ  1, the general case (Γ < 1) simplifies as h r  i rΓ2 Γ + (2γ − 1)  = FCas (PR ) ≈ − (1 − Γ) 2 ln 4 2   PR 1 + Γ PR 1 + Γ ≈− ln = ˜ (60) ACas 1 − Γ ACas (1 − Γ)2 where γ = 0.5772.. is Euler’s constant. The log probability is:    PR 1+Γ PR 1 + Γ log(˜) = log + log + log − ln ACas 1−Γ ACas (1 − Γ)2 and is valid below the knee point to β ≈ 1 + ln  PR ACas 1 +ln 1+Γ (1−Γ)2 PR ACas  2 1 (1−Γ) 4 1+Γ . The slope is found as β ≈ and it gradually approaches a Rayleigh slope (61) d log(˜ ) P , leading d log A R Cas R → for APCas 0. For Γ = 0 the model collapses to the well known double-Rayleigh model [8]. When ˜ < 1/e, the function in (60) can be inverted [33] via a single Lambert W branch [34]: 1 1−Γ −1  PR = FCas () ≈ −ACas ˜ ˜ 1 + Γ W−1 − 1−Γ (62) −1 For the singular case of Γ = 1(r1 = r2 ), simple deduction yields rCas = r1 r2 = FCas () = q √ −1 2 R and the slope β ≈ 21 is identical to FRayl ()2 = rRayl . Thus, FCas (PR ) = FRayl ( PR ) ∼ APRayl the singular case of a TW model. It can be concluded from Fig. 7 that a the log-log behavior of cascaded Rayleigh fading can be represented by two different slopes with a breakpoint. V. S IMPLIFIED A NALYSIS OF D IVERSITY S CHEMES In practice, attaining very high reliability levels with reasonable power can only happen by having high levels of diversity at the receiver. Our analysis has shown that the tail approximation at the URLLC levels mostly has the form given in (1), which can be used for simplified diversity analysis, in particular in cases in which the full PDF/CDF are not tractable. For small terminals, the main impairment towards exploiting multi-antenna (multi-branch) diversity is the branch power ratio (BPR), defined for a pair of the m−th and n−th antenna as BPRmn = Am An [38], [39]. On the other hand, when operating in higher frequency bands, the terminal dimensions become comparable or larger than the wavelength, which leads to low branch correlations. Therefore, in the following we assume that the receiver has M antennas 22 10 0 10 -1 ǫ 10 -2 10 -3 FCas , Γ =0 Γ =0.8 Γ =0.99 Γ=1 FLN , σdB =8dB σdB =12 dB σdB =16dB ǫ̃ 10 -4 10 -5 10 -6 -60 -50 -40 -30 -20 -10 0 10 PR /A dB Fig. 7. Cascaded Rayleigh (59) with equal power links (σ1 = σ2 = σ) - and Log-Normal (54) distributions, for different values of the parameters. Dotted curves: tail approx. (60), (55). that are not correlated, i.e., the received signals across antennas are independent non-identically distributed (i.n.i.d.) random variables (RV). In Selection Combining (SC), only the strongest signal among the M antennas is selected: PR,SC = max(P1 , .., PM ) (63) For independent branches, the CDF can be expressed as a simple product of the individual CDFs across branches: M Y  = FSC (PR ) = Fm (PR ) (64) m=1 When Maximum Ratio Combining (MRC) is used, the received power is: PR,MRC = M X Pm (65) m=1 The simplest distribution we consider, is the Rayleigh case. For uncorrelated Rayleigh branch signals, tractable expressions for the tail exist when all BPR=1 or all BPR6=1 [9], [40]. [40] suggests a simple approximation for the MRC PDF for low powers, valid for all BPRs; using the approximation in [40] one can readily obtain:  M Y M M 1 Y PR 1 PR αm  = FMRC (PR ) ∼ αm = = ˜ M ! m=1 Am M ! A1 BPRm1 |{z} |{z} m=1 {z } αMRC | {z } αMRC | ∼FSC which has simple product form as in SC, shifted by αMRC = ∼FSC 1 . M! (66) 23 Next, we derive an approximation of the CDF for general M-branch MRC at low powers (see Appendix B); the approximation is of the same form as (66): QM  β M Y PR m m=1 Γ(1 + βm ) = ˜  = FMRC (PR ) ≈ αm P Am Γ(1 + M m=1 βm ) m=1 | {z }| {z } αMRC ∼FSC = Q (67) Fm Note that this solution also splits into a MRC weighting term αMRC , which depends solely on the branch slopes β and correctly collapsing to 1 for M = 1, and a term similar to SC Q FSC = Fm , which involves the offsets α. The distribution specific parameters α, β for this simple expression, are given in Table I. Inserting ακµ/m and βκµ/m from Table I, does indeed produce the distribution specific MRC solution for shadowed κ − µ fading given in [26, (18)]. When all branch slopes are equal βm = β, we get:  βM Y M Γ(1 + β)M PR αm  = FMRC (PR ) ≈ = ˜ Γ(1 + βM ) A1 BPRβm m=1 {z }| | {z } αMRC (68) Q ∼FSC = Fm with BPRm = Am /A1 . A heuristic simplification is αMRC ∼ 1 , M !β M = 4 and . 1.5dB for M = 8, both at 10−6 probability and for with outage error . 1dB for 1 2 . β . 2. For the particular Rayleigh case of β = 1, the solution collapses to the known result of (66). Using (67), we can bound the tail approximation error using the approximation error functions derived previously, resulting in: φMRC (PR ) = (1 + φmax (PR ))M − 1, (69) with φmax (PR ) = max(φ1 (PR ), . . . , φM (PR )). Using Bernoulli approximation, we arrive at the intuitive expression φMRC (PR ) ≈ M φmax (PR ), which can be used for quick evaluation of the upper bound on the power PR for given error tolerance η. Finally, based on (67) it is easy to make a heuristic generic expansion by considering local log-log linear approximation of any CDF tail. This is e.g. the case for Log-Normal and Cascaded Rayleigh models 10 , where the branch slopes depend on the power levels β(PR /A), such that:  = FMRC (PR ) ≈ αMRC (β1 (PR )..βM (PR ))FSC (PR ) = ˜ (70) with αMRC given in (67) or simplified in (68) when all branches have the same log-log slope. With this structure and availability of slopes β(PR ), one can use the full CDFs in FSC . 10 or some of the more elaborate tails, like for κ − µ type models etc. 24 Fig. 8. MRC with M = 4 for i.i.d. TWDP, Log-Normal and Double-Rayleigh (Cascaded Rayleigh with Γ = 0). Also MRC with M = 3 for one branch of each of the distributions. Bolder curves are obtained by simulation, thin dotted lines are the tail approximations. (40), (60), (55) and MRC (70), (67). Bold dots represent 1dB deviation of tail vs. simulation. Fig. 8 shows Monte Carlo simulation with 108 samples of TWDP, Log-Normal and DoubleRayleigh (Cascaded Rayleigh with Γ = 0) distributions with different mean powers. Each distribution is further circularly shifted to provide uncorrelated copies for i.i.d. M = 4. It is observed how the single branch tail approximations (thin lines) follow the simulations up to the onset of the shoulders. The 4−branch MRC tail approximation shows very good fit at URLLC probabilities - bold dots indicate point of 1dB deviation to the simulation. Furthermore, from the log-normal and 3-branch cross-distribution MRC, it is observed that the heuristic expansion in (70) indeed provides very useful results. VI. D ISCUSSION AND C ONCLUSIONS We have investigated the properties of wireless channel models in the URLLC regime and developed approximations of the tail distributions. Furthermore, our analysis has shown that, for a wide range of practical models, the outage probability at URLLC levels depends on the minimal required decoding power through an exponent β which, for the case of Rayleigh fading is β ≈ 1. More importantly, it has also been shown that the outage probability also depends on a tail offset α, which is strongly dependent on specific specular and diffuse part combinations. 25 The previous URLLC studies [41] have adopted Rayleigh models. Our analysis reveal it is not so much the existence of such a power law outage model with exponent β, but rather the tail offset α, which needs attention in assessing the tail probability. We have seen that hyper-Rayleigh fading, can result in an exponent β = 1 2 at URLLC levels, which leads to very conservative, if at all useful, rate selection. Several models can lead to hyper-Rayleigh fading, despite the fact that the physical basis of the models are different, such as: cancellation of dominant paths, cascaded links, multiple scattering or shadowing. Hence, when conducting empirical studies that work with these models, one should account for the large uncertainty that occurs when assessing the models at the URLLC levels and collect proportionally large number of samples. Finally, we have provided a simplified analysis of MRC diversity for power-law tails, as well as a heuristic generic expansion. This paves the way for more elaborate diversity analysis, which is, on the other hand, vital for achieving URLLC operation with reasonable data rates. We have used a comprehensive list of channel models and in this sense the present paper can be considered as a reference work for the channel models in URLLC regime. The necessary next step is to relate the models to the experimental results that are relevant for the URLLC regime. This requires a careful design of the measurement procedures, considering that one needs to capture rare events and thus potentially use a large number of measurement samples. ACKNOWLEDGMENT This work was supported by the H2020 European Research Council (ERC Consolidator Grant Nr. 648382 WILLOW). A PPENDIX A D ERIVATION OF THE A PPROXIMATION E RROR F UNCTION Two-Wave Model: We derive the tail for the general case when ∆ ≤ 1 (which includes ∆ = 1 as a special case). By definition the tail is the solution to the following integral: Z √PR 2r r = √  2 dr. ATW (1−∆) r2 2 πATW ∆ − 1 − ATW We introduce the following variable x ≡ x(r): r 1 2 1−∆ x= r − ATW . ∆ ∆ (71) (72) 26 Using change of variables, the integral (71) can be written as follows: Z √PR? 2x r = 2 dx,  0 x2 πATW 1 − 1 − ATW (73) where PR? = x2 (PR ). Even though the CDF has a closed form expression, bounding the sum of the higher order terms of its series expansion is difficult. We use an alternative approach instead. Specifically, we expand the integrand into Taylor series in the interval [0, x), x ≥ 0, x → 0 using Lagrange form for the remainder (i.e., the sum of the remaining higher order terms): f (x) = f (0) + f 00 (δ) 2 f 0 (0) x+ x, 1! 2! (74) for δ ∈ (0, x). Then, we bound the remainder from above, relying on the fact that integrating will not change the inequality; we obtain the following: ! r Z √PR? 1 2 4(ATW − δ 2 ) = x2 dx + p 2 3 π A π (2ATW − δ ) TW 0 √ ! r Z PR? ? 1 2 4(ATW − PR ) x2 dx ≤ + p ? 3 π ATW π (2ATW − PR ) 0 p ? r (PR )3 2 p ? 4(ATW − PR? ) 1 PR + p = π ATW 3 π (2ATW − PR? )3 ! r r 1 2 p ? 4 ATW (ATW − PR? )PR? p = PR 1 + , π ATW 3 2 π (2ATW − PR? )3 (75) (76) (77) (78) Recognizing that (78) can be written as  ≤ ˜(1 + φ(PR )), we extract φ(PR ) as the second term in the brackets in (78), completing the derivation. Note that in (76) we used the fact that the multiplicative term in front of x2 increases monotonically with δ ∈ (0, x); hence we bound it p from above with x = PR? . Rayleigh Model: Deriving the approximation error function follows similar steps as in the TW case, except that we directly bound the Lagrange remainder of the Taylor series expansion of the tail in the interval [0, PR ), PR → 0. Hence, we obtain: PR P 2 − PR PR P2 PR = − 2R e ARayl ≥ − 2R = ARayl ARayl ARayl ARayl ARayl  PR 1− 2ARayl  , (79) which completes the derivation. Weibull Model: We use the inversion: PR = 1 AWei (− ln(1 − )) β , Γ(1 + 1/β) (80) 27 and the following bounds: + 2 ≥ − ln(1 − ) ≥ . (1 − ) (81) The upper bound in (81) has been derived by bounding from above the remainder of the Taylor expansion of the function − ln(1 − ) in the interval [0, ), i.e.: 2 2 , ≤ 2(1 − )2 (1 − ) (82) for  ≥ 0. Replacing (81) into (80), we get:   β1 1 1 AWei  AWei β β , ≥ PR ≥ Γ(1 + 1/β) 1− Γ(1 + 1/β) (83) which, after inverting for , can be written as: β  PR Γ(1 + 1/β) AWei 1   ˜ ≥  ≥ ˜ β = ˜ 1 − β  .   R R 1 + Γ(1 + 1/β) APWei 1 + Γ(1 + 1/β) APWei   (84) Rician, Nakagami-m and κ−µ Model: Due to space limitation, we derive the approximation error function only for the general κ − µ model; the corresponding error functions for the Rician and Nakagami-m models can be obtained as special cases.11 We use polynomial series expansion for the generalized Marcum Q-function via generalized Laguerre polynomials and write the CDF as follows [21]: −κµ =e ∞ X (µ−1) Ln (κµ) (−1) Γ(µ + n + 1) n=0 n  PR (κ + 1)µ Aκµ n+µ , (85) (α) where Ln (·) is the generalized Laguerre polynomial of degree n and order α. Recognizing that the first term in the above sum gives the power law approximation ˜, we obtain the following:  n+µ ∞ (µ−1) X PR (κµ) −κµ n Ln (86) | − ˜| = e (−1) (κ + 1)µ Γ(µ + n + 1) A κµ n=1  n+µ ∞ X |L(µ−1) PR (κµ)| n −κµ ≤e (κ + 1)µ (87) Γ(µ + n + 1) A κµ n=1  n+µ ∞ − κµ Γ(µ + n) PR e 2 X ≤ (κ + 1)µ (88) Γ(µ) n=1 n!Γ(µ + n + 1) Aκµ µ X  n κµ  ∞ 1 e− 2 PR PR ≤ (κ + 1)µ (κ + 1)µ (89) Γ(µ)µ Aκµ n! A κµ n=1   P R κµ (κ+1)µ A κµ − 1 = ˜e 2 e , (90) 11 Note that they can be derived separately using similar reasoning. 28 which completes the derivation. In (88) we used the following upper bound [21]: |L(α) n (x)| ≤ Γ(α + n + 1) x e2 , n!Γ(α + 1) (91) and in (89) we used: Γ(µ + n) (µ + n − 1)! 1 1 = = ≤ , Γ(µ + n + 1) (µ + n)! µ+n µ (92) for n ≥ 1. A PPENDIX B MRC FOR R ANDOM VARIABLES WITH P OWER -L AW TAILS A general M-branch MRC PDF solution for independent RV can be obtained through a convolution of the branch PDFs f1 ∗ f2 .. ∗ fM , e.g. through the multiplication of moment Q generating functions (MGF) Mm and inverse Laplace transform L−1 . However, approximating the full CDF this way, can result in too complex solutions to readily extract a simple tail approximation. However, it is sufficient to deal with branch tail PDFs only [42]. The lower tail PDF corresponding to (1) can be obtained as: dα PAR f (PR ) ≈ PR →0 dPR β  β 1 βPRβ−1 =α A (93) Using Laplace transform relation [43, ET I 137(1), Table 17.13] F (s) = L(f (t)) = 1/sν ↔ β f (t) = tν−1 /Γ(ν), the branch F (s) ≈ α A1 β Γ(β) . The i.n.i.d M-branch MRC CDF (for any sβ   Q M 1 −1 1 BPR or β combination), is established as FMRC (PR ) = L m=1 Fm (s) , where s is used s to produce the CDF from the inverse transform. Using the same Laplace relation as before, we arrive at  = FMRC (PR ) ≈ −1 L ! 1 s QM m=1 sβm · M Y  αm βm Γ(βm ) m=1 1 Am βm = ˜ (94) which after reordering of terms appears in the form given in (67). R EFERENCES [1] P. Popovski, “Ultra-reliable communication in 5G wireless systems,” in Proceedings of the 1st International Conference on 5G for Ubiquitous Connectivity (5GU), Levi, Finland, 2014, pp. 146–151. [2] P. Popovski, J. J. Nielsen, C. Stefanovic, E. de Carvalho, E. Strom, K. F. Trillingsgaard, A.-S. Bana, D. M. Kim, R. Kotaba, J. Park, and R. B. Sorensen, “Wireless Access for Ultra-Reliable Low-Latency Communication (URLLC): Principles and Building Blocks”, IEEE Network Magazine, accepted, 2017, available on Arxiv. 29 [3] P. Schulz, M. Matthe, H. Klessig, M. Simsek, et al., “Latency critical iot applications in 5g: Perspective on the design of radio interface and network architecture,” IEEE Communications Magazine, vol. 55, no. 2, pp. 70–78, 2017. [4] G. D. Durgin, T. S. Rappaport, and D. A. De Wolf, “New analytical models and probability density functions for fading in wireless communications,” IEEE Transactions on Communications, vol. 50, no. 6, pp. 1005–1015, 2002. [5] M. Rao, F. J. Lopez-Martinez, M.-S. Alouini, and A. Goldsmith, “MGF approach to the analysis of generalized two-ray fading models,” IEEE Transactions on Wireless Communications, vol. 14, no. 5, pp. 2548–2561, 2015. [6] G. Durisi, T. Koch, and P. Popovski, “Toward massive, ultrareliable, and low-latency wireless communication with short packets,” Proceedings of the IEEE, vol. 104, no. 9, pp. 1711–1726, 2016. [7] Digital cellular telecommunications system (Phase 2+); Radio transmission and reception, TS 145 005 V10.8.0, ETSI standard, 2014. [8] J. B. Andersen and I. Z. Kovacs, “Power distributions revisited,” in COST273 3rd Management Committee Meeting, 2002, pp. 17–18. [9] R. Vaughan and J. B. Andersen, Channels, propagation and antennas for mobile communications. IET, 2003, vol. 50. [10] J. Frolik, “On appropriate models for characterizing hyper-rayleigh fading,” IEEE Trans. on Wireless Com., vol. 7, no. 12, 2008. [11] M. Pent, “Orthogonal polynomial approach for the Marcum Q-function numerical computation,” Electronics Letters, vol. 4, no. 25, pp. 563–564, 1968. [12] Á. Baricz, “Bounds for modified Bessel functions of the first and second kinds,” Proceedings of the Edinburgh Mathematical Society (Series 2), vol. 53, no. 03, pp. 575–599, 2010. [13] S. A. Saberali and N. C. Beaulieu, “New expressions for TWDP fading statistics,” IEEE Wireless Communications Letters, vol. 2, no. 6, pp. 643–646, 2013. [14] R. Lorenz, “Theoretical distribution functions of multipath fading processes in the mobile radio and determination of their parameters by measurements,” Technisher Bericht (in German), vol. 455, 1979. [15] H. Hashemi, “The indoor radio propagation channel,” Proceedings of the IEEE, vol. 81, no. 7, pp. 943–968, 1993. [16] A. Bessate and F. El Bouanani, “A very tight approximate results of MRC receivers over independent Weibull fading channels,” Physical Communication, vol. 21, pp. 30–40, 2016. [17] M. Nakagami, “The m-Distribution, a general formula of intensity of rapid fading,” Proceedings of Statistical Methods in Radio Wave Propagation, Pergamon Press, pp. 3–36, June 18–20, 1958 [18] G. J. O. Jameson, “The incomplete gamma functions,” Math. Gazette, vol. 100, no. 548, pp. 298–306, 2016. [19] M. D. Yacoub, “The κ − µ distribution and the η − µ distribution,” IEEE Antennas and Propagation Magazine, vol. 49, no. 1, pp. 68–81, 2007. [20] Á. Baricz, “Tight bounds for the generalized Marcum Q-function,” J. Math. Anal. Appl., vol. 360, no. 1, pp. 265–277, 2009. [21] Sz. András, Á. Baricz and Y. Sun, “The generalized Marcum Q-function: an orthogonal polynomial approach,” Acta Univ. Sapientiae Math., vol.3, no. 1, pp. 60–76, 2011. [22] S. K. Yoo, S. L. Cotton, P. C. Sofotasios and S. Freear, “Shadowed Fading in Indoor Off-Body Communication Channels: A Statistical Characterization Using the κµ / Gamma Composite Fading Model,” in IEEE Trans. on Wireless Com., vol. 15, no. 8, pp. 5231-5244, 2016. [23] S. K. Yoo, S. L. Cotton, P. C. Sofotasios, M. Matthaiou, M. Valkama, and G K. Karagiannidis,“The κ − µ / Inverse Gamma Fading Model,” IEEE 26th Int. Symp. on Pers., Indoor, and Mob. Radio Com. (PIMRC), pp. 425-429, 2015 [24] S. K. Yoo, Fading in Wearable Communications Channels and its Mitigation. PhD [Dissertation]. Belfast, UK: Queens Univ. of Belfast, 2017. Accessed 19-11-2017. [Online]. Available: https://pure.qub.ac.uk/portal/files/130797348/thesis final.pdf. 30 [25] Hypergeometric 1F1, Wolfram Research, Inc.. Assessed 19-11-2017. [Online]. Available http://functions.wolfram.com/PDF/ Hypergeometric1F1.pdf [26] J. F. Paris, “Statistical Characterization of κ−µ Shadowed Fading,” IEEE Trans. on Veh. Tech., vol, 63, no. 2, pp. 518–526, 2014. [27] J. Choi and A. K. Rathie, “Certain Summation Formulas for Humbert´s Double Hypergeometric Series,” Com. Korean Math. Soc., vol. 30, no. 4, pp. 439–446, 2015. [28] C. M. Joshi and J. P. Arya, “Inequalities for certain confluent hypergeometric functions of two variables,” Indian J. pure appl. Math, vol. 13, no. 4, pp. 491–500, 1982. [29] E. Martos-Naya, J. M. Romero-Jerez, F. J. Lopez-Martinez and J. F. Paris. (2016) A MATLAB program for the computation of the confluent hypergeometric function Φ2 . Accessed 3-11-2017. [Online]. Available: https://riuma.uma.es//xmlui/handle/ 10630/12068. [30] M. K. Simon and M.-S. Alouini, Digital communication over fading channels. John Wiley & Sons, 2005, vol. 95. [31] Y. A. Chau and K. Y.-T. Huang, “On the second-order statistics of correlated cascaded rayleigh fading channels,” International Journal of Antennas and Propagation, vol. 2012, 2012. [32] J. N. Pierce and S. Stein, “Multiple diversity with nonindependent fading,” Proceedings of the IRE, vol. 48, no. 1, pp. 89–104, 1960. [33] Maplesoft, Maple User Manual, Maplesoft, Waterloo, ON Canada, 2017, version 18. [Online]. Available: http://www.maplesoft.com [34] R. M. Corless, G. H. Gonnet, D. E. G. Hare, D. J. Jeffrey, and D. E. Knuth, “On the LambertW function,” Advances in Computational Mathematics, vol. 5, no. 1, pp. 329–359, 1996. [35] H. M. Schöpf and P. H. Supancic, “On Bürmann’s theorem and its application to problems of linear and nonlinear heat transfer and diffusion,” The Mathematica Journal, vol. 16, pp. 1–44, 2014. [36] S. Winitzki. (2008) A handy approximation for the error function and its inverse. Assessed: 19-11-2017. [Online]. Available: https://sites.google.com/site/winitzki/sergei-winitzkis-files [37] C. S. Withers and S. Nadarajah, “A Generalized Suzuki Distribution,” IEEE Wireless Personal Communications, vol. 62, pp. 807–830, 2012. [38] V. Plicanic, B. K. Lau, A. Derneryd, and Z. Ying, “Actual diversity performance of a multiband diversity antenna with hand and head effects,” IEEE Transactions on Antennas and Propagation, vol. 57, no. 5, pp. 1547–1556, 2009. [39] B. R. Yanakiev, J. Ødum Nielsen, M. Christensen, and G. Frølund Pedersen, “Correlation measurements on small mobile devices,” in Antennas and Propagation (EUCAP), 2012 6th European Conference on. [40] M. Schwartz, W. R. Bennett, and S. Stein, Communication Systems and Techniques. [41] N. A. Johansson, Y.-P. E. Wang, E. Eriksson and M. IEEE, 2012, pp. 382–385. McGraw-Hill -IEEE Press, 1996. Hessler, “Radio access for ultra-reliable and low-latency 5G communications,” in International Conference on Communication Workshop (ICCW). IEEE, 2015, pp. 1184–1189. [42] P. Hitczenko and S. Montgomery-Smith, “A note on sums of independent random variables,” in Advances in Stochastic Inequalities, Contemp. Math. 234, T. Hill and C. Houdre, Eds. Providence R.I., USA: A.M.S., 1999, pp. 69–73. [43] I. Gradshteyn and I. Ryzhik, Table of Integrals, Series, and Products, 7th ed. USA: Academic Press/ Elsevier, 2007. [44] 3rd Generation Partnership Project (3GPP), “Study on scenarios and requirements for next generation access technologies,”, Technical Report TR 38.913 V14.3.0, retrieved 09 2017.
7cs.IT
1 Exploiting Multiple-Antenna Techniques for Non-Orthogonal Multiple Access Xiaoming Chen, Senior Member, IEEE, Zhaoyang Zhang, Member, IEEE, Caijun Zhong, arXiv:1706.00635v1 [cs.IT] 2 Jun 2017 Senior Member, IEEE, and Derrick Wing Kwan Ng, Member, IEEE Abstract This paper aims to provide a comprehensive solution for the design, analysis, and optimization of a multiple-antenna non-orthogonal multiple access (NOMA) system for multiuser downlink communication with both time duplex division (TDD) and frequency duplex division (FDD) modes. First, we design a new framework for multiple-antenna NOMA, including user clustering, channel state information (CSI) acquisition, superposition coding, transmit beamforming, and successive interference cancellation (SIC). Then, we analyze the performance of the considered system, and derive exact closedform expressions for average transmission rates in terms of transmit power, CSI accuracy, transmission mode, and channel conditions. For further enhancing the system performance, we optimize three key parameters, i.e., transmit power, feedback bits, and transmission mode. Especially, we propose a lowcomplexity joint optimization scheme, so as to fully exploit the potential of multiple-antenna techniques in NOMA. Moreover, through asymptotic analysis, we reveal the impact of system parameters on average transmission rates, and hence present some guidelines on the design of multiple-antenna NOMA. Finally, simulation results validate our theoretical analysis, and show that a substantial performance gain can be obtained over traditional orthogonal multiple access (OMA) technology under practical conditions. Index Terms NOMA, multiple-antenna techniques, performance analysis, power allocation, feedback distribution, mode selection. Xiaoming Chen ([email protected]), Zhaoyang Zhang ([email protected]), and Caijun Zhong ([email protected]) are with the College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou 310027, China. Derrick Wing Kwan Ng ([email protected]) is with the school of Electrical Engineering and Telecommunications, the University of New South Wales, NSW, Australia. 2 I. I NTRODUCTION Current wireless communication systems in general adopt various types of orthogonal multiple access (OMA) technologies for serving multiple users, such as time division multiple access (TDMA), frequency division multiple access (FDMA), and code division multiple access (CDMA), where one resource block is exclusively allocated to one mobile user (MU) to avoid possible multiuser interference. In practice, the OMA technologies are relatively easy to implement, albeit at the cost of low spectral efficiency. Recently, with the rapid development of mobile internet and proliferation of mobile devices, it is expected that future wireless communication systems should be able to support massive connectivity, which is an extremely challenging task for the OMA technologies with limited radio resources. Responding to this, non-orthogonal multiple access (NOMA) has been recently proposed as a promising access technology for the fifth-generation (5G) mobile communication systems, due to its potential in achieving high spectral efficiency [1]–[3]. The principle of NOMA is to exploit the power domain to simultaneously serve multiple MUs utilizing the same radio resources [4], [5], with the aid of sophisticated successive interference cancellation (SIC) receivers [6], [7]. Despite the adoption of SIC, inter-user interference still exists except for the MU with the strongest channel gain, which limits the overall system performance [8]. To address this issue, power allocation has been considered as an effective method to harness multiuser interference [9], [10]. Since the overall performance is limited by the MUs with weak channel conditions, it is intuitive to allocate more power to the weak MUs and less power to the strong MU in order to enhance the effective channel gain and minimize the interference to the weak MUs [11]. For the specific two-user case, the optimal power allocation scheme was studied in [12], and [13] proposed two sub-optimal power allocation schemes exploiting the Karush-Kuhn-Tucker (KKT) conditions, while the issue of quality of service (QoS) requirements of NONA systems was investigated in [14]. For the case with arbitrary number of users, the computational complexity of performing SIC increases substantially and the design of the optimal power allocation becomes intractable. To facilitate an effective system design, clustering and user pairing have been proposed [15], [16]. Generally speaking, multiple MUs with distinctive channel gains are selected to form a cluster, in which SIC is conducted 3 to mitigate the interference [17], [18]. In general, a small cluster consisting a small number of MUs implies low complexity of SIC, but leads to high inter-cluster interference. Thus, it makes sense to dynamically adjust the size of a cluster according to performance requirements and system parameters, so as to achieve a balance between implementation complexity and interference mitigation [19]. However, dynamic clustering is not able to reduce the inter-cluster interference, indicating the necessity of carrying out dynamic clustering in combination with efficient interference mitigation schemes. It is well known that the multiple-antenna technology is a powerful interference mitigation scheme [20]–[22], hence, can be naturally applied to NOMA systems [23], [24]. In [25], the authors proposed a beamforming scheme for combating inter-cluster and intra-cluster interference in a NOMA downlink, where the base station (BS) was equipped with multiple antennas and the MUs have a single antenna each. A more general setup was considered in [26], where both the BS and the MUs are multiple-antenna devices. By exploiting multiple antennas at the BS and the MUs, a signal alignment scheme was proposed to mitigate both the intra-cluster and intercluster interference. It is worth pointing out that the implementation of the two above schemes requires full channel state information (CSI) at the BS, which is usually difficult and costly in practice. To circumvent the difficulty in CSI acquisition, random beamforming was adopted in [27], which inevitably leads to performance loss. Alternatively, the work in [28] suggested to employ zero-forcing (ZF) detection at the multiple-antenna MUs for inter-cluster interference cancelation. However, the ZF scheme requires that the number of antennas at each MU is greater than the number of antennas at the BS, which is in general impractical. To effectively realize the potential benefits of multiple-antenna techniques, the amount and quality of CSI available at the BS plays a key role. In practice, the CSI can be obtained in several different ways. For instance, in time duplex division (TDD) systems, the BS can obtain the downlink CSI through estimating the CSI of uplink by leveraging the channel reciprocity. While in frequency duplex division (FDD) systems, the downlink CSI is usually first estimated and quantized at the MUs, and then is conveyed back to the BS via a feedback link. For both practical TDD and FDD systems, the BS has access to only partial CSI. As a result, there will be residual inter-cluster and intra-cluster interference. To the best of the authors’ knowledge, previous works only consider two extreme cases with full CSI or no CSI, the design, analysis 4 and optimization of multiple-antenna NOMA systems with partial CSI remains an uncharted area. Motivated by this, we present a comprehensive study on the impact of partial CSI on the design, analysis, and optimization of multiple-antenna NOMA downlink communication systems. Specifically, we consider heterogeneous downlink channels, and the BS equipped with arbitrarily multiple antennas has different CSI accuracies about the downlink channels. The major contributions of this paper are summarized as follows: 1) We design a general framework for multiple-antenna NOMA downlink communications including user clustering, CSI estimation, superposition coding, transmit beamforming, and SIC. In particular, the proposed framework is applicable in both TDD and FDD modes. 2) We analyze the performance of the proposed multiple-antenna NOMA, and derive exact expressions for the average transmission rates of each MU in an arbitrary cluster. The average transmission rate is a function of transmit power, CSI accuracy, transmission mode, and channel conditions. 3) We optimize three key parameters of multiple-antenna NOMA, namely, transmit power, feedback bits, and transmission mode. In particular, we present closed-form expressions for the power allocation and feedback distribution. For mode selection, we show that the mode of two MUs in a cluster is optimal in practical cases with moderate and high CSI accuracy, which provides theoretical justification for the two-user setup in the previous works [9]–[14], [23]–[28], [30], [31]. Finally, a low complexity joint optimization scheme of transmit power, feedback bits, and transmission mode is proposed. 4) Through asymptotic analysis of average transmission rates, several key insights are obtained. a) Imperfect CSI results in residual inter-cluster interference at MUs. Thus, there exists a performance gap between practical NOMA with imperfect CSI and ideal NOMA with perfect CSI. The performance gap is an increasing function of transmit power of information signal and a decreasing function of CSI accuracy. In order to maintain a constant performance, transmit energy of pilot sequence for channel estimation in TDD mode and spatial resolution in FDD mode should be increased as transmit 5 power of information signal grows. b) From the perspective of maximizing the sum rate, arranging all MUs in one cluster is optimal if there is no CSI at the BS, while the best option is to arrange one MU in each cluster if there is perfect CSI at the BS. c) In the interference-limited scenario, the average transmission rate for the 1st MU with the strongest channel gain in each cluster increases linearly proportionally to the number of feedback bits. Under the noise-limited condition, the average transmission rate is independent of CSI accuracy. d) In the interference-limited case, equal power allocation among all MUs asymptotically approaches the optimal performance. The rest of this paper is organized as follows: Section II gives a brief introduction of the considered NOMA downlink communication system, and designs the corresponding multipleantenna transmission framework. Section III first analyzes the average transmission rates in presence of imperfect CSI, and then proposes three performance optimization schemes. Section IV derives the average transmission rates in two extreme cases through asymptotic analysis, and presents some system design guidelines. Section V provides simulation results to validate the effectiveness of the proposed schemes. Finally, Section VI concludes the whole paper. Notations: We use bold upper (lower) letters to denote matrices (column vectors), (·)H to denote conjugate transpose, E[·] to denote expectation, k · k to denote the L2 -norm of a vector, d | · | to denote the absolute value, = to denote the equality in distribution, ⌊x⌋ to denote the maximum integer not larger than x, and C to denote the set of complex number. The acronym i.i.d. means “independent and identically distributed”, pdf means “probability density function”, and cdf means “cumulative distribution function”. II. S YSTEM M ODEL AND F RAMEWORK D ESIGN A. User Clustering Consider a downlink communication scenario in a single-cell system, where a base station (BS) broadcasts messages to multiple MUs, cf. Fig. 1. Note that the BS is equipped with M antennas, while the MUs have a single antenna each due to the size limitation. To strike a balance between the system performance and computational complexity in NOMA systems, 6 BS Fig. 1. A multiuser NOMA communication system with 4 clusters. it is necessary to carry out user clustering. In particular, user clustering can be designed from different perspectives. For instance, a signal-to-interference-plus-noise ratio (SINR) maximization user clustering scheme was adopted in [32] and quasi-orthogonal MUs were selected to form a cluster in [33]. Intuitively, these schemes perform user clustering by the exhaustive search method, resulting in high implementation complexity. In this paper, we design a simple user clustering scheme based on the information of spatial direction1 . Specifically, the MUs in the same direction but with distinctive propagation distances are arranged into a cluster. On one hand, the same direction of the MUs in a cluster allows the use of a single beam to nearly align all MUs in such a cluster, thereby facilitating the mitigation of the inter-cluster interference and the enhancement of the effective channel gain. On the other hand, a large gap of propagation distances avoids severe inter-user interference and enables a more accurate SIC at the MUs [34]– [36]. If two MUs are close to each other with almost equal channel gains, it is possible to assign them in different clusters by improving the spatial resolution via increasing the number of spatial beams and the number of BS antennas. Without loss of generality, we assume that the MUs are grouped into N clusters with K MUs in each cluster. To facilitate the following presentation, 1/2 we use αn,k hn,k to denote the M-dimensional channel vector from the BS to the kth MU in the nth cluster, where αn,k is the large-scale channel fading, and hn,k is the small-scale channel fading following zero mean complex Gaussian distribution with unit variance. It is assumed that αn,k remains constant for a relatively long period, while hn,k keeps unchanged in a time slot but 1 The spatial direction of users can be found via various methods/technologies such as GPS or user location tracking algorithms. 7 varies independently over time slots. B. CSI Acquisition For the TDD mode, the BS obtains the downlink CSI through uplink channel estimation. Specifically, at the beginning of each time slot, the MUs simultaneously send pilot sequences of τ symbols to the BS, and the received pilot at the BS can be expressed as YP = N X K q X P αn,k hn,k Φn,k + NP , τ Pn,k (1) n=1 k=1 P where Pn,k is the transmit power for pilot sequence of the kth MU in the nth cluster, NP is an additive white Gaussian noise (AWGN) matrix with i.i.d. zero mean and unit variance complex Gaussian distributed entries. Φn,k ∈ C1×τ is the pilot sequence sent from the kth MU in the nth cluster. It is required that τ > NK, such that the pairwise orthogonality that Φn,k ΦH i,j = 0 and Φn,k ΦH n,k = 1, ∀(n, k) 6= (i, j), can be guaranteed. By making use of the pairwise orthogonality, the received pilot can be transformed as YP Φ H n,k q P αn,k hn,k + NP ΦH = τ Pn,k n,k . (2) Then, by using minimum mean squared error (MMSE) estimation, the relation between the actual channel gain hn,k and the estimated channel gain ĥn,k can be expressed as hn,k = √ ρn,k ĥn,k + p 1 − ρn,k en,k , (3) where en,k is the channel estimation error vector with i.i.d. zero mean and unit variance complex Gaussian distributed entries, and is independent of ĥn,k . Variable ρn,k = 1 P α 1+τ Pn,k n,k P α τ Pn,k n,k P α 1+τ Pn,k n,k = 1− is the correlation coefficient between hn,k and ĥn,k . A large ρn,k means a high accuracy for channel estimation. Thus, it is possible to improve the CSI accuracy by increasing the transmit P power Pn,k or the length τ of pilot sequence. For the FDD mode, the CSI is usually conveyed from the MUs to the BS through a feedback link. Since the feedback link is rate-constrained, CSI at the MUs should first be quantized. Specifically, the kth MU in the nth cluster chooses an optimal codeword from a predetermined (1) Bn,k ) (2 quantization codebook Bn,k = {h̃n,k , . . . , h̃n,k (j) } of size 2Bn,k , where h̃n,k is the jth codeword 8 of unit norm and Bn,k is the number of feedback bits. Mathematically, the codeword selection criterion is given by j ⋆ = arg max (j) 2 Bn,k 1≤j≤2 hH n,k h̃n,k . (4) Then, the MU conveys the index j ⋆ to the BS with Bn,k feedback bits, and the BS recoveries (j ⋆ ) the quantized CSI h̃n,k from the same codebook. In other words, the BS only gets the phase information by using the feedback scheme based on a quantization codebook. However, as shown in below, the phase information is sufficient for the design of spatial beamforming. Similarly, the relation between the real CSI and the obtained CSI in FDD mode can be approximated as [39] h̃n,k = where h̃n,k = hn,k khn,k k √ ⋆ ̺n,k h̃n,k + p 1 − ̺n,k ẽn,k , (5) ⋆ is the phase of the channel hn,k , h̃n,k is the quantized phase information, ẽn,k Bn,k is the quantization error vector with uniform distribution, and ̺n,k = 1 − 2− M −1 is the associated correlation coefficient or CSI accuracy. Thus, it is possible to improve the CSI accuracy by increasing the size of quantization codebook for a given number of antennas M at the BS. C. Superposition Coding and Transmit Beamforming Based on the available CSI, the BS constructs one transmit beam for each cluster, so as to mitigate or even completely cancel the inter-cluster interference. To strike balance between system performance and implementation complexity, we adopt zero-force beamforming (ZFBF) at the BS. We take the deign of beam wi for the ith cluster as an example. First, we construct a complementary matrix H̄i 2 as: H̄i = [ĥ1,1 , · · · , ĥ1,K , · · · , ĥi−1,1 , · · · , ĥi−1,K , ĥi+1,1 , · · · , ĥN,K ]H . (6) Then, we perform singular value decomposition (SVD) on H̄i and obtain its right singular vectors ui,j , j = 1, · · · , Nu , with respect to the zero singular values, where Nu is the number of zero Nu P θi,j ui,j , where θi,j > 0 is a weight singular values. Finally, we can design the beam as wi = j=1 2 ⋆ ⋆ ⋆ ⋆ ⋆ ⋆ In FDD mode, the complementary matrix is given by H̄i = [h̃1,1 , · · · , h̃1,K , · · · , h̃i−1,1 , · · · , h̃i−1,K , h̃i+1,1 , · · · , h̃N,K ]H . 9 such that Nu P θi,j = 1. Thus, the received signal at the kth MU in the nth cluster is given by j=1 yn,k = √ αn,k hH n,k N X wi si + nn,k i=1 = √ αn,k hH n,k wn sn N q X H wi si + nn,k , + αn,k (1 − ρn,k )en,k (7) i=1,i6=n where si = K q P j=1 S S Pi,j si,j is the superposition coded signal with Pi,j and si,j being transmit power and transmit signal for the jth MU in the ith cluster, and nn,k is the AWGN with unit S variance. In general, Pi,j should be carefully allocated to distinguish the MUs in the power domain, which we will discuss in detail below. Note that Eq. (7) holds true due to the fact that p p H √ 3 hH ρn,k ĥn,k wi + 1 − ρn,k eH 1 − ρn,k eH n,k wi = n,k wi = n,k wi for ZFBF in TDD mode . With perfect CSI at the BS, i.e., ρn,k = 1, the inter-cluster interference can be completely cancelled. D. Successive Interference Cancellation Although ZFBF at the BS can mitigate partial inter-cluster interference from the other clusters, there still exists intra-cluster interference from the same cluster. In order to improve the received signal quality, the MU conducts SIC according to the principle of NOMA. Without loss of generality, we assume that the effective channel gains in the ith cluster have the following order: √ √ 2 H 2 | αi,1 hH i,1 wi | ≥ · · · ≥ | αi,K hi,K wi | . (8) It is reasonably assumed that the BS may know MUs’ effective gains through the channel quality indicator (CQI) messages, and then determines the user order in (8). Thus, in the ith cluster, the jth MU can always successively decode the lth MU’s signal, ∀l > j, if the lth MU can decode its own signal. As a result, the jth MU can subtract the interference from the lth MU in the received signal before decoding its own signal. After SIC, the signal-to-interference-plus-noise 3 √ In FDD mode, we have hH n,k wi = ̺n,k eH n,k wi , d √ ⋆ ̺n,k khn,k k(h̃n,k )H wi + √ 1 − ̺n,k khn,k kẽH n,k wi = √ d 1 − ̺n,k khn,k kẽH n,k wi = 1− where = denotes the equality in distribution. If ̺n,k = ρn,k , Eq. (7) also holds true in FDD mode. In the sequel, without loss of generality, we no longer distinguish between TDD and FDD. 10 ratio (SINR) at the kth MU in the nth cluster is given by γn,k = 2 αn,k |hH n,k wn | | {z k−1 X 2 S αn,k |hH n,k wn | Pn,k S Pn,j j=1 Intra-cluster interference } + αn,k (1 − ρn,k ) | N X i=1,i6=n 2 |eH n,k wi | {z K X , S Pi,l l=1 Inter-cluster interference (9) + |{z} 1 AWGN } where the first term in the denominator of (9) is the residual intra-cluster interference after SIC at the MU, the second one is the residual inter-cluster interference after ZFBF at the BS, and the third one is the AWGN. For the 1st MU in each cluster, there is no intra-cluster interference, since it can completely eliminate the intra-cluster interference. Note that in this paper, we assume that perfect SIC can be performed at the MUs. In practical NOMA systems, SIC might be imperfect due to a limited computational capability at the MUs. Thus, there exists residual intra-cluster interference from the weaker MUs even after SIC [37]. However, the study of the impact of imperfect SIC on the system performance is beyond the scope of this paper and we would like to investigate it in the future work. Moreover, the transmit power has a significant impact on the SIC and the performance of NOMA [38]. Thus, we will quantitatively analyze the impact of transmit power and then aim to optimize the transmit power for improving the performance in the following sections. III. P ERFORMANCE A NALYSIS AND O PTIMIZATION In this section, we concentrate on performance analysis and optimization of multi-antenna NOMA downlink with imperfect CSI. Specifically, we first derive closed-form expressions for the average transmission rates of the 1st MU and the other MUs, and then propose separate and joint optimization schemes of transmit power, feedback bits, and transmit mode, so as to maximize the average sum rate of the system. A. Average Transmission Rate We start by analyzing the average transmission rate of the kth MU in the nth cluster. First, we consider the case k > 1. According to the definition, the corresponding average transmission rate can be computed as Rn,k = E [log2 (1 + γn,k )] 11   H 2  αn,k |hn,k wn |    = E log2    " = E log2 −E log2 S Pn,j j=1 + αn,k (1 − ρn,k ) N P 2 |eH n,k wi | K P  S Pi,l + 1    K k−1 N P P P  H S S H 2 2 Pi,l + 1 αn,k |hn,k wn | Pn,j + αn,k (1 − ρn,k ) |en,k wi | 2 αn,k |hH n,k wn | " k P j=1 k X S Pn,j + αn,k (1 − ρn,k ) k−1 X S Pn,j j=1 N X i=1,i6=n + αn,k (1 − ρn,k ) l=1 l=1 i=1,i6=n j=1 2 αn,k |hH n,k wn | i=1,i6=n 2 |eH n,k wi | N X K X +1 l=1 2 |eH n,k wi | i=1,i6=n S Pi,l K X l=1 S Pi,l !# +1 !# . (10) Note that the average transmission rate in (10) can be expressed as the difference of two terms, which have a similar form. Hence, we concentrate on the derivation of the first term. k P 2 S For notational convenience, we use W to denote the term αn,k |hH Pn,j + αn,k (1 − n,k wn | j=1 ρn,k ) N P i=1,i6=n |eH n,k wi | K P 2 l=1 S Pi,l . To compute the first expectation, the key is to obtain the probability 2 density function (pdf) of W . Checking the first random variable |hH n,k wn | in W , since wn of 2 2 unit norm is designed independent of hn,k , |hH n,k wn | is χ distributed with 2 degrees of freedom 2 2 [40]. Similarly, |eH n,k wi | also has the distribution χ (2). Therefore, W can be considered as a weighted sum of N random variables with χ2 (2) distribution. According to [41], W is a nested finite weighted sum of N Erlang pdfs, whose pdf is given by fW (x) = N X i=1 where q ηn,k =  q i ΞN i, {ηn,k }N q=1 g(x, ηn,k ),  k P  S  α Pq,j  n,k K P  S   αn,k (1 − ρn,k ) Pq,l if q 6= n i g(x, ηn,k )= ΞN  q i, {ηn,k }N q=1 if q = n j=1 1 i ηn,k , l=1 exp − N −1 i Y (−1)N −1 ηn,k = N Q l s=1 ηn,k l=1 (11) x i ηn,k 1 i ηn,k ! , − 1 s+U(s−i) ηn,k !−1 , 12 and U(x) is the well-known unit step function defined as U(x ≥ 0) = 1 and zero otherwise. q It is worth pointing out that the weights ΞN are constant for given {ηn,k }N q=1 . Hence, the first expectation in (10) can be computed as Z ∞ E[log2 (1 + W )] = log2 (1 + x)fW (x)dx = 0 N X i=1 where Ei (x) = Rx q ΞN i, {ηn,k }N q=1 N  Z ∞ 0 log2 (1 + x)  1 X q ΞN i, {ηn,k }N = − q=1 exp ln(2) i=1 1 i ηn,k 1 i ηn,k ! exp − Ei − x ! i ηn,k ! 1 i ηn,k , dx (12) exp(t) dt t is the exponential integral function. Eq. (12) follows from [42, Eq. N K k−1 P P P S 2 S 2 |eH Pi,t (4.3372)]. Similarly, we use V to denote αn,k |hH Pn,j +αn,k (1−ρn,k ) n,k wn | n,k wi | −∞ j=1 i=1,i6=n in the second term of (10). Thus, the second expectation term can be computed as ! ! N  1 1 1 X v ΞN i, {βn,k }N Ei − i , E[log2 (1 + V )] = − v=1 exp i ln(2) i=1 βn,k βn,k where v βn,k =  k−1 P S   Pv,j α  n,k t=1 (13) if v = n j=1 K P   S  αn,k (1 − ρn,k ) Pv,l if v 6= n . l=1 Hence, we can obtain the average transmission rate for the kth MU in the nth cluster as follows ! ! N  1 1 1 X v Ei − i ΞN i, {βn,k }N Rn,k = v=1 exp i ln(2) i=1 βn,k βn,k ! ! N X  1 1 1 q ΞN i, {ηn,k }N Ei − i . (14) − q=1 exp i ln(2) i=1 ηn,k ηn,k Then, we consider the case k = 1. Since the first MU can decode all the other MUs’ signals in the same cluster, there is no intra-cluster interference. In this case, the corresponding average transmission rate reduces to Rn,1     N −1  1 1 X 1 v N −1 Ei − i = ΞN −1 i, {βn,1 }v=1 exp i ln(2) i=1 βn,1 βn,1 13 N  1 X q ΞN i, {ηn,1 }N − q=1 exp ln(2) i=1 where q = ηn,1 and v βn,1 =  S   αn,1 Pq,1  1 i ηn,1    1 Ei − i , ηn,1 (15) if q = n   αn,1 (1 − ρn,1 ) K P l=1 S Pq,l if q 6= n  K P  S   αn,1 (1 − ρn,1 ) Pv,l    αn,1 (1 − ρn,1 ) l=1 K P l=1 S Pv+1,l , if v < n . if v ≥ n Combing (14) and (15), it is easy to evaluate the performance of a multiple-antenna NOMA downlink with arbitrary system parameters and channel conditions. In particular, it is possible to reveal the impact of system parameters, i.e., transmit power, CSI accuracy, and transmission mode. B. Power Allocation From (14) and (15), it is easy to observe that with imperfect CSI, transmit power has a great impact on average transmission rates. On one hand, increasing the transmit power can enhance the desired signal strength. On the other hand, it also increases the interference. Thus, it is desired to distribute the transmit power according to channel conditions. To maximize the sum rate of the considered multiple-antenna NOMA system subject to a total power constraint, we have the following optimization problem: J1 : max S Pn,k s.t. C1 : C2 : N X K X Rn,k (16) n=1 k=1 N X K X S S Pn,k ≤ Ptot n=1 k=1 S Pn,k > 0, S where Ptot is the maximum total transmit power budget. It is worth pointing out that in certain scenarios, user fairness might be of particular importance. To guarantee user fairness, one can replace the objective function of J1 with the maximization of a weighted sum rate, where the 14 weights can directly affect the power allocation and thus the MUs’ rates. Unfortunately, J1 is not a convex problem due to the complicated expression for the objective function. Thus, it is difficult to directly provide a closed-form solution for the optimal transmit power. As a compromise solution, we propose an effective power allocation scheme based on the following important observation of the multiple-antenna NOMA downlink system: Lemma 1: The inter-cluster interference is dependent of power allocation between the clusters, while the intra-cluster interference is determined by power allocation among the MUs in the same cluster. Proof: A close observation of the inter-cluster interference αn,k (1 − ρn,k ) K P l=1 S Pi,l in (9) indicates that K P l=1 N P i=1,i6=n 2 |eH n,k wi | S Pi,l is the total transmit power for the ith cluster, which suggests that inter-cluster power allocation does not affect the inter-cluster interference. Inspired by Lemma 1, the power allocation scheme can be divided into two steps. In the first step, the BS distributes the total power among the N clusters. In the second step, each cluster individually carries out power allocation subject to the power constraint determined by the first step. In the following, we give the details of the two-step power allocation scheme. First, we design the power allocation between the clusters from the perspective of minimizing inter-cluster interference. For the ith cluster, the average aggregate interference to the other clusters is given by Ii = E " N K X X n=1,n6=i k=1 = N K X X n=1,n6=i k=1 where PiS = K P l=1 2 αn,k (1 − ρn,k )|eH n,k wi | ! K X S Pi,l l=1 αn,k (1 − ρn,k ) PiS , # (17) S Pi,l is the total transmit power of the ith cluster. Eq. (17) follows the fact that 2 E[|eH n,k wi | ] = 1. Intuitively, a large interference coefficient N P K P n=1,n6=i k=1 αn,k (1 − ρn,k ) means a more severe inter-cluster interference caused by the ith cluster. In order to mitigate the intercluster interference for improving the average sum rate, we propose to distribute the power proportionally to the reciprocal of interference coefficient. Specifically, the transmit power for 15 the ith cluster can be computed as PiS = N P K P n=1,n6=i k=1 N P l=1 N P αn,k (1 − ρn,k ) K P n=1,n6=l k=1 !−1 αn,k (1 − ρn,k ) S !−1 Ptol . (18) Then, we allocate the power in the cluster for further increasing the average sum rate. According to the nature of NOMA techniques, the first MU not only has the strongest effective channel gain for the desired signal, but also generates a weak interference to the other MUs. On the contrary, the Kth MU has the weakest effective channel gain for the desired signal, and also produces a strong interference to the other MUs. Thus, from the perspective of maximizing the sum of average rate, it is better to allocate the power based on the following criterion: S S S Pn,1 ≥ · · · ≥ Pn,k ≥ · · · ≥ Pn,K . (19) On the other hand, in order to facilitate SIC, the NOMA in general requires the transmit powers in a cluster to follow a criterion below [28]: S S S Pn,1 ≤ · · · ≤ Pn,k ≤ · · · ≤ Pn,K . (20) Under this condition, the MU performs SIC according to the descending order of the user index, namely the ascending order of the effective channel gain. Specifically, the kth MU cancels the interference from the Kth to the (k + 1)th MU in sequence. Thus, the SINR for decoding each interference signal is the highest, which facilitates SIC at MUs [38]. To simultaneously fulfill the above two criterions, we propose to equally distribute the powers within a cluster, namely S Pn,k = PnS /K. (21) Substituting (18) into (21), the transmit power for the kth MU in the nth cluster can be computed 16 as S Pn,k = N P K P i=1,i6=n j=1  K N P l=1 αi,j (1 − ρi,j ) !−1 S !−1  Ptol . N K P P  αi,j (1 − ρi,j ) (22) i=1,i6=l j=1 Thus, we can distribute the transmit power based on (22) for given channel statistical information and the CSI accuracy, which has a quite low computational complexity. Remarks: We note that path loss coefficient αn,k , ∀n, k, remain constant for a relatively long time, and it is easy to obtain at the BS via long-term measurement. Hence, the proposed power allocation scheme incurs a low system overhead, and can be implemented with low complexity. C. Feedback Distribution For the FDD mode, the accuracy of quantized CSI relies on the size of codebook 2Bn,k , where Bn,k is the number of feedback bits from the kth MU in the nth cluster. As observed in (14) and (15), it is possible to decrease the interference by increasing feedback bits. However, due to the rate constraint on the feedback link, the total number of feedback bits is limited. Therefore, it is of great importance to optimize the feedback bits among the MUs for performance enhancement. According to the received SNR in (9), the CSI accuracy only affects the inter-cluster interference. Thus, it makes sense to optimize the feedback bits to minimizing the average sum of inter-cluster interference given by # " N K K N X XX X S 2 Iinter = E Pi,l αn,k (1 − ̺n,k ) |eH n,k wi | n=1 k=1 = N X K X n=1 k=1 αn,k i=1,i6=n N X i=1,i6=n Bn,k PiS 2− M −1 . l=1 (23) 17 Hence, the optimization problem for feedback bits distribution can be expressed as J2 : min Bn,k N X K X s.t. C3 : αn,k n=1 k=1 N X K X n=1 k=1 N X Bn,k PiS 2− M −1 (24) i=1,i6=n Bn,k ≤ Btot , C4 : Bn,k ≥ 0, where Btot is an upper bound on the total number of feedback bits. J2 is an integer programming problem, hence is difficult to solve. To tackle this challenge, we relax the integer constraint on Bn,k . In this case, according to the fact that N X K X αn,k n=1 k=1 N X PiS 2 B − Mn,k −1 i=1,i6=n N Y K Y ≥ NK αn,k n=1 k=1  N P = NK 2  = NK 2 where the equality holds true only when αn,k K P Bn,k 1  NK N P PiS 2 B − Mn,k −1 i=1,i6=n k=1 − n=1 M −1 B − Mtot −1 N X 1  NK N Y K Y  1 ! NK αn,k n=1 k=1 N Y K Y n=1 k=1 PiS i=1,i6=n N X αn,k N X PiS i=1,i6=n 1 ! NK 1 ! NK , (25) Bn,k i=1,i6=n PiS 2− M −1 , ∀n, k are equal. In other words, the objective function in (24) can be minimized while satisfying the following condition: αn,k N X i=1,i6=n PiS 2 B − Mn,k −1  = 2 B − Mtot −1 1  NK N Y K Y αn,k n=1 k=1 N X PiS i=1,i6=n 1 ! NK . (26) Hence, based on the relaxed optimization problem, the optimal number of feedback bits for the kth MU in the nth cluster is given by Bn,k N K Btot 1 XX = log2 − NK NK i=1 j=1 αi,j N X l=1,l6=i PlS ! + log2 αn,k N X l=1,l6=n PlS ! . (27) Given channel statistical information and transmit power allocation, it is easy to determine the feedback distribution according to (27). Note that there exists an integer constraint on the number of feedback bits in practice, so we should utilize the maximum integer that is not larger than 18 Bn,k in (27), i.e., ⌊Bn,k ⌋, ∀n, k. Remarks: The number of feedback bits distributed to the kth MU in the nth cluster is determined by the average inter-cluster interference generated by the kth MU in the nth cluster with respect to the average inter-cluster interference of each MU. In other words, if one MU generates more inter-cluster interference, it would be allocated with more feedback bits, so as to facilitate a more accurate ZFBF to minimize the total interference. D. Mode Selection As discussed above, the performance of the multiple-antenna NOMA system is limited by both inter-cluster and intra-cluster interference. Although ZFBF at the BS and SIC at the MUs are jointly applied, there still exists residual interference. Intuitively, the strength of the residual interference mainly relies on the number of clusters N and the number of MUs in each cluster K. For instance, increasing the number of MUs in each cluster might reduce the inter-cluster interference, but also results in an increase in intra-cluster interference. Thus, it is desired to dynamically adjust the transmission mode, including the number of clusters and the number of MUs in each cluster, according to channel conditions and system parameters. For dynamic mode selection, we have the following lemma: Lemma 2: If the BS has no CSI about the downlink, it is optimal to set N = 1. On the other hand, if the BS has perfect CSI about the downlink, K = 1 is the best choice. Proof: First, if there is no CSI, namely ρn,k = 0, ∀n, k, ZFBF cannot be utilized to mitigate the inter-cluster interference. If all the MUs belong to one cluster, interference can be mitigated as much as possible by SIC. In the case of perfect CSI at the BS, ZFBF can completely the interference. Thus, it is optimal to arrange one MU in one cluster. In above, we consider two extreme scenarios of no and perfect CSI at the BS, respectively. In practice, the BS has partial CSI through channel estimation or quantization feedback. Thus, we propose to dynamically choose the transmission mode for maximizing the sum of average 19 transmission rate, which is equivalent to an optimization problem below: J3 : max N,K N X K X Rn,k (28) n=1 k=1 s.t. C5 : NK = Nu , C6 : N > 0, C7 : K > 0, where Nu is the number of MUs in the multiple-antenna NOMA system. J3 is also an integer programming problem, so it is difficult to obtain the closed-form solution. Under this condition, it is feasible to get the optimal solution by numerical search and the search complexity is O(N K ). In order to control the complexity of SIC, the number of MUs in one cluster is usually small, e.g., K = 2. Therefore, the complexity of numerical search is acceptable. E. Joint Optimization Scheme In fact, transmit power, feedback bits and transmission mode are coupled, and determine the performance together. Therefore, it is better to jointly optimize these variables, so as to further improve the performance of the multiple-antenna NOMA systems. For example, given a transmission mode, it is easy to first allocate transmit power according to (22), and then distribute feedback bits according to (27). Finally, we can select an optimal transmission mode with the largest sum rate. The complexity of the joint optimization is mainly determined by the mode selection. As mentioned above, if the number of MUs in one cluster is small, the complex of mode selection is acceptable. IV. A SYMPTOTIC A NALYSIS In order to provide insightful guidelines for system design, we now pursue an asymptotic analysis on the average sum rate of the system. In particular, two extreme cases are studied, namely, interference limited and noise limited. 20 A. Interference Limited Case S S With loss of generality, we let Pn,k = θn,k Ptot , ∀n, k, where 0 < θn,k < 1 is a power N P K P !−1 αv,j (1−ρv,j ) v=1,v6=n j=1 allocation factor. For instance, θn,k is equal to  K N P l=1 N P K P v=1,v6=l j=1 !−1   αv,j (1−ρv,j ) in the proposed S power allocation scheme in Section III.B. If the total power Ptot is large enough, the noise term of SINR in (9) is negligible. In this case, with the help of [42, Eq. (4.3311)], the average transmission rate of the kth MU (k > 1) in the nth cluster reduces to N Rn,k = N   1 X 1 X q v i i ΞN i, {ηn,k ΞN i, {βn,k }N }N q=1 ln(ηn,k ) − v=1 ln(βn,k ), ln(2) i=1 ln(2) i=1 (29) where we have also used the fact that N X ΞN i=1  q i, {ηn,k }N q=1 = N X i=1  v ΞN i, {βn,k }N v=1 = 1. (30) Similarly, the asymptotic average transmission rate of the 1st MU in the nth MU can be obtained as Rn,1 = N N −1     1 X 1 X q i v N −1 i ΞN i, {ηn,1 }N ln η − ΞN −1 i, {βn,1 }v=1 ln βn,1 . q=1 n,1 ln(2) i=1 ln(2) i=1 (31) Combining (29) and (31), we have the following important result: Theorem 1: In the region of high transmit power, the average transmission rate is independent S S S of Ptot , and there exists a performance ceiling regardless of Ptot , i.e., once Ptot is larger than a saturation point, the average transmission rate will not increase further even the transmit power increases. i i i i S Proof: According to the definitions, ηn,k and βn,k can be rewritten as ηn,k = ωn,k Ptot and i i S βn,k = ψn,k Ptot , where i ωn,k =  k P   α θi,j  n,k if i = n j=1 K P    αn,k (1 − ρn,k ) θi,l if i 6= n l=1 , 21 and i ψn,k =  k−1 P   θi,j  αn,k if i = n j=1 K P    αn,k (1 − ρn,k ) θi,l if i 6= n , l=1   q v N S respectively. Thus, ΞN i, {ηn,k }N and Ξ i, {β } N q=1 n,k v=1 are independent of Ptot . Hence, Rn,k in (29) can be transformed as N Rn,k  1 X q S i ΞN i, {ηn,k }N = q=1 (ln(Ptot ) + ln(ωn,k )) ln(2) i=1 N  1 X v S i − ΞN i, {βn,k }N v=1 (ln(Ptot ) + ln(ψn,k )) ln(2) i=1 N N   1 X 1 X q i v i ΞN i, {ηn,k }N ln(ω ) − ΞN i, {βn,k }N = q=1 n,k v=1 ln(ψn,k ), (32) ln(2) i=1 ln(2) i=1 where Eq. (32) follows the fact that N P i=1 we can rewrite Rn,1 in (31) as Rn,1 N  P  q v ΞN i, {ηn,k }N = ΞN i, {βn,k }N q=1 v=1 = 1. Similarly, i=1 N N −1     1 X 1 X q N i v N −1 i ΞN i, {ηn,1 }q=1 ln ωn,1 − ΞN −1 i, {βn,1 }v=1 ln ψn,1 , (33) = ln(2) i=1 ln(2) i=1 where i ωn,1 = and i ψn,1 =  S   αn,1 θi,1 if i = n   αn,1 (1 − ρn,1 ) K P l=1 S θi,l if i 6= n  K P  S   αn,1 (1 − ρn,1 ) θi,l    αn,1 (1 − ρn,1 ) l=1 K P l=1 S θi+1,l , if i < n . if i ≥ n S Note that both (32) and (33) are regardless of Ptot , which proves Theorem 1. Now, we investigate the relation between the performance ceiling in Theorem 1 and the CSI accuracy ρn,k . First, we consider Rn,k with k > 1. As ρn,k asymptotically approaches 1, the 22 inter-cluster interference is negligible. Then, Rn,k can be further reduced as !# !# " " k−1 k X X S 2 S 2 ideal Pn,j Pn,j − E log2 αn,k |hH Rn,k = E log2 αn,k |hH n,k wn | n,k wn | j=1 j=1  k P  ωn,j    j=1  = log2  k−1 . P  ψn,j (34) j=1 It is found that even with perfect CSI, the average transmission rate for the (k > 1)th MU is P  k ωn,j j=1 still upper bounded. The bound log2  k−1 P ψn,j j=1  is completely determined by channel conditions, and thus cannot be increased via power allocation. Differently, for the 1st MU, if the CSI at the BS is sufficiently accurate, the SINR γn,1 becomes high. As a result, the constant term 1 in the rate expression is negligible, and thus the average transmission rate can be approximated as        Rn,1 ≈E log2     2 S αn,1 |hH  n,1 wn | Pn,1  N K P P S  H 2 αn,1 (1 − ρn,1 ) |en,1 wi | Pi,l i=1,i6=n l=1 "   2 S − E log2 =E log2 αn,1 |hH n,1 wn | Pn,1 {z } | Ideal average rate | αn,1 (1 − ρn,1 ) N X i=1,i6=n {z 2 |eH n,1 wi | Rate loss due to imperfect CSI K X l=1 S Pi,l !# .(35) } In (35), the first term is the ideal average transmission rate with perfect CSI, and the second one is rate loss caused by imperfect CSI. We first check the term of the ideal average transmission rate, which is given by   S 2 ideal θn,1 |hH w | Rn,1 = E log2 αn,1 Ptot n n,1  C S θn,1 − = log2 αn,1 Ptot . ln(2) (36) Note that if there is perfect CSI at the BS, the average transmission rate of the 1st MU increases S proportionally to log2 (Ptot ) without a bound. However, as seen in (34), the (k > 1)th MU has an upper bounded rate under the same condition, which reconfirms the claim in Lemma 2 that it is optimal to arrange one MU in each cluster in presence of perfect CSI. Then, we investigate 23 the rate loss due to imperfect CSI, which can be expressed as loss Rn,1 = log2 αn,1 (1 − S ρn,1 )Ptot where µvn,1 =  N −1   1 X N −1 − ΞN −1 i, {µvn,1 }v=1 C − ln µin,1 , (37) ln(2) i=1  K P   θv,l     l=1 K P l=1 if v < n . θv+1,l if v ≥ n loss S Given a ρn,1 , the rate loss Rn,1 enlarges as the total transmit power Ptot increases. In order ideal to keep the same rate of increase to the ideal rate Rn,1 , the CSI accuracy ρn,1 should satisfy the following theorem: S Theorem 2: Only when (1 − ρn,1 )Ptot is equal to a constant ε, the average transmission rate of the 1st MU in the nth cluster with imperfect CSI remains a fixed gap with respect to the p ideal rate. Specifically, the transmit power for training sequence should satisfy Pn,1 = S /ε−1 Ptot αn,1 τ S in TDD systems, while the number of feedback bits should satisfy Bn,1 = (M − 1) log2 (Ptot /ε) in FDD systems. Proof: The proof is intuitively. By substituting ρn,1 = 1 − 1+τ P P1 Bn,1 n,1 αn,1 S into (1 − ρn,1 )Ptot =ε S for TDD systems and ̺n,1 = 1 − 2− M −1 into (1 − ̺n,1 )Ptot = ε for FDD systems, we can get p Pn,1 = S /ε−1 Ptot αn,1 τ S and Bn,1 = (M − 1) log2 (Ptot /ε), which proves Theorem 2. p Remarks: For the CSI accuracy at the BS, Pn,1 τ (namely transmit energy for training sequence) in TDD systems and Bn,1 M −1 (namely spatial resolution) in FDD systems are two crucial factors. Specifically, given a requirement on CSI accuracy, it is possible to shorten the length of training sequence by increasing the transmit power, so as to leave more time for data transmission in a time slot. However, in order to keep the pairwise orthogonality of training sequences, the length of training sequence τ must be larger than the number of MUs. In other words, the minimum value of τ is NK. Similarly, in FDD systems, it is possible to reduce the feedback bits by increasing the number of antennas M. Yet, in order to fulfill the spatial degrees of freedom for ZFBF at the BS, M must be not smaller than (N − 1)K + 1. This is because the beam wi for the ith cluster should be in the null space of the channels for the (N − 1)K MUs in the other N − 1 clusters. 24 Furthermore, substituting (36) and (37) into (35), we have Rn,1 ≈ − log2 (1 − ρn,1 ) + log2 (θn,1 ) − N −1 X i=1   N −1 ΞN −1 i, {µvn,1}v=1 log2 µin,1 . (38) Given a power allocation scheme, it is interesting that the bound of Rn,1 is independent of channel conditions. As analyzed above, it is possible to improve the average rate by improving the CSI accuracy. Especially, for FDD systems, we have the following lemma: Lemma 3: At the high power region with a large number of feedback bits, the average rate of the 1st MU increases linearly as the numbers of feedback bits increase. Bn,1 Proof: Replacing ρn,1 in (38) with ̺n,1 = 1 − 2− M −1 , Rn,1 is transformed as N −1 Rn,1 X   Bn,1 N −1 + log2 (θn,1 ) − ΞN −1 i, {µvn,1 }v=1 log2 µin,1 , ≈ M −1 i=1 (39) which yields Lemma 3. B. Noise Limited Case If the interference term is negligible with respect to the noise term due to a low transmit power, then the SINR γn,k , ∀n, k is reduced as 2 S γn,k = αn,k |hH n,k wn | Pn,k , (40) 2 2 which is equivalent to the interference-free case. As discussed earlier, |hH n,k wn | is χ (2) dis- tributed, then the average transmission rate can be computed as Z ∞  S Rn,k = αn,k x exp(−x)dx log2 1 + Pn,k 0 ! ! 1 1 Ei − S . = − exp S αn,k Pn,k Pn,k αn,k (41) Note that Eq. (41) is independent of the CSI accuracy, thus it is unnecessary to carry out channel estimation or CSI feedback in this scenario. Since both intra-cluster interference and inter-cluster interference are negligible, ZFBF at the BS and SIC at the MUs are not required, and all optimization schemes asymptotically approach the same performance. 25 TABLE I PARAMETER TABLE FOR (αn,k , ρn,k ), ∀n ∈ [1, 3] AND k ∈ [1, 2]. ❍❍ k n ❍❍ 1 2 1 2 3 (1.00, 0.90) (0.95, 0.85) (0.90, 0.80) (0.10, 0.70) (0.20, 0.75) (0.15, 0.80) V. S IMULATION R ESULTS To evaluate the performance of the proposed multiple-antenna NOMA technology, we present several simulation results under different scenarios. For convenience, we set M = 6, N = 3, K = 2, Btot = 12, while αn,k and ρn,k are given in Tab. I for all simulation scenarios without S extra specification. In addition, we use SNR (in dB) to represent 10 log10 Ptot . 2 Theoretical (R1,1) 1.8 Simulation (R ) 1,1 1.6 Theoretical (R ) Simulation (R ) 1,2 Average Rate (b/s/Hz) 1.4 1,2 1.2 1 0.8 0.6 0.4 0.2 0 −10 Fig. 2. −5 0 5 10 15 SNR (dB) 20 25 30 35 Comparison of theoretical expressions and simulation results. First, we verify the accuracy of the derived theoretical expressions. As seen in Fig. 2, the theoretical expressions for both the 1st and the 2nd MUs in the 1st cluster well coincide with the simulation results in the whole SNR region, which confirms the high accuracy. As the principle of NOMA implies, the 1st MU performs better than the second MU. At high SNR, the average rates of the both MUs are asymptotically saturated, which proves Theorem 1 again. Secondly, we compare the proposed power allocation scheme with the equal power allocation scheme and the fixed power allocation scheme proposed in [4]. Note that the fixed power allocation scheme distributes the power with a fixed ratio 1:4 between the two MUs in a cluster 26 6 Sum of Average Rate (b/s/Hz) 5 4 3 2 Adapive Power Allocation Equal Power Allocation Fixed Power Allocation 1 0 -10 -5 0 5 10 15 20 25 30 35 SNR (dB) Fig. 3. Performance comparison of different power allocation schemes. so as to facilitate the SIC. It is found in Fig. 3 that the proposed power allocation scheme offers an obvious performance gain over the two baseline schemes, especially in the medium SNR region. Note that practical communication systems in general operate at medium SNR, thus the proposed scheme is able to achieve a given performance requirement with a lower SNR. As the SNR increases, the proposed scheme and the equal allocation scheme achieve the same saturated sum rate, but the fixed allocation scheme has a clear performance loss. Next, we examine the advantage of feedback allocation for the FDD based NOMA system with equal power allocation, cf. Fig. 4. As analyzed in Section IV.B, at very low SNR, namely the noise-limited case, the average rate is independent of CSI accuracy, and thus the two schemes asymptotically approach the same sum rate. As SNR increases, the proposed feedback allocation scheme achieves a larger performance gain. Similarly, at high SNR, both the two schemes are saturated, and the proposed scheme obtains the largest performance gain. For instance, at SNR= 30 dB, there is a gain of more than 0.5 b/s/Hz. Furthermore, we investigate the impact of the total number of feedback bits on the average rates of different MUs at SNR= 35 dB. As shown in Fig. 5, the performance of the 1st MU is clearly better than that of the 2nd MU. Moreover, the average rate of the 1st MU is nearly a linear function of the number of feedback bits, which reconfirms the claims of Lemma 3. Then, we investigate the impact of transmission mode on the performance of the NOMA systems at SNR= 10 dB with equal power allocation in Fig. 6. To concentrate on the impact of 27 3 Sum of Average Rate (b/s/Hz) 2.5 2 1.5 1 Adaptive Feedback Allocation Equal Feedback Allocation 0.5 0 −10 Fig. 4. −5 0 5 10 15 SNR (dB) 20 25 30 35 Performance comparison of different feedback allocation schemes. 1.4 1.2 Average Rate (b/s/Hz) 1 0.8 ideal R1,2 0.6 R 1,1 R1,2 0.4 0.2 0 Fig. 5. 0 5 10 15 20 25 Total Number of Feedback (bits) 30 35 Asymptotic performance with a large number of feedback bits. transmission mode, we set the same CSI accuracy of all downlink channels as ρ. Note that we consider four fixed transmission modes under the same channel conditions in the case of 6 MUs in total. Consistent with the claims in Lemma 2, mode 4 with N = 1 and K = 6 achieves the largest sum rate at low CSI accuracy, while mode 1 with N = 6 and K = 1 performs best at high CSI accuracy. In addition, it is found that at medium CSI accuracy, mode 2 with N = 3 and K = 2 is optimal, since it is capable to achieve a best balance between intra-cluster interference and inter-cluster interference. Thus, we propose to dynamically select the transmission mode according to channel conditions and system parameters. As shown by the red line in Fig. 6, dynamic mode selection can always obtain the maximum sum rate. Finally, we exhibit the superiority of the proposed joint optimization scheme for the NOMA 28 7 Mode 1 (N=6, K=1) Mode 2 (N=3, K=2) Mode 3 (N=2, K=3) Mode 4 (N=1, K=6) Dynamic Mode Sum of Average Rate (b/s/Hz) 6 5 4 3 2 1 Fig. 6. 0 0.1 0.2 0.3 0.4 0.5 ρ 0.6 0.7 0.8 0.9 1 Performance comparison of different transmission modes. 9 Sum of Average Rate (b/s/Hz) 8 NOMA (Joint Optimization) NOMA (Fixed Scheme) OMA (TDMA) 7 6 5 4 3 2 1 Fig. 7. 0 0.1 0.2 0.3 0.4 ρ 0.5 0.6 0.7 0.8 0.9 Performance comparison of a joint optimization scheme and a fixed allocation scheme. systems at SNR= 10 dB. In addition, we take a fixed scheme based on NOMA and a time division multiple access (TDMA) based on OMA as baseline schemes. Specifically, the joint optimization scheme first distributes the transmit power with equal feedback allocation, then allocates the feedback bits based on the distributed power, finally selects the optimal transmission mode. The fixed scheme always adopts the mode 2 (N = 3, K = 2) with equal power and feedback allocation. The TDMA equally allocates each time slot to the 6 MUs, and utilizes maximum ratio transmission (MRT) based on the available CSI at the BS to maximize the rate. For clarity of notation, we use ρ to denote the CSI accuracy based on equal feedback allocation. In other words, the total number of feedback bits is equal to Btot = −K ∗ N ∗ (M − 1) ∗ log2 (1 − ρ). As seen in Fig. 7, the fixed scheme performs better than the TDMA scheme at low and high 29 CSI accuracy, and slightly worse at medium regime. However, the proposed joint optimization scheme performs much better than the two baseline schemes. Especially at high CSI accuracy, the performance gap becomes substantially large. For instance, there is a performance gain of about 3 b/s/Hz at ρ = 0.8, and up to more than 5 b/s/Hz at ρ = 0.9. As analyzed in Lemma 2 and confirmed by Fig. 6, when ρ is larger than 0.8, which is a common CSI accuracy in practical systems, mode 2 is optimal for maximizing the system performance. Thus, the joint optimization scheme is reduced to joint power and feedback allocation, which requires only a very low complexity. Thus, the proposed NOMA scheme with joint optimization can achieve a good performance with low complexity, and it is a promising technique for future wireless communication systems. VI. C ONCLUSION This paper provided a comprehensive solution for designing, analyzing, and optimizing a NOMA technology over a general multiuser multiple-antenna downlink in both TDD and FDD modes. First, we proposed a new framework for multiple-antenna NOMA. Then, we analyzed the performance, and derived exact closed-form expressions for average transmission rates. Afterwards, we optimized the three key parameters of multiple-antenna NOMA, i.e., transmit power, feedback bits, and transmission mode. Finally, we conducted asymptotic performance analysis, and obtained insights on system performance and design guidelines. R EFERENCES [1] Y. Saito, Y. Kishiyama, A. Benjebour, T. Nakamura, A. Li, and K. Higuchi, “Non-orthogonal multiple access (NOMA) for cellular future radio access,” in Proc. IEEE Veh. Technol. Conf. (VTC-Spring), Jun. 2013, pp. 1-5. [2] Z. Ding, Y. Liu, J. Choi, Q. Sun, M. Elkashlan, C-L. I, and H. V. Poor, “Application of non-orthogonal multiple access in LTE and 5G networks,” IEEE Commun. Mag., vol. 55, no. 2, pp. 185-191, Feb. 2017. [3] L. Dai, B. Wang, Y. Yuan, S. Han, C-L. I, and Z. Wang, “Non-orthogonal multiple access for 5G: solutions, challenges, opportunities, and future research trends,” IEEE Commun. Mag., vol. 53, no. 9, pp. 74-81, Sept. 2015. [4] Z. Ding, Z. Yang, P. Fan, and H. V. Poor, “On the performance of non-orthogonal multiple access in 5G systems with randomly deployed users,” IEEE Signal Process. Lett., vol. 21, no. 12, pp. 1501-1505, Dec. 2014. [5] Y. Yuan, Z. Yuan, G. Yu, C-H. Hwang, P-K. Liao, A. Li, and K. Takeda, “Non-orthogonal transmission technology in LTE evolution,” IEEE Commun. Mag., vol. 54, no. 7, pp. 68-74, Jul. 2016. [6] S. Timotheou and I. Krikidis, “Fairness for non-orthogonal multiple access in 5G systems,” IEEE Signal Process. Lett., vol. 22, no. 10, pp. 1647-1651, Oct. 2015. 30 [7] S-L. Shieh and Y-C. Huang, “A simple scheme for realizing the promised gains of downlink nonorthogonal multiple access,” IEEE Trans. Commun., vol. 64, no. 4, pp. 1624-1635, Apr. 2016. [8] P. Xu, Y. Yuan, Z. Ding, X. Dai, and R. Schober, “On the outage performance of non-orthogonal multiple access with 1-bit feedback,” IEEE Trans. Wireless Commun., vol. 15, no. 10, pp. 6716-6730, Oct. 2016. [9] Z. Yang, Z. Ding, P. Fan, and Z. Ma, “Outage performance for dynamic power allocation in hybrid non-orthogonal multiple access systems,” IEEE Commun. Lett., vol. 20, no. 8, pp. 1695-1698, Aug. 2016. [10] F. Fang, H. Zhang, J. Cheng, and V. C. M. Leung, “Energy-efficient resource allocation for downlink orthogonal multiple access (NOMA) network,” IEEE Trans. Commun., vol. 64, no. 9, pp. 3722-3732, Sept. 2016. [11] H. Tabassum, M. S. Ali, E. Hossain, M. J. Hossain, and D. I. Kim, “Non-orthogonal multiple access (NOMA) in cellular uplink and downlink: challenges and enabling techniques,” Aug. 2016. [Online]: http://128.84.21.199/abs/1608.05783. [12] J. Choi, “On the power allocation for a practical multiuser superposition scheme in NOMA systems,” IEEE Commun. Lett., vol. 20, no. 3, pp. 438-441, Mar. 2016. [13] C-L. Wang, J-Y. Chen, and Y-J. Chen, “Power allocation for downlink non-orthogonal multiple access system,” IEEE Wireless Commun. Lett., vol. 5, no. 5, pp. 532-535, Oct. 2016. [14] Z. Yang, Z. Ding, P. Fan, and N. Al-Dhahir, “A general power allocation scheme to guarantee quality of service in downlink and uplink NOMA systems,” IEEE Trans. Wireless Commun., vol. 15, no. 11, pp. 7244-7257, Nov. 2016. [15] Y. Liu, M. Elkashlan, Z. Ding, and G. K. Karagiannidis, “Fairness of user clustering in MIMO non-orthogonal multiple access systems,” IEEE Commun. Lett., vol. 20, no. 7, pp. 1465-1468, Jul. 2016. [16] H. Zhang, D-K. Zhang, W-X. Meng, and C. Li, “User pairing algorithm with SIC in non-orthogonal multiple access system,” in Proc. IEEE Intern. Conf. Commun. (ICC), May 2016, pp. 1-6. [17] J. Mei, L. Yao, H. Long, and K. Zheng, “Joint user pairing and power allocation for downlink non-orthogonal multiple access systems,” in Proc. IEEE Intern. Conf. Commun. (ICC), May 2016, pp. 1-6. [18] Z. Q. Al-Abbasi and D. K. C. So, “User-pairing based non-orthogonal multiple access (NOMA) system,” in Proc. IEEE Veh. Technol. Conf. (VTC-Spring), Apr. 2016, pp. 1-5. [19] Md S. Ali, H. Tabassum, and E. Hossain, “Dynamic user clustering and power allocation for uplink and downlink nonorthogonal multiple access (NOMA) systems,” IEEE Access, vol. 4, pp. 6325-6343, Aug. 2016. [20] H. Weingarten, Y. Steinberg, and S. S. Shamai, “The capacity region of the Gaussian multiple-input multiple-output broadcast channel,” IEEE Trans. Inf. Theory, vol. 52, no. 9, pp. 3936-3964, Sept. 2006. [21] M. A. Maddah-Ali, M. A. Sadrabadi, and A. K. Khandani, “Broadcast in MIMO systems based on a generalized QR decomposition: signaling and performance analysis,” IEEE Trans. Inf. Theory, vol. 54, no. 3, pp. 1124-1138, Mar. 2008. [22] A. D. Dabbagh and D. J. Love, “Precoding for multiple antenna Gaussian broadcast channels,” IEEE Trans. Signal Process., vol. 55, no. 7, pp. 3837-3850, Jul. 2007. [23] Q. Sun, S. Han, C-L. I, and Z. Pan, “On the ergodic capacity of MIMO NOMA systems,” IEEE Wireless Commun. Lett., vol. 4, no. 4, pp. 405-408, Oct. 2015. [24] J. Choi, “On the power allocation for MIMO-NOMA systems with layered transmission,” IEEE Trans. Wireless Commun., vol. 15, no. 5, pp. 3226-3237, May 2016. [25] Z. Chen, Z. Ding, and X. Dai, “Beamforming for combating inter-cluster and intra-cluster interference in hybrid NOMA systems,” IEEE Access, vol. 4, pp. 4452-4463, Aug. 2016. [26] Z. Ding, R. Schober, and H. V. Poor, “A general MIMO framework for NOMA downlink and uplink transmission based on signal alignment,” IEEE Trans. Wireless Commun., vol. 15, no. 6, pp. 4438-4454, Jun. 2016. 31 [27] K. Higuchi and Y. Kishyama, “Non-orthogonal access with random beamforming and intra-beam SIC for cellular MIMO downlink,” in Proc. IEEE Veh. Technol. Conf. (VTC-Fall), Sept. 2013, pp. 1-5. [28] Z. Ding, F. Adachi, and H. V. Poor, “The application of MIMO to non-orthogonal multiple access,” IEEE Trans. Wireless Commun., vol. 15, no. 1, pp. 537-552, Jan. 2016. [29] D. J. Love, R. W. Heath Jr., V. K. N. Lau, D. Gesbert, B. D. Rao, and M. Andrews, “An overview of limited feedback in wireless communication systems,” IEEE J. Sel. Areas Commun.,, vol. 26, no. 8, pp. 1341-1365, Oct. 2008. [30] N. Zhang, J. Wang, G. Kang, and Y. Liu, “Uplink non-orthogonal multiple access in 5G systems,” IEEE Commun. Lett., vol. 20, no. 3, pp. 458-461, Mar. 2016. [31] Y. Liu, Z. Ding, M. Elkashlan, and H. V. Poor, “Cooperative non-orthogonal multiple access with simultaneous wireless information and power transfer,” IEEE J. Sel. Areas Commun., vol. 34, no. 4, pp. 938-953, Mar. 2016. [32] N. Nonaka, Y. Kishiyama, and K. Higuchi, “Non-orthogonal multiple access using intra-beam superposition coding and SIC in base station cooperative MIMO cellular downlink,” in Proc. IEEE Veh. Technol. Conf. (VTC-Spring), Sept. 2014, pp. 1-5. [33] S. Ali, E. Hossain, and D. I. Kim, “Non-orthogonal multiple access (NOMA) for downlink multiuser MIMO systems: user clustering, beamforming, and power allocation,” IEEE Access, vol. 5, pp. 565-577, Oct. 2017. [34] M. B. Shahab, M. Irfan, M. F. Kader, and S. Y. Shin, “User pairing schemes for capacity maximization in nonorthogonal multiple access systems,” Wireless Commun. Mob. Comp., vol. 16, no. 17, pp. 2884-2894, Dec. 2016. [35] Z. Ding, P. Fan, and V. Poor, “Impact of user pairing on 5G non-orthogonal multiple access downlink transmissions,” IEEE Trans. Veh. Technol., vol. 65, no. 8, pp. 6010-6023, Aug. 2016. [36] M. B. Shahab, M. F. Kader, and S. Y. Shin, “A virtual user pairing scheme to optimally utilize the spectrum of unpaired users in non-orthogonal multiple access,” IEEE Signal Process. Lett., vol. 23, no. 12, pp. 1766-1770, Dec. 2016. [37] K. Saito, A. Benjebbour, Y. Kishiyama, Y. Okumura, and T. Nakamura, “Performance and design of SIC receiver for downlink NOMA with open-loop SU-MIMO,” in Proc. IEEE Intern. Conf. Commun. Workshop (ICCW), Jun. 2015, pp. 1161-1165. [38] M. B. Shahab, M. F. Kader, and S. Y. Shin, “On the power allocation of non-orthogonal multiple access for 5G wireless networks,” in Proc. Intern. Conf. Open Source Systems Technol. (ICOSST), Dec. 2016, pp. 89-94. [39] N. Jindal, “MIMO Broadcast channels with finite-rate feedback,” IEEE Trans. Inf. Theory, vol. 52, no. 11, pp. 5045-5060, Nov. 2006. [40] K. K. Mukkavilli, A. Sabharwal, E. Erkip, and B. Aazhang, “On beamforming with finite rate feedback in multiple-antenna systems,” IEEE Trans. Inf. Theory, vol. 49, no. 10, pp. 2562-2579, Oct. 2003. [41] G. K. Karagiannidis, N. C. Sagias, and T. A. Tsiftsis, “Closed-form statistics for the sum of squared Nakagami-m variates and its application,” IEEE Trans. Commun., vol. 54, no. 8, pp. 1353-1359, Aug. 2006. [42] I. S. Gradshteyn and I. M. Ryzhik, “Tables of intergrals, series, and products,” Acedemic Press, USA, 2007.
7cs.IT
Motion Planning with Gamma-Harmonic Potential Fields Ahmad A. Masoud Electrical Engineering Department, King Fahad University of Petroleum and Minerals, Dhahran 31261, Saudi Arabia, e-mail: [email protected] Abstract: This paper extends the capabilities of the harmonic potential field (HPF) approach to planning. The extension covers the situation where the workspace of a robot cannot be segmented into geometrical subregions where each region has an attribute of its own. The suggested approach uses a task-centered, probabilistic descriptor of the workspace as an input to the planner. This descriptor is processed, along with a goal point, to yield the navigation policy needed to steer the agent from any point in its workspace to the target. The approach is easily adaptable to planning in a cluttered environment containing a vector drift field. The extension of the HPF approach is based on the physical analogy with an electric current flowing in a nonhomogeneous conducting medium. The resulting potential field is known as the gamma-harmonic potential (GHPF). Proofs of the ability of the modified approach to avoid zero-probability (definite threat) regions and to converge to the goal are provided. The capabilities of the planner are demonstrated using simulation. Keywords: probabilistic robotics, autonomous agents, motion planning, navigation, risk minimization, harmonic potential Nomenclature HPF: Harmonic potential field GHPF: Gamma-harmonic potential field BVP: Boundary value problem LFC: Liapunov function candidate P(x): a probabilistic field describing of the fitness of the point x to carry-out a task S: So: Sa: O: T: u(x): xs: xT: N: D(t): H(x): L: L2: L@: ': V(x): F(x): J(x): Q(x): =: Ξ : workspace of the agent E: definite threat (P(x)=0) subset of S the admissible region of the probabilistic workspace (Sa=S-So) forbidden regions in deterministic worspaces an infinitesimally small set in S navigation policy starting point of motion target point the empty set the time paramerized trajectory laid by the planner the hessian matrix the gradient operator the laplacian operator the divergence operator boundary of the deterministic workspace potential field conductivity electric current density a vector field describing the drift force in the workspace Liapunov function time derivative of = minimum invariant set I. Introduction: Designing an autonomous agent is a challenging multi-disciplinary task [1]. Special attention is being paid to the propulsion, data acquisition and communication systems used by an agent. However, the biggest challenge seems to be in designing a proper planning module. The function of this module is to unite these sub-systems into one goal-oriented unit. There is a long list of conditions a planner must satisfy. These conditions are necessary in order to generate a sequence of instructions which the actuators of motion may execute to successful completion of an assignment. However, the conditions on handling and representing mission data seem to be the most stringent [2]. There are two core requirements a representation should satisfy. They are 1- compatibility with the manner in which data is being processed and action is being generated, 2- updatability of the representation. Updatability requires that the validity of the existing portion of the representation not be conditioned on the future data that could be received. This allows for the incremental construction of a representation. Another important issue is to increase the diversity of environment-related, operator-supplied information which the planner is capable of processing in order to yield the guidance signal. Most planners assume divisible environments that may be partitioned into subsets of homogeneous attributes. The most common scheme is to have a binary partition of admissible sets and forbidden ones. Binary partitions may be constructed using geometric structures like circles [3], occupancy maps, Voronoi partitions [4], grids and graphs [5], samples [9] and trees [6]. There are situations where the environment in which an agent is operating is not divisible. For example, a plane flying through turbulent atmosphere [7] experiences a degree of turbulence where clear space is diffused into turbulent space with no sharp boundaries separating the two. Also, in the case of mobile robots operating in rough terrains [8], the description should be based on the degree of difficulty in negotiating the terrain. It is highly unlikely that success can be attained by basing the actions on a binary geometric partition of the environment into admissible and forbidden regions. An alternative to the geometric approach is the use of a soft representation that consists of a field reflecting, at each point in the environment, the probability of achieving the task. Probabilistic representations are ideal for encoding the information in non-divisible environments. These representations may also be used to efficiently incorporate spatial ambiguity and the aging of information in the representation. For example, exact knowledge of the location of a one-dimensional point object may be represented as a probability distribution function (PDF) with an impulse function (figure-1). Convolution of this exact spatial knowledge with the proper blurring operator produces another representation with the ambiguity factored-in. Repeated application of the blurring operator leads to a uniform PDF representing the maximum state of ambiguity (i.e. the object could be anywhere). Figure-1: spatial ambiguity and aging of information Errors and uncertainty in the sensed signals are almost always present in navigation systems. Countermeasures [38] to reduce their effect on the generated trajectory have to be incorporated into the system for robust operation. Soft probabilistic representations are an effective approach to counteracting ambiguity. They have been adopted by many researchers in robotics. For example, in [10,11], probabilistic representations were used for path planning in the presence of sensory data ambiguity. This representation may be fed to a reinforcement learning-based, or an optimal control-based stage, to generate the navigation policy. It is noticed that the structure of the ambiguity field has a pronounced influence on the performance of the policy generator. The less structured the field is, the more time it takes the algorithm to converge. In [39], the ambiguity in determining the location of a target moving in an obstacle-free space was probabilistically modeled using an occupancy map. This map is fed into the potential field stage to direct the motion of the pursuer in a direction that would increase the probability of capturing the target. Fuzzy logic techniques [12] are used to derive navigation actions for robots based on soft probabilistic representations. While these techniques have proven their practicality, they are not provably correct (heuristics). Qualitative methods for dealing with uncertainty may also be found in [13]. Occasionally, a workspace contains a force external to the agent that influences its state. This force is known as the drift field. A drift field is usually treated as a source of disturbance whose influence should be suppressed by the agent’s low-level controller. With the advances in forecast technology [14], a drift field may be predicted for a considerable period ahead of time. These advances make it possible to consider drift as a source of information which the planner should take into consideration instead of being a source of disturbance that has to be suppressed. Being able to treat drift as a source of information has important applications in the area of energyexhaustive missions. When energy is a scarce resource, good planning does not only reduce the drain caused by drift, but may even use it as a source for powering the agent. In the past, several techniques were suggested for incorporating drift fields in the planning process. Most of these techniques use optimization or search-based methods for determining the minimum cost path that connects two points in the driftoccupied space while avoiding the obstacles. In [15], a genetic algorithm planner is used to lay a minimum energy trajectory for an AUV operating in turbulent waters. A tree-based trajectory planner [16] is used to plan a path for an aerial glider so that the path is laid along the component of the wind which exerts maximum lift. The A* search approach is used for planning a path for an AUV operating in a current field [17]. A symbolic wave expansion approach is developed to tackle planning in workspaces with dynamic current fields [18]. Other approaches may be found in [19,20]. This study suggests a method for generating the navigation policy for a robot using a goal point, and a taskcentered, probabilistic description of the environment, as inputs. The method is efficient, provably-correct and assumes no structure whatsoever on the descriptor used to represent the environment. The developed method can be adapted in a straightforward manner to motion planning in a cluttered workspace that is populated by a drift field. The suggested approach is called the gamma-harmonic [33] potential field planning approach. It is derived by extending the capabilities of the deterministic harmonic potential field (HPF) approach to motion planing [21,22,23] in order to tackle environments that can only be probabilistically described. Previous attempts to construct a probabilistic HPF planner focus on using HPF as a probability measure describing the danger of collision with obstacles in the environment. This strips the HPF approach from its welljustified role as a navigation policy generator and reduces it to a merely descriptive tool that is an input to a navigation policy generation stage [24,25]. This paper is organized in the following manner. In section II, the problem is stated along with its scope. In section III, the probabilistic, HPF-based planning approach is developed. In section IV, an interpretation of the task-centered, probabilistic descriptor is suggested for incorporating drift in the trajectory generation process. Section V provides a basic analysis of the behavior of the approach. Simulation results are in section VI and conclusions are placed in section VII. II. Problem Statement and Scope The objective of this work is to develop a navigation guidance policy (u) generator that can be used by a point mobile agent to convert an a priori known, differentiable probabilistic field (P) that describes the content of the workspace (S) and a target point (XT) into a well-behaved vector field that may be used by the agent to direct its motion in S . The probabilistic field describes, in a task-centered manner, the fitness of a point x in S to carry-out the intended task. Let So ( So d S) be the set of definite threat region defined as (1) Ωo = {x:P(x) = 0} . regardless of the structure of P, the first order dynamical system that is constructed using the navigation policy (2) x = u(x, P(x), x T ) ∀ x( 0) ∈ Ω − Ω o must yield a goal-seeking guidance action that enables a point agent to converge from anywhere in S-So to the target lim x(t) → x T . t→ ∞ (3) The navigation policy must also guarantee that definite threat regions (So) are avoided at all times: x(t) ∩ Ωo ≡ ϕ œt . (4) The trajectory D(t) which is the solution of the dynamical system in (2) is required to be analytic. It is also required that the risk (risk=1/P(x)) accumulated by moving along this trajectory ∞ 1 ∫0 (P( ρ(t))) ρ (t) dt (5) be reduced to a global minimum. To the best of this author’s knowledge, the suggested navigation policy generator is the first to directly generate a goal-seeking, guidance action from a probabilistic description of the environment. The approach amasses a considerable number of novel and important properties: 1- it can process inherently ambiguous data while making provably-correct statements on the resulting behavior of the agent, 2- the probabilistic descriptor has an intuitive format whose local effect on the generated action is discernable by the operator. This makes the policy generator a valuable component in a decision support system [40], 3- the probabilistic descriptor need not be generated from an underlying deterministic model of the environment. Therefore, its use extends beyond that of modeling the ambiguity and uncertainty of a deterministic model to dealing with inherently stochastic environments such as rough terrains, 4- the goal-seeking guidance action adds to the robustness of the planning process in the sense that if a disturbance throws the agent off-course, the new location contains another sequence of instructions that will lead the agent to the target. The analyticity of the solutions making up the navigation policy causes the deviation from the original path to be commensurate with the amount of disturbance the agent is subjected to, 5- forcing the solution trajectories to be analytic results in well-behaved, dynamically-friendly paths that can, with high probability, be converted by the control unit of the agent into a navigation control, 6- as will be demonstrated in this paper, the nature of the probabilistic descriptor, and the fact that it doesn’t have to satisfy any condition other than differentiability, have significant advantages. They make it possible, with little effort, to change the format of the environment data to suit other challenging planning problems, e.g., planning in the presence of drift fields for energy exhaustive missions. III. The Gamma-Harmonic Approach III.1: HPF: A Background Harmonic potential fields (HPFs) have proven to be effective tools for inducing, in an agent, an intelligent, emergent, embodied, context-sensitive and goal-oriented behavior (i.e. a planning action). A planning action generated by an HPF-based planner can operate in an informationally-open and organizationally-closed mode [26]. Thus, an agent is able to make decisions on-the-fly using on-line sensory data without relying on the help of an external agent. HPF-based planners can also operate in an informationally-closed, organizationally-open mode. This mode makes it possible to utilize existing data about the environment in generating the planning action as well as eliciting the help of external agents in managing the task at hand. Such features make it possible to adapt HPFs for planning in a variety of situations. For example, in [27] vector-harmonic potential fields were used for planning with robots having second order dynamics. In [28], the approach was configured to work in a pursuit-evasion planning mode, and in [29], the HPF approach was modified to incorporate joint constraints on regional avoidance and direction. The decentralized, multi-agent planning case was tackled using the HPF approach in [30]. The HPF approach was also found to be able to facilitate the integration of planners as subsystems in networked controllers [31]. It effectively addresses the real-life requirements needed to successfully network sensory, communication and control modules. A basic setting of the HPF approach is shown in (6). L2V(x)/0 X0S ∂V = 0 at x = ', subject to: V(xS) = 1, V(xT) = 0 , and ∂n A provably-correct path may be generated using the gradient dynamical system Solve: (6) x = -∇V(x). (7) where S is the workspace, ' is its boundary, n is a unit vector normal to ', xs is the starting point, and xT is the target point [29]. III.2: A Physical Metaphor: Equations (6) & (7) model the behavior of an electric current flowing in a homogeneous conductor with a conductivity F(x)= constant [32]. The conductor is populated by perfect insulators (F=0) occupying the forbidden regions surrounded by ' (figure-2). Figure-2: Physical metaphor for the planner The flow is described by the electric current density(J) J(x) = −σ (x)∇V(x) . (8) The homogeneous Neumann condition means that the current cannot penetrate the perfect conductor and has to move tangentially to it. The Laplce equation is simply a product of applying the continuity condition to the electric flow ∇ ⋅ ( −σ (x)∇V(x)) = 0 . (9) If F is constant, equation (9) reduces to the well-known Laplace operator in (6). The conductivity F may be used to represent how favorable a point in the robot’s space is to conducting motion. A F=0 means that the corresponding space does not support motion at all. On the other hand, a high value of F means that the corresponding space highly favors motion. III.3 The Gamma-Harmonic Planner It is possible to establish a tight analogy between the above situation and the situation where the agent’s environment is to be represented in a task-centered manner using a descriptor (P(x)). This descriptor specifies at each point in the space the agent’s ability to perform the assigned task. It does not matter what the causes are e.g. sensor problems, rough terrains, man-made hazzards etc. If the agent is expected to encounter difficulties operating in a certain region of space, a low value for P(x) is assigned to that region. A navigation policy that satisfies the conditions stated in II may be generated as LA(P(x)LV(x))/0 subject to x0S (10) V(xS) = 1, V(xT) = 0 A provably-correct , analytical trajectory may be constructed using the gradient dynamical system u(x) = -∇V(x). (11) The potential field (V) generated by (10) is known as the gamma-harmonic potential field (GHPF) [33]. The GHPF operator in (10) is strongly related to the ordinary Laplace operator and possesses a useful physical interpretation. First, the operator should be expanded as LA(P(x)LV(x)) = P(x)L2V(x) + LP(x)tLV(x)) =0, which leads to (12) L2V(x) = - (1/P(x)) (-LP(x)t (-LV(x))). Notice that -LV(x) is the direction at which motion is to be driven and -LP(x) is a vector pointing in the direction of increasing risk. Also, keep in mind that the Laplacian of a potential is the divergence of the gradient field which is physically defined as the outflow of flux when the volume shrinks to zero. If the Laplacian is negative, there is a sink in the closed region, i.e. motion is indebted by absorbing the gradient flux used to direct motion. If the Laplacian is positive, there is a source in the closed region, that is, motion is stimulated by aiding the flow of the gradient flux. When the Laplacian is zero (Lapace equation), the region is neutral towards the gradient flux. As can be seen, the GHPF operator may be viewed as an intelligent version of the HPF that is sensitive to the future ability of the agent to carry-out its task (figure-3). Figure-3: The modified operator exhibits intelligent behavior. IV. Adapting the planner to the drift case It is shown that P(x) may be easily interpreted in an opportunistic manner that exploits drift as a source for powering the agent. A planner that mitigates the effect of drift on motion is expected to satisfy the following conditions lim x(t) → x T , ∀x(0) ∈ Ω x(t) ∩ O ≡ φ ∀t t→ ∞ U= ∫ ∞ 0 Fc( Ψ (x(t)),−∇V(x(t))) ∇V(x(t)) dt (13) where the functional U has to be minimized, or reduced, to a satisfactory value. x 0RN, O is the set of forbidden regions (obstacles, '=MO), S is the subset of admissible space (worksapce), Q(x) is the field in S describing drift, U is a task-related cost functional constructed by accumulating point costs (Fc(x)) along the path of the agent from the starting point to the target and Fc is a point cost function constructed in aim with the aspect of interest to the operator. In an energy exhaustive mission, it is desirable that the obstaclefree path connecting the starting and end points has a drift component that is in-phase with the direction along which motion is heading. A choice of Fc is (figure-4) K ∇V T Ψ K Fc = (1 − cos(Θ)) = (1 + ). 2 ∇V Ψ 2 P(x) = K − Fc = A drift-sensitive descriptor is K ∇V T Ψ (1 − ). 2 ∇V Ψ (14) (15) where 1 is the angle between -LV and Q and K is a positive constant. Here P(x) is used to describe the utility of the drift at a certain point in space. Figure-4: The suggested utility function versus 1. The drift-sensitive GHPF operator is ∇V T Ψ ∇ V − ∇⋅( )∇V) ≡ 0. ∇V Ψ 2 (16) and the overall drift-sensitive navigation policy generator is ∇V T Ψ )∇V) ≡ 0. x0S (17) ∇V Ψ subject to V(xS) = 1, V(xT) = 0 , and P(x)/0 at x 0O, where O is the set of deterministic obstacles to ∇2V − ∇ ⋅ ( be avoided. The path may be generated using the dynamical system in (7). V. Stability Analysis: In this section, propositions along with their proofs are provided to explore the behavior of the GHPF approach suggested in section III. Proposition-1: The path generated by the PDE-ODE system in (10) will avoid regions that have zero probability of achieving the task (So). Proof: Assume a point x that is arbitrarily close to So (figure-5). Since P(x) is differentiable, its value may be assumed equal to zero. Using the identity LA(P(x)LV(x)) = LP(x)t LV(x) + P(x)L2V(x) = 0 when x is close to So, equation 18 reduces to (18) - (-LP(x)t )(-LV(x)) = 0. Note that -LP(x) points in the direction of increasing risk that leads to the region So, while -LV(x) is the direction along which motion is to be steered. In other words, in the vicinity of So, the planner will project motion tangentially to the boundary of the zero probability region. Hence, So will be avoided. Figure-5: Zero-P regions will be avoided, Proposition-2: A potential field that is generated using the boundary value problem in (10) cannot have any extrema local or global in its admissible workspace S-So. Proof: Note that from proposition-1, x will always stay in S-So where P(x) is greater than zero. The fact that no minima or maxima can occur in S-So may be inferred directly from the GHPF operator. Using the expansion in (12) and noting that at a critical point LV(x)=0, we have ∂ 2 V(x) ∇ ⋅ (P(x)∇V(x)) = P(x)∇ V(x) = P(x) ∑ =0 ∂ x 2i i =1 2 N (19) where x=[x1 x2 ... xN]t. At a local maxima, LV(x) will be negative in the whole local neighborhood surrounding x. This means that the second derivative of V(x) with respect to all xi’s has to be positive (i.e. V(x) at the critical point is concaved upward). Since P(x) in that region is finite and positive, the governing relationship in (10) will be violated. The same thing will happen at a minima. LV(x) will be positive in the whole local neighborhood surrounding x. Therefore, no extrema of V local or global can occur in S-So. Proposition-3: For an Sa (Sa = S - So)with a finite size, V(x) 0 Sa is a Liapunov function candidate (LFC). Proof: An LFC defined on a finite space must satisfy the following: 1- it must be at least continuous, 2- it must be positive in Sa (V(x) > 0, x 0 Sa), 3- its value must be zero at the target point (V(xT) =0). Since V(x) is forced to satisfy the differential condition in (10), it is analytic. Therefore, it satisfies the first condition. Since the global maximum happens at x=xs (V(xs)=1) and a global minimum at x=xT (V(xT)=0), the second and third conditions are satisfied. Proposition-4: If V(x) is constant at a subset of S, it is constant for all S. Proof: First assume that V(x) = C ( C is a constant) in T where Td S (figure-6). Then consider an infinitesimally expanded region T` that surrounds T. Finally, let xo be a point that lies on the boundary of T (MT) and xo+ a point on MT`. Then, the potential at xo+ may be written as V(xo+) = V(xo) + drA(-LV(xo)t n) (20) where dr is a differential element and n is a unit vector normal to MT. Since V is constant inside T, the gradient field degenerates to zero. Since the continuity relation (10) is enforced in both T and T`, equation (20) reduces to: V(xo+) = V(xo) = C . (21) By repeatedly applying the above procedure, it can be shown that the subregion T may be expanded to include all S. In other words, if V(x) is constant on a subregion of S, it will be constant for all S. Figure-6: Subregions of degenerate fields cannot occur, Definition-1: Let V(x) be at least twice differentiable scalar function (V(x): RN 6 R). A point xo is called a critical point of V if the gradient vanishes at that point (LV(xo)=0). Otherwise, xo is regular. A critical point is Morse if its Hessian matrix (H(xo)) is nonsingular. V(x) is Morse if all of its critical points are Morse [34]. Showing that the GHPF is Morse is a key requirement for generating a provably-correct navigation policy [45]. The critical points in a Morse function are saddle points that do not trap motion and prevent the agent from reaching its target. Proposition-5: If V(x) is a function defined in an N-dimensional space (RN) on an open set S and satisfies (10), then the Hessian matrix at every critical point of V is nonsingular, i.e. V is Morse. Proof: There are two properties of V that are used in the proof of proposition 5. 1- V(x) defined on an open set S contains no maxima or minima, local or global in S. An extrema of V(x) can only occur at the boundary of S, 2- if V(x) is constant in any open subset of S, then it is constant for all S. Let xo be a critical point of V(x) inside S. Since no maxima or minima of V exist inside S, xo has to be a saddle point. Let V(x) be represented in the neighborhood of xo using a second order Taylor series: V(x) = V(x O ) + ∇V(x O ) T (x − x O ) + 1 (x − x O ) T H(x O )(x − x O ) 2 *x-xo*<<1. (22) Since Xo is a critical point of V, we have V ' = V(x) - V(x O ) = 1 (x − x O ) T H(x O )(x − x O ) 2 *x-xo*<<1. (23) Notice that adding or subtracting a constant to V yields another potential field that satisfies the relation in (10). Using eigenvalue decomposition [35] ⎡λ1 ⎢0 1 ' T T⎢ V = (x − x O ) U ⎢. 2 ⎢ ⎣0 0 . 0 . ⋅ 0 . λ2 ⎡λ1 ⎤ ⎢0 ⎥ ⎥ U(x − x O ) = 1 ξ T ⎢ ⎥ 2 ⎢. ⎥ ⎢ λN ⎦ ⎣0 0 0 . 0 λ2 . 0 0 . ⋅ . ⎤ ⎥ ⎥ξ = 1 ⎥ 2 ⎥ λN ⎦ 0 0 . N ∑λ ξ i i i=1 2 (24) where U is an orthonormal matrix of eigenvectors, 8’s are the eigenvalues of H(xo), and >=[>1 >2 ..>N]T = U(x-xo). V` cannot be zero on any open subset S. Otherwise, it will be zero for all S which is not the case. This can only be true if and only if all the 8i’s are nonzero. In other words, the Hessian of V at a critical point Xo is nonsingular. This makes V a Morse function. Proposition-6: Let V(x) be the potential field generated using the BVP in (10). The trajectory of the dynamical system will globally, asymptotically converge to: x = −∇V(x) lim x → x T (25) x(0) = xs 0 Sa t→ ∞ (26) Proof of the above proposition is carried out using the LaSalle invariance principle [36]. Proof: Let = be the Liapunov function candidate Ξ = V(x) (27) 1 T  . x D(x)x + x T D(x)x 2 x = −∇V(x)  x)  = K ⋅ ∇V(x) x + The time derivative (27) is Ξ(x, (28) Substituting in the above equation yields (29) (30) T x = − ∇V(x) 2 LV will vanish at the target point (xT) and may have isolated critical points {xi} in Sa. This results in Ξ ≤ 0 ∀ x. (31) According to the LaSalle principle, any bounded solution of (25) will converge to the minimum invariant set E ⊂ {x T ∪ {x i }} . (32) Determining E requires studying the critical points of V(x) where LV(x)=0. According to proposition-2, xT is the only minimum (stable equilibrium point) V(x) can have. Besides xT , V(x) has other critical points {xci} at which LV=0. However, the Hessian at these points is non-singular, i.e. V(x) is Morse. From the above, we conclude that E contains only one point x = x T to which motion will converge. Proposition-7: Let D(t) be the time parameterized trajectory generated by the gradient dynamical system x = −∇V(x) x(0) = xs 0Sa (33) where D(0)= xs, D(4)=xT and V(x) is generated from the GHPF BVP in (10). Then D(t) minimizes the cost ∞ 1 ∫ P(x( t)) ∇V(x( t)) dt functional: 0 (34) Proof: The proof follows directly from using electric current to model motion and using the conductivity field to model cost (cost=1/fitness). It is well-known that electric current will flow along the minimum resistance (cost) path. For a mathematical proof, that is based on the calculus of variation, see [43]. It is also worth mentioning that in addition to the above, the suggested navigation policy globally minimizes the energy functional ∫ P(x) ∇V(x) dΩ 2 (35) Ω which is known as the Dirichlet principle [44]. VI. Results: In this section, simulation results are presented to demonstrate the capabilities of the GHPF approach. V.1: From Probabilistic to Deterministic Planning: There are several settings in which a harmonic potential field may be configured for navigation. Some of these settings are discussed in [31,32]. Each one of these configurations possesses distinct topological properties that are reflected in the integral and differential properties of the generated path. Figure-7 shows the control navigation policies of four different configurations for a simple rectangular environment. Homogeneous Neumann Nonhomogeneous Neumann Dirichlet Biharmonic Figure-7: Navigation policies, different settings, HPF approach. The following example demonstrates the ability of the GHPF planner to handle deterministic environments by simply binary quantizing P(x) into 0 and 1(figure-8). Figure-8: Probabilistic descriptor, Figure-9: Path and navigation policy - GHPF approach, The navigation control policy and the generated path are shown in figure-9. Unlike the deterministic case where the navigation policy degenerates inside the region to be avoided, the probabilistic HPF approach maintains the navigation field inside this region. This is of practical value because if a disturbance occurs throwing the robot inside a forbidden region, the robot can resume motion to the target instead of staying motionless in that area. V.2: Typical Non-divisible Environments: In figure-10, a more challenging environment is used to test the approach. The probabilistic descriptor is shown as a 3-dimensional plot. As can be seen, segmenting this map into regions suitable for navigation, and into others that are not, is very difficult, if at all possible. Moreover, this binary segmentation could result in creating isolated regions that are disconnected from the rest of the workspace. Figure-10: Probabilistic descriptor - surface map, The navigation control policy is generated directly from the probabilistic map in figure-10 by the GHPF planner without any conditioning to the signal. The generated path superimposed on the image of the probabilistic descriptor (the brighter the area, the more suitable it is for navigation) of the environment is shown in figure-11. Visual assessment of the path in figure-11 clearly shows its smoothness and reasonable length. The path passes mainly through bright areas which are best suited for navigation. The corresponding navigation policy is shown in figure-12. Figure-11: Generated path, Figure-12: Navigation policy. Figure-13 shows another nondivisible environment for different starting and end points. Figure-13: Different paths, same environment. V.3: Comparison With the D-Star Algorithm: The GHPF approach is compared to the D-star algorithm [41,42]. The D-star is the benchmark for planning in the presence of uncertainties. Figure-14 shows an intensity map reflecting the fitness of space for motion. The brighter the region the more suitable it is for motion to pass through. The trajectory generated by the D-star is shown in figure-15. The trajectory generated by the GHPF planner along with an alternative path resulting from adding an obstruction (zero fitness zone) is shown in figure-16 . As can be seen, the trajectories from both techniques have comparable lengths and seek the most suitable area for motion to pass through. However, the path from the GHPF is considerably smoother than that produced by the D-star algorithm. As a result, the GHPF path is favorable for a dynamical agent to traverse. One must keep in mind that the GHPF approach is a goal-seeking planner where as the D-star is a path following planner. If a disturbance throws the agent off-course, an agent using GHPF will still move towards the target in an acceptable manner while an expensive re-planning has to be used in the case of the D-star. Figure-14: The probabilistic environment. Figure-15: Path from the D-Star Algorithm Figure-16: Paths from the GHPF planner. V.4: Extreme Disorder: The capabilities of the GHPF planner in handling unstructured descriptors are tested using natural white noise as a representation of the environment (figure-17). Figure-18 shows the generated path superimposed on the image of the descriptor. One may observe the following: 1- the method works well although a highly unstructured, nondifferentiable white noise is used as the environment descriptor, 2- the path did converge to the target, 3- the path remained smooth with a reasonable length, 4- Although it is generally accepted that white noise means zero information, the path is low risk-sensitive. Figure-19 shows the path superimposed on two risk levels where the black areas of the first risk level are higher than 95% of the maximum of the environment and the other is higher than 85% of the maximum. As can be seen in both cases, the planner steered the path away from those regions. Figure-17: White noise, Figure-18: Path - random environment, Figure-19: Path versus risk levels. V.5: Drift-sensitive Planning The ability of the planner to process the drift data and the geometry of the space while generating a wellbehaved navigation policy is demonstrated in different drift scenarios. In figure-20, the planner tackles a drift that has a vortex form and is rotating in a counter-clockwise direction. Two paths are shown; one with the box interior empty and one with an obstruction added. As can be seen in both cases, an obstacle-free trajectory that exploits the presence of drift in powering motion is laid to the target. Figure-20: Trajectory in a ccw vortex In figure-21, a random correlated drift field is used instead of the vortex field. As can be seen, a smooth path with reasonable length respecting the boundary of the space is laid to the target. Figure-22 shows the difference between the trajectory heading and the angle of drift at different points along the path. One can see that the majority of the drift components are aiding motion. Notice that motion starts from a difficult area where there is no direction to proceed along without going against drift. Figure-21: Trajectory in a random drift field, Figure-22: Drift-heading differential in figure-21. Figure-23: Navigation policy in figure-21. Figure-23 shows the navigation policy corresponding to the drift in figure-21. Despite the random nature of drift, a smooth navigation policy is obtained for steering motion. In figure-24, an additional obstacle is added in the middle of the workspace for another case of random drift. As before, a smooth and safe path with reasonable length is generated. Figure-25 shows the difference between the heading of the trajectory and the angle of the drift. As can be seen, almost all the points of the trajectory lie in the energy harvesting zone with the absolute value of the difference being less than B/2. Figure-24: Trajectory in a random drift field. Figure-25: Drift-heading differential in figure-24. VII. Conclusion: In this paper, the gamma-harmonic potential field approach is suggested for tackling planning in environments with inherent uncertainty that deny an operator the ability to segment the workspace into geometric regions of homogeneous attributes. An interpretation of the GHPF approach is also suggested for handling planning in cluttered environments that are populated by a drift field. The suggested extensions are proof of principle that the HPF approach is capable of efficiently addressing the information diversity issue needed for a planner to tackle a realistic situation in a provably-correct manner. Acknowledgment: The author would like to thank King Fahad University of Petroleum and Minerals for its support of this work. References: [1] B. Siciliano, O. Khatib, “Handbook of Robotics”, Springer-Verlag, Berlin Heilderberg, 2008 [2] W. Woods, “What’s Important About Knowledge Representation”, Computer, Vol. 16, Issue 10, 1983, pp. 22-27 [3] H. Moravec, “Obstacle Avoidance and Navigation in the Real World by a Seeing Robot Rover”, CMURI, TR-3, 1980 [4] C. Jorgensen, W. Hamel, C. Weisbin, “Autonomous Navigation”, Byte, January 1986, pp. 223-235 [5] G. Giralt, R. Chatila, M. Vaisset, “An Integrated Navigation and Motion Control System for Autonomous, Multisensory, Mobile Robots”, Robotics Research: The first international symposium, 1990, pp. 420-443 [6] S. LaValle, “Planning Algorithms”, Cambridge University Press, 2006 [7] H. Stough, “Aircraft Weather Mitigation for the Next Generation Air Transportation Systems”, 23rd Conference on IIPS, Dec. 2004, 15.8, pp. 1-6. [8] K. Iagnemma, “S. Dubousky, “Mobile Robots in Rough Terrains”, Springer Tracts in Advanced Robotics, Vol. 12, 2004 [9] M. Unser, “Sampling-50 Years After Shannon”, Proc. IEEE Vol. 88, No. 4, pp. 469-587, April 2000. [10] S. Thrun, W. Burgard, D. Fox, “Probabilistic Robotics”, MIT Press, 2005 [11] R. Simmons, S. Koenig, “Probabilistic Robot Navigation in Partially Observable Environments”, Proceedings of the International Joint Conference on Artificial Intelligence ( IJCAI’90), 1995, pp. 10801087. [12] X. Yang, “Development and Application of Fuzzy Logic-Based Autonomous Robot Navigation Algorithms”, Ph.D Thesis, University of Toronto, Canada 2005. [13] S. Parsons, “Qualitative Methods for Reasoning Under Uncertainty”, MIT Press 2001. [14] J. Armstrong, “Long-Range Forecasting”. John Wiley 1985. 2nd Edition. [15] A. Alvarez, A. Caiti, R. Onken, "Evolutionary Path Planning for Autonomous Underwater Vehicles in a Variable Ocean", IEEE Journal of Oceanic Engineering, vol. 29, no. 2, April 2004, pp. 418-429 [16] Jack W. Langelaan, "Tree-based Trajectory Planning to Exploit Atmospheric Energy", 2008 American Control Conference, Westin Seattle Hotel, Seattle, Washington, USA, June 11-13, 2008 [17] Bartolome Garau, Alberto Alvarez, Gabriel Oliver, "Path Planning of Autonomous Underwater Vehicles in Current Fields with Complex Spatial Variability: an A* Approach", Proceedings of the 2005 IEEE, International Conference on Robotics and Automation Barcelona, Spain, April 2005 [18] M. Soulignac, P. Taillibert, M. Rueher, "Time-minimal Path Planning in Dynamic Current Fields", 2009 IEEE International Conference on Robotics and Automation, Kobe International Conference Center, Kobe, Japan, May 12-17, 2009 [19] Juan Carlos Rubio, Sean Kragelund, "The Transpacific Crossing: Long Range Adaptive Path Planning for UAVS Through Variable Wind Fields", Digital Avionics Systems Conference, 2003. DASC '03. The 22nd Publication Date: 12-16 Oct. 2003 Volume: 2, On page(s): 8.B.4- 81-12 vol [20] Hiroshi Kawano,, "Method for Designating the Wind Condition in MDP-based Motion Planning of Under-actuated Blimp type UAV", 2007 IEEE International Conference on Robotics and Automation Roma, Italy, 10-14 April 2007. [21] C. Connolly, R. Weiss, J. Burns, “Path Planning Using Laplace Equation”, IEEE Int. Conf. Robotics Automat., Cincinnati, OH, May 13-18, 1990, pp. 2102-2106. [22] J. Decuyper, D. Keymeulen, “A Reactive Robot Navigation System Based on a Fluid Dynamics Metaphor”, in Proc. Parallel Problem Solving from Nature First Workshop, H. Schwefel, R. Hartmanis Eds, Dortmund-Germany, Oct 1-3, 1990, pp. 356-362. [23] S. Akashita, S. Kawamura, K. Hayashi, “New Navigation Function Utilizing Hydrodynamic Potential for Mobile Robots”, IEEE Int. Workshop Intelli. Motion Contr., Istanbul, Turkey, Aug 20-22, 1990, pp. 413-417. [24] C. Connolly, “Harmonic Functions and Collision Probabilities”, The International Journal of Robotics Research, Vol. 16, No.4, pp. 497-507, August 1997. [25] J. Rossell, P. Iniguez, “Path Planning Using Harmonic Functions and Probabilistic Cell Decomposition”, Universitat Politechnica De Catalunya, IOC-DT-P-2004-16, Octubre, 2004. [26] A. Masoud, "An Informationally-Open, Organizationally-Closed Control Structure for Navigating a Robot in an Unknown, Stationary Environment" 2003 IEEE International Symposium on Intelligent Control, October 5-8, 2003, Houston, Texas, USA, pp. 614-619. [27] S. Masoud A. Masoud, "Constrained Motion Control Using Vector Potential Fields", The IEEE Transactions on Systems, Man, and Cybernetics, Part A: Systems and Humans. May 2000, Vol. 30, No.3, pp.251-272. [28] Ahmad A. Masoud, “A Hybrid PDE-ODE Controller for Intercepting an Intelligent, Well-informed Target in a Stationary, Cluttered Environment”, 44th IEEE Conference on Decision and Control, Seville-Spain, December 12-15, 2005. [29] S. Masoud, A. Masoud, " Motion Planning in the Presence of Directional and Obstacle Avoidance Constraints Using Nonlinear Anisotropic, Harmonic Potential Fields: A Physical Metaphor", IEEE Transactions on Systems, Man, & Cybernetics, Part A: systems and humans, Vol 32, No. 6, November 2002, pp. 705-723. [30] Ahmad A. Masoud, "Decentralized, Self-organizing, Potential field-based Control for Individually-motivated, Mobile Agents in a Cluttered Environment: A Vector-Harmonic Potential Field Approach", IEEE Transactions on Systems, Man, & Cybernetics, Part A: systems and humans, Vol. 37, No. 3, pp. 372-390, May 2007. [31] R. Gupta, A. Masoud, M. Chow, "A Network based, Delay-tolerant, Integrated Navigation System for a differential drive UGV using Harmonic Potential Field", Proceedings of the 45th IEEE Conference on Decision & Control Manchester Grand Hyatt Hotel San Diego, CA, USA, December 13-15, 2006, pp. 1870-1875. [32] J. Jackson , "Classical Electrodynamics", Wiley, 1989 [33] D. Ingerman, J. Morrow, "On a Characterization of the Kernel of the Dirichlet-to-Neumann Map for a Planar Region”, SIAM Journal on Mathematical Analysis, Vol. 29 , No. 1, Jan. 1998, pp. 106 - 115. [34] J. Milnor, “Morse Theory”, Princeton University Press, 1963. [35] G. Strang, "Linear Algebra and Its Applications", Brooks Cole, 3 edition, 1988. [36] J. LaSalle, “Some Extensions of Lyapunov’s Second Method”, IRE Transactions on Circuit Theory, CT-7, No. 4, pp. 520-527, 1960. [37] Ahmad A. Masoud, "A Harmonic Potential Field Approach with a Probabilistic Space Descriptor for Planning in Non-divisible Environments.", 2009 IEEE International Conference on Robotics and Automation, May 12 - 17, 2009, Kobe, Japan, pp. 3774-3779. [38] K. Deergha Rao, M. N. S. Swamy, E. I. Plotkin, "GPS Navigation with Increased Immunity to Modeling Errors", IEEE Transactions on Aerospace and Electronic Systems Vol. 40, No. 1 January 2004, pp. 2-11 [39] A. G. Shem, T. A. Mazzuchi, "Addressing Uncertainty in UAV Navigation Decision-Making", IEEE Transactions on Aerospace and Electronic Systems Vol. 44, No. 1 January 2008, pp. 295-313 [40] P. Keen,(1978). Decision support systems: an organizational perspective. Reading, Mass., Addison-Wesley Pub. Co. ISBN 0-201-03667-3 [41] A. Stentz, “Optimal and Efficient Path Planning for Partially-Known Environments”, Proceedings 1994 IEEE Robotics and Automation, 1994, 8-13 May 1994, pp. 3310-3317. [42] A. Stentz, “Optimal and Efficient Path Planning for Unknown and Dynamic Environments”, technical report CMU-RI-TR-93-20, the robotics institute, Carnegie Mellon University, 1993. [43] H. Schroer, “The Minimum Resistance”, 2002, Mire book, Harvard. [44] S. Bergman, M. Schiffer, “Kernel Functions and Elliptic Differential Equations in Mathematical Physics”, Academic Press Inc, New York, 1953. [45] D. Koditschek, “Exact Robot Navigation by Means of Potential Functions: Some Topological Considerations”, IEEE International conference on Robotics and Automation, Raleigh, N.C. 1987, pp. 1-6.
3cs.SY
arXiv:1607.07650v1 [math.GR] 26 Jul 2016 The concept of self-similar automata over a changing alphabet and lamplighter groups generated by such automata Adam Worynaa a Silesian University of Technology, Institute of Mathematics, ul. Kaszubska 23, 44-100 Gliwice, Poland Abstract Generalizing the idea of self-similar groups defined by Mealy automata, we itroduce the notion of a self-similar automaton and a self-similar group over a changing alphabet. We show that every finitely generated residually-finite group is self-similar over an arbitrary unbounded changing alphabet. We construct some naturally defined self-similar automaton representations over an unbounded changing alphabet for any lamplighter group K ≀ Z with an arbitrary finitely generated (finite or infinite) abelian group K. Keywords: rooted tree, changing alphabet, time-varying automaton, group generated by an automaton, lamplighter group 1. Introduction and main results The handling of various languages and approaches for describing groups of automorphisms of spherically homogeneous rooted trees brought many significant results. For example, the algebraic language of the so-called ’tableaux’ with ’truncated’ polynomials over finite fields, introduced by L. Kaloujnine [8], turned out to be effective in studying these groups as iterated wreath products. V. I. Sushchanskyy [16] uses this language to construct the pioneering examples of infinite periodic 2-generated p-groups (p > 2) as well as to produce factorable subgroups of wreath products of groups [17, 18]. A. V. Rozhkov [12], by using the notion of the cortage and studying the so-called Aleshin type groups, first gave an example of a 2-generated periodic group containing elements of all possible finite orders (see also [13, 14]). A. Ershler [2, 3] uses a combinatorial language of a time-varying automaton to discover new results concerning the growth and the so-called Fölner functions of groups. In the present paper, by introducing the notion of a self-similar automaton over a changing alphabet and the group generated by such an automaton, Email address: [email protected] (Adam Woryna) Preprint we make an attempt to show that the concept of a self-similar group with its faithful action by automorphisms on a homogeneous rooted tree (also called a regular rooted tree) can be naturally extended to an arbitrary spherically homogeneous rooted tree. As we show in Theorem 1, our combinatorial approach allows to characterize the class of finitely generated residually finite groups. We also provide (Theorem 2) naturally-defined representations for a large class of lamplighter groups for which the realization by the standard notion of a Mealy type automaton is not known. 1.1. Mealy automata and self-similar groups Self-similar groups appear in many branches of mathematics, including operators algebra, dynamical systems, automata theory, combinatorics, ergodic theory, fractals, and others. The most interesting and beautiful examples of selfsimilar actions are defined on the tree X ∗ of finite words over a finite alphabet X by using the notion of a transducer (also called a Mealy automaton), which, by definition, is a finite set Q together with the transition function ϕ : Q × X → Q and an output function ψ : Q × X → X. The convenient way defining and presenting a Mealy automaton is to draw its diagram. For example, the diagram in Figure 1 determines a Mealy automaton over the alphabet X = {0, 1}, with five states a, b, c, d, e, with a transition function given by the oriented edges and with an output function given by labelling of the vertices by the maps X → X, where id denotes the identity map and σ is a transposition: 0 7→ 1, 1 7→ 0. We d b 1 id id 1 1 0 0 id c 0 s id 0,1 0,1 e a Figure 1: The automaton generating the Grigorchuk group easily verify from this diagram the following equalities defining the transition and output functions of this automaton: ϕ(a, x) = ϕ(b, 0) = ϕ(e, x) = e, x ∈ X, a, ϕ(b, 1) = c, ϕ(c, 0) = a, ϕ(c, 1) = ψ(a, x) = d, ϕ(d, 0) = e, ϕ(d, 1) = b, σ(x), ψ(b, x) = ψ(c, x) = ψ(d, x) = ψ(e, x) = x, x ∈ X. Each state q ∈ Q of a Mealy automaton A = (X, Q, ϕ, ψ) defines a transformation of the tree X ∗ . Namely, given a word x0 x1 . . . xk ∈ X ∗ the automaton A 2 acts on the first letter x0 by the map which labels the vertex corresponding to the state q in the diagram of A, changes its state to the state given by the end of the arrow going out of q and labelled by x0 ; being in a new state the automaton reads the next letter x1 , transforms it and moves to the next state according to the above rule; and continues in this way to transform the remaining letters of the word. The induced transformation of the set X ∗ preserves the empty word and the beginnings of words and, if A is invertible (i.e. all the maps labelling vertices in the diagram of A are permutations of the alphabet X), it defines an automorphism of the tree X ∗ . The group generated by automorphisms corresponding to all states of an invertible automaton A with the composition of mappings as a product is called the group generated by this automaton. According to [4] (Definition 1.1, p. 129) an abstract group G is called self-similar over an alphabet X if there is a Mealy automaton A = (X, Q, ϕ, ψ) which generates a group isomorphic with G. For instance, the group G generated by the automaton depicted in Figure 1 is a fameous Grigorchuk group constituting one of the most interesting example among groups generated by all the states of a Mealy automaton and, as it turned out, in the whole group theory. This is also the first nontrivial example of a self-similar group. Groups generated by the states of a Mealy automaton form a remarkable class of groups containing various important examples connected to many interesting topics. For more about Mealy automata, groups generated by them, self-similar groups and for many open problems around these concepts see the survey paper [6] or [11]. 1.2. Time-varying automata and self-similarity over a changing alphabet A time-varying automaton, which is a natural generalization of a Mealy type automaton, allows to work not only with a fixed finite alphabet X but also with an infinite sequence X = (Xi )i≥0 = (X0 , X1 , . . .) of such alphabets, which further will be called a changing alphabet. We say that the changing alphabet X is unbounded if the sequence of cardinalities (|Xi |)i≥0 (the so-called branching index of X) is unbounded. By definition, a time-varying automaton over a changing alphabet X = (Xi )i≥0 is a quadruple A = (X, Q, ϕ, ψ), where Q is a finite set of states, ϕ = (ϕi )i≥0 and ψ = (ψi )i≥0 are sequences of, respectively, transition functions and output functions of the form ϕi : Q × Xi → Q, ψi : Q × Xi → Xi . The automaton A is called invertible if x 7→ ψi (q, x) is an invertible mapping of the set Xi for all i ≥ 0 and q ∈ Q. For every k ≥ 0 we also define the k-th shift of the automaton A that is a time-varying automaton Ak = (X(k) , Q, ϕ(k) , ψ(k) ), 3 where X(k) = (Xk+i )i≥0 , ϕ(k) = (ϕk+i )i≥0 , ψ(k) = (ψk+i )i≥0 . In particular, A is a Mealy automaton if and only if Ak = A for every k ≥ 0. At first, in the literature of time-varying automata only the notion of an automaton over a fixed alphabet was considered and merely the structural properties of such automata were studied (see [10]). We first define in [20] the class of time-varying automata over a changing alphabet X = (Xi )i≥0 (see also [21, 22, 23]). Such an alphabet defines a tree X ∗ of finite words, which is an example of a spherically homogeneous rooted tree. The tree X ∗ is homogeneous if and only if the branching index of X is constant. Moreover, any locally finite spherically homogeneous rooted tree T is isomorphic with X ∗ for a suitable changing alphabet X = (Xi )i≥0 . We introduced in [20] the notion of a group generated by an automaton over a changing alphabet and show that this combinatorial language is apt to define and study groups of automorphisms of spherically homogeneous rooted trees which may be not homogeneous. Specifically, if A = (X, Q, ϕ, ψ) is an invertible time-varying automaton over a changing alphabet X = (Xi )i≥0 , then for every k ≥ 0 and each state q ∈ Q we define, via transition and output functions, an ∗ automorphism qk of the tree X(k) and call the group G(Ak ) = hqk : q ∈ Qi as a group generated by the k-th shift of A. Definition 1. An invertible time-varying automaton A = (X, Q, ϕ, ψ) is called self-similar if for every k ≥ 0 the mapping qk 7→ q0 (q ∈ Q) induces an isomorphism G(Ak ) ≃ G(A). If A = (X, Q, ϕ, ψ) is a Mealy automaton, then we have qk = q0 for all k ≥ 0 and q ∈ Q. Thus every Mealy automaton is self-similar by the above definition. Definition 2. An abstract group G is called a self-similar group over the changing alphabet X = (Xi )i≥0 if there is a self-similar automaton A = (X, Q, ϕ, ψ) which generates a group isomorphic with G, i.e. G(A) ≃ G. Every self-similar group G over a changing alphabet X is an example of a finitely generated residually finite group (a group is called residually finite if there is a descending chain N0 ≥ N1 ≥ N2 ≥ . . . of normal subgroups of finite indexes and the trivial intersection). Indeed, the group Aut(X ∗ ) of all automorphisms of the tree X ∗ is an example of a residually finite group (stabilizers of consecutive levels of this tree are normal subgroups of finite indexes which intersect trivially). Thus G, as a group isomorphic with a subgroup of Aut(X ∗ ), is residually finite. Given an abstract group G and a changing alphabet X = (Xi )i≥0 , it is natural to ask about the automaton complexity of the group G over the alphabet X, i.e. about the minimal number of states in an automaton over X which 4 generates a group isomorphic with G. Obviously, for any changing alphabet X the automaton complexity of any group G over X is not smaller than the rank r(G) of this group, i.e. than the minimal cardinality of a generating set of G. Definition 3. If the number of states in an automaton A = (X, Q, ϕ, ψ) coincides with the rank of the group G(A), i.e. |Q| = r(G(A)), then the automaton A is called optimal. 1.3. The class of finitely generated self-similar groups over an unbounded changing alphabet In the theory of groups generated by Mealy automata the problem of classifying all groups generated by automata over a given alphabet X and with a given number of states is essential and intensively studied. For example, by the result from [6] there are only six groups (up to isomorphism) generated by 2-state Mealy automata over the binary alphabet, including the simplest example of an interesting group generated by a Mealy automaton, that is the lamplighter group C2 ≀ Z, where C2 denotes a cyclic group of order 2 and Z is the infinite cyclic group. Much effort put towards a classification of groups 1 0 s id 1 0 Figure 2: A Mealy automaton generating the lamplighter group C2 ≀ Z generated by 3-state Mealy automata over the binary alphabet. In the paper [1] was shown, among other things, that there are no more than 124 pairwise nonisomorphic groups defined by such automata. Only recently it was discovered that a free-nonabelian group of finite rank is self-similar. Namely, M. Vorobets and Ya. Vorobets showed in [19] that a Mealy automaton (the so-called Aleshin type automaton) depicted in Figure 3 generates a free group of rank 3. Moreover, it is still an open question whether or not there exists an optimal Mealy automaton which generates a free group of rank 2. Referring to this problem, we provided [23] for this group an explicit and naturally defined realization by an optimal automaton over a changing alphabet; see also Example 1. The first result of the present paper shows that the situation completely changes if we try to characterize all self-similar groups over an arbitrary unbounded changing alphabet. Namely, we show the following Theorem 1. Every finitely generated residually finite group G is self-similar over any unbounded changing alphabet and the corresponding self-similar automaton is optimal. 5 id 0,1 0 1 1 s s 0 Figure 3: A Mealy automaton generating a free group As we will see in the proof, the construction leading to Theorem 1 is based on the specific type of automaton which we call an automaton of a diagonal type. Definition 4. We say that a time-varying automaton A = (X, Q, ϕ, ψ) is of a diagonal type, or that A has transition functions defined in the diagonal way if ϕi (q, x) = q for all i ≥ 0, q ∈ Q and x ∈ Xi . Note that the proof of Theorem 1 is unconstructive and we can not explicitly define the output functions of the corresponding self-similar automaton for a given group G and a changing alphabet X = (Xi )i≥0 . Indeed, in that proof we choose for the construction of an automaton A which generates the group G an arbitrary generating set S as a set of states and we define the transition functions in the diagonal way. Next, we show that the output functions of A can be induced by some embedding g 7→ (g (0) , g (1) , . . .) of the group G into the infinite cartesian product S(X0 ) × S(X1 ) × . . . of symmetric groups on sets of letters; due to this embedding we define the output functions as follows: ψi (s, x) = s(i) (x) for all i ≥ 0, s ∈ S and x ∈ Xi . However, the above embedding arises simply from the fact that G embeds into the cartesian product G/N0 × G/N1 × . . ., where G = N0 ⊲ N1 ⊲ N2 ⊲ . . . is an infinite normal series with finite quotients G/Ni (i ≥ 0) and the trivial intersection. But the existence of such a normal series is equivalent to the assumption that G is residually finite. Hence, given a residually finite group G, this method does not tell us how to compute effectively the permutations s(i) ∈ S(Xi ). In particular, we know nothing about the cycle structure of these permutations. Moreover, in the case of a free group of rank 2 we tried to find in the paper [21] an explicit representation by an optimal automaton of a diagonal type. However, our computations led to a pretty complicated description of the mappings s(i) and, consequently, of output functions in such an automaton. In view of the above disadvantages as well as the previous results, it would be interesting to find and study explicit and naturally defined representations of a given group G by self-similar automata over a changing alphabet, especially if such representations are unknown by using the notion of a Mealy automaton. 6 1.4. Lamplighter groups as self-similar groups over a changing alphabet Let K be an arbitrary nontrivial finitely generated (finite or infinite) abelian group. By definition, the lamplighter group K ≀ Z is a semidirect product M K ⋊Z Z L with the action of the infinite cyclic group on the direct sum Z K by a shift. For the rank r(K ≀ Z) of this group we have: r(K ≀ Z) = r(K × Z) = r(K) + 1. In recent years groups of this form are intensively studied via their selfsimilar actions. For example, the automaton realization of the lamplighter group C2 ≀ Z from the paper [7] was used by R. Grigorchuk and A. Żuk to compute the spectral measure associated on random walks on this group, which leads to a counterexample to the strong Atiyach conjecture [5]. This construction was generalized [9, 15] by P. Silva and B. Steinberg in the concept of a reset Mealy automaton, where for every finite abelian group K the corresponding lamplighter group was realized as a group generated by a Mealy automaton in which the set of states and the alphabet coincides with K. Hence, the only known optimal realization of the group K ≀ Z by an automaton concerns the simplest case K = C2 . Here, it should also be mentioned that the question whether or not there exists a representation of a lamplighter group K ≀ Z with an infinite K by a Mealy automaton is still open. For the second result of this paper we come from a naturally defined and self-similar automaton D of a diagonal type over a changing alphabet X that generates the direct product K × Z (see Example 3). We discover that only by a simple manoeuvre on each transition function in D we can pass to a new automaton A and obtain in this way a self-similar optimal automaton realization for the lamplighter group K ≀ Z. Namely, we distinguish the state a in the automaton D and the letters xq,i ∈ Xi (i ≥ 0) for each state q 6= a and we define A as an automaton which operates like D, unless being in a state q 6= a it receives the letter xq,i in a moment i ≥ 0, then it moves to the state a. To define the changing alphabet X = (Xi )i≥0 and output functions in the automaton D we decompose the group K into a direct product of cyclic groups K = Zn1 × (Cr1 × . . . × Crn2 ), n1 , n2 ≥ 0 such that n1 + n2 = r(K). Let us denote n := r(K) = n1 + n2 . For every i ≥ 0 we consider arbitrary 2n pairwise disjoint cycles π1,i , . . . , πn,i , σ1,i , . . . , σn,i in the symmetric group on a finite set Xi which satisfy the following conditions: for each 1 ≤ s ≤ n the cycles πs,i (i ≥ 0) do not have a uniformly bounded length, as well as the cycles σs,i (i ≥ 0) for each 1 ≤ s ≤ n1 , and the length of each cycle σs,i (n1 < s ≤ n, i ≥ 0) does not depend on i and is equal to rs−n1 . In particular, we see that the arising changing alphabet X = (Xi )i≥0 is unbounded. 7 Now, taking n symbols b1 , . . . , bn different from a, we define the automaton D as an automaton of a diagonal type over the changing alphabet X = (Xi )i≥0 , with the set of states Q = {a, b1 , . . . , bn } and the output functions defined as follows:  αi (x), q = a, ψi (q, x) = βs,i (x), q = bs , 1 ≤ s ≤ n for all i ≥ 0, q ∈ Q and x ∈ Xi , where the permutations αi , βs,i are defined in the following way αi = π1,i · . . . · πn,i , βs,i = σs,i · αi . As we mention above, passing to the automaton A just needs to redefine transition functions of D. Explicitly, we define: A = (X, Q, ϕA , ψ), where  a, q = bs , x = xs,i , 1 ≤ s ≤ n, A ϕi (q, x) = q, otherwise, and xs,i := xbs ,i is a letter from the cycle πs,i . Our main result is the following Theorem 2. The automaton A is self-similar and the group G(A) is isomorphic with the lamplighter group K ≀ Z. In the proof of Theorem 2 we do not use the algebraic language of embedding into wreath products of groups (the so-called wreath recursion), which is common in studying self-similar groups. Instead of this, for any i ≥ 0 we di∗ of rectly provide some recursive formulae describing the action on the tree X(i) ∗ ∗ , where → X(i) automaton transformations ai , cs,i : X(i) cs,i = (bs )i · a−1 i , 1 ≤ s ≤ n. We show (Proposition 3 and Lemma 1) that for every i ≥ 0 any two transformations cs,i and cs′ ,i (1 ≤ s, s′ ≤ n) commute and the mapping cs,i 7→ κs induces an isomorphism between the group Ki = hci,1 , . . . , ci,n i and the group K, where the elements κs ∈ K (1 ≤ s ≤ n) form the standard generating set of K, that is κs = (0, . . . , 0, us , 0, . . . , 0), where us occurs in the s-th position and represents a generator of the corresponding cyclic factor of K. (k) Next, for each k ∈ Z we consider the conjugation Ki = ai−k Ki aki as well as the group (k) Hi = hKi : k ∈ Zi generated by all these conjugations, and we show (Proposition 4) that Hi is a (k) direct sum of its subgroups Ki , k ∈ Z, which gives the isomorphism Hi ≃ L Z K. Further, we show (Lemma 4) that the group Hi intersects trivially 8 with the cyclic group generated by ai and that this cyclic group acts on Hi by conjugation via the shift (Lemma 3). This implies that the group Gi = hHi , ai i = hc1,i , . . . , cn,i , ai i is a semidirect product of the subgroups Hi and hai i, and, consequently, that Gi is isomorphic with the lamplighter group K ≀ Z via the mapping ai 7→ u, cs,i 7→ ηs , 1 ≤ s ≤ n, where {u, η1 , . . . , ηn } is theLstandard generating set of K ≀ Z, that is u is a generator of Z in the product Z K ⋊ Z and M ηs = (. . . , 0, 0, 0, κs , 0, 0, 0, . . .) ∈ K, Z where each κs on the right hand side occurs in zero position. Finally, we obtain that the group G(Ai ) = hai , (b1 )i , . . . , (bn )i i generated by the i-th shift of the automaton A is isomorphic with Gi and the mapping ai 7→ u, (bs )i 7→ ηs · u induces the isomorphism G(Ai ) ≃ K ≀ Z. As we will see in the proof, all calculations verifying the above propositions and lemmas are entirely elementary. However, they are burdened with some technical formulae and the presentation of the proof seems to be quite formalistic. This contrasts with a transparent and intuitive idea behind the definition of the automaton A. On the other hand, our construction confirms one of the most intriguing phenomenon in the computational group theory which originates from dealing with the simplest examples of transducers, i.e. although they can generate various classic and important groups with rare and exotic properties, the derivation of such a group from the corresponding automaton is far from obvious and, quite often, requires sophisticated and complex approaches. It can be seen, for example, in the original proof from [7] finding the lamplighter group C2 ≀ Z as a self-similar group or in a long history discovering a free group behind the Aleshin type automaton. We hope that our construction will not only fill the gap in self-similar automaton representations for a large class of lamplighter groups, but also can serve as a useful tool in the further study of these groups, in particular, via their dynamics and geometry on the corresponding tree. We also think that it will help to better understand the connection between the structure of an automaton over a changing alphabet and the group it defines. 2. The tree X ∗ and its automaton transformations In the Introduction, we have already presented the notion of a changing alphabet X = (Xi )i≥0 , a time-varying automaton A = (X, Q, ϕ, ψ), their shifts X(k) , Ak (k ≥ 0) and a self-similar group over a changing alphabet. Now, we only recall the necessary notions and facts concerning the tree X ∗ of finite words ∗ ∗ (q ∈ Q, k ≥ 0). We over X and automaton transformations qk : X(k) → X(k) ∗ define the tree X as a disjoint sum of cartesian products X0 × X1 × . . . × Xi , i ≥ 0, 9 together with the empty word ǫ. Since the elements of a word w ∈ X ∗ are called letters, we will not separate them by commas and we will write w = x0 x1 . . . xk , where xi ∈ Xi are the letters. Thus for a given i ≥ 0 the above cartesian product consists of all words w ∈ X ∗ with the length |w| = i + 1. If w ∈ X ∗ and ∗ v ∈ X(|w|) , then wv denotes a concatenation of w and v. Obviously wv ∈ X ∗ . The set X ∗ has the structure of a spherically homogeneous rooted tree. The empty word is the root of this tree and two words are connected if and only if they are of the form w and wx for some w ∈ X ∗ and x ∈ X|w| . In particular, the i-th level X i (i ≥ 0) of the tree X ∗ (that is the set of vertices at the distance i from the root) consists of all words of the length i. There is a natural interpretation of an automaton A = (X, Q, ϕ, ψ) over a changing alphabet X = (Xi )i≥0 as a machine, which being in a moment i ≥ 0 in a state q ∈ Q reads a letter x ∈ Xi from the input tape, types the letter ψi (q, x) on the output tape, moves to the state ϕi (q, x) and proceeds further to the moment i + 1. This allows to define for every k ≥ 0 and q ∈ Q the ∗ ∗ ∗ transformation qk : X(k) → X(k) recursively: qk (ǫ) = ǫ and if xw ∈ X(k) with ∗ x ∈ Xk , w ∈ X(k+1) , then ′ (w), qk (xw) = x′ qk+1 (1) where x′ = ψk (q, x) and q ′ = ϕk (q, x). The transformation qk is called an ∗ automaton transformations of the tree X(k) (corresponding to the state q in the k-th transition of A). It is convenient to present an automaton A = (X, Q, ϕ, ψ) over a changing alphabet X = (Xi )i≥0 as a labeled directed locally finite graph with the set {(i, q) : i ≥ 0, q ∈ Q} as a set of vertices. Two vertices are connected with an arrow if and only if they are of the form (i, q) and (i + 1, ϕi (q, x)) for some i ≥ 0, q ∈ Q and x ∈ Xi . This arrow is labeled by x, starts from the vertex (i, q) and goes to the vertex (i + 1, ϕi (q, x)). Each vertex (i, q) is labeled by the mapping τi,q : Xi → Xi , τi,q : x 7→ ψi (q, x) (the so-called labelling of the state q in the i-th transition of A). Further, to make the graph of the automaton clear, we will substitute a large number of arrows connecting two given vertices and having the same direction for a one multi-arrow labeled by suitable letters and if the labelling of such a multi-arrow starting from a given vertex follows from the labelling of other arrows starting from this vertex, we will omit this labelling. Directly by the formula (1) we see that in case A invertible (i.e. each τi,q is a permutation of the set Xi ), every ∗ ∗ automaton transformation qi : X(i) → X(i) is also invertible. By the formula (1) we also see that qi preserves the lengths of words and common beginnings ∗ and hence it defines an endomorphisms of the tree X(i) . Moreover, the composition of two automaton transformations of the same tree is also an automaton transformation of this tree and the inverse mapping to the invertible automaton 10 x1,0 a0 b1,2 b1,1 b1,0 x1,1 b1,3 x1,2 a2 a1 a3 Figure 4: The graph of the automaton A generating K ≀ Z, case n = 1 transformation is also of this type (we can show this by using an operation of the composition of two time-varying automata over the same changing alphabet and the notion of an inverse automaton - see [20]). In particular, the set of all invertible automaton transformations of a tree X ∗ together with the composition of mappings as a product forms a proper subgroup of the group Aut(X ∗ ) of automorphisms of this tree, which is denoted by F A(X ∗ ) and called the group of finite-state automorphisms. Remark 1. Through the rest of the text we use the right action convention in writing the composition f g of two transformation f, g : X ∗ → X ∗ , i.e. the transformation applied first is written on the left and for any word w ∈ X ∗ we have the equality f g(w) = g(f (w)). 3. The proof of Theorem 1 Since the group G is countable and residually finite, there is an infinite normal series G = N0 ⊲ N1 ⊲ . . . with finite quotients G/Ni and the trivial intersection. Then the function g 7→ (gNi )i≥0 is an embedding of G into the infinite cartesian product G/N0 × G/N1 × . . . . (2) Let X = (Xi )i≥0 be an unbounded changing alphabet. Then there is an infinite increasing sequence 0 < t0 < t1 < . . . of integers such that for every i ≥ 0 the quotient G/Ni embeds into the product S(Xti−1 +1 ) × S(Xti−1 +2 ) × . . . × S(Xti ) of symmetric groups of consecutive sets of letters, where for i = 0 we take t(−1) = −1. Consequently, there is an embedding g 7→ (g (0) , g (1) , . . .), g ∈ G 11 of G into the cartesian product S(X0 ) × S(X1 ) × . . .. The above embedding ∗ ∗ allows to define the transformations ξg,i : X(i) → X(i) (i ≥ 0, g ∈ G) as follows: ∗ ξg,i (ǫ) = ǫ and if w = x0 x1 . . . xk ∈ X(i) , then ξg,i (w) = g (i) (x0 )g (i+1) (x1 ) . . . g (i+k) (xk ). In particular, for all g, g ′ ∈ G and i ≥ 0 we have ξgg′ ,i = ξg,i · ξg′ ,i , ξg−1 ,i = (ξg,i )−1 . (3) Let S be any generating set of G whose cardinality coincides with the rank of G. Let us define a time-varying automaton A = (X, S, ϕ, ψ) as follows: ϕi (s, x) = s, ψi (s, x) = s(i) (x) for all i ≥ 0, s ∈ S, x ∈ Xi . Let us fix i ≥ 0. Directly by the definition of A the transformation si defined by a state s ∈ S in the i-th transition of A acts ∗ on any word w = x0 x1 . . . xk ∈ X(i) in the following way si (w) = s(i) (x0 )s(i+1) (x1 ) . . . s(i+k) (xk ). In particular, we have si = ξs,i for every s ∈ S and, since the elements si for s ∈ S generate the group G(Ai ), we have G(Ai ) = {ξg,i : g ∈ G}. By the equalities (3) we see that the mapping G → G(Ai ), g 7→ ξg,i T is an epimorphism. Further, since k≥0 Nk = {1G}, for every g 6= 1G there are infinitely many k ≥ 0 such that gNk 6= Nk . Consequently g (k) 6= IdXk for ∗ for every g 6= 1G . Hence the infinitely many k ≥ 0, which implies ξg,i 6= IdX(i) above epimorphism is one-to-one. Consequently, the mapping s 7→ si (s ∈ S) induces an isomorphism G ≃ G(Ai ), i.e. the automaton A is an optimal selfsimilar realization of G. As we explained in the introduction, in spite of Theorem 1, it is not easy to find an explicit self-similar automaton realization for a given group G. Even if we are able to construct an automaton defining G, this automaton may not be necessarily self-similar. Example 1. Let (ri )i≥0 be a nondecreasing unbounded sequence of integers ri > 1 and let X = (Xi )i≥0 be a changing alphabet in which Xi = {1, 2, . . . , ri }. For each i ≥ 0 we define two cyclic permutations αi , βi of the set Xi : αi = (1, 2), βi = (1, 2 . . . , ri ). Let A = (X, {a, b}, ϕ, ψ) be an automaton in which the transition and output functions are defined as follows    q, x 6= 1, αi (x), q = a, a, x = 1, q = b, ϕi (q, x) = ψi (q, x) = βi (x), q = b.  b, x = 1, q = a, 12 b0 b1 1 a0 1 b2 a1 b3 1 1 1 a2 1 a3 Figure 5: A self-similar automaton from Example 1 According to [23] the group G(A) = ha0 , b0 i is a free nonabelian group of rank two generated freely by the transformations a0 , b0 . Directly by the construction of the automaton A we see that for every i ≥ 0 the automaton transformations ∗ ∗ ai , bi : X(i) → X(i) also generate a free nonabelian group of rank 2. In particular, the mapping ai 7→ a0 , bi 7→ b0 induces an isomorphism G(Ai ) ≃ G(A). Thus the automaton A is a self-similar realization of a free nonabelian group of rank 2 over the changing alphabet X. Example 2. Let the changing alphabet X = (Xi )i≥0 and the permutations αi , βi ∈ S(Xi ) be defined as in the previous example. Let A = (X, {a, b}, ϕ, ψ) be an automaton of a diagonal type with the output functions defined as follows: ψi (a, x) = αi (x) and ψi (b, x) = βi (x) for all i ≥ 0 and x ∈ Xi . The automaton b0 b1 b2 b3 a0 a1 a2 a3 Figure 6: An automaton from Example 2 A is not a self-similar realization of the group G = G(A) because A is not selfsimilar. Indeed, let t ≥ 0 be the smallest number such that rt ≥ 3 and t′ > t be the smallest number such that rt′ > rt . Then one can verify by straightforward calculations that W (at′ , bt′ ) 6= IdX(t∗ ′ ) and W (a0 , b0 ) = IdX ∗ , where the groupword W = W (a, b) is defined as follows W (a, b) = (abrt −1 ab−rt +1 )2 . Thus the mapping at′ 7→ a0 , bt′ 7→ b0 can not be extended to the isomorphism of groups G(At′ ) and G(A). Example 3. Let D be an automaton of a diagonal type defined in the introduction. For every i ≥ 0 let ai , (bs )i (1 ≤ s ≤ n) be automaton transfor∗ mations of the tree X(i) corresponding to the states of D in i-th transition. Obviously, the automaton D is invertible. By definition of transition and output functions in D we obtain the following recursions: ai (xw) = αi (x)ai+1 (w), 13 ∗ (bs )i (xw) = βs,i (x)bi+1 (w) for every x ∈ Xi and w ∈ X(i+1) . The group G(Di ) generated by the i-th shift of the automaton D is generated by the transformations ai and cs,i = (bs )i a−1 for 1 ≤ s ≤ n. Since σs,i = βs,i α−1 = α−1 i i i βs,i , we have by the above recursions: cs,i (xw) = σs,i (x)cs,i+1 (w) for x ∈ Xi and ∗ w ∈ X(i+1) . We consider in the infinite cartesian product S(X0 ) × S(X1 ) × . . . the subgroup hλi , γ1,i , . . . , γn,i i, where λi = (αi , αi+1 , . . .), γs,i = (σs,i , σs,i+1 , . . .). By the above recursions for ai and cs,i we see that the mapping ai 7→ λi , cs,i 7→ γs,i , 1 ≤ s ≤ n induces an isomorphism G(Di ) ≃ hλi , γ1,i , . . . , γn,i i. For each i ≥ 0 the permutations αi , σ1,i , . . . , σn,i are pairwise disjoint. Consequently, the mn 0 m1 group hλi , γ1,i , . . . , γn,i i is abelian and the product λm i γ1,i . . . γn,i is trivial if and only if o(λi ) | m0 and o(γs,i ) | ms for 1 ≤ s ≤ n, where o(g) denotes the order of an element g. Since αi = π1,i · . . . · πn,i , we see by the assumptions on the permutations πs,i and σs,i that o(λi ) = o(γ1,i ) = . . . = o(γn1 ,i ) = ∞ and o(γs,i ) = rs−n1 for n1 < s ≤ n. This implies that the automaton D is self-similar and it generates the group isomorphic with K × Z. 4. The proof of Theorem 2 Let A be an automaton defined in the introduction and for every i ≥ 0 let ai , ∗ (bs )i (1 ≤ s ≤ n) be automaton transformations of the tree X(i) corresponding to the states of A in its i-th transition. Obviously, the automaton A is invertible. Directly by formulae defining transition and output functions of A we obtain the following recursions:  βs,i (x)ai+1 (w), x = xs,i , ai (xw) = αi (x)ai+1 (w), (bs )i (xw) = βs,i (x)(bs )i+1 (w), x 6= xs,i −1 −1 ∗ for every x ∈ Xi and w ∈ X(i+1) . Hence a−1 i (xw) = αi (x)ai+1 (w) and the −1 transformations cs,i = (bs )i ai satisfy: cs,i (xw) = (bs )i a−1 (xw) = a−1 i ((bs )i (xw)) =  −1i ai (βs,i (x)ai+1 (w)), x = xs,i , = = a−1 (β (x)(b ) (w)), x 6= xs,i , s,i s i+1 i  −1 −1 αi (βs,i (x))ai+1 (ai+1 (w)), x = xs,i , = = −1 α−1 i (βs,i (x))ai+1 ((bs )i+1 (w)), x 6= xs,i ,  σs,i (x)w, x = xs,i , = σs,i (x)cs,i+1 (w), x 6= xs,i . But, if x = xs,i , then σs,i (x) = x as the cycles πs,i and σs,i are disjoint. In consequence, we have  xw, x = xs,i , cs,i (xw) = (4) σs,i (x)cs,i+1 (w), x 6= xs,i . 14 4.1. The groups Ki We consider the groups Ki = hc1,i , . . . , cn,i i, i ≥ 0. Proposition 3. For every i ≥ 0 the mapping cs,i 7→ κs (1 ≤ s ≤ n) induces an isomorphism Ki ≃ K. Proof. At first we show that for s 6= s′ the transformations cs,i and cs′ ,i commute. By (4) we obtain: cs,i cs′ ,i (xw) = cs′ ,i (cs,i (xw)) =  cs′ ,i (xw), = cs′ ,i (σs,i (x)cs,i+1 (w)), x = xs,i , x 6= xs,i . (5) If x = xs,i , then x 6= xs′ ,i and σs′ ,i (x) = x. Thus by (4) and (5) we have in this case cs,i cs′ ,i (xw) = cs′ ,i (xw) = σs′ ,i (x)cs′ ,i+1 (w) = xcs′ ,i+1 (w). Similarly, if x = xs′ ,i , then σs,i (x) = x and by (4) and (5) we obtain in this case cs,i cs′ ,i (xw) = cs′ ,i (σs,i (x)cs,i+1 (w)) = cs′ ,i (xcs,i+1 (w)) = xcs,i+1 (w). If x 6= xs,i and x 6= xs′ ,i , then σs,i (x) 6= xs′ ,i and consequently cs,i cs′ ,i (xw) = cs′ ,i (σs,i (x)cs,i+1 (w)) = σs,i σs′ ,i (x)cs,i+1 cs′ ,i+1 (w). We conclude   xcs′ ,i+1 (w), xcs,i+1 (w), cs,i cs′ ,i (xw) =  σs,i σs′ ,i (x)cs,i+1 cs′ ,i+1 (w), x = xs,i , x = xs′ ,i , otherwise.   xcs′ ,i+1 (w), ′ xcs,i+1 (w), cs ,i cs,i (xw) =  ′ σs ,i σs,i (x)cs′ ,i+1 cs,i+1 (w), x = xs,i , x = xs′ ,i , otherwise. (6) By analogy we have Thus we may write cs,i cs′ ,i (xw) =  cs′ ,i cs,i (xw), σs,i σs′ ,i (x)cs,i+1 cs′ ,i+1 (w), x ∈ {xs,i , xs′ ,i }, otherwise. (7) Since the cycles σs,i and σs′ ,i commute, we see by (7) that the transformations cs,i and cs′ ,i commute as well. −1 If x 6= xs,i , then σs,i (x) 6= xs,i and by the recursion (4) we have: c−1 s,i (xw) =  xw, −1 (x)c−1 σs,i s,i+1 (w), 15 x = xs,i , x 6= xs,i By easy inductive argument we obtain for any integer k:  xw, x = xs,i , cks,i (xw) = k k σs,i (x)cs,i+1 (w), x 6= xs,i . (8) By our assumption, for each 1 ≤ s ≤ n1 the cycles σs,i (i ≥ 0) do not have a uniformly bounded length. Hence, by the recursion (8) we see that the transformation cs,i is of infinite order for every 1 ≤ s ≤ n1 . Similarly, if n1 < s ≤ n, then the length of σs,i is equal to rs−n1 and consequently the order of cs,i is equal to rs−n1 . We define the subset V ⊆ Zn as follows: V = {(m1 , . . . , mn ) : 0 ≤ ms < rs−n1 for n1 < s ≤ n}. By above, we obtain the equality Ki = {CM,i : M ∈ V}, where for every M = (m1 , . . . , mn ) ∈ V we define mn 1 CM,i = cm 1,i · . . . · cn,i . We show that the above presentation is unique. To this end to each CM,i we associate the permutation ΣM,i ∈ Sym(Xi ) defined as follows: m1 · . . . · σnm,ni . ΣM,i = σ1,i In the following lemma, which is a generalization of the recursion (8), for every M ∈ V we denote by M (s) (1 ≤ s ≤ n) the element arising from M by the substitution of ms for 0. Lemma 1. We have CM,i (xw) =  xCM (s) ,i+1 (w), ΣM,i (x)CM,i+1 (w), x = xs,i , otherwise (9) ∗ for all x ∈ Xi , w ∈ X(i+1) and M ∈ V. Proof. (of Lemma 1) Let M = (m1 , . . . , mn ) and for every 1 ≤ s ≤ n let Ms ∈ V arises from M by the substitution of every component mi with i > s for 0. Since Mn = M and for every 1 ≤ s ≤ n we have ms 1 CMs ,i = cm 1,i · . . . · cs,i , it is sufficient to show that for every 1 ≤ s ≤ n the following recursion holds:  xCM (s′ ) ,i+1 (w), x = xs′ ,i , 1 ≤ s′ ≤ s, s CMs ,i (xw) = (10) ΣMs ,i (x)CMs ,i+1 (w), otherwise. 16 1 For s = 1 we have CMs ,i = CM1 ,i = cm 1,i and (10) coincides with the recursion (8). Let us fix s > 1 and let us assume that (10) is true for s − 1. Then we have CMs ,i (xw) s = cm s,i (CMs−1 ,i (xw)) = ( m cs,is (xCM (s′ ) ,i+1 (w)), s−1 = s cm s,i (ΣMs−1 ,i (x)CMs−1 ,i+1 (w)), x = xs′ ,i , 1 ≤ s′ < s, otherwise. ms But, if x = xs′ ,i for some 1 ≤ s′ < s, then σs,i (x) = x. Thus by the recursion (8) we have in this case ms ms s cm s,i (xCM (s′ ) ,i+1 (w)) = σs,i (x)cs,i+1 (CM (s′ ) ,i+1 (w)) = xCM (s′ ) ,i+1 (w). s−1 s−1 s m m1 s−1 (x) = x. Thus by the recursion (8) ·. . .·σs−1,i If x = xs,i , then ΣMs−1 ,i (x) = σ1,i we have in this case CMs ,i (xw) s = cm s,i (ΣMs−1 ,i (x)CMs−1 ,i+1 (w)) = s = cm s,i (xCMs−1 ,i+1 (w)) = = xCMs−1 ,i+1 (w) = = xCM (s) ,i+1 (w). s If x ∈ / {x1i , . . . , xs,i }, then ΣMs−1 ,i (x) 6= ΣMs−1 ,i (xs,i ) = xs,i . Thus by the recursion (8) we have in this case CMs ,i (xw) s = cm s,i (ΣMs−1 ,i (x)CMs−1 ,i+1 (w)) = ms s (ΣMs−1 ,i (x))cm = σs,i s,i+1 (CMs−1 ,i+1 (w)) = = ΣMs ,i (x)CMs ,i+1 (w). To sum up, we conclude the recursion (10), which finishes the proof of the lemma. mn 1 Now, let us assume that the product CM,i = cm 1,i · . . . · cn,i represents the identity mapping for some M = (m1 , . . . , mn ) ∈ V. Let n1 < s ≤ n. Since there is x0 ∈ Xi which belongs to the cycle σs,i and x0 ∈ / {x1,i , . . . , xn,i }, we ms (x0 ). Thus the divisibility have by Lemma 1: x0 = CM,i (x0 ) = ΣM,i (x0 ) = σs,i rs−n1 | ms holds and consequently ms = 0. If 1 ≤ s ≤ n1 , then there is i0 > i such that the length of the cycle σs,i0 is greater than |ms |. Let w = ∗ x0 x1 . . . xi0 −i ∈ X(i) be a word such that the last letter xi0 −i belongs to the / {x1,j+i , . . . , xn,j+i } for every 0 ≤ j ≤ i − i0 . By Lemma 1 cycle σs,i0 and xj ∈ we obtain CM,i (w) = ΣM,i (x0 )ΣM,i+1 (x1 ) . . . ΣM,i0 (xi0 −i ). ms (xi0 −i ). Consequently Since CM,i (w) = w, we have xi0 −i = ΣM,i0 (xi0 −i ) = σs,i 0 ms = 0, which finishes the proof of Proposition 3. 17 4.2. The groups Hi We consider the transformations dk,s,i = a−k · cs,i · aki , i ≥ 0, k ∈ Z, i 1 ≤ s ≤ n. For every i ≥ 0 we define the group Hi = hdk,s,i : k ∈ Z, 1 ≤ s ≤ ni. (k) For every i ≥ 0 and k ∈ Z we also define the subgroup Ki (k) Ki ≤ Hi , where −k k k = a−k i Ki ai = {ai CM,i ai : M ∈ V}. (k) By Proposition 3 we have the isomorphism Ki ≃ Ki ≃ K. (k) Proposition 4. The group Hi is a direct sum of its subgroups Ki and the mapping dk,s,i 7→ (. . . , 0, 0, κs , 0, 0, . . .), k ∈ Z, for k ∈ Z 1 ≤ s ≤ n, where the generator κs ∈ LK on the right hand side is in the k-th position, induces the isomorphism Hi ≃ Z K. Proof. By the recursion for ai and by (4) we easily obtain:  xw, x = αki (xs,i ), dk,s,i (xw) = σs,i (x)dk,s,i+1 (w), x 6= αki (xs,i ) ∗ . Further, by the recursion for all x ∈ Xi and w ∈ X(i+1) ′ obtain for all k, k ∈ Z, 1 ≤ s, s′ ≤ n, x ∈ Xi and w ∈ recursion:   xdk′ ,s′ ,i+1 (w), ′ ′ dk,s,i dk ,s ,i (xw) = xdk,s,i+1 (w),  σs,i σs′ ,i (x)dk,s,i+1 dk′ ,s′ ,i+1 (w), for ai and by (6) we ∗ X(i+1) the following x = αki (xs,i ), ′ x = αki (xs′ ,i ), otherwise, and hence dk,s,i dk′ ,s′ ,i (xw) =  dk′ ,s′ ,i dk,s,i (xw), σs,i σs′ ,i (x)dk,s,i+1 dk′ ,s′ ,i+1 (w), ′ x ∈ {αki (xs,i ), αki (xs′ ,i )}, otherwise. Consequently, the group Hi is abelian. For k ∈ Z, M ∈ V and i ≥ 0 we denote k dk,M,i = a−k i CM,i ai . (k) In particular, we have Ki = {dk,M,i : M ∈ V}. Lemma 2. If M = (m1 , . . . , mn ) ∈ V, then mn 1 dk,M,i = dm k,1,i · . . . · dk,n,i . 18 Proof (of Lemma 2). We have dk,M,i = mn k 1 a−k · cm 1,i · . . . · cn,i · ai = i = −k −k k k k n 2 1 · cm · cm (a−k · cm n,i · ai ) = 2,i · ai ) · . . . · (ai 1,i · ai ) · (ai i = · cn,i · aki )mn = (a−k · c1,i · aki )m1 · . . . · (a−k i i = mn 1 dm k,1,i · . . . · dk,n,i , which finishes the proof of Lemma 2. Since the group Hi is abelian, we see by Lemma 2 that every element of Hi can be presented as a product dk1 ,M1 ,i · . . . · dkt ,Mt ,i , t ≥ 1, (11) where k1 < k2 < . . . < kt , kj ∈ Z, Mj ∈ V, 1 ≤ j ≤ t. We call the weight of the product (11) as a number ||M1 || + . . . + ||Mt ||, where ||M || = |m1 | + . . . + |mn | for any M = (m1 , . . . , mn ) ∈ V. Now, we have to show that every element of Hi is represented uniquely by such a product. Suppose not. Then there are products (11) with nonzero weights defining the identity transformation. By Proposition 3 it must be t ≥ 2 in every such product. Let us choose t ≥ 2, i0 ≥ 0 and the triples (k1 , M1 , i0 ), . . . , (kt , Mt , i0 ) ∈ Z × V × {i0 }, such that the product dk1 ,M1 ,i0 · . . . · dkt ,Mt ,i0 defines the identity transformation and its nonzero weight is minimal. We can assume that each Mj ∈ V (1 ≤ j ≤ t) is a nonzero vector. For every i ≥ 0 we define the product Di = dk1 ,M1 ,i · . . . · dkt ,Mt ,i . There is a letter zi ∈ Xi (i ≥ 0) which does not belong to any of the cycles πs,i (1 ≤ s ≤ n). In particular zi 6= αki 1 (xs,i ) for every 1 ≤ s ≤ n. For every 1 ≤ s ≤ n we also have k ΣM1 ,i · ΣM2 ,i · . . . · ΣMj ,i (zi ) 6= αi j+1 (xs,i ), 1 ≤ j ≤ t. (12) Let us denote: zi′ = ΣM1 ,i · . . . · ΣMt ,i (zi ), i ≥ 0. ∗ X(i+1) For all i ≥ 0, k ∈ Z, M ∈ V, x ∈ Xi and w ∈ following recursion  xdk,M (s) ,i+1 (w), dk,M,i (xw) = ΣM,i (x)dk,M,i+1 (w), 19 we obtain by Lemma 1 the x = αki (xs,i ), otherwise. (13) The above recursion and the inequalities (12) imply ′ ′ Di (zi zi+1 . . . zi+j w) = zi′ zi+1 . . . zi+j Di+j+1 (w) (14) ∗ for all i, j ≥ 0 and w ∈ X(i+j+1) . Since Di0 defines the identity mapping, we obtain by (14) that for every i ≥ i0 the transformation Di also defines the identity mapping. Let 1 ≤ s ≤ n be such that the s-th component of M1 is not (s) equal to 0. In particular ||M1 || < ||M1 ||. Let ν ≥ i0 be such that the length L of the cycle πs,ν is greater than the sum |k1 | + . . . + |kt |. Let us consider the letter k1 (xs,ν ) ∈ Xν . z = αkν1 (xs,ν ) = πs,ν By the equality (13) we have dk1 ,M1 ,ν (zw) = zdk1 ,M (s) ,ν+1 (w) 1 (15) ∗ for every w ∈ X(ν+1) . For every 2 ≤ j ≤ t we obviously have: ΣM2 ,ν · . . . · ΣMj ,ν (z) = z. (16) z∈ / {αkνj (x1,ν ), . . . , αkνj (xn,ν )}. (17) For 2 ≤ j ≤ t we also have Indeed, suppose to othe contrary that k z = αkνj (xs′ ,ν ) = πs′j,ν (xs′ ,ν ) for some 2 ≤ j ≤ t and 1 ≤ s′ ≤ n. Then we have k k1 (xs,ν ) = πs′j,ν (xs′ ,ν ). πs,ν Since the cycles π1,ν , . . . , πn,ν are pairwise disjoint, we obtain s = s′ . Hence kj −k1 (xs,ν ) = xs,ν and consequently L | kj − k1 , which contradicts with the πs,ν inequality L > |k1 | + . . . + |kt |. By (13) and by (15)-(17) we obtain Dν (zw) = ∗ zD(w) for all w ∈ X(ν+1) , where D = dk1 ,M (s) ,ν+1 · dk2 ,M2 ,ν+1 · . . . · dkt ,Mt ,ν+1 . 1 Since Dν defines the identity mapping, the product D also defines the identity (s) mapping. Since ||M1 || < ||M1 || and each Mj (2 ≤ j ≤ t) is a nonzero vector, we see that D has a nonzero weight which is smaller than the weight of Di0 , contrary to our assumption. This finishes the proof of Proposition 4. Lemma 3. We have a−k · dk1 ,M1 ,i · . . . · dkt ,Mt ,i · aki = dk+k1 ,M1 ,i · . . . · dk+kt ,Mt ,i i for all i ≥ 0, t ≥ 1, k, kj ∈ Z and Mj ∈ V (1 ≤ j ≤ t). 20 Proof. The left side of the above equality is equal to (a−k · dk1 ,M1 ,i · aki ) · (a−k · dk2 ,M2 ,i · aki ) · . . . · (a−k · dkt ,Mt ,i · aki ). i i i Now, it suffices to see that the j-th factor (1 ≤ j ≤ t) of the above product is equal to a−k · dkj ,Mj ,i · aki i −kj = ai−k · (ai = ai −(k+kj ) k · CMj ,i · ai j ) · aki = k+kj · CMj ,i · ai = dk+kj ,Mj ,i . This finishes the proof of Lemma 3. ∗ Let us fix 1 ≤ s ≤ n. For all i, j ≥ 0 let us consider the words wi,j ∈ X(i) , where wi,j = xs,i xs,i+1 . . . xs,i+j . Lemma 4. For every i ≥ 0 the following hold: (i) the group Hi is contained in the stabilizer of any wi,j , (ii) for every nonzero integer k there is j ≥ 0 such that aki (wi,j ) 6= wi,j . Proof. For j ≥ 0 and M ∈ V we have ΣM,i+j (xs,i+j ) = xs,i+j . Thus by the equality (13) we see that for every j ≥ 0 the transformations dk,M,i+j (k ∈ Z, M ∈ V) stabilize the letter xs,i+j . Consequently, the elements dk,M,i (k ∈ Z, M ∈ V) stabilize the words wi,j (j ≥ 0). Since the elements generate the group Hi , we obtain (i). To show (ii) we see by the recursion for ai that for every j ≥ 0 we have: k k aki (wi,j ) = αki (xs,i ) . . . αki+j (xs,i+j ) = πs,i (xs,i ) . . . πs,i+j (xs,i+j ). Now, the claim follows from the assumption that the lengths of cycles πs,i+j (j ≥ 0) are not uniformly bounded. 4.3. The groups Gi Let us define the groups Gi = hai , c1,i , . . . , cn,i i = hKi , ai i, i ≥ 0. Directly by the definition of the group Hi we obtain Ki ≤ Hi ⊳ Gi . By Lemma 4 the group Hi intersects trivially with the cyclic group hai i. Thus Gi is a semidirect product of the subgroups Hi and L hai i. By Proposition 4 the subgroup Hi is isomorphic with the direct sum Z K via dk,s,i 7→ (. . . , 0, 0, κs , 0, 0, . . .), k ∈ Z, 1 ≤ s ≤ n, where κs ∈ K on the right hand side is in the k-th position. By Lemma 4 the cyclic group hai i is isomorphic, via ai 7→ u, with Z. By Lemma 3 this cyclic 21 group acts on Hi by L conjugation via the shift. Consequently, we obtain the isomorphism Gi ≃ Z K ⋊ Z = K ≀ Z, which is induced by the mapping ai 7→ u, cs,i 7→ ηs (1 ≤ s ≤ n). Further, for the group G(Ai ) generated by the i-th shift (i ≥ 0) of the automaton A we have G(Ai ) = hai , (b1 )i , . . . , (bn )i i. Since cs,i = (bs )i a−1 i , we obtain G(Ai ) = Gi and the mapping ai 7→ u, (bs )i 7→ ηs · u, 1 ≤ s ≤ n induces an isomorphism G(Ai ) ≃ K ≀ Z, which finishes the proof of Theorem 2. 5. Examples 5.1. The lamplighter group Z ≀ Z The lamplighter group Z ≀ Z is an example of a 2-generated torsion-free not finitely presented metabelian group. It is generated L by two elements u, η, where u is a generator of Z in the semidirect product Z Z ⋊ Z and η = (. . . , 0, 0, 0, u, 0, 0, 0, . . .), where u on the right hand side is in zero position. This group has the following k k′ presentation hu, η : [η u , η u ] = 1, k, k ′ ∈ Zi, where [x, y] = x−1 y −1 xy. Basing on the construction of the automaton A we can present the lamplighter group Z≀Z as a group defined by a 2-state self-similar automaton over the changing alphabet as follows. Let (ri )i≥0 be an arbitrary unbounded sequence of integers ri > 1 and let X = (Xi )i≥0 be the changing alphabet with Xi = {1, 2, . . . , 2ri }. Let A = (X, Q, ϕ, ψ) be an automaton with the 2-element set Q = {a, b} of internal states and the following sequences ϕ = (ϕi )i≥0 , ψ = (ψi )i≥0 of transition and output functions:   a, q = b, x = 1, αi (x), q = a, ϕi (q, x) = ψi (q, x) = q, otherwise, βi (x), q = b, where the permutations αi , βi of the set Xi may be defined, for example, as follows: αi = (1, 3, . . . , 2ri − 1), βi = (1, 3, . . . , 2ri − 1)(2, 4, . . . , 2ri ). By above, we see that the automaton A is self-similar and the group G(A) defined by A is isomorphic with the lamplighter group Z ≀ Z via a0 7→ u, c0 7→ η, where c0 = b0 a−1 0 . 22 5.2. The lamplighter groups Zn ≀ Z, n ≥ 1 Generalizing the above construction, we can describe for every positive integer n the wreath product Zn ≀ Z as a group defined by the following self-similar automaton A = (X, Q, ϕ, ψ): • X = (Xi )i≥0 , where Xi = {1, 2, . . . , 2nri } and (ri )i≥0 is an arbitrary unbounded sequence of integers greater than 1, • Q = {a, b0 , . . . , bn−1 },  a, q = bs , x = 2sri + 1, 0 ≤ s ≤ n − 1, • ϕi (q, x) = q, otherwise,  αi (x), q = a, • ψi (q, x) = βs,i (x), q = bs , 0 ≤ s ≤ n − 1, where the permutations αi , βs,i of the set Xi are defined as follows αi = n−1 Y (2sri + 1, 2sri + 3, . . . , 2sri + 2ri − 1), s=0 βs,i = αi · (2sri + 2, 2sri + 4, . . . , 2sri + 2ri ). 5.3. The lamplighter groups Cr ≀ Z, r > 1 For lamplighter groups of the form Cr ≀ Z (r > 1) the corresponding selfsimilar automaton representation A = (X, Q, ϕ, ψ) may be described as follows: • Xi = {1, 2, . . . , ri , ri + 1, ri + 2, . . . , ri + r}, where (ri )i≥0 is an arbitrary unbounded sequence of integers greater than 1, • Q = {a, b},  a, q = b, x = 1, • ϕi (q, x) = q, otherwise,  αi (x), q = a, • ψi (q, x) = βi (x), q = b, where the permutations αi , βi ∈ Sym(Xi ) may be defined, for example, as follows: αi = (1, 2, . . . , ri ), βi = (1, 2, . . . , ri )(ri + 1, ri + 2, . . . , ri + r). References [1] I. Bondarenko, R. Grigorchuk, R. Kravchenko, Y. Muntyan, V. Nekrashevych, Z. Sunic, D. Savchuk. Groups generated by 3-state automata over a 2-letter alphabet II, Journal of Mathematical Sciences Vol. 156, N 1 (2009), 187-208. 23 [2] A. Erschler. Piecewise automatic groups, Duke Math. J., 134:3 (2006), 591613. [3] A. Erschler. Automatically presented groups, Groups Geom. Dyn., (2007), 47-59. [4] R. Grigorchuk. Solved and Unsolved Problems Around One Group, in Infinite Groups: Geometric, Combinatorial and Dynamical Aspects. Progress in Mathematics Series, Vol. 248, L. Bartholdi, T. Ceccherini-Silberstein, T. Smirnova-Nagnibeda, A. Żuk (Eds.), 2005, 413 p. [5] R. Grigorchuk, P. Linnell, T. Schick and A. Żuk. On a question of Atiyah, C. R. Acad. Sci., Paris, Sr. I, Math. 331 (2000), 663-668. [6] R. Grigorchuk, V. Nekrashevych, V. Sushchanskii. Automata, Dynamical Systems and Groups, Proceedings of Steklov Institute of Mathematics, 231:128-203, 2000. [7] R. Grigorchuk, A. Żuk. The Lamplighter Group as a Group Generated by a 2-state Automaton and its spectrum, Geometriae Dedicata, 87:209-244, 2001. [8] L. Kaloujnine. La structure des p-groupes de Sylow des groupes symetriques finis, Ann. Sci. Ecole Norm. Sup. (3), 65:239-276, 1948. [9] M. Kambites, P. Silva, B. Steinberg. The spectra of lamplighter groups and Cayley machines, Geometriae Dedicata, Vol. 120, No. 1. (2006), pp. 193227. [10] B. Mikolajczak. Algebraic and structural automata theory, translated from the Polish Annals of Discrete Mathematics, Vol. 44 (North-Holland Publishing Co., Amsterdam, 1991). [11] V. Nekrashevych. Self-similar groups, volume 117 in Mathematical Surveys and Monographs. Amer. Math. Soc., Providence, RI, 2005. [12] A. V. Rozhkov. Theory of Aleshin type groups, Math. Notes Vol. 40, Number 5, (1986), 827-836. [13] A. V. Rozhkov. Conditions for finiteness in Aleshin-type groups, Algebra and Logic, Vol. 29, Number 6 (1990), 467-479. [14] A. V. Rozhkov. Stabilizers of corteges in Aleshin-type groups, Math. Notes Vol. 47, Number 3 (1990), 307-312. [15] P. Silva, B. Steinberg. On a class of automata groups generalizing lamplighter groups, Internat. J. Algebra Comput. 15 (2005), no. 5-6, 1213-1234. [16] V. I. Sushchanskyy. Periodic permutation p-groups and the unrestricted Burnside problem, Soviet Math. Dokl. vol 20 (1979), N4, 766-770. 24 [17] V. I. Sushchanskyy. Wreath products and periodic factorable groups, Math. USSR Sbornik, vol. 67 (1990), N2, 535-553. [18] V. I. Sushchanskyy. Wreath products and factorable groups, St. Petersburg Math, v. 6 (1995), N1, 173-198. [19] M. Vorobets, Ya. Vorobets. On a free group of transformations defined by an automaton, Geom. Dedicata 124 (2007) 237-249. [20] A. Woryna. On permutation groups generated by time-varying Mealy automata, Publ. Math. Debrecen, 67(1-2):115-130, 2005. [21] A. Woryna. Representations of a free group of rank two by time-varying Mealy automata, Discuss. Math. Gen. Algebra Appl. 25 (1) (2005) 119-134. [22] A. Woryna. On generation of wreath products of cyclic groups by two state time varying Mealy automata, Internat. J. Algebra Comput., 16(2):397-415, 2006. [23] A. Woryna. The concept of duality for automata over a changing alphabet and generation a free group by such automata, Theoret. Comput. Sci., 412(45):6420-6431, 2011. 25
4math.GR
HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH arXiv:1603.03671v5 [math.GR] 22 Sep 2017 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER Abstract. We show that any free product of two (non-trivial) countable groups, one of them being infinite, admits a faithful and homogeneous action on the Random Graph. We also show that a large class of HNN extensions or free products, amalgamated over a finite group, admit such an action and we extend our results to groups acting on trees. Finally, we show the ubiquity of finitely generated free dense subgroups of the automorphism group of the Random Graph whose action on it have all orbits infinite. Introduction The Random Graph (or Rado graph or Erdös-Rényi graph) is the unique, up to isomorphism, countable infinite graph R having the following property : for any pair of disjoint finite subsets (U, V ) of the set of vertices there exists a vertex adjacent to any vertex in U and not adjacent to any vertex in V . Using this property and a model theoretic device called “back-andforth” one can show that R is homogeneous: any graph isomorphism between finite induced subgraphs can be extended to a graph automorphism of R. Hence, the Random Graph plays the same role in graph theory as the Uryshon’s space does in metric spaces. The Random Graph has been popularized by Erdös and Rényi in a serie of papers between 1959 and 1968. They showed [ER63] that if a countable graph is chosen at random, by selecting edges independently with probability 21 from the set of 2-elements subsets of the vertex set, then almost surely the resulting graph is isomorphic to R. Erdös and Rényi conclude that this Theorem demolishes the theory of infinite random graphs (however, the world of finite random graphs is much less predictable). Since almost all countable graphs are isomorphic to R, Erdös and Renyi did not give an explicit construction of the Random Graph. However, by using the uniqueness property of R, it is clear that one may give many different explicit constructions. Such an explicit description was proposed by Rado [Ra64]. The uniqueness property of R may also be used to show many stability properties (if small changes are made on R then the resulting graph is still isomorphic to R) and to construct many automorphisms of R as well as group actions on R. The homogeneity of R means that its automorphism group Aut(R) is large: it acts transitively on vertices, edges and more generally on finite configurations of any given isomorphism type. We will view it as a closed subset of the Polish group (for the topology of pointwise convergence) of the bijections of the vertices. Hence, it has a natural Polish group topology. The goal of this paper is to understand the countable dense subgroups of Aut(R). The first construction of such subgroups was given in [Mac86], where Macpherson showed that Aut(R) contains a dense free subgroup on 2 generators. More generally, he showed that if M is a ℵ0 categorical structure, then Aut(M ) has a dense free subgroup of rank ℵ0 . Melles and Shelah This research has been partially supported by the ANR Blanc ANR-14-CE25-0004, acronym GAMME. 1 2 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER [MS94] proved that, if M is a saturated model of a complete theory T with |M | = λ > |T |, then Aut(M ) has a dense free subgroup of cardinality 2λ . By using the extension property for graphs, Bhattacharjee and Macpherson showed in [BM05] that Aut(R) has a dense locally finite subgroup. We call a group action Γ y R homogeneous if, for any graph isomorphism ϕ : U → V between finite induced subgraphs U, V of R, there exists g ∈ Γ such that g(u) = ϕ(u) for all u ∈ U . The homogeneity of R means exactly that Aut(R) y R is homogeneous. Moreover, it is easy to check that a subgroup G < Aut(R) is dense if and only if the action G y R is homogeneous. Hence, to understand the countable dense subgroups of Aut(R) one has to identity the class HR of all countable groups that admit a faithful and homogeneous action on R. Besides free groups ([Mac86], [MS94], [GK03] and [GS15]) and a locally finite subgroup ([BM05]), little is known on groups in HR . There are some obvious obstructions to be in the class HR : it is easy to deduce from the simplicity of Aut(R), proved in [Tr85], that any Γ ∈ HR must be icc and not solvable (see Corollary 2.14). Our first positive result is the following. We shall use the notion of highly core-free subgroups, which is a strengthening of core-freeness, introduced in [FMS15] and recalled in Section 1.2. As an example, let us note that a finite subgroup in an icc group is highly core-free. Theorem A. If Γ1 , Γ2 are non-trivial countable groups and Γ1 is infinite then Γ1 ∗ Γ2 ∈ HR . If Σ < Γ1 , Γ2 is a common finite subgroup such that Σ is highly core-free in Γ1 and, either Γ2 is infinite and Σ is highly core-free in Γ2 , or Γ2 is finite and [Γ2 : Σ] ≥ 2, then Γ1 ∗ Γ2 ∈ HR . Σ To prove Theorem A, we first show that any infinite countable group Γ admits a “nice” action on R (Corollary 2.8). To produce this explicit action on R we use an inductive limit process. Then starting from an action of Γ := Γ1 ∗ Γ2 y R and an automorphism α ∈ Aut(R), we construct a natural action πα : Γ y R and we show that the set {α ∈ Aut(R) : πα is faithful and homogeneous} is a dense Gδ in Aut(R) whenever the initial action is “nice enough” (Theorem 3.1). We follow the same strategy for amalgamated free products but we need to be more careful since we also have to realize the amalgamated free product relations. Using the same strategy, we prove an analogous result for HNN-extensions. Theorem B. Let H be an infinite countable group, Σ < H a finite subgroup and θ : Σ → H an injective group homomorphism. If both Σ and θ(Σ) are highly core-free in H then HNN(H, Σ, θ) ∈ HR . By Bass-Serre theory we obtain the following result. Corollary C. Let Γ be a countable group acting, without inversion, on a non-trivial tree T in the sense of [Se77]. If every vertex stabilizer of T is infinite and, for every edge e of T the stabilizer of e is finite and is a highly core-free subgroup of both the stabilizer of the source of e and the stabilizer of the range of e, then Γ ∈ HR . Finally, we study the ubiquity of dense free subgroups of Aut(R). Gartside and Knight [GK03] gave necessary and sufficient conditions for a Polish topological group to be “almost free1”, and gave applications to permutation groups, profinite groups, Lie groups and unitary 1A Polish group G is almost free if for all n ≥ 2, the set {(g , . . . , g ) ∈ Gn : g , . . . , g freely generates a 1 n 1 n free subgroup of G} is a dense Gδ in Gn . HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH 3 groups. In particular, they showed that if M is ℵ0 -categorical then Aut(M ) is almost free. There are abundant results on the ubiquity of free subgroups in various classes of groups. In particular, almost freeness of various oligomorphic2 groups has been shown in [Dr85], [Ka92], [GMR93], [Ca96] and [GK03]. We prove the following result. For k ≥ 2 and ᾱ = (α1 , . . . , αk ) ∈ Aut(R)k , we denote by hᾱi the subgroup of Aut(R) generated by α1 , . . . , αk , and we set Ak := {ᾱ ∈ Aut(R)k : hᾱi y R has all orbits infinite}. Since Ak is closed in Aut(R)k , it is a Baire space. Theorem D. For all k ≥ 2, the set of ᾱ = (α1 , . . . , αk ) ∈ Aut(R)k such that hᾱi is a free group with basis ᾱ and hᾱi y R is homogeneous with all orbits infinite is a dense Gδ in Ak . To prove Theorem D we use the “back-and-forth” device (Theorem 6.6). The paper is organized as follows. In Section 1 we introduce the notations used in the paper about graphs, random extensions and inductive limits. In Section 2 we introduce the Random Graph and we show how to extend any group action on a finite or countable graph to a group action on the Random Graph. We study the basic properties of the extension and the properties of groups acting homogeneously of R. We prove Theorem A in Section 3, Theorem B in Section 4, Corollary C in Section 5 and Theorem D in Section 6. 1. preliminaries 1.1. Graphs. A Graph G is a pair (V (G), E(G)), where V (G) is a set and E(G) ⊂ V (G)2 is a subset such that E(G) ∩ ∆ = ∅, where ∆ = {(u, u) ∈ V (G)2 : u ∈ V (G)}, and (u, v) ∈ E(G) ⇔ (v, u) ∈ E(G). For u, v ∈ V (G), we write u ∼ v if (u, v) ∈ E(G). Any subset U ⊂ V (G) has a natural structure of a graph, called the induced graph structure on U and denoted by GU , defined by V (GU ) = U and E(GU ) = E(G) ∩ (U × U ). Let G1 , G2 be two graphs. A morphism of graphs from G1 to G2 is a map π : V (G1 ) → V (G2 ) such that, for all u, v ∈ V (G1 ), u ∼ v =⇒ π(u) ∼ π(v). A morphism π is called an isomorphism of graphs if π is moreover bijective and its inverse π −1 : V (G2 ) → V (G1 ) is a morphism of graphs. A morphism of graphs π : G1 → G2 is called open if for all u, v ∈ V (G1 ), u ∼ v ⇔ π(u) ∼ π(v). Note that the inclusion ι of a subset U ⊂ V (G) gives an open and injective graph homomorphism ι : GU → G. Moreover, an injective graph homomorphism π : G1 → G2 induces an isomorphism between G1 and the induced graph on π(V (G1 )) if and only if π is open. A partial isomorphism of a graph G is an isomorphism between induced subgraphs of G. For a partial isomorphism ϕ, we write d(ϕ) its domain and r(ϕ) its range. We denote by P (G) the set of finite partial isomorphisms i.e. those partial isomorphism for which d(ϕ) is finite (hence r(ϕ) is also finite). Given a graph G, we write Aut(G) the group of isomorphisms from G to G. An action of a group Γ on the graph G is a group homomorphism α : Γ → Aut(G). We write Γ y G an action of Γ on G. Let S(X) be the group of bijections of a countable set X. It is a Polish group under the topology of pointwise convergence. Note that Aut(G) ⊂ S(V (G)) is a closed subgroup hence, Aut(G) is a Polish group. 2A permutation group G < S(X) is called oligomorphic if G has only finitely many orbits on X n for every n ∈ N. 4 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER 1.2. Group actions. The main purpose of this section is to introduce the Random Extension of a group action on a graph. This notion will be crucial to produce explicit actions on the Random Graph (Section 2.2). Many properties of an action are preserved by its Random Extension (Proposition 1.1). However, the freeness property is not preserved; except for actions of torsion-free groups. Nevertheless, a weaker -but still useful- property, that we call strong faithfulness, is preserved. An action Γ y X of a group Γ on a set X is called strongly faithful if, for all finite subsets F ⊂ X with X \ F 6= ∅ one has g(x) = x ∀x ∈ / F =⇒ g = 1. When X is finite, an action on X is strongly faithful if and only if it is free. When X is infinite, an action Γ y X is strongly faithful if and only if for all g ∈ Γ \ {1}, the set {x ∈ X : gx 6= x} is infinite. Note also that, for an action Γ y X with X infinite, one has: almost free (i.e. every non-trivial group element has finitely many fixed points) =⇒ strongly faithful =⇒ faithful. An action Γ y G of a group Γ on a graph G is called homogeneous if for all ϕ ∈ P (G) (recall that d(ϕ) and r(ϕ) are supposed to be finite), there exists g ∈ Γ such that gu = ϕ(u) for all u ∈ d(ϕ). It is easy to see that Γ y G is homogeneous if and only if the image of Γ in Aut(G) is dense. We say that Γ y G has infinite orbits (resp. is free, resp. is strongly faithful) if the action Γ y V (G) on the set V (G) has infinite orbits i.e. all orbits are infinite (resp. is free, resp. is strongly faithful). We say that Γ y G disconnects the finite sets if for all finite subsets F ⊂ V (G) there exists g ∈ Γ such that gF ∩ F = ∅ and, for all u, v ∈ F , gu ≁ v. Note that if Γ y G disconnects the finite sets then it has infinite orbits and the converse holds when E(G) = ∅. We say that Γ y G is singular if there exists u ∈ V (G) and g ∈ Γ such that gu ∼ u. The notion of a highly core-free subgroup Σ < Γ with respect to an action Γ y X on the set X has been introduced in [FMS15]. We will use here a similar notion for an action Γ y G on a graph G. Let Σ < Γ be a subgroup and Γ y G be an action on the graph G. We say that Σ is highly core-free with respect to Γ y G if for every finite subset F ⊂ V (G) there exists g ∈ Γ such that gF ∩ ΣF = ∅, gu ≁ v for all u, v ∈ F , Σgu ∩ Σgv = ∅ for all u, v ∈ F with u 6= v and σgu ≁ gv for all u, v ∈ F and all σ ∈ Σ \ {1}. It is clear that if Σ is highly core-free w.r.t. Γ y G then the action Γ y G disconnects the finite sets. We adopt a terminology slightly different from [FMS15]: if E(G) = ∅ then Σ is highly core-free with respect to the action Γ y G on the graph G in the sense explained above if and only if Σ is strongly highly core-free with respect to the action Γ y V (G) on the set in the sense of [FMS15, Definition 1.7]. In particular, if E(G) = ∅ and Γ y G is free then Σ is highly core-free with respect to Γ y G if and only if Σ is highly-core free in Γ (see [FMS15, Definition 1.1 and Lemma 1.6]. Note also that the trivial subgroup is highly core-free w.r.t. Γ y G if and only if the action Γ y G disconnects the finite sets and, if Σ is highly core-free with respect to Γ y G then every subgroup of Σ is highly core-free with respect to Γ y G. In particular, if there exists a subgroup Σ which is highly core-free with respect to Γ y G then the action Γ y G disconnects the finite sets. We refer to [FMS15] for examples of highly core-free subgroups and we just recall that any finite subgroup of an icc group is a highly core-free subgroup. In the sequel we always assume that G is a non-empty graph. The Random Extension of G is e = V (G) ⊔ Pf (V (G)), where Pf (X) denotes the set of non-empty the graph Ge defined by V (G) HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH 5 e = E(G) ⊔ (⊔U ∈P (V (G)) U × {U }) ⊔ (⊔U ∈P (V (G)) {U } × U ) finite subsets of a set X, and E(G) f f e is exactly the set of neighbors of the vertex U ∈ V (G). e Note that the i.e., U ⊂ V (G) ⊂ V (G) e defines an injective and open graph homomorphism ι : G → G. e inclusion ι : V (G) → V (G) Given an action Γ y G there is a natural action Γ y Ge for which the map ι is Γ-equivariant. e and, for U ∈ Pf (V (G)) ⊂ V (G) e and Indeed, we take the original action of Γ on V (G) ⊂ V (G) g ∈ Γ, we define the action of g on U by g · U := gU := {gu : u ∈ U } ∈ Pf (V (G)). This e U ∈ Pf (V (G)) and all g ∈ Γ we have defines an action of Γ on Ge since for all z ∈ V (G), z ∼ U ⇔ z ∈ U ⇔ gz ∈ gU ⇔ g · z ∼ g · U. It is clear that ι is Γ-equivariant. Proposition 1.1. The following holds. (1) Γ y Ge is faithful if and only if Γ y G is faithful. (2) Γ y Ge is non-singular if and only if Γ y G is non-singular. (3) Γ y Ge has infinite orbits if and only if Γ y G has infinite orbits. (4) Γ y Ge disconnects the finite sets if and only if Γ y G disconnects the finite sets. (5) If Γ has torsion then Γ y Ge is not free and if Γ is torsion-free, G is infinite and Γ y G is free then Γ y Ge is free. (6) If Γ y G is strongly faithful and G is infinite then Γ y Ge is strongly faithful. (7) Let Σ < Γ be a subgroup. Σ is highly core-free w.r.t. Γ y G if and only if it is highly e core-free w.r.t. Γ y G. Proof. The proofs of (1) to (6) are routine verifications and left to the reader. We prove e be a statement (7) below. Suppose that Σ is highly core-free w.r.t. Γ y G. Let F ⊂ V (G) finite set and write F = F1 ⊔ F2 , where F1 = F ∩ V (G) and F2 = F ∩ Pf (V (G)). Define the finite subset Fe = F1 ∪ (∪U ∈F2 U ) ⊂ V (G). Since Σ is highly core-free w.r.t. Γ y G there exists g ∈ Γ such that gFe ∩ ΣFe = ∅, u ≁ gv for all u, v ∈ Fe, Σgu ∩ Σgv = ∅ for all u, v ∈ Fe with u 6= v and σgu ≁ gv for all u, v ∈ Fe and all σ ∈ Σ \ {1}. Obviously, gF ∩ ΣF = ∅ and for all u, v ∈ F , gu ≁ v. Let us show that Σgu ∩ Σgv = ∅ for all u, v ∈ F with u 6= v. If both u, v are in F1 or if (u ∈ F1 and v ∈ F2 ) or (v ∈ F2 and u ∈ F1 ) this is obvious. So assume that U, V ∈ F2 . We need to show that ΣgU = ΣgV =⇒ U = V . We know that for all u ∈ U and all v ∈ V we have u 6= v =⇒ Σgu ∩ Σgv = ∅. Hence, if there exists x ∈ ΣgU ∩ ΣgV then we find σ, σ ′ ∈ Σ such that σgU = σ ′ gV . It follows that if u ∈ U then there exists v ∈ V such that σgu = σ ′ gv hence, Σgu = Σgv so we must have u = v ∈ V . Hence, U ⊂ V . Similarly, V ⊂ U so that U = V . It remains to show that σgu ≁ gv for all u, v ∈ F and σ ∈ Σ \ {1}. For σ ∈ Σ \ {1} it is clear that σgu ≁ gv for all u, v ∈ F2 and also for all u, v ∈ F1 . Let u ∈ F1 and V ∈ F2 . If σ ∈ Σ is such that σgu ∼ gV then there exists v ∈ V ⊂ Fe such that σgu = gv. Hence Σgu = Σgv which implies that u = v. It follows that σ = 1. The converse is obvious.  1.3. Inductive limits. Let Xn be a sequence of countable sets with injective maps ιn : Xn → Xn+1 . Let l ≥ n. Define the injective map ιl,n : Xn → Xl by ιl,n = ιl−1 ◦ · · · ◦ ιn if l > n, and ιn,n = id. Observe that for all n ≤ m ≤ l, ιl,m ◦ ιm,n = ιl,n . Let X∞ be the inductive limit i.e. X∞ = ⊔Xn / ∼, where x ∼ y if and only if there exists n, m ∈ N and l ≥ n, m such that x ∈ Xn , y ∈ Xm and ιl,n (x) = ιl,m (y). Observe that we have injections ι∞,n : Xn → X∞ , ι∞,n (x) = [x], where [x] denotes the class of the element x ∈ Xn ⊂ ⊔n Xn 6 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER for the equivalence relation described above. Those injections satisfy X∞ = ∪↑n ι∞,n (Xn ) and ι∞,n+1 ◦ ιn = ι∞,n for all n ∈ N. Actually X∞ is the unique set, up to a canonical bijection, such that there exists injections ι∞,n : Xn → X∞ satisfying X∞ = ∪↑n ι∞,n (Xn ) and ι∞,n+1 ◦ ιn = ιn for all n ∈ N. Note that X∞ is at most countable. Actually, it is finite if and only if for all n ∈ N, Xn is finite and there exists l ∈ N such that ιk (Xk ) = Xk+1 for all k ≥ l. Given a sequence of actions πn : Γ → S(Xn ) of a group Γ on the set Xn satisfying πn+1 (g) ◦ ιn = ιn ◦ πn (g) for all n ∈ N and g ∈ Γ, we define the inductive limit action π∞ : Γ → S(X∞ ) by π∞ (g) ◦ ι∞,n = ι∞,n ◦ πn (g) for all n ∈ N. The next proposition contains some standard observations on inductive limits. Since this results are well known and very easy to check, we omit the proof. Proposition 1.2. The following holds. (1) (2) (3) (4) If there exists n ∈ N such that Γ y Xn is faithful then Γ y X∞ is faithful. Γ y X∞ is free if and only if Γ y Xn is free for all n ∈ N. If Γ y Xn is strongly faithful for all n ∈ N then Γ y X∞ is strongly faithful. Γ y X∞ has infinite orbits if and only if Γ y Xn has infinite orbits for all n ∈ N. Let Gn be a sequence of countable graphs with injective graphs homomorphisms ιn : Gn → Gn+1 . We define the inductive limit graph G∞ by defining V (G∞ ) = ∪↑ ι∞,n (V (Gn )) as the inductive limit of the V (Gn ) with respect to the maps ιn and E(G∞ ) is the set of couples (u, v) ∈ V (G∞ )2 for which there exists k ≥ n, m with u = ι∞,n (u0 ), u0 ∈ V (Gn ), v = ι∞,m (v0 ), v0 ∈ V (Gm ) and (ιk,n (u0 ), ιk,m (v0 )) ∈ E(Gk ). We collect elementary observations on the inductive limit graph in the following proposition. Proposition 1.3. The following holds. (1) ι∞,n is a graph homomorphism for all n ∈ N. (2) If ιn is open for all n ∈ N then ι∞,n is open for all n ∈ N. (3) If Γ y G∞ disconnects the finite sets then Γ y Gn disconnects the finite sets for all n ∈ N. The converse holds when ιn is open for all n ∈ N. (4) If Γ y G∞ is non-singular then Γ y Gn is non-singular for all n ∈ N. The converse holds when ιn is open for all n ∈ N. (5) Let Σ < Γ be a subgroup. If Σ is highly core-free w.r.t. Γ y G∞ then Σ is highly core-free w.r.t. Γ y Gn for all n. The converse holds when ιn is open for all n. Proof. The proofs of (1) to (4), being routine verifications, are left to the reader. Let us prove (5). Suppose that Σ is highly core-free with respect to Γ y G∞ and let n ∈ N and F ⊂ V (Gn ) be a finite subset. Let F ′ = ι∞,n (F ) ⊂ V (G∞ ) and take g ∈ Γ such that gF ′ ∩ ΣF ′ = ∅, gu′ ≁ v ′ for all u′ , v ′ ∈ F ′ , Σgu′ ∩ Σgv ′ = ∅ for all u′ , v ′ ∈ F ′ with u′ 6= v ′ and σgu′ ≁ gv ′ for all u′ , v ′ ∈ F ′ and all σ ∈ Σ \ {1}. It follows that gF ∩ ΣF = ∅ and Σgu ∩ Σgv = ∅ for all u, v ∈ F with u 6= v. Moreover, ι∞,n (gu) ≁ ι∞,n (v) and ι∞,n (σgu) ≁ ι∞,n (gv) for all u, v ∈ F and all σ ∈ Σ \ {1}. By (1) we have gu ≁ v and σgu ≁ gv for all u, v ∈ F and all σ ∈ Σ \ {1}. Suppose now that ιn is open and Σ is highly core-free w.r.t. Γ y Gn for all n ∈ N. Let F ⊂ V (G∞ ) be a finite set and take n ∈ N large enough so that F = ι∞,n (F ′ ), where F ′ ⊂ V (Gn ) is a finite set. Take g ∈ Γ such that gF ′ ∩ ΣF ′ = ∅, gu′ ≁ v ′ and σgu′ ≁ gv ′ for all u′ , v ′ ∈ F ′ and all σ ∈ Σ \ {1} and Σgu′ ∩ Σgv ′ = ∅ for all u′ , v ′ ∈ F ′ with u′ 6= v ′ . Since ι∞,n injective we have gF ∩ ΣF = ∅ and Σgu ∩ Σgv = ∅ for all u, v ∈ F with u 6= v. HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH 7 Moreover, since ι∞,n is open (by (2)) we have gu ≁ v and σgu ≁ gv for all u, v ∈ F and all σ ∈ Σ \ {1}.  2. The Random Graph 2.1. Definition of the Random Graph. Given a graph G and subsets U, V ⊂ V (G) we define GU,V as the induced subgraph on the subsets of vertices V (GU,V ) := {z ∈ V (G) \ (U ∪ V ) : z ∼ u , ∀u ∈ U and z ≁ v , ∀v ∈ V }. Note that V (GU,V ) may be empty for some subsets U and V . To ease the notations we will denote by the same symbol GU,V the induced graph on V (GU,V ) and the set of vertices V (GU,V ). Definition 2.1. We say that a graph G has property (R) if, for any disjoint finite subsets U, V ⊂ V (G), GU,V 6= ∅. Note that a graph with property (R) is necessarily infinite. We recall the following well-known result (See e.g. [Ca97] or [Ca99]) that will be generalized later (Proposition 2.5). Proposition 2.2. Let G1 , G2 be two infinite countable graphs with property (R) and A ⊂ V (G1 ), B ⊂ V (G2 ) be finite subsets. Any isomorphism between the induced graphs ϕ : (G1 )A → (G2 )B extends to an isomorphism ϕ : G1 → G2 . There are many ways to construct a countable graph with property (R). Proposition 2.2 shows that a countable graph with property (R) is unique, up to isomorphism. Such a graph is denoted by R and called the Random Graph. It also follows from Proposition 2.2 that every ϕ ∈ P (R) admits an extension ϕ ∈ Aut(R) (i.e. ϕ|d(ϕ) = ϕ). Proposition 2.2 is also useful to show stability properties of the graph R as done in the next Proposition. Proposition 2.3. The following holds. (1) For all finite subset A ⊂ V (R), the induced subgraph RV (R)\A on V (R) \ A is isomorphic to R. (2) For all finite and disjoint U, V ⊂ V (R), the graph RU,V is isomorphic to R. (3) For all non-empty countable graph G, the inductive limit G∞ of the sequence G0 = G and Gn+1 = Gen (with ιn : Gn → Gn+1 the inclusion) is isomorphic to R. In particular, every countable graph is isomorphic to an induced subgraph of R. Proof. (1). Let U, V ⊂ V (R)\A be disjoint finite subsets. Apply property (R) with the disjoint finite subsets U ′ = U ⊔ A, V ⊂ V (R) and get z ∈ V (R) \ (U ′ ∪ V ) = (V (R) \ A) \ (U ∪ V ) such that z ∼ u for all u ∈ U and z ≁ v for all v ∈ V . Hence the countable graph induced on V (R) \ A has property (R). (2). Since RU,V is at most countable, it suffices to check that RU,V has property (R) and it is left to the reader. (3). Since G is either finite and non-empty or infinite countable, the inductive limit is infinite countable and it suffices to check that it has property (R). Since ιn is open for all n ∈ N it follows from Proposition 1.3 that we may and will assume that ι∞,n = id and (Gn )n is an increasing sequence of induced subgraphs of G∞ such that G∞ = ∪↑ Gn . Let U, V be two finite and disjoint subsets of the inductive limit. Let n large enough so that U, V V (Gn ). If U 6= ∅, we consider the element z = U ∈ Pf (V (Gn )) ⊂ V (Gn+1 ) ⊂ V (G∞ ) and if U = ∅, we consider any element z = {x} ∈ Pf (V (Gn )) ⊂ V (Gn+1 ) ⊂ V (G∞ ) for x ∈ V (Gn ) \ V . 8 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER Then, by definition of the Random extension, we have, for all u ∈ U , (u, z) ∈ E(Gn+1 ) and, for all v ∈ V , (v, z) ∈ / E(Gn+1 ). Since Gn+1 is the induced subgraph on V (Gn+1 ) we have z ∈ (G∞ )U,V . The last assertion also follows from Proposition 1.3 since the inclusion of G0 = G in the inductive limit is open.  Remark 2.4. The construction of Proposition 2.3, assertions (3), shows the existence of R. This construction may also be performed starting with any countable graph G, even G = ∅, and replacing, in the construction of the Random Extension, Pf (X) by all the finite subsets of X (even the empty one). The resulting inductive limit is again isomorphic to R. We shall need the following generalization of Proposition 2.2. The proof is done by using the “back-and-forth” device. Proposition 2.5. For k = 1, 2, let πk : Σ y R be two free and non-singular actions of the finite group Σ. For any partial isomorphism ϕ ∈ P (R) such that π1 (Σ)d(ϕ) = d(ϕ) and π2 (Σ)r(ϕ) = r(ϕ) and ϕπ1 (σ) = π2 (σ)ϕ for all σ ∈ Σ, there exists ϕ ∈ Aut(R) such that ϕπ1 (σ) = π2 (σ)ϕ for all σ ∈ Σ and ϕ|d(ϕ) = ϕ. ∞ Proof. Write V (R) \ d(ϕ) = ⊔∞ k=1 π1 (Σ)xk and V (R) \ r(ϕ) = ⊔k=1 π2 (Σ)yk . We define inductively pairwise distinct integers kn , pairwise distinct integers ln , subsets An = d(ϕ) ⊔ π1 (Σ)xk1 ⊔ . . . ⊔ π1 (Σ)xkn ⊂ V (R), Bn = r(ϕ) ⊔ π2 (Σ)yl1 ⊔ · · · ⊔ π2 (Σ)yln ⊂ V (R) and isomorphisms ϕn : An → Bn such that ϕn |d(ϕ) = ϕ, ϕn π1 (σ) = π2 (σ)ϕn for all σ ∈ Σ, An ⊂ An+1 , Bn ⊂ Bn+1 , ϕn+1 |An = ϕn and V (R) = ∪An = ∪Bn . Once it is done, the Proposition is proved since we can define ϕ ∈ Aut(R), x ∈ An 7→ ϕn (x) which obviously satisfies ϕπ1 (σ) = π2 (σ)ϕ for all σ ∈ Σ and ϕ|d(ϕ) = ϕ. Define A0 = d(ϕ), B0 = r(ϕ) and ϕ0 = ϕ. If step n is constructed and n is even define kn+1 = Min{k ≥ 1 : xk ∈ / An }, U = {x ∈ An : x ∼ xkn+1 }, V = An \ U and W = {y ∈ V (R) \ Bn : y ∼ ϕn (x) for all x ∈ U and y ≁ ϕn (x) for all x ∈ V }. By property (R), W is non-empty. Let ln+1 = Min{k ≥ 1 : π2 (Σ)yk ∩ W 6= ∅}. Replacing yln+1 by an element in π2 (Σ)yln+1 we may and will assume that yln+1 ∈ W . Define An+1 = An ⊔ π1 (Σ)xkn+1 and Bn+1 = Bn ⊔ π2 (Σ)yln+1 . By freeness, we may define a bijection ϕn+1 : An+1 → Bn+1 by ϕn+1 |An = ϕn and ϕn+1 (π1 (σ)xkn+1 ) = π2 (σ)ykn+1 , for all σ ∈ Σ. By construction, it is an isomorphism between the induced graphs. Indeed if x ∈ An is such that x ∼ π1 (σ)xkn+1 for some σ ∈ Σ then π1 (σ −1 )x ∼ xkn+1 so yln+1 ∼ ϕn (π1 (σ −1 )x) = π2 (σ −1 )ϕn (x) hence, ϕn+1 (x) = ϕn (x) ∼ π2 (σ)yln+1 = ϕn+1 (π1 (σ)xkn+1 ). Since ϕn is a graph isomorphism and since both actions are non-singular (which means that there is no edges on induced subgraphs of the form πk (Σ)z) it shows that ϕn+1 is also a graph isomorphism. By construction we also have that ϕn+1 π1 (σ) = π2 (σ)ϕn+1 for all σ ∈ Σ. If n is odd define ln+1 = Min{k ≥ 1 : yk ∈ / Bn }, U = {y ∈ Bn : y ∼ yln+1 }, V = Bn \ U and −1 W = {x ∈ V (G1 ) \ An : x ∼ ϕ−1 n (y) for all y ∈ U and x ≁ ϕn (y) for all y ∈ V }. By property (R), W is non-empty. Let kn+1 = Min{k ≥ 1 : xk ∈ W }. We may and will assume that xkn+1 ∈ W . Define An+1 = An ⊔ π1 (Σ)xkn+1 , Bn+1 = Bn ⊔ π2 (Σ)yln+1 . By construction, the map ϕn+1 : An+1 → Bn+1 defined by ϕn+1 |An = ϕn and ϕn+1 (π1 (σ)xkn+1 ) = π2 (σ)ykn+1 , σ ∈ Σ, is an isomorphism between the induced graphs. HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH 9 F Let us show that ∪An = d(ϕ) ⊔ ∞ n=1 π1 (Σ)xkn = V (R). It suffices to show that {kn : n ≥ ∗ 1} = N . Suppose that there exists s ∈ N, s 6= kn for all n ≥ 1. Since the elements kn are pairwise distinct, the set {k2n+1 : n ≥ 0} is not bounded. Hence, there exists n ∈ N such that s < k2n+1 . By definition, we have k2n+1 = Min{k ≥ 1 : xk ∈ / A2n }. However we have xs ∈ / A2n and s < k2n+1 , a contradiction. The proof of ∪Bn = V (R) is similar.  2.2. Induced action on the Random Graph. Let Γ y G be an action of a group Γ on a non-empty countable graph G and consider the sequence of graphs G0 = G and Gn+1 = Gen with the associated sequence of actions Γ y Gn . By Proposition 2.3, assertion (3), the inductive limit action defines an action of Γ on R. We call it the induced action of Γ y G on R. We will show in Corollary 2.7 that many properties on the action Γ y G are preserved when passing to the induced action Γ y R. However, freeness is not preserved and one has to consider a weaker notion that we call property (F ). Definition 2.6. We say that an action Γ y G has property (F ) if for all finite subsets S ∈ Γ \ {1} and F ⊂ V (G), there exists x ∈ V (G) \ F such that x ≁ u for all u ∈ F and gx 6= x for all g ∈ S. Note that any action with property (F ) is faithful and any free action on R has property (F ). Corollary 2.7. Let Γ y G be an action of a group Γ on a non-empty countable graph G and write Γ y R the induced action. The following holds. (1) If Γ y G is faithful then Γ y R is faithful. (2) Γ y G has infinite orbits if and only if Γ y R has infinite orbits. (3) Γ y G disconnects the finite sets if and only if Γ y R disconnects the finite sets. (4) Γ y G is non-singular if and only if Γ y R is non-singular. (5) Let Σ < Γ be a subgroup. Σ is highly core-free w.r.t. Γ y G if and only if Σ is highly core-free w.r.t. Γ y R. (6) If Γ y G is free, G is infinite and Γ is torsion free then Γ y R is free. (7) If Γ y G is strongly faithful, G is infinite then Γ y R has property (F ). Proof. The assertions (1) to (6) follow directly follows from Propositions 1.1, 1.2 and 1.3. Let fn . Let F ⊂ V (G∞ ) us prove (7). We recall that R = G∞ = ∪↑ Gn , where G0 = G and Gn+1 = G and S = {g1 , . . . , gn } ⊂ Γ \ {1} be finite subsets. Since Γ y G is strongly faithful we deduce, by Proposition 1.1 (6) and induction that Γ y GN is strongly faithful for all N . Let N ∈ N be large enough so that F ⊂ V (GN ). We can use strong faithfulness (and the fact that GN is infinite) to construct, by induction, pairwise distinct vertices y1 , . . . yn ∈ V (GN ) \ F such that yi 6= gj yj for all 1 ≤ i, j ≤ n. Define x := {y1 , . . . yn } ∈ Pf (V (GN )) ⊂ V (GN +1 ) ⊂ V (G∞ ) = V (R). By construction x ∈ V (R) \ F and gk x 6= x for all k. Since the inclusion GN +1 ⊂ G∞ = R is open, we also have x ≁ u for all u ∈ F .  Corollary 2.8. Every infinite countable group Γ admits an action Γ y R that is nonsingular, has property (F ) and disconnects the finite sets. If Γ is torsion-free then the action can be chosen to be moreover free. Proof. Consider the graph G defined by V (G) = Γ and E(G) = ∅ with the action Γ y G given by left multiplication which is free, has infinite orbits and hence disconnects the finite sets and is non-singular since E(G) = ∅. By Corollary 2.7 the induced action Γ y R has the required properties.  10 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER 2.3. The Random Extension with paramater. We now describe a parametrized version of the induced action. As explained before, freeness is not preserved when passing to the action on the Random Extension and also when passing to the induced action. However, it is easy to compute explicitly the fixed points of any group element in the Random Extension. Now, given an action Γ y G and some fixed group elements in Γ, one can modify the Random Extension by removing the fixed points of our given group elements to make them act freely on this modified version of the Random Extension. Then, the inductive limit process associated to this modified Random Extension will also produce, in some cases, the Random Graph with an action on it, for which our given group elements act freely. We call this process the parametrized Random Extension. Let π : Γ y G be an action of the group Γ on the graph G and F ⊂ Γ be a subset. From the action π, we have a canonical action Γ y Ge on the Random Extension. Remark 2.9. If G is infinite, for any g ∈ Γ, the set of fixed points of g for the action on the Random Extension is either empty or of the form : FixGe(g) = {A ∈ Pf (V (G)) of the form A = ⊔N i=1 hgixi with hgixi finite ∀i} ⊔ FixG (g). In particular, when the action π is free one has FixGe(g) = ∅ whenever g has infinite order and, if g has finite order, then FixGe(g) is the set of finite unions of hgi-orbits. e \ {Fix e(g) : g ∈ F }. Note that if gF g −1 = F for all Consider the induced subgraph on V (G) G −1 e g ∈ Γ then, since gFixGe(h) = FixGe(ghg ), the induced subgraph on V (G)\{Fix Ge(g) : g ∈ F } is globally Γ-invariant and we get an action of Γ on it by restriction for which the elements of F act freely by construction. However, for a general F , we cannot restrict the action and this is why we will remove more sets then the fixed points of elements of F . Assume from now that G is a graph and l ∈ N∗ . We define the graph Gel , the Random Extension of G with parameter l, as the induced subgraph on e V (Gel ) := V (G) ⊔ {U ∈ Pf (V (G)) : gcd(l, |U |) = 1} ⊂ V (G). Fix an action Γ y G. Since for any g ∈ Γ and any finite subset U ⊂ V (G) one has |gU | = |U |, the subgraph Gel is globally Γ-invariant and we get an action π el : Γ y Gel by restriction. Note that for any u ∈ V (G) one has {u} ∈ V (Gel ). It is clear that the inclusions of G in Gel and of Gel in Ge are Γ-equivariant open (and injective) graph homomorphisms. Proposition 2.10. Let π : Γ y G be an action and l ∈ N∗ . (1) π is faithful if and only if π el is faithful. (2) π is non-singular if and only if π el is non-singular. (3) π has infinite orbits if and only if π el has infinite orbits. (4) π disconnects the finite sets if and only if π el disconnects the finite sets. (5) Let Σ < Γ be a subgroup. Σ is highly core-free w.r.t. π if and only if Σ is highly core-free w.r.t. π el . (6) If π is strongly faithful and G is infinite then π el is strongly faithful. (7) Let Σ < Γ be a finite subgroup. If Σ y G is free and G is infinite then Σ y Ge|Σ| is free. Proof. Assertions (1) to (6) are obvious. Let us prove (7). Since any non-trivial element of Σ acts freely, it follows from Remark 2.9 that any finite subset U ⊂ V (G) in the set of fixed HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH 11 points of σ ∈ Σ is a finite union of sets of the form hσix, hence its size is a multiple of the order of σ and U ∈ / V (Ge|Σ| ).  We can now construct the induced action on R of the action π with paramter l. Define the g sequence of graphs G0 = G with action π0 = π of Γ on it and, for n ≥ 0, Gn+1 = (G n )l with l the action πn+1 = (f πn )l . Consider the inductive limit G∞ with the inductive limit action l : Γ y G l on it. We list the properties of π in the next Proposition. π∞ l ∞ Proposition 2.11. Let π : Γ y G be an action on a countable graph G and let l ∈ N∗ . l ≃ R. (1) If G is infinite then G∞ l is faithful. (2) π is faithful if and only if π∞ l has infinite orbits. (3) π has infinite orbits if and only if π∞ l disconnects the finite sets. (4) π disconnects the finite sets if and only if π∞ (5) Let Σ < Γ be a subgroup. Σ is highly core-free w.r.t. to π if and only if Σ is highly l . core-free w.r.t. to π∞ l is non-singular. (6) π is non-singular if and only if π∞ |Σ| (7) Let Σ < Γ be a finite subgroup. If Σ y G is free and G is infinite then Σ y G∞ is free. l has property (F ). (8) Suppose that Γ and G are infinite. If π is strongly faithful then π∞ Proof. The assertions (2) to (7) follow directly follows from Propositions 2.10, 1.2 and 1.3. (1). Since G is infinite it follows that Gn is infinite for all n and G∞ is infinite. Hence, it suffices to check that G∞ has property (R). Let U, V ⊂ V (G∞ ) be two finite subsets such that U ∩ V = ∅ and let n ∈ N such that U, V ⊂ V (Gn ). Since V (Gn ) \ V is infinite, we may find a finite subset x ⊂ V (Gn ) \ V such that U ⊂ x and gcd(|x|, |Σ|) = 1. Hence, |Σ| fn )|Σ| ) ⊂ V (G∞ x ∈ V ((G ) is such that x ∼ u for all u ∈ U and x ≁ v for all v ∈ V . l ) be finite subsets. Taking a larger S if (8). Let S = {g1 , . . . , gn } ⊂ Γ \ {1} and F ⊂ V (G∞ necessary, we may and will assume that gcd(n, l) = 1. We repeat the proof of Corollary 2.7, assertion (7) and we get pairwise distinct vertices y1 , . . . yn ∈ V (GN ), where F ⊂ V (GN ) such that the element x = {y1 , . . . , yn } ∈ Pf (V (GN )) ⊂ GeN satisfies the desired properties. Since l gcd(|x|, l) = 1, x ∈ V ((Gg  N )l ) ⊂ V (G∞ ). This concludes the proof. Let Σ < Γ be a finite subgroup of an infinite countable group Γ. By considering the induced action on R with parameter l = |Σ| of the free action by left multiplication Γ y Γ and view Γ as a graph with no edges, we obtain the following Corollary. Corollary 2.12. Let Σ < Γ be a finite subgroup of an infinite countable group Γ. There exists a non-singular action Γ y R with property (F ) such that the action Σ y R is free, the action H y R of any infinite subgroup H < Γ disconnects the finite sets and, for any pair of intermediate subgroups Σ′ < H ′ < Γ, if Σ′ is a highly core-free in H ′ then Σ′ is highly core-free w.r.t. H ′ y R. 2.4. Homogeneous actions on the Random Graph. Proposition 2.13. Let Γ y R be an homogeneous action. The following holds. (1) Γ y R disconnects the finite sets. (2) If N < Γ is normal then either N acts trivially or it acts homogeneously. 12 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER Proof. (1). Let F ⊂ V (R) be a finite set and write F = {u1 , . . . , un } where the vertices ui are pairwise distinct. We shall define inductively pairwise distinct vertices v1 , . . . vn ∈ V (R) \ F such that, for all 1 ≤ i, j ≤ n, ui ∼ uj ⇔ vi ∼ vj and ui ≁ vj . Take v1 ∈ R∅,F . Suppose that for a given 1 ≤ l ≤ n − 1 we have pairwise distinct vertices v1 , . . . , vl ∈ V (R) \ F such that, for all 1 ≤ i, j ≤ l, ui ∼ uj ⇔ vi ∼ vj and ui ≁ vj . Let U = {vi : 1 ≤ i ≤ l s.t. ui ∼ ul+1 } and V = {vi : 1 ≤ i ≤ l s.t. ui ≁ ul+1 }. Then U and V ⊔ F are finite and disjoint. Take vl+1 ∈ RU,V ⊔F . Then vl+1 ∼ vi ⇔ ul+1 ∼ ui and vl+1 ≁ ui for all 1 ≤ i ≤ l and vl+1 ∈ / F ⊔ {vi : 1 ≤ i ≤ l}. This concludes the construction of the vi by induction. By the properties of the vi , the map ϕ : {ui : 1 ≤ i ≤ n} → {vi : 1 ≤ i ≤ n} defined by ϕ(ui ) = vi is an isomorphism between the induced subgraphs. Since Γ y R is homogeneous there exists g ∈ Γ such that gui = vi for all 1 ≤ i ≤ n. It follows that gF ∩ F = ∅ and gu ≁ u′ for all u, u′ ∈ F . (2). Write N the closure of the image of N inside Aut(R). By [Tr85], the abstract group Aut(R) is simple and, since N is normal in Aut(R) one has either N = {1} or N = Aut(R).  Using the previous Proposition and arguing as in [MS13, Corollary 1.6] we obtain the following Corollary. Corollary 2.14. If Γ ∈ HR then Γ is icc and not solvable. 3. Actions of amalgamated free products on the Random Graph Let Γ1 , Γ2 be two countable groups with a common finite subgroup Σ and define Γ = Γ1 ∗ Γ2 . Suppose that we have a faithful action Γ y R and view Γ < Aut(R). Σ Let Z := {α ∈ Aut(R) : ασ = σα ∀σ ∈ Σ}. Note that Z is a closed subgroup of Aut(R), hence a Polish group. Moreover, for all α ∈ Z, there exists a unique group homomorphism πα : Γ → Aut(R) such that:  g if g ∈ Γ1 , πα (g) = −1 α gα if g ∈ Γ2 . When Σ is trivial, we have Z = Aut(R). In this section we prove the following result. Theorem 3.1. If Γ y R is non-singular and has property (F ), Σ is highly core-free w.r.t. Γ1 , Γ2 y R and Σ y R is free then the set O = {α ∈ Z : πα is homogeneous and faithful} is a dense Gδ in Z. In particular, for every countably infinite groups Γ1 , Γ2 we have Γ1 ∗Γ2 ∈ HR and, for any finite highly core-free subgroup Σ < Γ1 , Γ2 we have Γ1 ∗ Γ2 ∈ HR . Σ Proof. We separate the proof in two lemmas. Lemma 3.2. If Σ y R is free and non-singular and Σ is highly core-free w.r.t. Γ1 , Γ2 y R then the set U = {α ∈ Z : πα is homogeneous} is a dense Gδ in Z. Proof. Since U = ∩ϕ∈P (R) Uϕ , where Uϕ = {α ∈ Z : ∃g ∈ Γ such that πα (g)|d(ϕ) = ϕ} is obviously open, it suffices to show that Uϕ is dense for all ϕ ∈ P (R). Let ϕ ∈ P (R), α ∈ Z and F ⊂ V (R) a finite subset. It suffices to show that there exists γ ∈ Z and g ∈ Γ such that γ|F = α|F and πγ (g)|d(ϕ) = ϕ. Since Σ is highly core-free w.r.t. Γ1 y R, there exists g1 ∈ Γ1 such that g1 d(ϕ) ∩ ΣF = ∅, g1 x ≁ u and σg1 x ≁ g1 x′ for all u ∈ ΣF , x, x′ ∈ d(ϕ), σ ∈ Σ \ {1} and Σg1 x ∩ Σg1 x′ = ∅ for all x, x′ ∈ d(ϕ) with x 6= x′ . There exists also g2 ∈ Γ1 such that HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH 13 g2−1 r(ϕ) ∩ (ΣF ⊔ Σg1 d(ϕ)) = ∅ and g2−1 y ≁ u , σg2−1 y ≁ g2−1 y ′ for all u ∈ ΣF ⊔ Σg1 d(ϕ), y, y ′ ∈ r(ϕ), σ ∈ Σ \ {1} and Σg2−1 y ∩ Σg2−1 y ′ = ∅ for all y, y ′ ∈ r(ϕ) with y 6= y ′ . Define F ′ = α(ΣF ) ⊔ Σα(g1 d(ϕ)). Since Σ is highly core-free w.r.t. Γ2 y R, there exists h ∈ Γ2 such that hF ′ ∩ F ′ = ∅ and for all u, v ∈ F ′ and σ ∈ Σ \ {1}, hu ≁ v, σhu ≁ hv and Σhu ∩ Σhu′ = ∅ for all u, u′ ∈ F ′ with u 6= u′ . Define A = (ΣF ) ⊔ Σg1 d(ϕ) ⊔ Σg2−1 r(ϕ) and B = α(ΣF ) ⊔ α(Σg1 d(ϕ)) ⊔ Σhα(g1 d(ϕ)). Note that ΣA = A and, since α ∈ Z, ΣB = B. Since the Σ action is free, we can define a bijection γ0 : A → B by γ0 (u) = α(u) for u ∈ ΣF ⊔ Σg1 d(ϕ) and γ0 (σg2−1 ϕ(x)) = σhg1 x, for all x ∈ d(ϕ) and σ ∈ Σ. By construction, γ0 is an isometry and γ0 σ = σγ0 for all σ ∈ Σ. By Proposition 2.5 there exists an extension γ ∈ Z of γ0 . Note that γ|F = α|F moreover, with g = g2 hg1 ∈ Γ we have, for all x ∈ d(ϕ), πγ (g)x = g2 γ −1 hγg1 x = g2 γ −1 hα(g1 x) = g2 g2−1 ϕ(x) = ϕ(x).  Lemma 3.3. If Σ y R is free, Γ y R is non-singular and has property (F ) then the set V = {α ∈ Z : πα is faithful} is a dense Gδ in Z. Proof. Writing V = ∩g∈Γ\{1} Vg , where Vg = {α ∈ Z : πα (g) 6= id} is obviously open, it suffices to show that Vg is dense for all g ∈ Γ \ {1}. If g ∈ Γk \ {1} (k = 1, 2) then it is easy to see that Vg = Z. Hence it suffices to show that Vg is dense for all g reduced of length at least 2. Write g = gin . . . gi1 , where n ≥ 2 and gik ∈ Γik \ Σ reduced expression for g. Fix α ∈ Z and F ⊂ V (R) a finite subset. Define the finite sets ! n [ Fe := ΣF ∪ α(ΣF ) ∪ (gi . . . gi )−1 (ΣF ∪ α(ΣF ) ⊂ V (R), l 1 l=1 and σgik . . . gi1 : 1 ≤ l < k ≤ n, σ ∈ Σ} ⊂ Γ\{1}. S := {σgil . . . gi1 : 1 ≤ l ≤ n, σ ∈ Σ}∪{gi−1 . . . gi−1 1 l Using property (F ) for Γ y R we find a vertex x ∈ V (R)\ Fe such that x ≁ u for all u ∈ Fe and gx 6= x for all g ∈ S. In particular, the sets ΣF (resp. α(ΣF )) and Σx, Σgil . . . gi1 x for 1 ≤ l ≤ n are pairwise distincts the only F vertices on the induced Fnand, since Γ y R is non-singular, ′ subgraph on Y := ΣF ⊔( l=1 Σgil . . . gi1 x)⊔Σx (resp. Y := Σα(F )⊔( nl=1 Σgil . . . gi1 x)⊔Σx) are the ones with extremities in ΣF (resp. Σα(F )). Hence, the bijection γ0 : Y → Y ′ defined by γ0 |ΣF = α|ΣF and γ0 |Y \ΣF = id is a graph isomorphism between the induced subgraphs. By construction and since α ∈ Z we have γ0 σ = σγ0 for all σ ∈ Σ. By Proposition 2.5 there exists an extension γ ∈ Z of γ0 . Then γ|F = α|F and πγ (gin . . . gi1 )x = gin . . . gi1 x 6= x, since gin . . . gi1 ∈ S. Hence γ ∈ Vg .  End of the proof of the Theorem. The first part of the Theorem follows from Lemmas 3.2 and 3.3 since O = U ∩ V . The last part follows from the first part and Corollary 2.12.  We have a similar result when only one of the factors in the free product is infinite. Theorem 3.4. Suppose that Γ2 is finite such that [Γ2 : Σ] ≥ 2. If Γ y R is non-singular and has property (F ), Σ is highly core-free w.r.t. Γ1 y R and the action Γ2 y R is free then the set O = {α ∈ Z : πα is homogeneous and faithful} is a dense Gδ in Z. In particular, for every countably infinite group Γ1 , for every finite nontrivial group Γ2 we have Γ1 ∗ Γ2 ∈ HR and, for any commun finite subgroup Σ < Γ1 , Γ2 such that Σ is highly core-free in Γ1 and [Γ2 : Σ] ≥ 2 we have Γ1 ∗ Γ2 ∈ HR . Σ 14 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER Proof. We first prove the analogue of Lemma 3.2. Lemma 3.5. If Σ is highly core-free w.r.t. Γ1 y R, Γ2 is finite and Γ2 y R is free and non-singular such that [Γ2 : Σ] ≥ 2, then the set U = {α ∈ Z : πα is homogeneous} is a dense Gδ in Z. Proof. Since U = ∩ϕ∈P (R) Uϕ , where Uϕ = {α ∈ Z : ∃g ∈ Γ such that πα (g)|d(ϕ) = ϕ} is obviously open, it suffices to show that Uϕ is dense for all ϕ ∈ P (R). Let ϕ ∈ P (R), α ∈ Z and F ⊂ V (R) a finite subset. Since Σ is highly core-free w.r.t. Γ1 y R, there exists g1 ∈ Γ1 such that g1 d(ϕ) ∩ ΣF = ∅, g1 x ≁ u and σg1 x ≁ g1 x′ for all u ∈ ΣF , x, x′ ∈ d(ϕ), σ ∈ Σ \ {1} and Σg1 x ∩ Σg1 x′ = ∅ for all x, x′ ∈ d(ϕ) with x 6= x′ . There exists also g2 ∈ Γ1 such that g2−1 r(ϕ) ∩ (ΣF ⊔ Σg1 d(ϕ)) = ∅ and g2−1 y ≁ u , σg2−1 y ≁ g2−1 y ′ for all u ∈ ΣF ⊔ Σg1 d(ϕ), y, y ′ ∈ r(ϕ), σ ∈ Σ \ {1} and Σg2−1 y ∩ Σg2−1 y ′ = ∅ for all y, y ′ ∈ r(ϕ) with y 6= y ′ . In the sequel we write d(ϕ) = {x1 , . . . , xn }. Claim. There exists pairwise distinct vertices z1 , . . . , zn ∈ V (R) \ Γ2 α(F ) such that • • • • • Γ2 zi ∩ Γ2 zj = ∅, ∀1 ≤ i 6= j ≤ n; Σhzi ∩ Σzi = ∅, ∀h ∈ Γ2 \ Σ, ∀ 1 ≤ i ≤ n; hzi ≁ zj , ∀h ∈ Γ2 \ {1}, ∀1 ≤ i, j ≤ n; xi ∼ xj ⇔ zi ∼ zj , ∀1 ≤ i, j ≤ n; zi ≁ u, ∀u ∈ Γ2 α(F ), ∀1 ≤ i ≤ n. Proof of the Claim. We define inductively the vertices z1 , . . . , zn ∈ V (R) \ Γ2 α(F ). For n = 1 we take z1 ∈ R∅,Γ2 α(F ) \ Γ2 α(F ). Then z1 ∈ / Γ2 α(F ) and, for all u ∈ Γ2 α(F ) and for all h ∈ Γ2 , u ≁ z1 and hz1 ≁ z1 since Γ2 y R is non-singular. Moreover, since Γ2 y R is free, we have hz1 6= σzi for all h ∈ Γ2 \ Σ and all σ ∈ Σ. Suppose that, for a given 1 ≤ l ≤ n − 1, we have pairwise distinct vertices z1 , . . . , zl ∈ V (R) \ Γ2 α(F ) such that Γ2 zi ∩ Γ2 zj = ∅, ∀1 ≤ i 6= j ≤ l, Σhzi ∩ Σzi = ∅, ∀ h ∈ Γ2 \ Σ, ∀ 1 ≤ i ≤ l, hzi ≁ zj , ∀h ∈ Γ2 \ {1}, xi ∼ xj ⇔ zi ∼ zj and zi ≁ u, ∀u ∈ Γ2 α(F ) ∀1 ≤ i, j ≤ l. Define U = {zi : 1 ≤ i ≤ l s.t. xi ∼ xl+1 } and V = {zi : 1 ≤ i ≤ l s.t. xi ≁ xl+1 } ⊔ {hzi : 1 ≤ i ≤ l, h ∈ Γ2 \ {1}} ⊔ Γ2 α(F ). Since RU,V is infinite and Γ2 is finite, we may take   zl+1 ∈ RU,V \ Γ2 α(F ) ⊔ ⊔li=1 Γ2 zi . Then, for all u ∈ Γ2 α(F ), u ≁ zl+1 and, for all 1 ≤ i ≤ l, h ∈ Γ2 \ {1}, xi ∼ xl+1 ⇔ zi ∼ zl+1 , hzi ≁ zl+1 and, for all i 6= l + 1, Γ2 zi ∩ Γ2 zl+1 = ∅. Also, since Γ2 y R is free, we have Σhzl+1 ∩ Σzl+1 = ∅ for all h ∈ Γ2 \ Σ. This completes the proof. End of the proof of the Lemma. Write d(ϕ) = {x1 , . . . , xn } and define, for 1 ≤ k ≤ n, yk = ϕ(xk ). Let z1 , . . . zn be the elements obtained by the Claim. Take h ∈ Γ2 \ Σ. Then the sets α(ΣF ), Σzi for 1 ≤ i ≤ n, and Σhzi for 1 ≤ i ≤ n are pairwise disjoint. Moreover, u ≁ σzi , u ≁ σhzi and σ ′ zi ≁ σhzj for all u ∈ α(ΣF ), for all σ, σ ′ ∈ Σ and for all 1 ≤ i, j ≤ n. Define  A = ΣF ⊔ (⊔ni=1 Σg1 xi ) ⊔ ⊔ni=1 Σg2−1 yi and B = α(ΣF ) ⊔ (⊔ni=1 Σzi ) ⊔ (⊔ni=1 Σhzi ) HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH 15 and consider the induced graph structure on A and B. Note that ΣA = A , ΣB = B and the only vertices in A (resp. B) are the ones with extremities in ΣF (resp. α(ΣF )). Since Σ y R is free, we may define a bijection γ0 : A → B by γ0 (u) = α(u) for u ∈ ΣF and γ0 (σg1 xi ) = σzi , γ0 (σg2−1 yi ) = σhzi for all 1 ≤ i ≤ n and for all σ ∈ Σ which is a graph isomorphism satisfying γ0 σ = σγ0 for all σ ∈ Σ. By Proposition 2.5, there exists an extension γ ∈ Z of γ0 . Then γ|F = α|F and with g = g2 hg1 ∈ Γ we have, for all 1 ≤ i ≤ n, πγ (g)xi = g2 γ −1 hγg1 xi = g2 γ −1 hzi = g2 g2−1 yi = yi .  End of the proof of the Theorem 3.4. The first assertion of the Theorem follows from Lemmas 3.5 and 3.3. The last part follows from the first part and Corollary 2.12, where the group Σ in Corollary 2.12 is actually our group Γ2 , the group Σ′ is our group Σ and the group H ′ is our group Γ1 .  4. Actions of HNN extensions on the Random Graph Let Σ < H be a finite subgroup of a countable group H and θ : Σ → H be an injective group homomorphism. Define Γ = HNN(H, Σ, θ) the HNN-extension and let t ∈ Γ be the “stable letter” i.e. Γ is the universal group generated by Σ and t with the relations tσt−1 = θ(σ) for all σ ∈ Σ. For ǫ ∈ {−1, 1}, we write  Σ if ǫ = 1, Σǫ := θ(Σ) if ǫ = −1. Suppose that we have a faithful action Γ y R and view Γ < Aut(R). Define the closed (hence Polish space) subset Z = {α ∈ Aut(R) : θ(σ) = ασα−1 for all σ ∈ Σ} ⊂ Aut(R) and note that it is non-empty (since t ∈ Z). By the universal property of Γ, for each α ∈ Z there exists a unique group homomorphism πα : Γ → Aut(R) such that πα |H = idH and πα (t) = α. In this section we prove the following result. Theorem 4.1. If Γ y R is non-singular, has property (F ), Σǫ y R is free and Σǫ is highly core-free w.r.t. H y R for all ǫ ∈ {−1, 1} then the set O = {α ∈ Aut(R) : πα is faithful and homogeneous} is a dense Gδ in Z. In particular, for any finite subgroup Σ of an infinite countable group H such that Σǫ < H is highly core-free for all ǫ ∈ {−1, 1}, we have HNN(H, Σ, θ) ∈ HR . We separate the proof in two lemmas. Lemma 4.2. If, for all ǫ ∈ {−1, 1}, Σǫ y R is free, non-singular and Σǫ is highly core-free w.r.t. H y R then the set U = {α ∈ Aut(R) : πα is homogeneous} is a dense Gδ in Z. Proof. Since U = ∩ϕ∈P (R) Uϕ , where Uϕ = {α ∈ Z : ∃g ∈ Γ such that πα (g)|d(ϕ) = ϕ} is obviously open, it suffices to show that Uϕ is dense for all ϕ ∈ P (R). Let ϕ ∈ P (R), α ∈ Z and F ⊂ V (R) a finite subset. It suffices to show that there exists γ ∈ Z and g ∈ Γ such that γ|F = α|F and πγ (g)|d(ϕ) = ϕ. Since Σ is highly core-free w.r.t. H y R, there exists g1 ∈ H such that g1 d(ϕ) ∩ ΣF = ∅ and g1 x ≁ u, σg1 x ≁ g1 x′ for all u ∈ ΣF , x, x′ ∈ d(ϕ), σ ∈ Σ \ {1} and Σg1 x ∩ Σg1 x′ = ∅ for all x, x′ ∈ d(ϕ) with x 6= x′ . Since θ(Σ) is highly core-free w.r.t. H y R, there exists g2 ∈ H such that g2−1 r(ϕ) ∩ (α(ΣF ⊔ Σg1 d(ϕ))) = ∅ 16 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER and g2−1 y ≁ u, θ(σ)g2−1 y ≁ g2−1 y ′ for all u ∈ α(ΣF ⊔ Σg1 d(ϕ)), y, y ′ ∈ r(ϕ), σ ∈ Σ \ {1} and θ(Σ)g2−1 y ∩ θ(Σ)g2−1 y ′ = ∅ for all y, y ′ ∈ r(ϕ) with y 6= y ′ . In the sequel we write d(ϕ) = {x1 , . . . , xn } and yi = ϕ(xi ). Define Y = (⊔ni=1 Σg1 xi )⊔(⊔ni=1 Σα−1 (g2−1 yi ))⊔(ΣF ) and note that α(Y ) = (⊔ni=1 θ(Σ)α(g1 xi )) ⊔ (⊔ni=1 θ(Σ)(g2−1 yi )) ⊔ (θ(Σ)α(F )). Note that ΣY = Y and θ(Σ)α(Y ) = α(Y ). Define the bijection γ0 : Y → α(Y ) by γ0 (σg1 xi ) = θ(σ)g2−1 yi , γ0 (σα−1 (g2−1 yi )) = θ(σ)α(g1 xi ) for all 1 ≤ i ≤ n, σ ∈ Σ and γ0 |ΣF = α|ΣF . By construction γ0 is a graph isomorphism such that γ0 σ = θ(σ)γ0 for all σ ∈ Σ. By Proposition 2.5 there exists an extension γ ∈ Z of γ0 . Note that γ|F = α|F moreover, with g = g2 tg1 ∈ Γ we have, for all 1 ≤ i ≤ n, πγ (g)xi = g2 γg1 xi = g2 γ0 (g1 xi ) = g2 g2−1 yi = yi .  Lemma 4.3. If, for all ǫ ∈ {−1, 1}, Σǫ y R is free and Γ y R is non-singular and has property (F ) then the set V = {πα : πα is faithful} is a dense Gδ in Z. T Proof. Since V = g∈Γ\{1} Vg , where Vg = {α ∈ Z : πα (g) 6= id} is clearly open, it suffices to show that Vg is dense for all g ∈ Γ \ {1}. We may and will assume that g ∈ / H, since when g ∈ H we have Vg = Z. Write g = hn tǫn . . . tǫ1 h0 a reduced expression for g, where n ≥ 1, hk ∈ H and ǫk ∈ {−1, 1}, ∀1 ≤ k ≤ n. Fix α ∈ Z and F ⊂ V (R) be a finite subset. Define H1 =  Σh0 if ǫ1 = 1 Σt−1 h0 if ǫ1 = −1 e1 = and H  θ(Σ)th0 if ǫ1 = 1 θ(Σ)h0 if ǫ1 = −1. For 2 ≤ l ≤ n, define   if ǫl = 1 θ(Σ)thl−1 tǫl−1 . . . tǫ1 h0 if ǫl = 1 Σhl−1 tǫl−1 . . . tǫ1 h0 e and H = Hl = l θ(Σ)hl−1 tǫl−1 . . . tǫ1 h0 if ǫl = −1. Σt−1 hl−1 tǫl−1 . . . tǫ1 h0 if ǫl = −1 −1 e := ∪n H e e−1 e Let G := ∪nl=1 Hl , G l=1 l and S := G G ∪ G G ∪ {g} ⊂ Γ be finite subsets of Γ and Fe := ∪s∈G∪Ge s−1 (ΣF ∪ α(ΣF )) ⊂ V (R) a finite subset of V (R). Since Γ y R has property (F ) there exists x ∈ V (R) \ Fe such that x ≁ u for all u ∈ Fe and sx 6= x for all e l x, the sets ΣF and Yl (resp. s ∈ S \ {1}. In particular, defining Yl := Hl x and Yel := H α(ΣF ) and Yel ) for 1 ≤ l ≤ n are pairwise disjoint and, using moreover the fact Fnthat Γ y R is non-singular, the only vertices on the induced subgraph on Y := ΣF ⊔ ( l=1 Yl ) (resp. F on Ye := θ(Σ)α(F ) ⊔ ( nl=1 Yel )) are the ones with extremities in ΣF (resp. in θ(Σ)α(F )). It implies that the bijection γ0 : Y → Ye defined by γ0 |ΣF = α|ΣF and, for all 1 ≤ l ≤ n, γ0 |Yl = t|Yl is actually a graph isomorphism between the induced subgraphs. Since we clearly have γ0 σ = θ(σ)γ0 for all σ ∈ Σ, there exists, by Proposition 2.5 an extension γ ∈ Z of γ0 . Then γ satisfies γ|F = α|F and πγ (g)x = hn γ ǫn . . . γ ǫ1 h0 x = hn tǫn . . . tǫ1 h0 x = gx 6= x since g ∈ S. It follows that γ ∈ Vg .  End of the proof of Theorem 4.1. The first assertion follows directly from Lemmas 4.2 and 4.3 since O = U ∩ V and the last part follows from the first part and Corollary 2.12. 5. Actions of groups acting on trees on the random graph Let Γ be a group acting without inversion on a non-trivial tree. By [Se77], the quotient graph G can be equipped with the structure of a graph of groups (G, {Γp }p∈V(G) , {Σe }e∈E(G) ) where each Σe = Σe is isomorphic to an edge stabilizer and each Γp is isomorphic to a vertex stabilizer and such that Γ is isomorphic to the fundamental group π1 (Γ, G) of this graph of HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH 17 groups i.e., given a fixed maximal subtree T ⊂ G, the group Γ is generated by the groups Γp for p ∈ V(G) and the edges e ∈ E(G) with the relations e = e−1 , se (x) = ere (x)e−1 , ∀x ∈ Σe and e = 1 ∀e ∈ E(T ), where se : Σe → Γs(e) and re = se : Σe → Γr(e) are respectively the source and range group monomomorphisms. Theorem 5.1. If Γp is countably infinite, for all p ∈ V(G), Σe is finite and se (Σe ) is highly core-free in Γs(e) , for all e ∈ E(G), then Γ ∈ HR . Proof. Let e0 be one edge of G and G ′ be the graph obtained from G by removing the edges e0 and e0 . Case 1: G ′ is connected. It follows from Bass-Serre theory that Γ = HNN(H, Σ, θ) where H is fundamental group of our graph of groups restricted to G ′ , Σ = re0 (Σe0 ) < H is a subgroup and θ : Σ → H is given by θ = se0 ◦ re−1 . By hypothesis H is countably infinite, 0 Σ is finite and, since Σ < Γr(e0 ) (resp. θ(Σ) < Γs(e0 ) ) is a highly core-free subgroup, Σ < H (resp. θ(Σ) < H) is also a highly core-free subgroup. Thus we may apply Theorem 4.1 to conclude that Γ ∈ HR . Case 2: G ′ is not connected. Let G1 and G2 be the two connected components of G ′ such that s(e0 ) ∈ V(G1 ) and r(e0 ) ∈ V(G2 ). Bass-Serre theory implies that Γ = Γ1 ∗Σe0 Γ2 , where Γi is the fundamental group of our graph of groups restricted to Gi , i = 1, 2, and Σe0 is viewed as a highly core-free subgroup of Γ1 via the map se0 and as a highly core-free subgroup of Γ2 via the map re0 since se0 (Σe0 ) is highly core-free in Γs(e0 ) and re0 (Σe0 ) is highly core-free in Γr(e0 ) by hypothesis. Since Γ1 and Γ2 are countably infinite and Σe0 is finite, we may apply Theorem 3.1 to conclude that Γ ∈ HR .  6. Actions of free groups on the random graph Recall that P (R) denotes the set of isomorphisms of R between finite induced subgraphs d(ϕ), r(ϕ) ⊂ V (R) and note that P (R) has a natural structure of groupoid. In this Section, we prove Theorem D. The main tool, called an elementary extension, is a refinement of the “back and forth” method used to extend any partial isomorphism ϕ ∈ P (R) to an automorphism of R. 6.1. Elementary extensions. Let Φ ⊂ P (R) and F ⊂ V (R) be finite subsets. Let γ ∈ Φ. We construct a partial isomorphism γ̃ ∈ P (R) which extends γ as follows: first, we set [  d(ϕ) ∪ r(ϕ) , D := K \ d(γ), R := K \ r(γ). K := F ∪ ϕ∈Φ Note that D and R have the same cardinality and write D = {x1 , . . . , xm } and R = {v1 , . . . , vm }. We then successively find vertices u1 , . . . , um ∈ V (R) \ K such that: • the vertex uj is adjacent to all vertices u ∈ d(γ) such that vj ∼ γ(u) and all vertices uj ′ , with j ′ < j, such that vj ′ ∼ vj ; • it is not adjacent to any other vertices in K. After that, we successively find vertices y1 , . . . , ym ∈ V (R) \ (K ∪ {u1 , . . . , um }) such that: • the vertex yi is adjacent to all vertices y ∈ r(γ) such that xj ∼ γ −1 (y) and all vertices yi′ , with i′ < i, such that xi′ ∼ xi ; 18 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER • it is not adjacent to any other vertices in K, nor to vertices u1 , . . . , um . Finally, since the sets D = {x1 , . . . , xm }, d(γ), {u1 , . . . , um } are pairwise disjoint, and the sets {y1 , . . . , ym }, r(γ), R = {v1 , . . . , vm } are pairwise disjoint, we can define a partial bijection γ̃ : K ⊔ {u1 , . . . , um } → {y1 , . . . , ym } ⊔ K which extends γ by setting γ̃(xi ) = yi for all i = 1, . . . , m, and γ̃(z) = γ(z) for all z ∈ d(γ), and γ̃(uj ) = vj for all j = 1, . . . , n. Definition 6.1. The above γ̃ is called an elementary extension of the triple (γ, Φ, F ). Proposition 6.2. Any elementary extension γ̃ of (γ, Φ, F ) has the following properties: (1) γ̃ ∈ P (R), that is, γ̃ is an isomorphism between finite induced subgraphs; (2) F ⊂ d(γ̃) ∩ r(γ̃), and for all ϕ ∈ Φ, d(ϕ) ∪  r(ϕ) ⊂ d(γ̃) ∩ r(γ̃); (3) The sets γ̃ −1 d(γ) \ r(γ) , γ̃ r(γ) \ d(γ)  and K are pairwise disjoint and u ≁ v, ∀u ∈ γ̃ −1 d(γ) \ r(γ) , ∀v ∈ γ̃ r(γ) \ d(γ) ; S e e := (Φ \ {γ}) ∪ {e (4) Let Φ γ }. For any subset A of ϕ∈Φe (d(ϕ) ∪ r(ϕ)) let Ω (resp. Ω) e Then, any edge of R be the orbit of A under the groupoid generated by Φ (resp. Φ). −1 e contained in Ω \ Ω is the image under γ e or γ e of an edge in Ω. Proof. (1) It is clear from the construction that K is finite, as it is a finite union of finite sets. Hence, d(γ̃) = K ⊔ {u1 , . . . , um } and r(γ̃) = {y1 , . . . , ym } ⊔ K are finite. Let us now check that γ̃ is an isomorphism between finite induced subgraphs. Notice that one has d(γ̃) = {x1 , . . . , xm } ⊔ d(γ) ⊔ {u1 , . . . , um } and let u, x ∈ d(γ̃). We are going to check that u ∼ x ⇔ γ̃(u) ∼ γ̃(x) by distinguishing cases. If u and x are in the same component of the disjoint union {x1 , . . . , xm } ⊔ d(γ) ⊔ {u1 , . . . , um }, the equivalence follows readily from the construction and the fact that γ ∈ P (R). If x ∈ d(γ) and u = uj , then x ∼ uj ⇔ γ(x) ∼ vj ⇔ γ̃(x) ∼ γ̃(uj ) by the selection of the uj ’s. If x = xi and u ∈ d(γ), then xi ∼ u ⇔ yi ∼ γ(u) ⇔ γ̃(xi ) ∼ γ̃(u) again by the choice of the yi ’s. Finally, if x = xi and u = uj , then γ̃(uj ) = vj and γ̃(xi ) = yi . It follows from the selection of the uj ’s and yi ’s that uj 6∼ xi and vj 6∼ yi . Hence the equivalence holds. (2) It is clear since F ∪ d(ϕ) ∪ r(ϕ) for ϕ ∈ Φ are contained in K and K ⊂ d(e γ ) ∩ r(e γ ).   (3) Notice that γ̃ −1 d(γ) \ r(γ) = {u1 , . . . , um } and γ̃ r(γ) \ d(γ) = {y1 , . . . , ym }. Therefore the assertion follows from the construction. e\Ω ⊂ γ e 2 \ Ω2 such that x ∼ y. (4) Note that Ω e−1 (Ω \ r(γ)) ∪ e γ (Ω \ d(γ)). Let (x, y) ∈ Ω −1 If x, y ∈ γ e(Ω \ d(γ)) or x, y ∈ e γ (Ω \ r(γ)) the result is trivial. Note that we cannot have x∈γ e(Ω \ d(γ)) and y ∈ γ e−1 (Ω \ r(γ)) because otherwise there exists i and j such that x = yi and y = uj which implies that x ≁ y. Suppose now that x ∈ e γ (Ω \ d(γ)) and y ∈ Ω. Then there is some i such that x = yi but then yi ∼ y implies that y ∈ r(γ) and γ e−1 (y) ∼ e γ −1 (x). −1 −1 −1 −1 Since γ e (y) = γ (y) ∈ Ω and e γ (x) = γ e (yi ) = xi ∈ Ω, we are done. The other cases are proved in the same way.  6.2. “Treezation” of a free group action. We denote by Fk the free group on k generators a1 , . . . , ak . Given a tuple ᾱ = (α1 , . . . , αk ) ∈ Aut(R)k , we denote by α : Fk → Aut(R) the unique group homomorphism such that α(aj ) = αj for all j. HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH 19 In this section, given such a k-tuple ᾱ ∈ Aut(R)k and a finite set F of vertices of R, we explain how to get a k-tuple β̄ ∈ Aut(R)k such that, for all j and ε = ±1, the automorphisms αεj and βjε coincide on F , and, informally speaking, the associated Schreier graph (see definition below) looks like a tree far from F . First, we define, for 1 ≤ j ≤ k, the element β0,j ∈ P (R) by the restriction: β0,j := αj |α−1 (F )∪F : α−1 j (F ) ∪ F → F ∪ αj (F ). j   Sk S Write Fe = kj=1 α−1 j (F ) ∪ F ∪ αj (F ) = j=1 (d(β0,j ) ∪ r(β0,j )). Let B0 ⊂ P (R) be the groupoid generated by β0,j for 1 ≤ j ≤ k and note that, for all ϕ ∈ B0 , one has d(ϕ)∪r(ϕ) ⊂ Fe. Remark 6.3. Suppose that α has all orbits infinite. It implies that for all x ∈ Fe there exists g ∈ B0 and 1 ≤ j ≤ k such that x ∈ d(g) and gx ∈ / d(β0,j ) ∩ r(β0,j ). Indeed, since the α-orbit of x is infinite, there exists w ∈ Fk such that α(w)x ∈ / Fe . Write w = αǫinn . . . αǫi11 its reduced ǫ1 ǫn expression, where ǫl ∈ {−1, 1}. Since α(w)x ∈ / Fe it follows that x ∈ / d(β0,i . . . β0,i ). We may n 1 ǫ1 and will assume that x ∈ d(β0,i ) (otherwise the conclusion is obvious). Let 1 ≤ n 0 ≤ n−1 1 ǫn0 ǫn0 ǫ1 ǫ1 be the largest integer such that x ∈ d(β0,in . . . β0,i1 ). Defining g = β0,in . . . β0,i1 ∈ B0 and 0 0 j = in0 +1 we reach the conclusion. Let V (R) = {z0 , z1 , . . . } be an enumeration of V (R) and define inductively β̄l ∈ P (R)k by β̄0 = (β0,1 , . . . , β0,k ) and, if for some l ≥ 0 the element β̄l = (βl,1 , . . . , βl,k ) is defined, we denote by j(l) the unique element in {1, . . . k} such that j(l) ≡ l + 1 (mod k) and we define β̄l+1 = (βl+1,1 , . . . , βl+1,k ) where βl+1,j(l) is the elementary extension of the triple (βl,j(l) , {βl,1 , . . . , βl,k }, {zl }) and βl+1,j = βl,j for j 6= j(l). We will denote by Bl ⊂ P (R) the groupoid generated by βl,1 , . . . , βl,k . Define β̄ = (β1 , . . . , βk ) ∈ Aut(R)k , where βj is (well-) defined by βj (x) = βj,l (x) whenever x ∈ d(βj,l ) (and denote by β : Fk → Aut(R) the unique group morphism which maps the aj onto βj ). Definition 6.4. The k-tuple β̄ is said to be a treezation of ᾱ relatively to F . We denote by Gβ the Schreier graph of β: the vertices are V (Gβ ) = V (R) and for all x ∈ V (Gβ ) we have an oriented edge from x to βj (x) which is decorated by j + and an oriented edge from βj−1 (x) to x decorated by j − . More generally, given a groupoid H ⊂ P (R) generated by γ1 , . . . , γk ∈ P (R), we define GH the Schreier graph of H with the generating tuple (γ1 , ..., γk ) in the following way: V (GH ) = V (R) and, for all x ∈ V (GH ), there is an oriented edge decorated by j + from x to γj (x) whenever x ∈ d(γj ) and there is an oriented edge decorated by j − from γj−1 (x) to x whenever x ∈ r(γj ). For all l ≥ 0, we will denote by Gl the Schreier graph of Bl with the generating tuple β̄l . Given a graph G, for l ≥ 2, a minimal path in G from x1 ∈ V (G) to xl ∈ V (G) is a finite sequence of pairwise distinct vertices x1 , . . . xl such that xi ∼ xi+1 for all 1 ≤ i ≤ l − 1. When l ≥ 3 and x1 ∼ xl , we call it a minimal cycle. Recall that a1 , . . . ak are the canonical generators of Fk . Proposition 6.5. Any treezation β̄ of ᾱ satisfies the following properties. 20 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER For all l ≥ 0 the minimal cycles in Gl are all contained in Fe. The minimal cycles of Gβ are all in Fe. The minimal paths of Gβ with extremities in Fe are contained in Fe. If the orbits of α are infinite then the orbits of β are infinite. For all x, y ∈ β(Fk )(Fe), if (x, y) ∈ E(R) then there exists x0 , y0 ∈ Fe such that (x0 , y0 ) ∈ E(R) and w ∈ Fk such that β(w)x0 = x and β(w)(y0 ) = y. (6) Let x ∈ Fe and w = aǫinn . . . aǫi11 ∈ Fk with its reduced expression, where ǫl ∈ {−1, 1}. If there exists s such that βiǫss . . . βiǫ11 x ∈ / Fe, then for all s < t ≤ n the path (1) (2) (3) (4) (5) ǫ s+1 βiǫss . . . βiǫ11 x, βis+1 . . . βiǫ11 x, . . . , βiǫnn . . . βiǫ11 x ǫt−1 is a geodesic path in Gβ and d(βiǫtt . . . βiǫ11 x, Fe) = 1 + d(βit−1 . . . βiǫ11 x, Fe ) for all s < t ≤ n, where d is the graph distance. Proof. (1) The result is obvious for G0 since all the edges in G0 have their source and range in Fe. Now, observe that the edges of Gl are included in the edges of Gl+1 with the same label and any edge of Gl+1 which is not already an edge on Gl has an extremity which is not in any domain or range of any βl,j for 1 ≤ j ≤ k since βl+1,j(l) is an elementary extension of βl,j(l) and for j 6= j(l) one has βl+1,j = βj,l . Hence, the minimal cycles in Gl+1 are contained in Gl . This proves the result by induction on l. (2) It follows from (1) since any minimal cycle in Gβ is a minimal cycle in Gl for some l ≥ 0. (3) The proof is the same as the one of (1) and (2). (4) It follows from (2) that the induced graph structure on the complement of Fe coming from Gβ is a forest. Hence, every x ∈ / Fe has an infinite β-orbit. If x ∈ Fe, we find, by Remark 6.3, g ∈ B0 and 1 ≤ j ≤ n such that x ∈ d(g) and gx ∈ / d(β0,j ) ∩ r(β0,j ). Hence either βk,j (gx) ∈ / Fe −1 or βk,j (gx) ∈ / Fe. In both cases, we find an element w ∈ Fk such that β(w)(x) ∈ / Fe . By the first part of the proof, the β-orbit of x is infinite. (5) Let Ω be the orbit of Fe under the action β and, for l ≥ 0, Ωl be the orbit of Fe under the groupoid Bl . Since every edge of R which is in Ω is actually in Ωl for some l, it suffices to show that, for all l ≥ 0 and all x, y ∈ Ωl with (x, y) ∈ E(R), there exists g ∈ Bl and x0 , y0 ∈ Fe such that gx0 = x and gy0 = y. For l = 0 it is trivial since Ω0 = Fe and the proof follows by induction by using assertion (4) of Proposition 6.2. (6) The proof is a direct consequence of the following remark, which is itself a direct consequence of (2) and (3). For all x ∈ / Fe in the connected component of a point in Fe there exist a unique path in Gβ without backtracking x1 , x2 , . . . , xl such that • x1 = x, • x2 , . . . , xl−1 ∈ / Fe, • xl ∈ Fe. Moreover, this path is geodesic.  6.3. Proof of Theorem D. Let us fix an integer k ≥ 2. We still denote by Fk the free group on k generators a1 , . . . , ak , and by α the morphism Fk → Aut(R) associated to some ᾱ ∈ Aut(R)k . Let us mention that one has hᾱi = {α(w) : w ∈ Fk }. The set A = {ᾱ ∈ Aut(R)k : every hαi-orbit on the vertices is infinite} HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH 21 is closed in Aut(R)k , which is closed on S(R)k . Thus A is a Baire space. We also consider the subsets F = {ᾱ ∈ Aut(R)k : α is injective }; H = {ᾱ ∈ Aut(R)k : α is an homogeneous action on Aut(R)}. Now, Theorem D can be stated as follows. Theorem 6.6. The subset A ∩ F ∩ H is a dense Gδ in A. Proof. Let us prove that A ∩ H is a dense Gδ in A. To do this, we observe that A ∩ H = T k ϕ∈P (R) (A ∩ Hϕ ), where Hϕ = {ᾱ ∈ Aut(R) : ∃w ∈ Fk , ∀x ∈ d(ϕ), α(w)x = ϕx}. Since the A ∩ Hϕ ’s are all open because the Hϕ ’s are open in Aut(R)k , it suffices to prove that A ∩ Hϕ is dense for all ϕ ∈ P (R). Take an arbitrary ϕ ∈ P (R) and an arbitrary ᾱ = (α1 , . . . , αk ) ∈ A. We need to prove that, ±1 for any finite subset F ⊂ V (R), there exists ω̄ ∈ A ∩ Hϕ such that α±1 j and ωj coincide on F for all j. To prove this, we may assume that F contains both d(ϕ) and r(ϕ). S We take such an F and we set F̃ = kj=1 (α−1 j (F ) ∪ F ∪ αj (F )) as we did in Section 6.2. Then we take a treezation β̄ of ᾱ relatively to F . Let us recall that β has all orbits infinite since α does, and that, for all j, the automorphisms βj±1 and α±1 j coincide on F . By Neumann’s Lemma, there exists u ∈ Fk (we consider it as a reduced word) such that β(u)F̃ ∩ F̃ = ∅. Up to replacing u by some aj u, we may assume that u is moreover cyclically reduced. This is possible thanks to Proposition 6.5 (6). Since u2 is reduced, we also have β(u2 )F̃ ∩ F̃ = ∅, again by Proposition 6.5 (6). Let us now show that F̃ and β(u2 )F̃ are disconnected, that is, for all x ∈ F̃ and y ∈ β(u2 )F̃ , we have x ≁ y. Assume by contradiction that there exsit x ∈ F̃ and y ∈ β(u2 )F̃ such that x ∼ y. By Proposition 6.5 (5), there exist x0 , y0 ∈ F̃ and v ∈ Fk such that x = β(v)x0 and y = β(v)y0 . Using Proposition 6.5 (6), we see first that v = uv ′ (reduced expression) since y = β(v)y0 is in β(v)F̃ ∩ β(u2 )F̃ , and then that β(v)x0 = β(u)β(v ′ )x0 is not in F̃ . Indeed, if β(v ′ )x0 ∈ F̃ , then β(u)β(v ′ )x0 is not in F̃ , and if β(v ′ )x0 ∈ / F̃ , then β(u)β(v ′ )x0 is farer from ′ F̃ than β(v )x0 in the Schreier graph of β. Since we have β(v)x0 ∈ / F̃ , x ∈ F̃ and x = β(v)x0 , we have obtained a contradiction. S We now set K = v β(v)F̃ , where the union is taken over the prefixes of u2 (it is the union of trajectories (see Definition 3.1 in [MS13]) of points in F̃ in the Schreier graph of β) and S K̃ = kj=1 (βj−1 (K) ∪ K ∪ βj (K)). We also consider some integer s > 10 diam(K̃), with respect to the metric on the Schreier graph of β, and an index i such that the first letter of u −s 2 s 2 is different from a±1 i . Proposition 6.5 ensures that β(ai u )F̃ , K̃ and β(ai u )F̃ are pairwise disjoint. 2 s 2 Moreover, β(a−s i u )F̃ and β(ai u )F̃ are both disconnected from K̃. Indeed, assume by contradiction that there exist x ∈ K̃ and y ∈ β(asi u2 )F̃ such that x ∼ y. By Proposition 6.5 (5), there exist x0 , y0 ∈ F̃ and v ∈ Fk such that x = β(v)x0 and y = β(v)y0 . On the one hand, using Proposition 6.5 (6), we see first that v = asi uv ′ (reduced expression) since y = β(v)y0 is in β(v)F̃ ∩ β(asi u2 )F̃ , and then that the distance (in the Schreier graph of β) between β(v)x0 and F̃ is greater than 10 diam(K̃). Indeed, we get, as above, that β(uv ′ )x0 ∈ / F̃ , and 22 PIERRE FIMA, SOYOUNG MOON, AND YVES STALDER deduce that the distance between β(v)x0 = β(asi )β(uv ′ )x0 and F̃ is greater than s, which is itself greater than 10 diam(K̃). On the other hand, since x is in K̃, its distance is less or equal to diam(K̃). Hence, we get x 6= β(v)x0 , a contradiction. This proves that there are no edges between K̃ and β(asi u2 )F̃ . One gets similarly that there are no edges between K̃ and 2 β(a−s i u )F̃ . Thus we can define a partial graph isomorphism τ ∈ P (R) as follows: we fix some index i′ distinct from i (which exists since k ≥ 2) and set: 2 s 2 r(τ ) = (K ∪ βi′ K) ⊔ β(a−s • d(τ ) = (βi−1 ′ K ∪ K) ⊔ β(ai u )d(ϕ); i u )r(ϕ); • for x ∈ βi−1 ′ K ∪ K, we set τ x = βi′ x; 2 −2 −s • for x ∈ β(asi u2 )d(ϕ), we set τ x = β(a−s i u ) ◦ ϕ ◦ β(u ai )x. 2 s 2 Note that without the fact that β(a−s i u )F̃ and β(ai u )F̃ are both disconnected from K̃, the map τ could be only a bijection between vertices. Now, one can extend τ to an automorphism ωi′ , and set ωj = βj for other indices j. Then, −s 2 −s 2 for all j, ωj±1 , βj±1 and α±1 j coincide on F . Moreover β(ai u ) and ω(ai u ) coincide on F , and the same holds for β(asi u2 ) and ω(asi u2 ). Hence, one has −s 2 −s 2 −2 −s ω(u−2 a−s i ai′ ai u )x = β(u ai )τ β(ai u )x s 2 −2 s −s 2 = β(u−2 a−s i )β(ai u )ϕβ(u ai )β(ai u )x = ϕx for all x ∈ d(ϕ). In other words, ω̄ = (ω1 , . . . , ωk ) is in Hϕ . Finally, as ωi′ coincides with βi′ on βi−1 ′ K ∪ K and all βi -orbits are infinite outside this set, ω also has all orbits infinite, that is ω̄ is in A. Consequently, we obtain that ω is in A ∩ Hϕ , which concludes the proof that A ∩ H is dense Gδ in A. One can use a similar, but much easier, argument to show that A ∩ F is a dense Gδ in A. This is left to the reader.  References [BM05] M. Bhattacharjee and H.D. Macpherson, A locally finite dense group acting on the random graph, Forum Math. 17 (2005) 513–517. [Ca96] P. J. Cameron, Cofinitary permutation groups, Bull. London Math. Soc. 28 (1996) 1–17. [Ca97] P. J. Cameron, The random graph, The mathematics of Paul Erdos, II, Algorithms Combin. 14, Springer, Berlin (1997) 333–351. [Ca99] P.J. Cameron, Permutation groups, London Mathematical Society Student Texts. 45 Cambridge University Press, Cambridge (1999). [Dr85] M. Droste, Structure of partially ordered sets with transitive automorphism group, Mem. Amer. Math. Soc. 334 (1985). [ER63] P. Erdös and A. Rényi. Asymmetric graphs, Acta Math. Acad. Sci. Hungar. 14 (1963) 295–315. [FMS15] P. Fima, S. Moon and Y. Stalder. Hilghly transitive actions of groups acting on trees, Proc. Amer. Math. Soc. 143 (2015) 5083–5094. [GMR93] A. M. W. Glass, S. H. McCleary and M. Rubin, Automorphism groups of countable highly homogeneous partially ordered sets, Math. Z. 214 (1993) 55–66. [GK03] P.M. Gartside and R.W. Knight, Ubiquity of free groups, Bull. Lond. Math. Soc. 35 (2003) 624–634. [GS15] S. Glaband and F. Strobin, Large free subgroups of automorphisms groups of ultrahomogeneous spaces, Colloq. Math. 140 (2015) 279–295. [Ka92] R. Kaye, The automorphism group of a countable recursive saturated structure, Proc. London Math. Soc. (3) 65 (1992) 225–244. [Mac11] H.D. Macpherson, A survey of homogeneous structures. Discrete Math. (15) 311 (2011) 1599–1634. [Mac86] H.D. Macpherson, Groups of automorphisms of ℵ0 -categorical structures, Quart. J. Math. (2) 37 (1986) 449–465. HOMOGENEOUS ACTIONS ON THE RANDOM GRAPH 23 [MS94] G. Melles and S. Shelah, Aut(M) has a large dense free subgroup for saturated M, Bull. London Math. Soc. 26 (1994), 339–344. [MS13] S. Moon and Y. Stalder, Highly transitive actions of free products, Algebr. Geom. Topol. 13 (2013) 589–607 [Ra64] R. Rado, Universal graphs and universal functions, Acta Arith. 9 (1964) 393–407. [Se77] J.-P. Serre, Arbres, amalgames, SL2 , Astérisque 46, Société Mathématique de France (1977). [Tr85] J. K. Truss, The group of the countable universal graph, Math. Proc. Cambridge Philos. Soc. 98 (1985) 213–245. Pierre Fima Univ Paris Diderot, Sorbonne Paris Cité, IMJ-PRG, UMR 7586, F-75013, Paris, France, Sorbonne Universités, UPMC Paris 06, UMR 7586, IMJ-PRG, F-75005, Paris, France, CNRS, UMR 7586, IMJ-PRG, F-75005, Paris, France. E-mail address: [email protected] Soyoung Moon Université de Bourgogne, Institut Mathématiques de Bourgogne, UMR 5584 CNRS, BP 47870, 21078 Dijon cedex France E-mail address: [email protected] Yves Stalder Université Clermont Auvergne, CNRS, Laboratoire de Mathmatiques Blaise Pascal, F-63000 Clermont-Ferrand, France E-mail address: [email protected]
4math.GR
Alternating Optimization for Capacity Region of Gaussian MIMO Broadcast Channels with Per-antenna Power Constraint Thuy M. Pham, Ronan Farrell, and Le-Nam Tran arXiv:1704.01473v2 [cs.IT] 17 May 2017 Department of Electronic Engineering, Maynooth University, Maynooth, Co. Kildare, Ireland Email: {minhthuy.pham, ronan.farrell, lenam.tran}@nuim.ie Abstract—This paper characterizes the capacity region of Gaussian MIMO broadcast channels (BCs) with per-antenna power constraint (PAPC). While the capacity region of MIMO BCs with a sum power constraint (SPC) was extensively studied, that under PAPC has received less attention. A reason is that efficient solutions for this problem are hard to find. The goal of this paper is to devise an efficient algorithm for determining the capacity region of Gaussian MIMO BCs subject to PAPC, which is scalable to the problem size. To this end, we first transform the weighted sum capacity maximization problem, which is inherently nonconvex with the input covariance matrices, into a convex formulation in the dual multiple access channel by minimax duality. Then we derive a computationally efficient algorithm combining the concept of alternating optimization and successive convex approximation. The proposed algorithm achieves much lower complexity compared to an existing interiorpoint method. Moreover, numerical results demonstrate that the proposed algorithm converges very fast under various scenarios. Index Terms—MIMO, minimax duality, dirty paper coding, alternating optimization, successive convex optimization. I. I NTRODUCTION Since its invention in the mid-90s [1], [2], multiple-input multiple-output (MIMO) technology has been adopted in all modern mobile wireless networks. From a system design perspective, one of the most fundamental problems is to compute the capacity of the system of interest. For a single user MIMO (SU-MIMO) channel, pioneer studies proved that the capacity can be achieved by Gaussian input signaling [1], [2]. For multiuser MIMO scenarios, the seminal work of [3] showed that dirty-paper coding (DPC) in fact achieves the entire capacity region of Gaussian MIMO broadcast channels (BCs). The particular case of the sum capacity of MIMO BCs was studied in several pioneer studies including [3]–[7], The capacity of MIMO systems is investigated along with a certain type of constraint on the input covariance matrices. In this regard, we remark that all papers mentioned above assume a sum power constraint (SPC), and this usually leads to efficiently computational algorithms. For the sum-capacity computation, Viswanathan et al. [8] applied a steepest descend method, while Yu [9] proposed a dual decomposition-based algorithm. In this line of research, Jindal et al. presented a sum power iterative water-filling algorithm by exploiting the MACBC duality. The entire capacity region of MIMO BCs with a SPC was characterized in [10], [11], using conjugate gradient projection (CGP)- and pre-conditioned gradient projectionbased approaches. While SPC has been widely considered, it is less appealing in reality due to the fact that each antenna is usually equipped with a different power amplifier, which has its own power budget. Despite its practical and fundamental importance, the research on efficient methods for computing the capacity region of Gaussian MIMO BCs has been quite limited. For SU-MIMO, this problem was solved in [12], [13] by the socalled mode-dropping algorithm. In [3], it was shown that DPC still achieves the full capacity region of the MIMO BC under PAPC. However, finding the DPC region with PAPC is more numerically difficult than with a SPC. In fact, no closed-form design has been reported for the computation of the capacity region of the MIMO BC subject to PAPC. To the best of our knowledge, the only attempt to characterize the entire capacity region of the MIMO BC subject to PAPC was made in [14]. Specifically, the authors in [14] established a modified duality between the MAC and BC and transformed the input optimization problem in the BC into a minimax optimization problem in the dual MAC. Then resulting program is solved by a standard barrier interior-point routine. Thus, as a common property for the class of second order optimization methods, the algorithm proposed in [14] has computational complexity that does not favor large-scale antenna systems which are envisioned in next wireless communications generations. In this paper, we determine the capacity region of MIMO BCs under PAPC. In particular, the problem of interest is also known as weighted sum rate maximization (WSRMax) for MIMO BCs. As mentioned earlier, the capacity region can be achieved by DPC, but the resulting WSRMax problem is nonconvex. As a standard step, we apply the minimax duality presented in [14] to transform the WSRMax problem into a minimax program. However, unlike [14] which solves the resulting minimax problem by a barrier interiorpoint method, we take advantage of the problem specifics to propose an efficient algorithm that blends the concept of alternating optimization (AO) and successive convex approximation (SCA). Especially, in each iteration of the proposed algorithm, closed-form expressions based on conjugate gradient projection (CGP) method are derived. As a result, the complexity of the proposed algorithm is much lower than that of the barrier method in [14], and scales linearly with the number of users in the system, making it particularly suitable for large-scale networks. Numerical experiments are carried out to demonstrate that the proposed algorithm can converge rapidly, especially for networks of high signal-to-noise ratio, and the number of iterations required for convergence is quite insensitive to the number of users. The remainder of the paper is organized as follows. The system model is described in Section II followed by the proposed algorithm in Section III. Section IV provides the complexity analysis of the proposed algorithm while Section V presents the numerical results. Finally, we conclude the paper in Section VI. Notation: Standard notations are used in this paper. Bold lower and upper case letters represent vectors and matrices, respectively. I defines an identity matrix, of which the size can be easily inferred from the context; CM×N denotes the space of M × N complex matrices; H† and HT are Hermitian and normal transpose of H, respectively; Hi,j is the (i, j)th entry of H; |H| is the determinant of H; rank(H) stands for the rank of H; diag(x) denotes the diagonal matrix with diagonal elements being x. II. S YSTEM M ODEL Consider a K-user MIMO BC where the base station and each user have N and Mk antennas, respectively. The channel matrix for user k is denoted by Hk . Let s be the composite signal that combines the data for all users in the downlink. Then, the received signal at user k is expressed as (1) yk = Hk s + zk where zk is the Gaussian noise with distribution CN (0, IM ). When DPC is applied to achieve the capacity region, for a given user k, the interference caused by users j < k is completely canceled without affecting the optimality. As a result, the WSRMax under PAPC is formulated as maximize {Sk 0} subject to PK |I+H Pk S H† | k i=1 i k k=1 wk log | |I+Hk Pk−1 Si H† | i=1 k PK [S ] ≤ P , ∀i k i,i i k=1 (2) where Sk is the input covariance matrix for the kth user, Pi is the power constraint on antenna i, and wk is the weighting factor assigned to user k. Without loss of P optimality, we assume that 0 < w1 ≤ w2 ≤ ... ≤ wK and K k=1 wk = 1 in the following. Since (2) is a nonconvex problem, solving it directly is not a good option. However, we can exploit the MAC-BC duality to transform (2) into minimax program in the dual MAC which can be solved efficiently by the novel AO as presented in the next section. III. P ROPOSED S OLUTION Applying the modified MAC-BC duality introduced in [15], we can equivalently rewrite (2) as the following minimax optimization problem PK PK † min max k=1 ∆k log |Q + i=k Hi S̄i Hi | Q0 {S̄k 0} subject to −w log |Q| PKK k=1 tr(S̄k ) = P, tr(QP) = P, Q : diagonal (3) PN where ∆k = wk − wk−1 ≥ 0, P , i=1 Pi ; {S̄k } and Q are input covariance and noise covariance matrices in the dual MAC, respectively. As shown in [15], the objective in (3) is convex with Q  0 and concave with {S̄k  0}. Thus, there exists a saddle point for (3), which also solves (2). The minimax formulation in (3) also suggests a way to find {S̄k } and Q by AO. However, a pure AO algorithm is not guaranteed to converge. In fact, a counterexample was already given in [16]. In what follows, we propose an iterative algorithm based on combining AO and SCA, of which convergence can be proved. Let {S̄nk } be the solution to the following maximization problem in the nth iteration PK PK † n maximize k=1 ∆k log |Q + i=k Hi S̄i Hi | {S̄k 0} (4) PK subject to k=1 tr(S̄k ) = P. The above maneuver is nothing but a standard routine of optimizing {S̄k } when Q is held fixed. Problem (4) can be solved by off-the-shelf interior-point convex solvers but the complexity is not affordable for large-scale systems. In our numerical experiments, all of known (free and commercial) solvers fail to solve (4) on a relatively powerful desktop PC for N ≥ 100, regardless of the number of users. That is, interiorpoint methods are not an efficient approach to solving (4) for massive MIMO techniques which are likely to be adopted in 5G systems. To overcome this shortcoming, we now present an efficient method to solve (4) based PKon the CGP framework. To proceed, let S = {S̄k |S̄k ≻ 0, k=1 tr(S̄k ) = P } be the feasible set of (4). The main operation of a CGP method is to project a given {S̃k } onto S. Our motivation is that the projection of {S̃k } onto S can be reduced to a projection of a resulting vector onto a canonical simplex, which can be computed efficiently. The projection of {S̃k } onto the feasible set S is formulated as PK 2 minimize k=1 ||Ṡk − S̃k ||F {Ṡk 0} (5) PK subject to tr( Ṡ ) = P. k k=1 Let Uk D̃k U†k = S̃k be the EVD of S̃k , where Uk is unitary and D̃k is diagonal. Then we can write Ṡk = Uk Ḋk U†k for some Ḋk  0. Since Uk is unitary, it holds that tr(Ṡk ) = tr(Ḋk ) and that ||Ṡk − S̃k ||F = ||Ḋk − D̃k ||F . That is to say, (5) is equivalent to PK 2 minimize k=1 ||Ḋk − D̃k ||F {Ḋk 0} (6) PK subject to k=1 tr(Ḋk ) = P. It is easy to see that Ḋk must be diagonal to minimize the objective of (6). Next let d̄k = diag(Ḋk ), d̃k = diag(D̃k ), d̄ = [d̄T1 , d̄T2 , . . . , d̄TK ]T , and d̃ = [d̃T1 , d̃T2 , . . . , d̃TK ]T . Then (6) can be reduced to minimize d̄≥0 subject to 1 2 ||d̄ − d̃||22 1M̃ d̄ = P (7) P where M̃ = K 1 Mk . It is now clear that (7) is the projection onto a canonical simplex and efficient algorithms (similar to water-filling algorithms) can be found in [17]. The complete description of the proposed CGP method for solving (4) is provided in Algorithm 1. We note that similar approaches were also presented in [10], [11]. Algorithm 1: The proposed CGP algorithm for solving (4). Input: P , ǫ > 0 0 1 Initialization: τ = 1 + ǫ, m = 0, {S̄k } ∈ S. 2 while (τ > ǫ) do 3 Calculate the conjugate gradient G̃m k . 4 Choose an appropriate positive scalar sm and create m m S̃m = S̄m k + s G̃k . m 5 Project S̃k onto S to obtain Ṡm k . 6 Choose appropriate step size αm and set m m m S̄m+1 = S̄m k + α (Ṡk − S̄k ). k m+1 m † 7 τ = | tr(∇f (S̄k ) (S̄k − S̄m k ))|. 8 m := m + 1. 9 end Output: S̄k as the optimal solution to (4). Another main step of a CGP method is the computation of the conjugated gradient of the objective, as required in line 3 of Algorithm 1. The conjugate gradient, denoted as G̃m k , can be calculated as follows. First, we compute the gradient of the objective in (4) as −1  K k X X H†k . (8) H†i S̄i Hi ∆j Q + ∇f (S̄k ) = −Hk j=1 i=j Then the conjugate gradient direction is given by m m m−1 G̃m k = −∇f (S̄k ) + β G̃k where the parameter β [18] ( 0 m β = m (9) is the Fletcher choice of deflection 2 −||∇f (S̄m k )|| † ∇f (S̄m−1 )) tr((G̃m−1 ) k k m=0 m≥1 (10) For the step size in line 6 of Algorithm 1, we perform an Armijo line search [19] to determine appropriate value. For the importance case of the sum capacity of the MIMO BC, we note that more efficient approaches to solving (4) do exist. In fact, in this case (4) becomes PK maximize log |Qn + i=1 H†i S̄i Hi | {S̄k 0} (11) PK subject to tr( S̄ ) = P. k k=1 We remark that (11) is equivalent to the problem of finding the sum-capacity of a MAC with a SPC for which the sum power iterative water-filling proposed in [20] or the dual decomposition method in [9] have been shown to be particularly computationally efficient. We now turn our attention to the problem of finding Qn+1 . If a pure AO method is followed, we arrive at the optimization problem below: PK PK † minimize i=k Hi S̄i Hi | k=1 ∆k log |Q + Q0 (12) −wK log |Q| subject to tr(QP) = P However, as mentioned in [14] and also observed in [16], the convergence of such a naive AO method is not guaranteed. The novelty of our proposed AO algorithm is that, instead of optimizing the original objective in (12) which can lead to fluctuations, we opt to minimize an upper bound of the objective in (12). This is in light of the SCA principle, and will lead to a monotonic convergence as shown in the Appendix. To this end, by invoking the concavity of logdet function, we have the following inequality  P † n n −n Q − Qn log |Q + K i=k Hi S̄i Hi | ≤ log |Φk | + tr Φk (13) PK n −1 where Φnk = Qn + i=k H†i S̄ni Hi , Φ−n , (Φ ) . Thus, k k using the above upper bound, Qn+1 is found to be the optimal solution to the following problem PK ∆k  −n  minimize k=1 wK tr Φk Q − log |Q| Q0 (14) subject to tr(QP) = P. Since Q in (14) is diagonal, i.e., Q = diag(q), we can rewrite (14) as PN PK n minimize i=1 k=1 φki qi − log qi q>0 (15) PN subject to i=1 Pi qi = P,  ∆ −n  k Φk i,i , i.e., φnki is the ith diagonal element where φnki = w K ∆k Φ−n of w k . By setting the derivative of Lagrangian function K of (15) to zero, we obtain 1 qi = PK (16) n k=1 φki + γPi where γ ≥ 0 is the solution of the equation N X i=1 PN Pi PK k=1 φnki + γPi = P. (17) Pi φn ki +γPi − P . It’s easy to see that PK g(γ) is decreasing with γ. When γ = 0, since k=1 φnki ≤ qi−1 we have Denote g(γ) = i=1 N X i=1 PK k=1 Pi PK k=1 φnki ≥ N X Pi qi = P. (18) i=1 Therefore, (17) always has a unique solution, which can found efficiently, e.g., by the bisection or Newton method. The proposed algorithm based on AO is summarized in Algorithm 2. The main point of Algorithm 2 is to eliminate the possible ping-pong effect of the obtained objective by the use of the inequality in (13). The convergence proof of Algorithm 2 is provided in the Appendix. Note that the error tolerance τ of Algorithm 2 is only computed for n ≥ 1. Algorithm 2: Proposed algorithm for the computation of the capacity region of a MIMO BC based on AO. Input: Q := Q diagonal matrix of positive elements, ǫ>0 Initialization: Set n := 0 and τ = 1 + ǫ. while (τ > ǫ) do Solve (4) and denote the optimal solution by {S̄nk } For n ≥ 1 compute τ = |f DPC (Qn , {S̄nk }) − f DPC (Qn−1 , {S̄kn−1 })|, where f DPC (·) denotes the objective in (3). For each k, compute PK † n n −1 Φ−n . k = (Q + i=k Hi S̄i Hi ) n+1 Solve (14) to find Q . n := n + 1. end Output: Use the obtained {S̄nk }K k=1 and the BC-MAC transformation in [5] to find the optimal solution to (2). 0 1 2 3 4 5 6 7 8 IV. C OMPLEXITY A NALYSIS In this section, we analyze the complexity estimation of the proposed algorithm in terms of the number of flops. The results of flop counting for typical matrix operations such as EVD, matrix inversion are taken from [21] and [22]. Moreover,we treat every complex operation as 6 real flops as considered in [22], [23]. In the complexity analysis presented below, we only consider the main operations of high complexity in the overall complexity. A. Complexity of Algorithm 2 Algorithm 2 has two key procedures, one for finding {S̄k } and one for finding Q. Of all steps in Algorithm 1, the computation of the conjugate gradient direction given in (8) has the largest computation cost. Similarly, the complexity for finding Q is mostly due to the computation of Φ−n k defined in (13). Therefore, the per-iteration complexity of Algorithm 2 is O(KN 3 ). That is, the complexity of the proposed algorithm increases linearly with the number of users. B. Complexity of the Barrier Interior-Point Method in [14] The method presented in [14] is based on solving the KKT condition for (3). To find the gradient w.r.t each input covariance matrix for the logdet function, the complexity is 6N 3 . Then, to find a Newton direction, this algorithm still needs to solve a linear system of M (M + 1)/2 unknowns, + 3N )3 flops.1 In total, the perwhich requires 6( KM(M+1) 2 2 3 iteration complexity is O(K N + K 3 M 6 ). 1 Here we assume Mk = M for all k to simplify the notation. TABLE I P ER - ITERATION COMPLEXITY COMPARISON Algorithms Per-iteration complexity Algorithm 2 Barrier interior-point method [14] O(KN 3 ) O(K 2 N 3 + K 3 M 6 ) The complexity comparison is summarized in Table I. We remark that the complexity of Algorithm 2 increases linearly with the number of users K, which was also achieved in [20] for the sum-capacity with SPC. Thus, Algorithm 2 scales much better than the barrier interior-point method in [14] with the number of users. V. N UMERICAL R ESULTS This section provides numerical results to verify the proposed algorithm. If not otherwise mentioned, the number of transmit antennas at the BS is set to N = 5 and the number of receive antennas at each user is set to Mk = 2 for all k.. The total power P is varied from a low value to a high one to investigate the convergence rate of the proposed algorithm.The power constraint for each transmit antenna is set equally. The initial value Q0 in the proposed algorithm is set to the identity matrix. An error tolerance of ǫ = 10−6 is selected as the stopping criterion for the proposed algorithm. In the first simulation, we plot the convergence of the proposed algorithm for different value of the total transmit power for a set of randomly generated channel realizations. As can be seen, Fig. 1 shows that the convergence of the proposed algorithm is strictly monotonic as proved in the Appendix. An interesting observation is that the number of iterations decreases when the total power increases. As mentioned in the complexity analysis, the proposed algorithm has a desirable property, i.e., the per-iteration complexity increases linearly with the number of users K. This is an attractive property in the system of a large number of users which is usually the case for a massive MIMO setup. However, it is difficult, if not impossible, to analyze the convergence rate of the proposed algorithm with K by analytical expressions. Instead, we study the convergence property of the proposed algorithm with K by numerical experiments. For the purpose, we plot in Fig. 2 the average number of iterations required by the proposed algorithm to converge. As can be seen, the number of iterations for the proposed algorithm to converge is relatively insensitive to K . It is worth noting that for similar setups, the interior-point method in [14] requires at least 60 iterations to converge, while the proposed algorithm only takes 4 iterations even in the case of 50 users. This promising characteristic of the proposed algorithm makes it suitable for studying the capacity region of massive MIMO systems where the number of transmit antennas and/or the number of users can be very large. This point will be further elaborated in the next numerical experiment. Taking the advantage that the proposed algorithm has low complexity, in the last numerical experiment we characterize the capacity region of a massive MIMO system with PAPC. In 102 P = 0 dB P = 10 dB P = 20 dB 4 ZF, PAPC DPC, PAPC DPC, SPC R2 Duality gap 10−1 2 10−4 10−7 0 2 4 6 8 10 0 Iteration 0 1 4 15 R2 Average number of iterations 20 5 10 15 20 25 30 35 40 45 ZF, PAPC DPC, PAPC DPC, SPC 50 Number of users Fig. 2. Average number of iterations versus number of users. 5 10 5 2 4 (a) Number of users K = 2 5 P = 0 dB P = 20 dB 3 R1 Fig. 1. Convergence behavior of the proposed algorithm with K = 2. 3 2 0 0 1 2 3 4 5 R1 particular, we also consider achievable rate region of the wellknown ZF scheme [24], [25]. The purpose is to understand the performance of ZF (which is thought to be sub-optimal) in comparison with the capacity achieving coding scheme under some realistic channel models. To this end we consider a simple urban scenario using WINNER II B1 channel model [26], where a base station, equipped with N = 128 antennas, is located at the center of the cell and single-antenna receivers are distributed randomly. The total power at the BS is P = 46 dBm and each antenna is subject to equal power constraint, i.e., Pi = P/N for i = 1, 2, . . . , 128. As can be seen clearly in Fig. 3, there is a remarkable gap between the achievable rate region of ZF and the capacity region, especially when the number of users increases. This basically implies that ZF is still far from optimal for a practical number of transmit antennas. Our observation opens research opportunities in the future to strike the balance between optimal performance by DPC and low-complexity by ZF. VI. C ONCLUSIONS In this paper, we have considered the problem of computing the capacity region of Gaussian MIMO BCs subject to PAPC. (b) Number of users K = 8 Fig. 3. Comparison of capacity region of a massive MIMO setup with N = 128 and Mk = 1 for all k. For the case K = 8 users, the capacity region is projected on the first two users. Towards this end, the problem of WSRMax with PAPC has been solved by a low-complexity algorithm. We have first converted the noncovex problem of MIMO BC into an equivalent minimax problem in the corresponding dual MAC. Then a novel AO algoithm has been proposed to solve the resulting mimimax program in combination with the successive convex optimization principle. In particular, all the computation in the proposed algorithm is based on closed-form expressions. In addition, the simulation results have demonstrated a fast and stable convergence of the proposed algorithm, even for large-scale settings. ACKNOWLEDGEMENTS This work was supported by a research grant from Science Foundation Ireland and is co-funded by the European Regional Development Fund under Grant 13/RC/2077. A PPENDIX For notational convenience, we denote Q , {Q|Q : ˆ k = ∆k . Since the diagonal, Q  0, tr(QP) = P }, ∆ wK PK function log |Q + i=k H†i S̄i Hi | is jointly concave with Q and {S̄k }, the following inequality holds K X k=1 ˆ k log |Q + ∆ ≤ K X i=k K X k=1 ˆ k log | Qn + ∆ K X + H†i S̄i Hi | | + ˆk ∆ k=1 K X i=k K X i=k H†i S̄ni Hi | k=1 ˆ k log |Qn+1 + ∆ ≤ K X {z Φn k } i=k k=1 + K X i=k Since {S̄nk } = arg max {S̄k }∈S † n tr(HΦ−n k H (S̄i − S̄i )) (19) ˆk ∆ k=1 K X i=k  n+1 † − S̄ni )) . (20) tr(HΦ−n k H (S̄i PK k=1 ∆k log |Qn + PK i=k H†i S̄i Hi |, † n tr(HΦ−n k H (S̄i − S̄i )) ≤ 0 ˆk ∆ k=1 K X i=k n+1 † tr(HΦ−n − S̄ni )) ≤ 0 k H (S̄i k=1 ˆ k log |Qn+1 + ∆ ≤ K X k=1 K X (21) (22) i=k H†i S̄n+1 Hi | i  ˆ k log |Φn | + tr Φ−n (Qn+1 − Qn ) . (23) ∆ k k  } k=1 (24) n+1 solves (14) it holds that   ˆ k log |Φn | + tr Φ−n (Qn+1 − Qn ) − log |Qn+1 | ∆ k k k=1 K X k=1   ˆ k log |Φn | + tr Φ−n (Q − Qn ) − log |Q| (25) ∆ k k for all Q ∈ Q. For the special case Q := Qn , the inequality above is reduced to K X k=1   ˆ k log |Φnk | + tr Φ−n (Qn+1 − Qn ) − log |Qn+1 | ∆ k K X k=1 | ˆ k log |Φnk | − log |Qn | . (26) ∆ {z f DPC (Qn ,{S̄n k }) } Combining (24) and (26) results in f DPC (Qn , {S̄nk }) ≥ f DPC (Qn+1 , {S̄n+1 }). k It is easy to see that {f DPC (Qn , {S̄nk })} is bounded above, and thus {f DPC (Qn , {S̄nk })} is convergent. We also note that (13) is strict if Q 6= Qn . Consequently, the sequence {f DPC (Qn , {S̄nk })} is strictly decreasing unless it is convergent. Therefore, the continuity of f DPC (·) and the compactness of S and Q imply lim f DPC (Qn , {S̄nk }) = f DPC (Q∗ , {S̄∗k }). n→∞ R EFERENCES which leads to K X {z   ˆ k log |Φn |+tr Φ−n (Qn+1 −Qn ) −log |Qn+1 |. ∆ k k H†i S̄n+1 Hi | i for all {S̄k } ∈ S. For {S̄k } = {S̄n+1 } the above inequality k is equal to K X i=k H†i S̄n+1 Hi | − log |Qn+1 | i f DPC (Qn+1 ,{S̄n+1 }) k K X ≤ K X ≤ the optimality condition results in K X | K X  ˆ k log |Φn | + tr(Φ−n (Qn+1 − Qn )) ∆ k k K X k=1 ˆ k log |Qn+1 + ∆ Since Q for all Q ∈ Q and {S̄k } ∈ S. The above inequality comes PK from the first order approximation of log |Q + i=k H†i S̄i Hi | around the point (Qn , {S̄nk }). Substitute Q := Qn+1 and S̄k := S̄n+1 into the above equality, we have k K X K X ≤   ˆ k tr Φ−n (Q − Qn ) ∆ k k=1 K X Subtract both sides of the above inequality by log |Qn+1 | results in [1] E. Telatar, “Capacity of multi-antenna Gaussian channels,” Eur. Trans. Telecommun, vol. 10, pp. 585–598, Nov. 1999. [2] G. J. Foschini and M. J. Gans, “On limits of wireless communications in a fading environment when using multiple antennas,” Wireless Pers.Commun, vol. 6, pp. 311–335, Mar. 1998. [3] H. Weingarten, Y. Steinberg, and S. S. Shamai, “The capacity region of the Gaussian multiple-input multiple-output broadcast channel,” IEEE Trans. Inf. Theory, vol. 52, no. 9, pp. 3936 – 3964, Sep. 2006. [4] G. Caire and S. Shamai, “On the achievable throughput of a multiantenna Gaussian broadcast channel,” IEEE Trans. Inf. Theory, vol. 49, no. 7, pp. 1691–1706, Jul. 2003. [5] S. Vishwanath, N. Jindal, and A. Goldsmith, “Duality, achievable rates and sum-rate capacity of Gaussian MIMO broadcast channels,” IEEE Trans. Inf. Theory, vol. 49, pp. 2658–2668, Oct. 2003. [6] P. Viswanath and D. Tse, “Sum capacity of the vector Gaussian broadcast channel and uplink-downlink duality,” IEEE Trans. Inf. Theory, vol. 49, no. 8, pp. 1912–1921, Aug. 2003. [7] W. Yu and J. Cioffi, “Sum capacity of Gaussian vector broadcast channels,” IEEE Trans. Inf. Theory, vol. 50, no. 9, pp. 1875–1892, Sep. 2004. [8] H. Viswanathan, S. Venkatesan, and H. Huang, “Downlink capacity evaluation of cellular networks with known-interference cancellation,” IEEE J. Sel. Areas Commun., vol. 21, no. 5, pp. 802–811, Jun. 2003. [9] W. Yu, “Sum-capacity computation for the Gaussian vector broadcast channel via dual decomposition,” IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 754 –759, Feb. 2006. [10] J. Liu, Y. T. Hou, and H. D. Sherali, “On the maximum weighted sumrate of MIMO Gaussian broadcast channels,” in Proc. IEEE ICC, May 2008, pp. 3664 – 3668. [11] R. Hunger, D. A. Schmidt, M. Joham, and W. Utschick, “A general covariance-based optimization framework using orthogonal projections,” in Proc. IEEE SPAWC, Jul. 2008, pp. 76 – 80. [12] M. Vu, “MIMO capacity with per-antenna power constraint,” in Proc. IEEE GLOBECOM, Dec. 2011, pp. 1 – 5. [13] ——, “The capacity of MIMO channels with per-antenna power constraint,” CoRR, vol. abs - 1106 - 5039, 2011. [14] W. Yu and T. Lan, “Transmitter optimization for the multi-antenna downlink with per-antenna power constraints,” IEEE Trans. Signal Process., vol. 55, no. 6, pp. 2646–2660, Jun. 2007. [15] W. Yu, “Uplink-downlink duality via minimax duality,” IEEE Trans. Inf. Theory, vol. 52, no. 2, pp. 361–374, Feb. 2006. [16] T. M. Pham, R. Farrell, and L.-N. Tran, “Revisiting the MIMO capacity with per-antenna power constraint: Fixed-point iteration and alternating optimization,” IEEE Trans. Signal Process., submitted. [17] L. Condat, “Fast projection onto the simplex and the ℓ1 ball,” Mathematical Programming, Series A, vol. 158, no. 1, pp. 575 – 585, July 2016. [18] R. Fletcher, Practical Methods of Optimization, 2nd ed. New York, NY, USA: Wiley-Interscience, 1987. [19] L. Armijo, “Minimization of functions having Lipschitz continuous first partial derivatives.” Pacific J. Math., vol. 16, no. 1, pp. 1 – 3, 1966. [20] N. Jindal, W. Rhee, S. Vishwanath, S. Jafar, and A. Goldsmith, “Sum power iterative water-filling for multi-antenna Gaussian broadcast channels,” IEEE Trans. Inf. Theory, vol. 51, no. 4, pp. 1570–1580, Apr. 2005. [21] G. H. Golub and C. F. V. Loan, Matrix Computations, 3rd ed. The John Hopkins Univ. Press, 1996. [22] Z. Shen, R. Chen, J. Andrews, J. Heath, R.W., and B. Evans, “Low complexity user selection algorithms for multiuser MIMO systems with block diagonalization,” IEEE Trans. Signal Process., vol. 54, no. 9, pp. 3658–3663, Sep. 2006. [23] L.-N. Tran, M. Bengtsson, and B. Ottersten, “Iterative precoder design and user scheduling for block-diagonalized systems,” IEEE Trans. Signal Process., vol. 60, no. 7, pp. 3726–3739, Jul. 2012. [24] Q. Spencer, A. Swindlehurst, and M. Haardt, “Zero-forcing methods for downlink spatial multiplexing in multiuser MIMO channels,” IEEE Trans. Signal Process., vol. 52, no. 2, pp. 461–471, Feb. 2004. [25] K. Karakayali, R. Yates, G. Foschini, and R. Valenzuela, “Optimum zero-forcing beamforming with per-antenna power constraints,” in Proc. IEEE ISIT 2007, Jun. 2007, pp. 101–105. [26] P. Kyösti, J. Meinilä, L. Hentilä, X. Zhao, T. Jämsä, C. Schneider, M. Narandzić, M. Milojević, A. Hong, J. Ylitalo, V.-M. Holappa, M. Alatossava, R. Bultitude, Y. de Jong, and T. Rautiainen, “Winner II channel models,” tech. rep. D1.1.2 V1.2, IST-4-027756 WINNER II, 2007.
7cs.IT
arXiv:cs/0203023v1 [cs.CE] 19 Mar 2002 Agent Trade Servers in Financial Exchange Systems DAVID LYBÄCK Financial Market Systems, OM AB and MAGNUS BOMAN Swedish Institute of Computer Science (SICS) AB New services based on the best-effort paradigm could complement the current deterministic services of an electronic financial exchange. Four crucial aspects of such systems would benefit from a hybrid stance: proper use of processing resources, bandwidth management, fault tolerance, and exception handling. We argue that a more refined view on Quality-of-Service control for exchange systems, in which the principal ambition of upholding a fair and orderly marketplace is left uncompromised, would benefit all interested parties. Categories and Subject Descriptors: I.2.11 [Artificial Intelligence]: Distributed Artificial Intelligence—Intelligent Agents; J.4 [Social and Behavioral Sciences]: Economics; K.4.4 [Computers and Society]: Electronic Commerce—Distributed Commercial Transactions General Terms: Economics, Management, Performance Additional Key Words and Phrases: Trading agent, Agent server, Financial exchange service, Agent programming 1. BACKGROUND Exchange operators are now in a business environment shaped by the nature of electronic trading, competing for listings and order flow, and facing the needs of both institutional major traders and retail-level investors. An exchange operation excels when traders display their true willingness to trade. The transaction costs of trading may then shrink, to the benefit of the participants and indirectly also of the marketplace in the competitive environment. The electronic exchange is a high-confidence complex system that has to behave in a very well understood and predictable fashion. This implies that the system must withstand variable load, various failures in supporting systems, faulty user commands, et cetera, and must not by erroneous internal actions cause serious economic damage to the participants of the exchange. While the ideal is continuous matching of simple orders at the equilibrium point with a completely deterministic flow, many needs arise where this ideal cannot be upheld. The complexity of the market entails that decision situations are not repeatable: if a trader makes the wrong choice, very few conclusions can be drawn ex post to improve a similar Journal Name etc. Permission to make digital/hard copy of all or part of this material without fee for personal or classroom use provided that the copies are not made or distributed for profit or commercial advantage, the ACM copyright/server notice, the title of the publication, and its date appear, and notice is given that copying is by permission of the ACM, Inc. To copy otherwise, to republish, to post on servers, or to redistribute to lists requires prior specific permission and/or a fee. c 2017 ACM 0000-0000/2017/0000-0001 $5.00 ACM Journal Name, Vol. V, No. N, December 2017, Pages 1–11. 2 · D. Lybäck and M. Boman situation next time, since there is no identifiable next time. The ideal is also under threat from all sorts of failures: hardware, software, peripheral, and network failures. Such failures negatively affect price dissemination routines, market access opportunities, and bid feedback in general. Our starting point is thus not the ideal marketplace, but a real one, the resilience of which is pivotal. In a deterministic system, output messages are deterministically repeatable given a stochastic series of input events, i.e. some identified non-deterministic events will completely determine the nature and sequence of the output stream in a welldocumented fashion. This implies that if there is any choking in the system whatsoever, the input stream will have to buffer. A best-effort service on the other hand, uses as much resources as it is allocated or until there is no better solution possible. Prolonged execution usually yield “better” results for a user, but prolonging is not always an option. The notion of best-effort defines the service result delivered at a particular time as the best one found or best one possible under those given, unrepeatable circumstances, and the service level of the result cannot be impeached: it is rational given the boundary [Russell and Wefald 1991]. Introducing the concept of best-effort could open up many new possibilities, but care must be taken so that the maximal downside of using a best-effort service - as opposed to a completely deterministic service - is under control and can be described. It is vital that an add-on service that executes according to best-effort does not in any way compromise the overall ambition of a fair and orderly marketplace. So far, these concerns have hindered best-effort innovations in the financial domain. We acknowledge the importance of complete tractability in the financial exchange services and therefore propose a hybrid system where new best-effort services complement the current deterministic services. We start in section 2 by discussing the state-of-the-art, and thereafter current problems. In section 4, we propose solutions to those problems. In sections 5 and 6, we analyze the advantages and point forwards, respectively. 2. THE STATE-OF-THE-ART A strong trend in the software industry is towards Quality-of-Service management with anytime algorithms [Dean and Boddy 1988] for good usage of processing resources and network bandwidth, and clever procedures for graceful degradation whenever there is choking or an outright subsystem failure. An anytime algorithm is a constructive solution to the problem of how to implement best-effort services. 2.1 Processing Resources Contingency orders depend on other orders in the marketplace. Some of the functionality in contingency orders can be provided by the distributed local trading clients, but only the core exchange system can guarantee deals involving more than one trade, such as buying one stock and selling another one simultaneously. Apart from such combinations, one can also place a linked order, i.e., a number of single orders held together by a “spider object” with information about exclusive ORconditions in that set of orders. By using contingency orders, the trading tactic or need is hidden from other participants; only the added liquidity is displayed, to some extent. A programming technique used for contingency order evaluation is ACM Journal Name, Vol. V, No. N, December 2017. Agent Trade Servers in Financial Exchange Systems · 3 automatic insertion of virtual orders derived from the contingency order. 2.2 Bandwidth Management Network congestion is (to a reasonable extent) avoided through the use of dedicated leased lines to the exchange system gateways. Custom-built delay boxes are sometimes in place to neutralize systematic latency differences, in order to minimize unfairness between participants. In some systems, the matching service can throttle down automatically (brake the matching service somewhat to reduce dissemination congestion) if the network is choking, to smoothen a temporary peak. 2.3 Fault Tolerance Fault tolerance techniques provide services compliant with the specification after a fault has occured. For hardware errors, or catastrophic situations like flooding or fire, exchange systems provide hot fail-over mechanisms (based on active replication), i.e. a hot standby replica is ready to take over in a very short time after a serious problem occurred in the primary site. As the secondary has processed the same material as the primary, processing can continue easily (after catch-up and rollback, if needed). Thus, the fail-over procedure is designed to have a minimal impact on users. In fact, users might not even realize that a fail-over has taken place. 2.4 Exception Handling Exception handling deals with the undefined and unanticipated conditions that, if left unchecked, can propagate through the system and cause a serious fault. Selfverifying software is an ad hoc method used in many important systems already deployed in our society, including electronic exchange systems, phone switches, and aircraft. A typical exchange system uses a rollback procedure from disk-based transaction logs for a restart, and in problematic cases, the operations staff must identify and (technically) nullify the transaction causing the problem, to enable a successful restart of the system. Whereas some adaptive rate control procedures are implemented in high-end electronic exchange systems, no system has (to our knowledge) installed components that actively and explicitly base the service offering to the user on a best-effort principle, because one does not want to compromise the paramount paradigm of determinism. 3. THE PROBLEM In this section, we will focus on unsolved problems within the four key areas explained in the previous section. We build on our own experiences of programming real-time trading agents. Parts of this experience comes from participating in the Trading Agent Competition (TAC), see http://tac.eecs.umich.edu/. Compared to artificial markets, such as the Artificial Stock Market (see http: //ArtStkMkt.sourceforge.net/) or idealized games (see, e.g., [Rust et al. 1994]), TAC is a realistic, albeit simple, emulation of real agent electronic trading. In order to keep its focus on practical issues, TAC was set up not as a (complex) stock exchange, but as an intelligent assistant competition. The assistants acquire resources for eight different customers on real-time Internet auctions. ACM Journal Name, Vol. V, No. N, December 2017. 4 · D. Lybäck and M. Boman 3.1 Processing Resources Contingency order matching can easily degrade the performance of the matching service. Basic Quality-of-Service control can temporarily suspend the evaluation of certain complex orders to ensure the upkeeping of basic matching service performance at all times. This is a crude maneuver, and will not suffice as users expect their combination orders kept active at all times. Future order types with more conditions and parameters will also increase the expectations on the evaluation policy (cf. [Ygge and Akkermans 2000]). In effect, users will slowly start to expect a more refined view on Quality-of-Service, like in all areas of Information Technology (cf., e.g., the Open Financial Exchange effort [Cover 2001] in which financial data is exchanged over the Internet in XML). The non-deterministic nature of multi-threaded processes with anytime algorithms has so far effectively kept them banned from live exchanges, while they play an important role on artificial marketplaces. For instance, the constraint solver implemented by Mats Carlsson in our TAC’02 agent 006, computes optimal amounts of various goods using an anytime algorithm [Aurell et al. 2002]. On the other hand, there will arise temporal series of events in a full-scale, multithreaded non-deterministic matching service that are practically non-repeatable, since the operating system is involved in the algorithm. For example, a single context switch event may not necessarily occur at exactly the same time or in the same way when replaying a situation, and may thereby promote another flow of events. This inherent property of non-deterministic services makes simulation, verification, and debugging extremely hard. This also means that no guarantees of basic performance or temporally dependent completions can be given a priori in a non-deterministic service, which makes it inappropriate for many heavy participants, considering the financial values involved. In other words, one would like best-effort processing services in certain cases, to open up for evaluation of more complex orders, but one can not give up the deterministic stance for the matching engine software. 3.2 Bandwidth Management Congestion that impacts order transactions or crucial price updates will undoubtedly lead to business problems for the participant and eventually for the exchange itself. A systematic latency that is not fair or a partial failure that affects some participants is unacceptable, and may even close the exchange until the problem is resolved. Interactive traffic such as order submission is very sensitive to latency caused by the need to retransmit after a packet loss. Treating packet loss as an indication of congestion in the network as done in TCP is inappropriate for time-critical transactions - such as most order submissions today. General throttling is not an ideal solution for bandwidth management in all cases. If more traffic is to become best-effort throughput, there should to be an alternative where there is less business sensitivity to inevitable small delays. One may argue that future networks with Active Queue Management [Ramakrishnan et al. 2001] which does not use overflow as a the only indication of congestion, will somewhat reduce the impact of a loss on latency-sensitive flows, but this is far from enough, considering the second-by-second importance of quotes in highly liquid ACM Journal Name, Vol. V, No. N, December 2017. Agent Trade Servers in Financial Exchange Systems · 5 orderbooks. Further, many exchanges prioritize between similar orders based on their arrival timestamp in milliseconds. We therefore see a future need of an explicit rate control, and a refined view of congestion. Furthermore, at the retail level, the drawbacks of TCP/IP are already at hand, since end customers do typically not lease dedicated lines to their respective Internet brokers. 3.3 Fault Tolerance A fail-over mechanism is often not a sufficient solution, as a logical problem leading to a stop in the primary site will also cause the exact same problem in the secondary site. Redundant replication is the classical way to get reliability from a sub-system, and since it can not be applied to software logic errors, this has created a software problem, which will continue to affect dependability. 3.4 Exception Handling On the one hand, fault-detection checks, which make a whole partition grind to a halt just because of one problematic minor software logic situation somewhere deep in one of many orderbooks, is not the exception handling one would like in a modern software system. On the other hand, no mistakes must be made by the system whatsoever. As mentioned above, it is totally unacceptable that the system would perform erroneous executions in a repeating or even aggravating fashion. The operator prefers a technical stop compared to a long series of faulty deals, for obvious reasons. Therefore the programmer may add invariants at vital points in the code, to defuse any potentially dangerous problem immediately. The obvious problem with self-checking software is its lack of rigor and the unclear coverage. There is always a trade-off between control and performance. Consequently, one method and level of exception handling is not appropriate for all parts of the software. The lack of proper technologies for ultra-reliable systems has created an unsatisfying situation when it comes to exception handling. In summary, the demand for complete tractability and total determinism puts harsh restrictions on the services and architecture options. We believe that it is the lack of proper assurance techniques that currently keeps ambitions of best-effort execution from the exchange systems arena. 4. THE SOLUTION While the ability to control the rate at which objects are refreshed is a nice tool for ensuring optimal usage of available bandwidth in network-enabled games, it requires good design judgment if best-effort transport via Internet is to be used professionally in active financial marketplaces. We believe that best-effort transport is most reasonable when combined with best-effort execution of scenario-based orders, because of the much lower criticality of its client to front-end server interaction timing. The marketplace operator can consider our suggested technical solutions in a backcasting milieu. 4.1 The Agent Trade Server To build upon the robustness of the deterministic exchange system platform, we propose a complementary subsystem operating fully under the best-effort principle, both for its bandwidth management, processing resource management, fault ACM Journal Name, Vol. V, No. N, December 2017. 6 · D. Lybäck and M. Boman tolerance, and exception handling. We acknowledge the importance of continued complete tractability in the current services and therefore propose a hybrid system where the new, optional best-effort services are contained in a separate subsystem: the Agent Trade Server (ATS), complementing the current deterministic services. A cardinal use case for the new subsystem is when the user uploads a latent order combined with a trigger configuration, combined into a so-called Agent Instruction, to the ATS. The ATS, operating under the best-effort principle, communicates with the exchange system platform using one of the already available protocols. 4.2 Proximity Advantage The ATS computer will typically be placed physically close to the computer hosting the exchange system back-end processes (core), and between the ATS and the core will be a high-speed local connection (cf. Figure 1). The parts comprising the complete setup with Agent Trade Servers are: (1) Deterministic core with the matching service proper. (2) Multi-protocol gateway(s). (3) Best-effort appurtenance: one or many adjacent servers, running encapsulated complementary best-effort services. The appurtenance is split into three parts: (a) Core services extension(s), e.g., a computer running Combinatorial Solution Finder, or other supportive services. (b) Agent Trade Server(s) type I, with user-configurable agents spawned under control of a dispatch service, utilizing multi-threading or similar nonverified resource sharing under the operating system of the server. (c) Agent Trade Server(s) type II, with user-programmable agents executing under sandbox resource management, or similar sharing mechanisms. There is extensive shielding between individual agents, so that one erroneously programmed or extra demanding agent cannot affect others in an uncontrollable way. (4) External connections via gateway(s), typically through either leased lines or shared IP links. Using these connections, users run interactive sessions for order submission, and subscribe to multi-cast messages, such as price updates. (5) External connections to the best-effort appurtenance, typically via the Internet for the interactive sessions and the subscriptions of, for example, status updates. Note that servers in the appurtenance classified as 3a above do not have external connections, but only internal links with the core. 4.3 Agent Trade Server Operation The ATS will, upon receiving an agent instruction, create a new thread to handle the request, or revise a previously spawned thread or process. A precondition for the automatic activation of a latent order in the ATS is a Boolean operation on a set of scenario triggers. Inspirational generic scenarios can be proposed by the marketplace operator, and then adjusted by each trader or investor. The scenarios could be saved in personalized galleries, and optionally calibrated for each individual order. The agent in the ATS will continuously evaluate the triggers in its instruction against real-time market information received from the core. ACM Journal Name, Vol. V, No. N, December 2017. Agent Trade Servers in Financial Exchange Systems Fig. 1. · 7 The Agent Trade Servers setup. To avoid resonance effects, the bourse operator would typically not allow only one short-term fluctuation in an index value as a single trigger. Instead, the trader should set a scenario by combining at least two (not totally correlated) triggers. Note that a real-time scenario trigger must evaluate the current market situation within specified timing constraints, quite short-termed, possibly slightly trading-off the quality of its own analysis if needed. Thus, even the trigger in itself is under Quality-of-Service control. The user could be allowed to set a minimal acceptable quality of the best-effort service for each individual agent instruction (according to a measurement yet to be defined), below which the agent is not allowed to trade. Another challenge is to investigate what scenarios that a trader would like to trigger some latent orders, and how the quality of real-time triggering is to be measured. The operator would also be constrained with respect to integrity and security issues. The trust relationship between marketplace and customer will be put to harsh tests, as the customer in the typical case is reluctant to reveal her preferences. As most uploads encode true preferences, this information is ideally processed without human intervention, and must be treated as maximally sensitive data. The drawback for the trader with communicating trade intention early, can be more than offset by the inherent premium in the reliable, cheap, high-speed connection between the core and the ATS. The proximity and direct connection is effectively feeding the ATS with market data much faster than anyone else, thus ACM Journal Name, Vol. V, No. N, December 2017. 8 · D. Lybäck and M. Boman putting the Agents in the ATS at an advantage in certain cases. This may be a strong enough incentive to stimulate an early submission of orders via the ATS when appropriate. 4.4 Agent Programming Although we primarily envision a few archetypical agents to be offered by the exchange for traders to configure, it is not unreasonable in later generations of the system to open up so-called “sandbox areas” for user-created code to execute in. To encode trade logic into agents means that the code can be pro-active, autonomous, and encapsulated. These features make agents an interesting option for future complex order types. The user can also make it hard for unauthorized persons to grasp the preferences encoded in the agent, and make it virtually impossible to apply reverse engineering techniques. The programming tools for creating agents are increasing in number and quality, as is the number of agent programming consultants and brokers, and plenty of code is shared through online communities. Academic research has pursued this track for years, also as connected to a marketplace scenario. Examples of our own recent efforts in this field include agent-based simulations on real stock market data [Boman et al. 2001], and multi-agent systems situations, in which several artificial agents operate simultaneously on the same marketplace, subject to norms and other constraints [Boman et al. 2001]. Although there is substantial previous work in the academic field when it comes to agent-mediated commerce, we feel that it is often unrealistic in the way it commonly approaches agent-mediated trading [Weinhardt and Gomber 1999] and agent mobility [Ruddle 1999]. The agency gateway is a multiprotocol connectivity zone to reach users via various end-user devices. At SICS, ways of providing accessible APIs for end-user services have been produced [Bylund 2001]. International community-shaping efforts, such as .NET and the possibility of SOAP adaptation for web services are also likely to be important in the near future. It should be noted that agent/server communication also has its own set of problems. An agent server is under constant threat of overload, due to malicious intent or bad agent code. In the TAC server, agent requests which are looped without correct termination conditions produce problems. It is also difficult for the people responsible for the server to shut out agents due to erroneous behavior, since the reasons for the shutout should be communicated to the agent owners, which is sometimes impossible. As a result, agent owners might find their agent has not been running as expected, and this without notification, let alone explanation. In a financial exchange, this is totally unacceptable. In addition, a study made in the first phases of programming our TAC’01 agent 006 [Hauser and Orrhult 2002] analyzes the following latency concerns. —Round trip time for agent- or server-generated message —Receive interval (from message receipt start until finish) —Intrinsic data time bound (e.g., a bound on the validity of message data) —Agent deliberation time (e.g., for interpreting a message) —Throughput: number of messages per time unit or clock cycle —Time-slotting: agent planning activities for sending and receiving messages ACM Journal Name, Vol. V, No. N, December 2017. Agent Trade Servers in Financial Exchange Systems · 9 —Network topology, affecting message delivery time These concerns should be added to the usual performance considerations [Tanenbaum 1996], including: —Traffic congestion caused by temporary resource overload —Resource imbalance (e.g., between network and processor capacity) —Broken cables —Broken connections (e.g., overloaded hosts or routers) Moreover, the sheer amount and varibility of the code running on an agent server makes code verification a practical impossibility. To get a crude feel for the amount of code expected, consider that our TAC’01 agent RiskPro consisted of about 7000 lines of code, 3000 lines of which were devoted to communication with the TAC server, over TCP/IP [Boman 2001]. There is currently considerable effort being put into solving these problems, however, both in the form of market-driven and in the form of purely academic research. 5. THE ADVANTAGES 5.1 Processing Resources New scenario-based orders can be evaluated continuously in the ATS without degrading the basic matching service. This means that the exchange system as a whole can offer not only the matching itself, but also resources for tactical trading without negative impact on the basic services. Special limitations surround combination orders. Only the core can fully match a combination, but cumbersome combinations should be kept away from the core as long as possible. Moreover, instead of a large combination, the ATS can formulate a basket of smaller combinations in a quick sequence, accepting small residuals. That said, the major advantage of an ATS when it comes to processing resources, is the continuous scenario evaluations ongoing in the agents. Only when a latent order triggers, it will be formulated in detail and submitted to the core. The user could be allowed to set a minimal acceptable quality of the best-effort matching service for each individual order, according to a measurement yet to be defined, below which the service is automatically disregarding that order. Another challenge is to investigate how the quality of real-time triggering is to be measured. 5.2 Bandwidth Management We argue that regardless of how clever future transport protols may become, besteffort transport solutions between the marketplace and the end user are most reasonable when combined with best-effort evaluation of scenario-based orders, such as in the ATS, because of the much lower criticality of its client/server interaction timing. Furthermore, the more intentions that could be communicated in advance of unfolding market events, the more the network load smoothens, which reduces the need for expensive reserve capacity in the network. Also, traffic between the ATS and the core would not be hindered, or slowed down, by network load in the Wide Area Network or the Internet, due to the direct or LAN connection used. A setup ACM Journal Name, Vol. V, No. N, December 2017. 10 · D. Lybäck and M. Boman with an ATS may therefore reduce gateway and network loads slightly, which is especially beneficial when approaching heavy load situations. Throttling may then in certain situations be eluded. This in turn might slightly increase the number of completed deals per minute under those circumstances. 5.3 Fault Tolerance Keeping as much as possible of the new functionality out of the deterministic core altogether will drastically reduce the risk of introducing new code into the heart of the system. Containment of new functionality away from the core is in itself a strategy for upkeeping and increasing highest possible reliability while the system is under continuous development. A crashed ATS thread was presumably only evaluating a single agent instruction from one user, and can therefore be out of action for a few seconds (or even die) without widespread consequences, whereas the core in a sense is the marketplace, and must not fault. The best-effort nature of the ATS allows for many possibilities in a fail-over situation to degrade its services until fully recovered to normal operating status. A recovery unit can replace a failed process, and concentrate on agent restart based on the user-submitted agent instruction, and can discard processing states. 5.4 Exception Handling Error containment and exception handling will be in effect locally in the ATS, and internal problems are not to affect the core services. 6. CONCLUSION We have suggested an Agent Trade Server as a complement to an exchange system, without compromising the core service, and so: —open up for proper best-effort transport connectivity, fully in line with Internet design foundations and contemporary trends in the industry —prepare for the introduction of new order details and tactical trading instructions —provide a new level of containment to better defend the complex system from fatal software error propagation when new services are introduced in the future. Acknowledgements The authors would like to thank Johan Kummeneje and Lars Rasmusson for draft comments. Lybäck was supported through his research assignment at OM, while the VINNOVA project TAP on accessible autonomous software provided Boman with the time required for this study. REFERENCES Aurell, E., Boman, M., Carlsson, M., Eriksson, J., Finne, N., Janson, S., Kreuger, P., and Rasmusson, L. 2002. A constraint programming agent for automated trading. Submitted. Boman, M. 2001. Trading agents. AgentLink News 6, 15–17. Boman, M., Brouwers, L., Hansson, K., Jansson, C.-G., Kummeneje, J., and Verhagen, H. 2001. Artificial agent action in markets. Electronic Commerce Research 1, 1-2, 159–168. Boman, M., Johansson, S., and Lybäck, D. 2001. Parrondo strategies for artificial traders. In Intelligent Agent Technology, N. Zhong, J. Liu, S. Ohsuga, and J. Bradshaw, Eds. World Scientific, Singapore, 150–159. ACM Journal Name, Vol. V, No. N, December 2017. Agent Trade Servers in Financial Exchange Systems · 11 Bylund, M. 2001. Personal service environments—Openness and user control in user-service interaction. Tech. rep., Dept of Information Technology, Uppsala Univ. IT Licentiate Thesis 2001-008. Cover, R. 2001. The XML cover pages—Open Financial Exchange (OFX). Available at http: //xml.coverpages.org/ofx.html, last modified Jan 3, 2001. Dean, T. and Boddy, M. 1988. An analysis of time-dependent planning. In Proc AAAI’88. AAAI Press/MIT Press, St. Paul, MN, 49–54. Hauser, V. and Orrhult, J. 2002. Network monitoring for networking agents. M.S. thesis, Dept of Computer and Systems Sciences, Stockholm University. 02-01-DSV-SU. Ramakrishnan, K. K., Floyd, S., and Black, D. 2001. The Addition of Explicit Congestion Notification (ECN) to IP. Internet Engineering Task Force. RFC 3168, Proposed Standard. Ruddle, M. 1999. The application of mobile agents to share trading. M.S. thesis, Dept of Electrical and Electronic Engineering, Univ of London. Russell, S. and Wefald, E. 1991. Do the Right Thing–Studies in Limited Rationality. MIT Press, Cambridge, Mass. Rust, J., Miller, J. H., and Palmer, R. 1994. Characterizing effective trading strategies–insights from a computerized double auction tournament. Economic Dynamics and Control 18, 61–96. Tanenbaum, A. S. 1996. Computer Networks, 3rd ed. Prentice-Hall, New Jersey. Weinhardt, C. and Gomber, P. 1999. Agent-mediated off-exchange trading. In Proc HICSS-32. IEEE Computer Society. Ygge, F. and Akkermans, H. 2000. Resource-oriented multicommodity market algorithms. Autonomous Agents and Multi-Agent Systems 3, 1, 53–71. Received March 2002; revised...; accepted... ACM Journal Name, Vol. V, No. N, December 2017.
5cs.CE
THE NUMBER OF MONODROMY REPRESENTATIONS OF ABELIAN VARIETIES OF LOW p-RANK arXiv:1706.03435v1 [math.AG] 12 Jun 2017 BRETT FRANKEL Abstract. Let Ag be an abelian variety of dimension g and p-rank λ ≤ 1 over an algebraically closed field of characteristic p > 0. We compute the number of homomorphisms from π1ét (Ag , a) to GLn (Fq ), where q is any power of p. We show that for fixed g, λ, and n, the number of such representations is polynomial in q, and give an explicit formula for this polynomial. We show that the set of such homomorphisms forms a constructible set, and use the geometry of this space to deduce information about the coefficients and degree of the polynomial. In the last section we prove a divisibility theorem about the number of homomorphisms from certain semidirect products of profinite groups into finite groups. As a corollary, we deduce that when λ = 0, # Hom(π1ét (Ag , a), GLn (Fq )) #GLn (Fq ) is a Laurent polynomial in q. Contents Acknowledgments 1. Introduction 2. Counting Formulas 2.1. Some Linear Algebra 2.2. Polynomials and their Types 2.3. Counting 3. The Space of Representations 4. A Divisibility Theorem for Profinite Groups 4.1. Introduction and Statement of Theorem 4.2. Proof (first statement) 4.3. Proof of the Converse References 1 2 2 4 4 4 5 10 12 12 13 14 15 Acknowledgments This paper is derived from my Ph.D. thesis, and as such there are far too many people who have impacted this work than can be listed here. It is a pleasure to thank Ted Chinburg, my thesis advisor, for suggesting this problem and for countless discussions and guidance. Thanks also to David Harbater, Zach Scherr, and Ching-Li Chai for helpful discussions and answering my questions. I am grateful to Professor Fernando Rodriguez-Villegas for hosting me for a very productive week at the International Centre for Theoretical Physics. It was he who observed Corollary 3, and introduced me to both his theorem with Cameron Gordon [5] and the theorem of Frobenius, [3] from which one easily deduces Corollary 3. He also made me aware of Proposition 2. Bob Guralnick very helpfully referred me to both his paper with Sethuraman [6] and earlier work of Gerstenhaber [4]. Additional thanks are due to Nir Avni for comments on the draft and Sebastian Sewerin for help reading [3]. Many others have had a great influence on my mathematical development, which no doubt manifests itself in some of the work that follows. There are too many to list here, but I have attempted to acknowledge many of these people in the thesis document.[2] 1. Introduction In Mixed Hodge Polynomials of Character Varieties, [7] Hausel and RodriguezVillegas computed the number of homomorphisms from the fundamental group of a Riemann surface X of genus g into GLn (Fq ). By van Kampen’s theorem, this is equivalent to counting 2g-tuples of matrices, P (q) = # Hom(π1 (X, x), GLn (Fq )) = #{(X1 , Y1 , · · · , Xg , Yg ) : [X1 , Y1 ][X2 , Y2] · · · [Xn , Yn ] = 1}, where [Xi , Yi ] denotes the commutator XY X −1 Y −1 . They also consider a twisted version of the problem, computing R(q) = # Homtwist (π1 (X, x), GLn (Fq )) = #{(X1 , Y1 , · · · , Xg , Yg ) : [X1 , Y1 ][X2 , Y2] · · · [Xn , Yn ] = ζn }, where ζn is a primitive nth root of 1. The salient feature of both these counts is that, for fixed g and n, P and R are both polynomial functions of q. Another curious feature is that P (q) is divisible by the order of GLn (Fq ). This phenomenon is explained in [5]. In the twisted case, P GLn acts scheme-theoretically freely on Homtwist (π1 (X), GLn ), and the GIT quotient Mn = Homtwist (π1 (X, x), GLn )//P GLn is the moduli space of twisted homomorphisms from π1 (X, x) to GLn . Via a theorem of N. Katz, (in the appendix to [7]), the polynomial point-counting formula is equal to the E-polynomial E(Mn , T ) of Mn (C), which encodes information about the weight and Hodge filtrations on the cohomology of Mn . Baraglia and Hekmati [1] have more recently extended these methods to the moduli of untwisted representations. The results in this paper may be thought of as an analogue of the combinatorial results in [7] and [1], now applied to varieties of positive characteristic. The most natural analogue would be consider a curve C/F̄p in place of a Riemann surface. 2 However, in positive characteristic, the étale fundamental group is a much more subtle object. In particular, there is not a single curve of genus greater than 1 for which an explicit presentation of the fundamental group is known. [12] However, these fundamental groups have well-understood abelianization; the abelianization of the fundamental group of a curve is the fundamental group of the Jacobian variety, which is the dual of the Jacobian. This paper treats representations of fundamental groups of abelian varieties. Let Ag be an abelian variety of dimension g. The p-torsion points of A form a vector space over Fp of dimension 0 ≤ λ ≤ g. We call λ the p-rank of Ag . Let a ∈ A. The étale fundamental group of Ag is given by π1ét (Ag , a) ∼ = Y (Zℓ )2g × Zλp ℓ6=p ∼ = (Ẑ′ )2g−λ × Ẑλ , where Ẑ = lim Z/n ←− is the profinite completion of the group of integers and Ẑ′ = lim Z/n ←− (n,p)=1 is the group of integers completed away from the prime p. To lighten the notation, we will henceforth write π1 (Ag ) instead of π1ét (Ag , a). Let q be a power of p. A homomorphism π1 (Ag ) → GLn (Fq ) is determined by the image of the topological generators, so specifying a homomorphism is equivalent to choosing an ordered 2g-tuple of pairwise commuting matrices, such that the first 2g − λ have order not divisible by p. Theorems 1 and 2 explicitly compute the number of homomorphisms π1 (Ag ) → GLn (Fq ), where q is a power of p and the p-rank of Ag is either 0 or 1, respectively. Theorem 3 computes the number of homomorphisms up to conjugation in the p-rank 0 case. All three counting formulas are polynomial in q, depending on g and n but not on the characteristic. Section 3 considers the space of all such representation, and relates the geometry of this space to certain features of the polynomial formulas in the previous section. We show that the counting polynomial has integer coefficients, and give a lower bound for the degree, which we show to be exact when A is in elliptic curve. In section 4, we state and prove the following theorem. Theorem (Theorem 6). Let S be a set of primes (not necessarily finite), and let ẐS = lim Z/n, where the the inverse limit is taken over all natural numbers n not ←− divisible by any prime in S. Then for any topologically finitely generated profinite group Γ and finite group G, # Hom(Γ ⋊ ẐS , G) ∈ S−1 Z, |G| 3 where Γ ⋊ ẐS is any semidirect product of Γ and G, and S is the multiplicative set generated by the elements of S. Conversely, if Γ̃ is topologically finitely generated and # Hom(Γ̃, G) ∈ S−1 Z |G| for all finite groups G, then there exists a Γ with Γ̃ ∼ = Γ ⋊ ẐS . Theorem 6 is the profinite analogue of a divisibility theorem by Gordon and Villegas[5]. We deduce as a corollary that when the λ = 0, Hom(π1 (Ag ), GLn (Fq ))/#GLn (Fq ) is a Laurent polynomial in q. 2. Counting Formulas Our main reference in this section is Macdonald’s Symmetric Functions and Hall Polynomials. [10] Throughout, p will be a prime number, and q will denote a power of p. 2.1. Some Linear Algebra. For any matrix X ∈ GLn (Fq ) we have an associated Fq [T ]-module structure on Fnq , where T acts by X. If X and Y induce isomorphic module structures on Fnq , then the isomorphism of Fq [T ]-modules defines an element of GLn (Fq ), so X and Y are conjugate. Note that X ∈ GLn (Fq ) has order prime to p if and only if X acts semisimply on Fnq . That is, X is diagonalizable over F̄q . Since elements of order prime to p act semisimply, the associated Fq [T ]-module is a direct sum of simple modules of the form Fq [T ]/f (T ), where the f (T ) are irreducible but not necessarily distinct. Matrices of order prime to p are thus uniquely characterized, up to conjugation, by their characteristic polynomials. 2.2. Polynomials and their Types. Definition 1. A type Λ of n is a partition of n along with a refinement of its conjugate. That is, a partition λ of n along with partitions λi of the multiplicities of the entries i of λ. This is a slight generalization of what Macdonald calls a type in [10]. For example, the data λ = (5 3 3 3 3 3 2 2 2 1 1) = (5(1) 3(5) 2(3) 1(2) ), λ5 = (1), λ4 = ∅, λ3 = (2 2 1), λ2 = (2 1), λ1 = (2) give a type of 28. We shall write λ ⊢ n when λ is a partition of n, and Λ  n when Λ is a type of n. By Y f (i, r) (i,r)∈Λ we will mean the product taken over all pairs (i, r) such that i ∈ λ and r ∈ λi . We index over λ without multiplicity, but over r ∈ λi with multiplicity. So if   λ = (3 3 3 3 3 1 1) Λ = λ3 = (2 2 1) ,  λ1 = (2) then Y f (i, r) = f (3, 2)2f (3, 1)f (1, 2). (i,r)∈Λ 4 We associate to an n × n matrix X a type Λ  n by considering the characteristic polynomial cX (T ) of X. Factoring cX (T ) into irreducible factors gives a partition of λ ⊢ n, where the entries of λ are the degrees of the irreducible factors of cX (T ), counted with multiplicity. We then take λi to be the partition consisting of the multiplicity of each distinct degree-i factor of cX (T ). For example, over F3 , we associate λ = (3 3 1 1 1 1 1 1), λ3 = (2), λ1 = (3 3) to the polynomial (T 3 + 2T + 1)2 (T − 2)3 (T − 1)3 . Recall that the number of irreducible monic polynomials over Fq of degree i is 1X µ(k)q i/k , i k|i where µ is the Möbius function. [9] Definition 2. Denote by ψΛ (q) the number of monic polynomials p(T ) ∈ Fq (T ) with factorization type Λ. Note that ψΛ (T ) ∈ Q[T ], but in general ψΛ (T ) ∈ / Z[T ]. For instance, if λ = (1 1), λ = (1 1), then (q − 1)(q − 2) φΛ (q) = . 2 1 2.3. Counting. We now count the number of homomorphisms from π1 (Ag ) to GLn (Fq ), where the p-rank of Ag is 0. This is equivalent to counting 2g-tuples of commuting matrices (X1 , . . . , Xk ) such that each Xi has order prime to p. Since the combinatorial arguments below do not make use of the fact that 2g is even, we may state a slightly stronger theorem. Theorem 1. The number of ordered k-tuples of pairwise-commuting, semisimple, invertible matricies with entries in Fq is Y X Y X ··· ψΛ2 (q i1 ) ψΛ1 (q) #GLn (q) (i1 ,r1 )∈Λ1 Λ2 r1 Λ1 n ··· X ψΛk−1 (q i1 i2 ···ik−2 ) Λk−1 rk−2 Y X (ik−1 ,rk−1 )∈Λk−1 Λk rk−1 (i2 ,r2 )∈Λ2 i1 i2 ···ik−1 ψ (q ) Q Λk . i 1 #GLrk (q i2 ···ik ) (ik ,rk )∈Λk For instance, the number of commuting semisimple pairs (X, Y ) ∈ GL2 (Fq )2 is (q 3 + q 2 − q + 1) #GL2 (Fq ) = q 6 − 3q 4 + 2q 3 + q 2 − 2q + 1. q In GL3 (Fq ), the number of semisimple commuting pairs is q 6 − q 5 − q 4 + 2q 3 − q 2 + q − 1 q3 =q 12 − 2q 11 − q 10 + 4q 9 − q 8 − 4q 6 + 2q 5 + 3q 4 − 2q 3 + q 2 − 2q + 1. #GL3 (Fq ) Proof. We first prove the theorem for k = 2. Since semisimple matrices are characterized, up to conjugation, by their characteristic polynomials, we associate to each such conjugacy class a type Λ. There are by definition ψΛ (q) conjugacy classes of 5 type Λ. Suppose X1 is any matrix of order prime to p, with associated type Λ. A matrix Y commutes with X1 if and only if Y acts X-equivariantly on Fnq . That is, writing Fnq as a sum of simple modules M (Fq [T ]/fj (T ))rj , Fnq ∼ = j the action of Y is a Fq [T ]-automorphism of each (Fq [T ]/fj (T ))rj . Specifying such an action is given by an element of GLrj (Fqdeg fj ). The order of the centralizer of a semisimple matrix X1 with type Λ is thus Y Y (1) #GLrj (Fqdeg fj ) = #GLr (Fqi ). fj (i,r)∈Λ The matrix X2 commutes with X1 and also must be semisimple, so X2 acts semisimply on each (Fq [T ]/fj (T ))rj . A semisimple matrix Yj ∈ GLrj (Fqdeg fj ) is determined, up to conjugacy, by its characteristic polynomial, or equivalently by a type N = (ν, ν ℓ )  rj and a polynomial of type N in Fqdeg fj [T ]. By the same argument used to compute the cardinality of the centralizer of X1 , we see that index of the centralizer of Yj ∈ GLrj (Fqdeg fj ) is #GLrj (Fqdeg fj ) Q . #GLs (Fqdeg(fj )ℓ ) (ℓ,s)∈N The number of all such Yj is therefore   X #GLrj (Fqdeg fj )  deg fj ) Q (2) . ψN (q #GLs (Fqdeg(fj )ℓ ) N r j (ℓ,ν)∈N So, if X1 is semisimple with type Λ, the number of semisimple matrices X2 commuting with X1 is   YX #GLrj (Fqdeg fj )  deg fj ) Q ψN (q  #GLs (Fqdeg(fj )ℓ ) f N r j (3) = j (ℓ,s)∈N   Y X #GLr (Fqi )  i . ψN (q ) Q ) #GL (F iℓ s q N r (i,r)∈Λ (ℓ,s)∈N Given a semisimple matrix X ∈ GLn (Fq ), we have just computed both the number of matrices (1) that commute with X and the number of semisimple matrices (3) that commute with X. Note that both of these numbers depend only on the type associated to X, since the deg(fj )’s are the entries of λ, and the rj ’s are the entries of the corresponding λdeg fj ’s. 6 From these two computations, we see that the number of pairs (X1 , X2 ) of commuting semisimple matrices in GLn (Fq ) is   XX Y X #GLr (Fqi ) #GLn (Fq )  i Q (4)  ψN (q ) Q #GLr (Fqi ) #GLs (Fqiℓ ) Λn ψ(Λ) (i,r)∈Λ N r (i,r)∈Λ (ℓ,s)∈N where the second sum is taken over polynomials with type Λ. Since   Y X #GLr (Fqi ) #GLn (Fq )  i Q  ψN (q ) Q #GLr (Fqi ) #GLs (Fqiℓ ) N r (i,r)∈Λ (i,r)∈Λ (ℓ,s)∈N depends only on Λ, we may simplify (4) by replacing the second summation with multiplication by ψΛ (q). Thus (4) simplifies to (5) #GLn (Fq ) X ψΛ (q) Y X (i,r)∈Λ N r Λn Q ψN (q i ) . #GLs (Fqiℓ ) (ℓ,s)∈N We now continue by induction. Suppose the number of pairwise commuting (k − 1)-tuples of semisimple elements of GLn (Fq ) is Y Y X X ψΛ2 (q i1 ) ψΛ1 (q) #GLn (q) Λ1 n ··· X Λk−2 rk−3 ψΛk−2 (q i1 i2 ···ik−3 ) X Y (ik−2 ,rk−2 )∈Λk−2 Λk−1 rk−2 (i1 ,r1 )∈Λ1 Λ2 r1 ··· (i2 ,r2 )∈Λ2 i1 i2 ···ik−2 Q ψΛk−1 (q ) . #GLrk−1 (q i1 i2 ···ik−1 ) (ik−1 ,rk−1 )∈Λk−1 Suppose further that for each possible action of X1 , . . . , Xk−1 on Fnq , the action of Xk−1 on an isotypic summands of the Fq [X1 , . . . , Xk−2 ]-module Fnq has characteristic polynomials of type Λk−1 as above when these isotypic summands are viewed as Fqi1 i2 ···ik−2 -vector spaces. (We note that Λ2k−1 may be different for different summands; indeed, Λk−1  rk−2 , and the symbol rk−2 takes on different values in different terms of the above formula). Since Xk commutes with each Xℓ , 1 ≤ ℓ < k, then Xk acts on Fnq by acting on each isotypic summand, which by induction are Fqi1 i2 ···ik−2 [Xk−1 ]-modules, or Fqi1 i2 ···ik−1 vector spaces of dimension rk−1 . There are X ψΛk (q i1 i2 ···ik−1 ) Λk rk−1 conjugacy classes of semisimple elements of GLrk−1 (Fqi1 i2 ···ik−1 ). By (1), each conjugacy class has #GLrk−1 (Fqi1 i2 ···ik−1 ) Q #GLrk (Fqi1 i2 ···ik ) (ik ,rk )∈Λk 7 elements, for a total of F (rk−1) = X ψΛk (q i1 i2 ···ik−1 ) Λk rk−1 #GLrk−1 (Fqi1 i2 ···ik−1 ) Q #GLrk (Fqi1 i2 ···ik ) (ik ,rk )∈Λk r Fqk−1 i1 i2 ···ik−1 . So by induction, the number of pairwise-commuting possible Xk -actions on k-tuples of semisimple matrices in GLn (Fq ) is X Y X Y #GLn (q) ψΛ1 (q) ··· ψΛ2 (q i1 ) (i1 ,r1 )∈Λ1 Λ2 r1 Λ1 n ··· X Y ψΛk−1 (q i1 ···ik−2 ) Λk−1 rk−2 (ik−1 ,rk−1 )∈Λk−1 (i2 ,r2 )∈Λ2 1 F (rk−1 ). #GLrk−1 (q i1 ···ik−1 ) Canceling factors of #GLrk−1 (q i1 ···ik−1 ), the above formula simplifies to the statement of the theorem.  Observation. From the proof, we see that the formula in Theorem 1 is polynomial in q. This polynomial depends on n and k, but not on the characteristic. The next theorem computes # Hom(π1 (Ag ), GLn (Fq )) when the p-rank of Ag is 1. Remark. The polynomial behavior of the formula in Theorem 1 is strictly a samecharacteristic phenomenon, which fails even in the simplest case in characteristic ℓ 6= p. Consider the set Hom(Ẑ′ , Gm (Fℓn ). When ℓ 6≡ 1 mod p, there are infinitely many n for which ℓn 6≡ 1 mod p, and for such n # Hom(Ẑ′ , Gm (Fℓn ) = ℓn − 1. But ℓn − 1. When ℓ ≡ 1 for those n such that ℓn ≡ 1 mod p, # Hom(Ẑ′ , Gm (Fℓn ) s mod p, we may write ℓ = mp + 1 with p ∤ m. Then for those n such that p ∤ n, the p-adic valuation of ℓn − 1 is s, and # Hom(Ẑ′ , Gm (Fℓn ) = (ℓn − 1)/ps . When p | n, # Hom(Ẑ′ , Gm (Fℓn ) (ℓn − 1)/ps . Theorem 2. The number of ordered k-tuples of invertible, pairwise-commuting matrices (X1 , . . . , Xk−1, Y ) such that the Xi are all semisimple is Y X Y X ··· ψΛ2 (q i1 ) #GLn (q) ψΛ1 (q) (i1 ,r1 )∈Λ1 Λ2 r1 Λ1 n ··· X ψΛk−2 (q i1 i2 ···ik−3 ) Λk−2 rk−3 Y (i2 ,r2 )∈Λ2 (q i1 ···ik−2 − 1)(q i1 ···ik−2 )rk−2 −1 . (ik−2 ,rk−2 )∈Λk−2 Proof. As in the inductive step of the theorem, we assume the number of pairwise commuting (k − 2)-tuples of semisimple elements of GLn (Fq ) is X Y X Y #GLn (q) ψΛ1 (q) ··· ψΛ2 (q i1 ) Λ1 n ··· X Λk−3 rk−4 ψΛk−3 (q i1 i2 ···ik−4 ) X Y (ik−3 ,rk−3 )∈Λk−3 Λk−2 rk−3 (i1 ,r1 )∈Λ1 Λ2 r1 (i2 ,r2 )∈Λ2 i1 i2 ···ik−3 Q ψΛk−2 (q ) . #GLrk−2 (q i1 i2 ···ik−2 ) (ik−2 ,rk−2 )∈Λk−2 As before, for each possible action of X1 , . . . , Xk−2 on Fnq , the action of Xk−2 on an isotypic summands of the Fq [X1 , . . . , Xk−3]-module Fnq has characteristic polynomials 8 of type Λk−2 as above when these isotypic summands are viewed as Fqi1 i2 ···ik−3 -vector spaces. Thus it suffices to count the number of pairs Xk−1 , Y which act on each isotypic summand of the Fq [X1 , . . . , Xk−2 ]-module Fnq . By assumption, these summands are r of the form Fqk−2 i1 ···ik−2 . Let O be the orbit, under the conjugation action in GLrk−2 (Fqi1 ···ik−2 ), of a matrix of order prime-to-p. Given any Xk−1 , we will denote by A its restriction to the isotypic summand in question, and similarly B will denote the restriction of Y . For any A ∈ O, the number of nonsingular matrices B that commute with A is the order of the centralizer of A. The number of elements in O is the index of the centralizer of A. Thus there are precisely #GLrk−2 (Fqi1 ···ik−2 ) pairs (A, B) such that A and B commute and A ∈ O. So to compute the number of pairs (A, B) with [A, B] = 1 and |A| prime to p, we need only count the number of possible conjugacy classes A. These are in bijection with characteristic polynomials, of which there are (q i1 ···ik−2 )rk−2 −1 (q i1 ···ik−2 − 1). Thus the number of pairwise commuting k-tuples of invertible n × n matrices (X1 , . . . , Xk−1 , Y ), of which all but possibly the last have order prime to p, is #GLn (q) X ψΛ1 (q) X ψΛk−3 (q i1 ···ik−4 ) X Y (ik−3 ,rk−3 )∈Λk−3 Λk−2 rk−3 Λk−3 rk−4 X ψΛ2 (q i1 ) (i1 ,r1 )∈Λ1 Λ2 r1 Λ1 n ··· Y ψΛk−2 (q Q Y ··· (i2 ,r2 )∈Λ2 i1 ···ik−3 )#(A, B) , #GLrk−2 (q i1 ···ik−2 ) (ik−2 ,rk−2 )∈Λk−2 where #(A, B) = Y #GLrk−2 (Fq i1···ik−2 )(q i1 ···ik−2 )rk−2 −1 (q i1 ···ik−2 − 1). (ik−2 ,rk−2 )∈Λk−2 After cancellation we arrive at the statement of the theorem.  Theorem 3. The number of conjugacy classes of k-tuples of pairwise-commuting, invertible, semisimple matrices is X ψΛ1 (q) ··· Λk−1 rk−2 ψΛk−1 (q i1 i2 ···ik−2 ) #{(X1 , . . . , Xk )}/ ∼= Y X ··· ψΛ2 (q i1 ) (i1 ,r1 )∈Λ1 Λ2 r1 Λ1 n X Y Y (i2 ,r2 )∈Λ2 (q i1 ···ik−1 )rk−1 −1 (q i1 ···ik−1 − 1). (ik−1 ,rk−1 )∈Λk−1 Proof. Observe that Y ∈ GLn (Fq ) stabilizes (X1 , . . . , Xk ) if and only if it commutes with each Xℓ . Writing Z(X1 , . . . , Xk ) for the order of the stabilizer of (X1 , . . . , Xk ) 9 and applying the orbit-stabilizer lemma, X #{(X1 , . . . , Xk )}/ ∼ = {(X1 ,...,Xk )} = = Z(X1 , . . . , Xk ) #GLn (Fq ) 1 #GLn (Fq ) 1 #GLn (Fq ) X Z(X1 , . . . , Xk ) {(X1 ,...,Xk )} X 1 {(X1 ,...,Xk ,Y )} where the last sum is taken over k +1-tuples pairwise commuting semisimple matrices (X1 , . . . , Xk , Y ) with all Xℓ semisimple. The value of this sum was computed in Theorem 2.  3. The Space of Representations In this section, Ag will be an abelian variety with p-rank 0, as before. However, all statements apply if the p-rank of Ag is 1, mutatis mutandi. All schemes below are reduced. Recall that a subset of affine (or projective) space is said to be constructible if it may be expressed as a Boolean combination of Zariski-closed sets. [11] Fix n, and let R = Hom(π1 (Ag ), GLn (F̄q )). Recall that a representation of π1 (Ag ) is given by a choice of 2g invertible matrices X k , 1 ≤ i ≤ 2g, such that these matrices are pairwise commuting and each have order relatively prime to p. Assigning a matrix 2 (Xijk )1≤i,j≤n to each generator, we may view R as a subset of A2gn . We will now show that for fixed n and g, the set R = Hom(π1 (Ag ), GLn (F̄q )) of all representations may be considered a constructible set. Proposition 1. R is constructible. Proof. The requirement that the matrices pairwise commute defines a closed affine subscheme. Invertibility is an open condition, so the variety of pairwise commuting, invertible matrices is quasi-affine. We may further specify that the matrices are (simultaneously) diagonalizable with the statement that there exists an invertible matrix C (with coordinates C i,j , 1 ≤ i, j ≤ n) so that CX k C −1 is diagonal for each k. This last statement requires an existential quantifier, and thus does not necessarily define a subscheme, but it is a first order statement in the sense of model theory, and 2 hence the space of representations is a definable subset of A2gn . Since the theory of algebraically closed fields admits quantifier elimination, [11] every definable set is in fact constructible, i.e. a Boolean combination of closed subschemes.  Alternatively, since we are considering simultaneously diagonalizable matrices, one may write the diagonalization of each Xk as an element of a torus, and then define 2 a morphism GLn × G2gn → A2gn , (C, X1 , . . . , X2g ) 7→ (C −1 X1 C, . . . , C −1 X2g C). m By Chevalley’s Theorem, [11] the image of this morphism, which is equal to R, is constructible.  Since R is constructible, we may write R = (R1 − R1′ ) ∪ (R2 − R2′ ) ∪ . . . ∪ (Rt − Rt′ ), where each Ri is a projective variety and Ri′ is a (possibly empty) closed subscheme 10 of Ri . Alternatively, since R is contained in affine space, we may choose each Ri and Ri′ to be affine. Proposition 2. Let P (T ) be the polynomial such that P (q) = #R(Fq ), the number of Fq -rational points of R. Then P (T ) ∈ Z[T ]. This argument is not new, and can be found in the appendix of [7] by Katz. Proof. Chose q so that each Ri and Ri′ is defined over Fq . Let P (T ) = a0 + a1 T + . . . ad T d . Then the zeta function Z(R, T ) is defined to be ∞ ∞ X X n P (q n )T n /n) Z(R, T ) = exp( R(Fqn )T /n) = exp( n=0 n=0 On the other hand, each Ri and is rational, and Ri′ is a projective variety, each Z(Ri , T ) and Z(Ri′ , T ) Z(R, T ) = Y Z(Ri , T ) i Z(Ri′ , T ) . Writing (1 − α1 T ) · · · (1 − αr T ) (1 − β1 T ) · · · (1 − βs T ) in lowest terms and taking logs, we see that r s X X log(1 − βk T ) = log Z(R, T ) log(1 − αj T ) − Z(R, T ) = j=1 k=1 = ∞ X P (q n )T n /n n=0 = ∞ X d X ai q in T n /n. n=0 i=0 And so, taking derivatives, s X k=1 r ∞ d d X αj XX X ai q i βk ∂ − = Z(R, T ) = ai q in T n−1 = . 1 − βk T 1 − αj T ∂T 1 − qiT n=0 i=1 j=1 i=1 Comparing poles, each αj and βk must be a power of q. Comparing numerators we see that for each i, either ai is the number of k such that βk = q i , or −ai is the number of j for which αj = q i .  From the above proof, we see that the leading coefficient of P (T ) is the multiplicity of (1 − βs T ) in the denominator of Z(R, T ). Theorem 4 (R. Guralnick, B. Sethuraman [6]). Let V be the variety of commuting k-tuples of n × n matrices. If ∆ ⊂ V denotes the union of the discriminant loci, i.e. the set on which at least one matrix has a repeated eigenvalue, then the the closure of R − ∆ is an irreducible component of V . The dimension of this component is n2 + (k − 1)n. Corollary 1. The degree of P (T ) is at least n2 + (2g − 1)n. 11 Proof. Let D ⊂ V be the union of the determinant loci, the set on which at least one matrix is not invertible. Note that V ⊃ R ⊃ V − (∆ ∪ D). Then we may decompose R = (R1 − R1′ ) ∪ (R2 − R2′ ) ∪ . . . ∪ (Rt − Rt′ ), where R1 is the closure of V − ∆, R1′ = ∆ ∪ D, and all Ri ⊂ (∆ ∪ D) for all i > 1. By theorem 4, the closure of R1 − ∆ is irreducible, and is thus equal to the closure of R1 − (∆ ∪ D) because D is closed. From the proof of Proposition 2, the degree of P (T ) is supi dim Ri , and dim R1 = n2 + (k − 1).  Theorem 5 (Gerstenhaber [4]). The variety of commuting pairs of matrices is irreducible. Corollary 2. When g = 1, P (T ) is monic of degree n2 + (2g − 1)n. Proof. Since the variety V of pairs of commuting matrices is irreducible, we may write R1 = V and R1′ = ∆ ∪ D, and V ⊃ R ⊃ V − (∆ ∪ D) as in the proof of Corollary 1. All remaining Ri are contained in ∆∪D, which is a closed subvariety of an irreducible variety and therefore of strictly lower dimension.  4. A Divisibility Theorem for Profinite Groups 4.1. Introduction and Statement of Theorem. In “On the divisibility of # Hom(Γ, G) by |G|,” [5] Cameron Gordon and Fernando Rodriguez-Villegas prove that for a finitely-generated group Γ̃, Γ̃ has infinite abelianization if and only if it satisfies the divisibility condition in the paper’s title for all finite groups G. This extends a result of Louis Solomon [14] which proves the same, assuming Γ̃ has a presentation with more generators than relations. Using their notation, # will denote the cardinality of a set, while | · | will be the order of either a finite group or an element of a finite group. By homomorphism, we will always mean a continuous homomorphism, where where countable groups are given the discrete topology. We now state and prove the following profinite analogue of [5]: Theorem 6. Let S be a set of primes (not necessarily finite), and let ẐS = lim Z/n, ←− where the the inverse limit is taken over all natural numbers n not divisible by any prime in S. Then for any topologically finitely generated profinite group Γ and finite group G, # Hom(Γ ⋊ ẐS , G) ∈ S−1 Z, |G| where Γ ⋊ ẐS is any semidirect product of Γ and G, and S is the multiplicative set generated by the elements of S. Conversely, if Γ̃ is topologically finitely generated and # Hom(Γ̃, G) ∈ S−1 Z |G| for all finite groups G, then there exists a Γ with Γ̃ ∼ = Γ ⋊ ẐS . Remark. This result is precisely the profinite version of [5], because a finitely generated group has infinite abelianiztion if and only if it is of the form Γ ⋊ Z. 12 Corollary 3. Let P (T ) be as in the previous section, i.e. the unique polynomial such that P (q) = Hom(π1 (Ag ), GLn (Fq )) whenever q is a power of p, and denote by G(T ) the polynomial such that G(q) = #GLn (Fq ). Then P (T ) 1 ∈ Z[T, ]. #GLn (Fq ) T Proof. Choose S = {p}. Writing π1 (Ag ) ∼ = (Ẑ′ )2g ∼ = (Ẑ′ )2g−1 × Ẑ′ ∼ = (Ẑ′ )2g−1 × ẐS , we see that for all q, P (q)/#GLn (Fq ) is a rational number whose denominator is (T ) divisible only by the prime p. Thus T n PG(T ∈ Q[T ] for sufficiently large n. Since ) (T ) G(T ) is monic, T n PG(T ∈ Z[T ]. )  4.2. Proof (first statement). In this section we prove that for any topologically finitely generated Γ, the number of homomorphisms from Γ ⋊ ẐS to G, when divided by |G|, has denominator divisible only by primes in S. We first recall a theorem of Frobenius. [3] (See [8] for a short elementary proof.) Theorem 7 (Frobenius). If H is a finite group and n elements of H with order dividing n is a multiple of n. |H|, then the number of Lemma 1. Let G be a finite group, and H a subgroup of order pr m, with p ∤ m. Choose g ∈ NG (H) with |g| a power of p. Then the set {x ∈ Hg : |x| is a power of p} has cardinality divisible by pr . Proof. We may assume that g ∈ / H, for otherwise the lemma follows from Theorem 7, and then reduce to the case that G = hH, gi, since it suffices to prove the lemma for s this subgroup. Let ps be the least positive integer with g p ∈ H. By Theorem 7, the number of elements of G with p-power order is a multiple of r+s p , and the number of such elements in K = hH, g p i is a multiple of pr+s−1 . Thus the number of elements of G\K with p-power order is a multiple of pr+s−1 . Let x ∈ G\K be of p-power order. Then we may write x = yg t with y ∈ H and p ∤ t. The group Z/(pr+s )× acts on the set of of all such x, n · x = xn . The orbit s of x under this action is the set of generators of hxi, and hxi surjects onto hgi/hg p i. Each orbit inside G\K therefore contains the same number of elements in each coset of the form Hg u with p ∤ u. G\K is the union of ϕ(ps ) = ps−1 (p − 1) cosets of H, so the number of elements of Hg with p-power order is equal to #{x ∈ G\K : |x| is a power of p}/ϕ(ps ), and is thus divisible by pr .  For any p ∈ / S we may write Γ ⋊ ẐS ∼ = (Γ ⋊ ẐS∪{p} ) ⋊ Zp . Therefore, we have reduced the statement of the theorem to the following: Proposition 3. Suppose Γ is any topologically finitely generated profinite group, and let vp denote the p-adic valuation on the integers. Then   # Hom(Γ ⋊ Zp , G) vp ≥ 0. |G| The proof of the proposition is now a straightforward modification of the argument in [5]. 13 Proof. A homomorphism Φ : Γ ⋊ Zp is determined by its restriction φ = Φ and the Γ image of 1 ∈ Zp , which is an element g ∈ G such that the order of g is a power of p, subject to the condition that φ((−1)γ(1)) = g −1φ(γ)g for all γ ∈ Γ. In particular, this condition ensures that g normalizes φ(Γ), and thus normalizes the centralizer Cφ of φ(Γ). Observe that if a pair (φ, g) determines a well-defined homomorphism Γ⋊Zp → G as above, then so does (φ, x) if and only if |x| is a power of p and x ∈ Cφ g. Thus # Hom(Γ ⋊ Zp , G) X #{g ∈ G : |g| is a power of p, g −1φ(γ)g = φ((−1)γ1) ∀ γ ∈ Γ}, = φ where the sum is taken over all φ which are restrictions of homomorphisms from Γ ⋊ Zp , ie. those φ for which there exists such a g. We let G act by conjugation on the set of homomorphisms restricted to Γ. The stabilizer of φ under this action is Cφ . Denoting by the orbit of φ by [φ], each element of [φ] extends to the same number of homomorphisms on Γ ⋊ Zp . For each [φ] choose a representative φ and an element gφ ∈ G such that (φ, gφ ) determines a homomorphism. # Hom(Γ ⋊ Zp , G) X #{g ∈ G : |g| is a power of p, g −1φ(γ)g = φ((−1)γ1) ∀ γ ∈ Γ} = φ X |G| = #{g ∈ G : |g| is a power of p, g −1 φ(γ)g = φ((−1)γ1) ∀ γ ∈ Γ} |Cφ | [φ] X |G| #{x ∈ Cφ gφ : |x| is a power of p}. = |Cφ | [φ] The proposition follows by applying Lemma 1 to the coset Cφ gφ .  4.3. Proof of the Converse. We have the following proposition: Lemma 2. Let S and S be as above, and Γ̃ a topologically finitely generated profinite group. Suppose that for all finite groups G, # Hom(Γ̃, G)/|G| ∈ S−1 Z. Then Γ̃ has ẐS = lim Z/n ←− (n,p)=1 ∀p∈S as a quotient. Proof. Suppose Γ̃ does not surject onto lim Z/n. ←− (n,p)=1 ∀p∈S Then for some prime ℓ ∈ / S and sufficiently large m, Γ̃ does not surject onto Z/ℓm . All homomorphisms from Γ̃ to abelian ℓ-groups factor through the maximal proabelian ℓ-quotient of Γ̃, which is itself a quotient of (Z/ℓm−1 )r , where Γ̃ is topologically generated by r generators. But there are at most ℓr(m−1) homomorphisms 14 (Z/ℓm−1 )r → Z/ℓs . Taking s > r(m − 1) and G = Z/ℓs , we see that # Hom(Γ̃, G) is not a multiple of |G|, and thus is not an element of S−1 Z.  All that remains to be shown is that Γ̃ is a semidirect product, which follows from the proposition below. Proposition 4. Let f : H → ẐS be a continuous surjection of profinite groups. Then f has a continuous section with closed image. Proof. Let h ∈ H with f (h) = 1. Q Consider the < h > generated by Q closed nsubgroup p h. It is isomorphic to a product p∈I Zp × p∈I / Z/p , where I is a set of primes. The order of h, as a supernatural number, is divisible by the order of 1 ∈ ẐS , and so S is contained in the complement of I (See [13]). Thus < h > has a direct factor K isomorphic to ẐS , which is a closed subgroup of H. K is topologically generated by the image k of h under the projection map < h > → K, and f (k) = 1. So H has a closed subgroup K which is abstractly isomorphic to ẐS , and f sends a topological generator of K to a topological generator of ẐS . The restriction of f to K is thus an isomorphism, and therefore admits a section.  Remark. We have in fact proven a stronger statement than the theorem. Namely, it suffices to check that # Hom(Γ̃, G)/|G| ∈ S−1 Z only for groups of the form G = Z/pe , where p ∈ / S. References [1] D. Baraglia and P. Hekmati. Arithmetic of singular character varieties and their e-polynomials. To appear in: Procedings of the LMS, 20(3):241–247, 2017. [2] B. Frankel. Representations of fundamental groups of abelian varieties in characteristic p. Dissertations available from ProQuest, Paper AAI10134944. [3] F. G. Frobenius. Verallgemeinerung des Sylow’schen satzes. Sitzungsberichte der Königlich Preussischen Akademie der Wissenschaften zu Berlin, pages 981–993, 1895. [4] M. Gerstenhaber. On dominance and varieties of commuting matrices. Annals of Mathematics, 73:324–348, 1961. [5] C. Gordon and F. Rodriguez-Villegas. On the divisibility of #Hom(Γ, G) by |G|. Journal of Algebra, 350(1):300–307, 2012. [6] R. M. Guralnick and B. A. Sethuraman. Commuting pairs and triples of matrices and related varieties. Linear Algebra and its Applications, 310(1-3):139–148, 2000. [7] T. Hausel and F. Rodriguez-Villegas. Mixed Hodge polynomials of character varieties. Inventiones Mathematicae, 174:555–624, 2008. [8] M. I. Isaacs and G. R. Robinson. On a theorem of Frobenius: Solutions of xn = 1 in finite groups. The American Mathematical Monthly, 99(4):352–354, 1992. [9] S. Lang. Algebra. Number 217 in Graduate Texts in Mathematics. Springer-Verlag, New York, revised third edition, 2000. [10] I. G. MacDonald. Symmetric Functions and Hall Polynomials. Oxford Mathematical Monographs. Oxford University Press, New York, second edition, 1995. [11] D. Marker. Model Theory: An Introduction. Number 217 in Graduate Texts in Mathematics. Springer-Verlag, New York, 2000. [12] R. Pries and K. Stevenson. A survey of Galois theory of curves in characteristic p. Fields Institute Communications, 60:169–191, 2011. [13] J-P. Serre. Galois Cohomology. Springer Monographs in Mathematics. Springer-Verlag, Berlin, corrected second printing of the second edition, 2002. [14] L. Solomon. The solution of equations in groups. Archiv der mathematik, 20(3):241–247, 1969. 15 Department of Mathematics, Northwestern University. 2033 Sheridan Road, Evanston, IL 60208, USA [email protected] 16
4math.GR
arXiv:1708.02218v3 [cs.CV] 12 Feb 2018 Graph Classification with 2D Convolutional Neural Networks Antoine J.-P. Tixier Giannis Nikolentzos École Polytechnique [email protected] École Polytechnique and AUEB [email protected] Polykarpos Meladianos Michalis Vazirgiannis École Polytechnique and AUEB [email protected] École Polytechnique and AUEB [email protected] ABSTRACT 2 Graph learning is currently dominated by graph kernels, which, while powerful, suffer some significant limitations. Convolutional Neural Networks (CNNs) offer a very appealing alternative, but processing graphs with CNNs is not trivial. To address this challenge, many sophisticated extensions of CNNs have recently been introduced. In this paper, we reverse the problem: rather than proposing yet another graph CNN model, we introduce a novel way to represent graphs as multi-channel image-like structures that allows them to be handled by vanilla 2D CNNs. Experiments reveal that our method is more accurate than state-of-the-art graph kernels and graph CNNs on 4 out of 6 real-world datasets (with and without continuous node attributes), and close elsewhere. Our approach is also preferable to graph kernels in terms of time complexity. Code and data are publicly available1 . The state-of-the-art in graph classification is currently dominated by a family of methods referred to as graph kernels. Graph kernels compute the similarity between two graphs as the sum of the pairwise similarities between some of their substructures, and then pass the similarity matrix computed on the entire dataset to a kernel-based supervised algorithm such as the Support Vector Machine [7] to learn soft classification rules. Graph kernels mainly vary based on the substructures they use, which include random walks [10], shortest paths [2], and subgraphs [28], to cite only a few. While graph kernels have been very successful, they suffer significant limitations: L1: High time complexity. This problem is threefold: first, populating the kernel matrix requires computing the similarity between every two graphs in the training set (say of size N ), which amounts to N (N −1)/2 operations. The cost of training therefore increases much more rapidly than the size of the dataset. Second, computing the similarity between a pair of graphs (i.e., performing a single operation) is itself polynomial in the number of nodes. For instance, the time complexity of the shortest path graph kernel is O(|V1 | 2 |V2 | 2 ) for two graphs (V1 , V2 ), where |Vi | is the number of nodes in graph Vi . Processing large graphs can thus become prohibitive, which is a serious limitation as big networks abound in practice. Finally, finding the support vectors is O(N 2 ) when the C parameter of the SVM is small and O(N 3 ) when it gets large [4], which can again pose a problem on big datasets. L2: Disjoint feature and rule learning. With graph kernels, the computation of the similarity matrix and the learning of the classification rules are two independent steps. In other words, the features are fixed and not optimized for the task. L3: Graph comparison is based on small independent substructures. As a result, graph kernels focus on local properties of graphs, ignoring their global structure [24]. They also underestimate the similarity between graphs and suffer unnecessarily high complexity (due to the explosion of the feature space), as substructures are considered to be orthogonal dimensions [32]. KEYWORDS Deep Learning, Graphs Kernels, Neural Networks, Convolution, Social Networks, Bioinformatics. 1 GRAPH CLASSIFICATION Graphs, or networks, are rich, flexible, and universal structures that can accurately represent the interaction among the components of many natural and human-made complex systems. A central graph mining task is that of graph classification (not to be mistaken with node classification). The instances are full graphs and the goal is to predict the category they belong to. The applications of graph classification are numerous and range from determining whether a protein is an enzyme or not in bioinformatics, to categorizing documents in NLP, and social network analysis. Graph classification is the task of interest in this study. 1 link will be provided upon acceptance Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. Conference’17, July 2017, Washington, DC, USA © 2018 Association for Computing Machinery. ACM ISBN 978-x-xxxx-xxxx-x/YY/MM. . . $15.00 https://doi.org/10.1145/nnnnnnn.nnnnnnn LIMITATIONS OF GRAPH KERNELS 3 PROPOSED METHOD 3.1 Proposed method: overview We propose a simple approach to turn a graph into a multichannel image-like structure suitable to be processed by a traditional 2D CNN. The process (summarized in Figure 1) can be broken down into 3 steps: (1) graph node embedding, Figure 1: Our 3-step approach represents graphs as “images” suitable to be passed to vanilla 2D CNNs (continuous node attribute vectors can be passed as extra channels). Steps 1 & 2: graph node embeddings and compression with PCA. Step 3: computation and stacking of the 2D histograms. (2) embedding space compression, (3) repeated extraction of 2D slices from the compressed space and computation of a 2D histogram for each slice. The “image” representation of the graph is finally given by the stack of its 2D histograms (each histogram making for a channel). Note that the dimensionality of the final representation of a graph does not depend on its number of nodes or edges. Big and small graphs are represented by images of the same size. Our method addresses the limitations of graph kernels (in bold) in the following ways: L1. By converting all graphs in a given dataset to representations of the same dimensionality, and by using a classical 2D CNN architecture for processing those graph representations, our method offers constant time complexity at the instance level, and linear time complexity at the dataset level. Moreover, state-of-the-art node embeddings can be obtained for a given graph in linear time w.r.t. the number of nodes in the graph, for instance with node2vec [12]. L2. Thanks to the 2D CNN classifier, features are learned directly from the raw data during training to optimize performance on the downstream task. L3. Our approach capitalizes on state-of-the-art graph node embedding techniques that capture both local and global properties of graphs. In addition, we remove the need for handcrafted features. 3.2 Proposed method: details How to represent graphs as structures that verify the spatial dependence property? Convolutional Neural Networks (CNNs) are feedforward neural networks specifically designed to work on regular grids [18]. A regular grid is the d-dimensional Euclidean space discretized by parallelotopes (rectangles for d = 2, cuboids for d = 3, etc.). Regular grids satisfy the spatial dependence2 property, which is the fundamental premise on which local receptive fields and hierarchical composition of features in CNNs hold. Traditionally, a graph G(V , E) is encoded as its adjacency matrix A or Laplacian matrix L. A is a square matrix of dimensionality |V | × |V |, symmetric in the case of undirected graphs, whose (i, j)th entry Ai, j is equal to the weight of the edge ei, j between nodes vi and v j , if such an edge exists, or to 0 otherwise. On the other hand, the Laplacian matrix L is equal to D − A, where D is the diagonal degree matrix. One could initially consider passing one of those structures as input to a 2D CNN. However, unlike in images, where close pixels are more strongly correlated than distant pixels, adjacency and Laplacian matrices are not associated with spatial dimensions and the notion of Euclidean distance, and thus do not satisfy the spatial dependence property. As will be detailed next, we capitalize on graph node embeddings to address this issue. Step 1: Graph node embeddings. There is local correlation in the node embedding space. In that space, the Euclidean distance between two points is meaningful: it is inversely proportional to the similarity of the two nodes they represent. For instance, two neighboring points in the embedding space might be associated with two nodes playing the same structural role (e.g., of flow control), belonging to the same community, or sharing some other common property. Step 2: Alignment and compression with PCA. As state-of-theart node embedding techniques (such as node2vec) are neural, they are stochastic. Dimensions are thus recycled from run to run, which means that a given dimension will not be associated with the same latent concepts across graphs, or across several runs on the same graph. Therefore, to ensure that the embeddings of all the graphs in the collection are comparable, we apply PCA and retain the first d ≪ D principal components (where D is the dimensionality of the original node embedding space). PCA also serves an information maximization (compression) purpose. Compression is desirable as it greatly reduces the shape of the tensors fed to the CNN (for reasons that will become clear in what follows), and thus complexity, at the expense of a negligible loss in information. 2 the concept of spatial dependence is well summarized by: “everything is related to everything else, Step 3: Computing and stacking 2D histograms. We finally repeatedly extract 2D slices from the d-dimensional PCA node embedding space, and turn those planes into regular grids by discretizing them into a finite, fixed number of equally-sized bins, where the value associated with each bin is the count of the number of nodes falling into that bin. In other words, we represent a graph as a stack of d/2 2D histograms of its (compressed) node embeddings3 . As illustrated in Figure 2, the first histogram is computed from the coordinates but near things are more related than distant things” [31]. For instance in images, close pixels are more related than distant pixels. 3 our representation is unrelated to the widespread color histogram encoding of images. channel 1 channel 2 compressed embeddings counts 3 3 40 2 2 30 1 1 4 20 0 0 20 dimension 4 3 30 1 2 4 dimension 2 0 1 2 3 40 bivariate histogram 4 counts 4 bivariate histogram 0 compressed embeddings −2 −1 0 1 2 3 −2 −1 4 0 1 2 3 4 10 −2 0 −2 −2 −2 10 −2 −1 dimension 1 dimension 1 0 1 2 3 dimension 3 4 0 −2 −1 0 1 2 3 dimension 3 4 Figure 2: Node embeddings and image representation of graph ID #10001 (577 nodes, 1320 edges) from the REDDIT-12K dataset. of the nodes in the plane made of the first two principal directions, the second histogram from directions 3 and 4, and so forth. Note that using adjacent and following PCA dimensions is an arbitrary choice. It ensures at least that channels are sorted by decreasing order of informativeness. Using computer vision vocabulary, bins can be viewed as pixels, and the 2D slices of the embedding space as channels. However, in our case, instead of having 3 channels (R,G,B) like with color images, we have d/2 of them. That is, each pixel (each bin) is associated with a vector of size d/2, whose entries are the counts of the nodes falling into that bin in the corresponding 2D slice of the embedding space. Finally, the resolution of the image is determined by the number of bins of the histograms, which is constant for a given dataset across all channels. 4 EXPERIMENTAL SETUP 4.1 2D CNN Architecture We implemented a variant of LeNet-5 [18] with which we reached 99.45% accuracy on the MNIST handwritten digit classification dataset. As illustrated in Figure 3 for an input of shape (5,28,28), this simple architecture deploys four convolutional-pooling layers (each repeated twice) in parallel, with respective region sizes of 3, 4, 5 and 6, followed by two fully-connected layers. Dropout [30] is employed for regularization at every hidden layer. The activations are ReLU functions (in that, our model differs from LeNet-5), except for the ultimate layer, which uses a softmax to output a probability distribution over classes. For the convolution-pooling block, we employ 64 filters at the first level, and as the signal is halved through the (2,2) max pooling layer, the number of filters in the subsequent convolutional layer is increased to 96 to compensate for the loss in resolution. Input (5,28,28) ConvPoolDropx2 ConvPoolDropx2 ConvPoolDropx2 ConvPoolDropx2 filter shape: (3,3) (96,5,5) filter shape: (4,4) (96,4,4) filter shape: (5,5) (96,4,4) filter shape: (6,6) (96,3,3) ReLU (all) FlattenConcatDrop (6336) DenseDrop (128) Output (#classes) ReLU Softmax Figure 3: 2D CNN architecture used in our experiments. The number within parentheses refer to the output dimensions of the tensors. 4.2 Neural embeddings To learn node embeddings, we used the high performance C++ implementation4 of the node2vec algorithm [12]. node2vec applies the very fast Skip-Gram language model [20] to truncated biased random walks performed on the graph. The algorithm scales linearly with the number of nodes in the network. 4.3 Real-world datasets We conducted experiments on 6 real-world datasets which we describe in what follows, and in Table 1. In all datasets, graphs are unweighted, undirected, with unlabeled nodes, and the task is to predict the class they belong to. Classes are mutually exclusive. In the first five datasets, we can only learn from the topology of the network, while in the sixth, we also have access to continuous node attributes. SOCIAL NETWORK DATASETS5 [32]. Note: in all REDDIT datasets, a graph corresponds to a thread where nodes represent users, and there is an edge between two nodes if one of the two users responded to a comment from the other user. • REDDIT-B graphs are labeled according to whether they were constructed from Q&A or discussion communities, • REDDIT-5K & REDDIT-12K feature graphs taken respectively from 5 and 11 forums dedicated to specific topics, • COLLAB graphs are hop-1 neighborhoods of researchers from a scientific collaboration network (two researchers are linked if they co-authored a paper), and are labeled according to the subfield of Physics the corresponding researcher belongs to, • IMDB-B features hop-1 neighborhoods of actors and actresses selected from two movie collaboration networks corresponding to specific genres (action and romance), in which two actors are linked if they starred in the same movie. Graphs are labeled according to the genre they were sampled from. BIOINFORMATICS DATASET [3, 14] (with continuous node attributes). • PROTEINS_full proteins are represented as graphs where nodes are secondary structure elements (SSEs) and there is an edge between two nodes if they are neighbors in the amino-acid sequence or in 3D space. The goal is to predict whether the protein is an enzyme or not. Furthermore, each node is associated with a 29dimensional continuous vector representing chemical measurements. 4 https://github.com/snap-stanford/snap/tree/master/examples/node2vec 5 http://www.mit.edu/~pinary/kdd/datasets.tar.gz To incorporate this extra information into our images, we compressed the attribute vectors with PCA, retaining the same number of dimensions as for the node embeddings, and normalized them to have same range as the node embeddings. Finally, each node was represented by a single vector made of its compressed node embedding concatenated with its compressed and normalized continuous attribute vector. That is, we had d/2 channels for node embeddings and d/2 channels for node attributes, where d is the number of principal components retained in each case. 4.4 Baselines On the social network datasets (on which there are no continuous node attributes), we re-implemented two state-of-the-art graph kernels, the graphlet kernel and the Weisfeiler-Lehman (WL) kernel. • the graphlet kernel [28] computes the similarity between two graphs as the cosine of their count vectors. These vectors encode how many subgraphs of size up to a certain threshold can be found in each graph (each entry is an occurrence count). We sampled 2000 graphlets of size up to 6 from each graph. • the Weisfeiler-Lehman (WL) subtree kernel [29]. The WL kernel is actually a framework that operates on top of any graph kernel accepting node labels and boosts its performance by using the relabeling procedure of the WL test of isomorphism. More precisely, following the computation of the kernel value between the two graphs, vertex labels are updated based on the labels of their neighbors. This two-step process repeats for a certain number of iterations. The final kernel value is the sum of the values at each iteration. Since our graphs have unlabeled nodes, we use their degrees as labels. Furthermore, we used the WL framework with the subtree graph kernel [10], as it is very efficient with this kernel [29]. We also report for comparison purposes the performance of multiple state-of-the-art baselines (the experimental setting in each case is the same as ours): • Deep Graph Kernels [32]. To be fair, we exclude this baseline from the comparison on the bioinformatics dataset since it doesn’t make use of node attributes. For all datasets: • PATCHY-SAN (PSCN k = 10), which is the best performing graph CNN model presented in [23], • Deep Graph CNN (DGCNN) [33]. On the PROTEINS_full dataset (all baselines below take into account node attribute vectors): • HGK-SP Hash Graph Kernel with shortest-path base kernel [21], • HGK-WL Hash Graph Kernel with Weisfeiler-Lehman subtree base kernel [21], • Graph invariant kernels (GIK) [25] (the best performing variant), • GraphHopper [9] and baselines within (all using continuous node attributes): • PROP-diff propagation kernel with diffusion scheme [22], • PROP-WL propagation kernel with hashing-based label discretization and WL kernel update [22]. 4.5 Configuration Following the conventional experimental set-up, we used 10fold cross validation and repeated each fold 3 times in all our experiments. For the graphlet and WL kernels, we used a C-SVM classifier6 [26]. The C parameter of the SVM and the number of iterations in WL were jointly optimized on a 90-10  % partition of the training set of each fold by searching the grid (10−4 , 104 , len = 10); (2, 7, step = 1) . For our 2D CNN, we used Xavier initialization [11], a batch size of 32, and for regularization, a dropout rate of 0.3 and early stopping with a patience of 5 epochs (null delta). The categorical cross-entropy loss was optimized with Adam [15] (default settings). We implemented our model in Keras [6] version 1.2.27 with tensorflow [1] backend. The hardware used consisted in an NVidia Titan X Pascal GPU with an 8-thread Intel Xeon 2.40 GHz CPU and 16 GB of RAM, under Ubuntu 16.04.2 LTS 64-bit operating system and Python 2.7. The graph kernel baselines were run on an 8-thread Intel i7 3.4 GHz CPU, with 16 GB of RAM, under Ubuntu 16.06 LTS 64-bit operating system and Python 2.7. 4.6 Resolution, channels, and node2vec parameters Image resolution. In our initial experiments involving spectral embeddings, the coordinates of any node in any dimension belonged to the [−1, 1] range, due to the eigenvectors being unit-normed. Furthermore, inspired by the MNIST images which are 28 × 28 in size, and on which we initially tested our 2D CNN architecture, we decided to learn 2D histograms featuring 28 bins in each direction. This gave us a resolution of 28/(1−(−1)), that is, 14 pixels per unit (or simply 14:1). As it was giving good results, we stuck to similar values in our subsequent experiments making use of neural embeddings. The only other value we experimented with was 9:1. Number of channels. With the p and q parameters of node2vec held constant and equal to 1, we conducted a search on the coarse grid (14,9);(2,5) to get more insights about the impact of resolution and number of channels (respectively). When using 5 channels, the graphs with less than 10 nodes were removed, because for these graphs, we cannot get a 10-dimensional node embedding space (we cannot have more dimensions than data points). However, this represented only a couple of graphs overall. On the PROTEINS_full dataset, we only experimented with 2 node embeddings channels and 2 node attributes channels. Image size. On a given dataset, image size is calculated as the range |max(coordinates)−min(coordinates)|×resolution, where coordinates are the flattened node embeddings. For instance, on COLLAB with a resolution of 9:1, image size is equal to 37 × 37, since |2.78 − (−1.33)| × 9 ≈ 37. p,q,c, and dn2v node2vec parameters. With the best resolution and number of channels, we then tuned the return and in-out parameters p and q of node2vec. Those parameters respectively bias the random walks towards exploring larger areas of the graph or staying in local neighborhoods, allowing the embeddings to encode a similarity that interpolates between structural equivalence (two nodes acting as, e.g., flow controllers, are close to each other) and 6 http://scikit-learn.org/stable/modules/generated/sklearn.svm.SVC.html 7 https://faroit.github.io/keras-docs/1.2.2/ Max # vertices Min # vertices Average # vertices Max # edges Min # edges Average # edges # graphs Average diameter Average density (%) # classes Max class imbalance IMDB-B COLLAB REDDIT-B REDDIT-5K REDDIT-12K PROTEINS_full 136 12 19.77 1249 26 96.53 1000 1.861 52.06 2 1:1 492 32 74.49 40120 60 2457.78 5000 1.864 50.92 3 1:3.4 3782 6 429.61 4071 4 497.75 2000 9.72 2.18 2 1:1 3648 22 508.50 4783 21 594.87 4999 11.96 0.90 5 1:1 3782 2 391.40 5171 1 456.89 11929 10.91 1.79 11 1:5 620 4 39.05 1049 5 72.82 1113 11.57 21.21 2 1:1.5 Table 1: Statistics of the social network datasets (first 5 columns) and the bioinformatics dataset used in our experiments. homophily (two nodes belonging to the same community are close to each other). Following the node2vec paper, we tried 5 combinations of values for (p, q): (1, 1); (0.25, 4); (4, 0.25); (0.5, 2); (2, 0.5) . Note that p = q = 1 is equivalent to DeepWalk [27]. Since the graphs in the COLLAB and the IMDB-B datasets are very dense (> 50%, average diameter of 1.86), we also tried to set the context size c to smaller values of 1 and 2 (the default value is c = 10). The context size is used when generating the training examples (context,target) to pass to the skip-gram model: in a given random walk, it determines how many nodes before and after the target node should be considered part of the context. Finally, since the graphs in COLLAB, IMDB-B, and PROTEINS_full are small (20, 74, and 39 nodes on average), we experimented with lower values of dn2v , namely 12 and 4. dn2v is the dimension of the node embeddings learned by node2vec (default value is 128). The final values of p, q, c, and dn2v for each dataset are summarized in Table 2. Res. #Chann. p,q c,dn2v REDDIT-B 9:1 5 2,0.5 - REDDIT-5K 9:1 2 4,0.25 - REDDIT-12K 9:1 5 1,1 - COLLAB 9:1 5 0.5,2 2,12 IMDB-B 14:1 5 1,1 - PROTEINS_full 9:1 2/2⋆ 0.5,2 -,4 Table 2: Final resolution, number of channels, and p, q, c, and d n2v node2vec parameters for each dataset. ⋆ number of channels for node embeddings and continuous node attributes. - means default value(s). 5 RESULTS The classification accuracy of our approach in comparison to the baselines is reported in Table 3 for the social network datasets and Table 4 for the bioinformatics dataset. Our approach shows (statistically) significantly better than all baselines on the REDDIT-12K and REDDIT-B datasets, with large improvements of 6.81 and 2.82 in accuracy over the best performing competitor, respectively. We also reach best performance on the REDDIT-5K and PROTEINS_full datasets, with respective improvements in accuracy of 1.34 and 0.52 over the best performing baselines. In particular, the fact that we reach best performance on PROTEINS_full shows that our approach is flexible enough to leverage not only the topology of the network, but also continuous node attributes, in a very simple and unified way (we simply concatenate node embeddings with node attribute vectors). Note that when not using node attributes (only the first 2 node embeddings channels), the performance of our model decreases from 77.12 to 73.43 on PROTEINS_full (3.69 decrease), which proves that the skill of our model comes from leveraging both node embeddings and attributes. Finally, on the IMDB-B dataset, we get third place, very close (≤ 1.2) to the top performers (no statistically significant difference). The only dataset on which a baseline proved significantly better than our approach is actually COLLAB (WL graph kernel). On this dataset though, the WL kernel beats all models by a wide margin, and we are relatively close to the other Deep Learning approaches (≤ 2.43). 5.1 Runtimes Even if not directly comparable, we report in Table 5 kernel matrix computation time for the two graph kernel baselines, along with the time required by our 2D CNN model to perform one pass over the entire training set, i.e., the time per epoch. With respects to time complexity, our method is superior to graph kernels on several counts: first, unlike graph kernels, the time required by the 2D CNN to process one training example is constant (all images for a given dataset have the same size), while computing the kernel value for a pair of graphs depends on their size (polynomial in the number of nodes). It is true that a prerequisite for our approach is an embedding for all the graphs in the dataset, but node2vec scales linearly with the number of nodes in the graph. Therefore, on big graphs, our method is still usable, while graph kernels may not be. Also, node2vec is easily parallelizable over the collection, so one can take advantage of multi-core CPUs to considerably speed up the process. Second, with a 2D CNN, the time necessary to go through the entire training set only increases linearly with the size of the set, while populating the kernel matrix is quadratic, and finding the support vectors is then again at least quadratic. This means that on large datasets, our approach is also preferable to graph kernels. Examples of 2D CNN architectures much more complex than ours applied to millions of images in reasonable time abound in the recent computer vision literature. Processing such big Table 3: 10-fold CV average test set classification accuracy of our proposed method compared to state-of-the-art graph kernels and graph CNNs, on the social network datasets. ± is standard deviation. Best performance per column in bold. ⋆ indicates stat. sign. at the p < 0.05 level (our 2D CNN vs. WL) using the Mann-Whitney U test (https://docs.scipy.org/doc/scipy-0.19.0/reference/generated/scipy.stats.mannwhitneyu.html). Dataset Acc. Std. dev. REDDIT-B REDDIT-5K REDDIT-12K COLLAB IMDB-B Method (size=2,000;nclasses=2) (4,999;5) (11,929;11) (5,000;3) (1,000;2) Graphlet Shervashidze2009 WL Shervashidze2011 Deep GK Yanardag2015 PSCN k = 10 Niepert2016 DGCNN Zhang2018 2D CNN (our method) 77.26 (± 2.34) 78.52 (± 2.01) 78.04 (± 0.39) 86.30 (± 1.58) 89.12⋆ (± 1.70) 39.75 (± 1.36) 50.77 (± 2.02) 41.27 (± 0.18) 49.10 (± 0.70) 52.11 (± 2.24) 25.98 (± 1.29) 34.57 (± 1.32) 32.22 (± 0.10) 41.32 (± 0.42) 48.13⋆ (± 1.47) 73.42 (± 2.43) 77.82⋆ (± 1.45) 73.09 (± 0.25) 72.60 (± 2.15) 73.76 (± 0.49) 71.33 (± 1.96) 65.40 (± 5.95) 71.60 (± 5.16) 66.96 (± 0.56 ) 71.00 (± 2.29) 70.03 (± 0.86) 70.40 (± 3.85) 2D CNN our method 77.12 2.79 DGCNN Zhang18 75.54 0.94 PSCN k = 10 Niepert16 75.00 2.51 HGK-SP Morris16 75.14 0.47 HGK-WL Morris16 74.88 0.64 GIK Orsini15 76.6 0.6 GraphHopper Feragen13 74.1 0.5 PROP-diff Neumann12 73.3 0.4 PROP-WL Neumann12 73.1 0.8 Table 4: 10-fold CV average test set classification accuracy of our proposed method compared to state-of-the-art graph kernels and graph CNNs on the bioinformatics dataset (PROTEINS_full). REDDIT-B REDDIT-5K REDDIT-12K COLLAB IMDB-B Size, average (# nodes, # edges) 2000, (430,498) 4999, (509,595) 11929, (391,457) 5000, (74,2458) 1000, (20,97) 1113, (39,73) Input shapes (for our approach) (5,62,62) (2,65,65) (5,73,73) (5,36,36) (5,37,37) (4,70,70) (2,70,70)⋆ Graphlet Shervashidze2009 WL Shervashidze2011 551 645 5046 5087 12208 20392 3238 1579 275 23 - 6 16 52 5 1 1 2D CNN (our approach) PROTEINS_full Table 5: Runtimes in seconds, rounded to the nearest integer. For the graph kernel baselines, time necessary to populate the Kernel matrix (8-thread 3.4GHz CPU). For our model, time per epoch (Titan X Pascal GPU). ⋆ with and without using node attributes datasets with graph kernels would simply be intractable. In addition, the performance of Deep Learning models tends to significantly improve in the presence of large quantities of training data. 6 RELATED WORK Motivated by the outstanding performance recently reached by Convolutional Neural Networks (CNNs) in computer vision, e.g. [17? ], many research efforts have been devoted to generalizing CNNs to graphs. Indeed, CNNs offer a very appealing alternative to kernel-based methods. The parsimony achieved through weight sharing makes them very efficient, their time complexity is constant for each training example and linear with respect to the size of the dataset, and the extra expressiveness they bring might translate to significant accuracy gains. However, since convolution and pooling are natively defined for regular, low-dimensional grids such as images (2D Euclidean space discretized by rectangles), generalizing CNNs to graphs, which are irregular, non-Euclidean objects, is far from trivial. Possible solutions that can be found in the literature fall into two broad categories: spatial and spectral techniques [5]. Spectral approaches [8, 16] invoke the convolution theorem from signal processing theory to perform graph convolutions as pointwise multiplications in the Fourier domain of the graph. The basis used to send the graph to the Fourier domain is given by the SVD decomposition of the Laplacian matrix of the graph, whose eigenvalues can be viewed as “frequencies”. By contrast, spatial methods [23, 33? ] operate directly on the graph structure. For instance, in [23], the algorithm first determines the sequence of nodes for which neighborhood graphs (of equal size) are created. To serve as receptive fields, the neighborhood graphs are then normalized, i.e., mapped to a vector space with a linear order, in which nodes with similar structural roles in the neighborhood graphs are close to each other. Normalization is the central step, and is performed via a labeling procedure. A 1D CNN architecture is finally applied to the receptive fields. 6.1 Departure from previous work While the aforementioned sophisticated frameworks have made great strides, we showed in this paper that graphs can also be processed by vanilla 2D CNN architectures. This is made possible by the novel graph representation we introduce, which encodes graphs as stacks of 2D histograms of their node embeddings (and continuous attribute vectors, if available). Compared to the more complex approaches that involve different architectural and/or operational modifications, the main advantage of our method is its simplicity. Crucially, we show that this simplicity can be obtained without giving up accuracy: we indeed outperform graph CNN baselines by a wide margin on some datasets, and are very close elsewhere. graph kernels too, allowing to process bigger datasets featuring larger graphs. 6.2 REFERENCES Discussion Replacing the raw counts by the empirical joint probability density function, either by normalizing the histograms, or with a Kernel Density Estimate, significantly deteriorated performance. This suggests that keeping the absolute values of the counts is important, which makes sense, because some categories might be associated with larger or smaller graphs, on average. Therefore, preventing the model from using size information is likely to decrease accuracy. We also observed that increasing the number of channels to more than 5 does not yield better results (which makes sense, as channels contain less and less information), but that reducing this number improves performance in some cases, probably because it plays a regularization role. The main contribution of our study is a novel method for representing graphs as multi-channel image-like structures from their node embeddings, that allows them to be processed by 2D CNNs. How the embeddings are computed, and which 2D CNN architecture is used, does not matter. We hold this flexibility to be a major strength. First, the embedding-agnostic nature of our method means that it can be seamlessly extended to directed, weighted, or labeled graphs with continuous or categorical node/edge attributes, simply by using an embedding algorithm that accepts such graphs, e.g., [19]. The independence of our approach with respect to the image classification model used is another advantage. Here, we employed a vanilla 2D CNN architecture as it was offering an excellent tradeoff between accuracy and simplicity, but more recent models, such as the one of [13], may yield even better results. Above all, performance should improve as graph node embedding algorithms and CNN architectures for images improve in the future. Limitations. Even though results are very good out-of-the-box in most cases, finding an embedding algorithm that works well, or the right combination of parameters for a given dataset, can require some efforts. For instance, on COLLAB and IMDB-B (the only two datasets on which we do not reach best performance), we hypothesize that our results are inferior to that observed elsewhere because the default parameter values of node2vec may not be wellsuited to very dense graphs such as the ones found in COLLAB and IMDB-B (diameter< 2, density > 50). Optimizing the node2vec parameters on these datasets probably requires more than a coarse grid search. 7 CONCLUSION The main contribution of this paper is to show that CNN architectures designed for images can be used for graph processing in a completely off-the-shelf manner, simply by representing graphs as stacks of two-dimensional histograms of their node embeddings. Our approach is flexible and continuous node attributes can be taken into account by passing them as additional channels. Despite the simplicity of our approach, we reach better results than stateof-the-art graph kernels and graph CNN models on 4 real-world datasets out of 6. Furthermore, these good results were obtained with limited parameter tuning and by using a basic 2D CNN model. From a time complexity perspective, our approach is preferable to [1] Martín Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, et al. 2016. Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint arXiv:1603.04467 (2016). [2] K. M. Borgwardt and H. Kriegel. 2005. Shortest-path kernels on graphs. In Proceedings of the 5th International Conference on Data Mining. 74–81. [3] Karsten M Borgwardt, Cheng Soon Ong, Stefan Schönauer, SVN Vishwanathan, Alex J Smola, and Hans-Peter Kriegel. 2005. Protein function prediction via graph kernels. Bioinformatics 21, suppl 1 (2005), i47–i56. [4] Léon Bottou and Chih-Jen Lin. 2007. Support vector machine solvers. Large scale kernel machines 3, 1 (2007), 301–320. [5] Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. 2013. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203 (2013). [6] François Chollet et al. 2015. Keras. https://github.com/fchollet/keras. (2015). [7] Corinna Cortes and Vladimir Vapnik. 1995. Support-vector networks. Machine learning 20, 3 (1995), 273–297. [8] Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. 2016. Convolutional neural networks on graphs with fast localized spectral filtering. In Advances in Neural Information Processing Systems. 3837–3845. [9] Aasa Feragen, Niklas Kasenburg, Jens Petersen, Marleen de Bruijne, and Karsten Borgwardt. 2013. Scalable kernels for graphs with continuous attributes. In Advances in Neural Information Processing Systems. 216–224. [10] Thomas Gärtner, Peter Flach, and Stefan Wrobel. 2003. On graph kernels: Hardness results and efficient alternatives. In Learning Theory and Kernel Machines. Springer, 129–143. [11] Xavier Glorot and Yoshua Bengio. 2010. Understanding the difficulty of training deep feedforward neural networks.. In Aistats, Vol. 9. 249–256. [12] Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 855–864. [13] Gao Huang, Zhuang Liu, Kilian Q Weinberger, and Laurens van der Maaten. 2016. Densely connected convolutional networks. arXiv preprint arXiv:1608.06993 (2016). [14] Kristian Kersting, Nils M. Kriege, Christopher Morris, Petra Mutzel, and Marion Neumann. 2016. Benchmark Data Sets for Graph Kernels. (2016). http:// graphkernels.cs.tu-dortmund.de [15] Diederik Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014). [16] Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016). [17] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. 2012. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems. 1097–1105. [18] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradientbased learning applied to document recognition. Proc. IEEE 86, 11 (1998), 2278– 2324. [19] Lizi Liao, Xiangnan He, Hanwang Zhang, and Tat-Seng Chua. 2017. Attributed Social Network Embedding. arXiv preprint arXiv:1705.04969 (2017). [20] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781 (2013). [21] Christopher Morris, Nils M Kriege, Kristian Kersting, and Petra Mutzel. 2016. Faster kernels for graphs with continuous attributes via hashing. In Data Mining (ICDM), 2016 IEEE 16th International Conference on. IEEE, 1095–1100. [22] Marion Neumann, Novi Patricia, Roman Garnett, and Kristian Kersting. 2012. Efficient graph kernels by randomization. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 378–393. [23] Mathias Niepert, Mohamed Ahmed, and Konstantin Kutzkov. 2016. Learning convolutional neural networks for graphs. In Proceedings of the 33rd annual international conference on machine learning. ACM. [24] Giannis Nikolentzos, Polykarpos Meladianos, and Michalis Vazirgiannis. 2017. Matching Node Embeddings for Graph Similarity.. In AAAI. 2429–2435. [25] Francesco Orsini, Paolo Frasconi, and Luc De Raedt. 2015. Graph invariant kernels. In Proceedings of the Twenty-fourth International Joint Conference on Artificial Intelligence. 3756–3762. [26] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 12 (2011), 2825–2830. [27] Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. 2014. Deepwalk: Online learning of social representations. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, 701–710. [28] N. Shervashidze, T. Petri, K. Mehlhorn, K. M. Borgwardt, and S.V.N. Vishwanathan. 2009. Efficient Graphlet Kernels for Large Graph Comparison. In Proceedings of the International Conference on Artificial Intelligence and Statistics. 488–495. [29] N. Shervashidze, P. Schweitzer, E. J. Van Leeuwen, K. Mehlhorn, and K. M. Borgwardt. 2011. Weisfeiler-Lehman Graph Kernels. The Journal of Machine Learning Research 12 (2011), 2539–2561. [30] Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2014. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research 15, 1 (2014), 1929–1958. [31] Waldo R Tobler. 1970. A computer movie simulating urban growth in the Detroit region. Economic geography 46, sup1 (1970), 234–240. [32] Pinar Yanardag and SVN Vishwanathan. 2015. Deep graph kernels. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 1365–1374. [33] Muhan Zhang, Zhicheng Cui, Marion Neumann, and Yixin Chen. 2018. An End-to-End Deep Learning Architecture for Graph Classification. (2018).
1cs.CV
Applying deep learning to classify pornographic images and videos Mohamed N. Moustafa arXiv:1511.08899v1 [cs.CV] 28 Nov 2015 Department of Computer Science and Engineering, The American University in Cairo, New Cairo 11835, Egypt [email protected] Abstract. It is no secret that pornographic material is now a one-clickaway from everyone, including children and minors. General social media networks are striving to isolate adult images and videos from normal ones. Intelligent image analysis methods can help to automatically detect and isolate questionable images in media. Unfortunately, these methods require vast experience to design the classifier including one or more of the popular computer vision feature descriptors. We propose to build a classifier based on one of the recently flourishing deep learning techniques. Convolutional neural networks contain many layers for both automatic features extraction and classification. The benefit is an easier system to build (no need for hand-crafting features and classifiers). Additionally, our experiments show that it is even more accurate than the state of the art methods on the most recent benchmark dataset. Keywords: deep learning, convolutional neural networks, adult image classification 1 Introduction For one reason or another, we aim for filtering out adult images and videos. Most parents, for instance, strive to protect their young children and minors from accessing online porn web sites either intentionally or accidentally. Additionally, in educational and workplace settings, basic ethical and conduct standards dictate that such images become inaccessible to the community. There are many indicators that online pornographic content is increasing exponentially [1], especially in the past few years [11]. Given the continuous flood of uploaded images and videos to general social media networks, one can imagine the difficult and tedious job of filtering out inappropriate material manually by the sites administrators. Therefore, we need to turn our attention to a solution that automatically detects and isolates porn content. Obviously, the solution is a form of an intelligent machine that can analyze the text, audio, or visual signals. In this paper, we focus on the visual cue, being the most salient form of pornography. We propose a deep learning system that automatically analyzes images (and video frames) before classifying the content as regular or porn. Our 2 M. Moustafa proposed solution achieved the highest classification rate, to our knowledge, of more than 94%, on the most recent Pornography benchmark dataset [8]. 2 Related work Most of the existing porn detection research relies on image analysis, based on some form of machine intelligence. Usually, the intelligent processing pipeline includes a feature extractor followed by a classifier. The feature extractor abstracts important information from the input image and feeds it to the pre-trained classifier that takes the decision. Along with the features, mentioned below, Support Vector Machines (SVM) was the most frequently used classifier in recent methods. Historically, the most investigated feature was the human skin color [6]. If the input image contains too much skin colored pixels, it was taken as an indicator of nudity. However, skin color solely is not reliable since a face closeup image has a lot of skin pixels while being non-porn. So, researchers augmented the skin color with other constraints or shape features [10]. With the introduction of new computer vision models, porn detection algorithms became more accurate. For instance, Deselaers et al introduced classification accuracy improvement by adopting the visual bag-of-words (BoW) model [4]. BoW tries to extract the most common patches that exist on a set of training images. For a detailed survey of existing methods, please refer to [9]. Recently, Avila et al [2] introduced a BoW framework, the BossaNova representation, to classify pornographic videos. BossaNova relies on the HueSIFT descriptors. Since HueSIFT represents both color and shape, it outperformed standard BoW models that rely only on shape or edge cues. Additionally, Caetano et al [3] extended the BossaNova to be more suitable for video classification. They applied binary descriptors with BossaNova for even better accuracy on the same benchmark dataset. 3 Deep learning In parallel with the traditional route of handcrafting a set of features and a classifier, another route is witnessing a rising interest: deep learning. This route combines both features extraction and classification into one module with less involvement from the designer in terms of selecting features or classifier. The ultimate target is to have a generic architecture that can learn any problem from its data, thus coming closer in performance to the human brain. Many researchers are now paying more attention to deep learning systems, specifically after the Alex Krizhevsky et al (AlexNet) outstanding performance in the ILSVRC-2012 Imagenet competition [7]. Imagenet challenge includes classifying hundreds of thousands of images to 1000 different possible classes. Krizhevsky used a convolutional neural network (ConvNet) architecture to win this most challenging competition in the area of image classification. Since then, ConvNet appeared in solutions to numerous image analysis problems, e.g., pedestrian detection, traffic Applying deep learning to classify pornographic images and videos 3 sign recognition, image segmentation, image restoration, and object recognition. Recently, the winner of 2014 Imagenet competition was also a ConvNet based system: GoogLeNet [12]. In this paper we propose applying a system of ConvNets to solve the pornographic classification with higher accuracy than the reported state of the art performance [3]. 4 Proposed method We propose applying a combination of ConvNets to classify porn from regular images and video frames. We will first describe the slight modifications we propose to change in the existing AlexNet and GoogLeNet to suit our problem. This is followed by a proposed simple fusion of both networks. 4.1 ANet: AlexNet-based classifier Basically, we adopt the same architecture proposed in [7], sans the output layer. As depicted in figure 1, the net contains five convolutional layers of neurons followed by three fully-connected ones. Each convolution layer filters its input two dimensional vectors with a kernel, whose coefficients are calculated iteratively during the training process. The fully connected layers are simply implementing the dot product between the input and weight vectors, where each neuron in layer l is connected to all outputs of neurons in layer l − 1. All neurons have Rectified Linear Unit (ReLU) activation function to speed up the learning process. The output of the last fully-connected layer is fed to a 2-way softmax which produces a distribution over our two class labels: benign or porn. Fig. 1. ANet: AlexNet-based deep ConvNet porn image classifier. The first five layers (in green) are convolutional followed by three fully connected layers (in red), including the output layer that contains two neurons corresponding to our two classes ‘benign’ or ‘porn’. ANet contains almost 56 million parameters that need to be calculated during the learning process. In order to properly train such a large network, a 4 M. Moustafa huge dataset is needed. The collection and labelling of this dataset would prove impractical. So, instead of starting training from scratch, we trained only the last layer that we newly introduced to the reference network. We started from the reference AlexNet weights for the first seven layers of the ConvNet. Those weights are the results of training the original AlexNet on the Imagenet 1.2 million images from 1000 different classes [5]. During our training, we used benign and porn images as input to the first layer, computed the output of the seventh layer (the feature vector), and only changed the weights of the eighth layer in a supervised way given the ground truth label of the image. This way, we are using the ConvNet as a general feature extractor (the first seven layers) and as a classifier (the last layer). The last layer could also be replaced with any other classifier, e.g., Support Vector Machine (SVM). This fine-tune training method lets the network train faster with less numbers of training data than the full method as it has less parameters to adapt (those of the last layer only). To test a new image, we adjust it by rescaling and subtracting the mean, before feeding it as input to the trained ConvNet. The network has two normalized outputs corresponding to the confidence in each class that sum to 100%. The test image is labeled after the output neuron with the largest score. 4.2 GNet: GoogLeNet-based classifier GoogLeNet, shown in figure 2, comprises 22 layers [12]. It is much deeper than AlexNet and incorporates the inception concept, which performs dimensionality reduction and projection. Each inception module contains two convolutional layers. Again, we have to change the output layers to produce only two labels: benign or porn. We follow the same steps in modifying, training, and testing this ConvNet as we have explained in the ANet classifier subsection above. Fig. 2. GoogLeNet-based deep ConvNet porn image classifier. The first three layers (in green) are convolutional followed by nine inception modules (in yellow), each contains two convolutional layers. The output layer is fully connected and contains two neurons corresponding to our two classes ‘benign’ or ‘porn’. Applying deep learning to classify pornographic images and videos 4.3 5 AGNet: Ensemble-ConvNet classifier When solving a hard problem, what would be better than consulting an expert? It would be having more than one to consult! We expect that ANet and GNet are making different classification mistakes. Therefore, we believe that Our Ensemble-ConvNet , fusing both ANet and GNet scores, should result in a smoother decision that makes less mistakes. We describe in figure 3 the overall architecture of our proposed AGNet classifier. We have followed the same procedure as [7] to pre-process images before feeding it as input to the first convolution layer. The preprocessing step in this case includes rescaling the image to 256 x 256 RGB pixels, subtracting the mean image, and finally dividing it into one or more windows of 224 x 224 pixels each to increase the training data instances. Each of ANet and GNet tests the input image, and produces ‘benign’ and ‘porn’ scores. We combine those scores in the ‘Fusion’ block, and finally decide whether it belongs to the benign or porn classes based on some predetermined threshold. By default, this threshold is 50%. This fusion block could be as complex as another trained nonlinear classifier. On the other hand, it could be as simple as a linear weighted average of scores. We show in our experiments below that a simple weighted average with equal weights is enough to produce superior accuracy than each individual network. To test a video sequence, we extract key frames and test them individually and finally label the video based on majority voting. In the event that the number of benign and porn voted frames is equal, we decide based on the class with the largest sum of scores for all key frames. Fig. 3. Proposed AGNet porn image classifier. 5 Experiments and results We have implemented our proposed ANet, GNet, and AGNet classifiers in C++ relying on the open source frameworks CAFFE [5] to train and test the ConvNets. To put our results in comparison with other state of the arts methods, we have used a recent benchmark dataset described below. 6 5.1 M. Moustafa NDPI dataset The NPDI Pornography database contains nearly 80 hours of 400 pornographic and 400 non-pornographic videos. It has been collected by the NPDI group, Federal University of Minas Gerais (UFMG), Brazil [8]. The database consists of several genres of pornography and depicts actors of many ethnicities, including multi-ethnic ones. The NPDI group has collected the pornographic videos from websites which only host that kind of material. The non-pornographic videos include two subcategories: 200 videos chosen at random (which they called “easy”) and 200 videos (“difficult”), selected from textual search queries like “beach”, “wrestling”, “swimming”, which contain body skin but not nudity or porn. The videos are already segmented into shots. On the average, there are almost 20 shots per video. A key-frame has been selected as simply the middle-frame of each video shot. In total, there are 16,727 key-frames (shots). Tables 1 and 2 highlight some statistics and samples of this dataset respectively. To unify the cross-validation protocol, the dataset is divided into five folds by generating nearly 640 videos for training and 160 for testing on each fold. In our experiments, we have followed the same training/testing 5-folds cross-validation. Table 1. NPDI dataset content Class Videos Hours Shots per Video Porn 400 57 15.6 Non-porn (“easy”) 200 11.5 33.8 Non-porn (“difficult”) 200 8.5 17.5 All videos 800 77 20.6 5.2 Experiment: testing individual images In this experiment, we have tested each classifier on the NPDI videos key-frames as individual images. The objective of this experiment was threefold. First, it was a comparison between the full versus the fine-tune training of the ConvNets. Second, we wanted to check whether GoogLeNet was better than AlexNet in this problem as it was in the Imagenet categorization challenge. Third, we were testing the individual ConvNet classifier versus the ensemble one. We have tested many variations of the ‘fusion’ block mentioned in figure 3. Our findings do not show much differences in the results. Consequently, we decided to choose the equal weighted average of both ConvNets scores for simplicity. In total, we trained 5 ANets, 5 GNets using four of the NPDI five folds of videos and used the fifth one for testing. We recorded the average accuracy of classification for various values of the threshold from 0 to 100 to build Receiver Operating Characteristics (ROC) curves. ROC is very informative since the user can select his or her operating point based on the application. Simply put, at threshold 0, both correct benign and false porn rates will be 1.0, which Applying deep learning to classify pornographic images and videos 7 Table 2. NPDI dataset samples Porn Non-porn (“difficult”) Non-porn (“easy”) corresponds to the top right corner point of the ROC. Increasing the threshold tightens the classifier filtration with lowering both rates. On the other hand, at threshold 100, the classifier will not pass any image and all classification rates will be 0.0, which is the left bottom point of the full ROC. We can see in figure 4, that both ANet and GNet (which were fine-tune trained as described earlier in the previous section) produced higher ROC than their fully trained counterparts. For instance when the rate of classifying porn frames by mistake as benign was 0.1, i.e., 10%, the rate of classifying benign frames correctly stood at 0.68 and 0.87 respectively for ANet fully trained and ANet fine tuned. Similarly, rates for GNet fully trained and fine tuned were 0.80 and 0.91 respectively. The ROC curves also show that AGNet is more accurate than GNet, especially when the rate of false porn classification is very low or very high. 5.3 Experiment: testing videos In this experiment, we have considered the majority voting of all frames, that belong to the same video sequence, in the decision as described earlier in the AGNet section. We recorded the average correct classification rate of all 5 folds and the standard deviation. The reported correct classification rate is the average of both ‘correctly classifying benign as benign’ and ‘correctly classifying porn as porn’. We included in table 3 also the best accuracy numbers reported in the recent literature as a baseline for comparison. As expected, GNet is more accurate than ANet as it has deeper architecture. Additionally, AGNet (with simple equal weighted average) is slightly more accurate than GNet. Note also that AGNet produced smaller variance than either ANet or GNet. For the record, we also reported the results of AGbNet, where 8 M. Moustafa Rate of classifying Benign as Benign 1 0.95 0.9 0.85 0.8 0.75 ANet-Fulltrain GNet-Fulltrain ANet GNet AGNet 0.7 0.65 0.6 0.01 0.1 Rate of classifying Porn as Benign Fig. 4. ConvNet Benign/Porn image classifier ROC. 1 Applying deep learning to classify pornographic images and videos 9 the score ‘fusion’ is the larger of ANet and GNet score instead of the simple average. It was interesting to see the all ConvNet based classifiers outperformed the BossaNova state of the art methods [2] [3] in the average accuracy but with slightly larger standard deviation. This might be due to the random initialization of the neural networks weights during training. We believe that the variance would go lower if we were to add more networks to the ensemble and fine tune some training parameters. Table 3. Videos classification accuracy on NPDI benchmark dataset Approach Accuracy (%) BossaNova (HueSIFT) [2] 89.5±1 BossaNova VD (BinBoost16) [3] 90.9±1 Proposed ANet 92.01±3 Proposed GNet 93.7±3 Proposed AGNet 93.8±2 Proposed AGbNet 94.1±2 6 Conclusions We proposed applying convolutional neural networks to automatically classify pornographic images and videos. We showed that our proposed fully automated solution outperformed the accuracy of hand-crafted feature descriptors solutions. We are continuing our research to find an even better network architecture for this problem. Nevertheless, all the successful applications so far rely on supervised training methods. We expect a new wave of deep learning networks would emerge by combining supervised and unsupervised methods where a network can learn from its mistakes while in actual deployment. We believe further research can also be directed toward allowing machines to consider the context and overall rhetorical meaning of a video clip while relating them to the images involved. References 1. Anthony, S., Just how big are porn sites?, (2012), http://www.extremetech.com/ computing/123929-just-how-big-are-porn-sites 2. Avila, S., Thome, N., Cord, M., Valle, E., and Araujo, A., Pooling in image representation: The visual codeword point of view. Computer Vision and Image Understanding, 117(5), pp. 453 – 465 (2013) 3. Caetano, C., Avila, S., Guimaraes, S., and Araujo, A. Pornography detection using bossanova video descriptor. In Signal Processing Conference (EUSIPCO), 2014 Proceedings of the 22nd European, pp. 1681–1685 (2014) 10 M. Moustafa 4. Deselaers, T., Pimenidis, L., and Ney, H. Bag-of-visual-words models for adult image classification and filtering. In Pattern Recognition, 2008. ICPR 2008. 19th International Conference on, pp. 1–4 (2008) 5. Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadarrama, S., and Darrell, T. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093 (2014) 6. Jones, M.J., and Rehg, J.M. Statistical color models with application to skin detection. In Computer Vision and Pattern Recognition, 1999. IEEE Computer Society Conference on., vol. 1. IEEE (1999) 7. Krizhevsky, A., Sutskever, I., and Hinton, G. Imagenet classification with deep convolutional neural networks. In F. Pereira, C.J.C. Burges, L. Bottou, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pp. 1097– 1105. Curran Associates, Inc., (2012) 8. NPDI Pornography Database, (2013), https://sites.google.com/site/ pornographydatabase/ 9. Ries, C., and Lienhart, R. A survey on visual adult image recognition. Multimedia Tools and Applications, 69(3), pp. 661–688 (2014) 10. Rowley, H., Jing, Y., and Baluja, S. Large scale image-based adult-content filtering. In Int Conference on Comp Vis Theor Appl (VISAPP 2006), pp. 290–296 (2006) 11. Short, M., Black, L., Smith, A., Wetterneck, C., and Wells, D. A review of internet pornography use research: methodology and content from the past 10 years. Cyberpsychology, behavior and social networking, 15(1), pp. 13–23 (2012) 12. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich, A. Going deeper with convolutions. CoRR, abs/1409.4842 (2014)
9cs.NE
Astroinformatics c 2017 International Astronomical Union Proceedings IAU Symposium No. 325, 2017 M. Brescia, S. G. Djorgovski, E. Feigelson, G. Longo & S. Cavuoti. DOI: 00.0000/X000000000000000X What can the programming language Rust do for astrophysics? Sergi Blanco-Cuaresma1 and Emeline Bolmont2 1 arXiv:1702.02951v1 [astro-ph.IM] 9 Feb 2017 2 Observatoire de Genève, Université de Genève, CH-1290 Versoix, Switzerland. email: [email protected] NaXys, Department of Mathematics, University of Namur, 8 Rempart de la Vierge, 5000 Namur, Belgium. email: [email protected] Abstract. The astrophysics community uses different tools for computational tasks such as complex systems simulations, radiative transfer calculations or big data. Programming languages like Fortran, C or C++ are commonly present in these tools and, generally, the language choice was made based on the need for performance. However, this comes at a cost: safety. For instance, a common source of error is the access to invalid memory regions, which produces random execution behaviors and affects the scientific interpretation of the results. In 2015, Mozilla Research released the first stable version of a new programming language named Rust. Many features make this new language attractive for the scientific community, it is open source and it guarantees memory safety while offering zero-cost abstraction. We explore the advantages and drawbacks of Rust for astrophysics by re-implementing the fundamental parts of Mercury-T, a Fortran code that simulates the dynamical and tidal evolution of multi-planet systems. Keywords. Rust, programming languages, N-Body, simulations, exoplanets 1. Introduction Plenty of tools in astrophysics are developed using system programming languages such as Fortran, C or C++. These languages are known to provide high performance and fast executions but they rely heavily on the developer for concurrency and memory control, which may lead to common errors as shown in Fig. 1: a) access to invalid memory regions, b) dangling pointers and attempts to free already freed memory, c) memory leaks and, d) race conditions. This can produce random behaviors and affect the scientific interpretation of the results. The recently created language Rust prevents such problems and fields like bioinformatics (Köster, 2015) have already started to take advantage of it. Astroinformatics can benefit from it too. We first discuss the general principles behind this new language and what makes it attractive when compared to more traditional languages such as C or Fortran. We then show that this language can reach the same performance as a Fortran N-Body simulator, Mercury-T† (Bolmont et al., 2015), designed for the study of the tidal evolution of multi-planet systems. 2. Rust Mozilla Research, motivated by the development of a new web browser engine (Anderson et al., 2015), released in 2015 the first stable version of a new open source pro† http://www.emelinebolmont.com 1 2 Sergi Blanco-Cuaresma & Emeline Bolmont a) b) c) Read / Write 4th element Allocate memory Allocate memory 1 2 Free 3 Your vector Neighboor's memory Free a 2 time d) Shared data Task 1 Forget to free nd Access freed memory Allocate memory Read Write Out of memory! Read Task 2 Overwrite Figure 1. Typical memory and concurrency problems with system programming languages such as Fortran, C or C++. gramming language named Rust. It uses patterns coming from functional programming languages (Poss, 2014) and it is designed not only for performance and concurrency, but also for safety. Rust introduces concepts like ownership, borrowing and variable lifetime, which: - facilitates the automatic control of the lifetime of objects during compilation time. There is no need for manually freeing resources or for an automated garbage collector like in Java or Go; - prevents the access to invalid memory regions; - enforces thread-safety (race conditions cannot occur). These zero-cost abstraction features make Rust very attractive for the scientific community with high performance needs. In Rust, variables are non-mutable by default (unless the mutable keyword is used) and they are bound to their content (i.e, they own it or they have ownership of it). When you assign one variable to another (case ’a’ in Fig. 2), you are not copying the content but transferring the ownership†, so that the previous variable does not have any content (like when we give a book to a friend, we stop having access to it). This transfer takes also place when we call functions (case ’b’ in Fig. 2), and it is important to note that Rust will free the bound resource when the variable binding goes out of scope (at the end of the function call for case ’b’). Hence, we do not have to worry about freeing memory and the compiler will validate for us that we are not accessing a memory region that has already been freed (errors are caught at compilation time, before execution time). Additionally, apart from transferring ownership, we can borrow the content of a variable (case ’c’ in Fig. 2). In this case, two variables have the same content but none of them can be modified, thus protecting us from race conditions. Alternatively, we can borrow in a more traditional way (like when we borrow a book from a friend, he is expecting to get it back when we stop using it) like in case ’d’ in Fig. 2, where the function borrows the content of a variable, operates with it (in this case, it could modify its content) and returns it to the original owner (not destroying it as shown in case ’b’). Exceptionally, all these rules can be violated if we make use of unsafe blocks, which is strongly discouraged but necessary in certain situation (e.g., dealing with external libraries written in Fortran, C or C++). If present, unsafe blocks allow us to clearly identify parts of the code which should be carefully audited, keeping it isolated and not making the whole program unsafe by default like in Fortran, C or C++. 3. Assessment We explored the advantages and drawbacks of Rust for astrophysics by re-implementing the fundamental parts of Mercury-T (Bolmont et al., 2015), a Fortran code that simulates the dynamical and tidal evolution of multi-planet systems. † Except primitive types or types that implement the Copy trait. What can Rust do for astrophysics? Transfer ownership Transfer Compiler All errors are caught at compilation time. v2 v2 error: v1 doesn't own anything Lifetime starts let v1 = vec![1, 2, 3]; calc(v1); println!("{}", v1[0]); ownership b) v1 error: v1 doesn't own anything d) Transfer v v ownership v1 Borrowing c) Borrows Without mutability v1 let mut v1 = vec![1, 2, 3]; let v2 = &v1; v2 v1 v1 and v2 access the same memory println!("{} {}", v1[0], v2[0]); v2 v1 v1[0] = 10; error: v1/v2 can be read but none can be modified fn calc(v: Vec<i32>) {...} Lifetime ends (i.e., vector is freed) Lifetime starts let mut v1 = vec![1, 2, 3]; calc(&mut v1); println!("{}", v1[0]); ownership ownership a) let v1 = vec![1, 2, 3]; let v2 = v1; println!("{}", v1[0]); 3 v1 v1 v1 Borrows With mutability v Gives back Borrowing rules Two kinds of borrows, never both: 1. One or more non-mutable reference 2. Just one mutable reference fn calc(v: &mut Vec<i32>) { v[0] = 10; } Lifetime ends (i.e., vector is freed) Figure 2. Rust’s ownership and borrowing concepts help us overcome the problems of traditional system programming languages. We developed a simple N-Body dynamical simulator‡ (without tidal effects) based on a leapfrog integrator in Rust, Fortran, C and Go (which provide a garbage collector for memory management). The software design and implementation does not include any language-specific optimization that a developer with basic knowledge would not do. We compiled the four implementations with an optimization level 3 (rustc/gfortran/gcc compiler) and the standard compilation arguments in the case of Go. We selected the best execution time out of five for an integration of 1 million years and the results are in Table 1. For this particular problem, Rust is as efficient as Fortran, and both surpass C and Go implementations Table 1. Best execution times of pure N-body simulations, for an integration time of 1 million years using a leap-frog integrator. Fortran C Go Rust 0m13.660s 0m14.640s 2m32.910sa 4m26.240s a Time might be improved if language-specific optimizations were implemented. Based on Mercury-T, we implemented the additional acceleration produced by tidal forces between the star and its planets into our Rust and Fortran leapfrog integrators. To test the codes, we ran a simulation of 100 million years with the same initial conditions as the case 3 described in the Mercury-T article (Bolmont et al., 2015), hence a single planet with a rotation period of 24 hours, orbiting a brown dwarf (0.08 M ) at 0.018 AU with an eccentricity of 0.1. The results are shown in Fig. 3, the Rust and Fortran code are practically identical and they reproduce a similar behavior to what is shown in the Mercury-T article. Nevertheless, leapfrog is a very simple integrator and not very accurate. This can be seen in the eccentricity evolution, which is slightly different from the Mercury-T article and appears noisy. As an additional exercise, we implemented the WHFast integrator (Rein and Tamayo, 2015) in Rust (black line in Fig. 3). This better integrator leads to a better agreement with Mercury-T thus demonstrating that a high level of accuracy can also be achieved with Rust. ‡ http://www.blancocuaresma.com/s/ Sergi Blanco-Cuaresma & Emeline Bolmont 0.028 0.026 0.024 0.022 0.020 0.018 0.016 Eccentricity Planet rotation period (hours) Semi-major axis (AU) 4 Leapfrog Fortran Leapfrog Rust WHFast Rust 140 120 100 80 60 10 -1 10 -2 10 -3 2 10 10 3 10 4 10 5 10 6 Time (years) 10 7 10 8 Figure 3. Tidal evolution of a single planet orbiting a brown-dwarf with the three integrators. Top: evolution of the semi-major axis of the planet. Middle: evolution of the planet’s rotation period. Bottom: evolution of the eccentricity of the planet. 4. Conclusions We have shown the reliability of Rust as a programming language as opposed to Fortran, C or even Go. Rust allows the user to avoid common mistakes such as the access to invalid memory regions and race conditions. We have also shown that it is a competitive language in terms of speed and accuracy. The main challenge we experienced was the initial learning curve, it was necessary to really understand and get used to the ownership and borrowing concepts. Once the paradigm shift is done, the benefits are immediate. We therefore encourage the community to consider Rust as a language that will help us produce good quality, memory safe, concurrent and high-performance scientific code. References Anderson, B., Herman, D., Matthews, J., McAllister, K., Goregaokar, M., Moffitt, J. and Sapin, S. 2015, arXiv, 1505.07383 Bolmont, E. and Raymond, S. N. and Leconte, J. and Hersant, F. and Correia, A. C. M. 2015, A&A, 583, A116 Köster, J. 2015, Bioinformatics, 32, 444 Poss, R. 2014, arXiv, 1407.5670 Rein, H. and Tamayo, D. 2015, MNRAS, 452, 376-388
6cs.PL
1 Flatness-based control of a two-degree-of-freedom platform with pneumatic artificial muscles arXiv:1703.05503v1 [cs.SY] 16 Mar 2017 David Bou Saba, Paolo Massioni, Eric Bideaux, and Xavier Brun Abstract—Pneumatic artificial muscles are a quite interesting type of actuators which have a very high power-to-weight and power-to-volume ratio. However, their efficient use requires very accurate control methods which can take into account their complex dynamic, which is highly nonlinear. This paper consider a model of two-degree-of-freedom platform whose attitude is determined by three pneumatic muscles controlled by servovalves, which mimics a simplified version of a Stewart platform. For this testbed, a model-based control approach is proposed, based on accurate first principle modeling of the muscles and the platform and on a static model for the servovalve. The employed control method is the so-called flatness-based control introduced by Fliess. The paper first recalls the basics of this control technique and then it shows how it can be applied to the proposed experimental platform; being flatness-based control an open-loop kind of control, a proportional-integral controller is added on top of it in order to add robustness with respect to modelling errors and external perturbations. At the end of the paper, the effectiveness of the proposed approach is shown by means of experimental results. A clear improvement of the tracking performance is visible compared to a simple proportional-integral controller. Index Terms—Pneumatic artificial muscles, nonlinear control, flatness. I. I NTRODUCTION Pneumatic artificial muscles (PAMs) are a quite efficient type of actuators which feature high power-to-volume ratio, high pulling efforts at a relatively low price [7]. This makes their use quite interesting in many engineering and robotic applications, even if their control is problematic due the nonlinearity in their dynamic model as well as from the hysteresis phenomena which they feature. Pneumatic artificial muscles produce a contraction effort when inflated, which is a nonlinear function of both the internal pressure and the relative contraction of its length. Many theoretical models of PAMs can be found in the literature [7], [6], [15], and this paper will refer to the results of experimental tests [3] that average out the hysteresis phenomena and therefore can model the behaviour very accurately. The subject of this paper is a study of a two-degree-of-freedom platform, actuated by three pneumatic muscles. The objective is the synthesis of a model-based control law allowing the tracking of a reference trajectory for a wide operating range of the muscles. The platform is constrained to a limited operating domain due to mechanical constraints and to the fact that the muscles generate only pulling efforts. Furthermore, the system The authors are with Laboratoire Ampère, UMR CNRS 5005, INSA de Lyon, Université de Lyon, 69621 Villeurbanne CEDEX, France. {david.bou-saba, paolo.massioni, eric.bideaux, xavier.brun}@insa-lyon.fr. Corresponding author: Paolo Massioni, tel: +33(0)472436035, fax: +33(0)472438530. can be considered as overactuated (three actuators moving two degrees of freedom), which requires a control allocation strategy. The control of PAMs has been approached with several methods, which try to cope with the strong nonlinearities of its dynamics. The approaches found in the literature are mainly inherently nonlinear control methods [2], [16]; sliding mode controllers are one of the most common choices [1], [5], [13], also sometimes combined with adaptive or neural controllers [14], [12], or backstepping [11]. Sliding mode controllers in fact provide enough robustness with respect to the dynamical model which is considered as uncertain. In this work, a flatness-based control [8] is proposed, which exploits a model of all the elements involved and which also solves the over-actuation problem at the same time. The robustness with respect to model errors is provided by coupling the flatness-based controller with a proportional-integral (PI) controller feeding back the error with respect to the reference trajectory. The paper is structured as follows. Section II introduces the notation used throughout the paper. Section III describes the model of the platform and of all its elements, including the pneumatic artificial muscles. Section IV shows that a proper choice of measurements makes the platform a flat system, for which a flatness-based law is proposed. Section V concerns the problem of overactuation and how it is solved. At last, Section VI proposes some experimental results whereas Section VII draws the conclusions of the article. II. N OTATION AND DEFINITIONS Let R be the set of real number, and N the set of the strictly positive integers. For a matrix A, A⊤ denotes the transpose. Given two functions f (x), g(x) ∈ Rn , with x ∈ Rn , let the Lie (x) · g(x). derivative of f along g be defined as Lg f (x) = ∂f∂x ∂Lξ−1 f (x) g · g(x), with L0g f (x) = For ξ ∈ N, let Lξg f (x) = ∂x f (x). For all signals x depending from the time t, let x(ξ) (2) indicate its ξ-th time derivative, i.e. x(1) = dx = dt = ẋ, x 2 d x dt2 = ẍ, etc. All the symbols concerning the pneumatic muscle platform are defined in Table I. III. T HE PNEUMATIC PLATFORM A. Description The pneumatic platform studied in this paper is represented in Fig. 1 and Fig. 2. It consists of a metal plate fixed to a spherical hinge on top of a vertical beam; three pneumatic muscles controlled by servovalves are attached to the plate at 2 P0 θ0 D0 l0 α K εa εb k r T R J φ1 = −90◦ φ2 = 30◦ φ3 = 150◦ θx θy Pi vi Vi qi εi ε0 Fi Γ Atmospheric pressure Weave angle of the muscle at rest Diameter of the muscle at rest Length of the muscles at rest Experimentally determined power coefficient Experimentally determined coefficient Experimentally determined coefficient Experimentally determined coefficient Polytropic index of air Perfect gas constant Air temperature Muscle application point distance from center (constant) Momentum of inertia about an horizontal axis (constant) Angular position of the 1st muscle (constant) Angular position of the 2nd muscle (constant) Angular position of the 3rd muscle (constant) Angular position of the platform around x axis Angular position of the platform around y axis Absolute pressure inside the i-th muscle Voltage applied to the i-th servovalve Volume of the i-th muscle Mass flow into the i-th muscle Contraction of the i-th muscle Initial contraction of the muscle Force applied by the i-th the muscle Perturbation torques TABLE I S YMBOL DEFINITIONS . equally spaced points. Due to the muscles, and for simplicity, it can be considered that the platform has only two degrees of freedom, i.e. the two rotational angles (θx and θy ) with respect to horizontal axes passing through the hinge. An inclinometer provides measurements of such angles, and pressure sensors are located inside each muscle. This platform can be considered as a simplified version of a Stewart platform, a test bench on which control laws can be tried and evaluated before moving to more complex systems with more degrees of freedom. Inclinometer Platform Pneumatic muscle Servovalve Pressure sensor Fig. 1. The experimental platform. θy y y M3 M3 M2 M2 φ3 θx φ2 x x F3 F2 M1 z φ1 M1 F1 Fig. 2. Axonometric view and view from the top of the top plate, with definition of the axes x, y, z and the rotation angles θx and θy . M1 , M2 and M3 are the attachment points of the three pneumatic artificial muscles. B. Model This section reports the differential equations describing the system dynamic and the different assumptions made. A more detailed description of the complete model of the system is presented in [4], with all the assumptions and explanations (including those concerning how the hysteresis has been taken into account). The first elements to be modeled are the pneumatic muscles, which are supposed to be identical (they have the same length at rest l0 , the same initial contraction ε0 , etc.). The length contraction of each muscle (i = 1, 2, 3) can be written as: εi = R (cos φi sin θy − sin φi sin θx cos θy ) + ε0 l0 (3) Subsequently, the rate of contraction of each muscle is the time derivative of εi , i.e. Rh −θ̇x sin φi cos θx cos θy ε̇i = l0 (4) i +θ̇y (cos φi cos θy + sin φi sin θx sin θy )     F1 1 θ̈x = M (θx , θy ) F2  + Γ θ̈y J F3 (5) where the matrix M (θx , θy ) is given in equation (1) at the top of the next page. The term Γ = [Γx , Γy ]⊤ contains the torques that will not be modelled (as an arbitrary choice) and will be left to the feedback control to take care of. Such terms are either due to friction, or to gyroscopic couplings between the two axes, or to external forces acting on the platform. The friction terms are quite difficult to model exactly, whereas the gyroscopic couplings are quite small due to the fact that the platform keeps always almost horizontal and moves at relatively low angular velocities. This allows writing the platform around each axis as decoupled, according to (5) above. Such an equation can also be written as     1 F θ̈x (6) = E (θx , θy ) 1 + G (θx , θy ) F3 + Γ F2 θ̈y J where the matrices E (θx , θy ) and G (θx , θy ) are given in equation (2) at the top of the next page. This form separates the effect of the first two forces with respect to F3 , which makes it easier to approach the overactuation problem. 3 M (θx , θy ) =  R − sin φ1 cos θx cos θy J cos φ1 cos θy + sin φ1 sin θx sin θy − sin φ2 cos θx cos θy cos φ2 cos θy + sin φ2 sin θx sin θy − sin φ3 cos θx cos θy cos φ3 cos θy + sin φ3 sin θx sin θy     R − sin φ1 cos θx cos θy − sin φ2 cos θx cos θy  E (θx , θy ) =    J cos φ1 cos θy + sin φ1 sin θx sin θy cos φ2 cos θy + sin φ2 sin θx sin θy     − sin φ3 cos θx cos θy   G (θx , θy ) = R J cos φ3 cos θy + sin φ3 sin θx sin θy where εi (εi − εa ) εi + εb (8)   α πD02 3 (1 − εi ) 1 − 4 tan2 θ0 sin2 θ0 (9) L(εi ) = K and H(εi ) = with α, K, εa and εb experimentally determined constants. Considering that the operating range of the servovalves is for 1.25 bar 6 Pi 6 7 bar, the possible forces for each muscle are represented in Figure 3. Notice that only traction forces are possible (the muscles cannot push). (2) At last, the mass flow of gas qi entering each muscle is a nonlinear function of the pressure inside the muscle and the voltage vi fed to the servovalve. This function is considered as static, and it can be described by means of a polynomial approximation of experimental data [10] (graphically depicted in Figure 4). Mass flow (q i ) [Nl/min] In turn, each force due to pneumatic muscles can be modeled with the so called quasi-static model [7], [4], [3] as Fi (Pi , εi ) = H(εi )(Pi − P0 ) + L(εi ), (7) 150 200 100 100 50 0 -100 0 -200 5 -50 -100 1000 0 P = 1.25 bar Voltage (v i ) [V] i 900 P = 2 bar 800 P i = 3 bar i Traction force (F i ) [N] (1) -5 P i = 4 bar 1 3 4 5 6 7 -150 Pressure (P i ) [bar] P = 5 bar 700 2 i P i = 6 bar P i = 7 bar 600 Fig. 4. Mass flow of a servovalve as a function of voltage vi and absolute muscle pressure Pi . 500 400 Considering that at each time instant, Pi is measured by pressure sensors, it is possible to find the vi which gives the desired qi by a simple inversion of the polynomial function qi (Pi , vi ). 300 200 100 0 0 0.05 0.1 0.15 0.2 0.25 Contraction (ǫ i ) C. Control objectives Fig. 3. Traction force applied by a muscle as a function of the contraction εi and absolute pressure Pi . The pressure inside each muscle is modeled as   krT Pi ∂V (εi ) Ṗi = qi (Pi , vi ) − ε̇i Vi (εi ) rT ∂εi (10) where k is the polytropic index of the gas, r the perfect gas constant, T the temperature (considered constant), qi the mass flow of gas, and Vi the volume of the muscle, for which the following formula has been proposed [4], [3]  α 1 ∂V (1 − εi ) π (11) (εi ) = D02 l0 − 2 + (α + 1) ∂εi 4 tan2 θ0 sin θ0 where D0 , l0 are the diameter and length of the muscle at rest, and θ0 is the weave angle of the muscle fibers (a constant). The aim of this testbed is to demonstrate the ability to track any smooth trajectory of θx and θy . Trajectories of this kind can be chosen as infinitely differentiable piece-wise polynomial function. The angles of the platforms (θx , θy ) are physically constrained to be in the range [−15◦ , 15◦ ]. For these values, for the contractions εi the ranges are constrained within [−0.03, .21] (the muscles need to be contracted in order to apply a force), for which H(εi ) is never equal to 0. IV. M ODEL ANALYSIS AND CONTROL The accurate knowledge of the model allows the application of flatness-based control, at the condition of being able to prove that the system is flat. The relevant notions are recalled here. 4 A. Flatness and flatness-based control The notion of flat system and flatness-based control for nonlinear systems have been introduced in [8]. Basically, the “flatness” is a property of a dynamical system and a choice of its output y, as defined here. Definition 1 (Flat system - adapted from [8]): A dynamical system of equations ẋ = f (x) + g(x)u, with x ∈ Rn , u =∈ Rm is flat if there exist an Rm -valued map h, an Rn -valued map η, and an Rm -valued map θ such that y = h(x, u, u(1) , . . . , u(ν) ) (12) x = η(y, y (1) , . . . , y (ν−1) ) (13) u = θ(y, y (1) , . . . , y (ν) ) (14) for an appropriate value of ν ∈ N. The output y is then called “flat output”. The idea of flatness can be explained briefly as follows. If one can choose as many output variables yi as inputs (the system is square), such that it is possible to recover the state and the inputs from the derivatives of these output variables, then the system is flat and a flatness-based, open loop control law can be derived by system inversion (as explained later on). Two fundamental concepts for system inversion are characteristic index and coupling matrix. Definition 2 (Characteristic index): The characteristic index of the i-th component yi of y is the smallest ρi ∈ N for which Lgj Lρfi −1 hi 6= 0 for at least one value of j. Definition 3 (Coupling matrix): The coupling matrix ∆(x) is given by the expression:   Lg2 Lfρ1 −1 h1 . . . Lgm Lfρ1 −1 h1 Lg1 Lfρ1 −1 h1   ρ −1 Lg2 Lfρ2 −1 h2 . . . Lgm Lfρ2 −1 h2   Lg1 Lf2 h2 .  ∆(x)= .. .. .. ..  .   . . . Lg1 Lρfm −1 hm Lg2 Lρfm −1 hm ... Lgm Lρfm −1 hm It can be shown that  (ρ )   ρ1 L f h1 y1 1  (ρ2 )   Lρf2 h2  y2     = ∆(x)u +  .. ..    .   . ρ (ρ ) L f m hm ym m (15)    .  (16) The control law which has been applied to the testbed is based on the following theorem, which is a well-known result for which no proof is necessary here. Theorem 4 (Adapted from [8]): If a system of equations ẋ = f (x) + g(x)u with u ∈ Rm is flat (Definition 1) with respect to a flat output y = h(x) ∈ Rm with characteristic coefficients ρi , and if the matrix ∆(x) is invertible (at least locally), then it is possible to track a given smooth reference trajectory y(t) = h(x(t)) by employing the control law  (ρ )   ρ  L f 1 h1 y1 1  (ρ2 )   Lρ2 h    f 2   y2 −1  − . (17) u = ∆(x)   .. ..    .    . (ρ ) ym m Lρfm hm It is possible to prove that with this state trajectory, the system’s dynamic of each yi is linear (simply a chain of ρi integrators). B. Complete state-space model The state of the platform model can be chosen as x = [x1 , x2 , x3 , . . . x7 ]⊤ = [θx , θy , θ̇x , θ̇y , P1 , P2 , P3 ]⊤ , whereas the input vector is u = [q1 , q2 , q3 ]⊤ . By neglecting the perturbation term Γ, the system dynamic can then be expressed as follows. ẋ = f (x) + g(x)u (18) where f (x) =   x3       x4       − cos x1 cos x2 sin φ1 (H (ε1 ) (x5 − P0 ) + L (ε1 ))     − cos x cos x sin φ (H (ε ) (x − P ) + L (ε )) 1 2 2 2 6 0 2     − cos x1 cos x2 sin φ3 (H (ε3 ) (x7 − P0 ) + L (ε3 ))            (cos φ1 cos x2 + sin φ1 sin x1 sin x2 ) (H (ε1 ) (x5 − P0 ) + L (ε1 )),   + (cos φ2 cos x2 + sin φ2 sin x1 sin x2 ) (H (ε2 ) (x6 − P0 ) + L (ε2 ))    + (cos φ3 cos x2 + sin φ3 sin x1 sin x2 ) (H (ε3 ) (x7 − P0 ) + L (ε3 ))              a(ε1 , ε̇1 )(x5 − P0 )       a(ε2 , ε̇2 )(x6 − P0 )     a(ε3 , ε̇3 )(x7 − P0 ) (19) g(x) = [g1 (x), g2 (x), g3 (x)] with     0 0  0   0       0   0         g1 (x) =   0  , g2 (x) =  0  , b(ε1 )  0       0  b(ε2 ) 0 0 with     a (εi , ε̇i ) = −  0  0     0     g3 (x) =   0   0     0  b(ε3 ) (20) k ∂V (εi ) ε̇i V (εi ) ∂εi krT    b(εi ) = V (εi )  (21) C. Flatness of the model The system is flat if a vector flat output [y1 y2 y3 ]⊤ , according to (12), can be found. Such flat output has to fulfill both (13), i.e., it should be possible to express the state vector as a function of its time derivatives, and (14), i.e. it should be possible to express the input as a function of its derivatives. This paragraph shows that the choice   y1 = x1 y2 = x2 (22)  y3 = F3 = H(ε3 )(x7 − P0 ) + L(ε3 ) 5 plotted in Figure 5 never reaches zero in this interval. So ρ2 = 3. actually works in making the system flat. First consider condition (13); x1 , x2 , x3 and x4 can be obtained directly from y1 , y2 and their first degree time derivatives. Once x1 , x2 , x3 and x4 are known, all εi , H(εi ) and L(εi ) are determined as well. Since F3 is an output and H(ε3 ) 6= 0, x7 is immediately also determined. At last, x5 and x6 can be determined from ẋ3 and ẋ4 if the matrix  − cos x1 cos x2 sin φ1 − cos x1 cos x2 sin φ2 cos x2 cos φ1 +sin x1 sin x2 sin φ1 cos x2 cos φ2 +sin x1 sin x2 sin φ2  is invertible. The determinant of this matrix is cos2 x2 cos x1 (sin φ2 cos φ1 − sin φ1 cos φ2 ) which is never 0 in the range of θx = x1 , θy = x2 allowed for the platform (i.e. they never reach ±90◦ ). Secondarily, consider condition (13); a necessary condition for this is that the sum of the characteristic indices of the three outputs is the same as the number of states, i.e. 7. The computation of such indices leads to the following results. Fig. 5. Value of z as function of x1 = θx and x2 = θy . • • Output y1 Lg1 y3 = Lg2 y3 = 0; Lg3 y3 = b(ε3 )H(ε3 ) 6= 0 ⇒ ρ3 = 1. Lg1 y1 = Lg2 y1 = Lg3 y1 = 0 ⇒ ρ1 > 1; Lf y1 = x3 ; Lg1 Lf y1 = Lg2 Lf y1 = Lg3 Lf y1 = 0 ⇒ ρ1 > 2; L2f y1 = ẋ3 = − cos x1 cos x2 (sin φ1 (H(ε1 )(x5 − P0 )+ L(ε1 )) + sin φ2 (H(ε2 )(x6 − P0 ) + L(ε2 ))+ sin φ3 (H(ε3 )(x7 − P0 )+L(ε3 ))); Lg1 L2f y1 = − sin φ1 cos x1 cos x2 H(ε1 )b(ε1 ) Lg2 L2f y1 = − sin φ2 cos x1 cos x2 H(ε2 )b(ε2 ) Lg3 L2f y1 = − sin φ3 cos x1 cos x2 H(ε3 )b(ε3 ) It can be pointed out that Lg1 L2f y1 , Lg2 L2f y1 and Lg3 L2f y1 are never equal to 0 for the x1 and x2 within the valid range, so ρ1 = 3. • Output y3 Output y2 Lg1 y2 = Lg2 y2 = Lg3 y2 = 0 ⇒ ρ2 > 1; Lf y2 = x4 ; Lg1 Lf y2 = Lg2 Lf y2 = Lg3 Lf y2 = 0 ⇒ ρ2 > 2; L2f y2 = ẋ4 = (cos φ1 cos x2 + sin φ1 sin x1 sin x2 )(H(ε1 )(x5 − P0 ) + L(ε1 ))+(cos φ2 cos x2 +sin φ2 sin x1 sin x2 )(H(ε2 )(x6 − P0 ) + L(ε2 )) + (cos φ3 cos x2 + sin φ3 sin x1 sin x2 )(H(ε3 )(x7 − P0 ) + L(ε3 )); Lg1 L2f y2 = (cos φ1 cos x2 + sin φ1 sin x1 sin x2 )H(ε1 )b(ε1 ) Lg2 L2f y2 = (cos φ2 cos x2 + sin φ2 sin x1 sin x2 )H(ε2 )b(ε2 ) Lg3 L2f y2 = (cos φ3 cos x2 + sin φ3 sin x1 sin x2 )H(ε3 )b(ε3 ) The necessary condition of ρ1 + ρ2 + ρ3 = 7 is satisfied. The last step is to verify that the decoupling matrix  Lg1 L2f y1  ∆ = Lg1 L2f y2 Lg1 y3 Lg2 L2f y1 Lg2 L2f y2 Lg2 y3 (23) (24) is invertible. The expression of ∆ is made explicit in (25) at the top of the next page (with the shorthand notation of Hi = H(εi ), Li = L(εi ), bi = b(εi )). The determinant of this matrix is |∆| = H1 H2 H3 b1 b2 b3 m with m = − sin φ1 cos x1 cos x2 (cos φ2 cos x2 + sin φ2 sin x1 sin x2 ) + sin φ2 cos x1 cos x2 (cos φ1 cos x2 + sin φ1 sin x1 sin x2 ). The values of m as function of θx and θy in the valid interval are depicted in Figure 6 Accordingly, |∆| 6= 0, so the decoupling matrix is invertible over the operating range and the chosen output is proven to be flat. Notice that Lg2 L2f y2 can never be zero in the valid range, as the function z = cos φ2 cos x2 + sin φ2 sin x1 sin x2  Lg3 L2f y1 Lg3 L2f y2  Lg3 y3 Fig. 6. Values of m as function of θx and θy . 6  − sin φ1 cos x1 cos x2 H1 b1 ∆ = (cos φ1 cos x2 + sin φ1 sin x1 sin x2 ) H1 b1 0 − sin φ2 cos x1 cos x2 H2 b2 (cos φ2 cos x2 + sin φ2 sin x1 sin x2 ) H2 b2 0  − sin φ3 cos x1 cos x2 H3 b3 (cos φ3 cos x2 + sin φ3 sin x1 sin x2 ) H3 b3  (25) H3 b3 ... ... θx , θ˙ x , ¨θx , θ x , θy , θ˙ y , ¨θ y , θ y , F 3 , F˙ 3 qi servovalves platform inclinometer pressure sensors θx,y flatness controller Pi + muscle model vi F3 − error inverse polynomial map experimental setup PI + + desired qi controller Fig. 7. Global control scheme. D. Control law The platform can then be controlled in open-loop with the law  (ρ )    y1 1 L3f y 1    q = ∆−1 (x)  y 2(ρ2 )  − L3f y 2  + ∆−1 (x)w (26) (ρ ) Lf y 3 y3 3 where y i is the desired trajectory and w is an additional control term. Due to the presence of the perturbation terms which have been neglected (Γ), if w = 0 there will necessarily be a non-zero error ǫi = yi − yi . Considering that the dynamic of the system under this law is just a chain of integrators, by imposing to w a feedback law as a function of the ǫi closes the loop. In this case, a proportional-integral controller (PI) has been tuned. Figure 7 shows the overall control scheme. The flatness-based control is for some aspects, quite similar to feedback linearisation control [9]. The main differences lie in the fact that flatness-based requires a specific choice of flat output (whereas feedback linearisation can take any output, assuming the system is observable), and that it does not require a knowledge or measure of the state variable. On the other hand, the baseline flatness-based control is feedforward only, which requires the introduction of the additional feedback term w. V. S OLVING THE OVERACTUATION It can be pointed out that the platform is overactuated, in the sense that the three forces applied by the muscles are generating only two torques. To tell it in another way, the average value of the Fi is irrelevant for the platform’s dynamic; if a given F1 = F̃1 , F2 = F̃2 and F3 = F̃3 generate certain torques, then F1 = F̃1 + F0 , F2 = F̃2 + F0 and F3 = F̃3 + F0 will generate the same torques for any F0 . On the other hand, it is useful to have three muscles instead of two due to the fact that muscles can only pull and not push, i.e. their force range is quite limited as shown by Figure 3. The choice of F3 as one of the flat output can be then interpreted in the light of this fact: first of all, it would be impossible to resolve the state from the output if one of the forces (or pressure) is not measured, as their effects on the angles is the same up to a constant term. Secondarily, the flat control allows choosing the value of F3 , which lets one choose the best value in order to let all the muscles be in their valid force range. Consider that at each instant, the εi are determined by the instantaneous geometry, which implies that each muscle has a limited interval of possible applicable forces (see Figure 8). One can find the intersection of such intervals and call Fmin its minimum and Fmax its maximum. Under the reasonable hypothesis that the platform turns slowly (in any case it is constrained to angles smaller than 15◦ ), it can be assumed that F1 , F2 and F3 have to be close to the equilibrium values, i.e. F1 ≈ F2 ≈ F3 . For this reason, setting the reference for F3 as 1 (27) F 3 = (Fmax (ε1 , ε2 , ε3 ) + Fmin (ε1 , ε2 , ε3 )) 2 gives the best chances of having F2 and F3 within the realisable interval as well. For the experiment in the next section, the reference for F3 has been determined by the law in (27). VI. E XPERIMENTAL RESULTS In order to assess the performance of the proposed control approach, an experiment has been conducted on the platform. The proposed flatness-based control coupled with PI has been compared to a PI controller, empirically tuned to get the best apparent performances. The same reference trajectory (a combination of sinusoids) has been tested for both controllers. Figure 9 reports the results for the PI controller, whereas Figure 10 shows the flatness-based controller results. It is 7 600 6 P i = 1.25 bar P i = 7 bar 4 450 2 300 degrees Traction force (F i ) [N] ǫ2 ǫ3 150 Fmax 0 -2 ǫ1 -4 θ (measure) Fmin x reference for θ 0 0 0.05 0.1 0.15 0.2 x θ (measure) -6 0.25 y reference for θ y Contraction (ǫ i ) -8 Fig. 8. For a given position, the three contractions εi are given, so not all forces are possible for each muscle, but only those within the pressure range between 1.25 and 7 bar. One can compute Fmax (the maximum force that all muscles can exert) and Fmax (the minimum force that all muscles can exert). Setting F3 as the average between these two allows all muscles to apply the desired forces and to maximise their range. 0 5 10 15 20 25 30 time [s] Fig. 10. Trajectory tracking with flatness-based control plus a PI. 6.5 6 5.5 bar apparent from the picture that the feedforward action added by the flatness-based control greatly improves the tracking ability; in fact, it can be computed that the root mean square tracking errors (for θx and θy respectively) are 0.51 and 0.59 degrees for the PI case. With the flatness controller, these root mean square errors become less than half, i.e. 0.25 and 0.29 degrees respectively (consider also that the inclinometers’ output has a quantisation equivalent to 0.18 degrees). 5 4.5 4 P 6 P P 1 2 3 3.5 0 4 10 15 20 25 30 time [s] Fig. 11. Pressures inside the pneumatic artificial muscles during the flatnessbased control plus PI experiment. 2 degrees 5 0 VII. C ONCLUSION -2 θ (measure) x -4 reference for θ x θ (measure) y reference for θ y -6 0 5 10 15 20 25 30 time [s] Fig. 9. Trajectory tracking with simple PI control. Figure 11 shows the evolution of the pressures during the flatness-based controller test. Figure 12 shows the force of the three muscles during the same test; remember that the reference for F3 is determined with the overactuation-solving law proposed in Section V. Notice that the forces never saturate (and neither do the voltages or the pressure), which validates the proposed strategy. This paper has presented the successful application of a flatness-based controller to a platform featuring three PAMs. The experimental results clearly show a better trajectory tracking compared to a simple PI controller. Future research will look at the possibility of using PAMs for building a complete six-degree-of-freedom Stewart platform, and controlling it with the same approach. R EFERENCES [1] H. Aschemann and D. Schindele. Sliding-mode control of a high-speed linear axis driven by pneumatic muscle actuators. IEEE Transactions on Industrial Electronics, 55(11):3855–3864, 2008. [2] D.X. Ba, T.Q. Dinh, and K.K. Ahn. An integrated intelligent nonlinear control method for a pneumatic artificial muscle. IEEE/ASME Transactions on Mechatronics, 21(4):1835–1845, 2016. [3] E. Bideaux, S. Sermeno Mena, and S. Sesmat. Parallel manipulator driven by pneumatic muscles. In 8th International Conference on Fluid Power (8th IFK), Dresden, Germany, 2012. 8 280 F1 260 F2 F3 240 220 N 200 180 160 140 120 100 80 0 5 10 15 20 25 30 time [s] Fig. 12. Forces applied by the pneumatic artificial muscles (estimated from the pressures and the contractions) with flatness-based control plus a PI. [4] D. Bou Saba, E. Bideaux, X. Brun, and P. Massioni. A complete model of a two degree of freedom platform actuated by three pneumatic muscles elaborated for control synthesis. In BATH/ASME 2016 Symposium on Fluid Power and Motion Control, pages V001T01A004– V001T01A004. American Society of Mechanical Engineers, 2016. [5] D. Cai and Y. Dai. A sliding mode controller for manipulator driven by artificial muscle actuator. In Proceedings of the 2000 IEEE International Conference on Control Applications, pages 668–673. IEEE, 2000. [6] C.-P. Chou and B. Hannaford. Measurement and modeling of McKibben pneumatic artificial muscles. IEEE Transactions on Robotics and Automation, 12(1):90–102, 1996. [7] F. Daerden and D. Lefeber. Pneumatic artificial muscles: actuators for robotics and automation. European journal of mechanical and environmental engineering, 47(1):11–21, 2002. [8] M. Fliess, J. Lévine, P. Martin, and P. Rouchon. Flatness and defect of non-linear systems: introductory theory and examples. International Journal of Control, 61(6):1327–1361, 1995. [9] A. Isidori. Nonlinear control systems. Springer Science & Business Media, 2013. [10] O. Olaby, X. Brun, S. Sesmat, T. Redarce, and E. Bideaux. Characterization and modeling of a proportional value for control synthesis. In Proceedings of the JFPS International Symposium on Fluid Power, volume 2005, pages 771–776. The Japan Fluid Power System Society, 2005. [11] R.A. Rahman and N. Sepehri. Design and experimental evaluation of a dynamical adaptive backstepping-sliding mode control scheme for positioning of an antagonistically paired pneumatic artificial muscles driven actuating system. International Journal of Control, pages 1–26, 2016. [12] R.M. Robinson, C.S. Kothera, R.M. Sanner, and N.M. Wereley. Nonlinear control of robotic manipulators driven by pneumatic artificial muscles. IEEE/ASME Transactions on Mechatronics, 21(1):55–68, February 2016. [13] X. Shen. Nonlinear model-based control of pneumatic artificial muscle servo systems. Control Engineering Practice, 18(3):311–317, 2010. [14] G.L. Shi and W. Shen. Hybrid control of a parallel platform based on pneumatic artificial muscles combining sliding mode controller and adaptive fuzzy CMAC. Control Engineering Practice, 1(1):76–86, 2013. [15] B. Tondu and P. Lopez. Modeling and control of McKibben artificial muscle robot actuators. Control Systems, IEEE, 20(2):15–38, 2000. [16] X. Zhu, G. Tao, B. Yao, and J. Cao. Adaptive robust posture control of parallel manipulator driven by pneumatic muscles with redundancy. IEEE/ASME Transactions on Mechatronics, 13(4):441–450, August 2008.
3cs.SY
A Simple Semantics for Haskell Overloading J. Garrett Morris arXiv:1612.08199v1 [cs.PL] 24 Dec 2016 University of Edinburgh [email protected] Abstract 1.1 Dictionary-Passing and its Disadvantages As originally proposed, type classes provide overloading and adhoc definition, but can still be understood (and implemented) in terms of strictly parametric calculi. This is not true of subsequent extensions of type classes. Functional dependencies and equality constraints allow the satisfiability of predicates to refine typing; this means that the interpretations of equivalent qualified types may not be interconvertible. Overlapping instances and instance chains allow predicates to be satisfied without determining the implementations of their associated class methods, introducing truly nonparametric behavior. We propose a new approach to the semantics of type classes, interpreting polymorphic expressions by the behavior of each of their ground instances, but without requiring that those behaviors be parametrically determined. We argue that this approach both matches the intuitive meanings of qualified types and accurately models the behavior of programs. The semantics of type classes has primarily been given by translations from instance declarations and (implicit) overloading to dictionaries and (explicit) dictionary arguments. This parallels the treatment of implicit polymorphism by translation to systems with explicit polymorphism (such as System F), and shares similar challenges. For a simple example, in Haskell, the map function has the polymorphic type scheme (t → u) → [t] → [u]. In translating to System F, this could be interpreted as either Categories and Subject Descriptors D.3.1 [Programming Languages]: Formal Definitions and Theory—Semantics; F.3.2 [Logics and Meanings of Programs]: Semantics of Programming Languages—Denotational semantics Keywords overloading; type classes; semantics ∀t.∀u.(t → u) → [t] → [u] or ∀u.∀t.(t → u) → [t] → [u]. But these types are not equivalent: they express different orders of passing type arguments. There are various ways of addressing this discrepancy: for example, Mitchell [8] shows that, for any two translations of an implicitly typed scheme, there is a term (which he calls a retyping function) which transforms terms of one translation to terms of the other, while only manipulating type abstractions and applications. Similar issues arise in the semantics of type classes. For example, a function to compare pairs (t, u) for equality could be given either the type scheme (Eq t, Eq u) ⇒ (t, u) → (t, u) → Bool or the type scheme 1. Introduction Implicit polymorphism (as provided by the Hindley-Milner type systems in ML and Haskell) provides a balance between the safety guarantees provided by strong typing, and the convenience of generic programming. The Hindley-Milner type system is strong enough to guarantee that the evaluation of well-typed terms will not get stuck, while polymorphism and principal types allow programmers to reuse code and omit excessive type annotation. Type classes [16] play a similar role for overloading: they preserve strong typing (ruling out run-time failures from the use of overloaded symbols in undefined ways) without requiring that programmers explicitly disambiguate overloaded expressions. Since their introduction, type classes have seen numerous extensions, such as multiparameter type classes, functional dependencies [5], and overlapping instances [13]; a variety of practical uses, from simple overloading to capturing complex invariants and type-directed behavior; and, the adoption of similar approaches in other strongly-typed programming languages, including Isabelle and Coq. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. Haskell ’14, September 6, 2014, Gothenburg, Sweden. Copyright is held by the owner/author(s). Publication rights licensed to ACM. ACM 978-1-4503-3041-1/14/09. . . $15.00. http://dx.doi.org/10.1145/2633357.2633364 (Eq u, Eq t) ⇒ (t, u) → (t, u) → Bool. In a dictionary-passing translation, type classes are interpreted by tuples, called dictionaries, containing the type-specific implementations of each of the class methods. Class instances correspond to dictionary definitions, while predicates in types correspond to dictionary arguments. In the case of the Eq class, which has equality and inequality methods, we could define Eq dictionaries by EqDict t = (t → t → Bool, t → t → Bool). Even though the two types for pair equality above are equivalent in the implicitly overloaded setting, their dictionary-passing translations are not: the first corresponds to a function of type EqDict t → EqDict u → (t, u) → (t, u) → Bool, while the second corresponds to EqDict u → EqDict t → (t, u) → (t, u) → Bool, Again, approaches exist to address this discrepancy: for example, Jones shows [3] that there are conversion functions, similar to Mitchell’s retyping functions, to convert between different translations of the same overloaded term. Our own work began by exploring instance chains [10], a proposed extension to Haskell-like type class systems. In the course of this exploration, we discovered several difficulties with existing approaches to the semantics of overloading. Mismatch in expressivity. System F typing is significantly more expressive than the Hindley-Milner type systems it is used to model. In particular, even within the translation of an ML or Haskell type scheme, there are arbitrarily many expressions that do not correspond to any expressions of the source language. The problem is compounded when considering dictionary-passing translations of type classes. For example, there is no notion in Haskell of class instances depending on terms; on the other hand, there is no difficulty in defining a term of type Int → EqDict Int. Uses of such a term cannot be equivalent to any use of the methods of Eq. As a consequence, there are properties of source programs (for example, that any two instances of == at the same type are equal) that may not be provable of their dictionary-passing translation without reference to the specific mechanisms of translation. Predicates refine typing. Second, the notions of equivalence of System F and Haskell types diverge once the satisfiability of predicates can refine typing. For example, functional dependencies allow programmers to declare that some parameters of a class depend upon others; in the declaration class Elems c e | c → e where empty : : c insert : : e → c → c the dependency c → e captures the intuition that the type of a container’s elements are determined by the type of the container. Concretely, given two predicates Elems τ υ and Elems τ ′ υ ′ , if we know that τ = τ ′ , then we can conclude υ = υ ′ . This property is lost in the dictionary-passing translation. Dictionaries for Elems contain just their methods: ElemsDict c e = (c, e → c → c) As a consequence, there are types that are equivalent in Haskell, but are not interconvertible in the dictionary-passing interpretation. For example, the type (Elems c e, Elems c e′ ) ⇒ e → e′ → c is equivalent to the (simpler) type (Elems c e) ⇒ e → e → c as we must have that e = e′ for the qualifiers in the first type to be satisfiable. However, there is no corresponding bijection between terms of type ElemsDict c e → ElemsDict c e′ → e → e′ → c and terms of type ElemsDict c e → e → e → c. While we can construct a term of the second type given a term of the first, there is no parametric construction of a term of the first type from a term of the second. Non-parametric behavior. Finally, other extensions to class systems make it possible to define terms which have no translation to parametric calculi. For example, we could define a function invBool that negated booleans and was the identity on all other types. We begin by introducing a suitable class: class Univ t where invBool : : t → t There are several approaches to populating the class, using different extensions of the Haskell class system. Using overlapping instances [13], we could simply provide the two desired instances of the class, relying on the type checker to disambiguate them based on their specificity: instance Univ Bool where invBool = not instance Univ t where invBool = id Using instance chains, we would specify the ordering directly: instance Univ Bool where invBool = not else Univ t where invBool = id With either of these approaches, we might expect that the type of the class method invBool is (Univ t) ⇒ t → t. However, the predicate Univ τ is provable for arbitrary types τ . Thus, the above type is intuitively equivalent to the unqualified type t → t; however, there is no term of that type in a purely parametric calculus that has the behavior of method invBool. (In practice, this is avoided by requiring that invBool’s type still include the Univ predicate, even though it is satisfied in all possible instantiations; while this avoids the difficulties in representing invBool in a parametric calculus, it disconnects the meaning of qualified types from the satisfiability of their predicates.) 1.2 Specialization-based Semantics We propose an alternative approach to the semantics of type-class based implicit overloading. Rather than interpret polymorphic expressions by terms in a calculus with higher-order polymorphism, we will interpret them as type-indexed collections of (the interpretations of) monomorphic terms, one for each possible ground instantiation of their type. We call this a specialization-based approach, as it relates polymorphic terms to each of their (groundtyped) specializations. We believe this approach has a number of advantages. • First, our approach interprets predicates directly as restrictions of the instantiation of type variables, rather than through an intermediate translation. Consequently, properties of the source language type system—such as the type refinement induced by the Elems predicates—are immediately reflected in the semantics, without requiring the introduction of coercions. • Second, our approach naturally supports non-parametric exam- ples, such as class Univ, and avoids introducing artificial distinction between the semantics of expressions using parametric and ad-hoc polymorphism. • Third, because our approach does not need to encode overload- ing via dictionaries, it becomes possible to reason about class methods directly, rather than through reasoning about the collection of dictionaries defined in a program. Our approach builds on Ohori’s simple semantics for ML polymorphism [12], extended by Harrison to support polymorphic recursion [1]. In this paper, we introduce a simple overloaded language called H− (§2), and give typing and equality judgments in the presence of classes and class methods. We apply our specialization-based approach to give a denotational semantics of H− (§3), and show the soundness of typing and equality with respect to the denotational semantics (§4). We also develop two examples, to demonstrate the advantages of our approach. First, we consider a pair of definitions, one parametric and the other ad-hoc, defining operational equivalent terms. We show that the defined terms are related by our equality judgment (§2.3) and have the same denotations (§3.5). This demonstrates the flexibility of our approach, and the ability to reason about class methods directly (the second and third advantages listed above). Second, we extend H− with functional dependencies (§5), and establish the soundness of the (extended) typing and equality judgments, all without having to augment the models of terms. This demonstrates the extensibility of our approach, and the close connection between properties of source terms and properties of their denotations (the first advantage listed above). 2. The H− Language Figure 1 gives the types and terms of H− ; we write x to denote a (possibly empty) sequence of x’s, and if π is a predicate C τ , we will sometimes write class(π) for C. As in Jones’s theory of qualified types [2], the typical Hindley-Milner types are extended with x ∈ Var t ∈ TVar C Term variable Type variables Class names Term constants Type constants Instance names Types τ, υ Predicates Pred ∋ π Contexts P, Q Qualified types ρ Type schemes Scheme ∋ σ Expressions Expr ∋ M, N Class axioms Axiom sets Methods: Signatures Implementations Class contexts ::= ::= ::= ::= ::= ::= | Axiom ∋ α ::= A ⊂ k K d ∈ InstName (VAR ) t|K|τ →τ Cτ π τ |π⇒ρ ρ | ∀t.σ x | k | λx.M | M N µx.M | let x = M in N d : ∀t. P ⇒ π Axiom Si ∈ Var ⇀ Pred × Scheme Im ∈ InstName × Var ⇀ Expr Ψ ::= hA, Si, Imi (x : σ) ∈ Γ P | Γ ⊢A x : σ (→ E) (µ) qualified types ρ, capturing the use of predicates. We must also account for the definition of classes and their methods. One approach would be to expand the grammar of expressions to include class and instance declarations; such an approach is taken in Wadler and Blott’s original presentation [16]. However, this approach makes such definitions local, in contrast to the global nature of subsequent type class systems (such as that of Haskell), and introduces problems with principal typing (as Wadler and Blott indicate in their discussion). We take an alternative approach, introducing new top level constructs (axioms A, method signatures Si, and method implementations Im) to model class and instance declarations. We refer to tuples of top level information as class contexts Ψ, and will give versions of both our typing and semantic judgments parameterized by such class contexts. Note that this leaves implicit many syntactic restrictions that would be present in a full language, such as the requirement that each instance declaration provide a complete set of method implementations. 2.1 H− Typing We begin with the typing of H− expressions; our expression language differs from Jones’s only in the introduction of µ (providing recursion) . Typing judgments take the form P | Γ ⊢A M : σ, where P is a set of predicates restricting the type variables in Γ and σ, and A is the set of class axioms (the latter is the only significant difference between our type system and Jones’s). The typing rules for H− expressions are given in Figure 2. We write ftv(τ ) for the free type variables in τ , and extend ftv to predicates π, contexts P, and environments Γ in the expected fashion. Rules (⇒ I) and (⇒ E) describe the interaction between the predicate context P and qualified types ρ. Otherwise, the rules are minimally changed from the corresponding typing rules of most Hindley-Milner systems. We continue with the rules for predicate entailment in H− , given in Figure 3. The judgment d : P A π denotes that the axiom named d proves predicate π, given assumptions P and class axioms A. We use a dummy instance name, written -, in the case that the goal is one of the assumptions. We will omit the instance name if (as in the typing rules) the particular instance used is irrelevant. We write P A Q if there are d1 . . . dn such that di : P Qi , and A P to abbreviate ∅ A P. Our entailment relation differs from Jones’s entailment relation for type classes and from our prior systems [10] in two respects. First, our system is intentionally simplified (for example, we omit superclasses and instance chains). P | Γ ⊢A (λx.M) : τ → τ ′ (⇒ I) P | Γ ⊢A µx.M : τ P | Γ ⊢A M : σ P, π | Γ ⊢A M : ρ P | Γ ⊢A M : π ⇒ ρ P | Γ ⊢A M : π ⇒ ρ P A π P | Γ ⊢A M : ρ t 6∈ ftv(Γ, P) P | Γ ⊢A M : ∀t.σ (L ET ) P | Γ ⊢A N : τ P | Γ ⊢A (M N) : τ ′ P | Γ, x : τ ⊢A M : τ − Figure 1: Types and terms of H . P | Γ, x : τ ⊢A M : τ ′ P | Γ ⊢A M : τ → τ ′ (⇒ E) (∀ I) (→ I) P | Γ ⊢A M : σ (∀ E) P | Γ ⊢A M : ∀t.σ P | Γ ⊢A M : [τ /t]σ P | Γ, x : σ ⊢A N : τ P | Γ ⊢A (let x = M in N) : τ Figure 2: Expression typing rules of H− . (A SSUME ) (A XIOM ) π∈P -:P Aπ (d : ∀t.Q′ ⇒ π ′ ) ∈ A d:P S π′ = π A P A S Q′ π Figure 3: Predicate entailment rules of H− . Second, we do not attempt to capture all the information that would be necessary for an dictionary-passing translation; we will show that having just the first instance name is sufficient to determine the meanings of overloaded expressions. In the source code of a Haskell program, type class methods are specified in class and instance declarations, such as the following: class Eq t where (==) : : t → t → Bool instance Eq t ⇒ Eq [ t ] where xs == ys = . . . We partition the information in the class and instance declarations into class context tuples hA, Si, Imi. The logical content is captured by the axioms A; in this example, we would expect that there would be some instance name d such that (d : ∀t. Eq t ⇒ Eq [t]) ∈ A. Haskell’s concrete syntax does not name instances; for our purposes, we assume that suitable identifiers are generated automatically. The method signatures are captured in the mapping Si; we distinguish the class in which the method is defined (along with the corresponding type variables) from the remainder of the method’s type scheme. For this example, we would have Si(==) = hEq t, t → t → Booli. Note that we have not quantified over the variables appearing in the class predicate, nor included the class predicate in the type scheme t → t → Bool. Each predicate in the range of Si will be of the form C t for some class C and type variables t, as they arise from class definitions. The type scheme of a class member may quantify over variables or include predicates beyond those used in the class itself. For example, the Monad class has the following definition: (C TXT ) {π ≁ π ′ | (d : P ⇒ π), (d′ : P′ ⇒ π ′ ) ∈ A} {(P | Γ, xi : σxi ⊢A Im(y, d) : σy,d ) | hy, di ∈ dom(Im)} P | Γ, xi : σxi ⊢A M : σ {β} P | Γ ⊢hA,Si,Imi M : σ P | Γ, x : τ ⊢Ψ M : τ ′ P | Γ ⊢Ψ (λx.M)N ≡ [N/x]M : τ ′ {η} Figure 4: H− typing with class contexts. {µ} class Monad m where return : : a → m a (¿¿=) : : m a → (a → m b) → m b {L ET } Note that the variable a in the type of return is not part of the Monad constraint. Thus, we would have that P | Γ ⊢Ψ M : τ → τ ′ P | Γ, x : τ ⊢Ψ M : τ P | Γ ⊢Ψ µx.M ≡ [µx.M/x]M : τ P | Γ ⊢Ψ M : σ ∀t.Eq t ⇒ [t] → [t] → Bool. We introduce abbreviations for the type scheme of each method, in general and at each instance, assuming some class context hA, Si, Imi. For each method name x such that Si(x) = hπ, ∀u.ρi, we define the type scheme for x by: σx = ∀t.∀u. π ⇒ ρ, or, equivalently, writing ρ as Q ⇒ τ : σx = ∀t, u. (π, Q) ⇒ τ where, in each case, t = ftv(π). Similarly, for each method x as above, and each instance d such that • hx, di ∈ dom(Im); • (d : ∀t. P ⇒ π ′ ) ∈ A; and, • there is some substitution S such that S π = π ′ we define the type scheme for x in d by: σx,d = ∀t, u. (P, S Q) ⇒ S τ. P | Γ, x : σ ⊢Ψ N : τ P | Γ ⊢Ψ (let x = M in N) ≡ [M/x]N : τ {M ETHOD} {∀ I} x 6∈ fv(M) P | Γ ⊢Ψ λx.(Mx) ≡ M : τ → τ ′ Si(return) = hMonad m, ∀a.a → m ai. The method implementations themselves are recorded in component Im, which maps pairs of method and instance names to implementing expressions. To describe the typing of methods and method implementations, we begin by describing the type of each method implementation. This is a combination of the defining instance, including its context, and the definition of the method itself. For example, in the instance above, the body of the == method should compare lists of arbitrary type t for equality (this arises from the instance predicate Eq [t ] and the signature of ==), given the assumption Eq t (arising from the defining instance). That is, we would expect it to have the type P | Γ ⊢Ψ N : τ t 6∈ ftv(P, Γ) Si(x) = hπ, σi d:P Sπ P | Γ ⊢hA,Si,Imi x ≡ Im(x, d) : S σ {(P | Γ ⊢Ψ M ≡ N : [τ /t]σ) | τ ∈ GType} P | Γ ⊢Ψ M ≡ N : ∀t.σ {∀ E} {⇒ I} {⇒ E} P | Γ ⊢Ψ M ≡ N : ∀t.σ P | Γ ⊢Ψ M ≡ N : [τ /t]σ P, π | Γ ⊢Ψ M ≡ N : ρ P | Γ ⊢Ψ M ≡ N : π ⇒ ρ P | Γ ⊢Ψ M ≡ N : π ⇒ ρ P π P | Γ ⊢Ψ M ≡ N : ρ Figure 5: Equality for H− terms. 2.2 Equality of H− Terms In this section, we give an axiomatic presentation of equality for H− terms. Our primary concerns are the treatment of polymorphism and class methods; otherwise, H− differs little from standard functional calculi. As described in the introduction, our intention is to permit reasoning about class methods directly, without relying on either a dictionary-passing translation or a preliminary inlining step that resolves all method overloading. This results in two unusual aspects of our rules: • While our presentation gives equality for expressions, it relies critically on components of the class context hA, Si, Imi—the axioms A to determine which instance solves given constraints, and the method implementations Im to determine the behavior of methods. • The treatment of polymorphism cannot be completely paramet- Finally, we give a typing rule parameterized by class contexts in Figure 4; in xi : σxi , the xi range over all methods defined in the program (i.e., over the domain of Si). Intuitively, an expression M has type τ under hA, Si, Imi if: ric, and different equalities may be provable for the same term at different types; for example, we cannot hope to have uniform proofs of properties of the == method when it is defined differently at different types. • None of the class instances overlap. More expressive class sys- Equality judgments take the form P | Γ ⊢Ψ M ≡ N : σ, denoting that, assuming predicates P, variables typed as in Γ, and class context Ψ, expressions M and N are equal at type σ. To simplify the presentation, we have omitted equational assumptions; however, extending our system with assumptions and a corresponding axiom rule would be trivial. The rules are those listed in Figure 5, together with rules for reflexivity, symmetry, and transitivity of equality, and the expected α-equivalence and congruence rules for each syntactic form. Rules {β}, {η}, {µ} and {L ET } should tems will require more elaborate restrictions; we give an example when extending H− to support functional dependencies (§5). • Each method implementation Im(x, d) has the type σx,d (meth- ods are allowed to be mutually recursive). • The main expression has the declared type σ, given that each class method xi has type σxi . be unsurprising. Rules {⇒I} and {⇒E} mirror the corresponding typing rules, assuring that we can only conclude equalities about well-typed expressions. Rule {∀ E} should also be unsurprising: if we have proved that two expressions are equal at a quantified type, we have that they are equal at any of its instances. Rule {∀ I} is less typical, as it requires one subproof for each possible ground type (GType ranges over ground type expressions). Note that this is only non-trivial for terms involving overloading. Finally, rule {M ETHOD} provides (one step of) method resolution. Intuitively, it says that for some class method x at type σ, if instance d proves that x is defined at σ, then x is equal to the implementation of x provided by instance d. 2.3 Polymorphic Identity Functions In the introduction, we gave an example of a polymorphic function (invBool) that could be instantiated at all types, yet did not have parametric behavior. In this section, we will consider a function which does have parametric behavior, but is defined in an ad-hoc fashion. We will demonstrate that our treatment of equality allows us to conclude that it is equal to its parametric equivalent. Our particular example is the identity function. First, we give its typical definition: id1 : : t → t id1 x = x For our second approach, we intend an overloaded definition that is provably equal to the parametric definition. We could produce such a definition using instance chains: class Id2 ' t where id2 ' : : t → t instance (Id2 ' t , Id2 ' u) ⇒ Id2 ' ( t → u) where id2 ' f = id2 ' ◦ f ◦ id2 ' else Id2 ' t where id2 ' x = x This gives an ad-hoc definition of the identity function, defined at all types but defined differently for function and non-function types. Reasoning about this definition would require extending the entailment relation to instance chains, introducing significant additional complexity. We present simpler instances, but restrict the domain of types to achieve a similar result. class Id2 t where id2 : : t → t instance Id2 Int where id2 x = x instance (Id2 t , Id2 u) ⇒ Id2 ( t → u) where id2 f = id2 ◦ f ◦ id2 We will use Int to stand in for all base (non-function) types. It should be intuitive that, while they are defined differently, id1 x and id2 x should each evaluate to x for any integer or function on integers x. Correspondingly, given a class context Ψ that describes (at least) Id2, we can prove that ⊢Ψ id1 ≡ id2 : τ (we omit the empty context and empty assumptions) for any such type τ . The case for integers is direct: one application of {M ETHOD} is sufficient to prove ⊢Ψ id2 ≡ λx.x : Int → Int. For functions of (functions of. . . ) integers, the proof has more steps, but is no more complicated. For the simplest example, to show that ⊢Ψ id2 ≡ λx.x : (Int → Int) → (Int → Int), we use {M ETHOD} to show ⊢Ψ id2 ≡ λf .(id2 ◦ f ◦ id2) : (Int → Int) → (Int → Int). Relying on the usual definition of composition and {β}, we show ⊢Ψ λf .(id2 ◦ f ◦ id2) ≡ λf .λx.id2(f (id2 x)) : (Int → Int) → (Int → Int) Finally, by two uses of {M ETHOD} for id2 on integers, and {η}, we have ⊢Ψ λf .λx.id2(f (id2 x)) ≡ λf .f : (Int → Int) → (Int → Int) and thus the desired result. We cannot expect to prove that id1 ≡ id2 at all types (i.e., ⊢Ψ id1 ≡ id2 : ∀t.t → t) without limiting the domain of types. For example, there is no instance of Id2 at type Bool; therefore, we cannot prove any non-trivial equalities ⊢Ψ id2 ≡ M : Bool → Bool. However, if we were to restrict the grammar of types to those types for which Id2 is defined (that is, if we define that τ ::= Int | τ → τ ), then we could construct such an argument. To show that ⊢Ψ id2 ≡ λx.x : ∀t.t → t, we begin by applying {∀ E}, requiring a derivation ⊢Ψ id2 ≡ λx.x : τ → τ for each ground type τ . We could construct such a set of derivations by induction on the structure of types, using the argument for Int above as the base case, and a construction following the one for Int → Int for the inductive case. A similar approach applies to the formulation using instance chains (class Id2'): we could show that the first clause applied to functions, the second clause applied to any non-function type, and use induction over the structure of types with those cases. 3. A Simple Semantics for Overloading Next, we develop a simple denotational semantics of H− programs, extending an approach originally proposed by Ohori [12] to describe the implicit polymorphism of ML. As with the presentation of equality in the previous section, the primary new challenges arise from the definition of class methods and the treatment of overloading. We will demonstrate that the specialization-based approach is well-suited to addressing both challenges. In particular, it allows expressions to have different interpretations at each ground type without introducing additional arguments or otherwise distinguishing qualified from unqualified type schemes. 3.1 The Meaning of Qualified Types To describe the meaning of overloaded expressions, we must begin with the meaning of qualified types. Intuitively, qualifiers in types can be viewed as predicates in set comprehensions—that is, a class Eq denotes a set of types, and the qualified type ∀t.Eq t ⇒ t → t → Bool describes the set of types {t → t → Bool | t ∈ Eq}. However, most existing approaches to the semantics of overloading do not interpret qualifiers in this fashion: Wadler and Blott [16], for instance, translate qualifiers into dictionary arguments, while Jones [2] translates qualified types into a calculus with explicit evidence abstraction and application. Our approach, by contrast, preserves the intuitive notion of qualifiers. Given some class context Ψ = hA, Si, Imi, we define the ground instances ⌊σ⌋Ψ of an H− type scheme σ by: ⌊τ ⌋Ψ = {τ } ( ⌊ρ⌋Ψ if A π ⌊π ⇒ ρ⌋Ψ = ∅ otherwise [ ⌊∀t.σ⌋Ψ = ⌊[τ /t]σ⌋Ψ . τ ∈GType Equivalently, if we define GSubst(t) to be substitutions that map t to ground types and are otherwise the identity, we have ⌊∀t.P ⇒ τ ⌋Ψ = {S τ | S ∈ GSubst(t, A S P}. We will omit annotation Ψ when it is unambiguous. In the typing judgments for H− , predicates can appear in both types and contexts. To account for both sources of predicates, we adopt Jones’s constrained type schemes (P | σ), where P is a list of predicates and σ is an H− type scheme; an unconstrained type scheme σ can be treated as the constrained scheme (∅ | σ) (as an empty set of predicates places no restrictions on the instantiation of the variables in σ). We can define the ground instances of constrained type schemes by a straightforward extension of the definition for unconstrained schemes: if Ψ = hA, Si, Imi, then ⌊(P | ∀t.Q ⇒ τ )⌋Ψ = {S τ | S ∈ GSubst(t), A (P, S Q)}. 3.2 Type Frames for Polymorphism We intend to give a semantics for H− expressions by giving a mapping from their typing derivations to type-indexed collections of monomorphic behavior. We begin by fixing a suitable domain for the monomorphic behaviors. Ohori assumed an underlying typeframe semantics; his translations, then, were from implicitly polymorphic terms to the interpretations of terms in the simply-typed λ-calculus. Unfortunately, we cannot apply his approach without some extension, as type classes are sufficient to encode polymorphic recursion. However, we can adopt Harrison’s extension [1] of Ohori’s approach, originally proposed to capture polymorphic recursion, and thus also sufficient for type class methods. We begin by defining PCPO frames, an extension of the standard notion of type frames. A PCPO frame is a tuple T = hT type J·K, T term J·K, Tτ,υ , ⊑τ , ⊔τ , ⊥τ i, (where we will omit the type and term annotations when they are apparent from context) subject to the following six conditions. 1. For each ground type τ , T the interpretation of τ . type Jτ K is a non-empty set providing 2. For each typing derivation ∆ of Γ ⊢ M : τ and Γ-compatible environment η, T term J∆Kη is the interpretation of M in T type Jτ K. 3. Tτ,υ : T type Jτ → υK × T type Jτ K → T type JυK provides the interpretation of the application of an element of τ → υ to an element of τ . 4. For any f , g ∈ T type Jτ → υK, if, for all x ∈ T type Jτ K, Tτ,υ (f , x) = Tτ,υ (g, x), then f = g. 5. T term J·K and Tτ,υ respect the semantics of the simply-typed λcalculus. In particular: each type. We would expect its semantics to include the pair hInt → Int, T term J⊢ λx : Int.x : Int → IntKi to account for the Int → Int ground instance of its type scheme, the pair hBool → Bool, T term J⊢ λx : Bool.x : Bool → BoolKi to account for the Bool → Bool ground instance of its type scheme, and so forth. Note that if σ has no quantifiers, and so ⌊σ⌋Ψ = {τ } for some type τ , then we have T scheme JσKΨ = {{hτ, bi} | b ∈ T type Jτ K}, and so an element of T scheme Jτ K is a singleton map, not an element of T type Jτ K. Harrison proves that T JσK is itself a pointed CPO, justifying solving recursive equations in T JσK. Theorem 1 (Harrison). Let T be a PCPO frame. Then, for any type scheme σ, T JσK is a pointed CPO where: • For any f , g ∈ T JσK, f ⊑σ g ⇐⇒ (∀τ ∈ ⌊σ⌋. f (τ ) ⊑τ g(τ )); • The bottom element ⊥σ is defined to be {hτ, ⊥τ i | τ ∈ ⌊σ⌋}; and, • The least upper bound of an ascending chain {fi } ⊆ T JσK is {hτ, uτ i | τ ∈ ⌊σ⌋, uτ = ⊔τ (fi (τ ))}. We can define continuous functions and least fixed points for sets T JσK in the usual fashion: • A function f : T JσK → T Jσ ′ K is continuous if f (⊔σ Xi ) = ⊔σ ′ (f (Xi )) for all directed chains Xi in T JσK. • The fixed point of a continuous function f : T JσK → T JσK is defined by fix(f ) = ⊔σ (f n (⊥σ )), and is the least value such that fix(f ) = f (fix(f )). 3.3 Semantics for Overloaded Expressions We can now give denotations for (typing derivations of) H− expressions. For some type environment Γ and substitution S ∈ GSubst(ftv(Γ)), we define an S − Γ-environment η as a mapping from variables to values such that η(x) ∈ T J(S σ)K for each assignment (x : σ) in Γ. Given a PCPO frame T , a derivation ∆ of P | Γ ⊢A M : σ, a ground substitution S, and an environment η, we define the interpretation T J∆KSη by cases. We have included only a few, representative cases here. • Case (→E): we have a derivation of the form • If ∆ derives Γ ⊢ x : τ , then T J∆Kη = η(x); • If ∆ derives Γ ⊢ M N : υ, ∆M derives Γ ⊢ M : ∆1 = τ → υ and ∆N derives Γ ⊢ N : τ , then T J∆Kη = Tτ,υ (T J∆M Kη, T J∆N Kη); and, ∆= • If ∆λ derives Γ ⊢ λx : τ.M : τ → υ and ∆M derives Γ, x : τ ⊢ M : υ, then Tτ,υ (T J∆λ Kη, d) = T J∆M K(η[x 7→ d]). 6. Each set T Jτ K is a PCPO with respect to ⊑τ , ⊔τ and ⊥τ . The first five conditions are the standard requirements for type frames; the final condition relates the type frame and PCPO structures of a PCPO frame. Given a PCPO frame T , we can define the interpretation of a polymorphic type scheme σ as the mappings from the ground instances τ of σ to elements of T Jτ K. That is: T scheme JσKΨ = Π(τ ∈ ⌊σ⌋Ψ ).T type Jτ K. where we will omit the scheme and Ψ annotations when it is not ambiguous. For example, the identity function λx.x has the type scheme ∀t.t → t. Therefore, the semantics of the identity function is a map from the ground instances of its type (i.e., the types τ → τ ) to the semantics of the simply-typed identity function at .. . P | Γ ⊢A M : τ → τ ′ ∆2 = .. . P | Γ ⊢A N : τ P | Γ ⊢A (M N) : τ ′ Let υ = S τ and υ ′ = S τ ′ , and define T J∆KSη = {hυ ′ , Tυ,υ′ ((T J∆1 KSη)(υ → υ ′ ), (T J∆2 KSη)(υ))i}. • Case (⇒I): we have a derivation of the form .. . ∆1 = P, π | Γ ⊢A M : ρ ∆= P | Γ ⊢A M : π ⇒ ρ This rule excludes those cases in which the predicate does not hold; thus, we define: ( T J∆1 KSη if S P S π; T J∆KSη = ∅ otherwise. • Case (⇒E): we have a derivation of the form .. . ∆1 = P | Γ ⊢A M : π ⇒ ρ P ∆= P | Γ ⊢A M : ρ π This rule does not affect the semantics of expression M, and so we define: T J∆KSη = T J∆1 KSη. • Case (∀ I): we have a derivation of the form .. . ∆1 = P | Γ ⊢A M : σ t 6∈ ftv(P, Γ) ∆= P | Γ ⊢A M : ∀t.σ Intuitively, we interpret a polymorphic expression as the map from ground instances of its type to its interpretations at those types. As the interpretation of the subderivation ∆1 is already in the form of a such a map, we can interpret ∆ as the union of the meanings of ∆1 for each ground instantiation of the quantified variable t. Formally, we define [ T J∆KSη = T J∆1 K(S[t 7→ τ ])η. τ ∈GType • Case (∀ E): we have a derivation of the form ∆1 = ∆= .. . P | Γ ⊢A M : ∀t.σ P | Γ ⊢A M : [τ /t]σ S By definition, ⌊∀t.σ⌋ = τ ∈GType ⌊[τ /t]σ⌋, and so ⌊[τ /t]σ⌋ ⊆ ⌊∀t.σ⌋. Thus, the interpretation of ∆ is a subset of the interpretation of ∆1 ; writing f |Y for the restriction of a function f to some subset Y of its domain, we define: T J∆KSη = (T J∆1 KSη)|⌊[τ /t]σ⌋. 3.4 Expressions with Class Contexts To complete our semantics of H− programs, we must account for the meaning of class methods. Our approach is intuitively simple: we collect the meanings of the class methods from the method implementations in each instance, and use the meanings of the methods to define the meaning of the main expression. Formally, we extend the interpretation function from derivations of P | Γ ⊢A M : σ to derivations of P | Γ ⊢Ψ M : σ as follows: • Let ∆ be a derivation of P | Γ ⊢Ψ M : τ . Then we know that ∆ must begin with an application of (C TXT ) (Figure 4) with one subderivation ∆y,d = .. . P | Γ, xi : σxi ⊢A Im(y, d) : σy,d for each pair hy, di ∈ dom(Im) and a subderivation ∆M = .. . P | Γ, xi : σxi ⊢A M : τ for the main expression M. We enumerate the methods in the program as x1 , x2 , . . . , xm , and let Σ = T Jσx1 K × T Jσx2 K × · · · × T Jσxm K. For each method xi , we define a function fi : Σ → T Jσxi K, approximating its meaning, as follows: [ T J∆xi ,d KS(η[xj 7→ bj ]), fi (hb1 , b2 , . . . , bm i)Sη = hxi ,di∈dom(Im) and define function f : Σ → Σ, approximating the meaning of all the methods in the program, as f (b) = hf1 (b), f2 (b), . . . , fm (b)i. We can now define a tuple b, such that the component bi is the meaning of method xi , as follows: G n f (⊥Σ ). b= Σ Finally, we extend the interpretation function to programs by T J∆KSη = T J∆M KS(η[xi 7→ bi ]). 3.5 Polymorphic Identity Functions Revisited We return to our earlier example of polymorphic identity functions (§2.3). As before, we consider two definitions of identity functions, one given parametrically (id1) and one given by overloading (id2). In this section, we will show that the denotations of id1 and id2 agree at all types for which id2 is defined. By doing so, we provide an intuitive demonstration that our denotational semantics captures the meaning of ad-hoc polymorphic and agrees with our definition of equality for H− terms. We show that T Jid1K and T Jid2K have the same value at each point in the domain of T Jid2K; that is, that for any type τ ∈ GType such that Id2 τ , T Jid1K(τ → τ ) = T Jid2K(τ → τ ). We proceed by induction on the structure of τ . In the base case, we know that τ = K for some non-functional type K. As we have assumed Id2 τ , we must have that K = Int, and, from the instances for Id2, we have T Jid2K(K → K) = T Jid2K(Int → Int) = T J⊢ λx : Int.x : Int → IntK. As T Jid1K(Int → Int) = T J⊢ λx : Int.x : Int → IntK, we have T Jid1K(K → K) = T Jid2K(K → K). In the inductive case, we know that τ = τ0 → τ1 for some types τ0 and τ1 . From the assumption that Id2 (τ0 → τ1 ) and the instances for Id2, we can assume that Id2 τ0 , Id2 τ1 , and that T Jid2K(τ → τ ) = T J⊢ λf : (τ0 → τ1 ).M ◦ f ◦ N : τ → τ K for some simply typed expressions M and N such that T JMK = T Jid2K(τ0 → τ0 ) and T JNK = T Jid2K(τ1 → τ1 ). The induction hypothesis gives that T Jid2K(τ0 → τ0 ) = T Jid1K(τ0 → τ0 ) and that T Jid2K(τ1 → τ1 ) = T Jid1K(τ1 → τ1 ), and thus that T JMK = T J⊢ λx : τ1 .x : τ1 → τ1 K and T JNK = T J⊢ λx : τ0 .x : τ0 → τ0 K. By congruence, we have T Jid2K(τ → τ ) = T Jλf : (τ0 → τ1 ).(λx : τ1 .x)◦f ◦(λx : τ0 .x)K. Finally, assuming a standard definition of composition, and reducing, we have T Jid2K(τ → τ ) = T Jλf : (τ0 → τ1 ).f K = T Jλf : τ.f K = T Jid1K(τ → τ ). In our previous discussion of this example, we argued that if the set of types were restricted to those types for which Id2 held, then id1 and id2 were equal. We can show a similar result here, by showing that if we define that τ ::= Int | τ → τ , then T Jid1K = T Jid2K. We begin by showing that they are defined over the same domain; that is, that ⌊∀t. t → t⌋ = ⌊∀u.Id2 u ⇒ u → u⌋. By definition, we have • Case (∀ I). Because σ = ∀t.σ ′ , we have that ⌊σ⌋ = ⌊∀t. t → t⌋ = {τ → τ | τ ∈ GType} [ ⌊[τ /t]σ ′ ⌋, τ ∈GType and and thus that ⌊∀u. Id2 u ⇒ u → u⌋ = {τ → τ | τ ∈ GType, Id2 τ }. We show that Id2 τ for all types τ by induction on the structure of τ . In the base case, we know that τ = Int, and by the first instance of Id2 we have Id2 τ . In the inductive case, we know that τ = τ0 → τ1 for some types τ0 , τ1 . In this case, we have that [τ0 /t, τ1 /u]τ = t → u and by the induction hypothesis, that Id2 τ0 and Id2 τ1 . Thus, from the second instance of Id2, we can conclude that Id2 (τo → τ1 ), that is, that Id2 τ . Because Id2 τ for all ground types τ , we have {τ → τ | τ ∈ GType, T JσK = 4. Formal Properties The previous sections have outlined typing and equality judgments for H− terms, and proposed a denotational semantics for H− typings. In this section, we will relate these two views of the language. We begin by showing that the denotation of a typing judgment falls into the expected type. This is mostly unsurprising; the only unusual aspect of H− in this respect is the role of the class context. We go on to show that the equational judgments are sound; again, the unusual aspect is to do with polymorphism ({∀ I} and {∀ E}) and class methods ({M ETHOD}). The H− type system follows Jones’s original formulation of OML; we rely on several of his metatheoretical results, such as the closure of typing under substitution. Theorem 2 (Soundness of typing). Given a class context Ψ, if ∆ is a derivation of P | Γ ⊢Ψ M : σ, S is a substitution, and η is an (S Γ)-environment, then T J∆KSη ∈ T J(S P | S σ)KΨ . We will divide the proof into three pieces. First, we show the soundness of the judgment P | Γ ⊢A M : σ. Then, we will argue that the union of the implementations of a method has the type of the method itself. Finally, we can combine these results to argue the soundness of P | Γ ⊢Ψ M : σ. Lemma 3. Given a class context Ψ = hA, Si, Imi where A is non-overlapping, if ∆ is a derivation of P | Γ ⊢A M : σ, S is a substitution, and η is a (S Γ)-environment, then T J∆KSη ∈ T J(S P | S σ)KΨ . Proof. The proof is by induction over the structure of derivation ∆. The cases are straightforward; we include several representative examples. (Meta-variables ∆n are as in the definition of T J·K above.) • Case (⇒ I). Observe that ⌊(S(P, π) | S ρ)⌋ = ⌊(S P | S (π ⇒ ρ))⌋. As such, if T J∆1 KSη ∈ T J(S (P, π) | S ρ)KΨ , then we must also have that T J∆KSη ∈ T J(S P | S (π ⇒ ρ))KΨ . • Case (⇒ E). As entailment is (trivially) closed under substitu- tion, P π implies that S P S π for any substitution S; thus, we can conclude that ⌊(S P | S (π ⇒ ρ))⌋ = ⌊(S P | S ρ)⌋. Finally, assuming that T J∆1 KSη ∈ T J(S P | S (π ⇒ ρ))K, we can conclude that T J∆KSη ∈ T J(S P | S ρ)K. (T J[τ /t]σ ′ K). τ ∈GType Thus, assuming that for ground types τ , T J∆1 K(S[t 7→ τ ])η ∈ T J(S P | S σ ′ )K, we have ! [ ′ T J∆KSη ∈ T J(S P | S σ )K = T J(S P | S σ)K. τ ∈GType • Case (∀ E). Assuming that T J∆1 KSη ∈ T J(S P | S (∀t.σ ′ ))K, the same argument about ground types as in the previous case gives that T J∆KSη ∈ T J(S P | S σ)K. Id2 τ } = {τ → τ | τ ∈ GType}, and so T Jid1K and T Jid2K are defined over the same domain. We have already shown that T Jid1K and T Jid2K agree at all points at which they are defined, and so we conclude T Jid1K = T Jid2K. [ The interpretation of typings P | Γ ⊢Ψ M : σ depends on the interpretations of the class methods. We will begin by showing that the interpretation of each method is in the denotation of its type. To do so, we will demonstrate that the interpretation of the type scheme of a method is the union of the interpretation of the type schemes of its instances. This will show that the union of the implementations is in the type of the method, from which the desired result follows immediately. Lemma 4. The ground instances of the type scheme of a method x are the union of its ground instances at each of its instances. That is, [ ⌊σx,d ⌋. ⌊σx ⌋ = hx,di∈dom(Im) Proof. Let σx = ∀t.(π, Q) ⇒ τ , where x is a method of class(π). We prove that [ ⌊σx,d ⌋ ⌊σx ⌋ = hd,xi∈dom(Im) by the inclusions ⌊σx ⌋ ⊆ [ ⌊σx,d ⌋, [ ⌊σx,d ⌋. hx,di∈dom(Im) and ⌊σx ⌋ ⊇ hx,di∈dom(Im) We will show only the first inclusion; the second is by an identical argument. Fix some υ ∈ ⌊σx ⌋. By definition, there is some S ∈ GSubst(t) such that υ = S τ and S π, S Q. Because S π, there must be some (d : ∀u. P ⇒ π ′ ) ∈ A and substitution S′ ∈ GSubst(u) such that S π = S′ π ′ and S′ P. Now, we have that σx,d = ∀t′ .(P, T Q) ⇒ T τ for some substitution T; thus, there is some T ′ ∈ GSubst(t′ ) such that υ = T ′ (T τ ), S P = T ′ (T Q), and so υ ∈ ⌊σx,d ⌋. Lemma 5. The interpretation of the type scheme of a method x is the union of the interpretations of its type scheme at each instance. That is, [ T Jσx,d K. T Jσx K = hx,di∈dom(Im) Proof. Recall that T scheme Jσx K = Π(τ ∈ ⌊σx ⌋).T type Jτ K. From Lemma 4, we have that  T scheme Jσx K = Π τ ∈ [  ⌊σx,d ⌋ .T type Jτ K. hx,di∈dom(Im) As T type J·K is a function, this is equivalent to [ Π(τ ∈ ⌊σx,d ⌋).T type Jτ K, T scheme Jσx K = hx,di∈dom(Im) and finally, again from the definition of T scheme J·K, [ T scheme Jσx,d K. T scheme Jσx K = hx,di∈dom(Im) Proof of Theorem 2. Finally, we can extend the soundness of our semantics to include class contexts. From Lemmas 4 and 5, we know that the interpretations of the methods fall in the interpretations of their type schemes, and so if η is a S − Γ-environment, then η[xi 7→ bi ] is a S − (Γ, xi : σxi )-environment. From Theorem 3, we have that T J∆M KS(η[xi 7→ bi ) ∈ T J(S P | S σ)KΨ , and thus that T J∆KSη ∈ T J(S P | S σ)KΨ . We would like to know that the meaning of an expression is independent of the particular choice of typing derivation. Unfortunately, this is not true in general for systems with type classes. A typical example involves the read and show methods, which have the following type signatures read : : Read t ⇒ String → t show : : Show t ⇒ t → String We can construct an expression show◦read of type (Read t, Show t) ⇒ String → String, where variable t can be instantiated arbitrarily in the typing, changing the meaning of the expression. To avoid this problem, we adopt the notion of an unambiguous type scheme from Jones’s work on coherence for qualified types [3]. Definition 6. A type scheme σ = ∀~t.P ⇒ τ is unambiguous if ftv(P) ⊆ ftv(τ ). As long as we restrict our attention to unambiguous type schemes, we have the expected coherence result. For example, suppose that ∆ is a derivation of P | Γ ⊢A λx.M : σ. We observe that ∆ must conclude with an application of (→I), say at P0 | Γ ⊢A λx.M : τ → τ ′ , followed by a series of applications of (⇒I), (⇒E), (∀ I) and (∀ E). While these latter applications determine σ, we can see intuitively that each υ ∈ ⌊σ⌋ must be a substitution instance of τ → τ ′ , and that the interpretation of ∆ at each ground type must be the interpretation of an instance of the subderivation ending with (→I). We can formalize these two observations by the following lemma. Lemma 7. If σ = ∀t.Q ⇒ τ , and ∆1 . . . ∆n is a sequence of derivations such that: • • • • ∆1 is a derivation of P1 | Γ ⊢A M : τ1 ; ∆n is a derivation of P | Γ ⊢A M : σ; Each of ∆2 . . . ∆n is by (⇒I), (⇒E), (∀ I) or (∀ E); and, Each ∆i is the principal subderivation of ∆i+1 then (a) There is a substitution S such that τ = S τ1 and P ∪ Q ⊣⊢ S P1 ; and, (b) For all ground substitutions S, for all υ ∈ ⌊S σ⌋, there is a unique S′ such that T J∆n KSηυ = T J∆1 KS′ ηυ. The proof is by induction on n; the cases are all trivial. We can now characterize the relationship between different typings of M. Theorem 8 (Coherence of T J·K). If ∆ derives P | Γ ⊢A M : σ and ∆′ derives P′ | Γ′ ⊢A M : σ ′ , where σ and σ ′ are unambiguous, then for all substitutions S and S′ such that S P ⊣⊢ S′ P′ , S Γ = S′ Γ′ , and S σ = S′ σ ′ , and for all ground substitutions U, T J∆K(U ◦ S) = T J∆′ K(U ◦ S′ ). The proof is by induction over the structure of M. In each case, use of the inductive hypothesis is justified by Lemma 7(a), and the conclusion derived from the definition of T J·K and Lemma 7(b). As an immediate corollary, we have that if ∆ and ∆′ are two derivations of the same typing judgment, then T J∆K = T J∆′ K. We can also show that, if P | Γ ⊢A M : σ is a principal typing of M, with derivation ∆, and ∆′ derives P | Γ ⊢A M : σ ′ for any other σ ′ , then for each substitution S′ there is a unique S such that, for all environments η, T J∆KSη ⊇ T J∆′ KS′ η. Theorem 9 (Soundness of ≡). Given a class context Ψ, if σ is unambiguous, P | Γ ⊢Ψ M ≡ N : σ, and ∆M , ∆N are derivations of P | Γ ⊢Ψ M : σ, P | Γ ⊢Ψ N : σ, then T J∆M K = T J∆N K. Proof. The proof is by induction over the derivation of P | Γ ⊢Ψ M ≡ N : σ. The interesting cases are to do with polymorphism and overloading. • Case {⇒ I}. We have a derivation concluding P, π | Γ ⊢Ψ M ≡ N : ρ P | Γ ⊢Ψ M ≡ N : π ⇒ ρ Let ∆M , ∆N be typing derivations of P | Γ ⊢A M : π ⇒ ρ and P | Γ ⊢A N : π ⇒ ρ; without loss of generality (because of Theorem 8), assume that each is by (⇒I), with subderivations ∆′M , ∆′N of P, π | Γ ⊢Ψ M : ρ and P, π | Γ ⊢Ψ N : ρ. From the definition of T J·K, we have T J∆M K = T J∆′M K and T J∆N K = T J∆′N K. The induction hypothesis gives that T J∆′M K = T J∆′N K, and so we can conclude T J∆M K = T J∆N K. • Case {⇒ E}. We have a derivation concluding P | Γ ⊢Ψ M ≡ N : π ⇒ ρ P A π P | Γ ⊢Ψ M ≡ N : ρ where Ψ = hA, Si, Imi. As in the previous case, the interpretation of the typing derivations for P | Γ ⊢Ψ M : ρ and P | Γ ⊢Ψ M : π ⇒ ρ are equal, and similarly for the typing derivations for N, and thus the induction hypothesis is sufficient for the desired conclusion. • Case {∀ I}. We have a derivation concluding {(P | Γ ⊢Ψ M ≡ N : [τ /t]σ) | τ ∈ GType} P | Γ ⊢Ψ M ≡ N : ∀t.σ From the induction hypothesis, we can conclude that, given derivations ∆τM of P | Γ ⊢Ψ M : [τ /t]σ and ∆τN of P | Γ ⊢Ψ N : [τ /t]σ, T J∆τM K = T J∆τN K. Let ∆M derive P | Γ ⊢Ψ M : ∀t.σ (and, without loss ofSgenerality, assume ∆M is by (∀I)); we know that T J∆M K = τ ∈GType T J∆τM K. We argue similarly for derivations ∆N of P | Γ ⊢Ψ N : ∀t.σ, and conclude that T J∆M K = T J∆N K. • Case {∀ E}. We have a derivation concluding P | Γ ⊢Ψ M ≡ N : ∀t.σ P | Γ ⊢Ψ M ≡ N : [τ /t]σ Let ∆M , ∆N be derivations that M and N have type [τ /t]σ; without loss of generality, assume they are by (∀E), with subderivations ∆′M , ∆′N that M and N have type ∀t.σ. From the induction hypothesis, we know T J∆′M K = T J∆′N K, and from the definition of T J·K we know that T J∆M K ⊆ T J∆′M K and T J∆N K ⊆ T J∆′N K. Thus, we can conclude that T J∆M K = T J∆N K. • Case {M ETHOD}. We have a derivation of the form Si(x) = π, σ d:P A Sπ P | Γ ⊢hA,Si,Imi x ≡ Im(x, d) : S σ Let ∆M be the derivation of P | Γ ⊢Ψ x : S σ. From the definition of T J·K, we know that T J∆M KSη = T J∆′M KS(η[xi 7→ bi ]) where the xi are the class methods, the bi are their implementations, and ∆′M is the derivation of P | Γ, xi : σi ⊢A x : S σ. Since x is a class method, we know that η[xi 7→ bi ] maps x to some method implementation bj , and therefore that T J∆′M K ⊆ bj . We also knowSthat bj is the fixed point of a function fj (hb1 , . . . , bn i)Sη = d T J∆x,d′ KS(η[xi 7→ bi ]), where ∆x,d′ derives P | Γ ⊢A Im(x, d′ ) : σx,d′ and d is one of the di . Thus, we know that if ∆N derives P | Γ ⊢Ψ Im(x, d) : S σ, then T J∆N K ⊆ bj . Finally, as T J∆M K and T J∆N K are defined over the same domain, we have that T J∆M K = T J∆N K. 5. Improvement and Functional Dependencies In the introduction, we set out several ways in which extensions of type class systems went beyond the expressiveness of existing semantic approaches to overloading. In this section, we return to one of those examples, demonstrating the flexibility of our specialization-based approach to type-class semantics. Functional dependencies [5] are a widely-used extension of type classes which capture relationships among parameters in multiparameter type classes. Earlier, we gave a class Elems to abstract over common operations on collections: class Elems c e | c → e where empty : : c insert : : e → c → c The functional dependency c → e indicates that the type of a collection (c) determines the type of its elements (e). Practically speaking, this has two consequences: • A program is only valid if the instances in the program respect the declared functional dependencies. For example, if a program already contained an instance which interpreted lists as collections: instance Elems [ t ] t where . . . the programmer could not later add an instance that interpreted strings (lists of characters in Haskell) as collections of codepoints (for simplicity represented as integers): instance Elems [Char ] Int • Given two predicates Elems τ υ and Elems τ ′ υ ′ , if we know τ = τ ′ , then we must have υ = υ ′ for both predicates to be satisfiable. We now consider an extension of H− to support functional dependencies. Following Jones [4], we introduce a syntactic characterization of improving substitutions, one way of describing predicate-induced type equivalence. We then extend the typing and equality judgments to take account of improving substitutions. Finally, we show that the extended systems are sound with respect to our semantics. Importantly, we do not have to extend the models of terms, nor do we introduce coercions, or other intermediate translations. We need only show that our characterization of improving substitutions is sound to show that the resulting type equivalences hold in the semantics. 5.1 Extending H− with Functional Dependencies To account for the satisfiability of predicates in qualified types, Jones introduces the notion of an improving substitution S for a set of predicates P [4]. Intuitively, a S improves P if every satisfiable ground instance of P is also a ground instance of S P. Jones uses improving substitutions to refine the results of type inference while still inferring principal types. We will adopt a similar approach, but in typing instead of type inference. Syntax. We begin by extending the syntax of class axioms to include functional dependency assertions: Index sets Class axioms X, Y ⊆ N α ::= C : X Y | d : ∀t. P ⇒ π In the representation of functional dependency axioms, we treat the class parameters by index rather than by name. If A were the axioms for the example above, we would expect to have a dependency Elems : {0} {1} ∈ A. Any particular class name may appear in many functional dependency assertions, or in none at all. We adopt some notational abbreviations: if X is an index set, we write π =X π ′ to indicate that π and π ′ agree at least on those parameters with indices in X, and S similarly write π ∼X π ′ to indicate that S is a unifier for those parameters of π and π ′ with indices in X. Improvement. To account for improvement in typing, we need a syntactic characterization of improving substitutions. In the case of functional dependencies, this can be given quite directly. We can give an improvement rule as a direct translation of the intuitive description above: (F UNDEP ) (C : X P Cτ Y) ∈ A P Cυ S τ =X υ τ ∼Y υ A ⊢ S improves P For example, if we have some Q such that Q Elems τ υ and Q Elems τ υ ′ , then (F UNDEP ) says that the any unifying substitution U such that U υ = U υ ′ is an improving substitution for Q. If S is an improving substitution for P, then the qualified type schemes (P | σ) and (S P | S σ) are equivalent, and we should be able to replace one with the other at will in typing derivations. One direction is already possible: if a term has type σ, then it is always possible to use it with type S σ (by a suitable series of applications of (∀ I) and (∀ E)). On the other hand, there is not (in general) a way with our existing typing rules to use a term of type S σ as a term of type σ. We add a typing rule to support this case. (I MPR ) S P | S Γ ⊢A M : S σ A ⊢ S improves P P | Γ ⊢A M : σ As in the case of (⇒I) and (⇒E), (I MPR ) has no effect on the semantics of terms. Thus, if we have a derivation .. . ∆1 = S P | S Γ ⊢A M : S σ A ⊢ S improves P ∆= P | Γ ⊢A M : σ we define that T J∆KS′ η = T J∆1 KS′′ η, where S′′ ◦ S = S′ (the existence of such an S′′ is guaranteed by the soundness of (F UN DEP )). Finally, we add a rule to the equality judgment allowing us to use improving substitutions in equality proofs. {I MPR} S P | S Γ ⊢hA,Si,Imi M ≡ N : S σ A ⊢ S improves P P | Γ ⊢hA,Si,Imi M ≡ N : σ Validating Functional Dependency Axioms. We must augment the context rule to check that the axioms respect the declared dependencies. This can be accomplished by, first, refining the overlap check to assure that no axioms overlap on the determining parameters of a functional dependencies, and second, requiring that, for each dependency C : X Y and each instance P ⇒ π of class C, any variables in the positions Y are determined by the functional dependencies of P. Our formalization of the latter notion follows Jones’s development [6]. We define the closure of a set of variables J with respect to the functional dependencies F as the least set JF+ such that • J ⊆ JF+ ; and V ∈ F and U ⊆ JF+ , then V ⊆ JF+ . S We write ftvX (C τ ) to abbreviate x∈X ftv(τx ), define the instantiation of a functional dependency assertion C : X Y at a predicate π = C τ , as the dependency ftvX (π) ftvY (π), and write fd(A, P) for the set of the instantiation of each functional dependency assertion in A at each predicate in P. We can now define the verification conditions for axioms and the new version of (C TXT ), as follows. • If U {π ≁X π ′ | (d : P ⇒ π), (d′ : P′ ⇒ π ′ ), (class(π) : X Y) ∈ A} Theorem 11 (Soundness of typing). Given a class context Ψ, if ∆ is a derivation of P | Γ ⊢Ψ M : σ, S is a substitution, and η is an (S Γ)-environment, then T J∆KSη ∈ T J(S P | S σ)KΨ . Proof. We need only consider the (I MPR ) case. From Lemma 10, we have that if T improves P, then T J(P | σ)KΨ = T J(T P | T σ)KΨ , and so the result follows from the induction hypothesis. We extend our notion of ambiguity to take account of functional dependencies: it is enough for the variables in the predicates P to be determined by the variables of τ . Definition 12. A type scheme σ = ∀~t.P ⇒ τ is unambiguous (given class axioms A) if ftv(P) ⊆ ftv(τ )+ fd(A,P) . The previous definition of ambiguity is a special case of this definition, where fd(A, P) is always empty. As uses of (I MPR ) do not affect the semantics of terms, its introduction does not compromise coherence. Theorem 13. If σ is unambiguous and ∆1 , ∆2 are derivations of P | Γ ⊢Ψ M : σ, then T J∆1 K = T J∆2 K. ⊢ non-overlapping(A) {ftv(πY ) ⊆ ftv(πX )+ | (d : P ⇒ π), (class(π) : X fd(A,P) Y) ∈ A} ⊢ covering(A) (C TXT ) There is some type variable in ftvY (π0′ ) \ ftvX (π0′ ) that is not constrained by P. This contradicts ⊢ covering(A). There is some π ′ ∈ P such that S0 π ′ and S1 π ′ violate a functional dependency of class(π ′ ). The derivations of S0 π ′ and S1 π ′ must be shorter than the derivations of π0 , π1 , and so we have the desired result by induction. ⊢ non-overlapping(A) ⊢ covering(A) {(P | Γ, xi : σxi ⊢A Im(y, d) : σy,d ) | hy, di ∈ dom(Im)} P | Γ, xi : σxi ⊢A M : σ P | Γ ⊢hA,Si,Imi M : σ Theorem 14 (Soundness of ≡). Given a class context Ψ, if σ is unambiguous, P | Γ ⊢Ψ M ≡ N : σ, and ∆M , ∆N are derivations of P | Γ ⊢Ψ M : σ, P | Γ ⊢Ψ N : σ, then T J∆M KΨ = T J∆N KΨ . Proof. Again, we need consider only the {I MPR} case. Without loss of generality, assume ∆M and ∆N are by (I MPR ), with subderivations ∆′M and ∆′N . As the interpretations of ∆M and ∆N are equal to the interpretations of ∆′M and ∆′N , the result follows from the induction hypothesis. 5.2 Soundness The significant challenge in proving soundness of the extended rules is showing that when A ⊢ S improves P is derivable, S is an improving substitution for P. Once we have established that result, the remaining soundness results will be direct. We introduce notation for the satisfiable ground instances of predicates P: ⌊P⌋A = {S P | S ∈ GSubst(ftv(P)), A S P}. We can now formally describe an improving substitution. Lemma 10. Given a set of axioms A such that ⊢ non-overlapping(A) and ⊢ covering(A), if A ⊢ S improves P, then ⌊P⌋A = ⌊S P⌋A . Proof. By contradiction. Assume that A ⊢ S improves P; then we must have π0 , π1 such that A π0 , A π1 and there is a functional dependency (class(π) : X Y) ∈ A such that π0 =X π1 but π0 6=Y π1 . We proceed by induction on the heights of the derivations of A π0 , A π1 . • There are distinct axioms d : P ⇒ π0′ , d′ : P′ ⇒ π1′ ∈ A and substitutions S0 , S1 such that S0 π0′ = π0 and S1 π1′ = π1′ . But then S0 ◦ S1 is a unifier for π0′ ∼X π1′ , contradicting ⊢ non-overlapping(A). • There is a single axiom d : P ⇒ π0′ and substitutions S0 , S1 such that S0 π0′ = π0 and S1 π0′ = π1 . We identify two subcases. 6. Related Work The semantics of polymorphism, in its various forms, has been studied extensively over the past half century; however, the particular extensions of Haskell that motivated this work are recent, and have received little formal attention. Our approach was inspired by Ohori’s semantics of Core ML [12]. While Ohori’s approach describes the semantics of polymorphism, he does not represent polymorphic values directly, which leads to an unusual treatment of the typing of let expressions. Harrison extends Ohori’s approach to treat polymorphic recursion [1]; in doing so, he provides a representation of polymorphic values. Harrison suggests that his approach could be applied to type classes as well. Ohori’s approach to the semantics of ML is somewhat unusual; more typical approaches include those of Milner [7] and Mitchell and Harper [9]. Ohori identifies reasons to prefer his approach over either that of Milner or that of Mitchell and Harper: both approaches use a semantic domain with far more values than correspond to values of ML, either because (in the untyped case) those values would not be well-typed, or (in the explicit typed case) they differ only in the type-level operations. The semantics of type-class-based overloading has also received significant attention. Wadler and Blott [16] described the meaning of type classes using a dictionary-passing translation, in which overloaded expressions are parameterized by type-specific implementations of class methods. Applying their approach to the full Haskell language, however, requires a target language with more complex types than their source language. For example, in translating the Monad class from the Haskell prelude, the dictionary for Monad τ must contain polymorphic values for the return and (¿¿=) methods. In his system of qualified types [2], Jones generalized the treatment of evidence by translating from a language with overloading (OML) to a language with explicit evidence abstraction and application. Jones does not provide a semantics of the language with explicit evidence abstraction and application; indeed, such a semantics could not usefully be defined without choosing a particular form of predicate, and thus a particular form of evidence. Odersky, Wadler and Wehr [11] propose an alternative formulation of overloading, including a type system and type inference algorithm, and a ideal-based semantics of qualified types. However, their approach requires a substantial restriction to the types of overloaded values which rules out many functions in the Haskell prelude as well as the examples from our previous work [10]. Jones [5] introduced functional dependencies in type classes, and discusses their use to improve type inference; his presentation of improvement is similar to ours, but he does not augment typing as does our (I MPR ) rule. Sulzmann et al. [15] give an alternative approach to the interaction of functional dependencies and type inference, via a translation into constraint-handling rules; unfortunately, their presentation conflates properties of their translation, such as termination, with properties of the relations themselves. System FC [14] extends System F with type-level equality constraints and corresponding coercion terms. While we are not aware of any formal presentation of functional dependencies in terms of System FC , we believe that a formulation of our (F UNDEP ) rule in terms of equality constraints is possible. In contrast to our approach, System FC requires extending the domain of the semantics, while still requiring translation of source-level features (functional dependencies or GADTs) into features of the semantics (equality constraints). 7. Conclusion We have proposed an alternative approach to the semantics of overloading, based on interpreting polymorphic values as sets of their monomorphic interpretations, which avoids several problems with traditional translation-based approaches. We have applied this result to a simple overloaded calculus, and shown the soundness of its typing and equality judgments. Finally, we have argued that the approach is flexible enough to support extensions to the type system, such as allowing the use of improving substitutions in typing. We conclude by identifying directions for future work: • Practical class systems are richer than the one used in this paper. We would like to extend these results to fuller systems, including our prior work on instance chains. • Dictionary-passing provides both a semantics of overloading and an implementation technique. We would like to explore whether implementation techniques based on specialization can be used to compile practical languages. • We claim that our approach avoids making distinctions between some observationally equivalent terms (such as in the polymorphic identity function example). We would like to explore whether adequacy and full abstraction results for the underlying frame model can be extended to similar results for our semantics. • Our definition of equality provides η-equivalence; however, η equivalence is not sound for Haskell. We would like to explore either whether our approach can be adapted to a language without η-equivalence. Acknowledgments. We would like to thank: Mark Jones for initially suggesting Ohori’s semantics of ML polymorphism as a basis for understanding overloading; Jim Hook for proposing the polymorphic identity function example; and, Keiko Nakata for her helpful feedback on drafts of the paper. References [1] W. Harrison. A simple semantics for polymorphic recursion. In Proceedings of the 3rd Asian Symposium on Programming Languages and Systems, APLAS ’05, pages 37–51, Tsukuba, Japan, 2005. SpringerVerlag. [2] M. P. Jones. A theory of qualified types. In B. K. Bruckner, editor, Proceedings of the 4th European symposium on programming, volume 582 of ESOP’92. Springer-Verlag, Rennes, France, 1992. [3] M. P. Jones. Coherence for qualified types. Technical Report YALEU/ DCS/RR-989, Yale University, 1993. [4] M. P. Jones. Simplifying and improving qualified types. In Proceedings of the seventh international conference on Functional programming languages and computer architecture, FPCA ’95, pages 160– 169, La Jolla, California, USA, 1995. ACM. [5] M. P. Jones. Type classes with functional dependencies. In Proceedings of the 9th European Symposium on Programming Languages and Systems, ESOP ’00, pages 230–244, Berlin, Germany, 2000. SpringerVerlag. [6] M. P. Jones and I. S. Diatchki. Language and program design for functional dependencies. In Proceedings of the first ACM SIGPLAN symposium on Haskell, Haskell ’08, pages 87–98, Victoria, BC, Canada, 2008. ACM. [7] R. Milner. A theory of type polymorphism in programming. Journal of Computer and System Sciences, (17):348–375, 1978. [8] J. C. Mitchell. Polymorphic type inference and containment. Inf. Comput., 76(2-3):211–249, Feb. 1988. [9] J. C. Mitchell and R. Harper. The essence of ML. In Proceedings of the 15th ACM SIGPLAN-SIGACT symposium on Principles of programming languages, POPL ’88, pages 28–46, San Diego, California, USA, 1988. ACM. [10] J. G. Morris and M. P. Jones. Instance chains: Type-class programming without overlapping instances. In Proceedings of the 15th ACM SIGPLAN international conference on Functional programming, ICFP ’10, Baltimore, MD, 2010. ACM. [11] M. Odersky, P. Wadler, and M. Wehr. A second look at overloading. In Proceedings of the seventh international conference on Functional programming languages and computer architecture, FPCA ’95, pages 135–146, La Jolla, California, USA, 1995. ACM. [12] A. Ohori. A simple semantics for ML polymorphism. In Proceedings of the fourth international conference on Functional programming languages and computer architecture, FPCA ’89, pages 281–292, London, UK, 1989. ACM. [13] S. Peyton Jones, M. P. Jones, and E. Meijer. Type classes: an exploration of the design space. In Proceedings of the 1997 workshop on Haskell, Haskell ’97, Amsterdam, The Netherlands, 1997. [14] M. Sulzmann, M. M. T. Chakravarty, S. P. Jones, and K. Donnelly. System f with type equality coercions. In Proceedings of the 2007 ACM SIGPLAN international workshop on Types in language design and implementation, TLDI ’07, pages 53–66, Nice, France, 2007. ACM. [15] M. Sulzmann, G. J. Duck, S. Peyton Jones, and P. J. Stuckey. Understanding functional dependencies via constraint handling rules. JFP, 17(1):83–129, 2007. [16] P. Wadler and S. Blott. How to make ad-hoc polymorphism less ad hoc. In Proceedings of the 16th ACM SIGPLAN-SIGACT symposium on Principles of programming languages, POPL ’89, pages 60–76, Austin, Texas, USA, 1989. ACM.
6cs.PL
arXiv:1405.6963v1 [math.AC] 27 May 2014 PSEUDO-GORENSTEIN AND LEVEL HIBI RINGS VIVIANA ENE, JÜRGEN HERZOG, TAKAYUKI HIBI AND SARA SAEEDI MADANI Abstract. We introduce pseudo-Gorenstein rings and characterize those Hibi rings attached to a finite distributive lattice L which are pseudo-Gorenstein. The characterization is given in terms of the poset of join-irreducible elements of L. We also present a necessary condition for Hibi rings to be level. Special attention is given to planar and hyper-planar lattices. Finally the pseudo-Goresntein and level property of Hibi rings and generalized Hibi rings is compared with each other. Introduction Let K be a field. Naturally attached to a finite distributive lattice L is a K-algebra K[L] which nowadays is called the Hibi ring of L. This K-algebra was introduced by the third author in 1987, see [11]. In that paper it is shown that K[L] is a normal Cohen–Macaulay domain and that K[L] is Gorenstein if and only if the poset P of join-irreducible elements of L is pure. Let R be an arbitrary standard graded Cohen–Macaulay K-algebra with canonical module ωR . Then R is Gorenstein if and only if ωR is a cyclic module, and hence generated in a single degree. The condition on ωR may be weakened in different ways. If one only requires that the generators of ωR are of the same degree, then R is called a level ring, and if one requires that there is only one generator of least degree, then we call R a pseudo-Gorenstein ring. In this paper we intend to characterize the pseudo-Gorenstein and level Hibi rings K[L] in terms of P . For that purpose we use the basic fact, observed in [11], that a K-basis of K[L] can be described in terms of order reversing functions P̂ → Z≥0 , and that a K-basis of the canonical module ωL of K[L] can be described in terms of strictly order reversing functions P̂ → Z≥0 . Here P̂ = P ∪ {−∞, ∞} with −∞ < x < ∞ for all x ∈ P , see Section 2 for details. Since the property of K[L] to be level, Gorenstein or pseudo-Gorenstein does not depend on the field K, we simply say that L is level, Gorenstein or pseudo-Gorenstein if K[L] has this property. In Section 1 we briefly list conditions which are equivalent to pseudo-Gorenstein and describe the relation of this notion to that of level and Gorenstein. Since pseudoGorenstein rings can be identified by the property that the leading coefficient of the numerator polynomial of the Hilbert series is equal to 1, pseudo-Gorenstein rings are much easier accessible than level rings. Theorem 2.1 gives a full characterization 2010 Mathematics Subject Classification. Primary 05E40, 13C13; Secondary 06D99, 06A11. Key words and phrases. Distributive lattices, Hibi rings, pseudo-Gorenstein, level. The first author was supported by the grant UEFISCDI, PN-II-ID-PCE- 2011-3-1023. The paper was written while the forth author was visiting the Department of Mathematics of University Duisburg-Essen. She wants to express her thanks for its hospitality. 1 of Hibi rings which are pseudo-Gorenstein. Indeed, it is shown that L is pseudoGorenstein if and only if depth(x) + height(x) = rank P̂ for all x ∈ P . This is equivalent to say, that for any given x ∈ P there exists a chain of maximal length in P̂ passing through x. It may be of interest to notice that this property of the chains in P has its analogue in the fact that in an affine domain dim R = dim R/p+height p for all p ∈ Spec R. Though Theorem 2.1 characterizes Hibi rings which are pseudo-Gorenstein, it may nevertheless be difficult to apply this characterization efficiently, even for planar lattices. In Section 3 we introduce hyper-planar lattices which represent a natural extension of planar lattices to higher dimensions. They are defined by the property that their poset P of join-irreducible elements admits a canonical chain decomposition, that is, a decomposition into pairwise disjoint maximal chains. In general such a decomposition is not unique. For hyper-planer lattices we introduce a regularity condition with the effect that the height of an element in P is the same as the height of the element in the chain to which it belongs. Apart from a few exceptions we keep this regularity hypothesis on hyper-planar lattices throughout the rest of the paper. In Theorem 3.3 it is shown that a regular hyper-planar lattice is pseudo-Gorenstein if and only if all chains in a canonical chain decomposition of P have the same length. For simple planar lattices it is shown in Theorem 3.5 that the regularity condition is in fact indispensable. Indeed, it is proved that a simple planar lattice is pseudo-Gorenstein if and only if it is regular and the (two) chains in a canonical chain decomposition of P have the same length. Unfortunately this result cannot be extended to hyper-planar lattices as we show by an example. The study of level Hibi rings is more difficult. There is a very nice sufficient condition on P that guarantees that L is level. In [12, Theorem 3.3] Miyazaki showed that L is level if for all x ∈ P all chains in P̂ ascending from x have the same length, and he showed by an example that this condition is not necessary. Let P ∨ be the dual poset of P , i.e., the poset on the same set as P but with all order relation reversed, and let L∨ be the distributive lattice whose poset of join-irreducible elements is P ∨. Then it is easily seen that L is level if and only if L∨ is level. Therefore it follows from Miyazaki’s theorem (as remarked by him in his paper), that L is also level if for all x ∈ P , all chains in P̂ descending from x have the same length. Thus we call a finite poset P a Miyazaki poset if for all x ∈ P all chains in P̂ ascending from x have the same length or all chains in P̂ descending from x have the same length. Unfortunately L may be level, though P is not a Miyazaki poset, and this may happen even for regular planar lattices. On the other hand, in Theorem 4.1 it is shown that if L is an arbitrary finite distributive lattice which is level, then for all x, y ∈ P with x ⋗ y we must have that height(x) + depth(y) ≤ rank P̂ + 1. Here we use the notation “x ⋗ y” to express that x covers y, that is, x > y and for any z ∈ P either z > x or z < y. At present we do not know whether these inequalities for all covering pairs in P actually characterize the levelness of L. On other hand, it is shown in Theorem 4.3 that a regular planar lattice L is level if and only if these inequalities hold for all covering pairs in P . We apply this result to give in Theorem 4.6 an explicit description of those distributive lattices L whose poset P 2 of join-irreducible elements has a special shape which we call a butterfly. In this particular case it turns out that L is level if and only if the initial ideal in(IL ) defines a level ring where IL is the defining ideal of the Hibi ring K[L]. One may wonder whether the regularity condition in Theorem 3.5 is really needed. In the case of a planar lattice with only one inside corner the regularity hypothesis may indeed be dropped, as shown in Theorem 4.8. In the last section of this paper we study the pseudo-Gorenstein and level property of generalized Hibi rings. For a fixed field K, a poset P and any integer r one defines the so-called generalized Hibi ring Rr (P ) which is naturally attached to rmultichains of poset ideals in P , see [5]. For r = 2 one obtains the ordinary Hibi rings. In Theorem 5.1 it is shown that R2 (P ) is pseudo-Gorenstein if and only if Rr (P ) is pseudo-Gorenstein for some r ≥ 2, and that R2 (P ) is level if Rr (P ) is level for some r ≥ 2. 1. Pseudo-Gorenstein rings Let K be a field and R a Cohen-Macaulay standard graded K-algebra of dimension d with canonical module ωR . We choose a presentation R ∼ = S/I where 2 S = K[x1 , . . . , xn ] is a polynomial ring and I ⊂ m with m = (x1 , . . . , xn ). Furthermore, let F be the graded minimal free resolution of S/I. It is a simple exercise to see that the following conditions are equivalent: (i) Let a = min{i : (ωR )i 6= 0}. Then dimK (ωR )a = 1; (ii) Let y1 , . . . , yd be a maximal regular sequence of linear forms in R and set R̄ = R/(y1, . . . , yd )R. Furthermore, let b = max{i : R̄i 6= 0}. Then dimK R̄b = 1. (iii) Let HR (t) = P (t)/(1 − t)d be the Hilbert series of R. Then the leading coefficient of P (t) is equal to 1. (iv) The highest shift c in the resolution F appears in Fn−d and βn−d,c (S/I) = 1. We call R pseudo-Gorenstein if one (or all) of the above equivalent conditions hold. The ring R is called level if ωR is generated in a single degree. It is clear from (i) that a pseudo-Gorenstein ring is level if and only if it is Gorenstein. Let < be a monomial order and assume that S/ in< (I) is Cohen-Macaulay. It follows from (iv) that S/ in< (I) is pseudo-Gorenstein if and only if S/I is pseudoGorenstein. In particular, if I ⊂ S is a toric ideal such that S/I is Gorenstein and in< (I) is a squarefree monomial ideal, then S/ in< (I) is pseudo-Gorenstein. Here we use that S/ in< (I) is Cohen–Macaulay, if I is a toric ideal and in< (I) is squarefree, see [13, Corollary 6.6.18]. 2. The canonical module of a Hibi ring Let (L, ∧, ∨) be a finite distributive lattice. An element α ∈ L is called joinirreducible if α 6= min L and whenever α = β ∨ γ, then α = β or α = γ. Let P be the subposet of join-irreducible elements of L. By a well-known theorem of Birkhoff [1], one has that L ∼ = I(P ), where I(P ) is the lattice of poset ideals of P with the partial order given by inclusion and with union and intersection as join and meet 3 operation. Poset ideals of P are subsets α of P with the property that if x ∈ α and y ≤ x, then y ∈ α. In particular, ∅ is a poset ideal of P . Given a field K. The Hibi ring of L over K is the K-algebra K[L] generated by the elements α ∈ L and with the defining relations αβ − (α ∧ β)(α ∨ β). Identifying L with I(P ), it is shown in [11] that K[L] is isomorphic to the toric ring generated Q over K by the elements uα with α ∈ L, where uα is the monomial s x∈α tx in the polynomial ring K[s, tx : x ∈ P ]. Let P̂ be the poset P ∪ {∞, −∞} with −∞ < x < ∞ for all x ∈ P . A map v : P̂ → Z≥0 is called order reversing if v(x) ≤ v(y) for all x, y ∈ P̂ with x ≥ y, and v is a called strictly order reversing if v(x) < v(y) for all x, y ∈ P̂ with x > y. We denote by S(P̂ ) the set of all order reversing functions v on P̂ with v(∞) = 0, and by T (P̂ ) the set of all strictly order reversing functions v on P̂ with v(∞) = 0. It is shown in [11] that the toric ring K[L] has a K-basis consisting of the monomials (1) sv(−∞) Y txv(x) , v ∈ S(P̂ ), Y txv(x) , v ∈ T (P̂ ) x∈P and that the monomials (2) sv(−∞) x∈P form a K-basis of the canonical ideal ωL ⊂ K[L]. The (finite) set of elements v ∈ T (P̂ ) which correspond to a minimal set of generators of ωL will be denoted by T0 (P̂ ). It follows from this description of ωL that the property of K[L] to be pseudoGorenstein, Gorenstein or level does not depend on K. Thus we call L itself pseudoGorenstein, Gorenstein or level if K[L] has this property Note that K[L] is standard graded with deg sv(−∞) of the monomials in (1). Q v(x) x∈P tx = v(−∞) for each Before proceeding we recall some basic concepts and notation regarding finite posets. Let Q be an arbitrary poset. A nonempty subposet C of P which is totally ordered is called a chain in P . The length of C is defined to be |C| − 1, and denoted ℓ(C). The rank of Q, denoted rank Q, is defined to be the maximal length of a chain in Q. Let x ∈ Q. Then heightQ (x) (resp. depthQ (x)) is defined to be the maximal length of a chain descending (resp. ascending) from x in Q. In the case that Q = P̂ for some poset P , we omit the lower index and simply write height(x) and depth(x). Let x, y ∈ P . It is said that x covers y, denoted x ⋗ y, if x > y and there exists no z ∈ P such that x > z > y. Let as before L be a finite distributive lattice. Then L is called simple if there exist no elements α, β ∈ L with the property α ⋗ β and such that for each γ ∈ L with γ 6= α, β, we have γ > α or γ < β. Let P be the poset of join-irreducible elements of L. Then L is simple if and only if there exists no element x ∈ P which is comparable with all elements in P . 4 It is observed in [6] that min{v(−∞) : v ∈ T (P̂ )} = rank P̂ . Thus it follows that L is pseudo-Gorenstein if and only if there exists precisely one v ∈ T (P̂ ) with v(−∞) = rank P̂ , and that L is level if and only if for any v ∈ T (P̂ ) there exists v ′ ∈ T (P̂ ) with v ′ (−∞) = rank P̂ and such that v − v ′ ∈ S(P̂ ). In the following result we characterize pseudo-Gorenstein distributive lattices in terms of their poset of join-irreducible elements. Theorem 2.1. The distributive lattice L is pseudo-Gorenstein if and only if depth(x) + height(x) = rank P̂ for all x ∈ P. Proof. Suppose first that depth(x) + height(x) = rank P̂ for all x ∈ P . This implies that for any x ∈ P̂ there exists a chain C of length equal to rank P̂ with x ∈ C. Now let v be any strictly order reversing function on P̂ with v(∞) = 0 and v(−∞) = rank P̂ . Then for any y ∈ C we must have v(y) = depth(y). In particular, v(x) = depth(x). This shows that v is uniquely determined, and proves that L is pseudo-Gorenstein. Conversely, suppose that L is pseudo-Gorenstein. For all x ∈ P̂ we set v(x) = depth(x) and v ′ (x) = rank P̂ − height(x). Then both, v and v ′ , are strictly order reversing functions on P̂ with v(∞) = v ′ (∞) = 0 and v(−∞) = v ′ (−∞) = rank P̂ . Since L is pseudo-Gorenstein we have v = v ′ . This implies that depth(x) + height(x) = rank P̂ for all x ∈ P .  3. Hyper-planar lattices Let L be a finite distributive lattice and P its poset of join-irreducible elements. We call L a hyper-planar lattice, if P as a set is the disjoint union of chains C1 , . . . , Cd , where each Ci is a maximal chain in P . We call such a chain decomposition canonical. Of course in general an element x ∈ Ci may be comparable with an element y ∈ Cj for some j 6= i. If this is the case and if x ⋗ y, then we call the chain x ⋗ y (of length one) a diagonal of P (with respect to the given canonical chain decomposition). For example, the poset depicted in Figure 5 has two diagonals. If d = 2, we recover the simple planar lattices. A canonical chain decomposition of the poset P of join-irreducible elements for a hyper-planar lattice L is in general not uniquely determined. However we claim that if C1 ∪ C2 ∪ · · · ∪ Cs and D1 ∪ D2 ∪ · · · ∪ Dt are canonical chain decompositions of P , then s = t. Indeed, let max(Q) denote the set of maximal elements of a finite poset Q. Then (3) max(P ) = max(C1 ) ∪ max(C2 ) ∪ · · · ∪ max(Cs ) = max(D1 ) ∪ max(D2 ) ∪ · · · ∪ max(Dt ). Let max(Ci ) = {xi } for i = 1, . . . , s and max(Di ) = {yi} for i = 1, . . . , t. Then the elements xi as well as the elements yi are pairwise distinct, and it follows from (3) that {x1 , x2 , . . . , xs } = {y1 , y2 , . . . , yt}, 5 Hence we see that s = t. One would even expect that the (4) {ℓ(C1 ), ℓ(C2 ), . . . , ℓ(Cs )} = {ℓ(D1 ), ℓ(D2 ), . . . , ℓ(Dt )}, as multisets. This however is not the case. For the poset P displayed in Figure 1 we have the following two canonical chain decompositions C1 = a < b < c < d < e < f, C2 = g < h < i < j < k < l, and D1 = a < b < i < e < f, D2 = g < h < c < d < j < k < l. Thus we see that ℓ(C1 ) = ℓ(C2 ) = 5, while ℓ(D1 ) = 4 and ℓ(D2 ) = 6. Also, note that the chain i ⋗ b is a diagonal of P , depicted in Figure 1, with respect to the canonical chain decomposition C1 ∪ C2 , but not is not a diagonal of P with respect to the canonical chain decomposition D1 ∪ D2 . •l f• •k e• •j d• •i c• b• •h a• •g Figure 1. In order to guarantee that equality (4) is satisfied we have to add an extra condition on the hyper-planar lattice: let L be a hyper-planar lattice whose poset of join-irreducible elements is P . In what follows this will be our standard assumption and notation. We say that L is a regular hyper-planar lattice, if for any canonical chain decomposition C1 ∪ C2 ∪ . . . ∪ Cd of P , and for all x < y with x ∈ Ci and y ∈ Cj it follows that heightCi (x) < heightCj (y). Lemma 3.1. Let L be a regular hyper-planar lattice and C1 ∪ . . . ∪ Cd be a canonical chain decomposition of P . Then for all i and x ∈ Ci we have heightCi (x) = heightP (x). Proof. We proceed by induction on heightP (x). If heightP (x) = 0, then there is nothing to show. Now assume that heightP (x) > 0 and let y ∈ P covered by x with heightP (y) = heightP (x) − 1. Say, y ∈ Cj . Since heightP (y) = heightP (x) − 1 we may apply our induction hypothesis and obtain heightP (x) − 1 = heightP (y) = heightCj (y) < heightCi (x) ≤ heightP (x). This yields the desired conclusion.  6 Corollary 3.2. Let L be a regular hyper-planar lattice, and assume that besides of C1 ∪ . . . ∪ Cd there is still another canonical chain decomposition D1 ∪ D2 ∪ · · · ∪ Dd of P . Then (a) {ℓ(C1 ), ℓ(C2), . . . , ℓ(Cd )} = {ℓ(D1 ), ℓ(D2 ), . . . , ℓ(Dd )}, as multisets. (b) rank P = max{ℓ(C1 ), . . . , ℓ(Cd )}. (c) height(x) + depth(x) = rank P̂ for all Ci with ℓ(Ci ) = rank P and all x ∈ Ci . Proof. Let max(Ci ) = {xi } and max(Di ) = {yi } for i = 1, . . . , t. We have seen in the discussion before Lemma 3.1 that {x1 , x2 , . . . , xd } = {y1, y2 , . . . , yd }, Therefore, {heightP (x1 ), heightP (x2 ), . . . , heightP (xd )} = {heightP (y1 ), heightP (y2 ), . . . , heightP (yd )}, as multi-sets. By Lemma 3.1, heightP (xi ) = ℓ(Ci) and heightP (yi ) = ℓ(Di ). This together with the observation that rank P = max{heightP (x1 ), heightP (x2 ), . . . , heightP (xd )} proves (a) and (b). In order to prove (c) we observe that rank P̂ = ℓ(Ĉi ) = heightĈi (x) + depthĈi (x) ≤ height(x) + depth(x) ≤ rank P̂ .  Now we are able to characterize the regular hyper-planar lattices which are pseudoGorenstein. Theorem 3.3. Let L be a regular hyper-planar lattice and C1 ∪. . .∪Cd be a canonical chain decomposition of P . Then L is pseudo-Gorenstein if and only if all Ci have the same length. Proof. Suppose all Ci have the same length. Then Corollary 3.2 implies that ℓ(Ci ) = rank P̂ for all i. Let x ∈ P . Then x ∈ Ci for some i, and hence height(x)+depth(x) = rank P̂ , by Corollary 3.2. Therefore, by Theorem 2.1, L is pseudo-Gorenstein. Conversely, suppose that not all Ci have the same length. Then Corollary 3.2 implies that there exists one Ci with ℓ(Ci ) < rank P . As in the proof of Theorem 2.1 we consider the strictly order reversing function v(x) = depth(x) and v ′ (x) = rank P̂ − height(x). Let x = max(Ci ). Then v(x) = 1 and, since L is regular, v ′ (x) = rank P̂ − (ℓ(Ci ) + 1) > rank P̂ − rank P − 1 = 1. This shows that L is not pseudo-Gorenstein.  In [12] Miyazaki showed that L is level, if for all x ∈ P all maximal chains ascending from p have the same length, or all maximal chains descending from x have the same length. We call a poset P with this property a Miyazaki poset. Corollary 3.4. Let L be a regular hyper-planar lattice and let C1 ∪ C2 ∪ · · · ∪ Cd be a canonical chain decomposition of P . We assume that all Ci have the same length. Then the following conditions are equivalent: 7 (a) L is Gorenstein; (b) L is level; (c) P is a Miyazaki poset. Proof. By Theorem 3.3, L is pseudo-Gorenstein. Thus (a) and (b) are equivalent, as noticed in Section 1. The implication (c) ⇒ (b) follows by Miyazaki [12], and (a) ⇒ (c) follows by Hibi’s theorem [11] which says that L is Gorenstein if and only if P is pure.  For simple planar lattices Theorem 3.3 can be improved as follows. Theorem 3.5. Let L be a simple planar lattice, and let C1 ∪C2 be a canonical chain decomposition of P . Then the following conditions are equivalent: (a) L is pseudo-Gorenstein; (b) L is regular and the chains C1 and C2 have the same length. Proof. It suffices to prove (a) ⇒ (b) because the implication (b) ⇒ (a) is a special case of Theorem 3.3. Now for the proof of the implication (a) ⇒ (b) we use the characterization of pseudo-Gorenstein lattices via the Hilbert series of K[L], as given in (iii), Section 1. Namely, if HK[L] (t) = P (t)/(1 − t)d with d = dim K[L], then L is pseudo-Gorenstein if and only if the leading coefficient of P (t) is one. Note that L may be identified with a two-sided ladder inside an m × n rectangle with n ≤ m as shown in Figure 2. Now according to [7] the leading coefficient of P (t) is the number (m, n) L (0, 0) Figure 2. of maximal cyclic sublattices of L. Thus L is pseudo-Gorenstein if and only if L admits precisely one maximal cyclic sublattice. By a cyclic sublattice of L, we mean a sublattice inside the two-sided ladder in Figure 2 consisting of a chain of squares and edges as in the example shown in Figure 3. We call the cyclic sublattice maximal if it has the maximal number of squares among all cyclic sublattices contained in L. 8 • • • • • • • • • • • • • • • • Figure 3. In the first step of our proof we show by induction on n + m that if L is pseudoGorenstein, then m = n and the lower inside corners of L are below the diagonal connecting (0, 0) with (n, n), while the upper inside corners are above this diagonal. In other words, if (i, j) is a lower inside corner of L, then i ≥ j, while for an upper inside corner (i, j) we have i ≤ j. If this is the case, then we say that the inside corners do not cross the diagonal. If m+n = 2, then there is nothing to prove. Let L′ be the maximal sublattice of L (again viewed as a ladder) with the bottom and top elements (0, 0) and (m−1, n−1), respectively. Now, we consider two cases. Suppose first that the integral points of the square [(m − 2, n − 2), (m − 1, n − 1)] belong to L′ . In this case, L′ is simple since L is simple. We claim that L′ is pseudo-Gorenstein. Indeed, suppose this is not the case. Then L′ has at least two different maximal cyclic sublattices. Then because L is simple, the square [(m − 1, n − 1), (m, n)] is a subset of L. Therefore, each of these maximal sublattices may be extended in L with the square [(m − 1, n − 1), (m, n)], contradicting our assumption that L is pseudo-Gorenstein. Thus, the induction hypothesis implies that m − 1 = n − 1, so that m = n. Also, the induction hypothesis implies that the inside corners of L′ do not cross the diagonal, and hence this is also the case for L. Now, suppose that L′ does not contain the square [(m − 2, n − 2), (m − 1, n − 1)]. Then we may assume that for some j < n − 1 the ladder L contains the squares [(m − 1, k), (m, k + 1)] for all k with j ≤ k ≤ n − 1, but does not contain the squares [(i, k), (i + 1, k)] for all i and j with 1 ≤ i ≤ m − 2 and j ≤ k ≤ n − 1. Then K[L′ ] has the same h-vector as K[L′′ ] where L′′ is a sublattice of L which viewed as a ladder is contained in [(0, 0), (m − 1, j)]. Note that j 6= m − 1, because j < n − 1 ≤ m. Therefore, since L′′ is simple, our induction hypothesis implies that L′′ is not pseudo-Gorenstein. Hence there exist at least two maximal cyclic sublattices in L′′ , and each of these cyclic sublattices may be extended to maximal cyclic sublattices in L, which contradicts the fact that L is pseudo-Gorenstein. Thus this second case is not possible. 9 Now we are ready to prove (b): Let C1 = x1 < x2 < · · · < xm and C2 = y1 < y2 < · · · < yn . Then L viewed as a two-sided ladder contains the points (0, 0) (corresponding to the poset ideal ∅ of L), and (m, n) (corresponding to the poset ideal L). Since L is pseudo-Gorenstein, it follows that m = n, as we have seen before. Being regular is equivalent to the condition that the inside corners of a ladder L do not cross the diagonal connecting (0, 0) and (n, n). In fact, the joinirreducible elements of L establishing the chain C2 can be identified with the vertices of the ladder (as displayed in Figure 2) which are located on the vertical border lines of the upper border and are different from the inside corners and different from (0, 0), while the join-irreducible elements of L forming the chain C1 can be identified with the vertices of the ladder which are located on the horizontal border lines of the lower border of L and are different from the inside corners and different from (0, 0). After this identification let x = (i, j) ∈ C2 be and y = (k, l) ∈ C1 , Then heightC2 (x) = j and heightC1 (y) = k. Assume now that x > y. Then this implies that i ≥ k. Since the inside corners of L do not cross the diagonal we have j > i, and thus heightC2 (x) = j > i ≥ k = heightC1 (y). Similarly, one shows that heightC2 (x) < heightC1 (y), if x < y. This completes the proof.  Theorem 3.5 is not valid if the hyper-planar lattice is not planar, as the example displayed in Figure 4 demonstrates. Indeed, the lattice L corresponding to P is pseudo-Gorenstein, but in this example we only have one canonical chain decomposition, and the chains of this decomposition have different lengths. Moreover, L is not regular. • • • • • • • • Figure 4. We also would like to remark that in Theorem 3.5(b) the condition “regular” is required. Indeed the poset shown in Figure 5 is the poset of join-irreducible elements of a non-regular simple planar lattice L for which L is not pseudo-Gorenstein. 4. Level distributive lattices Throughout this section L will be a finite distributive lattice and P its poset of join-irreducible elements. In the previous section we recalled the fact that L is level if P is a Miyazaki poset. In his paper [12] Miyazaki mentioned the fact that his condition on P is only a sufficient condition. One may ask whether for hyper-planar lattices a stronger result is possible. We begin with a necessary condition for levelness which is valid for any distributive lattice. 10 • • • • • • • • • • Figure 5. Theorem 4.1. Suppose L is level. Then height(x) + depth(y) ≤ rank P̂ + 1 (5) for all x, y ∈ P with x ⋗ y. Proof. Let x, y ∈ P such that x covers y and suppose that height(x) + depth(y) > rank P̂ + 1. We have to show that L is not level. Our assumption implies that height(x) + depth(y) > rank P̂ + 1 ≥ height(x) + depth(x) + 1, and hence depth(y) > depth(x) + 1. We show that there exists w ∈ T0 (P̂ ) with w(−∞) > rank P̂ . This then proves that L is not level. Let depth(y) − depth(x) − 1 = α. Then α > 0. We define v : P̂ → Z≥0 as follows: v(z) = ( depth(z) + α, depth(z), if x ≥ z, z 6= y, otherwise. Then v ∈ T (P̂ ). If v ∈ T0 (P̂ ), then we are done, since v(−∞) = depth(−∞) + α = rank P̂ + α ≥ rank P̂ + 1. The last inequality follows from the fact that α > 0. On the other hand, if v 6∈ T0 (P̂ ), then there exists w ∈ T0 (P̂ ) with v − w ∈ S(P̂ ). It follows that 0 ≤ v(x) − w(x) ≤ v(y) − w(y) = depth(y) − w(y) ≤ 0. Hence w(x) = v(x) = depth(x) + α = depth(x) + depth(y) − depth(x) − 1 = depth(y) − 1. Let x = z0 > z1 > · · · > zk = −∞ 11 be a chain whose length is height(x). Then w(x) < w(z1 ) < · · · < w(zk ) = w(−∞), which implies that w(−∞) ≥ w(x) + height(x) = (depth(y) − 1) + height(x) > rank P̂ .  In his paper [12] Miyazaki remarked that for all z ∈ P all chains ascending from z have the same length if and only if for all x, y ∈ P with x ⋗ y, we have depth(y) = depth(x) + 1. Therefore, P is a Miyazaki poset if and only if depth(y) = depth(x) + 1 for all x, y ∈ P with x ⋗ y, or height(x) = height(y) + 1 for all x, y ∈ P with x ⋗ y. In either case L is level. Corollary 4.2. Suppose L is pseudo-Gorenstein and P satisfies the inequality (5) for all x, y ∈ P with x ⋗ y. Then depth(y) = depth(x) + 1 and height(x) = height(y) + 1 for all x, y ∈ P with x ⋗ y. In particular, L is level and hence Gorenstein. Proof. For all x, y ∈ P with x ⋗ y, we have height(x) + depth(x) = rank P̂ , since L is pseudo-Gorenstein. Thus, by the inequality (5), we have height(x) + depth(y) ≤ height(x) + depth(x) + 1, and hence depth(y) ≤ depth(x) + 1. On the other hand, clearly, we have depth(y) ≥ depth(x) + 1, which implies the first desired formula. The formula regarding height is similarly obtained. Therefore, L is level, because P is a Miyazaki poset.  As mentioned before, if P is not a Miyazaki poset, then L may nevertheless be level, and this may happen even if L is a regular simple planar lattices. Figure 6 shows a poset which is not a Miyazaki poset. However its ideal lattice is a regular simple planar lattice and is level. • • • • • Figure 6. The following result shows that for regular planar lattices the necessary condition for levelness formulated in Theorem 4.1 is also sufficient. Theorem 4.3. Let L be a regular planar lattice. Then the following conditions are equivalent: 12 (a) L is level; (b) height(x) + depth(y) ≤ rank P̂ + 1 for all x, y ∈ P with x ⋗ y; (c) for all x, y ∈ P with x ⋗ y, either depth(y) = depth(x) + 1 or height(x) = height(y) + 1. Remark 4.4. (i) Observe that a Miyazaki poset satisfies condition (c). On the other hand, Figure 6 shows a poset satisfying condition (c) which is not Miyazaki. (ii) Let C1 ∪ C2 be a canonical chain decomposition of P . The inequality in (b) and the equations in (c) are always satisfied for those x ⋗ y for which x and y belong to the same chain in the decomposition. Hence it suffices to check the inequality in (b) and equations in (c) only for diagonals. Indeed, this fact follows directly from Lemma 3.1. For instance, if x covers y and both belong to the same chain, then, by Lemma 3.1, height(x) = height(y) + 1. Thus height(x) + depth(y) = height(y) + 1 + depth(y) ≤ rank P̂ + 1. For (c) we only need to observe that the condition height(x) = height(y) + 1 for any x that covers y in the same chain, is always fulfilled, again by Lemma 3.1. Before proving Theorem 4.3 we will need the following result. Lemma 4.5. Let L be a regular planar lattice. Let C1 ∪ C2 be a canonical chain decomposition of P , and assume that ℓ(C1 ) = rank P (cf. Corollary 3.2). Suppose that P satisfies condition (b) of Theorem 4.3. Then for every v ∈ T0 (P̂ ) we have v(max(C1 )) = 1. Proof. Assume that v(max(C1 )) > 1. Then v(z) ≥ depth(z) + 1 for all z ∈ C1 . Let ( v(x) − 1, if v(x) ≥ depth(x) + 1 (I), ′ v (x) = v(x), if v(x) = depth(x) (II), for all x ∈ P̂ . We show that v ′ ∈ T (P̂ ) and v − v ′ ∈ S(P̂ ). Since v ′ 6= v, this will then show that v 6∈ T0 (P̂ ), a contradiction. Indeed, to see that v ′ ∈ T (P̂ ) we have to show that v ′ (x) < v ′ (y) for all x ⋗ y. If both x and y satisfy (I) or (II), then the assertion is trivial. If x satisfies (I) and y satisfies (II), then v ′ (x) = v(x) −1 < v(y) = v ′ (y), and if x satisfies (II) and y satisfies (I), then v(x) = depth(x) ≤ depth(y) − 1 ≤ v(y) − 2. Hence v(x) < v(y) − 1, and this implies that v ′ (x) < v ′ (y). It remains to be shown that v − v ′ ∈ S(P̂ ) which amounts to prove that v(x) − ′ v (x) ≤ v(y) − v ′ (y) for all x ⋗ y. For this we only need to show that we cannot have v ′ (x) = v(x) − 1 and v(y) = v ′ (y), or, equivalently, that v(x) ≥ depth(x) + 1 and v(y) = depth(y) is impossible. Assume to the contrary that there exist x ⋗ y with v(x) ≥ depth(x) + 1 and v(y) = depth(y) . Then y 6∈ C1 since v(z) ≥ depth(z) + 1 for all z ∈ C1 . Thus, we may either have x ∈ C1 and y ∈ C2 , or x, y ∈ C2 . In the first case, since height(x) + depth(y) ≤ rank P̂ + 1 by assumption, and since rank P̂ = height(x) + depth(x) due to the regularity of L (see Corollary 3.2), we get depth(y) ≤ depth(x) + 1, and hence depth(y) = depth(x) + 1. Therefore, depth(y) = depth(x) + 1 ≤ v(x) < v(y), a contradiction. 13 Finally, let x, y ∈ C2 . Since v(x) < v(y), it follows that depth(y) > depth(x) + 1. Therefore, the longest chain from y to ∞ cannot pass through x. This implies that there exists z ∈ C1 with z ⋗ y. As in the first case we then deduce that v(y) > depth(y). So we get again a contradiction.  Proof of Theorem 4.3. (a) ⇒ (b) follows from Theorem 4.1. (b) ⇒ (c): Let C1 ∪ C2 be a canonical chain decomposition of P with |C1 | ≥ |C2 |. If x, y ∈ C1 or x, y ∈ C2 , then by Lemma 3.1, it follows that height(x) = height(y) + 1. Next suppose that x ∈ C1 . Since L is regular, we may apply Corollary 3.2 and conclude that height(x) + depth(x) = rank P̂ . Thus, by (b), we get depth(y) ≤ depth(x) + 1. On the other hand, it is clear that depth(y) ≥ depth(x) + 1. So that depth(y) = depth(x) + 1. Finally, if y ∈ C1 , then by Corollary 3.2 we have height(y)+depth(y) = rank P̂ . As in the previous case, we conclude that height(x) = height(y) + 1. (c) ⇒ (b): If depth(y) = depth(x) + 1, then height(x) + depth(y) = height(x) + depth(x) + 1 ≤ rank P̂ + 1, and if height(x) = height(y) + 1, then height(x) + depth(y) = height(y) + depth(y) + 1 ≤ rank P̂ + 1. (b) ⇒ (a): As in Lemma 4.5 we let C1 ∪ C2 be a canonical chain decomposition of P , and may assume that ℓ(C1 ) = rank P ≥ ℓ(C2 ). Let v ∈ T0 (P̂ ). We will show that there exists v ′ ∈ T (P̂ ) with v ′ (−∞) = rank P̂ and such that v − v ′ ∈ S(P̂ ). Since v is a minimal generator it follows that v = v ′ , and we are done. In order to construct v ′ we consider the subposet Q of P which is obtained from P by removing the maximal elements max(C1 ) and max(C2 ). We define on Q̂ the strictly order reversing function u by u(∞) = 0, and u(z) = v(z) − 1 for all other z ∈ Q̂. We notice that the ideal lattice of Q is again a regular planar lattice satisfying (b). Indeed, assume that there exist x ⋗ y with x, y ∈ Q such that heightQ̂ (x) + depthQ̂ (y) > rank Q̂ + 1 = rank P̂ . Since heightQ̂ (x) = height(x) and depth(y) = depthQ̂ (y) + 1, it follows that height(x) + depth(y) = heightQ̂ (x) + depthQ̂ (y) + 1 > rank P̂ + 1, a contradiction. Therefore, by induction on the rank we may assume that the ideal lattice of Q is level. Hence there exists w ∈ T (Q̂) with w(−∞) = rank Q̂ = rank P̂ − 1 and such that u − w ∈ S(Q̂). Set v ′ (z) = 1 + w(z) for all z ∈ A = Q ∪ {−∞}. Then v ′ is a strictly order reversing function on A with v ′ (−∞) = rank P̂ and such that v − v ′ is order reversing on A. It remains to define v ′ (Ci ) for i = 1, 2 in a way such that v ′ ∈ T (P̂ ) and v−v ′ ∈ S(P̂ ). We have to set v ′ (max(C1 )) = 1 since v(max(C1 )) = 1, and of course v ′ (∞) = 0. Let x = max(C2 ) and let z ∈ C2 be the unique element with x⋗z. We set v ′ (x) = v(x)−u(z)+w(z) = v(x)−v(z)+1+w(z), and claim that this v ′ has the desired properties. Indeed, v ′ (x) = v(x) − (v(z) − 1 − w(z)) ≤ v(x) and v ′ (x) < 1 + w(z) = v ′ (z), since v(x) < v(z). If z is the only element covered by x, we are done. Otherwise, there exists y ∈ C1 with x ⋗ y and it remains to be shown that v ′ (y) > v ′ (x) = v(x) − v(z) + 1 + w(z). Suppose we know that 14 depthQ̂ (y) ≥ w(z), then v ′ (y) = w(y) + 1 ≥ depthQ̂ (y) + 1 > w(z) ≥ v ′ (x), as desired, since v(x) − v(z) + 1 ≤ 0. Thus in order to complete the proof we have to show that depthQ̂ (y) ≥ w(z). Since the ideal lattice of Q̂ is regular, this is equivalent to showing that (6) w(z) ≤ rank Q̂ − heightQ̂ (y). The assumption (b) and Corollary 3.2(c) imply that height(x) + depth(y) ≤ rank P̂ + 1 = height(y) + depth(y) + 1, so that height(x) ≤ height(y) + 1. This yields (7) height(x) = height(y) + 1 since height(x) ≥ height(y) + 1 is always valid. On the other hand, since L is regular, Lemma 3.1 implies that heightP (x) = heightC2 (x) = heightC2 (z) + 1 = heightP (z) + 1. This implies that height(x) = height(z) + 1. So together with (7) we then conclude that height(y) = height(z). Since heightQ̂ (y) = height(y) and height(z) = heightQ̂ (z), inequality (6) becomes w(z) ≤ rank Q̂ − heightQ̂ (z), and since w(−∞) = rank Q̂, this inequality indeed holds. This completes the proof of the theorem.  In the following theorem we discuss an example of a poset for which the conditions of Theorem 4.1 can be made more explicit. Let P be a finite poset with a canonical chain decomposition C1 ∪C2 with 2 ≤ |C1 | ≤ |C2 |. For i = 1, 2, let xi be the maximal and yi the minimal element of Ci . We call P a butterfly poset (of type (C1 , C2 )), if x1 ⋗ y2 and x2 ⋗ y1 are the only diagonals of P . Figure 6 displays a butterfly poset. Obviously, the ideal lattice of a butterfly poset is regular. For the next result, we need some notation. The Hibi ring K[L] can be presented as the quotient ring T /IL , where T is the polynomial ring over K in the variables xα with α ∈ L and where IL is generated by the binomials xα xβ − xα∨β xα∧β . In the following theorem, we consider a monomial order < given by a height reverse lexicographic monomial order, that is, the reverse lexicographic monomial order induced by a total ordering of the variables satisfying xα < xβ if heightL (α) > heightL (β). Theorem 4.6. Let P be a butterfly poset of type (C1 , C2 ), and L its ideal lattice. Then the following conditions are equivalent: (a) T /in< (IL ) is level; (b) L is level; (c) height(x) + depth(y) ≤ rank P̂ + 1 for all x, y ∈ P with x ⋗ y; (d) for all x, y ∈ P with x ⋗ y, either depth(y) = depth(x) + 1 or height(x) = height(y) + 1; (e) |C1 | = 2. 15 Proof. (a) ⇒ (b) is well-known. Conditions (b), (c) and (d) are equivalent by Theorem 4.3, since L is regular. (c) ⇒ (e): Since 2 ≤ |C1 | ≤ |C2 |, we have height(x1 ) = |C1|, depth(y2 ) = |C2 | and rank P̂ = |C2 | + 1. On the other hand, by condition (b), we have height(x1 ) + depth(y2 ) ≤ rank P̂ + 1. So, |C2 | ≤ 2, and hence |C2 | = 2. (e) ⇒ (a): It is shown in [11] that in< (IL ) is generated by the monomials xα xβ where α and β are incomparable elements of L. Thus, in< (IL ) is the StanleyReisner ideal of the order complex ∆ of L. It is known that ∆ is pure shellable, and hence Cohen-Macaulay, see [2, Theorem 6.1]. Since dim(T / in< (IL )) = dim(T /IL ) = |P | + 1, it follows that pd(T / in< (IL )) = pd(T /IL ) = |L| − |P | − 1. Moreover, since T / in< (IL ) and T /IL are Cohen-Macaulay, their regularity is given by the degree of the numerator polynomial of their Hilbert series. Hence, since both Hilbert series coincide, their regularity is the same and we obtain reg(T /in< (IL )) = reg(T /IL ) = |P | − rank P , by [6]. Since |C1 | = 2, the lattice L viewed as a ladder is of the form as shown in Figure 7. So, we see that |L| = 3|C2| + 1. Therefore, since |P | = |C2 | + 2, we see that pd(T /in< (IL )) = 2|C2| − 2 and reg(T /in< (IL )) = 2. (|C2 |, 2) (0, 0) Figure 7. So to prove (a), it is enough to show that β2|C2 |−2,j (T /in< (IL )) = 0, for all j < 2|C2 |. By Hochster’s formula, we have β2|C2 |−2,j (T /in< (IL )) = X W ⊆V,|W |=j f dimK H j−2|C2 |+1 (∆W ; K), where V is the vertex set of ∆, and where ∆W denotes as usual the subcomplex of ∆ f induced by W . For W ⊆ V with |W | = j < 2|C2 | − 1, dimK H j−2|C2 |+1 (∆W ; K) = 0, since j − 2|C2| + 1 < 0. Let W ⊆ V with |W | = 2|C2 | − 1. We show that f (∆ ; K) = 0 or equivalently ∆ is a connected simplicial complex. As the dimK H 0 W W (1) connectedness of ∆W is equivalent to the connectedness of the 1-skeleton ∆W of (1) ∆W , we show that the graph G = ∆W is connected. The vertices of G correspond to the lattice points of the elements in W which is a certain subset of the lattice points of the ladder displayed in Figure 7. Two vertices of G are adjacent if they are contained in a chain in the lattice L. In other words, two vertices (i, j) and (i′ , j ′) with j ≤ j ′ are adjacent in G if j = j ′ or i ≤ i′ . Let v = (i, j) and w = (i′ , j ′) be two nonadjacent vertices in G, so that we may assume j < j ′ and i′ < i. We show that there exists a path in G connecting v and w. 16 Case 1. There exists (s, t) ∈ W such that either s ≤ i′ , t ≤ j or s ≥ i, t ≥ j ′ . Then (s, t) is adjacent to both of v and w, and hence v and w are connected. Case 2. There exists s with i′ < s < i such that (s, j), (s, j ′) ∈ W . Then we get a path in G between v and w passing through (s, j) and (s, j ′ ). Assume that W does not satisfy the conditions given in Case 1 and Case 2. Let A = {(s, t) ∈ W : i′ < s < i, t = j or t = j ′ }. Then |A| ≤ i − i′ − 1. First suppose that j = 0 and j ′ = 1. Also, set A1 = {(s, 0) : i ≤ s ≤ |C2| − 1}, A2 = {(s, 1) : 0 ≤ s ≤ i′ } and A3 = {(s, 2) : 1 ≤ s < i}. Since we are not in Case 1 and Case 2, it follows that W ⊂ A1 ∪ A2 ∪ A3 ∪ A. Therefore, |W | ≤ |A1 | + |A2 | + |A3 | + |A| ≤ (|C2 | − i) + (i′ + 1) + (i − 1) + (i − i′ − 1) = |C2 | + i − 1 < 2|C2| − 1. The last inequality follows since i < |C2 |. This contradicts the fact that |W | = 2|C2 | − 1. Now suppose that j = 0 and j ′ = 2 and set A1 = {(s, 0) : i ≤ s ≤ |C2 | − 1}, A2 = {(s, 1) : 0 ≤ s ≤ |C2 |}, A3 = {(s, 2) : 1 ≤ s ≤ i′ }. We claim that |A2 ∩ W | ≤ |A2 | − 2, unless there is a path in G connecting v and w. Indeed, we may assume that either (i, 1) or (i′ , 1) belongs to W . If both vertices belong to W , then there is a path in G between v and w. Without loss of generality, we may assume that (i, 1) ∈ W and (i′ , 1) ∈ / W . In that case, non of the elements in the nonempty set {(s, 1) : s < i′ } belongs to W , because otherwise we get a path connecting v and w in G. This proves the claim. Since we are not in Case 1 and Case 2, we have W ⊂ A1 ∪ (A2 ∩ W ) ∪ A3 ∪ A. As we may assume that |A2 ∩ W | ≤ |A2 | − 2, we get |W | ≤ |A1 | + |A2 ∩ W | + |A3 | + |A| ≤ (|C2 | − i) + (|C2 | − 1) + (i′ ) + (i − i′ − 1) = 2|C2 | − 2, a contradiction. Finally, j = 1 and j ′ = 2 is similarly treated as the case j = 0 and j ′ = 1.  As a straightforward consequence of the next result it can be seen that the implication (a) ⇒ (c) in Theorem 4.3 is in general not valid for non-planar lattices, not even for hyper-planar lattices. Theorem 4.7. Let P = P1 ∪ P2 be a finite poset with the property that the elements of P1 and P2 are incomparable to each other, and suppose that P2 is a chain of length r. Let L be the ideal lattice of P . Then L is level for all r ≫ 0. Proof. Let L1 by the ideal lattice of P1 , and L2 that of P2 . It is observed in [10] and easy to see that K[L] ∼ = K[L1 ] ∗ K[L2 ] (which is the Segre product of K[L1 ] and K[L2 ]). By [8, Theorem 4.3.1] we have ωL ∼ = ωL1 ∗ ωL2 , where for graded K[L1 ]module M and a graded K[L2 ]-module N the homogeneous components of the Segre product M ∗ N are given by (M ∗ N)i = Mi ⊗K Ni for all i. Now if L2 is a chain of length r, then S = K[L2 ] is a polynomial ring of dimension r + 2, and hence ωL2 ∼ = S(−r − 2), see for example [3, Proposition 3.6.11 and Example 3.6.10]. Hence 17 we obtain (8) ωL ∼ = ωL1 ∗ S(−r − 2) ∼ = (ωL1 )≥r+2 , L where for a graded module M and any integer s we set M≥s = i≥s Mi . Let t be the highest degree of a generator in a minimal set of generators of ωL1 . Then (8) implies that ωL1 is generated in the single degree r + 2 if r + 2 ≥ t. Thus for any r ≥ t − 2 we see that L is level.  We would like to mention that the arguments given in the proof of Theorem 4.7 yield the following slightly more general result: for an arbitrary finite poset P we set γ(P ) = max{v(−∞) : v ∈ T0 (P̂ )}. Note that γ(P ) is the highest degree of a generator in a minimal set of generators of the canonical module of the ideal lattice of P . Now let P = P1 ∪ P2 and suppose that the elements of P1 and P2 are incomparable. Furthermore, assume that the ideal lattice L2 of P2 is level. Then the ideal lattice L of P is level if γ(P2 ) ≥ γ(P1 ). Computational evidence leads us to conjecture that the equivalent conditions given in Theorem 4.3 do hold for any planar lattice (without any regularity assumption). In support of this conjecture we have the following result. Theorem 4.8. Let L be a simple planar lattice whose poset P of join-irreducible elements has the single diagonal x⋗y with respect to a canonical chain decomposition. Then the following conditions are equivalent: (a) L is level; (b) height(x) + depth(y) ≤ rank P̂ + 1; (c) either depth(y) = depth(x) + 1 or height(x) = height(y) + 1. Observe that for all covering pairs u, v ∈ P which are different from the diagonal, conditions (b) and (c) are always satisfied. Proof of Theorem 4.8. (a) ⇒ (b) follows from Theorem 4.1. (b) ⇒ (c): Let C1 ∪ C2 be a canonical chain decomposition of P and x ⋗ y its unique diagonal with x ∈ C1 and y ∈ C2 . We define the integers a = |{z ∈ P̂ : z > x}|, b = |{z ∈ P̂ : z < x, z ∈ / C2 }|, c = |{z ∈ P̂ : z > y, z ∈ / C1 }| and d = |{z ∈ P̂ : z < y}|, see Figure 8. Then height(y) = d and depth(x) = a. On the other hand, depth(y) = max{c, a + 1} and height(x) = max{b, d + 1}. Suppose that none of the equalities in condition (c) hold. Then depth(y) > depth(x) + 1 and height(x) > height(y) + 1. So, max{c, a + 1} > a + 1, and hence c > a + 1 and depth(y) = c. Similarly, max{b, d + 1} > d + 1, and hence b > d + 1 and height(x) = b. Note that P̂ has just three maximal chains whose lengths are a + b, c + d and a + d + 1, so that rank P̂ = max{a + b, c + d}. Thus (b) implies b + c ≤ max{a + b + 1, c + d + 1}, a contradiction. Therefore one of the desired equalities in (c) hold for x, y. (c) ⇒ (a): Let C1 : −∞ = x′b < x′b−1 < · · · < x′1 < x < x1 < · · · < xa−1 < xa = ∞, ′ C2 : −∞ = yd′ < yd−1 < · · · < y1′ < y < y1 < · · · < yc−1 < yc = ∞. 18 ∞ • • • a c •x y• b • d • • −∞ Figure 8. The condition (c) guarantees that either a + 1 ≥ c or b ≤ d + 1. We may assume that a + 1 ≥ c. The case b ≤ d + 1 is treated similarly by replacing P by P ∨. Case 1. Suppose that a longest chain of P̂ is ′ −∞ = yd′ < yd−1 < · · · < y1′ < y < x < x1 < · · · < xa−1 < xa = ∞. In other words, one has b ≤ d + 1. The rank of P̂ = a + d + 1. Let v ∈ T (P̂ ) with v(−∞) > a + d + 1. We distinguish several cases: (i) Let v(x) > a. Then there are i and j with v(xi ) − v(xi+1 ) ≥ 2 and v(yj ) − v(yj+1) ≥ 2, where x0 = x and y0 = y. Let u ∈ S(P̂ ) with u(z) = 1 if either z ≤ xi or z ≤ yj and u(z) = 0, otherwise. Then v − u ∈ T (P̂ ). (ii) Let v(x) = a. In this case v(y) ≥ a + 1. ′ If v(y) = a+1, then there are i and j with v(x′i+1 )−v(x′i ) ≥ 2 and v(yj+1 )−v(yj′ ) ≥ 2, where x′0 = x and y0′ = y. Let u ∈ S(P̂ ) with with u(z) = 1 if either z ≤ x′i+1 or ′ z ≤ yj+1 and u(z) = 0, otherwise. Then v − u ∈ T (P̂ ). If v(y) > a+1, then there are i and j with v(yj )−v(yj+1) ≥ 2 and v(x′i+1 )−v(x′i ) ≥ 2, where x′0 = x and y0 = y. Let u ∈ S(P̂ ) with u(z) = 1 if either z ≤ x′i+1 or z ≤ yj and u(z) = 0, otherwise. Then v − u ∈ T (P̂ ). Case 2. Suppose that a longest chain of P̂ is −∞ = x′b < x′b−1 < · · · < x′1 < x < x1 < · · · < xa−1 < xa = ∞. In other words, one has b ≥ d + 1. The rank of P̂ = a + b. Let v ∈ T (P̂ ) with v(−∞) > a + b. We distinguish several cases: (i) Let v(x) > a. Then there are i and j with v(xi ) − v(xi+1 ) ≥ 2 and v(yj ) − v(yj+1) ≥ 2, where x0 = x and y0 = y. Let u ∈ S(P̂ ) with u(z) = 1 if either z ≤ xi or z ≤ yj and u(z) = 0, otherwise. Then v − u ∈ T (P̂ ). 19 (ii) Let v(x) = a. Then there is i with v(x′i+1 ) − v(x′i ) ≥ 2, where x′0 = x. Since ′ b ≥ d + 1, one has either v(y) − v(x) ≥ 2 or there is j with v(yj+1 ) − v(yj′ ) ≥ 2, where y0′ = y. If v(y) − v(x) ≥ 2, then there is j with v(yj ) − v(yj+1) ≥ 2, where y0 = y. Let u ∈ S(P̂ ) with u(z) = 1 if either z ≤ x′i+1 or z ≤ yj and u(z) = 0 otherwise. Then v − u ∈ T (P̂ ). ′ Suppose that there is j with v(yj+1 )−v(yj′ ) ≥ 2, where y0′ = y. Let u ∈ S(P̂ ) with ′ and u(z) = 0, otherwise. Then v − u ∈ T (P̂ ). u(z) = 1 if either z ≤ x′i+1 or z ≤ yj+1 The discussions in both cases show that every v ∈ T (P̂ ) which belongs to T0 (P̂ ) satisfies v(−∞) = rank(P̂ ). Hence L is level, as desired.  s m t n Figure 9. Corollary 4.9. Let I be the ladder determinantal ideal of 2-minors of the ladder as displayed in Figure 9. Then S/I is level if and only if min{m, n} ≤ s + t. Proof. First note that the ideal lattice of the poset with one diagonal depicted in Figure 8 can be identified with the one-sided ladder as shown in Figure 9 with n = c + d − 1, m = a + b − 1, s = a and t = d. Thus, by Theorem 4.8, S/I is level if and only if condition (c) of Theorem 4.8 holds. According to Figure 8, this is equivalent to say that b ≤ d + 1 or c ≤ a + 1, or equivalently m ≤ s + t or n ≤ s + t, which implies the assertion.  5. Generalized Hibi rings Let P = {x1 , . . . , xn } be a finite poset and r a positive integer. In Section 2 we identified the Hibi ring K[L] as a subring of the polynomial ring K[s, tx : x ∈ P ]. There is a different natural embedding of K[L] into a polynomial ring, namely into the polynomial ring K[sx , tx : x ∈ P ] where P is the set of join-irreducible elements of L, see [9]. For this embedding the generators of K[L] are the monomials Q Q uα = x∈α tx x∈P \α sx . This suggests an extension of the notion of Hibi rings as introduced in [5], see also [4]. 20 An r-multichain of I(P ) is a chain of poset ideals of length r, I : ∅ = α0 ⊆ α1 ⊆ α2 ⊆ · · · ⊆ αr = P. We define a partial order on the set Ir (P ) of all r-multichains of I(P ) by setting I ≤ I ′ if αk ⊆ αk′ for k = 1, . . . , r. Observe that the partially ordered set Ir (P ) is a distributive lattice, with meet and join defined as follows: for I : α0 ⊆ α1 ⊆ · · · ⊆ αr and I ′ : α0′ ⊆ α1′ ⊆ · · · ⊆ αr′ we let I ∧ I ′ be the multichain with (I ∧ I ′ )k = αk ∩ αk′ for k = 1, . . . , r, and I ∨ I ′ the multichain with (I ∨ I ′ )k = αk ∪ αk′ for k = 1, . . . , r. With each r-multichain I of Ir (P ) we associate a monomial uI in the polynomial ring K[xij : 1 ≤ i ≤ r, 1 ≤ j ≤ n] in rn indeterminates which is defined as Q uI = x1γ1 x2γ2 · · · xrγr , where xkγk = xℓ ∈γk xkℓ and γk = αk \ αk−1 for k = 1, . . . , r. The generalized Hibi ring Rr (P ) is the toric ring generated by the monomials uI with I ∈ Ir (P ). By what we said at the beginning of this section it is clear that K[L] = R2 (P ). In [5, Theorem 4.1] is shown that Rr (P ) can be identified with the ordinary Hibi ring K[Lr ] where Lr is the ideal lattice of Ir (P ), and further it is shown in [5, Theorem 4.3] that Ir (P ) is isomorphic to the poset Pr = P × Qr−1 where Qr−1 = [r − 1] with the natural order of its elements, and where P × Qr−1 denotes the direct product of the posets P and Qr−1 . In general the direct product P × Q of two posets P and Q is defined to be the poset which as a set is just the cartesian product of two sets P and Q and with partial order given by (x, y) ≤ (x′ , y ′) if and only if x ≤ x′ and y ≤ y ′. This identification of Rr (P ) with K[Lr ] was used in [5] to prove that for any given r ≥ 2, the Hibi ring K[L] is Gorenstein if and only if K[Lr ] is Gorenstein. We denote by type(R) the Cohen–Macaulay type of a Cohen–Macaulay ring R. It is defined to be the number of generators of ωR . Here we show Theorem 5.1. Let L be a finite distributive lattice, and r ≥ 2 an integer. Then (a) type(K[L]) ≤ type(K[Lr ]); (b) L is pseudo-Gorenstein if and only if Lr is pseudo-Gorenstein; (c) If Lr is level then L is level. Proof. (a) We need to show that |T0 (P̂ )| ≤ |T0 (P̂r )|. In order to prove this we define an injective map ι : T0 (P̂ ) → T0 (P̂r ). Given v ∈ T0 (P̂ ) we let ι(v)(x, i) = v(x) + (r − 1 − i) and ι(v)(∞) = 0 and ι(v)(−∞) = v(−∞) + (r − 2). Obviously, ι(v) 6= ι(w) for v, w ∈ T0 (P̂ ) with v 6= w and ι(v) ∈ T (P̂r ). Thus it remains to show that ι(v) actually belongs to T0 (Pˆr ). We set v ′ = ι(v), and show that if v ′ −u ∈ S(P̂r ) for some u ∈ T (P̂r ), then v ′ = u. For any w ∈ T (P̂r ) and for i ∈ [r − 1] we define the function wi on P̂ as follows: wi (x) = w(x, i) − (r − 1 − i) for all x ∈ P, wi (∞) = 0 and wi (−∞) = max{wi (x) : x ∈ P } + 1. Then wi ∈ T (P̂ ). Since v ′ − u ∈ S(P̂r ) it follows that v − ui = vi′ − ui ∈ S(P̂ ). Since v ∈ T0 (P̂ ) we see that v = ui for all i. This shows that v ′ = u. 21 (b) Let x ∈ P and i ∈ [r − 1]. We claim that heightP̂r (x, i) = heightP̂ (x) + (i − 1) and depthP̂r (x, i) = depthP̂ (x)+(r−i−1). If heightP̂r (x, i) = 1, then there is nothing to prove. Let heightP̂r (x, i) > 1 and let x = x0 > x1 > · · · > xd > −∞ be a maximal chain of length heightP̂ (x) in P̂ . Then (x, i) = (x0 , i) > (x1 , i) > · · · > (xd , i) > (xd , i − 1) > · · · > (xd , 1) > −∞ is a maximal chain of length heightP̂ (x) + (i − 1) in P̂r . It follows that heightP̂r (x, i) ≥ heightP̂ (x)+(i−1). To prove the other inequality we use induction on height. Let (x, i) = z0 > z1 > · · · > zt > −∞ be a maximal chain of length heightP̂r (x, i) in P̂r . Then z1 = (x, i − 1) or z1 = (x′ , i) where x′ is an element of P covered by x. If z1 = (x, i − 1), then by induction hypothesis, we get heightP̂r (x, i−1) ≤ heightP̂ (x)+(i−2), and hence heightP̂r (x, i) = heightP̂r (x, i−1)+ 1 ≤ heightP̂ (x) + (i − 1). If z1 = (x′ , i), then our induction hypothesis implies that heightP̂r (x′ , i) ≤ heightP̂ (x′ )+(i−1), and hence heightP̂r (x, i) = heightP̂r (x′ , i)+1 ≤ heightP̂ (x)+(i−1), where the last inequality follows from the fact that heightP̂ (x′ ) ≤ heightP̂ (x) − 1. So, heightP̂r (x, i) = heightP̂ (x) + (i − 1). A similar argument can be applied to prove the claimed formula regarding the depth. As a side result, we obtain that rank P̂r = rank P̂ + (r − 2). It follows that heightP̂ (x) + depthP̂ (x) = rank P̂ if and only if heightP̂r (x, i) + depthP̂r (x, i) = rank P̂r . Thus Theorem 2.1 yields the desired result. (c) Suppose that L is not level. Then there exists v ∈ T0 (P̂ ) with v(−∞) > rank P̂ . Then ι(v), as defined in the proof of part (a), belongs to T0 (P̂r ) and ι(v)(−∞) = v(−∞) + (r − 2) > rank P̂ + (r − 2) = rank P̂r . This shows that Lr is not level.  Note that if type(K[L]) = 1, then type(K[Lr ]) = 1 for all r ≥ 2, since by [5, Corollary 4.5], K[L] is Gorenstein if and only if K[Lr ] is Gorenstein. But the following example given in Figure 10 shows that the inequality in Theorem 5.1 (a) may be strict. Indeed, let P be the poset depicted in the left side of Figure 10. Then P3 is the poset which is shown in right side of Figure 10. It can be easily checked, by considering all possible strictly order reversing functions on P̂ and P̂3 which correspond to the minimal generators of the canonical module, that type(K[L]) = 2, but type(K[L3 ]) = 3. • • • • • • • • • Figure 10. In Theorem 5.1(c) we only could prove that if Lr is level, then L is level. However we expect that the other implication also holds. An indication that this might be true is Miyazaki’s theorem and the easy to prove fact that a poset P is Miyazaki if 22 and only if Pr is Miyazaki. Moreover, if the necessary condition for levelness given in Theorem 4.1 would also be sufficient, which indeed we expect, then one could also conclude that L is level if and only if Lr is level, since P̂ satisfies the inequalities (5) if and only if this is the case for P̂r . References [1] G. Birkhoff, Lattice Theory (3rd ed.), Amer. Math. Soc. Colloq. Publ. No. 25. Providence, R. I.: Amer. Math. Soc. [2] A. Björner, Shellable and Cohen-Macaulay partially ordered sets, Trans. Amer. Math. Soc. 260 (1980), 159–183. [3] W. Bruns, J. Herzog, Cohen-Macaulay rings, Revised Ed., Cambridge University Press, (1998). [4] V. Ene, J. Herzog, Gröbner bases in commutative algebra, Graduate Studies in Mathematics 130, AMS, (2012). [5] V. Ene, J. Herzog, F. Mohammadi, Monomial ideals and toric rings of Hibi type arising from a finite poset, Eur. J. Comb. 32 (2011), 404-421. [6] V. Ene, J. Herzog, S. Saeedi Madani, A note on the regularity of Hibi rings, preprint, arXiv: 1404.2554v2. [7] V. Ene, A. A. Qureshi, A. Rauf, Regularity of join-meet ideals of distributive lattices, Electron. J. Combin. 20 (3) (2013), #P20. [8] S. Goto, K.-i. Watanabe, On graded rings, I, J. Math. Soc. Japan 30(2) (1978), 179–213. [9] J. Herzog, T. Hibi, Distributive lattices, bipartite graphs and Alexander duality, J. Algebraic Combin. 22 (2005), 289-302. [10] J. Herzog, T. Hibi, G. Restuccia, Strongly Koszul algebras, Math. Scand. 86(2) (2000), 161– 178. [11] T. Hibi, Distributive lattices, affine semigroup rings and algebras with straightening laws, In: “Commutative Algebra and Combinatorics” (M. Nagata and H. Matsumura, Eds.), Adv. Stud. Pure Math. 11, North–Holland, Amsterdam, (1987), 93–109. [12] M. Miyazaki, A sufficient condition for a Hibi ring to be level and levelness of Schubert cycles, Comm. Alg. 35 (2007), 2894–2900. [13] R. Villarreal, Combinatorial optimization methods in commutative algebra, to appear. Viviana Ene, Faculty of Mathematics and Computer Science, Ovidius University, Bd. Mamaia 124, 900527 Constanta, Romania, and Simion Stoilow Institute of Mathematics of the Romanian Academy, Research group of the project ID-PCE-2011-1023, P.O.Box 1-764, Bucharest 014700, Romania E-mail address: [email protected] Jürgen Herzog, Fachbereich Mathematik, Universität Duisburg-Essen, Campus Essen, 45117 Essen, Germany E-mail address: [email protected] Takayuki Hibi, Department of Pure and Applied Mathematics, Graduate School of Information Science and Technology, Osaka University, Toyonaka, Osaka 5600043, Japan E-mail address: [email protected] Sara Saeedi Madani, Department of Pure Mathematics, Faculty of Mathematics and Computer Science, Amirkabir University of Technology (Tehran Polytechnic), 424, Hafez Ave., Tehran 15914, Iran, and School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P.O. Box 19395-5746, Tehran, Iran E-mail address: [email protected] 23
0math.AC
Towards Advanced Phenotypic Mutations in Cartesian Genetic Programming arXiv:1803.06127v1 [cs.NE] 16 Mar 2018 Roman Kalkreuth TU Dortmund University, Department of Computer Science, Dortmund, Germany [email protected] Abstract. Cartesian Genetic Programming is often used with a point mutation as the sole genetic operator. In this paper, we propose two phenotypic mutation techniques and take a step towards advanced phenotypic mutations in Cartesian Genetic Programming. The functionality of the proposed mutations is inspired by biological evolution which mutates DNA sequences by inserting and deleting nucleotides. Experiments with symbolic regression and boolean functions problems show a better search performance when the proposed mutations are in use. The results of our experiments indicate that the use of phenotypic mutations could be beneficial for the use of Cartesian Genetic Programming. Keywords: Cartesian Genetic Programming · Mutation · Phenotype 1 Introduction Genetic Programming (GP), as popularized by Koza [5,6,7], uses syntax trees as program representation. Cartesian Genetic Programming (CGP) as introduced by Miller et al. [10] offers a novel graph-based representation which in addition to standard GP problem domains, makes it easy to be applied to many graph-based applications such as electronic circuits, image processing, and neural networks. CGP is often used only with a point mutation as genetic operator. In contrast to tree-based GP for which advanced methods of crossover and mutation have been introduced and investigated, the use of advanced mutation techniques in CGP appears to be mostly unexplored. This significant lack of knowledge in CGP has been the major motivation for our work. Another motivation for our work has been the introduction of a phenotypic subgraph crossover technique for CGP by Kalkreuth et al. [4]. The experiments of Kalkreuth et al. show that the use of the subgraph crossover technique can be beneficial for the search performance of CGP. In standard tree-based GP, the simultaneous use of multiple types of mutation has been found beneficial by Kraft et al. [8] and Angeline et al. [1]. To our best knowledge, this principle has not been investigated in CGP yet. In this paper, we propose two phenotypic mutations for CGP and take a step towards advanced phenotypic mutations in CGP. Furthermore, we present a first empirical initial testing of both techniques. Section 2 of this paper describes CGP briefly and surveys previous work on phenotypic mutations in CGP. In section 3 we propose our new mutation techniques. Section 4 is devoted to the experimental results and the description of our experiments. In Section 5 we discuss the results of our experiments. Finally, section 6 gives a conclusion and outlines future work. 2 2.1 Related Work Cartesian Genetic Programming Cartesian Genetic Programming is a form of Genetic Programming which offers a novel graph-based representation. In contrast to tree-based GP, CGP represents a genetic program via genotype-phenotype mapping as an indexed, acyclic and directed graph. Originally the structure of the graphs was a rectangular grid of Nr rows and Nc columns, but later work also focused on a representation with at least one row. The genes in the genotype are grouped, and each group refers to a node of the graph, except the last one which represents the outputs of the phenotype. Each node is represented by two types of genes which index the function number in the GP function set and the node inputs. These nodes are called function nodes and execute functions on the input values. The number of input genes depends on the maximum arity Na of the function set. The last group in the genotype represents the indexes of the nodes which lead to the outputs. A backward search is used to decode the corresponding phenotype. The backward search starts from the outputs and processes the linked nodes in the genotype. In this way, only active nodes are processed during the evaluation. The number of inputs Ni , outputs No and the length of the genotype is fixed. Every candidate program is represented with Nr ∗ Nc ∗ (Na + 1) + No integers. Even when the length of the genotype is fixed for every candidate program, the length of the corresponding phenotype in CGP is variable which can be considered as a significant advantage of the CGP representation. Figure 2 exemplifies the CGP decoding procedure. CGP traditionally operates within a (1+λ) evolutionary algorithm (EA) in which λ is often chosen with a size of four. The new population in each generation consists of the best individual of the previous population and the λ created offspring. The breeding procedure is mostly done by a point mutation which swaps genes in the genotype of an individual in their permissible range by chance. Algorithm 1 exemplifies the functioning of the standard (1+λ)-EA. One of the most important technique is a special rule for the selection of the new parent. In the case when two or more individuals can serve as the parent, an individual which has not served as the parent in the previous generation will be selected as a new parent. This strategy is important because it ensures the diversity of the population and has been found highly beneficial for the search performance of CGP. Algorithm 1 Standard (1+λ)-CGP algorithm 1: procedure (1+λ)-CGP 2: Initialize(P) . Initialize parent individual 3: Evaluate(P) . Evaluate the fitness of the parent individual 4: Check(P) . Check if the parent individual meets the target fitness 5: while true do . Until termination criteria not triggered 6: Q ← breed(P) . Breed λ offsprings by mutation 7: Evaluate(Q) . Evaluate the fitness of the offsprings 8: if any individual of Q meets the target fitness then 9: return best individual of Q 10: end if 11: if any individual of Q has better fitness then P then 12: Replace P by the offspring with the best fitness 13: end if 14: end while 15: end procedure Fig. 1: Decoding procedure in Cartesian Genetic Programming 2.2 Advanced Mutation Techniques in CGP For an investigation of the length bias and the search limitation of CGP, a variation of the standard point mutation technique has been introduced by Goldman et al. [3]. The modified point mutation mutates exactly one active gene for all offspring. This called single active-gene mutation strategy (SAGMS) has been found beneficial for the search performance of CGP. The SAGMS can be seen as a form of phenotypic genetic operator since it respects only active function genes in the genotype which are part of the corresponding phenotype. Later work by Pedroni [11] utilized SAGMS to propose an explicitly neutral mutation operator which allows the user to directly control the mutation rate of inactive genes. Pedroni found that the proposed mutation is less sensitive to mutation rate and to reach perfect solutions more consistently than the standard mutation operator. 3 The Proposed Methods The proposed phenotypic mutations for CGP are inspired by biological evolution in which extra base pairs are inserted into a new place in the DNA sequence. This type of mutation is called insertion. Another mutation called deletion removes sections of the DNA sequence. Related to CGP, we adopt this principle by activating and deactivating randomly chosen function nodes. If a genome is selected for the insertion mutation, one inactive function node becomes active. In the case that all function nodes are active, the individual remains unchanged. Contrary, when deletion is performed, one active node becomes inactive. If the option that outputs can directly connect to the input nodes is disabled, a minimum number of active function nodes has to be defined. The activation and deactivation of the nodes is done by adjusting the connection genes of neighborhood nodes. Both mutation techniques work similarly to the single active-gene mutation strategy. The state of exactly one function node of an individual is changed. Since these forms of mutation can elicit strong changes in the behavior of the individuals, we apply an insertion rate and a deletion rate for each offspring which is selected for mutation. Fig. 2 and Fig. 3 exemplify the insertion and the deletion mutation techniques. As visible, the connection genes in the genotype are adjusted to deactivate or activate a particular function node in the phenotype. Fig. 2: The proposed insertion mutation technique Fig. 3: The proposed deletion mutation technique 4 4.1 Experiments Experimental Setup We performed experiments on symbolic regression and boolean functions. To evaluate the search performance of the mutations, we measured the number of generations until the CGP algorithm terminated successfully (generations-tosuccess) and the best fitness value which was found after a predefined number of generations (best-fitness-of-run). We used the standard (1+λ)-CGP algorithm with a population size of 5 individuals ( (1 + 4)-CGP) for our experiments. To classify the significance of our results, we used the Mann-Whitney-U-Test. The mean values are denoted a † if the p-value is less than the significance level 0.05 and a ‡ if the p-value is less than the significance level 0.01 compared to the use of mutation as the sole genetic operator. We performed 100 independent runs with different random seeds. For an empirical initial testing of the proposed mutation techniques we defined a range of the mutation ranges from 0% to 30% inclusive. The setting of important CGP parameters such as point mutation rate has been empirically determined and is oriented with former work. We used the insertion and deletion mutation technique in combination with the standard point mutation technique of CGP. The termination criteria are explained in the particular experiment. For our experiments, we performed a grid analysis of a set of popular benchmark problems which have been used in former GP and CGP experiments. We chose three well known boolean and symbolic regression problems. We analyzed a 4x 4 grid for each problem. The mutations rates for the insertion and deletion mutation techniques are denoted in 10% steps. For instance, a mutation rate of 10% is denoted as 0.1. We used a reducing fitness for our experiments. The minimum number of function nodes in the phenotype was set to a number of 4. 4.2 Boolean Functions For our first experiment in the boolean domain, we chose multiple boolean output problems as the digital adder, subtractor, and multiplier.Former work by White et al. [12] proposed these sort of problems as suitable alternatives to the overused parity problems. To evaluate the fitness of the individuals on the multiple output problems, we used compressed truth tables. We defined the fitness value of an individual as the number of different bits. When this number became zero, the algorithm terminated successfully. The configurations for the two-bit boolean functions are shown in Table 1. In this problem domain, we evaluated the problems with (generations-to-success). Table 2, 3 and 4 show the results of the grid analysis for the boolean problems. It is clearly visible that the use of the insertion and deletion mutation technique significantly reduces the number of generations until the termination criterion triggered. Table 1: Configuration of the boolean functions problems Property Add. 2 Bit Mul. 2 Bit Sub. 2 Bit 30 3 2 30 4 4 30 4 3 AND, OR XOR, AND? AND, OR XOR, AND? AND, OR, XOR NOR, AND? Point mutation rate 0.05 ? AND with one inverted input 0.05 0.05 Node count Number of inputs Number of outputs Function set 0.0 Deletion Rate 0.1 0.2 0.3 0.0 86494 49710‡ 36302‡ 46613‡ Insertion Rate 0.1 0.2 ‡ 54474 57109‡ 49894‡ 49767‡ 42709‡ 53105‡ 52043‡ 49382‡ 0.3 55678‡ 47868‡ 47358‡ 57263‡ Table 2: Mean number of generations for various settings of insertion and deletion mutation for the 2-Bit digital adder problem 0.0 Deletion Rate 0.1 0.2 0.3 0.0 17497 14682 18196 17978 Insertion Rate 0.1 0.2 ‡ 10995 12290† 12873† 15050 14115† 13440† 24898 15954 0.3 12260† 14761 15396 12414 Table 3: Mean number of generations for various settings of insertion and deletion mutation for the 2-Bit multiplier problem 0.0 Deletion Rate 0.1 0.2 0.3 0.0 14809 14327 14277 15436 Insertion Rate 0.1 0.2 10607‡ 10864‡ 11166† 12079† 9515‡ 11638† 8335‡ 12566† 0.3 11157‡ 13661 12562 9734‡ Table 4: Mean number of generations for various settings of insertion and deletion mutation for the 2-Bit subtractor problem 4.3 Symbolic Regression For symbolic regression, we have chosen three problems from the work of Clegg et al. [2] and McDermott et al. [9] for better GP benchmarks, and the Pagie-1 problem which has been proposed by White et al. [12] as an alternative to the heavily overused Koza-1 (“quartic”) problem. The functions of the problems are shown in Table 6. A training data set U[a, b, c] refers to c uniform random samples drawn from a to b inclusive and E[a, b, c] relates to a grid of points evenly spaced with an interval of c, from a to b inclusive. The configurations for the problems are shown in Table 5. In this problem domain, we evaluated the problems with the (best-fitness-of-run) method. The fitness of the individuals was represented by a cost function value. The cost function was defined by the sum of the absolute difference between the real function values and the values of an evaluated individual. We defined the termination criteria for the experiment with a cost function value less or equal than 0.01 and a predefined number of generations. We defined a budget of 10000 fitness evaluations for each run. Table 5: Configuration of the symbolic regression problems Property Node count Number of inputs Number of outputs Function set Point mutation rate Koza-2,3, Pagie-1 10 1 1 +, −, ∗, /, sin, cos, ln(|n|), e n 0.2 Table 6: Symbolic regression problems of the first experiment Problem Koza-2 Koza-3 Pagie-1 Objective Function x 5 − 2x 3 + x x 6 − 2x 4 + x 2 1/(1 + x −4 ) + 1/(1 + y −4 ) Vars 1 1 2 Training Set U[-1,1,20] U[-1,1,20] E[-5,5,0.4] Table 7, 8 and 9 show the results of the grid analysis for the symbolic regression problems. It is clearly visible that the use of the insertion and deletion mutation technique significantly improves the fitness after a predefined number of fitness evaluations. 5 Discussion Our experiments indicate that the insertion and deletion mutation may be beneficial for the search performance of CGP. Furthermore, our experiments showed beneficial effects in two different problem domains. The results of our 0.0 Deletion Rate 0.1 0.2 0.3 0.0 0, 33 0, 25 0, 25 0, 23 Insertion Rate 0.1 0.2 0.3 0, 20‡ 0, 18‡ 0, 17‡ 0, 21‡ 0, 17‡ 0, 20‡ 0, 20‡ 0, 18‡ 0, 19‡ 0, 21‡ 0, 21‡ 0, 21‡ Table 7: Mean best fitness of run for various settings of insertion and deletion mutation for the Koza-2 problem 0.0 Deletion Rate 0.1 0.2 0.3 0.0 0, 34 0, 32 0, 30 0, 32 Insertion Rate 0.1 0.2 0.3 0, 25‡ 0, 23‡ 0, 22‡ 0, 24‡ 0, 22‡ 0, 23‡ 0, 26‡ 0, 24‡ 0, 23‡ 0, 26 0, 24‡ 0, 22‡ Table 8: Mean best fitness of run for various settings of insertion and deletion mutation for the Koza-3 problem 0.0 Deletion Rate 0.1 0.2 0.3 0.0 195, 90 194, 12 192, 69 193, 93 Insertion Rate 0.1 0.2 186, 12 176, 23‡ 185, 55† 175, 03‡ 183, 32† 187, 43 191, 85 183, 04† 0.3 175, 45‡ 172, 52‡ 179, 07‡ 178, 39‡ Table 9: Mean best fitness of run for various settings of insertion and deletion mutation for the Pagie-1 problem experiments also show that the proposed phenotypic mutations can significantly contribute to the search performance for different types of fitness. However, in our experiments, we only investigated simple test problems which are well known in the field of GP and CGP. For more significant statements about the potential of the insertion and deletion mutation techniques in CGP a more detailed and comprehensive study is needed. This study should include a larger set of state-of-the-art benchmarks which have been proposed by McDermott et al. [9] and White et al. [12] Furthermore, this study should also include more problem domains. Beside to the question of contributions to the search performance of CGP, it has to be investigated in which way the proposed mutations contribute to the search performance. A suitable method to investigate the behavior and effect of both mutations would be an analysis of the exploration abilities of the search space. A method which could contribute to achieving more insight into the exploration behavior of both mutations would be an analysis of the fitness landscape. This analysis could also answer the question if our proposed mutations contribute to the overstepping of local optima. Another important point to discuss is the parametrization of the insertion and deletion mutation technique. On some problems, the simultaneous use of both mutations showed beneficial effects. However, our experiments also show that the sole use of one phenotypic mutation can also beneficial for the search performance. Therefore we think that a detailed investigation of different cases is needed in which the use of both or only one mutation is beneficial. The last point which should be discussed is the complexity of the insertion and deletion mutation. To activate and deactivate certain function nodes, both mutations require a permanent listing of the active function nodes and the corresponding structure of connections. Therefore we think that the runtime of both mutations should be investigated on a theoretical and practical level. The results of the runtime measurement should also be compared to the runtime of the respective fitness evaluation tasks. This type of comparison could be very helpful to get more clearness if the insertion and deletion mutation are really beneficial for the search performance of CGP. 6 Conclusion and Future Work Within this paper, we proposed two new phenotypic mutations and took a step towards advanced phenotypic mutations in CGP. A first initial empirical testing indicates that the use of both mutation techniques could be beneficial for the search performance of CGP. Our experiments clearly show that the use of the insertion and deletion mutation techniques can significantly improve the search performance of CGP for our tested problems. However, for more significant statements about the beneficial effects of the proposed mutations, a rigorous and comprehensive study on a larger set of problems is needed and should include the investigation of different CGP algorithms. Consequently, we will mainly focus on more detailed and comprehensive experiments in the future. These experiments will also include an analysis of the exploration abilities of CGP when the proposed mutations are in use. Another part of our future work is devoted to a detailed investigation of the parametrization of both mutation techniques. This will also include investigations in which way both mutations work together and if there are similar functional behaviors between different problems. References 1. Angeline, P.J.: An investigation into the sensitivity of genetic programming to the frequency of leaf selection during subtree crossover. In: Koza, J.R., Goldberg, D.E., Fogel, D.B., Riolo, R.L. (eds.) Genetic Programming 1996: Proceedings of the First Annual Conference. pp. 21–29. MIT Press, Stanford University, CA, USA (28–31 Jul 1996) 2. Clegg, J., Walker, J.A., Miller, J.F.: A New Crossover Technique for Cartesian Genetic Programming. In: GECCO ’07: Proceedings of the 9th annual conference on Genetic and evolutionary computation. vol. 2, pp. 1580–1587. ACM Press, London (7-11 Jul 2007) 3. Goldman, B.W., Punch, W.F.: Length bias and search limitations in cartesian genetic programming. In: Proceedings of the 15th Annual Conference on Genetic and Evolutionary Computation. pp. 933–940. GECCO ’13, ACM, New York, NY, USA (2013) 4. Kalkreuth, R., Rudolph, G., Droschinsky, A.: A New Subgraph Crossover for Cartesian Genetic Programming, pp. 294–310. Springer International Publishing, Cham (2017), https://doi.org/10.1007/978-3-319-55696-3_19 5. Koza, J.: Genetic Programming: A paradigm for genetically breeding populations of computer programs to solve problems. Technical Report STAN-CS-90-1314, Dept. of Computer Science, Stanford University (Jun 1990) 6. Koza, J.R.: Genetic Programming: On the Programming of Computers by Means of Natural Selection. MIT Press, Cambridge, MA, USA (1992) 7. Koza, J.R.: Genetic Programming II: Automatic Discovery of Reusable Programs. MIT Press, Cambridge Massachusetts (May 1994) 8. Kraft, D.H., Petry, F.E., Buckles, B.P., Sadasivan, T.: The use of genetic programming to build queries for information retrieval. In: Proceedings of the 1994 IEEE World Congress on Computational Intelligence. vol. 1, pp. 468–473. IEEE Press, Orlando, Florida, USA (27-29 Jun 1994) 9. McDermott, J., White, D.R., Luke, S., Manzoni, L., Castelli, M., Vanneschi, L., Jaśkowski, W., Krawiec, K., Harper, R., Jong, K.D., O’Reilly, U.M.: Genetic programming needs better benchmarks. In: Proceedings of the 14th international conference on Genetic and Evolutionary Computation Conference. pp. 791–798. GECCO ’08, ACM, Philadelphia (2012) 10. Miller, J.F., Thomson, P.: Cartesian Genetic Programming. In: Genetic Programming, Proceedings of EuroGP’2000. LNCS, vol. 1802, pp. 121–132. Springer-Verlag, Edinburgh (15-16 Apr 2000) 11. Pedroni, E.: An explicitly neutral mutation operator in cartesian genetic programming, https://equalparts.eu/about/ Neutral-Mutations-in-Cartesian-Genetic-Programming.pdf 12. White, D.R., McDermott, J., Castelli, M., Manzoni, L., Goldman, B.W., Kronberger, G., Jaskowski, W., O’Reilly, U.M., Luke, S.: Better GP Benchmarks: Community Survey Results and Proposals. Genetic Programming and Evolvable Machines 14(1), 3–29 (Mar 2013)
9cs.NE
A Metamodel of Unit Testing for Object-Oriented Programming Languages arXiv:0912.3583v1 [cs.PL] 18 Dec 2009 Martin Lévesque Dept. of Computer Science University of Quebec at Montréal Montréal, Québec, Canada Email: [email protected] ABSTRACT A unit test is a method for verifying the accuracy and the proper functioning of a portion of a program. This work consists to study the relation and the approaches to test Object-Oriented Programming (OOP) programs and to propose a metamodel that enables the programmer to write the tests while writing the source code to be tested. Categories and Subject Descriptors D.2.5 [Software Engineering]: Testing and Debugging General Terms Verification Keywords Object-Oriented Programming, Unit Testing, Metamodel 1. INTRODUCTION A unit test [10] is a method for verifying the accuracy and the proper functioning of a portion of a program. This work consists to study the relation and the approaches to test OOP programs and to propose metamodel that enables the programmer to write the tests while writing the source code to be tested. By considering the current approaches used in the industry, the integration and the writing of the tests are hard tasks: most likely, the tools available require to write additional components manually and most of the time, the tests are not near of the source code to be tested. This paper is organized as follows. Section 2 gives a state of the art about the relation between OOP programs and the tests. Section 3 describes the proposed metamodel of unit testing for OOP languages. Section 4 gives an example of what could looks like an OOP environment which uses the proposed metamodel. finally, Section 5 contains the conclusion and future possible work. 2. OOP AND UNIT TESTS This section contains the motivations of testing OOP languages differently compared to testing procedural languages. Then, the levels of abtraction for the tests are defined. Finally, current approaches to test OOP languages are discussed. 2.1 Motivations of testing OOP languages differently compared to testing procedural languages This section describes some reasons why an OOP language must be tested completely differently compared to procedural languages. Intuitively, specialization and aggregation combined to polymorphism increases the difficulty to detect errors during the integration of several components. The primary goal consists of decreasing the maximum effort to test a unit by reusing as much as possible the tests [9, 5]. 2.1.1 Test unit In procedural languages such as (C, Pascal, etc.), the principal units are the procedure and the module. However, the principal unit in an OOP language is the class [9, 1, 4]. This difference has a huge impact on the methodology to test the software since several new OOP concepts are nonexistent in the procedural paradigm. 2.1.2 Object and encapsulation Traditional techniques to test procedural languages are not necessary application to the object oriented paradigm since the encapsulation reduces or eliminates the possibility to test a specific state of an attribute of an object [9]. 2.1.3 Specialization and inheritance A method is defines by one of the following characteristic [9]: 1. Inherited from the parent class without redefinition. 2. Redefinition without any call to super. 3. Redefinition with a call to super. 4. A new method nonexistent in the parent classes. Existing works consider that the tests for cases 2, 3 and 4 must be retested completely compared to the first case which does not require any redefinition. 2.1.4 Polymorphism Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Copyright 200X ACM X-XXXXX-XX-X/XX/XX ...$10.00. Let p(z) be a method named p which take a parameter z where z contains a method m. Depending on the dynamic type of z, several different methods m when z.m is called by polymorphism. It has a huge impact on the tests since several permutations of dynamic types must be passed for the parameters in order cover more lines of code. 2.1.5 OOP language entity dimensions OOP language entities can be categorized as follows [1]: • Object: A class instance composed of several attributes and methods. An object represents a specific state. • Class: A model which factorizes the properties of its instances. • Class hierarchy: A class inherits the properties of its parent classes. • Package (or subsystem): A subsystem defines an interface presenting services. A subsystem regroups classes with strong semantic relations. Jin et al. has developed a mechanism which uses a metamodel which allows to test the coupling between units [8]. Figure 1: System of N classes uing JUnit. • System: Composed of a set of subsystems and links them together. 2.2 Levels of abstraction for the tests The principal levels of abstraction for the test are the white box, the black box and the grey box [4]: • White box: White box tests consider the implementation of the classes and the methods. • Black box: Black box tests does not take into account the implementation but only the interface units. • Grey box: Grey box tests take into account the white box and the black box levels. 2.3 Current approaches to test OOP languages The approaches used in practice and in the research community are based on specifications or on programs. 2.3.1 Specification-based The first approach to test an object-oriented program consists to use a certain specification language which describes what the packages and the classes are actually doing. With this specification, it is possible to generate several tests automatically in order to verify the correctness of the implementation. Barbey et al. [2] have used an object oriented specification called Concurrent Object-Oriented Petri Nets (CO-OPN/2). In such a language, an interface describing the attributes in the methods is defined. The axioms enable the programmer to define the comportements of the methods. The main problem in this work is that it does not take into account the characteristics of the OOP paradigm such as polymorphism, specialization, etc. Several researchers [6, 12] have defined that the state of an object is the result of a certain number of method calls. Doong et al. [6] have used a similar approach compared to the work of Barbey and al., however they specified algebraically the behavior of the classes. The verification of the correctness of a method is done by generating a sequence of method calls. Then axioms are used to validate the results. The main problems with the specification-based approach is that the programmer must manipulate an extra language only in order to test the software. Also, the tests are most likely not near the source code to be tested which is not good by considerating that, in practice, the tests are not updated when the software change. Figure 2: Loss of the semantic relations with a typical system using JUnit. 2.3.2 Program-based The program-based approach consists to get a certain confidence level by writing manually the tests in order to simulate the main possibilities of execution of the software. A standard in the industry is the JUnit Framework [3] well used for the Java programming language. The main components of the JUnit Framework are the tests suites and the tests cases. A test case regroups several tests of a class and the suite is used to contains several tests cases. Let say a system has N classes. Then if there is one test case for each class, then the system will consists of 2∗N +1 classes (Fig. 1). Another problem with JUnit is the loss of the semantic relations between test classes (Fig. 2). In fact, classes TestClass2 and TestClass3 could be defined to be the subclass of TestClass1 and reuse the code from TestClass1. However it would be done manually. We need something more automatic in order to reuse the tests. 3. METAMODEL OF UNIT TESTING FOR OOP The proposed metamodel (Fig. 3) is an extension of the metamodel defined by Ducournau et al. [7]. The main motivation of this extension is to add properties specifically for unit tests in order to write unit tests by taking advantage of the mechanisms and the forces of OOP. The relations defined in [7] are also used and expressed as follows: • Redefines (redef): Used to redefines a given entity. • Has (has): Used to know which knowledge is known by a certain entity. For example, it can be useful to know which tests are attached to a given local property. • Introduces (intro): Useful to know which entity has introduced a property or a test. • Belongs to (belongs): Useful to know which global entity is associated to a given local entity. • Defines (def): Used to know which entity has defined a certain property. Figure 3: Metamodel of Unit Testing for OOP. 3.1.1 Property tests Global Test Property. A test property (Test Package, Local Test Class or Local Test Property) represents, in a complete OOP environment using the proposed metalmodel, a block of unit test code associated to a unit (A Package, a class or a property). The idea of incorporating test properties inside a class in the metamodel was inspired by the D Programming Language [11]. However, they do not considered OOP mechanisms such as inheritance and they have not modeled several layers of granularity. Here is a brief list of the main features of the proposed metamodel: • Covered dimensions: The proposed model enables to test packages (and sub-packages), classes and class properties. • Level of abstraction to define tests: Grey box. Testing packages allows to link classes together regardless of the implementation, which corresponds to the black box level. However, testing classes and methods corresponds to the white box level since the tests and the implementation are regrouped within the same entity. • The program-based approach to test the software is used. • Test’s units are the package, the class and the property class. On the other hand, the central unit is the class. • Encapsulation: the white box tests allow to access to all properties compared to the black box tests which allow to access only to public properties. • Inheritance and specialization of tests are built by considering the real world and the type safety. A super class’ test is useful to test a subclass, so it is inherited by default. Let’s express the global test properties which belongs to a certain global property (gc ) of a class c having super classes. P arentsc represents the set of all super classes of c: ∀p∈P arentsc c ≺ p Let Gc be the global properties of c, Lgc the local properties of a given global property gc ∈ Gc of a class c and its super classes are defined by: [ Lgc = {ld |belongs(ld , gc ), ld ∈ Ld } As said, the proposed metamodel (Fig. 3) is an extension of the metalmodel proposed by Ducournau et al. [7]. However, properties for tests are the following: where belongs is a relation used to know if a local property belongs to a global property. Then the global test properties (GT Pgc ) of a given global global property is given by: [ GT Pgc = {gtp|has(lc , gtp), gtp ∈ GT Pc } • Properties to test packages. (3) lc ∈Lgc where GP Tc represents the global test properties of a class c. The relation has(lc , gtp) states that lc knows gtp. A Global Test Property (GTP) is either inherited from a local property redefinition or introduced by a local property. The introduced global tests (IGT Pgc ) represent a sub set of GT Pgc : [ {gtp|intro(d, gtp), d∈P arentsc ∪{c} ∃lc ∈Lgc has(lc , gtp), gtp ∈ GT Pc } ⊆ GT Pgc (4) where intro(d, gtp) means that d introduces gtp. Local Test Property. Let define all local test properties LT Pc for a given class c, the local test properties of a given global property gtc is defined by: • Properties to test classes. • Properties to test class properties. (2) d∈P arentsc ∪{c} • The specialization is used for the tests class and for the tests property in order to reuse 3.1 Class and properties (1) LT P Cgtc = [ {ltc |belongs(ltc , gtc )} ltc ∈LT Pc However, redefined tests must be removed: (5) LT Pgtc = LT P Cgtc − [ {t1 |∃t2 ∈LT P Cgtc redef (t2 , t1 )} t1 ∈LT P Cgtc (6) where redef (t2 , t1 ) is a relation which expresses that the test t2 redefines the test t1 . 3.1.2 Class tests Global Test Class. A global test class is either inherited from a super class or introduced by a given class. Let GT C be the global tests class of a system, GT Cc the global tests class of a class c are defined by: [ GT Cc = {t|∃c∈P arentsc ∪{c} intro(c, t)} Figure 4: Example of name conflict of tests with inheritance: by using two times the same global test property associated to the same local property of a super class. (7) t∈GT C Local Test Class. Let LT C be the local tests of a class. The local tests of a class c is given by: LT CCc = [ {lt|∃gtc∈GT Cc belongs(lt, gtc)} (8) lt∈LT C However, redefinitions must be removed: LT Cc = S LT CCc − {t |∃t2 ∈LT CCc redef (t2 , t1 )} 1 t1 ∈LT CCc (9) 3.2 Type-Safety In static typing languages, the redefinition of the return type must covariant and the redefinition of the parameters must be contravariant in order to have safe types. It does not poses any problem for the proposed metamodel since the metamodel is an extension of the existing properties and the notion of property and class is not altered. 3.3 Property Conflicts 3.3.1 Global Test Property Name conflict of tests with inheritance: by using two times the same global test property associated to the same local property of a super class. The example in Fig. 4 is an example of a name conflict of tests by using two times the same global test property associated to the same local property of a super class. In this example, two tests for the same method operation are defined (Test1Operation and Test2Operation). Both classes B and C define a new test Test2Operation which causes a name conflict since two global test property with the same exists for the method operation. Solutions: 1. Factorize B.Test2Operation and C.Test2Operation in the super class A if the test in class B is the same than the one defined in class C. 2. Redefine operation in B and C. Figure 5: Example of name conflict of global property test with multiple inheritance. 3. Move Test2Operation in class A and redefine the local test property associated to Test2Operation in class B and C. Multiple Inheritance. The second type of possible conflict with global test properties is with multiple inheritance. The example with multiple inheritance (Fig. 5) does not cause any conflict for the tests since each new global test property is associated to the redefined add method. However, the method add is redefined in classes B and C and there exist three solutions for this problem. The tests must follow one of these three methods of conflict resolution: 1. Rename method add in B and C (for example B.addB, C.addC). 2. Selection. In class D, the method add is redefined and a method between B.add and C.add is called and the associated tests of that selection are now associated also to D.add. 3. Unification. In class D, all tests from B and C are associated to D.add. 3.3.2 Local Test Property Local Test Property name conflict with inheritance. Let A be a super class of B and C which introduces a method add and a test associated (Fig. 6). The subclasses B and C redefine the local test property Test1Add. Without class D, there Figure 7: A very simple example: Cow, grass and mouse. Figure 6: Example of Local Test Property name conflict with inheritance. exists no conflict. However, with a class D which can inherit both B.Test1Add and C.Test1Add there is a name conflict: which test should be chosen by the compiler ? Depending on the test, three strategies exists with multiple inheritance: rename, select and unify. 3.4 Tests execution A compiler of an object oriented programming language can then execute automatically by using the meta information from the metamodel proposed (Algorithm 1). Algorithm 1: Executing the tests by using the metamodel (ExecTests). Data: P , a set of all packages. 1 foreach p ∈ P do 2 // Test all packages 3 foreach Test Package tp such that has(p, tp) do 4 Execute tp 5 end 6 foreach Class c such that has(p, c) do 7 // Test all classes 8 foreach ltc ∈ LT Cc do 9 Execute ltc 10 end 11 // Test all properties: 12 foreach Global Property gp such that has(c, gp) do 13 foreach Global Test Property such that gtp ∈ GT Pgp do 14 foreach Local Test Property such that ltp ∈ LT Pgtp do 15 Execute ltp 16 end 17 end 18 end 19 end 20 end Figure 8: Metamodel of the example for the classes Animal, Mouse and Cow. 4.1 First example: Cow, grass and mouse The example (Fig. 7) contains a super class Animal and two subclasses Cow and Mouse. The Animal class defines a method lastF oodEaten which returns the last food that the animal has recently eaten. That class also introduces a new test property T estN otN ull associated to the method lastF oodEaten. This test verify that the food exists when the animal has recently eaten something. The class Cow redefines the method lastFoodEaten and adds a new test T estGrassColor in order to verify that the grass has the right color (Green). The following listing presents the pseudocode of the first example: ✞ 1 2 3 4 5 6 7 8 9 12 13 EXAMPLE This section contains a very basic example showing the strengths of the proposed metamodel in a real OOP environment like Java. p u b l i c Food ( C o l o r c ) { c o l o r s . add ( c ) ; } 10 11 4. p u b l i c c l a s s Food { p r i v a t e L i s t <Color > c o l o r s = new A r r a y L i s t < C o l o r > ( ) ; p u b l i c Food ( ) { } 14 15 16 17 / / Returns t r u e i f the food i s c . p u b l i c boolean i s C o l o r ( Color c ) { /∗ . . . ∗/ } 18 } 20 21 22 23 24 25 26 f () 2 p u b l i c c l a s s G r a s s e x t e n d s Food { public Grass ( ) { c o l o r s . add ( C o l o r . Green ) ; } } defines a Property Test named T estN ame associated to the method f . Let say that a compiler tests those classes in the order of the declaration. Then, the following output would be printed: • Animal.lastFoodEaten() v1 • Animal.TestNotNull 27 28 29 30 p u b l i c c l a s s Animal { p r i v a t e Food l a s t F o o d E a t e n = n u l l ; • Mouse.lastFoodEaten() v1 31 Test TestNotNull { C u r r e n t . l a s t F o o d E a t e n = new Food ( C o l o r . Red ) ; 32 33 34 35 37 String output = t h i s . g e t C l a s s ( ) . getName ( ) + " . T e s t N o t N u l l " ; System . o u t . p r i n t l n ( o u t p u t ) ; 40 41 } p u b l i c Food l a s t F o o d E a t e n ( ) { String output = t h i s . g e t C l a s s ( ) . getName ( ) + " . l a s t F o o d E a t e n ( ) v1 " System . o u t . p r i n t l n ( o u t p u t ) ; 42 43 44 45 46 47 48 5. CONCLUSION AND FUTURE WORK There exist two approaches in order to test a OOP software: specification-based and program-based. The approach based on a specification enables the programmer to modelize the behavior of the objects, but this technique is not so interesting since it uses an extra language and the mechanisms used does not take into account the advantages of an OOP language. The proposed metamodel offers a new approach combining the advantages of the current approaches. The metamodel makes it possible to write unit tests inside the classes to be tested. It facilitates the programmer to write and maintain a software unit. The proposed metamodel contains test properties for the class properties, for the classes and for the packages. There exists specialization links between the tests properties. The metamodel enables the compiler to execute automatically the tests. The future work is to embed the proposed metamodel in a real OOP environment in order to validate the usability and the reliability. Then, existing systems using for example JUnit will be converted to a system using the proposed metamodel in order to compare the testing approaches. return lastFoodEaten ; } 51 } 53 54 55 56 p u b l i c c l a s s Mouse e x t e n d s Animal { } 57 58 59 60 61 62 63 p u b l i c c l a s s Cow e x t e n d s Animal { Test TestGrassColor { / / Ok , i t i s g r a s s . C u r r e n t . l a s t F o o d E a t e n = new G r a s s ( ) ; 64 Food f = C u r r e n t . l a s t F o o d E a t e n ( ) ; a s s e r t T r u e ( f . i s C o l o r ( C o l o r . Green ) ) ; System . o u t . p r i n t l n ( t h i s . g e t C l a s s ( ) . getName ( ) + " . TestGrassColor" ) ; 65 66 67 68 69 } public Grass lastFoodEaten ( ) { System . o u t . p r i n t l n ( t h i s . g e t C l a s s ( ) . getName ( ) + " . l a s t F o o d E a t e n ( ) v2 " ) ; 70 71 72 73 74 75 6. REFERENCES 76 i f ( lastFoodEaten instanceof Grass ) return ( Grass ) lastFoodEaten ; 77 78 79 return null ; 80 } 81 82 } ✆ ✝ Listing 1: Pseudocode of the example. The listing 1 contains a new keyword Current which correspond to a new instance of the current class, allocated only during the test. The following syntax: ✞ 1 T e s t TestName { . . . } • Cow.TestGrassColor We can clearly observe the benefit of using the proposed metamodel: the tests defined in a class are also used in the subclasses in order to cover more lines of code automatically. 49 50 • Cow.lastFoodEaten() v2 • Cow.lastFoodEaten() v2 38 39 • Mouse.TestNotNull • Cow.TestNotNull Food l a s t = C u r r e n t . l a s t F o o d E a t e n ( ) ; a s s e r t T r u e ( l a s t != n u l l ) ; 36 52 ✆ ✝ 19 [1] S. Barbey, M. Ammann, and A. Strohmeier. Open Issues in Testing Object-Oriented Software. In Karol FrÃijhauf, editor, ECSQ’94 (European Conference on Software Quality), Basel, Switzerland, October 17-20 1994, pages 257– 267. vdf Hochschulverlag AG an der ETH ZÃijrich, 1994. Also available as Technical Report EPFL-DI No 94/45. [2] S. Barbey, D. Buchs, and C. Péraire. A Theory of Specification-Based Testing for Object-Oriented Software. In EDCC-2: Proceedings of the Second European Dependable Computing Conference on Dependable Computing, pages 303–320, London, UK, 1996. Springer-Verlag. [3] R. A. Becker and W. S. Cleveland. Java Development With Ant, Chapitre 4 : Testing with JUnit. Manning, 2002. [4] E.V. Berard. Issues in the testing of object-oriented software. In Electro/94 International. IEEE Computer Society, 1994. [5] C. Bucanac. Object Oriented Testing Report. Technical Report 1998-12-09, IDE, University Of Karlskrona/Ronneby, December 1998. [6] R. Doong and P. G. Frankl. The ASTOOT approach to testing object-oriented programs. ACM Trans. Softw. Eng. Methodol., 3(2):101–130, 1994. [7] R. Ducournau, F. Morandat, and J. Privat. Modules and Class Refinement: a Meta-Modeling Approach to Object-Oriented Programming. Technical Report 07-021, LIRMM, 2007. [8] Z. Jin and A. J. Offutt. Coupling-based Integration Testing. In ICECCS ’96: Proceedings of the 2nd IEEE International Conference on Engineering of Complex Computer Systems, page 10, Washington, DC, USA, 1996. IEEE Computer Society. [9] T. A. Roger. Testing the Polymorphic Relationships of Object-Oriented Components. Technical Report ISE-TR-99-05, Department of Information and Software Engineering, George Mason University, February 1999. [10] P. Runeson. A Survey of Unit Testing Practices. IEEE Software, 23(4):22–29, 2006. [11] W. Bright. D programming Language, 2005. http://www.digitalmars.com/d/. [12] T. Xie. Automatic identification of common and special object-oriented unit tests. In OOPSLA ’04: Companion to the 19th annual ACM SIGPLAN conference on Object-oriented programming systems, languages, and applications, pages 324–325, New York, NY, USA, 2004. ACM.
6cs.PL
lpopt: A Rule Optimization Tool for Answer Set Programming arXiv:1608.05675v2 [cs.LO] 23 Aug 2016 Manuel Bichler, Michael Morak, and Stefan Woltran TU Wien, Vienna, Austria {surname}@dbai.tuwien.ac.at Abstract. State-of-the-art answer set programming (ASP) solvers rely on a program called a grounder to convert non-ground programs containing variables into variable-free, propositional programs. The size of this grounding depends heavily on the size of the non-ground rules, and thus, reducing the size of such rules is a promising approach to improve solving performance. To this end, in this paper we announce lpopt, a tool that decomposes large logic programming rules into smaller rules that are easier to handle for current solvers. The tool is specifically tailored to handle the standard syntax of the ASP language (ASP-Core) and makes it easier for users to write efficient and intuitive ASP programs, which would otherwise often require significant hand-tuning by expert ASP engineers. It is based on an idea proposed by Morak and Woltran (2012) that we extend significantly in order to handle the full ASP syntax, including complex constructs like aggregates, weak constraints, and arithmetic expressions. We present the algorithm, the theoretical foundations on how to treat these constructs, as well as an experimental evaluation showing the viability of our approach. 1 Introduction Answer set programming (ASP) [15,17,8,13] is a well-established logic programming paradigm based on the stable model semantics of logic programs. Its main advantage is an intuitive, declarative language, and the fact that, generally, each answer set of a given logic program describes a valid answer to the original question. Moreover, ASP solvers—see e.g. [14,1,12,2]—have made huge strides in efficiency. A logic program usually consists of a set of logical implications by which new facts can be inferred from existing ones, and a set of facts that represent the concrete input instance. Logic programming in general, and ASP in particular, have also gained popularity because of their intuitive, declarative syntax. The following example illustrates this: Example 1. The following rule naturally expresses the fact that two people are relatives of the same generation up to second cousin if they share a great-grandparent. uptosecondcousin(X, Y) :parent(X, PX), parent(PX, GPX), parent(GPX, GGP), parent(GPY, GGP), parent(PY, GPY), parent(Y, PY), X != Y. Rules written in an intuitive fashion, like the one in the above example, are usually larger than strictly necessary. Unfortunately, the use of large rules causes problems for current ASP solvers since the input program is grounded first (i.e. all the variables in each rule are replaced by all possible, valid combinations of constants). This grounding step generally requires exponential time for rules of arbitrary size. In practice, the grounding time can thus become prohibitively large. Also, the ASP solver is usually quicker in evaluating the program if the grounding size remains small. In order to increase solving performance, we could therefore split the rule in Example 1 up into several smaller ones by hand, keeping track of grandparents and greatgrandparents in separate predicates, and then writing a smaller version of the second cousin rule. While this is comparatively easy to do for this example, this can become very tedious if the rules become even more complex and larger, maybe also involving negation or arithmetic expressions. However, since current ASP grounders and solvers become increasingly slower with larger rules, and noting the fact that ASP programs often need expert hand-tuning to perform well in practice, this represents a significant entry barrier and contradicts the fact that logic programs should be fully declarative: in a perfect world, the concrete formulation should not have an impact on the runtime. In addition, to minimize solver runtime in general, it is therefore one of our goals to enable logic programs to be written in an intuitive, fully declarative way without having to think about various technical encoding optimizations. To this end, in this paper we propose the lpopt tool that automatically optimizes and rewrites large logic programming rules into multiple smaller ones in order to improve solving performance. This tool, based on an idea proposed for very simple ASP programs in [18], uses the concept of tree decompositions of rules to split them into smaller chunks. Intuitively, via a tree decomposition joins in the body of a rule are arranged into a tree-like form. Joins that belong together are then split off into a separate rule, only keeping the join result in a temporary atom. We then extend the algorithm to handle the entire standardized ASP language [10], and also introduce new optimizations for complex language constructs such as weak constraints, arithmetic expressions, and aggregates. The main contributions of this paper are therefore as follows: – we extend, on a theoretical basis, the lpopt algorithm proposed in [18] to the full syntax of the ASP language according to the ASP-Core-2 language specification [10]; – we establish how to treat complex constructs like aggregates, and propose an adaptation of the decomposition approach so that it can split up large aggregate expressions into multiple smaller rules and expressions, further reducing the grounding size; – we implement the lpopt algorithm in C++, yielding the lpopt tool for automated logic program optimization, and give an overview of how this tool is used in practice; and – we perform an experimental evaluation of the tool on the encodings and instances used in the fifth Answer Set Programming Competition which show the benefit of our approach, even for encodings already heavily hand-optimized by ASP experts. 2 Preliminaries General Definitions. We define two pairwise disjoint countably infinite sets of symbols: a set C of constants and a set V of variables. Different constants represent different values (unique name assumption). By X we denote sequences (or, with slight notational abuse, sets) of variables X1 , . . . , Xk with k > 0. For brevity, let [n] = {1, . . . , n}, for any integer n > 1. A (relational) schema S is a (finite) set of relational symbols (or predicates). We write p/n for the fact that p is an n-ary predicate. A term is a constant or variable. An atomic formula a over S (called S-atom) has the form p(t), where p ∈ S and t is a sequence of terms. An S-literal is either an S-atom (i.e. a positive literal), or an S-atom preceded by the negation symbol “¬” (i.e. a negative literal). For a literal `, we write dom(`) for the set of its terms, and var (`) for its variables. This notation naturally extends to sets of literals. For brevity, we will treat conjunctions of literals as sets. For a domain C ⊆ C, a (total or two-valued) S-interpretation I is a set of S-atoms containing only constants from C such that, for every S-atom p(a) ∈ I, p(a) is true, and otherwise false. When obvious from the context, we will omit the schema-prefix. A substitution from a set of literals L to a set of literals L0 is a mapping s : C∪V → C ∪ V that is defined on dom(L), is the identity on C, and p(t1 , . . . , tn ) ∈ L (resp. ¬p(t1 , . . . , tn ) ∈ L) implies p(s(t1 ), . . . , s(tn )) ∈ L0 (resp., ¬p(s(t1 ), . . . , s(tn )) ∈ L0 ). Answer Set Programming (ASP). A logic programming rule is a universally quantified reverse first-order implication of the form H(X, Y) ← B + (X, Y, Z, W) ∧ B − (X, Z), where H (the head), resp. B + (the positive body), is a disjunction, resp. conjunction, of atoms, and B − (the negative body) is a conjunction of negative literals, each over terms from C ∪ V. For a rule π, let H (π), B + (π), and B − (π) denote the set of atoms occurring in the head, the positive, and the negative body, respectively. Let B (π) = B + (π) ∪ B − (π). A rule π where H (π) = ∅ is called a constraint. Substitutions naturally extend to rules. We focus on safe rules where every variable in the rule occurs in the positive body. A rule is called ground if all its terms are constants. The grounding of a rule π w.r.t. a domain C ⊆ C is the set of rules ground C (π) = {s(π) | s is a substitution, mapping var (π) to elements from C}. A logic program Π is a finite set of logic programming rules. The schema of a program Π, denoted sch(Π), is the set of predicates appearing in Π. The active domain of Π, denoted adom(Π), with adom(Π) ⊂ C, is the set of constants appearing in Π. A program Π is ground if all its rules S are ground. The grounding of a program Π is the ground program ground (Π) = π∈Π ground adom(Π) (π). The (GelfondLifschitz) reduct of a ground program Π w.r.t. an interpretation I is the ground program Π I = {H (π) ← B + (π) | π ∈ Π, B − (π) ∩ I = ∅}. A sch(Π)-interpretation I is a (classical) model of a ground program Π, denoted I  Π if, for every ground rule π ∈ Π, it holds that I ∩ B + (π) = ∅ or I ∩ (H (π) ∪ B − (π)) 6= ∅, that is, I satisfies π. I is a stable model (or answer set) of Π, denoted I s Π if, in addition, there is no J ⊂ I such that J  Π I , that is, I is subset-minimal w.r.t. the reduct Π I . The set of answer sets of Π, denoted AS (Π), are defined as AS (Π) = {I | I is a sch(Π)-interpretation, and I s Π}. For a non-ground program Π, we define AS (Π) = AS (ground (Π)). When referring to the fact that a logic program is intended to be interpreted under the answer set semantics, we often refer to it as an ASP program. Tree Decompositions. A tree decomposition of a graph G = (V, E) is a pair T = (T, χ), where T is a rooted tree and χ is a labelling function over nodes t of T , with χ(t) ⊆ V called the bag of t, such that the following holds: (i) for each v ∈ V , there exists a node t in T , such that v ∈ χ(t); (ii) for each {v, w} ∈ E, there exists a node t in T , such that {v, w} ⊆ χ(t); and (iii) for all nodes r, s, and t in T , such that s lies on the path from r to t, we have χ(r) ∩ χ(t) ⊆ χ(s). The width of a tree decomposition is defined as the cardinality of its largest bag minus one. The treewidth of a graph G, denoted by tw (G), is the minimum width over all tree decompositions of G. To decide whether a graph has treewidth at most k is NP-complete [3]. For an arbitrary but fixed k however, this problem can be solved (and a tree decomposition constructed) in linear time [6]. Given a non-ground logic programming rule π, we let its Gaifman graph Gπ = (var (π), E) such that there is an edge (X, Y ) in E iff variables X and Y occur together in the head or in a body atom of π. We refer to a tree decomposition of Gπ as a tree decomposition of rule π. The treewidth of rule π is the treewidth of Gπ . 3 Rule Decomposition This section lays out the theoretical foundations for our rule decomposition approach. First, we recall the algorithm from [18], and then describe how it can be extended to handle three of the main extensions of the ASP language, namely arithmetic expressions, aggregates, and weak constraints (i.e. optimization statements), as defined in the ASP-Core language standard [10]. As demonstrated in Example 1, rules that are intuitive to write and read are not necessarily the most efficient ones to evaluate in practice. ASP solvers generally struggle with rules that contain many variables since they rely on a grounder-solver approach: first, the grounding of a logic program is computed by a grounder. As per the definition in Section 2, the size of the grounding can, in the worst case, be exponential in the number of variables. For large rules, the grounding step can already take a prohibitively large amount of time. However, the solver is also adversely affected by this blowup. In practice, this leads to long runtimes and sometimes the inability of the ASP system to solve a given instance. This also contributes to the fact that, while the syntax of ASP is fully declarative, writing efficient encodings still takes expert knowledge. It is therefore desirable to have a way to automatically rewrite such large rules into a more efficient representation. One way to do this is the rule decomposition approach, first proposed in [18], which we will briefly recall next. 3.1 Decomposition of Simple Rules Generally speaking, the approach in [18] computes the tree decomposition of a rule, and then splits the rule up into multiple, smaller rules according to this decomposition. While in the worst case this decomposition may not change the rule at all, in practice it is often the case that large rules can be split up very well. For instance, the large rule in Example 1 will be amenable for such a decomposition. Let us briefly recall the algorithm from [18] which we will refer to as the lpopt algorithm. For a given rule π, the algorithm works as follows: 1. Compute a tree decomposition T = (T, χ) of π with minimal width where all variables occurring in the head of π are contained in its root node bag. 2. For each node n, let temp n be a fresh predicate, and the same for each variable X in π and predicate dom X . Let Yn = χ(n) ∩ χ(pn ), where pn is the parent node of n. For the root node root, let temp root be the entire head of π, and, accordingly, Yroot = var (H (π)). Now, for a node n, generate the following rule: temp n (Yn ) ← {a ∈ B (π) ∪ {dom X (X) | var (a) ⊆ χ(n)} | a ∈ B − (π), X ∈ var (a), var (a) ⊆ χ(n), 6 ∃b ∈ B + (π) : var (b) ⊆ χ(n), X ∈ var (b)} ∪ {temp m (Ym ) | m is a child of n}. 3. For each X ∈ var (B − (π)), for which a domain predicate dom is needed to guarantee safety of a rule generated above, pick an atom a ∈ B + (π), such that X ∈ var (a) and generate a rule domX (X) ← a. Step 3 is needed because splitting up a rule may make it unsafe. In order to remedy this, a domain predicate is generated for each unsafe variable that arises due to the rule splitting in step 2. The following example illustrates how the algorithm works. Example 2. Given the rule π = h(X, W ) ← e(X, Y ), e(Y, Z), ¬e(Z, W ), e(W, X), a tree decomposition of π could look as follows (note that we write in each bag of the tree decomposition not just the variables as per definition but also all literals of rule π over these variables which is a more intuitive notation): h(X, W ), e(X, Y ), e(W, X) e(Y, Z), ¬e(Z, W ) Applying the lpopt algorithm to π with the tree decomposition above yields the following set of rules lpopt(π): domW (W ) ← e(W, X), temp(Y, W ) ← e(Y, Z), ¬e(Z, W ), domW (W ), and h(X, W ) ← e(X, Y ), e(W, X), temp(Y, W ), where temp is a fresh predicate not appearing anywhere else. Let Π be a logic program. When the above algorithm is applied to all rules in Π, resulting in a logic program lpopt(Π) as stated in [18], the answer sets of Π are preserved in the following way: when all temporary atoms are removed, each answer set of lpopt(Π) coincides with exactly one answer set from the original program Π. Furthermore, the size of the grounding now no longer depends on the rule size. In fact, it now only depends on the rule treewidth as the following result states: Theorem 1 ([18]). The size of ground (lpopt(Π)) is bounded by O(2k · n), where n is the size of Π, and k is the maximal treewidth of the rules in Π. The above theorem implies that the size of the grounding of a program Π, after optimization via the lpopt algorithm, is no longer exponential in the size of Π, but only in the treewidth of its rules. As [18] demonstrates, this decomposition approach already has a significant impact on the size of the grounding in practical instances. However, the ASP language standard [10] extends the ASP language with other useful constructs that the lpopt algorithm proposed in [18] cannot handle. These include arithmetic expressions, aggregates, and weak constraints. Looking at concrete, practical instances of ASP programs, e.g. the encodings used in recent ASP competitions [11], a large majority use such constructs. In the following, we will therefore extend the lpopt algorithm to be able to treat them in a similar way. 3.2 Treating Arithmetic Expressions Arithmetic expressions are atoms of the form X = ϕ(Y), that is, an equality with one variable (or constant number) X on the left-hand side, and an expression ϕ on the right-hand side, where ϕ is any mathematical expression built using the variables from Y, constant numbers, and the arithmetic connectives “+,” “-,” “*,” and “/.” In addition to the positive and negative body, a rule π may also contain a set of such arithmetic expressions describing a relationship between variables with the obvious meaning. Clearly, in order to adapt the rule decomposition approach to this it is easy to extend the definition of the graph representation of π to simply contain a clique between all variables occurring together in an arithmetic expression. The lpopt algorithm then works as described above up to step 2. However, a problem may arise when, in step 3 of the lpopt algorithm, a domain predicate domX (X) is to be generated. Consider the following example: Example 3. Let π be the rule a(X) ← ¬b(X, Y ), c(Y ), d (Z), X = Z + Z. A simple decomposition according to the lpopt algorithm may lead to the following rules: temp(X) ← ¬b(X, Y ), c(Y ), domX (X), and a(X) ← d (Z), X = Z + Z, temp(X). It remains to define the domain predicate domX . According to the original definition of lpopt, we would get domX (X) ← X = Z + Z which is unsafe. The conditions for safety of rules with arithmetic expressions are defined in the ASP language specification [10]. As Example 3 shows, in order for such expressions to work with the lpopt algorithm a more general approach to defining the domain predicates is needed in step 3. In fact, instead of choosing a single atom from the rule body to generate the domain predicate, in general a set of atoms and arithmetic expressions must be chosen. It is easy to see that if a rule π is safe then, for each variable X ∈ B (π), there is a set A of (positive) atoms and arithmetic expressions in the body of π that makes that variable safe. In step 3 of the lpopt algorithm, for a variable X we now choose such a set AX of body elements in a greedy fashion as follows: let S = {X} the set of variables that we need to make safe. For each variable S ∈ S, pick a (positive) atom from B (π) that makes S safe, add it to AX , and remove S from S. If no such atom exists in the body of π, greedily add the smallest arithmetic expression S = ϕ(Y) in B (π) to AX and let S = S \ {S} ∪ Y. Repeat this process until S is empty. Since π itself is safe and finite in size, the above procedure necessarily terminates. Finally, generate the rule domX (X) ← AX . It is easy to see that this rule is safe and describes the possible domain of variable X as required. Note also that this rule can not be split up futher as removing any single element of the rule would make it unsafe. Example 4. A correct domain predicate for Example 3 would be defined as follows: domX (X) ← X = Z + Z, d (Z). This ensures the proper safety of all rules generated by the lpopt algorithm. Note that the rule generated in Example 4 repeats most of the atoms that the second rule generated in Example 3 already contains. It is not immediately obvious how such situations can be remedied in general. Investigating this issue is part of ongoing work. 3.3 Treating Weak Constraints As defined in [10], a weak constraint π[k : t] is a constraint π annotated with a term k representing a weight and a sequence of terms t occurring in π. The intended meaning is that each answer set I is annotated by a total weight w(I), which is the sum over all k for each tuple of constants c that realize t in I and satisfy the body of π. Such a weak constraint can easily be decomposed by replacing π[k : t] with the rule π 0 = temp(k, t) ← B (π), where temp is a fresh predicate, and the weak constraint ⊥ ← temp(k, t)[k : t]. Finally, the lpopt algorithm is then applied to rule π 0 . This allows our rule decomposition approach also to be applied in an optimization context (i.e. where the task for the solver is to find optimal answer sets w.r.t. their weight). 3.4 Treating Aggregate Expressions An aggregate expression, as defined in [10], is an expression of the form t 4 #agg{t : ϕ(X)}, where t is a term; 4 ∈ {<, 6, =, 6=, >, >} is a builtin relation; agg is one of sum, count, max , and min; t = ht1 , . . . , tn i is a sequence of terms; and ϕ(X) is a set of literals, arithmetic expressions, and aggregate expressions, called the aggregate body. Aggregates may appear in rule bodies, or recursively inside other aggregates, with the following semantic meaning: Given an interpretation I, for each valid substitution s such that s(ϕ(X)) ⊆ I, take the tuple of constants s(t). Let us denote this set with T . Now, execute the aggregate function on T as follows: for #count, calculate |T |; for #sum, calculate Σt∈T t1 , where t1 is the first term in t; for #max and #min, take the maximum and minimum term appearing in the first position of each tuple in T , respectively. Finally, an aggregate expression is true if the relation 4 between term t and the result of the aggregate function is fulfilled. Extending the lpopt algorithm to aggregate expressions is again straightforward: The rule graph Gπ = (V, E) of a rule π containing aggregate expressions is defined as follows: Let V be the set of variables occurring in π outside of aggregate expressions. Let E be as before and, in addition, add, for each aggregate expression e, a clique between all variables var (e) ∩ V to E. Intuitively, the rule graph should contain, for each aggregate expression, a clique between all variables that appear in the aggregate and somewhere else in the rule. Variables appearing only in aggregates are in a sense “local” and are therefore not of interest when decomposing the rule. While the above transformation is straightforward, we can, however, go one step further and also decompose the inside elements of an aggregate expression. To this end, let t 4 #agg{t : ϕ(X, Y)} be an aggregate expression occurring in some rule π, where X are variables that occur either in t or somewhere else in π, and Y are variables occurring inside the aggregate only. Replace the aggregate expression with t 4 #agg{t : ψ(X, Z), temp(t, Z)}, and furthermore, generate a rule temp(t, Z) ← ψ(Y), ψ dom (Y), for some fresh predicate temp. Here, ψ contains all those atoms from ϕ that contain a variable from X, and ψ contains the rest. ψ dom contains domain predicates generated like in step 3 of the lpopt algorithm, as needed to make the temporary rule safe. The temporary rule can then be decomposed via lpopt. This is best illustrated by an example: Example 5. Let π be the following logic programming rule, saying that a vertex is “good” if it has at least two neighbours that, themselves, have a red neighbour: good (X) ← vertex (X), 2 6 #count{Y : edge(X, Y ), edge(Y, Z), red (Z)}. According to the above approach, the rule can now be split up as follows. Firstly, the aggregate is replaced: good (X) ← vertex (X), 2 6 #count{Y : edge(X, Y ), temp(Y )}, and furthermore, a temporary rule is created as follows: temp(Y ) ← edge(Y, Z), red (Z). The latter rule is now amenable for decomposition via the lpopt algorithm. Note that the above approach allows us to decompose, to a degree, even the insides of an aggregate, which, for large aggregate bodies, can lead to a further significant reduction in the grounding size. 3.5 Correctness The correctness of the above extensions to the original algorithm follows by the same arguments that prove the correctness of the original algorithm proposed in [18], and trivially from the construction for arithmetic expressions and safety. For the latter, note that for domain predicates of a variable X we explicitly select a set of atoms that make the variable safe, and that such a set always exists, since the original rule is safe. For the former two (namely weak constraints and aggregate expressions), the only thing that needs to be examined is the first step: replacing (part of) the body with a temporary predicate. But correctness of this is easy to see. Instead of performing all joins within the weak constraint or aggregate, we perform the join in a new, separate rule and project only relevant variables into a temporary predicate. The weak constraint or aggregate then only needs to consider this temporary predicate since, by construction, all other variables not projected into the temporary predicate do not play a role w.r.t. optimization or aggregation. Finally, the original algorithm from [18] extended to handle arithmetic expressions, for which correctness has already been established, is then applied to this new, separate rule. 3.6 Further Language Extensions The ASP-Core language specification [10], as well as the gringo grounder1 , allow further constructs like variable pooling, aggregates with multiple bodies, or with upper and lower bounds in the same expression, in addition to various extensions that amount to syntactic sugar. These constructs make the above explanations unnecessarily more tedious. However, from a theoretical point of view, all of these additional constructs can be normalized to one of the forms discussed in the previous subsections. Furthermore, as we shall see in the next section, we have implemented the lpopt algorithm to directly treat all standard ASP language constructs and certain other additions, like variable pooling. More details about this general approach, and the exact, but more tedious, algorithm details, can be found in [4]. 4 Implementation A full implementation of the algorithm and its extensions described in Section 3 is now available in the form of the lpopt tool, available with relevant documentation and examples at http://dbai.tuwien.ac.at/proj/lpopt. The following gives a quick outline of how to use the tool. lpopt accepts as its input any form of ASP program that follows the ASP input language specification laid out in [10]. The output of the program in its default configuration is a decomposed program that also follows this specification. In addition, the tool guarantees that no language construct is introduced in the output that was not previously present in the input (cf. Section 3). Therefore, for example, a program without aggregates will not contain any aggregates as a result of rule decomposition. The following is a description of the parameters of the tool: 1 http://potassco.sourceforge.net Usage: lpopt [-idbt] [-s seed] [-f file] [-h alg] [-l file] -d dumb: do not perform optimization -b print verbose and benchmark information -t perform only tree decomposition step -i ignore head variables when decomposing -h alg decomposition algorithm, one of {mcs, mf, miw (def)} -s seed initialize random number generator with seed. -f file the file to read from (default is stdin) -l file output infos (treewidth) to file In what follows, we will briefly describe the most important features of the tool. Tree Decomposition Heuristics. As stated in Section 2, computing an optimal tree decomposition w.r.t. width is an NP-hard problem. We thus make use of several heuristic algorithms, namely the maximum cardinality search (mcs), minimum fill (mf), and minimum induced width (miw) approaches described in [7], that yield tree decompositions that provide good upper bounds on the treewidth (i.e. on an optimal decomposition). It turns out that in practice, since rules in ASP programs are usually not overly large, these heuristics come close to, and often even yield, an optimal tree decomposition for rules. The heuristic algorithm to use for decomposition can be selected using the -h command line parameter. Since these heuristic approaches rely to some degree on randomization, a seed for the pseudo-random number generator can be passed along with the -s command line parameter. Measuring the Treewidth of Rules. Theorem 1 allows us to calculate an upper bound on the size of the grounding of the input program. In order to do this, the maximal treewidth of any rule in an ASP program must be known. The -l switch of the lpopt tool allows this to be calculated. It forces the tool to perform tree decompositions on all rules inside an input ASP program, simply outputting the maximal treewidth (or, more accurately, an upper bound; see above) over all of them into the given file, and then exiting. Clearly, when a single ASP rule is given as input, this switch will output a treewidth upper bound of that single rule. Recommended Usage Assuming that a file enc.lp contains the encoding of a problem as an ASP program and that a file instance.db contains a set of ground facts representing a problem instance, the recommended usage of the tool is as follows: cat enc.lp instance.db | lpopt | grounder | solver In the above command, grounder and solver are programs for grounding and for solving, respectively. One established solver that we will use in the next section for our experimental evaluation is clasp [14]. If clasp is used as a solver together with the lpopt tool, we generally recommend the use of the --sat-prepro flag, which often speeds up the solving process substantially for decomposed rules generated by lpopt (by considering the fact that the truth values of all temporary atoms generated by lpopt are determined exactly by the rule body, and need never be guessed). 5 Experimental Evaluation We have tested our lpopt tool and benchmarked the performance of grounding and solving of programs preprocessed with lpopt against non-preprocessed ones. All benchmarks were made on the instance sets of the fifth answer set programming competition 2014 2 , which, for most problem classes, provides two encodings, one from 2013, and one from 2014. The benchmarks have been run on a 3.5GHz AMD Opteron Processor 6308 with 192 GB of RAM to its disposal. We used the potassco software suite3 , namely gringo verison 4.5.3 as the grounder and clasp version 3.1.3 as the solver. A timeout of 300 seconds was set for solving, and 1000 seconds for grounding. Furthermore, as suggested in the previous section, clasp was called with the --sat-prepro flag enabled. In this paper, we will survey the most important results. Remark. One central aim of our tool is to improve solving performance for handwritten encodings by non-experts of ASP. In the spirit of a truly declarative language, it shouldn’t matter how an encoding is written as long as it is correct (i.e. w.r.t. runtime, there should not be a difference between “good” and “bad” encodings). In this respect, the ASP competition does not offer an optimal benchmark set since all encodings are extensively hand-tuned by ASP experts. However, as to the best of our knowledge there is no better-suited comprehensive benchmark set available, we will show that even for these extensively hand-tuned ASP competition encodings our tool can still find decompositions that decrease grounding size and improve solving performance. However, there are also encodings that are so perfectly hand-tuned that only trivial optimizations are possible with the current version of lpopt. Results. Let us first note that the runtime of lpopt itself, for all encodings in the benchmark set, was always less than what can be accurately measured on a computer system today. Applying our rule decomposition algorithm thus comes virtually for free for hand-written encodings. Out of the 49 encodings provided by the ASP competition, lpopt was able to syntactically rewrite 41 which indicates that, as mentioned above, even extensively hand-tuned programs can be further decomposed in an automated manner. The remaining eight encodings contained rules that were so small that no further decomposition was possible (i.e. their Gaifman graph was a clique of usually 3-4 nodes) and thus the output of lpopt was the original, unmodified encoding in these cases. In 27 of the 41 encodings rewritten by lpopt, the decompositions were trivial and had no significant impact on the solving performance. This is due to the fact that only rules that were already very small (and thus did not contribute much to the grounding size in the first place) could be decomposed. In five cases out of the 41 rewritten encodings, we noticed a decrease in solving performance (see the paragraph on limitations of lpopt below for an explanation) and in the remaining seven cases, the lpopt rewriting was able to speed up the solving process with substantial improvements in three of these seven. Two of those were the stable marriage problem encoding of 2013, and the permutation pattern matching encoding of 2014 which we will take a closer look at below. Full benchmark results for the entire dataset can be found in [4]. 2 3 https://www.mat.unical.it/aspcomp2014/ http://potassco.sourceforge.net (a) (b) Fig. 1: Benchmark results for the stable marriage 2013 instances. The horizontal axis represents the individual test instances, sorted by runtime without rule decomposition. As can be seen in Figure 1, both grounding and solving time decrease dramatically. Notice that the grounding time is, in general, directly correlated with the size of the respective grounding. With lpopt preprocessing, the grounding size decreases dramatically by a factor of up to 65. The grounder is thirty times faster when using preprocessing, and the solver about three times. This is because of the following constraint in the encoding that can be decomposed very well: :- match(M,W1), manAssignsScore(M,W,Smw), W1!=W, manAssignsScore(M,W1,Smw1), Smw>Smw1, match(M1,W), womanAssignsScore(W,M,Swm), womanAssignsScore(W,M1,Swm1), Swm>=Swm1. The constraint rule above is quite intuitive to read: There cannot be a man M and a woman W , such that they would both be better off if they were matched together, instead of being matched as they are (that is, to W 1 and M 1, respectively). It encodes, precisely and straightforwardly, the condition of a stable marriage. The 2014 encoding splits this rule up, making the encoding much harder to understand. However, with lpopt preprocessing, the grounding and solving performance matches that of the hand-tuned 2014 encoding. This again illustrates that the lpopt algorithm allows for efficient processing of rules written by non-experts that are not explicitly hand-tuned. (a) (b) Fig. 2: Benchmark results for permutation pattern matching 2014. The horizontal axis represents the individual test instances, sorted by runtime without rule decomposition. A second example of lpopt’s capabilities is the permutation pattern matching problem illustrated in Figure 2. The grounding time of the largest instance is 980 seconds without preprocessing and 17 seconds with preprocessing. This instance was also impossible to solve within the timeout window of 300 seconds without lpopt preprocessing, but finishing within 88 seconds when lpopt was run first. Other Use Cases. lpopt has also been employed in other works that illustrate its performance benefits. In particular, several solvers for other formalisms rely on a rewriting to ASP in order to solve the original problem. Such rewritings can easily lead to the generation of large rules that current ASP solving systems are generally unable to handle. For example, in [16] ASP rewritings for several problems from the abstract argumentation domain, proposed in [9], are implemented. In Section 4.6 of the thesis, the performance benefits of lpopt are clearly demonstated for these rewritings. Interestingly, these rewritings also make heavy use of aggregates which goes to show that lpopt also handles these constructs well. Another example is [5], where multiple rewritings for ΣP2 and ΣP3 -hard problems are proposed and then benchmarked, again showcasing that without lpopt these rewritings could not be solved by current ASP solvers in all but the most simple cases. Limitations. However, we also want to point out some limitations of the lpopt algorithm. When a domain predicate is used by the algorithm, the selection of atoms that generate this domain predicate is at the moment essentially random, since the greedy selection depends on the order of the atoms appearing in the rule. This approach, as discussed in Section 3, may thus not pick an optimal set of atoms. However, it depends on this selection how many ground rules this domain predicate rule will generate when passed to the grounder. Therefore, it may at the moment be the case that the increased grounding size caused by the domain predicate rules may destroy any benefit caused by splitting up the main rule. This is precisely what caused the increase in solving time for the five encodings out of 49 that lpopt was able to rewrite but where solving performance deteriorated. Clearly, this begs the question of what the best strategy is to select atoms to generate domain predicates. This is part of ongoing work. 6 Conclusions In this paper, we present an algorithm, based on a prototype from [18], that allows the decomposition of large logic programming rules into smaller ones that current state-ofthe-art answer set programming solvers are better equipped to handle. Our implementation handles the entire ASP-Core-2 language [10]. Benchmark results show that in practice, even for extensively hand-tuned ASP programs, our rule decomposition algorithm can improve solving performance significantly. Future work will include implementing this approach directly into state-of-the-art grounders like the gringo grounder used in our benchmarks, as well as further refining the algorithm w.r.t. selection of domain predicate atoms, as discussed at the end of Section 5. Acknowledgments. Funded by the Austrian Science Fund (FWF): Y698, P25607. References 1. Alviano, M., Dodaro, C., Faber, W., Leone, N., Ricca, F.: WASP: A native ASP solver based on constraint learning. In: Proc. LPNMR. pp. 54–66 (2013), http://dx.doi.org/10. 1007/978-3-642-40564-8_6 2. Alviano, M., Faber, W., Leone, N., Perri, S., Pfeifer, G., Terracina, G.: The disjunctive datalog system DLV. In: Datalog Reloaded. Revised Selected Papers. pp. 282–301 (2010), http://dx.doi.org/10.1007/978-3-642-24206-9_17 3. Arnborg, S., Corneil, D.G., Proskurowski, A.: Complexity of finding embeddings in a k-tree. SIAM J. Algeb. Discr. Meth. 8(2), 277–284 (1987), http://dx.doi.org/10.1137/ 0608024 4. Bichler, M.: Optimizing non-ground answer set programs via rule decomposition. BSc Thesis, TU Wien. http://dbai.tuwien.ac.at/proj/lpopt (2015) 5. Bichler, M., Morak, M., Woltran, S.: The power of non-ground rules in answer set programming. In: Proc. ICLP. To appear (2016), https://arxiv.org/abs/1608.01856 6. Bodlaender, H.L.: A linear-time algorithm for finding tree-decompositions of small treewidth. SIAM J. Comput. 25(6), 1305–1317 (1996), http://dx.doi.org/10. 1137/S0097539793251219 7. Bodlaender, H.L., Koster, A.M.C.A.: Treewidth computations i. upper bounds. Inf. Comput. 208(3), 259–275 (2010), http://dx.doi.org/10.1016/j.ic.2009.03.008 8. Brewka, G., Eiter, T., Truszczynski, M.: Answer set programming at a glance. Commun. ACM 54(12), 92–103 (2011), http://doi.acm.org/10.1145/2043174. 2043195 9. Brewka, G., Woltran, S.: GRAPPA: A semantical framework for graph-based argument processing. In: Proc. ECAI. pp. 153–158 (2014), http://dx.doi.org/10.3233/ 978-1-61499-419-0-153 10. Calimeri, F., Faber, W., Gebser, M., Ianni, G., Kaminski, R., Krennwallner, T., Leone, N., Ricca, F., Schaub, T.: ASP-Core-2 Input Language Format v2.03c (2015), https://www. mat.unical.it/aspcomp2013/ASPStandardization, accessed: 2016-06-27 11. Calimeri, F., Gebser, M., Maratea, M., Ricca, F.: Design and results of the fifth answer set programming competition. Artif. Intell. 231, 151–181 (2016), http://dx.doi.org/ 10.1016/j.artint.2015.09.008 12. Elkabani, I., Pontelli, E., Son, T.C.: SmodelsA - A system for computing answer sets of logic programs with aggregates. In: Proc. LPNMR. pp. 427–431 (2005), http://dx. doi.org/10.1007/11546207_40 13. Gebser, M., Kaminski, R., Kaufmann, B., Schaub, T.: Answer Set Solving in Practice. Synthesis Lectures on Artificial Intelligence and Machine Learning, Morgan & Claypool Publishers (2012), http://dx.doi.org/10.2200/S00457ED1V01Y201211AIM019 14. Gebser, M., Kaufmann, B., Schaub, T.: Conflict-driven answer set solving: From theory to practice. Artif. Intell. 187, 52–89 (2012), http://dx.doi.org/10.1016/j. artint.2012.04.001 15. Gelfond, M., Lifschitz, V.: The stable model semantics for logic programming. In: Proc. ICLP/SLP. pp. 1070–1080 (1988) 16. Heißenberger, G.: A System for Advanced Graphical Argumentation Formalisms. Master’s thesis, TU Wien (2016), www.dbai.tuwien.ac.at/proj/adf/grappavis/ 17. Marek, V.W., Truszczyński, M.: Stable Models and an Alternative Logic Programming Paradigm. In: The Logic Programming Paradigm – A 25-Year Perspective, pp. 375–398. Springer (1999) 18. Morak, M., Woltran, S.: Preprocessing of complex non-ground rules in answer set programming. In: Proc. ICLP. pp. 247–258 (2012), http://dx.doi.org/10.4230/LIPIcs. ICLP.2012.247
6cs.PL
The challenge of simultaneous object detection and pose estimation: a comparative study Daniel Oñoro-Rubio, Roberto J. López-Sastre, Carolina Redondo-Cabrera and Pedro Gil-Jiménez GRAM, University of Alcalá, Alcalá de Henares, 28805, Spain arXiv:1801.08110v1 [cs.CV] 24 Jan 2018 Abstract Detecting objects and estimating their pose remains as one of the major challenges of the computer vision research community. There exists a compromise between localizing the objects and estimating their viewpoints. The detector ideally needs to be viewinvariant, while the pose estimation process should be able to generalize towards the category-level. This work is an exploration of using deep learning models for solving both problems simultaneously. For doing so, we propose three novel deep learning architectures, which are able to perform a joint detection and pose estimation, where we gradually decouple the two tasks. We also investigate whether the pose estimation problem should be solved as a classification or regression problem, being this still an open question in the computer vision community. We detail a comparative analysis of all our solutions and the methods that currently define the state of the art for this problem. We use PASCAL3D+ and ObjectNet3D datasets to present the thorough experimental evaluation and main results. With the proposed models we achieve the state-of-the-art performance in both datasets. Keywords: Pose estimation, viewpoint estimation, object detection, deep learning, convolutional neural network 1. Introduction Over the last decades, the category-level object detection problem has drawn considerable attention. As a result, much progress has been realized, leaded mainly by international challenges and benchmarking datasets, such as the PASCAL VOC Challenges [1] or the ImageNet dataset [2]. Nevertheless, researchers soon identified the importance of not only localizing the objects, but also estimating their poses or viewpoints, e.g. [3, 4, 5, 6]. This new capability results fundamental to enable a true interaction with the world and its objects. For instance, a robot which merely knows the location of a cup but that cannot find its handle, will not be able to grasp it. In the end, the robotic solution needs to know a viewpoint estimation of the object to facilitate the inference of the visual affordance for the object. Also, in the augmented reality field, to localize and estimate the viewpoint of the objects, is a crucial feature in order to project a realistic hologram, for instance. Technically, given an image, these models can localize the objects, predicting their associated bounding boxes, and are also able to estimate the relative pose of the object instances in the scene with respect to the camera. Figure 1 shows an example, where the viewpoint of the object is encoded using just the azimuth angle. In the image, the target objects are the sofa and the bicycle. Their locations are depicted by their bounding boxes (in green), and their azimuth angles are represented by the blue arrow inside the yellow circle. The computer vision community rapidly detected the necessity of providing the appropriate annotated datasets, in order to Email address: [email protected], [email protected], [email protected] and [email protected] (Daniel Oñoro-Rubio, Roberto J. López-Sastre, Carolina Redondo-Cabrera and Pedro Gil-Jiménez) Preprint submitted to ArXiv Bycicle 180º Sofa 180º 0º 0º Figure 1: Object category detection and pose estimation example. In the image, the sofa and the bicycle are localized by the green bounding boxes. The blue arrow inside the yellow circles shows the azimuth angles of the objects, which is a form of viewpoint annotation. experimentally validate the object detection and pose estimations approaches. To date, several datasets have been released. Some examples are: 3D Object categories [4], EPFL Multiview car [7], ICARO [8], PASCAL3D+ [9] or ObjectNet3D [10]. Thanks to these datasets, multiple models have been experimentally evaluated. It is particularly interesting to observe how all the published approaches can be classified in two groups. In the first one, we find those models that decouple both problems (e.g. [11, 12, 13]), making first a location of the object, to later estimate its pose. In the second group we identify the approaches that solve both tasks simultaneously (e.g. [9, 14, 15]), because they understand that to carry out a correct location reJanuary 25, 2018 quires a good estimation of the pose, and vice versa. But the discrepancies do not end here. Unlike the problem of object detection, where the metric for the experimental evaluation is clear, being this the mean Average Precision (mAP) defined in the PASCAL VOC Challenge, for the problem of object detection and pose estimation, multiple metrics have been adopted. This is motivated by the fact that not all the models understand the viewpoint estimation problem in the same way. Some solutions, i.e. the discrete approaches, consider that this is a classification problem, when others, i.e. the continuous models, understand the pose as a continuous variable, whose estimation must be approached by solving a regression problem. This article is an attempt to provide a comparative study where these issues can be addressed. The main contributions of this work are as follows: must be solved separately [11, 12, 13]. The second group consists of the models where the detection and the viewpoint estimation are fully coupled [9, 14, 15, 16]. Within these two groups, one must note that while some models solve the pose estimation as a classification problem, i.e. the discrete approaches [15, 17], others treat the viewpoint estimation as a regression problem, i.e. the continuous solutions [12, 16, 18]. In this paper, we introduce three novel deep learning architectures for the problem of joint object detection and pose estimation. They all are extensions for the excellent Faster RCNN object detection model [19]. We have designed them to gradually decouple the object localization and pose estimation tasks. Our models significantly differ from previous deep learning based approaches for the same tasks. For instance, if we consider the work of Tulsiani et al. [17], we observe that their solution is based on a detector (using the R-CNN [20]), followed by a pose classification network, fully decoupling both tasks. On the contrary, all our architectures are trained fully end-to-end, performing a joint detection and viewpoint estimation. Moreover, the deep architectures implemented are different. Massa et al. [15] also propose a joint model. However, their approach is completely different. They base their design on the Fast R-CNN detector [21]. Technically, they modify the Fast R-CNN output to provide the detections based on an accumulative sum of scores that is provided by the pose classification for each object category. In a different manner, our solutions are based on the Faster R-CNN, which is a distinct architecture. Moreover, in our work we explore not only a modification of the output of the networks, but multiple architecture designs where we can gradually separate the branches of the network dedicated to the object localization and the viewpoint estimation tasks. Finally, this paper offers a detailed comparative study of solutions for the joint object detection and pose estimation problem. The study included in [22] focus on the different problem of object classification and pose estimation,i.e. they do not consider the object localization task. • We introduce three novel deep learning architectures for the problem of simultaneous object detection and pose estimation. Our models seek to perform a joint detection and pose estimation, trained fully end-to-end. We start with a model that fully integrates the tasks of object localization and object pose estimation. Then, we present two architectures that gradually decouple both tasks, proposing a final deep network where the integration is minimal. All our solutions are detailed in Sections 3.1 and 3.2. • All our architectures have been carefully designed to be able to treat the pose estimation problem from a continuous or from a discrete perspective. We simply need to change the loss functions used during learning. This is detailed in Section 3.3. Therefore, in our experiments, we carefully compare the performance of these two families of methods, reporting results using four different loss functions. Therefore, this paper aims to shed some light on which perspective is more appropriate, keeping the network architecture fixed. • Thanks to the proposed models, we are able to offer an experimental evaluation (see Section 4) designed to carefully analyze how coupled the detection and pose estimation tasks are, being this our final contribution. We also bring a detailed comparison with all the solutions that establish the state-of-the-art for the problem of object category detection and pose estimation. We carefully analyze all the models using two publicly available datasets: PASCAL3D+ [9] and ObjectNet3D [10]. 3. Simultaneous detection and pose estimation models In the following section, we formulate the learning problem for a joint detection and pose estimation. Then, we detail the proposed architectures, named: single-path, specific-path, and specific-network (Figure 2 shows an overview of all our designs). Technically, they all are extensions for the Faster RCNN approach [19]. Finally, we provide a detailed analysis of the loss functions used in our experimental evaluation. 2. Related Work Object category detection and viewpoint estimation is a growing research field. Several are the methods that have contributed to improve the state of the art. Like we just have said, we can organize in two groups all the approaches in the literature. In the first one, we find those models that understand that these two tasks, i.e. object localization and pose estimation, 3.1. Learning model for simultaneous detection and pose estimation Our goal is to learn a strong visual representation that allows the models to: localize the objects, classify them and estimate their viewpoint with respect to the camera. Furthermore, we consider an in the wild setting where multiple objects of a 2 Conv5 RPN ROI POOLING Bbox. Reg. Cls Pooled Features Pose Pose Pred. ... Fc7 Conv1 Fc6 Image of as the composition of a number of convolutional structure functions, which transform the input image to feature maps that are used to solve the target tasks. For the particular problem of simultaneous object detection and viewpoint estimation, our CNN prediction tˆ should be expressed as follows, Detection Faster R-CNN ROIS (a) Single-path architecture. Cls Detection Pooled Features (Pose) Pose ROI POOLING Pooled Features ROIS (b) Specific-path architecture. Image RPN Conv1 (Pose) ... Cls Detection (Det) ROIS Conv5 (Pose) Pose ... Pose Pred. (Det) Conv5 Bbox. Reg. Faster R-CNN Conv1 (1) zθ : X → RD represents the D-dimensional feature mapping that the network performs to the input images. Technically, it consists in the transformation of the input image xi into a feature that is used to feed the output layers of our models. We encode in θ the trainable weights of the deep architecture that allow the network to perform the mapping. In our solutions, the weights in θ define the hidden layers that are shared by all the tasks that the deep network needs to solve. FW corresponds the set of functions of the output layers. They take as input the deep feature map zθ (xi ). For the problem considered in this paper, our set of functions must address three different tasks: classification (y), object localization (β) and viewpoint estimation (φ). Therefore, FW = ( fWy y , fWβ β , fWφ φ ). fWy y with weights W y produces the prediction for the object category, i.e. tˆy . fWβ β predicts the object location tˆβ . Finally, fWφ φ is in charge of the prediction of the viewpoint tˆφ . According to the prediction model detailed in Equation 1, we define the following objective function to learn our multitask neural network: Pose Pred. Pooled Features (Det) ROI POOLING Bbox. Reg. Fc7 (Det) Fc7 (Pose) ROI POOLING Fc7 (Det) RPN Fc7 (Pose) Conv5 Fc6 (Det) ... Fc6 (Pose) Conv1 Fc6 (Det) Image Fc6 (Pose) Faster R-CNN tˆθ,W = FW ◦ zθ (xi ) . VGG-16 (c) Specific-network architecture. Figure 2: Proposed deep learning architectures for simultaneous object detection and pose estimation. arg min L(θ, W, S ) , θ,W variety of categories appear in real-world scenarios, with a considerable variability on the background, and where occlusions and truncations are the rule rather than the exception. Therefore, the supervised learning process starts from a trainN ing set S = {(xi , ti )}i=1 , where N is the number of training samples. For each sample i in the dataset, xi ∈ X represents the input image, and ti ∈ T , with ti = (yi , βi , φi ), encodes the annotations for the three tasks to solve: classification (yi ), object localization (βi ) and pose estimation (φi ). yi ∈ Y with Y = [1, 2, . . . , C, C + 1] describes the object class, being C the total number of object categories. Category C + 1 is used to consider a generic background class. βi ∈ R4 represents the bounding box localization of a particular object within image xi . Finally, φi ∈ R3 encodes the 3D viewpoint annotation for a particular object with respect to the camera position as a tuple of azimuth, elevation and zenith angles. We propose to learn a convolutional neural network (CNN) [23] for simultaneous object detection and pose estimation. Technically, these CNNs are a combination of three main features which let the model achieve a sort of invariance with respect to imaging conditions: local receptive fields, shared convolutional weights, and spatial pooling. Each unit in a layer receives inputs from a set of units located in a small neighborhood of the previous layer. In the forward pass of a CNN, each output feature is computed by the convolution of the input feature from the previous layer. Therefore, these deep networks can be thought (2) where the loss function follows the equation, L(θ, W, S ) = λ1 Ly (θ, W y , S )+λ2 Lβ (θ, W β , S )+λ3 Lφ (θ, W φ , S ) . (3) λi for i ∈ (1, 2, 3) corresponds to the scalar value that controls the importance of a particular loss during training. For the classification loss Ly we use a categorical cross-entropy function. A simple Euclidean loss is used for the object localization task loss Lβ . Finally, for the pose estimation loss Lφ multiple options are considered. We detail them in Section 3.3. 3.2. The proposed architectures 3.2.1. Single-path architecture Our first deep network design is the single-path architecture. It offers a natural extension of the Faster R-CNN model for the problem of simultaneous object detection and pose estimation. Technically, we simply add an extra output layer in order to predict the viewpoint of the object. To understand the extension proposed, we proceed with a description of the original Faster R-CNN pipeline. As it is shown in Figure 2a, the Faster R-CNN consists of three stages. The first stage is performed by the convolutional layers. An input image passes through the convolutional network, to be transformed into a deep feature map. The second stage is represented by the Region Proposal Network (RPN), which serves as 3 an “attention” mechanism during learning. Technically, it is a fully convolutional (sub)network, which takes an image feature map as input, and outputs a set of rectangular object proposals, with their corresponding objectness scores. To go into details, this RPN takes the feature map obtained from the last convolutional layer (e.g. convolution 5 in a VGG16-based architecture), and adds a new convolutional layer which is in charge of learning to generate regions of interest (ROIs). In the third stage, these ROIs are used for pooling those features that are passed to the last two fully-connected (FC) layers. Finally, the responses coming from the last FC layer are used by the model: 1) to classify the ROIs into background or object; and 2) to perform a final bounding box regression for a fine-grained localization of the object. In Figure 2a we represent these two tasks with the blocks named as “Cls” (for classification) and “Bbox. Reg.” (for the bounding box regression). Technically, the “Cls” module is implemented with a softmax layer, and the “Bbox. Reg.” layer is a linear regressor for the four coordinates that define a bounding box. In order to evaluate the capability of the Faster R-CNN for the task of pose estimation, guaranteeing a minimal intervention in the model architecture, we propose the single-path extension. It consists in the incorporation of an additional output layer (see box “Pose” in Figure 2a), connected to the last FC layer as well. The objective of this layer is to cast a prediction for the viewpoint, and to measure the loss for this task, propagating the appropriate gradients to the rest of the network during learning. For training this single-path model, we solve the objective loss function of Equation 2. We give the same weight to each task, i.e. λ1 = λ2 = λ3 = 1. Note that at this point, we do not specify whether the viewpoint estimation will be considered as a classification or regression problem. In this sense, different loss functions will be considered an evaluated in the experiments, in order to attain a high level of understanding of the simultaneous detection and pose estimation problem. the object classification and bounding box regression losses, i.e. λ1 = λ2 = 0 and λ3 = 1. 3.2.3. Specific-network architecture With our third architecture, named specific-network, we attempt to separate as much as possible the detection and pose estimation tasks within the same architecture. The key idea of this design is to provide a model with two networks that can be fully specialized in their respective tasks, while they are learned simultaneously and end-to-end. Consequently, as it is shown in Figure 2c, we design a model made of two independent networks: the detection network and the pose network. The detection network is in charge of fully performing the object localization task, as in the original design of the Faster R-CNN. The pose network must focus on the viewpoint estimation task, without any influence of the detection objective. Therefore, this network has now its own initial convolutional layers. To align the detection and pose estimation, the pose network receives the ROIs generated by the RPN module of the detection network. Technically, an input image is forwarded simultaneously into both convolutional networks. The second stage of the Faster R-CNN, i.e. the generation of ROIs by the RPN, occurs in the detection network only. These ROIs are shared with the pose network. Finally, each network pools its own features from the generated ROIs, feeds its FC layers with these features, and produces its corresponding outputs. Overall, we have an architecture with two specialized networks, that are synchronized to solve the object detection and pose estimation tasks in a single pass. For learning this model we follow the same procedure as for the specific-path. We train our detection network to solve the Equation 2 where λ3 = 0 and λ1 = λ2 = 1. The pose network is solved just for the pose problem, hence, λ1 = λ2 = 0 and λ3 = 1. The main difference with respect the specific-path model is that there are no shared features, so each network is fully specialized to solve its corresponding task. 3.2.2. Specific-path architecture The specific-path is our second approach. Our objective with this architecture is to explore the consequences of a slightly separation of the pose estimation task from the object class detection, learning specific deep features for each task. As it is shown in Figure 2b, the extension we propose for this second approach consists in adding two independent FC layers, which are directly connected to the pose estimation layer. Note that we do not change the rest of the architecture, i.e. both the initial convolutional layers and the RPN module are shared. The pooled features are used to feed the two groups of FC layers that form two types of features: one for the object detection task, and the other for the viewpoint estimation. Therefore, during training, each network FC path learns its specific features based on its gradients, while the rest of layers learn a shared representation. The model is learned solving the objective function shown in Equation 2. For the detection path, λ1 = λ2 = 1, and λ3 = 0. For the pose path we solve the Equation 2 getting rid of 3.2.4. Why have we chosen these designs? All our architectures are extensions of the Faster R-CNN approach [19]. Originally, the Faster R-CNN architecture was proposed to address the problem of object detection only. This model has systematically prevailing on all the detection benchmarks (e.g. PASCAL VOC [1], COCO [24] and ILSVRC detection [2]), where leading results are obtained by Faster R-CNN based models, albeit with deeper features (e.g. using deep residual networks [25]). So, following a simple performance criterion, we believe that the Faster R-CNN with its excellent results is a good choice. Our second criterion for the selection of this Faster R-CNN architecture is related with the main objective of our research: propose and evaluate solutions for the problem of simultaneous object detection and viewpoint estimation. Note that we neither address the problem of pose estimation in a classification setup in isolation (e.g. [22], where the object localization problem is not considered), nor decouple the object detection and pose estimation tasks (e.g. [17]). Our models seek to perform a joint 4 detection and pose estimation, trained fully end-to-end, and the Faster R-CNN architecture is an ideal candidate to extend. All our solutions perform a direct pooling of regions of interests in the images from the internal RPN of the Faster R-CNN. This way, we do not need to use any external process to hypothesize bounding boxes (e.g. Selective Search [26]), hence performing a truly end-to-end simultaneous object detection and pose estimation model, where the weights of the fully convolutional RPN learn to predict object bounds and objectness scores at each position, to maximize not only the object detection accuracy, but also the viewpoint estimation performance. Finally, we want to discuss our main arguments for the concrete extensions proposed in our architectures. Traditionally, the computer vision community working on the problem of pose estimation for object categories has been divided into two groups. Those that understand that the tasks of localizing objects and estimating their poses are decoupled tasks (e.g. [17, 12, 18, 13]), and those that advocate for jointly solving both tasks (e.g. [15, 27, 16, 14]). The architectures proposed in this paper move from a fully integration of both tasks, i.e. in the single-path, towards the specific-network model, where the integration is minimal. In this way, we can design an experimental evaluation to thoroughly analyze how coupled the detection and pose estimation tasks are. Moreover, all our experiments are carried on publicly available dataset which have been designed for the problem of detection and viewpoint estimation, therefore a direct comparison with previous methods that define the state of the art is also possible. orientation angles as points on a unit circle by the following transformation, p(α) = (sin(α), cos(α)), p(α) ∈ R2 . Probably, the simplest way to train the pose regressor is by using an Euclidean loss, as follows: N  2 1 X   φ p li − p fWφ φ ◦ zθ (xi ) . Lφ (θ, W , S ) = 2N i=1 φ A popular alternative to the Euclidean loss, is the Huber loss function, Lφ (θ, W φ , S ) = N   1 X log σlφ ( fWφ φ ◦ zθ (xi ) , i N i=1 1 N    φ  φ 2  1  PN   2 p li − p fW φ ◦ zθ (xi )  φ  φ  i=1    δ p li − p fW φ ◦ zθ (xi ) − 12 δ2     if p liφ − p fWφ φ ◦ zθ (xi ) ≤ δ, otherwise . (6) The advantage of this loss is that it tends to be more robust to outliers than the Euclidean loss. Finally, we propose to also use the continuous cyclic cosine cost function, which is widely used in the natural language processing literature [29]. It is defined as follows,  N p(liφ )p( fWφ φ ◦ zθ (xi )) 1 X  Lφ (θ, W , S ) = 1 − N i=1 p(liφ ) p( fWφ φ ◦ zθ (xi )) φ    . (7) 4. Experiments 4.1. Implementation details To perform our experiments, we have implemented all our models and loss functions using the deep learning framework Caffe [30]. The optimization is done by using the Stochastic Gradient Descent algorithm, with: a momentum of 0.9; a weight decay of 0.0005; and a learning rate of 0.001. The learning rate of the output layer for the pose estimation has been multiplied by a factor of 0.01, so as to guarantee that the network properly converges. We publicly release all our implementations1 . We follow the standard procedure of the Faster R-CNN [19] for training the models in an end-to-end fashion. This way, for each training iteration, just one image is taken and passed through the first set of convolutions. In a second step, a collection of 128 region proposals is generated. These regions are used to build the batch to feed the last set of FC layers. This batch contains 32 samples of foreground samples and 96 samples of background. For the experimental evaluation, we use two publicly available datasets, which have been especially designed for the evaluation of object detection and pose estimation models: PASCAL3D+ [9] and ObjectNet3D [10]. We strictly follow the experimental setup described in these datasets. In the following sections, more details are provided, as well as a thorough analysis of the results and main conclusions obtained. 3.3. Loss functions for pose estimation Unlike the well-defined object detection task, the viewpoint estimation problem has been traditionally considered from two different perspectives: the continuous and the discrete. Most methods in the literature adopt the discrete formulation. That is, they understand the pose estimation as a classification problem, relying on a coarse quantization of the poses for their multi-view object detectors (e.g. [11, 14, 27]). Only a few approaches consider that the pose estimation of categories is ultimately a continuous problem, i.e. a regression problem (e.g. [16, 18, 28]). In this paper, all our architectures are evaluated considering these two perspectives for the viewpoint estimation. When we want our models to consider discrete outputs for the pose estimation (the “Pose” layer in Figure 2), we integrate the following categorical cross-entropy loss function in Equation 2: Lφ (θ, W φ , S ) = − (5) (4) where N is the number of samples, and σlφ is the softmax function for the label liφ . When the pose estimation is considered from the continuous perspective, multiple adequate regression loss functions can be integrated. For all them, it is fundamental to deal with the circularity of the viewpoint. Therefore, we first represent the 4.2. Results in the PASCAL3D+ dataset PASCAL3D+ [9] dataset is one of the largest and most challenging datasets for the problem of object detection and pose 1 The link to download all the models and software to reproduce the results will be inserted once the paper gets accepted. 5 Airplane Bike Boat Car Train Monitor Figure 3: Some images of the PASCAL3D+ dataset. estimation. Technically, it consists of: 1) the images and annotations of the 12 rigid object categories provided with the PASCAL VOC 2012 dataset [1]; and 2) an additional set of 22,394 images taken from the ImageNet [2] dataset, for the same 12 categories. On average, it has more than 3000 instances per object category. The test set has 5823 images directly inherited from the PASCAL VOC 2012 test subset. Figure 3 shows some examples of images. One can clearly observe that the images provided contain objects “in the wild”. The standard PASCAL VOC annotation for all the objects (i.e. category label and bounding box), has been extended to provide a precise 3D pose. This has been done performing a manual alignment of the objects in the images with 3D CAD models. This way, azimuth, elevation and distance from the camera pose in 3D are provided for each object. For our analysis, we follow the official experimental setup of the PASCAL3D+ [9]. The evaluation metric for the object detection and pose estimation is the Average Viewpoint Precision (AVP). This AVP is similar to the Average Precision (AP) for object detection. To compute the AVP, every output of the detector is considered to be correct if and only if the bounding box overlap with the ground truth annotation is larger than 50% and the viewpoint estimation for the azimuth angle is correct. When we consider a discrete space for the viewpoint, the viewpoint estimation is correct if it coincides with the ground truth azimuth label. On the contrary, if the viewpoint belongs to a continuous space, then, two viewpoint labels are correct if the distance between them is smaller than a fixed threshold of 2π v , where v is the number of views. the Faster R-CNN model [19] using only the training set of the PASCAL VOC 2012 dataset. Note that the validation set of the original PASCAL VOC 2012 is now the test set proposed in the PASCAL3D+, therefore, we do not allow the Faster R-CNN to be pre-trained on it.For the rest of model weights that are not covered by the pre-trained models, we basically follow a standard random initialization. Here we simply want to explore what initialization procedure is the best option. Therefore, for this preliminary experiment, we just use our first architecture, the Single-path. The pose estimation is considered as a classification problem, using 360 discrete bins, and we employ the cross-entropy loss defined in Eq. 4. Init. strategy ImageNet PASCAL VOC 2012 mAP 49.5 63.6 mAVP 4 37.6 42.4 mAVP 8 32.0 32.2 mAVP 16 24.6 23.6 mAVP 24 20.2 18.9 Table 1: Effect of the network initialization strategy in the PASCAL3D+ for the Single-path architecture. Table 1 shows the main results using the described initialization strategies. In terms of object detection precision, i.e. mAP, the initialization of our model, using the PASCAL VOC 2012 datasets is the best option, by a considerable margin, with respect to the ImageNet based strategy. Interestingly, the mAP of our model (63.6) improves the state-of-the-art for the object detection task in the official PASCAL3D+ leaderboard 2 , where the best mAP is of 62.5 reported in [15]. In terms of a joint object detection and pose estimation, we also report the mAVP for different sets of views (4, 8, 16 and 24). The ImageNet based initialization reports slightly better results only for the more fine grained setups of 16 and 24 views. When just 4 or 8 views are considered, the initialization process using the PASCAL VOC 2012 is the best option, considering its high detection precision. This first experiment also reveals that it seems to be a trade-off between how good the system is localizing objects and how accurate the pose predictions are. Overall, we conclude that the best initialization strategy is clearly the 4.2.1. Network initialization analysis One of the most common practices in deep learning consists in initializing a deep network architecture with the weights of a model pre-trained in a big dataset, such as ImageNet [2], and then start a fine tunning process for a specific task, typically using a different dataset. For our problem of joint object detection and pose estimation, we also follow this popular recipe. In a nutshell, we fine tune our networks in the PASCAL3D+ dataset, using for the initialization of the weights two pre-trained models: the original VGG16 model [31] trained for the ImageNet dataset, and 2 Official PASCAL3D+ leaderboard is available at http://cvgl. stanford.edu/projects/pascal3d.html 6 Losses mAP mAVP 4 mAVP 8 mAVP 16 mAVP 24 Discrete (Eq. 4) Euclidean (Eq. 5) Huber (Eq. 6) Cyclic Cosine (Eq. 7) 63.6 64.3 64.5 55.6 42.4 47.9 46.1 42.1 32.2 34.7 31.5 32.2 23.6 23.2 20.2 22.5 18.9 17.6 15.2 17.5 of the network. However, we did not explore this option. Instead, we opted for applying the clipping gradient strategy [32], with a threshold value of 5. If we analyze now the mAVP, where both object detection and viewpoint estimation accuracies are considered, we can observe that, in general, the best performance is reported when the Euclidean loss based model is used. Moreover, within the group of continuous viewpoint estimation models, the Euclidean is the clear winner. Therefore, for the rest of the paper, when a continuous viewpoint model is learned, we use the Euclidean loss. Interestingly, the continuous approach wins the discrete model only when 4 and 8 set of views are considered. For 16 and 24 views, the discrete model retrieves a slightly better performance. In our experiments, we have noted that the continuous pose estimation approaches tend to offer smooth predictions that are concentrated around the most frequent viewpoint of the training set. However, the discrete approach, with a Softmax loss, does not suffer that much from this pose annotation bias. Figure 4 shows a detailed comparison of the performance between a discrete and a continuous approach for a pair of representative object categories: car and bus. Car is the class with the largest amount of samples in the PASCAL3D+ dataset, i.e. 1004 instances of non-difficult objects. The annotated views for cars are distributed quite homogeneously across all the poses, although they are slightly biased towards the frontal and rear views. Category bus provides only 301 samples, and the pose is clearly concentrated in the frontal view. For the category Car, Figures 4a and 4b show that the performance of both models (continuous and discrete) are comparable. The continuous pose model tends to get confused with nearby views, while the discrete approach reports more errors with opposite viewpoints. The scenario changes when one inspects the results for the Bus object category. Figures 4c and 4d show that the performance of the continuous model is slightly worse than the one of the discrete model. Like we detail above, the continuous model tends to concentrate its predictions around the pose annotated bias (i.e. the frontal). Observe the bar diagram in 4c, where most of the Rear views are assigned to Frontal views. We want to conclude this analysis, adding an additional dimension to the discussion: the influence (in the performance) of the evaluation metric used. The problem of simultaneous detection and pose estimation has not been associated with either a clear experimental evaluation process or an evaluation metric. Obviously, part of the problem is that discrete and continuous approaches, being of a different nature, have been evaluated in different ways. As a result, multiple evaluation metrics have been proposed, e.g. Pose Estimation Average Precision (PEAP) [5], Average Orientation Similarity (AOS) [33] and AVP [9]. We refer the reader to [34], where an extensive analysis of the different evaluation metrics is presented. We have compared the performance of the AVP and the AOS metrics. Our experiments reveal that the AVP metric tends to favor discrete approaches, while the AOS metric favors the continuous models. For instance, for the category bus, Figure 5 shows the precision-recall curves when the different metrics are used. When the AVP metric is used, the discrete approach (Dis- Table 2: Loss function analysis for the PASCAL3D+ dataset. Object detection and viewpoint estimation performances are reported. one based on the PASCAL VOC 2012 dataset. Therefore, for the rest of experiments, we follow this initialization strategy. 4.2.2. Discrete vs. Continuous approaches analysis As we have discussed in Section 3.3, the pose estimation problem can be treated following either a discrete approach, i.e. as a classification problem, or a continuous approximation, i.e. as a regression problem. One of the main objectives of our study is to shed light on this discussion. We have carefully designed all our architectures, so they all can consider a discrete and a continuous approximation to the pose estimation problem. We simply have to change the Pose estimation layer, and its associated loss function. Up to four different loss functions are analyzed in these experiments, one for the discrete case and three for the continuous approach. When the discrete scenario is considered, we follow the cross-entropy loss function in Equation 4. Technically, our architectures consider 360 different classes for the azimuth angle. For each category in the dataset (except for the background), we learn a specific pose estimator, therefore, we need to define a softmax function with a length of 360×C elements, where C is the number of classes. During learning, we have opted to mask the softmax layer, propagating only the error for the elements that correspond to the pose of the foreground class. For the continuous pose estimation problem, our networks learn to directly perform the regression of the two values corresponding to the conversion to polar coordinates the azimuth angle. We design our deep models to learn a particular regressor for each object category. And again, during learning, only the regressor that corresponds to the associated class label of the sample in the training batch, is allowed to propagate errors. Following this continuous setup, we analyze the three different loss functions introduced in Section 3.3: the Euclidean loss (Eq. 5), the Huber loss (Eq. 6), and the Cyclic cosine loss (Eq. 7). Table 2 shows the main results, when the different loss functions are used. Discrete, Euclidean and Huber losses exhibit a very similar detection performance (mAP). Only when the Cyclic Cosine loss is used, a substantial drop of the detection performance is reported. The reason we find to explain this fact is that during training, the Cyclic Cosine loss can eventually produce larger gradients than the detection loss. This issue causes that the learning process tends to focus more “attention” on the pose estimation task, obtaining a deep model with a worse object localization accuracy. A simple adjustment of the λ values in Eq. 3 did not properly work in our experiments. Another possibility could be to perform a power normalization of the gradients produced by the different losses at the same level 7 (a) Car - Euclidean loss. (b) Car - Softmax loss. (c) Bus - Euclidean loss. (d) Bus - Softmax loss. Figure 4: Viewpoint estimation performance detailed analysis. A comparison between continuous (with Euclidean loss) and discrete (with a Softmax loss) models for categories Car and Bus. (a) and (b) contain the results for the car category, while (c) and (d) show the results for the bus class. First row include pie charts showing the general performance of the models, where it is reported the percentage of: correct detections, confusions with opposite viewpoints, confusions with nearby poses, and the rest of errors (Other). Second row shows a detailed analysis, of the same type of errors, considering 8 set of viewpoints (F: Frontal, R-F: Right-Frontal, F-L: Frontal-Left, RE: Rear, RE-R: Rear-Right, L: Left, L-RE: Left-Rear and R: Right). 1 are more likely to occur in opposite views. Objectively, errors with close poses are not as important as errors associated with opposite poses. We believe that the continuous models could result more attractive for the problem we are dealing with. However, if the amount of training data is not large enough, and is not well balanced in terms of pose annotations, a discrete estimation model,i.e. based on a classifier, is the best option. This is the normal situation in all datasets, and also in the PASCAL3D+. Therefore, for the rest of our study, we opt for a discrete model. 0.9 0.8 0.7 Precision 0.6 0.5 Dis−AP = 0.7975 Dis−AOS = 0.6946 Dis−AVP (pi/12) = 0.6292 Cont−AP = 0.7922 Cont−AOS = 0.7195 Cont−AVP (pi/12) = 0.5799 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 Recall 0.6 0.7 0.8 0.9 1 4.2.3. Independent vs Joint object detection and pose estimation A quick reading of the scientific literature reveals two main models for tackling the problem of detecting and estimating the pose of object categories. On the one hand we find those who decouple both tasks. The detector is trained and executed separately to locate objects in the images. Subsequently, the pose estimator is responsible for associating a pose to the detected object. On the other hand, we have the models that are trained to solve both tasks together. In this section, we analyze the performance of these two families of works. To do so, we offer a detailed comparison of the proposed architectures in Section 3.2, with existing state-of-the-art models that belong to one family or another. We need to start this experimental evaluation making the following observations with respect to the three architecture proposed in this paper. Technically, our 3 network designs present a clear evolution in terms of the degree of coupling of the tasks of detection and pose estimation. Our Single-path approach clearly belongs to the joint family. Note that in this architecture, all the features of the network are shared for both tasks. With the Specific-path architecture we advance one step Figure 5: Detection and pose estimation performance for the bus category. A comparison based on evaluation metrics AOS and AVP, for both continuous (red tonalities) and discrete (blue tonalities) approaches. AVP) obtains a higher average precision, compared to the one reported for the continuous model (Cont-AVP). On the other hand, when the AOS metric is followed, the average precision is slightly superior for the continuous model, i.e. Dis-AOS < Cont-AOS. In any case, taking into account the observations made in [34], we would like to remark that the AOS metric is not an adequate measurement of the object detection and pose estimation problem. In [34], the authors show that this metric is dominated mainly by the detection performance, masquerading the pose estimation precision. Therefore, for the rest of our study, we choose to use an evaluation procedure based only on the AVP metric. Overall, based on these results, we conclude that continuous viewpoint estimation models tend to accumulate errors at nearby poses, while discrete pose estimation approaches errors 8 forward in the decoupling degree. It is a hybrid system, where the convolutional layer features are shared, while the FC layers are split into two paths: one for the object localization and one for the pose estimation. Finally we propose the Specificnetwork. Although it should be considered as an architecture belonging to the group of independent, we cannot forget that it actually proposes a new paradigm, where both networks, specialized in different tasks, can be trained end-to-end. Note that although the networks learn their characteristics in a decoupled way, the ROIs produced by the network in charge of the location are shared with the network for the estimation of the pose, which somehow conditions their learning. This end-toend methodology clearly differs from the rest of state-of-the-art independent models (e.g. [11] ). Table 3 shows the results for all of our architectures in the PASCAL3D+ dataset. Overall, our two independent models report a better performance than the Single-path architecture. For the specific case of 4 set of views, the best performance is given by the Specific-path model, which achieves the best AVP for 6 of 11 categories. For the rest of set of views (8, 16 and 24), the best performance is consistently achieved by our Specificnetwork architecture. The obtained results show that the independent approaches perform better than joint approaches. In Figure 6, we show some qualitative examples produced by our Specific-network architecture. We now compare our best model, i.e. the Specific-network, with the state-of-the-art models in Table 4. First of all, our Specific-network reports the best object detection results: see last column in Table 4. Depending on the the number set of views used for the evaluation in the PASCAL3D+ we can identify different winners, even from different families of methods. For instance, joint models retrieve the best results, in terms of mAVP, for 4, 8 and 24 views. For 16 views, it is the independent model in [11] the one reporting the best performance. Regarding all the results in Table 4 we can conclude that the independent approaches exhibit a better accuracy over most of the joint models. Note that the state-of-the-art for 24 view sets if achieved by the Craft-CNN [15], which uses synthetic CAD models during learning. This is also the case for the RenderCNN [27]. The rest of models, including ours, do not use any extra data in form of CAD models. Note that the Specific-network systematically reports a better performance than the RenderCNN, for instance. The Single-Shot approach [35] is the clear winner for 4 and 8 set of views, and the VP&KP [11] wins for 16 set of views. In all these scenarios, our Specific-network reports a higher detection accuracy than the winner model. This aspect is relevant, because the metric used tends to favor detectors with a lower localization precision. We refer the reader to the study in [34] for more details. In other words, the more detections that are retrieved by a model, the greater the likelihood that the objects for which pose estimations have to be assigned are objects that, being more difficult to detect, appear occluded or truncated, or that are too small, aspects which naturally complicate a correct estimation of the viewpoint. Every model comes with its own detector: VP&KP uses the R-CNN [37], Craft-CNN uses the Fast R-CNN [21], and we follow the Faster R-CNN architecture [19]. How can we evaluate the actual influence of the detector in the viewpoint estimation performance? In order to shed some light on this issue, we have decided to perform an additional experiment. We have taken the code of the VP&KP model provided by the authors. This model defines an independent type architecture, where two completely decoupled and different deep networks are used: one for detection, and one for the pose estimation. We start using our Specific-path model which has the best detection performance, and we run it over the training images. We then collect these detections on the training data to enrich the ground truth data. Note that we only collect those detections whose overlap with the original ground truth is grater than 70%. This is equivalent to the jittering technique applied in the original paper but taking into account the bounding box distribution of the detector. With this extended training data, we proceed to train the original pose estimator in [11]. For the test images, we recover our detections, and apply the described pose estimator on them. We call this pipeline: Improved VP&KP (Imp-VP&KP). Technically, the detector of the original VP&KP has been improved, using the Faster R-CNN now. As we can see in Table 5, our Improved VP&KP systematically reports better results than the original work. Moreover, in Figure 7 we present a comparison between the Imp-VP&KP and the results of the Craft-CNN [15] for 24 views. We can observe how by simply updating the object detector, the model of [11] can easily get the same performance as the Craft-CNN [15]. 4.2.4. The side effect of the pose estimation in the joint system The systems that address the object detection and pose estimation problems simultaneously, in principle, have multiple benefits, compared with the models that decouple both tasks. They are clearly more efficient, in terms of computational cost. Note that during training, for instance, both task are learned simultaneously. Moreover, for a test image, the localization of the object, and the estimation of its pose is obtained at the same time, not needing to process the images with a complex pipeline consisting of a detector followed by a viewpoint estimator. In a joint system, most of the operations are shared between tasks. In spite of these advantages, our experiments reveal that there is a trade-off between doing the object localization accurately and casting a precise estimation for the viewpoint. Ideally, a good detector should be invariant to the different poses of an object, e.g. it should correctly localize frontal and rear views of cars. This would push the detection models to learn representations that are not adequate to discriminate between the different poses, being this what a good pose estimator should learn. In Table 6 we report some results that can help us to understand the mentioned trade-off. The first two rows show the results reported by Massa et al. [15]. They show a comparison between their joint and independent approaches. Their independent solution clearly obtains a better performance for the object detection than the joint model, but also one can observe how the pose estimation precision, in terms of mAVP, decreases. 9 Methods Aero Single-path Specific-path Specific-network 78.1 78.5 77.8 Single-path Specific-path Specific-network 52.4 56.7 58.4 Single-path Specific-path Specific-network 42.9 47.2 51.3 Single-path Specific-path Specific-network 22.8 33.4 36.7 Single-path Specific-path Specific-network 18.1 26.0 22.9 Bike Boat Bus Car Chair Table MBike Sofa AP Object Detection 74.3 47.2 79.7 70.2 28.2 53.0 76.0 52.0 73.1 49.3 79.2 70.3 32.3 52.7 78.0 58.0 74.2 47.9 78.7 70.3 30.7 52.9 78.1 56.5 AVP 4 Views - Joint Object Detection and Pose Estimation 41.7 18.6 66.2 45.3 14.2 26.1 44.7 40.4 54.7 24.1 66.2 50.2 17.3 30.1 55.7 44.0 57.0 23.2 66.3 53.3 16.9 27.9 60.9 41.5 AVP 8 Views - Joint Object Detection and Pose Estimation 28.9 11.1 52.7 38.8 10.5 18.1 32.0 28.3 38.3 16.3 47.2 43.0 12.8 25.5 47.5 33.2 43.2 14.4 54.6 46.1 13.3 21.8 48.4 33.8 AVP 16 Views - Joint Object Detection and Pose Estimation 19.5 7.8 54.4 31.8 6.8 14.0 20.5 15.6 25.9 10.1 51.3 32.7 8.0 20.1 23.8 25.9 30.5 11.7 57.4 39.7 8.9 21.8 29.6 25.5 AVP 24 Views - Joint Object Detection and Pose Estimation 15.3 4.4 44.8 27.2 5.2 11.8 13.7 14.0 18.3 7.7 40.6 29.3 5.2 15.9 18.4 20.3 21.8 8.8 45.0 33.2 7.0 18.2 20.8 16.9 Train Monitor Avg. 79.5 77.9 77.7 60.8 64.6 62.7 63.6 64.9 64.3 63.7 61.6 60.1 52.9 60.4 52.6 42.4 47.4 47.1 50.2 53.4 49.4 40.5 43.5 41.7 32.2 37.1 38.2 42.9 38.0 38.0 23.6 32.5 31.9 23.6 27.4 30.2 36.9 36.7 33.4 16.9 24.4 21.8 18.9 22.1 22.7 Table 3: Object detection and pose estimation results in the PASCAL3D+ dataset. Comparison between all our architectures. In gray color we show our joint solution, i.e. the Single-path architecture. The remaining architectures (Specific-path and Specific-network) can be classified in the group of independent approaches. propose a generalization of the AVP. They basically extend the AVP to consider the prediction of the three angles provided in the annotation: azimuth, elevation and in-plane rotation. Technically, the solutions must provide an estimation for these three variables. Then, the corresponding predicted rotation matrix R̂ is constructed. The difference between the ground truth pose R, and the prediction encoded in R̂ is computed using a geodesic distance as follows: It is also interesting to observe, in the last rows of Table 6, how this trade-off between object detection and pose estimation performances also affects the model of Poirson et al. [35]. We can see that when they try to train their Single-Shot joint model to be more discriminative in terms of poses,i.e. increasing the number of sets of views from 4 to 24, the object detection accuracy tends to decrease. If we now analyze the performance reported by our solutions, from the Single-path to the Specific-network, we note that the detection performance slightly increases for our independent models, but we are able to also report a better performance in terms of pose estimation. We explain this fact with the type of deep architectures we have proposed. Both the Specific-path and the Specific-network can not be categorized as truly independent models: we do not completely decouple the tasks of object localization and pose estimation. Ours is an exercise or relaxing the amount of shared information between these tasks, which defines a training process able to enforce the networks to learn representations that are adequate for both tasks.   1 d(R, R̂) = √ || log R̂T R || . (8) 2 According to Xiang et al. in [10], for the AVP, an estimation is considered to be correct if d(R, R̂) < π6 . With respect to the technical implementation of our models, note that they cast a prediction for the three pose angles (azimuth, elevation and in-plane rotation) simultaneously. We repeat the same initialization procedure, using a pre-trained model on the ImageNet dataset. Again, we use the Stochastic Gradient Descent optimizer, with a momentum of 0.9, and the weight decay is set to 0.0005. This time we fix to 1 all the specific learning rates for each layer. The training is performed in an end-to-end fashion following the Faster R-CNN procedure [19]. 4.3. Results in the ObjectNet3D dataset In this work, we also perform a detailed experimental evaluation of our models in the large scale dataset for 3D object recognition ObjectNet3D [10]. It consists of 100 categories, 90.127 images and more than 200.000 annotated objects. This dataset has been carefully designed for the evaluation of the problems of object detection, classification, and pose estimation. Similarly to the PASCAL3D+, the object pose annotation is the result of the manual alignment of a 3D CAD model with the target object. Figure 8 shows some examples of this dataset. Like we describe in Section 4.1, we strictly follow the experimental setup detailed in [10]. Only the training data is used to learn the models. We then report our results using the validation and test sets. For the evaluation metric, Xiang et al. [10] 4.3.1. Discrete vs. Continuous approaches analysis In our experiments with the PASCAL3D+ dataset, one of the main conclusions obtained has been that the discrete pose estimation models, based on classifiers, give better results than continuous pose estimation models. When the number of training samples is not large enough, and the pose annotations are not well balanced, a discrete estimation model is generally the best option. Now, with the novel ObjectNet3D dataset, which provides more viewpoint annotations for more object categories, we have the opportunity to explore whether we can obtain a better performance for the continuous approaches. 10 Correct Prediction Bad Pose Bad Detection Figure 6: Qualitative results produced by the Specific-network in the PASCAL3D+ dataset. In green, we depict the ground truth annotations, while in red we show the results produced by our model. Rectangles correspond to the bounding boxes, while the arrows depict annotated orientations of the objects. 70 Craft-CNN 4.3.2. Comparison of our architectures In this section, we propose to analyze the performance of all our architectures, i.e. the Single-path, the Specific-path and the Specific-network, in this novel dataset. We only use the training set for learning the models, and the evaluation is carried in the validation set. Figure 9 shows that for this dataset, all our models report a very similar performance. Note how the AP reported for the object localization task is almost identical for the three networks, while for the pose estimation the Specificpath exhibits a slightly superior AVP. In any case, we conclude that for this dataset, there is no clear winner within our models. Therefore, now that the amount of training data in the ObjectNet3D dataset has increased considerably, it seems that there are no major differences between treating the problem of locating objects and estimating their pose jointly or separately. ImpVPKP 60 50 40 30 20 10 0 Figure 7: Comparison of Craft-CNN and the Imp-VP&KP experiment for 24 views. 4.3.3. A comparison with the State-of-the-art In this section, we provide a comparison with the state-ofthe-art models reported by Xiang et al. [10]. For the joint 2D detection and continuous 3D pose estimation task, they propose a modification of the Fast R-CNN [21] model, using two different base architectures: the VGG-16 [31] and the AlexNet [38]. Technically, they add a viewpoint regression FC branch just after the FC7 layer. Their network is trained to jointly solve three tasks: classification, bounding box regression and viewpoint regression. The FC layer for viewpoint regression is of size 3 × 101, i.e. , for each class, it predicts the three angles of azimuth, elevation and in-plane rotation. The smoothed L1 loss is used for viewpoint regression. Table 8 shows the comparison with the state-of-the-art models, but now in the test set of the ObjectNet3D dataset. On the first two rows of the table, we include the results of the VGG16 and the AlexNet based models reported in [9]. The last row shows the performance of our Specific-path model. First, note We follow the same procedure described in Section 4.2.2 for our previous Discrete vs. Continuous approaches analysis. We use our Single-path model, which is trained for a continuous pose estimation task, solving a regression problem using the Euclidean and Huber losses. When the discrete pose estimation problem is tackled, we simply learn a classifier employing the Softmax loss. Table 7 reports the obtained results of our Single-path architecture, trained on the training set, and evaluated over the validation set. In our experiments, we observe a similar performance among all the models, but this time the continuous pose estimation architectures exhibit a small advantage, like we have previously suggested. Therefore, for the rest of the experiments in this dataset, we use the continuous viewpoint architecture, employing the Huber loss. 11 Methods Aero VDPM [9] DPM-VOC+VP [14] Craft-CNN [15] Single-Shot [35] SubCNN [36] RenderCNN [27] VP&KP [11] Specific-network 34.6 37.4 64.6 61.4 50.0 63.1 58.4 VDPM [9] DPM-VOC+VP [14] Craft-CNN [15] Single-Shot [35] SubCNN [36] RenderCNN [27] VP&KP [11] Specific-network 23.4 28.6 58.7 48.8 44.5 57.5 51.3 VDPM [9] DPM-VOC+VP [14] Craft-CNN [15] Single-Shot [35] SubCNN [36] RenderCNN [27] VP&KP [11] Specific-network 15.4 15.9 46.1 28.0 27.5 46.6 36.7 VDPM [9] DPM-VOC+VP [14] Craft-CNN [15] Single-Shot [35] SubCNN [36] RenderCNN [27] VP&KP [11] Specific-network 8.0 9.7 42.4 33.4 20.7 21.5 37.0 22.9 Bike Boat Bus Car Chair Table MBike Sofa AVP 4 Views - Joint Object Detection and Pose Estimation 41.7 1.5 26.1 20.2 6.8 3.1 30.4 5.1 43.9 0.3 48.6 36.9 6.1 2.1 31.8 11.8 62.1 26.8- 70.0 51.4 11.3 40.7 62.7 40.6 60.4 21.1 63.0 48.7 23.8 17.4 60.7 47.8 50.5 15.1 57.1 41.8 15.7 18.6 50.8 28.4 59.4 20.3 69.8 55.2 25.1 24.3 61.1 43.8 57.0 23.2 66.3 53.3 16.9 27.9 60.9 41.5 AVP 8 Views - Joint Object Detection and Pose Estimation 36.5 1.0 35.5 23.5 5.8 3.6 25.1 12.5 40.3 0.2 38.0 36.6 9.4 2.6 32.0 11.0 56.4 19.9 62.4 42.2 10.6 34.7 58.6 38.8 36.3 16.4 39.8 37.2 19.1 13.2 37.0 32.1 41.1 10.1 48.0 36.6 13.7 15.1 39.9 26.8 54.8 18.9 59.4 51.5 24.7 20.5 59.5 43.7 43.2 14.4 54.6 46.1 13.3 21.8 48.4 33.8 AVP 16 Views - Joint Object Detection and Pose Estimation 18.4 0.5 46.9 18.1 6.0 2.2 16.1 10.0 22.9 0.3 49.0 29.6 6.1 2.3 16.7 7.1 39.6 13.6 56.0 36.8 6.4 23.5 41.8 27.0 23.7 10.0 50.8 31.4 14.3 9.4 23.4 19.5 25.8 6.5 45.8 29.7 8.5 12.0 31.4 17.7 42.0 12.7 64.6 42.7 20.8 18.5 38.8 33.5 30.5 11.7 57.4 39.7 8.9 21.8 29.6 25.5 AVP 24 Views - Joint Object Detection and Pose Estimation 14.3 0.3 39.2 13.7 4.4 3.6 10.1 8.2 16.7 2.2 42.1 24.6 4.2 2.1 10.5 4.1 37.0 18.0 59.6 43.3 7.6 25.1 39.3 29.4 29.4 9.2 54.7 35.7 5.5 23.0 30.3 27.6 16.4 7.9 34.6 24.6 9.4 7.6 19.9 20.0 22.0 4.1 38.6 25.5 7.4 11.0 24.4 15.0 33.4 10.0 54.1 40.0 17.5 19.9 34.3 28.9 21.8 8.8 45.0 33.2 7.0 18.2 20.8 16.9 Train Monitor mAVP mAP 10.7 11.1 65.9 55.9 46.1 59.4 60.1 34.7 32.2 61.3 62.3 58.2 55.4 52.6 19.5 23.8 50.7 47.5 39.7 49.1 47.1 26.8 27.0 61.0 60.7 56.9 56.9 64.3 10.9 9.8 61.2 44.4 39.1 53.3 49.4 27.4 28.6 49.7 26.9 46.5 45.6 41.7 18.7 21.5 45.1 31.9 32.9 44.5 38.2 29.9 28.3 60.4 60.7 56.9 56.9 64.3 22.1 20.2 38.8 30.7 29.7 42.5 38.0 16.3 19.9 36.4 27.8 31.4 32.9 31.9 15.6 17.3 33.3 24.5 24.2 36.0 30.2 30.0 28.3 60.0 60.7 56.9 56.9 64.3 20.0 20.7 48.1 44.1 32.7 28.0 43.9 33.4 11.2 12.9 28.4 34.3 18.2 19.8 22.7 21.8 12.1 13.6 34.4 28.8 19.3 19.8 31.1 22.7 29.5 27.1 59.9 59.3 60.7 56.9 56.9 64.3 Table 4: Comparison with the state-of-the-art in the PASCAL3D+ dataset.In gray color we show the joint solutions. that we are able to report a better detection performance than the AlexNet based solution in [9]. Second, although the VGG-16 based architecture of [9] reports the best detection results, if we simultaneously consider the object localization and viewpoint estimation accuracies, i.e. using the AVP metric, our model is the clear winner. This fact is particularly relevant, if we consider that the pose estimation is bounded by the detection performance, according to the evaluation metric used. Overall, this implies that our model is more accurate predicting poses. In a detailed comparison of our solution with the VGG-16 based architecture used in [9], we find the following differences that also help to explain the results obtained. First, while our model is trained fully end-to-end, the approach in [10] consists in training the Region Proposal Network of [19] first, and then using these proposals to fine-tune their model for the object detection and pose estimation tasks. Therefore, their model is mainly trained to optimize the detection performance, which explains why our Specific-path reports a slightly lower mAP. Second, there are significant differences in how the pose estimation is performed. In [10], a regressor is trained to directly predict viewpoint values in degrees. We, instead, decompose each angle into two polar coordinates. This decomposition naturally takes into account the cyclic nature of viewpoint angles. This explains why our Single-path model reports a better per- formance for the pose. We finally show some qualitative results for the ObjectNet3D dataset in Figure 10. 5. Conclusion In this work, we have presented a complete analysis of the state of the art for the problem of simultaneous object detection and pose estimation. We have designed an experimental validation, using the PASCAL3D+ and ObjectNet3D datasets, where we can evaluate the degree of coupling that exists among the tasks of object localization and viewpoint estimation. For doing so, we have introduced three deep learning architectures, which are able to perform a joint detection and pose estimation, where we gradually decouple these two tasks. With the proposed models we have achieved the state-of-the-art performance in both datasets. We have concluded that decoupling the detection from the viewpoint estimation task have benefits on the overall performance of the models. Furthermore, we have extended the comparative analysis of all our approaches considering the pose estimation as a discrete or a continuous problem, according to the two families of work in the literature. In our experiments, we have analyzed the main factors that need to be considered during the system design and 12 Methods Aero VP&KP [11] Imp-VP&KP 63.1 70.8 VP&KP [11] Imp-VP&KP 57.5 63.9 VP&KP [11] Imp-VP&KP 46.6 51.2 VP&KP [11] Imp-VP&KP 37.0 40.7 Bike Boat Bus Car Chair Table MBike Sofa Train AVP 4 Views - Joint Object Detection and Pose Estimation 59.4 20.3 69.8 55.2 25.1 24.3 61.1 43.8 59.4 66.2 37.9 75.5 61.6 17.7 39.5 68.9 49.6 67.0 AVP 8 Views - Joint Object Detection and Pose Estimation 54.8 18.9 59.4 51.5 24.7 20.5 59.5 43.7 53.3 61.4 29.0 63.3 56.2 15.8 32.8 65.3 42.0 60.6 AVP 16 Views - Joint Object Detection and Pose Estimation 42.0 12.7 64.6 42.7 20.8 18.5 38.8 33.5 42.5 43.2 20.4 68.9 47.3 17.7 30.1 40.8 36.5 44.7 AVP 24 Views - Joint Object Detection and Pose Estimation 33.4 10.0 54.1 40.0 17.5 19.9 34.3 28.9 43.9 36.4 15.8 58.5 45.8 10.7 28.5 35.9 28.3 49.5 Monitor Avg. 55.4 62.8 49.1 56.1 45.6 53.6 44.5 49.4 32.9 38.9 36.0 39.6 22.7 26.6 31.1 34.3 Table 5: VP&KP [11] vs. Imp-VP&KP experiment. Fork Iron Car Motorbike Plate Kettle Figure 8: ObjectNet3D image samples. Method mAP 1 mAVP Craft-CNN (AlexNet) [15] Ours Joint - 24 views (Single-path) Independent - 24 views (Specific-path) Independent - 24 views (Specific-network) Single-Shot [35] Joint - 4 views Joint - 8 views Joint - 16 views Joint - 24 views 48.6 51.6 21.1 20.5 63.6 64.9 64.3 18.9 22.1 22.7 61.0 60.4 60.0 59.3 50.7 45.1 33.3 28.8 0.8 0.7 Precision Joint - 24 views Independent - 24 views mAP 59.7 60.5 60.4 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 Recall 0.6 0.7 0.8 0.9 1 Figure 9: Object detection and pose estimation performance of our Single-path, Specific-path and Specific-network architectures in the ObjectNet3D dataset. Both AP and AVP metrics, with their associated precision-recall curves, are reported. Table 6: Analysis of the trade-off between object detection and pose estimation performance. Losses Discrete (Eq. 4) Euclidean (Eq. 5) Huber (Eq. 6) Single−path AP: 60.4 Single−path AVP: 41.5 Specific−path AP: 60.4 Specific−path AVP: 43.1 Specific−network AP: 59.8 Specific−network AVP: 40.1 0.9 mAVP 40.9 41.2 41.5 training. Despite the similar performance among the different approaches, we have observed a difference between the discrete and the continuous models. We conclude that the continuous approaches are more sensitive to the pose bias in the annotation than the discrete models, hence requiring bigger datasets. Table 7: Loss function analysis for the ObjectNet3D dataset. Object detection and viewpoint estimation performances are reported. 13 Correct Prediction Bad Pose Bad Detection Figure 10: Qualitative results produced by the Specific-path in the ObjectNet3D. In green, we depict the ground truth annotations, while in red we show the results produced by our model. Rectangles correspond to the bounding boxes, while the arrows depict annotated orientations of the objects. Method AlexNet [10] VGG-16 [10] Our mAP 54.2 67.5 64.2 mAVP 35.4 42.6 46.7 [8] R. López-Sastre, C. Redondo-Cabrera, P. Gil-Jiménez, S. MaldonadoBascón, ICARO: Image collection of annotated real-world objects, http://agamenon.tsc.uah.es/Personales/rlopez/data/icaro, 2010. [9] Y. Xiang, R. Mottaghi, S. Savarese, Beyond PASCAL: A benchmark for 3D object detection in the wild, in: IEEE Winter Conference on Applications of Computer Vision (WACV), 2014. [10] Y. Xiang, W. Kim, W. Chen, J. Ji, C. Choy, H. Su, R. Mottaghi, L. Guibas, S. Savarese, ObjectNet3D: A large scale database for 3D object recognition, in: ECCV, 2016. [11] S. Tulsiani, J. Malik, Viewpoints and keypoints, in: CVPR, 2015. [12] D. Glasner, M. Galun, S. Alpert, R. Basri, G. Shakhnarovich, Viewpointaware object detection and continuous pose estimation, Image and Vision Computing 30 (2012) 923–933. [13] C. Redondo-Cabrera, R. Lopez-Sastre, Because better detections are still possible: Multi-aspect object detection with boosted hough forest, in: BMVC, 2015. [14] B. Pepik, M. Stark, P. Gehler, B. Schiele, Teaching 3D geometry to deformable part models, in: CVPR, 2012. [15] F. Massa, R. Marlet, M. Aubry, Crafting a multi-task CNN for viewpoint estimation, in: BMVC, 2016. [16] C. Redondo-Cabrera, R. Lopez-Sastre, T. Tuytelaars, All together now: Simultaneous object detection and continuous pose estimation using a hough forest with probabilistic locally enhanced voting, in: BMVC, 2014. [17] S. Tulsiani, J. Carreira, J. Malik, Pose induction for novel object categories, in: ICCV, 2015. [18] M. Fenzi, L. Leal-Taixé, B. Rosenhahn, J. Ostermann, Class generative models based on feature regression for pose estimation of object categories, in: CVPR, 2013. [19] S. Ren, K. He, R. Girshick, J. Sun, Faster R-CNN: Towards real-time object detection with region proposal networks, in: NIPS, 2015. [20] R. Girshick, J. Donahue, T. Darrell, J. Malik, Rich feature hierarchies for accurate object detection and semantic segmentation, in: CVPR, 2014. [21] R. Girshick, Fast R-CNN, in: ICCV, 2015. [22] M. Elhoseiny, T. El-Gaaly, A. Bakry, A. M. Elgammal, A comparative analysis and study of multiview cnn models for joint object categorization and pose estimation, in: ICML, 2016. [23] Y. LeCun, B. E. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. E. Hubbard, L. D. Jackel, Handwritten digit recognition with a backpropagation network, in: NIPS, 1990, pp. 396–404. [24] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, C. L. Zitnick, Microsoft COCO: Common objects in context, in: ECCV, 2014. Table 8: Comparison with state-of-the-art models in the ObjectNet3D dataset. Acknowledgments This work is supported by project PREPEATE, with reference TEC2016-80326-R, of the Spanish Ministry of Economy, Industry and Competitiveness. We gratefully acknowledge the support of NVIDIA Corporation with the donation of the GPU used for this research. Cloud computing resources were kindly provided through a Microsoft Azure for Research Award. References References [1] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, A. Zisserman, The pascal visual object classes (VOC) challenge, International Journal of Computer Vision 88 (2010) 303–338. [2] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, Imagenet: A large-scale hierarchical image database, in: CVPR, 2009. [3] A. Thomas, V. Ferrari, B. Leibe, T. Tuytelaars, B. Schiele, L. Van Gool, Towards multi-view object class detection, in: CVPR, 2006. [4] L. Savarese, S.and Fei-Fei, View synthesis for recognizing unseen poses of object classes, in: ECCV, 2008. [5] R. J. Lopez-Sastre, T. Tuytelaars, S. Savarese, Deformable part models revisited: A performance evaluation for object category pose estimation, in: ICCV 2011, 1st IEEE Workshop on Challenges and Opportunities in Robot Perception, 2011. [6] S. Yingze-Bao, M. Sun, S. Savarese, Toward coherent object detection and scene layout understanding, Image and Vision Computing 29 (2011) 569 – 579. [7] M. Ozuysal, V. Lepetit, P. Fua, Pose estimation for category specific multiview object localization, in: CVPR, 2009. 14 [25] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: CVPR, 2016. [26] J. Uijlings, K. van de Sande, T. Gevers, A. Smeulders, Selective search for object recognition, International Journal of Computer Vision 104 (2013) 154–171. [27] H. Su, C. R. Qi, Y. Li, L. J. Guibas, Render for CNN: Viewpoint estimation in images using CNNs trained with rendered 3d model views, in: ICCV, 2015. [28] L. Beyer, A. Hermans, B. Leibe, Biternion nets: Continuous head pose regression from discrete training labels, in: Pattern Recognition, volume 9358 of Lecture Notes in Computer Science, Springer, 2015, pp. 157–168. [29] A. Singhal, Modern information retrieval: a brief overview, BULLETIN OF THE IEEE COMPUTER SOCIETY TECHNICAL COMMITTEE ON DATA ENGINEERING 24 (2001) 2001. [30] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, T. Darrell, Caffe: Convolutional architecture for fast feature embedding, arXiv preprint arXiv:1408.5093 (2014). [31] K. Simonyan, A. Zisserman, Very deep convolutional networks for largescale image recognition, CoRR abs/1409.1556 (2014). [32] R. Pascanu, T. Mikolov, Y. Bengio, On the difficulty of training recurrent neural networks, in: ICML, 2013. [33] A. Geiger, P. Lenz, R. Urtasun, Are we ready for autonomous driving? the KITTI vision benchmark suite, in: CVPR, 2012. [34] C. Redondo-Cabrera, R. J. Lopez-Sastre, Y. Xiang, T. Tuytelaars, S. Savarese, Pose estimation errors, the ultimate diagnosis, in: ECCV, 2016. [35] P. Poirson, P. Ammirato, C. Fu, W. Liu, J. Kosecka, A. C. Berg, Fast single shot detection and pose estimation, in: 3DV, 2016. [36] Y. Xiang, W. Choi, Y. Lin, S. Savarese, Subcategory-aware convolutional neural networks for object proposals and detection, in: IEEE Winter Conference on Applications of Computer Vision, 2017. [37] R. Girshick, J. Donahue, T. Darrell, J. Malik, Rich feature hierarchies for accurate object detection and semantic segmentation, in: CVPR, 2014. [38] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional neural networks, in: NIPS, 2012. 15
1cs.CV
Network Sketching: Exploiting Binary Structure in Deep CNNs Yiwen Guo1 , Anbang Yao1 , Hao Zhao2,1 , Yurong Chen1 1 Intel Labs China, 2 Department of Electronic Engineering, Tsinghua University arXiv:1706.02021v1 [cs.NE] 7 Jun 2017 {yiwen.guo,anbang.yao,hao.zhao,yurong.chen}@intel.com Abstract Convolutional neural networks (CNNs) with deep architectures have substantially advanced the state-of-the-art in computer vision tasks. However, deep networks are typically resource-intensive and thus difficult to be deployed on mobile devices. Recently, CNNs with binary weights have shown compelling efficiency to the community, whereas the accuracy of such models is usually unsatisfactory in practice. In this paper, we introduce network sketching as a novel technique of pursuing binary-weight CNNs, targeting at more faithful inference and better trade-off for practical applications. Our basic idea is to exploit binary structure directly in pre-trained filter banks and produce binaryweight models via tensor expansion. The whole process can be treated as a coarse-to-fine model approximation, akin to the pencil drawing steps of outlining and shading. To further speedup the generated models, namely the sketches, we also propose an associative implementation of binary tensor convolutions. Experimental results demonstrate that a proper sketch of AlexNet (or ResNet) outperforms the existing binary-weight models by large margins on the ImageNet large scale classification task, while the committed memory for network parameters only exceeds a little. 1. Introduction Over the past decade, convolutional neural networks (CNNs) have been accepted as the core of many computer vision solutions. Deep models trained on a massive amount of data have delivered impressive accuracy on a variety of tasks, including but not limited to semantic segmentation, face recognition, object detection and recognition. In spite of the successes, mobile devices cannot take much advantage of these models, mainly due to their inadequacy of computational resources. As is known to all, camera based games are dazzling to be operated with object recognition and detection techniques, hence it is eagerly anticipated to deploy advanced CNNs (e.g., AlexNet [15], VGG-Net [25] and ResNet [10]) on tablets and smartphones. Nevertheless, as the winner of ILSVRC-2012 com- Figure 1. Sketching a network model by exploiting binary structure within pre-trained filter banks, after which the full-precision model can be converted to an efficient one with binary (in black and light grey) connections. petition, AlexNet comes along with nearly 61 million realvalued parameters and 1.5 billion floating-point operations (FLOPs) to classify an image, making it resource-intensive in different aspects. Running it for real-time applications would require considerable high CPU/GPU workloads and memory footprints, which is prohibitive on typical mobile devices. The similar situation occurs on the deeper networks like VGG-Net and ResNet. Recently, CNNs with binary weights are designed to resolve this problem. By forcing the connection weights to only two possible values (normally +1 and −1), researchers attempt to eliminate the required floating-point multiplications (FMULs) for network inference, as they are considered to be the most expensive operations. In addition, since the real-valued weights are converted to be binary, these networks would commit much less space for storing their parameters, which leads to a great saving in the memory footprints and thus energy costs [8]. Several methods have been proposed to train such networks [1, 2, 22]. However, the reported accuracy of obtained models is unsatisfactory on large dataset (e.g., ImageNet) [22]. Even worse, since straightforwardly widen the networks does not guarantee any increase in accuracy [14], it is also unclear how we can make a trade-off between the model precision and expected accuracy with these methods. In this paper, we introduce network sketching as a new way of pursuing binary-weight CNNs, where the binary structures are exploited in pre-trained models rather than being trained from scratch. To seek the possibility of yielding state-of-the-art models, we propose two theoretical grounded algorithms, making it possible to regulate the precision of sketching for more accurate inference. Moreover, to further improve the efficiency of generated models (a.k.a., sketches), we also propose an algorithm to associatively implement binary tensor convolutions, with which the required number of floating-point additions and subtractions (FADDs)1 is likewise reduced. Experimental results demonstrate that our method works extraordinarily well on both AlexNet and ResNet. That is, with a bit more FLOPs required and a little more memory space committed, the generated sketches outperform the existing binaryweight AlexNets and ResNets by large margins, producing near state-of-the-art recognition accuracy on the ImageNet dataset. The remainder of this paper is structured as follows. In Section 2, we briefly introduce the related works on CNN acceleration and compression. In Section 3, we highlight the motivation of our method and provide some theoretical analyses for its implementations. In Section 4, we introduce the associative implementation for binary tensor convolutions. At last, Section 5 experimentally demonstrates the efficacy of our method and Section 6 draws the conclusions. appealing speedups (2× to 4×) with acceptable accuracy drop (≤ 1%). 2 Unlike the above mentioned ones, some research works regard memory saving as the top priority. To tackle the storage issue of deep networks, Gong et al. [6], Wu et al. [31] and Lin et al. [18] consider applying the quantization techniques to pre-trained CNNs, and trying to make network compressions with minor concessions on the inference accuracy. Another powerful category of methods in this scope is network pruning. Starting from the early work of LeCun et al’s [17] and Hassibi & Stork’s [9], pruning methods have delivered surprisingly good compressions on a range of CNNs, including some advanced ones like AlexNet and VGGnet [8, 26, 7]. In addition, due to the reduction in model complexity, a fair speedup can be observed as a byproduct. As a method for generating binary-weight CNNs, our network sketching is orthogonal to most of the existing compression and acceleration methods. For example, it can be jointly applied with low-rank based methods, by first decomposing the weight tensors into low-rank components and then sketching them. As for the cooperation with quantization-based methods, sketching first and conducting product quantization thereafter would be a good choice. 3. Network Sketching The deployment problem of deep CNNs has become a concern for years. Efficient models can be learnt either from scratch or from pre-trained models. Generally, training from scratch demands strong integration of network architecture and training policy [18], and here we mainly discuss representative works on the latter strategy. Early works are usually hardware-specific. Not restricted to CNNs, Vanhoucke et al. [28] take advantage of programmatic optimizations to produce a 3× speedup on x86 CPUs. On the other hand, Mathieu et al. [20] perform fast Fourier transform (FFT) on GPUs and propose to compute convolutions efficiently in the frequency domain. Additionally, Vasilache et al. [29] introduce two new FFTbased implementations for more significant speedups. More recently, low-rank based matrix (or tensor) decomposition has been used as an alternative way to accomplish this task. Mainly inspired by the seminal works from Denil et al. [4] and Rigamonti et al. [23], low-rank based methods attempt to exploit parameter redundancy among different feature channels and filters. By properly decomposing pretrained filters, these methods [5, 12, 16, 32, 19] can achieve In general, convolutional layers and fully-connected layers are the most resource-hungry components in deep CNNs. Fortunately, both of them are considered to be overparameterized [4, 30]. In this section, we highlight the motivation of our method and present its implementation details on the convolutional layers as an example. Fullyconnected layers can be operated in a similar way. Suppose that the learnable weights of a convolutional layer L can be arranged and represented as {W(i) : 0 ≤ i < n}, in which n indicates the target number of feature maps, and W(i) ∈ Rc×w×h is the weight tensor of its ith filter. Storing all these weights would require 32 × c × w × h × n bit memory, and the direct implementation of convolutions would require s × c × w × h × n FMULs (along with the same number of FADDs), in which s indicates the spatial size of target feature maps. Since many convolutional models are believed to be informational redundant, it is possible to seek their lowprecision and compact counterparts for better efficiency. With this in mind, we consider exploiting binary structures within L, by using the divide and conquer strategy. We shall first approximate the pre-trained filters with a linear span of certain binary basis, and then group the identical basis tensors to pursue the maximal network efficiency. Details 1 Without ambiguity, we collectively abbreviate floating-point additions and floating-point subtractions as FADDs. 2 Some other works concentrate on learning low-rank filters from scratch [27, 11], which is out of the scope of our paper. 2. Related Works are described in the following two subsections, in which we drop superscript (i) from W because the arguments apply to all the n weight tensors. 3.1. Approximating the Filters As claimed above, our first goal is to find a binary expansion of W that approximates it well (as P illustrated in m−1 Figure 2), which means W ≈ hB, ai = j=0 aj Bj , c×w×h×m in which B ∈ {+1, −1} and a ∈ Rm are the concatenations of m binary tensors {B0 , ..., Bm−1 } and the same number of scale factors {a0 , ..., am−1 }, respectively. We herein investigate the appropriate choice of B and a with a fixed m. Two theoretical grounded algorithms are proposed in Section 3.1.1 and 3.1.2, respectively. Figure 2. Approximate the real-valued weight tensor with a sum of scaled binary tensors. Algorithm 1 The direct approximation algorithm: Input: W: the pre-trained weight tensor, m: the desired cardinality of binary basis. Output: {Bj , aj : 0 ≤ j < m}: a binary basis and a series of scale factors. Initialize j ← 0 and Ŵj ← W. repeat Calculate Bj and aj by Equation (3). Calculate Ŵj+1 = Ŵj −aj Bj and update j ← j +1. until j reaches its maximal number m. under this circumstance, in which function sgn(·) calculates the element-wise sign of the input tensor, and t = c×w ×h. The above algorithm is summarized in Algorithm 1. It is considered to be heuristic (or greedy) in the sense that each Bj is selected to be the current optimum, regardless of whether it will preclude better approximations later on. Furthermore, some simple deductions give the following theoretical result. Theorem 1. For any m ≥ 0, Algorithm 1 achieves a reconstruction error e2 satisfying e2 ≤ kWk2 (1 − 1/t)m . 3.1.1 Direct Approximation For easy understanding, we shall first introduce the direct approximation algorithm. Generally, the reconstruction error (or approximation error, round-off error) e2 := kW − hB, aik2 should be minimized to retain the model accuracy after expansion. However, as a concrete decision problem, directly minimizing e2 seems NP-hard and thus solving it can be very time consuming [3]. In order to finish up in reasonable time, we propose a heuristic algorithm, in which Bj and aj are sequentially learnt and each of them is selected to be the current optimum with respect to the e2 minimization criterion. That is, 2 Bj , aj = arg min Ŵj − aB , (1) B∈B, a∈R in which B := {+1, −1}c×w×h , the norm operator k · k is defined as kXk := hX, Xi1/2 for any 3-D tensor X, and Ŵj indicates the approximation residue after combining all the previously generated tensor(s). In particular, Ŵj = W if j = 0, and j−1 X Ŵj = W − ak Bk (2) k=0 if j ≥ 1. It can be easily known that, through derivative calculations, Equation (1) is equivalent with Bj = sgn(Ŵj ) and aj = hBj , Ŵj i t (3) (4) Proof. Since Bj = sgn(Ŵj ), we can obtain that, X (l) hBj , Ŵj i = |ŵj | ≥ kŴj k, l (5) in which ŵj is an entry of Ŵj , with superscript (l) indicates its index. From Equation (2) and (5), we have kŴj+1 k2 = kŴj k2 − aj hBj , Ŵj i = kŴj k2 1− hBj , Ŵj i2 tkŴj k2 ! (6) ≤ kŴj k2 (1 − 1/t) . The result follows by applying Formula (6) for j varying from 0 to m − 1. 3.1.2 Approximation with Refinement We can see from Theorem 1 that, by utilizing the direct approximation algorithm, the reconstruction error e2 decays exponentially with a rate proportional to 1/t. That is, given a W with small size (i.e., when t is small), the approximation in Algorithm 1 can be pretty good with only a handful of binary tensors. Nevertheless, when t is relatively large, the reconstruction error will decrease slowly and the approximation can be unsatisfactory even with a large number of binary tensors. In this section, we propose to refine the direct approximation algorithm for better reconstruction property. Algorithm 2 Approximation with refinement: Input: W: the pre-trained weight tensor, m: the desired cardinality of binary basis. Output: {Bj , aj : 0 ≤ j < m}: a binary basis and a series of scale factors. Initialize j ← 0 and Ŵj ← W. repeat Calculate Bj and aj by Equation (3) and (7). Update j ← j + 1 and calculate Ŵj by Equation (2). until j reaches its maximal number m. Considering that, in Algorithm 1, both Bj and aj are chosen to minimize e2 with fixed counterparts. However, in most cases, it is doubtful whether B and a are optimal overall. If not, we may need to refine at least one of them for the sake of better approximation. On account of the computational simplicity, we turn to a specific case when B is fixed. That is, suppose the direct approximation has already produced B̂ and â, we hereby seek another scale vector a satisfying kW − hB̂, aik2 ≤ kW − hB̂, âik2 . To achieve this, we follow the least square regression method and get −1 T aj = BjT Bj Bj · vec(W), (7) in which the operator vec(·) gets a column vector whose elements are taken from the input tensor, and Bj gets [vec(B0 ), ..., vec(Bj )]. The above algorithm with scale factor refinement is summarized in Algorithm 2. Intuitively, the refinement operation attributes a memory-like feature to our method, and the following theorem ensures it can converge faster in comparison with Algorithm 1. Theorem 2. For any m ≥ 0, Algorithm 2 achieves a reconstruction error e2 satisfying 2 e ≤ kWk 2 m−1 Y j=0 1 1− t − λ(j, t)  , (8) in which λ(j, t) ≥ 0, for 0 ≤ j ≤ m − 1. Proof. To simplify the notations, let us further denote wj := vec(Wj ) and bj+1 := vec(Bj+1 ), then we can obtain by block matrix multiplication and inverse that,   Φ + Φψψ T Φ/r −Φψ/r T −1 (Bj+1 Bj+1 ) = , (9) −ψ T Φ/r 1/r in which Φ = (BjT Bj )−1 , ψ = BjT bj+1 , and r = bTj+1 bj+1 − ψ T Φψ. Consequently, we have the following equation for j = 0, ..., m − 1, ! Λ(bj+1 bTj+1 ) wj+1 = I − T wj , (10) bj+1 Λbj+1 by defining Λ := I − Bj ΦBjT . As we know, given positive semi-definite matrices X and Y , tr(XY ) ≤ tr(X)tr(Y ). Then, Equation (10) gives, kŴj+1 k2 ≤ kŴj k2 − = kŴj k2 − = kŴj k 2 wjT (bj+1 bTj+1 )Λ(bj+1 bTj+1 )wj (bTj+1 Λbj+1 )2 wjT (bj+1 bTj+1 )wj bTj+1 Λbj+1 1− hBj , Ŵj i2 bTj+1 Λbj+1 kŴj k2 ! . Obviously, it follows that, kŴj+1 k2 ≤ kŴj k2 (1 − 1/(t − λ(j, t))), (11) in which λ(j, t) := bTj+1 (I − Λ)bj+1 . Since λ(j, t) represents the squared Euclidean norm of an orthogonal projection of bj+1 , it is not difficult to prove λ(j, t) ≥ 0, and then the result follows. 3.2. Geometric Interpretation After expanding the pre-trained filters, we can group the identical binary tensors to save some more memory footprints. In this paper, the whole technique is named as network sketching, and the generated binary-weight model is straightforwardly called a sketch. Next we shall interpret the sketching process from a geometric point of view. For a start, we should be aware that, Equation (1) is essentially seeking a linear subspace spanned by a set of t-dimensional binary vectors to minimize its Euclidean distance to vec(W). In concept, there are two variables to be determined in this problem. Both Algorithm 1 and 2 solve it in a heuristic way, and the jth binary vector is always estimated by minimizing the distance between itself and the current approximation residue. What make them different is that, Algorithm 2 takes advantage of the linear span of its previous j−1 estimations for better approximation, whereas Algorithm 1 does not. Let us now take a closer look at Theorem 2. Compared with Equation (4) in Theorem 1, the distinction of Equation (8) mainly lies in the existence of λ(j, t). Clearly, Algorithm 2 will converge faster than Algorithm 1 as long as λ(j, t) > 0 holds for any j ∈ [0, m − 1]. Geometrically speaking, if we consider Bj (BjT Bj )−1 BjT as the matrix of an orthogonal projection onto Sj := span(b0 , ..., bj ), then λ(j, t) is equal to the squared Euclidean norm of a vector projection. Therefore, λ(j, t) = 0 holds if and only if vector bj+1 is orthogonal to Sj , or in other words, orthogonal to each element of {b0 , ..., bj } which occurs with extremely low probability and only on the condition of t ∈ {2k : k ∈ N}. That is, Algorithm 2 will probably prevail over Algorithm 1 in practice. 4. Speeding-up the Sketch Model Using Algorithm 1 or 2, one can easily get a set of mn binary tensors on L, which means the storage requirement of learnable weights is reduced by a factor of 32t/(32m + tm)×. When applying the model, the required number of FMULs is also significantly reduced, by a factor of (t/m)×. Probably, the only side effect of sketching is some increases in the number of FADDs, as it poses an extra burden on the computing units. In this section, we try to ameliorate this defect and introduce an algorithm to further speedup the binary-weight networks. We start from an observation that, yet the required number of FADDs grows monotonously with m, the inherent number of addends and augends is always fixed with a given input of L. That is, some repetitive FADDs exist in the direct implementation of binary tensor convolutions. Let us denote X ∈ Rc×w×h as an input sub-feature map and see Figure 3 for a schematic illustration. Bj1 +1 +1 −1 −1 Bj2 −1 +1 −1 +1 Bj1 ∨ Bj2 −1 0 0 +1 Table 1. Truth table of the element-wise operator ∨. Since Bj0 Y Bj1 produces ternary outputs on each index position, we can naturally regard X ∗ (Bj0 Y Bj1 ) as an iteration of switch ... case ... statements. In this manner, only the entries corresponding to ±1 in Bj0 YBj1 need to be operated in X, and thus acceleration is gained. Assuming that the inner-product of Bj0 and Bj1 equals to r, then (t − r)/2 + 1 and (t + r)/2 + 1 FADDs are still required for calculating Equation (12) and (13), respectively. Obviously, we expect the real number r ∈ [−t, +t] to be close to either t or −t for the possibility of fewer FADDs, and thus faster convolutions in our implementation. In particular, if r ≥ 0, Equation (12) is chosen for better efficiency; otherwise, Equation (13) should be chosen. 4.2. Constructing a Dependency Tree Figure 3. As highlighted in the rounded rectangles, with high probability, repetitive FADD exists in the direct implementation of binary tensor convolutions. 4.1. Associative Implementation To properly avoid redundant operations, we first present an associative implementation of the multiple convolutions X ∗ B0 , ..., X ∗ Bmn−1 on L, in which the connection among different convolutions is fully exploited. To be more specific, our strategy is to perform convolutions in a hierarchical and progressive way. That is, each of the convolution results is used as a baseline of the following calculations. Suppose the j0 -th convolution is calculated in advance and it produces X∗Bj0 = s, then the convolution of X and Bj1 can be derived by using X ∗ Bj1 = s + (X ∗ (Bj0 Y Bj1 )) × 2, (12) or alternatively, X ∗ Bj1 = s − (X ∗ (¬Bj0 Y Bj1 )) × 2, (13) in which ¬ denotes the element-wise not operator, and Y denotes an element-wise operator whose behavior is in accordance with Table 1. Our implementation works by properly rearranging the binary tensors and implementing binary tensor convolutions in an indirect way. For this reason, along with Equations (12) and (13), a dependency tree is also required to drive it. In particular, dependency is the notion that certain binary tensors are linked to specify which convolution to perform first and which follows up. For instance, with the depth-first-search strategy, T in Figure 4 shows a dependency tree indicating first to calculate X∗B1 , then to derive X ∗ B0 from the previous result, then to calculate X ∗ B3 on the base of X ∗ B0 , and so forth. By traversing the whole tree, all mn convolutions will be progressively and efficiently calculated. Figure 4. A dependency tree for our algorithm. It suggests an order under which the associative convolutions are to be performed. In fact, our algorithm works with a stochastically given tree, but a dedicated T is still in demand for its optimum performance. Let G = {V, E} be an undirected weighted graph with the vertex set V and weight matrix E ∈ Rmn×mn . Each element of V represents a single binary tensor, and each element of E measures the distance Algorithm 3 The associative implementation: Input: {Bj : 0 ≤ j < mn}: the set of binary tensors, X: the input sub-feature map, T : the dependency tree. Output: {yj : 0 ≤ j < mn}: the results of convolution. Get z = T .root and calculate yz.key = X ∗ Bz.key . Initialize the baseline value by s ← yz.key . repeat Search the next node of T and update z, s. Calculate yz.key by using Equation (12) or (13). until search ends. between two chosen tensors. To keep in line with the previous subsection, we here define the distance function of the following form   t+r t−r , , (14) d(Bj0 , Bj1 ) := min 2 2 in which r = hBj0 , Bj1 i indicates the inner-product of Bj0 and Bj1 . Clearly, the defined function is a metric on {−1, +1}c×w×h and its range is restricted in [0, t]. Recall that, we expect r to be close to ±t in the previous subsection. In consequence, the optimal dependency tree should have the shortest distance from root to each of its vertices, and thus the minimum spanning tree (MST) of G is what we want. From this perspective, we can use some off-the-shell algorithms to construct such a tree. Prim’s algorithm [21] is chosen in this paper on account of its linear time complexity with respect to |E|, i.e., O(m2 n2 ) on L. With the obtained T , one can implement our algorithm easily and the whole process is summarized in Algorithm 3. Note that, although the fully-connected layers calculate vector-matrix multiplications, they can be considered as a bunch of tensor convolutions. Therefore, in the binary case, we can also make accelerations in the fully-connected layers by using Algorithm 3. 5. Experimental Results In this section, we try to empirically analyze the proposed algorithms. For pragmatic reasons, all experiments are conducted on the famous ImageNet ILSVRC-2012 database [24] with advanced CNNs and the open-source Caffe framework [13]. The training set is comprised of 1.2 million labeled images and the test set is comprised of 50,000 validation images. In Section 5.1 and 5.2, we will test the performance of the sketching algorithms (i.e., Algorithm 1 and 2) and the associative implementation of convolutions (i.e., Algorithm 3) in the sense of filter approximation and computational efficiency, respectively. Then, in Section 5.3, we evaluate the whole-net performance of our sketches and compare them with other binary-weight models. Layer Name Conv1 Conv2 Conv3 Conv4 Conv5 Fc6 Fc7 Fc8 Filters 96 256 384 384 256 1 1 1 Params (b) ∼1M ∼10M ∼28M ∼21M ∼14M ∼1208M ∼537M ∼131M FLOPs ∼211M ∼448M ∼299M ∼224M ∼150M ∼75M ∼34M ∼8M Table 2. Details of the learnable layers in AlexNet [15], in which ”Conv2” is the most computationally expensive one and ”Fc6” commits the most memory (in bits). In all these layers, FLOPs consist of the same number of FADDs and FMULs. 5.1. Efficacy of Sketching Algorithms As a starting experiment, we consider sketching the famous AlexNet model [15]. Although it is the champion solution of ILSVRC-2012, AlexNet seems to be very resource-intensive. Therefore, it is indeed appealing to seek its low-precision and efficient counterparts. As claimed in Section 1, AlexNet is an 8-layer model with 61M learnable parameters. Layer-wise details are shown in Table 2, and the pre-trained reference model is available online 3 . Using Algorithm 1 and 2, we are able to generate binaryweight AlexNets with different precisions. Theoretical analyses have been given in Section 3, so in this subsection, we shall empirically analyze the proposed algorithms. In particular, we demonstrate in Figure 6 how ”energy” accumulates with a varying size of memory P commitment for P different approximators. Defined as 1 − e2 / kW k2 , the accumulative energy is negatively correlated with reconstruction error [32], so the faster it increases, the better. In the figure, our two algorithms are abbreviately named as ”Sketching (direct)” and ”Sketching (refined)”. To compare with other strategies, we also test the stochastically generated binary basis (named ”Sketching random”) as used in [14], and the network pruning technique [8] which is not naturally orthogonal with our sketching method. The scalar factors for ”Sketching (random)” is calculated by Equation (7) to ensure its optimal performance. We can see that, it is consistent with the theoretical result that Algorithm 1 converges much slower than Algorithm 2 on all learnable layers, making it less effective on the filter approximation task. However, on the other hand, Algorithm 1 can be better when compared with ”Sketching (random)” and the pruning technique. With small working memory, our direct approximation algorithm approximates better. However, if the memory size increases, pruning technique may converge faster to its optimum. 3 https://github.com/BVLC/caffe/tree/master/ models/bvlc_alexnet. As discussed in Section 4, parameter m balances the model accuracy and efficiency in our algorithms. Figure 6 shows that, a small m (for example 3) should be adequate for AlexNet to attain over 80% of the accumulative energy in its refined sketch. Let us take layer ”Conv5” and ”Fc6” as examples and see Table 3 for more details. Layer Name Conv2 sketch Fc6 sketch Energy (%) 82.9 94.0 Params (b) ∼0.9M ∼114M FMULs ∼560K ∼12K Table 3. With only 3 bits allocated, the refined sketch of AlexNet attains over 80% of the energy on ”Conv2” and ”Fc6”, and more than 10× reduction in the committed memory for network parameters. Meanwhile, the required number of FMULs is also extremely reduced (by 400× and ∼3000×) on the two layers. 5.2. Efficiency of Associative Manipulations The associative implementation of binary tensor manipulations (i.e., Algorithm 3) is directly tested on the 3-bit refined sketch of AlexNet. To begin with, we still focus on ”Conv2 sketch” and ”Fc6 sketch”. Just to be clear, we produce the result of Algorithm 3 with both a stochastically generated dependency tree and a delicately calculated MST, while the direct implementation results are compared as a benchmark. All the implementations require the same number of FMULs, as demonstrated in Table 3, and significantly different number of FADDs, as compared in Table 4. Note that, in the associative implementations, some logical evaluations and ×2 operations are specially involved. Nevertheless, they are much less expensive than the FADDs and FMULs [22], by at least an order of magnitude. Hence, their cost are not deeply analyzed in this subsection 4 . Implementation Direct Associative (random) Associative (MST) Conv2 sketch ∼672M ∼328M ∼265M Fc6 sketch ∼113M ∼56M ∼49M Table 4. The associative implementation remarkably reduces the required number of FADDs on ”Conv2 sketch” and ”Fc6 sketch”. From the above table, we know that our associative implementation largely reduces the required number of FADDs on ”Conv2 sketch” and ”Fc6 sketch”. That is, it properly ameliorates the adverse effect of network sketching and enables us to evaluate the 3-bit sketch of AlexNet without any unbearably increase in the required amount of computation. In addition, the MST helps to further improve its performance and finally get ∼2.5× and ∼ 2.3× reduc4 Since the actual speedups varies dramatically with the architecture of processing units, so we will not measure it in the paper. Figure 5. The associative implementation of binary tensor convolutions helps to gain 2× to 3× reductions in the required number of FADDs on all learnable layers of ”AlexNet sketch”. tions on the two layers respectively. Results on all learnable layers are summarized in Figure 5. 5.3. Whole-net Performance Having tested Algorithm 1, 2 and 3 on the base of their own criteria, it is time to compare the whole-net performance of our sketch with that of other binary weight models [1, 22]. Inspired by the previous experimental results, we still use the 3-bit (direct and refined) sketches for evaluation, as they are both very efficient and accurate. Considering that the fully-connected layers of AlexNet contain more than 95% of its parameters, we should try sketching them to an extreme, namely 1 bit. Similar with Rastegari et al. [22], we also keep the ’fc8’ layer (or say, the output layer) to be of its full precision and report the top-1 and top5 inference accuracies. However, distinguished from their methods, we sketch the ’conv1’ layer as well because it is also compute-intensive (as shown in Table 1). Model Reference Sketch (ref.) Sketch (dir.) BWN [22] BC [1] Params (b) ∼1951M ∼193M ∼193M ∼190M ∼189M Top-1 (%) 57.2 55.2 54.4 53.8 35.4 Top-5 (%) 80.3 78.8 78.1 77.0 61.0 Table 5. Network sketching technique generates binary-weight AlexNets with the ability to make faithful inference and roughly 10.1× fewer parameters than its reference (in bits). Test accuracies of the competitors are cited from the paper. An updated version of BWN gains significantly improved accuracies (top-1: 56.8% and top-5: 79.4%), but the technical improvement seems unclear. Just to avoid the propagation of reconstruction errors, we need to somehow fine-tune the generated sketches. Naturally, there are two protocols to help accomplish this task; one is known as projection gradient descent and the other is stochastic gradient descent with full precision weight update [1]. We choose the latter by virtue of its better convergency. The training batch size is set as 256 and the momentum is 0.9. We let the learning rate drops every 20,000 iterations from 0.001, which is one tenth of the original Figure 6. Network sketching approximates AlexNet well enough with a much smaller amount of committed memory, and the refinement operation helps to achieve better convergency on all of its learnable layers. value as set in Krizhevsky et al.’s paper [15], and use only the center crop for accuracy evaluation. After totally 70,000 iterations (i.e., roughly 14 epochs), our sketches can make faithful inference on the test set, and the refined model is better than the direct one. As shown in Table 5, our refined sketch of AlexNet achieves a top-1 accuracy of 55.2% and a top-5 accuracy of 78.8%, which means it outperforms the recent released models in the name of BinaryConnect (BC) [1] and Binary-Weight Network (BWN) [22] by large margins, while the required number of parameters only exceeds a little bit. Network pruning technique also achieves compelling results on compressing AlexNet. However, it demands a lot of extra space for storing parameter indices, and more importantly, even the optimal pruning methods perform mediocrely on convolutional layers [8, 7]. In contrast, network sketching works sufficiently well on both of the layer types. Here we also testify its efficacy on ResNet [10]. Being equipped with much more convolutional layers than that of AlexNet, ResNet wins the ILSVRC-2015 classification competition. There are many instantiations of its architecture, and for fair comparisons, we choose the type B version of 18 layers (as with Rastegari et al. [22]). A pre-trained Torch model is available online 5 and we convert it into an equivalent Caffe model before sketching 6 . For the fine-tuning process, we set the training batch size as 64 and let the learning rate drops from 0.0001. After 200,000 iterations (i.e., roughly 10 epochs), the generated 5 https://github.com/facebook/fb.resnet.torch/ tree/master/pretrained. 6 https://github.com/facebook/fb-caffe-exts. sketch represents a top-1 accuracy of 67.8% and a top-5 accuracy of 88.4% on ImageNet dataset. Refer to Table 6 for a comparison of the classification accuracy of different binary-weight models. Model Reference Sketch (ref.) Sketch (dir.) BWN [22] Params (b) ∼374M ∼51M ∼51M ∼28M Top-1 (%) 68.8 67.8 67.3 60.8 Top-5 (%) 89.0 88.4 88.2 83.0 Table 6. Network sketching technique generates binary-weight ResNets with the ability to make faithful inference and roughly 7.4× fewer parameters than its reference (in bits). The test accuracies of BWN are directly cited from its paper. 6. Conclusions In this paper, we introduce network sketching as a novel technology for pursuing binary-weight CNNs. It is more flexible than the current available methods and it enables researchers and engineers to regulate the precision of generated sketches and get better trade-off between the model efficiency and accuracy. Both theoretical and empirical analyses have been given to validate its efficacy. Moreover, we also propose an associative implementation of binary tensor convolutions to further speedup the sketches. After all these efforts, we are able to generate binary-weight AlexNets and ResNets with the ability to make both efficient and faithful inference on the ImageNet classification task. Future works shall include exploring the sketching results of other CNNs. References [1] M. Courbariaux, Y. Bengio, and J.-P. David. BinaryConnect: Training deep neural networks with binary weights during propagations. In Advances in Neural Information Processing Systems (NIPS), 2015. [2] M. Courbariaux, I. Hubara, D. Soudry, R. El-Yaniv, and Y. Bengio. Binarized neural networks: Training deep neural networks with weights and activations constrained to +1 or -1. In arXiv preprint arXiv:1602.02830, 2016. [3] G. Davis, S. Mallat, and M. Avellaneda. Adaptive greedy approximations. Constructive approximation, 13(1):57–98, 1997. [4] M. Denil, B. Shakibi, L. Dinh, M. Ranzato, and N. d. Freitas. Predicting parameters in deep learning. In Advances in Neural Information Processing Systems (NIPS), 2013. [5] E. L. Denton, W. Zaremba, J. Bruna, Y. LeCun, and R. Fergus. Exploiting linear structure within convolutional networks for efficient evaluation. In Advances in Neural Information Processing Systems (NIPS), 2014. [6] Y. Gong, L. Liu, M. Yang, and L. Bourdev. Compressing deep convolutional networks using vector quantization. arXiv preprint arXiv:1412.6115, 2014. [7] Y. Guo, A. Yao, and Y. Chen. Dynamic network surgery for efficient DNNs. In Advances in Neural Information Processing Systems (NIPS), 2016. [8] S. Han, J. Pool, J. Tran, and W. J. Dally. Learning both weights and connections for efficient neural networks. In Advances in Neural Information Processing Systems (NIPS), 2015. [9] B. Hassibi and D. G. Stork. Second order derivatives for network pruning: Optimal brain surgeon. In Advances in Neural Information Processing Systems (NIPS), 1993. [10] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. [11] Y. Ioannou, D. Robertson, J. Shotton, R. Cipolla, and A. Criminisi. Training CNNs with low-rank filters for efficient image classification. In International Conference on Learning Representations (ICLR), 2016. [12] M. Jaderberg, A. Vedaldi, and A. Zisserman. Speeding up convolutional neural networks with low rank expansions. In British Machine Vision Conference (BMVC), 2014. [13] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. In ACM Multimedia Conference (MM), 2014. [14] F. Juefei-Xu, V. N. Boddeti, and M. Savvides. Local binary convolutional neural networks. arXiv preprint arXiv:1608.06049, 2016. [15] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (NIPS), 2012. [16] V. Lebedev, Y. Ganin, M. Rakhuba, I. Oseledets, and V. Lempitsky. Speeding-up convolutional neural networks using fine-tuned CP-decomposition. arXiv preprint arXiv:1412.6553, 2014. [17] Y. LeCun, J. S. Denker, S. A. Solla, R. E. Howard, and L. D. Jackel. Optimal brain damage. In Advances in Neural Information Processing Systems (NIPS), 1989. [18] D. D. Lin, S. S. Talathi, and V. S. Annapureddy. Fixed point quantization of deep convolutional networks. In International Conference on Machine Learning (ICML), 2016. [19] B. Liu, M. Wang, H. Foroosh, M. Tappen, and M. Pensky. Sparse convolutional neural networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. [20] M. Mathieu, M. Henaff, and Y. LeCun. Fast training of convolutional networks through FFTs. arXiv preprint arXiv:1312.5851, 2013. [21] R. C. Prim. Shortest connection networks and some generalizations. Bell system technical journal, 36(6):1389–1401, 1957. [22] M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. XNOR-Net: Imagenet classification using binary convolutional neural networks. arXiv preprint arXiv:1603.05279v2, 2016. [23] R. Rigamonti, A. Sironi, V. Lepetit, and P. Fua. Learning separable filters. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2013. [24] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211–252, 2015. [25] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations (ICLR), 2014. [26] S. Srinivas and R. V. Babu. Data-free parameter pruning for deep neural networks. In British Machine Vision Conference (BMVC), 2015. [27] C. Tai, T. Xiao, Y. Zhang, X. Wang, and W. E. Convolutional neural networks with low-rank regularization. In International Conference on Learning Representations (ICLR), 2016. [28] V. Vanhoucke, A. Senior, and M. Z. Mao. Improving the speed of neural networks on CPUs. In Deep Learning and Unsupervised Feature Learning Workshop, Advances in Neural Information Processing Systems (NIPS Workshop), 2011. [29] N. Vasilache, J. Johnson, M. Mathieu, S. Chintala, S. Piantino, and Y. LeCun. Fast convolutional nets with fbfft: A GPU performance evaluation. In International Conference on Learning Representations (ICLR), 2015. [30] A. Veit, M. Wilber, and S. Belongie. Residual networks are exponential ensembles of relatively shallow networks. In Advances in Neural Information Processing Systems (NIPS), 2016. [31] J. Wu, C. Leng, Y. Wang, Q. Hu, and J. Cheng. Quantized convolutional neural networks for mobile devices. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. [32] X. Zhang, J. Zou, X. Ming, K. He, and J. Sun. Efficient and accurate approximations of nonlinear convolutional networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015.
9cs.NE
arXiv:1710.00892v1 [cs.LG] 2 Oct 2017 Rényi Differential Privacy Mechanisms for Posterior Sampling Joseph Geumlek University of California, San Diego [email protected] Shuang Song University of California, San Diego [email protected] Kamalika Chaudhuri University of California, San Diego [email protected] October 4, 2017 Abstract With the newly proposed privacy definition of Rényi Differential Privacy (RDP) in [15], we re-examine the inherent privacy of releasing a single sample from a posterior distribution. We exploit the impact of the prior distribution in mitigating the influence of individual data points. In particular, we focus on sampling from an exponential family and specific generalized linear models, such as logistic regression. We propose novel RDP mechanisms as well as offering a new RDP analysis for an existing method in order to add value to the RDP framework. Each method is capable of achieving arbitrary RDP privacy guarantees, and we offer experimental results of their efficacy. 1 Introduction As data analysis continues to expand and permeate ever more facets of life, the concerns over the privacy of one’s data grow too. Many results have arrived in recent years to tackle the inherent conflict of extracting usable knowledge from a data set without over-extracting or leaking the private data of individuals. Before one can strike a balance between these competing goals, one needs a framework by which to quantify what it means to preserve an individual’s privacy. Since 2006, Differential Privacy (DP) has reigned as the privacy framework of choice [6]. It quantifies privacy by measuring how indistinguishable the mechanism is across whether or not any one individual is in or out of the data set. This gave not just privacy semantics, but also robust mathematical guarantees. However, the requirements have been cumbersome for utility, leading to many proposed relaxations. One common relaxation is approximate DP, which allows arbitrarily bad events to occur with probability at most δ. A more recent relaxation is Rényi Differential Privacy (RDP) proposed in [15], which uses the measure of Rényi divergences to smoothly vary between bounding the average and maximum privacy loss. However, RDP has very few mechanisms compared to the more established approximate DP. We expand the RDP repertoire with novel 1 mechanisms inspired by Rényi divergences, as well as re-analyzing an existing method in this new light. Inherent to DP and RDP is that there must be some uncertainty in the mechanism; they can not be deterministic. Many privacy methods have been motivated by exploiting pre-existing sources of randomness in machine learning algorithms. One promising area has been Bayesian data analysis, which focuses on maintaining and tracking the uncertainty within probabilistic models. Posterior sampling is prevalent in many Bayesian methods, serving to introduce randomness that matches the currently held uncertainty. We analyze the privacy arising from posterior sampling as applied to two domains: sampling from exponential family and Bayesian logistic regression. Along with these analyses, we offer tunable mechanisms that can achieve stronger privacy guarantees than directly sampling from the posterior. These mechanisms work via controlling the relative strength of the prior in determining the posterior, building off the common intuition that concentrated prior distributions can prevent overfitting in Bayesian data analysis. We experimentally validate our new methods on synthetic and real data. 2 2.1 Background Privacy Model. We say two data sets X and X0 are neighboring if they differ in the private record of a single individual or person. We use n to refer to the number of records in the data set. Definition 1. Differential Privacy (DP) [6]. A randomized mechanism A(X) is said to be (, δ)differentially private if for any subset U of the output range of A and any neighboring data sets X and X0 , we have p(A(X) ∈ U ) ≤ exp () p(A(X0 ) ∈ U ) + δ. DP is concerned with the difference the participation of a individual might have on the output distribution of the mechanism. When δ > 0, it is known as approximate DP while the δ = 0 case is known as pure DP. The requirements for DP can be phrased in terms of a privacy loss variable, a random variable that captures the effective privacy loss of the mechanism output. Definition 2. Privacy Loss Variable [2]. We can define a random variable Z that measures the privacy loss of a given output of a mechanism across two neighboring data sets X and X0 . Z = log p(A(X) = o) p(A(X0 ) = o) (1) o∼A(X) (, δ)-DP is the requirement that for any two neighboring data sets Z ≤  with probability at least 1 − δ. The exact nature of the trade-off and semantics between  and δ is subtle, and choosing them appropriately is difficult. For example, setting δ = 1/n permits (, δ)-DP mechanisms that always violate the privacy of a random individual. However, there are other ways to specify that a random variable is mostly small. One such way is to bound the Rényi divergence of A(X) and A(X0 ). Definition 3. Rényi Divergence [2]. The Rényi divergence of order λ between the two distributions P and Q is defined as Z 1 Dλ (P ||Q) = log P (o)λ Q(o)1−λ do. (2) λ−1 2 As λ → ∞, Rényi divergence becomes the max divergence; moreover, setting P = A(X) and 1 Q = A(X0 ) ensures that Dλ (P ||Q) = λ−1 log EZ [e(λ−1)Z ], where Z is the privacy loss variable. Thus, a bound on the Rényi divergence over all orders λ ∈ (0, ∞) is equivalent to (, 0)-DP, and as λ → 1, this approaches the expected value of Z equal to KL(A(X)||A(X0 )). This leads us to Rényi Differential Privacy, a flexible privacy notion that covers this intermediate behavior. Definition 4. Rényi Differential Privacy (RDP) [15]. A randomized mechanism A(X) is said to be (λ, )-Rényi differentially private if for any neighboring data sets X and X0 we have Dλ (A(X)||A(X0 )) ≤ . The choice of λ in RDP is used to tune how much concern is placed on unlikely large values of Z versus the average value of Z. One can consider a mechanism’s privacy as being quantified by the entire curve of  values associated with each order λ, but the results of [15] show that almost identical results can be achieved when this curve is known at only a finite collection of possible λ values. 2.2 Posterior Sampling. In Bayesian inference, we have a model class Θ, and are given observations x1 , . . . , xn assumed to be drawn from a θ ∈ Θ. Our goal is to maintain our beliefs about θ given the observational data in the form of the posterior distribution p(θ|x1 , . . . , xn ). This is often done in the form of drawing samples from the posterior. Our goal in this paper is to develop privacy preserving mechanisms for two popular and simple posterior sampling methods. The first is sampling from the exponential family posterior, which we address in Section 3; the second is sampling from posteriors induced by a subset of Generalized Linear Models, which we address in Section 4. 2.3 Related Work. Differential privacy has emerged as the gold standard for privacy in a number of data analysis applications – see [8, 16] for surveys. Since enforcing pure DP sometimes requires the addition of high noise, a number of relaxations have been proposed in the literature. The most popular relaxation is approximate DP [6], and a number of uniquely approximate DP mechanisms have been designed by [1, 3, 7, 17] among others. However, while this relaxation has some nice properties, recent work [13, 15] has argued that it can also lead privacy pitfalls in some cases. Approximate differential privacy is also related to, but is weaker than, the closely related δ-probabilistic privacy [12] and (1, , δ)-indistinguishability [4]. Our privacy definition of choice is Rényi differential privacy [15], which is motivated by two recent relaxations – concentrated DP [9] and z-CDP [2]. Concentrated DP has two parameters, µ and τ , controlling the mean and concentration of the privacy loss variable. Given a privacy parameter α, z-CDP essentially requires (λ, αλ)-RDP for all λ. While [2, 9, 15] establish tighter bounds on the privacy of existing differentially private and approximate DP mechanisms, we provide mechanisms based on posterior sampling from exponential families that are uniquely RDP. RDP is also a generalization of the notion of KL-privacy [20], which has been shown to be related to generalization in machine learning. There has also been some recent work on privacy properties of Bayesian posterior sampling; however most of the work has focused on establishing pure or approximate DP. [5] establishes 3 conditions under which some popular Bayesian posterior sampling procedures directly satisfy pure or approximate DP. [19] provides a pure DP way to sample from a posterior that satisfies certain mild conditions by raising the temperature. [10, 21] provide a simple statistically efficient algorithm for sampling from exponential family posteriors. [14] shows that directly sampling from the posterior of certain GLMs, such as logistic regression, with the right parameters provides approximate differential privacy. While our work draws inspiration from all [5, 14, 19], the main difference between their and our work is that we provide RDP guarantees. 3 RDP Mechanisms based on Exponential Family Posterior Sampling In this section, we analyze the Rényi divergences between distributions from the same exponential family, which will lead to our RDP mechanisms for sampling from exponential family posteriors. 3.1 Background: Exponential Families This section will give a in-depth explanation of exponential families and the properties of them we exploit in our analysis. An exponential family is a family of probability distributions over x ∈ X indexed by the parameter θ ∈ Θ ⊆ Rd that can be written in this canonical form for some choice of functions h : X → R, S : X → Rd , and A : Θ → R: ! n n Y X S(xi )) · θ − n · A(θ) . (3) p(x1 , . . . , xn |θ) = ( h(xi ))exp ( i=1 i=1 We call h the base measure, S the sufficient statistics of x, and A as the log-partition function of this family. Note that the data {x1 , . . . , xn } interact with the parameter θ solely through the dot product of θ and the sum of their sufficient statistics. When the parameter θ is used in this dot product unmodified (as in (3)), we call this a natural parameterization. Our analysis will be restricted to the families that satisfy the following two properties: Definition 5. An exponential family is minimal if the coordinates of the function S are not almost surely linearly dependent, and the interior of Θ is non-empty. Definition 6. For any for ∆ ∈ R, an exponential family is ∆-bounded if ∆ ≥ sup ||S(x) − S(y)||. (4) x,y∈X This constraint can be relaxed with some caveats explored in the appendix. When a family is minimal, the R log-partition function A has many interesting characteristics. It can be defined as A(θ) = log X h(x)exp (S(x) · θ) dx, and serves to normalize the distribution. Its derivatives form the cumulants of the distribution, that is to say ∇A(θ) = κ1 = Ex|θ [S(x)] and ∇2 A(θ) = κ2 = Ex|θ [(S(x) − κ1 )(S(x) − κ1 )| ]. This second cumulant is also the covariance of S(x), which demonstrates that A(θ) must be a convex function since covariances must be positive semidefinite. 4 In Bayesian data analysis, we are interested in finding our posterior distribution over the parameter θ that generated the data. We must introduce a prior distribution p(θ|η) to describe our initial beliefs on θ, where η is a parameterization of our family of priors. p(θ|x1 , . . . , xn , η) ∝ p(x1 , . . . , xn |θ)p(θ|η) ∝( n Y h(xi ))exp ( i=1 n X (5) ! S(xi )) · θ − n · A(θ) p(θ|η) (6) i=1 ! n X ∝ exp ( S(xi ), n) · (θ, −A(θ)) p(θ|η) (7) i=1 (8) Qn Notice that we can ignore the ( i=1 h(xi )) as it is a constant that will be normalized out. If we let our prior take the form of another exponential family p(θ|η) = exp (T (θ) · η − B(η)) where R T (θ) = (θ, −A(θ)) and B(η) = log Θ exp (T (θ) · η) dθ, the we can perform these manipulations, n X p(θ|x1 , . . . , xn , η) ∝ exp ( S(xi ), n) · T (θ) + η · T (θ) − B(η) ! (9) i=1 !   n X ∝ exp η+( S(xi ), n) · T (θ) − B(η) (10) i=1 and see that expression (10) can be written as p(θ|η 0 ) = exp (T (θ) · η 0 − C(η 0 )) (11) Pn where η 0 = η + i=1 (S(xi ), 1) and C(η 0 ) is chosen such that the distribution is normalized. This family of posteriors is precisely the same exponential family that we chose for our prior. We call this a conjugate P prior, and it offers us an efficient way of finding the parameter of our posterior: n ηposterior = ηprior + i=1 (S(xi ), 1). Within this family, T (θ) forms the sufficient statistics of θ, and the derivatives of C(η) give the cumulants of these sufficient statistics. Beta-Bernoulli System. A specific example of an exponential family that we will be interested in is the Beta-Bernoulli system, where an individual’s data is a single i.i.d. bit modeled as a Bernoulli variable with parameter ρ, along with a Beta conjugate prior. p(x1 , . . . , xn |ρ) = n Y ρxi (1 − ρ)1−xi (12) i=1 ρ Letting θ = log( 1−ρ ) and A(θ) = log(1 + exp (θ)) = − log(1 − ρ) , we can rewrite the equation as follows: 5 p(x1 , . . . , xn |ρ) = n Y ( i=1 ρ xi ) (1 − ρ) 1−ρ n X ρ = exp xi log( ) + log(1 − ρ) 1 − ρ i=1 ! n X = exp ( xi ) · θ − A(θ) . (13) ! (14) (15) i=1 This system satisfies the properties we require, as this natural parameterization with θ is both minimal and ∆-bounded for ∆ = 1. As our mechanisms are interested mainly in the posterior, the rest of this section will be written with respect the family specified by equation (11). Now that we have the notation for our distributions, we can write out the expression for the Rényi divergence of two posterior distributions P and Q (parameterized by ηP and ηQ ) from the same exponential family. This expression allows us to directly compute the Rényi divergences of posterior sampling methods, and forms the crux of the analysis of our exponential family mechanisms. Observation 7. Let P and Q be two posterior distributions from the same exponential family that are parameterized by ηP and ηQ . Then, 1 log Dλ (P ||Q) = λ−1 Z ! P (θ)λ Q(θ)1−λ dθ = Θ C(ληP + (1 − λ)ηQ ) − λC(ηP ) + C(ηQ ). (16) λ−1 To help analyze the implication of equation (16) for Rényi Differential Privacy, we define some sets of prior/posterior parameters η that arise in our analysis. R Definition 8. We say a posterior parameter η is normalizable if C(η) = log Θ exp (T (θ) · η)) dθ is finite. Let E denote the set of all normalizable η for the conjugate prior family. Definition 9. Let pset(η0 , n) be the convex hull of all parameters η of the form η0 + n(S(x), 1) for x ∈ X . When n is an integer this represents the hull of possible posterior parameters after observing n data points starting with the prior η0 . Definition 10. Let Dif f be the difference set for the family, where Dif f is the convex hull of all vectors of the form (S(x) − S(y), 0) for x, y ∈ X . Definition 11. Two posterior parameters η1 and η2 are neighboring iff η1 − η2 ∈ Dif f . They are r-neighboring iff (η1 − η2 )/r ∈ Dif f . 3.2 Mechanisms and Privacy Guarantees We begin with our simplest mechanism, Direct Sampling, which samples according to the true posterior. This mechanism is presented as Algorithm 1. Even though Algorithm 1 is generally not differentially private [5], Theorem A.3 suggests that it offers RDP for ∆-bounded exponential families and certain orders λ. 6 Algorithm 1 Direct Posterior Require: η0 , {x1 , . . . , xn } Pn 1: Sample θ ∼ p(θ|η 0 ) where η 0 = η0 + i=1 (S(xi ), 1) Theorem 12. For a ∆-bounded minimal exponential family of distributions p(x|θ) with continuous log-partition function A(θ), there exists λ∗ ∈ (1, ∞] such Algorithm 1 achieves (λ, (η0 , n, λ))-RDP for λ < λ∗ . λ∗ is the supremum over all λ such that all η in the set η0 + (λ − 1)Dif f are normalizable. Corollary 1. For the Beta-Bernoulli system with a prior Beta(α0 , β0 ), Algorithm 1 achieves (λ, )-RDP iff λ > 1 and λ < 1 + min(α0 , β0 ). Notice the implication of Corollary 1: for any η0 and n > 0, there exists finite λ such that direct posterior sampling does not guarantee (λ, )-RDP for any finite . This also prevents (, 0)-DP as an achievable goal as well. Algorithm 1 is inflexible; it offers us no way to change the privacy guarantee. This motivates us to propose two different modifications to Algorithm 1 that are capable of achieving arbitrary privacy parameters. Algorithm 2 modifies the contribution of the data X to the posterior by introducing a coefficient r, while Algorithm 3 modifies the contribution of the prior η0 by introducing a coefficient m. These simple ideas have shown up before in variations: [19] introduces a temperature scaling that acts similarly to r, while [5, 14] analyze concentration constraints for prior distributions much like our coefficient m. Algorithm 2 Diffused Posterior Require: η0 , {x1 , . . . , xn }, , λ 1: Find r ∈ (0, 1] such that ∀r-neighboring ηP , ηQ ∈ pset(η0 , rn), Dλ (p(θ|ηP )||p(θ|ηQ )) ≤  Pn 2: Sample θ ∼ p(θ|η 0 ) where η 0 = η0 + r i=1 (S(xi ), 1) Theorem 13. For any ∆-bounded minimal exponential family with prior η0 in the interior of E, any λ > 1, and any  > 0, there exists r∗ ∈ (0, 1] such that using r ∈ (0, r∗ ] in Algorithm 2 will achieve (λ, )-RDP. Algorithm 3 Concentrated Posterior Require: η0 , {x1 , . . . , xn }, , λ 1: Find m ∈ (0, 1] such that ∀ neighboring ηP , ηQ ∈ pset(η0 /m, n), Dλ (p(θ|ηP )||p(θ|ηQ )) ≤  Pn 2: Sample θ ∼ p(θ|η 0 ) where η 0 = η0 /m + i=1 (S(xi ), 1) Theorem 14. For any ∆-bounded minimal exponential family with prior η0 in the interior of E, any λ > 1, and any  > 0, there exists m∗ ∈ (0, 1] such that using m ∈ (0, m∗ ] in Algorithm 3 will achieve (λ, )-RDP. Theorems 13 and 14 can be interpreted as demonstrating that any RDP privacy level can be achieved by setting r or m arbitrarily close to zero. A small r implies a weak contribution from the data, while a small m implies a strong prior that outweighs the contribution from the data. Setting 7 r = 1 and m = 1 reduces to Algorithm 1, in which a sample is released from the true posterior without any modifications for privacy. We have not yet specified how to find the appropriate values of r or m, and the condition requires checking the supremum of divergences across the possible pset range of parameters arising as posteriors. However, with an additional assumption this supremum of divergences can be efficiently computed. Theorem 15. Let e(ηP , ηQ , λ) = Dλ (p(θ|ηP )||p(θ|ηQ )). For a fixed λ and fixed ηP , the function e is a convex function over ηQ . If for any direction v ∈ Dif f , the function gv (η) = v | ∇2 C(η)v is convex over η, then for a fixed λ, the function fλ (ηP ) = supηQ r−neighboring ηP e(ηP , ηQ , λ) is convex over ηP in the directions spanned by Dif f . Corollary 2. The Beta-Bernoulli system satisfies the conditions of Theorem 15 since the functions gv (η) have the form (v (1) )2 (ψ1 (η (1) )+ψ1 (η (2) −η (1) )), and ψ1 is the digamma function. Both pset and Dif f are defined as convex sets. The expression supr−neighboring ηP ,ηQ ∈pset(η0 ,n) Dλ (p(θ|ηP )||p(θ|ηQ )) is therefore equivalent to the maximum of Dλ (p(θ|ηP )||p(θ|ηQ )) where ηP ∈ η0 + {(0, n), (n, n)} and ηQ ∈ ηP ± (r, 0). The higher dimensional Dirichlet-Categorical system also satsifies the conditions of Theorem 15. This result is located in the appendix. We can do a binary search over (0, 1] to find an appropriate value of r or m. At each candidate value, we only need to consider the boundary situations to evaluate whether this value achieves the desired RDP privacy level. These boundary situations depend on the choice of model, and not the data size n. For example, in the Beta-Bernoulli system, evaluating the supremum involves calculating the Rényi diverengence across at most 4 pairs of distributions, as in Corollary 2. In the d dimensional Dirichlet-Categorical setting, there are O(d3 ) distribution pairs to evaluate. Eventually, the search process is guaranteed to find a non-zero choice for r or m that achieves the desired privacy level, although the utility optimality of this choice is not guaranteed. If stopped early and none of the tested candidate values satisfy the privacy constraint, the analyst can either continue to iterate or decide not to release anything. 3.3 Extension: Public Data for Exponential Families The use of a conjugate prior makes the interaction of observed data versus the prior easy to see. The prior η0 can be expressed as (αχ, α), where χ is a vector expressing the average sufficient statistics of pseudo-observations and α represents a count of these pseudo-observations. After witnessing the n data points, the posterior becomes a prior that has averaged the data sufficient statistics into a new χ and added n to α. If the data analyst had some data in addition to X that was not privacy sensitive, perhaps from a stale data set for which privacy requirements have lapsed, then this data can be used to form a better prior for the analysis. Not only would this improve utility by adding information that can be fully exploited, it would also in most cases improve the privacy guarantees as well. A stronger prior, especially a prior farther from the boundaries where C(η) becomes infinite, will lead to smaller Rényi divergences. This is effectively the same behavior as the Concentrated Sampling mechanism, which scales the prior to imagine more pseudo-observations had been seen. This also could apply to settings in which the 8 analyst can adaptively pay to receive non-private data, since this method will inform us once our prior formed from this data becomes strong enough to sample directly at our desired RDP level. This also carries another privacy implication for partial data breaches. If an adversary learns the data of some individuals in the data set, the Direct Sampling mechanism’s privacy guarantee for the remaining individuals can actually improve. Any contributions of the affected individuals to the posterior become in effect yet more public data placed in the prior. The privacy analysis and subsequent guarantees will match the setting in which this strengthened prior was used. 3.4 Extension: Releasing the result of a Statistical Query Here we are given a sensitive database X = {x1 , . . . , xn } and a predicate φ(·) which maps each xi into thePinterval [0, 1]. Our goal is to release a Rényi DP approximation to the quantity: n F (X) = n1 i=1 φ(xi ). Observe that directly releasing F (X) is neither DP nor Rényi DP, since this is a deterministic algorithm; our goal is to release a random sample from a suitable distribution so that the output is as close to F (X) as possible. The task of releasing a privatized result of a statistical query can be embedded into our BetaBernoulli system. This allows the privatized statistical query release to be done using either Algorithm 2 or Algorithm 3. We can extend the Beta-Bernoulli model to allow the sufficient statistics S(x) to range over the interval [0, 1] instead of just the discrete set {0, 1}. This alteration still results in a ∆-bounded exponential family, and the privacy results hold. The sampled posterior will be a Beta distribution that will concentrate around the mean of the data observations and the pseudo-observations of the prior. The process is described in the Beta-Sampled Statistical Query algorithm. The final transformation maps the natural parameter θ ∈ (−∞, ∞) onto the mean of the distribution ρ ∈ (0, 1). Algorithm 4 Beta-Sampled Statistical Query Require: η0 , {x1 , . . . , xn }, f, , λ 1: Compute Xf = {f (x1 ), . . . , f (xn )}. 2: Sample θ via Algorithm 2 or Algorithm 3 applied to Xf with η0 , , and λ. exp(θ) 3: Release ρ = 1+exp(θ) . 4 RDP for Generalized Linear Models with Gaussian Prior In this section, we reinterpret some existing algorithms in [14] in the light of RDP, and use ideas from [14] to provide new RDP algorithms for posterior sampling for a subset of generalized linear models with Gaussian priors. 4.1 Background: Generalized Linear Models (GLMs) The goal of generalized linear models (GLMs) is to predict an outcome y given an input vector x; y is assumed to be generated from a distribution in the exponential family whose mean depends on x through E [y|x] = g −1 (w> x), where w represents the weight of linear combination of x, and g is 9 called the link function. For example, in logistic regression, the link function g is logit and g −1 is the sigmoid function; and in linear regression, the link functions is the identity function. Learning in GLMs means learning the actual linear combination w.  Specifically, the likelihood of y given x can be written as p(y|w, x) = h(y)exp yw> x − A(w> x) , where x ∈ X , y ∈ Y, A is the log-partition function, and h(y) the scaling constant. Given a dataset D = {(x1 , y1 ), . . . , (xn , yn )} of n examples with xi ∈ X and Qyni ∈ Y, our goal is to learn the parameter w. Let p(D|w) denote p({y1 , . . . , yn }|w, {x1 , . . . , xn }) = i=1 p(yi |w, xi ). We set the prior p(w) as a multivariate Gaussian distribution with covariance Σ = (nβ)−1 I, i.e., p(w) ∼ N (0, (nβ)−1 I). The posterior distribution of w given D can be written as   n nβkwk2 Y p(D|w)p(w) R ∝ exp − p(w|D) = p(yi |w, xi ). (17) 2 p(D|w0 )p(w0 )dw0 Rd i=1 4.2 Mechanisms and Privacy Guarantees First, we introduce some assumptions that characterize the subset of GLMs and the corresponding training data on which RDP can be guaranteed. Assumption 1. 1. X is a bounded domain such that kxk2 ≤ c for all x ∈ X , and xi ∈ X for all (xi , yi ) ∈ D. 2. Y is a bounded domain such that Y ⊆ [ymin , ymax ], and yi ∈ Y for all (xi , yi ) ∈ D.. 3. g −1 has bounded range such that g −1 ∈ [γmin , γmax ]. Then, let B = max{|ymin − γmax |, |ymax − γmin |}. Example: Binary Regression with Bounded X Binary regression is used in the case where y takes value Y = {0, 1}. There arethree common types of binary regression, logistic regression with g −1 (w> x) = 1/(1 + exp −w> x ), probit regression with g −1 (w> x) = Φ(w> x) where Φ  is the Gaussian cdf, and complementary log-log regression with g −1 (w> x) = 1 − exp −exp w> x . In these three cases, Y = {0, 1}, g −1 has range (0, 1) and thus B = 1. Moreover, it is often assumed for binary regression that any example lies in a bounded domain, i.e., kxk2 ≤ c for x ∈ X . Now we establish the privacy guarantee for sampling directly from the posterior in (17) in Theorem 17. We also show that this privacy bound is tight for logistic regression; a detailed analysis is in Appendix. Theorem 16. Suppose we are given a GLM and a dataset D of size n that satisfies Assumption 1, and a Gaussian prior with covariance Σ = (nβ)−1 I, then sampling with posterior in (17) satisfies 2 2 (λ, 2cnβB λ)-RDP for all λ ≥ 1. Notice that direct posterior sampling cannot achieve (λ, )-RDP for arbitrary λ and . We next present Algorithm 5 and 6, as analogous to Algorithm 3 and 2 for exponential family respectively, that guarantee any given RDP requirement. Algorithm 5 achieves a given RDP level by setting a stronger prior, while Algorithm 6 by raising the temperature of the likelihood. Algorithm 5 Concentrated Posterior Require: Dataset D of size n; Gaussian prior with covariance (nβ0 )−1 I; (λ, ). 2 2 B λ 1: Set β = max{ 2c n , β0 } in (17). 2: Sample w ∼ p(w|D) in (17). 10 Algorithm 6 Diffuse Posterior −1 Require: Dataset D of size n; Gaussian prior with covariance (nβ) q I; (λ, ). nβ 1: Replace p(yi |w, xi ) with p(yi |w, xi )ρ in (17) where ρ = min{1, 2c2 B 2 λ }. 2: Sample w ∼ p(w|D) in (17). It follows directly from Theorem 17 that under Assumption 1, Algorithm 5 satisfies (λ, )-RDP. Theorem 17. Suppose we are given a GLM and a dataset D of size n that satisfies Assumption 1, and a Gaussian prior with covariance Σ = (nβ)−1 I, then Algorithm 6 guarantees (λ, )-RDP. In fact, it guarantees (λ̃, λ λ̃)-RDP for any λ̃ ≥ 1. We show that the RDP guarantee in Theorem 16 is tight for logistic regression. Theorem 18. For any d > 1 and any n ≥ 1, there exists neighboring datasets D and D0 , each of size n, there exists λ0 , such that for any λ > λ0 , λ-Rényi Divergence between logistic regression c2 posteriors under D and D0 with Gaussian prior is larger than 2nβ (λ − 1). 2 This implies the tightness of (λ, 2c nβ λ), the RDP guarantee of posterior sampling for logistic regression posterior. 5 Experiments In this section, we present the experimental results for our proposed algorithms for both exponential family and GLMs. Our experimental design focuses on two goals – first, analyzing the relationship between λ and  in our privacy guarantees and second, exploring the privacy-utility trade-off of our proposed methods in relation to existing methods. 5.1 Synthetic Data: Beta-Bernoulli Sampling Experiments In this section, we consider posterior sampling in the Beta-Bernoulli system. We compare three algorithms. As a baseline, we select a modified version of the algorithm in [10], which privatizes the sufficient statistic of the data to create a privatized posterior. Instead of Laplace noise that is used by[10], we use Gaussian noise to do the privatization; [15] shows that if Gaussian noise with 2 variance σ 2 is added, then this offers an RDP guarantee of (λ, λ ∆ σ 2 ) for ∆-bounded exponential families. We also consider the two algorithms presented in Section 3.2 – Algorithm 2 and 3; observe that Algorithm 1 is a special case of both. 500 iterations of binary search were used to select r and m when needed. Achievable Privacy Levels. We plot the (λ, )-RDP parameters achieved by Algorithms 2 and 3 for a few values of r and m. These parameters are plotted for a prior η0 = (6, 18) and the data size n = 100 which are selected arbitrarily for illustrative purposes. We plot over six values {0.1, 0.3, 0.5, 0.7, 0.9, 1} of the scaling constants r and m. The results are presented in Figure 1. Our primary observation is the presence of the vertical asymptotes for our proposed methods. Recall that any privacy level is achievable with our algorithms given small enough r or m; these plots demonstrate the interaction of λ and . As r and m decrease, the  guarantees improve at each λ and even become finite at larger orders λ, but a vertical asymptote still exists. The results for the baseline are not plotted: it achieves RDP along any line of positive slope passing through the origin. 11 2 2 r = 0.1 r = 0.3 r = 0.5 r = 0.7 r = 0.9 direct posterior 1.5 1 m = 0.1 m = 0.3 m = 0.5 m = 0.7 m = 0.9 direct posterior 1.5 1 0.5 0.5 0 0 0 5 10 15 20 0 5 (a) Algorithm 2 10 15 20 (b) Algorithm 3 Figure 1: Illustration of Potential (λ, )-RDP Curves for Exponential Family Sampling. 70 0.2 0 -10 -5 0 (a) KL: λ = 2 < λ∗ Alg. 2 Alg. 3 Gauss.Mech. 0.4 0.2 50 40 30 20 10 0 -10 -5 40 30 20 0 -15 (b) KL: λ = 15 > λ∗ 50 10 0 0 Alg. 2 Alg. 3 Gauss.Mech. True Post. 60 - log-likelihood 0.4 KL divergence KL divergence Alg. 2 Alg. 3 Gauss.Mech. - log-likelihood 0.6 0.6 70 Alg. 2 Alg. 3 Gauss.Mech. True Post. 60 -10 -5 0 (c) − log p(XH ): λ = 2 -15 -10 -5 0 (d) − log p(XH ): λ = 15 Figure 2: Exponential Family Synthetic Data Experiments. Privacy-Utility Tradeoff. We next evaluate the privacy-utility tradeoff of the algorithms by plotting KL(P ||A) as a function of  with λ fixed, where P is the true posterior and A is the output distribution of a mechanism. For Algorithms 2 and 3, the KL divergence can be evaluated in closed form. For the Gaussian mechanism, numerical integration was used to evaluate the KL divergence integral. We have arbitrarily chosen η0 = (6, 18) and data set X with 100 total trials and 38 successful trials. We have plotted the resulting divergences over a range of  for λ = 2 in (a) and for λ = 15 in (b) of Figure 2. When λ = 2 < λ∗ , both Algorithms 2 and 3 reach zero KL divergence once direct sampling is possible. The Gaussian mechanism must always add nonzero noise. As  → 0, Algorithm 3 approaches a point mass distribution heavily penalized by the KL divergence. Due to its projection step, the Gaussian Mechanism follows a bimodal distribution as  → 0. Algorithm 2 degrades to the prior, with modest KL divergence. When λ = 15 > λ∗ , the divergences for Algorithms 2 and 3 are bounded away from 0, while the Gaussian mechanism still approaches the truth as  → ∞. In a non-private setting, the KL divergence would be zero. Finally, we plot log p(XH |θ) as a function of , where θ comes from one of the mechanisms applied to X. Both X and XH consist of 100 Bernoulli trials with proportion parameter ρ = 0.5. This experiment was run 10000 times, and we report the mean and standard deviation. Similar to the previous section, we have a fixed prior of η0 = (6, 18). The results are shown for λ = 2 in (c) and for λ = 15 in (d) of 2. These results agree with the limit behaviors in the KL test. This experiment is more favorable for Algorithm 3, as it degrades only to the log likelihood under the mode of the prior. In this plot, we have included sampling from the true posterior as a non-private baseline. 12 5.2 Real Data: Bayesian Logistic Regression Experiments We now experiment with Bayesian logistic regression with Gaussian prior on three real datasets. We consider three algorithms – Algorithm 5 and 6, as well as the OPS algorithm proposed in [19] as a sanity check. OPS achieves pure differential privacy when the posterior has bounded support; for this algorithm, we thus truncate the Gaussian prior to make its support the L2 ball of radius c/β, which is the smallest data-independent ball guaranteed to contain the MAP classifier. Achievable Privacy Levels. We consider the achievable RDP guarantees for our algorithms and OPS under the same set of parameters β, c, ρ and B = 1. [19] shows that with the truncated prior, 2 2 OPS guarantees 4cβ ρ -differential privacy, which implies (λ, 4cβ ρ )-RDP for all λ ∈ [1, ∞]; whereas 2 2 our algorithm guarantees (λ, 2cnβρ λ)-RDP for all λ ≥ 1. Therefore our algorithm achieves better RDP guarantees at λ ≤ 2n ρ , which is quite high in practice as n is the dataset size. Privacy-Utility: Test Log-Likelihood and Error. We conduct Bayesian logistic regression on three real datasets: Abalone, Adult and MNIST. We perform binary classification tasks: abalones with less than 10 rings vs. the rest for Abalone, digit 3 vs. digit 8 for MNIST, and income ≤ 50K vs. > 50K for Adult. We encode all categorical features with one-hot encoding, resulting in 9 dimensions for Abalone, 100 dimensions for Adult and 784 dimensions in MNIST. We then scale each feature to range from [−0.5, 0.5], and normalize each example to norm 1. 1/3 of the each dataset is used for testing, and the rest for training. Abalone has 2784 training and 1393 test samples, Adult has 32561 and 16281, and MNIST has 7988 and 3994 respectively. For all algorithms, we use an original Gaussian prior with β = 10−3 . The posterior sampling is done using slice sampling with 1000 burn-in samples. Notice that slice sampling does not give samples from the exact posterior. However, a number of MCMC methods are known to converge in total variational distance in time polynomial in the data dimension for log-concave posteriors (which is the case here) [18]. Thus, provided that the burn-in period is long enough, we expect the induced distribution to be quite close, and we leave an exact RDP analysis of the MCMC sampling as future work. For privacy parameters, we set λ = 1, 10, 100 and  ∈ {e−5 , e−4 , . . . , e3 }. Figure 3 shows the test error averaged over 50 repeated runs. More experiments for test log-likelihood presented in the Appendix. We see that both Algorithm 5 and 6 achieve lower test error than OPS at all privacy levels and across all datasets. This is to be expected, since OPS guarantees pure differential privacy which is stronger than RDP. Comparing Algorithm 5 and 6, we can see that the latter always achieves better utility. 6 Conclusion The inherent randomness of posterior sampling and the mitigating influence of a prior can be made to offer a wide range of privacy guarantees. Our proposed methods outperform existing methods in specific situations. The privacy analyses of the mechanisms fit nicely into the recently introduced RDP framework, which continues to present itself as a relaxation of DP worthy of further investigation. 13 0.5 0.4 Concentrated Diffuse OPS True Posterior 0.5 0.4 0.3 Concentrated Diffuse OPS True Posterior 0.5 Test error Concentrated Diffuse OPS True Posterior Test error Test error 0.6 0.4 0.3 0.2 0.3 0.1 0.2 Test error 0.6 0 2 Concentrated Diffuse OPS True Posterior 0.5 0.4 -4 -2 0 2 Concentrated Diffuse OPS True Posterior 0.5 0.4 0.3 -4 0.3 2 Concentrated Diffuse OPS True Posterior 0.5 0.4 -4 -2 0 Concentrated Diffuse OPS True Posterior 0.5 0.4 0.3 0.3 -2 0 (a) Abalone. 2 0.3 0.2 -4 -2 0 2 Concentrated Diffuse OPS True Posterior 0.5 0.4 0.3 0.2 0.1 0.2 -4 2 Concentrated Diffuse OPS True Posterior 0.4 2 Test error Test error 0.6 0 Test error -2 0 0.1 0.2 -4 -2 0.5 Test error -2 Test error -4 -4 -2 0 (b) Adult. 2 -4 -2 0 (c) MNIST 3vs8. Figure 3: Test error vs. privacy parameter . λ = 1, 10, 100 from top to bottom. 14 2 Acknowledgements This work was partially supported by NSF under IIS 1253942, ONR under N00014-16-1-2616, and a Google Faculty Research Award. References [1] M. Bun, K. Nissim, U. Stemmer, and S. Vadhan. Differentially private release and learning of threshold functions. In Foundations of Computer Science (FOCS), 2015 IEEE 56th Annual Symposium on, pages 634–649. IEEE, 2015. [2] M. Bun and T. Steinke. Concentrated differential privacy: Simplifications, extensions, and lower bounds. In Theory of Cryptography Conference, pages 635–658. Springer, 2016. [3] K. Chaudhuri, D. Hsu, and S. Song. The large margin mechanism for differentially private maximization. In Neural Inf. Processing Systems, 2014. [4] K. Chaudhuri and N. Mishra. When random sampling preserves privacy. In Annual International Cryptology Conference, pages 198–213. Springer, 2006. [5] C. Dimitrakakis, B. Nelson, A. Mitrokotsa, and B. I. Rubinstein. Robust and private Bayesian inference. In International Conference on Algorithmic Learning Theory, pages 291–305. Springer, 2014. [6] C. Dwork, K. Kenthapadi, F. McSherry, I. Mironov, and M. Naor. Our data, ourselves: Privacy via distributed noise generation. In Annual International Conference on the Theory and Applications of Cryptographic Techniques, pages 486–503. Springer, 2006. [7] C. Dwork and J. Lei. Differential privacy and robust statistics. In Proceedings of the forty-first annual ACM symposium on Theory of computing, pages 371–380. ACM, 2009. [8] C. Dwork, A. Roth, et al. The algorithmic foundations of differential privacy, volume 9. Now Publishers, Inc., 2014. [9] C. Dwork and G. N. Rothblum. arXiv:1603.01887, 2016. Concentrated differential privacy. arXiv preprint [10] J. Foulds, J. Geumlek, M. Welling, and K. Chaudhuri. On the theory and practice of privacypreserving bayesian data analysis. arXiv preprint arXiv:1603.07294, 2016. [11] F. Liese and K.-J. Miescke. Statistical decision theory. In Statistical Decision Theory, pages 1–52. Springer, 2007. [12] A. Machanavajjhala, D. Kifer, J. Abowd, J. Gehrke, and L. Vilhuber. Privacy: Theory meets practice on the map. In Data Engineering, 2008. ICDE 2008. IEEE 24th International Conference on, pages 277–286. IEEE, 2008. [13] F. McSherry. How many secrets do you have? https://github.com/frankmcsherry/blog/ blob/master/posts/2017-02-08.md, 2017. 15 [14] K. Minami, H. Arai, I. Sato, and H. Nakagawa. Differential privacy without sensitivity. In Advances in Neural Information Processing Systems, pages 956–964, 2016. [15] I. Mironov. Rényi differential privacy. arXiv preprint arXiv:1702.07476, 2017. [16] A. D. Sarwate and K. Chaudhuri. Signal processing and machine learning with differential privacy: Algorithms and challenges for continuous data. IEEE signal processing magazine, 30(5):86–94, 2013. [17] A. G. Thakurta and A. Smith. Differentially private feature selection via stability arguments, and the robustness of the lasso. In Conference on Learning Theory, pages 819–850, 2013. [18] S. Vempala. Geometric random walks: a survey. Combinatorial and computational geometry, 52(573-612):2, 2005. [19] Y.-X. Wang, S. E. Fienberg, and A. Smola. Privacy for free: Posterior sampling and stochastic gradient Monte Carlo. Proceedings of The 32nd International Conference on Machine Learning (ICML), pages 2493––2502, 2015. [20] Y.-X. Wang, J. Lei, and S. E. Fienberg. On-average kl-privacy and its equivalence to generalization for max-entropy mechanisms. In International Conference on Privacy in Statistical Databases, pages 121–134. Springer, 2016. [21] Z. Zhang, B. Rubinstein, and C. Dimitrakakis. On the differential privacy of Bayesian inference. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence (AAAI), 2016. 16 A A.1 Appendix Proofs of Exponential Family Sampling Theorems Our proofs will make extensive use of the definitions laid out in Section 3.1. We will however need an additional definition for a modified version of pset, and as well the set of possible updates to the posterior parameter that might arise from the data. Definition A.1. Let lpset(η0 , n, b) = pset(η0 , n) + bDif f . This is the set of posterior parameters that are b-neighboring at least one of the elements of pset(η0 , n) Definition A.2. Let U be the set of posterior updates for an exponential family, where U is the convex hull of all vectors of the form (S(x), 1) for x, y ∈ X . We begin by noting that observing a data set when starting at a normalizable prior η0 must result in a normalizable posterior parameter η 0 . Observation 1. In a minimal exponential family, for any prior parameter η0 , any n > 0, and any posterior update, every possible posterior parameter in the set η0 + nU is also normalizable. As C(η) must be a convex function for minimal families, this must apply to positive non-integer values of n as well. With this observation, we are ready to prove our result on the conditions under which sampling from our posterior gives a finite (λ, )-RDP guarantee. Theorem A.3. For a ∆-bounded minimal exponential family of distributions p(x|θ) with continuous log-partition function A(θ), there exists λ∗ ∈ (1, ∞] such Algorithm 1 achieves (λ, (η0 , n, λ))-RDP for λ < λ∗ . λ∗ is the supremum over all λ such that all η in the set η0 + (λ − 1)Dif f are normalizable. Proof: P Algorithm 1 samples directly from the posterior ηpost = η0 + i (S(xi ), 1). When applied to neighboring data sets X and X0 , it selects posterior parameters that are neighboring. The theorem can be reinterpreted as saying there exists λ∗ such that for λ < λ∗ we have sup Dλ (p(θ|ηP )||p(θ|ηQ )) < ∞. (18) neighboring ηP ,ηQ ∈pset(η0 ,n) For these two posteriors from the same exponential family, we can write out the Rényi divergence in terms of the log-partition function C(η). C(ληP + (1 − λ)ηQ ) − λC(ηP ) + C(ηQ ) (19) λ−1 We wish to show that this is bounded above over all neighboring ηP and ηQ our mechanism might generate, and will do so by showing that |C(η)| must be bounded every where it is applied in equation (19) if λ < λ∗ . To find this bound, we will ultimately show each potential application of C(η) lies within a closed subset of E, from which the continuity of C will imply an upperbound. Let’s begin by observing that ηP and ηQ must lie within pset(η0 , n) as they arise as posteriors for neighboring data sets X and X0 . The point ηL = ληP + (1 − λ)ηQ = ηP + (λ − 1)(ηP − ηQ ) might not lie within pset(η0 , n). However, we know ηP − ηQ lies within Dif f and that ηL − ηP is within (λ − 1)Dif f . This means for any neighboring data sets, ηP , ηQ , and ηL lie inside lpset(η0 , n, λ − 1). Dλ (p(θ|ηP )||p(θ|ηQ )) = 17 If λ < λ∗ , then η0 + (λ − 1)Dif f ⊆ E. The set η0 + (λ − 1)Dif f is potentially an open set, but the closure of this set must be within E as well, since we can always construct λ0 ∈ (λ, λ∗ ) where η0 + (λ0 − 1)Dif f ⊆ E, and the points inside η0 + (λ − 1)Dif f can’t converge to any point outside of η0 + (λ0 − 1)Dif f . Any point in η ∈ lpset(η0 , n, λ−1) can be broken down into three components using the definition of lpset: η = η0 + u + d, where u ∈ nU and d ∈ (λ − 1)Dif f . For any point in this lpset, we can therefore subtract off the component u to reach a point in the set η0 + (λ − 1)Dif f . With Observation 1, we can conclude that η is normalizable if η − u is normalizable, and therefore the closure of lpset(η0 , n, λ − 1) is a subset of E if η0 + (λ − 1)Dif f is a subset of E, which we have shown for λ < λ∗ . As C(η) is a continuous function, we know that the supremum of |C(η)| over the closure of lpset(η0 , n, λ − 1) must be finite. Remember that for any neighboring data sets, ηP ,ηQ , and ηL are inside lpset(η0 , n, λ − 1). Since |C(η)| is bounded over this lpset, so too must our expression for Dλ (p(θ|ηP )||p(θ|ηQ )) in equation (19). Therefore there exists an upper-bound for the order λ Rényi divergence across all pairs of posterior parameters selected by Algorithm 1 on neighboring data sets. This finite upper-bound provides a finite value for (η0 , n, λ) for which Algorithm 1 offers (λ, (η0 , n, λ))-RDP .  To prove our results for Algorithm 2 and Algorithm 3, we’ll need an additional result that bounds the Rényi divergence in terms of the Hessian of the log-partition function and the distance between the two distribution parameters. Lemma 3. For λ > 1, if ||∇2 C(η)|| < H over the set {ηP + x(ηP − ηQ )|x ∈ [−λ + 1, λ − 1]}, then Dλ (p(θ|ηP )||p(θ|ηQ )) ≤ ||ηP − ηQ ||2 Hλ (20) Proof: Define the function g(x) = C(ηP + xv) where x ∈ R and v = ηP − ηQ . This allows us to rewrite the Rényi divergence as g(1 − λ) − λg(0) + g(1) λ−1 Now we will replace g with its first order Taylor expansion Dλ (P ||Q) = g(x) = g(0) + xg 0 (0) + e(x) (21) (22) where e(x) is the approximation error term, satisfying |e(x)| ≤ x2 maxy∈[−x,x] g 00 (y)/2. This results in g(0) + (1 − λ)g 0 (0) + e(1 − λ) − λg(0) + g(0) + g 0 (0) + e(1) λ−1 e(1 − λ) =− + e(1) λ−1 (λ − 1)2 ≤ max g 00 (y)/2 + max g 00 (y)/2 . λ − 1 y∈[−λ+1,λ−1] y∈[−1,1] Dλ (p(θ|ηP )||p(θ|ηQ )) = 18 (23) (24) (25) Further, we can express g 00 in terms of C and v. g 00 (y) = v | ∇2 C(ηP + yv)v 2 (26) 2 ≤ ||ηP − ηQ || ||∇ C(ηP + yv)|| (27) 2 ≤ ||ηP − ηQ || H (28) Plugging in this bound on g 00 gives the desired result. Dλ (p(θ|ηP )||p(θ|ηQ )) ≤ (λ − 1)2 λ−1 max y∈[−λ+1,λ−1] g 00 (y)/2 + max g 00 (y)/2 (29) y∈[−1,1] ≤ (λ − 1)||ηP − ηQ ||2 H/2 + ||ηP − ηQ ||2 H/2 (30) 2 (31) 2 (32) ≤ ||ηP − ηQ || Hλ/2 ≤ ||ηP − ηQ || Hλ  We will also make use of the following standard results about the Hessian of the log-partition function of minimal exponential families, given in [11] as Theorem 1.17 and Corollary 1.19 and rephrased for our purposes. Theorem 4. (Theorem 1.17 from [11]) The log-partition function C(η) of a minimal exponential family is infinitely often differentiable at parameters η in the interior of the normalizable set E. Theorem 5. (Corollary 1.19 from [11]) For minimal exponential family, the Hessian of the logpartition function ∇2 C(η) is nonsingular for every parameter η in the interior of the normalizable set E. These results imply that the Hessian ∇2 C(η) must exist and be continuous over η in the interior of E, as well as having non-zero determinant. Theorem A.4. For any ∆-bounded minimal exponential family with prior η0 in the interior of E, any λ > 1, and any  > 0, there exists r∗ ∈ (0, 1] such that using r ∈ (0, r∗ ] in Algorithm 2 will achieve (λ, )-RDP. Proof: Pn Recall that Algorithm 2 uses the posterior parameter η 0 = η0 + r i (S(x), 1) where the data contribution has been scaled by r. Our first step of this proof is to show that there exists r0 ∈ (0, 1] such that the order λ Rényi divergences of the generated parameters are finite for r < r0 . Similar to the proof of Theorem A.3, we will do so by creating a closed set where C(η) is finite and that must contain ηP , ηQ , and ηL for any choice of neighboring data sets. On neighboring data sets, this generates r-neighboring parameters ηP and ηQ . The point ηL = ληP + (1 − λ)ηQ is therefore r(λ − 1)-neighboring ηP . These points must be contained in the set lpset(η0 , rn, r(λ − 1)) = η0 + rnU + r(λ − 1)Dif f . For any point in this set, we can subtract off the component in rnU to get to a modified prior that is r(λ − 1)-neighboring η0 . By the assumption that η0 is in the interior of E, there exists δ > 0 such that the ball B(η0 , δ) ⊆ E. δ For the choice r0 = 2(λ−1)∆ , for any r ∈ (0, r0 ), the modified prior we constructed for each point in 19 lpset(η0 , rn, r(λ−1)) is within distance r(λ−1)∆ of η0 and therefore within B(η0 , δ/2) ⊂ B(η0 , δ) ⊆ E. Observation 1 then allows us to conclude that every point η in lpset(η0 , rn, r(λ − 1)) has an open neighborhood of radius δ2 where C(η) is finite. This is enough to conclude that the closure of this lpset must also lie entirely within E, and C(η) is finite and continuous over this closed set. As in Theorem A.3, this suffices to show that the supremum of order λ Rényi divergences on neighboring data sets is bounded above. We have thus shown there exists r0 where the  of our (λ, )-RDP guarantee is finite for r < r0 . However, our goal was to achieve a specific  guarantee. Our proof of the existence of r∗ centers around the claim that there must exist a bound H for the Hessian of C(η) over all choices of r ∈ [0, r0 ). We can construct the set D = ∪r∈[0,r0 ] lpset(η0 , rn, r(λ − 1)), which will contain every possible ηP , ηQ , and ηL that might arise from any pair neighboring data sets and any choice of r in that interval. The previous argument still applies: each point in this union must have an open neighborhood of radius δ/2 that is a subset of E. This is enough to conclude that closure of D is also a subset of E. Theorem 4 implies ∇2 C(η) exists and is continuous on the interior of E, and this further implies that there must exist H such that for all η in this closure we have ||∇2 C(η)|| ≤ H. For any value r, we know that ηP and ηQ are r-neighboring, so we know ||ηP − ηQ || ≤ r∆. Since D contains lpset(η0 , rn, r(λ − 1)), the bound H must apply for all η in the set {ηP + x(ηP − ηQ )|x ∈ [−λ + 1, λ − 1]}. This allows us to use Lemma 3 to get the following expression: Dλ (p(θ|ηP )||p(θ|ηQ )) ≤ ||ηP − ηQ ||2 Hλ 2 ≤ r∆ Hλ. (33) (34) If we set r∗ = ∆2Hλ , then for r < r∗ the order λ Rényi divergence of Algorithm 2 is bounded above by , which gives us the desired result.  The concentrated mechanism is a bit more subtle in how it reduces the influence of the data, and so we need this result modified from Lemmas 9 and 10 in the appendix of [10]. These results are presented here in a way that matches our notation. It effectively states that if we start at a prior η0 satisfy mild but technical regularity assumptions, then the Hessians C(kη0 ) must converge to zero as k grows. In practical terms, this implies the covariance of our prior distribution must shrink as we increase the number of pseudo-observations. Definition A.5. Let Tη∗ = T (argmaxθ∈Θ η · T (θ)). This represents the mode of the sufficient statistics under the distribution p(T (θ)|η). Lemma 6. (Lemma 9 from [10]) If A(θ) is continuously differentiable and η0 is in the interior of E, then argmaxθ∈Θ η · T (θ) must be in the interior of Θ. Lemma 7. (Lemma 10 from [10]) If we have a minimal exponential family in which A(θ) is differentiable of all orders, there exists δ1 > 0 such that the ball B(η0 , δ1 ) is a subset of E, there exists δ2 > 0 and a bound L such that all the seventh order partial derivatives of A(θ) on the set Dη0 ,δ1 ,δ2 = {θ| minη∈B(η0 ,δ1 ) ||T (θ) − Tη∗ || < δ2 } are bounded by P , and the determinant of ∇2 A(θ) is bounded away from zero on Dη0 ,δ1 ,δ2 , then there exists real number V, K such that for k > K we have ∀η ∈ B(η0 , δ1 ) ||∇2 C(kη)|| < 20 V . k (35) Theorem A.6. For any ∆-bounded minimal exponential family with prior η0 in the interior of E, for any λ > 1, and any  > 0, there exists m∗ ∈ (0, 1] such that using m ∈ (0, m∗ ] in Algorithm 3 will achieve (λ, )-RDP. Proof: 0 −1 PnFor a fixed value of m, recall that Algorithm 30 selects the posterior parameter η = m η0 + i=1 (S(xi ), 1). For neighboring data sets X and X , the selected posterior parameters ηP , ηQ , and ηL = ληP + (1 − λ)ηQ lie within lpset(m−1 η0 , n, λ − 1) = m−1 η0 + nU + (λ − 1)Dif f . We start by showing that the conditions of Lemma 7 are met. As we assumed η0 is in the interior of E, there exists δ1 > 0 such that we have the ball B(η0 , δ1 ) ⊆ E. By Theorem 4, the log-partition function of the data likelihood A(θ) is differentiable of all orders, and Theorem 5 tells us that the Hessian ∇2 A(θ) is non-singular with non-zero determinant on the interior of Θ. This permits the application of Lemma 6, offering a mapping from η in the interior of E to their mode Tη∗ corresponding to a parameter θ in the interior of Θ. Knowing that A(θ) is infinitely differentiable on the interior of Θ further implies that the seventh order derivatives are well-behaved in a neighborhood around each mode resulting from this mapping. This provides the rest of the requirements for Lemma 7. Therefore there exists V and K such that the following holds ∀η ∈ B(η0 , δ1 ) : ||∇2 C(kη)|| ≤ V . k (36) We wish to show that ||∇2 C(η)|| must be bounded on the expanded set lpset(m−1 η0 , n, λ − 1) = m η0 + nU + (λ − 1)Dif f , and will do so by showing that for small enough m we can use equation (36) to bound the Hessians. Let α(η) denote the last coordinate of η. This represents the pseudo-observation count of this parameter, and notice that ∀u ∈ U : α(u) = 1 and ∀v ∈ Dif f : α(v) = 0. We are going to analyze the scaled set cm · lpset(m−1 η0 , n, λ − 1) where cm is a positive scaling constant that will depend on m. −1 cm · lpset(m−1 η0 , n, λ − 1) = cm m−1 η0 + cm nU + cm (λ − 1)Dif f (37) . For each η in this cm · lpset, we have α(η) = cm m−1 α(η0 ) + cm n · 1 + cm (λ − 1) · 0 = cm (m−1 α(η0 ) + n) . (38) 0) −1 Setting cm = m−1α(η η0 , n, λ − 1) we have α(η0 )+n thus guarantees that for all η in cm · lpset(m α(η) = α(η0 ). We want to know how far the points in this cm ·lpset are from η0 , so we simply subtract η0 to get a set Dm of vectors. These offset vectors have the form cm · lpset(m−1 η0 , n, λ − 1) − η0 and therefore lie in the set Dm = (cm m−1 − 1)η0 + cm nU + cm (λ − 1)Dif f . Using our expression of cm as a function of m, we can see the following limiting behavior: 21 (39) α(η0 ) =0 m−1 α(η0 ) + n m−1 α(η0 ) lim cm m−1 − 1 = lim −1 − 1 = 1 − 1 = 0. m→0 m→0 m α(η0 ) + n lim cm = lim m→0 (40) m→0 (41) These limits lets us take the limit of the size of the vectors in Dm as m → 0: lim sup ||v|| ≤ lim (cm m−1 − 1)||η0 || + cm n sup ||u1 || + cm (λ − 1) m→0 v∈Dm m→0 u1 ∈U ≤ 0 · ||η0 || + 0 · sup ||u1 || + 0 · u1 ∈U sup sup ||u2 || (42) u2 ∈Dif f ||u2 || (43) u2 ∈Dif f ≤ 0. (44) This limit supremum on Dm tells us that as m → 0, the maximum distance between points in the scaled set cm · lpset(m−1 η0 , n, λ − 1) and η0 gets arbitrarily small. This means there exists some m0 such that for m < m0 the scaled set cm · lpset(m−1 η0 , n, λ − 1) lies within B(η0 , δ1 ). This scaling mapping can be inverted, and it implies lpset(m−1 η0 , n, λ − 1) is contained within c1m B(η0 , δ1 ). Being contained within this scaled ball is precisely what we need to use equation (36) with k1 = cm . Equation (36) bounds ||∇2 C(η)|| ≤ Hm = V cm for all η in lpset(m−1 η0 , n, λ − 1), which in turn lets us use Lemma 3 to bound our Rényi divergences. Dλ (p(θ|ηP )||p(θ|ηQ )) ≤ ||ηP − ηQ ||2 Hm λ 2 ≤ ∆ V cm λ. (45) (46) As we have cm → 0 as m → 0, we know there must exist m∗ such that for m < m∗ we have cm ≤ ∆2V λ . This means the order λ Rényi divergences of Algorithm 3 on neighboring data sets is bounded above by , which gives us the desired result.  We have one last theorem to prove, the result claiming the Rényi divergences of order λ between ηP and its neighbors is convex, which greatly simplifies finding the supremum of these divergences over the convex sets being considered. Theorem A.7. Let e(ηP , ηQ , λ) = Dλ (p(θ|ηP )||p(θ|ηQ )). For a fixed λ and fixed ηP , the function e is a convex function over ηQ . If for any direction v ∈ Dif f , the function gv (η) = v | ∇2 C(η)v is convex over η, then for a fixed λ, the function fλ (ηP ) = sup e(ηP , ηQ , λ) (47) ηQ r−neighboring ηP is convex over ηP in the directions spanned by Dif f . Proof: First, we can show that for a fixed ηP and fixed λ, the choice of ηQ in the supremum must lie on the boundary of possible neighbors. This is derived from showing that Dλ (P ||Q) is convex over the choice of ηQ . 22 Consider once again the expression for our Rényi divergence, expressed now as the function e(ηP , ηQ , λ): e(ηP , ηQ , λ) = Dλ (P ||Q) = C(ληP + (1 − λ)ηQ ) − λC(ηP ) + C(ηQ ). λ−1 (48) Let ∇ηQ e(ηP , ηQ , λ) denote the gradient of the divergence with respect to ηQ . 1−λ ∇C(ληP + (1 − λ)ηQ ) λ−1 = ∇C(ηQ ) − ∇C(ληP + (1 − λ)ηQ ). ∇ηQ e(ηP , ηQ , λ) = ∇C(ηQ ) + (49) (50) We can further find the Hessian with respect to ηQ : ∇2ηQ e(ηP , ηQ , λ) = ∇2 C(ηQ ) − (1 − λ)∇2 C(ληP + (1 − λ)ηQ ). (51) By virtue of being a minimal exponential family, we know C is convex and thus ∇2 C is PSD everywhere. Combined with the fact that λ > 1, this is enough to conclude that ∇2ηQ e(ηP , ηQ , λ) is also PSD for everywhere with λ > 1. This means e(ηP , ηQ , λ) is a convex function with respect to ηQ for any fixed ηP and λ. We now wish to characterize the function fλ (ηP ), which takes a supremum over ηQ ∈ ηP + rDif f of e(ηP , ηQ , λ). fλ (ηP ) = sup e(ηP , ηQ , λ) (52) ηQ r−neighboring ηP We re-parameterize this supremum in terms of the offset b = ηQ − ηP . fλ (ηP ) = sup e(ηP , ηP + b, λ) (53) b∈rDif f Now for any fixed offset b, x we can find the expression for the Hessian of ∇2ηP e(ηP , ηP + b, λ). ∇2ηP e(ηP , ηP + b, λ) = ∇2 C(ηP + b) − 1 λ ∇2 C(ηP ) + ∇2 C(ηp + (1 − λ)b) λ−1 λ−1 (54) We wish to show this Hessian is PSD, i.e. for any vector v we have v | ∇2ηP e(ηP , ηP + b, λ)v is non-negative. We can rewrite this in terms of the function gv (η) introduced in the theorem statement. λ 1 v | ∇2ηP e(ηP , ηP + b, λ)v = gv (ηP + b) − gv (ηP ) + gv (ηp + (1 − λ)b) λ−1 λ−1   λ λ−1 1 = gv (ηP + b) − gv (ηP ) + gv (ηp + (1 − λ)b) λ−1 λ λ 23 (55) (56) 120 100 Alg. 2 Alg. 3 Gauss.Mech. True Post. 100 80 80 120 40 100 - log-likelihood 60 60 50 40 30 20 0 -12 -10 -8 -6 -4 -2 0 2 (a) ρ = 1/3 (high match with η0 ) 0 -16 60 40 0 10 -14 80 20 20 -20 -16 Alg. 2 Alg. 3 Gauss.Mech. True Post. 140 70 - log-likelihood - log-likelihood 160 Alg. 2 Alg. 3 Gauss.Mech. True Post. 90 -14 -12 -10 -8 -6 -4 -2 0 (b) ρ = 1/2 2 -20 -16 -14 -12 -10 -8 -6 -4 -2 0 2 (c) ρ = 2/3 (low match with η0 ) Figure 4: Utility Comparison for a fixed η0 but varying true population parameter λ We know λ−1 > 0 and that ηP must lie between ηP + b and ηP − (λ − 1)b. Our assumption that gv (η) is convex over η for all directions v then lets us use Jensen’s inequality to see that the expression (56) must be non-negative. This lets us conclude that v | (∇2ηP e(ηP , ηP + b, λ)v ≥ 0 for all v, and thus this Hessian is PSD for any ηP . This in turn means our divergence e(ηP , ηP + b, λ) is convex over ηP assuming a fixed offset b. We return to fλ (ηP ), and observe that it is a supremum of functions that are convex, and therefore it is convex as well.  A.2 Additional Beta-Bernoulli Experiments The utility of the prior-based methods (Algorithms 2 and 3) depends on how well the prior matches the observed data. Figure 4 shows several additional situations for the experimental procedure of measuring the log-likelihood of the data. In each case, the prior η0 = (6, 18) was used, and both X and XH had 100 data points. λ = 15 was fixed in these additional experiments. The only thing that varies is the true population parameter ρ. In (a), ρ = 1/3 closely matches the predictions of the prior η0 . In (b), ρ = 0.5, presented as an intermediate case where the prior is misleading. Finally, in (c), ρ = 2/3, which is biased in the opposite direction as the prior. In all cases, the proposed methods act conservatively in the face of high privacy, but in (a) this worst case limiting behavior still has high utility. Having a strong informative prior helps these mechanisms. The setting in which the prior is based off of a representative sample of non-private data from the same population as the private data is likely to be beneficial for Algorithms 2 and 3. One other case is presented in Figure 5, where ρ = 0.2 but the prior has been changed to η0 = (1, 2). λ is still 15, and the number of data points is still 100. This prior corresponds to the uniform prior, as it assigns equal probability to all estimated data means on (0, 1). It represents an attractive case on a non-informative prior, but also represents a situation in which privacy is difficult. In particular, λ∗ = 2 in this setting. When Algorithm 3 scales up this prior, it becomes concentrated around ρ = 0.2, so this setting also corresponds to a case where the true population parameter does not match well with the predictions from the prior. 24 140 Alg. 2 Alg. 3 Gauss.Mech. True Post. 120 100 - log-likelihood 80 60 40 20 0 -20 -16 -14 -12 -10 -8 -6 -4 -2 0 2 Figure 5: Utility Experiment for the non-informative uniform prior A.3 A.3.1 Application to other exponential families Dirichlet-Categorical The Categorical family is a higher dimension generalization of the Bernoulli family. Instead of just two possible values, (e.g. "failure" or "success", 0 or 1), a categorical variable is allowed to take any of d discrete values. The parameters of a categorical distribution assign a probability to each of the discrete values. These probabilities are constrained to sum to one in order to be a valid distribution, so this family of distribution can be described with only d − 1 parameters. Our propsed method works with this family as well, but the proof is a little more difficult due to the higher dimensions. Let the space of observations X = {1, 2, . . . , d}. The sufficient statistics of an observation x is a vector of indicator variables, S(x) = {I1 (x), . . . , Id−1 (x)}. Notice that Id (x) is not included, since it can be derived from the other coordinates of S(x). Including this last indicator variable would make Pd the family non-minimal where the sufficient statistics satsify the linear relationship i=1 Ii = 1. The conjugate prior family is the Dirichlet family. Under our construction of the conjugate prior, we want the parameter η to satisfy the relationship ηposterior = ηprior + (S(x), 1). This means that η is d dimensional, and the last coordinate of η measures an effective count of observations. Since each coordinate of S(x) is bounded by one, we also have the relationship that for any posterior, η (d) ≥ η (i) for i ∈ [d]. When d = 2, this derivation exactly matches the one from the Beta-Bernoulli system and it is ∆-bounded for ∆ = 1. √ For d > 2, this family is ∆-bounded for ∆ = 2. For any two observations, S(x) − S(y) is non-zero in atmost two locations, and each location has a difference of at most 1. Further, this Dirichlet-Categorical system satisfies the requirements of Theorem A.7. The necessary requirement is that for any direction v ∈ Dif f , the function gv (η) = v | ∇2 C(η)v is convex over η. For this system, we have an expression for C(η) : ! d−1 d−1     X X C(η) = log Γ(η (k) ) + log Γ(η (d) − η (k) ) − log Γ(η (d) ) (57) k=1 i=k 25 This value is merely the sum of the log-Gamma function applied to the count of observations at each value, Pd−1minus the log-Gamma function applied to the total count of observations. The expression η (d) − i=1 η (i) evaluates to the count of observations located at the implicit dth value, since η (d) carries the total count of observations seen. With this expression, we can calculate the gradient and Hessian. The digamma function ψ0 is the derivative of the log-Gamma function log (Γ(·)), and the trigamma function ψ1 is the derivative of the digamma function.     P ψ0 η (i) − ψ0 η (d) − d−1 η (k) i 6= d   k=1 ∇C(η)(i) = (58)  P d−1 (d) (k) (d) ψ0 η − − ψ0 η i=d k=1 η    Pd−1 (k)  (i) (d)  ψ η + ψ η − i = j 6= d  1 1 k=1 η     Pd−1 (k)   (d) ψ1 η − i 6= j, i 6= d, j 6= d  k=1 η     P d−1 ∇2 C(η)(i,j) = −ψ1 η (d) − k=1 η (k) (59) i 6= j, i = d, j 6= d    Pd−1 (i)   (d)  −ψ1 η − k=1 η i 6= j, i 6= d, j = d       P  d−1 (k) ψ1 η (d) − − ψ1 (η (d) ) i = j = d k=1 η When v ∈ Dif f , the last coordinate of v is zero since changing one observation does not change the total count of observations. This means the expression gv (η) = v | ∇2 C(η)v can ignore the last coordinate of v, as well as the last row and column of ∇2 C(η). This means we are only concerned with the entries matching the first two cases of equation (59). Let ṽ denote the vector formed by the first d − 1 coordinates of v. A careful examination the matrix M equal to the top d − 1 rows and and leftmost d − 1 columns of ∇2 C(η) reveals that M decomposes as ! d−1   X (d) (k) M = ψ1 η − η [1] + diag ψ1 (η (1) ), . . . , ψ1 (η (d−1) ) (60) k=1 where [1] is the d − 1 by d − 1 matrix where all entries are 1, and diag constructs a diagonal matrix from the given values. This means for all v ∈ Dif f , we have the following expression: gv (η) = v | ∇2 C(η)v (61) = ṽ | M ṽ = ṽ | (62) ψ1 (η (d) − d−1 X ! η (k) )[1] + diag(ψ1 (η (1) , . . . , η (d−1) ) ṽ (63) k=1 = ψ1 (η (d) − d−1 X η (k) ) (ṽ | [1]ṽ) + d−1 X ψ1 (η (i) )(ṽ (i) )2 (64) i=1 k=1 (65) With the fact that [1] is PSD and that (ṽ (i) )2 is always positive, the above calculations show that gv (η) is the sum of many applications of the digamma function ψ1 . Each of these applications 26 has a positive coefficient, and the function ψ1 is convex. This concludes the proof that gv (η) is convex over η for any v ∈ Dif f . (When d = 2, this expression for gv in fact matches the one derived from the Beta-Bernoulli system.) This means that the expression for the worst-case Rényi divergence between neighboring posterior parameters is convex, and so the maximum must be located at the boundaries. In this case, the pset is a shifted simplex, so the maximum must occur at one of the vertices. The potential pairs of posterior parameters that must be checked in order to evaluate the RDP guarantee grows as O(d3 ). A.3.2 Gaussian-Gaussian and non-∆-bounded families Another interesting setting is estimating the mean of a Gaussian variable when the variance is known. In this case, the conjugate prior is also a Gaussian distribution. This system satisfies the v | ∇2 C(η)v convexity requirement for Theorem A.7, since the variance 2 ∇ C(η) is constant when the final coordinate (the total count of observations) is fixed. Thus for any v ∈ Dif f , the function gv (η) is constant and therefore convex. However, this setting does not satisfy the ∆-bounded assumption. The observations can be arbitrarily large, and changing a single observation can therefore lead to arbitrarily large changes to posterior parameters and thus also arbitrarily large Rényi divergences between neighboring data sets. The exponential family mathematics behind our results did not directly depend on the ∆boundedness assumption. Instead, this bound was used only to bound the pset of possible posterior parameters in order to bound the distance ||ηP − ηQ || when considering neighboring data sets. This bounded pset then ensured our privacy guarantee was finite. For any given data set X, we can bound the Rényi divergence between the posterior from X and the posterior from any other data set X0 satisfying S(X) − S(X0 ) ≤ ∆. This is true even when the exponential family is not ∆-bounded. This permits two different approaches: we can relax the RDP framework further, protecting only data sets and a select bounded range of neighboring data sets rather than all the neighbors, or we can include a data preprocessing step that projects the observations onto a set with bounded sufficient statistics. The latter approach permits the use of the RDP framework without introducing further relaxations. For example, we could replace the observations X with X̃ = f (X) where the following function f was applied to each observation x in X:   −∆ x ≤ −∆ f (x) = x (66) −∆ < x < ∆   ∆ ∆≤x Although the statistical model still believes arbitrarily large observations are possible, the preprocessing projection step allows us to bound ||ηP − ηQ || ≤ ∆ where ηP is the posterior for f (X) and ηQ is the posterior for f (X0 ) with any neighboring set of observations X0 . This comes with the caveat that our model no longer matches reality, since it is unaware of the distortions introduced by our preprocessing step. We leads to a potential degradation of utility for the mechanism output, but our privacy guarantees will hold. If the data altered by f is sufficiently rare, these distortions should be minimal. 27 A.4 Proofs in Section 4 A.4.1 GLMs Privacy Proof In this section we prove Theorem 16, 17. Here we state and prove a more general version of the theorems. Consider any problem with likelihood in the form ! n X p(D|w) = exp − `(w, xi , yi ) i=1 and posterior of the following form Pn exp (− i=1 ρ`(w, xi , yi )) p(w) Pn , exp (− i=1 ρ`(w0 , xi , yi )) p(w0 )dw0 Rd p(w|D) = R (67) where in the case of logistic regression, ` is the logistic loss function. Then we have the following lemma. Lemma 1. Suppose `(·, x, y) is L-Lipschitz and convex, and − log p(w) is twice differentiable and 2 2 m-strongly convex. Posterior sampling from (67) satisfies (λ, 2ρmL λ)-RDP for all λ ≥ 1. Proof. (of Lemma 1) The proof follows from the same idea as in the proof of Theorem 7 of [14]. The basic idea is that the posterior distribution p(·|D) satisfies Logarithmic Sobolev inequality (LSI), p(w|D) which implies sub-Gaussian concentration on log p(w|D 0 ) ; and sub-Gaussian concentration implies RDP. Before the proof, we define LSI and introduce the relation between sub-Gaussian concentration and LSI. Definition A.8. A distribution µ satisfies the Log-Sobolev Inquality (LSI) with constant C if for any integrable function f ,         Eµ f 2 log f 2 − Eµ f 2 log Eµ f 2 ≤ 2CEµ k∇f k2 . C. Then for every L-Lipschitz Theorem A.9. (Herbst’s Theorem) If µ satisfies LSI with constant  function f , for any λ, E [exp (λ(f − Eµ [f ]))] ≤ exp Cλ2 L2 /2 . Lemma 2. Let U : Rd → R be a twice differential, m-strongly convex and integrable function. Let µ be a probability measure on Rd whose density is proportional to exp (−U ). Then µ satisfies LSI with constant C = 1/m. Now we prove RDP bound of posterior sampling from (67). Firstly, notice that negative of log of the prior, − log p(w), is twiceP differentiable, m-strongly n convex and integrable. And therefore negative of log of the posterior, ρ i=1 `(w, xi , yi ) − log p(w) is m-strongly convex. According to Lemma 2, distribution p(w|D) satisfies LSI with constant 1/m. p(w|D) Then, set f in Theorem A.9 as f (D, D0 , w) = log p(w|D 0 ) . Since the `(·, x, y) is L-Lipschitz, we know that f (D, D0 , w) is 2ρL-Lipschitz. According to Theorem A.9, for any λ ∈ R,     2 2 2 p(w|D) 0 − D (p(w|D)kp(w|D )) ≤ e2λ ρ L /m . Ew∼p(w|D) exp λ log KL p(w|D0 ) 28 Let a = 2ρ2 L2 /m. Equivalently, then for any λ ∈ R,     p(w|D) Ew∼p(w|D) exp λ log ≤ exp aλ2 + λDKL (p(w|D)kp(w|D0 )) . p(w|D0 ) And setting λ to λ − 1, we have     p(w|D) ≤exp a(λ − 1)2 + (λ − 1)DKL (p(w|D)kp(w|D0 )) Ew∼p(w|D) exp (λ − 1) log 0 p(w|D ) ≤exp ((λ − 1) (aλ + DKL (p(w|D)kp(w|D0 )) − a)) . If λ ≥ 1, the expectation is upper bounded by    0 exp (λ − 1) aλ + max D (p(w|D)kp(w|D )) − a . KL 0 d(D,D )=1 According to the definition of zCDP in [2], this implies zCDP with ρ= 2ρ2 L2 , m ξ= max 0 d(D,D )=1 2 DKL (p(w|D)kp(w|D0 )) − 2ρ2 L2 , m 2 2 2 which is equivalent to (λ, 2ρmL λ + maxd(D,D0 )=1 DKL (p(w|D)kp(w|D0 )) − 2ρmL )-RDP for any λ ≥ 1. p(w|D) Finally, we aim at bounding DKL (p(w|D)kp(w|D0 )). Let F (w) = p(w|D 0 ) . According to the definition of KL-divergence, we have DKL (p(w|D)kp(w|D0 )) = Ep(w|D) [log F ] = Ep(w|D0 ) [F log F ] − Ep(w|D0 ) [F ] Ep(w|D0 ) [log F ] , which, by setting f = √ F in Definition A.8 and having C = 1/m, can be upper bounded by DKL (p(w|D)kp(w|D0 )) ≤ h √ i 2 Ep(w|D0 ) k∇ F k22 . m We have k∇ log F k2 =ρk∇ log p(w|D) − ∇ log p(w|D0 )k2 =ρk∇ log(p(D|w)p(w)) − ∇ log(p(D0 |w)p(w))k2 =ρk∇ log p(D|w) − ∇ log p(D0 |w)k2 ≤2ρL, and therefore √ k∇ F k22 = k∇exp  1 log F 2  √ k22 =k F F ∇ log F k22 = k∇ log F k22 ≤ ρ2 L2 F. 2 4 29 (68) So the KL-divergence in (68) is upper bounded by 2ρ2 L2 2ρ2 L2 Ep(w|D0 ) [F ] = . m m 2 2 Therefore Bayesian logistic regression satisfies (λ, 2ρmL λ)-RDP for any λ. For readers familiar with the proof of Theorem 7 in [14], the proof here is exactly the same except that the tail bound of sub-Gaussian concentration in Equation 21 and consequently 25 there are replaced by the moment generating function bound. The reason for not using the tail bound to imply moment generating function bound is because that loses constant factor. For GLMs, we have `(w, x, y) = − log h(y) + A(w> x) − yw> x, and thus ∇w `(w, x, y) = (µ − y)x = (g −1 (w> x) − y)x. Then, by the condition in Theorem 16 and 17, k∇w `(w, x, y)k2 is upper bounded by Bc and `(·, x, y) is Bc-Lipschitz. A.4.2 Logistic Regression Tightness Proof. (of Theorem 18) We aim to upper bound Z Z p(w|D)λ p(D|w)λ dw = p(w) dw × p(w|D0 )λ−1 p(D0 |w)λ−1 R [ p(w)p(D0 |w)dw]λ−1 R [ p(w)p(D|w)dw]λ (69) for all λ > 1. For convenience, here we assume Y = {−1, 1} instead of 0, 1, and thus p(y|w, x) can be written > as 1/(1 + e−yw x ). Let σ 2 = (nβ)−1 denote the variance of the Gaussian prior. Firstly, we consider the case when |D| = 1. We will extend the analysis to |D| > 1 later. Consider any x ∈ Rd with kxk = c. Let D = {(x, y)} and D0 = {(x0 , y 0 )}, where x0 = x and 0 y = −y = −1. Firstly, we prove an equation that will be used later. Let ∗j be the j-th dimension of vector ∗. For any i and d, we have   Z  1 kwk2 exp − exp −iw> x dw (70) √ d 2 2σ d 2πσ 2 R ! Z d Y wj2 1 √ exp − 2 exp (−iwj xj ) dwj = 2σ 2πσ 2 R j=1 !   Z d Y i2 x2j σ 2 1 (wj + ixj σ 2 )2 √ = exp − exp dwj 2σ 2 2 2πσ 2 R j=1 ! d Y i2 x2j σ 2 = exp 2 j=1   2 2 i σ kxk22 . =exp 2 30 Now we will consider the two terms in (69) separately. For the first term, we have Z p(D|w)λ dw p(w) p(D0 |w)λ−1 Rd    Z 1 kwk2 (1 + exp −y 0 w> x0 )λ−1 =√ dw exp − d 2σ 2 (1 + exp (−yw> x))λ 2πσ 2 Rd    Z 1 kwk2 (1 + exp w> x )λ−1 =√ dw exp − d 2σ 2 (1 + exp (−w> x))λ 2πσ 2 Rd    Z 1 kwk2 exp (λ − 1)w> x =√ exp − dw. d 2σ 2 1 + exp (−w> x) 2πσ 2 Rd (71) Let S+ be any half-space of Rd and S− = Rd \S+ . The above equals to       Z Z 1 1 kwk2 exp (λ − 1)w> x kwk2 exp (λ − 1)w> x dw + √ dw exp − exp − √ d d 2σ 2 1 + exp (−w> x) 2σ 2 1 + exp (−w> x) 2πσ 2 S+ 2πσ 2 S− For any x and any w ∈ S− , we have −yw> x = y(−w)> x. By changing variable in the second integral from w to −w, the above equals to       Z Z 1 kwk2 exp (λ − 1)w> x 1 kwk2 exp −(λ − 1)w> x exp − dw + √ exp − dw √ d d 2σ 2 1 + exp (−w> x) 2σ 2 1 + exp (w> x) 2πσ 2 S+ 2πσ 2 S+ !    Z exp −(λ − 1)w> x exp (λ − 1)w> x 1 kwk2 =√ exp − + dw d 2σ 2 1 + exp (−w> x) 1 + exp (w> x) 2πσ 2 S+   λ−1 Z X  1 kwk2 exp −iw> x (−1)i+λ−1 dw =√ exp − d 2 2σ 2πσ 2 S+ i=−λ+1     Z −1 i+λ−1 Z X  (−1) kwk2 kwk2 (−1)λ−1 > = exp −iw x dw + √ dw exp − exp − √ d d 2σ 2 2σ 2 S+ 2πσ 2 2πσ 2 S+ i=−λ+1   λ−1 X (−1)i+λ−1 Z  kwk2 exp − + exp −iw> x dw. √ d 2 2σ S+ 2πσ 2 i=1 The middle term equals to (−1)λ−1 /2. And changing variable from w to −w in the first term, the above equals to   λ−1 X (−1)−i+λ−1 Z  kwk2 = exp − exp −iw> x dw √ d 2 2σ S− 2πσ 2 i=1   λ−1 X (−1)i+λ−1 Z  kwk2 (−1)λ−1 > exp − exp −iw x dw + + √ d 2σ 2 2 S+ 2πσ 2 i=1   Z λ−1 X (−1)−i+λ−1  (−1)λ−1 kwk2 > = exp −iw x dw + . exp − √ d 2σ 2 2 Rd 2πσ 2 i=1 31 Using the equation in (70) with the fact that kxk2 = c, the above equals to λ−1 X (−1) i=1 −i+λ−1  exp i2 σ 2 c2 2   2 2 2 λ−1 X (−1)λ−1 i σ c −i+λ−1 + = . (−1) exp 2 2 i=0 Now we consider the second term in (69). We have   Z Z 1 kwk2 1 p(w)p(D|w)dw = √ exp − dw. d 2 > x) 2σ 1 + exp (−yw d 2πσ 2 R Let S+ be any half-space of Rd and S− = Rd \S+ . The above equals to     Z Z 1 1 1 kwk2 kwk2 1 exp − dw + exp − dw. √ √ d d 2 > x) 2 > x) 2σ 1 + exp (−yw 2σ 1 + exp (−yw 2 2 S+ S− 2πσ 2πσ For any x and any w ∈ S− , we have −yw> x = y(−w)> x. By changing variable in the second integral from w to −w, the above equals to     Z Z kwk2 kwk2 1 1 1 1 exp − exp − dw + dw √ √ d d 2 > 2 2σ 1 + exp (−yw x) 2σ 1 + exp (yw> x) 2πσ 2 S+ 2πσ 2 S+    Z kwk2 1 1 1 exp − =√ + dw d 2σ 2 1 + exp (−yw> x) 1 + exp (yw> x) 2πσ 2 S+   Z 1 kwk2 =√ exp − dw d 2σ 2 2πσ 2 S+ 1 = . 2 Since the above result holds for any dataset D, we know that the second term of (69) equals to 2. Combining the two terms, (69) equals to 2 λ−1 X (−1) −i+λ−1  exp i=0 i2 σ 2 c2 2  . As λ becomes sufficiently large, this is lower bounded by   (λ − 1)2 σ 2 c2 exp . 2 Now we consider the case when |D| > 1. We aim to show that there exists D and D0 , such that the same results hold for them. Let D = {(x, y), (x1 , y 1 ), . . . , (xn−1 , y n−1 )} and D0 = {(x0 , y 0 ), (x1 , y 1 ), . . . , (xn−1 , y n−1 )}, i.e., D and D0 differ in the first example. Let S1 ⊂ [d], |S1 | < d be a set of dimensions and let S2 = [d]\S1 . Let x = x0 be a vector with non-zero dimensions only on S1 , i.e., xi = 0, ∀i ∈ S2 , and kxk2 = c. Let wS∗ denote the dimensions of w that belongs to S∗ . We then have kxS1 k2 = c. Let y 0 = −y = −1. Suppose ∀j ∈ [n − 1], xj has non-zero dimensions only at {1, . . . , d}\S∗ , i.e., xji = 0, ∀i ∈ S∗ . 32 We have Z p(w)p(D|w)dw   Z Y 1 kwk2 1 1 =√ exp − dw d 2 j w > x) j w > xj ) 2σ 1 + exp (−y 1 + exp (−y 2πσ 2 j   Z kwS1 k2 1 1  dwS1 exp − =√ 2 |S1 | 2σ 1 + exp −y j wS>1 xS1 2πσ 2   Z kwS2 k2 Y 1 1   dwS2 exp − ×√ 2 |S2 | 2σ j > j 2πσ 2 j 1 + exp −y wS2 xS2   Z 1 1 kwS2 k2 Y 1   dwS2 exp − = √ 2 2πσ 2 |S2 | 2σ 2 j w > xj 1 + exp −y j S2 S2 where the second step follows from the fact that x is non-zero only at S1 and sj is non-zero only at S2 , and the last step follows from the normalization term of |D| = 1 (second term of (69)). We also have Z p(D|w)λ p(w) dw p(D0 |w)λ−1    Z kwk2 (1 + exp −y 0 w> x0 )λ−1 Y 1 1 exp − =√ dw d 2 > x))λ j w > xj ) 2σ (1 + exp (−yw 1 + exp (−y 2 2πσ j    Z kwk2 (1 + exp w> x )λ−1 Y 1 1 exp − dw =√ d 2 > λ 2σ (1 + exp (−w x)) j 1 + exp (−y j w> xj ) 2πσ 2    Z 1 kwS1 k2 (1 + exp wS>1 xS1 )λ−1  dwS1 =√ exp − |S1 | 2σ 2 (1 + exp −wS>1 xS1 )λ 2πσ 2   Z kwS2 k2 Y 1 1  dwS2  exp − ×√ 2 |S2 | 2σ j > j 2πσ 2 j 1 + exp −y wS1 xS1      Z (λ − 1)2 σ 2 c2 kwS2 k2 Y 1 1   dwS2 ≤exp exp ×√ exp − 2 |S | 2 2 2σ j > j 2πσ 2 j 1 + exp −y wS1 xS1 where the last step follows from the calculation of the first term of |D| = 1 since x0 = x, kxS1 k2 = c. Combining term for xj s cancelled and the Renyi divergence is  them 2together,  the integration (λ−1) σ 2 c2 σ 2 c2 (λ−1) 1 = O( ), the same as that at |D| = 1. λ−1 log 2exp 2 2 A.5 Additional Experiments for GLMs In this section, we present more experimental results on the same datasets. We show the negative log-likelihood at λ ∈ {1, 10, 100} in Figure 6. We can see the same trend as that in test error. Both of our proposed algorithms achieves smaller negative log-likelihood, and the diffused algorithm achieves lower negative log-likelihood than the concentrate algorithm. 33 101 2 100 -2 0 100 -4 -2 0 (a) Abalone. -4 10 2 -2 0 100 -4 10 -2 0 100 -4 -2 0 (b) Adult. Test -log(likelihood) 100 -4 2 -2 0 2 Concentrated Diffuse OPS True Posterior 101 100 2 Concentrated Diffuse OPS True Posterior 1 Concentrated Diffuse OPS True Posterior 101 2 Concentrated Diffuse OPS True Posterior 1 2 Concentrated Diffuse OPS True Posterior 101 100 Test -log(likelihood) 0 Concentrated Diffuse OPS True Posterior -4 Test -log(likelihood) -2 10 Concentrated Diffuse OPS True Posterior 1 -4 Test -log(likelihood) -4 Test -log(likelihood) 10 0 Test -log(likelihood) 101 Test -log(likelihood) Test log(-log(likelihood)) Test -log(likelihood) Concentrated Diffuse OPS True Posterior -2 0 2 Concentrated Diffuse OPS True Posterior 101 100 -4 -2 0 2 (c) MNIST 3vs8. Figure 6: Negative log-likelihood vs. privacy parameter . λ = 1, 10, 100 from top to bottom. y-axis plotted in log scale. 34
2cs.AI
FQAM-FBMC Design and Its Application to Machine Type Communication Yinan Qi, Milos Tesanovic Samsung Electronics R&D Institute UK, Staines, Middlesex TW18 4QE, UK {yinan.qi, m.tesanovic}@samsung.com Abstract—In this paper, we propose a novel waveform design which efficiently combines two air interface components: Frequency and Quadrature-Amplitude Modulation (FQAM) and Filter Bank Multicarrier (FBMC). The proposed approach takes the unique characteristics of FQAM into consideration and exploits the design of prototype filters for FBMC to effectively avoid self-interference between adjacent subcarriers in the complex domain, thus providing improved performance compared with conventional solutions in terms of selfinterference, spectrum confinement and complexity with negligible rate loss. The resulting waveform properties are proven in this paper to be particularly suitable for Machine Type Communications (MTC) devices due to the observed reduced PAPR and lowered energy consumption. MTC has created a new eco-system that gives rise to a plethora of interesting applications and new business opportunities in the fifth-Generation (5G) mobile system and services – the enabling technology for the emerging paradigm of Internet of Things (IoT). Keywords—FQAM; QAM-FBMC; 5G; PHY; waveform design; MTC; IoT; spectral efficiency; PAPR; energy consumption I. INTRODUCTION Within the pivotal objective of developing the overall 5G RAN design, air interface design will play a crucial part since its components directly impact spectrum efficiency, out-ofband emissions, throughput and transmission delay, system capacity and reliability, as well as hardware complexity and energy efficiency [1]-[2]. With ambitious goals set for 5G of supporting services with different (and often diverging) requirements, a highly flexible 5G air interface design will be required to answer this demand. As MTC has been one of key driving forces to 5G, spectrally efficient support for heterogeneous services that have quite different requirements is becoming ever so important. Accordingly, several enabling airinterface design methods have been actively investigated to support flexible spectrum sharing. Machine Type Communication (MTC), also commonly known as Machine-to-Machine (M2M), refers to exchange of information (whether wireless or wired) to and from machines without human involvement [3]-[4]. Two important components of air interface design for MTC devices are waveform and modulation schemes. The choice of waveforms plays a major role in the successful realization of 5G. Cyclic Prefix (CP)-OFDM, as a waveform widely used in current wired/wireless communication systems, has been considered as one of the candidates for 5G waveforms. CP-OFDM achieves perfect orthogonality of subcarrier signals in the complex domain that allows trivial generation of transmit signal through IFFT, trivial separation of the transmitted data symbols at the receiver through FFT, and trivial adoption to MIMO channels [5]-[8]. It is well-localized in time domain and thus suitable for delay critical applications and Time Division Duplexing (TDD) based deployments. However, the spectral efficiency of CPOFDM is reduced because of the usage of CP and it has comparatively high out of band (OOB) leakage. In this regard, FBMC is proposed as an alternative [9]-[10]. FBMC enhances the fundamental spectral efficiency because of the well-localized time/frequency traits adopted from a pulse shaping filter per subcarrier, thus reducing the overhead of guard band required to fit in the given spectrum bandwidth, while meeting the spectrum mask requirement. Furthermore, the effectively increased symbol duration is suitable for handling the multi-path fading channels even without CP overhead. Consequently, the FBMC system can reduce the inherent overheads such as CP and guard-bands in CP-OFDM. FBMC is attractive in specific asynchronous scenarios, where Coordinated Multi-Point Transmission and Reception (CoMP) and Dynamic Spectrum Access (DSA) in a fragmented spectrum are employed to support the much higher traffic demand in 5G. However, to maintain the transmission symbol rate, the conventional FBMC system generally doubles the lattice density either in time or in frequency compared with OFDM while adopting OQAM [11]. In OQAM, in-phase and quadrature-phase modulation symbols are mapped separately with half symbol duration offset. Thus, OQAM-FBMC or SMT causes intrinsic interference that makes it not straightforward to apply conventional pilot designs and corresponding channel estimation algorithms as well as MIMO schemes as in CPOFDM systems. In this regard, QAM-FBMC system which can transmit the QAM symbols is proposed to enable fundamental spectral efficiency enhancement whilst keeping the signal processing complexity low [12]-[14]. Basically, QAM-FBMC use different prototype filters to odd- and even-numbered subcarriers to maintain complex domain orthogonality. QAMFBCM is shown to be able to remove intrinsic interference and thus the implementation of channel estimation and MIMO is straightforward. Another important component is modulation. QAM is envisioned to be adopted in 5G as it provides good Euclidian distance properties and it is easy to demodulate. Instead of the conventional QAM modulation, FQAM as a combination of QAM and Frequency Shift Keying (FSK) has been proposed in conjunction with OFDMA and it has demonstrated a significant performance gain especially for the cell-edge users [15]. With Figure 1 Conventional QAM-FBMC FQAM, the statistical distribution of inter-cell interference is likely to be non-Gaussian, especially for cell-edge users. It is proved that the worst-case additive noise in wireless networks with respect to the channel capacity has a Gaussian distribution. From this observation, one can expect that the channel capacity can be increased by FQAM which makes inter-cell interference non-Gaussian. In addition, the fact that only a small number of sub-carriers are activated in FQAM results in low PAPR so that the energy efficiency is improved, thereby makes it a suitable solution for MTC devices with stringent energy consumption requirements. However, trying to harness the similar benefits by applying FQAM in an FBMC-based system through direct integration of the two might not be able to fulfill the key objectives for modulation and waveform design for MTC devices: 1) low complexity, e.g., computation complexity and implementation complexity, 2) low self-interference, 3) good spectrum confinement, i.e., low out-of-band (OOB) emission, and 4) good frequency localization, i.e., small side-lobe. In this regard, direct implementation of FQAM in current QAMFBMC may cause following problems: • different prototype filters are needed for odd- and even-numbered subcarriers, which increases the implementation complexity; • when the spectrum confinement and frequency localization are good, there is strong selfinterference in the complex domain; • when self-interference is small, spectrum is not well confined. In this paper, we propose a new method to integrate FQAM and FBCM to achieve four main objectives at the expense of only a minor transmission rate loss. The rest of the paper is organized as follows. In section II of this paper, a system model is provided. Then the proposed solutions are introduced in Section III. Section IV is devoted to analyzing the performance of the proposed approach in terms of complexity, self-interference, spectrum confinement and frequency localization. Some numerical results are presented in section V and the paper is concluded in section VI. II. SYSTEM MODEL The conventional QAM-FBMC system separates adjacent subcarriers with B filter-banks keeping near orthogonality in complex domain. The transmitted signal can be expressed as  MB −1 2π n  j s  M /B  x ( n ) =   pb [n − kM ]   Db,s [k ] e  k =−∞ b = 0  s =0    +∞ B −1 (1) where Db,s[k] is the complex data symbol on the (sB+b)-th subcarrier in the k-th symbol, M is the total number of subcarriers and pb[n] is the b-th prototype filter. The time domain filter coefficients pb[n] is given by {n|0≤n<LM}, where L is called an overlapping factor. With L≥2, the QAM-FBMC symbols have duration LM and thus partially overlap. The QAM-FBMC system can be efficiently implemented using M/B-point inverse fast Fourier transform (IFFT), BL-times repetition, and time domain base filtering as shown in Figure 1 [13] for the case of L=2. FQAM is a combination of Frequency Shift Keying (FSK) and Quadrature Amplitude Modulation (QAM) [15]. A (MF, MQ) F-QAM symbol carries log2MF + log2MQ bits, where log2MF bits represent one of the frequency tones, and log2MQ represent one of the QAM symbols as shown in Figure 2. Figure 2 FQAM Three design options are proposed in the current QAMFBMC systems [12]-[13], where different prototype filters are used for odd- and even-numbered subcarriers, i.e., B=2. As a simple extension of the QAM implementations, i.e. without taking other issues (such as spectrum confinement and complexity into consideration), FQAM can be implemented in these solutions as Figure 3 Self-interference in frequency domain for scheme 1 x (n) = +∞ B −1  k =−∞ b = 0  MB −1 2π n  j s  pb [ n − kM ]  δ fb, s ,lb, s Db, s [ k ] e M / B  (2)  s =0     where fb,s∈{0, 1, …, MF-1} is the frequency-tone index based on FSK modulation for the (sB+b)-th subcarrier, lb,s= (sB+b) modulo MF, and δf,l is the Kronecker delta function defined as 1, 0, δ f ,l =  f = l, f ≠ l. (3) III. PROPOSED APPROACH In contrast to the current solution, we propose to use the same prototype filters for all subcarriers, meaning that FQAM can be implemented as x (n) = +∞ M −1  δ k =−∞ m = 0 f m ,lm p [ n − kM ] Dm [ k ] e j 2π n m M (4) where fm∈{0, 1, …, MF-1} is the frequency-tone index based on FSK modulation for the m-th subcarrier, lm= m modulo MF. The use of the same prototype filter for odd- and evennumbered subcarriers significantly reduces implementation complexity. The prototype filter should be carefully designed to provide good spectrum confinement and therefore PHYDYAS filter is employed [9]. In traditional QAM-FBMC, the level of self-interference mainly depends on the filter shape and spacing between two adjacent subcarriers, denoted as Δf. In FQAM-FBMC, only one out of MF subcarriers is activated to carry a QAM symbol and the rest of the subcarriers are deactivated. Therefore, the level of self-interference depends on the spacing between two active subcarriers carrying QAM symbols, which is a multiple of Δf, anywhere between 1×Δf and (MF-1)×Δf. In this regard, two approaches are proposed to eliminate self-interference. A. Scheme 1: FQAM(ASK)-FBMC With PHYDYAS filter, we know that as long as the spacing between two adjacent active subcarriers is larger than 2Δf, the self-interference could be extremely small [9]. As only one out of MF subcarriers is activated to carry a QAM symbol and the rest of the subcarriers are deactivated, it follows that as long as fm≠0 or MF-1, i.e., the subcarriers on the two edges are not activated, no self-interference will be generated from FQAM symbol n to adjacent FQAM symbol n±1 as shown in Figure 3 (a), where the spacing between two activated subcarriers is 5Δf. Therefore, QAM can be used in conjunction with FSK to form a FQAM symbol. On the contrary, if the subcarriers on the edge are activated as shown in Figure 3 (b), strong complex domain selfinterference will occur between two FQAM symbols demonstrated by the overlapping part of two activated edge subcarriers. In such a case, complex domain self-interference cannot be eliminated. Therefore, ASK modulation is proposed to be applied in either real or imaginary domain only on the edge subcarriers that might cause self-interference to ensure complex domain orthogonality and the modulation is therefore FASK. When ASK is applied, the transmission rate in terms of bits/symbol is reduced (since we use ASK instead of QAM) but the rate loss could be small depending on FSK modulation order. B. Scheme 2: Opportunistic FQAM-FBMC In the scheme 1, ASK is applied as long as an active subcarrier is possible to generate self-interference. In scheme 2, the spacing between adjacent active subcarriers are taken into consideration and QAM or ASK modulation is opportunistically applied. By doing this, the rate loss can be further reduced. Instead of only transmitting in real/imaginary domain, we propose that after FSK modulation, if the subcarrier at either edge of one FQAM symbol is activated, its adjacent FQAM symbol will be checked. There is a possibility that the adjacent FQAM symbol does not activate the subcarrier that may cause self-interference as shown in Figure 4. In such a case, QAM symbols can still be applied even when fm=0 or MF, i.e., the edge subcarrier is activated. Only when two adjacent subcarriers are activated at the same time, ASK modulation in real/imaginary domain is employed as in scheme 1. The transmission rate loss is reduced to RΔ = 1 log 2 M Q MF2 ( ) (8) For (MF,MQ)=(4,4), the average rate of the proposed scheme is 3.875 bit/symbol with 3.13% rate loss compared with normal rate 4 bit/symbol. When we have (MF,MQ)=(8,4), the average rate of the proposed scheme is 4.97 bit/symbol with only 0.7% rate loss compared with normal rate 5 bit/symbol. IV. PERFORMACE COMPARISONS There are three existing QAM-FBMC solutions with different prototype filter design: Figure 4 Opportunistic FQAM-FBMC It should be noted that when the new schemes are applied, the difference between eq. (2) and (4) is Dm[k] in (4) could be either QAM modulated complex signal or ASK modulated real/imaginary signal depending on the level of selfinterference generated by the active subcarriers. In addition, the demodulation of opportunistic FQAM-FBMC should take into account the constraints imposed by the modulation. C. Rate Loss Analysis In scheme 1, FSK always carries log2(MF) bits. For edge subcarriers, ASK carries 1/2log2(MQ) and the incidence of applying ASK is 2/MF, giving an average of log2(MQ)/MF bits in total for edge subcarriers. For non-edge subcarriers, QAM carries log2(MQ) and the incidence of such a case is (MF-2)/MF, so that total (MF-2)/MF*log2(MQ) bits can be conveyed on average. Summing the two together, the average transmission rate, in bits per symbol, is given as R = log 2 ( M F ) + M F −1 log 2 M Q MF ( ) (5) The transmission rate will be reduced due to replacing QAM with ASK but the rate loss could be small with increased FSK modulation order. The reduction is • Solution 1: PHYDYAS filter and its block interleaved variant are used for odd- and even-numbered subcarriers, respectively [12]. • Solution 2: Type-I filter is used for both odd- and even-numbered subcarriers [13]. • Solution 3: Two different Type-II filters are used for odd- and even-numbered subcarriers, respectively [14]. We will compare these state-of-the-art QAM-FBMC based solutions with our proposed FQAM-FBMC scheme in terms of self-interference, spectrum confinement and complexity in this section. A. Self-interference Comparisons The signal power and the total self-interference power on (b, s)-th subcarrier can be calculated as Ps = p( b , s ),( b , s ) [0] Pi =  2 L  2 ( b ', s ' ) ≠ ( b , s ) k =− ( L −1) p( b, s ),( b ', s ') [kM ] − p( b , s ),( b, s ) [0] 1 log 2 M Q MF ( ) (6) For example, for the proposed FQAM-FBMC design, when we have (MF,MQ)=(4,4), the rate of the proposed scheme is 3.5 bit/symbol with 12.5% rate loss compared with normal rate 4 bit/symbol. When we have (MF,MQ)=(8,4), the rate of the proposed scheme is 4.75 bit/symbol with only 5% rate loss compared with normal rate 5 bit/symbol. With the opportunistic approach, the transmission rate loss can be further reduced. Using similar analysis as in the first scheme, the transmission rate is given as R = log 2 ( M F ) + M F2 −1 log 2 M Q MF2 ( ) (7) , (9) respectively, and the self-signal-to-interference ratio (self-SIR) can define as γ self = 10 log10 RΔ = 2 Ps (dB) Pi (10) Conventional solution 2 and 3 only achieve around 15 and 20dB self-SIR based on (9) and (10) because of the strong selfinterference. With relatively low self-SIR, only low order modulation schemes can be applied, e.g., 4QAM and 16 QAM. On the contrary, conventional solution 1 can achieve more than 60dB self-SIR and therefore high order modulation, e.g., 64QAM, can be applied. Since it is possible to achieve more than 60dB self-SIR based on (9) and (10), in the proposed scheme the self-interference can be very small, even negligible. This is due to the fast fall-off rate provided by PHYDYAS filter. B. Spectrum Confinement and Frequency Localization Comparisons In the first solution, PHYDYAS filter is used for oddnumbered subcarriers and a block interleaved PHYDYAS filter is used for even-numbered subcarriers for the sake of complex domain orthogonality. However, the block interleaving will severely undermine the spectrum confinement performance as shown in Figure 5. Figure 7 AWGN Channel Figure 5 PSD for the odd-numbered subcarriers The high side-lobe lead to two consequences: 1) high intersubcarrier interference in asynchronous scenario and thus imposes very high requirement for synchronization; 2) more guard band at two edges of the entire spectrum, which will cause low spectral efficiency. Both have negative impacts on the implementation of MTC devices. The second solution uses a prototype filter with PSD similar to that of PHYDYAS filter as shown in Figure 6 and the same filter is used for both even- and odd-numbered subcarriers. Thus its spectrum confinement performance is similar with the proposed scheme. Figure 8 EVA Channel increases the implementation complexity compared to the proposed scheme with only one prototype filter. Therefore, the computation complexity of solution 1 and 3 is higher than the proposed scheme. V. NUMERICAL RESULTS Figure 6 PSD of Type-I and Type-II filters The PSD of Type-II filter in solution 3 is shown in Figure 6 as well. Clearly, it has a larger side-lobe and the spectrum is not confined as well as Type-I filter and the proposed scheme. C. Complexity From the complexity perspective, both solution 1 and solution 3 use two different prototype filters, which clearly In this section, we present numerical results to compare the performance of the proposed approaches with conventional ones. Here we assume M=100, L=4 and zero forcing (ZF) equalization is used for non-Gaussian channel. Uncoded bit error rate (BER) results for both AWGN channel (given as an initial study, applicable to conventional channels with static or low-speed users) and 3GPP Extended Vehicular A (EVA) channel with speed 50 km/h (provided to cover more realistic cases and high speeds) are presented as shown in Figure 7 and 8 [16]. The modulation order of FQAM is (MF,MQ)=(4,4), which is used to for all approaches. As can be seen from the figures, the BER performance of the conventional solutions is determined by self-SIR. Solution with the highest self-SIR (over 60 dB), i.e., solution 1 achieves the best performance and performance of solution 2 and 3 is degraded because of the lower self-SIR (20 and 15 dB). The proposed approach 1, where ASK is used to avoid selfinterference for the edge subcarriers, slightly outperforms the best conventional solution because the self-interference can be completely removed. The proposed approach 2 has slightly higher BER than the conventional solution 1 and achieves the same BER as conventional solution 2. Both new approaches suffer certain level of rate loss but the loss of the second approach is negligible. More importantly, both new approaches have lower complexity and better spectrum confinement. [3] [4] [5] [6] [7] [8] Finally, a qualitative comparison of the above schemes is presented in Table-I. [9] Table-I Comparisons [10] Complexity Selfinterference Spectrum confinement Frequency localization Solution 1 Medium Low Solution 2 Low High Solution 3 High Medium New schemes Low Low Bad Good Medium Good Bad Good Medium Good [11] [12] [13] VI. CONCLUSIONS There are some major requirements for the design of airinterfaces for MTC devices: 1) low complexity, 2) low energy consumption and 3) better spectrum confinement considering MTC devices might need to operate in an asynchronous manner. In this paper, we propose novel approaches to efficiently integrate FQAM and FBMC for MTC devices to fulfill these objectives. The PAPR of FQAM is lower because only 1 subcarrier is activated out of MF subcarriers; this additionally reduces energy consumption. Whereas two different prototype filters are used for odd- and even-numbered subcarriers in conventional QAM-FBMC, only one prototype filter is employed in the new approaches to reduce implementation complexity. In addition, the proposed solutions also achieve low self-interference by utilizing the unique characteristics of FQAM. Better spectrum confinement is achieved using the PHYDYAS prototype filter. ACKNOWLEDGEMENTS This work has been performed in the framework of the H2020 project METIS-II co-funded by the EU. The views expressed are those of the authors and do not necessarily represent the project. The consortium is not liable for any use that may be made of any of the information contained therein. REFERENCES [1] [2] Samsung 5G vision white paper, Feb. 2015, available at: http://www.samsung.com/global/businessimages/insights/2015/Samsung-5G-Vision-0.pdf Milos Tesanovic et.al., “Towards a flexible harmonised 5G air interface with multi-service, multi-connectivity support”, Proc. ETSI Air Interface Workshop, Sophia-Antipolis, January 2016. [14] [15] [16] M. Zorzi, A. Gluhak, S. Lange, and A. Bassi, “From Today’s INTRAnet of Things to Future INTERnet of Things: A Wireless and Mobility Related View”, IEEE Wireless Communications Magazine, vol. 17, no.6, pp. 44-51, Dec. 2010. Y. Qi, A. Ijaz, M. Webb, Y. Morioka, A. Quddus, Y. Ma, and R. Tafazolli, "On the Physical Layer Design for Low Cost Machine Type Communication in 3GPP LTE," in Proc. of VTC fall 2014. R. Van Nee and R. Prasad, OFDM for Wireless Multimedia Communications, Boston, MA: Artech House, 2000. Y. Li and G. L. Stüber, Eds., Orthogonal Frequency Division Multiplexing for Wireless Communications. New York, NY: SpringerVerlag, 2006. Air Interface for Fixed and Mobile Broadband Wireless Access Systems, IEEE Standard 802.16e, 2005. M. Morelli, C. C. Jay Kuo, and M. O. Pun, “Synchronization techniques for orthogonal frequency division multiple access (OFDMA): A tutorial review,” Proc. IEEE, vol. 95, no. 7, pp. 1394–1427, July 2007. M. Bellanger, et. al, “FBMC physical layer: a primer,” 06 2010. [Online]. Available: http://www.ict-phydyas.org B. Farhang-Boroujeny, “OFDM Versus Filter Bank Multicarrier,” IEEE Signal Processing Magazine, vol. 23, pp. 92-112, May 2011. C. Lele, P. Siohan and R. Legouable, “Channel estimation with scattered pilots in OFDM/OQAM,” in in SPAWC 2008, Recife, Brazil., pp. 286290, Jun. 2008. Hyungju Nam, et.al., “A New Filter-Bank Multicarrier System for QAM Signal Transmission and Reception,” in Proc. IEEE Intern. Conf. Commun. (ICC’14), Jun. 2014, pp. 5227–5232. Y. Yun, C. Kim, K. Kim, Z. Ho, B. Lee, and J. Seol, "A new waveform enabling enhanced QAM-FBMC systems," in Proc. of 2015 IEEE 16th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pp.116-120, Stockholm, Sweden, June 2015. K. Kim, Y. H. Yun, C. Kim, Z. Ho, Y. Cho, and J. Seol, “Pre-processing based soft-demapper for per-tone MIMO operation in QAM-FBMC systems,” in Proc. of 2015 IEEE 26th Annual International Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC), pp. 507-511, Hong Kong, China, Aug. 2015. Sungnam Hong et al., “Frequency and Quadrature- Amplitude Modulation for Downlink Cellular OFDMA Networks,” IEEE Journal on Selected Areas in Communication, vol. 32, no. 6, pp.1256-1267, Jun. 2014. 3GPP TS 36.104. "Base Station (BS) radio transmission and reception." 3rd Generation Partnership Project; Technical Specification Group Radio Access Network; Evolved Universal Terrestrial Radio Access (EUTRA). URL: http://www.3gpp.org.
7cs.IT
1 An SDR Implementation of a Visible Light Communication System Based on the IEEE 802.15.7 Standard arXiv:1803.07337v1 [eess.SP] 20 Mar 2018 ∗ Centre Jorge Baranda∗ , Pol Henarejos∗ , Ciprian G. Gavrincea∗ Tecnològic de Telecomunicacions de Catalunya (CTTC), Castelldefels, Barcelona (Spain) Email: {jorge.baranda, pol.henarejos, ciprian.gavrincea}@cttc.es Abstract—The aim of this paper is to present an implementation of a functional IEEE 802.15.7 real-time testbed based on the Software Defined Radio (SDR) concept. This implementation is built with low cost commercial off-the-shelf (COTS) analog devices and the use of Universal Software Radio Peripheral version 2 (USRP2) equipment combined with a generic object-oriented framework that takes advantage of several open source software libraries. The prototype is validated in a controlled laboratory environment with an over-the-air measurement campaign. I. I NTRODUCTION Visible Light Communications (VLC) is an emerging research field of free space optic (FSO) communications. The evolution of high-power white light emitting diodes (LEDs) within the last decades has led to the development of low cost lighting devices with better performance in terms of both energy efficiency and life expectancy. LEDs are expected to replace incandescent light bulbs and fluorescent lamps in the next generation illumination systems. Nonetheless, these devices can be used for applications beyond the illumination purpose [1]. The use of a LED to transmit data while serving as an illumination source has captured the interest of the research community as well as global standardization efforts [2]. The use of VLC technology presents many attractive advantages in terms of available bandwidth, non-interference with radio bands (critical especially due to the scarcity in some frequency bands such as ISM), potential spatial reuse, inherent protection against eavesdropping and, all of this with power efficient devices and without high deployment costs. In the last years, the focus of the research community has been centered in improving the data rates achievable by VLC systems, obtaining performances of 1.25 Gb/s [3]. Nevertheless, although there are some standards already published for VLC communications ([4], [5]), proof-of-concept implementation examples based on these standards are scarce. In this paper, we present, to the best of our knowledge, the first implementation of a real-time VLC system prototype based on the IEEE 802.15.7 specification [5] using low cost COTS analog devices. For this development, we have followed an SDR based design approach, similar to the one presented in [6]. The SDR based approach brings a lot of advantages to develop communication system testbeds, regarding a pure hardware (HW) based implementation. Among them, we would like to highlight: flexibility, modularity, reduced financial costs and shorter development time. These advantages are due to the fact that, in an SDR-based implementation, hardware problems are turned into software problems, allowing the use of novel algorithms almost immediately, accelerating the transition from simulation to demonstration. From the variety of available commercial SDR platforms [7], we find attractive the equipment provided by Ettus Research [8], specially the USRP2 and its predecessors, because of its superior tradeoff between price and performance and, the support provided by a large user community. Finally, this paper unveils the performance of the developed VLC system with an over-theair measurement campaign. II. IEEE 802.15.7 PHY LAYER BRIEF OVERVIEW The IEEE 802.15.7 standard [5] defines the physical (PHY) and the medium access control (MAC) layer for short-range optical wireless communications using visible light for indoor and outdoor applications. One of its main distinguishing points compared to other specifications, such as the Japan Electronics and Information Technology Industries (JEITA) specifications [4], is the support for dimming and flickering mitigation techniques [9]. These aspects are very important due to the negative physiological effects that flickering may have in humans and because dimming allows power savings and energy efficiency. The standard presents three different PHY layers which are grouped by their data rates. The operation range of PHY I is from 11.67 kb/s to 266.6 kb/s, PHY II data rates are comprised between 1.25 Mb/s and 96 Mb/s, while PHY III operates from 12 Mb/s to 96 Mb/s. The modulation formats are different among PHY I/II and PHY III. While PHY I and PHY II use on-off keying (OOK) and variable pulse-position modulation (VPPM), PHY III uses a particular modulation format called color shift keying (CSK), in which information is transmitted by means of multiple optical sources. Each PHY mode includes run length limiting (RLL) codes to achieve direct current (DC) balance, avoiding data induced flickering, and forward error correction (FEC) schemes to improve the link reliability for the envisaged environments. PHY I is optimized for low data rate, transmissions over long distances, e.g., outdoor applications such as traffic and vehicle lights; and PHY II/III are designed for high data rate indoor point-to-point applications. FEC schemes proposed by 2 the standard are a combination of Reed-Solomon (RS) and convolutional codes (CC). These schemes work reasonably well in the presence of hard-decisions performed by the clock data recovery circuits (CDR) and make the interface with the proposed RLL codes (Manchester, 4B6B and 8B10B) easier. III. D EVELOPMENT METHODOLOGY The development of this prototype started with the implementation of a reference encoder-decoder model in MATLAB according to the specifications of [5] for PHY I/II operating modes. Regarding the MAC layer, only a set of features of the header specification are included to simplify the communication link. The construction of the model was useful for a better understanding of the specification and constituted an important debugging tool which helped in the posterior implementation under the FlexiCom object-oriented framework. FlexiCom framework is born as an evolution of the uPHYLA framework [10]. As uPHYLA, FlexiCom framework is intended as a tool to implement in software the physical layer of any communication system. Hence, a practical demonstration of novel signal processing and communication algorithms with real digitized signals can be performed within a short development time. Nevertheless, the new created architecture for the FlexiCom framework provides enhanced graphical capabilities and simplicity to users and developers, as depicted in Fig. 1. flow of samples in the GNU Radio environment, the effort is focused on redesigning the encoder/decoder chains, identifying the minimal processing units and introducing low-complexity algorithms which improve the system latency. Finally, we would like to highlight the importance of the testing phase to verify the successive development steps. In the MATLAB programming step, the exhaustive test of all possible combinations allows the functional verification of the encoder and the decoder implementation. Moreover, the information provided by this initial testing phase helps to accelerate the multi-level testing phase (basic block unit testing, module testing and system testing) of the implementation under the FlexiCom framework by means of the CppUnit testing framework [13]. IV. S YSTEM M ODEL The current prototype implementation supports the modulation and demodulation of all the operating and data modes defined for PHY I at [5], both for OOK and VPPM schemes. These modes are included in Table I. Based on the SDR approach used in this implementation, the prototype is divided into two main subsystems: hardware and software subsystem, as shown in Fig. 2. The hardware subsystem is constituted by the elements labeled as LED, Photodiode, USRP TX and USRP RX, while the software subsystem is composed by the elements labeled as FlexiCom TX and FlexiCom RX. TABLE I PHY I O PERATING M ODES Modulation OOK VPPM RLL Code Optical FEC clock rate Outer Code (RS) Inner Code (CC) (15,7) 1/4 (15,11) 1/3 Manchester 200 KHz (15,11) 2/3 (15,11) none none none (15,2) none (15,4) none 4B6B 400 KHz (15,7) none none none Data Operating Rate mode 11.67 kb/s 1 24.44 kb/s 2 48.89 kb/s 3 73.3 kb/s 4 100 kb/s 5 35.56 kb/s 1 71.11 kb/s 2 124.4 kb/s 3 266.6 kb/s 4 A. Hardware Subsystem Fig. 1. Graphical Interface used to configure the VLC communication layout. FlexiCom platform uses Nokia’s Qt library [11] to offer the possibility of choosing among a set of communication system layouts. Each layout presents a customized interface for system configuration and performance metric visualization. Moreover, this framework runs over most operating systems (OS) such as Windows, Linux or Mac OSX. Once the system is configured, FlexiCom uses the GNU Radio library [12] to construct the communication system and to distribute the samples among the involved blocks in a continuous flow. In the second development phase, the MATLAB code is ported to the FlexiCom framework. Due to the stream-based The analog transmitter front-end is constituted by the USRP2 platform equipped with the LFTX daughterboard, an amplification stage, the LED driver circuit and a commercial high power white LED as the light emitting source. The software subsystem generates the stream of bits which are digital to analog (DAC) converted in the USRP2. The modulated signal provided by the SDR platform is amplified to be adjusted to the levels required to control the LED driver circuit. This circuit is composed of a power MOSFET transistor (STD12NF06L) which acts in switching mode, providing the needed current to the LED fixture. The light source is a commercial phosphorescent white LED module (OSTAR R LE CW E2B), consisting of a blue LED covered by a yellow phosphor layer designed for indoor lighting applications. The light source is equipped with a reflector that reduces the 3 viewing angle of the LED and concentrates the light beam. With the current reflector, the viewing angle is reduced to 30◦ (half optical transmitted power is achieved at 15◦ from the axis of the LED device). Fig. 3. Structure of the VLC IEEE 802.15.7 software transmitter. At the transmitter, the PHR and the PSDU units are generated independently according to the system configuration. Each unit passes through a PHY I modulator block, where the corresponding FEC scheme and RLL coding are applied (see Table I). After the modulation, both units are assembled into one stream of data and the TDP pattern is inserted. In the last block, before the USRP2 device, the FLP pattern and the idle pattern are attached to form the entire frame. Fig. 2. IEEE 802.15.7 based VLC Prototype. The analog receiver front-end is constituted by the USRP2 device equipped with the LFRX daughterboard and a commercial photodetector (THORLABS R PDA36A) equipped with a spherical concentrating lens. The detector consists of a PIN silicon photodiode with an active area of 13 mm2 and an optical sensitivity of 0.2-0.4 A/W in the visible range. The PDA36A device is equipped with a transimpedance amplifier (TIA) chain circuit with manual adjustable gain, ranging between 0 and 70 dB. This setting impacts in the receiver bandwidth. But this is not the only bandwidth limiting factor of this setup. Beside the limitations of the LED fabrication process [14], the high amount of current that is required to operate (∼ 700 mA) conditions the switching speed possibilities of the led driver circuit. B. Software Subsystem The software subsystem is constituted by two general purpose processors (GPPs), one for the modulation and one for the demodulation of the signals. An IEEE 802.15.7 frame is composed of three elements: the synchronization header (SHR), the physical header (PHR) and the physical service data unit (PSDU). In order to maintain the level of luminance and to avoid inter-frame flickering, an in-band idle pattern is inserted between frames [9]. The SHR contains the fast locking pattern (FLP) and the topology dependent pattern (TDP), used to lock the CDR circuit and perform the synchronization with the incoming message. The PHR contains frame information related to the length of the data unit, the used modulation and FEC scheme. The PSDU is the data unit, and also contains the MAC Header (MHR) in which parameters such as the sequence length and control format fields are included. As the current implementation features a broadcasting system, it modulates/demodulates the MHR headers according to the format specified in [5] for data frames. The building blocks of the physical layer transmitter and receiver are depicted in Fig. 3 and Fig. 4, respectively. Fig. 4. Structure of the VLC IEEE 802.15.7 software receiver. At the receiver, frame synchronization and timing recovery are performed at the first block. Frame recovery is achieved through the exploitation of repeated patterns at the SHR header. Timing recovery is the procedure to determine the optimum sampling instant required to decide the value of the incoming bits. Due to the limitations of the USRP2 hardware in terms of clock reconfigurability, timing recovery is performed using a non-data aided detector based on the maximum likelihood (ML) algorithm [15]. According to this detector, the optimum sample corresponds to the one that maximizes the energy of the received oversampled sequence of samples. As this prototype works with a binary modulation, many transitions are produced within a frame. The oversampling allows a better reconstruction of the waveform, so the transitions can be easily detected through the maximum energy criterion. Both algorithms have a limited complexity, allowing a feasible implementation in terms of system latency. Moreover, the support of Vector-Optimized Library of Kernels (VOLK) inside the GNU Radio library allows the use of Streaming SIMD Extensions (SSE), which boosts the execution of the synchronization block by a factor of four. Once the incoming flow of bits is synchronized, the PHR and the PSDU units are extracted and demodulated according to the system configuration. Finally the bits are parsed in order to extract the information corresponding to the different units and the validity of the received frame is checked. A frame is declared invalid if there is an error in the cyclic redundancy check (CRC) field, or the frame sequence number does not correspond to the previous frame. Fig. 5 shows the distribution of the relative computational load among the receiving processing blocks obtained with the OProfile tool [16] when using Operation mode 1 and OOK modulation, which is the more demanding operating mode in terms of computer processing. The execution of 4 the receiving chain requires only 25% of the total computer processing power. The synchronization block represents 55% of the processing, while the demodulation (Manchester + CC + RS decoding + Deinterleaver operations) accounts for the 35%. The remaining 10% is required to extract the PHR and the PSDU bit streams and to parse the information included within these units. Fig. 5. Distribution of the computational load among the different receiving processing blocks. V. R ESULTS Once all the system units are tested and validated individually, the last phase of the development is dedicated to characterize the performance of the complete implementation (baseband modulation/demodulation, signal conditioning (DAC/ADC) and analog optical front-ends) using over-the-air measurements in a controlled environment. The transmitter code runs in a PC laptop with 2 GB of RAM and an Intel Core 2 Duo I3 processor clocked at 1.5 GHz. The receiver code runs in a PC laptop with 4 GB of RAM and an Intel Quad-core I5 processor clocked at 2 GHz. Additional details of the test-system are summarized in Table II. TABLE II T ESTBED S YSTEM D ETAILS Feature Operating System Platform QT Version GNU Radio Version USRP2 HW revision USRP2 Firmware version Daugtherboard model Transmitter Windows 7 H.E. FlexiCom 4.8 3.6.2 4 3.0.4 LFTX Receiver Ubuntu 12.04 FlexiCom 4.8 3.6.2 4 3.0.4 LFRX In the first experiment, measurements are taken in a medium-sized room without windows where the transmitter and the receiver are horizontally aligned and separated in order to determine the system bit-error rate (BER) performance versus the distance. In this setup, the transmitter is moved in 50 cm steps and, at least, 50000 frames of 128 bits of data are transmitted to measure the system performance. The amplification of the TIA chain at the receiver is set to 0 dB. The BER performance versus distance for each operating mode of the PHY I specification using OOK modulation is presented in Fig. 6. The points presenting a BER value of 10−6 means that no error was found during the test. 10−6 constitutes the floor of our measurement because the amount of transmitted bits in each repetition is in the order of 106 . Fig. 6. Performance of the VLC IEEE 802.15.7 prototype for different operating modes of PHY I using OOK modulation. Prior to analyze the performance of the system, a comment about the tested distances is worthy to be done. According to Section II, PHY I is thought to be used for outdoor applications. In such scenarios, the light emitter is constituted by an array of LED devices, which are capable of providing much more luminous flux than the current setup herein presented. One must remember that the current setup uses a single luminary with a maximum output of 450 lumens, according to its datasheet, which is suitable for indoor illumination systems. As expected, the system performance degrades with the distance as a consequence of a lower SNR value at the receiver site. Operation mode 1 and 2, which present the strongest channel coding schemes among the considered operating modes (see Table I), obtain the best performance. Nonetheless, as the distance between the transmitter and the receiver increases, the impact of channel coding in less robust schemes disappears and the BER performance becomes almost equivalent to the scheme without channel coding scheme (Op. mode 5), impacting in the achieved effective data rate. We define the effective data rate as the total amount of correct information bits received without taking into account the overhead introduced by the different headers (SHR, PHR and MHR) and the idle-pattern. Table III shows the measured effective data rate for each considered operating mode using OOK modulation at a distance of two meters. TABLE III M EASURED EFFECTIVE DATA RATE AT 2 METERS DISTANCE Operating Mode Op. Mode 1 Op. Mode 2 Op. Mode 3 Op. Mode 4 Op. Mode 5 Effective data rate 7.9 kb/s 18.01 kb/s 24.63 kb/s 32.38 kb/s 35.22 kb/s In the second experiment, we tested the end-to-end performance of our prototype when transmitting and receiving 5 a video coded in MPEG-TS (Transport Stream) format in a medium-sized office room with standard illumination conditions. The transmitter extracts the information from an MPEGTS stream coded at an MPEG2 video compression bit rate of 25 kb/s and audio rate of 16 kb/s, and modulates it according to the frame structure explained in Section IV-B using the coding scheme corresponding to 73.3 kb/s of data rate (see Table I). At the receiver side, the data payload is demodulated and sent through a user data protocol (UDP) socket connection to the VideoLAN streaming software [17]. Additionally, the system clock data rate is set to 500 KHz, 2.5 times faster than the IEEE 802.15.7 specification of PHY I OOK modulation in order to cope with the extra signaling of the MPEG stream and to achieve on-line displaying. The current prototype implementation supports the on-line streaming at a distance of around 1.5 m with an effective measured data rate of about 107 kb/s with a packet error rate (PER) of 1e-3. In this setup, the amplification of the TIA chain at the receiver is set to 20 dB. Fig. 7 presents one captured image from the received video signal, together with some metrics captured with the FlexiCom framework: effective data rate, number of total/lost packets, packet error rate (PER) and quantity of correct/incorrect decoded packets (PHR and PSDU units). Fig. 7. Received video capture, for a 1.5m link (from “Gangnam Style”, PSY, 2012). VI. C ONCLUSIONS AND FUTURE WORK We have presented a functional system prototype implementing all the operating modes of the PHY I specification of the IEEE 802.15.7 standard for VLC. This development is based on the SDR concept, providing the prototype with enough flexibility and modularity to enhance its capabilities without requiring long development time and high financial costs. The prototype is completed with low cost COTS analog devices to perform the transmission and reception of the light signal. The IEEE 802.15.7 standard serves as a promising platform to start developing and introducing VLC solutions into the market for low-medium data rate point-to-point and broadcast systems in the field of intelligent transportation systems (ITS) or smart indoor locating (SIL). Future system development will be focused on two aspects: a) the integration of dimming functionalities into the communication framework when using OOK modulation by inserting compensation symbols within the frame without reducing the range of the communication system; and b) further research into the analog subsystem to improve the system bandwidth and to implement the operating modes of PHY II. From the software point of view, the MATLAB simulator and the FlexiCom modules are easily customizable. ACKNOWLEDGMENT The authors would like to thank Miquel Payaró, Christian Ibars and the Access Technologies area members of CTTC for their support for the implementation of the VLC system prototype. R EFERENCES [1] T. Komine and M. Nakagawa, “Fundamental Analysis for Visible-Light Communication System using LED Lights”, IEEE Transactions on Consumer Electronics, vol. 50, no. 1, pp. 100-107, Feb. 2004. [2] J. Vucic and K.-D. Langer, “High-Speed Visible Light Communications: State-of-the-Art”, in Optical Fiber Communication Conference (OFC’12), Los Angeles, USA, Mar. 2012, OSA Technical Digest (Optical Society of America, 2012). [3] C. Kottke, J. Hilt, K. Habel, J. Vucic and K.-D. Langer, “1.25 Gbit/s Visible Light WDM Link Based on DMT Modulation of a Single RGB LED Luminary”, in Proc. European Conference on Optical Communications, Amsterdam, The Netherlands, Jun. 2012, OSA Technical Digest (Optical Society of America, 2012). [4] Japan Electronics and Information Technology Industries Association (JEITA), Visible Light Communication CP-1221, CP-1222 specifications, Japan, 2007. [5] Institute of Electrical and Electronics Engineers, “Standard for Local and metropolitan area networks, Part 15.7: Short-Range Wireless Optical Communication Using Visible Light”, Rev. 802.15.7-2011, 2011. [6] M. Rahaim, T. Borogovac, T.D.C. Little, A. Mirvakili and V. Joyner, “Demonstration of a Software Defined Visible Light Communication System”, in Demo and Exhibits of International Conference on Mobile Computing and Networking, Las Vegas, USA, Sep. 2011. [7] R. Farrell, M. Sanchez and G. Corley, “Software-Defined Radio Demonstrators: An Example and Future Trends”, International Journal of Digital Multimedia Broadcasting, vol. 2009, Article ID 547650, 12 pages, 2009. doi:10.1155/2009/547650. [8] Ettus Research LLC, “Universal Software Radio Platform” [Online]. Available: http://www.ettus.com, Retrieved: March 11, 2013. [9] R.D. Roberts, S. Rajagopal and S.-K. Lim, “IEEE 802.15.7 Physical Layer Summary”, GLOBECOM Workshops, 2nd IEEE Workshop on Optical Wireless Communications, Houston, USA, Dec. 2011, pp. 772-776. [10] J.Baranda, P.Henarejos, Y.Grunenberger, and M. Nájar, “Prototyping with SDR: a quick way to play with next-gen communications systems”, International Symposium on Wireless Communication Systems (ISWCS’11), Aachen, Germany, Nov. 2011, pp.16-20. [11] “QT Homepage” [Online]. Available: http://qt.nokia.com/products/, Retrieved: March 11, 2013. [12] “GNU Radio, the GNU Software Radio project” [Online]. Available: http://gnuradio.org/redmine/projects/gnuradio/wiki, Retrieved: March 11, 2013. [13] “CppUnit Project Page” [Online]. Available: http://sourceforge.net/projects/cppunit/, Retrieved: March 11, 2013. [14] J. Vucic, C. Kottke, S. Nerreter, A. Bttner, K.-D.Langer, and J.W. Walewski, “ White Light Wireless Transmission at 200+ Mb/s Net Data Rate by Use of Discrete-Multitone Modulation”, IEEE Photon. Technol. Lett. vol.21, no. 20, pp. 1511-13, Oct. 2009. [15] U. Mengali and A.N. D’Andrea, “ Synchronization Techniques for Digital Receivers ”, Plenum Press, NY, 1997. [16] “OProfile Homepage” [Online]. Available: http://oprofile.sourceforge.net/news/, Retrieved: March 11, 2013. [17] “VideoLAN Homepage” [Online]. Available: http://www.videolan.org/index.html, Retrieved: March 11, 2013.
7cs.IT
arXiv: 1706.00679 Testing Gaussian Process with Applications to Super-Resolution arXiv:1706.00679v2 [math.ST] 14 Nov 2017 Jean-Marc Azaïs• , Yohann De Castro?∗ and Stéphane Mourareau◦ • Institut de Mathématiques de Toulouse Université Paul Sabatier, 118 route de Narbonne, 31062 Toulouse, France ? Laboratoire de Mathématiques d’Orsay Univ. Paris-Sud, CNRS, Université Paris-Saclay, 91405 Orsay, France ∗ Équipe MOKAPLAN INRIA, 2 rue Simone Iff, 75012 Paris, France ◦ Laboratoire d’Analyse et de Mathématiques Appliquées Univ. Paris-Est, Champs sur Marne, 77454 Marne La Vallée, France Abstract: This article introduces exact testing procedures on the mean of a Gaussian process X derived from the outcomes of `1 -minimization over the space of complex valued measures. The process X can thought as the sum of two terms: first, the convolution between some kernel and a target atomic measure (mean of the process); second, a random perturbation by an additive (centered) Gaussian process. The first testing procedure considered is based on a dense sequence of grids on the index set of X and we establish that it converges (as the grid step tends to zero) to a randomized testing procedure: the decision of the test depends on the observation X and also on an independent random variable. The second testing procedure is based on the maxima and the Hessian of X in a grid-less manner. We show that both testing procedures can be performed when the variance is unknown (and the correlation function of X is known). These testing procedures can be used for the problem of deconvolution over the space of complex valued measures, and applications in frame of the Super-Resolution theory are presented. As a byproduct, numerical investigations may demonstrate that our grid-less method is more powerful (it detects sparse alternatives) than tests based on very thin grids. MSC 2010 subject classifications: Primary 62E15, 62F03, 60G15, 62H10, 62H15; secondary 60E05, 60G10, 62J05, 94A08. Keywords and phrases: Hypothesis Testing, Gaussian Process, Kac-Rice formula, Super-Resolution. Preprint version of January 8, 2018 1. Introduction 1.1. Grid-less spike detection through the “continuous” LARS New testing procedures based on the outcomes of `1 minimization methods have attracted a lot of attention in the statistical community. Of particular interest is the so-called “Spacing test ”, that we referred to as S ST , based on the Least-Angle Regression Selection (LARS), that measures the significance of the addition of a new active variable along the LARS path, see [16, Chapter 6] for further details. Specifically, one is testing the relative distance between consecutive “knots ” of the LARS, for instance λ1,P and λ2,P . The first knot λ1,P is the maximal correlation between a response variable and P predictors. The second knot λ2,P is then the correlation between some residuals and P − 1 predictors, and so on. This approach is now well 1 Azaïs, De Castro and Mourareau/1 INTRODUCTION 2 referenced among the regularized methods of high-dimensional statistics and it can be linked to minimizing the `1 -norm over P coordinates, see for instance [16, Chapter 6]. In this paper, we focus on `1 -minimization over the space of signed measures and we ask for testing procedures based on these solutions. Indeed, in deconvolution problems over the space of measures [7]—e.g., Super-Resolution or line spectral estimation [8, 14, 12, 20, 11, 3]—one may observe a noisy version of a convolution of a target discrete measure by some known kernel and one may be willing to infer on the discrete measure. In this case, testing a particular measure is encompassed by testing the mean of some “correlation” process, see Section 5 for further details. In particular, remark that there is an uncountable number of predictors—while there were P predictors previously when inferring on vectors of RP —since we are looking at correlations between a response variable and vectors indexed by a continuum, say [0, 2π). For instance, assume that one observes some moments (or Fourier coefficients) of some discrete measure on the torus [0, 2π) and one is willing to infer on its support. A strategy would be to look at correlations between the response variable and the moment curve (1, t, t2 , . . . , tn ) (or (cos(kt) ± ı sin(kt))). In this framework, the LARS algorithm does not return a sequence of entries (among P possible coordinates) and phases as in high-dimensional statistics but rather a sequence of locations (among the continuum [0, 2π)) and phases. In this paper, we invoke the LARS to this framework—we referred to it as “continuous” LARS—for which an uncountable number of active variables may enter the model. We present this extension in Section 6 defining consecutive knots (λ1 , λ2 ). One can wonder: • Can the Spacing test be used in the frame of Super-Resolution? • Is there a grid-less procedure more powerful, in the sense of detecting spikes, than the Spacing tests constructed on thin grids? 1.5 1 0.5 0 -0.5 -1 -1.5 thin grids LARS from the process X(·) ( 1, 2 , R) from the Hessian R and S Rice Grid-less approach LARS from P grid points X(·) ( 1,P , testing procedures 2,P ) from (X(tp ))P p=1 S ST Grid approach Figure 1. The grid-less approach uses the Hessian and the first two “knots” (λ1 , λ2 ) of the “continuous” LARS to build the test statistics S Rice . We compare it to the grid approach that builds a test statistics S ST using the knots (λ1,P , λ2,P ) computed from a P points grid discretization (X(tp ))P p=1 of the continuous process X. Interestingly, as we will prove, the answer is no to the first question if no modifications of the test statistic is done. Furthermore, the way that the Spacing test can be fixed to be extended to a “grid-less” frame gives a new testing procedure S Rice that accounts for the distance between 3 1 INTRODUCTION consecutive knots (λ1 , λ2 ) with respect to value of the Hessian at some maximal point, see Figure 1 for a global view on our approach. 1.2. A comparative study When the predictors are normalized, the Spacing test (ST) statistics is given by the expression S ST (λ1,P , λ2,P ) := Φ(λ1,P ) Φ(λ2,P ) where Φ = 1 − Φ is the Gaussian survival function and Φ the standard normal cumulative distribution function. In the framework of high-dimensional statistics, this statistics is exactly distributed w.r.t. a uniform law on [0, 1] under the global null, namely S ST can be considered as the observed significance [22, 4]. It is clear that one should not use this testing procedure in the Super-Resolution framework since there is no theoretical guarantees in this case. Yet the practitioner may be tempted to replace (λ1,P , λ2,P ) by (λ1 , λ2 ) given by the “continuous” LARS. Unfortunately, this paper shows that the resulting test statistics S ST is non conservative in this frame, i.e., it makes too many false rejections and one should avoid using it in practice, see the green line in Figure 2. 1 1 1 0.9 0.9 0.9 0.8 0.8 0.8 0.7 0.7 0.7 0.6 0.6 0.6 0.5 0.5 0.5 0.4 0.4 0.4 0.3 0.3 0.3 0.2 0.2 0.2 0.1 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Figure 2. [Under the null] Comparison of the empirical cumulative distribution of the two statistics S Rice (blue line, see Theorem 6) and S ST (green line) under the null hypothesis when applied to the consecutive knots (λ1 , λ2 ) given by the “continuous” LARS in both cases. The diagonal (cdf of the uniform) is represented in dotted black line. The model is described by the Super-Resolution framework (see Section 5) with cutoff frequencies fc = 3, 5, 7 from left to right. The new test statistic S Rice is exactly distributed w.r.t. the uniform law on [0, 1] under the null hypothesis. To overcome this disappointing feature, one may be willing to consider thinner and thinner grids and look at the limit as P tends to infinity. In this case, one can show that λ1,P tends to the λ1 of “continuous” LARS, but λ2,P does not converge to λ2 , it converges to λ2 as shown in (6). This results in a limit test that is a randomized version of the Spacing test that we referred to as S Grid and presented in Theorem 3. The second approach is to take a thin grid and to use S ST . This approach is perfectly valid, this test statistics follows a uniform distribution under the null and it should be compared to our new testing procedure S Rice . This numerical investigation has been performed in the frame of Super-Resolution and it is presented in Figure 3, more details can be found in Section 5.2. Figure 3 gives the cumulative distribution functions of the test statistics under “sparse” alternatives, i.e., when true spikes are to be detected. The larger the power, the better the test Azaïs, De Castro and Mourareau/1 INTRODUCTION 4 detects spikes (abscissa represents the level of the test and the ordinate the probability to detect the spike). In these sets of experiments, we can note that the testing procedure S Rice based on some Hessian and the whole process X(·) is uniformly better than the spacing test even if one takes very thin grids. One can see that the power (the ability to detect sparse objects, Dirac masses here) of the grid methods seems to present a limit that is always improved by the continuous approach. 1 1 1 0.9 0.9 0.9 0.8 0.8 0.8 0.7 0.7 0.7 0.6 0.6 0.6 0.5 0.5 0.5 0.4 0.4 0.4 0.3 0.3 0.3 0.2 0.2 0.2 0.1 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0 0.2 0.4 0.6 0.8 1 1 1 1 0.9 0.9 0.9 0.8 0.8 0.8 0.7 0.7 0.7 0.6 0.6 0.6 0.5 0.5 0.5 0.4 0.4 0.4 0.3 0.3 0.3 0.2 0.2 0.2 0.1 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0 0.2 0.4 0.6 0.8 1 Figure 3. [Under the alternative] Empirical cumulative distribution under the alternative of the Rice test (blue) and the discrete grid tests with size 32 (green) , 102 (red), 322 (purple) and 502 (cyan) . The√alternative is defined by a single atom at a random location with a weight log(N ) = log(2fc + 1) (first row) or N (second row). In columns : fc = 3, 5, 7. 1.3. Toward a functional testing procedure In the seminal paper [21], the authors generalize the LARS test to regularized least squares problems with penalization P being a semi-norm over P coordinates but they do not cover the “continuum” framework under consideration here. On the technical side, their convex dual formulation leads to the study of a linear Gaussian process on which they invoke a Kac-Rice formula while our convex dual formulation involves a functional framework for which Kac-Rice formula cannot be applied. Furthermore, exact testing procedures in High-Dimensional statistics are based on PostSelection Inference techniques. In a nutshell, note that the sub-gradient of the `1 -norm is a vector with entries in [−1, 1] so that, in `1 -minimization, the selection event is described by the signs (sub-gradient entry value {±1}) and the active support (indices for which the sub-gradient is ±1) at the solution point through first order optimality conditions such as the Karun-Kush-Tucker conditions. Post-Selection Inference is then the ability to build a testing procedure that may depend on the aforementioned selection event. This approach 5 1 INTRODUCTION might seem meaningless in the frame of Super-Resolution since the selection events have measure zero—given Gaussian responses and a finite subset S ⊂ [0, 2π), the probability that the output support of `1 minimization is exactly S is zero. Indeed, specifying the location and the phase of the supremum of a Gaussian process is an event with null measure. Remark also that there is an uncountable number of selection events—as many as the number of points in the index of the process. So the tools of Post-Selection Inference cannot be applied here and one need a new theoretical analysis of testing procedures based on the outcomes of `1 -minimisation over the space of signed measures. For the first time, this paper paves the way to build new testing procedures in the framework of Super-Resolution theory (and line spectral estimation). In particular, we prove that we can rightfully construct global null exact testing procedures on the first two knots λ1 and λ2 of the “continuous” LARS when one has a continuum of predictors, see Theorems 3 and 6 and Figure 2. These two new procedures offer the ability to test the mean of any stationary Gaussian process with known correlation function Γ and C 2 -paths. Furthermore, one of these tests is unbiased, see Theorem 3 and they can be both Studentized, see Theorems 4 and 8, when variance σ 2 is unknown. 1.4. Outline The outline of the paper is described by the following (short) table of contents: 2 Notations and problem formulation . . . . . . . . . . . . . . . . . . . 2.1 Hypothesis testing problem . . . . . . . . . . . . . . . . . . . . 2.2 The first and second knots of a Gaussian process . . . . . . . . 2.3 Estimating the variance . . . . . . . . . . . . . . . . . . . . . . 3 Passing to the limit, the grid approach . . . . . . . . . . . . . . . . . 4 The Rice method: a grid-less approach . . . . . . . . . . . . . . . . . 4.1 The known variance case . . . . . . . . . . . . . . . . . . . . . . 4.2 The unknown variance case . . . . . . . . . . . . . . . . . . . . 5 Applications to the Super-Resolution Theory . . . . . . . . . . . . . 5.1 Framework and results . . . . . . . . . . . . . . . . . . . . . . . 5.2 A numerical study . . . . . . . . . . . . . . . . . . . . . . . . . 6 LARS for complex-valued Gaussian processes . . . . . . . . . . . . . 6.1 Cameron-Martin type Assumption on the mean and Notations . 6.2 Description of the LARS . . . . . . . . . . . . . . . . . . . . . . 6.3 The second knot . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4 Illustration: The two first knots of Super-Resolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 5 6 7 11 12 12 15 17 17 19 21 21 22 25 25 Notations and the formal problem formulation is described in Section 2. In Section 3, we present the test statistic S Grid which is constructed taking the limit of consecutive LARS knots (λ1,P , λ2,P ) on thinner and thinner grids (namely the number of predictors P tends to infinity). Section 4 is the theoretical construction of our grid-less test based on consecutive knots (λ1 , λ2 ) of the “continuous” LARS. The main result concerning the test statistic S Rice is presented in this section. Applications to spike detection in Super-Resolution are developed in Section 5. The general construction of the “continuous” LARS is given in Section 6. An appendix with the proofs can be found at the end of the paper. Azaïs, De Castro and Mourareau/2 NOTATIONS AND PROBLEM FORMULATION 6 2. Notations and problem formulation 2.1. Hypothesis testing problem In this paper, our purpose is to test the mean value of a stationary complex-valued Gaussian process Z with C 2 -paths indexed by [0, 2π). We assume that Z = A1 + ıA2 where A1 and A2 are two independent and identically distributed real-valued processes with C 2 -paths. Assume that the correlation function Γ of A1 (and A2 ) satisfies ∀t ∈ (0, 2π), |Γ(t)| < 1 (Anorm ) and let σ 2 := Var(A1 (·)) so that Cov(A1 (s), A1 (t)) = σ 2 Γ(t − s) . (1) We denote by T := [0, 2π)2 the 2-dimensional torus. Assume that we observe a real-valued process (X(z))z∈T indexed by T such that  ∀z ∈ T, X(z) := A1 (t) cos θ + A2 (t) sin θ = Re e−ıθ Z(t) , where z = (t, θ) and Re(·) denotes the real part of a complex number. Remark that observing X is equivalent to observe Z since we can recover Z from X and conversely. Furthermore, we may assume that the process (X(z))z∈T satisfies a.s. there is no point z ∈ T s.t. X 0 (z) = 0 and det(X 00 (z)) = 0, (Adegen ) where X 0 (z) and X 00 (z) denote the gradient and the Hessian of X at point z. Note that sufficient conditions for (Adegen ) are given by [5, Proposition 6.5] applied to (X(z))z∈T . In particular if the distribution of X 00 (t) is non degenerated, using [5, Condition (b) of Proposition 6.5], it implies that Assumption (Adegen ) is met. Note also that Assumption (Adegen ) is referred to as “Morse” process in [2]. Remark that (Anorm ) and (Adegen ) are mild assumptions ensuring that Z is a non-pathological process with C 2 -paths. This paper aims at testing the following hypotheses. H0 : “Z is centered ” against H1 : “Z is not centered ” . Subtracting the mean tested by the null hypothesis, remark that this framework encompasses any testing problem whose null hypothesis is a single hypothesis on the mean of Z. 2.2. The first and second knots of a Gaussian process As in high-dimensional statistics, we can define the first and second knots (λ1 , λ2 ) as follows. If we model some spatial correlation by means of the process X, the most correlated point zb ∈ T and the maximal correlation λ1 are respectively the argument maximum and the maximum of X defined by zb := arg max X(z) and λ1 := X(b z) . (2) z∈T Under Assumption (Anorm ), one can check that the argument maximum is almost surely a singleton, see Proposition 16. 7 2 NOTATIONS AND PROBLEM FORMULATION To construct the second knot, given a fixed z ∈ T, one can equivalently consider two regressions of X(y), as follows. • On the one hand, the regression on X(z) that will appear in the grid method of Section 3. Using a convenient normalisation related to the definition of the LARS knots, we set ∀y ∈ T \ {z}, X z (y) := where X(y) − X(z)ρ(z − y) X(y) − X(z) = X(z) + , 1 − ρ(z − y) 1 − ρ(z − y) ∀z ∈ T, ρ(z) := Γ(t) cos θ , is the correlation function of the stationary Gaussian process X. One can check that X z is a Gaussian process indexed by T \ {z} and independent of X(z). • On the other hand, the regression on (X(z), X 0 (z)) will be needed for convergence purposes in Section 4. With the convenient normalization, we set ∀y ∈ T \ {z}, X |z (y) := e −1 X 0 (z)i X(y) − ρ(z − y)X(z) + hρ0 (z − y), Λ . 1 − ρ(z − y) e := −ρ00 (0) is the variance-covariance where ρ0 is the gradient of the correlation function ρ and Λ matrix of the derivative process of X, namely X 0 . Since the derivative at zb is zero, note that X zb(·) = X |bz (·) and we define the second knot λ2 as yb := arg max X zb(y) and λ2 := X zb(b y ) = X |bz (b y) , (3) y∈T\{b z} where we prove that (b y , λ2 ) are well defined and that yb is almost surely unique, see Proposition 16 and Remark 7. Furthermore, the couple (b y , λ2 ) can be equivalently defined using the extension of the LARS to our framework, the interested reader may consult Section 6.3. 2.3. Estimating the variance When the variance σ 2 is unknown in (1), we precise here the assumptions and the estimator we use to estimate the variance. In this section, except for explicit examples, we consider a real valued Gaussian process Y not necessarily stationary defined on the 2-dimensional torus T. Let m ≥ 2 (possibly infinite) and assume that Y admits an order m Karhunen-Loève expansion in the sense that Y = m X i=1 ζi fi with Var(ζi ) = σ and ∀t ∈ T, 2 m X i=1 |fi (t)|2 = 1 , (KL(m)) where the equality holds in L2 (Ω) and (f1 , . . . , fm ) is a system of non-zero functions orthogonal on L2 (T). Through our analysis, we need to consider one of the following assumptions. • If m is finite, ∃(z1 , . . . , zm ) ∈ Tm pairwise distincts s.t. (Y (z1 ), . . . , Y (zm )) is non degenerated. (ND(m)) Azaïs, De Castro and Mourareau/2 NOTATIONS AND PROBLEM FORMULATION 8 • If m = ∞, ∀p ∈ N? , ∃(z1 , . . . , zp ) ∈ Tp pairwise distincts s.t. (Y (z1 ), . . . , Y (zp )) is non degenerated. ( ND(∞)) Recall that a Gaussian vector is called non-degenerated if its variance-covariance matrix is non-degenerated, i.e., it has full rank. Some examples of process Y satisfying (KL(m)) and (ND(m)) with m = ∞ are given by the normalized Brownian motion and any Gaussian stationary process with a spectrum that admits an accumulation point, see [10, Page 203]. For instance, the process corresponding to the Super-Resolution problem satisfies (KL(m)) and (ND(m)) with m finite, namely m is twice the number of observed frequencies, see Section 5. Definition 1. Let Y be a Gaussian process with constant variance σ 2 = Var(Y (·)) and satisfying Assumptions (KL(m)) and (ND(m)) with m finite. The quantity m 2 σ bKL (Y 1 X 2 ) := ζi , m i=1 is called the Karhunen-Loève estimator of σ 2 . 2 Remark 1. An explicit expression of the estimator σ bKL is always possible from some set of 0 pairwise disjoint points z1 , . . . , zm0 with m > m. We only need to check that the variancecovariance matrix of the (Y (z1 ), . . . , Y (zm0 )) has rank m. 2 Remark 2. Sufficiency considerations imply that σ bKL is an optimal unbiased estimator for the mean-squared error by Rao–Blackwell theorem. Given the aforementioned definition, we are now able to construct variance estimators for the process X. We assume that the complex Gaussian process Z that define X satisfies the following hypotheses for some N ∈ N. Z admits a complex Karhunen-Loève expansion of order N (KLZ (N )) and satisfies the following non-degeneracy conditions: ∀(t1 , . . . , tN ) ∈ [0, 2π)N pairwise distincts, (Z(t1 ), Z(t2 ), . . . , Z(tN )) is non degenerated and (NDZ (N )) 0 (Z(t1 ), Z (t1 ), Z(t3 ), . . . , Z(tN )) is non degenerated. Our aim is to build, for each z ∈ T, two estimators of the variance σ 2 independently from X(z) or (X(z), X 0 (z)). Indeed, in the following, we will distinguish two kind of statistics. The first one is the limit of the finite dimensional statistic S Grid , see Section 3. The second one is the case of the maximum over T, see Section 4. Both cases won’t use the same estimation of σ 2 . • In the grid situation, we define z Xnorm (y) := X(y) − ρ(z − y)X(z) p , 1 − ρ2 (z − y) 9 2 NOTATIONS AND PROBLEM FORMULATION where y belongs to T\{z}, ρ(·) denotes the correlation function of the process X and set 2 z σ bz2 := σ bKL Xnorm (·)  which is well defined, independent of X(z) and with constant variance σ 2 . Furthermore, it follows a σ 2 χ2 (2N −1)/(2N −1) distribution where χ2 (2N −1) denotes the chi-square distribution with 2N − 1 degrees of freedom, see Proposition 2. Finally, we consider the variance estimator (4) σ b2 = σ bzb2 , defined at point zb given by (2). • In the continuous case, we define e −1 X 0 (z)i X(y) − ρ(z − y)X(z) + hρ0 (z − y), Λ |z Xnorm (y) := q , e −1 ρ0 (z − y)i 1 − ρ2 (z − y) − hρ0 (z − y), Λ where y belongs to T \ {z} and set 2 2 |z σ b|z := σ bKL Xnorm (·)  which is well defined, independent of (X(z), X 0 (z)) and with constant variance σ 2 . Furthermore, it follows a σ 2 χ2 (2N − 3)/(2N − 3), see Proposition 2. Finally, we consider the variance estimator 2 σ b|2 = σ b|b (5) z, defined at point zb given by (2). Proposition 2. Let Z satisfy (KLZ (N )) and (NDZ (N )) and set z ∈ T then the following claims are true. (a) σ bz2 is well defined and follows a (b) 2 σ b|z is well defined and follows a σ 2 χ2 (2N −1) 2N −1 σ 2 χ2 (2N −3) 2N −3 distribution. distribution. |z z (·)/b σz is independent of σ bz , and the process Xnorm (·)/b σ|z is indepen(c) The process Xnorm dent of the random variable σ b|z . Proof. (a) Fix z = (t1 , θ1 ) ∈ T. Since Z satisfies (NDZ (N )), there exists (t2 , . . . , tN ) ∈ [0, 2π)N −1 pairwise different such that (Z(t1 ), Z(t2 ), . . . , Z(tN )) is non degenerated. Then, considering z1 = z, zN +1 = z + (0, π/2) and ∀i ∈ {2, . . . , N }, zi = (ti , θ1 ) and zN +i = (ti , θ1 + π/2), the vector V1 := (X(z1 ), . . . , X(z2N )) satisfies 2N = rank(X(z1 ), X(z2 ), . . . , X(z2N )) z z = rank(X(z1 ), Xnorm (z2 ), . . . , Xnorm (z2N )) z z = 1 + rank(Xnorm (z2 ), . . . , Xnorm (z2N )) z where rank denotes the rank of the covariance matrix of a random vector. Deduce that Xnorm (·) satisfies (ND(2N − 1)). This, in turn, implies that the 2N functions gi (·) = fi (·) − ρ(z − .)fi (z) Azaïs, De Castro and Mourareau/2 NOTATIONS AND PROBLEM FORMULATION 10 are in fact in a space of dimension 2N − 1 and a Gram-Schmidt orthogonalization in L2 (T) z z z gives (KL(2N − 1)) for the process Xnorm (·). Finally, from (Xnorm (z2 ), . . . , Xnorm (z2N )), we compute σ bz2 that follows the desired distribution. (b) In the case of the regression over (X(z), X 0 (z)), remark that ∂θ X(z) = X(t1 , θ1 + π/2) = X(z2 ) and ∂t X(z) = Re(e−ıθ1 Z 0 (t1 )) where ∂θ (resp. ∂t ) denote the partial derivative with respect to θ (resp. t). Because of hypothesis (NDZ (N )), the two vectors V1 and V2 := (X(z1 ), X(z2 ), Re(e−ıθ1 Z 0 (t1 )), Im(e−ıθ1 Z 0 (t1 )), X(z5 ), . . . , X(z2N )) have rank 2N so both are invertible functions of (Re(ζ1 ), Im(ζ1 ), . . . , Re(ζN ), Im(ζN )). In particular, Im(e−ıθ1 Z 0 (t1 )) is a linear combination of V2 . Let γ1 and γ2 be the coefficients associated to X(z3 ) and X(z4 ). By triangular combination, we deduce that the distribution of (X(z1 ), ∂θ X(z1 ), ∂t X(z1 ), γ1 X(z3 ) + γ2 X(z4 ), X(z5 ), . . . , X(z2N )) is non-degenerated and so that (γ1 , γ2 ) 6= (0, 0). Setting ψ such that γ1 γ2 cos(ψ) = p 2 and sin(ψ) = p 2 2 γ1 + γ2 γ1 + γ22 we get the non-degeneracy of (X(z1 ), ∂θ X(z1 ), ∂t X(z1 ), X(z2N +1 ), X(z5 ), . . . , X(z2N )) where z2N +1 = (t2 , θ1 + ψ). Finally, similarly to the proof of the previous point, regression, |z |z scaling and independence prove that the rank of (Xnorm (z5 ), . . . , Xnorm (z2N +1 )) is 2N − 3 so |z 2 is well defined and distributed that Xnorm (·) satisfies KL(2N −3) and ND(2N −3) and that σ b|z as σ 2 χ2 (2N −3) . 2N −3 (c) This is a direct consequence of the independence of the angle and the norm for each |z z marginal Gaussian vector build from Xnorm or Xnorm . Remark 3. When the complex process Z admits an infinite Karhunen-Loève decomposition, we need the following modified hypothesis ∀p ∈ N? , ∀(t1 , . . . , tp ) ∈ [0, 2π)p pairwise distincts, (Z(t1 ), Z(t2 ), . . . , Z(tp )) is non degenerated and (NDZ (∞)) 0 (Z(t1 ), Z (t1 ), Z(t3 ), . . . , Z(tp )) is non degenerated. Indeed, for every enter p ≥ 1, note that from the observation of the vector (Z(t1 ), Z(t2 ), . . . , Z(tp )) (resp. (Z(t1 ), Z 0 (t1 ), . . . , Z(tp ))) for pairwise disjoint points t1 , . . . , tp , we can construct an es2 (resp. σ 2 ), of the variance σ 2 with distribution σ 2 χ2 (2p − 1)/(2p − 1) (resp. timator, say σ b2p b|2p σ 2 χ2 (2p−3)/(2p−3)). Making p tend to infinity, classical concentration inequalities and Borel2 (resp. σ 2 ) converges almost surely to σ 2 . Thus the variance σ 2 Cantelli lemma prove that σ b2p b|2p 2 ) this observation, is directly observable from the entire path of X. We still denote σ bz2 (resp. σ b|z where z = z1 = (t1 , θ1 ). 11 3 PASSING TO THE LIMIT, THE GRID APPROACH 3. Passing to the limit, the grid approach The main idea of this section is to define a sequence of grids (Gn )n≥1 on T, to construct a sequence of test statistics (Sn )n≥1 from the values of the process X on Gn as in [4] and to pass to the limit as n → ∞. More precisely, we consider Gn to be the grid with mesh ∆n := (2π)2−n on T, zbn := arg max X(z) and λ1,n := max X(z) . z∈Gn z∈Gn It is the maximum of the process X when indexing by the grid. We can also define the maximum of the regression when indexing by the grid, namely λ2,n := max y∈Gn \{b zn } X zbn (y) . The Hessian at the maximum (2) on T is denoted by X 00 := X 00 (b z ). By Assumption (Adegen ), it is a random variable with values in the set of non degenerated negative definite matrices of size 2 × 2. We can define a non degenerated positive quadratic form (i.e., a metric) on R2 by kvkX 00 = −v > X 00 v, for v ∈ R2 . Using this metric, we can consider the corresponding Voronoi tessellation of Z2 . It is a regular partition of R2 by parallelograms, invariant by translations (1, 0) and (0, 1). Denote by Vo ⊂ [−1, 1]2 the Voronoi cell of the origin in this partition and by U := U(Vo ) the uniform distribution on this cell. We understand the law U as a conditional law with respect to X 00 and, conditionally to X 00 , this law is taken independent of (λ1 , λ2 ), see Lemma 12. Conditionally to X 00 , define the randomized statistics ( )  k − 2U  k> λ2 := λ2 ∨ λ1 + sup X 00 , (6) 1 e e 21 kk 2 2 k∈Z \{0} kΛ kk kΛ e 12 is the square root of Λ e = −ρ00 (0) and a ∨ b = max(a, b). A proof of the following where Λ result is given in Appendix A.1. Theorem 3. Under Assumptions (Anorm ) and (Adegen ), the randomized test statistics S Grid := Φ(λ1 /σ) ∼ U([0, 1]) , Φ(λ2 /σ) where Φ denotes the standard Gaussian survival function. Moreover, the test with p-value S Grid is unbiased: under the alternative H1 , it holds P{S Grid 6 α} > α for all α ∈ (0, 1). Theorem 3 shows in particular that the statistics—referred to as the Spacing test statistics in the introduction—given by Φ(λ1 /σ) (7) S ST = Φ(λ2 /σ) does not follows a U([0, 1]) distribution ans leads to a non-conservative test. Indeed, observe that almost surely λ2 ≤ λ2 so that S ST ≥ S Grid almost surely. Note that the two test statistics differ on the event {λ2 6= λ2 } = {λ2 < λ2 }. Now, when the variance σ 2 is unknown, we can build an estimator σ b2 defined in (4) and obtain a Studentized version of the previous theorem. Azaïs, De Castro and Mourareau/4 THE RICE METHOD: A GRID-LESS APPROACH 12 Theorem 4. Assume (Anorm ), (Adegen ), (KLZ (N )) and (NDZ (N )) where 2 6 N < ∞, then the following test statistics T Grid satisfies T Grid := F m−1 (λ1 /b σ)  ∼ U([0, 1]) F m−1 λ2 /b σ under the null hypothesis where m = 2N , Fm−1 is the Student cumulative distribution function b2 is defined by (4). with m − 1 degrees of freedom, F m−1 = 1 − Fm−1 its survival function and σ A proof can be found in Appendix A.2. Remark 4. Only the first point of (NDZ (N )) is required for the proof. Moreover, if m = +∞, the Student distribution is to be replaced by a standard normal distribution. 4. The Rice method: a grid-less approach In this section, we build our test statistic directly on the entire path of the process X in a grid-less manner. We assume that the process X satisfies Assumptions (Anorm ) and (Adegen ), and is centered. As in the preceding section, we consider λ1 and λ2 defined by (2) and (3) respectively. e so that the covariance function of X e is the correlation function ρ of X, We denote X = σ X e is the standardized version of X. Note that, by regression formulas and stationarity, namely X it holds   00 e e 0 (z)) = −Λ e X(z) e e (z) (X(z), X , ∀z ∈ T, E X e by the decomposition so that we can define the process R e 00 (z) = −Λ e X(z) e e X + R(z) e e e = −ρ00 (0) is the variance-covariance where R(z) and X(z) are independent for any z ∈ T and Λ 0 e (t). In particular, observe that matrix of X e z ) + R(b X 00 (b z ) = −ΛX(b z) , e z ). Using the Rice method of [5, Theorem 7.2] (see also [18]), it follows that where R(b z ) = σ R(b the maximum λ1 has for density w.r.t the Lebesgue measure on R+ at point ` > 0   e (cst)(−1)d E det(−ΛX(0) + R(0))1A` X(0) = `, X 0 (0) = 0 σ −1 φ(σ −1 `), where φ denotes the standard Gaussian density, A` is the event {X(y) 6 `, ∀y ∈ T} and (cst), as in the following, denotes a positive constant. The numerical values (cst) may vary from an occurence to another and it may depend on m and σ which are assumed fixed in our framework. 4.1. The known variance case We begin by the known variance case. The main observation is that the method of [5, Theorem 7.2] can be extended to compute the joint distribution of (λ1 , λ2 , R(b z )) as follows. • Denote S the set of symmetric matrices and pick a Borel set B on D := R2 × S. 13 4 THE RICE METHOD: A GRID-LESS APPROACH • For every z ∈ T, recall that ∀y ∈ T \ {z}, X |z (y) := and define e −1 X 0 (z)i X(y) − ρ(z − y)X(z) + hρ0 (z − y), Λ 1 − ρ(z − y) ∀z ∈ T, λz2 := sup X |z (y) . (8) y∈T\{z} Remark that, for fixed z ∈ T, λz2 is a.s. finite by Lemma 9, X |z (·) is independent of (X(z), X 0 (z)) and, by way of consequence, λz2 is independent of (X(z), X 0 (z)). Furthermore, note that since T is without boundary, for z = zb, one has X 0 (z) = 0 and λz2 = λ2 as defined by (3). • Observe that on the event {∀y 6= z, X(y) < X(z)} one has almost surely that z = zb, X(z) = λ1 , λz2 = λ2 and R(z) = R(b z ). Also, a simple computation shows that ∀z ∈ T s.t. X 0 (z) = 0, 1{z=bz } = 1{∀y6=z, X(y)<X(z)} = 1{0<λz2 <X(z)} , almost surely. Hence, by unicity of zb and recalling that the set {z ; X 0 (z) = 0} is finite under (Adegen ), we deduce that X 1{(X(z),λz2 ,R(z))∈B}∩{0<λz2 <X(z)} = 1{(λ1 ,λ2 ,R(bz ))∈B} . z:X 0 (z)=0 (n) • On D define smooth lower approximations ϕB converge when n goes to infinity i.e. ∀(`1 , `2 , r) ∈ Ω, of the indicator function of B that (n) ϕB (`1 , `2 , r) −→ 1{(`1 ,`2 ,r)∈B}∩{0<`2 <`1 } . • Apply Rice formula with weights [5, Theorem 6.4] (see also the proof of [5, Theorem 7.2]) to compute h X i (n) z E ϕB X(z), λ2 , R(z) z:X 0 (z)=0 Z = (cst) ZT = (cst) T h i (n) e E | det(−ΛX(z) + R(z))| ϕB (X(z), λz2 , R(z)) X 0 (z) = 0 dz h i (n) e E | det(−ΛX(z) + R(z))| ϕB (X(z), λz2 , R(z)) dz where the last equality relies on the fact that (X(z), λz2 , R(z)) is independent of X 0 (z). • Combining the previous observations and passing to the monotone limit as n tends to ∞ in the aforementioned Rice formula with weights, we get that   P λ1 , λ2 , R(b z) ∈ B h X i =E 1{(X(z),λz2 ,R(z))∈B}∩{0<λz2 <X(z)} z:X 0 (z)=0 Z h i e E | det(−ΛX(z) + R(z))|1{(X(z),λz2 ,R(z))∈B}∩{0<λz2 <X(z)} dz T h i e = (cst) E | det(−ΛX(0) + R(0))|1{(X(0),λ02 ,R(0))∈B}∩{0<λ02 <X(0)} , h i e = (cst) E det(−ΛX(0) + R(0))1{0<λ02 <X(0)} 1{(X(0),λ02 ,R(0))∈B} , = (cst) (9) Azaïs, De Castro and Mourareau/4 THE RICE METHOD: A GRID-LESS APPROACH 14 by stationarity and using that, on the event {0 < λ02 < X(0)}, the matrix −X 00 (0) = e ΛX(0) − R(0) belongs to the set of positive definite symmetric matrices, namely S+ . Before stating the key result on the joint density of (λ1 , λ2 , R(b z )) we need to introduce a dominating measure. First, recall that X(0) is independent of the pair (λ02 , R(0)). Then, e0 , R(0)) e0 is defined as in (8) for the process X. e e observe that (λ02 , R(0)) = σ × (λ where λ 2 2 0 e e Denote µ1 the law of (λ2 , R(0)) and note that it does not depend on σ. Denote µσ the law of (λ02 , R(0)) and remark that for any Borel set B of R × S, it holds µσ (σB) = µ1 (B). Eventually, remark that The law of (X(0), λ02 , R(0)) is dominated by Leb(R) ⊗ µσ . (10) where Leb(R) denotes the Lebesgue measure on R. As a consequence we can prove the following proposition. Proposition 5. The joint law L((λ1 , λ2 , R(b z ))) of (λ1 , λ2 , R(b z )) satisfies for all (`1 , `2 , r) ∈ 2 R × S, dL((λ1 , λ2 , R(b z ))) e 1 + r)1{0<` <` } σ −1 φ(σ −1 `1 ) , (`1 , `2 , r) = (cst) det(−Λ` 2 1 dLeb(R) ⊗ µσ where Leb(R) ⊗ µσ is defined by (10) and S denotes the set of symmetric matrices. Proof. Observe that the density at point `1 of X(0) with respect to the Lebesgue measure is σ −1 φ(σ −1 `1 ) and recall (10). Now, for any Borel set B of R2 × S, note that h i e E det(−ΛX(0) + R(0))1{(0<λ02 <X(0)} 1{(X(0),λ02 ,R(0))∈B} Z e 1 + r)1{0<` <` } σ −1 φ(σ −1 `1 )d`1 µσ (d(`2 , r)) = det(−Λ` 2 1 B thanks to (9), which prove the result. We can now state our result when the variance is known. Theorem 6. Set + ∀r ∈ S , ∀` > 0, Z +∞ Gr (`) := e + r)φ(uσ −1 )du , det(−Λu ` e denotes the Hessian of the correlation function ρ of X at the origin. Under Assumpwhere Λ tions (Anorm ) and (Adegen ), the test statistic S Rice := GR(bz ) (λ1 ) GR(bz ) (λ2 ) ∼ U([0, 1]) under the null H0 . Proof. Using Proposition 5, we know that the density of λ1 at `1 and conditional to (λ2 , R(b z )) = (`2 , r) is equal to e 1 + r)φ(σ −1 `1 )1{0<` <` } . (cst) det(−Λ` 2 1 4 THE RICE METHOD: A GRID-LESS APPROACH 15 It is well known that, if a random variable Z has for cumulative density function F then F(Z) follows an uniform distribution on [0, 1]. This implies that, conditionally to (λ2 , R(b z )) = (`2 , r), Gr (λ1 ) ∼ U([0, 1]). Gr (`2 ) Since the conditional distribution does not depend on (`2 , r), it is also the non conditional distribution and it yields GR(bz ) (λ1 ) ∼ U([0, 1]) , GR(bz ) (λ2 ) as claimed. 4.2. The unknown variance case e where σ > 0 is unknown. Assume that Z satisHence, suppose that we observe X = σ X fies (KLZ (N )) and (NDZ (N )), and set m = 2N . The regression of the Hessian on (X(z), X 0 (z)) reads now e e ∀z ∈ T, X 00 (z) = −ΛX(z) + σ R(z). because X 0 (z) is independent of (X(z), X 00 (z)) by stationarity. The variance being unknown, we estimate it using σ b|2 which is defined by (5). For fixed z ∈ T, by Claims (b) and (c) of Proposition 2, we know that the following random variables or random processes |z X(z) , X 0 (z) , Xnorm (·) and σ b|z σ b|z |z are mutually independent. As Xnorm (·) = hz (·) X |z (·) where hz (·) is a deterministic function and as Lemma 15 shows that R(z) can be expressed as radial limits of X |z (·) at point z, we get that  X |z (·) R(z)  and σ b|z are mutually independent, , X(z) , X 0 (z) , σ b|z σ b|z and by consequence X(z) , X 0 (z) ,  λz R(z)  2 , and σ b|z are mutually independent. σ b|z σ b|z We turn now to the Rice formula described previously and introduce the notation T2,z := λz2 and T2 := T2,bz . σ b|z Denote Leb(R2 ) the Lebesgue measure on R2 and let µ1 be the joint law of the couple of random variables (T2,0 , R(0)/b σ|0 ). Under H0 , note that X(0) is a centered Gaussian variable 2 with variance σ and σ b|0 /σ is distributed as a chi-distribution with m − 3 degrees of freedom, i.e., the law of density m−3 21− 2 m−4 s fχm−3 (s) = exp(−s2 /2) Γ m−3 2 Azaïs, De Castro and Mourareau/4 THE RICE METHOD: A GRID-LESS APPROACH 16 b|0 /σ, T2,0 , R(0)/b σ|0 ) has a denwhere Γ is the Gamma function. Then the quadruplet (X(0), σ sity with respect to Leb(R2 ) ⊗ µ1 at point (`1 , s, t2 , r) ∈ R3 × S equal to  2  s (m − 3) m−4 (cst) s exp − σ −1 φ(σ −1 `1 ). 2 Using the same method as for the proof of Proposition 5 we have the following proposition. Proposition 7. Assume that Z satisfies (Anorm ), (Adegen ), (KLZ (N )) and (NDZ (N )), and set m = 2N . Then, the joint distribution of λ1 , σ b| /σ, T2 , R(b z )/b σ| has a density with respect to Leb(R2 ) ⊗ µ1 at point (`1 , s, t2 , r) ∈ R3 × S+ equal to  2  s (m − 3) m−4 e (cst) det(−Λ`1 + σsr) s exp − φ(σ −1 `1 )1{0<σst2 <`1 } , 2 where (cst) is a positive constant that may depend on m and σ. Consequently, we derive the following result. Theorem 8. Assume that Z satisfies (Anorm ), (Adegen ), (KLZ (N )) and (NDZ (N )), and set m = 2N . For all r ∈ S+ , define H r (·) as ! r Z +∞  m − 1 e 1 + r fm−1 t1 ∀` > 0, H r (`) := det − Λt dt1 , m−3 ` where fm−1 is the density of the Student distribution with m − 1 degrees of freedom. Under the null H0 , the test statistic H R(bz ) (T1 ) T Rice := ∼ U([0, 1]), H R(bz ) (T2 ) where T1 := λ1 /b σ| , T2 = λ2 /b σ| and σ b| is defined by (5). `1 Proof. First, using Proposition 7 and the change of variable t1 = σs , the joint distribution of the quadruplet (T1 , σ b| /σ, T2 , R(b z )/b σ| )) at point (t1 , s, t2 , r) is given by   2 s (m − 3) m−3 e (cst) det(σs(−Λt1 + r))s exp − φ(st1 )1{0<t2 <t1 } 2   !2 r m − 3 m − 1 e 1 + r)sm−1 exp − s  φ(st1 )1{0<t <t } . = (cst) det(−Λt 2 1 m−1 2 Second, note that if X and Y are two independent random variables of density fX and fY then the density of X/Y satisfies Z fX/Y (z) = fX (zy)yfY (y)dy. R In our case, integrating over s and with the change of variable s ← s p (m − 1)/(m − 3), it 5 APPLICATIONS TO THE SUPER-RESOLUTION THEORY 17 holds  !2 m − 3 m − 1  ds φ(st1 )sm−1 exp − s m−1 2 + R ! r  2  Z m−1 s (m − 1) m−2 ss exp − ds = (cst) φ st1 m−3 2 R+ ! r Z m−1 χm−1 (s)ds φ st1 = (cst) sf √ m−3 m−1 R+ ! r m−1 . = fm−1 t1 m−3  Z r Putting together, the density of (T1 , T2 , R(b z )/b σ ) at point (t1 , t2 , r) is now given by ! r m − 1 e 1 + r)fm−1 t1 (cst) det(−Λt 1{0<t2 <t1 } , m−3 and we conclude using the same trick as the one of Theorem 6. 5. Applications to the Super-Resolution Theory 5.1. Framework and results Deconvolution over the space of complex-valued Radon measure has recently attracted a lot of attention in the “Super-Resolution” community—and its companion formulation in “Line spectral estimation”. A standard aim is to recover fine scale details of an image from few low frequency measurements—ideally the observation is given by a low-pass filter. The novelty in this body of work relies on new theoretical guarantees of the `1 -minimization over the space of Radon measures with finite support. Some recent works on this topic can be found in the papers [11, 7, 20, 8, 3, 13, 6, 12] and references therein. An important example throughout this paper is given by the Super-Resolution problem which can be stated as follows. Let ν 0 ∈ (M([0, 2π), C), k·k1 ) a complex-valued Radon measure on the one dimensional torus identified to [0, 2π) equipped with the natural circle-wise metric. Note that ||·||1 denotes the total variation norm on M([0, 2π)). The space (M([0, 2π), C), k·k1 ) can be defined as the topological dual space of continuous functions on [0, 2π) equipped with the L∞ -norm. Let N = 2fc + 1 where fc ≥ 1 is referred to as the “frequency cut-off”. Denote by DN the Dirichlet kernel defined by ∀t ∈ [0, 2π), DN (t) := sin(N t/2) . sin(t/2) Consider the linear operator FN : M([0, 2π), C) → CN that maps any complex-valued Radon measure ν onto its Fourier coefficients ck (ν) where Z ck (ν) := exp(−ıkx)ν(dx) T Azaïs, De Castro and Mourareau/5 APPLICATIONS TO THE SUPER-RESOLUTION THEORY 18 for integers k such that |k| 6 fc . Consider ζ = (ζk )k where ζk = ζ1,k + ıζ2,k and ζ`,k are i.i.d. standard Gaussian random variables for |k| 6 fc and ` = 1, 2. In the Super-Resolution frame, we observe a perturbed version of the Fourier coefficients, namely 1 y = √ FN (ν 0 ) + σζ . N ? —the dual operator of F , remark that we observe the trigonometric polynomial Applying FN N 1 ? (y) Z := √ FN N which reads as ∀t ∈ [0, 2π), 1 Z(t) = N Z T DN (t − x)ν 0 (dx) + σ fc X k=−fc 1 √ ζk exp(ıkt). N (11) Hence, one observes Z and infers on ν 0 assuming that it has finite support. To this purpose, consider the process X defined for all (t, θ) ∈ T by X(t, θ) := Re(e−ıθ Z(t)) = cos(θ) Re(Z(t)) + sin(θ) Im(Z(t)), (12) where Re and Im denote the real and imaginary part of a complex number. When ν 0 ≡ 0, remark that the processes A1 = Re(Z) and A2 = Im(Z) are two independent and identically distributed real-valued processes with C ∞ -paths. An elementary computation shows that X has correlation function ρ and A1 has correlation function Γ with ρ(z − y) = cos(θ − α)Γ(t − s) where Γ(t − s) = DN (t − s)/N for all z = (t, θ) and y = (s, α) in T. Remark that (Anorm ) holds true for Γ. In this case, we are testing ? H0 : “FN (FN (ν 0 )) ≡ 0 ” ? against H1 : “∃t ∈ [0, 2π), FN (FN (ν 0 ))(t) 6= 0 ” , or equivalently H0 : “ν 0 ≡ 0 ” against H1 : “∃t ∈ [0, 2π), ν 0 (t) 6= 0 ” . Subtracting the known measure ν 0 , remark that this framework encompasses testing problem whose null hypothesis is any single hypothesis H0 : “ν 0 ≡ ν0 ” against alternatives of the form H1 : “∃t ∈ [0, 2π), ν 0 (t) 6= ν0 (t) ”. Furthermore, we have the following propositions. First, we check that we can apply our results to the Super-Resolution process. Proposition 9. The process X defined by (12) satisfies Condition (KL(m)) and Condition (ND(m)) with m = 2N = 4fc + 2. 19 5 APPLICATIONS TO THE SUPER-RESOLUTION THEORY Then, we derive a first result when the noise level σ si known. Proposition 10. Under the null H0 , the test statistic Rice SSR = σ(α1 λ1 + α2 )φ(λ1 /σ) + (α1 σ 2 − α32 )Φ(λ1 /σ) ∼ U([0, 1]), σ(α1 λ2 + α2 )φ(λ2 /σ) + (α1 σ 2 − α32 )Φ(λ2 /σ) where Φ is the standard Gaussian cumulative distribution function , Φ = 1 − Φ its survival function, φ its density function, (λ1 , λ2 ) is defined by ((2), (3)) and  α1 = 31 fc (fc + 1),    fc   b  α = √1 P (k 2 − α ) × Re(y eı(kbt−θ) ), 1 2 k N k=−fc   fc P  b  1  k × Re(yk eı(kbt−θ) ).  α3 = √N k=−fc Finally, we have the following result when the noise level σ is unknown. Proposition 11. Under the null H0 , the test statistic Rice TSR = −1 α2 F α1 F m−3 (T1 ) + (α1 T1 + α2 )fm−3 (T1 ) − γm 3 m−1 (T1 ) ∼ U([0, 1]), −1 2 α1 F m−3 (T2 ) + (α1 T2 + α2 )fm−3 (T2 ) − γm α3 F m−1 (T2 ) where Fd is the Student cumulative distribution function with d degrees of freedom, F d = 1−Fd its survival function, fd its density function, T1 = λ1 /b σ| , T2 = λ2 /b σ| , σ b| is defined by (5) and m−3 m Γ Γ ) ) ( ( 2 2 γm = m−3 m−2 Γ( m−1 )Γ( m−2 ) . 2 2 A proof of these propositions can be found in Appendix A.4. 5.2. A numerical study 5.2.1. Computation of λ2 Rice ), we need To build our test statistic S Rice in the Super-Resolution context (namely SSR to compute three quantities. The first one is λ1 , the maximum of X(.) over the torus T. Its simple form allow us to use classical optimization routines, for instance fminsearch on MATLAB or optim on R both combined with global resolution options on T. The second one is R = R(b z ) which appears in the test statistic through the coefficients α1 , α2 and α3 that are simple functions of the observation y and zb. Finally, the third one is   X(y) − X(b z) zb . λ2 = λ2 = λ1 + max y∈T 1 − ρ(b z − y) Contrary to λ1 , there is some indetermination problem when y is close to zb. In particular, the approximation of zb is by definition not exact and the radial limits of X | are not numerically achieved. A way to get around that is the integral form of the remainder in Taylor’s theorem. In full generality, we compute Z 1  Q P 2   T 00 r Iq   (1 − h) (y − z b ) X (b z + h(y − z b ))(y − z b ) dh   q=0 zb 0 λ2 − λ1 = max Z 1 ' , Q  y∈T  P   T 00 2   (1 − h) (y − zb) ρ (b z + h(y − zb))(y − zb) dh r Jq 0 q=0 Azaïs, De Castro and Mourareau/5 APPLICATIONS TO THE SUPER-RESOLUTION THEORY 20 where Iq and Jq are given by Simpson’s rule [1, Page 886] applied to the numerator (resp. denominator), Q is a tuning parameter that control the approximation error and r = ||y − zb||22 . However, in the Super-Resolution case, elementary trigonometry identities give the following simpler form of the denominator r 2 fc X k=−fc 2 (k cos(α) − sin(α)) × sinc  r(k cos(α) − sin(α)) 2 2 where y − zb = (r cos(α), r sin(α)) and sinc denote the cardinal sine function, i.e.  sin x   x sinc(x) =   1 if x 6= 0, if x = 0, which is a numerically robust function. We conclude the optimization using the same routine as the one of λ1 . 5.2.2. Monte-Carlo experiment In this section we compare the cumulative distribution of several statistics of test in the case where the variance is known, namely • The statistics of the Rice test S Rice , given by Theorem 6, are displayed in blue. • The statistics of the Spacing test S ST , given by (7), are displayed in green. • The statistics of the Spacing test on grids Gn given by Φ(λ1,n )/Φ(λ2,n ) are displayed with a color that take the respective values green, red, purple and cyan for sizes equal to 32 , 102 , 322 , 502 . • The grid test, based on S Grid of Theorem 3 can be viewed as the limit of the discrete grid tests above as the size growths to infinity. As one can see in the figures, there is some evidence that this limit is numerically reached for a size n = 502 . We complete each graph by the diagonal to the cumulative distribution function of the uniform law on [0, 1] displayed in black. All the figures are based on 2000 simulations of the corresponding statistics. The first figure studies the distribution of S Rice and S ST under the Null. This figure is displayed in the introduction (see Figure 2). The second figure deals with the grid statistic and S Rice under various alternatives defined by a single spike and compares the power of the Rice test with the discrete grid tests, see Figure 3. Finally, the third figure performs the same study but with an alternative defined by two atoms, see Figure 4. Discussion Figure 2 suggests that the Spacing test is highly non-conservative which is a major drawback. For instance, when fc = 7, the empirical level of the Spacing test at a nominal level of 5% is in fact 11,3%, showing that this test is very non-conservative. For its part, the Rice test is exact as predicted by the theory. This numerical agreement prove that the numerical algorithm described in Section 5.2.1 is efficient. 21 6 LARS FOR COMPLEX-VALUED GAUSSIAN PROCESSES 1 1 1 0.9 0.9 0.9 0.8 0.8 0.8 0.7 0.7 0.7 0.6 0.6 0.6 0.5 0.5 0.5 0.4 0.4 0.4 0.3 0.3 0.3 0.2 0.2 0.2 0.1 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Figure 4. Same as Figure 2 except that (a) fc = 7 , (b) the alternative is defined by two spikes at random √ locations √ √ (with a constraint of separation) (c) the weights are now from left to right (log N, log N ); (log N, N ); ( N , N ). In Figure 3 and 4 we see that the power of the discrete grid tests may seem an increasing function of the number of points of the grid. This power seems to converge since the curves associated to 322 (purple) and 502 (cyan) are almost indistinguishable. This suggests that the Rice test (blue) is always more powerful than the discrete grid test or the limit grid test. Consequently, it seems unbiaised for any choice of alternative. In conclusion the Rice test seems to be the best choice even if we are still not able to prove theoretically that it is unbiased. 6. LARS for complex-valued Gaussian processes 6.1. Cameron-Martin type Assumption on the mean and Notations The algorithm presented here can be used for a large class of complex processes Z. In this section, we introduce new notation and a new set of assumptions. We consider a complexvalued Gaussian process Z indexed on a compact metric space K with covariance function K. Remark 5. Note that this model encompasses our framework setting K = [0, 2π) and K = 2σ 2 Γ with Γ the correlation of A1 defined in Section 2.1. We do not assume that the process is stationary in this section though our method can be applied to stationary processes as for instance in the Super Resolution frame, see Section 6.4. We assume that its covariance K is such that there exists σ > 0 such that ∀s 6= t ∈ K, K(t, t) = 2σ 2 and K(s, t) < 2σ 2 . (13) The scalar 2 accounts for the contribution of the real and the imaginary part of Z and σ 2 is the variance of the real part of Z. We assume that Z has continuous sample paths. We present here the underlying hypothesis on the mean of the Gaussian processes under consideration when using the LARS algorithm. This hypothesis is of Cameron-Martin type. Indeed, the main drawback that should be avoided is when the mean cannot be represented in the RKHS of the Gaussian process. We recall that we can define a reproducing Hilbert space of the covariance K, see [15, Chapter 2.6] for instance. Denote (H, h·, ·iH ) this complex Hilbert space. Also, we can invoke a Karhunen-Loève expansion of the process Z. Namely, there exist Azaïs, De Castro and Mourareau/6 LARS FOR COMPLEX-VALUED GAUSSIAN PROCESSES 22 i.i.d. complex standard normal variables (gj )j≥1 , a real orthonormal system (ej )j≥1 on L2 (K) and σj > 0 such that X X Z − EZ = σj gj ej and σj2 = 2σ 2 < ∞, j j where the identity holds almost surely in the Banach space of continuous functions on K equipped with the L∞ -norm. By Mercer’s theorem, we know that X ∀s, t ∈ K, K(s, t) = σj2 ej (s)ej (t) , j where the identity holds almost surely in the Banach space of continuous functions on K × K equipped with the L∞ -norm. We recall also that the Hilbert space H can be defined as H := nX j aj ej | X |aj |2 j σj2 <∞ o with the inner product X X X aj bj . h aj ej , bj ej iH = σj2 j j j We observe Z and we want to estimate its mean E Z. Remark that almost surely it holds Z − E Z ∈ H, where H is the closure of H in the space of continuous functions equipped with the infinity norm, see e.g. [15, Corollary 2.6.11]. Remark that H is also closed in L2 (K). Denoting by E the L2 orthogonal space of H, one has L2 (K) = H ⊕ E where the sum is orthogonal. We denote by P (resp. P ⊥ ) the orthogonal projection onto H (resp. E). Since almost surely Z − E Z ∈ H, remark that almost surely P ⊥ (Z) = P ⊥ (EZ) and this process can be observed and is deterministic. Without loss of generality, we assume that P ⊥ (EZ) = 0 subtracting P ⊥ (Z) to Z. Also, we assume that P(EZ) ∈ H . (14) Recall that P(EZ) = EZ and Assumption (14) gives that Z ∈ H using Z − E Z ∈ H. 6.2. Description of the “continuous” LARS We assume that Z ∈ H and, as mentioned above, this assumption is equivalent to Assumption (14). Following standard references, e.g., [16, Chapter 5.6], the Least-Angle Regression Selection (LARS) algorithm can be extended to Gaussian processes. To the best of our knowledge, the LARS for complex Gaussian processes has never been introduced and we present its formulation here for the first time. Actually, the presentation given in this section can be applied to any RKHS setting. It results in a description of the LARS in infinite dimensional feature spaces and this framework has been dealt in [19]. However, note that the paper [19] only concerns real signed measures and their “doubling” dimension trick [19, page 546] cannot be used when dealing with complex measures. In particular, their result cannot be invoked in Super-Resolution where it is of utmost importance to deal with complex measures. This section presents the “continuous” LARS for Super-Resolution. 6 LARS FOR COMPLEX-VALUED GAUSSIAN PROCESSES 23 The LARS is a variable selection algorithm giving a sequence ((λk , µk ))k≥1 where the knots are ordered such that λ1 ≥ λ2 ≥ . . . > 0 and µk ∈ (M(K, C), k · k1 ) is a complex-valued measure. We recall that the space (M(K, C), k · k1 ) is defined as the dual space of the space of continuous functions on K equipped with the L∞ -norm. Inspired by the Super-Resolution framework—presented in Section 5, we consider Z as some “correlation process” in the spirit of (11). In particular, the most correlated point can be defined by (2), namely λ1 := max |Z(t)| . t∈K Under Assumption (13), Proposition 16 shows that almost surely there exists a unique point t1 such that λ1 = |Z(t1 )|. Define the “active set” A as A1 = (t1 ). This active set is the support of the second fitted solution µ2 . Set the first fitted solution to µ1 = 0 and the first residual to Z1 = Z for initialization purposes. Observe that (15) Z1 (t1 ) = λ1 eıθ1 , |Z1 (t1 )| = λ1 , ∀t ∈ K, ∀t 6= t1 , Z Z1 (t) = Z(t) − K(u, t)µ1 (du) , K |Z1 (t)| < λ1 . From this point we proceed iteratively. For k ≥ 2, we assume that we have found (λk−1 , µk−1 ) and Ak−1 = (t1 , . . . , tk−1 ) such that Z K(u, t)µk−1 (du) , ∀t ∈ K, Zk−1 (t) := Z(t) − K ∀t ∈ Ak−1 , ∀t ∈ / Ak−1 , |Zk−1 (t)| = λk−1 , |Zk−1 (t)| < λk−1 . We want to add an other point tk to the active set and define a discrete measure µk supported on Ak−1 while keeping the above inequalities true. From this point proceed as follows. ◦ First, we solve the least-squares fit given by a = arg min Zk−1 (·) − c∈Ck−1 k−1 X i=1 ci K(ti , ·) 2 H This program can be solved in closed form since its amounts in fitting k − 1 weights in leastsquares regression. It holds that −1 a = Mk−1 (Zk−1 (t1 ), . . . , Zk−1 (tk−1 )) where we denote Mk−1 = (K(ti , tj ))1≤i,j≤k−1 and we assume that Mk−1 is invertible. If Mk−1 is not invertible then we stop. The complex vector a = (a1 , . . . , ak−1 ) is referred to as the “least-squares direction”. P Remark 6. Note that the function k−1 i=1 ai K(ti , ·) is the regression of Zk−1 onto the finite dimensional space Span{K(ti , ·) ; i = 1, . . . , k − 1}. Azaïs, De Castro and Mourareau/6 ◦ LARS FOR COMPLEX-VALUED GAUSSIAN PROCESSES 24 Then, for any 0 < λ ≤ λk−1 , define µ(λ) := µk−1 + (1 − λ/λk−1 ) k−1 X ai δti i=1 k−1 X Z (λ) (·) := Zk−1 (·) + (λ/λk−1 − 1) i=1 , ai K(ti , ·) , and observe that |Z (λ) (t)| = λ for all t ∈ {t1 , . . . , tk−1 }. Indeed, it holds Z (λ) (tj ) = Zk−1 (tj ) + (λ/λk−1 − 1) k−1 X ai K(ti , tj ) , i=1 > = Zk−1 (tj ) + (λ/λk−1 − 1)a Mk−1 (0, . . . , 0, |{z} 1 , 0, . . . , 0) , jth = λZk−1 (tj )/λk−1 , and recall that |Zk−1 (tj )| = λk−1 . Now, we keep track of the largest value of the “correlation” process |Z (λ) | on the complementary set of Ak−1 while moving λ from λk−1 toward zero. We define λk as the largest value for which there exists a point t ∈ / Ak−1 such that |Z (λ) |(t) = λ. Set  λk := max β > 0 ; ∃ t ∈ / Ak−1 , s.t. |Z (β) |(t) = β , (16) and tk := arg max |Z (λk ) |(s) . s∈A / k−1 If tk is not unique, we add all the solutions of (16) to the active set Ak . For sake of readability, we assume that tk is the only solution to (16). ◦ Update Ak = (t1 , . . . , tk ) , µk = µ(λk ) = µk−1 + (1 − λk /λk−1 ) k−1 X ai δti , i=1 Zk (·) = Z (λk ) (·) = Zk−1 (·) + (λk /λk−1 − 1) where, for all t ∈ K, Zk (t) = Zk−1 (t) + (λk /λk−1 − 1) = Z(t) − = Z(t) − Z K k−1 X k−1 X i=1 ai K(ti , ·) , ai K(ti , t) i=1 K(u, t)µk−1 (du) + (λk /λk−1 − 1) k−1 X ai K(ti , t) , i=1 Z K(u, t)µk (du) , K is the kth residual associated to the kth fitted solution µk . Remark also that ∀t ∈ {t1 , . . . , tk }, ∀t 6= {t1 , . . . , tk }, and update k to k + 1 to iterate the procedure. |Zk (t)| = λk , |Zk (t)| < λk , 25 6 LARS FOR COMPLEX-VALUED GAUSSIAN PROCESSES 6.3. The second knot First, observe that λ1 is defined as in (2) and that the two definitions agree. Indeed, recall that X(t, θ) = Re (e−ıθ Z(t)) so that max X = max |Z| at point zb = (t1 , θ1 ) with t1 as in (15). By optimality, it holds that λ1 = e−ıθ1 Z(t1 ). Then, the case k = 2 is interesting since λ2 is a statistic used in the test statistics described previously. We will see that the two definitions agree here again. For k = 2, it holds Z1 = Z and the least squares direction is given by a = Z(t1 )/(2σ 2 ) and Z (λ) by λ K(t1 , t) − 1)Z(t1 ) , λ1 2σ 2 K(t1 , t) = Z(t) + eıθ1 (λ − λ1 ) 2σ 2 Z (λ) (t) = Z(t) + ( Multiplying by e−ıθ and taking the real part, this latter can be equivalently written as Re (e−ıθ Z (λ) (t)) = X(z) + (λ − λ1 ) cos(θ1 − θ) K(t1 , t) , 2σ 2 where z = (t, θ) ∈ T. Now, recall that ρ(t, θ) := Γ(t) cos θ = cos(θ)K(0, t)/(2σ 2 ) to compute Re (e−ıθ Z (λ) (t)) = X(z) + (λ − λ1 )ρ(z − zb) . We deduce that Re (e−ıθ Z (λ) (t)) ≤ λ ⇔ X(z) − λ1 ρ(z − zb) ≤ λ(1 − ρ(z − zb)) X(z) − X(b z )ρ(z − zb) ⇔ ≤λ 1 − ρ(z − zb) ⇔ X zb(z) ≤ λ showing that the second knot λ2 is exactly the quantity defined in (3). 6.4. Illustration: The two first knots of Super-Resolution The Super-Resolution process is defined in (11). It satisfies Condition (KLZ (N )) and Condition (NDZ (N )) with N = 2fc +1. The first point is given by the maximum of the modulus of Z, see the red curve in Figure 5. Observe that Z1 = Z and the maximum satisfies Z1 (t1 ) = λ1 eıθ1 . Then, we compute λ DN (t1 − t) Z (λ) (t) = Z1 (t) + ( − 1)Z1 (t1 ) , λ1 2N σ 2 where DN denotes the Dirichlet kernel. For λ > λ2 , the maximum of |Z (λ) | is achieved at a unique point, namely t1 . For λ = λ2 , a second point achieves the maximum. This transition defines Z2 := Z (λ2 ) , see Figure 5. From this point, we can iterate fitting the least squares direction on the support {t1 , t2 } and decreasing |Z2 | while a third point achieves the maximum. Given the red curve in Figure 5, it was not obvious that the second knot would have been t2 since other local maxima seemed more significant than t2 on the red curve. Azaïs, De Castro and Mourareau/A PROOFS 26 Figure 5. LARS for Super-Resolution: we fit a Dirichlet kernel at the maximal correlation point t1 until the maximal correlation in the residual is matched at a second point t2 = 6 t1 . Appendix A: Proofs We denote for random variables, Xn = oP (rn ) and Yn = OP (rn ) (for rn 6= 0) means that rn−1 ||Xn || converges to 0 in probability and rn−1 ||Yn || is uniformly tight, respectively. Furthermore, we consider the following processes. • The stationary process X(z) = X(t, θ) defined on T with covariance function given by Cov(X(y), X(z)) = σ 2 ρ(z − y) where we recall the correlation function is given by ρ(z − y) = cos(θ − α)Γ(t − s), • For every z ∈ T, recall the regressions with respect to X(z) X(y) − X(z)ρ(z − y) X(y) − X(z) = X(z) + , 1 − ρ(z − y) 1 − ρ(z − y) X(y) − ρ(z − y)X(z) z p Xnorm (y) = . 1 − ρ2 (z − y) ∀y ∈ T \ {z}, X z (y) = • For every z ∈ T, recall the regressions with respect to (X(z), X 0 (z)) e −1 X 0 (z)i X(y) − ρ(z − y)X(z) + hρ0 (z − y), Λ , 1 − ρ(z − y) e −1 X 0 (z)i X(y) − ρ(z − y)X(z) + hρ0 (z − y), Λ |z Xnorm (y) = q . e −1 ρ0 (z − y)i 1 − ρ2 (z − y) − hρ0 (z − y), Λ ∀y ∈ T \ {z}, X |z (y) = In particular, recall that zb is defined by (2) so X 0 (b z ) = 0 and it yields that X zb = X |bz . A.1. Proof of Theorem 3 Since the variance is known, we consider without loss of generality that σ 2 = 1. Using the metric given by the quadratic form represented by X 00 , we can consider the closest point z n of the grid Gn to zb by z n = arg min kb z − ukX 00 = arg min hb z − u, −X 00 (b z )(b z − u)i . u∈Gn u∈Gn 27 A PROOFS The main claim is that, while it holds λ1,n → λ1 a.s., we don’t have the same result for λ2,n , see Lemma 14. We begin with the following preliminary result, which is related to the result of Azaïs-Chassan (see [9]). Lemma 12. Conditionally to X 00 , ∆−1 z − z n ) follows a uniform distribution on V0 and this n (b distribution is independent from λ1 and λ2 . Proof. Remark that zb has uniform distribution on T by stationarity and this distribution is independent from λ1 and λ2 . Let B be a Borelian in R2 . Remark that zb − z n ∈ ∆n V0 by definition of z n and note that z n ∈ ∆n Z2 . Conditionally to X 00 , it holds P{∆−1 z − z n ) ∈ B} = P{∆−1 z − z n ) ∈ B ∩ V0 } n (b n (b = P{(b z − z n ) ∈ ∆n (B ∩ V0 )} = P{b z ∈ ∆n (B ∩ V0 + Z2 )} . Since zb has uniform distribution on T and since V0 + Z2 is a partition of R2 , it holds that P{b z ∈ ∆n (B ∩ V0 + Z2 )} = P{b z ∈ 2π(B ∩ V0 + Z2 )} = Leb(R2 )(2π(B ∩ V0 )) , Leb(R2 )(2πV0 ) where Leb(R2 ) denotes the Lebesgue measure on R2 . Lemma 13. It holds that (a) X(b zn ) − X(z n ) = oP (∆2n ). (b) P{b zn = 6 z n } → 0 as n goes to ∞. (c) Let F be any measurable function, then F (b zn ) − F (z n ) tends to zero in probability at arbitrary speed. (d) Almost surely, one has z n → zb and zbn → zb as n goes to infinity. Proof. Let ε > 0. By definition of z n and since V0 ⊂ [−1, 1]2 , it holds that √ ||b z − z n || 6 2∆n , (17) almost surely. Since X has C 2 -paths and by Taylor expansion, one has X(b z ) − X(z n ) = (1/2)||b z − z n ||2X 00 + oP (∆2n ) (18) Since −X 00 is positive definite, there exists M > 0 sufficiently large such that (1/M )Id2 4 −X 00 4 M Id2 where 4 denotes the Lowner ordering between symmetric matrices. Then, it holds ∀z ∈ R2 , (1/M )||z||2 ≤ ||z||2X 00 6 M ||z||2 . (19) From (17), (18) and (19), we deduce that 0 6 X(b z ) − X(b zn ) 6 X(b z ) − X(z n ) = OP (∆2n ) , using the optimality of zb and zbn . (20) Azaïs, De Castro and Mourareau/A PROOFS 28 By compactness of T, uniqueness of optimum zb ∈ R2 and C 2 -continuity of X, there exists η > 0 and a neighborhood N0 ⊂ R2 of zb ∈ R2 such that X(b z ) − η > X(z) for any z ∈ / N0 and ∀z ∈ N0 , (1/4)||b z − z||2X 00 6 X(b z ) − X(z) 6 ||b z − z||2X 00 (21) using again a Taylor expansion as in (18). Using (20), it holds that, on an event of probability at least 1 − ε/4 and for n large enough, 0 6 X(b z ) − X(b zn ) ≤ η/2 implying that zbn ∈ N0 . Invoke (19), (20) and (21) to deduce that zb − zbn = OP (∆n ). Using Taylor formula again, we get that X(b z ) − X(b zn ) = (1/2)||b z − zbn ||2X 00 + oP (∆2n ) . (22) By optimality of zbn and z n and using (18) and (22), one gets 0 6 X(b zn ) − X(z n ) ≤ (1/2)(||b z − z n ||2X 00 − ||b z − zbn ||2X 00 ) + oP (∆2n ) . (23) z − zbn ||2X 00 ≤ 0, we get (a). Observing that ||b z − z n ||2X 00 − ||b 00 Conditionally to X and in the metric defined by || · ||X 00 , there exists η 0 > 0, such that the η 0 -neighborhood, denoted by Nη0 , of the boundary ∂V0 of V0 has relative volume (for the Lebesgue measure) less than ε/8. More precisely, Nη0 denotes the set of points in V0 ⊂ R2 with || · ||X 00 -distance less than η 0 to the boundary of V0 . In particular, ∀k ∈ Z2 \ {0}, ∀z ∈ V0 \ Nη0 , ||z||X 00 + η 0 6 ||z − k||X 00 , by Cauchy-Schwarz inequality. Using Lemma 12 and by homogeneity, we deduce that it holds ||b z − z n ||X 00 + η 0 ∆n 6 ||b z − g||X 00 , ∀g ∈ ∆n Z2 \ {0}, with probability at least 1 − ε/8. It follows that z − g||2X 00 , ||b z − z n ||2X 00 + (η 0 )2 ∆2n 6 ||b ∀g ∈ ∆n Z2 \ {0}, using that (a + b)2 > a2 + b2 for a, b > 0. Now, invoke (23) to get that 06− (η 0 )2 2 ∆n 1{z n 6=zbn } + oP (∆2n ) . 2 On these events, we get that, for n sufficiently large, zbn and z n must be equal except on an event of probability at most ε/4 + ε/8 ≤ ε. Furthermore, this result holds unconditionally in X 00 . We deduce that lim sup P{z n 6= zbn } 6 ε, proving (b). Note that (c) is a consequence of the fact that, for n sufficiently large, zbn and z n must be equal except on an event of arbitrarily small size. In particular, it shows that supk>n ||z k − zbk || converges towards zero in probability, which is equivalent to almost sure converge of z n − zbn towards zero. Claim (d) follows when remarking that (17) proves a.s. convergence of z n towards zb. Lemma 14. As n tends to infinity, λ2,n converges in distribution to λ2 . Proof. Let β ∈ R be such that 0 < β < 1/2, say β = 1/4. Let ε ∈ (0, 1). We can write λ2,n = λA,n ∨ λB,n with λA,n := max u∈Gn \{b zn } s.t. λB,n := max u∈Gn s.t. ku−b zn k6∆β n ku−b zn k>∆β n X zbn (u) =: max X zbn (u), u∈Gn,A X zbn (u) =: max X zbn (u). u∈Gn,B 29 A PROOFS We first prove that λB,n → λ2 as n tends to infinity in distribution. By compactness, remark that there exists a constant Cr > 0 such that 1 − ρ(u) > Cr kuk2 . (24) It also holds that X(u) − X(b zn ) , 1 − ρ(u − zbn ) X(u) − X(b z) X zb(u) = X(b z) + . 1 − ρ(u − zb) X zbn (u) = X(b zn ) + (25) (26) Let us look to the rhs of (25) and (26). By Claim (d) of Lemma 13 and the continuous mapping theorem, note that X(b zn ) converges toward λ1 = X(b z ) a.s. and we can omit these terms. It remains to prove that on Gn,B the second terms are equivalent. Because of Lemma 13, 1 − ρ(u − zbn ) converges to 1 − ρ(u − z n ) at arbitrary speed. Remember that (17) gives zb − z n = OP (∆n ) and it holds that ||u − zb|| > ∆βn , on Gn,B . It follows that there exists C > 0 such that 1 − ρ(u − zbn ) > C∆2β n and 1 − ρ(u − zb) > C∆2β n , (27) with probability greater than 1 − ε/2. As for the numerators, Eqs. (20) and (21) show that for all u ∈ Gn,B X(u) − X(b zn ) X(b z ) − X(b zn ) |X(b z ) − X(b zn )| −1 = = (cst) = OP (∆2−2β ) n 2 X(u) − X(b z) X(u) − X(b z) ||u − zb|| In this sense, we say that X(u)−X(b zn ) is uniformly equivalent to X(u)−X(b z ) on the grid Gn,B in probability. Using (27) and noticing that for any u ∈ T |ρ(u − zb) − ρ(u − zbn )| 6 ||ρ0 ||∞ ||b z − zbn || = OP (∆n ) , the same result holds for the denominators, namely 1 − ρ(u − zbn ) is uniformly equivalent to 1 − ρ(u − zb) on the grid Gn,B in probability. We deduce that X zbn (u) is uniformly equivalent to X zbn (u) on the grid Gn,B in probability and, passing tho their maximum, one can deduce that λB,n converges to λ2 in probability. We turn now to the study of the local part λA,n . Again, by Claim (c) of Lemma 13 we can replace zbn by z n in the numerator of the r.h.s in (25) and we forget the first term which limit is clearly λ1 almost surely. We perform a Taylor expansion at zb, it gives that X(u) − X(b z ) = (1/2)(u − zb)> X 00 (u − zb)(1 + oP (1)). for any u ∈ Gn,A . Since z n − zb = OP (∆n ), we also get that X(z n ) − X(b z ) = (1/2)(z n − zb)> X 00 (z n − zb) + oP (∆2n ). As for the denominator, invoke (17), (24) and Claim (c) of Lemma 13 to get that 1 − ρ(u − z n ) > 2Cr ∆2n , 1 − ρ(u − zbn ) = 1 − ρ(u − z n ) + oP (∆2n ),  e − z n ) (1 + oP (1)), and 1 − ρ(u − zbn ) = (1/2) (u − z n )> Λ(u Azaïs, De Castro and Mourareau/A PROOFS 30 e denotes the Hessian at point 0 of ρ. Putting all together yields where −Λ (u − z n )> X 00 (u + z n − 2b z) X(u) − X(b zn ) = (1 + oP (1)), > e 1 − ρ(u − zbn ) (u − z n ) Λ(u − z n ) for any u ∈ Gn,A . Now we know that, in distribution, zb − z n = ∆n U and we know that u − z n = k∆n with k belonging to a certain growing subset of Z2 which limit is Z2 . Finally, conditionally to X 00 , we obtain that max u∈Gn,A \{b zn } X zbn (u) −→ λ1 + k> (k − 2U) X 00 , 1 e 2 kk e 21 kk k∈Z2 \{0} kΛ kΛ sup in distribution. Eventually, consider the test statistic Sn := Φ(λ1,n )/Φ(λ   2,n ) and keep in mind that X u + (0, π) = −X(u) and that if u belongs to Gn , u + (0, π) also belongs. So Theorem 1 of [4] applies showing that, under the alternative, P{Sn 6 α} > α. It suffices to pass to the limit to get the desired result. A.2. Proof of Theorem 4 We use the same grid argument as for the proof of Theorem 3. Let t1 , t2 , . . . , tN be pairwise distinct points of [0, 2π), θ1 ∈ [0, 2π), m = 2N and set z1 = (t1 , θ1 ), . . . , zN = (tN , θ1 ), zN +1 = (t1 , θ1 + π/2), . . . , zm = (tN , θ1 + π/2). Because of the first assumption of (NDZ (N )), the distribution of (X(z1 ), . . . , X(zm )) is non degenerated. Consequently, following the proof of Proposition 2, we know that X z1 satisfies KL(m − 1) and ND(m − 1). Denote g1 , . . . , gm−1 the eigenfunctions of the KarhunenLoève (KL) representation of X (0,0) . Note that X z1 (·) has the same distribution as X (0,0) (.−z1 ) (stationarity) and that both are defined on the same space so the KL-eigenfunctions of X z1 are g1 (. − z1 ), . . . , gm−1 (. − z1 ). Now consider Az1 = (Azi,j1 )1≤i,j≤m−1 the matrix with entries Azi,j1 = gi (zj+1 − z1 ) which is invertible thanks to KL(m − 1) and ND(m − 1) and build so that     X z1 (z2 ) ζ1   .. .  z    = A 1  ..  . . X z1 (zm ) ζm−1 2 (X zbn (G )), the estimator of σ 2 One possible explicit expression, among many others, of σ bKL n norm on the grid Gn , is  X (0,0) (z2 − zbn )   ..   . (0,0) X (zm − zbn )  2 zbn σ bn2 := σ bKL (Xnorm (Gn )) = 1 m−1  Azbn −1 2 , 2 which is a composition of continuous functions of zbn . In particular, as zbn converges a.s. to zb (see Lemma 13, Claim (b)), we deduce that σ bn2 converges a.s. to σ bzb2 as n goes to infinity. 31 A PROOFS Finally, since the KL estimator is unique, this estimator coincide with the estimator σ b22 of [4] and Theorem 3 of [4] implies that  F m−1 λ1,n /b σn  ∼ U([0, 1]). F m−1 λ2,n /b σn Note that λn1 converges almost surely to λ1 and λ2,n converges in distribution to λ2 (see Lemma 14) to complete the proof. A.3. Proof of Proposition 9 (a). We can assume that Z defined by (11) is centered and, in this case, it holds fc σ X ζk exp(ıkt), ∀t ∈ [0, 2π) Z(t) = √ N k=−fc (28) where we recall that N = 2f c + 1 and ζk = ζk,1 + ıζk,2 for k = −fc , . . . , fc are independent standard complex Gaussian variables. Formula (28) shows that Z satisfies (KLZ (N )). (b). Let (t1 , . . . , tN ) ∈ [0, 2π) be pairwise differents, θ ∈ [0, 2π) and set      Z(t1 ) exp(−ıfc t1 ) . . . exp(ıfc t1 ) ζ1     ..   .. .. ..  =   .  =: At1 ,...,tN ζ, . . . Z(tN ) exp(−ıfc tN ) . . . exp(ıfc tN ) ζN where At1 ,...,tN is a Vandermonde matrix, invertible as soon as ti 6= tj for all i 6= j. This prove the first point of NDZ (N ). For the second assertion, consider h > 0 such that h < min1≤i<j≤N −1 (ti − tj ) and the Gaussian vector (Z(t1 ), . . . , Z(tN −1 ), Z(t1 + h))T =: At1 ,...,tN −1 ,t1 +h ζ, where the covariance matrix At1 ,...,tN −1 ,t1 +h satisfies Y det(A∗t1 ,...,t1 +h At1 ,...,t1 +h ) = 1≤i<j≤N −1 N (N −1)/2 =4 | exp(ıti ) − exp(ıtj )|2 Y 2 sin 1≤i<j≤N −1  N −1 Y j=1 ti − tj 2 | exp(ı(t1 + h)) − exp(ıtj )|2  NY −1 j=1 2 sin  tj − (t1 + h) 2  2 = sin (h/2) × gt1 ,...,tN −1 (h) where gt1 ,...,tN −1 (0) 6= 0 if (ti )1≤i≤N −1 are pairwise distincts. Finally, denote by Rh the linear transformation involving the first and the last coordinate such that     Z(t1 ) Z(t1 ) ..     ..     . . = R    h  Z(tN −1 )   Z(tN −1 )  Z(t1 +h)−Z(t1 ) Z(t1 + h) h and remark that sin2 (h/2) 1 = gt1 ,...,tN −1 (0) × 6= 0 2 h→0 h 4 lim det(A∗t1 ,...,t1 +h Rh∗ Rh At1 ,...,t1 +h ) = gt1 ,...,tN −1 (0) × lim h→0 giving the desired non degeneracy condition. Azaïs, De Castro and Mourareau/B AUXILIARY RESULTS 32 A.4. Proof of Proposition 10 and Proposition 11 Easy computations give the following results for φ(·), Z Z +∞ Z +∞ φ(t)dt = Φ(`), tφ(t)dt = φ(`), ` +∞ t2 φ(t)dt = `φ(`) + Φ(`), ` ` for fm−1 (·), Z +∞ fm−1 ` +∞ Z tfm−1 ` Z ! r m−3 m−1 t dt = F m−1 ` , m−1 m−3 !   r √ Γ m−3 (m − 3) m − 3 Γ m m−1 2 2   fm−3 (`), √ t dt = m−3 (m − 2) m − 1 Γ m−1 Γ m−2 2 2 !   r √  Γ m−3 m−1 (m − 3) m − 3 Γ m 2 2   × `fm−3 (`) + F m−3 (`) , √ t dt = m−1 m−2 m−3 (m − 2) m − 1 Γ 2 Γ 2 r +∞ t2 fm−1 ` m−1 m−3 ! r and for R, e z ) + R(b X 00 (b z ) = −ΛX(b z ),     −α2 α3 α1 0 X(b z) + , =− 0 1 α3 0 where  α1 = 31 fc (fc + 1),    fc   b  α = √1 P (k 2 − α ) × Re(y eı(kbt−θ) ), 1 2 k N k=−fc   fc P  b  1  k × Re(yk eı(kbt−θ) ).  α3 = √N k=−fc To conclude, use Proposition 9 to apply Theorem 6 and Theorem 8. Appendix B: Auxiliary results B.1. Regularity of X |z and new expression of R(z) Lemma 15. X |z (y) admits radials limits as y → z. More precisely for all λ in the unit sphere lim X |z (z + uλ) = u→0 Proof. As u tends to zero 1 − ρ(uλ) = λ> R(z)λ . e λ> Λλ u2 > e (λ Λλ + o(1)). 2 Moreover, a Taylor expansion gives X(z + uλ) = X(z) + uXλ0 (z) + u2 00 X (z) + op (u2 ), 2 λ 33 REFERENCES and e + op (u2 ), ρ0λ (uλ) = uρ00λ (0) + op (u2 ) = −uΛ where (Xλ0 , ρ0λ ) and (Xλ00 , ρ00λ ) are directional derivative and directional Hessian. By consequence, |z X (z + uλ) = = which tends to u2 >e 2 X(z)λ Λλ u2 2 e −1 X 0 (z)i + uX 0 (z) + + hρ0 (uλ), Λ λ u2 > e 2 (λ Λλ + o(1))   e + X 00 (z) + op (1) X(z)λ> Λλ λ u2 >e 2 (λ Λλ u2 00 2 Xλ (z) + op (u2 ) + o(1))   e λ> ΛX(z) + X 00 (z) λ e λ> Λλ e as u tends to 0 since Xλ00 (z) = λ> X 00 (z)λ. The result follows from X 00 (z) = −ΛX(z)+R(z). B.2. Maximum of a continuous process The following result is borrowed from [17, Theorem 3] and [23]. Proposition 16. Let {Y (t) ; t ∈ T } be a Gaussian process with continuous sample paths defined on a compact metric space T . Suppose in addition that: There is no two different points s, t ∈ T such that X(s) = X(t) a.s. (29) Then almost surely the maximum of X on T is attained at a single point. Observe that (Anorm ) implies (29). Remark 7. Proposition 16 can be applied to the process X |bz which is not continuous on a compact set. We use the “pumping method” as follows. Use (a) a parameterization of T as [0, 2π)2 , (b) polar coordinates for y ∈ T \ {b z } with origin at zb, (c) the change of parameter y = (ρ, θ) 7→ ((ρ + 1), θ) that transforms the non-compact set T \ {b z } into a compact set (we have inflated the “hole” {b z } into a ball centered around zb with radius one) on which the process X |bz is continuous thanks to Lemma 15. References [1] M. Abramowitz and I. Stegun. Handbook of Mathematical Functions. Dover, New York, fifth edition, 1964. [2] R. J. Adler and J. E. Taylor. Random fields and geometry. Springer Science & Business Media, 2009. [3] J.-M. Azaïs, Y. De Castro, and F. Gamboa. Spike detection from inaccurate samplings. Applied and Computational Harmonic Analysis, 38(2):177–195, 2015. REFERENCES/REFERENCES 34 [4] J.-M. Azaïs, Y. De Castro, and S. Mourareau. Power of the Spacing test for Least-Angle Regression. Bernoulli, 2016. [5] J.-M. Azaïs and M. Wschebor. Level sets and extrema of random processes and fields. John Wiley & Sons Inc., 2009. [6] T. Bendory, S. Dekel, and A. Feuer. Robust recovery of stream of pulses using convex optimization. Journal of Mathematical Analysis and Applications, 442(2):511–536, 2016. [7] K. Bredies and H. K. Pikkarainen. Inverse problems in spaces of measures. ESAIM: Control, Optimisation and Calculus of Variations, 19(01):190–218, 2013. [8] E. J. Candès and C. Fernandez-Granda. Towards a Mathematical Theory of Superresolution. Communications on Pure and Applied Mathematics, 67(6):906–956, 2014. [9] M. Chassan and J.-M. Azais. Discretization error for the maximum of a Gaussian field. Technical report, HAL, 2017. working paper or preprint. [10] H. Cramer and M. Leadbetter. Stationary and Related Stochastic Processes (1967). [11] Y. De Castro and F. Gamboa. Exact reconstruction using beurling minimal extrapolation. Journal of Mathematical Analysis and applications, 395(1):336–354, 2012. [12] V. Duval and G. Peyré. Exact support recovery for sparse spikes deconvolution. Foundations of Computational Mathematics, pages 1–41, 2015. [13] C. Fernandez-Granda. Support detection in super-resolution. In 10th international conference on Sampling Theory and Applications (SampTA 2013), pages 145–148, Bremen, Germany, July 2013. [14] C. Fernandez-Granda. Super-resolution of point sources via convex programming. Information and Inference, page iaw005, 2016. [15] E. Giné and R. Nickl. Mathematical foundations of infinite-dimensional statistical models, volume 40. Cambridge University Press, 2015. [16] T. Hastie, R. Tibshirani, and M. Wainwright. Statistical learning with sparsity: the lasso and generalizations. CRC Press, 2015. [17] M. A. Lifshits. On the absolute continuity of distributions of functionals of random processes. Theory of Probability & Its Applications, 27(3):600–607, 1983. [18] R. Lockhart, J. Taylor, R. J. Tibshirani, and R. Tibshirani. A significance test for the lasso. Annals of statistics, 42(2):413, 2014. [19] S. Rosset, G. Swirszcz, N. Srebro, and J. Zhu. `1 regularization in infinite dimensional feature spaces. Lecture Notes in Computer Science, 4539:544, 2007. [20] G. Tang, B. N. Bhaskar, P. Shah, and B. Recht. Compressed sensing off the grid. Information Theory, IEEE Transactions on, 59(11):7465–7490, 2013. [21] J. E. Taylor, J. R. Loftus, and R. J. Tibshirani. Inference in adaptive regression via the Kac–Rice formula. Ann. Statist., 44(2):743–770, 04 2016. [22] R. J. Tibshirani, J. Taylor, R. Lockhart, and R. Tibshirani. Exact post-selection inference for sequential regression procedures. Journal of the American Statistical Association, 111(514):600–620, 2016. [23] V. Tsirel’Son. The density of the distribution of the maximum of a Gaussian process. Theory of Probability & Its Applications, 20(4):847–856, 1976.
7cs.IT
New avenue to the Parton Distribution Functions: Self-Organizing Maps arXiv:0810.2598v2 [hep-ph] 2 Nov 2008 H. Honkanena,b,∗ , S. Liutia,† a Department of Physics, University of Virginia, P.O. Box 400714, Charlottesville, VA 22904-4714, USA b Department of Physics and Astronomy, Iowa State University, Ames, IA 50011, USA J. Carnahanc , Y. Loitierec , P. R. Reynoldsc c Department of Computer Science, School of Engineering, University of Virginia, P.O. Box 400740 Charlottesville, VA 22904-4740, USA Abstract Neural network algorithms have been recently applied to construct Parton Distribution Function (PDF) parametrizations which provide an alternative to standard global fitting procedures. We propose a technique based on an interactive neural network algorithm using Self-Organizing Maps (SOMs). SOMs are a class of clustering algorithms based on competitive learning among spatially-ordered neurons. Our SOMs are trained on selections of stochastically generated PDF samples. The selection criterion for every optimization iteration is based on the features of the clustered PDFs. Our main goal is to provide a fitting procedure that, at variance with the standard neural network approaches, allows for an increased control of the systematic bias by enabling user interaction in the various stages of the process. PACS numbers: 13.60.Hb, 12.38.Bx, 84.35.+i ∗ † [email protected] [email protected] 1 Introduction Modelling experimental data always introduces bias, in the form of either a theoretical or systematical bias. The former is introduced by researchers with the precise structure of the model they use, which invariably constrains the form of the solutions. The latter form of bias is introduced by algorithms, such as optimization algorithms, which may favour some results in ways which are not justified by their objective functions, but rather depend on the internal operation of the algorithm. In this paper we concentrate on high energy hadronic interactions, which are believed to be described by Quantum Chromodynamics (QCD). Because of the properties of factorization and asymptotic freedom of the theory, the cross sections for a wide number of hadronic reactions can be computed using perturbation theory, as convolutions of perturbatively calculable hard scattering coefficients, with non perturbative Parton Distribution Functions (PDFs) that parametrize the large distance hadronic structure. The extraction of the PDFs from experiment is inherently affected by a bias, which ultimately dictates the accuracy with which the theoretical predictions can be compared to the high precision measurements of experimental observables. In particular, the form of bias introduced by PDFs will necessarily impact the upcoming searches of physics beyond the Standard Model at the Large Hadron Collider (LHC). This situation has in fact motivated an impressive body of work, and continuous, ongoing efforts to both estimate and control PDFs uncertainties. Currently, the established method to obtain the PDFs is the global analysis, a fitting procedure, where initial scale Q0 ∼ 1GeV ≤ Qmin dat ansatze, as a function of the momentum fraction x, for each parton flavour i in hadron h are evolved to higher scales according to the perturbative QCD renormalization group equations. All the available observables e.g. the proton structure function, F2p (x, Q2 ), are composed of the candidate PDFs and comparison with the data is made with the help of some statistical estimator such as the global χ2 , 2 χ = Ne XX (Datai − Theori ) Vij−1 (Dataj − Theorj ) , expt. i,j=1 1 (1) where the error matrix Vij consists of the statistical and uncorrelated systematic errors, as well as of the correlated systematic errors when available. The parameters in the ansatze are then adjusted and the whole process repeated until a global minimum has been found. The modern PDF collaborations (CTEQ [1] and references within , MRST [2– 4], Alekhin [5, 6], Zeus [7] and H1 [8]) also provide error estimates for the PDF sets. They all rely on some kind of variant of the Hessian method (see e.g. [9] for details), which is based on a Taylor expansion of the global χ2 around it’s minimum. When only the leading terms are kept, the displacement of χ2 can be written in terms of Hessian matrix Hij , which consists of second derivatives of χ2 with respect to the parameter displacements, evaluated at the minimum. The error estimate for the parameters themselves, or for any quantity that depends on those parameters, can then be obtained in terms of the inverse of the Hessian matrix, (∆X)2 = ∆χ2 X ∂X ∂yi i,j H −1  ∂X . ij ∂y j (2) For details of PDF uncertainty studies see e.g. Refs. [10, 11]. The global analysis combined with Hessian error estimation is a powerful method, allowing for both extrapolation outside the kinematical range of the data and extension to multivariable cases, such as nuclear PDFs (nPDFs) [12–15]. In principle, when more data become available, the method could also be applied to Generalized Parton Distributions (GPDs), for which only model-dependent [16] or semi model-dependent [17, 18] solutions presently exist. However, there are uncertainties related to the method itself, that are difficult to quantify, but may turn out to have a large effect. Choosing global χ2 as a statistical estimator may not be adequate since the minimum of the global fit may not correspond to a minima of the individual data sets, and as a result the definition of ∆χ2 may be ambiguous. Estimates for the current major global analyses are that ∆χ2 = 50 − 100 is needed to obtain a ∼ 90% confidence interval [1, 2]. In principle this problem could be avoided by using the Lagrange multiplier method (see e.g.[19]), which does not assume quadratic behaviour for the errors around the minimum, instead of the Hessian method, 2 but this is computationally more expensive solution. Introducing a functional form at the initial scale necessarily introduces a parametrization dependence bias and theoretical assumptions behind the fits, such as s, s̄, c quark content, details of the scale evolution (e.g. higher order perturbative corrections, large/small x resummation), higher twists etc. as well as the data selection and treatment, e.g. kinematical cuts, all reflect into the final result of the analysis. Also, there may be systematical bias introduced by the optimization algorithm. The differences between the current global PDF sets tend to be larger than the estimated uncertainties [20], and these differences again translate to the predictions for the LHC observables, such as Higgs [21] or W ± and Z production cross sections [1]. A new, fresh approach to the PDF fitting has recently been proposed by NNPDF collaboration [22, 23] who have replaced a typical functional form ansatz with a more complex standard neural network (NN) solution and the Hessian method with Monte Carlo sampling of the data (see the references within [23] for the nonsinglet PDF fit and the details of the Monte Carlo sampling). Figure 1: Schematic diagram of a feed-forward neural network, from [24]. Neural network can be described as a computing solution that consists of interconnected processing elements, neurons, that work together to produce an output function. In a typical feed-forward NN (see Fig. 1) the output is given by the neurons in the last layer, as a non-linear function of the output of all neurons in the previous layer, which in turn is a function of the output of all neurons in the previous layer, and so on, starting from the first layer, which receives the input. For a NN with L layers and nl neurons in each PL−1 layer, the total number of the parameters is l=1 (nl nl+1 + nl+1 ). 3 In the beginning of the NNPDF fitting procedure a Monte Carlo sample of replicas of the experimental data is generated by jittering the central values of the data withing their errorbars using univariate Gaussian (or some other distribution if desired) random numbers for each independent error source. The number of the replicas is made so large that the Monte Carlo set of replicas models faithfully the probability distribution of the original data. For each replica a Genetic Algorithm (GA) fit is performed by first setting the NN parameters for each parton flavour to be fitted randomly, then making clones of the set of parameters, and mutating each of them randomly (multiple mutations). After scale evolution the comparison with the data is performed for all the clones, and the best clones are selected for a source of new clones, and the process repeated until the minimum for the χ2 has been found. Overfitting of the data is prevented by using only part of the data in the minimizing procedure, and using the other part to monitor the behaviour of the χ2 . When fitting PDFs one thus ends up with Nrep PDF sets, each initial scale parton distribution parametrized by a different NN. The quality of the global fit is then given by the χ2 computed from the averages over the sample of trained neural networks. The mean value of the parton distribution at the starting scale for a given value of x is found by averaging over the replicas, and the uncertainty on this value is the variance of the values given by the replicas. The NNPDF method circumvents the problem of choosing a suitable ∆χ2 , and it relies on GA which works on a population of solutions for each MC replica, thus having a lowered possibility of getting trapped in local minima. NN parametrizations are also highly complex, with large number of parameters, and thus unbiased compared to the ansatze used in global fits. The estimated uncertainties for NNPDF fits are larger than those of global fits, possibly indicating that the global fit uncertainties may have been underestimated. It should, however, be pointed out that the MC sampling of the data is not not tied to use of NNs, and it thus remains undetermined whether the large uncertainties would persist if the MC sampling was used with a fixed functional form. The complexity of NN results may also induce problems, especially when used in a purely automated fitting procedure. Since the effect of modifying individual NN parameters is unknown, the result may exhibit 4 strange or unwanted behaviour in the extrapolation region, or in between the data points if the data is sparse. In such a case, and in a case of incompatible data, the overfitting method is also unsafe to use. Implementation of information not given directly by the data, such as nonperturbative models, lattice calculations or knowledge from prior work in general, is also difficult in this approach. A possible method of estimating the PDF uncertainties could also be provided by Bayesian statistical analysis, as preliminarily studied in [25, 26] and explained in [27], in which the errors for the PDF parameters, or for an observable constructed from the PDFs, are first encapsulated in prior probabilities for an enlarged set of model parameters, and posterior distributions are obtained using computational tools such as Markov Chain Monte Carlo. Similar to NNPDF approach, this method allows for an inclusion of nonGaussian systematic errors for the data. In this introductory paper we propose a new method which relies on the use of Self-Organizing Maps (SOMs), a subtype of neural network. The idea of our method is to create means for introducing “Researcher Insight” instead of “Theoretical bias”. In other words, we want to give up fully automated fitting procedure and eventually develop an interactive fitting program which would allow us to “take the best of both worlds”, to combine the best features of both the standard functional form approach and the neural network approach. In this first step, we solely concentrate on single variable functions, free proton PDFs, but discuss the extension of the model to multivariable cases. In Section 2 we describe the general features of the SOMs, in Sections 3 and 4 we present two PDF fitting algorithms relying on the use of SOMs and finally in Section 5 we envision the possibilities the SOM method has to offer. 2 Self-Organizing Maps The SOM is a visualization algorithm which attempts to represent all the available observations with optimal accuracy using a restricted set of models. The SOM was developed by T. Kohonen in the early 1980’s ([28], see also [29]) to model biological brain functions, but has since then developed 5 into a powerful computational approach on it’s own right. Many fields of science, such as statistics, signal processing, control theory, financial analyses, experimental physics, chemistry and medicine, have adopted the SOM as a standard analytical tool. SOMs have been applied to texture discrimination, classification/pattern recognition, motion detection, genetic activity mapping, drug discovery, cloud classification, and speech recognition, among others. Also, a new application area is organization of very large document collections. However, applications in particle physics have been scarce so far, and mostly directed to improving the algorithms for background event rejection [30–32]. SOM consists of nodes, map cells, which are all assigned spatial coordinates, and the topology of the map is determined by a chosen distance metric Mmap . Each cell i contains a map vector Vi , that is isomorphic to the data samples used for training of the neural network. In the following we will concentrate on a 2-dimensional rectangular lattice for simplicity. A natural choice for the P topology is then L1 (x, y) = 2i=1 |xi − yi |, which also has been proved [33] to be an ideal choice for high-dimensional data, such as PDFs in our case. The implementation of SOMs proceeds in three stages: 1) initialization of the SOM, 2) training of the SOM and 3) associating the data samples with a trained map, i.e. clustering. During the initialization the map vectors are chosen such that each cell is set to contain an arbitrarily selected sample of either the actual data to be clustered, or anything isomorphic to them (see Fig. 2 for an example). The actual training data samples, which may be e.g. subset or the whole set of the actual data, are then associated with map vectors by minimizing a similarity metric Mdata . We choose Mdata = L1 . The map vector each data sample becomes matched against, is then the most similar one to the data sample among all the other map vectors. It may happen that some map vectors do no not have any samples associated with them, and some may actually have many. During the training the map vectors are updated by averaging them with the data samples that fell into the cells within a given decreasing neighbourhood, see Fig. 3. This type of training which is based on rewarding the winning node to become more like data, is called competitive learning. The initial value of a map vector Vi at SOM cell i then changes during the course of 6 Figure 2: A 2D grid SOM which cells get randomly associated with the type of data samples we would like to study, such as nonsinglet PDFs or observables. At this stage each cell gets associated with only one curve, the map vector. training as Vi (t + 1) = Vi (t) (1 − w(t) Nj,i (t)) + Sj (t) w(t) Nj,i(t) (3) where now Vi (t + 1) is the contents of the SOM cell i after the data sample Sj has been presented on the map. The neighbourhood, the radius, within which the map vectors are updated is given by the function Nj,i (t), centered on the winner cell j. Thus even the map vectors in those cells that didn’t find a matching data sample are adjusted, rewarded, to become more like 2 data. Typically Nj,i (t) = e−Mmap (j,i) /r(t) , where r(t) is a monotonously decreasing radius sequence. In the beginning of the training the neighbourhood may contain the whole map and in the end it just consists of the cell itself. Moreover, the updating is also controlled by w(t), which is a monotonously decreasing weight sequence in the range [0, 1]. As the training proceeds the neighbourhood function eventually causes the data samples to be placed on a certain region of the map, where the neighbouring map vectors are becoming increasingly similar to each other, and the weight sequence w(t) furthermore finetunes their position. In the end on a properly trained SOM, cells that are topologically close to each other will have map vectors which are similar to each other. In the final phase the actual data is matched against the map vectors of the trained map, and thus get distributed on the map according to the feature that was used 7 Figure 3: (Colour online) Each data sample Si is associated with the one map vector Vi it is most similar to. As a reward for the match, the winning map vector, as well as its neighbouring map vectors, get averaged with the data associated with the winning cell. as Mdata . Clusters with similar data now emerge as a result of unsupervised learning. For example, a map containing RGB colour triplets would initially have colours randomly scattered around it, but during the course of training it would evolve into patches of colour which smoothly blend with each other, see Fig. 4. This local similarity property is the feature that makes SOM suitable for visualization purposes, thus facilitating user interaction with the data. Since each map vector now represent a class of similar objects, the SOM is an ideal tool to visualize high-dimensional data, by projecting it onto a low-dimensional map clustered according to some desired similar feature. SOMs, however, also have disadvantages. Each SOM, though identical in size and shape and containing same type of data, is different. The clusters may also split in such a way that similar type of data can be found in several different places on the map. We are not aware of any mathematical or computational means of detecting if and when the map is fully trained, and whether there occurs splitting or not, other than actually computing the 8 Figure 4: (Colour online) SOM containing RGB colour triplets getting trained. Adapted from Ref. [34] similarities between the neighbouring cells and studying them. In this work we use the so-called batch-version of the training, in which all the training data samples are matched against the map vectors before the training begins. The map vectors are then averaged with all the training samples within the neighbourhood radius simultaneously. The procedure is repeated Nstep (free parameter to choose) times such that in every training step the same set of training data samples is associated with the evolving map and in Eq.(3) t now counts training steps. When the map is trained, the actual data is finally matched against the map vectors. In our study our training data are always going to be the whole data we want to cluster, and the last training step is thus the clustering stage. The benefit of the batch training compared to the incremental training, described earlier, is that the training is independent of the order in which the training samples are introduced on the map. 3 MIXPDF algorithm In this approach our aim is to both i) to be able to study the properties of the PDFs in a model independent way and yet ii) to be able to implement knowledge from the prior works on PDFs, and ultimately iii) to be able to guide the fitting procedure interactively with the help of SOM properties. At this stage it is important to distinguish between the experimental data 9 and the training data of the SOM. When we are referring to measured data used in the PDF fitting, such as F2 data, we always call it experimental data. The SOM training data in this study is going to be a collection of candidate PDF sets, produced by us, or something composed of them. A PDF set in the following will always mean a set of 8 curves, one for each independent ¯ s = s̄, c = c̄ and b = b̄ in this simplified parton flavour f = (g, uv , dv , ū, d, introductory study), that are properly normalized such that XZ 1 dxxff /p (x, Q2 ) = 1, (4) f 0 and conserve baryon number and charge Z 1 Z 1 2 dxfuv /p (x, Q ) = 2, dxfdv /p (x, Q2 ) = 1. 0 (5) 0 In order to proceed we have to decide how to create our candidate PDF sets, decide the details of the SOMs, details of the actual fitting algorithm, experimental data selection and details of the scale evolution. In this introductory paper our aim is not to provide a finalised SOMPDF set, but rather to explore the possibilities and restrictions of the method we are proposing. Therefore we refrain from using “all the possible experimental data” as used in global analyses, but concentrate on DIS structure function data from H1 [35], BCDMS [36, 37] and Zeus [38], which we use without additional kinematical cuts or normalization factors (except rejecting the data points below our initial scale). The parameters for the DGLAP scale evolution were chosen to be those of CTEQ6 (CTEQ6L1 for lowest order (LO)) [39], the initial scale being Q0 = 1.3 GeV. In next-to-leading order (NLO) case the evolution code was taken from [40] (QCDNUM17 beta release). We will start now with a simple pedagogical example, which we call MIXPDF algorithm, where we use some of the existing PDF sets as material for new candidate PDFs. At first, we will choose CTEQ6 [39], CTEQ5 [41], MRST02 [2, 42], Alekhin [5] and GRV98 [43] sets and construct new PDF sets from them such that at the initial scale each parton flavour in the range x = [10−5 , 1] is randomly selected from one of these five sets (we set the heavy flavours to be zero below their mass thresholds). The sumrules on this new 10 set are then imposed such that the original normalization of uv and dv are preserved, but the rest of the flavours are scaled together so that Eq.(4) is fulfilled. In this study we accept the <few% normalization error which results from the fact that our x-range is not x = [0, 1]. From now on we call these type of PDF sets database PDFs. We randomly initialize a small 5 × 5 map with these candidate database PDFs, such that each map vector Vi consists of the PDF set itself, and of the observables F2p (x, Q20 ) derived from it. Next we train the map with Nstep = 5 batch-training steps with training data that consists of 100 database PDFs plus 5 original “mother” PDF sets, which we will call init PDFs from now on. We choose the similarity criterion to be the similarity of observables F2p (x, Q2 ) with Mdata = L1 . The similarity is tested at a number of x-values (equidistant in logarithmic scale up to x ∼ 0.2, and equidistant in linear scale above that) both at the initial scale and at all the evolved scales where experimental data exist. On every training, after the matching, all the observables (PDFs) of the map vectors get averaged with the observables (PDFs, flavor by flavor) matched within the neighbourhood according to Eq. (3). The resulting new averaged map vector PDFs are rescaled again (such that uv and dv are scaled first) to obey the sumrules. From now on we will call these type of PDF sets map PDFs. The map PDFs are evolved and the observables at every experimental data scale are computed and compared for similarity with the observables from the training PDFs. After the training we have a map with 25 map PDFs and the same 105 PDF sets we used to train the map. The resulting LO SOM is shown in Fig. 5, with just F2p (x, Q20 )’s of each cell shown for clarity. The red curves in the figure are the map F2’s constructed from the map PDFs, black curves are the database F2’s and green curves are the init F2’s constructed from the init PDFs (CTEQ6, CTEQ5, MRST02, Alekhin and GRV98 parametrizations). It is obviously difficult to judge visually just by looking at the curves whether the map is good and fully trained. One hint about possible ordering may be provided by the fact that the shape of the F2p curve must correlate with the χ2 /N against the experimental data. The distribution of the χ2 /N values (no correlated systematic errors are taken into account for simplicity) of the map PDFs, shown in each cell , does indeed seem somewhat organized. 11 1 2 3 4 2.46 2.20 4.36 4.98 3.28 2.49 2.38 1.54 2.59 4.04 3.22 2.51 1.45 2.75 6.55 5.51 4.27 3.26 2.00 9.49 7.24 4.72 4.58 2.32 4 3 2 1 0 0 LO 2.44 1. iteration -5 10 10 -4 -3 10 10 -2 10 -1 1 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 Figure 5: (Colour online) Trained 1. iteration LO map. Red curves are map F2p ’s, green curves init F2p ’s and black curves rest of the database F2p ’s. The number in each cell is the χ2 /N of the map PDF. Table 1 lists the χ2 /N values the original init PDFs obtain within the MIXPDF framework described above. Comparison of these values with the values in Fig. 5 reveals that some of the map PDFs, as also some of the database PDFs, have gained a χ2 /N comparable to or better than that of the init PDFs. Inspired by the progress, we start a second fitting iteration by selecting the 5 best PDF sets from the 25+5+100 PDF sets of the first iteration as our new init PDFs (which are now properly normalized after the first iteration) to generate database PDFs for a whole new SOM. Since the best PDF candidate from the first iteration is matched on this new map as an unmodified init PDF, it is guaranteed that the χ2 /N as a function of the iteration either decreases or remains the same. We keep repeating the iterations until the χ2 /N saturates. Fig. 6 (Case 1) shows the χ2 /N as a function of iterations for the best PDF on the trained map, for the worst PDF on the map and for 12 PDF‡ Alekhin CTEQ6 CTEQ5 CTEQ4 MRST02 GRV98 LO χ2 /N 3.34 1.67 3.25 2.23 2.24 8.47 NLO χ2 /N 29.1 2.02 6.48 2.41 1.89 9.58 Table 1: χ2 /N for different MIXPDF input PDF sets against all the datasets used (H1, ZEUS, BCDMS, N=709). the worst of the 5 PDFs selected for the subsequent iteration as an init PDF. The final χ2 /N of these runs are listed in Table 2 (first row) as Case 1 and Fig. 7 shows these results (black solid line), together with original starting sets at the initial scale (note the different scaling for gluons in LO and in NLO figures). For 10 repeated LO runs we obtain χ¯2 /N=1.208 and σ=0.029. Let us now analyze in more detail how the optimization proceeds. Figs. 8,9 show the LO maps also for the 2. and 3. iterations. On the first iteration the init PDFs, the shapes of which were taken from existing parametrizations, fall in the cells (0,1) (CTEQ5), (1,3) (Alekhin), (1,4) (MRST02), (2,3) (CTEQ6) and (3,0) (GRV98), so the modern sets, Alekhin, MRST02 and CTEQ6, group close to each other, i.e. the shapes of the observables they produce are very similar, as expected. The best 5 PDFs selected as the 2. iteration init PDFs also come from this neighbourhood, 3 of them from the cell (1,3) and the other 2 from the cell (2,3). Two of these selected sets are map PDFs, two are database PDFs and also the original init PDF CTEQ6 survived for the 2. iteration. At the end of the 2. iteration the init PDFs, which originated from the neighbouring cells, are scattered in the cells (0,1), (0.3), (1,0) (CTEQ6), ‡ These are the χ2 /N for the initial scale PDF sets taken from the quoted parametrizations and evolved with CTEQ6 DGLAP settings, the heavy flavours were set to be zero below their mass thresholds, no kinematical cuts or normalization factors for the experimental data were imposed, and no correlated systematic errors of the data were used to compute the χ2 /N . We do not claim these values to describe the quality of the quoted PDF sets. 13 4.0 4.0 LO 3.5 NLO Case 1: best of 5 worst of 5 worst of map 5 5, Nstep=5 3.0 Case 1: Case 2: 3.0 Case 2: 2 2.0 best of 5 worst of 5 worst of map 2.5 2.0 1.5 1.5 1.0 1.0 0.5 0.5 0 1 2 3 4 5 6 7 8 Iteration 0 1 2 3 4 5 6 7 8 Iteration Figure 6: (Colour online) χ2 /N of the MIXPDF runs as a function of the iteration. (2,2) and (2,3) and the best 5 PDFs produced during this iteration are in the cells (4,1) (database PDF), (4,0) (map PDF), (3,2) (map PDF), (2,2) (database PDF) and (3,0) (map PDF). After the 3. iteration the above best 5 PDFs of 2. iteration are in the cells (2,0), (3,1), (0.2), (0.4) and (3,2) and the new best 5 PDFs produced are all map PDFs with 4 of them in neighbouring cell pairs. Map PDFs provide complicated linear combinations of the database PDFs and obviously play an important role in the algorithm. The size of the map dictates how much the neighbouring map vectors differ from each other. Since the PDFs in the same cell are not required to be similar, only the observables constructed from them are, a cell or a neighbourhood may in principle contain a spread of PDFs with a spread of χ2 /N’s. However, since our selection criteria for the init PDFs was based on the best χ2 /N only, it is inevitable that the observables on the map become increasingly similar as the iterations go by, and the χ2 /N flattens very fast as can be seen from Fig. 6. As a consequence we quickly lose the variety in the shapes of the PDFs as the iterations proceed, and on the final iteration all the PDFs on the map end up being very similar. 14 /N best of 5 worst of 5 worst of map 2.5 2 /N 3.5 best of 5 worst of 5 worst of map 5 5, Nstep=5 SOM 5x5 5x5 5x5 5x5 5x5 5x5 5x5 5x5 15x15 15x15 15x15 15x15 Nstep 5 5 5 5 5 5 10 40 5 5 5 5 # init 5 5 10 10 15 20 10 10 5 5 30 30 # database 100 100 100 100 100 100 100 100 900 900 900 900 Case 1 2 1 2 1 1 1 1 1 2 1 2 LO χ2 /N 1.19 1.37 1.16 1.49 1.16 1.17 1.16 1.20 1.22 1.31 1.16 1.25 NLO χ2 /N 1.28 1.44 1.25 1.43 1.45 1.30 1.25 l.53 Table 2: χ2 /N against all the datasets used (H1, ZEUS, BCDMS) for some selected MIXPDF runs. The MIXPDF algorithm obviously has several other weaknesses too. Among them are how the result would change if we started with another first iteration PDF selection, and what are the effects of changing the size of the map, number of the database PDFs and init PDF sets and size of Nstep ? In general, how much the final result depends on the choices that we make during the fitting process? Let us now study some of these questions a bit. Since we have chosen our evolution settings to be those of CTEQ6’s, it necessarily becomes a favoured set (although we don’t impose any kinematical cuts on the experimental data). Therefore we performed another LO and NLO runs, with CTEQ6 now replaced with CTEQ4 [44]. The results of these runs are reported in Table 2 (2. row) and in Fig. 6 (χ2 /N) and Fig. 7 (the PDFs) as Case 2 . The Case 2 clearly produces worse results. Without an input from CTEQ6 we automatically lose all the low gluons at small-x -type of results in NLO, for example. Fig. 10 addresses the issue of choosing different Nstep , for LO Case 1 and Case 2. The solid (dotted) lines show the best (worst) init PDF selected 15 4.0 4.0 Q=1.3 GeV CTEQ6 Alekhin MRST02 GRV98 CTEQ5 CTEQ4 LO 3.5 3.0 0.25*xg 2.5 Q=1.3 GeV CTEQ6 Alekhin MRST02 GRV98 CTEQ5 CTEQ4 NLO 0.45*xg 5 5, Nstep= 5 Case 1 Case 2 5 5, Nstep= 5 Case 1 Case 2 15 15, Nstep= 5 Case 1 Case 2 15 15, Nstep= 5 Case 1 Case 2 3.5 3.0 2.5 2.0 2.0 1.5 1.0 xuV xu 1.5 1.0 xuV xu 0.5 0.0 -5 10 0.5 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 7: (Colour online) MIXPDF results together with the input PDF sets. from each iteration for several Nstep selections. In Case 2 the best results are obtained with small number of training steps, whereas Case 1 does not seem to benefit from a longer SOM training. Keeping the stochastical nature of the process in our minds, we may speculate that the seemingly opposite behaviour for the Case 1 and Case 2 results from the fact that it is more probable to produce a good set of database PDFs in Case 1 than in Case 2. If the database is not so good to begin with, the longer training contaminates all the map PDFs with the low quality part of the database PDFs. Table 2 also showcases best results from a variety of MIXPDF runs where we have tried different combinations of SOM features. It is interesting to notice that increasing the size of the map and database does not necessarily lead to a better performance. Instead the number of the init PDFs on later iterations (always 5 for the first iteration) seem to be a key factor, it has to be sufficiently large to preserve the variety of database the PDFs but small enough to consist of PDFs with good quality. In our limited space of database candidates (∼ 65 = 7776 possibilities) the optimal set of variables 16 1 2 3 4 1.43 1.76 1.70 1.72 1.38 1.59 1.57 1.65 1.51 1.37 1.46 1.53 1.39 1.42 1.35 1.36 1.33 1.41 1.60 1.33 1.47 1.65 1.67 1.67 4 3 2 1 0 0 LO 1.41 2. iteration -5 10 10 -4 -3 10 10 -2 10 -1 1 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 Figure 8: (Colour online) Trained 2. iteration LO map, curves and numbers as in Fig.5. for the MIXPDF run should not be not impossible to map down. The method we used to generate the sample data is very simple indeed. The number of all the possible candidate database PDFs is not very large to begin with, so the quality of the final results strongly depends on the quality of the input for the SOM. Since the map PDFs are obtained by averaging with the training samples, and the non-valence flavours are scaled by a common factor when imposing the sumrules, the map PDFs tend to lie in between the init PDFs. Therefore map PDFs with extreme shapes are never produced, and thus never explored by the algorithm. A method which relies on sampling existing parametrizations on a SOM is inconvenient also because it is not readily applicable to multivariable cases. For the case of the PDFs it is sufficient to have a value for each flavour for a discrete set of x-values, but for a multivariable cases, such as nPDFs, or GPDs the task of keeping track of the grid of values for each flavour in each x for several different values of the additional variables (e.g. A and Z 17 1 2 3 4 1.26 1.26 1.28 1.36 1.35 1.26 1.26 1.31 1.33 1.36 1.29 1.25 1.30 1.37 1.36 1.32 1.27 1.35 1.35 1.41 1.29 1.27 1.30 1.42 4 3 2 1 0 0 LO 1.34 3. iteration -5 10 10 -4 -3 10 10 -2 10 -1 1 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 Figure 9: (Colour online) Trained 3. iteration LO map, curves and numbers as in Fig.5. for nuclei, and the skewness, ξ and the squared 4-momentum transfer, t for GPDs) is computationally expensive. In principle, SOM can keep track of the interrelations of the map vectors, and knowing the parametrization for the init PDFs, it would be possible to construct the parametrization for the map PDFs. That would, however, lead to very complicated parametrizations, and different nPDF, GPD etc. parametrizations are presently not even either measured or defined at the same initial scale. Despite of its problems, on a more basic level, MIXPDF does have the the desirable feature that it allows us to use SOM as a part of the PDF optimization algorithm in such a way that we cluster our candidate PDFs on the map, and select those PDFs which i) minimize a chosen fitness function, e.g. χ2 , when compared against experimental data, and ii) have some desired feature which can be visualized on the map or used as a clustering criterion. Therefore, in the following Section, we keep building on this example. 18 2.2 2.2 LO, 5 5 LO, 5 5 Nstep: Case 1 Case 2 best of 5, solid lines worst of 5, dashed lines 2.0 1.8 1.6 best of 5, solid lines worst of 5, dashed lines 1.4 1.4 1.2 1.2 1.0 1.0 0 1 2 3 4 5 6 7 8 Iteration 0 1 2 3 4 5 6 7 8 Iteration Figure 10: (Colour online) LO χ2 /N for 5 × 5 SOM runs with different Nstep . 4 ENVPDF algorithm Most of the problems with the MIXPDF algorithm originate from the need to be able to generate the database PDFs in an unbiased way as possible, and at the same time to have a variety of PDF candidates available at every stage of the fitting procedure. Yet, one needs to have control over the features of the database PDFs that are created. To accomplish this, we choose, at variance with the “conventional” PDFs sets or NNPDFs, to give up the functional form of PDFs and rather to rely on purely stochastical methods in generating the initial and training samples of the PDFs. Our choice is a GA-type analysis, in which our parameters are the values of PDFs at the initial scale for each flavour at each value of x where the experimental data exist. To obtain control over the shape of the PDFs we use some of the existing distributions to establish an initial range, or envelope, within which we sample the database PDF values. Again, we use the Case 1 and 2 PDF sets (CTEQ6, CTEQ5, CTEQ4, MRST02, Alekhin and GRV98) as an initialization guideline. We construct our initial PDF generator first to, for each flavour separately, select ran19 /N 2 5 10 20 30 40 2 /N 1.8 1.6 Nstep: 5 10 20 30 40 2.0 domly either the range [0.5, 1], [1.0, 1.5] or [0.75, 1.25] times any of the Case 1 (or 2) PDF set. Compared to MIXPDF algorithm we are thus adding more freedom to the scaling of the database PDFs. Next the initial generators generate values for each xdata § using uniform, instead of Gaussian, distribution around the existing parametrizations, thus reducing direct bias from them. Gaussian smoothing is applied to the resulting set of points, and the flavours combined to form a PDF set such that the curve is linearly interpolated from the discrete set of generated points. The candidate PDF sets are then scaled to obey the sumrules as in MIXPDF algorithm. In order to obtain a reasonable selection of PDFs to start with, we reject candidates which have χ2 /N > 10 (computed as in MIXPDF algorithm). To further avoid direct bias from the Case 1 and 2 PDFs, we don’t include the init PDFs into the training set for the first iteration as we did in MIXPDF case. For a N × N SOM we choose the size of the database to be 4N 2 . During the later iterations we proceed as follows: At the end of each iteration we pick from the trained N ×N SOM 2N best PDFs as the init PDFs. These init PDFs are introduced into the training set alongside with the database PDFs, which are now constructed using each of the init PDFs in turn as a center for a Gaussian random number generator, which assigns for all the flavours for each x a value around that same init PDF such that 1 − σ of the generator is given by the spread of the best PDFs in the topologically nearest neighbouring cells. The object of these generators is thus to refine a good candidate PDF found in the previous iteration by jittering it’s values within a range determined by the shape of other good candidate PDFs from the previous iteration. The generated PDFs are then smoothed and scaled to obey the sumrules. Sets with χ2 /N > 10 are always rejected. We learnt from the MIXPDF algorithm that it is important to preserve the variety of the PDF shapes on the map, so we also keep Norig copies of the first iteration generators in our generator mix. Table 3 lists results from a variety of such runs. The results do not seem § To ensure a reasonable large-x behaviour for the PDFs, we also generate with the same method values for them in a few x-points outside the range of the experimental data. We also require the PDFs, the gluons especially, to be positive for simplicity. 20 to be very sensitive to the number of SOM training steps, Nstep , but are highly sensitive to the number of first iteration generators used in subsequent iterations. Although the generators can now in principle produce an infinite number of different PDFs, the algorithm would not be able to radically change the shape of the database PDFs without introducing a random element on the map. Setting Norig > 0 provides, through map PDFs, that element, and keeps the algorithm from getting fixed to a local minimum. The ENVPDF algorithm is now more independent from the initial selection of the PDF sets, Case 1 or 2, than MIXPDF, since no values of e.g. the CTEQ6 set in the original generator are ever introduced on the map directly. SOM Nstep 5x5 5 5x5 10 5x5 20 5x5 30 5x5 40 5x5 5 5x5 20 5x5 5 5x5 10 5x5 15 15x15 5 15x15 5 Norig 2 2 2 2 2 0 0 2 2 2 6 6 Case 1 1 1 1 1 1 1 2 2 2 1 2 LO χ2 /N 1.04 1.10 1.10 1.10 1.08 1.41 1.26 1.14 1.12 1.18 1.00 1.13 NLO χ2 /N 1.08 1.25 1.07 1.18 Table 3: χ2 /N against all the datasets used (H1, ZEUS, BCDMS) for variety of ENVPDF runs. Fig. 11 shows the χ2 /N as a function of iteration for 5x5 LO and NLO, both Case 1 and Case 2, runs, where Nstep = 5 and Norig = 2. Clearly the ENVPDF runs take multiple number of iterations for the χ2 /N to level compared to the MIXPDF runs, and they are therefore more costly in time. With the ENVPDF algorithm, however, the χ2 /N keeps on slowly improving even after all the mother PDFs from the same iteration are equally good fits. For a larger 15x15 SOM the number of needed iterations remains as large. 21 3.0 3.0 LO 2.5 NLO Case 1: 5 5, Nstep=5 Case 1: 5 5, Nstep=5 best of 10 worst of 10 Case 2: Case 2: 2.0 2 best of 10 worst of 10 /N best of 10 worst of 10 2 /N 2.0 2.5 best of 10 worst of 10 1.5 1.5 1.0 1.0 0.5 0.5 0 5 10 15 20 25 30 35 40 45 50 Iteration 0 5 10 15 20 25 30 35 40 45 50 Iteration Figure 11: (Colour online) χ2 /N of the ENVPDF runs as a function of the iteration. Fig. 12 shows some of the Case 1 LO ENVPDF results at the initial scale Q = 1.3 GeV (left panel), and evolved up to Q = 3 GeV (right panel). The reference curves shown are also evolved as in MIXPDF. Although the initial scale ENVPDF results appear wiggly, they smooth out soon because of the additional well known effect of QCD evolution. In fact, the initial scale curves could be made smoother by applying a stronger Gaussian smoothing, but this is not necessary, as long as the starting scale is below the Qmin of the data. The evolved curves preserve the initially set baryon number scaling within 0.5% and momentum sumrule within 1.5% accuracy. Also, the results obtained from a larger map tend to be smoother since the map PDFs get averaged with a larger number of other PDFs. Studying the relation between the redundant wiggliness of our initial scale PDFs and possible fitting of statistical fluctuations of the experimental data is beyond the scope of this paper. The NLO Case 1 and 2 results are presented in Fig. 13. The trend of the results is clearly the same as in MIXPDF case, CTEQ6 is a favoured set, and especially the PDFs with gluons similar to those of CTEQ6’s have good χ2 /N. 22 We did not study the effect of modifying the width or the shape of the envelope in detail here, but choosing the envelope to be the wider or narrower than 1 − σ for the Gaussian generate seem to lead both slower and poorer convergence. Also, since we are clustering on the similarity of the observables, the same cell may in theory contain the best PDF of the iteration and PDFs which have χ2 /N as large as 10. Therefore the shape of the envelope should be determined only by the curves with promising shapes. 4.0 2.5 LO LO 3.5 Q=1.3 GeV CTEQ6 MRST02 CTEQ4 2.0 0.25*xg 1.5 Q=3.0 GeV CTEQ6 MRST02 CTEQ4 0.1*xg 5 5, Nstep= 5 Case 1 Case 2 5 5, Nstep= 5 Case 1 Case 2 15 15, Nstep= 5 Case 1 Case 2 15 15, Nstep= 5 Case 1 Case 2 3.0 2.5 2.0 1.0 1.5 xuV xu 1.0 xu 0.5 xuV 0.5 0.0 -5 10 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 12: (Colour online) LO ENVPDF results at the initial scale, and at Q = 3.0 GeV. Next we want to study the PDF uncertainty using the unique means the SOMs provide to us even for a case of PDFs without a functional form. Since we have only used DIS data in this introductory study, we are only able to explore the small-x uncertainty for now. Figs. 14 (LO) and 15 (NLO) showcase, besides our best results, the spread of all the initial scale PDFs with χ2 /N ≤ 1.2, that were obtained during a 5 × 5 (left panel) and 15 × 15 (right panel) SOM run. Since the number of such PDF sets is typically of the order of thousands, we only plot the minimum and maximum of the bundle 23 4.0 3.0 Q=1.3 GeV CTEQ6 MRST02 CTEQ4 2.5 2.0 15 15, Nstep= 5 Case 1 Case 2 NLO NLO 3.5 Q=3.0 GeV CTEQ6 MRST02 CTEQ4 0.25*xg 5 5, Nstep= 5 Case 1 Case 2 0.85*xg 3.0 5 5, Nstep= 5 Case 1 Case 2 2.5 2.0 1.5 15 15, Nstep= 5 Case 1 Case 2 1.5 1.0 xuV 1.0 xu xuV 0.5 xu 0.0 -5 10 0.5 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 13: (Colour online) NLO ENVPDF results at the initial scale, and at Q = 3.0 GeV. of curves. Since the total number of experimental datapoints used is ∼ 710, the spread ∆χ2 /N ∼ 0.2 corresponds to a ∆χ2 ∼ 140. Expectedly, the smallx gluons obtain the largest uncertainty for all the cases we studied. Even though a larger SOM with a larger database might be expected to have more variety in the shapes of the PDFs, the χ2 /N ≤ 1.2 spreads of the 5 × 5 and 15 × 15 SOMs are more or less equal sized (the apparent differences in sizes at Q = Q0 even out when the curves are evolved). Both maps therefore end up producing the same extreme shapes for the map PDFs although a larger map has more subclasses for them. Remarkably then, a single SOM run can provide a quick uncertainty estimate for a chosen ∆χ2 without performing a separate error analysis. Due to the stochastical nature of the ENVPDF algorithm, we may well also study the combined results from several separate runs. It is especially important to verify the stability of our results, to show that the results are indeed reproducible instead of lucky coincidences. Left panels of Figs 16 (LO) and 24 3.0 3.0 LO LO 2.5 2.5 Q=1.3 GeV CTEQ6 MRST02 2.0 Q=1.3 GeV CTEQ6 MRST02 5 5, Nstep= 5 Case 1 2 /N 1.2 0.25*xg 2.0 15 15, Nstep= 5 Case 1 2 /N 1.2 0.25*xg 1.5 1.5 1.0 1.0 xuV xuV xu xu 0.5 0.0 -5 10 0.5 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 14: (Colour online) LO ENVPDF best result and the χ2 /N ≤ 1.2 spread of results. 17 (NLO) present the best results, and the combined χ2 /N ≤ 1.2 spreads for 10 repeated 5 × 5, Nstep = 5 runs at the initial scale. The average χ2 /N and the standard deviation σ for these runs are in LO (NLO) are 1.065 and 0.014 (1.122 and 0.029), corresponding to ∆χ2 ∼ 10 (20) for LO (NLO). The right panels of the same Figs 16, 17 show the 10 best result curves and the χ2 /N ≤ 1.2 spreads evolved up to Q = 3.0 GeV. Clearly the seemingly large difference between the small-x gluon results at the initial scale is not statistically significant, but smooths out when gluons are evolved. Thus the initial scale wiggliness of the PDFs is mainly only a residual effect from our method of generating them and not linked to the overtraining of the SOM, and we refrain from studying cases where stronger initial scale smoothing is applied. Therefore our simple method of producing the candidate PDFs by jittering random numbers inside a predetermined envelope is surprisingly stable when used together with a complicated PDF processing that SOMs provide. 25 3.5 3.5 Q=1.3 GeV CTEQ6 MRST02 NLO 3.0 NLO 5 5, Nstep= 5 Case 1 2 /N 1.2 2.5 0.85*xg Q=1.3 GeV CTEQ6 MRST02 3.0 15 15, Nstep= 5 Case 1 2 /N 1.2 0.85*xg 2.5 2.0 2.0 1.5 1.5 1.0 1.0 xuV 0.5 xuV xu 0.5 xu 0.0 -0.5 -5 10 0.0 -0.5 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 15: (Colour online) NLO ENVPDF best result and the χ2 /N ≤ 1.2 spread of results. 5 Future of the SOMPDFs So far we have shown a relatively straightforward method of obtaining stochastically generated, parameter-free, PDFs, with an uncertainty estimate for a desired ∆χ2 . On every iteration using our competitive learning algorithm, the selection of the winning PDFs was based on the χ2 /N alone, and the fitting procedure was fully automated. In our MIXPDF algorithm the SOMs were used merely as a tool to create new combinations, map PDFs, of our input database. The ENVPDF algorithm also used the topology of the map to determine the shape of the envelope, within which we sampled the database PDFs. We reiterate that our initial study was aimed at observing and recording the behavior of the SOM as an optimization tool. Many of the features of our results could not in fact be predicted based on general assumptions. The proposed method can be extended much further than that. The automated version of the algorithm could be set to sample a vector consisting of PDF 26 4.0 3.0 LO LO 3.5 2.5 Q=1.3 GeV CTEQ6 MRST02 2.0 0.25*xg Q=3.0 GeV CTEQ6 MRST02 0.1*xg 5 5, Nstep= 5 Case 1 2 /N 1.2 3.0 5 5, Nstep= 5 Case 1 2 /N 1.2 2.5 2.0 1.5 1.5 1.0 xuV 0.5 1.0 xu xu xuV 0.5 0.0 -5 10 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 16: (Colour online) LO ENVPDF best results and the χ2 /N ≤ 1.2 spreads of results from 10 separate runs. parameters, instead of values of PDFs in each value of x of the data. That would lead to smooth, continuous type of solutions, either along the lines of global analyses, or NNPDFs using N SOMs for N Monte-Carlo sampled replicas of the data. Since the solution would be required to stay within an envelope of selected width and shape given by the map, no restrictions for the parameters themselves would be required. For such a method, all the existing error estimates, besides an uncertainty band produced by the map, would be applicable as well. What ultimately sets the SOM method apart from the standard global analyses or NNPDF method, however, are the clustering and visualization possibilities that it offers. Instead of setting Mdata = L1 and clustering according to the similarity of the observables, it is possible to set the clustering criteria to be anything that can be mathematically quantified, e.g. the shape of the gluons or the large-x behaviour of the PDFs. The desired feature of the PDFs can then be projected out from the SOM. Moreover, by combining the 27 4.0 3.5 Q=1.3 GeV CTEQ6 MRST02 NLO 3.0 NLO 3.5 5 5, Nstep= 5 Case 1 2 /N 1.2 2.5 Q=3.0 GeV CTEQ6 MRST02 0.25*xg 0.85*xg 3.0 5 5, Nstep= 5 Case 1 2 /N 1.2 2.0 2.5 2.0 1.5 1.0 1.5 xuV 0.5 1.0 xu xu xuV 0.0 -0.5 -5 10 0.5 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 -5 1 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 17: (Colour online) NLO ENVPDF best results and the χ2 /N ≤ 1.2 spreads of results from 10 separate runs. method with an interactive graphic user interface (GUI), it would be possible to change and control the shape and the width of the envelope as the minimization proceeds, to guide the process by applying researcher insight at various stages of the process. Furthermore, the uncertainty band produced by the SOM as the run proceeds, could help the user to make decisions about the next steps of the minimization. With GUI it would be e.g. possible to constrain the extrapolation of the NN generated PDFs outside the x-range of the data without explicitly introducing terms to ensure the correct smalland large-x behaviour as in NNPDF method (see Eq.(87) in [23]). The selection of the best PDF candidates for the subsequent iteration could then be made based on the user’s preferences instead of solely based on the χ2 /N. That kind of method in turn could be extended to multivariable cases such as nPDFs and even GPDs and other not so well-known cases, where the data is too sparse for stochastically generated, parameter-free, PDFs. Generally, any PDF fitting method involves a large number of flexible points 28 “opportunities for adapting and fine tuning”, which act as a source for both systematical and theoretical bias when fixed. Obvious optimization method independent sources of theoretical bias are the various parameters of the DGLAP equations, inclusion of extra sources of Q2 -dependence beyond DGLAP-type evolution and the data selection, affecting the coverage of different kinematical regions. SOMs themselves, and different SOMPDF algorithm variations naturally also introduce flexible points of their own. We explored a little about the effects of choosing the size of the SOM and the number of the batch training steps Nstep . There are also plenty of other SOM properties that can be modified, such as the shape of the SOM itself. We chose to use a rectangular lattice, but generally the SOM can take any shape desired. For demanding vizualisation purposes a hexagonal shape is an excellent choice, since the meaning of the nearest neighbours is better defined. The SOMPDF method, supplemented with the use of a GUI, will allow us to both qualitatively and quantitatively study the flexible points involved in the PDFs fitting. More complex hadronic matrix elements, such as the ones defining the GPDs, are natural candidates for future studies of cases where the experimental data are not numerous enough to allow for a model independent fitting, and the guidance and intuition of the user is therefore irreplaceable. The method we are proposing is extremely open for user interaction, and the possibilities of such a method are widely unexplored. Acknowledgements We thank David Brogan for helping us to start this project. This work was financially supported by the US National Science Foundation grant no.0426971. HH was also supported by the U.S. Department of Energy, grant no. DE-FG02-87ER40371. SL is supported by the U.S. Department of Energy, grant no. DE-FG02-01ER41200. 29 References [1] P. M. Nadolsky et al., arXiv:0802.0007 [hep-ph]. [2] A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Eur. Phys. J. C 23 (2002) 73 [arXiv:hep-ph/0110215]. [3] A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Phys. Lett. B 604 (2004) 61 [arXiv:hep-ph/0410230]. [4] A. D. Martin, W. J. Stirling, R. S. Thorne and G. Watt, Phys. Lett. B 652 (2007) 292 [arXiv:0706.0459 [hep-ph]]. [5] S. Alekhin, Phys. Rev. D 68 (2003) 014002 [arXiv:hep-ph/0211096]. [6] S. Alekhin, K. Melnikov and F. Petriello, Phys. Rev. D 74 (2006) 054033 [arXiv:hep-ph/0606237]. [7] S. Chekanov et al. [ZEUS Collaboration], Eur. Phys. J. C 42 (2005) 1 [arXiv:hep-ph/0503274]. [8] C. Adloff et al. [H1 Collaboration], Eur. Phys. J. C 30 (2003) 1 [arXiv:hep-ex/0304003]. [9] J. Pumplin et al., Phys. Rev. D 65 (2002) 014013 [arXiv:hepph/0101032]. [10] A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Eur. Phys. J. C 35 (2004) 325 [arXiv:hep-ph/0308087]. [11] A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Eur. Phys. J. C 28 (2003) 455 [arXiv:hep-ph/0211080]. [12] K. J. Eskola, H. Paukkunen and C. A. Salgado, JHEP 0807 (2008) 102 [arXiv:0802.0139 [hep-ph]]. 30 [13] K. J. Eskola, V. J. Kolhinen, H. Paukkunen and C. A. Salgado, JHEP 0705 (2007) 002 [arXiv:hep-ph/0703104]. [14] M. Hirai, S. Kumano and T. H. Nagai, Phys. Rev. C 70 (2004) 044905 [arXiv:hep-ph/0404093]. [15] M. Hirai, S. Kumano and T. H. Nagai, Phys. Rev. C 76 (2007) 065207 [arXiv:0709.3038 [hep-ph]]. [16] M. Vanderhaeghen, P. A. M. Guichon and M. Guidal, Phys. Rev. D 60 (1999) 094017 [arXiv:hep-ph/9905372]. [17] S. Ahmad, H. Honkanen, S. Liuti and S. K. Taneja, Phys. Rev. D 75 (2007) 094003 [arXiv:hep-ph/0611046]. [18] S. Ahmad, H. Honkanen, S. Liuti and S. K. Taneja, arXiv:0708.0268 [hep-ph]. [19] D. Stump et al., Phys. Rev. D 65 (2002) 014012 [arXiv:hep-ph/0101051]. [20] J. Pumplin, AIP Conf. Proc. 792, 50 (2005) [arXiv:hep-ph/0507093]. [21] A. Djouadi and S. Ferrag, Phys. Lett. B 586 (2004) 345 [arXiv:hepph/0310209]. [22] M. Ubiali, arXiv:0809.3716 [hep-ph]. [23] R. D. Ball et al., arXiv:0808.1231 [hep-ph]. [24] L. Del Debbio, S. Forte, J. I. Latorre, A. Piccione and J. Rojo [NNPDF Collaboration], JHEP 0703 (2007) 039 [arXiv:hep-ph/0701127]. [25] W. T. Giele, S. A. Keller and D. A. Kosower, arXiv:hep-ph/0104052. [26] G. Cowan, Prepared for 14th International Workshop on Deep Inelastic Scattering (DIS 2006), Tsukuba, Japan, 20-24 Apr 2006 31 [27] G. D’Agostini, “Bayesian Reasoning in Data Analysis: A Critical Introduction”, World Scientific Publishing, Singapore (2003). [28] T. Kohonen, Self-organizing Maps, Third Edition, Springer (2001). [29] http://www.cis.hut.fi/projects/somtoolbox [30] J. S. Lange, H. Freiesleben and P. Hermanowski, Nucl. Instrum. Meth. A 389 (1997) 214. [31] K. H. Becks, J. Drees, U. Flagmeyer and U. Muller, Nucl. Instrum. Meth. A 426 (1999) 599. [32] J. S. Lange, C. Fukunaga, M. Tanaka and A. Bozek, Nucl. Instrum. Meth. A 420 (1999) 288. [33] Charu C. Aggarwal , Alexander Hinneburg , Daniel A. Keim, Proceedings of the 8th International Conference on Database Theory, p.420-434, January 04-06, 2001 [34] http://www.ai-junkie.com/ann/som/som1.html and http://davis.wpi.edu/~ matt/courses/soms/ [35] C. Adloff et al. [H1 Collaboration], Eur. Phys. J. C 21 (2001) 33 [arXiv:hep-ex/0012053]. [36] A. C. Benvenuti et al. [BCDMS Collaboration], Phys. Lett. B 223 (1989) 485. [37] A. C. Benvenuti et al. [BCDMS Collaboration], Phys. Lett. B 237 (1990) 592. [38] S. Chekanov et al. [ZEUS Collaboration], Eur. Phys. J. C 21 (2001) 443 [arXiv:hep-ex/0105090]. 32 [39] J. Pumplin, D. R. Stump, J. Huston, H. L. Lai, P. Nadolsky and W. K. Tung, JHEP 0207 (2002) 012 [arXiv:hep-ph/0201195]. [40] http://www.nikhef.nl/~ h24/qcdnum/ [41] H. L. Lai et al. [CTEQ Collaboration], Eur. Phys. J. C 12 (2000) 375 [arXiv:hep-ph/9903282]. [42] A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Phys. Lett. B 531 (2002) 216 [arXiv:hep-ph/0201127]. [43] M. Gluck, E. Reya and A. Vogt, Eur. Phys. J. C 5 (1998) 461 [arXiv:hepph/9806404]. [44] H. L. Lai et al., Phys. Rev. D 55, 1280 (1997) [arXiv:hep-ph/9606399]. 33
5cs.CE
Published in Journal of Telematics and Informatics, Vol 4, No 1 (2016): March 2016 Abstract: http://section.iaesonline.com/index.php/JTI/article/view/188 One-Minute Derivation of The Conjugate Gradient Algorithm Muhammad Ali Raza Anjum Army Public College of Management and Sciences Rawalpindi, PAKISTAN e-mail: [email protected] Abstract One of the great triumphs in the history of numerical methods was the discovery of the Conjugate Gradient (CG) algorithm. It could solve a symmetric positive-definite system of linear equations of dimension N in exactly N steps. As many practical problems at that time belonged to this category, CG algorithm became rapidly popular. It remains popular even today due to its immense computational power. But despite its amazing computational ability, mathematics of this algorithm is not easy to learn. Lengthy derivations, redundant notations, and over-emphasis on formal presentation make it much difficult for a beginner to master this algorithm. This paper aims to serve as a starting point for such readers. It provides a curt, easy-to-follow but minimalist derivation of the algorithm by keeping the sufficient steps only, maintaining a uniform notation, and focusing entirely on the ease of reader. Keywords: Conjugate Gradient, Algorithm, Optimization 1. Introduction Importance of the Conjugate Gradient (CG) algorithm can hardly be overemphasized. It is everywhere: in optimization theory [1, 2], in adaptive filtering [3, 4], in machine learning [5, 6], in image processing [7, 8] and in many others [9-11]. Though it is popular, the algorithm is not easy to learn. The chief difficulty lies with its derivation. Reasons for this difficulty are as numerous as the reasons for its popularity. Too many approaches, too much variety in notation and over emphasis on form and presentation are just to name a few. Books do not try to resolve these issues. Research papers avoid resolving these issues. There must be an excellent cause for them to do so but they all essentially miss one point: to convey this fascinating algorithm to the aspiring student in a quick, easy and understandable way. Tutorials have made an attempt in this regard. Though they are scarce as well, two of them are truly remarkable. First one tries to systematically address the above mentioned problems but ends up in a long 64 page document in an attempt to explain “everything” [12]. As a result, derivation is almost impossible to follow. Second one gives a quick explanation to the general philosophy of the algorithm but leaves the derivation entirely to the reader [13]. This paper does not aim to do any of these. It neither attempts to explain the philosophy of algorithm nor “everything” of the algorithm. It will present a quick, easy, and understandable proof of the most difficult part of the algorithm: its mathematical derivation. By following this derivation, readers will acquire at least four benefits. Firstly, they will understand the derivation with minimum background knowledge. Only linear algebra will be required. Secondly, they will grasp it in minimum amount of time. Derivation presented is as short as possible. Thirdly, they will learn it with minimum effort. Onus of making things understandable is on author. Finally, once the derivation is finished, they will be able to implement it in a programming language of their choice. The results will be presented in an easy to program manner. 2. System Model Consider following system of linear equations. is a symmetric × matrix such that this system iteratively. = = . and are (1) × 1 vectors. We attempt to solve Page 1 of 5 Published in Journal of Telematics and Informatics, Vol 4, No 1 (2016): March 2016 Abstract: http://section.iaesonline.com/index.php/JTI/article/view/188 = is the estimate computed in the -th iteration. iteration. is the step-size. 3. Computation of Residues Subtracting the actual solution + (2) is the correction term required in the -th from both sides of Eq. (1), − This results in, = − = + + is the error in the -th iteration. At the startup, we make a guess for is purely arbitrary, there will be a residue . = Or, − = Multiplying both sides of Eq. (3) with , Replacing the result of Eq. (4) in Eq. (5), − Or, − = =− (4) = + (5) =− + (6) = (7) − = for which, . Since our guess for =− 3. First Major Assumption Now we multiply both sides of Eq. (7) with We choose value for − − =0 . This is the first crucial decision we make for CG algorithm. So Eq. (8) becomes, Finally for , 0= − Hence, Eqs. (2) and (7) become, = = (8) (9) (10) (11) = And, (3) + − (12) (13) 4. Second Major Assumption So far we have Eqs. (7) and (11) for the computation of and the respectively. Now we want to derive the from and . Let us define the following relationship between all three. = + (14) Page 2 of 5 Published in Journal of Telematics and Informatics, Vol 4, No 1 (2016): March 2016 Abstract: http://section.iaesonline.com/index.php/JTI/article/view/188 Note that this relationship is quite similar to the pattern in Eq. (12). Now, we multiply both sides of Eq. (14) with . = + Further multiplying both sides of Eq. (15) with We want to choose , = for which, (15) + (16) =0 (17) This is the second crucial decision for the CG algorithm. So Eq. (16) becomes, Finally for 0= , + (18) =− 5. Making the Algorithm Recursive It will be much more efficient if we can compute from because we already have from Eq. (11). For this purpose, we will exploit term in Eq. (18). This can be done by recalling Eq. (7) and multiplying it with on both sides. Now we see = − term on the right hand side of Eq. (19). Being a scalar term, = Substituting our assumption from Eq. (9) in Eq. (19), = 0− Or, =− Using Eq. (20) and remembering that 1 =− Finally, in terms of (22) 1 Substituting the result of Eq. (23) in (18), = , = 1 (20) (21) is symmetric, =− (19) 1 (23) (24) (25) Now we have Eq. (25) to compute from . Only other term required to compute Eq. (25) is . But we already have and from Eq. (13). If we can compute from ’s only, it will not only make Eq. (25) more efficient but also recursive. For this purpose, only the dominator term in Eq. (17) has to be expressed in terms of ’s for this purpose. So we recall (14) and substitute in place of + 1 in it. Multiplying both sides of (26) first with , = + (26) Page 3 of 5 Published in Journal of Telematics and Informatics, Vol 4, No 1 (2016): March 2016 Abstract: http://section.iaesonline.com/index.php/JTI/article/view/188 And then multiplying with = , + = (27) + Employing the assumption of Eq. (17) and re-arranging, = Or, +0 = (31) = Or, (32) = Re-arranging Eq. (32), (29) (30) = Substituting Eq. (30) in Eq. (11), (28) (33) = Eq. (33) fully expresses the denominator of Eq. (25) in terms of ’s. Substituting Eq. (33) in Eq. (25), = 1 = Finally, (35) = Eq. (35) expresses (34) in terms of ’s only and this completes our derivation of CG algorithm. 6. The Algorithm Itself Now we recap the results and list them in the form of an algorithm. 1. Make an initial guess . 2. Take = = − . 3. Compute using Eq. (32). = 4. Update using Eq. (12). 5. Update using Eq. (7). 6. Determine using Eq. (35). = = + − = 7. Update using Eq. (14). = + Page 4 of 5 Published in Journal of Telematics and Informatics, Vol 4, No 1 (2016): March 2016 Abstract: http://section.iaesonline.com/index.php/JTI/article/view/188 7. Conclusion An easy, step-by-step but minimalist derivation of the Conjugate Gradient algorithm is provided in this paper. By following this derivation, the readers, regardless of their background, will not only master this fascinating algorithm in no time but will also be able to program it for their own purposes which will provide them a better understanding of their own problems on which they are working. References [1] C. Li, L. Fang, and X. Cao, "Global Convergence of A Kind of Conjugate Gradient Method," TELKOMNIKA Indonesian Journal of Electrical Engineering, vol. 11, pp. 544-549, 2013. [2] C. Li, "A Modified Conjugate Gradient Method for Unconstrained Optimization," TELKOMNIKA Indonesian Journal of Electrical Engineering, vol. 11, pp. 6373-6380, 2013. [3] L. Wang and R. C. de Lamare, "Set-membership constrained conjugate gradient adaptive algorithm for beamforming," Signal Processing, IET, vol. 6, pp. 789-797, 2012. [4] Y. Liu, R. Ranganathan, M. T. Hunter, and W. B. Mikhael, "Complex adaptive LMS algorithm employing the conjugate gradient principle for channel estimation and equalization," Circuits, Systems, and Signal Processing, vol. 31, pp. 1067-1087, 2012. [5] G. Harik, "Method for an optimizing predictive model using gradient descent and conjugate residuals," ed: Google Patents, 2014. [6] J. Ngiam, A. Coates, A. Lahiri, B. Prochnow, Q. V. Le, and A. Y. Ng, "On optimization methods for deep learning," in Proceedings of the 28th International Conference on Machine Learning (ICML-11), 2011, pp. 265-272. [7] W. Huang, Y. Wang, and D. W. Rosen, "Inverse Surfacelet Transform for Image Reconstruction With Constrained-Conjugate Gradient Methods," Journal of Computing and Information Science in Engineering, vol. 14, p. 021005, 2014. [8] S. K. Jain, R. K. Ray, and A. Bhavsar, "A Comparative Study of Iterative Solvers for Image De-noising," in Proceedings of the 3rd International Conference on Frontiers of Intelligent Computing: Theory and Applications (FICTA) 2014, 2015, pp. 307-314. [9] H. Wu, T. Guilin University of Electronic, G. Wang, T. Guilin University of Electronic, F. Zhang, T. Guilin University of Electronic, et al., "Inversion of Surface Nuclear Magnetic Resonance by Regularization with Simulated Atomic Transition Method," TELKOMNIKA Indonesian Journal of Electrical Engineering, vol. 11, pp. 5055-5060, 2013. [10] Z. Chen, H. Li, and M. Rangaswamy, "Conjugate gradient adaptive matched filter," Aerospace and Electronic Systems, IEEE Transactions on, vol. 51, pp. 178-191, 2015. [11] S. Zhao, D. L. Jones, S. Khoo, and Z. Man, "Frequency-domain beamformers using conjugate gradient techniques for speech enhancement," The Journal of the Acoustical Society of America, vol. 136, pp. 1160-1175, 2014. [12] J. R. Shewchuk, "An introduction to the conjugate gradient method without the agonizing pain," ed: Carnegie-Mellon University. Department of Computer Science, 1994. [13] G. Strang, Computational science and engineering vol. 1: Wellesley-Cambridge Press Wellesley, 2007. Page 5 of 5
8cs.DS
arXiv:1511.02534v2 [math.ST] 31 Mar 2017 Order Determination of Large Dimensional Dynamic Factor Model Z. D. Baia , Chen Wangb , Ya Xuec and Matthew Hardingd Northeast Normal Universitya , University of Cambridgeb and University of California, Irvinec,d April 3, 2017 Abstract Consider the following dynamic factor model: Rt = Pq i=0 Λi ft−i + et , t = 1, ..., T , where Λi is an n × k loading matrix of full rank, {ft } are i.i.d. k × 1-factors, and et are independent n × 1 white noises. Now, assuming that n/T → c > 0, we want to estimate the orders k and q respectively. Define a random matrix Φn (τ ) = T 1 X (Rj R∗j+τ + Rj+τ R∗j ), 2T j=1 where τ ≥ 0 is an integer. When there are no factors, the matrix Φn (τ ) reduces to Mn (τ ) = T 1 X (ej e∗j+τ + ej+τ e∗j ). 2T j=1 When τ = 0, Mn (τ ) reduces to the usual sample covariance matrix whose ESD tends to the well known MP law and Φn (0) reduces to the standard spike model. Hence the number k(q + 1) can be estimated by the number of spiked eigenvalues of Φn (0). To obtain separate estimates of k and q , we have employed the spectral analysis of Mn (τ ) and established the spiked model analysis for Φn (τ ). 1 1 Introduction For a p × p random Hermitian matrix A with eigenvalues λj , j = 1, 2, · · · , p, the empirical spectral distribution (ESD) of A is defined as p 1X F (x) = I(λj ≤ x). p j=1 A The limiting distribution F of {F An } for a given sequence of random matrices {An } is called the limiting spectral distribution (LSD). Let {εit } be independent identically distributed (i.i.d) random variables with common mean 0, variance 1. Consider a high dimensional dynamic kP factor model with lag q, that is, Rt = qi=0 Λi ft−i + et , t = 1, ..., T , where Λi is an n×k loading matrix of full rank, {ft } are i.i.d. k × 1-factors with common mean 0, variance 1, whereas et corresponds to the noise component with et = (ε1t , · · · , εnt )′ . In addition, both components of et and ft are assumed to have finite 4th moment. This model can also be thought as an information-plus-noise type model (Dozier & Silverstein, 2007a, b; Bai & Silverstein, 2012). Here both n and T tend to ∞, with n/T → c for some c > 0. Compared with n and T , the number of factors k and that of lags q are fixed but unknown. An interesting and important problem to economists is how to estimate k and PT 1 ∗ ∗ √1 et and Mn (τ ) = q. To this end, define Φn (τ ) = 2T j=1 (Rj Rj+τ + Rj+τ Rj ), γt = 2T PT ∗ ∗ ∗ k=1 (γk γk+τ + γk+τ γk ), τ = 0, 1, · · · . Here stands for the transpose and complex conjugate of a complex number and τ is referred to be the number of lags. Denote Λ = (Λ0 , Λ1 , · · · , Λq )n×k(q+1) ,  fT +τ −1 · · · fτ +1  fT +τ   fT +τ −1 fT +τ −2 · · · fτ  τ F =  .. .. .. ..  . . . .   fT +τ −q fT +τ −1−q · · · fτ +1−q eτ = (eT +τ , eT +τ −1 , · · · , eτ +1 )n×T . Then we have that Φn (τ ) = 1 2T          , k(q+1)×T [(ΛFτ + eτ )(ΛF0 + e0 )∗ + (ΛF0 + e0 )(ΛFτ + eτ )∗ ] and Mn (τ ) = 2 1 2T (eτ e0∗ + e0 eτ ∗ ). Note that essentially, Mn (τ ) and Φn (τ ) are symmetrized auto-cross covariance matrices at lag τ and generalize the standard sample covariance matrices Mn (0) and Φn (0), respectively. The matrix Mn (0) has been intensively studied in the literature and it is well known that the LSD has an MP law (Marčenko and Pastur, 1967). Readers may refer to Jin et al. (2014) and Wang et al. (2015) for more details about the model. To estimate k and q, the following method can be employed. First, note that when τ = 0 and Cov(ft ) = Σf , the population covariance matrix of Rt is a spiked population model (Johnstone (2001), Baik and Silverstein (2006), Bai and Yao (2008)) with k(q+1) spikes. Therefore, k(q+1) can be estimated by counting the number of eigenvalues of Φn (0) that are larger than some phase transition point. Next, the separated estimation of k and q can be achieved by investigating the spectral property of Mn (τ ) for general τ ≥ 1, using the fact that the number of eigenvalues of Φn (τ ) that lie outside the support of the LSD of Mn (τ ) at lags 1 ≤ τ ≤ q is different from that at lags τ > q. Thus, the estimates of k and q can be separated by counting the number of eigenvalues of Φn (τ ) that lie outside the support of the LSD of Mn (τ ) from τ = 0, 1, 2, · · · , q, q + 1, · · · . Note that for the above method to work, the LSD of Mn (τ ) for general τ ≥ 1 must be known. This is derived in Jin et al. (2014). Moreover, it is required that no eigenvalues outside the the support of the LSD of Mn (τ ) so that if an eigenvalue of Φn (τ ) goes out of the support of the LSD of Mn (τ ), it must come from the signal part. Wang et al. (2015) proved such phenomenon theoretically. Both results are included in Section 2 for readers’ reference. The rest of the paper is structured as follows: Some known results are given in Section 2. Section 3 presents truncation of variables and Section 4 estimates k(q + 1). The estimation of q is provided in Section 5, from the which the estimation of k can also be obtained. Section 6 discusses the case when the variance of the noise part is unknown. A simulation study is shown in Section 7 and some proofs are presented in Appendix. Regrading the norm used in this paper, the norm applied to a vector is the usual Euclidean norm, with notation k ∗ k. For a matrix, two kinds of norm have been used. The operator norm, 3 denoted by k ∗ ko, is the largest singular value. For matrices of fixed dimension, the Kolmogorov norm, defined as the largest absolute value of all the entries, has been used, with notation k ∗ kK . 2 Some known results In this section, we present some known results. Lemma 2.1 (Burkholder (1973)). Let {Xk } be a complex martingale difference sequence with respect to the increasing σ-fields {Fn }. Then, for p ≥ 2, we have E| X   X X p/2 X k | p ≤ Kp E E(|Xk |2 |Fk−1) +E |Xk |p . Lemma 2.2 (Lemma A.1 of Bai and Silverstein (1998)). For X = (X1 , · · · , Xn )′ i.i.d. standardized (complex) entries, B n × n Hermitian nonnegative definite matrix, we have, for any p ≥ 1, E|X∗BX|p ≤ Kp  trB where Kp is a constant depending on p only. p  + rE|X1 |2p trBp , Lemma 2.3 (Jin et al. (2014)). Assume: (a) τ ≥ 1 is a fixed integer. (b) ek = (ε1k , · · · , εnk )′ , k = 1, 2, ..., T +τ , are n-dimensional vectors of independent standard complex components with sup1≤i≤n,1≤t≤T +τ E|εit |2+δ ≤ M < ∞ for some δ ∈ (0, 2], and for any η > 0, 1 η 2+δ nT n T +τ X X i=1 t=1 E(|εit |2+δ I(|εit | ≥ ηT 1/(2+δ) )) = o(1). (c) n/(T + τ ) → c > 0 as n, T → ∞. (d) Mn (τ ) = PT ∗ k=1 (γk γk+τ + γk+τ γk∗ ), where γk = 4 √1 2T ek . (2.1) D Then as n, T → ∞, F Mn (τ ) → Fc a.s. and Fc has a density function given by q 2 y 1 1 φc (x) = 2cπ 1+y0 0 − ( 1−c + √1+y )2 , |x| ≤ a, |x| 0 where a=   √ (1−c) 1+y1 , y1 −1 c 6= 1, 2, c = 1,  y0 is the largest real root of the equation: y 3 − (1−c)2 −x2 2 y x2 − 4 y x2 − 4 x2 = 0 and y1 is the only real root of the equation: ((1 − c)2 − 1)y 3 + y 2 + y − 1 = 0 (2.2) such that y1 > 1 if c < 1 and y1 ∈ (0, 1) if c > 1. Further, if c > 1, then Fc has a point mass 1 − 1/c at the origin. Note that as long as τ ≥ 1, Fc does not depend on τ. Lemma 2.4 (Bai and Wang (2015)). Theorem 2.3 still holds with the 2 + δ moment condition weakened to 2nd moment. Lemma 2.5 (Wang et al. (2015)). Assume: (a) τ ≥ 1 is a fixed integer. (b) ek = (ε1k , · · · , εnk )′ , k = 1, 2, ..., T + τ , are n-vectors of independent standard complex components with supi,t E|εit |4 ≤ M for some M > 0. (c) There exist K > 0 and a random variable X with finite fourth order moment such that, for any x > 0, for all n, T n T +τ 1 XX P(|εit | > x) ≤ KP(|X| > x). nT i=1 t=1 (d) cn ≡ n/T → c > 0 as n → ∞. (e) Mn = PT ∗ k=1 (γk γk+τ + γk+τ γk∗ ), where γk = √1 2T ek . (f) The interval [a,b] lies outside the support of Fc , where Fc is defined as in Lemma 2.3. Then P (no eigenvalues of Mn appear in [a, b] for all large n) = 1. 5 (2.3) 3 Truncation, centralization and standardization of variables As proved in Wang et al.(2015), we may assume that the εij ’s satisfy the conditions that |εij | ≤ C, Eεij = 0, E|εij |2 = 1, E|εij |4 < M (3.4) for some C, M > 0. For the truncation of variables in Fτ , first note that for a random variable X with E|X|4 < ∞,  P 2ℓ P |X| > 2ℓ/4 < ∞. Given E|Fτij |4 < ∞, i = 1, · · · , k(q + 1), j = 1, · · · , T , we have ∞ ℓ=1   Fτ , |Fτ | < T 1/4 ,  ij ij define F̂τij = F̂τ = F̂τij and  0, otherwise, Φ̂n (τ ) = 1 [(ΛF̂τ + eτ )(ΛF̂0 + e0 )∗ + (ΛF̂0 + e0 )(ΛF̂τ + eτ )∗ ]. 2T Then we have   P Φn (τ ) 6= Φ̂n (τ ), i.o.   = P Fτ 6= F̂τ , i.o.  ∞ [ ∞ \ = P  [ L=1 T =L i≤k(q+1) j≤T  2ℓ+1 ∞ X  [ ≤ lim P  L→∞ ℓ=L ≤   {|Fτij | ≥ T 1/4 }  [ T =2ℓ +1 i≤k(q+1) j≤2ℓ+1   {|Fτij | ≥ 2ℓ/4 }    ∞ X   [ τ ℓ/4  {|F | ≥ 2 } P ij   L→∞ lim ℓ=L ≤ k(q + 1) lim i≤k(q+1) j≤2ℓ+1 ∞ X ℓ+1 L→∞ → 0. 2 ℓ=L P |Fτ11 | ≥ 2ℓ/4  This completes the proof of truncation. Centralization and standardization can be justified in the 6 same way as in Appendix A of Wang et al. (2015). In what follows, we may assume that |Fτij | < T 1/4 , EFτij = 0, E|Fτij |2 = 1, E|Fτij |4 < M for some M > 0. 4 Estimation of k(q + 1) In this section, we will estimate k(q+1) by an investigation of the limiting properties of eigenvalues of Φn (0). For simplicity, rewrite Φn (0) = Φ(0), F0 = F and e0 = e. With these notations, we have Φ(0) = 1 (ΛF T + e)(ΛF + e)∗ and M(0) = 1 ee∗ . T When Λ = 0, Φ(0) reduces to M(0), which is a standard sample covariance matrix and thus its ESD tends to the famous MP law (Marčenko and Pastur, 1967). Suppose ℓ is an eigenvalue of Φ(0), then we have 0 = det |ℓI − Φ(0)| = det ℓI − M(0) − 1 1 1 ΛFe∗ − eF∗ Λ∗ − ΛFF∗ Λ∗ . T T T (4.1) Let B = (B1 : B2 ) be an n × n orthogonal matrix such that B1 = Λ(Λ∗ Λ)−1/2 and thus Λ∗ B2 = 0k(q+1)×(n−k(q+1)) . Then (4.1) is equivalent to det ℓIk(q+1) − T1 B∗1 (ΛF + e)(F∗ Λ∗ + e∗ )B1 − T1 B∗1 (ΛF + e)e∗ B2 − T1 B∗2 e(F∗ Λ∗ + e∗ )B1 ℓIn−k(q+1) − T1 B∗2 ee∗ B2 =0 (4.2) If we further assume that ℓ is not an eigenvalue of T1 B∗2 ee∗ B2 , then we have det |Ik(q+1) − 1 ∗ B (ΛF + e)D−1 (ℓ)(F∗ Λ∗ + e∗ )B1 | = 0, T 1 where D(ℓ) = ℓIT − T1 e∗ B2 B∗2 e. Denote H(ℓ) = ℓIT − T1 e∗ e, then we obtain  −1 1 1 ∗ −1 1 ∗ −1 ∗ ∗ B1 eD (ℓ)e B1 = I + B1 eH (ℓ)e B1 B∗1 eH−1 (ℓ)e∗ B1 . T T T (4.3) (4.4) Next, we have  1 ∗ 1 ∗ −1 B1 eH (ℓ)e∗ B1 = B1 − T I + ℓT (ℓIn − M(0))−1 B1 T T = −Ik(q+1) + ℓB∗1 (ℓIn − M(0))−1 B1 . 7 (4.5) Substitute (4.5) back to (4.4), and we have 1 ∗ −1 B eD (ℓ)e∗ B1 T 1 −1  −1 1 ∗ B1 (M(0) − ℓIn )−1 B1 = Ik(q+1) − ℓB∗1 (ℓIn − M(0))−1 B1 = Ik(q+1) + ℓ Write B1 = (b1 , · · · , bk(q+1) ), then we have kbi k = 1. By Lemma 6 in Bai, Liu and Wong (2011), we have b∗i (M(0) − ℓIn )−1 bi → m, a.s. b∗i (M(0) − ℓIn )−1 bj → 0, a.s., and for i 6= j where m = m(ℓ) = limn→∞ n1 tr(M(0) − ℓIn )−1 is the Stieltjes transform of the sample covariance with ratio index c = limn→∞ Tn . By Lemma 3.11 of Bai and Silverstein (2010), we have m satisfying p 1 − c − ℓ + (1 − ℓ − c)2 − 4ℓc m(ℓ) = . 2cℓ Therefore, we obtain 1  1 ∗ −1 Ik(q+1) . B1 eD (ℓ)e∗ B1 → 1 + T ℓm Next, we want to show, with probability 1 that 1 ∗ B ΛFD−1 (ℓ)e∗ B1 → 0 T 1 and 1 ∗ −1 B eD (ℓ)F∗ Λ∗ B1 → 0. T 1 Note that 1 ∗ B ΛFD−1 (ℓ)e∗ B1 T 1 −1 1 1 I + B∗1 eH−1 (ℓ)e∗ B1 B∗1 ΛFH−1 (ℓ)e∗ B1 = T T −1 1  ∗ B1 (M(0) − ℓIN )−1 B1 B∗1 ΛFH−1 (ℓ)e∗ B1 = ℓT −1 −1 1  ∗ 1 B1 (M(0) − ℓIN )−1 B1 = B∗1 ΛF ℓIT − e∗ e e∗ B1 . ℓT T 8 (4.6) Recall M(0) = 1 ee∗ . T Fix δ > 0 and let event A = {λmax (M(0)) ≤ (1 + √ c)2 + δ} and Ac be the complement. By Theorem 5.9 of Bai and Silverstein (2010), we have P(Ac ) = o(n−t ) for any t > 0. Suppose ℓ is an eigenvalue of Φ(0) larger than (1 + √ c)2 + 2δ. By the fact that M(0) 1 ∗ e e T have the same set of nonzero eigenvalues, we have, under A, that kℓIT − T1 e∗ eko ≥ −1 ℓ − k T1 e∗ eko ≥ δ > 0, and hence k ℓIT − T1 e∗ e ko ≤ 1δ . and Therefore, for any ε > 0, we have −1 1 1 P(k F ℓIT − e∗ e e∗ B1 kK ≥ ε) T T   −1 1 1 = E P(k F ℓIT − e∗ e e∗ B1 kK ≥ ε) e T T   −1 1 1 ≤ E P(k F ℓIT − e∗ e e∗ B1 kK ≥ ε) e, A + P(Ac ). T T Write F = (F̃1 , · · · , F̃k(q+1) )′ . For the first term, by Lemma 2.2, we have   1 1 ∗ −1 ∗ E P(k F ℓIT − e e e B1 kK ≥ ε) e, A T T   1 1 ∗ −1 ∗ 4r ≤ 4r 4r EE kF ℓIT − e e e B1 kK e, A ε T T h i 1 ∗ −1 ∗ 1 ∗ −1 ∗ 2r 1 ∗ e, A ≤ 4r 4r EE trF ℓIT − e e e B1 B1 e ℓIT − e e F ε T T T h k(q+1) i X −1 −1 2r 1 1 1 e, A F̃∗i ℓIT − e∗ e e∗ B1 B∗1 e ℓIT − e∗ e F̃i = 4r 4r EE ε T T T i=1 k(q+1) i [k(q + 1)]2r−1 h X  ∗ 1 ∗ −1 2r 1 ∗ −1 ∗ ∗ ≤ E e e e e F̃ e B B e ℓI − E e, A F̃ ℓI − i 1 T T 1 i ε4r T 4r T T i=1 k(q+1) K2r [k(q + 1)]2r−1 h X  1 ∗ −1 2r 1 ∗ −1 ∗ ∗ E e e e B B e ℓI − ee + tr ℓI − 1 T T 1 ε4r T 4r T T i=1 −1 −1  i 1 1 E|F11 |4r tr[(ℓIT − e∗ e e∗ B1 B∗1 e ℓIT − e∗ e ]2r A T T i 2r h  K2r [k(q + 1)] 2r ∗ ∗ 4r ∗ ∗ 2r ≤ E tre B1 B1 e + E|F11 | tr(e B1 B1 e) A δ 4r ε4r T 4r √  K2r [k(q + 1)]2r [(1 + c)2 + δ]2r  ∗ 2r 4r ∗ 2r E (trB B ) + E|F | tr(B B ) ≤ 11 1 1 1 1 δ 4r ε4r T 2r √ 2r 2 2r  K2r [k(q + 1)] [(1 + c) + δ] 2r 4 4r−4 4 k(q + 1) ≤ [k(q + 1)] + E|F | T 11 δ 4r ε4r T 2r ≤ 9 which is summable for r ≥ 1. Hence, we have shown with probability 1 that 1 ∗ B1 ΛFD−1 (ℓ)e∗ B1 → 0. T Similarly, we have with probability 1 that 1 ∗ −1 B1 eD (ℓ)F∗ Λ∗ B1 → 0. T Therefore, substituting into (4.3), we have 1 1 det | B∗1 ΛFD−1 (ℓ)F∗ Λ∗ B1 + Ik(q+1) | → 0. T ℓm(ℓ) (4.7) Using Bai, Liu and Wong (2011) again, we have the diagonal elements of the matrix T −1 FD−1 (ℓ)F∗ tend to −m(ℓ) and the off diagonal elements tend to 0. Here m(ℓ) is the Stieltjies transform of the LSD of T1 e∗ e and satisfies m(ℓ) = − 1−c + cm(ℓ). ℓ Thus, if Λ∗ Λ → Q, then (4.7) can be further simplified as det | − Qm(ℓ) + 1 Ik(q+1) | = 0. ℓm(ℓ) (4.8) If α is an eigenvalue of Q, and there is an ℓ belonging to the complement of the support of the LSD of M(0) such that α = 1 , ℓm(ℓ)m(ℓ) then ℓ is a solution of (4.8). From (4.6), we have cℓm2 (ℓ) − (1 − c − ℓ)m(ℓ) + 1 = 0, which implies  1−c + cm(ℓ) ℓ = −(1 − c)m(ℓ) + cℓm2 (ℓ) ℓm(ℓ)m(ℓ) = ℓm(ℓ) − = −(1 − c)m(ℓ) + (1 − c − ℓ)m(ℓ) − 1 = −ℓm(ℓ) − 1 p 1 − c − ℓ + (1 − ℓ − c)2 − 4ℓc = − −1 2c =: g(ℓ). 10 It is easy to verify that g ′ (ℓ) < 0, implying that ℓm(ℓ)m(ℓ) is decreasing. Also note that √ √ 1 > c, then we have ℓ > ℓm(ℓ)m(ℓ) = √1c when ℓ = (1 + c)2 . Therefore, if α = ℓm(ℓ)m(ℓ) √ √ √ (1+ c)2 . This recovers the result of Baik and Silverstein (2006). Note that [(1− c)2 , (1+ c)2 ] √ is the support of the MP law. Hence, if all the eigenvalues of Q are greater than c, we have k(q + 1) sample eigenvalues of Φn (0) goes outside the right boundary of the support of the MP law. Note that although the distribution of rest n − k(q + 1) sample eigenvalues follows the MP law with the largest sample eigenvalue converging to the right boundary, there is still a positive probability that the largest sample eigenvalue goes beyond the right boundary. Therefore, to completely separate the k(q + 1) spiked sample eigenvalues from the rest, the threshold is set √ as (1 + c)2 (1 + 2n−2/3 ). In other words, k(q + 1) can be estimated by the number of sample √ eigenvalues of Φn (0) greater than (1 + c)2 (1 + 2n−2/3 ). Remark 4.1 For factor models, the loading matrix is unknown. This, however, is not a concern in our estimation because compared with the noise matrix, the loading matrix is denominating, √ making the condition easily satisfied that all the eigenvalues of Q are greater than c. Remark 4.2 The rationale of choosing (1+2n−2/3 ) as the buffering factor of the criterion is that, √ according to Tracy-Widom law, the quantity of a non-spiked eigenvalue larger than (1 + c)2 has an order of n−2/3 . Therefore, it is good enough for us to choose (1 + 2n−2/3 ) to completely separate the spikes and the bulk eigenvalues. 5 Estimation of q Next, we want to split k and q. Let τ ≥ 1 be given and assume that ℓ is an eigenvalue of Φn (τ ). For simplicity, write Mn (τ ) = M and for t = 1, 2, · · · , T , define Ft = (ft , ft−1 , · · · , ft−q )′ such 11 that Rt = ΛFt + et . Then we have 0 = det |ℓI − Φn (τ )| T 1 X ΛFj F∗j+τ Λ∗ + ΛF∗j+τ F∗j Λ∗ = det ℓI − M − 2T j=1  +ej F∗j+τ Λ∗ + ΛF∗j+τ e∗j + e∗j+τ F∗j Λ∗ + ΛFj e∗j+τ . (5.1) Define B, B1 and B2 the same as in the last section. Multiplying B∗ from left and B from right to the above matrix and by Λ∗ B2 = 0k(q+1)×(n−k(q+1)) , we have (5.1) equivalent to 0 = det ℓIk(q+1) − S11 −S12 −S21 ℓIn−k(q+1) − S22 = det |ℓI − S22 | det |ℓI − Kn (ℓ)|, where S11 T 1 X ∗ B [(ΛFj + ej )(ΛFj+τ + ej+τ )∗ + (ΛFj+τ + ej+τ )(ΛFj + ej )∗ ]B1 = 2T j=1 1 S12 = T 1 X ∗ B1 (ΛFj e∗j+τ + ΛFj+τ e∗j )B2 + B∗1 MB2 2T j=1 S21 = S∗12 S22 = B∗2 MB2 Kn (ℓ) = S11 + S12 (ℓIn−k(q+1) − S22 )−1 S21 . Therefore, if ℓ is not an eigenvalue of S22 , by the factorization above, ℓ must be an eigenvalue of Kn (ℓ), i.e. det |Kn (ℓ) − ℓI| = 0. PT 1 ∗ ∗ Denote W = 2T j=1 (Fj ej+τ + Fj+τ ej ). By the assumptions of et ’s and Ft ’s, the random vector {Fj e∗j+τ + Fj+τ e∗j , j ≥ 1} is (q + 1)-dependent (see Page 224, Chung 2001). It then follows with probability 1 that WB1 = o(1) B∗1 W∗ = o(1) T 1 X (Fj F∗j+τ + Fj+τ F∗j ) = H(τ ) + o(1), 2T j=1 12 where  0 · · · 1 · · · . .  .. . . 0 1   .. H(τ ) =  . 0 1 0 . .. . . . 1 0  0 ··· 1 ···  0 ..  .   1 , ..   .  0 is of dimension k(q + 1) × k(q + 1) with two bands of 1’s of kτ -distance from the main diagonal. Therefore, we have a.s. S11 = B∗1 ΛH(τ )Λ∗ B1 + B∗1 MB1 + o(1) S12 = B∗1 (M + ΛW)B2 S21 = B∗2 (M + W∗ Λ∗ )B1 . Subsequently, we have a.s.  −1 Kn (ℓ) = B∗1 ΛH(τ )Λ∗ B1 + B∗1 MB1 + B∗1 (M + ΛW)B2 ℓI − B∗2 MB2 B∗2 (M + W∗ Λ∗ )B1 + o(1). Note that  −1 B∗1 MB1 + B∗1 MB2 ℓI − B∗2 MB2 B∗2 MB1 1 1 = B∗1 MB1 + B∗1 M B2 B∗2 M(I − B2 B∗2 M)−1 B1 ℓ ℓ 1 = B∗1 M(I − B2 B∗2 M)−1 B1 ℓ ∗ = ℓB1 M(ℓI − M + B1 B∗1 M)−1 B1  −1 −1 = ℓI − ℓ I + B∗1 M ℓI − M B1  −1 −1 = ℓI − B∗1 ℓI − M B1 13 and  −1 WB2 ℓI − B∗2 MB2 B∗2 W∗ = = T  −1 1 X ∗ ∗ ∗ (Fi γi+τ + Fi+τ γi )B2 ℓI − B2 MB2 B∗2 (γj+τ F∗j + γj F∗j+τ ) 2T i,j=1 T  −1  −1 i 1 Xh ∗ Fj γj+τ B2 ℓI − B∗2 MB2 B∗2 γj+τ F∗j + Fj+τ γj∗ B2 ℓI − B∗2 MB2 B∗2 γj F∗j+τ + 2T j=1 T  −1  −1 i 1 Xh ∗ Fj γj+τ B2 ℓI − B∗2 MB2 B∗2 γj F∗j+τ + Fj+τ γj∗ B2 ℓI − B∗2 MB2 B∗2 γj+τ F∗j + 2T j=1 T  −1 1 X ∗ ∗ (Fj+τ γj+2τ + Fj+2τ γj+τ )B2 ℓI − B∗2 MB2 B∗2 (γj+τ F∗j + γj F∗j+τ ) + 2T j=1 T  −1 1 X ∗ (Fj−τ γj∗ + Fj γj−τ )B2 ℓI − B∗2 MB2 B∗2 (γj+τ F∗j + γj F∗j+τ ) + 2T j=1 1 2T T X i,j=1 i6=j,i6=j±τ  −1 ∗ (Fi γi+τ + Fi+τ γi∗ )B2 ℓI − B∗2 MB2 B∗2 (γj+τ F∗j + γj F∗j+τ ) =: P1 + P2 + P3 + P4 + P5 . Next, we give a lemma on the quadratic form of γj . Lemma 5.1 Let i, j ∈ N be given, we have almost surely and uniformly in i and j that   cm p cm   −c22m2 −c22m2  −1 ≡ Cp , i = j ± pτ 1− 2x 1− 4x 1 1 γi∗ B2 ℓI − B∗2 MB2 B∗2 γj →   0, otherwise. The proof of the lemma is postponed in the Appendix. 14 First, we have E(P1 ) = E[E(P1 |γ1 , · · · , γT +τ )] T  −1 1 X nh ∗ E Fj γj+τ B2 ℓI − B∗2 MB2 = B∗2 γj+τ F∗j + 2T j=1  −1 i o Fj+τ γj∗ B2 ℓI − B∗2 MB2 B∗2 γj F∗j+τ γ1 , · · · , γT +τ T h −1 X i 1 ∗ ∗ ∗ ∗ Etr ℓI − B2 MB2 B2 (γj+τ γj+τ + γj γj )B2 Ik(q+1) = 2T j=1 T −1 1X ∗  Eγj B2 ℓI − B∗2 MB2 B∗2 γj Ik(q+1) = T j=1 = C0 Ik(q+1) + o(1). Similarly, we have 1 C1 H(τ ) + o(1), 2 1 E(P3 ) = C1 HL (τ ) + (C0 HL (2τ ) + C2 Ik(q+1) ) + o(1), 2 1 E(P4 ) = C1 HU (τ ) + (C0 HU (2τ ) + C2 Ik(q+1) ) + o(1). 2 E(P2 ) = Here HL (τ ) and HU (τ ) denote the lower and upper part of H(τ ) with the rest entries being 0. Furthermore, denote HL (0) ≡ HU (0) ≡ Ik(q+1) and hence H(0) = HL (0) + HU (0) = 2Ik(q+1) .   Consider i = j ± pτ for p = 0, 1, · · · , τq . When p = 0, we have E(P1) + E(P2) = 21 C0 H(0) + 12 C1 H(τ ) + o(1). When p = 1, we have E(P3) + E(P4) = 21 C2 H(0) + C1 H(τ ) + 12 C0 H(2τ ) + o(1). When p = 2, we have part of E(P5) is 21 C3 H(τ ) + C2 H(2τ ) + 12 C1 H(3τ ) + o(1). .. . 15 When p = q τ , we have part of E(P5) is q  q  q  1   1 H( τ − τ ) + C q  H( τ ) + C q  H( τ + τ ) + o(1) C q 2 τ +1 τ τ 2 τ −1 τ τ q  q  1   H( = τ − τ ) + C q  H( τ ) + o(1). C q +1 2 τ τ τ τ   When p = τq + 1, we have part of E(P5) is q  q 1   q  1 τ + 2τ ) + C q  H( τ + τ ) + C q  H( τ ) + o(1) C q H( +1 2 τ τ τ 2 τ +2 τ τ q  1   H( = τ ) + o(1). C q 2 τ +2 τ Next, we want to show that Pi → E(Pi ) a.s. Since all the Pi ’s are of finite dimension, it suffices to show the a.s convergence entry-wise. Denote the (u, v)-entry of a matrix A by A(u,v) . For  −1 i = 1, define αj = γj∗ B2 ℓI − B∗2 MB2 B∗2 γj . Then for any positive integer s, applying Lemma 2.1, we have E(|P1(i1,i2) − = E ≤ 2 αj+τ αj δ(i1 ,i2 ) − δ(i1 ,i2 ) |2s ) 2 2 T h n 1 X i α o2s αj+τ j αj+τ (Fj F∗j )(i1 ,i2 ) + αj (Fj+τ F∗j+τ )(i1 ,i2 ) − δ(i1 ,i2 ) − δ(i1 ,i2 ) 2T j=1 2 2 2s−1 T h 1 X i2s αj+τ E αj+τ (Fj F∗j )(i1 ,i2 ) − δ(i1 ,i2 ) + 2T j=1 2 22s−1 E T h 1 X i2s αj αj (Fj+τ F∗j+τ )(i1 ,i2 ) − δ(i1 ,i2 ) 2T j=1 2 T i2s 1 h1 X αj+τ (Fj F∗j )(i1 ,i2 ) − αj+τ δ(i1 ,i2 ) + E = 2 T j=1 T i2s 1 h1 X αj (Fj+τ F∗j+τ )(i1 ,i2 ) − αj δ(i1 ,i2 ) E 2 T j=1 T h io2s 1 n 1 X ∗ h (i1 ,i2 ) i (i1 ,i2 ) = Fj αj+τ Fj − tr αj+τ + E 2 T j=1 T h io2s 1 n 1 X ∗ h (i1 ,i2 ) i (i1 ,i2 ) F α Fj+τ − tr αj E . 2 T j=1 j+τ j 16 (5.2)   Here a(u,v) denotes the matrix with the (u, v)-entry being a and 0 elsewhere. By the truncation of εij and the fact that k(ℓI − B∗2 MB2 )−1 ko ≤ η −1 with η = ℓ − dc > 0, both |αj | and |αj+τ | are bounded from above, say, by C. Also notice that |Fτij | < T 1/4 and E|Fτij |4 < M. Similar to the proof of Lemma 9.1 in Bai and Silverstein (2010), we have T s h i h io2s C 2s X n1 X  (i1 ,i2 ) (i1 ,i2 ) F∗j αj+τ Fj − tr αj+τ M l l2s + l4s E ≤ s T j=1 T l=1 n1 E T T X j=1 F∗j+τ s h i h io2s C 2s X  (i1 ,i2 ) (i1 ,i2 ) M l l2s + l4s . αj Fj+τ − tr αj ≤ s T l=1 Substituting the above back to (5.2) and choosing s ≥ 2, we have P1(i1 ,i2 ) − αj αj+τ δ(i1 ,i2 ) − δ(i1 ,i2 ) = oa.s (1). 2 2 Again, by the almost sure and uniform convergence of αj and αj+τ to C0 , we have αj+τ αj δ(i1 ,i2 ) + δ(i1 ,i2 ) − C0 δ(i1 ,i2 ) = oa.s (1). 2 2 Therefore, we have shown that P1 − E(P1 ) = oa.s. (1). Results for i = 2, 3, 4, 5 can be shown in a similar way. Denote α = − cm 2 2 2 1− c2xm 1 and β = − cm 2 2 2 1− c4xm , then we have Cp = αβ p . Note that H(pτ ) = 0 for 1 17 p > [q/τ ], and we have, with probability 1 that  −1 WB2 ℓI − B∗2 MB2 B∗2 W∗ ∞ X 1 1 3 1 1 1 ( Cp−2 + Cp + Cp+2)H(pτ ) → ( C0 + C2 )H(0) + ( C1 + C3 )H(τ ) + 2 2 2 2 2 2 p=2 = = = ∞ i X αh β p−2 H(pτ ) (1 + β 2 )H(0) + (3β + β 3 )H(τ ) + (1 + β 2 )2 2 p=2 ∞ X i αh (1 + β 2 )H(0) + (3β + β 3 )H(τ ) + (1 + β 2 )2 β p−2 HL (pτ ) + HU (pτ ) 2 p=2 ∞ i X  αh (1 + β 2 )H(0) + (3β + β 3 )H(τ ) + (1 + β 2 )2 β p−2 JL (pτ ) + JU (pτ ) ⊗ Ik 2 p=2 ∞ i X  αh (1 + β 2 )H(0) + (3β + β 3 )H(τ ) + (1 + β 2 )2 β p−2 JpL (τ ) + JpU (τ ) ⊗ Ik 2 p=2  αh (1 + β 2 )H(0) + (3β + β 3 )H(τ ) + (1 + β 2 )2 J2L (τ )(I − βJL (τ ))−1 = 2  i +J2U (τ )(I − βJU (τ ))−1 ⊗ Ik . = Note that 1 [I − (I − βJL (τ ))](I − βJL (τ ))−1 β 1 = [(I − βJL (τ ))−1 − I] β 1 JL (τ )[(I − βJL (τ ))−1 − I] = β 1 1 = [(I − βJL (τ ))−1 − I] − JL (τ ). 2 β β JL (τ )(I − βJL (τ ))−1 = J2L (τ )(I − βJL (τ ))−1 Similarly, 1 JU (τ )[(I − βJU (τ ))−1 − I] β 1 1 [(I − βJU (τ ))−1 − I] − JU (τ ). = 2 β β J2U (τ )(I − βJU (τ ))−1 = 18 Therefore, we have   (1 + β 2 )2 J2L (τ )(I − βJL (τ ))−1 + J2U (τ )(I − βJU (τ ))−1 ⊗ Ik 1 1 = (1 + β 2 )2 2 (I − βJL (τ ))−1 + 2 (I − βJU (τ ))−1 β β   1 2 − 2 I − JL (τ ) + JU (τ ) ⊗ Ik β β   i 2 2h (1 + β ) −1 −1 = (I − βJ (τ )) 2I − βJ (τ ) − βJ (τ ) (I − βJ (τ )) ⊗ Ik L L U U β2 (1 + β 2 )2 (1 + β 2 )2 H(0) − H(τ ) − β2 β (1 + β 2 )2 (1 + β 2 )2 (1 + β 2 )2 =: G(τ ) − H(0) − H(τ ). β2 β2 β Hence, we have a.s.  −1 ∗ ℓI − B2 MB2 B∗2 W∗ Λ∗ B1  (1 + β 2 )2  (1 + β 2 )2  α 1/2 h 3 H(τ ) 1 + β2 − H(0) + 3β + β − Q → 2 β2 β i (1 + β 2 )2 + G(τ ) Q1/2 . β2 B∗1 ΛWB2 Last, we want to show that with probability 1,  −1 B∗1 ΛWB2 ℓI − B∗2 MB2 B∗2 MB1 → 0 and  −1 B∗1 MB2 ℓI − B∗2 MB2 B∗2 W∗Λ∗ B1 → 0. Note that −1  −1  1 B∗1 ΛWB2 ℓI − B∗2 MB2 B∗2 MB1 = B∗1 ΛW I − B2 B∗2 M B1 − B∗1 ΛWB1 ℓ and that −1  −1  1 W∗ Λ∗ B1 − B∗1 W∗Λ∗ B1 . B∗1 MB2 ℓI − B∗2 MB2 B∗2 Λ∗ WB1 = B∗1 I − MB2 B∗2 ℓ 19 Hence, by WB1 = oa.s. (1) and B∗1 W∗ = oa.s. (1), it suffices to show with probability 1 that,  −1 1 B∗1 ΛW I − B2 B∗2 M B1 → 0 ℓ and −1  1 W∗ Λ∗ B1 → 0. B∗1 I − MB2 B∗2 ℓ By B1 = Λ(Λ∗ Λ)−1/2 , we have −1  1 ∗ I − B2 B2 M B1 ℓ −1  1 ∗ ∗ 1/2 Λ(Λ∗ Λ)−1/2 = (Λ Λ) W I − B2 B2 M ℓ  −1 1 1/2 ∗ = Q W I − B2 B2 M ΛQ−1/2 . ℓ B∗1 ΛW By law of large numbers, we have with probability 1 that, = → = = −1  1 Λ W I − B2 B∗2 M ℓ T  −1 1 1 X (Fj ε∗j+τ + Fj+τ ε∗j ) I − B2 B∗2 M Λ 2T j=1 ℓ  −1 1 ∗ ∗ ∗ E(F1 ε1+τ + F1+τ ε1 ) I − B2 B2 M Λ ℓ     −1 1 ∗ ∗ ∗ Λ ε1 , · · · , εT+τ EE (F1 ε1+τ + F1+τ ε1 ) I − B2 B2 M ℓ o(1). Hence, we have with probability 1 B∗1 ΛW  −1 −1  1 1 ∗ ∗ 1/2 I − B2 B2 M B1 = Q W I − B2 B2 M ΛQ−1/2 = o(1). ℓ ℓ Similarly, −1  1 W∗ Λ∗ B1 = oa.s. (1). B∗1 I − MB2 B∗2 ℓ 20 Therefore, ℓ should satisfy  −1 −1 det Q1/2 H(τ )Q1/2 + B∗1 M − ℓI B1 +  (1 + β 2 )2  (1 + β 2 )2  α 1/2 h 2 3 H(τ ) 1+β − H(0) + 3β + β − Q 2 β2 β i (1 + β 2 )2 + G(τ ) Q1/2 → 0. 2 β Recall B1 = Λ(Λ∗ Λ)−1/2 . Our next goal is to find the limit of B∗1 M − ℓI −1 B1 . Define A = M − ℓI and Ak = A − (γk+τ + γk−τ )γk∗ − γk (γk+τ + γk−τ )∗ , then we have the following lemmas, with proofs given in the Appendix. Lemma 5.2 Let x ∈ Cn1 := {x ∈ Cn : kxk = 1} be given. For r ≥ 1, we have 2r E|γk∗ A−1 ≤ KT −r k x| for some K > 0. Lemma 5.3 For any x, y ∈ Cn1 , we have x∗ A−1 y → − x∗ y cm √ +ℓ 1−c2 m2 + 1−c2 m2 a.s. Finally, we have  cm(ℓ) p + ℓ Ik(q+1) 1 − c2 m2 (ℓ) + 1 − c2 m2 (ℓ) h  (1 + β 2 )2  (1 + β 2 )2  α 3 H(τ ) H(0) + 3β + β − + Q1/2 1 + β 2 − 2 β2 β i (1 + β 2 )2 G(τ ) Q1/2 = 0, + β2 det Q1/2 H(τ )Q1/2 −  or equivalently  (1 + β 2 )2  (1 + β 2 )2 α h 2 3 1 + β2 − H(τ ) H(0) + 3β + β − + 2 β2 β α  i  (1 + β 2 )2 cm(ℓ) p + + ℓ Q−1 = 0. G(τ ) − 2 2 2 2 2 β 1 − c m (ℓ) + 1 − c m (ℓ) det 21 (5.3) When τ > q, one has H(τ ) = 0, G(τ ) = 2I and (5.3) reduces to det α(1 + β 2 )I − or det   cm(ℓ) p + ℓ Q−1 = 0, 1 − c2 m2 (ℓ) + 1 − c2 m2 (ℓ) cm(ℓ) p (Q + I) + ℓI = 0. 1 − c2 m2 (ℓ) + 1 − c2 m2 (ℓ) Let λ be an eigenvalue of Q, then we have 1 ℓ cm(ℓ) p 1+ + = 0. λ λ 1 − c2 m2 (ℓ) + 1 − c2 m2 (ℓ) (5.4) When τ = q, (5.3) reduces to     α cm(ℓ) p det α(1 + β 2 )I + 1 + (3β + β 3 ) H(q) − + ℓ Q−1 = 0. 2 1 − c2 m2 (ℓ) + 1 − c2 m2 (ℓ) Writing  0 .  ..  . . H(q) =  .   0  1 ··· .. . .. . .. . 0  · · · 0 1  .. .. . . 0  . . . . ..  . . .  ⊗ Ik ,  . . . . ..  . . .  ··· ··· 0 one can easily verify that the eigenvalues of H(q) are 1, −1 and 0, with multiplicity k, k and k(q − 1), respectively. Suppose that H(q) and Q are commutative, that is, there is a common orthogonal matrix O simultaneously diagnalizing the two matrices, i.e., we have H(q) = ODH O′ and Q = ODQ O′ , where DH = diag[a1 , · · · , ak(q+1) ] and DQ = diag[λ1 , · · · , λk(q+1) ]. Then, (5.3) further reduces to  α (3β + β 3 ) aj 2 1 ℓ cm(ℓ) p 1+ + , = 2 2 2 2 λj λj 1 − c m (ℓ) + 1 − c m (ℓ)  1+ 22 j = 1, · · · , k(q + 1). Substituting α = − cm 2 2 2 1− c2xm 1 aj and β = − cm 2 2 2 1− c4xm 1 1 , for j = 1, · · · , k(q + 1), we have −1 1 p × 2 1 − c2 m2 (ℓ) + 1 − c2 m2 (ℓ) h 1 ℓi cm(ℓ) p 1+ + λj λj 1 − c2 m2 (ℓ) + 1 − c2 m2 (ℓ) =: gj (ℓ). = + (5.5) Notice that (5.4) is a special case of (5.5) for aj = 0. Note that when x is outside the support [−d(c), d(c)] of the LSD of Mn (τ ), m2 (x) 6= 0. Hence, we have cm1 (x)((1 − c − cxm1 (x))2 − c2 x2 m22 (x)) +x(1 − c2 m21 (x) + c2 m22 (x))(1 − c − cxm1 (x)) = 0. Let x ↓ d(c) := d and we have m2 (x) → 0 and m1 (x) → m1 (d) satisfying cm1 (d)(1 − c − cdm1 (d))2 + d(1 − c2 m21 (d))(1 − c − cdm1 (d)) = [1 − c − cdm1 (d)][cm1 (d)(1 − c − cdm1 (d)) + d(1 − c2 m21 (d))] = 0, from which we have m1 (d) = 1−c− √ (1−c)2 +8d2 . 4cd Rewrite cm(ℓ) 1 p 1+ 1 λj − + 2 1 − c2 m2 (ℓ) 1  −1 ℓ 1 p + + 2 1 − c2 m2 (ℓ) + 1 − c2 m2 (ℓ) λj := gj1(ℓ) + gj2 (ℓ). gj (ℓ) = 3 2 1 2 2 c m (ℓ) 2 We will show that gj (ℓ) is increasing over (d(c), ∞) by showing so are gj1 and gj2. By definition, over (d(c), ∞), m is an increasing function taking negative values and m2 is a decreasing function taking positive values. Hence, it is easy to see that gj1(ℓ) is increasing over (d(c), ∞). p For gj2(ℓ), define h(ℓ) = 1 − c2 m2 (ℓ) + 1 − c2 m2 (ℓ) and rewrite gj2 (ℓ) = ℓh(ℓ) h(ℓ) . It is easy λj (1+ 23 2 ) to see that h(ℓ) > 0 and h′ (ℓ) > 0 over (d(c), ∞). Hence we have ′ gj2 (ℓ) = [h(ℓ) + ℓh′ (ℓ)](1 + λj (1 + h(ℓ) )− 2 h(ℓ) 2 ) 2 ′ ℓh(ℓ) h 2(ℓ) = h2 (ℓ) 2 h(ℓ) 2 ) 2 h(ℓ) + ℓh′ (ℓ) + λj (1 + > 0. By symmetry, gj (ℓ) is increasing over (−∞, −d(c)) as well. Therefore, based on the sign of gj (d(c)), we have the following cases to consider. Case I. gj (d(c)) ≥ 0: i. If aj > gj (d(c)), then (5.5) has one solution in (d(c), ∞) and no solution in (−∞, −d(c)). ii. If g(d(c)) ≥ aj ≥ −gj (d(c)) = gj (−d(c)), then (5.5) has no solution in (d(c), ∞) and (−∞, −d(c)). iii. If aj < −gj (d(c)), then (5.5) has one solution in (−∞, −d(c)) and no solution in (d(c), ∞). Case II. gj (d(c)) < 0: i. If aj ≥ −gj (d(c)), then (5.5) has one solution in (d(c), ∞) and no solution in (−∞, −d(c)). ii. If g(d(c)) < aj < −gj (d(c)) = gj (−d(c)), then (5.5) has one solution in (d(c), ∞) and one solution in (−∞, −d(c)). iii. If aj ≤ gj (d(c)), then (5.5) has one solution in (−∞, −d(c)) and no solution in (d(c), ∞). Remark 5.1 In real application, compared with the noise component, the loading matrix Λ dominates. As a result, all the eigenvalues of Q = Λ∗ Λ are large (more precisely, they are of the same order as n). Hence, we can assume that Q−1 = 0. Thus the commutative assumption of H(q) and Q can be relaxed. Moreover, under this case, we always have gj (d(c)) < 0. Remark 5.2 For the same reason as stated before Remark 4.1, dc is replaced by (1 + anb )dc in practice. Simulation indicates a fit of a = 0.1, b = −1/3. Notice that all the eigenvalues of H(q + 1) are 0, while for H(q), k eigenvalues are 1 and k eigenvalues are −1, with the rest being 0. Making use of such difference and applying the above analysis to the cases that τ = q and τ = q + 1 gives an estimate of q. Together with the estimation of k(q + 1), we easily obtain the estimate of k. A numerical demonstration is given in the simulation. 24 6 Estimate of σ 2 The above estimation is based on the assumption that σ 2 , the variance of the noise part is given. In practice, it is often the case that σ 2 is unknown. To this end, we can estimate σ 2 by employing the properties of the MP law. More precisely, we first estimate the left boundary of the support of the MP law by the smallest sample eigenvalue of Φn (0), say λ̂1 (the eigenvalues are arranged in ascending order), and estimate the right boundary by √ (1+ c)2 √ λ̂ . (1− c)2 1 An iteration is then applied. 2 The initial estimator of σ 2 , say σ̂(0) , is obtained as the sample mean of the sample eigenvalues of Φn (0) that lie within the interval [λ̂m+1 , λ̂n−m ], where m ≥ 0 is such that λ̂n−m is the largest eigenvalue of Φn (0) less than √ (1+ c)2 (1+2n−2/3 ) √ λ̂1 . (1− c)2 For i ≥ 1, we obtain the updated estimator 2 σ̂(i) by taking the sample mean of the sample eigenvalues of Φn (0) that lie within the interval √ √ 2 2 2 2 , (1 + c)2 (1 + 2n−2/3 )σ̂(i−1) ]. The iteration stops once we have σ̂(ℓ−1) = σ̂(ℓ) [(1 − c)2 σ̂(i−1) 2 for some ℓ and our estimator σ̂ 2 := σ̂(ℓ) . As shown in the simulation, our estimation of k and q still works well with such estimator. 7 Simulation Table 1 presents a simulation about the result discussed above, displaying the largest 13 absolute values of the eigenvalues for lags τ from 0 to 5. Here Rt = q X Λi ft−i + et , t = 1, ..., T (7.1) i=0 where ft ’s are factors of length k; Λi , i = 0, ..., q is a constant time-invariant matrix of size n × k, et is the error term and q is the lag of the model. In addition, assume that: et are i.i.d.  random variables with et ∼ N (0, σ 2 In ) and ft are i.i.d. random variables with ft ∼ N 0, σf2 Ik , h i independent of et . Λi = Λi1 Λi2 . . . Λik where Λij i = 0, ..., q; j = 1, ..., k is a vector of length n and is given by Λij = β1n + εij where 1n is a vector of 1’s and εij are i.i.d. random variables with εij ∼ N (0, σε2 In ). For n = 450, T = 500, k = 2, q = 2, β = 1.0, 25 σf2 = 4, σ 2 = 1 and σε2 = 0.25, we have c = 0.9, bc = (1 + √ c)2 = 3.7974 and dc = 1.8573. Eigenvalues of Q are from 95 to 285, making Q−1 ∼ 0. When τ = 0, and σ 2 = 1 is known, using the phase transition point bc = (1 + √ c)2 σ 2 = 3.7974, we see that the number of spotted spikes is 6, which estimates k(q + 1). When for τ = q + 1, we have H(τ ) = 0. Moreover, as Q−1 ∼ 0, we have gj (d(c)) ∼ −0.4284 < 0. That is, our Case II (ii) applies for all the k(q + 1) 0 eigenvalues of H(q + 1), making the number of spikes 2k(q + 1) as verified by applying the phase transition point dc = 1.8573. For τ = q, H(τ ) has k eigenvalues of 1, k eigenvalues of −1 and k(q − 1) eigenvalues of 0 with Case II (i),(iii) and (ii) applicable, respectively. Thus, we have k + k + 2k(q − 1) = 2kq < 2k(q + 1) eigenvalues in this case. Again, this agrees with the use of the phase transition point dc = 1.8573. In other words, the number of spikes first jumps to 2k(q + 1) at τ = 3 which estimates q + 1. The estimation of k is obvious. When σ 2 = 1 is unknown, using technique as in Section 6, one has σ̂ 2 = 0.9894. It then √ follows that b̂c = (1 + c)2 (1 + 2n−2/3 )σ̂ 2 = 3.8851, dˆc = 1.8616 (with rescale factor 1 + 0.1n−1/3 ), which gives the same estimates as above. A Some proofs A.1 Proof of Lemma 5.1 Define Mk = M − γk (γk+τ + γk−τ )∗ − (γk+τ + γk−τ )γk∗ , and ∗ ∗ Mk,k+τ,··· ,k+lτ = Mk,k+τ,··· ,k+(l−1)τ − γk+(l+1)τ γk+lτ − γk+lτ γk+(l+1)τ , l ≥ 1. 26 τ =0 τ =1 τ =2 τ =3 τ =4 τ =5 10031.2366 6227.5906 2865.1554 640.5761 155.9377 128.6870 534.5839 363.7782 258.4859 48.9667 46.5224 92.3756 473.1639 325.8391 224.9343 22.7478 46.0225 53.7072 458.2226 305.6334 214.9755 21.6373 45.6650 26.3564 435.2661 13.1683 45.7319 21.3150 25.4884 25.6006 392.6272 11.1482 17.8374 19.2596 18.0820 19.3930 3.6928 9.0674 11.7423 10.3580 15.5876 14.6107 3.5809 7.9537 8.0837 9.9668 12.9088 12.0980 3.5449 1.7375 1.7988 9.5028 10.5568 8.5791 3.4579 1.7326 1.7895 8.5483 4.7840 7.0596 3.4312 1.7015 1.7388 5.5931 4.3896 4.5411 3.3829 1.6957 1.7242 3.5968 4.3843 3.6744 3.3701 1.6751 1.6724 1.8215 1.7944 1.7468 Table 1: Absolute values of the largest eigenvalues of the empirical covariance matrix at various lags with parameters: n = 450, T = 500, k = 2, q = 2, β = 1.0, σf2 = 4, σ 2 = 1 and σε2 = 0.25. √ Note that c = 0.9, bc = (1 + c)2 = 3.7974 and dc = 1.8573. When σ 2 = 1 is unknown, one has σ̂ 2 = 0.9894, b̂c = 3.8851 and dˆc = 1.8616. 27 Suppose that i ≥ j, then we have γi∗ B2  B∗2 MB2 ℓI − −1 B∗2 γj −1  ∗ )B2 B∗2 γj = γi∗ B2 ℓI − B∗2 Mj B2 − B∗2 (γj+τ + γj−τ )γj∗ B2 − B∗2 γj (γj+τ + γj−τ  −1 γi∗ B2 ℓI − B∗2 Mj B2 − B∗2 (γj+τ + γj−τ )γj∗ B2 B∗2 γj =  −1 ∗ 1 − (γj+τ + γj−τ )B2 ℓI − B∗2 Mj B2 − B∗2 (γj+τ + γj−τ )γj∗ B2 B∗2 γj   (ℓI−B∗2 Mj B2 )−1 B∗2 (γj+τ +γj−τ )γj∗ B2 (ℓI−B∗2 Mj B2 )−1 B∗2 γj γi∗ B2 (ℓI − B∗2 Mj B2 )−1 + 1−γj∗ B2 (ℓI−B∗2 Mj B2 )−1 B∗2 (γj+τ +γj−τ )   = (ℓI−B∗2 Mj B2 )−1 B∗2 (γj+τ +γj−τ )γj∗ B2 (ℓI−B∗2 Mj B2 )−1 ∗ ∗ −1 B∗2 γj 1 − (γj+τ + γj−τ )B2 (ℓI − B2 Mj B2 ) + 1−γ ∗ B2 (ℓI−B∗ Mj B2 )−1 B∗ (γj+τ +γj−τ ) j γi∗ B2 =  2 2 (ℓI−B∗2 Mj B2 )−1 +(ℓI−B∗2 Mj B2 )−1 B∗2 (γj+τ +γj−τ )γj∗ B2 (ℓI−B∗2 Mj B2 )−1   B∗2 γj  + oa.s. (1) ∗ 1−(γj+τ +γj−τ )B2 (ℓI−B∗2 Mj B2 )−1 +(ℓI−B∗2 Mj B2 )−1 B∗2 (γj+τ +γj−τ )γj∗ B2 (ℓI−B∗2 Mj B2 )−1 B∗2 γj =        − cm 2 2 2 1− c2xm 1 − cm 2 2 2 1− c2xm 1 + oa.s. (1), i=j γi∗ B2 (ℓI − B∗2 Mj B2 )−1 B∗2 (γj+τ + γj−τ ) + oa.s. (1), otherwise. Next, we have γi∗ B2 (ℓI − B∗2 Mj B2 )−1 B∗2 γj+τ ∗ ∗ = γi∗ B2 (ℓI − B∗2 Mj,j+τ B2 − B∗2 γj+2τ γj+τ B2 − B∗2 γj+τ γj+2τ B2 )−1 B∗2 γj+τ ∗ γi∗ B2 (ℓI − B∗2 Mj,j+τ B2 − B∗2 γj+2τ γj+τ B2 )−1 B∗2 γj+τ = ∗ ∗ 1 − γj+2τ B2 (ℓI − B∗2 Mj,j+τ B2 − B∗2 γj+2τ γj+τ B2 )−1 B∗2 γj+τ   ∗ (ℓI−B∗2 Mj,j+τ B2 )−1 B∗2 γj+2τ γj+τ B2 (ℓI−B∗2 Mj,j+τ B2 )−1 B∗2 γj+τ γi∗ B2 (ℓI − B∗2 Mj,j+τ B2 )−1 + ∗ 1−γj+τ B2 (ℓI−B∗2 Mj,j+τ B2 )−1 B∗2 γj+2τ   = ∗ (ℓI−B∗2 Mj,j+τ B2 )−1 B∗2 γj+2τ γj+τ B2 (ℓI−B∗2 Mj,j+τ B2 )−1 ∗ ∗ −1 B∗2 γj+τ 1 − γj+2τ B2 (ℓI − B2 Mj,j+τ B2 ) + 1−γ ∗ B2 (ℓI−B∗ Mj,j+τ B2 )−1 B∗ γj+2τ j+τ =  2 2  ∗ B2 (ℓI−B∗2 Mj,j+τ B2 )−1 B∗2 γj+τ γi∗ B2 (ℓI−B∗2 Mj,j+τ B2 )−1 +(ℓI−B∗2 Mj,j+τ B2 )−1 B∗2 γj+2τ γj+τ   ∗ ∗ B2 (ℓI−B∗2 Mj,j+τ B2 )−1 B∗2 γj+τ 1−γj+2τ B2 (ℓI−B∗2 Mj,j+τ B2 )−1 +(ℓI−B∗2 Mj,j+τ B2 )−1 B∗2 γj+2τ γj+τ =        − cm 2 2 2 1− c4xm 1 − cm 2 2 2 1− c4xm 1 + oa.s. (1), i=j+τ γi∗ B2 (ℓI − B∗2 Mj,j+τ B2 )−1 B∗2 γj+2τ + oa.s. (1), 28 otherwise. + oa.s. (1) γi∗ B2 (ℓI − B∗2 Mj B2 )−1 B∗2 γj−τ ∗ ∗ = γi∗ B2 (ℓI − B∗2 Mj,j−τ B2 − B∗2 γj−2τ γj−τ B2 − B∗2 γj−τ γj−2τ B2 )−1 B∗2 γj−τ ∗ γi∗ B2 (ℓI − B∗2 Mj,j−τ B2 − B∗2 γj−2τ γj−τ B2 )−1 B∗2 γj−τ ∗ ∗ 1 − γj−2τ B2 (ℓI − B∗2 Mj,j−τ B2 − B∗2 γj−2τ γj−τ B2 )−1 B∗2 γj−τ   ∗ (ℓI−B∗2 Mj,j−τ B2 )−1 B∗2 γj−2τ γj−τ B2 (ℓI−B∗2 Mj,j−τ B2 )−1 ∗ ∗ −1 γi B2 (ℓI − B2 Mj,j−τ B2 ) + B∗2 γj−τ ∗ 1−γj−τ B2 (ℓI−B∗2 Mj,j−τ B2 )−1 B∗2 γj−2τ   = ∗ (ℓI−B∗2 Mj,j−τ B2 )−1 B∗2 γj−2τ γj−τ B2 (ℓI−B∗2 Mj,j−τ B2 )−1 ∗ B∗2 γj−τ 1 − γj−2τ B2 (ℓI − B∗2 Mj,j−τ B2 )−1 + ∗ ∗ ∗ −1 1−γ B2 (ℓI−B Mj,j−τ B2 ) B γj−2τ = j−τ = γi∗ B2  2 ∗ (ℓI−B∗2 Mj,j−τ B2 )−1 +(ℓI−B∗2 Mj,j−τ B2 )−1 B∗2 γj−2τ γj−τ B2 (ℓI−B∗2 Mj,j−τ B2 )−1  2  B∗2 γj−τ  + oa.s. (1) ∗ ∗ B2 (ℓI−B∗2 Mj,j−τ B2 )−1 B∗2 γj−τ 1−γj−2τ B2 (ℓI−B∗2 Mj,j−τ B2 )−1 +(ℓI−B∗2 Mj,j−τ B2 )−1 B∗2 γj−2τ γj−τ = Note that − cm 2 2 2 1− c4xm − cm 2 1− = c2 m2 4x1 a x1 γi∗ B2 (ℓI − B∗2 Mj,j−τ B2 )−1 B∗2 γj−2τ + oa.s. (1). < 1, by induction, we have 1 γi∗ B2 (ℓI − B∗2 Mj B2 )−1 B∗2 γj−τ = oa.s. (1). Then result then follows by induction. By symmetry, it holds when i < j. The proof of the lemma is complete. A.2 Proof of Lemma 5.2 ′ 2 Let A−1 k x = b = (b1 , · · · , bn ) . Noting |εij | < C and E|εij | = 1, we have 2r E(γk∗ A−1 k x) n X 1 εki bi )2r = r r E( 2T i=1 r X X 1 ≤ r rE 2T l=1 1≤j <···<j ≤n i 1 = ≤ 1 2r T r K 2r T r E E r X l X 1 (2r)! i1 i1 εkj1 bj1 · · · εikjl l bijl1 i ! · · · il ! +···+i =2r 1 X l (2r)! i1 i1 εkj1 bj1 · · · εikjl l bijl1 i ! · · · il ! +···+i =2r, 1 X l=1 1≤j1 <···<jl ≤n i1 l i1 ≥2,··· ,il ≥2 r X X X (2r)! |bj1 |i1 · · · |bjl |il . i ! · · · i ! 1 l +···+i =2r, l=1 1≤j1 <···<jl ≤n i1 l i1 ≥2,··· ,il ≥2 29 By Pn j=1 2 |bj |2 = kA−1 k xk and Cauchy-Schwartz inequality, we have X ≤ ≤ (2r)! |bj1 |i1 · · · |bjl |il i ! · · · il ! +···+i =2r, 1 X 1≤j1 <···<jl ≤n i1 l i1 ≥2,··· ,il ≥2 n X X (2r)! ( |bj |2 )r i ! · · · il ! j=1 +···+i =2r, 1 i1 l i1 ≥2,··· ,il ≥2 2r 2r l2r ||A−1 k ||o ||x|| 2r ≤ l . η 2r Here η := ℓ − dc > 0. Therefore, we have 2r E|γk∗ A−1 ≤ KT −r k x| for some K > 0. The proof of the lemma is complete. A.3 Proof of Lemma 5.3 ∗ ∗ P ROOF. Let x, y ∈ Cn1 be given. Define Ak,k+τ = Ak − γk+τ γk+2τ − γk+2τ γk+τ , ∗ Ãk = Ak + γk (γk+τ + γk−τ )∗ and Ãk,k+τ = Ak − γk+2τ γk+τ . First we have x∗ A−1y − Ex∗ A−1 y T X  ∗ (Ek − Ek−1 ) A−1 − A−1 (ℓ) y = x k = k=1 T X k=1 ≡ T X k=1 (Ek − Ek−1 ) − ∗ −1  ∗ −1 x∗ A−1 x∗ Ã−1 k γk (γk+τ + γk−τ ) Ak y k (γk+τ + γk−τ )γk Ãk y − 1 + (γk+τ + γk−τ )∗ A−1 1 + γk∗ Ã−1 k γk k (γk+τ + γk−τ ) (Ek − Ek−1 ) − αk1 − αk2 ). 30 Using Lemma 2.1, we have, for i = 1, 2 E| T X k=1 (Ek − Ek−1 )αki |2l  T X ≤ Kl  E k=1  T X ≤ Kl′ E ≤ 2 Note that l Kl′ k=1  E Ak−1 Ek−1|(Ek − Ek−1 )αki |2 Ek−1 |Ek αki |2 + T X k=1 Ek−1 |αki | = (Ãk,k+τ + 2 !l T X k=1 + !l + T X k=1 Ek−1 |αki |2 T X k=1 ∗ γk+2τ γk+τ )−1  E|(Ek − Ek−1 )αki |2l  !l + T X k=1  E|Ek αki |2l + T X k=1  E|Ek−1 αki |2l  E|αki |2l  . = Ã−1 k,k+τ − (A.1) −1 ∗ Ã−1 k,k+τ γk+2τ γk+τ Ãk,k+τ ∗ Ã−1 1 + γk+τ k,k+τ γk+2τ . Hence, we have ∗ ∗ −1 γk+τ A−1 k = γk+τ Ãk,k+τ − −1 ∗ ∗ γk+τ Ã−1 k,k+τ γk+2τ γk+τ Ãk,k+τ ∗ Ã−1 1 + γk+τ k,k+τ γk+2τ = ∗ γk+τ Ã−1 k,k+τ ∗ Ã−1 1 + γk+τ k,k+τ γk+2τ . Next, we have ∗ ∗ −1 γk+τ Ã−1 k,k+τ = γk+τ Ak,k+τ − ∗ = γk+τ A−1 k,k+τ − −1 ∗ ∗ γk+τ A−1 k,k+τ γk+τ γk+2τ Ak,k+τ ∗ 1 + γk+2τ A−1 k,k+τ γk+τ cm ∗ γ A−1 + Rk1 , 2 k+2τ k,k+τ where Rk1 −1 ∗ ∗ γk+τ A−1 cm ∗ k,k+τ γk+τ γk+2τ Ak,k+τ −1 = γ A − ∗ 2 k+2τ k,k+τ 1 + γk+2τ A−1 k,k+τ γk+τ ! −1 −1 cm ∗ cm ∗ − γ A γ + γ A γ k+τ k,k+τ k+τ 2 2 k+2τ k,k+τ k+τ ∗ γk+2τ A−1 = k,k+τ . −1 ∗ 1 + γk+2τ Ak,k+τ γk+τ Substitute back, we obtain ∗ γk+τ A−1 k y = ∗ γk+τ A−1 k,k+τ y − ∗ A−1 γk+τ k,k+τ γk+τ ∗ A−1 1+γk+2τ k,k+τ γk+τ ∗ A−1 1 + γk+τ k,k+τ γk+2τ − ∗ γk+2τ A−1 k,k+τ y cm ∗ A−1 γ γ 2 k+2τ k,k+τ k+2τ 31 + Rk1 γk+2τ , (A.2) with Rk1 γk+2τ = cm 2 ∗ − γk+τ A−1 k,k+τ γk+τ + cm ∗ γ A−1 γ 2 k+2τ k,k+τ k+τ ∗ A−1 γk+2τ k,k+τ γk+τ 1+ ! ∗ γk+2τ A−1 k,k+τ γk+2τ . Note that lim n→∞ = 1 ∗ γk+τ A−1 k,k+τ γk+τ ∗ A−1 1 + γk+2τ k,k+τ γk+τ cm 2 cm − cm 2 2x1 = cm 2 x1 <1  ∗ A−1 1 + γk+τ k,k+τ γk+2τ − cm ∗ A−1 γ γ 2 k+2τ k,k+τ k+2τ + Rk1 γk+2τ  and ∗ lim 1 + γk+τ A−1 k,k+τ γk+2τ − n→∞ cm ∗ cm cm γk+2τ A−1 , k,k+τ γk+2τ + Rk1 γk+2τ = 1 − 2 2 2x1 −1 ∗ ∗ which is bounded. Using induction, we have |γk+τ A−1 k y| ≤ K|γk+τ Ak,k+τ y|. −1 ∗ ∗ ∗ −1 ∗ −1 Similarly, we have |γk−τ A−1 k y| ≤ K|γk−τ Ak,k−τ y|, |x Ak γk+τ | ≤ K|x Ak,k+τ γk+τ | and ∗ −1 |x∗ A−1 k γk−τ | ≤ K|x Ak,k−τ γk−τ |. 2 ∗ −1 −1 Therefore, by noting (γk+τ + γk−τ )∗ A−1 k γk = oa.s. (1), |εit | < C, E|εit | = 1 and x Ak Āk x being bounded, we have E T X k=1 Ek−1|αk2 |2 T X ≤ KE k=1 !l −1 2 Ek−1 |x∗ A−1 k γk γk+τ Ak y| !l T X 1 −1 2 Ek−1 x∗ Ak−1 Ā−1 = KE k x|γk+τ Ak y| 2T k=1 2l ≤ K max E|γk+τ A−1 k,k+τ y| k ≤ K Tl 32 !l and T X = k=1 T X k=1 E|αk2 |2l −1 2 E |x∗ A−1 k γk γk+τ Ak y| K 2l max E|γk+τ A−1 k,k+τ y| l−1 k T K ≤ T 2l−1 l ≤ −1 For i = 1, by Ã−1 k = Ak − ∗ −1 A−1 k γk (γk+τ +γk−τ ) Ak , −1 ∗ 1+(γk+τ +γk−τ ) Ak γk we have ∗ −1 x∗ Ã−1 k (γk+τ + γk−τ )γk Ãk y   ∗ −1  ∗ −1  A−1 A−1 ∗ −1 k γk (γk+τ + γk−τ ) Ak k γk (γk+τ + γk−τ ) Ak (γ + γ )γ A − y = x∗ A−1 − k+τ k−τ k k k 1 + (γk+τ + γk−τ )∗ A−1 1 + (γk+τ + γk−τ )∗ A−1 k γk k γk ∗ −1 = x∗ A−1 k (γk+τ + γk−τ )γk Ak y (γk+τ + γk−τ )∗ A−1 k (γk+τ + γk−τ ) ∗ −1 x Ak γk γk∗ A−1 k y 1 + (γk+τ + γk−τ )∗ A−1 γ k k γk∗ A−1 ∗ −1 k γk x∗ A−1 − k (γk+τ + γk−τ )(γk+τ + γk−τ ) Ak y −1 ∗ 1 + (γk+τ + γk−τ ) Ak γk γ ∗ A−1 γk (γk+τ + γk−τ )∗ A−1 k (γk+τ + γk−τ ) ∗ −1 + k k x Ak γk (γk+τ + γk−τ )∗ A−1 2 k y −1 ∗ 1 + (γk+τ + γk−τ ) Ak γk =: αk11 − αk12 − αk13 + αk14 − It is easy to see that work on αk11 and αk14 is the same as that on αk2 . For αk12 , by Cauchy-Schwartz’s inequality, we have E T X k=1 ≤ KE Ek−1 |αk12 |2 T X k=1 !l 2 ∗ −1 2 Ek−1|x∗ A−1 k γk | Ek−1 |γk Ak y| !l T X 1 −1 ∗ −1 −1 Ek−1 x∗ A−1 = KE k Āk xEk−1 y Āk Ak y 2 4T k=1 ≤ K Tl 33 !l −1 and by |εit | < C, E|εit |2 = 1 and x∗ A−1 k Āk x being bounded, we have T X k=1 ≤ K E|αk12 |2l T X k=1 2 ∗ −1 2 E |x∗ A−1 k γk | |γk Ak y| K 2l ≤ max E|γk A−1 k y| T l−1 k K . ≤ T 2l−1 l −1 ∗ ∗ ∗ −1 ∗ −1 By the fact that |γk±τ A−1 k y| ≤ K|γk±τ Ak,k±τ y| and |x Ak γk±τ | ≤ K|x Ak,k±τ γk±τ |, the similar result for αk13 follows by the same reason. Substituting all the above results into (A.1) and choosing l large enough, we have x∗ A−1 y − Ex∗ A−1 y → 0 a.s. Next, we want to show the convergence of Ex∗ A−1 y. By A= T X k=1 we have In = T X k=1 ∗ (γk γk+τ + γk+τ γk∗ ) − ℓIn ∗ (γk γk+τ A−1 + γk+τ γk∗ A−1 ) − ℓA−1 . Multiplying x∗ from left and y from right and taking expectation, we obtain ∗ x y = T X k=1 = T X k=1 ∗ (Ex∗ γk γk+τ A−1 y + Ex∗ γk+τ γk∗ A−1 y) − ℓEx∗ A−1 y Ex∗ (γk+τ + γk−τ )γk∗ A−1 y − ℓEx∗ A−1 y. (A.3) By A = Ãk + (γk+τ + γk−τ )γk∗ , Ãk = Ak + γk (γk+τ + γk−τ )∗ , (γk+τ + γk−τ )∗ A−1 k γk = oa.s. (1), 34 γk∗ A−1 k γk = cm 2 + oa.s. (1), we have γk∗ A−1 y γk∗ Ã−1 k y 1 + γk∗ Ã−1 k (γk+τ + γk−τ )   ∗ −1 A−1 k γk (γk+τ +γk−τ ) Ak γk∗ A−1 − y k 1+(γk+τ +γk−τ )∗ A−1 k γk   = ∗ −1 A−1 k γk (γk+τ +γk−τ ) Ak (γk+τ + γk−τ ) 1 + γk∗ A−1 −1 ∗ k − 1+(γ +γ ) A γ = k+τ = = k−τ k k ∗ −1 ∗ −1 γk∗ A−1 k y + γk Ak y(γk+τ + γk−τ ) Ak γk ∗ −1 [1 + (γk+τ + γk−τ )∗ A−1 k γk ][1 + γk Ak (γk+τ + γk−τ )] − cm 2 (γk+τ + γk−τ )∗ A−1 k y + oa.s. (1). c2 m2 1 − 2x1 ∗ −1 − γk∗ A−1 k γk (γk+τ + γk−τ ) Ak y ∗ −1 − γk∗ A−1 k γk (γk+τ + γk−τ ) Ak (γk+τ + γk−τ ) (A.4) Next, we have = = = ∗ ∗ γk+τ A−1 k yx γk+τ −1 −1 −1 ∗ ∗ ∗ ∗  (1 + γk+2τ A−1 k,k+τ γk+τ )γk+τ Ak,k+τ y − γk+τ Ak,k+τ γk+τ γk+2τ Ak,k+τ y −1 ∗ ∗ (1 + γk+2τ A−1 k,k+τ γk+τ )(1 + γk+τ Ak,k+τ γk+2τ ) ∗ ∗ γk+τ A−1 k,k+τ yx γk+τ + oa.s. (1) −1 ∗ ∗ A−1 1 − γk+τ k,k+τ γk+τ γk+2τ Ak,k+τ γk+2τ x∗ A−1 k,k+τ y + oa.s. (1) . 2 2 2T (1 − c4xm1 ) ∗ ∗ Similarly, we can show that γk−τ A−1 k yx γk−τ = − −1 ∗ ∗ γk+τ A−1 k,k+τ γk+τ γk+2τ Ak,k+τ γk+2τ x∗ A−1 k,k−τ y+oa.s. (1) 2 2 2T (1− c4xm )  x∗ γk+τ ∗ ∗ , γk+τ A−1 k yx γk−τ = oa.s. (1) 1 ∗ ∗ ∗ −1 ∗ −1 and γk−τ A−1 k yx γk+τ = oa.s. (1). Next, we will show that Ex Ak,k±τ y − Ex A y = o(1). By writing ∗ −1 ∗ −1 ∗ −1 ∗ −1 ∗ −1 Ex∗ A−1 k,k±τ y − Ex A y = Ex Ak,k±τ y − Ex Ak y + Ex Ak y − Ex A y, ∗ −1 it is sufficient to show Ex∗ A−1 k y − Ex A y = o(1). Note that ∗ −1 Ex∗ A−1 k y − Ex A y ∗ −1  ∗ −1  x∗ A−1 x∗ Ã−1 k γk (γk+τ + γk−τ ) Ak y k (γk+τ + γk−τ )γk Ãk y + E 1 + (γk+τ + γk−τ )∗ A−1 1 + γk∗ Ã−1 k γk k (γk+τ + γk−τ ) = Eαk1 + Eαk2 = E Previous calculation shows that E|αk1 | = o(1) and E|αk2 | = o(1). Substituting these back to (A.3) and (A.4), we finish proving the lemma. 35 References [1] Bai, Z.D., Liu H.X. and Wong, W.K. (2011) Asymptotic properties of eigenmatrices of a large sample covariance matrix. Ann. Appl. Probab. 21, 1994–2015. [2] Bai, Z.D. and Silverstein, J.W. (1998) No eigenvalues outside the support of the limiting spectral distribution of large-dimensional sample covariance matrices. Ann. Probab. 26, 316–345. [3] Bai, Z.D. and Silverstein, J.W. (2010) Spectral Analysis of Large Dimensional Random Matrices, 2nd ed. Springer Verlag, New York. [4] Bai, Z.D. and Yao, J.F. (2008) Central limit theorems for eigenvalues in a spiked population model. Ann. Inst. H. Poincar?????? Probab. Statist. Volume 44, Number 3, 447–474. [5] Bai, Z.D. and Wang, C. (2015) A note on the limiting spectral distribution of a symmetrized auto-cross covariance matrix. Statistics and Probability Letters, 96, 333 – 340. [6] Baik, Jinho and Silverstein, J. W. (2006) Eigenvalues of large sample covariance matrices of spiked population models. J. Multivariate Anal. 97(6), 1382-1408. [7] Burkholder, D.L. (1973) Distribution function inequalities for martingales. Ann. Probab. 1, 19–42. [8] Chung, K.L. (2001) A Course in Probability Theory, 3rd ed. Academic Press, New York. [9] Johnstone, I. (2001) On the distribution of the largest eigenvalue in principal components analysis. Ann. Statist. 29, 295 – 327. [10] Jin, B. S., Wang, C., Bai, Z.D., Nair, K.K. and Harding, M.C. (2014) Limiting spectral distribution of a symmetrized auto-cross covariance matrix. Ann. Appl. Probab. 24, 1199– 1225. 36 [11] Marčenko, V.A. and Pastur, L.A. (1967) Distribution of eigenvalues for some sets of matrices. Mat. Sb. 72, 507 – 536. [12] Wang, C., Jin, B. S., Bai, Z.D., Nair, K.K. and Harding, M.C. (2015) Strong limit of the extreme eigenvalues of a symmetrized auto-cross covariance matrix. Ann. Appl. Probab. 25, 3624 – 3683. 37
10math.ST
Coding-theorem Like Behaviour and Emergence of the Universal Distribution from Resource-bounded Algorithmic Probability Hector Zenil1,2,3∗ Liliana Badillo 4 Santiago Hernández-Orozco 2,3,4 and Francisco Hernández-Quiroz 5 1 Department of Computer Science, University of Oxford, Oxford, U.K. Algorithmic Dynamics Lab, Unit of Computational Medicine, SciLifeLab, Centre for Molecular Medicine, Department of Medicine Solna, Karolinska Institute, Stockholm, Sweden. 3 Algorithmic Nature Group, LABORES, Paris, France. 4 Posgrado en Ciencias e Ingenierı́a de la Computaci,́ Universidad Nacional Autonoma de México (UNAM) 5 Departamento de Matemáticas, Facultad de Ciencias, Universidad Nacional Autónoma de México (UNAM), Ciudad de México, México. arXiv:1711.01711v10 [cs.IT] 23 Mar 2018 2 Abstract Previously referred to as ‘miraculous’ in the scientific literature because of its powerful properties and its wide application as optimal solution to the problem of induction/inference, (approximations to) Algorithmic Probability (AP) and the associated Universal Distribution are (or should be) of the greatest importance in science. Here we investigate the emergence, the rates of emergence and convergence, and the Coding-theorem like behaviour of AP in Turing-subuniversal models of computation. We investigate empirical distributions of computing models in the Chomsky hierarchy. We introduce measures of algorithmic probability and algorithmic complexity based upon resource-bounded computation, in contrast to previously thoroughly investigated distributions produced from the output distribution of Turing machines. This approach allows for numerical approximations to algorithmic (Kolmogorov-Chaitin) complexitybased estimations at each of the levels of a computational hierarchy. We demonstrate that all these estimations are correlated in rank and that they converge both in rank and values as a function of computational power, despite fundamental differences between computational models. In the context of natural processes that operate below the Turing universal level because of finite resources and physical degradation, the investigation of natural biases stemming from algorithmic rules may shed light on the distribution of outcomes. We show that up to 60% of the simplicity/complexity bias in distributions produced even by the weakest of the computational models can be accounted for by Algorithmic Probability in its approximation to the Universal Distribution. Keywords algorithmic coding-theorem; Solomonoff’s induction; information theory; Shannon entropy; lossless compression; Levin’s semi-measure, computable algorithmic complexity; finite-state complexity; transducer complexity; context-free grammar complexity; linear-bounded complexity; time resource-bounded complexity. ∗ Corresponding author: hector.zenil [at] algorithmicnaturelab [dot] org 1 1 Motivation and Significance Algorithmic Probability (AP) and its associated Universal Distribution (UD) predicts the way in which strings distribute when ‘random’ computer programs are run. This algorithmic ‘law’ thus regulates the behaviour of the output distribution of computer programs. The Universal Distribution is the probability distribution that establishes how the output strings from a universal computer running a random computer program distribute. Formally, the Algorithmic Probability of a string AP (s) is defined by: X 1/2|p| APU (s) = (1) p:U (p)=s where the sum is over all halting programs p for which U , a prefix-free universal Turing machine, outputs the string s. A prefix-free universal Turing machine defines a set of valid programs such that the sum is bounded by Kraft’s inequality [17] and not greater than 1 (it is also called a semi-probability measure because some programs will not halt, and thus the sum of the probabilities is never really 1). An invariance theorem establishes that the choice of reference universal Turing machine introduces a vanishing bias as a function of string size: APU (s) ≤ cU,U 0 APU 0 (s) (2) 0 where cU,U 0 is a constant that depends on U and U (think of a compiler size translating in both directions) but is independent of s. Hence the reference U can safely be dropped in the long term. Yet this invariance theorem tells us nothing about the rate of convergence, thus making these numerical experiments more relevant and necessary. The Algorithmic Probability and the Universal Distribution represent the theoretically optimal response to the challenge of induction and inference, according to R. Solomonoff, one of the proponents of algorithmic information theory [33, 34, 35]. More recently, at a panel discussion at the World Science Festival in New York City on Dec 14, 2014, Marvin Minsky, one of the founding fathers of AI, said (own transcription): It seems to me that the most important discovery since Gödel was the discovery by Chaitin, Solomonoff and Kolmogorov of the concept called Algorithmic Probability, which is a fundamental new theory of how to make predictions given a collection of experiences, and this is a beautiful theory, everybody should learn it, but it’s got one problem, that is, that you cannot actually calculate what this theory predicts because it is too hard, it requires an infinite amount of work. However, it should be possible to make practical approximations to the Chaitin, Kolmogorov, Solomonoff theory that would make better 2 predictions than anything we have today. Everybody should learn all about that and spend the rest of their lives working on it. The Universal Distribution has also been characterized as miraculous, because of its performance in inference and prediction [24]. However, the calculation of both AP (s) is not computable. This has meant that for decades after its discovery, few attempts have been made to apply Algorithmic Probability to problems in general science. However, to be more precise AP (s) is upper semi-computable, which means that it can be approximated from below. AP is of fundamental interest to science, as it addresses the most pressing challenges in the areas of complexity, inference and causality, attempting to keep pushing towards better methods for approximating algorithmic probability. A recent new framework and a pipeline of relevant numerical methods have been advanced and have proven successful in many areas, ranging from cognition to graph complexity [9, 12, 7, 40, 41]. There are many properties of m that make it optimal [33, 34, 35, 24]. For example, the same Universal Distribution will work for any problem within a convergent error; it can deal with missing and multidimensional data; the data do not need to be stationary or ergodic; there is no under-fitting or over-fitting because the method is parameter-free and thus the data need not be divided into training and test sets; it is the gold standard for a Bayesian approach in the sense that it updates the distribution in the most efficient and accurate way possible with no assumptions. Several interesting extensions of resource-bounded Universal Search approaches have been introduced in order to make algorithmic probability more useful in practice [28, 29, 20, 14]. Some of these provide some theoretical bounds [1]. Certain of these approaches have explored the effect of relaxing some of the conditions (e.g. universality) on which Levin’s Universal Search is fundamentally based [37] or have introduced domain-specific versions (and thus versions of conditional AP). Here we explore the behaviour of explicitly weaker models of computation– of increasing computational power– in order to investigate asymptotic behaviour and the emergence of the Universal Distribution, and the properties of both the different models with which to approximate it and the actual empirical distributions that such models produce. The so-called Universal Search [15] is based on dovetailing all possible programs and their runtimes such that the fraction of time allocated to program p is 1/2|p| , where |p| is the size of the program (in number of bits). Despite the algorithm’s simplicity and remarkable theoretical properties, a potentially huge constant slowdown factor has kept it from being much used in practice. Some of the approaches to speeding it up have included the introduction of bias and making the search domain specific, which has at the same time limited the power of Algorithmic Probability. There are practical applications of AP that make it very relevant. If one could translate some of the power of Algorithmic Probability to decidable models (thus below Type-0 in the Chomsky hierarchy) without having to deal with the uncomputability of algorithmic complexity and algorithmic probability, it 3 would be effectively possible to trade computing power for predictive power. While trade-offs must exist for this to be possible (full predictability and uncomputability are incompatible), the question of finding a threshold for the coding-theorem to apply would be key to transferring some power through relaxing the computational power of an algorithm. If a smooth trade-off is found before the undecidability border of Turing completeness, it would mean that the partial advantages of Algorithmic Information Theory can be found and partially recovered from simpler models of computation in exchange for accuracy. Such simpler models of computation may model physical processes that are computationally rich but are subject to noise or are bounded by resources. More real-world approaches may then lead to applications such as in the reduction of conformational distributions of protein folding [6, 10] in a framework that may favour or rule out certain paths, thereby helping predict the most likely (algorithmic) final configuration. If the chemical and thermodynamic laws that drive these processes are considered algorithmic in any way, even under random interactions, e.g. molecular Brownian motion, the Universal Distribution may offer insights that may help us quantify the most likely regions if these laws in any sense constitute forms of computation below or at the Turing level that we explore here. This will appear more plausible if one considers the probabilistic bias affecting convergence [19], bearing in mind that we have demonstrated that biological evolution operating in algorithmic space can better explain some phenomenology related to natural selection [11]. 1.1 Uncomputability in complexity Here we explore the middle ground at the boundary and study the interplay between computable and non-computable measures of algorithmic probability connected to algorithmic complexity. Indeed, a deep connection between the algorithmic complexity (or Kolmogorov-Chaitin complexity) of an object s and AP of s was found and formalized by way of the algorithmic Coding theorem. The theorem establishes that the probability of s being produced by a random algorithm is inversely proportional to its algorithmic complexity (up to a constant) [16]: − log AP (s) = K(s) + O(1) (3) Levin proved that the output distribution established by Algorithmic Probability dominates (up to multiplicative scaling) any other distribution produced by algorithmic means as long as the executor is a universal machine, hence giving the distribution its ‘universal’ character (and its name: ‘Universal Distribution’). This so-called Universal Distribution is a signature of Turing-completeness. However, all processes that model or regulate natural phenomena may not necessarily be Turing universal. For example, some models of self-assembly may not be powerful enough to reach Turing-completeness, yet they display similar 4 output distributions to those predicted by the Universal Distribution by way of the algorithmic Coding theorem, with simplicity highly favoured by frequency of production. Noise is another source of power degradation that may preempt universality and therefore the scope and application of algorithmic probability. However, if some subuniversal systems approach the coding-theorem behaviour, these give us great prediction capabilities, and less powerful but computable algorithmic complexity measures. Here we ask whether such distributions can be partially or totally explained by importing the relation established by the coding theorem, and under what conditions non-universal systems can display algorithmic coding-theorem like behaviour. We produce empirical distributions of systems at each of the computing levels of the Chomsky hierarchy, starting from transducers (Type-3) as defined in [18], Context-free grammars (Type-2) as defined in [36], linear-bounded nondeterministic Turing machines (Type-1) as approximations to bounded KolmogorovChaitin complexity and a universal procedure from an enumeration of Turing machines (Type-0) as defined in [5, 21]. We report the results of the experiments and comparisons, showing the gradual coding-theorem-like behaviour at the boundary between decidable and undecidable systems. 2 Methods We will denote by T M (n, k) or simply (n, k) the set of all strings produced by all the Turing machines with n states and k symbols. 2.1 The Chomsky Hierarchy The Chomsky hierarchy is a strict containment hierarchy of classes of formal grammars equivalent to different computational models of increasing computing power. At each of the 4 levels, grammars and automata compute a larger set of possible languages and strings. The four levels, from weaker to stronger computational power, are: Type-3 The most restricted grammars generating the regular languages. They consist of rules with single non-terminal symbols on the left-hand side and strings of terminal symbols followed by at most one non-terminal symbol on the right-hand side. These types of rule are referred to as right linear (but a symmetrical left linear definition works as well). This level is studied by way of finite-state transducers (FST), a generalization of finitestate automata (FSA) that produce an output at every step, generating a relation between input strings and output strings. Though apparently more general, FST-recognized languages are the same as FSA-accepted sets. Hence both can represent this level. We use an enumeration of transducers introduced in [18] where they also proved an invariance theorem, thus demonstrating that the enumeration choice is invariant (up to a constant). 5 Type-2 Grammars that generate the context-free languages. These kinds of grammars are extensively used in linguistics. The languages generated by CFG grammars are exactly the languages that can be recognized by a non-deterministic pushdown automaton. We denote this level by CFG. We generated production rules for 40 000 grammars according to a sound scheme introduced in [36]. Type-1 Grammars that generate the context-sensitive languages. The languages described by these grammars are all languages that can be recognized by a linear-bounded automaton (LBA), a Turing machine whose tape’s length is bounded by a constant times the length of the input. An AP-based variation is introduced here, and we denote it by LBA/AP. Type-0 The least restricted grammar. Generates the languages that can be recognized by Turing machines, also called recursively enumerable languages. This is the level at which Turing-universality is achieved or required. We used previously generated distributions produced and reported in [5, 21] We also explore the consequences of relaxing the halting configuration (e.g. halting state) in models of universal computation (Type-0) when it comes to comparing their output distributions. 2.2 Finite-state complexity Formal language theory and algorithmic complexity had traditionally been at odds as regards the number of states, or the number of transitions in a minimal finite automaton accepting a regular language. In [3] a connection was established by extending the notions of Blum static complexity and encoded function space. The main reason for this lack of connection was that languages are sets of strings, rather than strings used for measures of algorithmic complexity, and a meaningful definition of the complexity of a language was lacking, as well as a definition of Finite-state algorithmic complexity. However, [18] offered a version of algorithmic complexity by replacing Turing machines with finite transducers. The complexity induced is called Finite-state complexity (FSA). Despite the fact that the Universality Theorem (true for Turing machines) is false for finite transducers, rather surprisingly, the invariance theorem holds true for finite-state complexity and, in contrast with descriptional complexities (plain and prefix-free), finite-state complexity is computable. Defined in [18] and analogous to the core concept (Kolmogorov-Chaitin complexity) of Algorithmic Information Theory (AIT)– based on finite transducers instead of Turing machines– finite-state complexity is computable, and there is no a priori upper bound for the number of states used for minimal descriptions of arbitrary strings. Consider a transducer T with the finite set of states Q = {1, . . . , n}. Then the transition function ∆ of T is encoded by a binary string σ (see [18] for details). The transducer T which is encoded by σ is called TσS0 , where S0 is the set of all strings in the form of σ. 6 In [18] it was shown that the set of all transducers could be enumerated by a regular language and that there existed a hierarchy for more general computable encodings. For this experiment we fix S = S0 . As in traditional AIT, where Turing machines are used to describe binary strings, transducers describe strings in the following way: we say that a pair (TσS , p), σ ∈ S, p ∈ B∗ , where B∗ is the set of all finite binary strings, is a description of the string s ∈ B∗ if and only if TσS (p) = s. The size of the description (TσS , p) is defined in the following way ||(TσS , p)|| = |σ| + |p|. Definition 2.1. ([18]) The finite-state complexity of s ∈ B∗ (that we will identify as FSA in the results) with respect to encoding S is defined by CS (s) = min {||(TσS , p)|| : TσS (p) = s}. σ∈S,p∈B∗ An important feature of traditional AIT is the invariance theorem, which states that complexity is optimal up to an additive constant and relies on the existence of a Universal Turing machine (the additive constant is in fact its size). In contrast with AIT, due to the non-existence of a “Universal transducer”, finite-state complexity includes the size of the transducer as part of the encoding length. Nevertheless, the invariance theorem holds true for finite-state complexity. An interesting consequence of the invariance theorem for finite-state complexity is the existence of an upper bound CS (s) ≤ |s| + 8 for all s ∈ B∗ , where 8 is the length of the string σ which encodes the identity transducer. Hence CS is computable. If Sn0 and Sn1 are encodings then CSn0 = f (CSn1 ) for computable function f [18]. An alternative definition of finite-state complexity based on Algorithmic Probability is as follows: Definition 2.2. ([18]) The finite-state complexity (denoted by FSA/AP in the results) of s ∈ B∗ with respect to encoding S is defined by CS (s) = − log Σσ∈S,p∈B∗ {||(TσS , p)|| : TσS (p) = s}. That is, the number of times that a string is accepted by a transducer (in this case, reported in the results, for encodings of size 8 to 22). 2.2.1 Building a finite-state empirical distribution We now define the construction of an empirical distribution using finite-state complexity. We introduce our alternative definition of algorithmic probability using transducers. Definition 2.3. (Finite-state Algorithmic Probability) Let S be the set of encodings of all transducers by a binary string in the form of σ. We then define the algorithmic probability of a string s ∈ B∗ as follows 7 APS (s) = X 2−(|σ|+|p|) . (σ,p):TσS =s For any string s, PS (s) is the algorithmic probability of s, computed for the set of encodings S. In the construction of the empirical distribution for finite-state complexity, we consider the set of strings τ ∈ B∗ such that τ = σ _ p (this is the concatenation of the binary strings σ and p), σ ∈ S, p ∈ B∗ and 8 ≤ |τ | ≤ 22. Hence |τ | = ||(TσS , p)||. Following [23] we define the empirical distribution function DS (n) (i.e. the probability distribution) as Definition 2.4. (Finite-state Distribution, FSA) DS (n) = |{(σ, p) : σ ∈ S, p ∈, 1}∗ , s ∈, 1}∗ , TσS (p) = s}| . |{(σ, p) : σ ∈ S, p ∈ B∗ , ||(TσS , p)|| = n}| In other words, DS (n) considers all strings τ ∈ B∗ of length n and determines whether τ = σ _ p such that σ ∈ S. Then DS (n) computes TσS (p) and counts the number of times that we have TσS (p) = s for every string s described by (σ, p) such that ||(TσS , p)|| = n 1 . We note that in the encoding σ ∈ S, a string ν ∈ B∗ occurring as the output of a transition in TσS contributes 2 · |ν| to the size ||(TσS , p)|| of a description of a string s. The decision to consider strings τ such that 8 ≤ |τ | was made based on the fact that in the encoding of the smallest transducer i.e. the one with the transition function, ∆(1, 0) = ∆(1, 1) = ∆(1, ε) (4) where ε is the empty string, the string ν (which occurs as the output of transitions in TσS ) has length |ν| = 4 and so contributes 2·|ν| = 8 to the size ||(TσS , p)|| of the description of a string s. 2.3 Context-free grammars In [36] Wharton describes an algorithm for a general purpose grammar enumerator which is adaptable to different classes of grammars (i. e. regular, contextfree, etc). We implemented this algorithm in the Wolfram Language with the purpose of enumerating context-free grammars over the terminal vocabulary {0, 1}, which are in Chomsky Normal Form. Before describing the implementation of the algorithm we define the following terminology: • • • • A grammar G is a 4-tuple (N, T, P, X). N is the non-terminal vocabulary. T is the terminal vocabulary. P is the set of productions. 1 Since S is in fact regular we could indeed use an enumeration of S, but in this instance we analyze all binary strings of length n. 8 • X is the start symbol. • V = N ∪ T is the vocabulary of G. • For any grammar G, n, t and p denote the cardinalities of N, T and P respectively. First, we define the structure of a grammar G. Let G be any grammar. Suppose we are given the non-terminal vocabulary N with an arbitrary ordering such that the first non-terminal is the start symbol X. The grammar G has a structure S which consists of a list of n integers. Each integer Si from S is the number of productions having the ith non-terminal on the left-hand side (according to the ordering of N ). Hence the cardinality p of the set of productions P satisfies Σni=0 Si = p. Now, let Γ be a class of grammars over a terminal vocabulary T . By Γc we denote the grammars in Γ with complexity c. We then enumerate Γ by increasing the complexity c. To every complexity class Γc corresponds a set of structure classes which is determined by n and p. Therefore a complexity class Γc is enumerated by enumerating each of its structure classes Γc,S (i. e. every structure S that constitutes Γc ). In addition, we need to define an ordered sequence R which consists of all possible right-hand sides for the production rules. The sequence R is ordered lexicographically (first terminals, then non- terminals) and is defined according to the class of grammars we want to enumerate. For example, suppose we are interested in enumerating the class of Chomsky Normal Form grammars over the terminal vocabulary T = {0, 1} and the non-terminal vocabulary N = {X, Y }, we then set R = {0, 1, XX, XY, Y X, Y Y }. 2.3.1 Implementation Given a complexity c, the algorithm described below (that we implemented in the Wolfram Language running on Mathematica) enumerates all the grammars according to [36] in a structure class Γc,S . 1. The complexity measure c is provided by the following pairing function in p and n: c = ((p + n − 1)(p + n − 2)/2) + n. 2. 3. 4. 5. In other words, given c, we apply the inverse of the above function in order to obtain the values of p and n. This function is implemented by the function pairingInverse[c]. The set of non-terminals N is generated by the function generateSetN[{n, p}]. The ordered sequence R is generated using the set of non-terminals N by the function generateSetR[N ]. The different structure classes Γc,S that correspond to complexity c are generated by the function generateStructureClasses[{n, p}]. All the possible grammars with the structure classes defined in the previous step are then generated. Each grammar has an associated matrix A. This is performed by function generateStructureMatricesA[S, Length[R]]. 9 6. The sequence R is used to generate the rules of the grammars by the function generateGrammars[matricesA, R]. 2.3.2 The CYK algorithm A procedure to decide if a string s is generated by a grammar G in polynomial time was implemented according to the Cocke-Younger-Kasami (CYK) algorithm. The CYK is an efficient worst-case parsing algorithm that operates on grammars in Chomsky normal form (CNF) in O(n3 · |G|), where n is the length of the parsed string and |G| is the size of the CNF grammar G. The algorithm considers every possible substring of the input string and decides s ∈ L(G), where L(G) is the language generated by G. The implementation was adapted from [26]. 2.4 CFG Algorithmic Probability We can now define the Algorithmic Probability of a string according to CFG as follows: Definition 2.5. (Context-free Algorithmic Probability, CFG) We define a CFG empirical distribution in terms of the Universal Distribution as follows: Ds (c) = |{G : s ∈ L(G)}| |{G}| where c = ((p + n − 1)(p + n − 2)/2) + n as defined in 2.3.1, L(G) is the language generated by G and |{G}| denotes the cardinality of the sample set of the grammars considered. For the results here reported |{G}| = 40 000, where all the grammars have a complexity at most c according to a structure class Γc,S [36]. ’Complexity’ here simply means a measure of the grammar size defined as a product of the grammar parameters. 2.5 Linear-bounded complexity In [1] it is shown that the time-bounded Kolmogorov distribution is universal (in the sense of convergence), and they cite the question of an analogue to the algorithmic Coding theorem as an open problem possibly to be tackled by exploiting the universality finding. On the other hand, in [2, 13] it has been shown that the time-bounded algorithmic complexity (being computable) is a Solovay function. These functions are an upper bound of algorithmic complexity (prefix-free version) and they give the same value for almost all strings. In [5, 21] we described a numerical approach to the problem of approximating the algorithmic complexity for short strings. This approach does an exhaustive execution of all deterministic 2-symbol Turing Machines, constructs an output frequency distribution, and then applies the Coding Theorem to approximate the algorithmic complexity of the strings produced. 10 For this experiment we follow the same approach using a less powerful model of computation, namely, linear-bounded automata (LBA). A LBA is basically a single tape Turing Machine which never leaves those cells on which the input was placed [23]. It is well known that the class of languages accepted by LBA is in fact the class of context-sensitive languages [23]. We use the same Turing Machine formalism as [5, 21], which is that of the Busy Beaver introduced by Rado [8], and we use the known values of the Busy Beaver functions. 2.6 Time complexity P T ime or P is the class of Turing Machines that produce an output in polynomial time with respect to the size of its input. When considering empty input strings (in machines with the same number of states), it is easy to see that this class is contained by the class defined by Linear Bounded Automata (LBA): if the number of the transition is bounded by a linear function, so are the number of cells it can visit. But it is important to note that LBA are not time restricted and can use non-deterministic transitions. Whether P T ime is a subset of N SP ACE(O(n)) (LBA) in other cases is an open question. Now, given that Turing Machines can decide context-free grammars in polynomial time (Subsection 2.3.2), P T ime is higher in the hierarchy than Type-2 languages. Within the context of this article, we will represent this class with the set of Turing Machines with 4 states and 2 symbols with no inputs whose execution time is upper-bounded by a fixed constant. We will cap our execution time by 27, 54, 81 and 107 for a total of 44 079 842 304 Turing machines, where 107 is the Busy Beaver value of the set. 2.7 The Chomsky hierarchy bounding execution time The definition of bounded algorithmic complexity is a variation of the unbounded version, as follows: Definition 2.6. (Linear-bounded Algorithmic Complexity, CFG) C t (s) = min{|p| : U (p) = s and U (p) runs in at most t steps} Being bounded by polynomial-sized tapes, the Turing Machines that decide context-sensitive grammars (type-1) can be captured in exponential time by deterministic Turing Machines. Exactly where each class of the Chomsky hierarchy is with respect to the time-based computational complexity classification is related to seminal open problems. For instance, a set equality between the languages recognized by linear-bounded automata and the ones recognized in exponential time would solve the P SP ACE 6= EXP T IM E question. Nevertheless, varying the allowed computation time for the CTM algorithm allows us to capture approximations to the descriptive complexity of an object with fewer computing resources, in a similar fashion to considering each member of the Chomsky hierarchy. 11 2.8 Non-halting models A first comparison between halting Turing machines and non-halting cellular automata was conducted in [39], where it was reported a strong correlation but also some differences, here we further investigate and compare the distributions of halting and non-halting models in connection to the emergence of algorithmic probability and the coding-theorem like behaviour in variations of Turing-complete models. Here, we also consider models with no halting configuration, such as cellular automata (nonH-CA) and Turing machines (nonH-TM) with no halting state as defined in [38], in order to assess whether or not they converge to the Universal Distribution defined over machines with a halting condition. For cellular automata we exhaustively ran all the 256 Elementary Cellular Automata [38] (i.e. closest neighbour and centre cell are taken into consideration) and all 65 536 so-called General Cellular Automata [38] (that is, with 2 neighbours to one side and one to the other, plus the centre cell). For Turing machines, we ran all 4096 (2,2) Turing machines with no halting state, and a sample of 65 536 (same number as CA) Turing machines in (3,2), also with no halting state. 2.9 Consolidating the empirical distributions In order to perform the comparisons among the distributions in each of the Chomsky hierarchy levels, it is necessary to consolidate cases of bias imposed by arbitrary choices of parameter models (e.g. starting from a tape with 0 or 1). This is because, for example, the string 0000 should occur exactly the same number of times as 1111 does and so on, because 0000 and 1111 should have the same algorithmic complexity. If s is the string and f (s) the frequency of production, we thus consolidate the algorithmic probability of s denoted by AP (s) as follows: AP (s) = f (s) + f (r(s)) + f (c(s)) + f (r(c(s)) |g| where |g|, by Burnside’s counting theorem, is the number of non-invariant transformations of s and r(s) is the reversion of s, e.g. 0001 becomes 1000, and c(s) is the negation of s, e.g. 0001 becomes 1110, for all empirical distributions for FSA, CFG, LBA and TM. It is worth noting that c and r do not increase the algorithmic complexity of s except by a very small constant, and thus there is no reason to expect either the model or the complexity of the strings to have been produced or selected to have different algorithmic complexities. More details on the counting method are given in [5]. Notice that this consolidation is not artificially imposed. It would be natural to run each computational model for each possible parameter from the model’s parameter space (e.g. running all TMs once on a ‘blank symbol’ as 0 and then again for the blank as 1) and the distribution would balance itself accordingly. So the application of the consolidating method is only to avoid unnecessary computation. 12 3 Results Here we report the shape and rank of strings in each of the empirical distributions produced by the models of computation, by increasing computational power according to the Chomsky hierarchy and by time computational hierarchy, including comparisons to the performance of classical Shannon entropy (over the bit strings, assuming a uniform distribution) and of lossless compression (using the Compress algorithm), as well as the behaviour of 2 non-halting models of computation. Table 1 reports the number of strings produced by each model. Chomsky Type 3 3 2 (2,0) (2,0) (2,0) 0 0 Computational Model FSA(8-22) FSA/AP(8-22) CFG(40K) LBA(27) LBA(54) LBA(81) LBA 107 = TM(4,2) TM(5,2) No. Strings 294 1038 496 847 1225 1286 1302 8190 Table 1: Strings of at most 12 bits produced at each level from a total of 8190 possible strings (all generated by (5, 2) except 2 [21]). LBA were simulated by halting time cutoffs, which is strictly less powerful than the TM model (Type-0) but strictly greater than CFG (Type-2). Each LBA is followed by its runtime cutoff. The cutoff value of 107 steps is the Busy Beaver for (4,2), and given that LBA are emulated by TM in (4,2) by letting them run up to 107, this means that they have exhausted the full TM space (4,2) and thus are equivalent. The empirical distribution for CFG was obtained by producing 40 000 grammars and checking whether each of the 1302 strings occurring in LBA 107 could be generated by any of these grammars. A Chomsky Type (2,0) means that the model is strictly between Type-0 and Type-2, thus representing Type-1. 3.1 Finite-state complexity The experiment consists of a thorough analysis of all strings τ ∈, 1}∗ that satisfy 8 ≤ |τ | ≤ 22. If the string τ satisfies τ = σ _ p (for some σ ∈ S) then we compute TσS (p) and generate a set of output strings. Then a frequency distribution is constructed from the set of output strings. On the other hand, we compute the finite-state complexity for strings s such that 0 ≤ s ≤ 8 (this is an arbitrary decision). 13 A B C Figure 1: Correlation plots (A) between the 2 models of Algorithmic Probability for FSA using the same enumeration (B) between FSA (both plain and FSA/APbased) and (C) both FSA and FSA/AP separately against Turing machines. 14 A B C D Figure 2: Correlation plots of the four Chomsky hierarchy types. (A) FSA v TM. (B) how FSA approximates TM before converging to a Kendall ranking correlation value below 0.9 (with LBA able to reach 1, see Fig. 5). (C) CFG v TM and (D) LBA v TM. 15 3.1.1 distributions DS (n) produced for 8 ≤ n ≤ 22 The results given in Table 2 indicate how many strings τ satisfy τ = σ _ p (such that σ encodes the transition table of some transducer TσS and p is an input for it) per string length. Running FSA is very fast; there is no halting problem and all stop very quickly. However, while FSA and CFG preserve some of the ranking of superior computational models and accelerate the appearance of ‘jumpers’ (long strings with low complexity), these weak models of computation do not generate the strings with the highest algorithmic complexity that appear in the tail of the distributions of more powerful models, as shown in Fig. 4. Size 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 Strings 256 512 1024 2048 4096 8192 16384 32768 65536 131072 262144 524288 1048576 2097152 4194304 Transducers 1 2 6 12 34 68 156 312 677 1354 2814 5628 11474 22948 46332 Table 2: Transducers per string length. For example, of 28 strings in DS (8), there is only one binary string of length 8 that encodes a transducer, which is the transducer with the transition function (4) (we refer to it as the smallest transducer). In the case of DS (9), we found that out of 29 strings, only two encode the smallest transducer with “0” and “1” as input. Again, the only string produced by this distribution is the empty string ε. DS (16) is the first distribution in which one of the strings encodes a transducer with two states. The finite-state complexity CS (s) of the strings produced by DS (16) shows 8 ≤ CS (s) ≤ 12 (see Table 14 and Table 15 in the Supplementary Material). DS (18) consists of 2814 transducers (see Table 3 and Table 4 in the Supplementary Material). The longest string produced is of length 6. The rest of the tables are in the Supplementary Material. The rest of the tables are in the Supplementary Material. 16 String ε 00 11 000 111 0 1 0000 1111 00000 11111 01 10 000000 0101 1010 111111 Probability 0.82800 0.02701 0.02701 0.01990 0.01990 0.01848 0.01848 0.01279 0.01279 0.00426 0.00426 0.00213 0.00213 0.00071 0.00071 0.00071 0.00071 Table 3: Example of a probability distribution for DS (18). Complexity 14 12 13 11 10 9 Frequency 1024 640 512 320 212 106 Table 4: Frequency of Finite-state complexity for strings produced by DS (18). 3.2 Computing finite-state complexity We performed another experiment in order to further analyze the characteristics of the finite-state complexity of all strings of length n. We summarize the results we obtained for computing finite-state complexity for each string s of length 0 ≤ n ≤ 8 in Table 5, while Table 6 shows the strings σ that encode transducers such that TσS (p) = s. 3.3 Context-free grammar distribution We created production rules for 298 233 grammars with up to 26 non-terminal symbols and used the first 40 000 of them to check how many of a set of 155 strings were produced by how many grammars for which we also had distribution numbers at all the other levelsa; FSA, LBA and TM. Table 7 shows the top 20 strings produced. 17 Complexity 4 7 8 9 10 11 12 13 14 15 16 Frequency 1 2 2 4 4 10 22 32 56 126 252 Table 5: Frequency of CS (s) for strings s of length 0 ≤ n ≤ 8. Transducer 0000 000100 0001010110 00010110 0001011100 0001011110 000110 00011100 0001110100 0001110110 0001111100 01000110 Frequency 1 8 1 4 1 1 8 4 1 1 1 480 Table 6: Frequency of the strings σ that encode transducers such that TσS (p) = s. 3.4 3.4.1 Emergence of the Universal Distribution Time-bounded Emergence Fig. 5 shows how LBA asymptotically approximate the Universal Distribution. As the results demonstrate (supported by Fig. 5), by varying the allowed execution time for the space of Turing Machines we can approximate the CTM distribution corresponding to each level of the Chomsky hierarchy. For instance, regular languages (Type-3 grammars) can be decoded in linear time, given that each transition and state in a finite automaton can be encoded by a corresponding state and transition in a Turing Machine. Context-free grammars (Type-2) can be decoded in polynomial time with parsing algorithms such as CYK (Subsection 2.3.2). 18 String 0 00 000 0000 00000 000000 00001 0001 00010 00011 001 0010 00100 00101 0011 00110 00111 01 010 0100 Frequency 5131 5206 5536 5480 5508 5508 2818 2810 2754 2764 2812 2692 2750 2744 2736 2730 2736 2688 2748 2742 Table 7: Top 20 most frequent strings generated by context-free grammars (CFG). Assuming that the algorithmic Coding theorem held, we can then transform the frequency into a computable algorithmic complexity estimation by CFG for purposes of comparisoh. 3.4.2 Rate of convergence of the distributions One opportunity afforded by this analysis is the assessment of the way in which other methods distribute strings by (statistical) randomness, such as Shannon entropy and other means of approximation of algorithmic complexity, such as lossless compression algorithms, in particular one of the most popular such methods, based on LZW (Compress). We can then compare these 2 methods in relation to estimations of a Universal Distribution produced by T M (4, 2). The results (see Fig. 6) of both entropy and Compression conform with the theoretical expectation. Entropy correlates best at the first level of the Chomsky hierarchy, that of FSA, stressing that the algorithmic discriminatory power of entropy to tell apart randomness from pseudo-randomness is limited to statistical regularities of the kind that regular languages would capture. Lossless compression, however, at least as assessed by one of the most popular methods underlying other popular lossless compression formats, outperformed Shannon entropy, but not by much, and it was at best most closely correlated to the output distribution generated by CFG. This does not come as a surprise, given that popular implementations of lossless compression are a variation of Shannon entropy generalized to blocks (variable-width window) that capture repetitions, 19 Figure 3: Comparison of empirical distributions of varying computational power, with TM in log plot (thus the only exponential) followed by LBA and then FSA and CFG when taking their fitting slope coefficients. Running CFG is very expensive because for every string we need to test whether it could be generated by all the grammars produced, which in our case means 1302 × 40 000, and while the CYK algorithm runs in polynomial time, the combinatorial explosion renders the AP-based model by CFG intractable. often followed by a remapping to use shorter codes for values with higher probabilities (dictionary encoding, Huffman coding)–hence effectively a basic grammar based on a simple rewriting rule system. We also found that, while non-halting models approximate the Universal Distribution, they start diverging from TM and remain correlated to LBA with lower runtimes, despite increasing the number of states. This may be expected from the over-representation of strings that non-halting machines would otherwise skip (defined as produced after halting for machines with halting configurations). 4 4.1 Some open questions Tighter bounds and suprema values We have provided upper and lower bounds for each model of computation, but current intervals seem to overlap for some measures of correlation. One question concerns the exact boundaries, especially how much closer to the Universal Distribution each supremum for each model of computation can take us, In other words, can they find the tighter bounds for the intervals at each level? 4.2 Computable error corrections There are some very interesting open questions to explore in the future. For example, whether computable corrections can be made to subuniversal distributions, e.g. as calculated by context-free grammars, in order to correct for trivial (and also computable) biases such as string length. Indeed, while CFG produced an interesting distribution closer to that of LBA and better than FSA, 20 Figure 4: Each model misses the most random strings according to the next most powerful model, with the greatest discrepancy against TM(5,2). For example, FSA and FSA/AP produce the same strings but they assign different values to each string. Here it is shown that FSA/AP outperforms FSA alone by more accurately identifying a greater number of random strings and therefore missing fewer of them compared to the next level (CFG). Moreover, FSA only produces 12 different complexity values among all possible strings produced, while FSA/AP produces 34. The plot also shows that CFG is more powerful than LBA with halting runtime 27 and similar to LBA 54. The progression of LBA towards the full power of TM is also noticeable. there is, for example, an over-representation of trivial long strings which can easily be corrected. The suspicion is that while one can apply these corrections and increase the speed of convergence to TM, the interesting cases are non-computable. A theoretical framework and a numerical analysis would be interesting to develop. 4.3 Sensitivity to choice of enumeration A question equivalent to that of the choice of programming language of the reference universal Turing machine is the question of the stability of the chosen enumerations for the different models of computation, both at the same and at different computational power levels. In [18] it is shown that FSA complexity is invariant to enumeration choice, and an open question is whether this happens at other levels of computational power between FSA and TM, e.g. CFG for some computational models of generative grammars (e.g. the one chosen in this paper). A striking outcome of the results reported here is that not only does the increase in computing power produce better approximations to the distribution produced by Turing machines, but that completely different models of computation—not only in terms of language description but also in terms of computational power— produce similar distributions, thus suggesting a greater independence of the choice of arbitrary parameters, such as enumeration 21 A B C D Figure 5: Trading computational power for accuracy. (A) Number of halting machines per maximum imposed runtime (B) Number of strings produced among all possible, showing that all strings up to length 9 were generated in (4, 2) (C) Output frequency distribution per running time produced in (4, 2) from highest to lowest frequency (D) Smooth density plot maximizing the differences of the output distributions at each runtime approximating the full power of Turing machines in (4, 2) versus strictly lower computational power when bounding the runtime. or model. In [39] we found such convergence in values and ranking in experiments comparing Turing machines, cellular automata and Post tag systems, and moreover these results suggested some sort of ‘natural behaviour’ defined as behaviour that is not artificially introduced with the purpose of producing a different-looking initial distribution (before converging per the invariance theorem for systems with such a property). Thus in [4] we proposed a measure of ‘algorithmicity’ in the sense of the Universal Distribution, quantifying how close or removed a method producing a distribution is from other approximations, in particular that of one of the most standard Turing machine models, the one used for the Busy Beaver [8], that we have shown is not a special case– several other variations of this model and completely different models of computation produce similar output distributions [42, 22, 4], including the results reported in this paper. However, one other open question would entail enumerating systems in different ways and numerically quantifying how many of them are convergent or divergent, to what degree the divergent diverge and for how long and under what conditions, and 22 Figure 6: Coding theorem-like behaviour and emergence of the Universal Distribution. Correlation in rank (distributions were sorted in terms of each other) of empirical output distributions as compared to the output distribution of T M (5, 2) as introduced and thoroughly investigated in [21]. A progression towards greater correlation is noticed as a function of increasing computational power. Bold black labels are placed at their Chomsky level and gray labels are placed within the highest correlated level. Shannon entropy and lossless compression (Compress) distribute values below or at about the first 2 Chomsky types, as expected. It is not surprising to see the LBA with runtime 107 further deviate in ranking, because LBA after 27 steps produced the highest frequency strings, which are expected to converge faster. Eventually LBA 107 (which is none other than TM(4,2)) will converge to TM(5,2). An empirical bound of non-halting models seems to be low LBA even when increasing the number of states (or symbols for CA). whether the convergent dominate. 4.4 Complexity of missing strings from non-halting models We have shown that for halting models of computation, decreasing the power of the computational model has the effect of missing some of the most algorithmically random strings that are produced in the model with the next highest computational power. As we have also seen, non-halting models seem to converge to lower runtime distributions (LBA, which even when they are highly 23 correlated to TMs, do not appear to asymptotically approach TMs but to remain closer to LBA. An interesting question to explore concerns the complexity of the strings missed by non-halting machines or which strings are under or over-represented. As opposed to the strings missed by halting machines below Turing universality, models of computation without halting configurations but still capable of Turing universality (defined on non-halting models) may miss different kinds of strings than those models with halting configurations but strictly weaker computational power (e.g. LBA). Do they miss more or fewer random strings as compared to halting models? 5 Conclusions It is interesting to explore and seek algorithmic Coding-theorem like behaviour in subuniversal systems to better understand the emergence, and rate of emergence, of properties related to algorithmic probability, in particular those established by the algorithmic Coding-theorem from the Universal Distribution–and thus of algorithmic complexity–for types of computational systems of different computational power, in particular below Turing universality. The results reported show that the closer a system is to Turing-completeness, the closer its output distribution is to the empirical distribution produced by a complete set of small Turing machines, and that finite approximations of algorithmic complexity from subuniversal systems [27] produce a gradual emergence of properties related to the Universal Distribution. The results also show improvements over current major tools for approximating algorithmic complexity, such as lossless compression algorithms. To our knowledge, it has not been possible to date to quantify or compare the performance of lossless compression algorithms as estimations of algorithmic complexity, as there was no other standard or alternative. The construction of empirical distributions does provide the means and constitutes an approach to evaluate performance in what we have named an algorithmicity test [39]. Compared to entropic and the most popular lossless compression algorithms, approximations to algorithmic complexity from finite approximations of Algorithmic Probability–even over weak models of computation–constitutes a major improvement over strings that are otherwise assigned a greater randomness content by traditional methods such as common lossless compression, Shannon entropy and equivalent statistical formulations. This algorithmic behavioural exploration is particularly relevant because processes in nature may be weaker than the full power of Turing machines, due to access to finite resources and because of physical degradation. For example, some biological and genetic processes such as DNA translation and RNA transcription can only decode and produce messages and sequences by reading in a single-strand direction and thus may be modelled by Context-free grammars (CFG). The results here suggest that, for example, RNA transcription would not sample uniformly the whole aminoacid space, but would favour the production of simple protein primary structures. In this way the approximation 24 to the Universal Distribution is not only descriptive but does play an active role in active (e.g. biological) processes. Given these types of questions, we also explored non-halting models, as natural processes may not have definite and distinguishable halting states, unlike, e.g., RNA transcription that, exceptionally, has a clear halting condition represented by the RNA polymerase enzyme reading the stop codons. We have shown that up to 60% of the simplicity vs complexity bias produced even by the weakest of the computational models (FST/FSA) in the Chomsky hierarchy (Type-3)–disregarding halting configuration–can be explained by or accounted for by Algorithmic Probability in the model output approximation to the Universal Distribution as tested by empirical distributions compared to the empirical distributions produced by large sets of small Turing machines. On the one hand, while producing empirical distributions from FSA is computationally cheap and produces a large set of strings mostly distributed in the shape of the Universal Distribution, it produces very few strings with different probability/complexity values and thus provides very little discriminatory power. On the other hand, producing strings at the level of CFG (Type-2) with grammars turned out to be computationally very expensive, despite the existence of a polynomial algorithm that can verify string production. However, when a set of strings needs to be checked against a set of grammars, the computational task becomes exponential by a combinatorial argument. Type-2, however, can also be simulated by not only limiting the tape length in LBAs but also forcing them to operate in only one tape direction, which would lead to a simulation similar to that of LBAs. Finally, we have demonstrated that LBAs (Type-1) are both the most direct and the fastest computational model approximating the empirical distributions produced by Turing machines (Type-0) that we generated. Acknowledgements L.B and F.H.-Q. were supported by the grant No. 221341 from SEP-Conacyt CB-2013-01, and H.Z. by grant No. 2015-05299 from the Swedish Research Council (Vetenskapsrådet). References [1] L. Antunes y L. Fortnow, Time-Bounded Universal Distributions, Electronic Colloquium on Computational Complexity, Report No. 144, 2005. [2] L. Bienvenu and R. Downey. Kolmogorov complexity and Solovay functions. In STACS, volume 3 of LIPIcs, pages 147–158. Schloss Dagstuhl- LeibnizZentrum fuer Informatik, 2009. [3] C. Campeanu, K. Culik II, K. Salomaa, S. Yu, State complexity of basic operations on finite languages. In: O. Boldt, H. Jürgensen, (eds.) WIA 1999, LNCS, vol. 2214, pp. 60–70. Springer, Heidelberg, 2001. 25 [4] J.-P. Delahaye, H. Zenil, Towards a stable definition of Kolmogorov-Chaitin complexity, arXiv:0804.3459 [cs.IT], 2008. [5] J.-P. Delahaye and H. Zenil, Numerical Evaluation of the Complexity of Short Strings: A Glance Into the Innermost Structure of Algorithmic Randomness, Applied Mathematics and Computation 219, pp. 63–77, 2012. [6] K. Dingle, S. Schaper, and A.A. Louis, The structure of the genotypephenotype map strongly constrains the evolution of non-coding RNA, Interface Focus 5: 20150053, 2015. [7] N. Gauvrit, F. Soler-Toscano, H. Zenil, Natural Scene Statistics Mediate the Perception of Image Complexity, Visual Cognition, vol. 22:8, pp. 1084– 1091, 2014. [8] T. Rado, “On non-computable functions” Bell System Technical Journal 41:3, 877–884, 1962. [9] N. Gauvrit, H. Zenil, F. Soler-Toscano, J.-P. Delahaye, P. Brugger, Human Behavioral Complexity Peaks at Age 25, PLoS Comput Biol 13(4): e1005408, 2017. [10] S.F. Greenbury, I.G. Johnston, A.A. Louis, S.E. Ahnert, J. R., A tractable genotype-phenotype map for the self-assembly of protein quaternary structure, Soc. Interface 11, 20140249, 2014. [11] S. Hernández-Orozco, H. Zenil, N.A. Kiani, Algorithmically probable mutations reproduce aspects of evolution such as convergence rate, genetic memory, modularity, diversity explosions, and mass extinction, arXiv:1709.00268 [cs.NE] [12] V. Kempe, N. Gauvrit, D. Forsyth, Structure emerges faster during cultural transmission in children than in adults, Cognition, 136, 247–254, 2014. [13] R. Hölzl, T. Kräling, W. Merkle, Time-bounded Kolmogorov complexity and Solovay functions, Theory Comput. Syst. 52:1, 80–94, 2013. [14] M. Hutter, A Theory of Universal Artificial Intelligence based on Algorithmic Complexity, Springer, 2000. [15] L.A Levin, Universal sequential search problems. Problems of Information Transmission, 9:265–266, 1973. [16] L.A. Levin. Laws of information conservation (non-growth) and aspects of the foundation of probability theory, Problems Information Transmission, 10(3):206–210, 1974. [17] L.G. Kraft, A device for quantizing, grouping, and coding amplitude modulated pulses, Cambridge, MA: MS Thesis, Electrical Engineering Department, Massachusetts Institute of Technology, 1949. 26 [18] Calude, C. S., Salomaa, K., & Roblot, T. K. (2011). Finite-state complexity. Theoretical Computer Science, 412(41), 5668-5677. [19] S. Schaper and A.A. Louis, The arrival of the frequent: how bias in genotype-phenotype maps can steer populations to local optima PLoS ONE 9(2): e86635, 2014. [20] B.R. Steunebrink, J. Schmidhuber, Towards an Actual Gödel Machine Implementation. In P. Wang, B. Goertzel, eds., Theoretical Foundations of Artificial General Intelligence, Springer, 2012. [21] Soler-Toscano, F., Zenil, H., Delahaye, J. P. & Gauvrit, N. (2014). Calculating Kolmogorov complexity from the output frequency distributions of small Turing machines. PloS one, 9(5), e96223. [22] F. Soler-Toscano, H. Zenil, J.-P. Delahaye and N. Gauvrit , Correspondence and Independence of Numerical Evaluations of Algorithmic Information Measures, Computability, vol. 2, no. 2, pp 125-140, 2013. [23] J.E. Hopcroft, & J.D. Ullman, Formal languages and their relation to automata, 1969. [24] W. Kirchherr and M. Li and P. Vitányi, The Miraculous Universal Distribution, Mathematical Intelligencer, 19, 7–15, 1997. [25] B.Y. Peled, V.K. Mishra, A.Y. Carmi, Computing by nowhere increasing complexity, arXiv:1710.01654 [cs.IT] [26] J. Rangel-Mondragon, Recognition and Parsing of Context-Free, http:// library.wolfram.com/infocenter/MathSource/3128/ Accessed on Aug 15, 2017. [27] F. Soler-Toscano, H. Zenil, A Computable Measure of Algorithmic Probability by Finite Approximations with an Application to Integer Sequences, Complexity (accepted). [28] J. Schmidhuber, Optimal Ordered Problem Solver, Machine Learning, 54, 211-254, 2004.. [29] J. Schmidhuber, V. Zhumatiy, M. Gagliolo, Bias-Optimal Incremental Learning of Control Sequences for Virtual Robots. In Groen, et al. (eds) Proceedings of the 8th conference on Intelligent Autonomous Systems, IAS8, Amsterdam, The Netherlands, pp. 658–665, 2004. [30] F. Soler-Toscano, H. Zenil, J.-P. Delahaye, and N. Gauvrit, Small Turing Machines with Halting State: Enumeration and Running on a Blank Tape. http://demonstrations.wolfram.com/ SmallTuringMachinesWithHaltingStateEnumerationAndRunningOnAB/. Wolfram Demonstrations Project. Published: January 3, 2013. 27 [31] R.J. Solomonoff. A formal theory of inductive inference: Parts 1 and 2. Information and Control, 7:1–22 and 224–254, 1964. [32] M. Li, and P. Vitányi, An Introduction to Kolmogorov Complexity and Its Applications, 3rd ed, Springer, N.Y., 2008. [33] R.J. Solomonoff, Complexity-Based Induction Systems: Comparisons and Convergence Theorems, IEEE Trans. on Information Theory, vol 24, No. 4, pp. 422–432, 1978. [34] R.J. Solomonoff, The Application of Algorithmic Probability to Problems in Artificial Intelligence, in L.N. Kanal and J.F. Lemmer (eds.), Uncertainty in Artificial Intelligence, pp. 473–491, Elsevier, 1986. [35] Solomonoff, R.J. A System for Incremental Learning Based on Algorithmic Probability, Proceedings of the Sixth Israeli Conference on Artificial Intelligence, Computer Vision and Pattern Recognition, Dec. 1989, pp. 515–527. [36] R.M. Wharton, Grammar enumeration and inference, Information and Control, 33(3), 253–272, 1977. [37] M. Wiering and J. Schmidhuber. Solving, POMDPs using Levin search and EIRA, In Proceedings of the International Conference on Machine Learning (ICML), pages 534–542, 1996. [38] S. Wolfram, A New Kind of Science, Wolfram Media, Champaign, IL., 2002. [39] H. Zenil and J-P. Delahaye, On the Algorithmic Nature of the World, In G. Dodig-Crnkovic and M. Burgin (eds), Information and Computation, World Scientific Publishing Company, 2010. [40] H. Zenil, F. Soler-Toscano, K. Dingle and A. Louis, Correlation of Automorphism Group Size and Topological Properties with Program-size Complexity Evaluations of Graphs and Complex Networks, Physica A: Statistical Mechanics and its Applications, vol. 404, pp. 341–358, 2014. [41] H. Zenil, N.A. Kiani and J. Tegnér, Methods of Information Theory and Algorithmic Complexity for Network Biology, Seminars in Cell and Developmental Biology, vol. 51, pp. 32-43, 2016. [42] H. Zenil, F. Soler-Toscano, J.-P. Delahaye and N. Gauvrit, TwoDimensional Kolmogorov Complexity and Validation of the Coding Theorem Method by Compressibility, PeerJ Computer Science, 1:e23, 2015. 28 Supplementary Material 5.0.1 Empirical distributions < DS (19) for FSA For DS (10) we have it that there are six strings that encode a transducer. In fact, two of them are different from the smallest transducer. However, as in the previous cases, the only string produced is the empty string ε. DS (11) consists of 12 strings with inputs of length one and three but the output of these transducers is ε. DS (12) contains 34 transducers with inputs whose lengths range from two to four. The output distribution still contains only the string ε. DS (13) is a more interesting distribution. It consists of 68 strings whose input has lengths 1, 3 and 5. Table 8 shows the probability distribution of the strings produced by this distribution. The finite-state complexity of the strings that comprise DS (13) is summarized in Table 9. String ε 0 1 Probability 0.94118 0.02941 0.02941 Table 8: Probability distribution of DS (13). Complexity 9 7 8 Frequency 32 20 16 Table 9: Frequency of finite-state complexity for strings produced by DS (13). DS (14) is a richer distribution than the previous one since it contains 156 strings that encode different transducers. Table 10 shows the different strings produced by this distribution. String ε 0 1 00 11 Probability 0.92308 0.02564 0.02564 0.01282 0.01282 Table 10: Probability distribution of DS (14). We note the following facts: • The length of the longest string produced is two. 29 • The string ε remains the one with the highest probability. • The finite-state complexity of the strings produced ranges from 7 to 10 (see Table 11). • DS (14) produces two strings of length 2 out of 22 , that is, “00” and “11”. Complexity 10 8 9 7 Frequency 64 40 32 20 Table 11: Frequency of Finite-state complexity for strings produced by DS (14). DS (15) is quite similar to DS (14) in terms of the strings it produces (see Table 10 and Table 13). String ε 0 1 00 11 000 111 Probability 0.88462 0.03205 0.03205 0.01923 0.01923 0.00641 0.00641 Table 12: Probability distribution of DS (15). Complexity 11 9 10 8 Frequency 128 80 64 40 Table 13: Frequency of Finite-state complexity for strings produced by DS (15). DS (17) shows an even more diverse set of strings produced (see Table 16). We have the following interesting facts, • The longest string produced is of length 5. • For the first time, a distribution produces all strings of length 2. 5.0.2 DS (19),DS (20),DS (21) and DS (22) Here are the strings that comprise each one of these distributions. 30 String ε 0 00 1 11 000 111 0000 1111 Probability 0.87592 0.02363 0.02363 0.02363 0.02363 0.01182 0.01182 0.00295 0.00295 Table 14: Probability distribution of DS (16). Complexity 12 10 11 9 8 Frequency 256 160 128 80 53 Table 15: Frequency of Finite-state complexity for strings produced by DS (16). String ε 0 1 00 11 000 111 0000 1111 00000 01 10 11111 Probability 0.83752 0.02806 0.02806 0.02511 0.02511 0.01773 0.01773 0.00739 0.00739 0.00148 0.00148 0.00148 0.00148 Table 16: Probability distribution of DS (17). String ε 0000 1111 000 111 Probability 0.78630 0.02109 0.02109 0.02066 0.02066 31 00 11 00000 11111 0 1 000000 111111 0000000 1111111 01 10 0101 1010 00000000 11111111 010101 101010 01010101 10101010 001 010 011 100 101 110 000000000 0000000000 00001 00010 00011 00100 00101 00110 00111 01000 01001 01010 0101010101 01011 01100 01101 01110 01111 10000 10001 0.01682 0.01682 0.01665 0.01665 0.01063 0.01063 0.00959 0.00959 0.00331 0.00331 0.00166 0.00166 0.00139 0.00139 0.00122 0.00122 0.00105 0.00105 0.00044 0.00044 0.00026 0.00026 0.00026 0.00026 0.00026 0.00026 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 32 10010 0.00009 10011 0.00009 10100 0.00009 10101 0.00009 1010101010 0.00009 10110 0.00009 10111 0.00009 11000 0.00009 11001 0.00009 11010 0.00009 11011 0.00009 11100 0.00009 11101 0.00009 11110 0.00009 111111111 0.00009 1111111111 0.00009 Table 22: Probability distribution of DS (21). String ε 0000 1111 000 111 00000 11111 000000 111111 00 11 0 1 0000000 1111111 00000000 11111111 0101 1010 01 10 010101 101010 33 Probability 0.78313 0.02180 0.02180 0.01744 0.01744 0.01727 0.01727 0.01442 0.01442 0.01416 0.01416 0.00622 0.00622 0.00587 0.00587 0.00276 0.00276 0.00160 0.00160 0.00138 0.00138 0.00125 0.00125 01010101 10101010 000000000 111111111 0000000000 1111111111 0101010101 1010101010 001 010 011 100 101 110 0001 0010 001001 0011 0100 010010 0110 011011 0111 1000 1001 100100 1011 101101 1100 1101 110110 1110 000000000000 000001 000010 000011 000100 000101 000110 000111 001000 001010 001011 001100 001101 001110 34 0.00073 0.00073 0.00043 0.00043 0.00030 0.00030 0.00026 0.00026 0.00017 0.00017 0.00017 0.00017 0.00017 0.00017 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00009 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 001111 010000 010001 010011 010100 010101010101 010110 010111 011000 011001 011010 011100 011101 011110 011111 100000 100001 100010 100011 100101 100110 100111 101000 101001 101010101010 101011 101100 101110 101111 110000 110001 110010 110011 110100 110101 110111 111000 111001 111010 111011 111100 111101 111110 111111111111 35 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 0.00004 Table 24: Probability distribution of DS (22). 5.0.3 Code in Python for finite-state complexity The program distributionTransducers.py is used to analyze all strings τ of length |τ | = n to determine whether τ satisfies τ = σ _ p (for σ ∈ S), and if so then the program computes TσS (p). This program generates a set of output strings (result-experiment-distribution.csv) from which we can construct an output frequency distribution. Example of execution: • python distributionTransducers.py 8 10 analyzes all strings of length 8 up to length 10. • python distributionTransducers.py 8 8 analyzes all strings of length 8. The file result-experiment-distribution.csv contains the following columns: • • • • • • • string: this corresponds to the strings τ discussed above. valid-encoding: takes value 1 in case τ = σ _ p and 0 otherwise. sigma: corresponds to string σ such that τ = σ _ p. string-p: corresponds to string p such that TσS (p) = s. num-states: number of states of transducer TσS (p). output: corresponds to string s such that TσS (p) = s. output-complexity: finite-state complexity of output string x. The program computeComplexityStrings.py computes the finite-state complexity for all strings of length n up to length m (this is the implementation of the algorithm described in [18]). This program generates the file resultcomplexity.csv which contains the following columns: • s: the string that the program is calculating the finite-state complexity for. • complexity: finite-state complexity of string s. • sigma: string σ such that TσS (p) = s. • string-p, string p such that TσS (p) = s. 36 Complexity 13 11 12 10 9 Frequency 512 320 256 160 106 Table 17: Frequency of finite-state complexity for strings produced by DS (17). String ε 000 111 00 11 0 1 0000 1111 00000 11111 000000 111111 01 10 0101 1010 0000000 001 010 010101 011 100 101 101010 110 1111111 Probability 0.80597 0.02345 0.02345 0.02274 0.02274 0.01812 0.01812 0.01706 0.01706 0.00817 0.00817 0.00284 0.00284 0.00178 0.00178 0.00107 0.00107 0.00036 0.00036 0.00036 0.00036 0.00036 0.00036 0.00036 0.00036 0.00036 0.00036 Table 18: Probability distribution of DS (19). ‘Jumpers’ as defined in [5] and [21] are apparent, those simple strings of relatively much greater length that climb the complexity ladder. 37 Complexity 15 13 14 12 11 10 Frequency 2048 1280 1024 640 424 212 Table 19: Frequency of Finite-state complexity for strings produced by DS (19). 38 String ε 0000 1111 00 000 11 111 0 00000 1 11111 000000 111111 01 10 0101 1010 0000000 1111111 010101 101010 00000000 11111111 0001 0010 0011 0100 01010101 0110 0111 1000 1001 10101010 1011 1100 1101 1110 Probability 0.80024 0.02144 0.02144 0.02092 0.02092 0.02092 0.02092 0.01185 0.01185 0.01185 0.01185 0.00593 0.00593 0.00174 0.00174 0.00157 0.00157 0.00139 0.00139 0.00070 0.00070 0.00035 0.00035 0.00017 0.00017 0.00017 0.00017 0.00017 0.00017 0.00017 0.00017 0.00017 0.00017 0.00017 0.00017 0.00017 0.00017 Table 20: Probability distribution of DS (20). 39 Complexity 16 14 15 13 12 11 10 Frequency 4096 2560 2048 1280 848 424 218 Table 21: Frequency of finite-state complexity for strings produced by DS (20). Complexity 17 15 16 14 13 12 11 Frequency 8192 5120 4096 2560 1696 848 436 Table 23: Frequency of Finite-state complexity for strings produced by DS (21). Complexity 18 16 17 15 14 13 12 11 Frequency 16384 10240 8192 5120 3392 1696 872 436 Table 25: Frequency of Finite-state complexity for strings produced by DS (22). 40
2cs.AI
Analyzing Large Biological Datasets with an Improved Algorithm for MIC Shuliang Wang, Yiping Zhao School of Software, Beijing Institute of Technology, Beijing 100081, China Abstract: A computational framework utilizes the traditional similarity measures for mining the significant relationships in biological annotations is recently proposed by Tatiana V. Karpinets et al. [2]. In this paper, an improved approximation algorithm for MIC (maximal information coefficient) named IAMIC is suggested to perfect this framework for discovering the hidden regularities between biological annotations. Further, IAMIC is the enhanced algorithm for approximating a novel similarity coefficient MIC with generality and equitability, which makes it more appropriate for data exploration. Here it is shown that IAMIC is also applicable for identify the associations between biological annotations. Keywords: IAMIC, biological annotations, data exploration 1. INTRODUCTION With the amount of biological data dramatic increases, more and more research community will be certainly in the process of working on annotations [1]. Many biological databases is currently collecting the annotation information of different biological objects at amazing rate, which requires appropriate analysis tools to put in place. Tatiana V. Karpinets et al. [2] recently proposed a novel approach for mining modular structure, relationships and regularities in lager biological datasets. This method applied the similarity measures, such as Pearson correlation coefficient to discover the hidden associations between pairwise annotations. Here we introduce a new measure to identify the relationship between pairs of annotations. The new measure is called IAMIC, the improved approximation algorithm for maximal information coefficient, which based on the algorithm of Reshef et al. [3], and it is an enhanced solution for detecting the relationships between variables in large data sets with better tradeoffs between equitability and time. In our work, we present a comparison between the proposed algorithm and the traditional similarity measures to uncover novel regularities between biological annotations. The rest of this paper is organized as follows. In section 2, the work that has been done previously in this field is summarized. In section 3, the description of applying IAMIC in the biological annotations is presented. In section 4, the experiment results are compared to the traditional similarity algorithms used by Tatiana V. Karpinets et al. [2]. 2. RELATED WORK To satisfy the increasing requirement of computational methods that process collected biological information into new knowledge, Tatiana V. Karpinets et al. [2] suggested an approach for analyzing the records of collected biological annotations in order to discover the associated annotations. This approach firstly converted the original annotations into the type-value formatted transactions which were conductive for preserving the semantic structure of data, and then calculated the records of annotations co-occurrence in the transactions to formulate a support matrix in which each row or column shown the co-occurrence values between a specific annotation and other annotations. Further, it used the one of the traditional similarity measures Pearson correlation coefficient to mine the relationship of annotations. Tatiana V. Karpinets et al. [2] applied the method to the GOLD datasets and demonstrated the ability to identify the significant relationships between biological annotations that especially do not co-occur with each other. Particularly, they discussed it may had some limitations as other statistical analysis shown, and then expected novel measures of similarity, such as maximal information coefficient (MIC) [3] could be utilized to discover hidden regularities in collected datasets. MIC is an interesting statistical measure proposed by Reshef et al. [3], which has always been attracting the eye of many fields due to its two properties: generality and equitability [4][5]. Compared to traditional similarity measures like Pearson correlation coefficient, MIC was more appropriate for a wide range of associations, not limited to specific relationships, and more equitable for no preference to certain functional types. MIC has been applied to biological terms successfully [6] [7] [8], including clinical data, genomics and virology applications. Although MIC was a great advance, it remained some drawbacks as presented in [9] [10] [11] and was argued less practical than distance correlation [13] [14] and HHG [15]. And then the later study of Reshef et al. [12] suggested that only the approximation algorithm for MIC resulted in these existed limitations rather than the intrinsic feature of MIC. Furthermore, we have been working on the improvement of the standard approximation algorithm for MIC to come more close to the true value of MIC. In this work, we want to apply the improved algorithm for MIC called IAMIC to the framework proposed by Tatiana V. Karpinets et al. [2] to uncover novel associations between biological annotations. 3. MINEING BIOLOGICAL DATASETS WITH IAMIC Here we are going to describe the main steps of mining biological annotations with an improved algorithm for calculating MIC named IAMIC. Firstly, we will introduce the principle of this algorithm for better understanding of whole work in our paper. 3.1 THE PRINCIPLE OF IAMIC Reshef et al.[3] stated that MIC is the metric value for the relationships between two variables. And he provided a MIC calculating algotithm, like a fitting method, which encapsulates the 2D splashes with grid partition. In other words, MIC is the value related to the grid partition that best reflects the true relationship between two variables. This fitting like grid partition method has the property of equitability naturally. It means that MIC will give similar scores to different functional relationships with similar noise levels or similar R2 (coefficient of determination). To calculate the MIC value, the standard algorithm proposed by Reshef et al. [3] simply equipartition on y-axis and then search the optimal grid partition on x-axis. That algorithm seems to reveal the most possible relationship between two variables. But in fact it is just an approximation for real MIC value. And that approximation will lead to a wrong answer. The simplest way to get the real MIC value is to enumerate all the possible grid partition, and find the best fitting one. However, the time expend is too high to follow. Wang et al. [16] proposed a fast and high accuracy method to approximate the real MIC value called IAMIC. This new algorithm attempts to find a better partition on y-axis through quadratic optimization instead of violence search. Specifically, IAMIC utilizes quadratic optimization on the largest value of each row of the characteristic matrix calculated by original algorithm in [3]. It retains the generality and improves the equitability in MIC, thus it is also more appropriate for mining the novel associations hidden in collected datasets than other similarity measures like Pearson correlation coefficient which has limitations for nonlinear relationships. 3.2 THE SIMPLE WORKFLOW OF APPLICATION After brief introduction about new algorithm IAMIC, it can more clearly present the workflow of mining biological annotations with this method. In our work, we will adopt the preprocessing process of datasets introduced by Tatiana V. Karpinets et al. [2]. It transfers the table formatted records to type-value formatted annotations. For each unique annotation, we calculate the values of co-occurrence with all other annotations in input datasets in order to build one row or column of a support matrix which is the object for later analysis by IAMIC. And then make use of new similarity measure IAMIC to quantify the associations between every pairwise annotation by computing their co-occurrence values recorded in the support matrix. Finally use the biological information visualization tool Cytoscape to show the data analysis results. The details of this workflow are presented as Figure 1. Work flow: (a) –step1 (b) –step2 (c) –step3 (d) ID Super-kingdom Group 1 Bacteria Firmicutes 2 Bacteria Firmicutes 3 Bacteria Alphaproteobacteria 4 Bacteria Betaproteobacteria 5 Bacteria Other_Bacteria Gram-stain + - Shape Cocci Rod Rod Rod Spiral (a) Collected table records ID1:{ Super-kingdom: Bacteria , Group: Firmicutes , Gram-stain:+ , Shape: Cocci } ID2:{ Super-kingdom: Bacteria , Group: Firmicutes , Gram-stain:- , Shape: Rod } ID3:{ Super-kingdom: Bacteria , Group: Alphaproteobacteria , Gram-stain:- , Shape: Rod } ID4:{ Super-kingdom: Bacteria , Group: Betaproteobacteria , Gram-stain:- , Shape: Rod } ID5:{ Super-kingdom: Bacteria , Group: Other- bacteria , (b) Type-value formatted records (c) Support matrix (d) Association matrix Gram-stain:- , Shape: Spiral } Figure 1: workflow of applying IAMIC for mining biological annotations In Figure 1, we present the workflow through simple example with four annotation types and five records corresponding to them. Step 1 shows the process of converting the collected table records (Figure 1(a)) to specific type-value formatted records (Figure 1(b)), and then generates a support matrix (Figure 1(c)) for unique annotations as shown in step 2. For instance, the value in the first row and second column of the support matrix refers to the number of records where annotation super_kingdom : Bacteria and annotation Group : Firmicutes co-occur, and especially, the values in the diagonal denote the number of records where annotations co-occur with itself, which are equal to the number of times in collected table recording such annotations. Step 3 uses IAMIC measure to estimate the similarity pairs of annotations by the support matrix produced in step 2 to create an association matrix (Figure 1(d)). 4. RESULTS AND COMPARISONS We apply the method demonstrated in Chapter 3 to analyze the biological annotations document that is sample datasets in program provided by Tatiana V. Karpinets et al. [2]. Simultaneously, we also use the traditional similarity measures such as Pearson correlation coefficient, Spearman’s rank correlation coefficient or Jaccard coefficient respectively to mining the same datasets. In this section, we are going to reveal these comparision results as shown in Figure 2. The results are generated by visualization tool Cytoscape with the association matrix, of which the vertexes refer to the annotations and the edges refer to associations estimated by similarity values between the annotations. There are 6782 collected table records, which contain 1109 unique annotations. We select 308 high frequency annotations that are found more than 6 records in the database during the experiments and screen out 2304 significant associations by setting p-value threshold 0.05. The comparision results from Figure 2 apparently show some advantages for IAMIC. Firstly, the measure IAMIC has better cluster feature than others. And some tight relationships between annotations are closer, unremarkable relations are far away, which bring us a direct information about which annotations are more likely to coexist. That is useful for analyzing the biological annotations. Secondly, it is easy to find that IAMIC lead to less isolated notation. That means IAMIC could mine the deeper relationship, which other method may not. Third, the more similar length between two annotations means a more disinterested measurement. That is to say, no matter what kind of relationships are, linear, curve, or even un-functional, IAMIC reveals their real relevance degree. (a) IAMIC (b) Pearson correlation coefficient (c) Spearman’s rank correlation coefficient (d) Jaccard(cosine) coefficient Figure 2: The comparision of analysis results 5. CONCLUSIONS In this paper, we prefer to use our improved approximation algorithm for MIC called IAMIC to mine the biological annotations with the work provided by Tatiana V. Karpinets et al. [2]. We believe IAMIC is the novel similarity measure which takes advantages of data exploration than traditional approaches of similarity, and verified it is also applicative to biological annotations experimentally. Finally, we hope a more appropriate similarity measure will be proposed in our next research, and it can be more conductive to uncover the hidden associations between increasing biological information. Acknowledgements This work was supported by National Natural Science Fund of China (61173061, 71201120), and Specialized Research Fund for the Doctoral Program of Higher Education (20121101110036). Reference [1] Howe D, Rhee SY et al. (2008) The future of biocuration. Nature 455: 47-50 [2] Karpinets TV, Park BH, Uberbacher EC (2012) Analyzing large biological datasets with association networks. Nucleic Acids Res 40: 17 [3] Reshef DN, Reshef YA, Finucane HK, Grossman SR, McVean G, Turnbaugh P, Lander ES, Mitzenmacher M, Sabeti PC (2011) Detecting novel associations in large data sets. Science 334 (6062): 1518-1524 [4] Speed T (2011) A correlation for the 21st century. Science 334 (6062): 1502-1503 [5] Stolovitzky G, Qiu P, Segal E, Noble B, Troyanskaya O, Simon N, Tibshirani R, Dougherty E (2012) Finding correlations in big data. Nat Biotechnol 30: 334–335 [6] Lin C, Canhao H, Miller T, Dligach D, Plenge RM, Karlson EW, Savova G (2012) Maximal information coefficient for feature selection for clinical document classification. ICML Workshop on Machine Learning for Clinical Data. [7] Das J, Mohammed J, Yu H (2012) Genome-scale analysis of interaction dynamics reveals organization of biological networks. Bioinformatics 28 (14): 1873-1878 [8] Anderson TK, Laegreid WW, Cerutti F, Osorio FA, Nelson EA, Christopher-Hennings J, Goldberg TL (2012) Ranking viruses: measures of positional importance within networks define core viruses for rational polyvalent vaccine development. Bioinformatics 28 (12): 1624-1632 [9] Simon N, Tibshirani R (2012) Comment on “Detecting novel associations in large data sets". Unpublished. http://www-stat.stanford.edu/_tibs/reshef/comment.pdf. Accessed 11 Nov. 2012 [10] Gorfine M, Heller R, Heller Y. (2012) Comment on “Detecting novel associations in large datasets". Unpublished. http://emotion.technion.ac.il/~gorfinm/files/science6.pdf. Accessed 11 Nov 2012 [11] Kinney JB, Atwal GS (2013) Equitability, mutual information, and the maximal information coefficient. ArXiv preprint arXiv:1301.7745 [12] Reshef D, Reshef Y, Mitzenmacher M, Sabeti P (2013) Equitability analysis of the maximal information coefficient, with comparisons. ArXiv preprint arXiv:1301.6314 [cs.LG] [13] Szekely G, Rizzo M, Bakirov N (2007) Measuring and testing independence by correlation of distances. Ann Stat 35: 2769-2794 [14] Szekely GJ, Rizzo ML (2009) Brownian distance covariance. Ann Appl Stat 3 (4): 1236-1265 [15] Heller R, Heller Y, Gorfine M (2012) A consistent multivariate test of association based on ranks of distances. ArXiv preprint arXiv:1201.3522 [stat.ME] [16] Shuliang W, Yiping Z, Yue S, Wenzhong S, (2013) An Improved Approximation Algorithm for MIC.
5cs.CE
THE STATUS OF THE ZASSENHAUS CONJECTURE FOR SMALL GROUPS arXiv:1609.00042v1 [math.RA] 31 Aug 2016 ANDREAS BÄCHLE∗ , ALLEN HERMAN† , ALEXANDER KONOVALOV, LEO MARGOLIS‡ , AND GURMAIL SINGH Abstract. We identify all small groups of order up to 288 in the GAP Library for which the Zassenhaus conjecture on rational conjugacy of units of finite order in the integral group ring cannot be established by an existing method. The groups must first survive all theoretical sieves and all known restrictions on partial augmentations (the HeLP+ method). Then two new computational methods for verifying the Zassenhaus conjecture are applied to the unresolved cases, which we call the quotient method and the partially central unit construction method. To the cases that remain we attempt an assortment of special arguments available for units of certain orders and the lattice method. In the end, the Zassenhaus conjecture is verified for all groups of order less than 144 and we give a list of all remaining cases among groups of orders 144 to 287. The Zassenhaus conjecture for torsion units of integral group rings states that any torsion unit of ZG should be conjugate in CG to a unit of ±G. We will abbreviate the conjecture as ZC1. Though this conjecture was made in the 1960s, it has only been established for a few families of groups. In this article, we study the status of the conjecture for groups of small order. This effort follows an earlier one by Höfert and Kimmerle that dealt with groups of order up to 71 [15]. Using recent theoretical and computational advances, we establish ZC1 for all groups of order less than 144, and give an accurate account of the groups of order 144 to 287 for which ZC1 cannot be established by the methods currently at our disposal (see Table 1.) The first step is to apply theoretical sieves. First we eliminate all nilpotent groups [21], cyclic-by-abelian groups [5], and groups whose derived subgroup is contained in a Sylow subgroup for some prime p [11, Theorem 1.2]. Groups of the form H × C2 for which ZH satisfies ZC1 can also be eliminated [15, Corollary 3.3], since we are inducting on group order. A further type of metabelian group described in [18] can also be sieved. Let ε denote the augmentation map on CG. A unit u of ZG is normalized if ε(u) =P 1. We will denote the normalized units of ZG by V(ZG). The partial augmentation of u = g ug g P on the conjugacy class xG of x ∈ G is εx (u) = y∈xG uy . An important connection between Date: August 31, 2016. 2010 Mathematics Subject Classification. 16U60, 16S34, 20C05, 20C10. Key words and phrases. Integral group ring, groups of units, Zassenhaus conjecture. ∗ This author is a postdoctoral researcher of the FWO (Research Foundation Flanders). † This author’s research has been supported by an NSERC Discovery Grant. ‡ This author’s research has been supported by a Marie Curie grant from EU project 705112-ZC. 1 2 BÄCHLE, HERMAN, KONOVALOV, MARGOLIS, AND SINGH torsion units in V(ZG), established in [18, Theorem 2.5], says that the torsion units satisfying the property: (*) for all n ≥ 0, there is only one conjugacy class xG in G such that εx (un ) 6= 0 are precisely the torsion units that are conjugate in CG to normalized trivial units. This result has inspired an approach that investigates the potential existence of normalized torsion units with more than one nonzero partial augmentation. The Luthar-Passi method [17], its modular extension by Hertweck [12], the Cohn-Livingstone and power congruence criteria (see [4, Remark 6]), and Hertweck’s criteria that εx (u) 6= 0 implies the order of x must divide the order of u [12, Theorem 2.3] produce several restrictions on partial augmentations that a normalized torsion unit of ZG must satisfy. The first and fourth authors recently produced the GAP package help [9] that implements this collection of methods, to which we refer as HeLP+ . After the sieve, the HeLP+ -method is applied to the remaining list of groups. For this step we use the GAP package [9], which utilizes the integer optimization programs normaliz [19] and 4ti2 [20]. Each HeLP+ -solution consists of the lists of partial augmentations for a possible normalized torsion unit and its powers. A nontrivial HeLP+ -solution is one for which εx (u) 6= 0 for more than one conjugacy class xG of G. For each nontrivial HeLP+ solution, we have to show that there is no torsion unit of V(ZG) that produces these partial augmentations. As this is an inductive approach, we can also eliminate a nontrivial HeLP+ solution for G by mapping it modulo a proper normal subgroup N to a nontrivial list of partial augmentations summing in G/N. If a normalized unit u with these partial augmentations exists in ZG, then its image would be a normalized unit with nontrivial partial augmentations in Z[G/N], whose existence would already be ruled out. Elimination by this “quotient method” is quite useful in groups that have several normal subgroups. The HeLP package also gives the spectrum for the torsion unit under each irreducible representation of the group. This information is equivalent to the partial augmentation information. It is used directly in the partially central construction and in the lattice method. In the partially central construction, we directly check if the central components of u in the Wedderburn decomposition of CG lie in the image of ZG. If they do not, the unit cannot be conjugate to an element of ZG. The partially central construction and the quotient method are implemented with software (see [8]). To the groups with nontrivial HeLP+ -solutions that remain after these computational procedures, we check a few more additional criteria available for units of particular orders by hand, and finally check if the lattice method of [3] can be applied. In the end, only a handful of groups of order between 144 and 287 survive all of these tests (see Table 1.) Our notation is based on GAP’s. We will refer to the group identified by SmallGroup(n,r) in GAP’s Small Groups Library as SG(n, r). For brevity we only record one nontrivial HeLP+ -solution in each Aut(G)-orbit. We denote conjugacy classes with their labels in GAP, and also use this notation for partial augmentations, so for example, (4a, 4c, 4f ) = (2, 1, −2) means the partial augmentations of the unit would be 2 on the class of elements of order 4 GAP labels with 4a by GAP, 1 on the class 4c, −2 on the class 4f , and 0 on all other conjugacy classes. ZC FOR SMALL GROUPS 3 Table 1. Unresolved nontrivial HeLP+ -solutions among groups of order up to 287. Structure GAP Id (144,117) C32 ⋊ D16 C32 ⋊ Q16 (144,119) (150,5) C52 ⋊ S3 (160,234) C24 ⋊ D10 (192,955) C24 ⋊ D12 3 (192,973) C2 ⋊ SG(24, 8) (192,974) C23 ⋊ SG(24, 8) (192,975) C23 : SG(24, 8) (192,976) C23 : SG(24, 8) (192,1489) C23 : S4 C23 : S4 (192,1490) (216,33) C33 ⋊ Q8 (216,35) (216,37) C33 ⋊ D8 C33 ⋊ D8 (216,153) (240,91) C32 ⋊ SL(2, 3) A5 ⋊ C4 unresolved cases: PAs[powers] (6b,6d,6e)=(-1,1,1)[2b,3b] (12a,12b,12d)=(1,1,-1)[2a,3a,4a,6a] (10a,10c)=(2,-1)[2a,5a] (2a,2b,2c)=(1,-1,1) (2a,2c,2f)=(1,1,-1),(1,-1,1) (8a,8bc,8d) = (1,-1,1)[2b,4db] (8a,8bc,8d) = (1,-1,1)[2c,4a] (8a,8bc,8d) = (1,-1,1)[2a,4fc] (8a,8bc,8d) = (1,-1,1)[2b,4eb] (8a,8b,8c)=(-1,1,1)[2b,4e] (8a,8b,8c) = (-1,1,1)[2c,4d] (12a,12c,12f) = (1,1,-1)[2a,3a,4a,6a] (0,2,-1)[2a,3c,4c,6c] (12a,12b)=(2,-1)[2c,3c,4d,6e] (6a,6e,6h) = (1,1,-1)[2a,3a], (0,2,-1)[2c,3c] (3a,3d,6a)= (1,-1,1),(-1,1,1)[3c,2a] (4a,4c,4d)=(1,0,-1,1)[2a] Remarks. (i). In Table 1, the nontrivial lists of partial augmentations of the unit u and the conjugacy classes in G of powers ud for d dividing the order o(u) of u are given. Since we only seek a counterexample of minimal unit order, we assume ud is always conjugate to an element of G for any proper divisor d of o(u). We use an order convention for multiple nontrivial solutions; so (8a, 8bc, 8d) = (1, −1, 1)[2a, 4f c] in Table 1 dentoes the information for the two solutions (8a, 8b, 8d) = (1, −1, 1)[2b, 4f ] and (8a, 8c, 8d) = (1, −1, 1)[2a, 4c]. (ii). SG(24, 8) ≃ C3 ⋊ D8 is one of three groups with this structure description. (iii). SG(240, 91) is a (central) quotient of GL(2, 5). ZC1 was verified for GL(2, 5) by Bovdi and Hertweck in [4]. However, it does not follow from their arguments that ZC1 holds for G. 1. Groups of order less than 288 that survive our sieves. We abbreviate the group identified as SmallGroup(n,r) in GAP’s Small Groups Library as SG(n, r). Our first theoretical sieve removes all nilpotent, cyclic-by-abelian, and p-byabelian groups, and groups satisfying case (a) of [18, Main Theorem]. Since we are looking for a minimal counterexample, we can also eliminate groups of the form C2 × H, where the group H is known to satisfy ZC1. After this sieve, 1121 groups of order less than 288 remain. To these groups we apply the HeLP+ method. This leaves a list of 99 groups. Next, we apply the quotient method. If the list of partial augmentations in a nontrivial HeLP+ -solution maps onto a nontrivial list of partial augmentations in any proper quotient 4 BÄCHLE, HERMAN, KONOVALOV, MARGOLIS, AND SINGH of the group, it is eliminated by our inductive argument. This leaves a list of 35 groups, which appear in Table 2. Table 2. Cases remaining after sieves, HeLP+ , and the quotient method. Order 48 72 96 144 150 160 168 192 200 216 240 GAP Id #, (unit order) 30(4) 40(6) 65(8), 186(4), 227(2) 117(6), 119(12), 182(6) 5(10) 234(2,4) 43(6) 182(4), 186(16), 955(2), 958(4,8) 973(8), 974(8), 975(8), 976(8), 987(4), 990(4), 1473(4), 1478(4), 1489(8), 1490(8), 1491(4), 1492(4), 1493(2) 43(10) 33(12), 35(12), 37(6), 153 (3,6), 161 (3) 91(4) Remark. The smallest groups our methods do not cover turn out to have order 144, so it is at order 288 where the quotient method becomes ineffective. 2. Partially central unit constructions The second, third, and fifth authors have developed a computer program that can construct torsion units of CG that have the partial augmentations and powers in a given HeLP-solution (see [8] and [10]). Each nontrivial HeLP+ -solution has accompanying spectral information. A complete set of explicit irreducible representations of the group is used to construct a unit of CG with the desired spectrum, and it follows it is conjugate in CG to any torsion unit with the same partial augmentations and powers. We then take e to be the sum of all the centrally primitive idempotents of CG for which ue is central in CGe. The centrality can be observed directly from the spectral information since u must represent diagonally under these irreducible representations. Finally, we determine a subset of Ge that gives a Z-basis of ZGe, and express ue in this basis. If ue 6∈ ZGe, then no CG-conjugate of u will ever be an element of ZG, and so this HeLP+ -solution will not produce a counterexample to ZC1. When this occurs, we simply say that the unit u is partially central. When constructing units, we write e3abc for the sum of the centrally primitive idempotents of CG corresponding to χ3a , χ3b , and χ3c . If we write e124 , then this means the sum of all centrally primitive idempotents of CG corresponding to all irreducible characters of degrees 1, 2, and 4. We use, for example, g4b to denote a group element in the conjugacy class of G labelled 4b by GAP, χ3d for the 4th irreducible character of degree 3 in GAP. ZC FOR SMALL GROUPS 5 SG(48,30), order 4. There is one Aut(G)-orbit of nontrivial HeLP+ -solutions that survive the quotient method. Its partial augmentations are (4a, 4c, 4d) = (−1, 1, 1), and its square lies in the class 2a. Its spectral information shows the central part of the unit is ue13cd . χ1abcd χ2a χ2b χ3a χ3b χ3c χ3d spec(X(u)) 1, −1, i, −i (i, −i) (1, −1) (1, 1, 1) (−1, −1, −1) (i, i, −i) (i, −i, −i) Using this spectral information and explicit irreducible representations constructed in GAP for each χ ∈ Irr(G), we can construct a representative of u in CG. With our notational conventions, the u our program constructs looks like u = g4b e123cd + g2a e3a − g2a e3b . With our program we then construct an explicit Z-basis of ZGe13cd and write ue13cd in terms of this basis. When we do this the coefficients we get are non-integer rational numbers, so this means u is partially central. So the partially central unit construction completes the verification of ZC1 for this group. As we have implemented this partially central check with our software, we just give a summary of the cases where it works. SG(96,65), order 8: There is one Aut(G)-orbit of nontrivial HeLP+ -solutions that survive the quotient method: (8a, 8d, 8g) = (1, −1, 1), with u2 ∈ 4a, u4 ∈ 2a. The central part of the unit is ue13ef gh and it does not lie in ZGe13ef gh . So ZC1 holds for G. SG(96,186), order 4: There is one Aut(G)-orbit of solutions left to resolve: (4b, 4g, 4h) = (−1, 1, 1), with u2 ∈ 2b. Our calculations show its central part ue13ef gh 6∈ ZGe13ef gh . So ZC1 holds for G. SG(96,227), order 2: One orbit of solutions to resolve: (2b, 2c, 2d) = (1, 1, −1). The unit we construct is partially central. So ZC1 holds for G. SG(192,182), order 4: Only one solution is left to resolve: (4c, 4d, 4g) = (1, 1, −1), u ∈ 2a. Since ue13gh 6∈ ZGe13gh , this verifies ZC1 for G. 2 SG(192,186), order 16: One solution is left to resolve: (16a, 16e, 16k) = (1, −1, 1), u2 ∈ 8a, u4 ∈ 4a, and u8 ∈ 2a. Since ue13ijkℓmnop 6∈ ZGe13ijkℓmnop , ZC1 holds for G. SG(192,955), order 2: There are five Aut(G)-orbits of order 2 solutions left to resolve: (2a, 2b, 2c, 2d, 2f ) = (1, 0, −1, 0, 1), (0, 1, −1, 0, 1), (0, 0, 1, 1, −1), (1, 0, 1, 0, −1), (0, 0, 2, 0, −1). For the first, second, and fourth of these, the central part of the unit we construct for them lies in ZGe, so the partially central method does not resolve them. For the third and fifth, the central part of the unit does not lie in ZGe, so these cases are resolved. The other three nontrivial HeLP+ -solutions for this group remain unresolved. SG(192,958), order 4: One solution left to resolve: (4b, 4g, 4j) = (1, −1, 1), u2 ∈ 2b. The unit we construct for this solution is partially central, so this verifies ZC1 for units of order 4. 6 BÄCHLE, HERMAN, KONOVALOV, MARGOLIS, AND SINGH SG(192,958), order 8: One orbit of solutions left to resolve: (8b, 8h, 8n) = (1, −1, 1), with u2 ∈ 4a, and u4 ∈ 2b. ue13abcdghij 6∈ ZGe13abcdghij , so the unit is partially central. This verifies ZC1 for G. SG(192,987), order 4: There are four Aut(G)-orbits of HeLP+ -solutions remaining to be resolved, which we can summarize as (4abcd, 4e, 4f ) = (1, 1, −1), u2 ∈ 2a. Our construction shows all four of these units are partially central. So this verifies ZC1 for G. SG(192,990), order 4: There are three Aut(G)-orbits of nontrivial solutions to consider: (4abc, 4d, 4e) = (1, 1, −1), with u2 ∈ 2c. All three are shown to be partially central by our program. So ZC1 is verified for G. SG(192,1473), order 4: One orbit of solutions to resolve: (4b, 4h, 4k) = (1, −1, 1), u2 ∈ 2c, which is shown to be partially central by our construction. So this resolves ZC1 for G. SG(192,1478), order 4: One orbit of solutions to resolve: (4c, 4i, 4m) = (1, −1, 1), u ∈ 2b, which is shown to be partially central by our construction. So this resolves ZC1 for G. 2 SG(192,1491), order 4: One orbit of solutions to resolve: (4a, 4d, 4e) = (1, −1, 1), u2 ∈ 2c. The unit our program constructs for this solution is partially central, so this verifies ZC1 for G. SG(192,1492), order 4: There are two orbits of solutions remaining: (4ab, 4c, 4e) = (1, −1, 1), with u2 ∈ 2b. The units our program constructs for these solutions are partially central, so this verifies ZC1 for G. SG(192,1493), order 2: There is one orbit of solutions to resolve, represented by (2a, 2b, 2e) = (−1, 1, 1). The unit our program gives for this solution is partially central, so this verifies ZC1 for G. 3. Methods for units of special order. The first of the special results we can apply is Proposition 4.2 of [11] concerning p-adic conjugacy of units. An easy integral consequence of this p-adic result is the following one, relevant for the p-subgroup version of the Zassenhaus conjecture (p-ZC3): Proposition 1. Let N be a normal p-subgroup of G. Suppose U is a finite subgroup of V(ZG) which maps to 1 under the natural map modulo N. Then U is conjugate in QG to a subgroup of N. SG(160,234), order 2: There are two orbits of solutions to consider: (2a, 2b, 2c) = (1, 1, −1) and (2b, 2c, 2d) = (1, 1, −1). For the latter, the classes 2b, 2c, and 2d are contained in a normal subgroup N of order 16, so we can apply Proposition 1 to resolve it. The other solution remains unresolved. SG(192,955), order 2: The three nontrivial solutions remaining to be resolved are (2a, 2b, 2c, 2f ) = (1, 0, −1, 1), (1, 0, 1, −1), (0, 1, −1, 1). The group has a normal subgroup N of order 32 that contains the classes 2b, 2c, and 2f . An application of Proposition 1 resolves the third of these solutions. The first two remain unresolved. ZC FOR SMALL GROUPS 7 The next lemma we can use is [14, Proposition 2]: Proposition 2. Let N be a normal p-subgroup of G. Suppose u is a torsion unit of ZG with augmentation 1 whose image modulo N has strictly smaller order than u. Then εg (u) = 0 for every g ∈ G whose p-part has strictly smaller order than the p-part of u. SG(72,40), order 6: The nontrivial HeLP+ solutions lie in two distinct Aut(G)-orbits, whose partial augmentations are: (2a, 2c, 6a) = (1, 1, −1), (−1, 1, 1). The group has a normal subgroup N of order 9, and the image of u modulo N in both cases would have order 2. Both of these cases fail the criteria of Proposition 2 at the prime p = 3. SG(144,182), order 6: The same reasoning applies. There are two orbits of nontrivial solutions, with (2a, 2b, 6a) = (1, 1, −1), (−1, 1, 1). There is a normal subgroup N of order 9 for which u mod N has order 2. So these partial augmentations of u also fail Proposition 2. SG(160,234), order 4: There is one orbit of nontrivial solutions to resolve: (2a, 4a) = (2, −1), u2 ∈ 2c. As noted earlier, G has a normal subgroup of order 16 containing g2c but not g2a . Modulo N, u will have order 2, so Proposition 2 applies. So normalized torsion units of order 4 in ZG will be rationally conjugate to elements of G. SG(168,43), order 6: There are ten nontrivial HeLP+ -solutions of order 6: (3a, 3b, 6a, 6b) = (1, 2, −1, −1), (1, −1, −1, 2), (2, 1, −1, −1), (−1, 1, 2, −1), (−1, 1, 1, 0), = (1, −1, 0, 1), (2, 1, −2, 0), (1, 2, 0, −2), (3, 0, −2, 0), (0, 3, 0, −2). The group G has a nontrivial normal subgroup N of order 8 generated by the class 2a. Since |G/N| = 21, any torsion unit u of ZG that has order 6 will be mapped modulo N to an element of order 3. By applying Proposition 2 with p = 2 we can eliminate every nontrivial solution with ε3a (u) 6= 0 or ε3b (u) 6= 0. So this verifies ZC1 for G. SG(200,43), order 10: There are two Aut(G)-orbits of nontrivial HeLP+ -solutions to resolve: (2a, 2c, 10a) = (−1, 1, 1), (1, 1, −1). G has a normal subgroup of order 25, so Proposition 2 tells us the partial augmentations of a normalized torsion unit of ZG with order 10 should vanish on elements of order 2. So these solutions are resolved, and this verifies ZC1 for G. We will need another result [6, Theorem 5.3] where ZC3 for p-subgroups is known. Proposition 3. Let G be a finite solvable group, and suppose L is the last nontrivial term of the lower central series of G. If p is a prime dividing |L| for which p4 does not divide |G|, then any finite p-subgroup of normalized torsion units is rationally conjugate to a subgroup of G. Of course, the last nontrivial term of the lower central series is the smallest normal subgroup L of G for which G/L is nilpotent. SG(216,153), order 3: For this group, if L is the last non-trivial term of the lower central series of G, then |L| = 72. By Proposition 3, normalized units of ZG with 2- or 3-power order are rationally conjugate to elements of G. SG(216,161), order 3: For this group, if L is the last non-trivial term of the lower central series of G, then |L| = 27. By Proposition 3, normalized units of ZG with 3-power order are rationally conjugate to elements of G. This verifies ZC1 for this group. 8 BÄCHLE, HERMAN, KONOVALOV, MARGOLIS, AND SINGH 5. Applying the Lattice Method. Let G = SG(216, 153), the special affine group F23 ⋊ SL(2, 3). After the arguments above there remain four Aut(G)-orbits of possibly non-trivial partial augmentations for units of order 6 in V(ZG). Two of these we will exclude using the lattice method introduced in [3]. We will use parts of the character table of G and the decomposition matrix of G given in table 3. The values of all characters on classes not given in the table are integral. Table 3. Character table of G χ1a χ2a χ3a χ8a 1a 2a 3a 3c 3d 6a 1 1 1 1 1 1 2 -2 -1 -1 -1 1 3 3 0 0 0 0 8 0 2 2 -1 0 Table 4. Decomposition matrix of G for the prime 3 χ1a χ2a χ3a χ8a ϕ1a ϕ2a ϕ3a 1 0 0 0 1 0 0 0 1 1 2 1 2a is the only class of involutions in G and in both possibilities of units we are going to study u2 is rationally conjugate to elements in 3c. Denote by ζ a primitive complex 3rd root of unity. Assume first that (ε3a (u), ε3d (u), ε6a(u)) = (−2, 2, 1). Then for representations D2a and D8a corresponding to the characters χ2a and χ8a respectively we find the following eigenvalues for u. D2a (u) ∼ diag(−ζ, −ζ 2), D8a (u) ∼ diag(ζ, ζ 2, ζ, ζ 2, −1, −1, −1, −1). By a theorem of Fong [16, Corollary 10.13] we may assume that the representations D2a and D8a are realized over a 3-adically complete discrete valuation ring which is unramified over the 3-adic integers. Denote by .̄ the reduction modulo the maximal ideal of R, also with respect to modules, and let k be the residue class field of R. Let L2a and L8a be RG-lattices corresponding to D2a and D8a respectively. Then when viewed as khūi-modules we have − − ∼ + by [3, Proposition 1.3] that L̄2a ∼ = L̄+ 2a ⊕ L̄2a and L̄8a = L̄8a ⊕ L̄8a are such that all the + + − composition factors of L̄2a and L̄8a are trivial while the composition factors of L̄− 2a and L̄8a are non-trivial. Moreover by [3, Propositions 1.3, 1.4] we know that L̄2a is a 2-dimensional indecomposable module while L̄− 8a is the direct sum of four 1-dimensional summands. But − since L̄2a is a sub- or factor module of L̄− 8a by the decomposition numbers in table 4, this contradicts the existence of u. Next assume (ε3a (u), ε3d(u), ε6a (u)) = (2, −2, 1). We will use similar notation as in the paragraph above. Then D3a (u) ∼ diag(1, ζ, ζ 2), D8a (u) ∼ diag(−ζ, −ζ 2, −ζ, −ζ 2 , 1, 1, 1, 1). ZC FOR SMALL GROUPS 9 Then viewed as khūi-modules we conclude, again by [3, Propositions 1.3, 1.4], that L̄+ 3a contains a direct indecomposable summand of dimension at least 2 while L̄+ is the direct 8a sum of four 1-dimensional modules. This again contradicts the decomposition numbers given in table 4. References [1] A. Bächle, A. Herman, A. Konovalov, L. Margolis, and G. Singh, ZCTU Source Code, https://github.com/drallenherman/ZCTU-Source-Code [2] A. Bächle and L. Margolis, HeLP – A GAP-package for torsion units in integral group rings, arXiv:1507.08174v3 [math.RT]. [3] A. Bächle and L. Margolis, Rational conjugacy of torsion units in integral group ring of non-solvable groups, Proc. Edinburgh Math. Soc., to appear; arXiv:1305.7419v3[mathRT]. [4] V. Bovdi and M. Hertweck, Zassenhaus conjecture for central extensions of Sn , J. Group Theory, 11 (1), (2008), 6374. [5] M. Caicedo, L. Margolis, and Á. del Rı́o, Zassenhaus conjecture for cyclic-by-abelian groups, J. London Math. Soc., (2), 88 (1), (2013), 65-78. [6] M. Dokuchaev and S. Juriaans, Finite subgroups in integral group rings, Can. J. Math., 48 (6), 1996), 1170-1179. [7] The GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.7.7; 2015, http://www.gap-system.org. [8] A. Bächle, A. Herman, A. Konovalov, L. Margolis, and G. Singh, ZCTU Source Code, https://github.com/alex-konovalov/ZCTU [9] A. Bächle and L. Margolis, HeLP – A GAP-package for torsion units in integral group rings, arXiv:1507.08174v3 [math.RT]. [10] A. Herman and G. Singh, Revisiting the Zassenhaus Conjecture on torsion units for the integral group rings of small groups, Proceedings. Math. Sciences. Indian Academy of Sciences, 125 (2), (2015), 167172. [11] M. Hertweck, On the torsion units of some integral group rings, Algebra Colloq., 13 (2), (2006), 329-348. [12] M. Hertweck, Partial augmentations and Brauer character values of torsion units in group rings, arXiv:math/0612429v2 [math.RA]. [13] M. Hertweck, Torsion units in integral group rings of certain metabelian groups, Proc. Edinburgh Math. Soc., 51, (2008), 363-385. [14] M. Hertweck, The orders of torsion units in integral group rings of finite solvable groups, Comm. Algebra, 36 (10), (2008), 3585-3588. [15] C. Höfert and W. Kimmerle, On torsion units of integral group rings of groups of small order. Groups, rings and group rings, 243252, Lect. Notes Pure Appl. Math., 248, Chapman & Hall/CRC, Boca Raton, FL, 2006. [16] I. M. Isaacs, Character theory of finite groups. Pure and Applied Mathematics, No. 69. Academic Press, New York-London, 1976. [17] I. S. Luthar and I. B. S. Passi, Zassenhaus conjecture for A5 . Proc. Indian Acad. Sci. Math. Sci. 99 (1989), no. 1, 15. [18] Z. Marciniak, J. Ritter, S. Sehgal, and A. Weiss, Torsion units in integral group rings of some metabelian groups, II, J. Number Theory, 25, (1987), 340-352. [19] normaliz W. Bruns, B. Ichim, and C. Söger, The power of pyramid decomposition in Normaliz. J. Symbolic Comput., 74 (2016), 513536. [20] 4ti2 team, 4ti2 – A software package for algebraic, geometric and combinatorial problems on linear spaces. Available at www.4ti2.de. [21] A. Weiss, Torsion units in integral group rings, J. Reine Angew. Math., 415 (1991), 175-187. 10 BÄCHLE, HERMAN, KONOVALOV, MARGOLIS, AND SINGH Vakgroep Wiskunde, Vrije Universiteit Brussel, Pleinlaan 2, 1050 Brussels, Belgium E-mail address: [email protected] Department of Mathematics and Statistics, University of Regina, Regina, Canada S4S 0A2 E-mail address: [email protected] School of Computer Science, University of St. Andrews, North Haugh, St. Andrews, Fife, KY16 SX9, U.K. E-mail address: [email protected] Departamento de matemáticas, Facultad de matemáticas, Universidad de Murcia, 30100 Murcia, Spain E-mail address: [email protected] Department of Mathematics and Statistics, University of Regina, Regina, Canada S4S 0A2 E-mail address: [email protected]
4math.GR
Latent nested nonparametric priors arXiv:1801.05048v1 [math.ST] 15 Jan 2018 Federico Camerlenghi Department of Economics, Management and Statistics, University of Milano–Bicocca and David B. Dunson Department of Statistical Science, Duke University and Antonio Lijoi and Igor Prünster∗ Department of Decision Sciences and BIDSA, Bocconi University and Abel Rodrı́guez Department of Applied Mathematics and Statistics, University of California at Santa Cruz ∗ A. Lijoi and I. Prünster are supported by the European Research Council (ERC) through StG ”N-BNP” 306406 1 Abstract Discrete random structures are important tools in Bayesian nonparametrics and the resulting models have proven effective in density estimation, clustering, topic modeling and prediction, among others. In this paper, we consider nested processes and study the dependence structures they induce. Dependence ranges between homogeneity, corresponding to full exchangeability, and maximum heterogeneity, corresponding to (unconditional) independence across samples. The popular nested Dirichlet process is shown to degenerate to the fully exchangeable case when there are ties across samples at the observed or latent level. To overcome this drawback, inherent to nesting general discrete random measures, we introduce a novel class of latent nested processes. These are obtained by adding common and group-specific completely random measures and, then, normalising to yield dependent random probability measures. We provide results on the partition distributions induced by latent nested processes, and develop an Markov Chain Monte Carlo sampler for Bayesian inferences. A test for distributional homogeneity across groups is obtained as a by product. The results and their inferential implications are showcased on synthetic and real data. Keywords: Bayesian nonparametrics; Completely random measures; Dependent nonparametric priors; Heterogeneity; Mixture models; Nested processes. 1 Introduction Data that are generated from different (though related) studies, populations or experiments are typically characterised by some degree of heterogeneity. A number of Bayesian nonparametric models have been proposed to accommodate such data structures, but analytic complexity has limited understanding of the implied dependence structure across samples. The spectrum of possible dependence ranges from homogeneity, corresponding to full exchangeability, to complete heterogeneity, corresponding to unconditional independence. It is clearly desirable to construct a prior that can cover this full spectrum, leading to a posterior that can appropriately adapt to the true dependence structure in the available data. This problem has been partly addressed in several papers. In Lijoi et al. (2014) a class of random probability measures is defined in such a way that proximity to full exchangeability or independence is expressed in terms of a r0, 1s–valued random variable. In the same spirit, a model decomposable into idiosyncratic and common components is devised in Müller et al. (2004). Alternatively, approaches based on Pólya tree priors are developed in Ma 2 & Wong (2011), Holmes et al. (2015), Filippi & Holmes (2017), while a multi–resolution scanning method is proposed in Soriano & Ma (2017). In Bhattacharya & Dunson (2012) Dirichlet process mixtures are used to test homogeneity across groups of observations on a manifold. A popular class of dependent nonparametric priors that fits this framework is the nested Dirichlet process of Rodrı́guez et al. (2008), which aims at clustering the probability distributions associated to d populations. For d “ 2 this model is ind Xi,1 , Xj,2 | pp̃1 , p̃2 q „ p̃1 ˆ p̃2 pp̃1 , p̃2 q | q̃ „ q̃ 2 , q̃ “ ÿ ωi δGi (1) iě1 where the random elements X` “ pXi,` qiě1 , for ` “ ř 1, 2, take values in a space X, the sequences pωi qiě1 and pGi qiě1 are independent, with iě1 ωi “ 1 almost surely, and the Gi ’s are i.i.d. random probability measures on X such that ÿ iid Gi “ wt,i δθt,i , θt,i „ P (2) tě1 for some non–atomic probability measure P on X. In Rodrı́guez et al. (2008) it is assumed that q̃ and the Gi ’s are realizations of Dirichlet processes while in Rodrı́guez & Dunson (2014) it is assumed they are from a generalised Dirichlet process introduced in Hjort (2000). Due to discreteness of q̃, one has p̃1 “ p̃2 with positive probability allowing for clustering at the level of the populations’ distributions and implying X1 „ X2 in such cases. The nested Dirichlet process has been widely used in a rich variety of applications, but it has an unappealing characteristic that provides motivation for this article. In particular, if X1 and X2 share at least one value, then the posterior distribution of pp̃1 , p̃2 q degenerates on tp̃1 “ p̃2 u, forcing homogeneity across the two samples. This occurs also in nested Dirichlet process mixture models in which the Xi,` are latent, and is not specific to the Dirichlet process but is a consequence of nesting discrete random probabilities. To overcome this major limitation, we propose a more flexible class of latent nested processes, which preserve heterogeneity a posteriori, even when distinct values are shared by different samples. Latent nested processes define p̃1 and p̃2 in (1) as resulting from normalisation of an additive random measure model with common and idiosyncratic components, the latter with nested structure. Latent nested processes are shown to have appealing distributional properties. In particular, nesting corresponds, in terms of the induced partitions, to a convex combination of full exchangeability and unconditional independence, the two extreme cases. This leads naturally to methodology for testing equality of distributions. 3 2 2.1 Nested processes Generalising nested Dirichlet processes via normalised random measures We first propose a class of nested processes that generalise nested Dirichlet processes by replacing the Dirichlet process components with a more flexible class of random measures. The idea is to define q̃ in (1) in terms of normalised completely random measures on the space P of probability measures on X. Let µ̃ be an almost ř surely finite completely random measure without fixed points of discontinuity, i.e. µ̃ “ iě1 Ji δGi where Gi are i.i.d. random probability measures on X with some fixed distribution Q on P. The corresponding Lévy measure on R` ˆ P is assumed to factorise as νpds, dpq “ c ρpsq ds Qpdpq (3) ş8 where ρ is some non–negative function such that 0 mint1, su ρpsq ds ă 8 and c ą 0. Since such a ν characterises µ̃ through its Lévy-Khintchine representation ż8´ ¯ ı ” ı ” ´λµ̃pAq 1 ´ e´λs ρpsq ds “: e´c QpAq ψpλq (4) E e “ exp ´ c QpAq 0 for any measurable A Ă P, we use the notation µ̃ „ CRMrν; Ps. The function ψ in (4) is also referred to as the Laplace exponent of µ̃. For a more extensive treatment ş8 of completely random measures, see Kingman (1993). If one additionally assumes that 0 ρpsq ds “ 8, then µ̃pPq ą 0 almost surely and we can define q̃ in (1) as µ̃ d (5) q̃ “ µ̃pPq This is known as a normalised random measure with independent increments, introduced in Regazzini et al. (2003), and is denoted as q̃ „ NRMIrν; Ps. The baseline measure, Q, of d µ̃ in (3) is, in turn, the probability distribution of q̃0 „ NRMIrν0 ; Xs, with q̃0 “ µ̃0 {µ̃0 pXq and µ̃0 having Lévy measure ν0 pds, dxq “ c0 ρ0 psq ds Q0 pdxq (6) ş8 ş8 for some non–negative function ρ0 such that 0 mint1, su ρ0 psq ds ă 8 and 0 ρ0 psqds “ 8. Moreover, Q0 is a non–atomic probability measure on X and ψ0 is the Laplace exponent of µ̃0 . The resulting general class of nested processes is such that pp̃1 , p̃2 q|q̃ „ q̃ 2 and is indicated by pp̃1 , p̃2 q „ NPpν0 , νq. The nested Dirichlet process of Rodrı́guez et al. (2008) is recovered by specifying µ̃ and µ̃0 to be gamma processes, namely ρpsq “ ρ0 psq “ s´1 e´s , so that both q̃ and q̃0 are Dirichlet processes. 4 2.2 Clustering properties of nested processes A key property of nested processes is their ability to cluster both population distributions and data from each population. In this subsection, we present results on: (i) the prior probability that p̃1 “ p̃2 and the resulting impact on ties at the observations’ level; (ii) equations for mixed moments as convex combinations of fully exchangeable and unconditionally independent special cases; and (iii) a similar convexity result for partially exchangeable partition probability function. The probability distribution of an exchangeable partition depends only on the numbers of objects in each group; the exchangeable partition probability function is the probability of observing a particular partition as a function of the group counts. Partial exchangeability is exchangeability within samples; the partially exchangeable partition probability function depends only on the number of objects in each group that are idiosyncratic to a group and common. Simple forms for the partially exchangeable partition probability function not only provide key insights into the clustering properties but also greatly facilitate computation. Before stating result (i), define ż8 ż8 q ´us p0q s e ρpsq ds, τq puq “ sq e´us ρ0 psq ds, τq puq “ 0 for any u ą 0, and agree that τ0 puq ” 0 p0q τ0 puq ” 1. Proposition 1. If pp̃1 , p̃2 q „ NPpν0 , νq, c “ QpPq and c0 “ Q0 pXq, then ż8 π1 :“ Ppp̃1 “ p̃2 q “ c u e´cψpuq τ2 puq du (7) 0 and the probability that any two observations from the two samples coincide equals ż8 p0q PpXj,1 “ Xk,2 q “ π1 c0 u e´c0 ψ0 puq τ2 puq ds ą 0. (8) 0 This result shows that the probability of p̃1 and p̃2 coinciding is positive, as desired, but also that this implies a positive probability of ties at the observations’ level. Moreover, (7) only depends on ν and not ν0 , since the latter acts on the X space. In contrast, the probability that any two observations Xj,1 and Xk,2 from the two samples coincide given in (8) depends also on ν0 . If pp̃1 , p̃2 q is a nested Dirichlet process, which corresponds to ρpsq “ ρ0 psq “ e´s {s, one obtains π1 “ 1{pc ` 1q and PpX1,1 “ X1,2 q “ π1 {pc0 ` 1q. The following proposition [our result (ii)] provides a representation of mixed moments as a convex combination of full exchangeability and unconditional independence between samples. 5 Proposition 2. If pp̃1 , p̃2 q „ NPpν0 , νq and π1 “ Ppp̃1 “ p̃2 q is as in (7), then ”ż ı E f1 pp1 qf2 pp2 qq̃pdp1 qq̃pdp2 q P2X ż “ π1 ż f1 ppqf2 ppqQpdpq ` p1 ´ π1 q PX (9) ż f1 ppqQpdpq PX f2 ppqQpdpq PX for all measurable functions f1 , f2 : P Ñ R` . This convexity property is a key property of nested processes. The component with weight 1 ´ π1 in (9) accounts for heterogeneity among data from different populations and it is important to retain this component also a posteriori in (1). Proposition 2 is instrumental to obtain our main result (iii) characterizing pn q the partially exchangeable random partition induced by X1 1 “ pX1,1 , . . . , Xn1 ,1 q and pn q X2 2 “ pX1,2 , . . . , Xn2 ,2 q in (1). To fix ideas consider a partition of the ni data of sampn q pn q ple Xi i into ki specific groups and k0 groups shared with sample Xj j (j ‰ i) with corresponding frequencies ni “ pn1,i , . . . , nki ,i q and qi “ pq1,i , . . . , qk0 ,i q. For example, p7q p4q X1 “ p0.5, 2, ´1, 5, 5, 0.5, 0.5q and X2 “ p5, ´2, 0.5, 0.5q yield a partition of n1 ` n2 “ 11 objects into 5 groups of which k1 “ 2 and k2 “ 1 are specific to the first and the second sample, respectively, and k0 “ 2 are shared. Moreover, the frequencies are n1 “ p1, 1q, n2 “ p1q, q1 “ p3, 2q and q2 “ p2, 1q. Let us start by analyzing the two extreme cases. For the fully exchangeable case (in the sense of exchangeability holding true across both samples), one obtains the exchangeable partition probability function ż8 ck0 pN q uN ´1 e´c0 ψ0 puq Φk pn1 , n2 , q1 ` q2 q “ ΓpN q 0 (10) k1 k2 k0 ź ź ź p0q p0q p0q ˆ τnj,1 puq τni,2 puq τqr,1 `qr,2 puq du j“1 r“1 i“1 ř having set N “ n1 ` n2 , k “ k0 ` k1 ` k2 and |a| “ pi“1 ai for any vector a “ pa1 , . . . , ap q P Rp with p ě 2. The marginal exchangeable partition probability functions for the individual sample ` “ 1, 2 are pn q Φk0 ``k` pn` , q` q pc0 qk0 `k` “ Γpn` q ż8 u n` ´1 0 e ´c0 ψ0 puq k ź̀ j“1 6 τnp0q puq j,` k0 ź r“1 τqp0q puq du r,` (11) Both (10) and (11) hold true with the constraints řk` j“1 nj,` ` řk0 r“1 qr,` “ n` and 1 ď k` `k0 ď p0q n` , for each ` “ 1, 2. Finally, the convention τ0 ” pnq pnq of the function Φk is zero, then it reduces to Φk´1 . 1 implies that whenever an argument p6q p6q For example, Φ3 p0, 2, 4q “ Φ2 p2, 4q. Both (10) and (11) solely depend on the Lévy intensity of the completely random measure and can be made explicit for specific choices. We are now ready to state our main result (iii). pn q pn q Theorem 1. The random partition induced by the samples X1 1 and X2 2 drawn from pp̃1 , p̃2 q „ NPpν0 , νq, according to (1), is characterised by the partially exchangeable partition probability function pN q pN q Πk pn1 , n2 , q1 , q2 q “ π1 Φk pn1 , n2 , q1 ` q2 q ` p1 ´ pn `|q |q pn `|q |q π1 q Φk0 1`k1 1 pn1 , q1 qΦk0 2`k2 2 pn2 , q2 q1t0u pk0 q (12) The two independent exchangeable partition probability functions in the second summand on the right–hand side of (12) are crucial for accounting for the heterogeneity across samples. However, the result shows that one shared value, i.e. k0 ě 1, forces the random partition to degenerate to the fully exchangeable case in (10). Hence, a single tie forces the two samples to be homogeneous, representing a serious limitation of all nested processes including the nDP special case. This result shows that degeneracy is a consequence of combining simple discrete random probabilities with nesting. In the following section, we develop a generalisation that is able to preserve heterogeneity in presence of ties between the samples. 3 Latent nested processes To address degeneracy of the partially exchangeable partition probability function in (12), we look for a model that, while still able to cluster random probabilities, can also take into pn q pn q account heterogeneity of the data in presence of ties between X1 1 and X2 2 . The issue is relevant also in mixture models where p̃1 and p̃2 are used to model partially exchangeable latent variables such as, e.g., vectors of means and variances in normal mixture models. To see this, consider a simple density estimation problem, where two-sample data of sizes n1 “ n2 “ 100 are generated from Xi,1 „ 1 1 Np5, 0.6q ` Np10, 0.6q 2 2 Xj,2 „ 7 1 1 Np5, 0.6q ` Np0, 0.6q. 2 2 This can be modeled by dependent normal mixtures with mean and variance specified in terms of a nested structure as in (1). The results, carried out by employing the algorithms detailed in Section 4, show two possible outcomes: either the model is able to estimate well the two bimodal marginal densities, while not identifying the presence of a common component, or it identifies the shared mixture component but does not yield a sensible estimate of the marginal densities, which both display three modes. The latter situation is displayed in Figure 1: once the shared component p5, 0.6q is detected, the two marginal distributions are considered identical as the whole dependence structure boils down to exchangeability across the two samples. (a) (b) Figure 1: Nested σ–stable mixture models: Estimated densities (blue) and true densities p100q p100q (red), for X1 in Panel (a) and for X2 in Panel (b). This critical issue can be tackled by a novel class of latent nested processes. Specifically, we introduce a model where the nesting structure is placed at the level of the underlying completely random measures, which leads to greater flexibility while preserving tractability. In order to define the new process, let M be the space of boundedly finite measures on X and Q the probability measure on M induced by µ̃0 „ CRMrν0 ; Xs, where ν0 is as in (6). 8 Hence, for any measurable subset A of X ż8´ ! ¯ ) ” ı ż ´λ mpAq ´λs ´λµ̃0 pAq e Qpdmq “ exp ´ c0 Q0 pAq 1´e ρ0 psq ds . E e “ M 0 Definition 1. Let q̃ „ NRMIrν; Ms, with νpds, dmq “ cρpsqds Qpdmq. Random probability measures pp̃1 , p̃2 q are a latent nested process if p̃` “ µ` ` µS µ` pXq ` µS pXq ` “ 1, 2, (13) where pµ1 , µ2 , µS q | q̃ „ q̃ 2 ˆ q̃S and q̃S is the law of a CRMrν0˚ ; Xs, where ν0˚ “ γ ν0 , for some γ ą 0. Henceforth, we will use the notation pp̃1 , p̃2 q „ LNPpγ, ν0 , νq. Furthermore, since p̃i “ wi µS µi ` p1 ´ wi q , µi pXq µS pXq where wi “ µi pXq , µS pXq ` µi pXq (14) each p̃i is a mixture of two components: an idiosyncratic component µi {µi pXq and a shared component µS {µS pXq. Here µS preserves heterogeneity across samples even when shared values are present. The parameter γ in the intensity ν0˚ tunes the effect of such a shared CRM. One recovers model (1) as γ Ñ 0. A generalisation to nested completely random measures of the results given in Propositions 1 and 2 is provided in the following proposition, whose proof is omitted. Proposition 3. If pµ1 , µ2 q | q̃ „ q̃ 2 , where q̃ „ NRMIrν; Ms as in Definition 1, then ż8 ˚ π1 “ Ppµ1 “ µ2 q “ c u e´cψpuq τ2 puq du (15) 0 and ”ż E M2 ż ı ˚ f1 pm1 q f2 pm2 q q̃ pdm1 , dm2 q “ π1 f1 pmq f2 pmq Qpdmq 2 M ` p1 ´ π1˚ q 2 ż ź `“1 for all measurable functions f1 , f2 : M Ñ R` . 9 M (16) f` pmq Qpdmq Proposition 4. If pp̃1 , p̃2 q „ LNPpγ, ν0 , νq, then Ppp̃1 “ p̃2 q “ Ppµ1 “ µ2 q. Proposition 4, combined with tp̃1 “ p̃1 u “ tµ1 “ µ2 u Y ptp̃1 “ p̃2 u X tµ1 ‰ µ2 uq, entails Prtp̃1 “ p̃2 u X tµ1 ‰ µ2 us “ 0 namely Pptp̃1 “ p̃2 u X tµ1 “ µ2 uq ` Pptp̃1 ‰ p̃2 u X tµ1 ‰ µ2 uq “ 1 and, then, the random variables 1 tp̃1 “ p̃2 u and 1 tµ1 “ µ2 u coincide almost surely. As a consequence the posterior distribution of 1 tµ1 “ µ2 u can be readily employed to test equality between the distributions of the two samples. Further details are given in Section 5. For analytic purposes, it is convenient to introduce an augmented version of the latent nested process, which includes latent indicator variables. In particular, pXi,1 , Xj,2 q | pp̃1 , p̃2 q „ p̃1 ˆ p̃2 , with pp̃1 , p̃2 q „ LNPpγ, ν0 , νq if and only if ind pXi,1 , Xj,2 q | pζi,1 , ζj,2 , µ1 , µ2 , µS q „ pζ1,i ˆ p2ζ2,j (17) pζi,1 , ζj,2 q | pµ1 , µ2 , µS q „ Bernpw1 q ˆ Bernpw2 q pµ1 , µ2 , µS q | pq̃, q̃S q „ q̃ 2 ˆ q̃S . The latent variables ζi,` indicate which random probability measure between p` and p0 “ pS generates each Xi,` , for i “ 1, . . . , n` . pn q pn q Theorem 2. The random partition induced by the samples X1 1 and X2 2 drawn from pp̃1 , p̃2 q „ LNPpγ, ν0 , νq, as in (17), is characterised by the partially exchangeable partition probability function pN q Πk pn1 , n2 , q1 , q2 q “ π1˚ ż8 ck0 p1 ` γqk ΓpN q k 2 ź̀ ź sN ´1 e´p1`γqc0 ψ0 psq ˆ 0 `“1 j“1 ` p1 ´ π1˚ q ÿ τnp0q psq j,` k0 ź p0q τqj,1 `qj,2 psqds j“1 I2 pn1 , n2 , q1 ` q2 , ζ ˚ q p˚q where ck0 γ k´k̄ I2 pn1 , n2 , q1 ` q2 , ζ q “ Γpn1 qΓpn2 q ż8ż8 un1 ´1 v n2 ´1 e´γc0 ψ0 pu`vq´c0 pψ0 puq`ψ0 pvqq ˚ 0 0 10 (18) k1 ź ˆ τnp0q pu j,1 ` p1 ´ ˚ qvq ζj,1 k2 ź τnp0q pp1 j,2 ´ ˚ qu ζj,2 p0q τqj,1 `qj,2 pu ` vqdudv j“1 j“1 j“1 ` vq k0 ź and the sum in the second summand on the right hand side of (18) runs over all the possible labels ζ ˚ P t0, 1uk1 `k2 . The partially exchangeable partition probability function (18) is a convex linear combination of an exchangeable partition probability function corresponding to full exchangeability across samples and one corresponding to unconditional independence. Heterogeneity across samples is preserved even in the presence of shared values. The above result is stated in full generality, and hence may seem somewhat complex. However, as the following examples show, when considering stable or gamma random measures, explicit expressions are obtained. When γ Ñ 0 the expression (18) reduces to (12), which means that the nested process is achieved as a special case. Example 1. Based on Theorem 2 we can derive an explicit expression of the partition structure of latent nested σ–stable processes. Suppose ρpsq “ σ s´1´σ {Γp1 ´ σq and ρ0 psq “ σ0 s´1´σ0 {Γp1 ´ σ0 q, for some σ and σ0 in p0, 1q. In such a situation it is easy to see that p0q π1˚ “ 1 ´ σ, τq puq “ σ0 p1 ´ σ0 qq´1 uσ0 ´q and ψ0 puq “ uσ0 . Moreover let c0 “ c “ 1, since the total mass of a stable process is redundant under normalization. If we further set ż1 wH1 ´1 p1 ´ wqH2 ´1 dw, Jσ0 ,γ pH1 , H2 ; Hq :“ σ0 ` p1 ´ wqσ0 sH 0 rγ ` w for any positive H1 , H2 and H, and k k0 2 ź̀ ź ź ξa pn1 , n2 , q1 ` q2 q :“ p1 ´ aqnj,` ´1 p1 ´ aqqj,1 `qj,2 ´1 , `“1 j“1 j“1 for any a P r0, 1q, then the partially exchangeable partition probability function in (18) may be rewritten as " p1 ´ σq σ pN q k´1 ` Πk pn1 , n2 , q1 , q2 q “ σ0 Γpkqξσ0 pn1 , n2 , q1 ` q2 q ΓpN q Γpn1 qΓpn2 q , . ÿ ˆ γ k´k̄ Jσ0 ,γ pn1 ´ n̄1 ` k̄1 σ0 , n2 ´ n̄2 ` k̄2 σ0 ; kq . p˚q 11 The sum with respect to ζ ˚ can be evaluated and it turns out that ” Bpk1 σ0 , k2 σ0 q σ0k´1 Γpkq ξσ0 pn1 , n2 , q1 ` q2 q 1 ´ σ ` σγ k0 ` q2 q “ Γpnq Bpn1 , n2 q śk2 ż 1 śk1 ni,2 ´σ0 nj,1 ´σ0 ı q i“1 r1 ` γp1 ´ wqs j“1 p1 ` γw Betapdw; k1 σ0 , k2 σ0 q ˆ ” ık 0 γ ` wσ0 ` p1 ´ wqσ0 pnq Πk pn1 , n2 , q1 where Betap ¨ ; a, bq stands for the beta distribution with parameters a and b, while Bpp, qq is the beta function with parameters p and q. As it is well–known, σ0k´1 Γpkq ξσ0 pn1 , n2 , q1 ` q2 q{ΓpN q is the exchangeable partition probability function of a normalised σ0 –stable process. Details on the above derivation, as well as for the following example, can be found in the Appendix. p0q Example 2. Let ρpsq “ ρ0 psq “ e´s {s. Recall that τq puq “ Γpqq{pu ` 1qq and ψ0 puq “ logp1 ` uq, furthermore π1˚ “ 1{p1 ` cq by standard calculations. From Theorem 2 we obtain the partition structure of the latent nested Dirichlet process " p1 ` γqk 1 pN q k Πk pn1 , n2 , q1 , q2 q “ ξ0 pn1 , n2 , q1 ` q2 qc0 1 ` c pc0 p1 ` γqqN , . k´k̄ ÿ c γ F pc ` n̄ , α, n ; α ` n , β ` n ; 1q ` 3 2 0 2 1 2 1 1`c pαqn2 pβqn1 p˚q where α “ pγ ` 1qc0 ` n1 ´ n̄1 , β “ c0 p2 ` γq and 3 F2 is the generalised hypergeometric function. In the same spirit as in the previous example, the first element in the linear convex combination above ck0 p1 ` γqk ξ0 pn1 , n2 , q1 ` q2 q{pc0 p1 ` γqqN is nothing but the Ewens’ sampling formula, i.e. the exchangeable partition probability function associated to the Dirichlet process whose base measure has total mass c0 p1 ` γq. 4 Markov Chain Monte Carlo algorithm We develop a class of Markov Chain Monte Carlo algorithms for posterior computation in latent nested process models relying on the partially exchangeable partition probability functions in Theorem 2, as they tended to be more effective. Moreover, the sampler is presented in the context of density estimation, where pn1 q Xj,` | pθ1 pn2 q , θ2 ind q „ hp ¨ ; θj,` q 12 ` “ 1, 2 pn1 q pXi,1 , Xj,2 q | pθ1 pn2 q , θ2 ind q „ hp ¨ ; θi,1 q ˆ hp ¨ ; θj,2 q pn q and the vectors θ` ` “ pθ1,` , . . . , θn` ,` q, for ` “ 1, 2 and with each θi,` taking values in Θ Ă Rb , are partially exchangeable and governed by a pair of pp̃1 , p̃2 q as in (17). The pn q pn q discreteness of p̃1 and p̃2 entails ties among the latent variables θ1 1 and θ2 2 that give rise to k “ k1 ` k2 ` k0 distinct clusters identified by pn q pn q pn q pn q • the k1 distinct values specific to θ1 1 , i.e. not shared with θ2 2 . These are denoted ˚ as θ1˚ :“ pθ1,1 , . . . , θk˚1 ,1 q, with corresponding frequencies n1 and labels ζ1˚ ; • the k2 distinct values specific to θ2 2 , i.e. not shared with θ1 1 . These are denoted ˚ , . . . , θk˚2 ,2 q, with corresponding frequencies n2 and labels ζ2˚ ; as θ2˚ :“ pθ1,2 pn q pn q ˚ • the k0 distinct values shared by θ1 1 and θ2 2 . These are denoted as θ0˚ :“ pθ1,0 , . . . , θk˚0 ,0 q, pn q with q` being their frequencies in θ` ` and shared labels ζ0˚ . As a straightforward consequence of Theorem 2, one can determine the joint distribution of the data X, the corresponding latent variables θ and labels ζ as follows pN q f px | θq Πk pn1 , n2 , q1 , q2 q k 2 ź̀ ź ˚ Q0 pdθj,` q (19) `“0 j“1 pN q where Πk ˚ ˚ u, u and Cr,`,0 :“ ti : θi,` “ θr,0 is as in (18) and, for Cj,` :“ ti : θi,` “ θj,` f px | θq “ k 2 ź̀ ź ź ˚ hpxi,` ; θj,` q k0 ź ź ˚ hpxi,` ; θr,0 q. r“1 iPCr,`,0 `“1 j“1 iPCj,` We do now specialise (19) to the case of latent nested σ–stable processes described in pn q Example 1. The Gibbs sampler is described just for sampling θ1 1 , since the structure pn q is replicated for θ2 2 . To simplify the notation, v ´j denotes the random variable v after the removal of θj,1 . Moreover, with T “ pX, θ, ζ, σ, σ0 , φq, weşlet T´θj,1 stand for T after deleting θj,1 , I “ 1tp̃1 “ p̃2 u and Q˚j pdθq “ hpxj,1 ; θq Q0 pdθq{ Θ hpxj,1 ; θq Q0 pdθq. Here φ denotes a vector of hyperparameters entering the definition of the base measure Q0 . The updating structure of the Gibbs sampler is as follows (1) Sample θj,1 from ÿ Ppθj,1 P dθ | T´θj,1 , I “ 1q “ w0 Q˚j,1 pdθq ` 13 wi,0 δtθ˚,´j u pdθq i,0 t ˚,´j i: ζi,0 “ζj,1 u ÿ ÿ wi,1 δtθ˚,´j u pdθq ` ` wi,2 δtθ˚,´j u pdθq i,1 ˚,´j i: ζi,1 “ζj,1 t u i,2 t ˚,´j i: ζi,2 “ζj,1 u ÿ Ppθj,1 P dθ | T´θj,1 , I “ 0q “ w01 Q˚j,1 pdθq ` ” ` 1t0u pζj,1 q w0 9 i,1 ˚,´j “ζj,1 u ti: ζi,1 k0 ı ÿ 1 1 wi,2 δtθ˚,´j u pdθq ` wr,0 δtθ˚,´j u pdθq ÿ i,2 ˚,´j “0u ti: ζi,2 where 1 wi,1 δtθ˚,´j u pdθq γ 1´ζj,1 σ0 k ´r hpxj,1 ; θq, 1`γ r,0 r“1 ˚,´j wi,` 9 pn´j i,` ´ σ0 qhpxj,1 ; θi,` q ` “ 1, 2 ´j ´j ˚,´j wi,0 9 pqi,1 ` qi,2 ´ σ0 qhpxj,1 ; θi,0 q ´j and, with a1 “ n1 ´ pn̄´j 1 ` ζj,1 q ` k̄1 σ0 and a2 “ n2 ´ n̄2 ` k̄2 σ0 , one further has w01 9 γ 1´ζj,1 σ0 k ´j Jσ0 pa1 ` ζj,1 σ0 , a2 ; k ´j ` 1q hpxj,1 ; θq, ˚,´j 1 wi,` 9 Jσ0 pa1 , a2 ; k ´j q pn´j i,` ´ σ0 q hpxj,` ; θj,` q ` “ 1, 2, ´j ´j ˚,´j 1 9 Jσ0 pa1 , a2 ; k ´j q pqi,1 ` qi,2 ´ σ0 q hpxj,1 ; θi,0 q. wi,0 ˚ from (2) Sample ζj,1 ˚ ˚ , I “ 1q “ Ppζj,1 “ x | T´ζj,1 γ 1´x 1`γ ˚ k´kx ´k̄0 ´k̄2 ˚ , I “ 0q 9 γ Ppζj,1 “ x | T´ζj,1 Jσ0 pn1 ´ nx ` kx σ0 , n2 ´ n̄2 ` k̄2 σ0 ; kq where x P t0, 1u, kx :“ x ` |ζ1˚,´j | and nx “ nj,1 x ` |ζ1˚,´j d n´j 1 |, where a d b denotes the component–wise product between two vectors a, b. Moreover, it should be stressed that, ˚ conditional on I “ 0, the labels ζr,0 are degenerate at x “ 0 for each r “ 1, . . . , k0 . (3) Update I from PpI “ 1 | T q “ 1 ´ PpI “ 0 | T q “ p1 ´ σqBpn1 , n2 q p1 ´ σqBpn1 , n2 q ` σJσ0 pā1 , ā2 ; kqp1 ` γqk 14 where ā1 “ n1 ´ n̄1 ` k̄1 σ0 and ā2 “ n2 ´ n̄2 ` k̄2 σ0 . This sampling distribution holds true pn q pn q ˚ ˚ “ 1. If this situation with label ζj,0 whenever θ1 1 and θ2 2 do not share any value θj,0 occurs, then PpI “ 1 | T q “ 1. (4) Update σ and σ0 from f pσ0 | T´σ0 , Iq 9 Jσ1´I pā1 , ā2 ; kq σ0k´1 κ0 pσ0 q 0 k 2 ź̀ ź p1 ´ σ0 qnj,` ´1 `“1 j“1 “ f pσ | T´σ , Iq 9 κpσq p1 ´ σq1t1u pIq ` σ1t0u pIq k0 ź p1 ´ σ0 qqr,1 `qr,2 ´1 r“1 ‰ where κ and κ0 are the priors for σ and σ0 , respectively. (5) Update γ from f pγ | T´γ , Iq 9 γ k´k̄ gpγq ” 1´σ ı 1 pIq ` σ J pā , ā ; kq 1 pIq σ0 1 2 t1u t0u p1 ` γqk where g is the prior distribution for γ. Finally, the updating of the hyperparameters depends on the specification of Q0 that is adopted. They will be displayed in the next section, under the assumption that Q0 is a normal/inverse–Gamma. The evaluation of the integral Jσ0 ph1 , h2 ; hq is essential for the implementation of the Markov Chain Monte Carlo procedure. This can be accomplished through numerical methods based on quadrature. However, computational issues arise when h1 and h2 are both less than 1 and the integrand defining Jσ0 is no longer bounded, although still integrable. For this reason we propose a plain Monte Carlo approximation of Jσ0 based on observing that ) ! 1 , Jσ0 ph1 , h2 ; hq “ Bph1 , h2 q E rγ ` W σ0 ` p1 ´ W qσ0 sh with W „ Betaph1 , h2 q. Then generating an i.i.d. sample tWi uLi“1 of length L, with Wi „ W , we get the following approximation L 1 1ÿ . Jσ0 ph1 , h2 ; hq « Bph1 , h2 q σ0 L i rγ ` Wi ` p1 ´ Wi qσ0 sh 15 5 Illustrations The algorithm introduced in Section 4 is employed here to estimate dependent random densities. Before implementation, we need first to complete the model specification of our latent nested model (13). Let Θ “ R ˆ R` and hp¨; pM, V qq be Gaussian with mean M and variance V . Moreover, as customary, Q0 is assumed to be a normal/inverse–Gamma distribution Q0 pdM, dV q “ Q0,1 pdV qQ0,2 pdM |V q with Q0,1 an inverse–Gamma probability distribution with parameters ps0 , S0 q and Q0,2 a Gaussian with mean m and variance τ V . Furthermore, the hyperpriors are τ ´1 „ Gampw{2, W {2q, m „ Npa, Aq, for some real parameters w ą 0, W ą 0, A ą 0 and a P R. In the simulation studies we have set pw, W q “ p1, 100q, pa, Aq “ ppn1 X̄ ` n2 Ȳ q{pn1 ` n2 q, 2q. The parameters τ and m are updated on the basis of their full conditional distributions, which can be easily derived, and correspond to ki 2 ÿ ˚ ´w k W ÿ pMi,j ´ mq2 ¯ ` , ` , L pτ |T´τ , Iq „ IG ˚ 2 2 2 2Vi,j i“0 j“1 ´R 1 ¯ L pm|T´m , Iq „ N , D D where R“ 2 ki ˚ Mi,j a ÿÿ ` , ˚ A i“0 j“1 τ Vi,j 2 D“ k i 1 1 ÿÿ ` . ˚ A i“0 j“1 τ Vi,j The model specification is completed by choosing uniform prior distributions for σ0 and σ. In order to overcome the possible slow mixing of the Pólya urn sampler, we include the acceleration step of MacEachern (1994) and West et al. (1994), which consists in resampling ˚ ki the distinct values pθi,j qj“1 , for i “ 0, 1, 2, at the end of every iteration. The numerical outcomes displayed in the sequel are based on 50, 000 iterations after 50, 000 burn–in sweeps. pn q pn q Throughout we assume the data X1 1 and X2 2 to be independently generated by two densities f1 and f2 . These will be estimated jointly through the MCMC procedure and the borrowing of strength phenomenon should then allow improved performance. An interesting byproduct of our analysis is the possibility to examine the clustering structure of each distribution, namely the number of components of each mixture. Since the expression 16 of the pEPPF (18) consists of two terms, in order to carry out posterior inference we have defined the random variable I “ 1tµ1 “µ2 u . This random variable allows to test whether the two samples come from the same distribution or not, since I “ 1tp̃1 “p̃2 u almost surely (see also Proposition 4). Indeed, if interest lies in testing H0 : p̃1 “ p̃2 versus H1 : p̃1 ­“ p̃2 , based on the Markov Chain Monte Carlo output, it is straightforward to compute an approximation of the Bayes factor BF “ PpI “ 1|Xq PpI “ 0q Ppp̃1 “ p̃2 |Xq Ppp̃1 ­“ p̃2 q “ Ppp̃1 ­“ p̃2 |Xq Ppp̃1 “ p̃2 q PpI “ 0|Xq PpI “ 1q leading to acceptance of the null hypothesis if BF is sufficiently large. In the following we first consider simulated datasets generated from normal mixtures and then we analyse the popular Iris dataset. 5.1 Synthetic examples pn q pn q We consider three different simulated scenarios, where X1 1 and X2 2 are independent and identically distributed draws from densities that are both two component mixtures of normals. In both cases ps0 , S0 q “ p1, 1q and the sample size is n “ n1 “ n2 “ 100. pn q pn q First consider a scenario where X1 1 and X2 2 are drawn from the same density Xi,1 „ Xj,2 „ 1 1 Np0, 1q ` Np5, 1q. 2 2 The posterior distributions for the number of mixture components, respectively denoted by K1 and K2 for the two samples, and for the number of shared components, denoted by K12 , are reported in Table 1. The maximum a posteriori estimate is highlighted in bold. The model is able to detect the correct number of components for each distribution as well as the correct number of components shared across the two mixtures. The density estimates, not reported here, are close to the true data generating densities. The Bayes factor to test pn q pn q equality between the distributions of X1 1 and X2 2 has been approximated through the Markov Chain Monte Carlo output and coincides with BF “ 5.85, providing evidence in favor of the null hypothesis. pn q pn q Scenario II corresponds to samples X1 1 and X2 2 generated, respectively, from Xi,1 „ 0.9 Np5, 0.6q ` 0.1 Np10, 0.6q Xj,2 „ 0.1 Np5, 0.6q ` 0.9 Np0, 0.6q. 17 scen. # comp. K1 I K2 K12 K1 II K2 K12 K1 III K2 K12 0 1 0 0 0 0 0 0 0 0 0 0 0 0.965 0 0 0 0 0 0.183 2 0.638 0.635 0.754 0.679 0.778 0.034 0.328 0.409 0.645 3 0.232 0.235 0.187 0.232 0.185 0.001 0.322 0.305 0.138 4 0.079 0.083 0.045 0.065 0.032 0 0.188 0.152 0.027 5 0.029 0.029 0.012 0.018 0.004 0 0.089 0.073 0.006 6 0.012 0.011 0.002 0.004 0.001 0 0.041 0.034 0.001 ě7 0.008 0.007 0.001 0.002 0 0 0.032 0.027 0 Table 1: Simulation study: Posterior distributions of the number of components in the first sample (K1 ), in the second sample (K2 ) and shared by the two samples (K12 ) corresponding to the three scenarios. The posterior probabilities corresponding to the MAP estimates are displayed in bold. Both densities have two components but only one in common, i.e. the normal distribution with mean 5. Moreover, the weight assigned to Np5, 0.6q differs in the two cases. The density estimates are displayed in Figure 2. The spike corresponding to the common component (concentrated around 5) is estimated more accurately than the idiosyncratic components (around 0 and 10, respectively) of the two samples nicely showcasing the borrowing of information across samples. Moreover, the posterior distributions of the number of components are reported in Table 1. The model correctly detects that each mixture has two components with one of them shared and the corresponding distributions are highly concentrated around the correct values. Finally the Bayes factor BF to test equality between the two distributions equals 0.00022 and the null hypothesis of distributional homogeneity is rejected. Scenario III consists in generating the data from mixtures with the same components pn q pn q but differing in their weights. Specifically, X1 1 and X2 2 are drawn from, respectively, Xi,1 „ 0.8 Np5, 1q ` 0.2 Np0, 1q Xj,2 „ 0.2 Np5, 1q ` 0.8 Np0, 1q, The posterior distribution of the number of components is again reported in Table 1 and again the correct number is identified, although in this case the distributions exhibit a 18 (a) (b) Figure 2: Estimated densities (blue) and true densities (red) for X in Panel (a) and Y in Panel (b). higher variability. The Bayes factor BF to test equality between the two distributions is 0.54, providing weak evidence in favor of the alternative hypothesis that the distributions differ. 5.2 Iris dataset Finally, we examine the well known Iris dataset, which contains several measurements concerning three different species of Iris flower: setosa, versicolor, virginica. More specifically, we focus on petal width of those species. The sample X has size n1 “ 90, containing 50 observations of setosa and 40 of versicolor. The second sample Y is of size n2 “ 60 with 10 observations of versicolor and 50 of virginica. Since the data are scattered across the whole interval r0, 30s, we need to allow for large variances and this is obtained by setting ps0 , S0 q “ p1, 4q. The model neatly identifies that the two densities have two components each and that one of them is shared as showcased by the posterior probabilities reported in Table 2. As for the Bayes factor, we obtain BF « 0 leading to the unsurprising conclusion that the two samples come from two different 19 distributions. The corresponding estimated densities are reported in Figure 3. # comp. K1 K2 K12 0 1 2 0 0 0.466 0 0.001 0.661 0 0.901 0.093 3 4 5 6 ě7 0.307 0.141 0.055 0.020 0.011 0.248 0.068 0.017 0.004 0.001 0.006 0 0 0 0 Table 2: Real data: Posterior distributions of the number of components in the first sample (K1 ), in the second sample (K2 ) and shared by the two samples (K12 ). The posterior probabilities corresponding to the MAP estimates are displayed in bold. Figure 3: Estimated densities for X (red) and Y (blue). We have also monitored the convergence of the algorithm that has been implemented. Though we here provide only details for the Iris dataset, we have conducted similar analyses also for each of the illustrations with synthetic datasets in Section 5.1. Notably, all the 20 examples with simulated data have experienced even better performances than those we are going to display henceforth. Figure 4 depicts the partial autocorrelation function for the sampled parameters σ and σ0 . The partial autocorrelation function apparently has an exponential decay and after the first lag exhibits almost negligible peaks. (a) (b) Figure 4: Plots of the partial autocorrelation functions for the parameters σ (a) and σ0 (b). We have additionally monitored the two estimated densities near the peaks, which identify the mixtures’ components. More precisely, Figure 5(a) displays the trace plots of the density referring to the first sample at the points 3 and 13, whereas Figure 5(b) shows the trace plots of the estimated density function of the second sample at the points 13 and 21. 6 Concluding remarks We have introduced and investigated a novel class of nonparametric priors featuring a latent nested structure. Our proposal allows flexible modeling of heterogeneous data and deals with problems of testing distributional homogeneity in two-sample problems. Even if our 21 (a) (b) pn q Figure 5: (a): trace plots of the density referring to X1 1 at the points 3 and 13; (b): trace pn q plots of the density referring to X2 2 at the points 13 and 21. treatment has been confined to the case d “ 2, we stress that the results may be formally extended to d ą 2 random probability measures. However, their implementation would be more challenging since the marginalization with respect to pp̃1 , . . . , p̃d q leads to considering all possible partitions of the d random probability measures. While sticking to the same model and framework which has been shown to be effective both from a theoretical and practical point of view in the case d “ 2, a more computationally oriented approach would be desirable in this case. There are two possible paths. The first, along the lines of the original proposal of the nested Dirichlet process in Rodrı́guez et al. (2008), consists in using tractable stick–breaking representations of the underlying random probabilities, whenever available to devise an efficient algorithm. The second, which needs an additional significant analytical step, requires the derivation of a posterior characterization of pp̃1 , . . . , p̃d q that allows sampling of the trajectories of latent nested processes and build up algorithms for which marginalization is not needed. Both will be the object of our future research. 22 Appendix 1 Proof of Proposition 1 Since pp̃1 , p̃2 q „ NPpν0 , νq, one has ż8 ż ż ż 2 µ̃ pdpq 2 “ u E e´uµ̃pPq µ̃2 pdpq du π1 “ E q̃ pdpq “ E 2 0 P P P µ̃ pPq (20) In order to get the result, we extend and adapt the techniques used in James et al. (2006). Indeed, it can be seen that “ ‰ E e´u µ̃pPX q µ̃2 pdpq “ e´c ψpuq c2 Q2 pdpq τ12 puq ` c Qpdpq τ2 puq . (21) ř ř Recall that Q is the probability distribution of the NRMI q̃0 “ jě1 ωj δθ̃j with jě1 ωj “ 1 iid almost surely and θ̃j „ Q0 . This means ř that Q is concentrated on the set of discrete probability measures on X. If p “ jě1 wj δθj P P is fixed, we set Wj,n :“ tωj1 “ w1 , . . . , ωjn “ wn u and Θj,n “ tθ̃j1 “ θ1 , . . . , θ̃jn “ θn u where j “ pj1 , . . . , jn q is a vector of positive integer. Then ´ ¯ ”ď ı Qptpuq “ P q̃0 “ p ď P pWj,n X Θj,n q (22) p˚q where the above union is taken over the set of all vectors j “ pj1 , . . . , jn q P Nn such that j1 ‰ ¨ ¨ ¨ ‰ jn . The upper bound in (22) is clearly equal to 0. This, combined with (21), yields ż ż E e´u µ̃pPX q µ̃2 pdpq “ c e´cψpuq τ2 puq PX Qpdpq “ c e´cψpuq τ2 puq PX and the proof is completed. Proof of Proposition 2 Let f1 “ 1A and f2 “ 1B , for some measurable subsets A and B of P. One has ż E f1 pp1 qf2 pp2 qq̃pdp1 qq̃pdp2 q “ E q̃pAqq̃pBq P2X “ E q̃ 2 pA X Bq ` E q̃pA X Bqq̃pB X Ac q ` E q̃pA X B c qq̃pBq. 23 It can now be easily seen that µ̃2 pA X Bq µ̃2 pPX q ż8 ” ı u e´cψpuq c2 Q2 pA X Bq τ12 puq ` c QpA X Bq τ2 puq du “ E q̃ 2 pA X Bq “ E 0 “ π1 Q2 pA X Bq ` p1 ´ π1 q QpA X Bq. On the other hand, if A X B “ ∅, we get ż8 c2 u e´c ψpuq τ12 puq du “ π1 QpAq QpBq E q̃pAqq̃pBq “ QpAqQpBq 0 To sum up, one finds that E q̃pAqq̃pBq “ π1 QpA X Bq ` p1 ´ π1 q rQ2 pA X Bq ` QpA X BqQpB X Ac q ` QpA X B c qQpBqs which boils down to (9). Now it is easy to prove that (9) is true when f1 and f2 are simple functions and, then, for all positive and measurable functions relying on the monotone convergence theorem. Proof of Theorem 1 pN q The partition probability function Πk pn1 , n2 , q1 , q2 q equals ż E Xk k1 ź n p̃1 j,1 pdx˚j,1 q j“1 k2 ź n p̃2 j,2 pdx˚j,2 q j“1 k0 ź q q p̃1j,1 pdzj˚ qp̃2j,2 pdzj˚ q (23) j“1 obtained by marginalizing with respect to pp̃1 , p̃2 q. Due to ś conditional independence of p̃1 and p̃2 , given q̃, the integrand in (23) can be rewritten as E 2`“1 h` pdx˚` , dz ˚ ; q̃q where, for each ` “ 1, 2, h` pdx˚` , dz ˚ ; q̃q “E k ” ź̀ n p̃` j,1` pdx˚j,` q j“1 ż k ź̀ “ k0 ź ˇ ı q̃ q ˇ p̃`j,` pdzj˚ qˇ j“1 n p` j,` pdx˚j,` q PX j“1 k0 ź j“1 24 q p`j,` pdzj˚ q q̃pdp` q A simple application of the Fubini–Tonelli theorem, then, yields ż ż pN q Πk pn1 , n2 , q1 , q2 q “ E f1 pp1 q f2 pp2 q q̃pdp1 q q̃pdp2 q Xk (24) P2X ś 0 qj,` ś ` nj,` p` pdzj˚ q and agree where, for each ` “ 1, 2, we have set f` pp` q :“ kj“1 p` pdx˚j,` q kj“1 ś0 that j“1 aj ” 1. In view of Proposition 2 the integrand in (24) boils down to ż ż P2X E f1 pp1 q f2 pp2 q q̃pdp1 q q̃pdp2 q “ π1 f1 ppq f2 ppq Qpdpq ` p1 ´ π1 q PX 2 ż ź `“1 f` ppq Qpdpq PX ” ı ” ı” ı “ π1 Ef1 pq̃0 q f2 pq̃0 q ` p1 ´ π1 q Ef1 pq̃0 q Ef2 pq̃0 q . In order to complete the proof it is now enough to note that, due to non–atomicity of Q0 , Ef1 pq̃0 qf2 pq̃0 q “ E k1 ź n q̃0 j,1 pdx˚j,1 q k2 ź n q̃0 j,2 pdx˚j,2 q q q̃0j,1 `qj,2 pdzj˚ q j“1 j“1 j“1 k0 ź is absolutely continuous with respect to Qk0 on Xk and dEf1 pq̃0 qf2 pq̃0 q ˚ ˚ ˚ pN q px1 , x2 , z q “ Φk pn1 , n2 , q1 ` q2 q k dQ0 for any vector px˚1 , x˚2 , z ˚ q whose k components are all distinct, and is zero otherwise. As for the second summand above, from Proposition 3 in James et al. (2009) one deduces that ” k1 k2 k0 ı” ı ź ź ź ˚ ˚ Ef1 pq̃0 q Ef2 pq̃0 q “ Q0 pdxj,1 q Q0 pdxj,2 q Q20 pdzj˚ q j“1 j“1 j“1 p|n |`|q1 |q 1 ˆ Φk0 `k 1 p|n |`|q2 |q 2 pn1 , q1 q Φk0 `k 2 pn2 , q2 q Then it is apparent that rEf1 pq̃0 qs rEf2 pq̃0 qs ! Qk0 and still by virtue of the non–atomicity of Q0 one has d rEf1 pq̃0 qs rEf2 pq̃0 qs ˚ ˚ ˚ p|n1 |`|q1 |q p|n2 |`|q2 |q px1 , x2 , z q “ Φk0 `k pn1 , q1 q Φk0 `k pn2 , q2 q 1t0u pk0 q k 1 2 dQ0 for any vector px˚1 , x˚2 , z ˚ q P Xk whose components are all distinct, and is zero otherwise. Note that if it were k0 ě 1, then some of the infinitesimal factors Q20 pdzj˚ q would not cancel and the above density would be exactly equal to zero. 25 Proof of Proposition 4 ř iid ω̃ δ , with Pp j η̃ j jě1 jě1 ω̃j “ 1q “ 1 and η̃j „ Q. ř ř pjq pjq Furthermore, η̃j is, in turn, a CRM η̃j “ kě1 ω̃k δX pjq where Pp kě1 ω̃k ă `8q “ 1 and Since q̃ „ NRMIrν; Ms, one has q̃ “ ř k pjq iid Xk „ µS “ ř Q0 , for any j “ 1, 2, . . . . An analogous representation holds true also for µS , i.e. p0q kě1 ω̃k δX p0q with the same conditions as above. From the assumptions one deduces k pjq pjq that the sequences pXk qkě1 and pω̃k qkě1 are independent also across different values of j, and Definition 1 entails, with probability 1, ˇ ı ” ˇ P pµ1 , µ2 , µS q P A1 ˆ A2 ˆ A0 ˇ q̃ “ q̃pA1 q q̃pA2 q q̃0 pA0 q which implies ˇ ¯ı µ2 ` µS µ1 ` µS ˇ “ ˇ q̃ µ1 pXq ` µS pXq µ2 pXq ` µS pXq ˇ ¯ ”ÿ ´ η̃j ` µS η̃i ` µS ˇ “E ω̃j ω̃i P “ ˇ q̃ η̃i pXq ` µS pXq η̃j pXq ` µS pXq i­“j ” ´ Ppp̃1 “ p̃2 q “ E P 8 ÿ ` ω̃i2 P i“1 ´ ˇ ¯ı η̃i ` µS η̃i ` µS ˇ “ ˇ q̃ η̃i pXq ` µS pXq η̃i pXq ` µS pXq ”ÿ ´ “E ω̃j ω̃i P i­“j 8 ˇ ¯ı ÿ η̃j ` µS η̃i ` µS ˇ “ ˇ q̃ ` E ω̃i2 . η̃i pXq ` µS pXq η̃j pXq ` µS pXq i“1 ř 2 For the second summand above one trivially has E 8 i“1 ω̃i “ Ppµ1 “ µ2 q. As for the first summand, a simple application of the Fubini–Tonelli theorem and the fact that ω̃j ě 1, for any j, yield the following upper bound ˇ ¯ı ”ÿ ´ η̃i ` µS η̃j ` µS ˇ E ω̃j ω̃i P “ ˇ q̃ η̃i pXq ` µS pXq η̃j pXq ` µS pXq i­“j “ ´ ÿ ” E ω̃j ω̃i P i­“j ÿ ď i­“j ´ P ˇ ¯ı η̃i ` µS η̃j ` µS ˇ “ ˇ q̃ η̃i pXq ` µS pXq η̃j pXq ` µS pXq ¯ η̃i ` µS η̃j ` µS “ . η̃i pXq ` µS pXq η̃j pXq ` µS pXq 26 The proof is completed by showing that this upper bound is zero. To this end, we fix piq piq positive integers i, j and n, consider the n–tuple of atoms pX1 , ¨ ¨ ¨ , Xn q referring to η̃i and correspondingly define the sets ! ) pjq piq pjq pjq Θ` :“ ω P Ω : X1 pωq “ X`1 pωq, . . . , Xnpiq pωq “ X`n pωq , ) ! p0q p0q piq p0q Θ` :“ ω P Ω : X1 pωq “ X`1 pωq, . . . , Xnpiq pωq “ X`n pωq for any ` “ p`1 , ¨ ¨ ¨ , `n q P Nn . It is then apparent that ´ P ¯ ” ı ď η̃j ` µ̃˚0 η̃i ` µ̃˚0 pjq p0q “ ď P pΘ Y Θ q ` ` η̃i pXq ` µ̃˚0 pXq η̃j pXq ` µ̃˚0 pXq `PNn ,` ­“` h1 h2 pjq p0q and this upper bound is equal to 0, because each of the events Θ` and Θ` in the above countable union has 0 probability in view of the non–atomicity of Q0 and independence. Proof of Theorem 2 pn q pn q Consider the partition induced by the sample X1 1 and X2 2 into k “ k1 ` k2 ` k0 groups with frequencies n` “ pn1,` , . . . , nk` ,` q, for ` “ 1, 2, and q̄ “ pq1,1 ` q2,1 , . . . , qk0 ,1 ` qk0 ,2 q. Recalling that p` “ µ` {µ` pXq, for ` “ 1, 2, the conditional likelihood is k 2 ź̀ ź k0 2 ź ź ˚ ˚ ˚ nj,` ζj,` ζ ˚ qr,` p1´ζj,` qnj,` p1´ζr,0 qpqr,1 `qr,2 q ˚ ˚ ˚ p` pdxj,` q pS pdxj,` q pS pdzr q p`r,0 pdzr˚ q r“1 `“1 j“1 `“1 where we take txj,` : j “ 1, . . . , k` u, for ` “ 1, 2, and tzr˚ : r “ 1, . . . , k0 u as the k1 ` k2 ` k0 distinct values in X. If we now let ´pu`vqµS pXq f0 pµS , u, vq :“ e k0 ź µS ˚ p1´ζr,0 qpqr,1 `qr,2 q pdzr˚ q r“1 f1 pµ1 , u, vq :“ e´uµ1 pXq k1 ź ζ ˚ nj,1 µ1j,1 `“1 j“1 pdx˚j,1 q k2 ź k0 ź ζ ˚ qr,1 µ1r,0 pdzr˚ q r“1 j“1 f2 pµ2 , u, vq :“ e´vµ2 pXq k 2 ź̀ ź ζ ˚ nj,2 µ2j,2 pdx˚j,2 q k0 ź r“1 j“1 27 ζ ˚ qr,2 µ2r,0 pdzr˚ q, ˚ qnj,` p1´ζj,` µ0 pdx˚j,` q and further take into account the probability distribution of the labels, conditional on pµS , µ1 , µ2 q, so that the the joint distribution of the random partition and of the corresponding labels ζ ˚˚ “ pζ1˚ , ζ2˚ , ζ0˚ q is pN q Πk pn1 , n2 , q1 , q2 ; ζ ˚˚ q 1 “ Γpn1 qΓpn2 q ż8ż8 u 0 n1 ´1 n2 ´1 v 0 2 ´ź ¯ E fi pµi , u, vq du dv, (25) i“0 where, for simplicity, we have set µ0 “ µS . Now, for any pu, vq P R2` , Proposition 3 implies 2 ź 2 ´ź 2 ˇ ˇ ¯ ¯ ´ ź ˇ ˚ ˇ E fi pµi , u, vq “ E E fi pµi , u, vqˇq̃, q̃ “ E E fi pµi , u, vqˇq̃ i“0 i“0 i“0 ” ı ż “ E f0 pµS , u, vq E f1 pm1 , u, vqf2 pm2 , u, vqq̃pdm1 qq̃pdm2 q M 2 2 ” ı! ” ź ı ı) ź “ E f0 pµS , u, vq π1˚ E E fi pµ̃0 , u, vq fi pµ̃0 , u, vq ` p1 ´ π1˚ q ” i“1 i“1 (26) k̄ Using the properties µS it is easy to show that E f0 pµS , u, vq ! Qk´ , 0 ř řk2 characterise řk1 ˚ that k0 ˚ ˚ where k̄ “ j“1 ζj,1 ` j“1 ζj,2 ` r“1 ζr,0 . Moreover drE f0 pµS , u, vqs k̄ dQk´ 0 pxq “ e ´γ c0 ψ0 pu`vq k̄ γ k´k̄ ck´ 0 2 ź ź τnp0q pu ` vq j,` k0 ź τqr,1 `qr,2 pu ` vq (27) r“1 `“1 j: ζj,` “0 k´k̄ for with all distinct components, and it is zero otherwise. If one notes that ś2 any x P X ˚ E f pµ̃ , u, vq vanishes when at least one of the ζr,0 ’s is non–zero, the other terms in i 0 i“1 (26) can be similarly handled and, after having marginalised with respect to pζ1˚ , ζ2˚ , ζ0˚ q, one has ÿ ÿ pN q Πk pn1 , n2 , q1 , q2 q “ π1˚ I1 pn1 , n2 , q1 ` q2 , ζ ˚˚ q ` p1 ´ π1˚ q I2 pn1 , n2 , q1 ` q2 , ζ ˚ q p˚˚q p˚q where the first sum runs over all vectors ζ ˚˚ “ pζ1˚ , ζ2˚ , ζ0˚ q P t0, 1uk and the second sum is over all vectors ζ ˚ “ pζ1˚ , ζ2˚ q P t0, 1uk´k0 . Moreover, ż8ż8 ck0 γ k´k̄ ˚˚ un1 ´1 v n2 ´1 e´p1`γqc0 ψ0 pu`vq I1 pn1 , n2 , q1 ` q2 , ζ q “ Γpn1 qΓpn2 q 0 0 28 k1 ź ˆ τnp0q pu j,1 ` vq j“1 k2 ź τnp0q pu j,2 ` vq j“1 k0 ź p0q τqj,1 `qj,2 pu ` vqdudv. j“1 One may further note that ÿ ˚˚ I1 pn1 , n2 , q1 ` q2 , ζ q “ k ÿ ÿ k̄“0 tζ ˚ : |ζ ˚ |“k̄u p˚q k1 ź ˆ τnp0q pu j,1 ck0 γ k´k̄ Γpn1 qΓpn2 q ` vq j“1 “ ck0 γ k´k̄ k̄ Γpn1 qΓpn2 q k1 ź ˆ τnp0q pu ` vq j,1 j“1 ˆ 0 τnp0q pu j,2 0 ` vq k0 ź p0q τqj,1 `qj,2 pu ` vq du dv j“1 ż8ż8 k2 ź un1 ´1 v n2 ´1 e´p1`γqc0 ψ0 pu`vq 0 0 τnp0q pu ` vq j,2 j“1 ck0 p1 ` γqk Γpn1 qΓpn2 q k1 ź un1 ´1 v n2 ´1 e´p1`γqc0 ψ0 pu`vq j“1 k ˆ ˙ ÿ k k̄“0 “ k2 ź ż8ż8 k0 ź p0q τqj,1 `qj,2 pu ` vqdudv j“1 ż8ż8 τnp0q pu j,1 un1 ´1 v n2 ´1 e´p1`γqc0 ψ0 pu`vq 0 0 ` vq j“1 k2 ź j“1 τnp0q pu j,2 ` vq k0 ź p0q τqj,1 `qj,2 pu ` vqdudv j“1 and a simple change of variable yields (18). Details on Examples 1 and 2 As for the latent nested σ–stable process, the first term in the expression of the pEPPF (18) turns out to be the EPPF of a σ0 –stable process multiplied by π1˚ “ 1 ´ σ, namely p1 ´ σq k0 2 k ź σ0k´1 Γpkq ź ź̀ p1 ´ σ0 qnj,` ´1 p1 ´ σ0 qqj,1 `qj,2 ´1 . ΓpN q `“1 j“1 j“1 As for the second summand in (18), the term I2 pn1 , n2 , q1 ` q2 , ζ ˚ q equals 29 σ0k γ k´k̄ ξσ pn1 , n2 , q1 ` q2 q Γpn1 qΓpn2 q 0 ż8ż8 un1 ´1 v n2 ´1 ˆ 0 0 exp t´γpu ` vqσ0 ´ uσ0 ´ v σ0 u dudv. pu ` vqN ´n̄1 ´n̄2 ´pk´k̄1 ´k̄2 qσ0 un̄1 ´k̄1 σ0 v n̄2 ´k̄2 σ0 The change of variables s “ u ` v and w “ u{pu ` vq, then, yields I2 pn1 , n2 , q1 ` q2 , ζ ˚ q “ σ0k´1 Γpkqγ k´k̄ ξσ pn1 , n2 , q1 ` q2 q Γpn1 qΓpn2 q 0 ż 1 n1 ´n̄1 `k̄1 σ0 ´1 p1 ´ wqn2 ´n̄2 `k̄2 σ0 ´1 w ˆ dw rγ ` wσ0 ` p1 ´ wqσ0 sk 0 pN q and the obtained expression for Πk follows. As far as the latent nested Dirichlet process is concerned, the first term in (18) coincides with the EPPF of a Dirichlet process having total mass c0 multiplied by π1˚ “ pc ` 1q´1 , i.e. k0 2 k ź rc0 p1 ` γqsk ź ź̀ 1 ¨ Γpnj,` q Γpqj,1 ` qj,2 q. 1 ` c pc0 p1 ` γqqN `“1 j“1 j“1 On the other hand, it can be seen that I2 pn1 , n2 , q1 ` q2 , ζ ˚ q equals k0 2 k ź ck0 γ k´k̄ ź ź̀ Γpnj,` q Γpqj,1 ` qj,2 q Γpn1 qΓpn2 q `“1 j“1 j“1 ż8ż8 un1 ´1 v n2 ´1 ˆ dudv. γc0 `N ´n̄1 ´n̄2 p1 ` uqn̄1 `c0 p1 ` vqn̄2 `c0 0 0 p1 ` u ` vq If p Fq pα1 , . . . , αp ; β1 , . . . , βq ; zq denotes the generalised hypergeometric series, which is defined as 8 ÿ pα1 qk . . . pαp qk z k F pα , . . . , α ; β , . . . , β ; zq :“ , p q 1 p 1 q pβ q . . . pβ q k! 1 k q k k“0 identity 3.197.1 in Gradshteyn & Ryzhik (2007) leads to rewrite I2 pn1 , n2 , q1 ` q2 , ζ ˚ q as follows ck0 γ k´k̄ Γpp1 ` γqc0 ` n1 ´ n̄1 q ξ0 pn1 , n2 , q1 ` q2 q Γpn1 qΓpp1 ` γqc0 ` N ´ n̄1 q 30 ż8 ˆ 0 un1 ´1 2 F1 pc0 ` n̄2 , n2 ; N ´ n̄1 ` c0 p1 ` γq; ´uqdu. p1 ` uqc0 p1`γq`n1 ´n̄2 On view of the formula 2 F1 pα, β; δ; zq “ p1´zq´α 2 F1 pα, δ ´β; δ; z{pz ´1qq and of the change of variable t “ u{p1 ` uq, the integral above may be expressed as ż1 tn1 ´1 p1 ´ tqc0 p1`γq`c0 ´1 ˆ 2 F1 pc0 ` n̄2 , c0 p1 ` γq ` n1 ´ n̄1 ; N ´ n̄1 ` c0 p1 ` γq; tqdt. 0 and, finally, identity 7.512.5 in Gradshteyn & Ryzhik (2007) yields the displayed closed pN q form of Πk . References Bhattacharya, A. & Dunson, D. (2012), ‘Nonparametric Bayes classification and hypothesis testing on manifolds’, J. Multivariate Anal. 111, 1–19. Filippi, S. & Holmes, C. C. (2017), ‘A Bayesian nonparametric approach for quantifying dependence beetween random variables’, Bayesian Analysis 12(4), 919–938. Gradshteyn, I. S. & Ryzhik, I. M. (2007), Tables of integrals, sums, series, and products, 7th edn, Academic Press. Hjort, N. L. (2000), Bayesian analysis for a generalized Dirichlet process prior, Technical report, University of Oslo. Holmes, C., Caron, F., Griffin, J. E. & Stephens, D. A. (2015), ‘Two–sample Bayesian nonparametric hypothesis testing’, Bayesian Analysis 10(2), 297–320. James, L. F., Lijoi, A. & Prünster, I. (2006), ‘Conjugacy as a distinctive feature of the Dirichlet process’, Scandinavian Journal of Statistics 33(1), 105–120. James, L. F., Lijoi, A. & Prünster, I. (2009), ‘Posterior analysis for normalized random measures with independent increments’, Scandinavian Journal of Statistics 36(1), 76–97. Kingman, J. F. C. (1993), Poisson processes, Oxford University Press. Lijoi, A., Nipoti, B. & Prünster, I. (2014), ‘Bayesian inference with dependent normalized completely random measures’, Bernoulli 20(3), 1260–1291. 31 Ma, L. & Wong, W. H. (2011), ‘Coupling optional Pólya trees and the two sample problem’, J. Amer. Statist. Assoc. 106(496), 1553–1565. MacEachern, S. N. (1994), ‘Estimating normal means with a conjugate style Dirichlet process prior’, Comm. Statist. Simulation Comput. 23(3), 727–741. Müller, P., Quintana, F. & Rosner, G. (2004), ‘A method for combining inference across related nonparametric Bayesian models’, J. R. Stat. Soc. Ser. B Stat. Methodol. 66(3), 735– 749. Regazzini, E., Lijoi, A. & Prünster, I. (2003), ‘Distributional results for means of random measures with independent increments’, Ann. Statist 31, 560–585. Rodrı́guez, A. & Dunson, D. B. (2014), ‘Functional clustering in nested designs: modeling variability in reproductive epidemiology studies’, Ann. Appl. Stat. 8(3), 1416–1442. Rodrı́guez, A., Dunson, D. B. & Gelfand, A. E. (2008), ‘The nested Dirichlet process’, J. Amer. Statist. Assoc. 103(483), 1131–1144. Soriano, J. & Ma, L. (2017), ‘Probabilistic multi-resolution scanning for two-sample differences’, J. R. Stat. Soc. Ser. B Stat. Methodol. 79(2), 547–572. West, M., Müller, P. & Escobar, M. D. (1994), Hierarchical priors and mixture models, with application in regression and density estimation, in ‘Aspects of uncertainty’, Wiley, Chichester, pp. 363–386. 32
10math.ST
Equivalence Classes of Optimal Structures in HP Protein Models Including Side Chains arXiv:0910.3848v1 [cs.CE] 20 Oct 2009 Martin Mann, Rolf Backofen, and Sebastian Will University of Freiburg, Bioinformatics, 79110 Freiburg, Germany, {mmann,backofen,will}@informatik.uni-freiburg.de Abstract. Lattice protein models, as the Hydrophobic-Polar (HP) model, are a common abstraction to enable exhaustive studies on structure, function, or evolution of proteins. A main issue is the high number of optimal structures, resulting from the hydrophobicity-based energy function applied. We introduce an equivalence relation on protein structures that correlates to the energy function. We discuss the efficient enumeration of optimal representatives of the corresponding equivalence classes and the application of the results. 1 Introduction Proteins are the central players in the game of life. They are involved in almost all processes in cells and organisms, comprising replication, metabolism, and movement. To be able to perform their specific functions, proteins have to adopt a certain fold or structure, which is encoded by the protein’s sequence. Thus, knowledge of a protein’s structure elucidates the mechanisms it is involved. Currently, it is not possible to calculate a protein’s functional fold from its sequence nor to simulate the whole folding process in detail. Simplified protein models are used to reduce the computational complexity. A common abstraction are lattice proteins [5, 6]. Here, the structure space a protein can adopt is discretized and allows for efficient folding simulations [8, 13]. Nevertheless, it is difficult to determine minimal energy structures, which represent the functional folds in such models. Even in the most simple Hydrophobic-Polar (HP) model [7], the optimal structure prediction problem stays computationally hard (NP-complete) [4]. Despite this complexity, a fast calculation of nonsymmetrical optimal structures in the HP model is possible using constraint programming techniques applied in the Constraint-based Protein Structure Prediction (CPSP) approach [3, 9, 10, 15]. Recently, we have introduced a significantly improved local search scheme for lattice protein folding simulations [14] using a full Miyazawa-Jernigan energy potential [11]. We take advantage of the efficient CPSP approach and initialize the folding simulations with optimal structures from the simpler HP model. This incorporates the phenomenon of hydrophobic collapse of protein structures, a driving force at the beginning of the folding process [1]. The already compact structures from the CPSP application form the starting point of the folding driven by more complex interactions. This scheme outperforms folding simulations using a standard initialization with random structures and yields better results within shorter simulation time [14]. To increase efficiency of local search methods, usually many optimization runs from different starting points are done. Since the set of all HP-optimal structures is usually too large as a starting set, we are interested in a smaller subset that still covers the structural diversity of the whole set as good as possible. We achieve this by enumerating optimal structures that maintain a given minimal distance to each other. Due to the hydrophobicity-focusing energy function, proteins in HP models show on average a huge number of optimal structures. Since polar residues do not contribute to the energy, optimal structures usually show a much higher variation in the placement of polar than hydrophobic residues. Here, we introduce an equivalence relation to partition the set of (optimal) structures into according classes. Two structures are defined to be equivalent, iff they do not differ in the placements of their hydrophobic residues. We introduce an extension to the CPSP approach that enables an efficient calculation of the number of equivalence classes of optimal structures via enumerating one representative per class. The approach is presented for backbone-only and side chain incorporating HP models. We show that a sequence’s number of representatives (later defined as core-degeneracy) is several magnitudes smaller than the overall number of all optimal structures (degeneracy). Thus, the set of optimal representatives is well placed to be used within the combined approach of CPSP and local search [14]. Furthermore, we propose another application of the equivalence classes: Since the equivalence relation is highly correlated to the HP energy function, the number of classes might be a better measure of structural stability than a sequences’ degeneracy [12]. 2 Preliminaries A lattice protein in the HP model is specified by its sequence S ∈ {H, P }n , where H and P denote hydrophobic and polar monomers, respectively. The structure positions are confined to nodes of a regular lattice L ⊆ Z3 . A valid backbone-only structure C ∈ Ln of length n is a self-avoiding walk (SAW) in the underlying lattice L, i.e. it holds connectivity ∀1≤i<n : (Ci − Ci+1 ) ∈ NL and self-avoidance ∀1≤i<j≤n : Ci 6= Cj , where NL denotes the set of distance vectors between neighbored points in L. An example is shown in Fig. 1a). The energy of a lattice protein structure is given by non-consecutive HH-contacts: ( X −1 : (Ci − Cj ) ∈ NL ∧ Si = Sj = H (1) E(S, C) = 0 : otherwise 1≤i<j≤n (i+1)<j An optimal structure minimizes the energy function. The number of optimal structures is denoted as degeneracy of a sequence and is an important measure of structural stability [12]. a) b) Fig. 1. Optimal structures of HPPHHPPPHPHHPHHPPHPHPPHHHPHHPPHPHPH in the facecentered-cubic lattice. (a) backbone-only model with energy -50, (b) side chain model with energy -55. Colors: green - H monomers, gray - P monomers, red - backbone in side chain models. Visualization by HPview from CPSP-package [10]. The CPSP-approach by Backofen and Will [3] enables the calculation of a sequence’s degeneracy without full structure space enumeration [15]. It utilizes the observation that optimal structures show a (nearly) optimal packing of H monomers. Thus, the CPSP-approach can be sketched in two major steps: 1. H-core construction: Given the number nH of H monomers from the target sequence S, all optimal packings of nH monomers are calculated. These optimal H-cores show the maximal number of contacts possible. For a fixed sequence S and the corresponding nH , we denote the set of optimal H-cores with O. The calculation of O is computationally difficult on its own and was solved by us using constraint programming [2, 3]. 2. Structure threading: Given S and O only structures are enumerated where the H monomers of S are confined to an optimal H-core O ∈ O, i.e. they are “threaded” through the H-cores. Since all O show the maximally possible number of contacts between H monomers, each resulting structure is optimal according to Eq. 1 as well. The structure threading is done by solving a Constraint Satisfaction Problem (CSP) for each O ∈ O as given below. Since step 1 depends only on the number of H monomers nH and no further property of any sequence, we can precalculate the H-cores for different nH and store them in a database. This significantly speeds up the approach and reduces the computation time to step 2, i.e. the structure threading. It might happen, that we find no appropriate structure threading for a sequence S and the according set of optimal H-cores O. Thus, we revert to the set of the best suboptimal H-cores O′ that show at least one contact less than an optimal H-core O ∈ O and iterate the procedure. Still it holds: the first successive structure threading is an optimal structure, since no H monomer packing with more contacts was found before. Further details on the CPSP approach in [3]. The CSPs solved in step 2 are given by (X , D, C), where we denote the set of variables X , their domains D, and a set of constraints C. For each monomer Si ∈ S a variable Xi ∈ X is introduced. The SAW is modeled by a sequence of binary neighboring constraints neigh(Xi , Xi+1 ) and a global alldiff(X ) to enforce the self-avoidingness. The optimal H-core O ∈ O is used to define the domains D: ∀i:Si =H : D(Xi ) = O and ∀i:Si =P : D(Xi ) = L \ O. Thus, if we find a solution of such a CSP, i.e. an assignment ai ∈ D(Xi ) for each variable that satisfies all constraints in C, it will minimize the energy function in Eq. 1, i.e. an optimal structure. 3 Representative Optimal Structures Revisiting the CSP we can see, that P monomers are constrained only by the SAW constraints. Imagine a sequence with a long tail of P monomers. Each valid placement of the subchain in front of the tail can be combined with a combinatorial number of possible SAWs of the tail. This leads to the immense degeneracy in the HP model. H Therefore, we set up an equivalence relation ∼ on structures (Eq. 2) that decomposes the set of all (optimal) structures into equivalence classes. In the following, the number of equivalence classes of optimal structures is denoted as core-degeneracy. As given by Eq. 2, structures from different equivalence classes differ in at least one H monomer placement. H C ∼ Ĉ ⇔ ∀i|Si =H : Ci = Ĉi . (2) The representative enumeration (that corresponds to core-degeneracy calculation) can be done via an extension of the CPSP approach presented in Sec. 2. Instead of calculating all optimal structures, we want to calculate only one representative per equivalence class. This has to be ensured at two stages: (I) the solutions of each single CSP for a given H-core have to be different according to Eq. 2, and (II) the solutions from two CSPs for two different H-cores have to H be different as well. The second condition (II) holds by definition, because ∼ is only defined on the H monomer placements that are constrained by different Hcores from O (differing in at least one position). In the following, we will discuss how to achieve the difference for solutions of a single CSP (I). Note that the core-degeneracy, i.e. the number of different placements of Hmonomers, or core-configurations, in optimal structures of a sequence, is not equal to the number of different H-cores, which are the sets of lattice points that are occupied by H-monomers. The latter number is easily obtained from the standard prediction algorithm, described in Sec. 2. It equals the number of cores, where the sequence is successfully threaded on. Restricted Search for Enumeration of Representatives The standard way to solve a CSP is a combination of domain filtering (i.e. constraint propagation) and depth first search. This results in a binary tree where each node represents a subproblem of the initial CSP (root) and edges represent the additional constraints added to derive the two subproblems from its predecessor node (CSP). The constraints c and ¬c added to derive the leave nodes of a certain CSP are often of the form c = (Xi ≡ d) by selecting a variable Xi from X and a value d ∈ D(Xi ) according to some heuristics. The constraint solver traverse the binary tree until a solution was found or an inconsistency of a constraint from C was detected. Therefore, a straightforward way to enumerate only one representative for each equivalence class can be sketched as follows: first, we restrict the search of the solving process onto the H associated variables. Then, we perform a single check for satisfiability, i.e. search for a single assignment of P monomer variables fulfilling all constraints in C. Thus, we get only one P monomer placement for a given H monomer assignment if any exists. The drawback of this approach is that we restrict the variable order of the search heuristics. But the performance of the CPSP approach mainly depends on the search heuristics applied to select a certain variable or value from its domain. It turned out that a mixed assignment of H and P associated variables yields the best runtimes. These heuristics can not be applied within the sketched procedure where we have to first assign H-associated variables, then P-associated ones. Thus, a lower CPSP performance is expected. But, we have to do less search which results in much faster runtimes than enumerating all optimal structures. 4 Representative Optimal Structures with Side Chains Recently, we have introduced the extension of the CPSP approach [9] to HP models including side chains [6]. Here, each amino acid of a protein sequence is represented by two monomers: Cib representing the backbone atoms, and Cis representing the atoms of the side chain. Beneath the SAW condition on the backbone monomers Cib , we constrain each side chain to be neighbored to its backbone, i.e. ∀1≤i≤n : (Cis − Cib ) ∈ NL . An example structure is given in Fig. 1b). The applied energy function E ′ exploits only HH-contacts of side chain monomers Cis : ( X −1 : (Cis − Cjs ) ∈ NL ∧ Si = Sj = H ′ s E (S, C ) = (3) 0 : otherwise 1≤i<j≤n Therefore, the side chain models show an even higher degeneracy than the backbone-only models discussed so far, since all backbone monomers Cib are H unconstrained by the energy function as well. Thus, an equivalence relation ≈ that focuses on the monomers constrained by the energy function is even more H striking in HP models including side chains. The relation ≈ is given by H (C b , C s ) ≈ (Ĉ b , Ĉ s ) ⇔ ∀i:Si =H : Cis = Ĉis (4) Therefore, we will enforce that structures from one equivalence class show the same H monomer side chain positioning. The CPSP approach for HP models including side chains differs only in the CSP formulation from the original approach for backbone-only models [9]. This allows for the application of the same approach discussed in the previous section to enumerate non-equivalent optimal structure representatives. Thus, we restrict search to the H associated side chain variables first and only check for satisfiability on the remaining variables. Fig. 2. Backbone-only models : Histogram of core-degeneracy (green) and degeneracy (red) with cut-off ≤ 106 . (Plots refer to 3D cubic lattice and sequence length 27.) 5 Results and Discussion We exemplify the enumeration of representatives for backbone-only and side chain models. We focus on the comparison of the resulting core-degeneracy of a sequence and its overall number of optimal structures, i.e. degeneracy, because we are interested in a reduced set of optimal structures, e.g. for local search initialization (see introduction). All following results are given for HP-sequences of length 27 in 3D cubic lattice. Since the enumeration and check of all 227 sequences (> 108 ) is computationally not feasible, we restrict each study to a large randomly chosen subset of 105 and 104 sequences, respectively. The program HPrep implements the approach from section 3. It is integrated into the CPSP-tools package [10] version 2.4.0 and available online1 . As discussed in Sec. 2, the structure threading step of the CPSP approach screens through a precomputed list of appropriate H-cores in decreasing number of contacts stored in a database. Therefore, it might occur that the available list from the database is exceeded without any solution, i.e. no optimal structure was computed. Still, the energy of the last H-core tried is a close lower bound on the energy this sequence can adopt. In the following, B denotes the subset of sequences where the current H-core database is not sufficient and thus the CPSP approach can give only a lower bound for now. The number of sequences in B is quite small. It is reasonable to assume that the degeneracy distribution among B is the same as for the remaining sequences or on average even higher. Backbone-only models We tested 105 random sequences in the backbone-only model in the 3D cubic lattice. Here, only 66% show a degeneracy below 106 . B comprises about 4% of the sequences. The remaining 30% can adopt even more than 106 structures with minimal energy. 1 http://cpsp.informatik.uni-freiburg.de Fig. 3. Models with side chains : Histogram of the core-degeneracy (green) and degeneracy (red) with cut-off ≤ 106 . Note: only 408 sequences out of 104 showed a degeneracy below 106 as given in the text. (Plots refer to 3D cubic lattice and sequence length 27.) Figure 2 summarizes the results: in red the degeneracy and in green the coredegeneracy distribution with cut-off 106 is presented. Thus, in red the degeneracy distribution comprises 66% of the sequences as given above. In contrast, all sequences show a number of optimal equivalence classes below 106 (in green)! The average degeneracy is reduced from 124800 (with cutoff 106 ) to a mean core-degeneracy of 4856. This reduction within two orders of magnitude results in reasonably small sets of representative structures e.g. to be utilized in local search initializations. Furthermore, the enumeration of representatives is on average six times faster than the enumeration of all optimal structures with a mean runtime of 2 seconds (Opteron 2356 - 2.3 GHz). This increase of small sets of representatives compared to the complete sets of optimal structures shows the advantage of the approach: core-degeneracy does not show the huge combinatorial explosion of degeneracy. This gets even more striking in HP models including side chains, as shown in the next section. Models including side chains The degeneracy in HP models including side chains is much higher than for backbone-only models. This results from the simple energy function (Eq. 3) that does not constrain the backbone or P monomers. Therefore, an immense number of optimal structures is present. From the 104 HP-sequences tested only 408 show a degeneracy below 106 . B comprises again about 3.1% of the sequences. When investigating core-degeneracy the picture changes completely: All of the sequences tested have less than 106 representatives. Figure 3 summarizes the distribution. The average number of representatives is about 1550, which is again at least three orders of magnitude smaller than the average degeneracy. Since we have only a very rough lower bound of 106 on the average degeneracy (due to the cut-off), the real reduction ratio is expected to be even higher. 6 Conclusions The introduced equivalence relations for HP models enables a energy function driven partitioning of structures. The presented CPSP approach extension en- ables an efficient calculation of representatives for all equivalence classes of optimal structures, i.e. calculation of a sequence’s core-degeneracy. Using our implementation HPrep, we showed that sequences show several orders of magnitudes less optimal equivalence classes than optimal structures. This is most striking in models including side chains. The sets of representatives are usually small. Furthermore, representatives show different hydrophobic core arrangements. Therefore, they are well placed to be used for the initialization of local search procedures that utilize more complex energy functions [14]. This emulates the hydrophobic collapse in the folding process. Since a sequence’s degeneracy is a measure of structural stability [12], we propose another application of our approach. The core-degeneracy might be used as a more reasonable measure of stability in the HP model compared to degeneracy. It ignores the HP model specific degeneracy blow-up due to unconstrained subchains of P monomers (see section 3). Thus, a structural stability analysis could be based on the presented equivalence classes instead of all possible structures. References 1. Vishwas R. Agashe, M. C. R. Shastry, and Jayant B. Udgaonkar. Initial hydrophobic collapse in the folding of barstar. Nature, 377:754–757, 1995. 2. Rolf Backofen and Sebastian Will. Optimally compact finite sphere packings — hydrophobic cores in the FCC. In Proc of CPM’01, volume 2089 of LNCS, pages 257–272, 2001. 3. Rolf Backofen and Sebastian Will. A constraint-based approach to fast and exact structure prediction in three-dimensional protein models. J Constraints, 11(1):5– 30, Jan 2006. 4. Bonnie Berger and Tom Leighton. Protein folding in the hydrophobic-hydrophilic (HP) model is NP-complete. J Comp Biol, 5:27–40, 1998. 5. K. A. Dill, S. Bromberg, K. Yue, K. M. Fiebig, D. P. Yee, P. D. Thomas, and H. S. Chan. Principles of protein folding – a perspective of simple exact models. Protein Science, 4:561–602, 1995. 6. S. Bromberg K. A. Dill. Side-chain entropy and packing in proteins. Protein Sci, 3(7):997–1009, 1994. 7. Kit F. Lau and Ken A. Dill. A lattice statistical mechanics model of the conformational and sequence spaces of proteins. Macromol., 22(10):3986–3997, 1989. 8. Martin Mann, Daniel Maticzka, Rhodri Saunders, and Rolf Backofen. Classifying protein-like sequences in arbitrary lattice protein models using LatPack. HFSP Journal, 2(6):396, 2008. 9. Martin Mann, Cameron Smith, Mohamad Rabbath, Marlien Edwards, Sebastian Will, and Rolf Backofen. CPSP-web-tool : a server for 3D lattice protein studies. Bioinformatics, 25(5):676–677, 2009. 10. Martin Mann, Sebastian Will, and Rolf Backofen. CPSP-tools - exact and complete algorithms for high-throughput 3D lattice protein studies. BMC Bioinformatics, 9:230, 2008. 11. S. Miyazawa and R. L. Jernigan. Residue-residue potentials with a favorable contact pair term and an unfavorable high packing density term, for simulation and threading. J Mol Biol, 256(3):623–44, 1996. 12. David Shortle, Hue S. Chan, and Ken A. Dill. Modeling the effects of mutations on the denatured states of proteins. Prot Sci, 1:201–215, 1992. 13. Kathleen Steinhöfel, Alexandros Skaliotis, and Andreas A. Albrecht. Stochastic protein folding simulation in the d-dimensional HP-model. In Proc of BIRD’07, pages 381–394. Springer, 2007. 14. Abu Dayem Ullah, Leonidas Kapsokalivas, Martin Mann, and Kathleen Steinhöfel. Protein folding simulation by two-stage optimization. In Proc. of ISICA’09, CCIS, Wuhan, China, Oct 2009. Springer. (accepted). 15. Sebastian Will and Martin Mann. Counting protein structures by DFS with dynamic decomposition. In Proc of WCB’06, pages 83–90, 2006.
5cs.CE
Nearest Common Ancestors: Universal Trees and Improved Labeling Schemes arXiv:1707.08807v1 [cs.DS] 27 Jul 2017 Fabian Kuhn∗ Konstantinos Panagiotou† Pascal Su‡ Abstract We investigate the nearest common ancestor (NCA) function in rooted trees. As the main conceptual contribution, the paper introduces universal trees for the NCA function: For a given family of rooted trees, an NCA-universal tree S is a rooted tree such that any tree T of the family can be embedded into S such that the embedding of the NCA in T of two nodes of T is equal to the NCA in S of the embeddings of the two nodes. As the main technical result we give explicit constructions of NCA-universal trees of size n2.318 for the family of rooted n-vertex trees and of size n1.894 for the family of rooted binary n-vertex trees. A direct consequence is the explicit construction of NCA-labeling schemes with labels of size 2.318 log2 n and 1.894 log2 n for the two families of rooted trees. This improves on the best known such labeling schemes established by Alstrup, Halvorsen and Larsen [SODA 2014]. Keywords: Rooted Trees, NCA, Nearest Common Ancestor, Lowest Common Ancestor, Universal Trees, Labeling Schemes, Embedding Schemes 1 Introduction The nearest common ancestori) (NCA) of two vertices u and v of a rooted tree T is the first common vertex of the paths connecting u and v to the root of T . Finding the nearest common ancestor appears as an essential operation in many algorithms and applications (see for example the survey by Alstrup, Gavoille, Kaplan, and Rauhe [2]). NCA-Universal Trees. The present paper introduces the notion of NCA-universal trees as a novel tool to study and algorithmically deal with the NCA function in rooted trees. We define an NCA-universal tree S for a family of rooted trees T as such that every tree T ∈ T can be embedded into S such that the NCA function is preserved by the embedding. More formally, an embedding of T into S is an injective mapping ϕT of V (T ) into V (S) such that the embedding function ϕT and the NCA function commute. ∗ Department of Computer Science, University of Freiburg, Germany. Department for Mathematics, University of Munich, Germany ‡ Department of Computer Science, ETH Zurich, Switzerland. i) In the literature, the nearest common ancestor of two vertices in a rooted tree is sometimes also referred to as the lowest or least common ancestor (LCA). † 1 NCA-Labeling Schemes. As an immediate application of an NCA-universal tree S for a family T of rooted trees, S directly implies an NCA-labeling scheme [2] for the family T . Generally, a labeling scheme is a way to preprocess the structure of a graph to later allow simple and fast queries. A labeling scheme consists of an encoder and a decoder, where the encoder must be able to label a family of graphs such that the decoder can answer queries, given just the labels and no additional information about the underlying graph. More specifically, an NCA-labeling scheme assigns a unique label to each node of a rooted tree T such that given the labels of two vertices u and v of T , it is possible to compute the label of the NCA of u and v in T . If an NCA-univeral tree S for a family T of rooted trees is given, we can get an NCA-labeling scheme for T as follows. Let |S| be the number of vertices of S and assume that the vertices of S are labeled from 0 to |S| − 1 in a arbitrary fixed way. Given an embedding of a tree T ∈ T into S, we then get the labeling of a vertex v of T by using the label of the vertex x of S to which v is embedded. The size of the labels (in bits) of the labeling scheme is therefore exactly dlog |S|e. We remark that throughout the paper, all logarithms are to base 2. Contribution. We show that the family of all rooted trees with at most n vertices has an NCAuniversal tree of size O(n2.318 ) and that the family of all binary rooted trees with at most n vertices has an NCA-univeral tree of size O(n1.894 ). This implies that the families of rooted n-vertex trees and of rooted n-vertex binary trees have labeling schemes with labels of size 2.318 log n and 1.894 log n, respectively. This improves on the best previous NCA-labeling schemes that were developed by Alstrup, Halvorsen and Larsen [3] and which require labels of size 2.772 log n for general rooted trees and of size 2.585 log n for binary rooted trees. In [3], it is also shown that any NCA-labeling scheme for general n-vertex rooted trees requires labels of size at least 1.008 log n. As we show how to explicitly construct the NCA-universal trees, our labeling schemes are constructive. Note that the best NCA-labeling schemes of [3] are not constructive and that the best previous constructive NCA-labeling scheme for n-vertex rooted trees requires labels of size 3 log n. Further, our NCA-labeling schemes are efficient, the embedding of a rooted tree into the constructed NCA-universal tree can be computed efficiently and a single query can be answered in time O(log2 n) (O(log n) for binary trees). We believe that our new NCA-labeling schemes are not only interesting because they improve upon the best existing schemes, but also because our approach leads to more intuitive and significantly simpler constructions. Related Work. Graph labeling schemes are an elegant way to store structural information about a graph. As every vertex is only assigned a small label, the information is stored in a completely distributed way and graph labelings therefore are particularly interesting in a distributed context, where labeling schemes are used for various kinds of graph queries [10, 15]. In addition, labeling schemes can be used in a context where extremely large graphs are processed and where accessing the data is expensive. To answer a pair-wise query, only the two labels of the corresponding vertices need to be accessed. The first labeling schemes that appear in the literature are adjacency labeling schemes (given the labels of two vertices, determine whether the vertices are adjacent). They were introduced among others by Breuer [6] and Folkman [7]. In the context of adjacency labeling schemes, it is well known that they are tightly connected to induced universal graphs. A graph G is an induced universal graph for a graph family H if G contains every graph H ∈ H as an induced subgraph. Induced universal graphs were first described by Rado [17] and Kannan et al. [14] noted the equivalence between adjacency labeling schemes and induced universal graphs. When considering the family of rooted trees we are interested in different queries such as whether 2 a vertex is an ancestor of the other. Ancestry labeling has been studied and labeling schemes of size log(n) + Θ(log(log(n))) are known to be tight [1]. If the tree has low depth, then a scheme of size log(n) + 2 log(d) + O(1) is known, where d is the depth of the tree [8]. For NCA-labeling schemes, a linear-size labeling scheme that answers queries in constant time was introduced by Harel and Tarjan in [12, 13]. In the following, there was a series of significant improvements in [16], [4], [9], [18], [5], [2] and most recently in [3]. In particular in [3], a lower bound for NCA labeling schemes of 1.008 log n is shown, which separates NCA-labelings, that need labels of size log n + Ω(log n), from ancestry labeling schemes, where labels of size log n + O(log log n) are sufficient. There is concurrent work by Gawrychowski and Łopuszański [11] who reach the exact same bounds for labeling schemes and construct almost identical universal trees as we do. The proof method is a bit different and in their paper they also have lowerbounds for the size of a universal tree for the NCA function. Outline. The rest of the paper is organized as follows. In the remainder of this section, we first formally define the problems and state our results in Sections 1.1 and 1.2. In Section 2, we first prove a simpler upper bound of O(n2 ) on the size of NCA-universal trees for the family of binary rooted trees. We extend the construction of Section 2 to obtain the stronger and more general results stated above in Section 3. Finally, in Section 4, we sketch how to efficiently implement our labeling scheme and in Section 5, we conclude the paper and discuss some open issues. 1.1 Definitions We next define the necessary graph-theoretic concepts and notation and we in particular formally introduce the notion of universal trees for the NCA function. In the following let Tn be the family of unlabeled rooted trees with at n vertices and Bn the family of unlabeled rooted binary trees with n vertices. Definition 1.1. In a rooted tree T = (V, E) a vertex u is an ancestor of a vertex v if u is contained in the (unique) path from v to the root of T . Note that according to the above definition, a node u is an ancestor of itself. Definition 1.2. Let T = (V, E) be a rooted tree. For a pair of vertices u and v their nearest common ancestor (NCA) ncaT (u, v) is the unique common ancestor that is furthest from the root of T . With this notation at hand we can define the notion of a universal tree for NCA. Definition 1.3. A rooted tree S is called an NCA-universal tree for a family of rooted trees T , if for every tree T ∈ T there is an embedding function ϕT : V (T ) 7→ V (S) such that ϕT commutes with the NCA function, i.e., for all u, v ∈ V (T ), ϕT (ncaT (u, v)) = ncaS (ϕT (u), ϕT (v)). Hence, the embedding has the property that the NCA of two nodes u and v of T is mapped to the NCA of ϕT (u) and ϕT (v) in S. Note that we do not require the root of T to be embedded to the root of S. In the following, a rooted tree that is universal for the NCA function is also called an NCA-universal tree. Definition 1.4. An NCA-labeling scheme for a family of rooted trees T is a pair of functions called the encoder (f ) and decoder (g) with f : {v|v ∈ T ∈ T } 7→ [m] and g : [m] × [m] 7→ [m] satisfying the following properties. 3 i) for every T ∈ T and every u, v ∈ V (T ), f (u) 6= f (v) and ii) for every T ∈ T and every u, v ∈ V (T ), g(f (u), f (v)) = f (ncaT (u, v)). For a node v ∈ T ∈ T , f (v) is called the label of v. The size of the labeling scheme defined by f and g is dlog me, i.e., the number of bits required to store the largest label. Given an NCA-universal tree S for a family of rooted trees T , we directly obtain an NCA-labeling scheme for T . Observation 1.5. Let S be an N -vertex rooted tree that is universal for the NCA function and the family T of rooted trees. Then, there exists an NCA-labeling scheme of size dlog N e for T . Proof. We assign unique names from 0 to N − 1 to the N vertices of S. Consider a tree T ∈ T and let f be an embedding of T into S. The label of a vertex v of T is the name assigned to vertex f (v) of S. Given the labels xu ∈ {0, . . . , N − 1} and xv ∈ {0, . . . , N − 1} of two nodes u and v of T , the decoder outputs the name xw ∈ {0, . . . , N − 1} of the NCA of the vertices u0 and v 0 with names xu and xv in S. 1.2 Main Results Our main result is an explicit construction of universal trees for the families of all trees and binary trees with n vertices. The same bounds can be found in the concurrent work of Gawrychowski and Łopuszański [11]. Theorem 1.6. Let n ∈ N. Then: • There is a rooted tree Sn of size less than n2.318 that is universal for the NCA function and the set Tn of rooted trees of size n. • There is a rooted tree Snbin of size less than n1.894 which is universal for the NCA function and the set Bn of rooted binary trees of size n. The proof can be found in Section 3. The direct implication of this for labeling schemes is summarized in the following statement. Theorem 1.7. For any n ∈ N there exists an N CA-labeling of size less than 2.318 log n and an N CA-labeling for binary trees of size less than 1.894 log n. This is an improvement of the current best known bound of 2.772 log n from [3]. Further for the specific case of binary trees of particular interest is that the constant is now below 2 and therefore will likely not be an integer. Proof of Theorem 1.7. This is exactly what we have shown in Observation 1.5 and therefore follows from Theorem 1.6. Although this is just an existential proof, from the construction we will see later, it is clear that a reasonably fast algorithmic implementation is possible and we will give a sketch in Section 4. 4 Figure 1: Recursive structure of the basic NCA-universal tree construction for rooted binary trees of size at most n. 2 Basic Universal Tree Construction The NCA-universal trees of Theorem 1.6 are constructed recursively. Before proving the general statements of Theorem 1.6, we describe a simpler, slightly weaker construction that provides an NCA-universal tree of size O(n2 ) for n-vertex binary trees. The full constructions required to prove Theorem 1.6 appears in Section 3. Theorem 2.1. For any n ∈ N there exists a rooted tree Sn of size less than n2 which is universal for the NCA function and the rooted binary trees of size at most n. Our recursive universal tree construction requires two kinds of NCA-universal trees. In addition to ordinary unlabeled rooted trees, we also need to define NCA-universal trees for the family of rooted binary trees where one leaf node is distinct (marked). Recall that Bn denotes the set of all n-vertex unlabeled rooted binary trees, so let Bn0 denote the family of unlabeled rooted binary trees on at most n vertices and with one marked leaf. Definition 2.2. A rooted tree S 0 with one marked leaf vertex w is called a NCA-universal tree for a family of rooted trees T 0 with one marked leaf if for every tree T ∈ T 0 , there exists an embedding function ϕT : V (T ) 7→ V (S 0 ) that maps the marked leaf of T to the marked leaf w of S 0 and where ϕT commutes with the NCA function, i.e., ϕT (ncaT (u, v)) = ncaS (ϕT (v), ϕT (u)) for all u, v ∈ T . As in Definition 1.3, we do not require that the root of T is mapped to the root of S 0 . Overview of the Construction. The construction of the NCA-universal tree for binary rooted trees is done recursively as illustrated in Figure 1. The universal tree Sn for binary trees of size at most n consists of three NCA-universal trees for binary trees of size at most n/2, where one of these three universal trees needs to work for the more general family trees with one marked leaf vertex. Universal trees for the family of n-vertex trees with a marked leaf are constructed recursively in a similar way. They consist of two NCA-universal trees for n/2-vertex binary trees with a marked leaf and of a single NCA-univeral tree for ordinary n − 1-vertex binary trees (see Figure 1). In order to show that the recursive construction of Figure 1 results in an NCA-universal tree we need to argue that any n-vertex binary tree T can be embedded. To achieve this, we show that any 5 rooted tree T has a vertex v such that v splits T into three subtrees of size at most n/2. Vertex v is then embedded to the vertex marked in red in the left part of Figure 1. The three subtrees of T induced by v are then embedded recursively into the three parts of the universal tree construction. Further, we need to show that any rooted tree T with a marked leaf can be partitioned in a similar way to be consistent with the recursive structure in the right part of Figure 1. In the following, we first give the basic technical lemmas required to partition n-vertex trees T into the required smaller subtrees. Based on these partitioning results, we the analyze the recursive NCA-universal tree construction in more detail and prove Theorem 2.1. As the same partitioning lemmas will also be needed in the general NCA-universal tree constructions in Section 3, they are stated more generally than what we require for the simple construction of the present section. Lemma 2.3. For every rooted n-vertex tree T and for every parameter λ ∈ (0, 1], there exists a vertex v ∈ V (T ) such that removing the edges from v to its children splits the tree into components such that each component rooted at a child of v has size at most b(1 − λ) · nc and such the remaining component containing v and the root of T has size at most dλ · ne. Proof. We determine v using the following simple iterative procedure. We initialize v to be the root of T . We stop the procedure as soon as v satisfies the conditions of the lemma. For some vertex u of T , let size(u) be the number of vertices in the subtree rooted at u. If v does not split the tree as required, we let w be the child vertex of v that maximizes size(w) and we set v := w. Since v goes from being the root of T to being a leaf of T during this process, the component containing the root goes from being of size 1 to a set of size n. We claim that for the last vertex v where the connected component of the root is still of size at most dλ · ne, the lemma holds. Clearly, the component with the root is of size at most dλ · ne and it thus suffices to show that all the subtrees of v are of size at most b(1 − λ)nc. Assume for contradiction that v has a child w such that size(w) ≥ 1 + b(1 − λ)nc. Then, removing all subtrees of w from T would result in a component of size at most n − b(1 − λ)nc = dλne and thus v would not be the last vertex for which the connected component of the root is of size at most dλne. For trees with a marked leaf we can get a similar tree splitting lemma. Lemma 2.4. Given a rooted n-vertex tree T with one marked leaf vertex w and a parameter λ ∈ 1 (0, 1]. If n ≥ 1−λ , there exists a vertex v ∈ V (T ) such that when removing the edges connecting v to its children, T is split into components satisfying the following properties. The component containing the root of T and vertex v has size at most dλne, the component containing the marked leaf w has size at most b(1 − λ)nc, and all other components have size at most n − 1. Proof. Let r bet the root vertex of T . We choose v to be last vertex on the path from r to w such that when removing the subtrees of v, the remaining component has size at most dλne. dλne < n, so v cannot be a leaf and thus v 6= w. Let v 0 be the root of the subtree of v containing w. To prove the lemma, it suffices to show that the subtree rooted at v 0 has size size(v 0 ) ≤ b(1 − λ)nc. For the sake of contradiction, assume that size(v 0 ) ≥ b(1 − λ)nc + 1. In this case, the total size of all subtrees of v 0 is at least b(1 − λ)nc and thus removing all subtrees of v 0 would leave a component of size at most dλne. This contradicts the assumption that v is the last vertex on the path from r to w for which this is true. Proof of Theorem 2.1. For every integer n ≥ 1, we show how to construct an NCA-universal tree Sn for the family Bn of n-vertex rooted binary trees and an NCA-universal tree Sn0 for the family Bn0 of n-vertex binary rooted trees with one marked leaf. We will prove by induction on n that |Sn | ≤ n2 and that |Sn0 | ≤ 2n2 − 1 for all n ≥ 1. 6 Figure 2: Embedding of an n-vertex binary tree T into Sn . For the induction base, note that S1 and S10 clearly need to only consist of a single vertex and we thus have |S1 | = |S10 | = 1. Thus, the bounds on |Sn | and |Sn0 | hold for n = 1. For the induction step, assume that n ≥ 2 and that |Sk | ≤ k 2 and |Sk0 | ≤ 2k 2 − 1 for all 1 ≤ k < n. We build the two NCA-universal trees Sn and Sn0 by using smaller NCA-universal trees as given in Figure 1. That 0 is, Sn is composed of one copy of Sdn/2e and two copies of Sbn/2c and Sn0 is composed of one copy 0 0 of Sdn/2e , Sbn/2c , and Sn−1 . We need to show that the constructed trees Sn and Sn0 are in fact NCA-universal trees and that they satisfy the required size bounds. We first show that the trees are or the right size. Using the induction hypothesis, we have Sd0 n e + 2 · Sb n2 c ≤ 2 · |Sn0 | = Sd0 n e + Sb0 n c 2 2 2 l n m2 j n k2 −1+2· ≤ n2 , and 2 2 l n m2 j n k2 + |Sn−1 | ≤ 2 · −1+2· − 1 + (n − 1)2 ≤ 2n2 − 1. 2 2 |Sn | = It thus remains to prove that the recursive construction of Sn and Sn0 allows to find a proper embedding for every T ∈ Bn into Sn and every T 0 ∈ Bn into Sn0 , respectively. To show this, we use Lemma 2.3 and Lemma 2.4. We first show how to construct an embedding ϕT of a binary n-vertex tree T ∈ Bn into Sn . For this purpose, we apply Lemma 2.3 with parameter λ = 1/2 to tree T . Let v ∈ V (T ) be the vertex of T that splits the tree such that the part containing the root of T and v has size at most dn/2e and such that all other components have size at most bn/2c. For the embedding take the splitting vertex v ∈ T given by Lemma 2.3 which will be embedded to the marked vertex of the copy of Sd0 n e (cf. Figure 1). Consider the three components 2 of T after splitting. By the induction hypothesis there is an embedding function to embed the child components (components below the splitting vertex v) into the two copies of Sb n2 c . Then we take the component with the root and the marked vertex to get a tree with a single marked leaf of size at most d n2 e and the induction hypothesis again provides with an embedding function of this component into Sd0 n e . The embedding is depicted in Figure 2. 2 To see that the NCA-function for any two vetices u1 and u2 and the described embedding function ϕT commute, we need that ncaSn (ϕT (u1 ), ϕT (u2 )) = ϕT (ncaT (u1 , u2 )). This can be verified through the following case analysis considering the three components after the splitting process. If u1 and u2 are in the same component, then we have embedded them into the same subtree and by the induction hypothesis, we have ϕT (ncaT (u1 , u2 )) = ncaSn (ϕT (u1 ), ϕT (u2 )). If u1 and u2 are in different child components, we have ncaT (u1 , u2 ) = v and the embedding is therefore also correct because is embedded to the vertex marked in red in Figure 2. Finally, if u1 is from the 7 Figure 3: Embedding of an n-vertex binary tree T 0 with a marked leaf into Sn0 . component containing the root and vertex v and u2 is a vertex from a child component, we have ncaT (u1 , u2 ) = ncaT (u1 , v) and similarly ncaSn (ϕT (u1 ), ϕT (u2 )) = ncaSn (ϕT (u1 ), ϕT (v)) and the embedding is therefore again correct by the induction hypothesis (applied to the partial embedding 0 into the subtree Sdn/2e ). For the family of n-vertex binary trees with a marked leaf, the embedding into the recursively constructed tree Sn0 (cf. Figure 1) works in similar way. Let T 0 be a binary tree of size at most n and with a marked leaf. We apply Lemma 2.4 with parameter λ = 1/2 to T 0 to obtain a vertex v ∈ V (T 0 ) that splits T 0 into a) a subtree of size at most dn/2e that contains the root of T 0 and that contains v as a leaf vertex, b) a subtree size at most bn/2c that is rooted at a child of v and contains the marked leaf of T 0 , and c) a (possibly empty) subtree of size at most n − 1 rooted at a child of v. The tree T 0 is embedded into Sn0 by embedding vertex v to the center red node separating the three recursive subtrees in Figure 1. The three subtrees resulting after splitting T 0 are embedded into the 0 0 three recursively constructed subtrees Sdn/2e , Sbn/2c , and Sn−1 in the natural way. The proof that the embedding is correct is done in the same way as for the embedding of T into Sn . The details of the embedding are illustrated in Figure 3. 3 General Universal Tree Construction The proof of the basic construction was handled in detail. We now adjust the construction to improve on the exponent in the size of Sn and to deal with general rooted trees. Throughout the section, we omit floor and ceiling functions. They do not change the calculations significantly, but hinder the readability of the proof. 8 Figure 4: General recursive construction of an NCA-universal tree for n-vertex rooted binary trees. Proof of Theorem 1.6. As suggested in Lemma 2.3 and Lemma 2.4, the adjustment of the basic construction can be made by choosing λ 6= 1/2 for the size of the splitted components. We start with the binary tree case. We apply the same induction as in the proof of Theorem 2.1. In the general case, we prove that |Sk | ≤ k β and |Sk0 | ≤ c · k β ∀ 1 ≤ k < n and some constants c and β that will be determined later. 0 and attach copies of S(1−λ)n and Sn/2 to the marked For the construction of Sn we take Sλn vertex for some λ ∈ (0, 1/2]. The construction of Sn0 remains the same as in Section 2. For an illustration, see Figure 4. Note that although in Lemma 2.3 the child components can be of size (1 − λ) · n, the two components together can have size at most n, so the smaller of the components is always of size at most n/2. We obtain 0 | + |S |Sn | ≤ |Sλn | ≤ c(λn)β + ((1 − λ)n)β + (1−λ)n | + |S n 2 β β |Sn0 | ≤ |S 0n | + |S 0n | + |Sn | ≤ c n2 + c n2 + nβ 2  n β 2 ! ≤ nβ ! ≤ cnβ . 2 We would like to choose β as small as possible. Note that with the second inequality, we obtain  n β  n β c +c + nβ ≤ cnβ 2 2 and we thus get that c ≥ 1−21β−1 . In our construction we are allowed to freely choose λ ∈ (0, 1/2]. We can thus choose β and λ such that β is minimized and following inequality is still satisfied: (1 − λ)nβ + cλnβ + · nβ ≤ nβ . 2 To achieve this, we choose λ = 0.296149... and the corresponding β ≤ 1.89311.... This proves the claim of Theorem 1.6 about binary rooted trees. In a general tree, a vertex can have many children. Therefore we adjust the construction to deal with this fact as shown in Figure 5. Take any λ ∈ (0, 1/2]. Let Sn be composed of a copy of 0 and attached to the marked vertex of that tree copies of S Sλn (1−λ)n , Sn/2 , Sn/3 , Sn/4 , etc., up to 0 0 S1 . Similarly, let Sn be composed of a copy of Sn/2 and attached to the marked vertex of that tree 0 , S , S copies of Sn/2 n n/2 , Sn/3 , Sn/4 , etc., up to S1 . 9 Figure 5: Construction of an NCA-univeral trees Sn for general rooted trees. For the embedding, note that we can sort the child components by size. Lemma 2.3 states that any child component is of size at most (1−λ)·n. In addition, the total size of all components cannot add up to more than the entire tree of size n. This implies that after ordering the components by size, the ith child component without a marked vertex is of size at most n/i. With the induction hypothesis that |Sk | ≤ k β and |Sk | ≤ c · k β , the recursion gives ! Pn β 0 |+ β β ≤ nβ n |Sn | ≤ |S(1−λ)n | + |Sλn i=2 |S i | ≤ ( (1 − λ)n) + c(λn) + (ζ(β) − 1)n , ! β β P |Sn0 | ≤ |S 0n | + |S 0n | + ni=1 |S ni | ≤ cnβ , ≤ c n2 + c n2 + ζ(β)nβ 2 2 where ζ(β) is the Riemann zeta function (ζ(β) = −β i≥1 i P ). Again we can deduce from the second ζ(β) . 1−2β−1 inequality that c ≥ By using λ = 0.341395..., we get that β ≤ 2.31757... . In both constructions, the fact that the NCA function and the embedding function commute follows in the same way as in the proof of Theorem 2.1. This concludes the proof of Theorem 1.6. 4 Implementation of the NCA-Labeling Scheme Theorem 4.1. The labeling schemes described in Section 3 can be constructed efficiently. Further, given two labels, the label of the nearest common ancestor can be determined in O(log2 n) time (in O(log n) time in the binary tree case). Proof Sketch. For R ∈ R we write [R] := {x : 1 ≤ x ≤ R}. In order to assign labels to the vertices of Sn we proceed as follows. Set s(n) = nβ , and recall that |Sn | ≤ s(n) and |Sn0 | ≤ cs(n). Moreover, as depicted in Figure 5 the tree Sn is composed out of the n + 1 trees 0 T0 = Sλn T1 = S(1−λ)n (1) T` = Sn/` , where 2 ≤ ` ≤ n. Define the corresponding counting sequence t−1 = 0 t0 = c s(λn) t1 = t0 + s((1 − λ)n) t` = t`−1 + s(n/`), where 2 ≤ ` ≤ n. 10 (2) For Sn0 we proceed similarly. As depicted in Figure 5 Sn0 is composed out of the n + 2 trees 0 0 T−1 = Sn/2 0 T00 = Sn/2 T`0 (3) = Sn/` , where 1 ≤ ` ≤ n. and the corresponding counting sequence is given by t0−2 = 0 t0−1 = cs(n/2) t00 = t−1 + cs(n/2) (4) t0` = t`−1 + s(n/`), where 1 ≤ ` ≤ n. Given these sequences, in order to assign labels to the vertices in Sn we assign to the vertices of Ti , 0 ≤ i ≤ n the labels in [ti ] \ [ti−1 ]. The assignment is performed recursively, in the sense that as soon the labels in Ti , 0 ≤ i ≤ n are assigned, they are translated by an additive dti−1 e, so that they all lie (with room to spare) in the required set [ti ] \ [ti−1 ]. The assignment is performed analogously for Sn0 , where we use the corresponding counting sequence instead. Given the label of a vertex in Sn , its location in the tree can be found with this preprocessing in O(log2 n) time. Indeed, in every step we have to decide in which of the at most n+2 subtrees we have to branch to; however, this can be decided with binary search on the sequences (ti )0≤i≤n or (t0i )−1≤i≤n . As the depth of the recursive construction of Sn is O(log n), the claim follows. 5 Conclusion We introduced NCA-universal trees and gave simple recursive constructions of such trees that in particular lead to improved NCA-labeling schemes for rooted trees. The paper leaves several interesting open questions. The current upper bound of 2.318 log n bits per label is still quite far from the 1.008 log n-bit lower bound proven in [3] and it remains an intriguing open problem to close this gap. In addition, given that NCA-universal trees provide an intuitive way to argue about NCA-labeling schemes, it is natural to ask whether the approach can lead to optimal NCA-labeling schemes or whether every NCA-labeling scheme for a given tree family can be turned into an equivalent one that can be characterized by an NCA-universal tree for the tree family. The following observation shows that NCA-universal trees are equivalent to a certain well-structured class of NCA-labeling schemes. We call an NCA-labeling scheme consistent if any three labels can occur together in some tree. More formally, an NCA-labeling scheme is called consistent if it satisfies the following three properties for any 3 possible labels x, y, and z. In the following, g is the decoder function. (I) If g(x, y) = z, then g(x, z) = z and g(y, z) = z (i.e., if z is the NCA of x and y, then z is an ancestor of x and y) (II) If g(x, y) = y and g(y, z) = z, then g(x, z) = z (i.e., if y is an ancestor of x and z an ancestor of y, then also z is an ancestor of x) (III) If g(x, y) = y and g(x, z) = z, then g(y, z) ∈ {y, z} (i.e., if y and z are ancestors of x, then z is an ancestor of y or y is an ancestor of z) 11 Theorem 5.1. Every NCA-universal tree S for a given family T of trees leads to a consistent NCA-labeling scheme for T with labels of size dlog |S|e. Conversely, every consistent NCA-labeling scheme for T and with `-bit labels induces an NCA-universal tree of size 2` for T . Proof. The first claim of the lemma is immediate because S is a tree and therefore any three vertices of S (i.e., any three labels) are consistent. For the second claim, define a directed graph G = (V, E) as follows. The vertex set V of G is the set of labels of the given NCA-labeling scheme. Assume that g is the decoder function of the labeling scheme. We add a directed edge from u ∈ V to v ∈ V if g(u, v) = v and there is no vertex w such that g(u, w) = w and g(w, v) = v (i.e., if v is the parent of u). We claim that G is a rooted tree. First observe that G is acyclic. Otherwise, by using Property (II) several times, we can find three vertices u, v, and w such that g(u, v) = v, g(v, w) = w, and g(w, u) = u. However from Property (II) we then also have g(u, w) = w, a contradiction. Second, we show that the out-degree of each vertex of G is at most 1. For contradiction, assume that there exists a vertex u that has out-going edges to v and w. We then have g(u, v) = v and g(u, w) = w and by Property (III) of consistent labeling schemes, we thus also have g(v, w) = w or g(w, v) = v. Thus, one of the two edges (u, v) and (u, w) cannot be in G. Finally, we show that there can be at most one vertex with out-degree 0. For the sake of contradiction assume that u and v both have out-degree 0 and let g(u, v) = w. By Property (I), we then also have g(u, w) = w and g(v, w) = w. If w 6= u, this implies that u has out-degree at least 1 and if w 6= v, it implies that v has out-degree at least 1. Hence, G is a rooted tree on the set of labels of the labeling scheme. Because the ancestry relationship of G is consistent with the labeling scheme, G is an NCA-universal tree for the family T. References [1] S. Abiteboul, H. Kaplan, and T. Milo. Compact labeling schemes for ancestor queries. In Proceedings of the Twelfth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’01, pages 547–556, Philadelphia, PA, USA, 2001. Society for Industrial and Applied Mathematics. [2] S. Alstrup, C. Gavoille, H. Kaplan, and T. Rauhe. Nearest common ancestors: A survey and a new distributed algorithm. In Proceedings of the Fourteenth Annual ACM Symposium on Parallel Algorithms and Architectures, SPAA ’02, pages 258–264, New York, NY, USA, 2002. ACM. [3] S. Alstrup, E. B. Halvorsen, and K. G. Larsen. Near-optimal labeling schemes for nearest common ancestors. In Proceedings of the Twenty-fifth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’14, pages 972–982, Philadelphia, PA, USA, 2014. Society for Industrial and Applied Mathematics. [4] M. A. Bender and M. Farach-Colton. The lca problem revisited. In LATIN, volume 1776, pages 88–94. Springer, 2000. [5] O. Berkman and U. Vishkin. Recursive star-tree parallel data structure. SIAM Journal on Computing, 22(2):221–242, 1993. [6] M. Breuer. Coding the vertexes of a graph. IEEE Transactions on Information Theory, 12(2):148–153, April 1966. 12 [7] M. A. Breuer and J. Folkman. An unexpected result in coding the vertices of a graph. Journal of Mathematical Analysis and Applications, 20(3):583 – 600, 1967. [8] P. Fraigniaud and A. Korman. Compact ancestry labeling schemes for xml trees. In Proceedings of the Twenty-first Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’10, pages 458–466, Philadelphia, PA, USA, 2010. Society for Industrial and Applied Mathematics. [9] H. N. Gabow, J. L. Bentley, and R. E. Tarjan. Scaling and related techniques for geometry problems. In Proceedings of the Sixteenth Annual ACM Symposium on Theory of Computing, STOC ’84, pages 135–143, New York, NY, USA, 1984. ACM. [10] C. Gavoille and D. Peleg. Compact and localized distributed data structures. Distributed Computing, 16(2):111–120, Sep 2003. [11] P. Gawrychowski and J. Łopuszański. Better labeling schemes for nearest common ancestors through minor-universal trees. arXiv:1707.06011, 2017. [12] D. Harel. A linear time algorithm for the lowest common ancestors problem. In Foundations of Computer Science, 1980., 21st Annual Symposium on, pages 308–319. IEEE, 1980. [13] D. Harel and R. E. Tarjan. Fast algorithms for finding nearest common ancestors. SIAM Journal on Computing, 13(2):338–355, 1984. [14] S. Kannan, M. Naor, and S. Rudich. Implicat representation of graphs. SIAM Journal on Discrete Mathematics, 5(4):596–603, 1992. [15] D. Peleg. Proximity-preserving labeling schemes and their applications. In Proc. 25th Int. Workshop on Graph-Theoretic Concepts in Comp. Sc., pages 30–41, 1999. [16] P. Powell. further improved lca algorithm. Plant Genome Data and Information Center collection on computational molecular biology and genetics, 1990. [17] R. Rado. Universal graphs and universal functions. Acta Arithmetica, 9(4):331–340, 0 1964. [18] B. Schieber and U. Vishkin. On finding lowest common ancestors: Simplification and parallelization. SIAM Journal on Computing, 17(6):1253–1262, 1988. 13 A A.1 Appendix Induction Basis For n equals one through four it is simple to find a universal tree of size at most n2 . For example: Figure 6: The root is always the top most vertex but the root of T ∈ Bn does not have to be embedded to the root of Sn Sn0 shall have size at most 2n2 . We want for any n a large tree Sn0 with a marked leaf such that we can embed any tree on n vertices with a marked leaf with the nca-function commuting. Again as example and induction basis: Figure 7: The larger red vertex is the marked leaf. It is easy to check that ∀T ∈ Bl and ∀T 0 ∈ Bl0 the NCA-query is equivalent to the NCA-query in Sl or Sl0 ∀l ∈ {1, .. , 4} resp. 14
8cs.DS
On computing tree and path decompositions with metric constraints on the bags Guillaume Ducoffe1,2 , Sylvain Legay3 , and Nicolas Nisse2,1 arXiv:1601.01958v1 [cs.CC] 8 Jan 2016 1 Univ. Nice Sophia Antipolis, CNRS, I3S, UMR 7271, 06900 Sophia Antipolis, France 2 INRIA, France 3 LRI, Univ Paris Sud, Université Paris-Saclay, 91405 Orsay, France Abstract We here investigate on the complexity of computing the tree-length and the tree-breadth of any graph G, that are respectively the best possible upper-bounds on the diameter and the radius of the bags in a tree decomposition of G. Path-length and path-breadth are similarly defined and studied for path decompositions. So far, it was already known that tree-length is NP-hard to compute. We here prove it is also the case for tree-breadth, path-length and path-breadth. Furthermore, we provide a more detailed analysis on the complexity of computing the tree-breadth. In particular, we show that graphs with tree-breadth one are in some sense the hardest instances for the problem of computing the tree-breadth. We give new properties of graphs with tree-breadth one. Then we use these properties in order to recognize in polynomial-time all graphs with tree-breadth one that are planar or bipartite graphs. On the way, we relate tree-breadth with the notion of k-good tree decompositions (for k = 1), that have been introduced in former work for routing. As a byproduct of the above relation, we prove that deciding on the existence of a k-good tree decomposition is NP-complete (even if k = 1). All this answers open questions from the literature. Keywords: tree-length; tree-breadth; path-length; path-breadth; k-good tree decompositions. 1 Introduction Context. It is a fundamental problem in metric graph theory [7] to embed a graph into a simpler metric space while minimizing the (multiplicative) distortion of the distances in the graph. In particular, minimum distortion embeddings of a graph into a tree or a path have practical applications in computer vision [48], computational chemistry and biology [38] as well as in network design and distributed protocols [36]. The two problems to embed a graph into a tree or a path with minimum distortion are NP-hard [2, 8, 43]. However, there exists a nice setting in order to approximate these two problems. More precisely, a series of graph parameters has been introduced in recent work in order to measure how much the distance distribution of a graph is close to a tree metric or a path metric [26, 29, 30]. We refer to [28, 30] for details about the relationships between these parameters and the two above-mentioned embedding problems. Here we study the complexity of computing these parameters, thereby solving open problems in the literature. The parameters that are considered in this note can be defined using the terminology of Robertson and Seymour tree decompositions [46]. Informally, a tree decomposition is a dividing of a graph G into “bags”: that are overlapping subgraphs that can be pieced together in a tree-like manner (formal definitions will be given in the technical sections of the paper). The shortestpath metric of G is “tree-like” when each bag of the tree decomposition has bounded diameter and bounded radius, where the distance taken between two vertices in a same bag is their distance in G. The tree-length [26] and the tree-breadth [29] of G are respectively the best possible upper-bounds on the diameter and the radius of the bags in a tree decomposition of G. Path-length [49] and path-breadth [30] are defined in the same fashion as tree-length and tree-breadth for path decompositions. In this paper, we focus on the complexity of computing the four parameters tree-length, tree-breadth, path-length and path-breadth. 1 Recent studies suggest that some classes of real-life networks – including biological networks and social networks – have bounded tree-length and tree-breadth [1]. This metric tree-likeness can be exploited in algorithms. For instance, bounded tree-length graphs admit compact distance labeling scheme [25] as well as a PTAS for the well-known Traveling Salesman problem [41]. Furthermore, the diameter and the radius of bounded tree-length graphs can be approximated up to an additive constant in linear-time [19]. In contrast to the above result, we emphasize that under classical complexity assumptions the diameter of general graphs cannot be approximated up to an additive constant in subquadratic-time, that is prohibitive for very large graphs [17]. Note that a large amount of the literature about tree decompositions rather seeks to minimize the size of the bags than their diameter. The tree-width [46] of a graph G is the best possible upperbound on the size of the bags in a tree decomposition of G. However, tree-length and the other parameters that are considered in this paper can differ arbitrarily from tree-width; we refer to [22] for a global picture on the relations between tree-length and tree-width. Furthermore, one aim of this paper is to complete the comparison between tree-width and path-width on one side, and treelength, tree-breadth, path-length and path-breadth on the other side, from the complexity point of view. Let us remind that computing the tree-width (resp. the path-width) is NP-hard [4, 39], however for every fixed k ≥ 1 there is a linear-time algorithm to decide whether a graph has tree-width at most k (resp., path-width at most k) [10, 14]. Related work. The complexity of computing tree-length, tree-breadth, path-length and pathbreadth has been left open in several works [26, 29, 30]. So far, it has been solved only for tree-length, that is NP-hard to compute. Tree-length and tree-breadth. It is NP-complete to decide whether a graph has tree-length at most k for every constant k ≥ 2 [42]. However, the reduction used for tree-length goes through weighted graphs and then goes back to unweighted graphs using rather elegant gadgets. It does not seem to us these gadgets can be easily generalized in order to apply to the other parameters that are considered in this note. On a more positive side, there exists a 3-approximation algorithm for tree-length [26]. In this aspect, it looks natural to investigate on the complexity of computing the tree-breadth, since any polynomial-time algorithm would imply an improved 2-approximation algorithm for tree-length. Path-length and path-breadth. There exist constant-factor approximation algorithms for pathlength and path-breadth [30]. Recently, the minimum eccentricity shortest-path problem – that is close to the computation of path-length and path-breadth – has been proved NP-hard [31]. Let us point out that for every fixed k, it can be decided in polynomial-time whether a graph admits a shortest-path with eccentricity at most k [31]. Our results will show the situation is different for path-length and path-breadth than for the minimum eccentricity shortest-path. Our contributions. On the negative side, we prove that tree-breadth, path-length and pathbreadth are NP-hard to compute. More precisely: • recognizing graphs with tree-breadth one is NP-complete; • recognizing graphs with path-length two is NP-complete; • recognizing graphs with path-breadth one is NP-complete. It is remarkable the last two results (for path-length and path-breadth) are obtained using the same reduction. Our reductions have distant similarities with the reduction that was used for tree-length. However, they do not need any detour through weighted graphs. We next focus our work on tree-breadth (although part of the results may extend to the three other parameters that are considered in this note). We give a more in-depth analysis on the complexity of computing this parameter. In particular, we prove it is equally hard to compute tree-breadth as to recognize graphs with tree-breadth one. Therefore, graphs with tree-breadth one are in some sense the hardest instances for the problem of computing the tree-breadth. The latter partly answers an open question from [29], where it was asked for a characterization of graphs with tree-breadth one. We also prove a few properties of graphs with tree-breadth one. In particular, graphs with tree-breadth one are exactly those graphs admitting a 1-good tree decomposition, that is a tree decomposition whose each bag has a spanning star. The more general notion of k-good tree decompositions was introduced in [40] to obtain new compact routing schemes. Note that as a byproduct of the above relation between 1-good tree decompositions and graphs with tree-breadth 2 one, we obtain that deciding on the existence of a k-good tree decomposition is an NP-complete problem even when k = 1. Finally, on the algorithmic side, we show how to recognize in polynomial time all graphs of tree-breadth one that are planar or bipartite. In particular, our recognition algorithm for planar graphs of tree-breadth one relies upon deep structural properties of these graphs. Definitions and useful notations are given in Section 2. All our results of NP-completeness are listed and proved in Section 3. Sections 4 and 5 are devoted to the computation of tree-breadth. In particular, in Section 5 we present and we prove correctness of an algorithm to recognize planar graphs with tree-breadth one. Finally, we discuss about some open questions in the conclusion (Section 6). 2 Definitions and preliminary results We refer to [24] for a survey on graph theory. Graphs in this study are finite, simple, connected and unweighted. Let G = (V, E) be a graph. For any X ⊆ V , let G[X] denote the subgraph of G induced by X. For any subgraph H of G, let NH (v) denote the set of neighbors of v ∈ V in H, and let NH [v] = NH (v) ∪ {v}. The distance distH (u, v) between two vertices u, v ∈ V in H is the minimum length (number of edges) of a path between u and v in a subgraph H of G. In what follows, we will omit the subscript when no ambiguity occurs. A set S ⊆ V is a dominating set of G if any vertex of V \ S has a neighbor in S. The dominating number γ(G) of a graph G is the minimum size of a dominating set of G. Tree decompositions and path decompositions of a graph. A tree decomposition (T, X ) of G is a pair consisting of a tree T and of a family X = (Xt )t∈V (T ) of subsets of V indexed by the nodes of T and satisfying: S • t∈V (T ) Xt = V ; • for any edge e = {u, v} ∈ E, there exists t ∈ V (T ) such that u, v ∈ Xt ; • for any v ∈ V , the set of nodes t ∈ V (T ) such that v ∈ Xt induces a subtree, denoted by Tv , of T . The sets Xt are called the bags of the decomposition. If no bag is contained into another one, then the tree decomposition is called reduced. Starting from any tree decomposition, a reduced tree decomposition can be obtained in polynomial-time by contracting any two adjacent bags with one contained into the other until it is no more possible to do that. In the following we will make use of the Helly property in our proofs: Lemma 1. [6, Helly property] T Let T be a tree and let T1 , T2 , . . . , Tk be a finite family of pairwise intersecting subtrees. Then, ki=1 Ti 6= ∅, or equivalently there is a node contained in all the k subtrees. Finally, let (T, X ) be a tree decomposition, it is called a path decomposition if T induces a path. Metric tree-likeness and path-likeness. All graph invariants that we consider in the paper can be defined in terms of tree decompositions and path decompositions. Let (T, X ) be any tree decomposition of a graph G. For any t ∈ V (T ), • the diameter of bag Xt equals maxv,w∈Xt distG (v, w); • the radius ρ(t) of a bag Xt equals minv∈V maxw∈Xt distG (v, w). The length of (T, X ) is the maximum diameter of its bags, while the breadth of (T, X ) is the maximum radius of its bags. The tree-length and the tree-breadth of G, respectively denoted by tl(G) and tb(G), are the minimum length and breadth of its tree decomposition, respectively. Let k be a positive integer, the tree decomposition (T, X ) is called k-good when each bag contains a dominating induced path of length at most k − 1. It is proved in [40] every graph G has a k-good tree decomposition for k = ch(G) − 1, with ch(G) denoting the size of a longest induced cycle of G. Finally, path-length, path-breadth and k-good path decompositions are similarly defined and 3 studied for the path decompositions as tree-length, tree-breadth and k-good tree decompositions are defined and studied for the tree decompositions. The path-length and path-breadth of G are respectively denoted by pl(G) and pb(G). It has been observed in [29, 30] that the four parameters tree-length, tree-breadth and pathlength, path-breadth are contraction-closed invariants. We will use the latter property in our proofs. Lemma 2 ( [29, 30]). For every G = (V, E) and for any edge e ∈ E: tl(G/e) ≤ tl(G), tb(G/e) ≤ tb(G) and pl(G/e) ≤ pl(G), pb(G/e) ≤ pb(G). Furthermore, it can be observed that for any graph G, tb(G) ≤ tl(G) ≤ 2 · tb(G) and similarly pb(G) ≤ pl(G) ≤ 2 · pb(G). Moreover, if a graph G admits a k-good tree decomposition, then tb(G) ≤ bk/2c + 1 and tl(G) ≤ k + 1. Before we end this section, let us prove the stronger equivalence, tb(G) = 1 if and only if G admits a 1-good tree decomposition. This result will be of importance in the following. Since a tree decomposition is 1-good if and only if each bag contains a spanning star, we will name the 1-good tree decompositions star-decompositions in the following. Definition 1. Let G = (V, E) be a connected graph, a star-decomposition is a tree decomposition (T, X ) of G whose each bag induces a subgraph of dominating number one, i.e., for any t ∈ V (T ), γ(G[Xt ]) = 1. Clearly, if a graph admits a star-decomposition, then it has tree-breadth at most one. Let us prove that the converse also holds. Lemma 3. For any graph G with tb(G) ≤ 1, every reduced tree decomposition of G of breadth one is a star-decomposition. In particular: • any tree decomposition of G of breadth one can be transformed into a star-decomposition in polynomial-time; • similarly, any path decomposition of G of breadth one can be transformed into a 1-good path decomposition in polynomial-time. Proof. Let (T, X ) be any reduced tree decomposition of G of breadth one. We will prove it is a star-decomposition. To prove it, let Xt ∈ X be arbitrary and let v ∈ V be such that maxw∈Xt distG (v, w) = 1, which exists because Xt has radius one. We now show that v ∈ Xt . Indeed, since the subtree Tv and the subtrees Tw , w ∈ Xt , pairwise intersect, then it comes by the T Helly Property (Lemma 1) that Tv ∩ w∈Xt Tw 6= ∅ i.e., there is some bag containing {v}∪Xt . As a result, we have that v ∈ Xt because (T, X ) is a reduced tree decomposition, hence γ(G[Xt ]) = 1. The latter implies that (T, X ) is a star-decomposition because Xt is arbitrary. Now let (T, X ) be any tree decomposition of G of breadth one. It can be transformed in polynomial-time into a reduced tree decomposition (T 0 , X 0 ) so that X 0 ⊆ X . Furthermore, (T 0 , X 0 ) has breadth one because it is the case for (T, X ), therefore (T 0 , X 0 ) is a star-decomposition. In particular, if (T, X ) is a path decomposition then so is (T 0 , X 0 ). Corollary 1. For any graph G, tb(G) ≤ 1 if and only if G admits a star-decomposition. 3 3.1 Intractability results Path-length and path-breadth This section is devoted to the complexity of all path-like invariants that we consider in this paper. Theorem 1. Deciding whether a graph has path-length at most k is NP-complete even if k = 2. In contrast to Theorem 1, graphs with path-length one are exactly the interval graphs [30], i.e., they can be recognized in linear-time. Theorem 2. Deciding whether a graph has path-breadth at most k is NP-complete even if k = 1. From the complexity result of Theorem 2, we will also prove the hardness of deciding on the existence of k-good path decompositions. 4 Theorem 3. Deciding whether a graph admits a k-good path decomposition is NP-complete even if k = 1. Proof. The problem is in NP. By Lemma 3, a graph G admits a 1-good path decomposition if and only if pb(G) ≤ 1, therefore it is NP-hard to decide whether a graph admits a 1-good path decomposition by Theorem 2. All of the NP-hardness proofs in this section will rely upon the same reduction from the Betweenness problem, defined below. The Betweenness problem, sometimes called the Total Ordering problem, is NP-complete [44]. In [34], it was used to show that the Interval Sandwich problem is NP-complete. What we here prove is that the Interval Sandwich problem remains NP-complete even if the second graph is a power of the first one, where the kth power Gk of any graph G is obtained from G by adding an edge between every two distinct vertices that are at distance at most k in G for every integer k ≥ 1. Indeed, a graph G has path-length at most k if and only if there is an Interval Sandwich between G and Gk (we refer to [42] for the proof of a similar equivalence between tree-length and the Chordal Sandwich problem). Problem 1 (Betweenness). Input: a set S of n elements, a set T of m ordered triples of elements in S. Question: Is there a total ordering of S such that for every triple t = (si , sj , sk ) ∈ T , either si < sj < sk or sk < sj < si ? Now, given an instance (S, T ) of the Betweenness problem, we will construct from S and T a graph GS,T as defined below. We will then prove that pl(GS,T ) ≤ 2 (resp. pb(GS,T ) ≤ 1) if and only if (S, T ) is a yes-instance of the Betweenness problem. Definition 2. Let S be a set of n elements, let T be a set of m ordered triples of elements in S. The graph GS,T is constructed as follows: • For every element si ∈ S, 1 ≤ i ≤ n, there are two adjacent vertices ui , vi in GS,T . The vertices ui are pairwise adjacent i.e., the set U = {ui | 1 ≤ i ≤ n} is a clique. • For every triple t = (si , sj , sk ) ∈ T , let us add in GS,T the vi vj -path (vi , at , bt , vj ) of length 3, and the vj vk -path (vj , ct , dt , vk ) of length 3. • Finally, for every triple t = (si , sj , sk ) ∈ T let us make adjacent at , bt with every ul such that l 6= k, similarly let us make adjacent ct , bt with every ul such that l 6= i. It can be noticed from Definition 2 that for any 1 ≤ i ≤ n, the vertex ui is adjacent to any vertex but those vj such that j 6= i, those at , bt such that si is the last element of triple t and those ct , bt such that si is the first element of triple t. We refer to Figure 1 for an illustration (see also Figure 2). Observe that GS,T has diameter 3 because the clique U dominates GS,T , therefore pl(GS,T ) ≤ 3 and we will show that it is hard to distinguish graphs with path-length two from graphs with path-length three. Similarly, the clique U dominates GS,T hence pb(GS,T ) ≤ 2, thus we will show that it is hard to distinguish graphs with path-breadth one from graphs with path-breadth two. Lemma 4. Let S be a set of n elements, let T be a set of m ordered triples of elements in S. If (S, T ) is a yes-instance of the Betweenness problem then pb(GS,T ) ≤ 1 and pl(GS,T ) ≤ 2, where GS,T is the graph that is defined in Definition 2. Proof. Since pl(GS,T ) ≤ 2 · pb(GS,T ) then we only need to prove that pb(GS,T ) ≤ 1. For convenience, let us reorder the elements of S so that for every triple (si , sj , sk ) ∈ T either i < j < k or k < j < i. It is possible to do that because by the hypothesis (S, T ) is a yes-instance of the Betweenness problem. If furthermore k < j < i, let us also replace (si , sj , sk ) with the inverse triple (sk , sj , si ). This way, we have a total ordering of S such that si < sj < sk for every triple (si , sj , sk ) ∈ T . Then, let us construct a path decomposition (P, X ) with n bags, denoted X1 , X2 , . . . , Xn , as follows. For every 1 ≤ i ≤ n, U ⊆ Xi and vi ∈ Xi . For every t = (si , sj , sk ) ∈ T , we add both at , bt into the bags Xl with i ≤ l ≤ j, similarly we add both ct , dt into the bags Xl with j ≤ l ≤ k. By construction, the clique U is contained in any bag of P and for every triple t = (si , sj , sk ) ∈ T we have at , bt , vi ∈ Xi and at , bt , ct , dt , vj ∈ Xj and ct , dt , vk ∈ Xk , therefore (P, X ) is indeed a path decomposition of GS,T . We claim that for every i, Xi ⊆ N [ui ], that will prove the lemma. Indeed if it were not the case for some i then by Definition 2 there should exist t ∈ T , j, k such that: either t = (si , sj , sk ) ∈ T 5 u1 t = (s i , s j , s k ), ... u3 u4 u i-1 c3 v1 c1 a1 d1 v2 b2 v3 d3 ui v5 at b4 bt a4 vj b3 uj ct dt d4 u i+1 u i+2 vk c4 U ... a3 v4 vi ... a2 d2 c2 b1 i<j<k u1 u2 u3 u5 u2 u k-1 uk uk+1 uk+2 ... Figure 1: The graph GS,T for S = [|1, 5|] and T = {(i, i + 1, i + 2) | 1 ≤ i ≤ 4}. Each colour corresponds to a given triple of T . For ease of reading, the adjacency relations between the vertices ui and the colored vertices at , bt , ct , dt are not drawn. un Figure 2: Adjacency relations in GS,T for one given triple t = (si , sj , sk ). and ct , dt ∈ Xi ; or t = (sk , sj , si ) ∈ T and at , bt ∈ Xi . But then by construction either at , bt are only contained in the bags Xl for k ≤ l ≤ j, or ct , dt are only contained in the bags Xl for j ≤ l ≤ k, thus contradicting the fact that either at , bt ∈ Xi or ct , dt ∈ Xi . Lemma 5. Let S be a set of n elements, let T be a set of m ordered triples of elements in S. If pb(GS,T ) ≤ 1 or pl(GS,T ) ≤ 2 then (S, T ) is a yes-instance of the Betweenness problem, where GS,T is the graph that is defined in Definition 2. Proof. Since pl(GS,T ) ≤ 2 · pb(GS,T ) then we only need to consider the case when pl(GS,T ) ≤ 2. Let (P, X ) be a path decomposition of length two, that exists by the hypothesis. Since the vertices vi are pairwise at distance 3 then the subpaths Pvi that are induced by the bags containing vertex vi are pairwise disjoint. Therefore, starting from an arbitrary endpoint of P and considering each vertex vi in the order that it appears in the path decomposition, this defines a total ordering over S. Let us reorder the set S so that vertex vi is the ith vertex to appear in the path-decomposition. We claim that for every triple t = (si , sj , sk ) ∈ T , either i < j < k or k < j < i, that will prove the lemma. By way of contradiction, let t = (si , sj , sk ) ∈ T such that either j < min{i, k} or j > max{i, k}. By symmetry, we only need to consider the case when j < i < k. In such case by construction the path between Pvj and Pvk in P contains Pvi . Let B ∈ Pvi , by the properties of a tree decomposition it is a vj vk -separator, so it must contain one of ct , dt . However, vertex vi ∈ B is at distance 3 from both vertices ct , dt , thus contradicting the fact that (P, X ) has length 2. We are now able to prove Theorems 1 and 2. Proof of Theorem 1. To prove that a graph G satisfies pl(G) ≤ k, it suffices to give as a certificate a tree decomposition of G with length at most k. Indeed, the all-pairs-shortest-paths in G can be computed in polynomial-time. Therefore, the problem of deciding whether a graph has path-length at most k is in NP. Given an instance (S, T ) of the Betweenness problem, let GS,T be as defined in Definition 2. We claim that pl(GS,T ) ≤ 2 if and only if the pair (S, T ) is a yes-instance of the Betweenness problem. This will prove the NP-hardness because our reduction is polynomial and the Betweenness problem is NP-complete. To prove the claim in one direction, if (S, T ) is a yesinstance then by Lemma 4 pl(GS,T ) ≤ 2. Conversely, if pl(GS,T ) ≤ 2 then (S, T ) is a yes-instance by Lemma 5, that proves the claim in the other direction. Proof of Theorem 2. To prove that a graph G satisfies pb(G) ≤ k, it suffices to give as a certificate a tree decomposition of G with breadth at most k. Indeed, the all-pairs-shortest-paths in G can be computed in polynomial-time. Therefore, the problem of deciding whether a graph has pathbreadth at most k is in NP. Given an instance (S, T ) of the Betweenness problem, let GS,T be as defined in Definition 2. We claim that pb(GS,T ) ≤ 1 if and only if the pair (S, T ) is a yes-instance 6 of the Betweenness problem. This will prove the NP-hardness because our reduction is polynomial and the Betweenness problem is NP-complete. To prove the claim in one direction, if (S, T ) is a yes-instance then by Lemma 4 pb(GS,T ) ≤ 1. Conversely, if pb(GS,T ) ≤ 1 then (S, T ) is a yes-instance by Lemma 5, that proves the claim in the other direction. To conclude this section, we strenghten the above hardness results with two inapproximability results. Indeed, it has to be noticed that for any graph parameter param, an α-approximation algorithm for param with α < 1 + k1 is enough to separate the graphs G such that param(G) ≤ k from those such that param(G) ≥ k + 1. Therefore, the two following corollaries follow from our polynomial-time reduction. Corollary 2. For every ε > 0, the path-length of a graph cannot be approximated within a factor 3 − ε unless P=NP. 2 Proof. Let GS,T be the graph of the reduction in Theorem 1. By Definition 2, it has diameter at most 3 and so pl(G  S,T ) ≤ 3. Since it is NP-hard to decide whether pl(GS,T ) ≤ 2, therefore it does not exist a 32 − ε -approximation algorithm for path-length unless P=NP. Corollary 3. For every ε > 0, the path-breadth of a graph cannot be approximated within a factor 2 − ε unless P=NP. Proof. Let GS,T be the graph of the reduction in Theorem 2. By Definition 2, the set U is a dominating clique and so pb(GS,T ) ≤ 2. Since it is NP-hard to decide whether pb(GS,T ) ≤ 1, therefore it does not exist a (2 − ε)-approximation algorithm for path-breadth unless P=NP. So far, there exists a 2-approximation algorithm for path-length and a 3-approximation algorithm for path-breadth [30]. Therefore, we let open whether there exist 23 -approximation algorithms for path-length and 2-approximation algorithms for path-breadth. 3.2 Tree-breadth We prove next that computing the tree-breadth is NP-hard. Theorem 4. Deciding whether a graph has tree-breadth at most k is NP-complete even if k = 1. Theorem 5. Deciding whether a graph admits a k-good tree decomposition is NP-complete even if k = 1. Proof. The problem is in NP. By Corollary 1, a graph G admits a star-decomposition if and only if tb(G) ≤ 1, therefore it is NP-hard to decide whether a graph admits a 1-good path decomposition by Theorem 4. In order to prove Theorem 4, we will reduce from the Chordal Sandwich problem (defined below). In [42], the author also proposed a reduction from the Chordal Sandwich problem in order to prove that computing tree-length is NP-hard. However, we will need different gadgets than in [42], and we will need different arguments to prove correctness of the reduction. Problem 2 (Chordal Sandwich). Input: graphs G1 = (V, E1 ) and G2 = (V, E2 ) such that E1 ⊆ E2 . Question: Is there a chordal graph H = (V, E) such that E1 ⊆ E ⊆ E2 ? The Chordal Sandwich problem is NP-complete even when the 2n = |V | vertices induce a perfect matching in Ḡ2 (the complementary of G2 ) [13, 35]. Perhaps surprisingly, the later constriction on the structure of Ḡ2 is a key element in our reduction. Indeed, we will need the following technical lemma. Lemma 6. Let G1 = (V, E1 ), G2 = (V, E2 ) such that E1 ⊆ E2 and Ḡ2 (the complementary of G2 ) is a perfect matching. Suppose that hG1 , G2 i is a yes-instance of the Chordal Sandwich problem. Then, there exists a reduced tree decomposition (T, X ) of G1 such that for every forbidden edge {u, v} ∈ / E2 : Tu ∩ Tv = ∅, Tu ∪ Tv = T , furthermore there are two adjacent bags Bu ∈ Tu and Bv ∈ Tv such that Bu \ u = Bv \ v. 7 Proof. Let H = (V, E) be any chordal graph such that E1 ⊆ E ⊆ E2 (that exists because hG1 , G2 i is a yes-instance of the Chordal Sandwich problem by the hypothesis) and the number |E| of edges is maximized. We will prove that any clique-tree (T, X ) of H satisfies the above properties (given in the statement of the lemma). To prove it, let {u, v} ∈ / E2 be arbitrary. Observe that Tu ∩ Tv = ∅ (else, {u, v} ∈ E, that would contradict that E ⊆ E2 ). Furthermore, let Bu ∈ Tu minimize the distance in T to the subtree Tv , let B be the unique bag that is adjacent to Bu on a shortest-path between Bu and Tv in T . Note that B ∈ / Tu by the minimality of distT (Bu , Tv ), however B may belong to Tv . Removing the edge {Bu , B} in T yields two subtrees T1 , T2 with Tu ⊆ T1 and Tv ⊆ T2 . In addition, we have that for every x ∈ V \ u such that Tx ∩ T1 6= ∅, {u, x} ∈ E2 since x 6= v and Ḡ2 is a perfect matching by the hypothesis. Similarly, we have that for every y ∈ V \ v such that Ty ∩ T2 6= ∅, {v, y} ∈ E2 . Therefore, by maximality of the number |E| of edges, it follows that T1 = Tu and T2 = Tv , and so, Tu ∪ Tv = T . In particular, B = Bv ∈ Tv . Finally, let us prove that Bu \ u = Bv \ v. Indeed, assume for the sake of contradiction that Bu \ u 6= Bv \ v. In particular, (Bu \ Bv ) \ u 6= ∅ or (Bv \ Bu ) \ v 6= ∅. Suppose w.l.o.g. that (Bu \ Bv ) \ u 6= ∅. Let H 0 = (V, E 0 ) be obtained from H by adding an edge between vertex v and every vertex of (Bu \ Bv ) \ u. By construction |E 0 | > |E|. Furthermore, H 0 is chordal since a clique-tree of H 0 can be obtained from (T, X ) by adding a new bag (Bu \ u) ∪ {v} in-between Bu and Bv . However, for every x ∈ (Bu \ Bv ) \ u we have that {x, v} ∈ E2 since x 6= u and Ḡ2 is a perfect matching by the hypothesis. As a result, E 0 ⊆ E2 , thus contradicting the maximality of the number |E| of edges in H. Proof of Theorem 4. The problem is in NP. To prove the NP-hardness,let hG1 , G2 i be any input of the Chordal Sandwich problem such that Ḡ2 is a perfect matching. The graph G0 is constructed from G1 as follows. First we add a clique V 0 of 2n = |V | vertices in G0 . Vertices v ∈ V are in one-to-one correspondance with vertices v 0 ∈ V 0 . Then, for every forbidden edge {u, v} ∈ / E2 , vertices u, v are respectively made adjacent to all vertices in V 0 \ v 0 and V 0 \ u0 . Finally, we add a distinct copy of the gadget Fuv in Figure 3, and we make adjacent suv and tuv to the two vertices u0 , v 0 (see also Figure 4 for an illustration). We will prove tb(G0 ) = 1 if and only if hG1 , G2 i is a yesinstance of the Chordal Sandwich problem. This will prove the NP-hardness because our reduction is polynomial and the Chordal Sandwich problem is NP-complete even when the 2n = |V | vertices induce a perfect matching in Ḡ2 (the complementary of G2 ) [13, 35]. u xuv suv tuv cuv yuv wuv zuv v Figure 3: The gadget Fuv . In one direction, assume tb(G0 ) = 1, let (T, X ) be a star-decomposition of G0 . Let H = (V, {{u, v} | Tu ∩ Tv 6= ∅}), that is a chordal graph such that E1 ⊆ E(H). To prove that hG1 , G2 i is a yes-instance of the Chordal Sandwich problem, it suffices to prove that Tu ∩ Tv = ∅ for every forbidden edge {u, v} ∈ / E2 . More precisely, we will prove that Tsuv ∩ Ttuv 6= ∅, for we claim that the latter implies Tu ∩ Tv = ∅. Indeed, assume Tsuv ∩ Ttuv 6= ∅ and Tu ∩ Tv 6= ∅. Since suv and tuv are both adjacent to u and v, therefore the four subtrees Tu , Tv , Tsuv , Ttuv pairwise intersect. By the Helly property (Lemma 1) Tu ∩ Tv ∩ Tsuv ∩ Ttuv 6= ∅, hence there is a bag containing u, v, suv , tuv but then it contradicts the fact that (T, X ) is a star-decomposition because no vertex dominates the four vertices. Therefore, Tsuv ∩ Ttuv 6= ∅ implies Tu ∩ Tv = ∅. Let us prove that Tsuv ∩ Ttuv 6= ∅. By contradiction, assume Tsuv ∩ Ttuv = ∅. Every bag B onto the path between Tsuv and Ttuv must contain cuv , xuv , furthermore N [cuv ] ∩ N [xuv ] = {suv , tuv }. Since, (T, X ) is a star-decomposition, the latter implies either suv ∈ B and B ⊆ N [suv ] or tuv ∈ B and B ⊆ N [tuv ]. 8 Consequently, there exist two adjacent bags Bs ∈ Tsuv , Bt ∈ Ttuv such that Bs ⊆ N [suv ] and Bt ⊆ N [tuv ]. Furthermore, Bs ∩ Bt is an suv tuv -separator by the properties of a tree decomposition. In particular, Bs ∩ Bt must intersect the path (yuv , wuv , zuv ) because yuv ∈ N (suv ) and zuv ∈ N (tuv ). However, Bs ⊆ N [suv ], Bt ⊆ N [tuv ] but N [suv ] ∩ N [tuv ] ∩ {yuv , wuv , zuv } = ∅, hence Bs ∩ Bt ∩ {yuv , wuv , zuv } = ∅, that is a contradiction. As a result, Tsuv ∩ Ttuv 6= ∅ and so, Tu ∩ Tv = ∅. x' yuv V' u' v' u v suv xuv wuv cuv tuv zuv G 1 x Figure 4: The graph G0 (simplified view). Conversely, assume that hG1 , G2 i is a yes-instance of the Chordal Sandwich problem. Since Ḡ2 is a perfect matching by the hypothesis, by Lemma 6 there exists a reduced tree decomposition (T, X ) of G1 such that for every forbidden edge {u, v} ∈ / E2 : Tu ∩ Tv = ∅, Tu ∪ Tv = T and there are two adjacent bags Bu ∈ Tu , Bv ∈ Tv so that Bu \ u = Bv \ v. Let us modify (T, X ) in order to obtain a star-decomposition of G0 . In order to achieve the result, we first claim that for every edge {t, t0 } ∈ E(T ), the bags Xt , Xt0 differ in exactly one vertex, that is, |Xt \ Xt0 | = 1 and similarly |Xt0 \ Xt | = 1. Indeed, Xt \ Xt0 6= ∅ because (T, X ) is reduced, so, let utt0 ∈ Xt \ Xt0 . Let vtt0 ∈ V be the unique vertex satisfying {utt0 , vtt0 } ∈ / E2 , that is well-defined because Ḡ2 is a perfect matching by the hypothesis. Note that vtt0 ∈ Xt0 because utt0 ∈ / Xt0 and Tutt0 ∪ Tvtt0 = T . Furthermore, vtt0 ∈ / Xt because utt0 ∈ Xt and Tutt0 ∩ Tvtt0 = ∅. By construction of (T, X ), there are two adjacent bags Butt0 ∈ Tutt0 , Bvtt0 ∈ Tvtt0 such that Butt0 \ utt0 = Bvtt0 \ vtt0 . Since utt0 ∈ Xt \ Xt0 and vtt0 ∈ Xt0 \ Xt , therefore, Xt = Butt0 and Xt0 = Bvtt0 , and so, Xt \ Xt0 = {utt0 } and Xt0 \ Xt = {vtt0 }. In the following, we will keep the above notations utt0 , vtt0 for every edge {t, t0 } ∈ E(T ) (in particular, utt0 = vt0 t and vtt0 = ut0 t ). Let us construct the star-decomposition (T 0 , X 0 ) of G0 as follows. S • For every node t ∈ V (T ), let St = Xt ∪ V 0 ∪ ( t0 ∈NT (t) {sutt0 vtt0 , tutt0 vtt0 }) (in particular, |St | = 2n + |Xt | + 2 · degT (t)). We will first construct a path decomposition of G0 [St ] whose bags are the sets Ytt0 = Xt ∪ V 0 ∪ {sutt0 vtt0 , tutt0 vtt0 } for every edge {t, t0 } ∈ E(T ) (note that the bags can be linearly ordered in an arbitrary way in the path decomposition). Furthermore, for every edge {t, t0 } ∈ E(T ), Ytt0 ⊆ N [u0tt0 ], where u0tt0 ∈ V 0 is the corresponding vertex to utt0 ∈ V in the clique V 0 (see Figure 5 for an illustration). Therefore, the above constructed path decomposition is a 1-good path decomposition. C v1 u2 C u3 u4 u1 v 2 u3 u4 u1 u 2 u3 s u1v1 u4 u1v1 u1 u 2 C u1 u 2 s u2v2 C t V' V' V' C t u2v2 u3 u4 s u3v3 C u1 u 2 t u3v3 u3 u4 V' s u4v4 C u1 u 2 t u4v4 u3 u4 C u1 u 2 u3 u4 C u3 v4 u1 u 2 v3 u4 Figure 5: The 1-good path decomposition (right) obtained from the central bag with degree four (left). • Then, we will connect the 1-good path decompositions together. More precisely, let us add an edge between the two bags Ytt0 and Yt0 t for every edge {t, t0 } ∈ E(T ) (see Figure 6 for an illustration). S In so doing, we claim that one obtains a star-decomposition of G0 [ t∈V (T ) St ]. Indeed, it is a tree decomposition since: – the clique V 0 is contained in all bags; 9 – for every {t, t0 } ∈ E(T ) the two vertices sutt0 vtt0 , tutt0 vtt0 are only contained in the two 0 0 0 adjacent bags Ytt0 and Yt0 t , furthermore utt0 , u0tt0 , vtt 0 ∈ Ytt0 and vtt0 , utt0 , vtt0 ∈ Yt0 t ; 0 – last, each vertex v ∈ V is contained in {Ytt0 | v ∈ Xt and t ∈ NT (t)} which induces a subtree since (T, X ) is a tree decomposition of G1 . Since in addition every bag Ytt0 , with {t, t0 } ∈ E(T ), is dominated by u0tt0 ∈ V 0 , this proves the claim that one obtains a star-decomposition. V' V' s ... u1v1 u1v1 u4 u1 u 2 s u2v2 C u1v1 u1 u 2 u2v2 u3 s u3v3 C t u4 u1 u 2 u4 v3 V' V' V' s ... C u3v3 u3 u1v1 u3v3 t v1 u 2 t s ... C t u3v3 u3 V' s u4v4 C t u4 u1 u 2 C t u4v4 u4 u3 u1 u 2 V' V' s ... u2v2 u2v2 u1 v 2 s ... C t u4 u3 u4v4 u4v4 u3 ... C t u1 u 2 u4 u3 v4 Figure 6: Connection of the 1-good path decomposition in Figure 5 to the neighbouring 1-good path decompositions. • In order to complete the construction, let us observe that for every forbidden edge {u, v} ∈ / E2 , there is a star-decomposition of Fuv \{u, v} with three leaf-bags {xuv , suv , tuv }, {yuv , suv , wuv } and {zuv , suv , wuv } and one internal bag of degree three Buv = {cuv , suv , tuv , wuv }. For every {t, t0 } ∈ E(T ), we simply connect the above star-decomposition of Futt0 vtt0 \ {utt0 , vtt0 } by making the internal bag Butt0 vtt0 adjacent to one of Ytt0 or Yt0 t (see Figure 7 for an illustration). s t u1v1 s s t 1 1 wu v s zu v t u1v1 3 3 zu v 3 3 3 3 1 1 s u1v1 2 2 s ... C v1 u 2 s V' V' s u1v1 3 3 u3v3 wu v 1 1 u1v1 wu v u3v3 c u1v1 wu v t t u3v3 c ... u3v3 yu v wu v wu v 1 1 t 3 3 1 1 1 1 u3v3 xu v u3v3 u1v1 yu v s t u3v3 u1v1 xu v u1v1 u3v3 t u3v3 u3 u4 u4v4 ... C u1 u 2 yu v 4 4 wu v 4 4 v3 s u4 u4v4 t xu v s u4v4 4 4 u4v4 V' V' s u1v1 t u1v1 s u2v2 C u1 u 2 V' u2v2 u3 s u3v3 C t u4 t u3v3 u1 u 2 u3 u4 u4v4 u4v4 u4 u2v2 u2v2 u1 v 2 t u2v2 s u4v4 t u4v4 u3 s u4v4 t u4v4 4 4 u3 wu v u4 4 4 V' ... C t t zu v u1 u 2 V' s ... u4v4 C t u3 c 4 4 V' s C u1 u 2 wu v u4 ... C u1 u 2 u3 v4 u2v2 c u2v2 t zu v 2 2 wu v 2 2 s u2v2 u2v2 wu v 2 2 s t u2v2 wu v yu v 2 2 2 2 u2v2 xu v 2 2 Figure 7: The respective star-decompositions of the gadgets Fui vi are connected to the other bags. By construction, the resulting tree decomposition (T 0 , X 0 ) of G0 is a star-decomposition, hence tb(G0 ) = 1. Recall that we can use our reduction from Definition 2 in order to prove that computing pathlength and path-breadth is NP-hard. By contrast, our reduction from Theorem 4 cannot be used to prove that tree-length is NP-hard to compute (in fact, the graph G0 resulting from the reduction has tree-length two). Finally, as in the previous Section 3.1, let us strenghten Theorem 4 with an inapproximability result. Corollary 4. For every ε > 0, the tree-breadth of a graph cannot be approximated within a factor 2 − ε unless P=NP. 10 4 General properties of graphs with tree-breadth one In Section 3.2, we prove that computing the tree-breadth is NP-hard. In particular, the recognition of graphs with tree-breadth one is NP-complete. In light of this result, we focus on graphs with tree-breadth one (in order to obtain a better understanding of what makes the problem difficult) Problem 3 (1-tree-breadth). Input: a connected graph G = (V, E) Question: tb(G) ≤ 1 ? In Lemma 7, we show that the problem of recognizing graphs with tree-breadth at most one is equivalent to the problem of computing tree-breadth. This further motivates our study of these graphs. Then, we will prove necessary conditions for a graph to be of tree-breadth one. • One is that all graphs with a star-decomposition have a domination elimination ordering (see Section 4.1). We will outline a few implications of this property. • Second, we will prove in Lemma 9 that if a graph G admits a star-decomposition then so do all the blocks of G, where the blocks here denote a particular case of induced subgraphs of G (e.g., see Definition 4). Finally, we will obtain from the latter result a polynomial-time algorithm to decide whether a bipartite graph has tree-breadth at most one (e.g., see Section 4.3). Definition 3. Let G be a graph with n vertices, denoted by v1 , v2 , . . . , vn , and let r be a positive integer. The graph G0r is obtained from G by adding a clique with n vertices, denoted by U = {u1 , u2 , . . . , un }, so that for every 1 ≤ i ≤ n, vertex ui is adjacent to BG (vi , r) = {x ∈ V (G) | distG (vi , x) ≤ r}. Lemma 7. For every graph G, for every positive integer r, let G0r be as defined in Definition 3, tb(G) ≤ r if and only if tb(G0r ) ≤ 1. Proof. If tb(G) ≤ r then we claim that starting from any tree decomposition (T, X ) of G with breadth at most r, one obtains a star-decomposition of G0r by adding the clique U in every bag Xt , t ∈ V (T ). Indeed, in such case for every bag Xt , t ∈ V (T ), by the hypothesis there is vi ∈ V (G) such that maxx∈Xt distG (vi , x) ≤ r, whence Xt ∪ U ⊆ NG0r [ui ]. Conversely, if tb(G0r ) ≤ 1 then we claim that starting from any tree decomposition (T 0 , X 0 ) of G0r with breadth at most one, one obtains a tree decomposition of G with breadth at most r by removing every vertex of the clique U from every bag Xt0 , t ∈ V (T 0 ). Indeed, in such case for every bag Xt0 , t ∈ V (T 0 ), by the hypothesis there is y ∈ V (G0r ) such that Xt0 ⊆ NG0r [y]. Furthermore, y ∈ {ui , vi } for some 1 ≤ i ≤ n, and so, since NG0r [vi ] ⊆ NG0r [ui ] by construction, therefore Xt0 \ U ⊆ NG0r (ui ) \ U = {x ∈ V (G) | distG (vi , x) ≤ r}. 4.1 Existence of specific elimination orderings Independently from the remaining of the section, let us prove some interesting properties of graphs with tree-breadth one in terms of elimination orderings. More precisely, a domination elimination ordering [23] of a graph G is a total ordering of its vertex-set, denoted by v1 , v2 , . . . , vn , so that for every 1 ≤ i < n, there is j > i satisfying that NG (vi ) ∩ {vi+1 , vi+2 , . . . , vn } ⊆ NG [vj ]. The existence of domination elimination orderings in some graph classes and their algorithmic applications has been studied in [18]. Let us prove that graphs with tree-breadth one all admit a domination elimination ordering. Lemma 8. Let G be such that tb(G) ≤ 1, G admits a domination elimination ordering. Proof. Assume G has at least two vertices (or else, the lemma is trivial). To prove the lemma, it suffices to prove the existence of u, v ∈ V (G) distinct such that N (v) ⊆ N [u] and tb(G \ v) ≤ 1 (then, the lemma follows by induction on the order of the graph). If G admits a universal vertex u, then one can pick v ∈ V (G) \ u arbitrary, N (v) ⊆ N [u] because u is universal in G, furthermore tb(G \ v) ≤ 1 because G \ v admits a universal vertex u. Else, G does not admit any universal vertex, let (T, X ) be a reduced tree decomposition of G of breadth one, that is a star-decomposition by Lemma 3. Let Xt , t ∈ V (T ) be a leaf. Since the tree decomposition is reduced, there must be v ∈ Xt satisfying Tv = {Xt }. Now there are two cases. 11 • Suppose there is u ∈ Xt \ v such that Xt ⊆ N [u]. Then, N (v) ⊆ Xt ⊆ N [u], and tb(G \ v) ≤ 1 because G \ v can be obtained from G by contracting the edge {u, v} and tree-breadth is contraction-closed by Lemma 2. • Else, Xt ⊆ N [v], and for every x ∈ Xt \ v, Xt 6⊆ N [x]. Let t0 ∈ V (T ) be the unique node adjacent to node t in T , that exists because G does not admit any universal vertex and so, T has at least two bags. Let us assume that for every x ∈ Xt \ v, x ∈ Xt ∩ Xt0 (for otherwise, N (x) ⊆ N [v] and tb(G \ x) ≤ 1 because G \ x can be obtained from G by contracting the edge {v, x} to v and tree-breadth is contraction-closed by Lemma 2). In particular, let u ∈ Xt0 satisfy Xt0 ⊆ N [u]. Then, N (v) = Xt ∩ Xt0 ⊆ N [u], furthermore tb(G \ v) ≤ 1 because (T \ t, X \ Xt ) is a star-decomposition of G \ v. Note that for a graph to have tree-breadth one, it must satisfy the necessary condition of Lemma 8 and this can be checked in polynomial-time. However, the existence of some domination elimination ordering is not a sufficient condition for the graph to have tree-breadth one. Indeed, every grid has a domination elimination ordering but the tree-length of the n×m grid is at least min{n, m}−1 [26] (recall that tl(G) ≤ 2tb(G) for any graph G). The existence of a domination elimination ordering has some interesting consequences about the graph structure. Let us recall one such a consequence about the cop-number of the graph. Corollary 5. For any graph G with tb(G) ≤ 1, G has cop-number ≤ 2 and the upper-bound is sharp. Proof. By Lemma 8, G admits a domination elimination ordering. Therefore, by [21, Theorem 4] G has cop-number ≤ 2. One can prove the sharpness of the upper-bound by setting G := C4 , the cycle with four vertices. 4.2 Properties of particular decompositions In the following, it will be useful not only to constrain the properties of the star-decomposition whose existence we are interested in, but also to further constrain the properties of the graph G that we take as input. Let us first remind basic terminology about graph separators. Definition 4. Let G = (V, E) be connected, a separator of G is any subset S ⊆ V such that G \ S has at least two connected components. In particular, a full component for S is any connected component C of G\S satisfying N (C) = S. A block is any induced subgraph G[C ∪ S] with S being a separator and C being a full component for S. Finally, a minimal separator is a separator with at least two full components. Our objective is to prove that if a graph G has tree-breadth one then so do all its blocks. In fact, we will prove a slightly more general result: Lemma 9. Let G = (V, E) be a graph, S ⊆ V be a separator, and W ⊆ V \ S be the union of some connected components of G \ S. If tb(G) = 1 and W contains a full component for S, then tb(G[W ∪ S]) = 1. More precisely if (T, X ) is a tree decomposition of G of breadth one, then (T, {Xt ∩ (W ∪ S) | Xt ∈ X }) is a tree decomposition of G[W ∪ S] of breadth one. Proof. Let (T, X ) be a tree decomposition of breadth one of G. Let us remove all vertices in V \(W ∪S) from bags in (T, X ), which yields a tree decomposition (T 0 , X 0 ) of the induced subgraph G[W ∪ S]. To prove the lemma, we are left to prove that (T 0 , X 0 ) has breadth one. Let Xt be a bag of (T 0 , X 0 ). By construction, Xt is fully contained into some bag of (T, X ), so it has radius one in G. Let v ∈ V be such that Xt ⊆ NG [v]. If v ∈ W ∪ S, then we are done. Else, since for all x∈ / S ∪ W, N (x) ∩ (S ∪ W ) ⊆ S (because S is a separator by the hypothesis), we must have that Xt ⊆ S. Let A ⊆ W be a full component for S, that exists by the hypothesis, and let TA be the subtree that is induced by the bags intersecting the component. Since we have that the subtree TA and the subtrees Tx , x ∈ Xt pairwise intersect — because for all x ∈ Xt ,x ∈ S and so, x has a T neighbour in A —, then by the Helly property (Lemma 1) TA ∩ x∈Xt Tx 6= ∅ i.e., there exists a bag in (T, X ) containing Xt and intersecting A. Moreover, any vertex dominating this bag must be either in S or in A, so in particular there exists u ∈ A ∪ S dominating Xt , which proves the lemma. 12 Lemma 9 implies that, under simple assumptions, a graph of tree-breadth one can be disconnected using any (minimal) separator, and the components must still induce subgraphs with treebreadth one. The converse does not hold in general, yet there are interesting cases when it does. Lemma 10. Let G = (V, E) be a graph, S ⊆ V be a clique-minimal-separator and A be a full component for S. Then, tb(G) = 1 if and only if both tb(G[A ∪ S]) = 1 and tb(G[V \ A]) = 1. Proof. By the hypothesis V \ (A ∪ S) contains a full component because S is a minimal separator. Therefore, if G has tree-breadth one, then so do G[A ∪ S] and G[V \ A] by Lemma 9. Conversely, suppose that we have both tb(G[A ∪ S]) = 1 and tb(G[V \ A]) = 1. Let (T 1 , X 1 ) be a tree decomposition of G[A ∪ S] with breadth one, let (T 2 , X 2 ) be a tree decomposition of G[V \ A] with i breadth one. Then for every i ∈ {1, 2} we have that since T S isi a clique the subtrees Ts , s ∈ S, pairwise intersect, so by the Helly Property (Lemma 1) s∈S Ts 6= ∅ i.e., S is fully contained into some bag of (T 1 , X 1 ) and it is fully contained into some bag of (T 2 , X 2 ). Moreover, (A ∪ S) ∩ (V \ A) = S, therefore a tree decomposition of G with breadth one can be obtained by adding an edge between some bag of (T 1 , X 1 ) containing S and some bag of (T 2 , X 2 ) containing S. Recall that computing the clique-minimal-decomposition of a graph G can be done in O(nm)time, where m denotes the number of edges [9]. By doing so, one replaces a graph G with the maximal subgraphs of G that have no clique-separator, a.k.a. atoms. Therefore, we will assume in the remaining of the proofs that there is no clique-separator in the graphs that we will study, we will call them prime graphs. 4.3 Application to bipartite graphs In this section, we describe an O(nm)-time algorithm so as to decide whether a prime bipartite graph has tree-breadth one. This combined with Lemma 10 proves that it can be decided in polynomial-time whether a bipartite graph has tree-breadth one. We will first describe a more general problem and how to solve it in polynomial-time. Tree decompositions with constrained set of bags. Our algorithm for bipartite graphs makes use of the correspondance between tree decompositions and triangulations of a graph. Indeed, recall that any reduced tree decomposition (T, X ) of a graph G is a clique-tree for some chordal supergraph H of G whose maximal cliques are the bags of X . Conversely, for any chordal supergraph H of G, every clique-tree of H is a tree decomposition of G whose bags are the maximal cliques of H [33]. Therefore as shown below, the following subproblem can be solved in polynomial-time: Problem 4. Input: a graph G, a family X of subsets of V (G). Question: Does there exist a tree T such that (T, X ) is a tree decomposition of G ? Let us assume w.l.o.g. that no subset of X is properly contained into another one. To solve Problem 4.3, it suffices to make every subset X ∈ X a clique in G, then to verify whether the resulting supergraph H of G is a chordal graph whose maximal cliques are exactly the sets in X . Since chordal graphs can be recognized in linear-time, and so can be enumerated their maximal cliques [32], therefore Problem 4.3 can be solved in polynomial-time. The algorithm for bipartite graphs. Now, given a bipartite graph G, we aim to exhibit a family X so that tb(G) = 1 if and only if there is a star-decomposition of G whose bags are X . By doing so, we will reduce the recognition of bipartite graph with tree-breadth at most one to the more general Problem 4.3. Lemma 11. Let G = (V0 ∪ V1 , E) be a prime bipartite graph with tree-breadth one. There is (T, X ) a star-decomposition of G such that either X = {N [v0 ] | v0 ∈ V0 }, or X = {N [v1 ] | v1 ∈ V1 }. Proof. Let (T, X ) be a star-decomposition of G, that exists by Lemma 3, minimizing the number |X | of bags. Suppose there is some v0 ∈ V0 , there is t ∈ V (T ) such that Xt ⊆ NG [v0 ] (the case when there is some v1 ∈ V1 , there is t ∈ V (T ) such that Xt ⊆ NG [v1 ] is symmetrical to this one). We claim that for every t0 ∈ V (T ), there exists v00 ∈ V0 satisfying Xt0 ⊆ NG [v00 ]. By contradiction, 13 let v0 ∈ V0 , v1 ∈ V1 , let t, t0 ∈ V (T ) be such that Xt ⊆ NG [v0 ], Xt0 ⊆ NG [v1 ]. By connectivity of the tree T we may assume w.l.o.g. that {t, t0 } ∈ E(T ). Moreover, NG (v0 ) ∩ NG (v1 ) = ∅ because G is bipartite. Therefore, Xt ∩ Xt0 ⊆ {v0 , v1 }, and in particular if Xt ∩ Xt0 = {v0 , v1 } then v0 , v1 are adjacent in G. However, by the properties of a tree decomposition this implies that Xt ∩ Xt0 is a clique-separator (either an edge or a single vertex), thus contradicting the fact that G is prime. Now, let v0 ∈ V0 be arbitrary. We claim that there is a unique bag Xt , t ∈ V (T ), containing v0 . Indeed, any such S bag Xt must satisfy Xt ⊆ NG [v0 ], whence the subtree Tv0 can be contracted into a single bag t∈Tv Xt without violating the property for the tree decomposition to be a star0 decomposition. As a result, the unicity of the bag Xt follows from the minimality of |X |. Finally, since Xt is unique and Xt ⊆ NG [v0 ], therefore Xt = NG [v0 ] and so, X = {N [v0 ] | v0 ∈ V0 }. We can easily deduce from Lemma 11 the following algorithm for deciding whether a prime bipartite graph G has tree-breadth one. Let (V0 , V1 ) be the (unique) bipartition of the vertex-set of G into two stable sets. Let X0 = {N [v0 ] | v0 ∈ V0 }, let X1 = {N [v1 ] | v1 ∈ V1 }. By Lemma 11, tb(G) = 1 if and only if one of (G, X0 ), (G, X1 ) is a yes-instance of Problem 4.3. 5 Algorithm for planar graphs We are now ready to present our main result. In this section, we describe a quadratic-time algorithm for deciding whether a prime planar graph has tree-breadth one. Overall, we claim that it gives us a quadratic-time algorithm for deciding whether a general planar graph has tree-breadth one. Indeed, the clique-decomposition of a planar graph takes O(n2 )-time to be computed, furthermore the disjoint union of the atoms has O(n + m) vertices [9], that is O(n) for planar graphs. Roughly, we will construct a star-decomposition of the graph by increments. The main principle of the recursive algorithm is to find a particular vertex, called leaf-vertex. Informally, it extracts a new bag of the star-decomposition from some ball around the leaf-vertex. Then, depending on the case, either the leaf-vertex vertex is removed or some edge is added or contracted. In both cases, the resulting graph remains prime and planar and has tree-breadth one if and only if the initial one has tree-breadth one. We prove that each inductive step takes a linear time. Moreover, we prove that there are at most a linear number of recursive iterations (Lemma 27). There are three kinds of leaf-vertices (e.g., see Figure 8). Definition 5. Let G = (V, E) be a graph. A vertex v is a leaf-vertex if one of the following conditions hold. Type 1. N (v) induces an av bv -path for some av , bv ∈ V \ {v}, denoted by Πv , of length at least 3 and there exists dv ∈ V \ {v} such that N (v) ⊆ N (dv ), i.e., dv dominates Πv . Type 2. N (v) induces a path, denoted by Πv = (av , bv , cv ), of length 2. Type 3. N (v) consists of two non adjacent vertices av and cv , and there exists bv ∈ (N (av ) ∩ N (cv )) \ {v}. av dv cv av bv av v v Type 1 cv Type 2 Figure 8: The three kinds of leaf-vertices. We are now ready to describe the algorithm. 14 bv v Type 3 cv 5.1 Algorithm Leaf-BottomUp Let G = (V, E) be prime planar graph. Assume G has at least 7 vertices (else, it is easy to conclude). Step 1 The first step is to find a leaf-vertex in G. In Section 5.4.1, we describe how to decide whether G has a leaf-vertex in linear-time. • if G has no leaf-vertex, then, by Theorem 7, no minimal separator of G induces a path of length 2. Therefore, by Lemma 20, tb(G) = 1 only if G has a star-decomposition with at most 2 bags. In that case, Algorithm Leaf-BottomUp checks whether it exists a star-decomposition with at most 2 bags, which can be done in quadratic time (see Lemma 26). If it exists, then tb(G) = 1. Otherwise, tb(G) > 1. • otherwise, let v be a leaf-vertex of G and go to Step 2 if v is of Type 1 and go to Step 3 otherwise. Step 2 Case v is of Type 1. Let Πv and dv be defined as in Definition 5. If V = N [v] ∪ {dv } then trivially tb(G) = 1. Else by Theorem 8, G0 is prime and planar, where G0 is the graph obtained from G \ v by contracting the internal nodes of Πv to a single edge, and tb(G) = 1 if and only if tb(G0 ) = 1. In that case, Algorithm Leaf-BottomUp is recursively applied on G0 . Step 3 Case v is of Type 2 or 3. Let av , bv , cv be defined as in Definition 5. In that case, Algorithm Leaf-BottomUp checks whether G \ v is prime. By Theorem 6, for any clique minimal separator S of G \ v (if any), there exists uv ∈ V \ {av , bv , cv , v} such that S = {bv , uv }. Therefore, this can be checked in linear time (by checking with a Depth-FirstSearch whether there is a cut-vertex of G \ {av , bv , cv , v} in the neighbors of bv ). If G \ v is prime then go to Step 3.1, else go to Step 3.2. Step 3.1 Case v is of Type 2 or 3 and G\v is prime. There are 4 cases that can be determined in linear-time. (a) Case |N (av ) ∩ N (cv )| ≥ 3 in G \ v, or there exists a minimal separator S ⊆ (N (av ) ∩ N (cv )) ∪ {av , cv } in G \ v and {av , cv } ⊆ S. By Theorem 9, tb(G) = 1 if and only if tb(G \ v) = 1. Since, moreover, G \ v is planar and prime, then Algorithm Leaf-BottomUp is recursively applied on G \ v. (b) Case: |N (av ) ∩ N (cv )| < 3 in G \ v and there is no minimal separator S ⊆ (N (av ) ∩ N (cv )) ∪ {av , cv } in G \ v such that {av , cv } ⊆ S. i Subcase: |N (av ) ∩ N (cv )| = 1 in G \ v. In that subcase, N (av ) ∩ N (cv ) = {v, bv } and, by Theorem 10, tb(G) = 1 if and only if G = C4 , a cycle with four vertices. Note that here it implies that tb(G) > 1 because G has at least 7 vertices. ii Subcase: |N (av ) ∩ N (cv )| = 2 in G \ v. In that subcase, let N (av ) ∩ N (cv ) = {v, bv , uv }. By Theorem 11, since G has more than 5 vertices, the graph G0 obtained from G by adding edges {v, uv } and {bv , v} is planar and prime, and moreover tb(G) = 1 if and only if tb(G0 ) = 1. In that case, Algorithm Leaf-BottomUp is recursively applied on G0 . Step 3.2 Case v is of Type 2 or 3 and G \ v has a clique separator. As mentioned in Step 3, in that case, there exists uv ∈ V \ {av , bv , cv , v} such that S = {bv , uv } is a minimal clique separator of G \ v. Moreover, by Theorem 6, G \ {av , bv , cv , v} is connected. By Theorem 12, tb(G) = 1 if and only if tb(G0 ) = 1 where G0 is obtained from G by adding the edge {v, bv } (if it were not already there). Moreover, G0 is prime and planar. Hence, we may assume that {v, bv } ∈ E (if not Algorithm Leaf-BottomUp adds it). Furthermore by Theorem 6, since G has more than 5 vertices, uv ∈ / N (av ) ∩ N (cv ). In the latter case, let us assume w.l.o.g. that uv ∈ / N (av ), that is either uv ∈ / N (av ) ∪ N (cv ) or uv ∈ N (cv ) \ N (av ). There are several cases to be considered. (a) Case uv ∈ / N (av ) ∪ N (cv ), or (N (uv ) ∩ N (av )) ∪ {v, cv } does not separate uv and av in G. By Theorem 13, G/vav is prime and planar, and tb(G) = 1 if and only if tb(G/vav ) = 1. In that case, Algorithm Leaf-BottomUp is recursively applied on G/vav , the graph obtained from G by contracting the edge {v, av }. 15 (b) Case uv ∈ N (cv ) \ N (av ), and (N (uv ) ∩ N (av )) ∪ {v, cv } separates uv and av in G. In that case, recall that G has at least 7 vertices. Again, Algorithm Leaf-BottomUp distinguishes several subcases. i Subcase N (bv ) = {v, av , cv , uv }. In that subcase, by Theorem 14, we can find in linear-time a vertex x ∈ (N (av ) ∩ N (uv )) \ {bv } such that G0 is planar, where G0 is obtained from G by adding the edge {bv , x}. Moreover, by Theorem 15, G0 /bv x (obtained by contracting {bv , x}) is prime and tb(G) = 1 if and only if tb(G0 /bv x) = 1. In that case, Algorithm Leaf-BottomUp is recursively applied on G0 /bv x. ii Subcase {v, av , cv , uv } ⊂ N (bv ) and N (bv )∩N (av )∩N (uv ) 6= ∅. In that subcase, |N (bv ) ∩ N (av ) ∩ N (uv )| = 1 by Lemma 23 and let x be this common neighbor. By Theorem 15, G/bv x (obtained by contracting {bv , x}) is prime and tb(G) = 1 if and only if tb(G/bv x) = 1. In that case, Algorithm Leaf-BottomUp is recursively applied on G/bv x. iii Subcase {v, av , cv , uv } ⊂ N (b) and N (bv ) ∩ N (av ) ∩ N (uv ) = ∅. In that subcase, by Theorem 16, there must be a unique x ∈ (N (av ) ∩ N (uv )) \ {bv } such that N (bv ) ∩ N (x) is a bv x-separator of G and |N (bv ) ∩ N (x)| ≥ 3 (or else, tb(G) > 1). • Suppose there is a leaf-vertex ` ∈ N (bv ) ∩ N (x). By Lemma 24, ` is of Type 1 or G \ ` is prime. In that case, go to Step 2 if ` is of Type 1, and go to Step 3.1 if ` has Type 2 or 3 (in both cases, ` takes the role of v). Note that we never go back to Step 3.2 in such case, so the algorithm cannot loop. • Otherwise, by Theorem 17, there exist y, z ∈ N (bv )∩N (x) two non-adjacent vertices, such that G0 is prime and planar, and tb(G) = 1 if and only if tb(G0 ) = 1, where G0 is obtained from G by adding the edge {x, y}. In that case, Algorithm Leaf-BottomUp is recursively applied on G0 . 5.2 5.2.1 Properties of prime planar graphs with tree-breadth one General lemmas We will first investigate on general properties of prime planar graphs. In particular, the following properties do not depend on the existence of a star-decomposition, therefore we do not use tree decompositions in our proofs. However, note that we refer to Definition 5 in Theorem 6. For clarity, we will separate the properties that hold for every biconnected planar graph from those that only hold for prime planar graphs. Properties of biconnected planar graphs. In order to obtain these properties, we will mostly rely on the notion of intermediate graphs, defined below. Definition 6. [15, Definition 6] Let G = (V, E) be a planar graph. We fix a plane embedding of G. Let F be the set of faces of this embedding. The intermediate graph GI = (V ∪ F, EI ) has vertex-set V ∪ F , furthermore E ⊆ EI and we add an edge in GI between an original vertex v ∈ V and a face-vertex f ∈ F whenever the corresponding vertex and face are incident in G (see Figure 9). Figure 9: A plane embedding of some planar graph (left) and the corresponding intermediate graph (right). Face-vertices are coloured in red. Note that an intermediate graph is planar. Furthermore, since a plane embedding can be constructed in linear-time [37], therefore so can be an intermediate graph. This is important for the quadratic-time complexity of Algorithm Leaf-BottomUp. To prove the correctness of Algorithm Leaf-BottomUp in the following, we will rely upon the following property of intermediate graphs. 16 Lemma 12. [15, Proposition 9] Let S be a minimal separator of some biconnected planar graph G = (V, E) and let C be a full component of G \ S. We fix a plane embedding of G. Then S corresponds to a cycle vS (C) of GI , of length 2|S| and with V ∩ vS (C) = S, and such that GI \ vS (C) has at least two connected components. Moreover, the original vertices of one of these components are exactly the vertices of C. In the following, we will rely upon two properties which both follow from Lemma 12. The first one is the following structural property of minimal separators of planar graphs. Corollary 6. Let S be a minimal separator of a biconnected planar graph G = (V, E). Then, S either induces a cycle or a forest of paths. Proof. Let us fix a plane embedding of G, let GI be the corresponding intermediate graph. Then, let CS be a smallest cycle of GI such that V ∩ CS = S, that exists by Lemma 12. To prove the corollary, it suffices to prove that CS is an induced cycle of GI . By contradiction, assume the existence of a chord xy of CS . Note that x ∈ S or y ∈ S because face-vertices are pairwise nonadjacent in GI . Therefore assume w.l.o.g. that x ∈ S. Let us divide CS in two cycles C1 , C2 such that C1 ∩ C2 = {x, y}. By the minimality of CS , S intersects both C1 \ C2 and C2 \ C1 . Therefore, let z1 , z2 ∈ S such that z1 ∈ C1 \ C2 and z2 ∈ C2 \ C1 . Finally, let A, B be two full components of G \ S. Observe that (A ∪ B) ∩ (C1 ∪ C2 ) = ∅ because V ∩ CS = S. Let us contract C1 , C2 in order to obtain the two triangles (z1 , x, y) and (z2 , x, y). In such case, there is a K3,3 -minor of GI with {A, B, y} and {x, z1 , z2 } being the respective sides of the bipartition, thus contradicting the fact that GI is planar. Therefore, CS is an induced cycle of GI and so, S induces a subgraph of a cycle in G, that is either a cycle or a forest of path. On the algorithmic side, one can also deduce from Lemma 12 the following corollary. Corollary 7. Let G be a biconnected planar graph, let S be a minimal separator of G. There is a planar supergraph GS of G with same vertex-set so that S either induces an edge (if |S| = 2) or a cycle of GS , and it can be constructed in linear-time. Figure 10: Addition of edges in a planar graph G so as to make a minimal separator of G induce a cycle. Proof. Let us fix a plane embedding of G, let GI be the corresponding intermediate graph. For every face-vertex f of GI , let us consider Sf = S ∩NGI (f ). We first claim that |Sf | ≤ 2. Indeed, let A, B be two full components of G \ S, let us contract them to any two vertices a ∈ A, b ∈ B. Then, there is a K3,|Sf | -minor of GI with {a, b, f } and Sf being the respective parts of the bipartition. Since GI is planar by construction, therefore, |Sf | ≤ 2. Now, the graph GS is constructed from G as follows (we refer to Figure 10 for an illustration of the proof). For every face-vertex f of GI , if Sf = (x, y) then we add the edge {x, y} in GS . Note that GS is a minor of GI and so, it is a planar graph. Moreover, by Lemma 12 there is a cycle of GI whose original vertices are exactly S and so, S induces a connected subgraph of GS . In particular if |S| = 2, then it must be an edge. Else, |S| > 2 and the connected subgraph GS [S] contains a cycle by construction. Since S is a minimal separator of GS by construction and GS [S] is not acyclic, it follows from Corollary 6 that S induces a cycle of GS . We will often make use of the routine of Corollary 7 in order to prove the quadratic-time complexity of Algorithm Leaf-BottomUp. Properties of prime planar graphs. Unlike the above Corollaries 6 and 7 (which hold for every biconnected planar graph), the following results only hold for prime planar graphs. We will make use of the following structural properties of prime planar graphs in order to prove the correctness of Algorithm Leaf-BottomUp. Lemma 13. Let G = (V, E) be a prime graph that is K3,3 -minor-free. Let v ∈ V , for every minimal separator S ⊆ NG (v) of the subgraph G \ v, S consists of two non-adjacent vertices. 17 Proof. Let S ⊆ NG (v) be a minimal separator of G \ v. There must exist two full components A and B of S in G \ (S ∪ {v}). Let us remove all nodes of the components of G \ (S ∪ {v}) but the ones in A or B. Then, let us contract A (resp., B) in a single vertex a (resp., b). We get a K3,|S| as a minor of G where {a, b, v} is one part of the bipartition, and so |S| ≤ 2. Finally, since S ∪ {v} is also a separator of G, then |S| ≥ 2 because otherwise S ∪ {v} would be an edge-separator. Therefore, |S| = 2 and it is a stable set because otherwise there would be a clique-separator of size 3 in G. Lemma 14. Let G be a prime planar graph, let the path Π = (a, b, c) be a separator of G, and let C be a component of G \ Π. Then, there is at most one common neighbour of a, b in C. Proof. First note that Π is induced or else it would be a clique-separator of G. Furthermore, a, c ∈ N (C 0 ) for every component C 0 of G \ (Π ∪ C) or else N (C 0 ) would be a clique-separator of G (either a vertex-separator or an edge-separator). In particular, it is always possible to make vertices a, c adjacent by contracting an arbitrary component of G \ (Π ∪ C). G\C a c b Q Q' u u' Figure 11: Case where the paths Q and Q0 are internally vertex-disjoint paths. By contradiction, let u, u0 ∈ N (a) ∩ N (b) ∩ C be distinct. We claim that there exists a uc-path Q in C ∪ {c} that does not contain u0 , because else the triangle a, b, u0 would separate u from Π, that contradicts the fact that G is prime. By symmetry, there also exists a u0 c-path Q0 in C ∪ {c} that does not contain u. There are two cases. • Q and Q0 are internally vertex-disjoint paths (see Figure 11 for an illustration). Let us contract Q \ c, Q0 \ c to the vertices u, u0 , let us contract an arbitrary component of G \ (Π ∪ C) in order to make vertices a, c adjacent, then let us contract a path from Q to Q0 in C (that exists, because C is connected by the hypothesis) in order to make vertices u, u0 adjacent. Then one obtains from a, b, c, u, u0 a K5 -minor, which contradicts the fact that G is planar. G\C a c b R y u Q Q' u' Figure 12: Case where the paths Q and Q0 intersect. • Q and Q0 intersect (see Figure 12 for an illustration). Let y ∈ (Q ∩ Q0 ) \ c be such that the uy-subpath of Q does not intersect Q0 . Let R be the yc-subpath of Q0 . We may assume w.l.o.g. that R ⊆ Q ∩ Q0 for the remaining of the proof, whence Q ∩ Q0 = R. Let us contract Q\R, Q0 \R, R\c in order to make vertices u, u0 , c adjacent to vertex y, then let us contract an arbitrary component of G \ (P ∪ C) in order to make vertices a, c adjacent. One obtains from a, b, c, u, u0 , y a K3,3 -minor with {a, b, y} being one side of the bipartition, that contradicts the fact that G is planar. 18 Lemma 15. Let G be a prime planar graph, let the path Π = (a, b, c) be a separator of G, and let C be a component of G \ Π. Suppose there is some vertex v ∈ C that is a common neighbour of a, b, c. Then, either C is reduced to v, or (a, v, c) is a separator of G. Furthermore, in the latter case, the path (a, v, c) separates vertex b from C \ v. Proof. Let us assume that C \ v 6= ∅. Let D be a connected component of G[C \ v]. Note that v ∈ N (D) because C is a connected component of G \ Π by the hypothesis. To prove the lemma, it suffices to prove that b ∈ / N (D). By contradiction, suppose that b ∈ N (D) (see Figure 13 for an illustration). Since v, b, a and v, b, c are pairwise adjacent and G has no clique-separator by the hypothesis, then necessarily N (D) = {a, b, c, v}. G\C a b D c v Figure 13: Existence of a component D ⊆ C \ v that is adjacent to b. Let us contract the component D to a single vertex x. Then, let C 0 be any component of G \ (Π ∪ C). We have that a, c ∈ N (C 0 ) or else N (C 0 ) would be a clique-separator of G (either a vertex-separator or an edge-separator). So, let us contract the component C 0 onto vertex a in order to make a and c adjacent. One obtains from a, b, c, v, x a K5 -minor, which contradicts the fact that G is planar. We recall that the gist of Algorithm Leaf-BottomUp is (informally) to try to remove a leaf-vertex v from G then to apply recursively the algorithm on G \ v. Because the algorithm is strongly dependent on the fact that G is prime, it is important to characterize the cases when G \ v is also prime. Indeed, new clique-decompositions are needed when G \ v is not prime, which may provoke a combinatorial explosion of the number of subgraphs to be considered. Therefore, before we conclude this section, let us characterize whenever there may be clique-separators in G \ v with v being a leaf-vertex. This will first require the following lemma. Lemma 16. Let G = (V, E) be a graph and let the path Π = (a, b, c) be a separator of G. Let C be the union of some components of G \ Π and let S be a separator of G[C ∪ Π]. Then, S is a separator in G or S separates a and c in G[C ∪ Π]. Moreover, in the latter case, G[C ∪ Π] \ S has exactly two components Ca and Cc containing a and c respectively. Proof. There are two cases. • Suppose there exists a component D of G[C ∪ Π] \ S such that NG (D) ⊆ C ∪ Π. Since D ∩ S = ∅, S 6= V \ D and NG (D) ⊆ S therefore S is a separator of G with D being a component of G \ S. • Else, every component D of G[C ∪ Π] \ S has a neighbour in V \ (C ∪ Π). The latter implies that D ∩ Π 6= ∅ for every component D of G[C ∪ Π] \ S because C is a union of components of G\Π by the hypothesis. In particular, since there exist at least two components of G[C ∪Π]\S then there must be one containing an endpoint of Π. W.l.o.g. assume there is a component Ca of G[C ∪ Π] \ S such that a ∈ Ca . Let Cc be any component of G[C ∪ Π] \ (S ∪ Ca ). We have that Cc ∩ N [Ca ] = ∅ because S separates Ca and Cc in G[C ∪ Π]. Therefore, a, b ∈ / Cc and so, c ∈ Cc because Cc ∩ Π 6= ∅. This finally proves that G[C ∪ Π] \ S has exactly two components Ca and Cc containing a and c respectively. Theorem 6. Let G = (V, E) be a prime planar graph, let v be a leaf-vertex of Type either 2 or 3 and let Πv = (av , bv , cv ) be as defined in Definition 5. Suppose that there exists a minimal separator S in G \ v that is a clique. Then, S = {uv , bv }, uv ∈ / Πv and the following hold: • V \ (Πv ∪ {v}) is a full component of G \ Πv . 19 • If uv ∈ N (av ) (resp. uv ∈ N (cv )), then av (resp. cv ) is simplicial in G \ v with neighbours {uv , bv }; • Furthermore uv ∈ / N (av ) ∩ N (cv ) unless V = Πv ∪ {uv , v}; uv av bv uv cv av v bv uv cv av v bv cv v Figure 14: Existence of a clique-separator in G \ v. Proof. Note that the subgraph G\v is planar and S is a minimal separator of G\v by the hypothesis, therefore by Corollary 6 either S induces a cycle or a forest of path. Since in addition S is a clique by the hypothesis, it follows that S either induces a singleton, an edge or a triangle. Since S is a clique and G is prime, S is not a separator of G, so by Lemma 16 with C = V \ (Πv ∪ {v}), S is an av cv -separator of G \ v. This both implies that bv ∈ S and S 0 := S ∪ {v} is a minimal av cv -separator of G. In particular, S being a strict subset of some minimal separator of G it cannot induce a cycle (by Corollary 6), hence it must induce either a singleton or an edge. Furthermore, still by Lemma 16 with C = V \(Πv ∪{v}) there exist exactly two components Ca , Cc of G\(S∪{v}), with av ∈ Ca , cv ∈ Cc . As a result, S \ bv 6= ∅, or else {av , bv }, {bv , cv } would be edge-separators of G, thus contradicting the hypothesis. Let S = {uv , bv }, uv ∈ / Πv ∪ {v}. If uv ∈ N (av ), then Ca \ av = ∅ (and so, av is simplicial in G \ v), for otherwise (av , uv , bv ) would be a clique-separator of G. Similarly, if uv ∈ N (cv ) then Cc \ cv = ∅ (and so, cv is simplicial in G \ v). In particular if uv ∈ N (av ) ∩ N (cv ) then Πv ∪ {uv , v} = V . Last, as there exists an av cv -path in every component C 0 of G \ (Πv ∪ {v}) because G has no clique-separator by the hypothesis, therefore uv ∈ C 0 . This implies that V \ (Πv ∪ {v}) is a full component of G \ Πv . 5.2.2 Constrained star-decompositions In the following, it will be useful to impose additional structure on the star-decompositions. In order to do that, we will prove properties on some pairs of vertices in the graph. Namely, we will prove that when x, y ∈ V satisfy a few technical conditions, then it can be assumed that Tx ∪ Ty is a subtree of the star-decomposition (T, X ). Lemma 17. Let G be a connected graph with tb(G) = 1, let x, y ∈ V (G) be non-adjacent (and x 6= y). Suppose the pair (x, y) satisfies that for every xy-separator S of G, if there is z ∈ / {x, y} that dominates S then z ∈ NG (x) ∩ NG (y). Then, there is a star-decomposition (T, X ) of G with Bx , By ∈ X , x ∈ Bx , y ∈ By and either Bx = By or Bx , By are adjacent in T . Moreover, in the latter case, Bx ⊆ N [x], By ⊆ N [y]. Proof. Consider a star-decomposition (T, X ) of G, that exists by Lemma 3. If x and y are not in a same bag, let Bx and By be the bags containing respectively x and y and as close as possible in T . By the properties of a tree decomposition, N (x) ∩ N (y) ⊆ Bx ∩ By . Hence, for any bag B between Bx and By in T , N (x) ∩ N (y) ⊆ B. • Case 1: If Bx and By are not adjacent in T , let B be any bag in the path between Bx and By in T . By the properties of a tree decomposition, B is an xy-separator. Moreover, let z ∈ B dominate the bag, by the hypothesis z ∈ N (x) ∩ N (y) because x, y ∈ / B. As a result, adding x and y in each bag B between Bx and By achieves a star-decomposition of G that has a bag containing both x, y. • Case 2: Now, let us assume that Bx and By are adjacent in T . Note that, if Bx ⊆ N [z] for some z ∈ N (x) ∩ N (y) (resp., if By ⊆ N [z] for some z ∈ N (x) ∩ N (y)) the result holds. Indeed, adding y in Bx (resp., x in By ) achieves a star-decomposition of G that has a bag containing both x, y. 20 So, let us consider the case when none of the two bags Bx , By is dominated by a vertex of N (x) ∩ N (y). Then, Bx \ x and By \ y are xy-separators by the properties of a tree decomposition. Let zx ∈ Bx , zy ∈ By satisfy Bx ⊆ N [zx ] and By ⊆ N [zy ]. By the hypothesis, zx ∈ {x} ∪ (N (x) ∩ N (y)) and zy ∈ {y} ∪ (N (x) ∩ N (y)). Thus it follows that zx = x and zy = y (or else, we are back to Case 1). Note that Bx ∩ By = N (x) ∩ N (y) in such a case. We will mostly use the following two weaker versions of Lemma 17 in our proofs. Corollary 8. Let G be a connected graph with tb(G) = 1, let x, y ∈ V (G) be non-adjacent (and x 6= y). Suppose there exists a minimal separator S ⊆ (N (x) ∩ N (y)) ∪ {x, y} in G and {x, y} ⊆ S. Then, there is a star-decomposition (T, X ) of G with Bx , By ∈ X , x ∈ Bx , y ∈ By and either Bx = By or Bx , By are adjacent in T . Moreover, in the latter case, Bx ⊆ N [x], By ⊆ N [y]. Proof. We claim that for every xy-separator S 0 of G, if there is z ∈ / {x, y} such that S 0 ⊆ N [z] then z ∈ N (x) ∩ N (y). Observe that if the claim holds, then the corollary follows from Lemma 17. To prove the claim, let S ⊆ (N (x) ∩ N (y)) ∪ {x, y} be a separator of G and {x, y} ⊆ S, that exists by the hypothesis. Note that for any full component C of G\S, the xy-separator S 0 must contain some vertex in C. Since there are at least two full components of G\S, then z ∈ S \(x, y) ⊆ N (x)∩N (y), that finally proves the claim. So far, the two above results in this section (Lemma 17 and Corollary 8) apply to general graphs with tree-breadth one. However, we will need the fact that the graph is planar for the following corollary. Corollary 9. Let G be a connected graph with tb(G) = 1, let x, y ∈ V (G) be non-adjacent (and x 6= y). Suppose G is K3,3 -minor-free and |NG (x) ∩ NG (y)| ≥ 3. Then, there is a star-decomposition (T, X ) of G with Bx , By ∈ X , x ∈ Bx , y ∈ By and either Bx = By or Bx , By are adjacent in T . Moreover, in the latter case, Bx ⊆ N [x], By ⊆ N [y]. Proof. We claim that for every xy-separator S of G, if there is z ∈ / {x, y} such that S ⊆ N [z] then z ∈ N (x) ∩ N (y). Observe that if the claim holds, then the corollary follows from Lemma 17. To prove the claim, first recall that |N (x) ∩ N (y)| ≥ 3. Since vertex z dominates S and S is an xy-separator, therefore, z dominates N (x) ∩ N (y) because N (x) ∩ N (y) ⊆ S. In such case, z ∈ N (x) ∩ N (y), or else, G admits a K3,|N (x)∩N (y)| -minor with {x, y, z} and N (x) ∩ N (y) being the respective of the bipartition, which contradicts the hypothesis. Before we conclude this section, let us emphasize a useful consequence of Corollary 8 regarding minimal 2-separators. Lemma 18. Let G = (V, E) with tb(G) = 1, let x, y ∈ V be non-adjacent such that S = {x, y} is a minimal separator of G (x 6= y). For every full component C of G\S, we have that N (x)∩N (y)∩C 6= ∅. Proof. Let (T, X ) be a star-decomposition of G, that exists by Lemma 3, minimizing the distance in T between the subtrees Tx and Ty (respectively induced by the bags containing x and y). There are two cases. • First, suppose that Tx ∩ Ty 6= ∅. For any full component C of G \ S, let TC be the subtree that is induced by all bags intersecting C. Because C is a full component, there must be an edge between x and a vertex of C, and this edge is in a bag of Tx ∩ TC . Similarly, there must be an edge between y and a vertex of C, and this edge is in a bag of Ty ∩ TC . As a result, the subtrees Tx , Ty , TC are pairwise intersecting, and so by the Helly property (Lemma 1) Tx ∩ Ty ∩ TC 6= ∅ i.e., there exists a bag Xt which contains S and it intersects C. Let z ∈ Xt dominate the bag. Note that z ∈ C ∪ S because it has to dominate some vertices in C and so, it cannot be in V \ (C ∪ S). Furthermore, recall that x, y are non-adjacent by the hypothesis. Therefore, z ∈ C ∩ N (x) ∩ N (y), and the result holds for any full component C of G \ S. 21 • Else, since S = {x, y} is a minimal separator and we assume (T, X ) to minimize the distance in T between Tx and Ty , by Corollary 8 there are two adjacent bags Bx , By such that x ∈ Bx \By dominates Bx , y ∈ By \ Bx dominates By . Since Bx ∩ By is an xy-separator by the properties of a tree-decomposition, then Bx ∩ By ∩ C 6= ∅ for every full component C of G \ S, that is N (x) ∩ N (y) ∩ C 6= ∅. 5.2.3 Bounded Treewidth Independently from Algorithm Leaf-BottomUp, let us introduce in this section another property of (not necessarily prime) planar graphs with tree-breadth one. More precisely, we prove these graphs have bounded treewidth. To prove this property, we will use the same terminology as for the previous subsections. Lemma 19. Let G be planar with tb(G) ≤ 1. Then, tw(G) ≤ 4 and the upper-bound is sharp. b u x y a v Figure 15: A planar graph G with tb(G) = 1 and tw(G) = 4. Proof. The treewidth of G is the maximum treewidth of its atoms [12], so, let us assume G to be a prime planar graph. Let (T, X ) be any star-decomposition of G, the graph H = (V, {{u, v} | Tu ∩Tv 6= ∅}) is chordal. Furthermore, if H 0 is a chordal graph with same vertex-set V and such that E(G) ⊆ E(H 0 ) ⊆ E(H) then any clique-tree of H 0 is still a star-decomposition of G. Therefore, we will assume w.l.o.g. that H is a minimal triangulation of G and (T, X ) is a clique-tree of H (in particular, (T, X ) is reduced). Additional properties of (T, X ) will be deduced from the latter assumption about H by using the results from [16]. Let us now prove the lemma by induction on |V (G)| (the base-case of the graph with a single vertex is trivial). • If |X | = 1, then G has some universal vertex u. Furthermore, since G is planar therefore, G \ u is outerplanar [47]. Consequently, tw(G \ u) ≤ 2 [11], so, tw(G) ≤ 3. • Suppose |X | = 2. Let X = {B, B 0 }. Since (T, X ) is assumed to be a clique-tree of some minimal triangulation H of G, therefore, B ∩ B 0 is a minimal separator [16]. Let us remind that by Corollary 7 there is a planar supergraph G0 of G with same vertex-set so that B ∩ B 0 induces either an edge or a cycle of G0 . Furthermore (T, X ) is also a star-decomposition of G0 , so, tb(G0 ) ≤ 1. In addition, tw(G) ≤ tw(G0 ). Recall that we can further assume G0 to be prime (or else, we apply the induction hypothesis on the atoms of G0 ), hence B ∩ B 0 induces a cycle of G0 of length at least four. Let B∆B 0 = (B \ B 0 ) ∪ (B 0 \ B). Since H is assumed to be a minimal triangulation and B, B 0 are leaves of a clique-tree of H, therefore, B \ B 0 is a (nonempty) dominating clique of the subgraph G[B], and similarly B 0 \ B is a (nonempty) dominating clique of the subgraph G[B 0 ] [16]. Thus, every vertex u ∈ B∆B 0 satisfies B ∩ B 0 ∈ N (u). Since |B ∩ B 0 | ≥ 4 because B ∩ B 0 induces a cycle of G0 of length at least four, therefore, |B∆B 0 | ≤ 2 or else there would be a K3,3 -minor of G0 with any three vertices of B ∩ B 0 being one part of the bipartition. As a result, since B ∩ B 0 induces a cycle, tw(G) ≤ tw(G0 ) ≤ 2 + |B∆B 0 | = 4. • Finally, suppose |X | ≥ 3. Let t ∈ V (T ) be an internal node, by the properties of a tree decomposition the bag Xt is a separator of G. Let b ∈ Xt satisfy Xt ⊆ NG [b]. Since G is prime and so, biconnected, therefore Xt \ b is a separator of G \ b. In such case, let us remind by Lemma 13 that there exist a, c ∈ Xt \ b non-adjacent such that {a, c} is a minimal separator of G \ b. In particular, the path Π = (a, b, c) is a separator of G. Let C1 , C2 , . . . , Cl 22 be the components of G \ Π. For every 1 ≤ i ≤ l, let Gi be obtained from G[Ci ∪ Π] by making the two endpoints a, c of Π adjacent. Note that Gi can be obtained from G by edge-contractions (because G is prime and so, a, c ∈ N (Cj ) for every 1 ≤ j ≤ l), therefore, tb(Gi ) ≤ 1 because tree-breadth is stable under edge-contractions (Lemma 2). In addition, tw(G) ≤ maxi tw(Gi ) because Π induces a triangle in every graph Gi by construction. As a result, for every 1 ≤ i ≤ l, since |V (Gi )| < |V (G)| by construction, therefore tw(Gi ) ≤ 4 by the induction hypothesis, whence tw(G) ≤ 4. Let G be constructed from the cycle (u, v, x, y) of length four by adding two vertices a, b such that NG (a) = NG (b) = {u, x, v, y} (see Figure 15 for an illustration). Since there exists a stardecomposition of G with two bags (respectively dominated by a, b), tb(G) ≤ 1. Moreover, G is 4-regular by construction, therefore tw(G) ≥ 4 [12]. This proves the sharpness of the upperbound. Note that since it is well-known that many difficult problems can be solved on bounded-treewidth graphs in linear-time, therefore, it may be the case that the recognition of planar graphs with treebreadth at most one can be simplified by using Lemma 19. However, we were unable to find a way to use it in our proofs (actually, the star-decomposition that can be computed using our algorithm may have unbounded width — because of leaf-vertices of Type 1). 5.3 5.3.1 Correctness of Algorithm Leaf-BottomUp Existence of a P3 -separator As a first step to prove correctness of Algorithm Leaf-BottomUp, let us prove correctness of Step 1. That is, we will prove that for every planar graph G with tb(G) = 1, G contains a leaf-vertex or G admits a star-decomposition with at most two bags. To prove this step, we will prove additional properties of the minimal separators of prime planar graphs with tree-breadth one. In the following, let P3 (G) be the set of (not necessarily minimal) separators of G that induce paths of length 2 (we will call them P3 -separators since they have three vertices). We will distinguish the case when P3 (G) 6= ∅ from the case when P3 (G) = ∅. Theorem 7. Let G be a prime planar graph with tb(G) = 1. If P3 (G) 6= ∅, then G has a leaf-vertex. Proof. Let Π = (a, b, c) ∈ P3 (G) minimize the size of a smallest component of G \ Π. We recall that {a, c} ∈ / E(G) because G is assumed to be prime by the hypothesis (the latter fact will be used in the following). Let C be any component of G \ Π of minimum size. Our aim is to prove the existence of some leaf-vertex v ∈ C (the latter dominating the component C), that will prove Theorem 7. Claim 1. There do not exist Π0 ⊆ Π ∪ C, C 0 ⊂ C such that Π0 ∈ P3 (G) and C 0 is a component of G \ Π0 . Proof. The claim follows from the minimality of C.  We will often use Claim 1 in the remaining of the proof. Let (T, X ) be a star-decomposition of G, that exists by Lemma 3. In particular, let Ta , Tc be the subtrees that are respectively induced by the bags containing a or c. Assume w.l.o.g. that (T, X ) minimizes the distance in T between the subtrees Ta and Tc . We will distinguish the case Ta ∩ Tc 6= ∅ from the case Ta ∩ Tc = ∅. Case Ta ∩ Tc 6= ∅. In such case, the subtrees Ta , Tb , Tc are pairwise intersecting and so, by the Helly property (Lemma 1) Ta ∩ Tb ∩ Tc 6= ∅. Let us remove all vertices in V \ (Π ∪ C) from bags in (T, X ). Let us call (T, X C ) the resulting tree decomposition of G[Π ∪ C]. Claim 2. (T, X C ) has breadth one. Proof. There are two cases to be considered. • If b has some neighbour in C, then C must be a full component of G \ Π, or else one of {a, b}, {b, c} should be a clique-separator thus contradicting the fact that G is prime by the hypothesis. In such case, the claim follows from Lemma 9. 23 • Else, b has no neighbour in C, and let D be the connected component of b in G \ (a, c). Let H be obtained from G by contracting D to b. By Lemma 2, tb(H) = 1. Let (T, X H ) be the tree decomposition of breadth one of H where for every t ∈ V (T ), XtH = Xt if Xt ∩D = ∅, XtH = (Xt \D)∪{b} else. Moreover, since b has no neighbour in C, D∩NG [C] = ∅ and so, H[C ∪ Π] = G[C ∪ Π] by construction. Finally, since {b} is a full component of H \ (a, c), therefore, by Lemma 9 applied to H, the tree decomposition (T, X C ) is indeed a tree decomposition of breadth one of G[C ∪ Π].  Let (T 0 , X 0 ) be any reduced tree decomposition obtained from (T, X C ). We point out that 6= ∅ by construction (because Ta ∩ Tb ∩ Tc 6= ∅). Furthermore, since by Claim 2 (T, X C ) has breadth one, therefore (T 0 , X 0 ) is a star-decomposition of G[C ∪ Π] by Lemma 3. Ta0 ∩ Tb0 ∩ Tc0 We will prove that C contains a leaf-vertex by contradiction. Informally, we will show, using the properties of the star-decomposition (T 0 , X 0 ), that if it is not the case that C contains a leaf-vertex, then P3 (G[C ∪ Π]) ∩ P3 (G) 6= ∅ and the latter contradicts Claim 1. In order to prove this, first note that a has at least one neighbour in C because G is prime by the hypothesis (indeed, (b, c) cannot be an edge-separator of G). We now distinguish between several subcases. • Case 1. There is u ∈ C such that u ∈ N (a) ∩ N (b) ∩ N (c) (e.g., see Figure 16). By Lemma 15, either C is reduced to u or there exist Π0 = (a, u, c) ∈ P3 (G), C 0 ⊆ C \ u and C 0 is a component of G \ Π0 . The latter case contradicts Claim 1, therefore, C is reduced to u and so u is a leaf-vertex of Type 2. G\ C a c b u C Figure 16: Case 1 Thus, from now on let us assume that no such vertex u exists. • Case 2. By contradiction, assume N (a)∩C ⊆ N (b)∩C. By Lemma 14, |N (a)∩N (b)∩C| ≤ 1, so, |N (a) ∩ C| = 1. Let u ∈ N (a) ∩ N (b) ∩ C be the unique neighbour of vertex a in C (see Figure 17). Since in such case we can assume that u ∈ / N (c) (for otherwise, we are back to Case 1), and vertex c has some neighbour in C because G is prime (and so, (a, b) cannot be an edge-separator of G), therefore, C is not reduced to vertex u. Then, Π0 = (u, b, c) ∈ P3 (G) because it separates a from C \ u, and so there is at least one component of G \ Π0 that is strictly contained into C by construction. This contradicts Claim 1, so, Case 2 cannot occur. G \C a c b u C Figure 17: Case 2 • Case 3. There is u ∈ C satisfying u ∈ N (a)\N (b). By the properties of a tree decomposition, there is some bag B 0 ∈ Ta0 ∩ Tu0 . Let v ∈ B 0 dominate the bag B 0 . By construction, v 6= b because u ∈ B 0 \ N (b), similarly v 6= c because a ∈ B 0 \ N (c). We will also prove later that v 6= a. Moreover, Π \ N [v] 6= ∅ (or else, we are back to Case 1), hence Π \ B 0 6= ∅. So let B be the bag adjacent to B 0 onto the unique path in T 0 from B 0 to Ta0 ∩ Tb0 ∩ Tc0 (we remind that the latter subtree is nonempty by construction). By the properties of the tree decomposition (T 0 , X 0 ), B ∩ B 0 is a separator of G[C ∪ Π]. Furthermore, a ∈ B ∩ B 0 . More generally Π ∩ B 0 ⊆ B ∩ B 0 by construction, therefore B ∩ B 0 is also a separator of G by Lemma 16. Let w ∈ B dominate this bag. Observe that w 6= c because a ∈ B ∩ B 0 . 24 We will prove that v ∈ C and v is a leaf-vertex. In order to prove these two results, we will need to prove that C ∪ Π is fully contained into the two adjacent bags B, B 0 (Claim 7). The latter will require intermediate claims. Claim 3. c ∈ B ∩ B 0 . Proof. Assume for the sake of contradiction that c ∈ / B ∩ B 0 (see Figure 18). Then, c ∈ / B0 0 0 because Π ∩ B ⊆ B ∩ B by construction. We will prove that the latter contradicts Claim 1. Indeed, first observe that G \ w is connected because G is prime and so, biconnected, by the hypothesis. In addition (B ∩ B 0 ) \ w is a (not necessarily minimal) separator of G \ w because it separates B 0 \ B from c. Let S ⊆ (B ∩ B 0 ) \ w be a minimal separator of G \ w. By Lemma 13, there exist x, y ∈ (B ∩ B 0 ) \ w non-adjacent such that S = {x, y}, and so, Π0 = (x, w, y) ∈ P3 (G). Note that Π0 6= Π, because we assume that c ∈ / B ∩ B 0 and so 0 0 c∈ / {x, y}. Moreover, since (T , X ) is a star-decomposition of G[C ∪ Π] by construction we have that Π0 ⊆ Π ∪ C, therefore x ∈ C or y ∈ C, because c ∈ / {x, y} and a, b are adjacent whereas x, y are non-adjacent. W.l.o.g. let x ∈ C. c b u x a v w y ... B B' Figure 18: Case c ∈ / B ∩ B0 Subclaim 3.1. Π0 is not an ac-separator. Proof. We refer to Figure 19 for an illustration of the proof. Let C 0 be any component of G \ (Π ∪ C). Observe that C 0 is fully contained into some component D of G \ Π0 , because Π0 ⊆ Π ∪ C. In addition, a, c ∈ N (C 0 ) because G is prime by the hypothesis (and so, neither a nor b nor c nor (a, b) nor (b, c) can be a separator of G). In particular, since we assume c∈ / B ∩ B 0 and so, c ∈ / Π0 , therefore, c ∈ D. As a result, either a ∈ Π0 or a, c ∈ D, that finally proves the subclaim. ◦ B' u v a y Π' w b D Π c C' Figure 19: Π0 is not an ac-separator. Let D be the component of G \ Π0 such that c ∈ D, that exists because we assume c ∈ / B ∩ B0 0 0 and so, c ∈ / Π . Since b, c are adjacent and Π is not an ac-separator by Claim 3.1, therefore, Π ⊆ Π0 ∪ D. Moreover, let us show that Claim 3.1 implies the existence of some D0 ⊂ C being a component of G \ Π0 , thus contradicting Claim 1. Indeed, let D0 be any component of G \ (Π0 ∪ D). Since G is prime by the hypothesis, x has some neighbour in D0 and so, D0 ∩ C 6= ∅ because x ∈ C and Π ∩ D0 = ∅ by construction. But then, D0 ⊆ C \ x, for the existence of some z ∈ D0 \ C would imply that D0 ∩ Π 6= ∅. To sum up, we conclude that it must be the case that c ∈ B ∩ B 0 .  We will use Claim 3 to prove that v ∈ C, as follows: Claim 4. v ∈ C. Furthermore, the two vertices b, v are non-adjacent. Proof. Recall that v ∈ C ∪ Π because (T 0 , X 0 ) is a star-decomposition of G[C ∪ Π] by construction. So we will only need to prove that v ∈ / Π. First, since a ∈ B ∩B 0 by construction 0 0 and c ∈ B ∩ B by Claim 3, therefore, a, c ∈ B ⊆ N [v]. The latter implies that v ∈ / {a, c} because a, c ∈ N [v] whereas a, c are non-adjacent. Furthermore, this implies b ∈ / N [v] because 25 we assume that Π 6⊆ N [v] (for otherwise, we are back to Case 1). As a result, v ∈ / Π, whence v ∈ C.  Then, we will need the following technical claim in order to prove that w = b (Claim 6). Claim 5. G[C ∪ Π] is prime. Proof. Suppose by contradiction there exists a clique-separator S of G[C ∪ Π]. Then, S could not be a separator of G because G is prime by the hypothesis. By Lemma 16, the latter implies that S is an ac-separator of G[C ∪ Π]. Therefore, the two vertices b, v ∈ N (a) ∩ N (c) must be in S, and so, since b, v are non-adjacent by Claim 4, the latter contradicts the fact that S is a clique.  Claim 6. w = b. Proof. Assume for the sake of contradiction w 6= b (see Figure 20). We will prove that it contradicts Claim 1. Indeed, the graph G[C ∪ Π] \ w is connected because G[C ∪ Π] is prime by Claim 5 and so, biconnected. In addition, (B ∩ B 0 ) \ w is a (not necessarily minimal) separator of G[C ∪ Π] \ w because it separates b from B 0 \ B (recall that b, v are non-adjacent by Claim 3, and so, b ∈ / B 0 ⊆ N [v]). Let S ⊆ (B ∩ B 0 ) \ w be a minimal separator of G[C ∪ Π] \ w. By Lemma 13, there exist x, y ∈ (B ∩ B 0 ) \ w non-adjacent such that S = {x, y}, and so, Π0 = (x, w, y) ∈ P3 (G[C ∪ Π]). Furthermore, b ∈ / Π0 ⊆ (B ∩ B 0 ) ∪ {w} ⊆ N [v] ∪ {w} and so, 0 Π cannot be an ac-separator of G[C ∪ Π], whence by Lemma 16 Π0 is a separator of G, and so, Π0 ∈ P3 (G). u a=x v b w c=y ... B B' Figure 20: Case c ∈ B ∩ B 0 and w 6= b. Let D ⊆ C ∪ Π be the component of G[C ∪ Π] \ Π0 containing vertex b. Note that Π ⊆ D ∪ Π0 because Π0 is not an ac-separator of G[C ∪Π]. Let D0 be any component of G[C ∪Π]\(Π0 ∪D), that exists because Π0 ∈ P3 (G[C ∪ Π]). Since N (D0 ) ⊆ C ∪ Π by construction, D0 ∩ Π = ∅ by construction and D0 is a component of G[C ∪ Π] \ Π0 , therefore, D0 is also a component of G \ Π0 . The latter contradicts Claim 1 because D0 ⊂ C.  Let S = {v, b} ∪ (B ∩ B 0 ). We are now able to prove that S = C ∪ Π (Claim 7). That is, C ∪ Π is fully contained in the two adjacent bags B, B 0 (respectively dominated by b, v). Claim 7. S = C ∪ Π. Proof. Assume by contradiction S 6= C ∪ Π, let D be a component of G[C ∪ Π] \ S (see Figure 21). Note that D ⊂ C because Π ⊂ S by construction. Furthermore, v, b ∈ / B ∩ B0 0 because w = b by Claim 6 and b ∈ / N (v) by Claim 3, so, B ∩ B is a (minimal) vb-separator of G[C ∪Π]. The latter implies v ∈ / N (D) or b ∈ / N (D) because D induces a connected subgraph, D ∩ B ∩ B 0 = ∅ by construction, and B ∩ B 0 is a bv-separator of G[C ∪ Π]. As a result, there exists z ∈ {v, b} such that N (D) \ z ⊆ B ∩ B 0 . x v a b c y D B' B ... Figure 21: Case z = v. Moreover let {z, z 0 } = {v, b}. G[C ∪ Π] \ z is connected because G|C ∪ Π] is prime by Claim 5, and so, biconnected. In addition, N (D) \ z is a minimal separator of G[C ∪ Π] \ z because it separates D from z 0 and N (D) \ z ⊆ B ∩ B 0 ⊆ N (z 0 ) by construction. By Lemma 13, one 26 obtains the existence of two non-adjacent vertices x, y ∈ B ∩ B 0 such that N (D) \ z = {x, y}, whence N (D) ⊆ {x, y, z}. Then, by construction Π0 = (x, z, y) ∈ P3 (G) with D ⊂ C being a component of G \ Π0 , that contradicts Claim 1.  By Claim 7, C ∪ Π = S (see Figure 22). Note that it implies that C ⊆ N [v] because C \ v = (B ∩ B 0 ) \ (a, c). In order to conclude that v is a leaf-vertex, we will finally prove in Claim 8 that either B ∩ B 0 = {a, c} or B ∩ B 0 induces a path. a v G \C b c Figure 22: Case c ∈ B ∩ B 0 , w = b and C ∪ Π = {v, b} ∪ (B ∩ B 0 ). Claim 8. If B ∩ B 0 6= {a, c}, then G[B ∩ B 0 ] is a path. Proof. Recall that b, v ∈ / B ∩ B 0 because w = b by Claim 6 and b ∈ / N (v) by Claim 3. Hence by the properties of a tree decomposition, B ∩ B 0 is a bv-separator of G[C ∪ Π]. Since v ∈ C by Claim 4, a ∈ B ∩ B 0 by construction and c ∈ B ∩ B 0 by Claim 3, therefore B ∩ B 0 is also a vb-separator of G. In particular, B ∩ B 0 is a minimal bv-separator of G because B ∩ B 0 ⊆ N (v) ∩ N (w) = N (v) ∩ N (b) (indeed, recall that w = b by Claim 6). By Corollary 6, B ∩ B 0 either induces a cycle or it induces a forest of paths. Subclaim 8.1. B ∩ B 0 does not induce a cycle. Proof. By contradiction, let B ∩ B 0 induce a cycle. Recall that B ∩ B 0 contains the pair of non-adjacent vertices a, c (because a ∈ B ∩ B 0 by construction and c ∈ B ∩ B 0 by Claim 3). Therefore, one can contract B ∩ B 0 until one obtains an induced quadrangle (a, x, c, y). Let us contract an arbitrary component of G \ (Π ∪ C) so as to obtain a vertex z. Note that a, c ∈ N (z) because G is prime by the hypothesis (indeed, neither a nor b nor c nor (a, b) nor (a, c) can be a separator of G). Then, let us contract the edge {a, z} to a. By doing so, one obtains a K3,3 -minor with {a, b, v} being one part of the bipartition and {x, y, c} being the other part. This contradicts the fact that G is planar by the hypothesis, therefore B ∩ B 0 does not induce a cycle. ◦ It follows from Claim 8.1 that B ∩ B 0 induces a forest of paths. Suppose for the sake of contradiction that B ∩B 0 induces a forest of at least two paths. Let x ∈ / {a, c} be the endpoint of some path in the forest, that exists because we assume that B 6= {a, c}. Observe that |N (x)| ≥ 2 because b, v ∈ N (x), and |N (x)| = |N (x) ∩ (C ∪ Π)| ≤ 3 because x is the endpoint of some path of B ∩ B 0 and x ∈ C. Furthermore, N (x) \ (b, v) ⊆ B ∩ B 0 ⊆ N (b) ∩ N (v), and so, if |N (x)| = 3 then N (x) induces a path. Let Π0 = N (x) if |N (x)| = 3, else Π0 = (b, a, v). By construction, Π0 ⊆ Π ∪ C is a separator of G with {x} ⊂ C being a component of G \ Π0 , thus contradicting Claim 1. Consequently, B ∩ B 0 induces a path.  By Claim 8, either B ∩ B 0 = {a, c} or B ∩ B 0 induces a path. Furthermore, B ∩ B 0 = N (v) because v ∈ C (Claim 4) and C ∪ Π = {v, b} ∪ (B ∩ B 0 ) (Claim 7). In particular, if B ∩ B 0 = {a, c} then v is a leaf-vertex of Type 3. Else, B ∩ B 0 induces a path and the latter implies that |B ∩ B 0 | ≥ 4 or else the path B ∩ B 0 would be a separator of G with {v} being a component of G \ (B ∩ B 0 ), thus contradicting Claim 1. As a result, since we also have that B ∩ B 0 ⊆ N (b) and b, v are non-adjacent by Claim 4, therefore, v is a leaf-vertex of Type 1. Case Ta ∩ Tc = ∅. Since Π is a separator of G and G is prime by the hypothesis, one of Π or Π \ b must be a minimal separator of G. Therefore, since (T, X ) is assumed to minimize the distance in T between Ta and Tc , by Corollary 8 there exist two bags Ba , Bc that are adjacent in T and such that a ∈ Ba \ Bc and c ∈ Bc \ Ba . Furthermore, a dominates Ba while c dominates Bc . Note that Ba ∩ Bc = N (a) ∩ N (c), so, b ∈ Ba ∩ Bc . In particular, by the properties of a tree decomposition this implies that S = N (a) ∩ N (c) is a minimal ac-separator of G. We will prove that C is reduced to a vertex (Claim 10), the latter being a leaf-vertex. Claim 9. C ⊆ S. 27 Proof. Assume for the sake of contradiction that C 6⊆ B ∩ B 0 . By the properties of a tree decomposition it comes that some vertex y ∈ C is separated from a or c by the set S = B ∩ B 0 = N (a) ∩ N (c). Say w.l.o.g. that S is an yc-separator. Let C 0 ⊂ C be the connected component containing y in G \ (S ∪ {a}). Since we have that G \ a is connected because G is prime by the hypothesis (and so, biconnected), that c ∈ / C 0 and N (C 0 ) \ a ⊆ S ∩ (C ∪ Π) ⊆ N (c) ∩ (C ∪ Π), then 0 it comes that N (C ) \ a is a minimal yc-separator of G \ a. So, by Lemma 13 there exist x0 , y 0 ∈ S such that N (C 0 ) \ a = {x0 , y 0 }. Therefore, Π0 = (x0 , a, y 0 ) ∈ P3 (G) and C 0 ⊂ C is a component of G \ Π0 , that contradicts Claim 1.  By Claim 9, C ⊆ S (see Figure 23 for an illustration). Since S is an ac-separator and for any component C 0 of G \ (Π ∪ C), a, c ∈ N (C 0 ) because G is prime, therefore S ∩ C 0 6= ∅. One thus obtains the following chain of strict subset containment relations C ⊂ C ∪ {b} ⊂ S. Furthermore, by Corollary 6, S either induces a cycle or a forest of paths, so, C being a strict connected subset of S, it must induce a path. In particular, C ∪ {b} also being a strict subset of S, either it induces a path or it is the union of the path induced by C with the isolated vertex b. b a c ... C Bc Ba Figure 23: Case Ta ∩ Tc = ∅. Claim 10. |C| = 1. Proof. Assume for the sake of contradiction that |C| ≥ 2. Since C induces a path, let us pick an endpoint v ∈ C that is not adjacent to vertex b (recall that C ∪ {b} being a strict subset of S, it does not induce a cycle). In such a case, N (v) induces a path Π0 ∈ P3 (G), with a, c ∈ Π0 and {v} ⊂ C is a component of G \ Π0 , thus contradicting Claim 1.  By Claim 10, C is reduced to a vertex v, that is either a leaf-vertex of Type 2 (if v ∈ N (b)) or of Type 3 (if v ∈ / N (b)). Note that in some cases, there may only exist leaf-vertices of only one Type (i.e., see respectively Figure 24, 25 and 26 for Types 1,2 and 3). Therefore, there is none of the three Types of leaf-vertices that can be avoided in our algorithm. v1 v1 v2 d v4 v2 d v1 v4 v2 d v4 v3 v3 v3 Figure 24: A planar graph G with tb(G) = 1 and all of its four leaf-vertices v1 , v2 , v3 , v4 of Type 1. Figure 25: A planar graph G with tb(G) = 1 and all of its four leaf-vertices v1 , v2 , v3 , v4 of Type 2. Figure 26: A planar graph G with tb(G) = 1 and all of its four leaf-vertices v1 , v2 , v3 , v4 of Type 3. Examples of planar graphs G with tb(G) = 1 and P3 (G) = ∅ include C4 , the cycle with four vertices. To prove correctness of Step 1, it now suffices to prove that all these graphs (with P3 (G) = ∅) admit a star-decomposition with at most two bags. Lemma 20. For any prime planar graph G, if tb(G) = 1 and P3 (G) = ∅, then G admits a star-decomposition with at most 2 bags. 28 Proof. By contradiction, let (T, X ) be a star-decomposition of G with at least three bags. Let t ∈ V (T ) be an internal node, by the properties of a tree decomposition the bag Xt is a separator of G. Let u ∈ Xt satisfy Xt ⊆ NG [u]. Since G is biconnected, therefore Xt \ u is a separator of G \ u. By Lemma 13, there exist x, y ∈ Xt \ u non-adjacent such that {x, y} is a minimal separator of G \ u. In such case, (x, u, y) ∈ P3 (G), which contradicts the fact that P3 (G) = ∅. 5.3.2 Case of leaf-vertex v of Type 1 Lemma 21. Let G be a prime planar graph and v be a leaf-vertex of Type 1. Let Πv be the path induced by N (v) and let av , cv be the ends of Πv . Suppose V (G) 6= N [v] ∪ {dv }. Then Π0 = (av , dv , cv ) ∈ P3 (G) and N [v] \ {av , cv } is a component of G \ Π0 . Proof. Let C be a component of G \ (N [v] ∪ {dv }), that exists by the hypothesis. By construction, v ∈ / N [C], so, N (C) ⊆ N (v) ∪ {dv } separates v from C. Furthermore, since G is prime by the hypothesis, there exist x, y ∈ N (C) non-adjacent. Note that dv ∈ / {x, y} because N (v) ⊆ N (dv ) by the hypothesis, hence x, y ∈ N (v). We claim that {x, y} = {av , cv }. By contradiction, suppose x ∈ / {av , cv }. Let us write Πv = (P, x, Q, y, R) with P, Q non-empty subpaths of Πv and R a (possibly empty) subpath of Πv . In such a case, the connected subsets S1 := {v}∪P, S2 := {dv }, S3 := {x}, S4 := Q and S5 := {y}∪C induce a K5 -minor of G, that contradicts the hypothesis that G is planar. Therefore, the claim is proved, that is, {x, y} = {av , cv }. To prove the lemma, it now suffices to prove that N (C) ∩ N (v) = {av , cv } for in such a case the result will hold for any component C 0 of G \ (N [v] ∪ {dv }). By contradiction, let x0 ∈ (N (C) ∩ N (v)) \ (av , cv ). Since |N (v)| ≥ 4 because v is a leaf-vertex of Type 1 by the hypothesis, therefore, x0 and av are non-adjacent or x0 and cv are non-adjacent. Let y 0 ∈ {av , cv } be non-adjacent to x0 . Since x0 , y 0 ∈ N (C) ∩ N (v) are non-adjacent, therefore, by the same proof as for the above claim {x0 , y 0 } = {av , cv }, that would contradict the assumption that x0 ∈ / {av , cv }. As a result, N (C) ⊆ (av , dv , cv ) and so, since the result holds for any component C 0 of G \ (N [v] ∪ {dv }), Π0 = (av , dv , cv ) ∈ P3 (G) with N [v] \ {av , cv } being a full component of G \ Π0 . Theorem 8. Let G be a prime planar graph and v be a leaf-vertex of Type 1. Let Πv be the path induced by N (v) and let av , cv be the ends of Πv . Suppose V (G) 6= N [v] ∪ {dv }. Then, the graph G0 , obtained from G \ v by contracting the internal vertices of Πv to a single edge, is prime and planar, and tb(G) = 1 if and only if tb(G0 ) = 1. av dv av cv cv dv x y v Figure 27: Contraction of the internal vertices of Πv to a single edge and removal of v. Proof. For the remaining of the proof, let Π0v = (av , x, y, cv ) be the path resulting from the contraction of the internal vertices of Πv to the edge {x, y} in G0 . By Lemma 21 (av , dv , cv ) ∈ P3 (G) with (N [v] \ (av , cv )) being a full component of G \ (av , dv , cv ). Consequently, NG0 (x) = {av , dv , y} and NG0 (y) = {cv , dv , x}. The graph G0 is a minor of G, that is a planar graph by the hypothesis, so, G0 is also planar. In order to prove that G0 is prime, by contradiction, let S be a minimal clique-separator of G0 . There are two cases to be considered. • Suppose x ∈ S or y ∈ S. In such case, S ⊆ (av , x, dv ), or S ⊆ (x, dv , y), or S ⊆ (y, dv , cv ). By Lemma 21 (av , dv , cv ) ∈ P3 (G) with(N [v] \ (av , cv )) being a full component of G \ (av , dv , cv ), and so, for every component C of G0 \ (Π0v ∪ {dv }) = G \ (N [v] ∪ {dv }) av , cv ∈ N (C) because G is prime by the hypothesis. In such case, since av ∈ / S or cv ∈ / S, therefore, G0 \ S is connected, that contradicts the assumption that S is a clique-separator of G0 . 29 • Else, x, y ∈ / S. Since av ∈ / S or cv ∈ / S because av and cv are non-adjacent in G0 , therefore, S 0 must be a separator of G \ (x, y) or else G0 \ S would be connected because Π0v induces a path of G0 (thus contradicting the assumption that S is a separator of G0 ). In such a case, since by Lemma 21 (av , dv , cv ) ∈ P3 (G) with(N [v] \ (av , cv )) being a full component of G \ (av , dv , cv ), since S is a separator of G0 \ (x, y) = G \ (N [v] \ (av , cv )) and since S is not a separator of G because G is prime by the hypothesis, therefore, by Lemma 16 there are exactly two components Ca , Cc in G0 \(S ∪{x, y}) with av ∈ Ca and cv ∈ Cc . However, Π0v is an av cv -path of G0 \ S, thus contradicting the assumption that S is a separator of G0 . As a result, G0 is a prime planar graph. Finally, let us prove tb(G) = 1 if and only if tb(G0 ) = 1. • If tb(G) = 1 then tb(G \ v) = 1 because N (v) ⊆ N (dv ) by the hypothesis, and so, tb(G0 ) = 1 because G0 is obtained from G \ v by edge-contractions and tree-breadth is contraction-closed (Lemma 2). • Conversely, let us prove that tb(G) = 1 if tb(G0 ) = 1. To prove this, let (T 0 , X 0 ) be a reduced star-decomposition of G0 , that exists by Lemma 3, minimizing the distance in T 0 between the two subtrees Ta0 v and Tc0v . In order to prove tb(G) = 1, it suffices to show how to construct a star-decomposition of G from (T 0 , X 0 ). We will prove as an intermediate claim that Ta0 v ∩ Tc0v 6= ∅. By contradiction, suppose Ta0 v ∩ Tc0v = ∅. Since by Lemma 21 (av , dv , cv ) ∈ P3 (G) with (N [v] \ (av , cv )) being a full component of G \ (av , dv , cv ), therefore, (av , dv , cv ) ∈ P3 (G0 ) with {x, y} being a full component of G0 \ (av , dv , cv ). Since we proved that G0 is prime, it follows that one of (av , cv ) or (av , dv , cv ) is a minimal separator of G0 . In such a situation, since (T 0 , X 0 ) is assumed to minimize the distance in T 0 between Ta0 v and Tc0v , therefore, by Corollary 8 there are two adjacent bags Ba0 v , Bc0 v such that av ∈ Ba0 v \ Bc0 v and cv ∈ Bc0 v \ Ba0 v respectively dominate Ba0 v and Bc0 v in G0 . However by the properties of a tree decomposition this implies that Ba0 v ∩ Bc0 v = N (av ) ∩ N (cv ) is an av cv -separator of G0 , thus contradicting the existence of the av cv -path Π0v . Therefore, the claim is proved and Ta0 v ∩ Tc0v 6= ∅. Recall that Ta0 v ∩Td0 v 6= ∅ and similarly Tc0v ∩Td0 v 6= ∅ by the properties of a tree decomposition. Hence, the subtrees Ta0 v , Tc0v , Td0 v are pairwise intersecting, and so, by the Helly property (Lemma 1), Ta0 v ∩ Td0 v ∩ Tc0v 6= ∅. Let us now proceed as follows so as to obtain a stardecomposition of G. Let us remove x, y from all bags in X 0 , that keeps the property for (T 0 , X 0 ) to be a star-decomposition because x and y are dominated by dv in G0 . Then, let us add two new bags B1 = N [v], B2 = N (v) ∪ {dv }, and finally let us make B1 , B2 pairwise adjacent and let us make B2 adjacent to some bag of Ta0 v ∩ Td0 v ∩ Tc0v . By construction, the resulting tree decomposition is indeed a star-decomposition of G, whence tb(G) = 1. 5.3.3 Proof of Step 3.1 (a) In the following three subsections ( 5.3.3, 5.3.4 and 5.3.5) we will prove correctness of the algorithm for the case of a leaf-vertex v of Type 2 or 3 and G \ v is prime ( Step 3.1). Our proofs in these subsections will mostly rely on Lemma 17. Let us first show how we can use Lemma 17 in order to prove correctness of Step 3.1 (a). Note that since we are in the case when G \ v is prime, we needn’t prove it in the following Theorem 9. Theorem 9. Let G = (V, E) be a prime planar graph, let v be a leaf-vertex of Type 2 or 3, and let Πv = (av , bv , cv ) be as in Definition 5. Suppose that |N (av ) ∩ N (cv )| ≥ 3 in G \ v, or there exists a minimal separator S ⊆ (N (av ) ∩ N (cv )) ∪ {av , cv } in G \ v and {av , cv } ⊆ S. Then, tb(G) = 1 if and only if tb(G \ v) = 1. Proof. First we prove that tb(G) = 1 implies that tb(G \ v) = 1, that is the easy part of the result. Let (T, X ) be a tree decomposition of G of breadth one, let (T, X 0 ) be such that for every node t ∈ V (T ), Xt0 = Xt \ v. Observe that (T, X 0 ) is a tree decomposition of G \ v. Furthermore, we claim that it has breadth one, indeed, for every t ∈ V (T ) such that Xt ⊆ NG [v], Xt0 ⊆ NG [bv ] because NG (v) ⊆ NG [bv ]. As a result, tb(G \ v) = 1. 30 Conversely, we prove that tb(G \ v) = 1 implies that tb(G) = 1. Let (T 0 , X 0 ) be a stardecomposition of G \ v minimizing the distance in T 0 between the subtrees Ta0 v and Tc0v . There are two cases. If Ta0 v ∩ Tc0v 6= ∅, then the subtrees Ta0 v , Tb0v , Tc0v are pairwise intersecting, hence by the Helly property (Lemma 1) Ta0 v ∩ Tb0v ∩ Tc0v 6= ∅, and so it suffices to make adjacent to any bag of Ta0 v ∩ Tb0v ∩ Tc0v the new bag NG [v] ⊆ {av , bv , cv , v} so as to obtain a star-decomposition of G. Else Ta0 v ∩ Tc0v = ∅ and so, by Corollary 9 if |N (av ) ∩ N (cv )| ≥ 3 in G \ v or by Corollary 8 else, there are two adjacent bags Ba0 v , Bc0 v such that av ∈ Ba0 v \ Bc0 v , bv ∈ Ba0 v ∩ Bc0 v ⊆ N (av ) ∩ N (cv ) and cv ∈ Bc0 v \ Ba0 v . Furthermore, av dominates Ba0 v while cv dominates Bc0 v . One obtains a star-decomposition of G simply by adding vertex v into bags Ba0 v and Bc0 v . 5.3.4 Proof of Step 3.1 (b) i The proof of this step is more involved than the proof of previous Step 3.1 (a). We will need the following intermediate lemma. Lemma 22. Let G = (V, E) be a prime graph with tb(G) = 1, let v be a leaf-vertex of Type 2 or 3 and let Πv = (av , bv , cv ) be as in Definition 5. Suppose that N (av ) ∩ N (cv ) = {v, bv } and V 6= Πv ∪ {v}. Then, NG [bv ] \ (av , cv , v) is an av cv -separator of G \ v. Proof. Let (T, X ) be a star-decomposition of G, that exists by Lemma 3, minimizing the distance in T between the subtrees Tav and Tcv . We claim that Tav ∩ Tcv 6= ∅, i.e., av , cv are in a same bag of the decomposition. By contradiction, let Tav ∩Tcv = ∅. Since G is prime and Πv is a separator of G, therefore, one of Πv or Πv \ bv is a minimal separator of G. Since (T, X ) minimizes the distance in T between Tav and Tcv , therefore, by Corollary 8 there exist two adjacent bags Bav , Bcv such that av ∈ Bav \ Bcv and cv ∈ Bcv \ Bav . Furthermore, vertices av and cv respectively dominate the bags Bav and Bcv . This implies Bav ∩ Bcv = NG (av ) ∩ NG (cv ) and so, NG (av ) ∩ NG (cv ) is a minimal av cv -separator of G by the properties of the tree decomposition. However, let C be any component of G \ (Πv ∪ {v}), that exists because V 6= Πv ∪ {v} by the hypothesis. Since G is prime, therefore, av , cv ∈ N (C) (or else, one of the cliques av or bv or cv or (av , bv ) or (bv , cv ) would be a clique-separator of G, thus contradicting the assumption that G is prime). Then, the av cv -separator NG (av ) ∩ NG (cv ) must contain some vertex of C, which contradicts the fact that NG (av ) ∩ NG (cv ) = {v, bv } by the hypothesis. As a result, we proved that Tav ∩ Tcv 6= ∅. Let H be the chordal supergraph of G such that (T, X ) is a clique-tree of H. Equivalently, every two vertices x, y ∈ V are adjacent in H if and only if they are in a same bag of X . In particular, av , cv are adjacent in H. Let S := NH (av ) ∩ NH (cv ). We claim that S is an av cv -separator of G. By contradiction, if it is not an av cv -separator of G, then there exists an av cv -path Pav cv of G which does not intersect S. Furthermore, Pav cv is a path of H because H is a supergraph of G, and it has length at least two because av , cv are non-adjacent in G. So, let Qav cv be taken of minimum length amongst all av cv -paths of length at least two in H that do not intersect S (the existence of such a path follows from the existence of Pav cv ). Observe that Qav cv may be not a path in G. By minimality of Qav cv , the vertices of Qav cv induce a cycle of H because av , cv are adjacent in H. Therefore, the vertices of Qav cv induce a triangle because H is chordal. However, this contradicts the fact that Qav cv does not intersect S = NH (av ) ∩ NH (cv ), so, the claim is proved. Finally, let us prove that S \ v ⊆ NG [bv ] \ (av , cv , v), that will conclude the proof that NG [bv ] \ (av , cv , v) is an av cv -separator of G \ v. For every vertex x ∈ S \ v, x ∈ NH (av ) ∩ NH (cv ), therefore, Tav ∩ Tx 6= ∅ and Tcv ∩ Tx 6= ∅ by construction of H. Since the subtrees Tav , Tcv , Tx are pairwise intersecting, by the Helly property (Lemma 1) Tav ∩ Tcv ∩ Tx 6= ∅, or equivalently there is some bag B ∈ Tav ∩ Tcv ∩ Tx . Let z ∈ B dominate the bag. Clearly, x ∈ NG [z]. Furthermore, z ∈ NG (av ) ∩ NG (cv ) because av , cv are non-adjacent in G. As a result, either z = bv or z = v. Since x 6= v and NG (v) ⊆ NG [bv ], we have that x ∈ NG [bv ] in both cases. Theorem 10. Let G = (V, E) be a prime planar graph with tb(G) = 1, v be a leaf-vertex of Type 2 or 3, and let Πv = (av , bv , cv ) be as in Definition 5. Suppose N (av ) ∩ N (cv ) = {bv , v}, and G \ v is prime, and there is no minimal separator S ⊆ (N (av ) ∩ N (cv )) ∪ {av , cv } in G \ v such that {av , cv } ⊆ S. Then, G = C4 , a cycle with four vertices. Proof. By contradiction, assume G 6= C4 . Since G is prime by the hypothesis, G has at least five vertices (the single other graph with four vertices and a leaf-vertex of Type 2 or 3 is the diamond, which is not prime). Equivalently, V 6= Πv ∪ {v}. By Lemma 22, this implies that N [bv ] \ (av , cv , v) is an av cv -separator of G \ v. Since G \ v is prime by the hypothesis, and so, biconnected, therefore, 31 G \ (bv , v) is connected, and so, N (bv ) \ (av , cv , v) 6= ∅ is an av cv -separator of G \ (bv , v). In particular, av , bv , cv ∈ N (V \ (Πv ∪ {v})). Moreover, we claim that V \ (Πv ∪ {v}) induces a connected subgraph (note that the latter implies that V \ (Πv ∪ {v}) is a full component of G \ Πv ). By contradiction, let C1 , C2 be distinct components of V \ (Πv ∪ {v}). Since G is prime, av , cv ∈ NG (C1 ) ∩ NG (C2 ) (or else, one of the cliques av or bv or cv or (av , bv ) or (bv , cv ) would be a clique-separator of G, thus contradicting the assumption that G is prime). Therefore, bv ∈ NG (C1 ) ∩ NG (C2 ) because N [bv ] \ (av , cv , v) is an av cv -separator of G \ v. It follows that Πv is a minimal separator of G \ v, that contradicts the hypothesis that there is no minimal separator S ⊆ (N (av ) ∩ N (cv )) ∪ {av , cv } in G \ v and {av , cv } ⊆ S. Consequently, V \ (Πv ∪ {v}) induces a connected subgraph. Let S 0 ⊆ N (bv ) \ (av , cv , v) be a minimal av cv -separator of G \ (bv , v). By Lemma 13, there exist x, y ∈ N (bv ) \ (av , cv , v) non-adjacent such that S 0 = {x, y}. Finally, let Π0 = (x, bv , y) and let A, C be the respective components of av , cv in G \ (Π0 ∪ {v}). Note that x, y ∈ N (A) ∩ N (C) because G \ v is prime by the hypothesis (indeed, neither x nor bv nor y nor (bv , x) nor (bv , y) can be a separator of G \ v). Let P be an xy-path of V \ (Πv ∪ {v}), that exists because V \ (Πv ∪ {v}) is connected. Also, let A0 ⊆ A and C 0 ⊆ C be the respective components of av , cv in G\(P ∪Π0 ∪{v}). Note that the subpath P \ (x, y) lies onto a unique component of G \ (Π0 ∪ {v}) because it does not intersect Πv ∪ {v} by construction, so, A0 = A or C 0 = C. By symmetry, assume that C 0 = C. There are two cases to consider. • Assume A0 = A (see Figure 28 for an illustration). Let us contract the internal vertices of P so as to make vertices x, y adjacent. Then, let us contract the components A, C to the two vertices av , cv , respectively. Finally, let us contract v to either av or cv . By construction, the five vertices av , bv , cv , x, y now induce a K5 , that contradicts the fact that G is planar by the hypothesis. P \ (x,y) P x y C av x A A y C C A av cv cv bv bv v v Figure 28: Case A0 = A (left). A K5 -minor is drawn (right), with edges resulting from contractions labeled in red. • Else, A0 6= A. Equivalently, P ⊆ A ∪ {x, y} (see Figure 29 for an illustration). Since A is connected, N (A0 ) ∩ (P \ (x, y)) 6= ∅. Let z ∈ N (A0 ) ∩ P . Let us contract the internal vertices of P to vertex z. Then, let us contract the components A0 and C 0 = C to the two vertices av , cv , respectively. Finally let us contract v to either av or cv . By construction, there is a K3,3 -minor whose sides of the bipartition are {av , x, y} and {bv , cv , z}, respectively, that contradicts the fact that G is planar by the hypothesis. P A\A' x P[x,z] y z A' z C av A' cv y x C P[z,y] C av cv bv bv v v Figure 29: Case A0 6= A (left). A K3,3 -minor is drawn (right), with each side of the bipartition being coloured differently. Edges resulting from contractions are labeled in red. Since both cases contradict the hypothesis that G is planar, therefore, G = C4 . 32 5.3.5 Proof of Step 3.1 (b) ii Theorem 11. Let G = (V, E) be a prime planar graph, let v be a leaf-vertex of Type 2 or 3, and let Πv = (av , bv , cv ) be as in Definition 5. Suppose that all of the following statements hold: • N (av ) ∩ N (cv ) = {v, bv , uv } with uv ∈ / {v} ∪ Πv ; • V 6= {av , bv , cv , uv , v}; • there is no minimal separator S ⊆ (N (av ) ∩ N (cv )) ∪ {av , cv } in G \ v and {av , cv } ⊆ S. Let G0 be the graph obtained from G by adding edges {v, uv } and {bv , v}, then tb(G) = 1 if and only if tb(G0 ) = 1. Moreover, G0 is planar and prime. uv av cv bv v Figure 30: Case of a leaf-vertex v of Type 3 with G \ v is prime and N (av ) ∩ N (cv ) = {uv , bv , v}. Red edges are those added by Algorithm Leaf-BottomUp. Proof. We will first prove that G \ v is prime. By contradiction, let S 0 be a minimal cliqueseparator of G \ v. By Theorem 6, there is wv 6= v such that S 0 = {bv , wv }, and by Lemma 16, S 0 must be an av cv -separator of G \ v. Then, it follows that wv = uv ∈ N (av ) ∩ N (cv ), whence V = {av , bv , cv , uv , v} by Theorem 6, that contradicts the hypothesis. Therefore, G \ v is prime. Let us prove that tb(G) = 1 implies that tb(G0 ) = 1. Let (T, X ) be a star-decomposition of G, which exists by Lemma 3, minimizing the distance in T between the subtrees Tav and Tcv . Since NG (v) ⊆ NG [bv ] then removing v from all bags leaves a tree decomposition of G \ v of breadth one. Up to reducing the tree decomposition, let (T 0 , X 0 ) be any reduced tree decomposition of G \ v that is obtained from (T, X ) by first removing v from the bags. Note that (T 0 , X 0 ) is a star-decomposition of G \ v by Lemma 3. Now, there are two cases. • Suppose Ta0 v ∩Tc0v 6= ∅. We will need to prove in this case that the two subtrees Ta0 v ∩Tb0v ∩Tc0v and Ta0 v ∩ Tu0 v ∩ Tc0v are nonempty and disjoint. Claim 11. bv , uv are non-adjacent in G. Proof. By contradiction, if it were the case that bv , uv are adjacent, then by Lemma 15, either uv is an isolated vertex of G \ (Πv ∪ {v}) — in which case, Πv ∈ P3 (G \ v) because we assume V 6= {av , bv , cv , uv , v} by the hypothesis —, or (av , uv , bv ) ∈ P3 (G \ v). Since G \ v is prime, it follows that one of Πv or Πv \ bv must be a minimal separator of G \ v, similarly one of (av , uv , cv ) or (av , cv ) must be a minimal separator of G \ v. Therefore, both cases contradict the hypothesis that there is no minimal separator S ⊆ (N (av ) ∩ N (cv )) ∪ {av , cv } in G \ v and {av , cv } ⊆ S, which proves that bv , uv are non-adjacent.  Recall that we are in the case when Ta0 v ∩ Tc0v 6= ∅. The subtrees Ta0 v , Tc0v , Tu0 v are pairwise intersecting, similarly the subtrees Ta0 v , Tc0v , Tb0v are pairwise intersecting. Therefore, by the Helly property (Lemma 1) Ta0 v ∩ Tb0v ∩ Tc0v 6= ∅ and Ta0 v ∩ Tu0 v ∩ Tc0v 6= ∅. Furthermore, Tav ∩ Tbv ∩ Tcv ∩ Tuv = ∅, because since bv , uv are non-adjacent by Claim 11 no vertex dominates all of {av , bv , cv , uv } in G, and so, Ta0 v ∩ Tb0v ∩ Tc0v ∩ Tu0 v = ∅. Claim 12. The subtrees Ta0 v ∩ Tb0v ∩ Tc0v and Ta0 v ∩ Tu0 v ∩ Tc0v are adjacent in T 0 . Proof. By contradiction, let B be an internal bag onto the path between both subtrees in T 0 , let z ∈ B dominate the bag. Note that av , cv ∈ B by the properties of the tree decomposition, z∈ / {av , cv } because av , cv are non-adjacent, and so, z ∈ N (av ) ∩ N (cv ) \ v = {uv , bv }. This 33 contradicts the fact that B ∈ / Ta0 v ∩ Tb0v ∩ Tc0v and B ∈ / Ta0 v ∩ Tu0 v ∩ Tc0v , therefore, the subtrees 0 0 0 0 0 0 Tav ∩ Tbv ∩ Tcv and Tav ∩ Tuv ∩ Tcv are adjacent in T 0 .  Finally, let B ∈ Ta0 v ∩ Tb0v ∩ Tc0v , B 0 ∈ Ta0 v ∩ Tu0 v ∩ Tc0v be adjacent, that exist by Claim 12. Observe that bv dominates B, uv dominates B 0 . To obtain a star-decomposition of G0 from (T 0 , X 0 ), it now suffices to add vertex v in B and B 0 , whence tb(G0 ) = 1. • Else, Ta0 v ∩Tc0v = ∅. This implies Tav ∩Tcv = ∅. Since the tree decomposition (T, X ) minimizes the distance in T between Tav and Tcv , G is planar and |N (av ) ∩ N (cv )| ≥ 3, therefore by Corollary 9, the subtrees Tav and Tcv are adjacent in T , whence the subtrees Ta0 v , Tc0v are also adjacent in T 0 . In particular, by Corollary 9 there exist two adjacent bags Ba0 v , Bc0 v ∈ X 0 such that av ∈ Ba0 v \ Bc0 v , Ba0 v ∩ Bc0 v = NG (av ) ∩ NG (cv ) \ v = {uv , bv }, cv ∈ Bc0 v \ Ba0 v . Furthermore, av dominates Ba0 v while cv dominates Bc0 v . Therefore, in order to obtain a star-decomposition of G0 from (T 0 , X 0 ), it now suffices to add vertex v in Ba0 v and Bc0 v – that yields exactly (T, X ) –, whence tb(G0 ) = 1. Before we can prove the equivalence, i.e., tb(G) = 1 if and only if tb(G0 ) = 1, we need to prove first that G0 is prime and planar. Claim 13. G0 is prime. Proof. Let S 0 be a clique-separator of G0 . Note that v ∈ S 0 by construction of G0 . Therefore, S 0 \ v is a clique-separator of G \ v, that contradicts the fact that G \ v is prime. Consequently, G0 is prime.  Claim 14. G0 is planar. Proof. Let us fix a plane embedding of G. By Jordan Theorem, the cycle induced by (av , bv , cv , uv ) separates the plane into two regions. Let G1 , G2 be respectively the subgraphs of G that are induced by all the vertices in each region. uv av c1 cv c2 bv Figure 31: Proof that the graph G0 of Theorem 11 is planar. We claim that either V \ (av , bv , cv , uv , v) ⊆ V (G1 ), or V \ (av , bv , cv , uv , v) ⊆ V (G2 ). Note that it will prove that G0 is planar, because then drawing vertex v onto the region that does not contain the set V \ (av , bv , cv , uv , v) yields a planar embedding of G0 . By contradiction, let C1 ⊆ V (G1 ), C2 ⊆ V (G2 ) be connected components of V \ (av , bv , cv , uv , v). Let Π0v = (av , uv , cv ). If one of Πv or Π0v belongs to P3 (G\v), then, there exists a minimal separator S ⊆ (N (av )∩N (cv ))∪ {av , cv } in G \ v and since G \ v is prime, {av , cv } ⊆ S. This would contradict the hypothesis, so, Πv , Π0v ∈ / P3 (G \ v). As a result, since (av , bv , cv , uv ) = Πv ∪ Π0v separates C1 from C2 , therefore, uv , bv ∈ N (C1 ) ∩ N (C2 ) (or else, Πv ∈ P3 (G \ v) or Π0v ∈ P3 (G \ v)). Let us remove all other components of V \ (av , bv , cv , uv , v) but C1 and C2 , and let us remove all edges between {av , cv } and C1 ∪ C2 if any (see Figure 31). Finally, let us contract C1 , C2 to the two vertices x1 , x2 . The cycle induced by (uv , x1 , bv , x2 ) separates the plane into two regions with av , cv being into different regions by construction. Vertex v must belong to one of the regions, but then it is a contradiction because v ∈ N (av ) ∩ N (cv ) by the hypothesis.  To conclude the proof, let us prove that conversely, tb(G0 ) = 1 implies that tb(G) = 1. Let (T 0 , X 0 ) be a star-decomposition of G0 minimizing the distance in T 0 between the subtrees Ta0 v and Tc0v . As an intermediate step, we claim that if removing vertex v from all bags of X 0 leaves a tree decomposition of G \ v of breadth one, then it implies that tb(G) = 1. To prove the claim, there are two cases to be considered. • If Ta0 v ∩ Tc0v 6= ∅, then the subtrees Ta0 v , Tb0v , Tc0v are pairwise intersecting, hence by the Helly property (Lemma 1) Ta0 v ∩ Tb0v ∩ Tc0v 6= ∅. Equivalently there is bag containing Πv , and so it suffices to remove v from all bags and then to make any bag containing Πv adjacent to the new bag NG [v] in order to obtain a tree decomposition of G of breadth one. 34 • Else, Ta0 v ∩ Tc0v = ∅. Since (T 0 , X 0 ) minimizes the distance in T 0 between the subtrees Ta0 v and Tc0v , G0 is planar by Claim 14 and av , cv have three common neighbours in G0 , therefore, by Corollary 9 there must exist two adjacents bags Ba0 v , Bc0 v such that av ∈ Ba0 v \Bc0 v , Ba0 v ∩Bc0 v = N (av ) ∩ N (cv ) and cv ∈ Bc0 v \ Ba0 v . Furthermore, vertex av dominates the bag Ba0 v , while vertex cv dominates the bag Bc0 v . As a result, removing vertex v from all bags but Ba0 v , Bc0 v leads to a tree decomposition of G of breadth one. Consequently, we are left to modify the tree decomposition (T 0 , X 0 ) so as to ensure that none of the bags is only dominated by vertex v in G0 , for if it is the case then removing v from all bags does leave a tree decomposition of G \ v of breadth one. We will call the latter property the removal property. Observe that if it is the case that (T 0 , X 0 ) does not satisfy the removal property, then there must be a bag B fully containing NG0 (v) because any strict subset of NG0 (v) is dominated by some vertex of G \ v. In particular, B = NG0 [v] because only vertex v dominates NG0 (v) in G0 , and so we can further assume that Tv0 = {B} without violating the property for (T 0 , X 0 ) to be a tree decomposition of G0 of breadth one. Therefore in the following, assume that (T 0 , X 0 ) is a reduced star-decomposition of G0 and Tv0 = {B}, that is always possible to achieve by Lemma 3 and above remarks. Since V 6= {av , bv , cv , uv , v} = NG0 [v] by the hypothesis, therefore, X 0 \ B 6= ∅. Let B 0 be adjacent to B in T 0 . Note that B ∩ B 0 6= {av , bv , cv , uv } because no other vertex than v dominates the subset {av , bv , cv , uv } in G0 . By the properties of a tree decomposition, B ∩ B 0 is a separator of G0 . Consequently, B ∩ B 0 is not a clique because G0 is prime by Claim 13. Furthermore, since B ∩ B 0 6= {av , bv , cv , uv } it holds that B \ (B 0 ∪ {v}) 6= ∅, consequently B ∩ B 0 is also a separator of G \ v. Since G \ v is prime, B ∩ B 0 cannot be any of (av , cv ), Πv or Π0v because by the hypothesis there is no minimal separator S ⊆ (N (av ) ∩ N (cv )) ∪ {av , cv } in G \ v and {av , cv } ⊆ S. It follows that B ∩ B 0 ⊆ {av , bv , uv } or B ∩ B 0 ⊆ {bv , cv , uv }. Let us substitute the bag B with the two adjacent bags B1 = {av , uv , bv , v}, B2 = {bv , cv , uv , v}, then we make adjacent all bags B 00 that were formerly adjacent to B to some bag amongst B1 , B2 containing B ∩B 00 . Note that B1 ⊆ N [av ] and that B2 ⊆ N [cv ]. Therefore, the resulting tree decomposition is a tree decomposition of G of breadth one such that v dominates no bag. 5.3.6 Case of leaf-vertex v of Type 2 or 3 and G \ v not prime The remaining subsections will be devoted to the proof of correctness of Step 3.2. In particular, this subsection is devoted to the proof that when G \ v is not prime one can only consider the case when the leaf-vertex v is of Type 2, i.e., v and bv are adjacent in G. Note that when v is of Type 3, then in general one cannot add an edge between v and bv without violating the property for the graph G to be planar, as shown in Figure 32. We will now prove that whenever we are in the conditions of Step 3.2, it is always possible to do so while preserving the planarity of the graph G and the property to be of tree-breadth one. bv av bv cv av v cv v Figure 32: A planar graph G with tb(G) = 1 (left), and a leaf-vertex v of Type 3 so that adding an edge between v and bv violates the property for the graph to be planar (right). In the latter case, one side of the bipartition of the K3,3 -minor is coloured red. Theorem 12. Let G be a prime planar graph. Let v be a leaf-vertex of Type 3 such that G \ v is not prime. Finally, let Πv = (av , bv , cv ) be as in Definition 5. Let G0 be obtained from G by adding the edge {v, bv }. Then, G0 is prime and planar, and tb(G) = 1 if and only if tb(G0 ) = 1. Proof. First, we prove that G0 is prime and planar. 35 • In order to prove that G0 is prime, by contradiction let S be a clique-separator of G0 . Since G0 is a supergraph of G, therefore S is a separator of G but it does not induce a clique in G. Hence, S contains the edge {v, bv }, and so either S ⊆ {av , bv , v} or S ⊆ {bv , cv , v}. Let C = V \ (Πv ∪ {v}), by Theorem 6, C is a full component of G \ Πv because G \ v is not prime. In particular, C is connected and av , cv ∈ N (C), that contradicts the fact that G0 \ S is unconnected. • Then in order to prove that G0 is planar, let us fix a plane embedding of G. The cycle induced by (av , bv , cv , v) separates the plane into two regions. To prove that G0 is planar, we claim that it suffices to prove that all vertices in C = V \ (Πv ∪ {v}) are in the same region, for then drawing the edge {bv , v} in the other region leads to a plane embedding of G0 . By contradiction, let x, y ∈ C be in different regions. By [15, Proposition 8], the cycle (av , bv , cv , v) is an xy-separator of G, that contradicts the fact that C is connected. Let us now prove that tb(G) = 1 implies that tb(G0 ) = 1. Let (T, X ) be a star-decomposition of G, that exists by Lemma 3, minimizing the distance in T between the subtrees Tav and Tcv . Let us remove vertex v from all bags, that leads to a tree decomposition (T, X−v ) of G \ v of breadth one because NG (v) ⊆ NG (bv ). Then, let (T 0 , X 0 ) be any reduced tree decomposition that is obtained from (T, X−v ), that is a star-decomposition of G \ v by Lemma 3. Now, there are two cases. If Ta0 v ∩Tc0v 6= ∅, then the subtrees Ta0 v , Tb0v , Tc0v are pairwise intersecting and so, by the Helly property (Lemma 1) Ta0 v ∩ Tb0v ∩ Tc0v 6= ∅. Hence one obtains a star-decomposition of G0 simply by making some bag of Ta0 v ∩ Tb0v ∩ Tc0v adjacent to the new bag NG0 [v] = {av , bv , cv , v}. Else, Ta0 v ∩ Tc0v = ∅, so, Tav ∩ Tcv = ∅. Since Πv ∈ P3 (G) and G is prime by the hypothesis, therefore, one of Πv or Πv \ bv must be a minimal separator of G. As a result, since (T, X ) is assumed to minimize the distance in T between the subtrees Tav and Tcv , by Corollary 8 there exist two adjacent bags Bav , Bcv ∈ X so that av ∈ Bav \ Bcv and cv ∈ Bcv \ Bav respectively dominate the bags Bav and Bcv . In such case, Bav ∩ Bcv = NG (av ) ∩ NG (cv ) and so, since bv , v ∈ Bav ∩ Bcv , (T, X ) is also a star-decomposition of G0 . So, in conclusion, tb(G0 ) = 1 in both cases. Conversely, let us prove that tb(G0 ) = 1 implies that tb(G) = 1. Let (T 0 , X 0 ) be a stardecomposition of G0 , that exists by Lemma 3, minimizing the distance in T 0 between the subtrees 0 Ta0 v and Tc0v . Let us remove vertex v from all bags, that leads to a tree decomposition (T 0 , X−v ) 0 of G \ v = G \ v of breadth one because NG0 [v] ⊆ NG0 [bv ]. Then, let (T, X ) be any reduced tree 0 decomposition that is obtained from (T 0 , X−v ), that is a star-decomposition of G \ v by Lemma 3. There are two cases. If Tav ∩ Tcv 6= ∅, then one obtains a star-decomposition of G simply by making some bag of Tav ∩ Tcv adjacent to the new bag NG [v] = {av , cv , v}. Else, Tav ∩ Tcv = ∅, so, Ta0 v ∩ Tc0v = ∅. Since Πv ∈ P3 (G0 ) and G0 is also prime, therefore, one of Πv or Πv \ bv must be a minimal separator of G0 . As a result, since (T 0 , X 0 ) is assumed to minimize the distance in T 0 between the subtrees Ta0 v and Tc0v , by Corollary 8 there exist two adjacent bags Ba0 v , Bc0 v ∈ X 0 so that av ∈ Ba0 v \ Bc0 v and cv ∈ Bc0 v \ Ba0 v respectively dominate the bags Ba0 v and Bc0 v . In such case, one obtains a star-decomposition of G by adding v in the two bags Ba0 v , Bc0 v . So, in conclusion, tb(G) = 1 in both cases. 5.3.7 Proof of Step 3.2 (a) Theorem 13. Let G be a prime planar graph, let v be a leaf-vertex of Type 2, Πv = (av , bv , cv ) be as in Definition 5, and let uv ∈ / Πv ∪ {v} be such that (bv , uv ) is an edge-separator of G \ v. Suppose av and uv are non-adjacent, and either cv and uv are non-adjacent or the subset NG (av ) ∩ NG (uv ) is not an av uv -separator in the subgraph G \ (cv , v). Then, G/vav (obtained by contracting {v, av }) is planar and prime and tb(G) = 1 if and only if tb(G/vav ) = 1. Proof. The graph G/vav is a contraction of the planar graph G, therefore it is planar. Let us prove that G/vav is prime. By contradiction, let S be a minimal clique-separator of G/vav . Since G/vav is a supergraph of G \ v, S is also a separator of G \ v. Furthermore, it is not an av cv -separator because av , cv are adjacent in G/vav , therefore, by Lemma 16 S is a separator of G. Since G is prime by the hypothesis, S does not induce a clique of G, whence av , cv ∈ S. However, since (bv , uv ) is not a separator of G because G is prime by the hypothesis, therefore by Lemma 16 (uv , bv ) is an av cv -separator of G \ v. So, NG (av ) ∩ NG (cv ) ⊆ {v, bv , uv }, that implies NG (av ) ∩ NG (cv ) = {v, bv } 36 uv av uv cv bv av v bv cv v uv uv av bv cv av v bv cv v Figure 33: Cases when Theorem 13 applies and the edge {v, av } can be contracted to av . because av and uv are non-adjacent by the hypothesis. In such a case S ⊆ Πv , but then V \(Πv ∪{v}) cannot be a full component of G \ Πv , thus contradicting Theorem 6. As a result, the graph G/vav is planar and prime. If tb(G) = 1 then tb(G/vav ) = 1 because tree-breadth is contraction-closed by Lemma 2. Conversely, let us prove that tb(G/vav ) = 1 implies tb(G) = 1. To show this, let (T, X ) be a stardecomposition of G/vav , that exists by Lemma 3, minimizing the number of bags |X | (in particular, (T, X ) is a reduced tree decomposition). Assume moreover (T, X ) to minimize the number of bags that are not contained into the closed neighbourhood of some vertex in G w.r.t. this property. Note that there is a bag of (T, X ) containing Πv , because since it is a clique of G/vav the subtrees Tav , Tbv , Tcv are pairwise intersecting and so, by the Helly property (Lemma 1), Tav ∩ Tbv ∩ Tcv 6= ∅. So, we can add in (T, X ) a new bag NG [v], and by making this bag adjacent to any bag of Tav ∩ Tbv ∩ Tcv one obtains a tree decomposition of G (not necessarily a star-decomposition). Consequently, we claim that to prove that tb(G) = 1, it suffices to prove that (T, X ) is a star-decomposition of G\v, for then the above construction leads to a star-decomposition of G. By contradiction, suppose it is not the case that (T, X ) is a star-decomposition of G \ v. Since G/vav and G \ v only differ in the edge {av , cv }, there must be a bag B of Tav ∩ Tcv that is only dominated by some of av , cv . We make the stronger claim that the bag B has a unique dominator, that is either av or cv . Since B is only dominated by some of av , cv , then in order to prove the claim by contradiction we only need to consider the case when B ⊆ NG/vav [av ] ∩ NG/vav [cv ]. Recall that NG/vav [av ]∩NG/vav [cv ] = {av , bv , cv } by the above remarks (because (uv , bv ) is an av cv -separator of G \ v), therefore either B = {av , bv , cv } or B = {av , cv }. In the first case (B = {av , bv , cv }) we have that B ⊆ N [bv ], thus contradicting the fact that B is only dominated by some of av , cv . However in the second case (B = {av , cv }) the bag B is strictly contained in any bag of the nonempty subtree Tav ∩ Tbv ∩ Tcv , thus contradicting the fact that (T, X ) is a reduced tree decomposition by minimality of |X |. Therefore, the claim is proved and so, the bag B has a unique dominator, that is either av or cv . Note that if B ⊆ NG/vav [cv ] then we may further assume that cv , uv are nonadjacent, or else by Theorem 6 NG/vav [cv ] = {av , bv , cv , uv } ⊆ N [bv ] and so, B ⊆ N [bv ], that would contradict the claim that B is only dominated by some of av , cv . In addition, since av and cv play symmetrical roles in the case when uv , cv are nonadjacent, let us assume w.l.o.g. that vertex av is the sole dominator of the bag B. In such a case, NG/vav (av ) ∩ NG/vav (cv ) = {bv } because (uv , bv ) is an av cv -separator of G \ v, so, since N (cv ) \ (Πv ∪ {v}) 6= ∅ because G is prime by the hypothesis, the existence of a bag B 0 containing vertex cv and adjacent to B follows. By the properties of a tree decomposition, B ∩ B 0 is a separator of G/vav Now, let Ca be the component of vertex av in G \ (bv , uv , v). Observe that cv ∈ / Ca because (uv , bv ) is an av cv -separator of G \ v. Since B ∩ B 0 ⊆ NG/vav [av ] ⊆ Ca ∪ Πv , therefore, B ∩ B 0 ∩ Ca 6= ∅ or else B ∩ B 0 would be a clique-separator in G/vav (impossible since it is a prime graph). There are several cases to be considered depending on the dominators of bag B0. • If av dominates B 0 then B, B 0 can be merged into one, thus contradicting the minimality of |X |; • Else, B 0 must be dominated by one of bv or uv because B ∩ B 0 ∩ Ca 6= ∅, cv ∈ (B ∩ B 0 ) \ Ca and (bv , uv ) separates cv from Ca . In fact, we claim that it cannot be dominated by vertex 37 uv . By contradiction, suppose that it is the case. Since av and uv are non-adjacent, therefore, av ∈ B \ B 0 and uv ∈ B 0 \ B. So, it follows by the properties of a tree decomposition that B ∩ B 0 is an av uv -separator of G/vav . However, B ∩ B 0 ⊆ N (av ) ∩ N (uv ), that contradicts the hypothesis that NG (av ) ∩ NG (uv ) is not an av uv -separator in the subgraph G \ (cv , v). Therefore, bv ∈ B 0 dominate the bag. Observe that if it were the case that there are at least two bags that are both adjacent to B and dominated by bv , then they could all be merged into one without violating the property for (T, X ) to be a star-decomposition. As a result, by minimality of |X |, B 0 is the unique bag that is both adjacent to B and dominated by bv , whence it is also the unique bag adjacent to B containing vertex cv . Let us substitute the two bags B, B 0 with B \ cv , B 0 ∪ {av }. Since NG/vav (av ) ∩ NG/vav (cv ) = {bv }, it is still a star-decomposition of G/vav with equal number of bags |X |. Furthermore, there is one less bag that is not contained in the closed neighbourhood of some vertex in G, thus contradicting the minimality of (T, X ). 5.3.8 Proof of Step 3.2 (b) i and Step 3.2 (b) ii In order to deal with all remaining cases, it will require us to further study the neighbourhood of vertex bv in the graph. Observe that in the following Theorem 14 we needn’t prove that the resulting graph G0 is prime because it will be proved in Theorem 15. Theorem 14. Let G be a prime planar graph, let v be a leaf-vertex of Type 2, Πv = (av , bv , cv ) be as in Definition 5, and let uv ∈ / Πv ∪ {v} be such that (bv , uv ) is an edge-separator of G \ v. Suppose uv ∈ N (cv ) \ N (av ), N (av ) ∩ N (uv ) is an av uv -separator of G \ (cv , v), and N (bv ) = {av , cv , uv , v}. Then, there exists x ∈ (N (av ) ∩ N (uv )) \ bv such that the graph G0 , obtained from G by adding the edge {bv , x}, is planar and satisfies tb(G0 ) = 1 if tb(G) = 1. Moreover, the vertex x can be found in linear-time. x av uv bv uv x cv av bv v v uv cv av bv x cv v Figure 34: Cases when one of Theorem 14 or Theorem 15 applies and vertex bv can be eventually contracted to another vertex. Proof. First, we claim that (av , uv ) is a minimal 2-separator of G. Indeed, by the hypothesis cv and uv are adjacent, therefore, by Theorem 6 NG (cv ) = {bv , uv , v}. In addition, N (bv ) = {av , cv , uv , v} by the hypothesis. Last, since G is prime by the hypothesis, therefore, N (av ) \ (Πv ∪ {v}) 6= ∅, and so, since av and uv are non-adjacent by the hypothesis, V (G) \ (av , bv , cv , uv , v) 6= ∅. As a result, (av , uv ) is a minimal 2-separator of G with {bv , cv , v} being a full component of G \ (av , uv ). Since N (av ) ∩ N (uv ) is an av uv -separator of G \ (cv , v) by the hypothesis, therefore, N (av ) ∩ N (uv )\bv 6= ∅, for it has to contain a vertex from every component of G\(av , bv , cv , uv , v). For now, let x ∈ N (av ) ∩ N (uv ) \ bv be arbitrary. Let us prove that tb(G) = 1 implies that tb(G0 ) = 1 where G0 is obtained by adding an edge between bv and x (for now, G0 may not be planar, depending on the choice for x). To prove this, let (T, X ) be a star-decomposition of G, that exists by Lemma 3, minimizing the distance in T between the subtrees Tav and Tuv . We claim that Tav ∩ Tuv 6= ∅. By contradiction, if Tav ∩ Tuv = ∅, then by Corollary 8, there are two bags Bav , Buv that are adjacent in T and such that av ∈ Bav \ Buv , uv ∈ Buv \ Bav respectively dominate Bav , Buv . However, this implies by the properties of a tree decomposition that Bav ∩ Buv ⊆ N (av ) ∩ N (uv ) is an av uv -separator of G. Since the av uv -path (av , v, cv , uv ) does not intersect N (av ) ∩ N (uv ), that is clearly a contradiction, and so, Tav ∩ Tuv 6= ∅. Furthermore, since there is a full component of G\(av , uv ) in the subgraph G\(bv , cv , v), therefore, by Lemma 9 the removal of vertices bv , cv , v from all bags in X leads to a tree decomposition (T, X − ) of breadth one of G \ (bv , cv , v). Let (T 0 , X 0 ) be a reduced star-decomposition obtained from (T, X − ), thar exists by Lemma 3. Since the subtrees Ta0 v , Tx0 , Tu0 v are pairwise intersecting 38 (because x ∈ N (av ) ∩ N (uv ) and Tav ∩ Tuv 6= ∅), therefore by the Helly property (Lemma 1) Ta0 v ∩ Tx0 ∩ Tu0 v 6= ∅. Let B ∈ Ta0 v ∩ Tx0 ∩ Tu0 v . To obtain a star-decomposition of G0 , it now suffices to make the bag B adjacent to the new bag NG0 [bv ] = {av , bv , cv , uv , v, x}. The above result holds for any choice of vertex x ∈ (NG (av )∩NG (uv ))\bv . Let us finally prove that one such a vertex x exists so that G0 is planar. Indeed, since N (av ) ∩ N (uv ) is an av uv -separator of G \ (cv , v) by the hypothesis, therefore, S := (N (av ) ∩ N (uv )) ∪ {v} is an av uv -separator of G, and in particular it is a minimal av uv -separator (because for every vertex s ∈ S, there is an av uv -path that intersects S only in s). By Corollary 7, it can be computed in linear-time a planar supergraph GS of G so that S induces a cycle of GS . Then, let NGS (bv ) ∩ S = {x, v}, by construction the graph G0 is planar for such a choice of vertex x. In Theorem 14, we show conditions so that vertex bv can be made adjacent to some other vertex of NG (av ) ∩ NG (uv ). Lemma 23 completes the picture by proving that if it is the case that NG (av ) ∩ NG (uv ) ∩ NG (bv ) 6= ∅, then |NG (av ) ∩ NG (uv ) ∩ NG (bv )| = 1 and vertex bv has exactly five neighbours. Lemma 23. Let G be a prime planar graph, let v be a leaf-vertex of Type 2, Πv = (av , bv , cv ) be as in Definition 5, and let uv ∈ / Πv ∪ {v} be such that (bv , uv ) is an edge-separator of G \ v. Suppose uv ∈ NG (cv ) \ NG (av ) and there exists x ∈ NG (av ) ∩ NG (uv ) ∩ NG (bv ). Then, NG (bv ) = {av , cv , uv , v, x}. uv x C bv cv av v Figure 35: Case when NG (bv ) 6= {av , cv , uv , v, x}. Proof. By contradiction, let C be a component of G \ (av , bv , cv , uv , v, x) such that bv ∈ N (C) (see Figure 35 for an illustration). By Theorem 6 NG (cv ) = {bv , uv , v}, therefore, cv , v ∈ / N (C). It follows that N (C) is a separator of G. In particular, N (C) ⊆ {av , bv , uv , x}, so, av , uv ∈ N (C) or else N (C) should be a clique-separator of the prime graph G. As a result, there is a K3,3 -minor with {av , bv , uv } and {C, x, {cv , v}} being the two sides of the bipartition. It contradicts the fact that G is planar by the hypothesis. Theorem 15. Let G be a prime planar graph, let v be a leaf-vertex of Type 2, Πv = (av , bv , cv ) be as in Definition 5, and let uv ∈ / Πv ∪ {v} be such that (bv , uv ) is an edge-separator of G \ v. Suppose uv ∈ N (cv ) \ N (av ), N (av ) ∩ N (uv ) is an av uv -separator of G \ (cv , v), and either N (bv ) = {av , cv , uv , v} or N (bv ) ∩ N (av ) ∩ N (uv ) 6= ∅. Then, there is x ∈ N (av ) ∩ N (uv ) such that one of the following must hold: • V (G) = {av , bv , cv , uv , v, x}, and G admits a star-decomposition with two bags NG [bv ], NG [x]; • or Π0 = (av , x, uv ) ∈ P3 (G), and let G0 be obtained from G by adding the edge {bv , x} (if it is not already present) then contracting this edge. The graph G0 is planar and prime, furthermore tb(G) = 1 if and only if tb(G0 ) = 1. Moreover, vertex x can be computed in linear-time. Proof. There are two cases. If NG (bv ) = {av , cv , uv , v}, then let x be set as in the statement of Theorem 14. Else, let x be the unique vertex of N (bv ) ∩ N (av ) ∩ N (uv ), that is well-defined by Lemma 23. Note that in both cases, vertex x can be computed in linear-time. In addition, N (bv ) ⊆ {av , cv , uv , v, x} (the latter property following from Lemma 23 when bv and x are adjacent, and being trivial else). Suppose for the proof that V (G) 6= {av , bv , cv , uv , v, x} (else, Theorem 15 is trivial). We claim that {bv , cv , v} is a component of G \ Π0 . Indeed, N (bv ) ⊆ Π0 ∪ {cv , v} by the hypothesis, and by Theorem 6 NG (cv ) = {bv , uv , v}. Since V (G) 6= {av , bv , cv , uv , v, x}, then it indeed follows that Π0 ∈ P3 (G), with {bv , cv , v} being a component of G \ Π0 . Let us prove that G0 is prime and planar. By Theorem 14, adding an edge between bv and x if it is not already present does not violate the property for the graph G to be planar. Therefore, G0 39 is planar because it is obtained by an edge-contraction from some planar graph. To prove that G0 is prime, by contradiction suppose the existence of a minimal clique-separator S 0 of G0 . Let us denote by x0 the vertex resulting from the contraction of the edge {bv , x}. Let S := S 0 if x0 ∈ / S 0 , S := (S 0 \ x0 ) ∪ {bv , x} else. By construction, S is a separator of G. In particular, S is not a clique because G is prime by the hypothesis. Therefore, S 6= S 0 , whence x0 ∈ S 0 or equivalently, x, bv ∈ S. We now claim that cv ∈ S ∩ S 0 or v ∈ S ∩ S 0 (possibly, v, cv ∈ S ∩ S 0 ). There are two cases. • Suppose that S \ bv is a separator of G. Then, S \ bv is not a clique because G is prime by the hypothesis. Since S \ (bv , x) = S 0 \ x0 is a clique, there must be some vertex of S \ (bv , x) = S ∩ S 0 that is adjacent to x0 in G0 but non-adjacent to x in G. Consequently, v ∈ S ∩ S 0 or cv ∈ S ∩ S 0 . • Else, S \ bv is not a separator of G. Recall that by construction, S is a separator of G. In particular, there must be two neighbours of bv in G that are separated by S in G. Since NG (bv ) \ x induces the path (av , v, cv , uv ), it follows that S must contain an internal node of the path, whence cv ∈ S ∩ S 0 or v ∈ S ∩ S 0 . However, in such case S 0 must be contained in one of (av , x0 , v), (v, x0 , cv ) or (cv , x0 , uv ), for it is a clique of G0 . In particular, let z ∈ {av , uv } \ S 0 . Since z has a neighbour in every component C 0 of G \ Π0 , {z} ∪ C 0 is not disconnected by S 0 in G0 . Furthermore, let us contract C 0 to z so as to make av and uv adjacent, S 0 intersects the resulting cycle (av , uv , cv , v) either in an edge (different from {av , uv }) or a single vertex because it is a clique of G0 , therefore, (av , uv , cv , v) \ S 0 is not disconnected by S 0 . Altogether, this contradicts the fact that S 0 is a separator of G0 , and so, G0 is prime. Finally, let us prove that tb(G0 ) = 1 if and only if tb(G) = 1. If tb(G) = 1, then let us assume bv and x to be adjacent (if they are not, then Theorem 14 ensures we can add the edge without violating the property for the graph to be of tree-breadth one). Then, tb(G0 ) = 1 because it is obtained by an edge-contraction from some graph with tree-breadth one and that tree-breadth is contraction-closed by Lemma 2. Conversely, let us prove that tb(G0 ) = 1 implies that tb(G) = 1. To prove this, let (T, X ) be a star-decomposition of G0 , that exists by Lemma 3, minimizing the distance in T between the subtrees Tav and Tuv . We claim that Tav ∩ Tuv 6= ∅. By contradiction, suppose Tav ∩ Tuv = ∅. Recall that (av , x0 , uv ) ∈ P3 (G0 ) (because Π0 ∈ P3 (G)) and G0 is prime, therefore one of (av , x0 , uv ) or (av , uv ) is a minimal separator of G0 . Since we assume the distance in T between Tav and Tuv to be minimized, by Corollary 8, there are two bags Bav , Buv that are adjacent in T so that av ∈ Bav \ Buv , uv ∈ Buv \ Bav respectively dominate Bav , Buv . However, by the properties of a tree decomposition Bav ∩ Buv ⊆ N (av ) ∩ N (uv ) is an av uv -separator of G0 , that is impossible due to the existence of the path (av , v, cv , uv ) in G0 that does not intersect N (av ) ∩ N (uv ). Therefore, Tav ∩ Tuv 6= ∅. Hence the subtrees Tav , Tx0 , Tuv are pairwise intersecting and so, by the Helly Property (Lemma 1), Tav ∩Tx0 ∩Tuv 6= ∅. Furthermore, NG0 [cv ]∪NG0 [v] ⊆ NG0 [x0 ] by construction. So, let us construct a tree decomposition of G of breadth one as follows. First, let us remove cv and v from all bags in X . Since NG0 [cv ] ∪ NG0 [v] ⊆ NG0 [x0 ], one obtains a tree decomposition of G0 \ (cv , v) of breadth one. Then let us replace x0 with x in all bags. Note that in so doing, one obtains a tree decomposition of G \ (bv , cv , v) of breadth one. Finally, let us make adjacent the new bag NG [bv ] with any bag of Tav ∩ Tx ∩ Tuv . The result is indeed a tree decomposition of G0 because NG [bv ] ⊆ {av , bv , cv , uv , v, x} and NG [cv ] ∪ NG [v] ⊆ NG [bv ]. 5.3.9 Proof of Step 3.2 (b) iii Theorem 16. Let G be a prime planar graph, let v be a leaf-vertex of Type 2, Πv = (av , bv , cv ) be as in Definition 5, and let uv ∈ / Πv ∪ {v} be such that (bv , uv ) is an edge-separator of G \ v. Suppose uv ∈ N (cv ) \ N (av ), N (av ) ∩ N (uv ) is an av uv -separator in the subgraph G \ (cv , v), N (bv ) 6= {av , cv , uv , v} and N (av ) ∩ N (bv ) ∩ N (uv ) = ∅. Then it can be computed in linear-time (a unique) x ∈ N (av ) ∩ N (uv ) such that if tb(G) = 1, N (bv ) ∩ N (x) is a bv x-separator and |N (bv ) ∩ N (x)| ≥ 3. Proof. Let W = (N (av ) ∩ N (uv )) ∪ {av , cv , uv , v}. By the hypothesis, N (bv ) 6= {av , cv , uv , v} and N (av ) ∩ N (bv ) ∩ N (uv ) = ∅, therefore, it exists a component C0 of G \ W such that bv ∈ N (C0 ). 40 We claim that there is x ∈ N (av ) ∩ N (uv ) ∩ N (C0 ) satisfying that N (C0 ) ⊆ (av , bv , x) or N (C0 ) ⊆ (uv , bv , x). Indeed, first observe that v, cv ∈ / N (C0 ) because by Theorem 6 NG (cv ) = {bv , uv , v}. Furthermore, av ∈ / N (C0 ) or uv ∈ / N (C0 ) because N (av ) ∩ N (uv ) is an av uv -separator of G \ (cv , v) by the hypothesis. So, let {z, z 0 } = {av , uv } satisfy z 0 ∈ / N (C0 ). Since G is prime by the hypothesis, and so, biconnected, G \ z is connected. Furthermore, N (C0 ) \ z ⊆ N (z 0 ) (by the definition of W ), therefore, N (C0 ) \ z is a minimal separator of G \ z. By Lemma 13 there exist s, t ∈ N (av ) ∩ N (uv ) non-adjacent such that N (C0 ) \ z = {s, t}. Since bv ∈ N (C0 ) by construction, therefore, let us set {s, t} = {bv , x}, that finally proves the claim. We claim in addition that x does not depend on the choice of the component C0 . By contradiction, let C, C 0 be two components of G \ W such that bv ∈ N (C) ∩ N (C 0 ) and let x, x0 ∈ N (av ) ∩ NG (uv ) be distinct and such that x ∈ N (C), x0 ∈ N (C 0 ). Then, there exists a K3,3 -minor with {av , bv , uv } and {{cv , v}, C ∪ {x}, C 0 ∪ {x0 }} being the sides of the bipartition, that contradicts the hypothesis that G is planar. Thus from now on, let x ∈ N (av ) ∩ N (uv ) \ bv be the unique vertex satisfying that for every component C of G \ W , if bv ∈ N (C) then x ∈ N (C). x uv C0 bv av cv v Figure 36: Component C0 such that bv , x ∈ N (C0 ). Recall that C0 is a fixed component of G \ W such that bv , x ∈ N (C0 ) (see Figure 36 for an illustration). Finally, assume for the remaining of the proof that tb(G) = 1 and let us prove that N (bv ) ∩ N (x) is a bv x-separator and |N (bv ) ∩ N (x)| ≥ 3. To prove it, we will only need to prove that N (bv ) ∩ N (x) is a bv x-separator of G. Indeed, in such a case N (bv ) ∩ N (x) ∩ C0 6= ∅, and so, |N (bv ) ∩ N (x)| ≥ 3 because av , uv ∈ N (bv ) ∩ N (x) and av , uv ∈ / C0 . Let (T, X ) be star-decomposition of G, that exists by Lemma 3, minimizing the distance in T between the subtrees Tbv and Tx . We claim that Tbv ∩ Tx = ∅. By contradiction, suppose Tbv ∩ Tx 6= ∅. Let us prove as an intermediate subclaim that Tav ∩ Tuv 6= ∅. By contradiction, let Tav ∩ Tuv = ∅. By the properties of a tree decomposition, every bag B onto the path in T between Tav and Tuv must contain N (av ) ∩ N (uv ) and at least one of v or cv . If cv ∈ B then B ⊆ N [uv ] since N (cv ) = {bv , uv , v} and x ∈ B. Similarly if v ∈ B then B ⊆ N [av ] since N (v) = {av , bv , cv } and x ∈ B. Consequently, there are two adjacent bags Bav , Buv such that av ∈ Bav \ Buv and uv ∈ Buv \ Bav respectively dominate Bav and Buv . However, by the properties of a tree decomposition, Bav ∩ Buv = N (av ) ∩ N (uv ) is an av uv -separator of G, thus contradicting the existence of the path (av , v, cv , uv ) in G. Therefore, it follows that Tav ∩ Tuv 6= ∅, that proves the subclaim. If Tav ∩ Tuv 6= ∅ and Tbv ∩ Tx 6= ∅ then the subtrees Tav , Tbv , Tuv , Tx are pairwise intersecting and so, it implies Tav ∩ Tuv ∩ Tbv ∩ Tx 6= ∅ by the Helly property (Lemma 1). However, let B 0 ∈ Tav ∩ Tuv ∩ Tbv ∩ Tx , no vertex in G can dominate B 0 because N (bv ) ∩ N (av ) ∩ N (uv ) = ∅ by the hypothesis, thus contradicting the fact that (T, X ) is a star-decomposition. As a result, we proved the claim that Tbv ∩ Tx = ∅. Finally, since there exists S ⊆ (N (bv ) ∩ N (x)) ∪ {bv , x} a minimal separator of G such that bv , x ∈ S (namely, S := N (C0 )), and (T, X ) is assumed to minimize the distance in T between Tbv and Tx , by Corollary 8 there exist two adjacent bags Bbv , Bx such that bv ∈ Bbv \ Bx , x ∈ Bx \ Bbv respectively dominate Bbv and Bx . By the properties of a tree decomposition, Bbv ∩ Bx = N (bv ) ∩ N (x) is indeed a bv x-separator of G. Lemma 24. Let G be a prime planar graph, let v be a leaf-vertex of Type 2, Πv = (av , bv , cv ) be as in Definition 5, and let uv ∈ / Πv ∪ {v} be such that (bv , uv ) is an edge-separator of G \ v. Suppose uv ∈ N (cv ) \ N (av ), N (av ) ∩ N (uv ) is an av uv -separator in the subgraph G \ (cv , v), N (bv ) 6= {av , cv , uv , v} and N (av ) ∩ N (bv ) ∩ N (uv ) = ∅. Assume furthermore that there is x ∈ N (av ) ∩ N (uv ), and there exists a leaf-vertex l ∈ N (bv ) ∩ N (x). Then, l is a leaf-vertex of Type 1, or l is a leaf-vertex of Type 2 or 3 and G \ l is prime. 41 uv x uv x av bv cv uv x l l l av cv bv av v v bv cv v Figure 37: Existence of a leaf-vertex in N (bv ) ∩ N (x). Proof. Suppose for the proof that l is not of Type 1 (else, Lemma 24 is trivial). Then, l is of Type 2 or 3, let Πl be as in Definition 5. Note that l 6= av because v, bv , x ∈ N (av ) do not induce a path, similarly l 6= uv because bv , cv , x ∈ N (uv ) do not induce a path. Furthermore by the hypothesis, bv and x are the two endpoints of Πl . Suppose by way of contradiction that there is a minimal clique-separator S of G \ l. Since G is prime by the hypothesis, by Lemma 16 S is a bv x-separator of G \ l. However, it implies that av , uv ∈ S, that contradicts the fact that S is a clique. As a result, G \ l is prime. Equipped with Lemma 24, we can assume from now on that there is no leaf-vertex that is adjacent to both vertices bv , x, or else it could be immediately processed by the algorithm. Theorem 17. Let G be a prime planar graph, let v be a leaf-vertex of Type 2, Πv = (av , bv , cv ) be as in Definition 5, and let uv ∈ / Πv ∪ {v} be such that (bv , uv ) is an edge-separator of G \ v. Suppose uv ∈ N (cv ) \ N (av ), N (av ) ∩ N (uv ) is an av uv -separator in the subgraph G \ (cv , v), N (bv ) 6= {av , cv , uv , v} and N (av ) ∩ N (bv ) ∩ N (uv ) = ∅. Assume furthermore that there is x ∈ N (av ) ∩ N (uv ) such that N (bv ) ∩ N (x) is a bv x-separator, |N (bv ) ∩ N (x)| ≥ 3, and there is no leaf-vertex in N (bv ) ∩ N (x). Then, there exist y, z ∈ N (bv ) ∩ N (x) non-adjacent such that the graph G0 , obtained from G by making y, z adjacent, is planar and prime, and it holds tb(G) = 1 if and only if tb(G0 ) = 1. Furthermore, the pair y, z can be computed in linear-time. uv x y z av bv cv v Figure 38: Illustration of Theorem 17. Proof. Let us first show how to find the pair y, z. Let W = {av , cv , v, uv }∪(N (av )∩N (uv )). Choose any component C0 of G \ W such that bv , x ∈ N (C0 ) and N (C0 ) ⊆ (av , bv , x) or N (C0 ) ⊆ (uv , bv , x) (the existence of such a component has been proved in Theorem 16). Note that N (bv )∩N (x)∩C0 6= ∅ since N (bv ) ∩ N (x) is a bv x-separator of G by the hypothesis. Then, let S := N (bv ) ∩ N (x). By the hypothesis S is a minimal separator of G and |S| ≥ 3, therefore, by Corollary 7 there is a planar supergraph GS of G so that S induces a cycle of GS . Furthermore, GS can be computed in linear-time. Let P be an av uv -path of the cycle GS [S] that intersects C0 . Since by the above claim av ∈ / NG (C0 ) or uv ∈ / NG (C0 ), therefore, there is y ∈ C0 ∩ V (P ), there is z adjacent to vertex y in P so that either z ∈ C1 for some component C1 of G \ (W ∪ C0 ) or z ∈ {av , uv } \ NG (C0 ). In particular, z ∈ / NG [C0 ] = C0 ∪ NG (C0 ). Moreover, the graph G0 , obtained from G by adding an edge between y and z, is planar by construction. Claim 15. G0 is prime. Proof. By contradiction, let X be a minimal clique-separator of G0 . Since G0 is a supergraph of G, X is a separator of G. As a result, y, z ∈ X because G is prime by the hypothesis. Let us prove as an intermediate step that N (y) ∩ N (z) = {bv , x}. There are two cases. If z ∈ {av , uv }, then let {z, z 0 } = {av , uv }. Since N (C0 ) ⊆ (z 0 , bv , x) (because z ∈ / NG (C0 )) and z, z 0 are non-adjacent by the 42 hypothesis, therefore, the claim immediately follows in this case. Else, z ∈ / {av , uv }. Let C1 be the component of G\(W ∪C0 ) containing z. In such case, bv , x ∈ N (C1 ) because z ∈ S by construction. Therefore, N (C1 ) ⊆ (av , bv , x) or N (C1 ) ⊆ (uv , bv , x) because the respective roles of components C0 , C1 are symmetrical in this case. Suppose by way of contradiction N (C0 ) = N (C1 ) = {s, bv , x} for some s ∈ {av , uv } and let {s, t} = {av , uv }. Then, there is a K3,3 -minor of G with {bv , x, s} and {C0 , C1 , {cv , v, t}} being the sides of the bipartition, that contradicts the hypothesis that G is planar. As a result, N (C0 ) ∩ N (C1 ) = {bv , x}, that finally proves the claim. Since X is assumed to be a clique of G0 and y, z ∈ X, it follows X ⊆ {bv , y, z} or X ⊆ {x, y, z}. Consequently, G[W \ X] is connected because av , uv ∈ / X is a dominating pair of W , bv , x ∈ N (av ) ∩ N (uv ) and bv ∈ / X or x ∈ / X. However, since y, z ∈ X and NG (y) ⊆ W ∪ C0 , then in such case there must be a component A of G \ X so that A ⊂ C0 . Since z ∈ / NG [C0 ] by construction, NG (A) ⊆ X \ z is a clique-separator of G, thus contradicting the hypothesis that G is prime. As a result, G0 is prime.  Now, let us prove tb(G) = 1 if and only if tb(G0 ) = 1. If tb(G) = 1, then let (T, X ) be star-decomposition of G, that exists by Lemma 3, minimizing the distance in T between Tav and Tuv . Let us prove that (T, X ) is a star-decomposition of G0 , whence tb(G0 ) = 1. To prove it, it is sufficient to prove Ty ∩ Tz 6= ∅. We will prove as an intermediate claim that Tav ∩Tuv 6= ∅. By contradiction, assume Tav ∩Tuv = ∅. Observe that Π0 = (av , bv , uv ) ∈ P3 (G) with {cv , v} being a full component of G \ Π0 . Therefore, since G is prime by the hypothesis, one of Π0 or Π0 \ bv is a minimal separator of G. Since (T, X ) is assumed to minimize the distance in T between Tav and Tuv , therefore, by Corollary 8 there are two adjacent bags Bav , Buv such that av ∈ Bav \Buv and uv ∈ Buv \Bav respectively dominate Bav and Buv . However, by the properties of a tree decomposition Bav ∩ Buv = N (av ) ∩ N (uv ) is an av uv -separator of G, that contradicts the existence of the path (av , v, cv , uv ) in G. Consequently the claim is proved, hence Tav ∩ Tuv 6= ∅. The latter claim implies Tbv ∩ Tx = ∅, for if Tbv ∩ Tx 6= ∅ then the subtrees Tav , Tbv , Tuv , Tx are pairwise intersecting, hence Tav ∩ Tbv ∩ Tx ∩ Tuv 6= ∅ by the Helly property (Lemma 1), that would contradict the fact that (T, X ) is a star-decomposition because N (av ) ∩ N (uv ) ∩ N (bv ) = ∅ by the hypothesis. Finally, since (x, y, bv , z) induces a cycle of G and Tbv ∩ Tx = ∅, therefore, by the properties of a tree decomposition Ty ∩ Tz 6= ∅, and so, (T, X ) is indeed a star-decomposition of G0 . Conversely, let us prove tb(G0 ) = 1 implies tb(G) = 1. To prove it, let (T 0 , X 0 ) be a stardecomposition of G0 , that exists by Lemma 3, minimizing the number |X 0 | of bags. Assume furthermore (T 0 , X 0 ) to minimize the number of bags B ∈ X 0 that are not contained into the closed neighbourhood of some vertex in G w.r.t. the minimality of |X 0 |. In order to prove tb(G) = 1, it suffices to prove that (T 0 , X 0 ) is a star-decomposition of G. We will start proving intermediate claims. Claim 16. av , uv ∈ / NG (y). Proof. By contradiction, assume the existence of z 0 ∈ {av , uv } so that z 0 and y are adjacent in G. In particular, z 0 6= z (since z ∈ / NG [C0 ]) and NG (C0 ) = {bv , x, z 0 } since either NG (C0 ) ⊆ {bv , x, av } or NG (C0 ) ⊆ {bv , x, uv }. Hence, the path (bv , z 0 , x) is a separator of G. Since y ∈ NG (bv ) ∩ NG (z 0 ) ∩ NG (x), by Lemma 15 either C0 is reduced to y or (bv , y, x) ∈ P3 (G) separates z 0 from C0 \ y. The case C0 \ y = ∅ implies that y is a leaf-vertex of Type 2, that contradicts the hypothesis that there is no leaf-vertex in NG (bv ) ∩ NG (x). Therefore, let (bv , y, x) ∈ P3 (G) separates z 0 from C0 \ y in G, and let C00 ⊆ C0 \ y be a component of G \ (bv , y, x) (such a component C00 exists because NG (C0 ) = {z 0 , bv , x}). Since G is prime, bv , x ∈ NG (C00 ) (indeed, neither bv nor y nor x nor (bv , y) nor (y, x) can be a separator of G). Therefore, NG (bv ) ∩ NG (x) ∩ C00 6= ∅ because NG (bv ) ∩ NG (x) is a bv x-separator of G by the hypothesis. Furthermore, y ∈ NG (C00 ) because C0 is connected. However in such case, there is a K3,3 -minor of G0 with {bv , x, y} and {av , C00 , uv } being the sides of the bipartition, that contradicts the fact that G0 is planar.  Claim 17. There is no vertex dominating the cycle (av , bv , uv , x) in G0 . Proof. By contradiction, if it were the case that such a vertex exists, then, since NG (bv ) ∩ NG (av ) ∩ NG (uv ) = ∅ by the hypothesis, the dominator should be y and furthermore z ∈ {av , uv }. In particular, y ∈ NG (bv ) ∩ NG (x) ∩ NG (z 0 ) with {z, z 0 } = {av , uv }, thus contradicting Claim 16.  Claim 18. Ta0 v ∩ Tu0 v 6= ∅. Proof. By contradiction, let Ta0 v ∩ Tu0 v = ∅. By the properties of a tree decomposition, every bag B onto the path in T 0 between Ta0 v and Tu0 v (including the endpoints) must contain NG0 (av ) ∩ 43 NG0 (uv ) and at least one of v or cv . Then, if cv ∈ B then B ⊆ NG0 [uv ] and so, B ∈ Tu0 v , since NG0 (cv ) = {bv , uv , v} and x ∈ B. Similarly if v ∈ B then B ⊆ NG0 [av ] and so, B ∈ Ta0 v , since NG0 (v) = {av , bv , cv } and x ∈ B. Consequently, there are two adjacent bags Bav , Buv such that av ∈ Bav \ Buv and uv ∈ Buv \ Bav respectively dominate Bav and Buv in G0 . However, by the properties of a tree decomposition, Bav ∩ Buv = NG0 (av ) ∩ NG0 (uv ) is an av uv -separator of G0 , thus contradicting the existence of the path (av , v, cv , uv ) in G0 . Therefore, it follows that Ta0 v ∩ Tu0 v 6= ∅, that proves the claim.  Claim 19. Tb0v ∩ Tx0 = ∅ Proof. Suppose for the sake of contradiction that Tb0v ∩ Tx0 6= ∅. By Claim 18, Ta0 v ∩ Tu0 v 6= ∅, and so, the subtrees Ta0 v , Tb0v , Tu0 v , Tx0 are pairwise intersecting. Hence by the Helly property (Lemma 1), Ta0 v ∩ Tb0v ∩ Tu0 v ∩ Tx0 6= ∅. However in such case, since (T 0 , X 0 ) is a star-decomposition of G0 there must be a vertex dominating the cycle (av , bv , uv , x) in G0 , thereby contradicting Claim 17.  As a result, Ta0 v ∩ Tu0 v 6= ∅ by Claim 18 and Tb0v ∩ Tx0 = ∅ by Claim 19. Finally, suppose by way of contradiction (T 0 , X 0 ) is not a star-decomposition of G. In such case, since G and G0 only differ in the edge {y, z}, there must exist B ∈ Ty0 ∩ Tz0 that is uniquely dominated by some of y, z in G0 . More precisely, let us prove that only one of y, z can dominate B. By contradiction, suppose B ⊆ NG0 [y]∩NG0 [z] = {bv , x, y, z} (indeed, y ∈ C0 whereas z ∈ / NG [C0 ]). Since Tb0v ∩ Tx0 = ∅ by Claim 19, B ⊆ {bv , y, z} or B ⊆ {x, y, z, }. Therefore, B is a clique of G0 . However, since B 6= V (G0 ), there is a bag B 0 adjacent to B and by the properties of a tree decomposition B ∩ B 0 is a clique-separator of G0 , thus contradicting the fact that G0 is prime by Claim 15. Consequently, either B ⊆ NG0 [y] or B ⊆ NG0 [z], and either B 6⊆ NG0 [y] or B 6⊆ NG0 [z]. In the following, let {s, t} = {y, z} satisfy B ⊆ NG0 [s], that is well-defined. Let B 0 be any bag adjacent to B so that t ∈ B 0 (such bag exists because y, z ∈ N (bv ) ∩ N (x), and bv ∈ / B or x ∈ /B because Tb0v ∩ Tx0 = ∅). There are three cases. • Suppose no vertex of bv , x, y, z dominates B 0 in G0 (see Figure 39 for an illustration). Since bv ∈ / B or x ∈ / B because Tb0v ∩ Tx0 = ∅, therefore, (B ∩ B 0 ) \ (bv , x, y, z) 6= ∅, or else by the properties of a tree decomposition that would be a clique-separator of G0 , thus contradicting the fact that G0 is prime by Claim 15. Let t0 ∈ (B∩B 0 )\(bv , x, y, z). Note that t0 and t are nonadjacent in G0 because t0 ∈ NG0 [s] and NG0 [y] ∩ NG0 [z] = {bv , x, y, z}. Let s0 ∈ B 0 dominate this bag. Note that s0 and s are non-adjacent in G0 because we assume s0 ∈ / {bv , x, y, z}, t ∈ NG0 (s0 ) and NG0 [y] ∩ NG0 [z] = {bv , x, y, z}. In particular, s0 6= t0 and (s, t0 , s0 , t) induces a path in G. By construction, y ∈ C0 and z ∈ / NG [C0 ], hence there must be some of s0 , t0 in NG (C0 ). Since NG (C0 ) ⊆ {av , bv , uv , x} and s0 , t0 ∈ / {bv , x, y, z}, therefore the pairs {s0 , t0 } and {av , uv } intersect. However, by Claim 16 av , uv ∈ / NG (y), similarly av , uv ∈ / NG (z), that contradicts the existence of the path (s, t0 , s0 , t) in G. Consequently, assume in the remaining cases that there is some vertex of bv , x, y, z dominating B 0 in G0 . t s t' B s' B' Figure 39: B ⊆ NG0 [s], B 0 ⊆ NG0 [s0 ]. • Suppose B 0 is dominated by one of y, z in G0 . We claim that B and B 0 are dominated by the same vertex of y, z, for if it were not the case B ∩ B 0 ⊆ NG0 [y] ∩ NG0 [z] = {bv , x, y, z}, and so, since by the properties of a tree decomposition B ∩ B 0 is a separator of G0 , and bv ∈ / B or x ∈ / B because Tb0v ∩ Tx0 = ∅, B ∩ B 0 should be a clique-separator of G0 , thus contradicting the fact that G0 is prime by Claim 15. However, in such a case bags B, B 0 could be merged into one while preserving the property for the tree decomposition to be a star-decomposition of G0 , that would contradict the minimality of |X 0 |. • Therefore, B 0 is dominated by some of bv , x. We claim that there is a unique such bag B 0 that is adjacent to B. By contradiction, let B 00 6= B 0 be adjacent to B and such that B 00 is also dominated by some of bv , x. In particular, if B 00 ∪ B 0 ⊆ N [bv ] or B 00 ∪ B 0 ⊆ N [x] then both bags B 0 , B 00 could be merged into one without violating the property for the tree decomposition to be a star-decomposition of G0 , that would contradict the minimality of |X 0 |. 44 Else, w.l.o.g. B 0 ⊆ N [bv ] and B 00 ⊆ N [x]. Since Tb0v ∩ Tx0 = ∅, av , uv ∈ N (bv ) ∩ N (x) and B 0 , B 00 are adjacent to B, therefore, by the properties of the tree decomposition av , uv ∈ B. However, av , uv ∈ / NG [y] by Claim 16, av and uv are non-adjacent, and either z ∈ {av , uv } or av , uv ∈ / NG [z] (by the same proof as for Claim 16), thus contradicting the fact that either B ⊆ NG0 [y] or B ⊆ NG0 [z]. Hence, the claim is proved and B 0 is assumed to be the unique bag adjacent to B such that B ⊆ N [bv ] or B ⊆ N [x]. In particular, B 0 is the unique bag adjacent to B containing vertex t (recall that {s, t} = {y, z} and B ⊆ NG0 [s], B 6⊆ NG0 [t]). Let us substitute the bags B, B 0 with B \ t, B 0 ∪ {s}. We claim that this operation keeps the property for (T 0 , X 0 ) to be a star-decomposition of G0 . To prove the claim, first note that the operation only modifies bags B and B 0 , furthermore B \ t ⊆ N [s] and B 0 ∪ {s} ⊆ N [bv ] or B 0 ∪ {s} ⊆ N [x]. Consequently, to prove the claim, it suffices to prove that the operation keeps the property for (T 0 , X 0 ) to be a tree decomposition of G0 (for in such a case, it is always a star-decomposition). Since Tt0 \ B is connected because B 0 is the only bag containing vertex t that is adjacent to the bag B, therefore, we are left to prove that there is no w ∈ NG0 (t) \ s such that Tw0 ∩ Tt0 = {B}. By contradiction, let w ∈ NG0 (t) \ s satisfy Tw0 ∩ Tt0 = {B}. Since w ∈ B ⊆ NG0 [s], therefore w ∈ NG0 (s)∩NG0 (t) = NG (y)∩NG (z) = {bv , x}. Moreover, w ∈ / B0 because t ∈ B 0 and we assume that Tw0 ∩ Tt0 = {B}. In such a case, since it is assumed that B 0 ⊆ N [bv ] or B 0 ⊆ N [x], and in addition Tx0 ∩ Tb0v = ∅, let us write {w, w0 } = {bv , x} such that w ∈ B \ B 0 , w0 ∈ B 0 \ B and B 0 ⊆ NG0 [w0 ]. By the properties of a tree decomposition, B ∩ B 0 is a bv x-separator of G0 , so, av , uv ∈ B ∩ B 0 . However, av , uv ∈ / NG (y) by Claim 16 and similarly av , uv ∈ / NG (z), that contradicts the fact that B ⊆ NG0 [s] for some s ∈ {y, z}. This finally proves the claim that substituting the bags B, B 0 with B \ t, B 0 ∪ {s} keeps the property for (T 0 , X 0 ) to be a star-decomposition of G0 . However, the above operation does not increase the number of bags |X 0 |, furthermore there is one less bag that is not contained in the closed neighbourhood of some vertex in G. This contradicts the minimality of (T 0 , X 0 ) w.r.t. these two properties. As a result, we proved by contradiction that (T 0 , X 0 ) is a star-decomposition of G, hence tb(G) = 1. 5.4 Complexity of Algorithm Leaf-BottomUp To complete this section, let us emphasize on some computational aspects of Algorithm Leaf-BottomUp, that will ensure the quadratic-time complexity of the algorithm. We here assume that the planar graph G is encoded with adjacency lists. Note that the adjacency lists can be updated in linear-time before each recursive call to the algorithm. We will need as a routine to test whether two vertices are adjacent in constant-time. In order to achieve the goal, the following result (relying upon the bounded degeneracy of planar graphs) will be used: Lemma 25 ( [20]). There exists a data structure such that each entry in the adjacency matrix of a planar graph can be looked up in constant time. The data structure uses linear storage, and can be constructed in linear time. 5.4.1 Finding a leaf-vertex At each call to the algorithm, it is first required to decide whether a leaf-vertex exists. If that is the case, then one such a vertex must be computed. Here is a way to achieve the goal in linear-time. Let us start computing the degree sequence of G, then let us order the vertices of the graph G by increasing degree. Finding a leaf-vertex of Type 1. Let v be any vertex of degree at least four. We claim that a necessary condition for v to be a leaf-vertex of Type 1 is that all but at most two neighbours of v have degree four. Indeed, if v is a leaf-vertex of Type 1, then let Πv , dv be defined as in Definition 5. By Lemma 21, either V (G) = N [v] ∪ {dv } or Π0 = (av , dv , cv ) ∈ P3 (G) and N [v] \ (av , cv ) is a full component of G \ Π0 . In both cases, all neighbours in N (v) \ (av , cv ) have degree four. • Therefore, let us count the number of neighbours of degree four in N (v), that can be done in O(deg(v))-time simply by traversing the adjacency list of vertex v (recall that the degree sequence of G has been computed). 45 • If there are all but at most two neighbours in N (v) that have degree four, then we claim that one can construct the induced subgraph G[N (v)] in O(deg(v))-time. Indeed, for every neighbour u ∈ N (v) that has degree four, let us test in constant-time for each of its four neighbours whether they are adjacent to vertex v — we only keep those for which it is the case in the adjacency list of u in G[N (v)]. Then, for every u ∈ N (v) that does not have degree four (there are at most two such vertices), let us construct the adjacency list of u in G[N (v)] simply by testing to which vertices in N (v) \ u it is adjacent — the latter takes constant-time by neighbour. • Once G[N (v)] has been computed, it is easy to check whether it is a path in O(|N (v)|) = O(deg(v))-time. • Finally, let u ∈ N (v) have degree four. Let us pick in constant-time any neighbour dv ∈ N (u) \ N (v) (note that such a vertex is unique if G[N (v)] induces a path). In order to decide whether v is a leaf-vertex of Type 1, it is now sufficient to test whether vertex dv is adjacent to every vertex in N (v) — that takes constant-time by neighbour. Finding a leaf-vertex of Type 2. Recall that a vertex is a leaf-vertex of Type 2 if and only if it has degree three and its three neighbours induce a path. Given any vertex of degree three, three adjacency tests are enough in order to determine whether its three neighbours induce a path — and each adjacency tests takes constant-time. Therefore, it can be checked in constant-time whether a vertex is a leaf-vertex of Type 2. Finding a leaf-vertex of Type 3. By Definition 5, a vertex v is a leaf-vertex of Type 3 if and only if it has degree two and its two neighbours are non-adjacent and they have at least two common neighbours (including v). Note that given a degree-two vertex, it can be checked whether its two neighbours are non-adjacent in constant-time. We now distinguish three cases. 1. First, suppose there is a vertex v such that N (v) = {x, y} and neighbour x is a degree-two vertex. In such case, let N (x) = {v, z}, in order to decide whether v is a leaf-vertex of Type 3, it is sufficient to test in constant-time whether y, z are adjacent. 2. Second, suppose there are two degree-two vertices v, v 0 that share the same two non-adjacent neighbours (i.e., N (v) = N (v 0 ) = {x, y} and x, y are non-adjacent). In such case, both vertices v, v 0 are leaf-vertices of Type 3 (this case may happen if for instance, G = K2,q with q ≥ 2). In order to check whether this case happens, it is sufficient to sort the pairs N (v) with v being a degree-two vertex in linear-time (for instance, using a bucket-sort). 3. Else, let V 0 contain every degree-two vertex v with two non-adjacent neighbours of degree at least three (if one of the two neighbours of v has degree two, we fall in the first case) W.l.o.g., every vertex v ∈ V 0 is uniquely determined by the pair N (v) composed of its two neighbours (or else, we fall in the second case). In such case, let us contract every v ∈ V 0 to one of its two neighbours. By doing so, we remove v and we make the two vertices in N (v) adjacent. Note that all these edge-contractions are pairwise independent. Let us call G0 the graph resulting from all edge-contractions, and let us call “virtual edges” any new edge resulting from an edge-contraction. Then, let us list all triangles in the resulting graph G0 , it can be done in linear-time [45]. By construction, v ∈ V 0 is a leaf-vertex of Type 3 if and only if the virtual edge resulting from its contraction belongs to a triangle in which it is the unique virtual edge. P Overall, finding a leaf-vertex in G takes O( v∈V deg(v))-time, that is O(n)-time because G is planar. 5.4.2 Existence of a star-decomposition with two bags Lemma 26. Let G be a planar graph, it can be decided in quadratic-time whether G admits a star-decomposition with one or two bags. Proof. G admits a star-decomposition with one bag if and only if there is a universal vertex in G, hence it can be decided in linear-time. Assume for the remaining of the proof that G does not admit a star-decomposition with less than two bags. We will consider two necessary conditions for some fixed pair x, y to be the dominators of the only two bags in some star-decomposition of G. For each of the two conditions, we will prove that all pairs satisfying the condition can be computed in quadratic-time. Then, we will conclude the proof by showing that the two conditions are sufficient to ensure the existence of a star-decomposition of G with two bags. 46 1. Recall that if it exists a star-decomposition of G with two bags, then by the properties of a tree decomposition every vertex of G must be contained in at least one bag. Therefore, if x, y are the only two dominators of the bags in some star-decomposition of G, they must be a dominating pair of G. It can be decided in O(deg(x) + deg(y))-time whether a fixed pair x, y is a dominating pair. So, overall, it takes O(n2 )-time to compute all dominating pairs of G with n being the order of the graph, for the graph is planar and so, it is a sparse graph. 2. Furthermore, recall that if it exists a star-decomposition of G with two bags, then by the properties of a tree decomposition every edge of G must be contained in at least one bag. Therefore, if there is a star-decomposition of G with two bags that are respectively dominated by x and y, then it must be the case that there does not exist any edge e = {u, v} so that u ∈ N [x] \ N [y] and v ∈ N [y] \ N [x] (else, such an edge could not be contained in any of the two bags). In order to decide whether the latter condition holds for some fixed pair x, y, it suffices to test whether every vertex of N [x] \ N [y] is non-adjacent to all vertices in N [y] \ N [x] — it takes constant-time per test and so, O(deg(x) · deg(y))-time in total. As P a result, computing all pairs x, y satisfying the condition requires O( x,y deg(x) · deg(y)) = P P O([ x deg(x)][ y deg(y)]) = O(n2 )-time because the graph G is planar and so, it is a sparse graph. Finally, let x, y satisfy the two above necessary conditions. We claim that (T, X ) with T being an edge and X = {N [x], N [y]} is a star-decomposition of G. Indeed, every vertex is contained into a bag because the pair x, y satisfies the first necessary condition. Furthermore, every edge has its both ends contained into a common bag because the pair x, y satisfies the second necessary condition. Last, all the bags containing a common vertex induce a subtree because there are only two bags. As a result, (T, X ) is a tree decomposition of G. Since each bag of X is respectively dominated by x or y, therefore (T, X ) is indeed a star-decomposition of G, that proves the claim, hence the lemma. Note that in any execution of Algorithm Leaf-BottomUp, it is verified at most once whether some planar graph admits a star-decomposition with one or two bags. 5.4.3 Upper-bound on the number of steps in the algorithm Lemma 27. Let G be a prime planar graph with n vertices and m edges. Then, there are at most 5n − m recursive calls to the Algorithm Leaf-BottomUp, that is O(n). Proof. First note that since G is planar by the hypothesis, 5n − m ≥ 0 and 5n − m = O(n). Let G0 with n0 vertices and m0 edges so that Algorithm Leaf-BottomUp is recursively applied on G0 when G is the input. Since there is at most one such a graph G0 (i.e., there is no more than one recursive call at each call of the algorithm), furthermore G0 is prime and planar, therefore, in order to prove the lemma it suffices to prove that 5n0 − m0 < 5n − m. To prove it, let us consider at which step of the algorithm the recursive call occurs. • If it is at Step 2, then G0 is obtained by removing a leaf-vertex of Type 1, denoted by v, and then contracting all the internal vertices in the path Πv (induced by N (v)) to a single edge. Therefore, n0 = n−deg(v)+3, m0 = m−3deg(v)+8 and so, 5n0 −m0 = 5n−m−(2deg(v)−7) < 5n − m because deg(v) ≥ 4. Thus, from now on let us assume we fall in Step 3, i.e., a leaf-vertex of Type 2 or 3 is considered, denoted by v. • If the recursion happens at Step 3.1 (a), then G0 is obtained by removing v. Therefore, n0 = n − 1 and either m0 = m − 3 (if v is of Type 2) or m0 = m − 2 (if v is of Type 3), hence m0 ≥ m − 3 and so, 5n0 − m0 ≤ 5n − m − 2 < 5n − m. • If it is at Step 3.1 (b), then we fall in Step 3.1 (b) ii (no recursion occurs in Step 3.1 (b) i), thus G0 is obtained by making v adjacent to the two vertices in (N (av ) ∩ N (cv )) \ v (including bv in the case when v is of Type 3). Therefore, n0 = n and either m0 = m + 1 (if v is of Type 2) or m0 = m+2 (if v is of Type 3), hence m0 ≥ m+1 and so, 5n0 −m0 ≤ 5n−m−1 < 5n−m. • Else, the recursion happens at Step 3.2. Recall that in such case, there exists a vertex uv such that (bv , uv ) is a clique-separator of G \ v. Adding an edge between v and bv if it does not exist, decreases 5n − m by 1, therefore from now on let us assume that v is a leaf-vertex of Type 2. 47 – if it is at Step 3.2 (a), then G0 is obtained by contracting the edge {v, av }. Therefore, n0 = n − 1, m0 = m − 2, hence 5n0 − m0 = 5n − m − 3 < 5n − m. – If it is at Step 3.2 (b) i, then G0 is obtained by adding an edge between bv and some vertex x ∈ (N (av )∩N (uv ))\bv then contracting this edge. Furthermore, N (bv ) = {av , cv , uv , v} in such case and cv , v ∈ / N (x). Therefore, n0 = n − 1, m0 = m − 2 and so, 5n0 − m0 = 5n − m − 3 < 5n − m. – If it is at Step 3.2 (b) ii, then G0 is obtained by contracting the edge {bv , x} where x ∈ N (av ) ∩ N (uv ) ∩ N (bv ). Furthermore, N (bv ) = {av , cv , uv , v, x} in such case and cv , v ∈ / N (x). Therefore, n0 = n − 1, m0 = m − 3 and so, 5n0 − m0 = 5n − m − 2 < 5n − m. – Finally, in all other cases the recursive call happens at Step 3.2 (b) iii. Then, G0 is obtained by adding an edge between two vertices y, z ∈ N (bv ) ∩ N (x) for some x ∈ (N (av )∩N (uv ))\bv . Therefore, n0 = n, m0 = m+1 and so, 5n0 −m0 = 5n−m−1 < 5n−m. 6 Conclusion and Open questions On the negative side, we proved the NP-hardness of computing five metric graph invariants (namely, tree-breadth, path-length, path-breadth, k-good tree and path decompositions) whose complexity has been left open in several works [26, 29, 30]. These results add up to the proof in [42] that it is NP-hard to compute the tree-length. We leave as a future work further study on the border between tractable and intractable instances for the problem of computing the above metric graph invariants. Especially, what are the graph classes for which it can be decided in polynomial-time whether a graph admits a star-decomposition ? In this paper, we partially answer to this question by proving that it is the case for bipartite graphs and planar graphs. Based on these two positive results, we conjecture that the problem is FixedParameter Tractable when it is parameterized by the clique-number of the graph (note that there is a large clique in all the graphs obtained from our polynomial-time reductions). Intermediate challenges could be to determine whether the problem is Fixed-Parameter Tractable when it is parameterized by the genus, the tree-width or the Hardwiger number. Finally, we notice that all our NP-hardness results imply that the above metric graph invariants cannot be approximated below some constant-factor. There remains a gap between our inapproximability results and the constant-ratio of the approximation algorithms in [26, 30]. Therefore, we leave as an interesting open question whether we can fill in this gap. References [1] M. Abu-Ata and F. Dragan. Metric tree-like structures in real-world networks: an empirical study. Networks, pages n/a–n/a, 2015. [2] R. Agarwala, V. Bafna, M. Farach, M. Paterson, and M. Thorup. On the approximability of numerical taxonomy (fitting distances by tree metrics). SIAM Journal on Computing, 28(3):1073–1085, 1998. [3] D. Aingworth, C. Chekuri, P. Indyk, and R. Motwani. Fast estimation of diameter and shortest paths (without matrix multiplication). SIAM Journal on Computing, 28(4):1167–1181, 1999. [4] S. Arnborg, D. Corneil, and A. Proskurowski. Complexity of finding embeddings in ak-tree. SIAM Journal on Algebraic Discrete Methods, 8(2):277–284, 1987. [5] P. Austrin, T. Pitassi, and Y. Wu. Inapproximability of treewidth, one-shot pebbling, and related layout problems. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, pages 13–24. Springer, 2012. [6] R. Balakrishnan and K. Ranganathan. A textbook of graph theory. Springer Science & Business Media, 2012. [7] H.-J. Bandelt and V. Chepoi. Metric graph theory and geometry: a survey. Contemporary Mathematics, 453:49–86, 2008. 48 [8] M. BÇŐdoiu, K. Dhamdhere, A. Gupta, Y. Rabinovich, H. Räcke, R. Ravi, and A. Sidiropoulos. Approximation algorithms for low-distortion embeddings into low-dimensional spaces. In Proceedings of the sixteenth annual ACM-SIAM symposium on Discrete algorithms, pages 119–128. Society for Industrial and Applied Mathematics, 2005. [9] A. Berry, R. Pogorelcnik, and G. Simonet. An introduction to clique minimal separator decomposition. Algorithms, 3(2):197–215, 2010. [10] H. Bodlaender. A linear-time algorithm for finding tree-decompositions of small treewidth. SIAM J. Comput., 25(6):1305–1317, 1996. [11] H. Bodlaender. A partial k-arboretum of graphs with bounded treewidth. Theoretical computer science, 209(1):1–45, 1998. [12] H. Bodlaender. Discovering treewidth. In SOFSEM 2005: Theory and Practice of Computer Science, 31st Conference on Current Trends in Theory and Practice of Computer Science, Liptovský Ján, Slovakia, January 22-28, 2005, Proceedings, pages 1–16, 2005. [13] H. Bodlaender, M. Fellows, and T. Warnow. Two strikes against perfect phylogeny. In Automata, Languages and Programming, 19th International Colloquium, ICALP92, Vienna, Austria, July 13-17, 1992, Proceedings, pages 273–283, 1992. [14] H. Bodlaender and T. Kloks. Efficient and constructive algorithms for the pathwidth and treewidth of graphs. Journal of Algorithms, 21(2):358–402, 1996. [15] V. Bouchitté, F. Mazoit, and I. Todinca. Chordal embeddings of planar graphs. Discrete Mathematics, 273(1):85–102, 2003. [16] V. Bouchitté and I. Todinca. Listing all potential maximal cliques of a graph. Theoretical Computer Science, 276(1):17–32, 2002. [17] S. Chechik, D. Larkin, L. Roditty, G. Schoenebeck, R. Tarjan, and V. Vassilevska Williams. Better approximation algorithms for the graph diameter. In Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1041–1052. SIAM, 2014. [18] V. Chepoi. On distance-preserving and domination elimination orderings. SIAM Journal on Discrete Mathematics, 11(3):414–436, 1998. [19] V. Chepoi, F. Dragan, B. Estellon, M. Habib, and Y. Vaxès. Diameters, centers, and approximating trees of delta-hyperbolicgeodesic spaces and graphs. In Proceedings of the Twentyfourth Annual Symposium on Computational Geometry, SCG ’08, pages 59–68, New York, NY, USA, 2008. ACM. [20] M. Chrobak and D. Eppstein. Planar orientations with low out-degree and compaction of adjacency matrices. Theor. Comput. Sci., 86(2):243–266, 1991. [21] N. Clarke and R. Nowakowski. Tandem-win graphs. Discrete Mathematics, 299(1):56–64, 2005. [22] D. Coudert, G. Ducoffe, and N. Nisse. Diameter of Minimal Separators in Graphs. Research Report RR-8639, Inria Sophia Antipolis ; I3S ; INRIA, Nov. 2014. [23] E. Dahlhaus, P. Hammer, F. Maffray, and S. Olariu. On domination elimination orderings and domination graphs. In Graph-Theoretic Concepts in Computer Science, 20th International Workshop, WG ’94, Herrsching, Germany, June 16-18, 1994, Proceedings, pages 81–92, 1994. [24] R. Diestel. Graph Theory, 4th Edition, volume 173 of Graduate texts in mathematics. Springer, 2012. [25] Y. Dourisboure, F. Dragan, C. Gavoille, and Y. Chenyu. Spanners for bounded tree-length graphs. Theor. Comput. Sci., 383(1):34–44, 2007. [26] Y. Dourisboure and C. Gavoille. Tree-decompositions with bags of small diameter. Discrete Mathematics, 307(16):2008–2029, 2007. [27] R. Downey and M. Fellows. Parameterized complexity. Springer Science & Business Media, 2012. [28] F. Dragan. Tree-like structures in graphs: A metric point of view. In Graph-Theoretic Concepts in Computer Science - 39th International Workshop, WG 2013, Lübeck, Germany, June 19-21, 2013, Revised Papers, pages 1–4, 2013. [29] F. Dragan and E. Köhler. An approximation algorithm for the tree t-spanner problem on unweighted graphs via generalized chordal graphs. Algorithmica, 69(4):884–905, 2014. 49 [30] F. Dragan, E. Köhler, and A. Leitert. Line-distortion, bandwidth and path-length of a graph. In Algorithm Theory–SWAT 2014, pages 158–169. Springer, 2014. [31] F. Dragan and A. Leitert. On the minimum eccentricity shortest path problem. In Algorithms and Data Structures, pages 276–288. Springer, 2015. [32] F. Gavril. Algorithms for minimum coloring, maximum clique, minimum covering by cliques, and maximum independent set of a chordal graph. SIAM Journal on Computing, 1(2):180–187, 1972. [33] F. Gavril. The intersection graphs of subtrees in trees are exactly the chordal graphs. Journal of Combinatorial Theory, Series B, 16(1):47–56, 1974. [34] M. Golumbic, H. Kaplan, and R. Shamir. On the complexity of dna physical mapping. Advances in Applied Mathematics, 15(3):251–261, 1994. [35] M. Golumbic, H. Kaplan, and R. Shamir. Graph sandwich problems. Journal of Algorithms, 19(3):449–473, 1995. [36] M. Herlihy, F. Kuhn, S. Tirthapura, and R. Wattenhofer. Dynamic analysis of the arrow distributed protocol. Theory of Computing Systems, 39(6):875–901, 2006. [37] J. Hopcroft and R. Tarjan. Efficient planarity testing. Journal of the ACM (JACM), 21(4):549– 568, 1974. [38] P. Indyk. Algorithmic applications of low-distortion geometric embeddings. In Proceedings of the 42Nd IEEE Symposium on Foundations of Computer Science, FOCS ’01, pages 10–, Washington, DC, USA, 2001. IEEE Computer Society. [39] T. Kashiwabara and T. Fujisawa. Np-completeness of the problem of finding a minimumclique-number interval graph containing a given graph as a subgraph. In Proc. Symposium of Circuits and Systems, 1979. [40] A. Kosowski, B. Li, N. Nisse, and K. Suchan. k-Chordal Graphs: from Cops and Robber to Compact Routing via Treewidth. Algorithmica, 72(3):758–777, 2015. [41] R. Krauthgamer and J. Lee. Algorithms on negatively curved spaces. In Foundations of Computer Science, 2006. FOCS’06. 47th Annual IEEE Symposium on, pages 119–132. IEEE, 2006. [42] D. Lokshtanov. On the complexity of computing treelength. Discrete Applied Mathematics, 158(7):820–827, 2010. [43] B. Monien. The bandwidth minimization problem for caterpillars with hair length 3 is npcomplete. SIAM Journal on Algebraic Discrete Methods, 7(4):505–512, 1986. [44] J. Opatrny. Total ordering problem. SIAM Journal on Computing, 8(1):111–114, 1979. [45] C. Papadimitriou and M. Yannakakis. The clique problem for planar graphs. Inf. Process. Lett., 13(4/5):131–133, 1981. [46] N. Robertson and P. Seymour. Graph minors. ii. algorithmic aspects of tree-width. Journal of algorithms, 7(3):309–322, 1986. [47] M. Sysło. Characterizations of outerplanar graphs. Discrete Mathematics, 26(1):47–53, 1979. [48] J. Tenenbaum, V. De Silva, and J. Langford. A global geometric framework for nonlinear dimensionality reduction. Science, 290(5500):2319–2323, 2000. [49] K. Umezawa and K. Yamazaki. Tree-length equals branch-length. Discrete Mathematics, 309(13):4656–4660, 2009. 50
8cs.DS
Robustness analysis of a Maximum Correntropy framework for linear regression arXiv:1703.04829v2 [cs.SY] 31 Aug 2017 Laurent Bako a a Laboratoire Ampère – Ecole Centrale de Lyon – Université de Lyon, France Abstract In this paper we formulate a solution of the robust linear regression problem in a general framework of correntropy maximization. Our formulation yields a unified class of estimators which includes the Gaussian and Laplacian kernel-based correntropy estimators as special cases. An analysis of the robustness properties is then provided. The analysis includes a quantitative characterization of the informativity degree of the regression which is appropriate for studying the stability of the estimator. Using this tool, a sufficient condition is expressed under which the parametric estimation error is shown to be bounded. Explicit expression of the bound is given and discussion on its numerical computation is supplied. For illustration purpose, two special cases are numerically studied. Key words: robust estimation, system identification, maximum correntropy, outliers. 1 Introduction Given a set of empirical observations generated by a system along with a class of parameterized candidate models, a parameter estimator is a function which maps the available data to the parameter space associated with the model class. A very desirable property for an estimator is that of robustness which characterizes a relative insensitivity of the estimator to deviations of the observed data from the assumed model. More specifically, this property is central in situations where the data are prone to non Gaussian noise or disturbances of possibly arbitrarily large amplitude (often called outliers). The quest for robust estimators has led to the development of many estimators such as the Least Absolute Deviation (LAD) [17,14,4,2], the least median of squares [16], the least trimmed squares [17], the class of M-estimators [11]. Evaluating formally to what extent a given estimator is robust requires setting a quantitative measure of robustness. Incidentally such a measure can serve as comparison criterion between different robust estimators. Generally, the robustness is assessed in term of the maximum proportion of outliers in the total data set that the estimator can handle while remaining stable (see for example the concept of breakdown point [17]). More recently the maximum correntropy [18,12,15] has Email address: [email protected] (Laurent Bako). 1 This paper was not presented at any IFAC meeting. Corresponding author L. Bako. Tel.: +33 472 186 452. Preprint submitted to Automatica emerged as an information-theoretic estimation framework which induces some robustness properties with respect to outliers. Although maximum correntropy estimation is closely related to M-estimation, its discovery has broadened the horizon of possibilities for designing robust identification schemes. As a matter of fact, it has been successfully applied to a variety of estimation problems such as linear/nonlinear regression, filtering, face recognition in computer vision [9,7,10]. Contribution. Although the maximum correntropy based estimators have been gaining an increasing success, the formal analysis of its robustness properties is still a largely open research question. In this paper we propose such an analysis for a class of maximum correntropy based estimators applying to linear regression problems. More precisely, the contribution of the current paper is articulated around the following three questions: • To what extent the maximum correntropy estimation framework is robust to outliers? By robustness, it is meant here a certain insensitivity of the estimator to large errors of possibly arbitrarily large magnitude. To address this question, we derive parametric estimation error bounds induced by the estimator in function of both the degree of richness of the regression data and on the fraction of outliers. In summary, we show that if the regression data enjoy some richness properties and if the number of outliers is reasonably small, then the parametric estimation error remains stable. September 4, 2017 Indeed the proportion of outliers that the estimator is capable to correct depends on how rich the regressor matrix is. Moreover, the estimation error appears to be a decreasing function of the richness measure. • How does richness of the training data set influence the robustness of the estimator and how to characterize it? We provide an appropriate characterization of the richness in terms of the cardinality of the regressor vectors which are strongly correlated to any vector of the regression space. As such however, this quantitative measure of richness is not computable at an affordable price. To alleviate this difficulty the paper proposes some estimates of this measure thus allowing for the approximation of the parametric estimation error bounds. • Does the maximum correntropy estimator (MCE) possess the exact recovery property? We show that unlike the LAD estimator, the MCE is not able to return exactly the true parameter vector once the measurement is affected by a single arbitrary nonzero error. The proof is given for the Gaussian kernel based estimator. 2 2.1 Robust regression problem The data-generating system Let {xt }t∈N and {yt }t∈N be some stochastic processes taking values respectively in Rn and R. They are assumed to be related by an equation of the form o yt = x⊤ t θ + vt , (1) where {vt }t∈N represent an unobserved error sequence; θo ∈ Rn is an unknown parameter vector. Eq. (1) may describe a static (memoryless) system or a dynamic one. In the latter case, we will conveniently assume that the so-called regressor (or explanatory vector) xt has the following structure xt = [ut ut−1 · · · ut−(n−1) ]⊤ , i.e., (1) is an FIR-type (Finite Impulse Response) system, with ut then denoting its input signal at time t. Assumption 1 The joint stochastic process {(xt , vt )}t∈N is independently and identically distributed. We note that an analysis of robustness of the maximum correntropy has been presented recently in [5,6]. However the analysis there is limited to the Gaussian kernel based correntropy and to a single parameter estimation problem. Moreover these works do not make clear how the properties of the data contribute to the robustness of the estimator. While this assumption can hold naturally for a static system, it might not be satisfied in some practical situations. For example, if (1) is a dynamic system (for instance, of FIR-type), this assumption is not satisfied 2 . But as will be seen, its only role is to highlight the correntropic origin of the estimation framework considered in this paper. Outline. The rest of this paper is organized as follows. Section 2 presents the robust regression problem and define the class of maximum correntropy estimators whose properties are to be studied in the paper. It also introduces the general setting of the paper. The main analysis results are developed in Section 3. In Section 4 we run numerical experiments to illustrate the richness measure and the evolution of the derived error bounds with respect to the amount of noise. Finally, Section 5 contains concluding remarks concerning this work. Assumption 2 The noise sequence {vt } satisfies the following: there is ε ≥ 0 such that if we define the index sets Iε0 = {t : |vt | ≤ ε} and Iεc = {t : |vt | > ε}, then the cardinality of Iε0 is "much larger" than that of |Iεc |. We will formalize latter in the paper what "much larger" can mean. Similarly as in [2], we can assume that vt is of the form vt = ft + et where {ft } is a sparse noise sequence in the sense that only a few elements of it are different from zero. However its nonzero elements are allowed to take on arbitrarily large values (called in this case, outliers). As to {et }, it is assumed to be a bounded and dense (i.e., not necessarily sparse) noise sequence of rather moderate amplitude. Notations. R is the set of real numbers; R+ is the set of real nonnegative numbers; N is the set of natural integers; C denotes the set of complex numbers. N will denote the number of data points and I = {1, . . . , N } the associated index set. For any finite set S, |S| refers to the cardinality of S. However, whenever x is a real (respectively complex) number, |x| will refer to the absolute value (respectively modulus) of x. For x = [x1 · · · xn ]⊤ ∈ Rn , kxkp will denote the p-norm N Problem. Given a finite collection Z N = {(xt , yt )}t=1 of measurements obeying the system equation (1), the robust regression problem of interest here is the one of finding a reliable estimate of the parameter vector θo despite the effect of arbitrarily large errors. of x defined by kxkp = (|x1 |p + · · · + |xn |p )1/p , for p ∈ {1, 2}, kxk∞ = maxi=1,...,n |xi |. The exponential of a real number z will be denoted exp(z) or ez according to visual convenience; ln(z) is the natural logarithm function. For a square and positive semi-definite matrix A, λmin (A) and λmax (A) denote respectively the minimal and maximal eigenvalues of A. Let θ denote a candidate parameter vector (PV) which we would like, ideally, to coincide with the true PV θo . Given xt and θ, the prediction we can make of yt is 2 Indeed this assumption can be relaxed to an appropriate notion of stationarity and ergodicity for the joint process {(xt , vt )} 2 ŷt (θ) = x⊤ t θ. It is then the goal of the estimation method to select θ such that yt and ŷt (θ) are close in some sense for any t. Closeness will be be measured in term of the so-called maximum correntropy between the measured output yt and the predicted value ŷt (θ). 2.2 PROOF. That ℓp satisfies P1-P3 is an obvious fact. As to Property P4, it follows from convexity. In effect the convexity of ℓp implies that for all (a, b) ∈ R2 , p |a + b| /2p = ℓp ((a + b)/2) ≤ 1/2ℓp (a) + 1/2ℓp(b). Multiplying by 2 gives 1/2p−1ℓp (a+b) ≤ ℓp (a)+ℓp (b), which by replacing a with a − b can be seen to be equivalent to P4 with αℓ = 1/2p−1 . ✷ Maximum correntropy estimation The correntropy is an information-theoretic measure of similarity between two arbitrary random variables [18,12]. More specifically, consider two random variables Y and Ŷ defined on the same probability space, and taking values in R. Let φℓ : R×R → R be a positive-definite kernel function on R (see e.g., [19, Chap. 2, p. 30] for a definition). The correntropy Vφℓ (Y, Ŷ ) between Y and Ŷ with respect to a kernel function φℓ , is defined by The correntropy maximization is an estimation framework where one tries to maximize the correntropy. In the regression problem stated above, we aim to find the parameter vector θ that maximizes 3 Vφℓ (yt , ŷt (θ)), the correntropy between yt and ŷt (θ) with respect to the kernel φℓ . In practice however the distribution 4 pxt ,yt is generally unknown so that one cannot evaluate the exact correntropy. As a consequence of this difficulty one would be content in practice with maximizing a sample estimate of the correntropy. Assume that we are given a N set Z N = {(xt , yt )}t=1 of data points sampled independently from the joint distribution pxt ,yt . Then in virtue of Assumption 1, an estimate of the correntropy is given by   Vφℓ (Y, Ŷ ) = EY,Ŷ φℓ (Y, Ŷ ) , where EY,Ŷ [·] refers to the expected value with respect to the joint distribution of (Y, Ŷ ). In a more explicit form, we have Z Z φℓ (y, ŷ)pY,Ŷ (y, ŷ)dydŷ (2) Vφℓ (Y, Ŷ ) = R N 1 X b φℓ (yk , ŷk (θ)) Vφℓ (yt , ŷt (θ)) = N k=1 R N   1 X exp −γℓ(yk − x⊤ = k θ) N with pY,Ŷ being the joint probability density function of (Y, Ŷ ). The correntropy constitutes a similarity measure between Y and Ŷ through the kernel φℓ . Although the original definition of correntropy in [18] fixes φℓ to be the Gaussian kernel, it is indeed possible to extend it to any positive definite kernel function. k=1 for all t ∈ I , {1, . . . , N }. Hence the maximum correntropy estimator (MCE) studied in this paper is the possibly set-valued map Ψmce : (Rn × R)N → Rn which maps the data to a parameter space, We consider in this paper a kernel function of the form φℓ (y, ŷ) = exp(−γℓ(y − ŷ)), (4) Ψmce (Z N ) = arg max Vbφℓ (yt , ŷt (θ)). (5) θ∈Rn (3) In the form (4)-(5) the MCE can be viewed as a particular instance of the prediction error estimation scheme  [13, Chap. 7] with prediction error measured by exp −γℓ(yk − x⊤ k θ) . Also, the performance index (4) is reminiscent of the risk-sensitive estimation cost which is used in control, adaptive filtering and parameter estimation [3,8]. But this latter approach, which roughly consists in the minimization of a sum of exponential of positive error terms, is not suitable for handling the effects of impulsive noise such as outliers. Although the focus of this paper is the analysis of the properties of the estimator (5), let us mention in passing that the underlying optimization problem in (5) is non convex. This implies that solving (5) numerically can be where γ > 0 is a user-specified parameter and ℓ : R → R+ is a function which satisfies the following properties: P1. ℓ is positive-definite: ℓ(a) ≥ 0 ∀a and ℓ(a) = 0 if and only if a = 0. P2. ℓ is symmetric: ℓ(−a) = ℓ(a). P3. ℓ is nondecreasing on R+ : ℓ(a) ≤ ℓ(b) whenever |a| ≤ |b|. P4. There exists αℓ > 0 such that ℓ(a − b) ≥ αℓ ℓ(a) − ℓ(b) ∀(a, b) ∈ R2 . Property P4 can be interpreted as a relaxed version of the triangle inequality property for ℓ. We can characterize a family of functions ℓ satisfying P1-P4 as follows. 3 By Assumption 1, Vφℓ (yt , ŷt (θ)) is indeed constant i.e., independent of t. Hence t refers here to an arbitrary time index. 4 To be precise, the interest is in pyt ,ŷt (θ) but this follows from pxt ,yt . Lemma 3 (Examples of functions obeying P1-P4) For any real number p ≥ 1, the function ℓp : R → R+ p defined by ℓp (x) = |x| satisfies the properties P1-P4. In particular, P4 is satisfied with αℓ = 1/2p−1. 3 challenging. However it can be interpreted iteratively as a weighted least squares problem in the case for example where φℓ is taken to be the Gaussian kernel. We will get back to this in Section 4. 3 We will refer to it as the correlation measure of the matrix X at the level α. It appears intuitively that ρα (X) is a decreasing function of α. Clearly, we get ρα (X) = 0 for α = 1 for finite N while ρα (X) = 1 for α = 0. For a given matrix, it would be interesting to be able to evaluate numerically the quantitative measure ρα (X) of richness. Indeed, this value will be required for numerical assessment of the error bound to be derived in Section 3.2. However computing exactly the value of ρα (X) is a hard combinatorial problem. Robustness properties of the MCE As discussed in the introduction, an estimator of the form (5) is intuitively thought (and empirically shown) to be endowed with some robustness properties. By this, we mean that it is able to keep behaving reasonably well when a certain fraction of the available data points are affected by noise components vt of possibly arbitrarily large magnitude. The question of main interest in this paper is to characterize quantitatively up to what extent the estimator defined in (5) can be insensitive to outliers. 3.1 We therefore discuss how to reach estimates of ρα (X) at an affordable cost. To this end, let X̃ = [x̃1 · · · x̃N ] ∈ Rn×N be the matrix obtained from X by normalizing its columns to unit 2-norm, i.e., x̃t = xt / kxt k2 for all t. Then introduce the number n o (9) σ(X) = minn X̃ ⊤ η ∞ s. t. kηk2 = 1 Data informativity η∈R As will be seen, the robustness property is inherited from both the structure of the estimator and the richness of the regression data. We are therefore interested in formalizing as well that richness and how it contributes to the robustness properties of the estimator. which is solely a function of the matrix X, hence the notation. Note that the so-defined σ(X) lies necessarily in the real interval [0, 1]. Moreover, it can be usefully √ 1/2 1/2 observed that σ(X) ≥ λmin (X̃ X̃ ⊤ )/ N , with λmin (·) referring to the square root of the minimum eigenvalue. Now for any t ∈ I consider the following index set o n p (10) 1 − δ2 Jt,α (X) = k ∈ I : |x̃⊤ k x̃t | ≥ To proceed with the analysis, let us introduce some notations. For convenience we make the following assumption. Assumption 4 The regressor sequence {xt } satisfies: xt 6= 0 for all t ∈ I. p √ where δ = 1 − α2 − 1 − σ(X)2 . It is assumed in the definition (10) that σ(X) ≥ α so that δ ≥ 0. For a given t, Jt,α (X) collects the indices of the regressors which are the most correlated to x̃t in the sense that the √ cosine of the angle they form with x̃t is larger than 1 − δ 2 . Finally, let Note that Assumption 4 is without loss of generality. Under this assumption, let us pose rx = min kxt k2 > 0, t∈I (6) with k·k2 denoting Euclidean norm. Upon dividing the system equation (1) by kxt k2 , we can even assume that rx = 1. Let α ∈ [0, 1] be a real number. For any η ∈ Rn , define the index set  Iα (X, η) = t ∈ I : |x⊤ (7) t η| ≥ α kxt k2 kηk2 vα (X) = 1 inf |Jt,α (X)| N t∈I (11) be the ratio between the minimum cardinality of the finite set Jt,α (X) over all t living in I and the number N of columns in X. Then we can estimate ρα (X) as follows. Proposition 5 Let X ∈ Rn×N be a real matrix. Then, for all α ∈ ]0, 1] with α ≤ σ(X), with X = [x1 · · · xN ] ∈ Rn×N a matrix formed with all the regressors. Finally, let ρα (X) be the ratio between the minimum cardinality that Iα (X, η) can attain over all possible values of η, and the number N of columns in X, i.e., 1 |Iα (X, η)| . (8) ρα (X) = infn η∈R N The number ρα (X) measures somehow the richness (or informativity/genericity) of the regression data. Intuitively, ρα (X) reflects a dense spanning of all directions of the vector space Rn by the vectors {xt }. For a given α > 0, it is desired that ρα (X) be as large as possible.  λ (X̃ X̃ ⊤ )  min vα (X) ≤ ρα (X) ≤ min 1, N α2 (12) with λmin (X̃ X̃ ⊤ ) denoting the minimum eigenvalue of the matrix X̃ X̃ ⊤ . The proof of this proposition uses the following lemma. Lemma 6 [20, Thm 5.14] Let x, y, z ∈ Cn be such that 4 x∗ x = y ∗ y = z ∗ z = 1 with x∗ denoting the conjugate transpose of x. Then inf η∈Rn p p p 1 − |x∗ y|2 ≤ 1 − |x∗ z|2 + 1 − |z ∗ y|2 . The key benefit of Proposition 5 is that it provides a method for estimating the measure ρα (X) defined in (8) at an affordable cost. Note however that while the upper bound in (12) can be computed easily, obtaining the lower bound vα (X) is still challenging. The reason is that this bound involves the number σ(X) in (9) whose numerical evaluation requires solving a nonconvex optimization problem. Nevertheless, it can be approximated through some heuristics, e.g. by solving a sequence of linear programs. Equality holds if and only if there exists β ∈ C such that either z = βx or z = βy. PROOF. [Proof of Proposition 5] The upper bound is immediate. To see this, let η0 6= 0 be the eigenvector associated with the smallest eigenvalue of X̃ X̃ ⊤ . Then 2 λmin (X̃ X̃ ⊤ ) kη0 k2 = ≥ N X t=1 Remark 7 In comparison to the classical concept of persistence of excitation (PE) in system identification, the richness property requiring that ρα (X) be large is a stronger property. In finite time, the quantitative persistence of excitation (called specifically sufficiency of excitation in this case) asks for the condition number 1/2 1/2 λmax (XX ⊤ )/λmin (XX ⊤ ) of XX ⊤ to be as close to 1 as possible. The PE condition appears to be a global property of the matrix X while the richness condition introduced here is a somewhat local property as it is basically counting the number of vectors xt pointing in any direction of the regression space. 2 (x̃⊤ t η0 ) X 2 (x̃⊤ t η0 ) 1 |Iα (X, η)| = ρα (X). ✷ N . t∈Iα (X,η0 ) ≥ |Iα (X, η0 )| α2 kη0 k22 It follows that ρα (X) = 1/N inf η∈Rn |Iα (X, η)| ≤ 1/N |Iα (X, η0 )| ≤ λmin (X̃ X̃ ⊤ )/(N α2 ). The upper inequality in (12) follows by additionally taking into consideration the obvious fact that ρα (X) ≤ 1. 3.2 We now prove the inequality vα (X) ≤ ρα (X). To begin with, note from (9) that for any η ∈ Rn satisfying kηk2 = 1, there exists t(η) ∈ I such that |η ⊤ x̃t(η) | ≥ σ(X). Consider an index k ∈ I, such that |x̃⊤ k x̃t(η) | ≥ h for some h ∈ [0, 1]. Then observe that |x̃⊤ η| ≥ α is equivalent to k Main results Equipped with the measure of informativity introduced above, we can now state the main result of this paper, which stands as follows. q p 2 1 − (x̃⊤ 1 − α2 . k η) ≤ Theorem 8 Let Iε0 = {t ∈ I : |vt | ≤ ε} and Iεc = {t ∈ I : |vt | > ε} with {vt } denoting the noise sequence in (1). Let ℓ be a function obeying P1-P4. Assume that the following condition is satisfied for some α ∈ ]0, 1], q q q 2 2 ≤ ⊤ x̃ 2+ 1 − (x̃⊤ 1 − (η 1 − (x̃⊤ η) ) t(η) k x̃t(η) ) k p p ≤ 1 − σ(X)2 + 1 − h2 . 0 1 −γℓ(ε) |Iε | > 1. (13) ρ (X) + e α N 1 + e−γℓ(ε)  Then for any θ⋆ ∈ Ψmce Z N with Z N being generated by system (1), it holds that On the other hand, by applying Lemma 6, we can write It follows that for |x̃⊤ k η| ≥ α to hold, it is sufficient that ℓ(αrx kθ⋆ − θo k2 ) ≤ p p p 1 − σ(X)2 + 1 − h2 ≤ 1 − α2 , √ which is equivalent to h ≥ 1 − δ 2 with √ in turn p 2 2 δ = 1 − α − 1 − σ(X) ∈ [0, 1] by the assumption that σ(X) ≥ α. Hence, for |x̃⊤ k η| to be greater √ than 1 − δ2. or equal to α, it is enough that |x̃⊤ x̃ | ≥ k t(η) This means that for a given t, k being in the index set Jt(η),α (X) defined in (10) is a sufficient condition n for |x̃⊤ k η| ≥ α for all η ∈ R such that kηk2 = 1. Therefore Jt(η),α (X) ⊂ Iα (X, η) hence implying that Jt(η),α (X) ≤ |Iα (X, η)|. Taking now the in1 Jt(η),α (X) ≤ fimum produces vα (X) ≤ inf η∈Rn N 1 ln(1/µℓ ), γαℓ (14) where µℓ = 1 + e−γℓ(ε) |Iε0 | N  1 ρ (X) −γℓ(ε) α 1 + e + ρα (X) − 1  0 −γℓ(ε) |Iε | −1 +e N (15) If in addition, ℓ is strictly increasing on R+ , then 1 −1 ℓ kθ − θ k2 ≤ αrx ⋆ 5 o   1 ln(1/µℓ ) . γαℓ (16) observations into the above inequality yields o  n 0   Iε exp − γℓ(ε) − |Iεc | exp − γℓ(ε)   ≤ Iε0 ∩ Iα (X, η ⋆ ) exp − γαℓ ℓ(αrx kη ⋆ k2 ) PROOF. Let  θ⋆ ∈ Ψmce Z N = arg max Vbφℓ (yt , ŷt (θ)). θ∈Rn Then for any θ ∈ Rn , it holds that N X t=1 + Iε0 ∩ Iαc (X, η ⋆ ) . N  X    ⋆ exp − γℓ(yt − x⊤ θ) ≤ exp − γℓ(yt − x⊤ t t θ ) By observing that Iε0 ∩ Iαc (X, η ⋆ ) = |Iε0 |− Iε0 ∩ Iα0 (X, η ⋆ ) , we can rearrange the above inequality in the form   e−γℓ(ε) Iε0 e−γℓ(ε) + Iε0 − N − Iε0 n o   ≤ Iε0 ∩ Iα (X, η ⋆ ) exp − γαℓ ℓ(αrx kη ⋆ k2 ) − 1 (17) Now by exploiting the definition of ρα (X), we can observe that t=1 Taking in particular θ = θo and invoking the system equation (1), it follows that X t∈Iε0   X   exp − γℓ(vt ) + exp − γℓ(vt ) t∈Iεc ≤ X t∈Iε0 +   ⋆ exp − γℓ(vt − x⊤ t η ) X t∈Iεc  exp − γℓ(vt − |Iε0 ∩ Iα (X, η ⋆ )| = |Iε0 | + |Iα (X, η ⋆ )| − |Iε0 ∪ Iα (X, η ⋆ )| ≥ |Iε0 | + N ρα (X) − N.  ⋆ x⊤ t η ) Moreover since where we have posed η ⋆ = θ⋆ − θo . This implies that X t∈Iε0 N ρα (X) + |Iε0 | ≥     X ⋆ c exp − γℓ(vt − x⊤ exp − γℓ(vt ) ≤ t η ) + |Iε | t∈Iε0 the assumption (13) guarantees that |Iε0 | + N ρα (X) − N > 0. Therefore since the term on the right hand side of (17) is negative, it holds that With |vt | ≤ ε for any t ∈ Iε0 , we have −γℓ(vt ) ≥ −γℓ(ε) by the symmetry of ℓ.As  properties P Pand nondecreasing a consequence, t∈I 0 exp − γℓ(vt ) ≥ t∈I 0 exp − ε ε  γℓ(ε) . On the other hand, by the fourth property of ⋆ ⊤ ⋆ the function ℓ, ℓ(vt − x⊤ t η ) ≥ αℓ ℓ(xt η ) − ℓ(vt ) ≥ ⊤ ⋆ ⋆ αℓ ℓ(xt η ) − ℓ(ε) hence implying that −γℓ(vt − x⊤ t η ) ≤ ⊤ ⋆ −γαℓ ℓ(xt η ) + γℓ(ε). Combining these observations allows us to write X   exp − γℓ(ε) − |Iεc | Then direct algebraic calculations lead to   µℓ ≤ exp − γαℓ ℓ(αrx kη ⋆ k2 ) ≤ 1 where µℓ is defined as in (15). Indeed, in virtue of the assumption (13), µℓ is positive. Hence we have    X ⋆ exp − γαℓ ℓ(x⊤ ≤ exp γℓ(ε) t η )  t∈Iε0 X t∈Iε0 ∩Iα (X,η ⋆ )  + exp γℓ(ε) X   ⋆ exp − γαℓ ℓ(x⊤ t η ) t∈Iε0 ∩Iαc (X,η ⋆ ) n o Iε0 e−γℓ(ε) + Iε0 − N − Iε0 o   n exp − γαℓ ℓ(αrx kη ⋆ k2 ) − 1 ≤ |Iε0 | + N ρα (X) − N e−γℓ(ε) t∈Iε0   = exp γℓ(ε) 1 N ρα (X) + e−γℓ(ε) |Iε0 |, 1 + e−γℓ(ε) ℓ(αrx kη ⋆ k2 ) ≤   ⋆ exp − γαℓ ℓ(x⊤ t η ) 1 ln(1/µℓ ). γαℓ Of course, if ℓ is monotonically increasing on R+ then it is invertible and the error bound in (16) follows. ✷ In the last equality we have partitioned the set Iε0 into Iε0 ∩Iα (X, η ⋆ ) and Iε0 ∩Iαc (X, η ⋆ ) with Iαc (X, η ⋆ ) being the complement of Iα (X, η ⋆ ) in I. Note from (7) that ⋆ ⋆ for all t ∈ Iε0 ∩ Iα (X, η ⋆ ), ℓ(x⊤ t η ) ≥ ℓ(αrx kη k2 ) so ⊤ ⋆ ⋆ that −γαℓ ℓ(xt η ) ≤ −γαℓ ℓ(αrx kη k2 ). Plugging these A few comments follow from this result. A key assumption of the theorem is condition (13). What it requires is on the one hand, that the proportion of outliers be somehow small and on the other hand, that the regression data X be rich in the sense that ρα (X) be large enough for a given nonzero α ∈ ]0, 1]. An important teaching of this condition is that the richer the data matrix X, the larger the number of outliers that can be corrected by the estimator. We can interpret (13) as a sufficient con- 6 dition for the stability of the estimator since it guarantees a bounded estimation error. We now discuss some special instances of Theorem 8 corresponding to two kernels which are frequently used for estimation. For convenience of the discussion, let us introduce the following notation. Let µ : R3+ → R+ A second comment concerns the amplitude of the error bound given in (16). For the purpose of making this bound small, we need the constant µℓ to be close to one. Again we see that this is favored by a small number of outliers and a rich data set. An interesting special case is when ℓ(ε) = 0, which occurs when the data are only affected by some outliers (ε = 0) and no dense noise. In this case the number µℓ defined in (15) reduces to µℓ = 1 − |Iε0 | N 1− µ(z, ε, α) = N whenever e−z 0 otherwise. |Iε0 | N which tend to suggest, since µℓ 6= 1, that no exact recovery might be achieved once the data are affected by a single outlier unless we consider in (16) the limit case when γ → +∞. A similar observation was made in [6] in a comparable context. We will prove below that the MCE does not possess the exact recovery property, at least in the case when ℓ(a) = a2 . In contrast, a robust estimator such as the LAD estimator (see, e.g., [1,2]) is able to achieve exact recovery under a relatively significant proportion of nonzero errors. 3.3 φ1 (y, ŷ) = exp (−γ1 |y − ŷ|) . Corollary 11 (Laplacian kernel) Let Iε0 be defined as in Theorem 8. Assume that the following condition is satisfied 0 1 −γ1 ε |Iε | ρ (X) + e >1 α −γ ε 1+e 1 N  − θ )) . Note that where wt (θ) = exp − γℓ(vt − the matrix on the left hand side of the equation above is finite regardless of the value of θ. Hence, in the event that θo ∈ Ψmce (Z N ), we would have t∈Iεc X (18) It is straightforward to see that the properties P1-P4 are satisfied by ℓ with αℓ = 1. Theorem 8 can be specialized to this case as follows. t∈Iεc wt (θo )vt xt = Laplacian kernel The Maximum Laplacian Correntropy estimator (MCEL) corresponds to the case where the function ℓ in (3) is taken to be such that ℓ(a) = |a|. As a result, the function φℓ takes the form PROOF. We start by observing that with ℓ being the square function, the cost V̂φℓ (yt , ŷt (θ)) is differentiable. Therefore, a necessary condition for θ to be in Ψmce (Z N ) is that ∇V̂φℓ (yt , ŷt (θ)) = 0, where ∇ refers to the gradient. This, by using the system equation (1) and exploiting the assumption that vt = 0 for t ∈ Iε0 , can be translated into ! N X X ⊤ wt (θ)xt xt (θ − θo ) = wt (θ)vt xt , X + 1+e1−z ρα (X)− 1 > 0 and µ(z, ε, α) = subject to α ≤ σ(X) and condition (13). Although such a minimum might not be easy to compute exactly, one can make the error bound a little tighter by performing for example some grid search. In the same manner one can envision optimizing the parameter γ of the estimator. Proposition 9 Let Assumption 4 hold and assume that for any t ∈ Iε0 , vt = 0. Take the function ℓ in (4) to be the square function, ℓ(a) = a2 . For all ε > 0, if |Iεc | ≥ 1 then there exists a sequence {vt } such that θo ∈ / Ψmce (Z N ) N when Z is generated from (1) under the action of {vt }. x⊤ t (θ |Iε0 | N Remark 10 The bound (16) allows for some degree of freedom in the choice of the parameter α. For a given function ℓ assumed to be invertible on R+ , a better bound can, in principle, be obtained as    1 −1 1 1 min ℓ ln γαℓ µ(γℓ(ε), ε, α) α∈]0,1] αrx + ρα (X) − 1 t=1 h i 0 1 −z |Iε | ρ (X)+e −1 α −z N + ρα (X) − 1 1 + e 1 + e−z |Iε0 | o (19) for some α ∈ ]0, 1]. Then for any θ⋆ ∈ arg maxθ∈Rn Vbφ1 (yt , ŷt (θ)) with φ1 defined as in (18), it holds that λt xt = 0 kθ⋆ − θo k2 ≤ t∈Iεc with λt = exp(−γℓ(vt ))vt . Clearly, it is possible to find a nonzero sequence {vt } which does not meet this condition. Hence θo cannot be in Ψmce (Z N ) for an arbitrary {vt : t ∈ Iεc } no matter how small the cardinality of Iεc is. ✷ 3.4   1 1 ln γ1 αrx µ(γ1 ε, ε, α) (20) Gaussian kernel The most used form of correntropy is the one based on the Gaussian kernel which, by omitting the normalizing 7 which can be written in the form (1) with θo = [0.5 −1 0.2]⊤ and xt = [ut ut−1 ut−2 ]⊤ . For the data- factor, can be written in the form φ2 (y, ŷ) = exp(−γ2 (y − ŷ)2 ) (21) iid generation experiment, assume that {ut } ∼ N (0, 1) i.e., {ut } is sampled independently and identically from a zero-mean Gaussian distribution of unit variance. As for the noise signal {vt }, it is defined as vt = et + ft with γ2 > 0. We will refer to the associated estimator as the maximum Gaussian correntropy estimator (MCEG). Here, the function ℓ is defined by ℓ(a) = a2 and according to Lemma 3, it satisfies the properties P1-P4. In particular, P4 is satisfied with αℓ = 1/2. Moreover ℓ is clearly monotonic on R+ . As a consequence, we get a corollary of Theorem 8 as follows. iid with {et } ∼ U([−ε, ε]) where U refers to the uniform distribution and {ft } is a sequence of sparse noise with only a few nonzero elements (which are otherwise not constrained in magnitude); the nonzero elements of {ft } are here sampled from N (50, 10). Corollary 12 (Gaussian kernel) Let Iε0 be defined as in Theorem 8. Assume that the following condition is satisfied 0 1 −γ2 ε2 |Iε | >1 2 ρα (X) + e −γ ε N 1+e 2 4.1 We generate N = 300 data pairs (xt , yt ) and carry out a comparison between three estimators: on the one hand, the maximum Laplacian correntropy estimator (MCEL) and the maximum Gaussian correntropy estimator (MCE-G) and on the other hand, the Least Absolute Deviation (LAD) estimator (which is also called ℓ1 estimator). Recall that MCE-G and MCE-L involve non convex optimization. Here they are heuristically implemented as a reweighted iterative least squares estimator and as a reweighted ℓ1 estimator respectively. The results are represented in Figure 1 in term of average estimation error. What this suggests is that for fixed values of the design parameters γ1 and γ2 (see Eqs (18) and (21) for the roles of these parameters), LAD and MCE-L enjoy a similar performance for small amount of noise. But as the noise level increases, LAD shows better stability capabilities than the MCE-L. Note that overall MCE-G tends to perform best in the setting of this experiment as long as the magnitude of the dense noise is reasonable (SNR larger than 2.8 dB). A possible justification for this is that squaring errors that contain outliers as in (21) cancel out their influence more forcefully than just taking their absolute value as in (18). (22) for some α ∈ ]0, 1]. Then for any θ⋆ ∈ arg maxθ∈Rn Vbφ2 (yt , ŷt (θ)), it holds that  1/2  2 1 1 . ln kθ − θ k2 ≤ αrx γ2 µ(γ2 ε2 , ε, α) ⋆ 3.5 o (23) A remark on the error-in-variables scenario We now consider the situation where only a noisy observation x̄t = xt + wt of the regressor vector xt in (1) is available for prediction. This scenario is referred to as the robust error-in-variable (EIV) regression problem. Then the predictor output is given by ŷt (θ) = x̄⊤ t θ. Indeed Theorem 8 remains valid for this case. To see this note that the system equation (1) can be rewritten as 4.2 where v̄t = vt − wt⊤ θo . Then clearly the theorem applies to the EIV scenario with {xt } and {vt }, replaced respectively by {x̄t } and {v̄t }. One limitation however in this case is that for a given ε ≥ 0, the cardinality of the set Iε0 = {t ∈ I : |v̄t | ≤ ε} is likely to be much smaller than in the situation where the regressors are noise-free. Numerical experiments The purpose of this section is to provide a numerical illustration of the richness measure (8) and of the estimation error bound (16). The system example considered for the experiment is of an FIR-type and is given by yt = 0.5ut − ut−1 + 0.2ut−2 + vt Estimation of richness measure ρα (X) We provide a graphical representation of how the informativity measure ρα (X) may, for a given data matrix X ∈ Rn×N , evolve with respect to the dimensions N/n of X and the demanded degree α of richness (See Figure 2). The estimated range for ρα (X) is based on Eq. (12). Here X is formed from an FIR-type of regressors with an input sampled from a zero-mean and unit variance Gaussian distribution. Our experiments in this specific study tend to suggest that ρα (X) is a non decreasing function of the ratio N/n and a decreasing function of α. Moreover, the estimated range (gray regions in Fig. 2) gets wider when n is large. o yt = x̄⊤ t θ + v̄t 4 Illustration of estimates 4.3 Estimates of error bounds The goal here is twofold: (i) illustrate the variation of the estimation error bounds with respect to the magnitude of the dense noise in the special cases (20) and (23); (24) 8 increasing rate of the bound corresponding to MCE-L is larger than that of MCE-G for the current setting. Note that the computation of bounds made here is not connected to the experiment of Section 4.1. placements 0.6 LAD 0.5 MCE-L MCE-G 2.5 0.4 0.3 Bound MCE-L Bound MCE-G 2 0.2 Bounds Error 0.7 0.1 1.5 1 0 0 0.2 0.4 0.6 0.8 1 1.2 ε 1.4 1.6 PSfrag replacements 0.5 Bounds Figure 1. Comparison of the estimators MCE-L, MCE-G and LAD. Evolution of the (average) estimation error in function of the level ε of noise. The range [0, 1.6] of ε corresponds indeed to a range of about [2, +∞] for the signal-to-noise ratio in this experiment. The results are obtained from a Monte-Carlo simulation of size 1000. For each experiment, the proportion of outliers is maintained fixed and equal to 50%. Design parameters: γ1 = 0.5 and γ2 = 0.25. 0 0 0.05 0.1 0.15 0.2 ε Figure 3. Illustration of increasing rates of the error bounds. Results obtained by assuming that γℓ(ε) = 0.2, |Iε0 |/N = 0.8, ρα (X) = 0.8, α = 0.6 in both cases. These values are chosen independently of a specific data matrix X but in such a way that (19) and (22) are satisfied. 1 Comparing theoretical bounds and empirical errors. It might be instructive to see how far away the theoretical error bounds may be from the empirical values. To study this aspect, let us consider a numerical experiment with a similar data-generating process as described in the beginning of Section 4. The dense noise level is set to ε = 0.05 which gives an SNR of about 25 dB and the proportion of outliers is set to 10% (which is small enough to enforce condition (13)). One difficulty in evaluating the theoretical bounds is that this requires evaluating ρα (X) which, as already discussed in Section 3.1, is a hard problem. Hence, ρα (X) is replaced here with the mean value of the lower and upper estimates displayed in (12). We then let the number N of data vary from 500 to 5000 and plot the empirical errors along with the bounds from (20) and (23) in Figure 4. ρα (X) 0.8 g replacements 0.6 0.4 0.2 0 0.2 0.4 α 0.6 0.8 (a) X ∈ R2×200 1 ρα (X) 0.8 g replacements 0.6 0.4 It is fair to observe that the theoretical bounds are conservative in the sense that they are generally higher than the true empirical errors. Here the ratio between the bounds and the true errors is about 30. Conservativeness is indeed a common feature for these types of results due to the various inequalities employed for the derivation. Nevertheless, the main interest of Theorem 8 is that it provides a sufficient condition for the robustness of the maximum correntropy estimator, a condition that depends explicitly on the degree of informativity of the regression data and on the proportion of outliers. Moreover, by expressing error bounds which involve explicitly the design parameters, the theorem gives insights into how to tune those parameters with the aim to improve estimation performance. A further remark one can make is that the general formula for the error bound in (16) has a kind of universal 0.2 0 0.2 0.4 α 0.6 0.8 (b) X ∈ R3×6000 Figure 2. Estimates of ρα (X) as a function of α ∈ [0, 1]. The true values of ρα (X) lie in the region depicted in gray color. (ii) assess how conservative the derived theoretical error bounds may be with respect to the empirical errors. Increasing rates of the bounds. If for each level ε of noise, we select the parameter γ such that the product γℓ(ε) is kept constant, then the error bounds corresponding to both MCE-L and MCE-G have a linear rate of change with respect to ε as depicted in Figure 3. The 9 [3] R. K. Boel, M. R. James, and I. R. Petersen. Robustness and risk-sensitive filtering. IEEE Transactions on Automatic Control, 47:451–461, 2002. 0 log10 (kθ ⋆ − θ o k2 ) -0.5 MCE-L MCE-G Bound MCE-L Bound MCE-G -1 -1.5 placements [4] E. Candès and P. A. Randall. Highly robust error correction by convex programming. IEEE Transactions on Information Theory, 54:2829–2840, 2006. [5] B. Chen, X. Liu, H. Zhao, N. Zheng, and J. Principe. Insights into the robustness of minimum error entropy estimation. IEEE Transactions on Neural Networks and Learning Systems (To appear), 2016. -2 [6] B. Chen, L.Xing, H. Zhao, B. Xu, and J. C. Principe. Robustness of maximum correntropy estimation against large outliers. https://arxiv.org/abs/1703.08065, 2017. -2.5 [7] B. Chen, L. Xing, J. Liang, N. Zheng, and J. C. Principe. Steady-state mean-square error analysis for adaptive filtering under the maximum correntropy criterion. IEEE Signal Processing Letters, 21:880–884, 2014. -3 0 1000 2000 3000 4000 5000 N [8] S. Dey and J. B. Moore. Risk-sensitive filtering and smoothing via reference probability methods. IEEE Transactions on Automatic Control, 42:1587–1591, 1997. Figure 4. Comparison of theoretical bounds and empirical errors kθ⋆ − θo k2 . Results obtained with an estimate of ρα (X) for α = 0.6, ε = 0.05 (i.e., SNR = 25 dB), γℓ(ε) = 0.2, |Iε0 |/N = 0.9. The results are obtained from a Monte-Carlo simulation of size 1000. [9] Y. Feng, X. Huang, L. Shi, Y. Yang, and J. A.K. Suykens. Learning with the maximum correntropy criterion induced losses for regression. Journal of Machine Learning Research, 16:993–1034, 2015. feature in the following sense: since the bound does not involve the magnitude of the true θo (for an FIR-type system for example), it is in principle valid regardless of θo . Hence the relative error will be as smaller as the norm of the to-be-estimated parameter vector θo is larger. [10] R. He, W.-S. Zheng, and B.-G. Hu. Maximum correntropy criterion for robust face recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33:1561–1576, 2011. [11] P. J. Huber and E. M. Ronchetti. Robust Statistics. A. John Wiley & Sons, Inc. Publication (2nd Ed), 2009. 5 Conclusion [12] W Liu, P. P. Pokharel, and J. C. Principe. Correntropy: Properties and aplications in non-gaussian signal processing. IEEE Transactions on Signal Processing, 55(11):5286–5298, 2007. In this paper we have proposed an analysis of the robustness properties of a correntropy maximization framework for regression problems. The class of estimators considered is quite general and include the Gaussian and Laplacian kernels as special cases. The contribution of the work consists in (i) deriving an appropriate notion of richness for the regression data; (ii) proving stability of the considered class of estimators under the derived richness condition when the data are subject to dense and sparse noise (outliers). Our main result states that if the regression data are rich enough and if the number of outliers is small in some sense, then the parametric estimation error is bounded. The results come with explicit bounds which, in default of being exactly computable, can be estimated with computable estimates. [13] L. Ljung. System Identification: Theory for the user (2nd Ed.). PTR Prentice Hall., Upper Saddle River, USA, 1999. [14] R. A. Maronna, R. D. Martin, and V. J. Yohai. Robust Statistics: Theory and Methods. John Wiley & Sons, Inc., 2006. [15] J. C. Principe, editor. Information Theoretic Learning: Renyi’s Entropy and Kernel Perspectives. New York: Springer Verlag, 2010. [16] P. J. Rousseeuw. Least median of squares regression. Journal of the American Statistical Association, 79:871–880, 1984. [17] P. J. Rousseeuw and A. M. Leroy. Robust Regression and Outlier Detection. John Wiley & Sons, Inc., 2005. [18] I. Santamaria, P. P. Pokharel, and J. C. Principe. Generalized correlation function: definition, properties, and application to blind equalization. IEEE Transactions on Signal Processing, 54:2187–2197, 2006. Acknowledgements [19] B. Schölkopf and A. J. Smola. Learning with Kernels Support Vector Machines, Regularization, Optimization and Beyond. MIT Press, Cambridge, MA, 2002. The author is grateful to the Associate Editor and the anonymous reviewers for constructive feedback. [20] F. Zhang. Matrix Theory: Basic Results and Techniques. Springer: New York, 1999. References [1] L. Bako. On a class of optimization-based robust estimators. IEEE Transactions on Automatic Control (To appear), 2017. [2] L. Bako and H. Ohlsson. Analysis of a nonsmooth optimization approach to robust estimation. Automatica, 66:132–145, 2016. 10
3cs.SY
Generating Natural Language Inference Chains Vladyslav Kolesnyk and Tim Rocktäschel and Sebastian Riedel University College London London, UK arXiv:1606.01404v1 [cs.CL] 4 Jun 2016 [email protected], {t.rocktaschel,s.riedel}@cs.ucl.ac.uk Abstract The ability to reason with natural language is a fundamental prerequisite for many NLP tasks such as information extraction, machine translation and question answering. To quantify this ability, systems are commonly tested whether they can recognize textual entailment, i.e., whether one sentence can be inferred from another one. However, in most NLP applications only single source sentences instead of sentence pairs are available. Hence, we propose a new task that measures how well a model can generate an entailed sentence from a source sentence. We take entailment-pairs of the Stanford Natural Language Inference corpus and train an LSTM with attention. On a manually annotated test set we found that 82% of generated sentences are correct, an improvement of 10.3% over an LSTM baseline. A qualitative analysis shows that this model is not only capable of shortening input sentences, but also inferring new statements via paraphrasing and phrase entailment. We then apply this model recursively to input-output pairs, thereby generating natural language inference chains that can be used to automatically construct an entailment graph from source sentences. Finally, by swapping source and target sentences we can also train a model that given an input sentence invents additional information to generate a new sentence. 1 Introduction The ability to determine entailment or contradiction between natural language text is essential for improving the performance in a wide range of natural language processing tasks. Recognizing Textual Entailment (RTE) is a task primarily designed to determine whether two natural language sentences are independent, contradictory or in an entailment relationship where the second sentence (the hypothesis) can be inferred from the first (the premise). Although systems that perform well in RTE could potentially be used to improve question answering, information extraction, text summarization and machine translation [6], only in few of such downstream NLP tasks sentence-pairs are actually available. Usually, only a single source sentence (e.g. a question that needs to be answered or a source sentence that we want to translate) is present and models need to come up with their own hypotheses and commonsense knowledge inferences. The release of the large Stanford Natural Language Inference (SNLI) corpus [2] allowed end-to-end differentiable neural networks to outperform feature-based classifiers on the RTE task [3, 4, 12, 15, 17]. In this work, we go a step further and investigate how well recurrent neural networks can produce true hypotheses given a source sentence. Furthermore, we qualitatively demonstrate that by only training on input-output pairs and recursively generating entailed sentence we can generate natural language inference chains (see Figure 1 for an example). Note that every inference step is interpretable as it is mapping one natural language sentence to another one. Our contributions are fourfold: (i) we propose Young blond woman putting her foot into a water fountain → A person is dipping her foot into water. → A person is wet. Figure 1: Example of a natural language inference chain that is generated from the first sentence. an entailment generation task based on the SNLI corpus (§2.1), (ii) we investigate a sequence-tosequence model and find that 82% of generated sentences are correct (§3.2), (iii) we demonstrate the ability to generate natural language inference chains trained solely from entailment pairs (§3.3), and finally (iv) we can also generate sentences with more specific information by swapping source and target sentences during training (§3.4). 2 Method In the section, we briefly introduce the entailment generation task and our sequence-to-sequence model. 2.1 Entailment Generation To create the entailment generation dataset, we simply filter the Stanford Natural Language Inference corpus for sentence-pairs of the entailment class. This results in a training set of 183, 416 sentence pairs, a development set of 3, 329 pairs and a test of 3, 368 pairs. Instead of a classification task, we can now use this dataset for a sequence transduction task. 2.2 Sequence-to-Sequence Sequence-to-sequence recurrent neural networks [14] have been successfully employed for many sequence transduction tasks in NLP such as machine translation [1, 5], constituency parsing [16], sentence summarization [13] and question answering [7]. They consist of two recurrent neural networks (RNNs): an encoder that maps an input sequence of words into a dense vector representation, and a decoder that conditioned on that vector representation generates an output sequence. Specifically, we use long short-term memory (LSTM) RNNs [8] for encoding and decoding. Furthermore, we experiment with wordby-word attention [1], which allows the decoder to search in the encoder outputs to circumvent the LSTM’s memory bottleneck. We use greedy decoding at test time. The success of LSTMs with attention in sequence transduction tasks makes them a natural choice as a baseline for entailment generation, and we leave the investigation of more advanced models to future work. 2.3 Optimization and Hyperparameters We use stochastic gradient descent with a minibatch size of 64 and the ADAM optimizer [9] with a first momentum coefficient of 0.9 and a second momentum coefficient of 0.999. Word embeddings are initialized with pre-trained word2vec vectors [10]. Out-of-vocabulary words (10.5%) are randomly initialized √ √ by sampling values uniformly from [− 3, 3] and optimized during training. Furthermore, we clip gradients using a norm of 5.0. We stop training after 25 epochs. 3 Experiments and Results We present results for various tasks: (i) given a premise, generate a sentence that can be inferred from the premise, (ii) construct inference chains by recursively generating sentences, and (iii) given a sentence, create a premise that would entail this sentence, i.e., make a more descriptive sentence by adding specific information. 3.1 Quantitative Evaluation We train an LSTM with and without attention on the training set. After training, we take the best model in terms of BLEU score [11] on the development set and calculate the BLEU score on the test set. To our surprise, we found that using attention yields only a marginally higher BLEU score (43.1 vs. 42.8). We suspect that this is due to the fact that generating entailed sentences has a larger space of valid target sequences, which makes the use of BLEU problematic and penalizes correct solutions. Hence, we manually annotated 100 random test sentences and decided whether the generated sentence can indeed be inferred from the source sentence. We found that sentences generated by an LSTM with attention are substantially more accurate (82% accuracy) than those generated from an LSTM baseline (71.7%). To gain more insights into the model’s capabilities, we turn to a thorough qualitative 1. He is driving a beautiful red car. → A man is driving a car. 2. Two men and a woman finishing a meal and drinks. → Three people are eating together. 3. A father holds a new born baby. → A man holds his baby. Figure 2: Valid sentences generated by the model. 1. A female gymnast in black and red being coached on bar skills. → A female is in a bar. 2. A dog drops a yellow disc on a beach. → The dog drops a red disc. 1. First time a computer Go program has beaten a 9-dan professional without handicaps. → A person is using a computer. 2. Competitors must be dressed in suitable ten nis attire that is almost entirely white. → The tennis player is wearing white. 3. The athletes run by the harbour during the Haspa Hamburg marathon. → The runners are outside. 4. Space travellers gaze up and strain to find the blue dot in their sky. → There are people looking at the sky. Figure 4: Generated sentences from external sources: (1) Wikipedia, (2) wimbledon.com, (3) a news article, (4) ”The Frontier is Everywhere” by Carl Sagan. Figure 3: Invalid sentences generated by the model. analysis of the attention LSTM model in the remainder of this paper. 3.2 Example Generations Figure 2 shows examples of generated sentences from the development set. Syntactic simplification of the input sentence seems to be the most common approach. The model removes certain parts of the premise such as adjectives, resulting in a more abstract sentence (see Figure 2.1). Figure 2.2 demonstrates that the system can recognize the number of subjects in the sentence and includes this information in the generated sentence. However, we did not observe such ’counting’ behavior for more than four subjects, indicating that the system memorized frequency patterns from the training set. Furthermore, we found predictions that hint to common-sense assumptions: if a sentence talks about a father holding a newborn baby, it is most likely that the newborn baby is his own child (Example 2.3). Limitations Two reappearing limitations of the proposed model are related to dealing with words that have a very different meaning but similar word2vec embeddings (e.g. colors), as well as ambiguous words. For instance, ’bar’ in Figure 3.1 refers to pole vault and not a place in which you can have a drink. Substituting one color by another one (Figure 3.2) is a common mistake. Out-of-Corpus Examples The SNLI corpus might not reflect the variety of sentences that can be encountered in downstream NLP tasks. In Figure 4 we present generated sentences for randomly selected examples of out-of-domain textual resources. They demonstrate that the model generalizes well to out-of-domain sentences, making it a potentially very useful component for improving systems for question answering, information extraction, sentence summarization etc. 3.3 Inference Chain Generation Next, we test how well the model can generate inference chains by repeatedly passing generated output sentences as inputs to the model. We stop once a sentence has already been generated in the chain. Figure 5 shows that this works well despite that the model was only trained on sentence-pairs. Furthermore, by generating inference chains for all sentences in the development set we construct an entailment graph. In that graph we found that sentences with shared semantics are eventually mapped to the same sentence that captures the shared meaning. 1. Two girls sit by a lake. → Two girls are sitting on a ledge in front of a lake. 1. People are celebrating a victory on the square. → People are celebrating. → People are happy. → The people are smil ing. ← People are smiling. ← Some people smiling. ← couple smiling ← A bride and groom smiles. ← A wedding party looks happy on the picture. 2. The athletes run at the start of the Haspa Hamburg marathon. → The runners are at the race. → There are runners at the race. → There are people running. → People are running. ← People are running outside. ← People are running away from the bear. Figure 5: Examples of inference chains where two premises (underlined) converge to the same sentence (highlighted). Figure 6: Topology of the generated entailment graph. A visualization of the topology of the entailment graph is shown in Figure 6. Note that there are several long inference chains, as well as large clusters of sentences (nodes) that are mapped (links) to the same shared meaning. 3.4 Inverse Inference By swapping the source and target sequences for training, we can train a model that given a sentence invents additional information to generate a new sentence (Figure 7). We believe this might prove useful to increase the language variety and 2. Two girls are playing chess → Two girls playing chess in a restaurant. 3. A woman is outside. → A woman is walking down a street in a city. → A woman in a black skirt walks down a city street. Figure 7: Generated sentences for the inverse task. complexity of AI unit tests such as the Facebook bAbI task [18], but we leave this for future work. 4 Conclusion and Future Work We investigated the ability of sequence-tosequence models to generate entailed sentences from a source sentence. To this end, we trained an attentive LSTM on entailment-pairs of the SNLI corpus. We found that this works well and generalizes beyond in-domain sentences. Hence, it could become a useful component for improving the performance of other NLP systems. We were able to generate natural language inference chains by recursively generating sentences from previously inferred ones. This allowed us to construct an entailment graph for sentences of the SNLI development corpus. In this graph, the shared meaning of two related sentences is represented by the first natural language sentence that connects both sentences. Every inference step is interpretable as it maps a natural language sentence to another one. Towards high-quality data augmentation, we experimented with reversing the generation task. We found that this enabled the model to learn to invent specific information. For future work, we want to integrate the presented model into larger architectures to improve the performance of downstream NLP tasks such as information extraction and question answering. Furthermore, we plan to use the model for data augmentation to train expressive neural networks on tasks where only little annotated data is available. Another interesting research direction is to investigate methods for increasing the diversity of the generated sentences. Acknowledgments We thank Guillaume Bouchard for suggesting the reversed generation task, and Dirk Weissenborn, Isabelle Augenstein and Matko Bosnjak for comments on drafts of this paper. This work was supported by Microsoft Research through its PhD Scholarship Programme, an Allen Distinguished Investigator Award, and a Marie Curie Career Integration Award. References [1] D. Bahdanau, K. Cho, and Y. Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2015. [2] S. R. Bowman, G. Angeli, C. Potts, and C. D. Manning. A large annotated corpus for learning natural language inference. In EMNLP, 2015. [3] S. R. Bowman, J. Gauthier, A. Rastogi, R. Gupta, C. D. Manning, and C. Potts. A fast unified model for parsing and sentence understanding. In ACL, 2016. [4] J. Cheng, L. Dong, and M. Lapata. Long short-term memory-networks for machine reading. arXiv preprint arXiv:1601.06733, 2016. [5] K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y. Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014. [6] I. Dagan, O. Glickman, and B. Magnini. The pascal recognising textual entailment challenge. In Machine learning challenges. evaluating predictive uncertainty, visual object classification, and recognising tectual entailment, pages 177–190. Springer, 2006. [7] K. M. Hermann, T. Kocisky, E. Grefenstette, L. Espeholt, W. Kay, M. Suleyman, and P. Blunsom. Teaching machines to read and comprehend. In NIPS, pages 1684–1692, 2015. [8] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural computation, 9 (8):1735–1780, 1997. [9] D. Kingma and J. Ba. Adam: A method for stochastic optimization. In ICLR, 2015. [10] T. Mikolov, I. Sutskever, K. Chen, G. S. Corrado, and J. Dean. Distributed representations of words and phrases and their compositionality. In NIPS, pages 3111–3119, 2013. [11] K. Papineni, S. Roukos, T. Ward, and W.-J. Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311–318, 2002. [12] T. Rocktäschel, E. Grefenstette, K. M. Hermann, T. Kočiskỳ, and P. Blunsom. Reasoning about entailment with neural attention. In ICLR, 2016. [13] A. M. Rush, S. Chopra, and J. Weston. A neural attention model for abstractive sentence summarization. In EMNLP, 2015. [14] I. Sutskever, O. Vinyals, and Q. V. Le. Sequence to sequence learning with neural networks. In NIPS, pages 3104–3112, 2014. [15] I. Vendrov, R. Kiros, S. Fidler, and R. Urtasun. Order-embeddings of images and language. In ICLR, 2016. [16] O. Vinyals, L. Kaiser, T. Koo, S. Petrov, I. Sutskever, and G. Hinton. Grammar as a foreign language. In NIPS, pages 2755–2763, 2015. [17] S. Wang and J. Jiang. Learning Natural Language Inference with LSTM. In NAACL, 2016. [18] J. Weston, A. Bordes, S. Chopra, and T. Mikolov. Towards ai-complete question answering: A set of prerequisite toy tasks. In ICLR, 2016.
9cs.NE
Approximating LZ77 via Small-Space Multiple-Pattern Matching Johannes Fischer∗1 , Travis Gagie2 , Pawel Gawrychowski†3 , and Tomasz Kociumaka‡3 arXiv:1504.06647v2 [cs.DS] 10 Sep 2015 1 TU Dortmund, Germany [email protected] 2 Helsinki Institute for Information Technology (HIIT), Department of Computer Science, University of Helsinki, Finland [email protected] 3 Institute of Informatics, University of Warsaw, Poland {gawry,kociumaka}@mimuw.edu.pl March 27, 2018 Abstract We generalize Karp-Rabin string matching to handle multiple patterns in O(n log n + m) time and O(s) space, where n is the length of the text and m is the total length of the s patterns, returning correct answers with high probability. As a prime application of our algorithm, we show how to approximate the LZ77 parse of a string of length n. If the optimal parse consists of z phrases, using only O(z) working space we can return a parse consisting of at most (1+ε)z phrases in O(ε−1 n log n) time, for any ε ∈ (0, 1]. As previous quasilinear-time algorithms for LZ77 use Ω(n/poly log n) space, but z can be exponentially small in n, these improvements in space are substantial. 1 Introduction Multiple-pattern matching, the task of locating the occurrences of s patterns of total length m in a single text of length n, is a fundamental problem in the field of string algorithms. The algorithm by Aho and Corasick [2] solves this problem using O(n+m) time and O(m) working space in addition to the space needed for the text and patterns. To list all occ occurrences rather than, e.g., the leftmost ones, extra O(occ) time is necessary. When the space is limited, we can use a compressed Aho-Corasick automaton [11]. In extreme cases, one could apply a linear-time constant-space single-pattern matching algorithm sequentially for each pattern in turn, at the cost of increasing the running time to O(n · s + m). Well-known examples of such algorithms include those by Galil and Seiferas [8], Crochemore and Perrin [5], and Karp and Rabin [13] (see [3] for a recent survey). It is easy to generalize Karp-Rabin matching to handle multiple patterns in O(n + m) expected time and O(s) working space provided that all patterns are of the same length [10]. To do this, we store the fingerprints of the patterns in a hash table, and then slide a window over the text maintaining the fingerprint of the fragment currently in the window. The hash table lets us check if the fragment is an occurrence of a pattern. If so, we report it and update the hash table so that every pattern is returned at most once. This is a very ∗ Supported by Academy of Finland grant 268324. done while the author held a post-doctoral position at Warsaw Center of Mathematics and Computer Science. ‡ Supported by Polish budget funds for science in 2013-2017 as a research project under the ‘Diamond Grant’ program. † Work 1 simple and actually applied idea [1], but it is not clear how to extend it for patterns with many distinct lengths. In this paper we develop a dictionary matching algorithm which works for any set of patterns in O(n log n + m) time and O(s) working space, assuming that read-only random access to the text and the patterns is available. If required, we can compute for every pattern its longest prefix occurring in the text, also in O(n log n + m) time and O(s) working space. In a very recent independent work Clifford et al. [4] gave a dictionary matching algorithm in the streaming model. In this setting the patterns and later the text are scanned once only (as opposed to read-only random access) and an occurrence needs to be reported immediately after its last character is read. Their algorithm uses O(s log `) space and takes O(log log(s + `)) time per character where ` is the length of the longest pattern ( m s ≤ ` ≤ m). Even though some of the ideas used in both results are similar, one should note that the streaming and read-only models are quite different. In particular, computing the longest prefix occurring in the text for every pattern requires Ω(m log min(n, |Σ|)) bits of space in the streaming model, as opposed to the O(s) working space achieved by our solution in the read-only setting. As a prime application of our dictionary matching algorithm, we show how to approximate the LempelZiv 77 (LZ77) parse [18] of a text of length n using working space proportional to the number of phrases (again, we assume read-only random access to the text). Computing the LZ77 parse in small space is an issue of high importance, with space being a frequent bottleneck of today’s systems. Moreover, LZ77 is useful not only for data compression, but also as a way to speed up algorithms [15]. We present a general approximation algorithm working in O(z) space for inputs admitting LZ77 parsing with z phrases. For any ε ∈ (0, 1], the algorithm can be used to produce a parse consisting of (1 + ε)z phrases in O(ε−1 n log n) time. To the best of our knowledge, approximating LZ77 factorization in small space has not been considered before, and our algorithm is significantly more efficient than methods producing the exact answer. A recent sublinear-space algorithm, due to Kärkkäinen et al. [12], runs in O(nd) time and uses O(n/d) space, for any parameter d. An earlier online solution by Gasieniec et al. [9] uses O(z) space and takes O(z 2 log2 z) time for each character appended. Other previous methods use significantly more space when the parse is small relative to n; see [7] for a recent discussion. Structure of the paper. Sect. 2 introduces terminology and recalls several known concepts. This is followed by the description of our dictionary matching algorithm. In Sect. 3 we show how to process patterns of length at most s and in Sect. 4 we handle longer patterns, with different procedures for repetitive and non-repetitive ones. In Sect. 5 we extend the algorithm to compute, for every pattern, the longest prefix occurring in the text. Finally, in Sect. 7, we apply the dictionary matching algorithm to construct an approximation of the LZ77 parsing, and in Sect. 6 we explain how to modify the algorithms to make them Las Vegas. Model of computation. Our algorithms are designed for the word-RAM with Ω(log n)-bit words and assume integer alphabet of polynomial size. The usage of Karp-Rabin fingerprints makes them Monte Carlo randomized: the correct answer is returned with high probability, i.e., the error probability is inverse polynomial with respect to input size, where the degree of the polynomial can be set arbitrarily large. With some additional effort, our algorithms can be turned into Las Vegas randomized, where the answer is always correct and the time bounds hold with high probability. Throughout the whole paper, we assume readonly random access to the text and the patterns, and we do not include their sizes while measuring space consumption. 2 Preliminaries We consider finite words over an integer alphabet Σ = {0, . . . , σ − 1}, where σ = poly(n + m). For a word w = w[1] . . . w[n] ∈ Σn , we define the length of w as |w| = n. For 1 ≤ i ≤ j ≤ n, a word u = w[i] . . . w[j] is called a subword of w. By w[i..j] we denote the occurrence of u at position i, called a fragment of w. A fragment with i = 1 is called a prefix and a fragment with j = n is called a suffix. 2 A positive integer p is called a period of w whenever w[i] = w[i + p] for all i = 1, 2, . . . , |w| − p. In this case, the prefix w[1..p] is often also called a period of w. The length of the shortest period of a word w is |w| denoted as per(w). A word w is called periodic if per(w) ≤ |w| 2 and highly periodic if per(w) ≤ 3 . The well-known periodicity lemma [6] says that if p and q are both periods of w, and p + q ≤ |w|, then gcd(p, q) is also a period of w. We say that word w is primitive if per(w) is not a proper divisor of |w|. Note that the shortest period w[1.. per(w)] is always primitive. 2.1 Fingerprints Our randomized construction is based on Karp-Rabin fingerprints; see [13]. Fix a word w[1..n] over an alphabet Σ = {0, . . . , σ−1}, a constant c ≥ 1, a prime number p > max(σ, nc+4 ), and choose x ∈ Zp uniformly at random. We define the fingerprint of a subword w[i..j] as Φ(w[i..j]) = w[i]+w[i+1]x+. . .+w[j]xj−i mod p. With probability at least 1 − n1c , no two distinct subwords of the same length have equal fingerprints. The situation when this happens for some two subwords is called a false-positive. From now on when stating the results we assume that there are no false-positives to avoid repeating that the answers are correct with high probability. For dictionary matching, we assume that no two distinct subwords of w = T P1 . . . Ps have equal fingerprints. Fingerprints let us easily locate many patterns of the same length. A straightforward solution described in the introduction builds a hash table mapping fingerprints to patterns. However, then we can only guarantee that the hash table is constructed correctly with probability 1 − O( s1c ) (for an arbitrary 1 constant c), and we would like to bound the error probability by O( (n+m) c ). Hence we replace hash table with a deterministic dictionary as explained below. Although it increases the time by O(s log s), the extra term becomes absorbed in the final complexities. Theorem 1. Given a text T of length n and patterns P1 , . . . , Ps , each of length exactly `, we can compute the the leftmost occurrence of every pattern Pi in T using O(n + s` + s log s) total time and O(s) space. Proof. We calculate the fingerprint Φ(Pj ) of every pattern. Then we build in O(s log s) time [16] a deterministic dictionary D with an entry mapping Φ(Pj ) to j. For multiple identical patterns we create just one entry, and at the end we copy the answers to all instances of the pattern. Then we scan the text T with a sliding window of length ` while maintaining the fingerprint Φ(T [i..i + ` − 1]) of the current window. Using D, we can find in O(1) time an index j such that Φ(T [i..i + ` − 1]) = Φ(Pj ), if any, and update the answer for Pj if needed (i.e., if there was no occurrence of Pj before). If we precompute x−1 , the fingerprints Φ(T [i..i + ` − 1]) can be updated in O(1) time while increasing i. 2.2 Tries A trie of a collection of strings P1 , . . . , Ps is a rooted tree whose nodes correspond to prefixes of the strings. The root represents the empty word and the edges are labeled with single characters. The node corresponding to a particular prefix is called its locus. In a compacted trie unary nodes that do not represent any Pi are dissolved and the labels of their incidents edges are concatenated. The dissolved nodes are called implicit as opposed to the explicit nodes, which remain stored. The locus of a string in a compacted trie might therefore be explicit or implicit. All edges outgoing from the same node are stored on a list sorted according to the first character, which is unique among these edges. The labels of edges of a compacted trie are stored as pointers to the respective fragments of strings Pi . Consequently, a compacted trie can be stored in space proportional to the number of explicit nodes, which is O(s). Consider two compacted tries T1 and T2 . We say that (possibly implicit) nodes v1 ∈ T1 and v2 ∈ T2 are twins if they are loci of the same string. Note that every v1 ∈ T1 has at most one twin v2 ∈ T2 . Lemma 2. Given two compacted tries T1 and T2 constructed for s1 and s2 strings, respectively, in O(s1 +s2 ) total time and space we can find for each explicit node v1 ∈ T1 a node v2 ∈ T2 such that if v1 has a twin in T2 , then v2 is its twin. (If v1 has no twin in T2 , the algorithm returns an arbitrary node v2 ∈ T2 ). 3 Proof. We recursively traverse both tries while maintaining a pair of nodes v1 ∈ T1 and v2 ∈ T2 , starting with the root of T1 and T2 satisfying the following invariant: either v1 and v2 are twins, or v1 has no twin in T2 . If v1 is explicit, we store v2 as the candidate for its twin. Next, we list the (possibly implicit) children of v1 and v2 and match them according to the edge labels with a linear scan. We recurse on all pairs of matched children. If both v1 and v2 are implicit, we simply advance to their immediate children. The last step is repeated until we reach an explicit node in at least one of the tries, so we keep it implicit in the implementation to make sure that the total number of operations is O(s1 + s2 ). If a node v ∈ T1 is not visited during the traversal, for sure it has no twin in T2 . Otherwise, we compute a single candidate for its twin. 3 Short Patterns To handle the patterns of length not exceeding a given threshold `, we first build a compacted trie for those patterns. Construction is easy if the patterns are sorted lexicographically: we insert them one by one into the compacted trie first naively traversing the trie from the root, then potentially partitioning one edge into two parts, and finally adding a leaf if necessary. Thus, the following result suffices to efficiently build the tries. Lemma 3. One can lexicographically sort strings P1 , . . . , Ps of total length m in O(m + σ ε ) time using O(s) space, for any constant ε > 0. √ Proof. We separately sort the m + σ ε/2 longest strings and all the remaining strings, and then merge both sorted lists. Note these longest strings can be found in O(s) time using a linear time selection algorithm. Long strings are sorted using insertion sort. If the longest common prefixes between adjacent (in the sorted order) strings are computed and stored, inserting Pj can be done in O(j + |Pj |) time. In more detail, let S1 , S2 , . . . , Sj−1 be the sorted list of already processed strings. We start with k := 1 and keep increasing k by one as long as Sk is lexicographically smaller than Pj while maintaining the longest common prefix between Sk and Pj , denoted `. After increasing k by one, we update ` using the longest common prefix between Sk−1 and Sk , denoted `0 , as follows. If `0 > `, we keep ` unchanged. If `0 = `, we try to iteratively increase ` by one as long as possible. In both cases, the new value of ` allows us to lexicographically compare Sk and Pj in√constant time. Finally, `0 < ` guarantees that Pj < Sk and √ we may terminate the procedure. Sorting the m + σ ε/2 longest strings using this√approach takes O(m + ( m + σ ε/2 )2 ) = O(m + σ ε ) time. The remaining strings are of length at most m each, and if there are any, then s ≥ σ ε/2 . We sort these strings by iteratively applying radix sort, treating each symbol from Σ as a sequence of 2ε symbols from {0, 1, . . . , σ ε/2 − 1}. Then a single radix sort takes time and space proportional to the number of strings involved plus the alphabet size, which is O(s + σ ε/2 ) = O(s). Furthermore, because the numbers of strings √ involved in the subsequent radix sorts sum up to m, the total time complexity is O(m+σ ε/2 m) = O(m+σ ε ). Finally, the merging takes time linear in the sum of the lengths of all the involved strings, so the total complexity is as claimed. Next, we partition T into O( n` ) overlapping blocks T1 = T [1..2`], T2 = T [` + 1..3`], T3 = T [2` + 1..4`], . . .. Notice that each subword of length at most ` is completely contained in some block. Thus, we can consider every block separately. The suffix tree of each block Ti takes O(` log `) time [17] and O(`) space to construct and store (the suffix tree is discarded after processing the block). We apply Lemma 2 to the suffix tree and the compacted trie of patterns; this takes O(` + s) time. For each pattern Pj we obtain a node such that the corresponding subword is equal to Pj provided that Pj occurs in Ti . We compute the leftmost occurrence Ti [b..e] of the subword, which takes constant time if we store additional data at every explicit node of the suffix tree, and then we check whether Ti [b..e] = Pj using fingerprints. For this, we precompute the fingerprints of all patterns, and for each block Ti we precompute the fingerprints of its prefixes in O(`) time and space, which allows to determine the fingerprint of any of its subwords in constant time. 4 In total, we spend O(m + σ ε ) for preprocessing and O(` log ` + s) for each block. Since σ = (n + m)O(1) , for small enough ε this yields the following result. Theorem 4. Given a text T of length n and patterns P1 , . . . , Ps of total length m, using O(n log ` + s n` + m) total time and O(s + `) space we can compute the leftmost occurrences in T of every pattern Pj of length at most `. 4 Long Patterns To handle patterns longer than a certain threshold, we first distribute them into groups according to the value of blog4/3 |Pj |c. Patterns longer than the text can be ignored, so there are O(log n) groups. Each group is handled separately, and from now on we consider only patterns Pj satisfying blog4/3 |Pj |c = i. We classify the patterns into classes depending on the periodicity of their prefixes and suffixes. We set ` = d(4/3)i e and define αj and βj as, respectively, the prefix and the suffix of length ` of Pj . Since 4 2 3 (|αj | + |βj |) = 3 ` ≥ |Pj |, the following fact yields a classification of the patterns into three classes: either Pj is highly periodic, or αj is not highly periodic, or βj is not highly periodic. The intuition behind this classification is that if the prefix or the suffix is not repetitive, then we will not see it many times in a short subword of the text. On the other hand, if both the prefix and suffix are repetitive, then there is some structure that we can take advantage of. Fact 5. Suppose x and y are a prefix and a suffix of a word w, respectively. If |x| + |y| ≥ |w| + p and p is a period of both x and y, then p is a period of w. Proof. We need to prove that w[i] = w[i + p] for all i = 1, 2, . . . , |w| − p. If i + p ≤ |x| this follows from p being a period of x, and if i ≥ |w| − |y| + 1 from p being a period of y. Because |x| + |y| ≥ |w| + p, these two cases cover all possible values of i. To assign every pattern to the appropriate class, we compute the periods of Pj , αj and βj using small space. Roughly the same result has been proved in [14], but for completeness we provide the full proof here. Lemma 6. Given a read-only string w one can decide in O(|w|) time and constant space if w is periodic and if so, compute per(w). Proof. Let v be the prefix of w of length d 12 |w|e and p be the starting position of the second occurrence of v in w, if any. We claim that if per(w) ≤ 12 |w|, then per(w) = p − 1. Observe first that in this case v occurs at a position per(w) + 1. Hence, per(w) ≥ p − 1. Moreover p − 1 is a period of w[1..|v| + p − 1] along with per(w). By the periodicity lemma, per(w) ≤ 21 |w| ≤ |v| implies that gcd(p − 1, per(w)) is also a period of that prefix. Thus per(w) > p − 1 would contradict the primitivity of w[1.. per(w)]. The algorithm computes the position p using a linear time constant-space pattern matching algorithm. If it exists, it uses letter-by-letter comparison to determine whether w[1..p − 1] is a period of w. If so, by the discussion above per(w) = p − 1 and the algorithm returns this value. Otherwise, 2 per(w) > |w|, i.e., w is not periodic. The algorithm runs in linear time and uses constant space. 4.1 Patterns without Long Highly Periodic Prefix Below we show how to deal with patterns with non-highly periodic prefixes αj . Patterns with non-highly periodic suffixes βj can be processed using the same method after reversing the text and the patterns. Lemma 7. Let ` be an arbitrary integer. Suppose we are given a text T of length n and patterns P1 , . . . , Ps such that for 1 ≤ j ≤ s we have ` ≤ |Pj | < 43 ` and αj = Pj [1..`] is not highly periodic. We can compute the leftmost and the rightmost occurrence of each pattern Pj in T using O(n + s(1 + n` ) log s + s`) time and O(s) space. 5 The algorithm scans the text T with a sliding window of length `. Whenever it encounters a subword equal to the prefix αj of some Pj , it creates a request to verify whether the corresponding suffix βj of length ` occurs at the appropriate position. The request is processed when the sliding window reaches that position. This way the algorithm detects the occurrences of all the patterns. In particular, we may store the leftmost and rightmost occurrence of each pattern. We use the fingerprints to compare the subwords of T with αj and βj . To this end, we precompute Φ(αj ) and Φ(βj ) for each j. We also build a deterministic dictionary D [16] with an entry mapping Φ(αj ) to j for every pattern (if there are multiple patterns with the same value of Φ(αj ), the dictionary maps a fingerprint to a list of indices). These steps take O(s`) and O(s log s), respectively. Pending requests are maintained in a priority queue Q, implemented using a binary heap1 as pairs containing the pattern index (as a value) and the position where the occurrence of βj is anticipated (as a key). Algorithm 1: Processing patterns with non-highly periodic αj . 1 2 3 4 5 6 7 8 for i = 1 to n − ` + 1 do h := Φ(w[i..i + ` − 1]) foreach j : Φ(αj ) = h do add a request (i + |Pj | − `, j) to Q foreach request (i, j) ∈ Q at position i do if h = Φ(βj ) then report an occurrence of Pj at i + ` − |Pj | remove (i, j) from Q Algorithm 1 provides a detailed description of the processing phase. Let us analyze its time and space complexities. Due to the properties of Karp-Rabin fingerprints, line 2 can be implemented in O(1) time. Also, the loops in lines 3 and 5 takes extra O(1) time even if the respective collections are empty. Apart from these, every operation can be assigned to a request, each of them taking O(1) (lines 3 and 5-6) or O(log |Q|) (lines 4 and 8) time. To bound |Q|, we need to look at the maximum number of pending requests. Fact 8. For any pattern Pj just O(1 + pending. n `) requests are created and at any time at most one of them is Proof. Note that there is a one-to-one correspondence between requests concerning Pj and the occurrences of αj in T . The distance between two such occurrences must be at least 13 `, because otherwise the period of αj would be at most 13 `, thus making αj highly periodic. This yields the O(1 + n` ) upper bound on the total number of requests. Additionally, any request is pending for at most |Pj | − ` < 13 ` iterations of the main for loop. Thus, the request corresponding to an occurrence of αj is already processed before the next occurrence appears. Hence, the scanning phase uses O(s) space and takes O(n + s(1 + n` ) log s) time. Taking preprocessing into account, we obtain bounds claimed in Lemma 7. 4.2 Highly Periodic Patterns Lemma 9. Let ` be an arbitrary integer. Given a text T of length n and a collection of highly periodic patterns P1 , . . . , Ps such that for 1 ≤ j ≤ s we have ` ≤ |Pj | < 34 `, we can compute the leftmost occurrence of each pattern Pj in T using O(n + s(1 + n` ) log s + s`) total time and O(s) space. The solution is basically the same as in the proof of Lemma 7, except that the algorithm ignores certain shiftable occurrences. An occurrence of x at position i of T is called shiftable if there is another occurrence 1 Hash tables could be used instead of the heap and the deterministic dictionary. Although this would improve the time complexity in Lemma 7, the running time of the algorithm in Thm. 12 would not change and failures with probability inverse polynomial with respect to s would be introduced; see also a discussion before Thm. 1. 6 of x at position i − per(x). The remaining occurrences are called non-shiftable. Notice that the leftmost occurrence is always non-shiftable, so indeed we can safely ignore some of the shiftable occurrences of the patterns. Because 2 per(Pj ) ≤ 32 |Pj | ≤ 89 ` < `, the following fact implies that if an occurrence of Pj is non-shiftable, then the occurrence of αj at the same position is also non-shiftable. Fact 10. Let y be a prefix of x such that |y| ≥ 2 per(x). Suppose x has a non-shiftable occurrence at position i in w. Then, the occurrence of y at position i is also non-shiftable. Proof. Note that per(y) + per(x) ≤ |y| so the periodicity lemma implies that per(y) = per(x). Let x = ρk ρ0 where ρ is the shortest period of x. Suppose that the occurrence of y at position i is shiftable, meaning that y occurs at position i − per(x). Since |y| ≥ per(x), y occurring at position i − per(x) implies that ρ occurs at the same position. Thus w[i − per(x)..i + |x| − 1] = ρk+1 ρ0 . But then x clearly occurs at position i − per(x), which contradicts the assumption that its occurrence at position i is non-shiftable. Consequently, we may generate requests only for the non-shiftable occurrences of αj . In other words, if an occurrence of αj is shiftable, we do not create the requests and proceed immediately to line 5. To detect and ignore such shiftable occurrences, we maintain the position of the last occurrence of every αj . However, if there are multiple patterns sharing the same prefix αj1 = . . . = αjk , we need to be careful so that the time to detect a shiftable occurrence is O(1) rather than O(k). To this end, we build another deterministic dictionary, which stores for each Φ(αj ) a pointer to the variable where we maintain the position of the previously encountered occurrence of αj . The variable is shared by all patterns with the same prefix αj . It remains to analyze the complexity of the modified algorithm. First, we need to bound the number of non-shiftable occurrences of a single αj . Assume that there is a non-shiftable occurrence αj at positions i0 < i such that i0 ≥ i − 21 `. Then i − i0 ≤ 12 ` is a period of T [i0 ..i + ` − 1]. By the periodicity lemma, per(αj ) divides i − i0 , and therefore αj occurs at position i0 − per(αj ), which contradicts the assumption that the occurrence at position i0 is non-shiftable. Consequently, the non-shiftable occurrences of every αj are at least 12 ` characters apart, and the total number of requests and the maximum number of pending requests can be bounded by O(s(1 + n` )) and O(s), respectively, as in the proof of Lemma 7. Taking into the account the time and space to maintain the additional components, which are O(n + s log s) and O(s), respectively, the final bounds remain the same. 4.3 Summary Theorem 11. Given a text T of length n and patterns P1 , . . . , Ps of total length m, using O(n log n + m + s n` log s) total time and O(s) space we can compute the leftmost occurrences in T of every pattern Pj of length at least `. Proof. The algorithm distributes the patterns into O(log n) groups according to their lengths, and then into three classes according to their repetitiveness, which takes O(m) time and O(s) space in total. Then, it applies either Lemma 7 or Lemma 9 on every class. It remains to show that the running times of all those calls sum up to the claimed bound. Each of them can be seen as O(n) plus O(|Pj | + (1 + |Pnj | ) log s) per every pattern Pj . Because ` ≤ |Pj | ≤ n and there are O(log n) groups, this sums up to O(n log n+m+s n` log s). Using Thm. 4 for all patterns of length at most min(n, s), and (if s ≤ n) Thm. 11 for patterns of length at least s, we obtain our main theorem. Theorem 12. Given a text T of length n and patterns P1 , . . . , Ps of total length m, we can compute the leftmost occurrence in T of every pattern Pj using O(n log n + m) total time and O(s) space. 5 Computing Longest Occurring Prefixes In this section we extend Thm. 12 to compute, for every pattern Pj , its longest prefix occurring in the text. A straightforward extension uses binary search to compute the length `j of the longest prefix of Pj occurring 7 in T . All binary searches are performed in parallel, that is, we proceed in O(log n) phases. In every phase we check, for every j, if Pj [1..1 + `j ] occurs in T using Thm. 12 and then update the corresponding `j accordingly. This results in O(n log2 n + m) total time complexity. To avoid the logarithmic multiplicative overhead in the running time, we use a more complex approach requiring a careful modification of all the components developed in Sections 3 and 4. Short patterns. We proceed as in Sect. 3 while maintaining a tentative longest prefix occurring in T for every pattern Pj , denoted Pj [1..`j ]. Recall that after processing a block Ti we obtain, for each pattern Pj , a node such that the corresponding substring is equal to Pj provided that Pj occurs in Ti . Now we need a stronger property, which is that for any length k, `j ≤ k ≤ |Pj |, the ancestor at string depth k of that node (if any) corresponds to Pj [1..k] provided that Pj [1..k] occurs in Ti . This can be guaranteed by modifying the procedure described in Lemma 2: if a child of v1 has no corresponding child of v2 , we report v2 as the twin of all nodes in the subtree rooted at that child of v1 . (Notice that now the string depth of v1 ∈ T1 might be larger than the string depth of its twin v2 ∈ T2 , but we generate exactly one twin for every v1 ∈ T1 .) Using the stronger property we can update every `j by first checking if Pj [1..`j ] occurs in Ti , and if so incrementing `j as long as possible. In more detail, let Ti [b..e] denote the substring corresponding to the twin of Pj . If |Ti [b..e]| < `j , there is nothing to do. Otherwise, we check whether Ti [b..(b + `j − 1)] = Pj [1..`j ] using fingerprints, and if so start to naively compare Ti [b + `j ..e] and Pj [`j + 1..|Pj |]. Because in the end P n j `j ≤ m, updating every `j takes O(s ` + m) additional total time. Theorem 13. Given a text T of length n and patterns P1 , . . . , Ps of total length m, using O(n log `+s n` +m) total time and O(s + `) space we can compute the longest prefix occurring in T for every pattern Pj of length at most `. Long patterns. As in Sect. 4, we again distribute all patterns of length at least ` into groups. However, now for patterns in the i-th group (satisfying blog4/3 |Pj |c = i), we set gi = d(4/3)i e and additionally require that Pj [1..gi ] occurs in T . To verify that this condition is true, we process the groups in the decreasing order of the index i and apply Thm. 1 to prefixes Pj [1..gi ]. If for some pattern Pj the prefix fails to occur in T , we replace Pj setting Pj := Pj [1..gi − 1]. Observe that this operation moves Pj to a group with a smaller index (or makes Pj a short pattern). Additionally, note that in subsequent steps the length of Pj decreases geometrically, so the total length of patterns for which we apply Thm. 1 is O(m) and thus the total running time of this preprocessing phase is O(n log n + m) as long as ` ≥ log s. Hence, from now on we consider only patterns Pj belonging to the i-th group, i.e., such that gi ≤ |Pj | < 34 gi and Pj [1..gi ] occurs in T . As before, we classify patterns depending on their periodicity. However, now the situation is more complex, because we cannot reverse the text and the patterns. As a warm-up, we first describe how to process patterns Pj with a non-highly periodic prefix αj = Pj [1..`]. While not used in the final solution, this step allows us to gradually introduce all the required modifications. Then we show to process all highly periodic patterns, and finally move to the general case, where patters are not highly periodic. Patterns with a non-highly periodic prefix. We maintain a tentative longest prefix occurring in T for every pattern Pj , denoted Pj [1..`j ] and initialized with `j = `, and proceed as in Algorithm 1 with the following modifications. In line 4, the new request is (i + `j − ` + 1, j). In line 6, we compare h with Φ(Pj [(`j +2−`)..(`j +1)]). If these two fingerprints are equal, we have found an occurrence of Pj [1..`j +1]. In such case we try to further extend the occurrence by naively comparing Pj [`j +1..|Pj |] with the corresponding fragment of T and incrementing `j as long as the corresponding characters match. For every Pj we also need to maintain Φ(Pj [(`j + 2 − `)..(`j + 1)]), which can be first initialized in O(`) time and then updated in O(1) time whenever `j is incremented. Because at any time at most one request is pending for every pattern Pj (and thus, while updating `j no such request is pending), this modified algorithm correctly determines the longest occurring prefix for every pattern with non-highly periodic αj . Lemma 14. Let ` be an arbitrary integer. Suppose we are given a text T of length n and patterns P1 , . . . , Ps such that, for 1 ≤ j ≤ s, we have ` ≤ |Pj | < 43 ` and αj = Pj [1..`] is not highly periodic. We can compute 8 the longest prefix occurring in T for every pattern Pj using O(n + s(1 + n` ) log s + s`) total time using O(s) space. Highly periodic patterns. As in Sect. 4.2, we observe that all shiftable occurrences of the longest prefix of Pj occurring in T can be ignored, and therefore it is enough to consider only non-shiftable occurrences of αj (by the same argument, because that longest prefix is of length at least |αj |). Therefore, we can again use Algorithm 1 with the same modifications. As for non-highly periodic αj , we maintain a tentative longest prefix Pj [1..`j ] for every pattern Pj . Whenever a non-shiftable occurrence of αj is detected, we create a new request to check if `j can be incremented. If so, we start to naively compare Pj [1..`j + 1] with the corresponding fragment of T . The total time and space complexity remain unchanged. Lemma 15. Let ` be an arbitrary integer. Given a text T of length n and a collection of highly periodic patterns P1 , . . . , Ps such that, for 1 ≤ j ≤ s, we have ` ≤ |Pj | < 43 `, we can compute the longest prefix occurring in T for every pattern Pj using O(n + s(1 + n` ) log s + s`) total time and O(s) space. General case. Now we describe how to process all non-highly periodic patterns Pj . This will be an extension of the simple modification described for the case of non-highly periodic prefix αj . We start with the following simple combinatorial fact. Fact 16. Let ` ≥ 3 be an integer and w be a non-highly periodic word of length at least `. Then there exists i such that w[i..i+`−1] is not highly periodic and either i = 1 or w[1..i+`−2] is highly periodic. Furthermore, such i can be found in O(|w|) time and constant space assuming read-only random access to w. Proof. If per(w[1..`]) > 31 `, we are done. Otherwise, choose largest j such that per(w[1..j]) = per(w[1..`]). Since w is not highly periodic, we have j < |w|. Thus per(w[1..j]) ≤ 31 ` but per(w[1..j + 1]) > 13 `. We claim that i = j + 2 − ` can be returned. We must argue that per(w[j + 2 − `..j + 1]) > 13 `. Otherwise, the periods of both w[1..j] and w[j + 2 − `..j + 1] are at most 31 `. But these two substrings share a fragment of length ` − 1 ≥ 23 `, so by the periodicity lemma their periods are in fact the same, and then the whole w[1..j + 1] has period at most 31 `, which is a contradiction. Regarding the implementation, we compute per(w[1..`]) using Lemma 6. Then we check how far the period of w[1..`] extends in the whole w naively in O(|w|) time and constant space. For every non-highly periodic pattern Pj we use Fact 16 to find its non-highly periodic substring of length `, denoted Pj [kj ...kj + ` − 1], such that kj = 1 or Pj [1..kj + ` − 2] is highly periodic. We begin with checking if Pj [1..kj + ` − 1] occurs in T using Lemma 7 (if kj = 1, it surely does because of how we partition the patterns into groups). If not, we replace Pj setting Pj := Pj [1..kj + ` − 2], which is highly periodic and can be processed as already described. From now on we consider only patterns Pj such that Pj [kj ..kj + ` − 1] is not highly periodic and Pj [1..kj + ` − 1] occurs in T . We further modify Algorithm 1 to obtain Algorithm 2 as follows. We scan the text T with a sliding window of length ` while maintaining a tentative longest occurring prefix Pj [1..`j ] for every pattern Pj , initialized by setting `j = kj + ` − 1. Whenever we encounter a substring equal to Pj [kj ..kj + ` − 1], i.e., Pj [kj ..kj + ` − 1] = T [i..i + ` − 1], we want to check if Pj [1..kj + ` − 1] = T [i − kj + 1..i + ` − 1] by comparing fingerprints of αj = Pj [1..`] and the corresponding fragment of T . This is not trivial as that fragment is already to the left of the current window. Hence we conceptually move two sliding windows of length `, corresponding to w[i + 31 `..i + 43 ` − 1] and w[i..i + ` − 1], respectively. Because kj ≤ |Pj | − ` ≤ 13 `, whenever the first window generates a request (called request of type I), the second one is still far enough to the left for the request to be processed in the future. Furthermore, because Pj [kj ..kj + ` − 1] is non-highly periodic and the distance between the sliding windows is 13 `, each pattern Pj contributes at most one pending request of type I at any moment and O(1 + n` ) such requests in total. Then, whenever a request of type I is successfully processed, we know that Pj [1..kj + ` − 1] matches with the corresponding fragment of T . We want to check if the occurrence of Pj [1..kj + ` − 1] can be extended to an occurrence of Pj [1..`j + 1]. To this end, we create another request (called request of type II) to check if Pj [`j +2−`..`j +1] matches the corresponding fragment of T . This request can be processed using the second window and, again because per(Pj [1..kj + ` − 1]) > 13 `, 9 Algorithm 2: Processing patterns with non-highly periodic Pj [kj ..kj + `−1]. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 for j = 1 to s do `j := kj + ` − 1 for i = 1 − 31 ` to n − ` + 1 do h1 := Φ(w[i + 13 `..i + 34 ` − 1]) h2 := Φ(w[i + `..i + ` − 1]) foreach j : Φ(Pj [kj ..kj + ` − 1]) = h1 do add a request (i + 13 ` − kj + 1, j) to Q1 foreach request (i, j) ∈ Q1 at position i do if h2 = Φ(P [1..`]) then add a request (i − ` + `j + 1, j) to Q2 remove (i, j) from Q1 foreach request (i − `, j) ∈ Q2 at position i do if h2 = Φ(Pj [`j + 2 − `..`j + 1]) then oj := i + ` − `j − 1 increment `j as long as Pj [`j + 1] = T [oj + `j ] remove (i, j) from Q2 each pattern contribues at most one pending request of type II at any moment and O((1 + n` )) such request in total. Finally, whenever a request of type II is successfully processed, we know that the corresponding `j can be incremented. Therefore, we start to naively compare the characters of Pj [`j + 1..|Pj ]] and the corresponding fragment of T . Since at that time no other request of type II is pending for Pj , such modified algorithm correctly computes all values `j (and the corresponding positions oj ). Lemma 17. Let ` be an arbitrary integer. Suppose we are given a text T of length n and patterns P1 , . . . , Ps such that, for 1 ≤ j ≤ s, we have ` ≤ |Pj | < 34 ` and Pj is non-highly periodic. We can compute the longest prefix occurring in T for every pattern Pj in O(n + s(1 + n` ) log s + s`) total time using O(s) space. By combining all the ingredients, we get the following theorem. Theorem 18. Given a text T of length n and patterns P1 , . . . , Ps of total length m, we can compute the longest prefix occurring in T for every pattern Pj using O(n log n + m) total time and O(s) space. Proof. We proceed as in Thm. 11 and 12, except that now we use Thm. 13, Lemma 15 and Lemma 17 instead of Thm. 4, Lemma 9 and Lemma 7, respectively. Additionally, we need O(n log n + m) time to distribute the long patterns into groups, which is absorbed in the final complexity. Finally, let us note that it is straightforward to modify the algorithm so that we can specify for every pattern Pj an upper bound rj on the starting positions of the occurrences. Theorem 19. Given a text T of length n, patterns P1 , . . . , Ps of total length m, and integers r1 , . . . , rs , we can compute for each pattern the maximum length `j and a position oj ≤ rj such that Pj [1..`j ] = T [oj ..(oj + `j − 1)], using O(n log n + m) total time and O(s) space. 6 Las Vegas Algorithms As shown below, it is not difficult to modify our dictionary matching algorithm so that it always verifies the correctness of the answers. Assuming that we are interested in finding just the leftmost occurrence of every pattern, we obtain an O(n log n + m)-time Las Vegas algorithm (with inverse-polynomial failure probability). In most cases, it suffices to naively verify in O(|Pj |) time whether the leftmost occurrence of Pj detected by the algorithm is valid. If it is not, we are guaranteed that the fingerprints Φ admit a false-positive. Since this event happens with inverse-polynomial probability, a failure can be reported. 10 This simple solution remains valid for short patterns and non-highly periodic long patterns. For highly periodic patterns, the situation is more complicated. The algorithm from Lemma 9 assumes that we are correctly detecting all occurrences of every αj so that we can filter out the shiftable ones. Verifying these occurrences naively might take too much time, because it is not enough check just one occurrence of every αj . Recall that per(αj ) ≤ 31 `. If the previous occurrence of αj was at position i ≥ i − 21 `, we will check if per(αj ) is a period of T [i0 ..i + ` − 1]. If so, either both occurrences (at position i0 and at position i) are false-positives, or none of them is, and the occurrence at position i0 can be ignored. Otherwise, at least one occurrence is surely false-positive, and we declare a failure. To check if per(αj ) is a period of T [i0 ..i + ` − 1], we partition T into overlapping blocks T1 = [1.. 32 `], T2 = [ 13 ` + 1.. 43 `], . . .. Let Tt = T [(t − 1) 13 ` + 1..(t + 1) 13 `] be the rightmost such block fully inside T [1..i + ` − 1]. We calculate the period of Tt using Lemma 6 in O(`) time and O(1) space, and then calculate how far the period extends to the left and to the right, terminating if it extends very far. Formally, we calculate the largest e < (t + 2) 31 ` and the smallest b > (t − 2) 13 ` − 12 ` such that per(Tt ) is a period of T [b..e]. This takes O(`) time for every t summing up to O(n) total time. Then, to check if per(αj ) is a period of T [i0 ..i + ` − 1] we check if it divides the period of per(Tt ) and furthermore r ≥ i + ` − 1 and ` ≤ i0 . Finally, we naively verify the the reported leftmost occurrences of Pj . Consequently, Las Vegas randomization suffices in Thoerem 12. For Thm. 18, we run the Las Vegas version of Thm. 12 with Pj [1..`j ] and Pj [1..`j + 1] as patterns to make sure that the former occur in T but the latter do not. For Thm. 19 we also use Thm. 12, but this time we need to see where the reported leftmost occurrences start compared to bounds rj . 7 Approximating LZ77 in Small Space A non-empty fragment T [i..j] is called a previous fragment if the corresponding subword occurs in T at a position i0 < i. A phrase is either a previous fragment or a single letter not occurring before in T . The LZ77-factorization of a text T [1..n] is a greedy factorization of T into z phrases, T = f1 f2 . . . fz , such that each fi is as long as possible. To formalize the concept of LZ77-approximation, we first make the following definition. Definition 20. Let w = g1 g2 . . . ga be a factorization of w into a phrases. We call it c-optimal if the fragment corresponding to the concatenation of any c consecutive phrases gi . . . gi+c−1 is not a previous fragment. A c-optimal factorization approximates the LZ77-factorization in the number of factors, as the following observation states. However, the stronger property of c-optimality is itself useful in certain situations. Observation 21. If w = g1 g2 . . . ga is a c-optimal factorization of w into a phrases, and the LZ77factorization of w consists of z phrases, then a ≤ c · z. We first describe how to use the dictionary matching algorithm described in Thm. 12 to produce a 2optimal factorization of w[1..n] in O(n log n) time and O(z) working space. Then, the resulting parse can be further refined to produce a (1 + ε)-optimal factorization in O(ε−1 n log n) additional time and the same space using the extension of the dictionary matching algorithm from Thm 18. 7.1 7.1.1 2-Approximation Algorithm Outline. Our algorithm is divided into three phases, each of which refines the factorization from the previous phase: Phase 1. Create a factorization of T [1..n] stored implicitly as z chains consisting of O(log n) phrases each. Phase 2. Try to merge phrases within the chains to produce an O(1)-optimal factorization. Phase 3. Try to merge adjacent factors as long as possible to produce the final 2-optimal factorization. 11 abbbabaabaabaabbbabaabbbbaabbab$ 1 0 0 0 1 1 0 0 1 1 0 0 1 0 0 0 0 1 1 0 0 0 1 0 0 1 0 0 0 0 1 0 1 1 1 0 1 0 0 0 Figure 1: An illustration of Phase 1 of the algorithm, with the “cherries” depicted in thicker lines. The horizontal lines represent the LZ77-factorization and the vertical lines depict factors induced by the tree. Longer separators are drawn between chains, whose lengths are written in binary with the least significant bits on top. Every phase takes O(n log n) time and uses O(z) working space. In the end, we get a 2-approximation of the LZ77-factorization. Phases 1 and 2 use the very simple multiple pattern matching algorithm for patterns of equal lengths developed in Thm. 1, while Phase 3 requires the general multiple pattern matching algorithm obtained in Thm. 12. 7.1.2 Phase 1. To construct the factorization, we imagine creating a binary tree on top the text T of length n = 2k – see also Fig. 1 (we implicitly pad w with sufficiently many $’s to make its length a power of 2). The algorithm works in log n rounds, and the i-th round works on level i of the tree, starting at i = 1 (the children of the root). On level i, the tree divides T into 2i blocks of size n/2i ; the aim is to identify previous fragments among these blocks and declare them as phrases. (In the beginning, no phrases exist, so all blocks are unfactored.) To find out if a block is a previous fragment, we use Thm. 1 and test whether the leftmost occurrence of the corresponding subword is the block itself. The exploration of the tree is naturally terminated at the nodes corresponding to the previous fragments (or single letters not occurring before), forming the leaves of a (conceptual) binary tree. A pair of leaves sharing the same parent is called a cherry. The block corresponding to the common parent is induced by the cherry. To analyze the algorithm, we make the following observation: Fact 22. A block induced by a cherry is never a previous fragment. Therefore, the number of cherries is at most z. Proof. The former part follows from construction. To prove the latter, observe that the blocks induced by different cherries are disjoint and hence each cherry can be assigned a unique LZ77-factor ending within the block. Consequently, while processing level i of the tree, we can afford storing all cherries generated so far on a sorted linked list L. The remaining already generated phrases are not explicitly stored. In addition, we also store a sorted linked list Li of all still unfactored nodes on the current level i (those for which the corresponding blocks are tested as previous fragments). Their number is bounded by z (because there is a cherry below every node on the list), so the total space is O(z). Maintaining both lists sorted is easily accomplished by scanning them in parallel with each scan of T , and inserting new cherries/unfactored nodes at their correct places. Furthermore, in the i-th round we apply Thm. 1 to at most 2i patterns of length Plog n n/2i , so the total time is i=1 O(n + 2i log(2i )) = O(n log n). Next, we analyze the structure of the resulting factorization. Let hx−1 hx and hy hy+1 be the two consecutive cherries. The phrases hx+1 . . . hy−1 correspond to the right siblings of the ancestors of hx and to the left siblings of the ancestors of hy (no further than to the lowest common ancestor of hx and hy ). This naturally partitions hx hx+1 . . . hy−1 hy into two parts, called an increasing chain and a decreasing chain to depict the behaviour of phrase lengths within each part. Observe that these lengths are powers of two, so the 12 structure of a chain of either type is determined by the total length of its phrases, which can be interpreted 0 as a bitvector with bit i0 set to 1 if there is a phrase of length 2i in the chain. Those bitvectors can be created while traversing the tree level by level, passing the partially created bitvectors down to the next level Li+1 until finally storing them at the cherries in L. At the end we obtain a sequence of chains of alternating types, see Fig. 1. Since the structure of each chain follows from its length, we store the sequence of chains rather the actual factorization, which might consist of Θ(z log n) = ω(z) phrases. By Fact 22, our representation uses O(z) words of space and the last phrase of a decreasing chain concatenated with the first phrase of the consecutive increasing chain never form a previous fragment (these phrases form the block induced by the cherry). 7.1.3 Phase 2. In this phase we merge phrases within the chains. We describe how to process increasing chains; the decreasing are handled, mutatis mutandis, analogously. We partition the phrases h` . . . hr within a chain into groups. For each chain we maintain an active group, initially consisting of h` , and scan the remaining phrases in the left-to-right order. We either append a phrase hi to the active group gj , or we output gj and make gj+1 = hi the new active group. The former action is performed if and only if the fragment of length 2|hi | starting at the same position as gj is a previous fragment. Having processed the whole chain, we also output the last active group. Fact 23. Within every chain every group gj forms a valid phrase, but no concatenation of three adjacent groups gj gj+1 gj+2 form a previous fragment. Proof. Since the lengths of phrases form an increasing sequence of powers of two, at the moment we need to decide if we append hi to gj we have |gj | ≤ |h` . . . hi−1 | < |hi |, so 2|hi | > |gj hi |, and thus we are guaranteed if we append gj , then gj hi is a previous factor. Finally, let us prove the aforementioned optimality condition, i.e., that gj gj+1 gj+2 is not a previous fragment for any three consecutive groups. Suppose that we output gj while processing hi , that is, gj+1 = hi . . . hi0 . We did not append hi to gj , so the fragment of length 2|hi | starting at the same position as gj is not a previous fragment. However, |gj gj+1 gj+2 | > |gj+1 gj+2 | ≥ |hi hi+1 | > 2|hi |, so this immediately implies that gj gj+1 gj+2 is not a previous fragment. The procedure described above is executed in parallel for all chains, each of which maintains just the length of its active group. In the i-th round only chains containing a phrase of length 2i participate (we use bit operations to verify which chains have length containing 2i in the binary expansion). These chains provide fragments of length 2i+1 and Thm. 1 is applied to decide which of them are previous fragments. The chains modify their active groups based on the answers; some of them may output their old active groups. These groups form phrases of the output factorization, so the space required to store them is amortized by the size of this factorization. As far as the running time is concerned, we observe that no more than min(z, 2ni ) Plog n chains participate in the i-th round . Thus, the total running time is i=1 O(n + 2ni log 2ni ) = O(n log n). To bound the overall approximation guarantee, suppose there are five consecutive output phrases forming a previous fragment. By Fact 22, these fragments cannot contain a block induced by any cherry. Thus, the phrases are contained within two chains. However, by Fact 23 no three consecutive phrases obtained from a single chain form a previous fragment. Hence the resulting factorization is 5-optimal. 7.1.4 Phase 3. The following lemma achieves the final 2-approximation: Lemma 24. Given a c-optimal factorization, one can compute a 2-optimal factorization using O(c · n log n) time and O(c · z) space. Proof. The procedure consists of c iterations. In every iteration we first detect previous fragments corresponding to concatenations of two adjacent phrases. The total length of the patterns is up to 2n, so this takes 13 O(n log n + m) = O(n log n) time and O(c · z) space using Thm. 12. Next, we scan through the factorization and merge every phrase gi with the preceding phrase gi−1 if gi−1 gi is a previous fragment and gi−1 has not been just merged with its predecessor. We shall prove that the resulting factorization is 2-optimal. Consider a pair of adjacent phrases gi−1 gi in the final factorization and let j be the starting position of gi . Suppose gi−1 gi is a previous fragment. Our algorithm performs merges only, so the phrase ending at position j − 1 concatenated with the phrase starting at position j formed a previous fragment at every iteration. The only reason that these factors were not merged could be another merge of the former factor. Consequently, the factor ending at position j − 1 took part in a merge at every iteration, i.e., gi−1 is a concatenation of at least c phrases of the input factorization. However, all the phrases created by the algorithm form previous fragments, which contradicts the c-optimality of the input factorization. 7.2 Approximation Scheme The starting point is a 2-optimal factorization into a phrases, which can be found in O(n log n) time using the previous method. The text is partitioned into 2ε a blocks corresponding to 2ε consecutive phrases. Every block is then greedily factorized into phrases. The factorization is implemented using Thm. 19 to compute the longest previous fragments in parallel for all blocks as follows. Denote the starting positions of the blocks by b1 < b2 < . . . and let ij be the current position in the i-th block, initially set to bi . For every i, we find the longest previous fragment starting at ij and fully contained inside T [ij ..bj+1 − 1] by computing the longest prefix of T [ij ..bj+1 − 1] occurring in T and starting in T [1..ij − 1], denoted T [ij ..ij + `j − 1]. Then we output every T [ij ..ij + `j − 1] as a new phrase and increase ij by `j . Because every block, by definition, can be factorized into 2ε phrases and the greedy factorization is optimal, this requires O( 1ε n log n) time in total. To bound the approximation guarantee, observe that every phrase inside the block, except possibly for the last one, contains an endpoint of a phrase in the LZ77-factorization. Consequently, the total number of phrases is at most z + 2ε a ≤ (1 + ε)z. Theorem 25. Given a text T of length n whose LZ77-factorization consists of z phrases, we can factorize T into at most 2z phrases using O(n log n) time and O(z) space. Moreover, for any ε ∈ (0, 1] in O(ε−1 n log n) time and O(z) space we can compute a factorization into no more than (1 + ε)z phrases. Acknowledgments. The authors would like to thank the participants of the Stringmasters 2015 workshop in Warsaw, where this work was initiated. We are particularly grateful to Marius Dumitran, Artur Jeż, and Patrick K. Nicholson. References [1] Rabin–Karp algorithm — Wikipedia, The Free Encyclopedia. [2] A. V. Aho and M. J. Corasick. Efficient string matching: An aid to bibliographic search. Commun. ACM, 18(6):333–340, 1975. [3] D. Breslauer, R. Grossi, and F. Mignosi. Simple real-time constant-space string matching. Theor. Comput. Sci., 483:2–9, 2013. [4] R. Clifford, A. Fontaine, E. Porat, B. Sach, and T. Starikovskaya. Dictionary matching in a stream. In N. Bansal and I. Finocchi, editors, ESA 2015, volume 8737 of LNCS, pages 361–372. Springer, Heidelberg, 2015. [5] M. Crochemore and D. Perrin. Two-way string matching. J. ACM, 38(3):651–675, 1991. [6] N. J. Fine and H. S. Wilf. Uniqueness theorems for periodic functions. P. Am. Math. Soc., 16(1):109–114, 1965. 14 [7] J. Fischer, T. I, and D. Köppl. Lempel Ziv computation in small space (LZ-CISS). In F. Cicalese, E. Porat, and U. Vaccaro, editors, Combinatorial Pattern Matching - 26th Annual Symposium, CPM 2015, Ischia Island, Italy, June 29 - July 1, 2015, Proceedings, volume 9133 of Lecture Notes in Computer Science, pages 172–184. Springer, 2015. [8] Z. Galil and J. I. Seiferas. Time-space-optimal string matching. J. Comput. Syst. Sci., 26(3):280–294, 1983. [9] L. Gasieniec, M. Karpinski, W. Plandowski, and W. Rytter. Efficient algorithms for Lempel-Ziv encoding (extended abstract). In R. G. Karlsson and A. Lingas, editors, SWAT 1996, volume 1097 of LNCS, pages 392–403. Springer, Heidelberg, 1996. [10] B. Gum and R. J. Lipton. Cheaper by the dozen: Batched algorithms. In V. Kumar and R. L. Grossman, editors, SDM 2001, pages 1–11. SIAM, Philadelphia, 2001. [11] W. Hon, T. Ku, R. Shah, S. V. Thankachan, and J. S. Vitter. Faster compressed dictionary matching. Theor. Comput. Sci., 475:113–119, 2013. [12] J. Kärkkäinen, D. Kempa, and S. J. Puglisi. Lightweight Lempel-Ziv parsing. In V. Bonifaci, C. Demetrescu, and A. Marchetti-Spaccamela, editors, SEA 2013, volume 7933 of LNCS, pages 139–150. Springer, Heidelberg, 2013. [13] R. M. Karp and M. O. Rabin. Efficient randomized pattern-matching algorithms. IBM J. Res. Dev., 31(2):249–260, 1987. [14] T. Kociumaka, T. Starikovskaya, and H. W. Vildhøj. Sublinear space algorithms for the longest common substring problem. In A. S. Schulz and D. Wagner, editors, ESA 2014, volume 8737 of LNCS, pages 605–617. Springer, Heidelberg, 2014. [15] M. Lohrey. Algorithmics on SLP-compressed strings: A survey. Groups Complexity Cryptology, 4(2):241– 299, 2012. [16] M. Ružić. Constructing efficient dictionaries in close to sorting time. In L. Aceto, I. Damgård, L. A. Goldberg, M. M. Halldórsson, A. Ingólfsdóttir, and I. Walukiewicz, editors, ICALP 2008, volume 5125, pages 84–95. Springer, Heidelberg, 2008. [17] E. Ukkonen. On-line construction of suffix trees. Algorithmica, 14(3):249–260, 1995. [18] J. Ziv and A. Lempel. A universal algorithm for sequential data compression. IEEE Trans. Inform. Theory, 23(3):337–343, 1977. 15
8cs.DS
Going Deeper in Spiking Neural Networks: VGG and Residual Architectures Abhronil Sengupta 1 Yuting Ye 2 Robert Wang 2 Chiao Liu 2 Kaushik Roy 1 arXiv:1802.02627v1 [cs.CV] 7 Feb 2018 Abstract Over the past few years, Spiking Neural Networks (SNNs) have become popular as a possible pathway to enable low-power event-driven neuromorphic hardware. However, their application in machine learning have largely been limited to very shallow neural network architectures for simple problems. In this paper, we propose a novel algorithmic technique for generating an SNN with a deep architecture, and demonstrate its effectiveness on complex visual recognition problems such as CIFAR-10 and ImageNet. Our technique applies to both VGG and Residual network architectures, with significantly better accuracy than the state-of-the-art. Finally, we present analysis of the sparse event-driven computations to demonstrate reduced hardware overhead when operating in the spiking domain. 1. Introduction Spiking Neural Networks (SNNs) are a significant shift from the standard way of operation of Artificial Neural Networks (Farabet et al., 2012). Most of the success of deep learning models of neural networks in complex pattern recognition tasks are based on neural units that receive, process and transmit analog information. Such Analog Neural Networks (ANNs) (Diehl et al., 2015), however, disregard the fact that the biological neurons in the brain (the computing framework after which it is inspired) processes binary spike-based information. Driven by this observation, the past few years have witnessed significant progress in the modeling and formulation of training schemes for SNNs as a new computing paradigm that can potentially replace ANNs as the next generation of Neural Networks. In addition to the fact that SNNs are inherently more biologically plausible, they offer the prospect of event-driven hardware operation. Spiking Neurons process input informa1 Department of Electrical and Computer Engineering, Purdue University, West Lafayette, IN, USA 2 Oculus Research, Facebook Research, Redmond, WA, USA. Correspondence to: Abhronil Sengupta <[email protected]>. tion only on the receipt of incoming binary spike signals. Given a sparsely-distributed input spike train, the hardware overhead (power consumption) for such a spike or eventbased hardware would be significantly reduced since large sections of the network that are not driven by incoming spikes can be power-gated (Chen et al., 1998). However, the vast majority of research on SNNs have been limited to very simple and shallow network architectures on relatively simple digit recognition datasets like MNIST (LeCun et al., 1998) while only few works report their performance on more complex standard vision datasets like CIFAR-10 (Krizhevsky & Hinton, 2009) and ImageNet (Russakovsky et al., 2015). The main reason behind their limited performance stems from the fact that SNNs are a significant shift from the operation of ANNs due to their temporal information processing capability. This has necessitated a rethinking of training mechanisms for SNNs. 2. Related Work Broadly, there are two main categories for training SNNs – supervised and unsupervised. Although unsupervised learning mechanisms like Spike-Timing Dependent Plasticity (STDP) are attractive for the implementation of lowpower on-chip local learning, their performance is still outperformed by supervised networks on even simple digit recognition platforms like the MNIST dataset (Diehl & Cook, 2015). Driven by this fact, a particular category of supervised SNN learning algorithms attempts to train ANNs using standard training schemes like backpropagation (to leverage the superior performance of standard training techniques for ANNs) and subsequently convert to event-driven SNNs for network operation (Cao et al., 2015; Zhao et al., 2015; Pérez-Carrasco et al., 2013; Diehl et al., 2015). This can be particularly appealing for NN implementations in low-power neuromorphic hardware specialized for SNNs (Merolla et al., 2014; Akopyan et al., 2015) or interfacing with silicon cochleas or event-driven sensors (Posch et al., 2014; 2011). Our work falls in this category and is based on the ANN-SNN conversion scheme proposed by authors in Ref. (Diehl et al., 2015). However, while prior work considers the ANN operation only during the conversion process, we show that considering the actual SNN operation during the conversion step is crucial for achieving minimal loss in classification accuracy. To Going Deeper in Spiking Neural Networks: VGG and Residual Architectures that effect, we propose a novel weight-normalization technique that ensures that the actual SNN operation is in the loop during the conversion phase. Note that this work tries to exploit neural activation sparsity by converting networks to the spiking domain for power-efficient hardware implementation and are complementary to efforts aimed at exploring sparsity in synaptic connections (Han et al., 2015a). 3. Main Contributions The specific contributions of our work are as follows: (i) As will be explained in later sections, there are various architectural constraints involved for training ANNs that can be converted to SNNs in a near-lossless manner. Hence, it is unclear whether the proposed techniques would scale to larger and deeper architectures for more complicated tasks. We provide, for the first time, proof of concept experiments that deep SNNs (extending from 16 to 34 layers) can provide competitive accuracies over complex datasets like CIFAR-10 and ImageNet. (ii) We propose a new ANN-SNN conversion technique that statistically outperforms state-of-the-art techniques. We report a classification error of 8.45% on the CIFAR-10 dataset which is the best-performing result reported for any SNN network, till date. For the first time, we report an SNN performance on the ImageNet 2012 validation set. We achieve a 30.04% top-1 error rate and 10.99% top-5 error rate for VGG-16 architectures. (iii) We explore Residual Network (ResNet) architectures as a potential pathway to enable deeper SNNs. We present insights and design constraints that are required to ensure ANN-SNN conversion for ResNets. We report a classification error of 12.54% on the CIFAR-10 dataset and a 34.53% top-1 error rate and 13.67% top-5 error rate on the ImageNet validation set. This is the first work that attempts to explore SNNs with residual network architectures. (iv) We demonstrate that SNN network sparsity significantly increases as the network depth increases. This further motivates the exploration of converting ANNs to SNNs for event-driven operation to reduce compute overhead. 4. Preliminaries 4.1. Input and Output Representation The main difference between ANN and SNN operation is the notion of time. While ANN inputs are static, SNNs operate based on dynamic binary spiking inputs as a function of time. The neural nodes also receive and transmit binary spike input signals in SNNs, unlike in ANNs, where the inputs and outputs of the neural nodes are analog values. In this work, we consider a rate-encoded network operation where the average number of spikes transmitted as input to the network over a large enough time window is approximately proportional to the magnitude of the original ANN inputs (pixel intensity in this case). The duration of the time window is dictated by the desired network performance (for instance, classification accuracy) at the output layer of the network. A Poisson event-generation process is used to produce the input spike train to the network. Every time-step of SNN operation is associated with the generation of a random number whose value is compared against the magnitude of the corresponding input. A spike event is triggered if the generated random number is less than the value of the corresponding pixel intensity. This process ensures that the average number of input spikes in the SNN is proportional to the magnitude of the corresponding ANN inputs and is typically used to simulate an SNN for recognition tasks based on datasets for static images (Diehl et al., 2015). Given a Poisson-generated spike train being fed to the network, spikes will be produced at the network outputs. Inference is based on the cumulative spike count of neurons at the output layer of the network over a given time-window. 4.2. ANN and SNN Neural Operation ANN to SNN conversion schemes usually consider Rectified Linear Unit (ReLU) as the ANN neuron activation function. For a neuron receiving inputs xi through synaptic weights wi , the ReLU neuron output y is given by, ! X y = max 0, wi .xi (1) i Although ReLU neurons are typically used in a large number of machine learning tasks at present, the main reason behind their usage for ANN-SNN conversion schemes is that they bear functional equivalence to an Integrate-Fire (IF) Spiking Neuron without any leak and refractory period (Cao et al., 2015; Diehl et al., 2015). Note that this is a particular type of Spiking Neuron model (Izhikevich, 2003). Let us consider the ANN inputs xi encoded in time as a spike train Xi (t), where E[Xi (t)] ∝ xi (for the rate encoding network being considered in this work). The IF Spiking Neuron keeps track of its membrane potential, vmem , which integrates incoming spikes and generates an output spike whenever the membrane potential cross a particular threshold vth . The membrane potential is reset to zero at the generation of an output spike. All neurons are reset whenever a spike train corresponding to a new image/pattern in presented. The IF Spiking Neuron dynamics as a function of time-step, t, can be described by the following equation, X vmem (t + 1) = vmem (t) + wi .Xi (t) (2) i Going Deeper in Spiking Neural Networks: VGG and Residual Architectures Let us first consider the simple case of a neuron being driven by a single input X(t) and a positive synaptic weight w. Due to the absence of any leak term in the neural dynamics, it is intuitive to show that the corresponding output spiking rate of the neuron is given by E[Y(t)] ∝ E[X(t)], with the proportionality factor being dependent on the ratio of w and vth . In the case when the synaptic weight is negative, the output spiking activity of the IF neuron is zero since the neuron is never able to cross the firing potential vth , mirroring the functionality of a ReLU. The higher the ratio of the threshold with respect to the weight, the more time is required for the neuron to spike, thereby reducing the neuron spiking rate, E[Y(t)], or equivalently increasing the time-delay for the neuron to generate a spike. A relatively high firing threshold can cause a huge delay for neurons to generate output spikes. For deep architectures, such a delay can quickly accumulate and cause the network to not produce any spiking outputs for relatively long periods of time. On the other hand, a relatively low threshold causes the SNN to lose any ability to distinguish between different magnitudes of the spike inputs P being accumulated to the membrane potential (the term i wi .Xi (t) in Eq. 2) of the Spiking Neuron, causing it to lose evidence during the membrane potential integration process. This, in turn, results in accuracy degradation of the converted network. Hence, an appropriate choice of the ratio of the neuron threshold to the synaptic weights is essential to ensure minimal loss in classification accuracy during the ANN-SNN conversion process (Diehl et al., 2015). Consequently, most of the research work in this field has been concentrated on outlining appropriate algorithms for threshold-balancing, or equivalently, weight normalizing different layers of a network to achieve near-lossless ANN-SNN conversion. 4.3. Architectural Constraints 4.3.1. B IAS IN N EURAL U NITS Typically neural units used for ANN-SNN conversion schemes are trained without any bias term (Diehl et al., 2015). This is due to the fact that optimization of the bias term in addition to the spiking neuron threshold expands the parameter space exploration, thereby causing the ANNSNN conversion process to be more difficult. Requirement of bias less neural units also entails that Batch Normalization technique (Ioffe & Szegedy, 2015) cannot be used as a regularizer during the training process since it biases the inputs to each layer of the network to ensure each layer is provided with inputs having zero mean. Instead, we use dropout (Srivastava et al., 2014) as the regularization technique. This technique simply masks portions of the input to each layer by utilizing samples from a Bernoulli distribution where each input to the layer has a specified probability of being dropped. 4.3.2. P OOLING O PERATION Deep convolutional neural network architectures typically consist of intermediate pooling layers to reduce the size of the convolution output maps. While various choices exist for performing the pooling mechanism, the two popular choices are either max-pooling (maximum neuron output over the pooling window) or spatial-averaging (twodimensional average pooling operation over the pooling window). Since the neuron activations are binary in SNNs instead of analog values, performing max-pooling would result in significant information loss for the next layer. Consequently, we consider spatial-averaging as the pooling mechanism in this work (Diehl et al., 2015). 5. Deep Convolutional SNN Architectures: VGG As mentioned previously, our work is based on the proposal outlined by authors in Ref. (Diehl et al., 2015). In order to ensure that a spiking neuron threshold is sufficiently high to distinguish different magnitude of the spike inputs, a worst case solution would be to set the threshold of a particular layer to the maximum of the summation of all the positive synaptic weights of neurons in that layer. However, such a “Model-Based Normalization” technique is highly pessimistic since all the fan-in neurons are not supposed to fire at every time-step (Diehl et al., 2015). In order to circumvent this issue, authors in Ref. (Diehl et al., 2015) proposed a “Data-Based Normalization” Technique wherein the neuron threshold of a particular layer is set equal to the maximum activation of all ReLUs in the corresponding layer (by passing the entire training set through the trained ANN once after training is completed). Such a “Data-Based” technique performed significantly better than the “Model-Based” algorithm in terms of the final classification accuracy and latency of the converted SNN (three-layered fully connected and convolutional architectures) for a digit recognition problem on the MNIST dataset (Diehl et al., 2015). Note that, this process is referred to as “weight-normalization” and “threshold-balancing” interchangeably in this text. As mentioned before, the goal of this work is to optimize the ratio of the synaptic weights with respect to the neuron firing threshold, vth . Hence, either all the synaptic weights preceding a neural layer are scaled by a normalization factor wnorm equal to the maximum neural activation and the threshold is set equal to 1 (“weight-normalization”), or the threshold vth is set equal to the maximum neuron activation for the corresponding layer with the synaptic weights remaining unchanged (“threshold-balancing”). Both operations are exactly equivalent mathematically. Going Deeper in Spiking Neural Networks: VGG and Residual Architectures 5.1. Proposed Algorithm: S PIKE-N ORM Driven by this observation, we propose a weightnormalization technique that adaptively balances the threshold of each layer by considering the actual operation of the SNN in the loop during the ANN-SNN conversion process. The algorithm normalizes the weights of the network sequentially for each layer. Given a particular trained ANN, the first step is to generate the input Poisson spike train for the network over the training set for a large enough time-window. The Poisson spike train allows us to record thePmaximum summation of weighted spike-input (the term i wi .Xi (t) in Eq. 2) that would be received by the first neural layer of the network. In order to minimize the temporal delay of the neuron and simultaneously ensure that the neuron firing threshold is not too low, we weightnormalize the first layer depending on the maximum spikebased input received by the first layer. After the threshold of the first layer is set, we are provided with a representative spike train at the output of the first layer which enables us to generate the input spike-stream for the next layer. The process is continued sequentially for all the layers of the network. The main difference between our proposal and prior work (Diehl et al., 2015) is the fact that the proposed weight-normalization scheme accounts for the actual SNN operation during the conversion process. As we will show in the Results section, this scheme is crucial to ensure nearlossless ANN-SNN conversion for significantly deep architectures and for complex recognition problems. 6. Extension to Residual Architectures 6.1. Architecture Constraints Residual network architectures were proposed as an attempt to scale convolutional neural networks to very deep Spatial Convolution No timedelay Spatial Convolution Same Threshold ReLU ReLU Spatial Convolution Spatial Convolution Time-Delay Addition Addition ReLU Output Output (a) Maximum SNN Activation However, the above algorithm leads us to the question: Are ANN activations representative of SNN activations? Let us consider a particular example, where the case of maximum activation is shown for a single ReLU. The neuron receives two inputs, namely 0.5 and 1. Let us consider unity synaptic weights in this scenario. Since the maximum ReLU activation is 1.5, the neuron threshold would be set equal to 1.5. However, when this network is converted to the SNN mode, both the inputs would be propagating binary spike signals. The ANN input, equal to 1, would be converted to spikes transmitting at every time-step while the other input would transmit spikes approximately 50% of the duration of a large enough time-window. Hence, the actual summation of spike inputs received by the neuron per time-step would be 2 for a large number of samples, which is higher than the spiking threshold (1.5). Clearly, some information loss would take place due to the lack of this evidence integration. ReLU of previous layer Input 14 12 10 8 6 4 2 0 0 (b) 5 10 #Layer 15 20 (c) Figure 1. (a) The basic ResNet functional unit, (b) Design constraints introduced in the functional unit to ensure near-lossless ANN-SNN conversion, (c) Typical maximum SNN activations for a ResNet having junction ReLU layers but the non-identity and identity input paths not having the same spiking threshold. While this is not representative of the case with equal thresholds in the two paths, it does justify the claim that after a few initial layers, the maximum SNN activations decay to values close to unity due to the identity mapping. layered stacks (He et al., 2016a). Although different variants of the basic functional unit have ben explored, we will only consider identity shortcut connections in this text (shortcut type-A according to the paper (He et al., 2016a)). Each unit consists of two parallel paths. The non-identity path consists of two spatial convolution layers with an intermediate ReLU layer. While the original ResNet formulation considers ReLUs at the junction of the parallel nonidentity and identity paths (He et al., 2016a), recent formulations do not consider junction ReLUs in the network architecture (He et al., 2016b). Absence of ReLUs at the junction point of the non-identity and identity paths was observed to produce a slight improvement in classification accuracy on the CIFAR-10 dataset1 . Due to the presence of the shortcut connections, important design considerations need to be accounted for to ensure near-lossless ANNSNN conversion. We start with the basic unit, as shown in Fig. 1(a), and point-wise impose various architectural constraints with justifications. 1 http://torch.ch/blog/2016/02/04/resnets.html Going Deeper in Spiking Neural Networks: VGG and Residual Architectures 6.1.1. R E LU S AT EACH JUNCTION POINT As we will show in the Results section, application of our proposed S PIKE-N ORM algorithm on such a residual architecture resulted in a converted SNN that exhibited accuracy degradation in comparison to the original trained ANN. We hypothesize that this degradation is attributed mainly to the absence of any ReLUs at the junction points. Each ReLU when converted to an IF Spiking Neuron imposes a particular amount of characteristic temporal delay (time interval between an incoming spike and the outgoing spike due to evidence integration). Due to the shortcut connections, spike information from the initial layers gets instantaneously propagated to later layers. The unbalanced temporal delay in the two parallel paths of the network can result in distortion of the spike information being propagated through the network. Consequently, as shown in Fig. 1(b), we include ReLUs at each junction point to provide a temporal balancing effect to the parallel paths (when converted to IF Spiking Neurons). An ideal solution would be to include a ReLU in the parallel path, but that would destroy the advantage of the identity mapping. 6.1.2. S AME THRESHOLD OF ALL FAN - IN LAYERS As shown in the next section, direct application of our proposed weight-normalization scheme still resulted in some amount of accuracy loss in comparison to the baseline ANN accuracy. Note that the weight-normalization scheme was applied by considering the immediate fan-in layer, i.e. the junction neuron layer is threshold balanced according to the activity of the non-identity neuron path. However, this layer also receives inputs from the previous junction neuron layer. Since the output spiking activity of a particular neuron is also dependent on the neuron threshold, all the fan-in neuron layers should have the same threshold to ensure that input spike information to the next layer is rate-encoded appropriately. However, the spiking threshold of the neuron layer in the non-identity path is dependent on the activity of the neuron layer at the previous junction. An observation of the typical threshold-balancing factors for the network without using this constraint (shown in Fig. 1(c)) reveal that the threshold-balancing factors mostly lie around unity after a few initial layers. This occurs mainly due to the identity mapping. The maximum summation of spike inputs received by the neurons in the junction layers are dominated by the identity mapping (close to unity). From this observation, we heuristically choose both the thresholds of the non-identity ReLU layer and the identity-ReLU layer equal to 1. However, the accuracy is still unable to approach the baseline ANN accuracy, which leads us to the third design constraint. 6.1.3. I NITIAL N ON -R ESIDUAL P RE -P ROCESSING L AYERS An observation of Fig. 1(c) reveals that the thresholdbalancing factors of the initial junction neuron layers are significantly higher than unity. This can be a primary reason for the degradation in classification accuracy of the converted SNN. We note that the residual architectures used by authors in Ref. (He et al., 2016a) use an initial convolution layer with a very wide receptive field (7 × 7 with a stride of 2) on the ImageNet dataset. The main motive behind such an architecture was to show the impact of increasing depth in their residual architectures on the classification accuracy. Inspired by the VGG-architecture, we replace the first 7 × 7 convolutional layer by a series of three 3 × 3 convolutions where the first two layers do not exhibit any shortcut connections. Addition of such initial non-residual pre-processing layers allows us to apply our proposed weight-normalization scheme in the initial layers while using a unity weight-normalization/thresholdbalancing factor for the later residual layers. As shown in the Results section, this scheme significantly assists in achieving classification accuracies close to the baseline ANN accuracy since after the initial layers, the maximum neuron activations decay to values close to unity because of the identity mapping. 7. Experiments 7.1. Datasets and Implementation We evaluate our proposals on standard visual object recognition benchmarks, namely the CIFAR-10 and ImageNet datasets. Experiments performed on networks for the CIFAR-10 dataset are trained on the training set images with per-pixel mean subtracted and evaluated on the testing set. We also present results on the much more complex ImageNet 2012 dataset that contains 1.28 million training images and report evaluation (top-1 and top-5 error rates) on the 50, 000 validation set. 224 × 224 crops from the input images are used for this experiment. We use VGG-16 architecture (Simonyan & Zisserman, 2014) for both the datasets. ResNet-20 configuration outlined in Ref. (He et al., 2016a) is used for the CIFAR-10 dataset while ResNet-34 is used for experiments on the ImageNet dataset. As mentioned previously, we do not utilize any batch-normalization layers. For VGG networks, a dropout layer is used after every ReLU layer except for those layers which are followed by a pooling layer. For Residual networks, we use dropout only for the ReLUs at the non-identity parallel paths but not at the junction layers. We found this crucial for achieving training convergence. Our implementation is derived from the Facebook ResNet implementation code for CIFAR and ImageNet datasets Going Deeper in Spiking Neural Networks: VGG and Residual Architectures available publicly2 . We use similar image pre-processing steps and scale and aspect-ratio augmentation techniques as used in (Szegedy et al., 2015). We report single-crop testing results while the error rates can be further reduced with 10-crop testing (Krizhevsky et al., 2012). Networks used for the CIFAR-10 dataset are trained on 2 GPUs with a batchsize of 256 for 200 epochs, while ImageNet training is performed on 8 GPUs for 100 epochs with a similar batchsize. The initial learning rate is 0.05. The learning rate is divided by 10 twice, at 81 and 122 epochs for CIFAR-10 dataset and at 30 and 60 epochs for ImageNet dataset. A weight decay of 0.0001 and a momentum of 0.9 is used for all the experiments. Proper weight initialization is crucial to achieve convergence in such deep networks without batch-normalization. For a non-residual convolutional layer (for both VGG and ResNet architectures) having kernel size k × k with n output channels, the weights are initialized from a normal distribution and q standard deviation k22n . However, for residual convolutional layers, the √standard deviation used for the normal distribution was k22n . We observed this to be important for achieving training convergence and a similar observation was also outlined in Ref. (Hardt & Ma, 2016) although their networks were trained without both dropout and batch-normalization. 7.2. Experiments for VGG Architectures Our VGG-16 model architecture follows the implementation outlined in 3 except that we do not utilize the batchnormalization layers. We used a randomly chosen minibatch of size 256 from the training set for the weightnormalization process on the CIFAR-10 dataset. While the entire training set can be used for the weight-normalization process, using a representative subset did not impact the results. We confirmed this by running multiple independent runs for both the CIFAR and ImageNet datasets. The standard deviation of the final classification error rate after 2500 time-steps was ∼ 0.01. All results reported in this article represent the average of 5 independent runs of the spiking network (since the input to the network is a random process). No notable difference in the classification error rate was observed at the end of 2500 time-steps and the network outputs converged to deterministic values despite being driven by stochastic inputs. For the SNN model based weight-normalization scheme (S PIKE-N ORM algorithm) we used 2500 time-steps for each layer sequentially to normalize the weights. Table 1 summarizes our results for the CIFAR-10 dataset. The baseline ANN error rate on the testing set was 8.3%. Since the main contribution of this work is to minimize the 2 3 https://github.com/facebook/fb.resnet.torch https://github.com/szagoruyko/cifar.torch loss in accuracy during conversion from ANN to SNN for deep-layered networks and not in pushing state-of-the-art results in ANN training, we did not perform any hyperparameter optimization. However, note that despite several architectural constraints being present in our ANN architecture, we are able to train deep networks that provide competitive classification accuracies using the training mechanisms described in the previous subsection. Further reduction in the baseline ANN error rate is possible by appropriately tuning the learning parameters. For the VGG-16 architecture, our implementation of the ANNmodel based weight-normalization technique, proposed by Ref. (Diehl et al., 2015), yielded an average SNN error rate of 8.54% leading to an error increment of 0.24%. The error increment was minimized to 0.15% on applying our proposed S PIKE-N ORM algorithm. Note that we consider a strict model-based weight-normalization scheme to isolate the impact of considering the effect of an ANN versus our SNN model for threshold-balancing. Further optimizations of considering the maximum synaptic weight during the weight-normalization process (Diehl et al., 2015) is still possible. Previous works have mainly focused on much shallower convolutional neural network architectures. Although Ref. (Hunsberger & Eliasmith, 2016) reports results with an accuracy loss of 0.18%, their baseline ANN suffers from some amount of accuracy degradation since their networks are trained with noise (in addition to architectural constraints mentioned before) to account for neuronal response variability due to incoming spike trains (Hunsberger & Eliasmith, 2016). It is also unclear whether the training mechanism with noise would scale up to deeper layered networks. Our work reports the best performance of a Spiking Neural Network on the CIFAR-10 dataset till date. Our network is the deepest convolutional architecture explored so far and our proposed algorithm statistically outperforms state-of-the-art algorithms for ANN-SNN conversion schemes in such architectures. The impact of our proposed algorithm is much more apparent on the more complex ImageNet dataset. The rates for the top-1 (top-5) error on the ImageNet validation set are summarized in Table 2. Note that these are single-crop results. The accuracy loss during the ANN-SNN conversion process is minimized by a margin of 0.57% by considering SNN-model based weight-normalization scheme. It is therefore expected that our proposed S PIKE-N ORM algorithm would significantly perform better than an ANNmodel based conversion scheme as the pattern recognition problem becomes more complex since it accounts for the actual SNN operation during the conversion process. This is the first work that reports competitive performance of a Spiking Neural Network on the entire 50, 000 ImageNet Going Deeper in Spiking Neural Networks: VGG and Residual Architectures Table 1. Results for CIFAR-10 Dataset ANN Error - SNN Error 10.68% Error Increment - 8-layered networks (Hunsberger & Eliasmith, 2016) 16.28% 16.46% VGG-16 (ANN model based conversion) 8.3% VGG-16 (S PIKE-N ORM) 8.3% 3-layered networks (Esser et al., 2016) Network Architecture ANN Error SNN Error Error Increment 29.48% (10.61%) 30.61% (11.21%) 1.13% (0.6%) 0.18% VGG-16 (ANN model based conversion) 8.54% 0.24% VGG-16 (S PIKE-N ORM) 29.48% (10.61%) 30.04% (10.99%) 0.56% (0.38%) 8.45% 0.15% 100 4 7.3. Experiments for Residual Architectures On the ImageNet dataset, we use the deeper ResNet-34 model outlined in Ref. (He et al., 2016a). The initial 7 × 7 convolutional layer is replaced by three 3 × 3 convolutional layers where the initial two layers are non-residual plain units. The baseline ANN is trained with an error of 29.31% while the converted SNN error is 34.53% at the end of 2500 timesteps. The results are summarized in Table. 3 and convergence plots for all our networks are provided in Fig. 3. It is worth noting here that the main motivation of exploring Residual Networks is to go deeper in Spiking Neural Networks. We explore relatively simple ResNet architecRef. (Hunsberger & Eliasmith, 2016) reports a performance of 48.2%(23.8%) on the first 3072-image test batch of the ImageNet 2012 dataset. 60 Basic Architecture 40 Constraint 2 20 0 0 Constraint 3 1000 2000 Timesteps Figure 2. Impact of the architectural constraints for Residual Networks. “Basic Architecture” does not involve any junction ReLU layers. “Constraint 1” involves junction ReLUs while “Constraint 2” imposes equal unity threshold for all residual units. Network accuracy is significantly improved with the inclusion of “Constraint 3” that involves pre-processing weight-normalized plain convolutional layers at the network input stage. 100 Error (%) Our residual networks for CIFAR-10 and ImageNet datasets follow the implementation in Ref. (He et al., 2016a). We first attempt to explain our design choices for ResNets by sequentially imposing each constraint on the network and showing their corresponding impact on network performance in Fig. 2. The “Basic Architecture” involves a residual network without any junction ReLUs. “Constraint 1” involves junction ReLUs without having equal spiking thresholds for all fan-in neural layers. “Constraint 2” imposes an equal threshold of unity for all the layers while “Constraint 3” performs best with two pre-processing plain convolutional layers (3 × 3) at the beginning of the network. The baseline ANN ResNet20 was trained with an error of 10.9% on the CIFAR10 dataset. Note that although we are using terminology consistent with Ref. (He et al., 2016a) for the network architectures, our ResNets contain two extra plain pre-processing layers. The converted SNN according to our proposal yielded a classification error rate of 12.54%. Weight-normalizing the initial two layers using the ANNmodel based weight-normalization scheme produced an average error of 12.87%, further validating the efficiency of our weight-normalization technique. 4 Constraint 1 80 Error (%) 2012 validation set. Table 2. Results for ImageNet Dataset 50 VGG-16 ResNet-20 CIFAR-10 100 Error (%) Network Architecture 80 VGG-16 ResNet-34 ImageNet 60 40 0 0 1000 2000 Timesteps 0 2000 1000 Timesteps Figure 3. Convergence plots for the VGG and ResNet SNN architectures for CIFAR-10 and ImageNet datasets are shown above. The classification error reduces as more evidence is integrated in the Spiking Neurons with increasing time-steps. Note that although the network depths are similar for CIFAR-10 dataset, the ResNet-20 converges much faster than the VGG architecture. The delay for inferencing is higher for ResNet-34 on the ImageNet dataset due to twice the number of layers as the VGG network. tures, as the ones used in Ref. (He et al., 2016a), which have an order of magnitude lower parameters than standard VGG-architectures. Further hyper-parameter optimizations or more complex architectures are still possible. While the accuracy loss in the ANN-SNN conversion process is more for ResNets than plain convolutional architectures, yet further optimizations like including more pre-processing initial layers or better threshold-balancing schemes for the residual units can still be explored. This work serves as the first work to explore ANN-SNN conversion schemes for Total Number of Spikes per Neuron Total Number of Spikes per Neuron Going Deeper in Spiking Neural Networks: VGG and Residual Architectures 10 VGG-16 5 0 6 4 1 3 9 11 7 5 Number of Layers 13 ResNet-34 2 0 20 30 10 Number of Layers Figure 4. Average cumulative spike count generated by neurons in VGG and ResNet architectures on the ImageNet dataset as a function of the layer number. 500 timesteps were used for accumulating the spike-counts for VGG networks while 2000 time-steps were used for ResNet architectures. The neural spiking sparsity increases significantly as network depth increases. Residual Networks and attempts to highlight important design constraints required for minimal loss in the conversion process. 7.4. Computation Reduction Due to Sparse Neural Events ANN operation for prediction of the output class of a particular input requires a single feed-forward pass per image. For SNN operation, the network has to be evaluated over a number of time-steps. However, specialized hardware that accounts for the event-driven neural operation and “computes only when required” can potentially exploit such alternative mechanisms of network operation. For instance, Fig. 4 represents the average total number of output spikes produced by neurons in VGG and ResNet architectures as a function of the layer for ImageNet dataset. A randomly chosen minibatch was used for the averaging process. We used 500 timesteps for accumulating the spikecounts for VGG networks while 2000 time-steps were used for ResNet architectures. This is in accordance to the convergence plots shown in Fig. 3. An important insight obtained from Fig. 4 is the fact that neuron spiking activity becomes sparser as the network depth increases. Hence, benefits from event-driven hardware is expected to increase as the network depth increases. While an estimate of the actual energy consumption reduction for SNN mode of operation is outside the scope of this current work, we provide an intuitive insight by providing the number of computations per synaptic operation being performed in the ANN versus the SNN. The number of synaptic operations per layer of the network can be easily estimated for an ANN from the architecture Table 3. Results for Residual Networks Dataset Network tecture CIFAR-10 ImageNet ResNet-20 ResNet-34 Archi- ANN Error SNN Error 10.9% 29.31% (10.31%) 12.54% 34.53% (13.67%) for the convolutional and linear layers. For the ANN, a multiply-accumulate (MAC) computation takes place per synaptic operation. On the other hand, a specialized SNN hardware would perform an accumulate computation (AC) per synaptic operation only upon the receipt of an incoming spike. Hence, the total number of AC operations occurring in the SNN would be represented by the layerwise product and summation of the average cumulative neural spike count for a particular layer and the corresponding number of synaptic operations. Calculation of this metric reveal that for the VGG network the ratio of SNN AC operations to ANN MAC operations is 1.975 while the ratio is 2.4 for the ResNet (the metric includes only ReLU/IF spiking neuron activations in the network). However, note the fact that a MAC operation involves an order of magnitude more energy consumption than an AC operation. For instance, Ref. (Han et al., 2015b) reports that the energy consumption in a 32-bit floating point MAC operation is 3.2pJ while the energy consumption is only 0.1pJ for an AC operation in 45nm technology. Hence, the energy consumption reduction for our SNN implementation is expected to be 16.2× for the VGG network and 13.3× for the ResNet in comparison to the original ANN implementation. 8. Conclusions and Future Work This work serves to provide inspiration to the fact that SNNs exhibit similar computing power as their ANN counterparts. This can potentially pave the way for the usage of SNNs in large scale visual recognition tasks, which can be enabled by low-power neuromorphic hardware. However, there are still open areas of exploration for improving SNN performance. A significant contribution to the present success of deep NNs is attributed to Batch-Normalization (Ioffe & Szegedy, 2015). While using bias less neural units constrain us to train networks without BatchNormalization, algorithmic techniques to implement Spiking Neurons with a bias term should be explored. Further, it is desirable to train ANNs and convert to SNNs without any accuracy loss. Although the proposed conversion technique attempts to minimize the conversion loss to a large extent, yet other variants of neural functionalities apart from ReLU-IF Spiking Neurons could be potentially explored to further reduce this gap. Additionally, further optimizations to minimize the accuracy loss in ANN-SNN conversion for ResNet architectures should be explored to scale SNN performance to even deeper architectures. Going Deeper in Spiking Neural Networks: VGG and Residual Architectures References Akopyan, Filipp, Sawada, Jun, Cassidy, Andrew, AlvarezIcaza, Rodrigo, Arthur, John, Merolla, Paul, Imam, Nabil, Nakamura, Yutaka, Datta, Pallab, Nam, Gi-Joon, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 34(10):1537–1557, 2015. Cao, Yongqiang, Chen, Yang, and Khosla, Deepak. Spiking deep convolutional neural networks for energyefficient object recognition. International Journal of Computer Vision, 113(1):54–66, 2015. Chen, Zhanping, Johnson, Mark, Wei, Liqiong, and Roy, W. Estimation of standby leakage power in cmos circuit considering accurate modeling of transistor stacks. In Low Power Electronics and Design, 1998. Proceedings. 1998 International Symposium on, pp. 239–244. IEEE, 1998. Diehl, Peter U and Cook, Matthew. Unsupervised learning of digit recognition using spike-timing-dependent plasticity. Frontiers in computational neuroscience, 9, 2015. Diehl, Peter U, Neil, Daniel, Binas, Jonathan, Cook, Matthew, Liu, Shih-Chii, and Pfeiffer, Michael. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In Neural Networks (IJCNN), 2015 International Joint Conference on, pp. 1–8. IEEE, 2015. Esser, Steven K, Merolla, Paul A, Arthur, John V, Cassidy, Andrew S, Appuswamy, Rathinakumar, Andreopoulos, Alexander, Berg, David J, McKinstry, Jeffrey L, Melano, Timothy, Barch, Davis R, et al. Convolutional networks for fast, energy-efficient neuromorphic computing. Proceedings of the National Academy of Sciences, pp. 201604850, 2016. Farabet, Clément, Paz, Rafael, Pérez-Carrasco, Jose, Zamarreño-Ramos, Carlos, Linares-Barranco, Alejandro, LeCun, Yann, Culurciello, Eugenio, SerranoGotarredona, Teresa, and Linares-Barranco, Bernabe. Comparison between frame-constrained fix-pixel-value and frame-free spiking-dynamic-pixel ConvNets for visual processing. Frontiers in neuroscience, 6, 2012. Han, Song, Mao, Huizi, and Dally, William J. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015a. Han, Song, Pool, Jeff, Tran, John, and Dally, William. Learning both weights and connections for efficient neural network. In Advances in neural information processing systems, pp. 1135–1143, 2015b. Hardt, Moritz and Ma, Tengyu. Identity matters in deep learning. arXiv preprint arXiv:1611.04231, 2016. He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, and Sun, Jian. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016a. He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, and Sun, Jian. Identity mappings in deep residual networks. In European Conference on Computer Vision, pp. 630–645. Springer, 2016b. Hunsberger, Eric and Eliasmith, Chris. Training spiking deep networks for neuromorphic hardware. arXiv preprint arXiv:1611.05141, 2016. Ioffe, Sergey and Szegedy, Christian. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456, 2015. Izhikevich, Eugene M. Simple model of spiking neurons. IEEE Transactions on neural networks, 14(6): 1569–1572, 2003. Krizhevsky, Alex and Hinton, Geoffrey. Learning multiple layers of features from tiny images. 2009. Krizhevsky, Alex, Sutskever, Ilya, and Hinton, Geoffrey E. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. LeCun, Yann, Bottou, Léon, Bengio, Yoshua, and Haffner, Patrick. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278– 2324, 1998. Merolla, Paul A, Arthur, John V, Alvarez-Icaza, Rodrigo, Cassidy, Andrew S, Sawada, Jun, Akopyan, Filipp, Jackson, Bryan L, Imam, Nabil, Guo, Chen, Nakamura, Yutaka, et al. A million spiking-neuron integrated circuit with a scalable communication network and interface. Science, 345(6197):668–673, 2014. Pérez-Carrasco, José Antonio, Zhao, Bo, Serrano, Carmen, Acha, Begona, Serrano-Gotarredona, Teresa, Chen, Shouchun, and Linares-Barranco, Bernabé. Mapping from Frame-Driven to Frame-Free Event-Driven Vision Systems by Low-Rate Rate Coding and Coincidence Processing–Application to Feedforward ConvNets. IEEE transactions on pattern analysis and machine intelligence, 35(11):2706–2719, 2013. Posch, Christoph, Matolin, Daniel, and Wohlgenannt, Rainer. A QVGA 143 dB dynamic range frame-free Going Deeper in Spiking Neural Networks: VGG and Residual Architectures PWM image sensor with lossless pixel-level video compression and time-domain CDS. IEEE Journal of SolidState Circuits, 46(1):259–275, 2011. Posch, Christoph, Serrano-Gotarredona, Teresa, LinaresBarranco, Bernabe, and Delbruck, Tobi. Retinomorphic event-based vision sensors: bioinspired cameras with spiking output. Proceedings of the IEEE, 102(10):1470– 1484, 2014. Russakovsky, Olga, Deng, Jia, Su, Hao, Krause, Jonathan, Satheesh, Sanjeev, Ma, Sean, Huang, Zhiheng, Karpathy, Andrej, Khosla, Aditya, Bernstein, Michael, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015. Simonyan, Karen and Zisserman, Andrew. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. Srivastava, Nitish, Hinton, Geoffrey E, Krizhevsky, Alex, Sutskever, Ilya, and Salakhutdinov, Ruslan. Dropout: a simple way to prevent neural networks from overfitting. Journal of machine learning research, 15(1):1929–1958, 2014. Szegedy, Christian, Liu, Wei, Jia, Yangqing, Sermanet, Pierre, Reed, Scott, Anguelov, Dragomir, Erhan, Dumitru, Vanhoucke, Vincent, and Rabinovich, Andrew. Going deeper with convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 1–9, 2015. Zhao, Bo, Ding, Ruoxi, Chen, Shoushun, LinaresBarranco, Bernabe, and Tang, Huajin. Feedforward categorization on AER motion events using cortex-like features in a spiking neural network. IEEE transactions on neural networks and learning systems, 26(9):1963– 1978, 2015.
1cs.CV
arXiv:1702.00561v1 [math.GR] 2 Feb 2017 Autocommuting probability of a finite group Parama Dutta and Rajat Kanti Nath∗ Department of Mathematical Sciences, Tezpur University, Napaam-784028, Sonitpur, Assam, India. Emails: [email protected] and [email protected] Abstract Let G be a finite group and Aut(G) the automorphism group of G. The autocommuting probability of G, denoted by Pr(G, Aut(G)), is the probability that a randomly chosen automorphism of G fixes a randomly chosen element of G. In this paper, we study Pr(G, Aut(G)) through a generalization. We obtain a computing formula, several bounds and characterizations of G through Pr(G, Aut(G)). We conclude the paper by showing that the generalized autocommuting probability of G remains unchanged under autoisoclinism. Key words: Automorphism group, Autocommuting probability, Autoisoclinism. 2010 Mathematics Subject Classification: 20D60, 20P05, 20F28. 1 Introduction Let G be a finite group and Aut(G) be its automorphism group. For any x ∈ G and α ∈ Aut(G) the element x−1 α(x), denoted by [x, α], is called an autocommutator in G. We write S(G, Aut(G)) to denote the set {[x, α] : x ∈ G and α ∈ Aut(G)} and K(G) := hS(G, Aut(G))i. We also write L(G) := {x : [x, α] = 1 for all α ∈ Aut(G)}. Note that K(G) and L(G) are characteristic subgroups of G known as the autocommutator subgroup and absolute center of G. It is easy to see that K(G) contains the commutator subgroup G′ and L(G) is contained in the center Z(G) of G. Further, L(G) = ∩ CG (α), where CG (α) = {x ∈ G : [x, α] = 1} is a subgroup of α∈Aut(G) G known as acentralizer of α ∈ Aut(G). The subgroups K(G) and L(G) were defined and studied by Hegarty in [6]. Let CAut(G) (x) := {α ∈ Aut(G) : α(x) = x} for x ∈ G then CAut(G) (x) is a subgroup of Aut(G) and CAut(G) (G) = {α ∈ Aut(G) : α(x) = x for all x ∈ G}. It follows that CAut(G) (G) = ∩ CAut(G) (x). x∈G ∗ Corresponding author 1 Let G be a finite group acting on a set Ω. In the year 1975, Sherman [13] introduced the probability that a randomly chosen element of Ω fixes a randomly chosen element of G. We denote this probability by Pr(G, Ω). If Ω = Aut(G) then Pr(G, Aut(G)) is nothing but the probability that the autocommutator of a randomly chosen pair of elements, one from G and the other from Aut(G), is equal to 1 (the identity element of G). Thus Pr(G, Aut(G)) = |{(x, α) ∈ G × Aut(G) : [x, α] = 1}| . |G|| Aut(G)| (1.1) Pr(G, Aut(G)) is called autocommuting probability of G. Sherman [13] considered the case when G is abelian and Ω = Aut(G). Note that if we take Ω = Inn(G), the inner automorphim group of G, then Pr(G, Inn(G)) gives the probability that a randomly chosen pair of elements of G commute. Pr(G, Inn(G)) is known as commuting probability of G. It is also denoted by Pr(G). The study of Pr(G) was initiated by Erdös and Turán [4]. After Erdös and Turán many authors have worked on Pr(G) and its generalizations (conf. [3] and the references therein). Somehow the study of Pr(G, Aut(G)) was neglected for long time. At this moment we have only a handful of papers on Pr(G, Aut(G)) (see [2], [8], [12]). In this paper, we study Pr(G, Aut(G)) through a generalization. In the year 2008, Pournaki and Sobhani [10] have generalized Pr(G) and considered the probability that the commutator of a randomly chosen pair of elements of G equals a given element g ∈ G. We write Prg (G) to denote this probability. Motivated by their work, in this paper, we consider the probability that the autocommutator of a randomly chosen pair of elements, one from G and the other from Aut(G), is equal to a given element g ∈ G. We write Prg (G, Aut(G)) to denote this probability. Thus, Prg (G, Aut(G)) = |{(x, α) ∈ G × Aut(G) : [x, α] = g}| . |G|| Aut(G)| (1.2) Notice that Pr1 (G, Aut(G)) = Pr(G, Aut(G)). Hence Prg (G, Aut(G)) is a generalization of Pr(G, Aut(G)). Clearly, Prg (G, Aut(G)) = 1 if and only if K(G) = {1} and g = 1 if and only if G = L(G) and g = 1. Also, Prg (G, Aut(G)) = 0 if and only if g ∈ / S(G, Aut(G)). Therefore, we consider G 6= L(G) and g ∈ S(G, Aut(G)) throughout the paper. In Section 2, we obtain a computing formula for Prg (G, Aut(G)) and deduce some of its consequences. In Section 3, we obtain several bounds for Prg (G, Aut(G)). In Section 4, we obtain some characterizations of G through Pr(G, Aut(G)). Finally, in the last section, we show that Prg (G, Aut(G)) is an invariant under autoisoclinism of groups. 2 A computing formula For any x, g ∈ G let Tx,g denotes the set {α ∈ Aut(G) : [x, α] = g}. Note that Tx,1 = CAut(G) (x). The following two lemmas play a crucial role in obtaining the computing formula for Prg (G, Aut(G)). Lemma 2.1. Let G be a finite group. If Tx,g 6= φ then Tx,g = σCAut(G) (x) for some σ ∈ Tx,g . Hence, |Tx,g | = |CAut(G) (x)|. 2 Proof. Let σ ∈ Tx,g and β ∈ σCAut(G) (x). Then β = σα for some α ∈ CAut(G) (x). We have [x, β] = [x, σα] = x−1 σ(α(x)) = [x, σ] = g. Therefore, β ∈ Tx,g and so σCAut(G) (x) ⊆ Tx,g . Again, let γ ∈ Tx,g then γ(x) = xg. We have σ −1 γ(x) = σ −1 (xg) = x and so σ −1 γ ∈ CAut(G) (x). Therefore, γ ∈ σCAut(G) (x) which gives Tx,g ⊆ σCAut(G) (x). Hence, the result follows. We know that Aut(G) acts on G by the action (α, x) 7→ α(x) where α ∈ Aut(G) and x ∈ G. Let orb(x) := {α(x) : α ∈ Aut(G)} be the orbit of x ∈ G. Then by orbit-stabilizer theorem, we have | orb(x)| = | Aut(G)| . |CAut(G) (x)| (2.1) Lemma 2.2. Let G be a finite group. Then Tx,g 6= φ if and only if xg ∈ orb(x). Proof. The result follows from the fact that α ∈ Tx,g if and only if xg ∈ orb(x). Now we state and prove the main theorem of this section. Theorem 2.3. Let G be a finite group. If g ∈ G then Prg (G, Aut(G)) = 1 |G|| Aut(G)| X |CAut(G) (x)| = x∈G xg∈orb(x) Proof. We have {(x, α) ∈ G × Aut(G) : [x, α] = g} = 1 |G| X x∈G xg∈orb(x) 1 . | orb(x)| ⊔ ({x} × Tx,g ), where ⊔ x∈G represents the union of disjoint sets. Therefore, by (1.2), we have X |G|| Aut(G)|Prg (G, Aut(G)) = | ⊔ ({x} × Tx,g )| = |Tx,g |. x∈G x∈G Hence, the result follows from Lemma 2.1, Lemma 2.2 and (2.1). Taking g = 1, in Theorem 2.3, we get the following corollary. Corollary 2.4. Let G be a finite group. Then Pr(G, Aut(G)) = X 1 | orb(G)| |CAut(G) (x)| = |G|| Aut(G)| x∈G |G| where orb(G) = {orb(x) : x ∈ G}. Corollary 2.5. Let G be a finite group. If CAut(G) (x) = {I} for all x ∈ G \ {1}, where I is the identity element of Aut(G), then Pr(G, Aut(G)) = 1 1 1 + − . |G| | Aut(G)| |G|| Aut(G)| Proof. By Corollary 2.4, we have |G|| Aut(G)| Pr(G, Aut(G)) = X |CAut(G) (x)| = | Aut(G)| + |G| − 1. x∈G Hence, the result follows. 3 P We also have |{(x, α) ∈ G × Aut(G) : [x, α] = 1}| = |CG (α)| and hence α∈Aut(G) Pr(G, Aut(G)) = 1 |G|| Aut(G)| X |CG (α)|. (2.2) α∈Aut(G) We conclude this section with the following two results. Proposition 2.6. Let G be a finite group. If g ∈ G then Prg−1 (G, Aut(G)) = Prg (G, Aut(G)). Proof. Let A = {(x, α) ∈ G × Aut(G) : [x, α] = g} and B = {(y, β) ∈ G × Aut(G) : [y, β] = g −1 }. Then (x, α) 7→ (α(x), α−1 ) gives a bijection between A and B. Therefore |A| = |B|. Hence the result follows from (1.2). Proposition 2.7. Let G and H be two finite groups such that gcd(|G|, |H|) = 1. If (g, h) ∈ G × H then Pr(g,h) (G × H, Aut(G × H)) = Prg (G, Aut(G))Prh (H, Aut(H)). Proof. Let X = {((x, y), αG×H ) ∈ (G × H) × Aut(G × H) : [(x, y), αG×H ] = (g, h)}, Y = {(x, αG ) ∈ G × Aut(G) : [x, αG ] = g} and Z = {(y, αH ) ∈ H × Aut(H) : [y, αH ] = h}. Since gcd(|G|, |H|) = 1, by [1, Lemma 2.1], we have Aut(G × H) = Aut(G) × Aut(H). Therefore, for every αG×H ∈ Aut(G × H) there exist unique αG ∈ Aut(G) and αH ∈ Aut(H) such that αG×H = αG × αH , where αG × αH ((x, y)) = (αG (x), αH (y)) for all (x, y) ∈ G × H. Also, for all (x, y) ∈ G × H, we have [(x, y), αG×H ] = (g, h) if and only if [x, αG ] = g and [y, αH ] = h. These leads to show that X = Y × Z. Therefore |X | |Y| |Z| = · . |G × H|| Aut(G × H)| |G|| Aut(G)| |H|| Aut(H)| Hence, the result follows from (1.1). 3 Some bounds We begin with the following lower bounds. Proposition 3.1. Let G be a finite group. Then |CAut(G) (G)|(|G|−|L(G)|) |G|| Aut(G)| (a) Prg (G, Aut(G)) ≥ |L(G)| |G| (b) Prg (G, Aut(G)) ≥ |L(G)||CAut(G) (G)| |G|| Aut(G)| + if g 6= 1. 4 if g = 1. Proof. Let C := {(x, α) ∈ G × Aut(G) : [x, α] = g}. (a) We have (L(G) × Aut(G)) ∪ (G × CAut(G) (G)) ⊆ C and |(L(G) × Aut(G)) ∪ (G × CAut(G) (G))| is equal to |L(G)|| Aut(G)| + |CAut(G) (G)||G| − |L(G)||CAut(G) (G)|. Hence, the result follows from (1.2). (b) Since g ∈ S(G, Aut(G)) we have C is non-empty. Let (y, β) ∈ C then (y, β) ∈ / L(G) × CAut(G) (G) otherwise [y, β] = 1. It is easy to see that the coset (y, β)(L(G) × CAut(G) (G)) having order |L(G)||CAut(G) (G)| is a subset of C. Hence, the result follows from (1.2). Proposition 3.2. Let G be a finite group. Then Prg (G, Aut(G)) ≤ Pr(G, Aut(G)). The equality holds if and only if g = 1. Proof. By Theorem 2.3, we have Prg (G, Aut(G)) = ≤ 1 |G|| Aut(G)| X |CAut(G) (x)| x∈G xg∈orb(x) X 1 |CAut(G) (x)| = Pr(G, Aut(G)). |G|| Aut(G)| x∈G The equality holds if and only if xg ∈ orb(x) for all x ∈ G if and only if g = 1. Proposition 3.3. Let G be a finite group and p the smallest prime dividing | Aut(G)|. If g 6= 1 then 1 |G| − |L(G)| < . Prg (G, Aut(G)) ≤ p|G| p Proof. By Theorem 2.3, we have Prg (G, Aut(G)) = 1 |G| X x∈G\L(G) xg∈orb(x) 1 | orb(x)| (3.1) noting that for x ∈ L(G) we have xg ∈ / orb(x). Also, for x ∈ G \ L(G) and xg ∈ orb(x) we have | orb(x)| > 1. Since | orb(x)| is a divisor of | Aut(G)| we have | orb(x)| ≥ p. Hence, the result follows from (3.1). The remaining part of this section is devoted in obtaining some lower and upper bounds for Pr(G, Aut(G)). The following theorem is an improvement of [8, Theorem 2.3 (ii)]. Theorem 3.4. Let G be a finite group and p the smallest prime dividing | Aut(G)|. Then |L(G)| p(|G| − |XG | − |L(G)|) + |XG | Pr(G, Aut(G)) ≥ + |G| |G|| Aut(G)| and (p − 1)|L(G)| + |G| |XG |(| Aut(G)| − p) Pr(G, Aut(G)) ≤ − , p|G| p|G|| Aut(G)| where XG = {x ∈ G : CAut(G) (x) = {I}}. 5 Proof. We have XG ∩ L(G) = φ. Therefore X |CAut(G) (x)| = |XG | + | Aut(G)||L(G)| + x∈G X |CAut(G) (x)|. x∈G\(XG ∪L(G)) For x ∈ G \ (XG ∪ L(G)) we have {I} 6= CAut(G) (x) 6= Aut(G) which implies p ≤ |CAut(G) (x)| ≤ | Aut(G)| . Therefore p X |CAut(G) (x)| ≥ |XG | + | Aut(G)||L(G)| + p(|G| − |XG | − |L(G)|) (3.2) x∈G and X |CAut(G) (x)| ≤ |XG | + | Aut(G)||L(G)| + x∈G | Aut(G)|(|G| − |XG | − |L(G)|) . (3.3) p Hence, the result follows from Corollary 2.4, (3.2) and (3.3). We have the following two corollaries. Corollary 3.5. Let G be a finite group. If p and q are the smallest primes dividing | Aut(G)| and |G| respectively then Pr(G, Aut(G)) ≤ In particular, if p = q then Pr(G, Aut(G)) ≤ p+q−1 . pq 2p−1 p2 ≤ 43 . Proof. Since G 6= L(G) we have |G : L(G)| ≥ q. Therefore, by Theorem 3.4, we have   p+q−1 p−1 1 +1 ≤ . Pr(G, Aut(G)) ≤ p |G : L(G)| pq Corollary 3.6. Let G be a finite group and p, q be the smallest primes dividing | Aut(G)| and |G| respectively. If G is non-abelian then Pr(G, Aut(G)) ≤ In particular, if p = q then Pr(G, Aut(G)) ≤ q2 + p − 1 . pq 2 p2 +p−1 p3 ≤ 58 . Proof. Since G is non-abelian we have |G : L(G)| ≥ q 2 . Therefore, by Theorem 3.4, we have   1 q2 + p − 1 p−1 Pr(G, Aut(G)) ≤ +1 ≤ . p |G : L(G)| pq 2 Theorem 3.7. Let G be a finite group. Then   1 |S(G, Aut(G))| − 1 Pr(G, Aut(G)) ≥ 1+ . |S(G, Aut(G))| |G : L(G)| The equality holds if and only if orb(x) = xS(G, Aut(G)) for all x ∈ G \ L(G). 6 Proof. For all x ∈ G \ L(G) we have α(x) = x[x, α] ∈ xS(G, Aut(G)). Therefore orb(x) ⊆ xS(G, Aut(G)) and so | orb(x)| ≤ |S(G, Aut(G))| for all x ∈ G \ L(G). Now, by Corollary 2.4, we have   X 1  X 1 1  Pr(G, Aut(G)) = + |G| | orb(x)| | orb(x)| x∈G\L(G) x∈L(G) |L(G)| 1 ≥ + |G| |G| X x∈G\L(G) 1 . |S(G, Aut(G))| Hence, the result follows. The following lemma is useful in obtaining the next corollary. Lemma 3.8. Let G be a finite group. Then, for any two integers m ≥ n, we have     n−1 1 m−1 1 1+ ≥ 1+ . n |G : L(G)| m |G : L(G)| If L(G) 6= G then equality holds if and only if m = n. Proof. The proof is an easy exercise. Corollary 3.9. Let G be a finite group. Then   |K(G)| − 1 1 1+ . Pr(G, Aut(G)) ≥ |K(G)| |G : L(G)| If G 6= L(G) then the equality holds if and only if K(G) = S(G, Aut(G)) and orb(x) is equal to S(G, Aut(G)) for all x ∈ G \ L(G). Proof. Since |K(G)| ≥ |S(G, Aut(G))|, the result follows from Theorem 3.7 and Lemma 3.8. Note that the equality holds if and only if equality holds in Theorem 3.7 and Lemma 3.8. This lower bound is also obtained in [2]. Note that the lower bound in Theorem 3.7 is better than that in Corollary 3.9. Also   p(|G| − |L(G)|) |K(G)| − 1 |L(G)| 1 + . 1+ ≥ |K(G)| |G : L(G)| |G| |G|| Aut(G)| Hence, the lower bound in Theorem 3.7 is also better than that in [8, Theorem 2.3(ii)]. 4 Some Characterizations In [2], Arora and Karan obtain a characterization of a finite group G if equality holds in Corollary 3.9. In this section, we obtain some characterizations of G if equality holds in Corollary 3.5 and Corollary 3.6. We begin with the following result. 7 Proposition 4.1. Let G be a finite group with Pr(G, Aut(G)) = p+q−1 for some pq primes p and q. Then pq divides |G|| Aut(G)|. If p and q are the smallest primes dividing | Aut(G)| and |G| respectively, then G ∼ = Zq . L(G) In particular, if G and Aut(G) are of even order and Pr(G, Aut(G)) = Z2 . 3 4 then G L(G) ∼ = Proof. By (1.1), we have (p + q − 1)|G|| Aut(G)| = pq|{(x, α) ∈ G × Aut(G) : [x, α] = 1}|. Therefore, pq divides |G|| Aut(G)|. If p and q are the smallest primes dividing | Aut(G)| and |G| respectively then, by Theorem 3.4, we have   p+q−1 1 p−1 ≤ +1 pq p |G : L(G)| which gives |G : L(G)| ≤ q. Hence, G L(G) ∼ = Zq . 2 +p−1 Proposition 4.2. Let G be a finite non-abelian group with Pr(G, Aut(G)) = q pq 2 for some primes p and q. Then pq divides |G|| Aut(G)|. If p and q are the smallest primes dividing | Aut(G)| and |G| respectively then G ∼ = Zq × Zq . L(G) In particular, if G and Aut(G) are of even order and Pr(G, Aut(G)) = Z2 × Z2 . 5 8 then G L(G) ∼ = Proof. By (1.1), we have (q 2 + p − 1)|G|| Aut(G)| = pq 2 |{(x, α) ∈ G × Aut(G) : [x, α] = 1}|. Therefore, pq divides |G|| Aut(G)|. If p and q are the smallest primes dividing | Aut(G)| and |G| respectively then, by Theorem 3.4, we have   1 p−1 q2 + p − 1 ≤ + 1 pq 2 p |G : L(G)| which gives |G : L(G)| ≤ q 2 . Since G is non-abelian we have |G : L(G)| = 6 1, q. Hence, G ∼ Z × Z . = q q L(G) We conclude this section with the following partial converses of Proposition 4.1 and 4.2. Proposition 4.3. Let G be a finite group. Let p, q be the smallest prime divisors of | Aut(G)| and |G| respectively and | Aut(G) : CAut(G) (x)| = p for all x ∈ G \ L(G). (a) If G ∼ = Zq then Pr(G, Aut(G)) = p+q−1 . L(G) (b) If G L(G) pq ∼ = Zq × Zq then Pr(G, Aut(G)) = 8 q 2 +p−1 . pq 2 Proof. Since | Aut(G) : CAut(G) (x)| = p for all x ∈ G \ L(G) we have |CAut(G) (x)| = | Aut(G)| for all x ∈ G \ L(G). Therefore, by Corollary 2.4, we have p Pr(G, Aut(G)) = = 1 |L(G)| + |G| |G|| Aut(G)| |CAut(G) (x)| x∈G\L(G) |L(G)| |G| − |L(G)| + . |G| p|G| Thus Pr(G, Aut(G)) = 1 p   p−1 +1 . |G : L(G)| (4.1) ∼ . = Zq then (4.1) gives Pr(G, Aut(G)) = p+q−1 pq q 2 +p−1 ∼ = Zq × Zq then (4.1) gives Pr(G, Aut(G)) = pq2 . (b) If G L(G) G L(G) 5 Autoisoclinism of groups (a) If X In the year 1940, Hall defined isoclinism between two groups (see [5]). In the year 1995, Lescot [7] showed that the commuting probability of two isoclinic finite groups are same. Later on Pournaki and Sobhani [10] showed that Prg (G) = Prβ(g) (H) if (α, β) is an isoclinism between the finite groups G and H. Following Hall, Moghaddam et al. [9] have defined autoisoclinism between two groups. Recall that two groups G G H and H are said to be autoisoclinic if there exist isomorphisms ψ : L(G) → L(H) , β : K(G) → K(H) and γ : Aut(G) → Aut(H) such that the following diagram commutes G L(G) ψ×γ × Aut(G) − −−−−→  a(G,Aut(G)) y K(G) where the maps a(G,Aut(G)) : Aut(H) → K(H) are given by G L(G) β − −−−−→ H L(H) × Aut(H)  a(H,Aut(H)) y K(H) × Aut(G) → K(G) and a(H,Aut(H)) : H L(H) × a(G,Aut(G)) (xL(G), α1 ) = [x, α1 ] and a(H,Aut(H)) (yL(H), α2 ) = [y, α2 ] respectively. In this case, the pair (ψ × γ, β) is called an autoisoclinism between the groups G and H. Recently, in [12], Rismanchain and Sepehrizadeh have shown that Pr(G, Aut(G)) = Pr(H, Aut(H)) if G and H are autoisoclinic finite groups. We conclude this paper with the following generalization of [12, Lemma 2.5]. Theorem 5.1. Let G and H be two finite groups and (ψ × γ, β) an autoisoclinism between them. Then Prg (G, Aut(G)) = Prβ(g) (H, Aut(H)). G Proof. Let Sg = {(xL(G), α1 ) ∈ L(G) ×Aut(G) : [x, α1 ] = g} and Tβ(g) = {(yL(H), α2 ) H ∈ L(H) × Aut(H) : [y, α2 ] = β(g)}. Since (ψ × γ, β) an autoisoclinism between G and 9 H, the mapping λ : Sg → Tβ(g) given by (xL(G), α1 ) 7→ (ψ(xL(G)), γ(α1 )) gives an one to one correspondence between Sg and Tβ(g) . Hence, |Sg | = |Tβ(g) |. Also |{(x, α1 ) ∈ G × Aut(G) : [x, α1 ] = g}| = |L(G)||Sg | (5.1) |{(y, α2 ) ∈ H × Aut(H) : [y, α2 ] = β(g)}| = |L(H)||Tβ(g) |. (5.2) and Therefore, by (1.1) and (5.1), we have Prg (G, Aut(G)) = |L(G)||Sg | . |G|| Aut(G)| Since |G/L(G)| = |H/L(H)| and | Aut(G)| = | Aut(H)| we have Prg (G, Aut(G)) = |L(H)||Tβ(g) | . |H|| Aut(H)| Hence, the result follows from (5.2) and (1.1). References [1] J. H. Christopher and L. R. Darren, Automorphism of finite abelian groups, Amer. Math. Monthly, 114, 917–923. [2] H. Arora and R. Karan, What is the probability an automorphism fixes a group element?, Comm. Algebra, 45(3), 1141–1150 (2017). [3] A. K. Das, R. K. Nath and M. R. Pournaki, A survey on the estimation of commutativity in finite groups, Southeast Asian Bull. Math., 37(2), 161–180 (2013). [4] P. Erdös and P. Turán, On some problems of a statistical group-theory. IV, Acta. Math. Acad. Sci. Hungar., 19, 413–435 (1968). [5] P. Hall, The classification of prime power groups, J. Reine Angew. Math., 182, 130–141 (1940). [6] P. V. Hegarty, The absolute centre of a group, J. Algebra, 169(3), 929–935 (1994). [7] P. Lescot, Isoclinism classes and commutativity degrees of finite groups, J. Algebra, 177, 847–869 (1995). [8] M. R. R. Moghaddam, F. Saeedi and E. Khamseh, The probability of an automorphism fixing a subgroup element of a finite group, Asian-Eur. J. Math. 4(2), 301308 (2011). [9] M. R. R. Moghaddam, M. J. Sadeghifard and M. Eshrati, Some properties of autoisoclinism of groups, Fifth International group theory conference, Islamic Azad University, Mashhad, Iran, 13-15 March 2013. [10] M. R. Pournaki and R. Sobhani, Probability that the commutator of two group elements is equal to a given element, J. Pure Appl. Algebra, 212, 727–734 (2008). [11] R. Rezaei, A. Erfanian and P. Lescot, A note on the relative commutativity degree of finite groups, Asian-Eur. J. Math., 7 (2014), DOI: 10.1142/S179355711450017X. [12] M. R. Rismanchian and Z. Sepehrizadeh, Autoisoclinism classes and autocommutativity degrees of finite groups, Hacet. J. Math. Stat. 44(4), 893–899 (2015). [13] G. J. Sherman, What is the probability an automorphism fixes a group element?, Amer. Math. Monthly, 82, 261–264 (1975). 10
4math.GR
A tight lower bound for Vertex Planarization on graphs of bounded treewidth Marcin Pilipczuk∗ arXiv:1511.08283v1 [cs.DS] 26 Nov 2015 Abstract In the Vertex Planarization problem one asks to delete the minimum possible number of vertices from an input graph to obtain a planar graph. The parameterized complexity of this problem, parameterized by the solution size (the number of deleted vertices) has recently attracted significant attention. The state-of-the-art algorithm of Jansen, Lokshtanov, and Saurabh [SODA 2014] runs in time 2O(k log k) · n on n-vertex graph with a solution of size k. It remains open if one can obtain a single-exponential dependency on k in the running time bound. One of the core technical contributions of the work of Jansen, Lokshtanov, and Saurabh is an algorithm that solves a weighted variant of Vertex Planarization in time 2O(w log w) · n on graphs of treewidth w. In this short note we prove that the running time of this routine is tight under the Exponential Time Hypothesis, even in unweighted graphs and when parameterizing by treedepth. Consequently, it is unlikely that a potential single-exponential algorithm for Vertex Planarization parameterized by the solution size can be obtained by merely improving upon the aforementioned bounded treewidth subroutine. 1 Introduction In the Vertex Planarization problem, given an undirected graph G and an integer k, our goal is to delete at most k vertices from the graph G to obtain a planar graph. If (G, k) is a YES-instance to Vertex Planarization, then we say that G is a k-apex graph. Since many algorithms for planar graphs can be easily generalized to near-planar graphs — k-apex graphs for small values of k — this motivates us to look for efficient algorithms to recognize k-apex graphs. In other words, we would like to solve Vertex Planarization for small values of k. By a classical result of Lewis and Yannakakis [8], Vertex Planarization is NP-hard when k is part of the input. Since one can check if a given graph is planar in linear time [4], Vertex Planarization can be trivially solved in time O(nk+1 ), where n = |V (G)|, that is, in polynomial time for every fixed value of k. However, such an algorithm is impractical even for small values of k; a question for a faster algorithm brings us to the realms of parameterized complexity. In the parameterized complexity, every problem comes with a parameter, being an additional complexity measure of input instances. The central notion is a fixed-parameter algorithm: an algorithm that solves an instance x with parameter k in time f (k)|x|O(1) for some computable function f . Such a running time bound, while still super-polynomial (the function f is usually exponential), is considered significantly better than say O(|x|k ), as it promises much faster algorithms for moderate values of k and large instances. We refer to recent textbooks [1, 2] for a more broad introduction to parameterized complexity. Due to the aforementioned motivation, it is natural to consider the solution size k as a parameter for Vertex Planarization, and ask for a fixed-parameter algorithm. Since, for a fixed value of k, the class of all k-apex graphs is closed under taking minors, the graph minor theory of Roberston and Seymour immediately yields a fixed-parameter algorithm, but with enormous dependency on the parameter in the running time bound.1 The quest for an explicit and faster fixed-parameter algorithm for Vertex Planarization has attracted significant attention in the parameterized complexity community ∗ Institute of Informatics, University of Warsaw, Poland, [email protected] this algorithm is non-uniform, that is, it requires an external advice depending on the parameter only. However, we can obtain a uniform algorithm using the techniques of Fellows and Langston [3]. 1 Formally, 1 in the recent years. First, Marx and Schlotter [12] obtained a relatively simple algorithm, with doublyexponential dependency on the parameter and n2 dependency on the input size in the running time bound. Later, Kawarabayashi [7] obtained a fixed-parameter algorithm with improved linear dependency on the input size, at the cost of worse dependency on the parameter. Finally, Jansen, Lokshtanov, and Saurabh [6] developed an algorithm with running time bound 2O(k log k) · n, improving upon all previous results. As noted in [6], a simple reduction shows that Vertex Planarization cannot be solved in time 2o(k) · nO(1) unless the Exponential Time Hypothesis fails. Informally speaking, the Exponential Time Hypothesis (ETH) [5] asserts that the satisfiability of 3-CNF formulae cannot be verified in time subexponential in the number of variables. In the recent years, a number of tight bounds for fixed-parameter algorithms have been obtained using ETH or the closely related Strong ETH; we refer to [9, 11] for an overview. In this light, it is natural to ask for tight bounds for fixed-parameter algorithms for Vertex Planarization. In particular, [6] asks for a single-exponential (i.e., with running time bound 2O(k) nO(1) ) algorithm. The core subroutine of the algorithm of Jansen, Lokshtanov, and Saurabh, is an algorithm that solves Vertex Planarization in time 2O(w log w) · n on graphs of treewidth w. A direct way to obtain a single-exponential algorithm for Vertex Planarization parameterized by the solution size would be to improve the running time of this bounded treewidth subroutine to 2O(w) · nO(1) . In this short note we show that such an improvement is unlikely, as it would violate the Exponential Time Hypothesis. Theorem 1. Unless the Exponential Time Hypothesis fails, there does not exist an algorithm that solves Vertex Planarization on n-vertex graphs of treewidth at most w in time 2o(w log w) nO(1) . In fact, our lower bound holds even for a more restrictive parameter of treedepth, instead of treewidth. While Theorem 1 does not exclude the possibility of a 2O(k) nO(1) -time algorithm for Vertex Planarization, it shows that to obtain such a running time one needs to circumvent the usage of boundedtreewidth subroutine on graphs of treewidth Ω(k) in the algorithm of Jansen, Lokshtanov, and Saurabh. The remainder of this paper is devoted to the proof of Theorem 1. 2 Lower bound We base our reduction on the framework for proving superexponential lower bounds introduced by Lokshtanov, Marx, and Saurabh [10]. For an integer k, by [k] we denote the set {1, 2, . . . , k}. Consequently, [k] × [k] is a k × k table of elements with rows being subsets of the form {i} × [k], and columns being subsets of the form [k] × {i}. We start from the following auxiliary problem. k × k Permutation Clique Parameter: k Input: An integer k and a graph G with vertex set [k] × [k]. Question: Is there a k-clique in G with exactly one element from each row and exactly one element from each column? As proven in [10], an 2o(k log k) -time algorithm for k × k Permutation Clique would violate ETH. Hence, to prove Theorem 1, it suffices to prove the following. Lemma 2. There exists a polynomial time algorithm that, given an instance (G, k) of k × k Permutation Clique, outputs an equivalent instance (H, `) of Vertex Planarization where the treedepth of the graph H is bounded by O(k). That is, as announced in the introduction, we in fact prove a stronger variant of Theorem 1, refuting an existence of a 2o(w log w) nO(1) -time algorithm for Vertex Planarization parameterized by the treedepth of the input graph. Recall that the treedepth of a graph G, denoted td(G), is always not smaller than the treewidth of G, and satisfies the following recursive formula. Lemma 3 ([13]). The treedepth of an empty graph is 0, and the treedepth of a one-vertex graph equals 1. The treedepth of a disconnected graph G equals the maximum of the treedepth of the connected components of G. The treedepth of a connected graph G is equal to td(G) = 1 + min td(G − {v}). v∈V (G) 2 Figure 1: Choice gadget C4 . The vertices ai are black and the vertices bj are white. A minimum solution that leaves b3 undeleted is marked with dashed circles. We refer to the textbook [14] for more information on treedepth. The rest of this section is devoted to the proof of Lemma 2. 2.1 One-in-many gadget We begin with a description of a gadget that allows us to encode a choice among many options. Given two vertices x and y, by introducing a K5 -edge xy we mean the following operation: we introduce three new vertices z1 , z2 and z3 and make x, y, z1 , z2 , z3 a clique. Note that in every solution to Vertex Planarization, at least one of the vertices of the set {x, y, z1 , z2 , z3 } needs to be deleted. As we do not add any more edges incident to any vertex zi , i = 1, 2, 3, we may safely restrict ourselves to solutions to Vertex Planarization that contain x or y and do not contain any of the vertices zi , i = 1, 2, 3. That is, we treat the vertices zi as undeletable vertices, and henceforth by a “solution to Vertex Planarization” we mean a solution not containing any such vertex. For an integer s ≥ 1, we define an s-choice gadget Cs as follows. We start with 3s+2 vertices denoted ai for 0 ≤ i ≤ 2s + 1 and bj for 1 ≤ j ≤ s. Then, for each 0 ≤ i < 2s + 1 we introduce a K5 -edge ai ai+1 and for each 1 ≤ j < s we introduce two K5 -edges bj a2j−1 and bj a2j . Any choice gadget created in the construction will be attached to the rest of the graph using the vertices bj ; informally speaking, in any optimal solution, exactly one vertex bj remains undeleted. We summarize the properties the s-choice gadget in the following lemma; see Figure 1 for an illustration. Lemma 4. For an s-choice gadget Cs , the following holds. 1. A minimum solution to Vertex Planarization on Cs consists of 2s vertices. 2. For every 1 ≤ j ≤ s there exists a minimum solution X to Vertex Planarization on Cs that contains all vertices bj 0 for j 0 6= j. 3. In every minimum solution to Vertex Planarization on Cs , at least one vertex bj remains undeleted. 4. The treedepth of the s-choice gadget is O(log s). Furthermore, the same treedepth bound holds for a graph constructed from an s-choice gadget by, for every 1 ≤ j ≤ s, introducing a constant-size graph Gj and identifying one vertex of Gj with the vertex bj . Proof. First, note that for every 1 ≤ j ≤ s, the set {bj 0 : j 0 6= j} ∪ {a2j 0 −1 : 1 ≤ j 0 ≤ j} ∪ {a2j 0 : j ≤ j 0 ≤ s} is a solution to Vertex Planarization on Cs of size 2s that contains all vertices bj 0 except for bj . Moreover, observe that, due to K5 -edges, for every 1 ≤ j ≤ s, any solution to Vertex Planarization on Cs needs to delete at least two vertices from the set {a2j−1 , a2j , bj }, and consequently has size at least 2s. This settles the first two claims. For the third claim, note that any vertex cover of a path of length 2s + 1 needs to contain at least s + 1 vertices, and consequently a solution to Vertex Planarization on Cs \ {bj : 1 ≤ j ≤ s} needs to contain at least s + 1 vertices. Thus, any solution to Vertex Planarization on Cs that contains {bj : 1 ≤ j ≤ s} contains at least 2s + 1 vertices. This settles the third claim. We prove the last claim by induction on s. For s = O(1), the gadget and the attached graphs Gj are of constant size, and the treedepth is constant. Otherwise, for an s-choice gadget Cs , we delete the three vertices z1 , z2 , z3 from the K5 -edge between a2bs/2c and a2bs/2c+1 . Note that the gadget splits into two connected components, both being subgraphs of a graph in question constructed from a ds/2e-choice gadget. The treedepth bound O(log s) follows from Lemma 3. 3 f1 uL 1 uL 2 v01 v11 v21 v31 u10 u11 u12 u13 uR 1 uR 2 f2 uL 3 v02 v12 v22 v32 uR 3 uL 4 u20 u21 u22 u23 uR 4 f3 uL 5 uL 6 v03 v13 v23 v33 u30 u31 u32 u33 uR 5 uR 6 f4 Figure 2: Part of the frame graph HF with its unique embedding. 2.2 Construction We now give a construction of the Vertex Planarization instance (H, `), given a k ×k Permutation Clique instance (G, k). Let m = |E(G)| and assume k ≥ 2. First, we introduce a frame graph HF . A ladder of length n is a 2n-vertex graph that consists of two paths v1 , v2 , . . . , vn and u1 , u2 , . . . , un together with edges vi ui for 1 ≤ i ≤ n. A cycle ladder of length n additionally contains edges vn v1 and un u1 , that is, v1 , v2 , . . . , vn and u1 , u2 , . . . , un are in fact cycles. The frame graph HF consists of two cycle ladders of length 2k, with vertex sets {viΓ , uΓi : 1 ≤ i ≤ 2k} for Γ ∈ {L, R}, and of k ladders of length 4 with vertex sets {viα , uα i : 0 ≤ i ≤ 3} for 1 ≤ α ≤ k, L L α R R α u0 , v2α−1 v3α , v2α u3 for each 1 ≤ α ≤ k (see Figure 2). Note that HF is connected with edges v2α−1 v0α , v2α 3-edge-connected and hence has a unique planar embedding. By f α we denote the face of the embedding of HF that is incident to all vertices viα , 0 ≤ i ≤ 3. Second, we introduce k vertices xβ , 1 ≤ β ≤ k. Our intention is to ensure that in any solution to Vertex Planarization on (H, `), no vertex xβ nor no vertex from the frame HF will be deleted, and each vertex xβ will be embedded into a different face f α . The choice of which vertex xβ is embedded into which face will correspond to a choice of the vertices of the clique in the instance (G, k) of k × k Permutation Clique. We now force such a behavior with some gadgets. For each i = 1, 2, 3, perform the following construction. For every 1 ≤ α, β ≤ k, introduce a vertex yiα,β incident to viα and xβ . Moreover, for every 1 ≤ β ≤ k, introduce a k-choice gadget C i,β and for each 1 ≤ α ≤ k identify the vertex bα of C i,β with yiα,β . See Figure 3. Informally speaking, by the properties of the k-choice gadget, each vertex xβ needs to select one face f α that will contain it in the planar embedding. The fact that the construction is performed three times ensures that no face f α is chosen by two vertices xβ , as otherwise a K3,3 -minor will be left in the graph. Let us now move to the description of the encoding of the edges of G. For an edge e ∈ E(G), let e = (p(e), γ(e))(q(e), δ(e)) where p(e) < q(e) (note that edges e with p(e) = q(e) are irrelevant to the problem, and hence we may assume there are no such edges). For 1 ≤ p < q ≤ k, we define E(p, q) = {e ∈ E(G) : (p(e), q(e)) = (p, q)}. For each edge e ∈ E(G), we introduce in H three vertices 4 vi1 vi2 vi3 vi4 C i,4 C i,1 x1 C i,2 C i,3 x2 x3 x4 Figure 3: Connections between the frame and vertices xβ that ensure that every vertex xβ is embedded into a different face f α . Vertices yiα,β are depicted white. xγ(e) xδ(e) c` e ce p(e) ca e q(e) v0 v0 Figure 4: Gadget introduced for an edge e ∈ E(G). The vertex c`e , marked white, is part of the choice gadget Ĉ p(e),q(e) . p(e) q(e) c`e , ce , cae , four edges v0 c`e , xγ(e) c`e , v0 cae , xδ(e) cae and two K5 -edges c`e ce and cae ce (see Figure 4). Moreover, for every 1 ≤ p < q ≤ k, introduce a |E(p, q)|-choice gadget Ĉ p,q and for every e ∈ E(p, q) identify c`e with a distinct vertex bj of Ĉ p,q . Informally speaking, for each edge e we need either to delete c`e and cae or only ce ; however, the second option is only possible if xγ(e) is embedded into fp(e) and at the same time xδ(e) is embedded into fq(e) . The choice gadget Ĉ p,q ensures that we can choose the second, cheaper option only once per each pair (p, q). We set ` = 3m + 6k 2 . This completes the description of the instance (H, `). Note that the budget ` is tight: it allows only to choose a minimum solution in all introduced choice gadgets, and one endpoint of each K5 -edge cae ce . 2.3 Treedepth bound Lemma 5. The treedepth of H is O(k). Proof. We use the recursive formula of Lemma 3. First, we delete from H all vertices of the frame HF and all vertices xβ . Note that we have deleted only 17k vertices in this manner. By Lemma 3, it suffices to show that every connected component of the remaining graph has treedepth O(k); we will in fact show a stronger bound of O(log k). Observe that the remaining graph contains two types of connected components. The first type are the k-choice gadgets C i,β for 1 ≤ i ≤ 3 and 1 ≤ β ≤ k; by Lemma 4, every such gadget has treedepth O(log k). The second type are the |E(p, q)|-choice gadgets Ĉ p,q for 1 ≤ p < q ≤ k, together with the vertices c`e , ce , cae and the K5 -edges between them. As |E(p, q)| ≤ k 2 for every 1 ≤ p < q ≤ k, by Lemma 4 the treedepth of these connected components is also O(log k). This finishes the proof of the lemma. 2.4 Equivalence In the following two lemmata we show the equivalence of the constructed Vertex Planarization instance (H, `) and the input k × k Permutation Clique instance (G, k), completing the proof of Lemma 2 and of Theorem 1. Lemma 6 (Completeness). If (G, k) is a YES-instance to k × k Permutation Clique, then (H, `) is a YES-instance to Vertex Planarization. 5 Proof. Let ρ : [k] → [k] be a solution to k×k Permutation Clique on (G, k), that is, ρ is a permutation of [k] and K := {(p, ρ(p)) : 1 ≤ p ≤ k} is a clique in G. Consider the following set X ⊆ V (H). 1. For each i = 1, 2, 3 and 1 ≤ α ≤ k, X contains a minimum solution (i.e., of size 2k) to Vertex α0 ,ρ(α) Planarization in the gadget C i,ρ(α) that contains all vertices yi for 1 ≤ α0 ≤ k except for α,ρ(α) yi . 2. For each 1 ≤ p < q ≤ k, denote by e(p, q) the unique edge in E(p, q) such that e(p, q) = (p, ρ(p))(q, ρ(q)). Then X contains a minimum solution to Vertex Planarization in the gadget Ĉ p,q that contains all vertices c`e for e ∈ E(p, q) except for c`e(p,q) , the vertex ce(p,q) and all vertices cae for e ∈ E(p, q) except for cae(p,q) . Note that we have introduced 3 · k · 2k = 6k 2 vertices in the first step and 3m vertices in the second step. Hence, |X| = 3m + 6k 2 = `. We now argue that H \ X is planar. It suffices to prove it for each 2-connected component of H \ X. Note that the claim is trivial or follows from Lemma 4 for each 2-connected component of H \ X except for the one that contains the frame HF . Consider the unique planar embedding of HF and embed into each face f α the vertex xρ(α) . Note that each vertex xβ is embedded into a different face. It is straightforward to verify that xρ(α) can be α,ρ(α) embedded into f α together with vertices yi for i = 1, 2, 3 and the vertices c`e or cae that correspond to the edges of G[K] incident to (α, ρ(α)). Note that all other vertices of H \ X lie in different 2-connected components than HF and, consequently, H \ X is planar. Lemma 7 (Soundness). If (H, `) is a YES-instance to Vertex Planarization, then (G, k) is a YESinstance to k × k Permutation Clique. Proof. Let X ⊆ V (H) be such that |X| ≤ ` and H \ X is planar. By Lemma 4, X needs to contain at least 2k vertices from each gadget C i,β for i = 1, 2, 3 and 1 ≤ β ≤ k; note that there are 3k such gadgets. Moreover, X needs to contain at least 2|E(p, q)| vertices from each gadget Ĉ p,q for 1 ≤ p < q ≤ k, which totals to at least 2m vertices. Finally, for each 1 ≤ i ≤ m, X needs to contain at least one vertex of the K5 -edge ce cae . As |X| ≤ ` = 6k 2 + 3m, we infer that |X| = ` and X contains a minimum solution to Vertex Planarization on each introduced choice gadget, and exactly one vertex from the pair {ce , cae } for each 1 ≤ i ≤ m. In particular, X does not contain any vertex of the frame graph HF , nor any vertex xβ , 1 ≤ β ≤ k. α(i,β),β By the properties of the choice gadget C i,β , there exists α(i, β) such that yi ∈ / X. Recall that X does not contain any vertex of HF , and HF is 3-edge-connected and, hence, admits a unique planar embedding depicted on Figure 2. As X does not contain xβ , we infer that α(i, β) = α(i0 , β) for every i, i0 ∈ {1, 2, 3}. Hence, we may suppress the argument i and henceforth analyze function α(β) such that α(β),β yi ∈ / X for every 1 ≤ β ≤ k and i = 1, 2, 3. Note that xβ needs to be embedded into the face f α(β) of HF in any planar embedding of H \ X. We now argue that α(·) is a permutation. By contradiction, let α(β) = α(β 0 ) for some β 6= β 0 . It is straightforward to verify that the following sets form a model of a K3,3 minor in H \ X, contradicting its planarity. 0 1. {viα , yiα,β , yiα,β } for i = 1, 2, 3; 0 2. {xβ } and {xβ }; α α 3. {uα 1 , u2 , u3 }. We infer that α(·) is a permutation of [k]. We claim K := {(α(β), β) : 1 ≤ β ≤ k} induces a clique in G. That is, that ρ := α−1 is a solution to k × k Permutation Clique on (G, k). Pick arbitrary 1 ≤ p < q ≤ k. Our goal is to prove that (p, ρ(p))(q, ρ(q)) ∈ E(G). Consider the choice gadget Ĉ p,q . By Lemma 4, there exists e ∈ E(p, q) such that c`e ∈ / X. Consequently, ce ∈ X due to the K5 -edge c`e ce , and thus cae ∈ / X. As c`e is adjacent to both v0p and xγ(e) , we infer that xγ(e) is embedded into the face f p and, consequently, α(γ(e)) = p. Symmetrically, we infer that α(δ(e)) = q. Thus, e = (p, ρ(p))(q, ρ(q)) ∈ E(G) and the lemma is proven. 6 Acknowledgements We thank Tomasz Kociumaka for numerous discussions on the complexity of vertex deletion problems to minor-closed graph classes. Research supported by Polish National Science Centre grant DEC-2012/05/D/ST6/03214. References [1] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk, M. Pilipczuk, and S. Saurabh. Parameterized Algorithms. Springer, 2015. [2] R. G. Downey and M. R. Fellows. Fundamentals of Parameterized Complexity. Texts in Computer Science. Springer, 2013. [3] M. R. Fellows and M. A. Langston. On search, decision, and the efficiency of polynomial-time algorithms. J. Comput. Syst. Sci., 49(3):769–779, 1994. [4] J. E. Hopcroft and R. E. Tarjan. Efficient planarity testing. J. ACM, 21(4):549–568, 1974. [5] R. Impagliazzo, R. Paturi, and F. Zane. Which problems have strongly exponential complexity? J. Comput. Syst. Sci., 63(4):512–530, 2001. [6] B. M. P. Jansen, D. Lokshtanov, and S. Saurabh. A near-optimal planarization algorithm. In C. Chekuri, editor, Proceedings of the Twenty-Fifth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2014, Portland, Oregon, USA, January 5-7, 2014, pages 1802–1811. SIAM, 2014. [7] K. Kawarabayashi. Planarity allowing few error vertices in linear time. In 50th Annual IEEE Symposium on Foundations of Computer Science, FOCS 2009, October 25-27, 2009, Atlanta, Georgia, USA, pages 639–648. IEEE Computer Society, 2009. [8] J. M. Lewis and M. Yannakakis. The node-deletion problem for hereditary properties is NP-complete. J. Comput. Syst. Sci., 20(2):219–230, 1980. [9] D. Lokshtanov, D. Marx, and S. Saurabh. Lower bounds based on the exponential time hypothesis. Bulletin of the EATCS, 105:41–72, 2011. [10] D. Lokshtanov, D. Marx, and S. Saurabh. Slightly superexponential parameterized problems. In D. Randall, editor, SODA, pages 760–776. SIAM, 2011. [11] D. Marx. What’s next? Future directions in parameterized complexity. In H. L. Bodlaender, R. Downey, F. V. Fomin, and D. Marx, editors, The Multivariate Algorithmic Revolution and Beyond - Essays Dedicated to Michael R. Fellows on the Occasion of His 60th Birthday, volume 7370 of Lecture Notes in Computer Science, pages 469–496. Springer, 2012. [12] D. Marx and I. Schlotter. Obtaining a planar graph by vertex deletion. Algorithmica, 62(3-4):807–822, 2012. [13] J. Nešetřil and P. O. de Mendez. Tree-depth, subgraph coloring and homomorphism bounds. Eur. J. Comb., 27(6):1022–1041, 2006. [14] J. Nešetřil and P. O. de Mendez. Sparsity - Graphs, Structures, and Algorithms, volume 28 of Algorithms and combinatorics. Springer, 2012. 7
8cs.DS
arXiv:1711.02953v1 [math.GR] 8 Nov 2017 Classification of pro-p PD2 pairs and the pro-p curve complex Gareth Wilkes November 9, 2017 Abstract We classify pro-p Poincaré duality pairs in dimension two. We then use this classification to build a pro-p analogue of the curve complex and establish its basic properties. Introduction A key object in the study of self-homeomorphisms of a surface Σ is the curve complex, a simpicial complex whose simplices are isotopy classes of certain multicurves on Σ. When working with the curve complex one often simplifies matters by assuming that a class of multicurves is in some simple standard form, and asserts that this is acceptable by noting that every multicurve may be taken to a standard form by means of some self-homeomorphism of Σ. One proves this by cutting Σ along the multicurve and using the classification of surfaces with boundary. Thus the curve complex may be viewed as built using this classification as a foundation. In this paper we will carry out this program in the realm of pro-p groups. The notions of ‘relative (co)homology’ and ‘Poincaré duality pair’ for profinite groups were defined and developed in the author’s paper [Wil17a]. In the first part of this paper we will classify those pro-p group pairs which are Poincaré duality pairs of dimension 2. The end result (Theorem 3.3) is closely related to (and indeed relies upon) the classification of Demushkin groups completed by Demushkin, Serre, and Labute [Dem61, Dem63, Ser62, Lab67]. In Section 4 we use this classification to classify the ways that the pro-p completion G of π1 Σ (for Σ a compact orientable surface) may split as a graph of pro-p groups with edge stabilisers Zp and show that, up to automorphism, the only possible splittings are the obvious geometric ones. Considering splittings as being dual to ‘simple closed curves’ this analysis yields a ‘pro-p curve complex’— a profinite space, containing the abstract curve complex, on which OutpGq has a natural action. This raises the question of what properties of automorphisms can be deduced from the action on this pro-p curve complex—in particular, is it possible for a pseudo-Anosov automorphism of π1 Σ to fix a point in the pro-p curve complex? It is perhaps somewhat amusing to compare the order of this program with the classification of discrete PD2 groups and PD2 pairs. In the discrete realm, the PD2 groups were shown to be precisely the surface groups by Bieri, Eckmann, Müller and Linnell [BE78, EM80, EL83]. As noted in Section 11 of [BE78] 1 this implies the classification of PD2 pairs via the classification of splittings of surface groups. In our present case the PD2 pairs are classified first, and then applied to study splittings. The author would like to thank his supervisor Marc Lackenby for his support. The author was supported by the EPSRC. Conventions. In this paper we adopt the following conventions. • The symbol p will denote a prime number unless otherwise specified. The symbol Zp denotes the p-adic integers and Fp the field with p elements. • Any appearance of p8 for p a prime should be interpreted as zero. • All groups will be pro-p groups and all subgroups closed unless otherwise specified. All homomorphisms will be continuous. • For group elements x and y, the conjugate y ´1 xy will be denoted xy and the commutator rx, ys will mean x´1 y ´1 xy “ x´1 xy . • Let G be a pro-p group and let M be a finite p-primary G-module. Then M ˚ will denote the Pontrjagin dual HompM, Qp {Zp q with action pg ¨ f qpmq “ f pg ´1 mq. • The automorphism group AutpZp q will be denoted Up . • For a pro-p group G and a homomorphism ρ : G Ñ Up we will denote the the module with underlying abelian group Zp and G-action χ by Zp pχq. The module Z{pm with action given by ρ will be denoted Mm pρq and its Pontrjagin dual by Im pρq. 1 Preliminaries 1.1 Automorphisms of Zp In this section we set up some notation and preliminaries concerning automorphisms of the p-adic integers and their quotient groups. First recall the structure of the automorphism group. Theorem 1.1 (Theorem 4.4.7 of [RZ00]). Let p be a prime. • If p ‰ 2 then Up “ Zˆp – Cp´1 ˆ Zp where the second factor is generated by the automorphism 1 ÞÑ 1 ` p. • U2 “ Zˆ2 – C2 ˆ Z2 “ x´1y ˆ x1 ÞÑ 1 ` 4y Note that for p ‰ 2 the image of any homomorphism from a pro-p group to Up meets Cp´1 trivially. Following [Lab67] we give the following notation to the pro-p subgroups of these automorphism groups. For 1 ď k ď 8 an integer set k Upkq p “ x1 ÞÑ 1 ` p y ď Up , rk`1s U2 “ x1 ÞÑ ´1 ` 2k`1 y ď U2 Let G be a pro-p group and let χ : G Ñ Up be a homomorphism. Note that since Zp has a unique subgroup of index pm for any m there is an induced map from G to AutpZ{pm q. We may define an invariant qpχq to be the highest power 2 pm of p such that G Ñ AutpZ{pm q is trivial. If this maximum does not exist (i.e. χ is trivial) then by convention we set qpχq “ p8 “ 0. Note that for any pro-p group G and any χ : G Ñ Up the image of G in AutpFp q is trivial. pmq One may easily show that if qpχq “ pm ‰ 2 then impχq “ Up . If qpχq “ 2 rf s pf q then the image of χ may be either x´1y ˆ U2 or U2 where f ě 2 or f “ 8. p1q r2s Note that U2 “ U2 . 1.2 Lower central q-series Let G be a pro-p group and let q “ pm for some 1 ď m ď 8. The lower central q-series of G is the sequence of closed subgroups of G defined by G1 “ G, Gj`1 “ Gqj rGj , Gs We may define grj pGq “ Gj {Gj`1 , grpGq “ ÿ grj pGq jě1 The fact that rGi , Gj s Ď Gi`j implies that grpGq is a graded Lie algebra over Zp {qZp , with Lie bracket induced by the commutator on homogeneous elements of grpGq. If q ‰ 0 then x ÞÑ xq induces a linear mapping φi : grj pGq Ñ grj`1 pGq for all j. If π denotes an indeterminate over Zp {qZp we may let the polynomial ring Zp {qZp rπs act on grpGq by defining π ¨ ξ “ φj pξq when ξ P grj pGq. By convention we extend this to the q “ 0 case by setting π “ 0 P Zp . This makes grpGq into an algebra over ř Zp {qZp rπs. It is a Lie algebra provided q “ 0 or p ‰ 2. When p “ 2 only ją1 grj pGq is a Lie algebra. This issue will not be of concern to us. All the identities concerning grj pGq are derived from commutator identites. We will record some here for later use. The proofs are elementary. Let u, v, w P G and k P Z. • ruv, ws “ ru, wsrru, ws, vsrv, ws • ru, vws “ ru, wsru, vsrru, vs, ws • for u P Gi , v P Gj we have ruk , vs ” ru, vsk ” ru, v k s modulo Gi`j`1 . 1.3 PDn groups and PDn pairs It would be neither necessary nor expedient to give a full account of the theory of profinite PDn groups at a prime p. For a proper discussion for profinite groups see [SW00], and for pro-p groups see [Ser13], Section I.4.5. Depending on your viewpoint the following is either a definition or a theorem. Definition 1.2. Let G be a pro-p group. Then G is a PDn group if • dim H k pG, Fp q is finite for all k; • dim H n pG, Fp q “ 1; and • for all k the cup product H k pG, Fp q ˆ H n´k pG, Fp q Ñ H n pG, Fp q is a non-degenerate bilinear form. 3 One may show (see Proposition 30 of [Ser13] or Theorem 4.5.6 of [SW00]) that these conditions are enough to guarantee a more extensive duality as follows. There is a unique homomorphism χ : G Ñ Up called the orientation character with the following properties. Let Ip pGq “ pZp pχqq˚ denote Qp {Zp equipped with the G action dual to χ. Then H n pG, Ip pGqq “ Qp {Zp and the cup product H k pG, M q b H n´k pG, HompM, Ip pGqqq Ñ H n pG, Ip pGqq induces isomorphisms H k pG, M q – H n´k pG, HompM, Ip pGqqq˚ – Hn´k pG, Zp pχq b M q for any finite p-primary G-module M . Here the last isomorphism is induced by Pontrjagin duality (see for example Proposition 6.3.6 of [RZ00]), together with the identity HompM, Zp pχq˚ q “ pZp pχq b M q˚ The theory of the (co)homology of a profinite group relative to a collection of subgroups was defined and studied in [Wil17a]. We will just state a restricted version of the definition here and leave it to the curious reader to read further. Definition 1.3 (Definition 2.1 of [Wil17a]). Let G be a pro-p group and S “ tS0 , . . . , Sb u be a finite family of closed subgroups pb ě 0q. Consider the short exact sequence b à 0 Ñ ∆G,S Ñ Zp rrG{Si ss Ñ Zp Ñ 0 i“0 where the final map is the augmentation. Then define K H‚`1 pG, S; M q “ TorG ‚ p∆G,S , M q, H ‚`1 pG, S, Aq “ Ext‚G p∆G,S , Aq where M is an inverse limit of finite p-primary G-modules, A is a direct limit of such modules and ∆K G,S denotes ∆G,S when considered as a right module via δ ¨ g “ g ´1 δ. The definition can actually be made for certain infinite families of subgroups as well. However in the case of duality pairs the family of peripheral subgroups is forced to be finite (Proposition 5.4 of [Wil17a]) so we will not concern ourselves with this. If we replace each Si by a conjugate Sigi then this leaves the cohomology invariant up to natural isomorphism by Proposition 2.9 of [Wil17a]. This cohomology theory is natural with respect to maps of pro-p group pairs. If pG, Sq and pG1 , S 1 q are pro-p group pairs with S “ tS0 , . . . , Sb u and S 1 “ tS01 , . . . , Sb1 1 u, then a map of pairs pG, Sq Ñ pG1 , S 1 q consists of a function f : t1, . . . , bu Ñ t1, . . . , b1 u and a group homomorphism φ : G Ñ G1 such that φpSi q Ď Sf1 piq for all i. In this circumstance one obtains natural maps H‚ pG, S; M q Ñ H‚ pG1 , S 1 ; M q, H ‚ pG1 , S 1 ; M q Ñ H ‚ pG, S; M q for any G1 -module M , regarded as a G-module via φ. See Proposition 2.6 of [Wil17a]. Two important facts we will use freely are Pontrjagin duality H k pG, S; M ˚ q “ Hk pG, S; M q˚ 4 and the existence of a cup product H k pG, S; M q b H n´k pG, N q Ñ H n pG, S; Lq for (direct limits of) finite p-primary G-modules L, M and N and a pairing M b N Ñ L. Again the following definition is more properly a theorem. Definition 1.4 (Theorem 5.15 of [Wil17a]). Let G be a pro-p group and S a collection of closed subgroups of G. Then pG, Sq is a PDn pair if • dim H k pG, S; Fp q is finite for all k; • dim H n pG, S; Fp q “ 1; and • for all k the cup product H k pG, S; Fp q ˆ H n´k pG, Fp q Ñ H n pG, S; Fp q is a non-degenerate bilinear form. Again for a PDn pair pG, Sq there exits a unique orientation character χ : G Ñ Up with similar properties to above. Namely if Ip pG, Sq “ Zp pχq˚ denotes Qp {Zp equipped with the G action dual to χ, then H n pG, S; Ip pG, Sqq “ Qp {Zp and the cup product H k pG, S; M q b H n´k pG, HompM, Ip pG, Sqqq Ñ H n pG, S; Ip pG, Sqq with respect to the evaluation pairing induces isomorphisms H k pG, S; M q – H n´k pG, HompM, Ip pG, Sqqq˚ – Hn´k pG, Zp pχq b M q for any finite p-primary G-module M . See Section 5.1 of [Wil17a] for more details on PDn pairs. Let us now state the immediate consequences of being a PD2 pair. Proposition 1.5. Let pG, Sq be a PD2 pair with orientation character χ. Then G is a finitely generated free pro-p group and S is a finite family of infinite procyclic subgroups of G. Moreover χ vanishes on each element of S. Proof. By Proposition 5.4 of [Wil17a], S is a finite family S “ tS0 , . . . , Sb u with each Si a PD1 group with orientation character χ|Si . The only pro-p PD1 group is Zp , which is orientable and so each Si is cyclic with χ|Si “ 1. Furthermore by Corollary 5.8 in [Wil17a] we find cdp pGq “ 1, so by (for example) Theorem 7.7.4 of [RZ00] G is a free pro-p group. It is finitely generated since H 1 pG, Fp q is finite. Proposition 1.6. Let pG, Sq be a PD2 pair with S “ tS0 , . . . , Sb u. Fix a generator s0 of S0 . Then the rank of G is n ` b for some n ě 0 and we may choose a basis tx1 , . . . , xn , s1 , . . . , sb u of G such that • si is a generator of Si for all i • s0 ” s1 ¨ ¨ ¨ sb modulo Gp rG, Gs 5 • ζi ps0 q “ 1 for all i, where ζi is the projection G Ñ Zp obtained by killing all the elements of our chosen basis other than si . Proof. By the long exact sequence in relative cohomology and duality (Proposition 3.5, Theorem 5.6 and Proposition 5.7 of [Wil17a]) we have a commuting diagram (with coefficients in Fp ) H 0 pGq p1,...,1q H2 pG, Sq Àb i“1 H 0 pSi q Àb H1 pSi q i“0 – whence Àb – i“0 H1 pGq H1 pSi q injects into H1 pGq. Therefore there is a basis B “ tx1 , . . . , xn , s1 , . . . , sb u of G where si is a generator of Si and s0 ” s1 ¨ ¨ ¨ sb modulo Gp rG, Gs. It follows that modulo rG, Gs we have s0 ” n ź i ¨ xpµ i b ź sλi i i“1 i“1 where the λi lie in Zˆp . Replacing each si by the generator sλi i of Si we may guarantee that ζi ps0 q “ 1 for all i. Note that the properties in the conclusion of the above theorem do not change if we replace each si (and Si ) a conjugate of itself for 1 ď i ď b. We will be needing a tool to identify their orientation character of a PD2 pair. Proposition 1.7. Let pG, Sq be a PDn pair at the prime p and let ρ : G Ñ Up . Then ρ coincides with the orientation character χ of pG, Sq if and only if |H n pG, S; Im pρqq| “ pm for every m P N Proof. By duality we have H n pG, S; Im pρqq˚ – HomG pIm pρq, Ip pG, Sqq “ HomG pZp pχq, Mm pρqq “ HomG pMm pχq, Mm pρqq Now if χ “ ρ then every group homomorphism from Z{pm to itself is an element of HomG pMm pχq, Mm pρqq so this group has size pm . Conversely if the size is pm then the identity map on Z{pm is an element of HomG pMm pχq, Mm pρqq, whence ρ and χ agree modulo pm . If they agree modulo pm for all m then χ “ ρ. 1.4 Demushkin groups Definition 1.8. A pro-p group G is a Demushkin group if • dim H 1 pG, Fp q ď 8 • dim H 2 pG, Fp q “ 1 6 • the cup product H 1 pG, Fp q ˆ H 1 pG, Fp q Ñ H 2 pG, Fp q is a non-degenerate bilinear form. The first two conditions may be rephrased as ‘G is a finitely generated onerelator pro-p group’. We call an element r of a finitely generated free pro-p group F a Demushkin word if F {xxryy is a Demushkin group. Demushkin groups were classified by Demushkin [Dem61, Dem63] (for p ‰ 2), by Serre [Ser62] (when p “ 2 and the rank of G is odd) and by Labute [Lab67] for the remaining cases. The only Demushkin group which is not a PD2 group is Z{2, although some authors do not admit Z{2 as a Demushkin group. The missing constraint to force G to be PD2 is that G has cohomological dimension 2. This follows from the definition of a Demushkin group provided G is infinite— see Section 9.1 of [Ser62]. By convention we declare the orientation character of Z{2Z to be the natural inclusion Z{2 Ñ x´1y Ď U2 . The classification may be stated as follows. Theorem 1.9. Let F be a finitely generated free pro-p group and let r be a Demushkin word in F . Let G “ F {xxryy and let χ : G Ñ Up be the orientation character of the PD2 group G. Then there is a basis x1 , . . . , xn of F such that r takes one of the following forms. • If q ‰ 2 then n is even and r “ xq1 rx1 , x2 s ¨ ¨ ¨ rxn´1 , xn s rf s • If q “ 2, n is even and impχq “ U2 for some f ě 2 then f r “ x2`2 rx1 , x2 s ¨ ¨ ¨ rxn´1 , xn s 1 pf q • If q “ 2, n is even and impχq “ x´1y ˆ U2 for some f ě 2 then f r “ x21 rx1 , x2 sx23 rx3 , x4 s ¨ ¨ ¨ rxn´1 , xn s pf q • If q “ 2, n is odd and impχq “ x´1y ˆ U2 for some f ě 2 then f r “ x21 x22 rx2 , x3 s ¨ ¨ ¨ rxn´1 , xn s Conversely all these forms of words are Demushkin words and the corresponding Demushkin groups have orientation characters as stated. 2 Capping off peripheral subgroups We will attack the classification problem by reducing the number of boundary components in order to make use of the classification of Demushkin groups as a base case. We do this largely because computing an explicit formula for the cup product on H 1 pG, Sq b H 1 pGq in terms of the Si is rather awkward, and it is easier to deal with the induction. We will need the following basic proposition concerning non-degeneracy of bilinear forms. The proof is an exercise in undergraduate algebra. 7 Proposition 2.1. Let V, W be vector spaces over a field F , and let B : V ˆW Ñ F be a bilinear form. Assume that we have direct sum decompositions V “ V 1 ‘ V 2, W “ W1 ‘ W2 and that Bpv 1 , w2 q “ 0 for all v 1 P V 1 , w2 P W 2 . Then B is non-degenerate if and only if the two bilinear forms B1 : V 1 ˆ W 1 Ñ F, B2 : V 2 ˆ W 2 Ñ F obtained by restriction from B are non-degenerate. We will first deal with the case b ě 1 as this only involves relative cohomology. Theorem 2.2. Let G be a free pro-p group of rank n ` b where n ě 0, b ě 1 and let S “ tSi ubi“0 be a family of cyclic subgroups of G. Suppose there exist generators si for the Si such that (1) s1 , . . . , sb generate a free factor of G and (2) s0 ” s1 ¨ ¨ ¨ sb modulo Gp rG, Gs. (3) Si X xxSb yy “ t1u for i ‰ b Define G to be the quotient of G by the normal subgroup generated by Sb , let φ : G Ñ G be the quotient map and let S i “ φpSi q for 0 ď i ă b. Let S “ tS i ub´1 i“0 . Note that pG, Sq satisfy the conditions (1) and (2) above. Then pG, Sq is a PD2 pair if and only if pG, Sq is a PD2 pair. When they are both PD2 pairs then their orientation characteristics χ, χ satisfy χ “ χφ. Remark. The assumption that Si X xxSb yy “ t1u for i ‰ b is equivalent to Si Ę xxSb yy since G is free, hence torsion-free. This is automatic for b ą 1 by the conditions (1) and (2). For the case when b “ 1 and S0 Ď xxS1 yy see Proposition 2.3. Proof. Extend s1 , . . . , sb to a free basis px1 , . . . , xn , s1 , . . . sb q for G and take px1 , . . . , xn , s1 , . . . sb´1 q as basis for G. There is a map of pairs pG, Sq Ñ pG, Sq induced by φ where we regard Sb as sent to 1 P Sb´1 . This gives a commutative diagram of long exact sequences in relative cohomology with Fp -coefficients. The set-up of the theorem and the fact that the coefficients are a field allow one to easily write down all the groups and maps in this sequence explicitly. In particular one finds that the natural map – Fp – H 2 pG, Sq ÝÑ H 2 pG, Sq – Fp is an isomorphism and that there are splittings H 1 pG, Sq “ H 0 pSb q ‘ H 1 pG, Sq, H 1 pGq “ H 1 pSb q ‘ H 1 pGq Here the maps from H 0 pSb q and H 1 pG, Sq to H 1 pG, Sq are the natural maps from the commutative diagram and the maps from H 1 pGq and H 1 pSb q to H 1 pGq are respectively induced by φ and by the retraction G Ñ Sb which kills all the elements in our chosen basis other than sb . 8 Since cup products are natural with respect to maps of pairs (Proposition 3.1 of [Wil17a]), the cup product on H 1 pG, Sq b H 1 pGq is the restriction of the cup product on H 1 pG, Sq b H 1 pGq. Furthermore Equation 3.6 in Section 3.2 of [Wil17a] gives a commutative pentagon H 0 pSb q b H 1 pGq H 0 pSb q b H 1 pSb q ! H 1 pG, Sq b H 1 pGq H 1 pSb q ! – 2 H pG, Sq This shows firstly that the restriction of the cup product on H 1 pG, Sq b H 1 pGq to H 0 pSb q b H 1 pSb q via the direct sum decompositions above agrees with the cup product on Sb and that the cup product of an element of H 0 pSb q with an element of H 1 pGq vanishes. We may now apply Proposition 2.1 to conclude that pG, Sq is a PD2 pair if and only if pG, Sq is a PD2 pair. We move on to consideration of the orientation characters. Since χpSb q “ 1 there is a unique map ρ : G Ñ Up such that χ “ ρφ. We must show that ρ is the orientation character of G. Let m be a natural number. Note that our choices of bases give a splitting G “ G > Sb . Let φ! : G Ñ G be the inclusion map and note that φφ! “ idG . Note also that χ|G “ ρ. By excision (Theorem 4.8 of [Wil17a]) the map φ! induces an isomorphism H 1 pG, Sb ; Im pχqq Ñ H 1 pG, 1; Im pρqq. Therefore since φφ! is the identity, φ induces an isomorphism in the other direction. Let S 1 “ tS0 . . . Sb´1 u. The map of pairs pG, S 1 q Ñ pG, Sq and Propositions 2.4 and 2.6 of [Wil17a] give a commutative diagram (with coefficients in Im pχq on the top row and Im pρq on the bottom row): H 1 pG, Sb q – H 1 pG, 1q H 1 pS 1 q H 2 pG, Sq 0 H 2 pG, S \ 1q 0 – H 1 pSq which demonstrates that |H 2 pG, S; Im pρqq| “ |H 2 pG, S \ 1; Im pρqq| “ |H 2 pG, S; Im pχqq| and we are done by Proposition 1.7. Proposition 2.3. Let G be a free pro-p group of rank n ` 1 where n ě 0 and let S “ tS0 , S1 u be a family of cyclic subgroups of G. Suppose there exist generators si for the Si such that (1) s1 generates a free factor of G and (2) s0 ” s1 modulo Gp rG, Gs. Assume further that S0 X xxS1 yy ‰ 1. If pG, Sq is a PD2 pair then n “ 0, hence G “ S0 “ S1 . Conversely if n “ 0 and G “ S0 “ S1 then pG, Sq is a PD2 pair. 9 Proof. Let G “ G{xxS1 yy. By considering the commuting diagram of long exact sequences for the map of pairs pG, Sq Ñ pG, t1u \ t1uq one may readily show that the maps H 1 pG, t1u \ t1uq Ñ H 1 pG, Sq, H 1 pGq Ñ H 1 pGq with coefficients Fp are respectively an isomorphism and an injection. Since G is free, H 2 pG, t1u \ t1uq “ 0 and the cup product on pG, t1u \ t1uq vanishes. Therefore the cup product of any element of H 1 pG, Sq with the n-dimensional subspace H 1 pGq of H 1 pGq is zero. This cup product is non-degenerate so n “ 0 as claimed. The converse statement is a case of Proposition 6.16 of [Wil17a]. We now tackle the somewhat more fiddly case when we ‘cap off’ the only peripheral subgroup. Theorem 2.4. Let G be a free pro-p group of rank n ą 0 and let S Ď rG, GsGp be a cyclic subgroup. Let G “ G{xxSyy and let φ be the quotient map. Then pG, Sq is a PD2 pair if and only if G is a Demushkin group. When this occurs the orientation characters χ and χ of pG, Sq and G are related by χ “ χφ. Proof. Firstly note that G is Demushkin if and only if H 2 pG, 1; Fp q “ Fp and the cup product H 1 pG, 1; Fp q b H 1 pG, Fp q Ñ H 2 pG, 1; Fp q is non-degenerate. This follows immediately from the natural isomorphisms H 1 pG, 1; Fp q – H 1 pG, Fp q, H 1 pG, 1; Fp q – H 2 pG, Fp q Now note that in our situation we have H 2 pG; Fp q – Fp unless S is trivial, in which case we are done. This follows since G is not free (since H 1 pGq “ H 1 pGq, if G were free the map φ would be an isomorphism and S trivial) so has nontrivial cohomology in dimension two (see Theorem 7.7.4 of [RZ00]). Furthermore the dimension of H 2 pGq is at most one since G is a one-relator pro-p group. See Section I.4.3 of [Ser13]. From the condition S Ď rG, GsGp and long exact sequences in cohomology one easily deduces that the natural map H 1 pG, 1; Fp q Ñ H 1 pG, S; Fp q is an isomorphism. Once we know that the corresponding map on H 2 is an isomorphism the first part of the theorem follows immediately. The long exact sequence implies that H 2 pG, Sq is also isomorphic to Fp . Verifying that the natural map H 2 pG, 1q Ñ H 2 pG, Sq is itself an isomorphism is rather technical and we delay it to the end of the proof. For now we move onto verifying that the orientation characters of pG, Sq and G agree. Let m be a natural number. We will make use of the Five Term Exact Sequence for the extension 1 Ñ R Ñ G Ñ G Ñ 1 (Corollary 7.2.5 of [RZ00]). Here the relevant portion takes the form H 1 pG, Im pρφqq Ñ H 1 pR, Z{pm qG Ñ H 2 pG, Im pρqq Ñ 0 (1) where R “ xxSyy and ρ : G Ñ Up is any homomorphism. Taking ρ “ 1 shows that the middle term is finite (using H 2 pG; Fp q “ Fp and dévissage to prove 10 that H 2 pG, Z{pm q is finite). Now consider varying ρ. By Proposition 1.7 the modulus of the final term attains its unique maximum when ρ agrees with χ modulo pm . On the other hand if there is a ρ such that the first map in (1) vanishes then this ρ will maximise the size of the final term, hence this ρ agrees with χ modulo pm . In particular, since χ|S vanishes we may write χ “ ρ0 φ for some ρ0 . There is now a commutative diagram induced by Poincaré duality – H 1 pG, Im pχqq H1 pG, S; Z{pm q H 1 pR, Z{pm qG 0 – H 1 pS, Z{pm q H0 pS; Z{pm q where the fact that the rightmost map is zero follows from the long exact sequence in relative homology, and the fact that the lower left map is injective is the statement that any G-invariant homomorphism from xxSyy to Z{pm which vanishes on S is trivial. We have also used the fact that Zp pχq b Im pχq is isomorphic to the trivial module Z{pm . It follows that the upper left map vanishes and therefore ρ0 agrees with χ and χ agrees with χφ modulo pm for every m as required. We must now tackle the final verification we have been avoiding, that the natural map H 2 pG, 1q Ñ H 2 pG, Sq is an isomorphism. Here and for the rest of the proof the coefficient group is assumed to be Fp . We approach H 2 pGq using the Five Term Exact Sequence as above, so it will be necessary to understand the cohomology of R. Firstly note that the full five term exact sequence in this case is tg – 0 Ñ H 1 pGq ÝÑ H 1 pGq Ñ H 1 pRqG ÝÑ H 2 G Ñ 0 So that the transgression map tg is an isomorphism. By the Corollary to Proposition 26 of [Ser13] we know H 1 pRqG – Fp , and as above the natural map H 1 pRqG Ñ H 1 pSq is injective, hence it is an isomorphism. We may also consider the long exact sequence in cohomology for the pair pG, Rq with coefficients in Fp . This is a sequence of vector spaces so applying any additive functor—for instance the functor p´qG taking G-invariants—preserves exactness. From the Five Lemma and the data already collected we find that the natural map from the resulting sequence to the long exact sequence for pG, Sq is an isomorphism. We now have a diagram of maps H 1 pRqG H 1 pRqG tg – H 2 pGq – δ – – φ ˚ H 2 pG, RqG H 1S – H 2 pG, Sq where the right hand square commutes. Here the bottom left horizontal map is induced by the map of pairs pG, Rq Ñ pG, 1q and the isomorphism H 2 pG, 1q “ H 2 pGq. Our aim is therefore to show that the composition along the bottom row is an isomorphism. It only remains to show that the left hand square actually commutes (in fact it will turn out that it commutes up to sign, but this is enough). This verification takes place on the level of chain complexes. 11 For a pro-p group H and a subgroup K, let C‚ pHq be the (inhomogeneous) bar resolution of H over Zp . See Section 6.2 of [RZ00]. For our purposes it is enough to recall that the lower stages are given by d d 1 0 Zp rrHssrrH 2 ss ÝÑ Zp rrHssrrHss ÝÑ Zp rrHss where the boundary maps are given on the basis elements rh1 , h2 s P H 2 and rhs P H by d1 prh1 , h2 sq “ h1 rh2 s ´ rh1 h2 s ` rh1 s, d0 prhsq “ h ´ 1 Furthermore there is a natural chain map indK H pC‚ pKqq ãÑ C‚ pHq which is an isomorphism on the zero level. Each quotient module C‚ pH, Kq is projective; indeed it is a free Zp rrHss-module on the pointed profinite space obtained by collapsing the closed subset K n Ď H n to a point. See Section 5.2 of [RZ00] for information on free modules on pointed profinite spaces. The long exact sequence in homology for the short exact sequence of chain complexes qt 0 Ñ indK H pC‚ pKqq ÝÑ C‚ pHq ÝÑ C‚ pH, Kq Ñ 0 shows that the final term is a resolution for ∆H,K (shifted by one degree). In our case of interest this gives us standard chain complexes which may be used to compute the various cohomology groups in play. In particular we have an identification C‚ pG, 1q ” C‚`1 pGq and the obvious map (of G-modules) qt C‚ pGq ÝÑ C‚ pG, Rq ÝÑ C‚ pGq gives the map H ˚ pGq Ñ H ˚ pG, Rq. Now let ζ : R Ñ Fp be a G-invariant homomorphism giving a generator of H 1 pRqG . Let σ : G Ñ G be a continuous section of the quotient map such that σp1q “ 1 and let πpgq “ gσpgRq´1 be the retraction G Ñ R so defined. The section σ exists by Proposition 2.2.2 of [RZ00]. The transgression tgpζq is defined in the following manner (see [Lab67], Section 1.4). The cochain ζπd2 : C2 pGq Ñ Fp may be shown to factor through C2 pGq, and this cochain gives tgpζq. It therefore also factors through C2 pG, Rq giving a cochain representing φ˚ tgpζq. Now let us compute the map δ : H 1 pRqG Ñ H 2 pG, Rq. Applying the Horseshoe Lemma (see [Wei95], Lemma 2.2.8) gives the commuting diagram of (par- 12 tial) resolutions C2 pG, Rq C2 pG, Rq ‘ C1 pGq ¨ ˝ d1 C1 pG, Rq qt 0 d0 ‚ C1 pG, Rq ‘ C0 pGq d0 C0 pGq pi˝d0 q‘ǫ d0 ∆G,R d1 C1 pGq ˛ i Zp rrG{Rss ǫ Zp The central column has a chain map to the standard resolution given by C2 pG, Rq ‘ C1 pGq C1 pG, Rq ‘ C0 pGq α‘β Zp rrGssrrRss Zp rrG{Rss ν‘id Zp rrGss Zp rrG{Rss where νprgsq “ σpgRq ´ 1 and αprg1 , g2 sq “ g1 g2 rπ̃pg2 qs ´ g1 g2 rπ̃pg1 g2 qs ` g1 rπ̃pg1 qs ´ g1 r1s where π̃pgq “ g ´1 σpgSq “ pπpgq´1 qg for some map β (which is irrelevant to what follows). We now readily see that δpζq “ ζα : C2 pG, Rq Ñ Fp is induced by the cochain ζ π̃d2 : C2 pGq Ñ Fp . As ζ π̃ “ ´ζπ we are done. 3 Classification In this section we will prove the main classification theorem. Definition 3.1. Let pG, Sq be a PD2 pair with S “ tS0 , . . . , Sb u. A basis for G in standard form (or simply standard form basis) is an ordered free generating set B “ tx1 , . . . , xn , s1 , . . . , sb u such that si generates some conjugate of Si for each i. Note that when b ě 1 any standard form basis of G{xxSb yy is the image of the first n ` b ´ 1 elements of some standard form basis of G. Definition 3.2. For integers n, b let G be a free group of rank n ` b on a generating set B “ tx1 , . . . , xn , s1 , . . . , sb u Let χ : G Ñ Up be a homomorphism vanishing on the si . Define the standard word r1 pn, b, χ; Bq P G according to the following cases. • If n is even and qpχq ‰ 2 then r1 pn, b, χ; Bq “ xq1 rx1 , x2 s ¨ ¨ ¨ rxn´1 , xn ss1 ¨ ¨ ¨ sb 13 rf s • If qpχq “ 2, n is even and impχq “ U2 for some f ě 2 then f r1 pn, b, χ; Bq “ x2`2 rx1 , x2 s ¨ ¨ ¨ rxn´1 , xn ss1 ¨ ¨ ¨ sb 1 pf q • If qpχq “ 2, n is even and impχq “ x´1y ˆ U2 for some f ě 2 then f r1 pn, b, χ; Bq “ x21 rx1 , x2 sx23 rx3 , x4 s ¨ ¨ ¨ rxn´1 , xn ss1 ¨ ¨ ¨ sb pf q • If qpχq “ 2, n is odd and impχq “ x´1y ˆ U2 for some f ě 2 then f r1 pn, b, χ; Bq “ x21 x22 rx2 , x3 s ¨ ¨ ¨ rxn´1 , xn ss1 ¨ ¨ ¨ sb If none of these cases holds we leave r1 pn, b, χ, Bq undefined. Note that whether or not r1 is defined depends only on n and impχq, not on b or B. When the basis B or the invariants pn, b, χq are clear from context we will omit them from the notation. The main theorem of this paper is the following classification theorem. Theorem 3.3. Let G be a free pro-p group of rank n ` b and S “ tS0 , . . . , Sb u be a finite family of closed procyclic subgroups where b ě 0, n ě 0 and n ` b ě 1. Fix a generator s0 of S0 . Then pG, Sq is a PD2 pair with orientation character χ if and only there exists a standard form basis B “ tx1 , . . . , xn , s1 , . . . , sb u of G such that s0 “ r1 pn, b, χ; Bq. Remark. Implicit in the final sentence is the statement that r1 pn, b, χq is not undefined. The backwards implication—that these forms of words do genuinely give PD2 pairs—follows immediately by an induction using Theorems 2.2 and 2.4 and Proposition 2.3 taken together with the classification of Demushkin groups (Theorem 1.9). The forwards implication will proceed in two steps. First we will use the results of the previous section to find a basic form for the word s0 . Then we will use the method of successive approximation to erode this word down to one of the forms above. From now on let G be a free pro-p group of rank n ` b and S “ tS0 , . . . , Sb u be a finite family of closed procyclic subgroups where b ě 0, n ě 0 and n ` b ě 1 and assume pG, Sq is a PD2 pair with orientation character χ. Set q “ qpχq and let pGj q be the lower central q-series. Fix the generator s0 of S0 . Proposition 3.4. For some basis B of G in standard form we have s0 ” r1 pn, b, χ; Bq modulo G3 (and r1 pn, b, χq is well-defined). Proof. We prove this by induction on b. The base cases are b “ 0—which holds by Theorem 2.4 and Theorem 1.9—and b “ 1, n “ 0, which holds by Proposition 2.3. Suppose that b ą 1 and that n ‰ 0 if b “ 1. Then we may apply Theorem 2.2 and an induction hypothesis to show that for some standard form basis B we have s0 P r1 pn, b ´ 1, χ; Bqxxsb yyG3 14 Let y be some element of xxsb yy such that s0 ” r1 pn, b ´ 1, χ; Bqy modulo G3 By Proposition 1.6, replacing sb by a power sµb for some µ P Zˆp we may assume that ζb pyq “ 1 where ζb is the projection of G onto Sb given by killing the other elements of B. We claim that modulo G3 we may write y as an element of the form sgb for some g P G. Consider an expression of G{G3 as an inverse limit of finite p-group quotients lim ÐÝ Pk . Let ψk : G{G3 Ñ Pk be the quotient map. For each k we have ψk pyq P xxψk psb qyy and since ζb pyq “ 1 we have ψk pyq ” ψk psb q modulo rPk , Pk s. Now since Pk is finite, we have ψk pyq P xxψk psb qyy. That is, there are gi P Pk and mi P Z (1 ď i ď r) such that ψk pyq “ “ r ź i“1 r ź pψk psb qmi qgi ψk psb qmi rψk psb qmi , gi s i“1 ř mi ś gi “ ψk psb q “ ψk psb q r ź rψk psb q, gimi s i“1 mi Here we have used the fact ř that ψk pyq ” ψk psb q modulo rPk , Pk s in the final line to show that ψk psb q mi “ ψk psb q. We have made frequent use of commutator identities and the fact that Pk was a quotient of G{G3 hence all third commutators in Pk vanish. Hence ψk pyq lies in the conjugacy class ψk psb qPk of sb for all k. Since these conjugacy classes form a surjective inverse system with inverse limit sG b , there does indeed exist g P G such that y “ sgb . Now replace sb by sgb so that s0 ” r1 pn, b ´ 1, χqsb “ r1 pn, b, χq modulo G3 . This concludes the proof. We will now use the method of successive approximation to improve congruence modulo G3 to equality. We will first examine the q ‰ 2 case to illustrate the method. The q “ 2 case simply requires more notation, together with some extra tidying up at the end. Proof of Theorem 3.3, q ‰ 2 case. Recall from Theorem 1.9 that if q ‰ 2 then n “ 2N is even. If q ‰ 0 then n ą 0. Let B “ tx1 , . . . , xn , s1 , . . . , sb u be a basis of G in standard form. If t “ pt1 , . . . , tn`b q P pGj´1 qn`b for j ě 3 then we may define t t B 1 “ B 1 ptq “ tx1 t1 , . . . , xn tn , s1n`1 , . . . , sbn`b u which is also a basis in standard form. One easily verifies that r1 pB 1 q ” r1 pBq modulo Gj 1 B so that there is a unique element dB j´1 ptq P Gn such that r1 pB q “ r1 pBqdj´1 ptq. Let the images of xi and si in gr1 pGq be denoted ξi and σi , let the image of ti in grj´1 pGq be τi and let the image of t in grj´1 pGqn`b be τ . Then the image 15 of dB j´1 ptq in grj pGq is ˆ ˙ b N ÿ ` ˘ ÿ q B δj´1 pτ q “ π ¨ τ1 ` rσi , τn`i s rτ2i´1 , ξ2i s ` rξ2i´1 , τ2i s ` rτ1 , ξ1 s ` 2 i“1 i“1 B as one may readily compute using commutator identities. This δj´1 is a Zp {qZp n`b linear homomorphism grj´1 pGq Ñ grj pGq. Up to certain minus signs in the σi terms and reordering some coordinates of grj´1 pGqn`b this is in fact the same as the map δj´1 described in Proposition 5 of [Lab67]. Hence by that B proposition we have impδj´1 q “ grj pGq. Therefore suppose that we have chosen a standard form basis Bj´1 such that s0 ” r1 pBj´1 q modulo Gj´1 Then s0 “ r1 pBj´1 qz for some z P Gj . There exists t P pGj´1 qn`b such that Bj´1 ptq equals the image of z in grj pGq. Form the new standard form basis δj´1 1 Bj “ Bj´1 ptq as above. Then Ş s0 ” r1 pBj q modulo Gj Since Gj “ t1u and since Bj´1 and Bj agree modulo Gj , for each i the sequence comprising the ith elements of the Bj converges to a limit in G as j Ñ 8. One may easily see that the set of limits B is a standard basis for G and s0 “ r1 pBq as required. We now move on to the q “ 2 case. The principle is the same, but the notation is more involved. Proof of Theorem 3.3, q “ 2 case. Let B “ tx1 , . . . , xn , s1 , . . . , sb u be a basis of G in standard form. For j ě 3, t “ pt1 , . . . , tn`b q P pGj´1 qn`b and ǫ P t0, 1ub and define j´1 B 1 “ B 1 pt, ǫq “ tx1 t1 , . . . , xn tn , ps1`2 1 ǫ1 tn`1 q j´1 , . . . , ps1`2 b ǫb tn`b q u which is again a basis in standard form. Again we have r1 pB 1 q ” r1 pBq modulo Gj , so that there exists a unique element dB j´1 pt, ǫq P Gj such that r1 pB 1 q “ r1 pBqdB j´1 pt, ǫq One may readily compute that the image of dB j´1 pt, ǫq in grj pGq is B δj´1 pτ q “ π ¨ τ1 ` rτ1 , ξ1 s ` N ÿ ` ˘ rτ2i´1 , ξ2i s ` rξ2i´1 , τ2i s i“1 ` b ÿ ` ˘ rσi , τn`i s ` ǫi π ¨ σi i“1 when n “ 2N is even and B δj´1 pτ q “ π ¨ τ1 ` rτ1 , ξ1 s ` N ÿ ` ˘ rτ2i , ξ2i`1 s ` rξ2i , τ2i`1 s i“1 ` b ÿ ` ˘ rσi , τn`i s ` ǫi π ¨ σi i“1 16 when n “ 2N ` 1 is odd—using the same notation as in the q ‰ 2 case. For either parity of n, Proposition 5 of [Lab67]—with minor notational tweaks— B shows that the image of δj´1 together with the elements π j´1 ¨ ξi generates grj pGq. Actually the cited proposition is slightly more precise than this, but we won’t need this precision here as it will be folded into the last step of the proof. As one final piece of notation, for λ P p4Z2 qn set epB, λq “ xλ1 1 ¨ ¨ ¨ xλnn Note that for any t, ǫ as above we have epB 1 pt, ǫq, λq ” epB, λq modulo Gj`1 where the extra precision modulo Gj`1 follows since the λi are divisible by 4. Now suppose that j ě 3 there is a basis Bj of G in standard form and some λj´1 P p4Z2 qn such that s0 “ epB, λqr1 pBqz where z P Gj . This holds for j “ 3 by Proposition 3.4. From above there exist t, ǫ and α such that j´1 z ” x12 α1 j´1 ¨ ¨ ¨ xn2 αn B dj´1 pt, ǫq modulo Gj`1 1 If Bj “ Bj´1 pt, ǫq then we have s0 j´1 ” epB, λj´1 qr1 pBqx12 α1 ” epBj , λj´1 ` 2j´1 αqr1 pBj q j´1 ¨ ¨ ¨ xn2 αn B dj´1 pt, ǫq modulo Gj`1 . As before the Bj converge to a standard basis for G; furthermore λj converges to some λ P p4Z2 qn . So passing to the limit we have found a standard-form basis B8 such that s0 “ epB8 , λj´1 qr1 pB8 q “ epB8 , λj´1 qr1 pn, 0, χ; X qs1 ¨ ¨ ¨ sb Here we have written B8 “ X \ Y where X denotes the first n elements of B8 . Now by Theorems 2.2 and 2.4 we find that epB8 , λj´1 qr1 pn, 0, χ; X q is in fact a Demushkin word in the free group G1 “ xX y, and moreover the image of the orientation character of G1 equals the image of χ. Therefore by Theorem 1.9 there is a free basis X 1 of G1 such that epB8 , λj´1 qr1 pn, 0, χ; X q “ r1 pn, 0, χ; X 1 q Then if B “ X 1 \Y then B is a basis of G in standard form and s0 “ r1 pn, b, χ; Bq as required. 4 Splittings and the pro-p curve complex An important tool in the study of surface automorphisms is the curve complex, which is a simplicial complex whose vertices are isotopy classes of essential simple closed curves on the surface. Translating into the language of group theory this may be rephrased as ‘a complex whose vertices are splittings of the 17 surface group over Z’. In this section we will define an analogous object for the pro-p completion of a surface group, on which the automorphism group of the pro-p group will act. One could no doubt study such an object for an arbitrary Demushkin group, but it is simpler and perhaps more interesting to restrict to the case of an orientable Demushkin group (i.e. the pro-p completion of an orientable surface group). For the rest of the section let Σ be a compact orientable surface which is not p ppq be the pro-p completion a sphere, disc or cylinder. Let Γ “ π1 Σ and let G “ Γ of Γ. Let S be a family of subgroups of G comprising the pro-p completion of one representative of the fundamental group of each boundary component. Definition 4.1. Define the relative automorphism group AutB pGq to be the group of automorphisms of G which send every S P S to a conjugate of itself. One may check that AutB pGq is a closed subgroup of AutpGq containing InnpGq. Define OutB pGq “ AutB pGq{ InnpGq. Similarly for Γ. Definition 4.2. Let ‫ג “ ג‬pΣq be the set of all equivalence classes of pairs pT, ρq with the following properties. • T is a pro-p tree and ρ : G Ñ AutpT q is a continuous left action of G on T by graph automorphisms such that GzT is finite. • For every e P EpT q the edge stabiliser stabρ peq is isomorphic to Zp • If GzT is not a loop, then for every vertex v P V pT q and every edge e incident to v the edge stabiliser stabρ peq is properly contained in the vertex stabiliser stabρ pvq. • Every S P S fixes a vertex of T , but does not fix any edge of T . Two such objects are regarded as equivalent if there is a G-equivariant graph isomorphism from one to another. The automorphism group AutB pGq acts on ‫ג‬pΣq on the right by pT, ρq ¨ ψ “ pT, ρ ˝ ψq where ψ P AutB pGq. Note that if ψ is an inner automorphism of G, then pT, ρq ¨ ψ is equivalent to pT, ρq so the action descends to an action of OutB pGq on ‫ג‬pΣq. The object ‫ג‬pΣq will be referred to as the pro-p curve complex of Σ. At present it is merely a set, but later we shall see that it has a topology making it into a profinite space such that the action of AutB pGq is continuous. We will usually abuse notation by ignoring the fact that elements of ‫ג‬pΣq are really equivalence classes and simply deal with the pairs pT, ρq themselves. Note that if one replaces G with Γ and ‘pro-p tree’ with ‘simplicial tree’ in the above definition then one recovers the curve complex CpΣq of Σ. For our purposes we will adopt the following definition. Definition 4.3. Given a set X, a simplicial structure on X is a partial ordering ď on X such that: • For every x P X there exist only finitely many y P X such that y ď x • Let X pnq be the set of elements x P X such that there are n ` 1 minimal elements of X which are smaller than or equal to x. Then for each x P X pnq , x is the join of these elements and the subposet ty P X | y ď xu is isomorphic to the face poset of an n-simplex. We refer to x P X pnq as an n-simplex of X. 18 The curve complex CpΣq has a natural AutB pΓq-equivariant simplicial structure which may be described as follows. For collections of isotopy classes of closed curves x, y P CpΣq, we say y ď x if y may be obtained by deleting some of the isotopy classes of closed curves in x. In terms of the action on dual trees, this corresponds to a Γ-equivariant epimorphism of trees. Therefore we define a poset structure on ‫ג‬pΣq as follows. Definition 4.4. For pT, ρq and pT 1 , ρ1 q in ‫ג‬pΣq, we declare pT, ρq ě pT 1 , ρ1 q if and only if there is a G-equivariant epimorphism of graphs T Ñ T 1 . Definition 4.5. We define a decomposition graph of Σ to be a finite connected graph X with two labelling functions n‚ , b‚ : V pXq Ñ N such that: ř • vPV pXq bv is the number of boundary components of Σ ř • rk Γ “ vPV pXq pnv ` bv q ` 2p1 ´ χpXqq ´ ǫ where ǫ is 1 if Σ has boundary or 0 if Σ is closed. • For a vertex v, if n “ 0 then either X is a loop and Σ is a torus or bv ` valpvq ą 2 Let D “ DpΣq be the set of decomposition graphs modulo isomorphism of labelled graphs. We will be aiming to prove the following theorem. Here we regard AutB pΓq as a subgroup of AutB pGq in the natural way. Theorem 4.6. Let Σ be a compact orientable surface and let CpΣq be the curve complex of Σ. There is an AutB pπ1 Σq-equivariant function i : CpΣq Ñ ‫ג‬pΣq such that: (1) i is an injection. If x, y P CpΣq then x ď y if and only if ipxq ď ipyq. Furthermore if z P ‫ג‬pΣq and z ď ipyq then z “ ipxq for some x ď y. (2) ď is a simplicial structure on ‫ג‬pΣq and i preserves the simplicial structure (3) There is a natural diagram of bijections i CpΣq{ AutB pΓq d¯ ‫ג‬pΣq{ AutB pGq δ̄ DpΣq Remark. Part (3) of the theorem is perhaps the most interesting part, and may be paraphrased as ‘any splitting of G with edge groups Zp is equivalent under AutB pGq to a geometric splitting’. Construction of i. Let x be an n-simplex in the curve complex—that is, an isotopy class of collections of n ` 1 disjoint essential simple closed curves on Σ which may be made disjoint. Such a collection (together with various choices of base-points) induces a graph-of-groups splitting of Γ where all edge groups are copies of Z, where peripheral subgroups are conjugate into vertex groups but not edge groups, and where edge groups are not equal to adjacent vertex groups except in the case of a torus with a single non-separating simple closed curve. 19 This splitting is p-efficient (that is, well-behaved in the pro-p topology on Γ) by, for instance, Proposition 3.8 of [Wil17b]. Therefore by Propositions 6.5.3 and 6.5.4 of [Rib17] there is an induced splitting of G as an injective graph of pro-p groups with procyclic edge groups. The standard tree dual to this splitting (see Theorem 6.5.2 of [Rib17]) is a representative of an element ipxq of ‫ג‬pΣq. Taking different choices of basepoints does not affect the equivalence class of ipxq in ‫ג‬pΣq. Furthermore the abstract tree T abs dual to the splitting of π1 Σ along x is naturally embedded as an abstract subgraph of T which is dense in the topology on T by Proposition 6.5.4 of [Rib17]. Construction of d and δ. There is a function d : CpΣq Ñ DpΣq defined as follows. Given a representative of an element of the curve complex (i.e. a collection C of disjoint essential simple closed curves on Σ) there is a graph X dual to these curves, with one vertex in each component of the complement of these curves and one edge for each curve. For a component v of Σ r C set nv to be half the genus of v and bv to be the number of boundary components of Σ lying in v. One easily verifies that this is a decomposition graph for Σ. This defines the map d. Clearly d is AutB pΓq-invariant and therefore gives a map d : CpΣq{ AutB pΓq Ñ DpΣq. By the standard theory of surfaces this is a bijection. We may also define a map δ : ‫ג‬pΣq Ñ DpΣq. Let pT, ρq represent an element of ‫ג‬pΣq. Then GzT “ X is a finite connected graph. Choose a maximal subtree Y of X and a function s : X Ñ T which is a section of the quotient map and such that sd0 pxq “ d0 pspxqq for all x P X and sd1 pxq “ d1 spxq for all x P Y . Such a map s is a fundamental 0-section in the language of [Rib17]. Given s we may form a graph of groups G “ pX, G‚ q as in Section 6.6 of [Rib17], whose vertex and edge groups Gx are the stabilisers of the points spxq. This is an injective graph of pro-p groups whose fundamental pro-p group is G. Every S P S is conjugate into exactly one of the Gv . Replace every S by some such conjugate and for v P V pXq let Sv be the collection of elements of S lying in Gv . Further let Ev be the collection of edge groups of G incident to v. Since pG, Sq is a PD2 pair, Theorem 5.18 of [Wil17a] shows that pGv , Sv Y Ev q is also an orientable PD2 pair for every v P V pxq. Therefore Gx is a free pro-p group of rank nv ` |Sv | ` |Ev | ´ 1 for some nv P N. Set bv “ |Sv |. We have now defined a decomposition graph δpT, ρq. One may also characterise these invariants as follows: |Ev | is the number of Gv -orbits of edges of T incident to spvq, and bv is the number of Gv -conjugacy classes of the conjugates of elements of S that lie in Gv . Finally nv is defined as rk Gv ´ |Sv | ´ |Ev | ` 1 as before. This shows not only that δpT, ρq is independent of the choices of Y and s but is also AutB pGq-invariant. Hence we also have a map δ : ‫ג‬pΣq{ AutB pGq Ñ DpΣq. Proof of (3). The first characterisation of δ in terms of graphs of groups shows immediately that δ ˝ i “ d. This immediately shows that δ is a surjection. So to show that i induces the required bijection it remains to show that δ is an injection. For a finite graph X let DX pΣq be the set of decomposition graphs whose underlying graph is X, and let ‫ג‬X pΣq be the set of elements pT, ρq of ‫ג‬pΣq with GzT – X. Then δ breaks into a disjoint union of maps ‫ג‬X pΣq{ AutB pGq Ñ DX pΣq and it suffices to prove that each of these is an injection. We do this 20 by induction on the number of edges of X, over all surfaces Σ simultaneously. The idea of the induction is simple. Each action on a tree may be represented by a graph of groups, which may be thought of as a smaller graph of groups plus one additional edge. The induction hypothesis allows us to identify the smaller graphs of groups, and we may then glue back up to recover the full decompositions. The precise proof requires much notation. Let pT, ρq, pT 1 , ρ1 q P ‫ג‬X pΣq and suppose δpT, ρq “ δpT 1 , ρ1 q. Choose an edge e of X and let the component(s) of X r e be X1 (and X2 ). If e is separating we order the indices so that d0 peq P X1 . Further choose graphs of groups as above corresponding to T and T 1 . Let the fundamental pro-p groups of the restricted graphs of groups over X1 (and X2 ) be G1 and G11 (and G2 and G12 respectively). After replacing the S P S by conjugates if necessary each member of S lies in exactly one of G1 or G2 . Let S1 consist of d0 pGe q (and d1 pGe q if e is non-separating) together with those elements of S lying in G1 . If e is separating then set S2 to be the elements of S lying in G2 together with d1 pGe q. Similarly define Si1 . Finally choose generators s0 and s10 of Ge and G1e . If e is non-separating denote a choice of stable letter for the HNN extension G1 >Ge by t so that t´1 d0 ps0 qt “ d1 ps0 q. Similarly for t1 . By collapsing X1 and X2 and applying Theorem 5.18 of [Wil17a] we find that pGi , Si q is a PD2 pair. The rank and number of boundary components may be read off immediately from δpT, ρq “ δpT 1 , ρ1 q. Both pairs are orientable so the classification of PD2 pairs (Theorem 3.3) shows that pGi , Si q and pG1i , Si1 q are both isomorphic as pairs (after possibly changing the peripheral groups by conjugacies) to the pro-p completion of the fundamental group of some compact r surface Σ. r of Now over Xi we have graph of groups decompositions Ui , Ui1 P ‫ג‬Xi pΣq pGi , Si q and pG1i , Si1 q. Since δpT, ρq “ δpT 1 , ρ1 q and the graphs of groups are restrictions of those for pG, Sq it follows that δpU q “ δpU 1 q. By induction we therefore have an isomorphism ψi : Gi Ñ G1i (relative to Si and Si1 ) such that ψi pGx q is a Gi -conjugate of G1x for each x P Xi . Applying an inner automorphism we may assume ψ1 pd0 ps0 qq “ d0 ps10 q and, if e is separating, that ψ2 pd1 ps0 qq “ d1 ps10 q. When e is separating we may therefore glue these maps together (by the universal property of amalgamated free products) to give ψ P AutB pGq such that ψpGx q is a G-conjugate of G1x for all x P X. Therefore the two graphs of groups decompositions pX, ψpG‚ qq and pX, G1‚ q differ only by the choice of 0-section, which does not affect the Bass-Serre tree (see Section 6.6 of [Rib17]). Therefore pT, ρq ¨ ψ “ pT 1 , ρ1 q as required. It only remains to show that in the case when e is separating we must show that the automorphism ψ of G1 extends across the HNN extension to give an automorphism of G. The construction above ensured that ψ1 pd0 ps0 qq “ d0 ps10 q and that ψ1 pd1 ps0 qq “ pd1 ps10 qλ qg for some λ P Up and g P G1 . We must show that λ “ 1, for then t ÞÑ t1 g gives the required extension of ψ1 to ψ P AutB pGq. 21 We have a diagram of maps ψ˚ – H1 pd0 pGe qq – conjptq H1 pd1 pGe qq H2 pG1 , S1 q ψ˚ – H1 pd0 pG1e qq H2 pG11 , S11 q – conjpt1 q conjpg´1 q˝ψ˚ – H1 pd1 pG1e qq with coefficients in Zp . Here conjpxq denotes a map induced by conjugacy by x P G. The two triangles in this diagram commute because of the segment ‹ H2 pG, S1 q ÝÑ H2 pG1 , S1 q ÝÑ H1 pGe q of the (dual of) the long exact sequence in Theorem 4.11 of [Wil17a], where the starred map is the composition d´1 ´d´1 0 H2 pG1 , S1 q Ñ H1 pS1 q Ñ H1 pd0 pGe qq ‘ H1 pd1 pGe qq ÝÝ1ÝÝÝÝ Ñ H1 pGe q Therefore the entire diagram commutes. Tracing d0 ps0 q around the outer rectangle shows that λ “ 1 as required. Proof of (1). Let us prove that i is an injection. Now suppose ipxq ď ipyq. Let the discrete tree dual to x be T abs and that dual to y be U abs , and let T and U be the trees for ipxq and ipyq. As mentioned above, T abs is embedded in T (and similarly for U ). Take a G-equivariant isomorphism f : T Ñ U “ X and identify GzT with GzU via f . Take a maximal subtree Y of X and fundamental 0-sections s : X Ñ T abs , s1 : X Ñ U abs lifting Y —which of course also provide 0-sections for T and U . Then f s and s1 are both fundamental 0-sections for U lifting the same tree Y . Then by Theorem 6.6.1 (or rather its proof) there is a Gequivariant isomorphism h : U Ñ U taking f s to s1 . Then since T abs “ Γ ¨ spXq and U abs “ Γ ¨ s1 pXq the map hf restricts to an equivariant graph isomorphism T abs Ñ U abs and we are done. To show that x ď y implies ipxq ď ipyq note that the graph of groups decomposition of Γ corresponding to x is obtained from the decomposition corresponding to y by collapsing some sub-graphs-of-groups; therefore the same is true of ipxq and ipyq, and this collapse gives the required equivariant epimorphism of trees as required. Also, if z ď ipyq then at at the level of graphs of groups z is obtained from ipyq by collapsing some sub-graphs-of-groups; such collapses are precisely those induced by collapses of the graph of discrete groups corresponding to y. This shows that z “ ipxq for some (unique) x ď y. Proof of (2). Part (2) follows from (1) and (3). For the notion of simplicial structure consists of conditions on ty P ‫ג‬pΣq | y ď xu for x P ‫ג‬pΣq. Part (1) and the simplicial structure on CpΣq show that these conditions hold when x P ipCpΣqq. Since the partial order is AutB pGq-invariant and every point in ‫ג‬pΣq may be translated to a point in ipCpΣqq via the action of AutB pGq this shows that the simplicial structure conditions hold at every point of ‫ג‬pΣq. This proves (2). 22 Notice that ‫ג‬pΣq has a natural topology making it into a profinite space on which AutB pGq acts continuously. For choosing any lift s : DpΣq Ñ ‫ג‬pΣq splitting the quotient map defines a surjection AutB pGq ˆ DpΣq Ñ ‫ג‬pΣq, to which we may give the quotient topology. This is homeomorphic to the disjoint union ğ ‫ג‬pΣq “ stabpspzqqz AutB pGq zPDpΣq so that ‫ג‬pΣq is a profinite space. Because DpΣq is finite this topology does not depend on the choice of s. References [BE78] Robert Bieri and Beno Eckmann. Relative homology and Poincaré duality for group pairs. Journal of Pure and Applied Algebra, 13(3):277– 319, 1978. [Dem61] S. P. Demushkin. The group of a maximal p-extension of a local field. Izvestiya Rossiiskoi Akademii Nauk. Seriya Matematicheskaya, 25(3):329–346, 1961. [Dem63] S. Demushkin. On 2-extensions of a local field. Sibirsk. Mat. Zh, 4(4):951–955, 1963. [EL83] Beno Eckmann and Peter Linnell. Poincaré duality groups of dimension two, II. Commentarii Mathematici Helvetici, 58(1):111–114, 1983. [EM80] Beno Eckmann and Heinz Müller. Poincaré duality groups of dimension two. Commentarii Mathematici Helvetici, 55(1):510–520, 1980. [Lab67] John P. Labute. Classification of Demushkin groups. Canad. J. Math, 19(1):106–132, 1967. [Rib17] Luis Ribes. Profinite Graphs and Groups, volume 66 of A Series of Modern Surveys in Mathematics. Springer, 2017. [RZ00] Luis Ribes and Pavel Zalesskii. Profinite groups. Springer, 2000. [Ser62] Jean-Pierre Serre. Structure de certains pro-p-groupes. Séminaire Bourbaki, 8:145–155, 1962. [Ser13] Jean-Pierre Serre. Galois cohomology. Springer Science & Business Media, 2013. [SW00] Peter Symonds and Thomas Weigel. Cohomology of p-adic analytic groups. In New horizons in pro-p groups, volume 184, pages 347–408. Birkhäuser Boston, 2000. [Wei95] Charles A. Weibel. An introduction to homological algebra. Number 38 in ‘Cambridge studies in advanced mathematics’. Cambridge university press, 1995. [Wil17a] Gareth Wilkes. Relative cohomology theory for profinite groups. arxiv preprint arXiv:1710.00730, 2017. 23 [Wil17b] Gareth Wilkes. Virtual pro-p properties of 3-manifold groups. Journal of Group Theory, 20(5):999–1023, 2017. 24
4math.GR
Asymptotics of selective inference arXiv:1501.03588v2 [math.ST] 4 Aug 2016 Xiaoying Tian, Jonathan Taylor∗ Department of Statistics Stanford University Sequoia Hall Stanford, CA 94305, USA e-mail: [email protected] [email protected] Abstract: In this paper, we seek to establish asymptotic results for selective inference procedures removing the assumption of Gaussianity. The class of selection procedures we consider are determined by affine inequalities, which we refer to as affine selection procedures. Examples of affine selection procedures include selective inference along the solution path of the LASSO, as well as selective inference after fitting the LASSO at a fixed value of the regularization parameter. We also consider some tests in penalized generalized linear models. Our result proves asymptotic convergence in the high dimensional setting where n < p, and n can be of a logarithmic factor of the dimension p for some procedures. Our method of proof adapts a method of Chatterjee (2005). AMS 2000 subject classifications: Primary 62M40; secondary 62J05. Keywords and phrases: selective inference, non-gaussian error, highdimensional inference, LASSO. 1. Introduction Selective inference is a recent research topic that studies valid inference after a statistical model is suggested by the data Fithian et al. (2014), Lee et al. (2013), Taylor et al. (2014, 2013). Classical inference tools break down at this point as the data used for the hypothesis test is allowed to be the data used to suggest the hypothesis. Specifically, instead of being given a priori, the hypothesis to test is dependent on the data, thus random. Formally, denoted by E ∗ = E ∗ (y, X) is the model selection procedure, which generates a set of hypotheses to test, or perhaps parameters for which to form intervals. It is useful to think of E ∗ as a point process with values in S, where S is some collection of questions of possible interest. Consider the following example, Suppose y|X ∼ G with y ∈ Rn , X ∈ Rn×p , X fixed. For any E ⊂ {1, . . . , p} define the functionals βj,E (G) = eTj argmin EG (ky − XE βE k2 |X) j ∈ E, β where ej is the unit vector with only the j-th entry being 1. Such functionals βj,E is essentially the best linear coefficients within the model consisting of only ∗ Supported in part by NSF grant DMS 1208857 and AFOSR grant 113039. 1 Tian and Taylor/Asymptotics of selective inference 2 variables in E. Then the collection of possily interesting questions are   S = {βj,E , j ∈ E} : E ⊂ {1, . . . , p} . The data (y, X) will then suggest a subset of interesting variables E, and E ∗ (y, X) designates the target for inference to be {βj,E , j ∈ E}, the best linear coefficient within a model consisting of only the variables in E. Previous literature has studied inference after different model selection procedures E ∗ . Notably Lee et al. (2013) proposed an exact test within the model suggested by LASSO, that is E ∗ = {βj,E , j ∈ E}, where E is the active set of the LASSO solution. The test is based upon a pivotal quantity which the authors prove to be distributed as Unif(0, 1) if the hypothesis to be tested is true. Thus such quantity Pj (y) can be used to test the hypothesis H0j : βj,E = 0, and control the “Type-I error” at level α, P (Pj (y) ≤ α | H0j is true) ≤ α. (1) By inverting such tests, Lee et al. (2013) can also construct valid confidence intervals for βj,E . It is of course worth noticing that either the hypothesis H0j or the parameters βj,E are random as E is suggested by the data. So the “Type-I error” (1) is not the classical Type-I error definition where the hypotheses are given a priori. Such inference framework is first considered in Berk et al. (2013), and we leave the philosophical discussions of such approach to Fithian et al. (2014). The means by which Lee et al. (2013) controls the “Type-I error” is through constructing the p-value functions Pj . Such construction is highly dependent on the assumption of normality of the error distribution. Other works like Lockhart et al. (2013), Taylor et al. (2014) used similar approaches. Compared to these previous work, we seek to remove the Gaussian assumption on the errors and establish asymptotic distributions of Pj in this work. We state the conditions under which Pj will be asymptotically distributed as Unif(0, 1), and thus Pj can be used as p-values to test the hypotheses and asymptotically control the “Type-I error” in (1). This allows asymptotically valid inference in the linear regression setting without normality assumptions. It also allows application of covariance test (Lockhart et al. 2013) in generalized linear models. 1.1. Related works Tibshirani et al. (2015) also considers uniform convergence of the statistics proposed by Taylor et al. (2014), but focuses mainly on the low dimensional case. In the high dimensional case, they have a negative result on the uniform convergence of the pivot. In this paper, we instead focus on the high dimensional case and state the conditions in which the pivot will converge. More specifically, n is allowed to be of a logarithmic factor of the dimension p for two common procedures introduced in Section 4. Tian and Taylor/Asymptotics of selective inference 3 In the works of Belloni et al. (2012), Meinshausen et al. (2012), Zhang & Zhang (2014), Javanmard & Montanari (2015), the authors proposed various ways of constructing confidence regions for the underlying parameters in the high-dimensional setting. One major difference between these works and our framework is that they try to achieve full model inference without using the data to choose a hypothesis. The advantage of such approach is robustness. But in the highdimensional setting, with tens of thousands of potential variables, it is natural to use the data to select hypotheses of interest and perform valid inference only for those hypotheses. In addition, some of the full model inference works require conditions of linear underlying model Meinshausen et al. (2012), Javanmard & Montanari (2015) which the framework of selective inference does not require. For more philosophical discussions on the comparisons of the two approaches, see Fithian et al. (2014). 1.2. Organization of the paper In Section 2, we formally introduce the methods for selective inference with certain model selection procedures, which we call affine selection procedures. In Section 3, we state the main theorem that will allow asymptotically valid inference. In Section 4, we will illustrate the applications of our results to two selective inference problems, selective inference after solving the LASSO at a fixed λ, and the covariance test for testing the global null in generalized linear models. We collect all the proofs in Section 5 and dicuss the directions of future research in Section 6. 2. Selective inference with affine selection procedures Suppose we have a design matrix X ∈ Rn×p , considered fixed, and ind yi |xi ∼ G(µ(xi ), σ 2 (xi )) (2) where xi is the i-th row of the matrix X and G(µ, σ 2 ) denotes any one-dimensional distribution with mean µ and variance σ 2 . We also denote µ(X) = (µ(x1 ), . . . , µ(xn )) and Σ(X) = diag(σ 2 (xi ), . . . , σ 2 (xn )), a diagonal matrix with σ 2 (xi ) as the diagonal entries. Some feature selection procedure is then applied on the data to select a subet E ⊆ {1, 2, . . . , p} and the target of inference will be E ∗ (y, X) = {βj,E , j ∈ E}. In general, we consider certain selection procedures called the affine selection procedures, Definition 1 (Affine selection procedure). Suppose a model selection procedure E ∗ : Rn × Rn×p → S, where S is a finite set of models, S = {E1 , . . . , E|S| }. We call E ∗ an affine selection procedure, if the selection event can be written as an affine set in the first argument of E ∗ . Formally, E ∗ is an affine selection Tian and Taylor/Asymptotics of selective inference 4 procedure if for each potential model to be selected E ∈ S, {E ∗ (z, X) = E} = {A(E, X)z ≤ b(E, X)} , (z, X) ∈ Rn × Rn×p . (3) where A ∈ Rk×n , b ∈ Rk and k ∈ N are dependent only on E and X. Moreover, the sets {A(Ei , X)z ≤ b(Ei , X)} ⊆ Rn , i = 1, . . . , |S| are disjoint or their intersections have measure 0 under the Lebesgue measure on Rn . Examples of affine selection procedures include selection procedures that are based on E, the set of variables chosen by the data and usually some other information1 . Various algorithms can be used to select E, e.g. E as the active set of the LASSO solution at a fixed λ (Lee et al. 2013), E as the first variable to enter the LASSO or LARS path (Lockhart et al. 2013), (more generally any ℓ1 penalized generalized linear models) or E as the k variables included at the k-th step of forward stepwise selection (Taylor et al. 2014). The works of Lee et al. (2013), Lockhart et al. (2013), Taylor et al. (2014, 2013) have constructed valid p-values when the family G is the Gaussian family. Formally, the pivotal function depends on the following quantities, 2.1. Notations The pivotal function is determined by the following functions. For any A ∈ Rk×n , b ∈ Rk , Σ ∈ Rn×n and η ∈ Rn , we define AΣη , η T Ση bj − (Az)j + αj η T z L(z; A, b, Σ, η) = max , αj <0 αj α = α(A, b, Σ, η) = bj − (Az)j + αj η T z . αj >0 αj U (z; A, b, Σ, η) = min (4) (5) (6) Furthermore, we define F (x; σ 2 , m, a, b) = Φ((x − m)/σ) − Φ((a − m)/σ) Φ((b − m)/σ) − Φ((a − m)/σ) which is the CDF of the univariate Gaussian law N (m, σ 2 ) truncated to the interval [a, b]. 2.2. A pivotal quantity with Gaussian errors Theorem 1 provides the construction of a pivotal function when the data is normally distributed and E ∗ is an affine selection procedure. We denote the 1 See Section 4 for details Tian and Taylor/Asymptotics of selective inference 5 response variables to be Y when G is the Gaussian family to distinguish it from y where G is a more general location-scale family. Note all distributions in this paper are conditional on X, that is the law we consider are either L(Y|X) or L(y|X). All random variables have access to X as if it were a constant. Theorem 1 (Lee et al. (2013)). Suppose X ∈ Rn×p and Y ∼ N (µ(X), Σ(X)), µ(X) ∈ Rn , Σ(X) ∈ Rn×n and E ∗ is an affine selection procedure on Rn ×Rn×p . Then any for any η : Rn → Rn measurable with respect to σ(E ∗ ) we have F (η(E ∗ )T Y; η(E ∗ )T Ση(E ∗ ), η(E ∗ )T µ, LE ∗ (Y), UE ∗ (Y)) E ∗ (Y, X) = E ∼ Unif(0, 1), (7) where E ∗ (z, X) = E ⇐⇒ A(E, X)z ≤ b(E, X) and LE (z) = L(z, A(E, X), b(E, X), Σ, η(E)) UE (z) = U (z, A(E, X), b(E, X), Σ, η(E)). (8) (9) Moreover, marginalizing over the selection procedure E ∗ , we have the following F (η(E ∗ )T Y; η(E ∗ )T Ση(E ∗ ), η(E ∗ )T µ, LE ∗ (Y), UE ∗ (Y)) ∼ Unif(0, 1). (10) The significance of Theorem 1 is that assuming the diagonal matrix Σ is known, the only unknown parameter for the pivotal quantity (10) is η T µ. To test the hypothesis H0 : η T µ = 0, we just need to plug in the value and then compute (10), which then can be used as a p-value to accept/reject the hypothesis. For example, if we take T η = XE (XE XE )−1 ej , (11) where ej is the unit vector with only the j-th entry being 1, η T µ = βj,E . The quantity in (10) is pivotal and can be used to test the hypothesis H0j : βj,E = 0, and control the “Type-I errpr” (1). Since X is fixed, we use the shorthand E ∗ (z) = E ∗ (z, X), A(E) = A(E, X), b(E) = b(E, X). 3. Asymptotics with non-Gaussian error Now if we remove the assumption that the error L(Y|X) = N (µ(X), Σ), the conclusion of Theorem 1 does not hold any more. The best we can hope for is a weak convergence result that the same pivotal quantities (10) would converge to Unif(0, 1) (as n → ∞). This requires some conditions on both the distribution L(y|X) and the selection procedure E ∗ . Our main contribution in this work, Theorem 3 establishes conditions on L(y|X) and E ∗ under which the pivotal quantity (10) is asymptotically distributed as Unif(0, 1). The main approach is to compare the distribution of the pivots (10) under the distribution L(y|X) with that under Gaussian distribution L(Y|X). In the latter case, the exact distribution is derived in Theorem 1. In the following, we establish the conditions where the above two distributions are comparable. Tian and Taylor/Asymptotics of selective inference 6 3.1. Bounding the influence function Note the pivotal quantity in (10) depends on y either through the linear functions η T y or the maximum/minimum of linear functions LE ∗ (y), UE ∗ (y). In approximating the exact Gaussian theory with asymptotic results a quantity analogous to a Lipschitz constant (in y) will be necessary, expressing the changes in η T y as well as the upper and lower bounds LE ∗ and UE ∗ . This, in some sense, describes the influence each yi can have on the pivotal quantity (10). For an affine selection procedure E ∗ : Rn × Rn×p → S, without loss of generality suppose E ∗ is surjective. Since E ∗ is affine, for any model E ∈ S, there are the associated A(E) and b(E) as defined in (3). We define M (E, η) = max 1≤i≤nrow(A(E)) 1≤j≤n A(E)ij + kη(E)k∞ , (A(E)Ση(E))i (12) ∗ M (E , η) = max M (E, η). E∈S We also define r(E) = nrow(A(E)), r(E ∗ ) = max nrow(A(E)). E∈S (13) The quantity M (E ∗ , η) measures the maximal influence any yi has on a smoothed version of the triple (η(E ∗ )T y, LE ∗ (y), UE ∗ (y)). As M (E ∗ , η) and r(E ∗ ) are critical in bounding the difference between L(y|X) and L(Y|X), it is important to get a sense of their size. Typically r(E ∗ ) is less than p, and we discuss the typical size of M (E ∗ , η) through the following simple example: Example 3.1. Suppose the design matrix X ∈ Rn×p is generated in the following way: we first generate each row independently from a distribution on Rp , and then normalize the column of X to have length 1. Suppose instead of using data to select a model, we just arbitrarily choose a subset E. This is equal to no selection at all, thus M (E, η) = kη(E)k∞ . If we want to perform inference for βj,E , we take T η = XE (XE XE )−1 ej , where ej is the unit vector with only the j-th coordinate being 1. Since we normalT ize the columns, it is not hard to verify (XE XE )−1 = Op (1), and maxij (Xij ) = −1/2 Op (n ), thus if the selected variables set always satisfies |E| ≪ n, η = Op (n−1/2 ). Therefore M (E ∗ , η) = Op (n−1/2 ). This is a very simple example which does not involve selection. In reality we will some meaningful selection procedure that uses the data so M (E ∗ , η) would involve A(E ∗ ) and b(E ∗ ) as well. However, we will see through examples in Section 4 that it is still reasonable to assume M (E ∗ , η) = O(n−1/2 ). The following theorem compares the distribution of (η(E ∗ )T y, LE ∗ (y), UE ∗ (y)) under L(y|X) and its Gaussian counterpart. Theorem 2. Fix X ∈ Rn×p . Suppose (y, Y) are defined conditionally independent given X on a common probability space such that Tian and Taylor/Asymptotics of selective inference 7 • L(y|X) has independent entries with mean vector µ and covariance matrix variance Σ and finite third moments bounded by γ; • L(Y|X) = N (µ, Σ). Suppose we are given η ∈ σ(E ∗ ), then given any bounded function W ∈ C (R3 ; R) with bounded derivatives satisfying ( ≥ 0 if v ≤ u ≤ w W (u, v, w) = = 0 else 3 there exists N = N (M (E ∗ , η), |S|, r(E ∗ ), W ), such that the following holds for n, p ≥ N ,     EW η(E ∗ )T y, LE ∗ (y), UE ∗ (y) − EW η(E ∗ )T Y, LE ∗ (Y), UE ∗ (Y)  15    4 ∗ ∗ 3 ≤ C(W, γ) log r(E )|S| nM (E , η) (14) where C(W, γ) is a constant depending only on the derivatives of W and γ, and η(E ∗ ) is η(E ∗ (y)) or η(E ∗ (Y)) depending on the context. As it is reasonable to assume M (E ∗ , η) = O(n−1/2 ), it is reasonable to assume the RHS of (14) goes to zero. Thus the distribution of (η(E ∗ )T y, LE ∗ (y), UE ∗ (y)) is close to that of (η(E ∗ )T Y, LE ∗ (Y), UE ∗ (Y)). In the following, we discuss the conditions under which the pivotal quantity (10) converges. 3.2. Smoothness of the pivot Note the bound in (14) also depends on C(W, γ), the derivatives of W . Thus besides the influence of each yi on (10), it is also necessary to control the smoothness of the (10). In particular, the pivot in (10) takes the form of a truncated Gaussian cdf. Moreover, the smoothness (derivatives) of the truncated Gaussian cdf F (x; σ 2 , m, a, b) can depend heavily on the truncation interval [a, b]. More specifically, a lower bound on the denominator of F (x; σ 2 , m, a, b) puts some constraints on the width of the interval [a, b] as well as its distance to the origin. In our context, a, b corresponds to the upper and lower bounds appearing in (10). Formally, we assume the following assumption: Assumption 1. Suppose we have Xn ∈ Rn×pn and yn ∈ Rn is generated according to (2), and Yn is generated independently (conditional on Xn ) from N (µ(Xn ), Σ(Xn )) a Gaussian distribution with the same means and variances. We also have affine selection procedures E ∗ = En∗ . We assume there exists δn → 0 such that P(UE ∗ (yn ) − LE ∗ (yn ) < δn ) → 0, P(UE ∗ (Yn ) − LE ∗ (Yn ) < δn ) → 0, P(min(|UE ∗ (yn )|, |LE ∗ (yn )|) > 1/δn ) → 0, P(min(|UE ∗ (Yn )|, |LE ∗ (Yn )|) > 1/δn ) → 0. (15) Tian and Taylor/Asymptotics of selective inference 8 The first two conditions in (15) puts a lower bound on the width of the truncation interval (LE ∗ (yn ), UE ∗ (yn )). The last two conditions makes sure the truncation will not appear too far from the origin and thus we will have reasonable behavior in the tail. δn is the rate at which the truncation interval will shrink (or the distance of the truncation interval to the origin). This rate will appear in the RHS of (14) and thus we impose a condition on (δn , M (En∗ , ηn ), r(En∗ ), |Sn |) to ensure the convergence of the pivot (10). 3.3. Main result Suppose we have Xn ∈ Rn×pn and yn ∈ Rn is generated according to (2). We denote its distribution as L(yn |Xn ). The convergence mentioned below is under this sequence of distributions {L(yn |Xn )}∞ n=1 . Theorem 3 (Convergence of the pivot). Suppose we have a sequence of yn generated as above with means µn = µ(Xn ), and variances Σn = Σ(Xn ) and have finite third moments. We also assume Assumption 1 is satisfied with a sequence of δn . Furthermore, let En∗ be a sequence of affine selection procedures, ηn = η(En∗ ), and the corresponding M (En∗ , ηn ), r(En∗ ) and Sn properly defined as in Section 3.1. Then if     4 1/δn6 · M (En∗ , ηn )3 · n log r(En∗ ) + log |Sn | → 0, as n → ∞, we have d P (ηnT yn ; ηnT Σn ηn , ηnT µn , LEn∗ , UEn∗ ) → Unif(0, 1), n → ∞, (16) where P (x; σ 2 , m, a, b) = 2 min(F (x; σ 2 , m, a, b), 1 − F (x; σ 2 , m, a, b)) is the twosided pivot. In the following section, we apply Theorem 3 to different selection procedures. 4. Examples We give two examples in this section as the applications of Theorem 3. The first example is to perform selective inference after solving the LASSO and the second is to test the global null in generalized linear models. In these two examples, we will explain why the selection procedure is affine, what is the data distribution L(yn |Xn ) and the quantities (δn , M (En∗ , ηn ), r(En∗ ), |Sn |). To ease the notations, we suppress the dependencies on n whenever possible. It is helpful to keep in mind that yn ∈ Rn , Xn ∈ Rn×pn . 4.1. Inference for LASSO with non-Gaussian errors Consider the linear model y = Xβ 0 + ǫ, iid X ∈ Rn×p , ǫi ∼ G(0, σ 2 ), (17) Tian and Taylor/Asymptotics of selective inference 9 where σ is known and the distribution G has finite third moments, but is not necessarily Gaussian. Tibshirani (1996) proposed the now famous LASSO. We get a sparse solution β̂ by solving 1 β̂ = minimize ky − Xβk22 + λkβk1 , (18) β∈Rp 2 where λ > 0 is the fixed regularization parameter. We choose λ as in Negahban et al. (2012). If we normalize √ the columns of X to have norm 1, Negahban et al. (2012) chooses λ to be O( log p). 4.1.1. Affine selection procedure As in Lee et al. (2013), we solve (18) and get a solution β̂. Now we consider the selection procedure based on (E, zE ), where E = supp(β̂), zE = sign(β̂E ), where β̂E is β̂ restricted to the active set E. Note this is different from the selection procedure based only on E but is closely related, for detailed discussion see Lee et al. (2013). The authors in Lee et al. (2013) proved such selection procedure is equivalent to the affine constraints A(E, zE )y ≤ b(E, zE ), where T T A(E, zE ) = −diag(zE )(XE XE )−1 XE , T b(E, zE ) = −λdiag(zE )(XE XE )−1 zE . (19) To test the hypothesis H0j : βj,E = 0 for any j ∈ E, we choose η to be as in (11). In this case, a simple calculation will put the number of possible states at |S| = 2p , which will √ cause the bound in (14) to blow up when p > n. However, the choice of λ = O( log p) (Negahban et al. 2012) together with other conditions will ensure |S| is polynomial in p with high probability. √ 4.1.2. Number of states |S| for λ = O( log p) Suppose X is column standardized to be mean zero and norm 1, we first introduce the restricted strong convexity condition for matrix X. Definition 2 (Restricted strong convexity Negahban et al. (2012)). We say X ∈ Rn×p satisfies the restricted strong convexity condition for index set A with constant m > 0 if kXvk22 ≥ mkvk22 , for all v ∈ {∆ ∈ Rp : k∆Ac k1 ≤ 3k∆A k1 }. Now we define the assumptions needed to ensure |S| is polynomial in p with high probability. Tian and Taylor/Asymptotics of selective inference 10 Assumption 2. X satisfies the restriced convexity condition for A = supp(β 0 ) with constant m, and φmax , the biggest eigenvalue of X T X is bounded by a constant Q. Assumption 3. ǫi are sub-Gaussian errors with known variance σ 2 . Assumption 4. The signal is sparse. More specifically, k = |supp(β 0 )| is bounded by a constant K. Following Negahban et al. (2012), Lemma 1 shows with the above assumptions, the effective size of |S| is polynomial in p with high probability. √ Lemma 1. With Assumptions 2-4, if we solve (18) with λ ≥ 4σ log p and get active set E, then with probability at least 1 − c1 exp(−c1 λ2 ), |E| ≤ 16Q2 ·K m2 where c1 is some constant that depends on m and the subgaussian constant of the error ǫ. Thus, with probability 1 − c1 exp(−c1 λ2 ), |S| ≤ pcK , c= 16Q2 . m2 The proof of Lemma 1 is deferred to the appendix. Having controlled |S|, now we need to get a bound for the influences. 4.1.3. Bounding the influence M (E ∗ , η) Assume we have normalized the design matrix X columnwise so that each column has norm 1. We further assume the following assumption on X, Assumption 5. Suppose we solve problem (18) with X and get the active set E. Let φmin be the smallest eigenvalue for submatrices of size less than n × |E|, more specifically, kXvk22 . φmin = min v∈Rp ,kvk0 ≤|E| kvk2 2 We assume φmin ≥ ν > 0. Lemma 2. Suppose X satisfies Assumption 5, then T T max (XE XE )−1 XE i,j  ij ≤ |E| · max |Xij |. i,j ν2 4.1.4. Choice of δn in Assumption 1 √ If we normalize the columns of X to have norm 1 and choose λ = O( log p) et al. (2012). Then we assume Assumption 1 is satisfied in (18) as in Negahban √ with δn = O(( log pn )−1−κ ), for any small κ > 0. Tian and Taylor/Asymptotics of selective inference 11 To avoid long passage and stay focused on the main topic, we illustrate that Assumption 1 is satisfied with such δn ’s in the following simplified setup. However, the approach can be adapted to include more general cases. Lemma 3. Suppose Assumption 2-4 are satisfied. We further assume that zE = T 1 and the matrix (XE XE )−1 is equicorrelated, i.e.   T T XE )−1 jj = τ > 0, (XE XE )−1 ii = (XE  T (XE XE )−1 ij  > 0, ∀i, j ∈ E, i 6= j. ρ= T X )−1 (XE E ii ), for Then if kβn0 k∞ = O(λn ), Assumption 1 is satisfied with δn = O(λ−1−κ n any κ > 0. Remark 1. Note if we do not assume zE = 1, the last two conditions in As) and the first two conditions sumption 1 are still satisfied with δn = O(λ−1−κ n can be satisfied with further assumptions. But we do not pursue the technical details here. 4.1.5. Convergence of selective tests in the Lasso problems Suppose we solve the Lasso problem (18) and get active set E, and want to test the hypotheses H0j : βj,E = 0, we can simply take η to be as in (11). Now we summarize the above results and apply Theorem 3 to get the following corollary √ Lemma 4. Suppose we solve the Lasso problem (18) with λn = 4σ log pn , and Assumption 1-5 are satisfied and the δn ’s in Assumption 1 is chosen as √ 1 1 1 (log pn )− 2 − 2 κ . If we further assume max |Xij | = O(n− 2 ), kβ 0 k∞ = O( log pn ), and there exists κ > 0 such that n−1/2 (log pn )(7+3κ) → 0, then the pivot in (16) calculated with the appropriate (ηn , LEn∗ , UEn∗ ) converges to Unif(0, 1). Furthermore, we can construct a test for βj,E based on this pivot that controls “Type-I error” (1) asymptotically. 4.2. Covariance test for ℓ1 -penalized generalized linear models One of the first results in selective inference was the covariance test Lockhart et al. (2013) which provided an asymptotic limiting distribution for the first step of the Lasso or LARS path. An exact version of this test under Gaussian errors was described in Taylor et al. (2013). In the following, we generalize the covariance test for generalized linear models. Suppose L(y|x) is in an exponential family. More specifically, p(y|x; β 0 ) = b(y) exp[(xT β 0 )y − Λ(xT β 0 )], (20) Tian and Taylor/Asymptotics of selective inference 12 where β 0 and x are p-dimensional vectors and Λ(η) is the cumulant generating function of the distribution. Suppose yi |xi are independently distributed according to the law above, where xi ’s are considered fixed. Then the ℓ1 penalized generalized linear regression can be expressed as X β̂λ∗ = argmin − log p(yi |xi ; β) + λkβk1 . (21) β∈Rp 1≤i≤n The covariance test for the global null H0 : β 0 = 0 is based upon the the first knot on the solution path of (21), which is largest score statistic (in absolute values) at β 0 = 0, n o λ1 = sup λ : β̂λ∗ 6= 0 = kX T (y − ∇Λ(0))k∞ . (22) The variable to achieve the maximum in (22) will be the first variable to enter the solution path. The covariance test can also be viewed as a test for the coefficient with (potentially) the largest absolute values. A guess for such variable is the first variable to enter the solution path of (21). In other words, covariance tests select the target of inference based on (j ∗ , s∗ ), where   ∗ ∗ T T (j , s ) = argmax |xj (y − ∇Λ(0))|, sign(xj ∗ (y − ∇Λ(0))) , (23) j and the test statistic is λ1 = |xTj∗ (y − ∇Λ(0))|. 4.2.1. Affine selection procedure The selection procedure is based on (j ∗ , s∗ ) defined in (23), it is easy to see that it is equivalent to xTk (y − ∇Λ(0)) ≤ s∗ xTj∗ (y − ∇Λ(0)), −xTk (y − ∇Λ(0)) ≤ s∗ xTj∗ (y − ∇Λ(0)), k = 1, . . . , p, k = 1, . . . , p. Writing in the form of A(j ∗ , s∗ )y ≤ b(j ∗ , s∗ ), we have  T   T  x1 − s∗ xTj∗ (x1 − s∗ xTj∗ )∇Λ(0) ..   ..       .  T .∗ T   T   xp − s xj ∗   (xp − s∗ xTj∗ )∇Λ(0)     . A= T b= ∗ T , T ∗ T  −x1 − s xj ∗  −(x1 + s xj ∗ )∇Λ(0)     .. ..     . . −xTp − s∗ xTj∗ −(xTp + s∗ xTj∗ )∇Λ(0) We notice that λ1 = s∗ xTj∗ (y − ∇Λ(0)). Thus to test the global H0 : β 0 = 0, we simply take η = s∗ xj ∗ . Tian and Taylor/Asymptotics of selective inference 13 The challenge in establishing a result for the covariance test for GLM is the lack of Gaussianity in the data distribution. The tools we develop in this paper, however, can circumvent this. But we first need to establish the resulting pivot ind which we can use to test the hypothesis H0 : β 0 = 0. Note that yi | xi ∼ 2 G(µ(xi ), σ(xi ) ), thus if G were normal distribution, we will have an exact pivot by applying Theorem 1. This result is also given in Taylor et al. (2013). Formally, we have the following corollary. ind Corollary 1 (Global test for Gaussian errors). Suppose Yi ∼ N (µi , σi2 ), X ∈ Rn×p fixed, define µ = (µ1 , . . . , µn ), Σ = diag{σ12 , . . . , σn2 }. After getting the first knot on the solution path of (21), we get (j ∗ , s∗ ) as defined in (23) and λ1 = |xTj∗ (Y − µ)|. Furthermore, we also define Θjk = xTj Σxk and L(j ∗ ,s∗ ) = s(xk − Θj ∗ k /Θj ∗ j ∗ xj ∗ )T (Y − µ) , 1 − ss∗ Θj ∗ k /Θj ∗ j ∗ (s,k):s∈{−1,1},k6=j ∗ , sup 1−ss∗ Θj∗ k /Θj∗ j∗ >0 U(j ∗ ,s∗ ) = Then, inf ∗ s(xk − Θj ∗ k /Θj ∗ j ∗ xj ∗ )T (Y − µ) , 1 − ss∗ Θj ∗ k /Θj ∗ j ∗ , (s,k):s∈{−1,1},k6=j 1−ss∗ Θj∗ k /Θj∗ j∗ <0     Φ UΘ(jj∗∗,sj∗∗ ) − Φ Θjλ∗1j∗     ∼ Unif(0, 1) Φ UΘ(jj∗∗,sj∗∗ ) − Φ LΘ(jj∗∗,sj∗∗ ) (24) Corollary 1 gives a pivot (24) which we can use to test the global null H0 : β 0 = 0 and control the “Type-I error” (1). In practice, we often normalized the columns of the design matrix X. In addition we may assume the observations yi ’s are independently distributed with the same marginal variance, i.e. Σ = σ 2 I, then U(j ∗ ,s∗ ) = ∞ and L(j ∗ ,s∗ ) simplifies to the second knot in the solution path λ2 , thus we have: 1 − Φ(λ1 ) ∼ Unif(0, 1). (25) 1 − Φ(λ2 ) For the pivot (25) to converge to Unif(0, 1), we need to consider the number of states |S|, the bound on the influence M (j ∗ , s∗ ) as well as the choice of δn in Assumption 1. 4.2.2. The conditions for the pivot to converge Since j ∗ ∈ {1, . . . , p}, and s∗ ∈ {−1, 1}, the number of possible states |S| are naturally bounded by 2p and r(E ∗ ) = 2p. We assume Σ = σ 2 I, and X are normalized columnwise to have norm 1. We first introduce the following condition on the design matrix X, which states that any two columns of X cannot be too correlated. Tian and Taylor/Asymptotics of selective inference 14 Assumption 6. Suppose there exists ρ > 0, such that xTi xj ≤ ρ2 < 1, ∀ i 6= j, i, j ∈ {1, 2, . . . , p}. Under Assumption 6, it is not hard to verify M (j ∗ , s∗ ) ≤ 2 maxij |Xij | . 1 − ρ2 Now we need √ to pick the δn ’s such that Assumption 1 holds. In particular, we choose δn = ( log pn )−1−κ , for some κ > 0. Now if we apply Theorem 3, we have the following result, Corollary 2. Suppose y|X is generated independently coordinate-wise through the distribution in (20) with the same marginal variance. Assume the columns 1 of X have norm 1, Assumption 6 is satisfied and maxij |Xij | = O(n− 2 ). Then if 7+3κ − 1 n 2 → 0, log pn the pivot converges to Unif(0, 1) under the global null H0 : β 0 = 0, 1 − Φ(λ1 ) d → Unif(0, 1). 1 − Φ(λ2 ) 5. Proof of the theorems Without loss of generality, we restrict our interest to the case µ = µ(X) = 0, Σ = Σ(X) = I. This is possible since any affine selection procedure E ∗ applied to data with mean µ(X) 6= 0 is equivalent to a centered affine selection procedure E ∗,0 applied to the centered data. Specifically, the linear part of E ∗,0 is the same as E ∗ and the offsets are related by b0 (E) = b(E) − A(E)µ. Further, note that all quantities in the theorems above are independent of b. Scaling of the errors is handled in a similar fashion. 5.1. Proof of Theorem 1 Analogous to the proof in Lee et al. (2013), we prove Theorem 1. Proof. To lighten notations, we suppress all dependencies on X as it is assumed known. Note that {E ∗ (Y) = E} = {A(E)Y ≤ b(E)}. Thus     d T ∗ T L η(E) Y|E (Y) = E = L η(E) Y|A(E)Y ≤ b(E) . Tian and Taylor/Asymptotics of selective inference 15 Dropping the dependence on E for the moment, {AY ≤ b} = {AY − E[AY|η T Y] ≤ b − E[AY|η T Y]}, = {E[AY|η T Y] ≤ b − (AY − E[AY|η T Y])} = {αη T Y ≤ b − AY + αη T Y} = {αj η T Y ≤ bj − (AY)j + αj η T Y, j = 1, . . . , k}. In other words, {AY ≤ b} = {A(E)Y ≤ b(E)} = {LE (Y) ≤ η T Y ≤ UE (Y)}, and     d T ∗ T T L η(E) Y|E (Y) = E = L η(E) Y | LE (Y) ≤ η(E) Y ≤ UE (Y) , Note also from the derivation above that (LE (Y), UE (Y)) is independent of η T Y for each E. Thus if we condition on E ∗ , UE ∗ (Y) and LE ∗ (Y), η(E ∗ )T Y is distributed as a Gaussian r.v. with mean 0 and variance kη(E ∗ )k2 truncated at UE ∗ and LE ∗ . Therefore, F (η T Y; kηk2 , 0, LE ∗ (Y), UE ∗ (Y))|E ∗ = E, LE ∗ (Y), UE ∗ (Y) ∼ Unif(0, 1). Considering that conditional on E ∗ , η(E ∗ )T Y is independent of UE ∗ and LE ∗ , we have (7). 5.2. Smoothing the maxima of affine functions In the proof of Theorem 2 and the related lemmas and corollaries, a technique developed by Chatterjee (2005) is frequently used. Roughly speaking, we want to study convergence of functions like LE and UE which can be expressed as maxima or minima of affine functions. These non-smooth functions are replaced by a smoothed surrogate at the cost of a factor appearing in their derivatives depending on the smoothing parameter. Specifically, we are interested in how this smoothing affects the following quantities. Definition 3. For any f ∈ C 3 (Rn ; Rq ), define ) ( r/p p n . λr (f ) = sup |∂l fk (x)| , 1 ≤ p ≤ r, l = 1, . . . , n, k = 1, . . . , q, x ∈ R (26) For any finite collection F of functions define λr (F ) = max λr (f ). f ∈F 3 Definition 4. For any g ∈ C (D, R) where D ⊆ R3 and any multi-index α = (α1 , α2 , α3 ), we define for r = 1, 2, 3, ) # " ( r/|α| ∂ |α| g(u, v, w) , (u, v, w) ∈ D, |α| ≤ r , 1 . Cr (g) = max sup ∂uα1 ∂v α2 ∂wα3 Tian and Taylor/Asymptotics of selective inference 16 Now we define the smoothed maxima operator. Definition 5. Let f : Rn → R3 and define f = maxv∈F v, where  F = z 7→ vj (z), vj ∈ C 3 (Rn , R3 ) is a finite collection of thrice differentiable functions vj ’s. The maximum is taken coordinate-wise. We define the smoothed maxima operator with parameter β as   X 1 exp (βvj ) ∈ C 3 (Rn , R3 ), (27) Γ(f, β) = log  β vj ∈F where the operators log and exp are applied coordinate-wise. Suppose the range of f, Γ(f, β), denoted as R(f ), R(Γ(f, β)) ⊆ D and let h = g ◦ f , hβ = g ◦ Γ(f, β), then Lemma 5 gives a bound on kh − hβ k∞ and λ3 (hβ ). Lemma 5. Assume the same notations as above, s = |F |, then for β ≥ 1 3 log s, β λ3 (hβ ) ≤ 13c · β 2 C3 (g)λ3 (F ), kh − hβ k∞ ≤ C1 (g) · (28) (29) where c is a universal constant. The proof of Lemma 5 will refer to the following lemma whose proof we leave in the Appendix. Lemma 6. For any f ∈ C 3 (Rn ; R3 ) and g ∈ C 3 (R3 , R), r = 1, 2, 3 λr (g ◦ f ) ≤ cCr (g) · λr (f ), ∀l = 1, 2, . . . , n, x ∈ Rn , where c is a universal constant. Now we prove Lemma 5. Proof. Note that for any u ∈ Rs    1 log exp β max uj 1≤j≤s β   s X 1 exp (βuj ) ≤ log  β j=1    1 ≤ log s exp β max uj ) 1≤j≤s β 1 = log s + max uj . 1≤j≤s β max uj = 1≤j≤s (30) Tian and Taylor/Asymptotics of selective inference 17 We take h = g ◦ f , and hβ = g ◦ Γ(f, β), |h(z) − hβ (z)| = |g ◦ f (z) − g ◦ Γ(f, β)(z)| ≤ 3C1 (g)kf (z) − Γ(f, β)(z)k∞ 1 ≤ 3C1 (g) · log s, β where the ∞ norm is the element-wise maximum absolute value. Thus we proved (28). Now let f = (f1 , f2 , f3 ) and vj = (v1j , v2j , v3j ), and define  Fi = z 7→ vij (z), vij ∈ C 3 (Rn , R) , i = 1, 2, 3. Theorem 1.3 in Chatterjee (2005) proved that λ3 (Γ(fi , β)) ≤ 13β 2 λ3 (Fi ), ∀β ≥ 1, i = 1, 2, 3. (31)   Note λ3 Γ(f, β) = maxi=1,2,3 λ3 Γ(fi , β) , and that λ3 (F ) = maxi=1,2,3 λ3 (Fi ), thus λ3 (Γ(f, β)) ≤ 13β 2 λ3 (F ). (32) This combined with Lemma 6 proves (29). 5.3. Proof of Theorem 2 To prove Theorem 2, we first prove the following lemma. Recall our reduction to the standard Gaussian N (0, I) in the beginning of Section 5.1. Lemma 7 is a simple adaption of Lindberg’s proof of the CLT. Lemma 7. Assuming the same notation as in Theorem 2, for any smooth function h ∈ C 3 (Rn ), |Eh(y) − Eh(Y)| ≤ 1 λ3 (h)n max(γ, E(|Yl |3 )) l 6 (33) We will prove Lemma 7 now. Proof. The proof proceeds by following the Lindberg proof of the CLT for h. Define y l = (y1 , y2 , . . . , yl , Yl+1 , . . . , Yn ), Y l = (y1 , y2 , . . . , yl−1 , Yl , . . . , Yn ), W l = (y1 , y2 , . . . , yl−1 , 0, Yl+1 , . . . , Yn ). We can break the absolute difference of the two expectations into n parts, |Eh(y) − Eh(Y)| ≤ n X l=1 |Eh(y l ) − Eh(Y l )|. (34) Tian and Taylor/Asymptotics of selective inference 18 Note that 1 h(y l ) − h(W l ) = ∂l h(W l )yl + ∂l2 h(W l )yl2 + Rl , 2 1 h(Y l ) − h(W l ) = ∂l h(W l )Yl + ∂l2 h(W l )Yl2 + T l , 2 where |Rl | ≤ 16 k∂l3 hk∞ |yl |3 , |T l | ≤ 61 k∂l3 hk∞ |Yl |3 . Moreover, because yl ’s and Yl ’s are independent, W l is independent of both yl and Yl . Continuing, we see the first and second order differences cancel out, |Eh(y l ) − Eh(Y l )| 1 2 E∂ h(W l )(yl2 − Yl2 ) + E|Rl − T l | 2 l 1 =|E∂l h(W l )E(yl − Yl )| + E∂l2 h(W l )E(yl2 − Yl2 ) + E|Rl − T l | 2 ≤|E∂l h(W l )(yl − Yl )| + =E|Rl − T l |. Combining the n parts, we have |Eh(y) − Eh(Y)| ≤ n X   1 λ3 (h) E(|Yl |3 ) + E(|yl |3 ) 6 l=1 1 ≤ λ3 (h)n max(γ, E(|Yl |3 )) l 6 Now we turn to the proof of Theorem 2. Proof. Note that since {A(Ei , X)y ≤ b(Ei )}, 1 ≤ i ≤ |S| are disjoint and for any state E {A(E, X)y ≤ b(E)} = {LE (y) ≤ η(E)T y ≤ UE (y)}. Therefore the quantity of interest is W (η(E ∗ )T y, LE ∗ (y), UE ∗ (y)) = max W (η(Ei )T y, LEi (y), UEi (y)) 1≤i≤|S| If we knew the above quantity was smooth with respect to the data, we can apply Lemma 7 directly. However, there are two non smooth expressions above: the maximum function over the states and in LEi and UEi . We smooth each and optimize over the smoothing parameter. Tian and Taylor/Asymptotics of selective inference 19 We first smooth over UE and LE . Note that ∀z ∈ Rn b(E)j − (A(E)z)j + αj η T z αj <0 αj LE (z) = max = max{v|v ∈ FL }, b(E)j − (A(E)z)j + αj η T z αj >0 αj UE (z) = min = min{v|v ∈ FU }, where FL , FU are the collections of affine functions     b(E)j − (A(E)z)j + αj η T z αj < 0 FL = vj : z 7→ αj     b(E)j − (A(E)z)j + αj η T z FU = vj : z 7→ αj > 0 . αj Finally, note that max(λ3 (FL ), λ3 (FU )) ≤ M (E, η)3 . (35) We define the smoothing parameter β = 1/δ, for some 0 < δ < 1. Then     X v (z) j , exp L̃E,δ (z) = Γ(LE (z), 1/δ) = δ log  δ vj ∈FL   (36)   X v (z) j , ŨE,δ (z) = −Γ(−UE (z), 1/δ) = −δ log  exp − δ vj ∈FU By Lemma 5, for any selection state E and z ∈ Rn , we see |LE (z) − L̃E,δ (z)| ≤ δ log r(E ∗ ) |UE (z) − ŨE,δ (z)| ≤ δ log r(E ∗ ) (37) Based on (37) and Lemma 5, we have |EW (η(E ∗ )T y, LE ∗ (y), UE ∗ (y)) − EW (η(E ∗ )T Y, LE ∗ (Y), UE ∗ (Y))| ≤|EW (η(E ∗ )T y, L̃E ∗ ,δ (y), ŨE ∗ ,δ (y)) − EW (η(E ∗ )T Y, L̃E ∗ ,δ (Y), ŨE ∗ ,δ (Y))| + 12δC1 (W ) log r(E ∗ ) =|E max W (η(Ei )T y, L̃Ei ,δ (y), ŨEi ,δ (y)) − E max W (η(Ei )T Y, L̃Ei ,δ (Y), ŨEi ,δ (Y))| Ei Ei ∗ + 12δC1 (W ) log r(E ) (38) The equality follows from the fact that UE ≥ ŨE,δ and LE ≤ L̃E,δ , for any state E and that W is supported on D = {(u, v, w)|v ≤ u ≤ w}. Therefore, W (η(E ∗ )T y, L̃E ∗ ,δ (y), ŨE ∗ ,δ (y)) = maxEi W (η(Ei )T y, L̃Ei ,δ (y), ŨEi ,δ (y)). Tian and Taylor/Asymptotics of selective inference 20 Next, we smooth the maximum over states. Define WE (z) = W (η T z, L̃E,δ (z), ŨE,δ (z)) E ∈ S, z ∈ Rn . We also define the smoothed maxima for max1≤i≤|S| WEi (z), " #  X WEi (z) . exp Hδ (z) = Γ( max WEi (z), 1/δ) = δ log δ 1≤i≤|S| i Thus by Lemma 5, |EW (η(E ∗ )T y, LE ∗ (y), UE ∗ (y)) − EW (η(E ∗ )T Y, LE ∗ (Y), UE ∗ (Y))| ≤12δ log |S| + 12δC1 (W ) log r(E ∗ ) + |EHδ (y) − EHδ (Y)|, (39) For any state E, define feδ,E : z 7→ (η(E)T z, L̃E,δ (z), ŨE,δ (z)). Theorem 1.3 in Chatterjee (2005) states that λ3 (Hδ ) ≤ 13 max λ3 (W ◦ feδ,Ei ). δ2 i (40) Per Lemma 5 and inequality (35) 13c λ3 (W ◦ feδ,E ) ≤ 2 C3 (W )M (E ∗ , η)3 . δ (41) From (33) in Lemma 7 together with (40) and (41), we have |EHδ (y) − EHδ (Y)| ≤ 169c C3 (W )nM (E ∗ , η)3 max(γ, E[Yl3 ]) 6δ 4 (42) All the above combined, we have |EW (η(E ∗ )T y, LE ∗ (y), UE ∗ (y)) − EW (η(E ∗ )T Y, LE ∗ (Y), UE ∗ (Y))| 169c 1 ≤12δ [log |S| + C1 (W ) log r(E ∗ )] + · C3 (W )nM (E ∗ , η)3 max(γ, E[Yl3 ]). 6 δ4 Notice that for the last inequality to hold, we require δ ≤ 1. But the optimal δ 5 = O(nM (E ∗ , η)3 /[log |S| + log r(E ∗ )]), which will go to 0 since the numerator shrinks to 0 while the denominator goes to ∞ as n, p → ∞. Therefore, the inequality holds. Optimizing over δ yields (14). 5.4. Proof of Theorem 3 Now let’s turn to the proof of our main result, Theorem 3, Proof. For the convenience of notation, we denote P (x; σ 2 , m, a, b) by P (x; a, b), omitting σ 2 , m in the following proof. Define D(δ) = {(x, a, b) : a ≤ x ≤ b, b − a ≥ δ, min(|b|, |a|) ≤ 1/δ} . Tian and Taylor/Asymptotics of selective inference We claim that for any small function Peδ such that 1 4 21 > δ > 0, we can find a thrice differentiable Peδ is supported on the set {a ≤ x ≤ b} Peδ − P ∞ ≤ (K1 + 1)δ on the set D(δ), 1 on the set D(δ), δ6 where K1 and K3 are defined in Corollary 3. The proof of the existence of such a function Peδ is left to Lemma 9 in the Appendix. Then for any positive, bounded function Ψ ∈ C 3 (R+ ; R+ ), Ψ(0) = 0, with bounded third derivatives, we have: C3 (Peδ ) ≤ K3 · |EΨ ◦ Peδ (y; LE ∗ , UE ∗ ) − EΨ ◦ P (y; LE ∗ , UE ∗ )|   ′ ≤ kΨ k∞ (K1 + 1)δ + 2 kΨk∞ · P UE ∗ (y) − LE ∗ (y) < δ   +2 kΨk∞ · P min(|UE ∗ (y)|, |LE ∗ (y)|) > 1/δ , |EΨ ◦ Peδ (Y; LE ∗ , UE ∗ ) − EΨ ◦ P (Y; LE ∗ , UE ∗ )|   ≤ kΨ′ k∞ (K1 + 1)δ + 2 kΨk∞ P UE ∗ (Y) − LE ∗ (Y) < δ   +2 kΨk∞ P min(|UE ∗ (Y)|, |LE ∗ (Y)|) > 1/δ , (43) On the other hand, we plug in Ψ ◦ Peδ as the W in Theorem 2, then for any sequence of Peδn ,  15  1 ∗ 4 ∗ 3 e e |EΨ ◦ Pδn (y) − EΨ ◦ Pδn (Y)| ≤ C K3 · 6 (log r(E )|S|) nM (E , η) . (44) δn If we choose a subsequence δn → 0, such that the right hand side of (44) goes to zero, then |EΨ ◦ P (y) − EΨ ◦ P (Y)| ≤|EΨ ◦ Peδn (y) − EΨ ◦ Peδn (Y)| + 2 kΨ′ k∞ (K1 + 1)δn   +2 kΨk∞ P UE ∗ (y) − LE ∗ (y) < δn   +2 kΨk∞ P min(|UE ∗ (y)|, |LE ∗ (y))| > 1/δn   +2 kΨk∞ P UE ∗ (Y) − LE ∗ (Y) < δn   +2 kΨk∞ P min(|UE ∗ (Y)|, |LE ∗ (Y)|) > 1/δn → 0. (45) Tian and Taylor/Asymptotics of selective inference 22 Note that the right hand side of (45) goes to 0 because of Assumption 1. Thus d we have the conclusion that P (y) → P (Y) ∼ Unif(0, 1). 6. Discussion This work proves a generic framework in which asymptotic results hold for many selective inference problems. It is, however, not directly applicable to some other procedures. Further work may include, 1. Fixed λ for generalized linear model. Our work derives a theory for inference after the affine selection procedure. However, inference for a fixed λ for the generalized linear regression is not an affine selection procedure. A plausible solution will be to approximate the loss function of GLM by a quadratic form and bound the difference between the quadratic form and the GLM loss function. However, this is still an open question. 2. Apply the result to nonparametric problems. It is a big step to remove the Gaussian assumptions required by Lee et al. (2013) which restricts our attention to Gaussian families. Without the Gaussian constraints, we can consider some exponential families and potentially some nonparametric problems as well. Acknowledgements We thank Jason Lee for the proof of Lemma 1 provides a polynomial bounds on the number of selection states. We also thank Will Fithian and Rob Tibshirani for discussions on potential applications of our result. We also thank the anonymous reviewer who has read our paper so carefully and provide constructive advice for reorganization. References Belloni, A., Chen, D., Chernozhukov, V. & Hansen, C. (2012), ‘Sparse models and methods for optimal instruments with an application to eminent domain’, Econometrica 80(6), 2369–2429. Berk, R., Brown, L., Buja, A., Zhang, K. & Zhao, L. (2013), ‘Valid post-selection inference’, The Annals of Statistics 41(2), 802–837. URL: http://projecteuclid.org/euclid.aos/1369836961 Chatterjee, S. (2005), ‘A simple invariance theorem’, arXiv preprint math/0508213 . Fithian, W., Sun, D. & Taylor, J. (2014), ‘Optimal inference after model selection’, arXiv:1410.2597 [math, stat] . arXiv: 1410.2597. URL: http://arxiv.org/abs/1410.2597 Javanmard, A. & Montanari, A. (2015), ‘De-biasing the lasso: Optimal sample size for gaussian designs’, arXiv preprint arXiv:1508.02757 . Lee, J. D., Sun, D. L., Sun, Y. & Taylor, J. E. (2013), ‘Exact inference after model selection via the lasso’, arXiv preprint arXiv:1311.6238 . Tian and Taylor/Asymptotics of selective inference 23 Lockhart, R., Taylor, J. E., Tibshirani, R. J. & Tibshirani, R. (2013), ‘A significance test for the lasso’, arXiv: 1301.7161 . Meinshausen, N., Meier, L. & Bühlmann, P. (2012), ‘P-values for highdimensional regression’, Journal of the American Statistical Association . Negahban, S. N., Ravikumar, P., Wainwright, M. J. & Yu, B. (2012), ‘A unified framework for high-dimensional analysis of MM-Estimators with decomposable regularizers’, Statistical Science 27(4), 538–557. URL: http://projecteuclid.org/euclid.ss/1356098555 Taylor, J., Lockhart, R., Tibshirani, R. J. & Tibshirani, R. (2014), ‘Postselection adaptive inference for least angle regression and the lasso’, arXiv:1401.3889 [stat] . URL: http://arxiv.org/abs/1401.3889 Taylor, J., Loftus, J., Tibshirani, R. & Tibshirani, R. (2013), ‘Tests in adaptive regression via the kac-rice formula’, arXiv preprint arXiv:1308.3020 . Tibshirani, R. (1996), ‘Regression shrinkage and selection via the lasso’, Journal of the Royal Statistical Society. Series B (Methodological) pp. 267–288. Tibshirani, R. J., Rinaldo, A., Tibshirani, R. & Wasserman, L. (2015), ‘Uniform asymptotic inference and the bootstrap after model selection’, arXiv preprint arXiv:1506.06266 . Zhang, C.-H. & Zhang, S. S. (2014), ‘Confidence intervals for low dimensional parameters in high dimensional linear models’, Journal of the Royal Statistical Society: Series B (Statistical Methodology) 76(1), 217–242. Appendix A: Proof of Lemma 6 Proof. Using the chain rules, we have the second derivatives with respect to xl , l = 1, 2, . . . , n as   X 3 3 X ∂g ∂ 2 fi ∂2g ∂fi ∂fj ∂2g ◦ f + = ∂x2l ∂fi ∂fj ∂xl ∂xl ∂fi ∂x2l i=1 i,j=1 (46) and the third derivatives with respect to xl , l = 1, 2, . . . , n as    2  3 3 X X ∂ 2g ∂fi ∂fj ∂fk ∂ fi ∂fj ∂3g ∂3g ◦ f +3 = ∂x3l ∂fi ∂fj ∂fk ∂xl ∂xl ∂xl ∂fi ∂fj ∂x2l ∂xl i,j=1 i,j,k=1 3 X ∂g ∂ 3 fi + . ∂fi ∂x3l i=1 (47) For r = 1, the conclusion is obviously true with the constant c = 3. For r = 2, the terms involving the partial derivatives of f are ∂ 2 fi ∂fi ∂fj · or , ∀i, j = 1, 2, . . . , n. ∂xl ∂xl ∂x2l Tian and Taylor/Asymptotics of selective inference 24 Note the first type of terms are bounded by λ2 (f )1/2 · λ2 (f )1/2 and the second type of terms are bounded by λ2 (f ). If we take c = 12, |∂l2 g ◦ f | ≤ cC2 (g)λ2 (f ). On the other hand, #2 " 3 X ∂g ∂fi 1 1 2 (48) ≤ [3C2 (g) 2 λ2 (f ) 2 ]2 ≤ 9C2 (g)λ2 (f ). |∂l g ◦ f | ≤ ∂f ∂x i l i=1 For r = 3, an equation similar to (48) will give us |∂l g ◦ f |3 ≤ 27C3 (g)λ3 (f ). Meanwhile,  3   X  2  2 3  3  2 X ∂g ∂ g ∂f ∂f ∂ f j i i  |∂l2 g ◦ f | =  + 2 ∂f ∂f ∂x ∂x ∂f ∂x i j l l i l i=1 i,j=1 3 2 h i 23 2 1 1 1 2 ≤ 9C3 (g) 3 (λ3 (f ) 3 · λ3 (f ) 3 ) + 3C3 (g) 3 (λ3 (f ) 3 ) 3 = 12 2 C3 (g)λ3 (f ). For the third derivatives ∂l3 g ◦ f , the terms that involve f are, ∂fi ∂fj ∂fk ∂ 2 fi ∂fj ∂ 3 fi · · or or · 2 ∂xl ∂xl ∂xl ∂xl ∂xl ∂x3l which are all bounded by λ3 (f ) and therefore λ3 (g ◦ f ) ≤ 57C3 (g)λ3 (f ). In summary, we can take c = 57. Appendix B: Existence of smooth approximation Pe We prove the existence of such functions as claimed in the proof of Theorem 3. Define P (x, a, b) = P (x; σ 2 , m, a, b). We first prove the following lemma, Lemma 8. Define   D(δ) = (x, a, b) : a ≤ x ≤ b, b − a ≥ min δ, 1 min(|b|, |a|)  . Then, on D(δ) for any δ < 1/4 we have 2 max 2 2 e−x /2 e−a /2 e−b /2 , , Φ(b) − Φ(a) Φ(b) − Φ(a) Φ(b) − Φ(a) !  ≤ C max δ −1 , min(|a|, |b|) for some universal constant C. Proof. Note that for any δ > 0 on D(δ) ∩ {(x, a, b) : sign(a) = sign(b)} we have e−x 2 /2 ≤ max(e−a 2 /2 2 , e−b /2 ) Tian and Taylor/Asymptotics of selective inference 25 so it suffices to prove that 2 2 max e−b /2 e−a /2 , Φ(b) − Φ(a) Φ(b) − Φ(a) ! ≤ Cδ −1 on D(δ) for δ < 1/4 as well as 2 e−x /2 ≤ Cδ −1 . (x,a,b)∈D(δ)∩{(x,a,b):sign(a)6=sign(b)} Φ(b) − Φ(a) sup Let’s consider this latter case first. For any δ > 0 on the set D(δ)∩{(x, a, b) : sign(a) 6= sign(b)} we have 2 e−x /2 1 ≤ Φ(b) − Φ(a) Φ(b) − Φ(a) 1 inf (a,b):sign(a)6=sign(b),b−a≥δ Φ(b) − Φ(a) √ 2π ≤ −δ2 /2 . δe ≤ Continuing, we further split the first case into two cases, i.e. D(δ)∩{(x, a, b) : sign(a) 6= sign(b)} and D(δ) ∩ {(x, a, b) : sign(a) = sign(b)}. For the first part, analogous to the analysis above, we have ! 2 2 1 e−a /2 e−b /2 ≤ max , Φ(b) − Φ(a) Φ(b) − Φ(a) Φ(b) − Φ(a) 1 inf (a,b):sign(a)6=sign(b),b−a≥δ Φ(b) − Φ(a) √ 2π ≤ −δ2 /2 . δe ≤ Now, we reduce to the case D(δ) ∩ {(x, a, b) : sign(a) = sign(b)} and without loss of generality, we consider the case where 0 < a < b. Note that for any δ > 0 on D(δ) ∩ {(x, a, b) : sign(a) = sign(b)} ∩ {0 < a < b} ! 2 2 2 e−a /2 e−a /2 e−b /2 ≤ sup max , , Φ(b) − Φ(a) Φ(b) − Φ(a) a:a<1/δ Φ(a + δ) − Φ(a) ! 2 2 2 e−a /2 e−b /2 e−a /2 ≤ sup , , max Φ(b) − Φ(a) Φ(b) − Φ(a) a:a>1/δ Φ(a + 1/a) − Φ(a) For δ < 1/4 and 0 < a < 1/δ, (2π)1/2 (Φ(a + δ) − Φ(a)) ≥ δe−(a+δ) = δe−a 2 2 /2 /2 −aδ−δ 2 /2 e Tian and Taylor/Asymptotics of selective inference Or, 26 2 2 e−a /2 ≤ δ −1 (2π)1/2 eaδ+δ /2 Φ(a + δ) − Φ(a) ≤ C1 δ −1 , where C1 = (2π)1/2 e1+1/32 . Similarly, for 0 < 1/δ < a, 1 −(a+1/a)2 /2 e a 2 2 1 = e−a /2 e−1−1/(2a ) a (2π)1/2 (Φ(a + 1/a) − Φ(a)) ≥ Or, 2 2 e−a /2 ≤ a(2π)1/2 e1+1/(2a ) Φ(a + 1/a) − Φ(a) ≤ a(2π)1/2 e1+1/(2a ≤ a(2π)1/2 e 1/2 1+1/32 where C1 = (2π) {0 < a < b}, e ) 1+ 12 δ 2 ≤ C1 a . Therefore, we have on D(δ)∩{(x, a, b) : sign(a) = sign(b)}∩ 2 max 2 2 e−a /2 e−b /2 , Φ(b) − Φ(a) Φ(b) − Φ(a) ! ≤ max   1 , min(|a|, |b|) . δ Remark 2. If we take D(δ) = {(x, a, b) : a ≤ x ≤ b, b − a ≥ δ, min(|a|, |b|) ≤ 1/δ} , then 2 max 2 2 e−x /2 e−a /2 e−b /2 , , Φ(b) − Φ(a) Φ(b) − Φ(a) Φ(b) − Φ(a) ! ≤ Cδ −1 , where C is a universal constant. Corollary 3. For δ < 1/4 and any multi-index α = (α1 , α2 , α3 ) we have sup (x,a,b)∈D(δ) C|α| (P ) ≤ K|α| δ −|α| . for constants Kl , l ≥ 1. Proof. We prove for α = (0, 0, 1), and similar proofs can be extend to other multi-index α as well. Since P (x, a, b) = 2 min(F (x; a, b), 1 − F (x; a, b)), we only need to prove for F (x; a, b). 1 Φ(x) − Φ(a) ∂F · √ · exp(−b2 /2). =− 2 ∂b [Φ(b) − Φ(a)] 2π Tian and Taylor/Asymptotics of selective inference 27 Therefore, Φ(x) − Φ(a) exp(−b2 /2) 1 1 ∂F ≤C . ≤√ · · ∂b δ 2π Φ(b) − Φ(a) Φ(b) − Φ(a) Finally, we put the lemma and the corollary together and prove the following lemma. Lemma 9. There exists a thrice differentible approximation Pe to P that satisfies, • Pe(x, a, b) is supported on {(x, a, b) : a ≤ x ≤ b}, 1 • C3 (Pe ) ≤ K3 6 on the set D(δ), δ • Pe ≤ (K1 + 1)δ on the set D(δ). ∞ Proof. Let Pδ be the smoothed version of P for the minimum function in P = 2 min(F, 1 − F ), and kPδ − P k∞ ≤ δ. Let Peδ = Pδ Iδ2 (x, a, b), where Iδ2 (x, a, b) is the smoothed version of the indicator function on {a ≤ x ≤ b}. Iδ2 (x, a, b) also satifies the condition that   x < a, or x > b, 0 2 Iδ (x, a, b) = 1 a + δ2 ≤ x ≤ b − δ2,   [0, 1] else. 1 , for some universal constant C. Thus it is δ6 not hard to verify that C3 (Peδ ) ≤ K3 δ16 . Iδ2 (x, a, b) also satisfies C3 (Iδ2 ) ≤ Peδ − P ∞ ≤ kP − P Iδ2 k∞ + kPδ Iδ2 − P Iδ2 k∞ ≤ sup a≤x≤a+δ 2 , or b−δ 2 ≤x≤b kP k∞ + δ ≤ C1 (P ) · δ 2 + δ ≤ (K1 + 1)δ. Appendix C: LASSO related proofs C.1. Proof of Lemma 1 We first introduce the following Lemma in Negahban et al. (2012). Tian and Taylor/Asymptotics of selective inference 28 Lemma 10. If we assume the same assumptions and notations as in Lemma 1, then with probability at least 1 − c1 exp(−c1 λ2 ), the following two inequalities hold: λ ≥ 2kXǫk∞, r 8 k log p 0 , kβ̂ − β k2 ≤ m n (49) (50) where k is the number of nonzero entries in β 0 and β̂ is the solution to (18) Proof of Lemma 1, Proof. According to the KKT conditions, ( xTj (y − X β̂) = λsign(β̂), |xTj (y − X β̂)| ≤ λ if j ∈ A, if j ∈ 6 A. According to Lemma 10, we assume both (49) and (50) hold. This happens with probability 1 − c1 exp(−c1 λ2 ). For any j, xTj (y − X β̂) = xTj (Xβ 0 − X β̂ + ǫ) = xTj X(β 0 − β̂) + xTj ǫ λ ≥ xTj X(β 0 − β̂) − . 2 Thus for j ∈ A, λ , 2 3λ . |xTj X(β̂ − β 0 )| ≤ 2 |xTj X(β̂ − β 0 )| ≥ kX T X(β̂ − β 0 )k22 = ≥ ≥ X j∈A X j∈A 2 X  2 xj X(β̂ − β 0 ) + xj X(β̂ − β 0 ) 2 xj X(β̂ − β 0 ) j6∈A λ2 |supp(β̂)|. 4 Also, kX T X(β̂ − β 0 )k22 ≤ kX T Xk22 kβ̂ − β 0 k22 2 1 2 4kλ kXk · 2 n2 m2 2 4kλ ≤ φ2max 2 . m ≤ Tian and Taylor/Asymptotics of selective inference 29 Combining the two inequalities, we have that |supp(β̂)| ≤ 16kφ2max m2 holds with probability 1 − c1 exp(−c1 λ2 ). C.2. Proof of Lemma 2 † T T Proof. Note XE = (XE XE )−1 XE . According to the assumption assumed in Lemma 2, φmin > ν. Thus for any possible active set E,   T max | (XE XE )−1 ij | ≤ 1/ν 2 . i,j The above result can be easily obtained using Singular Value Decomposition on XE . Therefore, we have † max |XE |ij ≤ |E|/ν 2 max |Xij | i,j i,j C.3. Proof of Lemma 3 Proof. Without loss of generality, we assume β 0 = 0. We first see that for any fixed E, and ηn chosen as in (11), the upper and lower bound simplies to UE = ∞, LE = max k∈E,k6=j λ · τ [1 + (|E| − 1)ρ] − β̄k,E + β̄j,E , ρ where β̄E ∈ R|E| is the least square estimator with the E variables, T T β̄E = (XE XE )−1 XE y. Note that the first two equations of (15) are automatically satisfied in this case. Without loss of generality, we assume LE ∗ > 0, and noticing LE ∗ ≤ maxE LE , we have P(LE ∗ (yn ) > 1/δn ) ≤ P(max LE (yn ) > 1/δn ). E Since maxE LE (yn ) is the maximum of at most pcK sub-Gaussian variables, thus κ the RHS is bounded by O(e−λn ) = O(p−κ ), which goes to 0. C.4. Proof of Lemma 4 Proof. Per Lemma 1, we have |En | ≤ cK, |Sn | ≤ pcK , Tian and Taylor/Asymptotics of selective inference 30 with probability at least 1 − c1 exp(−c1 λ2n ). We modify the selection procedure En∗ on the small probability event. More specifically, we define Een∗ as ( En∗ (yn , Xn ), if |En | ≤ cK, ∗ Een (yn , Xn ) = no selection, else. It is easy to see that Een∗ is also an affine selection procedure. which differs from En∗ only on the event {|En | > cK}. Thus the pivots formed with En∗ and Een∗ converge in probability,  P P (ηnT yn ; ηnT Σn ηn , ηnT µn , LEn∗ , UEn∗ )  c1 6= P (ηnT yn ; ηnT Σn ηn , ηnT µn , LEe∗ , UEe∗ ) ≤ c1 → 0. n n p Therefore, we only need to consider the asymptotic distribution of the pivot with Een∗ as the selection procedure. Note that for Een∗ , M (Een∗ , ηn ) ≤ cK · max |Xij |, ν2 r(Een∗ ) ≤ p, |Sn | ≤ pcK . Now with our choice of δn ’s, it is easy to rewrite the condition in Theorem 3 as n−1/2 (log pn )7+3κ → 0.
10math.ST
DeepMovie: Using Optical Flow and Deep Neural Networks to Stylize Movies Alexander G. Anderson1 , Cory P. Berg1 , Daniel P. Mossing1 , and Bruno A. Olshausen1 arXiv:1605.08153v1 [cs.CV] 26 May 2016 1 Redwood Center for Theoretical Neuroscience, University of California, Berkeley May 27, 2016 Abstract A recent paper by Gatys et al. [1] describes a method for rendering an image in the style of another image. First, they use convolutional neural network features to build a statistical model for the style of an image. Then they create a new image with the content of one image but the style statistics of another image. Here, we extend this method to render a movie in a given artistic style. The naive solution that independently renders each frame produces poor results because the features of the style move substantially from one frame to the next. The other naive method that initializes the optimization for the next frame using the rendered version of the previous frame also produces poor results because the features of the texture stay fixed relative to the frame of the movie instead of moving with objects in the scene. The main contribution of this paper is to use optical flow to initialize the style transfer optimization so that the texture features move with the objects in the video. Finally, we suggest a method to incorporate optical flow explicitly into the cost function. 1 Introduction and Related Research Automatic image stylization is a popular feature in image editors like Adobe Photoshop and social networking services like Instagram. However, many of these styles are limited in their complexity. Recent work has shown that we can use statistical correlations of neural network features to describe complex styles. Furthermore, these models can be then used to create new images with the same statistics. But a question remains: how do we apply these complex styles to video? This task is additionally challenging because the textures need to move smoothly with the objects in the scene to obtain an pleasing video. In this section, we review past work that motivated our approach of using optical flow in combination with neural network features to stylize movies. In a seminal paper on texture generation, Heeger and Bergen built a statistical model of textures using an image pyramid and then developed a method to render new images with those same statistics [2]. While this approach produces good results, the method is limited due to the use of hand-designed features. For alternative methods of texture synthesis, we direct the reader to a pair of papers on texturizing images and a more recent review [3, 4, 5]. Recent research on artificial neural networks has shown that the features produced by task-driven learning are better than hand-designed features. In particular, convolutional neural networks trained on image classification (e.g. AlexNet trained on IMAGENET [6], [7]) extract image features that are useful for many tasks such as fine grained image recognition, scene recognition, and image retrieval [8]. In order to understand the information captured by neural network features, a number of papers have used backpropagation to find an image that has a given set of neural network features [9, 10]. These advances in building useful image features and learning how to manipulate them has led to an explosion of research into generating artistic images using deep neural networks. Two recent notable examples 1 are Google’s post on inceptionism [11] and the pair of papers from Gatys et al. [12, 1]. Since then, there have been a number of papers that have sought to explore various aspects of the idea and to speed up the rendering (e.g. using an adversarial network approach) [13, 14, 15, 16, 17, 18, 19]. In order to apply these methods to movies, the textures that these methods overlay onto an image must move with the objects in the scene. The problem of detecting the motion of objects in a scene is well-studied in the field of computer vision and is often referred to as computing the optical flow. Thus we can use the optical flow fields in order to move the textures with the objects as they move in the movie. This paper is organized as follows. First, we review previous research, including the formulation of the Gatys et al. style transfer paper. Next, we extend this method using optical flow to encourage temporal continuity. Then we present our results on rendering a naturalistic movie and assess the success of our method as well as techniques for improving it. Finally, we end with some discussion and conclusions. 2 Theoretical Formulation First, we review the approach taken by Gatys et al. [1]. While past work used hand-designed features to build statistical models of textures, Gatys et al. had the idea to use neural network features instead. To be more precise, suppose that we have a convolutional neural network such as VGG-19 [20]. Denote the input image to the network as x (with the individual pixels of the image denoted as xu,v ) and the activations of l the lth convolutional layer for feature map j and pixel i as Fi,j (x) (suppose that there are J l feature maps l and I pixels in that layer). Then define the Gram matrix Glj,j 0 (x) = 1 X l l F (x)Fi,j 0 (x). I l i i,j (1) We can think of this matrix as computing the second order statistics between different neural network features because we are averaging over the spatial dimensions of the feature maps of a convolutional network. Part of the intuition of this approach is that complex, higher-order, statistics in the image space can be captured using second order correlations in the neural network feature space. The style transfer objective function takes in two images xs and xc , the style and content images, respectively and renders an image, xr : xr = argminx E(x) c (2) s E(x) = Lcontent (x, x ) + Lstyle (x, x ) + Ltv (x) X 1 X l l Lcontent (x, xc ) = λlc l l (F (x) − Fi,j (xc ))2 I J i,j i,j (3) (4) l 1 X l (G 0 (x) − Glj,j 0 (xs ))2 (J l )2 0 j,j j,j l X Ltv (x) = λtv (xu+1,v − xu,v )2 + (xu,v+1 − xu,v )2 Lstyle (x, xs ) = X λls (5) (6) u,v where λlc,s,tv are weights that determine the relative strength of the style, content, and total variation costs and xu,v denotes a pixel of the image. Now we describe several methods for extending this work to movies. Suppose that we have a movie, whose frames are xt for t = 0, 1, . . .. Since the Gram matrix cost is spatially invariant, there are many ways to overlay a texture onto the scene. In the language of optimization, there are many local minima and the initialization has a strong impact on the result of a gradient descent minimization procedure. We exploit this property to obtain temporal continuity in our movie. Let y = ST (x0 , xc , xs ) denote the output of a gradient descent based minimization of the style transfer objective function that uses x0 as the initialization point for the optimization, xc as the content image, and xs as the style image. Alternative methods for applying art style to movies that we implemented (the first two methods are baselines and the third method is the optical flow based method that works better than the baseline methods): 2 • Independent Initialization: yt = ST (xt , xt , xs ). • Rendered Previous Frame Initialization: yt = ST (yt−1 , xt , xs ). • Optical Flow Initialization: yt = ST (yt0 , xt , xs ), yt0 = F LOW (yt−1 , Ft−1,t ) where Ft−1,t denotes the optical flow field taking xt−1 to xt and F LOW applies an optical flow field to an image. This approach is summarized in Fig. 1. Additional methods for applying art style to movies that may improve results, but that we have not implemented yet: • Optical Flow Initialization and Loss: In addition to doing the above, we can add a term l(x, yt0 ) that penalizes the distance between the image that we want to render and the optical flowed version of the previous frame of the movie. For l, we can either use a robust loss or a squared error loss following the optical flow literature (eg. [21]). • Optical Flow with Back-Tracking: Instead of sequentially optimizing each frame, one can imagine creating an objective function that jointly optimizes over all frames. For instance, we can add the following term: X Lof,t (xt ) = l(xt , F LOW (yt0 , Ft0 ,t )) (7) t0 ∈N (t) where N (t) = {t − 1, t + 1}. The full objective function can be descended by pre-computing the optical flow fields and then descending with respect to each image sequentially and then doing additional passes to clean up the images. 3 Results In order to show the generality of our technique, we test our approach on a short clip from Star Wars: Episode V - The Empire Strikes Back. This clip has a number of properties that are interesting and challenging for our algorithm: 1. There are parts of the video where there is both large motion of the camera and large motion of the objects in the scene. 2. Surfaces are revealed and hidden as time flows forward. 3. There are special effects such as rapid flashes of light that create large variations in the overall brightness of the scene, temporarily occlude objects, and break the brightness constancy assumption. A video summary of our results can be found at https://youtu.be/nOrXeFBkP04. In Fig. 2, we show that our algorithm handles large camera and image motions well in comparison to two baselines: independent rendering of each frame, and initialization using the previous rendered frame. In the case of independent initialization, strokes in adjacent frames show little or no apparent relationship. The previous frame initialization method generates a video where characters appear to melt into the background as the camera pans. In the case of the flowed previous frame initialization, the strokes appear to move smoothly with the characters. In Fig. 3, we show that occluded surfaces are rendered properly. Also we see that special effects (here, the screen lighting up due to a laser blast) are lost to some extent, but the results still look reasonable. 4 Discussion and Conclusions Compared to many machine learning tasks that have clear benchmarks, computer generated art is a subjective task. On one hand, a quantitative (and arguably objective) measure of a stylized video is the sum of the squares of the difference between the flowed version of one frame and the next frame. Relatedly, we hypothesize that we could obtain more temporally consistent results by adding an explicit term in the cost function to regularize the style transfer. On the other hand, some variability from frame to frame contributes to the artistic effect. For instance, a video directed by Boris Seewald makes good use of a combination of temporal continuity in content but discontinuity in style http://www.thisiscolossal.com/2016/03/ 3 Style Transfer Algorithm In: Initialization, Content Image, Style Image Out: Styled Image Optical Flow Algorithm In: Frame k, Frame k+1 Out: Flow Field Image Flow Algorithm In: Flow Field, Image Out: Flowed Image Style Transfer Algorithm In: Initialization, Content Image, Style Image Out: Styled Image Time Figure 1: Overview of Our Approach: We begin by applying the style transfer algorithm to the first frame of the movie using the content image as the initialization. Next, we calculate the optical flow field that takes the first frame of the movie to the second frame. We apply this flow-field to the rendered version of the first frame and use that as the initialization for the style transfer optimization for the next frame. Note, for instance, that a blue pixel in the flow field image means that the underlying object in the video at that pixel moved to the left from frame one to frame two. Intuitively, in order to apply the flow field to the styled image, you move the parts of the image that have a blue pixel in the flow field to the left. 4 Figure 2: Initialization using optical flow improves results relative to baseline methods during camera panning: Each row shows several adjacent frames of the movie rendered according to a different method. Row One: Frames from the original movie at 10 Hz. Row Two: Independent Initialization - strokes in adjacent frames show little or no apparent relationship. Row Three: Previous Rendered Frame Initialization - brush strokes stay fixed relative to the frame of the movie and the characters appearing to melt into the background as the camera pans. Row Four: Flowed Previous Rendered Frame Initialization - strokes appear to move smoothly with the characters. 5 Figure 3: Optical Flow Method During Occlusions and Special Effects: Top: The content cost appropriately stylizes a surface that is gradually revealed due to a moving occluder. Bottom: The resulting rendered video does not capture the special effect and gets warped as a result. rotoscoped-music-video-boris-seewald-disco/. Continued work in this field will benefit from input from artists as to what are desirable properties of stylized movies. Along the lines of developing artistically pleasing results, one of the major limitations of this approach is that all of the different style attributes are wrapped up in a model of style that has many parameters, making it difficult to manipulate. For instance, simple properties such as the overall light-level of an image are entangled in the thousands of parameters in the gram-matrix texture model. We experimented with one approach of addressing this issue. In the supplementary figures, we use histogram matching to rescale the style image and then compute a new gram matrix (see Fig. 5). From a practical (eg. for artists) and theoretical perspective, it would be beneficial to figure out how to disentangle simple image properties from the more complex properties (such as brush strokes) that the neural style features capture. Another limitation is that in order to make these algorithms commercially useful, the rendering algorithm must be optimized, possibly adapting the approaches suggested in [15] or [13]. The current approach requires a relatively costly gradient descent procedure to render each frame. In conclusion, our work demonstrates that we can improve the temporal coherence of an artistically styled movie by making use of optical flow. While statistics based methods of rendering textures are promising, much work needs to be done in order to tease apart simple image properties to make these rendering algorithms useful for artists. 5 Statement on Contribution This project benefited from the collaboration of a number of people. AGA did the literature review, came up with the idea to stylize movies, and implemented the majority of the code (based on rewriting https: //github.com/anishathalye/neural-style). BAO originally suggested using optical flow to make the rendered image more temporally coherent. CPB helped in the debugging of the main algorithm, and ran the controls. DPM used histogram matching to better handle the special effects and contributed to the write-up. We thank Brian Cheung and Dylan Paiton from the Redwood Center for Theoretical Neuroscience 6 and Richard Zhang and Alyosha Efros from the Berkeley Computer Vision Group for useful feedback. References [1] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. A neural algorithm of artistic style. arXiv preprint arXiv:1508.06576, 2015. [2] David J Heeger and James R Bergen. Pyramid-based texture analysis/synthesis. In Proceedings of the 22nd annual conference on Computer graphics and interactive techniques, pages 229–238. ACM, 1995. [3] Alexei A Efros and Thomas K Leung. Texture synthesis by non-parametric sampling. In Computer Vision, 1999. The Proceedings of the Seventh IEEE International Conference on, volume 2, pages 1033– 1038. IEEE, 1999. [4] Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 341–346. ACM, 2001. [5] Li-Yi Wei, Sylvain Lefebvre, Vivek Kwatra, and Greg Turk. State of the art in example-based texture synthesis. In Eurographics 2009, State of the Art Report, EG-STAR, pages 93–117. Eurographics Association, 2009. [6] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105, 2012. [7] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211–252, 2015. [8] Ali Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. Cnn features off-the-shelf: an astounding baseline for recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 806–813, 2014. [9] Aravindh Mahendran and Andrea Vedaldi. Understanding deep image representations by inverting them. In Computer Vision and Pattern Recognition (CVPR), 2015 IEEE Conference on, pages 5188– 5196. IEEE, 2015. [10] M Cimpoi, S Maji, and A Vedaldi. Deep filter banks for texture recognition and segmentation. Computer Vision and Pattern Recognition (CVPR), 2015 IEEE Conference on, pages 3828–3836, 2015. [11] Alexander Mordvintsev, Michael Tyka, and Christopher Olah. Inceptionism: Going deeper into neural networks, google research blog. Retreived June, 17, 2015. [12] Leon A. Gatys, Alexander S. Ecker, and Matthias Bethge. Texture Synthesis Using Convolutional Neural Networks. Nips, pages 1–10, 2015. [13] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual Losses for Real-Time Style Transfer and Super-Resolution. 2016. [14] Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky. Texture networks: Feedforward synthesis of textures and stylized images. arXiv preprint arXiv:1603.03417, 2016. [15] Chuan Li and Michael Wand. Combining markov random fields and convolutional neural networks for image synthesis. arXiv preprint arXiv:1601.04589, 2016. [16] Tsung-Yu Lin and Subhransu Maji. Visualizing and understanding deep texture representations. arXiv preprint arXiv:1511.05197, 2015. 7 [17] Yang Lu, Song-Chun Zhu, and Ying Nian Wu. Learning frame models using cnn filters for knowledge visualization. arXiv preprint arXiv:1509.08379, 2015. [18] Rujie Yin. Content aware neural style transfer. arXiv preprint arXiv:1601.04568, 2016. [19] Emily L Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a laplacian pyramid of adversarial networks. In Advances in Neural Information Processing Systems, pages 1486– 1494, 2015. [20] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. [21] Deqing Sun, Stefan Roth, and Michael J Black. Secrets of optical flow estimation and their principles. In Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference on, pages 2432–2439. IEEE, 2010. [22] Martı́n Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mané, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viégas, Oriol Vinyals, Pete Warden, Martin Wattenberg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. Software available from tensorflow.org. [23] Diederik Kingma and Jimmy Ba. arXiv:1412.6980, 2014. Adam: A method for stochastic optimization. arXiv preprint [24] Dong C Liu and Jorge Nocedal. On the limited memory bfgs method for large scale optimization. Mathematical programming, 45(1-3):503–528, 1989. [25] Thomas Brox and Jitendra Malik. Large displacement optical flow: descriptor matching in variational motion estimation. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 33(3):500–513, 2011. 6 Supplementary Methods 1. The core style transfer algorithm was implemented using TensorFlow [22]. We used a pretrained version of VGG-19 [20] available from http://www.vlfeat.org/matconvnet/models/imagenet-vgg-verydeep-19. mat. 2. Experimentally we found that quasi-second order methods perform better than vanilla gradient descent. In particular, we use Adam with about 1000 iterations to optimize the objective function for each frame of the movie [23]. Others papers use L-BFGS [24] (eg. [1]). 3. For all experiments, we use conv4 2 for the content layer and conv1 1, conv2 1, conv3 1, conv4 1, conv5 1 for the style layers along the lines of [1]. However, we use different weight values. Roughly speaking, we set the weights so that each term in the cost function contributed at least five percent to the total cost. During optimization, we found that the content cost would still dominate. 4. In order to generate optical flow fields, we use a pre-compiled binary from T. Brox and J. Malik [25]. We note that these dense flow fields are much better than those produced by the built-in optical flow methods in OpenCV. Our results could be improved even more using a better optical flow algorithm. 5. We use the content image to initialize the optimization in the case of the first frame of the video or in the presence of manually detected scene cuts. 6. The movies are rendered at 10 frames per second. 8 7 7.1 Supplementary Figures Histogram matching improves color coherence across scenes In contrast to a static content image, a movie scene might have colors and textures which change in time. In extending style transfer from images to movies, one might wish to stylistically match the reference image with each frame, or with the movie scene as a whole. For example if the scene is brightly illuminated in one frame and dark in another, one might wish to match the two frames to light and dark colors in the image’s palette, respectively, rather than capturing the entire color palette in both. Further, the same object or character might appear as relatively dark or light compared with the background. However, for ease of recognition, one might wish to consistently render them with similar colors. Matching histograms between movie and image could allow one to capture a large part of the perceptually important color statistics. As an example, we first transfer the RGB histogram of the style image to that of the entire content video. Thus, we allow the colors to vary from frame to frame, while still matching the movie as a whole to the full range of colors. Subsequently, we transfer the histogram of each histogramtransferred video frame back to the original style image in order to allow the colors of the target style to vary in time. This procedure is outlined in Fig. 5a. This allows us, for example, to preserve the brightness of special effect blaster fire as in Fig. 5b, and the rendering of characters’ skin and hair color as in Fig. 5c. The improvement in color coherence comes at the price of a less diverse array of colors and reduction in image quality due to two successive histogram transfers. For related reasons, it also seems to disrupt continuity between frames due to artifactual changes in brightness. This could be improved by using source images with greater color depth. Additionally, similar effects could possibly be obtained without explicit histogram matching using a joint optimization approach as suggested in section 2. 9 Figure 4: Movie Frames and Optical Flow: Top: frames of the movie. Middle: Optical flow field of the movie. Bottom: Applying optical flow to the frames of the movie. 10 (a) (b) (c) Figure 5: a. In order to allow individual rendered frames to vary in their color palette as they do in the original, we apply histogram matching. First, we transfer the histogram of “Starry Night” to the entire movie viewed as an image (i.e., the histogram of the entire movie is matched to that of the painting). Then, we transfer the histogram of each frame of the transformed movie, to a frame of the new “style movie.” Thus, each frame of the style movie captures a subset of the painting’s original histogram. b. This allows us to capture brightness changes, as in special effects. From left to right, the original movie frame; the frame rendered using the optical flow approach; and the frame rendered using the optical flow approach with histogram matching. c. Similarly, we can preserve the color of objects and characters reappearing over the course of the movie. From left to right, two original frames depicting the same character; the frames rendered using the optical flow approach; and the frames rendered using optical flow and histogram matching. 11
9cs.NE
Asymptotic properties of a componentwise ARH(1) plug-in predictor Javier Álvarez-Liébana1 , Denis Bosq2 and M. Dolores Ruiz-Medina1 arXiv:1706.06498v1 [math.ST] 20 Jun 2017 2nd April 2018 1 Department of Statistics and O. R., University of Granada, Spain. LSTA, Université Pierre et Marie Curie-Paris 6, Paris, France. E-mail: [email protected] 2 Summary This paper presents new results on prediction of linear processes in function spaces. The autoregressive Hilbertian process framework of order one (ARH(1) process framework) is adopted. A componentwise estimator of the autocorrelation operator is formulated, from the moment-based estimation of its diagonal coefficients, with respect to the orthogonal eigenvectors of the auto-covariance operator, which are assumed to be known. Mean-square convergence to the theoretical autocorrelation operator, in the space of Hilbert-Schmidt operators, is proved. Consistency then follows in that space. For the associated ARH(1) plug-in predictor, mean absolute convergence to the corresponding conditional expectation, in the considered Hilbert space, is obtained. Hence, consistency in that space also holds. A simulation study is undertaken to illustrate the finite-large sample behavior of the formulated componentwise estimator and predictor. The performance of the presented approach is compared with alternative approaches in the previous and current ARH(1) framework literature, including the case of unknown eigenvectors. Key words: ARH(1) processes; consistency; functional prediction; mean absolute and quadratic convergence 1 Introduction. In the last few decades, an extensive literature on statistical inference from functional random variables has emerged, motivated, in part, by the statistical analysis of high-dimensional data, as well as data with a continuous (infinite-dimensional) nature (see [11, 12, 19, 25, 44, 45, 52, 53], among others). A selected set of new developments and branches of the functional statistical theory can be found in [10]. Particularly, in the current literature, we refer to the recent books [34] and [35], surveys [15], and Special Issue [29], and the references therein. These references include a nice summary on the statistics theory for functional data, contemplating covariance operator theory and eigenfunction expansion, perturbation theory, smoothing and regularization, probability measures on a Hilbert spaces, functional 1 principal component analysis, functional counterparts of the multivariate canonical correlation analysis, the two sample problem and the change point problem, functional linear models, functional test for independence, functional time series theory, spatially distributed curves, software packages and numerical implementation of the statistical procedures discussed, among other topics. In the functional regression model context, we particularly refer to the case where the predictor is a random function, and the response is scalar, since this topic has been widely developed. Different specifications of the regression operator arise, in several applied fields, such as chemometrics, biology, economics and climatology. An active research area in the literature is devoted to estimate the regression operator. To avoid the computational (high-dimensionality) limitations of the nonparametric approach, several parametric and semi-parametric methods have been proposed (see, for example, [23], and the refereces therein). In particular, in [23], to avoid the problem of high dimensionality, a combination of a spline approximation and the one-dimensional Nadaraya-Watson approach is proposed. The generalizations in the case of more regressors (all functional, or both functional and real) have been addressed in the nonparametric, semi-parametric, and parametric frameworks (see for an overview [1], [22], and [26]). For functional response and covariate, in the nonparametric regression framework, we refer to [24], where a functional version of the Nadaraya-Watson estimator is proposed for the estimation of the regression operator. Its pointwise asymptotic normality is derived as well. Resampling ideas are applied to overcome the difficulties arising in the estimation of the asymptotic bias and variance. Semifunctional partial linear regression, introduced in [2], allows the prediction of a real-valued random variable from a set of real explanatory variables, and a time-dependent functional explanatory variable. The asymptotic, and practical properties in real-data applications, of the estimators derived are studied as well. Motivated by genetic and environmental applications, in [14], a semi-parametric maximum likelihood method for the estimation of odds ratio association parameters is also developed in a high dimensional data framework. In the autoregressive Hilbertian time series framework, several estimation and prediction procedures have been proposed with the derivation of the corresponding asymptotic theory. In [39], the weakconvergence to the normal distribution of the estimator of the autocorrelation operator studied, based on projection into the theoretical eigenvectors, is established, under suitable conditions. In [11] and [13], the problem of prediction of linear processes in function spaces is addressed. In particular, sufficient conditions for the consistency of the empirical auto- and cross- covariance operators are obtained. The asymptotic normal distribution of the empirical auto-covariance operator is also derived. Projection into the empirical eigenvectors is considered, in the formulation of a componentwise estimator of the autocorrelation operator, in absence of information about the theoretical ones. Asymptotic properties of the empirical eigenvalues and eigenvectors are also analyzed. In [31], the efficiency of a componentwise estimator of the autocorrelation operator, based on projection into the empirical eigenvector system of 2 the auto-covariance operator, is obtained. In [40], consistency, in the space of bounded linear operators, of the formulated estimator of the autocorrelation operator, and of its associated ARH(1) plug-in predictor is proved. Sufficient conditions for the weak-convergence of the ARH(1) plug-in predictor to a Hilbert-valued Gaussian random variable are derived in [41]. High deflection results or large and moderate deviations for infinite-dimensional autoregressive processes are derived in [42]. The law of the iterated logarithm for the covariance operator estimator is obtained in [43]. The main properties for the class of autoregressive Hilbertian processes with random coefficients are investigated in [47]. Interesting extensions of the autoregressive Hilbertian framework can be found in [36], who offers a new estimate, based on the spectral decomposition of the autocorrelation operator, and not of the auto-covariance operator. The first generalization on autoregressive processes of order greater than one is proposed in [46], in order to improve prediction. ARHX (1) models, i.e., autoregressive Hilbertian processes with exogenous variables are formulated in [17]. In [30] and [31] a doubly stochastic formulation of the autoregressive Hilbertian process is studied. The ARHD model is introduced in [38], taking into account the regularity of trajectories through the derivatives. Recently, [16] introduced the conditional autoregressive Hilbertian processes (CARH processes), as a new class of processes, and developed parallel projection estimation methods to predict such processes. In the Banach-valued context, we refer to the papers by [6, 20, 50, 51], among others. In this paper, we assume that the autocorrelation operator belongs to the Hilbert-Schmidt class, and admits a diagonal spectral decomposition in terms of the orthogonal eigenvector system of the auto-covariance operator. That is the case, for example, of an autocorrelation operator defined as a continuous function of the autocovariance operator (see Remark 1 below). A componentwise estimator of the autocorrelation operator is formulated in terms of such an eigenvector system. In the derivation of the results presented in this paper we assume that the eigenvectors (φj , j ≥ 1) of the autocovariance operator C are known. That is the case, for example, of defining our random initial condition as the solution, in the mean-square sense, of a stochastic differential equation driven by white noise (e.g., the Wiener measure). However, beyond this case, the sparse representation and whitening properties of wavelet bases can be exploited to obtain a diagonal representation of the autocovariance and crosscovariance operators (see Remark 3 below), in terms of a common and known wavelet basis. Smoothing functional data with a suitable penalization norm in terms of wavelets also could lead to an empirical regularized diagonal approximation of the autocorrelation operator in terms of wavelets. Finally, we refer to shrinkage estimation of the autocovariance and cross covariance operators in terms of a suitable common wavelet basis, allowing a diagonal empirical approximation of both operators, in terms of such a basis (see also Section 6). Under the setting of assumptions established in this paper (see Section 2), convergence to the auto- 3 correlation operator, in L2 -sense in the space of Hilbert-Schmidt operators S (H) , i.e., convergence in the space L2S(H) (Ω, A, P) , is derived, for the formulated componentwise estimator. Consistency then follows in S (H) (see Section 3.1). Under the same setting of conditions, consistency in H of the associated ARH(1) plug-in predictor is obtained as well, from its convergence in the L1 -sense in the Hilbert space H, i.e., in the space L1H (Ω, A, P) (see Section 3.2). The Gaussian framework is analyzed in Section 4. In that framework, the numerical examples studied in Section 5 illustrate the behavior of the formulated componentwise autocorrelation operator estimator and associated predictor for large sample sizes. A comparative study with alternative ARH(1) prediction techniques (even in a non-diagonal scenario), based on componentwise parameter estimators of the autocorrelation operator (including the case of empirical eigenvectors), as well as based on kernel (nonparametric) functional estimators, and penalized, spline and wavelet, estimators, is also performed. Final conclusions and some guidelines for the application of the proposed approach from real-data are provided in Section 6. 2 Preliminaries. The preliminary definitions and lemmas, that will be applied in the derivation of the main results of this paper, in the context of ARH(1) processes, are now provided. In the following, let us denote by H a real separable Hilbert space. Recall that a zero-mean ARH(1) process X = (Xn , n ∈ Z) satisfies the equation (see [11], among others) Xn = ρ (Xn−1 ) + εn , n ∈ Z, (1) where ρ denotes the autocorrelation operator of process X, which belongs to the space L(H) of bounded linear operators, such that kρk kL(H) < 1, for k ≥ k0 , and for certain k0 ≥ 1, with k · kL(H) denoting the norm in the space L(H). The Hilbert-valued innovation process ε = (εn , n ∈ Z) is considered to be strong-white noise, i.e., ε is a Hilbert-valued zero-mean stationary process, with independent and identically distributed components in time, and with σε2 = Ekεn k2H < ∞, for all n ∈ Z. We restrict our attention here to the case where ρ is such that kρkL(H) < 1. The following assumptions will be made: Assumption A1. The autocovariance operator C = E[Xn ⊗ Xn ] = E[X0 ⊗ X0 ], for every n ∈ Z, is a positive self-adjoint and trace operator. It then admits the following 4 diagonal spectral representation, in terms of its eigenvectors (φj , j ≥ 1) C= ∞ X j=1 where ∞ X j=1 creasing Cj φj ⊗ φj , (2) Cj < ∞, with (Cj , j ≥ 1) denoting the system of real positive eigenvalues of C, arranged in deorder of their magnitudes C1 ≥ C2 ≥ · · · ≥ Cj ≥ Cj+1 ≥ · · · > 0. Assumption A2. The autocorrelation operator ρ is a self-adjoint and Hilbert-Schmidt operator, admitting the following diagonal spectral decomposition: ρ= ∞ X j=1 ρj φj ⊗ φj , ∞ X j=1 ρ2j < ∞, (3) where (ρj , j ≥ 1) is the system of eigenvalues of the autocorrelation operator ρ, with respect to the orthonormal system of eigenvectors (φj , j ≥ 1) of the autocovariance operator C. Note that, under Assumption A2, kρkL(H) = sup ρj < 1. j≥1 Remark 1 Assumption A2 holds, in particular, when operator ρ is defined as a continuous function of operator C (see [18], pp. 119-140). See also Remark 3 below. In the following, for any n ∈ Z, let D = E[Xn ⊗ Xn+1 ] = E[X0 ⊗ X1 ] be the cross-covariance operator of the ARH(1) process X = (Xn , n ∈ Z). Under Assumptions A1-A2, by projection of equation (1) into the orthonormal system (φj , j ≥ 1), we obtain, for each j ≥ 1, the following AR(1) equation: Xn,j = ρj Xn−1,j + εn,j , n ∈ Z, (4) where Xn,j = hXn , φj iH and εn,j = hεn , φj iH , for all n ∈ Z. Under Assumptions A1-A2, from equation (4), for each j ≥ 1, ρj = ρ(φj )(φj ) = φj , DC −1 (φj ) = hD(φj ), φj iH C = −1 (φj ), φj D E [Xn,j Xn−1,j ]  2  = j, Cj E Xn−1,j 5 H H ∀n ∈ Z, (5)  −1  2 ] and Xn,j = hXn , φj iH , j ≥ 1, since where Dj = hD(φj ), φj iH = E [Xn,j Xn−1,j ] , Cj−1 = [E Xn−1,j D= ∞ X j=1 Dj φj ⊗ φj , Dj = ρj Cj , j ≥ 1. (6) Let us now consider the Banach space L2H (Ω, A, P) of the classes of equivalence established in L2H (Ω, A, P) , the space of zero-mean second-order Hilbert-valued random variables (H-valued random variables) with finite seminorm given by kZkL2 H (Ω,A,P) = q E kZk2H , ∀Z ∈ L2H (Ω, A, P) . (7) That is, for Z, Y ∈ L2H (Ω, A, P) , Z and Y belong to the same equivalence class if and only if E kZ − Y kH = 0. In the next section, we will consider, in particular, H = S(H), the Hilbert space of Hilbert-Schmidt operators on a Hilbert space H, i.e., we will consider, the space L2S(H) (Ω, A, P) with the seminorm kY k2L2 S(H) (Ω,A,P) = E kY k2S(H) , ∀Y ∈ L2S(H) (Ω, A, P) . (8) For each n ∈ Z, let us consider the following biorthogonal representation of the functional value Xn of the ARH(1) process X = (Xn , n ∈ Z) , as well as of the functional value εn of its innovation process ε = (εn , n ∈ Z) : Xn = εn = ∞ ∞ X X p hXn , φj iH p Cj p Cj ηj (n)φj , φj = Cj j=1 j=1 ∞ X j=1 where ηj (n) = hXn ,φj iH √ Cj ∞ σj X hεn , φj iH σj ηej (n)φj , φj = σj j=1 X = √n,j and ηej (n) = Cj (9) hεn ,φj iH σj = εn,j σj , (10) for every n ∈ Z, and for each j ≥ 1. Here, under Assumptions A1-A2, for Rε = E[εn ⊗ εn ] = E[ε0 ⊗ ε0 ], n ∈ Z, Rε φj = σj2 φj , j ≥ 1, where, as before, (φj , j ≥ 1) denotes the system of eigenvectors of the autocovariance operator C, and X σj2 = σε2 = Ekεn k2H , for all n ∈ Z. j≥1 The following lemma provides the convergence, in the seminorm of L2H (Ω, A, P), of the series expan6 sions (9)-(10). Lemma 1 Let X = (Xn , n ∈ Z) be a zero-mean ARH(1) process. Under Assumptions A1-A2, for any n ∈ Z, the following limit holds 2 bn,M lim E Xn − X M→∞ bn,M = where X H = 0, (11) M X p Cj ηj (n)φj . Furthermore, j=1 E lim M→∞ h i   bn,M bn,M ⊗ Xn − X Xn − X 2 S(H) = 0. (12) Similar assertions hold for the biorthogonal series representation εn = ∞ X ∞ σj j=1 X hεn , φj iH σj ηej (n)φj . φj = σj j=1 Proof.  bn,M Under Assumption A1, from the trace property of C, the sequence X satisfies, for M sufficiently large, and L > 0, arbitrary,  M X p = Cj ηj (n)φj , M ≥ 1 j=1 2 bn,M+L − X bn,M k2 2 b b kX L (Ω,A,P ) = EkXn,M+L − Xn,M kH H = M+L X M+L X j=M+1 k=M+1 = M+L X j=M+1 since, under Assumption A1, ∞ X j=1 M+L X  bn,M X  Cj → 0, when M → ∞,  Cj < ∞, hence,  Cj converges to zero when M j=M+1 p p Cj Ck E[ηj (n)ηk (n)] hφj , φk iH M X j=1 (13)  Cj , M ≥ 1 is a Cauchy sequence. Thus, → ∞, for L > 0, arbitrary. From equation (13),  M X p Cj ηj (n)φj , M ≥ 1 is also a Cauchy sequence in L2H (Ω, A, P ). Thus, the sequence = j=1  bn,M , M ≥ 1 has finite limit in L2 (Ω, A, P ), for all n ∈ Z. X H 7 Furthermore, 2 bn,M lim E Xn − X M→∞ + lim M→∞ −2 lim H 2 = E kXn kH M X M X p p Cj Ch E [ηj (n)ηh (n)] hφj , φh iH j=1 h=1 M→∞ M M X X p 2 Cj = 0. Cj E [hXn , ηj (n)φj iH ] = σX − lim M→∞ j=1 j=1 (14) In the derivation of the identities in (13) and (14), we have applied that, for every j, h ≥ 1, Cφj = E [ηj (n)ηh (n)] = 2 σX = EkXn k2H = Cj φj , δj,h , ∞ X Cj < +∞, j=1  E hXn , hφj , φh iH = δj,h ,  p ηj (n)φj iH = Cj . (15) Moreover, from identities in (15), E = h i   bn,M bn,M ⊗ Xn − lim X Xn − lim X E[Xn ⊗ Xn ] + lim M→∞ −2 lim M→∞ = = M→∞ M→∞ M X j=1 M X M X j=1 h=1 2 S(H) p p Cj Ch φj ⊗ φh E[ηj (n)ηh (n)] p E[Xn ⊗ Cj ηj (n)φj ] 2 S(H)   M M X X E[Xn ⊗ Xn ] + lim  Cj φj ⊗ φj − 2 Cj φj ⊗ φj  M→∞ E[Xn ⊗ Xn ] − lim M→∞ j=1 M X j=1 j=1 2 S(H) 2 = 0. Cj φj ⊗ φj (16) S(H) In a similar way, we can derive convergence to εn , in L2H (Ω, A, P), of the series ∞ X j=1 σj ηej (n)φj , for every n ∈ Z, since ε is assumed to be strong-white noise, and hence, its covariance operator Rε is in the trace class. We can also obtain an analogous to equation (16).  8 In equations (9)-(10), for every n ∈ Z, E[ηj (n)] = 0, E[ηj (n)ηh (n)] = δj,h , j, h ≥ 1, n∈Z (17) E[e ηj (n)] = 0, E[e ηj (n)e ηh (n)] = δj,h , j, h ≥ 1, n ∈ Z. (18) Note that, from Assumption A2 for each j ≥ 1, (Xn,j , n ∈ Z) in equation (4) defines a stationary and invertible AR(1) process. In addition, from equations (9) and (15), for every n ∈ Z, and j, p ≥ 1, Xn = E [Xn,j Xn,p ] = 2 E kXn kH which implies that Cj = = σj2 , 1−ρ2j ∞ X Xn,j φj , j=1 ∞ X ∞ ∞ X X 2 ρkj ρhp E [εn−k,j εn−h,p ] = δj,p ρ2k j σj k=0 h=0 k=0 ∞ ∞ ∞ X X  2  X hC (φj ) , φj iH = Cj = E Xn,j = j=1 j=1 j=1 = δj,p σj2 , 1 − ρ2j 2 σX < ∞, (19) for each j ≥ 1. In particular, we obtain, for each j ≥ 1, and for every n ∈ Z, E [ηj (n)ηj (n + 1)] = " Xn,j Xn+1,j p E p Cj Cj = ∞ X ∞ X = ∞ X # = E [Xn,j Xn+1,j ] Cj ρk+h E [εn−k,j εn+1−h,j ] j k=0 h=0 Cj ρ2k+1 σj2 j k=0 = Cj σj2 ρj = ρj , Cj 1 − ρ2j Remark 2 From equation (4) and Lemma 1, keeping in mind that Cj = n ∈ Z. σj2 , 1−ρ2j (20) for each j ≥ 1, the following invertible and stationary AR(1) process can be defined as ηj (n) = ρj ηj (n − 1) + q 1 − ρ2j ηej (n), 0 < ρ2j ≤ ρj < 1, (21) where, for each j ≥ 1, (ηj (n), n ∈ Z) and (e ηj (n), n ∈ Z) are respectively introduced in equations (9)i h (10). In the following, for each j ≥ 1, we assume that E (e ηj (n))4 < ∞, for every n ∈ Z, to ensure ergodicity for all second-order moments, in the mean-square sense (see, for example, [33], pp. 192-193). 9 Furthermore, D = E[Xn ⊗ Xn+1 ] = = = ∞ X ∞ X j=1 p=1 ∞ X ∞ X   E hXn , φj iH hXn+1 , φp iH φj ⊗ φp   p p E hXn , φj iH hXn+1 , φp iH p p Cj Cp φj ⊗ φp Cj Cp j=1 p=1 ∞ X ∞ X p p Cj Cp E [ηj (n)ηp (n + 1)] φj ⊗ φp . j=1 p=1 (22) Remark 3 In particular, Assumption A2 holds if the following orthogonality condition is satisfied E [ηj (n)ηp (n + 1)] = δj,p , j, p ≥ 1, n ∈ Z, (23) where δj,p denotes the Kronecker Delta function. In practice, unconditional bases, like wavelet bases, lead to a sparse representation for functional data (see, for example, [48, 49, 56] for statistically-oriented treatments). Wavelet bases are also designed for sparse representation of kernels defining integral operators, in L2 spaces with respect to a suitable measure (see [37]). The Discrete Wavelet Transform (DWT) approximately decorrelates or whitens data (see [56]). In particular, operators C and D could admit an almost diagonal representation with respect to the self-tensorial tensorial product of a suitable wavelet basis. 3 Estimation and prediction results. A componentwise estimator of the autocorrelation operator, and of the associated ARH(1) plug-in predictor is formulated in this section. Their convergence to the corresponding theoretical functional values, in the spaces L2S(H) (Ω, A, P ) and LH (Ω, A, P ), is respectively derived as well. Their consistency in the spaces S(H) and H then follows. From equation (5), for each j ≥ 1, and for a given sample size n ≥ 1, one can consider the usual b n,j and C bn,j of Dj and Cj , in the AR(1) framework, given by respective moment-based estimators D b n,j D = n−2 1 X Xi,j Xi+1,j , n − 1 i=0 10 n−1 X bn,j = 1 C X2 . n i=0 i,j (24) The following truncated componentwise estimator of ρ is then formulated: ρbkn = kn X j=1 ρbn,j φj ⊗ φj , (25) where, for each j ≥ 1, ρbn,j b n,j D = = bn,j C 1 n−1 n−2 X Xi,j Xi+1,j i=0 n−1 X 2 1 Xi,j n i=0 n = n−1 n−2 X Xi,j Xi+1,j i=0 n−1 X 2 Xi,j i=0 . (26) Here, the truncation parameter kn indicates that we have considered the first kn eigenvectors associated with the first kn eigenvalues, arranged in decreasing order of their modulus magnitude. Furthermore, kn is such that kn < 1, n lim kn = ∞, n→∞ n > 1. (27) The following additional condition will be assumed on kn for the derivation of the subsequent results: Assumption A3. √ lim Ckn n = ∞. The truncation parameter kn in (25) is such that n→∞ Remark 4 Assumption A3 has also been considered in p. 217 of [11], to ensure weak consistency of the proposed estimator of ρ, as well as, in [39] (see Proposition 4, p. 902), in the derivation of asymptotic normality. From Remark 2, for each j ≥ 1, ηj = (ηj (n), n ∈ Z) in equation (21) defines a stationary and invertible AR(1) process, ergodic in the mean-square sense (see, for example, [5]). Therefore, in view of equations (17) and (20), for any j ≥ 1, there exist two positive constants Kj,1 and Kj,2 such that the following identities hold: lim " E 1− 1 n n→∞ lim n→∞ " E ρj − n−1 X #2 ηj2 (i) i=0 1 n 1 n−1 n−2 X (28) #2 ηj (i)ηj (i + 1) i=0 1 n 11 = Kj,1 , = Kj,2 . (29) Equations (28)-(29) imply, for n sufficiently large, Var Var ! n−1 e j,1 K 1X 2 ηj (i) ≤ , n i=0 n (30) ! n−2 e j,2 K 1 X ηj (i)ηj (i + 1) ≤ , n − 1 i=0 n (31) e j,1 and K e j,2 , for each j ≥ 1. Equivalently, for n sufficiently large, for certain positive constants K !2  n−1 X 1 η 2 (i)  ≤ E 1− n i=0 j  !2  n−1 X 1 E  ρj − ηj (i)ηj (i + 1)  ≤ n − 1 i=0  e j,1 K , n (32) e j,2 K , n (33) The following assumption is now considered. Assumption A4.   e j,1 + K e j,2 < ∞. S = sup K j≥1 Remark 5 From equation (26), applying Cauchy–Schwarz’s inequality, we obtain, for each j ≥ 1, |b ρn,j | = ≤ n n−1 n−2 X Xi,j Xi+1,j i=0 n−1 X 2 Xi,j i=0 n n−1 ≤ v un−2 X uX 2 n−2 2 t Xi+1,j Xi,j i=0 i=0 n−1 X 2 Xi,j i=0 v u n−2 uX u 2 Xi+1,j u n u n u i=0 ≤ a.s. u n−1 n − 1u X n−1 2 t Xi,j (34) i=0 3.1 Convergence in L2S(H) (Ω, A, P) In the following proposition, convergence of ρbkn to ρ, in the space L2S(H) (Ω, A, P) , is derived, under the setting of conditions formulated in the previous sections. Proposition 1 Let X = (Xn , n ∈ Z) be a zero-mean standard ARH(1) process. Under Assumptions A1-A4, the following limit holds: 2 lim kρ − ρbkn kL2 n→∞ S(H) 12 (Ω,A,P) = 0. (35) Specifically, kρ − 2 ρbkn kL2 (Ω,A,P) S(H) ≤ g(n), with g(n) = O 1 Ck2n n ! , n → ∞. (36) Remark 6 Corollary 4.3, in [11], p. 107, can be applied to obtain weak convergence results, in terms of weak expectation, using the empirical eigenvectors (see definition of weak expectation at the beginning of Section 1.3, in p. 27, in [11]). Proof. For each j ≥ 1, the following almost surely inequality is satisfied: b n,j Dj D − bn,j Cj C |ρj − ρbn,j | = b n,j b n,j − C bn,j D bn,j − D b n,j Cj + C bn,j D bn,j − D b n,j Cj Dj C Dj C = bn,j bn,j Cj C Cj C = b n,j bn,j − Cj D b n,j Dj − D C + bn,j Cj Cj C   1 b n,j . bn,j + Dj − D |b ρn,j | Cj − C ≤ Cj = (37) Thus, under Assumptions A1–A2, from equation (34), for each j ≥ 1, 2 (ρj − ρbn,j ) ≤ ≤ ≤ 2 1  b n,j bn,j + Dj − D |b ρ | C − C n,j j Cj2   2  2  2 2 bn,j + Dj − D b n,j (b ρ ) C − C n,j j Cj2 ! 2   2 2  2 n b n,j bn,j + Dj − D a.s., Cj − C Cj2 n−1 (38) which implies, for each j ≥ 1, i h 2 E (ρj − ρbn,j )2 ≤ 2 Cj  n n−1 2 !   2  2  b n,j bn,j . + E Dj − D E Cj − C Under Assumption A2, from equations (25) and (39), 13 (39) kρ − ρbkn k2L2 2 S(H) = E kρ − ρbkn kS(H) = (Ω,A,P) kn X ≤ j=1 +E 2 Cj2   n n−1 kn X j=1 2 b n,j Dj − D ∞ i h X   2 E ρ2j E (ρj − ρbn,j ) + j=kn +1  2  b E Cj − Cn,j 2  + ∞ X ρ2j j=kn +1 2   2  2 n b E C − C j n,j Ck2n j=1 n − 1  ∞ 2  X b n,j + ρ2j E Dj − D kn  X ≤ + 2 ≤  n n−1 Ck2n 2 j=kn +1  kn   ∞ 2  2  X X b b ρ2j . + + E Dj − Dn,j E Cj − Cn,j j=1 j=kn +1 (40) Furthermore, from (9) and (26), for j ≥ 1, bn,j C n−1 n−1 1X 1X 2 Xi,j = Cj ηj2 (i), n i=0 n i=0 = b n,j D (41) n−2 n−2 1 X 1 X Xi,j Xi+1,j = Cj ηj (i)ηj (i + 1), n − 1 i=0 n − 1 i=0 = (42) where, considering equation (6), Dj = E [Xn,j Xn+1,j ] = Cj E [ηj (n)ηj (n + 1)] = Cj ρj , (43) for each j ≥ 1. Equations (40)–(43) then lead to kρ − ρbkn k2L2 S(H) (Ω,A,P) ≤ +  2   !2  n−1 X 1 Cj2 E  1 − ηj2 (i)  n j=1 i=0  !2  n−2 X 1 ηj (i + 1)ηj (i)  +E  ρj − n − 1 i=0 2 n n−1 Ck2n ∞ X ρ2j . j=kn +1 14 kn X (44) For each j ≥ 1, and for n sufficiently large, considering equations (32)–(33), under Assumption A4 E kρ − 2 ρbkn kS(H) 2 ≤  n n−1 Ck2n 2S ≤ 2  n n−1 Ck2n n kn X e j,1 + K e j,2 K n Cj2 j=1 2 kn X Cj2 + j=1 ∞ X ! + ∞ X ρ2j j=kn +1 ρ2j . j=kn +1 (45) From the trace property of operator C, lim n→∞ kn X Cj2 = ∞ X j=1 j=1 Cj2 < ∞, (46) and from the Hilbert-Schmidt property of ρ, lim n→∞ ∞ X ρ2j = 0. (47) j=kn +1 Thus, in view of equations (45)–(47), kρ − 2 ρbkn kL2 (Ω,A,P) S(H) = E kρ − where 2S g(n) =  n n−1 Ck2n n 2 ρbkn kS(H) 2 kn X ≤ g(n) = O Ck2n n ! , n → ∞, (48) Cj2 + j=1 1 ∞ X ρ2j . (49) j=kn +1 Under Assumption A3, equations (48) and (49) imply lim kρ − ρbkn k2L2 n→∞ as we wanted to prove. S(H) (Ω,A,P) = 0,  Note that consistency of ρbkn in the space S (H) directly follows from equation (35) in Proposition 1. Corollary 1 Let X = (Xn , n ∈ Z) be a zero-mean standard ARH(1) process. Under Assumptions A1-A4, as n → ∞, kρ − ρbkn kS(H) →p 0, where, as usual, →p denotes the convergence in probability. 15 (50) 3.2 Consistency of the ARH(1) plug-in predictor. Let us consider L (H) , as before, denoting the space of bounded linear operators on H, with the norm kAkL(H) = sup X∈H kA (X)kH , kXkH (51) for every A ∈ L (H) . In particular, for each X ∈ H, kA (X)kH ≤ kAkL(H) kXkH . (52) bn = ρbkn (Xn−1 ) X (53) In the following, we denote by the ARH(1) plug-in predictor of Xn , as an estimator of the conditional expectation E[Xn |Xn−1 ] = bn = ρbkn (Xn−1 ) in H. ρ (Xn−1 ) . The following proposition provides the consistency of X Proposition 2 Let X = (Xn , n ∈ Z) be a zero-mean standard ARH(1) process. Under Assumptions A1-A4, lim E k(ρ − ρbkn ) (Xn−1 )kH = 0. (54) n→∞ Specifically, In particular, E k(ρ − ρbkn ) (Xn−1 )kH ≤ h (n) , h (n) = O  1 √ Ckn n  , n → ∞. k(ρ − ρbkn ) (Xn−1 )kH →p 0, (55) (56) where, as usual, →p denotes the convergence in probability. Proof. From (52) and Proposition 1, for n sufficiently large, the following almost surely inequality holds: bn ρ (Xn−1 ) − X H ≤ kρ − ρbkn kL(H) kXn−1 kH , bn = ρbkn (Xn−1 ) . Thus, where, as given in equation (53), X bn E ρ (Xn−1 ) − X H h i ≤ E kρ − ρbkn kL(H) kXn−1 kH . (57) (58) From Cauchy-Schwarz’s inequality, keeping in mind that, for a Hilbert-Schmidt operator K, it always 16 holds that kKkL(H) ≤ kKkS(H) , we have from equation (58), q q 2 2 bn E Xn − X ≤ E kρ − ρbkn kL(H) E kXn−1 kH H q q q 2 ≤ E kρ − ρbkn kS(H) E kXn−1 k2H = E kρ − ρbkn k2S(H) σX , 2 where, as before, σX = E kXn−1 k2H = ∞ X j=1 (59) Cj < ∞, for each n ∈ Z (see equation (15)). Since from Proposition 1 (see equation (36)), kρ − 2 ρbkn kL2 (Ω,A,P) S(H) ≤ g(n), 1 with g(n) = O Ck2n n ! , n → ∞, from equation (59), we obtain, where h (n) = σX E k(ρ − ρbkn ) (Xn−1 )kH ≤ h (n) , (60) lim E k(ρ − ρbkn ) (Xn−1 )kH = 0, (61) p g (n), with g (n) being given in (49). In particular, under Assumption A3, n→∞ which implies that bn k(ρ − ρbkn ) (Xn−1 )kH = ρ (Xn−1 ) − X H →p 0, n → ∞. (62)  4 The Gaussian case. In this section, we restrict our attention to the Gaussian ARH(1) context. In that context, we prove that, under Assumptions A1-A2 and Assumption A4 holds. From equation (17),  n−1  X 2 ηj (i)    i=0    = 1, E  n   We now compute the variance of n−1 X n ≥ 1. (63) ηj2 (i) i=0 n , applying the theory of quadratic forms for Gaussian vectors with correlated components. Specifically, for each j ≥ 1, and n ≥ 2, the n × 1 random vector ηjT = (ηj (0), . . . , ηj (n − 1)) is Multivariate Normal distributed, with null mean vector, and covariance 17 matrix  Σn×n where, under  1   ρ  j   = 0   .  ..   0 Assumptions ρj 0 ... ... 1 ρj 0 ... ρj .. . 1 .. . ρj .. . ... .. . ... 0 0 ρj A1–A2, we  0   0     0   ..  .    1 , (64) n×n have applied (20), i.e., E[ηj (i)ηj (i + 1)] = ρj , for every i ∈ Z, and j ≥ 1. It is well-known (see, for example, [32]) that the variance of a quadratic form defined from a multivariate Gaussian vector y ∼ N (µ, Λ), and a symmetric matrix Q is given by:   Var yT Qy = 2tr [QΛQΛ] + 4µT QΛQµ, (65) where tr denotes the trace. For each j ≥ 1, applying equation (65), with y = ηj , Λ = Σn×n , in (64), and Q = In×n , the n × n identity matrix, we obtain   Var ηjT In×n ηj = Var "n−1 X # ηj2 (i) i=0  = 2tr [Σn×n Σn×n ] = 2 n + 2(n − 1)ρ2j . (66) Furthermore, from equation (66), for each j ≥ 1,  n−1  X 2 ηj (i)      i=0    = 2 n + 2(n − 1)ρ2j = 2 + 4 1 − 1 ρ2j . Var    n2 n n n n2   (67) From equation (67), we then obtain   n−1 X 2 ηj (i)     i=0   lim Var   n→∞ n   2   ηj2 (i)       i=0   1 − = lim E    n→∞  n       = 2 +4 n→∞ n lim 18 n−1 X  1 1 − 2 n n  ρ2j = 0. (68) Equation (68) leads to lim n→∞  n−1  X 2 η (i)    i=0 j   Var    n   = 2 + 4ρ2j . 1 n (69) Hence, for each j ≥ 1, Kj,1 in equation (28) is given by Kj,1 = 2 + 4ρ2j ,   n−1 X 2 η (i)    i=0 j   ≤ 2+4  and, from equation (67), Var   n   1 n − 1 n2  (70) e j,1 ρ2j ≤ 2 + 4ρ2j ≤ 6. Thus, for every j ≥ 1, K in equation (30) satisfies e j,1 ≤ 6. K (71) Remark 7 Note that, from Lemma 1, for each j ≥ 1,   E ηej4 (i) = 3, ∀i ∈ Z. (72) Thus, the assumption considered in Remark 2 holds, and for each j ≥ 1, the AR(1) process ηj = (ηj (n), n ∈ Z) is ergodic for all second-order moments, in the mean-square sense (see pp. 192–193 of [33]). For n ≥ 2, and for each j ≥ 1, we are now going to compute Kj,2 in (29). The (n − 1) × 1 T T random vectors ηj⋆ = (ηj (0), . . . , ηj (n − 2)) and ηj⋆⋆ = (ηj (1), . . . , ηj (n − 1)) are Multivariate Normal distributed, with null mean vector, and covariance matrix  e (n−1)×(n−1) Σ       =      1 ρj 0 ... ... ρj 1 ρj 0 ... 0 .. . ρj .. . 1 .. . ρj .. . ... .. . 0 ... 0 0 ρj From equation (20), for each j ≥ 1, 19  0   0     0   ..  .    1 (n−1)×(n−1) . (73) E "n−2 X # ηj (i)ηj (i + 1) = i=0 n−2 X i=0   ρj = (n − 1)ρj = tr E ηj⋆ [ηj⋆⋆ ]T , (74) where     E ηj⋆ [ηj⋆⋆ ]T = E ηj⋆ ⊗ ηj⋆⋆ = ρj I(n−1)×(n−1) , (75) with, as before, I(n−1)×(n−1) denoting the (n − 1) × (n − 1) identity matrix. However, the variance of n−2 X ηj (i)ηj (i+1) depends greatly on the distribution of ηj⋆ and ηj⋆⋆ . In the Gaussian case, keeping in mind i=0 that ηj⋆ = (ηj (0), . . . , ηj (n − 2))T and ηj⋆⋆ = (ηj (1), . . . , ηj (n − 1))T are zero-mean multivariate Normal e (n−1)×(n−1) given in (73), and having cross-covariance matrix distributed vectors with covariance matrix Σ n−2 X ηj (i)ηj (i + 1), from (74)-(75), as follows: (75), we can compute the variance of i=0     Var [ηj⋆ ]T I(n−1)×(n−1) ηj⋆⋆ = E [ηj⋆ ]T I(n−1)×(n−1) ηj⋆⋆ [ηj⋆ ]T I(n−1)×(n−1) ηj⋆⋆ 2  − E [ηj⋆ ]T I(n−1)×(n−1) ηj⋆⋆ = n−2 X X n−2 i=0 p=0 = n−2 X E [ηj (i)ηj (i + 1)] i=0 + 2  E [ηj (i)ηj (i + 1)ηj (p)ηj (p + 1)] − E [ηj⋆ ]T I(n−1)×(n−1) ηj⋆⋆ n−2 X n−2 X n−2 X E [ηj (p)ηj (p + 1)] p=0 E [ηj (i)ηj (p)] E [ηj (i + 1)ηj (p + 1)] i=0 p=0 + n−2 X n−2 X i=0 p=0 2  E [ηj (i)ηj (p + 1)] E [ηj (i + 1)ηj (p)] − E [ηj⋆ ]T I(n−1)×(n−1) ηj⋆⋆   2  e (n−1)×(n−1) Σ e (n−1)×(n−1) = tr E[ηj⋆ ⊗ ηj⋆⋆ ] + tr Σ  2  T   − tr E[ηj⋆ ⊗ ηj⋆⋆ ] +tr E[ηj⋆ ⊗ ηj⋆⋆ ] E[ηj⋆ ⊗ ηj⋆⋆ ]       e (n−1)×(n−1) Σ e (n−1)×(n−1) + tr E[η ⋆ ⊗ η ⋆⋆ ] E[η ⋆ ⊗ η ⋆⋆ ] T = tr Σ j j j j = (n − 1) + 2(n − 2)ρ2j + (n − 1)ρ2j , where, from (75),        ⋆ ⋆⋆ ⋆ ⋆⋆ T E[ηj ⊗ ηj ] E[ηj ⊗ ηj ] =     ρ2j 0 ... ... 0 0 .. . ρ2j .. 0 .. . 0 .. . 0 ... ... .. . .. . From (76), 20 . .. . ρ2j       = ρ2j I(n−1)×(n−1) .    (76)  n−2  X ηj (i)ηj (i + 1)    i=0  (n − 1) + 2(n − 2)ρ2j + (n − 1)ρ2j = Var  .   n−1 (n − 1)2   Therefore, for each j ≥ 1, lim n→∞  n−2  X ηj (i)ηj (i + 1)    i=0    Var   n−1   1/n Thus, for each j ≥ 1, Kj,2 in (29) is given by Kj,2 = 1 + 3ρ2j . (77) (78)   n−2 X ηj (i)ηj (i + 1)     i=0 2 ≤  = 1+3ρj . From equation (77), Var  n−1    e j,2 in equation (31) satisfies 1 + 3ρ2j ≤ 4. Thus, for every j ≥ 1, K e j,2 ≤ 4. K (79) From equations (71) and (79), the constant S in Assumption A4 is such that S ≤ 6 + 4 = 10. 5 Simulations. A simulation study is undertaken to illustrate the behavior of the formulated componentwise estimator of the autocorrelation operator, and of its associated ARH(1) plug-in predictor for large sample sizes, in Section 5.1. By implementation of alternative ARH(1) plug-in prediction techniques, from the previous and current literature, a comparative study is also developed in Section 5.2. In the subsequent sections, we restrict our attention to the Gaussian ARH(1) framework. bn for large sample sizes 5.1 Behavior of ρb and X In this subsection, the empirical functional mean-squared errors, associated with the proposed componentwise estimator of the autocorrelation operator, and associated ARH(1) plug-in predictor are evaluated from repeated generations of a sequence of samples with increasing, sufficiently large, sizes. In the simulation of that samples, from an ARH(1) process, operators ρ, C and Rε are defined within the family of fractional elliptic polynomials, with constant coefficients, of the Dirichlet negative Laplacian operator on an interval. Let us first consider the Dirichlet negative Laplacian operator on an interval (a, b), (−∆)(a,b) , given by 21 (−∆)(a,b) (f ) (x) f (a) = − d2 f (x) , dx2 x ∈ (a, b) ⊂ R, = f (b) = 0. (80)   The eigenvectors (φj , j ≥ 1) , and eigenvalues λj (−∆)(a,b) , j ≥ 1 of (−∆)(a,b) satisfy, for each j ≥ 1,  = λj (−∆)(a,b) φj (x) , (−∆)(a,b) φj (x) x ∈ (a, b), φj (a) = φj (b) = 0. (81) For each j ≥ 1, the solution to equation (81) is given by (see [27], p. 6): φj (x) λj (−∆)(a,b)   2 πjx , sin b−a b−a π2 j 2 , j ≥ 1. (b − a)2 =  = ∀x ∈ [a, b] , j ≥ 1, (82) (83) We consider here operator C defined as (see Remark 1) C = (−∆)(a,b) −δ1 /2 , δ1 > 1. (84) From [18], pp. 119-140, the eigenvectors of C coincide with the eigenvectors of (−∆)(a,b) , and its eigenvalues (Cj , j ≥ 1) are given by:  −δ1 /2 Cj = λj (−∆)(a,b) =  π2 j 2 (b − a)2 −δ1 /2 = aj −δ1 , a=  π b−a −δ1 , δ1 > 1. (85) Additionally, considering ρ= " λ1 (−∆)(a,b)  (−∆)(a,b) − ǫ #−δ2 /2 , 1 < δ2 < 2, (86)  for certain positive constant ǫ < λ1 (−∆)(a,b) close to zero, ρ is a positive self-adjoint Hilbert-Schmidt operator (with a non extremely fast decay rate), whose eigenvectors coincide with the eigenvectors of (−∆)(a,b) , and whose eigenvalues (ρj , j ≥ 1) are such that ρj < 1, for every j ≥ 1, and ρ2j =  !−δ2 λj (−∆)(a,b)  , λ1 (−∆)(a,b) − ǫ ρ2j ∈ (0, 1) , 22 1 < δ2 < 2, j ≥ 1, (87) where, as before,    0 < λ1 (−∆)(a,b) ≤ λ2 (−∆)(a,b) ≤ · · · ≤ λj (−∆)(a,b) ≤ . . . .  From (19), the eigenvalues σj2 , j ≥ 1 of Rε are then defined, for each j ≥ 1, as σj2 = Cj 1 − ρ2j  = [λj −(δ1 /2+δ2 )  −δ1 /2 λj (−∆)(a,b) (−∆)(a,b) ] −  −δ2 . λ1 (−∆)(a,b) − ǫ (88) Note that Rε is in the trace class, since the trace property of C, and the fact that ρj < 1, for every j ≥ 1, implies ∞ X σj2 = ∞ X j=1 j=1 ∞  X Cj < ∞. Cj 1 − ρ2j < j=1 For this particular example of operator C, we have considered truncation parameter kn of the form kn = n1/α , (89) for a suitable α > 0, which, in particular, allows verification of (27). From equation (85),  −δ1 /2 √ √ Ckn n = λkn −∆(a,b) n=  πkn (b − a) −δ1 √ n, δ1 > 1. (90) From equation (89), Assumption A3 is then satisfied if 1/2 − δ1 > 0, α i.e., if α > 2δ1 > 2, since δ1 > 1. (91) Let us fix δ1 = 2.4 and δ2 = 1.1. Then, from equation (91), α > 48/10. In particular, the values α1 = 5 and α2 = 6 have been tested, in Table 1 below, for H = L2 ((a, b)), and (a, b) = (0, 4), where L2 ((a, b)) denotes the space of square integrable functions on (a, b). Generations have been performed under Assumptions A1–A4, for our particular choice of C and selection of kn in a diagonal scenario. The computed empirical truncated functional mean square error EMSEρbkn of the estimator ρbkn of ρ, with n being, as before, the sample size, is given by: EMSEρbkn ρbw n,j = = N kn 2 1 XX , ρj − ρbw n,j N w=1 j=1 bw D n,j = bw C n,j 23 1 n−1 1 n n−2 X w w Xi,j Xi+1,j i=0 n−1 X i=0 (92)  w 2 Xi,j , (93) where N denotes the number of simulations, and for each j = 1, . . . , kn , ρbw n,j represents the estimator of ρj , w w w based on the wth generation of the values X0,j , . . . , Xn−1,j , with Xi,j = hXiw , φj iH , for w = 1, . . . , 700, bn = ρbkn (Xn−1 ) , we compute the empirical version and i = 0, . . . , n − 1. For the plug-in predictor X UB(EMAEXbnkn ) of the derived upper bound (59), which, for each n ∈ Z, is given by UB(EMAEXbnkn ) = = 2 where σ bX = kn X v u kn N X u1 X 2 2 \ w t ρj − ρbw E Xn−1 n,j H N w=1 j=1 v u kn N X q u1 X 2 2 t σ b = ρj − ρbw EMSEρbkn σ bX , X n,j N w=1 j=1 (94) Cj . j=1 From N = 700 realizations, for each one of the values of the sample sizes (nt , t = 1, . . . , 20) = (15000 + 20000(t − 1), t = 1, . . . , 20) , the EMSEρbkn and UB(EMAEXbnkn ) values, for α = 5 and α = 6, are displayed in Table 1, where the abbreviated notations M SEρbkn,1 , for EMSEρbkn , and UBXb k n n,1 UB(EMAEXbnkn ), are used (see also Figures 1-2). 24 , for n kn,1 M SEρbkn,1 n1 = 15000 6 3.74 (10) n2 = 35000 8 n3 = 55000 UBXb k n n,1 −4 2.87 (10) 2.15 (10) −4 8 1.34 (10)−4 n4 = 75000 9 1.09 (10) n5 = 95000 9 9.48 (10)−5 n6 = 115000 10 8.31 (10) n7 = 135000 10 n8 = 155000 −4 kn,2 −2 4 2.45 (10) 2.21 (10) −2 5 1.75 (10)−2 1.35 (10) −4 1.71 (10) 6 1.03 (10)−4 1.51 (10)−2 1.57 (10) −2 6 7.55 (10) 1.29 (10) 1.47 (10)−2 6 5.86 (10)−5 1.14 (10)−2 −2 6 5.16 (10) 6.81 (10) −5 1.25 (10) −2 7 10 6.37 (10)−5 1.21 (10)−2 n9 = 175000 11 6.14 (10) n10 = 195000 11 5.34 (10)−5 n11 = 215000 11 4.67 (10) n12 = 235000 11 n13 = 255000 −5 −2 −2 −2 −5 1.07 (10) 4.86 (10) −5 1.04 (10) 7 3.88 (10)−5 9.66 (10)−3 1.19 (10) −2 7 3.87 (10) 9.65 (10) 1.11 (10)−2 7 3.42 (10)−5 8.79 (10)−3 −2 7 3.40 (10) −5 1.03 (10) 4.66 (10) −5 1.03 (10) −2 7 12 4.53 (10)−5 1.02 (10)−2 n14 = 275000 12 4.24 (10) n15 = 295000 12 3.72 (10)−5 n16 = 315000 12 3.62 (10) n17 = 335000 12 n18 = 355000 −5 −2 −2 −3 −5 8.74 (10) 2.92 (10) −5 8.12 (10) 7 2.77 (10)−5 7.95 (10)−3 9.95 (10) −3 8 2.77 (10) 7.94 (10) 9.32 (10)−3 8 2.67 (10)−5 7.76 (10)−3 −3 8 2.55 (10) −5 9.21 (10) 3.39 (10) −5 8.91 (10) −3 8 12 3.34 (10)−5 8.86 (10)−3 n18 = 375000 13 3.34 (10) n20 = 395000 13 3.12 (10)−5 −5 k n n,2 2.25 (10) 1.39 (10) −5 UBXb −4 −5 −5 M SEρbkn,2 −5 −3 −3 −3 −5 7.64 (10) 2.28 (10) −5 7.04 (10) 8 2.20 (10)−5 7.04 (10)−3 8.86 (10) −3 8 2.04 (10) 6.84 (10) 8.56 (10)−3 8 1.92 (10)−5 6.65 (10)−3 −5 Table 1: EM SEρbkn (here, M SEρbkn,i ), and UB(EMAEXbnkn ) (here, UBXb k n n,i −3 −3 −3 ) values, in (92)–(94), based on N = 700 simulations, for δ1 = 2.4 and δ2 = 1.1, considering the sample sizes nt = 15000+20000(t−1), t = 1, . . . , 20, and the corresponding kn,1 and kn,2 values, for α1 = 5 and α2 = 6 25 −4 8 −4 x 10 8 x 10 EMCE for operator, α = 5 1/T EMCE for operator, α = 6 1/T 3/4 3/4 (1 / T) (1 / T) 7 7 6 6 5 5 4 4 3 3 2 2 1 1 0 0 0.5 1 1.5 2 n 2.5 3 3.5 0 4 0 0.5 1 1.5 5 x 10 2 n 2.5 3 3.5 4 5 x 10 Figure 1: EM SEρbkn values (blue line), in (92)–(93), based on N = 700 simulations, for δ1 = 2.4 and δ2 = 1.1, considering the sample sizes T = nt = 15000 + 20000(t − 1), t = 1, . . . , 20, and the corresponding kn,1 and kn,2 values, for α1 = 5 (left-hand side) and α2 = 6 (right-hand side), against (1/T )3/4 = (1/nt )3/4 (black dot line) and 1/T = (1/nt ) (red dot line) 0.07 0.07 EMCE for predictor, α = 5 EMCE for predictor, α = 6 (1 / T)1/2 (1 / T)1/2 1/3 1/3 (1 / T) (1 / T) 0.06 0.06 0.05 0.05 0.04 0.04 0.03 0.03 0.02 0.02 0.01 0.01 0 0 0.5 1 1.5 2 n 2.5 3 3.5 0 4 5 x 10 0 0.5 1 1.5 2 n 2.5 3 3.5 4 5 x 10 Figure 2: UB(EMAEXbnkn ) values (blue line), in (94), based on N = 700 simulations, for δ1 = 2.4 and δ2 = 1.1, considering the sample sizes T = nt = 15000 + 20000(t − 1), t = 1, . . . , 20, and the corresponding kn,1 and kn,2 values, for α1 = 5 (left-hand side) and α2 = 6 (right-hand side), against (1/T )1/2 = (1/nt )1/2 (red dot line) and (1/T )1/3 = (1/nt )1/3 (black dot line) In the example considered, a one-parameter model of kn is selected depending on parameter α. In Example 2, in p. 286 in [31], with the same spirit, for an equivalent spectral class of operators C, a three parameter model is established for kn , to ensure convergence in quadratic mean in the space L(H) of the componentwise estimator of ρ constructed from the known eigenvectors of C. The numerical results displayed in Table 1 and Figures 1-2 illustrate the fact that the proposed componentwise estimator ρbkn presents a velocity of convergence to ρ, in quadratic mean in S(H), faster than n−1/3 , which corresponds to the optimal case for the componentwise estimator of ρ proposed in [31], in the case of known eigenvectors of C (see, in particular, Theorem 1, Remark 2 and Example 2 in [31]). For lager values of the parameters δ1 than 2.4, and α than 6, a faster velocity of convergence of ρbkn to ρ, in quadratic mean in the space S(H), will be obtained. However, larger sample sizes are required for larger values of α, in 26 order to estimate a given number of coefficients of ρ. A more detailed discussion about comparison of the rates of convergence of the ARH(1) plug-in predictors proposed in [4], [9], [11] and [31] can be found in the next subsection. 5.2 A comparative study In this section, the performance of our approach is compared with those ones given in [4, 9, 11, 31], including the case of unknown eigenvectors of C. In the last case, our approach and the approaches presented in [11] and [31] are implemented in terms of the empirical eigenvectors (see section 5.2.2). 5.2.1 Theoretical-eigenvector-based componentwise estimators Let us first compare the performance of our ARH(1) plug-in predictor, defined in (53), and the ones formulated in [11] and [31], in terms of the theoretical eigenvectors (φj , j ≥ 1) of C. Note that, in this first part of our comparative study, we consider the previous generated Gaussian ARH(1) process, with autocovariance and autocorrelation operators defined from equations (85) and (87), for different rates of convergence to zero of parameters Cj and ρ2j , j ≥ 1, with both sequences being summable sequences. Since we restrict our attention to the Gaussian case, conditions A1 , B1 and C1 , formulated in [11], pp. 211–212, are satisfied by the generated ARH(1) process. Similarly, Conditions H1 –H3 in p. 283 of [31] are satisfied by our generated Gaussian ARH(1) process. The estimators of ρ from [11] and [31] we are going to test here are given as follows. In Section 8.2 of [11] the following estimator of ρ is proposed ρbn (x) ρbn,l (x) = =  kn  X −1 kn b (x) = ρbn,l (x)φl , Π Dn Cn Π kn l=1 x ∈ H, n−2 kn 1 1 XX hφ , xiH Xi,j Xi+1,l , bn,j j n − 1 i=0 j=1 C (95) (96) in the finite dimensional subspace Hkn = sp (φ1 , . . . , φkn ) of H, where Πkn is the orthogonal projector over Hkn , and, as before, Xi,j = hXi , φj iH , for j ≥ 1. A modified estimator of ρ is studied in [31], Section 2, given by ρbn,a (x) ρbn,a,l (x) −1 bn,a where C (x) = kn X j=1 = kn   X −1 kn bn,a (x) = ρbn,a,l (x)φl , Π Πkn Dn C l=1 = x ∈ H, n−1 kn 1 1 XX  hφj , xiH Xi,j Xi+1,l ,  n − 1 i=1 j=1 max C bn,j , an (97) (98) 1  hφj , xiH φj (a.s.). Here, (an , n ∈ N) is such that (see Theorem 1  bn,j , an max C 27 in [31]) α Ckγn ≤ an ≤ βλkn , nε α > 0, 0 < β < 1, ε < 1/2, γ ≥ 1. (99) Tables 2-3 below display results for ρbkn = ρbkn , where ρbkn is given in equations (25)-(26) (see third column), ρbkn = ρbn , where ρbn is given in equations (95)–(96) (see fourth column), and ρbkn = ρbn,a , where ρbn,a is defined in (97)–(98) (see fifth column). Specifically, Tables 2-3 below provide the truncated, for two different kn rules, empirical values of Ekρ (Xn−1 ) − ρbkn (Xn−1 )kH , based on N = 700 generations of each one of the functional samples considered with size nt = 15000 + 20000(t − 1), t = 1, . . . , 20, when Cj and ρ2j , j ≥ 1, are defined as done above in equations (85) and (87). Specifically, in Table 2, the following parameter values have been considered: δ1 = 2.4 δ2 = 1.1, and kn = n1/α , for α = 6, according to our Assumption A3, (which is also considered in p. 217 of [11], to ensure weak consistency of the proposed estimator of ρ). While, in Table 3, the same empirical values are displayed for δ1 = 61 60 , δ2 = 1.1, and kn is selected according to Example 2, in p. 286 in [31]. Thus, in Table 3, 1−2ǫ kn = n δ1 (4+2γ) , γ ≥ 1, ǫ < 1/2. (100) In particular we have chosen γ = 2, and ǫ = 0.04δ1 . Note that, from Theorem 1 and Remark 1 in [31], for the choice made of kn in Table 3, convergence to ρ, in quadratic mean in the space L(H), holds for ρbn,a given in (98). 28 n kn Our Approach n1 = 15000 4 2.25 (10) n2 = 35000 5 n3 = 55000 Bosq (2000) −2 2.57 (10) 1.71 (10) −2 6 1.51 (10)−2 n4 = 75000 6 1.29 (10) n5 = 95000 6 1.14 (10) n6 = 115000 6 1.07 (10) n7 = 135000 7 n8 = 155000 Guillas (2001) −2 2.36 (10) 1.72 (10) −2 1.84 (10) 1.65 (10)−2 1.53 (10)−2 −2 1.46 (10) −2 1.20 (10) −2 1.10 (10) 1.04 (10) −2 7 9.66 (10)−3 n9 = 175000 7 9.65 (10) n10 = 195000 7 8.79 (10) n11 = 215000 7 8.74 (10) n12 = 235000 7 n13 = 255000 −2 −2 −2 1.37 (10) −2 1.16 (10) −2 1.11 (10) 1.06 (10) −2 1.07 (10) 9.91 (10)−3 1.01 (10)−2 −3 9.79 (10) −3 9.12 (10) −3 8.79 (10) 8.12 (10) −3 7 7.95 (10)−3 n14 = 275000 8 7.94 (10) n15 = 295000 8 7.76 (10) n16 = 315000 8 7.64 (10) n17 = 335000 8 n18 = 355000 −2 −2 −2 −2 −3 9.68 (10) −3 8.93 (10) −3 8.83 (10) 8.69 (10) −3 8.75 (10) 8.53 (10)−3 8.73 (10)−3 −3 8.52 (10) −3 8.49 (10) −3 7.88 (10) 7.04 (10) −3 8 7.04 (10)−3 n19 = 375000 8 6.84 (10) n20 = 395000 8 6.65 (10) −3 −3 −3 −3 −3 8.58 (10) −3 8.36 (10) −3 8.13 (10) 7.24 (10) −3 7.59 (10) 7.23 (10)−3 6.92 (10)−3 −3 6.89 (10) −3 6.67 (10) −3 −3 −3 −3 −3 6.90 (10) −3 −3 6.85 (10) −3 Table 2: Truncated empirical values of Ekρ (Xn−1 ) − ρbkn (Xn−1 )kH , for ρbkn given in equations (25)-(26) (third column), in equations (95)–(96) (fourth column), and in equations (97)–(98) (fifth column), based on N = 700 simulations, for δ1 = 2.4 and δ2 = 1.1, considering the sample sizes nt = 15000+20000(t−1), t = 1, . . . , 20, and the corresponding kn = n1/α values, for α = 6 (from Assumption A3), assuming that (φj , j ≥ 1) are known. 29 n kn Our Approach n1 = 15000 2 9.91 (10) n2 = 35000 3 n3 = 55000 Bosq (2000) −3 1.39 (10) 8.78 (10) −3 3 7.89 (10)−3 n4 = 75000 3 6.49 (10) n5 = 95000 3 6.36 (10) n6 = 115000 3 6.14 (10) n7 = 135000 3 n8 = 155000 Guillas (2001) −2 1.26 (10) 1.34 (10) −2 1.24 (10) 1.15 (10)−2 1.14 (10)−2 −3 1.01 (10) −3 9.09 (10) −3 7.65 (10) 5.91 (10) −3 3 5.73 (10)−3 n9 = 175000 3 5.44 (10) n10 = 195000 3 5.10 (10) n11 = 215000 4 5.01 (10) n12 = 235000 4 n13 = 255000 −2 −2 −2 8.58 (10) −3 8.29 (10) −3 7.26 (10) 7.03 (10) −3 6.69 (10) 6.77 (10)−3 6.54 (10)−3 −3 6.74 (10) −3 6.69 (10) −3 6.48 (10) 4.85 (10) −3 4 4.17 (10)−3 n14 = 275000 4 4.64 (10) n15 = 295000 4 4.55 (10) n16 = 315000 4 4.48 (10) n17 = 335000 4 n18 = 355000 −3 −3 −3 −3 −3 6.16 (10) −3 5.97 (10) −3 5.94 (10) 6.45 (10) −3 5.83 (10) 6.17 (10)−3 5.68 (10)−3 −3 5.99 (10) −3 5.94 (10) −3 5.69 (10) 4.38 (10) −3 4 4.16 (10)−3 n19 = 375000 4 3.91 (10) n20 = 395000 4 3.86 (10) −3 −3 −3 −3 −3 5.60 (10) −3 5.58 (10) −3 5.50 (10) 5.58 (10) −3 5.44 (10) 5.45 (10)−3 5.42 (10)−3 −3 5.34 (10) −3 5.29 (10) −3 −3 −3 −3 −3 5.32 (10) −3 −3 5.26 (10) −3 Table 3: Truncated empirical values of Ekρ (Xn−1 )− ρbkn (Xn−1 )kH , for ρbkn given in equations (25)–(26) (third column), in equations (95)–(96) (fourth column), and in equations (97)–(98) (fifth column), based 61 and δ2 = 1.1, considering the sample sizes nt = 15000 + 20000(t− 1), on N = 700 simulations, for δ1 = 60 t = 1, . . . , 20, and the corresponding kn given in (100). It can be observed in Table 2 a similar performance of the three methods compared with the truncation order kn satisfying our Assumption A3, with slightly worse results being obtained from the estimator defined in equations (97)–(98), specially for the sample size n8 = 155000. Furthermore, in Table 3, a better performance of our approach is observed for the smallest sample sizes (from n1 = 15000 until n4 = 75000). While, for the remaining largest sample sizes, slightly differences are observed, with a better performance of our approach, very close to the other two approaches presented in [11] and [31]. 30 5.2.2 Empirical-eigenvector-based componentwise estimators In this section, we address the case where (φj , j ≥ 1) are unknown, as usually it occurs in real-data problems. Specifically, an empirical version of the estimators of ρ formulated in equations (25)–(26), (95)–(96) and (97)–(98), consisting of replacing the theoretical eigenvectors (φj , j ≥ 1) by their empirical n−1 X Xi ⊗ Xi , n ≥ 1, is now studied. (random) counterparts (φn,j , j ≥ 1) , the eigenvectors of Cn = n1 i=0 Specifically, taking (Cn,j , j ≥ 1) as the eigenvalues of Cn , we consider 1 n−1 ρen,j = ρen (x) = ρen,a (x) = ρen,l (x) ρen,a,l (x) 1 n n−2 X ei+1,j ei,j X X i=0 n−1 X i=0 ei,j X 2 , ρekn = kn X j=1 ρen,j φn,j ⊗ φn,j , kn   X e kn (x) = e kn Dn Cn−1 Π Π ρen,l (x)φn,l , l=1 = n,a l=1 = x ∈ H, n−2 kn 1 XX 1 ei+1,l , ei,j X hφn,j , xiH X n − 1 i=0 j=1 Cn,j kn   X e kn (x) = e kn Dn C −1 Π ρen,a,l (x)φn,l , Π (101) (102) x ∈ H, n−2 kn 1 1 XX ei+1,l, ei,j X hφn,j , xiH X n − 1 i=0 j=1 max (Cn,j , an ) (103) ei,j = hXi , φn,j i , Π e kn denotes the orthogonal projector into the space where, for i ∈ Z, and j ≥ 1, X H e kn = sp (φn,1 , . . . , φn,kn ) . The Gaussian ARH(1) process is generated under Assumptions A1–A2, H as well as C1′ in p. 218 in [11]. Note that conditions A1 and B1′ in [11] already hold. Moreover, as given in Theorem 8.8 and Example 8.6, in p. 221 in [11], for Cj , ρ2j , j ≥ 1, defined as above in equations (85) and (87), with, in particular, δ1 = 2.4, and δ2 = 1.1, the estimator ρen converges almost surely to ρ under the condition 2 nCkn 2 −→ ∞,  kn X Log(n)  bj  j=1 √ √   −1 where b1 = 2 2 (C1 − C2 ) and bj = 2 2 max (Cj−1 − Cj )−1 , (Cj − Cj+1 )−1 , for j ≥ 2. Particularly, in Table 4 the truncation criteria from Assumption A3 is used, and in Table 4, kn ≃ Log(n) has been tested (see Example 8.6, in [11], p. 221). 31 n kn n1 = 15000 9 8.42 (10) n2 = 35000 10 5.51 (10) n3 = 55000 10 Our approach Bosq (2000) Guillas (2001) −2 1.0614 1.0353 −2 1.0186 1.0052 −2 1.0174 0.9986 −2 1.0153 0.9951 −2 1.0127 0.9883 −2 1.0113 0.9627 4.75 (10) n4 = 75000 11 4.43 (10) n5 = 95000 11 3.68 (10) n6 = 115000 11 3.51 (10) n7 = 135000 11 3.23 (10) −2 1.0081 0.9247 n8 = 155000 11 2.95 (10)−2 1.0066 0.9119 n9 = 175000 12 2.94 (10) −2 1.0057 0.9113 n10 = 195000 12 2.80 (10) −2 0.9948 0.8912 n11 = 215000 12 2.71 (10) −2 0.9017 0.8615 n12 = 235000 12 2.59 (10) −2 0.8896 0.8201 n13 = 255000 12 2.58 (10)−2 0.8783 0.8004 n14 = 275000 12 2.35 (10) −2 0.8719 0.7832 −2 0.8602 0.7780 −2 0.8424 0.7469 n15 = 295000 12 2.28 (10) n16 = 315000 12 2.27 (10) n17 = 335000 12 2.16 (10) −2 0.8217 0.7140 n18 = 355000 12 2.14 (10)−2 0.8001 0.7066 n19 = 375000 12 2.09 (10) −2 0.7778 0.6872 n20 = 395000 12 2.06 (10) −2 0.7693 0.6621 Table 4: Truncated empirical values of E kρ (Xn−1 ) − ρekn (Xn−1 )kH , for ρekn = ρekn given in equation (101) (third column), ρekn = ρen defined in equation (102) (fourth column) and ρekn = ρen,a defined in equation (103) (fifth column), based on N = 700 simulations, for δ1 = 2.4 and δ2 = 1.1, considering the sample sizes nt = 15000 + 20000(t − 1), t = 1, . . . , 20, and kn = Log(n). The time discretization step is assumed as ht = 0.08. A better performance of our estimator (101) in comparison with estimator (102), formulated in [11], and estimator (103), formulated in [31] (see Example 4 and Remark 4, in p. 291 in [31]) is observed below in Table 4. Note that, in particular, in Example 4 and Remark 4, in [31], p. 291, smaller values of kn than Log(n) are required for a given sample size n, to ensure convergence in quadratic mean, and, in particular, weak-consistency. However, considering a smaller discretization step size, applying truncation at kn = n1/6 , (i.e., α = 6), we obtain (see Table 5), for the same parameter values δ1 = 2.4 and δ2 = 1.1, better results than in Table 4, since a smaller number of coefficients of ρ (parameters) to be estimated is considered in Table 5, from a richer sample information coming from the smaller discretization step size considered. In particular, it can be observed in Table 5 below, a similar performance of the three 32 approaches studied, with a worse performance of the approach presented in [31], for n1 = 15000, since a too large number of coefficients of ρ to be estimated is required, for the smallest sample size. To illustrate this fact, in Table 6, the value kn = ⌈e′ n1/(8δ1 +2) ⌉, e′ = 17 10 proposed, in Example 4 and Remark 4, in [31], p. 291, is considered to compute the truncated empirical values of Ekρ(Xn−1 ) − ρekn (Xn−1 )kH , for ρekn defined in equation (101) (third column), for ρekn = ρen given in equation (102) (fourth column), and for ρekn = ρen,a in equation (103) (fifth column). With this truncation rule, requiring a smaller number of coefficients of ρ to be estimated for a given sample size, a better performance of the estimator proposed in [31] is obtained for the largest sample size tested n20 = 395000, with a similar performance of the three approaches tested, for the rest of sample sizes considered, with such a truncation rule. n kn Our approach n1 = 15000 4 9.88 (10) n2 = 35000 5 9.52 (10) n3 = 55000 6 9.12 (10) n4 = 75000 6 n5 = 95000 Bosq (2000) −2 9.25 (10) −2 9.07 (10) −2 8.92 (10) 8.48 (10) −2 6 7.61 (10)−2 n6 = 115000 6 7.05 (10) n7 = 135000 7 6.99 (10) n8 = 155000 7 6.70 (10) n9 = 175000 7 n10 = 195000 Guillas (2001) −2 0.1059 −2 9.86 (10) −2 9.39 (10) 8.64 (10) −2 8.98 (10) 8.30 (10)−2 8.46 (10)−2 −2 7.96 (10) −2 7.84 (10) −2 7.45 (10) 6.49 (10) −2 7 5.88 (10)−2 n11 = 215000 7 5.63 (10) n12 = 235000 7 5.30 (10) n13 = 255000 7 5.05 (10) n14 = 275000 8 n15 = 295000 −2 −2 −2 −2 8.04 (10) −2 7.82 (10) −2 7.40 (10) 7.03 (10) −2 7.07 (10) 6.74 (10)−2 6.80 (10)−2 −2 6.46 (10) −2 6.28 (10) −2 6.19 (10) 4.88 (10) −2 8 4.58 (10)−2 n16 = 315000 8 4.24 (10) n17 = 335000 8 3.86 (10) n18 = 355000 8 3.70 (10) n19 = 375000 8 n20 = 395000 8 −2 −2 −2 −2 −2 6.57 (10) −2 6.37 (10) −2 6.24 (10) 5.99 (10) −2 6.15 (10) 5.74 (10)−2 6.04 (10)−2 −2 5.52 (10) −2 5.24 (10) −2 5.02 (10) 3.55 (10) −2 3.46 (10)−2 −2 −2 −2 −2 −2 5.93 (10) −2 −2 5.70 (10) −2 5.53 (10) 4.88 (10) −2 5.36 (10) 4.70 (10)−2 5.23 (10)−2 −2 −2 −2 Table 5: Truncated empirical values of E kρ (Xn−1 ) − ρekn (Xn−1 )kH , for ρekn defined in equation (101) (third column), for ρekn = ρen given in equation (102) (fourth column), and for ρekn = ρen,a in equation (103) (fifth column), based on N = 200 (due to high-dimensionality) simulations, for δ1 = 2.4 and δ2 = 1.1, considering the sample sizes nt = 15000 + 20000(t − 1), t = 1, . . . , 20, and kn = n1/6 . The time discretization step size is now ht = 0.015. 33 n kn Our approach n1 = 15000 2 6.78 (10) n2 = 35000 2 n3 = 55000 Bosq (2000) −2 8.77 (10) 6.72 (10) −2 2 6.46 (10)−2 n4 = 75000 2 6.24 (10) n5 = 95000 2 5.42 (10) n6 = 115000 2 4.84 (10) n7 = 135000 2 n8 = 155000 Guillas (2001) −2 6.64 (10) 8.61 (10) −2 6.30 (10) 8.48 (10)−2 6.17 (10)−2 −2 8.20 (10) −2 7.84 (10) −2 7.34 (10) 4.27 (10) −2 2 3.64 (10)−2 n9 = 175000 3 3.51 (10) n10 = 195000 3 3.38 (10) n11 = 215000 3 3.16 (10) n12 = 235000 3 n13 = 255000 −2 −2 −2 5.76 (10) −2 5.03 (10) −2 4.56 (10) 6.95 (10) −2 3.94 (10) 6.60 (10)−2 3.65 (10)−2 −2 6.52 (10) −2 6.16 (10) −2 5.78 (10) 2.98 (10) −2 3 2.83 (10)−2 n14 = 275000 3 2.50 (10) n15 = 295000 3 2.23 (10) n16 = 315000 3 2.15 (10) n17 = 335000 3 n18 = 355000 −2 −2 −2 −2 −2 3.42 (10) −2 3.24 (10) −2 2.85 (10) 5.53 (10) −2 2.60 (10) 5.15 (10)−2 2.34 (10)−2 −2 4.85 (10) −2 4.46 (10) −2 4.30 (10) 2.06 (10) −2 3 1.98 (10)−2 n19 = 375000 3 1.89 (10) n20 = 395000 3 1.82 (10) −2 −2 −2 −2 −2 2.05 (10) −2 1.83 (10) −2 1.58 (10) 4.14 (10) −2 1.40 (10) 3.95 (10)−2 1.24 (10)−2 −2 3.77 (10) −2 3.70 (10) −2 −2 −2 −2 −2 1.05 (10) −2 −2 9.93 (10) −3 Table 6: Truncated empirical values of E kρ (Xn−1 ) − ρekn (Xn−1 )kH , for ρekn defined in equation (101) (third column), for ρekn = ρen given in equation (102) (fourth column), and for ρekn = ρen,a in equation (103) (fifth column), based on N = 200 (due to high-dimensionality) simulations, for δ1 = 2.4 and δ2 = 1.1, considering the sample sizes nt = 15000 + 20000(t − 1), t = 1, . . . , 20, and kn = ⌈e′ n1/(8δ1 +2) ⌉, e′ = 17 10 . The time discretization step size is now ht = 0.015. 5.2.3 Kernel-based non-parametric and penalized estimation In practice, curves are observed in discrete times, and should be approximated by smooth functions. In [9], the following optimization problem is considered: bi = argmin LX bi X 2 L2 bi (tj ) = Xi (tj ), , X j = 1, . . . , p, i = 0, . . . , n − 1, (104) where L is a linear differential operator of order d. Our interpolation is computed by Matlab smoothingspline method. Non-linear kernel regression is then considered, in terms of the smoothed functional 34 data, solution to (104), as follows: b hn X n n−2 X i=0 = ρbhn (x) = ρbhn (Xn−1 ),  bi+1 K  X  n−2 X i=0   K bi − x X 2 L2 hn bi − x X 2 L2 hn     , (105)   where K is the usual Gaussian kernel, and bi − x X 2 L2 = Z bi (t) − x(t))2 dt, (X i = 0, . . . , n − 2. Alternatively, in [9], prediction, in the context of Functional Autoregressive processes (FAR(1) processes), under the linear assumption on ρ, which is considered to be a compact operator, with kρk < 1, b1 , . . . , X bn , solving the optimization problem is also studied, from smooth data X  p n−1 2 1 X 1 X  b q,l (tj ) + l D2 X b q,l min Xi (tj ) − X i i bi ∈Hq n p X j=1 i=0 2 L2  , (106) where l is the smoothing parameter, Hq is the q-dimensional functional subspace spanned by the leading eigenvectors of the autocovariance operator C associated with its largest eigenvalues. Thus, smoothness and rank constrain is considered in the computation of the solution to the optimization problem (106). Such a solution is obtained by means of functional PCA. The following regularized empirical estimators of C and D are then considered, with inversion of C in the subspace Hq : n−1 X bq,l = 1 bi ⊗ X bi , C X n i=0 b q,l = D n−2 1 X b bi+1 . Xi ⊗ X n − 1 i=0 Thus, the regularized estimator of ρ is given by b−1 , b q,l C ρbq,l = D q,l and the predictor bnq,l = ρbq,l Xn−1 . X Due to computational cost limitations, in Table 7, the following statistics are evaluated to compare the performance of the two above-referred prediction methodologies: EM AE hbn = Xn p 2 1 X b hn (tj ) , Xn (tj ) − X n p j=1 35 (107) q,l EM AEX b = n p 2 1 X bnq,l (tj ) . Xn (tj ) − X p j=1 h n h (108) q,l EM AEX b EM AEXbn,1 EM AEXbn,2 −2 −2 8.85 (10) 8.99 (10) −2 8.43 (10) −2 8.69 (10) −2 −2 n n n −2 n1 = 750 8.57 (10) n2 = 1250 7.67 (10) n3 = 1750 7.15 (10) 7.12 (10) 8.05 (10) n4 = 2250 7.09 (10)−2 6.87 (10)−2 7.59 (10)−2 n5 = 2750 6.87 (10) n6 = 3250 6.52 (10) n7 = 3750 6.20 (10) n8 = 4250 −2 6.67 (10) −2 5.92 (10) −2 5.56 (10) 6.06 (10) −2 n9 = 4750 5.67 (10)−2 n10 = 5250 5.24 (10) n11 = 5750 5.01 (10) n12 = 6250 4.90 (10) n13 = 6750 4.87 (10) −2 −2 −2 7.31 (10) −2 7.28 (10) −2 7.13 (10) 5.32 (10) −2 7.06 (10) 5.25 (10)−2 6.47 (10)−2 −2 5.12 (10) −2 4.82 (10) −2 4.49 (10) −2 3.87 (10) −2 −2 −2 −2 −2 6.08 (10) −2 −2 5.75 (10) −2 5.33 (10) −2 4.97 (10) −2 −2 −2 h q,l Table 7: EM AEXbn,i , i = 1, 2, and EM AEX bn values (see (107) and (108), respectively), with q = 7, n based on N = 200 simulations, for δ1 = 2.4 and δ2 = 1.1, considering now the sample sizes nt = 750 + 500(t − 1), t = 1, . . . , 13, hn,1 = 0.1 and hn,2 = 0.3 It can be observed a similar performance of the kernel-based and penalized FAR(1) predictors, from smooth functional data, which is also comparable, considering one realization, to the performance obtained in Table 6, from the empirical eigenvectors. 5.2.4 Wavelet-based prediction for ARH(1) processes The approach presented in [4] is now studied. Specifically, wavelet-based regularization is applied to obtain smooth estimates of the sample paths. The projection onto the space VJ , generated by translations of the scaling function φJk , k = 0, . . . , 2J − 1, at level J, associated with a multiresolution analysis of H, is first considered. For a given primary resolution level j0 , with j0 < J, the following wavelet decomposition at J − j0 resolution levels can be computed for any projected curve ΦVJ Xi , in the space VJ , for i = 0, . . . , n − 1 : ΦVJ Xi = j0 2X −1 cij0 k φj0 k dijk ψjk , j=j0 k=0 k=0 cij0 k + j −1 J−1 X 2X = hΦVJ Xi , φj0 k iH , 36 dijk = hΦVJ Xi , ψjk iH . (109) For i = 0, . . . , n − 1, the following variational problem is solved to obtain the smooth estimate of the curve Xi : inf i f ∈H  ΦVJ Xi − f i 2 L2 + λ ΦVj⊥ f  ; f ∈H , 2 0 (110) where ΦVj⊥ denotes the orthogonal projection operator of H onto the orhogonal complement of Vj0 , and 0 for i = 0, 1 . . . n − 1, j0 2X −1 i f = j αij0 k φj0 k + −1 ∞ 2X X i βjk ψjk . j=j0 k=0 k=0 Using the equivalent sequence of norms of fractional Sobolev spaces of order s with s > 1/2, on a suitable interval (in our case, s = δ1 ), the minimization of (110) is equivalent to the optimization problem, for i = 0, . . . , n − 1, j0 2X −1 (αij0 k k=0 − cij0 k )2 + j j −1 J−1 X 2X (dijk − j=j0 k=0 i 2 βjk ) + −1 ∞ 2X X i 2 λ2js [βjk ] . (111) j=j0 k=0 The solution to (111) is given by, for i = 0, . . . , n − 1, i αd j0 k i βd j0 k i βd j0 k cij0 k , = k = 0, 1, . . . , 2j0 − 1, dijk = (1 + λ22sj ) = 0, (112) j = j0 , . . . , J − 1, k = 0, 1, . . . , 2j − 1, , j ≥ J, k = 0, 1, . . . , 2j − 1. (113) (114) In particular, in the subsequent computations, we have considered the following value of the smoothing parameter λ (see [3]): bM = λ    M M X X  σj2   Cj  j=1 j=1 n . The following smoothed data are then computed e bM = X i,λ removing the trend e an,λbM = 1 n n−1 X i=0 j0 2X −1 k=0 i φ αd j0 k j0 k + j −1 J−1 X 2X j=j0 k=0 i βd j0 k ψjk , e bM to obtain X i,λ e bM − e Yei,λbM = X an,λbM , i,λ 37 i = 0, . . . , n − 1, (115) for the computation of ρen,λbM (x) = ρen,λbM ,l (x) =  kn  X e kn D e bM C e−1 Π e kn (x) = Π ρen,λbM ,l (x)φeM l , x ∈ H, bM n,λ bM bM λ kn X j=1 e bM C n,λ = n,λ 1 n−1 n−2 X i=0 λ l=1 1 e bM Ye bM , hφeM j , xiH Yi,λ ,j i+1,λ ,l e Cn,λbM ,j n−1 1Xe Y bM ⊗ Yei,λbM , n i=0 i,λ (116) (117) E D E D e bM = C e bM φb bM . Table 8 displays the empirical truncated where Yei,λbM ,j = Yei,λbM , φbj,λbM , and C n,λ ,j n,λ j,λ approximation of Eke ρn,λbM (Xn−1 )−ρ(Xn−1 )kH , applying our approach and the above-described approach presented in [4], for kni = n1/αi , i = 1, 2, with α1 = 6, according to our Assumption A3, and α2 > 4δ1 , according to H4 : nCk4n → ∞ in p. 149 in [4]. In particular, we have considered δ1 = 2.4, and α2 = 10. From the results displayed in Table 8, it can be observed a similar performance for the two truncation rules implemented, and approaches compared, for the small sample sizes tested. A similar performance is also displayed by the approaches presented in [9], for such small sample sizes (see Table 7). n kn,1 O.A. [4] kn,2 O.A. [4] n1 = 750 3 0.0702 0.0911 1 0.0636 0.0589 n2 = 1250 3 0.0550 0.0873 2 0.0509 0.0429 n3 = 1750 3 0.0473 0.0803 2 0.0455 0.0394 n4 = 2250 3 0.0414 0.0795 2 0.0409 0.0377 n5 = 2750 3 0.0365 0.0734 2 0.0355 0.0349 n6 = 3250 3 0.0343 0.0719 2 0.0333 0.0307 n7 = 3750 3 0.0330 0.0675 2 0.0325 0.0293 n8 = 4250 4 0.0328 0.0672 2 0.0313 0.0286 n9 = 4750 4 0.0317 0.0664 2 0.0309 0.0256 n10 = 5250 4 0.0309 0.0636 2 0.0276 0.0229 n11 = 5750 4 0.0298 0.0598 2 0.0203 0.0196 n12 = 6250 4 0.0283 0.0583 2 0.0166 0.0153 n13 = 6750 4 0.0276 0.0555 2 0.0148 0.0137 Table 8: Truncated empirical values of Ekρ(Xn−1 ) − ρekn (Xn−1 )kH , with ρekn defined in equation (101), and of Eke ρn,λbM (Xn−1 ) − ρ(Xn−1 )kH ,, based on N = 200 simulations, for δ1 = 2.4 and δ2 = 1.1, bM , M = 50, and the corresponding considering the sample sizes nt = 750+500(t−1), t = 1, . . . , 13, using λ 1/αi kn,i = n , for α1 = 6 and α2 = 10. Here, O.A. means Our Approach and [4] means The approach presented in [4] 38 5.2.5 Non-diagonal autocorrelation operator The methodology proposed can still be applied in the case of a non-diagonal autocorrelation operator ρ, with respect to the autocovariance eigenvector system (Cj , j ≥ 1), with coefficients close to zero outside of the main diagonal. This fact is illustrated below in Table 9, where the Gaussian ARH(1) process generated has autocorrelation operator ρ, with coefficients ρj,h with respect to the basis (φj ⊗ φh , j, h ≥ 1) , given by ρ2j,j = ρ2j,j+a =   −δ2 λj (−∆)(a,b)     , λ1 (−∆)(a,b) − ǫ  0.01 , a = 1, 2, 3, 4, 5, 5a2 ρ2j+a,j = (118) 0.02 , a = 1, 2, 3, 4, 5, 5a2 (119) where ρ2j,h = 0 when h 6= j − 1, j, j + 1, and ρ2j,j+a = ρ2j+a,j = 0 when a ≥ 6. The coefficients of the auto-covariance operator Rε of the innovation process ε, with respect to the mentioned basis (φj ⊗ φh , j, h ≥ 1) , are defined as follows: 2 σj,j = 2 σj,j+a =  Cj 1 − ρ2j,j , (120) 0.015 , a = 1, 2, 3, 4, 5, 5a2 2 σj+a,j = 0.01 , a = 1, 2, 3, 4, 5, 5a2 (121) 2 2 2 where σj,h = 0 when h 6= j − 1, j, j + 1, and σj,j+a = σj+a,j = 0 when a ≥ 6. In Table 9, we have considered kn = ⌈n1/α ⌉, with α = 6, δ1 = 2.4, δ2 = 1.1 and time discretization step ht = 0.015. As expected, it can be observed a better performance of the approaches presented in [11] and [31] against our approach, in terms of the theoretical eigenvectors, for this non-diagonal autocorrelation operator case. Although the performance of our approach is not too bad for the announced case where the coefficients of ρ, with respect to (φj ⊗ φh , j, h ≥ 1), j 6= h, are close to zero. Hence, in the general non-diagonal autocorrelation operator case, our approach can be applied, when smoothing of data has been previously implemented, considering, for example, suitable penalized wavelet nonparametric regression (see, for instance, [3] and [4]), leading to a sparse, almost diagonal, representation of bn and D b n , in terms of suitable wavelet bases (see comments given in the next section). C 39 n kn Our approach Bosq (2000) n1 = 15000 4 0.5812 8.94 (10) n2 = 35000 5 0.5604 n3 = 55000 6 n4 = 75000 Guillas (2001) −2 0.1055 7.05 (10) −2 9.49 (10) 0.5480 6.67 (10)−2 9.14 (10)−2 6 0.5322 6.24 (10) n5 = 95000 6 0.5115 5.89 (10) n6 = 115000 6 0.4975 5.62 (10) n7 = 135000 7 0.4946 n8 = 155000 7 n9 = 175000 −2 −2 8.85 (10) −2 −2 8.47 (10) −2 8.04 (10) 5.57 (10) −2 7.66 (10) 0.4810 5.28 (10)−2 7.24 (10)−2 7 0.4735 5.01 (10) n10 = 195000 7 0.4608 4.90 (10) n11 = 215000 7 0.4424 4.69 (10) n12 = 235000 7 0.4250 n13 = 255000 7 n14 = 275000 −2 −2 −2 −2 6.78 (10) −2 −2 6.30 (10) −2 6.07 (10) 4.45 (10) −2 5.82 (10) 0.4106 4.25 (10)−2 5.54 (10)−2 8 0.4080 4.14 (10) n15 = 295000 8 0.3808 4.09 (10) n16 = 315000 8 0.3604 3.85 (10) n17 = 335000 8 0.3489 n18 = 355000 8 n19 = 375000 n20 = 395000 −2 −2 −2 −2 5.16 (10) −2 −2 4.81 (10) −2 4.53 (10) 3.56 (10) −2 4.29 (10) 0.3302 3.29 (10)−2 3.98 (10)−2 8 0.3204 2.90 (10) 8 0.3177 2.62 (10) −2 −2 −2 −2 3.75 (10) −2 −2 3.44 (10) −2 D D Table 9: Truncated empirical values of E ρ(Xn−1 ) − ρbN bN given in equations kn (Xn−1 ) H , for ρ kn (25)–(26) (third column), in equations (95)–(96) (fourth column), and in equations (97)–(98) (fifth column), from the non-diagonal data generated by equations (118)-(121), based on N = 200 (due to high-dimensionality) simulations, for δ1 = 2.4 and δ2 = 1.1, considering the sample sizes nt = 15000 + 20000(t − 1), t = 1, . . . , 20, and the corresponding kn = ⌈n1/α ⌉, α = 6 values (truncation criteria from Assumption A3), assuming that (φj , j ≥ 1) are known. The time discretization step is assumed as ht = 0.015. 6 Final comments. The present paper proves that, under Assumptions A1–A4, the formulated componentwise estimator ρbkn converges to ρ, in the space L2S(H) (Ω, A, P ). Its associated ARH(1) plug-in predictor converges, in the space L1H (Ω, A, P ), to the predictor E[Xn |Xn−1 ] = ρ(Xn−1 ). Therefore, both, componentwise functional parameter estimator and plug-in predictor, are consistent. In the numerical examples considered, for the special family of covariance operators studied, and for 40 the particular selection of kn made, the empirical truncated functional mean-square errors of ρbkn display a rate of convergence to zero faster than (1/n)3/4 , and slower than (1/n), for the large sample sizes bn present a decay studied. Furthermore, the empirical truncated functional mean absolute errors of X velocity faster than (1/n)3/8 (or, as displayed, faster than (1/n)3/9 ), and slower than (1/n)1/2 , for the large sample sizes studied. In practice, our approach can be implemented when the eigenvectors of the autocovariance operator C are known. In the situation pointed out in Remark 1, where ρ and C share the same eigenvector system, we refer to the case of physical phenomena, whose dynamics is described from a random initial condition satisfying a given stochastic differential equation, in the mean-square sense, with driven process defined by a white noise process in time (e.g., Gaussian white noise process). In this case, the system of eigenvectors of the auto-covariance operator is known, since coincide with the eigenvectors of the differential operator defining such a stochastic differential equation. Otherwise, as illustrated in the simulation study undertaken, they must be approximated by some empirical eigenvectors (e.g., the eigenvectors of the empirical autocovariance operator). We comment here some alternatives to obtain a suitable basis of empirical eigenvectors satisfying Assumptions A1-A2. Since wavelet bases are well suited for sparse representation of functions, recent work has considered combining them with sparsityinducing penalties, both for semiparametric regression ([57]), and for regression with functional or kernel predictors (see [57], [58] and [59], among others). The latter papers focused on l1 penalization, also known as the lasso (see [54]), in the wavelet domain. Alternatives to the lasso include the SCAD penalty (see [21]), and the adaptive lasso (see, for example, [60]). The l1 penalty in the elastic net criterion has the effect of shrinking small coefficients to zero. This can be interpreted as imposing a prior that favors a sparse estimate. The above mentioned smoothing techniques, based on wavelets, can be applied b1 , . . . X bn of the functional data X1 , . . . , Xn , whose empirical to obtain a smooth sparse representation X auto-covariance operator and cross-covariance operator n−1 X bi ⊗ X bi bn = 1 X C n i=0 bn = D n−2 1 X b bi+1 Xi ⊗ X n − 1 i=0 admits a diagonal representation in terms of wavelets. On the other hand, estimating a covariance matrix is an important task in applications where the number of variables is larger than the number of observations. In the literature, shrinkage approaches for estimating a high-dimensional covariance matrix are employed to circumvent the limitations of the sample covariance matrix. Particularly, a new family of nonparametric Stein-type shrinkage covariance estimators is proposed in [55] (see also references therein), whose members are written as a convex linear 41 combination of the sample covariance matrix and of a predefined invertible diagonal target matrix. These results can be applied to our framework, considering the shrinkage estimator of the auto-covariance C and cross-covariance operator D, with respect to a common suitable wavelet function basis, which can lead to a empirical diagonal representation of both operators, in terms of such wavelet basis. Acknowledgments This work has been supported in part by project MTM2015–71839–P (co-funded by Feder funds), of the DGI, MINECO, Spain. References [1] G. Aneiros-Pérez and P. Vieu, Semi-functional partial linear regression, Stat. Probab. Letters 76, 2006, pp. 1102–1110. [2] G. Aneiros-Pérez and P. Vieu, Nonparametric time series prediction: A semi-functional partial linear modeling, J. Multivariate Anal. 99, 2008, pp. 834–857. [3] C. Angelini, D. De Canditiis and F. Leblanc, Wavelet regression estimation in nonparametric mixed effect models, J. Multivariate Anal. 85, 2003, pp. 267–291. [4] A. Antoniadis and T. Sapatinas, Wavelet methods for continuous-time prediction using Hilbert-valued autoregressive processes, J. Multivariate Anal. 87, 2003, pp. 133–158. [5] M. S. Bartlett, On the theoretical specification and sampling properties of autocorrelated time series, Supplement to the Journal of the Royal Statistical Society 8, 1946, pp. 27–41. [6] N. Bensmain and T. Mourid, Estimateur ”sieve” de l’opérateur d’un processus ARH(1), C. R. Acad. Sci. Paris Sér. I Math. 332, 2001, pp. 1015–1018. [7] P. C. Besse and H. Cardot, Approximation spline de la prévision d’un processus fonctionnel autorégressif d’ordre 1, Canad. J. Statist. 24, 1996, pp. 467–487. [8] P. C. Besse, H. Cardot and F. Ferraty, Simultaneous non-parametric regressions of unbalanced longitudinal data, Comp. Stat. Data Anal. 24, 1997, pp. 255–270. [9] P. C. Besse, H. Cardot and D. B. Stephenson, Autoregressive Forecasting of Some Functional Climatic Variations, Scand. J. Statist. 27, 2000, pp. 673–687. [10] E. G. Bongiorno, A. Goia, E. Salinelli and P. Vieu, Contributions in infinite-dimensional statistics and related topics, Societá Editrice Esculapio, Bologna 2014. 42 [11] D. Bosq, Linear Processes in Function Spaces, Springer-Verlag, New York 2000. [12] D. Bosq, General linear processes in Hilbert spaces and prediction, J. Stat. Plann. Inference 137, 2007, pp. 879–894. [13] D. Bosq and D. Blanke, Inference and Predictions in Large Dimensions, John Wiley, Chichester 2007. [14] J. Chen, D. Lin and H. Hochner, Semiparametric maximum likelihood methods for analyzing genetic and environmental effects with case-control motherchild pair data, Biometrics 68, 2012, pp. 869–877. [15] A. Cuevas, A partial overview of the theory of statistics with functional data, J. Stat. Plann. Infer. 147, 2014, pp. 1–23. [16] J. Cugliari, Prévision non paramétrique de processus à valeurs fonctionnelles. Application à la consommation d’électricité, PhD thesis, University of Paris-Sud 11, Paris 2011. [17] J. Damon and S. Guillas, Estimation and simulation of autoregressive processes with exogenous variables, Stat. Inference 8, 2005, pp. 185–204. [18] R. Dautray and J. -L. Lions, Mathematical Analysis and Numerical Methods for Science and Technology Volume 3: Spectral Theory and Applications, Springer, New York 1990. [19] J. Dedecker and F. Merlevede, The conditional central limit theorem in Hilbert spaces, Stochastic processes and their applications 108, 2003, pp. 229–262. [20] H. Dehling and O. S. Sharipov, Estimation of mean and covariance operator for Banach space valued autoregressive processes with dependent innovations, Stat. Inference Stoch. Process. 8, 2005, pp. 137–149. [21] J. Fan and R. Li, Variable selection via nonconcave penalized likelihood and its oracle properties, J. Amer. Statist. Assoc. 96, 2001, pp. 1348–1360. [22] M. Febrero-Bande and W. Gonzalez-Manteiga, Generalized additive models for functional data, Test 22, 2013, pp. 278–292. [23] F. Ferraty, A. Goia, E. Salinelli and P. Vieu, Functional projection pursuit regression, Test 22, 2013, pp. 293–320. [24] F. Ferraty, I. V. Keilegom and P. Vieu, Regression when both response and predictor are functions, J. Multivariate Anal. 109, 2012, pp. 10–28. [25] F. Ferraty and P. Vieu, Nonparametric functional data analysis. Springer Series in Statistics, Springer, New York 2006. 43 [26] F. Ferraty and P. Vieu, Additive prediction and boosting for functional data, Comp. Stat. Data Anal. 53, 2009, pp. 1400–1413. [27] D. S. Grebenkov and B. T. Nguyen, Geometrical structure of Laplacian eigenfunctions, SIAM Review 55, 2013, pp. 601–667. [28] U. Grenander, Abstract Inference, Wiley, New York 1981. [29] A. Goia and P. Vieu, An introduction to recent advances in high/infinite dimensional statistics, J. Multivariate Anal. 146, 2016, pp. 1–6. [30] S. Guillas, Noncausality and functional discretization, limit theorems for an ARHX(1) process, C. R. Acad. Sci. Paris Sér. I Math. 331, 2000, pp. 91–94. [31] S. Guillas, Rates of convergence of autocorrelation estimates for autoregressive Hilbertian processes, Stat. Prob. Lett. 55, 2001, pp. 281–291. [32] J. Gurland, Quadratic forms in normally distributed random variables, Sankhya 17, 1956, pp. 37–50. [33] J. D. Hamilton, Time Series Analysis, Princeton University Press, Princenton 1994. [34] L. Horváth and P. Kokoszka, Inference for Functional Data with Applications, in: Springer Series in Statistics, Springer, New York 2012. [35] T. Hsing and R. Eubank, Theoretical Foundations of Functional Data Analysis, with an Introduction to Linear Operators, Wiley Series in Probability and Statistics, John Wiley & Sons, United Kingdom 2015. [36] V. Kargin and A. Onatski, Curve forecasting by functional autoregression, J. Multivariate Anal. 99, 2008, pp. 2508–2526. [37] S. Mallat, A Wavelet Tour of Signal Processing: The Sparse Way, 3rd ed. Academic Press, Burlington, MA 2009. [38] J. -M. Marion and B. Pumo, Comparison of ARH(1) and ARHD(1) models on physiological data, Ann. I.S.U.P. 48, 2004, pp. 29–38. [39] A. Mas, Normalité asymptotique de l’estimateur empirique de l’opérateur d’autocorrélation d’un processus ARH(1), C. R. Acad. Sci. Paris Sér. I Math. 329, 1999, pp. 899–902. [40] A. Mas, Consistance du prédicteur dans le modèle ARH(1): le cas compact, Ann. I.S.U.P. 48, 2004, pp. 39–48. 44 [41] A. Mas, Weak-convergence in the functional autoregressive model, J. Multivariate Anal. 98, 2007, pp. 1231–1261. [42] A. Mas and L. Menneteau, Large and moderate deviations for infinite dimensional autoregressive processes, J. Multivariate Anal. 87, 2003, pp. 241–260. [43] L. Menneteau, Some laws of the iterated logarithm in Hilbertian autoregressive models, J. Multivariate Anal. 92, 2005, pp. 405–425. [44] F. Merlevede, Lois des grands nombres et loi du logarithme itéré compacte pour des processus linéaires à valeurs dans un espace de Banach de type 2, C. R. Acad. Sci. Paris Sér. I Math. 323, 1996, pp. 521–524. [45] F. Merlevede, Résultats de convergence presque sûre pour l’estimation et la prévision des processus linéaires hilbertiens, C. R. Acad. Sci. Paris Sér. I Math. 324, 1997, pp. 573–576. [46] T. Mourid, Processus autorégressifs banachiques d’ordre supérieur, C. R. Acad. Sci. Paris Sér. I Math. 317, 1993, pp. 1167–1172. [47] T. Mourid, Processus autorégressifs hilbertiens à coefficients aléatoires, Ann. I.S.U.P. 48, 2004, pp. 79–85. [48] G. P. Nason, Wavelet methods in Statistics with R, Springer 2008. [49] R. T. Ogden, Essential Wavelets for Statistical Applications and Data Analysis, Birkhauser, Boston, MA 1997. [50] B. Pumo, Estimation et prévision de processus autorégressifs fonctionnels, PhD thesis, University of Paris 6, Paris 1992. [51] B. Pumo, Prediction of continuous time processes by C[0,1] -valued autoregressive process, Stat. Inference Stoch. Process. 1, 1998, pp. 297–309. [52] J. O. Ramsay and B. W. Silverman, Functional data analysis, second ed. Springer Series in Statistics. Springer, New York 2005. [53] M. D. Ruiz-Medina, Spatial functional prediction from spatial autorregresive Hilbertian processes, Environmetrics 23, 2012, pp. 119–128. [54] R. Tibshirani, Regression shrinkage and selection via the lasso, J. Roy. Statist. Soc. Ser. B, 58, pp. 267–288. [55] A. Touloumis, Nonparametric Stein-type Shrinkage Covariance Matrix Estimators in HighDimensional Settings, Comp. Stat. Data Anal. 83, 2015, pp. 251–261. 45 [56] B. Vidakovic, Nonlinear wavelet shrinkage with Bayes rules and Bayes factors, J. Amer. Statist. Assoc. 93, 1998, pp. 173–179. [57] M. P. Wand and J. T. Ormerod, Penalized wavelets: Embedding wavelets into semiparametric regression, Electron. J. Stat. 5, 2011, pp. 1654–1717. [58] Y. Zhao, H. Chen and R. T. Ogden, Wavelet-based weighted LASSO and screening approaches in functional linear regression, J. Comput. Graph. Statist., 2015, to appear. [59] Y. Zhao, R. T. Ogden and P. T. Reiss, Wavelet-based lasso in functional linear regression, J. Comput. Graph. Statist. 21, 2012, pp. 600–617. [60] H. Zou, The adaptive lasso and its oracle properties, J. Amer. Statist. Assoc. 101, 2006, pp. 1418– 1429. 46
10math.ST
arXiv:cs/0603028v1 [cs.PL] 8 Mar 2006 On the tree-transformation power of XSLT Wim Janssen Alexandr Korlyukovd Jan Van den Bussche∗ Abstract XSLT is a standard rule-based programming language for expressing transformations of XML data. The language is currently in transition from version 1.0 to 2.0. In order to understand the computational consequences of this transition, we restrict XSLT to its pure tree-transformation capabilities. Under this focus, we observe that XSLT 1.0 was not yet a computationally complete tree-transformation language: every 1.0 program can be implemented in exponential time. A crucial new feature of version 2.0, however, which allows node sets over temporary trees, yields completeness. We provide a formal operational semantics for XSLT programs, and establish confluence for this semantics. 1 Introduction XSLT is a powerful rule-based programming language, relatively widely used, for expressing transformations of XML data, and is developed by the W3C (World Wide Web Consortium) [2, 8, 17]. An XSLT program is run on an XML document as input, and produces another XML document as output. (XSLT programs are actually called “stylesheets”, as one of their main uses is to produce stylised renderings of the input data, but we will continue to call them programs here.) The language is actually in a transition period: the current standard, version 1.0, is being replaced by version 2.0. It is important to understand what the new features of 2.0 really add. In the present paper, we focus on the tree-transformation capabilities of XSLT. Indeed, XML documents are essentially ordered, node-labeled trees. ∗ Wim Janssen and Jan Van den Bussche are with the University of Hasselt, Belgium. Alexandr Korlyukov, who was with Grodno State University, Belarus, sadly passed away shortly after we agreed to write a joint paper. 1 From the perspective of tree-transformation capabilities, the most important new feature is that of “node sets over temporary trees”. We will show that this feature turns XSLT into a computationally complete treetransformation language. Indeed, as we will also show, XSLT 1.0 was not yet complete in this sense. Specifically, any 1.0 program can be implemented within exponential time in the worst case. Some programs actually express PSPACE-complete problems, because we will show that any linear-space turing machine can be simulated by an XSLT 1.0 program. To put our results in context, we note that the designers of XSLT will most probably regard the incompleteness of their language as a feature, rather than a defect. Indeed, in the requirements document for 2.0, turning XSLT into a general-purpose programming language is explicitly stated as a “non-goal” [3]. In that respect, our result on the completeness of 2.0 exposes (albeit in a narrow sense) a failure to meet the requirements! At this point we should be a little clearer on what we mean by “focusing on the tree-transformation capabilities of XSLT”. As already mentioned, XML documents are essentially trees where the nodes are labeled by arbitrary strings. We make abstraction of this string content by regarding the node labels as coming from some finite alphabet. Accordingly, we strip XSLT of its string-manipulation functions, and restrict its arithmetic to arbitrary polynomial-time functions on counters, i.e., integers in the range {1, 2, . . . , n} with n the number of nodes in the input tree. It is, incidentally, quite easy to see that XSLT 1.0 without these restrictions can express all computable functions on strings (or integers). Indeed, rules in XSLT can be called recursively, and we all know that arbitrary recursion over the strings or the integers gives us completeness. We will provide a formal operational semantics for the substantial fragment of XSLT discussed in this paper. A formal semantics has not been available, although the W3C specifications represent a fine effort in defining it informally. Of course we have tried to make our formalisation faithful to those specifications. Our semantics does not impose an order on operations when there is no need to, and as a result the resulting transition relation is non-deterministic. We establish, however, a confluence property, so that any two terminating runs on the same input yield the same final result. Confluence was not yet proven rigorously for XSLT, and can help in providing a formal justification for alternative processing strategies that XSLT implementations may follow for the sake of optimisation. 2 a b c a c b Figure 1: A data tree. 2 2.1 Data model Data trees Let Σ be a finite alphabet, including the special label doc. By a data tree we simply mean a finite ordered tree, in which the nodes are labeled by elements of Σ. Up to isomorphism, we can describe a data tree t by a string string (t) over the alphabet Σ extended with the two symbols { and }: if the root of t is labeled a and its sequence of top-level subtrees is t1 , . . . , tk , then string (t) = a{string (t1 ) . . . string(tk )} Thus, for the data tree shown in Figure 1, the string representation equals a{b{}c{a{}b{}}c{}}. A data forest is a finite sequence of data trees. Forests arise naturally in XSLT, and for uniformity reasons we need to be able to present them as data trees. This can easily be done as follows: Definition 1 (maketree). Let F be a data forest. Then maketree(F ) is the data tree obtained by affixing a root node on top of F , and labeling this root node with doc.1 2.2 Stores and values Let T be a supply of tree variables, including the special tree variable Input. We define: 1 The root node added by maketree models what is called the “document root” in the XPath data model [6], although we do not model it entirely faithfully, as we do not formally distinguish “document nodes” from “element nodes”. This is only for simplicity; it is no problem to incorporate this distinction in our formalism, and our technical results do not depend on our simplification. 3 Definition 2. A store is a finite set S of pairs of the form (x, t), where x ∈ T and t is a data tree, such that (1) Input occurs in S; (2) no tree variable occurs twice in S; and (3) all data trees occurring in S have disjoint sets of nodes. The tree assigned to Input is called the input tree; the other trees are called the temporary trees. Definition 3. A value over S is a finite sequence consisting of nodes from trees in S, and counters over S. Here, a counter over S is an integer in the range {1, 2, . . . , n}, where n is the total number of nodes in S. Values as defined above formalise the kind of values that can be returned by XPath expressions. XPath [1, 5] is a language that is used as a sublanguage in XSLT for the purpose of selecting nodes from trees. But XPath expressions can also return numbers, which is useful as an aid in making node selections (e.g., the i-th child of a node, or the i-th node of the tree in preorder). We limit these numbers to counters, in order to concentrate on pure tree transformations. 3 XPath abstraction Since the language XPath is already well understood [27, 13, 7, 14], and its study in itself is not our focus, we will work with an abstraction of XPath, which we denote by X . For our purposes it will suffice to divide the X -expressions in only two different types, which we denote by nodes and mixed. A value is of type nodes if it consists exclusively of nodes; otherwise it is of type mixed. In order to define the semantics of X , we need some definitions, which reflect those from the XPath specification. Let V be a supply of value variables, disjoint from T . Definition 4. An environment over S is a finite set E of pairs of the form (x, v), where x ∈ V and v is a value over S, such that no value variable occurs twice in E. Definition 5. A context triple over S is a triple (z, i, k) where z is a node from S or a counter over S, and i and k are counters over S such that i 6 k. We call z the context item, i the context position, and k the context size. Definition 6. A context is a triple (S, E, c) where S is a store, E is an environment over S, and c is a context triple over S. 4 If we denote the universe of all possible contexts by Contexts, the semantics of X is now given by a partial function eval on X × Contexts, such that whenever defined, eval (e, C) is a value over C’s store, and this value has the same type as e. Remark 3.1. A static type system, based on XML Schema [4, 25], can be put on contexts to ensure definedness of expressions [7], but we omit that as safety is not the focus of the present paper. In general we do not assume much from X , except for the availability of the following basic expressions, also present in real XPath: ˆ An expression ‘/*’, such that eval (/*, C) equals the root node of the input tree in C’s store. ˆ An expression ‘child::*’, such that eval (child::*, C) is defined whenever C’s context item is a node n, and then equals the list of children of n. 4 Syntax In this section, we define the syntax of a sizeable fragment of XSLT 2.0. The reader familiar with XSLT will notice that we have simplified and cleaned up the language in a few places. These modifications are only for the sake of simplicity of exposition, and our technical results do not depend on them. We discuss our deviations from the real language further in Section 5.5. Also, the concrete syntax of real XSLT is XML-based and rather unwieldy. For the sake of presentation, we therefore give a syntax of our own, which is non-XML, but otherwise follows the same lines as the real syntax. The grammar is shown in Figure 2. The only typing condition we need is that in an apply-statement or in a vcopy-statement, expr must be of type nodes. Also, no two different rules can have the same name, and the name in a call-statement must be the name of some rule. We will often identify a template M with its syntax tree. This tree consists of all occurrences of statements in M and represents how they follow each other and how they are nested in each other; we omit the formal definition. Observe that only cons-, foreach-, tree-, and if-statements can have children. Note also that, since a template is a sequence of statements, the syntax “tree” is actually a forest, i.e., a sequence of trees, but we will still call it a tree. 5 Program → Rule* Rule → template name match expr (mode name)? { Template } Template → Statement* Statement → cons label { Template } | apply expr (mode name)? | call name | foreach expr { Template } | val value variable expr | tree tree variable { Template } | vcopy expr | tcopy tree variable | if expr { Template } else { Template } Figure 2: Our syntax. The terminal symbol expr stands for an X expression; label stands for an element of our alphabet Σ; value variable and tree variable stand for elements of V and T , respectively; and name is self-explanatory. As usual we use * to denote repetition, ? to denote optionality, and use ( and ) for lexical grouping. Variable definitions happen through val- and tree-statements. We will need the notion of a statement being in the scope of some variable definition; this is defined in the standard way as follows. Definition 7. Let M be a template, and let S1 and S2 be two statements occurring in M . We say that S2 is in the scope of S1 if S2 is a right sibling of S1 in the syntax tree of M , or a descendant of such a right sibling. An illustration is in Figure 3. One final definition: Definition 8. Template M ′ is called a subtemplate of template M if M ′ consists of a sequence of consecutive sibling statements occurring in M . 5 Operational semantics Fix a program P and a data tree t. We will describe the semantics of P on input t as a rewrite relation ⇒ among configurations. Definition 9. A configuration consists of a template M together with a partial function that assigns a context to some of the statements of M (more 6 Figure 3: Depiction of a syntax tree. The nodes in the scope of the black node are those that are striped. precisely, the nodes of its syntax tree). The statements that have a context are called active; we require that the descendants of an inactive node are inactive too. Cons-statements are never active. We use the following notation concerning configurations: ˆ S ✁ γ denotes that S is a statement occurring in the template of configuration γ. ˆ If S ✁ γ, then γ(S) = C denotes that S is active in γ, having context C. ˆ If M is a subtemplate of a configuration γ, then M itself can be taken as a configuration by inheriting all the context assignments done by γ. We call such a configuration a subconfiguration. ˆ If M is a subconfiguration of γ, and γ ′ is another configuration, then γ(M ← γ ′ ) denotes the configuration obtained from γ by replacing M by γ ′ . The initial configuration is defined as follows. Definition 10. 1. The initial context equals  {(Input, t)}, ∅, (r, 1, 1) where r is the root of t. 2. The initial template equals the single statement ‘apply /*’. 3. The initial configuration consists of the initial template, whose single statement is assigned the initial context. 7 ˆ S = if e { Mtrue } else { Mfalse } ✁ γ γ(S) = C eval (e, C) 6= ∅ if γ ⇒ γ(S ← Mtrue ) ˆ S = if e { Mtrue } else { Mfalse } ✁ γ γ(S) = C eval (e, C) = ∅ if γ ⇒ γ(S ← Mfalse) Figure 4: Semantics of if-statements; ∅ denotes the empty sequence. The goal will be to rewrite the initial configuration into a terminal template; this is a configuration consisting exclusively of cons-statements. Observe that terminal templates can be viewed as data forests; indeed, simply by removing the cons’s from a terminal template, we obtain the string representation of a data forest. For the rewrite relation ⇒ we are going to define, terminal configurations will be normal forms, i.e., cannot be rewritten further. If, for two configurations γ0 and γ1 , we have γ0 ⇒ · · · ⇒ γ1 and γ1 is a normal form, we denote that by γ0 ⇒! γ1 . The relation ⇒ will be defined in such a way that if γ0 is the initial configuration and γ0 ⇒! γ1 , then γ1 will be terminal. Moreover, we will prove in Theorem 1 that each configuration γ0 has at most one such normal form γ1 . We thus define: Definition 11. Given P and t, let γ0 be the initial configuration and let γ0 ⇒! γ1 . Then the final result tree of applying P to t is defined to be maketree(γ1 ). In the above definition, we can indeed apply maketree, defined on data forests (Definition 1), to γ1 , since γ1 is terminal and we just observed that terminal templates describe forests. Note that the final result tree is only determined up to isomorphism. 5.1 If-statements If-statements are the only ones that generate control flow, so we treat them if by a separate rewrite relation ⇒, defined by the semantic rules shown in Figure 4. 8 if It is not difficult to show that ⇒ is terminating and locally confluent, whence confluent, so that every configuration has a unique normal form if w.r.t. ⇒ [26]. This normal form no longer contains any active if-statements. (Quite obviously, the most efficient way to get to this normal form is to work if out the if-statements top-down.) We write γ ⇒ ! γ ′ to denote that γ ′ is the if normal form of γ w.r.t. ⇒. Remark 5.1. Our main rewrite relation ⇒ is not terminating in general. The reason why we treat if-statements separately is to avoid nonsensical rewritings such as where we execute a non-terminating statement in the else-branch of an if-statement whose test evaluates to true. 5.2 Apply-, call-, and foreach-statements For the semantics of apply-statements, we need the following definitions. Definition 12 (ruletoapply). Let C be a context, let n be a node, and let m be a name. Then ruletoapply (C, n) (respectively, ruletoapply (C, n, m)) equals the template belonging to the first rule in P (respectively, with mode name equal to m) whose expr satisfies n ∈ eval (expr, C). If no such rule exists, both ruletoapply (C, n) and ruletoapply (C, n, m) default to the single-statement template ‘apply child::*’. Definition 13 (init). Let M be a template, and let C be a context. Then init(M, C) equals the configuration obtained from M by assigning context C to every statement in M , except for all statements in the scope of any variable definition, and all statements that are below a foreach-statement; all those statements remain inactive. We are now ready for the semantic rule for apply-statements, shown in Figure 5. We omit the rule for an apply-statement with a mode m: the only difference with the rule shown is that we use ruletoapply (ni , C, m). The semantic rule for foreach-statements is very similar to that for applystatements, and is also shown in Figure 5. For call-statements, we need the following definition. Definition 14 (rulewithname). For any name, let rulewithname (name) denote the template of the rule in P with that name. The semantic rule for a call-statement is then again shown in Figure 5. 9 ˆ S = apply e ✁ γ γ(S) = C = (S, E, c) eval (e, C) = (n1 , . . . , nk ) ruletoapply (ni , C) = Mi for i = 1, . . . , k init(Mi , (S, E, (ni , i, k))) = γi for i = 1, . . . , k if γ(S ← γ1 . . . γk ) ⇒ ! γ ′ γ ⇒ γ′ ˆ S = foreach e { M } ✁ γ γ(S) = C = (S, E, c) eval (e, C) = (z1 , . . . , zk ) init(M, (S, E, (zi , i, k))) = γi γ(S ← γ1 . . . γk ) γ ⇒ γ′ if ⇒! for i = 1, . . . , k γ′ ˆ S = call name ✁ γ γ(S) = C rulewithname (name) = M init(M, C) = γ1 if γ(S ← γ1 ) ⇒ ! γ ′ γ ⇒ γ′ Figure 5: Semantics of apply-, call-, and foreach-statements. 10 ˆ S = val x e ✁ γ γ(S) = C C(x : eval(e, C)) = C ′ updateset (γ, S) = M init(M, C ′ ) = γ1 if γ(SM ← γ1 ) ⇒ ! γ ′ γ ⇒ γ′ ˆ S = tree y { M } ✁ γ M is terminal γ(S) = C C(y : maketree (M )) = C ′ updateset (γ, S) = M ′ init(M ′ , C ′ ) = γ3 if γ(SM ′ ← γ3 ) ⇒ ! γ ′ γ ⇒ γ′ Figure 6: Semantics of variable definitions. 5.3 Variable definitions For a context C = (S, E, c), a value variable x, a value v, a tree variable y, and a data tree t, we denote by ˆ C(x : v) the context obtained from C by updating E with the pair (x, v); and by ˆ C(y : t) the context obtained from C by updating S with the pair (y, t). We also define: Definition 15 (updateset). Let γ be a configuration and let S ✁ γ. Let M be the template underlying γ. Let S1 , . . . , Sk be the right siblings of S in M , in that order. Let j be the smallest index for which Sj is active in γ; if all the Si are inactive, put j = k + 1. Then the template S1 . . . Sj−1 is denoted by updateset (γ, S). If j = 1 then this is the empty template. We are now ready for the semantic rules for variable definitions, shown in Figure 6. 11 c n4 a b n1 c n2 c a b a n3 b  ttemp forest ((n4 , n1 , n2 , n3 , n1 ), S) = cons cons cons cons cons c b c a b { cons a {} cons b {} } {} { cons a {} cons b {} } {} {} Figure 7: Illustration of Definitions 16 and 17. 5.4 Copy-statements The following definitions are illustrated in Figure 7. Definition 16 (forest). Let S be a store, and let (n1 , . . . , nk ) be a sequence of nodes from S. For i = 1, . . . , k, let ti be the data subtree rooted at ni . Then forest ((n1 , . . . , nk ), S) equals the data forest (t1 , . . . , tn ). Definition 17 (ttemp). Let F be a data forest. Then ttemp(F ) equals the terminal template describing F . We also need: Definition 18 (choproot). Let t be a data tree with top-level subtrees t1 , . . . , tk , in that order. Then choproot (t) equals the data forest (t1 , . . . , tk ). The semantic rules for copy-statements are now shown in Figure 8. 5.5 Discussion The final result of applying P to t (Definition 11) may be undefined for two very different reasons. The first, fundamental, reason is that the rewriting may be nonterminating. The second reason is that the rewriting may abort because the evaluation of an X -expression is undefined, or the tree variable in a tcopy-statement is not defined in the store. This second reason can easily be avoided by a type system on X , as already mentioned in Remark 3.1, together with scoping rules to keep track of which variables are visible in the XSLT program and which variables are used in the X -expressions. Such 12 ˆ S = vcopy e ✁ γ γ(S) = C = (S, E, c) eval (e, C) = (n1 , . . . , nk )  ttemp forest((n1 , . . . , nk ), S) = M γ ⇒ γ(S ← M ) ˆ S = tcopy y ✁ γ γ(S) = (S, E, c) (y, t) ∈ S ttemp(choproot (t)) = M γ ⇒ γ(S ← M ) Figure 8: Semantics of copy-statements. scoping rules are entirely standard, and indeed are implemented in the XSLT processor SAXON [16]. In the same vein, we have simplified the parameter passing mechanism of XSLT, and have omitted the feature of global variables. On the other hand, our mechanism for choosing the rule to apply (Definition 12) is more powerful than the one provided by XSLT, as ours is context-dependent. It is actually easier to define that way. As already mentioned at the beginning of Section 4, none of our technical results depend on the modifications we have made. Finally, we note that the XSLT processor SAXON evaluates variable definitions lazily, whereas we simply evaluate them eagerly. Again, lazy evaluation could have been easily incorporated in our formalism. Some programs may terminate on some inputs lazily, while they do not terminate eagerly, but for programs that use all the variables they define there is no difference. 5.6 Confluence Recall that we call a rewrite relation confluent if, whenever we can rewrite a configuration γ1 to γ2 as well as to γ3 , then there exists γ4 such that we can further rewrite both γ2 and γ3 into γ4 . Confluence guarantees that all terminating runs from a common configuration also end in a common configuration [26]. Since, for our rewrite relation ⇒, either all runs on some input are nonterminating, or none is, the following theorem implies that the same final result of a program P on an input t, if defined at all, will be 13 obtained regardless of the order in which we process active statements. Theorem 1. Our rewrite relation ⇒ is confluent. Proof. The proof is a very easy application of a basic theorem of Rosen about subtree replacement systems [24]. A subtree replacement system R is a (typically infinite) set of pairs of the form φ → ψ, where φ and ψ are descriptions up to isomorphism of ordered, node-labeled trees, where the node labels come from some (again typically infinite) set V . Let us refer to such trees as V -trees. Such a system R naturally induces a rewrite system ⇒R on V -trees: we have t ⇒R t′ if there exists a node n of t and a pair φ → ψ in R such that the subtree t/n is isomorphic to φ, and t′ = t(n ← ψ). Here, we use the notation t/n for the subtree of t rooted at n, and the notation t(n ← ψ) for the tree obtained from t by replacing t/n by a fresh copy of ψ. Rosen’s theorem states that if R is “unequivocal” and “closed”, then ⇒R is confluent. “Unequivocal” means that for each φ there is at most one ψ such that φ → ψ is in R. The definition of R being “closed” is a bit more complicated. To state it, we need the notion of a residue map from φ to ψ. This is a mapping r from the nonroot nodes of φ to sets of nonroot nodes of ψ, such that for m ∈ r(n) the subtrees φ/n and ψ/m are isomorphic. Moreover, if n1 and n2 are independent (no descendants of each other), then all nodes in r(n1 ) must also be independent of all nodes in r(n2 ). Now R being closed means that we can assign a residue map r[φ, ψ] to every φ → ψ in R in such a way that for any φ0 → ψ0 in R, and any node n of φ0 , if there exists a pair φ0 /n → ψ in R, then the pair φ0 (n ← ψ) → ψ0 (r[φ0 , ψ0 ](n) ← ψ) is also in R. Denoting the latter pair by φ1 → ψ1 , we must moreover have for each node p of φ0 that is independent of n, that r[φ1 , ψ1 ](p) = r[φ0 , ψ0 ](p). To apply Rosen’s theorem, we view configurations (Definition 9) as V trees, where V = Statements ∪ (Statements × Contexts). Here, Statements is the set of all possible syntactic forms of statements. So, given a configuration, we take the syntax tree of the underlying template, and label every inactive node by its corresponding statement, and every active node by its corresponding statement and its context in the configuration. (Since templates are sequences, we actually get V -forests rather than V -trees, but that is a minor fuss.) Now consider the subtree replacement system R consisting of all pairs γ → γ ′ for which γ ⇒ γ ′ as defined by our semantics, where γ consists of a single statement S0 , and the active statement being processed to get γ ′ is a direct child of S0 . Since our semantics always substitutes siblings 14 for siblings, it is clear that ⇒R then coincides with our rewrite relation ⇒. Since the processing of every individual statement is always deterministic (up to isomorphism of trees), R as just defined is clearly unequivocal. We want to show that R is closed. Thereto, we define residue maps r[γ, γ ′ ] as follows. The case where γ → γ ′ is the processing of an apply- or call-statement, is depicted in Figure 9 (top). The node being processed is shown in black. The subtemplates to the left and right are left untouched. Referring to the notation used in Figure 5, the newly substituted subtemplate γnew is such if if that γ1 . . . γk ⇒ ! γnew (for apply) or γ1 ⇒ ! γnew (for call). Indeed, since we if apply ⇒ ! at the end of every processing step, γ itself does not contain any active if-statements. We define r = r[γ, γ ′ ] as follows: ˆ For nodes n in γleft or γright , we put r(n) := {n′ }, where n′ is the corresponding node in γ ′ . ˆ For the black node b, we put r(b) := ∅. The main condition for closedness is clearly satisfied, because statements can be processed independently. Note that the black node has no children, let alone active children, which allows us to put r(b) = ∅. The condition on p’s is also satisfied, because both r[φ0 , ψ0 ] and r[φ1 , ψ1 ] will set r(p) to {p′ }. The case where γ → γ ′ is the processing of a foreach-statement is depicted in Figure 9 (middle). This case is analogous to the previous one. The only difference is that the black node now has descendants (M in the figure). Because the init function (Definition 13) always leaves descendants of a foreach node inactive, however, the nodes in M are inactive at this time, and we can put r(n) := ∅ for all of them. The case where γ → γ ′ is the processing of a val-statement is depicted in Figure 9 (bottom). Since all nodes in the update set are inactive by definition (Definition 15), we can again put r(n) := ∅ for all nodes in the update set. The case of a tree-statement is similar; now the black node again has descendants, but again these are all inactive (they are all consstatements). The case where γ → γ ′ is the processing of a copy-statement, finally, is again analogous. 6 Computational completeness As defined in Definition 11, an XSLT program P expresses a partial function from data trees to data forests, where the output forest is represented by a 15 γ′ γ → γright γleft γnew γleft γright apply/call γ′ γ → foreach γleft γright γleft γnew γright M γ′ γ → γleft updateset γright γleft γnew γright val Figure 9: Illustration to the proof of Theorem 1. tree by affixing a root node labeled doc on top (Definition 1). The output is defined up to isomorphism only, and P does not distinguish between isomorphic inputs. This leads us to the following definition: Definition 19. A tree transformation is a partial function from data trees to data trees with root labeled doc, mapping isomorphic trees to isomorphic trees. Using the string representation of data trees defined in Section 2.1, we further define: Definition 20. A tree transformation f is called computable if the string function f˜: string (t) 7→ string(f (t)) is computable in the classical sense. Up to now, we have assumed from our XPath abstraction X only the availability of the expressions ‘/*’ and ‘child::*’. For our proof of the following theorem, we need to assume the availability of a few more very simple expressions, also present in real XPath: ˆ y/*, for any tree variable y, evaluates to the root of the tree assigned to y. 16 ˆ //* evaluates to the sequence of all nodes in the store (it does not matter in which order). ˆ child::*[1] evaluates to the first child of the context item (which should be a node). ˆ following-sibling::*[1] evaluates to the immediate right sibling of the context node, or the empty sequence if the context node has no right siblings. ˆ Increment, decrement, and test on counters: the constant expression ‘1’, and the expressions ‘x+1’, ‘x-1’, and ‘x=1’ for any value variable x, which should consist of a single counter. If x has the maximal counter value, then x+1 need not be defined, and if x has value 1, then x-1 need not be defined. The test x=1 yields any nonempty sequence for true and the empty sequence for false. ˆ name()=’a’, for any a ∈ Σ, returning any nonempty sequence if the label of the context node is a, and the empty sequence otherwise. ˆ () evaluates to the empty sequence. We establish: Theorem 2. Every computable tree transformation f can be realised by a program. Proof. We can naturally represent any string s over some finite alphabet as a flat data tree over the same alphabet. We denote this flat tree by flattree (s). Its root is labeled doc, and has k children, where k is the length of s, such that the labels of the children spell out the string s. There are no other nodes. The proof now consists of three parts: 1. Program the transformation t 7→ flattree (string (t)). 2. Show that every turing machine (working on strings) can be simulated by some program working on the flattree representation of strings. 3. Program the transformation flattree (string (t)) 7→ t. The theorem then follows by composing these three steps, where we simulate a turing machine for f˜ in step 2. Note that the composition of three programs can be written as a single program, using a temporary tree to 17 template tree2string match (//*) { cons a { } cons lbrace { } apply (child::*) cons rbrace { } } Figure 10: From t to flattree (string (t)). pass the intermediate results, and using modes to keep the rules from the different programs separate. The programs for steps 1 and 3 are shown in Figures 10 and 11. For simplicity, they are for an alphabet consisting of a single letter a, but it is obvious how to generalise the programs. The real XSLT versions are given in the Appendix. We point out that these programs are actually 1.0 programs, so it is only for step 2 of the proof that we need XSLT 2.0. For step 2, we can represent a configuration of a turing machine A by two temporary trees left and right. At each step, variable right holds (as a flat tree) the content of the tape starting at the head position and ending in the last tape cell; variable left holds the reverse of the tape portion left of the head position. To keep track of the current state of the machine, we use value variables q for each state q of A, such that at each step precisely one of these is nonempty. (This is why we need the X -expression ().) Changing the symbol under the head to an a amounts to assigning a new content to right by putting in cons a {}, followed by copies of the nodes in the current content of right, where we skip the first one. Moving the head a cell to the right amounts to assigning a new content to left by putting in a node labeled with the current symbol, followed by copies of the nodes in the current content of left. We also assign a new content to right in the now obvious way; if we were at the end of the tape we add a new node labeled blank. Moving the head a cell to the left is simulated analogously. The only X -expressions we need here are the ones we have assumed to be available. The simulation thus consists of repeatedly calling a big if-then-else that tests for the transition to be performed, and performs that transition. We may assume A is programmed in such a way that the final output is produced starting from a designated state. In this way we can build up the final output string in a fresh temporary tree and pass it to step 3. 18 template doc match (/*) { apply (child::*[1]) } template string2tree match (//*) { cons a { apply (following-sibling::*[1]) mode dochildren } val counter (1) call searchnextsibling } template dochildren match (//*) mode dochildren { if name()=’lbrace’ { apply (following-sibling::*[1]) mode dochildren } else { if name()=’a’ { call string2tree } else { } } } template searchnextsibling match (//*) mode search { if name()=’lbrace’ { val counter (counter + 1) apply (following-sibling::*[1]) mode search } else { if name()=’a’ { apply (following-sibling::*[1]) mode search } else { val counter (counter - 1) if counter = 1 { apply (following-sibling::*[1]) mode dochildren } else { apply (following-sibling::*[1]) mode search } } } } Figure 11: From flattree (string (t)) to t. 19 7 XSLT 1.0 In this section we will show that every XSLT 1.0 program can be implemented in exponential time, in sharp contrast to the computational completeness result of the previous section. A fundamental difference between XSLT 1.0 and 2.0 is that in 1.0, X expressions are “input-only”, defined as follows. Definition 21. 1. Let C = (S, E, (z, i, k)) be a context. Let the input tree in S be t. Then we call C input-only if every value appearing in E is already a value over the store {(Input, t)}, and also (z, i, k) is like that. 2. By Ĉ, we mean the context ({(Input, t)}, E, (z, i, k)). So, Ĉ equals C where we have removed all temporary trees. 3. Now an X -expression e is called input-only if for any input-only context C for which eval (e, C) is defined, we have eval (e, C) = eval (e, Ĉ), and this must be a value over C’s input tree only. In other words, input-only expressions are oblivious to the temporary trees in the store; they only see the input tree. We further define: Definition 22. An input-only X -expression e is called polynomial if for each input-only context C, the computation of eval (e, C) can be done in time polynomial in the size of C’s input tree. We now define: Definition 23. A program is called 1.0 if it only uses input-only, polynomial X -expressions. Essentially, 1.0 programs cannot do anything with temporary trees except copy them using tcopy statements. We note that real XPath 1.0 expressions are indeed input-only and polynomial; actually, real XPath 1.0 is much more restricted than that, but for our purpose we do not need to assume anything more. In order to establish an exponential upper bound on the time-complexity of 1.0 programs, we cannot use an explicit representation of the output tree. Indeed, 1.0 programs can produce result trees of size doubly exponential in the size of the input tree. For example, using subsets of input nodes, ordered lexicographically, as depth counters, we can produce a full binary 20 a a a b c b c b b a b b a b Figure 12: Left, a data tree, and right, a DAG representation of it. tree of depth 2n from an input tree with n nodes. Obviously a doubly exponentially long output could never be computed in singly exponential time. We therefore use a DAG representation of trees: an old and well-known trick [22] that is also used in tree transduction [18], and that has recently found new applications in XML [10]. Formally, a DAG representation is a collection G of trees, where trees in G can have special leafs which are not labeled, and from which a pointer departs to the root of another tree in G. On condition that the resulting pointer graph is acyclic, starting from a designated “root tree” in G we can naturally obtain a tree by unfolding along the pointers. An illustration is shown in Figure 12. We establish: Theorem 3. Let P be an 1.0 program. Then the following problem is solvO(1) time: able in exponential, i.e., 2n Input: a data tree t Output: a DAG representation of the final result tree of applying P to t, or a message signaling non-termination if P does not terminate on t. Proof. We will generate a DAG representation G by applying modified versions of the semantic rules from Section 5. We initialise G with all the subtrees of t. These trees have no pointers. Each tree that will be added to G will be a configuration, which still has to be developed further into a final data tree with pointers, using the same modified rules. Because we will have to point to the newly added configurations later, we identify each 21 added configuration by a pair (name, C) where name is the name of a template rule in P and C is a context. In the description below, whenever we say that we “add” a configuration to G, identified by some pair (name, C), we really mean that we add it unless a configuration identified by that same pair already exists in G. The modifications are now the following. 1. When executing an apply-statement, we do not directly insert copies of the templates belonging to the rules that must be applied (the γi ’s in Figure 5). Rather, we add, for i = 1, . . . , k, the configuration γi′ to if G, where γi ⇒ ! γi′ . We identify γi′ by the pair (name i , Ci ), with name i the name of the rule γi comes from, and Ci = (S, E, (ni , i, k)) using the notation of Figure 5. Moreover, in place of the apply-statement we insert a sequence of k pointer nodes pointing to (name 1 , C1 ), . . . , (name k , Ck ), respectively. 2. When executing a call-statement call name under context C, we again do not insert γ1 (compare Figure 5), but add the configuration γ1′ to if G, where γ1 ⇒ ! γ1′ , and identify it by the pair (name, C). We then replace the statement by a pointer node pointing to that pair. 3. By making template rules from the bodies of all foreach-statements in P , we may assume without loss of generality that the body of every foreach-statement is a single call-statement. A foreach-statement is then processed analogously to apply- and call-statements. 4. As we did with foreach-statements, we may assume that the body of each tree-statement is a single call-statement. When executing a treestatement, we may assume that the call-statement has already been turned into a pointer to some pair (name 0 , C0 ). We then assign that pair directly to y in the new context C ′ (compare Figure 6); we no longer apply maketree. So, in the modified kind of store we use, we assign name–context pairs, rather than fully specified temporary trees, to tree variables. 5. Correspondingly, when executing a statement tcopy y, we now directly turn it into a pointer to the pair assigned to y. 6. Finally, when executing a vcopy-statement, we do not insert the whole forest generated by (n1 , . . . , nk ) in the configuration (compare Figure 8), but merely insert a sequence of k pointers to the input subtrees rooted at n1 , . . . , nk , respectively. 22 We initiate the generation of G by starting with the initial configuration as always. Processing that configuration will add the first tree to G, which serves as the root tree of the DAG representation. When all trees in G have been fully developed into data trees with pointer nodes, the algorithm terminates. In case P does not terminate on t, however, that will never happen, and we need a way to detect nontermination. Thereto, recall that every context consists of an environment E and a context triple c on the one hand, and a store S on the other hand. Since all X -expressions used are input-only, and thus oblivious to the store-part of a context (except for the input tree, which does not change), we are in an infinite loop from the moment that there is a cycle in G’s pointer graph where we ignore the store-part of the contexts. More precisely, this happens when from a pointer node in a tree identified by (name, C1 ) we can follow pointers and reach a pointer to a pair (name, C2 ) with the same name and where C1 and C2 are equal in their (E, c)-parts. As soon as we detect such a cycle, we terminate the algorithm and signal nontermination. Note that thus the algorithm always terminates. Indeed, since only inputonly X -expressions are used, all contexts that appear in the computation are input-only, and there are only a finite number of possible (E, c)-parts of input-only configuration over a fixed input tree. Let us analyse the complexity of this algorithm. Since all X -expressions used are polynomial, there is a natural number K such that each value that appears in a context is at most nK long, where n equals the number of nodes in t. Each element of such a length-nK sequence is a node or a counter over t, K so there are at most (2n)n different values. There are a constant c1 number K of different value variables in P , so there are at most ((2n)n )c1 different environments. Likewise, the number of different context triples is (2n)3 , K′ K so, ignoring the stores, there are in total at most (2n)3 · (2n)c1 n 6 2n different contexts, for some natural number K ′ > K. With a constant c2 number of different template names in P , we get a maximal number of K′ c2 2n different configurations that can be added to G before the algorithm will surely terminate. It remains to see how long it takes to fully rewrite each of those configurations into a data tree with pointers. A configuration initially consists of at most a constant c3 number of statements. The evaluation of X -expressions, which are polynomial, takes at most c3 nK time in total. Processing an apply- or a foreach-statement takes at most c3 nK modifications to the configuration and to G; for the other statements this takes at most c3 such operations. Each such operation, however, involves the handling of con23 texts, whose stores can become quite large if treated naively. Indeed, treestatements assign a context to a tree variable, yielding a new context which may then again be assigned to a tree variable, and so on. To keep this under control, we do not copy the contexts literally, but number them consecutively in the order they are introduced in G. A map data structure keeps track of this numbering. The stores then consist of an at most constant c4 number of assignments of pairs (name, context number) to tree variables. ′ K′ As there are at most 2n different contexts, each number is at most nK bits long. Looking up whether a given context is already in G, and if so, ′ K′ finding its number, takes O(log 2n ) = O(nK ) time using a suitable map data structure. K′ We conclude that the processing of G takes a total time of c2 2n · O(1) ′ O(nK ) = 2n , as had to be proven. A legitimate question is whether the complexity bound given by Theorem 3 can still be improved. In this respect we can show that, even within the limits of real XSLT 1.0, any linear-space turing machine can be simulated by a 1.0 program. Note that some PSPACE-complete problems, such as QBF-SAT [21], are solvable in linear space. This shows that the time complexity upper bound of Theorem 3 cannot be improved without showing that PSPACE is properly included in EXPTIME (a famous open problem). The simulation gets as input a flat tree representing an input string, and uses the n child nodes to simulate the n tape cells. For each letter a of the tape alphabet, a value variable cell a holds the nodes representing the tape cells that have an a. A value variable head holds the node representing the cell seen by the machine’s head. The machine’s state is kept by additional value variables state q for each state q, such that state q is nonempty iff the machine is in state q. Writing a letter in a cell, moving the head left or right, or changing state, are accomplished by easy updates on the value-variables, which can be expressed by real XPath 1.0 expressions. Choosing the right transition is done by a big if-then-else statement. Successive transitions are performed by recursively applying the simulating template rule until a halting state is reached. Remark 7.1. A final remark is that our results imply that XSLT 1.0 is not closed under composition. Indeed, building up a tree of doubly exponential size (as we already remarked is possible in XSLT 1.0), followed by the building up of a tree of exponential size, amounts to building up a tree of triply exponential size. If that would be possible by a single program, then a DAG representation of a triply exponentially large tree would be com- 24 putable in singly exponential time. It is well known, however, that a DAG representation cannot be more than singly exponentially smaller than the tree it represents. Closure under composition is another sharp contrast between XSLT 1.0 and 2.0, as the latter is indeed closed under composition as already noted in the proof of Theorem 2. 8 Conclusions W3C recommendations such as the XSLT specifications are no Holy scriptures. Theoretical scrutinising of W3C work, which is what we have done here, can help in better understanding the possibilities and limitations of various newly proposed programming languages related to the Web, eventually leading to better proposals. A formalisation of the full XSLT 2.0 language, with all the dirty details both concerning the language itself as concerning the XPath 2.0 data model, is probably something that should be done. We believe our work gives a clear direction how this could be done. Note also that XSLT contains a lot of redundancies. For example, foreach-statements are eliminable, as are call-statements, and the match attribute of template rules. A formalisation such as ours can provide a rigorous foundation to prove such redundancies, or to prove correct various processing strategies or optimisation techniques XSLT implementations may use. A formal tree transformation model denoted by TL, in part inspired by XSLT, but still omitting many of its features, has already been studied by Maneth and his collaborators [9, 19]. The TL model can be compiled into the earlier formalism of “macro tree transducers” [12, 23]. It is certainly an interesting topic for further research to similarly translate our XSLT formalisation (even partially) into macro tree transducers, so that techniques already developed for these transducers can be applied. For example, under regular expression types [15] (known much earlier under the name of “recognisable tree languages”), exact automated typechecking is possible for compositions of macro tree transducers, using the method of “inverse type inference” [20]. This method has various other applications, such as deciding termination on all possible inputs [19]. Being able to apply this method to our XSLT 1.0 formalism would improve the analysis techniques of Dong and Bailey [11], which are not complete. 25 Acknowledgment We are indebted to Frank Neven for his initial participation in this research. References [1] XML path language (XPath) version 1.0. November 1999. W3C Recommendation, [2] XSL transformations (XSLT) version 1.0. November 1999. W3C Recommendation, [3] XSLT requirements version 2.0. W3C Working Draft, February 2001. [4] XML schema. W3C Recommendation, October 2004. [5] XML path language (XPath) version 2.0. W3C Working Draft, April 2005. [6] XQuery 1.0 and XPath 2.0 data model. W3C Working Draft, April 2005. [7] XQuery 1.0 and XPath 2.0 formal semantics. W3C Working Draft, June 2005. [8] XSL transformations (XSLT) version 2.0. W3C Working Draft, April 2005. [9] G.J. Bex, S. Maneth, and F. Neven. A formal model for an expressive fragment of XSLT. Information Systems, 27(1):21–39, 2002. [10] P. Buneman, M. Grohe, and C. Koch. Path queries on compressed XML. In J.C. Freytag, P.C. Lockemann, et al., editors, Proceedings 29th International Conference on Very Large Data Bases, pages 141– 152. Morgan Kaufmann, 2003. [11] C. Dong and J. Bailey. Static analysis of XSLT programs. In K.D. Schewe and H.E. Williams, editors, Database technologies—Proceedings ADC 2004, pages 151–160. Australian Computer Society, 2004. [12] J. Engelfriet and H. Vogler. Macro tree transducers. Journal of Computer and System Sciences, 31(1):71–146, 1985. 26 [13] G. Gottlob, C. Koch, and R. Pichler. XPath processing in a nutshell. SIGMOD Record, 32(2):21–27, 2003. [14] J. Hidders, J. Paredaens, R. Vercammen, et al. A light but formal introduction to XQuery. In Z. Bellahsène, T. Milo, M. Rys, et al., editors, Database and XML Technologies—Proceedings XSym, volume 3186 of Lecture Notes in Computer Science, pages 5–20. Springer, 2004. [15] H. Hosoya and B.C. Pierce. XDuce: A statically typed XML processing language. ACM Transactions on Internet Technology, 3(2):117–148, 2003. [16] M. Kay. SAXON: The http://saxon.sourceforge.net. XSLT and XQuery processor. [17] M. Kay. XSLT 2.0 Programmer’s Reference. Wrox, 3rd edition, 2004. [18] S. Maneth. The complexity of compositions of deterministic tree transducers. In M. Agrawal and A. Seth, editors, FST TCS 2002 Proceedings, volume 2556 of Lecture Notes in Computer Science, pages 265–276. Springer, 2002. [19] S. Maneth, A. Berlea, T. Perst, and H. Seidl. XML type checking with macro tree transducers. In Proceedings 24th ACM Symposium on Principles of Database Systems, pages 283–294. ACM Press, 2005. [20] T. Milo, D. Suciu, and V. Vianu. Typechecking for XML transformers. Journal of Computer and System Sciences, 66(1):66–97, 2003. [21] C.H. Papadimitriou. Computational Complexity. Addison-Wesley, 1994. [22] M.S. Paterson and M.N. Wegman. Linear unification. Journal of Computer and System Sciences, 16(2):158–167, 1978. [23] T. Perst and H. Seidl. Macro forest transducers. Information Processing Letters, 89(3):141–149, 2004. [24] B.K. Rosen. Tree-manipulating systems and Church-Rosser theorems. Journal of the ACM, 20(1):160–187, 1973. [25] J. Siméon and P. Wadler. The essence of XML. In Proceedings 30th ACM Symposium on Principles of Programming Languages, pages 1–13. ACM Press, 2003. [26] Terese. Term Rewriting Systems. Cambridge University Press, 2003. 27 [27] P. Wadler. A formal semantics of patterns in XSLT and XPath. Markup Languages: Theory and Practice, 2(2):183–202, 2000. 28 A A.1 Real XSLT programs Figure 10 in real XSLT <xsl:transform xmlns:xsl="http://www.w3.org/1999/XSL/Transform" version="1.0"> <xsl:template name="tree2string" match="//*"> <a/> <lbrace/> <xsl:apply-templates select="child::*"/> <rbrace/> </xsl:template> </xsl:transform> A.2 Figure 11 in real XSLT <xsl:transform xmlns:xsl="http://www.w3.org/1999/XSL/Transform" version="1.0"> <xsl:template match="/doc"> <xsl:apply-templates select="child::*[1]"/> </xsl:template> <xsl:template name="string2tree" match="/doc//*"> <a> <xsl:apply-templates select="following-sibling::*[1]" mode="dochildren"/> </a> <xsl:call-template name="searchnextsibling"> <xsl:with-param name="counter" select="1"/> </xsl:call-template> </xsl:template> <xsl:template match="//*" mode="dochildren"> 29 <xsl:if test="name()=’lbrace’"> <xsl:apply-templates select="following-sibling::*[1]" mode="dochildren"/> </xsl:if> <xsl:if test="name()=’a’"> <xsl:call-template name="string2tree"/> </xsl:if> </xsl:template> <xsl:template name="searchnextsibling" match="//*" mode="search"> <xsl:param name="counter"/> <xsl:if test="name()=’lbrace’"> <xsl:apply-templates select="following-sibling::*[1]" mode="search"> <xsl:with-param name="counter" select="$counter + 1"/> </xsl:apply-templates> </xsl:if> <xsl:if test="name()=’a’"> <xsl:apply-templates select="following-sibling::*[1]" mode="search"> <xsl:with-param name="counter" select="$counter"/> </xsl:apply-templates> </xsl:if> <xsl:if test="name()=’rbrace’"> <xsl:if test="$counter=2"> <xsl:apply-templates select="following-sibling::*[1]" mode="dochildren"/> </xsl:if> <xsl:if test="$counter>2"> <xsl:apply-templates select="following-sibling::*[1]" mode="search"> <xsl:with-param name="counter" select="$counter - 1"/> </xsl:apply-templates> </xsl:if> </xsl:if> </xsl:template> </xsl:transform> 30
2cs.AI
arXiv:1711.01090v1 [math.GR] 3 Nov 2017 FACTORIZATIONS OF ALMOST SIMPLE GROUPS WITH A FACTOR HAVING MANY NONSOLVABLE COMPOSITION FACTORS CAI HENG LI AND BINZHOU XIA Abstract. This paper classifies the factorizations of almost simple groups with a factor having at least two nonsolvable composition factors. This together with a previous classification result of the authors reduces the factorization problem of almost simple groups to the case where both factors have a unique nonsolvable composition factor. Key words: factorizations; almost simple groups MSC2010: 20D40, 20D06, 20D08 1. Introduction For a group G, an expression G = HK with subgroups H and K of G is called a factorization of G, where H and K are called factors. Factorizations of groups play an important role not only in group theory but also in other branches of mathematics such as Galois theory [7, 9] and graph theory [14, 17]. A group G is almost simple with socle L if L 6 G 6 Aut(L) for some finite nonabelian simple group L. To study factorizations of finite groups, it is crucial to study those of almost simple groups. In 1987, Hering, Liebeck and Saxl [10] classified factorizations of exceptional groups of Lie type. A factorization G = HK is called a maximal factorization of G if both H and K are maximal subgroups of G. In 1990, Liebeck, Praeger and Saxl published the landmark work [15] classifying maximal factorizations of almost simple groups. Furthermore, when the socle is an alternating group, all the factorizations of such a group were determined in [15, Theorem D]. Based on the maximal factorizations in [15, Theorem C], Giudici [8] in 2006 determined the factorizations of sporadic groups. In 2010, all factorizations of almost simple groups for which one factor is maximal and the intersection of the two factors is trivial have been determined [17] by Liebeck, Praeger and Saxl. In [14], factorizations of almost simple groups with a solvable factor are classified. As a counter part, we classify in this paper (Theorem 1.1) factorizations of almost simple groups with a factor having at least two nonsolvable composition factors (counted with multiplicity). These two classification results then reduce the factorization problem of almost simple groups to the case where both factors have a unique nonsolvable composition factor (Corollary 1.2). Theorem 1.1. Let G be an almost simple group with socle L. Suppose that G has a factorization G = HK, where H has at least two nonsolvable composition factors and K is core-free in G. Then either (a) or (b) below holds. (a) An 6 G 6 Sn with n > 10, and one of the following holds: (a.1) H is a transitive permutation group of degree n, and An−1 6 K 6 Sn−1 ; 1 2 LI AND XIA (a.2) n = 10, H is a transitive permutation group of degree 10 such that (A5 × A5 ).2 6 H 6 S5 ≀ S2 , and K = SL2 (8) or SL2 (8).3; (a.3) n = 12, A7 × A5 6 H 6 S7 × S5 , and K = M12 ; (a.4) n = 24, A19 × A5 6 H 6 S19 × S5 , and K = M24 . (b) L is a symplectic group or an orthogonal group of plus type, and the triple (L, H ∩ L, K ∩ L) lies in Table 1 or Table 2, respectively, where Q 6 2. Table 1. row L H ∩L (Sp2a (2f b ) × Sp2a (2f b )).R.2, Sp4ℓ (2f ), f ℓ > 2 ab = ℓ and R 6 b × b (G2 (2f ℓ ) × G2 (2f ℓ )).R.2, Sp12ℓ (2f ) R6ℓ×ℓ Sp4ℓ (4), ℓ > 2 Sp2 (4) × Sp4ℓ−2 (4) Sp4ℓ (4), ℓ > 2 (Sp2 (4) × Sp2ℓ (4)).P , P 6 2 Sp8ℓ (2) (Sp2 (4) × Sp2ℓ (4)).P , P 6 [8] Sp4 (2f ), f > 3 odd (Sp2 (2f ) × Sp2 (2f )).P , P 6 2 U × Sp4 (2f ), Sp6 (2f ), f > 2 U nonsolvable and U 6 Sp2 (2f ) Sp12 (4) Sp2 (4) × Sp10 (4) Sp12 (4) (Sp2 (4) × G2 (4)).P , P 6 2 Sp24 (2) (Sp2 (4) × G2 (4)).P , P 6 [8] K ∩L Ex. 1 f Ω− 4ℓ (2 ).Q 3.2 2 3 4 5 6 7 8 9 10 f Ω− 12ℓ (2 ).Q 3.3 Sp2ℓ (16).Q Ω− 4ℓ (4).Q Ω− 8ℓ (2).Q Sz(2f ) max 3.7 3.8 max G2 (2f ) 3.12 G2 (16).Q Ω− 12 (4).Q Ω− 24 (2).Q 3.14 3.15 3.17 Table 2. row L K ∩L Ex. 1 Ω4ℓ−1 (q) 3.5 2 3 4 5 6 H ∩L (U × PSp2ℓ (q)).P , U nonsolvable, PΩ+ 4ℓ (q), ℓ > 2, q > 4 U 6 PSp (q) and P 6 gcd(2, ℓ, q − 1) 2 Ω+ (Sp2 (4) × Sp2ℓ (4)).P , P 6 22 8ℓ (2) (Sp2 (4c ) × Sp2ℓ (16)).P , Ω+ 8ℓ (4) c = 1 or 2 and P 6 22 U × G2 (2f ), f ), f > 2 Ω+ (2 12 U nonsolvable and U 6 Sp2 (2f ) + Ω24 (2) (Sp2 (4) × G2 (4)).P , P 6 22 (Sp2 (4c ) × G2 (16)).P , Ω+ 24 (4) c = 1 or 2 and P 6 22 Sp8ℓ−2 (2) 3.10 Sp8ℓ−2 (4) 3.11 Sp10 (2f ) 3.13 Sp22 (4) 3.16 Sp22 (4) 3.18 Here are some remarks on Theorem 1.1. (i) For each factorization G = HK described in Theorem 1.1, K ∩ L is almost simple, and either H has exactly two nonsolvable composition factors or (a.1) occurs. (ii) It is easy to see that each case in part (a) of Theorem 1.1 gives rise to a factorization G = HK. Also, for each row of Table 1 and Table 2 there is an example for such factorization G = HK, which can be found in the last column: the label 3.k indicates that such an example is given in Example 3.k while the label max indicates that there exists a maximal factorization (see Lemma 3.1). FACTORIZATIONS OF ALMOST SIMPLE GROUPS 3 Corollary 1.2. Let G be an almost simple group. Suppose G = HK with core-free subgroups H and K of G. Then interchanging H and K if necessary, one of the following holds. (a) H is solvable, and (G, H, K) is described in [14, Theorem 1.1]. (b) H has at least two nonsolvable composition factors, and (G, H, K) is described in Theorem 1.1. (c) Both H and K have a unique nonsolvable composition factor. The proof of Theorem 1.1 is sketched in Section 4 and completed in Sections 5 and 6. In Section 3, we construct factorizations as described in Table 1 and Table 2. Acknowledgements. This work was supported by NSFC grants 11231008 and 11771200. The second author is very grateful to South University of Science and Technology for the financial support of his visit to the first author. 2. Strategy of proof Hereafter, all the groups are assumed to be finite. The group-theoretic notation in this paper mostly follows [15]. There are several equivalent conditions for a group factorization as in Lemma 2.1 below. Lemma 2.1. Let H and K be subgroups of G. Then the following are equivalent. (a) G = HK. (b) G = H x K y for any x, y ∈ G. (c) |H ∩ K||G| = |H||K|. (d) |G| 6 |H||K|/|H ∩ K|. (e) H acts transitively on [G : K] by right multiplication. (f) K acts transitively on [G : H] by right multiplication. Remark. Lemma 2.1 is easy to prove but plays a fundamental role in the study of group factorizations: (i) due to part (b) we will consider conjugacy classes of subgroups when studying factorizations of a group;. (ii) given a group G and its subgroups H and K, in order to inspect whether G = HK we only need to compute the orders of G, H, K and H ∩ K by part (c) or (d), which enables us to search factorizations of a group efficiently in Magma [3]. Here are more observations about group factorizations. Lemma 2.2. Let H and K be subgroups of G, and L be a normal subgroup of G. If G = HK, then we have the following divisibilities. (a) |G| divides |H||K|. (b) |G| divides |H ∩ L||K||G/L|. (c) |L| divides |H ∩ L||K|. (d) |L| divides |H ∩ L||K ∩ L||G/L|. Lemma 2.3. Let H, K and M be subgroups of G. If G = HK and H 6 M, then M = H(K ∩ M). 4 LI AND XIA Lemma 2.4. Let H, K and M be subgroups of G. Suppose H 6 M and M = H(K ∩ M). Then G = HK if and only if G = MK. To construct more factorizations from existing ones, we introduce some diagrams for convenience: for any group G and subgroups H, K and M of G, the diagram ⑥⑥ ⑥⑥ ⑥⑥ ⑥ ⑥ H G❆ ❆❆ ❆❆ ❆❆ ❆ K means G = HK, while the diagram K H ❇❇ ❇❇ ❇❇ ❇❇ M ⑤ ⑤⑤ ⑤⑤ ⑤ ⑤⑤ means H ∩ K = M. Then the diagram ④ ④④ ④④ ④ ④④ G❈ ❈❈ ❈❈ ❈❈ ❈ H1 ❆ ❆❆ ❆❆ ❆❆ ❆ ... ⑥ ⑥⑥ ⑥⑥ ⑥ ⑥⑥ K1 ❆ ❆❆ ❆❆ ❆❆ ❆ ❆❆ ❆❆ ❆❆ ❆❆ ❆ Hℓ ... ⑥ ⑥⑥ ⑥⑥ ⑥ ⑥ ⑥⑥ ❆❆ ❆❆ ❆❆ ❆❆ ❆ Kℓ implies G = H1 K1 = H1 H2 K2 = · · · = H1 H2 · · · Hℓ Kℓ = H1 Kℓ . As mentioned in the Introduction, the maximal factorizations of almost simple groups have been classified by Liebeck, Praeger and Saxl [15]. In order to apply this result to investigate the general factorizations of an almost simple group G, say, we need to embed a given factorization G = HK to the a maximal factorization G = AB. This may be easily accomplished by taking arbitrary maximal subgroups A and B of G containing H and K respectively. However, such maximal subgroups A and B are not necessarily core-free even if H and K are core-free. For example, if HL < G then the maximal subgroup of G containing HL (and thus containing H) is not core-free in G. In fact, HL = G if and only if all maximal subgroup of G containing H are core-free in G. Lemma 2.5. Suppose that G is an almost simple group with socle L and G = HK with core-free subgroups H and K. Then there exist L E G∗ 6 G and a factorization G∗ = H ∗ K ∗ such that H ∗ ∩ L = H ∩ L, K ∗ ∩ L = K ∩ L and H ∗ L = K ∗ L = G∗ . Proof. Take G∗ = HL ∩ KL, H ∗ = H ∩ G∗ and K ∗ = K ∩ G∗ . Then L E G∗ 6 G, H ∗ ∩L = H ∩L and K ∗ ∩L = K ∩L. It follows that H ∗ and K ∗ are both core-free in G∗ since H and K are both core-free in G. By [16, Lemma 2(i)], we have G∗ = H ∗ K ∗ and H ∗ L = K ∗ L = G∗ . Thus, G∗ = H ∗ K ∗ is a factorization as desired.  FACTORIZATIONS OF ALMOST SIMPLE GROUPS 5 To describe a factorization G = HK of an almost simple group G with H and K core-free, if we are concerned with H ∩ Soc(G) and K ∩ Soc(G) instead of H and K, then Lemma 2.5 allows us to assume HL = KL = G. This is equivalent to assume that any maximal subgroups of G containing H and K, respectively, are core-free in G. Furthermore, the following lemma help to determine the factors H and K once we know H ∩ Soc(G) and K ∩ Soc(G) for a factorizations G = HK. Lemma 2.6. Suppose that G is an almost simple group with socle L and G = HK with core-free subgroups H and K. Then there exists a factorization G/L = H K such that H = (H ∩ L).H and K = (K ∩ L).K. Proof. Let H = HL/L and K = KL/L. Then by modulo L, we derive from G = HK that G/L = H K. Moreover, since H = HL/L ∼ = H/(H ∩ L) and ∼ K = KL/L = K/(K ∩ L), we have H = (H ∩ L).H and K = (K ∩ L).K. Hence the lemma follows.  Remark. In an alternative way, one may embed a factorization G = HK of an almost simple group G with H and K core-free into a factorization of G with both factors maximal among core-free subgroups of G. The latter is called a max− factorization of G and is classified in [16]. Suppose that G is an almost simple group with socle L, and G = HK is a factorization of G with H having at least two nonsolvable composition factors and K core-free. We first prove Theorem 1.1 in the case where L is a non-classical simple group. Lemma 2.7. If L is a non-classical simple group, then L = An with n > 10 and one of the following holds: (a) H is a transitive permutation group of degree n, and An−1 6 K 6 Sn−1 ; (b) n = 10, H is a transitive permutation group of degree 10 such that (A5 × A5 ).2 6 H 6 S5 ≀ S2 , and K = SL2 (8) or SL2 (8).3; (c) n = 12, A7 × A5 6 H 6 S7 × S5 , and K = M12 ; (d) n = 24, A19 × A5 6 H 6 S19 × S5 , and K = M24 . Proof. Let L be a non-classical simple group satisfying the assumption of Theorem 1.1. Consulting the classification of factorizations of exceptional groups of Lie type in [10], one sees that L is not an exceptional group of Lie type as H has at least two nonsolvable composition factors. Similarly, L is not a sporadic simple group by [8]. We thus have L = An acting naturally on a set Ω of n points. Since G has a subgroup H which has at least two nonsolvable composition factors, we know that n > 10. Then according to Theorem D and its Remark 2 in [15], one of the following cases appears. (i) H is k-homogeneous on Ω and An−k 6 K 6 Sn−k × Sk for some 1 6 k 6 5. (ii) An−k 6 H 6 Sn−k × Sk and K is k-homogeneous on Ω for some 1 6 k 6 5. (iii) n = 10, H is a transitive subgroup of S10 such that A5 × A5 ⊳ H 6 S5 ≀ S2 , and K = SL2 (8) or SL2 (8).3. Note that all the k-homogeneous permutation groups with k > 2 are known: the k-transitive permutation groups are listed, for example in [5, Tables 7.3 and 7.4], and the k-homogeneous but not k-transitive groups are classified in [12]. 6 LI AND XIA First assume case (i) appears. If k > 2, then the classification of k-homogeneous permutation groups shows that H has at most one nonsolvable composition factor, a contradiction. Hence k = 1 as in part (a). Next assume that case (ii) appears. Then k = 5 because H has at least two nonsolvable composition factors. As K An , we conclude from the classification of 5-homogeneous permutation groups [5, 12] that (n, K) = (12, M12 ) or (24, M24 ). This is described in part (c) or (d), respectively, of Theorem 1.1. Finally, case (iii) leads to part (b). Thus the lemma is true.  To complete the proof of Theorem 1.1, we may assume that L is a classical group of Lie type due to Lemma 2.7. Let A and B be maximal subgroups of G containing H and K, respectively. Then we may assume that A and B are both core-free in G by virtue of Lemma 2.5. We write these assumptions in the following hypothesis for later convenience. Hypothesis 2.8. Let G be an almost simple group with socle L classical of Lie type. Suppose that G = HK is a factorization of G with H having at least two nonsolvable composition factors and K core-free, and A and B are core-free maximal subgroups of G containing H and K, respectively. Notice that A has at least one nonsolvable composition factor since the subgroup H of A is nonsolvable. In Section 5, we deal with the case where A has at least two nonsolvable composition factors. Then in Section 6, we deal with the case where A has exactly one nonsolvable composition factor. In this case, the primitive prime divisors will play an important role in our analysis. Given positive integers a and n, a prime number r is called a primitive prime divisor of the pair (a, n) if r divides an − 1 but does not divide ai − 1 for any positive integer i < n. In other words, a primitive prime divisor of (a, n) is a prime number r such that a has order n in F× r . In particular, we have the observation in the next lemma, which will be used repeatedly (and sometimes implicitly) in this paper. Lemma 2.9. If r is a primitive prime divisor of (a, m), then m r −1 and so r > m. By an elegant theorem of Zsigmondy (see for example [2, Theorem IX.8.3]), (a, n) has a primitive prime divisor whenever a > 2 and n > 3 with (a, n) 6= (2, 6). For any positive integers a > 2 and n > 3, denote the set of primitive prime divisors of (a, n) by ppd(a, n) if (a, n) 6= (2, 6), and set ppd(2, 6) = {7}. For any positive integer m and prime number p, denote by mp the largest p-power that divides m. 3. Examples Let S be a classical group defined on a vector space V . If Z is the center of S such that S/Z is simple group and X is a subgroup of GL(V ) containing S as a normal subgroup, then for any subgroup Y of X, denote by ˆY the subgroup (Y ∩ S)Z/Z of S/Z. If S is a linear group, define Pk [S] to be the stabilizer of a k-space in S. Next assume that S is a symplectic, unitary or orthogonal group. (i) If S is transitive on the totally singular k-spaces, then let Pk [S] be the stabilizer of a totally singular k-space in S. FACTORIZATIONS OF ALMOST SIMPLE GROUPS 7 (ii) If S is not transitive on the totally singular k-spaces, then let Pk−1 [S] and Pk [S] be the stabilizers of totally singular k-spaces in the two different orbits of S. (iii) If S is not transitive on the totally singular k-spaces, then let Pk−1,k [S] be the stabilizer of a totally singular (k − 1)-space in S. (iv) If S is not transitive on the totally singular k-spaces, then let P1,k−1 [S] be the intersection P1 [S] ∩ Pk−1 [S], where the 1-space stabilized by P1 [S] lies in the k-space stabilized by Pk−1 [S]. (v) If S is not transitive on the totally singular k-spaces, then let P1,k [S] be the intersection P1 [S] ∩ Pk [S], where the 1-space stabilized by P1 [S] lies in the k-space stabilized by Pk [S]. For any non-degenerate k-space W , denote (i) Nk [S] = SW if either S is symplectic or unitary, or S is orthogonal of even dimension with k odd; (ii) Nεk [S] = SW for ε = ± if S is orthogonal and W has type ε; (iii) Nεk [S] = SW for ε = ± if S is orthogonal of odd dimension and W ⊥ has type ε. + For the above defined groups Pk [S], Pi,j [S], Nk [S], N− k [S] and Nk [S], we will simply − + write Pk , Pi,j , Nk , Nk and Nk , respectively, if the classical group S is clear from the context. Let G be an almost simple group with socle classical simple, and assume that no element of G induces a triality automorphism if Soc(G) = PΩ+ 8 (q). In [1], Aschbacher defined eight families C1 –C8 of subgroups of G. These groups are now called geometric subgroups of G, and described in more details by Kleidman and Liebeck [13]. The following lemma can be read off from Tables 1–4 of [15]. Lemma 3.1. Let G be an almost simple group with socle L classical. If G = AB is a maximal factorization of G such that A has at least two nonsolvable composition factors and B is core-free in G, then A has exactly two nonsolvable composition factors and (L, A ∩ L, B ∩ L) = (L, X, Y ) as in Table 3. Conversely, for any triple (L, X, Y ) in Table 3, there exists a maximal factorization G = HK satisfying Soc(G) = L, H ∩ L = X and K ∩ L = Y . Table 3. row 1 2 3 4 5 6 7 8 L Sp4ℓ (2f ), f ℓ > 2 Sp4ℓ (4), ℓ > 2 Sp4 (2f ), f > 3 odd Sp6 (2f ), f > 2 PΩ+ 4ℓ (q), ℓ > 3, q > 4 PΩ+ 8 (q), q > 5 odd Ω+ (2) 8 Ω+ 8 (4) X (Sp2ℓ (2f ) × Sp2ℓ (2f )).2 Sp2 (4) × Sp4ℓ−2 (4) (Sp2 (2f ) × Sp2 (2f )).2 Sp2 (2f ) × Sp4 (2f ) (PSp2 (q) × PSp2ℓ (q)). gcd(2, ℓ, q − 1) (PSp2 (q) × PSp4 (q)).2 (SL2 (4) × SL2 (4)).22 (SL2 (16) × SL2 (16)).22 Y f O− 4ℓ (2 ) Sp2ℓ (16).2 Sz(2f ) G2 (2f ) Ω4ℓ−1 (q) Ω7 (q) Sp6 (2) Sp6 (4) 8 LI AND XIA f Example 3.2. Let G = Sp4ℓ (2f ) with f ℓ > 2, and K = O− 4ℓ (2 ) < G. Then for any positive integers a and b such that ab = ℓ, there exists a subgroup H = Sp2a (2f b ) ≀ S2 = ((Sp2a (2f b ) × Sp2a (2f b )):2 of G such that G = HK, as in row 1 of Table 1. fb In fact, let M = Sp4a (2f b ):b be a C3 -subgroup of G. Then M ∩ K = O− 4a (2 ).b, and so |M ∩ K|||G| = |M||K|. It follows that G = MK by Lemma 2.1, and then we have the diagram Sp4ℓ (2f ) ▼▼▼ ▼▼▼ ▼▼▼ ▼ ♦♦ ♦♦♦ ♦♦♦ ♦ ♦ ♦ f O− 4ℓ (2 ) Sp4a(2f b):b ♣♣♣ ♣♣♣ ♣ ♣ ♣♣♣ ❖❖❖ ❖❖❖ ❖❖❖ ❖❖ Sp4a (2f b) ◆◆◆ ♦ ♦ ♦♦♦[15, Theorem A]◆◆◆◆ ♦ ♦ ◆◆◆ ♦♦ ◆ ♦♦♦ ♦ ♦♦♦ ♦♦♦ ♦ ♦ ♦♦ qq qqq q q qqq fb O− 4a (2 ):b fb O− 4a (2 ) Sp2a(2f b) ≀ S2 This gives a subgroup H = Sp2a (2f b ) ≀ S2 = ((Sp2a (2f b ) × Sp2a (2f b )):2 of G such that G = HK, proving Example 3.2. f Example 3.3. Let G = Sp12ℓ (2f ) and K = O− 12ℓ (2 ) < G. Then there exists a subgroup H = G2 (2f ℓ ) ≀ S2 = (G2 (2f ) × G2 (2f )):2 of G such that G = HK, as in row 2 of Table 1. To show that Example 3.3 holds, we first note Sp6 (2f ℓ ) = G2 (2f ℓ )Oε6 (2f ℓ ) for ε = ± by [15, Theorem A]. This implies that fℓ + fℓ Sp6 (2f ℓ ) × Sp6 (2f ℓ ) = (G2(2f ℓ ) × G2 (2f ℓ )(O− 6 (2 ) × O6 (2 )) + fℓ fℓ and so Sp6 (2f ℓ ) ≀ S2 = (G2 (2f ℓ ) ≀ S2 )(O− 6 (2 ) × O6 (2 )). Then the diagram Sp12ℓ(2f ) ▼▼▼ ♦ ▼▼ ♦♦♦ ♦♦♦ [15, Theorem A] ▼▼▼▼▼ ♦ ♦ ▼ ♦♦ f O− 12ℓ(2 ) Sp12(2f ℓ):ℓ ❧❧ ❧❧❧ ❧❧❧ ❧❧❧ ❧ ❧ ❧ Sp12 (2f ℓ) ❦ ❦❦❦ ❦❦❦ ❦❦❦ ❦ ❦ ❦❦ G2 (2f ℓ) ≀ S2 ❙❙❙ ❙❙❙ ❙❙❙ ❙❙❙ ❙❙ fℓ O− 6 (2 ) q qqq qqq q q q fℓ O− 12 (2 ):ℓ ♦ ♦♦♦ ♦♦♦ ♦ ♦ ♦♦ fℓ O− 12 (2 ):ℓ Sp6(2f ℓ) ≀ S2 ♥♥♥ ♥♥♥ ♥♥♥ ♥♥♥ ❘❘❘ ❘❘❘ ❘❘❘ ❘❘❘ ❘❘ ❖❖❖ ❖❖❖ ❖❖❖ ❖❖ × ❧ ❧❧❧ ❧❧❧ ❧❧❧ ❧❧❧ fℓ O+ 6 (2 ) This gives a factorization G = HK with H = G2 (2f ℓ ) ≀ S2 = (G2 (2f ) × G2 (2f )):2. Lemma 3.4. Let W be a 2-dimensional symplectic space over Fq with the nondegenerate symplectic form α and a standard basis w1 , w2 , and V be a 2m-dimensional symplectic space over Fq with the non-degenerate symplectic form β and a standard FACTORIZATIONS OF ALMOST SIMPLE GROUPS 9 basis u1 , v1 , . . . , um , vm . Define a quadratic form Q on W ⊗ V with the associated bilinear form: (W ⊗ V ) × (W ⊗ V ) → Fq , (w ⊗ v, w ′ ⊗ v ′ ) 7→ α(w, w ′)β(v, v ′) such that Q(w ⊗ v) = 0 for any w ∈ W and v ∈ V . Let B be the stabilizer in PΩ(W ⊗ V, Q) of hw1 ⊗ u1 + w2 ⊗ v1 i, and M = PSp2m (q) be a subgroup of PΩ(W ⊗ V, Q) that stabilizes W ⊗ V and acts trivially on W . Then, for m > 2 and q > 4, M ∩ B = {h ∈ PSp2ℓ (q) | uh1 = λu1, v1h = λv1 , λ2 = 1, λ ∈ F∗q }. Proof. Direct calculation shows that M ∩ B = {h ∈ PSp2m (q) | w1 ⊗ uh1 + w2 ⊗ v1h = λ(w1 ⊗ u1 + w2 ⊗ v1 ), λ ∈ F∗q } = {h ∈ PSp2m (q) | w1 ⊗ (uh1 − λu1 ) = w2 ⊗ (λv1 − v1h ), λ ∈ F∗q } = {h ∈ PSp2m (q) | uh1 = λu1 , v1h = λv1 , λ ∈ F∗q } = {h ∈ PSp2m (q) | uh1 = λu1 , v1h = λv1 , β(uh1 , v1h ) = β(u1 , v1 ), λ ∈ F∗q } = {h ∈ PSp2m (q) | uh1 = λu1 , v1h = λv1 , λ2 = 1, λ ∈ F∗q }, completing the proof.  Example 3.5. Let G = PΩ+ 4ℓ (q) with ℓ > 2 and q > 4, and K = N1 [G] = Sp4ℓ−2 (q). Then for any U 6 PSp2 (q), there exists a subgroup H = U × PSp2ℓ (q) of G such that G = HK, as in row 1 of Table 2. In fact, let W , V , Q, B and M be as defined in Lemma 3.4 with m = ℓ. Let G = PΩ(W ⊗ V, Q) and K = B (as there is only one conjugacy class of subgroups N1 [G] in G). Take H = U × M to be a subgroup of G stabilizing W ⊗ V which acts on W and V as U and M = PSp2ℓ (q), respectively. To prove G = HK, we only need to prove G = MK since M 6 H. By Lemma 3.4, we have |M ∩ K||G| = |{h ∈ PSp2ℓ (q) | uh1 = λu1 , v1h = λv1 , λ2 = 1, λ ∈ F∗q }||G| = |Sp2ℓ−2 (q)||G| = |PSp2ℓ (q)||K| = |M||K|. This shows that G = MK, proving Example 3.5. By [15, Theorem A], there is a maximal factorization G = AK with G = ΓSp4ℓ (4), − A = ΓO+ 4ℓ (4) and K = ΓO4ℓ (4). We generalize this in the next lemma by showing that A can be replaced by a smaller group Ω+ 4ℓ (4).2. Lemma 3.6. Let G = L ⋊ hφi = ΓSp4ℓ (4) with ℓ > 2, where L = Sp4ℓ (4) and φ is a field automorphism of order 2, and K = ΓO− 4ℓ (4) < G. Then there exists M = Ω+ (4) < L such that G = (M ⋊ hφi)K and (M ⋊ hφi) ∩ K = N1 [M]. 4ℓ Proof. From [15, 3.2.4(e)] we see that there is a subgroup A = ΓO+ 4ℓ (4) of G such that G = AK, A = (A∩L)⋊hφi and A∩K = N1 [A∩L] = Sp4ℓ−2 (4)×2. Taking M = + Soc(A ∩ L), we have M = Ω+ 4ℓ (4) < L, A ∩ L = O4ℓ (4) and A = (M ⋊ hφi)N1 [A ∩ L]. This implies that A = (M ⋊ hφi)(A ∩ K), and so G = (M ⋊ hφi)K by Lemma 2.4. Moreover, (M ⋊ hφi) ∩ K = (M ⋊ hφi) ∩ (A ∩ K) since M ⋊ hφi < A. Hence (M ⋊ hφi) ∩ K = (M ⋊ hφi) ∩ N1 [(A ∩ L)] = N1 [M], completing the proof.  10 LI AND XIA Example 3.7. Let G = ΓSp4ℓ (4) with ℓ > 2, and K = ΓO− 4ℓ (4) < G. Then there exists a subgroup H = (Sp2 (4) × Sp2ℓ (4)):2 of G such that G = HK, as in row 4 of Table 1. In fact, let φ and M be as in Lemma 3.6 with G = (M ⋊hφi)K and (M ⋊hφi)∩K = N1 [M]. As M = Ω+ 4ℓ (4), there is a maximal subgroup H = (Sp2 (4)⊗Sp2ℓ (4))⋊hφi of M ⋊ hφi. Then H = (Sp2 (4) × Sp2ℓ (4)):2, M ⋊ hφi = HM, and M = (H ∩ M)N1 [M] by [15, Theorem A]. It follows that M ⋊ hφi = H(H ∩ M)N1 [M] = HN1 [M]. Hence we have G = HK as shown in the diagram ΓSp4ℓ (4) ♥ ♥♥♥ ♥♥♥ ♥ ♥ ♥♥♥ ▼▼▼ ▼▼▼ ▼▼▼ ▼ PPP PPP PPP PPP q qqq qqq q q qq ΓO− 4ℓ (4) Ω+ 4ℓ (4) ⋊ hφi ❥❥ ❥❥❥❥ ❥❥❥❥ ❥ ❥ ❥ ❥❥❥ (Sp2(4) ⊗ Sp2ℓ(4)) ⋊ hφi N1 [Ω+ 4ℓ(4)] Example 3.8. Let G = Sp8ℓ (2) and K = O− 8ℓ (2) < G. Then there exists a subgroup H = (Sp2 (4) × Sp2ℓ (4)):2 of G such that G = HK, as in row 5 of Table 1. This is shown in the diagram Sp8ℓ(2) ❑❑ rr ❑❑ r r r r [15, Theorem A] ❑❑❑❑ r r ❑ rr (Sp2 (4) × ΓSp4ℓ(4) O− 8ℓ (2) ▲▲▲ ❧ t ❧ ❧ t ▲ ❧ t ❧ ▲ ❧❧❧ Example 3.7 ▲▲▲▲[15, 3.2.1(d)]tttt ❧❧❧ ▲ tt ❧❧❧ − Sp2ℓ(4)):2 ΓO4ℓ(4) Before stating the next lemma, recall from [15, Theorem A] that there is a maximal + factorization G = AK with G = Ω+ 4m (2), A = ΓO2m (4) and K = N1 [G]. We generalize this by showing in the lemma below that A can be replaced by a smaller group Ω+ 2m (4).2. Lemma 3.9. Let G = Ω+ 4m (2) with m > 2, K = N1 [G] = Sp4m−2 (2), A = + ΓO2m (4) < G and M = Soc(A) ⋊ hφi < G, where φ is the field automorphism of F4 of order 2. Then A∩K 6 O+ 2m (4), G = MK and M ∩K = N1 [Soc(A)] = Sp2m−2 (4). Proof. Let V be a 2m-dimensional orthogonal space over F4 with a non-degenerate quadratic form P of plus type, and let e1 , f1 , . . . , em , fm be a standard basis of V . Take A = O(V, P ) ⋊ hφi, so that M = Ω(V, P ) ⋊ hφi and φ fixes each of e1 , f1 , . . . , em , fm . Define a map Q from V to F2 by putting Q(u) = P (u) + P (u)2 for each u ∈ V . Then Q is a non-degenerate quadratic form of plus type on V while V is considered as a 4m-dimensional orthogonal space over F2 with respect to Q. Regard G as Ω(V, Q) and regard K as the stabilizer in G of e1 + ωf1 , where ω is a 2 2 generator of F× 4 . (Note that Q(e1 +ωf1 ) = P (e1 +ωf1 )+P (e1 +ωf1 ) = ω +ω = 1.) 2 Then A < G since both O(V, P ) and φ preserve P (u) + P (u) for each u ∈ V . FACTORIZATIONS OF ALMOST SIMPLE GROUPS 11 We first prove A ∩ K 6 O(V, P ). To see this, suppose on the contrary that there exists g ∈ O(V, P ) \ (A ∩ K). Then since A = O(V, P ) ⋊ hφi, we conclude that P gφ ∈ K. Writing (e1 + ωf1 )g = 2ℓ i=1 (ai ei + bi fi ) with a1 , b1 , . . . , a2ℓ , b2ℓ ∈ F4 , we deduce that !φ 2ℓ 2ℓ X X e1 + ωf1 = (e1 + ωf1 )gφ = (ai ei + bi fi ) = (a2i ei + b2i fi ). i=1 a21 = 2 b21 a2i i=1 b2i This means 1, = ω and = = 0 for 2 6 i 6 2ℓ, which is equivalent to a1 = 1, b1 = ω and ai = bi = 0 for 2 6 i 6 2ℓ. Now (e1 + ωf1 )g = 2ℓ X (ai ei + bi fi ) = e1 + ω 2f1 , i=1 2 but P (e1 +ωf1 ) = ω 6= ω = P (e1 +ω 2 f1 ), contradicting g ∈ O(V, P ). Consequently, A ∩ K 6 O(V, P ) = O+ 2m (4). By the above conclusion we know that M ∩ K 6 O(V, P ), and so M ∩ K = O(V, P ) ∩ M ∩ K = Ω(V, P ) ∩ K = N1 [Soc(A)] = Sp2m−2 (4). As a consequence, |M ∩ K||G| = |M||K|. This yields G = MK by Lemma 2.1.  Example 3.10. Let G = Ω+ 8ℓ (2) and K = N1 [G] = Sp8ℓ−2 (2). Then there exists a subgroup H = (Sp2 (4) × Sp2ℓ (4)):2 of G such that G = HK, as in row 2 of Table 2. In fact, let φ and M be as in Lemma 3.9 with m = 2ℓ, so that G = MK and M ∩ K = N1 [Soc(M)] = Sp4ℓ−2 (4). Taking H = (Sp2 (4) ⊗ Sp2ℓ (4)) ⋊ hφi to be a maximal subgroup of M = Ω+ 4ℓ (4) ⋊ hφi, we have H = (Sp2 (4) × Sp2ℓ (4)):2, M = HSoc(M), and Soc(M) = (H ∩ Soc(M))N1 [Soc(M)] by [15, Theorem A]. It follows that M = H(H ∩ Soc(M))N1 [Soc(M)] = HN1 [Soc(M)], and then the diagram ♥♥♥ ♥♥♥ ♥ ♥ ♥ ♥♥♥ Ω+ 8ℓ (2) ◆◆◆ ◆◆◆ ◆◆◆ ◆◆ Ω+ 4ℓ (4) ⋊ hφi ❥❥❥❥ ❥❥❥❥ ❥ ❥ ❥ ❥ ❥❥❥❥ PPP PPP PPP PPP Sp8ℓ−2(2) ♣♣♣ ♣♣♣ ♣ ♣ ♣♣♣ N1[Ω+ 4ℓ(4)] (Sp2 (4) ⊗ Sp2ℓ (4)) ⋊ hφi shows that G = HK. Example 3.11. Let G = ΓO+ 8ℓ (4) and K = N1 [G]. Then there exists a subgroup H = (Sp2 (4) × Sp2ℓ (16)):4 of G such that G = HK, as in row 3 of Table 2. Let W , V , Q and M be as in Lemma 3.4 with m = 2ℓ and q = 4, and take ei = w1 ⊗ ui , fi = w2 ⊗ vi , ei+2ℓ = w1 ⊗ vi and fi+2ℓ = w2 ⊗ ui for 1 6 i 6 2ℓ. Then e1 , f1 , . . . , e4ℓ , f4ℓ is a standard basis of the orthogonal space W ⊗ V . Let φ be the field automorphism of order 2 fixing each of e1 , f1 , . . . , e4ℓ , f4ℓ . Then we can regard G as O(W ⊗ V, Q) ⋊ hφi and regard K as the stabilizer in G of 12 LI AND XIA he1 + f1 i. Let A = (C × M) ⋊ hφi be the subgroup of G stabilizing W ⊗ V , where C = Sp2 (4) acts on V trivially, and denote the actions of A on W and V by π1 and π2 respectively. It is evident that Aπ1 = ΓSp2 (4) and Aπ2 = ΓSp4ℓ (4). Take D = Sp2ℓ (16):2 to be a C3 -subgroup of M such that (D ⋊ hφi)π2 = ΓSp2ℓ (16). Then there exists an element σ of M with (D ⋊ hφi)π2 = Soc(D) ⋊ hσφiπ2 = Sp2ℓ (16):4. As (σφ)π1 = φπ1 has order 2 and (σφ)π2 has order 4, we see that σφ has order 4. / Soc(D). Furthermore, σφ normalizes Also, hσφi ∩ Soc(D) = 1 since ((σφ)2 )π2 ∈ C × Soc(D) since both σ and φ normalize C × Soc(D). Thus we have a subgroup H = (C × Soc(D)) ⋊ hσφi = (Sp2 (4) × Sp2ℓ (16)):4 of A. We claim that (A ∩ K)π2 contains the stabilizer in Aπ2 of the non-degenerate 2space hu1 , v1 i. Since A ∩ K > hφi, it suffices to show that (A ∩ K)π2 contains the stabilizer of hu1, v1 i in M π2 = Sp4ℓ (4). Let g be an arbitrary element of M π2 that stabilizes hu1, v1 i. Then ug1 = au1 + bv1 and v1g = cu1 + dv1 for some a, b, c, d ∈ F4 with ad + bc = 1. Take x ∈ C and y ∈ M such that w1x = dw1 + cw2, w2x = bw1 + aw2 and y π2 = g, and let k = xy ∈ A. It is straightforward to verify that k π2 = g and k fixes e1 + f1 , which means k ∈ A ∩ K. This proves our claim. Now we have H π2 > Soc(D) ⋊ hσφiπ2 = (D ⋊ hφi)π2 = ΓSp2ℓ (16) and (A ∩ K)π2 > N2 [Aπ2 ]. It follows from [15, Theorem A] that Aπ2 = ΓSp4ℓ (4) has a factorization Aπ2 = H π2 (A ∩ K)π2 . Then since A contains C = ker(π2 ), we deduce that A = H(A ∩ K). Therefore, the diagram G = ΓO+ 8ℓ (4) ❖❖❖ ✐ ✐ ✐ ✐ ❖❖ ✐ ✐ ✐✐✐ [15, Theorem A] ❖❖❖❖❖ ✐ ✐ ✐ ❖❖ ✐✐✐✐ A = (Sp2 (4) × Sp4ℓ(4)):2 H ❯❯❯❯ ❯❯❯❯ ❯❯❯❯ ❯❯❯❯ ❯❯ ❦ ❦❦❦ ❦❦❦ ❦ ❦ ❦ ❦❦❦ ❦❦❦ K = N1 A∩K ♥ ♥♥♥ ♥♥♥ ♥ ♥ ♥♥♥ shows that G = HK, proving Example 3.11. Example 3.12. Let L = Sp6 (2f ) with f > 2, G be an almost simple group with socle L, H = U ×Sp4 (2f ) 6 N2 [L] with U 6 Sp2 (2f ), and K be a maximal subgroup of G such that K ∩ L = G2 (2f ). Then G = HK, as in row 7 of Table 1. By [19, 4.3], there is a 6-dimensional vector space V over F2f with basis x1 , x2 , x3 , x6 , x7 , x8 and a bilinear form β on V with Gram matrix 1        1 1 1 1 1        FACTORIZATIONS OF ALMOST SIMPLE GROUPS 13 such that L := Sp(V, β) = Sp6 (2f ) has a subgroup N = G2 (2f ) generated by r : (x1 , x2 , x3 , x6 , x7 , x8 ) 7→ (x1 , x3 , x2 , x7 , x6 , x8 ), s : (x1 , x2 , x3 , x6 , x7 , x8 ) 7→ (x2 , x1 , x6 , x3 , x8 , x7 ), T (λ, µ) : (x1 , x2 , x3 , x6 , x7 , x8 ) 7→ (λx1 , µx2 , λµ−1x3 , λ−1 µx6 , µ−1 x7 , λ−1 x8 ), A(α) : x7 7→ x7 + αx1 , x8 7→ x8 + αx2 and F (α) : x2 7→ x2 + αx1 , x6 7→ x6 + α2 x3 , x8 7→ x8 + αx7 , where λ runs over F∗2f , µ runs over F∗2f and α runs over F2f . Let M be the subgroup of L generated by s, (rs)3 , T (λ, µ), A(α), F (α), and φ be the semilinear transformation of V defined by (α1 x1 +α2 x2 +α3 x3 +α6x6 +α7 x7 +α8 x8 )φ = α12 x1 +α22x2 +α32x3 +α62x6 +α72x7 +α82x8 for α1 , α2 , α3 , α6 , α7 , α8 ∈ F2f . Then we can write G = L⋊hφe i and K = N ⋊hφe i for some divisor e of f . Moreover, M stabilizes hx3 , x6 i and hx1 , x2 , x7 , x8 i, respectively. Let A be the stabilizer in G of hx3 , x6 i whose actions on hx3 , x6 i and hx1 , x2 , x7 , x8 i are denoted by π1 and π2 , respectively. It follows that K ∩ L = G2 (2f ), and A ∩ L = (A ∩ L)π1 × (A ∩ L)π2 with (A ∩ L)π1 = Sp2 (2f ) and (A ∩ L)π2 = Sp4 (2f ). Note that f M 6 A ∩ K ∩ L while M = Ω+ 4 (2 ) is a maximal subgroup of K ∩ L. We conclude that A ∩ K ∩ L = M since K ∩ L does not stabilize hx3 , x6 i. Now as |A ∩ K ∩ L||L| = |M||L| = |A ∩ L||K ∩ L|, we have L = (A ∩ L)(K ∩ L) by Lemma 2.1. Regard H = U × (A ∩ L)π2 6 A ∩ L, U being a subgroup of (A ∩ L)π1 . Note that the subgroup of M generated by (rs)3 , T (λ, 1), F (α), where λ runs over F∗2f and α runs over F2f , acts on hx3 , x6 i as Sp2 (2f ). We have M π2 > (A ∩ L)π1 and so (A ∩ K ∩ L)π1 = (A ∩ L)π1 . Accordingly, H(A ∩ K ∩ L) > (A ∩ L)π2 (A ∩ K ∩ L) = (A ∩ L)π2 (A ∩ K ∩ L)π1 = (A ∩ L)π2 (A ∩ L)π1 = A ∩ L, which means A ∩ L = H(A ∩ K ∩ L). This together with H 6 A ∩ L and L = (A∩L)(K∩L) implies that L = H(K∩L). Hence HK = H(K∩L)hφe i = Lhφe i = G, which proves Example 3.12. f f Example 3.13. Let G = Ω+ 12 (2 ) with f > 2, and K = N1 [G] = Sp10 (2 ). Then for f f any U 6 Sp2 (2 ), there exists a subgroup H = U × G2 (2 ) of G such that G = HK, as in row 4 of Table 2. Let W , V , Q, B and M be as defined in Lemma 3.4 with m = 3 and q = 2f . Then we may let G = Ω(W ⊗ V, Q) and K = B. It follows from Lemma 3.4 that M ∩ K = Sp4 (2f ) < N2 [M]. Hence |M ∩ K||G| = |M||K|, and so G = MK by Lemma 2.1. Let N = G2 (2f ) < M, and H = U ⊗ N be a subgroup of G stabilizing W ⊗ V which acts on W and V as U and N, respectively. To prove that G = HK, it suffices to prove G = NK since N 6 H. In fact, Example 3.12 shows that M = N(M ∩ K), which together with G = MK leads to G = NK by Lemma 2.4. 14 LI AND XIA This implies G = HK, proving Example 3.13. We remark that here M is contained in a maximal subgroup A = Sp2 (2f ) ⊗ Sp6 (2f ) of G, so that we have the diagram qqq qqq q q qq qqq M H ⑤ ⑤⑤ ⑤⑤ ⑤ ⑤⑤ ✉ ✉✉ ✉✉ ✉✉ ✉ ✉✉ A ▼▼▼ ■■ ■■ ■■ ■■ ■ ▼▼▼ ▼▼▼ ▼▼ G ❍❍ ❍❍ ❍❍ ❍❍ ❍❍ ✈✈ ✈✈ ✈✈ ✈ ✈✈ K A∩K qq qqq q q qqq M ∩K Example 3.14. Let G = ΓSp12 (4) and H = N2 [G] = (Sp2 (4) × Sp10 (4)):2. Then there exists a subgroup K = G2 (16):4 of G such that G = HK, as in row 8 of Table 1. Let B = ΓSp6 (16) be a C3 -subgroup of G. According to [15, 3.2.1(a)] we have G = HB with H ∩ B = H ∩ Soc(B) = Sp2 (4) × Sp4 (16). Take K = G2 (16):4 to be a maximal subgroup of B (see [4, Table 8.29]). Then B = Soc(B)K and K ∩ Soc(B) = G2 (16). Moreover, Soc(B) = (H ∩ Soc(B))(K ∩ B) as Example 3.12 shows. Thereby we deduce B = (H ∩ Soc(B))K and thus B = (H ∩ B)K. This implies G = HK as shown in the diagram G = ΓSp12(4) H = N2 ❥❥ ❥❥❥❥ ❥❥❥❥ ❥ ❥ ❥ ❥❥❥❥ ❯❯❯❯ ❯❯❯❯ ❯❯❯❯ ❯❯❯❯ ❯ ❚❚❚❚ ❚❚❚❚ ❚❚❚❚ ❚❚❚❚ ❚ ✐ ✐✐✐✐ ✐✐✐✐ ✐ ✐ ✐ ✐ ✐✐✐✐ H ∩ B = Sp2(4) × Sp4 (16) B = ΓSp6 (16) ❘❘❘ ❘❘❘ ❘❘❘ ❘❘❘ ❘ K = G2(16):4 proving Example 3.14. Example 3.15. Let G = ΓSp12 (4) and K = ΓO− 12 (4) < G. Then there exists a subgroup H = (Sp2 (4) × G2 (4)):2 of G such that G = HK, as in row 9 of Table 1. Let φ and M be as in Lemma 3.6 with ℓ = 3, so that G = (M ⋊ hφi)K and (M ⋊ hφi) ∩ K = N1 [M]. Following the proof of Example 3.13 we obtain a subgroup H ∗ = Sp2 (4)×G2 (4) of M such that M = H ∗ N1 [M] and H ∗ is normalized by φ. Take H = H ∗ ⋊ hφi < M ⋊ hφi. We have H = (Sp2 (4) × G2 (4)):2 and M ⋊ hφi = HM. It follows from Lemma 2.4 that M ⋊ hφi = HN1 [M] since H ∩ M = H ∗ and FACTORIZATIONS OF ALMOST SIMPLE GROUPS 15 M = H ∗ N1 [M]. This implies G = HK as shown in the diagram G = ΓSp12(4) ❘❘❘ ❘❘❘ ❘❘❘ ❘❘❘ ❘ ❥❥❥❥ ❥❥❥❥ ❥ ❥ ❥ ❥❥❥❥ K = ΓO− 12 (4) M ⋊ hφi = Ω+ 12 (4):2 H ♥♥ ♥♥♥ ♥♥♥ ♥ ♥ ♥♥ ♥♥♥ ❚❚❚❚ ❚❚❚❚ ❚❚❚❚ ❚❚❚❚ ❚ N1 [M] ❧❧❧ ❧❧❧ ❧❧❧ ❧ ❧ ❧❧❧ proving Example 3.15. Example 3.16. Let G = Ω+ 24 (2) and K = N1 [G] = Sp22 (2). Then there exists a subgroup H = (Sp2 (4) × G2 (4)):2 of G such that G = HK, as in row 5 of Table 2. In fact, let φ and M be as in Lemma 3.9 with m = 6, so that G = MK and M ∩ K = N1 [Soc(M)]. Following the proof of Example 3.13 we obtain a subgroup H ∗ = Sp2 (4) × G2 (4) of Soc(M) such that Soc(M) = H ∗N1 [Soc(M)] and H ∗ is normalized by φ. Take H = H ∗ ⋊ hφi < M. We have H = (Sp2 (4) × Sp2ℓ (4)):2 and M = HSoc(M). This implies M = HN1 [Soc(M)] by Lemma 2.4 since H ∩Soc(M) = H ∗ and Soc(M) = H ∗ N1 [Soc(M)]. Then the following diagram shows that G = HK. G = Ω+ 24 (2) ♠ ♠♠♠ ♠♠♠ ♠ ♠ ♠ ♠♠♠ PPP PPP PPP PPP ◗◗◗ ◗◗◗ ◗◗◗ ◗◗◗ ◗ ♥ ♥♥♥ ♥♥♥ ♥ ♥ ♥♥♥ M = Ω+ 12(4):2 H rr rrr r r rr rrr K = N1 [G] N1[Soc(M)] Example 3.17. Let G = Sp24 (2) and K = O− 24 (2) < G. Then there exists a subgroup H = (Sp2 (4) × G2 (4)):2 of G such that G = HK, as in row 10 of Table 1. This is shown in the diagram Sp (2) O+ 24 (2) qq qqq q q qqq Ω+ 24(2) ▼ ♠ ▼▼▼ ♠ ♠ ♠ ♠ ▼ Example 3.16 ▼▼▼▼ ♠♠♠ ♠ ♠ ▼ ♠♠♠ (Sp2(4) × G2 (4)):2 24 ▼▼▼ ♦♦ ♦♦♦[15, Theorem A]▼▼▼▼ ♦ ♦ ▼▼▼ ♦ ♦♦♦ O− 24 (2) ❖❖❖ ❖❖❖ qqq ❖❖❖ [15, 3.2.4(e)] qqqq ❖❖ qqq + N1 [O24(2)] ♦♦ ♦ ♦ ♦ ♦♦♦ ♦♦♦ N1[Ω+ 24(2)] Example 3.18. Let G = ΓO+ 24 (4) and K = N1 [G]. Then there exists a subgroup H = (Sp2 (4) × G2 (16)):4 of G such that G = HK, as in row 6 of Table 2. 16 LI AND XIA Let W , V , Q and M be as in Lemma 3.4 with m = 6 and q = 4, and take ei = w1 ⊗ ui , fi = w2 ⊗ vi , ei+6 = w1 ⊗ vi and fi+6 = w2 ⊗ ui for 1 6 i 6 6. Then e1 , f1 , . . . , e12 , f12 is a standard basis of the orthogonal space W ⊗ V . Let φ be the field automorphism of order 2 fixing each of e1 , f1 , . . . , e12 , f12 . Then we can regard G as O(W ⊗ V, Q) ⋊ hφi and regard K as the stabilizer in G of he1 + f1 i. Let A = (C × M) ⋊ hφi be the subgroup of G stabilizing W ⊗ V , where C = Sp2 (4) acts on V trivially, and denote the actions of A on W and V by π1 and π2 respectively. It is evident that Aπ1 = ΓSp2 (4) and Aπ2 = ΓSp12 (4). Following the proof of Example 3.14 we see that there is a subgroup D = G2 (16):2 of M = Sp12 (4) with (D⋊hφi)π2 = G2 (16):4 and Aπ2 = (D⋊hφi)π2 N2 [Aπ2 ]. Accordingly, there exists an element σ of M such that (D ⋊ hφi)π2 = Soc(D) ⋊ hσφiπ2 = G2 (16):4. Since (σφ)π1 = φπ1 has order 2 and (σφ)π2 has order 4, we conclude that σφ has order 4. Also, hσφi ∩ Soc(D) = 1 as ((σφ)2 )π2 ∈ / Soc(D). Furthermore, σφ normalizes C × Soc(D) since both σ and φ normalize C × Soc(D). Thus we have a subgroup H = (C × Soc(D)) ⋊ hσφi = (Sp2 (4) × G2 (16)):4 of A. As in the proof of Example 3.11 we know that (A ∩ K)π2 contains the stabilizer in Aπ2 of the non-degenerate 2-space hu1, v1 i. Now (A ∩ K)π2 > N2 (Aπ2 ) and Aπ2 = H π2 N2 [Aπ2 ]. It follows that Aπ2 = H π2 (A ∩ K)π2 , and so A = H(A ∩ K) since A contains C = ker(π2 ). Therefore, the diagram G = ΓO+ 24 (4) ❖❖❖ ✐ ❖❖ ✐✐✐✐ ✐ ✐ ✐ ✐ ✐ [15, Theorem A] ❖❖❖❖❖ ✐ ✐ ✐ ✐ ❖❖ ✐ ✐✐ A = (Sp2(4) × Sp12(4)):2 H ❦❦❦ ❦❦❦ ❦❦❦ ❦ ❦ ❦❦❦ ❦❦❦ ❯❯❯❯ ❯❯❯❯ ❯❯❯❯ ❯❯❯❯ ❯❯❯ K = N1 A∩K ♥♥ ♥♥♥ ♥♥♥ ♥ ♥ ♥♥ shows that G = HK, proving Example 3.18. Corollary 3.19. For each L as in the second column of Table 1, there exist an almost simple group G with socle L and a factorization G = HK such that H ∩ L and K ∩ L are as described in the same row as L in Table 1. Proof. For row 3 and row 6 of Table 1, there is a maximal factorization G = HK (corresponding to row 2 and row 3 of Table 3 respectively) as Lemma 3.1 shows. For the other rows of Table 1, the assertion is true by the example indicated in the last column of Table 1.  4. At least two nonsolvable composition factors of A In this section we prove that under Hypothesis 2.8, if A has at least two nonsolvable composition factors then (L, H ∩L, K ∩L) lies in Table 1 or Table 2, as part (b) of Theorem 1.1 asserts. For Sp4 (2f ) 6 G 6 ΓSp4 (2f ) with f > 2, we have factorizations G = XY with f f f X ∩ Soc(G) = O− 4 (2 ) or Sz(2 ) and Y ∩ L = Sp2 (2 ) ≀ S2 (see [15, Theorem A]). The lemma below shows that, however, G = XY does not hold if Y is replaced by its index 2 subgroup N2 [G]. FACTORIZATIONS OF ALMOST SIMPLE GROUPS 17 Lemma 4.1. Let L = Sp4 (2f ) with f > 2, and L 6 G 6 ΓSp4 (2f ). Suppose G = XN2 [G] for some subgroup X of G. Then X > L. Proof. Let Y be a maximal subgroup of G containing T := N2 [G]. We have Y ∩ L = Sp2 (2f ) ≀ S2 and G = XY . It then follows from [15, Theorem A] that one of the following cases appears. (i) X > L. f (ii) X ∩ L 6 O− 4 (2 ). (iii) f is odd and X ∩ L 6 Sz(2f ). f If case (ii) appears, then the factorization G = XT would imply that ΓO− 4 (2 ) is transitive on the nonsingular 2-dimensional symplectic subspaces, which is not true. Suppose that case (iii) appears. Let Z be a maximal subgroup of G containing X such that Z ∩ L = Sz(2f ). According to [18], Z ∩ L contains an element g whose matrix with respect to a standard basis e1 , f1 , e2 , f2 for Sp4 (2f ) is   0 1  1 0 .   0 1 1 0 By the equivalence of statements (a) and (b) of Lemma 2.1, we may assume that T fixes he1 , f1 i. Thus Z ∩ T > hgi and so |Z ∩ T |2 > 2. Since G/L 6 f has odd order, we have |Z|2 = |Z ∩ L|2 and |T |2 = |T ∩ L|2 . Moreover, the factorization G = XT indicates G = ZT since X 6 Z. Thus we derive that |Z|2 |T |2 |Z ∩ L|2 |T ∩ L|2 |Z ∩ T |2 = = |G|2 |G|2 |Sz(2f )|2 |Sp2 (2f ) × Sp2 (2f )|2 = = 1, |Sp4 (2f )|2 a contradiction. Hence X > L as the lemma asserts.  We also need the following two technical lemmas. Lemma 4.2. Let L = Sp2ℓ (2f ) with f ℓ > 2, and L 6 G 6 ΓSp2ℓ (2f ). Suppose G = RX = RY for some nonsolvable subgroup R and subgroups X and Y such − f f fk that X ∩ L = O+ 2ℓ (2 ) and Y ∩ L = O2ℓ (2 ). Then either Sp2ℓ/k (2 ) 6 R ∩ L 6 Sp2ℓ/k (2f k ).k for some integer k dividing ℓ, or G2 (2f ℓ/3 ) 6 R ∩ L 6 G2 (2f ℓ/3 ).(ℓ/3) with ℓ divisible by 3. Proof. We use induction on ℓ. First suppose ℓ = 1. Take any r ∈ ppd(2, 2f ). Then r divides |G|/|X|, and so r divides |R| since G = RX. As r > f and |R|/|R ∩ L| 6 f , it follows that r divides |R ∩ L|. Thus, from the classification of subgroups of Sp2 (2f ) = SL2 (2f ) (see for example [11, II §8]) we conclude that R ∩ L = Sp2 (2f ) as R ∩ L is an nonsolvable subgroup of Sp2 (2f ) of order divisible by r. Next suppose that ℓ > 1 and the lemma holds for each smaller ℓ. Let T be a maximal subgroup of G containing R, and assume that T is core-free in G in view of Lemma 2.5. It follows that G = T X = T Y . Then by [15, Theorem A], either T ∩ L = Sp2a (2f b ).b with ab = ℓ and b prime, or T ∩ L = G2 (2f ) with ℓ = 3. Case 1. T ∩L = Sp2a (2f b ).b, where ab = ℓ and b is a prime divisor of ℓ. In this case, T is an almost simple group with socle Sp2a (2f b ), and we deduce from Lemma 2.3 18 LI AND XIA fb that T = R(X ∩T ) = R(Y ∩T ). Furthermore, it is clear that X ∩Soc(T ) = O+ 2a (2 ) − fb and Y ∩ Soc(T ) = O2a (2 ). Then by our inductive hypothesis, either (i) Sp2a/d (2f bd ) 6 R ∩ Soc(T ) 6 Sp2a/d (2f bd ).d for some divisor d of a, or (ii) G2 (2f ba/3 ) 6 R ∩ Soc(T ) 6 G2 (2f ba/3 ).(a/3) with a divisible by 3. Note R∩Soc(T ) 6 R∩L 6 (R∩Soc(T )).((T ∩L)/Soc(T )) = (R∩Soc(T )).b. For (i), we obtain by writing k = bd that Spℓ/k (2f k ) 6 R ∩ L 6 (Spℓ/k (2f k ).d).b = Spℓ/k (2f k ).k. For (ii), we have G2 (2f ℓ/3 ) 6 R ∩ L 6 (G2 (2f ℓ/3 ).(a/3)).b = G2 (2f ℓ/3 ).(ℓ/3). Hence the conclusion of the lemma holds. Case 2. ℓ = 3 and T ∩ L = G2 (2f ). In this case, G2 (2f ) 6 T 6 G2 (2f ).f , and T = R(X ∩T ) with X ∩T ∩L = SL3 (2f ).2 (see [15, 5.2.3(b)]). Since R is nonsolvable, we conclude R > Soc(T ) by [10], which implies R ∩ L = G2(2f ). This satisfies the conclusion of the lemma. The proof is thus completed.  Lemma 4.3. Let G be an almost simple group with socle L = PSp2ℓ (q), where ℓ > 3 and q > 4. Suppose G = XN2 [G] for some subgroup X of G. Then one of the following holds. (a) (b) (c) (d) X > L. ℓ = 3, q is even and X ∩ L = G2 (q). ℓ is even, q = 4 and Spℓ (q 2 ) 6 X ∩ L 6 Spℓ (q 2 ).2. ℓ = 6, q = 4 and G2 (q 2 ) 6 X ∩ L 6 G2 (q 2 ).2. Proof. If X is not core-free in G, then part (a) of the lemma holds. In the remaining of the proof we assume that X is core-free in G. Let T be a maximal subgroup of G containing X and let Y = N2 [G]. It follows from G = XY that G = T Y . By Lemma 2.5 we may assume that T is core-free in G. Then appealing to [15, Theorem A] we have the following two cases. (i) ℓ = 3, q is even and T ∩ L = G2 (q). (ii) ℓ is even, q = 4 and T ∩ L = Spℓ (q 2 ).2. Note that T = X(T ∩ Y ) by Lemma 2.3. First assume that case (i) appears. Here T is almost simple with socle G2 (q) and T ∩ Y ∩ L = Sp2 (q) × Sp2 (q) (see [15, 5.2.3(b)]). According to [10], the factorization T = X(T ∩ Y ) implies that X > Soc(T ). Consequently, X ∩ L = G2 (q) as described in part (b) of the lemma. Next assume that case (ii) appears. In this case, Spℓ (q 2 ) 6 T 6 ΓSpℓ (q 2 ) and T ∩ Y 6 N2 [T ] (see [15, 3.2.1(a)]). If X is not core-free in T , then part (c) of the lemma holds. Now assume that X is core-free in T . Then by [15, Theorem A], ℓ = 6 and X ∩ Soc(T ) 6 G2 (q). Applying the same argument as in case (i) (with G therein replaced by T here) to the factorization T = XN2 [T ] we conclude that X ∩ Soc(T ) = G2 (q). This implies G2 (q 2 ) 6 X ∩ L 6 G2 (q 2 ).2, as in part (d) of the lemma.  Now we can prove the main result of this section. FACTORIZATIONS OF ALMOST SIMPLE GROUPS 19 Lemma 4.4. Suppose Hypothesis 2.8 and that A has at least two nonsolvable composition factors, then (L, H ∩ L, K ∩ L) lies in Table 1 or Table 2. Proof. By Lemma 3.1, (L, A ∩ L, B ∩ L) = (L, X, Y ) as in Table 3. Let G = L.O with O 6 Out(L). We discuss the eight rows of Table 3 in order. Row 1. In this row, O 6 f and A ∩ L = (C × D).2 with C ∼ = D = Sp2ℓ (2f ). Let M = (C ×D).O = N2ℓ [G] be the subgroup of index 2 in A such that M ∩L = C ×D. − f f Since A ∩ B = (O+ 2ℓ (2 ) × O2ℓ (2 )).O (see [15, 3.2.4(b)]), we have A ∩ B 6 M. Applying Lemma 2.3, we deduce from G = HB that A = H(A ∩ B) and then M = (H ∩ M)(A ∩ B). Note that M 6 (C.O) × (D.O). Let π+ and π− be the projections from M to C.O and D.O respectively. Then M πε = (H ∩ M)πε (A ∩ B)πε for ε = ±. f If H 6 M, then G = N2ℓ [G]B, which implies that B = O− 4ℓ (2 ).O is transitive on the nonsingular symplectic subspaces of dimension 2ℓ, a contradiction. Hence there exists g ∈ H such that C g = D and so (H ∩ C)g = H ∩ D. As a consequence, (H∩(C×D))π+ ∼ = (H∩(C×D))π− = U for some U 6 Sp2ℓ (2f ), and (H∩M)πε 6 U.O for ε = ±. Since M πε = (H ∩ M)πε (A ∩ B)πε and (A ∩ B)πε = Oε2ℓ (2f ).O for ε = ±, we conclude from Lemma 4.2 that either Sp2ℓ/k (2f k ) 6 U 6 Sp2ℓ/k (2f k ).k for some integer k dividing ℓ, or G2 (2f ℓ ) 6 U 6 G2 (2f ℓ/3 ).(ℓ/3) with ℓ divisible by 3. In particular, U is an almost simple group. This implies H ∩(C ×D) > Soc(U)×Soc(U) as H has at least two nonsolvable composition factors. It follows that H ∩ L = (H ∩ (C × D)).2 = (Soc(U) × Soc(U)).R.2, where either Soc(U) = Sp2ℓ/k (2f k ) and R 6 k × k with k dividing ℓ, or Soc(U) = G2 (2f ℓ/3 ) and R 6 (ℓ/3) × (ℓ/3) with ℓ divisible by 3. Next we show that K ∩ L > f Ω− 4ℓ (2 ) and thus either row 1 or row 2 of Table 1 appears. f Suppose K ∩ L Ω− 4ℓ (2 ). Then we have a factorization B = (A ∩ B)K with f K core-free in B. Moreover, B is an almost simple group with socle Ω− 2ℓ (2 ), and + − f f A ∩ B = (O2ℓ (2 ) × O2ℓ (2 )).O is a maximal subgroup of B. However, inspection of [15, Tables 1–3] shows that no such factorization exists. This contradiction proves f that K ∩ L > Ω− 4ℓ (2 ), as desired. Row 2. From [15, Table 1] we see that G = L.2 and A = (C ×D).2 < (C.2)×(D.2), where C = Sp2 (4) and D = Sp4ℓ−2 (4). Let ϕ be the projection from A to D.2. Then since A = H(A ∩ B) by Lemma 2.3, we have Aϕ = H ϕ (A ∩ B)ϕ . Moreover, Aϕ is an almost simple group with socle Sp4ℓ−2 (4), and (A ∩ B)ϕ 6 N2 [Aϕ ]. Suppose that H ϕ is core-free in Aϕ . Then as Aϕ = H ϕ N2 [Aϕ ], we conclude from [15, Theorem A] that ℓ = 2 and H ϕ 6 G2 (4).2. Consequently, H ∩ L 6 C × M with M = G2 (4) < D. Applying Lemma 2.2 to the factorization G = HB we see that |H ∩ L| is divisible by |L|/|B| = |Sp8 (4)|/|Sp4 (16).4| = 4|G2(4)| = 4|M|. Since the smallest index of proper subgroups of M = G2 (4) is 416 (see [6]), which is larger than |C|/4, it follows that H ∩L > M. This in turn forces H ∩L = C ×M in order that H has at least two nonsolvable composition factors. However, A ∩ B > C (see [15, 3.2.1(a)]), whence |H ∩ B| > |C| > |C| |C||M||B| |H ∩ L||B| |H||B| = = > , 4 |L| |L| |G| contradicting the factorization G = HB. Thus, H ϕ is not core-free in Aϕ , and so H ∩ L > D. As H has at least two nonsolvable composition factors, we thereby 20 LI AND XIA obtain H ∩ L = C × D = Sp2 (4) × Sp4ℓ−2 (4). If K ∩ L > Sp2ℓ (16), then row 3 of Table 1 appears. Now assume K ∩ L Sp2ℓ (16). Note that B is almost simple with socle Sp2ℓ (16) and A ∩ B = Sp2 (4) × Sp2ℓ−2 (16) < N2 [B]. We have B = N2 [B]K with K core-free in B. Thereby we conclude from Lemma 4.3 that ℓ = 3 and G2 (16) 6 K ∩ L 6 G2 (16).2. This leads to row 8 of Table 1. Row 3. In this row, O 6 f and A ∩ L = (C × D).2 with C ∼ = D = Sp2 (2f ). Applying Lemma 2.2 to the factorization G = HB we see that f |H ∩ L| is divisible by |L|/|B ∩ L| = 22f (22f − 1)(2f + 1). Consequently, as |H ∩ L|/|H ∩ (C × D)| divides |A ∩ L|/|C × D| = 2, f |H ∩ (C × D)| is divisible by 22f −1 (22f − 1)(2f + 1). Moreover, |H ∩ (C × D)| divides |H ∩ C||D| = |H ∩ C||Sp2 (2f )|. It follows that f |H ∩ C| is divisible by 22f −1 (22f − 1)(2f + 1)/|Sp2 (2f )| = 2f −1 (2f + 1). Then by the classification of subgroups of Sp2 (2f ) we conclude that H ∩ C = Sp2 (2f ). For the same reason, H ∩ D = Sp2 (2f ). Therefore, H ∩ L > C × D, which means that H ∩ L = (C × D).P = (Sp2 (2f ) × Sp2 (2f )).P with P 6 2. Since G = AK, by Lemma 2.3 we have B = (A ∩ B)K. Note that B is almost simple with socle Sz(2f ) and A ∩ B = D2(2f −1) .O (see [15, 5.1.7(b)]). We then infer from [10] that K > Soc(B). Hence K ∩ L = Sz(2f ), as in row 6 of Table 1. Row 4. Let N, M, φ, A, π1 and π2 be as in the proof of Example 3.12. Since G has a unique conjugacy class of maximal subgroups isomorphic to B, we may assume that B = hN, φe i by Lemma 2.1(b), where e = |O|. Then A ∩ B = M ⋊ hφe i. By Lemma 2.3, we derive from G = HB and H 6 A that A = H(A ∩ B), so Aπ2 = H π2 (A ∩ B)π2 . Note that Aπ2 = Sp4 (2f ) has a graph automorphism fusing f its C2 -subgroups with C8 subgroups, and M π2 = Ω+ 4 (2 ). We thereby obtain a factorization Aπ2 γ = H π2 γ N2 [Aπ2 γ ] for some automorphism γ of Aπ2 . Hence by Lemma 4.1, H π2 > Soc(Aπ2 ) = Sp4 (2f ), and so H ∩ L = U × Sp4 (2f ) for some subgroup U of (A ∩ L)π1 = Sp2 (2f ). Moreover, U is nonsolvable because H ∩ L has at least two nonsolvable composition factors. Next we show that K ∩ L = G2 (2f ) and thus (L, H ∩ L, K ∩ L) lies in row 7 of Table 1. Suppose K ∩ L < B ∩ L = G2 (2f ). Then is K core-free in B. By Lemma 2.3, we derive from G = AK and K 6 B that B = (A ∩ B)K. However, as B is an almost simple group with socle G2 (2f ) and A ∩ B = (Sp2 (2f ) × Sp2 (2f )).O, it is seen in [10] that no such factorization B = (A ∩ B)K exists. This contradiction yields K ∩ L = G2 (2f ), as desired. Rows 5 and 6. In these two rows, L = PΩ+ 4ℓ (q) with either ℓ > 3 and q > 4 or ℓ = 2 and q > 5. Putting c = gcd(2, ℓ, q −1), we have A∩L = (C ×D).c with C = PSp2 (q) and D = PSp2ℓ (q). Let M be the subgroup of index c in A such that M ∩L = C ×D. Then A ∩ B 6 M and M 6 (C.O) × (D.O). Denote by ϕ the projection from M to D.O. Applying Lemma 2.3, we deduce from G = HB that A = H(A ∩ B) and then M = (H ∩ M)(A ∩ B). Thereby we have M ϕ = (H ∩ M)ϕ (A ∩ B)ϕ . Moreover, M ϕ /CM ϕ (D) is almost simple with socle D = PSp2ℓ (q) and (A ∩ B)ϕ 6 N2 [M ϕ ] FACTORIZATIONS OF ALMOST SIMPLE GROUPS 21 (see [15, 3.6.1(d)]). It follows that M ϕ = (H ∩ M)ϕ N2 [M ϕ ], and so one of the following four cases appears by Lemma 4.3 . (i) (H ∩ M)ϕ > D. (ii) ℓ = 3, q is even and (H ∩ M)ϕ ∩ D = G2 (q). (iii) ℓ is even, q = 4 and Spℓ (q 2 ) 6 (H ∩ M)ϕ ∩ D 6 Spℓ (q 2 ).2. (iv) ℓ = 6, q = 4 and G2 (q 2 ) 6 (H ∩ M)ϕ ∩ D 6 G2 (q 2 ).2. Note that H ∩ L has at least two nonsolvable composition factors as H has at least two nonsolvable composition factors. If case (i) appears, then H ∩ L = (U × PSp2ℓ (q)).P with P 6 gcd(2, ℓ, q − 1), where U is an nonsolvable subgroup of PSp2 (q). If case (ii) appears, then H ∩ L = U × G2 (q) with ℓ = 3 and q even, where U is an nonsolvable subgroup of Sp2 (q). For case (iii), H ∩ L = (Sp2 (4) × Spℓ (16)).P with ℓ even, q = 4 and P 6 2. For case (iv) we have H ∩ L = (Sp2 (4) × G2 (16)).P with ℓ = 6, q = 4 and P 6 2. Next we show that K ∩ L = Ω4ℓ−1 (q) and thus cases (i)–(iv) lead to rows 1, 4, 3, 6, respectively, of Table 2. Suppose K ∩ L 6= B ∩ L = Ω4ℓ−1 (q). From G = AK we derive that |K| is divisible by |G|/|A|, whence any prime in ppd(q, 4ℓ − 2) ∪ ppd(q, 2ℓ − 1) ∪ ppd(q, 4ℓ − 4) divides |K|. However, by Lemma 2.3 we have B = (A ∩ B)K, and [15, Theorem A] shows that no such group K satisfies these conditions. This contradiction implies K ∩ L = Ω4ℓ−1 (q), as desired. 2 Row 7. Here L = Ω+ 8 (2), H ∩ L 6 (SL2 (4) × SL2 (4)).2 and K ∩ L 6 Sp6 (2). In order that H ∩ L has at least two nonsolvable composition factors, we have H ∩L = (SL2 (4)×SL2 (4)).P with P 6 22 . Furthermore, computation in Magma [3] shows K ∩ L = Sp6 (2) in order that G = HK. Thus row 2 (with ℓ = 1) of Table 2 appears. 2 Row 8. In this row, L = Ω+ 8 (4), A∩L = (SL2 (16)×SL2 (16)).2 and B∩L 6 Sp6 (4). Then |O| divides 4, and thus Lemma 2.2 implies that |L| divides 4|K ∩ L||H ∩ L|. In particular, |L| divides 4|B ∩ L||H ∩ L|, which indicates that |H ∩ L| is divisible by 17. Then as H ∩ L is a subgroup of A ∩ L = (SL2 (16) × SL2 (16)).22 with at least two nonsolvable composition factors, we conclude that either H ∩ L = (SL2 (4) × SL2 (16)).P with P 6 2, or H ∩ L = (SL2 (16) × SL2 (16)).P with P 6 2 . For the former, since 4|K∩L| is divisible by |L|/|(SL2 (4)×SL2 (16)).2| = 217 · 33 · 52 · 7 · 13 · 17, we have K ∩ L = Sp6 (4), which leads to row 3 (with c = 1 and ℓ = 1) of Table 2. Now assume that H ∩ L = (SL2 (16) × SL2 (16)).P with P 6 22 . It follows that 4|K ∩ L| is divisible by |L|/|(SL2 (16) × SL2 (16)).22| = 214 · 33 · 52 · 7 · 13. Therefore, K ∩L = Sp6 (4) or G2 (4). If K ∩L = G2 (4), then |K| divides |Out(L)||K ∩ 2 L| = 4|G2(4)| = |Ω+ 8 (4)|/|(SL2 (16) × SL2 (16)).2 | = |L|/|A ∩ L| = |G|/|A|, which implies that A ∩ K = 1. However, there exists no such factorization G = AK according to [17, Theorem 1.1]. Thus we have K ∩ L = Sp6 (4), and so row 3 (with c = 2 and ℓ = 1) of Table 2 appears.  2 5. Exactly one nonsolvable composition factor of A In this section we deal with the case where the group A in Hypothesis 2.8 has exactly one nonsolvable composition factor. For a group X, the solvable radical of 22 LI AND XIA X, denoted by Rad(X), is the product of all the solvable normal subgroups of X. Note that if X has exactly one nonsolvable composition factor then X/Rad(X) is almost simple. Hence from the factorization A = H(A∩B) we obtain a factorization A/Rad(A) = (HRad(A)/Rad(A))((A ∩ B)Rad(A)/Rad(A)) of the almost simple group A/Rad(A) with the factor HRad(A)/Rad(A) still having at least two composition factors. We shall see in Lemma 5.1 that the other factor (A ∩ B)Rad(A)/Rad(A) is core-free. For a group X, let X (∞) be the smallest normal subgroup of X such that X/X (∞) is solvable. Lemma 5.1. Suppose that G is an almost simple group with socle classical of Lie type, and G = AB with subgroups A and B maximal and core-free in G. If A has exactly one nonsolvable composition factor and R = Rad(A), then A/R is almost simple and (A ∩ B)R/R is core-free in A/R. Proof. Let S be the unique nonsolvable composition factor of A. Then A/R is an almost simple group with socle S. Suppose for a contradiction that (A ∩ B)R/R contains Soc(A/R) = (A/R)(∞) = A(∞) R/R. We deduce that (A ∩ B)R > A(∞) R. Then since |(A ∩ B)R| divides |A ∩ B||R|, it follows that |A ∩ B||R| is divisible by |A(∞) R|. From G = AB we have |G| |A| |A||R| = = . |B| |A ∩ B| |A ∩ B||R| Therefore, |G|/|B| divides |A||A(∞) ∩ R| |A||Rad(A(∞) )| |A| |A||R| = = = (∞) . (∞) (∞) (∞) |A R| |A | |A | |A /Rad(A(∞) )| Since A(∞) /Rad(A(∞) ) is also an almost simple group with socle S, we see that |A(∞) /Rad(A(∞) )| is divisible by |S| and so |G|/|B| divides |A|/|S|. However, one concludes from [15, Theorem A] that no such factorization G = AB exists with |G|/|B| dividing |A|/|S|. Hence (A ∩ B)R/R does not contain Soc(A/R), which means that (A ∩ B)R/R is core-free in A/R.  The following lemma is needed later. + Lemma 5.2. Let Ω+ 4m (4) 6 G 6 O4m (4) with m > 2, and H be a subgroup of G such that G = HN1 [G]. If H is contained in a C4 -subgroup of G, then H > Sp2m (4). Proof. Let W , V , Q and M be as in Lemma 3.4 with q = 4, and regard G as + Ω(W ⊗ V, Q) or O(W ⊗ V, Q) according to G = Ω+ 4m (4) or O4m (4), respectively. Let A be the maximal C4 -subgroup of G containing H. Then G = AN1 [G], and so by [15, Theorem A] we may regard A as the subgroup of G stabilizing W ⊗ V . From [13] we see that A < Soc(G) and thus A = Sp2 (4) × M. Denote the projection of A onto M by ϕ. Since G = HN1 [G] and H 6 A, we deduce from Lemma 2.3 that A = H(A ∩ N1 [G]), which yields M = Aϕ = H ϕ (A ∩ N1 [G])ϕ . As (A ∩ N1 [G])ϕ 6 N2 [M] (see [15, 3.6.1(d)]), we further deduce that M = H ϕ N2 [M]. According to [15, Theorem A], M = Sp2m (4) has no maximal factorization with a factor N2 [M]. Hence H ϕ = M. This implies H > Sp2m (4) as H 6 A = Sp2 (4) × Sp2m (4).  FACTORIZATIONS OF ALMOST SIMPLE GROUPS 23 Based on the results above, we are able to prove Theorem 1.1 by induction on the order of G. The inductive hypothesis is as follows. Hypothesis 5.3. Suppose that for each almost simple group G1 of order properly dividing |G|, if G1 = H1 K1 is a factorization of G1 with H1 having at least two nonsolvable composition factors and K1 core-free, then (G1 , H1 , K1 ) is described in Theorem 1.1. Suppose both Hypothesis 2.8 and Hypothesis 5.3 in the rest of this section. We first give the candidates for (L, A ∩ L, B ∩ L) in the case where A has exactly one nonsolvable composition factor. Lemma 5.4. Either A has at least two nonsolvable composition factors, or the triple (L, A ∩ L, B ∩ L) lies in Table 4. Table 4. row 1 2 3 4 5 6 L Sp4ℓ (2f ), ℓ > 2 Sp8ℓ (2), ℓ > 1 Sp4ℓ (4), ℓ > 2 Ω+ 8ℓ (2), ℓ > 2 Ω+ 8ℓ (4), ℓ > 2 f Ω+ 8 (2 ), f > 2 A∩L B∩L fb f Sp4a (2 ).b with ab = ℓ and b prime O− 4ℓ (2 ) + − O8ℓ (2) O8ℓ (2) + O4ℓ (4) O− 4ℓ (4) + 2 Ω4ℓ (4).2 Sp8ℓ−2 (2) 2 Ω+ Sp8ℓ−2 (4) 4ℓ (16).2 f Sp6 (2 ) Sp6 (2f ) Proof. The maximal factorizations G = AB are classified by [15, Theorem A] and described in Tables 1–4 of [15]. Since H is nonsolvable, A is nonsolvable too. Thus, either A has at least two nonsolvable composition factors, or A has precisely one nonsolvable composition factor. Assume in the following that A has precisely one nonsolvable composition factor. Denote R = Rad(A), A = A/R, H = HR/R and A ∩ B = (A ∩ B)R/R. Then A is an almost simple group, and H has at least two nonsolvable composition factors. Moreover, applying Lemma 2.3 we derive from G = HB and H 6 A that A = H(A ∩ B). Hence A = H A ∩ B with A ∩ B core-free in A as Lemma 5.1 asserts. Now by Hypothesis 5.3, either Soc(A) = An for some n > 10, or Soc(A) lies in the second column of Table 1 or Table 2. Checking this condition of A for factorizations G = AB in Tables 1–4 of [15], we deduce that either (L, A ∩ L, B ∩ L) lies in Table 5 or one of the following cases appears. (i) L = Sp2m (2f ), A ∩ L = Sp2a (2f b ).b with ab = m and b prime, and B ∩ L = f O− 2m (2 ). − (ii) L = Sp2m (2), A ∩ L = O+ 2m (2), and B ∩ L = O2m (2). + (iii) L = Sp2m (4), A ∩ L = O2m (4), and B ∩ L = O− 2m (4). f f (iv) L = Sp6 (2 ), A ∩ L = P1 , and B ∩ L = G2 (2 ). (v) G = L = Sp8 (2), A = S10 , B = O− 8 (2). + f (vi) L = Ω2m (2 ) with m > 6 even, A ∩ L = N1 , and B ∩ L = ˆGUm (q).2. f + f 2 (vii) L = Ω+ 2m (2 ) with m > 6 even and f 6 2, A∩L = Ωm (4 ).2 , and B∩L = N1 . (viii) L = Ω− 10 (2), A ∩ L = A12 , and B ∩ L = P1 . f f f (ix) L = Ω+ 8 (2 ) with f > 2, A ∩ L = Ω7 (2 ), and B ∩ L = Ω7 (2 ). 24 LI AND XIA + (x) L = PΩ+ 8 (3), A ∩ L = Ω8 (2), and B ∩ L = Ω7 (3). Table 5. L PSLn (q) A∩L NL (PSpn (q)) B∩L P1 or Pn−1 PSLn (q) NL (PSpn (q)) Stab(V1 ⊕ Vn−1 ) PSp2m (q) PSp2m (q) PSp2a (q b ).b P1 P1 PSp2a (q b ).b PSp2m (2f ) Sp2a (2f b ).b f PSp2m (2f ) O+ 2m (2 ) PSp2m (2f ) Spm (4f ).2 f O+ 2m (2 ) Sp2a (2f b ).b N2 PSp2m (2) N2 PSp2m (4f ) Sp2m (2f ) Spm (4).2 f O− 2m (4 ) PSp8 (2) PSU2m (q) f PΩ− 2m (2 ) PΩ− 2m (2) PΩ− 2m (4) f PΩ+ 2m (2 ) + PΩ2m (2f ) PΩ+ 2m (q) f PΩ+ 2m (2 ) f PΩ+ 2m (2 ) O+ 8 (2) NL (PSp2m (q)) N1 N1 N1 N1 N1 P1 N1 N1 PSL2 (17) N1 ˆGUm (2f ) Ω− m (4).2 Ω− m (16).2 Pm or Pm−1 NL (Sp2 (2f ) ⊗ Spm (2f )) ˆGUm (q).2 ˆGLm (q).2 f 2 Ω+ m (4 ).2 PΩ+ 2m (4) f PΩ+ 16 (2 ) f PΩ+ 16 (2 ) + PΩ24 (2) f PΩ+ 8 (2 ) + f PΩ8 (2 ) f PΩ+ 8 (2 ) + f PΩ8 (4 ) PΩ+ 8 (3) PΩ+ 8 (3) Ω+ (4) 8 N+ 2 Ω9 (2f ).a N1 N1 Ω7 (2f ) Ω7 (2f ) Ω7 (2f ) Ω7 (4f ) Ω+ 8 (2) Ω+ 8 (2) Ω7 (4) ˆGUm (4).2 N1 Ω9 (2f ).a Co1 P1 , P3 or P4 f ˆ((2f + 1) × Ω− 6 (2 )).2 + f f ˆ((2 − 1) × Ω6 (2 )).2 f Ω− 8 (2 ) P1 , P3 or P4 P13 , P14 or P34 (SL2 (16) × SL2 (16)).22 condition n > 4 even and (n, q) 6= (6, 2) n > 4 even and (n, q) 6= (6, 2) m > 2a > 4 m > 3 and (m, q) 6= (4, 2) m > 2a > 4 m>4 m > 4 even and f 6 2 m > 6 even f 6 2 and mf 6= 3 (m, q) 6= (3, 2) m > 5 odd m > 6 even m > 4 even m>5 m > 6 even m > 6 even m>5 m > 6 even and f 6 2 m > 6 even a62 a62 f >2 f >2 f >2 e n n 2m 2m − 2 2m 2m − 2 2m 2m − 2 m 3 2m 2m − 2 2m − 2 2m − 2 2m − 2 2m − 2 2m − 4 2m − 2 2m − 2 2m − 4 8 14 22 6 3 6 6 6 6 6 Suppose that (L, A ∩ L, B ∩ L) lies in Table 5, so that L = PSLn (q), PSp2m (q), PSU2m (q) or PΩ± 2m (q). For each row in Table 5, take e to be the number as in the last column and take r ∈ ppd(q, e). Then one can verify that r divides |G|/|B| but not |R|. It follows that r divides H due to the factorization G = HB, and then r divides H. Further, r divides H ∩ Soc(A) as r does not divide |Out(Soc(A))|. FACTORIZATIONS OF ALMOST SIMPLE GROUPS 25 However, Hypothesis 5.3 requires H ∩ Soc(A) to be in the third column of Table 1 or Table 2, which implies that |H ∩ Soc(A)| is not divisible by r. This contradiction shows that (L, A ∩ L, B ∩ L) does not lie in Table 5. Thus, we only need to consider cases (i)–(x) in the following. First assume that case (i) appears. From Hypothesis 5.3 we deduce that either a is even, or a = 3 and H ∩ Soc(A) 6 Sp4 (2f b ) × Sp2 (2f b ). The former immediately leads to row 1 of Table 4, so we assume the latter now. Take r ∈ ppd(2, 3f b). Then as r is coprime to both |H| and |R|, r is coprime to |H|. However, r divides f |Sp6b (2f )|/|O− 6b (2 )| = |L|/|B ∩ L|, contradicting the factorization G = HB by Lemma 2.2. For case (ii), we deduce from Hypothesis 5.3 that m is divisible by 4, and so row 2 of Table 4 appears. For case (iii), we deduce from Hypothesis 5.3 that m is even, and so row 3 of Table 4 appears. Suppose that case (iv) occurs. Then either f > 2 and Soc(A) = Sp4 (2f ), or f = 1 and Soc(A) = A6 . From Hypothesis 5.3 we deduce that f > 2 and any prime r in ppd(2, 4f ) is coprime to |H|. It follows that r does not divide |H| since r does not divide |R|. However, r divides |Sp6 (2f )|/|G2(2f )| = |L|/|B ∩ L|, contradicting the factorization G = HB by Lemma 2.2. Suppose that case (v) occurs. Since G = HB and H 6 A, we have the factorization A = H(A ∩ B) by Lemma 2.3. However, A ∩ B = S7 × S3 by [15, 5.1.9], which implies that the factorization A = H(A ∩ B) does not satisfy Theorem 1.1. This is contrary to Hypothesis 5.3. Next suppose that case (vi) occurs. Then as 2m − 2 > 10, we deduce from Hypothesis 5.3 that 2m − 2 is divisible by 4, contrary to the condition that m is even. For case (vii), we deduce from Hypothesis 5.3 that m is divisible by 4, whence row 4 or 5 of Table 4 appears. Suppose that case (viii) occurs. In this case, we have (G, A) = (Ω− 10 (2), A12 ) or − (O10 (2), S12 ). Since G = HB and H 6 A, we have the factorization A = H(A ∩ B) by Lemma 2.3. However, A ∩ B = S8 × S4 by [15, 5.2.16], which implies that the factorization A = H(A ∩ B) does not satisfy Theorem 1.1. This is contrary to Hypothesis 5.3. For case (ix), in view of Ω7 (2f ) ∼ = Sp6 (2f ) we see that row 6 of Table 4 appears. Finally suppose that (x) occurs. Since G = HB and H 6 A, we have the factorization A = H(A ∩ B) by Lemma 2.3. However, A ∩ B ∩ L = 26 .A7 by [15, 5.1.15], which implies that the factorization A = H(A ∩ B) does not satisfy Theorem 1.1. This is contrary to Hypothesis 5.3.  In the forthcoming Lemmas 5.5–5.10, we analyze the six candidates for the triple (L, A ∩ L, B ∩ L) in Table 4 to show that (L, H ∩ L, K ∩ L) lies in Table 1 or Table 2. Note that we have A = H(A ∩ B) and B = (A ∩ B)K from the factorization G = HK. Moreover, we see in Table 4 that both A and B have exactly one nonsolvable composition factor, so both A/Rad(A) and B/Rad(B) are almost simple. Lemma 5.5. Suppose (L, A ∩ L, B ∩ L) lies in row 1 of Table 4. Then (L, H ∩ L, K ∩ L) lies in Table 1. 26 LI AND XIA f Proof. Here L = Sp4ℓ (2f ), A ∩ L = Sp4a (2f b ).b and B ∩ L = O− 4ℓ (2 ) with ab = ℓ − fb and b prime. Since A = H(A ∩ B) and A ∩ B ∩ L = O4a (2 ).b, we infer from Hypothesis 5.3 that Soc(A) is described in the second column of row 1, 2, 4, 5, 9 or 10 of Table 1. Take any r ∈ ppd(2, 4f ℓ). Since B = (A ∩ B)K and B is almost f simple with socle Ω− 4ℓ (2 ), we derive from [15, Theorem A] that either K > Soc(B) or K 6 N1 [B]. For the latter, r divides |H ∩L| because r divides |L| but not |N1 [B]|. Case 1. Let Soc(A) = Sp4a (2f b ) lie in the second column of row 1 of Table 1. In this case, H ∩ Soc(A) = (Sp2a0 (2f bb0 ) × Sp2a0 (2f bb0 )).R0 .2 with R0 6 b0 × b0 , where a0 and b0 are positive integers such that a = a0 b0 . Let a1 = a0 and b1 = b0 b. Then a1 b1 = a0 b0 b = ab = ℓ and H ∩ L = (Sp2a1 (2f b1 ) × Sp2a1 (2f b1 )).R.2 with R 6 b1 × b1 . In particular, |H ∩ L| is not divisible by r, and so K > Soc(B) by the conclusion of f the previous paragraph. It follows that K ∩ L = Ω− 4ℓ (2 ).Q with Q 6 2, and thus (L, H ∩ L, K ∩ L) is as described in row 1 of Table 1. Case 2. Let Soc(A) = Sp4a (2f b ) lie in the second column of row 2 of Table 1. Here a = 3a1 for some integer a1 and H ∩ Soc(A) = (G2 (2f ba1 ) × G2 (2f ba1 )).R1 .2 with R1 6 a1 × a1 . Let ℓ1 = a1 b. It follows that ℓ = ab = 3a1 b = 3ℓ1 , L = Sp12ℓ1 (2f ) and H ∩ L = (G2 (2f ℓ1 ) × G2 (2f ℓ1 )).R.2 with R 6 ℓ1 × ℓ1 . In particular, |H ∩ L| is f not divisible by r, which implies that K > Soc(B). Thus, K ∩ L = Ω− 4ℓ (2 ).Q = f Ω− 12ℓ1 (2 ).Q with Q 6 2, and so (L, H ∩ L, K ∩ L) is as described in row 2 of Table 1. Case 3. Let Soc(A) = Sp4a (2f b ) lie in the second column of row 4 of Table 1. Here a > 2, b = 2, f = 1 and H ∩ Soc(A) = (Sp2 (4) × Sp2a (4)).P with P 6 2. Thus, L = Sp4ℓ (2) = Sp8a (2). Since (H ∩ L)/(H ∩ Soc(A)) ∼ = (H ∩ L)Soc(A)/Soc(A) 6 (A ∩ L)/Soc(A) = b, we have Sp2 (4) × Sp2a (4) 6 H ∩ L 6 (Sp2 (4) × Sp2a (4)).[4]. In particular, |H ∩ L| is not divisible by r, and so K > Soc(B). Then we conclude that K ∩ L = Ω− 4ℓ (2).Q = Ω− (2).Q with Q 6 2 and (L, H ∩ L, K ∩ L) is as described in row 5 of Table 1. 8a fb Case 4. LetSoc(A) = Sp4a (2 ) lie in the second column of row 5 or 10 of Table 1. In this case we have f b = 1, contrary to the condition that b is prime. Case 5. Let Soc(A) = Sp4a (2f b ) lie in the second column of row 9 of Table 1. In this case, a = 3, b = 2, f = 1 and H ∩ Soc(A) = (Sp2 (4) × G2(4)).P with P 6 2. As H ∩ Soc(A) 6 H ∩ L 6 (H ∩ Soc(A)).b, we derive that H ∩ L = (Sp2 (4) × G2 (4)).P1 , where P1 6 22 . It follows that |H ∩ L| is not divisible by r, and then K > Soc(B). − Thus, K ∩ L = Ω− 4ℓ (2).Q = Ω24 (2).Q with Q 6 2, and so (L, H ∩ L, K ∩ L) lies in row 10 of Table 1.  Lemma 5.6. Suppose (L, A ∩ L, B ∩ L) lies in row 2 of Table 4. Then (L, H ∩ L, K ∩ L) lies in Table 1. − Proof. Here G = Sp8ℓ (2), A = O+ 8ℓ (2) and B = O8ℓ (2). Since A = H(A ∩ B) and A ∩ B = Sp8ℓ−2 (2) × 2 = N1 [A] (see [15, 3.2.4(e)]), Hypothesis 5.3 implies that Soc(A) is described in the second column of row 2 or 5 of Table 2. Take any r ∈ ppd(2, 8ℓ − 2). Since B = (A ∩ B)K and B is almost simple with socle Ω− 8ℓ (2), we derive from [15, Theorem A] that either K > Soc(B) or K ∩ Soc(B) 6 Ω− (4).2. 4ℓ For the latter, r divides |H ∩ L| because r divides |L| but not |Ω− 4ℓ (4).2|. Case 1. Let Soc(A) = Ω+ 8ℓ (2) lie in the second column of row 2 of Table 2. In this case, H ∩ Soc(A) = (Sp2 (4) × Sp2ℓ (4)).P with P 6 22 . Then since (H ∩ L)/(H ∩ Soc(A)) ∼ = (H ∩ L)Soc(A)/Soc(A) 6 (A ∩ L)/Soc(A) = 2, FACTORIZATIONS OF ALMOST SIMPLE GROUPS 27 we have Sp2 (4) × Sp2ℓ (4) 6 H ∩ L 6 (Sp2 (4) × Sp2ℓ (4)).[8]. In particular, |H ∩ L| is not divisible by r, which implies K > Soc(B) by the conclusion of the previous paragraph. Hence K ∩ L = Ω− 8ℓ (2).Q with Q 6 2 and so (L, H ∩ L, K ∩ L) lies in row 5 of Table 1. Case 2. Let Soc(A) = Ω+ 8ℓ (2) lie in the second column of row 5 of Table 2. In this case, ℓ = 3 and H ∩ Soc(A) = (Sp2 (4) × G2 (4)).P with P 6 22 . Since H ∩ Soc(A) 6 H ∩ L 6 (H ∩ Soc(A)).2, we have Sp2 (4) × G2 (4) 6 H ∩ L 6 (Sp2 (4) × G2 (4)).[8]. In particular, |H ∩ L| is not divisible by r, and so K > Soc(B). It follows that − K ∩ L = Ω− 8ℓ (2).Q = Ω24 (4).Q with Q 6 2, and thus (L, H ∩ L, K ∩ L) lies in row 10 of Table 1.  Lemma 5.7. Suppose (L, A ∩ L, B ∩ L) lies in row 3 of Table 4. Then (L, H ∩ L, K ∩ L) lies in Table 1. − Proof. Here ℓ > 2 and (G, A, B) = (ΓSp4ℓ (4), ΓO+ 4ℓ (4), ΓO4ℓ (4)) according to [15, Theorem A]. Since A = H(A ∩ B) and A ∩ B = N1 [A ∩ L] = Sp4ℓ−2 (4) × 2 (see [15, 3.2.4(e)]), we derive from Hypothesis 5.3 that Soc(A) lies in the second column of row 1, 3, 4 or 6 of Table 2. Take any r ∈ ppd(4, 4ℓ − 2). Since B = (A ∩ B)K and B is almost simple with socle Ω− 4ℓ (4), it follows from [15, Theorem A] that either K > Soc(B) or K ∩ Soc(B) 6 Ω− 2ℓ (16).2. For the latter, r divides |H ∩ L| because r divides |L| but not |Ω− (16).2|. 2ℓ Case 1. Let Soc(A) = Ω+ 4ℓ (4) lie in the second column of row 1 of Table 2. In this case, H ∩ Soc(A) = U × Sp2ℓ (4), where U is an nonsolvable subgroup of Sp2 (4). Therefore, U = Sp2 (4) and H ∩ Soc(A) = Sp2 (4) × Sp2ℓ (4). Since (H ∩ L)/(H ∩ Soc(A)) ∼ = (H ∩ L)Soc(A)/Soc(A) 6 (A ∩ L)/Soc(A) = 2, we deduce that Sp2 (4) × Sp2ℓ (4) 6 H ∩ L 6 (Sp2 (4) × Sp2ℓ (4)).2. In particular, |H ∩ L| is not divisible by r, and so K > Soc(B) by the conclusion of the previous paragraph. Hence K ∩ L = Ω− 4ℓ (4).Q with Q 6 2, and thus (L, H ∩ L, K ∩ L) lies in row 4 of Table 1. Case 2. Let Soc(A) = Ω+ 4ℓ (4) lie in the second column of row 3 or 6 of Table 2. In this case we have ℓ = 2ℓ1 for some integer ℓ1 . Moreover, either H ∩ Soc(A) = (Sp2 (4c ) × Sp2ℓ1 (16)).P , or ℓ1 = 3 and H ∩ Soc(A) = (Sp2 (4c ) × G2 (16)).P , where c = 1 or 2 and P 6 22 . As A ∩ B = N1 [A ∩ L] < A ∩ L (see [15, 3.2.4(e)]) and A = H(A ∩ B), we have a factorization A ∩ L = (H ∩ L)N1 [A ∩ L] ∗ ∗ of A ∩ L = O+ 8ℓ1 (4). By Lemma 2.5, there exist groups A and H such that + ∗ ∗ ∗ ∗ ∗ ∗ Ω+ 8ℓ1 (4) 6 A 6 O8ℓ1 (4), A = H Soc(A) = H N1 [A ] and H ∩Soc(A) = H∩Soc(A). Let M be a maximal subgroup of A∗ containing H ∗ . Then A∗ = MN1 [A∗ ], and M is core-free in A∗ . From Hypothesis 5.3 we see that M cannot have a linear, unitary or plus type orthogonal group as its unique nonsolvable composition factor. Thus one of the following occurs by [15, Theorem A]. (i) M ∩ Soc(A∗ ) = Sp2 (4) ⊗ Sp4ℓ1 (4). (ii) M ∩ Soc(A∗ ) = Ω4ℓ1 (16).22 . (iii) ℓ1 = 2 and M ∩ Soc(A∗ ) = Sp8 (4).a with a 6 2. (iv) ℓ1 = 1 and M ∩ Soc(A∗ ) = Sp6 (4). (v) ℓ1 = 1 and M ∩ Soc(A∗ ) = (Sp2 (16) × Sp2 (16)).22 . 28 LI AND XIA If (i) occurs, then H ∗ > Sp4ℓ1 (4) by Lemma 5.2, which is not possible since we have either H ∗ ∩ Soc(A) = (Sp2 (4c ) × Sp2ℓ1 (16)).P or ℓ1 = 3 and H ∗ ∩ Soc(A) = (Sp2 (4c )×G2 (16)).P . If M ∩Soc(A∗ ) is as described in (ii) or (v), then the argument in [15, 3.6.1(c)] shows that A∗ 6= MN1 [A∗ ], a contradiction. If (iii) occurs, then we see from Hypothesis 5.3 and the third column of Table 1 that H ∗ cannot have Sp2 (4c ) × Sp2ℓ1 (16) = Sp2 (4c ) × Sp4 (16) as a normal subgroup, a contradiction. Finally, suppose that (iv) occurs. Then M ∩ N1 [A∗ ] has a normal subgroup G2 (4) by [15, 5.1.15] while H ∗ has a normal subgroup Sp2 (4c ) × Sp2 (16). By Lemma 2.3, we deduce from A∗ = H ∗ N1 [A∗ ] and H ∗ 6 M that M = H ∗ (M ∩ N1 [A∗ ]). However, M does not have such a factorization according to Hypothesis 5.3, a contradiction. Case 3. Let Soc(A) = Ω+ 4ℓ (4) lie in the second column of row 4 of Table 2. In this case, ℓ = 3 and H ∩ Soc(A) = (Sp2 (4) × G2 (4)). Since H ∩ Soc(A) 6 H ∩ L 6 (H ∩Soc(A)).2, we have Sp2 (4) ×G2 (4) 6 H ∩L 6 (Sp2 (4) ×G2 (4)).2. In particular, − |H∩L| is not divisible by r, and so K > Soc(B). Hence K∩L = Ω− 4ℓ (4).Q = Ω12 (4).Q with Q 6 2, and then (L, H ∩ L, K ∩ L) lies in row 9 of Table 1.  Lemma 5.8. Suppose (L, A ∩ L, B ∩ L) lies in row 4 of Table 4. Then (L, H ∩ L, K ∩ L) lies in Table 2. + 2 Proof. Here L = Ω+ 8ℓ (2) with ℓ > 2, A ∩ L = Ω4ℓ (4).2 and B ∩ L = Sp8ℓ−2 (2). Furthermore, we have G = L in order that A is maximal in G. By Lemma 3.9, A ∩ B = Sp4ℓ−2 (4) × 2 < O+ 4ℓ (4). Then as A = H(A ∩ B), we derive from Hypothesis 5.3 that Soc(A) is described in the second column of row 1, 3, 4 or 6 of Table 2. Since B = (A ∩ B)K and |A ∩ B| has no prime divisor in the set ppd(2, 8ℓ − 2) ∪ ppd(2, 8ℓ − 6) ∪ ppd(2, 4ℓ − 1), we conclude that |K| is divisible by each prime in ppd(2, 8ℓ − 2) ∪ ppd(2, 8ℓ − 6) ∪ ppd(2, 4ℓ − 1). Hence by [15, Theorem A], the factorization B = (A ∩ B)K implies that K > Soc(B), and so K ∩ L = Sp8ℓ−2 (2). Case 1. Let Soc(A) = Ω+ 4ℓ (4) lie in the second column of row 1 of Table 2. In this case, H ∩ Soc(A) = U × Sp2ℓ (4), where U is an nonsolvable subgroup of Sp2 (4). Thus, U = Sp2 (4) and H ∩ Soc(A) = Sp2 (4) × Sp2ℓ (4). Since (H ∩ L)/(H ∩ Soc(A)) ∼ = (H ∩ L)Soc(A)/Soc(A) 6 (A ∩ L)/Soc(A) = 22 , we deduce that Sp2 (4) × Sp2ℓ (4) 6 H ∩ L 6 (Sp2 (4) × Sp2ℓ (4)).22. Therefore, (L, H ∩ L, K ∩ L) lies in row 2 of Table 2. Case 2. Let Soc(A) = Ω+ 4ℓ (4) lie in the second column of row 3 or 6 of Table 2. Since A = H(A ∩ B) and A ∩ B = Sp4ℓ−2 (4) × 2 6 N1 [O+ 4ℓ (4)], we derive that + + + O+ 4ℓ (4) = (H ∩ O4ℓ (4))(A ∩ B) = (H ∩ O4ℓ (4))N1 [O4ℓ (4)]. Then one obtains a contradiction along the same lines as in Case 2 of the proof of Lemma 5.7. Case 3. Let Soc(A) = Ω+ 4ℓ (4) lie in the second column of row 4 of Table 2. In this case, ℓ = 3 and H ∩ Soc(A) = (Sp2 (4) × G2 (4)). As H ∩ Soc(A) 6 H ∩ L 6 (H ∩ Soc(A)).22 , we have Sp2 (4) × G2 (4) 6 H ∩ L 6 (Sp2 (4) × G2 (4)).22 . Thus, (L, H ∩ L, K ∩ L) lies in row 5 of Table 2.  Lemma 5.9. Suppose (L, A ∩ L, B ∩ L) lies in row 5 of Table 4. Then (L, H ∩ L, K ∩ L) lies in Table 2. + 2 Proof. Here L = Ω+ 8ℓ (4) with ℓ > 2, A∩L = Ω4ℓ (16).2 and B ∩L = Sp8ℓ−2 (4). Since A = H(A ∩ B) and A ∩ B ∩ L = Sp4ℓ−2 (16), we derive from Hypothesis 5.3 that FACTORIZATIONS OF ALMOST SIMPLE GROUPS 29 Soc(A) is described in the second column of row 1 or 4 of Table 2. As B = (A∩B)K and |A ∩ B| has no prime divisor in ppd(4, 8ℓ − 2) ∪ ppd(4, 8ℓ − 6) ∪ ppd(4, 4ℓ − 1), we conclude that |K| is divisible by each prime in ppd(4, 8ℓ − 2) ∪ ppd(4, 8ℓ − 6) ∪ ppd(4, 4ℓ − 1). Then by [15, Theorem A], the factorization B = (A ∩ B)K implies that K ∩ L = Sp8ℓ−2 (4). Case 1. Let Soc(A) = Ω+ 4ℓ (16) lie in the second column of row 1 of Table 2. In this case, H ∩ Soc(A) = U × Sp2ℓ (16), where U is an nonsolvable subgroup of Sp2 (16). It follows that U = Sp2 (4) or Sp2 (16), and then H ∩ Soc(A) = Sp2 (4c ) × Sp2ℓ (16) with c = 1 or 2. Since (H ∩ L)/(H ∩ Soc(A)) ∼ = (H ∩ L)Soc(A)/Soc(A) 6 (A ∩ L)/Soc(A) = 22 , we have Sp2 (4c ) × Sp2ℓ (4) 6 H ∩ L 6 (Sp2 (4c ) × Sp2ℓ (4)).22 . Thus, (L, H ∩ L, K ∩ L) lies in row 3 of Table 2. Case 2. Let Soc(A) = Ω+ 4ℓ (16) lie in the second column of row 4 of Table 2. In this case, ℓ = 3 and H ∩ Soc(A) = Sp2 (4c ) × G2 (16) with c = 1 or 2. As H ∩ Soc(A) 6 H ∩L 6 (H ∩Soc(A)).22 , we have Sp2 (4c )×G2 (16) 6 H ∩L 6 (Sp2 (4c )×G2 (16)).22. Hence (L, H ∩ L, K ∩ L) lies in row 6 of Table 2.  Lemma 5.10. Suppose (L, A ∩ L, B ∩ L) lies in row 6 of Table 4. Then (L, H ∩ L, K ∩ L) lies in Table 2. f f ∼ Proof. Here L = Ω+ 8 (2 ) with f > 2 and A ∩ L = B ∩ L = Sp6 (2 ). Since A = H(A ∩ B) and A ∩ B ∩ L = G2 (2f ) (see [15, 5.1.15]), we derive from Hypothesis 5.3 that A ∩ L is described in the second column of row 7 of Table 1. It follows that H ∩ L = H ∩ (A ∩ L) = U × Sp4 (2f ), where U is an nonsolvable subgroup of Sp2 (2f ). In particular, |H ∩ L| is not divisible by any prime in ppd(2, 6f ) ∪ ppd(2, 3f ). In view of the factorization B = (A ∩ B)K we deduce from [15, Theorem A] that − f f f either K ∩ L = Sp6 (2f ), or K ∩ L is contained in O+ 6 (2 ), O6 (2 ), P1 [Sp6 (2 )] or N1 [Sp6 (2f )]. For the latter, there exists r ∈ ppd(2, 6f ) ∪ ppd(2, 3f ) dividing |L|/|K ∩ L|, which indicates that r divides |H ∩ L| by Lemma 2.2, a contradiction. Therefore, K ∩ L = Sp6 (2f ), and so (L, H ∩ L, K ∩ L) lies in row 1 of Table 2.  Now we are in the position to complete the proof of Theorem 1.1. Proof of Theorem 1.1. Suppose that G is an almost simple group with socle L and G = HK, where H has at least two nonsolvable composition factors and K is corefree. We prove Theorem 1.1 by induction on |G|. By Lemma 2.7 we may assume that L is a classical group. Moreover, by virtue of Lemma 2.5 we may let A and B be maximal subgroups of G containing H and K, respectively. Thus we have both Hypothesis 2.8 and Hypothesis 5.3. Then according to Lemma 5.4, either A has at least two nonsolvable composition factors, or the triple (L, A ∩ L, B ∩ L) lies in Table 4. If A has at least two nonsolvable composition factors, then Lemma 4.4 asserts that (L, H ∩ L, K ∩ L) lies in Table 1 or Table 2. If (L, A ∩ L, B ∩ L) lies in Table 4, then it follows from Lemmas 5.5–5.10 that (L, H ∩ L, K ∩ L) lies in Table 1 or Table 2. Hence Theorem 1.1 is true.  References [1] M. Aschbacher, On the maximal subgroups of the finite classical groups, Invent. Math., 76 (1984), no. 3, 469–514. 30 LI AND XIA [2] N. Blackburn and B. Huppert, Finite groups II, Springer-Verlag, Berlin-New York, 1982. [3] W. Bosma, J. Cannon and C. Playoust, The magma algebra system I: The user language, J. Symbolic Comput., 24 (1997), no. 3-4, 235–265. [4] J. N. Bray, D. F. Holt and C. M. Roney-Dougal, The maximal subgroups of the low-dimensional finite classical groups, Cambridge University Press, Cambridge, 2013. [5] P. J. Cameron, Permutation groups, London Mathematical Society Student Texts, 45, Cambridge University Press, Cambridge, 1999. [6] J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker and R. A. Wilson, Atlas of finite groups: maximal subgroups and ordinary characters for simple groups, Clarendon Press, Oxford, 1985. [7] M. Fried, R. Guralnick and J. Saxl, Schur covers and Carlitz’s conjecture, Israel J. Math. 82 (1993), no. 1-3, 157–225. [8] M. Giudici, Factorisations of sporadic simple groups, J. Algebra, 304 (2006), no. 1, 311–323. [9] R. M. Guralnick and J. Saxl, Monodromy groups of polynomials, Groups of Lie type and their geometries (Como, 1993), 125–150, London Math. Soc. Lecture Note Ser., 207, Cambridge Univ. Press, Cambridge, 1995. [10] C. Hering, M. W. Liebeck and J. Saxl, The factorizations of the finite exceptional groups of lie type, J. Algebra, 106 (1987), no. 2, 517–527. [11] B. Huppert, Endliche Gruppen. I, Die Grundlehren der Mathematischen Wissenschaften, Band 134, Springer-Verlag, Berlin-New York 1967. [12] W. M. Kantor, k-homogeneous groups, Math. Z., 124 (1972), 261–265. [13] P. B. Kleidman and M. W. Liebeck, The subgroup structure of the finite classical groups, Cambridge University Press, Cambridge, 1990. [14] C. H. Li and B. Xia, Factorizations of almost simple groups with a solvable factor, and Cayley graphs of solvable groups, submitted. Available online at https://arxiv.org/abs/1408.0350. [15] M. Liebeck, C. E. Praeger and J. Saxl, The maximal factorizations of the finite simple groups and their automorphism groups, Mem. Amer. Math. Soc., 86 (1990), no. 432. [16] M. W. Liebeck, C. E. Praeger and J. Saxl, On factorizations of almost simple groups, J. Algebra, 185 (1996), no. 2, 409–419. [17] M. Liebeck, C. E. Praeger and J. Saxl, Regular subgroups of primitive permutation groups, Mem. Amer. Math. Soc., 203 (2010), no. 952. [18] M. Suzuki, A new type of simple groups of finite order, Proc. Nat. Acad. Sci. U.S.A., 46 (1960), 868–870. [19] R. Wilson, The finite simple groups, Graduate Texts in Mathematics, 251. Springer, 2009. (Li) Department of Mathematics, South University of Science and Technology, Shenzhen, Guangdong 518055, P. R. China E-mail address: [email protected] (Xia) School of Mathematics and Statistics, The University of Melbourne, Parkville, VIC 3010, Australia E-mail address: [email protected]
4math.GR
Accepted by VLDB 2018 Model-Free Control for Distributed Stream Data Processing using Deep Reinforcement Learning Teng Li, Zhiyuan Xu, Jian Tang and Yanzhi Wang {tli01, zxu105, jtang02, ywang393}@syr.edu arXiv:1803.01016v1 [cs.DC] 2 Mar 2018 Department of Electrical Engineering and Computer Science Syracuse University Syracuse, NY 13244 ABSTRACT ∗ framework can quickly reach a good scheduling solution during online learning, which justifies its practicability for online control in DSDPSs. In this paper, we focus on general-purpose Distributed Stream Data Processing Systems (DSDPSs), which deal with processing of unbounded streams of continuous data at scale distributedly in real or near-real time. A fundamental problem in a DSDPS is the scheduling problem (i.e., assigning workload to workers/machines) with the objective of minimizing average end-to-end tuple processing time. A widelyused solution is to distribute workload evenly over machines in the cluster in a round-robin manner, which is obviously not efficient due to lack of consideration for communication delay. Model-based approaches (such as queueing theory) do not work well either due to the high complexity of the system environment. We aim to develop a novel model-free approach that can learn to well control a DSDPS from its experience rather than accurate and mathematically solvable system models, just as a human learns a skill (such as cooking, driving, swimming, etc). Specifically, we, for the first time, propose to leverage emerging Deep Reinforcement Learning (DRL) for enabling model-free control in DSDPSs; and present design, implementation and evaluation of a novel and highly effective DRL-based control framework, which minimizes average end-to-end tuple processing time by jointly learning the system environment via collecting very limited runtime statistics data and making decisions under the guidance of powerful Deep Neural Networks (DNNs). To validate and evaluate the proposed framework, we implemented it based on a widely-used DSDPS, Apache Storm, and tested it with three representative applications: continuous queries, log stream processing and word count (stream version). Extensive experimental results show 1) Compared to Storm’s default scheduler and the state-of-the-art model-based method, the proposed framework reduces average tuple processing by 33.5% and 14.0% respectively on average. 2) The proposed 1. INTRODUCTION In this paper, we focus on general-purpose Distributed Stream Data Processing Systems (DSDPSs) (such as Apache Storm [48] and Google’s MillWheel [3]), which deal with processing of unbounded streams of continuous data at scale distributedly in real or near-real time. Their programming models and runtime systems are quite different from those of MapReduce-based batch processing systems, such as Hadoop [22] and Spark [46], which usually handle static big data in an offline manner. To fulfill the real or near-real time online processing requirements, average end-to-end tuple processing time (or simply average tuple processing time [52]) is the most important performance metric for a DSDPS. A fundamental problem in a DSDPS is the scheduling problem (i.e., assigning workload to workers/machines) with the objective of minimizing average tuple processing time. A widely-used solution is to distribute workload over machines in the cluster in a round-robin manner [48], which is obviously not efficient due to lack of consideration for communication delay among processes/machines. We may be able to better solve this problem if we can accurately model the correlation between a solution and its objective value, i.e., predict/estimate average tuple processing time for a given scheduling solution. However, this is very hard and has not yet been well studied in the context of DSDPS. In distributed batch processing systems (such as MapReducebased systems), an individual task’s completion time can be well estimated [53], and the scheduling problem can be well formulated into a Mixed Integer Linear Programming (MILP) problem with the objective of minimizing the makespan of a job [13, 55]. Then it can be tackled by optimally solving the MILP problem or using a fast polynomial-time approximation/heuristic algorithm. However, in a DSDPS, a task or an application never ends unless it is terminated by its user. A scheduling solution makes a significant impact on the average tuple processing time. But their relationship is very subtle and complicated. It does not even seem possible to have a mathematical programming formulation for the scheduling problem if its objective is to directly minimize average tuple processing time. Queueing theory has been employed to model distributed stream database systems [34]. However, it does not work for ∗Accepted by VLDB 2018 1 DSDPSs due to the following reasons: 1) The queueing theory can only provide accurate estimations for queueing delay under a few strong assumptions (e.g, tuple arrivals follow a Poisson distribution, etc), which, however, may not hold in a complex DSDPS. 2) In the queueing theory, many problems in a queueing network (rather than a single queue) remain open problems, while a DSDPS represents a fairly complicated multi-point to multi-point queueing network where tuples from a queue may be distributed to multiple downstream queues, and a queue may receive tuples from multiple different upstream queues. In addition, a model-based scheduling framework has been proposed in a very recent work [25], which employs a model to estimate (end-to-end) tuple processing time by combing delay at each component (including processing time at each Processing Unit (PU) and communication delay between two PUs) predicted by a supervised learning method, Support Vector Regression (SVR) [14]. However, this model-based approach suffers from two problems: 1) In a very complicated distributed computing environment such as a DSDPS, (end-to-end) tuple processing time (i.e., end-to-end delay) may be caused by many factors, which are not fully captured by the proposed model. 2) Prediction for each individual component may not be accurate. 3) A large amount of high-dimensional statistics data need to be collected to build and update the model, which leads to high overhead. Hence, we aim to develop a novel model-free approach that can learn to well control a DSDPS from its experience rather than accurate and mathematically solvable system models, just as a human learns a skill (such as cooking, driving, swimming, etc). Recent breakthrough of Deep Reinforcement Learning (DRL) [33] provides a promising technique for enabling effective model-free control. DRL [33] (originally developed by a startup DeepMind) enables computers to learn to play games, including Atari 2600 video games and one of the most complicated games, Go (AlphaGo [45]), and beat the best human players. Even though DRL has made tremendous successes on game-playing that usually has a limited action space (e.g., moving up/down/left/right), it has not yet been investigated how DRL can be leveraged for control problems in complex distributed computing systems, such as DSDPSs, which usually have sophisticated state and huge action spaces. We believe DRL is especially promising for control in DSDPSs because: 1) It has advantages over other dynamic system control techniques such as model-based predictive control in that the former is model-free and does not rely on accurate and mathematically solvable system models (such as queueing models), thereby enhancing the applicability in complex systems with randomized behaviors. 2) it is capable of handling a sophisticated state space (such as AlphaGo [45]), which is more advantageous over traditional Reinforcement Learning (RL) [49]. 3) It is able to deal with time-variant environments such as varying system states and user demands. However, direct application of the basic DRL technique, such as Deep Q Network (DQN) based DRL proposed in the pioneering work [33], may not work well here since it is only capable of handling control problems with a limited action space but the control problems (such as scheduling) in a DSDPS usually have a sophisticated state space, and a huge action (worker threads, worker processes, virtual/physical machines, and their combinations) space (See Section 3.2 for greater details). Moreover, the exist- ing DRL methods [33] usually need to collect big data (e.g., lots of images for the game-playing applications) for learning, which are additional overhead and burden for an online system. Our goal is to develop a method that only needs to collect very limited statistics data during runtime. In this paper, we aim to develop a novel and highly effective DRL-based model-free control framework for a DSDPS to minimize average data processing time by jointly learning the system environment with very limited runtime statistics data and making scheduling decisions under the guidance of powerful Deep Neural Networks (DNNs). We summarize our contributions in the following: • We show that direct application of DQN-based DRL to scheduling in DSDPSs does not work well. • We are the first to present a highly effective and practical DRL-based model-free control framework for scheduling in DSDPSs. • We show via extensive experiments with three representative Stream Data Processing (SDP) applications that the proposed framework outperforms the current practice and the state-of-the-art. To the best of our knowledge, we are the first to leverage DRL for enabling model-free control in DSDPSs. We aim to promote a simple and practical model-free approach based on emerging DRL, which, we believe, can be extended or modified to better control many other complex distributed computing systems. The rest of the paper is organized as follows: We give a brief background introduction in Section 2. We then present design and implementation details of the proposed framework in Section 3. Experimental settings are described, and experimental results are presented and analyzed in Section 4. We discuss related work in Section 5 and conclude the paper in Section 6. 2. BACKGROUND In this section, we provide a brief background introduction to DSDPS, Storm and DRL. 2.1 Distributed Stream Data Processing System (DSDPS) In a DSDPS, a stream is an unbounded sequence of tuples. A data source reads data from external source(s) and emits streams into the system. A Processing Unit (PU) consumes tuples from data sources or other PUs, and processes them using code provided by a user. After that, it can pass it to other PUs for further processing. A DSDPS usually uses two levels of abstractions (logical and physical) to express parallelism. In the logical layer, an application is usually modeled as a directed graph, in which each vertex corresponds to a data source or a PU, and direct edges show how data tuples are routed among data sources/PUs. A task is an instance of a data source or PU, and each data source or PU can be executed as many parallel tasks on a cluster of machines. In the physical layer, a DSDPS usually includes a set of virtual or physical machines that actually process incoming data, and a master serving as the central control unit, which distributes user code around the cluster, scheduling tasks, and monitoring them for failures. At runtime, an application graph is executed on multiple worker processes running on multiple (physical or virtual) machines. Each machine is usually configured to have multiple slots. The number of slots indicates the number of worker processes that can be run on this machine, and can be pre-configured by the cluster operator based on hardware constraints (such as the number of CPU cores). Each worker process occupies a slot, which uses one or multiple threads to actually process data tuples using user code. Normally, a task is mapped to a thread at runtime (even it does not have to be this way). Each machine also runs a daemon that listens for any work assigned to it by the master. In a DSDPS, a scheduling solution specifies how to assign threads to processes and machines. Many DSDPSs include a default scheduler but allow it to be replaced by a custom scheduler. The default scheduler usually uses a simple scheduling solution, which assigns threads to pre-configured processes and then assigns those processes to machines both in a roundrobin manner. This solution leads to almost even distribution of workload over available machines in the cluster. In addition, a DSDPS usually supports several ways for grouping, which defines how to distribute tuples among tasks. Typical grouping policies include: fields grouping (based on a key), shuffle grouping (random), all grouping (one-to-all) and global grouping (all-to-one). When the message ID of a tuple coming out of a spout successfully traverses the whole topology, a special acker is called to inform the originating data source that message processing is complete. The (end-to-end) tuple processing time is the duration between when the data source emits the tuple and when it has been acked (fully processed). Note that we are only interested in processing times of those tuples emitted by data sources since they reflect the total end-to-end processing delay over the whole application. To ensure fault tolerance, if a message ID is marked failure due to acknowledgment timeout, data processing will be recovered by replaying the corresponding data source tuple. The master monitors heartbeat signals from all worker processes periodically. It re-schedules them when it discovers a failure. 2.2 Apache Storm Since we implemented the proposed framework based on Apache Storm [48], we briefly introduce it here. Apache Storm is an open-source and fault-tolerant DSDPS, which has an architecture and programming model very similar to what described above, and has been widely used by quite a few companies and institutes. In Storm, data source, PU, application graph, master, worker process and worker thread are called spout, bolt, topology, Nimbus, worker and executor, respectively. Storm uses ZooKeeper [54] as a coordination service to maintain it’s own mutable configuration (such as scheduling solution), naming, and distributed synchronization among machines. All configurations stored in ZooKeeper are organized in a tree structure. Nimbus (i.e., master) provides interfaces to fetch or update Storm’s mutable configurations. A Storm topology contains a topology specific configuration, which is loaded before the topology starts and does not change during runtime. 2.3 Deep Reinforcement Learning (DRL) A DRL agent can be trained via both offline training and online deep Q-learning [33, 45]. It usually adopts a DNN (known as DQN) to derive the correlation between each state-action pair (s, a) of the system under control and its value function Q(s, a), which is the expected cumulative (with discounts) reward function when system starts at state s and follows action a (and certain policy thereafter). Q(s, a) is given as: Q(s, a) = E ∞ hX i λk rt (st , at ) s0 = s, a0 = a , (2.1) t=0 where rt (·) is the reward, and λ < 1 is the discount factor. The offline training needs to accumulate enough samples of value estimates and the corresponding state-action pair (s, a) for constructing a sufficiently accurate DNN using either a model-based (a mathematical model) procedure or actual measurement data (model-free) [33]. For example, in game-playing applications [33], this procedure includes preprocessing game playing samples, and obtaining state transition samples and Q-value estimates (e.g., win/lose and/or the score achieved). The deep Q-learning is adopted for the online learning and dynamic control based on the offlinebuilt DNN. More specifically, at each decision epoch t, the system under control is at a state st . The DRL agent performs inference using the DNN to select action at , either the one with the highest Q-value estimate, or with a certain degree of randomness using the ǫ-greedy policy [41]. Using a neural network (or even a DNN) as a function approximator in RL is known to suffer from instability or even divergence. Hence, experience replay and target network were introduced in [33] to improve stability. A DRL agent updates the DNN with a mini-batch from the experience replay buffer [33], which stores state transition samples collected during training. Compared to using only immediately collected samples, uniformly sampling from the replay buffer allows the DRL agent to break the correlation between sequential generated samples, and learn from a more independently and identically distributed past experiences, which is required by most of training algorithms, such as Stochastic Gradient Descent (SGD). So the use of experience replay buffer can smooth out learning and avoid oscillations or divergence. Besides, a DRL agent uses a separate target network (with the same structure as the original DNN) to estimate target values for training the DNN. Its parameters are slowly updated every C > 1 epochs and are held fixed between individual updates. The DQN-based DRL only works for control problems with a low-dimensional discrete action space. Continuous control has often been tackled by the actor-critic-based policy gradient approach [44]. The traditional actor-critic approach can also be extended to use a DNN (such as DQN) to guide decision making [26]. A recent work [26] from DeepMind introduced an actor-critic method, called Deep Deterministic Policy Gradient (DDPG), for continuous control. The basic idea is to maintain a parameterized actor function and a parameterized critic function. The critic function can be implemented using the above DQN, which returns Q value for a given state-action pair. The actor function can also be implemented using a DNN, which specifies the current policy by mapping a state to a specific action. Both the experience replay and target network introduced above can also be integrated to this approach to ensure stability. 3. DESIGN AND IMPLEMENTATION OF THE PROPOSED FRAMEWORK In this section, we present the design and implementation details of the proposed framework. 3.1 Overview Master Custom Scheduler Machine Machine Processes DSDPS Processes Process …... Process Thread Thread Thread Thread State and Reward DRL-based Control Thread Database Transition Samples DRL Agent Scheduling Solution Figure 1: The architecture of the proposed DRLbased control framework We illustrate the proposed framework in Figure 1, which can be viewed to have two parts: DSDPS and DRL-based Control. The architecture is fairly simple and clean, which consists of the following components: 1) DRL Agent (Section 3.2): it is the core of the proposed framework, which takes the state as input, applies a DRL-based method to generating a scheduling solution, and pushes it to the custom scheduler. 2) Database: It stores transition samples including state, action and reward information for training (See Section 3.2 for details). 3) Custom Scheduler: It deploys the generated scheduling solution on the DSDPS via the master. Our design leads to the following desirable features: 1) Model-free Control: Our design employs a DRL-based method for control, which learns to control a DSDPS from runtime statistics data without relying on any mathematically solvable system model. 2) Highly Effective Control: The proposed DRL-based control is guided by DNNs, aiming to directly minimize average tuple processing time. Note that the current practise evenly distributes workload over machines; and some existing methods aim to achieve an indirect goal, (e.g., minimizing inter-machine traffic load [52]), with the hope that it can lead to minimum average tuple processing time. These solutions are obviously less convincing and effective than the proposed approach. 3) Low Control Overhead: The proposed framework only needs to collect very limited statistics data, i.e., just the average tuple processing time, during runtime for offline training and online learning (see explanations in Section 3.2), which leads to low control overhead. 4) Hot Swapping of Control Algorithms: The core component of the proposed framework, DRL agent, is external to the DSDPS, which ensures minimum modifications to the DSDPS, and more importantly, makes it possible to replace it or its algorithm at runtime without shutting down the DSDPS. 5) Transparent to DSDSP users: The proposed framework is completely transparent to DSDSP users, i.e., a user does not have to make any change to his/her code in order to run his/her application on the new DSDPS with the proposed framework. We implemented the proposed framework based on Apache Storm [48]. In our implementation, the custom scheduler runs within Nimbus, which has access to various information regarding executors, supervisors and slots. A socket is implemented for communications between the custom scheduler and the DRL agent. When an action is generated by the DRL agent, it is translated to a Storm-recognizable scheduling solution and pushed to the custom scheduler. Upon receiving a scheduling solution, the custom scheduler first frees the executors that need to be re-assigned and then adds them to the slots of target machines. Note that during the deployment of a new scheduling solution, we try to make a minimal impact to the DSDPS by only re-assigning those executors whose assignments are different from before while keeping the rest untouched (instead of deploying the new scheduling solution from scratch by freeing all the executors first and assigning executors one by one as Storm normally does). In this way, we can reduce overhead and make the system to re-stabilize quickly. In addition, to make sure of accurate data collection, after a scheduling solution is applied, the proposed framework waits for a few minutes until the system re-stabilizes to collect the average tuple processing time and takes the average of 5 consecutive measurements with a 10-second interval. 3.2 DRL-based Control In this section, we present the proposed DRL-based control, which targets at minimizing the end-to-end average tuple processing time via scheduling. Given a set of machines M, a set of processes P, and a set of threads N , a scheduling problem in a DSDPS is to assign each thread to a process of a machine, i.e., to find two mappings: N 7→ P and P 7→ M. It has been shown [52] that assigning threads from an application (which usually exchange data quite often) to more than one processes on a machine introduces inter-process traffic, which leads to serious performance degradation. Hence, similar as in [52, 25], our design ensures that on every machine, threads from the same application are assigned to only one process. Therefore the above two mappings can be merged into just one mapping: N 7→ M, i.e., to assign each thread to a machine. Let a scheduling solution be X =< xij >, i ∈ {1, · · · , N }, j ∈ {1, · · · , M }, where xij = 1 if thread i is assigned to machine j; and N and M are the numbers of threads and machines respectively. Different scheduling solutions lead to different tuple processing and transfer delays at/between tasks at runtime thus different end-to-end tuple processing times [25]. We aim to find a scheduling solution that minimizes the average end-to-end tuple processing time. We have described how DRL basically works in Section 2.3. Here, we discuss how to apply DRL to solving the above scheduling problem in a DSDPS. We first define the state space, action space and reward. State Space: A state s = (X, w) consists of two parts: the scheduling solution X, i.e., the current assignment of executors, and the workload w, which includes the tuple arrival rate (i.e., the number of tuples per second) of each data source. The state space is denoted as S. Workload is included in the state to achieve better adaptivity and sensitivity to the incoming workload, which has been validated by our experimental results. Action Space: An action is defined as a =< aij >, ∀i ∈ PM {1, ..., N }, ∀j ∈ {1, ..., M }, where j=1 aij = 1, ∀i, and aij = 1 means assigning thread i to machine j. The action space A is the space P that contains all feasible actions. Note that the constraints M j=1 aij = 1, ∀i ensure that each thread i can only be assigned to a single machine, and the size of action space |A| = M N . Note that an action can be easily translated to its corresponding scheduling solution. Reward : The reward is simply defined to be the negative average tuple processing time so that the objective of the DRL agent is to maximize the reward. Note that the design of state space, action space and reward is critical to the success of a DRL method. In our case, the action space and reward are straightforward. However, there are many different ways for defining the state space because a DSDPS includes various runtime information (features) [25], e.g., CPU/memory/network usages of machines, workload at each executor/process/ machine, average tuple processing delay at each executor/PU, tuple transfer delay between executors/PUs, etc. We, however, choose a simple and clean way in our design. We tried to add additional system runtime information into the state but found that it does not necessarily lead to performance improvement. Different scheduling solutions lead to different values for the above features and eventually different average end-to-end tuple processing times; and the tuple arrival rates reflect the incoming workload. These information turns out to be sufficient for representing runtime system state. We observed that based on our design, the proposed DNNs can well model the correlation between the state and the average end-to-end tuple processing time (reward) after training. A straightforward way to apply DRL to solving the scheduling problem is to directly use the DQN-based method proposed in the pioneering work [33]. The DQN-based method uses a value iteration approach, in which the value function Q = Q(s, a; θ) is a parameterized function (with parameters θ) that takes state s and the action space A as input and return Q value for each action a ∈ A. Then we can use a greedy method to make an action selection: πQ (s) = argmax Q(s, a; θ) (3.1) a∈A If we want to apply the DQN-based method here, we need to restrict the exponentially-large action space A described above to a polynomial-time searchable space. The most natural way to achieve this is to restrict each action to assigning only one thread to a machine. In this way, the size of the action space can be significantly reduced to |A| = N × M , which obviously can be searched in polynomial time. Specifically, as mentioned above, in the offline training phase, we can collect enough samples of rewards and the corresponding state-action pairs for constructing a sufficiently accurate DQN, using a model-free method that deploys a randomlygenerated scheduling solution (state), and collect and record the corresponding average tuple processing time (reward). Then in the online learning phase, at each decision epoch t, the DRL agent obtains the estimated Q value from the DQN for each at with the input of current state st . Then ǫgreedy policy [41] is applied to select the action at according to the current state st : with (1 − ǫ) probability, the action with the highest estimated Q value is chosen, or an action is randomly selected with probability ǫ. After observing immediate reward rt and next state st+1 , a state transition sample (st , at , rt , st+1 ) is stored into the experience replay buffer. At every decision epoch, the DQN is updated with a mini-batch of collected samples in the experience replay buffer using SGD. Although this DQN-based method can provide solutions to the scheduling problem and does achieve model-free control for DSDPSs, it faces the following issue. On one hand, as described above, its time complexity grows linearly with |A|, which demands an action space with a very limited size. On the other hand, restricting the action space may result in limited exploration of the entire exponentially-large action space and thus suboptimal or even poor solutions to the scheduling problem, especially for the large cases. The experimental results in Section 4 validate this claim. 3.2.1 The Actor-critic-based Method for Scheduling In this section, we present a method that can better explore the action space while keeping time complexity at a reasonable level. AK â f (·) Q(·) Figure 2: The actor-critic-based method We leverage some advanced RL techniques, including actor-critic method [49, 6] and the deterministic policy gradient [44], for solving the scheduling problem. Note that since these techniques only provide general design frameworks, we still need to come up with a specific solution to our problem studied here. The basic idea of the proposed scheduling method is illustrated in Figure 2, which includes three major components: 1) an actor network that takes the state as input and returns a proto-action â, 2) an optimizer that finds a set AK of K Nearest Neighbors (K-NN) of â in the action space, and 3) a critic network that takes the state and and AK as input and returns Q value for each action a ∈ AK . Then an action with the highest Q value can be selected for execution. The basic design philosophy of this method is similar to that of a rounding algorithm, which finds a continuous solution by solving a relaxed version of the original integer (i.e., discrete) problem instance and then rounds the continuous solution to a “close” feasible integer solution that hopefully offers an objective value close to the optimal. Specifically, the actor network f (s; θπ ) = â N is a function parameterized by θπ and f : S 7→ RM . â is returned as a proto-action that takes continuous values so â ∈ / A. In our design, we use a 2-layer fully-connected feedforward neural network to serve as the actor network, which includes 64 and 32 neurons in the first and second layer respectively and uses the hyperbolic tangent function tanh(·) for activation. Note that we chose this activation function because our empirical testing showed it works better than the other commonly-used activation functions. The hardest part is to find the K-NN of the proto-action, which has not been well discussed in related work before. Even though finding K-NN can easily be done in linear time, the input size is M N here, which could be a huge number even in a small cluster. Hence, enumerating all actions in A and doing a linear search to find the K-NN may take an exponentially long time. We introduce an optimizer, which finds the K-NN by solving a series of Mixed-Integer Quadratic Programming (MIQP) problems presented in the following: MIQP-NN: min : a s.t.: M X ka − âk22 aij = 1, ∀i ∈ {1, · · · , N }; (3.2) j=1 Similar to the actor network f (·), we employ a 2-layer fullyconnected feedforward neural network to serve as the critic network, which includes 64 and 32 neurons in the first and second layer respectively and uses the hyperbolic tangent function tanh(·) for activation. Note that the two DNNs (one for actor network and one for critic network) are jointly trained using the collected samples. Algorithm 1 The actor-critic-based method for scheduling 1: Randomly initialize critic network Q(·) and actor network f (·) with weights θQ and θπ respectively; ′ 2: Initialize target networks Q′ and f ′ with weights θQ ← Q π′ π θ ,θ ←θ ; 3: Initialize experience replay buffer B; /**Offline Training**/ 4: Load the historical transition samples into B, train the actor and critic network offline; /**Online Learning**/ 5: Initialize a random process R for exploration; 6: Receive a initial observed state s1 ; /**Decision Epoch**/ 7: for t = 1 to T do 8: Derive proto-action â from the actor network f (·); 9: Apply exploration policy to â: R(â) = â + ǫI; 10: Find K-NN actions AK of â by solving a series of MIQP-NN problems (described above); 11: Select action at = argmaxa∈AK Q(st , a); 12: Execute action at by deploying the corresponding scheduling solution, and observe the reward rt ; 13: Store transition sample (st , at , rt , st+1 ) into B; 14: Sample a random mini-batch of H transition samples (si , ai , ri , si+1 ) from B; 15: yi := ri +γ maxa∈Ai+1,K Q′ (si+1 , a), ∀i ∈ {1, · · · , H}, where Ai+1,K is the set of K-NN of f ′ (si+1 ); 16: Update the critic network Q(·) by minimizing the loss: H P 1 L(θQ ) = H [yi − Q(si , ai )]2 ; i=1 17: aij ∈ {0, 1}, ∀i ∈ {1, · · · , N }, ∀j ∈ {1, · · · , M }. In this formulation, the objective is to find the action a that is the nearest neighbor of the proto-action â. The constraints ensure that action a is a feasible action, i.e., a ∈ A. To find the K-NN, the MIQP-NN problem needs to be iteratively solved K times. Each time, one of the KNN of the proto-action will be returned, the corresponding values < aij > are fixed, then the MIQP-NN problem is updated and solved again to obtain the next nearest neighbor until all the K-NN are obtained. Note that this simple MIQP problem can usually be efficiently solved by a solver as long as the input size is not too large. In our tests, we found our MIQP-NN problem instances were all solved very quickly (within 10ms on a regular desktop) by the Gurobi Optimizer [19]. For very large cases, the MIQP-NN problem can be relaxed to a convex programming problem [11] and a rounding algorithm can be used to obtain approximate solutions. The set AK of K-NN actions are further passed to the critic network to select the action. The critic network Q(s, a; θQ ) is a function parameterized by θQ , which returns Q value for each action a ∈ AK , just like the above DQN. The action can then be selected as follows: πQ (s) = argmax Q(s, a; θQ ). a∈AK (3.3) Update the weights θπ of actor network f (·) using the sampled gradient: H P 1 ∇θ π f ≈ H ∇â Q(s, â)|â=f (si ) · ∇θ π f (s)|si ; i=1 18: Update the corresponding target networks: ′ ′ θQ := τ θQ + (1 − τ )θQ ; ′ ′ θπ := τ θπ + (1 − τ )θπ 19: end for We formally present the actor-critic-based method for scheduling as Algorithm 1. First, the algorithm randomly initializes all the weights θπ of actor network f (·) and θQ of critic network Q(·) (line 1). If we directly use the actor and critic networks to generate the training target values < yi > (line 15), it may suffer from unstable and divergence problems as shown in paper [6]. Thus, similar as in [6, 26], we create the target networks to improve training stability. The target networks are clones of the original actor or critic ′ networks, but the weights of the target networks θQ and ′ θπ are slowly updated, which is controlled by a parameter τ . In our implementation, we set τ = 0.01. To robustly train the the actor and critic networks, we adopt the experience replay buffer B [33]. Instead of training network using the transition sample immediately collected at each decision epoch t (from line 8 to line 12), we first store the sample into a replay buffer B, then randomly select a mini-batch of transition samples from B to train the actor and critic networks. Note that since the size of B is limited, the oldest sample will be discarded when B is full. The sizes of replay buffer and mini-batch were set to |B| = 1000 and H = 32 respectively in our implementation. The online exploration policy (line 9) is constructed as R(â) = â+ǫI, where ǫ is an adjustable parameter just as the ǫ in the ǫ-greedy method [41], which determines the probability to add a random noise to the proto-action rather than take the derived action from the actor network. ǫ decreases with decision epoch t, which means with more training, more derived actions (rather than random ones) will be taken. In this way, ǫ can tradeoff exploration and exploitation. The parameter I is a uniformly distributed random noise, each element of which was set to a random number in [0, 1] in our implementation. The critic network Q(·) is trained by the mini-batch samples from B as mentioned above. For every transition sample (si , ai , ri , si+1 ) in the mini-batch, first we obtain the proto-action âi+1 of the next state si+1 from the target actor network f ′ (si+1 ); second, we find K-NN actions Ai+1,K of the proto-action âi+1 by solving a series of MIQP-NN problems presented above; then we obtain the highest Q-value from the target critic network, maxa∈Ai+1,K Q′ (si+1 , a). To train critic network Q(si , ai ), the target value yi for input si and ai is given by the sum of the immediate reward ri and the discounted max Q-value (line 15). The discount factor γ = 0.99 in our implementation. A common loss function L(·) is used to train the critic network (line 16). The actor network f (·) is trained by the deterministic policy gradient method [44] (line 17). The gradient is calculated by the chain rule to obtain the expected return from the transition samples in the mini-batch with respect to the weights θπ of the actor network. The actor and critic networks can be pre-trained by the historical transition samples, so usually the offline training (line 4) is performed first, which is almost the same as online learning (lines 13–18). In our implementation, we first collected 10, 000 transition samples with random actions for each experimental setup and then pre-trained the actor and critic networks offline. In this way, we can explore more possible states and actions and significantly speed up online learning. 4. We implemented three popular and representative SDP applications (called topologies in Storm) to test the proposed framework: continuous queries, log stream processing and word count (stream version), which are described in the following. These applications were also used for performance evaluation in [25] that presented the state-of-the-art modelbased approach; thus using them ensures fair comparisons. Continuous Queries Topology (Figure 3): This topology represents a popular application on Storm. It is a select query that works by initializing access to a database table created in memory and looping over each row to check if there is a hit [8]. It consists of a spout and two bolts. Randomly generated queries are emitted continuously by the spout and sent to a Query bolt. The database tables are placed in the memory of worker machines. After taking queries from the spout, the Query bolt iterates over the database table to check if there is any matching record. The Query bolt emits the matching record to the last bolt named the File bolt, which writes matching records into a file. In our experiments, a database table with vehicle plates and their owners’ information including their names and SSNs was randomly generated. We also randomly generated queries to search the database table for owners of speeding vehicles, while vehicle speeds were randomly generated and attached to every entry. To perform a comprehensive evaluation, we came up with 3 different setups for this topology: small-scale, mediumscale and large-scale. In the small-scale experiment, a total of 20 executors were created, including 2 spout executors, 9 Query bolt executors and 9 File bolt executors. In the medium-scale experiment, we had 50 executors in total, including 5 spout executors, 25 Query bolt executors and 20 File bolt executors. For the large-scale experiment, we had a total of 100 executors, including 10 spout executors, 45 Query bolt executors and 45 File bolt executors. Continuous Queries Topology Spout Query Bolt File Bolt Output File Database Table PERFORMANCE EVALUATION In this section, we describe experimental setup, followed by experimental results and analysis. 4.1 Experimental Setup We implemented the proposed DRL-based control framework over Apache Storm [48] (obtained from Storm’s repository on Apache Software Foundation) and installed the system on top of Ubuntu Linux 12.04. We also used Google’s TensorFlow [50] to implement and train the DNNs. For performance evaluation, we conducted experiments on a cluster in our data center. The Storm cluster consists of 11 IBM blade servers (1 for Nimbus and 10 for worker machines) connected by a 1Gbps network, each with an Intel Xeon Quad-Core 2.0GHz CPU and 4GB memory. Each worker machine was configured to have 10 slots. Figure 3: Continuous Queries Topology Log Stream Processing Topology (Figure 4): Being one of the most popular applications for Storm, this topology uses an open-source log agent called LogStash [29] to read data from log files. Log lines are submitted by LogStash as separate JSON values into a Redis [39] queue, which emits the output to the spout. We used Microsoft IIS log files collected from computers at our university as the input data. The LogRules bolt performs rule-based analysis on the log stream, delivering values containing a specific type of log entry instance. The results are simultaneously delivered to two separate bolts: one is the Indexer bolt performing index actions and another is the Counter bolt performing counting actions on the log entries. For the testing purpose, we slightly revised the original topology to include two more bolts, two Database bolts, after the Indexer and Counter bolts respectively. They store the results into separate collections in a Mongo database for verification purpose. In our experiment, the topology was configured to have a total of 100 executors, including 10 spout executors, 20 LogRules bolt executors, 20 Indexer bolt executors, 20 Counter bolt executors and 15 executors for each Database bolt. Log Files Log Stream Processing Topology Redis Spout LogRules Bolt Log Stash Indexer Bolt Database Bolt Counter Bolt Database Bolt Mongo DB Figure 4: Log Stream Processing Topology Word Count Topology (stream version) (Figure 5): The original version of the topology is widely known as a classical MapReduce application that counts every word’s number of appearances in one or multiple files. The stream version used in our experiments runs a similar routine but with a stream data source. This topology consists of one spout and three bolts with a chain-like structure. LogStash [29] was used to read data from input source files. LogStash submits input file lines as separate JSON values into a Redis [39] queue, which are consumed and emitted into the topology. The text file of Alice’s Adventures in Wonderland [1] was used as the input file. When the input file is pushed into the Redis queue, the spout produces a data stream which is first directed to the SplitSentence bolt, which splits each input line into individual words and further sends them to the WordCount bolt. This bolt then counts the number of appearances using fields grouping. The Database bolt finally stores the results into a Mongo database. In the experiment, the topology was configured to have a total of 100 executors, including 10 spout executors, 30 SplitSentence bolt executors, 30 WordCount executors and 30 Database bolt executors. Input File Word Count Topology Redis Spout SplitSentence Bolt WordCount Bolt Database Bolt Mongo DB Log Stash Figure 5: Word Count Topology (stream version) 4.2 Experimental Results and Analysis In this section, we present and analyze experimental results. To well justify effectiveness of our design, we compared the proposed DRL-based control framework with the actor-critic-based method (labeled as “Actor-critic-based DRL”) with the default scheduler of Storm (labeled as “Default”) and the state-of-the-art model-based method proposed in a very recent paper [25] (labeled as “Model-based”) in terms of average (end-to-end) tuple processing time. Moreover, we included the straightforward DQN-based DRL method (described in Section 3.2) in the comparisons (labeled as “DQN-based DRL”). For the proposed actor-critic-based DRL method and the DQN-based DRL method, both the offline training and online learning were performed to train the DNNs to reach certain scheduling solutions, which were then deployed to the Storm cluster described above. The figures presented in the following show the average tuple processing time corresponding to the scheduling solutions given by all these methods in the period of 20 minutes. In addition, we show the performance of the two DRL methods over the online learning procedure in terms of the reward. For illustration and comparison purposes, we normalize and smooth the reward r−rmin values using a commonly-used method rmax (where r −rmin is the actual reward, rmin and rmax are the minimum and maximum rewards during online learning respectively) and the well-known forward-backward filtering algorithm [20] respectively. Note that in Figures 6, 8 and 10, time 0 is the time when a scheduling solution given by a well-trained DRL agent is deployed in the Storm. It usually takes a little while (10-20 minutes) for the system to gradually stabilize after a new scheduling solution is deployed. This process is quite smooth, which has also been shown in [25]. So these figures do not show the performance of the DRL methods during training processes but the performance of the scheduling solutions given by well-trained DRL agents. The rewards given by the DRL agents during their online learning processes are shown in Figures 7, 9 and 11. This process usually involves large fluctuations, which have also been shown in other DRL-related works such as [23, 26]. Continuous Queries Topology (Figure 3): We present the corresponding experimental results in Figures 6 and 7. As mentioned above, we performed experiments on this topology using three setups: small-scale, medium-scale and largescale, whose corresponding settings are described in the last subsection. From Figure 6, we can see that for all 3 setups and all the four methods, after a scheduling solution is deployed, the average tuple processing time decreases and stabilizes at a lower value (compared to the initial one) after a short period of 8 − 10 minutes. Specifically, in Figure 6(a) (small-scale), if the default scheduler is used, it starts at 3.71ms and stabilizes at 1.96ms; if the model-based method is employed, it starts at 3.22ms and stabilizes at 1.46ms; if the DQN-based DRL method is used, it starts at 3.20ms and stabilizes at 1.54ms; and if the actor-critic-based DRL method is applied, it starts at 2.94ms and stabilizes at 1.33ms. In this case, the actor-critic-based DRL method reduces the average tuple processing time by 31.4% compared to the default scheduler and by 9.5% compared to the model-based method. The DQN-based DRL method performs slightly worse than the model-based method. From Figure 6(b) (medium-scale), we can see that the average tuple processing times given by all the methods slightly go up. Specifically, if the default scheduler is used, it stabilizes at 2.08ms; if the model-based method is employed, it stabilizes at 1.61ms; if the DQN-based DRL method is used, it stabilizes at 1.59ms; and if the actor-critic-based DRL method is applied, it stabilizes at 1.43ms. Hence, in this case, the actor-critic-based DRL method achieves a performance improvement of 31.2% over the default scheduler 4 3.5 3 2.5 2 1.5 1 0 5 10 Running time (min) 15 4.5 Default Model−based DQN−based DRL Actor−critic−based DRL 3.5 Average tuple processing time (ms) Default Model−based DQN−based DRL Actor−critic−based DRL Average tuple processing time (ms) Average tuple processing time (ms) 4 3 2.5 2 1.5 1 0 20 (a) Small-scale 5 10 Running time (min) 15 (b) Medium-scale 20 Default Model−based DQN−based DRL Actor−critic−based DRL 4 3.5 3 2.5 2 1.5 0 5 10 Running time (min) 15 20 (c) Large-scale Figure 6: Average tuple processing time over the continuous queries topology 0.7 Actor−critic−based DRL DQN−based DRL 0.65 Normalized reward 0.6 0.55 0.5 0.45 0.4 0.35 0.3 0.25 0 500 1000 Decision epoch 1500 2000 Figure 7: Normalized reward over the continuous queries topology (large-scale) and 11.2% over the model-based method. The performance of DQN-based DRL method is still comparable to the modelbased method. From Figure 6(c) (large-scale), we can observe that the average tuple processing times given by all the methods increase further but still stabilize at reasonable values, which essentially shows that the Storm cluster undertakes heavier workload but has not been overloaded in this large-scale case. Specially, if the default scheduler is used, it stabilizes at 2.64ms; if the model-based method is employed, it stabilizes at 2.12ms; if the DQN-based DRL method is used, it stabilizes at 2.45ms; and if the actor-critic-based DRL method is applied, it stabilizes at 1.72ms. In this case, the actor-critic-based DRL method achieves a more significant performance improvement of 34.8% over the default scheduler and 18.9% over the model-based method. In addition, the performance of the DQN-based DRL method is noticeably worse than that of the model-based method. In summary, we can make the following observations: 1) The proposed actor-critic-based DRL method consistently outperforms all the other three methods, which well justifies the effectiveness of the proposed model-free approach for control problems in DSDPSs. 2) The performance improvement (over the default scheduler and the model-based method) offered by the proposed actor-critic-based DRL method become more and more significant with the increase of input size, which shows that the proposed model-free method works even better when the distributed computing environment becomes more and more sophisticated. 3) Direct application of DQN-based DRL method does not work well, especially in the large case. This method lacks a carefully-designed mechanism (such as the proposed MIQPbased mechanism presented in Section 3.2.1) that can fully discover the action space and make a wise action selection. Hence, in large cases with huge action spaces, random selection of action may lead to a suboptimal or even poor decision. We further exploit how the two DRL methods behave during online learning by showing how the normalized reward varies over time within T = 2000 decision epochs in Figure 7. We performed this experiment using the large-scale setup described above. From this figure, we can observe that both methods start from similar initial reward values. The DQN-based DRL method keeps fluctuating during the entire procedure and ends at an average award value of 0.44 (the average over the last 200 epochs); while the actor-criticbased DRL method experiences some fluctuations initially, then gradually climbs to a higher value. More importantly, the actor-critic-based DRL method consistently offers higher rewards compared to the DQN-based method during online learning. These results further confirm superiority of the proposed method during online learning. Moreover, we find that even in this large-scale case, the proposed actorcritic-based DRL method can quickly reach a good scheduling solution (whose performance have been discussed above) without going though a really long online learning procedure (e.g., several million epochs) shown in other applications [33]. This justifies the practicability of the proposed DRL method for online control in DSPDSs. Log Stream Processing Topology (Figure 4): We performed a large-scale experiment over the log stream processing topology, whose settings have been discussed in the last subsection too. We show the corresponding results in Figures 8 and 9. This topology is more complicated than the previous continuous queries topology, which leads to a longer average tuple processing time no matter which method is used. Default Model−based DQN−based DRL Actor−critic−based DRL 11 10 9 8 7 0 5 10 15 20 Figure 8: Average tuple processing time over the log processing topology (large-scale) 0.9 Actor−critic−based DRL DQN−based DRL Normalized reward 0.8 0.75 0.7 0.65 0.6 0.55 0.5 0 500 1000 1500 Decision epoch Figure 9: Normalized reward over the log processing topology (large-scale) Average tuple processing time (ms) 4.5 Default Model−based DQN−based DRL Actor−critic−based DRL 4 3.5 3 2.5 2 1.5 0 0.55 0.5 0.45 0.4 Running time (min) 0.85 Actor−critic−based DRL DQN−based DRL Normalized reward Average tuple processing time (ms) 12 5 10 Running time (min) 15 20 Figure 10: Average tuple processing time over the word count topology (large-scale) 0 500 1000 1500 Decision epoch Figure 11: Normalized reward over the word count topology (large-scale) In Figure 8, if the default scheduler is used, it stabilizes at 9.61ms; if the model-based method is employed, it stabilizes at 7.91ms; if the DQN-based DRL method is used, it stabilizes at 8.19ms; and if the actor-critic-based DRL method is applied, it stabilizes at 7.20ms. As expected, the proposed actor-critic-based DRL method consistently outperforms the other three methods. Specifically, it reduces the average tuple processing time by 25.1% compared to the default scheduler and 9.0% compared to the model-based method. Furthermore, the DQN-based DRL method performs worse than the model-based method, which is consistent with the results related to the continuous queries topology. Similarly, we show how the normalized reward varies over time within T = 1500 decision epochs in Figure 9. From this figure, we can make a similar observation that the actor-critic-based DRL method consistently leads to higher rewards compared to the DQN-based method during online learning. Obviously, after a short period of online learning, the proposed actor-critic-based DRL method can reach a good scheduling solution (discussed above) in this topology too. Word Count Topology (stream version) (Figure 5): We performed a large-scale experiment over the word count (stream version) topology and show the corresponding results in Figures 10 and 11. Since the complexity of this topology is similar to that of the continuous queries topology, all the four methods give similar average time processing times. In Figure 10, if the default scheduler is used, it stabilizes at 3.10ms; if the model-based method is employed, it stabilizes at 2.16 ms; if the DQN-based DRL method is used, it stabilizes at 2.29ms; and if the actor-critic-based DRL method is applied, it stabilizes at 1.70ms. The actor-critic-based DRL method results in a performance improvement of 45.2% over the default scheduler and 21.3% improvement over the model-based method. The performance of the DQN-based DRL method is still noticeably worse than the model-based method. We also show the performance of the two DRL methods during online learning in Figure 11, from which we can make observations similar to those related to the first two topologies. In addition, we compared the proposed model-free method with the model-based method under significant workload 10 4 4 Model−based Actor−critic−based DRL 3.5 3 2.5 2 1.5 0 10 20 30 Running time (min) 40 (a) Continuous queries topology 50 Model−based Actor−critic−based DRL 9.5 Average tuple processing time (ms) Average tuple processing time (ms) Average tuple processing time (ms) Model−based Actor−critic−based DRL 9 8.5 8 7.5 7 0 10 20 30 Running time (min) 40 50 (b) Log stream processing topology 3.5 3 2.5 2 1.5 0 10 20 30 Running time (min) 40 50 (c) Word count topology Figure 12: Average tuple processing time over 3 different topologies (large-scale) under significant workload changes changes on 3 different topologies (large-scale). For the continuous queries topology (large-scale), we can see from Figure 12(a) that when the workload is increased by 50% at 20 minute, the average tuple processing time of the actor-criticbased DRL method rises sharply to a relatively high value then gradually stabilizes at 1.76ms, while the model-based method rises sharply too and then stabilizes at 2.17ms. The spikes are caused by the adjustment of the scheduling solution. However, we can observe that once the system stabilizes, the proposed method leads to a very minor increase on average tuple processing time. Hence, it is sensitive to the workload change and can quickly adjust its scheduling solution accordingly to avoid performance degradation. This result well justifies the robustness of the proposed method in a highly dynamic environment with significant workload changes. Moreover, during the whole period, we can see that the proposed method still consistently outperforms the model-based method. We can make similar observations for the other two topologies from Figures 12(b)–12(c). 5. RELATED WORK In this section, we provide a comprehensive review for related systems and research efforts. Distributed Stream Data Processing Systems (DSDPS): Recently, a few general-purpose distributed systems have been developed particularly for SDP in a large cluster or in the cloud. Storm [48] is an open-source, distributed and fault-tolerant system that was designed particularly for processing unbounded streams of data in real or near-real time. Storm provides a directed graph based model for programming as well as a mechanism to guarantee message processing. S4 [42] is another general-purpose DSDPS that has a similar programming model and runtime system. MillWheel [3] is a platform designed for building lowlatency SDP applications, which has been used at Google. A user specifies a directed computation graph and application code for individual nodes, and the system manages persistent states and continuous flows of data, all within the envelope of the framework’s fault-tolerance guarantees. TimeStream [38] is a distributed system developed by Microsoft specifically for low-latency continuous processing of stream data in the cloud. It employs a powerful new abstraction called resilient substitution that caters to specific needs in this new computation model to handle failure recovery and dynamic reconfiguration in response to load changes. Spark Streaming [47] is a DSDPS that uses the core Spark API and its fast scheduling capability to perform streaming analytics. It ingests data in mini-batches and performs RDD transformations on those mini-batches of data. Other DSDPSs include C-MR [7], Apache Flink [16], Google’s FlumeJava [12], M3 [2], WalmartLab’s Muppet [24], IBM’s System S [32] and Apache Samza [43]. Modeling and scheduling have also been studied in the context of DSDPS and distributed stream database system. In an early work [34], Nicola and Jarke provided a survey of performance models for distributed and replicated database systems, especially queueing theory based models. In [30], scheduling strategies were proposed for a data stream management system, aiming at minimizing the tuple latency and total memory requirement. In [51], Wei et al. proposed a prediction-based Quality-of-Service (QoS) management scheme for periodic queries over dynamic data streams, featuring query workload estimators that predict the query workload using execution time profiling and input data sampling. The authors of [40] described a decentralized framework for pro-actively predicting and alleviating hot-spots in SDP applications in real-time. In [5], Aniello et al. presented both offline and online schedulers for Storm. The offline scheduler analyzes the topology structure and makes scheduling decisions; the online scheduler continuously monitors system performance and reschedules executors at runtime to improve overall performance. Xu et al. presented T-Storm in [52], which is a traffic-aware scheduling framework that minimizes inter-node and inter-process traffic in Storm while ensuring no worker nodes were overloaded, and enables fine-grained control over worker node consolidation. In [9], Bedini et al. presented a set of models that characterize a practical DSDPS using the Actor Model theory. They also presented an experimental validation of the proposed models using Storm. In [10], a general and simple technique was presented to design and implement priority-based resource scheduling in flow-graph-based DSDPSs by allowing application developers to augment flow graphs with priority metadata and by introducing an extensive set of priority schemas. In [31], a novel technique was proposed for resource usage estimation of SDP workloads in the cloud, which uses mixture density networks, a combined structure of neural networks and mixture models, to estimate the whole spectrum of resource usage as probability density functions. In a very recent work [25], Li et al. presented a topology-aware method to accurately predict the average tuple processing time for a given scheduling solution, according to the topology of the application graph and runtime statistics. For scheduling, they presented an effective algorithm to assign threads to machines under the guidance of the proposed prediction model. Unlike them, we aim to develop a model-free control framework for scheduling in DSDPSs to directly minimize the average tuple processing time by leveraging the state-of-the-art DRL techniques, which has not been done before. Deep Reinforcement Learning (DRL): DRL has recently attracted extensive attention from both industry and academia. In a pioneering work [33], Mnih et al. proposed Deep Q Network (DQN), which can learn successful policies directly from high dimensional sensory inputs. This work bridges the gap between high-dimensional sensory inputs and actions, resulting in the first artificial agent that is capable of learning to excel at a diverse array of challenging gaming tasks. The authors of [23] proposed double Q-learning as a specific adaptation to the DQN, which is introduced in a tabular setting and can be generalized to work with a large-scale function approximation. The paper [17] considered a problem of multiple agents sensing and acting with the goal of maximizing their shared utility, based on DQN. The authors designed agents that can learn communication protocols to share information needed for accomplishing tasks. In order to further extend DRL to address continuous actions and large discrete action spaces, Duan et al. [15] presented a benchmark suite of control tasks to quantify progress in the domain of continuous control; and Lillicrap et al. [26] proposed an actor-critic, model-free algorithm based on the policy gradient that can operate over continuous action spaces. Gu et al. [18] presented normalized advantage functions to reduce the sample complexity of DRL for continuous tasks. Arnold et al. [6] extended the methods proposed for continuous actions to make decisions within a large discrete action space. In [36], the authors employed a DRL framework to jointly learn state representations and action policies using game rewards as feedback for text-based games, where the action space contains all possible text descriptions. It remains unknown if and how the emerging DRL can be applied to solving complicated control and resource allocation problems in DSDPSs. In addition, RL/DRL has been applied to control of big data and cloud systems. In [35], the authors proposed a novel MapReduce scheduler in heterogeneous environments based on RL, which observes the system state of task execution and suggests speculative re-execution of the slower tasks on other available nodes in the cluster. An RL approach was proposed in [37] to enable automated tuning configuration of MapReduce parameters. Liu et al. [27] proposed a novel hierarchical framework for solving the overall resource resource allocation and power management problem in cloud computing systems with DRL. The control problems in MapReduce and cloud systems are quite different from the problem studied here. Hence, the methods proposed in these related works cannot be directly applied here. 6. CONCLUSIONS In this paper, we investigated a novel model-free approach that can learn to well control a DSDPS from its experience rather than accurate and mathematically solvable system models, just as a human learns a skill. We presented design, implementation and evaluation of a novel and highly effective DRL-based control framework for DSDPSs, which minimizes the average end-to-end tuple processing time. The proposed framework enables model-free control by jointly learning the system environment with very limited runtime statistics data and making decisions under the guidance of two DNNs, an actor network and a critic network. We implemented it based on Apache Storm, and tested it with three representative applications: continuous queries, log stream processing and word count (stream version). Extensive experimental results well justified the effectiveness of our design, which showed: 1) The proposed framework achieves a performance improvement of 33.5% over Storm’s default scheduler and 14.0% over the state-of-the-art model-based method on average. 2) The proposed DRL-based framework can quickly reach a good scheduling solution during online learning. 7. ACKNOWLEDGEMENT This research was supported in part by AFOSR grants FA9550-16-1-0077 and FA9550-16-1-0340. The information reported here does not reflect the position or the policy of the federal government. 8. REFERENCES [1] Alice’s Adventures in Wonderland, http://www.gutenberg.org/files/11/11-pdf.pdf [2] A.M. Aly, A. Sallam, B.M. Gnanasekaran, L. Nguyen-Dinh, W.G. Aref, M. Ouzzani, and A. Ghafoor, M3: Stream Processing on Main-Memory MapReduce, Proceedings of IEEE ICDE’2012, pp.1253–1256, 2012. [3] T. Akidau, A. Balikov, K. Bejiroglu, S. Chernyak, J. Haberman, R. Lax, S. McVeety, D. Mills, P. Nordstrom and S. Whittle, MillWheel: fault-tolerant stream processing at internet scale. PVLDB, 6(11): 1033–1044, 2013. [4] Q. Anderson, Storm real-time processing cookbook, PACKT Publishing, 2013. [5] L. Aniello, R. Baldoni and L. Querzoni, Adaptive online scheduling in Storm, Proceedings of ACM DEBS’2013. [6] G. D. Arnold, R. Evans, H. v. Hasselt, P. Sunehag, T. Lillicrap, J. Hunt, T. Mann, T. Weber, T. Degris and B. Coppin, Deep reinforcement learning in large discrete action spaces, arXiv: 1512.07679, 2016. [7] N. Backman, K. Pattabiraman, R. Fonseca and U. Cetintemel, C-MR: continuously executing MapReduce workflows on multi-core processors, Proceedings of MapReduce’12. [8] P. Bakkum and K. Skadron, Accelerating SQL database operations on a GPU with CUDA, Proceedings of the 3rd Workshop on General-Purpose Computation on GPU (GPGPU’10), pp. 94–103. [9] Ivan Bedini, Sherif Sakr, Bart Theeten, Alessandra Sala, Peter Cogan, Modeling performance of a parallel [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] streaming engine: bridging theory and costs, Proceedings of IEEE ICPE’2013, pp. 173–184. P. Bellavista, A. Corradi, A. Reale and N. Ticca, Priority-based resource scheduling in distributed stream processing systems for big data applications, Proceedings of IEEE/ACM International Conference on Utility and Cloud Computing, 2014, pp. 363–370. S. Boyd and L. Vandenberghe, Convex Optimization Cambridge University Press, 2004. C. Chambers, A. Raniwala, F. Perry, S. Adams, R. Henry, R. Bradshaw and N. Weizenbaum, FlumeJava: easy, efficient data-parallel pipelines, Proceedings of ACM PLDI’2010, pp. 363–375. F. Chen, M. Kodialam and T. V. Lakshman, Joint scheduling of processing and shuffle phases in MapReduce systems, Proceedings of IEEE Infocom’2012, pp. 1143–1151. H. Drucker, C. Burges, L. Kaufman, A. Smola and V. Vapnik, Support vector regression machines, Proceedings of NIPS’1996, pp. 155–161. Y. Duan, X. Chen, R. Houthooft, J. Schulman and P. Abbeel, Benchmarking deep reinforcement learning for continuous control, Proceedings of ICML’2016. Apache Flink, https://flink.apache.org/ J. N. Foerster, Y. M. Assael, N. d. Freitas and S. Whiteson, Learning to communicate with deep multi-agent reinforcement learning, Proceedings of NIPS’2016. S. Gu, T. Lillicrap, I. Sutskever and S. Levine, Continuous deep Q-Learning with model-based acceleration, Proceedings of ICML’2016. Gurobi Optimizer, http://www.gurobi.com/ F. Gustafsson, Determining the initial states in forward-backward filtering, IEEE Transactions on Signal Processing, Vol. 44, No. 4, 1996, pp. 988–992. J. Han, M. Kamber and J. Pei, Data Mining: Concepts and Techniques (3rd Edition), Morgan Kaufmann, 2011. Apache Hadoop, http://hadoop.apache.org/ H. v. Hasselt, A. Guez, and D. Silver, Deep reinforcement learning with double Q-learning, Proceedings of AAAI’2016. W. Lam, L. Liu, S. Prasad, A. Rajaraman, Z. Vacheri and A. Doan, Muppet: MapReduce-style processing of fast data, PVLDB, 5(12): 1814–1825, 2012. T. Li, J. Tang and J. Xu, Performance modeling and predicitive scheduling for distributed stream data processing, IEEE Transactions on Big Data, 2(4): 353–364, 2016. T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver and D. Wierstra, Continuous control with deep reinforcement learning, Proceedings of ICLR’2016. N. Liu, Z. Li, J. Xu, Z. Xu, S. Lin, Q. Qiu, J. Tang and Y. Wang, A hierarchical framework of cloud resource allocation and power management using deep reinforcement learning, Proceedings of ICDCS’2017. S. Loesing, M. Hentschel, T. Kraska and D. Kossmann, Stormy: an elastic and highly available streaming service in the cloud, Proceedings of the 2012 ACM Joint EDBT/ICDT Workshops, pp. 55–60. [29] Logstash - Open Source Log Management, http://logstash.net/ [30] Q. Jiang and S. Chakravarthy, Scheduling strategies for a data stream management system, Technical Report CSE-2003-30. [31] A. Khoshkbarforoushha, R. Ranjan, R. Gaire, P. P. Jayaraman, J. Hosking and E. Abbasnejad, Resource usage estimation of data stream processing workloads in datacenter clouds, 2015, http://arxiv.org/abs/1501.07020. Proceedings of ACM STOC’1997, pp. 654–663. [32] V. Kumar, H. Andrade, B. Gedik, and K.-L. Wu, Deduce: at the intersection of Mapreduce and stream processing, Proceedings of ACM EDBT’2010, pp. 657–662. [33] V. Mnih, et al. , Human-level control through deep reinforcement learning, Nature, 518(7540): 529–533, 2015. [34] M. Nicola and M. Jarke, Performance modeling of distributed and replicated databases, IEEE Transactions on Knowledge Discovery and Data Engineering, 12(4): 645–672, 2000. [35] N. Naik, A. Negi and V. Sastry, Performance improvement of MapReduce framework in heterogeneous context using reinforcement learning, Procedia Computer Science, 50: 169–175, 2015. [36] K. Narasimhan, T. D Kulkarni and R. Barzilay, Language understanding for text-based games using deep reinforcement learning, Proceedings of Conference on Empirical Methods in Natural Language Processing, 2015. [37] C. Peng, C. Zhang, C. Peng and J. Man, A reinforcement learning approach to map reduce auto-configuration under networked environment, International Journal of Security and Networks, 12(3): 135–140, 2017. [38] Z. Qian, et al. , Timestream: reliable stream computation in the cloud, Proceedings of EuroSys’2013. [39] Redis, http://redis.io [40] T. Repantis and V. Kalogeraki, Hot-spot prediction and alleviation in distributed stream processing applications, Proceedings of IEEE DSN’2008, pp. 346–355. [41] M. Restelli, Reinforcement learning - exploration vs exploitation, 2015, http://home.deib.polimi.it/restelli/MyWebSite/pdf/ rl5.pdf [42] Apache S4, http://incubator.apache.org/s4/ [43] Apache Samza, http://samza.apache.org/ [44] D. Silver, G. Lever, N. Heess, T. Degris, D. Wierstra and M. Riedmiller, Deterministic policy gradient algorithms, Proceedings of ICML’2014. [45] D. Silver, et al. , Mastering the game of Go with deep neural networks and tree search Nature, 529: 484–489, 2016. [46] Apache Spark, http://spark.apache.org/ [47] Spark Streaming — Apache Spark, http://spark.apache.org/streaming/ [48] Apache Storm, http://storm.apache.org/ [49] R. Sutton and A. Barto, Reinforcement learning: an introduction, MIT press Cambridge, 1998. [50] TensorFlow, https://www.tensorflow.org/ [51] Y. Wei, V. Prasad, S. Son and J. Stankovic, Prediction-based QoS management for real-time data streams, Proceedings of IEEE RTSS’2006. [52] J. Xu, Z. Chen, J. Tang and S. Su, T-Storm: traffic-aware online scheduling in Storm, Proceedings of IEEE ICDCS’2014, pp. 535–544. [53] M. Zaharia, A. Konwinski, A. D. Joseph, R. Katz, I. Stoica, Improving MapReduce performance in heterogeneous environments, Proceedings of OSDI’2008. [54] Apache Zookeeper, https://zookeeper.apache.org/ [55] Y. Zhu, et al. , Minimizing makespan and total completion time in MapReduce-like systems, Proceedings of IEEE Infocom’2014, pp. 2166–2174.
2cs.AI
УДК 004.9:66.013.512 В.В. Мигунов, Р.Р. Кафиятуллов ЗАЩИТА ИНФОРМАЦИИ В КОМПЛЕКСНОЙ САПР РЕКОНСТРУКЦИИ ПРОМЫШЛЕННЫХ ПРЕДПРИЯТИЙ Введение. По мере перехода к электронному представлению информации, которое делает возможным ее хранение и транспортировку на все более компактных носителях и передачу со все большими скоростями по каналам связи, задача защиты информации (ЗИ) в электронном виде приобретает особую актуальность. Возникающие здесь проблемы и методы очень разнообразны: от законодательных, административных и организационных до таких аспектов технической защиты, как предотвращение внедрения программвирусов или физическая защита помещений (п. 5.1.3. [1]). Теория и практика ЗИ в нашей стране развиваются как путем распространения глубоко разработанных методов защиты сведений, составляющих государственную тайну, на конфиденциальные сведения [1], так и на основе относительно новых разработок в сфере защиты коммерческой тайны [2, 3]. При этом в первую очередь рассматривается ЗИ от умышленных действий. Например, в п. 5.1.2. [1] говорится: "Основными направлениями защиты информации являются: обеспечение защиты информации от хищения, утраты, утечки, уничтожения, искажения и подделки за счет НСД и специальных воздействий; обеспечение защиты информации от утечки по техническим каналам при ее обработке, хранении и передаче по каналам связи" (НСД - несанкционированный доступ). В [3]: "В настоящей работе предпринята попытка возможно более полного охвата угроз безопасности субъектов информационных отношений. Однако следует иметь в виду, что научнотехнический прогресс может привести к появлению принципиально новых видов угроз и что изощренный ум злоумышленника способен придумать новые способы преодоления систем безопасности, НСД к данным и дезорганизации работы информационных систем". Лишь во вторую очередь принято рассматривать угрозы от неумышленных действия. В частности, для САПР эти угрозы и методы борьбы с ними изучены мало. В противоположность подходам к ЗИ, акцентированным на противодействии злому умыслу, настоящая работа посвящена изучению угроз информации, возникающих вследствие непредумышленных действий пользователей САПР, и методов ЗИ, ограниченных сферой влияния программного обеспечения САПР. В основе рассмотрения лежит практический опыт развития комплексной САПР реконструкции предприятий TechnoCAD GlassX, эксплуатируемой в проектно-конструкторском подразделении крупного предприятия химической промышленности [4] в течение 10 лет. Из особенностей такой САПР, охарактеризованных в [5], существенной для ЗИ явилась комплексность. Отметим, что именно непредумышленные действия являются основными угрозами информации. По сведениям из [2]: "... в 1998 году. Основные причины повреждений электронной информации распределились следующим образом: неумышленная ошибка человека - 52% случаев, умышленные действия человека - 10% случаев, отказ техники - 10% случаев, повреждения в результате пожара - 15% случаев, повреждения водой 10% случаев". В [3] приводятся данные, относящиеся к более позднему времени: "Согласно статистики, 65% потерь - следствие непреднамеренных ошибок. Пожары и наводнения можно считать пустяками по сравнению с безграмотностью и расхлябанностью". Угрозы информации в сфере влияния САПР. В первую очередь беспокоящие проектировщика угрозы информации в САПР – это угрозы порчи или утраты электронных документов (файлов с чертежами), приводящие к необходимости повторно выполнять уже сделанную работу. По опыту эксплуатации САПР выделены 4 угрозы, которые в первую очередь отмечаются пользователями: 1. Случайное удаление, изменение элементов чертежа. 2. Утеря результатов работы за период с последнего сохранения чертежа на диск вследствие отказов: сети электроснабжения, узлов компьютера, операционной системы и самой САПР. 3. Случайная подмена чертежа другим с тем же именем файла. Для идентификации чертежей проектировщики часто употребляют номер проектируемого корпуса. При этом монтажный чертеж трубопровода и схема автоматизации, созданные на разных рабочих местах, могут иметь одно и то же имя файла. При передаче чертежей между рабочими местами возникает угроза неумышленной замены чертежа. 4. Утрата уверенности в целостности чертежа, возникающая при отсутствии системы ответственного электронного архивирования (хранения) файлов чертежей. Чертеж может измениться в результате неосторожных действий при его просмотре или печати, при использовании его как заготовки для создания другого чертежа. Реализация ЗИ в САПР от этих угроз далее описана соответственно как "Откат", "Автосохранение", "Автоматическое резервное копирование" и "Электронная подпись". Откат. Откат, или возврат изменений (команды "Отменить" и "Вернуть") широко применяемый метод борьбы с 1й угрозой. Его теория глубоко развита для баз данных [6] как работа с транзакциями. В условиях комплексной САПР возникает специфика, связанная с тем, что при выходе на довольно продолжительное время в режим работы в специализированном проблемноориентированном расширении [7] пользователь занимается параметрическим проектированием, не имеет связи с чертежом и накапливает информацию не в нем, а в создаваемом параметрическом представлении (ПП). Примеры таких расширений: проектирование профилей наружных сетей водоснабжения и канализации, молниезащиты зданий и сооружений. Функции отката рассмотрим сначала для чертежа, затем для ПП его частей. Все операции разработки чертежа удается свести к двум первичным, атомарным - добавление и удаление элемента чертежа. Изменение типа линии, слоя, цвета и др. рассматриваются как удаление и последующее добавление в чертеж измененного элемента. Операции над множеством элементов (сдвиг, растяжение и др.) моделируются как множество одинаковых операций, каждая из которых выполняется над одним элементом. Вводится понятие "шаг изменения", которое включает все операции, проведенные пользователем над однократно выбранными элементами. Тем самым достигается соответствие интуитивного представления об "изменении" и количества возвращаемых за один шаг атомарных изменений - аналог транзакций в базах данных. На рис.1 каждый шаг обозначен параллелограммом, "стабильные" состояния чертежа флажками, а стрелки показывают направление изменений. Последовательность шагов при возврате изменений существенна, иначе: может потребоваться удаление элементов, которых в чертеже нет; пользователь не будет знать, какое изменение возвратится в первую очередь. Учитываются изменения чертежа в последнем сеансе работы с ним, возврат возможен до завершения сеанса. Рис.1. Модель пошаговых изменений - транзакций В рабочий файл последовательно записываются удаляемые и добавляемые геометрические элементы вместе с признаком - добавлен или удален элемент. В результате, как показано на рис.2, один шаг представляется в виде сцепленных в железнодорожный состав различных геометрических элементов, образующих как бы сумму атомарных изменений. Внутри одного шага последовательность элементов несущественна, если все элементы разные. Если же для удобства программирования приходится в один шаг включать, например, удаление и последующее восстановление одного и того же элемента, то порядок следования атомарных изменений при возврате должен быть строго обратным. Рис.2. Общая структура одного шага изменения чертежа Для возврата изменения делается проход по файлу, пока номер изменения совпадает с нужным. Для каждого из найденных элементов с этим номером изменения производится следующее: если он был удален, он добавляется в чертеж; если он был добавлен, то в чертеже ищется совпадающий с ним и удаляется из чертежа. Применяя команды "Отменить" и "Вернуть", можно как угодно долго перебирать состояния чертежа в текущем сеансе работы. ПП части чертежа есть комплекс ее свойств, или информационная модель. Видимая геометрическая часть генерируется по ПП. С точки зрения отката наиболее существенное отличие работы с ПП от работы непосредственно с чертежом заключается в интеллектуальных связях элементов друг с другом. При удалении одной трубы из ПП аксонометрической схемы трубопроводной системы автоматически удаляются все нанесенные на нее обозначения трубопроводной арматуры, элементов трубопроводов, тексты и сноски, указывающие на трубу и удаленные обозначения, выносные линии размеров. В этих условиях система отката не имеет явно очерченных границ того, что же изменяется за один шаг. Верхней границей является ПП полностью. В рамках модульной технологии разработки расширений САПР все ПП преобразуются к единой компактной форме, задаваемой адресом в памяти и длиной [7], поэтому заменяемым объектом выбрано все ПП. Это решение универсально для различных расширений САПР. При любом изменении новая версия ПП записывается в рабочий файл. Операции над множеством здесь возможны, но их не надо выражать в множестве актов удаления и добавления, каждому шагу изменений соответствует одно ПП. Отмена изменений и возврат реализуются путем перехода к предыдущей или к следующей версии ПП с полной его заменой. Ситуации соответствует рис.1 без детализации каждого шага по рис.2. Последовательность шагов изменения роли не играет, каждая записанная версия ПП дает полный комплект всех сведений для генерации изображения. Возможен скачок через несколько шагов. Автосохранение. Для борьбы с угрозой 2 применяется широко распространенный метод автоматического сохранения информации через заданный промежуток времени - автосохранение. Специфика комплексной САПР приводит к необходимости автосохранения не только чертежа, но и текущих параметрических представлений. С интервалом не менее заданного на диске автоматически создаются копии чертежа и ПП текущей задачи, предыдущая копия с диска удаляется. При нормальном завершении задачи или работы с чертежом автоматически созданные копии удаляются. В случае некорректного завершения работы (например, из-за отключения электропитания) эти копии остаются, и при следующей загрузке пользователю будет предложено восстановить чертеж и ПП задачи. Автоматическое резервное копирование. Защита от случайной подмены файла чертежа обеспечивается путем автоматического резервного копирования всех файлов, подвергшихся изменению за период с предыдущего копирования, в отдельный каталог диска с именем, определяемым датой создания резервных копий. Пользователь назначает дни недели, в которые следует проводить автоматическое резервное копирование, и при первом запуске САПР в эти дни оно осуществляется. Забота об удалении ставших ненужными резервных копий лежит на пользователе. Электронная подпись. Для сохранения и контроля целостности чертежа (борьба с угрозой 4) применяется метод электронной подписи. Он состоит в том, что при наличии хотя бы одной подписи запрещено любое изменение чертежа. Подписи защищены личными паролями. Наличие подписи на чертеже гарантирует совпадение его содержания с тем, что было в момент подписания. Снять подпись может любой пользователь, после чего чертеж можно изменять. Но при этом будет видно, что чертеж не подписан. Заключение. Описанные четыре угрозы информации и методы ее защиты в комплексной САПР реконструкции предприятий являются в значительной степени общими и для других программных средств редактирования пользовательских файлов, поскольку состав угроз определен практическими потребностями. Индивидуальные особенности реализация ЗИ продиктованы прежде всего комплексностью САПР, поддерживающей несколько информационных моделей объектов проектирования. БИБЛИОГРАФИЧЕСКИЙ СПИСОК 1. Специальные требования и рекомендации по технической защите конфиденциальной информации (СТР-К). Решение Коллегии Гостехкомиссии России № 7.2/02.03.01г. [Электронный ресурс] – Режим доступа: http://www.confidentiality.strongdisk.ru, свободный. – Загл. с экрана. – Яз. рус. 2. Беляев А.В. "Методы и средства защиты информации" (курс лекций). – ЧФ СПбГТУ, 2000. [Электронный ресурс] – Режим доступа: http://www.citforum.ru/internet/infsecure/its2000_03.shtml, свободный. – Загл. с экрана. – Яз. рус. 3. Домарев В.В. Безопасность информационных технологий. Системный подход. – Киев: ООО ТИД «Диасофт», 2004. – 992с. 4. Мигунов В.В. TechnoCAD GlassX − отечественная САПР реконструкции предприятия//САПР и графика, 2004, №№ 4, 5, 6. 5. Мигунов В.В. Особенности комплексной САПР реконструкции промышленных предприятий//Труды Международных научно-технических конференций "Интеллектуальные системы (IEEE AIS'04)" и "Интеллектуальные САПР (CAD-2004)". Научное издание в 3-х томах. – М: Из*д-во Физико-математической литературы, 2004, т.2, С.70-75. 6. Дейт, К.Дж. Введение в системы баз данных.: Пер с англ. – 6-е изд. – Киев: Диалектика, 1998. – 784с. 7. Мигунов В.В. Модульная технология разработки проблемноориентированных расширений САПР реконструкции предприятия//Известия Тульского государственного университета. Серия "Экономика. Управление. Стандартизация. Качество". – Вып. 1. – Тула: Изд-во ТулГУ, 2004. – С.56-66.
5cs.CE
AptaTRACE: Elucidating Sequence-Structure Binding Motifs by Uncovering Selection Trends in HT-SELEX Experiments Phuong Dao1,? , Jan Hoinka1,? , Yijie Wang1 , Mayumi Takahashi2 , Jiehua Zhou2 , Fabrizio Costa3 , John Rossi2 , John Burnett2 , Rolf Backofen3 , and Teresa M.Przytycka1 arXiv:1604.03081v1 [q-bio.QM] 5 Apr 2016 1 National Center of Biotechnology Information, National Library of Medicine, NIH, Bethesda MD 20894, USA [email protected] 2 Department of Molecular and Cellular Biology, Beckman Research Institute of City of Hope, CA, USA 3 Bioinformatics Group, Department of Computer Science, University of Freiburg, Georges-Kohler-Allee 106, 79110 Freiburg, Germany Abstract. Aptamers, short synthetic RNA/DNA molecules binding specific targets with high affinity and specificity, are utilized in an increasing spectrum of bio-medical applications. Aptamers are identified in vitro via the Systematic Evolution of Ligands by Exponential Enrichment (SELEX) protocol. SELEX selects binders through an iterative process that, starting from a pool of random ssDNA/RNA sequences, amplifies target-affine species through a series of selection cycles. HT-SELEX, which combines SELEX with high throughput sequencing, has recently transformed aptamer development and has opened the field to even more applications. HT-SELEX is capable of generating over half a billion data points, challenging computational scientists with the task of identifying aptamer properties such as sequence structure motifs that determine binding. While currently available motif finding approaches suggest partial solutions to this question, none possess the generality or scalability required for HT-SELEX data, and they do not take advantage of important properties of the experimental procedure. We present AptaTRACE, a novel approach for the identification of sequence-structure binding motifs in HT-SELEX derived aptamers. Our approach leverages the experimental design of the SELEX protocol and identifies sequence-structure motifs that show a signature of selection. Because of its unique approach, AptaTRACE can uncover motifs even when these are present in only a minuscule fraction of the pool. Due to these features, our method can help to reduce the number of selection cycles required to produce aptamers with the desired properties, thus reducing cost and time of this rather expensive procedure. The performance of the method on simulated and real data indicates that AptaTRACE can detect sequence-structure motifs even in highly challenging data. ? Equal contribution, these authors are listed in alphabetical order 1 Introduction Aptamers are short RNA/DNA molecules capable of binding, with high affinity and specificity, a specific target molecule via sequence and structure features that are complementary to the biochemical characteristics of the target’s surface. The utilization of aptamers in a multitude of biotechnological and medical sciences has recently dramatically increased. While only 80 aptamer related publications were added to Pubmed in the year 2000, this number has since roughly doubled every 5 years, with 207 records added in 2005 alone, 565 additional inclusions in 2010, and as many as 957 new manuscripts indexed in 2014. This astonishing trend is in part attributable to the considerable diversity of possible targets which span from small organic molecules [1], over transcription factors [2] and other proteins or protein complexes [3], to the surfaces of viruses [4] and entire cells [5]. This broad range of targets makes aptamers suitable candidates for a variety of applications ranging from molecular biosensors [6], to drug delivery systems [7], and antibody replacement [8] to just name a few. While the specifics vary depending on the target, aptamers are typically identified through the Systematic Evolution of Ligands by Exponential Enrichment (SELEX) protocol [9]. SELEX leverages the well established paradigm of in vitro selection by repetitively enriching a pool of initially random sequences (species) with those that strongly bind a target of interest. Specifically, based on the assumption that a large enough initial pool of randomized (oligo)nucleotides contains some species with favorable sequence and structure allowing for binding to the target, these binders are then selected for through a series of selection cycles. Each such cycle involves (a) incubating the pool with the target molecules, (b) partitioning target-bound species from non-binders and (c) removing the latter from the pool, followed by (d) elution of the bound fraction from the target, and (e) amplifying the remaining sequences via polymerase chain reaction (PCR) to form the input for the subsequent round. After a target-specific number of selection cycles, the final pool is then used to extract dominating, putatively high-affinity species, via traditional cloning experiments, computational analysis, and binding affinity assays. Depending on their intended application, favorable binders are often further post-processed in vitro to meet additional requirements such as improved structural stability or reducing the size of the aptamer to the relevant binding region. Another reason for the resurgence of interest in aptamer research relates to the utilization of affordable next-generation sequencing technologies along with traditional SELEX. This novel protocol, called HT-SELEX, combines Systematic Evolution of Ligands by Exponential Enrichment with high-throughput sequencing. In HT-SLELEX, after certain (or all) rounds of selection (including the initial pool), aptamer pools are split into two samples, the first of which serves as the starting point for the next cycle whereas the latter is sequenced. The resulting sequencing data, consisting of 2-50 million sequences per round, is then analyzed in silico in order to identify candidates that experience exponential enrichment throughout the selection [10,11]. The massive amount of sequencing data produced by HT-SELEX opens the opportunity for the study of many aspects of the protocol that were either not accessible in traditional SELEX or that could be realized more accurately given hundreds of millions of data points. One of the most challenging of these problems is the discovery of aptamer properties that facilitate binding to the target. However, development of universal methods for the analysis of HT-SELEX data is challenged by the vast diversity of selection conditions such as temperature, salt concentration and the number of targets in the solution to just name a few. Further, each of the stages (a-e) comprising one selection cycle can be accomplished by a variety of technologies. For instance, choosing between open PCR and droplet PCR for the amplification step has been shown to 1 have a great impact on the diversity of the amplification product [12,13,14]. Even more importantly, the complexity of the target molecule is also of great relevance. As a case in point, it has been shown that in vitro selection against transcription factors, and other molecules that are evolutionary optimized to efficiently recognize specific DNA/RNA targets, requires only a small number of rounds in order to produce high quality aptamers [15,2]. On the other side of the spectrum, in the case of CELL-SELEX, a variation of SELEX in which the pool is incubated with entire cells, the number of required selection cycles and the amount of non-specific binders that emerge during selection is significantly larger. Indeed, such a target can in general accommodate a multitude of binding sites, each exposing different binding preferences and leading to a parallel selection towards unrelated binding motifs. Current motif finding algorithms however, have not been designed with these challenges in mind and the need for the development of novel approaches that address the characteristics specific to the SELEX protocol has become highly relevant. Traditionally, motif discovery has been defined as the problem of finding a set of common sub-sequences that are statistically enriched in a given collection of DNA, RNA, or protein sequences. To date, a large variety of computational methods in this area has been published (see [16,17,18] for a comprehensive review). In one of the earlier works, Lawrence and Reilly [19] introduced an Expectation Maximization (EM) based algorithm for finding motifs from protein sequences. This approach has been consequently adopted by various other methods [20,21] including MEME [22] – one of the most widely used programs in this category. Lawrence et al. also introduced a Gibbs sampling approach for motif identification [23] which laid the grounds for other methods such as AlignACE [24], MotifSampler [25], and BioProspector [26] based on this general technique. In addition, numerous approaches have been designed based on efficient counting of all possible k-mers in a data set followed by a statistical analysis of their enrichment. Representatives for this category include Weeder [27], DREME [28], YMF [29], MDScan [30], and Amadeus [31]. Kuang et al. designed a kernel based technique around a set of similar k-mers with a small number of mismatches to extract short motifs in protein sequences [32]. Another group of algorithms that also allows for elucidation of motifs with mismatches is built on suffix tree techniques (Sagot [33], Pavesi et al. [34], and Leibovich [35]). Furthermore, regression based methods have been developed that take additional information, such as the affinity of the input sequences or the genomic regulatory contexts into account. These include, but are not limited to MatrixREDUCE [36], PREGO [37], ChIPMunk [38], and SeqGL [39]. For more information, we refer the reader to Weirauch et al. [40] for a comprehensive evaluation of many of the above techniques. Finally, a number of approaches for the identification of sequence motif in HT-SELEX data targeting transcription factors (TF-SELEX) have been published. One representative of this category is BEEML [41], which is, to our knowledge, the first computational method for finding motifs on this type of high-throughput sequencing data. Assuming the existence of a single binding motif, the method aims at fitting a binding energy model to the data which combines independent attributes from each position in the motif with higher order dependencies. Another method by Jolma et al. approaches the problem by using k-mers to construct a position weight matrix in order to infer the binding models [2,42]. Similarly, Orenstein et al. [43] also uses a k-mer approach based on frequencies from a single round of selection to identify binding motifs for transcription factor HT-SELEX data. Notably, despite of HT-SELEX’ capability of generating data from multiple rounds of selection, all currently existent methods are based on the analysis of only a single selection cycle. However, choosing the round for optimal motif elucidation is not always trivial, and while some effort has been made to address this question (see 2 for example Orenstein and Shamir [43], Jolma et al. [2]) this decision is ultimately left to the user. The search for motifs in the context of RNA sequences faces another dimension in complexity as binding of ssDNA and RNA molecules is known to be sequence and structure dependent. In particular, it has been proposed that binding regions in those molecules tend to be predominantly single stranded [44]. MEMERIS [45], for instance, leverages this assumption by weighting nucleotides according to their likelihood of being unpaired. These positional weights then guide MEME to focus the motif search on loop regions. In contrast, RNAcontext [46] divides the single stranded contexts into known secondary substructures such as hairpins, bulge loops, inner loops, and stems. Consequently, RNAcontext is capable of reporting the relative preference of the structural context along with the primary structure of the potential motif. Recently, Hoinka et al. introduced AptaMotif [47] a method to discover sequence–structure motifs from SELEX derived aptamers. This method utilizes information about the structural ensemble of aptamers obtained by enumerating of all possible structures within a user-defined energy range from the Minimum Free Energy (MFE) structure. By representing each aptamer by the set of its unique substructures (i.e. hairpins, bulge-loops, inner-loops, and multibranch loops), AptaMotif applies an iterative sampling approach combined with sequence-structure alignment techniques to identify high-scoring seeds which are consequently extended to motifs over the full data set. However, AptaMotif was designed for sequencing data obtained from traditional SELEX, under the assumption that this data predominantly consists of motif containing sequences. Subsequently, APTANI [48] extended AptaMotif to handle larger sequence collections via a set of parameter optimizations and sampling techniques, but it also expects a high ratio of motif occurrences. Still, none of the above mentioned methods address the full spectrum of challenges when analyzing data from HT-SELEX selections. First, none of these approaches, as currently implemented, scales well with the data sizes produced by modern high throughput sequencing experiments. Next, only a few of the methods consider the existence of secondary motifs while the majority operates under the assumption that only a singe primary motif is present in the data. This assumption might apply to TF-SELEX, but it cannot be generalized to common purpose HT-SELEX but where one should consider many motifs of possibly similar strength. Furthermore, secondary structure information, which has proven effective in guiding the motif search to biologically relevant binding sites, is not included in most of these methods. A notable exception is RNAContext which can handle relatively large data sets but suffers from the single motif assumption that cannot be easily removed. Finally, none of these approaches attempt to utilize the full scope of the information produced by modern HT-SELEX experiments that includes sequencing data from multiple rounds of selection. In order to close this gap, we have developed AptaTRACE, a method for the identification of sequence-structure motifs for HT-SELEX that utilizes the available data from all sequenced selection rounds, and which is robust enough to be applicable to a broad spectrum of RNA/ssDNA HT-SELEX experiments, independent of the target’s properties. Furthermore, AptaTRACE is not limited to the detection of a single motif but capable of elucidating an arbitrary number of binding sites along with their corresponding structural preferences. AptaTRACE approaches the sequence-structure motif finding problem in a novel and unique way. Unlike previous methods, it does not rely on aptamer frequency or its derivative - cycle-to-cycle enrichment. Aptamer frequency has been recently shown to be a poor predictor of aptamer affinity [10,49,50], and while cycle-to-cycle enrichment has shown a somewhat better performance, 3 the choice of the cycles to compare is not obvious and does not always allow for extraction of sequence-structure motifs. In contrast, our method builds on tracing the dynamics of the SELEX process itself to uncover motif-induced selection trends. We applied AptaTRACE to sequencing data obtained from realistically simulating SELEX over 10 rounds of selection (4 million sequences per round) with known binding motifs as well as to an in vitro cell-SELEX experiment over 9 selection cycles (40 million sequences per cycle). In both cases, our method was successful in extracting highly significant sequence-structure motifs while scaling well with the 10-fold increase in data size. 2 Results We start with a high-level outline of the method, followed by a more detailed description. Next, we use simulated data produced with a novel, extended version of our AptaSim program [51] to compare the performance of AptaTRACE to other methods that can handle similar data sizes or incorporate secondary structure into their models. Finally, we show our results of applying AptaTRACE to an in vitro selection consisting of high-throughput data from 9 rounds of cellSELEX [14]. 2.1 Top Level Description of the Algorithm Our method builds on accepted assumptions regarding the general HT-SELEX procedure. First, we assume that the affinity and specificity of aptamers are mainly attributed to a combination of localized sequence and structural features that exhibit complementary biochemical properties to a target’s binding site. Given a large number of molecules in the initial pool it is expected that such binding motifs are embedded in multiple, distinct aptamers. Consequently, during the selection process, aptamers containing these highly target-affine sequence-structure motifs will become enriched as compared to target non-specific sequences. Notably, under these assumptions, aptamers that contain only the sequence motif without the appropriate structural context are either not enriched at all or enriched to a much lower degree. The second critical assumption we make is the existence of a multitude of sequence-structure binding motifs that either compete for the same binding site, or are binding to different surface regions of the target [6]. Leveraging the above properties of the SELEX protocol, AptaTRACE detects sequencestructure motifs by identifying sequence motifs which undergo selection towards a particular secondary structure context. Specifically, we expect that in the initial pool the structural contexts of each k-mer are distributed according to a background distribution that can be determined from the data. However, for sequence motifs involved in binding, in later selection cycles, this distribution becomes biased towards the structural context favored by the binding interaction with the target site. Consequently, AptaTRACE aims at identifying sequence motifs whose tendency of residing in a hairpin, bugle loop, inner loop, multiple loop, danging end, or of being paired converges to a specific structural context throughout the selection. To achieve this, for each sequenced pool we compute the distribution of the structural contexts of all possible k-mers (all possible nucleotides sequences of length k) in all aptamers. Next, we use the relative entropy (KL-divergence) to estimate, for every k-mer, the change in the distribution of its secondary structure contexts (K-context distribution, for short) between any cycle to a later cycle. The sum of these KL-divergence scores over all pairs of selection cycles defines the context shifting score for a given k-mer. The context shifting score is thus an estimate of the selection towards the preferred structure(s). Complementing the context shifting 4 score is the K-context trace, which summarizes the dynamics of the changes in the K-context distribution over consecutive selection cycles. In order to assess the statistical significance of these context shifting scores, we additionally compute a null distribution consisting of context shifting scores derived from k-mers of all low-affinity aptamers in the selection. This background is used to determine a p-value for the structural shift for each k-mer. Predicted motifs are then constructed by aggregating overlapping k-mers under the restriction that the structural preferences in the overlapped region are consistent. Finally, Position Specific Weight Matrices (PWM), specifically their sequence logos representing these motifs, along with their motif context traces (the average k-context traces of the k-mers used in the PWM construction) are reported to the user. 2.2 Detailed Description of AptaTRACE AptaTRACE takes as input the sequencing results from all, or a subset of selection cycles from an HT-SELEX experiment and outputs a list of position specific weight matrices (PWMs) along with a visual representation of the motifs structural context shift throughout the selection. K-context and K-context Distribution. Any individual occurrence of a k-mer in an aptamer has a specific secondary structure context called K-context that depends on the structure of that particular aptamer. In what follows, let Ki be the i-th k-mer (using an arbitrary indexing of all 4k possible k-mers over the alphabet Ωs = {A, C, G, T }). In addition, let Rx be the set of unique aptamers sequenced in selection round x that have a frequency above a threshold α (this facilitates noise reduction - see computation of p-value below and Fig. 1, A). First, for every aptamer a of fixed length n in Rx , we use SFold [52] to estimate the probability for each nucleotide in a of being part of a hairpin (H), an inner loop (I), a bulge loop (B), a multi-loop (M), a dangling end (D), or being paired (P) (Fig. 1, B). Each aptamer a is hence associated with a matrix of dimension |ΩC | × n, where ΩC = {H, I, B, M, D, P }, in which rows correspond to a particular context C while each column contains the context probabilities of the corresponding nucleotide in a. Next, we define the K-context of a k-mer occurrence in aptamer a as the row-wise mean of the context probabilities over the matrix columns corresponding to the location of that k-mer in the aptamer sequence. Recall, that the main idea behind AptaTRACE is to track the changes in secondary structure preferences of k-mers over the selection cycles. Capturing these secondary structure preferences should therefor take the entirety of K-contexts from all occurrences of a k-mer in a particular selection cycle into account. Thus, we define the K-context distribution of a k-mer Ki in round x as the averaged secondary structure profile of all K-contexts of Ki over all aptamers in Rx . Formally, let Pxi (C), where C ∈ ΩC , be the average probability of the structural context C over all occurrences of the k-mer Ki in all aptamers that meet the threshold criteria in round x. Then, the K-context distribution of Ki in round x is the vector Pxi = [Pxi (H), Pxi (B), Pxi (I), Pxi (M ), Pxi (D), Pxi (P )], normalized such that all entries sum up to one. (Fig. 1 C). Analysis of the Shift of K-context Distributions during Selection. If a k-mer forms part of a sequence-structure binding motif, its K-context distribution is expected to shift towards the context C that is preferred for the binding interaction throughout the selection. In contrast, if a k-mer is not affected by selection, we expect little to no change in its context distribution over consecutive rounds. We can capture this dynamics for any k-mer Ki by its 5 Fig. 1. Schematic overview of our AptaTRACE method. (A) For each cycle, all sequences with frequency above a user defined threshold α are selected as input. (B) Computation of secondary structure probability profiles for each aptamer using SFOLD. For each nucleotide the profile describes the probability of residing in a hairpin, bugle loop, inner loop, multiple loop, danging end, or of being paired. (C) K-context and K-context distribution calculation for each k-mer. (D) Generation of the K-context trace for each k-mer. (E-G) K-mer ranking and statistical significance estimation. Given any two selection cycles the relative entropy (KL-divergence) is used to estimate the change in the distribution of its K-context distribution. The sum of these KL-divergence scores over all pairs of selection cycles defines the context shifting score for a given k-mer. In order to assess the statistical significance of these context shifting scores, a null distribution is computed consisting of context shifting scores derived from k-mers of all low-affinity aptamers in the selection (frequency ≤ α). This background is used to determine a p-value for the structural shift for each k-mer. Top scoring k-mers are selected as seeds. (H) Predicted motifs are constructed by aggregating k-mers overlapping with the seed under the restriction that the structural preferences in the overlapped region are consistent. (I) Position Specific Weight Matrices representing these motifs, along with their K-context traces are reported to the user. 6 so called K-context trace Ki = [P0i , P1i , . . . , Pm i ], defined as a vector tracking the K-context distribution over all m selection cycles (Fig 1 D). Our method consequently quantifies such shifts in the K-context distribution using the Kullback–Leibler divergence (relative entropy) – a measure for the difference between two probability distributions. Here, for any k-mer Ki the first distribution corresponds to the K-context distribution Pxi of an earlier round x and the second to the K-context distribution Pyi of a later selection cycle y. The KL-divergence between two appropriately chosen selection cycles might suffice, at least for some scenarios such as TF-SELEX, to capture the shifts in K-context distributions. In practice however, for larger and more complex targets the selection landscape tends to be more complicated with various aptamers achieving peak enrichment at different selection cycles. Thus makes it rather difficult to confidently chose such two presumably most informative cycles while ignoring remaining information. Therefore we compute the cumulative KL-divergence between all pairs of sequenced pools. In summary, we define the context shifting score score(ki ) for k-mer Ki as score(Ki ) = m−1 X m X DKL (Pyi kPxi ) x=1 y=x+1 = m−1 X m X  X  x=1 y=x+1 C∈ΩC  y P (C)  Pyi (C) × log xi Pi (C) where DKL (P kQ) is the Kullback–Leibler divergence between two discrete distributions P and Q. To ensure statistical accuracy, the context shifting scores are only calculated for all k-mers with a count of at least β individual occurrences in each pool (here, β = 100). Significance Estimation and p-value Computation. While the context shifting score establishes a ranking of the k-mers in order of their overall change in secondary structure context, it does not provide any information over the statistical significance of that shift, i.e. it cannot distinguish between changes in response to the true selection pressure and changes associated with background noise such as non-binding species. These background species however are expected to occur in very low numbers throughout the selection. We leverage this property by using the context shifting scores of the k-mers from these low-count aptamers to construct a null distribution that is used to identify the significant context shifting scores for the full data set. In detail, we include all k-mer occurrences from aptamers that are not included in the previous generation of the context profiles, i.e. all aptamers below or equal to the user defined threshold α. We note that the resulting null follows a log-normal distribution in our in vitro experiment as well as for the simulation data presented in this study (see Section 2.4). The above described procedure hence allows for the computation of a p-value for each Kcontext trace and we only retain those K-context traces with p-value below a user specifiable threshold (the default value is 0.01) (Fig.1 E-G). Elucidating Sequence-Structure Motifs and Sequence Logos. In the last step, AptaTRACE proceeds to extract the final motifs by clustering similar and overlapping k-mers with correlating, statistically significant structural shifts together (Fig 1 H). This allows to uncover sequence-structure motifs that might extend over the chosen k-mer size and to build PWMs that summarize the motifs. Motif construction is accomplished iteratively. Until all k-mers have 7 Table 1. Sequence-structure motifs identified by AptaTRACE from virtual SELEX given all 10 selection cycles including the initial pool as input. AptaTRACE was able to recover all 5 motifs. Shown here are the identified sequence logos, the k-mer that scored highest in significance used for construction of each motif (seed) and its p-value, the abundance of the motif in the final selection round (Frequency), the first cycle at which the motif was detected (C ∗ ), as well as the motif context trace throughout the selection from the initial pool to round 10. Sequence Logos Logo Seed Seed p-Value Frequency C∗ Motif Context Trace 0 1 2 3 4 5 6 7 8 9 10 TCGG GTGTACTA GTTAA GGAGCG GGAACTC AG G CC T TA AGTCGG 4.37E-36 24% 5 CC AT TA A D D D D D D D D D D D GTGTAC 1.91E-32 17.3% 5 CG G AA T G C T G B B B B M B M M M M M TGGTTA 2.19E-23 4.6% 9 D D D D D D D D D D D GGAGCG 2.83E-25 3.6% 8 T A G T M M M M M M M M M M M D D D D D GGAACT 1.12E-27 1.9% T C 10 D D D PPPPPPPPHHH HHHHHHHHPP P I I I I I I I I I I I M M M M M M M M M M M A C A PPPPPPPPPPP HHHHHHHHHHH I I I I I I I I I I I B B B B B B B B B B B GT A G TG HHH M M M M M M M M M M M B B B B B B B B B B B A AA PPPPPPPPPP P HHHHHHHH I I I I I I I I I I I C CC T D D D D D D D D D D D T A HHHHHH P I I I I I I I I I I I B B B B B B M M M M M C A H M M M M M M M M M M M B B B B B B B B B B B C TG T C A G C P PPPPPPPPPP HH HHHH PPPPPPP HPPP HHHHHHHH I I I I I I I I I I I B D D D B B B B B B B D D D B B B D D D D D been assigned to a cluster, the most significant k-mer is first selected as seed, and all similar and highly overlapping k-mers with a p-value below the defined threshold and with comparable structural context are aggregated to the cluster. The details of this relatively straightforward procedure are described in the Supplemenary Materials and Methods section B. In a last step, the resulting motifs are reported to the user via their PWMs, sequence logos and their motif context traces, defined as the averaged K-context traces of those k-mers constituting the PWM. The set of aptamer candidates that satisfy both, the primary and secondary structure properties of the motifs, sorted by their statistical significance or frequency of occurrence is also included in the output (Fig. 1 I). 2.3 Results on Simulated Data To test our new approach, we applied AptaTRACE to a data set generated by means of in-silico SELEX as no benchmarking set that could be used as a gold standard is currently available. To this end, we used an extension to our AptaSim program [51] designed to realistically simulate target-specific selection including, among other factors, species affinity, polymerase amplification and polymerase errors, and the effects of sampling from the selection pools for sequencing. Our current extension additionally allows for implanting sequence-structure motifs with well defined properties. We generated a data set of 4 million sequences per round containing 5 motifs (denoted here as motifs (a)-(e)), 5-8 nucleotides in length located predominantly in unpaired regions. Note that the motif sequence also occurs randomly in the background sequences, albeit in arbitrary structural contexts, and is hence not over-represented in the initial pool.Each motif was initially present in 100 different target-affine aptamer species and consequently selected for over 10 rounds of SELEX. A complete description of the simulation as well as the parameters used during in silico SELEX are available in Supplementary Material and Methods Sections A and E, respectively. We applied AptaTRACE, as well as DREME and RNAcontext to the data set to compare their capability of extracting these motifs. Since DREME and RNAcontext can only be applied to one selection round at a time, we provided these two approaches with data from the last selection 8 cycle alone, choosing the initial pool as background when required. AptaTRACE was applied to both, the reduced data set, as well as to all selection cycles. Notably, neither DREME nor RNAcontext are capable of handling 4 million sequences in a reasonable time frame, prompting us to sample 10% of aptamers from the last and the unselected round as the input for DREME, and the 10000 most frequent and least frequent sequences of the last selection cycle for RNAcontext. The full scope of parameters used for these methods during the comparison are detailed in Supplementary Material and Methods E. Since RNAcontext’s model assumes a single motif in the data, a direct comparison would not be fair for that software. Nonetheless, we examined the possibility of the method of identifying at least one binding site due to the large abundance of implanted motif (a) in the final selection round, however without success. Tab. 2 summarizes the results of AptaTRACE when applied to the full dataset, as well as the to last selection cycle only, compared to DREMEs performance. While DREME failed to identify the low-affinity motif (e) as well as the shorter but more targetaffine motif (c), AptaTRACE was able to recover all motifs in both test scenarios. A more detailed summary of the sequence logos extracted by our approach on the full data set, including their motif context traces and statistical significance, is available in Tab. 1. Interestingly, a visual inspection of the motif context trace (last column, Tab. 1) points to the possibility of capturing most of these motifs at earlier cycles. Indeed, computing the selection round in which a motif was first detected by AptaTRACE (column C ∗ , Tab.1), confirmed this expectation. Table 2. Comparison of AptaTRACE against other methods based on simulated data. AptaTRACE was applied to the entire dataset as well as to the last selection cycle only. While our method successfully identified all implanted motifs, DREME was only able of extracting 3 out of 5. We show the implanted motifs, their binding affinity used throughout the selection (B.A.) in the first two columns. The output PWMs produced by the tested methods that correspond to the implanted motifs are displayed in the remaining columns. (a) Motif B.A. GTGTACTA 75% AptaTRACE (all cycles) TCGG GTGTACTA TATGTC TAGGGC GGAGCG AG G CC T TA (b) AGTCGG 70% CC AT TA C TG CG G AA T T C A G C A (c) GTTAA 60% T (d) (e) GGAACTC GGAGCG 55% 50% A C T A 2.4 AptaTRACE (last cycle) DREME (last cycle) GTGTACTA GTGTACTA AGTCGG T C A G C A GC G TCG AA T C TCGG GTTAA G AACT GGAGCG AG G T CC TA CC AT TA G C T G C T A G T C T C T A AC T C A A G G G AA C T TA GGAACTC G A T T A AGC Results on Cell-SELEX Data Next, we applied AptaTRACE to the results of an in-vitro HT-SELEX experiment where the initial pool as well as 7 of 9 selection rounds have been sequenced, averaging 40 million aptamers per cycle (see Section C for a detailed description of the experimental procedure). We did not challenge DREME with this task, since this data set is 10-fold larger in size compared to the simulated selection, and even in the latter case DREME managed to only handle 10% of the data. AptaTRACE was able to successfully extract a total of 25 motifs, the five most frequent of which are shown in Tab. 3, and a full list is given in Supplementary Tab. 4. 9 Table 3. Five most frequent sequence-structure motifs as produced by AptaTRACE on CELL-SELEX data. The sequence logo as well as the most frequent k-mer constituting the logo (Logo Seed) and its p-value are depicted for each motif. The motif context trace for the sequenced cycles (0,1,3,5,6,7,8,9) is shown in the last column. Sequence Logo Logo Seed Seed p-Value Frequency Motif Context Trace 0 TGTGC TTATT GTTT A ATGTT GTGTCC C G CTGTG 5.31E-03 32.98 % G C T A T TTATT 7.16E-03 25.14 % G G C T C T MMMMM PPP HPPPPP HH MH H H I H H I I I I I I I M D D D D D D GTTTA 9.71E-03 16.77 % MMMPM PPP HPPPMP HH MH H H H H I I I I I I I I M M B B B B B B B B D D D D ATGTT 3.66E-03 15.27 % D D D D PPPPPPPP HHHMMMMM I H H H I H H I I M I I I I M M B B B B B B B B D D D D D D D GTGTC 6.61E-03 13.95 % G A HH M G A HHHHH PPP HPPPPP D D CA C 9 B B B B B B B B A T A CT 8 M M M AA G 7 B B B B B A A 6 M M M M M C CA C 5 B B I I I I B I GC A 3 I I I T A G G C 1 PPPPPPPP B B B I B B I H I B I I I I H M M M M I M B B H H H H H M H M M The context trace of these motifs hints towards two properties of the selection process. First, a clear selection towards single stranded regions for every extracted motif can be observed. It has always been stipulated that ssDNA/RNA binding motifs are most likely located in loop regions [44]. Indeed, this assumption was leveraged by MEMERIS, by imposing structural priors directing the motif search towards single stranded regions. In the case of AptaTRACE, no prior assumption of this type was made. The fact that despite a lack of such priors, motifs detected by AptaTRACE conform with the expected properties of RNA sequence-structure binding sites support their relevance for binding. Next, the trend of the structural preferences of these motifs emerges relatively early during the selection process indicating that, in conjunction with our method, the identification of biologically relevant binding sites in general purpose HT-SELEX data might be possible with fewer selection cycles. 3 Conclusion Aptamers have a broad spectrum of applications and are increasingly being used to develop new therapeutics and diagnostics. HT-SELEX, in contrast to the traditional protocol, provides data for a global analysis of the selection properties and for simultaneous discovery of an large number of candidates. This extensive amount of information has utility only in conjunction with suitable computational methods to analyze the data. Unlike in traditional SELEX, where only a handful of potential binders are retrieved and exhaustively tested experimentally, HT-SELEX returns a massive amount of sequencing data sampled from some, or all, selection rounds. This data consequently serves as the basis for the challenging task of identifying suitable binding candidates and for deriving their sequencestructure properties that are key for binding affinity and specificity. Except for the special case of TF binding aptamers, no previous tool addressing this task existed. The realization that a naive relationship between aptamer frequency and their binding affinity is not universally valid, further complicates this task. Several potential factors during any stage of the selection contribute to this complexity, including polymerase amplification biases, sequencing biases, contamination of foreign sequences, and non-specific binding. These factors prompted aptamer experts to consider cycle-to-cycle enrichment instead of frequency counts as a predictor for binding affinity. While cycle-to-cycle enrichment did increase the predictive power of these methods, it cannot bypass 10 problems related to amplification bias nor can it identify aptamer properties that drive binding affinity and specificity. In contrast, AptaTRACE is specifically designed to identify sequence-structure binding motifs in HT-SELEX data and thus to predict the features behind binding affinity and specificity. Importantly, rather than using quantitative information, it directly leverages the experimental design of the SELEX protocol and identifies motifs that are under selection through appropriately designed scoring functions. By focusing on local motifs that are selected for, AptaTRACE bypasses global biases such as the PCR bias which is typically related to more universal sequence properties such as the CG content. In addition, because AptaTRACE measures selection towards a given sequence-structure motif by the shift in the distribution of the structural context and not based on abundance, it can uncover statistically significant motifs that are selected for even when these only form a small fraction of the pool. This is an important property that can ultimately help to shorten the number of cycles needed for selection and thus to reduce the overall cost of the procedure. In testing on simulated data, AptaTRACE outperformed other methods, in part because these methods models were not specifically designed to handle these sort of sequences. Furthermore, no competitors exist that could be tested on in vitro data as none of the current programs scale to the amount of data points produced by HT-SELEX. While we currently have no gold standard for experimental data to measure the quality of the identified motifs, it is reassuring that the motifs converge structurally to loop regions, consistent with the accepted view where such binding sites reside. Sequence logos provide a convenient visualization of the selected motifs. For TF-binding, information used to derive these logos can be used to estimate binding energy [53,54]. However, for general HT-SELEX this connection is less immediate as one has to take into account the energy contribution from the structure component [55]. Perhaps even more importantly, aptamers binding to large cell surfaces are likely to be exposed to more binding opportunities than aptamers binding to single receptors and thus the number of resulting motifs can be expressed as a function of interaction probability and binding affinity. We hypothesize that the here presented K-context trace will be helpful in untangling some of these contributions. Finally, analysis of the K-context trace indicates that the selection signal can be identified at very early cycles. This suggests that, with deep enough sequencing, only a limited number of selection cycles might be required. Yet, this analysis also shows that the dynamics of Kcontext traces is not the same for all sequence-structure motifs. While most trends essentially stabilize at a relatively early cycle, some continue to grow. We hypothesize that this type of information can aid the identification of the most promising binders. Note that while we defined the K-context shifting score on all pairs of sequenced selection pools, it can also be used to focus the analysis to any part of the selection as long as it includes at least two cycles and hence to center on additional details of the selection dynamics. Other variants of the K-context shifting score (e.g. always using the initial pool as background/reference in the summation) can also prove informative, however full elucidation of this dynamic will require concerted computational and experimental effort. AptaTRACE is not only a powerful method to detect emerging sequence-structure motifs but also, a flexible tool to interrogate such selection dynamics. 11 References 1. Y. S. Kim and M. B. Gu. Advances in Aptamer Screening and Small Molecule Aptasensors. Advances in biochemical engineering/biotechnology, Jul 2013. PMID: 23851587. 2. A. Jolma, T. Kivioja, J. Toivonen, L. Cheng, G. Wei, M. Enge, M. Taipale, J. M. Vaquerizas, J. Yan, M. J. Sillanpaa, M. Bonke, K. Palin, S. Talukder, T. R. Hughes, N. M. Luscombe, E. Ukkonen, and J. Taipale. Multiplexed massively parallel SELEX for characterization of human transcription factor binding specificities. Genome Res., 20(6):861–873, Jun 2010. 3. A. Berezhnoy, C. A. Stewart, 2nd Mcnamara, J. O., William Thiel, Paloma Giangrande, Giorgio Trinchieri, and Eli Gilboa. Isolation and optimization of murine IL-10 receptor blocking oligonucleotide aptamers using high-throughput sequencing. Molecular therapy: the journal of the American Society of Gene Therapy, 20(6):1242–1250, Jun 2012. PMID: 22434135. 4. J. M. Binning, T. Wang, P. Luthra, R. S. Shabman, D. M. Borek, G. Liu, W. Xu, D. W. Leung, C. F. Basler, and G. K. Amarasinghe. Development of RNA aptamers targeting Ebola virus VP35. Biochemistry, Sep 2013. PMID: 24067086. 5. H. Shi, W. Cui, X. He, Q. Guo, K. Wang, X. Ye, and J. Tang. Whole Cell-SELEX Aptamers for Highly Specific Fluorescence Molecular Imaging of Carcinomas In Vivo. PloS one, 8(8):e70476, 2013. PMID: 23950940. 6. R. Zichel, W. Chearwae, G. S. Pandey, B. Golding, and Z. E. Sauna. Aptamers as a sensitive tool to detect subtle modifications in therapeutic proteins. PloS one, 7(2):e31948, 2012. PMID: 22384109. 7. R. K. Upadhyay. Nucleic Acid Aptamer-Guided Cancer Therapeutics and Diagnostics: the Next Generation of Cancer Medicine. Theranostics, 5(1):32–42, Jan 2015. 8. Macugen. Fda approves new drug treatment for age-related macular degeneration. 9. A. D. Ellington and J. W. Szostak. In vitro selection of RNA molecules that bind specific ligands. Nature, 346(6287):818–822, Aug 1990. PMID: 1697402. 10. J. Hoinka, A. Berezhnoy, Z.E. Sauna, E. Gilboa, and T.M. Przytycka. AptaCluster - A Method to Cluster HT-SELEX Aptamer Pools and Lessons from its Application. Res Comput Mol Biol., 8394(1):115–128, Jan 2014. 11. K. K. Alam, J. L. Chang, and D. H. Burke. FASTAptamer: A Bioinformatic Toolkit for High-throughput Sequence Analysis of Combinatorial Selections. Mol Ther Nucleic Acids, e230, Mar 2015. 12. M. U. Musheev and S. N. Krylov. Selection of aptamers by systematic evolution of ligands by exponential enrichment: addressing the polymerase chain reaction issue. Anal Chim Acta., 564(1):91–96, Mar 2005. 13. R. Yufa, S. M. Krylova, C. Bruce, E. A. Bagg, C. J. Schofield, and S. N. Krylov. Emulsion PCR significantly improves nonequilibrium capillary electrophoresis of equilibrium mixtures-based aptamer selection: allowing for efficient and rapid selection of aptamer to unmodified ABH2 protein. Anal Chem., 87(2):1441–1449, Jan 2005. 14. M. Takahashi, J. Zhou, J. J. Rossi, and J. C. Burnett. A Comparative Study for Aptamer Selection and PCR Bias Using Droplet Digital PCR versus Open PCR. Manuscript in Preparation, Oct. 15. G. V. Kupakuwana, 2nd Crill, J. E., M. P. McPike, and P. N. Borer. Acyclic identification of aptamers for human alpha-thrombin using over-represented libraries and deep sequencing. PloS one, 6(5):e19395, 2011. PMID: 21625587. 16. M. Tompa, N. Li, T. L. Bailey, G. M. Church, B. De Moor, E. Eskin, A. V. Favorov, M. C. Frith, Y. Fu, W. J. Kent, V. J. Makeev, A. A. Mironov, W. S. Noble, G. Pavesi, G. Pesole, M. Regnier, N. Simonis, S. Sinha, G. Thijs, J. van Helden, M. Vandenbogaert, Z. Weng, C. Workman, C. Ye, and Z. Zhu. Assessing computational tools for the discovery of transcription factor binding sites. Nat. Biotechnol., 23(1):137–144, Jan 2005. 17. M. K. Das and H. K. Dai. A survey of DNA motif finding algorithms. BMC Bioinformatics, 8 Suppl 7:S21, 2007. 18. F. Zambelli, G. Pesole, and G. Pavesi. Motif discovery and transcription factor binding sites before and after the next-generation sequencing era. Brief. Bioinformatics, 14(2):225–237, Mar 2013. 19. C. E. Lawrence and A. A. Reilly. An expectation maximization (EM) algorithm for the identification and characterization of common sites in unaligned biopolymer sequences. Proteins, 7(1):41–51, 1990. 20. S. Sinha, M. Blanchette, and M. Tompa. PhyME: a probabilistic algorithm for finding motifs in sets of orthologous sequences. BMC Bioinformatics, 5:170, Oct 2004. 21. J. E. Reid and L. Wernisch. STEME: efficient EM to find motifs in large data sets. Nucleic Acids Res., 39(18):e126, Oct 2011. 22. T. L. Bailey and C. Elkan. Unsupervised learning of multiple motifs in biopolymers using expectation maximization. Machine Learning, 21:51, 1995. 12 23. C. E. Lawrence, S. F. Altschul, M. S. Boguski, J. S. Liu, A. F. Neuwald, and J. C. Wootton. Detecting subtle sequence signals: a Gibbs sampling strategy for multiple alignment. Science, 262(5131):208–214, Oct 1993. 24. F. P. Roth, J. D. Hughes, P. W. Estep, and G. M. Church. Finding DNA regulatory motifs within unaligned noncoding sequences clustered by whole-genome mRNA quantitation. Nat. Biotechnol., 16(10):939–945, Oct 1998. 25. G Thijs, K Marchal, Lescotm M, S Rombauts, B. De Moor, Rouzé. P., and Y. Moreau. A gibbs sampling method to detect overrepresented motifs in the upstream regions of coexpressed genes. Journal of Computational Biology, 9(2), 2002. 26. X. Liu, D. L. Brutlag, and J. S. Liu. BioProspector: discovering conserved DNA motifs in upstream regulatory regions of co-expressed genes. Pac Symp Biocomput, pages 127–138, 2001. 27. G. Pavesi, P. Mereghetti, G. Mauri, and G. Pesole. Weeder Web: discovery of transcription factor binding sites in a set of sequences from co-regulated genes. Nucleic Acids Res., 32(Web Server issue):199–203, Jul 2004. 28. T. L. Bailey. DREME: motif discovery in transcription factor ChIP-seq data. Bioinformatics, 27(12):1653– 1659, Jun 2011. 29. S. Sinha and M. Tompa. Discovery of novel transcription factor binding sites by statistical overrepresentation. Nucleic Acids Res., 30(24):5549–5560, Dec 2002. 30. X. S. Liu, D. L. Brutlag, and J. S. Liu. An algorithm for finding protein-DNA binding sites with applications to chromatin-immunoprecipitation microarray experiments. Nat. Biotechnol., 20(8):835–839, Aug 2002. 31. C. Linhart, Y. Halperin, and R. Shamir. Transcription factor and microRNA motif discovery: the Amadeus platform and a compendium of metazoan target sets. Genome Res., 18(7):1180–1189, Jul 2008. 32. R. Kuang, E. Ie, K. Wang, K. Wang, M. Siddiqi, Y. Freund, and C. Leslie. Profile-based string kernels for remote homology detection and motif extraction. J Bioinform Comput Biol, 3(3):527–550, Jun 2005. 33. M.F. Sagot. Spelling approximate repeated or common motifs using a suffix tree. In LATIN: Latin American Symposium on Theoretical Informatics, 1998. 34. G. Pavesi, G. Mauri, and G. Pesole. An algorithm for finding signals of unknown length in DNA sequences. Bioinformatics, 17 Suppl 1:S207–214, 2001. 35. L. Leibovich, I. Paz, Z. Yakhini, and Y. Mandel-Gutfreund. DRIMust: a web server for discovering rank imbalanced motifs using suffix trees. Nucleic Acids Res., 41(Web Server issue):W174–179, Jul 2013. 36. B. C. Foat, A. V. Morozov, and H. J. Bussemaker. Statistical mechanical modeling of genome-wide transcription factor occupancy data by MatrixREDUCE. Bioinformatics, 22(14):e141–149, Jul 2006. 37. A. Tanay. Extensive low-affinity transcriptional interactions in the yeast genome. Genome Res., 16(8):962– 972, Aug 2006. 38. I. V. Kulakovskiy, V. A. Boeva, A. V. Favorov, and V. J. Makeev. Deep and wide digging for binding motifs in ChIP-Seq data. Bioinformatics, 26(20):2622–2623, Oct 2010. 39. M. Setty and C. S. Leslie. SeqGL Identifies Context-Dependent Binding Signals in Genome-Wide Regulatory Element Maps. PLoS Comput. Biol., 11(5):e1004271, May 2015. 40. M. T. Weirauch, A. Cote, R. Norel, M. Annala, Y. Zhao, T. R. Riley, J. Saez-Rodriguez, T. Cokelaer, A. Vedenko, S. Talukder, H. J. Bussemaker, Q. D. Morris, M. L. Bulyk, G. Stolovitzky, T. R. Hughes, P. Agius, A. Arvey, P. Bucher, C. G. Callan, C. W. Chang, C. Y. Chen, Y. S. Chen, Y. W. Chu, J. Grau, I. Grosse, V. Jagannathan, J. Keilwagen, S. M. Kie?basa, J. B. Kinney, H. Klein, M. B. Kursa, H. Lahdesmaki, K. Laurila, C. Lei, C. Leslie, C. Linhart, A. Murugan, A. My?ickova, W. S. Noble, M. Nykter, Y. Orenstein, S. Posch, J. Ruan, W. R. Rudnicki, C. D. Schmid, R. Shamir, W. K. Sung, M. Vingron, and Z. Zhang. Evaluation of methods for modeling transcription factor sequence specificity. Nat. Biotechnol., 31(2):126– 134, Feb 2013. 41. Y. Zhao, D. Granas, and G. D. Stormo. Inferring binding energies from selected binding sites. PLoS Comput. Biol., 5(12):e1000590, Dec 2009. 42. A. Jolma, J. Yan, T. Whitington, J. Toivonen, K. R. Nitta, P. Rastas, E. Morgunova, M. Enge, M. Taipale, G. Wei, K. Palin, J. M. Vaquerizas, R. Vincentelli, N. M. Luscombe, T. R. Hughes, P. Lemaire, E. Ukkonen, T. Kivioja, and J. Taipale. DNA-binding specificities of human transcription factors. Cell, 152(1-2):327–339, Jan 2013. 43. Y. Orenstein and R. Shamir. Hts-ibis: fast and accurate inference of binding site motifs from ht-selex data. bioRxiv, 2015. 44. C. Schudoma, P. May, V. Nikiforova, and D. Walther. Sequence-structure relationships in RNA loops: establishing the basis for loop homology modeling. Nucleic Acids Res., 38(3):970–980, Nov 2009. 45. M. Hiller, R. Pudimat, A. Busch, and R. Backofen. Using RNA secondary structures to guide sequence motif finding towards single-stranded regions. Nucleic Acids Res., 34(17):e117, 2006. 13 46. H. Kazan, D. Ray, E. T. Chan, T. A. Hughes, and Morris Q. RNAcontext: A New Method for Learning the Sequence and Structure Binding Preferences of RNA-Binding Proteins. PLOS CB, 2010. 47. J. Hoinka, E. Zotenko, A. Friedman, Z. E. Sauna, and T. M. Przytycka. Identification of sequence-structure RNA binding motifs for SELEX-derived aptamers. Bioinformatics, 28(12):i215–223, Jun 2012. 48. J. Caroli, C. Taccioli, A. De La Fuente, P. Serafini, and S. Bicciato. APTANI: a computational tool to select aptamers through sequence-structure motif analysis of HT-SELEX data. Bioinformatics, Sep 2015. 49. M. Cho, Y. Xiao, J. Nie, R. Stewart, Csordas A. T., Oh S. S., J. A. Thomson, and H. T. Soh. Quantitative selection of DNA aptamers through microfluidic selection and high-throughput sequencing. Proceedings of the National Academy of Sciences, 107:15373–15378, Jul 2010. 50. W. H. Thiel, T. Bair, A. S. Peek, X. Liu, J. Dassie, K. R. Stockdale, M. A. Behlke, F. J. Miller Jr, and P. H. Giangrande. Rapid Identification of Cell-Specific, Internalizing RNA Aptamers with Bioinformatics Analyses of a Cell-Based Aptamer Selection. PLOS one, Sep 2012. 51. J. Hoinka, A. Berezhnoy, P. Dao, Z. E. Sauna, E. Gilboa, and T. M. Przytycka. Large scale analysis of the mutational landscape in HT-SELEX improves aptamer discovery. Nucleic Acids Res, 43(12):5699–5707, Jul 2015. 52. Y. Ding and C. E. Lawrence. Statistical prediction of single-stranded regions in RNA secondary structure and application to predicting effective antisense target sites and beyond. Nucleic Acids Res., 29(5):1034–1046, Mar 2001. 53. P. V. Benos, A. S. Lapedes, and Stormo G. D. Probabilistic code for DNA recognition by proteins of the EGR family. J Mol Biol., 323(4):701–727, Nov 2002. 54. P. V. Benos, M. L. Bulyk, and G. D. Stormo. Additivity in protein-DNA interactions: how good an approximation is it? Nucleic Acids Res., 30(20):4442–4451, Oct 2002. 55. T. M. Przytycka and D. Levens. Shapely DNA attracts the right partner. Proc Natl Acad Sci U S A., 112(15):4516–4517, Apr 2015. 14 Supplementary Material And Methods A SELEX Simulation Details In order to create data sets with properties comparable to in-vitro HT-SELEX experiments, we designed a simulation scheme capable of mimicking target-specific selection, error-prone amplification, as well as the stochastic nature of partitioning each cycle into sequencing set and selection set. Furthermore, our simulation allows for the introduction of an arbitrary number of sequence-structure motifs of different sizes and affinities, and provides control over the binding strength of background sequences. The simulation builds on our AptaSIM software [51] and extends it to take secondary structure into account. For each aptamer a, we use |a|, to denote its frequency and abind to denote its biding affinity. For a user-defined pool size, a set of aptamers containing sequence-structure motifs of desired length is first produced using a second order hidden Markov model (HMM) trained with sequences from an in vitro HT-SELEX experiment. To this end, we generate a sequence from the HMM, predict its secondary structure and identify all single-stranded regions larger or equal to the motif size. Next, we substitute the nucleotides in one of these unpaired regions with the motif and verify, using SFold, that the secondary structure profile of the motif region is predominantly single stranded. If the average probability of any of the single stranded secondary structure contexts in the motif region falls below 60%, we discard this sequence, and otherwise add it to the pool. This procedure is repeated as many times as required to create the desired number of sequences containing the motifs. The remaining aptamers are sampled directly from the training data in order to ensure realistic secondary structure properties that sequences generated from the HMM might not necessarily possess. Using this set of sequences as input, we perform a weighted sampling without replacement according to the aptamers count and affinity. I.e., until the desired sample size is reached, we compute a weight w(a) for each aptamer a as |a| ∗ abind , a∈Rx |a| ∗ abind w(a) = P (1) sample a sequence according to this distribution and adjust the weights to account for the removed aptamer. The resulting pool is then amplified by means of virtual PCR in which the amplification efficiency of the polymerase, as well as the mutation rate is adjustable. In order to simulate the sampling effect, we inject a user-defined percentage of singleton aptamers with low affinity, sampled from the training data, into the pool, and reduce the pool size to its original size by weighted sampling without replacement where the weights are generated according to the aptamer counts. After each cycle, the pool is stored in fastq format and used as input for our AptaTRACE algorithm. B Generating Motif Logos and Choosing k A k-mer does not necessarily cover the motif completely. In addition, aptamers can contain slightly modified sequences in the same structural context. Such motif variants can arise in15 dependently or as a result of polymerase errors during the amplification stages of the SELEX protocol. Therefor, our method clusters these elements in an iterative manner. First, k-mers with significant context shift scores are sorted in decreasing order according to their k-mer frequencies in the last selection cycle. We then select the top k-mer as a cluster seed, iterate through each of the remaining k-mers, and use the following decision scheme to choose whether or not to include the k-mer into to the cluster. Fully overlapping k-mers must have at most one mismatch with the seed. For partially overlapping k-mers and k ≤ 6, the longest common substring (LCS) with the seed must be at least 4 nucleotides in length, while the LCS is required to be at least 5 for k ≥ 7. Furthermore, their K-context trace must also satisfy similar structural shifts that correlate with the seed to be included in the cluster. To examine the similarities of the context shifts, for each Ki , we identify the specific structural context C 0 that changes the most over the selection rounds: m−1 m X X 0 C = arg max f (C) = Pyi (C) − Pxi (C) C If C0 x=1 y=x+1 coincides with the dominant context of the seed, we include the K-context to the cluster. The number of clusters that results from this approach can also serve as a measure for an appropriate choice of the k-mer size for a particular data set. By choosing k such that it maximizes the number of the resulting clusters we ensure to capture all non-overlapping motifs in the pool. C CELL-SELEX Experimental Details Cell-based SELEX was performed using an RNA library containing a randomized 30-nt region flanked by fixed primer sequences. Cell-based selection was performed as previously described [1], by employing open PCR for DNA amplification during each selection round. Positive selection was performed on Hela cells transduced with a bicistronic lentiviral vector expressing the target surface receptor and GFP, while unmodified Hela cells, which lack expression of the target receptor, were used for negative selection. High throughput sequencing (HTS) was performed on the positive selection at rounds 0, 1, 3, 5, 6, 7, 8, and 9. Significant molecular enrichment of receptor-specific aptamers was observed after five rounds of selection. D AptaTRACE Results on in vitro Data Table 4: Full set of the sequence-structure motifs as produced by AptaTRACE on CELL-SELEX data. The sequence logo as well as the most frequent k-mer constituting the logo (Logo Seed) and its p-value are depicted for each motif. The K-context trace for the sequenced cycles (0,1,3,5,6,7,8,9) is shown in the last column. Sequence Logo Logo Seed Seed p-Value Frequency Motif Context Trace 0 TGTGC TTATT C G T A A 5.31E-03 32.98 % G 3 5 6 7 8 9 HHHHH PPP HPPPPP HH I I I T A T G C CTGTG G C 1 B B I I I I B I M M M M M M M M B B B B B TTATT 7.16E-03 25.14 % GC C CA A MMMMM PPP HPPPPP HH MH H H I H H I I I I I I I M M B B B B B B B B D D D D D D D D 16 T T TA ATGTT G A C GTTTA 9.71E-03 16.77 % A T A G ATGTT 3.66E-03 15.27 % D D D D PPPPPPP P HHHMMMMM I H H H I H H I I M I I I I M M CA C D D D G G HH B B B B B B B B D C CT MMMPM PPP HPPPMP MH H H H H I I I I I I I I M M AA G B B B B B B B B D D D D D D D TGTCC G T A G GTGTC 6.61E-03 13.95 % T C A PPPPPPPP B B B I B B I H I B I I I I H M M M M I M B B H H H H H M H M M AATTG TGGCC C GCTGC TAATGT ATTAA TGCGCC C GC T A AATAT CTTTG TGGTGC GTAAA GCGTG T TAAGT ATAA GTTAG ATAGT A AATTG 3.10E-03 12.76 % A G T CTGGC 2.46E-03 12.55 % A T T T G I TG A CAA G G CGCTG 7.07E-03 11.30 % A A G TAATG 5.70E-04 10.12 % ATTAA 2.64E-03 10.11 % T I I TGCGC 3.25E-04 9.73 % CTGCA 1.49E-03 7.61 % AATAT 1.90E-03 7.21 % T C C A C G C T G A I I I CTTTG 8.74E-03 6.59 % T G C PPPPPP PP HHMMMM H MH I H I H H H I I I I M M M I I B D B B B D D B D D D D B B B D TGGTG 4.65E-03 5.22 % HHHHH PPPP PPPP HH H I I I I I I I I M B M B B B B B B M B M M M M M GTAAA 2.85E-03 4.79 % MMMMM PPPP PPP HHH P I H H HHH I I I I I M M M I I B B B B B B B D D D D D D GCGTG 3.87E-03 4.08 % B D D HHHHH PPPP PPPP H HH I I I I I I I I B B B M M M M M M B B M B M B B TAAGT 8.04E-03 3.95 % MMMMM PPPP P HHH PPP I H H HHH I I I I I M M M C I I B B B B B B B B D D D D D D CATAA 8.02E-03 3.74 % T C G A HHHMMMMM PPPPPPPP MH HHHH I I I I I M M D A G T B D A G T TA A C M B B B B B B B B D D D D D D C A H M M G G H B C G G HHHH PPPP HPPPP H I I I B B M B M M A T M M M M M M I B I B I I I M B G T B B B B B M T A HHHHH PPP HPPPPP HH I I I I I I B I I B B M M G CC HHHMMMMM PPPHHHHH MPPPPP I I I I I I M M D C G A G I I I D AA C HHHPPPPP B B B B B B B B D D D D D D TA C PPPMMMMM MHHHHH I I I I I M M AG A TG G H D T C HHHHH PPPP PPPP H H I I I B B B B B B B B D D D D D D D G G D M M M T G M D B B B B B G T GG B B M M D I I B M I M I I M M B B M C CC M M D D A C AC A HHHHH PPPP PPPP HHH I I I M A C G D B I I B I I B B B B M M G A I I I I D D G C A HHHPPPPP B B B B B B B B D D D D D G C PPPMMMMM MH I I HHHH I I M M A CG D HHHMMMMM PPPPPHHH M H H PPP I I I I I I M M I I B B B B B B B B D D D D D D D D GTTAG 8.60E-03 3.70 % AC G PPPPPP PP HHMMMMM H H I H H H H I I I I I M I I M M C B B B B B B B B D D D D D D ATAGT 6.15E-03 2.91 % A CC MMMMM PPP HPPPP HH P M H HHHH I I I I I M M G I I I B B B B B B B B D D D D D D D 17 D TGCCGC T G CG C AA T TGCCG 2.78E-04 2.61 % C AA HHHHH PPPP PPPP H H I I I H TG I I I I I B B M M M B M M M M B B M B B B TGCTC GTATG CCTAT A A T T C A C G TGCTC 6.10E-03 2.52 % PPP PPPP P MMMM HH HM I H H H H H I I M I I I I I M B B B B B B B B M D D D D D D D GTATG 9.05E-03 2.30 % PPPPPPP P HHHMMMMM I HHHHH I I M M M I I I I I B B B B B B B B D D D D CCTAT 2.27E-03 1.53 % T M I I H H H M H H M B I I I I G I D D D D B D D B D B B B A D E D D PHP PPPPP HPHMMMMM I B B Parameters used in this Study AptaTRACE: We define all aptamers as low count that have a frequency lower than 3, i.e. we define α = 3. Simulation: We created a simulated dataset of 4 million sequences over 10 selection cycles. The initial round was designed to contain 5 motifs between 5-8 nucleotides in length and with binding affinities of 50%, 55%, 60%, 70%, and 75% respectively (the background affinities were randomly assigned in a range of 1-25%). Each motif was initially represented a total of 100 times in the pool. In each cycle, we sample a total of 40% of the pool, amplify the selected aptamers with an efficiency of 95% and a mutation rate of 5%, and introduce a total of 50% background sequences into the mix. Comparison with other Methods: The k-mer size for DREME was restricted to the range k = 5, . . . , 8 and we limited it to output the top 100 PWM models with an e-value of at most 0.01. For RNAcontext, we utilized the log of the frequency of each sequence to represent their binding affinities. Finally, we report the consensus sequence for each output PWM from DREME, RNAcontext and AptaTRACE. To compute the value for C ∗ , we consider a PWM as a matching an implanted motif if the sequence of the implanted motif is a substring of the consensus sequence of the PWM. F Implementation Details and Runtime Our AptaTRACE pipeline is implemented as a modular system in C++ and Java. RNA secondary structure profiles for all aptamers in the pool were predicted in parallel on the server farm at the National Center for Biotechnology Information, NIH, using 1000 Cores and 200MB or RAM per job. The simulated data required about 5 hours of wall clock time, whereas the in vitro selection data finished in roughly 4 days. The enumeration of k-mers and consequent extraction of k-contexts was implemented as a multi-threaded C++ library and is capable of processing the here presented results in approximately 10 hours for all choices of k between 5-8 on a large memory server with 100 CPUs. Finally, the context shifting scores, k-mer extraction, and clustering is implemented as a multi-threaded JAVA program requiring 3-4 hours, depending on the data size, for its completion. The source code, is currently available on request and will be hosted online at: www.ncbi.nlm.nih.gov/CBBresearch/Przytycka/index.cgi#aptatools 18
5cs.CE
arXiv:1602.08268v1 [q-bio.PE] 26 Feb 2016 Construction of Gene and Species Trees from Sequence Data incl. Orthologs, Paralogs, and Xenologs Marc Hellmuth University of Greifswald Department of Mathematics and Computer Science Walther- Rathenau-Strasse 47, D-17487 Greifswald, Germany, and Saarland University Center for Bioinformatics Building E 2.1, P.O. Box 151150, D-66041 Saarbrücken, Germany Email: [email protected] Nicolas Wieseke Leipzig University Parallel Computing and Complex Systems Group Department of Computer Science Augustusplatz 10, D-04109 Leipzig, Germany Email: [email protected] Abstract Phylogenetic reconstruction aims at finding plausible hypotheses of the evolutionary history of genes or species based on genomic sequence information. The distinction of orthologous genes (genes that having a common ancestry and diverged after a speciation) is crucial and lies at the heart of many genomic studies. However, existing methods that rely only on 1:1 orthologs to infer species trees are strongly restricted to a small set of allowed genes that provide information about the species tree. The use of larger gene sets that consist in addition of non-orthologous genes (e.g. so-called paralogous or xenologous genes) considerably increases the information about the evolutionary history of the respective species. In this work, we introduce a novel method to compute species phylogenies based on sequence data including orthologs, paralogs or even xenologs. 1 1 Introduction Sequence-based phylogenetic approaches heavily rely on initial data sets to be composed of 1:1 orthologous sequences only. To this end alignments of protein or DNA sequences are employed whose evolutionary history is believed to be congruent to that of the respective species, a property that can be ensured most easily in the absence of gene duplications or horizontal gene transfer. Phylogenetic studies thus judiciously select families of genes that rarely exhibit duplications (such as rRNAs, most ribosomal proteins, and many of the housekeeping enzymes). In the presence of gene duplications, however, it becomes necessary to distinguish between the evolutionary history of genes (gene trees) and the evolutionary history of the species (species trees) in which these genes reside. Recent advances in mathematical phylogenetics, based on the theory of symbolic ultrametrics [7], have indicated that gene duplications can also convey meaningful phylogenetic information provided orthologs and paralogs can be distinguished with a degree of certainty [24, 27, 28]. Here, we examine a novel approach and explain the conceptional steps for the inference of species trees based on the knowledge of orthologs, paralogs or even xenologs [24, 27, 28]. 2 Preliminaries We give here a brief summary of the main definitions and concepts that are needed. Graphs, Gene Trees and Species Trees An (undirected) graph G is a pair (V, E) with non-empty vertex set V and edge set E containing two-element subsets of V . A class of graphs that will play an important role in this contribution are cographs. A graph G = (V, E) is a cograph iff G does not contain an induced path on four vertices, see [12, 13] for more details. A tree T = (V, E) is a connected, cycle-free graph. We distinguish two types of vertices in a tree: the leaves which are contained in only one edge and the inner vertices which are contained in at least two edges. In order to avoid uninteresting trivial cases, we will usually assume that T has at least three leaves. A rooted tree is a tree in which one special (inner) vertex is selected to be the root. The least common ancestor lcaT (x, y) of two vertices x and y in a rooted tree T is the first (unique) vertex that lies on the path from x to the root and y to the root. We say that a tree T contains the triple xy|z if x, y, and z are leaves of T and the path from x to y does not intersect the path from z to the root of T . A set of triples R is consistent if there is a rooted tree that contains all triples in R. An event-labeled tree, usually denoted by the pair (T, t), is a rooted tree T together with a map t : V → M that assigns to each inner vertex an event m ∈ M . For two leaves x and y of an event-labeled tree (T, t) its least common ancestor lcaT (x, y) is therefore marked with an event t(lcaT (x, y)) = m, which we denote for simplicity by lcaT (x, y) = t m. ∧ 2 In what follows, the set S will always denote a set of species and the set G a set of genes. We write x ∈ X if a gene x ∈ G resides in the species X ∈ S. A species tree (for S) is a rooted tree T with leaf-set S. A gene tree (for G) is an event-labeled tree (T, t) that has as leaf-set G. We refer the reader to [43] for an overview and important results on phylogenetics. Binary Relations and its Graph- and Tree-Representations A (binary) relation R over (an underlying set) G is a subset of G × G. We will write ⌊G × G⌋irr := (G × G) \ {(x, x) | x ∈ G} to denote the irreflexive part of G × G. Each relation R has a natural representation as a graph GR = (G, ER ) with vertex set G and edges connecting two vertices whenever they are in relation R. In what follows, we will always deal with irreflexive symmetric relations, which we call for simplicity just relations. Therefore, the corresponding graphs GR can be considered as undirected graphs without loops, that is, {x} 6∈ ER and, additionally, {x, y} ∈ ER iff (x, y) ∈ R (and thus, (y, x) ∈ R). While Graph-Representations GR of R are straightforward and defined for all binary relations, tree-representations of R are a bit more difficile to derive and, even more annoying, not every binary relation does have a tree-representation. For each tree representing a relation R over G the leaf-set L(T ) is G and a specific event-label is chosen so that the least common ancestor of two distinct elements x, y ∈ G is labeled in a way that uniquely determines whether (x, y) ∈ R or not. That is, an eventlabeled tree (T, t) with events “0” and “1” on its inner vertices represents a (symmetric irreflexive) binary relation R if for all (x, y) ∈ ⌊G × G⌋irr it holds that lcaT (x, y) = t 1 if and only if (x, y) ∈ R. The latter definitions can easily be extended to arbitrary disjoint (irreflexive symmetric) relations R1 , . . . , Rk over G: An edge-colored graph GR1 ,...,Rk = (G, E := ∪ki=1 ERi ) represents the relations R1 , . . . , Rk if it holds that (x, y) ∈ Ri if and only if {x, y} ∈ E and the edge {x, y} is colored with “i”. Analogously, an event-labeled tree (T, t) with events “0” and “1, . . . , k” on its inner vertices represents the relations R1 , . . . , Rk if for all (x, y) ∈ ⌊G × G⌋irr it holds that lcaT (x, y) = t i if and only if (x, y) ∈ Ri , 1 ≤ i ≤ k. The latter implies that for all pairs (x, y) that are in none of the relations Ri we have lcaT (x, y) = t 0. In practice, the disjoint relations correspond to the evolutionary relationship between genes contained in G, as e.g. the disjoint relations Ro and Rp that comprise the pairs of orthologous and paralogous genes, respectively. ∧ ∧ ∧ Paralogy, Orthology, and Xenology The current flood of genome sequencing data poses new challenges for comparative genomics and phylogenetics. An important topic in this context is the reconstruction of large families of homologous proteins, RNAs, and other genetic elements. The distinction between orthologs, paralogs, and xenologs is a key step in any research program of this type. The distinction between orthologous 3 and paralogous gene pairs dates back to the 1970s: two genes whose least common ancestor in the gene tree corresponds to a duplication are paralogs; if the least common ancestor was a speciation event and the genes are from different species, they are orthologs [17]. The importance of this distinction is two-fold: On the one hand, it is informative in genome annotation and, on the other hand, the orthology (or paralogy) relation conveys information about the events corresponding to internal nodes of the gene tree [24] and about the underlying species tree [27, 28]. We are aware of the controversy about the distinction between orthologous and paralogous genes and their consequence in the context of gene function, however, we adopt here the point of view that homology, and therefore also orthology and paralogy, refer only to the evolutionary history of a gene family and not to its function [22, 20]. In contrast to orthology and paralogy, the definition of xenology is less well established and by no means consistent in the biological literature. Xenology is defined in terms of horizontal gene transfer (HGT), that refers to the transfer of genes between organisms in a manner other than traditional reproduction and across species. The most commonly used definition stipulates that two genes are xenologs if their history since their common ancestor involves horizontal gene transfer of at least one of them [18, 32]. In this setting, both orthologs and paralogs may at the same time be xenologs [32]. Importantly, the mathematical framework established for evolutionary “event”-relations, as the orthology relation [7, 24], naturally accommodates more than two types of events associated with the internal nodes of the gene tree. It is appealing, therefore, to think of a HGT event as different from both speciation and duplication, in line with [23] where the term “xenologous” was originally introduced. In this contribution, we therefore will consider a slight modification of the terms orthologs, paralogs and xenologs, so-called lca-orthologs, lcaparalogs and lca-xenologs. To this end, note that for a set of genes G, the evolutionary relationship between two homologous genes contained in G is entirely explained by the true evolutionary gene-history of these genes. More precisely, if T is a (known) tree reflecting the true genehistory together with the events that happened, that is, the labeling t that tags the inner vertices of T as a speciation, duplication or HGT event, respectively, then we can determine the three disjoint relations Ro , Rp and Rx comprising the pairs of so-called lca-orthologous, lca-paralogous and lca-xenologous genes, respectively, as follows: Two genes x, y ∈ G are • lca-orthologous, if lcaT (x, y) = t speciation; ∧ • lca-paralogous, if lcaT (x, y) = t duplication and ∧ • lca-xenologous, if lcaT (x, y) = t HGT. ∧ The latter also implies the edge-colored graph representation GRo ,Rp ,Rx , see Figure 1 for an illustrative example. In the absence of horizontal gene transfer, the relations lca-orthologs and lca-paralogs are equivalent to orthologs and paralogs as defined by Fitch [18]. We are aware of the fact that this definition of lca-“events” leads to a loss of information of the direction of the HGT event, i.e., the information 4 Ro = {dv | v ∈ {ab1 , ac2 , b1 c2 , b2 c2 } G \ {d}} ∪ Rx = {b3 c1 } Rp = ⌊G × G⌋irr \ (Ro ∪ Rx ) xy ∈ R⋆ means that (x, y)(y, x) ∈ R⋆ , with ⋆ ∈ {o, p, x} b2 b1 b3 a d c1 c3 c2 Figure 1: Example of an evolutionary scenario showing the “true” evolution of a gene family evolving along the species tree (shown as blue tube-like tree). The corresponding true gene tree T appears embedded in the species tree S. The speciation vertices in the gene tree (red circuits) appear on the vertices of the species tree (blue ovals), while the duplication vertices (blue squares) and the HGT-vertices (green triangles) are located on the edges of the species tree. Gene losses are represented with “x”. The true gene-tree T uniquely determines the relationships between the genes by means of the event at lcaT (x, y) of distinct genes x, y ∈ G. The pairs of lca-orthologous, -paralogous and -xenologous genes are comprised in the relations Ro , Rp and Rx , respectively. The graphrepresentation GRo ,Rp ,Rx is shown in the lower left part. Non-drawn edges indicate the paralogous genes. This graph clearly suggests that the orthologyrelation Ro is not a complete subgraph, and thus, does not cluster or partition the input gene set G. However, in all cases the subgraphs GRo , GRp and GRx are so-called cographs, cf. Thm. 1. 5 of donor and acceptor. However, for the proposed method and to understand the idea of representing estimates of evolutionary relationships in an event-labeled tree this information is not necessarily needed. Nevertheless, generalizations to tree-representations of non-symmetric relations or a mathematical framework for xenologs w.r.t. the notion of Fitch might improve the proposed methods. Remark 1. If there is no risk of confusion and if not stated differently, we call lca-orthologs, lca-paralogs, and lca-xenologs simply orthologs, paralogs and xenologs, respectively. Clearly, evolutionary history and the events of the past cannot be observed directly and hence, must be inferred, using algorithmic and statistical methods, from the genomic data available today. Therefore, we can only deal with estimates of the relations Ro , Rp and Rx . In this contribution, we use those estimates to reconstruct (a hypothesis of) the evolutionary history of the genes and, eventually, the history of the species the genes reside in. We wish to emphasize that the three relations Ro , Rp and Rx (will) serve as illustrative examples and the cases Rp = ∅ or Rx = ∅ are allowed. In practice, it is possible to have more than these three relations. By way of example, the relation containing the pairs of paralogous genes might be more refined, since gene duplications have several different mechanistic causes that are also empirically distinguishable in real data sets. Thus, instead of heaving a single relation Rp that comprises all paralogs, we could have different types of paralogy relations that distinguish between events such as local segmental duplications, duplications by retrotransposition, or whole-genome duplications [50]. 3 From Sequence Data to Species Trees In this section, we provide the main steps in order to infer event-labeled gene trees and species trees from respective estimated event-relations. An implementation of these steps by means of integer linear programming is provided in the software tool ParaPhylo [27]. The starting point of this method is an estimate of the (true) orthology relation Ro . From this estimate the necessary information of the eventlabeled gene trees and the respective species trees will be derived. 3.1 Orthology Detection The inference of the orthology relation Ro and lies at the heart of many reconstruction methods. Orthology inference methods can be classified based on the methodology they use to infer orthology into tree-based and graph-based methods, for an overview see e.g. [2, 14, 21, 34, 46]. Tree-based orthology inference methods rely on the reconciliation of a constructed gene tree (without event-labeling) from an alignment of homologous sequences and a given species tree, see e.g. [4, 19, 29, 47, 49]. Although tree-based approaches are often considered as very accurate given a species tree, it suffers from high computational costs and is hence 6 limited in practice to a moderate number of species and genes. A further limitation of those tree-reconciliation methods is that for many scenarios the species tree is not known with confidence and, in addition, all practical issues that complicate phylogenetic inference (e.g. variability of duplication rates, mistaken homology, or HGT) limit the accuracy of both the gene and the species trees. Intriguingly, with graph-based orthology inference methods it is possible in practice to detect the pairs of orthologous genes with acceptable accuracy without constructing either gene or species trees. Many tools of this type have become available over the last decade. To name only a few, COG [45], OMA [42, 3], eggNOG [31], OrthoMCL [37, 11], InParanoid [40], Roundup 2.0 [16], EGM2 [39] or ProteinOrtho [35] and its extension PoFF [36]. Graph-based methods detect orthologous genes for two (pairwise) or more (multiple) species. These methods consist of a graph construction phase and, in some cases, a clustering phase [46]. In the graph construction phase, a graph is inferred where vertices represent genes, and (weighted) edges the (confidence of) orthology relationships. The latter rely on pairwise sequence similarities (e.g., basic local alignment search tool (BLAST) or Smith-Waterman) calculated between all sequences involved and an operational definition of orthology, for example, reciprocal best hit (RBH), bi-directional best hit (BBH), symmetrical best hit (SymBeT) or reciprocal smallest distance (RSD). In the clustering phase, clusters or groups of orthologs are constructed, using e.g., single-linkage, complete-linkage, spectral clustering or Markov Cluster algorithm. However, orthology is a symmetric, but not a transitive relation, i.e., it does in general not represent a partition of the set of genes G. In particular, a set G′ of genes can be orthologous to another gene g ∈ G \ G′ but the genes within G′ are not necessarily orthologous to each other. In this case, the genes in G′ are called co-orthologs to gene g [33]. It is important to mention that, therefore, the problem of orthology detection is fundamentally different from clustering or partitioning of the input gene set. In addition to OMA and ProteinOrtho only Synergy, EGM2, and InParanoid attempt to resolve the orthology relation at the level of gene pairs. The latter two tools can only be used for the analysis of two species at a time, while Synergy is not available as standalone tool and therefore cannot be applied to arbitrary user-defined data sets. In particular, the use of orthology inference tools is often limited to the species offered through the databases published by their authors. An exception is provided by ProteinOrtho [35] and its extension PoFF [36], methods that we will use in our approach. These standalone tools are specifically designed to handle large-scale user-defined data and can be applied to hundreds of species containing millions of proteins at ones. In particular, such computations can be performed on off-the-shelf hardware [35]. ProteinOrtho and PoFF compare similarities of given gene sequences (the bit score of the blast alignment) that together with an an E-value cutoff yield an edgeweighted directed graph. Based on reciprocal best hits, an undirected subgraph is extracted (graph construction phase) on which spectral clustering methods are applied (clustering phase), to determine significant groups of orthologous genes. To enhance the prediction accuracy, the relative order of genes (synteny) can be used as additional feature for the 7 discrimination between orthologs and paralogs. To summarize, graph-based methods have in common, that the output is a set of (pairs of) putative orthologous genes. In addition, orthology detection tools often report some weight or confidence value w(x, y) for x and y to be orthologs or not. This gives rise to a symmetric, irreflexive binary relation bo = {(x, y) | x, y ∈ G are estimated orthologs} R (1) = {(x, y) | lcaT (x, y) = t speciation (in the estimated gene tree T )}. (2) ∧ 3.2 Construction of Gene Trees Characterization of Evolutionary Event Relations Assume we have given a “true” orthology relation Ro over G, i.e., Ro comprises all pairs of “true” orthologs, that is, if the true evolutionary history (T, t) of the genes would be known, then (x, y) ∈ Ro if and only if lcaT (x, y) = t speciation. As we will show, given such a true relation without the knowledge of the gene tree (T, t), it is possible to reconstruct the “observable discriminating part” of (T, t) using the information contained in Ro , resp., Rp only, at least in the absence of xenologous genes [24, 25]. In the presence of HGT-events, but given the “true” relations Ro and Rp it is even possible to reconstruct (T, t) using the information contained in Ro and Rp only [24, 25]. Note, for the set of pairs of (lca)xenologs Rx we have ∧ Rx = ⌊G × G⌋irr \ (Ro ∪ Rp ). Clearly, since we do not know the true evolutionary history with conbo , R bp , R bx of these true relations fidence, we always deal with estimates R Ro , Rp , Rx . In order to understand under which conditions it is possible bo , R bp , R bx , to infer a gene tree (T, t) that represents the disjoint estimates R we characterize in the following the structure of their graph-representation bo ∪ R bp ∪ R bx = ⌊G × G⌋irr , then G b b b is a comGRbo ,Rb p ,Rbx . Note, if R Ro , Rp , Rx plete edge-colored graph, i.e., for all distinct x, y ∈ G there is an edge {x, y} ∈ E s.t. {x, y} is colored with with “⋆” if and only if (x, y) ∈ R⋆ , ⋆ ∈ {o, p, x} The following theorem is based on results established by Böcker and Dress [7] and Hellmuth et al. [24]. Theorem 1 ([7, 24]). Let GR1 ,...,Rk be the graph-representation of the relations R1 , . . . , Rk over some set G. There is an event-labeled gene tree representing R1 , . . . , Rk if and only if (i) the graph GRi = (G, ERi ) is a cograph for all i ∈ {1, . . . , k} and (ii) for all three distinct genes x, y, z ∈ G the three edges {x, y}, {x, z} and {y, z} in GR1 ,...,Rk have at most two distinct colors. bp = ⌊G × G⌋irr \ R bo , Clearly, in the absence of xenologs and thus, if R we can ignore condition (ii), since at most two colors occur in GRbo ,Rbp . In the latter case, GRbo , resp., GRbp alone provide all information of the underlying gene tree. 8 bo or R bp , R bx and Theorem 1 implies that whenever we have estimates R we want to find a tree (T, t) that represents these relations we must ensure that neither GRbo , GRbp nor GRbx contains an induced path on four vertices and that there is no triangle (a cycle on three vertices) in GRbo ,Rb p ,Rbx where each edge is colored differently. However, due to noise in the data or mispredicted events of pairs of genes, the graph GRbo ,Rbp ,Rbx will usually violate condition (i) or (ii). A particular difficulty arises from the fact, bo only, and do not know how to that we usually deal with the estimate R distinguish between the paralogs and xenologs. bo , R bp , R bx to the One possibility to correct the initial estimates R ∗ ∗ ∗ “closest” relations Ro , Rp , Rx so that there is a tree representation of Ro∗ , Rp∗ , Rx∗ , therefore, could be the change of a minimum number of edgecolors in GRbo ,Rbp ,Rb x so that GR∗o ,R∗p ,R∗x fulfills Condition (i) and (ii). This problem was recently shown to be NP-complete [25, 26, 38]. Inference of Local Substructures of the Gene Tree Assume we have given (estimated or true) relations Ro , Rp , Rx so that the graphrepresentation GRo ,Rp ,Rx fulfills Condition (i) and (ii) of Theorem 1. We show now briefly, how to construct the tree (T, t) that represents Ro , Rp , Rx . Here we utilize the information of triples that are extracted from the graph GRo ,Rp ,Rx and that must be contained in any gene-tree (T, t) representing Ro , Rp , Rx . More precisely, given the relations R1 , . . . , Rk we define the set of triples TR1 ,...,Rk as follows: For all three distinct genes x, y, z ∈ G we add the triple xy|z to TR1 ,...,Rk if and only if the colors of the edge {y, z} and {x, z} are identical but distinct from the color of the edge {x, y} in GR1 ,...,Rk . In other words, for the given evolutionary relations Ro , Rp , Rx the triple xy|z is added to TRo ,Rp ,Rx iff the two genes x and z, as well as y and z are in the same evolutionary relationship, but different from the evolutionary relation between x and y. Theorem 2 ([7, 24]). Let GR1 ,...,Rk be the graph-representation of the relations R1 , . . . , Rk . The graph GR1 ,...,Rk fulfills conditions (i) and (ii) of Theorem 1 (and thus, there is a tree representation of R1 , . . . , Rk ) if and only if there is a tree T that contains all the triples in TR1 ,...,Rk . The importance of the latter theorem lies in the fact, that the wellknown algorithm BUILD [1, 43] can be applied to TR1 ,...,Rk to determine whether the set of triples TR1 ,...,Rk is consistent, and, if so, constructs a tree representation in polynomial-time. To obtain a valid event-label for such a tree T we can simply set t(lcaT (x, y)) = ⋆ if the color of the edge {x, y} in GRo ,Rp ,Rx is “⋆”, ⋆ ∈ {o, p, x} [24]. It should be stressed that the evolutionary relations do not contain the full information on the event-labeled gene tree, see Fig. 2. Instead, the constructed gene trees (T, t) are homeomorphic images of the (possibly true) observable gene tree (T ′ , t′ ) by collapsing adjacent events of the same type [24]. That is, in the constructed tree (T, t) all inner vertices that are connected by an edge will have different event-labels, see Fig 2. Those trees are also known as discriminating representation, cf. [7]. However, these discriminating representations contain and provide 9 the necessary information to recover the input-relations, are unique (up to isomorphism), and do not pretend a higher resolution than actually supported by the data. 3.3 Construction of Species Trees While the latter results have been established for (lca)-orthologs, -paralogs and -xenologs, we restrict our attention in this subsection to orthologous and paralogous genes only and assume that there are no HGT-events in the gene trees. We shall see later, that in practical computation the existence of xenologous genes does not have a large impact on the reconstructed species history, although the theoretical results are established for gene histories without xenologous genes. In order to derive for a gene-tree (T, t) (that contains only speciation and duplication events) a species tree S with which (T, t) can be reconciled with or simply spoken “embedded” into, we need to answer the question under which conditions there exists such a species tree for a given gene tree. A tree S = (W, F ) with leaf set S is a species tree for a gene tree T = (V, E) with leaf set G if there is a reconciliation map µ : V → W ∪ F that map the vertices in V to vertices or edges in W ∪ F . A reconciliation map µ maps the genes x ∈ G in T to the respective species X ∈ S in S the gene x resides in so that specific constraints are fulfilled. In particular, the inner vertices of T with label “speciation” are mapped to the inner vertices of S, while the duplication vertices of T are mapped to the edges in W so that the relative “evolutionary order” of the vertices in T is preserved in S. We refer to [28] for the full definition of reconciliation maps. In Fig. 1, the reconciliation map µ is implicitly given by drawing the species tree superimposed on the gene tree. Hence, for a given gene tree (T, t) we wish to efficiently decide whether there is a species tree in which (T, t) can be embedded into, and if so, construct such a species tree together with the respective reconciliation map. We will approach the problem of deriving a species tree from an event-labeled gene tree by reducing the reconciliation map from gene tree to species tree to rooted triples of genes residing in three distinct species. To this end we define a species triple set S derived from (T, t) that provides all information needed to efficiently decide whether there is a species tree S for (T, t) or not. Let Ro (T ) be the set of all triples ab|c that are contained in T s.t. a, b, c ∈ G reside in pairwise different species and lcaT (a, b, c) = t speciation, then set ∧ S := {AB|C : ∃ab|c ∈ Ro (T ) with a ∈ A, b ∈ B, c ∈ C}. It should be noted that by results established in [7, 24] it is possible to derive the triple set S directly from the orthology relation Ro without constructing a gene tree, cf. [24]: AB|C ∈ S if and only if (I) A, B and C are pairwise different species and there are genes a ∈ A, b ∈ B, c ∈ C so that either 10 a b1 b2 c2 b3 c1 c3 d A B C D Figure 2: Left, the homeomorphic image (T, t) of the observable gene tree (T ′ , t′ ) in Fig. 1 is shown. The true gene tree in Fig. 1 represents all extant as well extinct genes, all duplication, HGT and speciation events. Not all of these events are observable from extant genes data, however. In particular, extinct genes cannot be observed. Thus, the observable gene tree (T ′ , t′ ) is obtained from the original gene tree in Fig. 1 by removing all vertices marked with “x” together with their incident edges and, thereafter, removing all inner vertices that are contained in only two edges. The homeomorphic image (T, t) is obtained from (T ′ , t′ ) by contraction of the edge that connect the two consecutive duplication events. The species triple set S is {AB|C2 , AB|D3 , AC|D3 , BC|D9 }, where indices indicate the number of gene triples in Ro (T ) that support the respective species triple. In this example, the (unique, and thus minimally resolved) species tree S that contains all triples in S is shown in the right part. The species tree S is identical to the true species tree shown in Fig. 1 11 (IIa) (a, c), (b, c) ∈ Ro and (a, b) 6∈ Ro or (IIb) (a, c), (b, c), (a, b) ∈ Ro and there is a gene d ∈ G with (c, d) ∈ Ro and (a, d), (b, d) ∈ / Ro . Thus, in order to infer species triples a sufficient number of duplication events must have happened. The following important result was given in [28]. Theorem 3. Let (T, t) be a given gene tree that contains only speciation and duplication events. Then there is a species tree S for (T, t) if and only if there is any tree containing all triples in S. In the positive case, the species tree S and the reconciliation between (T, t) and S can be found in polynomial time. Interestingly, the latter theorem implies that the gene tree (T, t) can be embedded into any tree that contains the triples in S. Hence, one usually wants to find a species tree with a least number of inner vertices, as those trees constitute one of the best estimates of the phylogeny without pretending a higher resolution than actually supported by the data. Such trees are also called minimally resolved tree and computing such trees is an NP-hard problem [30]. Despite the variance reduction due to cograph editing, noise in the data, as well as the occasional introduction of contradictory triples as a consequence of horizontal gene transfer is unavoidable. The species triple set S collected from the individual gene families thus will not always be consistent. The problem of determining a maximum consistent subset of an inconsistent set of triples is NP-hard and also APX-hard, see [9, 48]. Polynomial-time approximation algorithms for this problem and further theoretical results are reviewed in [10]. The results in this subsection have been established for the reconciliation between event-labeled gene trees without HGT-events and inferred species. Although there are reconciliation maps defined for gene trees that contain xenologs and respective species trees [5, 6], a mathematical characterization of the species triples S and the existence of species trees for those gene trees, which might help also to understand the transfer events itself, however, is still an open problem. 3.4 Summary of the Theory The latter results show that it is not necessary to restrict the inference of species trees to 1:1 orthologs. Importantly, orthology information alone is sufficient to reconstruct the species tree provided that (i) the orthology is known without error and unperturbed by horizontal gene transfer and (ii) the input data contains a sufficient number of duplication events. Although species trees can be inferred in polynomial time for noise-free data, in a realistic setting, three NP-hard optimization problems need to be solved. We summarize the important working steps to infer the respective gene and species trees from genetic material. (W1) bo and set R bp = ⌊G × G⌋irr \ R bo . Compute the estimate R 12 (W2) Edit the graph GRbo to the closest cograph with a minimum number of edge edits to obtain the graph GRo . Note, Rp = ⌊G×G⌋irr \Ro . (W3) Compute the tree representation (T, t) w.r.t. Ro , Rp . (W4) Extract the species triple set Sb from Ro (T ). (W5) b Extract a maximal consistent triple set S from S. (W6) Compute a minimally resolved species tree S that contains all triples in S, and, if desired, the reconciliation map µ between (T, t) and S (cf. Thm. 3). In the presence of horizontal transfer, in Step (W1) the xenologous genes x, y are either predicted as orthologs or paralogs. Furthermore, in Step (W2) it suffices to edit the graph GRbo only, since afterwards the graph representation GRp with Rp = ⌊G × G⌋irr \ Ro and, thus GRo ,Rp fulfills the conditions of Thm. 1 [12, 24]. In particular, the graphs GRo ,Rp and GRp have then been obtained from GRbo ,Rbp , resp., GRbp with a minimum number of edge edits. The latter is due to the fact that the complement GRb o is the graph GRbp [12]. To extract the species triple set Sb in Step (W4), it suffices to choose the respective species triples using Condition (I) and (IIa)/(IIb), without constructing the gene trees and thus, Step (W3) can be ignored if the gene history is not of further interest. 4 Evaluation In [27] it was already shown that for real-life data sets the paralogy-based method produces phylogenetic trees for moderately sized species sets. The resulting species trees are comparable to those presented in the literature that are constructed by “state-of-the-art” phylogenetic reconciliation approaches as RAxML [44] or MrBayes [41]. To this end, genomic sequences of eleven Aquificales and 19 Enterobacteriales species were analyzed. Based on the NCBI gene annotations of those species, an orthology prediction was performed using ProteinOrtho. From that prediction, phylogenetic trees were constructed using the aforementioned orthology-paralogy-based approach (working steps (W2)-(W6)) implemented as integer linear program in ParaPhylo [27]. The advantage of this approach is the computation of exact solutions, however, the runtime scales exponentially with the number of input genes per gene family and the number of species. However, as there is no gold standard for phylogenetic tree reconstruction, three simulation studies are carried out to evaluate the robustness of the method. Using the Artificial Life Framework (ALF) [15], the evolution of generated gene sequences was simulated along a given branch length-annotated species tree, explicitly taking into account gene duplication, gene loss, and horizontal transfer events. For realistic species trees, the γ-proteobacteria tree from the OMA project [3] was randomly pruned to a size of 10 species while conserving the branch lengths. For additional details on the simulation see [27]. The reconstructed trees are then compared with the initial species trees, using the software TreeCmp [8]. In the provided box-blots (Fig. 3), tree distances are computed according to 13 the triple metric and normalized by the average distance between random Yule trees, see [27] for further evaluations. 0.8 0.4 0.0 tree distance gene families 100 paralogous noise 0.4 0.8 orthologous noise 0.0 tree distance homologous noise 300 500 # gene families 15 20 % noise ProteinOrtho 25 5 10 15 20 % noise lca xenology 25 5 10 15 20 % noise Fitch xenology 25 0.4 0.8 10 0.0 tree distance 5 0/ 0 5.5/ 15.7 10.6/ 15.3/ 0/ 5.5/ 10.6/ 15.3/ 0/ 28.7 39.4 0 15.7 28.7 39.4 0 % HGT events / % xenologous pairs of genes 5.5/ 15.7 10.6/ 15.3/ 28.7 39.4 Figure 3: Accuracy of reconstructed species trees (10 species) in simulated data sets; (Top) Dependence on the number of gene families; (Middle) Dependence of different noise models; (Down) Dependence on noise by HGT. The three simulation studies are intended to answer three individual questions. 1. How much data is needed to provide enough information to reconstruct accurate species trees? (cf. Fig. 3 (top)) 2. How does the method perform with noisy data? (cf. Fig. 3 (middle)) 3. What is the impact of horizontal gene transfer on the accuracy of the method? (cf. Fig. 3 (down)) To construct accurate species trees, the presented method requires a sufficient amount of duplicated genes. Assuming a certain gene duplication rate, the amount of duplicated genes correlates directly with the number of genes per species, respectively the number of gene families. The first simulation study (Fig. 3 (top)) is therefore performed with several numbers of gene families, varying from 100 to 500. The simulation 14 with ALF was performed without horizontal gene transfer and the phylogenetic trees are computed based on the unaltered orthology/paralogy relation obtained from the simulation, that is, the orthologs and paralogs can directly be derived from the simulated gene trees. It turned out that with an duplication rate of 0.005, which corresponds to approximately 8% of paralogous pairs of genes, 500 gene families are sufficient to produce reliable phylogenetic trees. With less gene families, and hence less duplicated genes, the trees tend to be only poorly resolved. For the second study the simulated orthology/paralogy relation of 1000 gene families was perturbed by different types of noise. (i) insertion and deletion of edges in the orthology graph (homologous noise), (ii) insertion of edges (orthologous noise), and (iii) deletion of edges (paralogous noise), see Fig. 3 (middle). In the three models an edge is inserted or removed with probability p ∈ {0.05, 0.1, 0.15, 0.2, 0.25}. It can be observed that up to noise of approximately 10% the method produces trees which are almost identical to the initial trees. Especially, in the case of orthology overprediction (orthologous noise) the method is robust even if 25% of the input data was disturbed. Finally, in the third analysis, data sets are simulated with different rates of horizontal gene transfer, see Fig. 3 (down). The number of HGT events in the gene trees are varied up to 15.3%, which corresponds to 39.4% of all pairs of genes (x, y) having at least one HGT event on the path from x to y in the generated gene tree, i.e., x and y are xenologous with respect to the definition of Fitch [18]. Firstly, the simulated gene sequences are analyzed using ProteinOrtho and the tree reconstruction is then performed based on the resulting orthology/paralogy prediction (Fig. 3 (down/left)). Secondly, we used both definitions of xenology, i.e., lca-xenologous and the notion of Fitch. Note, so far the reconstruction of species trees with ParaPhylo requires that pairs of genes are either orthologous or paralogous. Hence, we used the information of the lcaorthologs, -paralogs and -xenologs derived from the simulated gene trees. Fig. 3 (down/center) shows the accuracy of reconstructed species trees under the assumption that all lca-xenologs are “mispredicted” as lcaorthologs, in which case all paralogous genes are identified correctly. Fig. 3 (down/right) shows the accuracy of reconstructed species trees under the assumption that all xenologs w.r.t. the notion of Fitch are interpreted as lca-orthologs. The latter amounts to the “misprediction” of lca-xenologs and lca-paralogs, as lca-orthologs. However, all remaining lca-paralogs, are still correctly identified. For the orthology/paralogy prediction based on ProteinOrtho, it turned out that the resulting trees have a distance of approximately 0.3 to 0.4 to the initial species tree. Thereby, a distance of 1 refers not to a maximal distance, but to the average distance between random trees. However, the accuracy of the constructed trees appears to be independent from the amount of horizontal gene transfer. Hence, ProteinOrtho is not able to either identify the gene families correctly, or mispredicts orthologs and paralogs (due to, e.g., gene loss). In case that all paralogous genes are identified correctly, ParaPhylo produces more accurate trees. We obtain even more accurate species trees, when predicting all pairs of Fitch-xenologous genes as lca-orthologs, even with a large amount of HGT events. 15 5 Concluding Remarks The restriction to 1:1 orthologs for the reconstruction of the evolutionary history of species is not necessary. Even more, it has been shown that the knowledge of only a few correct identified paralogs allows to reconstruct accurate species trees, even in the presence of horizontal gene transfer. The information of paralogs is strictly complementary to the sources of information used in phylogenomics studies, which are always based on alignments of orthologous sequences. Hence, paralogs contain meaningful and valuable information about the gene and the species trees. Future research might therefore focus on improvements of orthology and paralogy infence tools, and mathematical frameworks for tree-representations of non-symmetric relations (since HGT is naturally a directed event), as well as a characterization of the reconciliation between gene and species trees in the presence of HGT. References [1] A. V. Aho, Y. Sagiv, T. G. Szymanski, and J. D. Ullman. Inferring a tree from lowest common ancestors with an application to the optimization of relational expressions. SIAM J. Comput., 10:405–421, 1981. [2] A M Altenhoff and C. Dessimoz. Phylogenetic and functional assessment of orthologs inference projects and methods. PLoS Comput Biol., 5:e1000262, 2009. [3] Adrian M. Altenhoff, Adrian Schneider, Gaston H. Gonnet, and Christophe Dessimoz. OMA 2011: orthology inference among 1000 complete genomes. Nucleic Acids Research, 39(suppl 1):D289–D294, 2011. [4] Lars Arvestad, Ann-Charlotte Berglund, Jens Lagergren, and Bengt Sennblad. Bayesian gene/species tree reconciliation and orthology analysis using mcmc. Bioinformatics, 19(suppl 1):i7–i15, 2003. [5] Mukul S. Bansal, Eric J. Alm, and Manolis Kellis. Efficient algorithms for the reconciliation problem with gene duplication, horizontal transfer and loss. Bioinformatics, 28(12):i283–i291, 2012. [6] Mukul S. Bansal, Eric J. Alm, and Manolis Kellis. Reconciliation revisited: Handling multiple optima when reconciling with duplication, transfer, and loss. Journal of Computational Biology, 20(10):738–754, 2013. [7] Sebastian Böcker and Andreas W. M. Dress. Recovering symbolically dated, rooted trees from symbolic ultrametrics. Adv. Math., 138:105– 125, 1998. [8] D. Bogdanowicz, K. Giaro, and B. Wróbel. Treecmp: Comparison of trees in polynomial time. Evolutionary Bioinformatics Online, 8:475, 2012. [9] J. Byrka, P. Gawrychowski, K. T. Huber, and S. Kelk. Worst-case optimal approximation algorithms for maximizing triplet consistency within phylogenetic networks. J. Discr. Alg., 8:65–75, 2010. 16 [10] J. Byrka, S. Guillemot, and J. Jansson. New results on optimizing rooted triplets consistency. Discr. Appl. Math., 158:1136–1147, 2010. [11] Feng Chen, Aaron J. Mackey, Christian J. Stoeckert, and David S. Roos. Orthomcl-db: querying a comprehensive multi-species collection of ortholog groups. Nucleic Acids Research, 34(suppl 1):D363– D368, 2006. [12] D. G. Corneil, H. Lerchs, and L. Steward Burlingham. Complement reducible graphs. Discr. Appl. Math., 3:163–174, 1981. [13] D. G. Corneil, Y. Perl, and L. K. Stewart. A linear recognition algorithm for cographs. SIAM J. Computing, 14:926–934, 1985. [14] Daniel A. Dalquen, Adrian M. Altenhoff, Gaston H. Gonnet, and Christophe Dessimoz. The impact of gene duplication, insertion, deletion, lateral gene transfer and sequencing error on orthology inference: A simulation study. PLoS ONE, 8(2):e56925, 02 2013. [15] Daniel A. Dalquen, Maria Anisimova, Gaston H. Gonnet, and Christophe Dessimoz. ALF–a simulation framework for genome evolution. Mol. Biol. Evol., 29(4):1115–1123, 2012. [16] Todd F. DeLuca, Jike Cui, Jae-Yoon Jung, Kristian Che St. Gabriel, and Dennis P. Wall. Roundup 2.0: Enabling comparative genomics for over 1800 genomes. Bioinformatics, 28(5):715–716, 2012. [17] W M Fitch. Distinguishing homologous from analogous proteins. Syst Zool, 19:99–113, 1970. [18] Walter M. Fitch. Homology: a personal view on some of the problems. Trends Genet., 16:227–231, 2000. [19] Shi G., M.-C. Peng, and T. Jiang. Multimsoar 2.0: An accurate tool to identify ortholog groups among multiple genomes. PLoS ONE, 6(6):e20892, 2011. [20] T. Gabaldón and EV. Koonin. Functional and evolutionary implications of gene orthology. Nat. Rev. Genet., 14(5):360–366, 2013. [21] Toni Gabaldón. Large-scale assignment of orthology: back to phylogenetics? Genome Biology, 9(10):235, 2008. [22] John Gerlt and Patricia Babbitt. Can sequence determine function? Genome Biology, 1(5):reviews0005.1–reviews0005.10, 2000. [23] G S Gray and W M Fitch. Evolution of antibiotic resistance genes: the DNA sequence of a kanamycin resistance gene from Staphylococcus aureus. Mol Biol Evol, 1:57–66, 1983. [24] M. Hellmuth, M. Hernandez-Rosales, K. T. Huber, V. Moulton, P. F. Stadler, and N. Wieseke. Orthology relations, symbolic ultrametrics, and cographs. Journal of Mathematical Biology, 66(1-2):399–420, 2013. [25] M. Hellmuth and N. Wieseke. On symbolic ultrametrics, cotree representations, and cograph edge decompositions and partitions. In Dachuan Xu, Donglei Du, and Dingzhu Du, editors, Computing and Combinatorics, volume 9198 of Lecture Notes in Computer Science, pages 609–623. Springer International Publishing, 2015. 17 [26] M. Hellmuth and N. Wieseke. On tree representations of relations and graphs: Symbolic ultrametrics and cograph edge decompositions. CoRR, abs/1509.05069, 2015. submitted to J. Comb. Opt. (Springer). [27] M. Hellmuth, N. Wieseke, M. Lechner, H.-P. Lenhof, M. Middendorf, and P. F. Stadler. Phylogenomics with paralogs. Proceedings of the National Academy of Sciences, 112(7):2058–2063, 2015. [28] M. Hernandez-Rosales, M. Hellmuth, N. Wieseke, K. T. Huber, V. Moulton, and P. F. Stadler. From event-labeled gene trees to species trees. BMC Bioinformatics, 13(Suppl 19):S6, 2012. [29] T. J. Hubbard and et al. Ensembl 2007. Nucleic Acids Research, 35(suppl 1):D610–D617, 2007. [30] J. Jansson, R.S. Lemence, and A. Lingas. The complexity of inferring a minimally resolved phylogenetic supertree. SIAM J. Comput., 41:272–291, 2012. [31] Lars Juhl Jensen, Philippe Julien, Michael Kuhn, Christian von Mering, Jean Muller, Tobias Doerks, and Peer Bork. eggnog: automated construction and annotation of orthologous groups of genes. Nucleic Acids Research, 36(suppl 1):D250–D254, 2008. [32] Roy A Jensen. Orthologs and paralogs – we need to get it right. Genome Biol, 2:8, 2001. [33] Eugene V. Koonin. Orthologs, paralogs, and evolutionary genomics1. Annual Review of Genetics, 39(1):309–338, 2005. PMID: 16285863. [34] David M. Kristensen, Yuri I. Wolf, Arcady R. Mushegian, and Eugene V. Koonin. Computational methods for gene orthology inference. Briefings in Bioinformatics, 12(5):379–391, 2011. [35] Marcus Lechner, Sven Findeiß, Lydia Steiner, Manja Marz, Peter F. Stadler, and Sonja J. Prohaska. Proteinortho: detection of (co)orthologs in large-scale analysis. BMC Bioinformatics, 12:124, 2011. [36] Marcus Lechner, Maribel Hernandez-Rosales, D. Doerr, N. Wiesecke, A. Thevenin, J. Stoye, Roland K. Hartmann, Sonja J. Prohaska, and Peter F. Stadler. Orthology detection combining clustering and synteny for very large datasets. PLoS ONE, 9(8):e105015, 08 2014. [37] Li Li, Christian J Stoeckert, and David S Roos. OrthoMCL: identification of ortholog groups for eukaryotic genomes. Genome research, 13(9):2178–2189, 2003. [38] Yunlong Liu, Jianxin Wang, Jiong Guo, and Jianer Chen. Complexity and parameterized algorithms for cograph editing. Theoretical Computer Science, 461(0):45 – 54, 2012. [39] Khalid Mahmood, Geoffrey I. Webb, Jiangning Song, James C. Whisstock, and Arun S. Konagurthu. Efficient large-scale protein sequence comparison and gene matching to identify orthologs and co-orthologs. Nucleic Acids Research, 40(6):e44–e44, 2012. [40] Gabriel Östlund, Thomas Schmitt, Kristoffer Forslund, Tina Köstler, David N Messina, Sanjit Roopra, Oliver Frings, and Erik LL Sonnhammer. InParanoid 7: new algorithms and tools for eukaryotic orthology analysis. Nucleic acids research, 38(suppl 1):D196– D203, 2010. 18 [41] F. Ronquist, M. Teslenko, P. van der Mark, D. L. Ayres, A. Darling, S. Hohna, B. Larget, L. Liu, M. A. Suchard, and J. P. Huelsenbeck. MrBayes 3.2: efficient Bayesian phylogenetic inference and model choice across a large model space. Syst. Biol., 61(3):539–542, 2012. [42] Adrian Schneider, Christophe Dessimoz, and Gaston H. Gonnet. Oma browserexploring orthologous relations across 352 complete genomes. Bioinformatics, 23(16):2180–2182, 2007. [43] Charles Semple and Mike Steel. Phylogenetics, volume 24 of Oxford Lecture Series in Mathematics and its Applications. Oxford University Press, Oxford, UK, 2003. [44] Alexandros Stamatakis. RAxML version 8: A tool for phylogenetic analysis and post-analysis of large phylogenies. Bioinformatics, 2014. [45] Roman L. Tatusov, Michael Y. Galperin, Darren A. Natale, and Eugene V. Koonin. The cog database: a tool for genome-scale analysis of protein functions and evolution. Nucleic Acids Research, 28(1):33– 36, 2000. [46] Kalliopi Trachana, Tomas A. Larsson, Sean Powell, Wei-Hua Chen, Tobias Doerks, Jean Muller, and Peer Bork. Orthology prediction methods: A quality assessment using curated protein families. BioEssays, 33(10):769–780, 2011. [47] Rene van der Heijden, Berend Snel, Vera van Noort, and Martijn Huynen. Orthology prediction at scalable resolution by phylogenetic tree analysis. BMC Bioinformatics, 8(1):83, 2007. [48] L. van Iersel, S. Kelk, and M. Mnich. Uniqueness, intractability and exact algorithms: reflections on level-k phylogenetic networks. J. Bioinf. Comp. Biol., 7:597–623, 2009. [49] Ilan Wapinski, Avi Pfeffer, Nir Friedman, and Aviv Regev. Automatic genome-wide reconstruction of phylogenetic gene trees. Bioinformatics, 23(13):i549–i558, 2007. [50] Jianzhi Zhang. Evolution by gene duplication: an update. Trends Ecol. Evol., 18:292–298, 2003. 19
8cs.DS
3D Face Reconstruction from Light Field Images: A Model-free Approach Mingtao Feng 1,2 Syed Zulqarnain Gilani1 Yaonan Wang 2 Ajmal Mian1 1 School of Computer Science and Software Engineering, The University of Western Australia arXiv:1711.05953v3 [cs.CV] 12 Jan 2018 {zulqarnain.gilani,ajmal.mian}@uwa.edu.au 2 College of Electrical and Information Engineering, Hunan University {mintfeng,yaonan}@hnu.edu.cn Abstract Reconstructing 3D facial geometry from a single RGB image has recently instigated wide research interest. However, it is still an ill-posed problem and most methods rely on prior models hence undermining the accuracy of the recovered 3D faces. In this paper, we exploit the Epipolar Plane Images (EPI) obtained from light field cameras and learn CNN models that recover horizontal and vertical 3D facial curves from the respective horizontal and vertical EPIs. Our 3D face reconstruction network (FaceLFnet) comprises a densely connected architecture to learn accurate 3D facial curves from low resolution EPIs. To train the proposed FaceLFnets from scratch, we synthesize photorealistic light field images from 3D facial scans. The curve by curve 3D face estimation approach allows the networks to learn from only 14K images of 80 identities, which still comprises over 11 Million EPIs/curves. The estimated facial curves are merged into a single pointcloud to which a surface is fitted to get the final 3D face. Our method is model-free, requires only a few training samples to learn FaceLFnet and can reconstruct 3D faces with high accuracy from single light field images under varying poses, expressions and lighting conditions. Comparison on the BU-3DFE and BU-4DFE datasets show that our method reduces reconstruction errors by over 20% compared to recent state of the art. Figure 1. Proposed pipeline for 3D face reconstruction from a single light field image. Using synthetic light field face images, we train two FaceLFnets for regressing 3D facial curves over their respective horizontal and vertical EPIs. The estimated depth maps are combined, using camera parameters, into a single pointcloud to which a surface is fitted to get the final 3D face. cation [15] and face animation [9, 49]. Reconstructing 3D facial geometry from RGB images is, therefore, receiving a significant interest from the research community. However, using a single RGB image to recover the 3D face is an ill-posed problem [31] since the depth information is lost during the projection process. In fact, many different 3D shapes can result in similar 2D projections. The scale and bas-relief ambiguities [6] are common examples. Most existing methods have resorted to the use of prior models such as the Basal Face Model (BFM) [43] and the Annotated Face Model(AFM) [12] to generate synthetic data with ground truth to train CNN [11, 40] models and to recover the model parameters at test time. However, model- 1. Introduction Three dimensional face analysis has the potential to address the challenges that confound its two dimensional counterpart such as variations in illumination, pose and scale [4]. This modality has achieved state-of-the-art performances on applications such as face recognition [14, 36, 39], syndrome diagnosis [16, 17, 47, 55], gender classifi1 based approaches are inherently biased and constrained to the space of the training data of the prior models. A 4D light field image captures the RGB color intensities at each pixel as well as the direction of incoming light rays. High resolution plenoptic cameras [2, 3] are now commercially available. Plenoptic cameras use an array of microlenses to capture many sub-aperture images arranged in an equally spaced rectangular grid. Unlike most 3D scanners that use active light projection and are hence restricted to indoor use, plenoptic cameras are passive and can instantly acquire light field images outdoors as well, in a single photographic exposure. The sub-aperture light field images have been exploited to improve the performance of many applications such as saliency detection [32], hyperspectral light field imaging [57], material classification [53], image segmentation [62] and image restoration [50, 56] and in particular, depth estimation [26, 48, 34, 52, 46]. This paper focuses on reconstructing 3D faces from light field images under a wide range of pose, expression and illumination variations. Various methods have been proposed to solve the illposed problem of reconstructing 3D facial geometry from a single RGB image [31, 40, 11, 51, 44, 29]. These methods all use one or more common techniques. For instance, Shape from Shading (SfS) uses the shading variation to reconstruct 3D faces but the caveat is that the method is sensitive to lighting and RGB image texture and even under near ideal conditions, suffers from the bas-relief ambiguity [6]. 3D Morphable Models (3DMM) [11, 40] project the 3D faces in a low-dimensional subspace. However, the models are confined to the linear space of their training data and do not generalize well to all face shapes [13]. Landmark based methods use facial keypoints to guide the reconstruction process but rely heavily on accurate localization of the landmarks. We propose a model-free approach (see Fig. 1) to reconstruct 3D faces directly from light field images using Convolutional Neural Networks (CNN). Our technique does not rely on model fitting or landmark detection. Training a CNN requires massive amount of photo-realistic labeled data. However, there is no publicly available 4D light field face dataset with corresponding ground truth 3D face models. We address this problem and propose a method of generating the training data. We use the BU-3DFE [58] and BU-4DFE datasets [60] to generate light field images from their ground truth 3D models. Figure 2 shows some examples. We randomly vary the light intensity and pose to make our dataset more realistic. Our dataset comprises approximately 19K photo-realistic light field images with ground truth depth maps 1 . Furthermore, we show that our method requires fewer training samples (facial identities) as it cap1 We use depth map to represent disparity map as they are related by light field camera parameters [22]. italizes on reconstructing 3D facial curves rather than the complete face at once. We believe that our synthesized dataset of 4D light field images with corresponding 3D facial scans can be applied to many other facial analysis problems such as pose estimation, recognition and alignment. Equipped with a rich light field image dataset, we propose a densely connected CNN architecture (FaceLFnet) to learn 3D facial curves from Epipolar Plane Images (EPIs). We train two networks separately using horizontal and vertical EPIs to increase the accuracy of depth estimation. The densenet architecture is preferred as it can accurately learn the subtle slopes in low resolution EPIs2 . FaceLFnets are trained using our synthetic light field face images for which the ground truth depth data is available. Once the face curve estimates are obtained independently from the horizontal and vertical FaceLFnets, we merge them into a single pointcloud based on the camera parameters and then use a surface fitting method to recover the final 3D face. The core idea of our work is a model-free approach, where the solution is not restricted to any statistical face space. This is possible by exploiting the shape information present in the Epipolar Plane Images. Our contribution are: (1) A model-free approach for 3D face reconstruction from a single light field image. Our method does not require face alignment or landmark detection and is robust to facial expressions, pose and illumination variations. Being model-free, our method also estimates the peripheral regions of the face such as hair and neck. (2) A training technique that does not require massive number of facial identities. Exploiting the EPIs, we demonstrate that the proposed FaceLFnet can learn from only a few identities (80) and still outperform the state-of-the-art methods by a margin of 26%. (3) A data syntheses technique for generating a light field face image dataset which, to the best of our knowledge, is the first of its kind. This dataset will contribute to solving other face analysis problems as well. 2. Related Work 3D face reconstruction from a single image has attracted significant attention recently. Shape-from-shading (SfS) has been a popular approach for this task [61, 37, 18]. For example, WenYi et al. [61] proposed a symmetric SfS method to obtain illumination-normalized image and developed a face recognition system. Roy et al. [37] proposed an improved SfS method to enhance the depth map combining the RGB image and rough depth image to create more details. Yudeog et al. [18] estimated lighting variations with both global and local light models. SfS approach was then applied with the estimated lighting models for accurate shape reconstruction. Reconstruction using SfS re2 Higher slope of lines in EPI corresponds to lower depth values. quires priors of reflectance properties and lighting conditions and suffers from the bas-relief ambiguity [6]. A 3D Morphable Model (3DMM) was introduced by Blanz and Vetter [7] which represents a 3D face as a linear combination of orthogonal basis vectors obtained by PCA over 100 male and 100 female identities. James et al. [8] extended the concept and proposed a statistical model combined with a texture model for fitting the 3DMM on face images in the wild. 3DMM has also been used in [38, 5, 42, 30] for face reconstruction. The main limitation of such methods is that the 3DMM cannot model every possible face. Moreover, it is unable to extract facial details like wrinkles and folds because such details are not encoded in the linear subspace. Recently, various attempts were made to integrate 3DMMs with CNN for facial geometry reconstruction from a single image. Elad et al. [40] employed an iterative CNN trained with synthetic data to estimate 3DMM vectors. The predicted geometry was then refined by the real-time shape-from-shading method. Matan et al. [41] extended the work [40] and introduced an end-to-end CNN framework that recovers the coarse facial shape using a CoarseNet, followed by a FineNet to refine the facial details. The two net parts are connected by a novel layer that renders the depth image from 3D mesh. Pengfei Dou et al. [11] proposed an end-to-end 3D face reconstruction method from a single RGB image. They trained a fusion-CNN with multi-task learning loss to simplify 3D face reconstruction into neutral and expressive 3D facial parameters estimation. Jourabloo et al. [29] proposed a 3DMM fitting method for face alignment, which uses a cascaded CNN to regress camera matrix and 3DMM parameters. Tuan Tran et al. [51] used multi image 3DMM estimates as ground truth and then trained a CNN to regress 3DMM shape and texture parameters from an input image. Kemelmacher el at. [31] used the input image as a guide to build a single reference model to align with the face image and then refined the reference model using SfS method. Tal et al. [19] used a 3D neutral face as reference model to approximate the RGB image for face frontalization. Matan et al. [44] proposed a translation network that learns two maps (a depth image and a correspondence map), used for non-rigid registration with a template face, from a single RGB image. Fine-tuning is then performed for reconstructing facial details. In contrast to SfS and model fitting based face reconstruction methods, we learn 3D face curves from EPIs of the light field image. Our method does not require face alignment, dense correspondence or model fitting steps and is robust to facial pose, expressions and illumination. To the best of our knowledge none of the existing methods is model-free and uses a prior face model at some stage of the reconstruction process. On the other hand our method is completely model-free. Similarly, we are unaware of any Figure 2. Central view examples of our rendered light field images. The ground truth 3D scans are aligned with the central view. To make the dataset rich in variations, the generated light field images use random backgrounds and differ extensively in ethnicity, gender, age, pose and illumination. existing technique that uses light field images for 3D face reconstruction. However, literature points to some research in shape reconstruction from light field images using deep learning. Heber et al. [20] presented a method for reconstructing the shape from light field images that applies a CNN for pixel wise depth estimation from EPI patches. Although this method produces accurate scene depth, it uses a carefully designed dataset containing drastic slope changes in the EPIs. This method is unsuitable for non-rigid facial geometry reconstruction as faces are generally smooth and their EPIs contain only subtle slope variations. Heber et al. [21] proposed a U-shaped network architecture that automatically learns from EPIs to reconstruct their corresponding disparity images. However, training the network requires disparity maps of all the light field sub-views as labels, which is unrealistic for real datasets. Our approach differs in three ways. Firstly, we use one full EPI as input and its corresponding depth values as labels to overcome the problem of inaccurate depth estimation in the presence of subtle slope variations in the EPIs. Secondly, we train networks using horizontal EPIs and vertical EPIs separately to obtain a more accurate combined 3D pointcloud. Finally, our method does not require disparity maps of all the light field sub-views. 3. Facial Light Field Image Dataset Generation The key to the success of CNN-based 3D face reconstruction from a single RGB image lies in the availability of large training datasets. However, there is no large scale dataset available that provides RGB face images and their corresponding high quality 3D models. Similarly, training a light field face reconstruction network requires a large-scale Figure 3. EPIs corresponding to the 3D face curves. (a) Horizontal and vertical EPIs are obtained between the central view and sub-aperture images that are in the same row and column. (b) and (c) Visualization of the relationship between depth curves and slopes of lines in horizontal and vertical EPIs respectively. light field face dataset with corresponding ground truth 3D facial scans. Over the past few years, the computer vision community has made considerable efforts to collect light field images [22, 53, 33, 35, 1] for different applications. The only public light field face dataset [45] captured by Lytro IllumTM camera consists of 100 identities with 20 samples per person. However, depth maps of this dataset are generated using the Lytro Desktop SoftwareTM and have low resolution as well as low depth accuracy. Therefore, this dataset is not suitable for training a network. In the absence of large-scale 4D light field face datasets, we propose to generate a dataset of light field face images with ground truth 3D models. For this purpose, we use the public BU-3DFE [58] and BU-4DFE [60] databases to generate light field face images. The former is used for training and testing whereas the latter is used only for testing only. The BU-3DFE dataset consists of 2,500 3D scans from 100 identities (56% female, 44% male), with an age range from 18 to 70 years and multiple ethnicities. Each subject is scanned in one neutral and 6 non-natural expressions each with four intensity levels. The BU-4DFE dataset contains 3D video sequences of 101 identities (58 female and 43 male) in six different facial expressions. We select the most representative frame of each expression sequence. As a result, our dataset contains 606 3D scans. These models contain shape details such as wrinkles of not only the fiducial area, but also the hair, ears and neck area which pose challenges for conventional 3D face reconstruction methods. All 3D models have RGB texture. To generate plausible synthetic light field face images, it is crucial to control the light field camera parameters, background and illumination properly during rendering. We use the open source Blender3 software and the light field camera tool proposed by Katrin Honauer et al. [22] for this purpose. We place a virtual light field camera in Blender with 15×15 3 http://www.blender.org Figure 4. Examples of EPIs and their corresponding 3D face curves.(a) Horizontal EPIs. (b) Vertical EPIs. micro-lenses and set its field of view to capture the 3D facial scans. Both BU-3DFE and BU-4DFE databases provide 3D facial models in the near frontal pose. We load the 3D models along with their textures in Blender and apply two rigid rotations (±15◦ ) in pitch and four in yaw (±15◦ and ±30◦ ). To synthesize photo realistic light field images, we apply randomly selected indoor and outdoor images as backgrounds. We place two lamps at different locations in the scene and randomly change their intensities to achieve lighting variations. The angular resolution of the synthetic light field image is 15 × 15 and the spatial resolution is 400×400. The ground truth depth maps are aligned with the central view of light field image. Examples of our synthetic light field images are shown in Figure 2. We implement a Python script4 in Blender on a 3.4 GHz machine with 8GB RAM to automatically generate the light field facial images. The process of synthesizing light field images can be parallelized since each sub-aperture image is rendered independently. In total, we use 80 identities from BU-3DFE dataset to synthesize 14,000 light field images with ground truth disparity maps. The remaining 20 subjects from BU-3DFE and all 101 subjects from the BU4DFE dataset are used as test data to generate 1,451 light field facial images for evaluation. 4. Proposed Method An overview of the proposed method for reconstructing facial geometry from light field image is shown in Figure 1 and the details follow. 4.1. Training Data A 4D light field image can be parameterized as L(u, v, x, y), where (x, y) and (u, v) represent the spatial and angular coordinates respectively [54]. When we fix v and y, then L(u, v ∗ , x, y ∗ ) defines a 2D horizontal EPI. Similarly the 2D vertical EPI can be represented as 4 The script for light field facial image synthesis will be made public. Figure 5. Our proposed FaceLFnet for learning 3D face curves from EPIs. It contains 4 dense blocks, followed by two fully connected layers. The layers between two neighboring blocks are defined as transition layers and change feature map sizes via convolution and pooling [23]. L(u∗ , v, x∗ , y) when we keep u and x constant. As shown in Figure 3, 2D EPIs demonstrate the linear characteristic of the light field image. The orientations of lines within the EPIs can infer the disparity of the corresponding 3D space points [54, 28, 59, 20, 21]. Equation (1) shows the relationship between the slope of the line and the disparity value where f is the light field camera parameter and k is the slope of the line. Z = −f × k, (1) As shown in Figure 3(b) and (c), EPIs correspond to the 3D facial curves from the ground truth. Different line slopes in the EPI indicate different curve shapes. We use 14,000 synthetic light field images corresponding to the 80 identities of BU-3DFE for training. All together we extract 11.2 Million horizontal and vertical EPIs as training samples. Figure 4 shows some example EPIs and their corresponding curves. Using EPI images as training data removes the need for a huge number of identities. Since each 3D face curve can be learned independently from its corresponding EPI, we are able to generate massive training data from a small number of 3D face scans. Note that we do not need any further data augmentation such as image inversion or multiple crops as our networks learn from the full EPIs. 4.2. FaceLFnet Architecture Each EPI in our case corresponds to a 3D face curve as shown in Figure 3 and 4. The goal is to predict the full 3D curve from the EPIs using deep learning. CNNs can learn slope information of the pixels from individual EPIs, however, pixel wise prediction is very challenging. Heber et al. [20] divided each EPI into patches for 3D scene estimation. The authors estimated the depth value from each EPI patch independently as it contained the information pertaining to a single line at the center of the patch. In our case, pixel wise estimation is not practical as our network must learn the inter-relationship between the lines in one full EPI to estimate the complete 3D curve. Furthermore, in case of light field images for faces, some EPI patches especially in the quasi planar facial areas are devoid of lines and hence do not contain enough depth information leading to inaccurate depth estimation. Therefore, we propose using a complete EPI for depth prediction in order to exploit the correlations of adjacent pixels and mitigate the problem of inaccurate depth estimation due to pixel wise prediction. The dimensions of each input EPI are 15 × 400 × 3 (horizontal/vertical sub-aperture images × horizontal/vertical image pixels × RGB channels). Such a low resolution in the first dimension and size disparity in the first two dimensions pose challenges as the information of the input EPIs will reduce rapidly in one dimension than the other when passed through a deep network. To extenuate this problem and inspired by the success of Gao et al. [23], we propose a light field face network for estimating facial geometry from EPIs. The architecture of our network is illustrated in Figure 5. It is based on DenseNet that consists of multiple dense blocks and transition layers. We use four dense blocks and change the softmax classifier to a regressor. Before passing the EPIs through the first dense block, a 16 channels convolution layer with 3 × 3 kernel size is used. For each dense block, we use three convolutional layers and set the growth rate to 12. We also use convolution followed by average pooling as transition layers between two adjacent dense blocks. The sizes of feature-map in the four dense blocks are 15 × 400, Layers Convolution Dense Block 1 Dense Block 4 Output Size 15 × 400 15 × 400 15 × 400 8 × 200 8 × 200 8 × 200 4 × 100 4 × 100 4 × 100 2 × 50 2 × 50 Regression Layer 400 Transition Layer 1 Dense Block 2 Transition Layer 2 Dense Block 3 Transition Layer 3 FaceLFnet 3 × 3 conv, stride 1 [3 × 3 conv, stride 1]×3 3 × 3 conv, stride 1 2 × 2 average pool, stride 2 [3 × 3 conv, stride 1]×3 3 × 3 conv, stride 1 2 × 2 average pool, stride 2 [3 × 3 conv, stride 1]×3 3 × 3 conv, stride 1 2 × 2 average pool, stride 2 [3 × 3 conv, stride 1]×3 4096 fully-connected 400 fully-connected EuclideanLoss Table 1. Our proposed FaceLFnet architecture. Note that each convolutional layer in the dense block corresponds to the sequence BN-ReLU. The growth rate of the four blocks is k = 12. Figure 6. Pose invariance. Columns one to four in each row respectively depict the input central view of the light field image, the ground truth 3D face, the reconstructed 3D face by our proposed method and the last two overlaid on each other. Figure 7. Expression invariance. As shown, our method can handle exaggerated expressions. Columns one to four in each row respectively depict the input central view of the light field image, the ground truth 3D face, the reconstructed 3D face by our proposed method and the last two overlaid on each other. 8 × 200, 4 × 100 and 2 × 50 respectively. The details of network configurations are given in Table 1. Both horizontal and vertical FaceLFnets are trained from scratch using the Caffe deep learning framework [27]. The initial learning rate is set to 0.0003 which is divided by 10 at 30000 and 50000 iterations. Our networks require only one epoch for convergence. The caffe model for the trained networks will be made public. 4.3. 3D Face Reconstruction The output of our horizontal and vertical FaceLFnets are 3D facial curves that together make up a 3D face. We combine all the horizontal and vertical curves (in our case 400 each) of a face to form a horizontal and a vertical depth map separately. The next step is to reconstruct a 3D face from the two depth maps. A naive way to reconstruct the face is to take the average of both depth maps. However, such a methodology results in reconstruction error as each curve was learned independently. To mitigate this problem we propose a technique to project the depth maps on a 2D surface. First of all, we convert both the depth maps to 3D pointclouds, using the camera parameters. Next we give a slight jitter to the horizontal pointcloud by translating it 1mm to the left on x-axis only. We fit a single surface of the form z(x, y) to both 3D pointclouds simultaneously using the gridfit algorithm [10]. Our method ensures that a smooth surface is fitted to the horizontal and vertical Figure 8. Invariance to illumination and skin color. Our method is robust to illumination variations and also works well in the case of dark skin (second row). pointclouds taking into account the correlation between the curves resulting in a smooth reconstructed 3D face. Figure 9. Qualitative results. The columns contain (in order) central view image, the ground truth 3D face, 3D face reconstructed by our method and 3D face reconstructed by Sela et al. [44]. 5. Experimental Results We now present the evaluation of our method for 3D face reconstruction from a single light field image on the 3D scans of the remaining 20 subjects from BU-3DFE and all 101 subjects from the BU-4DFE dataset. We compare our subjective results with the recent state-of-the-art algorithm [44] for qualitative evaluation. We also present quantitative comparison with VRN-Guided [25] and other stateof-the-art methods [44, 41, 64, 31, 63, 24] on both datasets. Note that the VRN-Guided method incorporates facial landmarks in their proposed VRN architecture whereas we follow a marker-less strategy. 5.1. Qualitative Evaluation For qualitative evaluation, we show our reconstruction results on light field images synthesized from BU3DFE [58] and BU-4DFE [60] databases. We also show the ground truth and predicted 3D face shapes overlaid on each other using the Scanalyze software. Figure 6 shows the reconstructed 3D faces under different poses to demonstrate that our method is robust to pose variations. Unlike model based algorithms for 3D face reconstruction [11, 44] from a single RGB image, our method can recover the 3D model of the full head including the peripheral regions such as hair and neck and sometimes even part of the clothing. Figure 7 shows our results under exaggerated expressions while Figure 8 shows our results under illumination changes. Note that our method is robust to variations in pose, expressions and illumination. We use the code provided by Sela et al. [44] for qualitative comparison of the reconstructed faces. Figure 9 shows 3D faces reconstructed from light field images using our method and 3D faces reconstructed from single central view RGB images using the recent state-of-the-art method proposed by Sela et al. [44]. Since [44] estimate only the facial region, we also crop our reconstructed faces for better visual comparison. As demonstrated, our method produces more visually accurate reconstructions in the global geometry compared to [44]. As compared to methods based on fine-tuning, our method can not capture fine details since we use the output of our network directly without complex post-processing steps. Our proposed method performs better than [44] because, firstly, [44] relies on a face detector and crops the input RGB image based on the detected coordinates while our method does not need any face detection or cropping. Secondly, [44] synthesized their training data from 3DMM parameters and thus their training images do not have the neck and hair regions etc. When the input images are far from the model space, the global face shape will be unsatisfactory at some key facial regions like mouth, nose and eyes as can be seen in Figure 9. Finally, Sela et al. [44] use non-rigid registration to fit the 3DMM to the coarse output of the proposed network. The model fitting process deforms the facial shape when the model and the coarse shape estimated by the network are quite different. 5.2. Quantitative Evaluation For quantitative comparison, we evaluate the 3D reconstruction on 3,500 light field images of 20 subjects from BU-3DFE [58] and 1,400 light field images of 101 subjects from the BU-4DFE dataset. To measure the affect of pose on the reconstruction accuracy, we use the 3,500 light field images from BU-3DFE dataset. There are 500 light field images for each pose. We use the Root Mean Square Error (RMSE) between the 3D point clouds of the estimated and ground truth reconstructions as a quantitative measure. Results of RMSE for different poses are depicted in Figure 10. Our method is robust to pose variations as the RMSE error increases by only 0.31mm when the pose is varied by 30 degrees. To measure the affect of facial expressions on reconstruction accuracy, we synthesize frontal images in different expressions (Angry, Disgust, Fear, Happy, Sad and Surprise) from the BU-4DFE dataset and measure the reconstruction errors. Figure 11 shows that the RMSE of 3D face reconstruction from our method is small even in the presence of exaggerated expressions. We compare the absolute depth error of our proposed method with the state-of-the-art in Table 2, which shows that our proposed 3D reconstruction outperforms all existing methods. We report depth errors evaluated by mean, standard deviation, median and the average ninety percent largest error. Note that for a fair comparison with Sela et al. [44] we report the results obtained on the same dataset directly from their paper instead of calculating the reconstruction errors from our implementation of their work. We also compare the results of our method with VRNGuided [25], 3DDFE [63] and EOS [24] methods using the BU-4DFE dataset [60]. We use the Normalized Mean Error (NME) metric proposed by Aarson [25] to report the results for comparison with existing methods. NME is defined as the average per vertex Euclidean distance between the estimated and the ground truth reconstruction normalized by the outer 3D interocular distance: n 1 X kxk − yk k2 , (2) NME = n d k=1 where n is the total number of vertices per facial mesh and d is the interocular distance. xk and yk represent the coordinates of vertices from the estimated and ground truth meshes respectively. The NME is calculated on the face region only. As shown in Table 3, our method outperforms the state-of-the-art. 6. Conclusion We presented a model-free approach for recovering the 3D facial geometry from a single light field image. We proposed FaceLFnet, a densely connected network architecture Figure 10. Reconstruction errors for different facial poses on the BU-3DFE dataset [58]. Note that the RMSE increases from 2.62 to 2.93 (by 0.31 mm only) under extreme pose variations. Figure 11. Reconstruction errors for different facial expressions on the BU-4DFE dataset [60]. The RMSE increases from 2.49 to 2.98 (by only 0.49 mm) under extreme expression variations. Sad has the highest error whereas surprise has the lowest because of more edges around the lips which favors EPI based reconstruction. Error in mm Mean SD Median 90% largest Kemelmacher et al. [31] 3.89 4.14 2.94 7.34 Zhu et al. [64] 3.85 3.23 2.93 7.91 Richardson et al. [41] 3.61 2.99 2.72 6.82 Matan et al. [44] 3.51 2.69 2.65 6.59 Ours 2.78 2.04 1.73 5.30 Table 2. Comparative results on the BU-3DFE dataset [58]. The absolute RMSE between ground truth and predicted shapes evaluated by mean, standard deviation, median and the average ninety percent largest error of the different methods are presented. NME 3DDFA[63] 5.14 EOS[24] 5.33 VRN-Guided[25] 4.71 Ours 3.72 Table 3. Reconstruction errors on the BU-4DFE dataset [60] in terms of NME defined in Eq. (2). ICP has been used to align the reconstructed face to the ground truth similar to [25]. that regresses the 3D facial curves over the Epipolar Plane Images. Using a curve by curve reconstruction approach, our method needs only a few training samples and yet generalizes well to unseen faces. We proposed a photo-realistic light field image synthesis method to generate a large-scale EPI dataset from a relatively small number of real facial identities. Our results show that 3D face reconstruction from light field images is more accurate and allows the use of a model-free approach which is robust to changes in pose, facial expressions, ethnicities and illumination. We conclude that light field cameras are a more appropriate choice as a passive sensor for 3D face reconstruction since they enjoy similar advantages to conventional RGB cameras in that they are point and shoot, portable and have low cost. These cameras are especially a better choice for medical applications where higher accuracy and model-free approaches are desirable. We will make our trained networks and dataset public which will become the first photo-realistic light field face dataset with ground truth 3D facial scans. Acknowledgments This research was supported by ARC grant DP160101458. The Titan Xp used for this research was donated by NVIDIA Corporation. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] http://lightfield.stanford.edu/. 4 https://www.lytro.com/. 2 https://www.raytrix.com/. 2 A. F. Abate, M. Nappi, D. Riccio, and G. Sabatino. 2D and 3D face recognition: A survey. Pattern Recognition Letters, 28(14):1885–1906, 2007. 1 O. Aldrian and W. A. Smith. Inverse rendering of faces with a 3d morphable model. IEEE transactions on pattern analysis and machine intelligence, 35(5):1080– 1093, 2013. 3 P. N. Belhumeur, D. J. Kriegman, and A. L. Yuille. The bas-relief ambiguity. International Journal of Computer Vision, 35(1):33–44, Nov 1999. 1, 2 V. Blanz and T. Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 187–194. ACM Press/AddisonWesley Publishing Co., 1999. 3 J. Booth, E. Antonakos, S. Ploumpis, G. Trigeorgis, Y. Panagakis, and S. Zafeiriou. 3d face morphable models ”in-the-wild”. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 3 C. Cao, Y. Weng, S. Lin, and K. Zhou. 3d shape regression for real-time facial animation. ACM Transactions on Graphics (TOG), 32(4):41, 2013. 1 J. D́Erico. Surface fitting using gridfit. In MATLAB Central File Exchange, 2008. 6 P. Dou, S. K. Shah, and I. A. Kakadiaris. End-to-end 3d face reconstruction with deep neural networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 1, 2, 3, 7 [12] T. Fang, X. Zhao, O. Ocegueda, S. K. Shah, and I. A. Kakadiaris. 3d/4d facial expression analysis: An advanced annotated face model approach. Image and vision Computing, 30(10):738–749, 2012. 1 [13] S. Z. Gilani, A. Mian, and P. Eastwood. Deep, dense and accurate 3D face correspondence for generating population specific deformable models. Pattern Recognition, 69:238–250, 2017. 2 [14] S. Z. Gilani, A. Mian, F. Shafait, and I. Reid. Dense 3D face correspondence. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 1 [15] S. Z. Gilani, K. Rooney, F. Shafait, M. Walters, and A. Mian. Geometric facial gender scoring: Objectivity of perception. PloS one, 9(6), 2014. 1 [16] P. Hammond. The use of 3d face shape modelling in dysmorphology. In Archives of disease in childhood, page 92(12), 2007. 1 [17] P. Hammond, C. Forster-Gibson, A. Chudley, et al. Face–brain asymmetry in autism spectrum disorders. Molecular Psychiatry, 13(6):614–623, 2008. 1 [18] Y. Han, J.-Y. Lee, and I. So Kweon. High quality shape from a single rgb-d image under uncalibrated natural illumination. In Proceedings of the IEEE International Conference on Computer Vision, pages 1617–1624, 2013. 2 [19] T. Hassner, S. Harel, E. Paz, and R. Enbar. Effective face frontalization in unconstrained images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4295–4304, 2015. 3 [20] S. Heber and T. Pock. Convolutional networks for shape from light field. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3746–3754, 2016. 3, 5 [21] S. Heber, W. Yu, and T. Pock. U-shaped networks for shape from light field. In BMVC, 2016. 3, 5 [22] K. Honauer, O. Johannsen, D. Kondermann, and B. Goldluecke. A dataset and evaluation methodology for depth estimation on 4d light fields. In Asian Conference on Computer Vision, pages 19–34. Springer, 2016. 2, 4 [23] G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger. Densely connected convolutional networks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 5 [24] P. Huber, G. Hu, R. Tena, P. Mortazavian, P. Koppen, W. J. Christmas, M. Ratsch, and J. Kittler. A multiresolution 3d morphable face model and fitting framework. In Proceedings of the 11th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications, 2016. 7, 8 [25] A. S. Jackson, A. Bulat, V. Argyriou, and G. Tzimiropoulos. Large pose 3d face reconstruction from a single image via direct volumetric cnn regression. In The IEEE International Conference on Computer Vision (ICCV), Oct 2017. 7, 8 [26] H.-G. Jeon, J. Park, G. Choe, J. Park, Y. Bok, Y.-W. Tai, and I. So Kweon. Accurate depth map estimation from a lenslet light field camera. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1547–1555, 2015. 2 [27] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia, pages 675–678. ACM, 2014. 6 [28] O. Johannsen, A. Sulc, and B. Goldluecke. What sparse light field coding reveals about scene structure. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3262– 3270, 2016. 5 [29] A. Jourabloo and X. Liu. Pose-invariant face alignment via cnn-based dense 3d model fitting. International Journal of Computer Vision, pages 1–17, 2017. 2, 3 [30] V. Kazemi, C. Keskin, J. Taylor, P. Kohli, and S. Izadi. Real-time face reconstruction from a single depth image. In 3D Vision (3DV), 2014 2nd international conference on, volume 1, pages 369–376. IEEE, 2014. 3 [31] I. Kemelmacher-Shlizerman and R. Basri. 3d face reconstruction from a single image using a single reference face shape. IEEE transactions on pattern analysis and machine intelligence, 33(2):394–405, 2011. 1, 2, 3, 7, 8 [32] N. Li, B. Sun, and J. Yu. A weighted sparse coding framework for saliency detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5216–5223, 2015. 2 [33] N. Li, J. Ye, Y. Ji, H. Ling, and J. Yu. Saliency detection on light field. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2014. 4 [34] H. Lin, C. Chen, S. Bing Kang, and J. Yu. Depth recovery from light field using focal stack symmetry. In Proceedings of the IEEE International Conference on Computer Vision, pages 3451–3459, 2015. 2 [35] K. Marwah, G. Wetzstein, Y. Bando, and R. Raskar. Compressive light field photography using overcomplete dictionaries and optimized projections. ACM Transactions on Graphics (TOG), 32(4):46, 2013. 4 [36] A. Mian, M. Bennamoun, and R. Owens. An efficient multimodal 2d-3d hybrid approach to automatic face recognition. IEEE transactions on pattern analysis and machine intelligence, 29(11), 2007. 1 [37] R. Or-El, G. Rosman, A. Wetzler, R. Kimmel, and A. M. Bruckstein. Rgbd-fusion: Real-time high precision depth recovery. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5407–5416, 2015. 2 [38] A. Patel and W. A. Smith. 3d morphable face models revisited. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 1327–1334. IEEE, 2009. 3 [39] C. Queirolo, L. Silva, O. Bellon, and M. Segundo. 3D face recognition using simulated annealing and the surface interpenetration measure. IEEE TPAMI, 32(2):206–219, 2010. 1 [40] E. Richardson, M. Sela, and R. Kimmel. 3d face reconstruction by learning from synthetic data. In 3D Vision (3DV), 2016 Fourth International Conference on, pages 460–469. IEEE, 2016. 1, 2, 3 [41] E. Richardson, M. Sela, R. Or-El, and R. Kimmel. Learning detailed face reconstruction from a single image. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 3, 7, 8 [42] J. Roth, Y. Tong, and X. Liu. Adaptive 3d face reconstruction from unconstrained photo collections. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4197–4206, 2016. 3 [43] A. Savran, N. Alyüz, H. Dibeklioğlu, O. Çeliktutan, B. Gökberk, B. Sankur, and L. Akarun. Bosphorus database for 3d face analysis. Biometrics and identity management, pages 47–56, 2008. 1 [44] M. Sela, E. Richardson, and R. Kimmel. Unrestricted facial geometry reconstruction using image-to-image translation. In The IEEE International Conference on Computer Vision (ICCV), Oct 2017. 2, 3, 7, 8 [45] A. Sepas-Moghaddam, V. Chiesa, P. L. Correia, F. Pereira, and J.-L. Dugelay. The ist-eurecom light field face database. In Biometrics and Forensics (IWBF), 2017 5th International Workshop on, pages 1–6. IEEE, 2017. 4 [46] H. Sheng, P. Zhao, S. Zhang, J. Zhang, and D. Yang. Occlusion-aware depth estimation for light field using multi-orientation epis. Pattern Recognition, 2017. 2 [47] D. W. Tan, S. Z. Gilani, M. T. Maybery, A. Mian, A. Hunt, M. Walters, and A. J. Whitehouse. Hypermasculinised facial morphology in boys and girls with autism spectrum disorder and its association with symptomatology. Scientific Reports, 7(1):9348, 2017. 1 [48] M. W. Tao, P. P. Srinivasan, J. Malik, S. Rusinkiewicz, and R. Ramamoorthi. Depth from shading, defocus, and correspondence using light-field angular coherence. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1940– 1948, 2015. 2 [49] J. Thies, M. Zollhofer, M. Stamminger, C. Theobalt, and M. Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2387–2395, 2016. 1 [50] J. Tian, Z. Murez, T. Cui, Z. Zhang, D. Kriegman, and R. Ramamoorthi. Depth and image restoration from light field in a scattering medium. In The IEEE International Conference on Computer Vision (ICCV), Oct 2017. 2 [51] A. Tuan Tran, T. Hassner, I. Masi, and G. Medioni. Regressing robust and discriminative 3d morphable models with a very deep neural network. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 2, 3 [52] T.-C. Wang, A. A. Efros, and R. Ramamoorthi. Occlusion-aware depth estimation using light-field cameras. In Proceedings of the IEEE International Conference on Computer Vision, pages 3487–3495, 2015. 2 [53] T.-C. Wang, J.-Y. Zhu, E. Hiroaki, M. Chandraker, A. A. Efros, and R. Ramamoorthi. A 4d light-field dataset and cnn architectures for material recognition. In European Conference on Computer Vision, pages 121–138. Springer, 2016. 2, 4 [54] S. Wanner and B. Goldluecke. Variational light field analysis for disparity estimation and super-resolution. IEEE transactions on pattern analysis and machine intelligence, 36(3):606–619, 2014. 4, 5 [55] A. J. Whitehouse, S. Z. Gilani, F. Shafait, A. Mian, D. W. Tan, M. T. Maybery, J. A. Keelan, R. Hart, D. J. Handelsman, M. Goonawardene, et al. Prenatal testosterone exposure is related to sexually dimorphic facial morphology in adulthood. In Proc. R. Soc. B, volume 282. The Royal Society, 2015. 1 [56] G. Wu, M. Zhao, L. Wang, Q. Dai, T. Chai, and Y. Liu. Light field reconstruction using deep convolutional network on epi. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 2 [57] Z. Xiong, L. Wang, H. Li, D. Liu, and F. Wu. Snapshot hyperspectral light field imaging. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 2 [58] L. Yin, X. Wei, Y. Sun, J. Wang, and M. J. Rosato. A 3d facial expression database for facial behavior re- search. In Automatic face and gesture recognition, 2006. FGR 2006. 7th international conference on, pages 211–216. IEEE, 2006. 2, 4, 7, 8 [59] S. Zhang, H. Sheng, C. Li, J. Zhang, and Z. Xiong. Robust depth estimation for light field via spinning parallelogram operator. Computer Vision and Image Understanding, 145:148–159, 2016. 5 [60] X. Zhang, L. Yin, J. F. Cohn, S. Canavan, M. Reale, A. Horowitz, and P. Liu. A high-resolution spontaneous 3d dynamic facial expression database. In Automatic Face and Gesture Recognition (FG), 2013 10th IEEE International Conference and Workshops on, pages 1–6. IEEE, 2013. 2, 4, 7, 8 [61] W. Y. Zhao and R. Chellappa. Illuminationinsensitive face recognition using symmetric shapefrom-shading. In Computer Vision and Pattern Recognition, 2000. Proceedings. IEEE Conference on, volume 1, pages 286–293. IEEE, 2000. 2 [62] H. Zhu, Q. Zhang, and Q. Wang. 4d light field superpixel and segmentation. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 2 [63] X. Zhu, Z. Lei, X. Liu, H. Shi, and S. Z. Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 146–155, 2016. 7, 8 [64] X. Zhu, Z. Lei, J. Yan, D. Yi, and S. Z. Li. Highfidelity pose and expression normalization for face recognition in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 787–796, 2015. 7, 8
1cs.CV
Group Decision Simulation 1 Evolutionary perspectives on collective decision making: Studying the implications of diversity and social network structure with agent-based simulations Hiroki Sayama Collective Dynamics of Complex Systems Research Group Departments of Bioengineering & Systems Science and Industrial Engineering Binghamton University, State University of New York, Binghamton, New York Phone: (607) 777-4439 Fax: (607) 777-5780 Email: [email protected] Shelley D. Dionne Center for Leadership Studies, School of Management Binghamton University, State University of New York, Binghamton, New York Francis J. Yammarino Center for Leadership Studies, School of Management Binghamton University, State University of New York, Binghamton, New York Abstract Collective, especially group-based, managerial decision making is crucial in organizations. Using an evolutionary theory approach to collective decision making, agent-based simulations were conducted to investigate how collective decision making would be affected by the agents’ diversity in problem understanding and/or behavior in discussion, as well as by their social network structure. Simulation results indicated that groups with consistent problem understanding tended to produce higher utility values of ideas and displayed better decision convergence, but only if there was no group-level bias in collective problem understanding. Simulation results also indicated the importance of balance between selection-oriented (i.e., exploitative) and variation-oriented (i.e., explorative) behaviors in discussion to achieve quality final decisions. Expanding the group size and introducing non-trivial social network structure generally improved the quality of ideas at the cost of decision convergence. Simulations with different social network topologies revealed that collective decision making on small-world networks with high local clustering tended to achieve highest decision quality more often than on random or scalefree networks. Implications of this evolutionary theory and simulation approach for future managerial research on collective, group, and multi-level decision making are discussed. Keywords: collective decision making, evolutionary theory, agent-based simulation, group homogeneity and heterogeneity, social networks Group Decision Simulation 2 Evolutionary perspectives on collective decision making: Studying the implications of diversity and social network structure with agent-based simulations Abstract Collective, especially group-based, managerial decision making is crucial in organizations. Using an evolutionary theory approach to collective decision making, agent-based simulations were conducted to investigate how collective decision making would be affected by the agents’ diversity in problem understanding and/or behavior in discussion, as well as by their social network structure. Simulation results indicated that groups with consistent problem understanding tended to produce higher utility values of ideas and displayed better decision convergence, but only if there was no group-level bias in collective problem understanding. Simulation results also indicated the importance of balance between selection-oriented (i.e., exploitative) and variation-oriented (i.e., explorative) behaviors in discussion to achieve quality final decisions. Expanding the group size and introducing non-trivial social network structure generally improved the quality of ideas at the cost of decision convergence. Simulations with different social network topologies revealed that collective decision making on small-world networks with high local clustering tended to achieve highest decision quality more often than on random or scalefree networks. Implications of this evolutionary theory and simulation approach for future managerial research on collective, group, and multi-level decision making are discussed. Keywords: collective decision making, evolutionary theory, agent-based simulation, group homogeneity and heterogeneity, social networks Group Decision Simulation 3 Evolutionary perspectives on collective decision making: Studying the implications of diversity and social network structure with agent-based simulations 1 Introduction Collective decision making plays an increasingly important role in society and organizations today (Mannes 2009; Kerr and Tindale 2004). In high-tech industries, for example, the number of engineers participating in the design of a single product can amount to hundreds or even thousands due to the increase of the product’s complexity far beyond each individual engineer’s capacity, which almost inevitably results in suboptimal outcomes (Klein et al. 2003a). Another example is the online collective decision making among massive anonymous participants via large-scale computer mediated communication networks, including collective website/product rating and common knowledge base formation (O’Reilly 2005). In these and related cases, participants and their societal or organizational structure may influence the final outcome of decision making processes. The complexity of the process is more pronounced when the participants are heterogeneous and are embedded in a topologically nonuniform network with differential distribution of power, as in most organizations and social systems. The dynamics of collective decision making in such conditions are poorly understood, and as such poses a significant challenge for the social and organizational sciences. Evidence of this challenge exists within the leadership, psychology and organizational behavior/management disciplines where collective dynamics, using both experimental and applied studies, generally emphasize linear statistical relationships between specific, narrowly defined team- or individual-level variables (Kerr and Tindale 2004; Salas et al. 2004; Dionne et al. 2012). Traditional studies seldom account for nonlinear dynamical processes that take place in a high-dimensional problem space and/or non-trivial social structure where interactions occur within a networked organizational structure. Abbott (2001) highlights this problem within the social sciences by discussing a “general linear reality,” where mainstream social science theories and methods treat linear models as actual representations of social systems. Further, Meyer et al. (2005) assert that reinforcing assumptions of linearity and equilibrium have “blocked the investigation of a family of interesting problems of great practical import (p. 456).” In other words, researchers point to an oversimplification of collective processes in general and collective decision making in particular. Examples of recent research not necessarily following a “general linear reality” to model inherent complexity in social systems are found within the complex systems research community, where social processes are studied using a mathematical/computational modeling approach (Simon 1955; Axelrod 1981; Epstein and Axtell 1996; Bar-Yam 1997; Bar-Yam 2004; Gilbert and Troitzsch 1999; Sterman 2000; Epstein 2006; Miller and Page 2007; Sayama, Farrell and Dionne 2011; Yamanoi and Sayama Group Decision Simulation 4 2012). Because emphasis is on emergent dynamical behavior of systems caused by nonlinear interactions among massive numbers of parts (a pervasive phenomenon also found in fields such as physics, biology, sociology, psychology, economics, engineering and computer science), advances in other scientific domains for modeling complex systems may benefit organizational research (Carroll and Burton 2000). Thus, the aim of this research is to enhance performance of groups and other entities involved in collective decision making. Collective decision making implies a larger clustering of individuals with interdependency based on shared expectations or hierarchy. Collectives can be complicated structures and include individuals, groups, functional business units and even larger industry alliances and networks (Dansereau, Alutto and Yammarino, 1984; Yammarino and Dansereau 2002; Yammarino et al. 2005). We seek to improve our understanding of both the dynamic nature of the collective decision process, as well as the influence of diversity and social connectivity issues related to decision making involving a number of participating group members. Employing evolutionary views in understanding decision making enables a uniform, straightforward explanation of many empirical findings about the effects of group composition and dynamics on group performance. Considering specific within-group level issues regarding the collaborative process of decision making also may offer clarity regarding the influence of group composition on performance. Borrowing from advances in complex systems enables a dynamic, withinlevel examination of an evolutionary collective decision process via computational modeling. To better understand decision making in collective and dynamic environments, we offer an alternative model of collective decision making based on evolutionary theory, examining the issues of diversity/heterogeneity in decision making groups and the effects of non-trivial social network structure among the group members on decision outcomes. We first explore how evolutionary theory can address complex changes over time by providing an explanatory framework for collective decision processes, and then discuss how specifying a targeted level of analysis can inform appropriate interpretation and limitations of decision making in dynamic environments. Finally, a computational agent-based model (Epstein 2006) with an evolutionary focus on collective decision making in groups and social networks is developed and tested, with diversity of problem understanding and behavioral patterns and social network structure manipulated as experimental variables. Specifically, four recommendations can be adapted to advance our theoretical understanding of collective decision making in complex social systems: 1) consider the impact of time; 2) study situations in flux; 3) incorporate nonlinear concepts into evolutionary theorizing; and 4) design multi-level research (Meyer et al. 2005) that takes into account complex social network topologies. These guidelines provide a starting point for investigating the complexity of collective decision making with an evolutionary and multi-level, network-oriented framework. Prior dynamical modeling in organizational research may have Group Decision Simulation 5 considered the impact of time and situations in flux; few if any, however, have included specific evolutionary and multi-level, network-oriented concepts. 2 Evolutionary theory and collective decision making Evolutionary theory describes adaptive changes of populations primarily by combining mechanisms of variation and selection (Futuyma 2005; Wilson 2005). The roles of these two mechanisms are similar to what was already discussed as “exploration” and “exploitation” in organizational learning literature (March 1991). In biological evolution, variation is caused primarily by internal genetic mechanisms (e.g., mutation and recombination) and plays an exploratory role that could potentially lead to a novel possibility of life form, but it usually reduces immediate competitiveness of a population. In contrast, selection is caused primarily by external environment (e.g., natural and sexual selection) and plays an exploitative role that enhances the presence of successful entities (genes, individuals, or groups) and eliminates unsuccessful ones, reducing the number of possibilities while potentially improving the overall competitiveness of the population. A dynamically maintained balance of the two mechanisms is the key to a successful evolutionary adaptation (Mitchell et al. 1991). We propose that decision making processes within a collective (such as a group or an organization) also may be viewed through a similar lens, by shifting the viewpoint from members’ personal properties (a more traditional psychological and decision making approach) to dynamical changes of ideas being discussed within the collective, where populations of potential ideas evolve via repetitive operations such as reproduction, recombination, mutation, and selection of ideas, conducted by participating members (Sayama and Dionne 2013a; Sayama and Dionne 2013b). Table 1 provides a brief summary of the evolutionary framework we propose by illustrating how some key evolutionary operators and concepts can be linked to decision factors. We take this approach because evolutionary theory provides a powerful theoretical framework that can readily address complex changes of systems over time in extremely highdimensional problem space, and also its explanatory mechanisms (heredity, variation, and selection) are theoretically clean-cut and easily accessible (Wilson 2005). Moreover, by shifting the viewpoint from individuals to ideas, a model could be liberated from the commonly used but somewhat artificial assumption that each individual always has his/her decision in mind. Rather, various ideas developed within and among participants are collectively reflected in the idea population, to which diverse withinindividual cognitive/behavioral patterns can be easily applied as a set of multiple evolutionary operators simultaneously acting on the same, shared idea population. -----------------------------Insert Table 1 Here -----------------------------2.1 Evolutionary operators and collective decision processes Group Decision Simulation 6 Many evolutionary operators (Mitchell 1996; Fogel 1995) can be conceived as a representation of diverse human behaviors in discussion and decision making processes (Sayama and Dionne 2013a; Sayama and Dionne 2013b; also see Table 1). For a selection-oriented example, replication of an idea is a form of positive selection, where the popularity of an idea is increased within the population of ideas. This action can be considered as advocacy of a particular idea under discussion. Similarly, criticism against an idea may also be considered a form of negative, subtractive selection. Criticism singles out an idea with poor utility and reduces its popularity within the population of ideas. Both positive and negative selections seek to narrow decision possibilities based on utilities (“fitness”) of ideas perceived by participants. For a variation-oriented example, random point mutation makes random changes to existing ideas by asking “what if”-type questions. Likewise, intelligent, or hill-climbing point mutation (Klein et al. 2003b), which is not present in biological evolution but may be relevant to include in collective decision making, initially begins like random point mutation, however several variations from the original idea may occur, with final selection of the idea with the highest perceived utility. Recombination represents the creation of a new idea by crossing multiple existing ideas. These variation-oriented evolutionary operators enhance the explorative capabilities of the population, but generally reduce their immediate fitness. To illustrate the typical evolutionary operators that may be present in collective decision making, consider a collective decision task for a design and marketing group from a toy maker that wants a stronger presence in the girls toys market. The company currently has a strong presence in the boys toys market with a construction block toy, but the toy’s predominantly primary and/or dark color blocks, rectangular and square shapes that construct buildings and vehicles, and generic figures that look like little men have not had significant appeal to girls. The team is charged with moving into the girls toys market to appeal to the other 50% of toy users. Here are some examples of evolutionary operators in action. 2.1.1 Replication One designer offers an idea to increase the color palate currently offered in the bricks, adding pastels and lighter colors. Another designer offers an idea to produce new shapes of blocks besides predominantly four-sided, regular blocks. Another designer thinks they should consider developing snap together jewelry. A marketing member of the team states they do not have an idea, but they really like the new color palate idea. This concurrence with a prior idea increases its popularity by adding an exact copy of an idea that already exists in the population, and is a form of advocacy. 2.1.2 Random Point Mutation One designer takes the new color palate idea and wonders “what if” figures could be color coordinated as well, and moreover, some figures could look like little women, not just little men. The Group Decision Simulation 7 copied idea (new color palate) is offered, but with a focus on including new figure designs (i.e., different gender) that could be outfitted in different colors. This random change to an existing idea represents random point mutation. 2.1.3 Intelligent Point Mutation Consider the idea generated above, where new figures of both men and women have color palates such as dark and bright colors on men figures and light and pastel colors on women figures. After a period of reflection of some of team’s ideas under consideration, a new idea emerges. For example, after the color palette idea emerges, each of the designers considers this new direction for a while. Then, someone suggests, with confidence, that adding accessories such as hats, shoes, hair and clothing, to customize the figures, would be the best way to go. His/her confidence comes from several internal trial-and-error thought processes inside him/her, which makes his/her suggestion more intelligent than other random “what-if” suggestions. 2.1.4 Recombination Consider a creativity exercise for the toy makers where all generated ideas are written down and placed into a jar. Two ideas are pulled from the jar randomly and the team is told to generate new ideas by linking them together. For example, a person may draw the ideas “new shapes for blocks” and “accessories for male and female figures” to develop an idea that accessories should snap on using the same technology that snap building blocks together. This idea means some new shapes (i.e., round blocks for hats, hair) need to be constructed and applied to the figures. This combines accessorizing figures with new shapes for building blocks (i.e., recombination of ideas). 2.1.5 Subtractive Selection In the discussion, someone may state that snap together jewelry moves too far away from the brand and therefore should be stricken from the idea list. This represents subtractive selection. This reduces the popularity of the idea in discussion so that its chance of being considered favorably may become lower. Moreover, besides the evolutionary operators, other evolutionary concepts can be adapted or applied to a collective decision making process as well. A group’s problem or decision space may be likened to a genetic possibility space. A potential idea , or set of choices for all aspects of a problem, may be akin to a genome. Similarly, a particular aspect of the problem could map to a locus on a genome, whereas a specific choice made for a particular aspect may be likened to allele (i.e., specific gene) on a locus. Adding to parallels in evolutionary framework application to decision making, consider that a set of potential ideas under discussion may map to the concept of population. Additionally, perceived (or real) utility value of a potential idea may represent fitness, and the increased utility value achieved by the idea population may be likened to adaptation. Group Decision Simulation 8 As summarized in Table 1, using an evolutionary theory perspective, we define collective decision making as ecologies of ideas over a social network habitat, where populations of potential ideas evolve via continual applications of evolutionary operators such as replication, recombination, mutation, selection, and migration, each conducted by participating group members. Thus, there appears to be an intuitive parallel between an evolutionary framework and a collective decision process. Applying an evolutionary theory to collective decision making seems consistent with the spirit of the Meyer et al. (2005) suggestions regarding improvement of research techniques to better reflect situations in flux and nonlinear concepts within an evolutionary framework. The next section addresses another important Meyer et al. (2005) concern: multi-level research. 3 Levels of analysis and evolution Evolutionary biologists Wilson and Wilson (2008) reiterate the link between adaptation and a specific regard for levels of analysis in reviewing the history of multi-level selection theory. Wilson and Wilson’s (2008) recent evolutionary perspective on multi-level selection challenges researchers to evaluate the balance between levels of selection, specifically where within-group selection is opposed by between-group selection. This deeper view of a multi-level evolutionary process can be applied to organizational research as well (Yammarino and Dansereau, 2011). Research on both levels of analysis within organizational behavior (Yammarino and Dansereau 2002; Dansereau et al. 1999; Klein et al. 1994; Dansereau et al. 1984) and on group collaborative processes (Chang and Harrington 2005, 2007; van Ginkel and van Knippenberg 2008) highlight the importance and value of explicitly viewing the heterogeneity and/or homogeneity of the group and/or collective. This homogeneity and heterogeneity perspective can be viewed as a within-level examination, where the entity of interest remains the group, but there can be at least two valid views at the collective level: homogeneity (what evolutionary theory refers to as a between-group focus) and heterogeneity (what evolutionary theory refers to as a withingroup focus). These two different perspectives for viewing groups are aligned with the concept of group wholes and group parts, a theoretical distinction developed within organizational behavior and applied to various research domains such as leadership and group dynamics (Yammarino and Dansereau 2002, 2011; Yammarino et al. 2005; Dansereau et al. 1999; Klein et al. 1994; Dansereau et al. 1984). A group wholes perspective indicates homogeneity within the group, and the relevant focus on the entity (i.e., groups) would be between groups, since differences within groups would be considered random (i.e., error). Conversely, a group parts perspective indicates heterogeneity within groups, and the relevant focus on the entity (i.e., groups) would be within groups, since differences between groups would be considered random (i.e., error). Group Decision Simulation 9 3.1.1 Homogeneity (group wholes, or between-group differences condition) The concept of differing perspectives on an entity can provide more specific insights regarding group processes, in that phenomena of interest may be more relevant when groups are homogeneous regarding their membership, but differ in characteristics from other groups. In this wholes condition, all members within a group possess the same (or at least very similar) characteristic, while in the next group all members possess some other characteristics that first group perhaps did not. Another view can be taken concerning amounts of a characteristic present, where members of a group would possess the same amount of a characteristic, while members of the next group also would possess the same characteristic, but all members would have more of that characteristic, or all members would have less of that characteristic. This view represents one patterning of characteristics and individuals in groups— homogeneity. 3.1.2 Heterogeneity (group parts, or within-group differences condition) From a contrasting perspective, phenomena of interest may be more relevant when groups are heterogeneous regarding their memberships. In this case, members within a group would have varying degrees of a characteristic, and the next group also would have members with varying degrees of a characteristic, and the same applies for all groups. This view is another patterning of characteristics and individuals in groups—heterogeneity. 3.2 Decision research and levels of analysis Precedent for a broadly applicable modeling approach has been established in the evolving architecture of problem-solving networks (Chang and Harrington 2007). This research enabled consideration of a generic problem-solving environment and assessment of emergence regularity of connectors within the problem environment. Moreover, Chang and Harrington’s research related to the modeling of both homogenous agents (2005) and heterogeneous agents (2007) is of interest to our work. Specifically, we use homogeneity and heterogeneity of groups as means for examining levels of analysis issues related to collective and/or group processes. Although Chang and Harrington’s (2005, 2007) modeling examines a more multi-level relationship between agents (individuals) and the larger environment, we are concerned with examining a withingroup, collective or collaborative decision process, where individuals would not be considered outside of the group. Our examination of a unique within-level evolutionary process, employing both within-group and between-group perspectives, is a novel view of collaborative decision making and advances the understanding of a collective environment. And, while Meyer and colleagues (2005) called for more multi-level research, this unique and deliberate view within a single level of analysis, but employing multiple perspectives (i.e., wholes/homogeneity and parts/heterogeneity) on that level, also advances social research in that it moves the field away from an oversimplified view of groups. Group Decision Simulation 10 A critical distinction of our research is that we are interested in examining a type of process occurring within the group over time, not necessarily the specific variables within the process. Dansereau, Yammarino and Kholes (1999) highlighted the nature of such research on differing perspectives of an entity and entity changes rather than on changes in specific variables over time. Because we are interested in the type of process occurring within the group during decision making, we agree with Dansereau and colleagues (1999) that the variables that characterize the level may change or remain stable, but the level of interest remains the same (in our case, the level of interest remains the group). Related, diversity and/or homogeneity and heterogeneity of groups and information sharing (Gruenfield et al. 1996; Gigone and Hastie 1993; Stasser and Stewart 1992; Stasser and Titus 1985) present an additional layer to the decision process that requires consideration. Nijstad and Kaps (2007) noted that homogeneity of preferences leads to a lack of sharing of unique information within a group, whereas preference diversity prevented premature consensus of the group and facilitated unbiased discussions of preferences. Lightle, Kagel and Arkes (2009) indicated individual heterogeneity in information recall may play an role in failure to identify hidden profiles within groups. Similarly, van Ginkel and van Knippenberg (2008) found that groups in decision tasks performed better when task representations emphasized information elaboration and the group acknowledged they shared the view of the task representation. These findings reinforced that groups tend to focus on finding common ground and reaching consensus, but highlighted the importance of understanding, as a group, the task representation. This shared understanding could be critical to group success and adaptation, and as such, we include an indicator of how well group members share a view of what constitutes the problem. Although advancements in decision research continue, many continue to focus on individual-level aspects related to a decision maker, such as how they adopt practical behavior rules (Maldonato 2007) or identification of performance moderator functions that may affect individual behaviors in simulated environments (Silverman et al. 2006). While multi-level implications exist in recent decision research (Nijstad and Kaps 2007; Van Ginkel and van Knippenberg 2008), there is limited specific focus on within-group level aspects of a decision process. Moreover, Maldonato (2007) notes there is likely no best way to view the decision process. As such, there may be some benefit to development of a preliminary model exploring the effect of membership similarity and differences on group-based decision processes from evolutionary and levels of analysis-based perspectives. Development of such a model advances understanding of collective decision making in that it builds on prior key decision research (Nijstad and Kaps 2007; Chang and Harrington 2007; Kock 1999, 2004; Knudsen and Levinthal 2007), incorporates the suggestions of improving organizational research offered by Meyer et al. (2005), and incrementally increases the complexity yet fuller understanding of the phenomena represented in prior collective decision models. The specific model assumptions are discussed in the next section. Group Decision Simulation 11 4 Modeling Dynamic Collective Decision Making Applying computational modeling to dynamical processes such as collective decision making may enable organizational researchers to more appropriately represent the potential nonlinearity of a collective process. For example, interdisciplinary exchange may have informed recent organizational research which includes several dynamical models proposed for collective decision making over social networks that consist of many interacting individuals (Klein et al. 2003a; Battiston et al. 2003a; Battiston et al. 2003b; Rodriguez and Steinbock 2004). These models, primarily an extension of models developed in theoretical physics, provide a novel, promising direction for research on group dynamics and collective decision making. A limitation of this research and more specifically its ability to model complex social systems, however, is the consideration of only simple problem spaces, typically made of binary or continuous numerical choices between 0 and 1. Increasingly complex nonlinear problem space has been modeled (Klein et al. 2003b; Rusmevichientong and Van Roy 2003, Klein et al. 2006) to consider interdependent networks of multiple aspects of a complex problem. This research, however, was not modeled in a collective, nontrivial societal context. This is not surprising because problems arise with collective decision models in that they commonly assume every individual agent has or makes his/her own decision. Following these assumptions, the collective decision making dynamic is represented as a process of propagation, interaction and modification of individual decisions. This is an over-simplified assumption compared to actual cognition processes and behavior of individuals and collectives (Lipshitz et al. 2001; Salas and Klein 2001). Individuals often keep multiple ideas in mind and may remain undecided during or even after a collective-level decision emerges. The collective decision forms not just through the interactions of individual decisions but also through the more active, dynamic exchanges of incomplete ideas and mental models being developed by every individual (Dionne, Sayama, Hao and Bush 2010). Such withinindividual mental and behavioral complexity has just begun to be included in computational models (c.f., Knudsen and Levinthal 2007; Dionne and Dionne 2008), and should be taken into account to a greater extent in order to investigate more detailed, dynamic aspects of collective decision making (Salas and Klein 2001). Although powerful in their ability to examine beyond the “general linear reality,” dynamical models of collective decision making are still at an initial, preliminary stage. However, if limitations in current collective decision research (such as those over-simplifications noted above) are straightforwardly addressed in a single model, the resulting model may be too complex to be useful and effective for scientific enquiries. Because so many details would be involved, the model may not be “transparent” enough to offer clear relationships between assumptions and outcomes (Miller and Page 2007, Adner et al. 2009, Ren, Carley and Argote 2006). Therefore, collective decision making models need to carefully Group Decision Simulation 12 balance the ability to represent the complexities of dynamical social interactions against the ability to straightforwardly explain collective decision processes and outcomes. In view of such contexts for computational models of social and organizational sciences, an agentbased model has already been proposed that applied the evolutionary framework introduced in the prior section to model collective decision making processes within a small-sized, well-connected social network structure (Sayama and Dionne 2013a, Sayama and Dionne 2013b). This model was used to conduct a specific within-level analysis on how homogeneity or heterogeneity of goals and decision utility functions among participants affect dynamics and the final outcomes of their collective decision making. We present here a new agent-based model that implements a systematic control of agents’ behavioral balance between selection-oriented and variation-oriented operators, together with much larger, non-trivial social network structure on which agents exchange ideas locally. We believe that our approach to social dynamics research can move the social sciences away from an oversimplified view in that it investigates nonlinear change in organizational research (Meyer et al. 2005). Moreover, examining a new theoretical framework is consistent with development of computational models, as Adner et al. (2009) recognize that simulation is generally an exercise in theory building. 4.1 4.1.1 Model assumptions Groups, or social networks Our model assumes that N agents are connected to a finite number of other agents via links through which ideas are exchanged. Each agent can memorize multiple ideas in its mind. Multiple copies of a single idea may be present, which represents a form of relative popularity for that idea to the agent. Each agent is initialized with a small number of randomly generated ideas in its mind at the beginning of a simulation. The agents begin to perform a set of actions on the population of ideas in their minds repeatedly for a fixed number of iterations. The order by which the agents take actions is randomized every time, but it is guaranteed that every agent does take exactly one action per iteration. This roundrobin format is commonly used in idea sharing phases with decision making techniques such as a nominal group technique and various brainstorm initiatives (Paulus and Yang 2000; Van de Ven and Delbecq 1974). As such, the number of actions performed in a simulation is a product of the number of agents N and the number of iterations T. While other group decision research has modeled hierarchical teams in decision models (c.f., Dionne and Dionne 2008), we make no assumptions regarding predetermined leadership and/or abilities within the team as several teams in organizations are self-led and share leadership responsibilities (Salas, Stagl and Burke 2004; Salas and Klein 2001). We investigate the potential impact of varying membership within the group (i.e., no assumption of identical abilities or uniform connectivities in general) on the Group Decision Simulation 13 potential pool of ideas. Since no single person is powerful enough to eliminate an idea from the group (i.e., shared leadership), we assumed that actions were performed on single copies of an idea, not the equivalence set of all idea replicates (described in detail below). 4.1.2 Utility functions The use of utility functions in collective decision research is a natural outgrowth of earlier research by Hollenbeck et al. (1995) noting team decision making theory can be considered an adaptation of individual decision models and decision alternatives can vary along a univariate continuum. This view supplies a multi-level (e.g., group parts and group wholes) perspective and allows for adaptation of individual utility functions throughout a collective decision process. Both factors can be represented and/or captured by collective decision computer models (c.f., Dionne and Dionne 2008). As such, the use of utility functions contributes to the development of this model as well. Agents develop and exchange ideas in an M-dimensional binary problem space, with a total of 2M possible ideas. For a simulation, every idea has an inherent utility value specified by a true utility function UT that is unavailable to any agent. Individual agents perceive idea utility values based on their own utility functions Uj constructed by adding noise to the master utility function UM. The master utility function UM may or may not be the same as UT, depending on the possibility of group-level bias (explained below). This initialization reflects the notion that today’s organizational problems are too complex for a single individual to solve (i.e., true utility value not available to any of group members), and therefore groups or collectives are assembled to solve problems and make decisions (Klein et al. 2003a; Salas and Klein 2001). Ideally, collectives function by bringing unique information from members (i.e., individual utility functions) together in such a way as to produce ideas that exceed an individual’s idea development capability (Kerr and Tindale 2004). We develop a semi-continuous assignment of utility values in the problem space in the following way. First, n representative ideas S = {vi} ( i = 1…n ) are randomly generated as strings of bits (zeros and ones), where each vi represents one idea of M bits. One idea is assigned the maximum utility value, 1, and another, the minimum utility value, 0. The remaining n – 2 ideas are assigned a random real value between 0 and 1. This action ensures that the entire range of utility values is always from 0 to 1, which makes it easier to compare different simulation results. The detailed shape of the distribution varies within this range for different simulation runs. The utility values of all possible ideas in the domain of the true utility function are defined by interpolation using the utility values of representative ideas in S. We use the Hamming distance as a measure of dissimilarity between two bit strings, which reflects the number of bits for which two strings vary (Hamming 1950). With this measure, the utility value of each possible idea v not present in S is calculated as a weighted average of the utility values of the representative ideas as follows: Group Decision Simulation 14 UT ∑ (v ) = n i =1 U T (vi ) ⋅ D(vi , v) −2 ∑ D(vi , v) −2 i =1 n (1) where v ∉ S is the idea in question, UT(vi) is the utility of a representative idea vi in S, and D(vi, v) is the Hamming distance between vi and v. This weighted average based definition gives the true utility function UT(v) a reasonably “smooth” utility value assignment in a high-dimensional problem space (i.e., similar ideas tend to have similar utility values in general). Such an underlying structure of the problem space is necessary for intelligent collective decision making to be better than unintelligent random trial and error. Note that the utility landscape construction method described above is different from that of Kauffman’s N-K fitness landscapes often used in management science (Kauffman 1993; Levinthal 1997; Rivkin 2000). We chose this approach because our method makes it easier and more straightforward to introduce group-level bias, i.e., discrepancy between the true and master utility functions. Group-level bias is simulated by adding random perturbation when the master utility function UM is constructed from the true utility function UT. Specifically, a bias β is included in the master utility function by flipping bits with probability 0.25β per bit on representative ideas in S, and by adding a random number in [–β, β] to utility values of the representative ideas. Their utility values are then renormalized to the range [0, 1]. The master utility function UM is then generated from these perturbed representative idea set in the same way as in Eq. (1). Adding bias changes fidelity of information at the group level, where β = 0 denotes perfect understanding of the problem (UM = UT) as a collective, while complete lack of understanding is asymptotically approached as bias increases. Moreover, each agent will unconsciously have a different set of utility values for the possible ideas of the problem. Individual utility functions Uj(v) ( j = 1…N ) are generated by adding random noise to the master utility function UM so that: U j (v) ∈ [max(U M (v) − ξ , 0), min(U M (v) + ξ ,1)] (2) for all v, where ξ is the parameter that determines the range of noise. Figure 1 shows an example of such individual utility functions in contrast to the master utility function. As bounded rational actors, agents are not aware of the full set of alternatives available to them, nor can agents fully specify potential actionpotential outcome causal linkages (Gavetti and Levinthal 2000). Therefore agents in our model are not aware of the entire structure of their own individual utility functions. They cannot tell what ideas would produce global maximum/minimum utility values, though they can retrieve a utility value from the function when a specific idea is given, which is a common assumption made in complex global optimization problems (Horst, Pardalos and Thoai 2000). -----------------------------Insert Figure 1 Here Group Decision Simulation 15 -----------------------------Although not explicitly aware of their entire utility function structure, agents in a homogeneous condition can represent a “group wholes” view, where agents have similar utility functions within the group, but across groups there exist different utility functions; yet all members of each particular group share a strong degree of similarity with their groups’ unique utility function. Conversely, a heterogeneous condition can represent a “group parts” view, where agents have different utility functions within the group, and these utility functions are generally not similar. As such, within each group unique and/or diverse utility functions prevail, but across groups, this pattern is not unique, as group after group exhibits this same type of uniqueness among its members. We recognize that a homogeneous group with no group-level bias would be unlikely in actual groups and collectives. Reduction of a group-level bias would be facilitated by different perspectives, expertise and experiences (i.e., diversity). While varying diversity on any number of dimensions (e.g., ethnic, gender, functional background, education, age) within teams has been studied in the literature (c.f., Kooij-de Bode, van Knippenberg and van Ginkel 2008; Williams and O’Reilly 1998; O’Reilly et al. 1998; Pelled, Eisenhardt and Xin 1999), research related to group performance has mixed reviews regarding the benefit of diversity within teams. While some diversity is thought to produce a more productive, functional conflict as opposed to an unproductive, relationship conflict (Jehn, Northcraft and Neale 1999), a meta-analysis on conflict (De Dreu and Weingart 2003) underscores that these various forms of conflict are all negatively related to group performance. Thus, group-level bias is included to assess potential issues associated with homogeneity within groups. 4.1.3 Evolutionary operators Our model uses agent behaviors reflecting either selection or variation as analogues for decision making behavior: replication, random point mutation, intelligent point mutation, recombination, and subtractive selection. While these five operators reflect common forms of action in evolution (Mitchell 1996; Fogel 1995), they also align with actions commonly found in brainstorming and normative decision making idea generation phases where the goal is to build new ideas from individually generated suggestions (Paulus and Yang 2000) (i.e. mutations and recombination) and idea evaluation phases where culling or supporting ideas (i.e., replication and/or subtraction) leads to final group idea selection and decision. Among those evolutionary operators, replication and subtractive selection use a preferential random search algorithm (Solis and Wets 1981), where rp ideas are randomly sampled from the idea population in the agent’s mind and ranked according to their perceived utility values, and then the agent selects the best (or worst) idea for replication (or subtractive selection). Note that the designs of the evolutionary operators used in this model are different from those used in earlier models (Sayama and Dionne 2013a, Sayama and Dionne 2013b), in order to make the variation and selection mechanisms Group Decision Simulation 16 more clearly separable. They are also extended so that their outcomes affect not only the agent’s own idea population but also those of its local neighbors on a social network, which represents the exchange of ideas through social ties. In other words, other agents can “hear” the focal agent’s opinion and update their own idea population according to it. Replication selects an idea from the agent’s idea population with the above-mentioned preferential random search algorithm, and then adds an exact copy of the idea back to its own idea population, as well as to those of its local neighbors (i.e., other agents that are connected to the agent executing replication). Replication therefore can neither produce a novel idea nor remove one, but it gently sways the ecology of the population by increasing the popularity of favorable existent ideas within a local neighborhood in the social network. This represents an advocacy of a particular idea under discussion. Random point mutation selects an idea from the population randomly, and then adds a copy of an idea with point mutations, flipping of bits at each aspect of a problem with a probability pm, to the agent’s idea population as well as to those of its neighbors. This represents an attempt of making random changes to the existing ideas, reflected in asking “what if” questions. Random point mutations help escape local maxima of a utility function in the problem space when a utility function is nonlinear and many-peaked. Intelligent point mutation selects an idea from the population randomly, makes several (rm) tentative offspring of the idea by adding random point mutations, and selects that of the highest perceived fitness for addition to the population. The selected idea is added to the agent’s idea population as well as to those of its neighbors. This represents a proposal of an improved idea derived from an existing idea under discussion. The intelligent point mutation can be useful in maximizing a utility function with one maximum by climbing monotone gradients, but it may perform poorly in a complex utility landscape. Recombination chooses two ideas from the agent’s idea population at random and then creates two offspring from the two parent ideas. Sexual reproduction is simulated with a multiple point cross-over recombination: parent ideas are aligned by aspects, for each of which there is a probability ps of switching their contents. Both of the two offspring are then added back to the agent’s idea population as well as to those of its neighbors. This represents a creation of new ideas from two existing ideas. Finally, subtractive selection uses a negative preferential random search algorithm to find the idea with low fitness from the agent’s idea population, and then one copy of that idea is deleted from there, as well as from the idea populations of its neighbors (if a neighbor agent does not have a copy of the idea nothing will happen). This operator modestly reduces the popularity or importance of the idea within the local neighborhood in the social network. This represents a criticism against a bad idea. Subtractive selection is the only operator that reduces the number of existing ideas and is therefore essential to groups attempting to attain convergence in the population distribution. Group Decision Simulation 17 Out of these five, replication and subtractive selection are selection-oriented operators, driving the exploitation in the discussion and decision making process. The other three (random/intelligent point mutations and recombination) are variation-oriented operators that increase the idea diversity and explore the problem space further. To systematically control and sweep the balance between the two evolutionary “forces” (selection/exploitation and variation/exploration), we introduced a global parameter p, which determines the behavioral tendency of agents. Specifically, each agent chooses an exploitative operator with probability p (or, an explorative operator with probability 1 - p; see Table 2). Setting p = 1 makes the agents completely selection-oriented, while p = 0 makes them fully exploratory. 4.1.4 Simulation settings Table 2 summarizes the parameter values used in our computer simulations. Most of those values were taken from earlier work (Sayama and Dionne 2013a, Sayama and Dionne 2013b), and were chosen so as to be reasonable in view of typical real collective decision making settings. We tested several variations for each parameter value, confirming that the results were not qualitatively different from the ones presented below. There are several experimental parameters that we varied in the three sets of experiments presented below. The first set of experiments manipulated β, group-level bias, and ξ, within-group noise. These two parameters were varied to represent different levels of accuracy and consistency of individual utility functions within a group. The second set of experiments varied p, the parameter that determines the balance between selection-oriented and variation-oriented operators in agents’ behaviors. The third set of experiments varied the size and topology of the group, by exponentially increasing the number of agents from N = 5, a small group whose size is within the optimal range for decision making teams (Kerr and Tindale, 2004; Salas, Stagl and Burke 2004), to N = 640, which forms a non-trivial social network. In all cases, the average node degree (i.e., average number of connections attached to a node) was always kept to four, which is a typical number of people one could have meaningful conversations with simultaneously. This assumption made the N = 5 case a fully connected network, while the network became increasingly sparse as N increased. For each specific value of N, three different network topologies were tested: random (RD), small-world (SW; Watts and Strogatz 1998) and scale-free (SF; Barabási and Albert 1999). For small-world networks, the link rewiring probability was set to 10%, which realizes the small-world property (Watts and Strogatz 1998) for relatively small-sized networks like those used in this study. These topological variations do not cause any effective differences for smaller N, but as N increases, their influences on network topology and dynamics of idea evolution begin to differentiate. -----------------------------Insert Table 2 Here Group Decision Simulation 18 -----------------------------4.2 Metrics of group performance Performance of a group is likely a multidimensional construct, as different authors have tested differing dimensions of group-based adaptation (c.f., LePine 2005; Kozlowski et al. 1999). For the purposes of collective decision making in organizational settings, the ability to converge on a decision is critical, as a group that cannot produce a decision likely fails in their task. In the meantime, convergence on a poor decision may be equally detrimental to a group as well, as mistakes could be costly. As such, it would seem that minimally the consideration of both convergence and decision quality would be needed to assess group performance. As required by increasingly complex organizational environments, groups and organizations need to converge quickly on decisions, and yet ensure these decisions have high efficacy related to solving perceived problems. We therefore used two separate performance metrics: One was the true utility value of the mode idea (the most supported idea) in the final population of ideas collected from all the agents’ minds, to measure the overall quality of collective decisions. This was selected as it is most likely that the most supported idea represents the group’s preferred idea, and once selected, this supported idea will be tested in the context of real-world problem solving. The other performance metric was the diversity of ideas remaining in the final population of ideas collected from all the agents’ minds, to measure the failure of the group to converge. This measurement is based on the classical definition of Shannon’s information entropy (Shannon 1948), m H = −∑ p ( xi ) log 2 p ( xi ) , (3) i =1 where m represents the number of different types of ideas in the final idea population, and p(xi) is the ratio of the number of the i-th type of idea to the total size of the final idea population. The theoretical maximum of H would be M, which occurs when all of 2M possible distinct ideas are equally represented. Since the entropy represents how many more bits would be needed to completely specify the single final collective decision, it can be assumed that M – H is a quantitative measure that intuitively means the number of aspects of the problem on which the group has formed a cohesive opinion. To rescale this to the range between 0 and 1, we used (M – H) / M as a measurement of the convergence of final collective decision. 5 Results 5.1 Effects of within-group noise and group-level bias We first conducted a computational experiment to examine the effects of increasing (a) withingroup noise, ξ, i.e., heterogeneity of individual utility functions within a group, and (b) group-level bias, β, i.e., discrepancy of the master utility function from the true utility function at a group level, on the Group Decision Simulation 19 overall group performance. For this experiment, the group was made of five agents with fully connected social network structure (i.e., everyone could talk to everyone else; a small group setting). We assumed that the agents were balanced in terms of their tendency between selection-oriented and variation-oriented behaviors in the discussion (i.e., p = 1/2). Figure 2 presents a summary of the results of simulations with within-group noise ξ and grouplevel bias β systematically varied. Each of the two performance metrics (i.e., level of convergence and utility of most supported idea, as described above) are visualized in a separate 3-D surface plot. We found that the level of convergence affected significantly by the within-group noise, while it was not affected at all by the group-level bias. On the other hand, the true utility of collective decisions degraded significantly when either the within-group noise or the group-level bias (or both) was increased. The true utility achieved by the most heterogeneous groups (ξ ~ 1.0) or the most biased groups (β ~ 1.0) dropped to about 0.5, which could be achieved just by random idea generation, meaning that there was no net improvement achieved during the discussion by those groups. This could be due to the within-group conflicts of interest (for greater ξ) and/or the lack of correct group-level understanding of the problem (for greater β). In real-world settings, the within-group noise and the group-level bias are not independent from each other. More specifically, if a group is assembled by gathering similar individuals with similar backgrounds, expertise and opinions, then the group tends to have less within-group conflicts but may risk of having a greater group-level bias. On the other hand, if a group is made of diverse individuals with different backgrounds, expertise and opinions, the group may have greater within-group conflicts but it may successfully reduce potential group-level bias and accomplish deeper discussion and better integration of ideas, as the diverse perspectives may represent the true nature of the problem more correctly (we confirmed this observation by numerical tests; to conserve space, results are not shown here). Therefore, there is a trade-off between minimizing the within-group heterogeneity and the grouplevel bias in a realistic setting. What kind of strategies of group formation will be optimal in order to maximize the true utility of collective decisions remains a non-trivial and problem-dependent question. -----------------------------Insert Figure 2 Here -----------------------------5.2 Effects of balance between selection-oriented and variation-oriented behaviors The “balanced” acts of agents assumed in the above experiment may be too ideal as a model of actual group members, because actual groups may have biased behavioral patterns as well. For example, some groups may be more prone to be critical, trying to purge bad ideas, while other groups may tend to promote combinations of multiple ideas in discussion. Examples of such behavioral patterns include Group Decision Simulation 20 organizational “cultures” shared by all group members, which is a plausible view of a factor that may influence group dynamics (Salas et al. 2004). We therefore ran another experiment to investigate the effects of balance between selectionoriented and variation-oriented behaviors patterns by systematically varying the parameter p. Greater values of p represent groups with more selection-oriented behaviors (i.e., advocacy and criticism), while smaller values of p represent groups with more variation-oriented behaviors (i.e., mutations and recombination). The group-level bias, β, was also varied as another experimental parameter, while the within-group noise, ξ, was fixed to 0.2 for this experiment. The group size and their network topology were the same as those in the first experiment. Figure 3 shows a summary of the results of the second experiment comparing group performances with different group behaviors, plotting two performance metrics in separate 3-D plots as used for Figure 2 (note that one of the axes is now for p, not for ξ). The effect of behavioral balance on the level of convergence is straightforward in that greater p (more selection-oriented behaviors) tended to promote convergence more. The effect of p on the utility of collective decisions, however, turned out not so trivial. While purely variation-oriented behaviors (p ~ 0.0) did not help increase the decision quality, neither purely selection-oriented behaviors (p ~ 1.0) did. There was a range of optimal balance (p = 0.7~0.9) where the groups achieved the highest decision quality. In the meantime, the effect of group-level bias is similar to that seen in Figure 2, so that the utility of collective decisions would be significantly lower if there was group-level bias. -----------------------------Insert Figure 3 Here -----------------------------5.3 Effects of group size and social network topology The first two experiments above assumed small, fully connected networks of agents. While their results produced useful implications for collective decision making in small group settings, they were not sufficient to generate insight into more general collective decision making dynamics on a larger nontrivial social environment, such as in a complex organization or on social media. We therefore conducted the third experiment in which the size of groups was increased from 5 to 640 in an exponential manner. For each size of the groups/networks, the average number of connections per agent (i.e., “degree” in network science terminology) were always kept to four, which was the same value as in the first two experiments above. The following values were used for other parameters: β = 0.0, ξ = 0.2, p = 0.5. In this experiment, larger groups were no longer considered a typical “group”, but rather they formed a more complex social/organizational network, perhaps more indicative of a “collective” in the Group Decision Simulation 21 organizational sciences. For each network size, we used the following three social network topologies. A new network topology was generated for each independent simulation run: • Random network (RD): A random network is a network in which connections are randomly assigned, which can be used as a random control condition. For our experiment, a total of 2N links were established between randomly selected pairs of agents. • Small-world network (SW; Watts and Strogatz 1998): A small-world network is a locally clustered (pseudo-)regular network, with a small number of global links introduced to reduce the effective diameter of the network significantly (i.e., a “small-world” effect). The small-world network may be considered a spatially extended network made of mostly local connections but with a few global connections. For our experiment, N agents were first arranged in a circle and each agent was connected to its nearest and second nearest neighbors so that the degree would be four for all. Then 10% of the links were randomly selected and either the origin or destination of each of those links was rewired to a randomly selected agent. • Scale-free network (SF; Barabási and Albert 1999): A scale-free network is a network in which the distribution of node degrees shows a power-law distribution. It represents a heterogeneous network made of a large number of poorly connected nodes and a few heavily connected “hubs”. Many real-world networks, including biological, engineered and social networks, were shown to be scale-free (Barabási 2009). While such networks show a small effective diameter like smallworld networks, they may not have high local clustering. For our experiment, a well-known preferential attachment algorithm (Barabási and Albert 1999) was used, starting with a fully connected network of five agents and then incrementally adding an agent by connecting it with two links to two existing agents selected preferentially based on their degrees, until the network size reached N. Figure 4 shows the effects of size and topology of networks on the decision outcomes. Without a surprise, the larger the group (or network) becomes, the harder it achieves convergence. Apparently there was no substantial difference between the three topological structures regarding their effects on the level of convergence. On the other hand, increasing group size had positive effects on the utility of the most supported idea within the group or on the social network. This can be understood in that, in a large network, agents can conduct different threads of discussions in parallel, which increases the chance for them to collectively find a better idea in the complex problem space. It is important for the agents to remain connected to each other so that the better ideas gradually spread over the network and widely accepted to become the more supported ideas. The same number of disconnected agents would not be able to achieve this kind of information aggregation and selection task. ------------------------------ Group Decision Simulation 22 Insert Figure 4 Here -----------------------------One particularly interesting phenomenon seen in Figure 4 is the difference in the utility of collective decisions between small-world networks and other two networks for larger N (N > 100). Figure 5 provides a more detailed view into this finding, showing the distributions of utilities of most supported ideas for 500 independent simulation runs for N = 640 under each of the three conditions. In each condition, the agents were able to find the truly best idea with utility 1.0 most of the time, but small-world networks facilitated such optimal decision making most frequently. The Mann-Whitney U test detected statistically significant differences between small-world and random (p < 0.003) as well as small-world and scale-free (p < 10-6) networks, while there was no significant difference between random and scalefree (p = 0.107) networks. The key distinctive feature of small-world networks that are not present in either random or scale-free networks is the local clustering. Such locally clustered social network structure helps agents in different regions in a network maintain their respective focus areas and engage in different local search, possibly enhancing the effective parallelism of collective decision making and therefore resulting in a greater number of successful decisions. In contrast, random and scale-free networks lack such local clustering, and the links in those networks are all “global”, mixing discussions prematurely and therefore reducing the effective parallelism of collective decision making. These observations have an interesting contrast with the fact that random and scale-free networks are highly efficient in information dissemination because of their global connectedness. Our results indicate that such efficiency of information dissemination may not necessarily imply the same for collective decision making. -----------------------------Insert Figure 5 Here -----------------------------6 Conclusions To improve our understanding of the dynamic nature of collective decision making, we developed an agent-based model and applied evolutionary operators as a means of illustrating how groups and collectives may move through a decision process based on ecologies of ideas over a social network habitat. Moreover, we considered various compositions of group members ranging from homogeneity to heterogeneity and examined the impact of team behaviors on the dynamic decision process as well. These explorations move toward a more realistic view of collective decision making within complex social systems, and answer calls (e.g., Meyer et al. 2005) for research that considers the impact of time and situations in flux, along with nonlinear, multi-level concepts incorporating evolutionary conceptual development. Group Decision Simulation 23 Our exploration revealed that the composition of the team has implications for decision making and likely considers the complex nature of asking several individuals to come together and agree on a direction that is best suited for the group/collective, rather than for each individual. Research on group diversity has found mixed results related to diversity and group performance issues such as creativity and decision effectiveness (De Dreu and West 2001; Jackson 1992; Nemeth 1986; 1992; Hoffman 1979; Gruenfeld et al. 1996; Harrison et al. 2002; Jehn and Mannix 2001; O’Reilly et al. 1989; Kraiger and Ford 1985), however our research indicates an important trade-off between reduction of within-group conflicts and mitigation of group-level bias. This means that the best team composition may depend greatly on specific problem settings. For example, if a team is tasked to work on a time-critical mission, then the convergence speed is key to their success and thus the emphasis should be placed more on the group homogeneity to avoid within-group conflicts. Or, if a team is formed to seek a truly high quality solution to a problem, then minimizing the possibility of group-level bias is critical for the team’s success, which may require increasing within-group diversity. Our results also implies that the balance between selection-oriented and variation-oriented behaviors may play an important role in collective decision making. While selection-oriented behaviors greatly promote convergence, they are not sufficient to achieve the highest possible utility. To improve the decision quality, the group also needs a good mixture of exploratory (variation-oriented) and exploitative (selection-oriented) behaviors. This also ties back to the diversity issue discussed above; a group may not necessarily benefit from diversity of individual problem understanding, but they can benefit from behavioral diversity of group members. In our simulations, the optimal balance between selection and variation was attained at p ~ 0.8 (i.e., 80% selection, 20% variation) but this particular balancing point may be problem dependent. Our results with social network structure illustrated intriguing effects of network topologies on decision quality, which was manifested particularly for larger networks. Small-world networks with spatially localized clusters tended to promote collective search of optimal ideas more often than random or scale-free networks, despite that the network size and the average degree were all identical. This finding offers another implication for the diversity in collective decision making: certain organizational structures may be more effective in generating and maintaining idea diversity in discussion, while other structures would tend to reduce idea diversity and promote premature convergence on suboptimal ideas more often. This is similar to the biological fact that certain geographical habitat structures can maintain greater biodiversity in evolutionary ecology. In the decision making context, this implies that not only within-group diversity or behavioral balance but also social network topologies could influence the dynamics of idea evolution in collective decision making processes. Group Decision Simulation 24 Using an evolutionary framework to model collective decision making processes, one can specifically examine the efficacy of a variety of decision processes employed by groups and collectives. The framework enables a means for direct comparison of various idea evolution paths within collective decision making, and enables an exploration of how the make-up and structure of teams could be critical depending on the overall requirements for decision making tasks. Furthermore, the evolutionary framework and subsequent computational model enables advancements in understanding collective decision making within a dynamic and complex social system. By employing an evolutionary framework we can explore the impact of time and situations in flux, and the modeling enables nonlinear exploration of processes. Finally, the multi-level, network-oriented nature of the research more appropriately models the potential differences in team composition and organizational topologies. This research adds to our understanding of the complex nature of collective decisions, and the potential pitfalls and caveats of employing various decision processes and designing teams in a heterogeneous and/or homogeneous manner. There are several limitations to our modeling study. For example, genetic operators may not exist in groups as “cleanly” as modeled in our simulation. We used simple parameterized settings to control the prevalence of operators, which may not be appropriate to represent the real human individual behavior in discussion. Also, our model considered only the heterogeneity of the utility functions of agents. In real human systems, the heterogeneity, or diversity, implies far richer concepts (Dionne et al. 2004), such as different backgrounds, professional expertise, and behavioral strategies, which were not present in our current model. To conduct a more comprehensive, systematic investigation of the homogeneity/heterogeneity issues, it would be critical to incorporate the heterogeneity of the participants’ domains of expertise, in addition to their utility functions. Having team members with diverse domains of expertise may further improve group performance. Furthermore, we tested only three typical social network topologies, but they are by no means an exhaustive list of possible organizational structures. Conducting experiments on more realistic social network topologies would add more realistic dynamics to the results. Computer simulations provide a mechanism to study complex, dynamic collective processes with relatively little cost to researchers and “subjects”, yet no level of model complexity could adequately capture the complexity of navigating decision making in real group environments (Adner et al. 2009). The interactive effects of levels of analysis, personality, knowledge, learning, group dynamics and the environment will likely remain the most complex of models. As such, to refine theory and direct and pinpoint empirical research, simulation is an excellent “first responder.” To offer practical ideas to organizations, group members and leaders, however, research needs to be conducted in such a way as to capture and analyze the actions of real groups, and assess theoretical hypotheses against empirical Group Decision Simulation 25 findings. Our simulations findings here could serve as a basis or direction for future experimental and field studies of decision making in various types of heterogeneous and homogeneous real-world groups and collectives. Acknowledgments: Research supported by the National Science Foundation (SES-0826711) Group Decision Simulation 26 References Abbot A (2001) Time Matters. University of Chicago Press, Chicago, Il Adner R, Polos L, Ryall M, Sorenson O (2009) Introduction to special topic forum: The case for formal theory. Acad Management Rev 34: 201-208 Axelrod R (1981) The evolution of cooperation. Sci 211: 1390-1396 Barabási AL, Albert R (1999) Emergence of scaling in random networks. Science 286: 509-512. Barabási AL (2009) Scale-free networks: a decade and beyond. Science 325: 412-413. Bar-Yam Y (1997) Dynamics of complex systems. Westview Press, Boulder, CO Bar-Yam Y (2004) Making things work: Solving complex problems in a complex world. NESCI Knowledge Press, Cambridge, MA Battiston S, Bonabeau E, Weisbuch G (2003a) Decision making dynamics in corporate boards. Physica A 322: 567-582 Battiston S, Weisbuch G, Bonabeau E (2003b) Decision spread in the corporate board network. Adv Complex Systems 6: 631-644 Carroll T, Burton RM (2000) Organizations and complexity: Searching for the edge of chaos. Comput Math Organ Theory 6: 319-337 Chang MH, Harrington JE Jr (2007) Discovery and diffusion of knowledge in an endogenous social network. Amer J Sociol 110(4): 937-976 Chang MH, Harrington JE Jr (2005) Innovators, imitators and the evolving architecture of problemsolving networks. Organ Sci 18(4): 648-666 Dansereau F, Alutto JA,Yammarino FJ (1984) Theory Testing in Organizational Behavior: The Varient Approach. Prentice-Hall, Englewood Cliffs, NJ Group Decision Simulation 27 Dansereau F,Yammarino FJ, Kohles J (1999) Multiple levels of analysis from a longitudinal perspective: Some implications for theory building. Acad Management Rev 24: 346-357 De Dreu CKW, Weingart LR (2003) Task and relationship conflict, team performance, and team member satisfaction: A meta-analysis. J Appl Psych 88: 741–749 De Dreu CKW, West MA (2001) Minority dissent and team innovation: The importance of participation in decision making. J Appl Psych 86: 1191-1201 Dionne SD, Akaishi J, Chen X, Gupta A, Sayama H, Yammarino FJ, Serban A, Hao C, Head HJ, Bush BJ (2012) Retrospective Relatedness Reconstruction: Applications to Adaptive Social Networks and Social Sentiment. Organ Res Methods 15(4): 663-692 Dionne SD, Dionne PJ (2008) Levels-based leadership and hierarchical group decision optimization: A simulation. Leadership Quart 19: 212-234 Epstein JM (2006) Generative Social Science: Studies in Agent-based Computational Modeling. Princeton University Press Epstein JM, Axtell R (1996) Growing Artificial Societies: Social Science from the Bottom Up. Brookings Institute Press, Washington, DC. Fogel DB (1995) Evolutionary Computation: Toward a New Philosophy of Machine Intelligence. IEEE Press, Piscataway, NJ Futuyma DJ (2005) Evolution. Sinauer Associates, Sunderland, MA Gavetti G, Daniel A Levinthal DA (2000) Looking Forward and Looking Backward: Cognitive and Experiential Search, Administrative Sci Quart 45: 113 – 137 Gilbert N, Troitzsch KG (1999) Simulation for the social scientist. Milton Keynes: Open University Press Group Decision Simulation 28 Gigone D, Hastie R (1993) The common knowledge effect: Information sharing and group judgment. J Personality Soc Psych 65: 959-974 Gruenfeld DH, Mannix EA, Williams KY, Neale MA (1996) Group composition and decision making: How member familiarity and information distribution affect process and performance. Organ Behavior Human Decision Processes 67: 1-15 Hamming RW (1950) Error detecting and error correcting codes. Bell System Tech J 26: 147-160 Harrison DA, Price KH, Gavin JH, Florey AT (2002) Time, teams, and task performance: Changing effects of surface-level and deep-level diversity on group functioning. Acad Management J 45: 1029-1045 Hoffman LR (1979) Applying experimental research on group problem solving to organizations. J Appl Behavioral Sci 15: 375-391 Hollenbeck JR, Ilgen DR, Sego DJ, Hedlund J, Major DA, Phillips J (1995) Multilevel theory of team decision making: Decision performance in teams incorporating distributed expertise. J Appl Psych 80(2): 292-316 Horst R, Pardalos PM, Thoai NV (2000) Introduction to Global Optimization, 2nd edn. Kluwer Academic Publishers, Dordrecht, Netherlands Jackson S (1992) Team composition in organizations. In: Worchel S, Wood W, Simpson J (eds) Group Process and Productivity. Sage Publications, London Jehn KA, Mannix EA (2001) The dynamic nature of conflict: A longitudinal study of intragroup conflict and group performance. Acad Management J 44: 238-251 Jehn KA, Northcraft GB, Neale MA (1999) Why differences make a difference: A field study of diversity, conflict, and performance in workgroups. Admin Sci Quart 44: 741–763 Kauffman SA (1993) The origin of order. Oxford University Press, NY Kerr NL, Tindale RS (2004) Group performance and decision making. Ann Rev Psych 55: 623-655 Group Decision Simulation 29 Klein M, Sayama H, Faratin P, Bar-Yam Y (2003a) The dynamics of collaborative design: Insights from complex systems and negotiations research. Concurrent Engrg: Res Appl 11: 201-209 Klein M, Faratin P, Sayama H, Bar-Yam Y (2003b) Protocols for negotiating complex contracts. IEEE Intelligent Systems 18(6): 32-38 Klein M, Faratin P, Sayama H, Bar-Yam Y (2006) An annealing protocol for negotiating complex contracts. In: Rennard J-P(ed) Handbook of Research on Nature Inspired Computing for Economics and Management, vol 2. Idea Group Publishing, Hershey, PA, Chapter XCVIII Klein KJ, Dansereau F, Hall RJ (1994) Levels issues in theory development, data collection, and analysis. Acad Management Rev 19: 195-229 Knudsen T, Levinthal DA (2007) Two faces of search: Alternative generation and alternative evaluation. Organ Sci 18(1): 39-54 Kock N (1999) Process Improvement and Organizational Learning: The Role of Collaboration Technologies. Idea Group Publishing, Hershey, PA Kock N (2004) The psychobiological model: Towards a new theory of computer-mediated communication based on Darwinian evolution. Organ Sci 15(3): 327-348 Kooij-de Bode HJM, van Knippenberg D, van Ginkel WP (2008) Ethnic diversity and distributed information in group decision making: The importance of information elaboration. Gr Dyn: Theory Res Pract 12: 307–320 Kozlowski SWJ, Gully SM, Nason ER, Smith EM (1999) Developing adaptive teams: A theory of compilation and performance across levels and time. In: Ilgen DR, Pulakos ED (eds) The Changing Nature of Performance: Implications for Staffing, Motivation and Development. Jossey-Bass, San Francisco, CA, pp 240-292 Kraiger K, Ford JK (1985) A meta-analysis of ratee-race effects in performance ratings. J Appl Psych 70: 56-65 Group Decision Simulation 30 LePine JA (2005) Adaptation of teams in response to unforeseen change: Effects of goal difficulty and team composition in terms of cognitive ability and goal orientation. J Appl Psych 90: 1153-1167 Levinthal DA (1997) Adaptation on rugged landscapes. Management Sci 43(7): 934-950 Lightle JP, Kagel JH, Arkes HR (2009) Information exchange in group decision making: The hidden profile problem reconsidered. Management Sci 55(4): 568-581 Lipshitz R, Klein G, Orasanu J, Salas E (2001) Focus Article: Taking stock of naturalistic decision making. J Behavioral Decision Making 14: 331-352 Maldonato M (2007) Undecidable decisions: Rationality limits and decision-making heuristics. World Futures: J General Evolution 63(1): 28-37 Mannes AE (2009) Are we wise about the wisdom of crowds? The use of group judgments in belief revision. Management Sci 55(8): 1267-1279 March JG (1991) Exploration and exploitation in organizational learning. Org Sci 2: 71-87. Meyer AD, Gaba V, Colwell KA (2005) Organizing far from equlibrium: Nonlinear change in organizational fields. Organ Sci 16(5): 456-473 Miller JH, Page SE (2007) Complex Adaptive Systems: An Introduction to Computational Models of Social Life. Princeton University Press, Princeton, NJ Mitchell M (1996) An Introduction to Genetic Algorithms. MIT Press, Cambridge, MA Mitchell M, Forrest S, Holland JH (1991) The royal road for genetic algorithms: Fitness landscapes and GA performance. In: Varela FJ, Bourgine P (eds) Toward a Practice of Autonomous Systems: Proceeds of the First European Conference on Artificial Life. Paris, France, pp 245-254 Nemeth CJ (1986) Differential contributions of majority and minority influence. Psych Rev 93: 23-32 Group Decision Simulation 31 Nemeth CJ (1992) Minority dissent as a stimulant to group performance. In: Worchel S, Wood W, Simpson J (eds) Group Process and Productivity. Sage Publications, London, pp 95-111 Nijstad BA, Kaps SC (2007) Taking the easy way out: Preference diversity decision strategies and decision refusals in groups. J Personality Soc Psych 94(5): 860-870 O’Reilly CA, Caldwell DF, Barnett WP (1989) Work group demography, social integration and turnover. Admin Sci Quart 34: 21-37 O’Reilly CA, Williams KY, Barsade S, Gruenfeld DH (1998) Group Demography and Innovation: Does Diversity Help? In M. A. Neale & E. A. Mannix (Ed), Research on Managing Groups and Teams Vol. 1: 183-207 O’Reilly T (2005) What is Web 2.0: Design patterns and business models for the next generation of software. IOP O’Reilly. http://www.oreillynet.com/pub/a/oreilly/tim/news/2005/09/30/what-is-web20.html. Accessed October 8, 2012 Paulus PB, Yang H-C (2000) Idea generation in groups: A basis for creativity in organizations. Organ Behavior Human Decision Processes 82: 76-87 Pelled LH, Eisenhardt KM, Xin KR (1999) Exploring the black box: An analysis of work group diversity, conflict and performance. Admin Sci Quart 44:1-28 Ren Y, Carley KM, Argote L (2006) The contingent effects of transactive memory: When is more beneficial to know what others know? Management Sci 52: 671-682 Rivkin JW (2000) Imitation of complex strategies. Management Sci 46(6): 824-844 Rodriguez, MA, Steinbock DJ (2004) A social network for societal-scale decision making systems. In NAACSOS 2004: Proc North Amer Association Comput Soc Organ Sci Conf, Pittsburgh, PA Rusmevichientong P, Van Roy B (2003) Decentralizing decision making in a large team with local information. Games Econom Behavior 43: 266-295 Group Decision Simulation 32 Salas E, Klein G (2001) Expertise and naturalistic decision making: An overview. In: E. Salas E, Klein G (eds) Linking Expertise and Naturalistic Decision Making. Lawrence Erlbaum Associates, Philadelphia, PA, pp 3-33 Salas E, Stagl KC, Burke CS (2004) 25 years of team effectiveness in organizations: Research themes and emerging needs. In: Cooper CL, Robertson IT (eds) International Review of Industrial and Organizational Psychology 19: 47-91 Sayama H, Dionne SD (2013a) Using evolutionary computation as models/tools for human decision making and creativity research. Proc 4th IEEE Symp Artif Life, pp.35-42. Sayama H, Dionne SD (2013b) Studying collective human decision making and creativity with evolutionary computation. Under review. Sayama H, Farrell DL, Dionne SD (2011) The effects of mental model formation on group decision making: An agent-based simulation. Complexity 16(3): 49-57. Shannon CE (1948) A mathematical theory of communication. Bell System Tech J 27: 379-423 Simon HA (1955) A behavioral model of rational choice. Amer Econom. Rev 69: 99-118 Silverman BG, Johns M, Corwell J, O’Brien K (2006) Human behavior models for agents in simulation games: Part 1: Enabling science with PMFserv. Presence 15(2): 139-162 Solis FJ, Wets, RJ-B (1981) Minimization by random search techniques. Math Oper Res 6: 19-30 Stasser G, Stewart D (1992) Discovery of hidden profiles by decision making groups: Solving a problem versus making a judgment. J Personality Soc Psych 57: 67-78 Stasser G, Titus W (1985) Pooling of unshared information in group decision making: Biased information sampling during discussion. J Personality Soc Psych 48: 1467-1478 Sterman JD (2000) Business dynamics: System thinking and modeling for a complex world. Irwin McGraw Hill, Boston, MA Group Decision Simulation 33 Van de Ven A, Delbecq AL (1974) The effectiveness of nominal, delphi, and interacting group decision making processes. Acad Management J 117: 605-621 van Ginkel WP, van Knippenberg D (2008) Group information elaboration and decision making: The role of shared task representations. Org Behavior Human Decision Processes 105(1): 89-97 Watts DJ, Strogatz SH (1998) Collective dynamics of ‘small-world’ networks. Nature 393: 440-442 Williams KY, O’Reilly CA (1998) Demography and diversity in organizations. Res Organ Behavior, 20: 77–140 Wilson DS (2005) Evolution for everyone: How to increase acceptance of interest in, and knowledge about evolution. Public Library Sci Biol 3(12): e364 Wilson DS, Wilson EO (2008) Evolution for the “good of the group.” Amer Sci 96: 380-389 Yamanoi J, Sayama H (2012) Post-merger cultural integration from a social network perspective: A computational modeling approach. Comput Math Org Theory 18: 1-22 Yammarino FJ, Dansereau F (eds) 2002 The many faces of multi-level issues. Research in Multi-Level Issues, vol 1. Elsevier Science,Oxford, UK Yammarino FJ, Dansereau F (2011) Multi-level issues in evolutionary theory, organization science, and leadership. Leadership Quart 22: 1042-1057 Yammarino FJ, Dionne SD, Chun JU, Dansereau F (2005) Leadership and levels of analysis: A state-ofthe-science review. Leadership Quart 16: 879-919 Group Decision Simulation 34 TABLE 1 Evolutionary Concepts Applied to Corresponding Decision Making Process Components Evolutionary Concept Decision Making Component Genetic possibility space Problem space (decision space) Genome Potential idea (a set of choices for all aspects of the problem) Locus on a genome Aspect of the problem Allele (specific gene) on a locus Specific choice made for an aspect Population A set of potential ideas being discussed Fitness Utility value of a potential idea (either perceived or real) Adaptation Increase of utility values achieved by an idea population Selection Narrowing of diversity of ideas based on their fitness Replication Increase of relative popularity of a potential idea in the discussion Recombination Production of a new potential idea by crossing multiple ideas Mutation Point-like change in an idea (possibly coming up with a novel idea) Group Decision Simulation 35 TABLE 2 Parameters and Symbols (Bold indicates experimental parameters varied) Parameter Value Meaning Parameters Related to Evolutionary Decision Process M 10 Problem space dimensionality n 10 Number of representative ideas to generate true/master utility functions rp 5 Number of sample ideas in preferential search algorithm rm 5 Number of offspring generated in intelligent point mutation pm 0.2 Random mutation rate per bit ps 0.4 Probability of random switching in recombination Probability for an agent to take selection-oriented actions p 0~1 p/2 Probability of replication - advocacy (increases popularity; selectionoriented) p/2 Probability of subtractive selection - criticism (decreases popularity; selection-oriented) (1-p)/3 Probability of random point mutation - minor modification of idea (variation-oriented) (1-p)/3 Probability of intelligent point mutation - improvement of existing idea (variation-oriented) (1-p)/3 Probability of recombination - generating new ideas from crossing multiple existing ideas (variation-oriented) Parameters Related to Team Characteristics Size of group/social network N 5~640 Network RD, SW, RD: random network, SW: small-world network, SF: scale-free network topology SF d 4 Average degree (average number of links connected to each agent) k 5 Number of initial randomly generated ideas in each agent’s mind Group-level bias 0~1 β Within-group noise 0~1 ξ T 60 Number of iterations Group Decision Simulation 36 FIGURE 1 Master and Individual Utility Functions Note: The master utility function with M = 10, generated from a representative set of idea utilities of size n = 10, is shown by black dots. An individual utility function by adding noise with ξ = 0.2 is shown by gray dots. The x-axis shows idea indices generated by interpreting bit strings as binary notations of an integer, i.e., all of different ideas are lined up along the horizontal axis and their utility values are plotted. It can be seen in the figure that the perturbed individual utility function (gray dots) maintains some structures of the master utility function (black dots), but they are not exactly the same, which represents the misunderstanding of the problem by the individual. Group Decision Simulation 37 FIGURE 2 Effects of Within-Group Noise and Group-Level Bias on Decision Convergence and Decision Quality Note: These plots summarize simulation results showing the effects of within-group noise (ξ) and group-level bias (β) on the level of convergence (left) and the true utility value of the most supported idea (right). Each dot represents an average result of 500 independent simulation runs. Left: Decision convergence was high when within-group noise was small, i.e., when groups were homogeneous in terms of the agents’ individual utility functions. Group-level bias did not show any substantial effect on decision convergence. Right: Quality decisions were made only if both within-group noise and group-level bias were small. As either increases, the quality of the most supported idea decreases. Group Decision Simulation 38 FIGURE 3 Effects of Balance between Selection-Oriented and Variation-Oriented Behaviors and Group-Level Bias on Decision Convergence and Decision Quality Note: These plots summarize simulation results showing the effects of group-level balance between selectionoriented and variation-oriented behaviors (p) and group-level bias (β) on the level of convergence (left) and the true utility value of the most supported idea (right). Each dot represents an average result of 500 independent simulation runs. Left: Decision convergence was high when group-level behavior was highly selection-oriented (p ~ 1), i.e., when groups used replication (advocacy) and subtractive selection (criticism) most of time. Group-level bias did not show any substantial effect on decision convergence. Right: Quality decisions were made only if the group-level behavior was maintained at an optimal balance between selection-oriented and variation-oriented behaviors (p = 0.7~0.9) while group-level bias was small. Group Decision Simulation 39 FIGURE 4 Effects of Group Size and Social Network Topology on Decision Convergence and Decision 0.45  0.40 0.35 0.30 0.25 0.20 0.15 Random Smallworld Scalefree                10 20    50 100 200 Group size       500 Utility of most supported idea Level of convergence Quality 0.90 0.88 0.86 0.84 0.82 0.80 0.78 0.76               10          20   Random Smallworld Scalefree 50 100 200 Group size 500 Note: These plots summarize simulation results showing the effects of group size (N) and social network topology (random, small-world or scale-free) on the level of convergence (left) and the true utility value of the most supported idea (right). Note the log scale for group size. Each dot represents an average result of 500 independent simulation runs. Left: Decision convergence decreased as the group size increased. Social network topology did not influence decision convergence very much. Right: For smaller groups (N < 100), the topological difference between random, small-world and scale-free networks did not have substantial effects, but for larger groups (N > 100), the small-world topology outperformed the other two topologies in terms of the quality of the most supported idea. Group Decision Simulation 40 FIGURE 5 Comparison of Distributions of Utilities of Most Supported Ideas Between Different Social Network Topologies Note: These plots summarize simulation results comparing the distributions of utilities of most supported ideas at the end of simulation between the three social network topologies (random, small-world or scale-free) for N = 640. The small-world network topology (middle) achieved the highest number of the maximal utility value (1.0) compared to the other two topologies, random (left) and scale-free (right). The MannWhitney U test revealed that there was a statistically significant difference between small-world and the other two (small-world v. random: p < 0.003, small-world v. scale-free: p < 10-6), while there was no statistically significant difference between random and scale-free topologies (p = 0.107).
9cs.NE
Combinatorial Secretary Problems with Ordinal Information∗ Martin Hoefer† Bojana Kodric‡ arXiv:1702.01290v3 [cs.DS] 25 Aug 2017 Abstract The secretary problem is a classic model for online decision making. Recently, combinatorial extensions such as matroid or matching secretary problems have become an important tool to study algorithmic problems in dynamic markets. Here the decision maker must know the numerical value of each arriving element, which can be a demanding informational assumption. In this paper, we initiate the study of combinatorial secretary problems with ordinal information, in which the decision maker only needs to be aware of a preference order consistent with the values of arrived elements. The goal is to design online algorithms with small competitive ratios. For a variety of combinatorial problems, such as bipartite matching, general packing LPs, and independent set with bounded local independence number, we design new algorithms that obtain constant competitive ratios. For the matroid secretary problem, we observe that many existing algorithms for special matroid structures maintain their competitive ratios even in the ordinal model. In these cases, the restriction to ordinal information does not represent any additional obstacle. Moreover, we show that ordinal variants of the submodular matroid secretary problems can be solved using algorithms for the linear versions by extending [18]. √ In contrast, we provide a lower bound of Ω( n/(log n)) for algorithms that are oblivious to the matroid structure, where n is the total number of elements. This contrasts an upper bound of O(log n) in the cardinal model, and it shows that the technique of thresholding is not sufficient for good algorithms in the ordinal model. 1 Introduction The secretary problem is a classic approach to model online decision making under uncertain input. The interpretation is that a firm needs to hire a secretary. There are n candidates and they arrive sequentially in random order for an interview. Following an interview, the firm learns the value of the candidate, and it has to make an immediate decision about hiring him before seeing the next candidate(s). If the candidate is hired, the process is over. Otherwise, a rejected candidate cannot be hired at a later point in time. The optimal algorithm is a simple greedy rule that rejects all candidates in an initial learning phase. In the following acceptance phase, it hires the first candidate that is the best among all the ones seen so far. It manages to hire the best candidate with optimal probability 1/e. Notably, it only needs to know if a candidate is the best seen so far, but no exact numerical values. Since its introduction [15], the secretary problem has attracted a huge amount of research interest. Recently, a variety of combinatorial extensions have been studied in the computer science literature [7], capturing a variety of fundamental online allocation problems in networks and markets, such as network design [28], resource allocation [24], medium access in networks [21], or competitive admission processes [12]. Prominently, in the matroid secretary problem [8], the elements of a weighted matroid arrive in uniform random order (e.g., weighted edges of an undirected graph G). The goal is to select a max-weight independent set of the matroid (e.g., a max-weight forest of G). The popular matroid secretary conjecture claims that for all matroids, there exists ∗ This work was supported by DFG Cluster of Excellence MMCI at Saarland University. of Computer Science, Goethe University Frankfurt/Main, Germany. [email protected] ‡ Max Planck Institute for Informatics and Saarland University, Saarland Informatics Campus, Germany. [email protected] † Institute 1 an algorithm with a constant competitive ratio, i.e., the expected total weight of the solution computed by the algorithm is at least a constant fraction of the total weight of the optimum solution. Despite much progress on special cases, the conjecture remains open. Beyond matroids, online algorithms for a variety of combinatorial secretary problems with downward-closed structure have recently been studied (e.g., matching [24, 28], independent set [21], linear packing problems [25] or submodular versions [18, 26]). The best known algorithms for matroid or matching secretary problems rely heavily on knowing the exact weight structure of elements. They either compute max-weight solutions to guide the admission process or rely on advanced bucketing techniques to group elements based on their weight. For a decision maker, in many applications it can be quite difficult to determine an exact cardinal preference for each of the incoming candidates. In contrast, in the original problem, the optimal algorithm only needs ordinal information about the candidates. This property provides a much more robust guarantee, since the numerical values can be arbitrary, as long as they are consistent with the preference order. In this paper, we study algorithms for combinatorial secretary problems that rely only on ordinal information. We assume that there is an unknown value for each element, but our algorithms only have access to the total order of the elements arrived so far, which is consistent with their values. We term this the ordinal model ; as opposed to the cardinal model, in which the algorithm learns the exact values. We show bounds on the competitive ratio, i.e., we compare the quality of the computed solutions to the optima in terms of the exact underyling but unknown numerical values. Consequently, competitive ratios for our algorithms are robust guarantees against uncertainty in the input. Our approach follows a recent line of research by studying the potential of algorithms with ordinal information to approximate optima based on numerical values [1, 3, 4, 10]. 1.1 Our Contribution We first point out that many algorithms proposed in the literature continue to work in the ordinal model. In particular, a wide variety of algorithms for variants of the matroid secretary problem with constant competitive ratios continue to obtain their guarantees in the ordinal model (see Table 1 for an overview). This shows that many results in the literature are much stronger, since the algorithms require significantly less information. Notably, the algorithm of [9] extends to the ordinal model and gives a ratio of O(log2 r) for general matroids, where r is the rank of the matroid. In contrast, the improved algorithms with ratios of O(log r) and O(log log r) [8, 16, 29] are not applicable in the ordinal model. For several combinatorial secretary problems we obtain new algorithms for the ordinal model. For online bipartite matching we give an algorithm that is 2e-competitive. We also extend this result to online packing LPs with at most d non-zero entries per variable. Here we obtain an O(d(B+1)/B )-competitive algorithm, where B is a tightness parameter of the constraints. Another extension is matching in general graphs, for which we give a 8.78-competitive algorithm. We give an O(α21 )-competitive algorithm for the online weighted independent set problem in graphs, where α1 is the local independence number of the graph. For example, for the prominent case of unit-disk graphs, α1 = 5 and we obtain a constant-competitive algorithm. For matroids, we extend a result of [18] to the ordinal model: The reduction from submodular to linear matroid secretary can be done with ordinal information for marginal weights of the elements. More specifically, we show that whenever there is an algorithm that solves the matroid secretary problem in the ordinal model on some matroid class and has a competitive ratio of α, there is also an algorithm for the submodular matroid secretary problem in the ordinal model on the same matroid class with a competitive ratio of O(α2 ). The ratio can be shown to be better if the linear algorithm satisfies some further properties. Lastly, we consider the importance of knowing the weights, ordering, and structure of the domain. For algorithms that have complete√ ordinal information but cannot learn the specific matroid structure, we show a lower bound of Ω( n/(log n)), even for partition matroids, where n is the number of elements in the ground set. This bound contrasts the O(log2 r)-competitive algorithm and indicates that learning the matroid structure is crucial in the ordinal model. Moreover, it 2 Matroid general 2 Ratio O(log r) Reference [9] k-uniform p 1 + O( 1/k), e [6, 15, 27] graphic cographic transversal regular 16 laminar √ 3 3e 2e 3e [28] [30] [13] [23] [14] 9e Table 1: Existing algorithms for matroid secretary problems that provide the same guarantee in the ordinal model. contrasts the cardinal model, where thresholding algorithms yield O(log r)-competitive algorithms without learning the matroid structure. For structural reasons, we present our results in a slightly different order. We first discuss the matroid results in Section 3. Then we proceed with matching and packing in Section 4 and independent set in Section 5. All missing proofs are deferred to the Appendix. 2 Preliminaries and Related Work In the typical problem we study, there is a set E of elements arriving sequentially in random order. The algorithm knows n = |E| in advance. It must accept or reject an element before seeing the next element(s). There is a set S ⊆ 2E of feasible solutions. S is downward-closed, i.e., if S ∈ S, then S ′ ∈ S for every S ′ ⊆ S. The goal is to accept a feasible solution that maximizes an P objective function f . In the linear version, each element has a value or weight we , and f (S) = e∈S we . In the submodular version, f is submodular and f (∅) = 0. In the linear ordinal model, the algorithm only sees a strict total order over the elements seen so far that is consistent with their weights (ties are broken arbitrarily). For the submodular version, we interpret the value of an element as its marginal contribution to a set of elements. In this case, our algorithm has access to an ordinal oracle O(S). For every subset S of arrived elements, O(S) returns a total order of arrived elements consistent with their marginal values f (e|S) = f (S ∪ {e}) − f (S). Given this information, we strive to design algorithms that will have a small competitive ratio f (S ∗ )/E[f (Salg )]. Here S ∗ is an optimal feasible solution and Salg the solution returned by the algorithm. Note that Salg is a random variable due to random-order arrival and possible internal randomization of the algorithm. In the matroid secretary problem, the pair M = (E, S) is a matroid. We summarize in Table 1 some of the existing results for classes of the (linear) problem that transfer to the ordinal model. The algorithms for all restricted matroid classes other than the graphic matroid assume a-priori complete knowledge of the matroid – only weights are revealed online. The algorithms do not use cardinal information, their decisions are based only on ordinal information. As such, they translate directly to the ordinal model. Notably, the algorithm from [9] solves even the general submodular matroid secretary problem in the ordinal model. 2.1 Related Work Our work is partly inspired by [4, 5], who study ordinal approximation algorithms for classical optimization problems. They design constant-factor approximation algorithms for matching and clustering problems with ordinal information and extend the results to truthful mechanisms. Our approach here differs due to online arrival. Anshelevich et al. [3] examine the quality of randomized social choice mechanisms when agents have metric preferences but only ordinal information is available to the mechanism. Previously, [1,10] studied ordinal measures of efficiency in matchings, for instance the average rank of an agent’s partner. The literature on the secretary problem is too broad to survey here. We only discuss directly related work on online algorithms for combinatorial variants. Cardinal versions of these problems have many important applications in ad-auctions and item allocation in online markets [22]. For 3 multiple-choice secretary, where we can select any k candidates, there are algorithms with ratios that are constant and asymptotically decreasing in k [6, 27]. More generally, the matroid secretary problem has attracted a large amount of research interest [8, 11, 16, 29], and the best-known algorithm in the cardinal model has ratio O(log log r). For results on specific matroid classes, see the overview in Table 1. Extensions to the submodular version are treated in [9, 18]. Another prominent domain is online bipartite matching, in which one side of the graph is known in advance and the other arrives online in random order, each vertex revealing all incident weighted edges when it arrives [28]. In this case, there is an optimal algorithm with ratio e [24]. Moreover, our paper is related to Göbel et al. [21] who study secretary versions of maximum independent set in graphs with bounded inductive independence number ρ. They derive an O(ρ2 )-competitive algorithm for unweighted and an O(ρ2 log n)-competitive algorithm for weighted independent set. In addition, algorithms have been proposed for further variants of the secretary problem, e.g., the temp secretary problem (candidates hired for a fixed duration) [19], parallel secretary (candidates interviewed in parallel) [17], or local secretary (several firms and limited feedback) [12]. For these variants, some existing algorithms (e.g., for the temp secretary problem in [19]) directly extend to the ordinal model. In general, however, the restriction to ordinal information poses an interesting challenge for future work in these domains. 3 Matroids 3.1 Submodular Matroids We start our analysis by showing that – in addition to algorithms for special cases mentioned above – a powerful technique for submodular matroid secretary problems [18] can be adjusted to work even in the ordinal model. More formally, in this section we show that there is a reduction from submodular matroid secretary problems with ordinal information (SMSPO) to linear matroid secretary problems with ordinal information (MSPO). The reduction uses Greedy (Algorithm 1) as a subroutine and interprets the marginal value when added to the greedy solution as the value of an element. These values are then forwarded to whichever algorithm (termed Linear) that solves the linear version of the problem. In the ordinal model, we are unable to see the exact marginal values. Nevertheless, we manage to construct a suitable ordering for the forwarded elements. Consequently, we can apply algorithm Linear as a subroutine to obtain a good solution for the ordinal submodular problem. Let M = (E, S) be the matroid, f the submodular function, and E the ground set of elements. The marginal contribution of element u to set M is denoted by f (u|M ) = f (M ∪ {u}) − f (M ). Since f can be non-monotone, Greedy in the cardinal model also checks if the marginal value of the currently best element is positive. While we cannot explicitly make this check in the ordinal model, note that f (u|M ) ≥ 0 ⇐⇒ f (M ∪ {u}) ≥ f (M ) = f (M ∪ {u′ }) for every u′ ∈ M . Since the ordinal oracle includes the elements of M in the ordering of marginal values, there is a way to check positivity even in the ordinal model. Therefore, our results also apply to non-monotone functions f . A potential problem with Algorithm 2 is that we must compare marginal contributions of different elements w.r.t. different sets. We can resolve this issue by following the steps of the Algorithm 1: Greedy [18] Input : ground set E Output: independent set M 1 2 3 4 Let M ← ∅ and E ′ ← E. while E ′ 6= ∅ do Let u ← maxu′ f (u′ |M ) and E ′ ← E ′ \ {u}; if (M ∪ {u} independent in M) ∧ (f (u|M ) ≥ 0) then add u to M ; 4 Algorithm 2: Online(p) algorithm [18] Input : n = |E|, size of the ground set Output: independent set Q ∩ N 1 Choose X from the binomial distribution B(n, 1/2). 2 Reject the first X elements of the input. Let L be the set of these elements. 3 Let M be the output of Greedy on the set L. 4 Let N ← ∅. 5 for each element u ∈ E \ L do 6 Let w(u) ← 0. 7 if u accepted by Greedy applied to M ∪ {u} then 8 With probability p do the following: 9 Add u to N . 10 Let Mu ⊆ M be the solution of Greedy immediately before it adds u to it. 11 w(u) ← f (u|Mu ). Pass u to Linear with weight w(u). 12 13 return Q ∩ N , where Q is the output of Linear. Greedy subroutine that tries to add new elements to the greedy solution computed on the sample. We use this information to construct a correct ordering over the marginal contributions of elements that we forward to Linear. Lemma 1. Let us denote by su the step of Greedy in which the element u is accepted when applied to M + u. Then su1 < su2 implies f (u1 |Mu1 ) ≥ f (u2 |Mu2 ). Proof. First, note that Mu1 ⊂ Mu2 when s1 > s2 . We denote by mu1 the element of M that would be taken in step su1 if u1 would not be available. Then we know that f (u1 |Mu1 ) ≥ f (mu1 |Mu1 ). Furthermore, since s1 < s2 , f (mu1 |Mu1 ) ≥ f (u2 |Mu1 ). Lastly, by using submodularity, we know that f (u2 |Mu1 ) ≥ f (u2 |Mu2 ). When su1 = su2 , then Mu1 = Mu2 so the oracle provides the order of marginal values. Otherwise, the lemma yields the ordinal information. Thus, we can construct an ordering for the elements that are forwarded to Linear that is consistent with their marginal values in the cardinal model. Hence, the reduction can be applied in the ordinal model, and all results from [18] continue to hold. We mention only the main theorem. It implies constant ratios for all problems in Table 1 in the submodular version. Theorem 1. Given an arbitrary algorithm Linear for MSPO that is α-competitive on a matroid class, there is an algorithm for SMSPO with competitive ratio is at most 24α(3α + 1) = O(α2 ) on the same matroid class. For SMSPO with monotone f , it can be improved to 8α(α + 1). 3.2 A Lower Bound Another powerful technique in the cardinal model is thresholding, where we first sample a constant fraction of the elements to learn their weights. Based on the largest weight observed, we pick a threshold and accept subsequent elements greedily if they exceed the threshold. This approach generalizes the classic algorithm [15] and provides logarithmic ratios for many combinatorial domains [8, 12, 21, 28]. Intuitively, these algorithms learn the weights but not the structure. We show that this technique does not easily generalize to the ordinal model. The algorithms with small ratios in the ordinal model rely heavily on the matroid structure. Indeed, in the ordinal model we show a polynomial lower bound for algorithms in the matroid secretary problem that learn the ordering but not the structure. Formally, we slightly simplify the setting as follows. The algorithm receives the global ordering of all elements in advance. It determines (possibly at random) a threshold position in the ordering. Then elements arrive and are accepted greedily if 5 1 k 2k 3k n−k n ... 1 0 1 1 1 .. . 1 0 0 0 0 .. . .. . .. . 0 0 0 0 1 0 0 ... 1 1 k .. . 1 1 Figure 1: Values for the family of instances de- Figure 2: One instance from the family described in the proof of Theorem 11, where the scribed in the proof of Theorem 11. position of the “valuable” ones is denoted by the thick segment. ranked above the threshold. Note that the algorithm does not use sampling, since in this case the only meaningful purpose of sampling is learning the structure. We call this a structure-oblivious algorithm. Theorem 2. Every structure-oblivious randomized algorithm has a competitive ratio of at least √ Ω( n/(log n)). Proof. In the proof, we restrict our attention to instances with weights in {0, 1} (for a formal justification, see Lemma 7 in the Appendix). We give a distribution of such instances on which √ every deterministic algorithm has a competitive ratio of Ω( n/(log n)). Using Yao’s principle, this shows the claimed result for randomized algorithms. All instances in the distribution are based on a graphic matroid (in fact, a partition matroid) of the following form. There is a simple path of 1 + k segments. The edges in each segment have weight of 0 or 1. We call the edges with value 1 in the last k segments the “valuable edges”. The total number of edges is the same in each instance and equals n + 1. All edges in the first segment have value 1 and there is exactly one edge of value 1 in all other segments (that being the aforementioned valuable edges). In the first instance there are in total k + 1 edges of value 1 (meaning that there is only one edge in the first segment). In each of the following instances this number is increased by k (in the i-th instance there are (i−1)·k +1 edges in the first segment) such that the last instance has only edges with value 1 (there are n − k + 1 edges in the first segment). The zero edges are always equally distributed on the last k path segments. The valuable edges are lower in the ordering than any non-valuable edge with value 1 (see Figure 1). Each of the instances appears with equal probability of nk (see Figure 2 for one example instance). A deterministic algorithm picks a threshold at position i. The expected value of the solution is i k k2 i k2 n kX k ≤1+ log ≤ log + 1 , E[w(Salg )] ≤ 1 + n ℓ n k n k ℓ=1 where log denotes the natural logarithm and the expression results from observing that the algorithm cannot obtain more than a value of 1 if its threshold i falls above the valuable 1’s. Otherwise it gets an additional fraction of k, depending on how close the threshold is positioned to the valuable 1’s. For instance, if the threshold is set between 1 and k positions below the valuable 1’s, the algorithm will in expectation select edges of total value of at least 1 + k/2. This follows from the random arrival order of the edges and the fact that the ratio of valuable to non-valuable edges that the algorithm is ready to accept is at least 1 : 2. Furthermore, we see that for this distribution √ of instances the optimal way to set a deterministic threshold is at the lowest position. Using k = n, a lower bound on the competitive ratio is  √  n n k = . n = Ω n n k2 k log k + k log n n log k + 1 6 Algorithm 3: Bipartite Matching Input : vertex set R and cardinality n = |L| Output: matching M 1 2 3 4 5 6 Let L′ be the first ⌊ ne ⌋ vertices of L, and M ← ∅; for each ℓ ∈ L \ L′ do L′ ← L′ ∪ {ℓ}; M (ℓ) ← greedy matching on G[L′ ∪ R]; Let e(ℓ) ← (ℓ, r) be the edge assigned to ℓ in M (ℓ) ; if M ∪ {e(ℓ) } is a matching then add e(ℓ) to M ; 4 4.1 Matching and Packing Bipartite Matching In this section, we study online bipartite matching. The vertices on the right side of the graph (denoted by R) are static and given in advance. The vertices on the left side (denoted by L) arrive sequentially in a random order. Every edge e = (r, ℓ) ∈ R × L has a non-negative weight w(e) ≥ 0. In the cardinal model, each vertex of L reveals upon arrival the weights of all incident edges. In the ordinal model, we are given a total order on all edges that have arrived so far, consistent with their weights. Before seeing the next vertex of L, the algorithm has to decide to which vertex r ∈ R (if any) it wants to match the current vertex ℓ. A match that is formed cannot be revoked. The goal is to maximize the total weight of the matching. The algorithm for the cardinal model in [24] achieves an optimal competitive ratio of e. However, this algorithm heavily exploits cardinal information by repeatedly computing max-weight matchings for the edges seen so far. For the ordinal model, our Algorithm 3 below obtains a competitive ratio of 2e. While similar in spirit, the main difference is that we rely on a greedy matching algorithm, which is based solely on ordinal information. It deteriorates the ratio only by a factor of 2. Here we assume to have access to ordinal preferences over all the edges in the graph. Note that the same approach works if the vertices provide correlated (ordinal) preference lists consistent with the edge weights, for every vertex from R and every arrived vertex from L. In this case, the greedy algorithm can still be implemented by iteratively matching and removing a pair that mutually prefers each other the most, and it provides an approximation guarantee of 2 for the max-weight matching (see, e.g., [2]). In contrast, if we receive only preference lists for vertices on one side, there are simple examples that establish super-constant lower bounds on the competitive ratio1 . Lemma 2. Let the random variable Av denote the contribution of the vertex v ∈ L to the output, i.e. weight assigned to v in M . Let w(M ∗ ) denote the value of the maximum-weight matching in G. For ℓ ∈ {⌈ ne ⌉, . . . , n},   ⌊ n ⌋ w(M ∗ ) E Aℓ ≥ e · . ℓ−1 2n Proof. We first show that e(ℓ) has a significant expected weight. Then we bound the probability of adding e(ℓ) to M . In step ℓ, |L′ | = ℓ and the algorithm computes a greedy matching M (ℓ) on G[L′ ∪ R]. The current vertex ℓ can be seen as selected uniformly at random from L′ , and L′ can be seen as ∗ ∗ ) ) and E[w(e(ℓ) )] ≥ w(M selected uniformly at random from L. Therefore, E[w(M (ℓ) )] ≥ nℓ · w(M 2 2n . Here we use that a greedy matching approximates the optimum by at most a factor of 2 [2]. 1 Consider a bipartite graph with two nodes on each side (named A,B and 1,2). If we only know that both A and B prefer 1 to 2, the ratio becomes at least 2 even in the offline case. Similar examples imply that the (offline) ratio must grow in the size of the graph. 7 Edge e(ℓ) can be added to M if r has not been matched already. The vertex r can be matched only when it is in M (k) . The probability of r being matched in step k is at most k1 and the order of the vertices in steps 1, . . . , k − 1 is irrelevant for this event. Pr[r unmatched in step ℓ] = Pr ℓ−1 ^ " k=⌈n/e⌉ r 6∈ e We now know that Pr[M ∪ e(ℓ) is a matching] ≥ lemma follows. (k) ⌊n/e⌋ ℓ−1 . # ≥ ℓ−1 Y k=⌈n/e⌉ ⌈n⌉ − 1 k−1 = e k ℓ−1 Using this and E[w(e(ℓ) )] ≥ w(M ∗ ) 2n , the Theorem 3. Algorithm 3 for bipartite matching is 2e-competitive. Proof. The weight of matching M can be obtained by summing over random variables Aℓ . " n # n n−1 X X ⌊n/e⌋ w(M ∗ ) ⌊n/e⌋ X 1 E[w(M )] = E Aℓ ≥ · = · w(M ∗ ) ℓ−1 2n 2n ℓ ℓ=1 Since ⌊n/e⌋ n ≥ 1 e − 1 n and ℓ=⌈n/e⌉ Pn−1 1 ℓ=⌊n/e⌋ ℓ ℓ=⌊n/e⌋ n ≥ ln ⌊n/e⌋ ≥ 1, it follows that E[w(M )] ≥  1 1 − e n  · w(M ∗ ) . 2 In the submodular version of the offline problem, the natural greedy algorithm gives a 3approximation [20]. It builds the matching by greedily adding an edge that maximizes the marginal improvement of f , which is the information delivered by the ordinal oracle. When using this algorithm as a subroutine for the bipartite matching secretary problem, the resulting procedure achieves a 12-approximation in the submodular case [26]. 4.2 Packing Our results for bipartite matching can be extended to online packing LPs of the form max cτ x s.t. Ax ≤ b and 0 ≤ x ≤ 1, which model problems with m resources and n online requests coming in random order. Each resource i ∈ [m] has a capacity bi that is known in advance, together with the number of requests. Every online request comes with a set of options, where each option has its profit and resource consumption. Once a request arrives, the coefficients of its variables are revealed and the assignment to the variables has to be determined. Formally, request j ∈ [n] corresponds to variables xj,1 , . . . , xj,K that represent K options. Each option k ∈ [K] contributes with profit cj,k ≥ 0 and has resource consumption P ai,j,k ≥ 0 for resource i. Overall, at most one option can be selected, i.e., there is a constraint k∈[K] xj,k ≤ 1, ∀j ∈ [n]. The objective is to maximize total profit while respecting the resource capacities. The offline problem is captured by the following linear program: X X X X ai,j,k xj,k ≤ bi i ∈ [m] cj,k xj,k s.t. max j∈[n] k∈[K] j∈[n] k∈[K] X k∈[K] xj,k ≤ 1 j ∈ [n] As a parameter, we denote by d the maximum number of non-zero entries in any column of the constraint matrix A, for which by definition d ≤ m. We compare the solution to the fractional ∗ optimum, which we denote by ratio will be expressed in terms of d and the j x . The competitive k bi capacity ratio B = mini∈[m] maxj∈[n],k∈[K] . ai,j,k 8 Algorithm 4: Packing LP Input : capacities b, total number of requests n, probability p = Output: assignment vector y 1 2 3 4 e(2d)1/B 1+e(2d)1/B Let L′ be the first p · n requests, and y ← 0; for each j ∈ / L′ do ′ ′ L ← L ∪ {j}; ′ x(L ) ← greedy assignment on the LP for L′ ; (L′ ) yj ← xj ; if ¬(A(y) ≤ b) then yj ← 0; 5 6 Kesselheim et al. [24] propose an algorithm that heavily exploits cardinal information – it repeatedly solves an LP-relaxation and uses the solution as a probability distribution over the options. Instead, our Algorithm 4 for the ordinal model is based on greedy assignments in terms of profits cj,k . More specifically, the greedy assignment considers variables xj,k in non-increasing order of cj,k . It sets a variable to 1 if this does not violate the capacity constraints, and to 0 otherwise. Theorem 4. Algorithm 4 for online packing LPs is O(d(B+1)/B )-competitive. 4.3 Matching in General Graphs Here we study the case when vertices of a general undirected graph arrive in random order. In the beginning, we only know the number n of vertices. Each edge in the graph has a non-negative weight w(e) ≥ 0. Each vertex reveals the incident edges to previously arrived vertices and their weights (cardinal model), or we receive a total order over all edges among arrived vertices that is consistent with the weights (ordinal model). An edge can be added to the matching only in the round in which it is revealed. The goal is to construct a matching with maximum weight. We can tackle this problem by prolonging the sampling phase and dividing the vertices into “left” and “right” vertices. Algorithm 5 first samples n/2 vertices. These are assigned to be the set R, corresponding to the static side of the graph in bipartite matching. The remaining vertices are assigned to be the set L. The algorithm then proceeds by sampling a fraction of the vertices of L, forming a set L′ . The remaining steps are exactly the same as in Algorithm 3. Theorem 5. Algorithm 5 for matching in general graphs is 12e/(e+1)-competitive, where 12e/(e+ 1) < 8.78. Algorithm 5: General Matching Input : vertex set V and cardinality n = |V | Output: matching M 1 2 3 4 5 6 7 Let R be the first ⌊ n2 ⌋ vertices of V ; n ⌋ vertices of V , and M ← ∅; Let L′ be the further ⌊ 2e ′ for each ℓ ∈ V \ L do L′ ← L′ ∪ {ℓ}; M (ℓ) ← greedy matching on G[L′ ∪ R]; Let e(ℓ) ← (ℓ, r) be the edge assigned to ℓ in M (ℓ) ; if M ∪ {e(ℓ) } is a matching then add e(ℓ) to M ; 9 Algorithm 6: Independent Set in Graphs with Bounded Local Independence Number p Input : n = |G|, p = α1 /(α1 + 1) Output: independent set of vertices S 1 2 3 4 5 6 Set k ← Binom(n, p), S ← ∅; Reject first k vertices of G, denote this set by G′ ; Build a maximal independent set of vertices from G′ greedily, denote this set by M1 ; for each v ∈ G \ G′ do w∗ ← max{w | N (v) ∩ M1 }; if (v > w∗ ) ∧ (S ∪ {v} independent set) then add v to S; Algorithm 7: Simulate p Input : n = |G|, p = α1 /(α1 + 1) Output: independent set of vertices S 1 2 3 4 5 6 7 8 9 5 Sort all vertices in G in non-increasing order of value; Initialize M1 , M2 ← ∅; for each v ∈ G in sorted order do if M1 ∪ {v} independent set then flip a coin with probability p of heads; if heads then M1 ← M1 ∪ {v}; else M2 ← M2 ∪ {v}; S ← M2 ; for each w ∈ S do if w has neighbors in S then remove w and all his neighbors from S; Independent Set and Local Independence In this section, we study maximum independent set in graphs with bounded local independence number. The set of elements are the vertices V of an underyling undirected graph G. Each vertex has a weight wv ≥ 0. We denote by N (v) the set of direct neighbors of vertex v. Vertices arrive sequentially in random order and reveal their position in the order of weights of vertices seen so far. The goal is to construct an independent set of G with maximum weight. The exact structure of G is unknown, but we know that G has a bounded local independence number α1 . Definition 1. An undirected graph G has local independence number α1 if for each node v, the cardinality of every independent set in the neighborhood N (v) is at most α1 . We propose Algorithm 6, which is inspired by the Sample-and-Price algorithm for matching in [28]. Note that Göbel et al. [21] construct a more general approach for graphs with bounded inductive independence number ρ. However, they only obtain a ratio of O(ρ2 log n) for the weighted version, where a competitive ratio of Ω(log n/ log2 log n) cannot be avoided, even in instances with constant ρ. These algorithms rely on ρ-approximation algorithms for the offline problem that crucially exploit cardinal information. Similar to [28], we reformulate Algorithm 6 into an equivalent approach (Algorithm 7) for the sake of analysis. Given the same arrival order, the same vertices are in the sample. Algorithm 7 drops all vertices from S that have neighbors in S while Algorithm 6 keeps one of them. Hence, E[w(SAlg )] ≥ E[w(SSim )]. In what follows, we analyze the performance of Simulate. The first 6 lemma follows directly from the definition of the local independence number. Lemma 3. E[w(M1 )] ≥ p · w(S ∗ ) α1 , where α1 ≥ 1 is the local independence number of G.   Lemma 4. E |N (v) ∩ M2 | v ∈ M2 ≤ α1 (1−p) p . 10 Proof. Let us denote by Xu1 and Xu2 the indicator variables for the events u ∈ M1 and u ∈ M2 respectively. Then, # " X X     E Xu2 v ∈ M2 Xu2 v ∈ M2 = E |N (v) ∩ M2 | v ∈ M2 = E u∈N (v) u∈N (v)   1−p 1−p X = E Xu1 v ∈ M2 ≤ · α1 p p u∈N (v) Theorem 6. Algorithm 7 for weighted independent set is O(α21 )-competitive, where α1 is the local independence number of the graph. Proof. By using Markov’s inequality and Lemma 4, Pr[|N (v) ∩ M2 | ≥ 1 v ∈ M2 ] ≤ and Pr[|N (v) ∩ M2 | < 1 v ∈ M2 ] > α1 · (1 − p)/p 1 − (α1 (1 − p)/p) . Thus, we can conclude that     1−p 1−p 1−p E[w(S)] ≥ 1 − α1 · · E[w(M2 )] ≥ 1 − α1 · · · w(S ∗ ) . p p α1 q 1 , which proves the theorem. The ratio is optimized for p = αα 1 +1 As a prominent example, α1 = 5 in the popular class of unit-disk graphs. In such graphs, our algorithm yields a constant competitive ratio for online independent set in the ordinal model. References [1] David Abraham, Robert Irving, Telikepalli Kavitha, and Kurt Mehlhorn. Popular matchings. SIAM J. Comput., 37(4):1030–1045, 2007. [2] Elliot Anshelevich and Martin Hoefer. Contribution games in networks. Algorithmica, 63(1– 2):51–90, 2012. [3] Elliot Anshelevich and John Postl. Randomized social choice functions under metric preferences. In Proc. 25th Intl. Joint Conf. Artif. Intell. (IJCAI), pages 46–59, 2016. [4] Elliot Anshelevich and Shreyas Sekar. Blind, greedy, and random: Algorithms for matching and clustering using only ordinal information. In Proc. 13th Conf. Artificial Intelligence (AAAI), pages 390–396, 2016. [5] Elliot Anshelevich and Shreyas Sekar. Truthful mechanisms for matching and clustering in an ordinal world. In Proc. 12th Conf. Web and Internet Economics (WINE), pages 265–278, 2016. [6] Moshe Babaioff, Nicole Immorlica, David Kempe, and Robert Kleinberg. A knapsack secretary problem with applications. In Proc. 10th Workshop Approximation Algorithms for Combinatorial Optimization Problems (APPROX), pages 16–28, 2007. [7] Moshe Babaioff, Nicole Immorlica, David Kempe, and Robert Kleinberg. Online auctions and generalized secretary problems. SIGecom Exchanges, 7(2), 2008. [8] Moshe Babaioff, Nicole Immorlica, and Robert Kleinberg. Matroids, secretary problems, and online mechanisms. In Proc. 18th Symp. Discrete Algorithms (SODA), pages 434–443, 2007. 11 [9] MohammadHossein Bateni, MohammadTaghi Hajiaghayi, and Morteza Zadimoghaddam. Submodular secretary problem and extensions. ACM Trans. Algorithms, 9(4):32, 2013. [10] Deeparnab Chakrabarty and Chaitanya Swamy. Welfare maximization and truthfulness in mechanism design with ordinal preferences. In Proc. 5th Symp. Innovations in Theoret. Computer Science (ITCS), pages 105–120, 2014. [11] Sourav Chakraborty and Oded Lachish. Improved competitive ratio for the matroid secretary problem. In Proc. 23rd Symp. Discrete Algorithms (SODA), pages 1702–1712, 2012. [12] Ning Chen, Martin Hoefer, Marvin Künnemann, Chengyu Lin, and Peihan Miao. Secretary markets with local information. In Proc. 42nd Intl. Coll. Automata, Languages and Programming (ICALP), volume 2, pages 552–563, 2015. [13] Nedialko Dimitrov and Greg Plaxton. Competitive weighted matching in transversal matroids. Algorithmica, 62(1-2):333–348, 2012. [14] Michael Dinitz and Guy Kortsarz. Matroid secretary for regular and decomposable matroids. SIAM J. Comput., 43(5):1807–1830, 2014. [15] Eugene Dynkin. The optimum choice of the instant for stopping a Markov process. In Sov. Math. Dokl, volume 4, pages 627–629, 1963. [16] Moran Feldman, Ola Svensson, and Rico Zenklusen. A simple O (log log(rank))-competitive algorithm for the matroid secretary problem. In Proc. 26th Symp. Discrete Algorithms (SODA), pages 1189–1201, 2015. [17] Moran Feldman and Moshe Tennenholtz. Interviewing secretaries in parallel. In Proc. 13th Conf. Electronic Commerce (EC), pages 550–567, 2012. [18] Moran Feldman and Rico Zenklusen. The submodular secretary problem goes linear. In Proc. 56th Symp. Foundations of Computer Science (FOCS), pages 486–505, 2015. [19] Amos Fiat, Ilia Gorelik, Haim Kaplan, and Slava Novgorodov. The temp secretary problem. In Proc. 23rd European Symp. Algorithms (ESA), pages 631–642, 2015. [20] Marshall Fisher, George Nemhauser, and Laurence Wolsey. An analysis of approximations for maximizing submodular set functions-II. In Polyhedral combinatorics, pages 73–87. Springer, 1978. [21] Oliver Göbel, Martin Hoefer, Thomas Kesselheim, Thomas Schleiden, and Berthold Vöcking. Online independent set beyond the worst-case: Secretaries, prophets and periods. In Proc. 41st Intl. Coll. Automata, Languages and Programming (ICALP), volume 2, pages 508–519, 2014. [22] MohammadTaghi Hajiaghayi, Robert Kleinberg, and David Parkes. Adaptive limited-supply online auctions. In Proc. 5th Conf. Electronic Commerce (EC), pages 71–80, 2004. [23] Patrick Jaillet, José Soto, and Rico Zenklusen. Advances on matroid secretary problems: Free order model and laminar case. In Proc. 16th Intl. Conf. Integer Programming and Combinatorial Optimization (IPCO), pages 254–265, 2013. [24] Thomas Kesselheim, Klaus Radke, Andreas Tönnis, and Berthold Vöcking. An optimal online algorithm for weighted bipartite matching and extensions to combinatorial auctions. In Proc. 21st European Symp. Algorithms (ESA), pages 589–600, 2013. [25] Thomas Kesselheim, Klaus Radke, Andreas Tönnis, and Berthold Vöcking. Primal beats dual on online packing LPs in the random-order model. In Proc. 46th Symp. Theory of Computing (STOC), pages 303–312, 2014. 12 [26] Thomas Kesselheim and Andreas Tönnis. Submodular secretary problems: Cardinality, matching, and linear constraints. In Proc. 20th Workshop Approximation Algorithms for Combinatorial Optimization Problems (APPROX), pages 16:1–16:22, 2017. [27] Robert Kleinberg. A multiple-choice secretary algorithm with applications to online auctions. In Proc. 16th Symp. Discrete Algorithms (SODA), pages 630–631, 2005. [28] Nitish Korula and Martin Pál. Algorithms for secretary problems on graphs and hypergraphs. In Proc. 36th Intl. Coll. Automata, Languages and Programming (ICALP), pages 508–520, 2009. [29] Oded Lachish. O(log log rank) competitive ratio for the matroid secretary problem. In Proc. 55th Symp. Foundations of Computer Science (FOCS), pages 326–335, 2014. [30] José Soto. Matroid secretary problem in the random-assignment model. SIAM J. Comput., 42(1):178–211, 2013. 13 A Appendix A.1 Proof of Theorem 4 The proof is based on the following lemma. Lemma 5. Let the random variable Aℓ denote the contribution of request xℓ to the output and cτ x∗ the value of the optimal fractional solution. For requests xj ∈ {pn + 1, . . . , n} it holds that B ! τ ∗  c x e(1 − p) . E[Aj ] ≥ 1 − d · p (d + 1)n (L′ ) (L′ ) (L′ ) τ ∗ c x Proof. If xj,k = 1, then as in proof of Lemma 2, we get E[cj xj ] = E[cj,k xj,k ] ≥ (d+1)n , where ′ the expectation is taken over the choice of the set L and the choice of the last vertex in the order of arrival. (L′ ) only if the capacity constraints can be respected. For the sake The algorithm sets yj to xj (L′ ) of analysis, we assume that the algorithm only sets yj to xj (L′ ) xj if every capacity constraint bi that (L′ ) (xj,k affects = 1 and ai,j,k 6= 0) is affected by at most B −1 previous requests. We bound the probability of a capacity constraint bi being affected in any preceding step s ∈ {pn + 1, . . . , j − 1}, for a fixed i: X (L′ ) Pr[(xj ′ is last in the order) ∧ (ai,j ′ ,k′ 6= 0)] Pr[bi affected by xs,k′ = 1] ≤ xj′ ∈{1,...,s} ≤ 1 s X xj′ ∈{1,...,s} B , s Pr[ai,j ′ ,k′ 6= 0] ≤ where the last step follows from y being a feasible solution throughout the run of the algorithm. (L′ ) Now, we bound the probability of not being able to set yj to xj : Y  X B Pr[bi is affected at least B times] ≤ s C⊆{pn+1,...,j−1}, |C|=B ≤ s∈C B    B  (1 − p)n B (1 − p)e · ≤ , pn p B (L′ ) so the probability of succeeding in setting yj to xj Pr[Ay ≤ b] ≥ 1 − d ·  is (1 − p)e p B , (L′ ) and there are at most d because we can do a union bound over all bi that are affected by xj such, since that is the maximal number of non-zero entries in any column of the constraint matrix A. (L′ ) Combining this with the inequality regarding the expected contribution of xj , we get the claimed result. Using Lemma 5, we get E[cτ y] = n X ℓ=pn+1 τ = ∗ E[Aℓ ] ≥ n X l=pn+1 1 c x · · d + 1 1 + e(2d)1/B B ! τ ∗ c x e(1 − p) 1−d· p (d + 1)n B !  cτ x∗ 1 ≥ 1−d· . (2d)1/B 2(d + 1)(1 + 2ed1/B )  14 Note that Theorem 4 contains the one-sided b-hypermatching problem as a special case. For the even more special case of b = 1 in the one-sided hypermatching, an algorithm was given in [28], which also works in the ordinal model. Our ratio in this special case is similar, but our approach extends to arbitrary capacities b ≥ 1. A.2 Proof of Theorem 5 The proof is based on the following lemma. Lemma 6. Let the random variable Aℓ denote the contribution of vertex ℓ ∈ ⌊n/2 + n/(2e)⌋ to the output, i.e. the weight of the edge assigned to ℓ in M . Then, E[Aℓ ] ≥ n ⌈ n2 + 2e ⌉ 1 w(M ∗ ) · · ℓ−1 2 n Proof. The proof is similar to the one of Lemma 5, with the additional observation that each edge is available with probability 12 . It is available only if the incident vertices are assigned to different sides of the bipartition. We now use the lemma to bound as follows: " n # n X X E[w(M )] = E Aℓ ≥ ℓ=1 ℓ=⌈n/2+n/(2e)⌉ ⌊n/2 + n/(2e)⌋ 1 w(M ∗ ) · · ℓ−1 2 n n−1 X 1 ⌊n/2 + n/(2e)⌋ w(M ∗ ) · · n 2 ℓ ℓ=⌊n/2+n/(2e)⌋    1 1 w(M ∗ ) 1 1 1+ − · · ≥ 2 e n 2 3 = A.3 Competitive Ratio and 0–1 Weights For worst-case bounds, we can restrict our attention to instances where all elements have cardinal weights in {0, 1}. These instances always result in the worst competitive ratio, as shown in the following lemma. Lemma 7. By converting an arbitrary weighted instance to an instance with weights in {0, 1}, the competitive ratio between the optimum solution and the solution computed by an algorithm based on ordinal information can only deteriorate. Proof. Without loss of generality, we assume that all elements of the original instance have distinct weights. We denote the elements chosen in the optimal solution by a∗1 , . . . , a∗k and the elements chosen by the algorithm by b1 , . . . , bm . The numbering respects the ordinal ordering of weights, i.e., a∗1 ≻ a∗2 ≻ . . . a∗k and b1 ≻ b2 ≻ · · · ≻ bm . The competitive ratio is w(a∗1 ) + · · · + w(a∗k ) OPT = . ALG w(b1 ) + · · · + w(bm ) This ratio can only increase if we change the weight of all elements that appear after a∗k in the global ordering to 0. This effectively shortens the set of elements with a contribution chosen by the algorithm to b1 , . . . , bℓ , for some suitable ℓ ≤ m. Furthermore, we change the weights of all elements between a∗i and a∗i+1 by decreasing them to a∗i+1 . We now denote the elements that the algorithm chose by c1 , . . . , cl , since their weights might have changed. Both of these changes do 15 not influence OPT, but they reduce the weight of the solution returned by the algorithm. We continue converting the instance, by focusing on w(a∗k ). Then, OPT w(a∗1 ) + · · · + w(a∗k ) w(a∗1 ) + · · · + w(a∗k ) = ≤ ALG w(b1 ) + · · · + w(bm ) w(b1 ) + · · · + w(bℓ ) A + w(a∗k ) w(a∗1 ) + · · · + w(a∗k ) = , ≤ w(c1 ) + · · · + w(cℓ ) B + r · w(a∗k ) where A = w(a∗1 ) + · · · + w(a∗k−1 ), B is the sum of the weights of all elements that the algorithm chose which are not equal to w(a∗k ) in the altered instance and r ∈ N0 . Taking the derivative for w(a∗k ),  A + w(a∗k ) B + r · w(a∗k ) ′ = B−r·A , (B + r · w(a∗k ))2 we either decrease w(a∗k ) to 0 or raise it to w(a∗k−1 ) (depending what makes the ratio increase, i.e., the sign of the derivative). We continue this procedure until all weights of the instance are equal to either to a∗1 or 0. Note that these changes preserve the global ordering. W.l.o.g., we can finally set w(a∗1 ) = 1. Note that we increased the ratio between the solution of the algorithm and the optimal solution for the original weights, when applying the transformed weights. Note that none of these transformations change the decisions of the algorithm. In contrast, the optimum solution for the transformed weights can only become better, which even further deteriorates the competitive ratio. 16
8cs.DS
arXiv:0908.3650v1 [cs.PL] 25 Aug 2009 Lazy mixin modules and disciplined effects Keiko Nakata February 27, 2018 Abstract Programming languages are expected to support programmer’s effort to structure program code. The ML module system, object systems and mixins are good examples of language constructs promoting modular programming. Among the three, mixins can be thought of as a generalization of the two others in the sense that mixins can incorporate features of ML modules and objects with a set of primitive operators with clean semantics. Much work has been devoted to build mixin-based module systems for practical programming languages. In respect of the operational semantics, previous work notably investigated mixin calculi in call-by-name and call-by-value evaluation settings. In this paper we examine a mixin calculus in a call-by-need, or lazy, evaluation setting. We demonstrate how lazy mixins can be interesting in practice with a series of examples, and formalize the operational semantics by adapting Ancona and Zucca’s concise formalization of call-by-name mixins. We then extend the semantics with constraints to control the evaluation order of components of mixins in several ways. The main motivation for considering the constraints is to produce side effects in a more explicit order than in a purely lazy, demand-driven setting. We explore the design space of possibly interesting constraints and consider two examples in detail. 1 Introduction Modularity is an important factor in the development of large programs. In particular programmer’s effort to logically organize program code is of great importance in the long run to maintain, debug and extend the program code. Many modern programming languages have mechanisms to support this effort by facilitating modular development of programs. Examples of these mechanisms are object systems and the ML module system. Being ML programmers, we enjoy the rich expressivity of the ML module system for modular programming. Nestable structures allow us to hierarchically organize namespaces and program code. With signature constraints, we control visibility of components of structures. In particular the combination of nesting and signature constraints offers fine grained visibility control, as witnessed, for example, by the Moby programming language [Fisher and Reppy(1999)]. Functors, which are functions on modules, facilitate code reuse in a modular way. 1 Although functors might not be as pervasive as nesting or signature constraints in our programs, they play a critical role in some contexts. Good examples are Map.Make and Set.Make functors as implemented in OCaml’s standard library. However ML does not have recursive modules. That is, neither recursive functions nor types can be defined across module boundaries. As a result of this constraint, programmers may have to consolidate conceptually separate components into a single module, intruding on modular programming [Russo(2001)]. Compared to ML, object systems have excellent support for recursion across object/class boundaries. Particularly in the presence of late-binding, object systems facilitate the development of extensible programs, where recursively defined types and functions may need to be extended together. Typical such scenarios are condensed in the notorious expression problem [Torgersen(2004)]. We expect a module system to support all the familiar features of ML modules as well as recursion between modules and late-binding. There are at least three approaches to design such a module system. Two of them are to extend the ML module system with recursion [Crary et al.(1999)Crary, Harper, and Puri] and to extend object systems with nesting, abbreviation and type members [Odersky et al.(2003)Odersky, Cremet, Röckl, and Zenger]. The third approach is to develop another form of a module system, namely mixins. In this paper we follow this third approach. The concept of mixins is first introduced in the context of object systems [Bracha(1992)], then is extended to in the context of ML-style modules. A mixin is a collection of named components, where each component can be either defined (bound to a definition) or deferred (declared without definition). Two key operations on mixins are the sum and freeze operations; the former takes two mixins and composes a new mixin by merging the two, and the latter resolves, or links, deferred components of a mixin to defined ones. The sum operation is reminiscent of functor application in ML and inheritance in object systems. However more flexibility is obtained by separating the resolution from the sum operation. The freeze operation is free to resolve a deferred component to a defined one independently of their names at any point, as long as their types match; notably it can liberally tie a recursive knot inside a mixin. Indeed mixins are designed to be a generalization of ML-style modules and objects, by incorporating features of both with a set of primitive operators with clean semantics [Bracha(1992), Ancona and Zucca(2002)]. Two challenging problems remain to replace ML-style modules with mixins, namely type checking and initialization. In this paper, we address the latter problem. Promising progress has been made in designing a type system for mixins with type components and a signature language to enforce type abstraction between mixins [Odersky et al.(2003)Odersky, Cremet, Röckl, and Zenger, Owens and Flatt(2006)]; we expect to benefit from the previous work for the former problem. Initialization of mixins poses an important design problem if we are to build a mixin-based module system on top of a call-by-value core language supporting arbitrary side-effects, such as the ML core language. The main difficulty stems 2 from unconstrained recursive definitions such as let rec x = x that the core language does not allow but that might result from the freeze operation. We need an initialization semantics for mixins which takes account of unconstrained recursion consistently with the call-by-value semantics of the core language; this is the subject of this paper. Ancona and Zucca’s mixin calculus [Ancona and Zucca(2002)], called CMS, is one of the most influential work in formalizing an operational semantics for mixins. The formalization is elegantly concise and the paper illustrates how mixins support various existing constructs for composing modules, found in the ML module system, object-systems and linking calculi [Cardelli(1997)], wellexplaining why we claim mixins are a generalization of other modular systems. Inspired by CMS, Hirschowitz and Leroy examined a mixin calculus in a callby-value setting to build a mixin-based module system on top of the ML core language. The original call-by-name semantics of CMS might not be well-suited for the call-by-value effectful ML core language; CMS admits recursive definitions such as let rec x = x, causing the evaluation to diverge when x is selected, and it can produce the same side-effect repeatedly. Neither behavior is consistent with the semantics of ML. In this paper we examine a mixin calculus in a call-by-need, or lazy, evaluation setting with a back-patching semantics. Broadly we model mixins as nestable records with lazy fields. In the simplest setting which we examine in Section 3, a component of a mixin is evaluated when it is selected. We explain how side-effects are duplicated and linearly produced by open and closed mixins respectively, while allowing both open and closed mixins to be merged indiscriminately; this tolerance of merging mixins of different status is our important design choice, which is different from previous proposals [Hirschowitz and Leroy(2005), Ancona et al.(2003)Ancona, Fagorzi, Moggi, and Zucca]. In Section 2.1, we exemplify how the tolerance can be useful. Then in Section 4, we extend the former semantics of Section 3 with an ability to constrain the evaluation order of components of mixins. While the former semantics admits the most flexible recursive initialization patterns for mixins, the flexibility can do harm because of the intrinsically implicit evaluation order determined at run-time in a demand-driven way. It may be interesting to constrain possible initialization patterns, making the evaluation order more explicit. For instance, we may want to enforce top-down evaluation order within mixins, where components of a mixin are evaluated following the textual definition order in the source program. Or, we may want to keep the invariant that once a component of a mixin is selected, all its components are eventually evaluated. Indeed our ultimate goal is to find the most beneficial constraint on the evaluation order that still admits interesting recursive initialization patterns, but that makes the evaluation order more explicit, thus more predictable to programmers. In this paper we do not propose a particular constraint; simply we do not yet have enough experience in programming with mixins to decide what constraint is best in practice. Hence we formalize the operational semantics so that it deals 3 with several constraints. Concretely, we give a constraint language and extend the semantics of Section 3 to evaluate components of mixins according to a given constraint expressed by the constraint language. We explore the design space of possibly interesting constraints and examine two particular constraints in detail as examples. Contributions of the paper are summarized as follows. We formalize the operational semantics for a lazy mixin calculus (Section 3), by moderately extending Ancona and Zucca’s formalization of a call-by-name mixin calculus [Ancona and Zucca(2002)]. We demonstrate how lazy mixins can be useful in practice through examples (Section 2). Then we extend the semantics to be able to control the evaluation order of components of mixins in several ways (Section 4) and exemplify concrete scenarios where particular evaluation strategies are enforced by constraints. We believe the ability to deal with several evaluation strategies is a novelty of the formalization and the formalization serves as a basis for exploring the design space. 2 Examples In this section, we introduce our lazy mixin calculus through a series of examples. Many constructs of the calculus come from CMS [Ancona and Zucca(2002)]. Examples are written in a more programmer-friendly surface syntax and we assume a small subset of the OCaml core language for the core language of the mixin calculus. We recall that the OCaml core language adopts a call-by-value evaluation strategy and supports arbitrary side-effects within (core) expressions. 2.1 MakeSet and MakeMultiSet mixins We start by looking at possible mixin equivalents to MakeSet and MakeMultiSet ML functors and their instances by the Key structure, as given in Figure 1. MakeSet and MakeMultiSet are functions on ML structures, or functors, for making sets and multi-sets of integers, internally represented as lists and lists of lists respectively. Both take as argument a structure containing a create function for producing integers out of nothing and a compare function for comparing given two integers. MakeSet extends this functionality to sets of integers and MakeMultiSet does for multi-sets. Then we apply the two functors to the Key structure to instantiate customized Set and MultiSet structures. Below is a mixin equivalent to the MakeSet functor: mixin MakeSet = { val create element : unit → int val compare element : int → int → int let create () = [ create element () ] let compare s1 s2 = ..... compare element ... } A mixin structure is a sequence of named components, where a component may be defined like create and compare or deferred like create element and com- 4 module MakeSet = functor (X : sig val create : unit → int val compare : int → int → int end) → struct let create () = [ X.create () ] let compare s1 s2 = ..... X.compare ... end module MakeMultiSet = functor(X : sig val create : unit → int val compare : int → int → int end) → struct let create () = [[ X.create () ]] let compare s1 s2 = .... X.compare ... end module Key = struct let count = ref (-1) let create () = incr count; !count let compare x y = if x = y then 0 else if x < y then 1 else (-1) end module Set = MakeSet(Key) module MultiSet = MakeMultiSet(Key) Figure 1: MakeSet and MakeMultiSet ML functors pare element. The bodies of defined components can refer to names of deferred components as if they were present. Similarly, below is a mixin equivalent to the MakeMultiSet functor: mixin MakeMultiSet = { val create element : unit → int val compare element : int → int → int let create () = [[ create element () ]] let compare s1 s2 = ..... compare element ... } We build a mixin equivalent to the Key structure in two steps: mixin FKey = { let count = ref (-1) let create key () = incr count; !count let compare key x y = if x = y then 0 else if x < y then 1 else (-1) } mixin Key = close(FKey) We distinguish two states of mixins: open and closed. Intuitively a closed mixin is a record with lazy fields, whereas an open mixin is a function which returns a record with lazy fields. Intended implications of this comparison are 1) projection of components is only possible from closed mixins, but not from open mixins; 2) an open mixin can be instantiated to create closed mixins; 3) side-effects contained in a closed mixin are produced exactly once, whereas side-effects in 5 an open mixin are produced as many times as the mixin is instantiated. FKey above is an open mixin, and Key is a closed mixin instantiated from FKey by the close operation. A projection Key.create key is legal, but FKey.create key is not. As expected, the counter Key.count is initialized to -1 exactly once. We may close the FKey mixin again: mixin Key2 = close(FKey) Counters Key.count and Key2.count are distinct from each other. For instance, calling Key.create does not increase Key2.count. The close operation is only available for mixins without holes, or mixins which do not contain deferred components. The freeze operation, possibly combined with the sum operation, is used to fill in the holes. For instance, below we merge Key and MakeSet mixins by the sum operation, then resolve the deferred components create element and compare element of MakeSet to the defined component create key and compare key of Key respectively by the freeze operation, to fill in the holes of MakeSet: mixin FSet’ = Key ← MakeSet mixin FSet = freezeψ (FSet’) where ψ is the following mapping1 : h create element 7→ create key; compare element 7→ compare key i Now FSet does not contain holes, hence we can close it: mixin Set = close(FSet) We do the same for MakeMultiSet in one step this time: mixin MultiSet = close(freezeψ (Key ← MakeMultiSet)) Set and MultiSet mixins are equivalent to Set and MultiSet ML structures as given in Figure 1. It is important to see differences resulting from closing FKey after merging it with MakeSet and MakeMultiSet as follows: mixin Set’ = close(freezeψ (FKey ← MakeSet)) mixin MultiSet’ = close(freezeψ (FKey ← MakeMultiSet)) Here FKey is instantiated twice. While Set and MultiSet share the same counter, Set’ and MultiSet’ have distinct counter’s. As a result, calling Set.create affects both results of the next calls to Set.create and MultiSet.create, while calling Set’.create only does the result of the next call to itself. 2.2 Widget mixins Many GUI programming APIs involve recursive initialization patterns to form mutually referential graphs among related widgets in the style of create-andconfigure, as Syme calls in [Syme(2005)]. In that paper he explains how recur1 Only in this section, we use angle brackets instead of square brackets to denote mappings to avoid confusion with list expressions. 6 sive initialization patterns are prominent in GUI programming and proposes a semi-safe lazy evaluation strategy for the ML core language to tame ML’s value recursion restriction; the restriction constrains right-hand sides of recursive definitions to be syntactic values, thus may hinder uses of sophisticated GUI programming APIs. In this subsection, we consider a GUI example similar to his in a lazy mixin setting. We assume given the following interface of the API: type form type formMenu type menuItem val createForm: string → form val createMenu: string → formMenu val createMenuItem: string → menuItem val toggle : menuItem → unit val setMenus : form ∗ formMenu list → unit val setMenuItems : formMenu ∗ menuItem list → unit val setAction : menuItem ∗ (unit → unit) → unit The API requires the create-and-configure initialization pattern where widgets are first created, then explicit mutation configures a relation between the widgets. Below we build boiler-plate mixins which encapsulate the create-andconfigure pattern: mixin Form = { val name : string val menus : formMenu list let form = createForm(name) let = setMenus(form, menus) } mixin Menu = { val name : string val items : menuItem list let menu = createMenu(name) let = setMenuItems(menu, items) } mixin MenuItem = { val name : string val other : menuItem let item = createMenuItem(name) let = setAction(item, fun () → toggle(other)) } Then we use them: 7 MyForm = hidename (freezehname7→namei (Form ← { let name = “Form” })) MyMenu = hidename (freezehname7→namei (Menu ← { let name = “Menu” })) MyItem1 = rename(hother7→item2i,hitem17→itemi) (hidename ( freezehname7→namei (MenuItem ← { let name = “Rice” }))) MyItem2 = rename(hother7→item1i,hitem27→itemi) (hidename ( freezehname7→namei (MenuItem ← { let name = “Grape” }))) MyGUI = close(freezeψ1 ( MyItem1 ← (MyItem2 ←(MyMenu ←MyForm)))) where ψ1 is the following mapping: h item1 7→ item1; item2 7→ item2; items 7→ [item1; item2] ; menus 7→ [menu]i Above we have introduced two new constructs. The hide operation hideX (M ) hides the component named X of the mixin M by making the component invisible outside. The rename operation rename(φ1 ,φ2 ) (M ) changes names of deferred and defined components of the mixin M by φ1 and φ2 respectively, where φ1 and φ2 are finite mappings on names of components2 . For instance, in the definition of MyItem1, the deferred component other is renamed to item2 and the defined component item to item1. Observe the opposite directions of mappings hother 7→ item2i for deferred components and hitem1 7→ itemi for defined ones. This adds flexibility of the rename operation in that when φ1 maps deferred components of distinct names to the same name, then the components can be resolved simultaneously, and that when φ2 maps two distinct names X1 and X2 to a single name X, then the component which was named X can now projected by either X1 or X2 . Thanks to the renaming, MyItem1 and MyItem2 can be merged to form MyGUI without causing name clash; a deferred component and a defined component of a mixin can have the same name, but a deferred (resp. defined) component must not have the same name as other deferred (resp. defined) components. Besides, we have used the freeze operation in a more flexible way by mapping a name of a deferred component to an expression composed of names of defined components, for example [item1; item2]. The above example builds a GUI application forming a widget containment hierarchy where MyForm contains MyMenu, which contains MyItem1 and MyItem2. MyItem1 and MyItem2 are mutually recursive; each toggles the activation state of the other. We hide the name component of mixins to be merged to avoid name clash. Anonymous (under-scored) components do not contribute to name clash; they would be implemented as syntax sugar via the hide operation. Renaming of other and item components of MyItem1 and MyItem2, both of which 2 In the formalization, we will distinguish α-convertible identifiers (internal names) and non-convertible names (external names). Then φ1 and φ2 are mappings on names, not on identifiers. 8 are derived from the same mixin MenuItem, is necessary for cross-connecting the deferred component of one to the defined component of the other. In the last line we sum up the constituent mixins and resolve deferred components to defined ones, configuring the widget containment hierarchy. This example suggests it can be useful to control the evaluation order of components of mixins, instead of evaluating them purely lazily in a demanddriven way, i.e. evaluating only the components that are projected. Indeed we would like to make sure the widget containment hierarchy has been properly configured before MyGUI is actively used. For that purpose, all the anonymous components of MyGUI must be evaluated before any of its components becomes externally accessible. In Section 4, we present the operational semantics which can enforce such constraint on the evaluation order. 2.3 A combinator library for marshallers The last example deals with marshaller combinators and is motivated by Syme’s paper again [Syme(2005)]. Kennedy introduced a functional-language combinator library for building marshallers and unmarshallers of data structures [Kennedy(2004)]. The essential ingredient of his proposal is the tying together of a marshaller and unmarshaller pair in a single value. Then the consistency of marshalling and unmarshalling is ensured by construction. The original proposal of Kennedy is implemented in Haskell. Porting the code to ML is mostly easy, except for a couple of wrinkles. The value recursion restriction of ML is a source for the wrinkles and requires cumbersome workaround for an ML version of the combinator. In this subsection, we rebuild a combinator library for marshallers using lazy mixins with ML as the core language and demonstrate how the use of lazy mixins avoids the value recursion problem. In the following examples we assume a richer mixin language where a mixin may contain deferred and defined types as components, although our formal development does not consider mixins with type components. Type checking of mixins is not in the scope of the paper. As far as our examples are concerned, existing type systems are sufficient [Flatt and Felleisen(1998)]. We consider a mixin-based combinator library with the specification given below. A mixin signature mixin M : { type t val x : t } → { type s = t ∗ t val y : s } specifies an open mixin with deferred type component t and value component x of type t, written in the left-hand side of the arrow, and with defined type component s satisfying type equation s = t ∗ t and value component y of type s, written in the right-hand side3 . The scope of type names declared in the left hand of the arrow extends to the right hand. 3 This signature language is designed only for the sake of the examples. A more practical signature language is proposed, for instance in [Owens and Flatt(2006)]. 9 type channel type α marshaller val marshal : α marshaller → α ∗ channel → unit val unmarshal : α marshaller → channel → α mixin PairMrshl : { type s1 type s2 val mrshl1 : s1 marshaller val mrshl2 : s2 marshaller } → { type t = s1 ∗ s2 val marshaller : t mrshl } mixin ListMrshl : { type elm val mrshl elm : elm marshaller } → { type t = elm list val mrshl : t marshaller } mixin InnerMrshl : { type src type trg val f : src → trg val g : trg → src val mrshl src : src marshaller } → { type t = trg val mrshl : t marshaller } mixin IntMrsh : {} → { type t = int val mrshl : t marshaller } mixin StringMrsh : {} → { type t = string val mrshl : t marshaller } IntMrsh and StringMrsh mixin have an empty deferred component. But they are still open mixins, thus needs to be closed for making their components accessible. The abstract type marshaller could be internally implemented as a record consisting of a marshalling action and unmarshalling action: type α marshaller = { marshal: α ∗ channel → unit; unmarshal : channel → α } Recall that it is important for consistently building marshallers and unmarshallers that a marshaller is a single value, but not two separate functions. In this way, users of the library can only build consistent marshaller/unmarshaller pairs. We do not present further details of how the library can be implemented. In the original paper [Kennedy(2004)], Kennedy explains an excellent implementation which lets the programmer control sharing of the marshaled data. Transposing marshaller combinators for constructed types such as pairs and lists, originally implemented as functions, to mixins is straightforward. For instance, a function-based marshaller combinator pairMrshl(mrshl1, mrshl2) of type (s1 marshaller) ∗ (s2 marshaller) → (s1 ∗ s2) marshaller for constructing a marshaller for a pair from marshallers of the components can be translated into a mixin as follows: mixin PairMrshl = { type s1 type s2 val mrshl1 : s1 marshaller val mrshl2 : s2 marshaller type t = s1 ∗ s2 let mrshl = (∗ the body of pairMrshl(mrshl1, mrshl2) ∗) } Now we turn to how to build custom-marshallers for user-defined data types. We first build marshallers for both a single file, represented as a pair of an integer 10 and string, and a list of files: type file = int ∗ string mixin FileMrshl = freeze*( (rename(∅,hs17→t;mrshl17→mrshli) (IntMrshl)) ← (rename(∅,hs27→t;mrshl27→mrshli) (StringMrshl)) ← PairMrshl)) mixin FilesMrshl = freeze*( (rename(∅,helm7→t;mrshl elm7→mrshli) (FileMrshl)) ← ListMrshl) The notation ∅ denotes an empty mapping. Above we have introduced a highlevel mixin construct freeze*, which resolves deferred components to the samenamed defined components if exists, then hides the defined components used. The formalization given in the next section does not include freeze*. For the surface language, we could implement it by combining freeze and hide operations with the help of the type system. Next we build marshallers for both a single folder and a list of folders, which form recursive data structures: type folder = { files: file list; subfldrs: folders } and folders = folder list Like Syme, we use an intermediate mixin in favor of conciseness. mixin FldrInnerMrshl = freeze*( { type src = folder type trg = folders let f (fls, fldrs) = { files = fls; subfldrs = fldrs } let g fld = (fld.files, fld.subfldrs) } ← InnerMrshl) Then to build marshallers for folder and folders, we follow their type definitions, by merging constituent mixins and resolving deferred components to defined ones: mixin FldrMrshl = close(freeze*( (rename(∅,hs17→t;mrshl17→mrshli) (FilesMrshl)) ← (hidet (rename(hs17→s1;s27→s2;mrshl17→mrshl1;mrshl27→mrshl2i , hmrshl src7→mrshli) (PairMrshl))) ← (rename(hmrshl src7→mrshl srci,helm7→t;mrshl elm7→mrshl;fldMrshl7→mrshli) (FldrInnerMrshl)) ← (rename(helm7→elm;mrshl elm7→mrshl elmi, hfldrsMrshl7→mrshl;mrshl27→mrshl;s27→ti) (ListMrshl)) Thus we have created marshaller FldrMrshl.fldrMrshl for a single folder and FldrMrshl.fldrsMrshl for a list of folders. In the surface language we could provide an identity mapping with appropriate domain to improve notational verbosity. For the above example, we preferred not to use identity mappings to make explicit how deferred components are resolved to defined ones. Specifying marshallers for recursive data types such as folder and folders is not problematic in the original Haskell context of Kennedy or in our lazy mixin context, essentially due to the laziness. But it is problematic in the context of the ML core language because of the value recursion restriction. The problem is well-discussed in the previous papers of Kennedy and Syme. In short, it is 11 due to ML’s intolerance of the following recursive definition4 let rec p = pairMrshl(p, intMrshl) where we assumed primitive marshaller intMrshl and marshaller combinator pairMrshl given in the library. Note that the abstraction of type marshaller is not the root of the problem. Even we exposed the underlying implementation of marshaller, we still face the problem since marshallers are pairs of functions. If we consider first-class mixins, i.e. mixins as core values, we could implement a marshaller as separate functions of marshal action and unmarshal action, while ensuring the consistency of constructed marshaller/unmarshaller pairs. Our formalization of a lazy mixin calculus is largely abstracted over the core language. The core language may support first-class mixins, but we preferred the current presentation in favor of generality by not assuming a richer core language. We also found the current presentation useful to demonstrate a possibly interesting scenario where lazy mixins and a call-by-value core language are combined to obtain more tolerant recursive definitions. 3 Lazy mixins We start by considering a simpler semantics, where a component of a mixin is evaluated when it is projected. The syntax of our lazy mixin calculus, named Lyre, is defined in Figure 2. We assume pairwise disjoint sets Idents of identifiers, Names of names, and Loc of locations. Components of a mixin are internally referred to by (α-convertible) identifiers, but externally accessed by (non-convertible) names. We use locations to formalize lazy evaluation. Notations For a finite mapping f , dom(f ) and ran(f ) respectively denote the domain and range of f . ∅ is an empty mapping, that is, dom(∅), cod (∅) and i∈I ran(∅) are empty sets. The notation ai 7→ bi denotes the finite mapping f such that, for all i ∈ I, f (ai ) = bi . It is only defined when, for all i, j ∈ I, i 6= j implies ai 6= aj . Throughout the paper, we only consider finite mappings, so simply say a mapping to mean a finite one. For a mapping f , f \x is the restriction of f to dom(f )\{x}. For mappings f, f ′ , we write f + f ′ for the union of f and f ′ . That is, dom(f + f ′ ) =dom(f ) ∪ dom(f ′ ), and for all x in dom(f + f ′ ), f (x) when x ∈ dom(f ) (f + f ′ )(x) = f ′ (x) when x ∈ dom(f ′ ) f + f ′ is defined only if dom(f ) ∩ dom(f ′ ) = ∅. The notation f ′ ◦ f denotes the mapping composition. It is defined only when ran(f ) ⊆ dom(f ′ ). We have explained most of the constructs for mixin expressions in the previous section. In the formalization, however, mixin structures take a more fundamental form. Precisely, a mixin structure, simply called structure hereafter, is a triple of input assignment ι, output assignment o and local binding 4 The example is not ideal in that ML does not allow recursive type definitions such as type t = t ∗ int. But the source of the problem should be clear. 12 x, y X, Y l ∈ ∈ ∈ Idents Names Loc identifiers names locations ::= ::= | | | | | | | | | ::= | | C|M [ι; o; ρ] M1 ← M2 φ1 ⊳ M ⊲ φ2 hide X (M ) freeze ψ (M ) close(M ) M.X x X l M.X | l x|X ... expressions mixin structure sum rename hide freeze close projection ::= xi 7→ Xi ::= i∈I Xi 7→ xi ρ ::= i∈I xi 7→ Ei φ ::= Xi 7→ Yi i∈I renaming ::= i∈I tying E M C ι o ψ i∈I Xi 7→ Ei location projection, location identifier, name input assignment output assignment local binding Figure 2: Syntax for Lyre 13 ρ. The local binding ρ is a mapping from identifiers to expressions and corresponds to the body of the mixin; if x is in dom(ρ), then x is a defined component of the mixin with ρ(x) being the defining expression. ρ(x) can refer to both defined and deferred components of the mixin via identifiers. Any identifier in dom(ρ) and dom(ι) is bound in ρ(x). ρ(x) must not contain names. The input assignment ι is a mapping from identifiers to names and corresponds to declarations of deferred components; a deferred component internally referred to by x is resolved by the name ι(x). dom(ρ) and dom(ι) must be disjoint. The output assignment o is a mapping from names to identifiers; a component of a mixin externally accessed by X is associated to o(X) inside the mixin. ran(o) must be a subset of dom(ρ) ∪ dom(ι). Structures are identified up to α-renaming of identifiers. The explicit distinction between identifiers and names allows identifiers to be renamed by α-conversion, while names remain immutable, thus making projection by name unambiguous [Hirschowitz and Leroy(2005), Lillibridge and Harper(1994)]. In agreement with the distinction of identifiers and names, the rename operation takes two mappings on names, and the hide operation takes a name as argument. The freeze operation takes a mapping from names to expressions, where expressions may contain names. Mixin expressions contain locations. We use locations in the operational semantics to implement lazy evaluation, but locations will not appear in the surface language. The formalization is mostly independent of the core language. We only assume the core language includes projection from mixin expressions, locations, identifiers and names. Again locations will not appear in the surface language. In Figure 3, we define the operational semantics for Lyre. A heap state σ is a mapping from locations to heap objects, which are either expressions, values or error. We formalize lazy evaluation by suspending and memorizing evaluation in heap states. We assume given core values v. Then values V are either structures or core values. We syntactically distinguish structures as mixin expressions, surrounded by square brackets, and as values, surrounded by angle brackets. The distinction lets us simplify the formalization. The judgment σ ⊢ E ↓ (V ; σ2 ) means that in heap state σ expression E evaluates into value V with heap state being σ2 . We assume given inference rules to deduce σ ⊢ C ↓ (V ; σ2 ) for core expressions other than projections or locations. i∈I Notations We write σ[li 7→ κi ] to denote a mapping extension. Precisely,  κi when l′ = li for some i ∈ I ı∈I σ[li 7→ κi ](l′ ) = ′ σ(l ) otherwise The notation is defined only if, for any i, j ∈ I, i 6= j implies li 6= lj . We may write σ[l 7→ κ] when I is a singleton. The notation E[li /xi ]i∈I denotes the substitution of li ’s for xi ’s in E for all i ∈ I. The notation is defined only if, for any i, j ∈ I, i 6= j implies xi 6= xj . For input assignment ι, tying ψ and output i∈I assignment o = Xi 7→ xi , o ◦ ψ ◦ ι is the mapping ρ such that dom(ρ) = dom(ι) and, for all x ∈ dom(ι), ρ(x) is the expression obtained from ψ◦ι(x) by replacing 14 V κ σ ::= ::= ∈ hι; o; ρi | v E | error | V Loc -fin-> κ values heap objects heap state mixin structure (1) σ ⊢ [ι; o; ρ] ↓ (hι; o; ρi; σ) sum σ ⊢ M1 ↓ (hι1 ; o1 ; ρ1 i; σ2 ) σ2 ⊢ M2 ↓ (hι2 ; o2 ; ρ2 i; σ3 ) dom(ρ1 ) ∩ dom(ρ2 ) = ∅ dom(ι1 ) ∩ dom(ι2 ) = ∅ (2) σ ⊢ M1 ← M2 ↓ (hι1 + ι2 ; o1 + o2 ; ρ1 + ρ2 i; σ3 ) rename σ ⊢ M ↓ (hι; o; ρi; σ2 ) (3) σ ⊢ φ ⊳ M ⊲ φ′ ↓ (hφ ◦ ι; o ◦ φ′ ; ρi; σ2 ) hide σ ⊢ M ↓ (hι; o; ρi; σ2 ) (4) σ ⊢ hide X (M ) ↓ (hι; o\X; ρi; σ2 ) freeze σ ⊢ M ↓ (hι1 + ι2 ; o; ρi; σ2 ) dom(ψ) = ran(ι1 ) dom(ψ) ∩ ran(ι2 ) = ∅ (5) σ ⊢ freeze ψ (M ) ↓ (hι2 ; o; (ρ + (o ◦ ψ ◦ ι1 )i; σ2 ) close i∈I σ ⊢ M ↓ (h∅; o; xi 7→ Ei i; σ2 ) ∀i ∈ I, li fresh (6) i∈I i∈I σ ⊢ close(M ) ↓ (h∅; o; xi 7→ li i; σ2 [li 7→ Ei [lj /xj ]j∈I ]) projection σ ⊢ M ↓ (hι; o; ρi; σ2 ) σ2 ⊢ (ρ ◦ o)(X) ↓ (V ; σ3 ) (7) σ ⊢ M.X ↓ (V ; σ3 ) location σ(l) = V (8) σ ⊢ l ↓ (V ; σ) σ(l) = E σ[l 7→ error] ⊢ E ↓ (V ; σ2 ) (9) σ ⊢ l ↓ (V ; σ2 [l 7→ V ]) Figure 3: Semantics 15 Xi ’s with xi ’s for all i ∈ I. o ◦ ψ ◦ ι is only defined when ran(ι) ⊆ dom(ψ) and, for all x ∈ dom(ι), any name appearing in ψ ◦ ι(x) is in dom(o). Let’s look at the inference rules. A structure evaluates into itself and the heap state is unchanged (rule (1)). The sum operation merges the two operand mixins (rule (2)). The side conditions ensure that identifiers do not collide and are always satisfiable by taking appropriate α-equivalent mixins. The effect of the rename operation is simply compositions of mappings (rule (3)). The hide operation narrows the domain of the output assignment (rule (4)). The freeze operation resolves deferred components according to the tying ψ (rule (5)). Precisely, for all x ∈ dom(ι1 ), x is resolved to the expression o ◦ ψ ◦ ι1 (x). ψ ◦ ι1 (x) may contain names, which are replaced with identifiers by o. The rule augments the local binding with o ◦ ψ ◦ ι1 and the input assignment of the resulting mixin diminishes accordingly. The rule (6) for the close operation is responsible for making mixins lazy. Firstly the operand mixin expression must be evaluated into a structure without holes. Then, for each defined identifier xi , a fresh heap location li is allocated to store the defining expression Ei , where any occurrence of xj ’s in Ei is substituted by lj ’s. The body of the resulting structure maps xi to li , thus access to xi is redirected to li . To evaluate projection M.X (rule (7)), M is first evaluated into a structure hι; o; ρi. Then the rule consults o for the associated identifier to X, thus determines the expression that M.X accesses by looking up the identifier in ρ. In the normal evaluation, i.e. when projection is made from a closed mixin, ρ ◦ o(X) returns a location, where the defining expression of o(X) is stored. The last two rules are for evaluating locations l, and are fairly standard. If the heap state contains a value at l (rule (8)), then the value is returned and the heap state is unchanged. When an expression E is stored at l, then it is evaluated. The heap state maps l to error during the evaluation of E, to avoid evaluating the same expression repeatedly and to signal an error for cyclic definitions. On completion, the heap state is updated with the resulting value. It may be useful to note that we do not need evaluation rules for names or identifiers, since in the normal evaluation names are substituted by expressions and identifiers by locations. As one may have noticed, there is potential that evaluation gets stuck. One serious source is when projection is made from an open mixin. Instead of preventing such a scenario at the operational semantics level, we leave it to the type system to eliminate the possibilities for evaluation to get stuck. Although we do not present a type system in this paper, straightforward adaptation of previous work such as [Flatt and Felleisen(1998)] is sufficient for this purpose. The type system would be able to eliminate other ill-typed scenarios such as an attempt to merge mixins with overlapping output names, i.e., dom(o1 ) ∩ dom(o2 ) 6= ∅ in rule (2). We have omitted inference rules for propagating error states, assuming when evaluation encounters an error during deduction, it immediately terminates signaling a runtime error. 16 3.1 Call-by-name and eager variants Small modifications to evaluation rules let the operational semantics model callby-name and eager evaluation strategies. This subsection presents those variants. To model a call-by-name strategy, we replace rules (8) and (9) with the single rule: σ ⊢ σ(l) ↓ (V ; σ2 ) (10) σ ⊢ l ↓ (V ; σ2 ) Heap states are not updated, thus expressions stored are re-evaluated whenever accessed. By eager evaluation strategy, we mean an evaluation strategy that evaluates all components of a mixin at once when the mixin is closed. This is also easily implemented by replacing rule (6) with: i∈I σ ⊢ M ↓ (h∅; o; xi 7→ Ei i; σ2 ) I = {1, 2, . . . , n} i∈I 7 Ei [lj /xj ]j∈I ] ∀i ∈ I, li fresh σ ′ = σ2 [li → ′ ′ σ ⊢ l1 ↓ (V1 ; σ1′ ) σ1′ ⊢ l2 ↓ (V2 ; σ2′ ) · · · σn−1 ⊢ ln ↓ (Vn ; σn′ ) i∈I σ ⊢ close(M ) ↓ ([∅; o; xi 7→ li ]; σn′ ) 4 (11) Lazy mixins and disciplined effects We extend the operational semantics of the previous section so that it takes account of constraints on the evaluation order of components of mixins. In examples of this section, we use a side-effecting function “print C”, which prints the resulting value of evaluating C, then returns the value, to visualize the evaluation order. 4.1 Design space of evaluation strategies There are several evaluation strategies that we found interesting to consider. Below we explain those strategies which we have in mind. Firstly we consider the following example: mixin M1 = close( { let c1 = print 1 let c2 = print 2 let c3 = print 3 let c4 = print 4 }) Top-down strategy We may want components of a mixin to be evaluated following the textual definition order in which they appear in the source program. This constraint will ensure, in the above example, that 1 is printed before 2, and 2 is before 3, and 3 is before 4. This strategy is reminiscent of ML’s strategy. Lazy-field and lazy-record strategies If we adopt top-down strategy, we have a choice on whether to make accessible components of a mixin immediately after they are evaluated, while other components of the mixin are still to be evaluated 17 consecutively. We call lazy-field strategy the strategy that allows such access and lazy-record strategy the strategy that does not allow. Intuitively lazy-field strategy treats a closed mixin as a record with lazy fields like { a1 = lazy (print 1); a2 = lazy (print 2) }, where “{“ and “}” are record constructors of the core language here, while lazy-record strategy does as a lazy record like lazy ({ a1 = print 1; a2 = print 2 }). For instance, let’s consider executing the program: mixin M2 = close ({ let c1 = 1 let c2 = 2 ∗ M3.c1 }) mixin M3 = close ({ let c1 = 3 + M2.c1 }) let main = M2.c2 We assume, for the explanatory purpose, that a program consists of a sequence of mixin definitions plus a core value definition named main, where the top-level mixin bindings can be mutually recursive. The execution of the program evaluates the defining expression of main. For the execution of the above program to succeed, M2.c1 should be accessible to M3.c1 before evaluation of M2.c2 is completed. Hence the execution succeeds with lazy-field strategy, but it fails with lazy-record strategy. Clearly lazy-field strategy is more permissive. Although restrictive, lazy-record strategy is interesting to consider particularly in the presence of finer grained accessibility control, as we will see below. Internal and external accessibilities It can be useful to change accessibility to components of a mixin depending on whether the access is made inside the same mixin or outside. For instance, let us consider the following program: mixin M4 = close( { let c1 = 1 + 2 let c2 = c1 + 4 let c3 = print “ok” }) let main = M4.c2 For the execution to succeed, M4.c1 should be accessible to M4.c2 immediately after M4.c1 is evaluated but before evaluation of M4.c2 is completed. Hence we need lazy-field strategy inside M4. However we may want components of M4 to become accessible outside, only after all components of M4 have been evaluated. In other words, we may want lazy-record strategy outside M4 to make sure “ok” is necessarily printed before any component of M4 becomes externally accessible. Indeed, in the widget mixins example of Section 2.2, we envisaged this internally-lazy-field externally-lazy-record strategy. For instance, the component menu of MyGUI should be accessible inside once it is created, to configure the widget containment hierarchy via functions setMenuItems and setMenus. However menu should not be accessed outside MyGUI, before the configuration is completed, i.e. all components, including anonymous ones, of MyGUI are evaluated. Internally-lazy-field externally-lazy-record strategy is also potentially interesting in the light of our experience in programming with ML modules; it is sometimes convenient to include anonymous side-effecting expressions at the end of a structure for the debugging purpose or to properly initialize mutable components of the structure. As well as design choices among above strategies, we have a choice on how to propagate constraints when merging or closing mixins. In contrast to this vast design space, we do not have enough experience in programming with mixins. 18 Plainly we cannot determine at present which strategy is most beneficial in practice. Hence we keep our formalization open to strategies. That is, the formalization is abstracted over constraints on strategies and can be instantiate to express particular strategies. 4.2 The constraint language We use binary relations as our constraint language to express various evaluation strategies. We identify binary relations on any set P as sets of pairs of elements in P . For instance, the relation {(c1, c2) (c2, c3) (c3, c4)} expresses top-down strategy in the first example. The relation {(c1, c3) (c2, c4)} stipulates that c1 should be evaluated before c3 and c2 before c4. There is no constraint on the evaluation order between c1 and c2 or c3 and c2. Hence this relation does not fix the evaluation order in a unique way. For instance in the first example, either of the output “1 2 3 4” or “ 2 1 3 4” is compatible with the relation. To deal with more expressive strategies, we distinguish three sorts of identifiers: (ordinary) identifiers x for controlling the evaluation order; internal identifiers x̂ for internal accessibility; external identifiers x for external accessibility. For instance, the relation {(x1 , xˆ2 )} stipulates that the component x1 must be evaluated before the component x2 becomes accessible inside the same mixin. Similarly {(x1 , x2 )} stipulates that x1 must be evaluated before x2 becomes accessible outside. We will formalize how relations on these three sorts of identifiers control the evaluation order and accessibility later, when we define the operational semantics. Below we give a descriptive explanation on how those three sorts of identifiers can be used to express strategies we proposed above. If we assign the relation {(c1, c2)} as the constraint to M2 in the second example, the execution of the program succeeds. The relation stipulates nothing about how M2.c1 becomes accessible, implying M2.c1 is accessible both inside M2 and outside immediately after it is evaluated. If we assign the relation {(c1, c2), (c2, c1)} as the constraint to M2, the execution fails. The relation stipulates that M2.c1 becomes accessible outside M2 only after M2.c2 has been evaluated. However M3.c1 needs to access M2.c1 to be evaluated, and M2.c2 to M3.c1; there is a circular dependency. Top-down internally-lazy-field externally-lazy-record strategy in the third example is expressed by assigning the relation {(c1, c2), (c2, c3), (c3, c1), (c3, c2)} as the constraint to M4. We remark that there are implicit constraints ˆ and (c1, c1) (but not (c1, ˆ c1)). imposed by the semantics, such as (c1, c1) The reason is simple: the component c1 must be evaluated before it becomes accessible. Hence if the relation included (c1, c1), which introduces a circular dependency together with the implicit constraint (c1, c1), the evaluation would fail. Precisely, the operational semantics signals an error. The relation {(c1, c2), ˆ (c3, c2) ˆ (c3, c1), (c3, c2)} in the third example expresses top(c2, c3), (c3, c1), down internally-lazy-record externally-lazy-record strategy. We have omitted ˆ since it is anyway induced from (c2, c3) including constraints such as (c2, c1), ˆ by transitivity. Transitivity is imposed by the semantics: if c2 must and (c3, c1) ˆ then naturally c2 must be evaluated be evaluated before c3, and c3 before c1, 19 mixin structure (12) σ; Θ ⊢ [ι; o; ρ; θ] ↓ (hι; o; ρ; θi; σ; Θ) sum σ; Θ ⊢ M1 ↓ (hι1 ; o1 ; ρ1 ; θ1 i; σ2 ; Θ2 ) σ2 ; Θ2 ⊢ M2 ↓ (hι2 ; o2 ; ρ2 ; θ2 i; σ3 ; Θ3 ) dom(ρ1 ) ∩ dom(ρ2 ) = ∅ dom(ι1 ) ∩ dom(ι2 ) = ∅ (13) σ; Θ ⊢ M1 ← M2 ↓ (hι1 + ι2 ; o1 + o2 ; ρ1 + ρ2 ; ν(dom(ι1 ) ∪ dom(ρ1 ), θ1 , dom(ι2 ) ∪ dom(ρ2 ), θ2 )i; σ3 ; Θ3 ) rename σ; Θ ⊢ M ↓ (hι; o; ρ; θi; σ2 ; Θ2 ) (14) σ; Θ ⊢ φ1 ⊳ M ⊲ φ2 ↓ (hφ1 ◦ ι; o ◦ φ2 ; ρ; θi; σ2 ; Θ2 ) hide σ; Θ ⊢ M ↓ (hι; o; ρi; σ2 ; Θ2 ) (15) σ; Θ ⊢ hide X (M ) ↓ (hι; o\X; ρi; σ2 ; Θ2 ) freeze σ; Θ ⊢ M ↓ (hι1 + ι2 ; o; ρ; θi; σ2 ; Θ2 ) dom(ψ) = ran(ι1 ) dom(ψ) ∩ ran(ι2 ) = ∅ (16) σ; Θ ⊢ freeze ψ (M ) ↓ (hι2 ; o; ρ + (o ◦ ψ ◦ ι1 ); θi; σ2 ; Θ2 ) close i∈I σ; Θ ⊢ M ↓ (h∅; o; xi 7→ Ei ; θi; σ2 ; Θ2 ) ∀i ∈ I, li li′ li′′ fresh Θ3 = θ[li /xi ]i∈I [li′ /x̂i ]i∈I [li′′ /xi ]i∈I i∈I i∈I i∈I i∈I σ; Θ ⊢ close(M ) ↓ (h∅; o; xi 7→ li′′ ; µ({xi | i ∈ I}, θ)i; σ2 [li 7→ Ei [lj′ /xj ]j∈I ][li′ 7→ li ][li′′ 7→ li ]; Θ2 ∪ Θ3 ) projection σ; Θ ⊢ M ↓ (hι; o; ρ; θi; σ2 ; Θ2 ) σ2 ; Θ2 ⊢ (ρ ◦ o)(X) ↓ (V ; σ3 ; Θ3 ) (18) σ; Θ ⊢ M.X ↓ (V ; σ3 ; Θ3 ) location σ(l) = V (19) σ; Θ ⊢ l ↓ (V ; σ; Θ) (l′ , l) ∈ Θ σ(l) = E σ[l 7→ error]; Θ\{(l′, l)} ⊢ l′ ↓ (V ′ ; σ2 ; Θ2 ) σ; Θ ⊢ l ↓ (V ; σ3 ; Θ3 ) l 6≺ Θ σ2 [l 7→ E]; Θ2 ⊢ l ↓ (V ; σ3 ; Θ3 ) σ(l) = E σ[l 7→ error]; Θ ⊢ E ↓ (V ; σ ′ ; Θ′ ) (21) σ; Θ ⊢ l ↓ (V ; σ ′ [l 7→ V ]; Θ′ ) Figure 4: Semantics with constraint ˆ before c1. 4.3 Operational semantics We extend a structure with a binary relation on ordinary, internal and external identifiers. Precisely, a structure [ι; o; ρ; θ] is now 4-tuple of an input assignment ι, output assignment o, local binding ρ, and local constraint θ, where θ is a binary relation on dom(ι) ∪ dom(ρ) ∪ {x | x ∈ dom(ι) ∪ dom(ρ)} ∪ {x̂ | x ∈ dom(ι) ∪ dom(ρ)}. Otherwise the syntax is unchanged from Figure 2. We use X as a metavariable for sets of ordinary, internal and external identifiers. We assume given two functions µ(X1 , θ1 ) and ν(X1 , θ1 , X2 , θ2 ), where θ1 and θ2 are binary relations on X1 and X2 respectively. µ(X1 , θ1 ) returns a binary relation on X1 and ν(X1 , θ1 , X2 , θ2 ) does on X1 ∪ X2 . The operational semantics uses µ(X1 , θ1 ) to build local constraints for mixins instantiated by the close operation, where X1 is the set of deferred and defined identifiers and θ1 the 20 (17) (20) local constraint of the operand mixin. Similarly it uses ν(X1 , θ1 , X2 , θ2 ) to build local constraints for mixins composed by the sum operation, where X1 (resp. X2 ) is the set of deferred and defined identifiers and θ1 (resp. θ2 ) is the local constraints of the left (resp. right) right operand mixin. By not fixing the interpretations of µ or ν, we keep the formalization neutral of how to propagate local constraints when closing or merging mixins. In Figure 4, we present the operational semantics which takes account of constraints. We use Θ as a metavariable for global constraints, or binary relations on locations. The judgment σ; Θ ⊢ E ↓ (V ; σ ′ ; Θ′ ) means that under global constraint Θ with heap state σ expression E evaluates into V , where the heap state and global constraint have evolved into σ ′ and Θ′ . As well as heap states, global constraints evolve during evaluation, since new constraints are added when a mixin is closed (rule (17)). The notation Θ\{(l′ , l)} denotes set subtraction. The notation l 6≺ Θ denotes the condition that there is not l′ such that (l′ , l) ∈ Θ. Compared to previous evaluation rules in Figure 3, modifications are made on rules (2) (6), and (9), which we explain in turn. As explained above, the sum operation uses the function ν to build a local constraint for the composed mixin (rule (13)). When closing a mixin (rule (17)), three fresh locations li , li′ , li′′ are created for each defined component xi of the mixin, in order to separately control the evaluation order by li ’s, internal accessibility by li′ ’s, and external accessibility by li′′ ’s. Four important points to be understood are as follows: 1) The local binding of the resulting mixin maps xi ’s to li′′ ’s, i.e. locations for external access; 2) Evaluation of Ei is suspended at li , where xi ’s in Ei are substituted by li′ ’s, i.e. locations for internal access; 3) The new heap state maps li′ ’s and li′′ ’s for internal and external accesses to li ’s to connect the accesses to underlying expressions; 4) The local constraint θ is transported to the global constraint, by instantiating ordinary, internal and external identifiers to the corresponding locations. The notation θ[li /xi ]i∈I [li′ /x̂i ]i∈I [li′′ /xi ]i∈I denotes the binary relation on locations obtained from θ by substituting li ’s for xi ’s , li′ ’s for x̂i ’s, and li′′ ’s for xi ’s, where the substitution is performed by regarding a binary relation as a set of pairs. The local constraint for the resulting mixin is build by the function µ. We have introduced two new rules (20) and (21) for evaluating locations, replacing the previous rule (9), to take the global constraint into account. Rule (20) considers the case where the global constraint stipulates that l′ should be evaluated before l, as described by the side-condition (l′ , l) ∈ Θ. The rule evaluates l′ first, while updating the heap state to map l to error and removing (l′ , l) from Θ; if the evaluation of E ′ attempts to evaluate l against the global constraint, an error is signaled. On completion of the evaluation, E is restored at l and the rule retries to evaluate l. The same rule (20) may be applied again, when there is another location which should be evaluated before l. Otherwise rule (21) is applicable. When there is no location which must be evaluated before l, as described by the side condition l 6≺ Θ in rule (21), then E is evaluated immediately. 21 It shall be informative to note that in rule (20) the heap state σ is updated with error to enforce the evaluation order stipulated by the constraint, while in rule (21) to detect ill-founded recursion. An important ingredient of the formalization is that evaluation of projection M.X always goes through a location. This facilitates to control the evaluation order of components of mixins, in terms of a binary relation on locations. It is easily proved that if σ(l) = V then l 6≺ Θ. We also remark that the operational semantics, in particular the evaluation order, is not necessarily deterministic, depending on the constraint. For instance, if we assign a local constraint {(c1, c3), (c2, c4)} to the mixin M1 of the first example in Section 4.1, we can build deductions which result in the outputs “1 2 3 4” and “2 1 3 4”. Non-determinism is not abnormal, but should be thought of as underspecification like the underspecified evaluation order of function parameters in some programming languages. The operational semantics can be proved sane in the sense that evaluation does not diverge due to ill-founded recursion. In other words, when evaluation diverges, heap states are extended infinitely. A heap state is extended only when a mixin is closed (rule (17)). Hence the heap explosion implies that the close operation is used infinitely often. This is similar to a situation where infinite recursion of function calls exhausts the stack. 4.4 An example We present an example by instantiating the constraint to express one particular evaluation strategy. The strategy is motivated by our previous work on examining a lazy evaluation strategy for recursive ML-style module [Nakata(2009)]. The purpose of the example is not to advocate this particular strategy, but to deliver the better intuition about how to use the constraint. For a structure [ι; o; ρ; θ], we let θ be the relation: {(xi , xj ) | xi ∈ dom(ρ) ∩ IdsC , xj ∈ dom(ρ), i < j} ∪ {(xi , xj ) | xi ∈ (dom(ι) ∪ dom(ρ)) ∩ IdsC , xj ∈ dom(ι) ∪ dom(ρ)} Above we have assumed for any xi , xj , if i < j then the definition or declaration for xi textually precedes that for xj in the source program. IdsC denotes the set of identifiers bounds to core expressions. I.e. we assume Idents consists of disjoint sets of identifiers to be bound to core expressions and mixin expressions, respectively. We give interpretations to functions µ and ν as follows: µ(X , θ) = ∅ ν(X , θ, X ′ , θ′ ) = θ ∪ θ′ ∪ {(xi , xj | xi ∈ (X ∪ X ′ ) ∩ IdsC , xj ∈ X ∪ X ′ } The strategy is close to top-down internally-lazy-field externally-lazy-record strategy, where sub-mixins are evaluated lazily. Below we explain the strategy in detail. 1. Preceding core components of the same and enclosing mixins are evaluated 22 first. Enclosing mixins may contain other sub-mixins, whose core fields need not be evaluated first. This is enforced by including in the local constraint of a structure the relation {(xi , xj ) | xi ∈ dom(ρ) ∩ IdsC , xj ∈ dom(ρ), i < j}. Note that xj may be bound to a mixin expression, and in order to evaluate the expression all core components preceding to it must have been evaluated. Thus the constraint controls the evaluation order not only of core components of the same mixin but also of those of enclosing mixins. 2. Components of a mixin become accessible inside the same mixin immediately after they are evaluated, but are accessible outside only after all the core components have been evaluated. This is enforced by not having constraints mentioning internal identifiers and by including in the local constraint of a structure the relation {(xi , xj ) | xi ∈ (dom(ι) ∪ dom(ρ)) ∩ IdsC , xj ∈ dom(ι) ∪ dom(ρ)}. Observe how we propagate the constraint on external accessibility when merging mixins by including the relation {(xi , xj | xi ∈ (X ∪ X ′ ) ∩ IdsC , xj ∈ X ∪ X ′ } in the result of ν. 3. While core components are evaluated following the textual definition order, there is no constraint on the evaluation order between components of merged mixins; ν does not introduce (xi , xj ) or (xi , xj ) for any xi ∈ X and xj ∈ X ′ . 4. We do not leave any constraint after a mixin is closed; it is enough to enforce the evaluate order once. This is reflected in the interpretation of µ. As we have said, we do not intend to justify ourselves in choosing this strategy, because we do not have enough programming experience to do so. Nevertheless we motivate the strategy below, aiming at posing questions to readers about possible concerns one may face in the quest of better design choices. The first condition ensures that backward references to core components of the same and enclosing mixins are necessarily proper values, but not suspensions or error’s. This provides programmers with a safety guarantee on backward references to core components, which seems useful for ML-initiated programmers. Interestingly this design choice also bears a similarity to Java’s class initialization policy [Gosling et al.(2005)Gosling, Joy, Steele, and Bracha]. We have already motivated in Section 4.1 the combination of internally-lazyfield and externally-lazy-record strategies; internally-lazy-field strategy leaves flexibility in intra-mixin recursion, while externally-lazy-record strategy keeps the evaluation stable towards outside. A possible drawback of externally-lazyrecord strategy, compared to externally-lazy-field strategy, is that we may loose flexibility in inter-mixin recursion. However we are less concerned by inter-mixin recursion, since we believe mixins are designated to support flexible intra-mixin recursion; notably the sum and freeze operations are useful for taking fix-points inside a mixin. We are less confident in the choice of ν. But it could be useful to keep the evaluation order independent between separately defined mixins; it is unlikely 23 that a programmer can foresee in which order components of other mixins to be merged with should/could be evaluated. 4.5 An extension There is a strategy which we want to consider, but our constraint language is not expressive enough for it. The strategy is a variant on top-down internallylazy-field externally-lazy-field strategy, where we impose an extra constraint, called trigger-constraint, that all components of a mixin must be evaluated at once before the first access to a component of the mixin returns. For instance, let us consider the following program: mixin M1 = { let c1 = print 1 let c2 = M2.c2 let c3 = print (c1 + c2) let c4 = print 5 } mixin M2 = { let c1 = M1.c1 let c2 = print (c1 + 1) let c3 = print 4 } let main = M1.c3 According to the above proposed strategy, the execution succeeds and “1 2 4 3 5” is printed. (Recall that “print (1+2)” returns 3, after printing 3). Here is why. 1. Top-down strategy ensures the printing orders “1 3 5” and “2 4” independently. 2. Thanks to internally-lazy-field strategy, M1.c3 and M2.c2 are successfully evaluated. 3. Thanks to externally-lazy-field strategy, evaluation of M2.c1 succeeds. Note that M2 is forced to evaluate by the access from M1.c2. Hence with externally-lazy-record strategy, the evaluation fails, since M1.c1 is not yet accessible outside then. 4. Finally and importantly, the trigger-constraint ensures that M2.c3 is evaluated before evaluation of M1.c2 returns and that M1.c4 is before evaluation of main returns. This explains why 4 is printed immediately after 2, and 5 is after 3. This strategy comes from our previous work on lazy recursive modules [Nakata(2009)]. In a recursive modules setting, inter-module recursion is important, hence we wanted externally-lazy-field strategy. We found the trigger-constraint useful to enforce our design policy that once a module is accessed, all its component are eventually evaluated. We want to consider the same strategy as our previous proposal in a mixin context, too. Observe that the trigger-constraint only lets the access to M1.c3 trigger evaluation of M1.c4, but M1.c1 is already accessible both inside and outside once it is evaluated. We cannot include any of conˆ in the local constraint of M1, since the first straints (c4, c3), (c4, c3), or (c4, c3) one is inconsistent with top-down strategy, the second with externally-lazy-field strategy, the third with internally-lazy-field strategy. 24 Π = (Θ, ∆) {l, l1 . . . ln } ∈ ∆ σ; (Θ, ∆\{{l, l1 . . . ln }}) ⊢ l ↓ (V ; σ ′ ; Π′ ) ′ ; Π′n−1 ⊢ ln ↓ (Vn ; σn′ ; Π′n ) σ ; Π ⊢ l1 ↓ (V1 ; σ1′ ; Π′1 ) σ1′ ; Π′1 ⊢ l2 ↓ (V2 ; σ2′ ; Π′2 ) · · · σn−1 ′ ′ σ; Π ⊢ l ↓ (V ; σn′ ; Π′n ) (22) l 6∈∈ ∆ (l′ , l) ∈ Θ σ(l) = E σ[l 7→ error]; (Θ\{(l′ , l)}, ∆) ⊢ l′ ↓ (V ′ ; σ2 ; Π2 ) σ2 [l 7→ E]; Π2 ⊢ l ↓ (V ; σ3 ; Π3 ) (23) σ; (Θ, ∆) ⊢ l ↓ (V ; σ3 ; Π2 ) Π = (Θ, ∆) l 6∈∈ ∆ l 6≺ Θ σ(l) = E σ[l 7→ error]; Π ⊢ E ↓ (V ; σ ′ ; Π′ ) (24) σ; Π ⊢ l ↓ (V ; σ ′ [l 7→ V ]; Π′ ) i∈I σ; Π ⊢ M ↓ (h∅; o; xi 7→ Ei ; πi; σ2 ; Π2 ) π = (θ, δ) Π2 = (Θ, ∆) ∀i ∈ I, li li′ li′′ fresh Θ′ = θ[li /xi ]i∈I [li′ /x̂i ]i∈I [li′′ /xi ]i∈I ∆′ = δ[li /xi ]i∈I [li′ /x̂i ]i∈I [li′′ /xi ]i∈I i∈I i∈I i∈I i∈I σ; Π ⊢ close(M ) ↓ (h∅; o; xi 7→ li′′ ; µ({xi | i ∈ I}, π)i; σ2 [li 7→ Ei [lj′ /xj ]j∈I ][li′ 7→ li ][li′′ 7→ li ]; (Θ ∪ Θ′ , ∆ ∪ ∆′ )) Figure 5: Extension with trigger-constraint 4.5.1 Formalization To gain extra expressivity to deal with the trigger-constraint, we extend a local constraint to be a pair of a binary relation on identifiers and a set of sets of identifiers. Precisely, a structure [ι; o; ρ; π] now contains a local constraint π, which is a pair (θ, δ) of a binary relation θ on dom(ι) ∪ dom(ρ) ∪ {x | x ∈ dom(ι) ∪ dom(ρ)} ∪ {x̂ | x ∈ dom(ι) ∪ dom(ρ)} and a set δ of sets of elements in dom(ι) ∪ dom(ρ). θ expresses constraints on the evaluation order and accessibilities as before. δ expresses the trigger-constraint; if X is in δ, then evaluation of all components bound to identifiers in X is triggered at once when any of the components is accessed for the first time. For instance when δ contains dom(ρ), then all defined components of the structure, but deferred ones, are evaluated at once when any of the defined components is accessed for the first time. Accordingly, a global constraint Π is now a pair (Θ, ∆) of a binary relation Θ on locations and set ∆ of sets of locations. The functions µ and ν need to be extended to operate on pairs of a binary relation and set of sets. To take account of the trigger-constraint, We replace rules (20) and (21) by rules (22), (23) and (24), and rule (17) by rule (25) as given in Figure 5. The notation l 6∈∈ ∆ denotes the condition that ∆ does not contain a set containing l. The additional work is to check, before evaluating a location l, whether there are locations whose evaluation is triggered by l. Rule (22) considers the case where l triggers evaluation of li ’s, as described by the side condition {l, l1 . . . ln } ∈ ∆. The rule bears responsibility for evaluating li ’s and {l, l1 . . . ln } is discharged from ∆. Except for the side condition l 6∈∈ ∆, rules (23) and (24) are identical to previous rules (20) and (21); the side condition ensures that all trigger constraints involving l have been handled. It is easily proved that (19) and (22) are exclusive to each other. Rule (25) is a technical adjustment, since the rule now needs to transport a pair of a binary relation and set of sets from the local constraint to the global constraint. The notation δ[li /xi ]i∈I [li′ /x̂i ]i∈I [li′′ /xi ]i∈I denotes the set of sets of locations obtained from δ by substituting li ’s for xi ’s , li′ ’s for x̂i ’s, and li′′ ’s for xi ’s. We do not repeat 25 (25) class A { val a1 = ... } class B extends A { val b1 = ... val b2 = ... } class C extends B { val c1 = .. } let c = new C Figure 6: A class hierarchy the other rules; the necessary modification is to replace Θ’s by Π’s. 4.5.2 An example As an example of how to use the trigger-constraint, we present a possible object initialization strategy in class-based object-oriented languages. The purpose of the example is not to explain an encoding of object-oriented language features such as inheritance and overriding, which is already examined in previous work [Ancona and Zucca(2002)]. Here we focus on the aspect of object initialization, by regarding components of a mixin as (instance) fields of an object. In this scenario, open mixins correspond to classes and closed mixins to objects. We shall not need the full expressivity of Lyre to model standard class-based object-oriented languages. For specificity, we restrict ourselves to a fragment of the calculus satisfying the following conditions: • Structures do not contain sub-mixins. This implies we do not consider inner classes. • The sum operation only takes open mixins as operands. The operation corresponds to inheritance. Then this is standard, since usually classes do not inherit from objects, or vice versa. In the sum construct M1 ← M2 , we assume the left operand mixin M1 corresponds to the superclass and the right operand mixin M2 to the inheriting class. For a structure [ι; o; ρ; (θ, δ)], we assign internally-lazy-record externallylazy-record strategy. The initialization order of fields within an object is kept unspecified. That is, we let θ be the relation: {(xi , xj ) | xi , xj ∈ (dom(ι) ∪ dom(ρ))} ∪ {(xi , xˆj ) | xi , xj ∈ (dom(ι) ∪ dom(ρ))} We let δ be the singleton: {(dom(ι) ∪ dom(ρ))} The trigger-constraint δ above stipulates that all the fields of an object must be initialized at once. We give interpretations to functions µ and ν as follows: µ(X , π) = (∅, ∅) ν(X , (θ, δ), X ′ , (θ′ , δ ′ )) = (θ ∪ θ′ ∪ {(xi , xj ) | xi ∈ X , xj ∈ X ′ }, {(X ∪ X ′ )}) 26 The interpretation of ν makes the strategy interesting. Viewed as object initialization, the strategy is described as follows. 1. Fields are initialized following the class hierarchy. That is, fields of superclasses are initialized before those of sub-classes. For instance in Figure 6, the field a1 of the object c is initialized before b1 or b2, and b1 and b2 are before c1. This is imposed by including the constraint {(xi , xj ) | xi ∈ X , xj ∈ X ′ } in the result of the first element of ν. 2. Any field inherited from a superclass becomes accessible both inside and outside, once all the fields from the superclass are initialized. Hence in the example, c.b1 and c.b2 become accessible immediately after both have been initialized, but before c.c1 is initialized. Note that neither of c.b1 nor c.b2 is accessible even inside before both have been initialized. This is specified by the local constraint assigned to a structure. 3. We assume the new operation for instantiating objects from classes is equivalent to the close operation followed by access to some field of the mixin closed. For instance we may assume the Object class which resides in the root of the class hierarchy, i.e. a superclass of any class, contains a special field named init for that purpose. Then the strategy enforces the standard object initialization policy where all fields of an object are initialized when it is created. This is where we use the trigger-constraint. The result {(X ∪ X ′ )} of the second element of ν ensures that all field of an object are initialized at once when init is accessed, since it triggers initialization of the fields of inheriting classes, when the init field is initialized. In fact this strategy, in particular the characteristic described second above, is inspired by the object initialization strategy of the F# programming language [Syme and Margetson(2009)] 5 . F# places restrictions on possible object initialization patterns, enhancing safety by eliminating programming styles which are often awkward sources for null-pointer exceptions. As a result, the strategy is less flexible than that of some object-oriented languages such as Java, yet it can still support common programming idioms found in object-oriented programming. In short, we found F#’s strategy an interesting example of disciplined evaluation orders. 5 Related Work One difference between previous work and the present work is that each of previous work examined an evaluation strategy, while we explored the design space of lazy evaluation strategies and give the operational semantics which can deal with several strategies. Most related to our work is Ancona and Zucca’s call-by-name 5 We believe the strategy presented is close to F#’s. Yet the exact strategy is under-specified in the language documentation. 27 mixin calculus [Ancona and Zucca(2002)] and Hirschowitz and Leroy’s call-byvalue mixin calculus [Hirschowitz and Leroy(2005)]. Ancona et al. investigated the interaction between mixins and computational effect, using a monadic metalanguage as semantic basis [Ancona et al.(2003)Ancona, Fagorzi, Moggi, and Zucca]. We will review the three in detail below. S. Fagorzi and E. Zucca proposed Rcalculus to allow projection from open mixins in a consistent way [Fagorzi and Zucca(2007)]; this is a design direction we have not explored. Our previous work proposed a lazy evaluation strategy for recursive ML-style modules, which we explained in Section 4.5. Design questions we encountered there motivated strategies we proposed in Section 4.1. We do not look back at the history of mixins, but only mention a few of influential papers [Bracha(1992), Ancona and Zucca(2002), Hirschowitz and Leroy(2005), Duggan and Sourelis(1996), Flatt and Felleisen(1998)]. In particular, type systems which eliminate unsound scenarios such as to close mixins having holes or to select a component from an open mixin, have been well-investigated. Those type system proposals are orthogonal and complementary to the presented work. Call-by-name mixins Ancona and Zucca formalized a mixin calculus, named CMS, with call-by-name evaluation using small-step semantics [Ancona and Zucca(2002)]. The formalization is concise and allows equational reasoning of mixins. However the call-by-name semantics might not be suitable to be used with a call-by-value core language allowing arbitrary side-effects, such as the ML core language, since with the call-by-name semantics evaluation of let rec x = x diverges when x is selected and the same side-effect can be produced repeatedly. Large part of the formalization in Section 3 is borrowed from CMS. We adapted their small-step semantics to big-step semantics. Technically we made a small but important modification; in CMS an output assignment maps names to expressions, whereas in Lyre names to identifiers. In this way we avoid duplicating expressions in the freeze operation. Call-by-value mixins Hirschowitz and Leroy formalized a mixin calculus with call-by-value evaluation [Hirschowitz and Leroy(2005)]. Generally call-by-value mixins result in the simplest evaluation order, while lazy mixins are more flexible in handling recursion. For instance, the marshallers example in Section 2.3 relies on the laziness to specify marshallers for recursive data types. We are motivated to distinguish open and closed mixins by their work. Apart from the difference of call-by-value and lazy, Lyre differs from their calculus in the sum operation in that we allow mixins to be merged independently of whether they are closed or open, while they only consider the sum operation on open mixins. Our design choice is motivated to keep flexibility in sharing side effects. For instance, it was useful in MakeSet and MakeMultiSet mixins example from Section 2.1; the sharing of counter between Set and MultiSet is achieved by merging the closed mixin Key with MakeSet and MakeMultiSet. It should be noted that one important objective of Herschowitz and Leroy’s work is to statically ensure initialization safety of mixins. The objective of our paper is not about static guarantees of initialization safety. 28 Effectful mixins and equational reasoning Ancona et al. examined the interaction between mixins and computational effect, by means of a recursive monadic binding [Ancona et al.(2003)Ancona, Fagorzi, Moggi, and Zucca]. The sum operation only takes open mixins as argument in their calculus. They separate computational components from non-computational ones, where the former are evaluated exactly once, while the latter are re-evaluated whenever they are projected. Computational components of a mixin are evaluated at once when the mixin is closed by the doall operation, in a similar way to the rule (11) from Section 3.1. To the best of our understanding, their evaluation strategy is top-down internally-lazy-field externally-lazy-record strategy, where closed submixins are evaluated immediately. When mixins are merged, the computational components of the right-operand mixin are evaluated before those of the leftoperand mixin. The separation of computational components lets them retain the CMS equational reasoning. 6 Conclusion We have formalized the operational semantics for a mixin calculus with lazy evaluation. We started by considering a simpler semantics where a component of a mixin is evaluated when it is projected. Then we extended the semantics to impose various restrictions on the evaluation order of and accessibilities to components of mixins, by adding local constraint to mixin structures and global constraint to the evaluation judgment. We have kept the formalization neutral of constraints and it can be instantiated to express several evaluation strategies. As well as we considered the design space of strategies, we took a closer look at two particular strategies. We do not claim the formalization is expressive enough to deal with all interesting evaluation strategies. At the same time, we do not think it technically difficult to extend the formalization to express more strategies. As the extension we made in Section 4.5 exemplifies, we may well gain more expressivity by extending the constraint language and by adding more evaluation rules. Indeed a moot point was to keep the formalization not too complicated without giving up too much expressivity to deal with interesting strategies. Acknowledgement I am grateful to Xavier Leroy for the valuable advice and fruitful discussions throughout the development of this work. References [Ancona and Zucca(2002)] D. Ancona and E. Zucca. A calculus of module systems. Journal of Functional Programming, 12, 2002. [Ancona et al.(2003)Ancona, Fagorzi, Moggi, and Zucca] D. Ancona, S. Fagorzi, E. Moggi, and E. Zucca. Mixin Modules and Computational Effects. In Proc. ICALP, 2003. 29 [Bracha(1992)] G. Bracha. The Programming Language Jigsaw: Mixins, Modularity and Multiple inheritance. PhD thesis, University of Utah, 1992. [Cardelli(1997)] L. Cardelli. Program Fragments, Linking, and Modularization. In Proc. POPL, pages 266–277, 1997. [Crary et al.(1999)Crary, Harper, and Puri] K. Crary, R. Harper, and S. Puri. What is a recursive module? In Proc. PLDI, pages 50–63, 1999. [Duggan and Sourelis(1996)] D. Duggan and C. Sourelis. Mixin modules. In Proc. ICFP, 1996. [Fagorzi and Zucca(2007)] S. Fagorzi and E. Zucca. A Calculus of Open Modules: Call-by-need Strategy and Confluence,. Mathematical Structures in Computer Science, 17, 2007. [Fisher and Reppy(1999)] K. Fisher and J. Reppy. The design of a class mechanism for moby. In Proc. PLDI. ACM Press, 1999. [Flatt and Felleisen(1998)] M. Flatt and M. Felleisen. Units: Cool Modules for HOT Languages. In Proc. PLDI. ACM Press, 1998. [Gosling et al.(2005)Gosling, Joy, Steele, and Bracha] J. Gosling, B. Joy, G. Steele, and G. Bracha. The Java Language Specification, Third Edition, chapter 12.4. Prentice Hall, 2005. [Hirschowitz and Leroy(2005)] T. Hirschowitz and X. Leroy. Mixin modules in a call-by-value setting. ACM Transactions on Programming Languages and Systems, 27(5):857–881, 2005. [Kennedy(2004)] A. Kennedy. Functional Pearls: Pickler Combinators. Journal of Functional Programming, 14(6):727–739, 2004. [Lillibridge and Harper(1994)] M. Lillibridge and R. Harper. A Type-Theoretic Approach to Higher-Order Modules with Sharing. In Proc. POPL, pages 123–137, 1994. [Nakata(2009)] K. Nakata. Lazy modules. http://cs.ioc.ee/~keiko/, 2009. Draft paper available at [Odersky et al.(2003)Odersky, Cremet, Röckl, and Zenger] M. Odersky, V. Cremet, C. Röckl, and M. Zenger. A nominal theory of objects with dependent types. In Proc. ECOOP, 2003. [Owens and Flatt(2006)] S. Owens and M. Flatt. From Structures and Functors to Modules and Units. In Proc. ICFP, 2006. [Russo(2001)] C. Russo. Recursive Structures for Standard ML. In Proc. ICFP, pages 50–61. ACM Press, 2001. 30 [Syme(2005)] D. Syme. Initializing mutually referential abstract objects: The value recursion challenge. In Proc. Workshop on ML, 2005. [Syme and Margetson(2009)] D. Syme and J. Margetson. The F# Programming Language. Software and documentation available on the Web, http://research.microsoft.com/en-us/um/cambridge/projects/fsharp/ [Torgersen(2004)] M. Torgersen. The Expression Problem Revisited. In Proc. ECOOP, volume 3086, 2004. 31
6cs.PL
On the Transformation Capability of Feasible Mechanisms for Programmable Matter Othon Michail1 , George Skretas2 , and Paul G. Spirakis3 1 2 arXiv:1703.04381v1 [cs.DS] 13 Mar 2017 3 Department of Computer Science, University of Liverpool, UK [email protected] Computer Engineering and Informatics Department (CEID), University of Patras, Greece [email protected] Department of Computer Science, University of Liverpool, UK Computer Technology Institute and Press “Diophantus” (CTI), Patras, Greece [email protected] Abstract In this work, we study theoretical models of programmable matter systems. The systems under consideration consist of spherical modules, kept together by magnetic forces and able to perform two minimal mechanical operations (or movements): rotate around a neighbor and slide over a line. In terms of modeling, there are n nodes arranged in a 2-dimensional grid and forming some initial shape. The goal is for the initial shape A to transform to some target shape B by a sequence of movements. Most of the paper focuses on transformability questions, meaning whether it is in principle feasible to transform a given shape to another. We first consider the case in which only rotation is available to the nodes. Our main result is that deciding whether two given shapes A and B can be transformed to each other, is in P. We then insist on rotation only and impose the restriction that the nodes must maintain global connectivity throughout the transformation. We prove that the corresponding transformability question is in PSPACE and study the problem of determining the minimum seeds that can make feasible, otherwise infeasible transformations. Next we allow both rotations and slidings and prove universality: any two connected shapes A, B of the same order, can be transformed to each other without breaking connectivity. The worst-case number of movements of the generic strategy is Ω(n2 ). We improve this to O(n) parallel time, by a pipelining strategy, and prove optimality of both by matching lower bounds. In the last part of the paper, we turn our attention to distributed transformations. The nodes are now distributed processes able to perform communicate-compute-move rounds. We provide distributed algorithms for a general type of transformations. Keywords and phrases programmable matter, transformation, reconfigurable robotics, shape formation, complexity, distributed algorithms 1 Introduction Programmable matter refers to any type of matter that can algorithmically change its physical properties. “Algorithmically” means that the change (or transformation) is the result of executing an underlying program. Depending on the implementation, the program could either be a centralized algorithm capable of controlling the whole programmable matter system (external control) or a decentralized protocol stored in the material itself and executed by various sub-components of the system (internal control). For a concrete example, imagine a material formed by a collection of spherical nanomodules kept together by magnetic forces. Each module is capable of storing (in some internal representation) and executing a simple program that handles communication with nearby modules and that controls the module’s electromagnets, in a way that allows the module to rotate or slide over neighboring modules. © Othon Michail, George Skretas, and Paul G. Spirakis; licensed under Creative Commons License CC-BY Leibniz International Proceedings in Informatics Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany 2 On the Transformation Capability of Feasible Mechanisms for Programmable Matter Such a material would be able to adjust its shape in a programmable way. Other examples of physical properties of interest for real applications would be connectivity, color [27, 5], and strength of the material. Computer scientists, nanoscientists, and engineers are more and more joining their forces towards the development of such programmable materials and have already produced some first impressive outcomes (even though it is evident that there is much more work to be done in the direction of real systems), such as programmed DNA molecules that self-assemble into desired structures [30, 13] and large collectives of tiny identical robots that orchestrate resembling a single multi-robot organism (Kilobot system) [31]. Other systems for programmable matter include the Robot Pebbles [21], consisting of 1cm cubic programmable matter modules able to form 2-dimensional (usually abbreviated “2D”) shapes through selfdisassembly, and the Millimotein [24], a chain of programmable matter which can fold itself into digitized approximations of arbitrary 3-dimensional (usually abbreviated “3D”) shapes. Ambitious long-term applications of programmable materials include molecular computers, collectives of nanorobots injected into the human circulatory system for monitoring and treating diseases, or even self-reproducing and self-healing machines. Apart from the fact that systems work is still in its infancy, there is also an apparent lack of unifying formalism and theoretical treatment. The following are some of the very few exceptions aiming at understanding the fundamental possibilities and limitations of this prospective. The area of algorithmic self-assembly tries to understand how to program molecules (mainly DNA strands) to manipulate themselves, grow into machines and at the same time control their own growth [13]. The theoretical model guiding the study in algorithmic self-assembly is the Abstract Tile Assembly Model (aTAM) [35, 29] and variations. Recently, a model, called the nubot model, was proposed for studying the complexity of self-assembled structures with active molecular components [36]. This model “is inspired by biology’s fantastic ability to assemble biomolecules that form systems with complicated structure and dynamics, from molecular motors that walk on rigid tracks and proteins that dynamically alter the structure of the cell during mitosis, to embryonic development where large-scale complicated organisms efficiently grow from a single cell” [36]. Another very recent model, called the Network Constructors model, studied what stable networks can be constructed by a population of finite-automata that interact randomly like molecules in a well-mixed solution and can establish bonds with each other according to the rules of a common small protocol [28]. The development of Network Constructors was based on the Population Protocol model of Angluin et al. [2], that does not include the capability of creating bonds and focuses more on the computation of functions on inputs. A very interesting fact about population protocols is that they are formally equivalent to chemical reaction networks (CRNs), “which model chemistry in a well-mixed solution and are widely used to describe information processing occurring in natural cellular regulatory networks” [14]. Also the recently proposed Amoebot model, “offers a versatile framework to model self-organizing particles and facilitates rigorous algorithmic research in the area of programmable matter” [10, 12, 11]. An indication of the potential that the research community sees in this effort, is the 1st Dagstuhl Seminar on “Algorithmic Foundations of Programmable Matter”, which took place in June 2016 and attracted leading scientist (both theoreticians and practitioners) from Algorithms, Distributed Computing, Robotics, and DNA Self-Assembly, with the aim at joining their forces to push forward this emerging subject. Each theoretical approach, and to be more precise, each individual model, has its own beauty and has lead to different insights and developments regarding potential programmable matter systems of the future and in some cases to very intriguing technical problems and O. Michail, G. Skretas, and P. G. Spirakis open questions. Still, it seems that the right way for theory to boost the development of more refined real systems is to reveal the transformation capabilities of mechanisms and technologies that are available now, rather than by exploring the unlimited variety of theoretical models that are not expected to correspond to a real implementation in the near future. In this paper, we follow such an approach, by studying the transformation capabilities of models for programmable matter, which are based on minimal mechanical capabilities, easily implementable by existing technology. 1.1 Our Approach We study a minimal programmable matter system consisting of n cycle-shaped modules, with each module (or node) occupying at any given time a cell of the 2D grid (no two nodes can occupy the same cell at the same time). Therefore, the composition of the programmable matter systems under consideration is discrete. Our main question throughout is whether an initial arrangement of the material can transform (either in principle, e.g., by an external authority, or by itself) to some other target arrangement. In more technical terms, we are provided with an initial shape A and a target shape B and we are asked whether A can be transformed to B via a sequence of valid transformation steps. Usually, a step consists either of a valid movement of a single node (in the sequential case) or of more than one nodes at the same time (in the parallel case). We consider two quite primitive types of movement. The first one, called rotation, allows a node to rotate 90° around one of its neighbors either clockwise or counterclockwise (see, e.g., Figure 7 in Section 3) and the second one, called sliding, allows a node to slide by one position “over” two neighboring nodes (see, e.g., Figure 2 in Section 5). Both movements succeed only if the whole direction of movement is free of obstacles (i.e., other nodes blocking the way). More formal definitions are provided in Section 2. One part of the paper focuses on the case in which only rotation is available to the nodes and the other part studies the case in which both rotation and sliding are available. The latter case has been studied to some extent in the past in the, so called, metamorphic systems [16, 17, 15], which makes those studies the closest to our approach. For rotation only, we introduce the notion of color-consistence and prove that if two shapes are not color-consistent then they cannot be transformed to each other. On the other hand color-consistence does not guarantee transformability as there is an infinite set of pairs (A, B) such that A and B are color consistent but still they cannot be transformed to each other. At this point, observe that if A can be transformed to B then the inverse is also true, as all movements considered in this paper are reversible. We distinguish two main types of transformations: those that are allowed to break the connectivity of the shape during the transformation and those that are not and call the corresponding problems Rot-Transformability and RotC-Transformability. We prove that RotCTransformability is a proper subset of Rot-Transformability by showing that a line-folding problem is in Rot-Transformability\RotC-Transformability. Our main result regarding Rot-Transformability is that Rot-Transformability ∈ P. To prove polynomial-time decidability, we prove that two shapes A and B are transformable to each other iff both A and B have at least one movement available (without any movement available, a shape is blocked and can only trivially transform to itself). Therefore, transformability reduces to checking the availability of a movement in the initial and target shapes. The idea is that if a movement is available in a shape A, then there is always a way to extract from A a 2-line (i.e., two neighboring nodes). Such a 2-line can move freely in any direction and can also extract further nodes to form a 4-line. A 4-line in turn can also move freely to any direction and is also capable of extracting nodes from the shape and transferring them, 3 4 On the Transformation Capability of Feasible Mechanisms for Programmable Matter one at a time, to any desired target position. In this manner, the 4-line can transform A to a line with leaves around it that is color-consistent to A (based on a proposition that we prove, stating that any shape has a corresponding color-consistent line-with-leaves). Similarly, B, given that it is color-consistent with A, can be transformed by the same approach to exactly the same line-with-leaves, and then, by reversibility, it follows that A and B can be transformed to each other by using the line-with-leaves as an intermediate. This set of transformations do not guarantee the preservation of connectivity during the transformation. That is, even though the initial and target shapes considered are connected shapes, the shapes formed at intermediate steps of the transformation may very well be disconnected shapes. We next study RotC-Transformability, in which again the only available movement is rotation, but now connectivity of the material has to be preserved throughout the transformation. The property of preserving the connectivity is expected to be a crucial property for programmable matter systems, as it allows the material to maintain coherence and strength, to eliminate the need for wireless communication, and, finally, enables the development of more effective power supply schemes, in which the modules can share resources or in which the modules have no batteries but are instead constantly supplied with energy by a centralized source (or by a supernode that is part of the material itself). Such benefits can lead to simplified designs and potentially to reduced size of individual modules. We first prove that RotC-Transformability ∈ PSPACE. The rest of our results here are strongly based on the notion of a seed. This stems from the observation that a large set of infeasible transformations become feasible by introducing to the initial shape an additional, and usually quite small, seed; i.e., a small shape that is being attached to some point of the initial shape. In particular, we prove that a 3-line seed, if placed appropriately, is sufficient to achieve folding of a line (otherwise impossible). We then investigate seeds that could serve as components capable of traveling the perimeter of an arbitrary connected shape A. Such shapes are very convenient as they are capable of “simulating” the universal transformation techniques that are possible if we have both rotation and sliding movements available (discussed in the sequel). To this end, we prove that all seeds of size ≤ 4 cannot serve for this purpose, by proving that they cannot even walk the perimeter of a simple line shape. Then we focus on a 6-seed and prove that such a seed is capable of walking the perimeter of a large family of shapes, called discrete-convex shapes. This is a first indication, that there might be a large family of shapes that can be transformed to each other with rotation only and without breaking connectivity, by extracting a 6-seed and then exploiting to transfer nodes to the desired positions. To further support this, we prove that the 6-seed is capable of performing such transfers, by detaching pairs of nodes from the shape, attaching them to itself, thus forming an 8-seed and then being still capable to walk the perimeter of the shape. Next, we consider the case in which both rotation and sliding are available and insist on connectivity preservation. We first provide a proof that this combination of simple movements is universal w.r.t. transformations, as any pair of connected shapes A and B of the same order, can be transformed to each other without ever breaking the connectivity throughout the transformation (a first proof of this fact had already appeared in [15]). This generic transformation requires Θ(n2 ) sequential movements in the worst case. By a potential-function argument we show that no transformation can improve on this worstcase complexity for some specific pairs of shapes and this lower bound is independent of connectivity preservation; it only depends on the inherent transformation-distance between the shapes. To improve on this, either some sort of parallelism must be employed or more O. Michail, G. Skretas, and P. G. Spirakis powerful movement mechanisms, e.g., movements of whole sub-shapes in one step. We investigate the former approach, and prove that there is a pipelining general transformation strategy that improves the time to O(n) (parallel time). We also give a matching Ω(n) lower bound. On the way, we also show that this parallel complexity is feasible even if the nodes are labeled, meaning that individual nodes must end up in specific positions of the target-shape. Afterwards, we propose a distributed algorithm that transforms any compact shapes into a line using the rotation-sliding movement without breaking the connectivity of the shape. We note that a unique leader is required, each node has 4 ports and we aim to minimise the memory as much as possible. The communication is synchronous with each node broadcasting messages to its neighbours each turn. Following this, we propose an algorithm that transforms any shape into a line. We have the same requirements and communication and our goal again to minimize the amount of memory required in the system. In Section 1.2 we discuss further related literature. Section 2 brings together all definitions and basic facts that are used throughout the paper. In Section 3, we study programmable matter systems equipped only with rotation movement. In Section 4, we insist on rotation only, but additionally require from the material to maintain connectivity throughout the transformation. In Section 5, we investigate the combined effect of rotation and sliding movements. Connectivity can always be preserved in this case. Section 6 focuses on distributed transformations having access to both rotation and sliding. Finally, in Section 7 we conclude and give further research directions that are opened by our work. 1.2 Further Related Work Mobile and Reconfigurable Robotics. There is a very rich literature on mobile and reconfigurable robotics. In mobile (swarm) robotics systems and models, as are, for example, the models for robot gathering [6, 25] and deployment [33] (cf., also [19]), geometric pattern formation [34, 8], and connectivity preservation [7], the modules are usually robots equipped with some mobility mechanism making them free to move in any direction of the plane (and in some cases even continuously). In contrast, we only allow discrete movements relative to neighboring nodes. Modular self-reconfigurable robotic systems form an area on their own, focusing on aspects like the design, motion planning, and control of autonomous robotic modules [4, 38, 1, 37]. The model considered in this paper bears similarities to some of the models that have appeared in this area. The main difference is that we follow a more computation-theoretic approach, while the studies in this area usually follow a more applied perspective. Puzzles. Puzzles are combinatorial one-player games, usually played on some sort of board. Typical questions of interest are whether a given puzzle is solvable and finding the solution with the fewest number of moves. Answers to these questions range from being in P up to PSPACE-hard or even undecidable when some puzzles are generalized to the entire plane with unboundedly many pieces [9, 22]. Famous examples of puzzles are the Fifteen Puzzle, Sliding Blocks, Rush Hour, Pushing Blocks, and Solitaire. Even though none of these is equivalent to the model considered here, the techniques that have been developed for solving and characterizing puzzles may turn very useful in the context of programmable matter systems. Actually, in some cases, such puzzles show up as special cases of the transformation problems considered here (e.g., the Fifteen Puzzle may be obtained if we restrict a transformation of node-labelled shapes to take place in a 4x4 square region). Passive Systems. Most of the models discussed so far including the model under considera- 5 6 On the Transformation Capability of Feasible Mechanisms for Programmable Matter tion in this paper, are active models, meaning that the movements are in the complete control of the algorithm. In contrast, in passive models the underlying algorithm cannot control the movements but in most cases it can decide in some way which movements to accept and which not. The typical assumption is that the movements are controlled by a scheduler (possibly adversarial), which represents some dynamicity of the system or the environment. Population Protocols [2, 3] and variants are a typical such example. For example, in Network Constructors [28] nodes move around randomly due to the dynamicity of the environment and when two of them interact the protocol can decide whether to establish a connection between them; that is, the protocol has some implicit control of the system’s dynamics. Another passive model, inspired from biological multicellular processes, was recently proposed by Emek and Uitto [18]. Most models from the theory of algorithmic self-assembly, like the Abstract Tile Assembly Model (aTAM) [35, 29], fall also in this category. In this paper, we are only concerned with active systems. Hybrid models combining active capabilities and passive dynamics, remain an interesting open research direction. 2 Preliminaries The programmable matter systems considered in this paper operate on a 2-dimensional square grid. As usual, each position (or cell) of the grid is uniquely referred to by its x and y coordinates, where x ≥ 0 corresponds to the row and y ≥ 0 to the column. Such a system consists of a set V of n modules, called nodes throughout. Each node may be viewed as a spherical module fitting inside a cell of the grid. At any given time, each node u ∈ V occupies a cell o(u) = (ox (u), oy (u)) = (i, j) (omitting the time index for simplicity here and also whenever clear from context) and no two nodes may occupy the same cell. In some cases, when a cell is occupied by a node we may refer to that cell by a color, e.g., black, and when a cell is not occupied (i.e., it is empty) we usually refer to it as white. At any given time t, the positioning of nodes on the grid defines an undirected neighboring relation E(t) ⊂ V × V , where {u, v} ∈ E iff ox (u) = ox (v) and |oy (u) − oy (v)| = 1 or oy (u) = oy (v) and |ox (u) − ox (v)| = 1, that is, if u and v are either horizontal or vertical neighbors on the grid, respectively. It is immediate to observe that every node can have at most 4 neighbors at any given time. A more informative way to define the system at a given time t, and thus often more convenient, is as a mapping Pt : N≥0 × N≥0 → {0, 1} where Pt (i, j) = 1 iff cell (i, j) is occupied by a node. At any given time t, Pt−1 (1) defines a shape. Such a shape is called connected if E(t) defines a connected graph. A connected shape is called convex if for any two occupied cells, the line that connects their centers does not pass through an empty cell. We call a shape discrete-convex if for any two occupied cells, belonging either to the same row or the same column, the line that connects their centers does not pass through an empty cell; i.e., in the latter we exclude diagonal lines. In general, shapes can transform to other shapes via a sequence of one or more movements of individual nodes. Time consists of discrete steps (or rounds) and in every step, zero or more movements may occur, possibly following a computation sub-step either centralized or distributed, depending on the application. In the sequential case, at most one movement may occur per step, and in the parallel case any number of “valid” movements may occur in parallel. 1 We consider two types of movements: (i) rotation and (ii) sliding. In both 1 By “valid”, we mean here subject to the constraint that their whole movement paths correspond to pairwise disjoint sub-areas of the grid. O. Michail, G. Skretas, and P. G. Spirakis 7 movements, a single node moves relative to one or more neighboring nodes as we explain now. A single rotation movement of a node u is a 90° rotation of u around one of its neighbors. Let (i, j) be the current position of u and let its neighbor be v occupying the cell (i − 1, j) (i.e., lying below u). Then u can rotate 90° clockwise (counterclockwise) around v iff the cells (i, j + 1) and (i − 1, j + 1) ((i, j − 1) and (i − 1, j − 1), respectively) are both empty. By rotating the whole system 90°, 180°, and 270°, all possible rotation movements are defined analogously. See Figure 1. i i−1 rotation is possible j after rotation j+1 Figure 1 Rotation to the right and down. A node on the black dot (in row i − 1) and empty cells at positions (i, j + 1) and (i − 1, j + 1) are required for this movement. Then an example movement is given. A single sliding movement of a node u is a one-step horizontal or vertical movement “over” a horizontal or vertical line of (neighboring) nodes of length 2. In particular, if (i, j) is the current position of u, then u can slide rightwards to position (i, j + 1) iff (i, j + 1) is not occupied and there exist nodes at positions (i − 1, j) and (i − 1, j + 1) or at positions (i + 1, j) and (i + 1, j + 1), or both. Precisely the same definition holds for up, left, and down sliding movements by rotating the whole system 90°, 180°, and 270° counterclockwise, respectively. Intuitively, a node can slide one step in one direction, if there are two consecutive nodes either immediately “below” or immediately “above” that direction that can assist the node slide (see Figure 2). 2 i+1 i i−1 sliding is possible j after sliding j+1 Figure 2 Sliding to the right. Either the two blues (dots in row i + 1) or the two blacks (dots in row i − 1) and an empty cell at position (i, j + 1) are required for this movement. Then an example movement with the two blacks is given. Let A and B be two shapes. We say that A transforms to B via a movement m (which m can be either a rotation or a sliding), denoted A → B, if there is a node u in A such that if u applies m, then the shape resulting after the movement is B (possibly after rotations and translations of the resulting shape, depending on the application). We say that A transforms m in one step to B (or that B is reachable in one step from A), denoted A → B, if A → B for some movement m. We say that A transforms to B (or that B is reachable from A) and write A B, if there is a sequence of shapes A = C0 , C1 , . . . , Ct = B, such that Ci → Ci+1 2 Observe that there are plausible variants of the present definition of sliding, such as to slide with nodes at (i − 1, j) and (i + 1, j + 1) or even with a single node at (i − 1, j) or at (i + 1, j). In this paper, though, we only focus on our original definition. 8 On the Transformation Capability of Feasible Mechanisms for Programmable Matter for all i, 0 ≤ i < t. We should mention that we do not always allow m to be any of the two possible movements. In particular, in Sections 3 and 4 we only allow m to be a rotation, as we there restrict attention to systems in which only rotation is available. We shall clearly explain what movements are permitted in each part of the paper. I Proposition 1. The relation “transforms to” (i.e., ‘ ’) is a partial equivalence relation. Proof. The relation ‘ ’ is a binary relation on shapes. To show that it is a partial equivalence relation, we have to show that it is symmetric and transitive. For symmetricity, we have to show that for all shapes A and B, if A B then B A. It suffices to show that for all A, B, if A → B then B → A, meaning that every one-step transformation (which can be either a single rotation or a single sliding) can be reversed. For the rotation case, this follows by observing that a rotation of a node u can be performed iff there are two consecutive empty positions in its trajectory. When u rotates, it leaves its previous position empty, thus, leaving in this way two consecutive positions empty for the reverse rotation to become enabled. The argument for sliding is similar. For transivity, we have to show that for all shapes A, B, and C, if A B and B C then A C. By definition, A B if there is a sequence of shapes A = C0 , C1 , . . . , Ct = B, such that Ci → Ci+1 for all i, 0 ≤ i < t and B C if there is a sequence of shapes B = Ct , Ct+1 , . . . , Ct+l = C, such that Ci → Ci+1 for all i, t ≤ i < t + l. So, for the sequence A = C0 , C1 , . . . , Ct = B, Ct+1 , . . . , Ct+l = C it holds that Ci → Ci+1 for all i, 0 ≤ i < t + l, that is, A C. J When the only available movement is rotation, there are shapes in which no rotation can be performed (we will see such examples in Section 3). If we introduce a null rotation, then every shape may transform to itself by applying the null rotation. That is, reflexivity is also satisfied, and, together with symmetricity and transivity from Proposition 1, “transforms to” (by rotations only) becomes an equivalence relation. I Definition 1. Let A be a connected shape. Color black each cell of the grid that is occupied by a node of A. A cell (i, j) is part of a hole of A if every infinite length single path starting from (i, j) (moving only horizontally and vertically) necessarily goes through a black cell. Color black also every cell that is part of a hole of A, to obtain a compact black shape A0 (i.e., one with no holes in it). Consider now polygons defined by unit-length line segments of the grid. Define the perimeter of A as the minimum-area such polygon that completely encloses A0 in its interior. The fact that the polygon must have an interior and an exterior follows directly from the Jordan curve theorem [23]. I Definition 2. Now, color red any cell of the grid that has contributed at least one of its line-segments to the perimeter and is not black (i.e., is not occupied by a node of A). Call this the cell-perimeter of shape A. See Figure 3 for an example. I Definition 3. The external surface of a connected shape A, is a shape B, not necessarily connected, consisting of all nodes u ∈ A such that u occupies a cell defining at least one of the line-segments of A’s perimeter. I Definition 4. The extended external surface of a connected shape A, is defined by adding to A’s external surface all nodes of A whose cell shares a corner with A’s perimeter (for example, the black node just below the hole, in Figure 3). I Proposition 2. The extended external surface of a connected shape A, is itself a connected shape. O. Michail, G. Skretas, and P. G. Spirakis Figure 3 The perimeter (polygon of unit-length line segments colored blue) and the cell-perimeter (cells colored red) of a shape A (white spherical nodes; their corresponding cells have been colored black). The dashed black cells correspond to a hole of A. Proof. The perimeter of A is connected, actually, it is a cycle. This connectivity is preserved by the extended external surface, as whenever the perimeter moves straight, we have two horizontally or vertically neighboring nodes on the extended external surface and whenever it makes a turn, we either stay put or preserve connectivity via an intermediate diagonal node (from those nodes used to extend the external surface). J Observe, though, that the extended external surface is not necessarily a cycle. For example, the extended external surface of a line-shape is equal to the shape itself (and, therefore, a line). 2.1 Problem Definitions We here provide formal definitions of all the transformation problems that are considered in this work. Rot-Transformability. Given an initial shape A and a target shape B (usually both connected), decide whether A can be transformed to B (usually, under translations and rotations of the shapes) by a sequence of rotation only movements. RotC-Transformability. The special case of Rot-Transformability in which A and B are connected shapes and, additionally, connectivity must be preserved throughout the transformation. RS-Transformability. Given an initial shape A and a target shape B (usually both connected), decide whether A can be transformed to B (usually, under translations and rotations of the shapes) by a sequence of rotation and sliding movements. Minimum-Seed-Determination. Given an initial shape A and a target shape B (usually only with rotation available and a proof that A and B are not transformable to each other without additional assumptions) determine a minimum-size seed and an initial positioning of that seed relative to A that makes the transformation from A to B feasible. There are several meaningful variations of this problem. For example, the seed may or may not form 9 10 On the Transformation Capability of Feasible Mechanisms for Programmable Matter part of the target shape or the seed may be used as an intermediated step to show feasibility with “external” help and then be able to show that, instead of externally providing it, it is possible to extract it from the initial shape A via a sequence of moves. We will clearly indicate which version is considered in each case. In the above problems, the goal is to show feasibility of a set of transformation instances and, if possible, to provide an algorithm that decides feasibility. 3 In the last part of the paper, we consider distributed transformation tasks. There, the nodes are distributed processes able to perform communicate-compute-move rounds and the goal is to program them so that they (algorithmically) self-transform their initial arrangement to a target arrangement. Distributed-Transformability. Given an initial shape A and a target shape B (usually by having access to both rotation and sliding), the nodes (which are now distributed processes), starting from A, must transform themselves to B by a sequence of communicationcomputation-movement rounds. In the distributed transformations, we mostly consider the case in which A can be any connected shape and B is a spanning line, i.e., a linear arrangement of all the nodes. 3 Rotation In this section, the only permitted movement is 90° rotation around a neighbor. Consider a black and red checkered coloring of the 2D grid, similar to the coloring of a chessboard. Then any shape S may be viewed as a colored shape consisting of b(S) blacks and r(S) reds. Call two shapes A and B color-consistent if b(A) = b(B) and r(A) = r(B) and call them color-inconsistent otherwise. Call a transformation from a shape A to a shape C color-preserving if A and C are color consistent. Observe now, that if A → B, then A and B are color-consistent, because a rotation can never move a node to a position of different color than its starting position. This implies that if A C, then A and C are color-consistent, because any two consecutive shapes in the sequence are color-consistent. We conclude that: I Observation 1. The rotation movement is color-preserving. Formally, A C (restricted to rotation only) implies that A and C are color-consistent. In particular, every node beginning from a black (red) position of the grid, will always be on black (red, respectively) positions throughout a transformation consisting only of rotations. Based on this property of the rotation movement, we may call each node black or red throughout a transformation, based only on its initial coloring. The above observation gives a partial way to determine that two shapes A and B cannot be transformed to each other by rotations. I Proposition 3. If two shapes A and B are color-inconsistent, then it is impossible to transform one to the other by rotations only. We now show that the inverse is not true, that is, it does not hold that any two color-consistent shapes can be transformed to each other by rotations. This is trivial for disconnected shapes, as any collection of isolated nodes cannot move at all, and either we 3 An immediate next goal is to devise an algorithm able to compute an actual transformation or even compute or approximate the optimum transformation (usually with respect to the number of moves). We leave these as interesting open problems. O. Michail, G. Skretas, and P. G. Spirakis consider only the cardinalities of the colors, in which case any two such shapes of equal cardinalities correspond to the same shape, or we also consider the precise positions of the nodes on the grid (e.g. by their relative distances), in which case no two such shapes can be transformed to each other. Thus, we show a counterexample for the case of connected shapes. We begin with a proposition relating the number of black and red nodes in a connected shape. I Proposition 4. A connected shape with k blacks has at least d(k − 1)/3e and at most 3k + 1 reds. Proof. For the upper bound, observe that a black can hold up to 4 distinct reds in its neighborhood, which implies that k blacks can hold up to 4k reds in total, even if the blacks were not required to be connected to each other. To satisfy connectivity, every black must share a red with some other black (if a black does not satisfy this, then it cannot be connected to any other black). Any such sharing reduces the number of reds by at least 1. As at least k − 1 such sharings are required for each black to participate in a sharing, it follows that we cannot avoid a reduction of at least k − 1 in the number of reds, which leaves us with at most 4k − (k − 1) = 3k + 1 reds. For the lower bound, if we invert the roles of blacks and reds, we have that l reds can hold at most 3l + 1 blacks. So, if k is the number of blacks, it holds that k ≤ 3l + 1 ⇔ l ≥ (k − 1)/3 and due to the fact that the number of reds must be an integer, we conclude that for k blacks the number of reds must be at least d(k − 1)/3e. J I Proposition 5. There is a generic connected shape, called line-with-leaves, that has a color-consistent version for any connected shape. In other words, for k blacks it covers the whole range of reds from d(k − 1)/3e to 3k + 1 reds. Proof. Consider a bi-color line starting with a black node and ending to a black node, such that all k blacks are exhausted, as shown in Figure 4. To do this, k − 1 reds are needed in order to alternate blacks and reds on the line. Next, “saturate” every black (i.e. maximize its degree) by adding as many red nodes as it can fit around it (recall that the maximum degree of every node is 4). The resulting saturated shape has k blacks and 3k + 1 reds. This shape covers the 3k + 1 upper bound on the possible number of reds. By removing red leaf-nodes (i.e., of degree 1) one after the other, we can achieve the whole range of numbers of reds, from k − 1 to 3k + 1 reds. It suffices to restrict attention to the range from k to 3k + 1 reds. Take now any connected shape A and color it in such a way that red is the majority color, that is l ≥ k, where l is the number of reds and k is the number of blacks (there is always a way to do that). From the upper bound of Proposition 4, l can be at most 3k + 1, so we have k ≤ l ≤ 3k + 1 for any connected shape A, which falls within the range that the line-with-leaves can represent. Therefore, we conclude that any connected shape A has a color-consistent shape B from the line-with-leaves family. J Figure 4 A saturated line-with-leaves shape, in which there are k = 5 blacks and 3k + 1 = 16 reds. 11 12 On the Transformation Capability of Feasible Mechanisms for Programmable Matter I Proposition 6. There is an infinite set of pairs (A, B) of connected shapes, such that A and B are color-consistent but cannot be transformed to each other by rotations only. Proof. For shape A, take a rhombus as shown in Figure 5, consisting of k 2 blacks and (k +1)2 reds, for any k ≥ 2. In this shape, every black node is “saturated”, meaning that it has 4 neighbors, all of them necessarily red. This immediately excludes the blacks from being able to move, as all their neighboring positions are occupied by reds. But the same holds for the reds, as all potential target-positions for a rotation are occupied by reds. Thus, no rotation movement can be applied to any such shape A and A can only be transformed to itself (by null rotations). By Proposition 5, any such A has a color-consistent shape B from the family of line-with-leaves shapes, such that B is not equal to A (actually in B several blacks may have degree 3 in contrast to A where all blacks have degree 4). We conclude that A and B are distinct color-consistent shapes which cannot be transformed to each other, and there is an infinite number of such pairs, as the number k 2 of black nodes of A can be made arbitrarily large. J Figure 5 A rhombus shape, consisting of k2 = 9 blacks and (k + 1)2 = 16 reds. Propositions 3 and 6 give a partial characterization of pairs of shapes that cannot be transformed to each other. Observe that the impossibilities proved so far, hold for all possible transformations based on rotation only, i.e., they do not restrict the transformation in any way as would be, for example, to not allow the transformation to break the connectivity of the shape at any time. A small shape of particular interest is a bi-color pair or 2-line. Such pairs can move easily in any direction, which makes them very useful components of transformations. One way to simplify some transformations would be to identify as many such pairs as possible in a shape and treat them in a different way than the rest of the nodes. A question in this respect is whether all the minority-color nodes of a connected shape can be completely to (distinct) nodes of the majority color. We show that this is not true. I Proposition 7. There is an infinite family of connected shapes, such that, if A is a shape in the family of size n, then any matching of A leaves at least n/8 nodes of each color unmatched. Proof. See Figure 6. J Recall that Rot-Transformability is the language of all transformation problems between connected shapes that can be solved by rotation only and RotC-Transformability is its subset obtained by the restriction that the transformation should not break the connectivity of the shape at any point during the transformation. We begin by showing that the inclusion between the two languages is strict, that is, there are strictly more feasible O. Michail, G. Skretas, and P. G. Spirakis Figure 6 The counterexample. transformations if we allow connectivity to break. We prove that by showing that there is a feasible transformation in Rot-Transformability\RotC-Transformability. I Theorem 5. RotC-Transformability ⊂ Rot-Transformability. Proof. RotC-Transformability ⊆ Rot-Transformability is immediate, as any transformation that does not break the shape’s connectivity is also a valid transformation for Rot-Transformability. So, it suffices to prove that there is a transformation problem in Rot-Transformability\RotC-Transformability. Consider a (connected) horizontal line of any even length n, and let u1 , u2 , . . . , un be its nodes. The transformation asks to fold the line onto itself, forming a double-line of length n/2 and width 2, i.e., a n/2 × 2 rectangle. It is easy to observe that this problem is not in RotC-Transformability for any n > 4: the only nodes that can rotate without breaking connectivity are u1 and un , but any of their two possible rotations only enables a rotation that will bring the nodes back to their original positions. This means that, if the transformation is not allowed to break connectivity, then such a shape is trapped in a loop in which only the endpoints can rotate between three possible positions, therefore it is impossible to fold a line of length greater than 4. On the other hand, if connectivity can be broken, we can perform the transformation by the following simple procedure, consisting of n/4 phases: In the beginning of every phase i ∈ {1, 2, . . . , bn/4c}, pick the nodes u2i−1 , u2i , which shall at that point be the two leftmost nodes of the original line. Rotate u2i−1 once clockwise, to move above u2i , then u2i three times clockwise to move to the right of u2i−1 (the first of these three rotations breaks connectivity and the third restores it), and then rotate u2i−1 twice clockwise to move to the right of u2i , then u2i twice clockwise to move to the right of u2i−1 and repeat this alternation until the pair that moves to the right meets the previous pair, which will be when u2i−1 becomes the left neighbor of u2i−2 on the upper line of the rectangle under formation, or, in case i = 1, when u2i−1 goes above un (see Figure 7). If n/4 is not an integer, then perform a final phase, in which the leftmost node of the original line is rotated once clockwise to move above its right neighbor, and this completes folding. J This means that allowing the connectivity to break enables more transformations, and this motivates us to start from this simpler case. But we already know from Proposition 6, that even in this case an infinite number of pairs of shapes cannot be transformed to each other. Aiming at a general transformation, we ask whether there is some minimal addition 13 14 On the Transformation Capability of Feasible Mechanisms for Programmable Matter 1 3 4 2 5 6 7 8 9 10 Figure 7 Line folding. to a shape that would allow it to transform. The solution turns out to be as small as a 2-line seed lying initially somewhere “outside” the boundaries of the shape (e.g., just below the lowest row occupied by the shape). Based on the above assumptions, we shall now prove that any pair of color-consistent connected shapes A and B can be transformed to each other. Recall from the discussion before Proposition 7, that 2-line shapes can move freely in any direction. The idea is to use this 2-line in order to extract from the shape another 2-line, and use the two 2-lines together as a 4-line seed. The 4-line can also move freely in all directions. Then we shall use the 4-line as a transportation medium for those nodes that cannot move alone. In particular, we partition the nodes of the shape into those that can leave the shape as part of a 2-line and those that cannot. The latter nodes require the help of the 4-line to move them by carrying them, one at a time, in the form of a shape of order 5, which can only move diagonally (due to color-preservation of Proposition 3). We exploit these mobility mechanisms to transform A into a uniquely defined shape from the line-with-leaves family of Proposition 5 (meaning that any two color-consistent shapes are matched to the same shape from the family). But if any connected shape A with an extra 2-line can be transformed to its color-consistent line-with-leaves version with an extra 2-line, then this also holds inversely due to reversibility of rotations (discussed in the proof of Proposition 1), and it follows that any A can be transformed to any B by transforming A to its line-with-leaves version LA and then inverting the transformation from B to LB = LA . I Theorem 6. If connectivity can break and there is a 2-line seed provided “outside” the initial shape, then any pair of color-consistent connected shapes A and B can be transformed to each other by rotations only. Proof. Without loss of generality (due to symmetry and the 2-line’s unrestricted mobility), it suffices to assume that the seed is provided somewhere below the lowest row l occupied by the shape A. We show how A can be transformed to LA with the help of the seed. We O. Michail, G. Skretas, and P. G. Spirakis 15 define LA as follows: Let k be the cardinality of the minority color, let it be the black color. As there are at least k reds, we can create a horizontal line of length 2k, i.e., u1 , u2 , . . . , u2k , starting with a black, i.e., u1 is black, and alternating blacks and reds. In this way, the blacks are exhausted. The remaining ≤ (3k + 1) − k = 2k + 1 reds are then added as leaves of the black nodes, starting from the position to the left of u1 and continuing counterclockwise, i.e., below u1 , below u3 , ..., below u2k−1 , above u2k−1 , above u2k−3 , and so on. This gives the same shape from the line-with-leaves family, for all color-consistent shapes (observe that the leaf to the right of the line is always placed). LA shall be constructed on rows l − 5 to l − 3 (not necessarily inclusive), with u1 on row l − 4 and a column j preferably between those that contain A. First, extract a 2-line from A, from row l, so that the 2-line seed becomes a 4-line seed. To see that this is possible for every shape A of order at least 2, distinguish the following two cases: (i) If the lowest row has a horizontal 2-line, then the 2-line can leave the shape without any help and approach the 2-seed. (ii) If not, then take any node u of row l. As A is connected and has at least two nodes, u must have a neighbor v above it. The only possibility that the 2-line u,v is not free to leave A is when v has both a left and a right neighbor. Figure 8 shows how this can be resolved with the help of the 2-line seed (now the 2-line seed approaches and extracts the 2-line). 2-line unable to leave 2-line seed binds (a) (b) (c) 4-line constructed (d) (e) Figure 8 Extracting a 2-line with the help of the 2-line seed. To transform A to LA , given the 4-line seed, do the following: While the minority color (color chosen for u1 ) is still present in A: If on the current lowest row occupied by A, there is a 2-line that can be extracted alone and move towards LA , then perform the shortest such movement that attaches the 2-line to the right endpoint of LA ’s line u1 , u2 , . . .. If not, then use the 4-line to extract a single node from the lowest row of A. If that node fits to the right endpoint of LA ’s line, place it there, otherwise, transfer it to an unoccupied position below row l − 7 to be used later. Once the minority color has been exhausted from A, alternate the two colors until u2k−3 has been placed (u2k−1 and u2k will only be placed in the end as they are part of the 4-line). To do this, use the 4-line to transfer nodes from A and from the “repository” maintained below LA . When this occurs, if there are no more nodes left, run the termination phase, otherwise transfer the remaining nodes with the 4-line, one after the other, and attach them around the line of LA , beginning from the position to the left of u1 counterclockwise, as decribed above (skipping the position u2k ). 16 On the Transformation Capability of Feasible Mechanisms for Programmable Matter Termination phase: the line-with-leaves is ready, apart from positions u2k−1 , u2k which require a 2-line from the 4-line. If the position above u2k−1 is empty, then extract a 2-line from the 4-line and transfer it to the positions u2k−1 , u2k . This completes the transformation. If the position above u2k−1 is occupied by a node u2k+1 , then place the whole 4-line vertically with its lowest endpoint on u2k (as in Figure 9). Then rotate the top endpoint counterclockwise, to move above u2k+1 , then rotate u2k+1 clockwise around it to move to its left, then rotate the node above u2k counterclockwise to move to u2k−1 , and finally restore u2k+1 to its original position. This completes the construction (the 2-line that always remains can be transferred in the end to a predefined position). J (a) (b) (c) (d) (e) (f) Figure 9 The termination phase of the transformation. The natural next question is to what extent can the 2-line seed assumption be dropped. Clearly, by Proposition 6, this cannot be always possible. The following corollary gives a sufficient condition to drop the 2-line seed assumption, without looking deep into the structure of the shapes that satisfy it. I Corollary 7. Assume rotations only and that connectivity can break. Let A and B be two color-consistent connected shapes such that each one of them can self-extract a 2-line. Then A and B can be transformed to each other. We remind that a rotation move in a grid can occur towards 4 directions: N orthEast(1), SouthEast(2), SouthW est(3), N orthW est(4). In order for the first move to occur a node has to be present North OR East but not both. The same requirements apply for moves 2, 3 and 4 respectively. If the connectivity of the shape can be broken and two nodes, A and B, are next to each other and A can perform a rotation using B, then B can perform a rotation using A if the connectivity of the shape can be broken. I Lemma 8. A 2-seed can be extracted from a shape iff a single rotation move is available on the shape. Proof. If a move is available on a shape but not on the perimeter, that move can be transferred to the perimeter through transformations. Let us consider a shape that has only two holes which are next to each other. We will call them cell A and E respectively. Without loss of generality let us consider that cell O. Michail, G. Skretas, and P. G. Spirakis A is west of E. We name S the cell south of A and SE the cell south of E. Now we propose the following method. The node residing in cell S rotates to the cell E and then the node in cell SE rotates to the cell A. After these two moves, cell S is renamed to A and SE is renamed to E. The cell south of the new A and E are named S and SE respectively. This method can be repeated indefinitely until the two white cells reach the end of the grid. We have shown how two white cells can“travel” south. By reversing the method the two white cells can travel north. The two white cells can travel east and west with a simple transformation before the method. After naming the four cells above, the node in cell S rotates to E. After this step we have two white cells, A and S. Now rename S into E and E into S. Now repeat the method and the two white positions will start travelling east. For the opposite direction, rotate the node in position SE to cell A, and rename A into SE and SE into A. Now repeat the method and the white cells can travel east. By using a combination of the above steps, the two white cells can move freely through the grid and reach any place. Now consider a shape where there are more than 2 holes but at least two are next to each other. We will show that the two white cells that are side by side can travel to the perimeter of the shape using the above method even if they reach other white cells. Without loss of generality suppose that the two white cells are the southernmost pair travelling south. If the travelling nodes ever meet a white cell south of them, we just need to show that we can turn this cell (S) from a white one to a black one. Thus we perform the following act: Check if there is a node west of A. If there is, move him south of A. Note that the cell west of S is always a black node because we cannot have two white cells next to each other south of A. If not check if there is a node north of A. Note that there is always a node north of A, else a move would never be available which is prohibited. Now move the node north of A to the west of A then south of A. This move is available if there is a node northwest of A. If there is not, move the node north of A, east of A then south of A. This move is available only if there is a node northeast of A. If there is not, move the node east to the northeastern cell of A then east of A then south of A. If there is not one, we reach the following shape.See figure 10 The first node available northeast of A or northwest of A can be moved with rotations to the cell south of A. If a node is not available on either of those lines then either the connectivity of the shape is breached because we know that there are nodes north of A which have to be connected with the rest of the shape, or the A cell is not part of the shape. Both of those are not allowed so there is always a node northeast or northwest of A. Thus there is always a way to fill the cell south of A. In a similar fashion if the cell south of E was white, we could always fill it. A 2-seed can be extracted from a shape if a single rotation move is available on the perimeter of the shape. Without the loss of generality suppose that nodes A and B are east-west to each other respectively, they are the southernmost nodes with a move available and none of them have any nodes in the two cells directly south of them. This means that the other can move as well. If node A can perform a rotation to move south of B then afterwards B can perform a rotation to move west of A. Then B can rotate south of A and A west of B. This four step method can be repeated forever until either one of them finds a node south. If one of them finds a node south, called C i.e. A find a node south of him then B moves north of A and A moves east of C. Then A and C perform the four step method. If the two nodes keep repeating this eventually they will disconnect from the shape as a 2-seed. 17 18 On the Transformation Capability of Feasible Mechanisms for Programmable Matter If a move is not available a 2-seed cannot be extracted. If a move is not available then no node can perform a rotation move. that no node can begin the process to extract himself as part of a 2-seed. This means J A Figure 10 Proof for 2seed extraction to the perimeter. I Theorem 9. Rotation-Transformability belongs to P. Proof. In Lemma 8, we proved that we can extract a 2-seed from a shape iff a move is initially available. By Theorem 6, if both shapes A and B have a 2-seed available then they can be transformed to each other. It follows that two shapes A and B can be transformed to each other iff both have a move available. Now we define a n × n grid where any shape with n nodes can fit in. The time it takes for an algorithm to check if one of the shapes has a move available is O(n). If for example the algorithm checks each individual node, that takes O(1) time and, therefore, O(n) time for n nodes. So for two shapes it takes O(n) time to check if a move is available in each of the shapes. Thus, the problem belongs to P. If the two shapes, A and B, are the same, then they can trivially transform to each other without any moves. An algorithm can check this by simply mapping the grid of the first shape, which takes O(n) time, and then check the second shape to see if the black cells match. If it ever finds a black cell that does not exist on the first shape, or it finds a white cell when it expected a black cell, then it decides that the two shapes are not the same. This process takes O(n) time because it is equals to the time it takes to visit every n node. Thus, it takes O(n) time to check if A = B. J 4 Rotation and Connectivity Preservation In this section, we restrict our attention to transformations that transform a connected shape A to one of its color-consistent shapes B, without ever breaking the connectivity of the shape on the way. As already mentioned in the introduction, connectivity preservation is a very desirable property for programmable matter, as, among other positive implications, it guarantees that communication between all nodes is maintained, it minimizes transformation failures, requires less sophisticated actuation mechanisms, and increases the external forces required to break the system apart. We begin by proving that RotC-Transformability can be decided in deterministic polynomial space. O. Michail, G. Skretas, and P. G. Spirakis I Theorem 10. RotC-Transformability is in PSPACE. Proof. We first present a nondeterministic Turing machine (NTM) N that decides Transformability in polynomial space. N takes as input two shapes A and B, both consisting of n nodes and at most 4n edges. A reasonable representation is in the form of a binary n × n matrix (representing a large enough sub-area of the grid) where an entry is 1 iff the corresponding position is occupied by a node. Given the present configuration C, where C = A initially, N nondeterministically picks a valid rotation movement of a single node. This gives a new configuration C 0 . Then N replaces the previous configuration with C 0 in its memory, by setting C ← C 0 . Moreover, N maintains a counter moves (counting the number of moves performed so far), with maximum value equal to the total number of possible shape 2 configurations, which is at most 2n in the binary matrix encoding of configurations. To set up such a counter, N just have to reserve for it n2 (binary) tape-cells, all initialized to 0. Every time N makes a move, as above, after setting a value to C 0 it also increases moves by 1, i.e., sets moves ← moves + 1. Then N takes another move and repeats. If it ever holds that C 0 = B (may require N to perform a polynomial-space pattern matching on the n × n matrix to find out), then N accepts. If it ever holds that the counter is exhausted, that is, all its bits are set to 1, N rejects. If A can be transformed to B, then there must be a transformation beginning from A and producing B, by a sequence of valid rotations, without ever repeating a shape. Thus, some branch of N ’s computation will follow such a sequence 2 and accept, while all non-accepting branches will reject after at most 2n moves (when moves reaches its maximum value). If A cannot be transformed to B, then all branches will reject 2 after at most 2n moves. Thus, N correctly decides Transformability. Every branch of N , at any time, stores at most to shapes (the previous and the current), which requires O(n2 ) 2 space in the matrix representation, and a 2n -counter which requires O(n2 ) bits. It follows that every branch uses space polynomial in the size of the input. So, far we have proved that Transformability is decidable in nondeterministic polynomial (actually, linear) space. By applying Savitch’s theorem [32] 4 , we conclude that Transformability is also decidable in deterministic polynomial space (actually, quadratic), i.e., it is in PSPACE. J Recall that in the line folding problem, the initial shape is a (connected) horizontal line of any even length n, with nodes u1 , u2 , . . . , un , and the transformation asks to fold the line onto itself, forming a double-line of length n/2 and width 2. As part of the proof of Theorem 5, it was shown that if n > 4, then it is impossible to solve the problem by rotation only (if n = 4, it is trivially solved, just by rotating each endpoint above its unique neighbor). In the next proposition, we employ again the idea of a seed to show that with a little external help the transformation becomes feasible. I Proposition 8. If there is a 3-line seed v1 , v2 , v3 , horizontally aligned over nodes u3 , u4 , u5 of the line, then the line can be folded. Proof. We distinguish two cases, depending on whether we want the seed to be part of the final folded line or not. If yes, then we can either use a 4-line seed directly, over nodes u3 , u4 , u5 , u6 , or a 3-line seed but require n to be odd (so that n + 3 is even). If not, then n must be even. We show the transformation for the first case, with n odd and a 3-line seed (the other cases can be then treated with minor modifications). 4 Informally, Savitch’s theorem establishes that any NTM that uses f (n) space can be converted to a deterministic TM that uses only f 2 (n) space. Formally, it establishes that for any function f : N → N, where f (n) ≥ log n, NSPACE(f (n)) ⊆ SPACE(f 2 (n)). 19 20 On the Transformation Capability of Feasible Mechanisms for Programmable Matter We first show a simple reduction from an odd line with a 3-line seed starting over its third node to an even line with a 4-line seed starting over its third node. By rotating u1 clockwise over u2 , we obtain the 4-line seed u1 , v1 , v2 , v3 . It only remains to move the whole seed two positions to the right (by rotating each of its 2-lines clockwise around themselves). In this manner, we obtain an even-length line u2 , . . . , un and a 4-line seed starting over its third node, without breaking connectivity. Therefore, in what follows we may assume that the initial shape is an even-length line u1 , u2 , . . . , un with a 4-line seed v1 , v2 , v3 , v4 horizontally aligned over nodes u3 , u4 , u5 , u6 . See Figure 11. J 1 1 u1 v1 v2 v3 3 v2 v4 u2 2 u1 v1 v3 v4 u2 5 4 4 v2 v1 v1 v1 v1 6 v2 u1 v1 v3 7 v4 u2 9 v2 v 8 2 v4 u2 10 u1 u2 u1 12 u1 u1 9 11 10 v3 v3 v4 v1 v4 11 10 u2 u2 v3 v3 v1 v1 v2 13 12 u1 v1 11 v4 13 12 v2 v2 v4 v4 u2 v2 v3 v4 v1 u2 u1 v3 v2 v1 u2 u1 v3 v2 13 v4 2-line leaves right v1 v4 Figure 11 The main subroutine of line folding with connectivity preservation. We believe that in order to transform one shape to another we first need to find a seed that can both move on the perimeter of a shape and being able to reach every possible cell of the perimeter. We call this for simplicity: traverse the perimeter. After this is guaranteed we want the seed to be able to extract nodes and move them gradually to specific cells of the O. Michail, G. Skretas, and P. G. Spirakis 21 perimeter in order to create the desired shape. Thus the seed could actually simulate the rotation-sliding movement. We begin with the smallest seed possible and try to tackle the problem of moving on the perimeter of a line. Note that we do not allow the nodes of the shape to move in order to simplify and strengthen the model. I Proposition 9. A 2-seed cannot traverse the perimeter a line without breaking the connectivity. Proof. Observe figure 12, shape number 1. The 2-seed has reached the end of the line and now it tries to move east of the line and then south of it. Note that the black node has 2 possible moves. It can either perform a single move and stop above the red node, or perform two subsequent moves and stop east of the red node. No matter the choice, the red node then is not able to move because any possible move would break the connectivity of the shape. See figure 12, shapes number 2, 3. Thus the black node has to stay in place and only the red node can move now. Observe now that the red node is trapped in a loop of 2 possible moves (excluding the act of moving above the black node which would not allow us to try and move under the line): become the new endpoint or move under the end of the line. The first case leads necessarily to the second case because it is the only legal move available (excluding the move of looping back). See figure 13. But when we each the second case, once more we are limited into looping back to the initial positions. Thus a 2-seed cannot traverse the perimeter of a line without breaking the connectivity J 1 2 3 Figure 12 Black node movement for 2-seed. 1 2 3 Figure 13 Red node movement for 2-seed). I Proposition 10. A 4-seed cannot traverse the perimeter of a line without breaking the connectivity. Proof. Consider the last time, tlast, that the black and red nodes in rows≤i increases from 2 to 3. This means that either a black or a red moved at tlast from i + 1 to i. From now on, none of those 3 nodes can go back to rows>i and there is one node remaining in rows>i. 22 On the Transformation Capability of Feasible Mechanisms for Programmable Matter Actually that node u must necessarily be in row i + 1, otherwise the connectivity would have broken. So no node from rows≤i can return to rows>i anymore and there is a single node u remaining in row r + 1. We begin by finding the possible shapes that meet the above requirements. The rotation of the node at tlast was necessarily clockwise, as the closest counterclockwise move to the line is from (i + 1, j + 1) to (i, j), but it requires (i, j) to be empty before rotating, but then 2 nodes in rows≤i and only one additional (u) in rows>i cannot support connectivity. We will now distinguish the tlast into cases. If u is a black node: If u is at position (i + 1, j − 2) then it is stuck forever (blue node cannot move and the other black and red cannot go up any more to carry u. It also cannot be at (i + 1, j + 2) as this does not permit a clockwise move of a red from (i + 1, j + 1), so it has to be at (i + 1, j). See figure 15 shape number 2. Node u is connected to A only via the red below it, which therefore cannot move unless u moves first(because no node can return to row>i any more to support u via another path. But the only way for u to move is for the black southeast node to move first, which in turn cannot move unless the rightmost red moves up which is impossible as no node may return to row>i (that red node can move down but then the only available movement is to return to its previous position. If u is a red node: It cannot be at (i + 1, j − 3) as before and it cannot be at (i + 1, j + 1) as the rotation at tlast was then necessarily from (i + 1, j) which is blocked by u. Observe that the clockwise rotation could not have been from (i + 1, j + 2). The only way to support connectivity in this case with 2 nodes in rows≤i and 2 in rows>i, is by having the following shape but then a clockwise rotation of the upper black is impossible. Therefore, if u is a red node it has to be at (i + 1, j − 1). See figure 15 shape number 1. Either nodes in rows≤i cannot move at all, or if the bottom black is far away, the rightmost black is trapped in a loop going down and then up to its original positions as before. Therefore the 4-seed cannot traverse the perimeter of a line without breaking the connectivity. J J j-3 j-2 j-1 j j+1 j+2 i+2 i+1 i i-1 i-2 Figure 14 Starting position of 4-seed on a line. I Proposition 11. A 6-seed can traverse the perimeter of a discrete-convex shape without breaking the connectivity. Proof. Consider a folded 6-seed occupying cells (i, j), (i, j + 1), (i, j + 2) and (i − 1, j), (i − 1, j + 1), (i − 1, j + 2). Since the shape is discrete-convex, iff there is any node present in cells (i, j + 3) or (i − 1, j + 3), there can be no node present in cells (i, j − 1) or (i − 1, j − 1). For the same reason if there is a node in cells (i − 2, j) or (i − 2, j + 1) or (i − 2, j + 2) O. Michail, G. Skretas, and P. G. Spirakis 23 1 2 Figure 15 Black and red cases for node u). there can be no node in cells (i + 1, j) or (i + 1, j + 1) or (i + 1, j + 2).In order to place this seed at those cells, one of the neighbouring cells has to be occupied by a node. Without loss of generality, suppose that the 6-seed tries to move east. There are 4 distinct cases for this move. Note that in the following 4 cases we assume the absolute minimum amount of neighbouring nodes. If at any case there were more present at the shape, the rotations would be the exact same without any modification or problem. A node occupies cell (i, j + 3). In order for this shape to be discrete-convex, a node has to be present in cell (i − 1, j + 3). In this case the 6-seed has to move north and performs the rotations described in figure 16, 17, 18, 19 if a node is present in cell (i + 1, j + 3) and the orientations described in figures 20 if a node is not present in cell (i + 1, j + 3) in order to keep moving. A node occupies cell (i − 1, j + 3) and no node occupies cell (i, j + 3). In this case the 6-seed performs the rotations described in figure 21 in order to climb the step. Note that since the 6-seed begins and ends the move while preserving its shape, it is guaranteed that any number of steps can be climbed this way. A node occupies cell (i − 2, j + 3) and no node occupies cell (i − 1, j + 3) or (i, j + 3). In this case the 6-seed performs the rotations described in figure 22 rotations in order to slide east. No nodes occupy cells (i − 2, j + 3), (i − 1, j + 3) and (i, j + 3). In this case the 6-seed performs the rotations described in figure 23 in order to reach a shape that matches the conditions of the step case. Therefore the 6-seed can now perform a climb move in order to continue. We can replicate the results for south, west, north directions by simply rotating the whole shape by 90, 180, 270 degrees respectively. J 1 2 2 i,j 2 1 1 2 1 2 3 1 2 1 4 5 6 7 Figure 16 Method for going north. I Proposition 12. An 8-seed can traverse the perimeter of a discrete-convex shape without breaking the connectivity. Proof. Consider a folded 8-seed occupying cells (i, j) (i, j + 1) (i, j + 2) (i, j + 3) and (i − 1, j), (i − 1, j + 1), (i − 1, j + 2), (i − 1, j + 3). Since the shape is discrete-convex, iff there is any 24 On the Transformation Capability of Feasible Mechanisms for Programmable Matter 1 2 3 i,j 2 4 2 1 2 1 1 1 2 3 4 5 6 Figure 17 Method for going north with node at (i-2,j+2). 1 2 2 1 i,j 3 1 2 1 2 3 4 Figure 18 Method for going north with node at (i-2,j+1). 1 1 2 1 3 2 i,j 1 2 2 1 2 3 4 5 6 Figure 19 Method for going north with node at (i-2,j). 2 2 3 1 1 1 1 2 4 3 2 3 4 3 1 2 3 4 5 Figure 20 Method for climbing north. 1 2 1 i,j 2 2 1 2 4 2 3 3 1 1 1 2 3 Figure 21 Method for climbing a step. 4 5 6 7 O. Michail, G. Skretas, and P. G. Spirakis 6 4 5 3 25 2 1 1 2 Figure 22 Method for sliding. 2 1 4 2 1 1 2 3 4 5 3 4 1 2 3 6 2 1 7 Figure 23 Method for transforming. node present in cells (i, j + 4) or (i − 1, j + 4), there can be no node present in cells (i, j − 1) or (i − 1, j − 1). For the same reason if there is a node in cells (i − 2, j) or (i − 2, j + 1) or (i − 2, j + 2) or (i − 2, j + 3) there can be no node in cells (i + 1, j) or (i + 1, j + 1) or (i + 1, j + 2) or (i + 1, j + 3).In order to place this seed at those cells, one of the neighbouring cells has to be occupied by a node. Without loss of generality, suppose that the 8seed tries to move east. There are 4 distinct cases for this move. Note that in the following 4 cases, if not mentioned, we assume the absolute minimum amount of neighbouring nodes. If at any case there were more present at the shape, the rotations would be the exact same without any modification or problem. A node occupies cell (i, j + 4). In order for this shape to be discrete-convex, a node has to present in cell (i − 1, j + 4). In this case the 8seed has to move north and performs the rotations described in figure 24 if a node is present in cell (i + 1, j + 4); and the orientations described in figure 25 if a node is not present in cell (i + 1, j + 4) in order to keep moving. A node occupies cell (i − 1, j + 4) and no node occupies cell (i, j + 4). In this case the 8-seed performs the rotations described in figure 26 and in figure 27 order to climb the step. Note that since the 8-seed begins and ends the move while preserving its shape, it is guaranteed that any number of steps can be climbed this way. A node occupies cell (i − 2, j + 4) and no node occupies cell (i − 1, j + 4) or (i, j + 4). In this case the 8-seed performs the rotations described in figure 28 rotations in order to slide east. No nodes occupy cells (i − 2, j + 4), (i − 1, j + 4) and (i, j + 4). In this case the 8-seed performs the rotations described in figure 29 in order to reach a shape that matches the conditions of the first case. Therefore the 8-seed can now perform a climb move in order to continue. We can replicate the results for south, west, north directions by simply rotating the whole shape by 90, 180, 270 degrees respectively. J Our goal here was to show that since both a 6-seed and an 8-seed can traverse the perimeter of any discrete-convex shape, then a 6-seed may be able to start extracting 2 nodes 26 On the Transformation Capability of Feasible Mechanisms for Programmable Matter 1 2 1 2 1 2 1 1 2 2 3 4 5 Figure 24 Method for going north. 3 2 1 1 3 2 2 2 4 1 3 5 4 1 6 3 4 5 2 1 3 2 1 1 2 3 4 5 6 4 1 3 2 1 7 8 9 Figure 25 Method for climbing north. 3 1 2 4 2 1 2 3 1 4 3 1 2 3 1 2 3 4 5 Figure 26 Method for climbing a step with a node at cell (i-2,j+2). 3 1 2 4 1 1 4 2 1 4 2 3 1 2 4 1 5 1 3 2 2 3 2 1 5 2 7 8 Figure 27 Method for climbing a step. 2 3 3 3 2 1 4 1 1 3 2 9 10 11 6 O. Michail, G. Skretas, and P. G. Spirakis 27 1 3 1 1 4 2 2 2 4 3 2 3 1 2 1 3 1 2 3 4 5 6 Figure 28 Method for sliding. 3 1 1 1 2 4 5 2 4 2 2 2 3 1 3 1 1 4 2 1 2 3 2 3 3 1 4 4 1 3 5 2 2 6 1 3 2 4 1 1 3 2 7 8 3 4 1 2 13 14 Figure 29 Method for transforming. 9 10 11 1 2 12 28 On the Transformation Capability of Feasible Mechanisms for Programmable Matter at a time from the shape A , move them as an 8-seed at a designated cell, leave them there, and continue this loop while creating i.e. a line with leaves. Afterwards we could perform the same method for shape B. If we succeeded in both shapes, then we could transform one to another. 5 Rotation and Sliding In this section, we study the combined effect of rotation and sliding movements. We shall prove that rotation and sliding together, are transformation-universal, meaning that they can transform any given shape to any other shape of the same size without ever breaking the connectivity during the transformation. It would be useful for the reader to recall Definitions 1, 2, 3, and 4 and Proposition 2, from Section 2, as the results that follow make extensive use of them. As the perimeter is a (connected) polygon, it can be traversed by a particle walking on its edges (the unit-length segments). We now show how to “simulate” the particle’s movement and traverse the cell-perimeter by a node, using rotation and sliding only. I Lemma 11. If we place a node u on any position of the cell-perimeter of a connected shape A, then u can walk the whole cell-perimeter and return to its original position by using only rotations and slidings. Proof. We show how to “simulate” the walk of a particle moving on the edges of the perimeter. The simulation implements the following simple rules: 1. If the current line-segment traversed by the particle concerns the same red cell as the one of the immediately previous line-segment traversed, then stay put. 2. If not: a. If the two consecutive line-segments traversed form a line-segment of length 2, then move by sliding one position in the same direction as the particle. b. If the two consecutive line-segments traversed are perpendicular to each other, then move by a single rotation in the same direction as the particle. It remains to prove that u can indeed always perform the claimed movements. (1) is trivial. For (2.a), a line-segment of length 2 on the perimeter is always defined by two consecutive blacks to the interior and two consecutive empty cells to the exterior (belonging to the cell-perimeter), therefore, u can slide on the empty cells. For (2.b), there must be a black in the internal angle defined by the line-segments and an empty cell diagonally to it, in the exterior (for an example, see the right black node on the highest row containing nodes of A, in Figure 3, Section 2). Therefore, rotation can be performed. J Next, we shall prove that u need not be an additional node, but actually a node belonging to the shape, and in particular one of those lying on the shape’s boundary. I Lemma 12. Let A be a connected shape of order at least 2. Then there is a subset R of the nodes on A’s external surface, such that |R| ≥ 2 and for all u ∈ R, if we completely remove u from A, then the resulting shape A0 = A − {u} is also connected. Proof. If the extended external surface of A contains a cycle, then such a cycle must necessarily have length at least 4 (due to geometry). In this case, any node of the intersection of the external surface (non-extended) and the cycle can be removed without breaking A’s connectivity. If the extended external surface of A does not contain a cycle, then it corresponds to a tree graph which by definition has at least 2 leaves, i.e., nodes of degree O. Michail, G. Skretas, and P. G. Spirakis exactly 1. Any such leaf can be removed without breaking A’s connectivity. In both cases, |R| ≥ 2. J I Lemma 13. Pick any u ∈ R (R defined on a connected shape A as above). Then u can walk the whole cell-perimeter of A0 = A − {u} by rotations and slidings. Proof. It suffices to observe that u already lies on the cell-perimeter of A0 . Then, by Lemma 11, it follows that such a walk is possible. J We are now ready to state and prove the universality theorem of rotations and slidings. I Theorem 14. Let A and B be any connected shapes, such that |A| = |B| = n. Then A and B can be transformed to each other by rotations and slidings, without breaking the connectivity during the transformation. Proof. It suffices to show that any connected shape A can be transformed to a spanning line L by rotations and slidings only and without breaking connectivity during the transformation. If we show this, then A can be transformed to L and B can be transformed to L (as A and B have the same order, therefore correspond to the same spanning line L), and by reversibility of these movements, A and B can be transformed to each other via L. Pick the rightmost column of the grid containing at least one node of A, and consider the lowest node of A in that column. Call that node u. Observe that all cells to the right of u are empty. Let the cell of u be (i, j). The final constructed line will start at (i, j) and end at (i, j + n − 1). The transformation is partitioned into n − 1 phases. In each phase k, we pick a node from the original shape and move it to position (i, j + k), that is, to the right of the right endpoint of the line formed so far. In phase 1, position (i, j + 1) is a cell of the cell-perimeter of A. So, even if it happens that u is a node of degree 1, by Lemma 12, there must be another such node v ∈ A that can walk the whole cell-perimeter of A0 = A − {v} (the latter, due to Lemma 13). As u 6= v, (i, j + 1) is also part of the cell-perimeter of A0 , therefore, v can move to (i, j + 1) by rotations and slidings. As A0 is connected (by Lemma 12), A0 ∪ {(i, j + 1)} is also connected and also all intermediate shapes were connected, because v moved on the cell-perimeter and, therefore, it never disconnected from the rest of the shape during its movement. In general, the transformation preserves the following invariant. At the beginning of phase k, 1 ≤ k ≤ n − 1, there is a connected shape S(k) (where S(1) = A) to the left of of column j (j inclusive) and a line of length k − 1 starting from position (i, j + 1) and growing to the right. Restricting attention to S(k), there is always a v 6= u that could move to position (i, j + 1) if it were not occupied. This implies that before the final movement that places it on (i, j + 1), v must have been in one of (i + 1, j) and (i + 1, j + 1), if we assume that v always walks in the clockwise direction. Observe now that from each of these positions v can perform zero or more right slidings above the line in order to reach the position above the right endpoint of the line. When this occurs, a final clockwise rotation makes v the new right endpoint of the line. The only exception is when v is on (i + 1, j + 1) and there is no line to the right of (i, j) (this implies the existence of a node on (i + 1, j), otherwise connectivity of S(k) would have been violated). In this case, v just performs a single downward sliding to become the right endpoint of the line. J I Theorem 15. The transformation of Theorem 14 requires Θ(n2 ) movements in the worst case. 29 30 On the Transformation Capability of Feasible Mechanisms for Programmable Matter Proof. Consider a ladder shape of order n, as depicted in Figure 30. The strategy of Theorem 14 will choose to construct the line to the right of node u. The only node that can be selected to move in each phase without breaking the shape’s connectivity is the top-left node. Initially, this is v, which must perform dn/2e movements to reach its position to the right of u. In general, the total number of movements M , performed by the transformation of Theorem 14 on the ladder, is given by M= = lnm 2 lnm 2 +2· (n−3)/2 l X i=1 nm +i 2 (n−3)/2 X (n − 2) + 2 · i i=1 = Θ(n2 ). J v 1 2 3 1 4 2 5 3 6 4 7 5 u 7 6 8 u 8 9 10 11 v Figure 30 Transforming a ladder into a spanning line. Theorem 15 shows that the above generic strategy is slow in some cases, as is the case of transforming a ladder shape into a spanning line. We shall now show that there are pairs of shapes for which any strategy and not only this particular one, may require a quadratic number of steps to transform one shape to the other. I Definition 16. Define the potential of a shape A as its minimum “distance” from the line L, where |A| = |L|. The distance is defined as follows: Consider any placement of L relative to A and any pairing of the nodes of A to the nodes of the line. Then sum up the Manhattan distances 5 between the nodes of each pair. The minimum sum between all possible relative placements and all possible pairings is the distance between A and L and also A’s potential. In case the two shapes do not have an equal number of nodes, then any matching is not perfect and the distance can be defined as infinite. Observe that the potential of the line is 0 as it can be totally aligned on itself and the sum of the distances is 0. I Lemma 17. The potential of the ladder is Θ(n2 ). Proof. We prove it for horizontal placement of the line, as the vertical case is symmetric. Any such placement leaves either above or below it at least half of the nodes of the ladder 5 The Manhattan distance between two points (i, j) and (i0 , j 0 ) is given by |i − i0 | + |j − j 0 |. O. Michail, G. Skretas, and P. G. Spirakis (maybe minus 1). W.l.o.g. let it be above it. Every two nodes, the height increases by 1, therefore there are 2 nodes at distance 1, 2 at distance 2,. . ., 2 at distance n/4. Any matching between these nodes and the nodes of the line gives for every pair a distance at least as large as the vertical distance between the ladder’s node and the line, thus, the total distance is at least 2 · 1 + 2 · 2 + ... + 2 · (n/4) = 2 · (1 + 2 + ... + n/4) = (n/4) · (n/4 + 1) = Θ(n2 ). We conclude that the potential of the initial ladder is Θ(n2 ). J I Theorem 18. Any transformation strategy based on rotations and slidings and performing a single movement per step, requires Θ(n2 ) steps to transform a ladder into a line. Proof. To show that Ω(n2 ) movements are needed to convert the ladder to a line, it suffices to observe that the difference in their potentials is that much and that one rotation or one sliding can decrease the potential by at most 1. J I Remark. The above lower bound is independent of connectivity preservation. It is just a matter of the total distance based on single distance-one movements. Finally, it is interesting to observe that such lower bounds can be computed in polynomial time, because there is a polynomial-time algorithm for computing the distance between two shapes. I Proposition 13. Let A and B be connected shapes. Then their distance d(A, B) can be computed in polynomial time. Proof. The algorithm picks a node u ∈ B, a cell c of the grid occupied by a node v ∈ A, and an orientation o ∈ north, east, south, west and draws a copy of the shape B, starting with u on c and respecting the orientation o. Then, it constructs (in its memory) a complete weighted bipartite graph (X, Y ), where X and Y are equal to the node-sets of A and B, respectively. The weight w(x, y) for x ∈ X and y ∈ Y is defined as the distance from x to y (given the drawing of shape B relative to shape A). To compute the minimum total distance pairing of the nodes of A and B for this particular placement of A and B, the algorithm computes a minimum cost perfect matching of (X, Y ), e.g., by the Kuhn-Munkres algorithm (a.k.a. the Hungarian algorithm) [26], and the sum of the weights of its edges k, and sets dist = min{d, k}. Then the algorithm repeats for the next selection of u ∈ B, cell c occupied by a node v ∈ A, and orientation o. In the end, the algorithm gives dist as output. To see that dist = d(A, B), observe that the algorithm just implements the procedure for computing the distance, of Definition 16, with the only differences being that it does not check all pairings of the nodes, instead directly computes the minimum-cost pairing, and that it does not try all relative placements of A and B but only those in which A and B share at least one cell of the grid. To see that this selection is w.l.o.g., assume that a placement of A and B in which no cell is shared achieves the minimum distance and observe that, in this case, A could be shifted one step “closer” to B, strictly decreasing their distance and, thus, contradicting the optimality of such a placement. As the relative placements of A and B are 4n2 and the Kuhn-Munkres algorithm is a polynomial-time algorithm (in the size of the bipartite graph), we conclude that the algorithm computes the distance in polynomial time. J To give a faster transformation either pipelining must be used (allowing for more than one movement in parallel) or more complex mechanisms that move sub-shapes consisting of many nodes, in a single step. 31 32 On the Transformation Capability of Feasible Mechanisms for Programmable Matter 5.1 Parallelizing the Transformations We now maintain the connectivity preservation requirement but allow an unbounded number of rotation and/or sliding movements to occur simultaneously in a single step. I Proposition 14. There is a pipelining strategy that transforms a ladder into a line in O(n) parallel time. Proof. Number the nodes of the ladder 1 through n starting from the top and following the ladder’s connectivity until the bottom-right node is reached. These gives an even-numbered upper diagonal and an odd-numbered lower diagonal. Node 1 moves as in Theorem 15. Any even node 2 ≤ w < n − 1 starts moving as long as its upper odd neighbor has reached the same level as w (e.g., node 2 first moves after node 1 has arrived to the right of node 3). Any odd node 1 < z < n starts moving as long as its even left neighbor has moved one level down (e.g., node 3 first moves after node 2 has arrived to the right of 5). After a node starts moving, it moves in every step as in Theorem 15 (but now many nodes can move in parallel, implementing a pipelining strategy). It can be immediately observed that any node i starts after at most 3 movements of node i − 1 (actually, only 2 movements for even i), so after roughly at most 3n steps, node n − 2 starts. Moreover, a node that starts, arrives at the right endpoint of the line after at most n steps, which means that after at most 4n = O(n) steps all nodes have taken their final position in the line. J Proposition 14 gives a hint that pipelining could be a general strategy to speed-up transformations. We next show how to generalize this technique to any possible pair of shapes. I Theorem 19. Let A and B be any connected shapes, such that |A| = |B| = n. Then there is a pipelining strategy that can transform A to B (and inversely) by rotations and slidings, without breaking the connectivity during the transformation, in O(n) parallel time. Proof. The transformation is a pipelined version of the sequential transformation of Theorem 14. Now, instead of picking an arbitrary next candidate node of S(k) to walk the cell-perimeter of S(k) clockwise, we always pick the rightmost clockwise node vk ∈ S(k), that is, the node that has to walk the shortest clockwise distance to arrive at the line under formation. This implies that the subsequent candidate node vk+1 to walk, is always “behind” vk in the clockwise direction and is either already free to move or is enabled after vk ’s departure. Observe that after at most 3 clockwise movements, vk cannot block any more the way of vk+1 on the (possibly updated) cell-perimeter. Moreover, the clockwise move of vk+1 , only introduces a gap in its original position, therefore it only affects the structure of the cell-perimeter “behind” it. The strategy is to start the walk of node vk+1 as soon as vk is no longer blocking its way. As in Proposition 14, once a node starts, it moves in every step, and again any node arrives after at most n movements. It follows, that if the pipelined movement of nodes cannot be blocked in any way, after 4n = O(n) steps all nodes must have arrived at their final positions. Observe now that the only case in which pipelining could be blocked is when a node is sliding through a (necessarily dead-end) “tunnel” of height 1 (such an example is the red tunnel on the third row from the bottom, in Figure 3). To avoid this, the nodes shortcut the tunnel by visiting only its first position (i, j) and then simply skipping the whole walk inside it (that walk would just return them to position (i, j) after a number of steps). J We next show that even if A and B are labeled shapes, that is, their nodes are assigned the indices 1, . . . , n (uniquely, i.e., without repetitions), we can still transform the labeled A O. Michail, G. Skretas, and P. G. Spirakis to the labeled B with only a linear increase in parallel time. We only consider transformations in which the nodes never change indices in any way (e.g., cannot transfer them, or swap them), so that each particular node of A must eventually occupy (physically) a particular position of B (the one corresponding to its index). I Corollary 20. The labeled version of the transformation of Theorem 19 can be performed in O(n) parallel time. Proof. Recall from Theorem 14 that the line were constructed to the right of some node u. That node was the lowest node in that column, therefore, there is no node below u in that column. The procedure of Theorem 19, if applied on the labeled versions of A and B will result in two (possibly differently) labeled lines, corresponding to two permutations of 1, 2, . . . , n, call them πA and πB . It suffices to show a way to transform πA to πB in linear parallel time, as then labelled A is transformed to πA , then πA to πB , and then πB to B (by reversing the transformation from B to πB ), all in linear parallel time. To do this, we actually slightly modify the procedure of Theorem 19, so that it does not construct πA in the form of a line, but in a different form that will allow us to transform it fast to πB without breaking connectivity. What we will construct is a double line, with the upper part growing to the right of node u as before and the lower part starting from the position just below u and also growing to the right. The upper line is an unordered version of the left half of πB and the lower line is an unordered version of the right half of πB . To implement the modification, when a node arrives above u, as before, if it belongs to the upper line, it goes to the right endpoint of the line as before, while if it belongs to the lower line, it continues its walk in order to teach the right endpoint of the lower line. When the transformation of labeled A to the folded line is over, the procedure has to order the nodes of the folded line and then unfold in order to produce πB . We first order the upper line in ascending order. While we do this, the lower line stays still in order to preserve the connectivity. When we are done, we order the lower line in descending order, now keeping the upper line still. Finally, we perform a parallel right sliding of the lower line (requiring linear parallel time), so that its inverse permutation ends up to the right of the upper line, thus forming π. It remains to show how the ordering of the upper line can be done in linear parallel time without breaking connectivity. To do this, we simulate a version of the odd-even sort algorithm (a.k.a. parallel bubble sort) which sorts a list of n numbers with O(n) processors in O(n) parallel time. The algorithm progresses in odd and even phases. In the odd phases, the odd positions are compared to their right neighbor and in the even phases to their left neighbor and if two neighbors are ever found not to respect the ordering a swap of their values is performed. In our simulation, we break each phase into two subphases as follows. Instead of performing all comparisons at once, as we cannot do this and preserve connectivity, in the first subphase we do every second of them and in the second subphase the rest so that between any pair of nodes being compared there are 2 nodes that are not being compared at the same time. Now if the comparison between the ith and the i + 1 node indicates a swap, then i + 1 rotates over i + 2, i slides right to occupy the previous position of i + 1, and finally i + 1 slides left over i and then rotates left around i to occupy i’s previous position. This swapping need 4 steps and does not break connectivity. The upper part has n/2 nodes, each subphase takes 4 steps to swap everyone (in parallel), each phase has 2 sub-phases, and O(n) phases are required for the ordering to complete, therefore, the total parallel time is O(n) for the upper part and similarly O(n) for the lower part. This completes the proof. J 33 34 On the Transformation Capability of Feasible Mechanisms for Programmable Matter An immediate observation is that a linear-time transformation does not seem satisfactory for all pairs of shapes. To this end, take a square S and rotate its top-left corner u, one position clockwise, to obtain an almost-square S 0 . Even though, a single counter-clockwise rotation of u suffices to transform S 0 to S, the transformation of Theorem 19 may go all the way around and first transform S 0 to a line and then transform the line to S. In this particular example, the distance between S and S 0 , according to Definition 16, is 2, while the generic transformation requires Θ(n) parallel time. So, it is plausible to ask if any transformation between two shapes A and B can be performed in time that grows as a function of their distance d(A, B). We show that this cannot always be the case, by presenting two shapes A and B with d(A, B) = 2, such that A and B require Ω(n) parallel time to be transformed to each other. I Proposition 15. There are two shapes A and B with d(A, B) = 2, such that A and B require Ω(n) parallel time to be transformed to each other. Proof. The two shapes, a black and a red one, are depicted in Figure 31. Both shapes form a square which is empty inside and also open close to the middle of its bottom side. The difference between the two shapes is the positioning of the bottom “door” of length 2. The red shape has it exactly in the middle of the side, while the black shape has it shifted one position to the left. Equivalently, the bottom side of the red shape is “balanced”, meaning that it has an equal number of nodes in each side of the vertical dashed axis that passes through the middle of the bottom, while the black shape is “unbalanced” having one more node to the right of the vertical axis than to its left. To transform the black shape into the red one, a node must necessary cross either the vertical or the horizontal axis. Because, if nothing of the two happens, then, no matter the transformation, we won’t be able to place the axes so that the running shape has two pairs of balanced quadrants, while, on the other hand, the red shape satisfies this, by pairing together the two bottom quadrants and the two upper quadrants. Clearly, no move can be performed in the upper quadrants initially, as this would break the shape’s connectivity. The only nodes that can move initially are u and w and no other node can ever move unless first approached by some other node that could already move. Observe also that u and w cannot cross the vertical boundary of their quadrants, unless with help of other nodes. But the only way for a second node to move in any of these quadrants (without breaking connectivity) is for either u or w to reach the corner of their quadrant which takes at least n/8 − 2 steps and then another n/8 steps for any (or both) of these nodes to reach the boundary, that is, at least n/4 − 2 steps, which already proves the required Ω(n) parallel-time lower bound (even a parallel algorithm has to pay the initial sequential movement of either u or w). J 6 Distributed Transformations with Rotation and Sliding LineTransformationProblem: Transform any connected shape into a shape where all nodes are either in the x axis or the y axis. I Definition 21. A node on a black cell of the grid is called a potential hole node if removing him would create a non-compact shape. I Definition 22. We define a local-info based movement (lib movement), a movement that a leader decides to perform without consulting the whole network. I Proposition 16. No algorithm based on lib movements can solve the line transformation problem without breaking the connectivity. O. Michail, G. Skretas, and P. G. Spirakis w v 35 u Figure 31 Counterexample for distance Proof. Observe the following shapes. See figure 32 If an algorithm performed a lib movement at the first shape, the same algorithm would have to perform the same movement at the second shape because it cannot distinguish the two shapes. That movement would break the connectivity on the second shape therefore no algorithm based on lib movements could solve the line formation problem on both shapes. J Figure 32 Lib movement counterexample I Definition 23. A shape is called compact when it has no holes. I Lemma 24. If all nodes in a connected shape have two or more neighbours, then there it at least one cycle present. Proof. Assume a connected graph where every node has at least 2 neighbours and there is no cycle present. Every node has 2 edges and we have n nodes. The sum of all edges is 2n/2 = n because we have double counted every edge. Now a connected graph without a cycle is called a tree and a tree has n − 1 edges. If and edge is added (n edges) the tree creates a cycle. J I Lemma 25. In compact shapes, there is always a lib movement that does not break the connectivity. Proof. Observe that there is no shape where all nodes have 3 or 4 neighbours and when a node has only one neighbour it can always perform a lib movement. Suppose that a shape exists where no lib movements are available. Each node in this shape has at least 2 neighbours so there is at least one cycle. If this cycle is compact, then a node can consult his neighbours to see if it is a bridge in order to move. (HeIt can do this by asking whether the two neighbours it has, have a different node as a common neighbor). This is a lib movement because the information required is local based. Therefore a lib movement is always available. J 36 On the Transformation Capability of Feasible Mechanisms for Programmable Matter . The difference between a move and a travel action is, that the leader moving means that it swaps his settings to a neighbouring node, while a leader traveling means that it moves his current node. label ← 0, phase ← 0, check ← 0, state ← 0, buck ← 0, num ← 0, line ← 0 while r = 1 do p0 ← north, p1 ← east, p2 ← south, p3 ← west, orien(node) = 1; Send north, east, south, west to p0 , p1 , p2 , p3 respectively; end while while r = 2, 3 AND buck = 0 do Send tick = up, right, down, lef t; Send num = 0, 1, 0, −1 to p0 , p1 , p2 , p3 respectively; Receive (tick 0 , num0 , ack 0 ); Do for every tick 0 , num0 received; if num0 > num then line = tick 0 ; num = num0 ; end if if ack 0 = null then state + +; else state − −; end if if state ≥ 2 then f ollow the path described on the line; end if if line has been reached then buck = 1; end if end while while r = 2, 3, .. AND buck = 1 do Receive(f lag 0 , qu0 ); if label = 0 then move west; f lag(node) = 1; if no node available west then label + +; end if end if if label = 1 then move east; f lag(node) = 1; if no node available east then label + +; end if end if if f lag 0 = 0 AND f lag = 0 then if label = 2 then if neighbours = 1 then label + +; else if neighbours = 2 AND neighbours not opposite then if neighbours are p0 , p1 then send qu = 1 to p0 ; label + +; end if if neighbours are p1 , p2 then send qu = 1 to p1 ; label + +; end if 1: procedure Compact-Line(Leader) 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: 27: 28: 29: 30: 31: 32: 33: 34: 35: 36: 37: 38: 39: 40: 41: 42: 43: 44: 45: 46: 47: 48: 49: 50: O. Michail, G. Skretas, and P. G. Spirakis 51: 52: 53: 54: 55: 56: 57: 58: 59: 60: 61: 62: 63: 64: 65: 66: 67: 68: 69: 70: 71: 72: 73: 74: 75: 76: if neighbours are p2 , p3 then send qu = 1 to p2 ; label + +; end if if neighbours are p3 , p0 then send qu = 1 to p3 ; label + +; end if else move; end if end if if label = 3 AND phase = 0 then if qu0 = 1 then phase = 1; else label − −; move; end if end if if label = 3 AND phase = 1 then travel; end if end if if f lag 0 = 1 OR f lag = 1 then f lag = 1; phase = 0; label = 2; move; end if end while end procedure 37 38 On the Transformation Capability of Feasible Mechanisms for Programmable Matter 1: procedure Compact-Line(Non-Leader) 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: f lag ← 0, orien ← 0, mark ← 0 while orien = 0 do Receive(north0 , south0 , east0 , west0 , tick 0 , num0 , qu0 ) if p0 receives south OR p1 receives west OR p2 receives north OR p3 receives east then orien = 1; end if if p0 receives east OR p1 receives south OR p2 receives west OR p3 receives north then pn = p(n−1)mod3 ; orien = 1; end if if p0 receives north OR p1 receives east OR p2 receives south OR p3 receives swest then pn = p(n−2)mod3 ; orien = 1; end if if p0 receives west OR p1 receives north OR p2 receives east OR p3 receives south then pn = p(n−3)mod3 ; orien = 1; end if Sendnorth, east, south, west to p0 , p1 , p2 , p3 respectively; end while while orien = 1 do if mark = 0 then Receive(tick 0 , num0 ); Send tick =conc(tick 0 , up), tick =conc(tick 0 , right), tick =conc(tick 0 , down), tick =conc(tick 0 , lef t) to p0 , p1 , p2 p3 respectively; Send num = num0 + 0, num = num0 + 1, num = num0 + 0, num = num0 − 1 to p0 , p1 , p2 , p3 respectively; Send ack = 1 to the node who sent you tick; mark = 1; path = m where m is the number of the port that received the tick 0 − num0 ; message; end if O. Michail, G. Skretas, and P. G. Spirakis 27: 28: 29: 30: 31: 32: 33: 34: 35: 36: 37: 38: 39: 40: 41: 42: 43: 44: 45: 46: 47: 48: 49: 50: 51: 52: 53: 54: 55: 56: 57: 58: 59: 60: 61: 62: 63: 64: 65: if mark=1 then Receive(ack 0 , tick 0 , qu0 ); if tick 0 was not null then Send(ack = 1) to ppath ; else if ack 0 was not null then Send(ack = 1) to ppath ; end if if qu = 0 then if p0 has neighbour then Sendqu = approve; else Sendqu = reject; end if end if if qu = 1 then if p1 has neighbour then Sendqu = approve; else Sendqu = reject; end if end if if qu = 2 then if p2 has neighbour then Sendqu = approve; else Sendqu = reject; end if end if if qu = 3 then if p3 has neighbour then Sendqu = approve; else Sendqu = reject; end if end if Sendf lag to p1 ; end if end while end procedure 39 40 On the Transformation Capability of Feasible Mechanisms for Programmable Matter I Theorem 26. The Compact Line algorithm solves the Line Formation problem for any compact starting shape without breaking the connectivity. Algorithm Description: The operation of the algorithm is split into 3 stages. The orientation stage(1), the line marking stage(2) and the movement stage(3). The first stage consists of the leader starting from a random node. It sets the orientation for the current node by marking ports 0,1,2,3 as “north”, “east”, “south”, “west” respectively. It then sends the orientation to all neighbours. All nodes receiving the orientation change their ports to coincide with the one the leader defined, and then propagate the message to their neighbours. Once a node sets its orientation once, it ignores all messages that concern it. In the second stage the leader searches for the rightmost node. It begins by broadcasting two messages to all neighbours: tick and num. The tick message consists of the direction the message was sent to. The num message is a number which starts as 0 and each time it is propagated through nodes, we add the following number: 0 for north , +1 for east, 0 for south and −1 for west. When a non leader node receives these messages, it propagates them to its neighbours after appending up, right, down, left, for neighbours 0,1,2,3 respectively, to the tick message and after adding the number to the num message following the method mentioned above. The node also sends a message called ack to the node who sent the tick and num. It then stores the node (path node) who sent the tick and marks himself. When a marked node receives a tick-num message it sends them to the path node along with an ack message. When the leader receives a num-tick message, it compares the num it received with the num0 it has in store. If the one it received is bigger, it replaces the num0 with num and keeps the tick0 message it received. Now, if the leader does not receive an ack for two consecutive rounds it starts following the path it has stored in the variable named line. Once it reaches the destination it marks the current node and starts moving west, marking all nodes in its path. It then returns to the node it marked first. The leader has now marked a designated line where it will move all other nodes to. This ends phase 2. The third stage consists of a loop being performed until all nodes form a line. Loop description: The leader moves randomly to nodes checking if they are on the correct line (f lag = 1). If it finds one and receives a message (flag0 =1), the leader marks it. If it finds one does not receive a message (flag=1), it checks two things. First it checks if the node has only one neighbour. Secondly it checks if the node has two neighbours not opposite to each other. If it does complete the second requirement, it sends a qu message to one of them asking it if the 2 nodes who are neighbours to it (the leader), have another common neighbour. The node then answers approve or reject. If any of those two checks are true (one neighbour, approve) the leader moves in a random fashion. Once it receives a message flag0 =1, it marks the node. That finishes the loop description. Proof. The goal is to show that any compact shape will always end up in a line. We need to show that the connectivity will be preserved throughout the transformation and the shape will not get stuck in a shape which is not a line. The first two phases cannot break the connectivity because no movement takes place. The leader is moving between nodes. The third phase consists of the loop. The loop has three phases. The lib movement, the search for the line, and the placement on the line. Observe that if a node performs a lib movement and iy does not break the connectivity, all its subsequent moves will not break it as long as no other node has moved. This guarantees that if the lib movement preserves the connectivity, the second and third phase will preserve it as well. Now in both checks performed by the algorithm, it is ensured that the node has no bridges so any movement it performs will not break the connectivity. O. Michail, G. Skretas, and P. G. Spirakis The first two phases will not turn the shape into an undesired one, because no movement takes place. All we have to show is that the loop won’t do it as well. The first case is whether a lib movement will always be available. Observe that there is no shape where all nodes have 3 or 4 neighbours. In this protocol we have a lib movement available when a node has only 1 neighbour or when a cluster of 4 nodes creates a square. Suppose that there is a shape that has no lib movements available. Every node has at least 2 neighbours. Such a shape can only be a cycle consisting of all nodes. This cycle is either a non compact shape or it consists of only 4 nodes. But both of those circumstances are prohibited because we have established that we are talking about compact shapes where there are no lib movements. Thus a lib movement is always available. The second case is whether the moving node will always find the right point of rightmost node of the line. Since the moving node is moving in random it will not get stuck in a loop. The rightmost node of the line is always accessible from any node on the shape. The third problem is whether the move on the rightmost part of the line creates a non compact shape. Since the line begins from the rightmost part of the shape, each time a node moves east of it, it has only one neighbour which is the rest of the line. So it cannot create a non-compact shape. J I Theorem 27. The Line Transform algorithm solves the Line Formation problem for any starting shape without breaking the connectivity. 7 Conclusions and Further Research There are many open problems related to the findings of the present work. We here restricted attention to the two extremes, in which the transformation either preserves connectivity or is free to break it arbitrarily. A compromise could be to allow some restricted degree of connectivity breaking, like necessarily restoring it in at most k ≥ 0 steps (a special case of this had been already proposed as an open question in [15]). There are other meaningful “good” properties that we would like to maintain throughout a transformation. An interesting example, is the strength of the shape. One of the various plausible definitions is as the minimum strength sub-shape of the shape (i.e., its weakest part; could possibly be captured by some sort of minimum geometric cuts). Then, a strength-preserving transformation would be one that reaches the target shape while trying to maximize this minimum. In the transformations considered in this paper, there was no a priori constraint on the maximum area that a transformation is allowed to cover or on the maximum dimensions that its intermediate shapes are allowed to have. It seems in general harder to achieve a particular transformation if any of these restrictions is imposed. For example, the generic transformation of Theorem 6 requires some additional space below the shape and the transformations of Theorems 14 and 19 convert any shape first to a spanning line, whose maximum dimension is √ n, even though the original shape could have a maximum dimension as small as n. Another interesting fact about restricting the boundaries is that in this way we get models equivalent to several interesting puzzles. For example, if the nodes are labeled, the initial shape is a square with a single empty cell, and the boundaries are restricted to the dimensions of the square, we get a generalization of the famous 15-puzzle (see, e.g., [9] for a very nice exposition of this and many more puzzles and 2-player games). Techniques developed in the context of puzzles could prove valuable for analyzing and characterizing discrete programmable matter systems. We intentionally restricted attention to very minimal actuation mechanisms, namely rotation and sliding. More sophisticated mechanical operations would enable a larger set 41 42 On the Transformation Capability of Feasible Mechanisms for Programmable Matter 1: procedure Line-Transform(Leader) 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: 27: 28: 29: 30: 31: 32: 33: 34: 35: 36: 37: 38: 39: 40: 41: 42: 43: 44: 45: 46: 47: label ← 0, phase ← 0, check ← 0, state ← 0, buck ← 0, num ← 0, line ← 0 while r = 1 do p0 ← north, p1 ← east, p2 ← south, p3 ← west, orien(node) = 1; Send north, east, south, west to p0 , p1 , p2 , p3 respectively; end while while r = 2, 3 AND buck = 0 do Send tick = up, right, down, lef t and num = 0, 1, 0, −1 to p0 , p1 , p2 , p3 respectively; Receive (tick 0 , num0 , ack 0 ); Do for every tick 0 , num0 received; if num0 > num then line = tick 0 ; num = num0 ; end if if ack 0 = null then state + +; else state − −; end if if state ≥ 2 then f ollow the path described on the line; end if if line has been reached then buck = 1; end if end while while r = 2, 3, .. AND buck = 1 do Receive(f lag 0 , qu0 ); if label = 0 then move west; f lag(node) = 1; if no node available west then label + +; end if end if if label = 1 then move east; f lag(node) = 1 if no node available east then label + +; end if end if if ( thenf lag 0 = 0 AND f lag = 0) if label = 2 then if neighbours = 1 then label + +; else if neighbours = 2 AND neighbours not opposite then if neighbours are p0 , p1 then send qu = 1 to p0 ; label + +; end if O. Michail, G. Skretas, and P. G. Spirakis if neighbours are p1 , p2 then send qu = 1 to p1 ; label + +; 50: end if 51: if neighbours are p2 , p3 then 52: send qu = 1 to p2 ; label + +; 53: end if 54: if neighbours are p3 , p0 then 55: send qu = 1 to p3 ; label + +; 56: end if 57: else if neighbour = 2 AND neighboursopposite then 58: if neighbours are p0 , p2 then 59: Sendqu = 2 to p0 ; label + +; 60: end if 61: if neighbours are p1 , p3 then 62: Sendqu = 3 to p1 ; label + +; 63: end if 64: else 65: move; 66: end if 67: end if 68: if label = 3 AND phase = 0 then 69: if ack 0 is null then 70: wait + +; 71: end if 72: if ack 0 is not null then 73: wait − −; 74: end if 75: if wait = 2 then 76: label − −; move; 77: end if 78: if qu0 = k where k is the number of the port from where the qu’ was received then 79: phase = 1; Sendclear = 1 to all ports; 80: end if 81: end if 82: if label = 3 AND phase = 1 then 83: travel; 84: end if 85: end if 86: if f lag 0 = 1 OR f lag = 1 then 87: f lag = 1; phase = 0; label = 2; travel; 88: end if 89: end while 90: end procedure 48: 49: 43 44 On the Transformation Capability of Feasible Mechanisms for Programmable Matter 1: procedure Line-Transform(Non-Leader) 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: 27: 28: 29: 30: 31: 32: 33: 34: 35: 36: 37: 38: f lag ← 0, orien ← 0, mark ← 0 while orien = 0 do Receive(north0 , south0 , east0 , west0 , tick 0 , num0 , qu0 ) if p0 receives south OR p1 receives west OR p2 receives north OR p3 receives east then orien = 1; end if if p0 receives east OR p1 receives south OR p2 receives west OR p3 receives north then pn = p(n−1)mod3 ; orien = 1; end if if p0 receives north OR p1 receives east OR p2 receives south OR p3 receives swest then pn = p(n−2)mod3 ; orien = 1; end if if p0 receives west OR p1 receives north OR p2 receives east OR p3 receives south then pn = p(n−3)mod3 ; orien = 1; end if Sendnorth, east, south, west to p0 , p1 , p2 , p3 respectively; end while while orien = 1 do if mark = 0 then Receive(tick 0 , num0 ); Send tick =conc(tick 0 , up), tick =conc(tick 0 , right), tick =conc(tick 0 , down), tick =conc(tick 0 , lef t) to p0 , p1 , p2 p3 respectively; Send num = num0 + 0, num = num0 + 1, num = num0 + 0, num = num0 − 1 to p0 , p1 , p2 , p3 respectively; Send ack = 1 to the node who sent you tick mark = 1; path = m where m is the number of the port that received the tick 0 − num0 ; message; end if if mark=1 then Receive(ack 0 , tick 0 , qu0 ); if tick 0 was not null then Send(ack = 1) to ppath ; else if ack 0 was not null then Send(ack = 1) to ppath ; end if if qu0 = 0, 1, 2, 3 then path = m where m is the number of the port that received the qu0 ; message; Sendqu = qu0 to all ports; Sendack = 1 to portpath ; mark = 2; end if Send f lag to p1 ; end if O. Michail, G. Skretas, and P. G. Spirakis 39: 40: 41: 42: 43: 44: 45: 46: 47: 48: 49: if mark = 2 then Sendf lageast; Receive(ack 0 , tick 0 , clear0 ); if ack 0 was not blank then Send ack = 1 to ppath end if if clear0 = 1 from ppath then Sendclear = 1 to all ports; path = null; mark = 1; end if end if end while end procedure of transformations and possibly also reduce the time complexity. Such examples, could be the ability of a node to become inserted between two neghboring nodes (while pushing them towards opposite directions). This could enable parallel mergings of two lines of length n/2 into a line of length n in a single step (an, thus, for example, transforming a square to a line in polylogarithmic time). Another, is the capability of rotating whole lines of nodes (like rotating arms, see, e.g., [36]). There are also some promising specific technical questions. We do not yet know what is the complexity of RotC-Transformability. The fact that a 6-seed is capable of transfering pairs of nodes to desired positions, suggests that shapes having such a seed in their exterior or being capable of self-extracting such a seed, will possibly be able to transform to each other. Even if this turns out to be true, it is totally unclear whether transformations involving at least one of the rest of the shapes are feasible. Moreover, we didn’t study the problem of computing or approximating the optimum transformation. It seems that the problem is computationally hard. A possible approach to prove NP-hardness would be by proving NP-hardness of Rectilinear Graphic TSP (could be via a reduction from Rectilinear Steiner Tree or Rectilinear TSP, which are both known to be NP-complete [20]) and then giving a reduction from that problem to the problem of a 2-seed exploring a set of locations on the grid. Finally, regarding the distributed transformations, there are various interesting variations of the model considered here, that would make sense. One of them is to assume nodes that are oblivious w.r.t. their orientation. 45 46 On the Transformation Capability of Feasible Mechanisms for Programmable Matter References 1 Greg Aloupis, Nadia Benbernou, Mirela Damian, Erik D Demaine, Robin Flatland, John Iacono, and Stefanie Wuhrer. Efficient reconfiguration of lattice-based modular robots. Computational geometry, 46(8):917–928, 2013. 2 Dana Angluin, James Aspnes, Zoë Diamadi, Michael J. Fischer, and René Peralta. Computation in networks of passively mobile finite-state sensors. Distributed Computing, 18(4):235–253, March 2006. 3 Dana Angluin, James Aspnes, David Eisenstat, and Eric Ruppert. The computational power of population protocols. Distributed Computing, 20(4):279–304, November 2007. 4 Zack Butler, Keith Kotay, Daniela Rus, and Kohji Tomita. Generic decentralized control for lattice-based self-reconfigurable robots. The International Journal of Robotics Research, 23(9):919–937, 2004. 5 Xuli Chen, Li Li, Xuemei Sun, Yanping Liu, Bin Luo, Changchun Wang, Yuping Bao, Hong Xu, and Huisheng Peng. Magnetochromatic polydiacetylene by incorporation of fe3o4 nanoparticles. Angewandte Chemie International Edition, 50(24):5486–5489, 2011. 6 Mark Cieliebak, Paola Flocchini, Giuseppe Prencipe, and Nicola Santoro. Solving the robots gathering problem. In International Colloquium on Automata, Languages, and Programming, pages 1181–1196. Springer, 2003. 7 Alejandro Cornejo, Fabian Kuhn, Ruy Ley-Wild, and Nancy Lynch. Keeping mobile robot swarms connected. In Proceedings of the 23rd international conference on Distributed computing, DISC’09, pages 496–511, Berlin, Heidelberg, 2009. Springer-Verlag. 8 Shantanu Das, Paola Flocchini, Nicola Santoro, and Masafumi Yamashita. Forming sequences of geometric patterns with oblivious mobile robots. Distributed Computing, 28(2):131–145, April 2015. 9 Erik D Demaine. Playing games with algorithms: Algorithmic combinatorial game theory. In International Symposium on Mathematical Foundations of Computer Science, pages 18– 33. Springer, 2001. 10 Zahra Derakhshandeh, Shlomi Dolev, Robert Gmyr, Andréa W Richa, Christian Scheideler, and Thim Strothmann. Brief announcement: amoebot–a new model for programmable matter. In Proceedings of the 26th ACM symposium on Parallelism in algorithms and architectures (SPAA), pages 220–222, 2014. 11 Zahra Derakhshandeh, Robert Gmyr, Alexandra Porter, Andréa W Richa, Christian Scheideler, and Thim Strothmann. On the runtime of universal coating for programmable matter. In International Conference on DNA-Based Computers, pages 148–164. Springer, 2016. 12 Zahra Derakhshandeh, Robert Gmyr, Andréa W Richa, Christian Scheideler, and Thim Strothmann. An algorithmic framework for shape formation problems in self-organizing particle systems. In Proceedings of the Second Annual International Conference on Nanoscale Computing and Communication, page 21. ACM, 2015. 13 David Doty. Theory of algorithmic self-assembly. Communications of the ACM, 55:78–88, 2012. 14 David Doty. Timing in chemical reaction networks. In Proc. of the 25th Annual ACM-SIAM Symp. on Discrete Algorithms (SODA), pages 772–784, 2014. 15 Adrian Dumitrescu and János Pach. Pushing squares around. In Proceedings of the twentieth annual symposium on Computational geometry, pages 116–123. ACM, 2004. 16 Adrian Dumitrescu, Ichiro Suzuki, and Masafumi Yamashita. Formations for fast locomotion of metamorphic robotic systems. The International Journal of Robotics Research, 23(6):583–593, 2004. O. Michail, G. Skretas, and P. G. Spirakis 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 Adrian Dumitrescu, Ichiro Suzuki, and Masafumi Yamashita. Motion planning for metamorphic systems: Feasibility, decidability, and distributed reconfiguration. IEEE Transactions on Robotics and Automation, 20(3):409–418, 2004. Yuval Emek and Jara Uitto. Dynamic networks of finite state machines. In International Colloquium on Structural Information and Communication Complexity, pages 19–34. Springer, 2016. Paola Flocchini, Giuseppe Prencipe, and Nicola Santoro. Distributed computing by oblivious mobile robots. Synthesis lectures on distributed computing theory, 3(2):1–185, 2012. Michael R Garey, Ronald L Graham, and David S Johnson. Some NP-complete geometric problems. In Proceedings of the eighth annual ACM symposium on Theory of computing, pages 10–22. ACM, 1976. Kyle Gilpin, Ara Knaian, and Daniela Rus. Robot pebbles: One centimeter modules for programmable matter through self-disassembly. In Robotics and Automation (ICRA), 2010 IEEE International Conference on, pages 2485–2492. IEEE, 2010. Robert A Hearn and Erik D Demaine. PSPACE-completeness of sliding-block puzzles and other problems through the nondeterministic constraint logic model of computation. Theoretical Computer Science, 343(1-2):72–96, 2005. Camille Jordan. Cours d’analyse de l’École polytechnique, volume 1. Gauthier-Villars et fils, 1893. Ara N Knaian, Kenneth C Cheung, Maxim B Lobovsky, Asa J Oines, Peter SchmidtNeilsen, and Neil A Gershenfeld. The milli-motein: A self-folding chain of programmable matter with a one centimeter module pitch. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1447–1453. IEEE, 2012. Evangelos Kranakis, Danny Krizanc, and Euripides Markou. The mobile agent rendezvous problem in the ring. Synthesis Lectures on Distributed Computing Theory, 1(1):1–122, 2010. Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83–97, 1955. Yunfeng Lu, Yi Yang, Alan Sellinger, Mengcheng Lu, Jinman Huang, Hongyou Fan, Raid Haddad, Gabriel Lopez, Alan R Burns, Darryl Y Sasaki, et al. Self-assembly of mesoscopically ordered chromatic polydiacetylene/silica nanocomposites. Nature, 410(6831):913–917, 2001. Othon Michail and Paul G. Spirakis. Simple and efficient local codes for distributed stable network construction. Distributed Computing, 29(3):207–237, 2016. Paul W. K. Rothemund and Erik Winfree. The program-size complexity of self-assembled squares. In Proceedings of the 32nd annual ACM symposium on Theory of computing (STOC), pages 459–468, 2000. Paul WK Rothemund. Folding dna to create nanoscale shapes and patterns. Nature, 440(7082):297–302, 2006. Michael Rubenstein, Alejandro Cornejo, and Radhika Nagpal. Programmable self-assembly in a thousand-robot swarm. Science, 345(6198):795–799, 2014. Walter J Savitch. Relationships between nondeterministic and deterministic tape complexities. Journal of computer and system sciences, 4(2):177–192, 1970. Masahiro Shibata, Toshiya Mega, Fukuhito Ooshita, Hirotsugu Kakugawa, and Toshimitsu Masuzawa. Uniform deployment of mobile agents in asynchronous rings. In Proceedings of the 2016 ACM Symposium on Principles of Distributed Computing, pages 415–424. ACM, 2016. Ichiro Suzuki and Masafumi Yamashita. Distributed anonymous mobile robots: Formation of geometric patterns. SIAM J. Comput., 28(4):1347–1363, March 1999. Erik Winfree. Algorithmic Self-Assembly of DNA. PhD thesis, California Institute of Technology, June 1998. 47 48 On the Transformation Capability of Feasible Mechanisms for Programmable Matter 36 37 38 Damien Woods, Ho-Lin Chen, Scott Goodfriend, Nadine Dabby, Erik Winfree, and Peng Yin. Active self-assembly of algorithmic shapes and patterns in polylogarithmic time. In Proceedings of the 4th conference on Innovations in Theoretical Computer Science, pages 353–354. ACM, 2013. Yukiko Yamauchi, Taichi Uehara, and Masafumi Yamashita. Brief announcement: pattern formation problem for synchronous mobile robots in the three dimensional euclidean space. In Proceedings of the 2016 ACM Symposium on Principles of Distributed Computing, pages 447–449. ACM, 2016. Mark Yim, Wei-Min Shen, Behnam Salemi, Daniela Rus, Mark Moll, Hod Lipson, Eric Klavins, and Gregory S Chirikjian. Modular self-reconfigurable robot systems [grand challenges of robotics]. IEEE Robotics & Automation Magazine, 14(1):43–52, 2007.
8cs.DS
arXiv:1710.01368v1 [math.AG] 3 Oct 2017 PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA ANNE LONJOU Résumé. L’action du groupe de Cremona de rang 2 sur un espace hyperbolique de dimension infinie est récemment devenue centrale dans l’étude de ce groupe. Guidée par l’analogie avec l’action de PSL(2, Z) sur le demi-plan de Poincaré, nous exhibons un domaine fondamental pour cette action en considérant un pavage de Voronoï. Nous étudions ensuite les cellules adjacentes à une cellule de Voronoï donnée, ainsi que les cellules qui partagent des points communs au bord à l’infini. Abstract. The action of the Cremona group of rank 2 on an infinite dimensional hyperbolic space is the main recent tool to study the Cremona group. Following the analogy with the action of PSL(2, Z) on the Poincaré half-plane, we exhibit a fundamental domain for this action by considering a Voronoï tessellation. Then we study adjacent cells to a given cell, as well as cells that share common points in the boundary at infinity. Table des matières Introduction Remerciements 1. Préliminaires 2. Construction d’un pavage de Voronoï 3. Cellule de Voronoï associée à l’identité 4. Cellules adjacentes à la cellule associée à l’identité 5. Cellules quasi-adjacentes à la cellule associée à l’identité Références 1 5 6 19 26 44 49 57 Introduction Le groupe de Cremona de rang 2 sur un corps k, noté Bir(P2k ), est le groupe des transformations birationnelles du plan projectif. Dans l’étude qui nous intéresse, l’action du groupe modulaire PSL(2, Z) sur le demi-plan de Poincaré s’est avérée un guide précieux. Dans la suite de l’introduction, nous décrivons diverses analogies entre PSL(2, Z) et le groupe de Cremona. Date: 5 octobre 2017. 2010 Mathematics Subject Classification. 14E07, 20F65. 1 PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA Générateurs. Soient  0 S= 1   −1 1 , T = 0 0   1 1 et U = 1 1 2  −1 0 trois éléments de PSL(2, Z). Les matrices S et U sont respectivement d’ordre 2 et 3 alors que la matrice T est d’ordre infini. Le groupe modulaire est de présentation finie. Il est par exemple engendré par les matrices S et U et les relations sont engendrées par les relateurs S 2 = I2 et U 3 = I2 , ou encore par les matrices S et T avec les relations engendrées par S 2 et (T S)3 : PSL(2, Z) = hS, U | S 2 et U 3 i ≃ hS, T | S 2 et (T S)3 i. Bien que le groupe de Cremona possède un système de générateurs connu, ce groupe n’est pas de type fini et ceci est valable sur n’importe quel corps (voir [Can17, Proposition 3.6]). Lorsque le corps k est algébriquement clos, d’après le théorème de Noether-Castelnuovo, le groupe de Cremona est engendré par le groupe des automorphismes du plan projectif isomorphe à PGL(3, k) et par l’involution quadratique standard qui dans une carte affine s’écrit : σ : (x, y) 799K ( x1 , y1 ). Un autre système de générateurs qui se trouve être plus pratique dans certaines circonstances est PGL(3, k) ainsi que le sous-groupe de Jonquières. Un élément du groupe de Jonquières est une application préservant le pinceau de droites {y = constante}. À noter que l’application σ appartient au groupe de Jonquières. Action sur un espace hyperbolique. Considérons le plan hyperbolique H2 . Un de ses modèles est le demi-plan de Poincaré qui est défini comme le sous-espace du plan complexe constitué des nombres complexes de partie imaginaire strictement positive : H2 = {z ∈ C | Im(z) > 0}. Il est muni de la métrique définie de la façon suivante. Pour tous z1 , z2 ∈ H2 ,   (Im(z1 ) − Im(z2 ))2 + (Re(z1 ) − Re(z2 ))2 d(z1 , z2 ) = argcosh 1 + . 2 Im(z1 ) Im(z2 )   a b 2 Le groupe modulaire agit sur H par isométries : pour tout ∈ PSL(2, Z) et c d 2 pour tout point z ∈ H ,   az + b a b . ·z = c d cz + d Les points à l’infini ou appelés également points au bord sont les points dont la partie imaginaire est nulle ainsi que le point noté ∞ qui permet de compactifier la droite réelle {y = 0}. Pour cette action, les matrices S et U sont des √ isométries elliptiques puisqu’elles fixent respectivement les points i ∈ H2 et 21 (1 + i 3) ∈ H2 . La matrice T est une isométrie parabolique car elle fixe un unique point au bord, le point ∞. Une matrice est hyperbolique si elle fixe deux points à l’infini comme  √ √ 1 2 par exemple la matrice qui fixe les deux points réels −1 − 2 et −1 + 2. 2 5 Un autre de ses modèles est le modèle de l’hyperboloïde. Il est défini comme la nappe d’hyperboloïde H2 := {(x, y, z) ∈ R3 | x2 − y 2 − z 2 = −1 et x > 0}, PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 3 munie de la distance d(u, v) = argcosh B(u, v), pour tous u, v ∈ H2 où B(·, ·) est la forme bilinéaire associée à la forme quadratique précédente. Le groupe de Cremona agit également par isométries sur un espace hyperbolique qui est un analogue de dimension infinie au modèle de l’hyperboloïde de H2 . L’espace de Picard-Manin associé à P2 est la limite inductive des groupes de Picard des surfaces obtenues en éclatant toute suite finie de points de P2 , infiniment proches ou non. Il est muni d’une forme d’intersection de signature (1, ∞). En considérant une nappe d’hyperboloïde, nous pouvons lui associer un espace hyperbolique de dimension infinie, noté H∞ . Les éléments de PGL(3, k) sont tous elliptiques, ceux du groupe de Jonquières sont elliptiques lorsque les degrés de leurs itérées sont bornés et paraboliques sinon. L’application de Hénon hn : (x, y) 7→ (y, y n − x) est un exemple d’élément hyperbolique. Non-simplicité du groupe de Cremona. Le groupe PSL(2, Z) n’est pas un groupe simple. En effet, pour tout entier N > 1, le sous-groupe Γ(N ) = {A ∈ PSL(2, Z) | A ≡ ± I2 (mod N )}, est un sous-groupe distingué de PSL(2, Z). En fait, il possède également de nombreux sous-groupes distingués d’indice infini puisque c’est un groupe SQ-universel, c’est-à-dire que tout groupe dénombrable se plonge dans un quotient de PSL(2, Z). En 2013, en faisant agir le groupe de Cremona sur H∞ , S. Cantat et S. Lamy montrent que lorsque le corps est algébriquement clos, le groupe de Cremona n’est pas simple (voir [CL13]). Ils construisent des sous-groupes propres dont tous les éléments sont de grands degrés. Pour montrer cela, ils élaborent une variante de la théorie de petite simplification. Récemment, nous avons étendu ce résultat à un corps quelconque (voir [Lon16]), en utilisant un théorème de petite simplification dû à F. Dahmani, V. Guirardel et D. Osin (voir [DGO17]) établi dans le cadre général de groupes agissant par isométries sur des espaces Gromov-hyperboliques. Nous sommes dans ce cadre d’une part car l’espace H2 est hyperbolique au sens de Gromov. En effet, chacun de √ ses triangles vérifie la propriété suivante : tout côté est contenu dans le ln(1 + 2)-voisinage de la réunion de ses deux autres côtés. ∞ 2 ∞ D’autre part, un √ triangle de H vivant dans une copie de H , l’espace H est lui aussi ln(1 + 2)-hyperbolique. En fait, comme conséquence des résultats de F. Dahmani, V. Guirardel et D. Osin, nous obtenons dans [Lon16] des propriétés plus précises que la seule non-simplicité, à savoir que pour k un corps quelconque, le groupe Bir(P2k ) contient des sous-groupes distingués libres, et est SQ-universel. Pavage de Voronoï. L’espace H∞ est central dans l’étude du groupe de Cremona, cependant un domaine fondamental pour l’action du groupe de Cremona sur H∞ n’avait jusqu’alors pas été étudié. Un domaine fondamental est un sous-espace fermé de H∞ tel que son orbite sous l’action du groupe de Cremona recouvre H∞ et que les éléments de l’orbite de son intérieur soient deux à deux disjoints. En fait, il va s’avérer plus commode de travailler avec les cellules de Voronoï. Le but de cet article est de construire et étudier un pavage de Voronoï associé à cette action. Lorsqu’un groupe agit isométriquement et discrètement sur un espace métrique géodésique et qu’il existe un point de l’espace dont le stabilisateur est réduit au neutre, « les cellules de Voronoï » sont un outil naturel pour construire un domaine fondamental pour cette action. PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 4 Définition. Soit P un ensemble discret de points d’un espace métrique géodésique X. À tout point p de P nous associons un ensemble de X, noté V(p) et appelé cellule de Voronoï associée au point p, constitué des points de X qui sont plus proches de p que des autres points de P : V(p) = {x ∈ X | pour tout q ∈ P, d(x, p) ≤ d(x, q)}. Les points p sont les « centres » des cellules de Voronoï. En effet, en considérant l’orbite d’un point dont le stabilisateur est réduit au neutre, nous obtenons un ensemble discret de points de notre espace métrique. La cellule de Voronoï associée à un point de cette orbite est un domaine fondamental. Revenons à l’action du groupe modulaire sur le demi-plan de Poincaré H2 . Il est bien connu que l’ensemble des points qui se situent au-dessus de la géodésique dont les points à l’infini sont −1 et 1, et qui ont une partie réelle supérieure ou égale à − 12 et inférieure ou égale à 21 est un domaine fondamental. Considérons l’orbite du point 2i sous l’action de PSL(2, Z). Les cellules de Voronoï correspondant à cet ensemble de points coïncident avec l’orbite du domaine fondamental donné plus haut. Cependant, si nous considérons l’orbite du point i sous l’action de PSL(2, Z), la cellule de Voronoï correspondant au point i n’est plus le domaine fondamental de l’action du groupe modulaire sur H2 , mais l’orbite de ce dernier par le stabilisateur de i qui est isomorphe à Z/2Z. L’objet de la section 2, est de construire un pavage de Voronoï associé à l’action du groupe de Cremona sur H∞ . Notons ℓ ∈ H∞ la classe d’une droite de P2 . Nous considérons les cellules de Voronoï associées à l’orbite de ℓ par l’action du groupe de Cremona. Le stabilisateur de ℓ étant PGL(3, k), nous identifions deux applications qui diffèrent par un automorphisme : f ∼ g ⇔ il existe a ∈ PGL(3, k), f = g ◦ a. La cellule de Voronoï associée à ℓ correspond à l’orbite d’un domaine fondamental sous l’action de PGL(3, k). Remarquons que deux applications appartenant à la même classe d’équivalence donnent la même cellule de Voronoï. Soit f appartenant à Bir(P2 ), nous notons V(f ) la cellule associée à f et l’application f est appelée « germe ». Dans cette section nous nous restreignons à un sous-espace convexe de H∞ qui est essentiellement l’enveloppe convexe de l’orbite de ℓ sous l’action de Bir(P2 ). Nous prenons en fait un convexe E un peu plus gros pour des raisons techniques (voir la définition de E à la sous-section 2.1). Nous étudions ensuite, lors de la section 3, les classes appartenant à la cellule de Voronoï associée à l’application identité. Le théorème principal de cette partie dit qu’il suffit que les classes soient plus proches de ℓ que de l’orbite de ℓ sous l’action des applications de Jonquières (et non pas de toutes les applications du groupe de Cremona). Théorème A. Une classe c appartient à la cellule V(id) si et seulement si pour toute application de Jonquières j, d(j# (c), ℓ) ≥ d(c, ℓ). En fait, le théorème 3.8 est plus précis que cela. Il suffit de vérifier l’inégalité pour certaines applications de Jonquières. Dans la section 4, nous déterminons les cellules non disjointes de la cellule V(id) (Corollaire 4.7) que nous appelons « cellules adjacentes ». Les germes de telles PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 5 cellules sont de deux types. Ceux qui sont de caractéristique Jonquières, c’est-àdire les applications du groupe de Cremona telles qu’il existe deux points p et q dans P2 et qui envoient le pinceau de droites passant par le point p sur le pinceau de droites passant par le point q. L’autre type de germes des cellules adjacentes à la cellule V(id) sont les applications du groupe de Cremona qui possèdent au plus huit points-base en position presque générale. Un ensemble de points {p0 , p1 , . . . , pr } est dit en position presque générale si d’une part pour chaque 0 ≤ i ≤ r le point pi vit soit dans P2 soit dans une surface dominant P2 qui est obtenue en éclatant seulement un sous-ensemble de points de {p0 , . . . , pr } et si d’autre part aucune des trois conditions suivantes n’est satisfaite : quatre des points de cet ensemble sont alignés, sept des points de cet ensemble sont sur une conique, deux des points de cet ensemble sont adhérents à un troisième point de cet ensemble. Théorème B. L’ensemble des germes des cellules adjacentes à la cellule V(id) est constitué de toutes : • les applications de caractéristique Jonquières, • les applications qui possèdent au plus 8 points-base en position presque générale. Nous déterminons également les classes qui se trouvent à l’intersection entre la cellule de Voronoï associée à ℓ et une ou plusieurs de ses cellules adjacentes (Théorème 4.1). Enfin dans la section 5 nous étudions les cellules qui possèdent une classe en commun à l’infini avec la cellule associée à l’identité. Revenons un instant à l’action du groupe modulaire sur le demi-plan de Poincaré et considérons le pavage de Voronoï associé à l’orbite du point 2i. Les cellules quasi-adjacentes à la cellule associée à 2i sont toutes les cellules obtenues comme image par le sous-groupe engendré par z → z + 1 de la cellule associée à 2i. Le théorème principal de cette section est Théorème C. L’ensemble des germes des cellules quasi-adjacentes à la cellule V(id) est constitué de toutes : • les applications de caractéristique Jonquières, • les applications qui possèdent au plus 9 points-base en position presque générale. De même que dans le cas précédent, nous étudions les classes à l’infini qui se trouvent dans l’intersection des bords à l’infini de la cellule associée à ℓ et d’une ou de plusieurs de ses cellules quasi-adjacentes. Une motivation pour cette étude était la question de l’hyperbolicité de certains graphes associés au groupe de Cremona, en particulier un graphe dû à Wright. Par rapport à l’analogie avec le groupe modulaire, ce graphe joue le rôle de l’arbre de Bass-Serre associé à la structure de produit amalgamé de PSL(2, Z). Dans un article à venir (voir également [Lon17]), nous ferons le lien entre le graphe de Wright et le graphe dual du pavage de Voronoï, et nous étudierons son hyperbolicité au sens de Gromov. Remerciements Je remercie vivement Stéphane Lamy, mon directeur de thèse, pour sa grande disponibilité et ses relectures minutieuses. Je remercie également les rapporteurs de PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 6 ma thèse, Charles Favre et Yves de Cornulier, pour leurs remarques qui ont permis de rendre certains passages plus clairs. 1. Préliminaires Dans cet article le corps de base, noté k, est algébriquement clos. Les surfaces considérées sont projectives et lisses. Le groupe de Cremona Bir(P2k ) est le groupe des applications birationnelles de 2 P = P2k vers lui-même. Un élément du groupe de Cremona s’écrit : f: P2 99K P2 [x : y : z] 799K [f0 (x, y, z) : f1 (x, y, z) : f2 (x, y, z)], où f0 , f1 , f2 ∈ k[x, y, z] sont des polynômes homogènes, de même degré et sans facteur commun. Nous appelons degré de f le degré des polynômes homogènes : deg(f ) := deg(fi ) pour i ∈ {0, 1, 2}. Par exemple, l’application quadratique standard σ : [x : y : z] 99K [yz : xz : xy] est un élément du groupe de Cremona. C’est une application de Jonquières puisqu’elle préserve le pinceau de droites passant par un point. En fait elle préserve trois pinceaux de droites, ceux passant par les points [0 : 0 : 1], [0 : 1 : 0] et [1 : 0 : 0]. Remarquons que l’inverse d’une application de Jonquières est encore une application de Jonquières. Soit S une surface. Une surface S ′ domine S s’il existe un morphisme birationnel allant de S ′ vers S. Considérons S1 et S2 deux surfaces dominant S et π1 et π2 leur morphisme respectif vers S. Nous disons que deux points p1 ∈ S1 et p2 ∈ S2 sont équivalents si π1−1 ◦ π2 est un isomorphisme local sur un voisinage de p2 et envoie p2 sur p1 . L’espace des bulles (« Bubble space » en anglais), noté B(S), est l’union de tous les points de toutes les surfaces dominant S modulo cette relation d’équivalence. Le théorème suivant est dû à O. Zariski et permet de décomposer toute application birationnelle comme composée d’éclatements et d’inverses d’éclatements. Théorème 1.1. Soient S1 et S2 deux surfaces, et f : S1 99K S2 une application birationnelle. Alors il existe une troisième surface S3 et deux composées d’éclatements π : S3 → S1 et σ : S3 → S2 telles que le diagramme suivant commute : S3 π S1 σ f S2 . Les points éclatés lors de la résolution minimale de f sont appelés points-base de f et vivent dans l’espace des bulles. Ils correspondent aux points-base du système linéaire associé à f . Cet ensemble de points est noté Bs(f ). À tout point-base de f est associé un entier appelé multiplicité et qui correspond à la multiplicité du système linéaire en ce point, c’est-à-dire à la plus petite multiplicité en ce point des polynômes définissant f . Le groupe de Cremona agit sur un hyperboloïde dans l’espace de Picard-Manin, noté H∞ . Dans ces préliminaires, nous rappelons la définition de H∞ ainsi que l’action du groupe de Cremona sur H∞ . Nous énonçons aussi des définitions et PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 7 des propriétés sur les points base des applications du groupe de Cremona qui nous seront utiles par la suite. 1.1. Hyperboloïde dans l’espace de Picard-Manin. Nous rappelons ici la construction de l’espace de Picard-Manin. Nous définissons ensuite l’action du groupe Bir(P2 ) sur cet espace. Enfin, nous nous intéressons à un sous-espace de l’espace de Picard-Manin qui est un espace hyperbolique de dimension infinie. Plus de précisions se trouvent dans [BC16, Section 4], [CL13, Part II.4], [Can11, Section 3] et [Lon17, Section 1.2.3]. 1.1.1. Espace de Picard-Manin. Soit S une surface. Nous considérons le groupe de Néron-Severi associé à S et tensorisé par R. Nous le notons encore N1 (S). C’est donc le groupe des diviseurs à coefficients réels sur S à équivalence numérique près. Il est muni d’une forme bilinéaire symétrique, la forme d’intersection. Pour tout diviseur D sur S nous notons {D}S sa classe de Néron-Severi ou {D} s’il n’y a pas d’ambiguïté sur la surface. Si π : S ′ −→ S est un morphisme birationnel entre deux surfaces, alors le tiré en arrière π ∗ : N1 (S) ֒→ N1 (S ′ ) qui à la classe d’un diviseur associe la classe de sa transformée totale, est un morphisme injectif qui préserve la forme d’intersection. De plus, N1 (S ′ ) est isomorphe à π ∗ (N1 (S)) ⊕ ( ⊕ R{Ep∗ }), p∈Bs(π −1 ) où Bs(π ) est l’ensemble des points-base de π −1 (infiniment proches ou pas) et Ep∗ est la transformée totale, vue dans S ′ , du diviseur exceptionnel Ep obtenu en éclatant le point p. Cette somme est orthogonale relativement à la forme d’intersection. Considérons la limite inductive des groupes de Néron-Severi des surfaces S ′ dominant S : Z C (S) = lim N1 (S ′ ), −1 −→ S ′ →S où l’indice C fait référence aux b-diviseurs de Cartier (pour plus de précisions voir [Fav10]). Remarquons que pour toute surface S ′ dominant S, le groupe N1 (S ′ ) est plongé dans Z C (S). En fait, si nous considérons un diviseur D sur S, à chaque surface S ′ dominant S, nous pouvons lui faire correspondre une classe de Néron-Severi {D}S ′ dans N1 (S ′ ). Ces éléments sont tous identifiés dans Z C (S) et correspondent à une classe d de Z C (S) notée en lettre minuscule. Exemple 1.2. Considérons un point q appartenant au diviseur exceptionnel Ep , issu de l’éclatement d’une surface S au point p. Notons Sp la surface obtenue en éclatant le point p et Sp,q celle en éclatant successivement les points p et q. La classe ep correspond à {Ep }Sp dans N1 (Sp ) et à {Ẽp + Eq }Sp,q dans N1 (Sp,q ) où Ẽp est la transformée stricte de Ep dans Sp,q . Définissons la forme d’intersection sur Z C (S). Pour cela, considérons c et d deux éléments de Z C (S). Il existe une surface S1 dominant S telles que les classes c et d correspondent respectivement à {C}S1 et {D}S1 dans N1 (S1 ). La forme d’intersection est donnée par : c · d = {C}S1 · {D}S1 . Elle ne dépend pas du choix de la surface S1 . PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 8 Par la suite, nous nous intéressons à l’espace de Hilbert défini par X X λ2p < ∞ et {D0 }S ∈ N1 (S)}, λp ep | λp ∈ R, Z(S) = {{D0 }S + p∈B(S) p∈B(S) que nous appelons l’espace de Picard-Manin (voir [CL13] et [Can11] ou encore [BFJ08]). C’est le complété L2 de Z C (S). Ses éléments sont appelés « classes de Picard-Manin » ou plus simplement « classes ». Les classes ep (nous gardons les notations introduites dans l’exemple 1.2) où p est un point de S ou d’une surface dominant S, sont d’auto-intersection −1, orthogonales deux à deux et orthogonales à N1 (S). La forme d’intersection est donc de signature (1, ∞) et préserve la décomposition orthogonale : M  Rep . Z(S) = N1 (S) ⊕ p∈B(S) Tout morphisme birationnel π : S ′ → S induit un isomorphisme π# de Z(S ′ ) vers Z(S) qui consiste à considérer comme exceptionnelles au-dessus de S les classes eq qui étaient dans l’espace de Néron-Severi de S ′ : −1 π# : Z(S) = N1 (S) ⊕ ( ⊕ Rep ) −→ Z(S ′ ) = N1 (S ′ ) ⊕ ( p∈B(S) {D0 }S + P λp ep p∈B(S) 7→ ⊕ r∈B(S) −1 r ∈Bs(π / ) {D̃0 }S ′ + P q∈Bs(π −1 ) Rer ) (mq (D0 ) + λq )eq P + λr er , r∈B(S) −1 r ∈Bs(π / ) où D̃0 est la transformée stricte dans S ′ de D0 et mq (D0 ) la multiplicité de D0 au point q. Dans le cas où la surface considérée est P2 , nous notons simplement Z l’espace de Picard-Manin associé : Z = {nℓ + X p∈B(P2 ) λp ep | n, λp ∈ R, où ℓ est la classe de la droite dans P2 . X p∈B(P2 ) λ2p < ∞}, 1.1.2. Forme canonique. Notons Z L1 l’ensemble des classes de Picard-Manin qui sont L1 : X X |λp | < ∞}. λp ep ∈ Z | Z L1 = {c = nℓ + p∈B(P2 ) p∈B(P2 ) Nous définissons la forme canonique kP2 sur Z L1 , comme la forme linéaire définie par : X λp , pour tout c ∈ Z L1 , kP2 (c) = kP2 ·c = −3n − p∈B(P2 ) où kP2 = −3ℓ + ! ep . Plus généralement, en notant KS le diviseur canonique P ep . d’une surface S dominant P2 , la forme canonique s’écrit kS = KS + P p∈B(P2 ) p∈B(S) PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 9 1.1.3. Action du groupe de Cremona sur l’espace de Picard-Manin. Considérons une résolution de f ∈ Bir(P2 ) : S π P2 σ f P2 . Le groupe Bir(P2 ) agit sur Z(P2 ) via l’application (f, c) 7→ f# (c) où f# est définie par f# = σ# ◦ (π# )−1 . Remarquons que (f# )−1 = (f −1 )# et que l’action de f préserve la forme d’intersection : pour tous c1 , c2 ∈ Z, f# (c1 ) · f# (c2 ) = c1 · c2 . Remarque 1.3. La forme canonique est constante sur l’orbite d’une classe par le groupe de Cremona, c’est-à-dire pour toute classe c ∈ Z L1 et pour tout f ∈ Bir(P2 ), nous avons : kP2 ·c = kP2 ·f# (c). Soit d le degré de f et notons p0 , p1 , . . . , pr−1 ses points-base de multiplicité respective {mi }0≤i≤r−1 et q0 , q1 , . . . qr−1 ceux de f −1 de multiplicité {m′i }0≤i≤r−1 . L’action de f sur ℓ et sur les classes {epj }0≤j≤r−1 est donnée par : f# (ℓ) = dℓ − r−1 X m′i eqi , (1.4) i=0 f# (epj ) = mj ℓ − r−1 X ai,j eqi . (1.5) i=0 Remarquons que les coefficients ai,j correspondent au nombre d’intersection des transformées totales des diviseurs exceptionnels obtenus en éclatant respectivement les points pj et qi , dans la résolution de f . Par exemple, si j une application de Jonquières de degré d > 1. Notons p0 et q0 les points-base maximaux respectifs de j et j−1 et p1 , . . . , p2d−2 et q1 , . . . , q2d−2 les petits points-base respectifs de j et j−1 , nous avons :  2d−2 P  j# (ℓ) = dℓ − (d − 1)eq0 − eqi    i=1 2d−2 P . (1.6) eqi j# (ep0 ) = (d − 1)ℓ − (d − 2)eq0 −    i=1   j# (epi ) = ℓ − eq0 − eqi pour 1 ≤ i ≤ 2d − 2 Remarque 1.7. D’après [AC02, Proposition 2.2.21], les coefficients ai,j sont positifs pour tous 0 ≤ i, j ≤ r − 1. Remarque 1.8. L’action de f sur ℓ correspond au système linéaire associé à la transformée par f −1 d’une droite ne passant pas par les points-base de f . Si tous les points-base de f sont dans P2 alors pour chaque point pj il existe une courbe contractée par f −1 sur le point pj de degré mj et passant avec multiplicités ai,j aux points qj . PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 10 Ces informations sur f se lisent dans sa matrice caractéristique :   d m0 m1 ... mr−1  −m′0 −a0,0 −a0,1 ... −a0,r−1     −m′1 −a −a . . . −a1,r−1  1,0 1,1 .    .. .. .. .. ..   . . . . . ′ −mr1 −ar−1,0 −ar−1,1 . . . −ar−1,r−1 Celles pour f −1 se lisent en ligne en changeant le signe des mi et des m′i : −1 f# (ℓ) = dℓ − −1 f# (eqi ) = m′i ℓ r−1 X mj epj , j=0 r−1 X − ai,j epj . j=0 Pour plus de détails sur la matrice caractéristique, nous renvoyons à [AC02, Section 2.4]. La première ligne de la matrice caractéristique de f s’appelle la caractéristique de f . Elle est notée (d; m0 , . . . , mr−1 ). Il ne nous reste plus qu’à regarder l’action de f sur les classes ep où p ∈ B(P2 ) n’est pas un point-base de f . Remarque 1.9. Si f est un isomorphisme d’un voisinage U de p ∈ P2 sur un voisinage V de f (p) ∈ P2 alors f# (ep ) = ef (p) . Sinon, quitte à éclater des points-base de f et des points-base de f −1 , il existe un point q sur une surface dominant P2 tel que l’application induite par f envoie p sur q et est un isomorphisme local entre des voisinages de p et de q et donc f# (ep ) = eq . Considérons une classe c de l’espace de Picard-Manin Z : c = nℓ − r−1 X i=0 λi epi − X λp ep . 2 p∈B(P ) p∈supp(f / ) L’action de f sur une classe c de l’espace de Picard-Manin s’obtient par linéarité. ! ! r−1 r−1 r−1 X X X X λp f# (ep ) ai,j eqi − λj mj ℓ − m′i eqi − f# (c) = n dℓ −  = nd − r−1 X j=0 i=0 j=0 i=0  λj mj  ℓ − r−1 X i=0  nm′i − r−1 X j=0 p∈B(P2 ) p∈supp(f / )  λj ai,j  eqi − X p∈B(P2 ) p∈supp(f / ) λp f# (ep ). (1.10) PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 11 1.1.4. Espace hyperbolique de dimension infinie dans l’espace de Picard-Manin. À présent, considérons l’espace H∞ (S) = {c ∈ Z(S) | c · c = 1 et c · d0 > 0}, où d0 ∈ N1 (S) est une classe ample. Muni de la distance définie par d(c, c′ ) = argcosh(c · c′ ) pour tous c, c′ ∈ H∞ (S), c’est un espace hyperbolique de dimension infinie. Nous nous intéressons plus particulièrement à H∞ (P2 ) que nous notons H∞ . Tout élément de H∞ est de la forme X X λ2p = 1. λp ep où n > 0 et n2 − nℓ + p∈B(P2 ) p∈B(P2 ) Comme le groupe de Cremona agit sur l’espace de Picard-Manin et que l’action préserve la forme d’intersection, pour montrer que le groupe de Cremona agit sur H∞ , il suffit de montrer que f# (c) · ℓ > 0. Comme f#−1 (ℓ) ∈ H∞ et que le nombre d’intersection entre deux classes est supérieur ou égal à 1, nous avons comme attendu : f# (c) · ℓ = c · f#−1 (ℓ) ≥ 1. 1.2. Propriétés des applications du groupe de Cremona. Dans cette soussection, nous nous intéressons aux propriétés que vérifient les points-base et les multiplicités des applications du groupe de Cremona. Pour cela, nous introduisons dans un premier temps un vocabulaire général utile pour les points-base d’une application, mais pas seulement. Enfin, nous nous concentrons sur les applications de Jonquières. 1.2.1. Vocabulaire. Soient π : S ′ → S une composée d’éclatements, avec p l’un des points éclatés dans cette suite, et Ep la transformée stricte sur S ′ du diviseur exceptionnel obtenu en éclatant le point p. Tout point q de S ′ et appartenant à Ep est dit adhérent à p. Nous notons cette relation q → p. Plus généralement, tout point q de S ′ et appartenant à la transformée totale π ∗ (Ep ) de Ep , est dit voisin de p. Considérons un ensemble P de points de B(P2 ). Nous disons qu’il est préconsistant si tous ses points vivent sur une surface obtenue en éclatant uniquement des points de P à partir de P2 . Tout point de cet ensemble qui est adhérent à un seul point de P est dit libre contrairement à un point adhérent à deux points distincts de P qui est appelé satellite. Ces deux dernières définitions sont relatives à la donnée d’un ensemble pré-consistant de points. Remarquons qu’un point peut être adhérent à au plus deux points. Un point adhérent et libre est habituellement appelé voisin du premier ordre, mais nous n’utilisons pas cette terminologie. Nous disons qu’une suite pré-consistante de points forme une tour si chaque point de la suite est adhérent au point précédent. Exemple 1.11. Soit p0 un point de P2 . Éclatons-le. Notons p1 un point sur le diviseur exceptionnel Ep0 obtenu en éclatant le point p0 . Éclatons le point p1 . Considérons deux points distincts p2 et p3 sur le diviseur exceptionnel Ep1 tel que p2 appartienne également à la transformée stricte de Ep0 . PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA π0−1 (Ep0 ) Ep0 p0 • πp0 • 12 p1 πp1 • p2 p3 • Ep1 L’ensemble des points {p0 , p1 , p2 , p3 } est pré-consistant. Les points p1 , p2 et p3 sont voisins de p0 . Le point p1 est un point libre adhérent à p0 alors que le point p2 est satellite et adhérent aux points p0 et p1 . Le point p3 est libre non adhérent à p0 et il est adhérent à p1 . Pour plus de facilité, nous représentons ce genre de situation par un graphe où les sommets sont les points. Et il y a une flèche orientée entre deux sommets si celui du dessus est adhérent à celui du dessous. Ainsi la situation précédente se réécrit : p2 p3 p1 p0 Trois points ou plus sont dits alignés s’ils appartiennent à une droite de P2 ou à la transformée stricte d’une droite de P2 . Une suite pondérée est la donnée de points de B(P2 ) où chacun est muni d’une multiplicité réelle positive. Soit p un point d’une suite pondérée P . La différence entre la multiplicité mp , associée au point p, et la somme des multiplicités des points de P adhérents à p est appelée l’excès du point p : X mp − mq . q∈P q→p Une suite pondérée est dite consistante si l’excès en tout point de cet ensemble est positif. Remarquons qu’une suite de points est pré-consistante s’il est possible de pondérer tous les points de cet ensemble de façon strictement positive de sorte que si nous associons à tout autre point de B(P2 ) une multiplicité nulle alors l’excès en tous les points de B(P2 ) est positif ou nul. Exemple 1.12. Si nous reprenons l’exemple précédent 1.11, l’ensemble des points {p0 , p1 , p2 , p3 } est pré-consistant car ils peuvent être respectivement pondérés par 3, 2, 1, 1 et alors tous les points de B(P2 ) ont un excès positif ou nul. Par contre les points p0 , p2 , p3 ne forment pas un ensemble de points pré-consistant. En effet, quelque soit le choix de multiplicité associé aux points p2 et p3 l’excès au point p1 est toujours strictement négatif. 1.2.2. Propriété des points-base et de leur multiplicité. Cette partie regroupe plusieurs résultats de base qui seront utilisés par la suite. Une application du groupe de Cremona est dite de caractéristique Jonquières si sa caractéristique est la même que celle d’une application de Jonquières à savoir : (d; d − 1, 12d−2 ) où d est le degré de l’application. En fait ce sont des applications qui envoient un pinceau de droites sur un autre pinceau de droites. Les relations suivantes sont obtenues à partir de simples calculs d’intersection. La plupart d’entre elles sont connues et peuvent par exemple se trouver dans [AC02]. PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 13 Lemme 1.13. Soit f ∈ Bir(P2 ). Notons (d; m0 , . . . , mr−1 ) la caractéristique de f , (d; m′0 , . . . , m′r−1 ) celle de f −1 et (ai,j )0≤i,j≤r−1 les coefficients de la sous-matrice de la matrice caractéristique ne contenant pas la première ligne et la première colonne. Nous avons les relations : r−1 P mi = 3d − 3. 1) 2) i=0 r−1 P m2i = d2 − 1. i=0 r−1 P a2i,j = m2j + 1. i=0 r−1 P m′i ai,j = dmj . i=0 3) Pour tout 0 ≤ i ≤ r − 1, mi ≤ d − 1. 4) Si m0 = d − 1 alors f est une application de caractéristique Jonquières. r−1 P ai,j = 3mj − 1. 5) 6) 7) 8) i=0 r−1 P i=0 ai,j ai,k = mj mk . Remarque. Les équations 1) et 2) sont souvent appelées « équations de Noether ». Une façon d’énoncer la seconde égalité est de dire que le système linéaire associé à f est homaloïdal, c’est-à-dire que deux courbes de ce système ont un unique point d’intersection hors des points-base. Démonstration. Les égalités 1) et 5) s’obtiennent respectivement en intersectant la forme canonique contre f#−1 (ℓ) et ℓ puis contre f# (epj ) et epj et en utilisant la remarque 1.3. Auto-intersecter f#−1 (ℓ) et f# (epj ), puis intersecter f# (epj ) contre f# (epk ) et f# (epj ) contre f# (ℓ) nous donne respectivement 2), 6), 7) et 8). Le point 3) découle de 2). Étudions le point 4). Si m0 = d − 1 alors en utilisant les égalités 1) et 2) nous obtenons que r−1 X i=1 mi (mi − 1) = 0, par conséquent, les multiplicités sont toutes égales à 1. En utilisant à nouveau le point 1), nous obtenons que r = 2d − 1 ce qui permet de conclure que f est de caractéristique Jonquières.  La preuve du lemme suivant est une application directe des relations de Noether vues dans le lemme 1.13. Lemme 1.14. Soit f ∈ Bir(P2 ) de degré d et admettant r points-base. 1) Si d ≥ 4 alors r ≥ 6. 2) Si r ≤ 8 alors d ≤ 17. Démonstration. Soient {mi }0≤i≤r−1 les multiplicités des points-base de f . L’inégalité de Cauchy-Schwarz donne : X 2 r−1 r−1 X mi ≤r m2i . i=0 i=0 PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 14 Ainsi d’après les équations de Noether 1) et 2) du lemme 1.13, nous obtenons : 9(d − 1)2 ≤ r(d2 − 1), qui implique en simplifiant par le facteur d − 1 9(d − 1) ≤ r(d + 1). Comme la suite ud = 9(d−1) d+1 est croissante, si d ≥ 4 alors 5 < u4 ≤ ud ≤ r et f a donc au moins six points-base. Si r ≤ 8 alors 9(d − 1) ≤ 8d + 8 qui nous donne d ≤ 17 comme attendu.  Une application f ∈ Bir(P2 ) est dite symétrique si elle est de degré strictement supérieur à 1 et si toutes ses multiplicités sont égales. Lemme 1.15 ([AC02, Lemma 2.5.5]). La caractéristique d’une application symétrique f ∈ Bir(P2 ) est forcément de l’une des formes suivantes : (2; 13 ), (5; 26 ), (8; 37 ) ou (17; 68 ). En particulier, la multiplicité m des points-base de f est liée au degré d par la relation d+1 . m= 3 Remarque 1.16. Soit f ∈ Bir(P2 ) de degré d. La somme des trois plus grandes multiplicités de ses points-base est supérieure ou égale à d + 1. En fait, il y a égalité si et seulement si f est une application symétrique ou de caractéristique Jonquières. Cela s’obtient en utilisant uniquement les égalités 1) et 2) du lemme 1.13. Proposition 1.17 (Positivité des excès). Soit f ∈ Bir(P2 ) les excès des points-base de f sont tous positifs : X mq . pour tout p ∈ Bs(f ), mp ≥ q∈Bs(f ) q→p Démonstration. Nous avons vu que la multiplicité mp d’un point-base p de f correspond à la multiplicité au point p des courbes du système linéaire associé à f . Considérons une courbe C générale de ce système linéaire. D’après [Sha94, p.252], la multiplicité de C au point p, notée mp (C) est supérieure ou égale à la somme des multiplicités de cette courbe aux points adhérents à p. Par conséquent, nous avons l’égalité attendue : X X mp = mp (C) ≥ mq .  mq (C) ≥ q∈B(P2 ) q→p q∈Bs(f ) q→p Proposition 1.18 (Bézout). Soit f ∈ Bir(P2 ) de degré d dont les points-base sont les points {pi }0≤i≤r−1 de multiplicité respective {mi }0≤i≤r−1 . Pour toute courbe C ′ de degré d′ passant par les points pi avec multiplicité µi nous avons : dd′ − r−1 X i=0 mi µi ≥ 0. PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 15 Remarque 1.19. Une application de caractéristique Jonquières de degré d et de point-base de multiplicité maximale p0 , ne possède pas deux points-base adhérents à un même troisième point-base différent de p0 sinon cela contredirait la positivité des excès (Proposition 1.17) en ce point puisque sa caractéristique est (d; d − 1, 12d−2 ). De plus, par le même argument, il y a au plus d − 1 points adhérents au point de multiplicité maximale. Par la proposition 1.18, deux points de multiplicité 1 ne peuvent pas être alignés avec le point de multiplicité maximale. Un ensemble de points de B(P2 ) est dit en position presque générale s’il est pré-consistant et qu’aucune des situations suivantes n’est satisfaite : • quatre des points de cet ensemble sont alignés, • sept des points de cet ensemble sont sur une conique, • deux des points de cet ensemble sont adhérents à un troisième point. C’est en fait une terminologie utilisée par Dolgachev [Dol12, p.397]. Cependant, comme il s’intéresse aux surfaces del Pezzo, il impose en plus le fait qu’il y ait au plus 8 points. Une surface S est faiblement del Pezzo si son diviseur canonique est numériquement effectif et big, c’est à dire que pour toute courbe C de S le nombre d’intersection entre la courbe C et le diviseur canonique est positif ou nul, et que ce dernier est d’auto-intersection strictement positive : C · KS ≥ 0 et K2S > 0. Proposition 1.20 ([Dol12, Corollary 8.1.17]). Une surface rationnelle est faiblement del Pezzo si et seulement si elle a été obtenue en éclatant k ≤ 8 points de P2 en position presque générale. Lemme 1.21. Soit f ∈ Bir(P2 ) une application ayant au plus 8 points-base. Les points-base de f sont en position presque générale si et seulement si les points-base de f −1 le sont. Démonstration. D’après la proposition 1.20, une surface est faiblement del Pezzo si et seulement si la suite de points éclatés satisfait les conditions du lemme. La surface obtenue en éclatant les points-base de f et de f −1 étant la même nous obtenons le résultat.  1.2.3. Générateurs du groupe de Cremona. Soient f ∈ Bir(P2 ) et p0 un de ses points de multiplicité maximale m0 . En suivant la terminologie de [AC02], nous appelons complexité de f le nombre d − m0 cf = . 2 Un point-base p de f différent de p0 est dit majeur si sa multiplicité mp est strictement supérieure à la complexité de f : d − m0 . 2 Remarquons que si f possède plusieurs points de multiplicité maximale, si nous changeons le choix du point maximal, la complexité de f reste identique ainsi que le nombre de points-base majeurs. Nous notons Maj(f ) cet ensemble et h son cardinal. L’involution standard de Cremona, qui s’écrit dans une carte locale : mp > 1 1 σ : (x, y) 99K ( , ) x y PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 16 est un exemple d’application qui appartient au groupe de Jonquières. Elle a une importance particulière dans le groupe de Cremona, puisque comme le corps de base est algébriquement clos elle engendre avec le groupe des automorphismes de P2 tout le groupe. Le théorème qui suit n’est plus valable si le corps de base n’est pas algébriquement clos. Théorème 1.22 (M. Noether et G. Castelnuovo). Le groupe de Cremona Bir(P2 ) est engendré par σ et PGL(3, k). M. Noether fut le premier a énoncer ce théorème à la fin du 19e siècle. Cependant, la première preuve exacte est due à G. Castelnuovo. L’idée de « la preuve » de M. Noether est la suivante. Considérons une application f . En pré-composant par une application quadratique dont les trois points-base sont trois points-base de f de plus grande multiplicité, le degré de f diminue. En réitérant ce procédé le degré devient 1 et la composée est une application linéaire. Cependant il n’existe pas toujours une telle application quadratique. En effet, une application quadratique ne peut pas avoir deux de ses points-base adhérents au troisième (Remarque 1.19). Or il est possible que les points-base de multiplicité maximale d’une application soit dans cette configuration. C’est le cas notamment des automorphisme polynomiaux, comme par exemple (x, y) 7→ (y 3 − x, y). Ce problème n’avait pas été vu par M. Noether car le formalisme des points adhérents n’existait pas. Dans [Ale16], J. Alexander corrige la preuve de M. Noether en introduisant la complexité d’une application. S’il existe une application quadratique dont les points-base sont le point-base maximal de f et deux points-base majeurs alors la complexité diminue strictement. Sinon, il faut utiliser un troisième point qui vit dans P2 . Dans ce cas, la complexité reste identique mais le nombre de points-base majeurs diminue. La récurrence se fait sur ces deux entiers positifs. Une conséquence de ce théorème est que la réunion du groupe de Jonquières et du groupe des automorphismes de P2 est également un système de générateurs du groupe de Cremona. Ce résultat est plus faible mais parfois plus maniable. Il a également l’avantage de pouvoir décomposer une application f ∈ Bir(P2 ) en produit de transformations de Jonquières de sorte que le degré augmente à chaque pré-composition. Théorème 1.23 ([AC02, Theorem 8.3.4]). Toute application de Cremona f est composée d’applications de Jonquières ji et d’un élément a ∈ PGL(3, k) : f = a ◦ jn ◦ · · · ◦ j1 , de sorte qu’à chaque pré-composition par une application de Jonquières le degré augmente strictement, pour tout 1 ≤ i ≤ n − 1 : deg(a ◦ jn ◦ · · · ◦ ji+1 ◦ ji ) > deg(a ◦ jn · · · ◦ ji+1 ). Dans [AC02, Theorem 8.3.4], M. Alberich-Carramiñana n’énonce pas la seconde partie du théorème, qu’elle démontre pourtant. Il est facile de voir qu’en précomposant f par une application de Jonquières dont l’inverse a pour point-base maximal le point-base de f de multiplicité maximale, et ses petits points-base sont parmi les points-base majeurs de f , alors le degré diminue strictement. La partie difficile est de montrer qu’il existe une telle application de Jonquières j. Elle montre que c’est le cas si les petits points-base de j−1 sont choisis comme étant tous les pointsbase majeurs de f , à l’exception d’un lorsque le cardinal des points-base majeurs de f est impair. L’algorithme qu’elle utilise et qui est dû à G. Castelnuovo consiste à PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 17 considérer une telle application de Jonquières j1 . Puis elle considère f ◦ j1−1 et réitère le procédé jusqu’à obtenir un automorphisme. 1.2.4. Support d’une transformation de Jonquières. Dans cette partie, nous déterminons dans un cas très spécial, celui où il existe au moins un point adhérent à un autre point, à quelles conditions un ensemble pré-consistant de points de B(P2 ) est le support d’une application de Jonquières. La preuve proposée ici est plus simple que celle présentée dans [AC02], mais elle ne traite que d’un cas particulier. Le problème dans le cas général est de faire attention au fait que les points-base ne doivent pas se trouver sur une courbe de sorte qu’en éclatant les points et en contractant les fibres, la courbe deviennent d’auto-intersection strictement plus petite que −1. Par exemple, trois points alignés ne forment pas le support d’une quadratique. Lemme 1.24. Soient f ∈ Bir(P2 ) et p un de ses points-base de multiplicité maximale. Parmi les points-base majeurs de f , le nombre de points-base adhérents à p est inférieur ou égal au nombre de points-base non adhérents à p. Démonstration. Nous reprenons lors des deux premiers points des calculs connus et faits par exemple dans la Section 8.2 de [AC02, Lemma 8.2.3 et 8.2.6]. Notons r le nombre de points-base de f et h le nombre de points-base majeurs de f . • D’après le théorème de Bézout (Proposition 1.18), nous avons d ≥ mi + m0 pour tout 1 ≤ i ≤ r − 1, ce qui implique : 2 cf ≥ mi . • Multiplions par cf l’équation 1) du lemme 1.13 et soustrayons cela à l’équation 2) du même lemme : r−1 X i=0 mi (mi − cf ) = d2 − 1 − 3d cf +3 cf = d(d − 3 cf ) + 3 cf −1 = (m0 + 2 cf )(m0 − cf ) + 3 cf −1, ce qui implique : r−1 X i=1 mi (mi − cf ) = 2 cf (m0 − cf ) + 3 cf −1. Comme pour i > h les points ne sont pas majeurs nous avons h X i=1 mi (mi − cf ) > 2 cf (m0 − cf ), ce qui donne finalement par le premier point h X i=1 (mi − cf ) > m0 − cf . • Notons ha le nombre de points-base de f majeurs et adhérents à p0 . Quitte à réordonner les indices, l’équation précédente se ré-écrit : ha X i=1 mi + h X i=ha +1 mi > m0 + (h − 1) cf . PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 18 Par positivité des excès en p0 pour f (Proposition 1.17) et par le premier point de la preuve : m0 + (h − ha )2 cf > m0 + (h − 1) cf ce qui entraîne : h 2 2(h − ha ) ≥ h. Par conséquent ≥ ha ce qui signifie qu’il y a au plus autant de points-base majeurs adhérents à p0 que de points-base majeurs non adhérents.  Lemme 1.25. Considérons un ensemble pré-consistant de points {pi }0≤i≤2δ de B(P2 ) satisfaisant a) exactement δ points sont adhérents à p0 , b) pour toute paire (i, j) où 1 ≤ i < j ≤ 2δ, les points pi , pj et p0 ne sont pas alignés, c) si deux points de cet ensemble sont adhérents à un même troisième point pi de cet ensemble alors i = 0. Alors il existe une application de Jonquières de degré δ + 1 qui possède cet ensemble comme points-base et dont p0 est le point-base majeur. Démonstration. Lorsque nous éclatons le point p0 , il n’y a pas deux points sur une même fibre d’après la condition b). Nous éclatons ensuite les δ points adhérents, et nous contractons les transformées strictes des fibres correspondantes. Par la condition c), il n’y a toujours pas deux points sur une même fibre. La surface obtenue est une surface de Hirzebruch Fδ+1 de section exceptionnelle d’auto-intersection −(δ + 1). Ensuite, nous éclatons un point non adhérent puis contractons la fibre passant par ce point. En faisant cela, nous obtenons la surface de Hirzeburch dont la section exceptionnelle est d’auto-intersection un de plus par rapport à la surface précédente et il n’y a toujours pas deux points sur une même fibre. Ainsi en répétant cette opération pour les δ points non adhérents nous obtenons une surface de Hirzeburch F1 et en contractant la section exceptionnelle nous obtenons une application de Jonquières de degré δ + 1.  Proposition 1.26. Soit f ∈ Bir(P2 ) possédant un point-base majeur adhérent à son (ou à un des ses) point-base maximal p0 . Pour tout sous-ensemble de l’ensemble des points-base majeurs de f , pré-consistant, de cardinal pair 2δ et possédant un nombre de points adhérents à p0 égal à δ, il existe une application de Jonquières de degré δ + 1 ayant p0 comme point-base maximal et cet ensemble de points comme autres points-base. De plus, un tel sous-ensemble existe toujours. Démonstration. Nous cherchons à utiliser le lemme 1.25. Pour cela, plusieurs points sont à vérifier : • L’hypothèse sur f et le lemme 1.24 justifient le fait qu’il existe un sousensemble de l’ensemble des points-base majeurs de f qui possède autant de points adhérents à p0 que de points non adhérents. • D’après le théorème de Bézout 1.18, il n’existe pas deux points majeurs alignés avec le point p0 . • Il n’y a pas deux points majeurs adhérents à un point majeur. Sinon, par la proposition 1.17, il existerait un point majeur p1 de multiplicité m1 > d − m0 . Mais dans ce cas, il y aurait un point q libre et adhérent à p0 (si p1 est voisin de p0 ) ou dans P2 qui serait de multiplicité strictement supérieure à d − m0 . Ceci contredit PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 19 le théorème de Bézout (proposition 1.18) en considérant la droite passant par les points p0 et q. Ainsi, d’après le lemme 1.25, il existe une application de Jonquières de degré δ + 1 ayant p0 comme point-base de multiplicité maximal et possédant cet ensemble comme petits points-base.  Remarque 1.27. Sans la condition qu’il y ait autant de points adhérents à p0 que de points non adhérents ce n’est pas possible de conclure. En effet, considérons les points {p0 , . . . , p6 } de P2 tels qu’il n’y ait pas deux points alignés avec p0 , que les points p1 , . . . , p4 soient alignés et que les points p5 et p6 n’appartiennent pas à la droite contenant les points p1 , . . . , p4 . • p0 • p6 • p1 • p5 • p2 • p3 • p4 Ils constituent le support d’une application de Jonquières de degré 4 et de pointbase maximal p0 . Mais si nous considérons seulement les points p0 , p1 , p2 , p3 , p4 ils ne forment plus le support d’une application de Jonquières. En effet, si nous éclatons les points p0 et p1 et que nous contractons la droite passant pas ces deux points, nous nous retrouvons sur P1 × P1 . En éclatant ensuite le point p2 , nous nous retrouvons sur F1 mais les deux points restants sont sur la section exceptionnelle. 2. Construction d’un pavage de Voronoï Considérons l’action du groupe de Cremona sur H∞ . L’objet de cette section est de construire un pavage de Voronoï associé à cette action. Dans la sous-section 2.1, nous nous restreignons à un sous-espace convexe E de H∞ contenant l’enveloppe convexe de l’orbite de la droite ℓ. Nous ne nous restreignons pas seulement à l’enveloppe convexe de l’orbite de ℓ car il est parfois difficile de vérifier qu’un élément y appartient. Pour cette raison, nous élargissons l’espace à étudier. Nous construisons alors un pavage de Voronoï de la façon suivante. Nous considérons l’orbite de la classe de la droite ℓ. Cela nous donne un ensemble discret de points de E. Cependant le stabilisateur de ℓ est PGL(3, k). Nous identifions donc deux applications qui diffèrent par un automorphisme : f ∼ g ⇔ il existe a ∈ PGL(3, k), f = g ◦ a. Nous notons f une telle classe. Par construction, toute application d’une même classe d’équivalence agit de la même façon sur ℓ. Pour toute classe f où f ∈ Bir(P2 ), nous associons une cellule de Voronoï notée V(f ) et définie comme : V(f ) = {c ∈ E | d(c, f# (ℓ)) ≤ d(c, g# (ℓ)) pour tout g ∈ Bir(P2 )}. Remarquons que toutes les applications du groupe de Cremona appartenant à la même classe d’équivalence indexent la même cellule de Voronoï. Les applications f ′ ∈ f sont appelées les germes associés à V(f ) et les classes f# (ℓ) sont les centres des cellules de Voronoï. Une cellule de Voronoï ne correspond pas tout-à-fait à un domaine fondamental mais c’est l’orbite d’un domaine fondamental sous l’action de PGL(3, k). PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 20 Nous définissons dans une première sous-section l’espace sur lequel nous faisons agir le groupe de Cremona, puis nous montrons que les cellules de Voronoï recouvrent l’espace construit et enfin nous montrons que les cellules de Voronoï ne s’accumulent pas sur une cellule. 2.1. Restriction à un sous-espace. Considérons l’espace hyperbolique H∞ de dimension infinie obtenu comme hyperboloïde dans l’espace de Picard-Manin qui a été construit à la sous-section 1.1.4. Nous considérons le sous-espace suivant. Définition 2.1. L’ensemble E est le sous-espace de H∞ constitué des classes X λp ep (n réel ≥ 1) c = nℓ − p∈B(P2 ) satisfaisant : 1) λp ≥ 0 pour tout p ∈ B(P2 ), 2) la positivité contre la classe anti-canonique : X λp ≥ 0, 3n − p∈B(P2 ) 3) la positivité des excès de tout point p ∈ B(P2 ) : X λp − λq ≥ 0, q∈B(P2 ) q→p 4) la condition de Bézout : pour toute courbe de P2 de degré d passant avec multiplicité µp en chaque point p ∈ B(P2 ) : X λp µp ≥ 0. nd − p∈B(P2 ) L’inégalité de la propriété 2) implique que les classes de E sont en fait L1 . Par conséquent, la forme canonique est bien définie sur les classes de E et la propriété 2) revient à demander que les classes soient positives contre la forme anti-canonique qui est l’opposé de la forme canonique : − kP2 ·c ≥ 0. Lorsqu’une classe satisfait cette inégalité, nous disons qu’elle est positive contre la classe anti-canonique. Le coefficient n est appelé le degré de c et les coefficients λp les multiplicités de c aux points p ∈ B(P2 ). Les points pour lesquels la multiplicité associée est strictement positive constituent le support de c. Cet ensemble est noté supp(c). Remarquons que par définition de H∞ celui-ci est dénombrable. Grâce à la propriété 3), le support de c est un ensemble pré-consistant de points, au sens de la définition introduite dans la section 1.2.1. Remarque 2.2. Par définition, une classe c ∈ E appartient à H∞ et possède donc un degré supérieur ou égal à 1 et strictement supérieur à chacune de ses multiplicités : X λ2p ≤ n2 − λ2q pour tout q ∈ B(P2 ). 1 = c2 = n 2 − p∈B(P2 ) PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 21 Notons Z >0 le sous-espace de l’espace de Picard-Manin constitué des classes d’auto-intersection strictement positive. Une classe de Z >0 satisfaisant les points 1), 2), 3) et 4) de la définition 2.1 est dite proportionnelle à une classe de E. Une telle classe appartient à la demi-droite ouverte issue de la classe nulle et passant par une classe de E. Plus généralement, étant donné un sous-ensemble F de l’espace de Picard-Manin, une classe d’auto-intersection strictement positive est dite proportionnelle à une classe de F si elle appartient à la demi-droite ouverte issue de la classe nulle et passant par une classe de F . L’application de normalisation : η : Z >0 → H∞ (2.3) x x 7→ p . B(x, x) envoie toute classe de Z >0 sur sa classe proportionnelle dans H∞ . Proposition 2.4. L’action du groupe de Cremona stabilise E. Démonstration. Le groupe de Cremona étant engendré par PGL(3, k) et l’involution standard σ (Théorème 1.22) il suffit de vérifier que l’action de ces applications préserve E. C’est le cas pour les automorphismes. Soit X λp ep ∈ E. c = nℓ − p∈B(P2 ) Montrons que σ# (c) appartient à E. Notons p0 , p1 et p2 les points-base de σ (qui sont en fait [0 : 0 : 1], [0 : 1 : 0] et [1 : 0 : 0]) et q0 , q1 et q2 ceux de σ −1 . Bien que σ soit une involution nous les différencions pour plus de clarté. D’après l’équation (1.10), nous avons : σ# (c) = (2n − λp0 − λp1 − λp2 )ℓ − (n − λp1 − λp2 )eq0 − (n − λp0 − λp2 )eq1 X λp σ# (ep ). − (n − λp0 − λp1 )eq2 − p∈B(P2 ) p∈{p / 0 ,p1 ,p2 } Le groupe de Cremona agit sur H∞ (voir le paragraphe 1.1.4) par conséquent la classe σ# (c) appartient à H∞ et possède donc un degré strictement positif. Intéressons-nous aux autres points de la définition. 1) Comme c ∈ E les multiplicités λp sont positives ou nulles. Montrons que c’est le cas des multiplicités de σ# (ℓ) pour les points q0 , q1 et q2 . Les points p0 et p1 sont dans P2 il existe donc une droite L passant par ces deux points et comme c satisfait la condition de Bézout (Définition 2.1.4)), nous avons : n − λp0 − λp1 ≥ 0. Les deux autres multiplicités sont positives avec le même argument. 2) La condition de positivité contre l’anti-canonique est satisfaite puisque la forme canonique est constante sur l’orbite de c sous l’action du groupe de Cremona (Remarque 1.3). 3) Montrons que les excès de σ# (c) sont positifs en tout point q ∈ B(P2 ). Soit q ∈ B(P2 ). Deux cas se présentent selon si q est un point-base de σ −1 ou pas. Dans le cas où q n’est pas un point-base de σ −1 , il existe p ∈ B(P2 ) tel que −1 ep = σ# (eq ). Quitte à éclater au départ et à l’arrivée, σ −1 induit un isomorphisme local envoyant q sur p et par conséquent les points adhérents à q sont envoyés sur des points adhérents à p et donc la positivité des excès est préservée. PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 22 Il nous reste à considérer les excès au dessus des points q0 , q1 et q2 . Notons S la surface obtenue en éclatant les points p0 , p1 et p2 . Les points adhérents à q0 sont les points appartenant à la transformée stricte de la droite L passant par les points p1 et p2 sur S. Ainsi la droite L passe par les points p1 et p2 ainsi que par les points adhérents à q0 . Par conséquent en considérant cette droite et le fait que c satisfait la condition de Bézout (Définition 2.1.4)), nous avons : X X λp ) ≥ 0. λp = n − (λ1 + λ2 + (n − λ1 − λ2 ) − 2.1.4) p7→q0 p7→q0 Ainsi, l’excès en q0 est positif. Il en est de même pour les points q1 et q2 . 4) Montrons que σ# (c) satisfait la condition de Bézout pour toute courbe de P2 . Considérons dans un premier temps une courbe C1 qui n’est contractée par σ −1 . −1 Notons c1 sa classe dans Picard-Manin. Dans ce cas, la classe σ# (c1 ) est la classe de Picard-Manin correspondant à la courbe de P2 obtenue comme adhérence de σ −1 (C1 \ {q0 , q1 , q2 }). Par conséquent son nombre d’intersection avec c est positif ou nul puisque c satisfait la condition de Bézout : −1 σ# (c) · c1 = c · σ# (c1 ) ≥ 0. 2.1.4) Considérons à présent les droites contractées par σ −1 comme par exemple la droite passant par les points q1 et q2 . Alors nous obtenons : 2n − λp0 − λp1 − λp2 − (n − λp0 − λp2 ) − (n − λp0 − λp1 ) = λp0 ≥ 0. Il en est de même des deux autres droites contractées. Par conséquent σ# (c) satisfait également la condition de Bézout (Définition 2.1.4)). Nous avons ainsi montré que l’action de σ préserve E et par conséquent que l’action du groupe de Cremona stabilise E.  Proposition 2.5. L’espace E est un sous-espace convexe fermé de H∞ . Démonstration. Les conditions 1) et 4) définissent des demi-espaces fermés de l’espace de Picard-Manin. Comme les multiplicités des classes de E sont positives d’après la condition 1), la condition 2) se réécrit de la façon suivante : X pour toute partie finie F de B(P2 ), λp ≤ 3n. p∈F Par conséquent la condition 2) correspond à une intersection infinie de demi-espaces fermés qui est un fermé. Nous raisonnons pareil pour la condition 3). En intersectant ce fermé avec l’hyperboloïde H∞ nous obtenons que E est fermé comme annoncé.  Proposition 2.6. L’enveloppe convexe de l’orbite de ℓ sous l’action du groupe de Cremona est incluse dans E. Démonstration. Comme par la proposition 2.5 l’espace E est convexe, il suffit de montrer que toutes les classes de l’orbite de ℓ sous l’action du groupe de Cremona sont dans E. De plus comme d’après la proposition 2.4 l’action du groupe de Cremona préserve E, il suffit de vérifier que la classe de la droite ℓ appartient à E. La classe ℓ appartient à H∞ et vérifie le point 1). La classe de la droite contre la classe anti-canonique vaut 3 ainsi la positivité contre l’anti-canonique est vérifiée (point 2)). Les points 3) et 4) sont également vérifiés.  PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 23 Remarque 2.7. Soit p ∈ P2 . Il existe t0 ∈ [0, 1[ tel que pour tout t0 ≤ t < 1 la classe √ 1 (ℓ − tep ) n’est pas dans l’enveloppe convexe de l’orbite de ℓ puisque l’inter1−t2 section de cette classe contre la classe anti-canonique est strictement supérieure à 3. 2.2. Pavage. Montrons que les cellules de Voronoï recouvrent l’espace E (Corollaire 2.11). Proposition 2.8. Pour toute classe c appartenant à E, les deux infimums suivants sont atteints :  1) inf  d(c, f# (ℓ)) | f ∈ Bir(P2 ) 2) inf d(c, f# (ℓ)) | f ∈ Bir(P2 ) et c ∈ / V(f ) . P Démonstration. Soit c = nℓ − λp ep une classe de E. Montrons que l’infimum suivant est atteint : p∈B(P2 )  inf d(c, f# (ℓ)) | f ∈ Bir(P2 ) . (2.9) Considérons les classes de l’orbite de la droite ℓ se trouvant dans la boule fermée de E centrée en c et de rayon argcosh(n). La classe ℓ est dans cette boule. De plus, par l’inégalité triangulaire, toute application f telle que la classe f# (ℓ) se situe dans cette boule est de degré au plus cosh(2 argcosh(n)) = 2n2 − 1. Notons B l’ensemble de ces applications. Il n’y a qu’un nombre fini de caractéristiques pour des applications du groupe de Cremona de degré au plus 2n2 − 1. Remarquons que si le support de c est fini alors l’infimum est atteint. Nous considérons donc le cas où c est de support infini. Notons C1 , . . . , Ck cet ensemble de caractéristiques. Pour 1 ≤ j ≤ k, posons Cj = (dj ; mj0 , . . . , mjrj ). Notons Bj l’ensemble des applications de B ayant Cj pour caractéristique. Montrons que pour tout 1 ≤ j ≤ k  (2.10) inf d(c, f# (ℓ)) | f ∈ Bj est atteint. Fixons 1 ≤ j ≤ k. Raisonnons par l’absurde et supposons qu’il existe une suite (fi )i∈N d’applications de Bj telle que la suite (d(c, fi # (ℓ)))i∈N soit strictement décroissante. Pour chaque i, nous regardons le point base q de fi−1 de multiplicité mj0 . Nous notons λ0,i la multiplicité λq de c associée au point q. À noter que si q n’est pas un point du support de c, la multiplicité λ0,i est nulle. Nous faisons de même pour chaque 0 ≤ s ≤ rj . Nous obtenons ainsi rj + 1 suites {λs,i }i∈N de coefficients. Si nous l’ordonnons, la suite des multiplicités de c est décroissante et tend vers 0. Par conséquent quitte à prendre des sous-suites nous pouvons supposer que chacune des rj +1 suites est ou bien constante ou bien strictement décroissante. Or la suite (d(c, fi # (ℓ)))i∈N étant strictement décroissante cela implique que la suite rj P ms λs,i est strictement croissante ce qui est absurde. Par conséquent, pour Si = s=0 tout 1 ≤ j ≤ k il existe une application f j ∈ Bj pour laquelle l’infimum (2.10) est atteint :  d(c, f#j (ℓ)) = inf d(c, f# (ℓ)) | f ∈ Bj . Notons j0 ∈ {1, . . . , k} un indice tel que d(c, f#j0 (ℓ)) = min d(c, f#j (ℓ)). 1≤j≤k L’infimum (2.9) est atteint pour f . Le point 2) se démontre de la même façon. j0  PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 24 Corollaire 2.11. Les cellules de Voronoï pavent l’espace E. Démonstration. Soit c ∈ E. Le point 1) de la proposition 2.8 implique qu’il existe f ∈ Bir(P2 ) telle que c appartient à V(f ).  2.3. Non-accumulation des cellules. Dans cette partie, nous montrons que les cellules de Voronoï ne s’accumulent pas sur cellule : pour tout f ∈ Bir(P2 ), pour toute classe c ∈ V(f ), il existe ε > 0 tel que les cellules de Voronoï ne contenant pas c mais tel que c soit dans leur ε-voisinage sont en nombre fini. La condition « ne contenant pas c » est nécessaire car nous verrons que dans certains cas c appartient à un nombre infini de cellules. Proposition 2.12. Pour tout c ∈ E, il existe ε > 0 tel que pour tout f ∈ Bir(P2 ) soit c appartient à V(f ), soit la distance entre c et V(f ) est strictement supérieure àε: d(c, V(f )) > ε. La notation « d(c, V(f )) » est un abus de notation et signifie l’infimum sur toutes les classes c′ de V(f ) des distances entre c et c′ : d(c, V(f )) = inf{d(c, c′ ) | c′ ∈ V(f )}. Démonstration. Quitte à faire agir le groupe de Cremona sur le pavage de Voronoï, nous pouvons supposer que c appartient à la cellule V(id). Si la classe c est à l’intérieur (pour le pavage de Voronoï) de V(id), tout ε strictement positif tel que la boule fermée centrée en c et de rayon ε est incluse dans l’intérieur (pour le pavage) de V(id) convient. Intéressons-nous à présent aux classes P λp ep ∈ E une classe apqui sont au bord de la cellule V(id). Soit c = nℓ − p∈B(P2 ) partenant au bord de la cellule V(id). D’après le point 2) de la proposition 2.8, l’infimum suivant est atteint Posons D := inf{d(c, f# (ℓ)) | f ∈ Bir(P2 ) et c ∈ / V(f )}. D − argcosh(n) . 6 Comme c appartient à V(id), ε est strictement positif. Montrons que pour toute application f ∈ Bir(P2 ) telle que c n’appartient pas à V(f ), nous avons : ε= d(c, V(f )) > ε. Raisonnons par l’absurde et supposons qu’il existe f ∈ Bir(P2 ) telle que c n’appartient pas à V(f ) et qu’il existe c′ ∈ V(f ) tel que d(c, c′ ) ≤ ε. Nous avons alors par définition de D et par l’inégalité triangulaire la contradiction suivante : D ≤ d(c, f# (ℓ)) ≤ d(c, c′ ) + d(c′ , f# (ℓ)) ≤ ε + d(c′ , ℓ) ≤ ε + d(c′ , c) + d(c, ℓ) ≤ 2ε + argcosh(n) D 2 argcosh(n) + 3 3 D 2D + = D. < 3 3 ≤  PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 25 Nous disons qu’un segment de E traverse une cellule de Voronoï s’il existe un sous-segment contenu dans l’intérieur (pour le pavage) de cette cellule. Le corollaire suivant dit qu’un segment ne traverse qu’un nombre fini de cellules de Voronoï. Plus généralement, il donne un analogue de ce résultat lorsque le segment reste dans le bord de Voronoï de plusieurs cellules. Corollaire 2.13. Soit [c, c′ ] un segment géodésique de E. Toute suite (ci )i∈I de classes du segment [c, c′ ] satisfaisant les trois points suivants est finie : • c0 = c, • ci < ci+1 et • pour tout i ≥ 1 il existe fi ∈ Bir(P2 ) satisfaisant ci ∈ V(fi ) et ci−1 ∈ / V(fi ) En particulier, un segment géodésique ne traverse qu’un nombre fini de cellules de Voronoï. Démonstration. Paramétrons le segment [c, c′ ] par γ(t) pour t ∈ [0, 1] avec γ(0) = c et γ(1) = c′ . Nous construisons une suite de points ci = (γ(ti )) par le procédé de récurrence suivant. Initialisons en posant t0 = 0 et c = γ(0). Pour i ≥ 1, si ti−1 ∈ [0, 1[ est déjà construit, nous posons ti = inf{t ∈]ti−1 , 1] | ∃f ∈ Bir(P2 ), γ(t) ∈ V(f ) et γ(ti−1 ) ∈ / V(f )}. Remarquons que les classes de cette suite appartiennent à plusieurs cellules par construction. De plus, pour tout 1 ≤ i et pour toute application f telle que V(f ) contient la classe ci mais pas la classe ci−1 , la cellule V(f ) ne contient pas les classes cj pour j ≤ i − 1 par convexité des cellules de Voronoï. Pour tout 0 ≤ t ≤ 1, d’après la proposition 2.12 et le paragraphe précédent, il existe εt > 0 tel que dans l’intervalle ouvert centré en γ(t) et de diamètre 2ε il existe au plus une classe de la suite {ci }i∈I . Ces ouverts forment un recouvrement du segment géodésique compact [c, c′ ], par conséquent il existe un sous-recouvrement fini. Or chacun de ces ouverts contient au plus un point de la suite. Il y a donc un nombre fini d’éléments dans cette suite.  2.4. Bord à l’infini. Le bord de H∞ est constitué des classes d’auto-intersection nulle. Ces classes peuvent être vues comme limites de classes de Picard-Manin vivant dans l’hyperplan affine d’équation {ℓ = 1} et proportionnelles à des classes de H∞ . Nous définissons le bord à l’infini de E, noté ∂∞ E comme l’ensemble des classes du bord de H∞ qui vérifient les points 1) à 4) de la définition 2.1 de E. De même que précédemment ces classes sont limites de classes de Picard-Manin vivant dans l’hyperplan affine d’équation {ℓ = 1} et proportionnelles à des classes de E. 2.5. Classes symétriques. Une classe dont les r plus grandes multiplicités sont égales et la (r + 1)-ème est strictement plus petite est appelée r-symétrique. Si de plus, c ne possède pas d’autre multiplicité, la classe est dite pure. Lemme 2.14. Soit c une classe r-symétrique avec la plus grande multiplicité qui est égale au tiers du degré de c : r−1 c = nℓ − nX ep − 3 i=0 i X λp ep , 2 p∈B(P ) p∈{p / 0 ,...,pr−1 } / {p0 , . . . , pr−1 }. Nous avons alors : où λp < n3 pour tout p ∈ 1) Si c ∈ E alors r ≤ 8. PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 26 2) Si c ∈ ∂∞ E alors r = 9 et c est une classe 9-symétrique pure. Démonstration. Considérons une telle classe c. Dans le cas où c ∈ E, si r ≥ 9 alors nous obtenons la contradiction suivante : X n2 λ2p ≤ 0. − 1 = c2 ≤ n 2 − 9 9 p∈{p / 0 ,...,pr−1 } Si maintenant c ∈ ∂∞ E alors c2 = 0 et par le même argument, nous obtenons r ≤ 9. Dans le cas où r = 9, notons que la classe c est forcément symétrique pure. Supposons que r ≤ 8. Alors il existe p ∈ / {p0 , . . . , pr−1 } tel que λp > 0 puisque : n2 − r n2 − 9 X λ2p = 0. (2.15) p∈{p / 0 ,...,pr−1 } Comme c appartient au bord à l’infini, nous avons par définition : X n λp . 0 ≤ − kP2 ·c = 3n − r − 3 (2.16) p∈{p / 0 ,...,pr−1 } 2 En multipliant l’égalité 2.15 par 3, nous avons 3n2 = 3r n9 + 3 P p∈{p / 0 ,...,pr−1 } λ2p . En multipliant l’équation 2.16 par n qui est positif et en remplaçant dans cette équation 3n2 par la valeur que nous venons d’obtenir, nous avons X X X n2 n2 λp (3λp − n), λ2p − r − n λp = 0 ≤ 3r + 3 9 3 p∈{p / 0 ,...,pr−1 } ce qui contredit le fait que λp < attendu. p∈{p / 0 ,...,pr−1 } n 3 p∈{p / 0 ,...,pr−1 } pour tout p ∈ / {p0 , . . . , pr−1 }. Ainsi r = 9 comme  Remarquons que ce lemme est l’analogue du lemme 1.15 sur les caractéristiques des applications symétriques. Cependant, comme les multiplicités des classes ne sont pas entières cela enlève de la rigidité, d’où l’hypothèse imposée sur les coefficients. 3. Cellule de Voronoï associée à l’identité Le groupe de Cremona agissant transitivement sur les cellules de Voronoï, il suffit d’étudier la cellule V(id) pour comprendre toutes les cellules. Dans cette section, nous caractérisons les classes se trouvant dans la cellule V(id) (Proposition 3.6 et Théorème 3.8). Par définition, une classe c est dans la cellule de Voronoï associée à l’identité si elle ne peut pas être rapprochée de ℓ en faisant agir un élément de Bir(P2 ) : c ∈ V(id) ⇔ pour tout f ∈ Bir(P2 ), f# (c) · ℓ ≥ c · ℓ. Avant cela, introduisons des notations utilisées dans cette section et par la suite. Nous avons vu que le support d’une classe c ∈ E (ou plus généralement dans H∞ ) est dénombrable. Par facilité d’écriture, étant donnée une classe c ∈ E, nous renumérotons toujours dans cette section les points de son support et les multiplicités correspondantes, de sorte qu’ils soient indicés par un sous-ensemble I ′ de N qui est soit un intervalle commençant à 0 soit N tout entier : X c = nℓ − λi epi . i∈I ′ PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 27 Remarquons que pour alléger les notations, nous écrivons λi au lieu de λpi . Par la suite, lorsque nous ferons agir une application f ∈ Bir(P2 ) sur c, nous aurons également besoin d’une notation pour les points-base de f . S’ils sont dans le support de c, la notation des points-base de f sera induite par celle sur les points du support de c. Si ce n’est pas le cas, nous aurons besoin d’introduire un ensemble fini J disjoint de N, tel que pour j ∈ J, λj est la multiplicité nulle de c correspondant au point-base pj de f qui n’est pas dans le support de c. De fait, pour tout i ∈ I ′ et pour tout j ∈ J, λi > λj . Ainsi, pour alléger les notation, I correspondra suivant le contexte à I ′ ou à I ′ ∪J et les indices de J seront considérés comme plus grands que ceux de I ′ . Une classe est dite ordonnée lorsque la numérotation induit un ordre décroissant sur les multiplicités de c : pour tout i, j ∈ I tels que i < j, λi ≥ λj . Considérons l’action d’une application f ∈ Bir(P2 ) de degré d sur une classe P c = nℓ − i∈I λi epi ∈ E. Pour tout i ∈ I tel que pi est un point-base de f , nous notons mi sa multiplicité pour f en tant que point-base. Alors nous avons l’égalité suivante : f# (c) · ℓ − c · ℓ = c · f#−1 (ℓ) − c · ℓ X = c · (dℓ − pi ∈supp(f ) = (d − 1)n − mi ) − n X (3.1) mi λi pi ∈supp(f ) Dans le cas où la somme des trois plus grandes multiplicités d’une classe ordonnée est inférieure ou égale à son degré, le lemme suivant permet de vérifier rapidement que cette classe appartient à V(id). P Lemme 3.2. Soit f ∈ Bir(P2 ) de degré d et c = nℓ − i∈I λi epi ∈ E. Notons, pour tout i ∈ I tel que pi ∈ supp(f ), mi la multiplicité du point-base pi . Considérons le multi-ensemble E constitué des indices de I, chacun apparaissant avec multiplicité mi . Alors, il est possible de partitionner E en d − 1 triplets et d’avoir ainsi, en notant T l’ensemble de ces triplets : X n − λi − λj − λk . f# (c) · ℓ − c · ℓ = {{i,j,k}}∈T De plus, il est possible de choisir la partition de E de sorte que chaque triplet de E soit constitué de trois indices deux à deux distincts. Rappelons que les indices de I correspondent à des points du support de c ou du support de f . Démonstration. D’après l’équation (3.1), nous avons l’égalité : X mi λi . f# (c) · ℓ − c · ℓ = (d − 1)n − (⋆) pi ∈supp(f ) L’égalité du lemme 1.13.1) 3(d − 1) = X mi pi ∈supp(f ) implique que nous pouvons constituer d − 1 triplets dans E afin d’arranger le terme de droite de (⋆) comme une somme de d − 1 termes de la forme n − λi − λj − λk , PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA où {{i, j, k}} ∈ T . Ainsi (d − 1)n − X mi λi = pi ∈supp(f ) X {{i,j,k}}∈T 28 n − λi − λj − λk . De plus, pour chaque i, l’inégalité du lemme 1.13.3) mi ≤ d − 1, permet de répartir les mi multiplicités λi de sorte qu’il n’y en ait pas deux dans le même terme de la somme et donc que les triplets de T soient formés d’indices deux à deux distincts. Une façon de former ces triplets est de faire une répartition en suivant. Nous répartissons le premier indice i0 de multiplicité non nulle mi0 dans les mi0 premiers triplets. Puis nous répartissons le second indice i1 dans les mi1 triplets suivant en revenant au premier quand nous avons compléter le dernier. En itérant ce procédé on obtient la partition en triplets attendue.  Lemme 3.3. Soit c ∈ E une classe de degré n. Supposons que l’une des deux conditions suivantes est réalisée. 1) Les points pi , pj et pk sont alignés. 2) Les points pi , pj et pk sont le support d’une application quadratique q et q# (c) · ℓ ≥ c · ℓ. Alors, en notant λi , λj et λk les multiplicités respectives, éventuellement nulles, des points pi , pj et pk pour c, nous avons : n − λi − λj − λk ≥ 0. Démonstration. Si les points pi , pj et pk sont alignés alors en considérant la droite passant par ces trois points, nous avons d’après l’inégalité de Bézout (Définition 2.1.4)) le résultat attendu. Si les points pi , pj et pk sont les points-base de l’application quadratique q, l’équation (3.1) nous donne : n − λi − λj − λk = q# (c) · ℓ − c · ℓ ≥ 0 ce qui prouve le lemme sous la seconde condition.  Une classe ordonnée qui possède au moins trois points dans son support est appelée spéciale si les points p1 et p2 sont adhérents à p0 et si ses trois plus grandes multiplicités vérifient : 0 > n − λ0 − λ1 − λ2 . Remarque 3.4. En particulier, une classe spéciale ne possède qu’une seule multiplicité maximale. En effet, si c est une classe spéciale alors λ0 > n2 puisque par positivité des excès (Définition 2.1.3)) λ0 ≥ λ1 + λ2 et donc 0 > n − λ0 − λ1 − λ2 ≥ n − 2λ0 . La conclusion s’obtient avec l’inégalité de Bézout (Définition 2.1.4)). L’objet des énoncés qui suivent est de caractériser les classes appartenant à V(id). Nous étudions dans un premier temps (Lemme 3.5 et Proposition 3.6) les classes non spéciales. D’une part car la preuve est facile puisque n’importe quel choix de partition de E en triplets d’indices deux à deux disjoints convient. D’autre part car elle donne une caractérisation immédiate pour contrôler si une classe non spéciale ordonnée appartient à V(id) ou pas. Il suffit de vérifier que la somme des trois premières multiplicité est plus petite que le degré. Dans un second temps nous nous intéressons aux classes spéciales (Théorème 3.8). PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 29 P Lemme 3.5. Soit c = nℓ − i∈I λi epi ∈ E une classe ordonnée. Si la somme des trois plus grandes multiplicités de c est inférieure ou égale à son degré : alors c ∈ V(id). n − λ0 − λ1 − λ2 ≥ 0, Démonstration. Considérons une telle classe c. L’hypothèse ainsi que la décroissance des multiplicités de c impliquent que pour tout triplet {{i, j, k}} d’indices deux à deux distincts de I, n − λi − λj − λk ≥ 0. Ainsi, par le lemme 3.2, pour tout f ∈ Bir(P2 ), f# (c) · ℓ ≥ c · ℓ. Par conséquent c ∈ V(id) comme attendu.  Si une classe c ordonnée n’est pas spéciale alors soit les points p0 , p1 et p2 sont alignés, soit ils forment le support d’une application quadratique, soit les points p1 et p2 sont adhérents à p0 et n − λ0 − λ1 − λ2 ≥ 0. Par positivité des excès en p0 (Définition 2.1.3)) il n’y a pas d’autre possibilité. P Proposition 3.6. Soit c = nℓ − i∈I λi epi ∈ E une classe ordonnée. 1) Si les points p0 , p1 et p2 sont alignés alors n− λ0 − λ1 − λ2 ≥ 0 et c ∈ V(id). 2) Si les points p0 , p1 et p2 sont le support d’une application quadratique alors la classe c appartient à V(id) si et seulement si n − λ0 − λ1 − λ2 ≥ 0. 3) Si les points p1 et p2 sont adhérents à p0 et si n − λ0 − λ1 − λ2 ≥ 0 alors c ∈ V(id). Remarque 3.7. Une classe c ordonnée et non spéciale qui appartient à V(id) a un degré supérieur ou égal à la somme de ses trois plus grandes multiplicités : n − λ0 − λ1 − λ2 ≥ 0. Démonstration. 1) Le premier point découle des lemmes 3.3 et 3.5. 2) Soit q une application quadratique ayant p0 , p1 et p2 comme points-base. Si c ∈ V(id), la définition de cellule de Voronoï implique qu’en particulier, −1 q# (c) · ℓ = c · q# (ℓ) ≥ c · ℓ. Le lemme 3.3 permet d’obtenir l’inégalité souhaitée. La réciproque découle du lemme 3.5. 3) Le troisième point découle du lemme 3.5.  Comme nous le verrons dans les exemples 3.27 et 3.28, il existe des classes spéciales appartenant à V(id). Par conséquent, la caractérisation précédente, à savoir que la somme des trois plus grandes multiplicités d’une classe ordonnée doit être inférieure ou égale au degré de la classe pour que celle-ci appartienne à V(id), ne caractérise pas toutes les classes de V(id). Ceci nous obligera par la suite à faire une disjonction de cas suivant si les classes sont spéciales ou non. De plus, il est difficile d’obtenir un critère aussi simple que dans le cas des classes non spéciales pour vérifier qu’une classe spéciale appartient à V(id). Le théorème 3.8 permet de réduire les applications birationnelles à tester pour vérifier qu’une classe spéciale est dans V(id). Il suffit de faire agir les applications de caractéristique Jonquières ayant comme point-base maximal le point correspondant à la plus grande multiplicité de c. En fait ce n’est pas si surprenant que le cas des classes spéciales soit PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 30 plus compliqué que le cas des classes non spéciales. Les trois premiers points du support des classes spéciales sont précisément dans la configuration qui avait été oubliée par M. Noether et qui ont posé problème dans sa preuve du théorème de Noether-Castelnuovo (voir le commentaire fait après le théorème 1.22). P Théorème 3.8. Soit c = nℓ − i∈I λi epi ∈ E une classe spéciale ordonnée. 1) La classe c est dans V(id) si et seulement si pour toute application de Jonquières j et de point-base maximal p0 nous avons j# (c) · ℓ ≥ c · ℓ. 2) De plus, si c ∈ V(id) ∩ V(f ) alors f est une application de caractéristique Jonquières dont l’inverse a comme point-base maximal p0 . En notant pi1 , . . . , pi2d−2 les petits points-base de f −1 , c vérifie : X 1 n = λ0 + λi . d−1 i∈{i1 ,...,i2d−2 } Si f est une application de caractéristique Jonquières alors elle envoie un pinceau de droites sur un autre pinceau de droites. Par conséquent, en la pré-composant par un automorphisme, c’est une application de Jonquières. Ainsi il existe un représentant de f qui est une application de Jonquières. Le reste de cette section est consacré à la preuve de ce théorème, et sera complété par deux exemples de classes spéciales. La démonstration est technique et est décomposée en faits pour faciliter la lecture. L’implication rapide est faite en premier. Pour montrer la réciproque, l’astuce est d’écrire la condition f# (c) · ℓ ≥ c · ℓ, sous forme d’une somme mettant en jeu les multiplicités de c qui sont supportées par les points-base de f de sorte que tous les termes de la somme soient positifs. Démonstration. Prouvons dans un premier temps le point 1). L’implication découle de la définition de cellule de Voronoï. En effet, une classe est dans la cellule associée à l’identité si et seulement si pour toute application f ∈ Bir(P2 ), f# (c) · ℓ ≥ c · ℓ. Par conséquent, si c ∈ V(id) alors en particulier, pour toute application de Jonquières j de point-base maximal p0 nous avons : j# (c) · ℓ ≥ c · ℓ. Intéressons nous à la réciproque du point 1). Supposons que pour toute application de Jonquières j de point-base maximal p0 , nous ayons : j# (c) · ℓ ≥ c · ℓ. (3.9) 2 Nous devons montrer que pour toute application f ∈ Bir(P ), 2 f# (c) · ℓ ≥ c · ℓ. Fixons f ∈ Bir(P ) de degré d dont les r points-base {pi } sont de multiplicité respective {mi }. La numérotation est induite par celle déjà faite sur les points pi du support de c : i ≤ j si et seulement si λi ≥ λj . Rappelons que si le point pi n’est pas dans le support de c alors nous posons λi = 0. D’après l’égalité (3.1) et le PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 31 lemme 3.2, et en gardant ses notations, nous devons montrer que la somme suivante est positive : X X n − λi − λj − λk ≥ 0. (⋆⋆) mi λi = (d − 1)n − pi ∈supp(f ) {{i,j,k}}∈T Le reste de la preuve consiste à choisir, lorsque c’est possible, les triplets de T de sorte que chaque terme soit positif. Lorsque ça ne sera pas le cas, nous partitionnerons également E avec des sous-ensembles de tailles différentes afin d’obtenir également des sommes de termes positifs. Fait 3.10. Pour tout triplet d’indices {i1 , i2 , i3 } tels que 0 < i1 < i2 ≤ i3 , nous avons : n − λi1 − λi2 − λi3 ≥ 0. De plus, si ce terme est nul alors pour tout 1 ≤ j ≤ i3 , λj = n − λ0 . Démonstration. Comme la classe c est spéciale les points p1 et p2 sont adhérents à p0 . Par positivité des excès en p0 pour c et par l’inégalité de Bézout pour la droite passant par les points p0 et p2 , nous avons : λi1 + λi2 + λi3 ≤ λ1 + 2λ2 ≤ λ0 + λ2 ≤ n. 2.1.3) 2.1.4) Si n − λi1 − λi2 − λi3 = 0 alors en considérant le cas d’égalité dans les inégalités ci-dessus nous obtenons λi3 = λ2 = n − λ0 . Nous concluons par décroissance des λi .  Fait 3.11. Si le point p0 n’est pas un point-base de multiplicité maximale pour f alors f# (c) · ℓ > c · ℓ. Démonstration. Si le point p0 n’est pas un point-base de multiplicité maximale pour f , alors il existe un point-base de f dans P2 , noté pr , tel que d−1 ≥ mr > m0 . 1.13.3) Nous posons comme convention que si p0 n’est pas un point-base de f , m0 = 0. Nous pouvons réécrire le terme de gauche de (⋆⋆) : X mi λi = m0 (n − λ0 − λr ) (d − 1)n − pi ∈supp(f ) + (d − 1 − m0 )n X mi λi − pi ∈supp(f ) i∈{0,r} / − (mr − m0 )λr . Nous complétons la seconde ligne à l’aide des (mr − m0 ) λr restants. C’est possible puisque nous avons vu que d − 1 ≥ mr . En répartissant les λi de la troisième ligne nous obtenons une partition en triplets similaire à celle du lemme 3.2. Cependant, comme nous avons imposé que certains triplets sont composés des indices 0 et r et d’autres d’indice r nous devons vérifier que c’est encore possible de former des triplets d’indices deux à deux disjoints. Ordonnons les triplets à remplir avec en PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 32 premier celui où il y a un indice r imposé puis ensuite ceux où aucun indice n’est imposé. Nous ne considérons pour le moment que ces triplets là. Nous répartissons les λi de la troisième ligne à l’aide de la « répartition en suivant » vue dans la preuve du lemme 3.2 en commençant au premier triplet où aucun indice n’est imposé. Pour chaque indice i tel que mi > d − 1 − m0 , une fois les d − 1 − m0 premiers indices i répartis nous complétons les triplets contenant les indices 0 et r. Puis nous répartissons l’indice suivant à partir de là où nous nous étions arrêtés dans la répartition en suivant. Si à la fin il reste des indices nous les mettons dans les triplets commençant par les indices 0 et r. Tous les termes de la somme sont ainsi soit de la forme n − λ0 − λr − λk avec k ∈ / {0, r} soit de la forme n − λi1 − λi2 − λi3 avec 0 < i1 < i2 < i3 . Montrons que les deux types de termes sont positifs ou nuls et qu’il existe au moins un terme strictement positif. • Par le fait 3.10, les (d−1−m0) termes de la forme n−λi1 −λi2 −λi3 sont positifs ou nuls. Montrons qu’il existe au moins un de ces termes qui est strictement positif. Comme d−1 > m0 et mr > m0 , il existe un terme où λr apparaît : n−λi1 −λi2 −λr avec 0 < i1 < i2 et r ∈ / {0, i1 , i2 }. Ce terme est strictement positif. En effet, sinon par le fait 3.10, nous aurions λ1 = λr = n − λ0 , ce qui implique que n − λ0 − λ1 − λr = λ0 − n < 0. Or les points p0 et pr étant distincts dans P2 , les points p0 , p1 et pr sont soit le support d’une application quadratique, soit alignés ce qui contredit le lemme 3.3. • Intéressons-nous aux m0 autres termes : n − λ0 − λr − λk où k ∈ / {0, r}. Il existe un point pk1 tel que pk soit voisin de, ou égal à pk1 et que les points p0 , pr et pk1 soient le support d’une application quadratique ou soient alignés. Ainsi par le lemme 3.3 et par la positivité de l’excès au point pk1 (Définition 2.1.3)), nous avons : n − λ0 − λr − λk ≥ n − λ0 − λr − λk1 ≥ 0. 3.3 2.1.3) Par conséquent, avec l’ensemble des triplets T définis ainsi, les termes de (⋆⋆) sont positifs et au moins un terme est strictement positif ce qui implique que cette somme est strictement positive.  Il nous reste à étudier le cas où le point p0 est de multiplicité maximale pour f qui va se révéler être le cas le plus difficile. Considérons les sous-ensembles disjoints d’indices des points-base de f suivants : • MIN est l’ensemble des indices i tels que pi ∈ supp(f ) et est un point 0 mineur de c, c’est-à-dire λi ≤ n−λ 2 . • ADH est l’ensemble des indices i ∈ / MIN tels que pi ∈ supp(f ) et est un point adhérent à p0 . • VNA est l’ensemble des indices i ∈ / MIN tels que pi ∈ supp(f ) et est un point voisin et non adhérent à p0 . Fait 3.12. Les ensembles MIN, ADH et VNA correspondent à une partition des points-base de f hors p0 . Démonstration. Ces ensembles d’indices étant disjoints, il suffit de montrer que tout point-base de f hors p0 est indicé par un élément de ADH ∪ VNA ∪ MIN. Tous les points-base de f voisins du point p0 sont indicés par un élément de ces trois sous-ensemble. Montrons que les points de P2 ou voisins d’un point de P2 PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 33 p m2 p m1 pm pr p0 Figure 1. Les points pm1 et pm2 sont adhérents au point pm . différent de p0 sont mineurs pour c. Soit pm ∈ P2 alors les points p0 , p1 , pm sont soit alignés soit le support d’une application quadratique donc d’après le lemme 3.3 nous obtenons l’inégalité : 0 ≤ n − λ0 − λ1 − λm ≤ n − λ0 − 2λm , 3.3 qui permet de conclure que le point pm est mineur pour c. De plus, tout point pm′ voisin de pm est mineur puisque λm′ ≤ λm par positivité des excès pour c en pm et en tous les points qu’il faut éclater pour obtenir p′m (Définition 2.1.3)).  Remarquons que par définition de l’ensemble MIN, tout terme de la forme suivante est positif : n − λ0 − λi − λj ≥ 0, pour tous i, j ∈ MIN. Notons que les indices i et j ne sont pas forcément distincts. Si l’ensemble ADH est vide, il en est de même de VNA par positivité des excès pour c et pour f (Définition 2.1.3) et Proposition 1.17). Par conséquent, dans ce cas-là tous les indices hors 0 sont dans MIN et la somme de (⋆⋆) est positive car tous les termes le sont. Supposons donc que ADH est non vide. Le but dans ce qui suit est d’arranger de façon adéquate les multiplicités des points indicés par ADH et par VNA afin d’obtenir une somme dont chaque terme est positif. Fait 3.13. Chaque point voisin de p0 possède au plus un point adhérent indicé par ADH ∪ VNA, les autres, s’ils existent, sont indicés par MIN. Démonstration. Soient pm1 et pm2 deux points adhérents au point pm voisin de p0 , et tels que λm2 ≤ λm1 (voir Figure 1). Notons pr , possiblement égal au point pm , le point libre et adhérent à p0 tel que pm soit voisin de pr . Alors d’après l’inégalité de Bézout pour la droite passant par les points p0 et pr et la positivité de l’excès en pr et en pm , nous avons : 0 ≤ n − λ0 − λr ≤ n − λ0 − λm ≤ n − λ0 − λm1 − λm2 ≤ n − λ0 − 2λm2 . 2.1.4) 2.1.3) 2.1.3) Par conséquent, m2 ∈ MIN.  Remarquons que le fait 3.13 implique que les points indicés par VNA sont tous libres. En effet, si un point pm2 est satellite cela signifie que pm2 est adhérent à deux points pm1 et pm et que pm1 est également adhérent à pm (voir Figure 2). Par PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 34 p m2 p m1 pm p0 Figure 2. Le point pm2 est un point satellite. le fait 3.13 et par positivité des excès en pm1 nous obtenons que m2 ∈ MIN. Par contre, les points indicés par ADH peuvent être libres ou satellites. Remarque 3.14. Un point libre indicé par ADH est un point infiniment proche du premier ordre de p0 . De plus, tout point satellite indicé par ADH est voisin d’un point libre indicé par ADH. Par exemple sur la figure 3, les points pi1 , pi2 , pi3 et pi4 sont adhérents au point p0 . Les points pi2 et pi3 sont des points satellites et sont voisins du point pi1 qui est libre. Le point pi4 est lui aussi libre. pi3 pi2 pi1 pi4 p0 Figure 3. Une configuration de points adhérents à p0 Fait 3.15. Supposons qu’il existe un sous-ensemble {p0 , pj1 , . . . , pj2δ } de l’ensemble des points-base de f qui est le support d’une application de Jonquières ayant pour point-base maximal p0 . Notons jmin ∈ {j1 , . . . , j2δ } un indice tel que λjmin = min{λi | i ∈ {j1 , . . . , j2δ }}. Alors pour tout i ≥ jmin , l’indice i est dans l’ensemble MIN. Démonstration. Soit j une application de caractéristique Jonquières ayant cet ensemble de points comme points-base. Notons qu’elle est de degré δ + 1. Nous avons alors par hypothèse sur c : 0 ≤ j# (c) · ℓ − c · ℓ = δn − δλ0 − (3.9) (3.1) 2δ X i=1 λji ≤ δ(n − λ0 − 2λjmin ). Par conséquent l’indice jmin est dans MIN, et par décroissance des λi , c’est le cas de tous les indices plus grands que jmin .  PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 35 Fait 3.16. Parmi les points-base de f voisins de tout point libre indicé par ADH il y a au moins autant de points indicés par ADH que par VNA. En particulier, le cardinal de VNA est strictement inférieur à celui de ADH. Démonstration. Soit a0 ∈ ADH tel que pa0 est un point libre. Notons a1 ≤ · · · ≤ ak les indices de ADH et v1 ≤ · · · ≤ vr les indices de VNA tels que les points pa1 , . . . , pak , pv1 , . . . , pvr soient les voisins de pa0 . Raisonnons par l’absurde et supposons que k soit strictement plus petit que r : k < r. Par le fait 3.13, ces points forment une tour où chaque point est adhérent au précédent mais n’est adhérent à aucun autre point de la tour, excepté de p0 lorsqu’ils sont indicés par ADH (voir Figure 4). Considérons l’ensemble pré-consistant {p0 , pa0 , pa1 , . . . , pak , pv1 , . . . , pvk+1 }. D’après le lemme 1.25, il existe une application de Jonquières de degré k + 2 ayant cet ensemble de points comme points-base et p0 comme point-base maximal. Par le fait 3.15, vk+1 , . . . , vr ∈ MIN ce qui est la contradiction attendue.  pvr pv1 pak pa1 pa0 p0 Figure 4. Tour de points au-dessus de p0 S’il existe un point-base majeur de f indicé par ADH alors nous introduisons une dernière catégorie d’indice non disjointe des précédentes appelée JONQ. Construisons cet ensemble. • tous les indices de ADH correspondant à des points majeurs pour f sont dans JONQ. • Ensuite nous ajoutons un à un les indices des points majeurs de f , dans l’ordre décroissant des λi , de façon à ce que l’ensemble reste consistant à chaque étape. • Nous nous arrêtons lorsqu’il y a autant de points adhérents à p0 que de points non adhérents à p0 . Cela arrive forcément d’après le lemme 1.24. Remarque 3.17. D’après la proposition 1.26, il existe une application de Jonquières ayant p0 comme point-base maximal et possédant l’ensemble des points indicés par JONQ comme points-base de multiplicité 1. Fait 3.18. Tout point-base majeur de f non indicé par JONQ est indicé par MIN. Démonstration. Raisonnons par l’absurde et supposons qu’il existe ps un point-base majeur de f tel que s ∈ / JONQ ∪ MIN. Comme les indices de ADH sont tous dans JONQ, la seule possibilité est que s ∈ VNA. Considérons un indice r de JONQ tel PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 36 que pour tout j ∈ JONQ, λr ≤ λj . Par la remarque 3.17 et le fait 3.15, l’indice r est dans MIN et non pas dans ADH. Par conséquent, c’est l’un des derniers points ajoutés et comme nous rangeons les indices dans JONQ dans l’ordre décroissant des λi cela signifie que λs ≤ λr . Par conséquent, s appartient à MIN, ce qui nous donne la contradiction attendue.  Soit 2δ le cardinal de l’ensemble JONQ. Nous renommons par j1 , . . . , j2δ l’ensemble des indices de JONQ de sorte que mj1 ≥ · · · ≥ mj2δ et que pour tout 1 ≤ i ≤ 2δ, l’ensemble de points {pj1 , . . . , pji } muni de leur multiplicité respective est consistant. S’il n’existe pas de point-base majeur de f indicé par ADH alors JONQ est un ensemble vide et l’entier δ ainsi que les multiplicités mji seront considérés comme nuls dans les calculs qui suivent. Commençons à réorganiser l’équation (⋆⋆). Par le fait 3.12, il est possible de répartir les termes de la façon suivante : X mi λi = mj2δ (δn − δλ0 − λj1 − · · · − λj2δ ) (d − 1)n − pi ∈supp(f ) + (m0 − δmj2δ )(n − λ0 ) + (d − 1 − m0 )n − − − 2δ−1 X i=1 (mji − mj2δ )λji X (3.19) mi λi i∈(VNA∪ADH)\JONQ X mi λi . i∈MIN\JONQ Lorsque nous sommes dans le cas des trois premières lignes, les termes en les mi sont appelés les poids des termes en les multiplicités λi . Notre but, à présent, est de répartir les λi des sommes des trois dernières lignes afin que tous les termes soient positifs. Posons pour tout point-base pi de f : ( mi − mj2δ si i ∈ JONQ m̃i = . mi sinon Considérons les indices i de ADH dont la multiplicité m̃i est non nulle. Notons-les a1 , . . . , ak de sorte que m̃a1 ≥ · · · ≥ m̃ak . Faisons de même pour les indices de VNA et notons les v1 , . . . , vr de sorte que m̃v1 ≥ · · · ≥ m̃vr . Malgré le fait 3.16, k n’est pas forcément strictement supérieur à r. En effet il se pourrait qu’une situation analogue à l’exemple suivant arrive. Considérons la tour de points p0 , p1 , p2 , p3 incluse dans l’ensemble des points-base de f où chacun est adhérent au précédent, où p2 est également adhérent à p0 (voir Figure 5) et où 1, 2 ∈ ADH et 3 ∈ VNA. Si nous supposons de plus que 1, 2 ∈ JONQ, 3 ∈ / JONQ, que m1 = m2 = mj2δ et que les autres points-base de f sont dans P2 alors k = 0 alors que r = 1. Cette situation arrive si le point p3 est un point-base de f qui n’est pas majeur pour f mais qui est majeur pour c. PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 37 p3 p2 p1 p0 Figure 5 Fait 3.20. Si r ≥ k > 0 alors il existe au moins r − k + 1 couples d’indices, notés (ak+t , vit )1≤t≤r−k+1 tels que tous les indices sont deux à deux distincts et pour 1≤t≤r−k+1 : • • • ak+t ∈ ADH \ {a1 , . . . , ak }, vit ∈ {vi }1≤i≤r , pvit est voisin de pak+t . Remarquons que la condition ak+t ∈ ADH \ {a1 , . . . , ak } est équivalente à ak+t ∈ ADH ∩ JONQ et m̃ak+t = 0. Remarque 3.21. L’hypothèse k > 0 est indispensable car sinon nous pourrions avoir la situation précédent le fait 3.20. Dans ce cas-là, k = 0 et r = 1. Il y a effectivement deux points indicés par ADH qui ont une multiplicité m̃i nulle mais il n’existe qu’un point indicé par VNA et voisin des points p1 et p2 . Démonstration. Notons pi1 , . . . , pin les points libres indicés par ADH. Tous les points indicés par ADH ∪ VNA sont voisins ou égaux à ces points-là d’après la remarque 3.14. Pour tout j ∈ {1 . . . , n}, considérons le point pij ainsi que ses points voisins indicés par ADH ∪ VNA. D’après le fait 3.13, ils forment une tour où chacun est adhérent uniquement au point précédent, sauf les points indicés par ADH qui sont également adhérents au point p0 . Notons kij et rij le cardinal des indices des points voisins ou égaux à pij dont la multiplicité m̃i est non nulle et qui sont respectivement indicés par ADH et VNA. Si kij ≤ rij alors d’après le fait 3.16, cela implique qu’il existe au moins rij −kij +1 points distincts de cette tour indicés par ADH qui ont une multiplicité m̃i nulle. Dans le cas où kij est non nul, rij − kij + 1 ≤ rij donc il existe au moins rij − kij + 1 points distincts de cette tour indicés par VNA qui ont une multiplicité m̃i non nulle. Si kij = 0 alors il y en a rij = rij − kij . Par conséquent dans les deux cas il existe rij − kij points distincts de cette tour indicés par VNA qui ont une multiplicité m̃i non nulle et autant indicés par ADH qui ont une multiplicité m̃i nulle. Nous pouvons en ajouter un de plus dans chaque famille lorsque kij est non nul. Notons J l’ensemble des indices j ∈ {1, . . . , n} tels que kij ≤ rij . Comme nous avons supposé k non nul cela signifie qu’il existe au moins un élément s ∈ {1 . . . , n} PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 38 tel que kis est non nul. S’il y en a plusieurs, nous en choisissons un. Posons : Cs =  P (rij − kij ) ris − kis + 1 +    j∈J si s ∈ J j6=s P    (rij − kij ) sinon . j∈J Nous avons montré que nous pouvons choisir au moins Cs couples ayant les propriétés de l’énoncé. Dans les deux cas de la définition de Cs , comme rij − kij < 0 pour j ∈ / J, nous avons r−k = X j∈J (rij − kij ) + X j ∈J / (rij − kij ) < Cs , par conséquent nous pouvons choisir r − k + 1 couples comme dans l’énoncé. Fait 3.22. Nous avons m̃a1 ≤ d−m0 2 et m̃v1 ≤  d−m0 2 . Démonstration. Montrons le résultat pour m̃a1 le résultat s’obtient de la même manière pour m̃v1 . Si le point pa1 n’est pas un point-base majeur pour f , le résultat est immédiat puisque m̃a1 = ma1 . Si le point pa1 est un point point-base majeur de f alors a1 ∈ JONQ sinon nous aurions une contradiction avec le fait 3.18 qui impliquerait que a1 ∈ MIN. Comme l’ensemble JONQ est pré-consistant, il existe un indice i dans JONQ ∩ ADH, possiblement égal à a1 tel que pa1 est voisin de pi et tel que le point pi est 0 libre. Raisonnons par l’absurde et supposons que ma1 − mj2δ = m̃a1 > d−m 2 . Par positivité de l’excès au point pi pour f (Proposition 1.17), et comme pj2δ est un point-base majeur pour f , nous avons : mi ≥ ma1 > d − m0 + mj2δ > d − m0 . 2 Nous obtenons une contradiction avec le théorème de Bézout (Proposition 1.18) en considérant la droite passant par les points p0 et pi , puisque le point pi est, par hypothèse, libre et adhérent à p0 .  Le fait 3.22 implique que d − m0 − m̃a1 − m̃v1 ≥ 0. Nous pouvons donc réarranger les termes en considérant deux cas, suivant si d − m0 − m̃a1 − m̃v1 est nul ou pas. PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 39 i) Si d − m0 − m̃a1 − m̃v1 ≥ 1 alors nous répartissons les termes de (3.19) de la façon suivante : X (∗jonq ) mi λi = mj2δ (δn − δλ0 − λj1 − · · · − λj2δ ) (d − 1)n − pi ∈supp(f ) + (m0 − δmj2δ )(n − λ0 )  + m̃ak (n − λa1 − · · · − λak )     + (m̃ak−1 − m̃ak )(n − λa1 − · · · − λak−1 )  ..   .     + (m̃a1 − m̃a2 )(n − λa1 )  + m̃vr (n − λv1 − · · · − λvr )     + (m̃vr−1 − m̃vr )(n − λv1 − · · · − λvr−1 )  ..   .     + (m̃v1 − m̃v2 )(n − λv1 ) + (d − 1 − m0 − m̃a1 − m̃v1 )n X − m̃i λi . (∗λ0 ) (∗adh ) (∗vna ) (∗n ) i∈MIN Remarquons que les expressions des lignes (∗adh ) et (∗vna ) sont respectivement égales à r k X X m̃vi λvi . m̃ai λai et m̃v1 n − m̃a1 n − i=1 i=1 Si r = 0 la somme s’écrit de la même façon sans les termes de (∗vna ). De même, si k = 0 il n’y a plus les termes de (∗adh ). Montrons que les termes de chaque ligne exceptée la dernière, sont positifs. Ensuite, nous répartirons les λi où i ∈ MIN de sorte que chaque terme reste positif. ∗ Les poids sont positifs. Les points pj1 , . . . , pjδ sont adhérents à p0 , par conséquent, par décroissance des multiplicités mji et par positivité des excès en p0 pour f , nous avons : δ X m0 ≥ mji ≥ δmj2δ . 1.17 i=1 Ainsi, le poids m0 − δmj2δ est positif. Les autres poids le sont par hypothèse de décroissance des multiplicités et par hypothèse du cas i). ∗ Les termes en λi sont positifs. Le terme en λi de (∗jonq ) est positif par l’hypothèse faite sur c (équation (3.9)) et par la remarque 3.17 et l’équation (3.1). Considérons les termes de (∗adh ). En utilisant la positivité des excès en p0 pour c, nous avons pour 1 ≤ i ≤ k : n − λa1 − λa2 − · · · − λai ≥ n − λ0 > 0. 2.2 2.1.3) (3.23) Par le fait 3.16 et par positivité des excès pour c en chaque point adhérent (Proposition 2.1.3)), nous avons : X X λi ≤ λi . i∈VNA i∈ADH PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 40 Ainsi les termes de (∗vna ) sont strictement positifs, pour 1 ≤ i ≤ r : n − λv1 − λv2 − · · · − λvi ≥ n − X i∈VNA λi ≥ n − X i∈ADH λi ≥ n − λ0 > 0. 2.2 (3.24) ∗ Répartissons les λi où i ∈ MIN. Nous voulons répartir les λi où i ∈ MIN, de sorte que les termes soient comme dans le lemme 3.2, de la forme n − λi − λj − λk . Cependant, si k ≥ 4 ou r ≥ 4, certains termes (∗adh ) et (∗vna ) sont surchargés, c’est-à-dire qu’ils contiennent plus de trois multiplicités λi , par conséquent d’autres termes devront être sous-chargés. Nous répartissons ainsi les multiplicités λi où i ∈ MIN dans les termes de (∗λ0 ), (∗n ) et les deux derniers de (∗adh ) et de (∗vna ). Les termes que nous venons de compléter sont positifs par la définition de l’ensemble MIN. Nous venons de montrer que si c satisfait la condition (3.9) alors pour toute application f qui a pour point-base maximal p0 et dont les points-base satisfont i), f# (c) · ℓ ≥ c · ℓ. ii) Plaçons nous dans le cas où d − m0 − m̃a1 − m̃v1 = 0. Le fait 3.22 implique que m̃a1 = m̃v1 = d − m0 . 2 En particulier, k et r sont strictement positifs. Nous ne pouvons pas répartir les termes de la somme comme dans le cas i) puisque le poids de (∗n ) est à présent négatif : −1 − m0 − m̃a1 − m̃v1 = −1. Si r ≥ k, nous reprenons le résultat et les notations du fait 3.20. Posons J = {it }1≤t≤r−k+1 l’ensemble des indices apparaissant dans le fait 3.20. Notons J c son complémentaire. Les ensembles J et J c forment une partition de l’ensemble des indices {1, . . . , r}. Dans le cas où J c est non vide, nous notons t0 un de ses éléments. S’il est vide les termes de (∗v ) n’existent pas et λvt0 = 0. Si r < k alors J est vide et J c contient les r indices {1, . . . , r}. Par rapport à la répartition des termes dans la somme du cas i), nous voulons répartir une fois les multiplicités λvi dans d’autres termes afin que le poids de (∗n ) soit positif. Pour cela, nous surchargeons le terme (∗av ) afin de ne pas avoir à compléter les termes (∗v ). Nous arrangeons donc les termes de la somme de la façon suivante : PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA (d − 1)n − X pi ∈supp(f ) 41 mi λi = mj2δ (δn − δλ0 − λj1 − · · · − λj2δ ) + n − λa1 − λa2 − · · · − λak − λvt0 − + (m̃ak − 1)(n − λa1 − · · · − λak ) .. . t∈J c \{t0 } X λvt (∗av ) t∈J      − · · · − λak−1 )  + (m̃ak−1 − m̃ak )(n − λa1 + (m̃a1 − m̃a2 )(n − λa1 ) X (n − λ0 − λvt ) + (∗jonq )       (∗adh ) (∗v ) + (m̃vr − 1)(n − λv1 − · · · − λvr ) + (m̃vr−1 − m̃vr )(n − λv1 .. . + (m̃v1 − m̃v2 )(n − λv1 )      − · · · − λvr−1 )  + (m0 − Card(J c ) + 1 − δmj2δ )(n − λ0 ) + (d − 1 − m0 − m̃a1 − m̃v1 + 1)n X − mi λi .       (∗vna ) (∗λ0 ) (∗n ) i∈MIN Remarquons que les expressions des lignes (∗av ) et (∗adh ) sont égales à : m̃a1 n − k X i=1 m̃ai λai − λvt0 − celle de la ligne (∗v ), si elle existe, est égale à : (Card(J c ) − 1)(n − λ0 ) − et celles des lignes (∗vna ) est égale à : (m̃v1 − 1)n − r X i=1 X λvt , t∈J X λvt , t∈J c \{t0 } (m̃vi − 1)λvi . Montrons que nous pouvons compléter les termes incomplets à l’aide des λi où i ∈ MIN afin que chaque terme soit positif. ∗ Les poids sont positifs. Montrons dans un premier temps que le cardinal de J c est inférieur ou égal à k − 1 : ( r ≤k−1 si r < k Card(J c ) = . r − (r − k + 1) = k − 1 sinon En utilisant cela ainsi que la positivité des excès en p0 pour f , nous obtenons : k X X X m̃ai + mj2δ ≥ k + δmj2δ > Card(J c ) + δmj2δ . m0 ≥ mi = 1.17 i∈ADH i=1 i∈JONQ∩ADH PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 42 Ainsi le poids m0 − Card(J c ) + 1 − δmj2 δ est strictement positif. Les autres le sont par décroissance des m̃i et par l’hypothèse faite dans le cas ii). ∗ Les termes en λi sont positifs. Les termes en λi de (∗jonq ), (∗adh ) et (∗vna ) sont positifs par les mêmes arguments que dans le cas i) puisqu’ils n’ont pas changé. S’il existe, le terme de (∗v ) est positif d’après l’inégalité de Bézout (Proposition 2.1.4)). Il reste à montrer la positivité du terme en λi de (∗av ). Si ℓ ≥ k alors d’après le fait 3.20, il existe des points pak+1 , . . . , paℓ+1 adhérents à p0 , distincts des {pai }1≤i≤k et tels que pour t ∈ J, pvit est voisin de pak+t . Ainsi par positivité des excès audessus de chaque point adhérent : X X λvt ≤ λak+t . t∈J t∈J Notons pa le point libre adhérent à p0 tel que le point pvt0 est voisin de pa . Nous avons par positivité des excès pour les points p0 et pa , et par l’inégalité de Bézout : n − λa1 − · · · − λak − λvt0 − X t∈J λvt ≥ n − k X i=1 λak − X t∈J ≥ n − λ0 − λvt0 2.1.3) ≥ n − λ0 − λa 2.1.3) ≥ 0, 2.1.4) λak+t − λvt0 (3.25) et donc le terme en λi de (∗av ) est positif. Si r < k alors J est vide et nous obtenons le même résultat en utilisant la même inégalité qu’au-dessus en enlevant la somme sur J. ∗ Montrons que nous pouvons laisser incomplets les termes de (∗v ), s’ils existent. Si k = 1 alors J c est vide et il n’y a par conséquent rien à faire. Sinon le terme de (∗av ) possède k + Card(J) + 1 − 3 = k − 2 + Card(J) multiplicités λi en trop qui compensent les Card(J c ) − 1 ≤ k − 2 multiplicités manquantes. ∗ Complétons les termes de la somme avec les λi où i ∈ MIN. D’après le point précédent, les termes à compléter sont les deux dernières lignes de (∗adh ) et (∗vna ), ainsi que les termes (∗λ0 ) et (∗n ). Par la définition de MIN tous les termes complétés sont positifs. Ainsi la somme est positive comme attendue. Ceci conclut le cas ii) et donc le point 1) du théorème 3.8. Montrons maintenant le point 2) du théorème 3.8. Supposons maintenant que c ∈ V(id) ∩ V(f ). Nous voulons montrer que f est de caractéristique Jonquières dont le point-base de multiplicité maximale de l’inverse est p0 . Par définition : c · ℓ = c · f# (ℓ) = f#−1 (c) · ℓ. Ainsi, d’après le fait 3.11, le point p0 doit être un point-base de f −1 de multiplicité maximale. De plus, la somme (⋆⋆) appliquée à f −1 doit être nulle. Suivant si nous nous trouvons dans le cas i) ou dans le cas ii), nous utilisons la répartition des termes faite dans i) ou dans ii). Comme c ∈ V(id), nous avons déjà montré, dans PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 43 la première partie de la preuve, que tous les termes de ces sommes sont positifs ou nuls. Nous devons étudier à quelles conditions ils sont en fait tous nuls. a) Les points-base de f −1 ne peuvent pas satisfaire la condition du cas ii). Supposons que ce soit le cas. Alors, le terme (∗av ) est nul ce qui implique en prenant le cas d’égalité des inégalités (3.25) que λvt0 = n − λ0 . Considérons la suite minimale de points du support de f −1 , hors p0 , qu’il a fallu éclater pour obtenir pvt0 et notons-les pi1 , . . . , pik , pik+1 où pik+1 = pvt0 . Comme vt0 ∈ VNA, par positivité des excès, les indices i1 , . . . , ik appartiennent à VNA ∪ ADH. D’après le fait 3.13, chacun est adhérent uniquement au point précédent sauf les points indicés par ADH qui sont également adhérents à p0 . Notons k ′ le plus petit entier tel que pik′ ne soit pas adhérent à p0 . Considérons k ′ − 2 points de P2 (pas forcément dans le support de c) tels que deux points de cette famille ne soient pas alignés avec p0 . D’après le lemme 1.25 ces points ainsi que les points p0 , pi1 , . . . , pi′k forment le support d’une application de Jonquières j de degré k ′ . Nous obtenons la contradiction suivante en utilisant la positivité des excès au-dessus des points {pis }1≤s≤k : 0 ≤ j# (c) · ℓ − c · ℓ (3.9) ≤ (k ′ − 1)n − (k ′ − 1)λ0 − λi1 − · · · − λi′k (3.1) ≤ (k ′ − 1)n − (k ′ − 1)λ0 − k ′ λvt0 2.1.3) (3.26) = −(n − λ0 ) < 0. 2.2 Par conséquent, dans le cas ii) il n’existe pas de f telle que c ∈ V(id) ∩ V(f ). b) Les points-base de f −1 doivent donc satisfaire les conditions du cas i). Tous les termes de la somme du cas i) étant nuls, c’est en particulier le cas du terme m̃ak (n − λa1 − λa2 − · · · − λak ) de (∗adh ). Par hypothèse sur le choix de l’indice k, m̃ak est non nul, c’est donc n − λa1 − λa2 − · · · − λak qui est nul. L’équation (3.23) implique que k ≤ 2 puisque pour k = 1 et k = 2 les termes ont été complétés à l’aide des λi où i ∈ MIN. Mais le fait 3.10 implique qu’alors λi = n− λ0 ce qui est absurde puisque i ∈ MIN. Ainsi k = 0 et il n’y a pas de termes (∗adh ). Nous obtenons de même qu’il n’y a pas de termes (∗vna ). Par le fait 3.10, le terme en λi de (∗n ) est strictement positif. Par conséquent, pour que (∗n ) soit nul, il faut que son poids qui est égal à d − 1 − m0 le soit. Ceci implique que m0 = d − 1. Par conséquent f −1 est une application de caractéristique Jonquières et dont le point-base maximal est p0 d’après le point 4) du lemme 1.13. De plus, en notant pi1 , . . . , pi2d−2 les petits points-base de f −1 , la condition c · ℓ = j#−1 · ℓ se réécrit grâce à l’équation (3.1) : X 1 λi . n = λ0 + d−1 Ceci achève la preuve du théorème 3.8. i∈{i1 ,...,i2d−2 }  Exemple 3.27. Il existe des classes spéciales dans la cellule V(id). Considérons p0 ∈ P2 et deux points libres et adhérents à p0 notés p1 et p2 . Considérons la classe 1 c = (7ℓ − 4ep0 − 2ep1 − 2ep2 ). 5 Montrons qu’elle appartient à E. Elle est d’auto-intersection 1. Les excès sont positifs. Elle est positive contre l’anti-canonique. Montrons qu’elle satisfait également PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 44 la condition de Bézout. C’est clair pour les droites. Pour les coniques aussi (à noter que la seule conique passant par les trois points est la réunion de deux droites). Il existe une cubique C passant avec multiplicité 2 en p0 et avec multiplicité 1 aux points p1 et p2 . Considérons une courbe D de P2 passant avec multiplicité µi aux points pi . Nous avons alors d’après le théorème de Bézout : 3d = C · D ≥ 2µ0 + µ1 + µ2 . Par conséquent, 7d ≥ 6d ≥ 4µ0 + 2µ1 + 2µ2 . Et la condition de Bézout est vérifiée pour toute courbe. Ainsi, la classe c appartient à E (Définition 2.1). De plus, les points p1 et p2 sont adhérents à p0 et le degré de c est strictement plus petit que la somme de ses trois plus grandes multiplicités. Par conséquent, c est une classe spéciale. Montrons que c ∈ V(id). D’après le théorème 3.8, il suffit de vérifier que pour toute application j de caractéristique Jonquières dont l’inverse a pour point-base maximal p0 nous avons j# (c) · ℓ ≥ ℓ · c ⇔ c · j−1 # (ℓ) ≥ c · ℓ. Il n’existe pas d’application quadratique dont les points-base ont cette configuration (par positivité des excès). Soit q une application quadratique dont les points-base de l’inverse sont les points p0 , p1 et un troisième noté q. Dans ce cas, nous avons 1 8 7 c · q# (ℓ) = (14 − 4 − 2) = > = c · ℓ. 5 5 5 Pour tout d ≥ 3, il existe une application de Jonquières j de degré d telle que les points p0 , p1 et p2 soient des points-base de j−1 avec p0 point-base maximal. Dans ces cas, nous avons : 3d 7 1 > . c · j# (ℓ) = (7d − 4(d − 1) − 2 − 2) = 5 5 5 Par conséquent, c appartient à la cellule identité et n’appartient à aucune autre cellule. L’exemple suivant montre que les germes de cellules contenant une classe spéciale en commun avec la cellule V(id) ne sont pas forcément des applications de caractéristique Jonquières dont les points-base sont en configuration spéciale. Exemple 3.28. Reprenons la même configuration que dans l’exemple précédent. Considérons deux points p1 et p2 libres et adhérents à un point p0 ∈ P2 . Posons : 1 c = √ (7ℓ − 4ep0 − 3ep1 − ep2 ). 23 En faisant exactement les mêmes étapes que dans l’exemple précédent, excepté pour l’inégalité de Bézout où nous devons en plus remarquer que d ≥ µ1 nous montrons dans un premier temps que la classe c est dans E. Puis en faisant les mêmes calculs d’intersection, nous montrons que c appartient à V(id) et aux cellules dont le germe est une application quadratique dont les points-base de l’inverse sont les points p0 , p1 et un troisième point. De plus, c n’appartient à aucune autre cellule. 4. Cellules adjacentes à la cellule associée à l’identité Deux cellules de Voronoï sont dites adjacentes si leur intersection est non vide. Dans cette partie, nous déterminons les germes des cellules de Voronoï qui sont adjacentes à la cellule V(id) (Théorème 4.1) et nous caractérisons les classes se trouvant dans de telles intersections. PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA Théorème 4.1. Soit c = nℓ − P i∈I 45 λi epi ∈ E une classe ordonnée. 1) Si c est une classe spéciale appartenant à V(id) ∩ V(f ) alors f est une application de caractéristique Jonquières dont l’inverse a pour point-base maximal p0 . De plus en notant pi1 , . . . , pi2d−2 les petits points-base de f −1 , c vérifie : X 1 n = λ0 + λi . d−1 i∈{i1 ,...,i2d−2 } 2) Dans le cas où c n’est pas une classe spéciale, quitte à permuter l’ordre des points pi qui ont même multiplicité λi , nous avons c ∈ V(id) ∩ V(f ) si et seulement si nous sommes dans l’une des trois situations suivantes. a) L’application f est quadratique, p0 , p1 et p2 sont les points-base de f −1 et n = λ0 + λ1 + λ2 où λ2 peut éventuellement être nulle. b) L’application f est de caractéristique Jonquières non quadratique, p0 est le point-base de multiplicité maximale de f −1 , les points {pi }1≤i≤2d−2 sont ses autres points-base et c = nℓ − λ0 ep0 − n−λ0 2 2d−2 X n − λ0 X epi − λi epi 2 i=1 i≥2d−1 où λ0 ≥ ≥ λi pour tout i ≥ 2d − 1. c) L’application f n’est pas de caractéristique Jonquières et possède r ∈ {6, 7, 8} points-base en position presque générale, les points {pi }0≤i≤r−1 sont les points-base de f −1 et r−1 X nX epi − λi epi c = nℓ − 3 i=0 i≥r où n 3 ≥ λi pour tout i ≥ r. Remarque 4.2. Par le lemme 1.21, les points-base de f sont en position presque générale si et seulement si il en est de même des points-base de f . Remarque 4.3. Une classe c non-spéciale qui contient seulement deux points dans son support et telle que la somme de ses deux plus grandes multiplicités est égale à son degré, appartient à une infinité de cellules de Voronoï. En effet, c est au bord de V(id) et appartient à toutes les cellules V(q) où q est une application quadratique telle que le support de c est inclus dans le support de q−1 . Ces cellules sont en nombre infini puisque il y a un nombre infini d’applications quadratiques modulo PGL(3, k). Une classe spéciale peut également appartenir à une infinité de cellules de Voronoï, comme par exemple dans 3.28. Dans les autres cas, les classes n’appartiennent qu’à un nombre fini de cellules de Voronoï. Démonstration. Si c est une classe spéciale et c ∈ V(id)∩V(f ) alors d’après le point 2) de la proposition 3.8, f −1 doit être de caractéristique Jonquières et doit avoir p0 comme point-base P de multiplicité maximale. Soit c = nℓ − λi epi ∈ E une classe non spéciale et ordonnée. La classe c ∈ V(f ) si et seulement si c · ℓ = c · f# (ℓ) = f#−1 (c) · ℓ. PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 46 Notons d le degré de f −1 et mi les multiplicités de ses points-base pi . D’après l’équation (3.1) et en reprenant les notations et le résultat du lemme 3.2, il existe un ensemble T de d − 1 triplets d’indices des points-base de f −1 dont chaque triplet est formé d’indices deux à deux disjoints et tel que l’égalité ci-dessus est équivalente à X X n − λi − λj − λk . (4.4) 0 = (d − 1)n − mi λi = pi ∈supp(f −1 ) {{i,j,k}}∈T Dans la somme de droite les multiplicités λi apparaissent mi fois. Comme la classe c n’est pas spéciale, elle appartient à V(id) si et seulement si pour tout i < j < k : n − λi − λj − λk ≥ n − λ0 − λ1 − λ2 ≥ 0. 3.7 (4.5) Ainsi c ∈ V(id) ∩ V(f ) si et seulement si tous les termes de la somme de droite de (4.4) sont nuls. Si nous sommes dans l’un des trois cas a), b) ou c) alors c appartient à V(id) ∩ V(f ). Montrons à présent l’implication et supposons que c appartient à V(id) ∩ V(f ). Distinguons les cas suivant si f est une application quadratique, si f est une application de caractéristique Jonquières non quadratique et enfin si f n’est pas de caractéristique Jonquières. a) Si f est une application quadratique dont l’inverse a pour points-base pi0 , pi1 , pi2 alors la somme (4.4) s’écrit n − λi0 − λi1 − λi2 = 0. Les multiplicités λi0 , λi1 et λi2 sont les trois plus grandes multiplicités de c car sinon : 0 = n − λi0 − λi1 − λi2 > n − λ0 − λ1 − λ2 , ce qui contredirait l’équation (4.5). Ainsi, les points p0 , p1 et p2 sont les points-base de f −1 et n = λ0 + λ1 + λ2 . b) Dans le cas où f est de caractéristique Jonquières de degré strictement supérieur à 2, notons λi0 la multiplicité pour c correspondant au point-base maximal de f −1 et λi1 , . . . , λi2d−2 les multiplicités pour c correspondant aux points-base pi1 , . . . , pi2d−2 de f −1 de sorte que λi1 ≥ · · · ≥ λi2d−2 . La somme (4.4) se réécrit : 0= d−1 X j=1 (n − λi0 − λi2j−1 − λi2j ). Comme tous les termes sont positifs ou nuls, ils sont en fait nuls. Ainsi, nous n−λ obtenons que λi1 = · · · = λi2d−2 = 2 i0 . De même que précédemment pour ne pas contredire l’équation (4.5), nous devons avoir λi0 = λ0 et λ1 = λi1 comme attendu. c) Considérons enfin le cas où f n’est pas de caractéristique Jonquières. Notons λi0 , λi1 , . . . , λir−1 les multiplicités de c correspondant aux points-base pi0 , . . . , pir−1 de f −1 de sorte que λi0 ≥ λi1 ≥ · · · ≥ λir−1 . Comme f n’est pas une application de Jonquières elle est de degré au moins 4 et par le lemme 1.14.1) elle possède au moins six points-base. De plus, par les points 3) et 4), pour tout i, mi ≤ d − 2. Par conséquent, nous pouvons arranger les termes de la somme (4.4) de sorte qu’il y en ait un de la forme : n−λi0 −λi1 −λi2 et un autre de la forme : n−λir−3 −λir−2 −λir−1 où r − 3 > 2. Comme c appartient à V(f ), ils sont tous les deux nuls : n = λi0 + λi1 + λi2 = λir2 + λir−1 + λir . PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 47 Par décroissance des λij cela signifie qu’ils sont tous égaux à n3 . Si maintenant ce ne sont pas les multiplicités maximales pour c cela implique en particulier que λ0 > n3 par conséquent : n n − λ0 − λ1 − λ2 < n − 3 = 0. 3 Or, ceci contredit l’équation (4.5). Ainsi, nous avons montré que si c appartient à V(id) ∩ V(f ) alors c est r′ -symétrique avec r′ ≥ r, c a pour multiplicité maximale n −1 sont inclus dans les points du support de c ayant la 3 et les points-base de f multiplicité maximale, comme attendu. De plus, d’après le lemme 2.14, une telle classe r′ -symétrique vérifie r′ ≤ 8, par conséquent f −1 ne possède pas plus de 8 points-base. Si f −1 possède quatre points-base alignés ou sept points-base sur une conique cela contredit l’inégalité de Bézout (Définition 2.1.4)) et donc cela contredit le fait que c ∈ E. De même si f −1 possède deux points-base adhérents à un même troisième point-base nous obtenons une contradiction avec la positivité des excès  (Définition 2.1.3)). La remarque 4.2 permet de conclure. Exemple 4.6. Voici des exemples de classes appartenant à l’intersection entre V(id) et une autre cellule. • La classe : ! r 2d−2 d+1 1 X d−1 2 ep ℓ− ep0 − cj = d+1 2 2 2 i=1 i appartient à V(id) ∩ V(j) où j est une application de caractéristique Jonquières de degré d dont l’inverse a pour point-base maximal p0 et pour autres points-base {pi }1≤i≤2d−2 . • La classe : r−1 X 3 1 cf = √ epi ℓ− √ 9−r 9 − r i=0 appartient à V(id) ∩ V(f ) où f −1 une application ayant au plus 8 points-base {pi }0≤i≤r−1 qui sont en position presque générale (r ≤ 8). Le corollaire 4.7 résume tous les germes des cellules adjacentes à V(id). Corollaire 4.7. L’ensemble des germes des cellules adjacentes à V(id) est constitué de toutes : 1) les applications de caractéristique Jonquières, 2) les applications ayant au plus 8 points-base, qui ne sont pas de caractéristiques Jonquières et dont les points-base de l’inverse sont en position presque générale. Remarque 4.8. Les conditions du point 2) du corollaire 4.7 sont nécessaires. Soient p0 , . . . , p4 ∈ P2 tels que trois points ne soient pas alignés. Ce sont les points-base d’une application de Jonquières j de degré 3 et de point-base maximal p0 . Soit L la droite passant par les points p1 et p2 . Considérons trois points formant le support d’une application quadratique q1 dont deux sont sur j(L) mais sont différents des points-base de j−1 . Alors l’application q1 ◦ j est une application ayant huit pointsbase dont quatre alignés et de caractéristique (6; 4, 24 , 13 ). Soit C une conique passant par les points p1 , . . . , p4 . Considérons trois points formant le support d’une application quadratique q2 qui sont sur j(C) mais qui ne sont pas des points-base de j−1 . L’application q2 ◦ j est une application ayant huit points-base dont sept sur une conique. PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 48 Enfin, si maintenant nous composons l’application de Jonquières de degré trois (x, y) 7→ (y 3 − x, y) par une application quadratique dont les trois points-base sont dans P2 différents du point [1 : 0 : 0], la composée possède huit points-base dont deux adhérents à un même troisième. Ces trois applications, bien qu’ayant huit points-base, ne sont pas des germes de cellules adjacentes à l’identité d’après le corollaire 4.7. Le corollaire suivant est une conséquence du corollaire 4.7 et du lemme 1.14.2). Corollaire 4.9. Les germes des cellules adjacentes à la cellule V(id) qui ne sont pas des applications de caractéristique Jonquières sont de degré inférieur ou égal à 17. Remarquons que si V(f ) et V(g) sont deux cellules adjacentes, le segment géodésique [f# (ℓ), g# (ℓ)] n’est pas forcément inclus dans l’union des cellules V(f ) ∪ V(g). Exemple 4.10. Considérons une application f dont les points-base de l’inverse sont en position presque générale et qui a pour caractéristique (4, 33 , 13 ). Les classes dans l’intersection de V(f ) et V(id) sont 6-symétriques, ce qui n’est le cas d’aucune classe du segment [f# (ℓ), ℓ]. Par conséquent le segment traverse au moins une autre cellule de Voronoï. Cependant si f est une application de caractéristique Jonquières ou une application symétrique cela n’arrive jamais. C’est l’objet du lemme suivant. Rappelons qu’une application est symétrique si elle est de degré strictement supérieur à 1 et si les multiplicités de ses points-base sont toutes égales. Lemme 4.11. Soit f une application symétrique ou de caractéristique Jonquières. Alors nous avons l’inclusion [ℓ, f# (ℓ)] ⊂ V(id) ∪ V(f ). Démonstration. Si j est une application de Jonquières de degré d, dont l’inverse a pour point-base maximal p0 et {p1 , . . . , p2d−2 } comme petits points-base. Notons s   2d−2 X 1 epi . (d + 1)ℓ − (d − 1)ep0 − cj = 2(d + 1) i=1 Si s est une application symétrique de degré d dont l’inverse possède r points-base {p0 , . . . , pr−1 }, nous posons r   r−1 1X d+1 ℓ− ep . cs = 2 3 i=0 i Les classes cj et cs sont les normalisations des milieux des segments de l’espace de l’espace de Hilbert ambiant reliant respectivement les classes ℓ et j# (ℓ) et ℓ et s# (ℓ). Par conséquent les classes cj et cs appartiennent respectivement aux segments géodésiques [ℓ, j# (ℓ)] et [ℓ, s# (ℓ)]. L’application s−1 est aussi une application symétrique de degré d et d’après le lemme 1.15 la multiplicité de ses points-base est égale à d+1 3 . Par conséquent, elle ne peut pas posséder trois points-base alignés, ni six points-base sur une conique car sinon cela contredirait l’inégalité de Bézout (Proposition 1.18). De même, il ne peut pas y avoir deux points-base adhérents à un même troisième sinon cela contredirait la positivité des excès pour s−1 (Proposition 1.17). Ainsi en utilisant le corollaire 4.7 et le théorème 4.1, cj ∈ V(id) ∩ V(j) PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 49 et cs ∈ V(id) ∩ V(s). Ainsi, cj est à égale distance de ℓ et de j# (ℓ) et cs de ℓ et de s# (ℓ) et comme elles appartiennent respectivement aux segments géodésiques [ℓ, j# (ℓ)] et [ℓ, s# (ℓ)] ce sont les milieux de ces segments. Ainsi, ces segments sont inclus dans la réunion des deux cellules de Voronoï.  5. Cellules quasi-adjacentes à la cellule associée à l’identité Rappelons que le bord à l’infini de E est constitué des classes du bord de H∞ , c’est-à-dire des classes d’auto-intersection nulle et de degré strictement positif, qui sont limites d’une suite de classes vivant dans l’hyperplan {ℓ = 1} et proportionnelles à des classes de E. Nous définissons de même le bord à l’infini d’une cellule de Voronoï V(f ). Ce sont les classes du bord à l’infini de E qui sont limites de classes vivant dans l’hyperplan {ℓ = 1} et qui sont proportionnelles à des classes de V(f ). Deux cellules ayant une classe en commun dans leur bord à l’infini sont dites quasi-adjacentes. Dans cette section, le but est de caractériser les applications dont la cellule de Voronoï est quasi-adjacente à la cellule associée à l’identité. Avant cela, nous étudions les classes candidates à être dans de telles intersections et nous montrons au passage que les cellules adjacentes à V(id) sont quasi-adjacentes à V(id) (Corollaire 5.6). Lemme 5.1. Soit {p0 , . . . , p2d−2 } le support d’une application de Jonquières de degré d et de point-base maximal p0 . Pour tout n strictement positif et pour tout d−1 d n ≤ λ0 < n la classe : cλ0 = nℓ − λ0 ep0 − 2d−2 n − λ0 X epi ∈ Z(P2 ) 2 i=1 est d’auto-intersection positive et est proportionnelle à une classe de E. Démonstration. Pour prouver que pour tout d−1 d n ≤ λ0 < n la classe cλ0 est proportionnelle à une classe de E, nous devons montrer que cette classe satisfait les points 1), 2), 3) et 4) de la définition 2.1. Pour λ0 < n, les multiplicités de la classe cλ0 sont toutes positives et donc cette classe satisfait 2.1.1). Montrons que pour n strictement positif fixé, et pour λ0 ∈ [ d−1 d n ≤ λ0 < n[, cλ0 est d’auto-intersection positive et vérifie les points 2), 3) et 4) de la définition 2.1. • Posons  n − λ 2 0 f (λ0 ) = c2λ0 = n2 − λ20 − (2d − 2). 2 ′ Alors f (λ0 ) = −λ0 (d+1)+n(d−1). Ainsi la fonction f est croissante sur l’intervalle n(d−1) [ n3 , n(d−1) d+1 ] puis décroissante sur l’intervalle [ d+1 , n[. De plus, lim f (λ0 ) = 0. Par λ0 →n d−1 conséquent, pour d−1 d+1 n ≤ d n ≤ λ0 < n, cλ0 est d’auto-intersection positive. • La condition 2) qui est la positivité contre l’anti-canonique correspond à : 3n − λ0 − (d − 1)(n − λ0 ) ≥ 0. Elle est vérifiée sans restriction sur λ0 lorsque d = 2 et sinon λ0 doit satisfaire : n(d − 4) λ0 ≥ . d−2 De plus, pour d > 2 nous avons λ0 ≥ d−1 d n. d−1 d ≥ d−4 d−2 et cette condition est satisfaite pour PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 50 • Nous devons vérifier à quelle condition sur λ0 , les excès en tous les points du support de cλ0 sont positifs (Définition 2.1.3)). D’après la remarque 1.19, il suffit juste de regarder l’excès au-dessus du point p0 . De plus, il y a au plus d − 1 points adhérents à p0 . Nous voulons avoir : n − λ0 . λ0 ≥ (d − 1) 2 d−1 Ceci est satisfait lorsque λ0 ≥ d+1 n, ce qui est le cas lorsque λ0 ≥ d−1 d n. • Nous nous intéressons à présent à l’inégalité de Bézout (Définition 2.1.4)). Soit D une courbe de degré δ passant avec multiplicités {µi }0≤i≤2d−2 en les points {pi }0≤i≤2d−2 . Nous cherchons à savoir pour quelles valeurs de λ0 , le terme suivant est positif :  2d−2    2d−2 2d−2 1 X 1 X n − λ0 X (5.2) µi = n δ − µi + λ0 µi − µ0 . nδ − λ0 µ0 − 2 2 i=1 2 i=1 i=1 Si toutes les multiplicités {µi }1≤i≤2d−2 sont nulles alors le terme (5.2) est égal à nδ − λ0 µ0 > n(δ − µ0 ) ≥ 0 et est positif. Considérons le cas où au moins l’un des µi est strictement positif pour 1 ≤ i ≤ 2d − 2. Par positivité des excès des points par lesquels passe la courbe D il existe alors un point de P2 ou un point libre et adhérent à p0 , noté pik ∈ {pi }1≤i≤2d−2 tel que µik > 0. En considérant la droite passant par les points p0 et pik le théorème de Bézout pour la courbe D implique que δ ≥ µ0 + µik . En particulier, δ est strictement supérieur à µ0 : δ > µ0 . P2d−2 P2d−2 1 ∗ Si δ ≥ 21 i=1 µi et 2 i=1 µi ≥ µ0 alors (5.2) est positif pour tout λ0 positif donc en particulier pour d−1 < n. d n ≤ λ0P P2d−2 2d−2 1 1 ∗ Considérons le cas où δ ≥ 2 i=1 µi et 2 i=1 µi < µ0 . Alors, pour tout λ0 positif le terme (5.2) est supérieur ou égal à :   2d−2 1 X µi ≥ 0. (n − λ0 ) µ0 − 2 i=1 ∗ C’est donc en particulier le cas pour d−1 d n ≤ λ0 < n. P2d−2 P2d−2 1 1 Si δ < 2 i=1 µi alors 2 i=1 µi > µ0 . Par conséquent, pour que (5.2) soit positif, il faut que nous ayons : 1 P2d−2 i=1 µi − δ 2 . (5.3) n > λ0 ≥ nt où t = 1 P2d−2 i=1 µi − µ0 2 Cependant, le facteur t dépend de la courbe D et nous voulons un coefficient λ0 uniforme. En particulier, il ne faudrait pas qu’il existe une suite de courbes {Di } telle que la suite de facteurs {ti } associés converge vers 1. Montrons que nous pouvons majorer t indépendamment de D. Considérons une courbe C du système linéaire associé à l’application de Jonquières. En appliquant le théorème de Bézout entre C et D, nous obtenons : 0 ≤ δd − µ0 (d − 1) − Cela se réécrit : −δ ≤− 2d−2 X µi . i=1 2d−2 d−1 1 X µ0 − µi . d d i=1 (5.4) PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 51 Par conséquent, nous avons : 2d−2  1 1  2d−2 X d−1 1 X µi − µi − δ ≤ − µ0 2 i=1 d i=1 d (5.4) 2 2d−2 d−2 X d−1 µi − µ0 2d i=1 d   2d−2 d−1 1 X µi − µ0 . < d 2 i=1 = d−1 Ainsi, t est majoré par d−1 d et pour tout d n ≤ λ0 < n, λ0 satisfait l’équation (5.3). Finalement, nous avons montré que pour tout λ0 tel que d−1 d n ≤ λ0 < n, la classe cλ0 est d’auto-intersection strictement positive et est proportionnelle à une classe de E comme annoncé.  La proposition suivante fournit des exemples de cellules quasi-adjacentes à la cellule associée à l’identité. Proposition 5.5. 1) Soit j une application de caractéristique Jonquières dont l’inverse a pour point-base maximal p0 ∈ P2 . Alors, nous avons : ℓ − ep0 ∈ ∂∞ V(id) ∩ ∂∞ V(j). 2) Soit f une application dont les r ≤ 8 points-base {p0 , . . . , pr−1 } sont en position presque générale. Alors pour toute complétion en un ensemble de neuf points {p0 , . . . , p8 } en position presque générale, nous avons : 3ℓ − 8 X i=0 epi ∈ ∂∞ V(id) ∩ ∂∞ V(f ). Démonstration. 1) Soit j une application de caractéristique Jonquières de degré d dont l’inverse a pour point-base maximal p0 ∈ P2 et pour petits points-base {pi }1≤i≤2d−2 . Fixons n positif. Soit n3 ≤ λ0 < n et considérons la classe cλ0 = nℓ − λ0 ep0 − 2d−2 n − λ0 X epi ∈ Z(P2 ). 2 i=1 D’après le lemme 5.1, pour tout λ0 tel que d−1 d n ≤ λ0 < n, cette classe est d’autointersection strictement positive et est proportionnelle à une classe de E. Par conséquent, la normalisation η(cλ0 ) de la classe cλ0 introduite 2.3 est d’auto-intersection 1 et appartient à E pour tout d−1 d n ≤ λ0 < n. Le théorème 4.1 implique que η(cλ0 ) c appartient à l’intersection V(id)∩V(j). Par conséquent, la classe nλ0 est proportionnelle à une classe appartenant à l’intersection des cellules V(id) et V(j). En prenant la limite quand λ0 tend vers n nous obtenons : ℓ − ep0 ∈ ∂∞ V(id) ∩ ∂∞ V(j) comme annoncé. 2) Considérons pour tout 0 < ε ≤ 1 la classe cε = 3ℓ − r−1 X i=0 epi − (1 − ε) 8 X i=r epi . PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 52 Son auto-intersection est strictement positive : 9 − r − (1 − ε)2 (9 − r) > 0. Nous pouvons par conséquent considérer sa normalisation η(cε ) qui est d’auto-intersection 1, qui est positive contre l’anti-canonique et dont les multiplicités et le degré sont positifs. Elle vérifie donc les points 1) et 2) de la définition 2.1. Le point 3) est vérifié car pour tout 0 < ε ≤ 1, les points du support de la classe cε sont en position presque générale et par conséquent les excès sont tous positifs. Par le même argument, l’inégalité de Bézout est vérifiée pour les droites et les coniques. Il reste à vérifier que l’inégalité de Bézout est satisfaite pour les courbes D de degré d supérieur ou égal à 3 et passant avec multiplicité µi par les points pi pour 0 ≤ i ≤ 8. Comme les points {p0 , . . . , p7 } sont en position presque générale, d’après la proposition 1.20, la surface obtenue en éclatant ces points est faiblement del Pezzo. Avec le même argument que dans la preuve du corollaire 4.7, le théorème de Riemann-Roch implique que la forme anti-canonique sur cette surface correspond à un pinceau de cubiques. Par conséquent, il existe une cubique C de ce pinceau passant par le point p8 . D’après l’inégalité de Bézout appliquée aux courbes C et D, nous avons : 8 r−1 8 X X X µi . µi + (1 − ε) µi ≥ 3d ≥ i=0 i=0 i=r Cela montre que la classe cε satisfait le point 2.1.4) et achève donc de prouver que cε est proportionnelle à une classe de E. En utilisant le théorème 4.1, nous obtenons finalement que pour tout 0 < ε ≤ 1, la classe cε est proportionnelle à une classe vivant dans l’intersection V(id) ∩ V(f ). En passant à la limite quand ε tend vers 0, nous avons : 8 X epi ∈ ∂∞ V(id) ∩ ∂∞ V(f ) 3ℓ − i=0 qui est le résultat annoncé.  Corollaire 5.6. Une cellule adjacente à V(id) est quasi-adjacente à V(id). Démonstration. D’après le corollaire 4.7, il n’y a que deux sortes de cellules adjacentes à V(id), celles dont le germe est de caractéristique Jonquières et celles dont le germe est une application dont les r ≤ 8 points-base de l’inverse sont en position presque générale. Par la proposition 5.5, elles sont toutes quasi-adjacentes à V(id).  Remarque 5.7. définition du bord à l’infini ∂∞ E, nous obtenons que les classes PPar 8 de type 3ℓ − i=0 epi appartiennent à E si et seulement si l’ensemble {p0 , . . . , p8 } est en position presque générale. Lemme 5.8. Soit c ∈ ∂∞ V(id) ∩ ∂∞ V(f ). Si c n’est pas une classe spéciale alors c est une classe 1-symétrique pure ou 9-symétrique pure. Sinon c est une classe 1-symétrique. P Démonstration. Supposons c = nℓ− i∈I λi epi ordonnée. Si c est une classe spéciale alors les points p1 et p2 sont adhérents au point p0 . Par positivité des excès en p0 , la multiplicité λ1 doit être strictement plus petite que p0 . Par conséquent c est une classe 1-symétrique. Considérons maintenant le cas où la classe c n’est pas spéciale. Par définition du bord à l’infini, c est limite d’une suite de classes proportionnelles à des classes de V(id) ∩ V(f ). Quitte à prendre une sous-suite nous pouvons supposer que les classes PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 53 de cette suite sont soit toutes spéciales soit toutes non spéciales. Si ces classes sont spéciales le degré de chacune des classes de cette suite est strictement inférieur à la somme de ses trois plus grandes multiplicités. Or c n’est pas une classe spéciale donc à la limite c satisfait n = λ0 + λ1 + λ2 . Si toutes les classes de cette suite ne sont pas spéciales alors le degré de chacune est égal à la somme de ses trois plus grandes multiplicités (Théorème 4.1). Ainsi, à la limite la classe c possède la même propriété : n = λ0 + λ1 + λ2 . (5.9) Si la classe c est k-symétrique avec k ≥ 3, alors ses multiplicités maximales sont égales au tiers du degré. Nous sommes dans les conditions du lemme 2.14 et c est 9-symétrique pure. Supposons à présent que c est k-symétrique pour k ≤ 2. Comme c2 = 0, nous avons X n2 = λ2i . (5.10) i∈I Par conséquent, si λ2 = 0 il en est de même de λ1 et donc la classe c est 1-symétrique pure. Montrons que λ2 est forcément nul. Supposons le contraire. Nous avons :   X  X  λi = λ2 2n − λ2 3n − λi i≥0 (5.9) i≥3 = −n2 + λ20 + λ21 + λ22 + 2nλ2 + X (5.10) = −(n − λ2 )2 + λ20 + λ21 + 2λ22 + X i≥3 = −(λ0 + λ1 )2 + λ20 + λ21 + 2λ22 + (5.9) = −2λ0 λ1 + 2λ22 + X i≥3 λi (λi − λ2 ) i≥3 λi (λi − λ2 ) X i≥3 (5.11) λi (λi − λ2 ) λi (λi − λ2 ) Comme la classe c est k-symétrique pour k ≤ 2, λ2 < λ0 . De plus, comme les multiplicités sont décroissantes le dernier terme de (5.11) est strictement négatif. P λi puisque λ2 est strictement positif. Ceci implique qu’il en est de même de 3n − i≥0 Mais cela contredit le fait qu’une classe au bord est positive contre l’anti-canonique.  Remarquons qu’il n’existe peut-être pas de classe spéciale au bord à l’infini de V(id). Nous n’avons ni réussi à en construire une ni réussi à montrer qu’il n’en existait pas. Proposition 5.12. L’ensemble des classes non spéciales au bord à l’infini de V(id) et qui sont également dans le bord à l’infini d’une cellule quasi-adjacente à V(id) est constitué des classes suivantes : 1) ℓ − ep0 où p0 ∈ P2 , 8 P epi où les points de l’ensemble {p0 , . . . , p8 } sont en position presque 2) 3ℓ − générale. i=0 Démonstration. D’après la proposition 5.5, les cas 1) et 2) sont dans l’intersection entre le bord à l’infini de V(id) et celui d’une cellule adjacente. Le lemme 5.8 nous PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 54 dit que les seules classes non spéciales vérifiant l’hypothèse de la proposition sont 1-symétrique pure ou 9-symétrique pure. Une classe 9-symétrique pure doit avoir comme support un ensemble de 9 points en position presque générale car sinon cela  contredirait le fait que cette classe appartient à ∂∞ E (Remarque 5.7). Proposition 5.13. Soient p0 ∈ P2 et f ∈ Bir(P2 ). La classe ℓ − ep0 appartient à l’intersection ∂∞ V(id) ∩ ∂∞ V(f ) si et seulement si f est une application de caractéristique Jonquières dont l’inverse a pour point-base maximal p0 . Démonstration. L’implication indirecte découle de la proposition 5.5.1). Soit f ∈ Bir(P2 ) telle que ℓ − ep0 soit dans l’intersection ∂∞ V(id) ∩ ∂∞ V(f ). Si la cellule V(f ) est une cellule adjacente à la cellule associée à l’identité alors pour toute classe c ∈ V(id) ∩ V(f ) nous avons par convexité des cellules de Voronoï l’inclusion suivante : [c, ℓ − ep0 [ ⊂ V(id) ∩ V(f ). D’après le théorème 4.1, si f n’est pas une application de caractéristique Jonquières les classes de V(id) ∩ V(f ) sont au moins r-symétriques où r ≥ 6 est le nombre de points-base de f . Par conséquent, en notant p0 , . . . , pr les points-base de f −1 toute classe multiple d’une classe du segment géodésique [c, ℓ−ep0 [ vit dans le sous-espace vectoriel de l’espace de Picard-Manin obtenu comme intersection des hyperplans {λp0 = λp1 }, {λp0 = λp2 }, . . ., {λp0 = λpr−1 }. Or, la classe ℓ − ep0 n’appartient pas au bord à l’infini de ce sous-espace ce qui mène à une contradiction. Avec le même type d’argument nous obtenons une contradiction si nous supposons que f est une application de caractéristique Jonquières dont le point-base maximal de f −1 n’est pas le point p0 . Montrons que f ne peut pas être le germe d’une cellule quasi-adjacente non adjacente à V(id). Raisonnons par l’absurde et supposons le contraire. L’espace E est convexe et les cellules de Voronoï recouvrent E. De plus, les cellules adjacentes à la cellule V(id) et contenant la classe ℓ − ep0 sont des applications de caractéristiques Jonquières dont le point-base maximal de l’inverse est p0 . Nous pouvons par conséquent supposer que la cellule V(f ) est adjacente à une cellule V(j) où j est un représentant qui est une application de Jonquières et dont l’inverse a pour point-base maximal p0 . Faisons agir à présent l’application j−1 . Elle fixe la classe à l’infini ℓ − ep0 ainsi la cellule V(j−1 ◦ f ) est une cellule adjacente de la cellule identité et elle possède la classe ℓ − ep0 dans son bord à l’infini. Par le point précédent, nous obtenons que j−1 ◦ f est une application de caractéristique Jonquières dont l’inverse a pour point-base maximal p0 . C’est donc aussi le cas de f ce qui mène à une contradiction. Ainsi, V(f ) est forcément adjacente à V(id) et f est de caractéristique Jonquières et son inverse possède le point p0 comme point-base maximal.  Remarque 5.14. S’il existe une classe spéciale c ∈ V(id) ∩ V(f ) et si V(f ) est adjacente à V(id) alors f −1 est une application de caractéristique Jonquières dont le point-base maximal est le point de multiplicité maximale de c. Cela découle du théorème 4.1 et du fait qu’il n’existe pas une suite de classes non spéciales de V(id) ∩ V(f ) convergeant vers c puisque leur degré est égal à la somme de leurs trois plus grandes multiplicités contrairement à c. Par le même argument que dans la preuve de la proposition 5.13, ce sont les seules applications dont les cellules de Voronoï contiennent cette classe. PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 55 Lemme 5.15. Soient {p0 , . . . , p8 } un ensemble de points en position presque généP rale et c = 3ℓ − 8i=0 epi ∈ E. Soit f ∈ Bir(P2 ). Si le support de f −1 est inclus dans le support de c, alors le support de f −1 (c) contient 9 points en position presque générale et X X f#−1 (ep ). eq − f#−1 (c) = 3ℓ − q∈supp(f ) p∈supp(c) −1 p∈supp(f / ) Démonstration. Notons {mq }q∈supp(f ) (respectivement {m′p }p∈supp(f −1 ) ) les multiplicités des points-base de f (respectivement de f −1 ) et {ap,q }q∈supp(f ) les autres coefficients de la matrice caractéristique. Nous avons d’après l’équation (1.10) :    X X X  3mq − ap,q eq f#−1 (c) = 3d − m′p ℓ − q∈supp(f ) p∈supp(f −1 ) p∈supp(f −1 ) − X f#−1 (ep ). p∈supp(c) −1 p∈supp(f / ) Or d’après les égalités 1) et 5) du lemme 1.13, nous avons :  P m′p = 3  3d − p∈supp(f −1 ) P  aq,p = 1, pour tout q ∈ supp(f ) 3mq − p∈supp(f −1 ) ce qui implique que f#−1 (c) = 3ℓ − X q∈supp(f ) eq − X f#−1 (ep ) p∈supp(c) −1 p∈supp(f / ) comme attendu. Comme le groupe de Cremona préserve E et ∂∞ E, la classe f#−1 (c) appartient à ∂∞ E. D’après la remarque 5.7, cela signifie que le support de cette classe est un ensemble de points en position presque générale.  Proposition 5.16. Soit {p0 , . . . , p8 } un ensemble de points en position presque générale. L’application f satisfait 3ℓ − 8 X i=0 epi ∈ ∂∞ V(id) ∩ ∂∞ V(f ) si et seulement si f possède au plus neuf points-base et le support de son inverse est inclus dans {p0 , . . . , p8 }. P8 Démonstration. Posons c = 3ℓ − i=0 epi . D’après la proposition 5.12 elle appartient au bord à l’infini de V(id). Montrons la réciproque. Si le support de f −1 est inclus dans le support de c alors le lemme 5.15 implique que la classe f#−1 (c) est une classe 9-symétrique pure dont le support est constitué de 9 points en position presque générale. Par la proposition 5.12, nous obtenons que la classe f#−1 (c) appartient au bord à l’infini de V(id). Par conséquent, c appartient au bord à l’infini de V(f ) et nous pouvons en conclure que comme attendu. c ∈ ∂∞ V(id) ∩ ∂∞ V(f ), PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 56 Montrons l’implication directe, elle se prouve P8comme l’implication de la proposition 5.13. Soit f ∈ Bir(P2 ) tel que c = 3ℓ − i=0 epi ∈ ∂∞ V(id) ∩ ∂∞ V(f ). Si f est le germe d’une cellule adjacente à la cellule V(id), par convexité des cellules de Voronoï nous obtenons pour tout x ∈ V(id) ∩ V(f ) : [x, c[ ⊂ V(id) ∩ V(f ). D’après le théorème 4.1, toute classe multiple d’une classe de [x, c[ vit dans le sousespace de Picard-Manin obtenu comme intersection d’hyperplans déterminés par les points-base de f −1 . Par conséquent c vit dans le bord à l’infini de l’intersection entre ce sous-espace et E. Ceci implique que l’ensemble des points-base de f −1 est inclus dans l’ensemble {p0 , . . . , p8 }. Supposons maintenant que f n’est pas la germe d’une cellule adjacente à l’identité. Comme le pavage de Voronoï recouvre E qui est de plus convexe, nous pouvons supposer qu’il existe g ∈ Bir(P2 ) telle que V(g) soit adjacente à V(id) et à V(f ) et que V(g) possède la classe c dans son bord à l’infini. Par le point précédent, le support de g −1 est inclus dans le support de c. Faisons agir g −1 . D’après le lemme 5.15, X X −1 −1 g# (ep ) eq − g# (c) = 3ℓ − q∈supp(g) p∈supp(c) −1 p∈supp(g / ) et son support est constitué de 9 points en position presque générale. D’après ce que nous avons montré dans le paragraphe précédent, les points-base de l’inverse −1 de g −1 ◦ f sont contenus dans le support de g# (c). Les points-base de f −1 ◦ g sont inclus dans l’union de l’ensemble des points-base de g et de l’ensemble des points images par g −1 des points-base de f −1 qui ne sont pas des points-base de g −1 . Par conséquent les points base de f −1 sont inclus dans le support de c comme attendu.  Le corollaire suivant se déduit de la remarque 5.14 et des propositions 5.12, 5.13 et 5.16. Corollaire 5.17. L’ensemble des germes des cellules quasi-adjacentes à V(id) est constitué de : • toutes les applications de caractéristique Jonquières, • de toutes les applications ayant au plus 9 points-base et dont les points-base de l’inverse sont en position presque générale. Corollaire 5.18. Soient f, g ∈ Bir(P2 ) telles que les cellules V(f ) et V(g) soient quasi-adjacentes à la cellule associée à l’identité. Notons p0 , . . . , pk−1 l’union des points-base de f −1 et de g −1 . Les trois cellules V(id), V(f ) et V(g) ont une classe commune dans leur bord à l’infini si et seulement si nous sommes dans l’un des deux cas suivants : • f −1 et g −1 sont deux applications de caractéristique Jonquières ayant le même point-base maximal, ou un même point-base maximal dans le cas où l’une est quadratique, • les points {p0 , . . . , pk−1 } sont en position presque générale et k ≤ 9. Démonstration. Supposons que les cellules V(id), V(f ) et V(g) possèdent une classe c commune au bord à l’infini et que f −1 et g −1 ne soient pas des applications de caractéristique Jonquières ayant le même point-base maximal. Si la classe c PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 57 est spéciale alors par la remarque 5.14, les applications f et g doivent être des applications de Jonquières de même point-base ce qui contredit notre hypothèse. La classe c n’est donc pas spéciale. Ainsi, la proposition 5.12 nous dit qu’il n’y a que deux possibilités pour cette classe. Les propositions 5.13 et 5.16 impliquent que c est une classe 9-symétrique pure dont le support contient les supports de f −1 et de g −1 : c = 3ℓ − 8 X i=0 epi ∈ ∂∞ V(f ) ∩ ∂∞ V(g) ∩ ∂∞ V(id). Ainsi, k ≤ 9 et, d’après la remarque 5.7, les points du support de c (et par conséquent {p0 , . . . , pk−1 }) sont en position presque générale. Montrons la réciproque. Si f −1 et g −1 sont des applications de caractéristique Jonquières et de même point-base maximal p0 alors la classe ℓ − ep0 appartient au bord à l’infini des trois cellules par la proposition 5.5.1). Considérons le cas où f −1 et g −1 ne sont pas des applications de caractéristique Jonquières ayant même pointbase maximal. Notons {p0 , . . . , pk−1 } l’union des points-base de f −1 et de g −1 , et supposons que cet ensemble soit en position presque générale et k ≤ 9. Alors nous pouvons compléter cette famille de sorte que {p0 , . . . , p8 } soit un ensemble de points 8 P en position presque générale. D’après la remarque 5.7, la classe 3ℓ− epi appartient i=0 au bord à l’infini de V(id), et d’après la proposition 5.16 nous obtenons qu’elle appartient également au bord à l’infini de V(g) et de V(f ) comme annoncé.  Références M. Alberich-Carramiñana – Geometry of the plane Cremona maps, Lecture Notes in Mathematics, vol. 1769, Springer-Verlag, Berlin, 2002. [Ale16] J. W. Alexander – « On the factorization of Cremona plane transformations », Trans. Amer. Math. Soc. 17 (1916). [BC16] J. Blanc & S. Cantat – « Dynamical degrees of birational transformations of projective surfaces », J. Amer. Math. Soc. 29 (2016), no. 2, p. 415–471. [BFJ08] S. Boucksom, C. Favre & M. Jonsson – « Degree growth of meromorphic surface maps », Duke Math. J. 141 (2008), no. 3, p. 519–538. [Can11] S. Cantat – « Sur les groupes de transformations birationnelles des surfaces », Annals of Mathematics. Second Series 174 (2011), p. 299–340. , « The Cremona groups », À paraître dans Proceedings of 2015 Summer Institute [Can17] on Algebraic Geometry, AMS Proceedings of Symposia in Pure Mathematics (2017). [CL13] S. Cantat & S. Lamy – « Normal subgroups in the Cremona group », Acta Math. 210 (2013), p. 31–94. [DGO17] F. Dahmani, V. Guirardel & D. Osin – « Hyperbolically embedded subgroups and rotating families in groups acting on hyperbolic spaces », Mem. Amer. Math. Soc. 245 (2017), no. 1156, p. v+152. [Dol12] I. V. Dolgachev – Classical algebraic geometry, Cambridge University Press, Cambridge, 2012, A modern view. [Fav10] C. Favre – « Le groupe de Cremona et ses sous-groupes de type fini », Astérisque (2010), no. 332, p. Exp. No. 998, vii, 11–43, Séminaire Bourbaki. Volume 2008/2009. Exposés 997–1011. [Lon16] A. Lonjou – « Non simplicité du groupe de Cremona sur tout corps », Ann. Inst. Fourier 66 (2016), no. 5, p. 2021–2046. [Lon17] , Groupe de Cremona et espaces hyperboliques, Thèse, 2017. [Sha94] I. R. Shafarevich – Basic algebraic geometry. 1, second éd., Springer-Verlag, Berlin, 1994, Varieties in projective space, Translated from the 1988 Russian edition and with notes by Miles Reid. [AC02] PAVAGE DE VORONOÏ ASSOCIÉ AU GROUPE DE CREMONA 58 Universität Basel, Departement Mathematik und Informatik, Spiegelgasse 1„ 4051 Basel, Switzerland E-mail address: [email protected]
4math.GR
arXiv:1402.3488v2 [cs.DS] 17 Sep 2015 A Unifying Model for Representing Time-Varying Graphs Klaus Wehmuth National Laboratory for Scientific Computing (LNCC) Av. Getúlio Vargas, 333 25651-075 – Petrópolis, RJ – Brazil [email protected] Eric Fleury ENS de Lyon / INRIA – Université de Lyon UMR CNRS – ENS Lyon – UCB Lyon 1 – INRIA 5668 46, allée d’Italie 69364 – Lyon Cedex 07, France [email protected] Artur Ziviani National Laboratory for Scientific Computing (LNCC) Av. Getúlio Vargas, 333 25651-075 – Petrópolis, RJ – Brazil [email protected] Abstract Graph-based models form a fundamental aspect of data representation in Data Sciences and play a key role in modeling complex networked systems. In particular, recently there is an ever-increasing interest in modeling dynamic complex networks, i.e. networks in which the topological structure (nodes and edges) may vary over time. In this context, we propose a novel model for representing finite discrete Time-Varying Graphs (TVGs), which are typically used to model dynamic complex networked systems. We analyze the data structures built from our proposed model and demonstrate that, 1 for most practical cases, the asymptotic memory complexity of our model is in the order of the cardinality of the set of edges. Further, we show that our proposal is an unifying model that can represent several previous (classes of) models for dynamic networks found in the recent literature, which in general are unable to represent each other. In contrast to previous models, our proposal is also able to intrinsically model cyclic (i.e. periodic) behavior in dynamic networks. These representation capabilities attest the expressive power of our proposed unifying model for TVGs. We thus believe our unifying model for TVGs is a step forward in the theoretical foundations for data analysis of complex networked systems. 1 Introduction Data representation is a fundamental aspect in the field of Data Science [1, 2, 3]. In this context, graph models for representing complex networked systems find broad applicability in several different areas, ranging from techno-social systems [4] to computational systems biology [5]. In a typical graph model, nodes denote objects of the domain of interest (e.g., individuals or genes), and edges express interactions between these objects. Characteristics of objects and their interactions can be represented as node and edge properties, in a similar way to the attributes of relations. Indeed, advanced graph analytics is at the core of the new field of Network Science [6, 7, 8]. Much of the utility of the graph abstraction, however, actually resides in the fact that it can represent relations between a set of objects as well as their connectivity properties, which derive from the notion of paths in a straightforward way without the need of further assumptions that are not explicit in the graph abstraction itself. Based upon graph representations, there is a lot of studies focused on investigating dynamic processes, such as random walks or information diffusion, over complex networks represented by graphs [9, 10, 11, 12]. More recently, there is an ever-increasing interest in investigating not only the process dynamics on networks, but also the dynamics of networks, i.e. when the network structure (nodes and edges) may vary over time [13, 14, 15, 16]. Indeed, Braha and Bar-Yam [17] indicate that new insights can be obtained from the dynamical behavior, including a dramatic time dependence of the role of nodes that is not apparent from static (time aggregated) analysis of node connectivity and network topology. Analyzing the dynamics of networks, however, brings a difficulty since the original graph abstraction was not created considering time 2 relations between nodes. As a consequence, the need to extend the basic graph abstraction to include time relations between nodes arose, leading to many models for representing Time-Varying Graphs (TVGs), also known in the literature as temporal or time-dependent networks [18, 20, 19, 21, 22, 23, 24, 25, 26]. Although recent TVG models appear extending the basic graph concept to include time relations (Section 4 discusses related work), they are nonetheless not general enough to satisfy the needs of different complex networked systems and also in many cases rely on assumptions that are not explicitly expressed in the model. For instance, in models based on snapshots (i.e., a series of static graphs), as those in [20, 19], it is implicitly assumed that a node in a given snapshot is connected to ‘itself’ in the next snapshot, making it possible to extend the transitivity of edges over time. This assumption, however, is not made explicit in the snapshot model. Therefore, when analyzed without this implicit assumption, a snapshot model is a sequence of disconnected graphs and therefore no connectivity is possible between different time instants. The need to handle this assumption, which is not explicitly part of the model, brings difficulties since the structure of the model by itself is no longer sufficient to properly represent its behavior, making the semantics, applicability, interpretation, and analysis of such models more complex. In this paper, we propose a new unifying model for representing finite discrete TVGs. Our proposed model is sufficiently general to capture the needs of distinct dynamic networks [20, 19, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31], whereas not requiring any further assumption that is not already explicitly contained in the model itself. Further, our model aims at preserving the strictly discrete nature of the basic graph abstraction, while also allowing to properly represent time relations between nodes. The model we propose is based upon our recent work [32, 33], which shows that structures such as multilayer and timevarying networks are in fact isomorphic to directed static graphs. Our proposal in this paper, however, is specifically tailored for use with TVGs. Furthermore, we also demonstrate the unifying properties of our proposed model for representing TVGs by describing how it represents several previous (classes of) models for dynamic networks found in the recent literature, which in general are unable to represent each other. In contrast to previous TVG models, our proposal is able to intrinsically model cyclic (i.e. periodic) behavior in complex dynamic networks. These representation features attest the expressive power of our proposed unifying model for TVGs. Finally, we also demonstrate that, for most practical cases, the asymptotic memory complexity of our TVG model is determined by the cardinality of the set of edges in the model. We thus believe our unifying model for 3 TVGs is a step forward in the theoretical foundations for data analysis of complex networked systems.1 This paper proceeds as follows. Section 2 introduces our proposed unifying model for representing TVGs and its main properties. Section 3 discusses data structures to properly represent TVGs using our model. In Section 4, we show how our unifying model can be used to represent previous models for dynamic networks while these models in general are unable to represent each other. Finally, we conclude in Section 5. 2 Proposed model for representing TVGs The TVG model we introduce in this paper is a particular case of a MultiAspect Graph (MAG) [32], in which the vertices and time instants are the key features (i.e., aspects) to be represented by the model. A MAG is a structure capable of representing multilayer and time-varying networks, while also having the property of being isomorphic to a directed graph. The MAG structural form is similar to the multilayer structure recently presented by [35], since in both cases the proposed structure has a construction similar to an even uniform hypergraph associated with an adjacency concept similar to the one of simple directed graphs. Formally, a MAG can be defined as an object H “ pA, Eq, where E is a set of edges and A is a finite list of sets, each of which is called an aspect. In our case, for modeling a TVG, we have two aspects, namely vertices and time instants, i.e. |A| “ 2. For the sake of simplicity, this 2-aspect MAG can be regarded as representing a TVG with an object H “ pV, E, T q, where V is the set of nodes, T is the set of time instants, and E Ď V ˆ T ˆV ˆ T is the set of edges. As a matter of notation, we denote V pHq as the set of all nodes in H, EpHq the set of all edges in H, and T pHq the set of all time instants in H. An edge e P EpHq is defined as an ordered quadruple e “ pu,ta , v,tb q, where u, v P V pHq are the origin and destination nodes, respectively, while ta ,tb P T pHq are the origin and destination time instants, respectively. Therefore, e “ pu,ta , v,tb q should be understood as a directed edge from node u at time ta to node v at time tb . If one needs to represent an undirected edge in the TVG, both pu,ta , v,tb q and pv,tb , u,ta q should be in EpHq. We also define four canonical projections, each projection mapping a dynamic edge into each one of its components: 1A preliminary version of this paper appears as a technical report [34]. 4 π1 : EpHq Ñ V pHq pu,ta , v,tb q ÞÑ u, π2 : EpHq Ñ T pHq pu,ta , v,tb q ÞÑ ta , π3 : EpHq Ñ V pHq pu,ta , v,tb q ÞÑ v, π4 : EpHq Ñ T pHq pu,ta , v,tb q ÞÑ tb . An edge e “ pu,ta , v,tb q in our model may be classified into four classes depending on its temporal characteristic: 1. Spatial edges connect two nodes at the same time instant, e is in the form of e “ pu,ta , v,ta q, where u ‰ v; 2. Temporal edges connect the same node at two distinct time instants, e is in the form of e “ pu,ta , u,tb q, where ta ‰ tb ; 3. Mixed edges connect distinct nodes at distinct time instants, e is in the form of e “ pu,ta , v,tb q, where u ‰ v and ta ‰ tb ; 4. Spatial-temporal self-loop edges connect the same node at the same time instant, e is in the form of e “ pu,ta , u,ta q. The definition of a TVG H is as general as possible and does not impose any order on the time set T pHq. One may, however, stick to the classical time notion and impose a total order on T pHq. In this context, where T pHq has a linear order, both mixed or temporal dynamic edges e “ pu,ta , v,tb q can also be classified as progressive or regressive depending on the order of their temporal components. Dynamic edges that are originated at an earlier time instant and destined to a later time instant are progressive (ta ă tb ), whereas dynamic edges originated at a later time instant and destined to an earlier time instant are regressive (ta ą tb ). 5 Regressive edges are particularly useful for creating cyclic TVGs, which in turn can be applied to model networks with a cyclic periodic behavior. Further, we define a temporal node as an ordered pair pu,ta q, where u P V pHq and ta P T pHq. The set V T pHq of all temporal nodes in a TVG H is given by the cartesian product of the set of nodes and the set of time instants, i.e. V T pHq “ V pHq ˆ T pHq. As a notation note, a temporal node is represented by the ordered pair that defines it, e.g. pu,ta q. The usage of the object H “ pV, E, T q to represent a TVG is formally described in the technical report [34]. Therein, the representation of the TVG based on temporal nodes is proven to be isomorphic to a directed static graph. This is an important theoretical result since this allows the use of the isomorphic directed graph as a tool to analyze both the properties of a TVG and the behavior of dynamic processes over a TVG, as done is this work. This model is then shown to unify the representation of several previous (classes of) models for TVGs of the recent literature, which in general are unable to represent each other, as we show in further detail in Section 4. 3 Algebraic representations and structures In this section, we discuss ways to properly represent a TVG using our proposed model. Similarly to static graphs, a TVG can be fully represented by an algebraic structure, like the MAG structure from which our TVG model is derived [33]. We thus first present a TVG algebraic representation based on the adjacency tensor in Section 3.1. In this work, we adopt matrix-based representations, in particular the adjacency matrix (Section 3.2). We also present the TVG incidence matrix in Section 3.3. In order to illustrate such representations, we use the TVG W presented in Figure 1, where spatial edges are represented by solid arrows and temporal edges by dashed arrows. Finally, in Section 3.4, we analyze the memory complexity for storing a TVG using our model. 3.1 TVG adjacency tensor The TVG adjacency tensor follows from the adjacency matrix widely used to represent static graphs. However, since the dynamic edges used in the TVG are ordered quadruples, this representation has as a consequence to be done by means of a 4th order tensor. 6 t t 0 t 1 2 0 1 0 1 0 1 2 3 2 3 2 3 Figure 1: TVG W : Example for the algebraic representation. We then define the adjacency tensor of a TVG H as a 4th order tensor ApHq with dimension |V pHq| ˆ |T pHq| ˆ |V pHq| ˆ |T pHq| that has an entry for every possible dynamic edge in H. Each dynamic edge present in the TVG H is represented by a non-zero entry in the adjacency tensor ApHq, while all other entries have a zero value. The non-zero entries represent the weight of the corresponding dynamic edge in the represented TVG. In the case of a unweighted TVG, the nonzero entries corresponding to the dynamic edges present in the TVG have value 1. u,ta The notation ApHqv,t is used to identify the entry corresponding to the dynamic b edge pu,ta , v,tb q of the TVG H. As an example, the adjacency tensor of the TVG W depicted in Figure 1 has dimension 4 ˆ 3 ˆ 4 ˆ 3, having a total of 144 entries. For instance, the pair of dynamic spatial edges connecting node 0 at time t0 to node 3 at time t0 is 3,t0 0 represented by the entries ApW q0,t 3,t0 and ApW q0,t0 , where both carry value 1 as the TVG W is unweighted. Note that, even though the TVG adjacency tensor ApHq has dimension |V pHt q|ˆ |T pHt q| ˆ |V pHt q| ˆ |T pHt q|, only the entries corresponding to dynamic edges present in the TVG H have non-zero values. 3.2 TVG adjacency matrix Since every MAG has a directed static graph that is isomorphic to it [32], the same holds for our TVG model, since it is a particular specialized case of a MAG. Consequently, it follows that the TVG can be represented by an adjacency matrix. In the more general environment represented by a MAG, a companion tuple 7 is used in order to properly identify and position each temporal node of the isomorphic graph in the adjacency matrix. Since the case we present in this work is restricted to MAGs with 2 aspects, it follows that the companion tuple is reduced to a pair, which in the first entry has the number of nodes and the second entry has the number of time instants. For instance, considering the TVG example of Figure 1, the companion tuple associated with its adjacency matrix is p4, 3q, since there are 4 nodes and 3 time instants. The function of the companion tuple is only to ensure that the order by which the temporal nodes are placed in the adjacency matrix is the one shown in Figure 2. Since in the case where the number of aspects is restricted to 2 this placement can be easily achieved, in this work we do not further mention the companion tuple. To get the TVG adjacency matrix, we only need to consider that each temporal node pu,ta q can be thought of as a node in a static graph. This static graph has |V | ˆ |T | nodes and, as a consequence, its adjacency matrix has |V | ˆ |T | ˆ |V | ˆ |T | “ |V |2 ˆ |T |2 entries. Since the non-zero entries of this matrix correspond to the dynamic edges of the TVG, further analysis show that this matrix is usually sparse and can therefore be stored in an efficient way. Figure 2: The matrix form of the adjacency tensor of TVG W . Figure 2 shows the matrix representation obtained for the illustrative TVG W 8 shown in Figure 1. From Figure 2, we highlight that the matrix form of the TVG adjacency tensor has interesting structural properties. First, each one of the four nodes (identified as 0, 1, 2, and 3) of the TVG W clearly appears as a separate entity in each of the three time instants (t0 , t1 , and t2 ) that compose the TVG W . Second, the main block diagonal (lightly shaded) contains the entries corresponding to the spatial edges at each time instant. In these three blocks the entries corresponding to the spatial edges of the TVG carry value 1. Third, the unshaded entries at the off-diagonal blocks correspond to the temporal edges. The eight progressive temporal edges present at the TVG W are indicated by the value 1 on the first superior diagonal. Finally, the dark shaded entries are the ones that correspond to the mixed edges. Since no mixed edges are present in the example TVG W , all these entries contain value 0. Further, we remark that the entries corresponding to progressive (mixed and temporal) edges are above the main block diagonal, whereas the edges corresponding to regressive edges appear below the main block diagonal. All these structural properties derive from the order adopted for representing the nodes and time instants present in the TVG and can be readily verified in the matrix form in a quite convenient way. Note that the procedure used to obtain the matrix form of the adjacency tensor is the well-known matricization or unfolding of a tensor [36]. In fact, since each entry of the matrix corresponds to an entry in the tensor, the process can be reversed, obtaining the corresponding tensor from its matrix form. 3.3 TVG incidence matrix The incidence matrix of a TVG is the incidence matrix corresponding to the directed static graph obtained though the temporal node representation presented in more detail in [34]. For instance, considering the TVG W shown in Figure 1, there is a temporal edge connecting node 0 at time t0 to node 0 at time t1 . Without loss of generality, we can label this edge as e0 and assign it the coordinate 0 for the edge dimension of the incidence tensor. As a consequence, this edge is represented in the incidence tensor by setting C00,0 “ ´1 and C10,0 “ 1. Taking the temporal edge connecting node 0 at time t1 to node 0 at time t2 and labeling it as e1 , its representation on the incidence tensor is C10,1 “ ´1 and C20,1 “ 1. By repeating this procedure for each edge in the TVG, the corresponding incidence tensor is created. Since the TVG W has 16 edges, being 8 temporal edges and 8 spatial edges, 4 nodes, and 3 time instants, the corresponding incidence tensor has a total of 4 ˆ 16 ˆ 3 “ 192 entries, from which 160 carry value 0, 16 carry value ´1, and 16 value 1. Note that the resulting matrix has 12 rows (for 3 times and 4 nodes) 9 and 16 columns, one for each edge in the TVG, totalizing 12 ˆ 16 “ 192 entries. Figure 3 shows the matrix representation of the incidence tensor corresponding to the TVG W shown in Figure 1. In this representation, the eight temporal edges present in the TVG are labeled as edges e0 to e7 and the 8 spatial edges as edges e8 to e15 . Considering this labeling, it is straightforward to verify the correspondence of the incidence tensor in matrix form to the TVG. Figure 3: The matrix form of the incidence tensor of TVG W . 3.4 Memory complexity for storing a TVG In this subsection, we analyze the memory complexity for storing a TVG. This analysis is valid for the TVG representations based on adjacency and incidence as well as their corresponding algebraic structures, which have all been discussed in Sections 3.1 to 3.3. In the general case, TVGs may have disconnected nodes or unused time instants. A disconnected node is defined as a node that has no dynamic edge incident to it, i.e. a node with no connection at any time instant. In a similar way, an unused time instant in a TVG is defined as a time instant at which there is no dynamic edge originated from or destined to it. In most practical cases, however, a TVG is expected to have none or very few disconnected nodes as well as a unused time instants as compared with the total number of nodes or time instants, respectively. We show that when a TVG has no disconnected nodes and unused time instants, or if the number of disconnected nodes and unused time instants is significantly lower than the number of dynamic edges, then the amount of storage 10 needed to represent a TVG is determined by the number of dynamic edges in the TVG. In other words, we show in this section that in most practical cases the memory complexity MC pHq for storing a given TVG H is MC pHq “ Θp|EpHq|q, where EpHq is the set of dynamic edges in the TVG H. The first step to achieve this is to show that the set of connected nodes and the set of used time instants of a TVG can be recovered from the set of dynamic edges. Lemma 1 shows this by using the dynamic edge definition and the canonical projections defined in Section 2. Lemma 1. In any given TVG H, the set of connected nodes and the set of used time instants can be recovered from the set EpHq that contains the dynamic edges of H. Proof. Let VC pHq be the set of connected nodes on TVG H. We now show that VC pHq can be constructed from the set EpHq. Since VC pHq contains only connected nodes, for every u P VC pHq there is at least one dynamic edge incident to u. Let eu P EpHq be a dynamic edge incident to node u. Then, either eu is of the form pu, ¨, ¨, ¨q or of the form p¨, ¨, u, ¨q, and therefore, either u “ π1 peu q or u “ π3 peu q. Therefore, we can write VC pHq as ď tπ1 peq, π3 pequ . VC pHq “ ePEpHq We now use a similar reasoning to recover the set of used time instants from the set EpHq. Let TU pHq be the set of used time instants in H and let tu be a used time instant. Then, there is at least one dynamic edge eu P EpHq of the form p¨,tu , ¨, ¨q or p¨, ¨, ¨,tu q such that tu “ π2 peu q or tu “ π4 peu q. Hence, we can write TU pHq as ď tπ2 peq, π4 pequ . TU pHq “ ePEpHq Since Lemma 1 shows that the both sets of connected nodes and used time instants in a given TVG H can be recovered from the set of dynamic edges EpHq, we can conclude that such information is redundant and does not need to be stored to represent a TVG. Therefore, to represent any given TVG H, it suffices to store the set of dynamic edges EpHq as well as the sets of disconnected nodes and unused time instants of the TVG. 11 We now demonstrate in Theorem 1 that when the sets of disconnected nodes and unused time instants of a TVG are significantly smaller than the set of dynamic edges, which is actually expected in most practical cases, then the asymptotic memory complexity of the TVG representation is determined by the cardinality of the set of dynamic edges. Theorem 1. If the set of disconnected nodes and the set of unused time instants are smaller than the set of dynamic edges, then the memory complexity MC pHq for storing a given TVG H is determined by the size of the set of dynamic edges, i.e. MC pHq “ Θp|EpHq|q. Proof. Let H be an arbitrary TVG, EpHq its set of dynamic edges, VC pHq its set of connected nodes, VN pHq its set of disconnected nodes, TU pHq its set of used time instants, TN pHq its set of unused time instants, and MC pHq the memory complexity for storing the TVG H. Note that V pHq “ VC pHqYVN pHq and T pHq “ TU pHq Y TN pHq, whereas VC pHq XVN pHq “ H and TU pHq X TN pHq “ H. Since Lemma 1 shows that VC pHq and TU pHq can be recovered from EpHq, it follows that to store a representation of H in memory, it suffices to store EpHq, VN pHq, and TN pHq. We now analyze the asymptotic bounds for the memory complexity for storing these sets, together with the assumption that |VN pHq| ` |TN pHq| ă |EpHq|. • Lower bound for MC pHq: Since for storing the set EpHq it is necessary at least to store an ordered quadruple for each dynamic edge e P EpHq, the memory needed is c1 ˆ |EpHq|, where c1 is an integer constant. To store the set VN pHq it is necessary to store all nodes u P VN pHq and therefore the memory needed is c2 ˆ |VN pHq|, while to store the set TN pHq it is necessary to store all time instants u P TN pHq, leading to a memory need of c3 ˆ |TN pHq|. Therefore, the total memory need is at least c1 ˆ |EpHq| ` c2 ˆ |VN pHq| ` c3 ˆ |TN pHq|, and MC pHq “ Ωp|EpHq| ` |VN pHq| ` |TN pHq|q. Finally, from our assumption that |VN pHq|`|TN pHq| ă |EpHq|, we conclude that MC pHq “ Ωp|EpHq|). • Upper bound for MC pHq: On the upper bound analysis, we see that the information to be stored is at most the same three sets EpHq, VN pHq, and TN pHq, which had to be stored in the lower bound analysis. Therefore, we conclude that MC pHq “ Op|EpHq|). Since MC pHq “ Ωp|EpHq|q and also MC pHq “ Op|EpHq|q, we finally conclude that MC pHq “ Θp|EpHq|q, if |VN pHq| ` |TN pHq| ă |EpHq|. 12 Corollary 1. The complete expression of the memory complexity for a given TVG H is MC pHq “ Θp|VN pHq| ` |TN pHq| ` |EpHq|q. Proof. It follows from the proof of Theorem 1 that the formal and complete expression of the lower bound memory complexity for storing a given TVG H is MC pHq “ Ωp|VN pHq| ` |TN pHq| ` |EpHq|q, where VN pHq its set of disconnected nodes, TN pHq its set of unused time instants, and EpHq the set of dynamic edges, while the upper bound is MC pHq “ Ωp|VN pHq| ` |TN pHq| ` |EpHq|q. We therefore conclude that complete form of the memory complexity is MC pHq “ Θp|VN pHq| ` |TN pHq| ` |EpHq|q. Since in most practical cases the amount of disconnected nodes and unused time instants is significantly smaller than the number of dynamic edges (i.e., |VN pHq| ` |TN pHq| ! |EpHq|), we can conclude that the expected amount of memory needed to store a representation of a given TVG H is determined by the size of the set of dynamic edges |EpHq|, i.e. MC pHq “ Θp|EpHq|q as states Theorem 1. Furthermore, any given TVG H for most practical applications can typically be expected to be sparse, i.e. the number of dynamic edges is significantly smaller than the squared number of temporal nodes (|EpHq| ! |V T pHq|2 ), thus allowing its storage in a compact form, similar to the compressed forms used for sparse matrices [37]. 4 A unifying TVG model In this section, we show that the TVG model we propose can be used to represent many previous models found in the literature. Further, these models are not always capable of representing each other and none of them has the same representation range of the unifying model we propose. In order to assert that all the considered models can be directly represented by our model, we show that each of these models is in fact a subset of what is representable in our model and can therefore be easily represented in our model, maintaining any previously obtained result as at least valid on a special case. To achieve this, we show that each of the studied models can be represented in our TVG model by using only a subset of the types of dynamic edges available on our model (see Section 2). We achieve this by presenting the general structure of the matrix form of the adjacency tensor of the TVG that represents the model under study. We establish a set of four nodes V “ t0, 1, 2, 3u and a set of three time instants T “ tt0 ,t1 ,t2 u to be used to illustrate all of the following analyses. This is 13 done without loss of generality, since the same procedure could be applied to any set of nodes and time instants present in a TVG. To show the representation in our model of the different TVG models found in the literature, we group them into classes that can be represented in our unifying model in similar ways and analyze these classes in the remaining of this section. 4.1 Models based on snapshots Some snapshot models for TVGs adopt an aggregate graph and a sequence of successive state sub-graphs that represent the network in a discrete way as time passes. Some examples are the models proposed by Ferreira [19] and Xuan et al. [20]. We also consider in this same class of snapshot models the models proposed by Holme [21] and Holme and Saramäki [26], where edges are represented as triples of the form pi, j,tq meaning the existence of a contact between nodes i and j at time t. Still under this same class, we also consider the models proposed by Tang et al. [23, 38, 39], as well as all other models in which the TVG is proposed as a sequence of static graphs (i.e., the snapshots), each of them representing the TVG at a given time instant. Snapshot models are widely used in the literature and in general give an intuitive and straightforward notion of TVGs. Nevertheless, the snapshot models also demand some assumptions to be made without having them explicitly constructed in the model. For instance, it is usually assumed in such snapshot models that the nodes have a sort of memory that allows the transitivity induced by the edges to propagate on each node over time. This means that a path can be constructed passing through a given node even if this node is disconnected from all others during a period of time, meaning that the node is capable of retaining the edge transitivity during the period of disconnection. Even though this behavior is straightforward and intuitively expected in many cases, it is not constructed within the model and has to be assumed as an additional (external) property of the model, resulting that this assumption has also to be incorporated into the algorithms used with those model and thus making these algorithms dependent of these external assumptions. From the analysis of the snapshot models, we remark that clearly they only use edges connecting nodes in the same given time instant. This is consistent with the concept of spatial dynamic edges proposed in our model. In this way, all the TVGs constructed in this class of snapshot models can be represented in our model by using only spatial dynamic edges, given that all the necessary nodes and time instants are present in our model. Thus, a TVG with four nodes and 14 three time instants in this class of snapshot model would be represented in our model by a TVG whose adjacency tensor in matrix form is as the one presented in Figure 4. The entries containing “*” may have non-zero values, indicating the potential presence of a spatial edges. Note that all the entries of this kind are located in the main block diagonal of the adjacency tensor in matrix form, which actually corresponds to the snapshots. Figure 4: Snapshot models represented by our unifying TVG model. In our general representation, we allow spatial and temporal self-loops (i.e. dynamic edges connecting a node to itself at the same time instant). This kind of edge is represented on the main diagonal of the matrix, just as it would be on an adjacency matrix of a static graph. Any TVG of the snapshot class can be represented in this straightforward way in our proposed unifying model. Actually, snapshot models are in fact a subspace of the representable TVGs in our model, making clear that this whole class of snapshot models is rather a subset of the representable TVGs in our model. Further, in snapshot models, each snapshot is formally disconnected from each other (although arguably implicitly connected), since the temporal connections are not explicitly constructed in the snapshot models. The number of dynamic edges used to represent a TVG of this kind is the 15 same as the number of edges present in the original snapshot-based representation. Therefore, the memory complexity for storing TVGs in our model is the same found on the original snapshot-based representation, which is compatible with the memory complexity discussed in Section 3.4. 4.2 Models based on continuous time intervals (CTI) The class discussed in this subsection includes models that use a presence function defined over continuous time intervals (i.e. t P R` ), such as the continuous time version of TVG proposed by Casteigts et al. [24]. We further assume that the presence function used in such models is constructed in such a way that every time interval (or their union) has a non-zero and finite measure. Although this assumption is not explicitly stated in the original paper, it is consistent with all examples and the reasoning present therein. It is also important to remind that the model we construct is a discrete version of the continuous time interval model that nonetheless retains all information present on the original model based on continuous time intervals. In order to represent TVGs based on continuous time intervals in our unifying model, some representations are possible depending on the target application of the model that defines the semantics associated with an edge in the TVG. An example with three possible representations is presented in Figure 5 and each of the these three representations is explained in the following. If the semantic of a time interval pta ,tb s associated with an edge e between two nodes u and v is that the edge exists from the beginning of the interval (i.e. from time ta ) until the end of the interval (i.e. until time tb ), it is possible to represent the existence of such an edge by using a mixed dynamic edge. This is consistent with the original semantics, where the edge e exists in the interval delimited by the mixed edge (i.e. from ta until tb ). A concrete example of this first representation is shown comparing Figures 5(a) and 5(b). The edge between nodes 0 and 1 present at the time interval p1, 15s shown in Figure 5(a) is represented by the mixed dynamic edge p0,t1 , 1,t15 q in Figure 5(b). Similarly, the edge between nodes 1 and 2 at the time interval p5, 7s is represented by the mixed dynamic edge p1,t5 , 2,t7 q. In the case an edge is present at different time intervals, we represent this edge by having one mixed edge for each time interval. Further, if the edge between nodes u and v is bidirectional, this edge is represented at each time interval by the pair of edges pu,ta , v,tb q and pv,ta , u,tb q. Note that once the edges present in the TVG are defined for all time intervals, it is possible to construct the set T 16 0 (1,15] 1 (5,7] 2 (a) TVG model based on continuous time intervals t1 t5 t7 t15 0 0 0 0 1 1 1 1 2 2 2 2 (b) First representation: Use of mixed edges t1 t5 t7 t15 t1 t5 t7 t15 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 (c) Second representation: Snapshots (d) Third representation: Use of temporal and spatial edges Figure 5: Converting continuous time intervals into a discrete TVG. of time instants based on the dynamic edges present in the set E. From this, we conclude that the TVGs in this class modeled by continuous time intervals can be represented in our model using only progressive mixed dynamic edges. Figure 6 shows the possible non-zero entries for a TVG with four nodes and three time instants based on the model of continuous time intervals using the first representation we are describing. Clearly from Figure 6, this kind of TVG is a subspace (and therefore a subset) of the TVGs representable by our proposed unifying model. Thus, we conclude that TVGs of this class are particular cases of the TVGs representable in our model. Even though this first representation by our model carries all information present in the original TVG based on continuous time intervals, it still relies on an assumption that is not explicitly present in this first representation. Namely, this assumption concerns that the time instants present on each mixed dynamic edge represent the time intervals on which the edge exists. Note that the same happens in the original model based on continuous time intervals. For example, using this assumption, a path between nodes 0 and 2 exists at any given time in17 Figure 6: The first representation of TVGs based on continuous time intervals. stant during the interval between t5 and t7 . To understand that this assumption is made in a manner that does not directly follows as a property of the structure of the TVG, notice that it makes the transitivity of the relation implied by an edge more difficult and cumbersome to determine. This happens because an external processing is needed to verify the presence of the edge (either by the time intervals indicated on the mixed dynamic edge or by the set of intervals on the continuous time model) in order to be able to establish its transitivity to another edge. To establish the existence of a path connecting two nodes, it is necessary to compute the intersections of the time intervals on which each edge exists. The resulting path then exists on the time interval obtained by this intersection. Therefore, the existence of a path connecting nodes in this kind of representation can be non-intuitive and computationally expensive to determine, thus impacting considerations about connectivity, reachability, communicability, and any other property derived from the transitivity of edges in a TVG. A second representation (Figure 5(c)) comes from the realization that the continuous time intervals model is in fact a continuos form of the snapshot model, in which edges are present at infinite (uncountable) time instants. Therefore, a natural way to represent the CTI model in a discrete form is to have the time instants 18 set T formed by the beginning and end instants of each interval present on the CTI model, and then placing the corresponding edge at each time instant for which an edge is present in the CTI model. The result of this is a snapshot representation of the CTI model, as shown in Figure 5(c), where the edge between nodes 0 and 1 is present at instants t0 ,t5 ,t7 , and t15 , while the edge between nodes 1 and 2 is present at instants t5 and t7 . This representation has the same characteristics we highlighted for the snapshot class of models. Although the TVG connectivity can be readily determined at each time instant, the snapshots (i.e. time instants) are formally disconnected from each other. Hence, if any sort of connectivity between time instants is to be considered, it has to be stated as an additional assumption, which is not directly encoded in the representation. A third representation (Figure 5(d)) can be derived from the second one (snapshots), by formally placing temporal dynamic edges where the connectivity between successive time instants is desired. In our example, this leads to the representation shown in Figure 5(d), where temporal nodes are used to connect nodes 0 and 1 between time instants t0 ,t5 ,t7 , and t15 and to connect nodes 1 and 2 from time instant t5 to t7 . This reflects the fact that in the original CTI model the edge p0, 1q was present at the interval pt0 ,t15 s, while the edge p1, 2q was present at the interval pt5 ,t7 s. Figure 7 shows the possible non-zero entries on the matrix form of the adjacency tensor of a TVG of this class using this third representation. Figure 7: The third representation of TVGs based on continuous time intervals. 19 In order to determine the memory complexity of representing a TVG based on continuous time intervals in our proposed unifying model, we state the following proposition, based on the third representation, presented in Figure 5(d), which uses the largest number of dynamic edges: Proposition 1. A TVG based on continuous time intervals with n nodes, m edges, and η continuous time intervals defining these edges can be represented in our model with Opη 2 q dynamic edges. Proof. For each time interval in the continuous time representation, a mixed dynamic edge is created on our model. Since each mixed edge requires two time instants to be defined, it follows that the representation in our model requires at most 2η time instants. Therefore, in our model we have |T | “ 2η in the worst case. To decompose a mixed dynamic edge into spatial and temporal dynamic edges, at most |T | spatial dynamic edges are needed to connect the two nodes on the mixed edge in all possible time instants. Further, at most 2p|T | ´ 1q temporal dynamic edges are needed to connect the nodes over all possible time instants. Therefore, to fully decompose a mixed edge, at most 2p|T | ´ 1q ` |T | “ 3|T | ´ 2 dynamic edges are needed. Hence, as η mixed dynamic edges are needed in the representation, to fully decompose them into spatial and temporal edges, at most ηp3|T | ´ 2q dynamic edges are needed. Expanding and substituting |T | “ 2η for the worst case, we have ηp3|T | ´ 2q “ ηp6η ´ 2q “ 6η 2 ´ 2η. We therefore conclude that the number of dynamic edges needed is Opη 2 q. From this, we further conclude that, for a TVG with η continuous time intervals, in our model we have |E| “ Opη 2 q. Therefore, the memory complexity of the representation based on continuous time instants is Opη 2 q, which is compatible with the memory complexity discussed in Section 3.4. 4.3 Models based on spatial and temporal edges (STE) Some models like the one proposed by Kostakos [22] are based on the idea that a class of links represent instantaneous iterations between distinct nodes while 20 other class represent a waiting state of a given node. These concepts are formalized in our proposed unifying model by spatial and temporal dynamic edges. In our model, these dynamic edges are fully formalized and can be used to make a unambiguous representation of this kind of TVG. Figure 8 depicts the entries that may be non-zero on the matrix form of the adjacency tensor of a TVG of this class having four nodes and three time instants. It is again clear that this class of TVGs can be thus represented as a subset of the TVGs that can be represented by the unifying model we propose. Figure 8: Representation of TVGs based on spatial and temporal edges. 4.4 Models based on temporal and mixed edges (TME) Some works found in the recent literature, such as the one by Kim and Anderson [25], loosely suggest the use of edges connecting nodes at different time instants. TVGs of this class can be represented in our model using only temporal and mixed dynamic edges. Figure 9 shows the matrix form of the adjacency tensor of a TVG of this class. It can be seen that this is also a particular case of the TVGs that can be represented using our unifying model. 21 Figure 9: Representation of TVGs based on temporal and mixed edges. 4.5 Overview on the unifying representation of previous models Table 1 shows a representation map that indicates if a (class of) TVG model(s) is able to represent another (class of) TVG model(s). We compare the unifying representation model we propose with the models based on (i) snapshots, (ii) continuous time intervals (CTI), (iii) spatial and temporal edges (STE), and (iv) temporal and mixed edges (TME), which have been presented in Sections 4.1 to 4.4, respectively. Snapshot models can only represent TVGs of the continuous time intervals (CTI) class, since they are in fact a continuos time version of the snapshot model. Direct representation of other models by snapshots is not possible because the snapshot model lacks the notion of temporal edges. Models based on continuous time intervals (CTI) and spatial and temporal edges (STE) can represent snapshot models because they support the notion of spatial edges as well as they are mutually able to represent each other. Models based on temporal and mixed edges (TME) are able to represent models based on continuous time intervals (CTI) in a discretized way. Remark that the converse is not true, i.e. models based on CTI are unable to 22 Table 1: Representation map between TVG models: An entry is checked if the (class of) model(s) in the row is able to represent the (class of) model(s) in the column. Snapshots CTI STE TME Unifying model Snapshots X X CTI X X X STE X X X TME X X Unifying model X X X X X represent models based on TME as they lack the notion of mixed edges. Furthermore, note that models based on TME are unable to represent in all cases models with spatial edges, such as the ones based on snapshots and STE. This is because, although a mixed edge could be seen as a composition of a spatial and a temporal edge, a mixed edge is unable to represent a single spatial edge, thus preventing the representation of these cases. Overall, we remark that considering the previous classes of TVG models, none is able to represent all others. This basically happens because the kind of edges present in one model not necessarily can be transformed into the kind of edges present in another model or class of models. In contrast, we have shown along this section that all classes of previous TVG models we consider can be represented in the unifying model we propose, whereas these previous TVG models not necessarily are able to represent each other, as shown in Table 1. Additionally, none of the previous TVG models makes use of regressive edges, which could be used in our proposal to intrinsically model cyclic (i.e. periodic) behavior in dynamic networks. Therefore, we highlight that none of the previous analyzed models have all the representation capabilities and thus the same expressive power of our proposed unifying model for TVGs. 5 Summary and outlook In this paper, we have proposed a novel model for representing finite discrete Time-Varying Graphs (TVGs). We have shown that our model is simple, yet flexible and efficient for the representation and modeling of dynamic networks. The proposed model preserves the discrete nature of the basic graph abstraction and has algebraic representations similar to the ones used on a regular graph. More23 over, we have also shown that our unifying model has enough expressive power to represent several previous (classes of) models for dynamic networks found in the recent literature, which in general are unable to represent each other, and also to intrinsically represent cyclic (i.e. periodic) behavior of dynamic networks. To further illustrate the flexibility of our model, we remark that our model can also be used to represent time schedules using a TVG. For example, in this case, the nodes can be thought of as locations and mixed edges as the amount of time taken to move between locations. Such representation can model scheduled arrivals and departures in transportation systems allowing, for instance, the evaluation whether a connection is feasible or the delivery time for logistics management. We have analyzed the proposed model proving that if the TVG nodes can be considered as independent entities at each time instant, the analyzed TVG is isomorphic to a directed static graph. This basic theoretical result has provided the ground for achieving other theoretical results that show that some properties of the analyzed TVG can be inferred from the temporal node representation of that TVG. This is an important set of theoretical results because this allows the use of the isomorphic directed graph as a tool to analyze both the properties of a TVG and the behavior of dynamic processes over a TVG. We have also demonstrated that, for most practical cases, the asymptotic memory complexity of our TVG model is determined by the cardinality of the set of edges. Further, from the basic definition of the model we proposed for representing TVGs, we have derived some basic properties such as communicability and connectivity using only properties that follow directly from the underlying structure of the model, such as the transitivity of the relation induced by the dynamic edges. As a consequence, in contrast to previous works, our model can be used without the need of external assumptions, meaning that all properties are derived from explicit properties of the TVG, in the same way that happens with static graphs. As future work, we intend to apply our proposed TVG model in the analysis of different dynamic complex networks as well as of properties over such dynamic networks. An example of such analysis is the recent work by Costa et al. [40]. In that work, authors analyze the concept of time centrality for information diffusion in dynamic complex networks using the TVG model we propose in this paper, thus investigating the relative importance of different time instants to start a information diffusion process. 24 Acknowledgment The cooperation among authors has been supported by STIC-AmSud. K. Wehmuth and A. Ziviani acknowledge that their work was partially funded by the Brazilian funding agencies CAPES, CNPq, and FAPERJ as well as the Brazilian Ministry of Science, Technology, and Innovation (MCTI). E. Fleury acknowledges the support by MILYON. References [1] V. Dhar, “Data science and prediction,” Communication of the ACM, vol. 56, no. 12, pp. 64–73, Dec. 2013. [2] C. A. Mattmann, “Computing: A vision for data science,” Nature, vol. 493, no. 7433, pp. 473–475, Jan. 2013. [3] H. V. Jagadish, J. Gehrke, A. Labrinidis, Y. Papakonstantinou, J. M. Patel, R. Ramakrishnan, and C. Shahabi, “Big data and its technical challenges,” Communications of the ACM, vol. 57, no. 7, pp. 86–94, Jul. 2014. [4] A. Vespignani, “Predicting the behavior of techno-social systems,” Science, vol. 325, no. 5939, pp. 425–428, Jul. 2009. [5] H. Kitano, “Computational systems biology,” Nature, vol. 420, no. 6912, pp. 206–210, Nov. 2002. [6] T. G. Lewis, Network science: Theory and applications. Wiley, 2009. [7] L. Kocarev, “Network science: A new paradigm shift,” IEEE Network, vol. 24, no. 6, pp. 6–9, Nov. 2010. [8] U. Brandes, G. Robins, A. McCraine, and S. Wasserman, “What is network science?” Network Science Journal, vol. 1, no. 1, pp. 1–15, Apr. 2013. [9] R. Pastor-Satorras and A. Vespignani, “Epidemic spreading in scale-free networks,” Physical Review Letters, vol. 86, pp. 3200–3203, Apr. 2001. [10] D. Kempe, J. Kleinberg, and E. Tardos, “Maximizing the spread of influence through a social network,” in Proc. of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pp. 137–146, Aug. 2003. 25 [11] A. Barrat, M. Barthelemy, and A. Vespignani, Dynamical Processes on Complex Networks. Cambridge University Press, 2008. [12] M. Iliofotou, M. Faloutsos, and M. Mitzenmacher, “Exploiting dynamicity in graph-based traffic analysis: Techniques and applications,” in Proc. of the ACM International Conference on Emerging Networking Experiments and Technologies (ACM CoNEXT), pp. 241–252, Dec. 2009. [13] N. Ganguly, A. Deutsch, and A. Mukherjee, Eds., Dynamics on and of complex networks. Springer, 2009. [14] D. Figueiredo, P. Nain, B. Ribeiro, E. de Souza e Silva, and D. Towsley, “Characterizing continuous time random walks on time varying graphs,” ACM SIGMETRICS Performance Evaluation Review, vol. 40, no. 1, pp. 307– 318, Jun. 2012 [15] P. Holme and J. Saramäki, Eds., Temporal Networks. Springer, 2013. [16] P. Holme, “Modern temporal network theory: arXiv:1508.01303, 33 p., Aug. 2015. A colloquium,” [17] D. Braha and Y. Bar-Yam, “From centrality to temporary fame: Dynamic centrality in complex networks,” Complexity, vol. 12, no. 2, pp. 59–63, Nov. 2006. [18] S. A. Hill and D. Braha, “Dynamic model of time-dependent complex networks,” Physical Review E, vol. 82, no. 4, Oct. 2010. [19] A. Ferreira, “Building a reference combinatorial model for MANETs,” IEEE Network, vol. 18, no. 5, pp. 24–29, Oct. 2004. [20] B. B. Xuan, A. Ferreira, and A. Jarry, “Evolving graphs and least cost journeys in dynamic networks,” in Proc. of the Symposium on Modeling and Optimization in Mobile, Ad Hoc and Wireless Networks (WiOpt), Mar. 2003. [21] P. Holme, “Network reachability of real-world contact sequences,” Physical Review E, vol. 71, no. 4, pp. 1–8, Apr. 2005. [22] V. Kostakos, “Temporal graphs,” Physica A: Statistical Mechanics and its Applications, vol. 388, no. 6, pp. 1007–1023, Mar. 2009. 26 [23] J. Tang, M. Musolesi, C. Mascolo, and V. Latora, “Temporal distance metrics for social network analysis,” in Proc. of the ACM Workshop on Online Social Networks (WOSN). pp. 31–36, Aug. 2009. [24] A. Casteigts, P. Flocchini, W. Quattrociocchi, and N. Santoro, “Time-varying graphs and dynamic networks,” International Journal of Parallel, Emergent and Distributed Systems, vol. 27, no. 5, pp. 387–408, Oct. 2012. [25] H. Kim and R. Anderson, “Temporal node centrality in complex networks,” Physical Review E, vol. 85, no. 2, Feb. 2012. [26] P. Holme and J. Saramäki, “Temporal networks,” Physics Reports, vol. 519, no. 3, pp. 97–125, Oct. 2012. [27] J. I. Alvarez-Hamelin, E. Fleury, A. Vespignani, and A. Ziviani, “Complex dynamic networks: Tools and methods (guest editorial),” Computer Networks, vol. 56, no. 3, pp. 967–969, Dec. 2012. [28] C. Sengul, A. C. Viana, and A. Ziviani, “A survey of adaptive services to cope with dynamics in wireless self-organizing networks,” ACM Computing Surveys, vol. 44, no. 4, pp. 23:1–23:35, Sep. 2012. [29] A. Guimarães, A. B. Vieira, A. P. C. Silva, and A. Ziviani, “Fast centralitydriven diffusion in dynamic networks,” in Proc. of the Workshop on Simplifying Complex Networks for Practitioners (SIMPLEX), World Wide Web conference (WWW), pp. 821–828, May 2013. [30] L. Martinet, C. Crespelle, and E. Fleury, “Dynamic contact network analysis in hospital wards,” in Proc. of the Workshop on Complex Networks (CompleNet), pp. 241–249, Mar. 2014. [31] Q. Wang, “Link prediction and threads in email networks,” in Proc. of the International Conference on Data Science and Advanced Analytics (DSAA), pp. 470–476, Oct. 2014. [32] K. Wehmuth, E. Fleury, and A. Ziviani, “On MultiAspect Graphs,” arXiv:1408.0943, 30 p., Aug. 2014. [33] K. Wehmuth, E. Fleury, and A. Ziviani, “MultiAspect Graphs: Algebraic representation and algorithms,” arXiv:1504.07893, 61 p., Aug. 2015. 27 [34] K. Wehmuth, A. Ziviani, and E. Fleury, “A Unifying Model for Representing Time-Varying Graphs,” INRIA Technical Report no. RR-8466, 39 p., Jan. 2014. [35] M. Kivelä, A. Arenas, M. Barthelemy, J. P. Gleeson, Y. Moreno, and M. A. Porter, “Multilayer networks,” Journal of Complex Networks, Jul. 2014. [36] T. G. Kolda, “Multilinear operators for higher-order decompositions,” Sandia National Laboratories, Technical Report, Apr. 2006. [37] R. Barrett, M. Berry, T. F. Chan, J. Demmel, J. M. Donato, J. Dongarra, V. Eijkhout, R. Pozo, C. Romine, and H. V. der Vorst, Templates for the solution of linear systems: Building blocks for iterative methods. SIAM, 1994. [38] J. Tang, S. Scellato, M. Musolesi, C. Mascolo, and V. Latora, “Small-world behavior in time-varying graphs,” Physical Review E, vol. 81, no. 5, pp. 81– 84, May 2010. [39] J. Tang, M. Musolesi, C. Mascolo, V. Latora, and V. Nicosia, “Analysing information flows and key mediators through temporal centrality metrics,” in Proc. of the ACM EuroSys Workshop on Social Network Systems (SNS), pp. 1–6, Apr. 2010. [40] E. C. Costa, A. B. Vieira, K. Wehmuth, A. Ziviani, and A. P. C. Silva, “Time centrality in dynamic complex networks,” arXiv:1504.00241, 12 p., Apr. 2015. 28
8cs.DS
A co-located partitions strategy for parallel CFD–DEM couplings Gabriele Pozzetti, Xavier Besseron, Alban Rousset, Bernhard Petersa,∗ a Campus Belval, Université du Luxembourg 6, Avenue de la Fonte, L-4364 Esch-sur-Alzette Luxembourg arXiv:1802.05029v1 [cs.CE] 14 Feb 2018 Abstract In this work, a new partition-collocation strategy for the parallel execution of CFD–DEM couplings is investigated. Having a good parallel performance is a key issue for an Eulerian-Lagrangian software that aims to be applied to solve industrially significant problems, as the computational cost of these couplings is one of their main drawback. The approach presented here consists in co-locating the overlapping parts of the simulation domain of each software on the same MPI process, in order to reduce the cost of the data exchanges. It is shown how this strategy allows reducing memory consumption and inter-process communication between CFD and DEM to a minimum and therefore to overcome an important parallelization bottleneck identified in the literature. Three benchmarks are proposed to assess the consistency and scalability of this approach. A coupled execution on 280 cores shows that less than 0.1% of the time is used to perform inter-physics data exchange. 1. Introduction Eulerian-Lagrangian couplings are nowadays widely used to address engineering and technical problems. In particular, CFD–DEM couplings have been successfully applied to study several configurations ranging from mechanical [14, 13], to chemical [11] and environmental [8] engineering. CFD–DEM coupled simulations are normally very computationally intensive, and, as already pointed out in [10], the execution time represents a major issue for the applicability of this numerical approach to complex scenarios. Therefore, optimizing the parallel performance of such a coupling is a fundamental step for allowing large-scale numerical solutions of industrial and technical problems. The parallelization of Eulerian-Lagrangian software is, however, rather delicate. This is mainly due to the fact that the optimal partitioning strategies for those wireframes are different, and the memory requirement of a coupled solution can represent a major performance issue. Furthermore, since the coupling normally affects an extended domain region (often the whole computational domain), the amount of information that is required to be exchanged is normally important. For this reason, highly efficient coupling approaches for boundary problems, like the one proposed in [4] may suffer for the extensive communication layer. At the same time, due to the Eulerian-Lagrangian nature of the coupling, mesh-based communication as the one proposed in [9] cannot, by themselves take care of the information exchange. One of the earliest attempt to parallelize a DEM algorithm was proposed in [21], where the authors distributed inter-particle contacts among processors on a machine featuring 512 cores. In their scheme, all particle data was stored in every process resulting in a memory-intensive computation that leads to a speedup of 8.73 for 512 cores with a 1672 particles assembly. A later work [12] showed how by reducing the DEM inter-process communication, a speedup of ∼11 within 16 process-computation of 100k particles was obtainable. This proved how, for the sole Lagrangian software, the memory usage and the inter-process communication can profoundly affect the parallel performance of the code. The problem of memory consumption and inter-process communication becomes even more important when a Lagrangian code is coupled with an Eulerian one as the single pieces of software generally reach the optimum performance with different load and partitioning strategies. This can lead to massive inter-process communication ∗ Principal corresponding author. Email: [email protected], [email protected] Preprint submitted to Elsevier February 15, 2018 that can deeply affect the performance of the overall algorithm. In [5], the authors proposed a mirror domain method for ensuring the correct passage of information between Lagrangian bubbles and an underlying fluid flow in a parallel execution. This method consists in distributing global knowledge of all the Eulerian and Lagrangian domains, by letting a process performing only a part of the computation, and then distributing the information via gather-scatter operations. It was shown in [5], how this strategy allows a correct access to information and an almost linear speedup up to 4 processes for a case with 648000 CFD cells and 100000 particles. Nevertheless, the strategy wasn’t able to offer significant advantages when operating on more than 32 processes. Therefore, its application is limited to medium-small scale problems. In order to cope with this problem, in [10], the authors proposed a parallelization strategy based on a Jostle graph for the DEM part and an independent parallelization for the CFD part. This allowed reducing memory consumption while keeping an high inter-physics communication load. The resulting coupling was observed to scale better than what previously seen in the literature obtaining a speedup of ∼35 for 64 processes. Nevertheless, the coupled code was shown to perform significantly worse than the sole DEM part, proving how the inter-physics communication between the Eulerian and the Lagrangian part can induce significant performance issues. This issue was underlined by [1], where a communication strategy based on non-distributed memory was implemented in order to reduce the communication costs. The results proved how this strategy can indeed reduce the DEM and the inter-physics communication, yet its advantages are limited to the usage of ∼30 processes though not allowing very-large-scale simulations. An important attempt to study large-scale problems with an Eulerian-Lagrangian coupling was presented in [19]. In this contributions, the authors propose a coupling between a highly efficient code for molecular dynamics and a well-known open-source software for CFD. The proposed decomposition strategy for CFD and molecular dynamics code are completely independent from each other, and all the required data for the inter-physics exchange is provided through a complex communication layer. It is shown how the specific couple of codes can tackle large-scale problems and scale over more than one hundred of processes. Nevertheless, when a large number of processes is used, the inter-physics communication becomes very important, taking up to 30% of the computation time. Nowadays, there is still no reference solution for the parallel execution of CFD–DEM couplings, and even though several strategies have been proposed, they have been focusing on specific software. The aim of the current work is to investigate the performance of a parallelization strategy for a CFD–DEM four-way coupling that aims to minimize the memory consumption and the inter-physics communication. The article is structured as follows. Firstly, the Lagrangian and Eulerian parts of this coupling are presented, and the systems of equations solved by the DEM and CFD part described. Secondly, the partitioning strategy based on co-located partitions is introduced. Finally, three benchmark cases are proposed in order to assess the consistency and performances of the proposed parallelization strategy. In this work, we refer to the coupling between the XDEM platform [16, 15], which is used to treat the Lagrangian entities, and the OpenFOAM [22] libraries, which is used to resolve the fluid flow equations in a Eulerian reference. The parallelization strategy is presented in its general way and can be implemented for a generic CFD–DEM coupling independently from the specific software. To the best of the authors’ knowledge, the partitioning strategy for CFD–DEM couplings here proposed is the first one that succeeds in keeping the inter-physics communication negligible while running over hundreds of process. 2. Methodology 2.1. Discrete Element Method (DEM) The Discrete Element Method (DEM) is a well-established approach for granular flows [7]. The XDEM platform [16, 3, 13] aims to extend the application of the DEM by attaching chemical/thermodynamical variables on the particles, and coupling the DEM core with codes for computational fluid dynamics (CFD). More in general, the coupling between CFD and DEM is a common strategy for addressing industrially relevant problems [13, 11]. In this work, the DEM module of the XDEM platform is used to evolve a set of particles describing physical entities in the 2 Table 1: List of variables Symbol Variable Symbol Variable uf µf Fb Ffpi ρf Ap fluid velocity fluid viscosity body force fluid-particle interaction force fluid density particle surface area p uc C σf ρp ε pressure compression velocity surface curvature fluid stress tensor particle density local porosity Mcoll torque acting on a particle due to collision events particle mass particle moment of inertia contact force acting on a particle contact force acting on the particle i due to the collision with particle j force rising from the particle-fluid interaction Fg gravitational force Mext u external torque acting on a particle particle velocity φ particle orientation ω particle angular velocity m I Fcoll Fi j Fdrag presence of a multiphase fluid. Defining with xi the positions, mi the masses, and φ the orientations, one can write d2 xi dt 2 d2 Ii 2 φ i dt mi = Fcoll + Fdrag + Fg , (1) = Mcoll + Mext , (2) In equation 1, the term Fc indicates the collision force, Fcoll = ∑ Fi j (x j , u j , φ j , ω j ), (3) i6= j with u j the velocity of particle j, and ω the angular velocity. The term Mcoll indicates the torque acting on the particle due to collisions Mcoll = ∑ Mi j (x j , u j , φ j , ω j ), (4) i6= j with Mi j the torque acted from particle j to particle i. The term Fdrag takes into account the force rising from the interaction with the fluid, and Fg corresponds to the gravitational force. For what concerns Fdrag the analytical expression of the term can be obtained from the literature. Many authors have focused on its description [6, 2] that is normally provided in the form of a semi empirical law Fdrag = β (u f − u p ), (5) β = β (u f − u p , ρ f , ρ p , d p , A p , µ f , ε), (6) with u f , u p the fluid and particle velocity respectively, ρ f , ρ p the respective densities, d p , A p the particle characteristic length and area, µ f the fluid viscosity, and ε the porosity, defined as the ratio between the volume occupied by the fluid and the total volume of the CFD cell. For the sake of generality, we took β as described in [18]. During the parallel execution of XDEM, the simulation domain is geometrically decomposed in regularly fixed-size cells that are used to distribute the workload between the processes. Every process has global knowledge of the domain structure and decomposition, but only performs the calculation, and holds knowledge, for the particles 3 that belong to its sub-domain. The load partitioning between processes has been shown to be often crucial for the parallel performance of XDEM [17], and different partitioning strategies are available within the XDEM platform. In this contribution, we investigate the parallel performance of XDEM when operated with geometrically uniform partitions for the coupling with OpenFOAM. In figure 8, standalone XDEM executions compare the different partitioning algorithms, including the geometrically uniform approach, in order to assess their influence on the specific cases used in this work. 2.2. Computational Fluid Dynamics (CFD) For the fluid solution, we refer to an unsteady incompressible flow through porous media with forcing terms arising from particle phase, as described in [16, 15, 14]. This system must fulfill the Navier-Stokes equations in the form, ∇ · εu f = ∂ε , ∂t ∂ εu f + ∇ · (εu f u f ) = −ε∇p + TD + Fb + Ffpi , ∂t  TD = ∇ · ε µ f ∇u f + ∇T u f , (7) (8) (9) with u f the fluid velocity, p the fluid pressure, µ f the fluid viscosity, Fb a generic body force, Ffpi the fluid-particle interaction force, that is the counterpart of Fdrag , which is here treated with the semi-implicit algorithm proposed in [23]. This set of equations is solved with the OpenFOAM libraries [22], which parallelization is based on domain decomposition. The CFD domain is split into sub-domains assigned to each process available at run time, over each of them a separate copy of the code is run. The exchange of information between processes is performed at boundaries through a dedicated patch class as described in [22]. In this contribution, the CFD domain is partitioned according to the DEM domain, as described in section 2.3. In figure 5, we compare standalone OpenFOAM execution with different partitioning algorithm to assess the influence of the partitioning on our specific cases. 2.3. Co-located partitions strategy for domain decomposition Since CFD–DEM couplings based on volume averaging technique only consider local exchanges of mass, momentum, and energy, the Lagrangian (DEM) entity will be affected only from the fluid characteristic of an area close to its center. At the same time, the fluid flow will be only locally interested by the action of particles present in the local region. In this contribution, we are taking advantage of theses characteristics to design an optimized partitioning strategy for our CFD–DEM simulation. This partitioning strategy is illustrated on figure 1: 1. In our problem, the CFD (OpenFOAM) and DEM (XDEM) domains are overlapping, totally or partially. As stated earlier, interactions between the fluid and the particles happen for objects closely located in the simulation space. 2. In consequence, our partitioning strategy considers the 2 domains together and its goal is to maintain the locality of the objects in the simulation. This means the domain elements co-located in the simulation space are assigned to the same process. In this way, the partition i of the CFD and the partition i of the DEM domain will be coinciding. 3. For the parallel execution, partitions are distributed to the computing nodes. It is important to make sure that partition i of the CFD domain and partition i of the DEM domain are actually located in the computing node, or even better, in the same MPI process. 4. With this partitioning strategy, the parallel execution benefits from the resulting communication pattern. Inter-partition intra-physics data exchanges, i.e. the same that occur in a standalone execution of OpenFOAM or XDEM, are performed with the MPI communication layer using native implementation. This ensures good portability of the approach as the intra-physics communication implementation is not changed. 4 Figure 1: Co-located partitions strategy for a coupled CFD–DEM simulation: 1. OpenFOAM and XDEM domains are overlapping. 2. Domain elements co-located in the simulation space are assigned to the same partitions. 3. Partitions are distributed to the computing nodes. 4. The parallel execution benefits from the resulting communication pattern. 5 Intra-partition inter-physics data exchanges are implemented with simple read/write operation in memory or direct library function calls because the two partitions share the same intra-process memory space. These data exchanges, which represent the majority of the data exchange due to the co-located partitions strategy, are much faster than the ones based on traditional communication layer like MPI. Inter-partition inter-physics data exchanges are achieved using the MPI communication layer. Thanks to the co-located partitions strategy, they are reduced to a minimum, or non-existent if the partitions are perfectly aligned. As a result, our approach has many advantages compared to the other work in the literature (cf section 1). Firstly, we rely on the distribution of the simulation domain among the computing nodes. Every partition of the computational domain keeps only its local data, and therefore reduces memory consumption. This is important as it allows simulating large-scale problems that would not fit in the memory of a single computing node. Secondly, by running the CFD and DEM code in the same process (by linking the two libraries into one executable), the local inter-physics data exchanges can be performed by direct memory accesses and avoid the overhead of an additional communication layer. Thanks to our partitioning strategy that imposes the co-location constraint on the CFD and DEM sub-domains, we ensure that the majority of the inter-physics data exchange will occur in that way. These advantages are meant to reduce the overhead of a parallel execution and to simplify the structure of the interface between the different codes, that will be identical for parallel and sequential execution. In this study, we implement the proposed strategy by using a simple partitioning algorithm that enforces the co-location constraint. This algorithm controls the partitioning of both the CFD and DEM domain. For the sake of simplicity, the global simulation space is here partitioned with a uniform strategy based on a geometric decomposition. This consists in splitting the domain into regions of similar volume making the spatial volume of the domains, that are assigned to each process, balanced. In this way, every process will handle a similar volume of the computational domain holding data and performing the computation for the CFD and DEM part occupying the specific spatial region. While this strategy is not optimal for a generic case, it greatly simplifies the implementation and the analysis of the results. The result of this partitioning is then enforced into each software bypassing any internal they may use. For OpenFOAM, this is achieved using the manual decomposition method which allows using a predefined partitioning defined in a text file. In this way, our partitioning algorithm considers the global simulation space of two physics software as a whole, not differencing if a point in space belongs to the CFD mesh or the DEM domain. This strategy is significantly different from what proposed in previous works [19], where the partitioning of CFD and DEM domains was aiming to optimize the load-balance of each standalone software, rather than the communication cost of the coupled execution. 3. Experimental results 3.1. Experimental methodology In order to assess the validity of our approach and evaluate the scalability of the proposed strategy, we have setup and executed three benchmarks. The first benchmark, One particle traveling across a processes boundary presented in section 3.2, tests the equivalence of the results between sequential run and parallel run with the current parallelization strategy. This is done to assess the validity of our approach and implementation by checking the continuity of the results when the particle travels across process boundary even without inter-partition inter-physics communication. The second benchmark, One million particles in two million cells proposed in section 3.3, investigates the effects of an heavy inter-physics communication on the performance of a standalone software. This is done to show how the proposed parallelization strategy allows maintaining the performance of the original single solver in the presence of an heavy inter-physics communication even with hundreds of processes. The third benchmark, Ten Million Particles in one Million cells described in section 3.4, studies the parallel performance of a coupled solution in case of an heavy coupled case. This is done to show how our solution can handle highly costly simulations and at the same time allows to resolve a main issue underlined in the literature linked to the inter-physics communication. 6 The experiments were carried out using the Iris cluster of the University of Luxembourg [20] which provides 168 computing nodes for a total of 4704 cores. The nodes used in this study feature a total a 128 GB of memory and have two Intel Xeon E5-2680 v4 processors running at 2.4 GHz, that is to say a total of 28 cores per node. The nodes are connected through a fast low-latency EDR InfiniBand (100Gb/s) network organized in a fat tree. We used OpenFOAM-Extend 3.2 and XDEM version b6e12a86, both compiled with Intel Compiler 2016.1.150 and parallel executions were performed using Intel MPI 5.1.2.150 over the InfiniBand network. To ensure the stability of the measurement, the nodes were reserved for an exclusive access. Additionally, each performance value reported in this section is the average of at least hundred of measurements. The standard deviation showed no significant variation in the results. 3.2. One particle traveling across a process boundary Figure 2: One particle traveling across process boundary: Setup. A first basic test is proposed, featuring a particle traveling across a boundary between processes. This is done in order to test the equivalence of the results obtained in sequential and parallel execution. The particle, initially at rest within the domain assigned to process 0, is accelerated by the fluid according to the law of equation 5. The resulting drag force pushes the particle across the boundary with process 1 causing it to be transferred from the sub-domain 0 to the sub-domain 1. As shown in figure 2, the boundary conditions imposed on the fluid domain are a uniform Dirichlet at the inlet, no-slip at the wall, and reference pressure at the outlet. The CFD domain is a square channel of dimensions of 1m, 0.2m and 0.2m, and is discretized with 240 identical cubic cells. In figure 3, the normalized particle velocity and acceleration, and the normalized solid volume and L1 norm of the fluid acceleration are proposed as a function of times. All the quantity are normalized by dividing them by their maximum value, so that they can be displayed on the same plot. The particle crosses the process boundary at 0.32s leaving the domain assigned to process 0 and entering into the domain assigned to process 1. It can be observed how the particle velocity and drag force are continuous across the processes boundary. This shows how the information on the fluid velocity at the particle position is correctly exchanged between CFD and DEM code in the whole domain, including the regions between boundaries. Similarly, the porosity and acceleration fields projected by the particle into the Eulerian grid do not suffer discontinuities when the particle switches between processes. In particular, it can be noticed how the results obtained with the parallel execution are perfectly matching the one obtained by running the code in sequential for both particle and fluid quantities. This shows how even without keeping global knowledge of the whole computational domain, continuity across processes of both Eulerian and Lagrangian quantities can be achieved. This allows greatly limiting both the memory usage and the inter-process communication. 3.3. One million particles in two million cells This second test aims to evaluate the influence of an intense intra-physics communication load on the scalability of the code. For that, a coupling between OpenFOAM for the CFD part and a dummy DEM software, named DummyDEM, is used. DummyDEM, is a modified version of XDEM whose purpose is to trigger all the necessary data exchange, but does not perform any actually DEM computation, i.e. the particles are not moving. 7 Normalized Velocity, Acceleration Normalized particle velocity and acceleration as a function of time 1 0.8 Particle Velocity (SEQ) Particle Velocity (PAR) Particle Accel. (SEQ) Particle Accel. (PAR) 0.6 0.4 0.2 Process boundary 0 0 0.1 0.2 0.3 0.4 0.5 0.6 Time [s] (a) Normalized Field Variables Normalized solid volume and fluid acceleration as a function of time 1.2 1 Solid Volume (SEQ) Solid Volume (PAR) Fluid Accel. (SEQ) Fluid Accel. (PAR) 0.8 0.6 0.4 0.2 Process boundary 0 0 0.1 0.2 0.3 0.4 0.5 0.6 Time [s] (b) Figure 3: One particle traveling across process boundary: comparison between the sequential (SEQ) and parallel (PAR) results for CFD and DEM variables. Continuity across the process boundary can be observed. 8 Figure 4: One million particles in two million cells: Setup. This case is similar to the one proposed in the previous section, but in order to increase the amount of data exchange it contains a swarm of one million of particles suspended in a channel flow discretized with two millions of identical cubical CFD cells (figure 4). The boundary conditions imposed on the fluid domain are as in the previous test the standard inlet, wall and outlet conditions on respectively the channel inlet, the surrounding walls and the outlet section. 3.3.1. Standalone OpenFOAM performance 250 ● ● Ideal ● OpenFOAM using METIS 200 ● OpenFOAM using SCOTCH Speedup OpenFOAM using Uniform ● 150 ● ● 100 ● ● 50 ● 0 ● 28 56 84 112 140 168 196 224 252 Number of cores Figure 5: One million particles in two millions cells: comparison between the scalability of the pure OpenFOAM partitioned using SCOTCH, METIS, and uniform partitioning algorithm. Similar scalability properties can be observed. In figure 5, we first present the scalability of the standalone OpenFOAM run obtained by decomposing the domain with the native SCOTCH and METIS partitioners, and the uniform partitioner that is used in the coupling. One can notice how the behavior of OpenFOAM is not highly dependent on the partitioner adopted. For that case, this can be explained considering that the problem is rather uniform and therefore the optimal partitioning would be close to the uniform one. 3.3.2. Coupled OpenFOAM–DummyDEM performance In a first step to understand the performance of the coupled CFD–DEM approach, we first decided to study the behavior of the coupling in case of uniformly distributed load, so that the final cost is not influenced by dynamic load balancing. In order to obtain this, our coupling OpenFOAM–DummyDEM performs all the coupling data exchange but our simplified DEM implementation, DummyDEM, does not execute any DEM calculation. As a result, in this testing coupling OpenFOAM–DummyDEM case, we have at every timestep: • data exchange from CFD to DEM: the fluid velocity is used to calculate the drag force applying on the particles • data exchange from DEM to CFD: the distribution of the particles is used to calculate the porosity field of the CFD domain 9 However, DummyDEM will not integrate the particle position which will remain stationary in the simulation domain. Furthermore, this allows adopting the same timestep for the CFD and DEM parts, further simplifying the analysis of the results. The resulting case is therefore equivalent to a pure CFD case, on the top of which a inter-physics communication is triggered. This allows studying the effects of the data exchange on the scalability of the software. ● 250 ● 200 Ideal ● OpenFOAM using Uniform ● Speedup DummyDEM using Uniform 150 ● OpenFOAM−DummyDEM coupled ● ● 100 ● ● 50 ● 0 ● 28 56 84 112 140 168 196 224 252 Number of cores Figure 6: One million particles in two millions cells: speedup of the coupled OpenFOAM–DummyDEM in comparison with standalone OpenFOAM and standalone DummyDEM. The scalability of the coupled OpenFOAM–DummyDEM is similar to the standalone OpenFOAM: the overhead of data exchange of the coupling is negligible. The poor performance of DummyDEM is due to the artificially reduced amount of computation. In figure 6, we evaluate the scalability of our OpenFOAM–DummyDEM coupling approach, in comparison with standalone OpenFOAM and standalone DummyDEM execution. The aim of this is to underline the influence of the presence of an interface between the Lagrangian and Eulerian domain, on the overall algorithm scalability. One can observe how in this case, the parallel performance of the DummyDEM part is poor. This is explained considering that the decision of not integrating the particle motion, which makes the DEM computation rather inexpensive, leading to the flat scalability profile observed in figure 6. This also offers the possibility of studying the performance of the coupled solver in a case where the performance of the two basic software is markedly different. In this configuration, for an execution on 252 cores, we measured that 81% of the execution time is spent in OpenFOAM code, 18% in DummyDEM and less than 1% doing inter-physics data exchange. As a consequence, the performance of the OpenFOAM–DummyDEM coupling is similar to the one of the standalone OpenFOAM execution. This shows how the presence of the interface between the two codes, that in this case is coupling one million of particles with two millions of cells, does not affect the performance significantly. This can be explained considering that, in this case, the generated partitions are perfectly aligned and, as a result, it does not introduce any inter-physics inter-process communication, and the most of the inter-physics data exchanges use direct memory accesses and avoid any communication layer. This also proves how, by using the co-located partitions strategy, the presence of an intensive inter-physics data exchange does not affect the parallel performance of the software. This is an important step for parallel execution of coupled software, as it should be ensured that its performance matches the one of a heavy loaded standalone algorithm without an excessive overhead. 3.4. Ten Million Particles in one Million cells The third test case consists in a layered bed of ten million particles moving in the presence of a carrier gas. This case was inspired by [19] and chosen to show the capability of the coupling to treat a very large amount of DEM particles while keeping the inter-physics communication cost low. As shown in figure 7, the test case is of the form proposed in [19] with a domain size of 480mm x 40mm x 480 mm, featuring a layered bed of 10 million of particles moving under the action of an incompressible flow. The 10 (a) (b) Figure 7: Ten million particles in one million cells: Setup. 11 boundary conditions for the fluid solution are of constant Dirichlet at the inlet (2m/s), non-slip at the wall boundaries and reference atmospheric pressure at the outlet. In [19] was pointed out that with heavy coupled cases, when partitioning the two domains independently, the inter-physics communication becomes more and more important with the rise of computing processes, reaching up to 30% of the whole cost of the simulation when using more than 200 computing processes. This means that, when using an independent domain partitioning from each software, the coupled simulation will be negatively affected by the coupling interface, leading to a coupled execution that will perform worse than the single-code one when operated on a high number of computing nodes. This behavior was also identified in [10] where the authors clearly showed how the coupled execution was performing worse than both of the standalone software. 3.4.1. Standalone XDEM and standalone OpenFOAM performance ● 280 ● Ideal ● XDEM using METIS ● XDEM using SCOTCH ● 196 Speedup XDEM using Uniform ● XDEM using Zoltan RIB ● ● 112 ● ● ● 28 28 56 84 112 140 168 196 224 252 280 Number of cores Figure 8: Ten million particles in one million cells: Speedup of standalone XDEM for different partitioning algorithms. ● 280 ● Ideal ● OpenFOAM using SCOTCH ● OpenFOAM using Uniform ● Speedup 196 ● ● ● 112 ● ● 28 ● 28 56 84 112 140 168 196 224 252 280 Number of cores Figure 9: Ten million particles in one million cells: Speedup of standalone OpenFOAM for different partitioning algorithms. In figure 8, the scalability performance of the pure DEM execution obtained with different partitioning algorithms is proposed. One can observe how the results are affected by the choice partitioning algorithm, but the general behavior 12 is rather similar. This can be explained considering that the setup for this case is rather homogeneous and uniformly distributed. Similarly, in figure 9, the performance of the pure CFD execution obtained with a uniform partitioning and a classic OpenFOAM partitioner (SCOTCH) are compared. Once more, the trends are rather similar and only minor differences can be noticed. 3.4.2. Coupled OpenFOAM–XDEM performance ● 280 ● ● Ideal OpenFOAM using Uniform 196 ● XDEM using Uniform ● Speedup OpenFOAM−XDEM coupled ● ● ● 112 ● ● 28 ● 28 56 84 112 140 168 196 224 252 280 Number of cores Figure 10: Ten million particles in one million cells: Speedup of the coupled OpenFOAM–XDEM approach in comparison with standalone XDEM and OpenFOAM: The coupled execution performs better than the standalone XDEM even for more than 200 processes. ● % of Total Execution time 0.075 ● 0.050 0.025 0.000 28 56 ● ● 84 112 ● ● 140 168 ● ● ● ● 196 224 252 280 Number of cores Figure 11: Ten million particles in one million cells: Percentage of time spent for inter-physics data exchange for an increasing number of processes. Less than 0.1% of the time is spent in the coupling part and the proposed solution scales well. In figure 10, we evaluate the performance of our coupled OpenFOAM–XDEM approach in comparison with standalone XDEM and standalone CFD, all executed with a uniform partitioning. It is interesting to notice how the parallel execution of the coupled code is not worse than the one of the single code when operating with more than 200 processes. This can be explained by the repartition of the computational cost: more than 92% is spent in the XDEM part, less than 8% in OpenFOAM and the inter-physics data exchanges represented less than 0.1% as shown 13 on figure 11. It appears that, in this case, the cost of the coupling is negligible and the behavior of the coupled solution is driven by the most computation intensive part of the code, namely XDEM. This represents a major improvement compared to previous work. In particular, instead of rising as it was happening for [19], the inter-physics data exchange reduces in magnitude when using our co-located partitions strategy. This can be explained considering that no inter-process communication is involved, and the amount of the data exchanged per each process reduces. In this way, instead of becoming one of the limiting factors for the scalability of a coupled execution, this communication benefits from being executed in parallel when a co-located partitions strategy is adopted. This represents an important advantage for large-scale applications that need to run over hundreds of processes in order to produce results within a reasonable computational time. 4. Discussion The main novelty of the current approach presented in section 2.3 is to introduce a new partitioning strategy for CFD and DEM domains. This strategy aims to co-locate partitions from the CFD and DEM domain that will need to perform information exchange, in the same MPI process. This is radically different from what previously proposed in the literature, where most of the works focused on optimizing the load-balancing of each software independently from each other while we focus first on reducing the inter-physics communication. We implemented this strategy by defining a partitioning algorithm that forces CFD and DEM to be co-located. In the cases proposed in this paper, the co-location constraint is completely fulfilled, meaning that all the information required for the inter-physics exchanges are present within the same MPI process. Therefore, the inter-physics interprocess communications are completely absent. As shown in section 3.4, this strategy allows overcoming the interphysics communication bottleneck that was identified in [19]. Nevertheless, adopting a co-location constraint introduces a significant limitation in the partitioning flexibility of the two domains. First, an hybrid partitioning algorithm must be adopted that can optimize the executions of the two codes. In uniform cases, as the one proposed in this work, a uniform partitioning algorithm allows achieving reasonably good parallel performances. On the other hand, in cases in which the DEM and CFD parts have significantly different parallel requirements (as, for instance, when the computational load of the DEM part is clustered in a specific region of the domain, while the CFD load is uniform) a better partitioning algorithm that optimizes the coupled execution must be designed. We want to point out how, for more complex partitioning algorithms, a perfect fulfillment of the co-location constraint might be difficult or not possible. In this cases, one might consider relaxing the co-location constraint allowing the partitions to be non-perfectly coinciding. This will require the introduction of an inter-physics interpartition communication layer that will cause a communication overhead with respect to the cases here presented. This cost will be null for perfectly coinciding partitions, and maximum for completely non-coinciding partitions. Therefore, an optimal partitioning algorithm for CFD–DEM couplings must consider both the cost associated with intra and inter-physics communication in order not to be excessively penalized by any bottleneck. Alternatively, in several cases, it might not be always necessary to have such a complex hybrid partitioning algorithm. When one of the software is clearly dominating in term of computational load, it could be sufficient to partition its domain with one of its native algorithms, and decide the partitions of the second software to match the decision of the dominant one, respecting the co-location principle proposed in this work. 5. Conclusion A partitioning strategy for CFD–DEM couplings has been investigated. It consists of the usage of co-located partitions for the CFD and the DEM domains. This allows performing the bulk information exchange between the two codes locally, reducing the global communication time and in particular minimizing the inter-physics parallel communication. This strategy was implemented by using a simple uniform partitioning algorithm that forces the partitions to be co-located. Three benchmark cases have been introduced in order to assess the consistency and performances of the proposed strategy. Experimental results were carried out using the coupling between the XDEM framework and the OpenFOAM libraries. The validity of the results have been assessed by comparing sequential and parallel 14 executions, and scalability performance over hundreds of processes has been reported showing that less than 0.1% of the time is used for inter-physics data exchange. The main advantage of the proposed strategy consists in the reduction of the cost associated with the inter-physics communication, that is a fundamental step toward the large-scale computation, as it allows maintaining good parallel performances when operating with hundreds of computing processes. This results proved how the inter-physics communication must be taken into account when choosing a partitioning algorithm for CFD–DEM couplings in order to avoid an important communication bottleneck. In order to improve the applicability our work to more generic scenarios, an enhanced flexibility in the domain partitioning can be provided. For this, dedicated partitioning algorithms must be designed to find the good tradeoff between a good load-balancing within each software, their own communication requirements and the volume of inter-physics data exchange. This will be studied in future works. Acknowledgement This research is in the framework of the project DigitalTwin, supported by the programme Investissement pour la compétitivité et emploi - European Regional Development Fund (Grant agreement: 2016-01-002-06). The experiments presented in this paper were carried out using the HPC facilities of the University of Luxembourg1 . [1] Amit Amritkar, Surya Deb, and Danesh Tafti. Efficient parallel cfd-dem simulations using openmp. Journal of Computational Physics, 256:501 – 519, 2014. [2] O.O. Ayeni, C.L. Wu, K. Nandakumar, and J.B. Joshi. Development and validation of a new drag law using mechanical energy balance approach for DEM–CFD simulation of gas–solid fluidized bed. Chemical Engineering Journal, 302:395–405, 2016. [3] Mehdi Baniasadi, Maryam Baniasadi, and Bernhard Peters. Coupled cfd-dem with heat and mass transfer to investigate the melting of a granular packed bed. Chemical Engineering Science, 178:136 – 145, 2018. [4] S. Buis, A. Piacentini, and D. Declat. Palm: A computational framework for assembling high-performance computing applications. Concurrency and Computation: Practice and Experience, 18(2):231–245, 2006. [5] D. Darmana, N.G. Deen, and J.A.M. Kuipers. Parallelization of an eulerlagrange model using mixed domain decomposition and a mirror domain technique: Application to dispersed gasliquid two-phase flow. Journal of Computational Physics, 220(1):216 – 248, 2006. [6] Wei Du, Xiaojun Bao, Jian Xu, and Weisheng Wei. Computational fluid dynamics (CFD) modeling of spouted bed: Assessment of drag coefficient correlations. Chemical Engineering Science, 61(5):1401–1420, 2006. [7] H. J. Herrmann. Physics of granular media. Chaos, Solitons & Fractals, 6:203–212, 1995. [8] L. Jing, C. Y. Kwok, Y. F. Leung, and Y. D. Sobral. Extended CFD–DEM for free-surface flow with multi-size granules. International Journal for Numerical and Analytical Methods in Geomechanics, 40(1):62–79, 2016. NAG-14-0182.R1. [9] Wolfgang Joppich and M Kürschner. Mpccia tool for the simulation of coupled applications. Concurrency and computation: Practice and Experience, 18(2):183–192, 2006. [10] D.K. Kafui, S. Johnson, C. Thornton, and J.P.K. Seville. Parallelization of a lagrangianeulerian dem/cfd code for application to fluidized beds. Powder Technology, 207(1):270 – 278, 2011. [11] Amir Houshang Mahmoudi, Florian Hoffmann, Miladin Markovic, Bernhard Peters, and Gerrit Brem. Numerical modeling of self-heating and self-ignition in a packed-bed of biomass using {XDEM}. Combustion and Flame, 163:358–369, 2016. [12] Algirdas Maknickas, Arnas Kačeniauskas, Rimantas Kačianauskas, Robertas Balevičius, and Algis Džiugys. Parallel dem software for simulation of granular media. Informatica, 17(2):207–224, April 2006. [13] Peters, Bernhard and Pozzetti, Gabriele. Flow characteristics of metallic powder grains for additive manufacturing. EPJ Web Conf., 140:13001, 2017. [14] Gabriele Pozzetti and Bernhard Peters. Evaluating Erosion Patterns in an abrasive water jet cutting using XDEM. Advances in Powder Metallurgy & Particulate Materials, pages 191–205, 2017. [15] Gabriele Pozzetti and Bernhard Peters. On the choice of a phase interchange strategy for a multiscale DEM-VOF Method. AIP Conference Proceedings, 1863, 2017. [16] Gabriele Pozzetti and Bernhard Peters. A multiscale dem-vof method for the simulation of three-phase flows. International Journal of Multiphase Flow, 99:186 – 204, 2018. [17] Alban Rousset, Xavier Besseron, and Bernhard Peters. Parallelizing xdem: Load-balancing policies and efficiency, a study. International Conference on Particle-Based Methods (PARTICLES17), 2017. [18] L. Schiller and Z. Naumann. A Drag Coefficient Corre-lation,. VDI Zeitung, Vol. 77,:318–320, 1935. [19] Rui Sun and Heng Xiao. SediFoam: A general-purpose, open-source CFD–DEM solver for particle-laden flow with emphasis on sediment transport. Computers & Geosciences, 89:207–219, 2016. [20] S. Varrette, P. Bouvry, H. Cartiaux, and F. Georgatos. Management of an Academic HPC Cluster: The UL Experience. In Proc. of the 2014 Intl. Conf. on High Performance Computing & Simulation (HPCS 2014), pages 959–967, Bologna, Italy, July 2014. IEEE. 1 http://hpc.uni.lu 15 [21] David W. Washington and Jay N. Meegoda. Micro-mechanical simulation of geotechnical problems using massively parallel computers. International Journal for Numerical and Analytical Methods in Geomechanics, 27(14):1227–1234, 2003. [22] H. G. Weller, G. Tabor, H. Jasak, and C. Fureby. A tensorial approach to computational continuum mechanics using object-oriented techniques. Computers in Physics, 12(6):620–631, 1998. [23] Heng Xiao and Jin Sun. Algorithms in a Robust Hybrid CFD-DEM Solver for Particle-Laden Flows. Communication Computer Physics, 2010. 16
5cs.CE
arXiv:1702.07448v3 [math.ST] 30 Nov 2017 Optimal Bayesian Minimax Rates for Unconstrained Large Covariance Matrices Kyoungjae Lee1 and Jaeyong Lee2 1 Department of Applied and Computational Mathematics and Statistics, The University of Notre Dame 2 Department of Statistics, Seoul National University December 4, 2017 Abstract We obtain the optimal Bayesian minimax rate for the unconstrained large covariance matrix of multivariate normal sample with mean zero, when both the sample size, n, and the dimension, p, of the covariance matrix tend to infinity. Traditionally the posterior convergence rate is used to compare the frequentist asymptotic performance of priors, but defining the optimality with it is elusive. We propose a new decision theoretic framework for prior selection and define Bayesian minimax rate. Under the proposed framework, we obtain the optimal Bayesian minimax rate for the spectral norm for all rates of p. We also considered Frobenius norm, Bregman divergence and squared log-determinant loss and obtain the optimal Bayesian minimax rate under certain rate conditions on p. A simulation study is conducted to support the theoretical results. Key words: Bayesian minimax rate; Convergence rate; Decision theoretic prior selection; Unconstrained covariance. 1 Introduction Estimating covariance matrix plays a fundamental role in multivariate data analysis. Many statistical methods in multivariate data analysis such as the principle component analysis, canonical 1 correlation analysis, linear and quadratic discriminant analysis require the estimated covariance matrix as the starting point of the analysis. In the risk management and the longitudinal data analysis, the covariance matrix estimation is a crucial part of the analysis. The log-determinant of covariance matrix is used for constructing hypothesis test or quadratic discriminant analysis [2]. Suppose we observe a random sample Xn = (X1 , . . . , Xn ), Xi ∈ Rp , i = 1, . . . , n, from the p-dimensional normal distribution with mean zero and covariance matrix Σ, i.e. iid X1 , . . . , Xn | Σ ∼ Np (0, Σ). We assume the zero mean and focus on the covariance matrix. With advance of technology, data arising from various areas such as climate prediction, image processing, gene association study, and proteomics, are often high dimensional. In such high dimensional settings, it is often natural to assume that the dimension of the variable p tends to infinity as the sample size n gets larger, i.e. p = pn −→ ∞ as n −→ ∞. This assumption can be justified as follows. First, when p is large in comparison with n, often the limiting scenario with p tending to infinity approximates closer to the reality than that with p fixed. Second, in many cases we can postulate the reality is infinitely complex and involves infinitely many variables, and with limited resources and time, we can collect only a portion of variables and observations. If we have more resources to collect more data, it is natural to collect more observations as well as more variables, i.e. to increase both n and p. When p tends to infinity as n −→ ∞, the traditional covariance estimator is not optimal [32]. The sparsity or bandable assumptions on large matrices have been used frequently in the literature. Many researchers have studied the large sample properties under the restrictive matrix classes. [6] considered the bandable covariance/precision classes and studied the convergence rate of banding estimator on those classes. [44] derived the convergence rate for precision matrices via sparse Cholesky factors and showed that it is the minimax rate under the Frobenius norm. In addition, the minimax convergence rates for the sparse or bandable covariance matrices were established by [11], [12, 13] and [45]. For a comprehensive review on the convergence rate for the covariance and precision matrices, see [10]. The posterior convergence rate has been investigated by [36], [4], and [21]. [36] showed that their continuous shrinkage priors are optimal for the sparse covariance estimation under the 2 spectral norm in the sense that the posterior convergence rate is quite close to the frequentist √ minimax rate. They achieved a nearly minimax rate upto a log n term under the spectral norm and sparse assumption even when n = o(p). [4] considered Bayesian banded precision matrix estimation using graphical models. They obtained the posterior convergence rate of the precision matrix under matrix `∞ norm when log p = o(n). [21] developed a prior distribution for the sparse PCA and showed that it achieves the minimax rate under the Frobenius norm. They also derived the posterior convergence rate under the spectral norm. Most of the previous works on the Bayesian estimation of large covariance matrix concentrate on the constrained covariance or precision matrix. To the best of our knowledge, only [22] considered asymptotic results for large unconstrained covariance matrix under the “large p and large n” setting. However, they attained the Bernstein-von Mises theorems under somewhat restrictive assumptions on the dimension p. In this paper, we fill the gap in the literature. At first, we propose a new decision theoretic framework to define Bayesian minimax rate. The posterior convergence rate is the primary concept when the asymptotic optimality is studied in the Bayesian sense. But it is not completely satisfactory. The following is a quote from [24] which they write just after defining the posterior convergence rate. ‘We defined “a” rather than the rate of contraction, and hence logically any rate slower than a contraction rate is also a contraction rate. Naturally we are interested in a fastest decreasing sequence n , but in general this may not exist or may be hard to establish. Thus our rate is an upper bound for a targeted rate, and generally we are happy if our rate is equal to or close to an “optimal” rate. With an abuse of terminology we often make statements like “n is the rate of contraction.” ’ In the proposed new decision theoretic framework, a probability measure on the parameter space is an action and a prior is a decision rule for it gives a probability measure (the posterior) for a given data set. In this setup, we define the convergence rate and the Bayesian minimax rate. We investigate the Bayesian minimax rates for unconstrained large covariance matrix. We consider four losses for the covariance inference: spectral norm, Frobenius norm, Bregman divergence and squared log-determinant loss. For the spectral norm, we have the complete result of the Bayesian minimax rate. We show that the Bayesian minimax rate is min(p/n, 1) for all rates 3 of p. For the Frobenius norm and Bregman divergence, we show the Bayesian minimax lower √ bound is p · min(p, n)/n for all rates of p, but obtained the upper bound under the constraint √ √ p ≤ n. Thus, under the condition p ≤ n, the Bayesian minimax rate is p2 /n. We also show that the Bayesian minimax rate under the squared log-determinant loss is p/n when p = o(n). The rest of the paper is organized as follows. In section 2, we define the model, the covariance classes we consider, and introduce some notations. We propose the new decision theoretical framework and define the Bayesian minimax rate. The Bayesian minimax rates under the spectral norm, the Frobenious norm, the Bregman matrix divergence, and the squared log-determinant loss are presented in section 3. A simulation study is given in section 4. The discussion is given in section 5, and the proofs are given in Supplementary Material ([35]). 2 Preliminaries 2.1 The Model and the Inverse-Wishart Prior Suppose we observe a random sample from the p-dimensional normal distribution X1 , · · · , Xn | Σn iid ∼ Np (0, Σn ), (1) where Σn is a p × p positive definite matrix, and p is a function of n such that p = pn −→ ∞ as n −→ ∞. The true value of the covariance matrix is denoted by Σ0 or Σ0n , which is dependent on n. For the prior of the covariance matrix Σn in model (1), we consider the inverse-Wishart prior Σn ∼ IWp (νn , An ), (2) where νn > p − 1, An is a p × p positive definite matrix for a proper prior. The mean of Σn is An /(νn − p − 1). The condition νn > p − 1 is needed for the distribution to have a density in the space of p × p positive definite matrices. If νn is an integer with νn ≤ p − 1, (2) defines a singular distribution on the space of p × p positive semidefinite matrices [43]. We also consider the truncated inverse-Wishart prior. The inverse-Wishart prior with parameter ν and A whose eigenvalues are restricted in [K1 , K2 ] with 0 < K1 < K2 is denoted by IWp (ν, A, K1 , K2 ). The truncated inverse-Wishart prior was adopted for technical reason. By Lemma E.1, to connect the Frobenius norm with Bregman matrix divergence, the eigenvalues 4 of argument matrices have to be bounded. The truncated inverse-Wishart prior guarantees that the posterior covariance matrix has bounded eigenvalues. 2.2 Matrix Norms and Notations We define the spectral norm (or matrix `2 norm) for matrices by kAk := sup kAxk2 , kxk2 =1 P where k · k2 denotes the vector `2 norm defined by kxk2 := ( pi=1 x2i )1/2 , x = (x1 , . . . , xp )T ∈ Rp p and A is p × p matrix. The spectral norm is the same as λmax (AT A) or λmax (A) if A is symmetric, where λmax (B) denotes the largest eigenvalue of B. The Frobenius norm is defined by  1 2 p X p X 2 kAkF :=  aij  , i=1 j=1 where A = (aij ) is a p × p matrix. It is the same as p tr(AT A), where tr(B) denotes the trace of B. The Frobenius norm is the vector `2 norm with p × p matrices treated as p2 -dimensional vectors. The Bregman divergence [7] is originally defined for vectors, but it can be extended to the real symmetric matrices. Let φ be a differentiable and strictly convex function that maps real symmetric p × p matrices to R. The Bregman divergence with φ between two real symmetric matrices is defined as Dφ (A, B) := φ(A) − φ(B) − tr[(∇φ(B))T (A − B)], where A and B are real symmetric matrices and ∇φ is the gradient of φ, i.e., ∇φ(B) = (∂φ(B)/∂Bi,j ). In this paper, we consider a class of φ such that φ(X) = Pp i=1 ϕ(λi ) where ϕ is a differentiable and strictly convex real-valued function and λi ’s are the eigenvalues of A. Furthermore, we assume that ϕ satisfies the following properties for some constant τ1 > 0: (i) ϕ is a twice differentiable and strictly convex function over λ ∈ (τ1 , ∞); (ii) there exist some constants C > 0 and r ∈ R such that |ϕ(λ)| ≤ Cλr for all λ ∈ (τ1 , ∞); and 5 (iii) for any positive constants τ > τ1 , there exist some positive constants ML and MU such that ML ≤ ϕ00 (λ) ≤ MU for all λ ∈ [τ1 , τ ]. The above class of Bregman matrix divergences includes the squared Frobenius norm, von Neumann divergence and Stein’s loss. For their use in statistics and mathematics, see [13], [18] and [34]. If ϕ(λ) = λ2 , the Bregman divergence is the squared Frobenius norm Dφ (A, B) = kA−Bk2F . If ϕ(λ) = λ log λ−λ, it is the von Neumann divergence Dφ (A, B) = tr (A log A − A log B − A + B) , where log A is the matrix logarithm, i.e., A = V DV T is mapped to log A = V log DV T . Here, D = diag(di ) is a p × p diagonal matrix where di is the ith eigenvalue of A, and V = [V1 , · · · , Vp ] is a p × p orthogonal matrix where Vi is an eigenvector of A corresponding to the eigenvalue di . If ϕ(λ) = − log λ, the Bregman divergence is the Stein’s loss Dφ (A, B) = tr(AB −1 ) − log det(AB −1 ) − p. The Stein’s loss is the Kullback-Leibler divergence between two multivariate normal distributions with means zero and covariance matrices A and B, respectively. Finally, we introduce some notations for asymptotic analysis which will be used subsequently. For any positive sequences an and bn , we say an  bn if there exist positive constants c and C such that c ≤ an /bn ≤ C for all sufficiently large n. We define an = o(bn ), if an /bn → 0 as n → ∞ and an = O(bn ), if there exist positive constants N and M such that |an | ≤ M |bn | for all d n ≥ N . For any random variables Xn and X, Xn −→ X means the convergence in distribution. For any real symmetric matrix A, A > 0 (A ≥ 0) means that the matrix A is positive definite (nonnegative definite). We denote δA as the dirac measure at A. 2.3 A Class of Covariance Matrices Let Cp denote the set of all p × p covariance matrices. For any positive constants τ, τ1 and τ2 , define the class of covariance matrix C(τ ) = Cp (τ ) := {Σ ∈ Cp : kΣk ≤ τ, Σ ≥ 0}, C(τ1 , τ2 ) = Cp (τ1 , τ2 ) := {Σ ∈ Cp : λmin (Σ) ≥ τ1 , kΣk ≤ τ2 }, where λmin (Σ) is the smallest eigenvalue of Σ. Throughout the paper, we consider the model (1) and assume that the true covariance matrix belongs to C(τ ) or C(τ1 , τ2 ). 6 Often the subgaussian property is used to relax the Gaussian distribution assumption. The distribution of random vector X has subgaussian property with variance factor τ > 0, if 2 /(2τ ) P (|v T (X − EX)| > t) ≤ e−t for all t > 0 and kvk = 1. The subgaussian property with variance factor τ implies kVar(X)k ≤ 2τ . In the literature, the subgaussian distribution is frequently used as a basic assumption, for examples, [11], [12, 13] and [45]. If X follows a multivariate normal distribution, kΣk ≤ τ is a sufficient condition for X to have the subgaussian property. 2.4 Decision Theoretic Prior Selection Let d(Σ, Σ0 ) be a pseudo-metric that measures the discrepancy between two covariance matrices Σ and Σ0 . A sequence n −→ 0 is called a posterior convergence rate at the true parameter Σ0 if for any Mn −→ ∞, π(d(Σ, Σ0 ) ≥ Mn n | Xn ) −→ 0 in PΣ0 -probability as n −→ ∞. The convergence rate is measured by the rate of n , which allows that the posterior contraction probability converges to zero in probability PΣ0 , where PΣ0 is iid the distribution for random sample (X1 , . . . , Xn ) ∼ Np (0, Σ0 ). In the literature, the posterior is said to achieve the minimax rate if its convergence rate is the same as the frequentist minimax rate ([36]; [21]; [29]). Since the posterior convergence rate cannot be faster than the frequentist minimax rate ([28]), it is often called the optimal rate of posterior convergence ([40]; [38]). However, its definition is elusive as the quote from [24] indicates. As an alternative framework for the evaluation of the prior and the posterior, we take a frequentist decision theoretical approach. For each n, the parameter space is Cp and the action space is the set of all probability measures on Cp . After the data Xn is collected, the posterior π(·|Xn ) is computed for the given prior π and the posterior takes a value in the action space. In this setup, the prior can be considered as a decision rule, because the prior and observations together produce the posterior. A probability measure in the action space will be used as a posterior for the inference, but it does not have to be generated from a prior. We define the loss 7 and risk function of the parameter Σ0 and the prior π as L(Σ0 , π(·|Xn )) := Eπ d(Σ, Σ0 )|Xn ), R(Σ0 , π) := EΣ0 L(Σ0 , π(·|Xn )) = EΣ0 Eπ d(Σ, Σ0 )|Xn ). Note that the risk function measures the performance of the prior π. To distinguish them from the usual loss and risk, we call the above loss and risk as posterior loss (P-loss) and posterior risk (P-risk). The P-risk itself is not new. For example, the P-risk was also used in [14] for density estimation on the unit interval. There are a couple of benefits of the proposed decision theoretic prior selection. First, the decision theoretic prior selection makes the definition of the minimax rate of the posterior mathematically concrete. Although the minimax rate of the posterior is used frequently, it has been used without a rigorous definition. The frequentist minimax rate is used as a proxy of the desired concept. Second, in the study of the posterior convergence rate, the scale of the loss function needs to be carefully chosen so that the posterior consistency holds. But in the proposed decision theoretic prior selection, the inconsistent priors can be compared without any conceptual difficulty. Thus, the scale of the loss function does not need to be chosen. We now define the minimax rate and convergence rate for P-loss. Let Πn be the class of all priors on Σn . A sequence rn is said to be the minimax rate for P-loss (P-loss minimax rate) or simply the Bayesian minimax rate for the class Cp∗ ⊂ Cp and the space of the prior distributions Π∗n ⊂ Πn , if inf sup EΣ0 L(Σ0 , π(·|Xn ))  rn . π∈Π∗n Σ0 ∈C ∗ p A prior π∗ is said to have a convergence rate for P-loss (P-loss convergence rate) or convergence rate an , if sup EΣ0 L(Σ0 , π ∗ (·|Xn )) . an , Σ0 ∈Cp and, if an  rn where rn is the minimax rate for P-loss, π ∗ is said to attain the minimax rate for P-loss or the Bayesian minimax rate. If it is clear from context, we will drop P-loss and refer them as the minimax rate and the convergence rate. For a given inference problem, we wish to find a prior π ∗ which attains the minimax rate for P-loss. Remark The P-loss convergence rate implies the posterior convergence rate by Proposition A.1 in Supplementary Material ([35]). By obtaining the P-loss convergence rate, we also get the 8 traditional posterior convergence rate. The converse may not be true, because for certain loss functions, the P-loss may not even converge to 0 while the posterior convergence rate converges to 0. Remark The P-loss convergence rate is slower than or equal to the frequentist minimax rate by Proposition A.2 in Supplementary Material ([35]). To obtain a P-loss minimax lower bound, the mathematical tools for frequentist minimax lower bound can be used. Remark If we assume that the prior class Πn includes the data dependent priors, the P-loss minimax rate is the same as the frequentist minimax rate. Take π = δΣ̂∗ where Σ̂∗ is an estimator attaining the frequentist minimax rate. Then, π attains the frequentist minimax rate and thus attains the Bayesian minimax rate. However, the data-dependent prior is not acceptable for legitimate Bayesian analysis unless the prior is dependent on ancillary statistics. Even if Πn does not contain data-dependent priors, in most cases the frequentist and P-loss minimax rates are the same. However, if we consider a restricted class of priors, the P-loss minimax rate might differ from the usual frequentist minimax rate. In such cases, the frequentist minimax rate will not be a natural concept to study the asymptotic properties of the posterior. See Remark in subsection 3.2. 3 Bayesian Minimax Rates under Various Matrix Loss Functions 3.1 Bayesian Minimax Rate under Spectral Norm In this subsection, we show that the Bayesian minimax rate for covariance matrix under the spectral norm is min(p/n, 1). We also show that the prior  n n πn (Σn ) = IWp (Σn | νn , An )I p ≤ + δIp (Σn )I p > 2 2  (3) attains the Bayesian minimax rate for the class C(τ1 , τ2 ) under the spectral norm, where IWp (Σ | νn , An ) is the inverse-Wishart distribution, νn > p − 1 and An is a p × p positive definite matrix. We have the complete result for all values of n and p. The Bayesian minimax rate holds for any 9 n and p, regardless of their relationship. The number 1/2 in the prior (3) can be replaced by any number in (0, 1) and the prior still renders the minimax rate. The main result of the section is given in Theorem 3.1 whose proof is given in Supplementary Material ([35]). We divide the proof into two parts: lower bound and upper bound parts. First, we show that the lower bound of the frequentist minimax rate is min(p/n, 1), which may be of interest in its own right, and it in turn implies that min(p/n, 1) is a Bayesian minimax lower bound. After that, the P-loss convergence rate with the prior (3) is derived, which is the same as the Bayesian minimax lower bound when νn2 = O(np) and An = Sn . Consequently, we obtain the following theorem by combining these two results. Throughout the paper, Πn is the class of all priors on Σn ∈ Cp as we have defined in subsection 2.4. Theorem 3.1 Consider the model (1). For any positive constants τ1 < τ2 , inf sup π∈Πn Σ0 ∈C(τ1 ,τ2 ) EΣ0 Eπ (kΣn − Σ0 k2 | Xn )  min p n  ,1 . Furthermore, the prior (3) with νn2 = O(np) and kAn k2 = O(np) attains the Bayesian minimax rate. Remark The proof for the lower bound holds even for τ1 and τ2 depending on n and possibly for τ1 −→ 0 and τ2 −→ ∞ as n −→ ∞. In such cases, the rate of the minimax lower bound is τ22 · min (p/n, 1). For details, see Theorem B.1 in the Supplementary Material ([35]). Note that τ2 affects the minimax lower bound, while τ1 does not. A similar phenomenon occurs for estimation of sparse spiked covariance matrices. See Theorem 4 of [10]. We have complete results of the Bayesian minimax rate under the spectral norm. In words, the results above do not have any condition on the rate of p and n. For a given rate of p, we obtained the Bayesian minimax rate. When p grows the same rate as n, the above theorem shows that estimating the covariance under the spectral norm is hopeless. Indeed, this can be seen from the form of the prior (3). When p ≥ n/2, the point mass prior δIp gives the Bayesian minimax rate. In words, you can not do better than the useless point mass prior δIp . Applying techniques used in the proof of the upper bound, one can show that the prior (3) also gives the same P-loss convergence rate for precision matrix. 10 Corollary 3.2 Consider the model (1) and prior (3) with νn2 = O(np) and kAn k2 = O(np). For any positive constants τ1 < τ2 , sup Σ0 ∈C(τ1 ,τ2 ) −1 2 EΣ0 Eπ (kΣ−1 n − Σ0 k | Xn ) ≤ c · min p n ,1  for all sufficiently large n and some constant c > 0. We remark here that [22] derived a posterior convergence rate for unconstrained covariance matrix under the spectral norm when p = o(n). In this paper, we obtained a P-loss convergence rate which implies the stronger convergence than a posterior convergence rate, for any n and p. [22] also attained a posterior convergence rate for precision matrix under p2 = o(n). In this paper, Corollary 3.2 gives a P-loss convergence rate for any n and p. 3.2 Bayesian Minimax Rate under Frobenius Norm Throughout this subsection, τ > 0 can depend on n and possibly τ −→ ∞ as n −→ ∞. In this subsection, we show that the rate of the Bayesian minimax lower bound for covariance matrix √ under Frobenius norm is τ 2 · min(p, n) · p/n for the class C(τ ), and the inverse-Wishart prior √ attains the Bayesian minimax lower bound when p ≤ n. The following theorem gives the Bayesian minimax lower bound. The proof of Theorem 3.3 is given in Supplementary Material ([35]). In the proof of the theorem, we prove that the lower √ bound of the frequentist minimax rate is τ 2 · min(p, n) · p/n as a by-product. Theorem 3.3 Consider the model (1). For any τ > 0, inf sup EΣ0 Eπ (kΣn − Σ0 k2F | Xn ) ≥ c · τ 2 · π∈Πn Σ0 ∈C(τ ) √ p · min(p, n) n for all sufficiently large n and some constant c > 0. Theorem 3.4 Consider the model (1) and prior (2) with νn > 0 and An > 0 for all n. If νn = p and kAn k2 = O(n), for any τ > 0, sup EΣ0 Eπ (kΣn − Σ0 k2F | Xn ) ≤ c · τ 2 · Σ0 ∈C(τ ) p2 n for some constant c > 0 and all sufficiently large n. Furthermore, if p ≤ √ n, νn2 = O(np) and kAn k2 = O(np) is the necessary and sufficient condition for achieving the rate p2 /n. 11 Note that if τ > 0 is a fixed constant, from the relationship between the spectral norm and Frobenius norm, one can obtain a P-loss convergence rate min(p, n) · p/n instead of p2 /n in Theorem 3.4. However, in this case, one should restrict the parameter space to C(τ1 , τ2 ) instead of the more general parameter space C(τ ). In practice, we recommend using νn = p and small An such as An = Op or An = Ip , where Op denotes a p × p zero matrix because it guarantees the rate p2 /n regardless of the relation between n and p. Note that the Jeffreys prior [31] π(Σn ) ∝ det(Σn )−(p+2)/2 , the independence-Jeffreys prior [42] π(Σn ) ∝ det(Σn )−(p+1)/2 and the prior proposed by [23] π(Σn ) ∝ det(Σn )−p satisfy the above conditions. They can be viewed as inverse-Wishart priors, IW (νn , An ), with parameters (1, Op ), (0, Op ) and (p − 1, Op ), respectively. Furthermore, the IW (p + 1, Sn ) prior, whose mean is Sn , also satisfies the conditions in Theorem 3.4. By Theorem 3.4 and Theorem 3.3, we have the Bayesian minimax rate τ 2 ·p2 /n for covariance √ matrix under the Frobenius norm when p ≤ n. Thus, with the inverse-Wishart prior, we attain the Bayesian minimax rate under the Frobenius norm. Theorem 3.5 Consider the model (1). If p ≤ inf √ n, for any τ > 0, sup EΣ0 Eπ (kΣn − Σ0 k2F | Xn )  τ 2 · π∈Πn Σ0 ∈C(τ ) p2 . n Furthermore, νn2 = O(np) and kAn k2 = O(np) is the necessary and sufficient condition for the √ prior (2) to achieve the Bayesian minimax rate when p ≤ n. Remark In section 2.4, we have said that the Bayesian minimax rate can be different from the frequentist minimax rate when a restricted prior class is considered, and that the frequentist minimax rate will not be a natural concept to address the asymptotic properties of the posteriors 12 from a restricted prior class. We give an example here. Consider a prior class Π∗n = {π ∈ √ IWp (νn , An ) : νn ≥ n, An ∈ Cp } and assume p ≤ n. It is easy to check that inf sup EΣ0 Eπ (kΣn − Σ0 k2F | Xn )  τ 2 · p, π∈Π∗n Σ0 ∈C(τ ) from the proof of Theorem 3.4. Note that the obtained P-loss minimax rate differs from the usual frequentist minimax rate, τ 2 · p2 /n. 3.3 Bayesian Minimax Rate under Bregman matrix Divergence In this section, we obtain the Bayesian minimax rate under a certain class of Bregman matrix divergences. Let Φ be the class of differentiable and strictly convex real-valued functions satisfying (i)-(iii) conditions in the subsection 2.2, and let DΦ be the class of Bregman matrix divergences P Dφ where φ(X) = pi=1 ϕ(λi ) for symmetric matrix X and ϕ ∈ Φ. To achieve the Bayesian minimax convergence rate for Bregman matrix divergences, we use the truncated inverse-Wishart distribution IWp (νn , An , K1 , K2 ) whose eigenvalues are all in [K1 , K2 ] for some positive constants K1 < K2 . The density function of IWp (νn , An , K1 , K2 ) is given by 1 π n,K1 ,K2 (Σn ) = −1 det(Σn )−(ν+p+1)/2 e− 2 tr(An Σn ) I(Σn ∈ C(K1 , K2 )) R 0−1 1 0 −(ν+p+1)/2 e− 2 tr(An Σn ) dΣ0 n C(K1 ,K2 ) det(Σn ) (4) where νn > p − 1 and An is a p × p positive definite matrix. Theorem 3.6 Consider the model (1). If p ≤ inf sup π∈Πn Σ0 ∈C(τ1 ,τ2 ) √ n, for any positive constants τ1 < τ2 EΣ0 Eπ (Dφ (Σn , Σ0 ) | Xn )  p2 n for all Dφ ∈ DΦ . Furthermore, the prior (4) with νn2 = O(np), kAn k2 = O(np), K1 < τ1 and √ K2 > τ2 achieves the Bayesian minimax rate when p ≤ n. To extend the minimax result for the squared Frobenius norm to the Bregman matrix divergence, the posterior distribution for Σn and the true covariance Σ0 should be included in the class C(K1 , K2 ) and C(τ1 , τ2 ), respectively, for some positive constants K1 < τ1 and K2 > τ2 . The truncated inverse-Wishart prior was needed to restrict the posterior distribution for Σn within the class C(K1 , K2 ). In practice, we recommend using sufficiently small K1 and large K2 . 13 According to the above theorem, the minimax convergence rate for the class DΦ is equivalent to that for the Frobenius norm if we consider the parameter space C(τ1 , τ2 ). Moreover, the truncated inverse-Wishart prior IWp (νn , An , K1 , K2 ) achieves the Bayesian minimax rate. The proof of the theorem is given in Supplementary Material ([35]). 3.4 Bayesian Minimax Rate of Log Determinant of Covariance Matrix In this subsection, we establish the Bayesian minimax rate for the log-determinant of the covariance matrix under squared error loss. The frequentist minimax lower bound was derived by [9]. We prove that the inverse-Wishart prior achieves the Bayesian minimax rate when p = o(n). The estimator of the log-determinant of the covariance matrix can be used as a basic ingredient for constructing hypothesis test or the quadratic discriminant analysis [2]. The logdeterminant of the covariance matrix is needed to compute the quadratic discriminant function for multivariate normal distribution 1 1 − log det Σ − (x − µ)T Σ−1 (x − µ) 2 2 where x is the random sample from Np (µ, Σ). Furthermore, the differential entropy of Np (µ, Σ) is given by p p log(2π) log det Σ + + , 2 2 2 so the estimation of the differential entropy is equivalent to estimation of the log-determinant of the covariance matrix, when we consider the multivariate normal distribution. The differential entropy has various applications including independent component analysis (ICA), spectroscopy, image analysis, and information theory. See [5], [19], [30] and [16]. [9] showed that the minimax rate for the log-determinant of the covariance matrix under squared error loss is p/n and their estimator achieves this optimal rate when p = o(n). On the Bayesian side, [41] and [26] suggested a Bayes estimator for log-determinant of the covariance matrix of the multivariate normal. They proposed using the inverse-Wishart prior and showed that the posterior mean minimizes expected Bregman divergence. In this subsection, we support their argument by showing that the inverse-Wishart prior achieves the P-loss minimax rate for log-determinant of the covariance matrix under squared error loss. Thus, we show that the inverse-Wishart prior gives the optimal result in the Bayesian sense. We also show the sufficient conditions for achieving the Bayesian minimax rate. The following theorem presents 14 the Bayesian minimax rate for the log-determinant of the covariance matrix under the squared error loss. The proof of the theorem is given in Supplementary Material ([35]). Theorem 3.7 Consider the model (1). If p = o(n), we have inf sup EΣ0 Eπ ((log det Σn − log det Σ0 )2 | Xn )  π∈Πn Σ0 ∈Cp p . n Furthermore, prior (2) with νn2 = O(n/p) and An = Op attains the Bayesian minimax rate. Remark One can also show that the optimal minimax convergence rate is achieved by using the prior (2) with νn2 = O(n/p), An = cn Sn and c2n = O(n/p). Remark [22] showed the Bernstein-von Mises result for the log-determinant of covariance, which implies a posterior convergence rate. However, they considered a restrictive parameter space C(τ1 , τ2 ) and the stronger condition p3 = o(n). In this paper, the more general parameter space Cp and weaker condition p = o(n) are sufficient for the stronger result, a P-loss convergence rate. 4 Simulation study In this section, we support our theoretical results by a simulation study. The simulations for three loss functions, spectral norm, square of scaled Frobenius norm and squared log-determinant loss, were conducted. We compare the performance of the minimax priors with those of some frequentist estimators. We choose the posterior mean as a Bayesian estimator. The posterior mean obtained from the minimax prior attains the minimax rate in Theorem B.2, Theorem 3.4 and Theorem 3.7 by the Jensen’s inequality. We generated dataset X1 , . . . , Xn from Np (0, Σ0 ) where true covariance matrix Σ0 was either diagonal or full covariance matrix. A full covariance matrix is a covariance matrix which does not have any restriction on its elements such as sparsity or banding. In the diagonal covariance iid setting, the true covariance is Σ0 = diag(σ0,ii ) where σ0,ii ∼ U nif (0, 5). In the full covariance setting, we made the true covariance Σ0 = V T V where V = (vij ) is a p × p matrix with iid vij ∼ N (0, 5/p). In the simulation study, the dimensions of the true covariance matrices are 25, 50, 100 and 200, and the numbers of data n are either n = p2 or n = dp3/2 e. For each 15 setting, we generated a true covariance once for which we generated 100 data sets and calculated estimators of the covariance. For the spectral norm and square of scaled Frobenius norm loss, we computed the posterior p mean of the inverse-Wishart prior, IW (νn , An ), for comparison. We chose νn = 2, n/p, p and n to see the effect of the νn , but fixed An = Op to remove the prior effect on the structure of the covariance estimate. By Theorems B.2 and 3.4, when n = p2 , the inverse-Wishart prior with p νn = 2, n/p and p are minimax priors, while that with νn = n is not. We also computed the sample covariance Sn and the tapering estimator Σ̂k [11] for comparison. As mentioned before, the sample covariance matrix is a Bayesian estimator using inverse-Wishart prior with νn = p+1 √ and An = Op , which satisfies the conditions in Theorem 3.4. We used k = n as the threshold of tapering estimator. It corresponds to α = 0 in [11], which gives the minimal sparse constraint for the covariance matrix in their class. Figure 1 summarizes the simulation results for the spectral norm. Each point of the plot was calculated by 100 1 X b (s) k kΣ0 − Σ n 100 s=1 b (s) where Σ n is the estimate of the true covariance Σ0 in s-th simulation. The first and second rows of Figure 1 show the results when the true covariance matrix is a diagonal and full covariance, respectively; the left and right columns are the results when n = p2 and n = dp3/2 e, respectively. The inverse-Wishart prior with νn = p and the sample covariance performed well in all cases. They are either the best or comparable to the best. When n = dp3/2 e, the truncated inverseWishart prior with νn = n is not minimax, and the simulation results show that it performed p the worst or the second to the worst. The inverse-Wishart priors with νn = 2 and n/p are minimax, and thus their risks decrease as n −→ ∞ in all cases, but their performance are slightly worse than that with νn = p. The tapering estimator Σ̂k performed the best in diagonal settings because it gives zero to many of upper and lower diagonal elements or shrink them toward zero. However, in the full covariance settings, it performed the worst or close to the worst for the same reason. Figure 2 summarizes the simulation results for Frobenius norm. Each point of the plot was 16 Figure 1: The risks for the Bayes estimator with IW (νn , Op , K), the sample covariance Sn and tapering estimator Σ̂k under the spectral norm loss function. The true covariances were generated in diagonal setting (top row) and full covariance setting (bottom row). The number of the observation was chosen by either n = p2 (left column) or n = dp3/2 e (right column). calculated by 100 1 X1 b (s) k2 kΣ0 − Σ n F 100 p s=1 where b (s) Σ n is the estimate of the true covariance Σ0 in s-th simulation. The results are quite similar to the spectral norm case. For the square of log-determinant loss, we chose the maximum likelihood estimator (MLE) log det Sn and the uniformly minimum variance unbiased estimator (UMVUE) for comparison. 17 Figure 2: The risks for the Bayes estimator with IW (νn , Op , K), the sample covariance Sn and tapering estimator Σ̂k under the squared Frobenius norm loss function. The true covariances were generated in diagonal setting (top row) and full covariance setting (bottom row). The number of the observation was chosen by either n = p2 (left column) or n = dp3/2 e (right column). The UMVUE of log det Σ is given by log det Sn + p log n 2 − p−1 X j=0  ψ n−k 2  where ψ is the digamma function which is defined by ψ(x) = d/dz log Γ(z)|z=x where Γ is the gamma function. See [1] for more details. We tried the same settings for inverse-Wishart prior as p before. Note that for n = p2 and n = dp3/2 e, the choices νn = 2 and n/p satisfy the sufficient condition in Theorem 3.7 while νn = p and n do not. The posterior mean of the log-determinant 18 Figure 3: The squared log-determinant loss function plot. The true covariances were generated in diagonal setting (top row) and full covariance setting (bottom row). The number of the observation was chosen by either n = p2 (left column) or n = dp3/2 e (right column). for the inverse-Wishart prior is    p−1  n X n + νn − k An + p log − ψ . log det Sn + n 2 2 j=0 Thus, the UMVUE is the same as the Bayesian estimator using inverse-Wishart prior with νn = 0 and An = Op , which satisfies the sufficient condition in Theorem 3.7. Figure 3 summarizes the simulation results for log-determinant. Each point of the plot was calculated by 100 (s) 1 X \ (log det Σ0 − log det Σn )2 100 s=1 where (s) \ log det Σn is the estimate of log det Σ in s-th simulation and Σ0 is the true covariance. 19 The top and bottom rows are for the diagonal and full true covariance cases, respectively; the left and right columns are for n = p2 and dp3/2 e, respectively. For the squared log-determinant loss, the inverse-Wishart priors with νn = 2 and p n/p are minimax, while those with νn = p and n are not. The UMVUE or the Bayes estimator of the the inverse-Wishart priors with νn = 0 performed the best in all cases. The inverse-Wishart priors p with νn = 2 and n/p performed comparable to the UMVUE. Interestingly, the inverse-Wishart priors with νn = p, which was the best under the spectral norm, performed worst in all cases. When n = dp3/2 e, the results for νn = p do not appear in the Figure 3 because of its large risk values. This signifies the fact that we need to choose different prior parameter for different loss function. 5 Discussion In this paper, we develop a new framework for the Bayesian minimax theory, and introduce Bayesian minimax rate and P-loss convergence rate. The proposed decision theoretic framework gives an alternative way to distinguish the good priors from the inadequate ones and makes the definition of the minimax rate of the posterior clear. We obtain the Bayesian minimax rates for the normal covariance model under the various loss functions: spectral norm, the squared Frobenius norm, Bregman matrix divergence and squared log-determinant loss for large covariance estimation. We show that the inverse-Wishart prior or truncated inverse-Wishart prior attains the Bayesian minimax rate. The simulation results support the theory obtained. A Basic properties of P-loss convergence rate A frequentist minimax lower bound is defined as a lower bound of inf sup EΣ0 (d(Σ̂, Σ0 )) Σ̂ Σ0 ∈Cp where Σ̂ denotes an arbitrary estimator of Σ0 , and we say rn is the frequentist minimax rate for the class Cp and the space of the estimators of Σ0 , if inf sup EΣ0 (d(Σ̂, Σ0 ))  rn . Σ̂ Σ0 ∈Cp 20 Propositions A.1 and A.2 state two basic properties of P-loss convergence rate and the Bayesian minimax rate. Proposition A.1 For any Σ0 ∈ Cp , a P-loss convergence rate at Σ0 is a posterior convergence rate at Σ0 . Proof Suppose that the rate of the P-loss convergence rate at Σ0 ∈ Cp is n , i.e., EΣ0 Eπ (d(Σ, Σ0 )|Xn )  n . For a sequence Mn −→ ∞ and δ > 0, PΣ0 (π(d(Σ, Σ0 ) ≥ Mn n | Xn ) > δ) PΣ0 (Eπ (d(Σ, Σ0 ) | Xn ) > δMn n ) 1 ≤ EΣ Eπ (d(Σ, Σ0 ) | Xn ) δMn n 0 −→ 0, as n → ∞. ≤ The first and second inequalities follow from the Markov inequality. Proposition A.2 A frequentist minimax lower bound for Σ0 is also a P-loss minimax lower bound for any loss function d(·, Σ0 ), i.e., inf sup EΣ0 Eπ (d(Σ, Σ0 ) | Xn ) ≥ inf sup EΣ0 (d(Σ̂, Σ0 )), π∈Πn Σ0 ∈Cp Σ̂ Σ0 ∈Cp where Σ̂ denotes an arbitrary estimator of Σ0 . Proof Note that the P-risk is always equal or larger than the posterior convergence rate by Markov’s inequality, and the frequentist minimax rate is a lower bound for the posterior convergence rate ([28]). Thus, the frequentist minimax rate is also a lower bound for the P-loss minimax rate. B Proof of Theorem 3.1 We divide the proof of Theorem 3.1 into two parts: the lower bound part (Theorem B.1) and the upper bound part (Theorem B.2). For Theorem B.1, we have a quite strong result in sense that it holds even for τ1 and τ2 depending on n and possibly τ1 −→ 0 and τ2 −→ ∞ as n −→ ∞. 21 Theorem B.1 Consider the model (1). For any positive constants τ1 < τ2 , for both fixed p and p −→ ∞ as n −→ ∞, inf sup π∈Πn Σ0 ∈C(τ1 ,τ2 ) EΣ0 Eπ (kΣn − Σ0 k2 | Xn ) ≥ c · τ22 · min p n  ,1 for all sufficiently large n and some constant c > 0. Theorem B.2 Consider the model (1) and prior (3) with νn2 = O(np) and kAn k2 = O(np). For any positive constants τ1 < τ2 , sup Σ0 ∈C(τ1 ,τ2 ) EΣ0 Eπ (kΣn − Σ0 k2 | Xn ) ≤ c · min p n  ,1 for all sufficiently large n and some constant c > 0. B.1 Proof of Theorem B.1 Lemma B.3-B.5 are used to prove Theorem B.1. The proofs of Lemma B.3 and Lemma B.4 are straightforward, and they are omitted here. Lemma B.3 Let fi be the density function of p-dimensional Np (0, Σi ), i = 0, 1, 2. If Σ−1 1 + −1 Σ−2 2 − Σ0 is a positive definite matrix, Z Rp f1 f2 −1/2 . dx = [det(Ip − Σ−2 0 (Σ1 − Σ0 )(Σ2 − Σ0 ))] f0 √ Lemma B.4 Define U := {u ∈ Rp : ui = ±1/ p, i = 1, · · · , p}. For any u, v ∼ U nif (U), d hu, vi ≡ 2B/p − 1 where B ∼ Bin(p, 1/2). Lemma B.5 Let P0 , P1 ∈ P where P is a set of all probability measures on X and let f0 R and f1 be their density functions, respectively. Define ξ = ξ(P0 , P1 ) := X f12 /f0 dx and set θi = θ(Pi ), i = 0, 1, where θ is a functional defined on P. Then inf max{E0 (δ − θ0 )2 , E1 (δ − θ1 )2 } ≥ δ (θ1 − θ0 )2 , (1 + ξ 1/2 )2 where δ denotes any estimator of θ and Ei represents the expectation with respect to Pi , i = 0, 1. 22 Proof For given estimator δ which satisfies R(δ, θ0 ) = E|δ(X) − θ|2 ≤ 2 , we have R(δ, θ1 ) ≥ (|θ1 − θ0 | − ξ 1/2 )2 by [8]. Choose  = |θ1 − θ0 |/(1 + ξ 1/2 ) so that 2 = (|θ1 − θ0 | − ξ 1/2 )2 . If E0 (δ − θ0 )2 ≤ 2 , we have max{E0 (δ − θ0 )2 , E1 (δ − θ1 )2 } ≥ E1 (δ − θ1 )2 ≥ 2 = (θ1 − θ0 )2 (1 + ξ 1/2 )2 If E0 (δ − θ0 )2 ≥ 2 , we have max{E0 (δ − θ0 )2 , E1 (δ − θ1 )2 } ≥ E0 (δ − θ0 )2 ≥ 2 = (θ1 − θ0 )2 . (1 + ξ 1/2 )2 Hence, inf max{E0 (δ − θ0 )2 , E1 (δ − θ1 )2 } ≥ δ (θ1 − θ0 )2 1 (1 + ξ 2 )2 . Proof of Theorem B.1 It suffices to show that inf sup Σ̂n Σ0 ∈C(τ1 ,τ2 ) EΣ0 kΣ̂n − Σ0 k2 ≥ c0 · τ22 · min p n ,1  for some constant c0 > 0 because by the Jensen’s inequality, inf sup π∈Πn Σ0 ∈C(τ1 ,τ2 ) EΣ0 Eπ (kΣn − Σ0 k2 | Xn ) ≥ inf sup π∈Πn Σ0 ∈C(τ1 ,τ2 ) ≥ inf sup Σ̂n Σ0 ∈C(τ1 ,τ2 ) EΣ0 kΣ̃n − Σ0 k2 EΣ0 kΣ̂n − Σ0 k2 , where Σ̃n := Eπ (Σn | Xn ). Assume n ≥ p and define n o √ u = (u1 , . . . , up ) ∈ Rp : ui = ±1/ p, i = 1, · · · , p , n o  τ2  Θ := Σ ∈ Rp×p : Σ = Ip + uuT , u ∈ U 1+ U := p with  = c p/n ≤ 1 for some small c > 0 satisfying  ≤ τ2 /τ1 −1. Let P0n = N (0, 2−1 τ2 Ip )n , P1n = P 2−p Σ∈Θ N (0, Σ)n and let f0n and f1n be their density functions, respectively. Note that kΣk = 23 τ2 and λmin (Σ) = (1 + )−1 τ2 ≥ τ1 for any Σ ∈ Θ, thus Θ ⊂ C(τ1 , τ2 ) for some small c > 0. By the above Lemma B.5, inf sup Σ̂n Σ0 ∈C(τ1 ,τ2 ) EΣ0 kΣ̂n − Σ0 k2 ≥ inf sup ≥ inf sup Σ̂n Σ0 ∈C(τ1 ,τ2 ) δ Σ0 ∈C(τ1 ,τ2 ) ≥ inf max EΣ0 (kΣ̂n k − kΣ0 k)2 EΣ0 (δ − kΣ0 k)2 EΣ0 (δ − kΣ0 k)2 δ Σ0 ∈{2−1 τ2 Ip }∪Θ ≥ inf max(Ef0n (δ − (1 + )−1 τ2 )2 , Ef1n (δ − τ2 )2 ) δ τ22 2 , 4(1 + ξ 1/2 )2 R where δ denotes any estimator of kΣ0 k and ξ := (f1n )2 /f0n . The fourth inequality follows from Z 2 inf max EfΣ (δ − kΣk) = inf max (δ(x) − τ2 )2 fΣn (x)dx δ Σ∈Θ δ Σ∈Θ Z 1 X ≥ inf p (δ(x) − τ2 )2 fΣn (x)dx δ 2 Σ∈Θ Z = inf (δ(x) − τ2 )2 f1n (x)dx ≥ δ = inf Ef1n (δ − τ2 )2 δ where fΣn is the density function of N (0, Σ)n . Now we calculate ξ. Z (f1n )2 ξ = f0n P Z (2−p Σ∈Θ fΣn )2 = f0n X Z fΣn fΣn 1 1 2 = 22p f0n Σ1 ,Σ2 ∈Θ X Z fΣ fΣ n 1 1 2 = 22p f0 Σ1 ,Σ2 ∈Θ = 1 X det[(Ip − 2 uuT vv T )]−n/2 22p = 1 X (1 − 2 (uT v)2 )−n/2 22p u,v∈U u,v∈U = E(1 − 2 hu, vi2 )−n/2 ≤ E(exp(2n2 hu, vi2 )), where u, v ∼ U nif (U). The fifth equality is derived from Lemma B.3. We will show that ξ ≤ C for some constant C > 0 for all sufficiently large n. If p does not grow to infinity, i.e., p ≤ C for 24 some constant C > 0, the last term bounded above easily, E(exp(2n2 hu, vi2 )) ≤ exp(2c2 p) ≤ d √ √ d exp(2c2 C). If p tends to infinity, by the Lemma B.4, note that phu, vi ≡ p(2B/p − 1) −→ N (0, 1) as n −→ ∞ where B ∼ Bin(p, 1/2). Note also that we have "  2 #! 2 1 E exp cp B−1 −→ E(exp(cZ 2 )) = √ p 1 − 2c by Theorem 1 of [33] for 0 < c < 1/2, Z ∼ N (0, 1). In our setting, consider Fp as the distribution p function of p(2B/p − 1)2 . Thus, we get the followings by taking  = c p/n for some small c > 0 such that 2c < 1/2, " ≤ ξ −→  E exp 2cp √ 2 B−1 p 2 #! 1 , 1 − 4c as n −→ ∞. Hence, we have inf sup EΣ0 kΣ̂n − Σ0 k2 ≥ Σ̂n Σ0 ∈C(τ1 ,τ2 ) p τ22 2 ≥ c0 · τ22 · n 4(1 + ξ 1/2 )2 for some c0 > 0 which proves the lower bound when n ≥ p. Now, assume n < p and define n o 1 Un := u ∈ Rn : ui = ± √ , i = 1, . . . , n n         Σn 0 τ2 T   In + uu , u ∈ Un . Θ := Σ= : Σn =   1+ 0 Ip−n Earlier result shows that inf sup EkΣ̂n − Σ0 k2 ≥ inf max EkΣ̂n − Σ0 k2 Σ̂n Σ0 ∈Θ Σ̂n Σ0 ∈C(τ1 ,τ2 ) ≥ c0 · τ22 · n = c0 · τ22 n for some c0 > 0. B.2 Proof of Theorem B.2 Lemma B.6 Let Ωn ∼ Wp (νn , νn−1 An ) with νn > p and positive definite matrix An , for all n ≥ 1 and kAn k ≤ τn for all sufficiently large n. Then, there exist positive constants c1 and c2 such that   2 2 P(kΩn − An k ≥ x) ≤ 5p e−c1 νn x /τn + e−c2 νn x/τn 25 for all x > 0. Proof There exist vj with kvj k2 = 1 for j = 1, . . . , 5p , such that kAk ≤ 4 · sup |vjT Avj | j≤5p for any p × p symmetric matrix A (Page 2141 of [11]). Thus, we have P(kΩn − An k ≥ x) ≤ P(kAn kkA−1/2 Ωn A−1/2 − Ip k ≥ x) n n ≤ P(kAn−1/2 Ωn A−1/2 − Ip k ≥ x/τn ) n ! ≤ P 4 · sup |vjT (An−1/2 Ωn An−1/2 − Ip )vj | ≥ x/τn j≤5p   ≤ 5p sup π |vjT (A−1/2 Ωn A−1/2 − Ip )vj | ≥ x/(4τn ) n n j≤5p   2 2 ≤ 5p e−c1 νn x /τn + e−c2 νn x/τn . −1/2 The last inequality follows from Lemma 2.4 and Theorem 3.2 of [39] because An −1/2 Ω n An ∼ Wp (νn , νn−1 Ip ). Lemma B.7 Let Ωn ∼ Wp (νn , νn−1 Ip ) with cνn ≥ p for some constant 0 < c < 1. Then , π(λmax (Ωn ) ≥ c1 ) ≤ 2e−νn /2 , √ 2 π(λmin (Ωn ) ≤ c2 ) ≤ 2e−νn (1− p/νn ) /8 for any constant c1 ≥ (2 + p p p/νn )2 and 0 < c2 ≤ (1 − p/νn )2 /4. Proof It follows from Corollary 5.35 in [20], p √ 2 p/νn + t/ νn ) ≤ 2e−t /2 , p √ 2 ≤ 1 − p/νn − t/ νn ) ≤ 2e−t /2 π(λmax (Ωn )1/2 ≥ 1 + (5) π(λmin (Ωn )1/2 (6) for any t ≥ 0. If we choose t = √ νn for (5), it gives the first inequality π(λmax (Ωn ) ≥ (2 + If we choose t = p p/νn )2 ) ≤ 2e−νn /2 . p p √ νn (1 − p/νn − (1 − p/νn )/2) > 0 for (6), it gives the second inequality π(λmin (Ωn ) ≤ (1 − p p/νn )2 /4) ≤ 2e−νn (1− ≤ 2e−νn (1− 26 √ √ p/νn −(1− p/νn )2 /8 . √ p/νn )/2)2 /2 Proof of Theorem B.2 We prove the upper bound for p ≤ n/2 case first. Note that EΣ0 Eπ (kΣn − Σ0 k | Xn ) ≤ EΣ0 Eπ (kΣn − Σ̆n k | Xn ) + EΣ0 kΣ̆n − Σ0 k, (7) where Σ̆n := (nSn + An )/(n + νn ). Consider the first term of right hand side (RHS) of (7). EΣ0 Eπ (kΣn − Σ̆n k | Xn ) h i = EΣ0 Eπ (kΣn − Σ̆n k | Xn )I(kΣ̆n k ≤ C1 and kΣ̆−1 k ≤ C ) 2 n h i + EΣ0 Eπ (kΣn − Σ̆n k | Xn )I(kΣ̆n k > C1 or kΣ̆−1 k > C ) 2 n (8) (9) for any constant C1 and C2 . The integrand of (8) is bounded by Eπ (kΣn − Σ̆n k | Xn )I(kΣ̆n k ≤ C1 and kΣ̆−1 n k ≤ C2 ) −1 ≤ Eπ (kΣn kkΣ−1 n Σ̆n − Ip k | Xn )I(kΣ̆n k ≤ C1 and kΣ̆n k ≤ C2 ) −1 1/2 1/2 −1 ≤ Eπ (kΣn kkΣ̆−1/2 kkΣ̆1/2 n n Σn Σ̆n − Ip kkΣ̆n k | Xn )I(kΣ̆n k ≤ C1 and kΣ̆n k ≤ C2 ) p −1 1/2 −1 ≤ C1 C2 · Eπ (kΣn kkΣ̆1/2 n Σn Σ̆n − Ip k | Xn )I(kΣ̆n k ≤ C1 and kΣ̆n k ≤ C2 ) h i1/2 p ≤ C1 C2 · Eπ (kΣn k2 | Xn )I(kΣ̆n k ≤ C1 and kΣ̆−1 k ≤ C ) 2 n h i1/2 −1 1/2 2 × Eπ (kΣ̆1/2 . n Σn Σ̆n − Ip k | Xn ) To show that h i p EΣ0 Eπ (kΣn − Σ̆n k | Xn )I(kΣ̆n k ≤ C1 and kΣ̆−1 p/n, k ≤ C ) . 2 n 1/2 1/2 it suffices to prove that Eπ (kΣ̆n Σ−1 n Σ̆n − Ip k2 | Xn ) . p/n and Eπ (kΣn k2 | Xn )I(kΣ̆n k ≤ C1 and kΣ̆−1 n k ≤ C2 ) = O(1). Note that −1 1/2 2 Eπ (kΣ̆1/2 n Σn Σ̆n − Ip k | Xn ) Z ∞ −1 1/2 2 ≤ π(kΣ̆1/2 n Σn Σ̆n − Ip k ≥ u | Xn )du + x x Z ∞  √  ≤ 5p e−C3 (n+νn )u + e−C4 (n+νn ) u du + x ≤ x 5p e−C3 (n+νn )x √ √ √ 5p · 2 xe−C4 (n+νn ) x 5p · 2e−C4 (n+νn ) x + + +x C3 (n + νn ) C4 (n + νn ) C42 (n + νn )2 27 (10) for any x > 0 and some positive constants C3 and C4 by Lemma B.6. If we choose x = C5 · p/n for some large C5 > 0, the rate of (10) is p/n. Note that Eπ kΣn k2 | Xn   ≤ Eπ kΣn k2 I(kΣn k > C6 ) | Xn + C62  1/2 ≤ Eπ kΣn k4 | Xn [π (kΣn k > C6 | Xn )]1/2 + C62 , by Hölder’s inequality. One can easily show that Eπ (kΣn k4 | Xn ) is bounded above by p5 up to some constant factor because kΣn k ≤ tr(Σn ) and Eπ (tr(Σn )4 | Xn )I(kΣ̆n k ≤ C1 and kΣ̆−1 n k ≤ C2 ) . p4 . Also note that  −1 π (kΣn k > C6 | Xn ) = π λmin (Σ−1 n ) < C6 | Xn   −1 1/2 | X ) ≤ π λmin (Σ̆n1/2 Σ−1 Σ̆ ) < k Σ̆ kC n n n n 6   −1 1/2 ≤ π λmin (Σ̆n1/2 Σ−1 n Σ̆n ) < C1 C6 | Xn ) √ 2 ≤ 2e−(n+νn )(1− p/(n+νn )) /8 for some constant C6 ≥ C1 · 4(1 − p of (8) is smaller than p/n. p 1/2)−2 by Lemma B.7. Thus, we have shown that the rate Now, we show that the rate of (9) is smaller than p p/n. Note that (9) is bounded by h i EΣ0 Eπ (kΣn − Σ̆n k | Xn )I(kΣ̆n k > C1 or kΣ̆−1 k > C ) 2 n h  i ≤ EΣ0 Eπ (kΣn k | Xn ) + kΣ̆n k I(kΣ̆n k > C1 or kΣ̆−1 n k > C2 ) h i h i ≤ EΣ0 Eπ (kΣn k | Xn )I(kΣ̆n k > C1 ) + EΣ0 Eπ (kΣn k | Xn )I(kΣ̆−1 k > C ) 2 n h i h i + EΣ0 kΣ̆n kI(kΣ̆n k > C1 ) + EΣ0 kΣ̆n kI(kΣ̆−1 n k > C2 ) Since Σ̆n = (nSn + An )/(n + νn ) and Σ0 ∈ C(τ1 , τ2 ), we have   kAn k > C1 PΣ0 (kΣ̆n k > C1 ) ≤ PΣ0 kSn k + n + νn   kAn k = PΣ0 kSn k > C1 − n + νn    kAn k −1 ≤ PΣ0 kS̄n k > τ2 C1 − n + νn −1/2 where S̄n := Σ0 −1/2 Sn Σ0 (11) ∼ Wp (n, n−1 Ip ). Then, (11) is bounded by 2e−n/2 for some constant 28 C1 > 0 by Lemma B.7. Similarly, for some constant C2 ,   n + νn −1 −1 · kSn k > C2 PΣ0 (kΣ̆n k > C2 ) ≤ PΣ0 n   νn  −1  = PΣ0 λmin (Sn ) < 1 + C2   n ν  n C2−1 ≤ PΣ0 λmin (S̄n ) < τ1−1 1 + n √ −n(1− p/n)2 /8 , ≤ 2e by Lemma B.7. It is easy to show that (n + νn )p kΣ̆n k n + νn − p − 1 Eπ (kΣn k | Xn ) ≤ and Z ∞ h i h i EΣ0 kΣ̆n kI(kΣ̆n k > C1 ) = PΣ0 kΣ̆n kI(kΣ̆n k > C1 ) ≥ u du Z0 ∞ = PΣ0 (kΣ̆n k ≥ u)du C1    Z ∞ kAn k du. ≤ PΣ0 kS̄n k ≥ τ2−1 u − n + νn C1 p √ q −1 By applying t = n( τ2 (u − kAn k/(n + νn )) − 1 − p/n) to the tail inequality (5), we have Z ∞  PΣ0 kS̄n k ≥ 1 ZC∞ ≤ 1 ZC∞ ≤ τ2−1   kAn k u− du n + νn √ 2 √ −1 2e−n( τ2 (u−kAn k/(n+νn ))−1− √ 2e−n uC7 /2 p/n) /2 du du C1 √ ≤ √ C1 −√C1 C7 n/2 1 − C1 C7 n/2 e + e C7 n 2C72 n2 for some constant C7 > 0. Also note that h i h  i1/2 EΣ0 kΣ̆n kI(kΣ̆−1 ≤ EΣ0 kΣ̆n k2 · PΣ0 kΣ̆−1 n k > C2 ) n k > C2 h i1/2 √ 2 ≤ EΣ0 kΣ̆n k2 · 2e−n(1− p/n) /8 29 and kAn k2 + 2EΣ0 kSn k2 (n + νn )2 Z ∞  kAn k2 2 2 sup P kS k ≥ u du + n Σ 0 2 n (n + νn ) 0 Z ∞  √ PΣ0 kS̄n k ≥ u/τ2 du C8 + 9 ZC∞ √ 1/4 √ 2 C8 + 2e−n(u / τ2 −1− p/n) /2 du 9 ZC∞ √ C8 + 2e−nC10 u/2 du EΣ0 kΣ̆n k2 ≤ 2 ≤ ≤ ≤ ≤ C9 for some positive constants C8 , C9 and C10 by applying the tail inequality (5). Thus, we have p shown that the rate of (9) is faster than p/n. For the second term of RHS of (7), note that   kAn k . n + νn p Since νn2 = O(np) and kAn k2 = O(np), it is trivial that νn /(n + νn ) . p/n and kAn k/(n + p p νn ) . p/n. One can show that EΣ0 kSn − Σ0 k ≤ EΣ0 kS̄n − Ip k · kΣ0 k . p/n by Lemma B.6. EΣ0 kΣ̆n − Σ0 k ≤ EΣ0 kSn − Σ0 k + νn n + νn EΣ0 kSn k + Furthermore, it is easy to prove that EΣ0 kSn k . 1 because we have proved EΣ0 kΣ̆n k2 . 1. Thus, p we have EΣ0 kΣ̆n − Σ0 k . p/n. For the case p > n/2, we have EΣ0 Eπ (kΣn − Σ0 k | Xn ) = kIp − Σ0 k ≤ kIp k + kΣ0 k = 1 + τ2 which has the same rate with min(p/n, 1). Proof of Theorem 3.2 It suffices to consider the case p ≤ n/2 because the other part is trivial. Note that −1 EΣ0 Eπ kΣ−1 n − Σ0 k | X n    −1 −1 −1 ≤ EΣ0 Eπ kΣ−1 n − Σ̆n k | Xn + EΣ0 kΣ̆n − Σ0 k h   i −1 −1 = EΣ0 Eπ kΣ−1 − Σ̆ k | X I(k Σ̆ k ≤ C ) n 1 n n n h   i −1 −1 + EΣ0 Eπ kΣ−1 − Σ̆ k | X I(k Σ̆ k > C ) n 1 n n n (12) −1 + EΣ0 kΣ̆−1 n − Σ0 k. (14) 30 (13) For the term (12), we have h   i −1 −1 EΣ0 Eπ kΣ−1 − Σ̆ k | X I(k Σ̆ k ≤ C ) n 1 n n n   1/2 ≤ C1 · EΣ0 Eπ kΣ̆n1/2 Σ−1 n Σ̆n − Ip k | Xn p . n by the argument (10) in the proof of Theorem B.2. For the term (13), note that ≤ ≤ . ≤ . h   i −1 −1 EΣ0 Eπ kΣ−1 − Σ̆ k | X I(k Σ̆ k > C ) n 1 n n n h  i  −1 −1 EΣ0 Eπ kΣ−1 k | X + k Σ̆ k I(k Σ̆ k > C ) n 1 n n n   i h  −1 1/2 −1 −1 EΣ0 Eπ kΣ̆1/2 n Σn Σ̆n k | Xn + 1 kΣ̆n kI(kΣ̆n k > C1 ) h i −1 p · EΣ0 kΣ̆−1 n kI(kΣ̆n k > C1 ) h i1/2  1/2 2 −1 p · EΣ0 kΣ̆−1 k · P k Σ̆ k > C 1 Σ0 n n √ 2 p2 · e−n(1− p/n) /16 by Lemma B.7. The last term (14) is bounded above by −1 −1 −1 EΣ0 kΣ̆−1 n − Sn k + EΣ0 kSn − Σ0 k. By the Woodbury formula, it is easy to show that −1 EΣ0 kΣ̆−1 n − Sn k ≤ . ≤ . 1 νn −1 −1 −1 · EΣ0 kSn−1 k + 2 · EΣ0 kSn−1 (A−1 n + n Sn )Sn k n n 1/2   νn 1  −1 −1 −1 2 1/2 + 2 EΣ0 kSn−1 k4 · EΣ0 k(A−1 n + n Sn ) k n n 1/2  1/2 1 νn · EΣ0 kSn k2 + EΣ0 kSn−1 k4 n n νn n and   −1/2 −1/2 −1 k . E kS kkΣ S Σ − I k EΣ0 kSn−1 − Σ−1 n p Σ n 0 0 0 0 h i1/2   1/2 −1/2 −1/2 ≤ EΣ0 kSn−1 k2 · EΣ0 kΣ0 Sn Σ0 − Ip k2 r p . n from the arguments used in the proof of Theorem B.2. 31 C Proof of Theorem 3.3 Before we prove Theorem 3.3, we define the total variation affinity and the L1 -distance between measures. L1 -distance Let P and Q be probability measures with density functions p and q with respect to a σ-finite measure ν, respectively. Let Z kP ∧ Qk := p ∧ q dν be the total variation affinity between P and Q, and Z kP − Qk1 := |p − q| dν be the L1 -distance between P and Q. Lemma C.1 (Assouad’s Lemma) Let the parameter set Θ = {0, 1}k , d be a pseudo-metric and T be any estimator of ψ(θ) based on the observation X from Pθ with θ ∈ Θ. Let H(θ, θ0 ) = Pk i=1 |θi − θi0 |. Then for all s > 0 max 2s Eθ ds (T, ψ(θ)) ≥ θ∈Θ ds (ψ(θ), ψ(θ0 )) k H(θ, θ0 ) 2 H(θ,θ0 )≥1 min min H(θ,θ0 )=1 kPθ ∧ Pθ0 k. For the proof of Assouad’s lemma, see [3]. Lemma C.2 For any p × p symmetric matrix B such that Ip + tB is a positive definite matrix for any t ∈ [0, 1] and kBkF is small, log det(Ip + B) = tr(B) − R where 0 ≤ R ≤ ckBk2F for some positive constant c. Proof of lemma C.2 Using the notation Ip = (eij ) = (e1 , . . . , ep ), let e := vec(Ip ) := (eT1 , . . . , eTp )T ∈ 2 2 2 Rp . In the same way, let b := vec(B) := (bT1 , . . . , bTp )T ∈ Rp . Define a function h : Rp → R by h(vec(A)) := log det(A), for any p × p positive definite matrix A. Then, the Taylor expansion yields 1 log det(Ip + B) = h(e + b) = h(e) + h0 (e)T b + bT h00 (e + tb)b 2 1 = h0 (e)T b + bT h00 (e + tb)b 2 32 for some t ∈ [0, 1], where |bT h00 (e + tb)b| ≤ kbk22 · kh00 (e + tb)k. Note that ∂ ∂A log det(A) = (A−1 )T [37], so h0 (a) = vec((A−1 )T ) and 0 T h (e) b = p X p X eji bij i=1 j=1 = tr(B). We need to prove that −ckbk22 ≤ bT h00 (e + tb)b/2 ≤ 0 for some constant c > 0. Since h(a) = log det(A) is concave on positive definite matrices [15], h00 (a) is a negative semidefinite matrix for all positive definite A. Thus, bT h00 (e + tb)b ≤ 0. Furthermore, kh00 (e + tb)k is a continuous function on t ∈ [0, 1] because Ip + tB is a positive definite matrix for any t ∈ [0, 1]. Thus, kh00 (e + tb)k/2 ≤ c for some constant c > 0 uniformly on t ∈ [0, 1]. Proof of Theorem 3.3 We follow closely the line of a proof in [11]. By the Jensen’s inequality, inf sup EΣ0 Eπ (kΣn − Σ0 k2F | Xn ) ≥ π∈Π Σ0 ∈C(τ ) inf sup EΣ0 kΣ̃n − Σ0 k2F π∈Π Σ0 ∈C(τ ) ≥ inf sup EΣ0 kΣ̂n − Σ0 k2F Σ̂n Σ0 ∈C(τ ) ≥ inf sup EΣ0 kΣ̂n − Σ0 k2F Σ̂n Σ0 ∈A for any A ⊂ C(τ ), where Σ̃n = Eπ (Σ n | Xn ). We show that for any τ > 0 and A ⊂ C(τ ), inf sup EΣ0 kΣ̂n − Σ0 k2F ≥ c · τ2 · Σ̂n Σ0 ∈A √ p · min(p, n) n for some constant c > 0. Note that τ can depend on n and possibly τ −→ ∞ as n −→ ∞. Without loss of generality, we assume τ > 1. Define   h i c1 √ A := Σ(θ) : Σ(θ) = c2 Ip + (θij I(1 ≤ |i − j| < k)) , θij = θji ∈ {0, 1}, i, j = 1, 2, . . . , p , n √ √ where k = min(p, n), c1 = min(1/3, 1/(3 2c4 )) and c2 = τ /(1 + c1 ). The constant c4 > 0 will be defined later. For any Σ(θ) ∈ A, kΣ(θ)k = = = ≤ ≤   c1 sup c2 x Ip + √ (θij I(1 ≤ |i − j| < k) x n kxk=1   c1 T √ (θij I(1 ≤ |i − j| < k) x c2 + sup c2 x n kxk=1   c1 c2 √ (θij I(1 ≤ |i − j| < k) c2 + n   c1 c2 √ (θij I(1 ≤ |i − j| < k) c2 + n 1 c1 c2 c2 + √ k. n T 33 By the definition of k, c1 and c2 , it follows kΣ(θ)k ≤ τ . Thus, we have A ⊂ C(τ ). Note that symmetric and diagonally dominant matrix Σ(θ) = (σij (θ)) , i.e., σii (θ) > p X |σij (θ)|, j6=i is a positive definite. See, for example, [27]. Also note that Σ(θ) − λIp , for all 0 < λ < (1 − 2c1 )c2 is a diagonally dominant matrix, thus, is positive definite. This implies that the minimum eigenvalue of Σ(θ), λmin (Σ(θ)) > λ for all 0 < λ < (1 − 2c1 )c2 , which in turn, implies λmin (Σ(θ)) ≥ (1 − 2c1 )c2 ≥ c2 3 kΣ(θ)−1 k = λmin (Σ(θ))−1 ≤ 3 . c2 because c1 ≤ 1/3. Thus, By Assouad’s lemma, inf sup EΣ0 kΣ̂n − Σ0 k2F ≥ Σ̂n Σ0 ∈A where H(θ, θ0 ) := kΣ(θ) − Σ(θ0 )k2F (2p − k)(k − 1) · · min kPθ ∧ Pθ0 k H(θ, θ0 ) 4 H(θ,θ0 )≥1 H(θ,θ 0 )=1 1 22 min Pp i>j,1≤|i−j|<k 0 |. The first factor of the RHS is given by |θij − θij kΣ(θ) − Σ(θ0 )k2F min H(θ, θ0 ) H(θ,θ0 )≥1 = min ( c√1 cn2 )2 H(θ,θ0 )≥1 P 1≤|i−j|<k (θij H(θ, θ0 ) 0 )2 − θij 2c21 c22 n  2 2 τ c1 = 2 1 + c1 n = 0 ∈ {0, 1} and because θij , θij P 1≤|i−j|<k (θij 0 )2 = 2H(θ, θ 0 ). The second factor of the RHS is − θij of rate kp. The proof of the theorem will be completed, if we show that lim inf min n→∞ H(θ,θ0 )=1 kPθ ∧ Pθ0 k ≥ c3 for some constant c3 > 0. Since kPθ − Pθ0 k1 = 2 − 2kPθ ∧ Pθ0 k, 34 it it suffices to prove, when H(θ, θ0 ) = 1, kPθ − Pθ0 k21 < 1, for all sufficiently large n. Then, we have lim inf min n→∞ H(θ,θ0 )=1 kPθ ∧ Pθ0 k > 1/2. Note that by Pinsker’s inequality [17], kPθ − Pθ0 k21 ≤ 2K(Pθ0 , Pθ )   = n · tr(Σ(θ0 )Σ(θ)−1 ) − log det(Σ(θ0 )Σ(θ)−1 ) − p where K(Pθ0 , Pθ ) := R (15) dP log( dPθθ0 )dPθ0 is the Kullback-Leibler divergence. Define A1 := Σ(θ0 ) − Σ(θ), then (15) can be written as   n · tr(A1 Σ(θ)−1 ) − log det(Ip + A1 Σ(θ)−1 ) h i = n · tr(Σ(θ)−1/2 A1 Σ(θ)−1/2 ) − log det(Ip + Σ(θ)−1/2 A1 Σ(θ)−1/2 ) . Consider the diagonalization of Σ(θ)−1 , Σ(θ)−1 = U V U T where U is an orthogonal matrix and V is a diagonal matrix. Since H(θ, θ0 ) = 1, kΣ(θ)−1/2 A1 Σ(θ)−1/2 k2F = kU V 1/2 U T A1 U V 1/2 U T k2F = kV 1/2 U T A1 U V 1/2 k2F ≤ kV k2 kU T A1 U k2F = kΣ(θ)−1 k2 kA1 k2F 32 2c2 c2 2c2 ≤ 2 · 1 2 = 32 · 1 . n n c2 Note that kΣ(θ)−1/2 A1 Σ(θ)−1/2 k ≤ kΣ(θ)−1/2 A1 Σ(θ)−1/2 k2F ≤ 32 · 2c21 /n ≤ 2/3 for any n ≥ 3 because c1 ≤ 1/3. Then Ip + tΣ(θ)−1/2 A1 Σ(θ)−1/2 is a positive definite matrix for any t ∈ [0, 1] and kΣ(θ)−1/2 A1 Σ(θ)−1/2 k2F is small, so we have log det(Ip + Σ(θ)−1/2 A1 Σ(θ)−1/2 ) = tr(Σ(θ)−1/2 A1 Σ(θ)−1/2 ) − Rn where 0 ≤ Rn ≤ c4 kΣ(θ)−1/2 A1 Σ(θ)−1/2 k2F for some constant c4 > 0 by Lemma C.2. Note that the constant c4 does not depend on c1 as long as c1 ≤ 1/3 and n ≥ 3. Thus, we have kPθ − Pθ0 k21 ≤ nRn √ such that Rn ≤ c4 kΣ(θ)−1/2 A1 Σ(θ)−1/2 k2F for all large n. Since we choose c1 = min(1/3, 1/(3 2c4 )), it completes the proof. 35 D Proof of Theorem 3.4 Proof of Theorem 3.4 Let Σ̃n = (σ̃n,ij ) := Eπ (Σn | Xn ). Note that EΣ0 Eπ (kΣn − Σ0 k2F | Xn ) p X p X = EΣ0 Eπ (σn,ij − σ0,ij )2 | Xn  i=1 j=1 p X p X = π  EΣ0 Var σn,ij | Xn + i=1 j=1 p X p X EΣ0 σ̃n,ij − σ0,ij i=1 j=1 =: T1 + T2 . Let Bn = (bn,ij ) := T1 = ≤ Pn T k=1 Xk Xk p X p X i=1 j=1 p X p X EΣ0  (n + νn − p + 1)b2 + (n + νn − p − 1)bn,ii bn,jj  n,ij (n + νn − p)(n + νn − p − 1)2 (n + νn − p − 3)  EΣ0 i=1 j=1 ≤ = + An . If n + νn − p ≥ 6, we have  2(n + νn − p)bn,ii bn,jj (n + νn − p)(n + νn − p − 1)2 (n + νn − p − 3) p X p   X 8 E b b n,ii n,jj Σ0 (n + νn − p)3 8 (n + νn − p)3 i=1 j=1 p X p X (CovΣ0 (bn,ii , bn,jj ) + EΣ0 bn,ii · EΣ0 bn,jj ) . i=1 j=1 The remaining steps are given by T1 ≤ = = ≤ p X p q  X 8 Var (b ) · Var (b ) + E b · E b n,ii n,jj n,ii n,jj Σ Σ Σ Σ 0 0 0 0 (n + νn − p)3 8 (n + νn − p)3 i=1 j=1 p X p  X  2nσ0,ii · σ0,jj + (nσ0,ii + an,ii ) · (nσ0,jj + an,jj ) i=1 j=1 p p p p X X X X 2 2  8 2 σ (n + 2n) + 2n σ a + a 0,ii 0,ii n,jj n,ii (n + νn − p)3 i=1 i=1 j=1 i=1   8 2 2 2 2 2 2 (n + 2n)p kΣ k + 2np kΣ k · kA k + p kA k . 0 0 n n (n + νn − p)3  Since Σ0 ∈ C(τ ), we have the upper bound for T1 , T1 ≤   8 2 2 2 2 2 2 (n + 2n)τ p + 2np τ kA k + p kA k . n n (n + νn − p)3 36 2 Similar to T1 , we can compute the T2 part by T2 = = p X p X  EΣ0 i=1 j=1 p X p  X 2 bij − σ0,ij n + νn − p − 1 VarΣ0 ( i=1 j=1 h i2  bn,ij bn,ij ) + EΣ0 ( − σ0,ij ) n + νn − p − 1 n + νn − p − 1 p X p  2 +σ h (−ν + p + 1)σ + a i2  X n(σ0,ij 0,ii σ0,jj ) n 0,ij n,ij + = (n + νn − p − 1)2 n + νn − p − 1 i=1 j=1 ≤ + p X p X 2n (σ0,ii σ0,jj ) (n + νn − p − 1)2 2 (n + νn − p − 1)2 i=1 j=1 p X p  X  2 (νn − p − 1)2 σ0,ij + a2n,ij . i=1 j=1 Since kΣ0 k2F ≤ pkΣ0 k2 , T2 ≤ ≤ p  X  2 2 2 2 2 n σ + (ν − p − 1) kΣ k + kA k 0,ii n 0 n F F (n + νn − p − 1)2 i=1   2 2 2 2 2 2 . np kΣ k + (ν − p) pkΣ k + pkA k 0 n 0 n (n + νn − p − 1)2 Thus, the upper bound of the rate for T2 is   4 2 2 2 2 2 . τ np + τ (ν − p) p + pkA k n n (n + νn − p)2 T2 ≤ We have the upper bound of the rate for the P-loss convergence rate sup EΣ0 Eπ (kΣn − Σ0 k2F | Xn ) Σ0 ∈C(τ )   c 2 2 2 2 2 2 n p τ + np kA kτ + p kA k n n (n + νn − p)3   c 2 2 2 2 2 + np τ + (νn − p) pτ + pkAn k (n + νn − p)2 ≤ (16) for some constant c > 0. Now, we get the upper bound sup EΣ0 Eπ kΣn − Σ0 k2F | Xn  Σ0 ∈C(τ ) ≤ c · τ2 · p2 n if we assume νn = p and kAn k2 = O(n). √ If we assume p ≤ n, each term in (16) should be smaller than τ 2 · p2 /n to obtain the minimax rate. Under this condition, νn2 = O(np) and kAn k2 = O(np) is the necessary and sufficient condition to attain the minimax rate τ 2 · p2 /n. 37 E Proof of Theorem 3.6 To obtain the minimax posterior rate of the Bregman divergence, we need the following lemma from [13]. Lemma E.1 Suppose that the eigenvalues of the real symmetric matrices X and Y lie in [τ1 , τ2 ] for some constants 0 < τ1 < τ2 . Then, there exist positive constants c1 < c2 depending on τ1 and τ2 such that c1 kX − Y k2F ≤ Dφ (X, Y ) ≤ c2 kX − Y k2F for all Dφ ∈ DΦ . Proof of Theorem 3.6 Let Σ̆n := (nSn + An )/(n + νn ). Then, EΣ0 Eπ n,K1 ,K2 (Dφ (Σn , Σ0 ) | Xn ) h n,K ,K i  1 2 ≤ C1 · EΣ0 Eπ kΣn − Σ0 k2F | Xn I(Σ̆n ∈ / C(C2 , C3 )) h n,K ,K i  1 2 + C1 · EΣ0 Eπ kΣn − Σ0 k2F | Xn I(Σ̆n ∈ C(C2 , C3 )) (17) (18) for some constant C1 > 0 and any positive constants C2 < C3 by Lemma E.1. Set C2 = τ1 (1−2c̃)2 and C3 = τ2 (1 + 2c̃)2 for some small constant c̃ > 0. Note that (17) is bounded by h n,K ,K i  1 2 EΣ0 Eπ kΣn − Σ0 k2F | Xn I(Σ̆n ∈ / C(C2 , C3 )) h n,K ,K i  1 2 ≤ EΣ0 Eπ p · kΣn − Σ0 k2 | Xn I(Σ̆n ∈ / C(C2 , C3 )) ≤ 2p(K22 + τ 2 )PΣ0 (Σ̆n ∈ / C(C2 , C3 )). Since Σ0 ∈ C(τ1 , τ2 ), PΣ0 (Σ̆n ∈ C(C2 , C3 )) is bounded below by    C3 kAn k νn  C2 ≤ λmin (S̄n ) & λmax (S̄n ) ≤ 1− (19) PΣ0 1+ n τ1 τ2 C3 (n + νn ) √ where S̄n ∼ Wp (n, n−1 Ip ). By applying Corollary 5.35 in [20] with t = c̃ n, (19) is bounded below by 1 − 2e−c̃ 2 n/2 for all sufficiently large n. Thus,   p2 2 . p · PΣ0 Σ̆n ∈ / C(C2 , C3 ) ≤ 2pe−c̃ n/2  n Note that the integrand of (18) is bounded by Eπ Z = ≤ n,K1 ,K2  kΣn − Σ0 k2F | Xn dIW (Σn | n + νn , nSn + An )I(Σn ∈ C(K1 , K2 )) kΣn − Σ0 k2F Rp dΣn 0 0 C(K1 ,K2 ) dIWp (Σn | n + νn , nSn + An )dΣn  1 · Eπ kΣn − Σ0 k2F | Xn π(Σn ∈ C(K1 , K2 ) | Xn ) 38 where π(Σn | Xn ) is a density function of IWp (n + νn , nSn + An ). If we show that π(Σn ∈ C(K1 , K2 ) | Xn )−1 I(Σ̆n ∈ C(C2 , C3 )) ≤ 2 for all sufficiently large n, the rate of (18) is p2 /n by Theorem 3.4. Note that I(Σ̆n ∈ C(C2 , C3 )) π(Σn ∈ C(K1 , K2 ) | Xn ) = I(Σ̆n ∈ C(C2 , C3 )) π(K1 ≤ λmin (Σn ) ≤ λmax (Σn ) ≤ K2 | Xn ) ≤ π(K1 ≤ ≤ I(Σ̆n ∈ C(C2 , C3 )) −1/2 −1/2 −1/2 −1/2 λmin (Σ̆n Σn Σ̆n )λmin (Σ̆n ) ≤ λmax (Σ̆n Σn Σ̆n )λmax (Σ̆n ) 1 π(K2−1 C3 1/2 ≤ 1/2 where Σ̆n Σ−1 n Σ̆n 1/2 1/2 λmin (Σ̆n Σ−1 n Σ̆n ) 1/2 1/2 −1 ≤ λmax (Σ̆n Σ−1 n Σ̆n ) ≤ K1 C2 | Xn ) ≤ K2 | Xn ) , | Xn ∼ Wp (n + νn , (n + νn )−1 Ip ). Note that if K1 < τ1 and K2 > τ2 , we always can find the small constant c̃ > 0 satisfying K1 ≤ C2 (1 + 2c̃)−2 = τ1 {(1 − 2c̃)/(1 + 2c̃)}2 and K2 ≥ C3 (1 − 2c̃)−2 = τ2 {(1 + 2c̃)/(1 − 2c̃)}2 . Then, by applying Corollary 5.35 in [20] with √ 2 t = c̃ n + νn , the last term is bounded above by (1 − 2e−c̃ (n+νn )/2 )−1 for all sufficiently large n. Since (1 − 2e−c̃ F 2 (n+ν n )/2 )−1 ≤ 2 for all sufficiently large n, it completes the proof. Proof of Theorem 3.7 Proof of Theorem 3.7 The minimax lower bound part is given at Theorem 3 of [9], so we prove here the upper bound part only. Let νn2 = O(n/p) and An = Op . Note that if Σ ∼ d Q 2 2 IWp (ν, A), it implies det(AΣ−1 ) ≡ p−1 k=0 χν−k where χν−k ’s are independent chi-square random variables with the degree of freedom ν − k (page 180 of [25]. Then, EΣ0 Eπ ((log det Σn − log det Σ0 )2 | Xn ) π = EΣ0 E ((log det nSn − log det Σ0 − p−1 X log χ2n+νn −k )2 | Xn ). k=0 Define Tn := log det Sn −τn,p , τn,p := Pp−1 k=0  ψ((n−k)/2)−log(n/2) and ψ(x) := d/dz log Γ(z)|z=x . Then, we have   p−1 X EΣ0 Eπ (log det nSn − log det Σ0 − log χ2n+νn −k )2 | Xn k=0 2 ≤ 2 · EΣ0 Tn − log det Σ0 X 2 p−1 2 + 2·E [ψ((n − k)/2) + log 2 − log χn+νn −k ] , k=0 39 (20) (21) where the last expectation is with respect to the chi-square random variables. The first term (20) has the upper bound EΣ0 Eπ  Tn − log det Σ0 2 | Xn   p 10p ≤ −2 log 1 − + . n 3n(n − p) (22) by Theorem 2 of [9]. The RHS of (22) has the asymptotic rate p/n because p = o(n). Using the facts, E(log χ2ν ) = ψ(ν/2) + log 2 and Var(log χ2ν ) = ψ 0 (ν/2), we can separate (21) into two parts:  !2  p−1   X n−k ψ | Xn  EΣ0 Eπ  + log 2 − log χ2n+νn −k 2 k=0 ! p−1 X ≤ 2 · VarΣ0 log χ2n+νn −k  (23) k=0 + 2·   !2 p−1   X n−k n + νn − k ψ . −ψ 2 2 (24) k=0 Note that ψ 0 (ν) = ν −1 + θ(2ν 2 )−1 + θ(6ν 3 )−1 for ν > 1 and 0 < θ < 1 (page 169 of [9]). Applying the above facts to (23), we can show that ! p−1  p−1 X X 2 = VarΣ0 log χn+νn −k k=0 2 2θ 4θ + + 2 n + νn − k (n + νn − k) 3(n + νn − k)3 k=0    p−1  X 1 7 ≤ −2 log 1 − + n + νn − k 3(n + νn − k)2 k=0   7 p p ≤ −2 log 1 − + · n + νn 3 n  (25) for 0 < θ < 1. In the second line, we use the inequality x + θx2 ≤ − log(1 − x) + x2 /2 for 0 < x < 1. Note that the RHS of (25) has the asymptotic rate p/n if p = o(n). For (24), we use the following property of digamma function, ψ(x + 1) − ψ(x) = x−1 . Thus, we have  2 νn e−1   !2 p−1   p−1 d X 2 X X n + νn − k 2 n−k  ψ −ψ ≤  2 2 n − k + 2x k=0 k=0 x=0 !2 νn + 2 ≤ log 1 + n−k−2 k=0   2 νn + 2 ≤ p log 1 + . n−p−2 p−1 X  Note that (26) has the asymptotic rate p/n if νn2 = O(n/p) and p = o(n). 40 (26) Combining (22)-(26), we have  p EΣ0 Eπ ((log det Σn − log det Σ0 )2 | Xn ) ≤ −C1 log 1 − n   2 p νn + 2 2 + C2 · + C3 · p log 1 + n n−p−2 for all sufficiently large n with n > p and some positive constants C1 , C2 and C3 . Since we assume p = o(n) and νn2 = O(n/p), EΣ0 Eπ ((log det Σn − log det Σ0 )2 | Xn ) ≤ c · p n for all sufficiently large n and some constant c > 0. References [1] Nabil Ali Ahmed and DV Gokhale. Entropy expressions and their estimators for multivariate distributions. IEEE Trans. Inform. Theory, 35(3):688–692, 1989. [2] T.W. Anderson. An Introduction to Multivariate Statistical Analysis. Wiley Series in Probability and Statistics. Wiley, 2003. [3] Patrice Assouad. Deux remarques sur l’estimation. C. R. Acad. Sci. Paris Sér. I Math., 296(23):1021–1024, 1983. [4] Sayantan Banerjee and Subhashis Ghosal. Posterior convergence rates for estimating large precision matrices using graphical models. Electron. J. Stat., 8(2):2111–2137, 2014. [5] Jan Beirlant, Edward J Dudewicz, László Györfi, and Edward C van der Meulen. Nonparametric entropy estimation: An overview. Int. J. Math. Stat. Sci., 6(1):17–39, 1997. [6] Peter J Bickel and Elizaveta Levina. Regularized estimation of large covariance matrices. Ann. Statist., 36(1):199–227, 2008b. [7] Lev M. Bregman. The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming. USSR Comput. Math. Math. Phys., 7(3):200–217, 1967. 41 [8] Lawrence D Brown and Mark G Low. A constrained risk inequality with applications to nonparametric functional estimation. Ann. Statist., 24(6):2524–2535, 1996. [9] T Tony Cai, Tengyuan Liang, and Harrison H Zhou. Law of log determinant of sample covariance matrix and optimal estimation of differential entropy for high-dimensional gaussian distributions. J. Multivariate Anal., 137:161–172, 2015. [10] T Tony Cai, Zhao Ren, and Harrison H Zhou. Estimating structured high-dimensional covariance and precision matrices: Optimal rates and adaptive estimation. Electron. J. Stat., 10(1):1–59, 2016. [11] T Tony Cai, Cun-Hui Zhang, and Harrison H Zhou. Optimal rates of convergence for covariance matrix estimation. Ann. Statist., 38(4):2118–2144, 2010. [12] T Tony Cai and Harrison H Zhou. Minimax estimation of large covariance matrices under l1 norm. Statist. Sinica, 22(4):1319–1378, 2012a. [13] T Tony Cai and Harrison H Zhou. Optimal rates of convergence for sparse covariance matrix estimation. Ann. Statist., 40(5):2389–2420, 2012b. [14] Ismaël Castillo. On bayesian supremum norm contraction rates. The Annals of Statistics, 42(5):2058–2091, 2014. [15] Thomas M Cover and A Thomas. Determinant inequalities via information theory. SIAM J. Matrix Anal. Appl., 9(3):384–392, 1988. [16] Thomas M. Cover and Joy A. Thomas. Elements of Information Theory. Wiley-Interscience, New York, NY, USA, 1991. [17] I Csiszár. Information-type measures of difference of probability distributions and indirect observations. Studia Sci. Math. Hungar., 2:299–318, 1967. [18] Inderjit S Dhillon and Joel A Tropp. Matrix nearness problems with bregman divergences. SIAM J. Matrix Anal. Appl., 29(4):1120–1146, 2007. [19] Edward J Dudewicz and Walter Mommaerts. Maximum entropy methods in modern spectroscopy: a review and an empiric entropy approach. In conference proceedings on The 42 frontiers of statistical scientific theory & industrial applications (Vol. II), pages 115–160. American Sciences Press, 1991. [20] Yonina C Eldar and Gitta Kutyniok. Compressed sensing: theory and applications. Cambridge University Press, 2012. [21] Chao Gao and Harrison H Zhou. Rate-optimal posterior contraction for sparse pca. Ann. Statist., 43(2):785–818, 2015. [22] Chao Gao and Harrison H Zhou. Bernstein-von mises theorems for functionals of the covariance matrix. Electronic Journal of Statistics, 10(2):1751–1806, 2016. [23] Seymour Geisser and Jerome Cornfield. Posterior distributions for multivariate normal parameters. J. R. Stat. Soc. Ser. B. Stat. Methodol., 25:368–376, 1963. [24] Subhashis Ghosal and Aad van der Vaart. Fundamentals of Nonparametric Bayesian Inference. Cambridge University Press, 2017. [25] NR Goodman. The distribution of the determinant of a complex wishart distributed matrix. Ann. Math. Statistics, 34(1):178–180, 1963. [26] Maya Gupta and Santosh Srivastava. Parametric bayesian estimation of differential entropy and relative entropy. Entropy, 12(4):818–843, 2010. [27] D.A. Harville. Matrix Algebra From a Statistician’s Perspective. Springer, 2008. [28] N.L. Hjort, C. Holmes, P. Müller, and S.G. Walker. Bayesian Nonparametrics. Cambridge Series in Statistical and Probabilistic Mathematics. Cambridge University Press, 2010. [29] Marc Hoffmann, Judith Rousseau, and Johannes Schmidt-Hieber. On adaptive posterior concentration rates. Ann. Statist., 43(5):2259–2295, 2015. [30] Aapo Hyvärinen. New approximations of differential entropy for independent component analysis and projection pursuit. In Proceedings of the 1997 Conference on Advances in Neural Information Processing Systems 10, NIPS ’97, pages 273–279, Cambridge, MA, USA, 1998. MIT Press. [31] H. Jeffreys. Theory of Probability. Oxford, Oxford, England, third edition, 1961. 43 [32] Iain M Johnstone and Arthur Yu Lu. On consistency and sparsity for principal components analysis in high dimensions. J. Amer. Statist. Assoc., 104(486):682–693, 2009. [33] W Kozakiewicz. On the convergence of sequences of moment generating functions. Ann. Math. Statistics, 18:61–69, 1947. [34] Brian Kulis, Mátyás A Sustik, and Inderjit S Dhillon. Low-rank kernel learning with bregman matrix divergences. J. Mach. Learn. Res., 10:341–376, 2009. [35] Kyoungjae Lee and Jaeyong Lee. Supplementary material for “optimal bayesian minimax rates for unconstrained large covariance matrices”. 2017. [36] Debdeep Pati, Anirban Bhattacharya, Natesh S Pillai, and David Dunson. Posterior contraction in sparse bayesian factor models for massive covariance matrices. Ann. Statist., 42(3):1102–1130, 2014. [37] Kaare Brandt Petersen and Michael Syskind Pedersen. The matrix cookbook. Technical University of Denmark, 7:15, 2008. [38] Veronika Rocková. Bayesian estimation of sparse signals with a continuous spike-and-slab prior. Submitted manuscript, pages 1–34, 2015. [39] L Saulis and VA Statulevic̆ius. Limit Theorems for Large Deviations, volume 73. Springer Science & Business Media, 1991. [40] Weining Shen and Subhashis Ghosal. Adaptive bayesian procedures using random series priors. Scand. J. Stat., 42(4):1194–1213, 2015. [41] Santosh Srivastava and Maya R Gupta. Bayesian estimation of the entropy of the multivariate gaussian. In 2008 IEEE International Symposium on Information Theory, pages 1103–1107. IEEE, 2008. [42] Dongchu Sun and James O Berger. Objective bayesian analysis for the multivariate normal model. Bayesian Statistics, 8:525–547, 2007. [43] Harald Uhlig. On singular wishart and singular multivariate beta distributions. Ann. Statist., 22(1):395–405, 1994. 44 [44] Nicolas Verzelen. Adaptive estimation of covariance matrices via cholesky decomposition. Electron. J. Stat., 4:1113–1150, 2010. [45] Lingzhou Xue and Hui Zou. Minimax optimal estimation of general bandable covariance matrices. J. Multivariate Anal., 116:45–51, 2013. 45
10math.ST
The Complex Negotiation Dialogue Game Romain Laroche Microsoft Maluuba, Montréal, Canada [email protected] arXiv:1707.01450v1 [cs.AI] 5 Jul 2017 Abstract This position paper formalises an abstract model for complex negotiation dialogue. This model is to be used for the benchmark of optimisation algorithms ranging from Reinforcement Learning to Stochastic Games, through Transfer Learning, One-Shot Learning or others. 1 Introduction A negotiation is defined as a bargaining process between two or more parties (each with its own aims, needs, and viewpoints) seeking to discover a common ground and reach an agreement to settle a matter of mutual concern or resolve a conflict. From a dialogue point of view, one distinguishes negotiation dialogue from standard dialogue by the mutual sharing of information1 , by its required user adaptation2 , and by the non-stationarity induced by its non fully cooperative structure: the user and system objectives correlate but also differ to some extent, and they are consequently adversely co-adapting. Research on negotiation dialogue experiences a growth of interest. At first, Reinforcement Learning (Sutton and Barto, 1998), the most popular framework for dialogue management in dialogue systems (Levin and Pieraccini, 1997; Laroche et al., 2009; Lemon and Pietquin, 2012), was applied to negotiation with mitigated results (English and Heeman, 2005; Georgila and Traum, 2011; Lewis et al., 2017), because the non-stationary policy of the opposing player prevents those algorithms from converging consistently. Then, Multi-Agent Reinforcement Learning (Bowling and Veloso, 2002) was applied but also with convergence 1 whereas standard dialogue mainly relies on discovering the user information or intent, 2 whereas standard dialogue, such as form filling applications, is rather indifferent to the user’s characteristics, difficulties (Georgila et al., 2014). Finally, recently, Stochastic Games (Shapley, 1953) were applied successfully (Barlier et al., 2015), with convergence guarantees, but only for zero-sum games, which is inconsistent with dialogue since most tasks are cooperative. Here, we extend (Laroche and Genevay, 2017)’s abstraction of the negotiation dialogue literature applications: (di Eugenio et al., 2000; English and Heeman, 2005) consider sets of furniture, (Afantenos et al., 2012; Efstathiou and Lemon, 2014; Georgila et al., 2014; Litman et al., 2016; Lewis et al., 2017) resource trading, and (Putois et al., 2010; Laroche et al., 2011; El Asri et al., 2014; Genevay and Laroche, 2016; Laroche and Féraud, 2017) appointment scheduling. Indeed, these negotiation dialogue problems are cast into a generic agreement problem over a shared set of options. The goal for the players is to reach an agreement and select an option. This negotiation dialogue game can be parametrised to make it zero-sum, purely cooperative, or general sum. However, (Laroche and Genevay, 2017) only consider elementary options: they are described through a single entity. We formalise in this paper the game for options that are compounded in the sense that they are characterised by several features. For instance, Tuesday morning is defined by two features: the day and the moment of the day. Considering compounded options naturally leads to richer expressions, and therefore to a larger set of actions: I’m available whenever on Tuesday, or I’d prefer in the afternoon. Since the options are uttered in a compounded way, as opposed to their elementary definition in (Laroche and Genevay, 2017), the state representation also becomes more complex. This extension allows more realistic dialogues, and more challenging Reinforcement Learning, Multi-Agent Reinforcement Learning, and Stochastic Games policy training. 2 The Negotiation Dialogue Game This section recalls the negotiation dialogue game as described in (Laroche and Genevay, 2017). The goal for each participant is to reach an agreement. The game involves a set of m players P = {pi }i∈[1,m] . With m > 2, the dialogue game is said to be multi-party (Asher et al., 2016; ?). The players consider n options (in resource trading, an option is an exchange proposal, in appointment scheduling, it is a time-slot), and the cost to agree on an option τ is cτi randomly sampled from distribution δi ∈ ∆R+ to agree on it. Players also have a utility ω i ∈ R+ for reaching an agreement. For each player, a parameter of cooperation with the other players αi ∈ R is introduced. As a result, player pi ’s immediate reward at the end of the dialogue is: X Ri (siT ) = ω i − cτi + αi (ω j − cτj ) (1) j6=i where siT is the last state reached by player pi at the end of the dialogue, and τ is the agreed option. If players fail to agree, the final immediate rewards Ri (siT ) = 0 for all players pi . If at least one player pj misunderstands and agrees on a wrong option τ j which was not the one proposed by the other players, this is even worse: each player pi gets the cost of selecting option τ i without the reward of successfully reaching an agreement: X j Ri (siT ) = −cτi i − αi cτ j (2) j6=i The values of αi give a description of the nature of the players, and therefore of the game as modelled in game theory (Shapley, 1953). If αi < 0, player pi is said to be antagonist: he has an interest in making the other players lose. In particular, if m = 2 and α1 = α2 = −1, it is a zero-sum game. If αi = 0, player pi is said to be selfcentred: he does not care if the other player is winning or losing. Finally, if αi > 0, player pi is said to be cooperative, and in particular, if ∀i ∈ [1, m], αi = 1, the game is said to be fully cooperative because ∀(i, j) ∈ [1, m]2 , Ri (siT ) = Rj (sjT ). From now on, and until the end of the article, we suppose that there are only m = 2 players: a system ps and a user pu . They act each one in turn, starting randomly by one or the other. They have four possible actions. ACCEPT(τ ) means that the user accepts the option τ (independently from the fact that τ has actually been proposed by the other player; if it has not, this induces the use of Equation 2 to determine the reward). This act ends the dialogue. R EF P ROP(τ ) means that the user refuses the proposed option and proposes instead option τ . A SK R EPEAT means that the player asks the other player to repeat his proposition. And finally, E ND D IAL denotes the fact that the player does not want to negotiate anymore, and terminates the dialogue. Understanding through speech recognition of system ps is assumed to be noisy with a sentence error rate SERsu after listening to a user pu : with probability SERsu , an error is made, and the system understands a random option instead of the one that was actually pronounced. In order to reflect human-machine dialogue reality, a simulated user always understands what the system says: SERus = 0. We adopt the way (Khouzaimi et al., 2015) generates speech recognition confidence scores: scorereco = 1 where X ∼ N(c, 0.2) given a user pu , 1+e−X two parameters (c⊥u , c⊤u ) with c⊥u < c⊤u are defined such that if the player understood the right option, c = c⊤u otherwise c = c⊥u . The further apart the normal distribution centres are, the easier it will be for the system to know if it understood the right option, given the score. 3 Allowing compounded options This section extends the negotiation dialogue game recalled in Section 2 with compounded options. Each option τ is now characterised by a set of ℓ features: τ = {fτk }k∈J1,ℓK , with fτk ∈ F k . Not all feature combinations might form a valid option, but for the sake of simplicity, we consider that the set of the n options contain all of them and that the cost for inconsistent ones is infinite. This way, we can express that an option is invalid but that the user is not aware of it. The cost of an option needs to be revisited consequently. The costs of two options that only differ by a feature are similar in general. Without loss of generality, we define the cost of one player pi for agreeing on a given option τ as follows: ciτ = ĉiτ + ℓ X ĉik , (3) k=1 where ĉik is the cost of agreeing on feature f k and ĉiτ is the cost for selecting this option in particular. In an appointment scheduling negotiation task, the feature related costs ĉik can generally be considered as null: there is no correlation between being booked on Monday morning and being available on Tuesday morning. Most of the constraints are therefore expressed in the ĉiτ term. On the opposite, in a furniture set application, the preferences are expressed on specific features of the furniture: colour, price, etc. In this case, the constraints mainly lie in ĉik terms. The option-as-features definition naturally induces new ways of expressing one’s preferences over the option set. The P ROP F EA k k 1 2 TURES(f , f , . . . ) dialogue act replaces the previously defined R EF P ROP(τ ): it means that the speaker wants the k1th , k2th , . . . features to be set to values f k1 , f k2 , . . . A SK R EPEAT still asks to repeat the whole last utterance, but its partial version is added: A SK PARTIAL R EPEAT(k1 , k2 , . . . ) consists in asking to repeat values of features k1 , k2 , . . . ACCEPT still accepts the last grounded option, but it can only be performed once all features have been grounded. Its partial version is also introduced: PARTIAL ACCEPT(k1 , k2 , . . . ) determines an agreement on the last grounded value of features k1 , k2 , . . . The compounded options imply complex actions, which in turn imply a complex understanding model: the sentence level understanding rate and score need to be extended. The sentence error rate SERsu is therefore replaced with a feature error rate F ERsu . The same speech recognition confidence score generation is used at the feature level, meaning that, at each P ROP F EA TURES, A SK PARTIAL R EPEAT, and PARTIAL AC CEPT acts, the player receives an array of feature values (or feature names), each associated with a confidence score. 4 Potential use of the complex negotiation dialogue game (Genevay and Laroche, 2016) already used the simple negotiation dialogue game to study Knowledge Transfer for Reinforcement Learning (Taylor and Stone, 2009; Lazaric, 2012) applied to dialogue systems (Gašic et al., 2013; Casanueva et al., 2015). It appears in this paper that the optimal policies are rather simple. Making the interaction process more intricate and more reality reflecting allows to put the computational tractability of the methods to the test. Following the same purpose, one-shot learning (Schaal et al., 1997; Fei-Fei et al., 2006) may also be used for negotiation dialogues. Cooperative co-adaptation in dialogue has been tackled only in one previous article: (Chandramohan et al., 2012). Similarly, but for the adversary case, the negotiation dialogue game offers a good empirical test bed for a generalisation to the general-sum games of (Barlier et al., 2015). We believe that this line of research is complementary with the more applied one of (Lewis et al., 2017) that work on real human dialogues and are more focused on dealing with natural language within a negotiation task. Their mitigated results indicate that negotiation generalisation over simulated users to real users is difficult, even when the simulated user is trained on human data. References Stergos Afantenos, Nicholas Asher, Farah Benamara, Anaı̈s Cadilhac, Cédric Dégremont, Pascal Denis, Markus Guhe, Simon Keizer, Alex Lascarides, Oliver Lemon, et al. 2012. Developing a corpus of strategic conversation in the settlers of catan. In SeineDial 2012-The 16th Workshop On The Semantics and Pragmatics Of Dialogue. Nicholas Asher, Julie Hunter, Mathieu Morey, Farah Benamara, and Stergos D. Afantenos. 2016. Discourse structure and dialogue acts in multiparty dialogue: the stac corpus. In Proceedings of the 11th Edition of Language Resources and Evaluation Conference (LREC). Merwan Barlier, Julien Perolat, Romain Laroche, and Olivier Pietquin. 2015. Human-machine dialogue as a stochastic game. In Proceedings of the 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue (Sigdial). Michael Bowling and Manuela Veloso. 2002. Multiagent learning using a variable learning rate. Artificial Intelligence 136(2):215–250. Inigo Casanueva, Thomas Hain, Heidi Christensen, Ricard Marxer, and Phil Green. 2015. Knowledge transfer between speakers for personalised dialogue management. In Proceedings of the 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue (Sigdial). Senthilkumar Chandramohan, Matthieu Geist, Fabrice Lefèvre, and Olivier Pietquin. 2012. Co-adaptation in Spoken Dialogue Systems. In Proceedings of the 4th International Workshop on Spoken Dialogue Systems (IWSDS). Paris, France, page 1. Barbara di Eugenio, Pamela W. Jordan, Richmond S. Thomason, and Johanna D. Moore. 2000. The agreement process: an empirical investigation of humanhuman computer-mediated collaborative dialogs. International Journal of Human-Computer Studies 53(6):1017 – 1076. Ioannis Efstathiou and Oliver Lemon. 2014. Learning non-cooperative dialogue behaviours. In Proceedings of the 15th Annual Meeting of the Special Interest Group on Discourse and Dialogue (Sigdial). Layla El Asri, Remi Lemonnier, Romain Laroche, Olivier Pietquin, and Hatim Khouzaimi. 2014. Nastia: Negotiating appointment setting interface. In Proceedings of the 9th Edition of Language Resources and Evaluation Conference (LREC). Michael S English and Peter A Heeman. 2005. Learning mixed initiative dialogue strategies by using reinforcement learning on both conversants. In Proceedings of the conference on Human Language Technology (HLT). Li Fei-Fei, Rob Fergus, and Pietro Perona. 2006. Oneshot learning of object categories. IEEE transactions on pattern analysis and machine intelligence 28(4):594–611. Milica Gašic, Catherine Breslin, Matthew Henderson, Dongho Kim, Martin Szummer, Blaise Thomson, Pirros Tsiakoulis, and Steve Young. 2013. Pomdpbased dialogue manager adaptation to extended domains. In Proceedings of the 14th Annual Meeting of the Special Interest Group on Discourse and Dialogue (Sigdial). Aude Genevay and Romain Laroche. 2016. Transfer learning for user adaptation in spoken dialogue systems. In Proceedings of the 15th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS). International Foundation for Autonomous Agents and Multiagent Systems. Kallirroi Georgila, Claire Nelson, and David Traum. 2014. Single-agent vs. multi-agent techniques for concurrent reinforcement learning of negotiation dialogue policies. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (ACL). Kallirroi Georgila and David R Traum. 2011. Reinforcement learning of argumentation dialogue policies in negotiation. In Proceedings of the 11th Annual Conference of the International Speech Communication Association (Interspeech). pages 2073– 2076. Hatim Khouzaimi, Romain Laroche, and Fabrice Lefevre. 2015. Optimising turn-taking strategies with reinforcement learning. In Proceedings of the 16th Annual Meeting of the Special Interest Group on Discourse and Dialogue (Sigdial). Romain Laroche and Raphaël Féraud. 2017. Algorithm selection of off-policy reinforcement learning algorithm. arXiv preprint arXiv:1701.08810 . Romain Laroche and Aude Genevay. 2017. The negotiation dialogue game. In Dialogues with Social Robots, Springer, pages 403–410. Romain Laroche, Ghislain Putois, Philippe Bretier, Martin Aranguren, Julia Velkovska, Helen Hastie, Simon Keizer, Kai Yu, Filip Jurcicek, Oliver Lemon, and Steve Young. 2011. D6.4: Final evaluation of classic towninfo and appointment scheduling systems. Report D6 4. Romain Laroche, Ghislain Putois, Philippe Bretier, and Bernadette Bouchon-Meunier. 2009. Hybridisation of expertise and reinforcement learning in dialogue systems. In Proceedings of the 9th Annual Conference of the International Speech Communication Association (Interspeech). pages 2479–2482. Alessandro Lazaric. 2012. Transfer in reinforcement learning: a framework and a survey. In Reinforcement Learning, Springer, pages 143–173. Oliver Lemon and Olivier Pietquin. 2012. Data-Driven Methods for Adaptive Spoken Dialogue Systems: Computational Learning for Conversational Interfaces . Springer. Esther Levin and Roberto Pieraccini. 1997. A stochastic model of computer-human interaction for learning dialogue strategies. In Proceedings of the 5th European Conference on Speech Communication and Technology (Eurospeech). Mike Lewis, Denis Yarats, Yann N. Dauphin, Devi Parikh, and Dhruv Batra. 2017. Deal or no deal? end-to-end learning for negotiation dialogues. arXiv preprint arXiv:1706.05125 . Diane Litman, Susannah Paletz, Zahra Rahimi, Stefani Allegretti, and Caitlin Rice. 2016. The teams corpus and entrainment in multi-party spoken dialogues . Ghislain Putois, Romain Laroche, and Philippe Bretier. 2010. Online reinforcement learning for spoken dialogue systems: The story of a commercial deployment success. In Proceedings of the 11th Annual Meeting of the Special Interest Group on Discourse and Dialogue. Citeseer, pages 185–192. Stefan Schaal et al. 1997. Learning from demonstration. Advances in neural information processing systems pages 1040–1046. Lloyd S Shapley. 1953. Stochastic games. Proceedings of the National Academy of Sciences of the United States of America 39(10):1095. Richard S Sutton and Andrew G Barto. 1998. Reinforcement learning: An introduction, volume 1. MIT press Cambridge. Matthew E Taylor and Peter Stone. 2009. Transfer learning for reinforcement learning domains: A survey. Journal of Machine Learning Research 10:1633–1685.
2cs.AI
arXiv:1503.00374v2 [cs.DS] 31 Aug 2016 A Randomized Algorithm for Approximating the Log Determinant of a Symmetric Positive Definite Matrix Christos Boutsidis ∗ Petros Drineas † Eugenia-Maria Kontopoulou § Prabhanjan Kambadur ‡ Anastasios Zouzias ¶ Abstract We introduce a novel algorithm for approximating the logarithm of the determinant of a symmetric positive definite (SPD) matrix. The algorithm is randomized and approximates the traces of a small number of matrix powers of a specially constructed matrix, using the method of Avron and Toledo [AT11]. From a theoretical perspective, we present additive and relative error bounds for our algorithm. Our additive error bound works for any SPD matrix, whereas our relative error bound works for SPD matrices whose eigenvalues lie in the interval (θ1 , 1), with 0 < θ1 < 1; the latter setting was proposed in [HMS15]. From an empirical perspective, we demonstrate that a C++ implementation of our algorithm can approximate the logarithm of the determinant of large matrices very accurately in a matter of seconds. 1 Introduction Given a matrix A ∈ Rn×n , the determinant of A, denoted by det(A), is one of the most important quantities associated with A. Since its invention by Cardano and Leibniz in the late 16th century, the determinant has been a fundamental mathematical concept with countless applications in numerical linear algebra and scientific computing. The advent of Big Data, which are often represented by matrices, increased the applicability of algorithms that compute, exactly or approximately, matrix determinants; see, for example, [LZL05, ZLLW08, ZL07, dBEG08, HSD+ 13] for machine learning applications (e.g., gaussian process regression) and [LP01, KL13, FHT08, PB97, PBGS00] for several data mining applications (e.g., spatial-temporal time series analysis). Formal definitions of the determinant include the well-known formulas derived by Leibniz and Laplace; however, neither the Laplace nor the Leibniz formula can be used to design an efficient, polynomial-time, algorithm to compute the determinant of A. To achieve this goal, one should rely on other properties of the determinant. For example, a standard approach would be to leverage the so-called LU matrix decomposition or the Cholesky decomposition for symmetric ∗ Email: [email protected] Purdue University. West Lafayette, IN. Email: [email protected] ‡ Bloomberg L.P. New York, NY. Email: [email protected] § Purdue University. West Lafayette, IN. Email: [email protected] ¶ Swisscom. Zurich, Switzerland. Email: [email protected] † 1 positive definite matrices (SPD) to get an O(n3 ) deterministic algorithm to compute the determinant of A. (Recall that an SPD matrix is a symmetric matrix with strictly positive eigenvalues.) In this paper, we are interested in approximating the logarithm of the determinant of a symmetric positive definite (SPD) matrix A. The logarithm of the determinant, instead of the determinant itself, is important in several settings [LZL05, ZLLW08, ZL07, dBEG08, HSD+ 13, LP01, KL13, FHT08, PB97, PBGS00]. Definition 1. [L OG D ET P ROBLEM DEFINITION ] Given an SPD matrix A ∈ Rn×n , compute, exactly or approximately, logdet (A). Note that since all the eigenvalues of A are strictly positive, the determinant of A is strictly positive. The best exact algorithm for the above problem simply computes the determinant of A in cubic time and takes its logarithm. Few approximation algorithms have appeared in the literature, but they either lack a proper theoretical convergence analysis or do not work for all SPD matrices. We will discuss prior work in detail in Section 1.2. 1.1 Our contributions We present a fast approximation algorithm for the problem of Definition 1. Our main algorithm (Algorithm 3) is randomized and its running time is   O nnz(A) mε−2 + log n log(1/δ) , where nnz(A) denotes the number of non-zero elements in A, 0 < δ < 1 denotes the failure probability of our algorithm, and (integer) m > 0 and (real) ε > 0 are user-controlled accuracy parameters that are specified in the input of the algorithm. The first step of our approximation algorithm uses the power method to compute an approximation to the dominant eigenvalue of A. This value will be used in a normalization (preconditioning) step in order to compute a convergent matrix-Taylor expansion. The second step of our algorithm leverages a truncated matrix-Taylor expansion of a suitably constructed matrix in order to compute an approximation of the log determinant. This second step leverages a randomized trace estimation algorithm from [AT11]. \ (A) be the value returned by our approximation algorithm (Algorithm 3); let logdet (A) Let logdet be the true log determinant of A; let λi (A) denote the i-th eigenvalue of A for all i = 1, . . . , n with λ1 (A) ≥ λ2 (A) ≥ . . . ≥ λn (A) > 0; and let κ (A) = λ1 (A)/λn (A) be the condition number of A. Our main result, proven in Lemma 6, is that if    1 m ≥ 7κ (A) log , (1) ε then, with probability at least 1 − 2δ, \ (A) − logdet (A) ≤ 2εΓ, logdet where Γ= n X i=1   λ1 (A) log 7 · . λi (A) 2 (2) We now take a careful look at the above approximation bound. First, given our choice of m in eqn. (1), the running time of the algorithm becomes   O nnz(A) κ (A) log (1/ε) ε−2 + log n log(1/δ) . (3) Thus, the running time of our algorithm increases linearly with the condition number of A. The error of our algorithm scales with Γ, a quantity that is not immediately comparable to logdet (A). It is worth noting that the Γ term increases logarithmically with respect to the ratios λ1 (A)/λi (A) ≥ 1. An obvious, but potentially loose upper bound for the sum of those ratios, is Γ= n X i=1   λ1 (A) log 7 · ≤ n · log (7κ(A)) . λi (A) (4) Our second result handles the family of SPD matrices whose eigenvalues all lie in the interval (θ1 , 1), with 0 < θ1 < 1; this setting was proposed in [HMS15]. In this case, a simplified version of Algorithm 3 returns a relative error approximation to the log-determinant of the input matrix. Indeed, Lemma 8 proves that, with probability at least 1 − δ, \ (A) − logdet (A) ≤ 2ε|logdet (A) |. logdet The running time of the simplified algorithm is   log(1/ε) log(1/δ) O nnz(A) . ε2 θ 1 (5) Finally, we implemented our algorithm in C++ and tested it on several large dense and sparse matrices. Our dense implementation runs on top of Elemental [PMVdG+ 13], a linear algebra library for distributed matrix computations with dense matrices. Our sparse implementation runs on top of Eigen 1 , a software library for sparse matrix computations. Our code is available to download on Github (see Section 5 for details and a link to our code). 1.2 Related Work The most relevant result to ours is the work in [BP99]. Barry and Pace [BP99] described a randomized algorithm for approximating the logarithm of the determinant of a matrix with special structure that we will describe below. They show that in order to approximate the logarithm of the determinant of a matrix A, it suffices to approximate the traces of Dk , for k = 1, 2, 3... for a suitably constructed matrix D. Specifically, [BP99] deals with approximations to SPD matrices A of the form A = In − αD, where 0 < α < 1 and all eigenvalues of D are in the interval [−1, 1]. \ (A) that is close to Given such a matrix A, the authors of [BP99] seek to derive an estimator logdet logdet (A). [BP99] proved (using the so-called Martin expansion [Mar92]) that log(det(A)) = − k=1 1 ∞   X αk  k  tr Dk − tr D . k k m X αk k=m http://eigen.tuxfamily.org/ 3 They considered the following estimator:   p  m  k T k  X 1 X α zi D zi   . \ logdet (A) = −n   Tz p k z  i i i=1  k=1 | {z } Vi All Vi for i = 1 . . . p are random variables and the value of p controls the variance of the estimator. The algorithm in [BP99] constructs vectors zi ∈ Rn whose entries are independent identically distributed standard Gaussian random variables. The above estimator ignores the trailing terms of the Martin expansion and only tries to approximate the first m terms. [BP99] presented the following approximation bound: s m−1 n·α σ2 \ (A) − logdet (A) ≤ logdet + 1.96 · , (m + 1)(1 − α) p where σ 2 is the variance of the random variable Vi . The above bound fails with probability at most 0.05. We now compare the results in [BP99] with ours. First, the idea of using the Martin expansion [Mar92] to relate the logarithm of the determinant and traces of matrix powers is present in both approaches. Second, the algorithm of [BP99] is applicable to SPD matrices that have special structure, while our algorithm is applicable to any SPD matrix. Intuitively, we overcome this limitation of [BP99] by estimating the top eigenvalue of the matrix in the first step of our algorithm. Third, our error bound is much better that the error bound of [BP99]. To analyze our algorithm, we used the theory of randomized trace estimators of Avron and Toledo [AT11], which relies on stronger measure-concentration inequalities than [BP99], which uses the weaker Chebyshev’s inequality. A similar idea using Chebyshev polynomials appeared in the paper [PL04]; to the best of our understanding, there are no theoretical convergence properties of the proposed algorithm. Applications to Gaussian process regression appeared in [LZL05, ZLLW08, ZL07]. The work of [Reu02] uses an approximate matrix inverse to compute the n-th root of the determinant of A for large sparse SPD matrices. The error bounds in this work are a posteriori and thus not directly comparable to our bounds. [HAB14] provides a strong worst-case theoretical result which is, however, only applicable to Symmetric Diagonally Dominant (SDD) matrices. The algorithm is randomized and guarantees \ (A) − logdet (A) ≤ ε · n, for a user specified error parameter that, with high probability, logdet ε > 0. This approach also uses the Martin expansion [Mar92] as well as ideas from preconditioning systems of linear equations with Laplacian  matrices [ST04]. The algorithm of [HAB14] runs in 3 2 −2 time O nnz(A)ε log (n) log (nκ(A)/ε) . To compare to our approach, we need to combine the suboptimal upper bound for Γ from eqn. (4) with the bound of eqn. (2). Then, we can run Algorithm 3 with input ε ε0 = , log(7κ(A)) 4 instead of ε to guarantee that the final error of our approximation will be bounded by ε0 n. Then, we can observe that the running time of [HAB14] depends logarithmically on the condition number of the input matrix A, whereas our algorithm has a linear dependency on the condition number. Notice, however, that our method is applicable to any SPD matrix while the method in [HAB14] is applicable only to SDD matrices; given current state-of-the-art on Laplacian preconditioners it looks hard to extend the approach of [HAB14] to general SPD matrices. Independently and in parallel with our work, [HMS15] presented an algorithm using Stochastic Chebyshev Expansions for the log-determinant problem. The algorithm is very similar in spirit to our approach, using the Chebyshev instead of the Taylor expansion and achieves relative-error guarantees for a special class of SPD matrices, namely matrices whose eigenvalues all lie in the interval (θ1 , 1 − θ1 ) for some 0 < θ1 < 1/2. As we already discussed, our algorithm also achieves a relative error bound under q such an assumption; the only difference is that the running time of [HMS15] is proportional to 1 θ1 log θ11 , whereas the running time of our approach (see eqn. (5)) is proportional to θ11 . This slightly improved running time might be due to the use of the Stochastic Chebyshev Expansions. However, importantly, our algorithm works for any SPD matrix, with arbitrary eigenvalues. Not surprisingly, the added generality comes with a loss in accuracy and the relative error bound becomes an additive error bound. Finally, two very recent papers [HMAS16, SAI16]2 presented algorithms to approximate the logdet of a matrix, highlighting the renewed importance of the topic. The work of [SAI16] presents a very novel approach to approximate the logdet of a positive semi-definite matrix, using a randomized subspace iteration approach. To the best of our understanding, the relevant bounds in their work (Theorem 2 in [SAI16]) are not directly comparable to our bounds. The work of [HMAS16] follows the lines of [HMS15] and leverages the use of Chebyshev approximations to propose novel estimators for the trace of a matrix function. Among the many exciting applications of the proposed approach is an additive-error approach to approximate the logdet of any square non-singular matrix; the algorithm needs as inputs upper and lower bounds for all the singular values of the input matrix. Similar to the running time of our additive error algorithm in eqn. (3), the time complexity of the proposed algorithm depends on the condition number of the input matrix (see Corollary 7 of [HMAS16]). We conclude by noting that common algorithms for the determinant computation assume floating point arithmetic and do not measure bit operations. If the computational cost is to be measured in bit operations, the situation is much more complicated and an exact computation of the determinant, even for integer matrices, is not trivial. We refer the interested reader to [EGV00] for more details. 2 Preliminaries 2.1 Notation Let A, B, . . . denote matrices and let a, b, . . . denote column vectors. In is the n×n identity matrix; 0m×n is the m × n matrix of zeros; tr (A) is the trace of a square matrix A; the Frobenius and 2 Both papers appeared after an earlier version of this paper was posted on ArXiv on March 2015 and cite this earlier version of our work. 5 P the spectral matrix-norms are: kAk2F = i,j A2ij and kAk2 = maxkxk2 =1 kAxk2 . We denote the determinant of a matrix A by det(A) and the (natural) logarithm of the determinant of A by logdet (A). We use log x to denote the natural logarithm of x . Finally, given an event E, P [E] denotes the probability of the event. For an SPD matrix A ∈ Rn×n , log [A] is an n × n matrix defined as: log [A] = UDUT , where U ∈ Rn×n contains the eigenvectors of A and D ∈ Rn×n is diagonal with entries being log(λ1 (A)), log(λ2 (A)), . . . , log(λn (A)). Let x be a scalar variable that satisfies |x| < 1. Then, using the Taylor expansion, log(1 − x) = − ∞ X xk k=1 k . A matrix-valued generalization of this identity is the following statement. Lemma 1. Let A ∈ Rn×n be a symmetric matrix whose eigenvalues all lie in the interval (−1, 1). Then, log(In − A) = − ∞ X Ak k=1 2.2 k . Power method The first step in our algorithm for approximating the determinant of an SPD matrix is to obtain an estimate for the largest eigenvalue of the matrix. Given an SPD matrix A ∈ Rn×n we will use the power-method (Algorithm 1) to obtain an accurate estimate of its largest eigenvalue. This estimated eigenvalue is denoted by λ̃1 (A). Algorithm 1 Power method, repeated q times. • Input: SPD matrix A ∈ Rn×n , integers q, t > 0 • For j = 1, . . . , q 1. Pick uniformly at random a vector xj0 ∈ {+1, −1}n 2. For i = 1, . . . , t • xji = A · xji−1 3. Compute: λ̃j1 (A) = T xjt Axjt T xjt xjt • Return: λ̃1 (A) = maxj=1...q λj1 (and the corresponding vector xt = xjt ) Algorithm 1 requires O(qt(n + nnz(A))) arithmetic operations to compute λ̃1 (A). Lemma 2 (see [Tre11] for a proof) argues that any λ̃j1 (A) is close to λ1 (A). 6 Lemma 2. For any fixed j = 1 . . . q, and for any t > 0, ε > 0, with probability at least 3/16, T (1 − ε) xjt Axjt λ (A) ≤ = λ̃j1 (A). 1 jT j 1 + 4n(1 − ε)2t xt xt  √  Let e = 2.718 . . . and let ε = 1 − (1/e) and t = log 4n ; then, with probability at least 3/16, for any fixed j = 1 . . . q, 1 1 λ1 (A) ≤ λ1 (A) ≤ λ̃j1 (A). 6 2e It is now easy to see that the largest value λ̃1 (A) (and the corresponding vector xt ) fails to satisfy the inequality (1/6)λ1 (A) ≤ λ̃1 (A) with probability at most    q 3 q 13 1− = ≤ δ, 16 16 where the last inequality follows by setting q = d4.82 log(1/δ)e ≥ log(1/δ)/ log(16/13). Finally, we note that, from the min-max principle, λ̃1 (A) ≤ λ1 (A). We summarize the above discussion in the following lemma.  √  Lemma 3. Let λ̃1 (A) be the output of Algorithm 1 with q = d4.82 log(1/δ)e and t = log 4n . Then, with probability at least 1 − δ, 1 λ1 (A) ≤ λ̃1 (A) ≤ λ1 (A). 6  The running time of Algorithm 1 is O (n + nnz(A)) log(n) log 1δ . 2.3 Trace estimation Even though computing the trace of a square n × n matrix requires only O(n) arithmetic operations, the situation is more complicated when A is given through a matrix function, e.g., A = X2 , for some matrix X and the user only observes X. For situations such as these, Avron and Toledo [AT11] analyzed several algorithms to estimate the trace of A. Algorithm 2 and Lemma 4 present the relevant results from their paper. Lemma 4. Let A ∈ Rn×n be an SPD matrix, let 0 < ε < 1 be an accuracy parameter, and let 0 < δ < 1 n be a failure  probability.2 If g1 , g2 , . . . , gp ∈ R are independent random standard Gaussian vectors, then, for p = 20 log(2/δ)/ε , with probability at least 1 − δ, p 1X > tr (A) − gi Agi ≤ ε · tr (A) . p i=1 The above lemma is immediate from Theorem 5.2 in [AT11]. 7 Algorithm 2 Randomized Trace Estimation • Input: SPD matrix A ∈ Rn×n , accuracy 0 < ε < 1, and failure probability 0 < δ < 1.   1. Let p = 20 log(2/δ)/ε2 2. Let g1 , g2 , . . . , gp be a set of independent Gaussian vectors in Rn 3. Let γ = 0 4. For i = 1, . . . , p • γ = γ + gi> Agi 5. γ = γ/p • Return: γ 3 Additive error approximation for general SPD matrices Lemma 5 is the starting point of our main algorithm for approximating the determinant of a symmetric positive definite matrix. The message in the lemma is that computing the log determinant of an SPD matrix A reduces to the task of computing the largest eigenvalue of A and the trace of all the powers of a matrix C related to A. Lemma 5. Let A ∈ Rn×n be an SPD matrix. For any α with λ1 (A) < α, define B := A/α and C := In − B. Then,  ∞ X tr Ck logdet (A) = n log(α) − . k k=1 Proof. Observe that B is an SPD matrix with kBk2 < 1. It follows that logdet (A) = log(αn det(A/α)) n Y = n log(α) + log ! λi (B) i=1 = n log(α) + n X log(λi (B)) i=1 = n log(α) + tr (log [B]) . Here, we used standard properties of the determinant, standard properties of the logarithm function, and the fact that (recall that B is an SPD matrix), tr (log [B]) = n X λi (log [B]) = i=1 n X log(λi (B)). i=1 Now, tr (log [B]) = tr (log [In − (In − B)]) = tr − ∞ X (In − B)k k=1 8 k !  ∞ X tr Ck =− . k k=1 (6) The second equality follows by the Taylor expansion because all the eigenvalues of C = In − B are contained3 in (0, 1) and the last equality follows by the linearity of the trace operator. 3.1 Algorithm Lemma 5 indicates the following high-level procedure for computing the logdet of an SPD matrix A: 1. Compute some α with λ1 (A) < α. 2. Compute C = In − A/α. 3. Compute the trace of all the powers of C. To implement the first step in this procedure we use the power iteration from the numerical linear algebra literature (see Section 2.2). The second step is straightforward. To implement the third P k step, we keep a finite number of summands in the expansion ∞ tr C . This step is important k=1 since the quality of the approximation, both theoretically and empirically, depends on the number of summands (denoted with m) that will be kept. On the other hand, the running time of the algorithm increases with m. Finally, to estimate the traces of the powers of C, we use the randomized algorithm of Section 2.3. Our approach is described in detail in Algorithm 3; notice that step 7 in Algorithm 3 is an efficient way of computing ! p m X X 1 \ (A) := n log(α) − logdet gi> Ck gi /k. p k=1 3.2 i=1 Error bound The following lemma proves that Algorithm 3 returns an accurate approximation to the logdet of A. \ (A) be the output of Algorithm 3 on inputs A, m, and ε. Then, with probability at Lemma 6. Let logdet least 1 − 2δ,   m  1 \ logdet (A) − logdet (A) ≤  + 1 − · Γ, 7κ (A)     P (A) where Γ = ni=1 log 7 · λλ1i (A) . If m ≥ 7κ (A) log 1ε , then \ (A) − logdet (A) ≤ 2Γ. logdet Proof. First, note that using our choice for α in Step 3 of Algorithm 3 and applying Lemma 3, we get that, with probability at least 1 − δ, 7 λ1 (A) < λ1 (A) ≤ α ≤ 7λ1 (A), 6 3 Indeed, λi (C) = 1 − λi (B) and 0 < λi (B) < 1 for all i = 1 . . . n. 9 (7) Algorithm 3 Randomized Log Determinant Estimation 1: INPUT: A ∈ Rn×n , accuracy parameter ε > 0, and integer m > 0. 2: Compute λ̃1 (A) using Algorithm 1 with (integers)t = O (log n) and q = O (log(1/δ)) 3: Pick α = 7λ̃1 (A) 4: Set C = In − A/α   5: Set p = 20 log(2/δ)/ε2 6: Let g1 , g2 , . . . , gp ∈ Rn be i.i.d. random Gaussian vectors. 7: For i = 1, 2 . . . , p (i) (i) (i) • v1 = Cgi and γ1 = gi> v1 • For k = 2, . . . , m (i) (i) (i) (i) 1. vk := Cvk−1 . (i) 2. γk = gi> vk (Inductively γk = gi> Ck gi ) • EndFor EndFor  P  P (i) p 1 \ (A) = n log(α) − m 9: OUTPUT: logdet /k k=1 p i=1 γk 8: The strick inequality at the leftmost side of the above equation follows since all eigenvalues of A are strictly positive. Let’s call the event that the above inequality holds E1 ; obviously, P [E1 ] ≥ 1 − δ   (and thus P Ē1 ≤ δ). We condition all further derivations on E1 holding and we manipulate \ (A) − logdet (A) as follows: ∆ = logdet ∆= ≤ = m X k=1 m X k=1 p X 1 p | i=1 p 1X > k gi C gi p ! 1 p ! i=1 p X gi> Ck gi /k − /k − i=1 gi> ∞ X k=1 m X   tr Ck /k  k tr C  /k + k=1 m X ! Ck /k gi − tr k=1 m X ! Ck /k + k=1 {z   tr Ck /k k=m+1 ∞ X   tr Ck /k . k=m+1 } ∆1 ∞ X | {z ∆2 } Below, we bound the two Westart with ∆1 : the idea is to apply  P termsk ∆1 and ∆2 separately. 2 . Let E denote the probability that Lemma 4 on the matrix m C /k with p = 20 log(2/δ)/ε 2 k=1   Lemma 4 holds; obviously, P [E2 ] ≥ 1 − δ (and thus P Ē2 ≤ δ) given our choice of p. We condition all further derivations on E2 holding as well to get ! ! m ∞ X X Ck /k ≤ ε · tr Ck /k . ∆1 ≤ ε · tr k=1 k=1 10  In the last inequality we used the fact that C is a positive matrix, hence for all k, tr Ck > 0. The second term ∆2 is bounded as follows: ∞ X ∆2 = =  k tr C k=m+1 ∞ X  ∞ X /k ≤ k=m+1  ∞ X  tr Cm · Ck−m /k ≤ k=m+1 ∞ X  1− λn (A) α  tr Ck−m /k ≤ kCm k2 · k=m+1 m    kCm k2 · tr Ck−m /k k=m+1 = kCm k2 · ≤   tr Ck /k ∞ X   tr Ck /k k=1 · ∞ X   tr Ck /k. k=1 In the first inequality, we used the triangle inequality and the fact that C is a positive matrix. In the second inequality, we used the following fact4 : given two positive semidefinite matrices A, B of the same size, tr (AB) ≤ kAk2 · tr (B) . In the last inequality, we used the fact that λ1 (C) = 1 − λn (B) = 1 − λn (A)/α. Combining the bounds for ∆1 and ∆2 gives  m  X   ∞ tr Ck λ (A) n \ (A) − logdet (A) ≤  + 1 − · . logdet α k k=1 We have already proven in Lemma 5 that  ∞ X tr Ck = −tr (log [B]) = n log(a) − logdet (A) . k k=1 Notice that the assumption of Lemma 5 (namely, λ1 (A) < α) is satisfied from the inequality of eqn. (7). We further manipulate the last term as follows: n Y n log(a) − logdet (A) = n log(α) − log( λi (A)) i=1 = n log(α) − n X log(λi (A)) i=1 = = n X i=1 n X i=1 4 This follows from Von Neumann’s trace inequality. 11 (log (α) − log (λi (A)))  log α λi (A)  . Collecting our results together, we get: \ (A) − logdet (A) ≤ logdet     X   n λn (A) m α + 1− · . log α λi (A) i=1 Using the inequality of eqn. (7) (only the upper bound on α is needed here) proves the first inequalx ity of the lemma. To prove the second inequality, we use the well-known fact that 1 − x−1 ≤ e−1 (where e = 2.718 . . . and x > 0) and our choice for m. Finally, recall that we conditioned all derivations on events E1 and E2 both holding, which can be bounded as follows:       P [E1 ∩ E2 ] = 1 − P Ē1 ∪ Ē2 ≥ 1 − P Ē1 − P Ē1 ≥ 1 − 2δ. The first inequality in the above derivation follows from the union bound. 3.3 Running time Step 2 takes O(nnz(A) log(n) log(1/δ)) time; we assume that nnz(A) ≥ n, since otherwise the determinant of A would be trivially equal to zero. For each k > 0, vk = Ck gi . The algorithm inductively computes vk and gi> Ck gi = gi> vk for all k = 1, 2, . . . , m. Given vk−1 , vk and gi> Ck gi can be computed in nnz(C) and O(n) time, respectively. Notice that nnz(C) ≤ n + nnz(A). Therefore, step 7 requires O(p · m · nnz(A)) time. Since p = O(ε−2 log(1/δ)), the total cost is    m  1 O nnz(A) · 2 + log n · log . ε δ 4 Relative error approximation for SPD matrices with bounded eigenvalues In this section, we argue that a simplified version of Algorithm 3 achieves a relative error approximation to the logdet of the SPD matrix A, under the assumption that all the eigenvalues of A lie in the interval (θ1 , 1), where 0 < θ1 < 1. This is a mild generalization of the setting introduced in [HMS15]. Given the upper bound on the largest eigenvalue of A, the proof of the following lemma (which is the analog of Lemma 5) is straightforward. Lemma 7. Let A ∈ Rn×n be an SPD matrix whose eigenvalues lie in the interval (θ1 , 1), for some 0 < θ1 < 1. Let C := In − A; then,  ∞ X tr Ck logdet (A) = − . k k=1 Proof. Similarly to the proof of Lemma 5, logdet (A) = log n Y ! λi (A) = i=1 n X i=1 12 log(λi (A)) = tr (log [A]) . Now, tr (log [A]) = tr (log [In − (In − A)]) = tr − ∞ X (In − A)k ! k k=1  ∞ X tr Ck =− . k k=1 The second equality follows by the Taylor expansion since all the eigenvalues of C = In − A are contained in the interval (0, 1). 4.1 The algorithm and the relative error bound We simplify Algorithm 3 as follows: we skip steps 2 and 3 and in step 4 we set C = In − A. The following lemma proves that in this special case the modified algorithm returns a relative error approximation to the log determinant of the input matrix A. \ (A) be the output of the (modified) Algorithm 3 on inputs A and ε. Then, with Lemma 8. Let logdet probability at least 1 − δ, \ (A) − logdet (A) ≤ 2ε · |logdet (A)| . logdet \ (A) − logdet (A) as follows: Proof. Similarly to the proof of Lemma 6, we manipulate ∆ = logdet ∆= m X k=1 ≤ = m X k=1 p X 1 p | i=1 p 1X > k gi C gi p ! /k − i=1 p 1X p /k − i=1 gi>   tr Ck /k k=1 ! gi> Ck gi ∞ X m X  k tr C  /k + k=1 m X ! Ck /k gi − tr m X ! Ck /k + k=1 k=1 {z   tr Ck /k k=m+1 ∞ X   tr Ck /k . k=m+1 } ∆1 ∞ X | {z ∆2 } We now bound the twoP terms ∆1 and ∆2 separately. We start   with ∆1 : the idea is to apply m k 2 Lemma 4 on the matrix k=1 C /k with p = 20 log(2/δ)/ε . Hence, with probability at least 1 − δ (this is the only probabilistic event in this lemma and hence 1 − δ is a lower bound on the success probability of the lemma): ! ! m ∞ X X ∆1 ≤ ε · tr Ck /k ≤ ε · tr Ck /k . k=1 k=1  In the last inequality we used the fact that C is a positive definite matrix, hence for all k, tr Ck > 13 0. Bounding ∆2 follows the lines of the proof of Lemma 6: ∆2 = ≤ ∞ X k=m+1 ∞ X ∞ X   tr Ck /k =   tr Cm · Ck−m /k k=m+1  m k−m kC k2 · tr C  m /k = kC k2 · k=m+1 ≤ kCm k2 · ∞ X   tr Ck /k ≤ (1 − λn (A))m k=1 ∞ X   tr Ck−m /k k=m+1 ∞  X  tr Ck /k . k=1 In the last inequality, we used the fact that λ1 (C) = 1 − λn (A). Combining the bounds for ∆1 and ∆2 gives  ∞ X tr Ck m \ (A) − logdet (A) ≤ ( + (1 − λn (A)) ) · . logdet k k=1 We have already proven in Lemma 7 that  ∞ X tr Ck = −tr (log [A]) = −logdet (A) . k k=1 Collecting our results, we get: \ (A) − logdet (A) ≤ (ε + (1 − λn (A))m ) · |logdet (A)| . logdet Using 1 − λn (A) < 1 − θ1 , we conclude that \ (A) − logdet (A) ≤ (ε + (1 − θ1 )m ) · |logdet (A)| . logdet Setting  1 m= · log θ1   1 ε x and using 1 − x−1 ≤ e−1 (where e = 2.718 . . . and x > 0), guarantees that (1 − θ1 )m ≤ ε and concludes the proof of the lemma. We conclude by discussing the running  time of thesimplified  Algorithm 3, which is equal to log(1/δ) log(1/ε) and m = O O(p · m · nnz(A)). Since p = O , the running time becomes θ1 ε2  O  log(1/ε) log(1/δ) nnz(A) . ε2 θ 1 14 5 Experiments The goal of our experimental section is to establish that our approximation to logdet (A) (as computed by Algorithm 3) is both accurate and fast for both dense and sparse matrices. The accuracy of Algorithm 3 is measured by comparing its result against the exact logdet (A) computed via the Cholesky factorization. The rest of this section is organized as follows: in Section 5.1, we describe our software for approximating logdet (A); in Section 5.2 we describe the computational environment that we used; and in Sections 5.3 and 5.4 we discuss experimental results for dense and sparse SPD matrices, respectively. 5.1 Software We developed high-quality, shared- and distributed-memory parallel C++ code for the algorithms listed in this paper. All of the code that was developed for this paper is hosted at https: //github.com/pkambadu/ApproxLogDet. In it’s current state, our software supports: (1) ingesting dense (binary and text format) and sparse (binary, text, and matrix market format) matrices, (2) generating large random SPD matrices, (3) computing both approximate and exact spectral norms of matrices, (4) computing both approximate and exact traces of matrices, and (5) computing both approximate and exact log determinants of matrices. Currently, we support both Eigen [GJ+ 10] and Elemental [PMVdG+ 13] matrices. The Eigen software package supports both dense and sparse matrices, while the Elemental software package mostly supports dense matrices and only recently added support for sparse matrices (pre-release). As we wanted the random SPD generation to be fast, we have used parallel random number generators from Random123 [SMDS11] in conjunction with Boost.Random. 5.2 Environment All our experiments were run on “Nadal”, a 60-core machine, where each core is an Intel R Xeon R E7-4890 machine running at 2.8 Ghz. Nadal has 1 TB of RAM and runs Linux kernel version 2.632. For compilation, we used GCC 4.9.2. We used Eigen 3.2.4, OpenMPI 1.8.4, Boost 1.55.7, and the latest version of Elemental at https://github.com/elemental. For experiments with Elemental, we used OpenBlas, which is an extension of GotoBlas [GVDG08], for its parallel prowess; Eigen has built-in the BLAS and LAPACK packages. 5.3 Dense Matrices Data Generation. In our experiments, we used two types of synthetic SPD matrices. The first type were diagonally dominant SPD matrices and were generated as follows. First, we created X ∈ Rn×n by drawing n2 entries from a uniform sphere with center 0.5 and radius 0.25. Then, we generated a symmetric matrix Y by setting Y = 0.5 ∗ (X + X> ). Finally, we ensured that the desired matrix A is positive definite by adding the value n to each diagonal entry [Cur09] of Y: A = Y + nIn . We call this method randSPDDenseDD. 15 The second approach generates SPD matrices that are not diagonally dominant. We created X, D ∈ Rn×n by drawing n2 and n entries, respectively, from a uniform sphere with center 0.5 and radius 0.25; D is a diagonal matrix with small entries. Next, we generated an orthogonal random matrix Q = qr (X). Thus, Q is an orthonormal basis for X. Finally, we generated A = QDQT . We call this method randSPDDense. randSPDDense is more expensive than randSPDDenseDD, as it requires an additional O(n3 ) computations for the QR factorization and the matrix-matrix product. Evaluation. To evaluate the runtime of Algorithm 3 against a baseline, we used the Cholesky decomposition to compute the logdet (A). More specifically, we computed A = LLT and returned logdet (A) = 2 · logdet (L). Since Elemental provides distributed and shared memory parallelism, we restricted ourselves to experiments with Elemental matrices throughout this section. Note that we measured the accuracy of the approximate algorithm in terms of the relative error to ensure that we have numbers of the same scale for matrices with vastly different values for logdet (A); we defined the relative error e as e = 100(x − x̃)/x, where x is the true value and x̃ is the approximation. Similarly, we defined the speedup s as s = tx /tx̃ , where tx is the time needed to compute x and tx̃ is the time needed to compute the approximation x̃. Results. For dense matrices, we first used synthetic matrices generated using randSPDDense; these are relatively ill-conditioned matrices. We experimented with values of n (number of rows and columns of A) in the set {5, 000, 7, 500, 10, 000, 12, 500, 15, 000}. The three key points pertaining to these matrices are shown in Figure 1. First, we discuss the effect of m, the number of terms in the Taylor series used to approximate logdet (A); Figure 1(a) depicts our results for the sequential case. On the y-axis, we see the relative error, which is measured against the exact logdet (A) as computed via the Cholesky factorization. We observe that for these ill-conditioned matrices, for small values of m (less than four) the relative error is high. However, for all values of m ≥ 4, we observe that the error drops significantly and stabilizes. We note that in each iteration, all random processes were re-seeded with new values; we have plotted the error bars throughout Figure 1. The standard deviation for both accuracy and time was consistently small; indeed, it is not visible to the naked eye at scale. To see the benefit of approximation, we look at Figure 1(b) together with Figure 1(a). For example, at m = 4, for all matrices, we get at least a factor of two speedup. As n gets larger, the speedups of the approximation also increase. For example, for n = 15, 000, the speedup at m = 4 is nearly six-fold. In terms of accuracy, Figure 1(a) shows that at m = 4, the relative error is approximately 4%. This speedup is expected as the Cholesky factorization requires O(n3 ) operations; Algorithm 3 only relies on matrix-matrix products where one of the matrices has a small number of columns (equal to p), which is independent of n. Finally, we discuss the parallel speedup in Figure 1(c), which shows the relative speedup of the approximate algorithm with respect to the baseline Cholesky algorithm. For this evaluation, we set m = 4 and varied the number of processes, denoted by np, from 1 to 60. The main take away from Figure 1(c) is that the approximate algorithm provides nearly the same or increasingly better speedups relative to a parallelized version of the exact (Cholesky) algorithm. For example, for n = 15, 000, the speedups for using the approximate algorithm are consistently better that 6.5x. The absolute values for logdet (A) and timing along with the baseline numbers for this experiment are given in Table 1. We report the numbers in Table 1 at m = 4 at which point, we have low relative error. For the second set of dense experiments, we generated diagonally dominant matrices using 16 (a) Accuracy vs. m (b) Speedup vs. m (c) Parallel Speedup Figure 1: Panels 1(a) and 1(b) depict the effect of m (see Algorithm 3) on the accuracy of the approximation and the time to completion, respectively, √ for dense matrices generated by randSPDDense. For all the panels, p = 60 and t = 2 log 4n. The baseline for all experiments was the Cholesky factorization, which was used to compute the exact value of logdet (A). For panels 1(a) and 1(b), the number of cores, np, was set to one. The last panel 1(c) depicts the relative speedup of the approximate algorithm when compared to the baseline solver (at m = 4). Elemental was used as the backend for these experiments. For the approximate algorithm, we report the mean and standard deviation of ten iterations. n 5000 7500 10000 12500 15000 exact -3717.89 -5474.49 -7347.33 -9167.47 -11100.9 logdet (A) mean -3546.920 -5225.152 -7003.086 -8734.956 -10575.16 std 8.10 8.73 7.79 17.43 15.09 exact 2.56 7.98 18.07 34.39 58.28 time (secs) mean std 1.15 0.0005 2.53 0.0015 4.47 0.0006 7.00 0.0030 10.39 0.0102 √ Table 1: Accuracy and sequential running times (at p = 60, m = 4 and t = log 4n) for dense random matrices generated using randSPDDense. Baselines were computed using the Cholesky factorization; mean and standard deviation are reported over ten iterations. randSPDDenseDD; we were able to quickly generate and run benchmarks on matrices of sizes n × n with n in the set {10, 000, 20, 000, 30, 000, 40, 000} due to the relatively simpler procedure involved in matrix generation. In this set of of experiments, due to the diagonal dominance, all matrices were well-conditioned. The results of our experiments on these well-conditioned matrices are presented in Figure 2 and show a marked improved over the results presented in Figure 1. First, notice that very few terms of the Taylor series (i.e., small m) are sufficient to get high accuracy approximations; this is apparent in Figure 2(a). In fact, we see that even at m = 2, we are near convergence and at m = 3, for most of the matrices, we have near-zero relative error. This experimental result, combined with Figure 2(b) is particularly encouraging; at m = 2, we seem to not only have a nearly lossless approximation of logdet (A), but also have at least a fivefold speedup. Similarly to Figure 1, the speedups are better for larger matrices. For example, for n = 40, 000, the speedup at m = 2 is nearly twenty-fold. We conclude our analysis by presenting Figure 2(c), which similarly to Figure 1(c), points out that at any level of parallelism, Algorithm 3 17 (a) Accuracy vs. m (b) Speedup vs. m (c) Parallel Speedup Figure 2: Panels 2(a) and 2(b) depict the effect of m (see Algorithm 3) on the accuracy of the approximation and the time to completion, respectively, for diagonally dominant dense √ random matrices generated by randSPDDenseDD. For all the panels, p = 60 and t = 2 log 4n. The baseline for all experiments was the Cholesky factorization, which was used to compute the exact value of logdet (A). For panels 2(a) and 2(b), the number of cores, np, was set to one. The last panel 2(c) depicts the relative speedup of the approximate algorithm when compared to the baseline solver (at m = 2). Elemental was used as the backend for these experiments. For the approximate algorithm, we report the mean and standard deviation over ten iterations. n 10000 20000 30000 40000 exact 92103.1 198069.0 309268.0 423865.0 logdet (A) mean 92269.5 198397.4 309763.8 424522.4 std 5.51 9.60 20.04 14.80 time (secs) exact mean 18.09 2.87 135.92 12.41 448.02 30.00 1043.74 58.05 std 0.01 0.02 0.12 0.05 √ Table 2: Accuracy and sequential running times (at p = 60, m = 2, and t = 2 log 4n) for diagonally dominant dense random matrices generated using randSPDDenseDD. Baselines were computed using the Cholesky factorization; mean and standard deviation are reported over ten iterations. maintains its relative performance over the exact (Cholesky) factorization. The absolute values for logdet (A) and the corresponding running times, along with the baseline for this experiment are presented in Table 2. We report the numbers in Table 1 at m = 2, at which point we have a low relative error. 5.4 Sparse Matrices Data Synthesis. To generate a sparse, synthetic matrix A ∈ Rn×n , with nnz non-zeros, we use a Bernoulli distribution to determine the location of the non-zero entries and a uniform distribution to generate the values. First, we completely fill the n principle diagonal entries. Next, we generate (nnz − n)/2 index positions in the upper triangle for the non-zero entries by sampling from a Bernoulli distribution with probability (nnz−n)/(n2 −n). We reflect each entry across the principle diagonal to ensure that A is symmetric and we add n to each diagonal entry to ensure that A is 18 name n nnz area of origin thermal2 ecology2 ldoor thermomech TC boneS01 1228045 999999 952203 102158 127224 8580313 4995991 42493817 711558 5516602 Thermal 2D/3D Structural Thermal Model reduction exact 1.3869e6 3.3943e6 1.4429e7 -546787 1.1093e6 logdet (A) approx mean std 1.3928e6 964.79 3.403e6 1212.8 1.4445e7 1683.5 -546829.4 553.12 1.106e6 247.14 time (sec) approx exact mean 31.28 31.24 18.5 10.47 117.91 17.60 57.84 2.58 130.4 8.48 m 149 125 33 77 125 Table 3: Description of the SPD matrices from the University of Florida sparse matrix collection [DH11] that were used in our experiments. All experiments were run sequentially (np = 1) using Eigen. Accuracy results for Algorithm 3 are reported using both the mean and the standard deviation over ten iterations at (with t = 5 and p = 5); we only report the mean for the running times, since the standard deviation is negligible. The exact logdet (A) was computed using the Cholesky factorization. SPD (actually, A is also diagonally dominant). Real Data. To demonstrate the prowess of Algorithm 3 on real-world data, we used SPD matrices from the University of Florida’s sparse matrix collection [DH11]. The complete list of matrices from this collection used in our experiments, as well as a brief description of each matrix, is given in columns 1–4 of Table 3. Evaluation. It is tricky to pick any single method as the “exact method” to compute the logdet (A) for a sparse SPD matrix A. One approach would be to use direct methods such as Cholesky decomposition of A [Dav06, Gup00]. For direct methods, it is difficult to derive an analytical solution for the number of operations required for the factorization as a function of the number of non-zero entries of the matrix, as this is highly dependent on the structure of the matrix [GKK97]. In the distributed setting, one also needs to consider the volume of communication involved, which is often the bottleneck. Alternately, we can use iterative methods to compute the eigenvalues of A [Dav75] and use the eigenvalues to compute logdet (A). It is clear that the worst case performance of both the direct and iterative methods is O(n3 ). However, iterative methods are typically used to compute a few eigenvalues and eigenvectors: therefore, we chose to use the Cholesky factorization based on matrix reordering to compute the exact value of logdet (A). It is important to note that both the direct and iterative methods are notoriously hard to implement, which comes to stark contrast with the almost trivial implementation of Algorithm 3, which is also readily parallelizable. Results. The true power of Algorithm 3 lies in its ability to approximate logdet (A) for sparse A. The Cholesky factorization can introduce O(n2 ) non-zeros during factorization due to fill-in; for many problems, there is insufficient memory to factorize a large, sparse matrix. In our first set of experiments, we wanted to show the effect of m on: (1) convergence of logdet (A), and (2) cost of the solution. To this end, we generated sparse, diagonally dominant SPD matrices of size n = 106 and varied the sparsity from 0.1% to 1% in increments of 0.25%. We did not attempt to compute the exact logdet (A) for these synthetic matrices — our aim was to merely study the speedup √ with m for different sparsities, while t and p were held constant at 2 log 4n and 60 respectively. The results are shown in Figure 3. Figure 3(a) depicts the convergence of logdet (A) measured as a relative error of the current estimate over the final estimate. As can be seen — for well conditioned 19 (a) Convergence with m (b) Cost as a function of m Figure 3: Panels 3(a) and 3(b) depict the effect of the number of terms in the Taylor expansion, m, (see Algorithm 3) on the convergence to the final solution and the time to completion of the approximation. The matrix size was fixed at n = 106 and sparsity was varied as 0.1%, 0.25%, √ 0.5%, 0.75%, and 1%. Experiments were run sequentially (np = 1) and we set p = 60, t = 2 log 4n. For panel 3(a), the baseline is the final value of logdet (A) at m = 25. For panel 3(b), the baseline is the time to completion of the approximation algorithm with m = 1. Eigen was used as the backend for these experiments. matrices — convergence is quick. Figure 3(b) shows the relative cost of increasing m; here the baseline is m = 1. Therefore, the additional cost incurred by increasing m is linear when all other parameters are held constant. The results of running Algorithm 3 on the UFL matrices are shown in Table 3. The numbers reported for the approximation are the mean and standard deviation over ten iterations, t = 5, and p = 5 5 . The value of m was varied between one and 150 in increments of five to select the best average accuracy. The matrices shown in Table 3 have a nice structure, which lends itself to nice reorderings and therefore an efficient computation of the Cholesky factorization. We see that even in such cases, the performance of Algorithm 3 is commendable due to its lower algorithmic complexity; ldoor is the only exception as the approximation takes longer to compute than the Cholesky factorization. In the case of thermomech TC, we achieve good accuracy while achieving a 22x speedup. 6 Conclusions Prior work has presented approximation algorithms for the logarithm of the determinant of a symmetric positive definite matrix; those algorithms either do not work for all SPD matrices, or do not admit a worst-case theoretical analysis, or both. In this work, we presented an approximation algorithm to compute the logarithm of the determinant of a SPD matrix that comes with strong theoretical worst-case analysis bounds and can be applied to any SPD matrix. A simplification of our algorithm delivers relative-error approximation guarantees for a popular special case of SPD 5 We experimented with different p, t and settled on the smallest values that did not result in loss in accuracy. 20 matrices. Using state-of-the-art C++ numerical linear algebra software packages for both dense and sparse matrices, we demonstrated that the proposed approximation algorithm performs remarkably well in practice in serial and parallel environments. 21 References [AT11] H. Avron and S. Toledo. Randomized Algorithms for Estimating the Trace of an Implicit Symmetric Positive Semi-definite Matrix. J. ACM, 58(2):8, 2011. [BP99] Ronald Paul Barry and R Kelley Pace. Monte carlo estimates of the log determinant of large sparse matrices. Linear Algebra and its applications, 289(1):41–54, 1999. [Cur09] Paul F Curran. On a variation of the gershgorin circle theorem with applications to stability theory. 2009. [Dav75] Ernest R Davidson. The iterative calculation of a few of the lowest eigenvalues and corresponding eigenvectors of large real-symmetric matrices. Journal of Computational Physics, 17(1):87–94, 1975. [Dav06] Timothy A Davis. Direct methods for sparse linear systems, volume 2. SIAM, 2006. [dBEG08] Alexandre d’Aspremont, Onureena Banerjee, and Laurent El Ghaoui. First-order methods for sparse covariance selection. SIAM Journal on Matrix Analysis and Applications, 30(1):56–66, 2008. [DH11] Timothy A Davis and Yifan Hu. The university of florida sparse matrix collection. ACM Transactions on Mathematical Software (TOMS), 38(1):1, 2011. [EGV00] Wayne Eberly, Mark Giesbrecht, and Gilles Villard. On computing the determinant and smith form of an integer matrix. In Foundations of Computer Science, 2000. Proceedings. 41st Annual Symposium on, pages 675–685. IEEE, 2000. [FHT08] Jerome Friedman, Trevor Hastie, and Robert Tibshirani. Sparse inverse covariance estimation with the graphical lasso. Biostatistics, 9(3):432–441, 2008. [GJ+ 10] Gaël Guennebaud, Benoı̂t Jacob, et al. Eigen v3. http://eigen.tuxfamily.org, 2010. [GKK97] Anshul Gupta, George Karypis, and Vipin Kumar. Highly scalable parallel algorithms for sparse matrix factorization. Parallel and Distributed Systems, IEEE Transactions on, 8(5):502–520, 1997. [Gup00] Anshul Gupta. Wsmp: Watson sparse matrix package (part-i: direct solution of symmetric sparse systems). IBM TJ Watson Research Center, Yorktown Heights, NY, Tech. Rep. RC, 21886, 2000. [GVDG08] Kazushige Goto and Robert Van De Geijn. High-performance implementation of the level-3 blas. ACM Transactions on Mathematical Software (TOMS), 35(1):4, 2008. [HAB14] Timothy Hunter, Ahmed El Alaoui, and Alexandre Bayen. Computing the logdeterminant of symmetric, diagonally dominant matrices in near-linear time. arXiv preprint arXiv:1408.1693, 2014. 22 [HMAS16] Insu Han, Dmitry Malioutov, Haim Avron, and Jinwoo Shin. Approximating the Spectral Sums of Large-scale Matrices using Chebyshev Approximations. jun 2016. [HMS15] Insu Han, Dmitry Malioutov, and Jinwoo Shin. Large-scale log-determinant computation through stochastic chebyshev expansions. In David Blei and Francis Bach, editors, Proceedings of the 32nd International Conference on Machine Learning (ICML15), pages 908–917. JMLR Workshop and Conference Proceedings, 2015. [HSD+ 13] Cho-Jui Hsieh, Mátyás A Sustik, Inderjit S Dhillon, Pradeep K Ravikumar, and Russell Poldrack. Big & quic: Sparse inverse covariance estimation for a million variables. In Advances in Neural Information Processing Systems, pages 3165–3173, 2013. [KL13] Prabhanjan Kambadur and Aurelie Lozano. A parallel, block greedy method for sparse inverse covariance estimation for ultra-high dimensions. In Proceedings of the Sixteenth International Conference on Artificial Intelligence and Statistics, pages 351– 359, 2013. [LP01] James P LeSage and R Kelley Pace. Spatial dependence in data mining. In Data Mining for Scientific and Engineering Applications, pages 439–460. Springer, 2001. [LZL05] WE Leithead, Yunong Zhang, and DJ Leith. Efficient gaussian process based on bfgs updating and logdet approximation. In the 16th IFAC world congress, 2005. [Mar92] RJ Martin. Approximations to the determinant term in gaussian maximum likelihood estimation of some spatial models. Communications in Statistics-Theory and Methods, 22(1):189–205, 1992. [PB97] R Kelley Pace and Ronald Barry. Quick computation of spatial autoregressive estimators. Geographical analysis, 29(3):232–247, 1997. [PBGS00] R Kelley Pace, Ronald Barry, Otis W Gilley, and CF Sirmans. A method for spatial– temporal forecasting with an application to real estate prices. International Journal of Forecasting, 16(2):229–246, 2000. [PL04] R Kelley Pace and James P LeSage. Chebyshev approximation of log-determinants of spatial weight matrices. Computational Statistics & Data Analysis, 45(2):179–196, 2004. [PMVdG+ 13] Jack Poulson, Bryan Marker, Robert A Van de Geijn, Jeff R Hammond, and Nichols A Romero. Elemental: A new framework for distributed memory dense matrix computations. ACM Transactions on Mathematical Software (TOMS), 39(2):13, 2013. [Reu02] Arnold Reusken. Approximation of the determinant of large sparse symmetric positive definite matrices. SIAM Journal on Matrix Analysis and Applications, 23(3):799– 818, 2002. 23 [SAI16] Arvind K. Saibaba, Alen Alexanderian, and Ilse C. F. Ipsen. Randomized Matrixfree Trace and Log-Determinant Estimators. page 37, may 2016. [SMDS11] John K Salmon, Mark A Moraes, Ron O Dror, and David E Shaw. Parallel random numbers: as easy as 1, 2, 3. In High Performance Computing, Networking, Storage and Analysis (SC), 2011 International Conference for, pages 1–12. IEEE, 2011. [ST04] Daniel A Spielman and Shang-Hua Teng. Nearly-linear time algorithms for graph partitioning, graph sparsification, and solving linear systems. In Proceedings of the thirty-sixth annual ACM symposium on Theory of computing, pages 81–90. ACM, 2004. [Tre11] Luca Trevisan. Graph Partitioning and Expanders. Handout 7, 2011. [ZL07] Yunong Zhang and William E Leithead. Approximate implementation of the logarithm of the matrix determinant in gaussian process regression. journal of Statistical Computation and Simulation, 77(4):329–348, 2007. [ZLLW08] Yunong Zhang, WE Leithead, DJ Leith, and L Walshe. Log-det approximation based on uniformly distributed seeds and its application to gaussian process regression. Journal of Computational and Applied Mathematics, 220(1):198–214, 2008. 24
8cs.DS
Nonfractional Memory: Filtering, Antipersistence, and Forecasting J. Eduardo Vera-Valdés∗ Department of Mathematical Sciences, Aalborg University, and CREATES. arXiv:1801.06677v1 [math.ST] 20 Jan 2018 Abstract The fractional difference operator remains to be the most popular mechanism to generate long memory due to the existence of efficient algorithms for their simulation and forecasting. Nonetheless, there is no theoretical argument linking the fractional difference operator with the presence of long memory in real data. In this regard, one of the most predominant theoretical explanations for the presence of long memory is cross-sectional aggregation of persistent micro units. Yet, the type of processes obtained by cross-sectional aggregation differs from the one due to fractional differencing. Thus, this paper develops fast algorithms to generate and forecast long memory by cross-sectional aggregation. Moreover, it is shown that the antipersistent phenomenon that arises for negative degrees of memory in the fractional difference literature is not present for crosssectionally aggregated processes. Pointedly, while the autocorrelations for the fractional difference operator are negative for negative degrees of memory by construction, this restriction does not apply to the cross-sectional aggregated scheme. We show that this has implications for long memory tests in the frequency domain, which will be misspecified for cross-sectionally aggregated processes with negative degrees of memory. Finally, we assess the forecast performance of high-order AR and ARF IM A models when the long memory series are generated by cross-sectional aggregation. Our results are of interest to practitioners developing forecasts of long memory variables like inflation, volatility, and climate data, where aggregation may be the source of long memory. Keywords: Nonfractional memory, long memory, fractional difference, antipersistence, forecasts. JEL classification: C15, C22, C53. ∗ E-mail: [email protected] Webpage: https://sites.google.com/view/veravaldes/ Address: Skjernvej 4A, 9220 Aalborg East, Denmark. Preprint submitted to arXiv January 23, 2018 1. Introduction Long memory has been a topic of interest in econometrics since Granger’s (1966) study on the shape of the spectrum of economic variables. Granger found that long-term fluctuations in economic variables if decomposed into frequency components are such that the amplitudes of the components decrease smoothly with decreasing period. As shown by Adenstedt (1974), this type of behaviour implies long lasting autocorrelations. The presence of long memory in the data can have perverse effects in estimation and forecasting methods if not included into the modelling scheme, see Beran (1994), and Beran et al. (2013). The autoregressive fractionally integrated moving average (ARF IM A) class of models has become one of the most popular methods to model long memory in the time series literature. They have the appeal of bridging the gap between the stationary autoregressive moving average (ARM A) models and the nonstationary autoregressive integrated moving average (ARIM A) model. ARF IM A models rely on the fractional difference operator to introduce long memory behaviour. Nonetheless, there are currently no economic or financial reasonings implying the fractional difference operator. One of the reasons behind the reliance of the time series literature on generating long memory by the fractional difference operator is the existence of efficient algorithms for their simulation and forecasting. In general, these type of algorithms are not available for other long memory generating schemes. Thus, in this paper, we develop algorithms for memory generation and forecasting by cross-sectional aggregation à la Granger (1980). We then contrast the properties of cross-sectionally aggregated processes to the ones obtained by the fractional difference operator. It is shown that cross-sectional aggregated processes are more flexible than fractionally differenced ones, in the sense that more short term dynamics may be included. Moreover, we show that the antipersistent property of negative autocorrelation for negative degrees of memory does not apply to long memory generated by cross-sectional aggregation. We show that this has repercussions for long memory estimators in the frequency domain. Finally, this paper evaluates the forecasting power of the ARF IM A, and high-order AR models when forecasting long memory generated by cross-sectional aggregation. It finds that high-order AR models beat a pure fractional difference process, I(d), in terms of forecasting performance for some cases. Nonetheless, allowing for short term dynamic in the form of an ARF IM A(1, d, 0) model produces comparable forecast performance as high-order AR models while relying on fewer parameters. This paper proceeds as follows. In Section 2, we present the fractional difference operator commonly used to model long memory in the time series literature. Section 3 discusses crosssectional aggregation as the theoretical explanation behind the presence of long memory in the data and develops a fast algorithm for its generation. Moreover, it contrasts properties of the fractional difference operator against the cross-sectional aggregation scheme. Section 4 discusses the antipersistent property in the context of cross-sectionally aggregated processes. Section 5 constructs minimum squared error forecasts, and studies the theoretical performance of highorder AR, and ARF IM A models when forecasting long memory generated by cross-sectional aggregation. Section 6 concludes. 2. The Fractional Difference Operator The ARF IM A specification due to Granger and Joyeux (1980), and Hosking (1981) has become the standard model to study long memory in the time series literature. They extended the ARM A model to include long memory dynamics by introducing the fractional difference operator (1 − L)d xt = εt , 2 where εt is a white noise process, and d ∈ (−1/2, 1/2). Following the standard binomial expansion, they decompose the fractional difference operator, (1 − L)d , to generate a series given by xt = ∞ X πj εt−j , (1) j=0 with coefficients πj = Γ(j + d)/(Γ(d)Γ(j + 1)) for j ∈ N. Using Stirling’s approximation, it can be shown that these coefficients decay at a hyperbolic rate, πj ≈ j d−1 , which in turn translates to slowly decaying autocorrelations. We write xt ∼ I(d) to denote a process generated by equation (1), that is, an integrated process with parameter d. The properties of the ARF IM A model have been well documented in, among others, Baillie (1996), and Beran et al. (2013). Moreover, fast algorithms have been developed to generate series using the fractional difference operator, see Jensen and Nielsen (2014). Thus, the ARF IM A model has become the canonical construction for modelling and forecasting long memory in the time series literature. Even though the ARF IM A provides a good representation of long memory, and bridges the gap between the stationary ARM A models and the non-stationary ARIM A model, to the best of our knowledge, there is no theoretical reasoning linking the ARF IM A model with the memory found in real data. That is, in contrast with the complete market hypothesis implying that stock prices follow a random walk, or capital depreciation suggesting an autoregressive process, there are no economic or financial arguments for the fractional difference operator to occur in real data. In this regard, the next section presents the most common theoretical motivation behind long memory in the time series literature, cross-sectional aggregation. 3. Cross-Sectional Aggregation Granger (1980), in line with the work of Robinson (1978) on autoregressive processes with random coefficients, showed that aggregating AR(1) processes with coefficients sampled from a Beta distribution can produce long memory. Define a cross-sectional aggregated series as N 1 X xt = √ xi,t , N i=1 (2) where the N individual series are generated as xi,t = αi xi,t−1 + i,t i = 1, 2, · · · , N ; where εi,t is an i.i.d. process with E[2i,t ] = σ2 . Moreover, αi2 is sampled from a Beta distribution with parameters a, b > 0, with density given by B(α; a, b) = 1 αa−1 (1 − α)b−1 B(a, b) for α ∈ (0, 1), where B(·, ·) is the Beta function. Granger showed that, as N → ∞, the autocorrelation function of xt decays at a hyperbolic rate with parameter d = 1 − b/2. Thus, xt has long memory. The cross-sectional aggregation result has been extended in several directions, including to allow for general ARM A processes, as well as to other distributions. See for instance, Linden (1999), Oppenheim and Viano (2004), and Zaffaroni (2004). As argued by Haldrup and Vera Valdés (2017), maintaining the Beta distribution allows us to have closed form representations. 3 Furthermore, Beran et al. (2010) proposed a method to estimate the parameters from the Beta distribution from the individual series, xi,t above. In applied work, cross-sectional aggregation has been cited as source of long memory for many series. To name but a few, it has been proposed for inflation, output, and volatility; see Balcilar (2004), Diebold and Rudebusch (1989), and Osterrieder et al. (2015). Even though a cross-sectional aggregated process has long memory, Haldrup and Vera Valdés (2017) show that the resulting series does not belong to the ARF IM A class of processes. Pointedly, they show that the fractionally differenced cross-sectional aggregated process does not follow an ARM A specification. The next subsections expands on the cross-sectional aggregation literature by developing fast algorithms for its generation, and contrasting its theoretical properties to the ones of the fractional difference operator. 3.1. Nonfractional Memory Generation As can be seen by its definition, Equation (2), generating cross-sectionally aggregated processes is computationally demanding. For each cross-sectionally aggregated process, we need to simulate a vast number of AR(1) processes, which are in turn computationally demanding. Haldrup and Vera Valdés (2017) suggest that to get a good approximation, the cross-sectional dimension should increase as the sample size. The computational demands are thus particularly large for Monte Carlo type of analysis on cross-sectionally aggregated processes. In what follows, we use the theoretical autocorrelations of the cross-sectionally aggregated process to present an algorithm that makes generation of long memory by cross-sectional aggregation comparable to fractional differencing in terms of computational requirements. Denote xt ∼ CSA(a, b) to a series generated by cross-sectional aggregation of autoregressive parameters sampled from the Beta distribution B(a, b), Equation (2). The notation makes explicit the origin of the memory by cross-sectional aggregation and its dependence on the two parameters of the Beta distribution. Theorem 1 develops a fast algorithm to generate cross-sectionally aggregated processes. Theorem 1. Let xt ∼ CSA(a, b) defined as in (2), then xt can be computed as the first T elements of the (2T − 1) × 1 vector T −1 F (F̄ φ̃ F̄ ε̃), where F̄ is the discrete Fourier transform, T −1 F is the inverse transform, and ’ ’ denotes multiplication element by element. Furthermore, z̃ is a (2T −1)×1 vector given by z̃ := [z0 , z1 , · · · , zT −1 , 0, · · · , 0] for zt = φt , εt where εj ∼ i.i.d.N (0, σ2 ), and φj = (B(p + j, q)/B(p, q))1/2 , ∀j ∈ N. Proof: See appendix A. The theorem is an application of the circular convolution theorem for the coefficients associated with the cross-sectionally aggregated process. In this sense, it is in line with the algorithm of Jensen and Nielsen (2014) for the fractional difference operator and thus achieves equivalent computational efficiency. The difference in computation times between the standard simulation of a cross-sectional aggregated times series, Equation (2), and the fast implementation in Theorem 1 are very large for all sizes. For small series (T ≈ 102 ) the gains are in the hundreds of times faster, while in the thousands for medium sized series (T ≈ 104 ). This is of course not surprising given that, for a sample of size T , the number of computations needed for the standard implementation is of order N T 2 , with N the cross-sectional dimension, which, as argued before, should increase as T does. Meanwhile, the computational requirements for the fast implementation are of order T log (T ). Codes implementing the algorithm for memory generation by cross-sectional aggregation are available in Appendix B, and on the author’s github repository at github.com/everval/Nonfractional. 4 To get a better understanding of the dynamics of the long memory by cross-sectional aggregation, the next sections compare it against the pure fractional noise, that is, a fractionally differenced white noise. However, note that we can allow more short-term dynamic by adding AR and M A filters to both specifications. 3.2. Nonfractional Memory and Fractional Difference Operator One notable difference between a cross-sectionally aggregated process and a fractionally differenced one is the number of parameters needed for its generation. Fractionally differenced processes rely on only one parameter to model the entire series, whilst a cross-sectionally aggregated process uses two. As we will see below, this gives the cross-sectional aggregation procedure more flexibility. Let γI(d) (·), and γCSA(a,b) (·) be the autocorrelation functions of an I(d), and a CSA(a, b) process, respectively. That is, γI(d) (k) = Γ(k + d)Γ(1 − d) ≈ k 2d−1 Γ(k − d + 1)Γ(d) as k → ∞ (3) B(a + k/2, b − 1) ≈ k 1−b as k → ∞, (4) B(a, b − 1) which show that both processes have hyperbolic decaying autocorrelations. As argued by Granger (1980), the asymptotic behaviour of the autocorrelation function for a cross-sectionally aggregated process, Equation (4), only depends on the second argument of the Beta function. In this context, both processes show the same rate of decay for the autocorrelation function for b = 2(1 − d). Moreover, the first argument of the Beta function allows us to introduce short term dynamics. Figure 1 shows the autocorrelation function of cross-sectional aggregated processes for different values of ‘a’, the first parameter of the Beta distribution. The figure shows that as the first parameter gets larger, so does the autocorrelation function for the initial lags. Thus, ‘a’ acts as a short memory modulator. γCSA(a,b) (k) = Figure 1: Autocorrelation function for an CSA(a, b) processes for different values of ‘a’ while having the same asymptotic behaviour. Section 5 will look at the forecasting performance of ARF IM A models when working with cross-sectionally aggregated processes, but for now suppose we are interested in looking in the other direction. That is, assume we have a process generated by the fractional difference operator and we want to approximate it with a cross-sectionally aggregated process. We show that we can use the first argument to generate such a series. Consider the loss function L(k, a, d) := k X 2 γI(d) (i) − γCSA(a,2(1−d)) (i) , i=0 5 (5) which measures the squared difference between autocorrelations at the first k lags for crosssectionally aggregated and fractionally differenced processes with the same long memory dynamics, hence we set b = 2(1 − d) for the cross-sectionally aggregated process. Minimizing (5) with respect to ‘a’, allows us to find the cross-sectional aggregated process that best approximates an I(d) one up to lag k, while having the same long term dynamics. Given the different forms of the autocorrelation functions, (3), and (4), there is in general not a value of ‘a’ that minimizes (5) for all values of k; for instance, mina L(2, a, 0.2) = 0.118, while mina L(30, a, 0.2) = 0.121. Moreover, it will depend on d. Nonetheless, selecting a medium sized k, say k ≈ 10, the approximation turns out to be quite 4 good in general. In Figure 2, we present a filtered {εt }10 t=1 ∼ N (0, 1) vector using the fractional difference operator with long memory parameter d = 0.2, and using the cross-sectional aggregated algorithm with parameters a = 0.12, and b = 1.6, so that they show similar short and long memory behaviours. Figure 2: Filtered series and autocorrelation function for a fractional differenced process I(0.2), and a cross-sectional aggregated one CSA(0.12, 1.6). The figure shows that the filtered series are quite similar. This behaviour can further be seen on the autocorrelation functions showing similar dynamics. Thus, the figure shows that it is possible to generate cross-sectionally aggregated processes that closely mimic ones due to fractional differencing. 4. Nonfractional Memory and Antipersistent Processes It is well known in the long memory literature that the fractional difference operator implies that the autocorrelation function is negative for negative degrees of memory, d ∈ (−1/2, 0). This can be seen in (3) where the sign of γI(d) (k) depends on Γ(d) in the denominator, which is negative for d ∈ (−1/2, 0). Furthermore, the behaviour of the spectral density for a fractionally differenced process near the origin is given by fI(d) (λ) ≈ c0 λ−2d as λ → 0, (6) where c0 is a constant. Thus, fI(d) (λ) → 0 as λ → 0, that is, the fractional difference operator for negative degrees of memory imply a spectral density collapsing to zero at the origin. These properties, among other related ones, has been named antipersistence in the literature. We argue that cross-sectionally aggregated processes do not share these features. First, Equation (4) shows that the autocorrelation function for the cross-sectionally aggregated process only depends on the Beta function, which is always positive. Figure 3 shows the autocorrelation function for both fractionally differenced and cross-sectional aggregated processes for a negative degree 6 of memory, d = −0.2. The figure shows that, even though both processes show the same rate of decay in their autocorrelation functions, they have opposite signs. Figure 3: Autocorrelation functions for a fractional differenced process, I(−0.2), and a cross-sectional aggregated one, CSA(0.09, 2, 4). The right plot shows a close-up for lags 70 to 110. Then, given the positive sign for all autocorrelations, Theorem 2 shows that the spectral density for the cross-sectionally aggregated process for negative degrees of memory converges to a positive constant. Theorem 2. Let xt ∼ CSA(a, b) defined as in (2) with b ∈ (2, 3) so that the long memory parameter is in the negative range, d ∈ (−1/2, 0), then, the spectral density of xt at the origin is positive. That is, fCSA(a,b) (0) = ca,b > 0, (7) where ca,b depends on the parameters of the Beta distribution. Proof: See appendix A. Figure 4 shows the periodogram, an estimate of the spectral density, for CSA(a, b) and I(d) processes of size T = 104 averaged for 104 replications. The figure shows that the periodogram for both processes show similar patterns for positive degrees of memory, both diverging to infinity at the same rate. Nonetheless, for negative degrees of memory, the periodogram collapses to zero as the frequency goes to zero for the fractionally differenced process, while it converges to a constant for the cross-sectionally aggregated process. Moreover, the latter property has implications for estimation and inference. In particular, tests for long memory in the frequency domain will be affected. These tests are based on the rate to which the periodogram goes to zero as an estimator for long memory using the log-periodogram regression, see Geweke and Porter-Hudak (1983), and Robinson (1995). The log-periodogram regression is given by log(fˆX (λj )) = a − 2d log(λj ) + uj , where fˆX (·) is the periodogram, a is a constant, and uj is the error term. From (6), note that the log-periodogram regression provides an estimate of the long memory parameter, d, for fractionally differenced processes. Tests in the frequency domain use this expression to estimate the degree of memory. Nonetheless, as Theorem 2, and Figure 4 show, these tests will be misspecified for long memory by cross-sectional aggregation for negative degrees of memory. To illustrate the misspecification problem, Table 1 reports the degree of long memory estimated by the method of Geweke and Porter-Hudak, 1983, GP H hereinafter, for several degrees of memory for both fractionally differenced, and cross-sectionally aggregated processes. 7 Figure 4: Mean periodogram of a fractional differenced process, I(d), [top], and a cross-sectional aggregated process, CSA(0.2, 2(1 − d)), [bottom]. A sample size of T = 10, 000 was used, with 10,000 replications, and d = 0.4, −0.4. Table 1: Mean, and standard deviation in parentheses, of estimated long memory by the GP H method for CSA(0.2, 2(1 − d)) and I(d) processes. We use the standard T 1/2 bandwidth, where the sample size is T = 104 , and with 104 replications. d = 0.2 d = −0.2 d = −0.4 d = 0.4 CSA(a, b) I(d) CSA(a, b) I(d) CSA(a, b) I(d) CSA(a, b) I(d) 0.4062 0.4034 0.2628 0.2011 0.1036 -0.1985 0.0653 -0.3927 (0.0701) ( 0.0697) (0.0697) (0.0696) (0.0687) (0.0685) (0.0695) (0.0717) The table shows that the estimator is relatively close to the true memory for both processes when the memory is positive, if slightly overshooting it for the cross-sectionally aggregated process, as reported by Haldrup and Vera Valdés (2017). This contrasts to the case of negative memory, where the table shows that the estimator remains precise for the fractionally differenced series, while it is incapable of detecting the long memory in the cross-sectionally aggregated processes, with estimators not statistically different from zero. This is of course not surprising in light of Theorem 2. In sum, the lack of the antipersistent property in cross-sectionally aggregated processes shows that care must be taken when estimating long memory for negative degrees of memory if the long memory generating mechanism is not the fractional difference operator. Further analysis of negative degrees of memory and antipersistence is a line of inquiry open for further research. 5. Nonfractional Memory Forecasting In the time series literature, some effort has been directed to assess the performance of the ARF IM A type of models when forecasting long memory processes. For instance, Ray (1993) calculates the percentage increase in mean-squared error (M SE) from forecasting I(d) series with AR models. She argues that the M SE may not increase significantly, particularly when we do not know the true long memory parameter. Crato and Ray (1996) compare the forecasting performance 8 of ARF IM A models against ARM A alternatives and find that ARF IM A models are in general outperformed by ARM A alternatives. Moreover, Man (2003) argues that an ARM A(2, 2) model compares favourably to an I(d) for short-term forecasts of long memory time series with fractionally differenced structure. One thing that these forecasting comparison studies have in common is the underlying assumption that long memory is generated by an ARF IM A process. In this context, a priori the forecasting exercises assume that a fractionally differenced process is the correct specification. As previously discussed, even though cross-sectional aggregation does generate long memory, the series does not follow an ARF IM A specification. The question remains whether an ARF IM A model serves as a good approximation for forecasting purposes. The next subsection computes the minimum square error forecasts for cross-sectional aggregated processes, CSA(a, b). With those as benchmark, the following subsections evaluate the forecasts performance of high order AR, and ARF IM A models on CSA(a, b) processes. 5.1. Minimum Square Error Forecasts Theorem 3 computes the minimum mean square error forecasts for a series generated by crosssectional aggregation. Theorem 3. Let xt ∼ CSA(a, b) defined as in (2), and let h ∈ N, then the minimum mean square error forecast h periods ahead, x̂t+h , can be computed as1 x̂T +h = T X φj νT −j+h , (8) j=h where νi is computed as νi = xi − i X φj νi−j ∀i ∈ 0, 1, · · · , T , j=1 with φj as in Theorem 1. Proof: See Appendix A. The theorem relies on the M A(∞) representation to construct the forecasts. Algorithms for computing the minimum mean square forecasts, (8), are presented in Appendix B, and are available on the author’s github repository at github.com/everval/Nonfractional. Theorem 3 allows for the construction of forecasts for the correct theoretical specification. Note that in the theorem we have assumed that we know the parameters of the the Beta distribution for the cross-sectionally aggregated process; for empirical applications we can use Beran et al.’s (2010) method to estimate them. In the following, we will use these computations to assess the forecasting performance of high-order AR, and ARF IM A models when working on long memory generated by cross-sectional aggregation. 5.2. Forecasts With AR(p) Models This subsection computes the forecasting efficiency loss of an AR(p) model when working with a CSA(a, b) process. Theorem 4 obtains the parameter estimates for an AR(p) model fitted to a CSA(a, b) process, and computes the efficiency loss for one-step ahead forecasts. 1 The Theorem assumes a Type II process analogous to samples of fractionally differenced processes, see Davidson (2009). In particular, it assumes νj = 0, ∀j < 0. 9 Theorem 4. Let xt ∼ CSA(a, b) defined as in (2), and estimate an AR(p) given by (1 − α1 L − α2 L2 − · · · − αp Lp )xt = ut , then   1 α1  α2   γCSA(a,b) (1)     ..  =  ..  .   .  αP −1   · · · γCSA(a,b) (p − 1) γCSA(a,b) (1)   · · · γCSA(a,b) (p − 2)  γCSA(a,b) (2)   , .. .. ..    . . . γCSA(a,b) (1) 1 .. . γCSA(a,b) (p − 1) γCSA(a,b) (p − 2) · · · 1 γCSA(a,b) (p) with γCSA(a,b) (·) defined as in 4. Furthermore, the one-step ahead forecast error variance relative to the minimum square error forecasts, ζAR(p) , is given by  ζAR(p) = B(a, b − 1) B(a, b)    1+ p X ! αi2 +2 i=1 p X i=1  γCSA(a,b) (i) −αi + p−i X  αj αj+i  . j=1 Proof: See Appendix A. As an example, consider estimating an AR(1) model to forecast a cross-sectionally aggregated process. From Theorem 4, the autoregressive parameter is α1 = γCSA(a,b) (1) = B(a + 1/2, b − 1)/B(a, b−1), while the one-step ahead forecast error variance is ζAR(1) = (B(a, b−1)/B(a, b))(1− 2 γCSA(a,b) (1)), which shows that the efficiency loss is a nonlinear function of both parameters of the CSA(a, b) process. To get a better sense of the nonlinearity, Figure 5 presents both the autoregressive parameter, α1 , and the efficiency loss, ζAR(p) , while varying the first parameter of the cross-sectional aggregated process, ‘a’, with fixed b = 1.8. On the one hand, the figure shows that as ‘a’ increases, so does the autoregressive parameter, this is in line with the discussion above regarding the first parameter as a short memory regulator. On the other hand, the one-step ahead forecast error variance shows a maximum around a = 0.5, with a forecast error variance of almost 15%. That is, more than double the reported by Man (2003) for an I(0.1) process. Thus, the figures suggest that we could expect worse performance using an AR(1) process to forecast a CSA(a, b) process than a pure I(d) one. Figure 5: Estimated autoregressive parameter and forecast error variance of fitted AR(1) on CSA(a, b) processes relative to the optimum forecasts while varying ‘a’, for different degrees of memory. Table 2 shows the one-step ahead forecast error variance of fitted AR(1) models on CSA(a, b) processes for different values of a, b. The losses are in line with the ones computed by Man (2003) for the I(d) case, increasing as the memory of the process increases. Moreover, as noted by Ray (1993) and Man (2003), we can reduce the one-step ahead forecast error variance by allowing for more lags in the AR specification. To get a sense of the improvements we can achieve, Table 2 also 10 Table 2: One-step ahead forecast error variance of fitted AR(1) and AR(20) models on CSA(a, b) processes relative to the optimum forecasts shown in (8). The brackets show the associated long memory parameter d. AR(1) AR(20) a\b [d] 1.8 [0.1] 1.6 [0.2] 1.4 [0.3] 1.2 [0.4] 1.1 [0.45] 1.8 [0.1] 1.6 [0.2] 1.4 [0.3] 1.2 [0.4] 1.1 [0.45] 0.1 0.5 0.9 1.3 1.7 1.085 1.145 1.129 1.110 1.095 1.110 1.172 1.146 1.122 1.104 1.154 1.211 1.170 1.137 1.114 1.257 1.273 1.202 1.156 1.126 1.387 1.320 1.223 1.168 1.133 1.071 1.111 1.099 1.086 1.075 1.085 1.123 1.107 1.091 1.079 1.104 1.137 1.115 1.096 1.082 1.129 1.153 1.124 1.101 1.085 1.144 1.161 1.128 1.103 1.086 presents the one-step ahead forecast error variance of AR(20) models fitted to CSA(a, b) processes for different values of a, b. As the table shows, increasing the order of the AR model can greatly reduce the one-step ahead forecast error variance, particularly for larger degrees of long memory. 5.3. Forecasts With Fractional Models This section studies the forecast performance of fractional models when working on crosssectional aggregated processes. In particular, we compute the one-step ahead forecast error variance of the pure I(d) model and of an ARF IM A(1, d, 0) allowing for more short term dynamics. Theorem 5. Let xt ∼ CSA(a, b) defined as in (2), then the one-step ahead forecast error variance of the d = 1 − b/2 fractional difference of the series, ζI(d) , is given by ζI(d) = γz (0). Furthermore, estimate an ARF IM A(1, d, 0) by fitting an AR(1) model to the d = 1 − b/2 fractional difference of the series, then the autoregressive parameter, αI , and the one-step ahead forecast error variance, ζARF IM A(1,d,0) , are given by αI = γz (1) , γz (0) ζARF IM A(1,d,0) = γz (0)2 − γz (1)2 . γz (0)2 Where in both expressions γz (k) is given by   γ ∗ (k) 1 γz (k) = B(a, b − 1) (F1 (k) − 1) + B(a + , q − 1)F2 (k) , B(a, b) 2 with γ ∗ (k) = σε2 Γ(1 + 2d) Γ(−d − k) , Γ(−d)Γ(1 + d) Γ(1 + d − k) and      1 − d + k −d + k 2+d+k 1+d+k F1 (k) := F 1, a, , , a + b − 1, , ,1 + 2 2 2 2      1 − d − k −d − k 2+d−k 1+d−k F 1, a, , , a + b − 1, , ,1 , 2 2 2 2 −d + k F2 (k) := ∗ 1+ d + k     1 1−d+k 2−d+k 1 2+d+k 3+d+k F 1, p + , , , a+b− , , ,1 2 2 2 2 2 2 −d − k + ∗ 1+d−k 11  F 1 1−d−k 2−d−k 1, a + , , 2 2 2     1 2+d−k 3+d−k , a+b− , ,1 , , 2 2 2 where d = 1 − b/2, and F [·] is the generalized hypergeometric function. Proof: See Appendix A. Table 3 presents the one-step ahead forecast error variance of fitted ARF IM A(1, d, 0), and I(d) models on CSA(a, b) processes for different values of a, b. The table also shows the estimated autoregressive parameter of the fitted ARF IM A(1, d, 0) model. As the table shows, the pure fractional differenced process can be quite bad at forecasting CSA(a, b) processes, specially for higher values of ‘a’. Once again, relating the first parameter functioning as a short memory regulator. Once we allow for more short term dynamics in the form of an ARF IM A(1, d, 0) model, the forecasting performance is much in line with the one from high order AR models. Thus, it greatly helps to allow for some short term dynamics in the modelling scheme. Table 3: One-step ahead forecast error variance, and estimated autoregressive parameter in parenthesis, of fitted I(d) and ARF IM A(1, d, 0) models on CSA(a, b) processes. The brackets show the associated long memory parameter d. I(d) ARF IM A(1, d, 0) a\b [d] 1.8 [0.1] 1.6 [0.2] 1.4 [0.3] 1.2 [0.4] 1.1 [0.45] 1.8 [0.1] 0.1 1.077 1.084 1.112 1.158 1.186 0.5 1.345 1.253 1.202 1.176 1.169 0.9 1.615 1.435 1.318 1.240 1.212 1.3 1.880 1.611 1.431 1.309 1.263 1.7 2.138 1.778 1.538 1.374 1.312 1.072 ( 0.067 1.127 ( 0.402 1.118 ( 0.555 1.104 ( 0.642 1.093 ( 0.699 1.6 [0.2] 1.083 ) ( -0.019 ) 1.132 ) ( 0.312 ) 1.121 ) ( 0.468 ) 1.107 ) ( 0.559 ) 1.096 ) ( 0.620 ) 1.4 [0.3] 1.2 [0.4] 1.1 [0.45] 1.103 ( -0.091 ) 1.138 ( 0.229 ) 1.123 ( 0.384 ) 1.108 ( 0.475 ) 1.097 ( 0.536 ) 1.131 ( -0.153 ) 1.147 ( 0.156 ) 1.125 ( 0.305 ) 1.107 ( 0.393 ) 1.095 ( 0.451 ) 1.147 ( -0.180 ) 1.152 ( 0.123 ) 1.125 ( 0.268 ) 1.106 ( 0.352 ) 1.093 ( 0.408 ) Comparing the results from Tables 2 and 3, we see that, as was the case for long memory series generated by the fractional difference operator, a high order AR(p) can be a good model for forecasting long memory at short horizons. Yet, the bias-variance trade-off has to be assessed given the number of estimated parameters. As an alternative, the ARF IM A(1, d, 0) produces similar results while relying in only two parameters. 6. Conclusions Even though there is no theoretical argument linking the presence of long memory in the data with the fractional difference operator, fractionally differenced processes remain the most popular construction in the long memory time series literature. This may be due to the existence of efficient algorithms for their simulation and forecasting. Thus, this paper presents fast algorithms to generate long memory by a theoretically based mechanism that do not rely on the fractional difference operator, cross-sectional aggregation. The cross-sectional aggregated process is then contrasted to the fractional difference operator. In particular, the paper analyses the antipersistent phenomenon. It is proven that, for negative degrees of memory, while the autocorrelations are negative by definition for the fractional difference operator, this restriction does not apply to the cross-sectional aggregated scheme. Furthermore, the paper shows that the lack of antipersistence for cross-sectional aggregated processes 12 has implications for long memory estimators in the frequency domain which will be misspecified in general. Moreover, this work evaluated the efficiency loss of using high-order AR and ARF IM A models to forecast long memory series generated by cross-sectional aggregation. It finds that, at short horizons, high-order AR models beat a pure fractional difference process, I(d), in terms of forecasting performance. Nonetheless, allowing for short term dynamic in the form of an ARF IM A(1, d, 0) model produces comparable forecast performance as high-order AR models while relying on less parameters. The results of this paper can be used in the context of Monte Carlo simulations of long memory estimators and forecasts. Of particular interest is the analysis of long memory in inflation, one of the primer examples of cross-sectional aggregation producing long memory due to the way the data is computed. Moreover, the results allow for financial econometrics and climate econometrics models to incorporate long memory forecasts consistent with the theoretical argument posed by cross-sectional aggregation. Acknowledgements I would like to thank Niels Haldrup for the careful reading of this article and all the insightful comments. References Adenstedt, R. K. (1974). On Large-Sample Estimation for the Mean of a Stationary Random Sequence. The Annals of Statistics, 2(6):1095–1107. Baillie, R. T. (1996). Long memory processes and fractional integration in econometrics. Journal of Econometrics, 73(1):5–59. Balcilar, M. (2004). Persistence in inflation: does aggregation cause long memory? Emerging Markets Finance and Trade, 40(5):25–56. Beran, J. (1994). Statistics for long-memory processes. Chapman & Hall. Beran, J., Feng, Y., Ghosh, S., and Kulik, R. (2013). Long-Memory Processes: probabilistic theories and Statistical Methods. Springer. Beran, J., Schtzner, M., and Ghosh, S. (2010). From short to long memory: Aggregation and estimation. Computational Statistics and Data Analysis, 54(11):2432–2442. Crato, N. and Ray, B. K. (1996). Model selection and forecasting for long-range dependent processes. Journal of Forecasting, 15(2):107–125. Davidson, J. (2009). When is a time series I(0)? In Castle, J. and Shepherd, N., editors, The Methodology and Practice of Econometrics, chapter 13. Oxford University Press. Diebold, F. X. and Rudebusch, G. D. (1989). Long Memory and Persistence in Agregate Output. Journal of Monetary Economics, 24(2):189–209. Geweke, J. and Porter-Hudak, S. (1983). The estimation and application of long memory time series models. Journal of Time Series Analysis, 4(4):221–238. Granger, C. W. (1966). The Typical Spectral Shape of an Economic Variable. Econometrica, 34(1):150–161. Granger, C. W. (1980). Long memory relationships and the aggregation of dynamic models. Journal of Econometrics, 14(2):227–238. Granger, C. W. and Joyeux, R. (1980). An Introduction to Long Memory Time Series Models and Fractional Differencing. Journal of Time Series Analysis, 1(1):15–29. Haldrup, N. and Vera Valdés, J. E. (2017). Long memory, fractional integration, and cross-sectional aggregation. Journal of Econometrics, 199(1):1–11. Hosking, J. R. M. (1981). Fractional differencing. Biometrika, 68(1):165–176. Jensen, A. N. and Nielsen, M. Ø. (2014). A Fast Fractional Difference Algorithm. Journal of Time Series Analysis, 35(5):428–436. Linden, M. (1999). Time series properties of aggregated AR(1) processes with uniformly distributed coefficients. Economics Letters, 64(1):31–36. Man, K. S. (2003). Long memory time series and short term forecasts. International Journal of Forecasting, 19(3):477–491. Oppenheim, G. and Viano, M. C. (2004). Aggregation of random parameters ornstein-uhlenbeck or ar processes: Some convergence results. Journal of Time Series Analysis, 25(3):335–350. 13 Osterrieder, D., Ventosa-Santaulària, D., and Vera-Valdés, J. E. (2015). Unbalanced Regressions and the Predictive Equation. CREATES Research Paper, 9. Ray, B. K. (1993). Modeling Long Memory Processes for Optimal Long Range Prediction. Journal of Time Series Analysis, 14(5):511–525. Robinson, P. M. (1978). Statistical Inference for a Random Coefficient Autoregressive Model. Scandinavian Journal of Statistics, 5(3):163–168. Robinson, P. M. (1995). Log-Periodogram Regression of Time Series with Long Range Dependence. The Annals of Statistics, 23(3):1048–1072. Zaffaroni, P. (2004). Contemporaneous aggregation of linear dynamic models in large economies. Journal of Econometrics, 120(1):75–102. Appendix A. Proofs Proofs of Theorem 1, and Theorem 2 Lemma 1 of Haldrup and Vera Valdés (2017), which is in turn an application of the Central Limit Theorem, show that in the limit, the cross-sectionally aggregated process, Equation (2), allows an M A(∞) representation. Using this representation, Theorem 1 is an application of the circular convolution Theorem, see Jensen and Nielsen (2014). Moreover, note that the spectral density for xt evaluated at the origin is given by ∞ σε2 X fX (0) = φj 2π 2 , j=0 where φj are the coefficients of the M A(∞) representation given by φj = the Beta function. Thus, fX (0) = ∞ σε2 X φj 2π j=0 = 2 ∞ X σε2 B(a + j, b)1/2 = 2πB(a, b) j=0 ∞ σε2 Γ(b) X Γ(a + j)1/2 2πB(a, b) Γ(a + j + b)1/2 j=0 2 2 B(a+j,b)1/2 , B(a,b)1/2 with B(·, ·), ∞ X σε2 Γ(a + j)1/2 Γ(b)1/2 = 2πB(a, b) Γ(a + j + b)1/2 j=0 ∞ σε2 Γ(b) X −b/2 ≈ j 2πB(a, b) j=0 2 2 = σε2 Γ(b) ζ(−b/2)2 2πB(a, b) = ca,b < ∞, where Γ(·), ζ(·) are the Gamma and EulerRiemann zeta function, respectively. Moreover, note that all terms in the expression are positive and thus ca,b > 0. Proof of Theorem 3 Assuming a Type II process, the Lemma shows that xt can be expressed as      1 φ1 φ2 · · · φT xT νT xT −1  0 1 φ1 · · · φT −1  νT −1       xT −2  0 0 1 · · · φT −2  νT −2  = ,    ..   .. .. .. . . ..   ..   .  . .   . . .  . ν0 0 0 0 ··· 1 x0 (9) with φt , νt defined in Theorem 1. Solving for νj , this represents a system with an upper triangular matrix to which fast algorithms exist. Once we have solved it, the minimum mean square error forecast h periods ahead can be easily obtained. 14 Proofs of Theorem 4, and Theorem 5 P∞ Let xt ∼ CSA(a, b) and denote Φ(z:,t ) := j=0 φj zt−j the function that applies the filter associated to the M A(∞) representation of the CSA(a, b) process to z:,t = (zt , zt−1 , · · · ). That is, xt = ∞ X φj νt−j := Φ(ν:,t ), j=0 where νj ∼ N (0, σε2 ) and φj = (B(p + j, q)/B(p, q))1/2 , ∀j ∈ N. Let ut be the residual of a fitted AR(p) process to xt , i.e., ut = (1 − α1 L − α2 L2 − · · · − αp Lp )xt , where L is the lag operator and {α1 , · · · , αp } are the fitted parameters. Hence, ut = (1 − α1 L − α2 L2 − · · · − αp Lp )xt = (1 − α1 L − α2 L2 − · · · − αp Lp )Φ(ν:,t ) = Φ(ν:,t ) − α1 Φ(ν:,t−1 ) − α2 Φ(ν:,t−2 ) − · · · − αp Φ(ν:,t−p ) = νt φ0 + νt−1 (φ1 − α1 φ0 ) + · · · + νt−s (φs − · · · − αp φs−p ) + · · · ∞ X = φ̃j νt−j , j=0 where  φ0     φ   1 − α1 φ0 φ1 − α1 φ1 − α2 φ0 φ˜j :=  ..   .    φj − · · · − αp φj−p j=0 j=1 j=2 .. . j≥p Thus, var(ut )σε = ∞ X 2 φ˜j j=0 = φ20 2 + (φ1 − α1 φ0 ) + · · · + ∞ X (φj − α1 φj−1 − · · · − αp φj−p )2 j=p     ∞ ∞ X X =  φ2j  (1 + α12 + · · · + αp2 ) + 2  φj φj−1  (−α1 + α1 α2 + · · · + αp−1 αp ) j=0  +2  j=1 ∞ X   φj φj−2  (−α2 + α1 α3 + · · · + αp−2 αp ) + 2  j=2 ∞ X  φj φj−p  (−αp ) j=p  B(a, b − 1)  (1 + α12 + · · · + αp2 ) + 2γCSA(a,b) (1)(−α1 + α1 α2 + · · · + αp−1 αp ) B(a, b)  + · · · + 2γCSA(a,b) (p)(−αp ) .  = Which shows that minimizing with respect to {α1 , · · · , αp } yields the Yule-Walker system of equations, the result follows. Finally, for Theorem 5, the first part regarding the autocorrelation function of a fractionally differenced cross-sectional aggregated process is Theorem 3 in Haldrup and Vera Valdés (2017). For the second part, the proof follows the same steps as for the pure AR(p) case above replacing the autocorrelation function of the pure CSA(a, b) process for the one of the fractionally differenced one. 15 B. Code Implementation Nonfractional Memory Generation R code c s a d i f f = function ( x , p , d ) { iT = length ( x ) n = nextn ( 2 ∗iT − 1 , 2 ) k = 0 : ( iT −1) b = ( beta ( p+k , d ) /beta ( p , d ) ) ˆ ( 1 / 2 ) dx = f f t ( f f t ( c ( x , rep ( 0 , n − iT ) ) ) ∗ f f t ( c ( b , rep ( 0 , n − iT ) ) ) , inverse = T) / n ; return (Re( dx [ 1 : iT ] ) ) } Matlab code function [ cx ] = c s a d i f f ( x , p , q ) T = size (x , 1 ) ; np2 = 2 . ˆ nextpow2( 2 ∗T−1); c o e f s = [ 1 , ( beta ( p +(1:T−1) , q ) . / . . . beta ( p , q ) ) . ˆ ( 1 / 2 ) ] ; cx = i f f t ( f f t ( x , np2 ) . ∗ f f t ( c o e f s ’ , np2 ) ) ; cx = cx ( 1 : T, : ) ; end Nonfractional Memory Forecasts R code get e r r c s a = function ( x , p , q ) { iT = length ( x ) k = 0 : ( iT −1) b = ( beta ( p+k , q ) /beta ( p , q ) ) ˆ ( 1 / 2 ) zen = c ( 1 , rep ( 0 , iT −1)) ma = pracma : : T o e p l i t z ( b , zen ) v j s = solve (ma, x ) return ( v j s ) } f o r e c a s t c s a = function ( x , p , q , h ) { iT = length ( x ) k = seq ( 1 , ( iT ) , 1 ) b = ( beta ( p+k , q ) /beta ( p , q ) ) ˆ ( 1 / 2 ) e r r = get e r r c s a ( x , p , q ) e r r v = rev ( e r r ) f x = rep ( 0 , h ) for ( i i i n 1 : h ) { f x [ i i ] = crossprod ( b [ i i : iT ] , e r r v [ 1 : ( iT− i i + 1 ) ] ) } return ( f x ) } 16 Matlab code function [ v j s ] = g e t e r r c s a ( x , p , q ) T = size (x , 1 ) ; c o e f s = [ 1 , ( beta ( p +(1:T−1) , q ) . / beta ( p , q ) ) . ˆ ( 1 / 2 ) ] ; zen = [ 1 , zeros ( 1 ,T− 1 ) ] ; ma = t o e p l i t z ( c o e f s , zen ) ; v j s = ma\x ; end function [ f x ] = f o r e c a s t c s a ( x , p , q , h ) T = size (x , 1 ) ; c o e f s = [ 1 , ( beta ( p +(1:T−1) , q ) . / beta ( p , q ) ) . ˆ ( 1 / 2 ) ] ; coefs = fliplr ( coefs ); err = get err csa (x , p , q ) ; f x = zeros ( h , 1 ) ; for i i = 1 : h f x ( i i , 1 ) = c o e f s ( 1 , 1 : T− i i ) ∗ e r r ( i i +1:T , 1 ) ; end end 17
10math.ST
IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES 1 Learning to Play Othello with Deep Neural Networks arXiv:1711.06583v1 [cs.AI] 17 Nov 2017 Paweł Liskowski, Wojciech Jaśkowski, Krzysztof Krawiec Abstract—Achieving superhuman playing level by AlphaGo corroborated the capabilities of convolutional neural architectures (CNNs) for capturing complex spatial patterns. This result was to a great extent due to several analogies between Go board states and 2D images CNNs have been designed for, in particular translational invariance and a relatively large board. In this paper, we verify whether CNN-based move predictors prove effective for Othello, a game with significantly different characteristics, including a much smaller board size and complete lack of translational invariance. We compare several CNN architectures and board encodings, augment them with state-of-theart extensions, train on an extensive database of experts’ moves, and examine them with respect to move prediction accuracy and playing strength. The empirical evaluation confirms high capabilities of neural move predictors and suggests a strong correlation between prediction accuracy and playing strength. The best CNNs not only surpass all other 1-ply Othello players proposed to date but defeat (2-ply) Edax, the best open-source Othello player. Index Terms—deep learning, convolutional neural networks, Othello I. I NTRODUCTION Neural networks, particularly the deep and convolutional ones (CNNs), excel at recognizing virtually all kinds of patterns [32]. They naturally generalize to arbitrary dimensionality, which makes them suitable for analyzing raster images, time series, video sequences, and 3D volumetric data in medical imaging [38]. These properties make CNNs powerful for solving a wide range of challenging tasks at unprecedented performance level, ranging from image classification [12], to object localization [10], object detection [27], image segmentation [23], and even visual reinforcement learning in 3D environments [19]. The structural analogy between 2D raster images and board states in games is natural, so no wonder that it has been exploited within the neural paradigm in the past (e.g., [42]) . In most cases however, neural nets served there as board evaluation functions employed for searching game trees. It is only recently that massive computing and efficient learning algorithms for deep CNNs enabled their use as direct move predictors capable of achieving the level of play in the most challenging games previously thought to be exclusive only to human players. This gave rise to a line of Go-playing programs, including the paramount achievement of AlphaGo, the superhuman-level Go-playing program [33]. The authors are with the Institute of Computing Science, Poznan University of Technology, Poznan, Poland. W. Jaśkowski did this work while staying at IDSIA Dalle Molle Institute for Artificial Intelligence Research, Manno TI, Switzerland; e-mails: {pliskowski, wjaskowski, kkrawiec}@cs.put.poznan.pl. CNNs’ success in Go was possible because it has been known for a long time that Go players, rather than performing a “mental” tree search of future game states, rely heavily on detecting patterns in board states. This, together with a high branching factor of 250, and a large board renders tree search approaches ineffective for Go. Here we ask whether CNNs have a practical potential for games of a small branching factor and much smaller board size, for which the minimax-style tree search algorithms perform well and achieve human-level performance. To this aim, we consider learning CNN-based move predictors for the game of Othello, a long-standing benchmark for AI [18]. This problem not only diverges from Go, but is also very different from analysis of images in computer vision: the input rasters are tiny (8 × 8 ternary board states), free from noise and distortions typical for real-world settings (sensor noise, lighting, perspective, etc.), and the desirable invariances involve axial symmetries rather than translation and scaling. Last but not least, every single piece on the board matters and influences the decision. CNNs have been hardly ever evaluated in such settings. Our contributions include i) an experimental study of different CNN-based architectures of deep neural networks for Othello; ii) state-of-the-art move prediction accuracy on the French Othello league game dataset WThor; iii) state-of-theart 0-ply policy for Othello (no look ahead) that significantly improves over the previous approaches; iv) an in-depth analysis of the characteristics of trained policies, including the relationship between move prediction accuracy and winning odds, strength and weaknesses at particular game stages, and confrontation with opponents that employ game tree search at different depths. The gathered evidence suggests that several of the proposed neural architectures are best-to-date move predictors for Othello. II. R ELATED W ORK CNNs have been introduced by Fukushima in Neocognitron [9] and occasionally used for selected image analysis tasks for the next two decades [22]. The attempts of applying them to broader classes of images remained however largely unsuccessful. The more recent advent of deep learning brought substantial conceptual breakthroughs, enabling training of deeper and more complex networks. This, combined with cheap computing power offered by GPUs revolutionized machine perception and allowed achieving unprecedented classification accuracy, exemplified by the chain of ever-improving neural architectures assessed on the ImageNet database [21], [12]. 2 As signaled in the Introduction, the structural analogy between natural images and game boards is quite obvious for humans and as such could not remain unnoticed for long. This led to early attempts of applying CNNs to board games [8]. Soon it became clear that the deep learning paradigm may be powerful enough to be used together with game tree search techniques. This claim led to DeepMind’s AlphaGo, the first superhuman-level Go playing system [33], which combined supervised learning, policy gradient search and Monte Carlo Tree Search. Othello has for a long time been a popular benchmark for computational intelligence methods [29], [35], [5], [25], [26], [42], [40], [28], [16], [30], [18]. All strong Othello-playing programs use a variant of the minimax search [7] with a board evaluation function. Past research suggested that more can be gained by improving the latter than the former; that is why recently the focus was mostly on training 1 look-ahead (a.k.a. 1-ply) agents using either self-play [26], [18], fixed opponents [20], [16], or expert game databases [30]. Multiple ways of training the agents have been proposed: value-based temporal difference learning [25], [42], [34], (co)evolution [26], [31], [17], [15], and hybrids thereof [39], [40]. Designing a board evaluation function involves finding a good function approximator. One of the most successful function approximators for board games are n-tuple networks. Interestingly, they were originally proposed for optical character recognition [4], and their reuse for board games is yet another sign that the analogy between images and piece arrangements on a board is feasible. They were employed for the first time for Othello under the name of tabular value functions by Buro [7] in his famous Logistello program and later popularized by Lucas [24]. Neural networks have also been used for Othello as function approximators, e.g., in [42], but, to the best of our knowledge, CNNs have never been used for this purpose. As we are not aware of any published move predictors for Othello, in the experimental part of this paper we compare our method to a representative sample of 1-ply strategies. A comparison of the state-of-the-art 1-ply players for Othello can be found in [18]. Among the published works, the best 1-ply player to date was obtained by Coevolutionary CMAES (covariance matrix adaptation evolution strategies) [18] and systematic n-tuple networks. Among the multi-ply Othello strategies, one of the contemporary leaders is Edax1 , an open source player that uses tabular value functions, one for each stage of the game. Edax is highly optimized (bitboard move generator) and uses deep tree minimax-like search (negasout with multi-probcut tree search [6]). Unfortunately, a description of how its evaluation function was obtained is not publicly available. One of the other top-5 Othello players to date is RL14_iPrefN that has been trained using preference learning on the French Othello League expert games dataset [30], reportedly obtaining 53.0% classification accuracy. In this paper, we significantly improve over this baseline. IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES (a) Othello initial board state. Black to move. (b) Board state after black’s move. White to move. Figure 1: Othello boards with legal moves marked as shaded locations. III. OTHELLO Othello is a perfect information, zero-sum, two-player strategy game played on an 8 × 8 board. There are 64 identical pieces which are white on one side and black on the other. The game begins with each player having two pieces placed diagonally in the center of the board (Fig. 1a). The black player moves first, placing a piece on one of four shaded locations, which may lead to the board state in Fig. 1b. A move is legal if the newly placed piece is adjacent to an opponent’s piece and causes one or more of the opponent’s pieces to become enclosed from both sides on a horizontal, vertical or diagonal line. The enclosed pieces are then flipped. Players alternate placing pieces on the board. The game ends when neither player has a legal move, and the player with more pieces on the board wins. If both players have the same number of pieces, the game ends in a draw. Despite simple rules, the game of Othello is far from trivial. The number of legal positions is approximately 1028 and the game tree has in the order of 1058 nodes [1], which precludes any exhaustive search method. Othello is also characterized by a high temporal volatility: a high number of pieces can be flipped in a single move, dramatically changing the board state. These features and the fact that the game has not yet been solved makes Othello an interesting test-bed for computational intelligence methods. IV. L EARNING TO P REDICT E XPERT M OVES A. The Classification Problem Since Othello is a deterministic and fully-observable environment, a policy π : X → Y is here a mapping from the space of board states X to the space of actions Y = {1, . . . , K} that correspond one-to-one to the K = 60 board positions. π(x) represents the desired move for board state x ∈ X. Given a training set of N examples of good (e.g., experts’) moves in X × Y , we can formulate the problem of finding π as a multi-class classification problem. A typical approach to such problems is training a parameterized probability function p(y|x; θ), that represents the confidence 1 http://abulmo.perso.neuf.fr/edax/ PAWEŁ LISKOWSKI et al.: LEARNING TO PLAY OTHELLO WITH DEEP NEURAL NETWORKS 3 ... ... ... ... Othello board 2 ⇥ 8 ⇥ 8 input 3 ⇥ 8 ⇥ 8 feature maps 3 ⇥ 8 ⇥ 8 feature maps output layer Figure 2: Exemplary CNN for move prediction. The board is fed into the network as an 8 × 8 two-channel binary image, where ones indicate the locations of the player’s (first channel) and opponent’s pieces (second channel). Units in a feature map (colored squares) receive data from 3 × 3 RFs (in the same color) in the image (or in a previous layer). The network uses zero-padding on the edges (not shown), so maps in all layers are 8 × 8. For clarity, each hidden layer has only 3 feature maps. of making a move y in state x. To estimate p, one needs to minimize the cross-entropy loss: L(θ) = − K N X X i=1 k=1 (k) yi (k) ln p(yi |xi ; θ), (k) where y = 1 if y = k and 0 otherwise. The task of a training algorithm is to estimate the optimal θ and so find θ̂, a policy parameterization that brings L(θ) sufficiently close to its global optimum. Once found, the player chooses the moves of highest confidence: πθ̂ (x) = arg max p(y|x; θ̂), y∈Legal(x) where Legal(x) ⊂ Y is the subset of moves that are legal in board state x. If π determines the move to make without explicit simulation of the future game course (using, e.g., minimax-style algorithms), we refer to it as move predictor. One-to-one correspondence between board locations and moves in Othello greatly facilitates the design of move predictors. For other board games that involve heterogeneous pieces that move around the board, the space of moves Y is more complex and requires a more sophisticated representation. B. CNNs for Move Prediction In the simplest scenario, the board states in X could be encoded as 64 ternary variables and fed into any probability estimator. However, generic probability estimators are not biased towards the kind of structure that relates board state variables in Othello. To see that such a structure exists, it is enough to imagine the board variables being randomly permuted, which would completely obfuscate the game for human players (while being still the same policy learning task under the formulation in the previous section). This shows that capturing spatial patterns of the player’s and opponent’s pieces is essential, and makes it particularly appropriate to take advantage of learning systems that have such capabilities, i.e., treat a board state as an image. Convolutional neural networks (CNNs) are nowadays the leading machine learning tool for image analysis (cf. the work cited in the introduction). In the following, we briefly describe the CNN architecture as we tailored it to move prediction in Othello, referring readers interested in this topic to other comprehensive studies [2], [32]. For clarity, we present here only the core CNN architecture and augment it with extensions in the subsequent experimental section. A CNN is a composite of multiple elementary processing units, each equipped with weighted inputs and one output, performing convolution of input signals with weights and transforming the outcome with some form of nonlinearity. The units are arranged in rectangular feature maps (grids) and spatially aligned with the raster of the input image. The spatial arrangement of units is the primary characteristic that makes CNNs suitable for processing visual information. Feature maps are arranged into layers, where maps in a given layer fetch signals only from the previous layer (see also Fig. 2). Our CNNs take as input the representation of the current board state, encoded in two 8 × 8 binary matrices. The 1s indicate the locations of the player’s pieces in the first matrix and opponent’s pieces in the second matrix. We refer to this encoding as pieces. In the encoding called vmoves, we add a third matrix in which we mark with 1s the player’s legal moves, to facilitate picking the valid moves during game playing. Following recent research conducted in Go [33], we define also the ones encoding, which includes an additional regularizing channel in pieces encoding that contains only ones. A board state represented in the above way forms an 8 × 8 binary image comprising two (for pieces), or three (for vmoves and ones) channels, and is fetched by the feature maps in the first layer of the CNN. Each unit in a feature map receives data only from its receptive field (RF), a small rectangle of neighboring board locations, however from all channels, and uses a separate weight set for each channel (see exemplary RFs in Fig. 2). Analogously in the subsequent layers, a unit fetches data from the corresponding RFs in all feature maps in the previous layer, and treats them with separate weight sets. The RFs of neighboring units in a map are offset by a stride. Board size, RF size and stride together determine the dimensions of a feature map. For instance, a feature map with 4 IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES Table I: CNN architectures used in the experiments. Layer names are followed by numbers of feature maps. Conv4 conv64 → conv64 → conv128 → conv128 → f c128 → f c60 Conv6 conv64 → conv64 → conv128 → conv128 → conv256 → conv256 → f c128 → f c60 Conv8 conv64 → conv64 → conv128 → conv128 → conv256 → conv256 → conv256 → conv256 → f c128 → f c60 3 × 3 RFs with stride 1 needs only 36 units when applied to an 8 × 8 board, because six RFs of width 3 each, overlapping by two rows, span the entire board width (and analogously for height). The local connectivity substantially reduces the number of weights and facilitates capturing of spatial patterns. Units within the same feature map share their weights, and so calculate the same local feature, albeit from a different part of the board. This reduces the number of parameters even further and makes the extracted features equivariant. For instance, the 6 × 6 feature map mentioned above has only 9 · b + 1 parameters for a b-channel input (3 × 3 = 9 plus one threshold per feature map). We use rectified linear units (ReLU) to process convolution outcomes, with the nonlinearity defined as f (x) = max(x, 0). ReLUs make networks learn faster than squeezing-function units (e.g., sigmoid or tanh) and are particularly effective in networks with many layers, effectively alleviating the vanishing gradient problem, because gradient backpropagates through them undistorted for all positive excitations, in contrast to a small interval of excitations for the squeezing units. Our basic architectures are the networks composed of 4, 6, and 8 convolutional layers shown in Table I. In each architecture, the input is passed through a stack of convolutional layers composed of 64, 128, or 256 feature maps, with 3×3 RFs and the stride set to 1. Since the input is very small, we pad it with zeros on the grid border (so that the RFs can reach also beyond the input grid) in order to preserve the spatial resolution after convolution. For the same reason, we do not use any form of pooling, i.e., aggregating the outputs of multiple units by other means than convolution, as there is little rationale for that for small input grids [36]. Consequently, all feature maps in all layers have 8 × 8 units. In all architectures, the stack of convolutional layers is followed by two fully-connected (FC) layers. The hidden layer consists of 128 units with ReLU activations, while the output layer has K = 60 units corresponding to 60 possible move candidates. Each output y (k) corresponds to a single board location, and the desired moves are encoded using the binary 1-of-K scheme (a.k.a. one-hot encoding). For the output values to be interpreted as probabilities they must sum to unity, which is achieved using the softmax transformation: ey ok = P j (k) ey(j) , The classifier selects the move with the highest probability, i.e., ô = arg maxk∈K ok . The architectures presented in Table I resulted from a series of preliminary experiments, where we found out that increasing network depth beyond eight convolutional layers does not improve the performance while significantly lengthening the training time. Also, small 3 × 3 RFs and deeper networks turn out to be more efficient in terms of trading-off runtime and accuracy than using fewer but larger RFs. V. E XPERIMENTS IN M OVE P REDICTION ACCURACY A. Experimental Setup In this section, we consider networks as move predictors and do not engage them in actual games. The key performance indicator is prediction accuracy, i.e. the percentage of correctly predicted moves. We compare various configurations of the neural move predictors described in Section IV trained on examples of moves extracted from games played by humans (Section V-A2) . 1) Datasets: Created in 1985, the WThor database2 contains the records of 119,339 games played by professional Othello players in various French tournaments. From those records, we extracted all board states accompanied with the color of the player to move and the move chosen by the player in that particular state. The resulting set of 6,874,503 (board, color, move) triples is referred to as Original. We then removed the duplicates, obtaining 4,880,413 unique (board, color, move) triples, referred to as Unique dataset in the following. Note that both datasets may contain inconsistent examples, i.e. the same board state associated with different moves. A perfect classifier can achieve accuracy of 91.16% on Original and 97.49% on Unique. The color attribute is used only to invert the pieces on the boards when the white player is to move, so that all states in the dataset are seen from the black player’s perspective. Deep neural networks are known to perform particularly well when trained on large volumes of data. In order to augment the datasets, we take advantage of Othello’s invariance to board symmetries. For each board state in the Original and Unique datasets, we produce its all seven reflections and rotations. This results respectively in two symmetric datasets: Original-S (54,996,024 examples) and Unique-S (39,019,056 examples), where the latter was cleared also of the duplicates resulting from symmetric transformations. To assess the generalization capabilities of move predictors, we partition each dataset into disjoint training and testing sets. For the asymmetric datasets (Original and Unique), we allocate 25% of examples for testing. The symmetric datasets are much larger, so we allocate only 5% of examples for testing, so that all test sets are similar in size. 2) Training: For training, we use stochastic gradient descent with 0.95 momentum. Units’ weights are initialized using He’s method [11] while the biases (offsets) are initialized with zeroes. The learning rate is initially set to 0.1, and then halved twice per epoch. The loss function is regularized by the L2 norm with the weight of 5 · 10−4 . Learning is stopped after 2 http://www.ffothello.org/informatique/la-base-wthor/ PAWEŁ LISKOWSKI et al.: LEARNING TO PLAY OTHELLO WITH DEEP NEURAL NETWORKS Table II: Prediction accuracy of networks trained on four different training sets and evaluated on the Original test set. Pieces encoding was used. Architecture Conv4 Conv6 Conv8 Original 55.1 57.2 58.1 Original-S 56.9 58.7 60.1 Unique 55.9 57.9 58.9 Architecture Conv4 Conv6 Conv8 pieces ones vmoves 57.2 59.2 60.5 57.3 59.1 60.6 57.3 59.0 59.0 Table IV: Probability [%] of choosing a valid move on UniqueS. Unique-S 57.2 59.1 60.5 Table III: Prediction accuracy on Unique-S test set for different board encodings and number of convolutional layers. 24 epochs of training for the asymmetric datasets (171, 600 batches of examples) and 6 epochs (434, 400 batches) for the symmetric datasets. The implementation is based on the Theano framework [41], which performs all computation on a GPU in singleprecision arithmetic. The experiments were conducted on an Intel Core i7-4770K CPU with NVIDIA GTX Titan GPU. Training time varied from single hours to dozens of hours depending on configuration, e.g. nearly 62 hours for Conv8 trained on Unique-S. B. Results 1) Effect of Data Augmentation: Table II presents the testset prediction accuracy of three basic network architectures trained on particular training sets with the pieces encoding. For fairness, the networks are compared on the same data, i.e. the Original test set. Training on the datasets augmented by symmetries systematically leads to a higher prediction accuracy. The impact of removing duplicates is also consistently positive, albeit not so strong. These two trends together render Unique-S most promising, so in the following experiments we use this dataset only. 2) Impact of Board Encoding and the Number of Layers: Table III presents the prediction accuracy of the three architectures when trained on Unique-S with various encodings of the board state. This time, the networks are tested on the UniqueS test set, which explains the minor differences w.r.t. Table II, where Original test set was used for testing. The accuracy clearly correlates with the number of layers, so it is tempting to reach for even deeper architectures. However, a 10-layer architecture analogous to Conv8 failed to further improve the prediction accuracy. Concerning the comparison of encodings, pieces and ones seem to perform on par, so the former should be preferred as it does not require an additional input layer/channel. The relative underperformance of vmoves is more puzzling as it is hard to see how giving a learner a hint about moves’ validity could deteriorate its performance. In an attempt to investigate this outcome, we evaluate the networks with respect to their ability to discern the valid and invalid moves. In Table IV, we present the test-set accuracy of particular networks, where 100% indicates that the network’s most excited output 5 Architecture pieces ones vmoves Conv4 Conv6 Conv8 99.95 99.99 99.99 99.95 99.99 99.99 100.0 100.0 100.0 Table V: The influence of dropout, batch normalization, and bagging on prediction accuracy [%]. Pieces encoding was used. Architecture Conv8 (baseline) Conv8+dropout Conv8+BN Conv8+BN+bagging Prediction accuracy Prediction time [×10−4 s] 60.5 58.8 62.7 64.0 1.6 1.6 1.9 19.4 is always among the valid moves. All considered networks and encodings bring this indicator very close to perfection, with the networks using vmoves making no mistakes at all in that respect. This suggests that telling apart the valid moves from the invalid ones is very easy regardless of configuration, and providing this information explicitly in vmoves is not only unnecessary but may distract the training process. Therefore, we use the pieces encoding in all experiments that follow. 3) Impact of Regularization and Bagging: In Table V, we report the impact of a few techniques that tend to improve generalization performance of CNNs. Dropout consists in disabling a number of (here: 50 percent) randomly selected units for the individual batches of training examples, which forces a network to form multiple alternative pathways from inputs to outputs. This usually makes networks more robust and reduces overfitting, particularly in domains where input data are inherently redundant, like natural images, typically composed of at least thousands of pixels. The states of Othello board are however very different: there are only 64 inputs, and every single piece counts when choosing the move to make. There is virtually no redundancy in the input. As Table V suggests, dropout prevents the Conv8 network from capturing the board state in full and so hampers its predictive accuracy. The purpose of batch normalization (BN) is to address the internal covariate shift [14] by standardizing the outputs of individual network units over batches of examples, which was demonstrated to be beneficial in many contexts. This is corroborated in our case with over two percent gain in accuracy, so we consider this configuration as a baseline throughout the rest of this paper. Finally, we attempt to reduce the classifier’s variance by employing bootstrap aggregation (bagging). We train 10 networks on bootstrapped samples drawn from the training set and average their corresponding outputs to determine the move to make. When applied to the Conv8+BN network (Table V), bagging boosts the performance to 64%. The price paid for that is a 10-fold increase in not only the training time3 , but 3 The mean prediction times were computed using batch size of 256. Making decision for a single board state takes on average 5-20 times more GPU time. 6 IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES Table VI: Prediction accuracy of residual networks. Conv8+BN (baseline) ResNet-32 ResNet-32-np ResNet-32-np-p 62.7 57.4 59.2 59.5 1.9 1.0 1.0 1.0 ResNet-56 ResNet-56-np ResNet-56-np–p 58.3 60.1 60.5 1.8 1.8 1.8 DenseNet-40 DenseNet-100 59.9 61.9 4.8 37.2 WideNet-40-4 62.2 9.4 Top-1 Accuracy # valid moves Prediction time [×10−4 s] 1 3 5 7 9 11 13 15 17 19 Accuracy Architecture 0.8 0.6 0.4 0.2 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 0.0 Move # valid moves 1 3 5 7 9 11 13 15 17 19 Top-2 Accuracy 1.0 0.8 0.6 0.4 0.2 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 0.0 Move # valid moves Top-3 Accuracy 1 3 5 7 9 11 13 15 17 19 also the prediction time, which will become relevant in the later parts of the study, when confronting our strategies with other opponents. 4) Architectures with Skip-Connections: The more recent DL research brought very deep network architectures containing dozens of layers [37], [12], which became the state-ofthe-art in image classification. Training such networks is only possible due to skip connections that form shortcuts in the pathways between layers, typically adding the representation (feature map) learned in a given layer to the representation learned in one of the consecutive layers. This, particularly when repeated multiple times across the network, opens new ways to achieving the training goal: rather than learning a mapping H such that y(x) = H(x), a network learns a residual F such that y(x) = F (x) + x. Skip connections alleviates the vanishing gradient problem as the gradient can backpropagate unchanged through many layers. Table VI presents the predictive accuracy of several skipconnection architectures. A ResNet-n network [12] consists of n 2 residual blocks, each comprising a nonlinear layer (ReLU activation) followed by a linear layer, shortcut with a skip connection that adds the input of the first layer to the output of the second one. Originally designed for image analysis, ResNets perform downsampling that reduces the input dimensionality; as there is arguably little need for that for our very small 8 × 8 input, we consider a variant stripped off that feature, ResNet-n-np. As expected, maintaining the dimensionality of the input throughout learning improves the accuracy of the smaller (ResNet-32) and larger (ResNet-56) architecture. Simple adding of layer’s outputs is possible only when they have the same dimensions. Otherwise, the authors of ResNets propose to implement the skip connections as linear mappings. The ResNet-n-np-p configurations in Table VI implement that feature, which slightly improves the predictions. ResNets’ skip connections can be considered first-order as each of them spans the inputs and outputs of the same residual block. In DenseNets [13], skip connections are being introduced between each pair of layers, so in a network of l of them. We trained layers (or residual blocks) there are l(l−1) 2 two DenseNet architectures. The smaller one (DenseNet-40) managed to perform on par with the other networks considered in this section, while the larger one (DenseNet-100) significantly outperformed them, almost reaching the performance level of the baseline. 1.0 1.0 0.8 0.6 0.4 0.2 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48 50 52 54 56 58 60 Move Figure 3: Top-k prediction accuracy factored w.r.t. the number of move in a game and the number of valid moves. Finally, a very recent work on architectures dubbed WideNet [43] revealed that the number of layers (and thus parameters) in residual networks can be significantly reduced with little or no harm for accuracy by making the residual blocks more expressive. This can be achieved by increasing the number of convolution layers within a block, increasing the number of feature maps, or both. Our instance of this architecture comprises 3 stacks of residual blocks, each stack composed of 6 blocks. Each residual block contains two 3×3 convolutional layers. The layers in consecutive stacks are equipped with respectively 64, 128, and 256 feature maps (cf. Table 1 in [43] for k = 4), so the network is four times wider than the original ResNet. As the last row of Table VI demonstrates, this WideNet configuration fares the best among the skipconnections architectures considered here, which suggests that aggregating multiple features at various stages of processing is essential for successful move prediction. However, WideNet40-4 is also one of the slowest networks when queried. Overall however, contrary to the expectations, neither of the very-deep skip-connection networks did improve over the Conv8+BN baseline. 5) Prediction Accuracy Analysis: The move prediction task is inherently multimodal: to play well, correct predictions have to be made along the entire course of the game, for 0.0 PAWEŁ LISKOWSKI et al.: LEARNING TO PLAY OTHELLO WITH DEEP NEURAL NETWORKS the board states typical for the initial game stages as well as those occurring in the endgame. It is thus interesting to ask how difficult it is for a network to perform well at particular stages of the game. Here, we answer this question in terms of prediction accuracy. In Fig. 3, we present the top-k prediction accuracy for our baseline network Conv8+BN, factored with respect to the number of a move in a game and the number of legal moves in a given state. Given four pieces in the initial board state, the former factor may vary from 5 to 60; however we start with 6 as the fifth move is always the same under the Othello symmetries (which we take into account here). The color of a datapoint reflects the probability that the correct move (as per the WThor test set) is among the top k indicated by the network. The figure reveals that the prediction accuracy remains largely consistent throughout the game. Only in the game’s beginning the network’s predictions are less accurate, which we attribute to a relatively small amount of training data for Unique-S at this game stage (due to many identical states). The other aspect at play, suggested by the differences between the top-1 and top-3 graphs, is the large fraction of inconsistent examples for the early game states, when there are many alternative paths to success and players tend to follow them accordingly to their personal habits and preferences. The high quality of predictions in the later game stages (also when the number of available moves is high) is particularly encouraging, as even a single move towards the end of a game can drastically change the state of the board – the feature that Othello owes its name to. VI. P LAYING S TRENGTH A NALYSIS Prediction accuracy considered in the previous section is only a surrogate measure of performance, since we are ultimately interested in the playing strength. In this section, we let our predictors play against other players. A. Experimental Setup We consider a suite of thirteen 1-ply players proposed in previous research, gathered and compared in our former study [18]. The suite consists of players with board evaluation functions encoded by weighted piece counter and n-tuple networks, trained by different methods including hand-design, temporal difference learning, evolution, and coevolution (cf. Section II). In addition to that suite of opponents, we consider also Edax ver. 4.3.24 , a strong, open source, highly optimized Othello program. Edax can be configured to play at a given ply (search depth) n, which we denote as Edax-n. Since our move predictors as well as all 14 opponents (including Edax) are deterministic, we resort to the following technique in order to obtain robust estimates of the odds of winning. Rather than playing a single game that starts from the initial board, we use a set of 1000 6-ply opening positions prepared by Runarsson and Lucas [30]. Starting from each such state, two games are played, with the players switching roles. For a single game, a player can obtain 1, 0.5 or 0 points 4 http://abulmo.perso.neuf.fr/edax/4.3/edax.4-3-2.zip 7 Table VII: The winning rates of the best performing move predictors against the 1-ply opponent players. Note that while CoCMAES is included in the 13 opponents suite, Edax-1 is not. Architecture ResNet-56-np-p DenseNet-sym-100 WideNet Conv8+BN Conv8+BN+bagging Prediction accuracy 60.5 61.9 62.2 62.7 64.0 Winning rate Average CoCMAES Edax-1 (13 opponents) -4+2x2 94.3 87.7 80.7 94.1 87.5 79.0 96.3 92.2 88.1 96.7 93.2 87.5 96.6 93.5 87.6 on the win, draw, or loss, respectively. We report the winning rate, i.e. the percentage of points possible to score in 2000 games. B. Playing Strength Evaluation We selected a representative sample of the five best predictors from Section V and confronted them with all 13 opponents from our suite as well as with Edax-1. Table VII presents their average winning rates against the 13 players, and separately the winning rates against CoCMAES-4+2x2 that proved the best in [18] and against Edax-1. All networks turn out to be significantly stronger than any of the existing 1-ply look-ahead players, achieving around 90% winning rate against them. We find this interesting, given that move predictors are essentially 0-ply strategies, as they do not perform any look-ahead. Conv8+BN achieves the highest average winning rate, and is only slightly worse than Conv8+BN+bagging when playing against CoCMAES-4+2x2 and than WideNet when playing against Edax-1. This overall performance corroborates its high prediction accuracy. The roughly monotonous relationship between the prediction accuracy and the winning rate observed in Table VII lets us ask whether such a tendency holds in general. In Fig. 4, we plot the latter against the former for all networks evaluated in this paper, when confronted with the models in the suite and with Edax-1. The graph reveals a very strong linear dependency (determination coefficient r2 = 0.997 and r2 = 0.999, respectively). Fitting linear models suggests that in the considered range of prediction accuracy (57-64 percent), each percent point of the prediction accuracy brings 1.7 percent points of the winning rate against the suite of 13 players, and 3.4 percent points against Edax-1. Obviously, this trend cannot hold globally, as illustrated by Conv8+BN+bagging (the rightmost two data points). Although this might be an outlier, it may also suggest that 63 percent is approximately the point at which further increases of prediction accuracy do not translate into better winning rates anymore. C. Effect of Data Augmentation on Playing Strength In Section V-B1, we found out that the training set stripped of duplicates and augmented with symmetries (Unique-S) gives rise to the best move prediction accuracy. However, prediction accuracy is only a proxy of what we really care about – the winning rate. It is thus not obvious whether an analogous claim holds for playing strength, so, in this 8 IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES Figure 4: Networks’ move prediction accuracy vs. winning rate against all 13 opponent players (red dots) and against Edax (blue triangles). Table VIII: The winning rates against the 13 players from the suite for the basic neural predictors trained on different datasets Architecture Conv4 Conv6 Conv8 Original 77.9 84.6 85.4 Original-S 85.2 89.0 91.8 Unique 80.5 86.5 88.6 Unique-S 86.0 90.0 93.1 section we let the basic convolutional predictors from Section V-B play against the opponents from the suite. Other settings remain as in the previous section. The winning rates presented in Table VIII corroborate the outcomes in terms of prediction accuracy presented earlier in Table II. Both duplicate removal and augmentations have significantly positive effects on the playing strength. This correlation shows again that, at least in certain intervals, move prediction accuracy is a reliable predictor of the playing strength (cf. Fig. 4). factor of Othello is 10, so for search depth 5 Edax’s decision are based of tens of thousands of board states each, while our neural predictor performs quite well with just one. Search depth impacts the mean time required by Edax to make a move, which we report in the last column of the table. As Edax is highly optimized, it is able to make moves at a rate comparable to the prediction time of the Conv8 CNN (1.9×10−4 s) for search depths up to 3. Deeper search takes Edax longer. On the other hand, however, the times reported for neural predictors have been assessed in batches of 256 board states, the common efficiency means in GPU computing. When asked to predict a move for a single board state, CNNs are 5-20 times slower. Table IX includes also Edax prediction accuracy on the Original dataset, which clearly grows with the depth of search. However, increasing the depth from 4 to 5 does not improve the accuracy, despite the associated increase in headto-head matches against Conv8+BN. Also, Edax’s prediction accuracy (59% at n = 5) is significantly lower than the 64% demonstrated by Conv8+BN. It seems thus that Edax plays in a different way than human players, whom CNNs try to mimic. E. Analysis of Game Stages As much as we find the performance of our 0-ply move predictors impressive, the previous experiments show that deeper lookaheads are essential for achieving high winning rates. In our last experiment, we seek to find out which game stage may benefit the most when supplementing our move predictor with game tree search (and conversely, which game stage is most problematic for neural predictors). To this aim, we split the game into four stages according to intervals of move numbers: [1, 15], [16, 30], [31, 45], and [46, 60]. For each stage, we design a hybrid strategy that employs Edax-5 in that D. Playing against Game Tree Search stage, while using the Conv8+BN policy to make decisions Section VI-B revealed that the 0-ply neural move predictors in the remaining stages. We let those hybrid strategies play outperform the 1-ply opponents, and do so by a large margin against Edax-1..5. The results summarized in Fig. 5 confirm that letting Edax (∼ 90 percent vs. the 50 percent tie point). It is interesting to see how that margin decreases when the opponents are make moves at any game stage improves the performance of allowed for deeper game-tree searches. To investigate this, we the hybrid. The gains are largest for n = 3; as demonstrated confront Conv8+BN with Edax-n for search depth n ∈ [2, 5] in Section VI-D, for smaller n Edax is relatively easy to beat (using the same protocol involving 2000 games), and gather already by pure Conv8+BN, so extending it with Edax-5 at the results in Table IX. Remarkably, our predictor maintains a single game stage does not improve the odds for winning superiority when Edax is allowed for 2-ply lookahead, though much. Conversely, for n > 3 the opponent Edax-n becomes the odds of its winning are now only roughly 6:4. Starting so strong that it is hard to increase the odds of winning either. The graphs reveal also that there is more to gain at later from search depth 3, Conv8+BN is more likely to lose than to win, and for 5-ply search its odds for winning drop to 1:9. This stages of the game regardless of the opponent’s depth. We is however still impressive, given that the average branching suspect that this is due to the combinatorial nature of the game, which makes the CNN’s pattern recognition capabilities less relevant for success in the last game stage. Also, given that Table IX: The performance of Conv8+BN vs. Edax-n. Othello games last at most 60 moves, a search depth 5 in the n Winning rate Edax move time [×10−4 s] Edax prediction accuracy last game stage enables Edax to often reach the final game 1 87.5 1.0 47.9 states and so obtain robust assessments of potential moves. 2 59.4 1.1 53.3 Last but not least, certain fraction of moves derived from 3 34.7 1.6 55.2 4 19.8 5.0 59.8 WThor can be suboptimal and so might have ‘deceived’ the 5 10.3 10.0 59.7 networks. 3/31/2017 https://docs.google.com/spreadsheets/d/1hgrDWaqWOA7NCvDhqKZgDt4LYEpA3Hd0wex3Gbo_XC0/pubchart?oid=124067… PAWEŁ LISKOWSKI et al.: LEARNING TO PLAY OTHELLO WITH DEEP NEURAL NETWORKS 1-ply 2-ply 3-ply 4-ply Winning Rate Gain [%] 15 10 9.35 7.18 5 5-ply 14.25 11.77 10.9 9.92 11.05 10.47 7.6 6.35 4.7 4.25 1.95 12.57 12.43 11.85 8.5 6.6 9 arbitrary complex combinatorial patterns occurring in entire board states. A natural follow-up step is thus considering fullyconnected networks that do not suffer from that limitation; however, they may be also much more costly in training and tend to overfit. In relation to that, an interesting study regarding fully-connected networks for Othello with up to 3 hidden layers was performed in [3]. 4.18 2.53 VIII. C ONCLUSIONS Our study brings evidence that deep neural networks are a viable methodology for training Othello agents. They do 0 15 30 45 not require explicit knowledge on game rules or strategies, nor extensive game-tree searches, nor explicitly compare inStage [move number] dividual future board states or actions. Rather than that, they Figure 5: The gain of winning rate of Conv8+BN when achieve a high level of playing by learning the associations hybridized with Edax-5 for a particular stage of the game. between the patterns observed in board states and the desired Winning rates are assessed independently against Edax-1..5. moves. We may claim thus that, at least to some extent, the trained neural predictors embody the knowledge of the Othello domain – a statement that would be hard to defend for VII. D ISCUSSION game-tree search techniques, which owe their power largely The deep CNN-based move predictor with its nearly 63% to algorithmic efficiency. On the other hand, the supervised approach proposed here of accuracy turned to be much better than n-tuples-based preference learning [30], which achieved only 53% on the cannot be deemed as ‘knowledge-free’ in the same sense as same dataset. Our Conv-8-BN wins against it in 93.9% of for instance evolutionary or reinforcement learning methods. games. We obtained a similar winning rate of 93.2% also Here, the essential know-how is hidden in the expert games against the CoCMAES player [18], the strongest to-date 1- used for training. CNNs power lies in the ability to extract and incorporate this knowledge into a system that is able to ply Othello player (cf. Table VII). Despite Othello’s relatively low branching factor and small generalize beyond the training examples, something no other board, CNNs proved to be supreme move predictors, managing method in the past has been shown to do well enough to to efficiently capture the expertise materialized in millions achieve the expert level of play in Othello. Given the supreme performance of CNNs, this work may of humans’ moves. We find that impressive given that small, discrete-valued board states are arguably very different from be worth extending in several directions. The networks that rely more on the fully-connected layers, mentioned in Section natural images5 . Seeing a convolutional network to perform so well in a VII, are one such possibility. Logistello [7] and Edax use board game that is characterized with little translation invari- different evaluation functions for each stage of the game (cf. ance is intriguing, as the main motivation for using convolution SectionVI-E), and such ‘factorization by game course’ could is precisely the translation invariance. This suggests that also lead to a better CNN agent. Another, arguably simple certain patterns observed in Othello boards can be to some extension is to add the information about the player to move, extent translation-invariant, or at least can be detected using which we expect to be beneficial given that Othello is a non‘higher-order convolutions’ implemented by a stack of several symmetric game. Last but not least, given the usefulness of convolutional features evidenced here, it would be interesting https://docs.google.com/spreadsheets/d/1hgrDWaqWOA7NCvDhqKZgDt4LYEpA3Hd0wex3Gbo_XC0/pubchart?oid=1240676136&… 1/1 convolutional layers intertwined with nonlinearities. Concerning limitations of the approach, we could not im- to verify their capabilities in other contexts than supervised prove upon the result obtained with the Conv-8-BN network learning of move predictors, e.g. as state or state-action evalby using deeper networks with skip connections. This suggests uation functions trained with reinforcement learning methods. that what works for high-dimensional image data does not alACKNOWLEDGMENTS ways succeeds for low-dimensional problems such as Othello. P. Liskowski acknowledges the support from grant An interesting observation concerning training is that the large networks hardly overfit to the data. The classification 2014/15/N/ST6/04572 funded by the National Science Cenaccuracy on the training set is typically within one percent tre, Poland. W. Jaśkowski acknowledges the support from point of the accuracy on the test set, suggesting that the Ministry of Science and Higher Education grant “Mobilnetworks we considered here cannot memorize the training ity Plus” no 1296/MOB/IV/2015/0. K. Krawiec was supdata entirely. We may conclude thus that there is still room ported by the National Science Centre, Poland, grant for improving their architectures. In particular, this may be no. 2014/15/B/ST6/05205. a sign of an inherent limitation of convolutional processing, R EFERENCES where the limited in size RFs may be insufficient to capture 0 5 Note that the images considered in computer vision hardly ever contain fewer than 100 pixels. [1] Victor L Allis. Searching for solutions in games and artificial intelligence. PhD thesis, University of Limburg, Maastricht, The Netherlands, 1994. 10 [2] Yoshua Bengio. Learning deep architectures for AI. Foundations and trends in Machine Learning, 2(1):1–127, 2009. [3] Kevin J Binkley, Ken Seehart, and Masafumi Hagiwara. A study of artificial neural network architectures for Othello evaluation functions. Information and Media Technologies, 2(4):1129–1139, 2007. [4] Woodrow Wilson Bledsoe and Iben Browning. Pattern recognition and reading by machine. In Proc. Eastern Joint Comput. Conf., pages 225– 232, 1959. [5] M Buro. An evaluation function for Othello based on statistics. Technical report, NEC, Princeton, NJ, 1997. [6] Michael Buro. Experiments with multi-probcut and a new high-quality evaluation function for othello. Games in AI Research, pages 77–96, 1997. [7] Michael Buro. Experiments with Multi-ProbCut and a new high-quality evaluation function for Othello. In H.J. van den Herik and H. Iida, editors, Games in AI Research, pages 77–96. University Maastricht, 2000. [8] Christopher Clark and Amos Storkey. Teaching deep convolutional neural networks to play go. arXiv preprint arXiv:1412.3409, 2014. [9] Kunihiko Fukushima and Sei Miyake. Neocognitron: A self-organizing neural network model for a mechanism of visual pattern recognition. In Competition and cooperation in neural nets, pages 267–285. Springer, 1982. [10] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580–587, 2014. [11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on Imagenet classification. In Proceedings of the IEEE international conference on computer vision, pages 1026–1034, 2015. [12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770– 778, 2016. [13] Gao Huang, Zhuang Liu, Kilian Q Weinberger, and Laurens van der Maaten. Densely connected convolutional networks. arXiv preprint arXiv:1608.06993, 2016. [14] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. [15] Wojciech Jaśkowski. Systematic n-tuple networks for othello position evaluation. ICGA Journal, 37(2):85–96, June 2014. [16] Wojciech Jaśkowski. Systematic n-tuple networks for position evaluation: Exceeding 90% in the othello league. Technical Report RA06/2014, arXiv:1406.1509, Institute of Computing Science, Poznan University of Technology, Poznań, Poland, 2014. [17] Wojciech Jaśkowski, Paweł Liskowski, Marcin Szubert, and Krzysztof Krawiec. Improving coevolution by random sampling. In Christian Blum, editor, GECCO’13: Proceedings of the 15th annual conference on Genetic and Evolutionary Computation, pages 1141–1148, Amsterdam, The Netherlands, July 2013. ACM. [18] Wojciech Jaśkowski and Marcin Szubert. Coevolutionary CMA-ES for knowledge-free learning of game position evaluation. IEEE Transactions on Computational Intelligence and AI in Games, 8(4):389–401, 2016. [19] Michał Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaśkowski. ViZDoom: A Doom-based AI Research Platform for Visual Reinforcement Learning. In IEEE Conference on Computational Intelligence and Games, pages 341–348, Santorini, Greece, Sep 2016. IEEE. [20] Krzysztof Krawiec and Marcin Grzegorz Szubert. Learning n-tuple networks for Othello by coevolutionary gradient search. In Proceedings of the 13th annual conference on Genetic and evolutionary computation - GECCO ’11, pages 355–362, New York, New York, USA, jul 2011. ACM Press. [21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105, 2012. [22] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. [23] Paweł Liskowski and Krzysztof Krawiec. Segmenting retinal blood vessels with deep neural networks. IEEE transactions on medical imaging, 35(11):2369–2380, 2016. [24] Simon M. Lucas. Learning to play Othello with N-tuple systems. Australian Journal of Intelligent Information Processing Systems, Special Issue on Game Technology, 9(4):01–20, 2007. IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES [25] Simon M. Lucas and Thomas P. Runarsson. Temporal difference learning versus co-evolution for acquiring Othello position evaluation. In IEEE Symposium on Computational Intelligence and Games, pages 52–59. IEEE, 2006. [26] Edward P Manning. Using Resource-Limited Nash Memory to Improve an Othello Evaluation Function. IEEE Transactions on Computational Intelligence and AI in Games, 2(1):40–53, 2010. [27] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 779–788, 2016. [28] Michiel Van Der Ree and Marco Wiering. Reinforcement learning in the game of Othello: Learning against a fixed opponent and learning from self-play. In Adaptive Dynamic Programming And Reinforcement Learning (ADPRL), 2013 IEEE Symposium on, pages 108—-115, 2013. [29] P.S. Rosenbloom. A world-championship-level othello program. Artificial Intelligence, 19(3):279–320, 1982. [30] Thomas Runarsson and Simon Lucas. Preference Learning for Move Prediction and Evaluation Function Approximation in Othello. Computational Intelligence and AI in Games, IEEE Transactions on, 6(3):300– 313, 2014. [31] Spyridon Samothrakis, S Lucas, TP Runarsson, and David Robles. Coevolving Game-Playing Agents: Measuring Performance and Intransitivities. IEEE Transactions on Evolutionary Computation, 17(2):213– 226, 2013. [32] Jürgen Schmidhuber. Deep learning in neural networks: An overview. Neural networks, 61:85–117, 2015. [33] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, 2016. [34] Ioannis E. Skoulakis and Michail G. Lagoudakis. Efficient Reinforcement Learning in Adversarial Games. In 2012 IEEE 24th International Conference on Tools with Artificial Intelligence, pages 704–711. IEEE, nov 2012. [35] R. Smith and B. Gray. Co-adaptive genetic algorithms: An example in othello strategy. In 7th Florida AI Research Symposium, 1994. [36] Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806, 2014. [37] Rupesh Kumar Srivastava, Klaus Greff, and Jürgen Schmidhuber. Highway networks. arXiv preprint arXiv:1505.00387, 2015. [38] Maciej Szkulmowski, Daniel Ruminski, Paweł Liskowski, Bartosz Wieloch, Krzysztof Krawiec, Bartosz Sikorski, and Maciej D Wojtkowski. Oct retinal angiography using neural networks. Investigative Ophthalmology & Visual Science, 57(12):454–454, 2016. [39] Marcin Szubert, Wojciech Jaśkowski, and Krzysztof Krawiec. Learning Board Evaluation Function for Othello by Hybridizing Coevolution with Temporal Difference Learning. Control and Cybernetics, 40(3):805–831, 2011. [40] Marcin Szubert, Wojciech Jaśkowski, and Krzysztof Krawiec. On scalability, generalization, and hybridization of coevolutionary learning: a case study for othello. IEEE Transactions on Computational Intelligence and AI in Games, 5(3):214–226, 2013. [41] Theano Development Team. Theano: A Python framework for fast computation of mathematical expressions. arXiv e-prints, abs/1605.02688, May 2016. [42] Sjoerd van den Dries and Marco A. Wiering. Neural-Fitted TD-Leaf Learning for Playing Othello With Structured Neural Networks. IEEE Transactions on Neural Networks and Learning Systems, 23(11):1701– 1713, November 2012. [43] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016.
1cs.CV
For a study of definable representations of algebraic groups Adrien Deloro arXiv:1510.06302v2 [math.LO] 4 Mar 2016 7th March 2016 Abstract We classify irreducible SL2 (K)-modules of low Morley rank (≤ 4 rk K) as a first step towards a more general conjecture. 1 Introduction It was more than fifty years since Robert Steinberg published his celebrated “tensor product” theorem. Fact (Steinberg’s Tensor Product Theorem, [15]). Let G be a semisimple algebraic group of characteristic p 6= 0 and rank ℓ, and let R denote the set of pℓ irreducible rational projective representations of G in which the high weight λ satisfies 0 ≤ λ(a) ≤ (p − 1) for all simple roots a. Let αi i denote the automorphism t 7→ tp of the universal field as well as the corresponding automorphism αi of G, and for R ∈ R let R denote the composition of αiN and R. Then every irreducible rational projective representation of G can be written uniquely as Riαi . Conversely, every such product yields an irreducible rational projective representation of G. As conjectured by Steinberg, the rationality assumption was later removed by Borel and Tits following their work on “abstract homomorphisms” between groups of K-points [2]. This of course requires consideration of all field automorphisms and not only iterates of the Frobenius map in the above theorem, but the set R remains unchanged [2, Corollaire 10.4]. We hope the reader to be familiar with the significance of Steinberg’s tensor theorem. For we wish in the present paper to conjecture an analogue in a different setting; such an analogue would be the first stone of a bridge linking representation theory to model theory. As a matter of fact, we suggest to study the category of definable representations of the semisimple algebraic groups — and since our expected reader is not an expert in mathematical logic, (s)he will find all necessary explanations after the statements below. Conjecture. Let (G, V ) be a module of finite Morley rank, where G is isomorphic to the group GK of K-points of some semisimple algebraic group G. Suppose that the action is faithful and irreducible. Then there are a definable, compatible K-vector space structure on V and definable field automorphisms ϕi such that V ≃ ⊗ϕi Ri , definably, as a K[G]-module. We prove a basic case of this conjecture (the precise statement is in §1.5). Theorem. Work in the setting of the conjecture. Suppose that K does not have characteristic 2 or 3, that G ≃ SL2 (K) and that rk V ≤ 4 rk K, where rk denotes the Morley rank. Then the desired conclusion holds. MSC 2010: 20F11. Keywords: Groups of finite Morley rank; Modules of finite Morley rank; Definable representations 1 Partial results had already been obtained: identification of the natural action when rk V ≤ 2 rk K in [7] (with an application to identification of the standard action of GLn (K) [1]), and identification of the adjoint action when rk V ≤ 3 rk K in [5]. One easily imagines that 3 rk K = rk SL2 (K) is a critical value past which things become much more difficult. Not only does our proof begin to overcome such difficulties, it also provides much quicker arguments for [7, 5] (see §1.6). And now let us explain what we mean. 1.1 Model Theory: Ranked Universes Model theory is concerned with mathematical structures in both a wide (general) and narrow (the restriction to “first-order”) sense. But we do not need a general introduction as we shall focus on groups. An excellent reference describing our setting is [4]. For the model theorist, algebraic groups as functors hardly exist: only affine groups of K-points can be examined, one by one. Such objects are equipped with a collection of finitary relations — among which the equality and the group law of course, but there could be more. This we call a “group structure”. Now on a general group structure G model theory studies the class of so-called definable sets [4, §4.1.1]: singletons, Cartesian powers of G, and the graphs of relations are definable, the class is closed under boolean combinations, under projections, and under quotients by definable equivalence relations. This reflects the logician’s will to work only with “first-order” formulas (an approach we shall not need here). This also loosely mimicks the behaviour of the constructible class in affine algebraic groups. Of course the geometry of the constructible class in an affine algebraic group over an algebraically closed field is well-understood, whereas for a general, abstract group structure, nothing similar should be expected. One ingredient is obviously missing to bring our setting closer to affine algebraic group theory: the Zariski topology. But this is not in the scope of model theory, and we shall rely on no topological methods. The loss is serious as geometers imagine. Instead we merely use a coarse analogue of the Zariski dimension, known as the Morley rank. The Morley rank, in general an ordinal-valued function, is defined by a standard induction: • a non-empty, finite definable set has rank 0; • a definable set A has rank ≥ α + 1 iff there are infinitely many disjoint definable subsets Bi ⊆ A with rk Bi ≥ α; • to deal with limit ordinals take suprema. Now it so happens (at least in the case of a group structure; other mathematical objects are not all so well-behaved model-theoretically) that if G is a group structure the Morley rank of which is finite, then it satisfies the following properties [4, §4.1.2], where f : A ։ B is any definable surjection between definable sets: • the rank is definable: for any k, the set {b ∈ B : rk f −1 (b) = k} is definable; • the rank is additive: if the latter set equals A then rk A = rk B + k; • uniformity phenomena (called “elimination of infinite quantifiers” by logicians): there is an integer m such that for any b ∈ B, the fiber f −1 (b) either is infinite or of cardinality ≤ m. All standard dimension computations [4, §4.2] are therefore allowed, though we fall short of topological methods. As finiteness of the Morley rank is inherited by definable subsets, the phrases “let G be a group of finite Morley rank” and “let G be a group definable in a finite Morley rank universe” describe the same setting. 1.2 Ranked Groups A ranked group structure enjoys a number of properties highly reminiscent of algebraic group theory: • the existence of a connected component [4, §5.2] H ◦ for each definable subgroup H ≤ G; 2 • generation by so-called indecomposable definable sets, similar to generation by irreducible subsets in algebraic groups [4, Theorem 5.26, known as Zilber’s Indecomposibility Theorem] (indecomposibility is a clever way to imitate topological irreducibility without a topology: a definable set is indecomposable if no definable subgroup can partition it using a finite number k > 1 of translates); • the ability to retrieve fields from definable, connected, infinite, soluble, non-nilpotent groups, as in the following fundamental fact. Fact A (Zilber’s Field Theorem: [4, Theorem 9.1]). Let G = A ⋊ H be a group of finite Morley rank where A and H are infinite definable abelian subgroups. Suppose the action is faithful and H-irreducible, viz. A has no definable, proper, infinite, H-invariant subgroup. Then the ring of endomorphisms of A generated by H is a definable field K, for which A ≃ K+ definably and H ֒→ K× in its action on K+ . Now an infinite field structure of finite Morley rank is necessarily algebraically closed by a theorem of Macintyre [4, Theorem 8.1] (note for the geometer: K may have Morley rank > 1). In particular there are no extensions of infinite definable fields in our setting, and consequently infinite fields of finite Morley rank have no definable groups of automorphisms [4, Theorem 8.3] (note: definable field automorphisms however exist, but only in characteristic p [4, p. 129]: it is not known whether they are all relative powers of the Frobenius map, which will make our conjecture interesting). But model theorists have been extremely plafyful with subgroups of K+ , resp. K× , which are not to be thought as “minimal” with respect to definability. Some of our arguments will look quite convoluted to the geometer, precisely because of this. Quite remarkably, we need virtually no other reprequisites than the above material — which we take as a strong indication that we are at the beginning of something not only new, but also of interest. 1.3 Ranked Modules A ranked module (G, V, ·) (or simply (G, V ) when there is no ambiguity) is a semi-direct product G ⋉ V such that: • G ⋉ V has finite Morley rank; • G and V are definable and infinite; • and V is abelian and connected. Be very careful that it is a requirement on V to be both definable and connected. All related notions are therefore connected versions: an irreducible module, a composition series, and so on. But in practice working in the definable-connected category is completely intuitive. The classification of ranked modules could be next to impossible. As a matter of fact, despite highly regarded efforts towards a well-known classification conjecture for the infinite simple groups of finite Morley rank, such abstract groups are not determined. But in the present paper we handle a very special case, where G is known to be isomorphic with SL2 (K). In particular, our work is almost disjoint in spirit from the ambitious classification programme we just mentioned, and definitely not about the so-called Cherlin-Zilber algebraicity conjecture. Now if SL2 (K) is definable inside a group of finite Morley rank, then so is K (the reader not familiar with definability should exercise here as it involves all techniques we shall use in the article); by the previous subsection, it is algebraically closed. Hence, if (G, V ) is a (faithful and irreducible) ranked module with G of the form SL2 (K), or more generally GK for G a semisimple algebraic group, then K is an algebraically closed field. In such an ideal world, how can V fail to be a representation of the algebraic group G? 1.4 On the Conjecture Let us repeat our conjecture. 3 Conjecture. Let (G, V ) be a module of finite Morley rank, where G is isomorphic to the group GK of K-points of some semisimple algebraic group G. Suppose that the action is faithful and irreducible. Then there are a definable, compatible K-vector space structure on V and definable field automorphisms ϕi such that V ≃ ⊗ϕi Ri , definably, as a K[G]-module. 1. Null Characteristic. Like Steinberg’s original theorem, the conjecture degenerates in characteristic 0, a case where it is already known. Fact B ([5, Lemma 1.4]). In a universe of finite Morley rank, consider the following definable objects: a field K, a quasi-simple algebraic group over K, a torsion-free abelian group V , and a faithful action of G on V for which V is G-irreducible. Then V ⋊ G is algebraic with respect to some K-vector space structure on V . One should note that this is a weak analogue of the equivalent statement in the so-called o-minimal category, where prior knowledge of G is not even required. Fact C (not used; Macpherson-Mosley-Tent; [13, Proposition 4.1]). Let (G, V ) be a module definable in an o-minimal structure. Suppose that the action is faithful and irreducible. Then there is a definable real closed field R such that V has definably the structure of a vector space over R, and G is definably isomorphic to a subgroup of GL(V ) in the standard action on V . Furthermore, if G is connected then its image in GLm (R) is R-semialgebraic. 2. Linearity. Let us return to the ranked setting. In positive characteristic, little is known. Needless to say, Steinberg’s slightly computational proof, and the more modern approach through Frobenius kernels as schemes [6] are both completely out of our methodological scope. And we confess to have no general strategy for the moment. Whether definable K-linearity of V would suffice is not completely clear: by the Borel-Tits corollary to Steinberg’s theorem, all irreducible linear representations of GK are known and of the desired form — but the definability of the field automorphisms used as twists is not fully obvious albeit highly expectable. We do however suspect that for given G = GK and V ≃ ⊗i ϕi Ri as above, ϕ2 the structures (K; ϕ , . . . , ϕϕn1 ) and (G, V, ·) should be inter-definable. It is a topic we shall return 1 to some day. But in any case, we see no general path to linearity, a much harder question. Our proof proceeds by defining the K-vector space structure manually, that is, with some hint of what the twist will be. 3. Field automorphisms. We wish to make a few more general comments. Fact D (not used; Macpherson-Pillay, Poizat; [14, Théorème 2]). Definable in a universe of finite Morley rank, let K be a field of prime characteristic and G ≤ GLn (K) be a simple subgroup. Then there are definable field automorphisms ϕ1 , . . . , ϕk such that G is definable in the structure (K; +, ·, ϕ1 , . . . , ϕk ). However remarkable this phenomenon and the proof Poizat gave, one hardly sees how to use it in our arguments. But we take it as supporting evidence for the conjecture: given a definable representation, the relevant category seems to be the rational category augmented by a finite number of twists. We wish to go further into mere speculation. In Steinberg’s classical result one computes the p-adic expansion of the highest weight; what would be the similar computation here? Observe how any two definable field automorphisms of a field of finite Morley rank commute over the algebraic closure of the prime field, hence everywhere; in particular, the ring Fp [Autdef (K)] generated by definable automorphisms of K is commutative. So computing in this ring would make (some) sense. This also do we take as a strong indication that our conjecture is the right statement. And now let us be specific. 4 1.5 The Result and its Legitimacy Theorem. Definable in a universe of finite Morley rank, let K be a field, G = SL2 (K), and V be an irreducible G-module of rank ≤ 4 rk K. Suppose the characteristic of K is neither 2 nor 3. Then there is a definable K-vector space structure on V such that V , as a K[G]-module, V is definably isomorphic to one of the following: • Nat G, the natural representation; • Ad G ≃ Sym2 Nat G, the adjoint representation; • Sym3 Nat G, the rational representation of degree 4 on K[X 3 , X 2 Y, XY 2 , Y 3 ]; • Nat G ⊗ χ (Nat G), a twist-and-tensor representation for some definable χ ∈ Aut(K). We do not know what happens in characteristics 2 and 3. But we suspect things would be much harder, and despite the challenge we confess that for the time being our interests lie more in higher Morley rank than in lower characteristic. The theorem is proved in §3, but it requires the tools developed in §2. We must legitimate our result — which we agree is but a very first step. However we consider it a breakthrough; there are three critical steps in the analysis: 1. leave the category of rational representations of SL2 (K); 2. remove bounds on the Morley rank of the SL2 (K)-module V ; 3. handle the other semisimple algebraic groups. As we said, until now only the case G ≃ SL2 (K) and rk V ≤ 3 rk K was known [5]. Here we successfully leave the rational category, which requires a complete change of methods. We view completion of 1 as evidence for the maturity of (some of) our methods. There are no claims now on how to pass the remaining steps. 1.6 On our Methods Let us describe the main features of the argument. • As opposed to [5] which handled the rk V ≤ 3 rk K case (see in particular [5, Step 5 in §2 and the Proposition in §3]), we are now working with modules greater in rank than the group. Consequently there can be no useful genericity argument since orbits of elements are all small when compared to the module. The present study entirely substitutes genericity arguments with much better techniques given in §2. Some of these were found by Tindzogho Ntsiri in his PhD [17], also [18]; what we cannot extend we faithfully reproduce for the sake of future reference. • We convert the abstract tools into methods based on the structure of V as a T -module (where T is an algebraic torus) and explained in §2.3: we shall even give a considerably shorter proof of [5] in §3.3. It is not unreasonable to hope that these methods can be generalised further. • The main issue with our proof is that we reproduce some of the unpleasant computations developed for SL2 (Z)-modules of short nilpotence length [9]. Now one should take intrinsic limitations on the unipotence length with care: [9, Proposition 2], which we briefly touch upon in §3.1, explains why computations which look possible in unipotence length ≤ 5 have a geometric obstruction in longer length. Although this geometric phenomenon simply cannot survive strong model-theoretic assumptions, for the moment there is no strategy other than the “short length miracle” of [9, Theorem 1]. In our opinion this is the main obstacle to jumping to identification of all rational representations of SL2 (K) among its definable modules. 5 But perhaps model theory, or a combination of finite group theory and model theory, could substitute such computations. This would open the way to something general and step 2 in §1.5 might then prove tractable. • In the twisted case we untensor by hand a linear structure we defined by hand. As we said in the previous subsection, one could naively hope to divide a solution to any case of the conjecture into two tasks: – proving linearity (in as general a context as possible); – proving in general that if V ≃ ⊗ϕi Ri is a definable GK -module, then the ϕi /ϕ1 are definable from G ⋉ V (see §1.4.2). But as we strongly suspect there is no direct path to linearity, analysing the conjecture in this two-step fashion hardly helps. This certainly reflects on our methods for the tensor case (Proposition 3.6 of §3.5): we isolate weight spaces and define the field action following what the twist should be; only after doing so do we actually retrieve the field automorphism. 2 General Tools The proof of our main Theorem will be in §3; this preliminary section contains tools whose interest goes beyond the rank 4k case. In §2.1 we state Tindzogho Ntsiri’s semi-simple analogue of Maschke’s Theorem. In §2.2 we enhance and simplify his systematic study of covariant, bi-additive maps between definable fields. These general results are applied to SL2 (K)-modules of finite Morley rank in §2.3. Before we start, we remind the reader that by convention, all modules are definable and connected; this affects the notions of irreducibility and composition series. 2.1 Abstract Tools 1: Actions of Tori Fact E ([10, Lemma G (iv)]). In a universe of finite Morley rank, let A be a definable, abelian group and R be a group acting on A by definable automorphisms. Let A0 ≤ A be a definable, R-invariant subgroup. Suppose that A has exponent p and R is a definable, soluble, p⊥ group. Then A = [A, R] ⊕ CA (R), ([A, R] ∩ A0 ) = [A0 , R], CA (R) covers CA/A0 (R), and CR (A) = CR (A0 , A/A0 ). It certainly is superfluous to explain the relevance of a definable version of Maschke’s Theorem. The result below is from Tindzogho Ntsiri’s PhD thesis [17]. Fact F (Tindzogho Ntsiri, [17, §5.2.1, Proposition 5.23] or [18, Corollary 2.17]). In a universe of finite Morley rank, let T be a definable, connected, soluble, p⊥ group and V be a T -module of exponent p. Suppose CV (T ) = 0. Then every T -submodule admits a direct complement T -module. Remark. The author wonders whether anything similar might exist in characteristic 0. This is mostly for art’s sake as the conjecture is settled there by Fact B. 2.2 Abstract Tools 2: Covariant Bi-additive Maps We move to studying T -covariant, bi-additive maps; the reason for doing so is the following. When analysing an arbitrary SL2 (K)-module, one will try to capture weight spaces as T -minimal modules; if the action of T is not trivial, Zilber’s Field Theorem recovers an algebraically closed field from every such module. But the interpretable field has no reason to be isomorphic to K, nor even to be the same on the various T -minimal modules. The hope is that linking the various weight spaces via the commutation map with a unipotent subgroup normalised by T should force interpretable fields to be definably isomorphic with K. A first result in this direction was developed (entirely by Cherlin) for [5] and served as a basis of Tindzogho Ntsiri’s work. 6 In a sense, the material in this section is again from [17]; but as opposed to Fact F, here we can generalise Tindzogho Ntsiri’s conclusions while simplifying his proof. The resulting statement is as follows. Theorem 2.1 (cf. [17, §5.1.1, Théorème 5.7] and [18, Theorem 3.2]). All objects (including functions and actions) are supposed to be definable in a universe of finite Morley rank. Let K, L, M be fields and U, V, W be vector spaces over K, L, M, respectively. Suppose there is a group T acting (not necessarily faithfully) by scalar multiplication on U , V , W , i.e. as a subgroup of K× , L× , M× . Let κ : T → K× , λ : T → L× and µ : T → M× be the associated group homomorphisms, and let ρ = κ × λ × µ : T → K× × L× × M× . Let β : U × V → W be a non-zero bi-additive map and suppose that β is plus-or-minusT -covariant, i.e. there are εK , εL , εM ∈ {±1} such that for all (t, u, v) ∈ T × U × V , one has β(tεK · u, tεL · v) = tεM · β(u, v). Then: • ker ρ = ker κ ∩ ker λ = ker κ ∩ ker µ; • if ker κ/ ker ρ is infinite, then L ≃ M; • if both ker κ/ ker ρ and im κ are infinite, then K ≃ L ≃ M; • if ker µ/ ker ρ is infinite, then K ≃ L; • if both ker µ/ ker ρ and im µ are infinite, then K ≃ L ≃ M. The proof will take place in §2.2.3; we start with easy remarks. 2.2.1 Two Remarks on Fields Notation. For a field F, let F′ = F \ {0} ∪ {∞} and define ∗ : F′ × F′ → F′ by:  ab for a, b ∈ F \ {0} with a 6= −b  a ∗ b = a+b a ∗ (−a) = ∞ for a ∈ F \ {0}  a ∗ ∞ = ∞ ∗ a = a for a ∈ F′ Extend · to F′ by a · ∞ = ∞ · a = ∞. Observation 1. (F′ , ∗, ·) is a field isomorphic to (F, +, ·) via the function ιF (a) = a−1 (with the convention 0−1 = ∞ and conversely); if F is definable then so are F′ and this isomorphism. Observation 2. Let F1 and F2 be fields, and denote by πi : F1 × F2 → Fi the projection maps. If a subring R ≤ F1 × F2 is a field, then π1 (R) ≃ π2 (R) (actually R stands for the graph of one such field isomorphism; alternatively, π1 (R) ≃ R ≃ π2 (R)). In particular F1 and F2 have the same characteristic. If in addition F1 , F2 , and R are definable in a theory of finite Morley rank with R infinite, then F1 ≃ F2 . Proof. It suffices to see that R is a one-to-one, functional relation. Since R is an additive subgroup of F1 × F2 , it suffices to see that (0, y) ∈ R entails y = 0: which is obvious since R is a field. For the second claim, bear in mind that there can be no extensions of infinite, definable fields for all such are algebraically closed [4, Theorem 8.1]. ♦ 2.2.2 An Example to Understand The following will not be used anywhere in the present article, but it involves all ideas underlying the proof of Theorem 2.1. Lemma. All objects (including functions and actions) are supposed to be definable in a universe of finite Morley rank. Let K, L be fields and U, V be vector spaces over K, L, respectively. Suppose there is a group T acting (not necessarily faithfully) by scalar multiplication on U, V , i.e. as a subgroup of K× , L× . 7 Let κ : T → K× and λ : T → L× be the associated group homomorphisms, and let ρ = κ × λ : T → K× × L× . Let α : U → V be a non-zero additive map and suppose that α is either T -covariant, i.e. α(t · u) = t · α(u), or T -contravariant, i.e. α(t · u) = t−1 · α(u). Then: • ker ρ = ker κ = ker λ; • if im κ (or equivalently, im λ) is infinite, then K ≃ L definably. Proof. Since α is non-zero, there is u0 ∈ U with α(u0 ) 6= 0. Let t ∈ ker κ. Then λ(t) · α(u0 ) = α(κ(t)±1 · u0 ) = α(u0 ) 6= 0, where the value ±1 depends on whether α is T -covariant or T contravariant; hence λ(t) = 1. Conversely, if t ∈ ker λ \ ker κ, then letting u1 = (κ(t) − 1)−1 · u0 one finds: α(u0 ) = α((κ(t) − 1) · u1 ) = λ(t)±1 · α(u1 ) − α(u1 ) = 0 which is a contradiction. We prove the second claim for a covariant map. Let FK,L = {(k, ℓ) ∈ K × L : ∀u ∈ U, α(k · u) = ℓ · α(u)}. One easily checks that FK,L is a definable subfield of the ring K × L. Also let GK,L = F× K,L . If im κ ≃ T / ker κ ≃ im λ is infinite, then so is ρ(T ) ≤ GK,L , and therefore FK,L too. It follows K ≃ L. The proof of the contravariant case is even more instructive. Let FK,L′ = {(k, ℓ) ∈ K × L \ {0} : ∀u ∈ U, ℓ · α(k · u) = α(u)} ∪ {(0, ∞)}. We claim that FK,L′ is a definable subfield of K × L′ ; perhaps the only non-trivial claim is stability under the additive law. So suppose (ki , ℓi ) ∈ K × L′ for i ∈ {1, 2}; the cases where ℓi = ∞ or ℓ1 + ℓ2 = 0 are easily dealt with, so we assume not. Then for u ∈ U : ℓ1 ℓ2 · (α(k1 · u) + α(k2 · u)) ℓ1 + ℓ2 ℓ1 ℓ2 · α(u) + · α(u) = ℓ1 + ℓ2 ℓ1 + ℓ2 = α(u) (ℓ1 ∗ ℓ2 ) · α((k1 + k2 ) · u) = as desired. Also let GK,L′ = {(k, ℓ) ∈ K× × L′× : ∀u ∈ U, ℓ · α(k · u) = α(u)}; clearly GK,L′ = F× K,L′ . If im κ is infinite, then by contravariance so are ρ(T ) ≤ GK,L′ and FK,L′ . Hence K ≃ L′ ≃ L. Using the same argument twice is unsatisfactory. So suppose again that α is T -contravariant. By means of the inverse isomorphism ιL : L ≃ L′ , view V as a vector space over L′ . Then α : U → V is T -covariant. Since the other assumptions and the statements are invariant under this change, the contravariant case actually reduces to the covariant one (first claim included). 2.2.3 Proof of Theorem 2.1 Proof of Theorem 2.1. Throughout the proof we assume εK = εL = εM = 1; we explain why in the end. The proof that ker ρ = ker κ ∩ ker λ = ker κ ∩ ker µ is essentially as in the previous Lemma. Since β is non-zero there is (u0 , v0 ) ∈ U × V with w0 = β(u0 , v0 ) 6= 0. If t ∈ ker κ ∩ ker λ, then µ(t) · w0 = β(κ(t) · u0 , λ(t) · v0 ) = w0 6= 0, so µ(t) = 1. Conversely for t ∈ (ker κ ∩ ker µ) \ ker λ, let v1 = (λ(t) − 1)−1 · v0 . Then: β(u0 , v0 ) = β(u0 , (λ(t) − 1) · v1 ) = β(κ(t) · u0 , λ(t) · v1 ) − β(u0 , v1 ) = µ(t) · β(u0 , v1 ) − β(u0 , v1 ) = 0 which is a contradiction. We now need a theory of notations. Notation. • Let FK,M = {(k, m) ∈ K × M : ∀(u, v) ∈ U × V, β(k · u, v) = m · β(u, v)}. • Let FK,L′ = {(k, ℓ) ∈ K × (L \ {0}) : ∀(u, v) ∈ U × V, β(k · u, ℓ · v) = β(u, v)} ∪ {(0, ∞)}. 8 • Let GK,L,M = {(k, ℓ, m) ∈ K× × L× × M× : ∀(u, v) ∈ U × V, β(k · u, ℓ · v) = m · β(u, v)}. • Let πK× : K× × L× × M× → K× denote projection on K× (define πL× and πM× likewise). • Let GK,M = GK,L,M ∩ ker πL× = {(k, 1, m) : (k, m) ∈ F× K,M }. • Let GK,L′ = GK,L,M ∩ ker πM× = {(k, ℓ, 1) : (k, ℓ) ∈ F× K,L′ }. One should also define FL,M and GL,M similarly. As in the above Lemma, FK,M is a definable subfield of K × M and FK,L′ is one of K × L′ . Clearly GK,L,M is a subgroup of K× × L× × M× ; by the covariance assumption, ρ(T ) ≤ GK,L,M . We are finally ready for the proof. Suppose ker κ/ ker ρ to be infinite. This means that ρ(ker κ) ≤ GK,L,M ∩ ker πK× = GL,M is infinite, and therefore so is FL,M : it follows that L ≃ M. Suppose that both ker κ/ ker ρ and im κ are infinite. Here again FL,M is infinite, so it surjects onto L: hence πL (FL,M ) = L and πL× (GL,M ) = L× . But since im κ is infinite too, GL,M has infinite index in ρ(T ) · GL,M ≤ GK,L,M . Therefore ker πL× ∩ GK,L,M = GK,M is infinite; so is FK,M , proving K ≃ M. Now suppose that ker µ/ ker ρ is infinite. This means that ρ(ker µ) ≤ GK,L′ is infinite, and therefore so is FK,L′ : it follows that K ≃ L′ ≃ L. Finally suppose that both ker µ/ ker ρ and im µ are infinite. Here again πL× (FK,L′ ) = L× (one should remember that L× and L′× have the same underlying set). But in addition, im µ is infinite, so GK,L′ has infinite index in ρ(T ) · GK,L′ ≤ GK,L,M : proving that ker πL× ∩ GK,L,M = GK,M is infinite, and K ≃ M. It remains to cover the other cases. Whenever εF = −1, view the corresponding F-vector space as a vector space over F′ via the inverse isomorphism ιF : F ≃ F′ . Then β is covariant in the usual sense β(t · u, t · v) = t · β(u, v). Remarks. • Tindzogho Ntsiri has another result [17, §5.1.2, Proposition 5.15 and Lemme 5.17] or [18, Proposition 3.10]: assuming the characteristic (which is obviously common to K, L, M, for instance by our method) is p and κ, λ, µ are onto, one has K+ ≃ L+ ≃ M+ and K× ≃ L× ≃ M× . We shall not use this. • The author does not care for n-additive maps with n ≥ 3. 2.3 Applied Tools: SL2 (K)-Modules We want some general results on SL2 (K)-modules of finite Morley rank; actually, part of the analysis requires less than the entire SL2 (K)-module structure and we shall try to clarify matters. This results in the notion of a non-trivial T -line, Definition 2.9 below. An important feature of our study is that we restrict to positive characteristic settings, as Fact B in §1.4 entirely settles the definable representation theory of algebraic groups in characteristic 0. From here on we adopt the following notations. Notation 2.2. • Let K be a definable field of positive characteristic p and G = SL2 (K); • let U be a maximal unipotent subgroup of G, B = NG (U ) be the associated Borel subgroup, and T ≤ B be an algebraic torus; • let w ∈ NG (T ) \ T (an element of order at most 4 inverting T ). Finally recall that H-module means: definable, connected, H-module. Fact G (from [3, Lemma B]). If V is a U -module, then the semi-direct product U ⋉ V is nilpotent. (One may however observe that this could fail in characteristic 0.) Notation 2.3. Let V be a U -module. 9 • Let Z0 = 0 and Zj+1 /Zj = CV◦ /Zj (U ). (As a consequence of Zilber’s Indecomposibility Theorem, [U, Zj+1 ] ≤ Zj .) • Let Yj = Zj /Zj−1 . • Let ℓU (V ) be the least integer with Zn = V . We call ℓU (V ) the U -length, or unipotence length, of V . (This is not the same measure as in the proof of Fact F.) By Fact G, the series (Zj ) is strictly increasing until one reaches ZℓU (V ) = V . Notice that if V is actually a B-module, then so are Zj and Yj . Lemma 2.4 (from [5, §2.1, Step 7]). Let V be a B-module. Then [U, CV (T )] ≤ [T, V ]. Proof. For brevity we let C = CV (T ) and D = [T, V ]. Let c0 ∈ C and u ∈ U \ {1}. Since V = C ⊕ D by Fact E, there is (c, d) ∈ C × D with [c0 , u] = c + d. Now for any t ∈ T , [c0 , t · u] = c + t · d ∈ hci + D. Since T · u generates U , one finds [c0 , U ] ≤ hci + D; by Zilber’s Indecomposibility Theorem and since the characteristic is finite, actually [c0 , U ] ≤ D, as desired. Lemma 2.5. Let V be a B-module of U -length exactly n and 2 ≤ j ≤ n. Then T cannot centralise both Yj and Yj−1 (see Notation 2.3). More precisely, if T centralises Yj−1 then CYj (T ) = 0. Proof. Let β : U × Yj → Yj−1 be given by: β(u, z + Zj−1 ) = [u, z] + Zj−2 It is routine to check that β is well-defined, bi-additive, and T -covariant. We claim that the right radical of β (usually denoted U ⊥ ) is finite. For if z + Zj−1 is βorthogonal to U , then [U, z] ≤ Zj−2 and z ∈ CV /Zj−2 (U ), the connected component of which is exactly Zj−1 . Now assuming that T centralises Yj−1 , let x ∈ CYj (T ) and (u, t) ∈ U × T ; one finds: β(t · u − u, x) = t · β(u, x) − β(u, x) = 0 Since [T, u] = U , we get β(U, x) = 0 for arbitrary x ∈ CYj (T ), forcing finiteness of the latter, and its triviality by connectedness of Yj and Fact E. Lemma 2.6. Let V be an irreducible G-module of U -length exactly n. Then Z1 ∩ w · Zn−1 is finite. Proof. We claim that V0 = hG · (Z1 ∩ w · Zn−1 )i ≤ Zn−1 . For let v ∈ Z1 ∩ w · Zn−1 and g ∈ G; by the Bruhat decomposition G = B ⊔ BwU there are two cases. • If g ∈ B, write g = tu with obvious notations; then g · v = t · v ∈ Z1 ≤ Zn−1 . • If g ∈ BwU then g = bwu and g · v = bw · v ∈ Zn−1 again. Since Zn−1 is proper in the irreducible G-module V , the G-module V0 is finite. Lemma 2.7. Let V be an irreducible G-module of U -length exactly n. Then T does not centralise Yn = Zn /Zn−1 = V /Zn−1 . Proof. Suppose it does. By Lemma 2.6, w · Z1 ∩ Zn−1 is finite, so for some finite K the quotient w · Z1 /K embeds into Yn as a T -module. It follows that T centralises w · Z1 /K: by connectedness of Z1 , T centralises w · Z1 . Also notice that by assumption and Fact E, [T, V ] ≤ Zn−1 . Now let u ∈ U be such that (uw)3 = 1 (see Notation 3.1 below if necessary), let i stand for the central involution of G (it is harmless to let i = 1 in characteristic 2), and a ∈ Z1 . Then: uwu · a = uwa = wa + [u, wa] = (wuw)−1 · a = wu−1 wa = ia + w[u−1 , wa]. Notice that [u, wa] ∈ [U, Zn ] ≤ Zn−1 . Moreover, since wa ∈ w · Z1 ≤ CV (T ), one has [u−1 , wa] ∈ [U, CV (T )] ≤ [T, V ] by Lemma 2.4, so w[u−1 , wa] ∈ w · [T, V ] = [T, V ] ≤ Zn−1 as observed. As a consequence wa = ia + w[u−1 , wa] − [u, wa] ∈ w · Z1 ∩ Zn−1 which is finite by Lemma 2.6, against Z1 being infinite. 10 Lemma 2.8. Let V be a B-module. For j 6= 1, Yj from Notation 2.3 is the direct sum of CYj (T ) and of non-central T -minimal modules of the same rank as K. If V is actually an irreducible G-module, then the same holds of j = 1. Proof. We already know from Fact E that Yj = CYj (T ) ⊕ [T, Yj ], so by Maschke’s Theorem (Fact F) the statement is about T -minimal submodules of the latter summand, if any. Let L ≤ [T, Yj ] be one such. Take again the bi-additive, T -covariant function β : U × Yj → Yj−1 from Lemma 2.5; remember that the right radical is finite. So the restriction β : U × L → Yj−1 is non-zero. Let M be a T -minimal subquotient of Yj−1 such that π ◦ β : U × L → M still remains non-zero. Now using the action of T on L, M respectively and Zilber’s Field Theorem, interpret fields L, M (possibly M = Fp if T centralises M ), then prepare to apply Theorem 2.1. Observe from the structure of Borel subgroups of SL2 (K) that here ker κ = Z(G) has order at most 2. Suppose rk L 6= rk K. Notice that ker◦ λ 6= 1; this is clear if rk L < rk K, and a consequence of Wagner’s Torus Theorem [19] if rk L > rk K as T ≃ K× cannot embed properly into L× in positive characteristic. Hence ker◦ λ and ker λ/ ker ρ are infinite. But also im◦ λ 6= 1 since otherwise T centralises L, against its definition. By Theorem 2.1, K ≃ L, against rk L = rk L 6= rk K. This shows rk L = rk K. If V is actually an irreducible G-module and j = 1, use Lemma 2.6 to embed w · Z1 into V /Zn−1 = Yn as a T -module (up to a finite kernel). Definition 2.9. A non-trivial T -line is a non-central, T -minimal module of same rank as K. For instance, it is a consequence of Lemmas 2.7 and 2.8 that Zn contains a non-trivial T -line not included in Zn−1 . Corollary 2.10. If V is an irreducible G-module, then [T, V ] is a direct sum of an even number of non-trivial T -lines. Proof. Let us refine the non-maximal B-series (Zj ) into a composition T -series: one gets T -central factors and non-trivial T -lines. The number of the latter must be even since w normalises [T, V ] and a w-invariant, T -minimal module must be centralised by T , as there are no definable groups of automorphisms of fields of finite Morley rank [4, Theorem 8.3]. The careful reader will observe that one ought to pay more attention to factors in composition series than we just did: since we work in the connected category, uniqueness in a Jordan-Hölder property can hold only up to isogeny. But in the case of connected T -modules, this is harmless. Remarks. • A weaker result had been obtained by Tindzogho Ntsiri [17, §5.2.2, Proposition 5.31] or [18, Theorem 4.6] under the assumption that SL2 (K) acts faithfully on V . • These results are far from optimal. The main problem is that the centraliser CV (T ) could be “scattered” among the various Yj , i.e. CYj (T ) could be non-zero for several indices, something prohibited in the algebraic setting. This for the moment seems to be the major obstruction to understanding the B-structure of SL2 (K)-modules of arbitrary rank, and we view it as a crucial investigation to lead in the near future. 3 Proof of the Main Theorem Let us state our result again. Theorem. Definable in a universe of finite Morley rank, let K be a field, G = SL2 (K), and V be an irreducible G-module of rank ≤ 4 rk K. Suppose the characteristic of K is neither 2 nor 3. Then there is a definable K-vector space structure on V such that V , as a K[G]-module, V is definably isomorphic to one of the following: • Nat G, the natural representation; 11 • Ad G ≃ Sym2 Nat G, the adjoint representation; • Sym3 Nat G, the rational representation of degree 4 on K[X 3 , X 2 Y, XY 2 , Y 3 ]; • Nat G ⊗ χ (Nat G), a twist-and-tensor representation for some definable χ ∈ Aut(K). Let us confess again that we do not know what happens in low characteristic. 3.1 Notations and Facts The characteristic 0 case is covered by Fact B from §1.4, so let us turn to positive characteristics. We keep and extend Notations 2.2 (see §2.3); as a matter of fact we need to coordinatise G = SL2 (K). Notation 3.1. For λ ∈ K+ (resp., K× ), let:    t 1 λ uλ = , tλ = λ 1 t−1 λ  , w=  1 −1  We let i = t−1 be the central involution of G (by assumption, p 6= 2); of course w2 = i. Notice that letting U = {uλ : λ ∈ K+ } ≃ K+ and T = {tλ : λ ∈ K× } ≃ K× is consistent with Notation 2.2. Also notice that tµ uλ t−1 µ = uλµ2 . Finally check that: uλ wuλ−1 wuλ w = tλ For brevity let u = u1 ; then the above relation simplifies into (uw)3 = 1. Notation 3.2. Let V be an irreducible G-module of rank ≤ 4 rk K. For brevity let k = rk(K). Notation 3.3. For λ ∈ K, let ∂λ = aduλ : V → V map v to [uλ , v] = uλ v − v. For brevity let ∂ = ∂1 . Hence ∂λ = uλ − 1 in End(V ). Notice that tµ ∂λ t−1 µ = ∂λµ2 . The rank 4k analysis starts here. The structure of the proof is rather straightforward. We first bound the unipotence length in Proposition 3.4 (§3.2). Small rank cases, where rk V ≤ 3k, are known by [7] and [5]; we however give shorter proofs in §3.3. Then real things begin. We cover Sym3 Nat G in Proposition 3.5 (§3.4), and then Nat G ⊗ χ (Nat G) in Proposition 3.6 (§3.5). We shall rely on a non-model-theoretic investigation of Symn Nat G. In the following, composition series are in the classical sense (no definability nor connectedness required). Fact H ([9, Theorem 1]). Let V be a Q[SL2 (Z)]-module. Suppose that for every unipotent element u ∈ SL2 (Z), (u − 1)5 = 0 in End(V ). Then V has a composition series each factor of which is a direct sum of copies of Q ⊗Z Symk Nat SL2 (Z) for k ∈ {0, . . . , 4}. Although Fact H is stated for SL2 (Z)-modules, its highly computational proof readily adapts to Fp [SL2 (Fp )]-modules, provided p is greater than the unipotence length; incidently the author does not know whether a character-theoretic proof exists in that case. One should however refrain from using Fact H. As indicated by [9, Proposition 2], there are deep geometric reasons which prevent the computation from extending to length ≥ 7, at least for Q[SL2 (Z)]-modules. So we are dissatisfied with this result, albeit next-to-optimal. A result we are satisfied with is the following. Fact I ([9, Theorem 2]). Let n ≥ 2 be an integer and K be a field of characteristic 0 or ≥ 2n + 1. Suppose that K is 2(n − 1)!!-radically closed. Let G = SL2 (K) and V be a G-module. Let K1 be the prime subfield and G1 = SL2 (K1 ). Suppose that V is a K1 -vector space such that V ≃ ⊕I Symn−1 Nat G1 as K1 [G1 ]-modules. Then V bears a compatible K-vector space structure for which one has V ≃ ⊕J Symn−1 Nat G as K[G]-modules. Moreover, the linear structure is definable over G by construction. 12 3.2 Bounding the Length Proposition 3.4. V as in Notation 3.2 has U -length at most 4. Proof. If there are four non-trivial T -lines (see Definition 2.9) we are done for rank reasons. So suppose there are only two. By Lemma 2.7 at least one occurs in the final factor Yn ; by Lemma 2.5 the other one cannot occur more than two steps afar, and cannot be more than two steps away from Z0 . The worst case is therefore where the two lines occur at stages 2 and 4: ℓU (V ) ≤ 4. 3.3 Warming up: Shorter Proofs in Rank 3k Since one can also measure technical progress in a topic by the ability to produce shorter proofs of existing theorems, we shall now revisit two earlier results. Fact J (rank 2k analysis: [7, Theorem B]). If V as in Notation 3.2 has rank ≤ 2k, then there is a definable K-vector space structure such that V ≃ Nat G. The proof will appear only marginally shorter than the original argument, but it is self-contained as opposed to [7] which relied on work by Timmesfeld [16, Theorem 3.4]. (From the latter we extract only the core computation, the group SL2 (K) being already identified and coordinatised.) Proof. For rank reasons, it follows from Corollary 2.10 that [T, V ] is the direct sum of exactly two non-trivial T -lines. As a consequence, V = Z2 has U -length 2 and CV (T ) = 0. Now let a ∈ Z1 ; with a look at Notation 3.3, set b = ∂wa = [u, wa] ∈ [U, V ] ≤ Z1 , and c = ∂wb ∈ Z1 . One finds: (uw)2 · a = uw(wa + b) = ia + wb + c = (uw)−1 · a = iwu−1 a = iwa so w(ia − b) = ia + c ∈ Z1 ∩ w · Z1 ≤ CV (U, wU w−1 ) = CV (G) ≤ CV (T ) = 0. Hence b = ia. Now for arbitrary λ ∈ K× , one still has tλ a ∈ Z1 so: uλ2 wa = tλ utλ−1 wa = tλ uwtλ a = tλ (wtλ a + itλ a) = wa + itλ2 a and since K is algebraically closed, one finds uλ wa = wa + itλ a. We may now define the vector space structure. On Z1 let λ · a = tλ a; on w · Z1 let λ · (wa) = w(λ · a). This is obviously additive in a and multiplicative in λ; moreover, since U centralises [U, V ] ≤ Z1 , (λ + µ) · a = tλ+µ a = i[uλ+µ , wa] = i(uλ · [uµ , wa] + [uλ , wa]) = i2 (tµ a + tλ a) = λ·a+µ·a so this does define a linear structure on Z1 ⊕ w · Z1 . (We already observed that the sum is direct indeed.) We claim that Z1 ⊕ w · Z1 is then a K[G]-module. First observe that it is hw, U i = G-invariant alright. By construction w acts linearly. Linearity of U is obvious on Z1 . Now for (λ, µ) ∈ K2 and a ∈ Z1 , one has: uµ (λ · wa) = uµ w(λ · a) = w(λ · a) + itµ tλ a = λ · wa + λ · (itµ a) = λ · uµ a which proves linearity of U on w · Z1 too. By irreducibility, V = Z1 ⊕ w · Z1 which is clearly isomorphic to Nat G. Remarks. • The argument applies to an irreducible G-module of U -length exactly 2 such that CV (T ) = 0, without mentioning the rank nor even its finiteness: this is a special case of Facts H and I. • In the case of an irreducible G-module of U -length exactly 2 such that CV (T ) = 0, there are no restrictions on the characteristic, which may be 0, 2 or 3 in the above computation. 13 • In order to remove from the latter the CV (T ) = 0 assumption in a finite Morley rank setting, one can follow [7, Claim 5.7] or its generalisation to SLn (K) [3, Proposition 2.1, Claim 1]. Fact K (rank 3k analysis: [5]). If V as in Notation 3.2 has rank ≤ 3k, then there is a definable Kvector space structure such that either V ≃ Nat G or V ≃ Sym2 Nat G (the adjoint representation). Proof. We may assume rk V > 2k; then ℓU (V ) ≥ 3 since otherwise V /CV (G) has rank 2k, and V has too by finiteness of CV (G). Here again, [T, V ] is the direct sum of two non-trivial T -lines; it follows that CV (T ) is infinite; in particular, the central involution i acts as 1. Notice that [U, CV (T )] 6= 0 as otherwise U centralises CV (T ) = w · CV (T ), so G = hU, wU w−1 i centralises CV (T ); by irreducibility the latter is therefore finite (notice that we did not reduce to CV (G) = 0), a contradiction. Hence [U, CV (T )] 6= 0; on the other hand equality does not hold in the inclusion [U, CV (T )] ≤ [T, V ] given by Lemma 2.4: as otherwise [T, V ] is hU, wi = G-invariant, so by irreducibility again, V = [T, V ], against CV (T ) 6= 0. The conclusion of the above analysis is that [U, CV (T )] consists in exactly one non-trivial T -line L; it follows that [T, V ] = L ⊕ w · L. Since L is U -invariant and T -minimal, it is centralised by U ; hence L ≤ Z1 . By Lemma 2.7 there is a non-trivial T -line in Yn , where n > 2 is the unipotence length, and in particular there is none in Yn−1 , . . . , Y2 : so by Lemma 2.5, n − 1 = 2; V is a cubic module in the sense [U, U, U, V ] = 0. Let us study its structure further. First we see that T centralises Y2 . A priori, by Fact E, one should expect Z1 = L ⊕ X1 with X1 = CZ1 (T ). But by Lemma 2.6, w · X1 embeds (modulo a finite kernel) into Y3 : both w · X1 and Y2 are centralised by T , which violates Lemma 2.5. So X1 = 0 and Z1 = L = [U, CV (T )]. Likewise, [T, Y3 ] = Y3 is a non-trivial T -line. Rank reasons then yield V = L ⊕ CV (T ) ⊕ w · L. It is time to identify the action. Over the prime field Fp one could rely on Fact H but this is not the spirit, so we perform a subset of the mentioned computations. In view of the decomposition V = L ⊕ CV (T ) ⊕ w · L, there are additive maps c : L → CV (T ) and ℓ : L → L such that for any a ∈ L = Z1 , one has uwa = wa + c(a) + ℓ(a). Recall from Notation 3.3 that ∂ stands for adu = u − 1 ∈ End(V ). Hence: uwuwa = uw(wa + c(a) + ℓ(a)) = a + wc(a) + ∂wc(a) + wℓ(a) + c(ℓ(a)) + ℓ(ℓ(a)) −1 = (uw) a = wa Projecting onto w · L, one finds ℓ(a) = a; projecting onto CV (T ), wc(a) = −c(ℓ(a)) = −c(a). If c(a) = 0 for a 6= 0, then actually uwa = wa + a and therefore wa = (uw)−1 a = (uw)2 a = uw(wa + a) = a + wa + a, a contradiction to the characteristic not being 2. So c : L → CV (T ) is injective, and therefore rk CV (T ) ≥ k; c is also surjective, so the previous computation shows that w inverts CV (T ). Now let d(a) = ∂c(a) ∈ [U, CV (T )] = L. Find: uwuwc(a) = −uw(c(a) + d(a)) = c(a) + d(a) − wd(a) − c(d(a)) − d(a) = (uw)−1 c(a) = wc(a) − wd(a) = −c(a) − wd(a) so c(d(a)) = 2c(a). Since the map c : L → CV (T ) is injective, d(a) = 2a. It is now clear that V is a direct sum of copies of Sym2 Nat G1 , where G1 = hu, wi = SL2 (Fp ). We then apply Fact I. Since the resulting K-linear structure is definable, there is a definable isomorphism with Sym2 Nat G. (After reading again the end of the argument of [5], we have no clue what Cherlin meant that day.) Remarks. • Here we did assume that the characteristic was neither 0 (at the beginning of the argument, when we used Lemma 2.4) nor 2 (at the end of the argument, for our computations); it could however be 3. 14 • The characteristic 2 case is dealt with quickly in [5]; remember that the characteristic 0 case is most efficiently handled by the model theory of Fact B. • The argument does not classify all cubic SL2 (K)-modules of finite Morley rank (if it did there would be no Proposition 3.6). What it does classify is those cubic modules with descending coherence degree κ(V ) ([8, §5]) equal to 1, i.e. with Z1 = ker ∂; there are no assumptions on the rank. The interested reader should try to prove directly that ker ∂ = Z1 in the rk V ≤ 3k configuration; even our quickest argument goes through splitting V = L ⊕ CV (T ) ⊕ w · L. 3.4 The Quartic Analysis Proposition 3.5. If V as in Notation 3.2 has U -length exactly 4, then there is a definable K-vector space structure such that V ≃ Sym3 Nat G. Before the proof it could be good to remind the reader that Sym3 Nat G is naturally isomorphic to the representation of SL2 (K) in the 4-dimensional space K[X 3 , X 2 Y, XY 2 , Y 3 ] of homogeneous polynomials in X, Y with degree 3, and to encourage him to perform a few calculations there (exponentiating the action of the Lie algebra is an alternative to direct computation). Proof of Proposition 3.5. 3.5.1 — Notation. For j = 1 . . . 4, let Žj = (Zj ∩ w · Zn+1−j )◦ (here n = 4). 3.5.2 — Claim. V = ⊕4j=1 Žj . Proof. If V = [T, V ], the claim follows from 2.10: in that case and for a rank reason, V is a direct sum of exactly four non-trivial T -lines, and since the U -length is exactly four, one sees that each Yi is a non-trivial T -line. Then intersections are easily taken. Notice that V = [T, V ] would follow from the next Claim 3.5.3 — but we do not know the action of the central involution yet. So let us assume towards a contradiction that CV (T ) 6= 0. Now there are exactly two non-trivial T -lines. Since T does not centralise Y4 by Lemma 2.7, and in view of Lemma 2.5, these lines are in Y2 and Y4 (an argument already used for Proposition 3.4). At this point T centralises Y1 and Y3 , so by Lemma 2.5 again, [T, Y4 ] = Y4 is a non-trivial T -line. On the other hand, by Lemma 2.6, w · Z1 embeds modulo a finite kernel into Y4 , so w · Z1 is a non-trivial T -line: Z1 as well, a contradiction. ♦ 3.5.3 — Claim. The central involution i acts as −1. Proof. The awful computation underlying Fact H (more specifically, [9, §1.2.7]) is not fully conclusive since one could imagine a globally quartic PSL2 (Z)-module which splits into cubic-by-cubic. So let us compute. Suppose i acts as 1; fix a ∈ Z1 . Noticing w · Ž2 = Ž3 and given the decomposition of Claim 3.5.2, there are additive maps α : Z1 → Z1 , and β, γ : Z1 → Ž2 such that: uwa = wa + wγ(a) + β(a) + α(a) Hence: uwuwa = a + γ(a) + ∂γ(a) + wβ(a) + ∂wβ(a) + wα(a) + wγ(α(a)) + β(α(a)) + α(α(a)) = (uw)−1 a = wa Projecting onto Ž4 one finds α(a) = a, then projecting onto Ž3 one finds γ(a) = −β(a), so finally ∂wβ(a) = ∂β(a) − 2a. Then letting δ(a) = ∂β(a) − 2a ∈ Z1 : uwuwβ(a) = uw(wβ(a) + δ(a)) = β(a) + ∂β(a) + wδ(a) − wβ(δ(a)) + β(δ(a)) + δ(a) −1 = (uw) β(a) = wβ(a) − w∂β(a) 15 Now this yields δ(a) = −∂β(a) = ∂β(a) − 2a, and ∂β(a) = a = −δ(a). Hence ∂wa = −wβ(a) + β(a) + a entails ∂ 2 wa = 0. We just proved ∂ 2 (w · Z1 ) = 0; conjugating by T , one has ∂λ2 (w · Z1 ) = 0 for any λ ∈ K; since the characteristic is not 2, ∂λ ∂µ (w · Z1 ) = 0 for any (λ, µ) ∈ K2 . This means w · Z1 ≤ Z2 . Now Lemma 2.6 forces the U -length to be 2, a contradiction. ♦ 3.5.4 — Claim. For λ ∈ K× there is an abelian group homomorphism βλ : Z1 → Ž2 such that for any a ∈ Z1 : uλ wa = wa + wt−1 λ βλ (a) + βλ (a) − tλ a As a matter of fact, βλ2 (a) = tλ β1 (tλ a). Proof. We shall first do it over the prime field, i.e. for λ = 1, and vary the field and values of λ afterwards. So for the moment set λ = 1; this is a quick computation, very similar to Claim 3.5.3 (as a matter of fact, one could reach both Claims simultaneously but the exposition may not benefit from doing so). Fix a ∈ Z1 . Noticing w · Ž2 = Ž3 and given the decomposition of Claim 3.5.2, one should expect additive maps α : Z1 → Z1 , and β, γ : Z1 → Ž2 such that: uwa = wa + wγ(a) + β(a) + α(a) Bearing in mind that (uw)3 = 1, one finds: uwuw · a = u(−a − γ(a) + wβ(a) + wα(a)) = −a − γ(a) − ∂γ(a) + wβ(a) + ∂wβ(a) + wα(a) + wγ(α(a)) + β(α(a)) + α(α(a)) = (uw)−1 a = −wu−1 a = −wa Projecting onto Ž4 = w·Z1 , this yields α(a) = −a; projecting onto Ž3 , one finds γ(a) = −β(α(a)) = β(a). Now let λ ∈ K× and ℓ ∈ K× be a square root. Then: uλ wa = tℓ ut−1 ℓ wa = tℓ uwtℓ a = tℓ (wtℓ a + wβ(tℓ a) + β(tℓ a) − tℓ a) = wa + tℓ wβ(tℓ a) + tℓ β(tℓ a) − tλ a = wa + wt−1 λ tℓ β(tℓ a) + tℓ β(tℓ a) − tλ a So one should let βλ (a) = tℓ β(tℓ a) to find the desired decomposition. ♦ Model theory, more specifically connectedness arguments, will play a role in the next Claim. Without finiteness of the rank, it might fail (the brave should see [9, §1.2.8]). 3.5.5 — Claim. βλ : Z1 → Ž2 is a group isomorphism. Proof. Let a ∈ Z1 , and introduce the definable sets A = {µ ∈ K× : βµ (a) = 0} ∪ {0} and X = K+ \ A. Suppose A 6= 0 and fix some λ ∈ A \ {0}, ie. with βλ (a) = 0. Thanks to Claim 3.5.4, one has for arbitrary µ ∈ K× : uµ uλ wa = uµ (wa − tλ a) = wa + wt−1 µ βµ (a) + βµ (a) − tµ (a) − tλ (a) = uλ+µ wa = wa + wt−1 λ+µ βλ+µ (a) + βλ+µ (a) − tλ+µ a Projecting onto Ž2 , one finds βλ+µ (a) = βµ (a). Hence A is a subroup of K+ . Let µ ∈ X, i.e. with βµ (a) 6= 0, if there is such a thing. Returning to the equation and now −1 projecting onto Ž3 , one finds tµ−1 βµ (a) = t−1 λ+µ βλ+µ (a) = tλ+µ βµ (a). Hence t(λ+µ)µ−1 centralises βµ (a) 6= 0. Then by T -minimality of Ž2 and Zilber’s Field Theorem, t(λ+µ)µ−1 centralises all of Ž2 . 16 Therefore X ⊆ {µ ∈ K× : t(λ+µ)µ−1 ∈ CT (Ž2 )}. But CT (Ž2 ) is proper (hence not of maximal rank) in the connected group T , and the map µ 7→ 1 + µλ is injective on X, so X is not of maximal rank in K+ . Hence A is of maximal rank, but it also is a subgroup. This proves A = K+ [4, Corollary 5.13 if not obvious], and βµ (a) = 0 for any µ ∈ K× . In particular, uµ wa = wa − tµ a for all µ ∈ K× ; now hG · ai is easily seen to be a quadratic module, hence by irreducibility, a = 0. A rank argument then proves the claim. ♦ 3.5.6 — Claim. For a ∈ Z1 , one has ∂β(a) = −3a. Proof. Return to the decomposition obtained in Claim 3.5.4: letting β = β1 , one has uwa = wa + wβ(a) + β(a) − a. Then: uwuwa = uw(wa + wβ(a) + β(a) − a) = −a − β(a) − ∂β(a) + wβ(a) + ∂wβ(a) − wa − wβ(a) − β(a) + a = (uw)−1 a = −wu−1 a = −wa So that ∂wβ(a) = 2β(a) + ∂β(a). Now let q = ∂β(a) ∈ Z1 . Then: uwuwβ(a) = uw(wβ(a) + 2β(a) + q) = −β(a) − q + 2wβ(a) + 4β(a) + 2q + wq + wβ(q) + β(q) − q = wq + wβ(q + 2a) + β(q + 3a) −1 = (uw) β(a) = −w(β(a) − q) = wq − wβ(a) This forces β(q + 3a) = 0, so by Claim 3.5.5, q = −3a, as desired. ♦ 3.5.7 — Claim. There is a K-vector space structure on V for which V ≃ Sym3 Nat G. Proof. At this stage we could rely on Fact H to determine the structure of V as a G1 -module, where G1 = SL2 (Fp ). The argument would be as follows. First, we compute the descending coherence degree κ(V ) in the sense of [8, §5]; more precisely we show κ(V ) = 1, viz. ker ∂ = Z1 . One inclusion is clear; for the other, let v ∈ ker ∂. In view of Claim 3.5.2, we may suppose v = wa + wb + c for a ∈ Z1 and (b, c) ∈ Ž22 . It is not hard to see that ∂wb = −β(∂b) + ∂b, so that 0 = ∂v = wβ(a) + β(a) − a − β(∂b) + ∂b + ∂c. Projecting onto Ž3 , one has β(a) = 0 whence a = 0 by Claim 3.5.5; then projecting onto Ž2 , β(∂b) = 0 whence ∂b = 0; finally ∂c = 0. But since β : Z1 → Ž2 is onto by Claim 3.5.6, one sees that β(∂b) = −3b = 0, so b = c = 0, and v = 0. This proves ker ∂ = Z1 . Now by Fact H, more specifically [9, §1.2.8] (“n = 4, i = −1”) there, as a G1 -module, V has a composition series (in the classical sense of the term) 0 ≤ V1 ≤ V2 ≤ V with V1 and V2 /V1 isomorphic to sums of copies of Nat G1 , and V /V2 a sum of copies of Sym3 Nat G1 . However, the quadratic radical over G1 is G1 -generated by ker ∂ ∩ w · ker ∂ 2 = Z1 ∩ w · Z2 , and by Lemma 2.6, the latter is contained in CV (T ) = 0, so V2 = 0 and V ≃ V /V2 has the desired structure under G1 . As a matter of fact we do not need [9, Theorem 1] since we already know that for a ∈ Z1 , hold uwa = wa + wβ(a) + β(a) − a and ∂β(a) = −3a: from there and the decomposition of Claim 3.5.2, constructing a G1 -isomorphism with Sym3 Nat G1 is actually immediate. In order to recover a K-linear structure, we now use Fact I (more accurately, [9, Theorem 2] and the remark following it to justify working in characteristic 6= 2, 3). ♦ This completes the proof of Proposition 3.5. 17 3.5 The Tensor Analysis And now let us do the twist. Proposition 3.6. If V as in Notation 3.2 has rank > 3k and U -length ≤ 3, then there are a definable K-vector space structure and a definable automorphism χ of K such that V ≃ Nat G ⊗ χ (Nat G) (twist-and-tensor). Before the proof begins it could be useful to work out the details of the target module. Example 3.7. Let ϕ, ψ ∈ Aut(K). Let K2 have basis:     1 0 e1 = , e2 = 0 1 Then in the twisted representation ϕ (Nat G), one has:    uλ · e1 tλ · e1 = ϕ(λ)e1 w · e1 = −e2 , , uλ · e2 tλ · e2 = ϕ(λ)−1 e2 w · e2 = e1 Now ϕ (Nat G) ⊗ ψ (Nat G) by:  w · e1 ⊗ e1    w · e1 ⊗ e2 w · e2 ⊗ e1    w · e2 ⊗ e2  uλ · e1 ⊗ e1    uλ · e1 ⊗ e2  uλ · e2 ⊗ e1   uλ · e2 ⊗ e2 = e1 = ϕ(λ)e1 + e2 has basis (e1 ⊗ e1 , e1 ⊗ e2 , e2 ⊗ e1 , e2 ⊗ e2 ), and the action there is given = = = = = = = =  e2 ⊗ e2 tλ · e 1 ⊗ e 1    −e2 ⊗ e1 tλ · e 1 ⊗ e 2 , −e1 ⊗ e2  t λ · e2 ⊗ e1   e1 ⊗ e1 tλ · e 2 ⊗ e 2 = = = = ϕ(λ)ψ(λ)e1 ⊗ e1 ϕ(λ)ψ(λ)−1 e1 ⊗ e2 , ϕ(λ)−1 ψ(λ)e2 ⊗ e1 −1 −1 ϕ(λ) ψ(λ) e2 ⊗ e2 e1 ⊗ e1 ψ(λ)e1 ⊗ e1 + e1 ⊗ e2 ϕ(λ)e1 ⊗ e1 + e2 ⊗ e1 ϕ(λ)ψ(λ)e1 ⊗ e1 + ϕ(λ)e1 ⊗ e2 + ψ(λ)e2 ⊗ e1 + e2 ⊗ e2 The reader should keep these in mind during the forthcoming proof. Proof of Proposition 3.6. 3.6.1 — Claim. The central involution acts as 1. Proof. The value of the central involution could be read from [9]; we shall argue more directly. Suppose i = −1 and let a1 ∈ Z1 . Let b2 = ∂wa1 ∈ [U, V ] ≤ Z2 . Then: uwuw · a1 = uw(wa1 + b2 ) = −a1 + wb2 + ∂wb2 = (uw)−1 · a1 = −wa1 This suggests to write b2 as q2 − a1 , with q2 ∈ Z2 . Hence: −wa1 = −a1 + wq2 − wa1 + ∂wq2 − (q2 − a1 ) and therefore uwq2 = q2 ; in particular wq2 ∈ Z2 . Now let r1 = ∂q2 ∈ Z1 ; observe how q2 = (uw)−1 q2 = −w(q2 − r1 ) = wr1 − wq2 , so wr1 = (1 + w)q2 ∈ w · Z1 ∩ Z2 . The latter is finite by Lemma 2.6; since 1 + w is a bijection, there are finitely many possible values for q2 . But the map a1 7→ q2 is additive from the connected group Z1 , so q2 = 0. This means that for any a1 ∈ Z1 one has ∂wa1 = −a1 ∈ Z1 , and therefore ∂ 2 w · Z1 = 0. We argue as in the end of Claim 3.5.3 of Proposition 3.5 to reach a contradiction. ♦ 3.6.2 — Claim. V = Z1 ⊕ L ⊕ w · L ⊕ w · Z1 , where Z1 and L are non-trivial T -lines, and L ≤ Z2 . Proof. We shall first show that Z2 ∩w·Z2 is infinite by an argument similar to Claim 3.6.1. Suppose Z2 ∩ w · Z2 is finite. Then for a1 ∈ Z1 , letting b2 = ∂wa1 ∈ Z2 and c2 = ∂wb2 ∈ Z2 : uwuwa1 = a1 + wb2 + c2 = wa1 so a1 − b2 ∈ Z2 ∩ w · Z2 which we assumed finite. Here again, a1 7→ a1 − b2 is additive, so b2 = a1 ; then uwa1 = wa1 + a1 , implying a1 = (uw)3 a1 = 3a1 + 2wa1 , and since the characteristic is not 2, 18 wa1 = −a1 ∈ Z1 ∩ w · Z1 which is finite by Lemma 2.6, against Z1 being infinite. So Z2 ∩ w · Z2 is infinite. Now suppose that CV (T ) 6= 0. Notice that this reduces the number of non-trivial T -lines to exactly 2. Since T does not centralise Y3 by Lemma 2.7, not both these lines are in Z2 ; so by Lemma 2.5 there is exactly one non-trivial T -line contained in Z2 , call it L. Still assuming CV (T ) 6= 0, suppose in addition L ≤ Z1 . Then Y2 is centralised by T ; by Lemma 2.5, CY3 (T ) = 0. But since w · Z1 embeds into Y3 (as a T -module and modulo a finite kernel), we find CZ1 (T ) = 0, that is, Z1 = L. On the other hand, rk CV (T ) = rk V − 2k > k. The map ∂ : CV (T ) → V has image ∂(CV (T )) ≤ [U, CV (T )] ≤ [T, V ] by Lemma 2.4, but also im ∂ ≤ [U, V ] ≤ Z2 , so im ∂ ≤ ([T, V ] ∩ Z2 )◦ = L of rank exactly k, and ∂ : CV (T ) → V has infinite kernel: there are infinitely many c ∈ CV (T ) with [u, c] = 0. Applying T , any such c is actually centralised by U : hence (CV (T ) ∩ Z1 )◦ 6= 0, contradicting Z1 = L, a non-trivial T -line. So, under the assumption CV (T ) 6= 0, the only non-trivial T -line in Z2 is not contained in Z1 : hence T centralises Z1 . By Lemma 2.5 again, CY2 (T ) = 0, and this shows Z2 = Z1 ⊕ L. So if an element v lies in Z2 ∩ w · Z2 , it has decompositions v = z + ℓ = wz ′ + ℓ′ with z, z ′ ∈ Z1 and ℓ, ℓ′ ∈ L. Now ℓ − wℓ′ ∈ CV (T ) ∩ [T, V ] = 0, and ℓ = wℓ′ ∈ L ∩ w · L = 0. Hence v = z = wz ′ ∈ Z1 ∩ w · Z1 which is finite by Lemma 2.6: we just contradicted Z2 ∩ w · Z2 being infinite. As a conclusion, CV (T ) = 0 and V is therefore the direct sum of non-trivial T -lines, the number of which can only be 4. But remember that Ž2 = (Z2 ∩ w · Z2 )◦ is infinite. Being a proper hT, wi-submodule of V , it is the sum of exactly two non-trivial T -lines: Ž2 = L⊕w ·L. Notice that Z1 ∩ Ž2 ≤ Z1 ∩w ·Z2 which is finite by Lemma 2.6, hence trivial since CV (T ) = 0. This way we reach V = Z1 ⊕L⊕w·L⊕w·Z1 . ♦ 3.6.3 — Claim. For λ ∈ K× there is an abelian group isomorphism βλ : Z1 ≃ L such that for any a ∈ Z1 : uλ wa = wa − tλ wβλ (a) + βλ (a) + tλ a As a matter of fact, βλ2 (a) = tλ β1 (tλ a). Finally, β = β1 : Z1 → L and ∂ : L → Z1 are converse bijections. Proof. This is very much like Claim 3.5.4 in the proof of Proposition 3.5. Fix a ∈ Z1 . Given the decomposition of Claim 3.6.2, one should expect additive maps α : Z1 → Z1 and β, γ : Z1 → L such that: uwa = wa + wγ(a) + β(a) + α(a) Bearing in mind that (uw)3 = 1, one finds: uwuw · a = u(a + γ(a) + wβ(a) + wα(a)) = a + γ(a) + ∂γ(a) + wβ(a) + ∂wβ(a) + wα(a) + wγ(α(a)) + β(α(a)) + α(α(a)) −1 = (uw) a = wu−1 a = wa (1) Projecting onto w · Z1 , this yields α(a) = a; projecting onto w · L, one finds γ(a) = −β(a). Moving from β1 = β to βλ is exactly as in Claim 3.5.4 from the proof of Proposition 3.5 Additivity of βλ is obvious; we did not prove that βλ is an isomorphism, which will be a consequence of the next. ♦ 3.6.4 — Claim. One has ∂λ βλ (a) = tλ a and ∂λ−1 wβλ (a) = −a. For b ∈ L, one has ∂λ wb = −∂λ tλ b. Proof. Equation (1) from Claim 3.6.3 now rewrites as ∂wβ(a) − ∂β(a) + 2a = 0. Let r = ∂β(a) ∈ [U, L] ≤ Z1 , then compute again: uwuwβ(a) = uw(wβ(a) + r − 2a) = β(a) + r + w(r − 2a) − wβ(r − 2a) + β(r − 2a) + (r − 2a) =w −1 −1 u β(a) = wβ(a) − wr 19 so projecting onto w·Z1 we find 2r = 2a; the characteristic not being 2, this results in a = r = ∂β(a) and ∂wβ(a) = −a. Incidently, this shows that the additive map β : Z1 → L is injective, and therefore surjective as well. The inverse map is clearly ∂ : L → Z1 . −1 Now for λ ∈ K, one has ∂λ βλ (a) = tℓ ∂t−1 ℓ tℓ β(tℓ a) = tλ a and ∂λ−1 wβλ (a) = tℓ ∂tℓ wtℓ β(tℓ a) = −a. Finally let b ∈ L; since βλ−1 : Z1 → L is a bijection, there is a ∈ Z1 with b = βλ−1 (a). Hence: ∂λ wb = ∂λ wβλ−1 (a) = −a = −tλ tλ−1 a = −tλ ∂λ−1 βλ−1 a = −∂λ tλ b ♦ 3.6.5 — Claim. In End(L), β∂λ + β∂µ = β∂λ+µ and (β∂λ )(β∂µ ) = β∂λµ for any (λ, µ) ∈ K2 . Proof. This will be tedious; the author strongly suspects that the second half of the argument could be made shorter with more culture on Jordan rings; incidently, it is not surprising to see Jordan rings appear in a cubic action (see the preprint [11] with the proviso that our setting smoothes many arguments). Let ψλ = β∂λ : L → L and ψ : K → End(L) map λ to ψλ . In general one has ∂λ+µ = ∂λ +∂µ +∂λ ∂µ ; here, since L ≤ Z2 , ψ is additive. What we must now prove is that ψ is multiplicative as well. Fix λ and µ ∈ K× . Then in view of Claims 3.6.3 and 3.6.4:  uλ uµ wa = uλ wa − wt−1 µ βµ (a) + βµ (a) + tµ a = wa − wt−1 λ βλ (a) + βλ (a) + tλ a −1 − wt−1 µ βµ (a) − ∂λ wtµ βµ (a) + βµ (a) + ∂λ βµ (a) + tµ a = uλ+µ wa = wa − wt−1 λ+µ βλ+µ (a) + βλ+µ (a) + tλ+µ a so projecting on the various summands one finds: −1 −1 t−1 λ βλ (a) + tµ βµ (a) = tλ+µ βλ+µ (a); βλ (a) + βµ (a) = βλ+µ (a); tλ a − ∂λ wt−1 µ βµ (a) + ∂λ βµ (a) + tµ a = tλ+µ a (2) But we also know from Claim 3.6.4 that tλ+µ a = ∂λ+µ βλ+µ (a); hence using im βλ = L ≤ Z2 : tλ+µ a = ∂λ+µ βλ+µ (a) = ∂λ βλ (a) + ∂λ βµ (a) + ∂µ βλ (a) + ∂µ βµ (a) = tλ a + ∂λ βµ (a) + ∂µ βλ (a) + tµ a −1 Putting this with Equation (2), one gets ∂µ βλ (a) = −∂λ wt−1 µ βµ (a), which equals ∂λ tλ tµ βµ (a) by Claim 3.6.4. Hence: ∂µ βλ (a) = ∂λ tλµ−1 βµ (a) (3) This was unpleasant but the rest of the argument goes slightly better, although we suspect there should be something shorter. We make a series of claims. 1. tλ = ∂λ βλ (as functions on Z1 ). This is from Claim 3.6.4. 2. If ℓ2 = λ then βλ ∂ = tℓ ψλ tℓ . −1 Recall from Claim 3.6.3 that tℓ β = βλ t−1 ℓ . Therefore tℓ ψλ tℓ = tℓ β∂λ tℓ = βλ tℓ ∂λ tℓ = βλ ∂. 3. If ℓ2 = λ and m2 = µ then ψµ tℓ ψλ tℓ = ψλ tλm−1 ψµ tm . This reduces to the previous claim and Equation (3) composed with β on the left and ∂ on the right: ψµ tℓ ψλ tℓ = β∂µ βλ ∂ = β∂λ tλµ−1 βµ ∂ = ψλ tλµ−1 tm ψµ tm . 4. If ℓ2 = λ then tℓ ψλ = ψλ tℓ . Set ℓ = 1 in the previous, and find ψµ = t−1 m ψµ tm . 20 5. ψµ tλ ψλ = ψλ tλ ψµ . With the last two items, ψµ ψλ tλ = ψλ tλ ψµ ; now ψλ and tλ commute. 6. ψµλ2 = ψλ ψµ ψλ . This is since ψµλ2 = β∂µλ2 = βtλ ∂µ t−1 λ , so with items 1 (which applies since ∂µ (L) = Z1 ) and 2, ψµλ2 = β∂λ βλ ∂µ t−1 = ψλ βλ ∂β∂µ t−1 = ψλ tℓ ψλ tℓ ψµ t−1 λ λ λ ; in view of the previous −1 2 commutation properties, one gets ψµλ2 = ψλ tλ ψµ t−1 λ = ψλ ψµ tλ ψλ tλ = ψλ ψµ ψλ . Whether one can conclude at this early stage is not clear to us. Hua’s famous theorem [12] on (anti-)automorphisms of skewfields would require End(L), or at least a subring, to be a division ring (which is not obvious even taking T -minimality into account); exploiting the identity ψλ2 = ψλ2 would require it to be commutative. Perhaps there is an algebraic argument, which we replace by basic model-theory. 7. For any polynomial P ∈ Fp [X], ψP (λ) = P (ψλ ). First show ψλn = ψλn by setting µ = λn in the previous, then use additivity of ψ. 8. For any two (λ, µ) ∈ Fp , ψλ ψµ = ψµ ψλ . This is because there are ν ∈ Fp and polynomials P, Q ∈ Fp [X] with λ = P (ν) and µ = Q(ν). 9. For any two (λ, µ) ∈ K, ψλ ψµ = ψµ ψλ . For λ ∈ Fp , Cλ = {µ ∈ K : ψλ ψµ = ψµ ψλ } is an infinite, definable field, whence equal to K. So for λ ∈ K, Cλ is an infinite, definable field, whence equal to K. 10. ψ is a ring homomorphism. This follows from ψλ ψµ = ψµ ψλ and ψλ2 = ψλ2 . This completes the proof of Claim 3.6.5. ♦ Fortunately the rest is now just painting from nature. 3.6.6 — Notation. We define an action of K  λ·b    λ · wb λ · ∂b    λ · w∂b on V by letting, for (λ, b) ∈ K × L: = = = = β∂λ b w(λ · b) ∂(λ · b) w(λ · ∂b) 3.6.7 — Claim. This defines a K-vector space structure on V for which G is linear. Proof. First bear in mind that since ∂ realises a bijection L ≃ Z1 , and since V = Z1 ⊕L⊕w·L⊕w·Z1, this is well-defined. Additivity in the module element is obvious. Additivity and multiplicativity in λ need only be proved on L, and are exactly Claim 3.6.5. It remains to prove that G is linear. By construction, w is. We turn to linearity of u, which amounts to that of ∂. The latter is obvious on Z1 and L. Now if wb ∈ w · L then by Claim 3.6.4, ∂w(λ · b) = −∂(λ · b) = −λ · ∂b = λ · ∂wb and ∂ is also linear on w · L. Notice that β : Z1 → L is linear. For if a ∈ Z1 , then there is b ∈ L with a = ∂b or equivalently b = β(a), so that: β(λ · a) = β(λ · ∂b) = β(∂(λ · b)) = λ · b = λ · β(a) Finally, for a ∈ Z1 : ∂(λ · wa) = ∂w(λ · a) = −wβ(λ · a) + β(λ · a) + (λ · a) = −wλ · β(a) + λ · β(a) + λ · a = −λ · wβ(a) + λ · β(a) + λ · a = λ · ∂wa 21 So far we proved linearity of hw, ui; now to linearity of T . For b ∈ L one has, thanks to Claim 3.6.5: tµ (λ · b) = β∂tµ (λ · b) = βtµ ∂µ−2 β∂λ b = βtµ ∂β∂µ−2 β∂λ b = βtµ ∂β∂λµ−2 b = βtµ ∂λµ−2 b = β∂λ tµ b = λ · (tµ b) This is linearity of T on L; linearity on w · L follows at once. Now for a ∈ Z1 , letting b = β(a), one has λ · a = ∂(λ · b) = ∂λ b = ∂λ β(a), so: tµ λ · a = tµ ∂λ β(a) = ∂β∂λµ2 tµ β(a) = ∂β∂λ β∂µ2 tµ β(a) = ∂λ βtµ ∂β(a) = λ · (tµ a) This is linearity of T on Z1 ; linearity on w · Z1 follows, proving global linearity of hw, u, T i = G. ♦ By linearity, G ≃ PSL2 (K) embeds definably into GL4 (K); Fact D we quoted in §1.4 guarantees definability of G in the pure field K together with a finite number of automorphisms. This is certainly consistent with the desired identification but does not settle matters. Another remarkable information is the use of the Borel-Tits positive answer [2, Corollaire 10.4] to Steinberg’s [15, Conjecture 1.2]. We know at this stage that V either is the rational module of Proposition 3.5 (which we already ruled out) or a tensor product Nat G ⊗ χ (Nat G) for some field automorphism χ. The problem is that we also announced definability of χ: and for the moment, we see no better argument than retrieving it by a direct computation. There ought to be something more general here and we shall return to the topic some day. 3.6.8 — Notation. Let b0 ∈ L \ {0} be fixed. For λ ∈ K let χ(λ) be the unique µ ∈ K with ∂µ b0 + ∂λ wb0 = 0. 3.6.9 — Claim. The map χ is a well-defined, definable automorphism of K not depending on b0 ; finally, V ≃ χ (Nat G) ⊗ Nat G. Proof. First notice that the function U → Z1 mapping uλ to ∂λ b0 is a definable bijection. So χ is a well-defined, definable function on K. Additivity is obvious since L ⊕ w · L ≤ Z2 . Now remember from Claim 3.6.4 that for b ∈ L, one has ∂λ wb = −∂λ tλ b, so that: χ(λ) · b0 = β∂χ(λ) b0 = −β∂λ wb0 = β∂λ tλ b0 = λ · tλ b0 In particular, by linearity of T : χ(λµ) · b0 = (λµ) · tλµ b0 = λ · (µ · tλ tµ b0 ) = λ · (tλ (µ · tµ b0 )) = χ(λ) · (χ(µ) · b0 ) This proves multiplicativity of χ, which is an automorphism of K. Incidently, if one were to take b1 = t · b0 for some t ∈ T , one would find:  β ∂χ(λ) b1 + ∂λ wb1 = χ(λ) · b1 − β∂λ tλ b1 = χ(λ) · tb0 − λ · tλ tb0 = t(χ(λ) · b0 − λ · tλ b0 ) = 0 But β is injective, so this means that the function χ does not depend on the precise choice of b0 in T · b0 = L \ {0}. It remains to see that we have a definable isomorphism V ≃ χ (Nat G) ⊗ Nat G. The latter has basis (e1 ⊗ e1 , e1 ⊗ e2 , e2 ⊗ e1 , e2 ⊗ e2 ) in the notations of Example 3.7; the reader should take ϕ = χ and ψ = Id there. Let a0 = ∂b0 ∈ Z1 , so that b0 = β(a0 ); clearly (a0 , b0 , w · b0 , w · a0 ) forms a basis of V as a K-vector space. Now consider the map:  7→ e1 ⊗ e1   a0  b0 7→ e1 ⊗ e2 f: w · b0 7→ −e2 ⊗ e1    w · a0 7→ e2 ⊗ e2 22 and extend linearly. We claim that f is G-covariant. The action of w and u is obviously preserved. So it suffices to study the action of T . Remember that tλ b0 = χ(λ) λ · b0 . Hence tλ a0 = tλ ∂b0 = χ(λ) ∂λ2 tλ b0 = λ ∂β∂λ2 b0 = λχ(λ)a0 . Therefore: f (tλ a0 ) = λχ(λ)f (a0 ) = λχ(λ)e1 ⊗ e1 = tλ f (a0 ) whereas: f (tλ b0 ) = χ(λ) χ(λ) f (b0 ) = e1 ⊗ e2 = tλ f (b0 ) λ λ Covariance on wa0 and wb0 follows. ♦ This concludes the proof of Proposition 3.6 and the rank 4k analysis. Remarks. • In order to convince the reader that despite our efforts matters are not perfectly clear, we do not know if one can classify cubic SL2 (K)-modules of finite Morley rank, i.e. those satisfying ℓU (V ) = 3. • Several years ago the author had tried something without model theory: there are many obstacles. First, the decomposition of Claim 3.6.2 cannot be granted. Second, it is not clear that injectivity of β entails its surjectivity in Claim 3.6.4. Also, but this may be due to our lack of algebraic culture, multiplicativity of ψ in Claim 3.6.5 apparently holds only over the algebraic closure of the prime field (to be more precise, the argument works over any quadratically closed, algebraic extension of Fp ). Last but not least, in order to “untensor” definably, one needs to construct χ as in Notation 3.6.8, which requires λ 7→ ∂λ b0 to be a bijection K ≃ Z1 ; the latter is not clear without rank arguments. This is the reason why we finally prefered to work in finite Morley rank. One’s first love is not easily forgotten. References [1] Ayşe Berkman and Alexandre Borovik. Groups of finite Morley rank with a pseudoreflection action. J. Algebra, 368:237–250, 2012. [2] Armand Borel and Jacques Tits. Homomorphismes “abstraits” de groupes algébriques simples. Ann. of Math. (2), 97:499–571, 1973. [3] Alexandre Borovik and Adrien Deloro. Rank 3 Bingo. J. Symbolic Logic, 2015. To appear. [4] Alexandre Borovik and Ali Nesin. Groups of finite Morley rank, volume 26 of Oxford Logic Guides. The Clarendon Press - Oxford University Press, New York, 1994. Oxford Science Publications. [5] Gregory Cherlin and Adrien Deloro. Small representations of SL2 in the finite Morley rank category. Journal of Symbolic Logic, 77(3):919–933, 2012. [6] Edward Cline, Brian Parshall, and Leonard Scott. On the tensor product theorem for algebraic groups. J. Algebra, 63(1):264–267, 1980. [7] Adrien Deloro. Actions of groups of finite Morley rank on small abelian groups. Bull. Symb. Log., 15(1):70–90, 2009. [8] Adrien Deloro. Symmetric powers of Nat sl2 (K). Communications in Algebra, 2015. To appear. [9] Adrien Deloro. Symmetric powers of Nat SL2 (K). Journal of Group Theory, 2015. To appear. 23 [10] Adrien Deloro and Éric Jaligot. Involutive automorphisms of N◦◦ -groups. Submitted, 2015. [11] Matthias Grüninger. On cubic action of a rank one group. Preprint. arXiv:1106.2310v2, 2011. [12] Loo-Keng Hua. On the automorphisms of a sfield. Proc. Nat. Acad. Sci. U. S. A., 35:386–389, 1949. [13] Dugald Macpherson, Alex Mosley, and Katrin Tent. Permutation groups in o-minimal structures. 62(2):650–670, 2000. [14] Bruno Poizat. Quelques modestes remarques à propos d’une conséquence inattendue d’un résultat surprenant de Monsieur Frank Olaf Wagner. J. Symbolic Logic, 66(4):1637–1646, 2001. [15] Robert Steinberg. Representations of algebraic groups. Nagoya Math. J., 22:33–56, 1963. [16] Franz Georg Timmesfeld. Abstract Root Subgroups and Simple Groups of Lie-type, volume 95 of Monographs in Mathematics. Birkhäuser, Basel, 2001. [17] Jules Tindzogho Ntsiri. Étude de quelques liens entre les groupes de rang de Morley fini et les groupes algébriques linéaires. PhD thesis, Université de Poitiers, Poitiers, 2013. [18] Jules Tindzogho Ntsiri. Structure of an SL2 -module of finite Morley rank. Submitted, 2016. [19] Frank Wagner. Fields of finite Morley rank. J. Symbolic Logic, 66(2):703–706, 2001. 24
4math.GR
SIMPLE DUAL BRAIDS, NONCROSSING PARTITIONS AND MIKADO BRAIDS OF TYPE Dn arXiv:1703.09599v2 [math.GR] 28 Jul 2017 BARBARA BAUMEISTER AND THOMAS GOBET Abstract. We show that the simple elements of the dual Garside structure of an Artin group of type Dn are Mikado braids, giving a positive answer to a conjecture of Digne and the second author. To this end, we use an embedding of the Artin group of type Dn in a suitable quotient of an Artin group of type Bn noticed by Allcock, of which we give a simple algebraic proof here. This allows one to give a characterization of the Mikado braids of type Dn in terms of those of type Bn and also to describe them topologically. Using this topological representation and Athanasiadis and Reiner’s model for noncrossing partitions of type Dn which can be used to represent the simple elements, we deduce the above mentioned conjecture. AMS 2010 Mathematics Classification: : 20F36, 20F55. Keywords. Coxeter groups, Artin-Tits groups, dual braid monoids, Garside theory, noncrossing partitions. Contents 1. Introduction 2. Artin groups of type Dn inside quotients of Artin groups of type Bn 2.1. Coxeter groups and Artin groups 2.2. Embeddings of Coxeter groups 2.3. Embeddings of Artin groups 3. Mikado braids of type Bn and Dn 3.1. Mikado braids of type Bn eBn 3.2. Mikado braids of type Dn inside A 4. Dual braid monoids 4.1. Noncrossing partitions 4.2. Dual braid monoids 4.3. Standard Coxeter elements in WDn 5. Simple dual braids of type Dn are Mikado braids 5.1. Outline of the proof 5.2. Graphical model for noncrossing partitions 5.3. The diagram Nx and the braid βx 5.4. Simple dual braids are Mikado braids References The second author was partially funded by the ANR Geolie ANR-15-CE40-0012. 2 3 3 3 4 6 6 7 7 7 7 8 9 9 9 11 17 18 2 BARBARA BAUMEISTER AND THOMAS GOBET 1. Introduction The dual braid monoid Bc∗ of a Coxeter system (W, S) of spherical type was introduced by Bessis [5] and depends on the choice of a standard Coxeter element c ∈ W (a product of all the elements of S in some order). It is generated by a copy Tc of the set T of reflections of W , that is, elements which are conjugates to elements of S. As a Garside monoid, it embeds into its group of fractions, which was shown by Bessis to be isomorphic to the Artin group AW corresponding to W . Unfortunately, this isomorphism is poorly understood, and the proof of its existence requires a case-by-case argument [5, Fact 2.2.4]. The aim of this note is to study properties of the simple elements Div(c) in Bc∗ viewed inside AW in case W is of type Dn and to show that they are Mikado braids, that is, that they can be represented as a quotient of two positive canonical lifts of elements of W . These braids appeared in work of Dehornoy [13] in type An and in work of Dyer [14] for arbitrary Coxeter systems and have many interesting properties. For example, they satisfy an analogue of Matsumoto’s Lemma in Coxeter groups [14, Section 9]. We refer the reader to [14, Section 9], [12, Section 4] (there the Mikado braids are called rational permutation braids, while the terminology Mikado braids rather refers to braids viewed topologically; it is shown however in [12] that both are equivalent) or [15, Section 3.2] for more on the topic. Another important property is that their images in the IwahoriHecke algebra H(W ) of the Coxeter system (W, S) have positivity properties; let us be more precise. There is a natural group homomorphism a : AW −→ H(W )× . If β ∈ AW is a Mikado braid and if we express its image a(β) in the canonical basis {Cw | w ∈ W } of the Hecke algebra, then the coefficients are Laurent polynomials with positive coefficients (see [12, Section 8]). This is one of the main motivations for studying Mikado braids, and showing that simple dual braids are Mikado braids. This last property was conjectured for an arbitrary Coxeter system (W, S) of spherical type in [12], and shown to hold in all the irreducible types different from Dn [12, Theorems 5.12, 6.6, 7.1]. In the classical types An and Bn , the conjecture is proven using a topological characterization of Mikado braids: it can be seen on any reduced braid diagram (resp. symmetric braid diagram in type Bn ) whether a braid is a Mikado braid or not. The present paper gives topological models for Mikado braids of type Dn , similar to those given in types An and Bn in [12], and solves the above conjecture in the remaining type Dn : Theorem 1.1. Let c be a standard Coxeter element in a Coxeter group (W, S) of type Dn . Then every element of Div(c) is a Mikado braid. As a consequence, every simple dual braid in every spherical type Artin group is a Mikado braid, the reduction to the irreducible case being immediate. Licata and Queffelec recently informed us that they also have a proof of the conjecture in types A, D, E with a different approach using categorification [18]. To prove the conjecture, we proceed as follows. Firstly, we explicitly realize the Artin group ADn of type Dn as an index two subgroup of a quotient of the Artin group ABn of type Bn . The existence of such a realization, which is of independent interest, is not new: it was noticed by Allcock [1, Section 4]. We give a simple proof of it here (Proposition 2.7). This allows to realize elements of ADn topologically by Artin braids. We then characterize Mikado braids of type Dn as the images of those Mikado braids of type Bn which surject onto elements of WDn ⊆ WBn under the canonical map from ABn onto WBn (Theorem 3.3). This implies that Mikado braids of type Dn satisfy a nice topological condition, and gives a model for their study in terms of symmetric SIMPLE DUAL BRAIDS, NONCROSSING PARTITIONS AND MIKADO BRAIDS OF TYPE Dn 3 Artin braids, because elements of ABn can be realized as symmetric Artin braids on 2n strands (see Section 3.1). Using Athanasiadis and Reiner’s graphical model [3] for c-noncrossing partitions of type Dn (which are in canonical bijections with the simple elements Div(c) of Bc∗ ; we denote this bijection by x 7→ xc , where x is a c-noncrossing partition), we attach to every such noncrossing partition x an Artin braid βx of type Bn , whose image in the above mentioned quotient is precisely the element xc ∈ ADn (Section 5). Using the topological characterization of Mikado braids of type Bn from [12], we then prove that βx is a Mikado braid of type Bn (Proposition 5.3), which concludes by the above mentioned characterization of Mikado braids of type Dn (Theorem 5.5). Acknowledgments. We thank Luis Paris for useful discussions with the second author and Jon McCammond for pointing out the reference [20]. 2. Artin groups of type Dn inside quotients of Artin groups of type Bn 2.1. Coxeter groups and Artin groups. This section is devoted to recalling basic facts on Coxeter groups and their Artin groups. We refer the reader to [9, 17] or [8] for more on the topic. A Coxeter system (W, S) is a group W generated by a set S of involutions subject to additional braid relations, that is, relations of the form st · · · = ts · · · for s, t ∈ S, s 6= t. Here st · · · denotes a strictly alternating product of s and t, and the number mst of factors in the left hand side equals the number mts of factors in the right hand side. We have mst ∈ {2, 3, . . . } ∪ {∞}, the case mst = ∞ meaning that there is no relation between s and t. Let ℓ : W → Z≥0 be the length function with respect to the set of generators S. Finite irreducible Coxeter groups are classified in four infinite families of types An , Bn , Dn , I2 (m) and six exceptional groups of types E6 , E7 , E8 , F4 , H3 , H4 . If X is a given type, we denote by (WX , SX ) a Coxeter system of this type. The Artin group AW attached to the Coxeter system (W, S) is generated by a copy S of the elements of S, subject only to the braid relations. This gives rise to a canonical surjection π : AW ։ W induced by s 7→ s. If W has type X, we simply denote AW by AX . The canonical map π has a set-theoretic section W ֒→ AW built as follows: let w = s1 s2 · · · sk be a reduced expression for w, that is, we have si ∈ S for all i = 1, . . . , k and k = ℓ(w). Then the lift s1 s2 · · · sk in AW is independent of the chosen reduced expression, and we therefore denote it by w. This is a consequence of the fact that in every Coxeter group, one can pass from any reduced expression of a fixed element w to any other just by applying a sequence of braid relations. The element w is the canonical positive lift of w. 2.2. Embeddings of Coxeter groups. Let (WBn , SBn ) be a Coxeter system of type Bn . We will identify it with the signed permutations group as follows: let S−n,n be the group of permutations of [−n, n] = {−n, −n + 1, . . . , −1, 1, . . . , n} and define WBn := {w ∈ S−n,n | w(−i) = −w(i), for all i ∈ [−n, n]}. Then setting s0 := (−1, 1) and si = (i, i + 1)(−i, −i − 1) for all i = 1, . . . , n − 1 we get that SBn = {s0 , s1 , . . . , sn−1 } is a simple system for WBn (see [8, Section 8.1]). Let (WDn , SDn ) be a Coxeter group of type Dn . Recall that WDn can be realized as an index two subgroup of WBn as follows: setting t0 = s0 s1 s0 , ti = si for all i = 1, . . . , n−1 we have that SDn := {t0 , t1 , . . . , tn−1 } is a simple system for the Coxeter group WDn = ht0 , t1 , . . . , tn−1 i of type Dn (see [8, Section 8.2]). In the following, a Coxeter group of type Dn will always be viewed inside WBn , with the above identifications. 4 BARBARA BAUMEISTER AND THOMAS GOBET 2.3. Embeddings of Artin groups. We assume the reader to be familiar with Artin groups attached to Coxeter groups and refer to [11, Chapter IX] for basic results. Notice that there are two surjective maps qB : ABn −→ AAn−1 , qD : ADn −→ AAn−1 defined as follows: if we denote by {σ1 , . . . , σn−1 } the set of standard Artin generators of the n-strand Artin braid group AAn−1 , then qB (s0 ) = 1, qB (si ) = σi for i 6= 0, while qD (t0 ) = σ1 , qD (ti ) = σi for all i 6= 0 (see [10, Section 2.1]). Both maps qB and qD are split and one can write AXn ∼ = ker(qX ) ⋊ AAn−1 for X ∈ {B, D}. Crisp and Paris showed that the embedding of WDn in WBn which we recalled in Subsection 2.2 does not come from an embedding ϕ : ADn −→ ABn such that qD = qB ◦ ϕ [10, Proposition 2.6]. In this section we show that there is an embedding of eBn of ABn ; this embedding can be seen as a natural lift of the ADn inside a quotient A embedding of Coxeter groups and has the expected properties (see Lemma 2.8). This is mostly a reformulation of results of Allcock [1, Sections 2 and 4], but we will give a simple algebraic proof of this fact here. eBn to be the quotient of ABn by the smallest normal subgroup Definition 2.1. Define A 2 containing s0 . It follows immediately from this definition that the canonical map πn : ABn ։ WBn eBn via two surjective maps πn,1 : ABn ։ A eBn and πn,2 : A eBn ։ WBn . factors through A Remark 2.2. In [20, Definition 3.3], a similar group, called the middle group, is considered. It is defined as the quotient of ABn by the smallest normal subgroup containing s21 (as a consequence, every s2i for i ≥ 1 is equal to 1 in the quotient since si lies in the same conjugacy class as s1 ). eBn , for all si ∈ SBn . Set Denote by s′i , i = 0, . . . , n − 1 the image of si ∈ ABn in A ′ ′ ′ ′ ′ ′ t0 = s0 s1 s0 and ti = si for i = 1, . . . , n − 1. Lemma 2.3. The elements t′0 , t′1 , . . . t′n−1 satisfy the braid relations of type Dn , that is, we have t′0 t′1 = t′1 t′0 , t′0 t′2 t′0 = t′2 t′0 t′2 , t′i t′i+1 t′i = t′i+1 t′i t′i+1 for all i = 1, . . . , n − 2, t′i t′j = t′j t′i if |i − j| > 1 and {i, j} = 6 {0, 2}. Proof. All the relations except the second one are immediate consequences of the type Bn braid relations satisfied by the s′0 , s′1 , . . . , s′n−1 . For the second relation we have ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ ′ t′0 t′2 t′0 = s′0 s′1 s′0 s′2 s′0 s′1 s′0 = s′0 s′1 s′2 0 s2 s1 s0 = s0 s1 s2 s1 s0 = s0 s2 s1 s2 s0 = s′2 s′0 s′1 s′0 s′2 = t′2 t′0 t′2 .  An immediate corollary is eBn defined by ιn (ti ) = Corollary 2.4. There is a group homomorphism ιn : ADn −→ A ′ ti for all i = 0, . . . , n − 1. We have the following situation Lemma 2.5. There is a commutative diagram ABn π n,1 eBn o //A ❊❊ ❊❊ ❊❊ πn,2 B ❊❊❊  πn ""  WB o n where πnD : ht′0 , . . . , t′n−1 i ′ ′ ? _ ht0 , . . . , tn−1 i oo ADn D πn  _? WDn −→ WDn is defined by πnD (t′i ) = ti for all i = 0, . . . , n − 1. SIMPLE DUAL BRAIDS, NONCROSSING PARTITIONS AND MIKADO BRAIDS OF TYPE Dn 5 Remark 2.6. In Proposition 2.7 below we will show that the map ιn is injective; hence πnD is in fact simply the canonical surjection ADn ։ WDn . eBn factors Proof. We have to show that the composition of πn,2 and ht′0 , . . . , t′n−1 i ֒→ A through WDn . It suffices to show that the image of t′i under this composition is precisely ti (viewed inside WBn via the embedding WDn ֒→ WBn ) for all i = 0, . . . , n − 1, which is immediate.  Proposition 2.7. The homomorphism ιn is injective and ht′0 , . . . , t′n−1 i is a subgroup eBn of index two. Hence ADn can be identified with the subgroup of A eBn generated of A ′ by the ti , i = 0, . . . , n − 1. Proof. We first notice that, as an immediate consequence of Lemma 2.5, the subgroup eBn is proper since WDn is a proper subgroup of WBn . U := ht′0 , . . . , t′n−1 i ⊆ A ′ As s0 interchanges t′0 and s′0 t′0 s′0 = t′1 , and as s′0 commutes with t′i for i = 2, . . . n − 1, the involution s′0 normalizes U and induces on U an automorphism of order 2 (which eBn . is in fact an outer automorphism). Therefore, U = ιn (ADn ) is of index 2 in A Next we determine a presentation of U using the Reidemeister-Schreier algorithm (see for instance [19]). We take as a Schreier-transversal T := {1, s′0 } for the right eBn . This yields the generating set cosets of U in A {ts′i ts′i −1 | t ∈ T and 0 ≤ i ≤ n − 1} = {t′i | 0 ≤ i ≤ n − 1} eBn . Application of this algorithm and where x is the representative of U x in T for x ∈ A of Tietze-transformations (see [19]) then precisely yields the braid relations as stated in Lemma 2.3. This shows that ιn is injective.  eBn with ADn and we set From now on we identify the subgroup ht′0 , t′1 , . . . , t′n−1 i ⊆ A eBn , the map qB factors ti = t′i for all i = 0, . . . , n − 1. Note that by definition of A eBn , giving rise to a surjection qeB : A eBn −→ AA . Then we have through A n−1 Lemma 2.8. The map ιn satisfies qeB ◦ ιn = qD . Proof. We have qD (t0 ) = σ1 and (e qB ◦ ιn )(t0 ) = qeB (s′0 s′1 s′0 ) = qB (s0 )qB (s1 )qB (s0 ) = qB (s1 ) = σ1 . For i ≥ 1 we have qD (ti ) = σi = qB (si ) = qeB (s′i ) = (e qB ◦ ιn )(ti ).  Definition 2.9. Given x ∈ WDn , we denote by xD the canonical positive lift of x in eBn ) and by xB the canonical positive ADn (which we will systematically view inside A lift of x in ABn . Proposition 2.10. Let x ∈ WDn . We have πn,1 (xB ) = xD . Proof. Let ti1 ti2 · · · tik be an SDn -reduced expression of x in WDn . Replacing t0 by s0 s1 s0 and ti by si for i = 1, . . . , n − 1 we get a word in the elements of SBn for x. Note that this may not be a reduced expression for x in WBn . It suffices to show that one can transform the above word into a reduced expression for x in WBn just by applying braid relations of type Bn and the relation s20 = 1. We prove the above statement by induction on k. If k = 1 then the claim holds since ti , i ≥ 1 is replaced by si while t0 is replaced by s0 s1 s0 which is SBn -reduced. Hence assume that k > 1. By induction the claim holds for x′ = ti2 · · · tik . By [8, Propositions 8.1.2, 8.2.2] one has that sj , j ≥ 1 is a left descend of x′ in WBn if and only if it is a left descent of x′ in WDn . Hence we can assume that ti1 = t0 and that it is the only left descent of x in WDn . Firstly, assume that s0 is a left descent of x′ in WBn , hence s0 is not a left descent of s0 x′ . We claim that it suffices to show that s1 is not a left descent of s0 x′ : indeed, 6 BARBARA BAUMEISTER AND THOMAS GOBET it implies that ℓ(s0 s1 s0 x′ ) = ℓ(s0 x′ ) + 2 (where ℓ is the length function in WBn ) by the lifting property (see [8, Corollary 2.2.8(i)]). Moreover by induction we can get every SBn -reduced decomposition of x′ using only the claimed relations, hence we can by induction get a reduced expression for x′ starting with s0 with these relations. The only additional relation to apply to get a reduced decomposition of x is the deletion of the s20 = 1 which appears when appending s0 s1 s0 at the left of such a reduced expression of x′ . Hence assume that s1 s0 x′ < s0 x′ in WBn , i.e., that s1 is a left descent of s0 x′ . By [8, Proposition 8.1.2] it follows that x′−1 s0 (1) > x′−1 s0 (2) which implies that x′−1 (−1) > x′−1 (2), hence −x′−1 (2) > x′−1 (1). But by [8, Proposition 8.2.2] it precisely means that t0 is a left descent of x′ , a contradiction. Now assume that s0 is not a left descent of x′ in WBn . Then s1 is not a left descent of x′ in WBn , otherwise using [8, Proposition 8.1.2] again it would be a left descent of ti1 x′ in WBn , hence in WDn by [8, Proposition 8.2.2], a contradiction. It follows that a reduced expression for y = s1 s0 x′ in WBn is obtained by concatenating s1 s0 at the left of a reduced expression for x′ (which we can obtain by induction). If s0 y > y then we are done, while if s0 y < y then by Matsumoto’s Lemma we can obtain a reduced expression of y starting with s0 just by applying type Bn braid relations. Deleting the s20 at the beginning of the word we then have a reduced expression of x.  Remark 2.11. The fact that reduced expressions of an element x ∈ WDn can be transformed into reduced expressions in WBn as we did in the proof above had been noticed by Hoefsmit in his thesis [16, Section 2.3] without a proof. The fact that ADn eBn also implies that the corresponding Iwahori-Hecke can be realized as a subgroup of A algebra H(WDn ) of type Dn embeds into the two-parameter Iwahori-Hecke algebra H(WBn ) of type Bn where the parameter corresponding to the conjugacy class of s0 is specialized at 1. This is precisely what Hoefsmit uses to study representations of Iwahori-Hecke algebras of type Dn using the representation theory of those algebras in type Bn . 3. Mikado braids of type Bn and Dn 3.1. Mikado braids of type Bn . We recall from [12] the following Definition 3.1. Let (W, S) be a finite Coxeter system with Artin group AW . An element β ∈ AW is a Mikado braid if there exist x, y ∈ W such that β = x−1 y. We denote by Mik(W ) (or Mik(X) if W is of type X) the set of Mikado braids in AW . We briefly recall results from [12, Section 6.2] on topological realizations of Mikado braids in type Bn which will be needed later on. The Artin group ABn embeds into AA2n−1 , which is isomorphic to the Artin braid group on 2n strands. Labeling the strands by −n, . . . , −1, 1, . . . , n, every simple generator in SBn ⊆ Sn,−n is then lifted to an Artin braid as follows. The generator s0 exchanges the strands 1 and −1, while the generator si , i = 1, . . . , n − 1 exchanges the strands i and i + 1 as well as the strands −i and −i − 1 (in both crossings, the strand coming from the right passes over the strand coming from the left, like in the right picture in Figure 3). Those braids in AA2n−1 which are in ABn are precisely those braids which are fixed by the automorphism which exchanges each crossing i, i + 1 by a crossing −i, −i − 1 of the same type, for all i. We call these braids symmetric. There is the following graphical characterization of Mikado braids in ABn Theorem 3.2 ([12, Theorem 6.3]). Let β ∈ ABn . The following are equivalent (1) The braid β is a Mikado braid, that is, there are x, y ∈ WBn such that β = x−1 y. SIMPLE DUAL BRAIDS, NONCROSSING PARTITIONS AND MIKADO BRAIDS OF TYPE Dn 7 (2) There is an Artin braid in AA2n−1 representing β, such that one can inductively remove pairs of symmetric strands, one of the two strands being above all the other strands (so that the symmetric one is under all the other strands). Note that in the second item above, we remove pairs of strands instead of single strands so that at each step of the process, the obtained braid is still symmetric (hence in ABn ). eBn . The aim of this subsection is to prove 3.2. Mikado braids of type Dn inside A the following result, relating Mikado braids of type Dn to Mikado braids of type Bn : eBn are precisely the Theorem 3.3. The Mikado braids of type Dn viewed inside A images of those Mikado braids of type Bn which surject onto elements of WDn , that is, we have Mik(Dn ) = {πn,1 (β) | β ∈ Mik(Bn ) and πnB (β) ∈ WDn }. eBn . Then there exist x, y ∈ WDn such that γ = (xD )−1 yD . Proof. Let γ ∈ Mik(Dn ) ⊆ A Note that by Lemma 2.5 we have πn,2 (γ) = x−1 y ∈ WDn . But by Proposition 2.10 we have γ = πn,1 (β) where β = (xB )−1 yB ∈ Mik(Bn ), which shows the first inclusion. Conversely, let β ∈ Mik(Bn ) such that πnB (β) ∈ WDn . We have to show that πn,1 (β) ∈ Mik(Dn ). By definition there are x, y ∈ WBn such that β = (xB )−1 yB . Since πnB (β) = x−1 y ∈ WDn , if either x or y is in WDn then both of them are in WDn in which case we are done by Proposition 2.10. Hence assume that x, y ∈ / WDn . Since WDn is a subgroup of WBn of index two and s0 ∈ / WDn there are x′ , y ′ ∈ WDn such that ′ B (the exponent depending on whether x = s0 x′ , y = s0 y ′ . If follows that xB = s±1 0 x B ′ e s0 x > x or not) and yB = s±1 0 y . Hence since the image of s0 in ABn has order two, D −1 ′ D ′ using Proposition 2.10 again we have πn,1 (β) = (x ) y which concludes.  4. Dual braid monoids 4.1. Noncrossing partitions. Let (W, S) be a Coxeter system of spherical type. Let S T = w∈W wSw−1 denote the set of reflections in W and ℓT : W −→ Z≥0 the corresponding length function. A standard Coxeter element in (W, S) is a product of all the elements of S. Given u, v ∈ W , we can define a partial order ≤T on W by u ≤T v ⇔ ℓT (u) + ℓT (u−1 v) = ℓT (v). In this case we say that u is a prefix of v. Let c be a standard Coxeter element. The set NC(W, c) of c-noncrossing partitions consists of all the x ∈ W such that x ≤T c. The poset (NC(W, c), ≤T ) is a lattice, isomorphic to the lattice of noncrossing partitions when W = WAn ∼ = Sn+1 . See [2] for more on the topic. Remark 4.1. There are several (unequivalent) definitions of Coxeter elements (see for instance [4, Section 2.2]). The above definitions still make sense for more general Coxeter elements, but for the realization of the dual braid monoids (which are introduced in the next section) inside Artin groups the Coxeter element is required to be standard (see [12, Remark 5.11]). 4.2. Dual braid monoids. We recall the definition and properties of dual braid monoids. For a detailed introduction to the topic the reader is referred to [5, 12] or [11]. Dual braid monoids were introduced by Bessis [5], generalizing definitions of Birman, Ko and Lee [7] and Bessis, Digne and Michel [6] to all the spherical types. Let (W, S) be a finite Coxeter system. Denote by T the set of reflections in W and by 8 BARBARA BAUMEISTER AND THOMAS GOBET AW the corresponding Artin-Tits group. Let c be a standard Coxeter element in W . Bessis defined the dual braid monoid attached to the triple (W, T, c) as follows. Take as generating set a copy Tc := {tc | t ∈ T } of T and set Bc∗ := htc ∈ Tc | tc ∈ Tc , tc t′c = (tt′ t)c tc if tt′ ≤T ci The defining relations of Bc∗ are called the dual braid relations with respect to c. We mention some properties of Bc∗ , which can be found in [5]. The monoid Bc∗ is infinite and embeds into AW . In fact, Bc∗ is a Garside monoid, hence it embeds into its group of fractions Frac(Bc∗ ) and the word problem in Frac(Bc∗ ) is solvable. Bessis showed that Frac(Bc∗ ) is isomorphic to AW , but his proof requires a case-by-case analysis (see [5, Fact 2.2.4]) and the isomorphism is difficult to understand explicitly. More precisely, the embedding Bc∗ ⊆ AW sends sc to s for every s ∈ S. In [12, Proposition 3.13], a formula for the elements of Tc (which are the atoms of the monoid Bc∗ ) as products of the Artin generators is given, but it does not give in general a braid word of shortest possible length. Example 4.2. Let (W, S) be of type A2 and c ∈ W be the Coxeter element s1 s2 where si = (i, i + 1). Then we have the dual braid relation (s1 )c (s2 )c = (s1 s2 s1 )c (s1 )c . Hence inside AW , the atom (s1 s2 s1 )c corresponding to the non-simple reflection s1 s2 s1 is equal to s1 s2 s−1 1 . As every Garside monoid, Bc∗ has a finite set of simple elements, which form a lattice under left divisibility. They are defined as follows. For x ∈ NC(W, c), let x = t1 t2 · · · tk be a T -reduced expression of x, that is, a reduced expression as product of reflections. Then Bessis showed that the element xc := (t1 )c (t2 )c · · · (tk )c ∈ Bc∗ is independent of the choice of the reduced expression of x and therefore well-defined as a consequence of a dual Matsumoto property [5, Section 1.6]. The Garside element is the lift cc of c and the set Div(c) of simple elements (that is, of (left) divisors of cc ) is given by Div(c) := {xc | x ∈ NC(W, c)}. There is an isomorphism of posets (NC(W, c), ≤T ) ∼ = (Div(c), ≤), x 7→ xc , where ≤ is the left-divisibility order in Bc∗ . In general, we are only able to determine the elements of Div(c) as words in the classical Artin generators S of AW by an inductive application of the dual braid relations. It is therefore difficult to study properties of elements of Div(c) viewed inside AW . Note that the composition Bc∗ ֒→ AW ։ W sends every product (t1 )c (t2 )c · · · (tk )c , ti ∈ T to t1 t2 · · · tk . 4.3. Standard Coxeter elements in WDn . In this subsection, we characterize standard Coxeter elements in WDn in terms of signed permutations. This will be needed to introduce graphical representations of c-noncrossing partitions of type Dn in Section 5.2. Recall that WDn ⊆ WBn and that w(−i) = −w(i), for all i ∈ [−n, n] and all w ∈ WBn . In WBn , cycles of the shape (i1 , . . . , ir , −i1 , . . . , −ir ) are abbreviated by [i1 , . . . ir ] and called balanced cycles, and those of type (i1 , . . . , ir )(−i1 , . . . , −ir ) by ((i1 , . . . , ir )) and called paired cycles. The set of reflections in WDn is T := TDn := {(i, j)(−i, −j) | i, j ∈ {−n, . . . , n}, i 6= ±j}, and every w ∈ WDn can be written as a product of disjoint cycles in which there is an even number of balanced cycles (see [3, Section 2]). Lemma 4.3. An element c ∈ WDn (n ≥ 3) is a standard Coxeter element if and only if c = (i1 , −i1 )(i2 , . . . , in , −i2 , . . . , −in ) where {i1 , . . . , in } = {1, 2, 3, . . . , n}, i1 ∈ {1, 2} and the sequence i2 · · · in is first increasing, then decreasing. SIMPLE DUAL BRAIDS, NONCROSSING PARTITIONS AND MIKADO BRAIDS OF TYPE Dn 9 Proof. The proof is by induction on n. The case n = 3 is easy to check by hand. Let c be a standard Coxeter element in WDn , n ≥ 4. Then either sn c or csn is a standard Coxeter element in WDn−1 , in which case induction and a straightforward computation shows that c is of the required form. Conversely if c is of the above form, then since (i1 , −i1 ) commutes with sn either sn c or csn is of the above form in WDn−1 , hence is a standard Coxeter element in WDn−1 , implying that c is a standard Coxeter element in WDn .  Elements in NC(WDn , c) will be described below via a graphical representation. 5. Simple dual braids of type Dn are Mikado braids The aim of this section is to show Theorem 1.1, that is, that simple dual braids of type Dn are Mikado braids. 5.1. Outline of the proof. The proof proceeds as follows. • Step 1. We describe in Section 5.2 a pictural model for the elements x ∈ NC(WDn , c) which is due to Athanasiadis and Reiner [3]. In this model the element x is represented by a diagram consisting of non-intersecting polygons joining labeled points on a circle. The labeling depends on the choice of the standard Coxeter element c, more precisely, we first require to write the Coxeter element as a signed permutation (as in Lemma 4.3). • Step 2. We slightly modify the diagram from Step 1 associated to x ∈ NC(WDn , c) to obtain a new diagram Nx consisting of non-intersecting polygons joining labeled points on a circle. The only difference with the AthanasiadisReiner model is that there is a point with two labels in the latter, which we split in two different points. As we will see, the diagram Nx is not unique in general, but we will show that all the information which we will use from the diagram Nx is independent of the chosen diagram representing x. From this new diagram Nx , we build a topological braid βx lying in an Artin group ABn of type Bn (viewed inside AA2n−1 , hence βx is a symmetric braid on 2n strands). We first explain how to define the diagram Nx for elements of TDn ⊆ NC(WDn , c) and we then do it for all x ∈ NC(WDn , c). eBn , for t ∈ TDn , lie in ADn and • Step 3. We show that the braids πn,1 (βt ) ∈ A satisfy the dual braid relations with respect to c. This will follow from the more general statement that if x ≤T xt ≤T c with t ∈ TDn , then πn,1 (βx )πn,1 (βt ) = πn,1 (βxt ). This property and the fact that πn,1 (βs ) = s for all s ∈ SDn will be enough to conclude that πn,1 (βx ) is equal to the simple dual braid xc for all x ∈ NC(WDn , c) (this is explained in the proof of Corollary 5.2). In particular we also show that πn,1 (βx ) does not depend on the choice of the diagram Nx . • Step 4. We show that the braid βx , x ∈ NC(WDn , c) is a Mikado braid in ABn by using the topological characterization of [12]. Recall that βx is defined graphically, as an Artin braid on 2n strands. Together with Step 3 and Theorem 3.3, it follows that xc = πn,1 (βx ) is a Mikado braid, which proves Theorem 1.1. 5.2. Graphical model for noncrossing partitions. Athanasiadis and Reiner found a graphical model for noncrossing partitions of type Dn . We present it here (with slightly different conventions). First we explain how to label a circle depending on the choice of the standard Coxeter element c. Given a standard Coxeter element c = (i1 , −i1 )(i2 , . . . , in , −i2 , . . . , −in ) in WDn , where the notation is as in Lemma 4.3 and where i2 = −n, we place 2n − 2 points (labeled by i2 , . . . , in , −i2 , . . . , −in ) on a circle as follows: point −n is at the top of the 10 BARBARA BAUMEISTER AND THOMAS GOBET circle while point n is at the bottom. The remaining points all have distinct height depending on their label: if i < j then point i is higher than point j. Moreover, when going along the circle in clockwise order starting at i2 = −n, the points must be met in the order i2 i3 · · · in (−i2 )(−i3 ) · · · (−in ). Finally, we add a point at the center of the circle, labeled by ±i1 . Athanasiadis and Reiner showed that c-noncrossing partitions are those for which there exists a graphical representation as follows (in their description, we have i1 = n; this corresponds to a choice of Coxeter element which is not standard, however by conjugation we can assume it to be standard an to have i1 ∈ {1, 2}. The c-noncrossing partition lattices are isomorphic for all Coxeter elements c). Given x ∈ NC(WDn , c), consider its cycle decomposition inside S−n,n and associate to each cycle the polygon given by the convex hull of the points labeled by elements in the support of the cycle. It results in a noncrossing diagram, i.e., the various obtained polygons do not intersect, with two possible exceptions: if there is a polygon Q of x with i1 ∈ Q, −i1 ∈ / Q, then −Q is also a polygon of x. Thus the two polygons Q and −Q will have the middle point in common (Note that since x is a signed permutation, for every polygon P of x we have that −P is also a polygon of x, possibly with P = −P ). The second case appears when the decomposition of x has a product of factors of the form [j][i1 ] for some j 6= ±i1 . In this case to avoid confusion with the noncrossing representation of the reflection ((j, i1 )) (or ((j, −i1 ))) we have to choose an alternative way of representing this product. Note that the cycle [j] should be considered as a polygon P such that P = −P . By analogy with the situation where there is such a polygon and where the point ±i1 lies inside P , we represent [j] by two curves both joining j to −j and not intersecting except at the points ±j, in such a way that the point ±i1 lies between these two curves. Conversely, to every noncrossing diagram with the above properties, one can associate an element x of NC(WDn , c) as follows: we send each polygon P with labels j1 , j2 , . . . , jk (read in clockwise order) to the cycle (j1 , j2 , . . . , jk ) except in case P = −P . Each single point with label i is sent to the one-cycle (i) except i1 in case there is a polygon P with P = −P (in which case ±i1 lie inside P ). In this last case, if P is labeled by j1 , j2 , . . . , jk then we send it to the product of cycles (i1 , −i1 )(j1 , j2 , . . . , jk ) (like in the middle example of Figure 1). The element x is then the product of all the cycles associated to all the polygons of the noncrossing diagram (note that they are disjoint). Note that when the middle point lies in two different polygons, one has to specify in which polygon the label i1 lies. Examples are given in Figure 1 and we refer to [3] for more details. b b b b b −2 b b 2 b b b b b b b −8 −7 −6 −5 −4 −3 b b −1 1 3 4 5 6 7 8 b b b b b 2 −2 b b b b b b b b −8 −7 −6 −5 −4 −3 b b −1 1 3 4 5 6 7 8 −8 −7 −6 −5 −4 −3 b b b b b b b 2 −2 b b b b b b b Figure 1. Examples of noncrossing diagrams for x1 = ((1, −8))((7, 5, −2)), x2 = ((8, 7, 5))[6, 3, 1][2], x3 = ((6, 3, −4)) ∈ NC(WDn , c). −1 1 3 4 5 6 7 8 SIMPLE DUAL BRAIDS, NONCROSSING PARTITIONS AND MIKADO BRAIDS OF TYPE Dn 11 5.3. The diagram Nx and the braid βx . To define the diagram Nx , we slightly modify the labeling of the circle given in the previous section by splitting the point ±i1 into two points placed on the vertical axis of the circle consistently with their labels (all the points should be placed such that the point i is higher than the point j if i < j). An example is given in Figure 2 and we call this labeling the c-labeling of the circle. The idea is then to start from Athanasiadis and Reiner’s graphical representation of x ∈ NC(WDn , c) and just split the middle point into two points. For convenience we may represent the polygons by curvilinear polygons since in some cases, because of the splitting it might not be possible to have the polygons not intersecting each other. Depending on the situation we will add an edge joining the two points i1 and −i1 : we explain more in details below how to draw the diagrams Nx , first when x is a reflection, then in general. b b b b b b b b b b b b b b b b −8 −7 −6 −5 −4 −3 −2 −1 1 2 3 4 5 6 7 8 Figure 2. Example of a c-labeling in type D8 . Here c = t1 t3 t5 t7 t6 t4 t2 t0 = (2, −2)[−8, −7, −5, −3, −1, 4, 6] and i1 = 2. 5.3.1. Pictures for reflections. Reflections are all of the form t = c1 c2 , where c1 and c2 are two 2-cycles with opposite support. If c1 = (i, j), we will draw a curvilinear “polygon” with two edges both joining i to j. We then orient the polygon in counterclockwise order. We do the same for c2 = (−i, −j) in such a way that the second curvilinear polygon does not intersect the first one. In some cases, there is not a unique way of drawing two such curvilinear polygons with the condition that the resulting diagram should be noncrossing. We explain how to do it in the next paragraph by separating the set of reflections into three classes. Firstly, assume that supp(c1 ) = {i, j} ⊆ {1, . . . , n}, then Nt is drawn as in the left picture of Figure 3. Now assume that supp(c1 ) = {i, −j} with i ∈ {1, . . . , n}\{i1 }, j ∈ {−1, . . . , −n}\{−i1 }. In that case, we draw the two curvilinear polygons in such a way that the two middle points labeled by ±i1 lie between them, as done in Figure 4. The last case is the case where c1 = (i1 , j) with j ∈ {−1, . . . , −n}\{−i1 }. In that case, there are two ways of drawing the curvilinear polygon (see the left pictures of Figure 5). We can choose any of the two pictures for Nt . Starting from such a noncrossing diagram, we then associate an Artin braid βt on 2n strands to it, by first projecting the noncrossing diagram to the right (as done in the left pictures of Figures 3 and 4), i.e., putting all the points on the same vertical line, obtaining a new graph for the noncrossing partition. This new graph can then be viewed as a braid diagram, viewed from the bottom: a curve joining point k to point ℓ corresponds to a k-th strand ending at ℓ, while single points without a curve starting or ending at them correspond to unbraided strands. If a point has nothing at its right (resp. at its left), it means that the corresponding unbraided strand is above all the others (resp. below all the others). The points lying right to (resp. left to) a curve correspond to an unbraided strand lying above (resp. below) the strand corresponding 12 BARBARA BAUMEISTER AND THOMAS GOBET to that curve. See the above mentioned Figures. Note that in the case of Figure 5, the two braids βt obtained from the two different diagrams Nt are distinct in ABn , but their eBn are the same because we can invert the crossings images πn,1 (βt ) in the quotient A corresponding to the generator s0 (because of the relation s20 = 1 which holds in the quotient). b b −8 −7 −6 −5 −4 −3 −2 −1 1 2 3 4 5 6 7 8 b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b Figure 3. The diagram Nt for t = (3, 6)(−3, −6) and the braid βt . b b b b b b b b b b b b b b b b −8 −7 −6 −5 −4 −3 −2 −1 1 2 3 4 5 6 7 8 b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b Figure 4. The diagram Nt for t = (3, −5)(−3, 5) and the braid βt . We now generalize the above picturial process, by associating a (possibly non unique) noncrossing diagram Nx and an Artin braid βx to every x ∈ NC(WDn , c). 5.3.2. Pictures for noncrossing partitions. To obtain a noncrossing diagram Nx with oriented curvilinear polygons from x as we did for reflections in the previous section, we proceed as follows: we orient every polygon of the noncrossing partition in counterclockwise order (note that this is the opposite orientation to the one given by the corresponding cycle of x, that is, an arrow j2 → j1 means that the cycle of x sends j1 to j2 ; hence this orientation corresponds to x−1 ). Polygons reduced to a single edge are replaced by curvilinear polygons with two edges as we did for reflections in Section 5.3.1. Again we split the points with labels ±i1 into two points with labels −i1 and i1 respectively as in Figure 2. In the case where the middle point in the Athanasiadis-Reiner model has no edge starting at it and does not lie inside a symmetric polygon, then the two points i1 and −i1 have no edge starting at them in the new diagram. In the case where there are two distinct polygons P and −P sharing the middle point, they are separated so that each SIMPLE DUAL BRAIDS, NONCROSSING PARTITIONS AND MIKADO BRAIDS OF TYPE Dn 13 b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b −8 −7 −6 −5 −4 −3 −2 −1 1 2 3 4 5 6 7 8 −8 −7 −6 −5 −4 −3 −2 −1 1 2 3 4 5 6 7 8 b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b Figure 5. Two diagrams Nt for t = (2, −7)(−2, 7) and the corresponding Artin braids βt . Note that the two Artin braids on the right are eBn . equal in the quotient A point lies in the correct curvilinear polygon (see Figure 6): there might be several nonisotopic diagrams which work when separating P from −P (in case P is a 2-cycle we precisely get what we already noticed and explained in Figure 5). A similar argument eBn of the various Artin braids to the one given in Figure 5 shows that the images in A βx obtained from the distinct diagrams Nx at the end of the process explained below will be equal. In case there is a symmetric polygon P = −P or a factor [j][i1 ] in x, we add a curvilinear polygon with two edges joining −i1 to i1 , oriented in counterclockwise order (Recall that in the noncrossing representation of [j][i1 ], the factor [j] is already represented by a curvilinear “polygon” with two edges and the point i1 inside it. Here we orient this polygon in counterclockwise order as in all other cases). b b b b b b −2 b b 2 b −8 −7 −6 −5 −4 −3 b −1 1 b b b b b b b b b b b b b b b 3 4 5 6 7 8 b b b b b b −8 −7 −6 −5 −4 −3 −2 −1 1 2 3 4 5 6 7 8 Figure 6. Splitting of two polygons with common middle point. If one has the diagram Nx with oriented curvilinear polygons as in Figure 6 on the right, we proceed exactly as we did for reflections in Section 5.3.1 to obtain βx : firstly, we put all the black points on a vertical line and project the noncrossing diagram to obtain a picture as in the left pictures in Figures 3 and 4; this diagram gives the Artin braid βx viewed from the bottom. We illustrate this process for the noncrossing diagram of the element x2 = ((8, 7, 5))[6, 3, 1][2] of Figure 1 in Figure 7. Note that as a consequence of this procedure, the orientation we put on polygons, which as we 14 BARBARA BAUMEISTER AND THOMAS GOBET already noticed at the beginning of the subsection is not the one corresponding to x but to x−1 , defines the permutation induced by the strands of βx . The fact that the permutation induced by the strands of βx is x−1 rather than x comes from the fact that our convention is to concatenate Artin braids from top to bottom. Note that we can always recover the braid from the middle diagram without ambiguity, because all the strands either strictly go up or down, except possibly in one case: in case i1 = 2 and x = (1, −1)(2, −2) ∈ NC(WDn , c), then the strands joining 1 to −1 and −1 to 1 do not strictly go up or down. In that case we represent the braid as done in Figure 8 in the next subsection. b −8 −7 −6 −5 −4 −3 −2 −1 1 2 3 4 5 6 7 8 b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b Figure 7. The Artin braid βx2 where x2 = ((8, 7, 5))[6, 3, 1][2] is as in Figure 1. The strands corresponding to the cycle [6, 3, 1] are drawn in blue. Note that it is a Mikado braid in ABn . −8 −7 −6 −5 −4 −3 −2 −1 1 2 3 4 5 6 7 8 b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b b Figure 8. The Artin braid βt for t = (1, −1)(2, −2) in case i1 = 2. In this way, we associate to every noncrossing partition x ∈ NC(WDn , c) an Artin braid βx ∈ ABn . For some x there are several possible βx ∈ ABn as illustrated in SIMPLE DUAL BRAIDS, NONCROSSING PARTITIONS AND MIKADO BRAIDS OF TYPE Dn 15 Figure 5, but they have the same image under πn,1 , hence πn,1 (βx ) is well-defined. We have Proposition 5.1. Let x ∈ WDn , t ∈ TDn such that x ≤T xt ≤T c. Then πn,1 (βx βt ) = πn,1 (βxt ). Proof. The situation x ∈ NC(WDn , c), t ∈ T and x ≤T xt ≤T c precisely corresponds to a cover relation in the noncrossing partition lattice of type Dn . These covering relations were described by Athanasiadis and Reiner [3, Section 3]: there are three families of covering relations. Setting y = xt, we have that x is obtained from y by replacing one or two balanced cycles or one paired cycle as follows: [j1 , j2 , . . . , jk ] 7→ [j1 , . . . , jℓ ]((jℓ+1 , . . . , jk )), 1 ≤ ℓ < k ≤ n − 1, ((j1 , j2 , . . . , jk )) 7→ ((j1 , . . . , jℓ ))((jℓ+1 , . . . , jk )), 1 ≤ ℓ < k ≤ n − 1, [j1 , . . . , jℓ ][jℓ+1 , . . . , jk ] 7→ ((j1 , . . . , jk )), 1 ≤ ℓ < k ≤ n − 1. Note that in the last case, we have either ℓ = 1 and j1 = ±i1 or k = ℓ + 1 and jk = ±i1 since x is a noncrossing partition. Indeed, the noncrossing partition has at most one polygon P with P = −P , in which case the middle point lies inside P . We have to show that the braid that we obtain by the concatenation βx ⋆ βt has the eBn as βxt . It is easy to deduce from the noncrossing representations same image in A Nx what the result of the concatenation of two such braids is. By the process explained above, the noncrossing diagram itself can be considered as an Artin braid, viewed inside a circle or rather a cylinder. An edge of a curvilinear polygon represents a strand, and the orientation indicates the startpoint and the endpoint of that strand. Consider the case where the cover relation xt 7→ x is the first one above, that is, it consists of breaking a symmetric polygon into a symmetric polygon and two opposite cycles. This means that xt has the two symmetric factors [i1 ] and [j1 , j2 , . . . , jk ] while x has the same factors as xt except that the two symmetric factors are replaced by [i1 ][j1 , . . . , jℓ ]((jℓ+1 , . . . , jk )) for some ℓ ∈ {1, . . . , k − 1} and t = ((jℓ , jk )). We have k ≥ 2. All the other polygons of x and xt have support disjoint from {±j1 , . . . , ±jk }, hence when concatenating βx ⋆βt it is graphically clear that they will stay unchanged: indeed, these polygons are disjoint from the two curvilinear polygons associated to the reflection t. Hence we can assume that xt = [i1 ][j1 , j2 , . . . , jk ] and x = [i1 ][j1 , . . . , jℓ ]((jℓ+1 , . . . , jk )). The situation is depicted in Figure 9 below. In the concatenated diagram, the strand starting at j1 first goes to −jℓ inside βx , then the strand starting at −jℓ goes to −jk inside βt . Hence the result is that the strand starting at j1 goes to −jk , and can be drawn as in the diagram on the right since there is no obstruction for such an isotopy. Similarly, the strand starting at −jℓ+1 first goes to −jk , then to jℓ , hence is isotopic to the strand which goes directly from −jℓ+1 to −jℓ as drawn in the picture on the right. The same happens on the other side, while all other strands stay unchanged. It follows that the result of the concatenation corresponds to the diagram on the right, which is precisely the diagram Nxt associated to xt. Hence we have the claim in the case where the cover relation is the one described, with k ≥ 2. We have to show the same for the other two cover relations. We also treat the case of the last cover relation and leave the second one to the reader. Note that in the case where the cover relation is given by [j1 , . . . , jℓ ][jℓ+1 , . . . , jk ] 7→ ((j1 , . . . , jk )), 16 BARBARA BAUMEISTER AND THOMAS GOBET b ... b b −jℓ −jk j1 b −i1 ... b j2 −j1 i1 b jℓ b b −j1 b jk−1 b b i1 b j2 b jℓ b b jℓ+1 b jk−1 ... j1 ... ... jk b −jk ... b −i1 ... b jk b b −j2 b −jk−1 b −jℓ ... b −j2 b −jℓ+1 b b ... −jk−1 b −jℓ+1 jℓ+1 ... Figure 9. Concatenating diagrams corresponding to the cover relation [i1 ][j1 , . . . , jk ] 7→ [i1 ][j1 , . . . , jℓ ]((jℓ+1 , . . . , jk ). we have either ℓ = 1 and j1 = ±i1 or ℓ + 1 = k and jk = ±i1 . Assume that ℓ + 1 = k and jk = −i1 , the case where jk = i1 as well as the cases where ℓ = 1, j1 = ±i1 are similar. We have x = ((j1 , . . . , jℓ , −i1 )), t = ((jℓ , i1 )). In this case, there are two possible diagrams Nx for x and the same holds for Nt (see Figure 5 for an illustration in the case where the noncrossing partition is a reflection). Since the corresponding braids βx obtained from the two different diagrams Nx have the same image under πn,1 we can choose any diagrams among the two, but the diagram Nt has to be chosen to be compatible with the diagram Nx if we want to do the same proof as for the first cover relation. One of the two situations is represented in Figure 10. Arguing as in the first case we then get the diagram on the right of the figure for the concatenation βx ⋆ βt . This diagram is the diagram Nxt up to the orientation of the two curves joining i1 to −i1 : but changing their orientation corresponds to inverting a middle crossing in βxt eBn thanks to the relation which gives rise to a braid which has the same image in A 2 s0 = 1. This proves the claim. −j2 b ... b b −j1 i1 b −i1 b i1 b b ... b b −jℓ b −j1 b jℓ −j2 b ... −jℓ −i1 b b j1 jℓ j2 b ... j1 b j2 Figure 10. Concatenating diagrams corresponding to the cover relation [j1 , . . . , jℓ ][−i1 ] 7→ ((j1 , . . . , jℓ , −i1 )).  Corollary 5.2. Let x ∈ NC(WDn , c). Then πn,1 (βx ) = xc . SIMPLE DUAL BRAIDS, NONCROSSING PARTITIONS AND MIKADO BRAIDS OF TYPE Dn 17 Proof. Recall that SDn = {(1, −2)(−1, 2)} ∪ {(i, i + 1)(−i, −i − 1) | i = 1, . . . , n − 1}. By construction of the braid βt from the diagram Nt we have that πn,1 (βs ) = s for all s ∈ SDn , and it is a general fact that sc = s for every simple reflection s. Hence we have the claim in case x is in SDn and in particular πn,1 (x) lies in ADn . Since by Proposition 5.1 the elements πn,1 (βt ) with t ∈ TDn satisfy the dual braid relations with respect to c, we claim that πn,1 (βt ) = tc for all t ∈ TDn . Indeed, for all t ∈ TDn , we can always find s ∈ SDn such that either st ≤T c or ts ≤T c, say, st ≤T c, and ℓS (sts) < ℓS (t) (this can be seen for instance using the noncrossing representation of t). It follows that we have the dual braid relation πn,1 (βs )πn,1 (βt ) = πn,1 (βsts )πn,1 (βs ). Arguing by induction on ℓS (t), we have that πn,1 (βq ) = qc for every reflection q occurring in the above equality except possibly t. Thanks to the dual braid relation sc tc = (sts)c sc we get that πn,1 (βt ) = tc and in particular that πn,1 (βt ) ∈ ADn . Now for x ∈ N C(WDn , c) arbitrary we can use Proposition 5.1 as well as the fact that x ≤T xt ≤T c, t ∈ TDn , implies that (xt)c = xc tc (see the end of Subsection 4.2) to get by induction on ℓT (x) that πn,1 (βx ) = xc .  5.4. Simple dual braids are Mikado braids. In all the examples drawn in the figures given in the previous sections, we see that the Artin braids βx resulting from simple dual braids are Mikado braids: they indeed satisfy the topological condition given by the point (2) of Theorem 3.2. This is the main statement which we want to prove here. Proposition 5.3. Let x ∈ NC(WDn , c). Then βx ∈ ABn is a Mikado braid. Proof. As βx ∈ ABn , it suffices to verify the point (2) of Theorem 3.2. Note that except in case x = (1, −1)(2, −2) and i1 = 2 (in which case the braid βx which is drawn in Figure 8 is obviously Mikado), the diagram which we obtained from Nx by putting all the dots on the same vertical line (as done in Figures 3 and 4; we call this diagram a vertical diagram) has the following property: each oriented curve joining two points either strictly increases or strictly decreases, and every two such distinct curves never cross. The first property follows from the fact that the diagram is obtained from Nx by projecting to the right a curve which is already either strictly increasing or strictly decreasing, while the second follows from the fact that the polygons in Nx do not cross. In such a diagram, consider a curve joining two points and going up with respect to the orientation, with no other curve lying at its right. It follows from the discussion in the paragraph above that it always exists. Every single point lying at the right of such a curve corresponds to a vertical unbraided strand in βx which lies above all the other strands. Therefore, every such point can be removed in the vertical diagram, and the symmetric point lying at the left of the curve which is symmetric to the original curve can be removed simultaneously: it corresponds to removing a vertical unbraided strand lying above all the other strands in βx , and simultaneously removing the symmetric unbraided strand lying below all the other strands, giving a new braid βx′ lying in ABn−1 since we removed a symmetric pair of strands. After removing all such points in the vertical diagram, the original curve has nothing at its right, hence corresponds to a strand which lies above all the other strands, and we can therefore remove it, as well as its symmetric strand. Again we obtain an element which lies in an Artin group of type Bm for a smaller m. Going on inductively, we can remove every strand corresponding to a curve, with a braid which stays symmetric at each step. If after removing the last curve we still have points, these correspond to vertical unbraided strands which can be removed. This concludes by Theorem 3.2. We illustrate the above procedure in Example 5.4 below. 18 BARBARA BAUMEISTER AND THOMAS GOBET  Note that we could define more generally vertical diagrams (not necessarily corresponding to simple dual braids) and associate to them an Artin braid, which would therefore always be Mikado. Example 5.4. We illustrate the procedure given in the proof of Proposition 5.3 in case x is the element x2 = ((8, 7, 5))[6, 3, 1][2] from Figure 1. The vertical diagram and the braid βx are given in Figure 7. The blue curve joining 6 to −1 in the vertical diagram has no other curve lying at its right. There is only the single point 4, which corresponds in βx to a strand which lies above all the others, with the symmetric strand −4 lying below all the others. Removing the pair of strands 4 and −4, we get a symmetric braid on 14 strands, hence in AB7 . We can then remove the strand corresponding to the original curve joining 6 to −1 as well as its symmetric strand, since there is no remaining strand lying above it. Going on inductively we eventually remove all pairs of strands. As a corollary we get the main result Theorem 5.5. Let x ∈ NC(WDn , c). Then xc is a Mikado braid. Proof. By Corollary 5.2 we have that πn,1 (βx ) = xc for every x ∈ NC(WDn , c). But by Proposition 5.3, βx is a Mikado braid in ABn . Applying Theorem 3.3 we get that xc = πn,1 (βx ) is a Mikado braid in ADn .  References [1] D. Allcock, Braid pictures for Artin groups, Trans. Amer. Math. Soc. 354 (2002), 3455-3474. [2] D. Armstrong, Generalized noncrossing partitions and combinatorics of Coxeter groups, Mem. Amer. Math. Soc. 202 (2009), No. 949. [3] C.A. Athanasiadis and V. Reiner, Noncrossing partitions for the group Dn , SIAM J. Discrete Math. 18 (2004), no. 2, 397-417 (electronic). [4] B. Baumeister, T. Gobet, K. Roberts, and P. Wegener, On the Hurwitz action in finite Coxeter groups, J. of Group Theory 20 (2017), 103-131. [5] D. Bessis, The dual braid monoid, Ann. Sci. École Normale Supérieure 36 (2003), 647-683. [6] D. Bessis, F. Digne, and J. Michel, Springer theory in braid groups and the Birman-Ko-Lee monoid, Pacific J. Math. 205 (2002), 287–309. [7] J. Birman, K.H. Ko, and S.J. Lee, A New Approach to the Word and Conjugacy Problems in the Braid Groups, Adv. in Math. 139 (1998), 322–353. [8] A. Björner and F. Brenti, Combinatorics of Coxeter groups, GTM 231, Springer, 2005. [9] N. Bourbaki, Groupes et algèbres de Lie, chapitres 4,5 et 6, Hermann (1968). [10] J. Crisp and L. Paris, Artin groups of type B and D, Adv. Geom. 5 (2005), no. 4, 607-636. [11] P. Dehornoy, F. Digne, D. Krammer, E. Godelle, and J. Michel. Foundations of Garside theory, Tracts in Mathematics 22, Europ. Math. Soc. (2015). [12] F. Digne and T. Gobet, Dual braid monoids, Mikado braids and positivity in Hecke algebras, Math. Z. 285 (2017), no. 1-2, 215-238. [13] P. Dehornoy, Three-dimensional realizations of braids, J. London Math. Soc. 60 (1999), 108–132. [14] M.J. Dyer, Modules for the dual nil Hecke ring, http://www3.nd.edu/~dyer/papers/nilhecke.pdf. [15] T. Gobet, Twisted filtrations of Soergel bimodules and linear Rouquier complexes, J. Algebra 484 (2017), 275-309. [16] P.N. Hoefsmit, Representations of Hecke Algebras of Finit Groups with BN-Pairs of Classical type, PhD thesis, University of British Columbia, August 1974. [17] J. Humphreys, Reflection groups and Coxeter groups, Cambridge Studies in Advanced Mathematics 29, Cambridge University Press (1990). [18] T. Licata and H. Queffelec, Braid groups of type ADE, Garside structures, and the categorified root lattice, preprint (2017), https://arxiv.org/pdf/1703.06011.pdf. [19] R.C. Lyndon and P.E. Schupp, Combinatorial group theory, Springer, Berlin, 1977. SIMPLE DUAL BRAIDS, NONCROSSING PARTITIONS AND MIKADO BRAIDS OF TYPE Dn 19 [20] J. McCammond and R. Sulvay, https://arxiv.org/abs/1312.7770. Artin groups of euclidean type, preprint (2013), Barbara Baumeister, Fakultät für Mathematik, Universität Bielefeld, Postfach 10 01 31, 33501 Bielefeld, Germany. E-mail address: [email protected] Thomas Gobet, Institut Élie Cartan de Lorraine, Université de Lorraine, site de Nancy, B.P. 70239, 54506 Vandoeuvre-lès-Nancy Cedex, France E-mail address: [email protected]
4math.GR
Local Parametric Estimation in High Frequency Data∗ arXiv:1603.05700v3 [q-fin.ST] 10 Mar 2017 Yoann Potiron** and Per Mykland*** ** Faculty of Business and Commerce, Keio University. 2-15-45 Mita, Minato-ku, Tokyo, 108-8345, Japan. Phone: +81 3 5418 6571. Email: [email protected], website: http://www.fbc.keio.ac.jp/~potiron *** Department of Statistics, The University of Chicago. 5734 S. University Avenue, Chicago IL, 60637. Phone: + 1 773 702 8044/8333. Fax: + 1 773 702 9810. Email: [email protected], website: http://galton.uchicago.edu/~mykland This version: March 13, 2017 Abstract In this paper, we give a general time-varying parameter model, where the multidimensional parameter, which possibly includes jumps, follows a local martingale. The quantity of interest is defined as the integrated value over time of the parameter process RT Θ = T −1 0 θt∗ dt. We provide a local parametric estimator (LPE) of Θ and conditions under which we can show the central limit theorem. The framework is restricted to the specific convergence rate n1/2 . Several examples of LPE, in which √ the conditions are shown to be satisfied assuming that the microstructure noise is Op (1/ n), are provided: estimation of volatility, powers of volatility, high-frequency covariance and volatility when incorporating trading information. The LPE considered in those cases are variations of the maximum likelihood estimator. We also treat the case of time-varying model with uncertainty zones and the time-varying MA(1). Keywords: high-frequency data; integrated volatility ; local martingale; local parametric estimator; market microstructure noise; powers of volatility; quasi-maximum likelihoodestimator MSC2010 subject classifications: 62G05, 62G20, 62M09, 91B84. ∗ Financial support from the National Science Foundation under grant DMS 14-07812 and from Keio University is greatly acknowledged. We are indebted to Simon Clinet, Takaki Hayashi, Dacheng Xiu, participants of the seminars in Berlin and Tokyo and conferences in Osaka, Toyama, the SoFie annual meeting in Hong Kong, the PIMS meeting in Edmonton for valuable comments, which helped in improving the quality of the paper. 1 1 Introduction Modeling dynamics is very important in various fields, including finance, economics, physics, environmental engineering, geology and sociology. Time-varying parametric models can deal with a special problem in dynamics, namely, the temporal evolution of systems. The extensive literature on time-varying parameter models and local parametric methods include and are not limited to Fan and Gijbels (1996), Hastie and Tibshirani (1993) or Fan and Zhang (1999) when regression and generalized regression models are involved, locally stationary processes following the work of Dahlhaus (1997, 2000), Dahlhaus and Rao (2006), or any other timevarying parameter models, e.g. Stock and Watson (1998) and Kim and Nelson (2006). In this paper, we propose to specify local parametric methods in the particular context of high-frequency observations for a general class of problems. Local methods have been used very much in the high-frequency data literature, see e.g. Mykland and Zhang (2009, 2011), Kristensen (2010), Reiß (2011) or Jacod and Rosenbaum (2013) among many others. If we define T as the horizon time, the target quantity in this monograph is defined as the integrated parameter Z 1 T ∗ Θ := θ ds, (1) T 0 s which can be equal to the volatility, the covariation between several assets, the variance of the microstructure noise, the friction parameter of the model with uncertainty zones (see Section 4.5 for more details), the time-varying parameters of the MA(1) model, etc. To estimate the integrated parameter, we estimate the local parameter on each block by using the parametric estimator on the observations within the block and take a weighted sum of the local parameter estimates, where each weight is equal to the corresponding block length. We call the obtained estimator the local parametric estimator (LPE). In Section 3, we investigate conditions under which we can establish the corresponding central limit theorem (CLT) with convergence rate n1/2 , where n is the (possibly expected) number of observations. The framework is such that the local block length vanishes asymptotically. The main contribution of this paper is not so much the general technology in itself described in this section, which is rather simple and natural, as the six applications which are derived easily as a corollary to the introduced theory along with their respective central limit theory work. Those applications are discussed extensively in Section 4 where we also provide the associated CLTs. As far as the authors know, all those results are new. Depending on the considered example, the LPE is useful for one or several of the following reasons: robustness: the LPE is robust to time-varying parameters (such as the noise variance, η from the model with uncertainty zones, the parameters of the MA(1) process) which are usually assumed constant. This is the case of all our examples, except for Example 4. efficiency: the LPE turns out to be more efficient than the global estimator or existing concurrent approaches. This is the case of Example 4. In addition, the LPE is conjectured to be efficient in all our examples except for Example 5. 2 definition of new estimators: It can be the case that the estimator doesn’t work globally but that the LPE provides a good candidate as in Example 2 and Example 4. We describe the six examples in what follows. To estimate integrated volatility under noisy observations, Xiu (2010) studied the quasi-maximum likelihood-estimator (QMLE), showed the corresponding asymptotic theory when the variance of the noise is fixed and obtained a convergence rate n1/4 , which is optimal (see Gloter and Jacod (2001)). The behavior of the QMLE remains unknown when the variance of the noise goes to 0. When assuming that it is √ Op (1/ n), we show that the LPE of the QMLE is optimal (with rate n1/2 ) and furthermore robust to time-varying noise variance. Another important problem, which goes back to Barndorff-Nielsen and Shephard (2002), is the estimation of higher powers of volatility. When there is microstructure noise, Jacod et al. (2010) show that the pre-averaging approach can be used, although not working when estimating the integral of σtk with odd k. To estimate the powers of volatility, we define a LPE where the local estimates are powers of the QMLE of volatility. Under the assumption √ on small noise Op (1/ n), we show that this estimator is optimal and robust to time-varying noise variance. In particular, as far as the authors know, this provides the first estimator of the tricity (i.e. the integrated third moment of volatility) with noisy observations under which the central limit theory is investigated. Furthermore, this is an example where the global approach doesn’t work as the QMLE is only consistent when estimating volatility. The third example is the estimation of the high-frequency covariance with the QMLE from Aït-Sahalia et al. (2010). As for the volatility problem and under the assumption on snall noise, we show that the LPE of QMLE is optimal and robust to time-varying noise variance. A more recent problem is the estimation of volatility when incorporating trading information. To do that, Li et al. (2016) assume that the noise is a parametric function of trading √ information with a remaining noise component of order Op (1/ n). Their strategy consists in first estimating the parametric part of the noise, and then take the sum of square pre-estimated efficient returns. They also advocate for the use of the QMLE after price pre-estimation although they don’t provide the associated limit theory. We show that the latter approach, when considering the LPE of QMLE, is optimal and provides a better asymptotic variance (AVAR) than the former technique. In addition, a modification of the local estimator as in Example 2 allows us to estimate higher powers of volatility. A concurrent ultra high frequency approach to model the observed price was given in Robert and Rosenbaum (2011, 2012), who introduced the semiparametric model with uncertainty zones where η is the 1-dimensional friction parameter, observation times are endogenous and observed prices lie on a tick grid. As most likely correlated with the volatility, it is natural to consider ηt as a time-varying parameter. We provide a formal model extension and establish the according limit theory of the LPE of the estimator considered in their work. In addition, our empirical illustration seems to indicate that ηt is indeed time-varying. In the last example, we consider an application in time series and introduce a time-varying MA(1) model with null mean. The time series is observed in high frequency on [0, T ] and θt∗ 3 corresponds to the two-dimensional parameter of the MA(1) process. We show that the LPE of the MLE is optimal and document that it outperforms the global MLE and other concurrent approaches in finite sample. As the spectrum of applications is diverse, the locally parametric model (LPM) introduced in Section 2 is very general. In particular, in the case where we model the price of an asset as a continuous efficient Itô process, it allows for endogeneity, auto-correlated time-varying noise, correlation between the efficient return and the noise, as well as multidimensional asynchronous observations. The disadvantage of such approach is that the definition of the LPM and the corresponding CLT conditions are rather involved, and thus we advise the reader to consult Section 2 and Section 3 with a simple setting in mind. In addition, we specify the conditions when observations are regular in Section 3.4. The remaining of this paper is organized as follows. The LPM is introduced in the following section. Conditions for the CLT are stated in Section 3. We give the examples in Section 4. We investigate the finite sample performance of the LPE in a time-varying MA(1) model in Section 5. An empirical illustration on the model of uncertainty zones is carried out in Section 6. We conclude in Section 7. Proofs and theoretical arguments are discussed in the Appendix. 2 2.1 The Locally Parametric Model (LPM) Vocabulary We need first to fix the vocabulary for the rest of this paper. Parametric model will refer to the (non time-varying) parametric model of the econometrician, and not to the time-varying parameter model. Correspondingly, parametric estimator stands for the parametric estimator (in the parametric model). 2.2 Data-generating mechanism We assume that we infer from the dr -dimensional vectors {R1,n , · · · , RNn ,n }, which are functions of the observations where Nn can be random, the observation times are such that τ0,n := 0 < τ1,n < · · · < τNn ,n ≤ T and τi−1,n , τi,n is the time block corresponding to (d ) Ri,n . We further assume that the last component of Ri,n , Ri,nr , is equal to the time increment (1) ∆τi,n . As an example when dr := 2, Ri,n can be defined as the (possibly log) returns of the original observations. As such, we will refer abusively to Ri,n as returns in the rest of the paper, but they should be thought as observable quantities. We assume that the returns Ri,n depend on the underlying parameter process θt∗ . In particular, the sampling times can depend on θt∗ . 2.3 The parameter  We will assume that the p-dimensional parameter process θt∗ := (θt∗ )(1) , · · · , (θt∗ )(p) , which is restricted to lie in K, a (not necessarily compact) subset of Rp , is a continuous local martingale 4 of the form dθt∗ := σtθ dWtθ (2) where σtθ is a random nonnegative process (of dimension p × p), and Wtθ a standard pdimensional Brownian motion. The parameter θt∗ can be for example equal to the volatility, the high-frequency covariance, the time-varying noise variance, the friction parameter η from the model with uncertainty zones, the parameters of the MA(1) process, etc. It is actually possible to include finite activity jumps and a drift in the parameter (see Section 4). We don’t assume any independence between θt∗ and the other quantities driving the observations, such as the Brownian motion of the efficient price process, the volatility of the efficient price process, the microstructure noise, etc. In particular, there can be leverage effect (see e.g. Wang and Mykland (2014), Aït-Sahalia et al. (2014)). Also, the arrival times τi,n and the parameter θt∗ can be correlated, i.e. there is endogeneity in sampling times. In the rest ofP this paper, for any u-dimensional vector ν := (ν (1) , · · · , ν (u) ), we will use the notation | ν |:= uk=1 |ν (k) | when referring to the Manhattan norm. Also, for any θ ∈ K, we define the 1-dimensional vector θ+ as the minimum value of the 1-dimensional sub-parameters restricted to be positive. We assume that | θt∗ | is locally bounded and (θt∗ )+ is locally bounded away from 0. Furthermore, we assume that the volatility of the parameter, σtθ , is locally bounded. 2.4 Asymptotics There are commonly two choices of asymptotics in the literature: the high-frequency asymptotics, which makes the number of observations explode on [0, T ], and the low-frequency asymptotics, which takes T to infinity. We chose the former one. Investigating the low-frequency implementation case is beyond the scope of this paper1 . 2.5 The model for the returns We insist on the fact that the reader shouldn’t spend too much time on this general version of the model and advise her to go directly to Section 2.6 for one key example. Alternatively, we advise the reader curious to see where the general model comes from to consult Appendix I prior to reading this section. In what follows, the general time-varying parameter model is given in (4), whereas the parametric model is provided in (3). We want to allow a general model for the returns Ri,n . To do that, we introduce the random non-observed quantities Qi,n , which take values into the general space Qn . We also define Mi,n := (Qi,n , Ri,n ), and assume that Mi,n take values on the space Mn 2 , which is a subset of Qn × Rdr . Let m be a nonnegative integer (which can be infinite) which will stand 1 If we set down the asymptotic theory in the same way as in p.3 of Dahlhaus (1997), we conjecture that the results of this paper would stay true. 2 Mn is assumed to be a Borel space. 5 for the order of memory in the model. We define the m initial values M−(m−1),n , · · · , M0,n of the Markov chain. Also, we introduce the m-dimensional "memory" vectors of past quantities Mi,n := (Mi,n , · · · , Mi−(m−1),n ), which take values on a space Mm,n (subset of Mm n ). Finally, we will refer to Mi,n := (Qi,n , Ri,n ) respectively for the unobserved part and the observed part of Mi,n . The parametric model is assumed to be expressed as Mi,n := Fn (Mi−1,n , Ui,n , θ∗ ), (3) where Fn (x, y, z) is a Mn -valued non-random function3 , the random innovation Ui,n are independent and identically distributed (IID) with distribution U for a fixed n (but the distribution can depend on n) and independent of the past information. We assume that the time-varying parameter model can be expressed as4 Mi,n = Fn (Mi−1,n , Ui,n , {θs∗ }τi−1,n ≤s≤τi,n ). (4) The interpretation of Qi,n is as follows. Among all the non-observed information we have from the past, Qi,n are the quantities which matter in the distribution of the future returns. In particular, the parametric model (3) allows for auto-correlation in returns (up to a lag possibly bigger than m). 2.6 A simple setting: regular observations with additive IID noise We provide here one key example. Assume that the observations are following the additive model Zτi,n = Xτi,n + i,n , where Xt = σt dWt is the efficient price and i,n the IID noise independent from Xt . We also assume that observations are regular τi,n = iT /n. In that case the parameter is θt∗ = σt2 , the leverage effect is possible, and the returns are simply defined as Ri,n = Zτi,n − Zτi−1,n , the unobserved quantity Qi,n = i,n , and Ui,n = ({Ws }τi−1,n ≤s≤τi,n , i,n ). The order of the model is m = 1. The form of Fn can be defined from the expression Z τi,n   σs dWs + i,n − i−1,n i,n , τi−1,n = Fn  Z τi−1,n i−1,n ,   σs dWs + i−1,n − i−2,n , ({Ws }τi−1,n ≤s≤τi,n , i,n ), {σs2 }τi−1,n ≤s≤τi,n . τi−2,n 3 We assume that for i any positive integer, Ui,n ∈ Un where Un is a Borel space and that Fn (x, y, z) is defined on Mm,n × Un × Cp (R+ ). Additionally, we assume that Fn (x, y, z) is a jointly measurable Mn -valued function such that for any (Mn , Un , θt ) ∈ Mm,n × Un × Cp (R+ ), we have E | Fn (Mn , Un , θt ) |< ∞ 4 In all generality, we can assume Mi,n = Fn (Mi−1,n , Ui,n , {θs∗ }0≤s≤τi,n ), and the central limit theorem stays true under this more general assumption. In that case, the conditions of Section 3 would be slightly altered. Nonetheless, we choose to work under (4) to keep the analysis more tractable in Section 3. 6 2.7 Estimation The approach taken here is usual in high-frequency data. We define the block size as hn , and the number of blocks as Bn := pNn h−1 n q. For i = 1, · · · , Bn we define the spot parameter’s average on the ith block as R Ti,n Θi,n := ∗ Ti−1,n θs ds ∆Ti,n , (5) b i,n . Then, we take where Ti,n := min(τihn , T ) and its corresponding parametric estimator as Θ b i,n and obtain an estimator of the integrated spot process the weighted sum of Θ Bn X b n := 1 b i,n ∆Ti,n . Θ Θ T (6) i=1 Note that each block includes exactly hn observations, except for the last one, which might include fewer observations. We call (6) the local parametric estimator (LPE). 3 The central limit theorem We investigate in the following of this section the limit distribution. We advise the reader that the notations and conditions are very much involved although the reasoning is rather standard. This is due to the generality of the problem, in particular from the possible endogeneity of observation times. We specify and simplify the conditions in the regular observation times case in Section 3.4. The reason why we expose the following conditions is that the central limit theory of the examples from Section 4 is obtained as a corollary5 of this section. We insist on the fact that different approaches could have led to a different set of conditions. 1 The following methods are specified6 to the rate of convergence n 2 . Formally, we aim to find the limit distribution of 1 n 2 T −1 Bn X  b i,n − Θi,n ∆Ti,n . Θ (7) i=1 Specifically, we want to show that (7) converges stably 7 to a limit distribution. 5 By corollary, we mean that they are a corollary of the obtained conditions in this section together with results from their own central limit theory work 6 It is possible to specify the problem with a general rate of convergence, but all the considered examples from this paper are with rate n1/2 . 7 One can look at definitions of stable convergence in Rényi (1963), Aldous and Eagleson (1978), Chapter 3 (p. 56) of Hall and Heyde (1980), Rootzén (1980), and Section 2 (pp. 169-170) of Jacod and Protter (1998). 7 3.1 Stable convergence We give the definition of stable convergence and the according condition. Definition. (stable convergence) A sequence of integrable variable Zn known with information J is said to converge weakly to an integrable random variable Z if for all E ∈ J ,     E Zn 1E → E Z1E This mode of convergence, which is a bit stronger than the regular convergence in distribution, is due for the statistical purpose. Because we will obtain in the variance limit of (7) random quantities, we need the stable convergence to infer the same way we would do it if the variance limit was nonrandom. Since the stable convergence needs a corresponding information J to be defined with, we need to be more specific about the definition of J . We will be needing the following technical assumption, which can be easily verified on all the considered examples. The idea goes back to Heath (1977). We define Ii,n the information up to time τi,n as follows. We assume that Ii,n is a (discrete-time) filtration on (Ω, F, P ) such that {θs∗ }0≤s≤τi,n and Ui,n are adapted to Ii,n . Also, we assume that M0,n is I0,n -measurable. Finally, the assumption "Ui,n independent of the past information" can be formally expressed as Ui,n independent of Ii−1,n . Condition (E0). Ii,n can be extended into Ji,n 8 , where Ji,n is the interpolated information (c) (c) of a continuous information Jt , i.e. Ji,n = Jτi,n .   In the following of this paper, when using the conditional expectation Eτ Z 9 , we will (c) (c) refer to the conditional expectation of Z knowing Jτ . Finally, we consider J := JT information to go with stable convergence. 3.2 the Preliminary definitions and outline The parametric estimator takes the following form θ̂k,n := θ̂k,n (r1,n ; · · · ; rk,n ; r0,n ). (8) Let i = 1, · · · , Bn be the block number. We define the Markov chain elements on the ith block j Mi,n := M(i−1)hn +j,n for j = 1, · · · , hn . We also define the initial vector of the ith block as −(m−1) 0 ,··· ,M M0i,n := (Mi,n i,n ) := M(i−1)hn ,n . We assume that b i,n := θ̂h ,n (R1 ; · · · ; Rhn ; R0 ). Θ i,n i,n n i,n We alo introduce the following definition. 8 9 It means that Ji,n is a discrete filtration and for any i nonnegative integer Ii,n ⊂ Ji,n (c) The underlying assumption is that τ is a Jt -stopping time. 8 (9) (1),i  (1),i Definition. EM,n For any block number i = 1, · · · , Bn and any M > 0, EM,n is defined as 1 , · · · , U hn ) ∈ E the bounded space of time-varying parameter θt such that (θt , Ui,n M,n . i,n Consider the block number i = 1, · · · , Bn . Let M > 0, M ∈ Mm,n , θ ∈ KM and (1),i j,M,θ j,M,θt θt ∈ EM,n . For any j = −(m − 1), · · · , hn we define Mi,n and Mi,n the Markov chain approximations with initial vector M and fixed parameter θ (respectively with time-varying −(m−1),M,θ 0,M,θ parameter process θt ). The initial vectors are defined as (Mi,n , · · · , Mi,n ) := M −(m−1),M,θt and (Mi,n 0,M,θt , · · · , Mi,n ) := M. Also, we define the m-dimensional "memory" vectors j−(m−1),M,θ j−(m−1),M,θ j,M,θ j,M,θt t t as Mj,M,θ := (Mi,n , · · · , Mi,n ) and Mj,M,θ := (Mi,n , · · · , Mi,n ). Fii,n i,n nally, for any positive integer j, the jth element of the ith block Markov chain approximations are obtained by the same recurrence relations as (3) and (4)  j,M,θ j−1,M,θ i Mi,n := Fn Mi,n , Uj,n ,θ , (10)  j,M,θt j−1,M,θt j Mi,n := Fn Mi,n , Ui,n , {θs }τ j−1 ≤s≤τ j . (11) i,n i,n ˆ M with initial vector M ∈ M . We introduce now the infeasible estimator on the ith block Θ̃ m,n i,n The infeasible estimator is defined as 1,M,Θ̃i,n h ,M,Θ̃i,n 0,M,Θ̃ ˆ M := θ̂ Θ̃ ; · · · ; Ri,nn ; Ri,n i,n ). hn ,n (Ri,n i,n (12) Since there is endogeneity, observation times matter in this work. When approximating the returns on a block holding the parameter θt∗ constant, we also induce a change in the observation times. For that reason, we introduce the following definitions. We define the length of the with starting vector M and with non time-varying parameter θ as Phnith block M,θ j,M,θ (dr ) ∆Ti,n := j=1 (Ri,n ) and the length of the ith block with starting vector M and with P n j,M,θt (dr ) t (Ri,n ) . time-varying parameter θt as ∆TM,θ := hj=1 i,n Let the nonrandom known vector M∗n := (Q∗n , R∗n ) ∈ Mm,n , which is chosen by the econometrician to carry the analysis of (7). For instance, in the case of an additive IID noisy model as studied in Section 2.6, we can simply choose M∗n := (0, 0), i.e. the starting return  b i,n − Θi,n ∆Ti,n as and noise are null. We can decompose Θ ˆ M0i,n ∆TM0i,n ,Θ̃i,n  + Θ̃ ˆ M0i,n ∆TM0i,n ,Θ̃i,n − Θ̃ ˆ M∗n ∆TM∗n ,Θ̃i,n  b i,n ∆Ti,n − Θ̃ Θ i,n i,n i,n i,n i,n i,n    M∗n ,Θ̃i,n M∗n ,Θ̃i,n M∗n ˆ + Θ̃i,n − Θ̃i,n ∆Ti,n + Θ̃i,n ∆Ti,n − ∆Ti,n + Θ̃i,n − Θi,n ∆Ti,n (13) where the first term is the error in estimation due to the use of the approximated model (12) instead of the time-varying parameter model (4), the second term is the error made when taking M∗n instead of M0i,n as initial value of the block in the mixture of the parametric model (12), the third term corresponds to the error of the estimation of the constant parameter by the underlying approximations starting with a fixed initial value M∗n , the fourth term corresponds to the error of arrival times approximation when considering a parametric model starting with 9 initial value M∗n and the last term is the error made by holding the process parameter constant ˆ M0i,n is a mixture of the parametric model with parameter Θ̃ and on each block. Note that Θ̃ i,n i,n M0i,n . a mixture of starting value It is instructive to consider (13) when we assume that the time-varying parameter model is equal to the parametric model and that the observation times are regular. In that case, all the previously defined block length approximations are equal to ∆Ti,n . The first term, the fourth term and the fifth term are equal to 0 by definition. Additionally, we can hope that asymptotically the initial values of the blocks will be forgotten and thus that under right conditions Bn 1 X ˆ M0 ˆ M∗n ∆T ≈ 0. Θ̃ i,n − Θ̃ n2 i,n i,n i,n i=1 1 Finally, if we assume that we know the convergence rate n 2 and the limit distribution N (0, Vθ∗ ) 1  of the parametric estimator, for any i = 1, · · · , Bn we have that hn2 Θ̂i,n − θ∗ ∆Ti,n ≈ N (0, Vθ∗ )∆Ti,n and thus we can hope that 1 2 n T −1 Bn X  b i,n − θ∗ ∆Ti,n ≈ N (0, Vθ∗ ) Θ i=1 under right assumptions (in particular on the block size hn ). The time-varying parameter model and endogenous arrival time case will be treated in a similar way than the parametric model case. Formally, we will be providing in what follows conditions such that Bn  1 X P n2 Θ̃i,n − Θi,n ∆Ti,n → 0, (14) i=1 1 n2 n Bn X 1 2 M∗ ,Θ̃i,n Θ̃i,n ∆Ti,nn i=1 Bn X − ∆Ti,n  (15)  Z  M∗n ,Θ̃i,n st-D M∗n ˆ Θ̃i,n − Θ̃i,n ∆Ti,n → T 0 i=1 1 P → 0, n 2 T −1 T 1 2 Vθs∗ ds N (0, 1), Bn X P ˆ M0i,n ∆TM0i,n ,Θ̃i,n − Θ̃ ˆ M∗n ∆TM∗n ,Θ̃i,n  → Θ̃ 0, i,n i,n i,n i,n (16) (17) i=1 1 n2 Bn X P ˆ M0i,n ∆TM0i,n ,Θ̃i,n  → b i,n ∆Ti,n − Θ̃ Θ 0. i,n i,n (18) i=1 Note that if the sampling times are not regular, Vθ∗ will most likely be different from the variance of the parametric estimator. For instances of this when estimating volatility, one can look at Corollary 2.30 on p.154 in Mykland and Zhang (2012), where the "Asymptotic Quadratic Variation of Time" alters the variance of the RV estimator. 10 3.3 Main result In the following, we provide six conditions which we describe in what follows. Condition (E1) - Condition (E4) are used to prove (14) - (16). Moreover, (17) is proven with Condition (E5). Finally, we can show (18) if we assume Condition (E6). The first two assumptions are due to the endogeneity of sampling times and are not needed in the regular case specified in Section 3.4. Condition (E1). The observation times are such that for k = 1, 2, 4, 8   sup Eτi−1,n (∆τi,n )k = Op (n−k ). (19) 1≤i≤Nn Remark 1. (block length) As consequence of (19) when k = 1, we have that the  an obvious  block length is such that E ∆Ti,n = O(hn n−1 ). We make a second assumption on the observation times, which is due to non regularity of arrival times. We make sure that when the initial Markov-chain is fixed to M∗n , we can bound the difference in length between the approximated block and the true block, uniformly in the block initial value parameter θ0 and the path of the parameter process θt . Let M > 0 and (i) i = 1, · · · , Bn . For any θt ∈ EM,n , M ∈ Mm,n and N ∈ Mm,n , we define the difference in block length expectation as   M,N,θt t 0 Ei,n := E ∆TM,θ − ∆TN,θ i,n i,n , and the difference variance as   M,N,θt t 0 := Var ∆TM,θ − ∆TN,θ Vi,n i,n . i,n Condition (E2). For any M > 0, we have Bn X i=1 M∗n ,M0i,n ,θt sup (i) θt ∈EM,n Bn X Ei,n M∗n ,M0i,n ,θt sup Vi,n 1 = op (n− l ), 2 = op (n− l ). (20) (21) (i) i=1 θt ∈EM,n The following condition is the main assumption of this section. On examples in Section 4, it is verified using the central limit theory proofs of the respective estimators. Because we allow for endogeneity of observation times, the conditions are more involved. We advise the reader to consult Condition (E3’) in Section 3.4 for the regular observation times case. We introduce the    M,θ 1,M,θ hn ,M,θ definition of the bias on the ith block Bi,n := E θ̂hn ,n (Ri,n ; · · · ; Ri,n ; R) − θ ∆TM,θ i,n for any M ∈ Mm,n and any θ ∈ K. 11 Condition (E3). For any parameter θ ∈ K, we assume that there exists a covariance matrix Vθ > 0 such that for any M > 0, we have Vθ is bounded for any θ ∈ KM and uniformly in θ ∈ KM and in i = 1, · · · , Bn  1   1,M∗ ,θ h ,M∗ ,θ M∗ ,θ  M∗ ,θ  Var hn2 θ̂hn ,n (Ri,n n ; · · · ; Ri,nn n ; R∗n ) − θ ∆Ti,nn = Vθ E ∆Ti,nn T hn n−1 + o(h2n n−2 )     ∗ ∗ ∗ 1,M ,θ h ,M ,θ M ,θ 4 E θ̂hn ,n (Ri,n n ; · · · ; Ri,nn n ; R∗n ) − θ ∆Ti,nn = O(h2n n−4 ) M∗ ,θ Bi,nn = o(hn n − 23 ) (22) (23) (24) We make a fourth assumption, which is on the block size hn . In practice, Condition (E4) provides us with the maximum block size hn to use for constant approximation of parameter. Note that the condition corresponds exactly to the block size order found in Mykland and Zhang (2011), who investigated how constant we can hold volatility in a small neighborhood in the case of regular observations of the price following a continuous Itô-process. Condition (E4). The block size hn is such that 1 n− 2 hn = o(1). (25) The next condition assumes that through uniformity in the parameter value and future parameter path, we can bound the discrepancy of the parametric model starting with two different initial vectors. Let M > 0. For any θ ∈ KM , M1 ∈ Mm,n and M2 ∈ Mm,n , we introduce the difference bias on the ith block as h M1 ,M2 ,θ 1,M1 ,θ hn ,M1 ,θ 1 ,θ Ei,n := E θ̂hn ,n (Ri,n , · · · , Ri,n ; R1 )∆TM i,n i 1,M2 ,θ hn ,M2 ,θ 2 ,θ − θ̂hn ,n (Ri,n , · · · , Ri,n ; R2 )∆TM , i,n and the difference variance as h M1 ,M2 ,θ 1,M1 ,θ hn ,M1 ,θ 1 ,θ Vi,n := Var θ̂hn ,n (Ri,n , · · · , Ri,n ; R1 )∆TM i,n i 1,M2 ,θ hn ,M2 ,θ 2 ,θ − θ̂hn ,n (Ri,n , · · · , Ri,n ; R2 )∆TM , i,n Condition (E5). For any M > 0 we have Bn X M0 ,M∗n ,θ sup Ei,ni,n i=1 θ∈KM Bn X M0 ,M∗n ,θ sup Vi,n i,n i=1 θ∈KM 12 1 = op (n− 2 ), (26) = op (n−1 ). (27) The last condition is very similar to Condition (E5), except that we are not looking at the discrepancy induced by different initial vectors but rather the difference between the timevarying parameter model and the parametric model when initial vectors are equal. For that reason, we provide the following definitions. Let i = 1, · · · , Bn be the block number. For any (1),i M > 0, any θt ∈ EM,n and any M ∈ Mm,n , we introduce the difference bias on the ith block as  M,θt 1,M,θt hn ,M,θt t Ei,n := E θ̂hn ,n (Ri,n , · · · , Ri,n ; R)∆TM,θ i,n  1,M,θ0 hn ,M,θ0 0 , − θ̂hn ,n (Ri,n , · · · , Ri,n ; R)∆TM,θ i,n as well as the variance of the difference  M,θt 1,M,θt hn ,M,θt t Vi,n := Var θ̂hn ,n (Ri,n , · · · , Ri,n ; R)∆TM,θ i,n  1,M,θ0 hn ,M,θ0 0 − θ̂hn ,n (Ri,n , · · · , Ri,n ; R)∆TM,θ . i,n Condition (E6). For any M > 0, we have Bn X sup (1),i M0 ,θt 1 Ei,ni,n = op (n− 2 ), (28) = op (n−1 ). (29) i=1 θt ∈EM,n Bn X i=1 M0 ,θt sup Vi,n i,n (1),i θt ∈EM,n We now state the main theorem of this paper, which investigates the limit distribution of (13). Theorem 1. (Central Limit Theorem) Assume Condition (E0) - Condition (E6). Then, stably in law as n → ∞, Z T 1   1 2 b n − Θ → T −1 n2 Θ Vθs∗ ds N (0, 1). (30) 0 Remark 2. (parametric model with regular observations) Note that in the case where observations are regular and the time-varying parameter model is equal to the parametric model b n is equal to the variance of the with true parameter θ∗ , the asymptotic variance (AVAR) of Θ parametric model, i.e. 1  1 b n − Θ → V 2∗ N (0, 1). n2 Θ θ Remark 3. (subset parameter estimation) In practice, one can be interested in estimating only a p0 -dimensional integrated parameter Ξ, where p0 ∈ {1, · · · , p − 1} and Ξ is a subset of Θ. If we define in analogy with the quantities depending on θt∗ the p0 -dimensional subquantities b i,n Ξ b n as well as the p0 × p0 -dimensional matrix Vξ , the Central Limit Theorem ξt∗ , ξˆk,n , Ξi,n , Ξ still holds in that case under the same assumptions Z T  1  1 2 −1 b n 2 Ξn − Ξ → T Vθs∗ ds N (0, 1). 0 13 Remark 4. (estimating the asymptotic variance) If the practitioner doesn’t have a (parametric) variance estimator at hand and that her parametric estimator can be written as in Mykland and Zhang (2016), one can use the techniques of the cited paper to obtain a variance estimate. Investigating if such techniques would work in our setting is beyond the scope of this paper. If she has a variance estimator v̂hn ,n , then for any i = 1, · · · , Bn she can estimate the ith 1 ; · · · ; Rhn ; R0 ), and the asymptotic variance as the block variance Vbi,n as Vbi,n := v̂hn ,n (Ri,n i,n i,n weighted sum Vbn = T −2 Bn X Vbi,n ∆Ti,n . (31) i=1 Under conditions similar to the ones in Appendix I and the ones from this current section, we can obtain the consistency of (31). 3.4 When observation times are regular We specify Theorem 1 when observations are regular, i.e. τi,n = iT n . In that case, note that Condition (E1) and Condition (E2) are straightforwardly satisfied. We also have that Condition (E4) remains unchanged. Finally, we adapt the other conditions in the following. Condition (E3’). For any parameter θ ∈ K, we assume that there exists a covariance matrix Vθ > 0 such that for any M > 0, we have Vθ is bounded for any θ ∈ KM and uniformly in θ ∈ KM and in i = 1, · · · , Bn  1  1,M∗ ,θ h ,M∗ ,θ Var hn2 θ̂hn ,n (Ri,n n ; · · · ; Ri,nn n ; R∗n ) − θ = Vθ T hn n−1 + o(h2n n−2 )    ∗ ∗ 4 1,M ,θ h ,M ,θ E θ̂hn ,n (Ri,n n ; · · · ; Ri,nn n ; R∗n ) − θ = O(h−2 n )   ∗ ∗ 1 1,M ,θ h ,M ,θ E θ̂hn ,n (Ri,n n ; · · · ; Ri,nn n ; R∗n ) − θ = o(n− 2 ) (32) (33) (34) Condition (E5’). For any M > 0 we have Bn X h 1,M0 ,θ hn ,M0 ,θ sup E θ̂hn ,n (Ri,n i,n , · · · , Ri,n i,n ; R0i,n ) i=1 θ∈KM 1,M∗n ,θ −θ̂hn ,n (Ri,n Bn X h ,M∗n ,θ , · · · , Ri,nn i −1 1 ; R∗n ) = op (hn 2 n 2 ), (35) h 1,M0 ,θ hn ,M0 ,θ sup Var θ̂hn ,n (Ri,n i,n , · · · , Ri,n i,n ; R0i,n ) i=1 θ∈KM 1,M∗n ,θ −θ̂hn ,n (Ri,n h ,M∗n ,θ , · · · , Ri,nn 14 i ; R∗n ) = op (h−1 n n), (36) Condition (E6’). For any M > 0, we have Bn X sup (1),i  1,M0 ,θt hn ,M0 ,θt E θ̂hn ,n (Ri,n i,n , · · · , Ri,n i,n ; R0i,n ) i=1 θt ∈EM,n 1,M0i,n ,θ0 −θ̂hn ,n (Ri,n Bn X i=1 sup Var (1),i θt ∈EM,n  hn ,M0i,n ,θ0 , · · · , Ri,n 1,M0i,n ,θt θ̂hn ,n (Ri,n 1,M0i,n ,θ0 −θ̂hn ,n (Ri,n ; R0i,n ) hn ,M0i,n ,θt , · · · , Ri,n hn ,M0i,n ,θ0 , · · · , Ri,n  −1 1 = op (hn 2 n 2 ), (37) ; R0i,n )  ; R0i,n ) = op (h−1 n n). (38) We now provide the corresponding theorem. Theorem 2. (Central Limit Theorem with regular observation times) Assume Condition (E0), Condition (E3’), Condition (E4), Condition (E5’) and Condition (E6’). Then, stably in law as n → ∞, Z T  1  1 2 b n − Θ → T −1 Vθs∗ ds N (0, 1). (39) n2 Θ 0 4 Examples This section provides some applications of the theory introduced in Section 3. To show that the conditions are satisfied, we use the theory work from each considered estimator. In all those examples this is relatively straightforward to show that conditions hold (see the proofs in Appendix II), indicating that our theory can be applied easily in those cases. The corresponding CLTs provided in this section are all new. We choose five examples with regular observations in which it is sufficient to show the conditions of Theorem 2. We further consider the model with uncertainty zones where there is endogeneity in observation times and in which we need to verify the more general conditions of Theorem 1. Two further examples of application of those techniques include our own recent work. Due to the high complexity of those instances, the conditions of Theorem 1 are not straightforwardly satisfied, but they provided a basis of work which significantly simplified the proofs. Potiron and Mykland (2017) introduced a bias-corrected Hayashi-Yoshida estimator (Hayashi and Yoshida (2005)) of the high-frequency covariance and showed the corresponding CLT under endogenous and asynchronous observations. To model duration data, Clinet and Potiron (2017a) built a time-varying parameter extension to the Hawkes self-exciting process, derived the bias-corrected MLE and showed the CLT of the corresponding LPE. In all the examples considered we have that the convergence rate n1/2 is optimal from Gloter and Jacod (2001). Thus the CLTs provided in the following of this section are all optimal. In the case of (42) in Example 1, (44) in Example 2, Theorem 5 (ii) and Theorem 15 6 (ii) in Example 4, Theorem 8 in Example 6, we also have a parametric estimator which achieves the Cramér-Rao bound of efficiency locally, thus we conjecture that the obtained variance is the nonparametric efficient bound. In the case of (41) in Example 1, i.e. when √ estimating volatility assuming that the noise is very small i,n = op (1/ n), the asymptotic R R T T variance is equal to 6T −1 0 σs4 ds, whereas the efficient bound 2T −1 0 σs4 ds is attained by the RV. This increases the variance by a factor of 3, which is also observed on the MLE (when assuming the volatility is constant) when misspecified on a model which doesn’t incorporate microstructure noise (see, e.g., Section 2.4 pp. 1486-1487 in Barndroff-Nielsen et al. (2008)). Rt Example 4.1-4.5 feature an underlying continuous-Itô price model dXt = 0 σu dWu (Example 4.6 considers a time series without any underlying price process). We discuss now how we can add a drift, allow for finite activity jump in both the price process and volatility and stochastic observation times when needed. We first show how to add a drift component. By Girsanov theorem, together with local arguments (see, e.g., pp. 158-161 in Mykland and Zhang (2012)), we can weaken the price and volatility local-martingale assumption by allowing them to follow an Itô-process (of dimension 2 in case of volatility or powers of volatility estimation, dimension 4 when considering the highfrequency covariance problem), with a volatility matrix locally bounded and locally bounded away from 0, and drift which is also locally bounded. It is also easy to see that we can allow for finite activity jumps in θt∗ and in Xt with a b i,n to a certain threshold truncation technique. To do that, one might cap the values of Θ e i,n = min(Θ b i,n , S). Consider S which doesn’t depend on n. More specifically, we define Θ ∗ Jn ⊂ {1, · · · , Bn } the set of blocks where there is at least one jump in θt and/or Xt . As the number of blocks B → ∞, the cardinal of Jn is at most finite, and thus we have that Nn X 1X e i,n ∆Ti,n ≈ 1 e i,n ∆Ti,n . Θ Θ T T i=1 i∈J / n It is then immediate to adapt the proof of the CLT, and it also works for Example 4.6 (in the case of the parameter jump only). Actually, when jumps occur only in the parameter, we don’t require to truncate the estimator. On the other hand, if infinitely many jumps are possible in both the price process and the parameter, the theoretical development is beyond the scope of this paper. In Example 4.1-4.4, the assumption on regular observation times can be weaken to stochastic times. To do that, we can follow the analysis of Barndorff-Nielsen et al. (2008, Section 5.3, pp. 1505-1507). To express their setting as a LPM, we assume that the observation times are of the form τi,n = Γi/(nT ) , (40) Rt where Γt is a stochastic process satisfying Γt = 0 Γ̃2u du, with Γ̃t a strictly positive parameter of the LPM. We can then construct a (change of time) process X̃t = XΓt so that for X̃t the 16 observations are regular. In view of Dambis Dubins-Schwarz theorem (see, e.g., Theorem 1.6 on p. 181 in Revuz and Yor (1999)) we have that [X]T = [X̃]ΓT . In addition, it is immediate to see that Condition (E1)-(E2) hold in that case. Alternatively one can assume that the quadratic variation of time (see, e.g., Assumption A on p. 1939 in Mykland and Zhang (2006)) exists and that observation times are independent of the price process. Under suitable assumptions, we can also show that Condition (E1)-(E2) hold in that case. Another nice corollary about the introduced theory is that we can obtain the CLT of the b i,n ). powers of the integrated parameter g(θt∗ ) effortlessly when using the local estimates g(Θ Essentially, the proof uses on each block a Taylor expansion as in the delta method. Rather than stating a general abstract result, we apply the technique on the local QMLE in Example 4.2 and on an adapted estimator from Li et al. (2016) in Example 4.4 to estimate the higher powers of volatility. It is clear that the same technique would work in the other examples too, although we don’t state the corresponding theorems. Finally, as the parametric estimator must satisfy bias conditions (24), (26) and (28), it is useful to consider in some instances a bias-corrected (BC) version of it which provides the b (BC) . The BC LPE is then constructed as estimate on the ith block Θ i,n Bn 1X (BC) b b (BC) ∆Ti,n . Θn = Θ i,n T i=1 Moreover, it is easy to see that Condition (E0), which is a natural assumption on the underlying filtration, is immediately satisfied in all the examples. Condition (E4), i.e. hn = o(n1/2 ), is not model specific and thus there is nothing further to show in the following examples. 4.1 Estimation of volatility with the QMLE We assume the setting of Section 2.6. We further assume that the noise has the form 1 i,n := n−α vt2 γt , where α ≥ 1/2, the noise variance vt is time-varying, and γt are IID with null-mean and √ unity variance. In other words we have i,n = Op (1/ n). The parameter process is defined as the two-dimensional volatility and noise variance process θt∗ = (σt2 , vt ) and thus Θ =  RT RT T −1 0 σt2 dt, T −1 0 vt dt . Correspondingly we work locally with the QMLE introduced in b n = (σ̂n2 , v̂n ). Xiu (2010, p. 236) and we introduce the notation for the corresponding LPE Θ b (BC) We also consider the bias-corrected version of the QMLE Θ , where the procedure to n construct the unbiased estimator is given in Appendix II Section 7.6.1. In numerical simulations under a realistic framework, this bias is not observed even with small values of n (see Section 6 in Xiu (2010) and Section 7 in Clinet and Potiron (2017b)), and thus it is safe to b n = (σ̂ 2 , v̂n ) in practice. use Θ n 17 Xiu (2010) showed the CLT of the QMLE when vt is non time-varying and α = 0. In the same setting Clinet and Potiron (2017b) showed that the asymptotic variance can be smaller when using the LPE with Bn = B fixed and documented that in finite sample the LPE was advantageous over the global QMLE. As in those papers where the asymptotics is such that the noise variance doesn’t shrink to 0 and the convergence rate is n1/4 , the techniques of this paper don’t apply. However the techniques allow us to investigate how the LPE behaves in a different asymptotics where the behavior of the QMLE and the LPE remains unknown, i.e. when the noise √ variance is Op (1/ n) and Bn tends to +∞. Moreover, we allow for heteroskedasticity in noise variance. Finally, in the case where the noise variance goes to 0 at the same speed as the variance of the returns, i.e. α = 1/2, we can also retrieve the integrated variance noise. In accordance with the setting of this paper, the convergence rate of both the volatility and the noise is n1/2 . Although the AVAR of the global QMLE is unknown in this setting, we expect that the LPE is more efficient in analogy with what was found in Clinet and Potiron (2017b). To verify the conditions for the CLT, we use heavily the asymptotic results of the QMLE (see Theorem 6 in Xiu (2010)) and the MLE in the low-frequency asymptotics (see Proposition 1 on p. 369 in Aït-Sahalia et al. (2005)). The result is formally embedded in the following theorem. Theorem 3. (QMLE) (i) We assume that α > 12 . Then, stably in law as n → ∞, n 1 2  σ̂n2 −T −1 Z T  σs2 ds  → 6T −1 Z T 1 2 σs4 ds N (0, 1). (41) 0 0  1 b (BC) (ii) When α = 12 , we have stable convergence in law of T n 2 Θ − Θ to a mixed normal n random variable with zero mean and asymptotic variance given by ! p p   RT RT 3 4v + σ 2 ds 4 − 0 σs4 + 2σs2 vs + 4σs3 4vs + σs2 ds s s 0 2σs + 4σs . (42) p   R 1 T 2 σ 2 + 2v + σ 2 ds 2v + σ • 4v + σ s s s s s s s 2 0 4.2 Estimation of powers of volatility Here the parameter is θt∗ = g(σt2 ) with g not being the identity function. We are concerned RT with the estimation of powers of volatility Θ = T −1 0 g(σt2 )dt under microstructure noise √ with variance Op (1/ n) in the same setting as in Section 4.1. The problem was introduced in Barndorff-Nielsen and Shephard (2002). They showed that the case g(x) = x2 is related to the asymptotic variance of the realized volatility. One can also consult Barndorff-Nielsen et al. (2006), Mykland and Zhang (2012, Proposition 2.17, p. 138) and Renault et al. (2013) for related developments. All those studies assume no microstructure noise. When there is microstructure noise, Jacod et al. (2010) use the pre-averaging method. In the special case of quarticity, one can also look at Mancino and Sanfelici (2012) and Andersen 18 et al. (2014). The pre-averaging approach only works when estimating σt2k with k a positive integer. As far as the authors know, the problem when the power is odd hasn’t been investigated so far. In particular, the question of how to estimate tricity under microstructure noise stays open. Under no microstructure noise, block estimation (Mykland and Zhang (2009, Section 4.1, p. 1421-1426)) has the ability to make the mentioned estimators approximately or fully efficient. 2 and then take a The path followed to do that is to first estimate locally the volatility σ̂i,n 2 Riemann sum of g(σ̂i,n ). See also Jacod and Rosenbaum (2013) for related work. In the same spirit when allowing for microstructure noise, we propose to use locally the 2 ), where σ̂ 2 is the QMLE estimate of the volatility on the ith block. The estimation g(σ̂i,n i,n b n = Bn−1 PBn g(σ̂ 2 ). Also, the BC estimator is defined global estimate is thus defined as Θ i,n i=1 PBn 2 )(BC) ). The theorem is given in what follows. The proof uses a −1 b (BC) g((σ̂ as Θ = B n n i,n i=1 local delta method and then follows the proof of Theorem 3. To reflect on the powerfulness of the local approach, the reader can note that the global QMLE is estimating the wrong quantity when g is different from the identity function, except when the volatility is constant. To see why this is the case, we consider the estimation of RT quarticity (i.e. with g(x) = x2 ) and we note that a global QMLE would estimate g( 0 σt2 dt), RT which is except when volatility is constant different from 0 σt4 dt. Indeed, the extensive empirical work in Andersen et al. (2014) indicates that the two quantities are very different in practice. Theorem 4. (powers of volatility) (i) We assume that α > 12 . Then, stably in law as n → ∞, n 1 2   b Θn − Θ → 6T −1 T Z g 0 1 2 (σs2 ) σs4 ds 2 N (0, 1). (43) 0 (ii) When α = 12 , we have n 1 2   b (BC) Θ − Θ → T −1 n Z T g 0 (σs2 ) 2σs4 + 4σs3 p   4vs + σs2 ds N (0, 1). (44) 0 4.3 Estimation of high-frequency covariance with the QMLE To estimate integrated covariance under noisy observations and asynchronous observations, Aït-Sahalia et al. (2010) introduced a QMLE based on a synchronization of observation times. It is clear that their Generalized Synchronization Method can be expressed as a LPM. In view of the close connection between their proposed estimator (2) on p. 1506 and the QMLE studied in Section 4.1, the conditions of our work can be verified and thus Theorem 2 (p. 1506) of the authors can be adapted with the LPE in a framework similar to Section 4.1, i.e. when the noise variance is O(n−1/2 ) and time-varying. 19 4.4 Estimation of volatility and higher powers of volatility incorporating trading information To incorporate all the information available in high frequency data (e.g. in addition to transaction prices, we also observe the trading volume, the type of trade i.e. buyer or seller initiated, more generally bid/ask information from the limit order book), Li et al. (2016) consider the model where the noise is partially observed through a parametric function Zτi,n = Xτi,n + i,n = Xτi,n + g(Ii,n , ν) + ˜i,n , where Ii,n is the vector of information at time τi,n and ˜i,n is the noisy part of the original noise i,n . Here again the observation times are assumed to be regular, i.e. τi,n = iT /n. The authors assume that ˜i,n is with mean 0, finite standard deviation and that n Var[˜ i,n ] → √ v, which in turn implies that ˜i,n = Op (1/ n). To embed this assumption in our LPM framework, there is no harm assuming that 1 ˜i,n = n−α v 2 γt , where α ≥ 1/2 and γt are IID with null-mean and unity variance. They estimate ν and the underlying price as N ν̂ = arg min ν n 1X bτ = Zτ − g(Ii,n , ν̂). (∆Zτi,n − ∆g(Ii,n , ν))2 , X i,n i,n 2 i=1 The authors then estimate the integrated volatility with ERVext = Nn Nn X X bτ )2 + 2 b τ ∆X bτ (∆X ∆X i,n i,n i−1,n i=1 i=2 and show the according central limit theory. Under suitable assumptions, they obtain the optimal convergence rate n1/2 and the AVAR when T = 1: Z 1 Z 1 (ERV ) 4 AV AR =6 σt dt + 8v σt2 dt + 8v 2 . 0 0 They also consider another estimator (which they call E-QMLE) which consists in using the QMLE from Xiu (2010), which we considered as a local estimator in Example 4.1, on the bτ . They indicate that the E-QMLE might yield a smaller AVAR estimated observations X i,n (see their discussion on p. 38), and they report in their numerical study that its finite sample performance is comparable to ERVext (see Table 2 in p. 41). They don’t investigate the corresponding central limit theory. With the theory provided in our paper, we cannot investigate the E-QMLE, but rather the bτ . To keep notation of our paper, we E-(LPE of QMLE), i.e. we apply Example 4.1 on X i,n 20 b n the E-(LPE of QMLE) estimator of volatility and Θ b n(BC) its bias-corrected version denote Θ (i.e. E-(BC LPE of QMLE)). The AVARs obtained in Theorem 5 are the same as in Theorem 3. This is due to the fact that the estimation of ν is very accurate featuring n as a rate of convergence and thus the pre-estimation doesn’t impact the AVAR. This was already the case for the ERVext (see the proof of Theorem 3 on pp. 46-47 in Li et al. (2016)). Recalling that the LPE of QMLE is conjectured to be more efficient than the QMLE, in particular this implies that E-(LPE of QMLE) is also conjectured to be more efficient than EQMLE. In Figure 1, we can see that E-(LPE of QMLE) highly improves the AVAR compared to the ERVext . The improvement gets bigger as the noise of ˜i,n increases. When setting the volatility and the noise variance as in the setting of the numerical study in Li et al. (2016), the ratio of AVARS is equal to 0.7. When we further assume no jumps in volatility, this ratio goes to 0.2. When choosing a bigger noise variance 1.44e − 07 which remains reasonable, this ratio is lower than 0.01. The overall picture is clearly in favor of the E-(LPE of QMLE). We provide the theorem of this estimator in what follows. Theorem 5. (E-(LPE of QMLE)) Under Assumption A in Li et al. (2016, p. 7): (i) we assume that α > 21 . Then, stably in law as n → ∞, n 1 2  b n − T −1 Θ Z T  σs2 ds  → 6T −1 0 T 1 2 σs4 ds N (0, 1). (45) p   12 4v + σs2 ds N (0, 1). (46) Z 0 (ii) when α = 12 , we have stable convergence in law of n 1 2   b (BC) − Θ → T −2 Θ n T Z 2σs4 + 4σs3 0 We discuss now briefly how to estimate the higher powers of volatility, i.e. when θt∗ = g(σt2 ) with g not being the identity function. We consider the estimator from Example 4.2, i.e. the 2 ), where σ̂ 2 is the QMLE estimate of the volatility on the ith block. local estimation g(σ̂i,n i,n bτ The difference with Example 4.2 is that this estimator is used on the estimated price X i,n based on the information. The theorem is given in what follows. Theorem 6. (powers of volatility) Under Assumption A in Li et al. (2016, p. 7): (i) We assume that α > 12 . Then, stably in law as n → ∞, n 1 2   b n − Θ → 6T −1 Θ T Z 1 2 2 g 0 (σs2 ) σs4 ds N (0, 1). (47) 0 (ii) When α = 12 , we have n 1 2   b (BC) Θ − Θ → T −1 n Z T g 0 (σs2 ) 2σs4 + 4σs3 0 21 p   4v + σs2 ds N (0, 1). (48) 4.5 Estimation of volatility using the model with uncertainty zones We introduce a time-varying friction parameter extension to the model with uncertainty zones introduced in Robert and Rosenbaum (2011) to estimate volatility. There is microstructure noise in this model. Let Zτi,n ,n be the last traded price. In a frictionless market, we can assume that a trade with 1-tick increment of the price will occur when the efficient price process crosses the mid-tick value Zτi,n ,n + α2n (and a trade with 1-tick decrement of the price when crossing the value Zτi,n ,n − α2n ). In practice, the absolute value of the increment (or the decrement) of the observed price can be bigger than the tick size αn . For that reason, the authors introduced the discrete variables Li,n that stand for the absolute size, in tick number, of the next price change. Thus, the next traded price which corresponds to a price change will be Zτi+1,n = Zτi,n ± αn Li,n . They also introduced a continuous (possibly multidimensional) time-varying parameter χt , and assume that conditional on the past, Li,n take values on {1, · · · , m} with Pτi,n (Li,n = k) = pk (χτi,n ) , k = 1, · · · , m, for some unknown positive differentiable with bounded derivative functions pk . We assume that the time-varying parameter is a null-drift Itô-process dχt := σtχ dWtχ , where the volatility σtχ is continuous. Also, the frictions induce that the transactions will not occur exactly when the efficient process is equal to the mid-tick values. For this purpose, in the notation of Robert and Rosenbaum (2012), let 0 < η < 1 be a parameter that quantifies the aversion to price change. The frictionless non-realistic case would correspond to η := 0. Conversely, if η is closer to 1, (α) the agents are very averse to trade. If we let Xt be the value of Xt rounded to the nearest multiple of α, the sampling times are defined recursively as τ0,n := 0 and for any positive integer i as n  1 n) τi,n := inf t > τi−1,n : Xt = Xτ(α − αn Li−1,n − + η i−1,n 2 o  1 n) or Xt = Xτ(α + αn Li−1,n − + η . i−1,n 2 Correspondingly, the observed price is assumed to be equal to the rounded efficient price (α ) Zτi,n := Xτi,nn . To express the model with uncertainty zones as a LPM, we consider that θt∗ := (σt2 , η, χt ). We are not interested in estimating directly χt and thus we follow Remark 3 and consider the subparameter ξt∗ = (σt , η) to be estimated. Following the setting of this paper, we will assume that η is a time-varying parameter ηt and we will extend the model with uncertainty zones in (49). We call this extension the time-varying friction parameter model with uncertainty zones. The sampling times are defined recursively as τi,n := 0 and for any positive integer i n  1 n) τi,n := inf t > τi−1,n : Xt = Xτ(α − αn Li,n − + ητi−1,n i−1,n 2 22 n) or Xt = Xτ(α + αn Li,n − i−1,n o 1 + ητi−1,n . 2 (49) The idea behind the time-varying friction model with uncertainty zones is that we hold the parameter ηt constant between two observations. RT RT The target quantity is the integrated parameter Θ := ( 0 σt2 dt, 0 ηt dt). We define the unobserved returns as Qi,n := Li,n and the observed returns as Ri,n := (∆Zτi,n , ∆τi,n ). The order of memory is m := 1. The random innovation is defined as the two-dimensional path Ui,n := ((∆W[τi−1,n ,t] )t≥τi−1,n , (∆W[τ0 i−1,n ,t] )t≥τi−1,n ) where the jump size Li,n is a function of Wt0 , which is a Brownian motion independent of the other quantities and which was introduced in p.11 of Robert and Rosenbaum (2012). We also define Z t Z t o n  (1) σs du(1) = −α q − 1 + 2η or σ du = α q w+ := inf t > 0 : n 0 s n s s 0 and 0 Z t Z t n o (1) w := inf t > 0 : σs dus = −αn q or σs du(1) = α q − 1 + 2η . n 0 s − 0 0 Furthermore, we define w := w+ 1{r(1) >0} + w− 1{r(1) <0} and y := m X k−1 k n  hX  X io − 12 k1 φ u(2) w ∈ p χ , p χ j w j w w j=1 k=1 (50) j=1 where φ is the cumulative function of the standard normal distribution. The right hand-side in (50) can be found in p.11 of Robert and Rosenbaum (2012). We can express Z w    (1) (2) Fn (q, r(1) , r(2) ), (ut , ut ), (σt2 , ηt , χt ) := y, αn q sign σs du(1) ,w . s 0 We estimate on each block the volatility and the friction parameter using respectively a d αn ,t and η̂αn ,t (see p. 8 in Robert and Rosenbaum slight modification of the estimator RV d αn ,t depends on η̂αn ,t , which we denotes (2012) for definitions). Actually, we can see that RV d αn ,t (η̂αn ,t ) in what follows. First, we define the new estimator of η: RV (mod) η̂α,t := m X (mod) (mod) λα,t,k uα,t,k (51) k=1 with (a) (mod) λα,t,k (c) Nα,t,k + Nα,t,k := P m j=1 (a) (c) Nα,t,j + Nα,t,j  and (mod) uα,t,k !  (c)   1  Nα,t,k := max 0, min 1, k −1 +1 , (a) 2 Nα,t,k 23 (mod) (a) where we assume that C/0 := ∞, and in particular uα,t,k = 1 when Nα,t,k = 0. Second, we define d (mod) = RV d αn ,t (η̂ (mod) ). RV αn ,t αn ,t We choose to work with this modified estimator for two reasons. First, the definition of η̂α,t (a) was unclear when Nα,t,k = 0 in p. 8 of Robert and Rosenbaum (2012). Second, the finite (m) sample bias of η̂α,t is slightly smaller with the modified estimator. Asymptotically, they are equivalent and thus all the theory provided in Robert and Rosenbaum (2012) can be used to prove the following theorem. Theorem 7. (Time-varying friction parameter model with uncertainty zones) Stably in law as n → ∞, αn−1   b n − Θ → T −1 Θ Z 0 T 1 2 Vθs∗ ds N (0, 1), (52) where Vθ can be straightforwardly inferred from the definition of Lemma 4.19 in p. 26 of Robert and Rosenbaum (2012). Remark 5. (convergence rate) Note that, equivalently, the convergence rate in (52) is the usual 1 n 2 when n corresponds to the expected number of observations. One can consult Remark 4 in Potiron and Mykland (2017) for more details. 4.6 Application in time series: the time-varying MA(1) We first specify the LPM for a one-dimensional time series. In that case, we assume that the observation times are regular of the form τi,n = ni T , and thus that the number of observations is Nn = n. We assume that the "returns" Ri,n stand for time series observations. Recalling that Mi,n = (Qi,n , Ri,n ), we assume that the time-varying time series can be expressed as the interpolation of θt∗ via Mi,n = Fn (Mi−1,n , Ui,n , θτ∗i−1,n ), (53) where θt∗ is assumed to be independent of all the innovations. Numerous time series10 are of the parametric form (3) and admits a time-varying parameter extension following (53). We provide details for the MA(1). In this specific example, several time-varying extensions are possible and we choose to work with the time-varying parameter model Ri,n = µτi−1,n + √ √ κτi−1,n λi,n + βτi−1,n κτi−1,n λi−1,n , where λi,n are standard normally-distributed white noise error terms, and κt is the time-varying variance. The three-dimensional parameter is defined as θt∗ := (µt , βt , κt ) ∈ R2 × R+ ∗. 10 We can actually show that any time series in state space form can be expressed with a corresponding Fn function. 24 To express the MA(1) as the parametric model (3) and a time-varying parameter model (4), we fix both the unobserved quantity and innovation equal to the white noise Qi,n = λi,n and Ui,n = λi,n , and the order of the model m = 1. We can define now the function  √ √ Fn m := (q, r), u, (µ, β, κ) = (q, µ + κu + β κq). We have thus expressed the MA(1) as a LPM. We discuss now how to estimate the parameters. ForRsimplicity, R T we assume in what follows T that µt = 0. The target quantity is thus equal to Θ = ( 0 βt dt, 0 κt dt). The local estimator is the regular MLE applied on the MA(1) model (Hamilton (1994), Section 5.4). On each block (of size hn ), the MLE bias is of order h−1 n (Tanaka (1984)) and thus the bias conditions (34), (36) and (38) are not satisfied. Nonetheless, we can correct for the bias up to the order O(h−2 n ) (BC) b b b as follows. We define the bias-corrected estimator as Θi,n = Θi,n − b(Θi,n , hn ), where the bias function b(θ, h) can be derived following the techniques in Tanaka (1984). In particular b (BC) satisfies the bias conditions if hn is chosen such that n1/4 = o(hn ). In this implies that Θ i,n practice this bias can be obtained by Monte-Carlo simulations (see our simulation study). In the parametric case and in a low frequency asymptotics where T → ∞ and observations times are 0, ∆, · · · , T = n∆ with ∆ > 0, known results (see, e.g., the proof of Proposition I in bκ pp. 391-393 (Ait-Sahalia et al. (2005)) show that the asymptotic variance of the MLE (β, b) is such that  1/2 1 − β2 0 1/2 b n ((β, κ b) − (β, κ)) → N (0, 1). 0 2κ2 The following theorem provides the time-varying version of the asymptotic theory when T is fixed. Theorem 8. (Time-varying MA(1)) We assume that n1/4 = o(hn ). Then, stably in law as n → ∞, n 5 1 2  b (BC) Θ −Θ → n T −1 RT 0 (1 − βs2 )ds 0 0 RT 0 ! !1 2κ2s ds 2 N (0, 1). Numerical study in the time-varying MA(1) 5.1 Goal of the study To investigate the finite sample performance of the LPE, we consider the time-varying MA(1) with null-mean introduced in Section 4.6. The goal of the study is twofold. First, we want to investigate how the LPE performs compared to some standard concurrent approaches. Second, we want to discuss about the choice of the tuning parameter hn in practice. We consider the following simple concurrent approaches: 25 MLE : the global MLE when considering that the parameters are not time-varying on [0, T ]. Fitting Recent Observations (FRO): This approach consists in fitting the MLE on a recent sub-block with less observations (e.g. on [TF , T ] where TF > 0) so that the parameter is roughly constant on that block. LPE h̃n : This approach consists in using the LPE (of MLE) with another choice of h̃n as block size. b n(BC) = Θ b i,n − b(Θ b i,n , hn ), we can either compute and To compute the BC estimator Θ implement the function b(θ, h) or carry out Monte-Carlo simulations to compute b(θ, h) for any (θ, h) prior to the numerical study. We choose the latter option as this allows to get also rid of bias terms which appears in the Taylor expansion in a higher order than O(h−2 ). Indeed, although those terms vanish asymptotically, they can pop up in a finite sample context. To be more specific, we first compute the sample mean for a grid of parameter values and block length (θ, h) with, say, 100,000 Monte Carlo paths11 of the parametric model. Then on each b i,n , hn ). block, we estimate the bias by b(Θ We discuss here what we expect theoretically from the bias-correction. In view of the b i,n on first approximation as the sum decomposition (13), we can disentangle the bias of Θ of two terms, namely the bias of the parametric estimator and the bias due to the fact that the parameter is time-varying. The former can be corrected by the econometrician, and we b (BC) accordingly. On the contrary, as the path of the define the bias-corrected local estimate Θ i,n parameter is unknown, we cannot correct for the latter. This is one reason why we have to 1 work with (up to constant terms) a hn < n 2 . The theory shows that the normalized latter bias will vanish asymptotically under that condition. Conversely, the econometrician who chooses 1 to work locally with hn > n 2 will most likely obtain a significant latter bias which she cannot identify, and correcting for the former bias might not improve the estimation in that case. 5.2 Model design We recall that the time-varying parameter is θt∗ = (βt , κt ). We set T = 1, which stands for one day (or one week, one month). We fix the number of observations n = 10000. We consider one toy model where the parameters move around a target parameter deterministically. We assume that the noisy parameter follows a cos function θt∗ = ν + A cos( 2πtδ T ), where ν = (β, κ) (β) (κ) is the parameter, A = (A , A ) corresponds to the amplitude, and δ = (δ (β) , δ (κ) ) stands for the number of oscillations on [0, T ]. It is clear that Θ = (β, κ). We fix the parameter ν = (0.5, 1) and the amplitude A = (0.2, 0.4). We also set a setting with a small number of oscillations δ = (4, 4) and one with a bigger number of oscillations δ = (10, 10). We simulate M = 1, 000 Monte-carlo paths. 11 Actually, the number of Monte Carlo paths can be significantly lower when h increases 26 In view of Theorem 8 and Condition (E4), the tuning parameter hn should (up to constant terms) satisfy n1/4 < hn < n1/2 . In our case, we have that n1/4 = 10 and n1/2 = 100. We decide thus to choose hn = 100 (the results would be similar with hn around 100). For the concurrent approach, we let h̃n ∈ {25, 500, 1000, 2000, 5000}. For the FRO approach, we choose TF = 0.95 (this choice is arbitrary but different values would yield to similar results), which means that we consider the last 500 observations to fit the MLE. 5.3 Results The results are reported in Table 1 when δ = (4, 4) and Table 2 when δ = (10, 10). First, note that there is not much qualitative difference between the two settings. Second, as expected from the theory, the LPE performs at its best with the choice hn = 100, and the bias-corrected version is much better. Moreover, the LPE100 outperforms all the concurrent approaches. The case h̃n = 25 allows us to check what can happen when we have blocks with very few observations. The bias-corrected estimator performs well to estimate κ, but somehow the bias-correction to estimate β doesn’t provide better estimates. This is most likely due to the fact that we have not enough observations on each block. The estimation made with h̃n = 500 is very decent in the case where we don’t have many oscillations. The bias-corrected estimator is actually not as good. This corroborates the theory that when hn >> 100 the main source of bias is due to the parameter which is time-varying rather than the parametric estimator bias itself. If we have a bigger number of oscillations, the estimates are not as accurate. When using bigger hn , we see the same pattern, and the accuracy of the estimation decreases as hn increases. The global MLE performs relatively good to estimate β, but have a strong bias in κ. This indicates that even in a simple deterministic model which oscillates around the target value, the MLE can’t be trusted. Finally, the FRO is far off. There are several problems coming with that approach. The estimated quantity is θT∗ , which can be much different from ν. Also, the block size to use is arbitrary, and the standard deviation is bigger. Remark 6. (block size) Condition (E4) provides the asymptotic order to use for the tuning parameter hn . Thus, it gives a rule of thumb to use on finite sample, but it is left to the practitioner to ultimitalely choose hn . If the parametric estimator is badly biased, the practitioner should increase the value of hn . Also, if the parameter process θt∗ seems not to be moving a lot, hn can be chosen to be bigger. In our simulation study, this rule of thumb could be trusted. In our empirical illustration, we can see that the estimated volatility is robust to the 1/2 value of hn if we choose hn ≈ Nn . As n can be chosen such that n = Nn , this indicates that the rule of thumb seems to be robust to the actual choice of hn in our empirical study too. 27 6 Empirical illustration in the model with uncertainty zones In this section, we are interested in estimating the integrated parameter in the time-varying friction parameter model with uncertainty zones introduced in Section 4.5. We recall that the parameter of interest is defined as ξt∗ = (σt2 , ηt ). We are looking at Orange (ORA.PA) stock price traded actively on the CAC 40 on Monday March 4th, 2013. To prevent from opening and closing effect, we assume that T := 1, consider that t = 0 corresponds to 9am and that t = T stands for 4pm. The number of trades corresponding to a change of price during this time period is equal to Nn = 3306, and the tick size αn = 0.001 euro. When assuming the non time-varying model with friction parameter equal to η, we can (mod) estimate the friction parameter with η̂αn ,T . We estimate the standard deviation as (mod) ŝn := ŝn (η̂α,T ), where the formal definition of the standard deviation sn (η) and the standard deviation estimate (mod) ŝn (η) can be found in Appendix II Section 7.12. We find that η̂αn ,T = 0.155 and ŝn = 0.008. (mod) Note that in finite sample, η̂αn ,T is biased and that this bias can be estimated following the estimate B̂ introduced in Appendix II Section 7.12. Because we choose hn big enough in this section, the bias is very small and will not alter the values of the estimators and the results of the tests in what follows. b i,n := (σ̂ 2 , η̂i,n ) the estimate on the ith block. We also define the standard We define Θ i,n deviation estimate of η̂i,n as ŝi,hn := si,hn (η̂i,n ). Note that ŝi,hn is not block-dependent except for the last block, which is removed from the analysis in the following because the number of observations can be too small in that block. We introduce the notation ŝhn := ŝ1,hn . Figure 2 shows the evolution over time of η̂i,n for different values of hn . Based on those estimates and the standard deviation estimate ŝhn , we compute the associated chi-square statistic χ2n := BX n −1  (mod) η̂i,n − η̂αn ,T 2 ŝhn i=1 . Under the null hypothesis which states that ηt is non time-varying, χ2n follows approximately a chi-square distribution with Bn − 1 degrees of freedom. We report χ2n for different values of hn in Table 3. The obtained values indicate that we have strong evidence against the null hypothesis, thus it provides us very good reasons to use the techniques of this paper. This analysis has been carried out on other days and other stocks. Our conclusion is that in almost every situation, the friction parameter is time-varying. Finally, using bootstrap values instead of the theoretical normal approximation wouldn’t change the picture. b n := (σ̂n2 , η̂n ) for different values of hn following the time-varying We now compute Θ 1/2 friction parameter model with uncertainty zones. Because Nn ≈ 57.5 we choose to work with hn = 43, · · · , 63. We can see in Figure 3 that we obtain different estimates of volatility using 28 the techniques of this work compared to the estimates of the model with uncertainty zones, which is one reason why it is crucial to use a proper time-varying model for ηt . The estimates of the model with uncertainty zones seems to underestimate the integrated volatility. In addition, the RV estimator, which ignores the microstructure noise, seems to be overestimating the integrated volatility. This is what to be expected in theory and thus indicates that the correction made by σ̂n is reasonable. Finally, the estimates are very similar for different values of hn , which seems to indicate that the method is robust to small block size variation. 7 Conclusion In this paper, we have introduced a general framework to provide theoretical tools to build central limit theorems of convergence rate n1/2 . We have applied successfully the method to investigate estimation of volatility (possibly under trading information), higher powers of volatility, high-frequency covariance, and time-varying parameters in the model with uncertainty zones and time-varying parameters of the MA(1). This allowed us to obtain estimators robust to time-varying quantities, more efficient and/or new estimators of quantities (such as in the case of higher powers of volatility). Subsequently, we believe that many other examples can be solved using the framework of our paper, which is simple and natural, either as a direct corollary of the respective central limit theory work (as in the case of all the examples discussed in our paper) or as a basis of work for the proofs. The latter was successfully done in our related papers Potiron and Mykland (2017) and Clinet and Potiron (2017a). This was possible as the framework is transparent and the conditions can be altered easily in view of the direct proofs. For this reason, we have chosen to allow for general models possibly including endogeneity, microstructure noise correlated with the other quantities, etc. References [1] Aït-Sahalia, Y., J. Fan, C.D. Wang and X. Yang (2015). Estimation of the continuous and discontinuous leverage effects, Available at SSRN. [2] Aït-Sahalia, Y., P.A. Mykland and L. Zhang (2005). How often to sample a continuoustime process in the presence of market microstructure noise, Review of Financial Studies 18, 351-416. [3] Andersen, T.G., D. Dobrislav and E. Schaumburg (2014). A robust neighborhood truncation approach to estimation of integrated quarticity, Econometric Theory,30(1),3-59. [4] Barndorff-Nielsen, O.E. and N. Shephard (2002). Econometric analysis of realized volatility and its use in estimating stochastic volatility models. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 64.2, 253-280. 29 [5] Barndorff-Nielsen, O.E., S.E. Graversen, J. Jacod, M. Podolskij and N. Shephard (2006). A central limit theorem for realised power and bipower variations of continuous semimartingales. From stochastic calculus to mathematical finance, Springer Berlin Heidelberg, 33-68. [6] Barndorff-Nielsen, O.E., P.R. Hansen, A. Lunde, and N. Shephard (2008). Designing realized kernels to measure the ex post variation of equity prices in the presence of noise. Econometrica 76(6), 1481-1536. [7] Clinet, S. and Y. Potiron (2017a). Estimating the Integrated Parameter of the TimeVarying Parameter Self-Exciting Process. arXiv preprint arXiv:1607.05831. [8] Clinet, S. and Y. Potiron (2017b). Efficient asymptotic variance reduction when estimating volatility in high frequency data. arXiv preprint arXiv:1701.01185. [9] Dahlhaus, R. (1997). Fitting time series models to nonstationary processes, Ann. Statist. 25(1), 1-37. [10] Dahlhaus, R. (2000). A likelihood approximation for locally stationary processes, Ann. Statist., 1762-1794. [11] Dahlhaus, R. and S.S. Rao (2006). Statistical inference for time-varying ARCH processes, Ann. Statist. 34(3), 1075-1114. [12] Fan, J. and I. Gijbels (1996). Local polynomial modelling and its applications: monographs on statistics and applied probability 66. Vol. 66. CRC Press. [13] Fan, J. and W. Zhang (1999). Statistical estimation in varying coefficient models. Ann. Statist., 1491-1518. [14] Gloter, A. and J. Jacod (2001). Diffusions with measurement errors. I. Local asymptotic normality. ESAIM: Probability and Statistics 5, 225-242. [15] Hall, P. and C.C. Heyde (1980). Martingale Limit Theory and its Application. Academic Press, Boston. [16] Hamilton, J.D. (1994). Time series analysis. Vol. 2. Princeton: Princeton university press. [17] Hastie T. and R. Tibshirani (1993). Varying-coefficient models. J. R. Stat. Soc. Ser. B. Stat. Methodol. 757-796. [18] Hayashi, T. and N. Yoshida (2005). On covariance estimation of non-synchronously observed diffusion processes. Bernoulli 11, 359-379. [19] Jacod, J., M. Podolskij and M. Vetter (2010). Limit theorems for moving averages of discretized processes plus noise. The Annals of Statistics 38.3, 1478-1545. 30 [20] Jacod, J. and P. Protter (1998). Asymptotic error distributions for the Euler method for stochastic differential equations. Ann. Probab. 26, 267-307. [21] Jacod, J. and P. Protter (2012). Discretization of Processes. Springer. [22] Jacod, J. and M. Rosenbaum (2013). Quarticity and other functionals of volatility: efficient estimation. Ann. Statist. 41(3), 1462-1484. [23] Jacod, J. and A. Shiryaev (2003). Limit Theorems For Stochastic Processes (2nd ed.). Berlin: Springer-Verlag. [24] Kim, C. J. and Nelson, C. R. (2006). Estimation of a forward-looking monetary policy rule: A time-varying parameter model using ex post data. Journal of Monetary Economics, 53(8), 1949-1966. [25] Kristensen, D. (2010). Nonparametric filtering of the realized spot volatility: A kernelbased approach. Econometric Theory, 26.01, 60-93. [26] Li, Y., S. Xie and X. Zheng (2016). Efficient estimation of integrated volatility incorporating trading information. Journal of econometrics 195.1, 33-50. [27] Mancino, M.E. and S. Sanfelici (2012). Estimation of quarticity with high-frequency data. Quantitative finance, 12(4),607-622. [28] Mykland, P.A., and L. Zhang (2006) ANOVA for Diffusions and Itô Processes. Annals of Statistics 34, 1931-1963. [29] Mykland, P.A. and L. Zhang (2009). Inference for Continuous Semimartingales Observed at High Frequency. Econometrica 77, 1403-1445. [30] Mykland, P.A. and L. Zhang (2011). The Double Gaussian Approximation for High Frequency Data. Scand. J. Stat. 38, 215-236. [31] Mykland, P.A. and L. Zhang (2012). The econometrics of High Frequency Data. In M. Kessler, A. Lindner and M. Sørensen (eds.), Statistical Methods for Stochastic Differential Equations, 109-190. Chapman nad Hall/CRC Press. [32] Mykland, P.A. and L. Zhang (2016). Assessment of Uncertainty in High Frequency Data: The Observed Asymptotic Variance. Forthcoming in Econometrica. Available at SSRN 2475620. [33] Potiron, Y. and P.A. Mykland (2017). Estimation of integrated quadratic covariation with endogenous sampling times. Journal of Econometrics, 197, 20-41. [34] Reiß, M (2011). Asymptotic equivalence for inference on the volatility from noisy observations. Ann. Statist. 39(2), 772-802. 31 [35] Renault, E., C. Sarisoy and B.J.M. Werker (2012). Efficient estimation of integrated volatility and related processes. Econometric Theory. 1-40. [36] Revuz, D. and M. Yor (1999) Continuous Martingales and Brownian motion. 3rd ed., Germany: Springer. [37] Robert, C.Y. and M. Rosenbaum (2011). A new approach for the dynamics of ultra-highfrequency data: the model with uncertainty zones. Journal of Financial Econometrics 9, 344-366. [38] Robert, C.Y. and M. Rosenbaum (2012). Volatility and covariation estimation when microstructure noise and trading times are endogenous. Math. Finance 22(1), 133-164. [39] Stock, J. H. and M. W. Watson (1998). Median unbiased estimation of coefficient variance in a time-varying parameter model. J. Amer. Statist. Assoc., 93(441), 349-358. [40] Tanaka K. (1984). An asymptotic expansion associated with the maximum likelihood estimators in ARMA models. J. R. Stat. Soc. Ser. B. Stat. Methodol. 58-67. [41] Wang, C.D. and P.A. Mykland (2014). The Estimation of Leverage Effect with High Frequency Data. J. Amer. Statist. Assoc., 109, 197-215. [42] Xiu, D. (2010). Quasi-maximum likelihood estimation of volatility with high frequency data. J. Econometrics 159, 235-250 . 32 Appendix I: consistency in a simple model The purpose of this appendix is to provide an outline of the LPM and the conditions of the CLT by investigating the simpler problem of consistency in the case of a simple model. The reader should refer to it for a deeper understanding of Section 2 and Section 3. Two toy examples, the estimation of volatility with regular not noisy observations, and the estimation of the rate of a Poisson process, are discussed extensively. Morever, techniques of proofs are also mentioned throughout the section. The obtained conditions are illustrative. Proofs of the conditions along with proofs that conditions hold in the two toy examples can be found in Appendix II. Finally, some detailed mathematical definitions of the objects can be found throughout this part too. The simple model We focus on a simple setting in this section. First, we work with dr := 2. Also, we assume that the observations occur at equidistant time intervals ∆τn := Tn , so that τi,n = ni T and (2) thus Ri,n := ∆τn is deterministic. For the rest of this appendix, we will forget about the second component of the returns Ri,n , which doesn’t provide us any further information, and pretend that Ri,n is real-valued. The parametric model is assumed to be very simple. It assumes that there exists a parameter θ∗ ∈ K such that Ri,n are independent and identically distributed (IID) random functions of θ∗ . If we introduce Ui,n an adequate IID sequence of random variables with distribution U , we can express the returns as  Ri,n := Fn Ui,n , θ∗ , (54) where Fn (x, y) is a non-random function. In (54), Ui,n can be seen as the random innovation. Since θt∗ can in fact be time-varying, Ri,n don’t necessarily follow (54) in the time-varying parameter model. A formal time-varying generalization of (54) will be given in (57). In general, Ri,n are neither identically distributed nor independent. Ri,n are not even necessarily conditionally independent given the true parameter process θt∗ , as we can see in the following two toy examples. Example 1. (estimatingRvolatility) Consider when θt∗ := σt2 (the volatility is thus assumed to τi,n follow (2)), and Ri,n := τi−1,n σs dWs , where Wt is a standard 1-dimensional Brownian motion. ∗ 2 In this case, the parameter space is K := R+ ∗ . The parametric model assumes θ := σ and that the distribution of the returns is Ri,n := σ∆Wτi,n , where ∆Wτi,n := Wτi,n − Wτi−1,n is the increment of the Brownian motion between the (i − 1)th observation time and the ith observation time and σ 2 is the fixed volatility. Under that assumption, the returns are IID. Under the time-varying parameter model, Ri,n are clearly not necessarily IID, and they are also not necessarily conditionally independent given the whole volatility process σt2 if there is a leverage effect. Example 2. (estimating the rate of a Poisson process) Suppose an econometrician observes data on the number of events (such as trades) in an arbitrary asset, and thinks the number 33 of events happening between 0 and t, Nt , follows a homogeneous Poisson process with rate λ. The parameter rate θt∗ := λt will be assumed to follow (2), with possibly a null-volatility σtθ = 0 if the homogeneity assumption turns out to be true. Because the econometrician doesn’t have access to the raw data, she can’t observe directly the exact time of each event. Instead, she only observes the number of events happening on a period (for instance a tenminute block) [τi−1,n , τi,n ), that is Ri,n = Nτ−i,n −Nτi−1,n . If the econometrician’s assumption of homogeneity is true, the returns are IID. In case of heterogeneity, Nt will be a inhomogeneous Poisson process, and the returns Ri,n will most likely be neither identically distributed nor independent. We need to introduce some notation and definitions. On a given block i = 1, · · · , Bn the 1 , · · · , Rhn . Formally, it means that Rj := R observed returns will be called Ri,n (i−1)hn +j,n for i,n i,n j 1 , · · · , R̃hn any j = 1, · · · , hn . In analogy with Ri,n , we introduce the approximated returns R̃i,n i,n j on the ith block. We also introduce the corresponding observation times τi,n := τ(i−1)hn +j,n for hn 0 j = 0, · · · , hn . Note that τi,n = τi−1,n . Finally, for j = 1, · · · , hn we define the time increment j j j−1 between the (j − 1)th return and the jth return of the ith block as ∆τi,n := τi,n − τi,n . We provide a time-varying generalization of the parametric model (54) as well as a formal expression for the approximated returns. To deal with the former, we assume that in general  Ri,n := Fn Ui,n , {θs∗ }τi−1,n ≤s≤τi,n . (55) The time-varying parameter model in (55) is a natural extension of the parametric model (54) because the returns Ri,n can depend on the parameter process path from the previous sampling time τi−1,n to the current sampling time τi,n . As Ri,n depend on the parameter path, it seems natural to allow Ui,n to be themselves process paths. For example, when the parameter is equal to the volatility process θt∗ := σt2 , we will assume that Ui,n are equal to the underlying Brownian motion Wt path (see Example 3 for more details). Also, as Ui,n are random innovation, they should be independent of the parameter process path past, but not on the current parameter path. In the case of volatility, it means that we allow for the leverage effect. A simple particular case of (55) is given by  Ri,n := Fn Ui,n , θτ∗i−1,n , (56) i.e. the returns depend on the parameter path only through its initial value. The time-varying friction parameter model with uncertainty zones introduced in Section 4.5 and the time-varying Hitting Constant Boundaries model defined in Appendix II and are defined as a mix of (55) and (56). Finally, the approximated returns R̃i,n follow a mixture of the parametric model (54) with initial block parameter value. We are now providing a formal definition of our intuition. We assume that  j j Ri,n := Fn Ui,n , {θs∗ }τ j−1 ≤s≤τ j , (57) i,n i,n  j j R̃i,n := Fn Ui,n , Θ̃i,n , (58) 34 j where the random innovation Ui,n take values on a space Un that can be functional12 and j that can depend on n, Ui,n are IID for a fixed n but the distribution can depend on n, and Fn (x, y) is a non-random function13 . Note that (57) is a mere re-expression of (55) using a different notation. For any block i = 1, · · · , Bn and for any observation time j = 0, · · · , hn of j j 14 the information up to time τi,n . The crucial assumption is that the ith block, we define Ii,n j Ui,n has to be independent of the past information15 (and in particular of Θ̃i,n ). Note that j we don’t assume any independence between the random innovation Ui,n and the parameter j process {θs∗ }τ j−1 ≤s≤τ j . We provide directly the definitions of Fn and Ui,n in the two toy i,n i,n examples. Example 3. (estimating volatility) In this case, Un is defined as the space C1 [0, ∆τn ] of continj uous paths parametrized by time t ∈ [0, τn ], Ui,n := {∆W[τ j−1 ,s] }τ j−1 ≤s≤τ j are the Brownian i,n i,n i,n motion increment path processes between two consecutive observation times. We assume that (Wtθ , Wt ) is jointly a (possibly non-standard) 2-dimensional Brownian motion. Thus, the ranj dom innovation Ui,n are indeed independent of the past in view of the Markov property of Rτ 1 Brownian motions. We also define Fn (ut , θt ) := 0 n θs2 dus . We thus obtain that the returns j R τi,n j j are defined as Ri,n := j−1 σs dWs and that the approximated returns R̃i,n := στ 0 ∆W[τ j−1 ,τ j ] τi,n i,n i,n i,n are the same quantity when holding the volatility constant on the block. Example 4. (estimating the rate of a Poisson process) We assume that the rate of the (possibly inhomogeneous) Poisson process is αn λt , where αn is a non time-varying and non-random quantity such that αn ∆τn := 1. In this case, we assume that Un is the space of increasing paths on R+ starting from 0 which takes values in N and whose jumps are equal to 1. We also assume that for any path in Un , the number of jumps is finite on any compact of R+ . j Ui,n can be defined as standard Poisson processes {Nti,j,n }t≥0 , independent of each other. We j R τi,n also have Fn (ut , θt ) := uR τn αn θs dus . Thus, if we let tji,n := j−1 αn λs ds, the returns are the τi,n 0 12 Un is a Borel space, for example the space C1 [0, ∆τn ] of 1-dimensional continuous paths parametrized by time t ∈ [0, τn ]. 13 Let Cp (R+ ) be the space of p-dimensional continuous paths parametrized by time t ∈ R+ , which is a Borel space. Consequently, Un × Cp (R+ ) is also a Borel space. We assume that Fn (x, y) is a jointly measurable real-valued function on Un × Cp (R+ ). Note that the advised reader will have seen that a priori {θs∗ }τ j−1 ≤s≤τ j i,n i,n j−1 is defined on Cp [0, τn ] (after translation of the domain by −τi,n ) in (57) and Θ̃i,n is a vector in (58), whereas + both should be defined on the space Cp (R ) according to the definition. We match the definitions by extending them as continuous paths on R+ . Formally, if θt ∈ Cp [0, τn ], we extend it as θt := θτn for all t > τn . Similarly, if θ ∈ K, we extend it as θt := θ for all t ≥ 0. 14 In what follows, we will be using the term information to refer to the mathematical object of filtration. Let (Ω, F, P ) be a probability space. Define the sorted information {Ik,n }k≥0 such that for any non-negative j integer k that we can decompose as k = (i − 1)hn + j where i ∈ {1, · · · , Bn } and j ∈ {0, · · · , hn }, Ik,n := Ii,n . ∗ We assume that Ik,n is a (discrete-time) filtration on (Ω, F, P ). In addition, we assume that {θs }0≤s≤τ j and i,n j j Ui,n are Ii,n -measurable. j−1 15 past information means up to time τi,n 35 time-changed Poisson processes j Ri,n = N i,j,n , j (59) ti,n j R̃i,n = Nαn ∆τ j i,j,n i,n λτ 0 i,n . (60) Consistency In the following of this section and Section 3, we will make the block size hn go to infinity hn → ∞. (61) Furthermore, we will make the block length ∆Ti,n vanish asymptotically. Because we assumed observations are regular in this section, this can be expressed as hn n−1 → 0. (62) b n as We can rewrite the consistency of Θ Bn X  P b i,n − Θi,n ∆Ti,n → Θ 0. (63) i=1 b i,n can be found in (67). In order to show (63), we can where the formal definition of Θ b i,n − Θi,n ) into the part related to misspecified distribution error, decompose the increments (Θ the part on estimation of approximated returns error and the evolution in the spot parameter error b̃  + Θ b̃ − Θ̃  + Θ̃ − Θ , b i,n − Θi,n = Θ b i,n − Θ Θ (64) i,n i,n i,n i,n i,n b̃ , which is defined formally in (68), is the parametric estimator used on the underwhere Θ i,n lying non-observed approximated returns. It is not a feasible estimator and appears in (64) only to shed light on the way we can obtain the consistency of the estimator in the proofs. We first deal with the last error term in (64), which is due to the non-constancy of the spot parameter θt∗ . Note that Bn X  Θ̃i,n − Θi,n ∆Ti,n = i=1 Bn X ∗ θT ∆Ti,n i−1,n i=1 Z Ti,n −  θs∗ ds (65) Ti−1,n and thus we deduce from Riemann-approximation16 that Bn X  P Θ̃i,n − Θi,n ∆Ti,n → 0. i=1 16 see i.e. Proposition 4.44 in p.51 of Jacod and Shiryaev (2003) 36 (66) To deal with the other terms in (64), we assume that for any positive integer k, the practitioner has at hand an estimator θ̂k,n := θ̂k,n (r1,n ; · · · ; rk,n ), which depends on the input of returns {r1,n ; · · · ; rk,n }. On each block i = 1, · · · , Bn we estimate the local parameter as  b i,n := θ̂h ,n R1 ; · · · ; Rhn . (67) Θ i,n n i,n b̃ is defined as the same parametric estimator with approximated The non-feasible estimator Θ i,n returns as input instead of observed returns  b̃ := θ̂ hn 1 Θ i,n hn ,n R̃i,n ; · · · ; R̃i,n . (68) j Note that (68) is infeasible because the approximated returns R̃i,n are non-observable quantities. Example 5. (estimating volatility) The estimator is the scaled usual RV, i.e. θ̂k,n (r1,n ; · · · ; rk,n ) P 2 . Note that θ̂ := T −1 k −1 n kj=1 rj,n k,n can also be asymptotically seen as the MLE (see the discussion pp. 112-115 in Mykland and Zhang (2012)). Example 6. (estimating the rateP of a Poisson process) The estimator to be used is the return k −1 mean θ̂k,n (r1,n ; · · · ; rk,n ) := k j=1 rj,n . For any M > 0, we introduce the following bounded space for the parameter value. Definition. (KM ) For any M > 0, we define KM := {θ ∈ K : | θ |≤ M and θ+ ≥ M −1 }, which is a bounded subset of K. In order to tackle the second term in (64), we make the assumption that the parametric estimator is L1 -convergent, locally uniformly in the model parameter θ if we actually observe returns coming from the parametric model. This can be expressed in the following condition. Condition (C1). Let the innovation of a block (V1,n , · · · , Vhn ,n ) be IID with distribution U . For any M > 0, h i sup E θ̂hn ,n (Fn (V1,n , θ); · · · ; Fn (Vhn ,n , θ)) − θ → 0. θ∈KM Remark 7. (practicability) Under Condition (C1), results on regular conditional distributions17 give us that the error made on the estimation of the underlying non-observed returns tends to 0, i.e. Bn X P b̃ − Θ̃ ∆T → Θ 0. i,n i,n i,n (69) i=1 This proof technique is the main idea of the paper. Regular conditional distributions are used to deduce results on the time-varying parameter model using uniform results in the parametric 17 see for instance Leo Breiman (1992), see Appendix II Section 7.2 for more details. 37 model. It considerably simplifies the work of the econometrician, in the sense that Condition (C1) is most likely easier to verify than (69). Indeed, Condition (C1) is "parametric" with non-random parameter, whereas (69) is "nonparametric" and each term of the sum Θ̃i,n is a mixture of parameter. Moreover, the uniform convergence in Condition (C1) shouldn’t be an additional problem at all once the econometrician has the (non-uniform) convergence because the parameter θ ∈ KM , which is a bounded space. Remark 8. (consistency) Note that L1 -convergence is slightly stronger than the simple consistency of the parametric estimator. Nonetheless, in most applications, we will have both. To deal with the first term of (64), we need to make sure that we can control the discrepancy between the estimate made on the observed returns and the estimate made on the underlying approximated returns, uniformly in the initial parameter and in the future path of the parameter process. First, we need to introduce a new definition. Definition. (EM,n ) For any M > 0, we define EM,n the product space of all null-drift continuous p-dimensional Itô-process and innovation (θt , Vn ), where Vn := (V1,n , · · · , Vhn ,n ), such that the initial value θ0 is non-random, θt ∈ KM for all 0 ≤ t ≤ T , the volatility of θt is bounded by M for all 0 ≤ t ≤ T , and for any j = 1, · · · , hn we have Vj,n independent of the path past {θs }0≤s≤τ j−1 . 1,n In analogy with KM which is a bounded space for the parameter value, EM,n can be seen as a bounded space for the parameter path (in the parameter value and the parameter volatility at each time 0 ≤ t ≤ T ). Note that each path in EM,n is random (but with non-random starting value), whereas elements of KM are non-random. We can now express the new following assumption. Condition (C2). We have sup   E θ̂hn ,n Fn (V1,n , θ0 ), · · · , Fn (Vhn ,n , θ0 ) (θt ,Vn )∈EM,n −θ̂hn ,n  Fn (V1,n , {θs }0≤s≤τ1,n ) 1 ), · · · , Fn (Vhn ,n , {θs } hn −1 τ ≤s≤τ hn 1,n  P → 0. 1,n Remark 9. (practicability) Condition (C2) implies18 that the error due to the local model approximation vanishes in the limit, i.e. Bn X P b̃ )∆T → b i,n − Θ (Θ 0. i,n i,n (70) i=1 Here again, the main argument of the proof is regular conditional distributions. In analogy with Remark 7, the econometrician is expected to have less trouble verifying Condition (C2) 18 see Appendix II - Section 7.2 for more details 38 than (70). Indeed, the parameter has a non-random initial value (and a random path) in Condition (C2), but a mixture of initial value στθ 0 (and a random path) in (70). Note i−1,n that here again the uniformity in Condition (C2) is not expected to add any difficulty to the problem because we look at expected values and EM,n is a well-chosen bounded space (in the parameter value and the parameter volatility). We can now summarize the theorem on consistency in this very simple case where observations occur at equidistant time intervals and returns are IID under the parametric model. Under Condition (C1) and Condition (C2), we have the consistency of (6), i.e. P bn → Θ. Θ (71) We obtain the consistency in the couple of toy examples19 . Remark 10. (LPE equal to the parametric estimator) The reader will have noticed that in the couple of examples, the LPE is equal to the parametric estimator. This is because in those very basic examples, the parametric estimator is linear, i.e. for any positive integer k and l = 1, · · · , k − 1 θ̂k,n (r1,n ; · · · ; rk,n ) = l k−l θ̂l,n (r1,n ; · · · ; rl,n ) + θ̂k−l,n (rl+1,n ; · · · ; rk,n ) k k (72) In more general examples, (72) will break, and we will obtain two distinct estimators. Challenges There are three empirical reasons why the presentation above is too simple. In practice, the observed returns can be autocorrelated, noisy and there can be endogeneity in sampling times. Accordingly, we build the general LPM in Section 2. Also, we investigate the limit distribution in Section 3. 19 see Appendix II for proofs 39 Appendix II: proofs 7.1 Preliminaries Since | θt∗ | is locally bounded and (θt∗ )+ is locally bounded away from 0, we can follow standard localisation arguments (see, e.g., pp. 160 − 161 of Mykland and Zhang (2012)) and assume without loss of generality that there exists M > 0 such that θt∗ ∈ KM for all 0 ≤ t ≤ T . Furthermore, because we assume that the volatility of the parameter σtθ is locally bounded, we can use the same techniques and assume without loss of generality that there exists σ + > 0 such that σtθ ≤ σ + for all 0 ≤ t ≤ T . Finally, we fix some notation. In the following of this paper, we will be using C for any constant C > 0, where the value can change from one line to the next. 7.2 Proof of the consistency (71) Proof (C1) ⇒(69) It suffices to show that (C1) implies that h i b̃ − Θ̃ sup E Θ = op (1). i,n i,n (73) i≥0 By (58) and (68), we can build gn such that we can write b̃ − Θ̃ hn 1 Θ i,n i,n = gn (Ui,n , · · · , Ui,n , Θ̃i,n ), where gn is a jointly measurable real-valued function such that hn 1 , Θ̃i,n ) < ∞. , · · · , Ui,n E gn (Ui,n We have that ii i h h h hn hn 1 1 E gn (Ui,n , · · · , Ui,n , Θ̃i,n ) = E E gn (Ui,n , · · · , Ui,n , Θ̃i,n ) Θ̃i,n Z  =E gn (u, Θ̃i,n )µω (du) 1 , · · · , U hn ) given Θ̃ where µω (du) is a regular conditional distribution for (Ui,n i,n (see, e.g., i,n Breiman (1992)). From Condition (C1), we obtain (73). Proof (C2) ⇒(70) It is sufficient to show that (C2) implies that h i ˆ sup E Θ̂i,n − Θ̃ = op (1). i,n i≥0 40 (74) (2) By (57), (58), (67) and (68), we can build gn such that we can write ˆ hn (2) 1 ∗ Θ̂i,n − Θ̃ i,n = gn (Ui,n , · · · , Ui,n , {θs }τ 0 0 i−1,n ≤s≤τi,n , Θ̃i,n ). We compute h i h h ii ˆ | = E E g (2) (U 1 , · · · , U hn , {θ∗ } 0 E |Θ̂i,n − Θ̃ 0 , Θ̃i,n )|Θ̃i,n i,n n i,n s τi−1,n ≤s≤τi,n i,n Z  (2) =E gn (v, Θ̃i,n )µω (dv) = op (1). 1 , · · · , U hn , {θ ∗ } where µω (dv) is a regular conditional distribution for (Ui,n s τ0 i,n 0 i−1,n ≤s≤τi,n ) given Θ̃i,n and where we used Condition (C2) in the last equality. 7.3 Proof of Consistency in Example 1 Let’s show Condition (C1) first. For any M > 0, the quantity  θ̂hn ,n Fn (V1,n , θ); · · · ; Fn (Vhn ,n , θ) − θ can be uniformly in {θ ∈ KM } bounded by hn X C (∆V[0;τn ],j,n )2 T −1 − 1 . (75) j=1 We can prove that (75) tends to 0 in probability as a straightforward consequence of Theorem I.4.47 of p.52 in Jacod and Shiryaev (2003). To show Condition (C2), let M > 0. It is sufficient to show that the following quantity " # Z τj hn X 1,n 2 2 −1 nhn E θ0 ∆V[0;τn ],j,n − θs dVs−τ j−1 ,j,n (76) j−1 τ1,n j=1 1,n goes to 0 uniformly in {(θt , Vn ) ∈ EM,n }. Using Conditional Burkholder-Davis-Gundy inequality (BDG, see inequality (2.1.32) of p. 39 in Jacod and Protter (2012)), (76) can be bounded uniformly by " # Z τj hn X 1,n −1 Chn E θ0 ∆V[0;τn ],j,n − θs dVs−τ j−1 ,j,n . (77) j−1 τ1,n j=1 1,n We can also bound uniformly (77) by Ch−1 n hn X  j 1/2  (∆τ1,n ) E {z } | j=1 O(n−1/2 ) | 41  sup j−1 j s∈[τ1,n ,τ1,n ] {z |θ0 − θs |  op (n−1/2 ) } where we used BDG another time to obtain op (n−1/2 ). 7.4 Proof of Consistency in Example 2 Condition (C1) can be shown easily. Similarly, Condition (C2) is a direct consequence of the definition in (59), (60) together with (62). 7.5 Proof of Theorem 1 (central limit theorem) We show (14) We aim to show that Ei,n := Bn X  1 P n 2 Θ̃i,n − Θi,n ∆Ti,n → 0. | {z } i=1 (78) ei,n   Note that ETi−1,n ei,n = 0 and thus that Ei,n is a discrete martingale. We compute the limit  2  P n of n B i=1 ETi−1,n ei,n . We have n Bn X Bn h X   ETi−1,n ETi−1,n e2i,n = n ≤ Cn 2 l Ti,n Ti−1,n i=1 i=1 Z 2 i ∗ (θu∗ − θT )du i−1,n Bn X   1 ETi−1,n (∆Ti,n )4 2 | {z } i=1 Op (h2n n−2 )  ETi−1,n  sup Ti−1,n ≤s≤Ti,n | {z ∗ (θs∗ − θT )4 i−1,n Op (hn n−1 )  12 } = op (1), where we used Conditional Cauchy-Schwarz in the inequality, (19) of Condition (E1) together with BDG  2to obtain the big taus, and (25) in the last equality. Because we showed PBninequality that n i=1 ETi−1,n ei,n tends to 0 in probability, we obtain (78). We show (16) Without loss of generality, we can assume that the prameter θt∗ is a 1-dimensional process. Because for example the parametric estimator can be biased, in all generality, Ai,n :=  1 ˆ M∗ M∗ ,Θ̃ n n 2 Θ̃ − Θ̃ ∆T n i,n is not the increment term of a discrete martingale. Thus, we i,n i,n i,n need first to compensate it in order to apply usual discrete martingale limit theorems. Let Bi,n := Ai,n − ETi−1,n [Ai,n ]. We want to use Corollary 3.1 of pp. 58 − 59 in Hall and Heyde (1980). First, note that by Condition (E0), condition (3.21) in p.58 of Hall and Heyde (1980) 42 is satisfied. We turn now to the two other conditions of the corollary in the two following steps. First condition : We will show in this step that for all  > 0, Bn X h i P 2 ETi−1,n Bi,n 1{Bi,n >} → 0. (79) i=1 The conditional Cauchy-Schwarz inequality gives us that each term of the sum in (79) can be bounded by   4    12 ETi−1,n Bi,n ETi−1,n 1{Bi,n >} . (80) To show that the right term of the sum in the right-hand side of (80) vanishes uniformly, we use regular conditional distribution together with (19), (22) and (25). We apply regular conditional distribution together with (23) on the left term of the sum in the right-hand side of (80). Then, we take the sum of all the terms in the right-hand side of (80) and we can prove (79). Second condition : We will prove that Bn X Z  2  P ETi−1,n Bi,n →T T Vθs∗ ds. 0 i=1 (81) By regular conditional distribution and (22), we have that Bn X Bn X  2   ETi−1,n Bi,n = T ETi−1,n VθT∗ i=1 M∗ ,Θ̃i,n  i−1,n i=1 ∆Ti,nn + op (1). In view of (21), the conditional Cauchy-Schwarz inequality and the boundedness of Vθ , we get Bn X h ETi−1,n VθT∗ i=1 M∗ ,Θ̃i,n i−1,n ∆Ti,nn i 1− l20 = hn 2 n l −1 Bn X h ETi−1,n VθT∗ i=1 i−1,n i ∆Ti,n + op (1). Using Lemma 2.2.11 of Jacod and Protter (2012) together with conditional Cauchy-Schwarz inequality, (19) and the boundedness of Vθ , we obtain T Bn X i=1 h ETi−1,n VθT∗ i−1,n Bn i X 1− 20 2 ∆Ti,n = hn l n l −1 T VθT∗ i=1 i−1,n ∆Ti,n + op (1). We can apply now Proposition I.4.44 (p. 51) in Jacod and Shiryaev (2003) and we get T Bn X i=1 VθT∗ P i−1,n Z ∆Ti,n → T 0 43 T Vθs∗ ds. We are interested in the stable convergence of the sum of Ai,n terms, but by using Corollary 3.1 of pp. 58 − 59 in Hall and Heyde (1980), we only obtain the stable convergence of the increment martingale terms Bi,n . We will show now that the sum of the conditional means   PBn Sn := i=1 ETi−1,n Ai,n tends to 0 in probability. An application of (24) together with regular conditional distribution will give us the convergence to 0 of Sn . We show (15) We want to prove 1 n2 Bn X M∗ ,Θ̃i,n Θ̃i,n ∆Ti,nn  P − ∆Ti,n → 0. (82) i=1 We define the conditional expectation of the terms in the sum of (82): h i M∗ ,Θ̃ Ci,n := ETi−1,n Θ̃i,n ∆Ti,nn i,n − ∆Ti,n . In analogy with the previous part, we can rewrite the term on the left-hand side of (82) as 1 n2 Bn X Bn X   − ∆Ti,n − Ci,n + Ci,n . {z } i=1 M∗ ,Θ̃i,n Θ̃i,n ∆Ti,nn i=1 | Di,n P Note that ki=1 Di,n is a discrete martingale, and thus to show that it vanishes asymptotically, it is sufficient to show n Bn X  2  P ETi−1,n Di,n → 0. (83) i=1 Regular conditional distribution and (21) implies (83). Similarly, regular conditional distribuP n P tion together with (20) enables us to deduce B i=1 Ci,n → 0. Thus, we proved (82). We show (17) and (18) The proof is very similar to the previous part, using respectively Condition (E5) and Condition (E6) instead of Condition (E2). 7.6 7.6.1 Proof of Theorem 3 (QMLE) Algorithm to construct the unbiased estimator b (BC) We describe here the algorithm to obtain Θ . Note that the bias-correction is only required n when α = 1/2. 44 1. We compute the local QMLEs. 2. From Theorem 6 (p. 241) in Xiu (2010), we compute the corresponding W1 and W2 . 3. We change some entries of the matrices to ensure unbiased estimates when using formula (21) and (22) in the aforementioned theorem. 4. We compute the unbiased local QMLE using the formula (21) and (22) with the corrected matrices. b (BC) 5. The bias-corrected LPE Θ is taken as the mean of local bias-corrected estimates. n 7.6.2 Proof We want to show that the conditions of Theorem 2 are satisfied. We start with the case α > 12 . The key result is Theorem 6 in Xiu (2010, p. 241). We show first Condition (E3’). We can see easily from the key result that if we choose Vθt∗ = 6σt2 , then (32) is satisfied. We can verify the second equation (33) straightforwardly. As for the bias condition (34), we can see that as the noise shrinks faster than the order of the returns to 0, then the bias tends to the sum of the diagonal elements of W1 defined in (23) in Xiu (2010, p. 241) minus unity. This equals 0 and thus (34) is satisfied. Condition (E5’) is satisfied with M∗n = (0, 0) as the behavior of the QMLE doesn’t depend much on the initial condition. Finally, Condition (E6’) can be shown combining the aforementioned theorem with the rationale in Appendix II Section 7.3. We consider now the case α = 1/2, i.e. when both the noise variance and the returns are b (BC) of the same rate. In that case, we need to use the bias-corrected estimator Θ so that we n can verify the conditions of Thorem 2. The key result here is Proposition 1 (p. 369) along with its proof (p. 391-393) in Aït-Sahalia et al. (2005). Condition (E3) is then satisfied using similar techniques of proof along with the fact that we have reduced the bias of the estimator. Condition (E5’) with M∗n = (0, 0) and Condition (E6’) are satisfied using similar techniques to the proof in the case α > 1/2 along with the fact that the estimator bias is reduced. 7.7 Proof of Theorem 4 (powers of volatility) We aim to show that we can verify the conditions of Theorem 2. The idea is to use a Taylor expansion as in the delta method. Then conditions will be satisfied following the proof of Theorem 3. More specifically, for Condition (E3) the Taylor expansion can be expressed as 1,M∗n ,σ 2 g θ̂hn ,n (Ri,n h1/2 n 1,M∗n ,σ 2 = h1/2 θ̂hn ,n (Ri,n n h ,M∗n ,σ 2 ; · · · ; Ri,nn h ,M∗n ,σ 2 ; · · · ; Ri,nn 45   ; R∗n ) − g(σ 2 )  ; R∗n ) − σ 2 g 0 (σ 2 ) + op (hn n−1 ). For Condition (E5’) we have similarly and uniformly 1,M0i,n ,σ 2 g(θ̂hn ,n (Ri,n 1,M0i,n ,σ 2 = (θ̂hn ,n (Ri,n hn ,M0i,n ,σ 2 , · · · , Ri,n hn ,M0i,n ,σ 2 , · · · , Ri,n 1,M∗n ,σ 2 ; R0i,n )) − g(θ̂hn ,n (Ri,n 1,M∗n ,σ 2 ; R0i,n ) − θ̂hn ,n (Ri,n h ,M∗n ,σ 2 , · · · , Ri,nn h ,M∗n ,σ 2 , · · · , Ri,nn ; R∗n )) ; R∗n ))g 0 (σ 2 ) −1/2 +op (h1/2 ). n n Finally, for Condition (E6’) we also have uniformly 1,M0i,n ,σt2 g θ̂hn ,n (Ri,n 1,M0i,n ,σt2 = θ̂hn ,n (Ri,n hn ,M0i,n ,σt2 , · · · , Ri,n hn ,M0i,n ,σt2 , · · · , Ri,n   1,M0 ,σ 2 hn ,M0 ,σ 2 ; R0i,n ) − g θ̂hn ,n (Ri,n i,n 0 , · · · , Ri,n i,n 0 ; R0i,n ) 1,M0i,n ,σ02 ; R0i,n ) − θ̂hn ,n (Ri,n hn ,M0i,n ,σ02 , · · · , Ri,n  ; R0i,n ) g 0 (σ02 ) −1/2 +op (h1/2 ). n n 7.8 Proof of Theorem 5 (E-(LPE of QMLE)) The strategy of the proof consists in showing that the estimation error in ν doesn’t affect asymptotically the behavior of the QMLE so that we can apply directly Theorem 3. To do that, the key results will be Theorem 3(i) (p. 37) in Li et al. (2016) and Theorem 6 (p. 241) in Xiu (2010). bτ = Zτ −g(Ii,n , ν̂) and we define the n-dimensional vector Ybn = (X bτ − We recall that X 1,n i,n i,n b0 , · · · , X bT − X bτ X ). We also define Y = ((X − X ) + (˜  −  ˜ ), · · · , (X − X n τ 0 1,n 0,n τ T n−1,n 1,n n−1,n ) + (˜ n,n − ˜n−1,n )) and δn = (g(I1,n , ν̂) − g(I0,n , ν̂)) − (g(I1,n , ν) − g(I0,n , ν)), · · · , g(In,n , ν̂) − g(In−1,n , ν̂)) − (g(In,n , ν) − g(In−1,n , ν))). It is clear that Ybn = Yn + δn . (84) b n is the LPE of QMLE on Ybn and we define Θ e n as the LPE of QMLE Finally, we recall that Θ on Yn . Consider the case α < 1/2 (the case α = 1/2 is done following the same line of reasoning). The goal is to show that stably in distribution Z T    1 2 1/2 b −1 (85) n Θn − Θ → 6T σs4 ds N (0, 1). 0 We decompose the left hand-side term in (85) as b n − Θ) = n1/2 (Θ e n − Θ) + n1/2 (Θ bn − Θ e n) . n1/2 (Θ | {z } | {z } An On the account of Theorem 3, we have that An → (6T −1 P show that Bn → 0, then this implies (85). 46 Bn RT 0 1 σs4 ds) 2 N (0, 1). Thus, if we can P We show now that Bn → 0. We define Mn the set of real n × n matrices. In view of Theorem 6 (p. 241) in Xiu (2010), there exists a function M : K → Mn × Mn θ 7→ (M (1) (θ), M (2) (θ)) b i,n = Ybn0 M (Θ b i,n )Ybn and Θ̃i,n = Yn0 M (Θ̃i,n )Yn , where we define for any θ ∈ K and such that Θ any n dimensional vector Y : Y 0 M (θ)Y = (Y 0 M (1) (θ)Y, Y 0 M (2) (θ)Y ). We have that b n − Θ̃i,n ), Bn = n1/2 (Θ b i,n )Ybn − Yn0 M (Θ̃i,n )Yn ), = n1/2 (Ybn0 M (Θ b i,n )(Yn + δn ) − Y 0 M (Θ̃i,n )Yn ), = n1/2 ((Yn + δn )0 M (Θ n 1/2 0 0 b b i,n )Yn + Yn0 M (Θ b i,n )δn + δn0 M (Θ b i,n )δn )), = n (Yn (M (Θi,n ) − M (Θ̃i,n ))Yn + (δn M (Θ b i,n ) − M (Θ̃i,n ))Yn + op (1), = n1/2 Yn0 (M (Θ = op (1). where we used (84) in the third equality, Assumption A along with Theorem 3(i) in Li et al. (2016) in the fifth equality, and Theorem 6 in Xiu (2010) along with Assumption A, Theorem 3(i) in Li et al. (2016) in the sixth equality. 7.9 Proof of Theorem 6 (powers of volatility) The proof follows the proof of Theorem 5 along with the proof of Theorem 4. 7.10 Proof of Theorem 7 (Time-varying friction parameter model with uncertainty zones) We choose M∗n = (1, 1, 1). Condition (E1) follows exactly from Corollary 4.4 (p. 14) in Robert t 0 and Rosenbaum (2012). We show now Condition (E2). We can decompose ∆TM,θ − ∆TN,θ i,n i,n into  N,θt  0 0 0 ∆TM,θ − ∆TN,θ + ∆TN,θ . (86) i,n i,n i,n − ∆Ti,n (1)  We deal with the first term in (86). We can see that under the parametric model Li,n , sign(Ri,n ) follows a discrete Markov chain on the space {1, · · · , m} × {−1, 1}. Following the same line 47 of reasoning as in the proof of Lemma 14 in Potiron and Mykland (2017), we can easily show that   1 0 0 E ∆TM,θ − ∆TN,θ sup = op (hn n−1− l ), i,n i,n (θt ,V,M,N)∈EM,n ×M2m,n sup (θt ,V,M,N)∈EM,n ×M2m,n   2 0 0 Var ∆TM,θ − ∆TN,θ = op (hn n−1− l ). i,n i,n We turn now to the second term in (86). Using the same idea as in the proof of Lemma 11 in Potiron and Mykland (2017), we deduce  1 N,θt  0 E ∆TN,θ = op (hn n−1− l ), i,n − ∆Ti,n (87)   2 t 0 = op (hn n−1− l ). Var ∆TM,θ − ∆TN,θ i,n i,n (88) sup (θt ,V,N)∈EM,n ×Mm,n sup (θt ,V,N)∈EM,n ×Mm,n We turn now to Condition (E3). Let M > 0. Because in the model with uncertainty M∗ ,θ (1)  zones ∆Ti,nn can be written as a sum of Markov chain Li,n , sign(Ri,n ) with a finite fourth moment, we have uniformly in θ ∈ KM and in i = 1, · · · , Bn that  10  1,M∗ ,θ h ,M∗ ,θ M∗ ,θ  Var hnl θ̂hn ,n (Ri,n n ; · · · ; Ri,nn n ; R∗n ) − θ ∆Ti,nn    10 1,M∗ ,θ h ,M∗ ,θ M∗ ,θ  = Var hnl θ̂hn ,n (Ri,n n ; · · · ; Ri,nn n ; R∗n ) − θ E ∆Ti,nn )2 + op (h2n n−2 ),  10   1,M∗ ,θ h ,M∗ ,θ M∗ ,θ  M∗ ,θ = Var hnl θ̂hn ,n (Ri,n n ; · · · ; Ri,nn n ; R∗n ) − θ E ∆Ti,nn ∆Ti,nn + op (h2n n−2 ), (1) (2) M∗ ,θ = Sθ,n Sθ,n ∆Ti,nn T hn n−1 + op (h2n n−2 ), with   10 1,M∗ ,θ h ,M∗ ,θ (1) Sθ,n := Var hnl θ̂hn ,n (Ri,n n ; · · · ; Ri,nn n ; R∗n ) − θ  (2) M∗ ,θ  and Sθ,n := E ∆Ti,nn T −1 h−1 n n. By Lemma 4.19 in p. 26 of Robert and Rosenbaum (2012) in the special case where the volatility is constant, we obtain the existence and the value of (1) (1) (1) Sθ such that Sθ,n → Sθ . Also, by Corollary 4.4 in p. 14 of Robert and Rosenbaum (2012), (2) (2) (2) (1) (2) there exists Sθ such that Sθ,n → Sθ . If we define Vθ = Sθ Sθ , (22) is satisfied. Moreover, (23) and (24) can be verified easily. Condition (E5) can be proven the same way as the first term in (86). Finally, we can prove Condition (E6) using (87), (88) and similar arguments to the proof of Lemma 13 in Potiron and Mykland (2017). 7.11 Proof of Theorem 8 (Time-varying MA(1)) The key result for this proof is the connection between the MA(1) process and the observations in the model described in Section 4.1 in the case α = 1/2. Such connection can be seen in 48 view of the proof of Proposition I (pp. 391-393) in Aït-Sahalia et al. (2005). More specifically, we can use Taylor expansions to re-express this estimator as the estimator in Section 4.1 and then use Theorem 3 (ii) to conclude. Similar Taylor expansions were already obtained in the proof of Theorem 4, and we will not further explain the details in this specific case. We verify now the conditions hold. Condition (E3) is obtained via Taylor expansions and Theorem 3 (ii), along with the fact that n1/4 = o(hn ) for the bias condition. We use the same results to prove Condition (E5’)-(E6’). 7.12 Estimation of the friction parameter bias and standard deviation in the model with uncertainty zones In this section, we provide the friction parameter bias and standard deviation used in our empirical illustration. The notation of Section 4.5 and Section 6 are in force. We define ŝn (η) := V̂ , where an expression of V̂ will be provided at the end of this section. First, we assume that the absolute jump size is constant equal to the tick size, i.e. Li,n := 1. In view of (51), we have  N (c)  α,t,1 (m) η̂α,t := min 1, . (a) 2Nα,t,1 (a) (c) We also have by definition that the number of alternations is Nα,t,1 = Nn − Nα,t,1 . If we assume that Nn is non-random, then 2η (c) ), (89) Nα,t,1 ∼ Bin(Nn , 2η + 1 where Bin(n, p) is a binomial distribution with n observations and probability p. Let B ∼ 2η ). We can define the bias Bin(Nn , 2η+1    B B := E min 1, −η 2(Nn − B) and the variance as   V := Var min 1,  B . 2(Nn − B) B and V can be computed easily numerically. If Nn is random, we can work conditional on Nn . As the sampling times are endogenous, (c) 2η (89) is not true in that case. Nonetheless, we can still approximate Nα,t,1 by Bin(Nn , 2η+1 ) if the number of observations is large enough. We now turn out to the general case, where Li,n can be different from 1. For k = 1, · · · , m (c) we define pk := 2η+k−1 2η+k and we let Bk be an independent sequence of distribution Bin(Nα,t,k + (a) Nα,t,k , pk ), and   1   Bk Ck := max 0, min 1, k − 1 + 1 . (a) (c) 2 Nα,t,k + Nα,t,k − Bk 49 P (m) (m) The distribution of η̂α,t can be approximated by the distribution of m i=1 λα,t,k Ck , and we can    Pm P (m) 2 (m)  estimate the bias as B̂ := m i=1 (λα,t,i ) Var Ck . i=1 λα,t,i E Ck and the variance as V̂ := 50 Figure 1: AVAR of ERVext and E-(LPE of QMLE) as a function of the noise variance, i.e. the variance of ˜i,n . The horizon time is set to T = 1 (which corresponds to 6.5 hours of intraday trading). On the left hand-side, we follow exactly the setting of the numerical study in Li et al. (2016), where σt2 = 0.000125 if 0.05 ≤ t < 0.95 and σt2 = 15 ∗ 0.000125 otherwise. There is on average one observation a second, which corresponds to n = 23, 400. On the right hand-side, the setting is the same except that we remove the jumps in volatility and consider σt2 = 0.000125 for 0 ≤ t ≤ 1. 51 Figure 2: Evolution of η̂i,n for different values of hn . The red line corresponds to the global (mod) (mod) estimate η̂αn ,T . The blue lines are one (local) standard deviation ŝhn away from η̂αn ,T . The (mod) purple lines are two (local) standard deviations away from η̂αn ,T . 52 Figure 3: Estimated volatility with the LPE for different values of hn . The red line corresponds to the RV estimator. The blue line stands for the model with uncertainty zones volatility d αn ,T . estimator RV estimator MLE FRO LPE BC LPE LPE BC LPE LPE BC LPE LPE BC LPE LPE BC LPE LPE BC LPE block size 25 25 100 100 500 500 1000 1000 2000 2000 5000 5000 β sample bias -0.0052 0.2913 -0.0168 -0.0172 0.0035 -0.0010 -0.0021 -0.0049 -0.0030 -0.0056 -0.0032 -0.0055 -0.0052 -0.0060 β s.d. 0.0085 0.0355 0.0131 0.0132 0.0083 0.0082 0.0094 0.0095 0.0099 0.0100 0.0102 0.0101 0.0087 0.0087 κ sample bias 0.1041 0.1666 -0.0985 -0.0062 -0.0256 -0.0065 0.0073 0.0098 0.0425 0.0438 0.1029 0.1035 0.1037 0.1044 κ s.d. 0.0148 0.0355 0.0096 0.0097 0.0096 0.0096 0.0101 0.0104 0.0125 0.0126 0.0143 0.0143 0.0148 0.0147 Table 1: In this table, we report the sample bias and the standard deviation for the different estimators in the case of a small number of oscillations δ = (4, 4). The parameter (β, κ) = (0.5, 1). The number of Monte-carlo simulations is 1,000. 53 estimator MLE FRO LPE LPE BC LPE LPE BC LPE LPE BC LPE LPE BC LPE LPE BC LPE LPE BC block size 25 25 100 100 500 500 1000 1000 2000 2000 5000 5000 β sample bias -0.0069 0.0065 -0.0148 -0.0155 0.0017 0.0012 -0.0053 -0.0085 -0.0071 -0.0115 -0.0071 -0.0108 -0.0071 -0.0093 β s.d. 0.0105 0.0391 0.0183 0.0184 0.0092 0.0092 0.0094 0.0094 0.0102 0.0102 0.0106 0.0106 0.0106 0.0106 κ sample bias 0.1094 0.0882 -0.0876 0.0046 -0.0164 0.0039 0.1046 0.1086 0.1078 0.1098 0.1087 0.1096 0.1087 0.1090 κ s.d. 0.0222 0.0678 0.0144 0.0143 0.0183 0.0183 0.0219 0.0221 0.0216 0.0217 0.0220 0.0221 0.0220 0.0219 Table 2: In this table, we report the sample bias and the standard deviation for the different estimators in the case of a bigger number of oscillations δ = (10, 10). The parameter (β, κ) = (0.5, 1). The number of Monte-carlo simulations is 1,000. hn 50 100 150 200 250 300 350 400 450 Bn 67 34 23 17 14 12 10 9 8 Chi Sq. Stat 719 268 155 116 109 68.5 90.6 91.5 42.6 Dg. Fr. 66 33 22 16 13 11 9 8 7 p-value 0 0 0 0 0 0 0 0 6e−7 Table 3: Summary chi-square statistics χ2n based on the block size hn . Note that since the number of observations of the last block is arbitrary, the last block estimate η̂Bn ,n is not used to compute the chi-square statistic. 54
10math.ST
Improving Max-Sum through Decimation to Solve Loopy Distributed Constraint Optimization Problems J. Cerquides1 , R. Emonet2 , G. Picard3 , and J.A. Rodriquez-Aguilar1 arXiv:1706.02209v1 [cs.MA] 7 Jun 2017 1 IIIA-CSIC, Campus UAB, 08193 Cerdanyola, Catalonia, Spain {cerquide,jar}@iiia.csic.es 2 Université de Lyon, Laboratoire Hubert Curien UMR CNRS 5516, France [email protected] 3 MINES Saint-Etienne, Laboratoire Hubert Curien UMR CNRS 5516, France [email protected] Abstract. In the context of solving large distributed constraint optimization problems (DCOP), belief-propagation and approximate inference algorithms are candidates of choice. However, in general, when the factor graph is very loopy (i.e. cyclic), these solution methods suffer from bad performance, due to non-convergence and many exchanged messages. As to improve performances of the Max-Sum inference algorithm when solving loopy constraint optimization problems, we propose here to take inspiration from the belief-propagation-guided decimation used to solve sparse random graphs (k-satisfiability). We propose the novel DeciMaxSum method, which is parameterized in terms of policies to decide when to trigger decimation, which variables to decimate, and which values to assign to decimated variables. Based on an empirical evaluation on a classical BP benchmark (the Ising model), some of these combinations of policies exhibit better performance than state-of-the-art competitors. 1 Introduction In the context of multi-agent systems, distributed constraint optimization problems (DCOP) are a convenient to model coordination issues agents have to face, like resource allocation, distributed planning or distributed configuration. In a DCOP, agents manage one or more variables they have to assign a value (e.g. a goal, a decision), while taking into account constraints with other agents. Solving a DCOP consists in making agents communicate as to minimize the violation of these constraints. Several solution methods exist to solve such problems, from complete and optimal solutions, to approximate ones. When dealing with larger scales (thousands of variables), approximate methods are solutions of choice. Indeed, complete methods, like ADOPT or DPOP, suffer exponential computation and/or communication cost in general settings [10,15]. As a consequence, in some large settings, approximate methods are better candidates, as evidenced by the extensive literature on the subject (see [1] for a complete review). One major difficulty for approximate method to solve DCOP is the presence of cycles in the constraint graph (or factor graph). Among the aforementioned methods, inference-based ones, like Max-Sum [3] and its extensions like [16], have demonstrated good performance even on loopy settings. However, there exists some cases, with numerous loops or large induced width of the constraint graph, where they perform badly, which translates into a larger number of messages, a longer time to convergence and a final solution with bad quality. One original approach to cope with loopy graphs is to break loops by decimating variables during the solving process. Decimation is a method inspired by statistical physics, and applied in belief-propagation, which consists in fixing the value of a variable, using the marginal values as the decision criteria to select the variable to decimate [13]. The decimation is processed regularly after the convergence of a classical belief-propagation procedure. In [11], decimation has been used in the constraint satisfaction framework, for solving centralized k-satisfiability problems [11]. Inspired by this concept, we propose a general framework for applying decimation in the DCOP setting. Other works proposed Max-Sum_AD_VP as to improve Max-Sum performance on loopy graphs [20]. The idea is to perform the inference mechanism through an overlay directed 2 Cerquides et al. acyclic graph, to remove loops, and to alternating the direction of edges at a fixed frequency as to improve the sub-optimal solution found with the previous direction. One mechanism within one of these extensions, namely value propagation, can be viewed as a temporary decimation. Against this background, the main goal of this paper is to propose a general framework for installing decimation in Max-Sum for solving DCOP. More precisely, we make the following contributions: 1. We propose a parametric solution method, namely DeciMaxSum, to implement decimation in Max-Sum. It takes three fundamental parameters for decimation: (i) a policy stating when to trigger decimation, (ii) a policy stating which variables to decimate, and (iii) a policy stating which value to assign to decimated variables. The flexibility of DeciMaxSum comes from the fact that any policy from (1i) can be combined with any policy from (1ii) and (1iii). 2. We propose a library of decimation policies; some inspired by the state-of-the-art and some original ones. Many combinations of policies are possible, depending on the problem to solve. 3. We implement and evaluate some of these combinations of decimation policies on classical DCOP benchmarks (meeting scheduling and Ising models), against state-of-the-art methods like standard Max-Sum and Max-Sum_AD_VP. The rest of the paper is organized as follows. Section 2 expounds some background on DCOP and expounds the decimation algorithm from which our algorithm DeciMaxSum is inspired. Section 3 defines the general framework of DeciMaxSum, and several examples of decimation policies. Section 4 presents results and analyses of experimenting DeciMaxSum, with different combinations of decimation policies, against Max-Sum and Max-Sum_AD_VP. Finally, Section 5 concludes this paper with some perspectives. 2 Background This section expounds the DCOP framework and some related belief-propagation algorithms from the literature are discussed concerning the mechanisms to handle cycles in constraint graphs. 2.1 Disributed Constraint Optimization Problems One way to model the coordination problem between smart objects is to formalize the problem as a distributed constraint optimization problem. Definition 1 (DCOP). A discrete Distributed Constraint Optimization Problem (or DCOP) is a tuple hA, X , D, C, µi, where: A = {a1 , . . . , a|A| } is a set of agents; X = {x1 , . . . , xN } are variables owned by the agents; D = {Dx1 , . . . , DxN } is a set of finite domains, such that variable xi takes values in Dxi = {v1 , . . . , vk }; C = {u1 , . . . , uM } is a set of soft constraints, where each ui defines a utility ∈ R ∪ {−∞} for each combination of assignments to a subset of variables Xi ⊆ X (a constraint is initially known only to the agents involved); µ : X → A is a function mapping variables to their associated agent. A solution to the DCOP is an assignment X ∗ = {x∗1 , . . . , x∗N } to all variables that maximizes the overall sum of costs4: M X um (Xm ) (1) m=1 4 Note that the notion of cost can be replaced by the notion of cost ∈ R∪{+∞}. In this case, solving a DCOP is a minimization problem of the overall sum of costs. DeciMaxSum: Decimation in MaxSum 3 As highlighted in [1], DCOPs have been widely studied and applied in many reference domains, and have many interesting features: (i) strong focus on decentralized approaches where agents negotiate a joint solution through local message exchange; (ii) solution techniques exploit the structure of the domain (by encoding this into constraints) to tackle hard computational problems; (iii) there is a wide choice of solutions for DCOPs ranging from complete algorithms to suboptimal algorithms. A binary DCOP can be represented as a constraint graph, where vertices represent variables, and edge represent binary constraints. In the case of n-ary constraints, a DCOP can be represented as a factor graph: an undirected bipartite graph in which vertices represent variables and constraints (called factors), and an edge exists between a variable and a constraint if the variable is in the scope of the constraint. Definition 2 (Factor Graph). A factor graph of a DCOP as in Def. 1, is a bipartite graph F G = hX , C, Ei, where the set of variable vertices corresponds to the set of variables X , the set of factor vertices corresponds to the set constraints C, and the set of edges is E = {eij | xi ∈ Xj }. When the graph representing the DCOP contains at least a cycle, we call it a cyclic DCOP; otherwise, it is acyclic. A large literature exists on algorithms for solving DCOPs which fall into two categories. On the one hand, complete algorithms like ADOPT and its extensions [9], or inference algorithms like DPOP [15] or ActionGDL [19], are optimal, but mainly suffer from expensive memory (e.g. exponential for DPOP) or communication (e.g. exponential for ADOPT) load –which we may not be able to afford in a constrained infrastructure, like in sensor networks. On the other hand, approximate algorithms like Max-Sum [3] or MGM [8] have the great advantage of being fast with a limited memory print and communication load, but losing optimality in some settings –e.g. Max-Sum is optimal on acyclic DCOPs, and may achieve good quality guarantee on some settings. The aforementioned algorithms mainly exploit the fact that an agent’s utility (or constraint’s cost) depends only on a subset of other agents’ decision variables, and that the global utility function (or cost function) is a sum of each agent’s utility (constraint’s cost). In this paper, we are especially interested in belief-propagation-based algorithms, like Max-Sum, where the notion of marginal values describes the dependency of the global utility function on variables. 2.2 From Belief-Propagation to Max-Sum Belief propagation (BP), i.e. sum-product message passing method, is a potentially distributed algorithm for performing inference on graphicalQmodels, and can operate on factor graphs representing a product of M factors [7]: F (x) = M m=1 fm (Xm ) . The sum-product algorithm provides an efficient local message passing procedure to compute the marginal functions of all variables simultaneously. The marginal function, Pzn (xn ) describes the total dependency of the global function F (x) on variable xn : zn (xn ) = {x0 },n0 6=n F (Xn0 ). BP operates iteratively propagating messages mi→j (tables associating marginals to each value of variables) along the edges of the factor graph.When the factor graph is a tree, BP algorithm computes the exact marginals and converge in a finite number a steps depending on the diameter of the graph [7]. Max-product is an alternative version of sum-product which computes the maximum value instead of the sum. Built as a derivative of max-product, Max-Sum is an approximate algorithm to solve DCOP [3]. The main evolution is the way messages are assessed, to pass from product to sum operator through logarithmic translation. And as a consequence, Max-Sum computes an assignment X ∗ 4 Cerquides et al. Algorithm 1: The BP-guided decimation algorithm from [11] Data: A factor graph representing a k-satisfiability problem Result: A feasible assignment X ∗ or FAIL 11 initialize BP messages U ←∅ for t = 1, . . . , n do run BP until the stopping criterion is met choose xi ∈ X \ U uniformly at random compute the BP marginal zi (xi ) choose x∗i distributed according to zi fix xi = x∗i U ← U ∪ {xi } simplify the factor graph if a contradiction is found, return FAIL 12 return X ∗ 1 2 3 4 5 6 7 8 9 10 that maximizes the DCOP objective in Equation 1. Depending on the DCOP to solve, MaxSum may be used with two different termination rules: (i) continue until convergence (no more exchanged messages, because when a variables or a factor receives twice the same message from the same emitter it does not propagates); (ii) propagate message for a fixed number of iterations per agent. Max-Sum is optimal on tree-shaped factor graphs, and still perform well on cyclic settings. But there exist problems for which Max-Sum does not converge or converge to a sub-optimal state. In fact, on cyclic settings [3] identify the following behaviors: (i) agents converge to fixed states that represent either the optimal solution, or a solution close to the optimal, and the propagation of messages ceases; (ii) agents converge as above, but the messages continue to change slightly at each update, and thus continue to be propagated around the network; (iii) neither the agents’ preferred states, nor the messages converge and both display cyclic behavior. As to improve Max-Sum performance on cyclic graphs, [20] proposed two extensions to Max-Sum: (i) Max-Sum_AD which operates Max-Sum on a directed acyclic graph built from the factor graph, and alternates direction at a fixed rate (a parameter of the algorithm); (ii) MaxSum_AD_VP which operates Max-Sum_AD and propagates current values of variables when sending Max-Sum messages so that factors receiving the value only consider this value instead of the whole domain of the variable. These two extensions, especially the second one, greatly improves the quality of the solution: Max-Sum_AD_VP found solutions that approximate the optimal solution by a factor of roughly 1.1 on average. However, the study does not consider the number of exchanged messages, or the time required to converge and terminate MaxSum_AD_VP. 2.3 BP-guided Decimation In this paper, we propose to take inspiration from work done in computational physics [13], as to cope with cyclicity in DCOP. Notably, [5] introduced the notion of decimation in constraint satisfaction, especially k-satisfiability, where variables are binary, xi ∈ {0, 1}, and each constraint requires k of the variables to be different from a specific k-uple. Authors proposed a class of algorithms, namely message passing-guided decimation procedure, which consists in iterating the following steps: (1) run a message passing algorithm, like BP ; (2) use the result to choose a variable index i , and a value x∗i for the corresponding variable; (3) replace the constraint satisfaction problem with the one obtained by fixing xi to x∗i . The BP-guided decimation procedure is shown in Algorithm 1, whose performances are analysed in [11,13]. DeciMaxSum: Decimation in MaxSum 5 BP-guided decimation operates on the factor graph representing the k-satisfiability problem to solve. At each step, the variable to decimate is randomly chosen among the remaining variables. The chosen variable xi is assigned a value determined by random sampling according to its marginal zi . After decimation, the factor graph is simplified: some edges are no more relevant, and factors can be sliced (columns corresponding to removed variables are deleted). In some settings, BP-guided decimation may fail, if random choices assign a value to a variable which is not consistent with other decimated variables. Some comments can be made on this approach. First, relying on marginal values is a key feature, and is the core of the “BP-guided” nature of this method. Marginal values are exploited to prune the factor graph. Second, while in the seminal work of [11], this procedure is used to solve satisfiability problems, the approach can easily be implemented to cope with optimization problems. For instance, the inference library libDAI proposes an implementation of decimation for discrete approximate inference in graphical models [12], which was amongst the three winners of the UAI 2010 Approximate Inference Challenge5. 2.4 State of a Factor Graph Representation The previous BP-based algorithm operates on factor graph representing the problem. “Operates” means that the algorithms create a data structure representing the factor graph which evolves with time : marginal values change, variables disappear, messages are sent/received, etc. Commonly, the logical representation of a factor graph is a set of nodes connected depending on the connectivity of the graph. Each such node has a state which stores some useful values. Definition 3. The current state F Gt at time t of a factor graph F G = hX , C, Ei is the composition of all the current states of the data structures used by the BP-based algorithm to operate on the related factor graph, including the marginal values zi , the messages mi→j , the set of decimated variables U, and other algorithm-specific data. We can consider that for a given problem, many factor graph states may exist. We denote S the set of possible factor graph states, and S(F G) ⊂ S the set of possible states for the factor graph F G. 3 DeciMaxSum: Extending Max-Sum with Decimation While mainly designed as a centralized algorithm and studied on k-SAT problems, BP-guided decimation could be utilized for solving DCOP with a few modifications. To the best of our knowledge, this approach has never been proposed for improving Max-Sum algorithm. Here we expound the core contribution of this paper, namely the DeciMaxSum framework and its components. 3.1 Principles The main idea is to extend the BP-guided decimation algorithm from [11] in order to define a more general framework, in which other BP-based existing algorithms could fit. First, the main focus is decimation, which means assigning a value to a variable as to remove it from the problem. As the name suggests, there is no way back when a variable has been decimated –unlike search algorithms, where variable assignments can be revised following a backtrack, for instance. Therefore, triggering decimation is an impacting decision. This is why our framework 5 http://www.cs.huji.ac.il/project/UAI10/ 6 Cerquides et al. is mainly based on answering three questions: (i) when is decimation triggered, (ii) which variable(s) to decimate, (iii) which value to assign to the decimated variable(s)? Several criteria can be defined for answering each question, and the DeciMaxSum specifies such criteria as decimation policies, that are fundamental parameters of the decimation procedure. Definition 4 (Decimation Policy). A decimation policy is a tuple π = hΘ, Φ, Υ, Λi where: – Θ : S → {0, 1} is the condition to trigger the decimation process, namely the trigger policy, – Φ : S → 2X is a filter policy which selects some candidate variables to decimate, – Υ : X × S → {0, 1} is the condition to perform decimation on a variable, namely perform policy, – Λ : X × S → DX is the assignment policy, which assigns a value to a given variable. A rich population of decimation-based algorithm can be modeled through this framework by combining decimation policies. For instance, one can consider a DeciMaxSum instance, which (i) triggers decimation once BP has converged, (ii) chooses randomly a variable to decimated within the whole set of non-decimated variables, and (iii) samples the value of the decimated variable depending on its marginal values (used as probability distribution). By doing so, we result in the classical BP-guided decimation algorithm from [11] . However, as many more decimation policies can be defined and combined, we fall into a more general framework generating a whole family of algorithms. 3.2 DeciMaxSum as an Algorithm We can summarize the DeciMaxSum framework using Algorithm 2. It is a reformulation of BPguided decimation, parameterized with a decimation policy. Here decimation is not necessarily triggered at the convergence (or time limit) of BP. Criterion Θ may relies on other components of the state of the factor graph. Contrary to classical BP-guided decimation, there may be several variables to decimate at the same time (like in some variants of DSA or MGM) and that variables can be chosen in an informed manner (and not randomly), using criterion Υ . Values assigned to decimated variables, are not necessarily chosen stochastically, but are assigned using the function Λ that can be deterministic (still depending on the current state of the FG). Since, here we’re not in the k-satisfiability case, but in an optimization case, there is no failure (only suboptimality), contrary to Algorithm 1. Finally, once all variables have been decimated, the output consists in decoding the state F Gt , i.e. getting the values assigned to decimated variables. This means that finally DeciMaxSum is performing decoding while solving the problem, which is not a common feature in other DCOP algorithms, like classical Max-Sum or DSA. Indeed, once these algorithms halt, a decoding phase must be performed to extract the solution from the variables’ states. While presented as a classical algorithm, let us note that decimation is meant to be implemented in a distributed and concurrent manner, depending on the decimation policy components. The rest of the section details and illustrates each of these decimation policies component with some examples. 3.3 Triggering Decimation (Θ criterion) In the original approach proposed by [11], decimation is triggered once BP has converged. In a distributed settings and diffusing algorithms like BP, this can be implemented using termination detection techniques.  1, if s is quiescent def (2) Θconverge (s) = 0, otherwise DeciMaxSum: Decimation in MaxSum 7 Algorithm 2: The DeciMaxSum framework as an algorithm Data: A factor graph F G = hX , C, Ei, a decimation policy π = hΘ, Φ, Υ, Λi Result: A feasible assignment X ∗ 1 2 3 4 5 6 7 8 9 10 initialize BP messages U ←∅ while U 6= X do run BP until decimation triggers, i.e. Θ(F Gt ) = 1 choose variables to decimate, X 0 = {xi ∈ Φ(F Gt ) | Υ (xi , F Gt )} for xi ∈ X 0 do xi ← Λ(xi , F Gt ) U ← U ∪ {xi } simplify F Gt // Sect. 3.3 // Sect. 3.4 // Sect. 3.5 // remove variables, slice factors return X ∗ by decoding U This trigger consists in detecting the quiescence of the current state of the factor graph. This means no process is enabled to perform any locally controlled action and there are no messages in the channels [6]. Algorithms like DijkstraScholten can detects such global state by implementing a send/receive network algorithm, based on the same graph than F G [6]. Note that such techniques generates extra communication load for termination detection-dedicated messages. Due to the Max-Sum behavior on loopy factor graphs, convergence may not be reached [20]. The common workaround is to run BP for a fixed number of iterations in case there is no convergence. Setting this time limit (namely LIMIT) might be really problem-dependent.  1, if time(s) = LIMIT def (3) Θtime (s) = 0, otherwise In synchronous settings (all variables and factors are executed synchronously, step by step), getting the iteration number of the current state of the FG, time(s), can done in a distributed manner, as usually done in Max-Sum. In the asynchronous case, one can either (i) use a shared clock, or (ii) count locally outcoming messages within each variables, and once a variable has sent a limit number of messages, decimation is triggered. In some settings with strong time or computation constraints (e.g. sensor networks [3], internet-of-things [17]), waiting convergence is not affordable. Indeed, BP may generate a lot of messages. Therefore, we may consider decimating before convergence at a fixed rate (e.g. each 10 iterations), or by sharing a fixed iteration budget amongst the variables (e.g. each 1000 iterations divided by the number of variables). We can even consider a varying decimation speed (e.g. faster at the beginning, and lower at the end, as observed in neural circuits in the brain [14]).  1, if time(s) mod f (s) = 0 def (4) Θfrequency (s) = 0, otherwise where f is a function of the current state of the system, for instance : – f (s) = RATE, with a predefined decimation frequency, – f (s) = BUDGET/|X |, with a predefined computation budget, – f (s) = 2 × time(s), for an decreasing decimation frequency. Finally, another approach could be to trigger decimation once a loop in the FG is detected. Indeed, decimation is used here to cope with loops, so decimating variables, which could potentially break loops, seems a good approach. 8 Cerquides et al. def Θloop (s) =  1, if ∃xi ∈ X , |loop(xi )| > 1 0, otherwise (5) where loop(xi ) is the set of agents in the same first loop that xi just discovered. Detecting loops in the FG can be implemented during BP, by adding some metadata on the BP messages, like done in the DFS-tree construction phase of algorithms like DPOP or ADOPT. 3.4 Deciding the Subset of Variables to Decimate (Φ and Υ criteria) Now our system has detected decimation should be triggered, the following question is “which variables to decimate?” In [11], the variable is chosen randomly in a uniform manner, while in [12], the variable with a the maximum entropy over its marginal values (the most determined variable) is selected. Obviously, exploiting the marginal values, build throughout propagation is a good idea. From which subset choosing the candidate variables to decimate? Both [11] and [12] select the only variable to decimate amongst the whole set of non-decimated variables (cf. line 5 in Algorithm 1). Here, Φ criterion is specified as follows: def Φall (s) = X \ U (6) However, this selection on the whole set of variables can be discussed when using local decimation triggers, like loop detection. In such case, selecting the variables to decimate within the agents in the loop, or the one which detected the loop sounds better. Another approach is to consider selecting agents depending on the past state of the system. For instance, if a variable has been decimated, good future candidates for decimation could be its direct neighbors in the FG: def (7) Φneighbors (s) ={x ∈ X \ U | neighbors(x) ∩ U 6= ∅} with neighbors(xi ) = {xj ∈ X | j 6= i, ∃eik , ekj ∈ E}. Which criteria to decide whether the variable decimate? Now, we have to specify the Υ criterion used to decide which candidates decimate. In [11], it is fully random: it does not depends on the current state of the variables. It corresponds to make each variable roll a dice and choosing the greatest draw:  1, if ∀xj 6= xi ∈ X , rand(xi ) > rand(xj ) def (8) Υmax_rand (xi , s) = 0, otherwise where rand(x) stands for the output of a random number generator (namely sample) using a uniform distribution (e.g. U [0, 1]). In [12], the variable with the maximal entropy over its marginal values is selected. This means the variable for which marginal values seems to be the most informed, in the Shannon’s Information Theory sense, is chosen:  1, if ∀xj 6= xi ∈ X , H(zi (xi )) > H(zj (xj )) def Υmax_entropy (xi , s) = (9) 0, otherwise P with H(zk (xk )) = − d∈Dk zk (xk )(d) log(zk (xk )(d)). From this, other criteria can be derived. For instance, instead of using entropy, one can consider the maximal normalized marginal value: DeciMaxSum: Decimation in MaxSum def Υmax_marginal (xi , s) = ( 1, if ∀xj 6= xi ∈ X , max(zi (xi )(d)) > max(zj (xj )(d)) d∈Di 0, otherwise d∈Di 9 (10) If several variables can be decimated at the same time, one may consider selecting the set of variable having an entropy or a normalized marginal value greater than a given threshold, to only decimate variable which are “sufficiently” determined. Hence, this approach requires setting another parameter (namely THRESHOLD):  1, if H(zi (xi )) > THRESHOLD def (11) Υthreshold_entropy (xi , s) = 0, otherwise Of course, many combination of the aforementioned criteria, and other criteria could be considered in our framework. We don’t discuss here criteria like in DSA which does not rely on marginal values, but on stochastic decision. Which subset of variables the decision to decimate a variable depends on? Behind this question lies the question of coordinating the variable selection. Indeed, if computing criterion Υ does not depend on the decision of other variables, the procedure is fully distributable at low communication cost, as for policies like (11). At the contrary, if the decision requires to be aware of the state of other variables, as for policies like (8), (9) and (10), the procedure will require some system-scale coordination messages. In [11] and [12], decimation only concerns all the variables, from which only one will be chosen. This requires a global coordination, or a distributed leader election protocol which may require an underlying network (ring, spanning tree, etc.), like the one used for quiescence detection, to propagate election messages [6]. In some cases, the decimation decision might be at local scale, when variables will make their decision depending on the decision of their direct neighbors, or variables in the same loop. In this case, less coordination messages will be required. For instance, if considering decimating variables in a loop, only variables in the loop will implement a leader election protocol. All policies, from (8) to (10), could be extended in the same manner, by replacing X by loop(xi ), neighours(xi ), or any subset of X . For instance:  1, if ∀xj 6= xi ∈ loop(xi ), rand(xi ) > rand(xj ) def (12) Υmax_rand_loop (xi , s) = 0, otherwise 3.5 Deciding the Values to Assign To Decimated Variables (Λ criterion) Now variables to decimate have been selected, the question is “which values to assign?” Usually, in BP-based algorithms, the simplest way to select values for variables, after propagation, is to assign values with maximal marginal value (or utility). [12] is using such a criterion for inference: def Λmax_marginal (xi , s) = argmax zi (xi )(d) (13) d∈Di While, the policy is deterministic, in [11] the choice of the value is a random choice using the marginal values as a probability distribution: def Λsample_marginal (xi , s) = sample(zi (xi )) (14) Once again, these are only some examples of policies exploiting BP, and one can easily specify many more. 10 Cerquides et al. 4 Experiments In this section we evaluate the performance of different combinations of decimation policies in DeciMaxSum, on a classical optimization model (Ising model), against classical Max-Sum [3] and its extension Max-Sum_AD_VP [20], we have implemented in our own framework. 4.1 Ising Model Since we are interested in evaluating our algorithms in the presence of strong dependencies among the values of variables, we evaluate them on Ising model which is a widely used benchmark in statistical physics [4]. We use here the same settings than [18]. Here, constraint graphs are rectangular grids where each binary variable xi is connected to its four closer neighbors (with toroidal links which connect opposite sides of the grid), and is constrained by a unary cost ri . The weight of each binary constraint rij is determined by first sampling a value κij from a uniform distribution U [−β, β] and then assigning ( κij if xi = xj rij (xi , xj ) = −κij otherwise The β parameter controls the average strength of interactions. In our experiments we set β to 1.6. The weight for each unary constraint ri is determined by sampling κi from a uniform distribution U [−0.05, 0.05] and then assigning ri (0) = κi and ri (1) = −κi . 4.2 Results and Analysis In this section we analyse results of different DeciMaxSum combinations to solve squared-shape Ising problems with side size varying from 10 to 20 (e.g. 100 to 400 variables). We implemented the following combinations: – 11 DeciMaxSum instances with different decimation policies using the following criteria: • trigger policies (Θ criterion): ∗ Θconverge (from equation 2, noted converge), ∗ rate-based Θfrequency (from equation 4, noted 2-periodic, 3-periodic, 5-periodic, 10periodic, 20-periodic, and 100-periodic), ∗ budget-based Θfrequency (from equation 4, noted periodic), • filter policy (Φ criterion): ∗ the one that selects the whole set of variables as potential variables to decimate (i.e. Φall from equation 6), • perform policies (Υ criterion): ∗ Υmax_rand (from equation 8, noted random), ∗ Υmax_entropy (from equation 9, noted max_entropy), • assignment policies (Λ criterion): ∗ deterministic Λmax_marginal (from equation 13, noted deterministic), ∗ sampled Λsample_marginal (from equation 14, noted sampling), – MaxSum, as defined in [3], – MaxSum_AD, as defined in [20], – MaxSum_AD_VP, as defined in [20], – Montanari-Decimation, as defined in [11], – Mooij-Decimation, as defined in [12]. DeciMaxSum: Decimation in MaxSum (a) Final total cost 11 (b) Final number of messages Fig. 1: Performances of DeciMaxSum and other solution methods on Ising problems (average of 10 problems per generator’s parameter set, and average of 3 runs per problem). Figure 1 presents two performance metrics (final total cost and total number of exchanged messages). Considering optimality of the final solutions obtained by the different solution methods and DeciMaxSum instances, what appears is that very fast decimation combined with a deterministic decimation of the most determined variable (max_entropy) presents the best cost. Besides, very fast decimation also imply that few messages are exchanged compared to other solution methods, since decimation cuts message propagations. However, all the solution methods (except Montanari-Decimation and Mooij-Decimation) tend to a comparable number of exchanged messages. 5 Conclusions In this paper we have investigated how to extend Max-Sum method for solving distributed constraint optimization problems, by taking inspiration from the decimation mechanisms used to solve k-satisfiability problems by belief-propagation. We propose a parametric method, namely DeciMaxSum, which can be set up with different decimation policies stating when to trigger decimation, which variables to decimate, and which value to assign to decimated variables. In this paper, we propose a library of such policies that can be combined to produce different versions of DeciMaxSum. Our empirical results on different benchmarks show that some combinations of decimation policies outperform classical Max-Sum and its extension Max-Sum_AD_VP, specifically design to handle loops. DeciMaxSum outputs better quality solutions in a reasonable number of message propagation. There are several paths to future research. First, we only explore a limited set of decimation policies. We wish to investigate more complex ones, especially policies trigger when loops are detected by agents. In fact, since our overarching goal is to cope with loops, detecting them at the agent level seems a reasonable approach to initiate decimation in a cyclic network. This approach will require agents to implement cycle-detection protocol, by sending message history, while propagating marginals. In such a setting, several decimation election may arise concurrently in the graph. Second, we would like to generalize DeciMaxSum framework to consider MaxSum_AD_VP as a particular case of decimation: iterated decimation. Finally, we plan to applied 12 Cerquides et al. DeciMaxSum on real world applications, with strong loopy nature, like the coordination of smart objects in IoT [17] or decentralized energy markets in the smart grid [2]. References 1. Cerquides, J., Farinelli, A., Meseguer, P., Ramchurn, S.D.: A tutorial on optimization for multi-agent systems. The Computer Journal 57(6), 799–824 (2014), http://dx.doi.org/10.1093/comjnl/bxt146 2. Cerquides, J., Picard, G., Rodríguez-Aguilar, J.: Designing a marketplace for the trading and distribution of energy in the smart grid. In: 14th International Conference on Autonomous Agents and Multiagent Systems (AAMAS). pp. 1285–1293. International Foundation for Autonomous Agents and Multiagent Systems (2015), http://www.aamas-conference. org/Proceedings/aamas2015/forms/contents.htm#I4 3. Farinelli, A., Rogers, A., Petcu, A., Jennings, N.R.: Decentralised coordination of low-power embedded devices using the max-sum algorithm. In: International Conference on Autonomous Agents and Multiagent Systems (AAMAS’08). pp. 639–646 (2008), http://dl.acm.org/citation.cfm?id=1402298.1402313 4. Kolmogorov, V.: Convergent tree-reweighted message passing for energy minimization. IEEE Transactions on Pattern Analysis and Machine Intelligence 28(10), 1568–1583 (Oct 2006) 5. Krzakala, F., Montanari, A., Ricci-Tersenghi, F., Semerjian, G., Zdeborova, L.: Gibbs states and the set of solutions of random constraint satisfaction problems. Proceedings of the National Academy of Science 104, 10318–10323 (Jun 2007) 6. Lynch, N.: Disributed Algorithms. Morgan Kaufmann (1996) 7. Mackay, D.J.C.: Information Theory, Inference and Learning Algorithms. Cambridge University Press, first edition edn. (Jun 2003) 8. Maheswaran, R., Pearce, J., Tambe, M.: Distributed algorithms for dcop: A graphical-game-based approach. In: Proceedings of the 17th International Conference on Parallel and Distributed Computing Systems (PDCS), San Francisco, CA. pp. 432– 439 (2004) 9. Modi, P.J., Shen, W., Tambe, M., Yokoo, M.: ADOPT: Asynchronous Distributed Constraint Optimization with Quality Guarantees. Artificial Intelligence 161(2), 149–180 (2005) 10. Modi, P., Shen, W., Tambe, M., Yokoo, M.: ADOPT: Asynchronous distributed constraint optimization with quality guarantees. Artificial Intelligence Journal (2005) 11. Montanari, A., Ricci-Tersenghi, F., Semerjian, G.: Solving constraint satisfaction problems through belief propagationguided decimation. CoRR abs/0709.1667 (2007), http://arxiv.org/abs/0709.1667 12. Mooij, J.M.: libDAI: A free and open source C++ library for discrete approximate inference in graphical models. Journal of Machine Learning Research 11, 2169–2173 (Aug 2010), http://www.jmlr.org/papers/volume11/mooij10a/ mooij10a.pdf 13. Mézard, M., Montanari, A.: Information, Physics, and Computation. Oxford University Press (2009) 14. Navlakha, S., Barth, A.L., Bar-Joseph, Z.: Decreasing-rate pruning optimizes the construction of efficient and robust distributed networks. PLOS Computational Biology 11(7), 1–23 (07 2015), http://dx.doi.org/10.1371%2Fjournal. pcbi.1004347 15. Petcu, A., Faltings, B.: A scalable method for multiagent constraint optimization. In: International Joint Conference on Artificial Intelligence (IJCAI’05). pp. 266–271 (2005) 16. Rogers, A., Farinelli, A., Stranders, R., Jennings, N.: Bounded approximate decentralised coordination via the max-sum algorithm. Artificial Intelligence 175(2), 730 – 759 (2011), http://www.sciencedirect.com/science/article/ pii/S0004370210001803 17. Rust, P., Picard, G., Ramparany, F.: Using message-passing DCOP algorithms to solve energy-efficient smart environment configuration problems. In: International Joint Conference on Artificial Intelligence (IJCAI). AAAI Press (2016) 18. Vinyals, M., Pujol, M., Rodríguez-Aguilar, J., Cerquides, J.: Divide and coordinate: solving dcops by agreement. In: International Joint Conference on Autonomous Agents and Multi-Agent Systems (AAMAS’10). pp. 149–156. IFAAMAS, Canada (2010) 19. Vinyals, M., Rodriguez-Aguilar, J., Cerquides, J.: Constructing a unifying theory of dynamic programming dcop algorithms via the generalized distributive law. Autonomous Agents and Multi-Agent Systems 22(3), 439–464 (2010), http://dx. doi.org/10.1007/s10458-010-9132-7 20. Zivan, R., Peled, H.: Max/min-sum distributed constraint optimization through value propagation on an alternating DAG. In: van der Hoek, W., Padgham, L., Conitzer, V., Winikoff, M. (eds.) International Conference on Autonomous Agents and Multiagent Systems, AAMAS 2012, Valencia, Spain, June 4-8, 2012 (3 Volumes). pp. 265–272. IFAAMAS (2012), http://dl.acm.org/citation.cfm?id=2343614
2cs.AI
arXiv:1410.5170v3 [math.ST] 1 Jul 2016 Robust and Efficient Parameter Estimation based on Censored Data with Stochastic Covariates Abhik Ghosh and Ayanendranath Basu Indian Statistical Institute [email protected], [email protected] Abstract Analysis of random censored life-time data along with some related stochastic covariables is of great importance in many applied sciences like medical research, population studies and planning etc. The parametric estimation technique commonly used under this set-up is based on the efficient but non-robust likelihood approach. In this paper, we propose a robust parametric estimator for the censored data with stochastic covariates based on the minimum density power divergence approach. The resulting estimator also has competitive efficiency with respect to the maximum likelihood estimator under pure data. The strong robustness property of the proposed estimator with respect to the presence of outliers is examined and illustrated through an appropriate simulation study in the context of censored regression with stochastic covariates. Further, the theoretical asymptotic properties of the proposed estimator are also derived in terms of a general class of M-estimators based on the estimating equation. Keywords: Censored Data; Robust Methods; Linear Regression; Density power divergence; M-Estimator; Exponential Regression Model, Accelerated Failure Time Model. 1 Introduction It is often necessary to analyze life-time data in many applied sciences including medical sciences, population studies, planning etc. For these survival analyses problems, researchers often cannot observe the full data because some of the respondents may leave the study in between or some may be still alive at the end of the study period. Statistical modelling of such data involves the idea of censored distributions and random censoring variables. Mathematically, let Y1 , . . . , Yn be n independent and identically distributed (i.i.d.) observations from the population with unknown life-time distribution GY . We assume that the observations are censored by a censoring distribution GC independent of GY and C1 , . . . , Cn denote n i.i.d. sample observations from GC . We only observe the portion of Yi s (right) censored by Ci s, i.e., we observe Zi = min (Yi , Ci ) and δi = I(Yi ≤ Ci ), i = 1, . . . , n, where I(A) denote the indicator function of the event A. Based on these data (Zi , δi ), our aim is to do inference about the lifetime distribution GY . Suppose Z(i,n) denotes the i-th order statistic in {Z1 , · · · , Zn } and δ[i,n] is the value of corresponding δ (i-th concomitant). The famous product-limit (non-parametric) estimator of GY under this set-up had been derived by Kaplan 1 and Meier (1958), which is given by d G Y (y) = 1 − n  Y i=1 δ[i,n] 1− n−i+1 I(Z(i,n) ≤y) . It can be seen that, under suitable assumptions, the above product-limit estimator is in fact the maximum likelihood estimator of the distribution function in presence of censoring and enjoys several optimum properties. Many researchers have proved such properties and also extended it for different complicated inference problems with censored data; for example, see Petersen (1977), Chen et al. (1982), Campbell and Földes (1982), Wang et al. (1986), Tsi et al. (1987), Dabrowska (1988), Lo et al. (1989), Zhou (1991), Cai (1998), Satten and Datta (2001) among many others. In this present paper we further assume the availability of a set of uncensored covariables X ∈ Rp that are associated with our target response Y ; i.e., for each respondent i we have observed the values Xi along with (Zi , δi ). These covariables are generally the demographic conditions of the subject or some measurable indicator of the response variable (e.g., medical diagnostic measures like blood pressure, hemoglobin content, etc., for clinical trial responses). Let us assume that the distribution function of these i.i.d. covariates is GX and their joint distribution function with Y is G so that Z Z GY (y) = G(x, y)dx = GY |X=x (y)GX (x)dx, where GY |X=x is the conditional distribution function of Y given X = x. Instead of inferring about the response Y alone, here we are more interested in obtaining the association between response and covariables through the conditional distribution GY |X ; the distribution GX of covariates is often of interest but sometimes it may act as a nuisance component too. Let us denote the i-th concomitant of X associated with Z(i,n) by X[i,n] . Under this set-up, Stute d (1993) has extended the Kaplan-Meier product limit (KMPL) estimator G Y (y) to obtain a nonparametric estimate of the joint multivariate distribution function G given by b y) = G(x, n X i=1 Win I(X[i,n] ≤ x, Z(i,n) ≤ y) where the weights are calculated as Win δ[j,n] i−1  δ[i,n] Y n−j . = n−i+1 n−j +1 j=1 Note that when there is no censoring at all, i.e., δi = 1 for all i, then Win = n1 for each i so b y) coincide with respective empirical distribution functions. Further, above that d GY (y) and G(x, estimator is also self-adjusted in presence of any ties in the data. These give the framework for non-parametric inference based on the censored data with covariates. Stute (1993, 1996) proved b y) and several asymptotic properties like strong consistency, asymptotic distribution of G(x, related statistical functionals. Several non-parametric and semi-parametric inference procedures b y) are widely used in real life applications. using G(x, However, for many applications in medical sciences, one may know the parametric form of the distribution of the survival time (censored responses) possibly through previous experience 2 in similar context (similar drugs or similar diseases may have studied in the past). In such cases, the use of a fully parametric model is much more appropriate over the semi-parametric or non-parametric models. Many advantages of a fully parametric model for the regression with censored responses had been illustrated in Chapter 8 of Hosmer et al. (2008) which include – (a) greater efficiency due to the use of full likelihood, (b) more meaningful estimates of clinical effects with simple interpretations, (c) prediction of the response variable from the fitted model etc. The most common parametric models used for the analysis of survival data are the exponential, Weibull or log-logistic distributions. Many researchers had used such parametric models to analyze survival data more efficiently; see for example, Cox and Oakes (1984), Crowder et al. (1991), Collett (2003), Lawless (2003), Klein and Moeschberger (2003) among others. The robustness issue with the survival data, on the other hand, has got prominent attention very recently. The size and availability of survival data has clearly been growing in recent times in biomedical and other industrial studies which often may contain few erroneous observations or outliers and it is very difficult to sort out those observations in presence of complicated censoring schemes. Some recent attempts have been made to obtain robust parametric estimators based on survival data without any covariates. For example, Wang (1999) has derived the properties of M-estimators for univariate life-time distributions and Basu et al. (2006) have developed a more efficient robust parametric estimator by minimizing the density power divergence measure (Basu et al., 1998). These estimators, along with the automatic control for the effect of outlying observations, provide a compromise between the most efficient classical parametric estimators like maximum likelihood or method of moments and the inefficient non-parametric or semiparametric approaches provided there is no significant loss of efficiency under pure data. The present paper extends this idea to develop such robust estimators for the model parameters under censored response with covariates. It does not follow directly from the existing literature as we need to change the laws of large number and central limit theorem for censored data suitably in the presence of covariates. It is to be noted that, in this paper we consider a fully parametric model for survival data with covariates, which is not the same as the usual semi-parametric or nonparametric regression models like the Cox proportional hazard model (Cox, 1972) or the Buckley-James linear regression (Buckley and James, 1979; Ritov, 1990). The latter methods are generally more robust with respect to model misspecification but less efficient compared to a fully parametric models. Further, many recent attempts have been made to develop inference under such semiparametric models that are robust also with respect to outliers in the data; e.g., Zhou (1992), Bednarski (1993), Kosorok et al. (2004), Bednarski and Borowicz (2006), Salibian-Barrera and Yohai (2008), Farcomeni and Viviani (2011) etc. However, no such work has been done to develop robust inference under fully parametric regression models with survival response. The few works that closely relate to the proposal of our present paper are by Zhou (2010), Locatelli et al. (2011) and Wang et al. (2015), who have proposed some robust solutions for a particular case of accelerated failure time regression model without any parametric assumptions on the stochastic covariates; the first two papers propose the M-estimators and S-estimators respectively for the semi-parametric AFT models with parametric location-scale error and the third one extends the M-estimators further to achieve robustness with respect to leverage points and simultaneous robust estimation of the error variance. In Zhou (2010) there is the possibility of dependence between the error and stochastic explanatory variables, but independence is also allowed in which case it considers the same model set-up as in Locatelli et al. (2011). However, none of these approaches considered fully parametric models with suitable assumptions on the marginal covariate distributions. The present paper fill this gap in the literature of survival anal3 ysis by proposing a simple yet general robust estimation criterion with more efficiency under any general parametric model for both the censored response and the stochastic covariates. Further, the proposal of the present paper is fully general with respect to model assumptions and can be easily extended to the semi-parametric models considered in the existing literature; we will show that the existing versions of the M-estimators of Zhou (2010) and Wang et al. (2015) can be considered as a special case of our proposal under the semi-parametric extension. In that sense, our proposed method in the present paper will give a complete general framework for all possible modeling options of censored data with stochastic covariates along with the possibility of more efficient inference through fully parametric covariate distributions. The rest of the paper is organized as follows. We start with a brief review of background b y) and the minimum density concepts and results about the non-parametric estimator G(x, power divergence estimators in Section 2. Next we consider a general parametric set-up for the censored lifetime data with covariates as described above and propose the modified minimum density power divergence estimator for the present set-up in Section 3; its application in the context of simple linear and exponential regression models with censored response and for the general parametric accelerated failure time model are also described in this section. In Section 4, we derive theoretical asymptotic properties for a general class of estimators containing the proposed minimum density power divergence estimator under the present set-up; this general class of estimators is indeed a suitable extension of the M-estimators. The global nature of our proposal and its generality are discussed in Section 5 along with the illustration of this extension in the semi-parametric set-up. Section 6 contains the robustness properties of the proposed MDPDE and the general M-estimators examined through the influence function analysis for both fully parametric and semi-parametric set-ups. The performance of the proposed minimum density power divergence estimator in terms of both efficiency and robustness is illustrated through some appropriate simulation studies in Section 7. Some remarks on the choice of the tuning parameter in the proposed estimator are presented in Section 8, while the paper ends with a short concluding remark in Section 9. 2 2.1 Preliminary Concepts and Results b y) Asymptotic Properties of G(x, One of the main barriers to derive any asymptotic results based on survival data was the unavailability of limit theorems, like law of iterated logarithm, law of large number, central limit theorem etc., under censorship. This problem has been solved in the recently decades mainly through the works of Stute and Wang; see Stute and Wang (1993), Stute (1995) for such limit theorems for the censored data without covariates and Stute (1993, 1996) for similar results in presence of covariables. In this section, we briefly describe some results from the later works with covariates that will be needed in this paper. Assume the set-up of life-time variable Y censored by an independent censoring variable C as discussed in Section 1. Denote Z = min(Y, C); the distribution function of Z is given by GZ = 1 − (1 − GY )(1 − GC ). In order to have the limiting results for this set-up, we need to make the following basic assumptions: (A1) The life-time variable Y and the censoring variable C are independent and their respective distribution functions GY and GC have no jump in common. 4 (A2) The random variable δ = I(Y ≤ C) and X are conditionally independent given Y, i.e. whenever the actual life-time is known the covariates provide no further information on censoring. More precisely, P (Y ≤ C|X, Y ) = P (Y ≤ C|Y ). Now, consider a real valued measurable function φ from Rp+1 to Rk and define Sn = n X Win φ(X[i,n] , Z(i,n) ) = i=1 Z b φ(x, y)G(dx, dy). (1) This functional Sn forms the basis of several estimators under this set-up. The results presented below describe its strong consistency and distributional convergence; see Stute (1993, 1996) for their proofs and details. Proposition 2.1 [Strong Consistency (Stute, 1993)] Suppose that φ(X, Y ) is integrable and Assumptions (A1) and (A2) hold for the above mentioned set-up. Then we have, with probability one and in the mean, Z Z φ(X, τGZ )dP, (2) φ(X, Y )dP + I(τGZ ∈ A) lim Sn = n→∞ Y =τGZ Y <τGZ where τGZ denote the least upper bound for the support of GZ given by τGz = inf {z : GZ (z) = 1}, and A denotes the set of atoms (jumps) of GZ . The above convergence can be written in a simpler form, by defining  G(x, y) if y < τ GZ e G(x, y) = G(x, τGZ −) + I(τGZ ∈ A)G(x, {τGZ }) if y ≥ τ GZ . Then, the convergence in (2) yields Z Z b e e say. φ(x, y)G(dx, dy) = φ(x, y)G(dx, dy) = S, lim n→∞ Further, note that the independence of Y and C gives τGZ = min(τGY , τGC ), where τGY and τGC are the least upper bound of the supports of GY and GC respectively. So, whenever τGY < τGC or e coincides with G and the estimator Sn becomes τGC = ∞, the modified distribution function G R a strongly consistent estimator of its population counterpart S = φ(x, y)G(dx, dy). Further, it follows that the Glivenko-Cantelli type strong uniform convergence of Ĝ to G holds under assumptions (A1) and (A2); see Corollary 1.5 of Stute (1993). Proposition 2.2 [Central Limit Theorem (Stute, 1996, Theorem 1.2)] Consider the above mentioned set-up with assumption (A2) and suppose that the measurable function φ(X, Y ) satisfies R (A3) [φ(X, Z)γ0 (Z)δ]2 dP < ∞, (A4) R e |φ(X, Z)| C 1/2 (w)G(dx, dw) < ∞, 5 where Z and  G0Z (dz ′ ) γ0 (z) = exp , with 1 − GZ (z ′ ) 0 Z w− GC (dz ′ ) C(w) = . [1 − GC (z ′ )][1 − GZ (z ′ )] 0 Then we have, as n → ∞, where z− √ G0Z (z) = P (Z ≤ z, δ = 0), D e → n(Sn − S) N (0, Σφ ), Σφ = Cov [φ(X, Z)γ0 (Z)δ + γ1 (Z)(1 − δ) − γ2 (Z)] , (3) (4) where γ1 and γ2 are vectors of the same length as φ and are defined as Z 1 e11 (dx, dw), I(z < w)φ(x, w)γ0 (w)G γ1 (z) = 1 − GZ (z) ZZ I(v < z, v < w)φ(x, w)γ0 (w) 0 e11 (dx, dw), and γ2 (z) = GZ (dv)G [1 − GZ (v)]2 e11 (x, z) = P (X ≤ x, Z ≤ z, δ = 1). with G Note that a consistent estimator of the above asymptotic variance can be obtained by using the corresponding sample covariance and by replacing the distribution functions in the definitions of γ0 , γ1 and γ2 by their respective empirical estimators. In this context, we should note that Assumption (A2) is strictly stronger than the usual assumptions in regression analysis for censored life-time data (Begun et al., 1983). This can b as a particular case of the inverse of the probability of be seen by writing Stute’s estimate G censoring weighted (IPCW) statistic, where the censoring weights are estimated by the marginal Kaplan-Meier estimator for the censoring time. This may lead to some biased inference when the censoring distribution depends on the covariates, but in such cases we cannot have robust results unless moving to the semi-parametric models like Cox regression. Further, Robins and Rotnitzky (1992) also assumed this stronger condition (A2) to develop a more efficient IPCW statistic under the semi-parametric set-up (see also Van der Laan and Robins, 2003). Zhou (2010) and Wang et al. (2015) have also considered the same assumption (A2) for robust estimation under semiparametric accelerated failure time models. So, while considering the fully parametric set-up throughout this paper, we continue with the assumption (A2) for deriving any asymptotic result; clearly it does not restrict the practical use of the proposed method in finite samples. 2.2 The Density Power Divergence and Corresponding Estimators The density power divergence based statistical inference has become quite popular in recent days due to its strong robustness properties and high asymptotic efficiency without using any non-parametric smoothing. The density power divergence measure between any two densities g and f (with respect to some common dominating measure) is defined in terms of a tuning parameter α ≥ 0 as (Basu et al., 1998),     Z  1 1+α 1 α 1+α   f g+ g , for α > 0, f − 1+ α α Z (5) dα (g, f ) =   g log(g/f ), for α = 0. 6 When we have n i.i.d. samples Y1 , . . . , Yn from a population with true density function g, modeled by the parametric family of densities F = {fθ : θ ∈ Θ ⊂ Rp }, the minimum density power divergence estimator (MDPDE) of the parameter θ is to be obtained by minimizing the density power divergence between the data and the model family; or equivalently by minimizing Z fθ1+α (y)dy 1+α − α Z fθα(y)dGn (y) = Z n fθ1+α (y)dy − 1+α1 X α fθ (Yi ), α n (6) i=1 with respect to θ; here Gn is the empirical distribution function based on the sample. See Basu et al. (1998, 2011) for more details and other properties of the MDPDEs. It is worthwhile to note that the MDPDE corresponding to α = 0 coincides with the maximum likelihood estimator (MLE); the MDPDEs become more robust but less efficient as α increases, although the extent of loss is not significant in most cases with small positive α. Thus the parameter α gives a trade-off between robustness and efficiency. Hong and Kim (2001) and Warwick and Jones (2005) have presented some data-driven choices for the selection of optimal tuning parameter α. The MDPDE has been applied to several statistical problems and has been extended suitably for different types of data. For example, Kim and Lee (2001) have extended it to the case of robust estimation of extreme value index, Lee and Song (2006, 2013) have provided extensions in the context of GARCH model and diffusion processes respectively and Ghosh and Basu (2013, 2014) have generalized it to the case of non-identically distributed data with applications to the linear regression and the generalized linear model. In the context of survival analysis, Basu et al. (2006) have extended the concept of MDPDE for censored data without any covariates to obtain a robust and efficient estimator. Based on n i.i.d. right censored observations Y1 , . . . , Yn as above, Basu et al. (2006) have proposed to use the Kaplan-Meier product limit estimator d GY in place of the empirical distribution function Gn in (6) and derived the properties of the corresponding MDPDEs. In the next section, we will further generalize this idea to obtain robust estimators for a joint parametric model based on censored data with covariates. 3 3.1 The Minimum Density Power Divergence Estimation (MDPDE) under Random Censoring with Covariates General Parametric Models and Estimating Equations Let us consider the set-up of Section 1. We are interested in making some inference about the distribution of the lifetime variable Y and its relation with the covariates (through the distribution GY |X ) based on the survival data with covariates (Zi , δi , Xi ). Sometimes one may also be interested in the distribution GX of the covariates. As noted earlier, this paper focuses on the parametric approach of inference; so we assume two model family of distributions for GY |X and GX given by FX = {Fθ (y|X) : θ ∈ Θ ⊆ Rq } and F′ = {FX,γ (x) : γ ∈ Γ ⊆ Rr } respectively. Then the target parameters of interest are θ and γ which we will estimate jointly based on (Zi , δi , Xi ). The case of known γ can be easily derived from this general case or from the work of Basu et al. (2006). The most common and popular method of estimation is the maximum likelihood estimator (MLE) that is obtained by maximizing the probability of the observed data (Zi , δi , Xi ) with respect to the parameters (θ, γ). However, in spite of several optimal properties, the MLE has well-known drawback of the lack of robustness. As noted in the previous section, the minimum density power divergence estimator can be used as a robust alternative to the MLE 7 with no significant loss in efficiency under pure data for several common problems. Deriving the motivation from these, specially from the work of Basu et al. (2006), here we consider the minimum density power divergence estimator (MDPDE) of (θ, γ) obtained by minimizing the objective function (6) for the joint density of the variables (Y, X) and a suitable estimator of this joint distribution. Let us denote the density of Fθ (y|X) by fθ (y|X) and so on. Then the joint model density of (X, Y ) is fθ (y|x)fX,θ (x). As an estimator of their true joint distribution b y), because of its optimality properties as described in Section 2.1. G we use the KMPL G(x, Thus, for any α > 0, the objective function to be minimized with respect to (θ, γ) is given by ZZ ZZ 1+α 1+α 1+α b y) Hn,α (θ, γ) = fθ (y|x) fX,γ (x) dxdy − fθ (y|x)α fX,γ (x)α dG(x, α ZZ n 1+α X Win fθ (Z(i,n) |X[i,n] )α fX,γ (X[i,n] )α . (7) = fθ (y|x)1+α fX,γ (x)1+α dxdy − α i=1 For  the case α =0, the MDPDE of (θ, γ) is to be obtained by minimizing the objective function lim Hn,α(θ, γ) , or equivalently α↓0 Hn,0 (θ, γ) = − n X i=1   Win log fθ (Z(i,n) |X[i,n] )fX,γ (X[i,n] ) . (8) The estimator obtained by minimizing (8) is nothing but the maximum likelihood estimator of (θ, γ) under the present set-up. Therefore, the proposed MDPDE is indeed a generalization of the MLE. The estimating equations of the MDPDE of (θ, γ) are then given by ∂Hn,α (θ, γ) = 0, ∂θ ∂Hn,α (θ, γ) = 0. ∂γ α ≥ 0. For α > 0, routine differentiation simplifies the estimating equations to yield ζθ − n X i=1 Win uθ (Z(i,n) , X[i,n] )fθ (Z(i,n) |X[i,n] )α fX,γ (X[i,n] )α = 0, ζγ − where n X i=1 Win uγ (X[i,n] )fθ (Z(i,n) |X[i,n] )α fX,γ (X[i,n] )α = 0, ζθ = ζγ = ZZ (9) (10) uθ (y, x)fθ (y|x)1+α fX,γ (x)1+α dxdy, ZZ uγ (x)fθ (y|x)1+α fX,γ (x)1+α dxdy, n X Win uθ (Z(i,n) , X[i,n] ) = 0, ∂ ln f (x) γ θ (y|x) with uθ (y, x) = ∂ ln f∂θ and uγ (x) = being the score functions corresponding to θ and ∂γ γ respectively. For α = 0, the corresponding estimating equation obtained by differentiating Hn,0 has the simpler form given by i=1 n X Win uγ (X[i,n] ) = 0, i=1 8 which can also be obtained from Equations (9) and (10) by substituting α = 0; note that, at α = 0, ζθ = 0 and ζγ = 0. Therefore, Equations (9) and (10) represent the estimating equations for all MDPDEs with α ≥ 0. Definition 3.1 Consider the above mentioned set-up. The minimum density power divergence estimator of (θ, γ) based on the observed data (Zi , δi , Xi ), i = 1, . . . , n is defined by the simultaneous root of the equations (9) and (10). If there are multiple roots of these equations, the MDPDE will be given by the root which minimizes the objective function (7) for α > 0, or (8) for α = 0. Clearly, the MDPDE is Fisher consistent by its definition and the estimating equations (9) and (10) are unbiased at the model. Further, there will not be any problem of root selection in case of multiple roots, which is a general issue in the inferences based on estimating equation. This is because we have a proper objective function in the case of MDPDE. Note that the MDPDE estimating equations (9) and (10) can also be written as ZZ b y) = ψ(y, x; θ, γ)dG(x, n X Win ψ(Z(i,n) , X[i,n] ; θ, γ) = 0, (11) i=1 where ψ(y, x; θ, γ) = (ψ1,α (y, x; θ, γ), ψ2,α (y, x; θ, γ))T with ψ1,α (Y, X; θ, γ) = ζθ − uθ (Y, X)fθ (Y |X)α fX,γ (X)α , ψ2,α (Y, X; θ, γ) = ζγ − uγ (X)fθ (Y |X)α fX,γ (X)α .  (12) This particular estimating equation (11) is similar to that of the M-estimator for i.i.d. nonb is the empirical distribution censored data with covariates (in fact they become the same if G function). So, extending the concept of Wang (1999), we can define the general M-estimator of θ and γ based on any general function ψ(y, x; θ, γ) : R × Rp × Rq × Rr 7→ Rq × Rr (13) as the solution of the estimating equation (11). However, to make it an unbiased estimating equation, we consider only the ψ-functions for which ZZ ψ(y, x; θ, γ)dG(x, y) = 0. (14) Definition 3.2 Consider the above mentioned parametric set-up for censored data with stochastic covariates. Also, consider a general ψ-function as in (13) satisfying the condition (14). An M-estimator of (θ, γ) corresponding to this general ψ-function based on the observed data (Zi , δi , Xi ), i = 1, . . . , n, is defined as the root of the estimating equation (11). Note that any general M-estimator is also Fisher consistent and is based on an unbiased estimating equation, by definition. But, in general, they may suffer from the problem of multiple roots and need a proper numerical techniques (like bootstrapping) to get a well-defined Mestimator. So, in this paper, we restrict our attention mainly to examine the performances of the robust MDPDE corresponding to the particular ψ-function defined in (12). However, we derive theoretical asymptotic properties of the general M-estimators in Section 4 and deduce the properties of MDPDEs from those general results in Section 4.2. 9 3.2 Application (I): Fully Parametric version of Linear Regression model We first consider the simplest problem of linear regression with censored responses and stochastic covariables. Precisely, we assume the linear regression model (LRM) Yi = XiT θ + ǫi , i = 1, . . . , n, (15) where Yi is the censored response (generally the lifetime), Xi is a p-variate stochastic auxiliary variable associated with the response, θ is the vector of unknown regression coefficients and ǫi is the error in specified linear model. We assume that the error ǫi s are independent and identically distributed with distribution function Fe and Xi s are independent of the errors having distribution function FX,γ . Generally we can assume both symmetric error distributions like the normal as well as asymmetric error distributions like the exponential; however the second group is used in most reliability applications. Then, the conditional distribution of the response variable Yi given Xi is Fθ (y|Xi ) = Fe (y − XiT θ). Now, we consider the incomplete censored observations (Zi , δi ) as defined in Section 1 and use them to estimate (θ, γ) robustly and efficiently. This inference problem clearly belongs to the general set-up considered in the previous subsection and frequently arises in reliability studies and other applied researches. We can obtain a robust solution to this problem through the proposed MDPDEs, obtained by just solving the estimating equations (9) and (10). Here we present the detail working for one particular example of model families FX and F′ . The case of other model families can also be tackled similarly. Suppose the response variable is exponentially distributed with mean depending on the co variates as E[Y |X] = X T θ; then FX = Exp(X T θ) : θ ∈ Rp where Exp(τ ) represents the exponential distribution with mean τ . Also, for simplicity, let us assume that the auxiliary variables are independent to each other and normally distributed so that F′ = {Np (γ, Ip ) : γ ∈ Rp }. In this case, the objective function Hn,α (θ, γ) of the MDPDE has a simpler form given by n e−αψi (θ,γ) 1+α X (1 + α)−3/2 (0) W ψ (θ, γ) − in T θ)α (2π)α/2 (2π)α/2 α i=1 (X[i,n]   n X 1 T Win ψi (θ, γ) + log(X[i,n] θ) + log(2π) , Hn,0 (θ, γ) = 2 Hn,α (θ, γ) = and α > 0, (16) (17) i=1 R 1 where ψ (0) (θ, γ) = (xT θ)−α Np (x, γ, 1+α Ip )dx and ψi (θ, γ) = Z(i,n) T X[i,n] θ + 21 (X[i,n] − γ)T (X[i,n] − γ). Note that the integral ψ (0) (θ, γ) is just the expectation of a simple function of multivariate normal random variable; so it can be computed quite easily using standard numerical integration techniques. Therefore, we can simply minimize the above objective functions by any numerical algorithm to obtain the MDPDE at any α ≥ 0. Alternatively, we can obtain the MDPDE by solving the estimating equations (9) and (10), which can also be simplified in this particular situation as ! n X Z 1 α (i,n) Win e−αψi (θ,γ) ψ̄ (0) (θ, γ) = − X[i,n] , T θ)(1+α) T θ)2 (1 + α)5/2 (X[i,n] (X[i,n] i=1 n h i X (X[i,n] − γ) α (0) (0) Win e−αψi (θ,γ) , ψ̄ (θ, γ) − γψ (θ, γ) = T θ)α 5/2 (1 + α) (X[i,n] i=1 R 1 where ψ̄ (0) (θ, γ) = (xT θ)−α xNp (x, γ, 1+α Ip )dx. The case of α = 0 (MLE) can be simplified further where the estimator of γ becomes independent of the parameter θ. To see this, we 10 simplify the above estimating equations at α = 0 as   T θ n Z(i,n) − X[i,n] X Win X[i,n] = 0, T θ)2 (X [i,n] i=1 n X i=1  Win X[i,n] − γ = 0. Pn W X i=1 Pn in [i,n] , i=1 Win Solving the second equation, we get that γ b= which clearly does not depend on θ, as is expected from the theory of maximum likelihood inference. The semi-parametric version of this model has been considered in Zhou (2010), where no assumption has been made about the distribution of the i.i.d. sequences {Xi } and {ǫi }. In that paper, an M-estimator of θ has been proposed by solving the estimating equation n X i=1   T θ = 0, Win ψ0 Z(i,n) − X[i,n] (18) for suitable choices of ψ0 . Although this M-estimator is different from the proposed MDPDE, it in fact belongs to our general class of M-estimators (Definition 3.2) as we will show in detail in section 5. Further, we will show in Section 6 that our MDPDE with α > 0 under fully parametric set-up is robust with respect to both the outliers in response and the leverage points in the explanatory variables; but the M-estimator of Zhou (2010) is not robust with respect to the leverage points. 3.3 Application (II): Fully Parametric Exponential Regression model for Medical Sciences The simple linear regression model considered in the previous subsection is the most popular inference problem under the set-up considered in this paper. However, although it is simple and potentially applicable in several real life problems, for the purpose of serving the typical applications in the medical science this simple linear model is rarely used. The reason is that in almost all medical applications the support of the distribution of the censoring times is shorter than the support of the lifetimes (we cannot follow the patients until they die). For this reason the linear model is usually not identifiable and in order to become applicable the setup of the proposed method needs to introduce a truncation time, say τ , such that the probability to be uncensored by time τ is strictly greater than zero for all x. We can suitably extend the proposed MDPDE to cover these assumptions through some more routine calculations, which we leave for the readers. In this section we present an alternative multiplicative model with exponential error for the applications in medical sciences. This particular model, known as the exponential regression model, is widely used and most popular in the medical sciences and related applications. More precisely, let us assume the multiplicative regression model for the survival times (responses) Yi , i = 1, . . . , n as T Yi = eXi θ × ǫi , i = 1, . . . , n, (19) where Xi is a p-variate stochastic auxiliary variable associated with the response, θ is the vector of unknown regression coefficients and ǫi is the error in the specified linear model. Such a multiplicative model ensures the positivity of the the response variables, which are generally lifetime in most applications. In the exponential regression model, we assume that the error variable ǫ is exponentially distributed with mean 1. Then the conditional distribution of the response 11 T variable Y given the covariate X is also withomean E[Y |X] = eX θ ; so considering n exponential T the notations of Section 1, FX = Exp(eX θ ) : θ ∈ Rp . Also, the Xi s are independent of the errors having distribution function FX,γ . Our objective is to estimate (θ, γ) robustly and efficiently based on the incomplete (censored) observations (Zi , δi , Xi ) as defined in Section 1. Again this inference problem belongs to the general set-up of Section 3.1 so that the proposed MDPDEs provide a robust solution to it. In the case of independent and normally distributed covariates with F′ = {Np (γ, Ip ) : γ ∈ Rp }, we can simplify the objective function Hn,α(θ, γ), to be minimized in order to obtain the MDPDE, as α(γ T θ)+ α2 (θ T θ) n 2(1+α) e (1 + α) X Hn,α (θ, γ) = Win e−αΓi (θ,γ) − 3/2 α/2 (1 + α) (2π) (2π)α/2 α i=1   n X 1 Win Γi (θ, γ) + log(2π) , Hn,0 (θ, γ) = 2 and α > 0, (20) (21) i=1 XT θ T T θ) + 1 (X where Γi (θ, γ) = Z(i,n) e [i,n] + (X[i,n] [i,n] − γ) (X[i,n] − γ). This objective function can 2 be easily minimized using any standard numerical techniques for any α ≥ 0. The estimating equations (9) and (10) can also be simplified for this particular case of ERM with normal covariates. For α > 0, they have the form n X −αΓi (θ,γ) Win e i=1   (X T θ) Z(i,n) e [i,n] + 1 X[i,n] = n X i=1 Win e−αΓi (θ,γ) X[i,n] − γ  =   α2 α αθ (θ T θ) α(γ T θ)+ 2(1+α) γ + , e 5/2 (1 + α) (1 + α) α2 α2 θ α(γ T θ)+ 2(1+α) (θ T θ) e . (1 + α)5/2 At α = 0 (MLE), these estimating equations further simplifies to n X i=1   (X T θ) Win Z(i,n) e [i,n] + 1 X[i,n] = 0, n X i=1  Win X[i,n] − γ = 0, which again produce the same estimator of γ as in the case of the LRM and independent of the parameter θ. 3.4 Application (III): Fully Parametric Version of Accelerated Failure Time (AFT) Models The exponential regression model considered in the previous section can be linearized by taking natural logarithm of the response time: log(Yi ) = XiT θ + ǫ∗i , (22) where ǫ∗i = log(ǫ) follows the standard extreme value distribution. This model can be generalized by considering some alternative distribution for ǫ∗i , but with mean 0. When ǫ∗i follows an extreme value distribution with mean 0 and scale parameter σ, then Yi follows a Weibull distribution and the resulting regression model is known as the Weibull regression model (WRM). Other common distributions for ǫ∗i are logistic (survival time has log-logistic distribution), normal (survival times 12 are log-normal) etc. In such models, the covariate has a multiplicative effect of the response life-time and hence they are generally known as the accelerated failure time (AFT) model. Consider a general location scale model family {fµ,σ (x) = f ((x − µ)/σ)} for some known function f and let ǫ∗i have density f0,σ . Then, given the covariate X, Yi has density   1 log(y) − (X T β) f . y σ This is the general form of the parametric AFT regression model; taking f as standard extreme value distribution it simplifies to the Weibull regression model and so on. Suppose the distribution of the covariates Xi is modeled by the family FX,γ as in the earlier cases so that our target becomes the estimation of θ = (β, σ) and γ = µ based on the censored observations (Zi , δi , Xi ) as defined in Section 1. Now we can again minimize the objective function (7) with respect to the parameters to obtain their robust MDPDE. The exact form of this objective function can be obtained easily for any particular choice of f and any standard numerical algorithm provide us with the solution of this optimization problem. Further, as we will see through the numerical illustrations in Section 7, the proposed MDPDE provides highly robust solutions in presence of outliers in data with little loss in efficiency at small positive α. In the case of the accelerated failure time models, the robustness of the MDPDE is directly comparable with the alternative proposal of Locatelli et al. (2011). Note that, contrary to the Locatelli et al. (2011) approach, our proposed estimator can estimate the parameter (γ) in the distribution of covariates simultaneously with θ = (β, σ); although someone might not see it to be a big advantage as the parameter γ can be estimated separately in many cases. However, the major advantage of our proposal is its generality and computational simplicity for any kind of parametric model with censored survival data with stochastic covariates. We will see in Section 5 that, even if we ignore the estimation of the parameters of the marginal distribution of covariates, our proposal contains the existing proposals of Locatelli et al. (2011), a similar proposal of Zhou (2010) under the same set-up and also their extension in Wang et al. (2015). This vast generality is the major strength of our proposal over the existing literature; see Section 5 for more detailed and general discussions. 4 Asymptotic Properties Consider the models and set-up described in Section 3.1. First, we derive the asymptotic properties of the general M-estimator (θ̂n , γ̂n ) of (θ, γ) as defined in Definition 3.2 based on a (random) censored sample of size n. Let us assume that the true distributions belong to the corresponding model families with (θ0 , γ0 ) being the true parameter value. Define ZZ n X b y) = Win ψ(Z(i,n) , X[i,n] ; θ, γ), λn (θ, γ) = ψ(y, x; θ, γ)dG(x, i=1 and λG (θ, γ) = ZZ ψ(y, x; θ, γ)dG(x, y). Then λn (θ, γ) is the empirical version of λG (θ, γ); also by definition λn (θ̂n , γ̂n ) = 0 and λG (θ0 , γ0 ) = 0. In order to prove the asymptotic consistency and normality of the general Mestimator (θ̂n , γ̂n ), we use the results of Section 2. So, we will assume that the assumptions (A1) to (A4) holds true with φ replaced by ψ. Further, let us assume 13 e and G coincides. (A5) Either τGY < τGC or τGC = ∞ so that G (A6) The variance matrix Σψ , as defined in (4) with φ replaced by ψ, exists finitely (with all entries finite). Then Proposition 2.1 and Proposition 2.2 give the strong consistency and asymptotic normality of λn (θ, γ), which are summarized in the following lemma. Lemma 4.1 Consider the above set-up with an integrable function ψ and let (θ0 , γ0 ) be the true parameter value. Then, (i) Under Assumptions (A1), (A2) and (A5) we have, with probability one, lim λn (θ0 , γ0 ) = λG (θ0 , γ0 ) n→∞ (ii) Under Assumptions (A2) to (A6), the asymptotic distribution of ZZ √ √ b − G](x, y) n [λn (θ, γ) − λG (θ, γ)] = n ψ(y, x; θ, γ)d[G is normal with mean 0 and variance matrix Σψ . 4.1 Strong Consistency and Asymptotic Normality of General M-estimators Wang (1999) has proved the strong consistency and asymptotic normality results for the Mestimators based on only censored variable with no covariables. This section extend the theory to the case where covariables are present along with the censored response. The extensions are in the line of the corresponding results with no censoring (see Huber, 1981; Serfling, 1980). Further note that whenever γ is known the asymptotic properties of θ̂n follows from just a routine application of the results derived in Wang (1999); so here we assume γ to be known and derive the joint distribution of M-estimator of θ and γ. These results provide a general (asymptotic) theoretical framework to study the properties of a wide class of estimators of (θ, γ) depending on the estimating equations. Denote the j-th component of the function ψ by ψj for j = 1, . . . , q + r. Also, consider following (stronger) conditions on the nature of the function ψ. (A7) ψ(y, x; θ, γ) is continuous in (θ, γ) and also bounded. (A8) The population estimating equation λG (θ, γ) = 0 has an unique root given by (θ0 , γ0 ). (A9) There exists a compact set C in Rq × Rr satisfying ZZ inf ψj (y, x; θ, γ)dG(x, y) > 0, (θ,γ)∈C / j = 1, . . . , q + r. Now let us start with the strong consistency of the M-estimator (θ̂n , γ̂n ) by an extension of Theorem 3 of Wang (1999, page 307) under above conditions. The proof follows in the same line of Wang (1999) by replacing the corresponding SLLN, given in Proposition 1 of Wang (1999), by the part (i) of Lemma 4.1 in the present context; hence it is omitted for simplicity of presentation. 14 Theorem 4.2 Consider the above set-up with Assumptions (A1), (A2), (A5), (A7) and (A8). Then we have the following results. (i) There exists a sequence of M-estimators {(θ̂n , γ̂n )} satisfying the empirical estimating equation λn (θ, γ) = 0 that converges with probability one to (θ0 , γ0 ). (ii) Further if (A9) also holds true, then any sequence of M-estimators {(θ̂n , γ̂n )} satisfying λn (θ, γ) = 0 converges with probability one to (θ0 , γ0 ). Note that the first part (i) of Theorem 4.2 is just a multivariate extension of Lemma B of Serfling (1980, page 249) from the complete data case to the present case of censored data with covariates. Further, the additional condition (A9) in part (ii) makes any sequence of Mestimators satisfying the estimating equation (11) to eventually fall in a compact neighborhood of (θ0 , γ0 ). This result, even with the stronger conditions, becomes really helpful when the empirical estimating equation λn (θ, γ) = 0 has multiple roots and one could obtain different M-estimator sequences by applying different numerical equation solving techniques. Part (ii) of Theorem 4.2 ensures that all theses sequences of M-estimators will be strongly consistent for the unique root (θ0 , γ0 ) of the equation λG (θ, γ) = 0. Next we turn our attention to the asymptotic normality of M-estimators. In this regard, we will first present a useful lemma in terms of any real valued function g(y, x, θ0 , γ0 ). This is again a suitable extension of Lemma 1 of Wang (1999, page 307) to the present set-up and the proof follows similarly by replacing the corresponding SLLN (Proposition 1 of Wang) by Part (i) of Lemma 4.1. Assume the following condition about the function g(y, x, θ0 , γ0 ). (A10) For a real valued function g(y, x, θ0 , γ0 ), at least one of the following holds: (i) g(y, x, θ, γ) is continuous at (θ0 , γ0 ) uniformly in (y, x). (ii) As δ → 0, ZZ sup {(θ,γ):||(θ,γ)−(θ0 ,γ0 )||≤δ} |g(y, x, θ, γ) − g(y, x, θ0 , γ0 )| dG(x, y) = hδ → 0. (Here || · || denotes the Euclidean norm). (iii) g is continuous in (y, x) for for any fixed (θ, γ) in a neighborhood of (θ0 , γ0 ), and lim (θ,γ)→(θ0 ,γ0 ) ||g(y, x, θ, γ) − g(y, x, θ0 , γ0 )||v = 0. (Here || · ||v denotes the total variation norm). RR (iv) g(y, x, θ, γ)dG(x, y) is continuous at (θ, γ) = (θ0 , γ0 ), and g is continuous in (y, x) for (θ, γ) in a neighborhood of (θ0 , γ0 ), and lim (θ,γ)→(θ0 ,γ0 ) (v) RR ||g(y, x, θ, γ) − g(y, x, θ0 , γ0 )||v < ∞. g(y, x, θ, γ)dG(x, y) is continuous at (θ, γ) = (θ0 , γ0 ), and ZZ ZZ P g(y, x, θ, γ)dĜ(x, y) → g(y, x, θ, γ)dG(x, y) < ∞, uniformly for (θ, γ) in a neighborhood of (θ0 , γ0 ). 15 RR Lemma 4.3 Suppose g(y, x, θ0 , γ0 ) is a real valued function with g(y, x, θ0 , γ0 )dG(x, y) < ∞. Assume that the conditions (A1), (A2) and (A10) hold for g. Then, for any sequence P (θ̂n , γ̂n ) →(θ0 , γ0 ), we have ZZ ZZ P g(y, x, θ̂n , γ̂n )dĜ(x, y) → g(y, x, θ0 , γ0 )dG(x, y). Theorem 4.4 Consider the above set-up and assume that ψ is differentiable with respect to (θ, γ) in a neighborhood of (θ0 , γ0 ) and the matrix ZZ ∂ ΛG (θ0 , γ0 ) = ψ(y, x, θ, γ) dG(x, y), (23) ∂(θ, γ) (θ,γ)=(θ0 ,γ0 ) exists finitely and is non-singular. Further assume that the assumptions of Lemma 4.3 hold for g(y, x, θ, γ) = Λij G (θ0 , γ0 ), the (i, j)-th element of ΛG (θ0 , γ0 ), with i, j = 1, . . . , q + r. Then, under Assumptions (A2) to (A6) we have, for any sequence of M-estimators {(θ̂n , γ̂n )} satisfying λn (θ, γ) = 0 that converges in probability to (θ0 , γ0 ),   i √ h D n (θ̂n , γ̂n ) − (θ0 , γ0 ) → N 0, ΛG (θ0 , γ0 )−1 Σψ (G)ΛG (θ0 , γ0 )−1 . Proof: Since ψ is differentiable in (θ, γ) , so is the function λn (θ, γ). So an application of multivariate mean value theorem yields h i λn (θ̂n , γ̂n ) − λn (θ0 , γ0 ) = ΛĜ (ζ1n , ζ2n ) (θ̂n , γ̂n ) − (θ0 , γ0 ) , with ||(ζ1n , ζ2n ) − (θ0 , γ0 )|| < ||(θ̂n , γ̂n ) − (θ0 , γ0 )||. Further, by definition, λn (θ̂n , γ̂n ) = 0 and λG (θ0 , γ0 ) = 0. Hence we get, Z Z   −1 b − G](x, y) . (θ̂n , γ̂n ) − (θ0 , γ0 ) = − ΛĜ (ζ1n , ζ2n ) ψ(y, x; θ, γ)d[G However, it follows from Lemma 4.3 that each term of ΛĜ (ζ1n , ζ2n ) convergence in probability to the corresponding term of ΛG (θ0 , γ0 ). Then, an application of Slutsky’s theorem and Part (ii) of Lemma 4.1 completes the proof of the theorem.  It is to be noted that the asymptotic normality of the M-estimators require more conditions than that required for its strong consistency in terms of differentiability properties of the ψ function, but it avoid the strong assumptions (A7) – (A9) used in Theorem 4.2. In fact, to obtain the asymptotic distributional convergence of any sequence of M-estimators in this case, it is just enough to ensure their convergence to the true parameter value in probability. All the related conditions used here are in the same spirit with that used in Wang (1999) have been no covariables are present and were discussed extensively in that paper. Finally, note that the estimating equation of any general M-estimator can be solved through an appropriate numerical technique but the complexity in terms of the iterative procedure increases extensively for a complicated non-linear ψ-function. However, one can show that, for 16 √ the Newton-Raphson algorithm, if we start the iterations with some n-consistent estimator of (θ, γ) then the estimator obtained by just one iteration, known as the one-step M-estimator, will have the same asymptotic distribution as the fully iterated M-estimator even in case of censored data with covariables as considered here. This is a well-known property of the M-estimator in case of complete data. The following theorem present this precisely for our case; the proof follows by an argument similar to that of Theorem 6 of Wang (1999) replacing Proposition 1 and 2 of that paper by Part (i) and Part (ii) of Lemma 4.1 respectively. √ Theorem 4.5 Suppose the conditions of Theorem 4.4 hold true and let (θen , γ en ) is any n(1) (1) consistent estimate of the true parameter value (θ0 , γ0 ). Then, the one-step M-estimator (θn , γn ), defined as h i−1 (θn(1) , γn(1) ) = (θen , γ en ) − ΛĜ (θen , γ en ) λn (θen , γ en ), (24) has the same distribution as that of the M-estimator (θ̂n , γ̂n ) derived in Theorem 4.4. 4.2 Properties of the MDPDE Note that, the MDPDE is a particular M-estimator with the ψ-function given by (12) and so all the results derived in the previous subsection for general M-estimators also hold true for the MDPDEs. In particular MDPDEs are strongly consistent and asymptotically normal under the assumptions considered in Theorems 4.2 and 4.4. However, in this particular case of MDPDEs, we can closely investigate the required assumptions for the particular form of the ψ-function. Note that assumptions (A1), (A2) and (A5) are related to the censoring scheme under consideration and others are about the special structure of the ψ-function. Further, in this particular case of MDPDE, the ψ-function depends on the model density and its score function. So, conditions (A3), (A4) and (A6) can easily be shown to hold for most statistical models by using the existence of finite and continuous second order moments of the score functions with respect to the true distribution G. Similar differentiability conditions on the model and score functions further ensure the assumptions of Lemma 4.3. So, the asymptotic normality of the MDPDEs follows from Theorem 4.4 for most models provided we can prove its consistency. However, assumptions (A6)–(A9), required to prove the strong consistency in Theorem 4.2, are rather difficult one and may not always hold for the assumed model. Noting that, the asymptotic normality of MDPDEs, as obtained in Theorem 4.4, does not require its strong consistency (only convergence in probability is enough), we now present an alternative approach to prove the (weak) consistency for the particular case of MDPDEs under some simpler conditions. This approach is essentially due ot Lehmann (1983), and has been used by Basu et al. (1998) to prove the asymptotic properties of the MDPDEs under i.i.d. complete data and extended by many researchers later in the context of different inference problems. Here, we extend their approach further for the present case of censored data with covariates. Let us also relax the assumption that the true distribution G belongs to the model family in the sense of assumption (D1) below. Define ZZ 1+α fθ (Y |X)α fX,γ (X)α , V (Y, X; θ, γ) = fθ (y|x)1+α fX,γ (x)1+α dxdy − α so that the MDPDE of (θ, γ) is to be obtained by minimizing ZZ b y), Hn (θ, γ) = V (y, x; θ, γ)dG(x, 17 with respect to the parameters. Further, the ψ-function for the MDPDEs as given by Equation (12) satisfies ψ1 (Y, X; θ, γ) = ∂V (Y, X; θ, γ) , ∂θ ψ2 (Y, X; θ, γ) = ∂V (Y, X; θ, γ) . ∂γ (25) Now, let us assume the following conditions: (D1) The supports of the distributions Fθ and FX,γ for any value of X are independent of the parameters θ and γ respectively. The true distribution G(x, y) is also supported on the set A = {(x, y) : fθ (y|x)fX,γ (x) > 0}, on which the true density g is positive. (D2) There exists an open subset ω of the parameter space that contains the best fitting parameter (θ0 , γ0 ) and for all (θ, γ) ∈ ω and for almost all (x, y) ∈ A, the densities fθ and fX,γ are thrice continuously differentiable with respect to θ and γ respectively. RR RR (D3) The integrals fθ (y|x)1+α fX,γ (x)1+α dxdy and fθ (y|x)α fX,γ (x)α dG(x, y) can be differentiated three times and the derivatives can be taken under the integral sign. Further the ψ-function under consideration is finite. (D4) The matrix ΛG (θ0 , γ0 ) = = ZZ ZZ ∂ ψ(y, x, θ, γ) ∂(θ, γ) dG(x, y) (θ,γ)=(θ0 ,γ0 ) ∂2 V (y, x, θ, γ) ∂(θ, γ)2 dG(x, y), (θ,γ)=(θ0 ,γ0 ) exists finitely and is non-singular. (D5) For all (θ, γ) ∈ ω, each of the third derivatives of V (y, x, θ, γ) with respect to (θ, γ) is bounded by a function of (x, y), independent of (θ, γ), that has finite expectation with respect to the true distribution G. Theorem 4.6 Under Assumptions (A1), (A2), (A5) and (D1)–(D5), there exists a sequence of solutions {(θ̂n , γ̂n )} of the minimum density power divergence estimating equations (9) and (10) with probability tending to one, that is consistent for the best fitting parameter (θ0 , γ0 ). (Then, the asymptotic normality of this sequence {(θ̂n , γ̂n )} follows from Theorem 4.4 under the assumptions of that theorem.) Proof: We follow a similar argument to that in the proof of Theorem 6.4.1(i) of Lehman (1983). Consider the behavior of Hn (θ, γ), as a function of (θ, γ), on a sphere Qa having center (θ0 , γ0 ) and radius a. Then, to prove the existence part, it is enough to show that, for sufficiently small a, Hn (θ, γ) > Hn (θ0 , γ0 ), (26) with probability tending to one, for any point (θ, γ) on the surface of Qa . Hence, for any a > 0, Hn (θ, γ) has a local minimum in the interior of Qa and the estimating equations of the MDPDE have a solution {(θ̂n (a), γ̂n (a)} within Qa , with probability tending to one. 18 Now a Taylor series expansion of Hn (θ, γ) around (θ0 , γ0 ) yields q+r X ∂Hn (θ, γ) (ζi − ζi0 ) Hn (θ0 , γ0 ) − Hn (θ, γ) = − ∂ζi i=1 q+r X (θ,γ)=(θ0 ,γ0 ) ∂ 2 Hn (θ, γ) ∂ζi ζj − 1 2 + q+r 1 X ∂ 3 Hn (θ, γ) (ζi − ζi0 )(ζj − ζj0 )(ζk − ζk0 ) 6 ∂ζi ζj ζk (ζi − ζi0 )(ζj − ζj0 ) i,j=1 (θ,γ)=(θ0 ,γ0 ) i,j,k=1 = S1 + S2 + S3 , (say), (θ,γ)=(θ ∗ ,γ ∗ ) (27) where ζi and ζi0 are the i-th component of the parameter vectors (θ, γ) and (θ0 , γ0 ) respectively for all i = 1, . . . , q+r, and (θ ∗ , γ ∗ ) lies in between (θ, γ) and (θ0 , γ0 ) with respect to the Euclidean norm. By a direct extension of the arguments presented in the proof of Theorem 3.1 of Basu et al. (2006), we get, with probability tending to one, on Qa , |S1 | < (q + r)a3 , for all a > 0; S2 < −ca2 , for all a < a0 with some c, a0 > 0; and |S3 | < ba3 , for all a > 0 with some b > 0, using the assumptions (D1)–(D5) and Lemma 4.1 whenever necessary. Combining these, we get max(S1 + S2 + S3 ) < −ca2 + (b + q + r)a3 , c proving (26) holds. which is less that zero whenever a < b+q+r Finally, to show that one can choose a root of the estimating equations of MDPDEs independent of the radius a, consider the sequence of roots closest to the best fitting parameter (θ0 , γ0 ), which exists by continuity of Hn (θ, γ) as a function of (θ, γ). This sequence will also be consistent completing the proof of the theorem.  Note that Assumptions (D1)–(D5) are easier to check compared to the (stronger) Assumptions (A6)–(A9) and are the routine extensions of the corresponding assumptions [(A1)–(A5)] of Basu et al. (2006). 4.3 Other M-estimators with Different ψ-Functions Although our main focus in this paper is to study one particular M-estimator, namely the minimum density power divergence estimator (MDPDE), it opens the scope of many different M-estimators through the general results derived in Section 4.1. This general framework of parameter estimation based on some suitable estimating equation is well studied in case of complete data and several optimum robustness properties of these M-estimators has been proved for different classes of weight function; for example, see Huber (1981) and Hampel et al. (1986). In fact, there exists different class of ψ-function generating optimum solution in case of different problems. For example, in case of estimating the location parameter in a symmetric distribution, the ψ functions, that are odd in the targeted parameter, lead to such optimum M-estimation. However, as pointed out in Wang (1999), an optimum ψ-function for the complete data might not enjoy similar optimality for the censored data, even if there is no covariable presence. 19 The main reason is that the lifetime variables are not usually symmetric and neither belong to a location-scale family; rather it is usually asymmetric. The case of censored data with covariates, as considered here, is much more complicated and we can not directly pick a ψfunction from the theory of complete data. Wang (1999) presented some example of ψ-functions in the context of censored data with no covariates that can be extended in the present case with several covariables. However, their usefulness and optimality both in terms of efficiency and robustness need to be verified for the censored data cases with or without covariates. There need a lot of research in this area to suggest an optimum ψ-function under any suitable criteria of robustness or efficiency based on censored data. However, we believe that the minimum density power divergence estimator proposed here is quite sufficient for most practical situations since it produces highly robust estimators with only a slight loss in efficiency compared to the maximum likelihood estimator (as described in Section 7). Further, the estimating equation of MDPDEs can be solved by any simple numerical technique quite comfortably and has a simple interpretation in terms of the density power divergence. Thus, although some future research work may provide suitable ψ-function satisfying some optimality criteria with complicated form or estimation procedure, the MDPDE will still have its importance in many practical scenarios due to its simplicity. 5 Global nature and the Semi-parametric Extensions Although the main focus of our paper is the MDPDE under fully parametric set-up, the Mestimator defined in Definition 3.2 and its asymptotic theory derived in the previous section is completely general in the sense that it can also be applied to any semi-parametric or even non-parametric set-ups. To see this, just note that the general M-estimator is defined in terms of a ψ function that only need to satisfy Equation (14). Therefore, one can also consider the ψ functions, ψ(y, x; θ), involving no parametric assumptions on the distribution of x (and hence independent of parameter γ) and define the M-estimator as before based on the corresponding estimating equation; that estimator will also follow the general asymptotic theory developed in this paper. Further, in this case, we might generalize our requirement (14) for such ψ functions by considering integral with respect to only the conditional distribution GY |X of Y given X as follows (since φ doesn’t include any distributional part of X): Z ψ(y, x; θ)dGY |X (y) = 0. (28) In this general sense, the existing estimators of Zhou (2010) and Wang et al. (2015) become particular members of our class of general M-estimators with some specific choice of ψ function without distributional assumptions on X. In particular the choice  (29) ψ(y, x; θ) = ψ0 y − xT θ , under the set-up considered in Section 3.2 (except the distributional assumption on X) generates the estimator proposed in Zhou (2010). Then the asymptotic results of Zhou et al (2010) directly follows from our general theory of Section 4; in particular, Theorem 3.3 of Zhaou (2010) follows from our Theorem 4.4. Similarly, the proposal of Wang et al. (2015) can also be though of as a special case of our general M-estimators under the set-up of 3.4 (except the distributional assumption on X) with 20 the ψ function  ψ(y, x; θ) =  ψ0   ω(x)(y−xt β) xω(x)   σ t β) χ ω(x)(y−x σ  , (30) where χ(s) = sψ0 (s)−1, ω(x) is some suitable weights and ψ0 is some suitable function as given in Wang et al. (2015). Once again, all the asymptotic results of their paper follow from our general theory presented in Section 4. For example, Theorem 3.2 and 3.3 of Wang et al. (2015) follow from our Theorem 4.2 and 4.4. respectively under the above mentioned set-up. A numerical comparison of our MDPDE with the estimator of Wang et al. (2015) has been provided later in Section 7.2 through an interesting real data example. However, the proposed MDPDE, a special M-estimator with the ψ function given by (12), involve the assumed density of the covariates X. So it cannot be applied directly to the semiparametric settings where no distributional assumption has been made. But, we can easily extend our definition of MDPDE for the semi-parametric cases by considering the density power divergence between the conditional densities fθ (Y |X) of Y given X instead of considering the joint density of Y and X. The ψ function corresponding to this extended MDPDE under semiparametric set-up can be seen to have the form = ζeθ (X) − uθ (Y, X)fθ (Y |X)α , (31) R where ζeθ (x) = uθ (y, x)fθ (y|x)1+α dy. Clearly, this ψ function, ψα (y, x; θ), corresponding to the extended MDPDE satisfies the stronger condition (28) and hence also satisfies (14). Thus, all the properties derived in Section 4 continue to hold under suitable modification for the semiparametric set-up. With this modification, the proposed MDPDE can now be applied to any semi-parametric set-up including the linear regression set-up of Zhou et al. (2010), as considered in Section 3.2 with fully parametric assumptions. 6 Robustness: Influence Function Analysis The influence function (Hampel et al., 1986) of an estimator is a popular tool to measure its classical robustness properties. It measures the stability of the estimator under infinitesimal contamination yielding a first order approximation of the bias due to that small contamination in data. More precisely, if Tψ (G) = (Tψθ (G), Tψγ (G)) denotes the statistical functional for the M-estimator corresponding to ψ (which satisfies Equation (14)), then the influence function of this estimator is defined as IF ((y0 , x0 ); Tψ , G) = ∂ Tψ (Gǫ ) ∂ǫ = lim ǫ=0 ǫ↓0 Tψ (Gǫ ) − Tψ (G) , ǫ where Gǫ = (1 − ǫ)G + ǫ∧(x0 ,y0 ) is the contaminated distribution with ǫ being the contamination proportion and ∧(x0 ,y0 ) being the degenerate distribution at the contamination point (x0 , y0 ). If the influence function is bounded in the contamination points (x0 , y0 ), the bias under infinitesimal contamination cannot become arbitrarily large even when the contamination is very far from the data center; hence the estimator will be robust with respect to the data contamination. A straightforward albeit lengthy differentiation of the estimation equation (Equation (14) with G replaced by Gǫ and (θ, γ) replaced by Tψ (Gǫ )) yields the form of the influence function of our general M-estimators, which is presented in the following theorem. 21 Theorem 6.1 Under the above mentioned set-up, IF ((y0 , x0 ); Tψ , G) = ΛG (Tψθ (G), Tψγ (G))−1 ψ(y0 , x0 ; Tψθ (G), Tψγ (G)). (32) Clearly, whenever we choose the ψ function to be bounded with respect to y and x, the influence function of the corresponding M-estimator will be bounded in both y0 and x0 ; hence the estimator will be robust with respect to both the outlier y0 in response variables as well as the leverage point x0 in the explanatory variables. However, if we have ψ function bounded only in y and not in x (like the ψ functions of the classical M-estimators under normal linear regression without censoring) the resulting estimator will be robust only with respect to outliers in response but may not be robust with respect to leverage points. In particular, the above theorem also provides the influence function of the proposed MDPDE under fully parametric models by just using the ψ function given in (12). Note that, for most common parametric models, this particular ψ function is bounded in both y and x whenever α > 0 implying the robust nature of the MDPDE with α > 0 for both the outliers in responses and covariates. However, at α = 0, the ψ function of the corresponding MDPDE (which is the same as the MLE) is proportional to the score functions which are generally unbounded for most parametric models and prove their non-robust nature. However, the MDPDE under the semi-parametric extension has a different ψ function, given in (31), which is not bounded in x for all α ≥ 0 under common parametric families; but it is generally bounded in y for α > 0. Hence the semi-parametric MDPDE with α > 0 are only robust with respect to outliers in the response but not robust under leverage points. 7 7.1 Numerical Illustrations Simulation Study Consider the exponential regression model with randomly censored data and normal covariables as discussed in the previous subsection. For simulation exercise, we consider only one covariable so that X is a univariate normal random variable with mean γ (scalar) and variance 1. Then a covariate sample of size n is generated from N (γ, 1) distribution and given the value x of the covariate we simulate the (lifetime) response variable from an exponential distribution with mean θx under a random censoring scheme; the true values of the parameters are taken to be θ = 1 and γ = 5. Here we consider the simple exponential censoring distribution, but the censoring rate is determined to keep the expected proportion of censoring at 10 or 20% under the true distribution. Under the exponential censoring distribution with mean τ , i.e., C ∼ Exp(τ ), the expected proportion of censoring under the true distribution Exp(θx) can be seen to be P (Y > C) = θx . τ + θx So to make this proportion equal to 10% or 20%, we need to take τ = 9θx and τ = 4θx respectively (with θ = 1 for our simulation study). Then we compute the MDPDE of (θ, γ) numerically and repeat the process 1000 times to obtain the empirical estimates for the total absolute bias (sum of the absolute biases of θ and γ) and the total MSE (sum of the MSEs of θ and γ) of the MDPDE with respect to the target value (1, 5). One can measure the performance of the proposed MDPDE with respect to several choice of tuning parameter α and with the maximum likelihood estimator (MLE) at α = 0 by comparing these empirical bias and MSEs. 22 At first we consider only the pure sample without any contamination and compare the efficiencies of the MDPDEs for different α with the MLE (at α = 0). The empirical estimates of efficiency are computed from the total MSEs and are reported in Table 1 along with the total absolute bias for different α and different censoring proportions. It is clear from the table that the efficiency of the MDPDE decreases as α increases but the loss in efficiency is not so significant at smaller positive values of α. Further, for any fixed α both the total absolute bias and MSE increase as the censoring proposing increases. Table 1: Empirical Summary measures for the MDPDEs under no contamination Total Abs. Bias Total MSE Relative Efficiency Cens. Prop. 10% 20% 10% 20% 10% 20% 0.00 0.3848 0.4260 0.1363 0.1860 100% 100% 0.01 0.3695 0.4192 0.1490 0.1954 91% 95% 0.10 0.4067 0.4616 0.1577 0.2078 86% 90% α 0.30 0.4659 0.5472 0.1873 0.2656 73% 70% 0.50 0.5145 0.6122 0.2198 0.3097 62% 60% 0.70 0.5474 0.6518 0.2458 0.3424 55% 54% 1.00 0.5801 0.6890 0.2773 0.3782 49% 49% Next, to examine the robustness of the proposed MDPDEs over the MLE, we repeat the above simulation study but with 5, 10, 15 or 20% contamination in the response variable and covariates. For contamination in response variable, we generate them from an Exp(5x) distribution (θ = 5) under the same censoring scheme as before; for contamination in the covariates, we simulate observations from another normal distribution with mean 10 and variance 1. The empirical bias and MSE of the estimators are reported in Tables 2 and 3 respectively. Clearly, note that the total absolute bias as well as the total MSE increases for any fixed α as the contamination proportion increases. However, these changes are rather drastic at smaller values of α and stabilize as α increases. In other words, the MDPDE with larger α ≥ 0.3 can successfully ignore the outliers to generate robust inference. Table 2: Empirical total absolute bias of the MDPDEs for different contamination proportions Cens. Prop. 10% 20% Cont. Prop 5% 10% 15% 20% 5% 10% 15% 20% 0.00 0.513 0.741 1.093 1.594 1.006 0.759 0.865 1.090 0.01 0.506 0.667 1.011 1.476 0.953 0.706 0.790 1.000 23 0.10 0.392 0.437 0.794 1.169 0.747 0.522 0.719 1.102 α 0.30 0.270 0.258 0.554 1.009 0.651 0.412 0.559 1.166 0.50 0.220 0.223 0.518 0.878 0.723 0.393 0.481 1.092 0.70 0.199 0.216 0.534 0.816 0.786 0.405 0.449 1.035 1.00 0.186 0.217 0.578 0.781 0.856 0.415 0.436 1.003 Table 3: Empirical total MSE of the MDPDEs for different contamination proportions Cens. Prop. 10% 20% 7.2 Cont. Prop 5% 10% 15% 20% 5% 10% 15% 20% 0.00 0.265 0.453 0.908 2.282 0.794 0.441 0.689 1.208 0.01 0.257 0.425 0.848 2.167 0.743 0.415 0.680 1.171 0.10 0.139 0.220 0.497 1.585 0.425 0.248 0.499 1.153 α 0.30 0.081 0.101 0.259 0.913 0.305 0.166 0.309 1.008 0.50 0.070 0.079 0.229 0.714 0.353 0.144 0.249 0.898 0.70 0.071 0.077 0.244 0.629 0.410 0.143 0.235 0.838 1.00 0.083 0.081 0.298 0.602 0.482 0.147 0.242 0.833 Real Data Application : Heart Transplant data We will now apply our proposed MDPDEs to an interesting real data example with the semiparametric model assumptions, which will illustrate the performance of the proposed semiparametric extension described in Section 5 along with its applicability in real life scenarios. The data set considered is from the popular Stanford heart transplant program described in details in Clark et al. (1971) and contains the following survival information of 158 patients (Crowley and Hu, 1977; Escobar and Meeker Jr, 1992): ID number of patients (“ID”), survival or censoring time (“TIME”), censoring status (dead or alive), patient’s age at first transplant in years (“AGE”) and the T5 mismatch score (“T5-MS”). The dataset, available from the ‘survival’ library of R, was analyzed statistically by many authors including Brown et al. (1973), Turnbull et al. (1974), Mantel and Byar (1974), and Miller and Halpern (1982). Recently, it has also been used to illustrate the performances of robust estimates under semi-parametric AFT models by Salibian-Barrera and Yohai (2008), Locatelli et al. (2011) and Wang et al. (2015). The latest robust estimator of Wang et al. (2015), namely, the KMW-GM estimator, has been seen to work best for this dataset while using the model log(TIME) = β0 + β1 (AGE) + β2 (T5-MS) + σǫ. Here, we will apply our proposed MDPDE with different tuning parameters α with the same parametric model as above and the assumption that ǫ ∼ N (0, 1) and illustrate the superior performance of our proposal over the KMW-GM estimator of Wang et al. (2015). As noted in Wang et al. (2015), there are three potential outliers in the dataset corresponding to the ID 2, 16 and 21, where the patients have unexpectedly shorter survival times. This finding is also consistent with the results from previous analyses of the dataset and so we also treat these three data points as outliers and compute our MDPDEs twice; once with the full data set and once after removing these outliers. However, since the estimates of the parameters (β0 , β1 , β2 , σ) differ only slightly in the two cases with and without outliers, we will report teh relative variation in the estimates in order to check the extent of their robustness. Following Wang et al. (2015), we define the relative variation as RelativeVariation = 24 |θ̂full − θ̂cleaned | |θ̂full | , where θ̂full is the estimated parameter value based on the full data set and θ̂cleaned is the parameter estimate based on the cleaned data after removing the three outliers. The relative variations obtained for each of the parameters are reported in Table 4 for our proposed MDPDEs with different tuning parameters and α and also for the KMW-GM estimator of Wang et al. (2015). It can be seen clearly from the table that the MDPDEs of most of the parameters are much more stable and have less relative variation compared to the KMW-GM estimator of Wang et al. (2015) for α ≥ 0.4; this clearly shows the greater robustness of our proposal compared to the existing robust method. Further, note that the relative variation is quite high at α = 0 which is the non-robust maximum likelihood estimator. As α increases the relative variations of all the parameters decrease significantly which again shows the significant gain in robustness of our proposal with increasing α. Table 4: The relative variation of the MDPDEs at α and the KMW-GM estimates of Wang et al. (2015) with and without outliers for the Heart Transplant Data α β0 β1 β2 σ 0 0.0191 0.8198 0.3622 0.0886 0.0200 0.3450 0.3189 0.0786 0.05 0.0202 0.1477 0.2727 0.0707 0.1 0.2 0.0191 0.0041 0.1825 0.0570 0.0172 0.0347 0.1079 0.0463 0.3 0.4 0.0153 0.0445 0.0558 0.0384 0.0139 0.0452 0.0236 0.0328 0.5 0.0118 0.0409 0.0039 0.0259 0.7 0.9 0.0108 0.0369 0.0088 0.0225 0.0106 0.0354 0.0077 0.0215 1 KMW-GM 0.0153 0.0034 0.0807 0.0573 8 On the Choice of Tunning Parameter α in MDPDE A crucial issue for applying the proposed MDPDE in any real-life problem is the choice of tuning parameter α. As we have seen that the robustness and efficiency of the MDPDEs depend crucially on the tuning parameter α, it needs to be chosen carefully in practice where we have no idea regarding the contamination and censoring proportions. The simulation study presented in Section 7 gives some indication in this direction. We have seen that the MDPDEs with larger α ≥ 0.3 are robust enough to successfully address the problem of outliers; the robustness increases as α increases. On the other hand, the efficiency of the MDPDEs under pure data is seen to decreases as α increases, but there is no significant loss in efficiency at smaller positive values of α near 0.3. So, we recommend to use a value of the tuning parameter α near 0.3 to get a fair compromise between efficiency and robustness whenever the amount of contamination is not known in practice. This is in-line with the empirical suggestions given by Basu et al. (2006) in the context of MDPDE based on censored data with no covariables. However, these empirical suggestions need further justification based on more elaborative simulation and theoretical aspects. In case of complete data, some such justifications of the data driven choice of α is given by Hong and Kim (2001) and Warwick and Jones (2005). Their work might have been generalized to the case of censored data, although it is not very easy, in order to solve this issue 25 of selecting α. We hope to pursue this in our future research. 9 Conclusion The present paper proposes the minimum density power divergence estimator under the parametric set-up for censored data with covariables to generate highly efficient and robust inference. The applicability of the proposed technique is illustrated through appropriate theoretical results and simulation exercise in the context of censored regression with stochastic covariates. Further, the paper provide the asymptotic theory for a general class of estimators based on the estimating equation which opens the scope of studying many such estimators in the context of censored data in presence of some stochastic covariates. Acknowledgments: The authors gratefully acknowledge the comments of two anonymous referees which led to an improved version of the manuscript. References [1] Basu, S., Basu, A., and Jones, M. C. (2006). Robust and efficient parametric estimation for censored survival data. Annals of the Institute of Statistical Mathematics, 58(2), 341–355. [2] Basu, A., Harris, I. R., Hjort, N. L., and Jones, M. C. (1998). Robust and efficient estimation by minimising a density power divergence. Biometrika, 85(3), 549–559. [3] Bednarski, T. (1993) Robust estimation in the Cox regression model. Scand. J. Statist., 20, 213–225. [4] Bednarski, T. and Borowicz, F. (2006). coxrobust: Robust Estimation in Cox Model. R package version 1.0. [5] Begun, J. M., Hall, W. J., Huang, W. M., Wellner, J. A. (1983). Information and Asymptotic Efficiency in Parametric-Nonparametric Models. Annals of Statistics, 11, 432–452. [6] Brown, B. W., Jr., Hollander, M., and Korwar, R. M. (1973). Nonparametric Test of independence for censored data with application to Heart Transplant Studies. Florida State University Conference on Reliability and Biometry. [7] Buckley, J., and James, I., (1979). Linear regression with censored data. Biometrika, 66, 429–436. [8] Cai, Z. (1998). Asymptotic properties of Kaplan-Meier estimator for censored dependent data. Statistics and probability letters, 37(4), 381–389. [9] Campbell, G., and Földes, A. (1982). Large sample properties of nonparametric bivariate estimators with censored data. Nonparametric statistical inference, 1, 103-121. [10] Chen, Y. Y., Hollander, M., and Langberg, N. A. (1982). Small-sample results for the Kaplan-Meier estimator. Journal of the American Statistical Association, 77, 141-144. [11] Clark, D. A., Stinson, E. B., Griepp, R. B., Schroeder, J. S., Shumway, N. E., and Harrison, D. C. (1971). Cardiac Transplantation in Man. VI. Prognosis of Patients Selected for Cardiac Transplantation. Annals of Internal Medicine, 75, 15–21. 26 [12] Collett, D. (2003). Modelling Survival Data in Medical Research. Chapman Hall, London, U.K. [13] Cox, D.R. (1972). Regression models and life tables (with discussion). Journal of Royal Statistical Society, Series B. 34, 187–220. [14] Cox, D. R., and Oakes, D. (1984). Analysis of Survival Data. Chapman Hall, London, U.K. [15] Crowder, M. J., Kimber, A. C., Smith, R. L., and Sweeting, T. J. (1991). Statistical Analysis of Reliability Data. Chapman Hall, London, U.K. [16] Crowley, J. and Hu, M. (1977). Covariance analysis of heart transplant survival data. Journal of the American Statistical Association, 72, 27–36. [17] Dabrowska, D. M. (1988). Kaplan-Meier estimate on the plane. The Annals of Statistics, 16(4), 1475–1489. [18] Escobar, L. A. and Meeker Jr, W. Q. (1992). Assessing influence in regression analysis with censored data. Biometrics, 48, 507–528. [19] Farcomeni, A. and Viviani, S. (2011) Robust estimation for the Cox regression model based on trimming. Biometrical Journal, 53(6), 956–973. [20] Ghosh, A., and Basu, A. (2013). Robust estimation for independent non-homogeneous observations using density power divergence with applications to linear regression. Electronic Journal of statistics, 7, 2420-2456. [21] Ghosh, A., and Basu, A. (2014). Robust Estimation in Generalized Linear Models : The Density Power Divergence Approach. Test, doi:10.1007/s11749-015-0445-3. [22] Hampel, F. R., E. Ronchetti, P. J. Rousseeuw, and W. Stahel (1986). Robust Statistics: The Approach Based on Influence Functions. New York, USA: John Wiley & Sons. [23] Hong, C. and Kim, Y. (2001), Automatic selection of the tuning parameter in the minimum density power divergence estimation. Journal of the Korean Statistical Society, 30, 453–465. [24] Hosmer, D. W., Lemeshow, S. and May, S. (2008). Applied Survival Analysis: Regression Modeling of Time-to-Event Data. John Wiley & Sons. [25] Huber, P. J. (1981). Robust Statistics. John Wiley & Sons. [26] Kaplan, E. L., and Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American statistical association, 53 (282), 457–481. [27] Kim, M., and Lee, S. (2008). Estimation of a tail index based on minimum density power divergence. Journal of Multivariate Analysis, 99(10), 2453–2471. [28] Klein, J.P. and Moeschberger, M.L. (2003). Survival Analysis Techniques for Censored and Truncated Data, Second Edition. Springer-Verlag, New York. [29] Kosorok, M.R., Lee, B.L. and Fine, J.P. (2004). Robust inference for univariate proportional hazards frailty regression models. Annals of Statistics, 32, 1448–1491. 27 [30] Lawless, J.F. (2003). Statistical Models and Methods for Lifetime Data, Second Edition. John Wiley & Sons, Inc. New York. [31] Lee, S., and Song, J. (2009). Minimum density power divergence estimator for GARCH models. Test, 18(2), 316–341. [32] Lee, S., and Song, J. (2013). Minimum density power divergence estimator for diffusion processes. Annals of the Institute of Statistical Mathematics, 65(2), 213-236. [33] Lehmann, E. L. (1983). Theory of Point Estimation. John Wiley & Sons. [34] Lo, S. H., Mack, Y. P., and Wang, J. L. (1989). Density and hazard rate estimation for censored data via strong representation of the Kaplan-Meier estimator. Probability theory and related fields, 80(3), 461-473. [35] Locatelli, I., Marazzi, A., Yohai, V. J. (2011). Robust accelerated failure time regression. Computational Statistics and Data Analysis. 55, 874–887. [36] Mantel, N. and Byar, D. P. (1974). Evaluation of Response-Time data involving transient states: An illustration using Heart-Transplant data. Journal of the American Statistical Association, 69, 81–86. [37] Miller, R. and Halpern, J. (1982). Regression with censored data. Biometrika, 69, 521–531. [38] Peterson Jr, A. V. (1977). Expressing the Kaplan-Meier estimator as a function of empirical subsurvival functions. Journal of the American Statistical Association, 72 (360a), 854–858. [39] Ritov, Y. (1986). Estimation in a Linear Regression Model with Censored Data. The Annals of Statistics, 18(1), 303–328. [40] Robins, J. M., and Rotnitzky, A. (1992). Recovery of information and adjustment for dependent censoring using surrogate markers. In AIDS Epidemiology, 297–331. Birkh´’auser Boston. [41] Satten, G. A., and Datta, S. (2001). The KaplanMeier estimator as an inverse-probabilityof-censoring weighted average. The American Statistician, 55(3), 207–210. [42] Salibian-Barrera, M., and Yohai, V. J. (2008). High breakdown point robust regression with censored data. The Annals of Statistics, 36(1), 118–146. [43] Serfling, R. J. (1980). Approximation Theorems of Mathematical Statistics. New York, USA: John Wiley & Sons. [44] Stute, W. (1993). Consistent estimation under random censorship when covariables are present. Journal of Multivariate Analysis, 45(1), 89–103. [45] Stute, W. (1995). The central limit theorem under random censorship. The Annals of Statistics, 23(2), 422–439. [46] Stute, W. (1996). Distributional convergence under random censorship when covariables are present. Scandinavian Journal of Statistics, 23(4), 461–471. [47] Stute, W., and Wang, J. L. (1993). The strong law under random censorship. The Annals of Statistics, 21(3), 1591–1607. 28 [48] Turnbull, B. W., Brown, B. W., Jr., and Hu, M. (1974). Survivorship analysis of Heart Transplant data. Journal of the American Statistical Association, 69, 74–80. [49] Tsai, W. Y., Jewell, N. P., and Wang, M. C. (1987). A note on the product-limit estimator under right censoring and left truncation. Biometrika, 74(4), 883–886. [50] Van der Laan, M. J., and Robins, J. M. (2003). Unified methods for censored longitudinal data and causality. Springer. [51] Wang, J. L. (1999). Asymptotic Properties of M-Estimators Based on Estimating Equations and Censored Data. Scandinavian journal of statistics, 26(2), 297–318. [52] Wang, M. C., Jewell, N. P., and Tsai, W. Y. (1986). Asymptotic properties of the product limit estimate under random truncation. The Annals of Statistics, 14(4), 1597–1605. [53] Warwick, J., and Jones, M. C. (2005). Choosing a robustness tuning parameter. Journal of Statistical Computation and Simulation, 75(7), 581–588. [54] Zhou, M. (1991). Some properties of the Kaplan-Meier estimator for independent nonidentically distributed random variables. The Annals of Statistics, 19(4), 2266–2274. [55] Zhou, M. (1992). M-estimation in censored linear models. Biometrika, 79(4), 837-841. 29
10math.ST
1 Pattern Recognition Letters journal homepage: www.elsevier.com Video Semantic Object Segmentation by Self-Adaptation of DCNN Seong-Jin Park∗∗, Ki-Sang Hong arXiv:1711.08180v1 [cs.CV] 22 Nov 2017 Department of Electrical Engineering, POSTECH, Namgu Pohang, Republic Korea ABSTRACT This paper proposes a new framework for semantic segmentation of objects in videos. We address the label inconsistency problem of deep convolutional neural networks (DCNNs) by exploiting the fact that videos have multiple frames; in a few frames the object is confidently-estimated (CE) and we use the information in them to improve labels of the other frames. Given the semantic segmentation results of each frame obtained from DCNN, we sample several CE frames to adapt the DCNN model to the input video by focusing on specific instances in the video rather than general objects in various circumstances. We propose offline and online approaches under different supervision levels. In experiments our method achieved great improvement over the original model and previous state-of-the-art methods. c 2016 Elsevier Ltd. All rights reserved. 1. Introduction Semantic segmentation assigns all pixels in an image to semantic classes; it gives finely-detailed pixel-level information to visual data and can build a valuable module for higher-level applications such as image answering, event detection, and autonomous driving. Conventional semantic segmentation techniques for images have been mostly built using handcrafted features on conditional random fields (CRFs) (Russell et al., 2009; Ladickỳ et al., 2010; Yao et al., 2012). Recently deep convolutional neural networks (DCNNs) have achieved great success in classification (Krizhevsky et al., 2012; Simonyan and Zisserman, 2014; Szegedy et al., 2015; He et al., 2015), so they have been widely applied to semantic segmentation approaches (Long et al., 2015; Chen et al., 2014; Noh et al., 2015; Zheng et al., 2015; Lin et al., 2015; Liu et al., 2015; Yu and Koltun, 2015). A video consists of a sequence of images, so image-based models can be applied to each of them. For instance, there was an attempt to apply object and region detectors pre-trained on still images to a video (Zhang et al., 2015); they used pretrained object detectors to generate rough object region proposals in each frame. Similarly, we adopt an approach that employs an image-trained model to process a video, but instead of a conventional object detector, we apply a DCNN semantic segmentation model to each frame. However, the DCNN model author: Tel.: +82-54-279-2881; e-mail: [email protected] (Seong-Jin Park) ∗∗ Corresponding (a) (b) Fig. 1. Problems when the pre-trained DCNN model is applied directly to a video frame (top: a frame, bottom: the result). Different colors of results represent different classes. (a) Objects segmented into different classes. (b) Label wavers between visually-similar categories (from left to right: frame 1 - mixed, frame 31 - dog, frame 61 - horse). Best viewed in color. can show spatially inconsistent labels as previously reported in (Qi et al., 2015) when it is applied to an image. This inconsistency is exacerbated for video due to various factors such as motion blur, video compression artifacts, and sudden defocus (Kalogeiton et al., 2015). When the model is applied directly to video, the labeling result for an object can be segmented into different classes, and can waver between visually confusing categories among frames (Fig. 1). Human vision also experiences 2 Video frames DCNN Semantic Segmentation Adapt model Globally confident Output Select frames Set labels Locally confident Generate self-adapting dataset Fig. 2. Main framework of our method. Best viewed in color. such difficulty of recognition under certain circumstances. Our framework is motivated by the following question: ‘How does a human recognize a confusing object?’ When a human has difficulty identifying the object, she can guess its identity by using learned models and focusing on the object for a while. Then, she can recognize the object by referring to the moments during which it is unambiguous; i.e., she tunes her model to the specific object while regularizing small appearance changes. In a way analogous to this process, our framework takes advantage of multiple frames of a video by emphasizing confidentlyestimated (CE) frames to adapt a pre-trained model to a video. The key idea of our method is to propagate the belief of CE frames to the other frames by fine-tuning DCNN model; we apply a pre-trained DCNN model to each frame to guess the object’s class, and collect frames in which the estimation is globally confident or locally confident. Then, we use the set of CE frames as a training set to fine-tune the pre-trained model to the instances in a video (Fig. 2). We restrict the DCNN model to be video-specific rather than general purpose; i.e., we make the model focus on the specific instances in each video. In our procedures, we only use the label of CE regions, and let the uncertainly-estimated (UE) regions be determined by the CE frames. We also incorporate weak labels (i.e., manuallyannotated class labels of objects in the video) to prevent a few incorrect labels from degrading the model. Our procedures to generate a self-adapting dataset and to use CE frames to update the model can recover the uncertain or misclassified parts of UE frames that include multiple objects. We also propose an online approach that is implemented in a way similar to object tracking, because object tracking and online video object segmentation are closely related in that both tasks should trace the appearance change of an object while localizing it. Then we combine the batch and online results to improve the motion-consistency of segmentation. We validate our proposed method on the Youtube-ObjectDataset (Prest et al., 2012; Jain and Grauman, 2014; Zhang et al., 2015). In experiments our model greatly improved the pre-trained model by mitigating its drawback even when we do not use the weak labels. 2. Related work Recent image semantic segmentation techniques have been propelled by the great advance of DCNN for the image classification task (Krizhevsky et al., 2012; Simonyan and Zisserman, 2014; Szegedy et al., 2015; He et al., 2015). Based on the classification network, Long et al. (2015) extended a convolutional network to a fully-convolutional end-to-end training framework for pixel-wise dense prediction. Chen et al. (2014) used a hole algorithm to efficiently compute dense feature maps, and combined the output of the network into a fully-connected CRF. Several follow-up studies proposed more-sophisticated combinations of CRF framework with DCNNs (Zheng et al., 2015; Lin et al., 2015; Liu et al., 2015). Yu and Koltun (2015) proposed a modified architecture that aggregates multi-scale context by using dilated convolutions specifically designed for dense prediction. Due to the difficulty of pixel-wise annotation for frames, most video semantic object segmentation techniques have been built on a weakly-supervised setting that is given only videolevel class labels of objects appearing in a video. Hartmann et al. (2012) trained weakly-supervised classifiers for several spatiotemporal segments and used graphcuts to refine them. Tang et al. (2013) determined positive concept segments by using a concept-ranking algorithm to compare all segments in positive videos to the negative videos. Liu et al. (2014) proposed a label transfer scheme based on nearest neighbors. The natural limitation of the weakly-supervised approach is that it has no information about the location of target objects. Because a video is a sequence of images, Zhang et al. (2015) used object detectors that had been pre-trained on still images, and applied them to a video to localize object candidates; in each frame the method generates several object proposals by using object detectors that correspond to the given labels and by using rough segmentation proposals based on objectness. Although the object detection gives the spatial information as a bounding box around objects that have a semantic class label, the information is not sufficient for pixel-wise segmentation. Thus we use a DCNN semantic segmentation model pre-trained on images to give the pixel-wise spatial extent of the object and its semantic label at the same time, and adapt the image-based DCNN model to the input video. In contrast to most existing approaches that focus on modeling temporal consistency at pixel or region levels, our framework does not necessarily assume that the neighboring frames should be similar, and because it samples several frames that may capture different appearances of an object, our framework is relatively insensitive to sudden changes of the object. Another related topic is video object segmentation with semisupervised video (Ali et al., 2011; Ramakanth and Babu, 2014; Tsai et al., 2012; Badrinarayanan et al., 2010; Jain and Grauman, 2014; Fathi et al., 2011), which is given pixel-wise annotation of certain frames. Especially, Jain and Grauman (2014) proposed a method that employs supervoxels to overcome the myopic view of consistency in pairwise potentials by incorporating additional supervoxel-level higher-order potential. Fathi et al. (2011) developed an incremental self-training framework by iteratively labeling the least uncertain frame and updating 3 similarity metrics. The framework is similar to ours in that we update a model based on previously-estimated frames, although we neither assume pixel-wise annotation nor require superpixels, which often wrongly capture the boundary of the object. We also update the pre-trained DCNN model, in contrast to the method of Fathi et al. (2011) that updates a simple similarity metric based on hand-crafted features such as SIFT and a color histogram. 3. Method: AdaptNet We assume that a video includes at least a few CE frames and that they are helpful to improve the results of the UE frames. The main idea of our method is to propagate the belief of those CE frames by fine-tuning DCNN model. Thus our main framework consists of the following steps: selection of CE frames, label map generation, and adaptation of the model to the input video. We describe the detailed algorithm of these steps in the following subsections. We first propose a batch (offline) framework, then an online framework. We combine those two results to improve the motion-consistency of segmentation by incorporating optical flow. At the end of the section, we mention simple extensions to the unsupervised batch algorithm. 3.1. Batch Let F denote a set of frame indices, and W denote a set of given weak labels of the input video. We begin by applying a pre-trained DCNN model θ to each frame f ∈ F , then use softmax to compute the probability P(xi |θ) that the i-th pixel is a member of each class xi ∈ O, where O denotes the set of object classes and background. The semantic label map S can be computed using the argmax operation for every i-th pixel: S (i) = arg max xi P(xi |θ). To adapt the DCNN model θ to the input video, we collect a self-adapting dataset G that consists of CE frames and corresponding label maps. We collect globally-CE and locally-CE frames and compute the respective label maps Gg and Gl to construct the self-adapting dataset. The procedures to select the frames and to compute the labels are described in the following and summarized in Algorithm 1. We first perform connectedcomponent analysis on each class map of S to generate a set R of object candidate regions. For each k-th segmented label map Rk ∈ R we measure the confidence C(Rk ) of the estimated regions, where the C(·) operator takes a label map as input and computes the average probability that the pixels labeled as objects in the label map have the corresponding class labels. Then we generate the label map Ggf by setting the label of the region only when its confidence exceeds a high threshold to . We also set the background label for every pixel for which the probability P(xi = bg|θ) of being background exceeds threshold tb . To complete Ggf , the remaining uncertain regions must be processed. For this purpose, we let the remaining pixels have the “ignored” label. The uncertain “ignored” pixels are not considered during computation of the loss function for model update. We also ignore all pixels that have labels that are not in Algorithm 1 AdaptNet-Batch 1: Given: DCNN model θ, a set of weak labels W 2: Local best confidence d = 0 3: for f ∈ F do 4: Initialize Ggf , Glf to ”ignored” labels 5: Compute P(x|θ) and S = arg maxx P(x|θ) 6: Compute set R of connected components in S 7: for Rk ∈ R do 8: if S (i) < W, i ∈ Rk then continue 9: if C(Rk ) > to then Set Ggf (i) = S (i), ∀i ∈ Rk 10: Set Glf (i) = S (i), ∀i ∈ Rk 11: Set Ggf (i) = Glf (i) = 0, ∀i, s.t.P(xi = bg|θ) > tb 12: if C(Ggf ) > 0 then G ← G ∪ {Ggf } 13: if C(Glf ) > d then 14: Update t = f and d = C(Glf ) 15: if f mod τb = 0 then 16: if Ggt < G then 17: G ← G ∪ {Glt } 18: Initialize d = 0 19: Finetune DCNN model θ to θ0 using the set G the set W of given weak labels of a video. We add globallyCE frames with Ggf that has at least one confident region (i.e., C(Ggf ) > 0) to the self-adapting dataset G. Because the selected frames might be temporally unevenly distributed, the model can be dominated by frames that are selected during a short interval. To mitigate the resulting drawback and regularize the model, we also select the locally-CE frames that have best object confidences during every period τb although the frames do not include globally-CE object regions. We determine the locally-CE frame and its label map Gl as follows: we generate a label map Glf for every frame f by keeping the label of all pixels only if the label S (i) is included in W, while setting the background as before. We measure the confidence of a frame by computing C(Glf ) and we regard as locally-CE the frame that has the highest confidence during every section of τb frames. If the locally-CE frame is not already selected as a globally-CE frame, we add it to the self-adapting dataset G. Given the self-adapting dataset G constructed from the above procedures, we finally adapt the DCNN model θ to the video by fine-tuning the model to θ0 based on the dataset. Then, we compute the new label map by applying θ0 to every frame. 3.2. Online The main difference between the batch and online frameworks is the generation of the self-adapting dataset. We adopt an online framework similar to object tracking using DCNN (Nam and Han, 2015) because object tracking and online video object segmentation are closely related in that they should trace an object’s appearance change while localizing object’s region. Nam and Han (2015) pre-trained domain-independent layers from training videos and collected two sets of frames (i.e., logterm and short-term sets) to fine-tune the model to be domainspecific. Similarly, to update the model periodically we collect 4 O.F. Sb So c(m f , m f +1 ) Fig. 3. Motion-consistent combination. Top row: consecutive frames of a video and color-coded optical flows (O.F.) between frames. Middle row: corresponding results of batch (S b ) model. Bottom row: corresponding results of online model (S o ). We compute the consistency c(m f , m f +1 ) between those results by following optical flow, then find the most consistent path of selected models. Best viewed in color. Algorithm 2 AdaptNet-Online 1: Given: DCNN model θ, a set of weak labels W 2: Local best confidence d = 0 3: for f ∈ F do 4: Initialize Ggf , Glf to ”ignored” labels 5: Compute P(x|θ) and S = arg maxx P(x|θ) 6: Compute set R of connected components in S 7: for Rk ∈ R do 8: if S (i) < W, i ∈ Rk then continue 9: if C(Rk ) > to then Set Ggf (i) = S (i), ∀i ∈ Rk 10: Set Glf (i) = S (i), ∀i ∈ Rk 11: Set Ggf (i) = Glf (i) = 0, ∀i, s.t.P(xi = bg|θ) > tb 12: if C(Ggf ) > 0 then 13: if |Tl | > τl then Tl .dequeue 14: Tl .enqueue(Ggf ) 15: if C(Glf ) > d then 16: Update t = f and d = C(Glf ) 17: if f mod τ s = 0 then 18: if Ggt < Tl then 19: if |T s | > τ s then T s .dequeue 20: T s .enqueue(Glt ) 21: Initialize d = 0 22: if f mod τb = 0 then 23: Finetune DCNN model θ to θ0 using T s ∪ Tl 24: Set θ ← θ0 latter; i.e., the batch algorithm makes its decision after processing the whole video. However, videos may exist in which the online algorithm shows better results for certain local frames. Thus we can combine the two results to improve the motionconsistency of segmentation by incorporating dense optical flow. We cast the combination as the problem of selecting the best model in every frame as follows: let m f ∈ {S bf , S of }, ∀ f be a variable that selects a labeled result between batch S bf and online S of , and c(m f , m f +1 ) measure a consistency between two consecutive labeled frames. We can formulate the motionconsistent model selection problem as argmax m X c(m f , m f +1 ), f where m = {m1 , m2 , ..., m|F | } is the set of selected models. We measure the consistency c(m f , m f +1 ) by the overlap o(m f , m f +1 ) of object regions between consecutive labeled frames warped by following dense optical flow (Farnebäck, 2003) (Fig. 3). Because the optical flow can be noisy we give a small preference  for the transition from batch to batch result. That is,   o(m f , m f +1 ) +  if m f = S bf     ∧ m f +1 = S bf +1 c(m f , m f +1 ) =      o(m f , m f +1 ) otherwise. Note that this problem can be easily solved using dynamic programming. and maintain two sets: Tl and T s of frames instead of G; Tl maintains τl globally-CE frames and T s maintains τ s locallyCE frames separately. Tl is implemented as a priority queue to collect globally-CE frames; T s is a basic queue to deal with local variations. After collecting two sets of frames for a certain period τb , we use both Tl and T s as the self-adapting dataset to update the parameters of the model θ. We iterate those procedures until the end of the video. The detailed procedures are described in Algorithm 2. Note that the first 11 lines in the algorithm are the same as the batch procedures. 3.3. Motion-Consistent Combination The batch algorithm generally works better than the online algorithm, because the former uses global update with larger pool of CE frames and has longer-range dependency than the 3.4. Unsupervised Video We briefly mention our method for processing an unsupervised video. Our framework can be easily applied to unsupervised videos by bypassing line 8 in Algorithm 1. This deletion means that we do not care whether the class actually appears in the video, thus we set all the labels of CE regions even if the labels are incorrect. We found that most of the videos processed in this way show similar results to those of weakly-supervised video, because the labels of pixels determined with very high probability usually correspond to the correct labels. Nevertheless, a few exceptions that correspond to incorrect labels occur, which can degrade the model and decrease the accuracy compared with the weakly-supervised setting. 5 Table 1. Intersection-over-union overlap on Youtube-Object-Dataset 2014 (Jain and Grauman, 2014) Base-context Base-front-end SCF (Jain and Grauman, 2014) Our-Unsupv-batch Our-Weak-on Our-Weak-batch Our-Weak-comb Our-Unsupv-CRF Our-Weak-on-CRF Our-Weak-batch-CRF Our-Weak-comb-CRF Aero 0.808 0.828 0.863 0.829 0.819 0.830 0.830 0.844 0.837 0.844 0.844 Bird 0.642 0.725 0.810 0.783 0.774 0.788 0.788 0.808 0.794 0.810 0.810 Boat 0.627 0.657 0.686 0.699 0.686 0.708 0.710 0.710 0.690 0.723 0.725 Car 0.746 0.797 0.694 0.812 0.791 0.817 0.817 0.822 0.797 0.827 0.827 3.5. Post-processing Because the output of DCNN is insufficient to exactly delineate the object, we use the fully-connected CRF (Koltun, 2011). We simply use the output of DCNN for the unary term and use colors and positions of pixels for the computation of pairwise terms as Chen et al. (2014) did. Finally, we refine the label map through morphological operations (i.e., dilation and erosion). 4. Experiments 4.1. Implementation details We tested the ‘front-end’ and ‘context’ DCNN models pretrained in Yu and Koltun (2015) and observed that the front-end model shows better results than the context model under our setting on Youtube-Object-Dataset (Table 1). Thus we used the front-end model as our baseline model θ. The front-end model is a modified version of the VGG-16 network (Simonyan and Zisserman, 2014) and is extended for dense prediction. In practice, we resize each frame such that its long side is 500 pixels, then pad the frame by reflecting it about each image boundary to be 900×900 pixels. We use threshold values to = 0.75 and tb = 0.8; the value for background is set slightly higher than for foreground to leave room for additional foreground pixels (e.g., pixels around objects). We set the local period τb = 30 for both algorithms and τl = 10, τ s = 5 for our online algorithm, and the small preference  = 0.02 for the motion-consistent combination. For model update, we fine-tuned all the layers with dropout ratio of 0.5 for the last two layers and iterated for the number of frames in the self-adapting dataset with batch size of 1. We set the learning rate to 0.001, momentum to 0.9, and weight decay to 0.0005. Due to the lack of a validation dataset, we use the fixed CRF parameters used in (Noh et al., 2015) during post-processing. Our implementation is based on the Caffe library (Jia et al., 2014) equipped with an Nvidia GTX Titan X GPU. 4.2. Evaluation We evaluate the proposed method on the Youtube-ObjectDataset (Jain and Grauman, 2014; Zhang et al., 2015) that contains subset of classes in the PASCAL VOC 2012 segmentation dataset, on which the baseline model is pre-trained. The Cat 0.622 0.616 0.589 0.688 0.676 0.688 0.713 0.696 0.694 0.698 0.722 Cow 0.646 0.646 0.686 0.675 0.680 0.685 0.696 0.688 0.690 0.700 0.709 Dog 0.670 0.671 0.618 0.701 0.710 0.732 0.732 0.717 0.726 0.745 0.745 Horse 0.414 0.462 0.540 0.505 0.540 0.589 0.595 0.514 0.553 0.610 0.611 Motor 0.570 0.674 0.609 0.705 0.693 0.711 0.711 0.714 0.704 0.722 0.722 Train 0.607 0.624 0.663 0.702 0.679 0.718 0.718 0.702 0.673 0.729 0.729 Avg. 0.635 0.670 0.672 0.710 0.705 0.727 0.731 0.722 0.716 0.741 0.744 Youtube-Object-Dataset was originally constructed by Prest et al. (2012). Jain and Grauman (2014) annotated pixel-level ground truth for every 10-th frame for the first shot of each video. The dataset consists of 126 videos with 10 object classes. Due to inconsistent numbers of annotations, Zhang et al. (2015) modified the dataset by resampling 100 frames (sampled every other frame) for each video and annotating missed frames for one in every 10 frames. We use those two versions of dataset to compare with the two existing methods respectively. We use the intersection-over-union overlap to measure the accuracy. The results on the dataset used in (Jain and Grauman, 2014) are shown in Table 1. We first report the accuracies of two baseline DCNN models proposed in (Yu and Koltun, 2015) denoted by Base-front-end and Base-context in the table. Note that the Base-front-end model (Yu and Koltun, 2015) showed higher accuracy (0.670) for the dataset under our setting than did the Base-context model (0.635), which attaches several context layers to the front-end model. It is interesting that the front-end model applied to each frame of unsupervised videos showed almost the same average accuracy as the results of SCF (Jain and Grauman, 2014), which is built based on semi-supervised video. Our method with weak-supervision (Our-Weak-comb-CRF) further improved the accuracy to 0.744, which exceeds that of SCF (average 0.672). Our online algorithm under weaksupervision (Our-Weak-on) improved the accuracy by 3.5% and the batch algorithm (Our-Weak-batch) improved it by 5.7% which is better than the online algorithm due to the global update. Note that our motion-consistent combination (Our-Weakcomb) achieved small improvement on a few classes by selecting results that are more motion-consistent, which shows temporally consistent video segmentation result. The great improvement of our model over the baseline model mostly originates from the correction of confusing parts of frames by the newly-updated model based on CE frames. Some representative results of ours are shown in Figure 41 . We also report our results without any supervision (Our-Unsupv-batch) explained in Sec 3.4 to show the efficiency of our method. It achieved slightly lower accuracy (0.710) than the algorithm with weak1 More video results and the datasets we used are available at the project webpage: https://seongjinpark.github.io/AdaptNet/ 6 Table 2. Intersection-over-union overlap on Youtube-Object-Dataset 2015 (Zhang et al., 2015) (Zhang et al., 2015) Base-front-end Our-Unsupv-batch Our-Weak-on Our-Weak-batch Our-Weak-comb Our-Unsupv-CRF Our-Weak-on-CRF Our-Weak–batch-CRF Our-Weak-comb-CRF Aero 0.758 0.786 0.791 0.795 0.794 0.794 0.808 0.813 0.809 0.809 Bird 0.608 0.727 0.766 0.773 0.786 0.786 0.791 0.796 0.809 0.807 Boat 0.437 0.632 0.681 0.665 0.685 0.684 0.695 0.672 0.699 0.698 Car 0.711 0.866 0.876 0.883 0.870 0.870 0.879 0.884 0.870 0.870 supervision, but had 3.8% higher average accuracy than SCF. We report only the batch algorithm for unsupervised videos because the online algorithm is more vulnerable than the batch algorithm to incorrect labels, especially for the short videos that the dataset includes. Post-processing increased the accuracy by about 1.1 ∼ 1.4%. Our result achieved state-of-the-art overlap accuracy for all classes except ‘aeroplane’. We observed that the slightly inferior accuracy on this class occurs mainly because our method gives less accurate object boundary for the class. This problem occurs because we used the same CRF parameters for all classes as we could not cross-validate the parameters for each class, whereas SCF is given manual delineation of the object in the first frame. In Table 2, we also report our results on the dataset used in Zhang et al. (2015) to compare with the framework that uses models pre-trained on images to segment weaklysupervised video. Zhang et al. (2015) constructed the dataset (Youtube-Object-Dataset 2015) by modifying the YoutubeObject-Dataset 2014. Because neither the dataset nor source code is provided by the authors, we manually built the dataset by following the procedures explained in the paper. Our results in Table 2 show tendencies similar to those in Table 1, from the online to the combined model that greatly improves the baseline model. In this analysis the difference between the accuracies of our model and that of existing method was much larger than in Table 1 because Zhang et al. (2015) used a conventional object detector on weakly-supervised video, whereas we use a semantic segmentation model based on DCNN. 4.3. Limitation The limitation of our method occurs when a video does not meet our assumption that at least one frame has a correct label or that at least one object region that corresponds to the pretrained object classes is estimated. Such cases mostly occur due to very small size of objects in an image. The absence of a frame to improve the other frames yields the same result as the baseline model. We plan to consider these problems in our future work. 5. Conclusion We proposed a novel framework for video semantic object segmentation that adapts the pre-trained DCNN model to the Cat 0.465 0.583 0.679 0.641 0.667 0.668 0.683 0.643 0.674 0.675 Cow 0.546 0.657 0.711 0.691 0.738 0.738 0.729 0.702 0.756 0.756 Dog 0.555 0.632 0.656 0.684 0.734 0.737 0.669 0.703 0.751 0.754 Horse 0.549 0.403 0.432 0.488 0.567 0.580 0.441 0.507 0.580 0.595 Motor 0.424 0.635 0.689 0.658 0.694 0.694 0.690 0.671 0.695 0.695 Train 0.358 0.626 0.650 0.652 0.672 0.673 0.651 0.631 0.651 0.645 Avg. 0.541 0.655 0.693 0.693 0.721 0.722 0.704 0.702 0.729 0.730 input video. To fine-tune the extensively-trained model to be video-specific, we constructed a self-adapting dataset that consists of several frames that help to improve the results of the UE frames. In experiments the proposed method improved the results by using the fine-tuned model to re-estimate the misclassified parts. It also achieved state-of-the-art accuracy by a large margin. We plan to extend the framework for semi-supervised video to increase the accuracy. We also expect that the efficient self-adapting framework can be applicable to generate a huge accurately-labeled video dataset, and thus be used to progress image semantic segmentation. References Ali, K., Hasler, D., Fleuret, F., 2011. Flowboostappearance learning from sparsely annotated video, in: Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on, IEEE. pp. 1433–1440. Badrinarayanan, V., Galasso, F., Cipolla, R., 2010. Label propagation in video sequences, in: Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference on, IEEE. pp. 3265–3272. Chen, L.C., Papandreou, G., Kokkinos, I., Murphy, K., Yuille, A.L., 2014. Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062 . Farnebäck, G., 2003. Two-frame motion estimation based on polynomial expansion, in: Scandinavian conference on Image analysis, Springer. pp. 363– 370. Fathi, A., Balcan, M.F., Ren, X., Rehg, J.M., 2011. Combining self training and active learning for video segmentation, in: Eds. Jesse Hoey, Stephen McKenna and Emanuele Trucco, In Proceedings of the British Machine Vision Conference (BMVC 2011), pp. 78–1. Hartmann, G., Grundmann, M., Hoffman, J., Tsai, D., Kwatra, V., Madani, O., Vijayanarasimhan, S., Essa, I., Rehg, J., Sukthankar, R., 2012. Weakly supervised learning of object segmentations from web-scale video, in: Computer Vision–ECCV 2012. Workshops and Demonstrations, Springer. pp. 198–208. He, K., Zhang, X., Ren, S., Sun, J., 2015. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385 . Jain, S.D., Grauman, K., 2014. Supervoxel-consistent foreground propagation in video, in: Computer Vision–ECCV 2014. Springer, pp. 656–671. Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadarrama, S., Darrell, T., 2014. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093 . Kalogeiton, V., Ferrari, V., Schmid, C., 2015. Analysing domain shift factors between videos and images for object detection. arXiv preprint arXiv:1501.01186 . Koltun, V., 2011. Efficient inference in fully connected crfs with gaussian edge potentials. Adv. Neural Inf. Process. Syst . Krizhevsky, A., Sutskever, I., Hinton, G.E., 2012. Imagenet classification with deep convolutional neural networks, in: Advances in neural information processing systems, pp. 1097–1105. 7 Aeroplane Bird Boat Car Cat Cow Dog Horse Motorbike Train Fig. 4. Representative results of proposed method compared with baseline model. The results of Top: Base-front-end, Middle: Our-Weak-comb, Bottom: Our-Weak-comb-CRF. Semantic labels are overlaid on images with different colors corresponding to different class labels. We only highlight the boundary of correct class. Best viewed in color. Ladickỳ, L., Sturgess, P., Alahari, K., Russell, C., Torr, P.H., 2010. What, where and how many? combining object detectors and crfs, in: European conference on computer vision, Springer. pp. 424–437. Lin, G., Shen, C., Reid, I., et al., 2015. Efficient piecewise training of deep structured models for semantic segmentation. arXiv preprint arXiv:1504.01013 . Liu, X., Tao, D., Song, M., Ruan, Y., Chen, C., Bu, J., 2014. Weakly supervised multiclass video segmentation, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 57–64. Liu, Z., Li, X., Luo, P., Loy, C.C., Tang, X., 2015. Semantic image segmentation via deep parsing network, in: Proceedings of the IEEE International Conference on Computer Vision, pp. 1377–1385. Long, J., Shelhamer, E., Darrell, T., 2015. Fully convolutional networks for semantic segmentation, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3431–3440. Nam, H., Han, B., 2015. Learning multi-domain convolutional neural networks for visual tracking. arXiv preprint arXiv:1510.07945 . Noh, H., Hong, S., Han, B., 2015. Learning deconvolution network for semantic segmentation, in: Proceedings of the IEEE International Conference on Computer Vision, pp. 1520–1528. Prest, A., Leistner, C., Civera, J., Schmid, C., Ferrari, V., 2012. Learning object class detectors from weakly annotated video, in: Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, IEEE. pp. 3282– 3289. Qi, X., Shi, J., Liu, S., Liao, R., Jia, J., 2015. Semantic segmentation with object clique potential, in: Proceedings of the IEEE International Conference on Computer Vision, pp. 2587–2595. Ramakanth, S.A., Babu, R.V., 2014. Seamseg: Video object segmentation using patch seams, in: Computer Vision and Pattern Recognition (CVPR), 2014 IEEE Conference on, IEEE. pp. 376–383. Russell, C., Kohli, P., Torr, P.H., et al., 2009. Associative hierarchical crfs for object class image segmentation, in: 2009 IEEE 12th International Conference on Computer Vision, IEEE. pp. 739–746. Simonyan, K., Zisserman, A., 2014. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 . Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A., 2015. Going deeper with convolutions, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1–9. Tang, K., Sukthankar, R., Yagnik, J., Fei-Fei, L., 2013. Discriminative segment annotation in weakly labeled video, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2483–2490. Tsai, D., Flagg, M., Nakazawa, A., Rehg, J.M., 2012. Motion coherent tracking using multi-label mrf optimization. International journal of computer vision 100, 190–202. Yao, J., Fidler, S., Urtasun, R., 2012. Describing the scene as a whole: Joint object detection, scene classification and semantic segmentation, in: Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, IEEE. pp. 702–709. Yu, F., Koltun, V., 2015. Multi-scale context aggregation by dilated convolutions. arXiv preprint arXiv:1511.07122 . Zhang, Y., Chen, X., Li, J., Wang, C., Xia, C., 2015. Semantic object segmentation via detection in weakly labeled video, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3641–3649. Zheng, S., Jayasumana, S., Romera-Paredes, B., Vineet, V., Su, Z., Du, D., Huang, C., Torr, P.H., 2015. Conditional random fields as recurrent neural networks, in: Proceedings of the IEEE International Conference on Computer Vision, pp. 1529–1537.
1cs.CV
Sufficient Markov Decision Processes with Alternating Deep Neural Networks Longshaokan Wang1 , Eric B. Laber1 , Katie Witkiewitz2 1 Department 2 Department of Statistics, North Carolina State University, Raleigh, NC, 27695, U.S.A. of Psychology, University of New Mexico, Albuquerque, NM, 87106, U.S.A. arXiv:1704.07531v2 [stat.ME] 17 Mar 2018 Abstract Advances in mobile computing technologies have made it possible to monitor and apply datadriven interventions across complex systems in real time. Recent and high-profile examples of data-driven decision making include autonomous vehicles, intelligent power grids, and precision medicine through mobile health. Markov decision processes are the primary mathematical model for sequential decision problems with a large or indefinite time horizon; existing methods for estimation and inference rely critically on the correctness of this model. Mathematically, this choice of model incurs little loss in generality as any decision process evolving in discrete time with observable process states, decisions, and outcomes can be represented as a Markov decision process. However, in some application domains, e.g., mobile health, choosing a representation of the underlying decision process that is both Markov and low-dimensional is non-trivial; current practice is to select a representation using domain expertise. We propose an automated method for constructing a low-dimensional representation of the original decision process for which: (P1) the Markov decision process model holds; and (P2) a decision strategy that leads to maximal mean utility when applied to the low-dimensional representation also leads to maximal mean utility when applied to population of interest. Our approach uses a novel deep neural network to define a class of potential process representations and then searches within this class for the representation of lowest dimension which satisfies (P1) and (P2). We illustrate the proposed method using a suite of simulation experiments and application to data from a mobile health intervention targeting smoking and heavy episodic drinking among college students. 1 Introduction Sequential decision problems arise in a wide range of application domains including autonomous vehicles (Bagnell and Schneider, 2001), finance (Bäuerle and Rieder, 2011), logistics (Zhang and Dietterich, 1995), robotics (Kober et al., 2013), power grids (Riedmiller et al., 2000), and healthcare (Chakraborty and Moodie, 2013). Markov decision processes (MDPs) (Bellman, 1957; Puterman, 2014) are the primary mathematical model for representing sequential decision problems with an indefinite time horizon (Bertsekas and Tsitsiklis, 1996; Sutton and Barto, 1998; Bather, 2000; Si, 2004; Powell, 2007; Wiering and Van Otterlo, 2012). This class of models is quite general as almost any decision process can be made into an MDP by concatenating data over multiple decision points (see Section 2 for a precise statement); however, coercing a decision process into the MDP framework in this way can lead to high-dimensional system state information that is difficult to model effectively. One common approach to construct a low-dimensional decision process from a high-dimensional MDP is to create a finite discretization of the space of possible system states and to treat the resultant process as a finite MDP (Gordon, 1995; Murao and Kitamura, 1997; Sutton and Barto, 1998; Kamio et al., 2004; Whiteson et al., 2007). However, such discretization can result in a significant loss of information and can be difficult to apply when the system state information is continuous and high-dimensional. Another common approach to dimension reduction is to construct a low-dimensional summary of the underlying system states, e.g., by applying principal components analysis (Jolliffe, 1986), multidimensional scaling (Borg and Groenen, 1997), or by constructing a local linear embedding (Roweis and Saul, 2000). These approaches can identify a low-dimensional representation of the system state but, as we shall demonstrate, they need not retain salient features for making good decisions. The preceding methods seek to construct a low-dimensional representation of a high-dimensional MDP with the goal of using the low-dimensional representation to estimate an optimal decision strategy, i.e., one that leads to maximal mean utility when applied to the original process; however, they offer no guarantee that the resulting process is an MDP or that a decision strategy estimated using data from the low-dimensional process will perform well when applied to the original process. We derive sufficient conditions under which a low-dimensional representation is an MDP, and that 1 an optimal decision strategy for this low-dimensional representation is optimal for the original process. We develop a hypothesis test for this sufficient condition based on the Brownian distance covariance (Székely et al., 2007; Székely and Rizzo, 2009) and use this test as the basis for selecting a low-dimensional representation within a class of deep neural networks. The proposed estimator can be viewed as a novel variant of deep neural networks for feature construction in MDPs. In Section 2, we review the MDP model for sequential decision making and define an optimal decision strategy. In Section 3, we derive conditions under which a low-dimensional representation of an MDP is sufficient for estimating an optimal decision strategy for the original process. In Section 4, we develop a new deep learning algorithm that is designed to produce low-dimensional representation that satisfies the proposed sufficiency condition. In Section 5, we evaluate the performance of the proposed method in a suite of simulated experiments. In Section 6, we illustrate the proposed method using data from a study of a mobile health intervention targeting smoking and heavy episodic drinking among college students (Witkiewitz et al., 2014). A discussion of future work is given in Section 7. 2 Setup and Notation We assume that the observed data are n on S1i , A1i , Ui1 , S2i , . . . , ATi , UiT , STi +1 independent and identically distributed copies of the trajectory which comprise n i=1 1 1 1 (S , A , U , S2 , . . . , AT , U T , ST +1 ) where: T ∈ N denotes the observation time; St ∈ Rpt denotes a summary of information collected up to time t = 1, . . . , T ; At ∈ A = {1, . . . , K} denotes the decision made at time t = 1, . . . , T ; and U t = U t (St , At , St+1 ) is a real-valued deterministic function of (St , At , St+1 ) that quantifies the momentary “goodness” of being in state St , making decision At , and subsequently transitioning to state St+1 . We assume throughout that supt |U t | ≤ M with probability one for some fixed constant M . In applications like mobile health, the observed data might be collected in a pilot study with a preset time horizon T (Maahs et al., 2012; Witkiewitz et al., 2014); however, the intent is to use these data to estimate an intervention strategy that will maximize some measure of cumulative utility when applied over an indefinite time horizon (Ertefaie, 2014; Liao et al., 2015; Luckett et al., 2016). Thus, we assume that (S1 , A1 , U 1 , S2 , . . . , AT , U T , ST +1 ) comprises the first 2 T observations of the process (S1 , A1 , U 1 , S2 , . . .). Furthermore, we assume (A0) that this infinite process is Markov and homogeneous in that it satisfies  P t+1 S ∈G t+1 t t 1 A ,S ,...,A ,S 1    t+1 t+1 t t =P S ∈G A ,S , (1) for all (measurable) subsets G t+1 ⊆ dom St+1 and t ∈ N and that the probability measure in (1) e t = (St , At−1 , . . . , St−mt ), does not depend on t. For any process (S1 , A1 , S2 , . . .) one can define S   e 1 , At , S e t , . . . satisfies (A0); to see this, note that the result where mt is chosen so that process S holds trivially for mt = t − 1. Furthermore, by augmenting the state with a variable for time, i.e., defining the new state at time t to be (Set , t), one can ensure that the probability measure in (A0) does not depend on t. In practice, mt is typically chosen to be a constant, as letting the dimension of the state grow with time makes extrapolation beyond the observed time horizon, T , difficult. Thus, hereafter we assume that the domain of the state is constant over time, i.e., dom St = S ⊆ Rp for all t ∈ N. Furthermore, we assume that the utility is homogeneous in time, i.e., U t = U (St , At , St+1 ) for all t ∈ N. A decision strategy, π : S → A, is a map from states to decisions so that, under π, a decision maker presented with St = st at time t will select decision π(st ). We define an optimal decision strategy using the language of potential outcomes (Rubin, 1978). We use an overline to denote history so that at = (a1 , . . . , at ) and st = (s1 , . . . , st ). The set of potential outcomes is O∗ =  ∗t t−1 S (a ) t≥1 where S∗t (at−1 ) is the potential state under at−1 and we have defined S∗1 (a0 ) = S1 .  Thus, the potential utility at time t under at is U S∗t (at−1 ), at , S∗(t+1) (at ) . The potential state P Q under a decision strategy, π, is S∗t (π) = at−1 S∗t (at−1 ) t−1 v=1 1π{S∗v (av−1 )}=av , and the potential    utility under π is U ∗t (π) = U S∗t (π), π S∗t (π) , S∗(t+1) (π) . Define the discounted mean utility under a decision strategy, π, as V (π) = E  X  t≥1   γ t−1 U ∗t (π) ,  where γ ∈ (0, 1) is a discount factor that balances the trade-off between immediate and long-term utility. Given a class of decision strategies, Π, an optimal decision strategy, π opt ∈ Π, satisfies 3 V (π opt ) ≥ V (π) for all π ∈ Π.   t t−1 = at−1 . To characterize π opt in terms of Define µt (at ; st , at−1 ) = P At = at S = st , A the data-generating model, we make the following assumptions for all t ∈ N: (C1) consistency, St = S∗t (A t−1 t ); (C2) positivity, there exists  > 0 such that µt (at ; S , A t t−1 one for all at ∈ A; and (C3) sequential ignorability, O∗ ⊥ At S , A t−1 ) ≥  with probability . These assumptions are standard in data-driven decision making (Robins, 2004; Schulte et al., 2014). Assumptions (C2) and (C3) hold by design in a randomized trial (Liao et al., 2015; Klasnja et al., 2015) but are not verifiable in the data for observational studies. Under these assumptions, the joint distribution  T of S∗t (π) t=1 is non-parametrically identifiable under the data-generating model for any decision strategy π and time horizon T . In our application, these assumptions will enable us to construct low-dimensional features of the state that retain all relevant information for estimating π opt without having to solve the original MDP as an intermediate step. 3 Sufficient Markov Decision Processes If the states St are high-dimensional it can be difficult to construct a high-quality estimator of the optimal decision strategy; furthermore, in applications like mobile health, storage and computational resources on the mobile device are limited, making it desirable to store only as much information as is needed to inform decision making. For any map φ : S → Rq define Stφ = φ(St ). t t+1 t We say that φ induces a sufficient MDP for π opt if (A , Sφ , U ) contains all relevant information t t+1 in (A , S t , U ) about π opt . Given a policy πφ : dom Stφ → A define the potential utility under πφ as Uφ∗t (πφ ) = X t n  o Y U S∗t at−1 , at , S∗(t+1) at 1πφ {S∗v (av−1 )}=av . at v=1 φ The following definition formalizes the notion of inducing a sufficient MDP. Definition 3.1. Let Π ⊆ AS denote a class of decision strategies defined on S and Πφ ⊆ ASφ a class of decision strategies defined on Sφ = dom Stφ ⊆ Rq . We say that the pair (φ, Πφ ) induces a sufficient MDP for π opt within Π if the following conditions hold for all t ∈ N: 4 t t+1 t (SM1) the process (A , Sφ , U ) is Markov and homogeneous, i.e.,     t t+1 t+1 t t+1 t t S , A = P S ∈ G S , A P St+1 ∈ G φ φ φ φ φ φ for any (measurable) subset Gφt+1 ⊆ Rq and this probability does not depend on t; (SM2) there exists π opt ∈ arg maxπ∈Π V (π) which can be written as π opt = πφopt ◦ φ, where πφopt ∈ nP o t−1 U ∗t (π ) . γ arg maxπφ ∈Πφ E φ t≥1 φ n on T +1 T T Thus, given observed data, (Si , Ai , Ui ) i=1 and class of decision strategies, Π, if one can find a pair (φ, Πφ ) which induces a sufficient MDP for π opt within Π, then it suffices to store only the n on T +1 T T Sφ,i , Ai , Ui reduced process . Furthermore, existing reinforcement learning algorithms i=1 (e.g., Sutton and Barto, 1998; Szepesvári, 2010) can be applied to this reduced process to construct an estimator of πφopt and hence π opt = πφopt ◦ φ. If the dimension of Stφ is substantially smaller than that of St , then using the reduced process can lead to smaller estimation error as well as reduced storage and computational costs. In some applications, it may also be desirable to have φ be a sparse function of St in the sense that it only depends on a subset of the components of St . For example, in the context of mobile health, one may construct the state, St , by concatenating measurements taken at time points t, t − 1, . . . , t − m, where the look-back period, m, is chosen conservatively based on clinical judgement to ensure that the process is Markov; however, a datadriven sparse feature map might identify that a look-back period of m0  m is sufficient thereby reducing computational and memory requirements but also generating new knowledge that may be of clinical value. The remainder of this section will focus on developing verifiable conditions for checking that (φ, Πφ ) induces a sufficient MDP. These conditions are used to build a data-driven, low-dimensional, and potentially sparse sufficient MDP.  | Define Yt+1 = U t , (St+1 )| for all t ∈ N. The following result provides a conditional independence criterion that ensures a given feature map induces a sufficient MDP; this criterion can be seen as an MDP analog of nonlinear sufficient dimension reduction in regression (Cook, 2007; Li et al., 2011). A proof is provided in the Supplemental Materials. Theorem 3.2. Let (S1 , A1 , U 1 , S2 , . . .) be an MDP that satisfies (A0) and (C1)-(C3). Suppose 5 that there exists φ : S → Rq such that Yt+1 ⊥ ⊥ St Stφ , At , (2) then, (φ, Πφ,msbl ) induces a sufficient MDP for π opt within Πmsbl , where Πmsbl is the set of measurable maps from S into A and Πφ,msbl is the set of measurable maps from Rq into A. The preceding result could be used to construct an estimator for φ so that (φ, Πφ,msbl ) induces a sufficient MDP for π opt within Πmsbl as follows. Let Φ denote a potential class of vector-valued functions on S. Let pbn (φ) denote a p-value for a test of the conditional independence criterion (2) based on the mapping φ, e.g., one might construct this p-value using conditional Brownian distance correlation (Wang et al., 2015) or kernel-based tests of conditional independence (Fukumizu et al., 2007). Then, one could select φbn to be the transformation of lowest dimension among those within n o the set φ ∈ Φ : dbn (φ) ≥ τ , where τ is a fixed significance level, e.g., τ = 0.10. However, such an approach can be computationally burdensome especially if the class Φ is large. Instead, we will develop a procedure based on a series of unconditional tests that is computationally simpler and allows for a flexible class of potential transformations. Before presenting this approach, we first describe how the conditional independence criterion in the above theorem can be applied recursively to potentially produce a sufficient MDP of lower dimension. The condition Yt+1 ⊥ ⊥ St Stφ , At is overly stringent in that it requires Stφ to capture all the information about Yt+1 contained within St regardless of whether or not that information is useful for decision making. However, given a sufficient MDP (S1φ , A1 , U 1 , S2φ , . . .), one can apply the above theorem to this MDP to obtain further dimension reduction; this process can be iterated until no n  | o| further dimension reduction is possible. For any map φ : S → Rq , define Ytφ = U t , St+1 . φ The following result is proved in the Supplemental Materials. Corollary 3.3. Let (S1 , A1 , U 1 , S2 , . . .) be an MDP that satisfies (A0) and (C1)-(C3). Assume that there exists φ0 : S → Rq0 such that (φ0 , Πφ0 ,msrbl ) induces a sufficient MDP for π opt within 6 Πmsrbl . Suppose that there exists φ1 : Rq0 → Rq1 such that for all t ∈ N Yt+1 ⊥ Stφ0 Stφ1 ◦φ0 , At , φ0 ⊥ (3) then (φ1 ◦φ0 , Πφ1 ◦φ0 ,msrbl ) induces a sufficient MDP for π opt within Πmsrbl . Furthermore, for k ≥ 2, ⊥ ⊥ Stφ denoting φk ◦ φk−1 ◦ · · · ◦ φ0 as φk , if there exists φk : Rqk−1 → Rqk such that Yt+1 φ k−1 then (φk , Πφ k ,msrbl ) induces a sufficient MDP for π opt k−1 Sφt , At , k within Πmsrbl . We now state a simple condition involving the residuals of a multivariate regression that can be used to test the conditional independence required in each step of the preceding corollary. In our implemenation we use residuals from a varient of deep neural networks that is suited to sequential decision problems (see Section 4). The following result is proved in the Supplemental Materials. Lemma 3.4. Let (S1 , A1 , U 1 , S2 , . . .) be an MDP that satisfies (A0) and (C1)-(C3). Suppose that there exists φ : S → Rq such that at least one of the following conditions hold: (i) o n  ⊥ ⊥ St At , Yt+1 − E Yt+1 Stφ , At (ii)  o  n  ⊥ ⊥ Yt+1 , Stφ At , St − E St Stφ then Yt+1 ⊥ ⊥ St Stφ , At . The preceding result can be used to verify the conditional independence condition required by Theorem (3.2) and Corollary (3.3) using unconditional tests of independence within levels of At ; in our simulation experiments, we used Brownian distance covariance for continuous states (Székely et al., 2007; Székely and Rizzo, 2009) and a likelihood ratio test for discrete states, though other choices are possible (Gretton et al., 2005a,b). Application of these tests requires modification to account for dependence over time within each subject. One simple approach, the one we follow here, is to compute a separate test at each time point and then to pool the resultant p-values using a pooling procedure that allows for general dependence. For example, let Gt = g(St , At , St+1 ) ∈ Rd1 and H t = h(St ) ∈ Rd2 be known features of (St , At , St+1 ) and St . Let Pn denote the empirical 7 measure. To test Gt ⊥ ⊥ H t using the Brownian distance covariance, we compute the test statistic     b t = ||Pn exp i ς | Gt + %| Ht − Pn exp iς | Gt Pn exp i%| Gt ||2 T n ω     Z Pn exp i ς | Gt + %| Ht  − Pn exp iς | Gt  Pn exp i%| H t 2 Γ 1+d1 Γ 1+d2 2 2 d%dς, = (d +d +2)/2 d +1 d +1 ||ς|| 1 ||%|| 2 π 1 2 b n as estimated with and subsequently compute the p-value, say pbtn , using the null distribution of T permutation (see Székely et al., 2007; Székely and Rizzo, 2009, for details). For each u = 1, . . . , T , let (u) (u) pbn denote the uth order statistic of pb1n , . . . , pbTn and define the pooled p-value pbun,pooled = T pbn /u. For each u = 1, . . . , T it can be shown that pbun,pooled is valid p-value (Rüger, 1978), e.g., u = 1 corresponds to the common Bonferroni correction. In our simulation experiments, we set u = bT /20 + 1c across all settings. 3.1 Variable screening The preceding results provide a pathway for constructing sufficient MDPs. However, while the criteria given in Theorem 3.2 and Lemma 3.4 can be used to identify low-dimensional structure in the state, they cannot be used to eliminate certain simple types of noise variables. For example,  let Bt t≥1 denote a homogeneous Markov process that is independent of (S1 , A1 , U 1 , S2 . . .), and    e t = (St )| , (Bt )| | . Clearly, the e 1 , A1 , U 1 , S e 2 , . . . , where S consider the augmented process S  optimal policy for the augmented process does not depend on Bt t≥1 , yet, Yt+1 need not be e t given St . To remove variables of this type, we develop a simple conditionally independent of S screening procedure that can be applied prior to constructing nonlinear features as described in the next section. The proposed screening procedure is based on the following result which is proved in the Supplemental Materials. Theorem 3.5. Let (S1 , A1 , U 1 , S2 , . . .) be an MDP that satisfies (A0) and (C1)-(C3). Suppose that there exists φ : S → Rq such that Yt+1 ⊥ ⊥ St Stφ , At , φ 8 (4) then, (φ, Πφ,msbl ) induces a sufficient MDP for π opt within Πmsbl , where Πmsbl is the set of measurable maps from S into A and Πφ,msbl is the set of measurable maps from Rq into A. This result can be viewed as a stronger version of Theorem 3.2 in that the required conditional et) = independence condition is weaker; indeed, in the example stated above, it can be seen that φ(S St satisfies (4). However, because φ appears in both Yt+1 and Stφ , constructing nonlinear features φ using this criterion is more challenging as the residual-based conditions stated in Lemma 3.4 can no longer be applied. Nevertheless, this criterion turns out to be ideally suited to screening procedures wherein the functions φ : Rp → Rq are of the form φ(st )j = stkj for j = 1, . . . , q, where {k1 , . . . , kq } is a subset of {1, . . . , p}. n o For any subset J ⊆ {1, . . . , p}, define StJ = Sjt j∈J  and YtJ = U t , (StJ )| . Let J1 denote the smallest set of indices such that U t depends on St and St+1 only through StJ1 and St+1 J1 conditioned n o on At . For k ≥ 2, define Jk = 1 ≤ j ≤ p : Sjt ⊥ 6 ⊥ YtJk−1 At . Let K denote the smallest value for which JK−1 = JK , such a K must exist as Jk−1 ⊆ Jk for all k, and define φscreen (St ) = StJK . The following results shows that φscreen induces a sufficient MDP; furthermore, Corollary 3 shows that such screening can be applied before nonlinear feature construction without destroying sufficiency. Theorem 3.6. Let (S1 , A1 , U 1 , S2 , . . .) be an MDP that satisfies (A0) and (C1)-(C3), and let J1 , . . . , JK , φscreen be as defined above. Assume that for any two non-empty subsets, J, J 0 ⊆ {1, . . . , p}, t t 6 ⊥ Yt+1 At . Then, Yt+1 ⊥ St Stφscreen , At . if StJ ⊥ 6 ⊥ Yt+1 φscreen ⊥ J0 J 0 A then there exists j ∈ J such that Sj ⊥ The condition that joint dependence implies marginal dependence (or equivalently, marginal independence implies joint independence) ensures that screening one variable at a time will identify the entire collection of important variables; this condition could be weakened by considering sets of multiple variables at a time though at the expense of additional computational burden. Algorithm 1 gives a schematic for estimating φscreen using the Brownian distance covariance to test for dependence. The inner for-loop (lines 4-7) of the algorithm can be executed in parallel and thereby scaled to large domains. 9 Algorithm 1 Screening with Brownian Distance Covariance n on T +1 T T Input: p-value threshold τ ; max number of iterations Nmax ; data (Si , Ai , Ui ) ; set of all indices D = {1, 2, . . . , p = dim(St )}. t+1 1: Set J0 = ∅, and YJ = {U t } 0 2: for k = 1, . . . , Nmax do 3: Set Jk = Jk−1 4: for each j ∈ D \ Jk−1 do 5: Perform dCov test on Sjt and YJt+1 within levels of At k−1 6: if p-value ≤ τ then 7: Set Jk = Jk ∪ {j} 8: if Jk = Jk−1 then 9: Set K = k, stop. Output: JK 4 i=1 Alternating Deep Neural Networks For simplicity, we assume that S = Rp . We consider summary functions φ : Rp → Rq that are representable as multi-layer neural networks (Anthony and Bartlett, 2009; LeCun et al., 2015; Goodfellow et al., 2016). Multi-layer neural networks have recently become a focal point in machine learning research because of their ability to identify complex and nonlinear structure in highdimensional data (see Goodfellow et al., 2016, and references therein). Thus, such models are ideally suited for nonlinear feature construction; here, we present a novel neural network architecture for estimating sufficient MDPs. We use criteria (i) in Lemma (3.4) to construct a data-driven summary function φ, therefore we also require a model for the regression of Yt+1 on Stφ and At ; we also use a multi-layer neural network for this predictive model. Thus, the model can be visualized as two multi-layer neural networks: one that composes the feature map φ and another that models the regression of Yt+1 on Stφ and At . A schematic for this model is displayed in Figure 1. Let Φ : R → [0, 1] denote a continuous and monotone increasing function and write Φ◦ to denote the vector-valued function obtained by elementwise application of Φ, i.e., Φ◦j (v) = Φ(vj ) where v ∈ Rd . The neural network for the feature map is parameterized as follows. Let r1 , . . . , rM1 ∈ N be such that r1 = p. The first layer of the feature map network is L1 (s; Σ1 , η1 ) = Φ◦ (Σ1 s + η1 ), where Σ1 ∈ Rr2 ×r1 and η1 ∈ Rr2 . 10 Recursively, for k = 2, . . . , M1 , define Lk (s; Σk , ηk , . . . , Σ1 , η1 ) = Φ◦ {Σk Lk−1 (s; Σk−1 , ηk−1 , . . . , Σ1 , η1 ) + ηk } , where Σk ∈ Rrk ×rk−1 and ηk ∈ Rk . Let θ1 = (ΣM1 , ηM1 , . . . , Σ1 , η1 ) then the feature map under θ is φ(s; θ1 ) = LM1 (s; θ1 ) = LM1 (s; ΣM1 , ηM1 , . . . , Σ1 , η1 ). Thus, the dimension of the feature map is rM1 . The neural network for the regression of Yt+1 on Stφ and At is as follows. Let rM1 +1 , . . . , rM1 +M2 ∈ N be such that rM1 +M2 = p + 1. For each a ∈ A define LM1 +1,a (s; θ1 , ΣM1 +1,a , ηM1 +1,a ) = Φ◦ {ΣM1 +1,a φ(s; θ1 ) + ηM1 ,a }, where ΣM1 +1,a ∈ RrM1 +1 ×rM1 and ηM1 ,a ∈ RrM1 +1 . Recursively, for k = 2, . . . , M2 and each a ∈ A define LM1 +k,a (s; θ1 , ΣM1 +k,a , ηM1 +k,a , . . . , ΣM1 +1,a , ηM1 +1,a ) = Φ◦ {ΣM1 +k,a LM1 +k−1,a (s; θ1 , ΣM1 +k−1,a , ηM1 +k−1,a , . . . , ΣM1 +1,a , ηM1 +1,a ) + ηM1 +k,a } , where ΣM1 +k,a ∈ RrM1 +k ×rM1 +k−1 and ηM1 +k ∈ RrM1 +k . For each a ∈ A, define θ2,a = (ΣM1 +M2 ,a , ηM1 +M2 ,a , . . . , ΣM1 +1,a , ηM1 +1,a ), and write θ2 = {θ2,a }a∈A . The postulated model for   E Yt+1 Stφ = stφ , At = at under parameters (θ1 , θ2 ) is LM1 +M1 (s; θ1 , θ2,at ). We use penalized least squares to construct an estimator of (θ1 , θ2 ). Let Pn denote the empirical measure and define Cnλ (θ1 , θ2 ) = Pn T X v r1 uX X u r2 t t+1 2 t ||LM1 +M2 (S ; θ1 , θ2,At ) − Y || + λ Σ21,`,j , t=1 j=1 `=1   λ ,θ λ bλ and subsequently θb1,n 2,n = arg minθ1 ,θ2 Cn (θ1 , θ2 ), where λ > 0 is a tuning parameter. The qP r2 2 term `=1 Σ1,`,j is a group-lasso penalty (Yuan and Lin, 2006) on the `th column of Σ1 ; if the `th column of Σ1 shrunk to zero then Stφ does not depend on the `th component of St . Computation of   λ ,θ bλ θb1,n 2,n also requires choosing values for λ, M1 , M2 , and r2 , . . . , rM1 −1 , rM1 +1 , . . . , rM1 +M2 −1 , (recall that r1 = p, rM1 +M2 = p + 1, and rM1 is the dimension of the feature map and is therefore considered separately). Tuning each of these parameters individually can be computationally burdensome, especially when M1 + M2 is large. In our implementation, we assumed 11 r2 = r3 = · · · = rM1 −1 = rM1 +1 = · · · = rM1 +M2 −1 = K1 and M1 = M2 = K2 ; then, for each fixed value of rM1 we selected (K1 , K2 , λ) to minimize cross-validated cost. Algorithm 2 shows the process for fitting this model; the algorithm uses subsampling to improve stability of the underlying sub-gradient descent updates (this is also known as taking minibatches, see Lab, 2014; Goodfellow et al., 2016, and references therein). Figure 1: Schematic for alternating deep neural network (ADNN) model. The term ‘alternating’ refers to the estimation algorithm which cycles over the networks for each treatment a ∈ A. To select the dimension of the feature map we choose the lowest dimension for which the Brownian distance covariance test of independence between Yt+1 − LM1 +M2 (St ; θb1,n , θb2,At ,n ) and St fails to reject at a pre-specified error level τ ∈ (0, 1). Let φb1n be the estimated feature map s 7→ n o b1 = j ∈ {1, . . . r1 } : Σ b 2 6= 0 for some ` ∈ {1, . . . , r2 } to be the elements of LM1 (s; θb1 ). Define R n 1,`,j n o St that dictate Stb1 ; write Stb1 as shorthand for Sjt . One may wish to iterate the foregoing 1 Rn φn b j∈R n estimation procedure as described in Corollary 3.3. However, because the components of Stb1 are φn each a potentially nonlinear combination of the elements of Stb1 , R therefore a sparse feature map n defined on the domain of Stb1 may not be any more sparse in terms of the original features. Thus, φn when iterating the feature map construction algorithm, we recommend using the reduced process n o T +1 T T n and the input; because the sigma-algebra generated by Stb1 contains the sigmaSRb1 ,i , Ai , Ui n Rn i=1 algebra generated by Stb1 , this does not incur any loss in generality. The above procedure can be φn 12 Algorithm 2 Alternating Deep Neural Networks Input: K1 , K2 ∈ N, λ ≥ 0; feature map dimension r1 ; data o n Tuning parameters T +1 T T n ; batch size proportion ν ∈ (0, 1); gradient-descent step-size {αb }b≥1 ; error Si , A i , U i i=1 1: 2: 3: 4: 5: (1) (1) tolerance  > 0; max number of iterations Nmax ; and initial parameter values θb1,n , θb2,n .  Set Da = (i, t) : Ati = a and na = #Da for each a ∈ A and t = 1, . . . , T for b = 1, . . . , Nmax do for each a ∈ A do Draw a random batch Ba of size bνna c without replacement from Da Compute a sub-gradient of the cost on batch Ba v   r2 r1 uX n o X X u 1 2 (b) (b) t  Λ(b) Σ21,`,j  LM1 +M2 Sti ; θb1,n , θb2,a,n − Yt+1 +λ a =∇ i bνna c j=1 (i,t)∈Ba 6: Compute a sub-gradient descent update     (b+1) (b) θb1,n θb1,n  =  + αb Λ(b) a (b+1) (b) θb θb 2,a,n 2,a,n (b+1) (b) Set θb2,a0 ,n = θb2,a0 ,n for all a0 6= a n o n o (b) (b) (b+1) (b+1) 8: If maxa Cnλ θb1,n , θb2,a,n − Cnλ θb1,n , θb2,a,n ≤  stop.     (b+1) (b+1) Output: θb1,n , θb2,n = θb1,n , θb2,n 7: 13 `=1 iterated until no further dimension reduction occurs. 5 Simulation Experiments We evaluate the finite sample performance of the proposed method (pre-screening with Brownian distance covariance + iterative alternating deep neural networks, which we will simply refer to as ADNN in this section) using a series of simulation experiments. To form a basis for comparison, we consider two alternative feature construction methods: (PCA) principal components analysis, so that the estimated feature map φbPCA (s) is the projection of s onto the first k principal components   P | of T −1 Tt=1 Pn St − Pn St St − Pn St ; and (tNN) a traditional sparse neural network, which can be seen as a special case of our proposed alternating deep neural network estimator where there is only 1 action. In our implementation of PCA, we choose the number of principal components, k, corresponding to 90% of variance explained. We do not compare with sparse PCA for variable selection, because based on preliminary runs, the principal components that explain 50% of variance already use all the variables in our generative model. In our implementation of tNN, we build a separate tNN for each a ∈ A, where (λ, K1 , K2 , r1 ) are tuned using cross-validation, and take the union of selected variables and constructed features. Note that there is no other obvious way to join the constructed features from tNN but to simply concatenate them, which will lead to inefficient dimension reduction especially when |A| is large, whereas we will see that ADNN provides a much more efficient way to aggregate the useful information across actions. We evaluate the quality of a feature map, φ, in terms of the marginal mean outcome under the o n T +1 T T n using Q-learning estimated optimal regime constructed from the reduced data Si , Ai , Ui i=1 with function approximation (Bertsekas and Tsitsiklis, 1996; Murphy, 2005); we use both linear function approximation and non-linear function approximation with neural networks. A description of Q-learning as well as these approximation architectures are described in the Supplemental Materials. 14 We consider data from the following class of generative models, as illustrated in Figure 2: S1 ∼ Normal64 (0, 0.25I64 ) ; A1 , . . . , AT ∼i.i.d. Bernoulli (0.5) ;  t+1 t+1 S4i−3 , S4i−2 ∼i.i.d. Normal (1 − At )g(Sit ), 0.01(1 − At ) + 0.25At ;  t+1 t+1 S4i−1 , S4i ∼i.i.d. Normal At g(Sit ), 0.01At + 0.25(1 − At ) ; U t ∼ Normal{(1 − At )[2{g(S1t ) + g(S2t )} − {g(S3t ) + g(S4t )}] + At [2{g(S3t ) + g(S4t )} − {g(S1t ) + g(S2t )}], 0.01}; for i = 1, 2, . . . , 16. Figure 2: Relationship between St and Yt+1 in the generative model, which depends on the action. First 16 variables determine the next state. First 4 variables determine the utility. The above class of models is indexed by g : R → R which we vary across the following maps: identity g(u) = u, truncated quadratic g(u) = min{u2 , 3}, and truncated exponential g(u) = min{exp(u), 3}, where the truncation is used to keep all variables of relatively the same scale across time points. Additionally, we add 3 types of noise variables, each taking up about 1 3 of total noises added: (i) dependent noise variables Djt , which are generated the same way as above except that they don’t affect the utility; (ii) white noises Wkt , which are sampled independently from Normal(0, 0.25) at each time point; and (iii) constants Clt , which are sampled independently from Normal(0, 0.25) at t = 1 and remain constant over time. More precisely, let m be the total number 15 of noise variables, then Dj1 , Wk1 , Cl1 ∼i.i.d. Normal (0, 0.25) ;  t+1 t+1 D4i−3 , D4i−2 ∼i.i.d. Normal (1 − At )g(Dit ), 0.01(1 − At ) + 0.25At ;  t+1 t+1 D4i−1 , D4i ∼i.i.d. Normal At g(Dit ), 0.01At + 0.25(1 − At ) ; Wkt ∼i.i.d. Normal (0, 0.25) ; Clt = Cl1 ; for j = 1, 2, . . . , bm/3c; k = 1, 2, . . . , dm/3e; l = 1, 2, . . . , dm/3e. It can be seen that the first 16 variables, the first 4 variables, and {g(S1t ), g(S2t ), g(S3t ) + g(S4t )}| all induce a sufficient MDP. the foregoing class of models is designed to evaluate the ability of the proposed method to identify low-dimensional and potentially nonlinear features of the state in the presence of action-dependent transitions and various noises. For each Monte Carlo replication, we sample n = 30 i.i.d. trajectories of length T = 90 from the above generative model. The results based on 500 Monte Carlo replications are reported in Table 1 - 3. In addition to reporting the marginal mean outcome under the policy estimated using Q-learning with both function approximations, we also report: (nVar) the number of selected variables; and (nDim) the dimension of the feature map. The table shows that (i) ADNN produces substantially smaller nVar and nDim compared with PCA or tNN in all cases; (ii) ADNN is robust to the 3 types of noises; (iii) when fed into the Q-learning algorithm, ADNN leads to considerably better marginal mean outcome than PCA and the original states under non-linear models; and (iv) ADNN is able to construct features suitable for Q-learning with linear function approximation even when the utility function and transition between states are non-linear. 6 Application to BASICS-Mobile We illustrate the proposed methedology using data on the effectiveness of BASICS-Mobile, a behavioral intervention delivered via mobile device, targeting heavy drinking and smoking among college students (Witkiewitz et al., 2014). Mobile interventions are appealing because of their 24-hour availability, anonymity, portability, increased compliance, and accurate data recording (Heron and 16 Model nNoise 0 linear 50 200 Feature map Linear Q NNQ nVar nDim st (st1 , st2 , st3 , st4 )| φbADNN (st ) φbtNN (st ) φbPCA (st ) st (st1 , st2 , st3 , st4 )| φbADNN (st ) φbtNN (st ) φbPCA (st ) st (st1 , st2 , st3 , st4 )| φbADNN (st ) φbtNN (st ) φbPCA (st ) 3.36(0.012) 3.34(0.012) 3.21(0.018) 3.38(0.012) 3.34(0.012) 3.31(0.012) 3.31(0.012) 3.26(0.014) 3.32(0.012) 3.34(0.012) 2.17(0.016) 3.33(0.012) 3.29(0.013) 3.34(0.012) 3.33(0.013) 3.31(0.012) 3.31(0.013) 3.34(0.013) 3.30(0.012) 3.30(0.012) 3.27(0.012) 3.29(0.013) 3.32(0.013) 3.29(0.013) 3.28(0.013) 2.98(0.035) 3.31(0.013) 3.32(0.012) 3.27(0.013) 3.11(0.028) 64 4 4.1(0.01) 16.0(0.00) 64 114 4 5.6(0.08) 37.0(0.00) 114 264 4 10.2(0.12) 87.4(0.11) 264 64 4 3.1(0.02) 34.4(0.13) 50.0(0.00) 114 4 4.6(0.08) 86.0(0.19) 85.8(0.02) 264 4 7.5(0.09) 157.8(0.48) 166.0(0.02) Table 1: Comparison of feature map estimators under linear transition and different number of noise variables (nNoise) in terms of: marginal mean outcome using Q-learning with linear function approximation (Linear Q); Q-learning with neural network function approximation (NN Q); the number of selected variables (nVar); and the dimension of the feature map (nDim) Smyth, 2010). BASICS-Mobile enrolled 30 students and lasted for 14 days. On the afternoon and evening of each day, the student is asked to complete a list of self-report questions, and then either an informational module or a treatment module is provided. A treatment module contains 1-3 mobile phone screens of interactive content, such as comparing the student’s smoking level with the levels of their peers, or guiding the student to manage their smoking urges. A treatment module is generally more burdensome than an informational module, and may be less effective if, for example, the student’s stress level is high; furthermore, excessive treatment can cause habituation and disengagement from the intervention. An optimal intervention will assignment a treatment module if and when it is needed without diminishing engagement. In our original formulation of this decision problem as an MDP, the state comprises of 15 variables capturing baseline information, current answers to the self-report questions, a weekend indicator, age, past attempts to quit smoking, current smoking urge, and current stress level; the action is whether a treatment module gets assigned; the reward is the negative of the cigarettes 17 Model nNoise 0 quad 50 200 Feature map Linear Q NNQ nVar nDim st (st1 , st2 , st3 , st4 )| φbADNN (st ) φbtNN (st ) φbPCA (st ) st (st1 , st2 , st3 , st4 )| φbADNN (st ) φbtNN (st ) φbPCA (st ) st (st1 , st2 , st3 , st4 )| φbADNN (st ) φbtNN (st ) φbPCA (st ) 3.08(0.062) 2.54(0.056) 6.63(0.038) 6.94(0.027) 2.97(0.064) 2.96(0.054) 2.58(0.057) 6.76(0.032) 6.98(0.031) 3.09(0.061) 1.28(0.030) 2.52(0.056) 6.87(0.034) 6.76(0.044) 3.09(0.062) 2.64(0.073) 6.75(0.046) 6.97(0.034) 6.54(0.068) 2.50(0.067) 1.69(0.064) 6.76(0.042) 6.99(0.030) 6.53(0.064) 2.00(0.067) 0.88(0.031) 6.68(0.050) 6.92(0.033) 6.03(0.075) 0.96(0.033) 64 4 4.1(0.02) 15.3(0.04) 64 114 4 6.4(0.06) 36.5(0.03) 114 264 4 14.3(0.14) 84.1(0.11) 264 64 4 2.4(0.04) 37.1(0.22) 51.2(0.02) 114 4 5.3(0.09) 88.3(0.22) 87.1(0.03) 264 4 12.5(0.23) 152.4(0.36) 167.4(0.03) Table 2: Comparison of feature map estimators under quadratic transition Model nNoise 0 exp 50 200 Feature map Linear Q NNQ nVar nDim st (st1 , st2 , st3 , st4 )| φbADNN (st ) φbtNN (st ) φbPCA (st ) st (st1 , st2 , st3 , st4 )| φbADNN (st ) φbtNN (st ) φbPCA (st ) st (st1 , st2 , st3 , st4 )| φbADNN (st ) φbtNN (st ) φbPCA (st ) 8.73(0.008) 9.20(0.006) 9.30(0.018) 9.44(0.005) 9.10(0.016) 8.78(0.008) 9.19(0.006) 9.32(0.018) 9.43(0.005) 8.89(0.014) 8.71(0.008) 9.19(0.006) 9.37(0.016) 9.41(0.005) 8.66(0.014) 8.78(0.012) 9.43(0.004) 9.45(0.004) 9.29(0.009) 9.02(0.023) 8.77(0.012) 9.43(0.005) 9.43(0.005) 9.18(0.012) 8.99(0.020) 8.73(0.012) 9.44(0.004) 9.41(0.006) 9.06(0.016) 9.02(0.022) 64 4 4.3(0.13) 16.0(0.00) 64 114 4 5.4(0.03) 37.0(0.00) 114 264 4 7.8(0.09) 93.4(0.10) 264 64 4 2.4(0.13) 42.3(0.17) 14.2(0.018) 114 4 2.4(0.04) 81.5(0.28) 37.2(0.02) 264 4 3.3(0.12) 152.4(0.42) 91.4(0.02) Table 3: Comparison of feature map estimators under exponential transition smoked at the next time point; the goal is to find a strategy that minimizes cumulative cigarette rate. 18 Two students with large amounts of missing data are excluded. All other missing values are imputed with the fitted value from a local polynomial regression of the state variable on time t. We treat all the variables as ordinal, partitioning some of them (see the Supplemental Materials for a complete description). We estimate φbADNN (st ) wherein conditional independence is checked via condition (i) in Lemma 3.4. The dimension of φbADNN (st ) is set to be the smallest dimension for which φbADNN (st ) fails to reject this independence condition at level τ = 0.05; this procedure resulted in a feature of dimension six. To increase the interpretability of the constructed feature map, we constrained the dimension reduction network to have no hidden layers. Under this constraint, φbADNN (st ) is a linear transformation of st followed by application of Φ◦ which was set to be the arctangent function. A plot of the weights of the 15 original variables in the linear transformation for each component of the feature map is useful in interpreting the learned feature map; see Figure 3 for an example. weights for the 1st new variable weights for the 2nd new variable 0.5 0.0 0.0 −0.5 −0.5 −1.0 −1.0 −1.5 wkend age male past90Smk quitSmk baseCigs urge wouldSmk smkMood fixMind stress tired cigs drinks lastCigs wkend age male past90Smk quitSmk baseCigs urge wouldSmk smkMood fixMind stress tired cigs drinks lastCigs −1.5 Figure 3: Weights of the original variables in the first two components of the estimated feature map. We estimate the optimal strategy using Q-learning applied to the learned feature map. Comparing the estimated parameters for treatment and no treatment, while examining the plots of weights, we can give a sense of how the original variables impact the optimal treatment assignment. For instance, the 1st parameter in the Q-function for treatment is smaller than the one for no treatment, which suggests that the 1st new variable contributes to the decision to apply treatment by being 19 small, i.e., if it is the weekend and a student’s stress level is low then the estimated policy is more likely to provide treatment. This agrees with the intuition that a treatment module would be more effective when the student is not busy or stressed. Similarly, it can be seen that previous attempts to quit smoking is positively associated with providing treatment, with the possible explanation that individuals with prior quit attempts tend to be more severe and in need of frequent treatment. 7 Discussion Data-driven decision support systems are being deployed across a wide range of application domains including medicine, engineering, and business. MDPs provide the mathematical underpinning for most data-driven decision problems with an infinite or indefinite time horizon. While the MDP model is extremely general, choosing a parsimonious representation of a decision process that fits the MDP model is non-trivial. We introduced the notion of a feature map which induces a sufficient MDP and provided an estimator of such a feature map based on a variant of deep neural networks. There are several important ways in which this work can be extended; we mention two of the most pressing here. We considered estimation from a batch of i.i.d. replicates; however, in some applications it may be desirable to estimate a feature map online as data accumulate. In such cases, a data-driven, and hence evolving, feature map of the state will be stored complicating estimation. Furthermore, because the proposed algorithm sweeps through the observed data multiple times it is not suitable for real-time estimation. Another important extension is to states with complex data structures, e.g., images and text, such data are increasingly common in health, engineering, and security applications. Existing neural network architectures designed for such data (Krizhevsky et al., 2012; Dahl et al., 2012; Simonyan and Zisserman, 2014) could potentially be integrated into the proposed feature map construction algorithm. References Anthony, M. and Bartlett, P. L. Neural Network Learning: Theoretical Foundations. cambridge university press (2009). 20 Bagnell, J. A. and Schneider, J. G. “Autonomous helicopter control using reinforcement learning policy search methods.” In International Conference on Robotics and Automation, volume 2, 1615–1620. IEEE (2001). Bather, J. Decision theory: An Introduction to Dynamic Programming and Sequential Decisions. John Wiley & Sons (2000). Bäuerle, N. and Rieder, U. Markov Decision Processes with Applications to Finance. Springer Science & Business Media (2011). Bellman, R. “A Markovian decision process.” Technical report, DTIC Document (1957). Bengio, Y. “Learning deep architectures for AI.” Foundations and Trends in Machine Learning, 2(1):1–127 (2009). Bertsekas, D. and Tsitsiklis, J. Neuro-Dynamic Programming. Belmond, MA: Athena Scientific (1996). Bertsekas, D. P. Dynamic Programming and Optimal Control , volume 1. Athena Scientific Belmont, MA (1995). Borg, I. and Groenen, P. Modern Multidimensional Scaling: Theory and Applications. SpringerVerlag (1997). Bourlard, H. and Kamp, Y. “Auto-association by multilayer perceptrons and singular value decomposition.” Biological Cybernetics, (59):291–294 (1998). Bowling, M., Ghodsi, A., and Wilinson, D. “Action respecting embedding.” In ICML ’05: Proceedings of the 22nd International Conference on Machine Learning. New York, NY, USA: ACM (2005). Chakraborty, B. and Moodie, E. E. M. Statistical Methods for Dynamic Treatment Regimes. Springer (2013). Cook, R. D. “Fisher lecture: Dimension reduction in regression.” Statistical Science, 1–26 (2007). 21 Dahl, G. E., Yu, D., Deng, L., and Acero, A. “Context-dependent pre-trained deep neural networks for large-vocabulary speech recognition.” IEEE Transactions on Audio, Speech, and Language Processing, 20(1):30–42 (2012). Ertefaie, A. “Constructing Dynamic Treatment Regimes in Infinite-Horizon Settings.” arXiv preprint arXiv:1406.0764 (2014). Filar, J. and Vrieze, K. Competitive Markov Decision Process. Springer-Verlag (1997). Fukumizu, K., Gretton, A., Sun, X., and Schölkopf, B. “Kernel Measures of Conditional Dependence.” In NIPS , volume 20, 489–496 (2007). Goodfellow, I. J., Bengio, Y., and Courville, A. Deep Learning. MIT Press (2016). Gordon, G. “Stable function approximation in dynamic programming.” Technical Report CMUCS-95-103, Computer Science Department, Carnegie Mellon University (1995). Gretton, A., Bousquet, O., Smola, A., and Schölkopf, B. “Measuring statistical dependence with Hilbert-Schmidt norms.” In International conference on algorithmic learning theory, 63–77. Springer (2005a). Gretton, A., Herbrich, R., Smola, A., Bousquet, O., and Schölkopf, B. “Kernel methods for measuring independence.” Journal of Machine Learning Research, 6(Dec):2075–2129 (2005b). Hadsell, R., Chopra, S., and LeCun, Y. “Dimensionality reduction by learning an invariant mapping.” In Proceedings of CVPR ’06 . Washington, DC, USA: IEEE Computer Society (2006). Heron, K. E. and Smyth, J. M. “Ecological momentary interventions: Incorporating mobile technology into psychosocial and health behaviour treatments.” British Journal of Health Psychology, 15(Pt 1):1–39 (2010). Hinton, G. E. “A practical guide to training restricted Boltzmann machines.” UTML Tech Report 2010-003, University of Toronto (2010). Hinton, G. E. and Salakhutdinov, R. “Reducing the dimensionality of data with neural networks.” Science, 313(5786):504–507 (2006). 22 Jolliffe, T. Principal Component Analysis. Springer-Verlag (1986). Kaelbling, L. P., Littman, M. L., and Moore, A. W. “Reinforcement learning: A survey.” Journal of artificial intelligence research, 4:237–285 (1996). Kamio, T., Soga, S., Fujisaka, H., and Mitsubori, K. “An adaptive state space segmentation for reinforcement learning using fuzzy-ART neural network.” MWSCAS , 3 (2004). Klasnja, P., Hekler, E. B., Shiffman, S., Boruvka, A., Almirall, D., Tewari, A., and Murphy, S. A. “Microrandomized trials: An experimental design for developing just-in-time adaptive interventions.” Health Psychology, 34(S):1220 (2015). Kober, J., Bagnell, J. A., and Peters, J. “Reinforcement learning in robotics: A survey.” The International Journal of Robotics Research, 0278364913495721 (2013). Krizhevsky, A., Sutskever, I., and Hinton, G. E. “Imagenet classification with deep convolutional neural networks.” In Advances in neural information processing systems, 1097–1105 (2012). Lab, L. “Deep learning tutorial.” Technical report, University of Montreal (2014). LeCun, Y., Bengio, Y., and Hinton, G. “Deep learning.” Nature, (521):436–444 (2015). Li, B., Artemiou, A., and Li, L. “Principal support vector machines for linear and nonlinear sufficient dimension reduction.” The Annals of Statistics, 3182–3210 (2011). Liao, P., Klasnja, P., Tewari, A., and Murphy, S. A. “Sample size calculations for micro-randomized trials in mHealth.” Statistics in medicine (2015). Luckett, D. J., Laber, E. B., Kahkoska, A. R., Maahs, D. M., Mayer-Davis, E., and Kosorok, M. R. “Estimating dynamic treatment regimes in mobile health using V-learning.” arXiv preprint arXiv:1611.03531 (2016). Maahs, D. M., Mayer-Davis, E., Bishop, F. K., Wang, L., Mangan, M., and McMurray, R. G. “Outpatient assessment of determinants of glucose excursions in adolescents with type 1 diabetes: proof of concept.” Diabetes technology & therapeutics, 14(8):658–664 (2012). 23 Mahadevan, S. “Learning representations and control in Markov decision processes: New frontiers.” Foundations and Trends in Machine Learning, 1(4):403–565 (2008). Mnih, V., Kavukcuoglu, K., Silva, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller, M., Fidjeland, A. K., Ostrovski, G., Petersen, S., Beattie, C., Sadik, A., Antonoglou, I., King, H., Kumaran, D., Wierstra, D., Legg, S., and Hassabis, D. “Human-level control through deep reinforcement learning.” Nature, (518):529–533 (2015). Murao, H. and Kitamura, S. “Q-Learning with adaptive state segmentation (QLASS).” In Proceedings of CIRA (1997). Murphy, S. A. “A generalization error for Q-learning.” Journal of Machine Learning Research, 6(Jul):1073–1097 (2005). Murphy, S. A., Deng, Y., Laber, E. B., Maei, H. R., Sutton, R., and Witkiewitz, K. “A batch, offpolicy actor-critic algorithm for optimizing the average reward.” eprint (2016). ArXiv:1607.05047 [stat.ML]. Nielsen, M. A. Neural Networks and Deep Learning. Determination Press (2015). Powell, W. B. Approximate Dynamic Programming: Solving the curses of dimensionality, volume 703. John Wiley & Sons (2007). Puterman, M. L. Markov Decision Processes: Discrete Stochastic Dynamic Programming. John Wiley & Sons (2014). Riedmiller, M., Moore, A., and Schneider, J. “Reinforcement learning for cooperating and communicating reactive agents in electrical power grids.” In Workshop on Balancing Reactivity and Social Deliberation in Multi-Agent Systems, 137–149. Springer (2000). Robins, J. M. “Optimal structural nested models for optimal sequential decisions.” In Proceedings of the second seattle Symposium in Biostatistics, 189–326. Springer (2004). Roweis, S. T. and Saul, L. K. “Nonlinear dimensionality reduction by locally linear embedding.” Science, (290):2323–2326 (2000). 24 Rubin, D. B. “Bayesian inference for causal effects: The role of randomization.” The Annals of statistics, 34–58 (1978). Rüger, B. “Das maximale Signifikanzniveau des Tests “Lehne H0 ab, wenn k under n gegebenen Tests zur Ablehnung führen.” Metrika, 25:171–178 (1978). Sardag, A. and Akin, H. L. “ARKAQ-Learning: Autonomous state space segmentation and policy generation.” ISCIS (2005). Schulte, P. J., Tsiatis, A. A., Laber, E. B., and Davidian, M. “Q-and A-learning methods for estimating optimal dynamic treatment regimes.” Statistical science: a review journal of the Institute of Mathematical Statistics, 29(4):640 (2014). Si, J. Handbook of Learning and Approximate Dynamic Programming, volume 2. John Wiley & Sons (2004). Simonyan, K. and Zisserman, A. “Very deep convolutional networks for large-scale image recognition.” arXiv preprint arXiv:1409.1556 (2014). Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. “Dropout: a simple way to prevent neural networks from overfitting.” Journal of Machine Learning Research, 15:1929–1958 (2014). Sutton, R. and Barto, A. Reinforcement Learning: An Introduction. MIT Press (1998). Sýkora, O. “State-space dimensionality reduction in Markov decision processes.” In WDS ’08 Proceedings of Contributed Papers, Part I, 165–170 (2008). Székely, G. and Rizzo, M. “Brownian distance covariance.” The Annals of Applied Statistics, 3:1236–1265 (2009). Székely, G., Rizzo, M., and Bakirov, N. “Measuring and testing dependence by correlation of distances.” The Annals of Statistics, 35(6):2769–2794 (2007). Szepesvári, C. “Algorithms for reinforcement learning.” Synthesis lectures on artificial intelligence and machine learning, 4(1):1–103 (2010). 25 Vincent, P., Larochelle, H., Lajoie, I., Bengio, Y., and Manzagol, P. A. “Stacked denoising autoencoders: learning useful representations in a deep network with a local denoising criterion.” Journal of Machine Learning Research, 11:3371–3408 (2010). Wang, X., Pan, W., Hu, W., Tian, Y., and Zhang, H. “Conditional distance correlation.” Journal of the American Statistical Association, 110(512) (2015). Whiteson, S., Taylor, M. E., and Stone, P. “Adaptive tile coding for value function approximation.” AI Technical Report AI-TR-07-339, University of Texas at Austin (2007). Wiering, M. and Van Otterlo, M. “Reinforcement learning: state-of-the-art.” Adaptation, Learning, and Optimization, 12 (2012). Witkiewitz, K., Desai, S. A., Bowen, S., Leigh, B. C., Kirouac, M., and Larimer, M. E. “Development and evaluation of a mobile intervention for heavy drinking and smoking among college students.” Psychology of Addictive Behaviors, 28(3):639–650 (2014). Yuan, M. and Lin, Y. “Model selection and estimation in regression with grouped variables.” Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(1):49–67 (2006). Zhang, W. and Dietterich, T. G. “A reinforcement learning approach to job-shop scheduling.” In IJCAI , volume 95, 1114–1120. Citeseer (1995). 8 Supplemental Materials Q-learning t Let (A , S t+1 t , U ) be an MDP. The value of a state-action pair under a policy π, referred to as the Q-function, is the discounted mean utility if the current state is s, current action is a, and nP o t−1 U ∗t (π)|S1 = s, A1 = a , where γ ∈ (0, 1) the agent follows π afterwards: Qπ (s, a) = E γ t≥1 is the discount factor. An optimal policy π opt yields the largest value for every state-action pair. Denote the corresponding Q-function as Qopt (s, a) = Qπ 26 opt (s, a) = max Qπ (s, a). If Qopt (s, a) is π known for all (s, a), then an optimal policy can be defined as π opt (s) = argmax Qopt (s, a). Qopt a∈A satisfies the Bellman Optimality Equations (BOE): opt Q  t (s, a) = E U + γ max Q 0 opt a ∈A (S t+1 0 t t  , a ) | S = s, A = a . In practice, one often cannot obtain Qopt by solving the above equations, because computing the right-hand-side requires the underlying model of the MDP, which is often unknown. Besides, solving a huge linear system can be costly. Q-learning is a stochastic optimization algorithm that doesn’t require knowing the transition model or solving a linear system. The update step for the classic Q-learning, Watkin’s Q-learning (Sutton and Barto, 1998), for finite state and action spaces, is as follows: Qk+1 (st , at ) ← Qk (st , at ) + α{ut + γ max Qk (st+1 , a) − Qk (st , at )}, a where α is the learning rate. If the state space is continuous, one may approximate Q(s, a) with a parametric function F (s, a; θ) and update the parameters instead: θ k+1 ← θ k + α{ut + γ max F (st+1 , a; θ k ) − F (st , at ; θ k )} · ∇θ F (st , at ; θ k ). a Proof of Theorem 3.2 t t+1 t Proof. First we show that the process (A , Sφ , U ) induced by (φ, Πφ,msbl ) satisfies (SM1). For any t ∈ N and measurable subset G ∈ Rq , t t P (St+1 ∈ Gφt+1 |Sφ , A ) φ t t t t t t = E{P (St+1 ∈ Gφt+1 |St , Sφ , A )|Sφ , A } φ = E{P (St+1 ∈ Gφt+1 |St , Stφ , At )|Sφ , A } (by Markov property of the original process) φ t t = E{P (St+1 ∈ Gφt+1 |Stφ , At )|Sφ , A } (by (2)) φ = P (St+1 ∈ Gφt+1 |Stφ , At ) φ 27 t t t t Also note that E{P (St+1 ∈ Gφt+1 |St , Sφ , A )|Sφ , A } does not depend on t by homogeneity of the φ original process. Thus the induced process is Markov and homogeneous. Next we show that the induced process satisfies (SM2). Let Qopt (s, a) be defined as before. Then we have Qopt (s, a) = E[U t + γ max Qopt (St+1 , a0 ) | St = s, At = a] 0 a (by BOE) = E[U t + γ max Qopt (St+1 , a0 ) | St = s, Stφ = sφ , At = a] 0 a = E[U t + γ max Qopt (St+1 , a0 ) | Stφ = sφ , At = a] 0 a = Qopt φ (sφ , a), (by (2)) and opt π opt (s) = argmax Qopt (s, a) = argmax Qopt φ (sφ , a) = πφ (sφ ). a a Proof of Corollary 3.3 Proof. By assumption (φ0 , Πφ0 ,msrbl ) induces a sufficient MDP for π opt within Πmsrbl , then by t t+1 t ◦ φ0 . definition the process (A , Sφ0 , U ) is Markov and homogeneous, and there exists π opt = πφopt 0 Define φk = φk ◦ · · · ◦ φ0 . By (3) and Theorem 3.2, (φ1 , Πφ1 ,msrbl ) induces a sufficient MDP for t t πφopt within Πφ0 ,msrbl . Then the process (A , Sφ , U ) is Markov and homogeneous, and there exists 0 1 ◦ φ0 = πφopt ◦ φ1 . Therefore (φ1 , Πφ πφopt = πφopt ◦ φ1 . Thus π opt = πφopt 0 0 1 1 MDP for π opt within Πmsrbl . Proof of Lemma 3.4 Proof. We show that (i) ⇒ Yt+1 ⊥ ⊥ St |Stφ , At : {Yt+1 − E(Yt+1 |Stφ , At )} ⊥ ⊥ St |At ⇒{Yt+1 − E(Yt+1 |Stφ , At )} ⊥ ⊥ (St , Stφ )|At ⇒{Yt+1 − E(Yt+1 |Stφ , At )} ⊥ ⊥ St |Stφ , At ⇒Yt+1 ⊥ ⊥ St |Stφ , At . 28 1 ,msrbl ) induces a sufficient The 1st implication follows from the fact that Stφ is a transformation of St . The 2nd implication follows from the fact that X ⊥ ⊥ Y, Z ⇒ f (X|Y, Z) = f (X) = f (X|Z) ⇒ X ⊥ ⊥ Y |Z for random variables X, Y, and Z. The 3rd implication follows from the fact that E(Yt+1 |Stφ , At ) is constant conditional on Stφ and At . Similarly, one can show that (ii) ⇒ Yt+1 ⊥ ⊥ St |Stφ , At . Proof of Theorem 3.5 t t+1 t Proof. First we show that the process (A , Sφ , U ) induced by (φ, Πφ,msbl ) satisfies (SM1). For any t ∈ N and measurable subset G ∈ Rq , t t ∈ Gφt+1 |Sφ , A ) P (St+1 φ t t t t t t = E{P (St+1 ∈ Gφt+1 |St , Sφ , A )|Sφ , A } φ = E{P (St+1 ∈ Gφt+1 |St , Stφ , At )|Sφ , A } (by Markov property of the original process) φ t t ∈ Gφt+1 |Stφ , At )|Sφ , A } (by (4)) = E{P (St+1 φ = P (St+1 ∈ Gφt+1 |Stφ , At ) φ t t t t ∈ Gφt+1 |St , Sφ , A )|Sφ , A } does not depend on t by homogeneity of the Also note that E{P (St+1 φ original process. Thus the induced process is Markov and homogeneous. Next we show that the induced process satisfies (SM2). Define Qopt,1 (st , at ) :=E{U (St , At , St+1 ) | St = st , At = at } t t t t = E{U (Stφ , At , St+1 φ ) | S = s ,A = a } t t t t = E{U (Stφ , At , St+1 φ ) | Sφ = s φ , A = a } = Qopt,1 (stφ , at ) φ 29 (by (4)) For T ≥ 2, define Qopt,T (st , at ) := E{U (St , At , St+1 ) + γ max Qopt,T −1 (St+1 , a0 ) | St = st , At = at } 0 a −1 t+1 0 = E{U (St , At , St+1 ) + γ max Qopt,T (Sφ , a ) | St = st , At = at } φ 0 a 0 t t t t = E{U (Stφ , At , St+1 Qφopt,T −1 (St+1 φ ) + γ max φ , a ) | Sφ = s φ , A = a } 0 a (by induction) (by (4)) = Qopt,T (stφ , at ) φ From now on we use U t = U (St , At , St+1 ) = U (Stφ , At , St+1 φ ) for short. Claim: supat ,st |Qopt,T (st , at ) − Qopt (st , at )|= O(γ T ). Given that the utilities are bounded, we have supst ,at ,st+1 |ut |≤ C1 , and consequently, supst ,at |Qopt (st , at )|≤ C2 , for some constants C1 and C2 . sup Qopt,1 (st , at ) − Qopt (st , at ) at ,st = sup E{U t | St = st , At = at } − E{U t + γ max Qopt (St+1 , a0 ) | St = st , At = at } 0 a t t a ,s = sup γ E{max Qopt (St+1 , a0 ) | St = st , At = at } 0 a t t a ,s ≤ γ C2 = O(γ). For T ≥ 2, assume that supat ,st |Qopt,T −1 (st , at ) − Qopt (st , at )|≤ γ T −1 C2 , then sup max Qopt,T −1 (st , a0 ) − max Qopt (st , a0 ) ≤ γ T −1 C2 , a0 a0 st and sup Qopt,T (st , at ) − Qopt (st , at ) at ,st Qopt,T −1 (St+1 , a0 ) − U t − γ max Qopt (St+1 , a0 ) | St = st , At = at } = sup E{U t + γ max 0 0 a a t t a ,s = sup γ E{max Qopt,T −1 (St+1 , a0 ) − max Qopt (St+1 , a0 ) | St = st , At = at } 0 0 a a t t a ,s ≤ γ (γ T −1 C2 ) = O(γ T ), (by induction) 30 which proves the claim. Therefore, limT →∞ Qopt,T (s, a) = Qopt (s, a) for all s and a. Similarly, limT →∞ Qopt,T (sφ , a) = Qopt φ φ (sφ , a) for all sφ and a. And we have π opt (s) = argmax Qopt (s, a) = argmax lim Qopt,T (s, a) a a T →∞ opt = argmax lim Qopt,T (sφ , a) = argmax Qopt φ φ (sφ , a) = πφ (sφ ). a T →∞ a Proof of Theorem 3.6 Proof. Let D be the set of all indices. Under the assumption that joint dependence implies marginal ⊥ ⊥ St |StJK , At . Because JK−1 = JK , ⊥ ⊥ StD\JK |At . Thus YJt+1 dependence, by construction YJt+1 K−1 K−1 the result follows. How the variables from BASICS-Mobile are partitioned The variable that records the the number of cigarettes smoked in between reports, CIGS, ranges from 0 to 20. The values imputed by local polynomial regression have decimals and are rounded to the nearest integers. CIGS has 18 unique values after rounding, which is still the most among all variables. We divide CIGS by 20 to rescale it to [0, 1], and the rescaled unique values of CIGS will be used as the levels for all variables. All other variables are rescaled to [0, 1] and rounded to the nearest level. 31
10math.ST
A TRACTABLE FAULT DETECTION AND ISOLATION APPROACH FOR NONLINEAR SYSTEMS WITH PROBABILISTIC PERFORMANCE arXiv:1408.1767v2 [math.OC] 22 Jan 2016 PEYMAN MOHAJERIN ESFAHANI AND JOHN LYGEROS Abstract. This article presents a novel perspective along with a scalable methodology to design a fault detection and isolation (FDI) filter for high dimensional nonlinear systems. Previous approaches on FDI problems are either confined to linear systems or they are only applicable to low dimensional dynamics with specific structures. In contrast, shifting attention from the system dynamics to the disturbance inputs, we propose a relaxed design perspective to train a linear residual generator given some statistical information about the disturbance patterns. That is, we propose an optimization-based approach to robustify the filter with respect to finitely many signatures of the nonlinearity. We then invoke recent results in randomized optimization to provide theoretical guarantees for the performance of the proposed filer. Finally, motivated by a cyber-physical attack emanating from the vulnerabilities introduced by the interaction between IT infrastructure and power system, we deploy the developed theoretical results to detect such an intrusion before the functionality of the power system is disrupted. 1. Introduction The task of FDI in control systems involves generating a diagnostic signal sensitive to the occurrence of specific faults. This task is typically accomplished by designing a filter with all available information as inputs (e.g., control signals and given measurements) and a scalar output that implements a non-zero mapping from the fault to the diagnostic signal, which is known as the residual, while decoupling unknown disturbances. The concept of residual plays a central role for the FDI problem which has been extensively studied in the last two decades. In the context of linear systems, Beard and Jones [Bea71, Jon73] pioneered an observer-based approach whose intrinsic limitation was later improved by Massoumnia et al. [MVW89]. Following the same principles but from a game theoretic perspective, Speyer and coauthors thoroughly investigated the approach in the presence of noisy measurements [CS98, DS99]. Nyberg and Frisk extended the class of systems to linear differential-algebraic equation (DAE) apparently subsuming all the previous linear classes [NF06], which recently also studied in the context of stochastic linear systems [EFK13]. This extension greatly enhanced the applicability of FDI methods since the DAE models appear in a wide range of applications, including electrical systems, robotic manipulators, and mechanical systems. For nonlinear systems, a natural approach is to linearize the model at an operating point, treat the nonlinear higher order terms as disturbances, and decouple their contributions from the residual by employing robust techniques [SF91, HP96]. This strategy only works well if either the system remains close to the chosen operating point, or the exact decoupling is possible. The Date: January 25, 2016. The authors are with the Automatic Control Laboratory, ETH Zürich, 8092 Zürich, Switzerland. Emails: {mohajerin,lygeros}@control.ee.ethz.ch. 1 A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance former approach is often limited, since in the presence of unknown inputs the system may have a wide dynamic operating range, which in case linearization leads to a large mismatch between linear model and nonlinear behavior. The latter approach was explored in detail by De Persis and Isidori, who in [PI01] proposed a differential geometric approach to extend the unobservibility subspaces of [Mas86, Section IV], and by Chen and Patton, who in [CP82, Section 9.2] dealt with a particular class of bilinear systems. These methods are, however, practically limited by the need to verify the required conditions on the system dynamics and transfer them into a standard form, which essentially involve solving partial differential equations, restricting the application of the method to relatively low dimensional systems. Motivated by this shortcoming, in this article we develop a novel approach to FDI which strikes a balance between analytical and computational tractability, and is applicable to high dimensional nonlinear dynamics. For this purpose, we propose a design perspective that basically shifts the emphasis from the system dynamics to the family of disturbances that the system may encounter. We assume that some statistical information of the disturbance patterns is available. Following [NF06] we restrict the FDI filters to a class of linear operators that fully decouple the contribution of the linear part of the dynamics. Thanks to the linearity of the resulting filter, we then trace the contribution of the nonlinear term to the residual, and propose an optimizationbased methodology to robustify the filter to the nonlinearity signatures of the dynamics by exploiting the statistical properties of the disturbance signals. The optimization formulation is effectively convex and hence tractable for high dimensional dynamics. Some preliminary results in this direction were reported in [MVAL12], while an application of our approach in the presence of measurement noise was successfully tested for wind turbines in [SMEKL13]. The performance of the proposed methodology is illustrated in an application to an emerging problem of cyber security in power networks. In modern power systems, the cyber-physical interaction of IT infrastructure (SCADA systems) with physical power systems renders the system vulnerable not only to operational errors but also to malicious external intrusions. As an example of this type of cyber-physical interaction we consider here the Automatic Generation Control (AGC) system, which is one of the few control loops in power networks that are closed over the SCADA system without human operator intervention. In earlier work [MVM+ 10, MVM+ 11] we have shown that, having gained access to the AGC signal, an attacker can provoke frequency deviations and power oscillations by applying sophisticated attack signals. The resulting disruption can be serious enough to trigger generator out-of-step protection relays, leading to load shedding and generator tripping. Our earlier work, however, also indicated that an early detection of the intrusion may allow one to disconnect the AGC and limit the damage by relying solely on the so-called primary frequency controllers. In this work we show how to mitigate this cyber-physical security concern by using the proposed FDI scheme to develop a protection layer which quickly detects the abnormal signals generated by the attacker. This approach to enhancing the cyber-security of power transmission systems led to an EU patent sponsored by ETH Zurich [MEVAL]. The article is organized as follows. In Section 2 a formal description of the FDI problem as well as the outline of the proposed methodology is presented. A general class of nonlinear models is described in Section 3. Then, reviewing residual generation for the linear models, we develop an optimization-based framework for nonlinear systems in Section 4. Theoretical guarantees are also provided in the context of randomized algorithms. We apply the developed methodology to the AGC case study in Section 5, and finally conclude with some remarks and directions for 2 P. Mohajerin Esfahani and J. Lygeros ݀ ݂ ‫ݕ‬ System Dynamics FDI Filter ‫ݎ‬ ‫ݑ‬ Figure 1. General configuration of the FDI filter future work in Section 6. For better readability, the technical proofs of Sections 4.2 and 4.3 are moved to the appendices. Notation. The symbols N and R+ denote the set of natural and nonnegative real numbers, respectively. Let A ∈ Rn×m be an n × m matrix with real values, A| ∈ Rm×n be its transpose, and kAk2 := σ(A) where σ is the maximum singular value of the matrix. Given a vector v := [v1 , · · · , vn ]| , the infinite norm is defined as kvk∞ := maxi≤n |vi |. Let G be a linear matrix  transfer function. Then kGkH∞ := supω∈R σ G(jω) , where σ is the maximum singular value of the matrix G(jω). The function space W n denotes the set of piece-wise continuous (p.w.c) functions taking values in Rn , and WTn is the restriction of W n to the time interval [0, T ], which is RT endowed with the L2 -inner product, i.e., he1 , e2 i := 0 e|1 (t)e2 (t)dt with the associated L2 -norm p kekL2 := he, ei. The linear operator p : W n → W n is the distributional derivative operator. d In particular, if e : R+ → Rn is a smooth mapping then p[e(t)] := dt e(t). Given a probability Nn n := space (Ω, F, P), we denote the n-Cartesian product space by Ω i=1 Ω and the respective product measure by Pn . 2. Problem Statement and Outline of the Proposed Approach In this section, we provide the formal description of the FDI problem as well as our new design perspective. We will also outline our methodology to tackle the proposed perspective. 2.1. Formal Description. The objective of the FDI design is to use all information to generate a diagnostic signal to alert the operators to the occurrence of a specific fault. Consider a general dynamical system as in Figure 1 with its inputs categorized into (i) unknown inputs d, (ii) fault signal f , and (iii) known inputs u. The unknown input d represents unknown disturbances that the dynamical system encounters during normal operation. The known input u contains all known signals injected to the system which together with the measurements y are available for FDI tasks. Finally, the input f is a fault (or an intrusion) which cannot be directly measured and represents the signal to be detected. The FDI task is to design a filter whose input are the known signals (u and y) and whose output (known as the residual and denoted by r) differentiates whether the measurements are a consequence of some normal disturbance input d, or due to the fault signal f . Formally speaking, the residual can be viewed as a function r(d, f ), and the FDI design is ideally translated as the mapping requirements (1a) d 7→ r(d, 0) ≡ 0, (1b) f 7→ r(d, f ) 6= 0, 3 ∀d A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance where condition (1a) ensures that the residual of the filter, r, is not excited when the system is perturbed by normal disturbances d, while condition (1b) guarantees the filter sensitivity to the fault f in the presence of any disturbance d. The state of the art in FDI concentrates on the system dynamics, and imposes restrictions to provide theoretical guarantees for the required mapping conditions (1). For example, the authors in [NF06] restrict the system to linear dynamics, whereas [HKEY99, PI01] treat nonlinear systems but impose necessary conditions in terms of a certain distribution connected to their dynamics. In an attempt to relax the perfect decoupling condition, one may consider the worst case scenario of the mapping (1) in a robust formulation as   min γ   γ,F RP : (2) s.t. r(d, 0) ≤ γ, ∀d ∈ D    f 7→ r(d, f ) 6= 0, ∀d ∈ D, where D is set of normal disturbances, γ is the alarm threshold of the designed filter, and the minimization is running over a given class of FDI filters denoted by F. Note that the residual r is influenced by the choice of the filter in F, but we omit this dependence for notational simplicity. In view of formulation (2), an alarm is only raised whenever the residual exceeds γ, i.e., the filter avoids any false alarm. This, however, comes at the cost of missed detections of the faults whose residual is not bigger than the threshold γ. In the literature, the robust perspective RP has also been studied in order for a trade-off between disturbance rejection and fault sensitivity for a certain class of dynamics, e.g., see [CP82, Section 9.2] for bilinear dynamics and [FF12] for multivariate polynomial systems. 2.2. New Design Perspective. Here we shift our attention from the system dynamics to the class of unknown inputs D. We assume that the disturbance signal d comes from a prescribed probability space and relax the robust formulation RP by introducing probabilistic constraints instead. In this view, the performance of the FDI filter is characterized in a probabilistic fashion. Assume that the signal d is modeled as a random variable on the prescribed probability space (Ω, F, P), which takes values in a metric space endowed with the corresponding Borel sigmaalgebra. Assume further that the class of FDI filters ensures the measurability of the mapping d 7→ r where r also belongs to a metric space. In light of this probabilistic framework, one may quantify the filter performance from different perspectives; in the following we propose two of them:     γ γ  min  min   γ,F γ,F    CP : (3) AP : s.t. E J kr(d, 0)k ≤ γ s.t. P kr(d, 0)k ≤ γ ≥ 1 − ε       f 7→ r(d, f ) 6= 0, ∀d ∈ D, f 7→ r(d, f ) 6= 0, ∀d ∈ D, where E[ · ] in AP is meant with respect to the probability measure P, and k · k is the corresponding norm in the r space. The function J : R+ → R+ in AP and ε ∈ (0, 1) in CP are design parameters. To control the filter residual generated by d, the payoff function J is required to be in class K∞ , i.e., J is strictly increasing and J(0) = 0 [Kha92, Definition 4.2, p. 144]. The decision variables in the above optimization programs are F, a class of FDI filters which is chosen a priori, and γ which is the filter threshold; we shall explain these design parameters more explicitly in subsequent sections. 4 P. Mohajerin Esfahani and J. Lygeros Two formulations provide different probabilistic interpretations of fault detection. The program AP stands for “Average Performance” and takes all possible disturbances into account, but in accordance with their occurrence probability in an averaging sense. The program CP stands for “Chance Performance” and ignores an ε-fraction of the disturbance patterns and only aims to optimize the performance over the rest of the disturbance space. Note that in the CP perspective, the parameter ε is an additional design parameter to be chosen a priori. Let us highlight that the proposed perspectives rely on the probability distribution P, which requires prior information about possible disturbance patterns. That is, unlike the existing literature, the proposed design prioritizes between disturbance patterns in terms of their occurrence likelihood. From a practical point of view this requirement may be natural; in Section 5 we will describe an application of this nature. 2.3. Outline of the Proposed Methodology. We employ randomized algorithms to tackle the formulations in (3). We generate n independent and identically distributed (i.i.d.) scenarios (di )ni=1 from the probability space (Ω, F, P), and consider the following optimization problems as random counterparts of those in (3):   γ   γ  min  min γ,F   γ,F  Pn 1 f : f : s.t. max kr(di , 0)k ≤ γ (4) AP CP s.t. i=1 J kr(di , 0)k ≤ γ n i≤n       f 7→ r(d, f ) 6= 0, ∀d ∈ D f 7→ r(d, f ) 6= 0, ∀d ∈ D, f and CP f are naturally stochastic as they depend on Notice that the optimization problems AP n the generated scenarios (di )i=1 , which is indeed a random variable defined on n-fold product probability space (Ωn , F n , Pn ). Therefore, their solutions are also random variables. In this work, we first restrict the FDI filters to a class of linear operators in which the random programs (4) are effectively convex, and hence tractable. In this step, the FDI filter is essentially robustified to n signatures of the dynamic nonlinearity. Subsequently, invoking existing results on randomized optimization, in particular [Han12, MSL15], we will provide probabilistic guarantees on the relation of programs (3) and their probabilistic counterparts in (4), whose precision is characterized in terms of the number of scenarios n. We should highlight that the true empirical approximation of the chance constraint in CP is Pn indeed n1 i=1 1 ≥ 1 − ε, where 1 is the indicator function. This approximation, ||r(di ,0)||≤γ as opposed to the one proposed in (4), leads to a non-convex optimization program which is, in general, computationally intractable. In addition, note that the design parameter ε of CP in (3) f in (4). However, as we will clarify in does not explicitly appear in the random counterpart CP 4.3, the parameter ε contributes to the probabilistic guarantees of the design. 3. Model Description and Basic Definitions In this section we introduce a class of nonlinear models along with some basic definitions, which will be considered as the system dynamics in Figure 1 throughout the article. Consider the nonlinear differential-algebraic equation (DAE) model (5) E(x) + H(p)x + L(p)z + F (p)f = 0, where the signals x, z, f are assumed to be piece-wise continuous (p.w.c.) functions from R+ into Rnx , Rnz , Rnf , respectively; we denote the spaces of such signals by W nx , W nz , W nf , respectively. 5 A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance Let nr be the number of rows in (5), and E : Rnx → Rnr be a Lipschitz continuous mapping. The operator p is the distributional derivative operator [Ada75, Section I], and H, L, F are polynomial matrices in the operator p with nr rows and nx , nz , nf columns, respectively. In the setup of Figure 1, the signal x represents all unknowns signals, e.g., internal states of the system dynamics and unknown disturbances d. The signal z contains all known signals, i.e., it is an augmented signal including control input u and available measurements y. The signal f stands for faults or intrusion which is the target of detection. We refer to [Shc07] and the references therein for general theory of nonlinear DAE systems and the regularity of their solutions. One may extend the space of functions x, z, f to Sobolev spaces, but this is outside the scope of our study. On the other hand, if these spaces are restricted to the (resp. right) smooth functions, then the operator p can be understood as the classical (resp. right) differentiation operator. Throughout this article we will focus on continuous-time models, but one can obtain similar results for discrete-time models by changing the operator p to the time-shift operator. We will think of the matrices H(p), L(p) and F (p) above either as linear operators on the function spaces (in which case p will be interpreted as a generalized derivative operator as explained above) or as algebraic objects (in which case p will be interpreted as simply a complex variable). The reader is asked to excuse this slight abuse of the notation, but the interpretation should be clear from the context. Let us first show the generality of the DAE framework of (5) by the following example. Consider the classical nonlinear ordinary differential equation (  GẊ(t) = EX X(t), d(t) + AX(t) + Bu u(t) + Bd d(t) + Bf f (t) (6)  Y (t) = EY X(t), d(t) + CX(t) + Du u(t) + Dd d(t) + Df f (t) where u( · ) is the input signal, d( · ) the unknown disturbance, Y ( · ) the measured output, X( · ) the internal variables, and f ( · ) a faults (or an attack) signal to be detected. Parameters G, A, Bu , Bd , Bf , Du , Dd , and Df are constant matrices and functions EX , EY are Lipschitz continuous mappings with appropriate dimensions. One can easily fit the model (6) into the DAE framework of (5) by defining     X Y x := , z := , d u         EX (x) −pG + A Bd 0 Bu Bf E(x) := , H(p) := , L(p) := , F (p) := . EY (x) C Dd −I Du Df Following [NF06], with a slight extension to a nonlinear dynamics, let us formally characterize all possible observations of the model (5) in the absence of the fault signal f :  (7) M := z ∈ W nz ∃x ∈ W nx : E(x) + H(p)x + L(p)z = 0 ; This set is known as the behavior of the system [PW98]. Definition 3.1 (Residual Generator). A proper linear time invariant filter r := R(p)z is a residual generator for (5) if for all z ∈ M, it holds that lim r(t) = 0. t→∞ Note that by Definition 3.1 the class of residual generators in this study is restricted to a class of linear transfer functions where R(p) is a matrix of proper rational functions of p. 6 P. Mohajerin Esfahani and J. Lygeros Definition 3.2 (Fault Sensitivity). The residual generator introduced in Definition 3.1 is sensitive to fault fi if the transfer function from fi to r is nonzero, where fi is the ith elements of the signal f . One can inspect that Definition 3.1 and Definition 3.2 essentially encode the basic mapping requirements (1a) and (1b), respectively. 4. Fault Detection and Isolation Filters The main objective of this section is to establish a scalable framework geared towards the design perspectives AP and CP as explained in Section 2. To this end, we first review a polynomial characterization of the residual generators and its linear program formulation counterpart for linear systems (i.e., the case where E(x) ≡ 0). We then extend the approach to the nonlinear model (5) to account for the contribution of E( · ) to the residual, and subsequently provide probabilistic performance guarantees for the resulting filter. 4.1. Residual Generators for Linear Systems. In this subsection we assume E(x) ≡ 0, i.e., we restrict our attention to the class of linear DAEs. One can observe that the behavior set M can alternatively be defined as  M = z ∈ W nz NH (p)L(p)z = 0 , where the collection of the rows of NH (p) forms an irreducible polynomial basis for the left null-space of the matrix H(p) [PW98, Section 2.5.2]. This representation allows one to describe the residual generators in terms of polynomial matrix equations. That is, by picking a linear combination of the rows of NH (p) and considering an arbitrary polynomial a(p) of sufficiently high order with roots with negative real parts, we arrive at a residual generator in the sense of Definition 3.1 with transfer operator (8) R(p) = a−1 (p)γ(p)NH (p)L(p) := a−1 (p)N (p)L(p), where γ(p) is a polynomial row vector representing a linear combination of the rows of NH (p). Note that the role of γ(p) is implicitly taken into consideration by N (p) := γ(p)NH (p). The above filter can easily be realized by an explicit state-space description with input z and output r. Multiplying the left hand-side of (5) by a−1 (p)N (p) leads to r = −a−1 (p)N (p)F (p)f. Thus, a sensitive residual generator, in the sense of Definition 3.1 and Definition 3.2, is characterized by the polynomial matrix equations (9a) N (p)H(p) = 0, (9b) N (p)F (p) 6= 0, where (9a) implements condition (1a) above (cf. Definition 3.1) while (9b) implements condition (1b) (cf. Definition 3.2). Both row polynomial vector N (p) and denominator polynomial a(p) can be viewed as design parameters. Throughout this study we, however, fix a(p) and aim to find an optimal N (p) with respect to a certain objective criterion related to the filter performance. In case there are more than one faults (nf > 1), it might be of interest to isolate the impact of one fault in the residual from the others. The following remark implies that the isolation problem is effectively a detection problem. 7 A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance Remark 4.1 (Fault Isolation). Consider model (5) and suppose nf > 1. In order to detect only one of the fault signals, say f1 , and isolate it from the other faults, fi , i ∈ {2, · · · , nf }, one may consider the detection problem for the same model but in new representation   x e E(x) + [H(p) F (p)] ˜ + L(p)z + F1 (p)f = 0, f where F1 (p) is the first column of F (p), and Fe(p) := [F2 (p), · · · , Fnf (p)], and f˜ := [f2 , · · · , fnf ]. In light of Remark 4.1, one can build a bank of filters where each filter aims to detect a particular fault while isolating the impact of the others; see [FKA09, Theorem 2] for more details on fault isolation. Next, we show how to transform the matrix polynomial equations (9) into a linear programming framework. Lemma 4.2. Let N (p) be a feasible polynomial matrix of degree dN for the inequalities (9), where dH dF dN X X X H(p) := Hi pi , F (p) := Fi pi , N (p) := Ni pi , i=0 nr ×nx i=0 nr ×nf i=0 1×nr , Fi ∈ R , and Ni ∈ R are constant matrices. Then, the polynomial and Hi ∈ R matrix inequalities (9) are equivalent, up to a scalar, to (10a) N̄ H̄ = 0, (10b) N̄ F̄ ∞ ≥ 1, where k · k∞ is the infinity vector norm, and  N̄ := N0 N1 ··· NdN  H1 ··· HdH 0 ··· H0 ··· .. . H0 HdH 0 .. ··· H1 .. . 0 H1 . ···  F0  0 F̄ :=  .  .. F1 ··· FdF 0 ··· F0 F1 .. . FdF 0 .. 0 ··· ··· .. . F0  H0  0 H̄ :=   .  .. 0 0 F1 . ···  0 ..  .  ,  0  HdH  0 ..  .  .  0  FdF Proof. It is easy to observe that N (p)H(p) = N̄ H̄[I pI · · · pi I]| , i := dN + dH , N (p)F (p) = N̄ F̄ [I pI · · · pj I]| , j := dN + dF . Moreover, in light of the linear structure of equations (9), one can simply scale the inequality (9b) and arrive at the assertion of the lemma.  Strictly speaking, the formulation in Lemma 4.2 is not a linear program, due to the non-convex constraint (10b). It is, however, easy to show that the characterization (10) can be understood as a number of linear programs, which grows linearly in the degree of the filter: 8 P. Mohajerin Esfahani and J. Lygeros Lemma 4.3. Consider the sets th n Nj := N̄ ∈ Rnr (dN +1) o N̄ H̄ = 0, N̄ F̄ vj ≥ 1 , ↓ j  | vj := 0, · · · , 1, · · · , 0 , Sm and let N := j=1 Nj where m := nf (dF + dN + 1) is the number of columns of F̄ (the parameters H̄, F̄ , nf , dF , dN are as considered in Lemma 4.2). Then, the set characterized by (10) is equivalent to N ∪ −N . Proof. Notice that kN̄ F̄ k∞ ≥ 1 if and only if there exists a coordinate j such that N̄ F̄ vj ≥ 1 or N̄ F̄ vj ≤ −1. Thus, the proof readily follows from the fact that each of the set Nj focuses on a component of the vector N̄ F̄ in (10b).  Fact 4.4. There exists a solution N (p) to (9) if and only if Rank [H(p) F (p)] > Rank H(p). Fact 4.4 provides necessary and sufficient conditions for the feasibility of the linear program formulation in Lemma 4.2; proof is omitted as it is an easy adaptation of the one in [FKA09, Corollary 3]. 4.2. Extension to Nonlinear Systems. In the presence of nonlinear terms E(x) 6= 0, it is straightforward to observe that the residual of filter (8) consists of two terms: (11) r := R(p)z = − a−1 (p)N (p)F (p)f − a−1 (p)N (p)E(x) . {z } | {z } | (i) (ii) Term (i) is the desired contribution of the fault f and is in common with the linear setup. Term (ii) is due to the nonlinear term E( · ) in (5). Our aim here will be to reduce the impact of E(x) while increasing the sensitivity to the fault f . To achieve this objective, we develop two approaches to control each of the two terms separately; in both cases we assume that the degree of the filter (i.e., dN in Lemma 4.2) and the denominator (i.e., a(p) in (11)) are fixed, and the aim is to design the numerator coefficients (i.e., N (p) in (11)). Approach (I) (Fault Sensitivity). To focus on fault sensitivity while neglecting the contribution of the nonlinear term, we assume that the system operates close to an equilibrium point xe ∈ Rnx . Even though in case of a fault the system may eventually deviate substantially from its nominal operating point, if the FDI filter succeeds in identifying the fault early the system will not have time to deviate too far. Hence, one may hope that a filter based on linearizing the system dynamics around the equilibrium would suffice. Then we assume, without loss of generality, that lim x→xe E(x) 2 = 0, kx − xe k2 where k · k2 stands for the Euclidean norm of a vector. If this is not the case, the linear part of E( · ) can be extracted and included in the linear part of the system. To increase the sensitivity of the linear filter to the fault f , we revisit the linear programming formulation (10) and seek a feasible numerator N (p) such that the coefficients of the transfer function N (p)F (p) attain maximum values within the admissible range. This gives rise to the 9 A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance following optimization problem:   max   N̄ F̄ N̄ (12) s.t.    ∞ N̄ H̄ N̄ =0 ≤1 ∞ where the objective function targets the contribution of the signal f to the residual r. Let us recall that N̄ F̄ is the vector containing all numerator coefficients of the transfer function f 7→ r. The second constraint in (12) is added to ensure that the solutions remain bounded; note that thanks to the linearity of the filter this constraint does not influence the performance. Though strictly speaking (12) is not a linear program, in a similar fashion as in Lemma 4.3 it is easy to transform it to a family of m different linear programs, where m is the number of columns of F̄ . How well the filter designed by (12) will work depends on the magnitude of the second term in (11), which is due to the nonlinearities E(x) and is ignored in (12). If the term generated by E(x) is large enough, the filter may lead to false alarms, whereas if we set our thresholds high to tolerate the disturbance generated by E(x) in nominal conditions, the filter may lead to missed detections. A direct way toward controlling this trade-off involving the nonlinear term will be the focus of the second approach. Approach (II) (Robustify to Nonlinearity Signatures). This approach is the main step toward the theoretical contribution of the article, and provides the principle ingredients to tackle the proposed perspectives AP and CP introduced in (3). The focus is on term (ii) of the residual (11), in relation to the mapping (1a). The idea is to robustify the filter against certain signatures of the nonlinearity during nominal operation. In the following we restrict the class of filters to the feasible solutions of polynomial matrix equations (9), characterized in Lemma 4.2. Let us denote the space of all p.w.c. functions from the interval [0, T ] to Rn by WTn . We equip this space with the L2 -inner product and the corresponding norm Z T p kekL2 := he, ei, he, gi := e| (t)g(t)dt, e, g ∈ WTn . 0 Consider an unknown signal x ∈ WTnx . In the context of the ODEs (6) that means we excite the system with the disturbance d( · ) for the time horizon T . We then stack d( · ) together with nr 1 the internal state X( · ) to introduce x := [ X d ]. We define the signals ex ∈ WT and rx ∈ WT as follows:  (13) ex (t) := E x(t) , rx (t) := −a−1 (p)N (p)[ex ](t), ∀t ∈ [0, T ]. The signal ex is the “nonlinearity signature” in the presence of the unknown signal x, and the signal rx is the contribution of the nonlinear term to the residual of the linear filter. Our goal now is to minimize krx kL2 in an optimization framework in which the coefficients of polynomial N (p) are the decision variables and the denominator a(p) is a fixed stable polynomial with the degree at least the same as N (p). Lemma 4.5. Let N (p) be a polynomial row vector of dimension nr and degree dN , and a(p) be a stable scalar polynomial with the degree at least dN . For any x ∈ WTnx there exists ψx ∈ n (d +1) such that WT r N (14a) rx (t) = N̄ ψx (t), (14b) kψx kL2 ≤ Ckex kL2 , ∀t ∈ [0, T ] p C := nr (dN + 1)ka−1 kH∞ , 10 P. Mohajerin Esfahani and J. Lygeros where N̄ is the vector collecting all the coefficients of the numerator N (p) as introduced in Lemma 4.2, and the signals ex and rx are defined as in (13). Proof. See Appendix I.1.  Given x ∈ WTnx and the corresponding function ψx as defined in Lemma 4.5, we have Z T 2 | krx kL2 = N̄ Qx N̄ , (15) Qx := ψx (t)ψx| (t)dt. 0 We call Qx the “signature matrix ” of the nonlinearity signature t 7→ ex (t) resulting from the unknown signal x. Given x and the corresponding signature matrix Qx , the L2 -norm of rx in (13) can be minimized by considering an objective which is a quadratic function of the filter coefficients N̄ subject to the linear constraints in (10):   min N̄ Qx N̄ |   N̄ (16) s.t. N̄ H̄ =0    ≥1 N̄ F̄ ∞ The program (16) is not a true quadratic program due to the second constraint. Following Lemma 4.3, however, one can show that the optimization program (16) can be viewed as a family of m quadratic programs where m = nf (dF + dN + 1). In the rest of the subsection, we establish an algorithmic approach to approximate the matrix Qx for a given x ∈ WTnx , with an arbitrary high precision. We first introduce a finite dimensional subspace of WT1 denoted by B := span{b0 , b1 , · · · , bk }, (17) Nnr where the collection of bi : [0, T ] → R is a basis for B. Let B nr := i=1 B be the nr Cartesian product of the set B, and TB : WTnr → B nr be the L2 -orthogonal projection operator onto B nr , i.e., (18) TB (ex ) = k X βi? bi , i=0 β ? := arg min ex − β k X β i bi L2 i=0 Let us remark that if the basis of B is orthonormal (i.e., hbi , bj i = 0 for i 6= j), then βi? = RT bi (t)ex (t)dt; we refer to [Lue69, Section 3.6] for more details on the projection operator. 0 Assumption 4.6. We stipulate that (i) The basis functions bi of subspace B are smooth and B is closed under the differentiation d operator p, i.e., for any b ∈ B we have p[b] = dt b ∈ B. (ii) The basis vectors in (17) are selected from an L2 -complete basis for WT1 , i.e., for any e ∈ WTnr , the projection error e − TB (e) L2 can be made arbitrarily small by increasing the dimension k of subspace B. The requirements of Assumptions 4.6 can be fulfilled for subspaces generated by, for example, the polynomial or Fourier basis. Thanks to Assumption 4.6(i), the linear operator p can be 11 A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance viewed as a matrix operator. That is, there exists a square matrix D with dimension k + 1 such that (19) p[B(t)] = d B(t) = DB(t), dt B(t) := [b0 (t), · · · , bk (t)]| . In Section 5.2 we will provide an example of such matrix operator for the Fourier basis. By virtue of the matrix representations of (19) we have   β? dN dN  β?D  X X   (20) Ni β ? Di B = N̄ D̄B, D̄ :=  .  , Ni pi β ? B = N (p)TB (ex ) = ..   i=0 i=0 β ? DdN where the vector β ? := [β0? , · · · , βk? ] is introduced in (18). If we define the positive semidefinite matrix G := [Gij ] of dimension k + 1 by Gij := a−1 (p)[bi ], a−1 (p)[bj ] , (21) we arrive at (22) a−1 (p)N (p)TB (e) 2 L2 = N̄ QB N̄ | , QB := D̄GD̄| , where D̄ and G are defined in (20) and (21), respectively. Note that the matrices G and D are built by the data of the subspace B and denominator a(p), whereas the nonlinearity signature only influences the coefficient β ? . The above discussion is summarized in Algorithm 1 with an emphasis on models described by the ODE (6), while Proposition 4.7 addresses the precision of the approximation scheme. Proposition 4.7 (Signature Matrix Approximation). Consider an unknown signal x : [0, T ] → Rnx in WTnx and the corresponding nonlinearity signature ex and signature matrix Qx as defined in (13) and (15), respectively. Let (bi )i∈N ⊂ WT1 be a family of basis functions satisfying Assumptions 4.6, and let B be the finite dimensional subspace in (17). If kex − TB (ex )kL2 < δ, where TB is the projection operator onto B nr , then  (23) Qx − QB 2 < C̄δ, C̄ := 1 + 2kex kL2 Cka−1 kH∞ , where QB is obtained by (22) (the output of Algorithm 1), and C is the same constant as in (14b). Proof. See Appendix I.1.  Remark 4.8 (Multi Signatures Training). In order to robustify the FDI filter to more than one unknown signal, say {xi ( · )}ni=1 , one may introduce an objective function as an average cost  Pn N̄ n1 i=1 Qxi N̄ | or the worst case viewpoint maxi≤n N̄ Qxi N̄ | , where Qxi is the signature matrix corresponding to xi as defined in (15). 4.3. Proposed Methodology and Probabilistic Performance. The preceding subsection proposed two optimization-based approaches to enhance the FDI filter design from linear to nonlinear system dynamics. Approach (I) targets the fault sensitivity while neglecting the nonlinear term of the system dynamics, and Approach (II) offers a QP framework to robustify the residual with respect to signatures of the dynamic nonlinearities. Here our aim is to achieve a 12 P. Mohajerin Esfahani and J. Lygeros Algorithm 1 Computing the signature matrix Qx in (15) (i) Initialization of the Filter Paramters: (a) Select a stable filter denominator a(p), a numerator degree dN not higher than a(p) order, and horizon T (b) Select a basis {bi }ki=1 ⊂ WT1 satisfying Assumptions 4.6 (c) Compute the differentiation matrix D in (19) (d) Compute the matrix G in (21) 1 (ii) Identification of the Nonlinearity Signature: (a) Input the disturbance pattern d( · ) for time horizon T (b) Solve (6) under inputs d( · ) and f ≡ 0 to obtain the internal state X( · ) (c) Set the unknown signal x(t) := [X | (t), d| (t)]|  |  | (d) Set the nonlinearity signature ex (t) := EX x(t) , EY| x(t) (iii) Computation of the Signature Matrix (a) Compute β ? from (18) (in case of orthonormal basis βi? = RT bi (t)ex (t)dt) 0 (b) Compute D̄ from (20) (c) Ouput QB := D̄GD̄| in (22) reconciliation between these two approaches. We subsequently provide theoretical results from the proposed solutions to the original design perspectives (3). Let (di )ni=1 ⊂ D be i.i.d. disturbance patterns generated from the probability space (Ω, F, P). For each di , let xi be the corresponding unknown signal with the associated signature matrix Qxi as defined in (15). In regard to the average perspective AP, we propose the two-stage (random) optimization program  min γ    γ,N̄     s.t. N̄ H̄ =0 f (24a) AP1 : N̄ F̄ ∞ ≥1    p  n  P  1   J N̄ Qxi N̄ | ≤γ n i=1  max N̄ F̄ ∞      N̄  N̄ H̄ =0  s.t. f2 : (24b) AP N̄ ∞ ≤1      n p  P  1  J kN̄1? k∞ N̄ Qxi N̄ | ≤ γ1?  n i=1 where J : R+ → R+ is an increasing and convex payoff function, and in the second stage (24b) N̄1? and γ1? are the optimizers of the first stage (24a), i.e., the programs (24) need to be 1A conservative but easy-to-implement approach is to set G an identity matrix with dimension k + 1. 13 A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance solved sequentially in a lexicographic (multi-objective) sense [MA04]. Let us recall that the filter coefficients can always be normalized with no performance deterioration. Hence, it is straightforward to observe that the main goal of the second stage is only to improve the coefficients of N̄ F̄ (concerning the fault sensitivity) while the optimality of the first stage (concerning the robustification to nonlinearity signatures) is guaranteed. Similarly, we also propose the following two-stage program for the perspective CP:   min γ   γ,N̄     s.t. N̄ H̄ =0 f1 : CP (25a) N̄ F̄ ∞ ≥1        max N̄ Qxi N̄ | ≤ γ i≤n (25b) f2 : CP  max   N̄     s.t.       N̄ F̄ ∞ N̄ H̄ N̄ =0 ≤1 ∞ kN̄1? k2∞ max N̄ Qxi N̄ i≤n  | ≤ γ1? Remark 4.9 (Computational Complexity). In view of Lemma 4.3, all the programs in (24) and (25) can be written as families of convex programs, and hence are tractable. It is, however, f is J(α) := α2 , the computational complexity worth noting that in case the payoff function of AP of the resulting programs in (24) is independent of the number of scenarios n, since the problems effectively reduce to a quadratic programming with a constraint involving the average of all the Pn respective signature matrices (i.e., n1 i=1 Qxi ). This is particularly of interest if one requires to train the filter for a large number of scenarios. Clearly, the filter designed by programs (24) and (25) is robustified to only finitely many most likely events, and as such, it may remain sensitive to disturbance patterns which have not been observed in the training phase. However, thanks to the probabilistic guarantees detailed in the sequel, we shall show that the probability of such failures (false alarm) is low. In fact, the tractability of our proposed scheme comes at the price of allowing for rare threshold violation of the filter. The rest of the subsection formalizes this probabilistic bridge between the program (24) (resp. (25)) and the original perspective AP (resp. CP) in (3) when the class of filters is confined to the linear residuals characterized in Lemma 4.2. For this purpose, we need a technical measurability assumption which is always expected to hold in practice. Assumption 4.10 (Measurability). We assume that the mapping D 3 d 7→ x ∈ WTnx is measurable where the function spaces are endowed with the L2 -topology and the respective Borel sigma-algebra. In particular, x can be viewed as a random variable on the same probability space as d. Assumption 4.10 is referred to the behavior of the system dynamics as a mapping from the disturbance d to the internal states. In the context of ODEs (6), it is well-known that under mild assumptions (e.g., Lipschitz continuity of EX ) the mapping d 7→ X is indeed continuous [Kha92, Chapter 5], which readily ensures Assumption 4.10. f Here we study the asymptotic behavior of the empirical 4.3.1. Probabilistic performance of AP.   average of E J(krk) uniformly in the filter coefficients N̄ , which allows us to link the solutions 14 P. Mohajerin Esfahani and J. Lygeros  of programs (24) to AP. Let N := N̄ ∈ Rnr(dN +1) : kN̄ k∞ ≤ 1 function of AP in (3) as the mapping φ : N × WTnx → R+ :   (26) φ(N̄ , x) := J krx kL2 = J kN̄ ψx kL2 , and consider the payoff where the second equality follows from Lemma 4.5. Theorem 4.11 (Average Performance). Suppose Assumption 4.10 holds and the random variable x is almost surely bounded2. Then, the mapping N̄ 7→ φ(N̄ , x) is a random function. Moreover, if (xi )ni=1 ⊂ WTnx are i.i.d. random variables and en is the uniform empirical average error n n1 X  o (27) en := sup φ(N̄ , xi ) − E φ(N̄ , x) , n i=1 N̄ ∈N then, (i) the Strong Law of Large Numbers (SLLN) holds, i.e., lim en = 0 almost surely. n→∞ √ (ii) the Uniform Central Limit Theorem (UCLT) holds, i.e., nen converges in law to a Gaussian variable with distribution N (0, σ) for some σ ≥ 0. Proof. See Appendix I.2 along with required preliminaries.  The following Corollary is an immediate consequence of the UCLT in Theorem 4.11 (ii). Corollary 4.12. Let assumptions of Theorem 4.11 hold, and en be the empirical average error (27). For all ε > 0 and k < 12 , we have  lim Pn nk en ≥ ε = 0, n→∞  where Pn denotes the n-fold product probability measure on Ωn , F n . f The formulation CP in (3) is known as chance con4.3.2. Probabilistic performance of CP. strained program which has received increasing attention due to recent developments toward f in (4)) in a convex tractable approaches, in particular via the scenario counterpart (cf. CP setting [CC06, CG08]. These studies are, however, not directly applicable to our problem due to the non-convexity arising from the constraint kN̄ F̄ k∞ ≥ 1. Here, following our recent work [MSL15], we exploit the specific structure of this non-convexity and adapt the scenario approach accordingly.  Let N̄n? , γn? be the optimizer obtained through the two-stage programs (25) where N̄n? is the filter coefficients and γn? represents the filter threshold; n is referred to the number of disturbance patterns. Given the filter N̄n? , let us denote the corresponding filter residual due to the signal x by rx [N̄n∗ ]; this is a slight modification of our notation rx in (13) to specify the filter coefficients. To quantify the filter performance, one may ask for the probability that a new unknown signal x 2 violates the threshold γn? when the FDI filter is set to N̄n? (i.e., the probability that rx [N̄n∗ ] L2 > γn∗ ). In the FDI literature such a violation is known as a false alarm, and from the CP standpoint its occurrence probability is allowed at most to the ε level. In this view the performance of the filter can be quantified by the event n   o  2 (28) E N̄n? , γn∗ := P rx [N̄n∗ ] L2 > γn∗ > ε . 2This assumption may be relaxed in terms of the moments of x, though this will not be pursued further here. 15 A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance f solution from the original perspective CP. The event (28) accounts for the feasibility of the CP Note that the measure P in (28) is referred to x whereas the stochasticity of the event stems  from the random solutions N̄n? , γn∗ .3 Theorem 4.13 (Chance Performance). Suppose Assumption 4.10 holds and (xi )ni=1 are i.i.d. f and random variables on (Ω, F, P). Let N̄n? ∈ Rnr (dN +1) and γn∗ ∈ R+ be the solutions of CP, n n measurable in F . Then, the set (28) is F -measurable, and for every β ∈ (0, 1) and any n such that  2  nf (dF + dN + 1) n≥ ln + nr (dN + 1) + 1 , ε β where dN is the degree of the filter and nf , nr , dF are the system size parameters of (5), we have   Pn E N̄n? , γn∗ < β. Proof. See Appendix I.2.  5. Cyber-Physical Security of Power Systems: AGC Case Study In this section, we illustrate the performance of our theoretical results to detect a cyber intrusion in a two-area power system. Motivated by our earlier studies [MVM+ 10, MVM+ 11], we consider the IEEE 118-bus power network equipped with primary and secondary frequency control. While the primary frequency control is implemented locally, the secondary loop, referred also as AGC (Automatic Generation Control), is closed over the SCADA system without human operator intervention. As investigated in [MVM+ 10], a cyber intrusion in this feedback loop may cause unacceptable frequency deviations and potentially load shedding or generation tripping. If the intrusion is, however, detected on time, one may prevent further damage by disconnecting the AGC. We show how to deploy the methodology developed in earlier sections to construct an FDI filter that uses the available measurements to diagnose an AGC intrusion sufficiently fast, despite the presence of unknown load deviations. 5.1. Mathematical Model Description. In this section a multi-machine power system, based only on frequency dynamics, is described [Andb]. The system is arbitrarily divided into two control areas. The generators are equipped with primary frequency control and each area is under AGC which adjusts the generating setpoints of specific generators so as to regulate frequency and maintain the power exchange between the two areas to its scheduled value. 5.1.1. System description. We consider a system comprising n buses and g number of generators. Let G = {i}g1 denote the set of generator indices and A1 = {i ∈ G | i in Area 1}, A2 = {i ∈ G | i in Area 2} the sets of generators that belong to Area 1 and Area 2, respectively. Let also Lktie = {(i, j)|i, j edges of a tie line from area k to the other areas}, where a tie line is a line connecting the two independently controlled areas and let also K = {1, 2} be the set of the indices of the control areas in the system. Using the classical generator model every synchronous machine is modeled as constant voltage source behind its transient reactance. The dynamic states of the system are the rotor angle δi (rad), the rotor electrical frequency fi (Hz) and the mechanical power (output of the turbine) 3The measure P is, with slight abuse of notation, the induced measure via the mapping addressed in Assump- tion 4.10. 16 P. Mohajerin Esfahani and J. Lygeros Pmi (M W ) for each generator i ∈ G. We also have one more state that represents the output of the AGC ∆Pagck for each control area k ∈ K. We denote by EG ∈ Cg a vector consisting of the generator internal node voltages EGi = for i ∈ G. The phase angle of the generator voltage node is assumed to coincide with 0 the rotor angle δi and |EGi | is a constant. The voltages of the rest of the nodes are included in VN ∈ Cn , whose entries are VN i = |VN i |∠θi for i = 1, . . . , n. To remove the algebraic constraints that appear due to the Kirchhoff’s first law for each node, we retain the internal nodes (behind the transient reactance) of the generators and eliminate the rest of the nodes. This could be achieved only under the assumption of constant impedance loads since in that way they can be included in the network admittance matrix. The node voltages can then be linearly connected to the internal node voltages, and hence to the dynamic state δi . This results in a reduced admittance matrix that corresponds only to the internal nodes of the generators, where the power flows are expressed directly in terms of the dynamic states of the system. The resulting model of the two area power system is described by the following set of equations. 0 |EGi |∠δi δ̇i = 2π(fi − f0 ), f0 1 (Pmi − Pei (δ) − (fi − f0 ) − ∆Ploadi ), 2Hi SBi Di 1 sat sat (P 0 = + vak ∆Pp,a + wak ∆Pagc,k − Pm,ak ), k Tch,ak m,ak X X = ckj (fj − f0 ) + bkj (Pmj − Pej (δ) − ∆Ploadj ) f˙i = Ṗm,ak ∆Ṗagc,k j∈Ak j∈Ak − 1 TNk gk (δ, f ) − Cpk hk (δ, f ) − Kk sat (∆Pagc,k − ∆Pagc,k ). TNk where i ∈ G, ak ∈ Ak for k ∈ K. Supperscript sat on the AGC output signal ∆Pagc,k and on the primary frequency control signal ∆Pp,ak highlights the saturation to which the signals are subjected. The primary frequency control is given by ∆Pp,i = −(fi − f0 )/Si . Based on the reduced admittance matrix, the generator electric power output is given by Pei = g X red EGi EGj (Gred ij cos(δi − δj ) + Bij sin(δi − δj )). j=1 dgk 0 ) and hk := Moreover, gk := (i,j)∈Lk (Pij − PT12 dt , where the power flow Pij , based on the tie initial admittance matrix of the system, is given by P Pij = |VNi ||VNj |(Gij cos(θi − θj ) + Bij sin(θi − θj )) All undefined variables are constants, and details on the derivation of the models can be found in [MVAL12]. The AGC attack is modeled as an additive signal to the AGC signal. For instance, if the attack signal is imposed in Area 1, the mechanical power dynamics of Area 1 will be modified as  1 0 sat sat (Pm,a + va1 ∆Pp,a + wa1 ∆Pagc + f (t) − Pm,a1 ), Ṗm,a1 = 1 1 1 Tch,a1 The above model can be compactly written as ( Ẋ(t) = h(X(t)) + Bd d(t) + Bf f (t) (29) Y (t) = CX(t), 17 A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance Figure 2. Stochastic load fluctuation and prediction error [Anda, p. 59]  | where X := {δi }1:g , {fi }1:g , {Pm,i }1:g , {∆Pagci }1:2 ∈ R3g+2 denotes the internal states vector comprising rotor angles δi , generators frequencies fi , generated mechanical powers  | Pm,i , and the AGC control signal ∆Pagci for each area. The external input d := {∆Ploadi }1:g represents the unknown load disturbances (discussed in the next subsection), and f represents the intrusion signal injected to the AGC of the first area. We assume that the measurements of all the frequencies and generated mechanical power are available, i.e., Y = {fi }1:g , {Pm,i }1:g ]| ∈ R2g . The nonlinear function h( · ) and the constant matrices Bd , Bf and C can be easily obtained by the mapping between the analytical model and (29). To transfer the ODE dynamic expression (29) into the DAE (5) it suffices to introduce " # X − Xe x := , z := Y − CXe d # " # " # " −pI + A Bd 0 h(X) − A(X − Xe ) , , H(p) := , L(p) := E(x) := −I 0 C 0 " # Bf F (p) := , 0 ∂h where Xe is the equilibrium of (29), i.e., h(Xe ) = 0, and A := ∂X . Notice that by the X=Xe above definition, the nonlinear term E( · ) only carries the nonlinearity of the system while the linear terms of the dynamic are incorporated into the constant matrices H, L, F . This can always be done without loss of generality, and practically may improve the performance of the scheme, as the linear terms can be fully decoupled from the residual of the filter. 5.1.2. Load Deviations and Disturbances. Small power imbalances arise during normal operation of power networks due, for example, to load fluctuation, load forecast errors, and trading on electricity market. Each of these sources give rise to deviations at different time scale. High frequency load fluctuation is typically time uncorrelated stochastic noise on a second or minute time scale, whereas forecast errors usually stem from the mismatch of predicted and actual consumption on a 15-minute time scale. Figure 2 demonstrates two samples of stochastic load fluctuation and forecast error which may appear at two different nodes of the network [Anda, p. 59]. The trading on the electricity market also introduces disturbances, for example, in an hourly framework (depending on the market). To capture these sources of uncertainty we consider a space of disturbance patterns comprising combinations of sinusoids at different frequency ranges (to model short term load fluctuation and mid-term forecast errors) and step functions (to model long-term abrupt changes due to the market). The space of load deviations (i.e., the disturbance patterns D in our FDI setting) is 18 P. Mohajerin Esfahani and J. Lygeros then described by (30) ∆Pload (t) := α0 + η X αi sin(ωi t + φi ), t ∈ [0, T ], i=1 where the parameters (αi )ηi=0 , (ωi )ηi=1 (φi )ηi=1 , and η are random variables whose distributions Pη induce the probability measure on D. We assume that i=0 |αi |2 is uniformly bounded with probability 1 to meet the requirements of Theorem 4.11. 5.2. Diagnosis Filter Design. To design the FDI filter, we set the degree of the filter dN = 7, the denominator a(p) = (p + 2)dN , and the finite time horizon T = 10 sec. Note that the degree of the filter is significantly less than the dimension of the system (29), which is 59. This is a general advantage of the residual generator approach in comparison to the observer-based approach where the filter order is effectively the same as the system dynamics. To compute the signature matrix Qx , we resort to the finite dimensional approximation QB in Proposition 4.7. Inspired by the class of disturbances in (30), we first choose Fourier basis with 80 harmonics ( cos( 2i ωt) i : even 2π bi (t) := (31) , ω := , i ∈ {0, 1, · · · , 80}. i+1 T ωt) i : odd sin( 2 We should emphasize that there is no restriction on the basis selection as long as Assumptions 4.6 are fulfilled; we refer to [MVAL12, Section V.B] for another example with a polynomial basis. Given the basis (31), it is easy to see that the differentiation matrix D introduced in (19) is   0 0 0 ··· 0 0 0 0 ω ··· 0 0      ··· 0 0  0 −ω 0  D = . .. .. .. ..  .. . .  .. . . . .    0 0 0 0 80ω  0 0 0 · · · −80ω 0 We can also compute offline (independent of x) the matrix G in (21) with the help of the basis (31) and the denominator a(p). To proceed with Qx of a sample ∆Pload we need to run the system dynamic (29) with the input d( · ) := ∆Pload and compute x(t) := [X(t)| , ∆Pload (t)]| where X is the internal states of the system. Given the signal x, we then project the nonlinearity signature  t 7→ ex (t) =: E x(t) onto the subspace B (i.e., TB (ex )), and finally obtain Qx from (22). In the following simulations, we deploy the YALMIP toolbox [Lof04] to solve the corresponding optimization problems. 5.3. Simulation Results. 5.3.1. Test system. To illustrate the FDI methodology we employed the IEEE 118-bus system. The data of the model are retrieved from a snapshot available at [ref]. It includes 19 generators, 177 lines, 99 load buses and 7 transmission level transformers. Since there were no dynamic data available, typical values provided by [AF02] were used for the simulations. The network was arbitrarily divided into two control areas whose nonlinear frequency model was developed in the preceding subsections. Figure 3 depicts a single-line diagram of the network and the boundaries of the two controlled areas where the first and second area contain, respectively, 12 and 7 generators. 19 A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance Area 2 Area 1 Figure 3. IEEE 118-bus system divided into two control areas 20 20 ΔP ΔP Load Load 5 Attack 10 0 0 5 10 (i) Load deviation at node 5 (dash) and AGC attack (solid) 0.2 15 Residual 0 −0.2 0 0 0 5 10 (i) Load deviation at node 5 (dash) and AGC attack (solid) 0.2 15 Residual 0 5 (ii) Linearized Model: E(x) ≡ 0 10 2 15 Residual 0 −2 0 5 Attack 10 −0.2 0 5 (ii) Linearized Model: E(x) ≡ 0 10 0.2 15 Residual 0 5 (iii) Full Nonlinear Model 10 15 −0.2 0 5 (iii) Full Nonlinear Model 10 15 (a) Performance of the filter neglecting the nonlinear (b) Performance of the filter trained for the step signaterm tures Figure 4. Performance of the FDI filters with step inputs 5.3.2. Numerical results. In the first simulation we consider the scenario that an attacker manipulates the AGC signal of the first area at Tack = 10 sec. We model this intrusion as a step signal equal to 14 M W injected into the AGC in Area 1. To challenge the filter, we also assume that a step load deviation occurs at Tload = 1 sec at node 5. In the following we present the results of two filters: Figure 4(a) shows the filter based on formulation (12) in Approach (I), which basically neglects the nonlinear term; Figure 4(b) shows the proposed filter in (24) based on AP perspective where the payoff function is J(α) := α2 ; see Remark 4.9 why such a payoff function is of particular interest. We validate the filters performance with two sets of measurements: first the measurements obtained from the linearized dynamic (i.e. E(x) ≡ 0); second the measurements obtained from the full nonlinear model (29). As shown in Fig. 4(a)(ii) and Fig. 4(b)(ii), both filters work perfectly well with linear dynamics measurements. It even appears that the first filter seems 20 P. Mohajerin Esfahani and J. Lygeros 1 0.9 0.8 0.7 ρ 0.6 0.5 0.4 0.3 0.2 0.1 0 0 100 200 300 400 500 600 Disturbance Samples 700 800 900 1000 Figure 5. The indicator ρ defined in (32) more sensitive. However, Fig. 4(a)(iii) and Fig. 4(b)(iii) demonstrate that in the nonlinear setting the first filter fails whereas the robustified filter works effectively similar to the linear setting. In the second simulation, to evaluate the filter performance in more realistic setup, we robustify the filter to random disturbance patterns, and then verify it with new generated samples. To measure the performance in the presence of the attack, we introduce the following indicator: max kr(t)k∞ (32) ρ := t≤Tack max kr(t)k∞ , t≤T where r is the residual (11), and Tack is when the attack starts. Observe that ρ ∈ [0, 1], and the lower ρ the better performance for the filter, e.g., in Fig. 4(a)(iii) ρ = 1, and in Fig. 4(b)(iii) ρ ≈ 0. In the training phase, we randomly generate five sinusoidal load deviations as described in (30), and excite the dynamics for T = 10 sec in the presence of each of the load deviations individually. Hence, in total we have n = 19 × 5 = 95 disturbance signatures. Then, we compute f in (24) with the payoff function J(α) := α2 and these the filter coefficients by virtue of AP 95 samples. In the operation phase, we generate two new disturbance patterns with the same distribution as in the training phase and run the system in the presence of both load deviations simultaneously at two random nodes for the horizon T = 120 sec. Meanwhile, we inject an attack signal at Tack = 110 sec in the AGC, and compute the indicator ρ in (32). Figure 5 demonstrates the result of this simulation for 1000 experiments. 6. Conclusion and Future Directions In this article, we proposed a novel perspective toward the FDI filter design, which is tackled via an optimization-based methodology along with probabilistic performance guarantees. Thanks to the convex formulation, the methodology is applicable to high dimensional nonlinear systems in which some statistical information of exogenous disturbances are available. Motivated by our earlier works, we deployed the proposed technique to design a diagnosis filter to detect the AGC malfunction in two-area power network. The simulation results validated the filter performance, particularly when the disturbance patterns are different from training to the operation phase. 21 A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance The central focus of the work here is to robustify the filter to certain signatures of dynamic nonlinearities in the presence of given disturbance patterns. As a next step, motivated by applications that the disruptive attack may follow certain patterns, a natural question is whether the filter can be trained to these attack patterns. From the technical standpoint, this problem in principle may be different from the robustification process since the former may involve maximization of the residual norm as opposed to the minimization for the robustification discussed in this article. Therefore, this problem offers a challenge to reconcile the disturbance rejection and the fault sensitivity objectives. The proposed methodology in this study is applicable to both discrete and continuous-time dynamics and measurements. In reality, however, we often have different time-setting in different parts, i.e., we only have discrete-time measurements while the system dynamics follows a continuous-time behavior. We believe this setup introduces new challenges to the field. We recently reported heuristic attempts toward this objective in [ETML13], though there is still a need to address this problem in a rigorous and systematic framework. Acknowledgment The authors are grateful to M. Vrakopoulou and G. Andersson for the help on the AGC case study. The first author also thanks G. Schildbach for fruitful discussions on randomized algorithms. I. Appendix I.1. Proofs of Section 4.2. Let us start with a preliminary required for the main proof of this section. PdN Lemma I.1. Let N (p) := i=0 Ni pi be an Rnr row polynomial vector with degree dN , and a(p) be a stable polynomial with the degree at least dN . Let N̄ := [N0 N1 · · · NdN ] be the collection of the coefficients of N (p). Then, p e N̄ k∞ , e := nr (dN + 1) ka−1 kH . a−1 N H∞ ≤ Ck C ∞ Proof. Let b(p) := (I.1)  Let b̄ := b0 (I.2) b1 PdN i be a polynomial scaler function. By H∞ -norm definition we have PdN 2 2i b(jω) 2 2 i=0 |bi | ω ≤ sup . a−1 b H∞ = sup |a(jω)|2 ω∈[0,∞) ω∈(−∞,∞) a(jω)  · · · bdN . It is then straightforward to inspect that ( dN X (dN + 1)kb̄k2∞ if ω ∈ [0, 1] 2 2i |bi | ω ≤ 2 2dN (dN + 1)kb̄k∞ ω if ω ∈ (1, ∞) i=0 i=0 bi p Therefore, (I.1) together with (I.2) yields to a−1 b 2 H∞ ≤ (dN + 1) a−1 2 kb̄k2∞ . H∞ Now, taking the dimension of the vector N (p) into consideration, we conclude the desired assertion.  22 P. Mohajerin Esfahani and J. Lygeros Proof of Lemma 4.5. Let ` ≥ dN be the degree of the scalar polynomial a(p). Then, taking advantage of the state-space representation of the matrix transfer function a−1 (p)N (p), in particular the observable canonical form [ZD97, Section 3.5], we have Z t rx (t) = C e−A(t−τ ) Bex (τ )dτ + Dex (t), 0 1×` where C ∈ R is a constant vector, A ∈ R`×` is the state matrix depending only on a(p), `×nr and B ∈ R and D ∈ R1×nr are matrices that depend linearly on all the coefficients of the numerator N̄ ∈ Rnr (dN +1) . Therefore, it can be readily deduced that (14a) holds for some n (d +1) function ψx ∈ WT r N . In regard to (14a) and the definition (13), we have (I.3) kN̄ ψx kL2 = krx kL2 = a−1 (p)N (p)ex L2 ≤ a−1 N H∞ e N̄ k∞ kex kL , kex kL2 ≤ Ck 2 where the first inequality follows from the classical result that the L2 -gain of a matrix transfer function is the H∞ -norm of the matrix [ZD97, Theorem 4.3, p. 51], and the second inequality follows from Lemma I.1. Since (I.3) holds for every N̄ ∈ Rnr (dN +1) , then p e x kL , kψx kL2 ≤ nr (dN + 1) Cke 2 which implies (14b).  Proof of Proposition 4.7. Observe that by virtue of the triangle inequality and linearity of the projection mapping we have  e N̄ k∞ δ, krx kL2 − a−1 (p)N (p)TB (ex ) L2 ≤ a−1 (p)N (p) ex − TB (ex ) L2 ≤ Ck where the second inequality follows in the same spirit as (I.3) and kex − TB (ex )kL2 ≤ δ. Note that by definitions of Qx and QB in (15) and (22), respectively, we have  2 e N̄ k∞ δ Ck e N̄ k∞ δ + 2krx kL N̄ (Qx − QB )N̄ | = krx k2L2 − a−1 (p)N (p)TB (ex ) L ≤ Ck 2 2   e 2 kN̄ k2∞ δ δ + 2kex kL ≤ Cka−1 kH kN̄ k22 δ 1 + 2kex kL ≤C 2 ∞ 2 where the inequality of the first line stems from the simple inequality |α2 − β 2 | ≤ |α − β| 2|α| +  |α − β| , and C is the constant as in (14b).  I.2. Proofs of Section 4.3. To prove Theorem 4.11 we need a preparatory result addressing the continuity of the mapping φ in (26). Lemma I.2. Consider the function φ as defined in (26). Then, there exists a constant L > 0 such that for any N̄1 , N̄2 ∈ N and x1 , x2 ∈ WTnx where kxi kL2 ≤ M , we have  φ(N̄1 , x1 ) − φ(N̄2 , x2 ) ≤ L N̄1 − N̄2 ∞ + kx1 − x2 kL2 . Proof. Let LE be the Lipschitz continuity constant of the mapping E : Rnx → Rnr in (13). We modify the notation of rx in (13) with a new argument as rx [N̄ ], in which N̄ represents the filter coefficients. Then, with the aid of (I.3), we have sup sup krx [N̄ ]kL2 ≤ kxkL2 ≤M N̄ ∈N sup e E kN̄ k∞ kxkL ≤ M f, sup CL 2 f := CL e E M, M kxkL2 ≤M N̄ ∈N e is introduced in Lemma I.1. As the payoff function J is convex, it is then where the constant C f] [Ber09, Proposition 5.4.2, p. 185]; we denote Lipschitz continuous over the compact set [0, M this Lipschitz constant by LJ . Then for any N̄i ∈ N and kxi kL2 ≤ M , i ∈ {1, 2}, we have, φ(N̄1 , x1 ) − φ(N̄2 , x2 ) ≤ LJ rx1 [N1 ] L2 23 − rx2 [N2 ] L2 (I.4) A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance   ≤ LJ rx1 [N1 ] − rx1 [N2 ] L + rx1 [N2 ] − rx2 [N2 ] L 2 2   e x kL kN1 − N2 k∞ + Cke e x − ex kL kN2 k∞ ≤ LJ Cke 1 2 1 2 2  e E M kN1 − N2 k∞ + kx1 − x2 kL . ≤ LJ CL 2 where (I.4) follows from (I.3) and the fact that the mapping (N̄ , ex ) 7→ rx [N̄ ] is bilinear.  Proof of Theorem 4.11. By virtue of Lemma I.2, one can infer that for every N̄ ∈ N the mapping x 7→ φ(N̄ , x) is continuous, and hence measurable. Therefore, φ(N̄ , x) can be viewed as a random variable for each N̄ ∈ N , which yields to the first assertion, see [Bil99, Chapter 2, p. 84] for more details. By uniform (almost sure) boundedness and again Lemma I.2, the mapping N̄ 7→ φ(N̄ , x) is uniformly Lipschitz continuous (except on a negligible set), and consequently first moment continuous in the sense of [Han12, Definition 2.5]. We then reach (i) by invoking [Han12, Theorem 2.1]. For assertion (ii), note that the compact set N is finite dimensional, and thus admits a logarithmic ε-capacity in the sense of [Dud99, Section. 1.2, p. 11]. Therefore, the condition [Dud99, (6.3.4), p. 209] is satisfied. Since the other requirements of [Dud99, Theorem 6.3.3, p. 208] are readily fulfilled by the uniform boundedness assumption and Lemma I.2, we arrive at the desired UCLT assertion in (ii).  To keep the paper self-contained, we provide a proof for Theorem 4.13 in the following, but refer the interested reader to [MSL15, Theorem 4.1] for a result of a more general setting. Proof of Theorem 4.13. The measurability of E is a straightforward consequence of the measurability of [N̄n? , γn? ] and Fubini’s Theorem [Bil95, Theorem 18.3, p. 234]. For notational simplicity, we introduce the following notation. Let ` := nr (dN + 1) + 1 and define the function f : R` × WTnx → R f (θ, x) := N̄ Qx N̄ | − γ, θ := [N̄ , γ]| ∈ R` , where Qx is the nonlinearity signature matrix of x as defined in (15), and θ is the augmented vector collecting all the decision variables. Consider the convex sets Θj ⊂ R` th n Θj := θ = [N̄ , γ]| o N̄ H̄ = 0, N̄ F̄ vj ≥ 1 , ↓ j  | vj := 0, · · · , 1, · · · , 0 , where the size of vj is m := nf (dF + dN + 1). Note that in view of Lemma 4.3, we can replace the Sm characterization of the filter coefficients in (10) with θ ∈ j=1 Θj . We then express the program f 1 in (25a) as follows: CP in (3) and its random counterpart CP   |  min c| θ min c θ   m   S m S θ∈ Θ j θ∈ Θj f1 : CP : CP j=1 j=1       s.t. max f (θ, xi ) ≤ 0, s.t. P f (θ, x) ≤ 0 ≥ 1 − ε i≤n where c is the constant vector with 0 elements except the last which is 1. It is straightforward f in (25) is the same as the to observe that the optimal threshold γn? of the two-stage program CP f optimal threshold obtained in the first stage CP1 . Thus, it suffices to show the desired assertion 24 P. Mohajerin Esfahani and J. Lygeros f 1 . Now, consider considering only the first stage. Let θn? := [N̄n? , γn? ] denote the optimizer of CP f m sub-programs denoted by CP (j) and CP (j) for j ∈ {1, · · · , m}:  (  min c| θ min c| θ θ∈Θj θ∈Θj f CP (j) : CP (j) :   s.t. max f (θ, xi ) ≤ 0, s.t. P f (θ, x) ≤ 0 ≥ 1 − ε i≤n f (j) by θ∗ . Note that for all j, the set Θj is deterministic Let us denote the optimal solution of CP n,j f (j) is feasible if (not affected by x) and convex, and the corresponding random program CP f Θj 6= ∅, thanks to the min-max structure of CP (j). Therefore, we can readily employ the existing results of the random convex problems. Namely, by [CG08, Theorem 1] we have `−1    X n i ∗ Pn E(θn,j ) < ε (1 − ε)n−i , ∀j ∈ {1, · · · , m} i i=0 m ∗ where E is introduced in (28). Furthermore, it is not hard to inspect that θn∗ ∈ θn,j . Thus, j=1 Sm ∗ ∗ E(θn ) ⊆ j=1 E(θn,j ) which yields `−1   m m  X [ X   n i ∗ ∗ ε (1 − ε)n−i . Pn E(θn,j ) <m Pn E(θn∗ ) ≤ Pn E(θn,j ) ≤ i i=0 j=1 j=1 Now, considering β as an upper bound, the desired assertion can be obtained by similar calculation as in [Cal09] to make the above inequality explicit for n in terms of ε and β.  References [Ada75] [AF02] [Anda] [Andb] [Bea71] [Ber09] [Bil95] [Bil99] [Cal09] [CC06] [CG08] [CP82] [CS98] [DS99] Robert A. Adams, Sobolev spaces, Academic Press [A subsidiary of Harcourt Brace Jovanovich, Publishers], New York-London, 1975, Pure and Applied Mathematics, Vol. 65. P. M. Anderson and A. A. Fouad, Power System Control and Stability, IEEE Computer Society Press, 2002. Gran Andersson, Dynamics and control of electric power systems, Power System Laboratory, ETH Zurich,. , Power system analysis, Power System Laboratory, ETH Zurich,. R. V. Beard, Failure accommodation in linear systems through self-reorganization, Ph.D. thesis, Massachusetts Inst. Technol., Cambridge, MA, 1971. Dimitri P. Bertsekas, Convex Optimization Theory, Athena Scientific, 2009. MR 2830150 (2012f:90001) Patrick Billingsley, Probability and measure, third ed., Wiley, 1995. MR 1324786 (95k:60001) , Convergence of Probability Measures, second ed., Wiley Series in Probability and Statistics: Probability and Statistics, John Wiley & Sons Inc., New York, 1999. MR 1700749 (2000e:60008) Giuseppe C. Calafiore, A note on the expected probability of constraint violation in sampled convex programs, 18th IEEE International Conference on Control Applications Part of 2009 IEEE Multiconference on Systems and Control, july 2009, pp. 1788 – 1791. Giuseppe C. Calafiore and Marco C. Campi, The scenario approach to robust control design, IEEE Trans. Automat. Control 51 (2006), no. 5, 742–753. MR 2232597 (2007a:93075) M. C. Campi and S. Garatti, The exact feasibility of randomized solutions of uncertain convex programs, SIAM J. Optim. 19 (2008), no. 3, 1211–1230. MR 2460739 (2009j:90081) J. Chen and R. Patton, Robust model based faults diagnosis for dynamic systems, Dordrecht: Kluwer Academic Publishers, New York, 1982. Walter H. Chung and Jason L. Speyer, A game-theoretic fault detection filter, IEEE Trans. Automat. Control 43 (1998), no. 2, 143–161. Randal K Douglas and Jason L Speyer, H bounded fault detection filter, Journal of guidance, control, and dynamics 22 (1999), no. 1, 129–138. 25 A Tractable FDI Approach for Nonlinear Systems with Probabilistic Performance [Dud99] R. M. Dudley, Uniform Central Limit Theorems, Cambridge Studies in Advanced Mathematics, vol. 63, Cambridge University Press, Cambridge, 1999. MR 1720712 (2000k:60050) [EFK13] Daniel Eriksson, Erik Frisk, and Mattias Krysander, A method for quantitative fault diagnosability analysis of stochastic linear descriptor models, Automatica 49 (2013), no. 6, 1591–1600. [ETML13] Erasmia Evangelia Tiniou, Peyman Mohajerin Esfahani, and John Lygeros, Fault detection with discrete-time measurements: an application for the cyber security of power networks, 52nd IEEE Conference Decision and Control, Dec 2013, pp. 194–199. [FF12] Giuseppe Franzè and Domenico Famularo, A robust fault detection filter for polynomial nonlinear systems via sum-of-squares decompositions, Systems & Control Letters 61 (2012), no. 8, 839–848. [FKA09] Erik Frisk, Mattias Krysander, and Jan Aslund, Sensor placement for fault isolation in linear differential-algebraic systems, Automatica 45 (2009), no. 6, 364–371. [Han12] Lars Peter Hansen, Proofs for large sample properties of generalized method of moments estimators, Journal of Econometrics 170 (2012), no. 2, 325–330. MR 2970318 [HKEY99] H. Hammouri, M. Kinnaert, and E.H. El Yaagoubi, Observer-based approach to fault detection and isolation for nonlinear systems, Automatic Control, IEEE Transactions on 44 (1999), no. 10, 1879 –1884. [HP96] M. Hou and R.J. Patton, An lmi approach to H− /H∞ fault detection observers, Control ’96, UKACC International Conference on (Conf. Publ. No. 427), vol. 1, sept. 1996, pp. 305 – 310 vol.1. [Jon73] H. L. Jones, Failure detection in linear systems, Ph.D. thesis, Massachusetts Inst. Technol., Cambridge, MA, 1973. [Kha92] Hassan K. Khalil, Nonlinear systems, Macmillan Publishing Company, New York, 1992. MR 1201326 (93k:34001) [Lof04] J. Lofberg, Yalmip : a toolbox for modeling and optimization in matlab, Computer Aided Control Systems Design, 2004 IEEE International Symposium on, sept. 2004, pp. 284 –289. [Lue69] David G. Luenberger, Optimization by vector space methods, John Wiley & Sons Inc., New York, 1969. [MA04] R Timothy Marler and Jasbir S Arora, Survey of multi-objective optimization methods for engineering, Structural and multidisciplinary optimization 26 (2004), no. 6, 369–395. [Mas86] Mohammad-Ali Massoumnia, A geometric approach to the synthesis of failure detection filters, IEEE Trans. Automat. Control 31 (1986), no. 9, 839–846. [MEVAL] Peyman Mohajerin Esfahani, Maria Vrakopoulou, Goran Andersson, and John Lygeros, Intrusion detection in electric power networks, Patent applied for EP-12005375, filed 24 July 2012. [MSL15] Peyman Mohajerin Esfahani, Tobias Sutter, and John Lygeros, Performance bounds for the scenario approach and an extension to a class of non-convex programs, IEEE Transactions on Automatic Control 60 (2015), no. 1, 46–58. [MVAL12] Peyman Mohajerin Esfahani, Maria Vrakopoulou, Goran Andersson, and John Lygeros, A tractable nonlinear fault detection and isolation technique with application to the cyber-physical security of power systems, 51st IEEE Conference Decision and Control, Dec 2012, Full version: http://control. ee.ethz.ch/index.cgi?page=publications;action=details;id=4196, pp. 3433–3438. [MVM+ 10] Peyman Mohajerin Esfahani, Maria Vrakopoulou, Kostas Margellos, John Lygeros, and Goran Andersson, Cyber attack in a two-area power system: Impact identification using reachability, American Control Conference, 2010, pp. 962–967. [MVM+ 11] , A robust policy for automatic generation control cyber attack in two area power network, 49th IEEE Conference Decision and Control, 2011, pp. 5973–5978. [MVW89] Mohammad-Ali Massoumnia, G. C. Verghese, and A. S. Willsky, Failure detection and identification, IEEE Transaction on Automatic Control 34 (1989), no. 3, 316–321. [NF06] Mattias Nyberg and Erik Frisk, Residual generation for fault diagnosis of system described by linear differential-algebraic equations, IEEE Transaction on Automatic Control 51 (2006), no. 12, 1995– 2000. [PI01] Claudio De Persis and Alberto Isidori, A geometric approach to nonlinear fault detection and isolation, IEEE Trans. Automat. Control 46 (2001), no. 6, 853–865. [PW98] Jan Willem Polderman and Jan C. Willems, Introduction to mathematical systems theory, Texts in Applied Mathematics, vol. 26, Springer-Verlag, New York, 1998, A behavioral approach. [ref] Power systems test case archive, college of engineering,university of washington, URL: http://www.ee.washington.edu/research/pstca/. 26 P. Mohajerin Esfahani and J. Lygeros [SF91] R. Seliger and P.M. Frank, Fault diagnosis by disturbance-decoupled nonlinear observers, Proceedings of the 30th IEEE Conference on Decision and Control, 1991, pp. 2248–2253. [Shc07] A. A. Shcheglova, Nonlinear differential-algebraic systems, Sibirsk. Mat. Zh. 48 (2007), no. 4, 931– 948. MR 2355385 (2009c:34002) [SMEKL13] Bratislav Svetozarevic, Peyman Mohajerin Esfahani, Maryam Kamgarpour, and John Lygeros, A robust fault detection and isolation filter for a horizontal axis variable speed wind turbine, American Control Conference (ACC), 2013, June 2013, pp. 4453–4458. [ZD97] Kemin Zhou and John C. Doyle, Essentials of robust control, Prentice Hall, September 1997. 27
3cs.SY
1 Secure Group Testing arXiv:1607.04849v3 [cs.IT] 16 Jan 2018 Alejandro Cohen Asaf Cohen Abstract—The principal goal of Group Testing (GT) is to identify a small subset of “defective” items from a large population, by grouping items into as few test pools as possible. The test outcome of a pool is positive if it contains at least one defective item, and is negative otherwise. GT algorithms are utilized in numerous applications, and in many of them maintaining the privacy of the tested items, namely, keeping secret whether they are defective or not, is critical. In this paper, we consider a scenario where there is an eavesdropper (Eve) who is able to observe a subset of the GT outcomes (pools). We propose a new non-adaptive Secure Group Testing (SGT) algorithm based on information-theoretic principles, which keeps the eavesdropper ignorant regarding the items’ status. Specifically, when the fraction of tests observed by Eve is 0 ≤ δ < 1, we prove that the number of tests required for both correct reconstruction at the legitimate user (with high 1 probability) and negligible information leakage to Eve is 1−δ times the number of tests required with no secrecy constraint for fixed K regime. By a matching converse, we completely characterize the Secure GT capacity. Moreover, we consider a computationally efficient decoding algorithm, and prove that for δ < 1/2 the number of tests required, without any constraint on 1 times the number of tests required with no K, is at most 1/2−δ secrecy constraint. I. I NTRODUCTION The classical version of Group Testing (GT) was suggested during World War II in order to identify syphilis-infected draftees while dramatically reducing the number of required tests [1]. Specifically, when the number of infected draftees, K, is much smaller than the population size, N , instead of examining each blood sample individually, one can conduct a small number of of pooled samples. Each pool outcome is negative if it contains no infected sample, and positive if it contains at least one infected sample. The problem is thus to identify the infected draftees via as few pooled tests as possible. Figure 1 (a)-(c) depicts a small example. Since its exploitation in WWII, GT has been utilized in numerous fields, including biology and chemistry [2], [3], communications [4]–[7], sensor networks [8], pattern matching [9] and web services [10]. GT has also found applications in the emerging field of Cyber Security, e.g., detection of significant changes in network traffic [11], Denial of Service attacks [12] and indexing information for data forensics [13]. Many scenarios which utilize GT involve sensitive information which should not be revealed if some of the tests leak (for instance, if one of the several labs to which tests have been distributed for parallel processing is compromised). However, in GT, leakage of even a single pool-test outcome may reveal significant information about the tested items. If A. Cohen, A. Cohen and O. Gurewitz are with the Department of Communication Systems Engineering, Ben-Gurion University of the Negev, BeerSheva 84105, Israel (e-mail: [email protected]; [email protected]; [email protected]). S. Jaggi is with the Department of Information Engineering, Chinese University of Hong Kong ([email protected]). Parts of this work were presented at the IEEE International Symposium on Information Theory, ISIT 2016. Sidharth Jaggi Omer Gurewitz the test outcome is negative it indicates that none of the items in the pool is defective; if it is positive, at least one of the items in the pool is defective (see Figure 1 (d) for a short example). Accordingly, it is critical to ensure that a leakage of a fraction of the pool-tests outcomes to undesirable or malicious eavesdroppers does not give them any useful information on the status of the items. It is very important to note that protecting GT is different from protecting the communication between the parties. To protect GT, one should make sure that information about the status of individual items is not revealed if a fraction of the test outcomes leaks. However, in GT, we do not want to assume one entity has access to all pool-tests, and can apply some encoding function before they are exposed. We also do not want to assume a mixer can add a certain substance that will prevent a third party from testing the sample. To protect GT, one should make sure that without altering mixed samples, if a fraction of them leaks, either already tested or not, information is not revealed. While the current literature includes several works on the privacy in GT algorithms for digital objects [13]–[16], these works are based on cryptographic schemes, assume the testing matrix is not known to all parties, impose a high computational burden, and, last but not least, assume the computational power of the eavesdropper is limited [17], [18]. Information theoretic security considered for secure communication [18], [19], on the other hand, if applied appropriately to GT, can offer privacy at the price of additional tests, without keys, obfuscation or assumptions on limited power. Due to the analogy between channel coding and group-testing regardless of security constraints, [20], [21], in Section II we present an extensive survey of the literature on secure communication as well. Main Contribution In this work, we formally define Secure Group Testing (SGT), suggest SGT algorithms based on information-theoretic principles and analyse their performance. In the considered model, there is an eavesdropper Eve who might observe part of the vector of pool-tests outcomes. The goal of the test designer is to design the tests in a manner such that a legitimate decoder can decode the status of the items (whether the items are defective or not) with an arbitrarily small error probability. It should also be the case that as long as Eve the eavesdropper gains only part of the output vector (a fraction δ - a bound on the value of δ is known a priori to the test designer, but which specific items are observed is not), Eve cannot (asymptotically, as the number of items being tested grows without bound) gain any significant information on the status of any of the items. We propose a SGT code and corresponding decoding algorithms which ensure high of reliability (with high probability over the test design, the legitimate decoder should be able to estimate the status of each item correctly), as well as strong 2 Secure Group-Test Model - Erasure Eavesdropper ∧ Items N Testing Matrix X Y (t) Main Channel T uses Eavesdropper Channel δ 0 1−δ 1 0 Z (t ) Test Outcomes N Eve ? δ 1 Figure 2: An analogy between a wiretap erasure channel and the corresponding SGT model. (a) (c) (b) (d) Figure 1: Classical group testing: An example of test results, a simple decoding procedure at the legitimate decoder and the risk of leakage. The example includes 7 items, out of which at most one defective (the second one in this case; unknown to the decoder). Three pooled tests are conducted. Each row dictates in which pooled tests the corresponding item participates. (a) Since the first result is negative, items 1 and 6 are not defective. (b) The second result is positive, hence at least one of items 2 and 4 is defective. (c) Based on the last result, as item 4 cannot be defective, it is clear that 2 is defective. Note that decoding in this case is simple: any algorithm which will simply rule out each item whose row in the matrix is not compatible with the result will rule out all but the second item, due to the first and last test results being negative, thus identifying the defective item easily. (d) An eavesdropper who has access to part of the results (the first two) can still infer useful information. Our goal is construct a testing matrix such that such an eavesdropper remains ignorant. secrecy conditions (as formally defined in Section III) - which ensures that essentially no information about the status of individual items leaks to Eve. Our first SGT code and corresponding decoding algorithm (based on Maximum Likelihood (ML) decoding) requires a number of tests that is essentially information-theoretically optimal in N, K and δ (as demonstrated in Section VI by corresponding information-theoretic converse that we also show for the problem). The second code and corresponding decoding algorithm, while requiring a constant factor larger number of tests than is information-theoretically necessary (by a factor that is a function of δ), is computationally efficient. It maintains the reliability and secrecy guarantees, yet requires only O(N 2 T ) decoding time, where T is the number of tests. We do so by proposing a model, which is, in a sense, analogous to a wiretap channel model, as depicted in Figure 2. In this analogy the subset of defective items (unknown a priori to all parties) takes the place of a confidential message. The testing matrix (representing the design of the pools - each row corresponds to the tests participated in by an item, and each column corresponds to a potential test) is a succinct representation of the encoder’s codebook. Rows or disjunctive unions of rows of this testing matrix can be considered as codewords. The decoding algorithm is analogous to a channel decoding process, and the eavesdropped signal is the output of an erasure channel, namely, having only any part of the transmitted signal from the legitimate source to the legitimate receiver. In classical non-adaptive group-testing, each row of the testing matrix comprises of a length-T binary vector which determines which pool-tests the item is tested in. In the SGT code constructions proposed in this work, each item instead corresponds to a vector chosen uniformly at random from a pre-specified set of random and independent vectors. Namely, we use stochastic encoding, and each vector corresponds to different sets of pool-tests an item may participate in. For each item the lab picks one of the vectors in its set (we term the set associated with item j as “Bin j”) uniformly at random, and the item participates in the pool-tests according to this randomly chosen vector. The set (“Bin”) is known a priori to all parties, but the specific vector chosen by the encoder/mixer is only known to the encoder/mixer, and hence is not a shared key/common randomness in any sense. A schematic description of our procedure is depicted in Figure 4. Accordingly, by obtaining a pool-test result, without knowing the specific vectors chosen by the lab for each item, the eavesdropper may gain only negligible information regarding the items themselves. Specifically, we show that by careful design of the testing procedure, even though the pool-tests in which each item participated are chosen randomly and even though the legitimate user does not know a-priori in which pool-tests each item has participated, the legitimate user will, with high probability over the testing procedure, be able to correctly identify the set of defective items, while the eavesdropper, observing only a subset of the pool-test results, will have no significant information regarding the status of the items. The structure of this work is as follows. In Section II, we present an extensive survey and summarize the related work. In Section III, a SGT model is formally described. Section IV includes our main results, with the direct proved in Section V and converse proved in Section VI. Section VII 3 describes a computationally efficient algorithm, and proves an upper bound on its error probability. Section VIII concludes the paper. II. BACKGROUND AND R ELATED W ORK A. Group-testing Group-testing comes in various flavours, and the literature on these is vast. At the risk of leaving out much, we reference here just some of the models that have been considered in the literature, and specify our focus in this work. 1) Performance Bounds: GT can be non-adaptive, where the testing matrix is designed beforehand, adaptive, where each new test can be designed while taking into account previous test results, or a combination of the two, where testing is adaptive, yet with batches of non-adaptive tests. It is also important to distinguish between exact recovery and a vanishing probability of error. To date, the best known lower bound on the number of K2 tests required (non-adaptive, exact recovery) is Ω( log K log N ) [22]. The best known explicit constructions were given in [23], resulting in O(K 2 log N ). However, focusing on exact recovery requires more tests, and forces a combinatorial nature on the problem. Settling for high probability reconstructions allows one to reduce the number of tests to the order of K log N .1 For example, see the channel-coding analogy given in [20]. A similar analogy to wiretap channels will be at the basis of this work as well. In fact, probabilistic methods with an error probability guarantee appeared in [24], without explicitly mentioning GT, yet showed the O(K log N ) bound. Additional probabilistic methods can be found in [25] for support recovery, or in [26], when an interesting phase transition phenomenon was observed, yielding tight results on the threshold (in terms of the number of tests) between the error probability approaching one or vanishing. 2) A Channel Coding Interpretation: As mentioned, the analogy to channel coding has proved useful [20]. [21] defined the notion of group testing capacity, that is, the value of log (N ) limN →∞ T K under which reliable algorithms exist, yet, over which, no reliable reconstruction is possible. A converse result for the Bernoulli, non-adaptive case was given in [27]. Strong converse results were given in [28], [29], again, building on the channel coding analogy, as well as converses for noisy GT [30]. In [31], adaptive GT was analyzed as a channel coding with feedback problem. 3) Efficient Algorithms: A wide variety of techniques were used to design efficient GT decoders. Results and surveys for early non-adaptive decoding algorithms were given in [32]–[34]. Moreover, although most of the works described above mainly targeted fundamental limits, some give efficient algorithms as well. In the context of this work, it is important to mention the recent COMP [35], DD and SCOMP [36] algorithms, concepts from which we will use herein. 1 A simple information theoretic argument explains a lower bound. There N N are K defectives out of N items, hence K possibilities to cover: log K bits of information. Since each test carries at most one bit, this is the amount of tests required. Stirling’s approximation easily shows that for K  N , the leading factor of that is K log(N/K). B. Secure communication It is very important to note that making GT secure is different from making communication secure, as remarked in Section I. Now, we briefly survey the literature in secure communication, since many of the ideas/models/primitives in secure communication will have analogues in secure grouptesting. 1) Information-theoretic secrecy: In a secure communication setting, transmitter Alice wishes to send a message m to receiver Bob. To do so, she is allowed to encode m into a (potentially random) function x = f (m), and transmit x over a medium. It is desired that the eavesdropper Eve should glean no information about m from its (potentially noisy) observation z. This information leakage is typically measured via the mutual information between m and z. The receiver Bob should be able to reconstruct m based on its (also potentially noisy) observation of x (and, potentially, a shared secret that both Bob and Alice know, but Eve is ignorant of). There are a variety of schemes in the literature for information-theoretically secure communications.2 Such schemes typically make one of several assumptions (or combinations of these): • Shared secrets/Common randomness/Symmetric-key encryption: The first scheme guaranteed to provide information-theoretic secrecy was by [37], who analyzed the secrecy of one-time pad schemes and showed that they ensure perfect secrecy (no leakage of transmitted message). He also provided lower bounds on the size of this shared key. The primary disadvantage of such schemes is that they require a shared key that is essentially as large as the amount of information to be conveyed, and it be continually refreshed for each new communication. These requirements typically make such schemes untenable in practice. • Wiretap secrecy/Physical-layer secrecy: Wyner et al. [19], [38] first considered certain communication models in which the communication channel from Alice to Eve is a degraded (noisier) version of the channel from Alice to Bob, and derived the information-theoretic capacity for communication in such settings. These results have been generalized in a variety of directions. See [18], [39], [40] for (relatively) recent results. The primary disadvantage of such schemes is that they require that it be possible to instantiate communication channels from Alice to Bob that are better than the communication channel from Alice to Eve. Further, they require that the channel parameters of both channels be relatively well known to Alice and Bob, since the choice of communication rate depends on these parameters. These assumptions make such schemes also untenable in practice, since on one hand Eve may deliberately situate herself to have 2 Security in general has many connotations for instance, in the informationtheory literature it can also mean a scheme that is resilient to an active adversary, for instance a communication scheme that is resilient to jamming against a malicious jammer. In this work we focus our attention on passive eavesdropping adversaries, and aim to ensure secrecy of communications visa-vis such adversaries. We shall thus henceforth use the terms security and secrecy interchangeably. 4 • a relatively clear view of Alice’s transmission than Bob, and on the other hand there are often no clear physicallymotivated reasons for Alice and Bob to know the channel parameters of the channel to Eve. Public discussion/Public feedback: A very nice result by Maurer ( [41] and subsequent work - see [18] for details) significantly alleviated at least one of the charges level against physical-layer security systems, that they required the channel to Bob to be “better” than the channel to Eve. Maurer demonstrated that feedback (even public feedback that is noiselessly observable by Eve) and multiround communication schemes can allow for informationtheoretically secure communication from Alice to Bob even if the channel from Alice to Bob is worse than the channel from Alice to Eve. Nonetheless, such public discussion schemes still require some level of knowledge of the channel parameters of the channel to Eve. 2) Cryptographic security: Due to the shortcomings highlighted above, modern communication systems usually back off from demanding information-theoretic security, and instead attempt to instantiate computational security. In these settings, instead of demanding small information leakage to arbitrary eavesdroppers, one instead assumes bounds on the computational power of the eavesdropper (for instance, that it cannot computationally efficiently invert “one-way functions”). Under such assumptions one is then often able to provide conditional security, for instance with a public-key infrastructure [42], [43]. Such schemes have their own challenges to instantiate. For one, the computational assumptions they rely on are sometimes unfounded and hence sometimes turn out to be prone to attack [17], [18], [44]. For another, the computational burden of implementing cryptographic primitives with strong guarantees can be somewhat high for Alice and Bob [45]. C. Secure Group-Testing On the face of it, the connection between secure communication and secure group-testing is perhaps not obvious. We highlight below scenarios that make these connections explicit. Paralleling the classification of secure communication schemes above, one can also conceive of a corresponding classification of secure GT schemes. 1) Information-theoretic schemes: • Shared secrets/Common randomness/Symmetric-key encryption: A possible scheme to achieve secure group testing, is to utilize a shared key between Alice and Bob. For example, consider a scenario in which Alice the nurse has a large number of blood samples that need to be tested for the presence of a disease. She sends them to a lab named Eve to be tested. To minimize the number of tests done via the lab, she pools blood samples appropriately. However, while the lab itself will perform the tests honestly, it can’t be trusted to keep medical records secure, and so Alice keeps secret the identity of the people tested in each pool. 3 3 Even in this setting, it can be seen that the number of diseased individuals can still be inferred by Eve. However, this is assumed to be a publicly known/estimable parameter. • Given the test outcomes, doctor Bob now desires to identify the set of diseased people. To be able to reconstruct this mapping, a relatively large amount of information (the mapping between individuals’ identities and pools tested) needs to be securely communicated from Alice to Bob. As in the one-time pad secure communication setting, this need for a large amount of common randomness makes such schemes unattractive in practice. Nonetheless, the question is theoretically interesting, and some interesting results have been recently reported in this direction by [13]–[16]. Wiretap secrecy/Physical-layer secrecy: This is the setting of this paper. Alice does not desire to communicate a large shared key to Bob, and still wishes to maintain secrecy of the identities of the diseased people from “honest but curious” Eve. Alice therefore does the following two things: (i) For some δ ∈ (0, 1), she chooses a 1/δ number of independent labs, and divides the T pools to be tested into 1/δ pool sets of T δ pools each, and sends each set to a distinct lab. (ii) For each blood pool, she publicly reveals to all parties (Bob, Eve, and anyone else who’s interested) a set S(t) of possible combinations of individuals whose blood could constitute that specific pool t. As to which specific combination from S(t) of individuals the pool actually comprises of, only Alice knows a priori - Alice generates this private randomness by herself, and does not leak it to anyone (perhaps by destroying all trace of it from her records). The twin-fold goal is now for Alice to choose pool-sets and set of S(t) for each t ∈ [T ] to ensure that as long as no more than one lab leaks information, there is sufficient randomness in the set of S(t) so that essentially no information about the diseased individuals identities leaks, but Bob (who has access to the test reports from all the 1/δ labs) can still accurately estimate (using the publicly available information on S(t) for each test t) the disease status of each individual. This scenario closely parallels the scenario in Wyner’s Wiretap channel. Specifically, this corresponds to Alice communicating a sequence of T test outcomes to Bob, whereas Eve can see only a δ fraction of test outcomes. To ensure secrecy, Alice injects private randomness (corresponding to which set from S(t) corresponds to the combination of individuals that was tested in test t) into each test - this is the analogue of the coding schemes often used for Wyner’s wiretap channels. Remark 1. It is a natural theoretical question to consider corresponding generalizations of this scenario with other types of broadcast channels from Alice to Bob/Eve (not just degraded erasure channels), especially since such problems are well-understood in a wiretap security context. However, the physical motivation of such generalizations is not as clear as in the scenario outlined above. So, even though in principle the schemes we present in Section III can be generalized to other broadcast channels, to keep the presentation in this paper clean we do not pursue these generalizations here. 5 • Remark 2. Note that there are other mechanisms via which Alice could use her private randomness. For instance, she could deliberately contaminate some fraction of the tests she sends to each lab with blood from diseased individuals. Doing so might reduce the amount of private randomness required to instantiate secrecy. While this is an intriguing direction for future work, we do not pursue such ideas here. Public discussion/Public feedback: The analogue of a public discussion communication scheme in the secure group-testing context is perhaps a setting in which Alice sends blood pools to labs in multiple rounds, also known as adaptive group testing in the GT literature. Bob, on observing the set of test outcomes in round i, then publicly broadcasts (to Alice, Eve, and any other interested parties) some (possibly randomized) function of his observations thus far. This has several potential advantages. Firstly, adaptive group-testing schemes (e.g. [36]) significantly outperform the best-known nonadaptive group-testing schemes (in terms of smaller number of tests required to identify diseased individuals) in regimes where K = ω(N 1/3 ). One can hope for similar gains here. Secondly, as in secure communication with public discussion, one can hope that multi-round GT schemes would enable information-theoretic secrecy even in situations where Eve may potentially have access to more test outcomes than Bob. Finally, such schemes may offer storage/computational complexity advantages over non-adaptive GT schemes. Hence this is an ongoing area of research, but outside the scope of this paper. 2) Cryptographic secrecy: As in the context of secure communication, the use of cryptographic primitives to keep information about the items being tested secure has also been explored in sparse recovery problems - see, for instance [13]–[16]. Schemes based on cryptographic primitives have similar weaknesses in the secure GT context as they do in the communication context, and we do not explore them here. Throughout the paper, we use boldface to denote matrices, capital letters to denote random variables, lower case letters to denote their realizations, and calligraphic letters to denote the alphabet. Logarithms are in base 2 and hb (·) denotes the binary entropy function. Figure 3 gives a graphical representation of the model. In general, and regardless of security constraints, non-adaptive GT is defined by a testing matrix T X = [X1T ; X2T ; . . . ; XN ] ∈ {0, 1}N ×T , where each row corresponds to a separate item j ∈ {1, . . . , N }, and each column corresponds to a separate pool test t ∈ {1, . . . , T }. For the j-th item, XjT = {Xj (1), . . . , Xj (T )} is a binary row vector, with the t-th entry Xj (t) = 1 if and only if item j participates in the t-th test. If Aj ∈ {0, 1} denotes an indicator function for the j-th item, determining whether it belongs to the defective set, i.e., Aj = 1 if j ∈ K and Aj = 0 otherwise, the (binary) outcome of the t ∈ {1, . . . , T } pool test Y (t) equals Y (t) = N _ Xj (t)Aj = j=1 _ Xd (t), d∈K W where is used to denote the boolean OR operation. In SGT, we assume an eavesdropper who observes a noisy vector Z T = {Z(1), . . . , Z(T )}, generated from the outcome vector Y T . In the erasure case considered in the work, the probability of erasure is 1 − δ, i.i.d. for each test. That is, on average, T δ outcomes are not erased and are accessible to the eavesdropper via Z T . Therefore, in the erasure case, if Bt ∈ {1, ?} is an erasure indicator function for the t-th pool test, i.e., Bt = 1 with probability δ, and Bt =? with probability 1 − δ, the eavesdropper observes   N _ Z(t) = Y (t)Bt =  Xj (t)Aj  Bt , t = 1, . . . , T. j=1 III. P ROBLEM F ORMULATION In SGT, a legitimate user desires to identify a small unknown subset K of defective items from a larger set N , while minimizing the number of measurements T and keeping the eavesdropper, which is able to observe a subset of the tests results, ignorant regarding the status of the N items. Let N = |N |, K = |K| denote the total number of items, and the number of defective items, respectively. As formally defined below, the legitimate user should (with high probability) be able to correctly estimate the set K; on the other hand, from the eavesdroppers perspective, this set should be “almost”  N uniformly distributed over all possible K sets. We assume that the number K of defective items in K is known a priori to all parties - this is a common assumption in the GT literature [3].4 4 If this is not the case, [46], [47] give methods/bounds on how to “probably approximately” correctly learn the value of K in a single stage with O(log N ) tests.  N } the index of the subset Denote by W ∈ W , {1, . . . , K of defective items. We assume W is uniformly distributed, that is, there is no a priori bias towards any specific subset.5 Further, denote by Ŵ (Y T ) the index recovered by the legitimate decoder, after observing Y T . In this work, we assume that the mixer may use a randomized testing matrix. In this case, the random bits used are know only to the mixer, and are not assumed to be shared with the decoder. In other words, the “codebook” which consists of all possible testing matrices is known to all parties, Alice, Bob and Eve. However, if the mixer choose a specific X, the random value is not shared with Bob or Eve. We refer to the codebook consisting of all 5 This is a common probabilistic model for the set of defectives in grouptesting. Another model, called Probabilistic Group Testing, assumes that items are defective with probability K/N . Yet another model assumes that any set of size at most K (rather than exactly K) instantiates with equal probability. In many group-testing scenarios results for one probabilistic model for the set of defectives can be translated over to other scenarios, so we focus on the model presented above, where exactly K items are defective. 6 Figure 3: Noiseless non-adaptive secure group-testing setup. possible matrices, together with the decoder at Bob’s side as SGT algorithm. As we are interested in the asymptotic behavior, i.e., in “capacity style” results, with a focus on the number of tests T (as a function of N and K) required to guarantee high probability of recovery as the number of items N grows without bound. For simplicity, in the first part of this work, we focus primarily on the regime where K is a constant independent of N . In Section VII, we give an algorithm which applies to any K.6 The following definition lays out the goals of SGT algorithms. Definition 1. A sequence of SGT algorithms with parameters N, K and T is asymptotically (in N) reliable and weakly or strongly secure if, (1) Reliable: The probability (over the index W ) of incorrect reconstruction of W at the legitimate receiver converges to zero. That is, lim P (Ŵ (Y T ) 6= W ) = 0. N →∞ (2) Weakly secure: One potential security goal is so-called weak information-theoretic security against eavesdropping. Specifically, if the eavesdropper observes Z T , a scheme is said to be weakly secure if 1 I(W ; Z T ) = 0. T →∞ T lim (3) Strongly secure: A stronger notion of security is so-called strong information-theoretic security against eavesdropping. Specifically, if the eavesdropper observes Z T , a scheme is said to be strongly secure if lim I(W ; Z T ) = 0. T →∞ Remark 3. Note that strong security implies that in the limit the distribution over Z T is essentially statistically independent of the distribution over W . Specifically, the KL divergence between pZ T ,W and pZ T pW converges to 0. Remark 4. While weak security is a much weaker notation of security against eavesdropping than strong security, and indeed is implied by strong security, nonetheless we consider it in this work for the following reason. Our impossibility result will show that even guaranteeing weak security requires at least a 6 Following the lead of [27], in principle, many of our results in this section as well can be extended to the regime where K = o(N 1/3 ), but for ease of presentation we do not do so here). certain number of tests, and our achievability results will show that essentially the same number of tests suffices to guarantee strong security. Hence both our impossibility and achievability results are with regard to the corresponding “harder to prove” notion of security. To conclude, the goal in this work is to design (for parameters N and K) an N × T measurement matrix (which is possibly randomized) and a decoding algorithm W (Y T ), such that on observing Y T , the legitimate decoder can (with high probability over W ) identify the subset of defective items, and yet, on observing Z T , the eavesdropper learns essentially nothing about the set of defective items. IV. M AIN R ESULTS Under the model definition given in Section III, our main results are the following sufficiency (direct) and necessity (converse) conditions, characterizing the maximal number of tests required to guarantee both reliability and security. The proofs are deferred to Section V and Section VI. A. Direct (Sufficiency) The sufficiency part is given by the following theorem. Theorem 1. Assume a SGT model with N items, out of which K = O(1) are defective. For any 0 ≤ δ < 1, if   1+εK N −K log , (1) T ≥ max i=1,...,K 1 − δ i i for some ε>0 independent of N and K, then there exists a sequence of SGT algorithms which are reliable and secure. That is, as N → ∞, both the average error probability approaches zero exponentially and an eavesdropper with leakage probability δ is kept ignorant. The construction of the SGT algorithm, together with the proofs of reliability and secrecy are deferred to Section V. In fact, in Section V we actually prove that the maximal error probability decays to 0. However, a few important remarks are in order now. First, rearranging terms in eq. (1), we have   1 (1 + ε)K N −K T ≥ max log . 1 − δ i=1,...,K i i That is, compared to only a reliability constraint, the number of tests required for both reliability and secrecy is increased 1 by the multiplicative factor 1−δ , where, again, δ is the leakage probability at the eavesdropper. 7 The result given in Theorem 1 uses an ML decoding at the legitimate receiver. The complexity burden in ML, however, prohibits the use of this result for large N. In Theorem 3, we suggest an efficient decoding algorithm, which maintains the reliability and the secrecy results using a much simpler decoding rule, at the price of only slightly  more tests. Using an upper bound on log N −K , the maximization in i Theorem 1 can be solved easily, leading to simple bound on T with tight scaling and only a moderate constant. Corollary 1. For SGT with parameters K << N and T , reliability and secrecy can be maintained if 1+ε K log(N − K)e. 1−δ  Proof. Substituting log N −K ≤ i log (N −K)e , the maximum i i over i is easily solved. T ≥ Note that together with the converse below, this suggests  K log N T =Θ , and, a Θ (K log N ) result for δ bounded 1−δ away from 1. D. Efficient Algorithms Under the SGT model definition given in Section III, we further consider a computationally efficient algorithm at the legitimate decoder. Specifically, we analyze the Definite NonDefective (DND) algorithm (originally called Combinatorial Orthogonal Matching Pursuit (COMP)), considered for the non-secure GT model in the literature [35], [36]. The theorem below states that indeed efficient decoding (with arbitrarily small error probability) and secrecy are possible, at the price of a higher T . Interestingly, the theorem applies to any K, and not necessarily only to K = O(1). This is, on top of the reduced complexity, an important benefit of the suggested algorithm. Theorem 3. Assume a SGT model with N items,  out of which K are defective. Then, for any δ < 21 1 − lnK2 , there exists an efficient decoding algorithm, requiring O(N 2 T ) operations, such that if the number of tests satisfies T ≥ 1 2 (1 1+ K log N − lnK2 ) − δ its error probability is upper bounded by B. Converse (Necessity) The necessity part is given by the following theorem. Theorem 2. Let T be the minimum number of tests necessary to identify a defective set of cardinality K among population of size N while keeping an eavesdropper, with a leakage probability δ < 1, ignorant regarding the status of the items. Then, if T1 I(W ; Z T ) < , one must have:   N 1 − T T ≥ log , K 1−δ Pe ≤ N − . The construction of the DND GT algorithm, together with the proofs of reliability and secrecy are deferred to Section VII. Clearly, the benefits of the algorithm above come at the price of additional tests and a smaller range of δ it can handle. V. C ODE C ONSTRUCTION AND A P ROOF FOR T HEOREM 1 Returning to the analogy in [21] between channel capacity and group testing, one might define by  Cs the (asymptotic) N minimal threshold value for log K /T , above which no reliable and secure scheme is possible. Under this definition, the result in this paper show that Cs = (1 − δ)C, where C is the capacity without the security constraint. Clearly, this can be written as Cs = C − δC, In order to keep the eavesdropper, which obtains only a fraction δ of the outcomes, ignorant regarding the status of the items, we randomly map the items to the tests. Specifically, as depicted in Figure 4, for each item we generate a bin, containing several rows. The number of such rows corresponds to the number of tests that the eavesdropper can obtain, yet, unlike wiretap channels, it is not identical to Eve’s capacity, and should be normalized by the number of defective items. Then, for the j-th item, we randomly select a row from the j-th bin. This row will determine in which tests the item will participate. In order to rigorously describe the construction of the matrices and bins, determine the exact values of the parameters (e.g., bin size), and analyze the reliability and secrecy, we first briefly review the representation of the GT problem as a channel coding problem [20], together with the components required for SGT.  N A SGT code consists of an index set W = {1, 2, . . . K }, its w-th item corresponding to the w-th subset K ⊂ {1, . . . , N }; A discrete memoryless source of randomness (R, pR ), with known alphabet R and known statistics pR ; An encoder, f : W × R → XSw ∈ {0, 1}K×T raising the usual interpretation as the difference between the capacity to the legitimate decoder and that to the eavesdropper [18]. Note that as the effective number of tests Eve sees is Te = δT , her GT capacity is δC. which maps the index W of the defective items to a matrix XTSw of codewords, each of its rows corresponding to a different item in the index set Sw , w ∈ W, |Sw | = K. The need for a stochastic encoder is similar to most encoders ensuring where T =  + ˜T , with ˜T → 0 as T → ∞. The lower bound is derived using Fano’s inequality to address reliability, assuming a negligible mutual information at the eavesdropper, thus keeping an eavesdropper with leakage probability δ ignorant, and information inequalities bounding the rate of the message on the one hand, and the data Eve does not see on the other. Compared with the lower bound without security constraints, it is increased by the multiplicative factor 1 1−δ . C. Secrecy capacity in SGT 8 Random selection of Test Vector per Item Bin Matrixes of code-words 11000……...0...0101 1 01001……...1...1001 2 M M N-Items 00011……...0...1101 M 1 2 M N 01101……...1...1011 1 10101……...1...0011 2 Secure Group testing Matrix 01001……...1...1001 10001……...0...0110 M M M 10001……...0...0110 M M M M ⇒ 10110……...1...0100 1 11001……...1...1100 2 10110……...1...0100 T N YSecure (t ) = ∨ X j (t ) Aj , 00110……...0...0001 M T j =1 t = 1KT Figure 4: Binning and encoding process for a SGT code. information theoretic security, as randomness is required to confuse the eavesdropper about the actual information [18]. Hence, we define by RK the random variable encompassing the randomness required for the K defective items, and by M the number of rows in each bin. Clearly, M K = H(RK ). At this point, an important clarification is in order. The lab, of course, does not know which items are defective. Thus, operationally, it needs to select a row for each item. However, in the analysis, since only the defective items affect the output (that is, only their rows are ORed together to give Y T ), we refer to the “message” as the index of the defective set w and refer only to the random variable RK required to choose the rows in their bins. In other words, unlike the analogous communication problem, in GT, nature performs the actual mapping from W to XTSw . The mixer only mixes the blood samples according to the (random in this case) testing matrix it has. A decoder at the legitimate user is a map Ŵ : Y T → W. The probability of error is P (Ŵ (Y T ) 6= W ). The probability that an outcome test leaks to the eavesdropper is δ. We assume a memoryless model, i.e., each outcome Y (t) depends only on the corresponding input XSw (t), and the eavesdropper observes Z(t), generated from Y (t) according to p(Y T , Z T |XSw ) = T Y 2) Testing: For each item j, the mixer/lab selects uniformly at random one codeword xT (m) from the j-th bin. Therefore, the SGT matrix contains N randomly selected codewords of length T , one for each item, defective or not. Amongst is an unknown subset XSTw , with the index w representing the true defective items. An entry of the j-th random codeword is 1 if the j-item is a member of the designated pool test and 0 otherwise. 3) Decoding at the Legitimate Receiver: The decoder looks for a collection of K codewords XSTŵ , one from each bin, for which Y T is most likely. Namely, p(Y (t)|XSw (t))p(Z(t)|Y (t)). P (Y T |XSTŵ ) > P (Y T |XSTw ), ∀w 6= ŵ. Then, the legitimate user (Bob) declares Ŵ (Y T ) as the set of bins in which the rows ŵ reside. A. Reliability Let (S 1 , S 2 ) denote a partition of the defective set S into disjoint sets S 1 and S 2 , with cardinalities i and K − i, respectively.7 Let I(XS 1 ; XS 2 , Y ) denote the mutual information between XS 1 and XS 2 , Y , under the i.i.d. distribution with which the codebook was generated and remembering that Y is the output of a Boolean channel. The following lemma is a key step in proving the reliability of the decoding algorithm. Lemma 1. If the number of tests satisfies  i log N −K M i , T ≥ (1 + ε) · max i=1,...,K I(XS 1 ; XS 2 , Y ) then, under the codebook above, as N → ∞ the average error probability approaches zero. Next, we prove Lemma 1, which extends the results in [20] to the codebook required for SGT. Specifically, to obtain a bound on the required number of tests as given in Lemma 1, we first state Lemma 2, which bounds the error probability of the ML decoder using a Gallager-type bound [48]. Definition 2. The probability of error event Ei in the ML decoder defined, as the event of mistaking the true set for a set which differs from it in exactly i items. Lemma 2. The error probability P (Ei ) is bounded by −T P (Ei ) ≤ 2 log Eo (ρ)−ρ K Mi log ( ) (N −K i ) i − −K T T T ! , t=1 We may now turn to the detailed construction and analysis. 1) Codebook Generation: Choose M such that log2 (M ) = T (δ − 0 )/(K) where the error exponent Eo (ρ) is given by " X X X Eo (ρ) = − log Y ∈{0,1} XS 2 ∈{0,1} P (XS 1 ) XS 1 ∈{0,1} #1+ρ 0 0 for some  > 0.  will QT affect the equivocation. Using a distribution P (X T ) = i=1 P (xi ), for each item generate M independent and identically distributed codewords xT (m), 1 ≤ m ≤ M . The codebook is depicted in the left hand side of Figure 4. Reveal the codebook to Alice and Bob. We assume Eve may have the codebook as well. p(Y, XS 2 |XS 1 ) 1 1+ρ , 0 ≤ ρ ≤ 1. 7 This partition helps decompose the error events into classes, where in class i one already knows K − i defective items, and the dominant error event corresponds to missing the other i. Thus, it is easier to present the error event as one “codeword” against another. See Appendix A for the details. 9 Proof of Lemma 1. For this lemma, we follow the derivation in [20]. However, due the different code construction, the details are different. Specially, for each item there is a bin of codewords, from which the decoder has to choose. Define   i log Ki log N −K M K i − − . f(ρ) = Eo (ρ) − ρ T T T We wish to show that T f(ρ) → ∞ as N → ∞. Note that log Ki is a constant for the fixed K regime. Thus for large T we have limT →∞ f(0) = 0. Since the function f(ρ) is differentiable and has a power series expansion, for a sufficiently small α, by Taylor series expansion in the neighborhood of ρ ∈ [0, α] we have f(ρ) = f(0) + ρ df |ρ=0 +Θ(ρ2 ). dρ Now, ∂Eo |ρ=0 ∂ρ XX X = [ P (XS 1 )p(Y, XS 2 |XS 1 ) log p(Y, XS 2 |XS 1 ) Y − XS 2 XS 1 X P (XS 1 )p(Y, XS 2 |XS 1 ) XS 1 = I(Xs1;Xs2,Y) 0.8 XS 1 (1 + ε) log N −K i  Mi < I(XS 1 ; XS 2 , Y ) T for some constant ε > 0, the exponent is positive for large enough T and we have P (Ei ) → 0 as T → ∞ for ρ > 0. Using a union bound one can show that taking the maximum over i will ensure a small error probability in total. 10 20 30 40 50 1 <= i <= K 60 70 80 90 100 90 100 Figure 5: Mutual Information Bound 1800 Necessary T with h (e-1) constant, N = 40000 b 1600 Proof of Claim 1. First, note that Necessary T without h (e-1) constant, N = 40000 b 1400 (a) Sufficient T with hb(e-1) constant, N = 40000 -1 I(X XS 2hb)(e+ I(XSN1=;40000 Y |XS 2 ) Sufficient ) constant, S 1 ; XS 2 , Y ) = I(X S 1T ;without 1200 1000 = H(Y |XS 2 ) − H(Y |XS ) 0 10 20 (b) 30 40 50 60 70 80 = q K−i H(q1i<= ) i <= K    1 1 = q K−i q i log i + 1 − q i log , q (1 − q i ) where equality (a) follows since the rows of the testing matrix are independent, and (b) follows since H(Y |XS ) is the uncertainty of the legitimate receiver given XS , thus when observing the noiseless outcomes of all pool tests, this uncertainty is zero. Also, note that the testing matrix is random and i.i.d. with distribution (1 − q, q), hence the probability for i zeros is q i . Then, under a fixed input distribution for the testing matrix ln(2) (p = ln(2) K , q = 1 − K ) and large K it is easy to verify that the bounds meet at the two endpoint of i = 1 and i = K, yet the mutual information is concave in i thus the bound is obtained. This is demonstrated graphically in Figure 5. Applying Claim 1 to the expression in Lemma 1, we have  i  i log N −K M log N −K M i i ≤ . i I(XS 1 ; XS 2 , Y ) K P (XS 1 )p(Y, XS 2 |XS 1 ) Hence, if 0 2000 P (XS 1 )p(Y, XS 2 |XS 1 )] p(Y, XS 2 |XS 1 ) = I(XS 1 ; XS 2 , Y ). P (XS 1 )p(Y, XS 2 |XS 1 ) i/K I(Xs1;Xs2,Y) 0.4 0 XS 2 XS 1 log P 0.6 0.2 XS 1 XXX Y X 1 Number of tests In Appendix A we analyze the bound provided in Lemma 2. Note that there are two main differences  Kcompared to nonN secured GT. First, the decoder has M possible subsets K  N of codewords to choose from, K for the number of possible bins and M K for the number of possible rows to take in each bin. Thus, when fixing the error event, there are N −K Mi i subsets to confuse the decoder. Moreover, due to the bin structure of the code, there are also many “wrong” codewords which are not the one transmitted on the channel, hence create a decoding error codeword-wise, yet the actual bin decoded may still be the right one. δ−K Hence, substituting M = 2T K , a sufficient condition for reliability is     1+ε N −K i T ≥ max log + T (δ −  ) K i 1≤i≤K i K K Rearranging terms results in T ≥ max 1≤i≤K   1 1+ε N −K log , 1 − δ + K − εδ + εK i/K i The expression I(XS 1 ; XS 2 , Y ) in Lemma 1 is critical to understand how many tests are required, yet it is not a function of the problem parameters in any straight forward mannar. We now bound it to get a better handle on T . where by reducing K and εK we increase the bound on T , and with some constant ε > 0. Noting that this is for large K and N , and that ε is independent of them, achieves the bound on T provided in Theorem 1 and reliability is established. Claim 1. For large K, and under a fixed input distribution ln(2) for the testing matrix ( ln(2) K , 1 − K ), the mutual information between XS 1 and (XS 2 , Y ) is lower bounded by B. Information Leakage at the Eavesdropper I(XS 1 ; XS 2 , Y ) ≥ i . K We now prove the security constraint is met. Hence, we wish to show that I(W ; Z T )/T → 0, as T → ∞. Denote by 10 CT the random codebook and by XST the set of codewords corresponding to the true defective items. We have,  1 1 T T T T I(W ; Z |CT ) = T I(W, RK ; Z |CT ) − I(RK ; Z |W, CT )  (a) 1 = T I(XST ; Z T |CT ) − I(RK ; Z T |W, CT ) = T1 (I(XST ; Z T |CT ) − H(RK |W, CT ) + H(RK |Z T , W, CT ))  (b) 1 = T I(XST ; Z T |CT ) − H(RK ) + H(RK |Z T , W, CT ) 1 1 (c) =I(XS ; Z|CT ) − K log M + H(RK |Z T , W, CT ) T  T  (d) (e) δ − 0 1 1 ≤δ − K T + H(RK |Z T , W, CT ) ≤ T + 0 , T K T where T → 0 as T → ∞. (a) is since there is a 1 : 1 correspondence between (W, RK ) and XST ; (b) is since RK is independent of W and CT ; (c) is since in this direct result the codebook is defined by the construction and is memoryless, as well as the channel; (d) is since by choosing an i.i.d. distribution for the codebook one easily observes that I(XS ; Z|CT ) ≤ δ. Finally, (e) is for the following reason: Given W and the codebook, Eve has a perfect knowledge regarding the bins from which the codewords were selected. It requires to see whether she can indeed estimate RK . Note that the channel Eve sees in this case is the following multiple access channel: each of the defective items can be considered as a “user” with M messages to transmit. Eve’s goal is to decode the messages from all users. This is possible if the rates at which the users transmit are within the capacity region of this MAC. Indeed, this is a (binary) Boolean MAC channel, followed by a simple erasure channel. The sum capacity cannot be larger than δ, and this sum capacity is easily achieved by letting one user transmit at a time, or, in our case, where the codebook is randomly i.i.d. distributed, under a fixed 1 input distribution p = 1 − 2 K . Since actually, we use input ln(2) distribution for the testing matrix of ( ln(2) K , 1 − K ), and 1 ln(2) limK→∞ 1 − 2 K / K = 1, for large K, each user obtain the same capacity; namely, each user sees a capacity of δ/K. Remark 5. Under non-secure GT, it is clear that simply adding tests to a given GT code (increasing T ) can only improve the performance of the code (in terms of reliability). A legitimate decoder can always disregard the added tests. For SGT, however, the situation is different. Simply adding tests to a given code, while fixing the bin sizes, might make the vector of results vulnerable to eavesdropping. In order to increase reliability, one should, of course, increase T , but also increase the bin sizes proportionally, so the secrecy result above will still hold. This will be true for the efficient algorithm suggested in Section VII as well. Remark 6. To establish the weak secrecy constraint we set δ−0 M to be 2T K , where the readability is archived without any constraint on 0 . However, in Appendix B to establish the 0 T δ+ strong secrecy constraint we require M to be 2 K . Remark 7. Note that since 1 1 1 H(RK |Z T , W, CT ) = H(RK ) − I(RK ; Z T , W, CT ) T T T 1 1 = log M − I(RK ; Z T , W, CT ), T T any finite-length approximation for I(RK ; Z T , W, CT ) will give a finite length approximation for the leakage at the eavesdropper. For example, one can use the results in [49], to show that the leakage can be approximated as √1T + 0 . VI. C ONVERSE (N ECESSITY ) In this section, we derive the necessity bound on the required number of tests. Let Z̄ denote the random variable corresponding to the tests which are not available to the eavesdropper. Hence, Y = (Z, Z̄). By Fano’s inequality, if Pe → 0, we have H(W |Y ) ≤ T 0T , where 0T → 0 as T → ∞. Moreover, the secrecy constraint implies I(W ; Z) ≤ T 00T , (2) where 00T → 0 as T → ∞. Consequently,   N log = H(W ) K = I(W ; Y ) + H(W |Y ) (a) ≤ I(W ; Z, Z̄) + T 0T = I(W ; Z) + I(W ; Z̄|Z) + T 0T (b) ≤ (c) I(W ; Z̄|Z) + T 0T + T 00T ≤ I(XSw ; Z̄|Z) + T (0T + 00T ) = H(Z̄|Z) − H(Z̄|XSw , Z) + T (0T + 00T ) (d) ≤ H(Z̄) + T (0T + 00T ) where (a) follows from Fano’s inequality and since Y = (Z, Z̄), (b) follows from (2), (c) follows from the Markov chain W → XSTw → Y → Z and (d) is since conditioning reduces entropy. We now evaluate H(Z̄). Denote by Ē the set of tests which are not available to Eve and by Ēγ the event {|Ē| ≤ T (1 − δ)(1 + γ)} for some γ > 0. We have H(Z̄) = P (Ēγ )H(Z̄|Ēγ ) + P (Ēγc )H(Z̄|Ēγc ) ≤ T (1 − δ)(1 + γ) + T P (Ēγc ) ≤ T (1 − δ)(1 + γ) + T 2−T (1−δ)f (γ) , where the last inequality follows from the Chernoff bound for i.i.d. Bernoulli random variables with parameter (1 − δ) and is true for some f (γ) such that f (γ) > 0 for any γ > 0. Thus, we have   N ≤ T (1 − δ)(1 + γ) + T 2−T (1−δ)f (γ) + T (0T + 00T ). log K That is,   N 1 − T log , T ≥ 1−δ K for some T such that T → 0 as T → ∞. This completes the converse proof. 11 VII. E FFICIENT A LGORITHMS defective items. As a result, we have Pe The achievability result given in Theorem 1 uses a random codebook and ML decoding at the legitimate party. The complexity burden in ML, however, prohibits the use of this result for large N . In this section, we derive and analyze an efficient decoding algorithm, which maintains the reliability result using a much simpler decoding rule, at the price of only slightly more tests. The secrecy constraint, as will be clear, is maintained by construction, as the codebook and mixing process do not change compared to the achievability result given before. Moreover, the result in this section will hold for any K, including even the case were K grows linearly with N. Specifically, we assume the same codebook generation and the testing procedure given in Section V, and analyze the Definite Non-Defective (DND) algorithm, previously considered for the non-secure GT in the literature [35], [36]. The decoding algorithm at the legitimate user is as follows. Bob attempts to match the rows of X with the outcome vector Y T . If a particular row j of X has the property that all locations t where it has 1, also corresponds to a 1 in Y (t), then that row can correspond to a defective item. If, however, the row has 1 at a location t where the output has 0, then it is not possible that the row corresponds to a defective item. The problem, however, when considering the code construction in this paper for SGT, is that the decoder does not know which row from each bin was selected for any given item. Thus, it takes a conservative approach, and declares an item as defective if at least one of the rows in its bin signals it may be so. An item is not defective only if all the rows in its bin prevent it from being so. It is clear that this decoding procedure has no false negatives, as a defective item will always be detected. It may have, though, false positives. A false positive may occur if all locations with ones in a row corresponding to a non-defective item are hidden by the ones of other rows corresponding to defective items and selected by the mixer. To calculate the error probability, fix a row of X corresponding to a nondefective item (a row in its bin). Let j1 ; . . . ; jk index the rows of X corresponding to the K defective items, and selected by the mixer for these items (that is, the rows which were actually added by the Boolean channel). An error event associated with the fixed row occurs if at any test where that row has a 1, at least one of the entries Xj1 (t), . . . , Xjk (t) also has a 1. The probability for this to happen, per column, is p(1 − (1 − p)K ). Hence, the probability that a test result in a fixed row is hidden from the decoder, in the sense that it cannot be declared as non defective due to a specific column, is p(1 − (1 − p)K ) + (1 − p) = 1 − p(1 − p)K . Since this should happen for all T columns, T the error probability for a fixed row is 1 − p(1 − p)K . Now, to compute the error probability for the entire procedure we take a union bound over all M (N − K) rows corresponding to non- ≤ (a) ≤ = (b) ≤ (c) ≤ ≤ T M (N − K) 1 − p(1 − p)K βK log N  y K y  1− MN 1 − K K βK log N   y K−1  y y 1− 1− MN 1 − K K K   β log N !  K y y 1− K   MN 1− Key −y y M N e−y(1− K )e β log N y −y 1 M N 1−y(1− K )e β ln 2 (d) = M N 1− 2 β(1− (e) = 2β(δ−) log N N 1− 2 β(1− = N β(δ−) N 1− 2 β(1− K ln 2 1 N 1−β ( 2 (1− K )−δ) . ≤ 1 ln 2 K ) 1 1 ln 2 ln 2 K ) ) (3) In the above, (a) follows by taking p = y/K and setting T as βK log N , for some positive y and β, to be defined. (b) follows since e−y ≤ (1 − y/n)n−1 for small y > 0 and any integer n > 0. In the sequence below, we will use it with y = ln 2, for which it is true. (c) follows since e−x ≥ (1 − x/n)n for x > 0 and any integer n > 0. (d) follows by choosing y = ln 2. (e) δ− is by setting M = 2T K and substituting the value for T . The result in (3) can be interpreted as follows. As long as δ, the leakage probability at the eavesdropper, is smaller than ln 2 1 2 (1 − K ), choosing T = βK log N with a large enough β results in an exponentially small error probability. For example, for large enough K and δ = 0.25, one needs β > 4, that is, about 4K log N tests to have an exponentially small (with N ) error probability while using an efficient decoding algorithm. To see the dependence of the error probability on the number of tests, denote     1 ln 2 =β 1− − δ − 1. 2 K Then, if the number of tests satisfies T ≥ 1 2 (1 1+ K log N − lnK2 ) − δ one has Pe ≤ N − . Thus, while the results under ML decoding (Theorem 1) show 1 that any value of δ < 1 is possible (with a 1−δ toll on T compared to non-secure GT), the analysis herein suggests that using the efficient algorithm, one can have a small error probability only for δ < 1/2, and the toll on T is greater 1 than 1 −δ . This is consistent with the fact that this algorithm 2 is known to achieve only half of the capacity for non-secure GT [36]. However, both these results may be due to coarse analysis, and not necessarily due to an inherent deficiency in the algorithm. Remark 8 (Complexity). It is easy to see that the algorithm runs over all rows in the codebook, and compares each one 12 Number of items N = 500, Number of defective items K = 3, Erasure Probability at Eve 1- delta = 0.1, Number of iterations = 8000 100 90 to any defective set Sw of size K, hence the probability of error computed decays exponentially for any “message W ”. Success (percentage) 80 70 60 50 40 Success (%) at the Receiver with DND decoder Success (%) at the Receiver with ML decoder Success (%) at Eve with DND decoder Success (%) at Eve with ML decoder DND bound on T ML lower bound on T ML upper bound on T 30 20 10 0 10 20 30 40 50 60 70 80 90 100 110 120 T - Number of pool tests Figure 6: Definite Non-Defective and ML simulation results. to the vector of tests’ results. The length of each row is T . δ There are N items, each having about 2 K T rows in its bin. Since T = O(K log N ), we have O(N 2 ) rows in total. Thus, the number of operations is O(N 2 T ) = O(KN 2 log N ). This should be compared to O(KN log N ) without any secrecy constraint. Figure 6 includes simulation results of the secure DND GT algorithm proposed, compared with ML decoding and the upper and lower bounds on the performance of ML. VIII. C ONCLUSIONS In this paper, we proposed a novel non-adaptive SGT algorithm, which with parameters N, K and T is asymptotically reliable and secure. Specifically, when the fraction of tests observed by Eve is 0 ≤ δ < 1, we prove that the number of tests required for both correct reconstruction at the legitimate user (with high probability) and negligible mutual information 1 times the number of tests required at Eve’s side is 1−δ with no secrecy constraint. We further provide sufficiency and necessity bounds on the number of tests required in the SGT model to obtains both, reliability and secrecy constraints. Moreover, we analyze in the proposed secure model, computationally efficient algorithms at the legitimate decoder, previously considered for the non-secure GT in the literature which identify the definitely non-defective items. A PPENDIX A E RROR P ROBABILITY B OUND In this section, we give a direct proof for Lemma 1. Specifically, we bound the error probability of the maximum likelihood decoder, and show that under the condition on T in the lemma, the error probability is indeed approaching 0 as N → ∞. Under the assumption that all items are equality likely to be defective, we consider the error probability given that the first K items are defective, that is, Sw=1 is the defective set. Denote this probability by Pe|1 . We have Pe|1 ≤ K X P (Ei ), i=1 where Ei is the event of a decoding error in which the decoder declares a defective set which differs from the true one (S1 ) in exactly i items. Note that the discussion below will apply In general, we follow the derivation in [20]. However, there is a key difference. In the code construction suggested in Section V, for each item there are M possible codewords (a “bin” of size M ). Only one of these codewords is selected by the mixer to decide in which pool tests the item will participate. Thus, when viewing this problem as a channel coding problem, if an item is defective, one and only one codeword out of its bin is actually transmitted (and summed with the codewords of the other K − 1 defective items). Since the decoder does not know which codewords were selected in each bin (the randomness is known only to the mixer), there are multiple error events to consider. E.g., events where the decoder choose the wrong codeword for some items, yet identified parts of the bins correctly, and, of course, events where the codeword selected was from a wrong bin. This complicates the error analysis significantly. Moreover, we wish to employ the correction suggested in [50], which results in a simpler yet stricter bound. Consider the event Ei . Clearly, Ei can be broken down into two disjoint events. The first is Ei and the event that the codewords selected for the correct K − i items are the true transmitted ones, and the second is the event of both Ei and the event that at least one of the codewords selected for the correct items is wrong. Denote the first event as Ei0 . Now, consider the case where we have Ei , that is, a correct decision on K − i items, yet, out of these K − i items, the decoder identified only j codewords right, 0 ≤ j < K − i, and for the rest, it identified a wrong codeword in the right bin. We claim that the probability for this event is exactly 0 ), as the decoder has to mistake X0 for XS1 , where P (EK−j both are collections of K codewords, and X0 shares exactly j codewords with XS1 . Thus, we have: Pe|1 ≤ = K X P (Ei ) i=1 K  X P (Ei0 ) + i=1 ≤ 2K K−i−1 X  j=0 K X   K −i−1 0 P (EK−j ) j P (Ei0 ). (4) i=1 The last inequality in (4) is loose, yet it suffices for the proof in the regime where K = O(1). A more delicate bounding technique might be required if K might grow with N . We now bound P (Ei0 ). Particularly, we will establish the following lemma. Lemma 3. The error probability P (Ei0 ) is bounded by −T P (Ei0 ) ≤ 2 log Eo (ρ)−ρ K log ( ) Mi (N −K i i ) − T T ! , 13 where the error exponent Eo (ρ) is given by " X X X Eo (ρ) = − log Y ∈{0,1} XS 2 ∈{0,1}  K There are K−i sets S1,w , and the summation does not depend on which set is it, hence, we get XX ρ iρ K P (Ei0 ) ≤ (N −K p1 (XS 1 ,Y T ) (i) i ) M P (XS 1 ) XS 1 ∈{0,1} Y T XS1 #1+ρ 1 1+ρ p(Y, XS 2 |XS 1 )  , 0 ≤ ρ ≤ 1. X  Proof. Denote by A = {w ∈ W : |S1c ,w | = i, |Sw | = K} the set of indices corresponding to sets of K items that differ from the true defective set S1 in exactly i items. S1c ,w for w ∈ W denotes the set of items which are in Sw but not in S1 . We have ≤ Pr[Ei0 |w0 = 1, XS1 , Y T ] X X pw (Y T ,XS1,w |XS1c ,w )s Q(XS1c ,w ) T s X X X Q(XS1c ,w ) S1,w S1c ,w XS1c ,w (5) pw (Y T ,XS1,w |XS1c ,w )s p1 (Y T ,XS1,w |XS1,wc )s (b)  ≤ X i (N −K i )M S1,w X Q(XS1c ,w ) XS1c ,w (c) ρ iρ ≤ (N −K i ) M X X S1,w ,XS1,w |XS1,wc ) pw (Y T ,XS1,w |XS1c ,w )s ρ p1 (Y T ,XS1,w |XS1,wc )s Q(XS1c ,w ) XS1c ,w pw (Y T ,XS1,w |XS1c ,w )s ρ p1 (Y T ,XS1,w |XS1,wc )s for all s > 0 and 0 ≤ ρ ≤ 1. Note that (a) is since the probability is less than 1 and can be raised to the power of ρ. (b) is critical. It follows from the symmetry of codebook construction, namely, the inner summation depends only on the codewords in XS1c ,w but not the ones in S1c ,w . Due to the code’sconstruction, i.e., its binning structure, there are exactly N −K M i possible sets of codewords to consider for S1c ,w . i (c) follows as the sum of positive numbers raised to the ρ-th power is smaller than the sum of the ρ-th powers. We now continue similar to [20], substituting the conditional error probability just derived in a summation over all codewords and output vectors. We have XX P (Ei0 ) = p1 (XS 1 , Y T ) Pr[Ei0 |w0 = 1, XS1 , Y T ] XS1 Y T ρ iρ ≤ (N −K i ) M X XX p1 (XS 1 ,Y T ) S1,w Y T XS1  X  XS1c ,w Q(XS1c ,w ) s pw (Y T ,XS1,w |XS1c ,w ) p1 (Y T ,XS1,w |XS1,wc )s Y  X XS1,w 1/(1+ρ) (XS1,w ,Y |XS1,wc ) Q(XS1c ,w )p1 1+ρ #T . XS1,wc −T log E0 (ρ)−ρ (N −K i ) + i log M T T ! − log (Ki ) ! T . , Pr[Ei |w0 = 1, XS1 , Y T ] (a)  X X X pw (Y T ,XS1,w |XS1c ,w )s ρ Q(XS1c ,w ) ≤ T s p1 (Y  which continue similar to [20], results in " X X 0 N −K ρ K P (Ei ) ≤ ( i ) M iρ ( i ) P (Ei0 ) ≤ 2 where (5) is exactly [20, eq. (25)], as when considering Ei0 we assume the decoder not only got K − i items right, but also the correct codeword in each such bin. Hence S1,w S1c ,w XS1c ,w p1 (Y T ,XS1,w |XS1,wc )s ρ Thus, we have ,XS1,w |XS1,wc ) p1 (Y w∈A XS1c ,w ≤ Q(XS1c ,w ) XS1c ,w pw (Y T ,XS1,w |XS1c ,w )s ρ  . A PPENDIX B I NFORMATION L EAKAGE AT THE E AVESDROPPER (S TRONG S ECRECY ) We wish to show that I(W ; Z T ) → 0. Denote by CT the random codebook and by XTSw the set of codewords corresponding to the true defective  items. N We assumed W ∈ {1, . . . , K } is uniformly distributed, that is, there is no a-priori bias to any specific subset. Further, the codebook includes independent and identically distributed codewords. The eavesdropper, observing Z T , wish to decode the true K independent and identically distributed codewords,  N which correspond to the defective items, one of K subsets. To analyze the information leakage at the eavesdropper, we note that the channel Eve sees in this case is the following multiple access channel. Each of the items can be considered as a “user” with M specific codewords. Eve’s goal is to identify the active users. Note the Eve’s channel can be viewed as (binary) Boolean MAC, followed by a BEC(1 − δ). The sum capacity to Eve cannot be larger than δ. In fact, since the codebook is randomly i.i.d. distributed, under a fixed ln(2) input distribution for the testing matrix ( ln(2) K , 1 − K ), Eve can obtain from each active user a rate of at most δ/K. Consequently, Eve may obtain a sub-matrix Z̃ of possibly transmitted codewords, where from each codeword Eve sees at most a capacity of δ/K. However, in this analysis, we help Eve even further to get the true sub-matrix Z̃, with the maximum rate δ/K. That is, we assume Eve gets the full rate, identifies the codewords of the users except for the erasures. We give Eve the power not to be confused from the existence of other, N − K users which did not transmit eventually. Providing this information to Eve only makes her stronger and thus a coding scheme that succeeds to keep Eve ignorant, will also succeed against the original Eve. 14 as P r(XjT ∈ C ∩ XjT ∈ Sh(Z̃j̃T , d)) = Vol(Sh(Z̃j̃T , d)) 2T  δ+0  2(1− K )T . 2T δ = T For each item we have M = 2 K codewords. Thus, on average, the number of codewords Eve sees on a shell, per defective item is |{w : XjT (w) ∈ Sh(Z̃j̃T , d)}| Z̃j̃T Figure 7: Codewords exactly lie in a ball around of radius δ d ≈ (1− K )T ), where j is the codeword index and 1 ≤ i ≤ N is the bin index. Now, once Eve obtains the true sub-matrix Z̃, since the codebook and the subsets of the items was generated independently and identically, we will analyze the information that leaks at the eavesdropper from each possibly transmitted codeword in Z̃ separately. Namely, per original codeword XjT , δ T from the T outcomes are not erased and are on average, K accessible to the eavesdropper via Z̃j̃T . Thus, out of the M independent and identically distributed codewords, there is an exponential number of codewords, that from eavesdropper perspective, may have participated in XTSw and could have resulted in the same Z̃j̃T . These consistent codewords are exactly those δ that lie in a ball around Z̃j̃T of radius d ≈ (1− K )T as depicted in Figure 7. The eavesdropper does not know what is the codeword XjT in XTSw , which was selected by the mixer and she even does not know what dH = kXjT − Z̃j̃T k is exactly (where k · k is the δ )T . hamming distance), other than the fact that dH ≈ (1 − K However, we help Eve by providing d and by choosing a  small yet exponential set of codewords (of size 2T K , for an appropriately small ) from the codebook C, chosen from the set of all codewords at distance d from Z̃j̃T , with the additional guarantee that it also contains the true XjT . We refer to this set as the oracle-given set XTOracel . Again, providing this information to Eve only makes her stronger and thus a coding scheme that succeeds to keep Eve ignorant, will also succeed against the original Eve. Conditioned on Eves view Z̃j̃T and XTOracel , each of the codewords in XTOracel is equally likely to have been participated in the pool tests. Nonetheless, Eve still has a reasonable  amount of uncertainty about which of the 2T K codewords was actually participated indeed, this is the uncertainty that we leverage in our analysis. We define, for any Z̃j̃T and d, a ball and a shell as Vol(Z̃j̃T , d) = {XjT : dH (XjT , Z̃j̃T ) ≤ d}, Sh(Z̃j̃T , d) = {XjT : dH (XjT , Z̃j̃T ) = d}. Hence, we define the probability for a codeword to fall in shell  δ+0 K  δ 0 · 2(1− K )T = 2T K . T 2 Hence, we can conclude that, on average, for every item, Eve has quite a few options in a particular shell. Now that we established that the average number of code0 T T TK words per item |{w : Xj (w) ∈ Sh(Z̃j̃ , d)}| is 2 , we wish to calculate the probability that the actual number of options deviates from the average by more than ε. We define = 2 T 0 EC1 (Z̃j̃T , d) := P r{(1 − ε)2T K ≤ 0 |{w : XjT (w) ∈ Sh(Z̃j̃T , d)}| ≤ (1 + ε)2T K }. Let us fix a pair Z̃j̃T and d for large enough T . By the Chernoff bound, and taking union bound over Z̃j̃T and d, we have 0 0 T K P r(EC1 (Z̃j̃T , d)) ≥ 1 − 2−ε 2 . Due to the super exponential decay in T , even when we take a union bound over all the codewords and all shells, we get that the probability that a codeword will have |{w : XjT (w) ∈ 0 Sh(Z̃j̃T , d)}| options significantly different than 2T K is very small. Actually, with high probability, Eve has almost the same number of options per item. Hence, for Eve all the codewords are almost equiprobable. In other words, Eve distribution on the items converges super-exponentially fast to a uniform one, hence H(W |Z T ) → H(W ) and we have I(W ; Z T ) → 0. R EFERENCES [1] R. Dorfman, “The detection of defective members of large populations,” The Annals of Mathematical Statistics, vol. 14, no. 4, pp. 436–440, 1943. [2] D.-Z. Du and F. K. Hwang, Combinatorial group testing and its applications. World Scientific, 1999, vol. 12. [3] A. J. Macula, “Probabilistic nonadaptive group testing in the presence of errors and DNA library screening,” Annals of Combinatorics, vol. 3, no. 1, pp. 61–69, 1999. [4] M. K. Varanasi, “Group detection for synchronous Gaussian codedivision multiple-access channels,” Information Theory, IEEE Transactions on, vol. 41, no. 4, pp. 1083–1096, 1995. [5] M. Cheraghchi, A. Karbasi, S. Mohajer, and V. Saligrama, “Graphconstrained group testing,” IEEE Transactions on Information Theory, vol. 58, no. 1, pp. 248–262, 2012. [6] S. Wu, S. Wei, Y. Wang, R. Vaidyanathan, and J. Yuan, “Partition information and its transmission over Boolean multi-access channels,” Information Theory, IEEE Transactions on, vol. 61, no. 2, pp. 1010– 1027, 2015. [7] ——, “Achievable partition information rate over noisy multi-access Boolean channel,” in Information Theory (ISIT), 2014 IEEE International Symposium on. IEEE, 2014, pp. 1206–1210. 15 [8] W. U. Bajwa, J. D. Haupt, A. M. Sayeed, and R. D. Nowak, “Joint source–channel communication for distributed estimation in sensor networks,” Information Theory, IEEE Transactions on, vol. 53, no. 10, pp. 3629–3653, 2007. [9] R. Clifford, K. Efremenko, E. Porat, and A. Rothschild, “k-mismatch with dont cares,” in European Symposium on Algorithms. Springer, 2007, pp. 151–162. [10] W.-T. Tsai, Y. Chen, Z. Cao, X. Bai, H. Huang, and R. Paul, “Testing web services using progressive group testing,” in Content Computing. Springer, 2004, pp. 314–322. [11] G. Cormode and S. Muthukrishnan, “What’s new: finding significant differences in network data streams,” IEEE/ACM Transactions on Networking (TON), vol. 13, no. 6, pp. 1219–1232, 2005. [12] Y. Xuan, I. Shin, M. T. Thai, and T. Znati, “Detecting application denial-of-service attacks: A group-testing-based approach,” Parallel and Distributed Systems, IEEE Transactions on, vol. 21, no. 8, pp. 1203– 1216, 2010. [13] M. T. Goodrich, M. J. Atallah, and R. Tamassia, “Indexing information for data forensics,” in Applied Cryptography and Network Security. Springer, 2005, pp. 206–221. [14] M. J. Atallah, K. B. Frikken, M. Blanton, and Y. Cho, “Private combinatorial group testing,” in Proceedings of the 2008 ACM symposium on Information, computer and communications security. ACM, 2008, pp. 312–320. [15] M. J. Freedman, K. Nissim, and B. Pinkas, “Efficient private matching and set intersection,” in Advances in Cryptology-EUROCRYPT 2004. Springer, 2004, pp. 1–19. [16] Y. Rachlin and D. Baron, “The secrecy of compressed sensing measurements,” in Communication, Control, and Computing, 2008 46th Annual Allerton Conference on. IEEE, 2008, pp. 813–817. [17] “Chapter 16: Cryptographic problems and philosophies,” in Stream Ciphers and Number Theory, ser. North-Holland Mathematical Library, C. D. Thomas W. Cusick and A. Renvall, Eds. Elsevier, 2004, vol. 66, pp. 391 – 406. [Online]. Available: http://www.sciencedirect. com/science/article/pii/S0924650904800187 [18] M. Bloch and J. Barros, Physical-Layer Security: From Information Theory to Security Engineering. Cambridge University Press, 2011. [19] A. D. Wyner, “The wire-tap channel,” Bell Syst. Tech. J., vol. 54, pp. 1355–1387, 1975. [20] G. K. Atia and V. Saligrama, “Boolean compressed sensing and noisy group testing,” Information Theory, IEEE Transactions on, vol. 58, no. 3, pp. 1880–1901, 2012. A minor corection appered in vol. 61, no. 3, pp. 15071507, 2015. [21] L. Baldassini, O. Johnson, and M. Aldridge, “The capacity of adaptive group testing,” in Information Theory Proceedings (ISIT), 2013 IEEE International Symposium on. IEEE, 2013, pp. 2676–2680. [22] Z. Füredi, “Onr-cover-free families,” Journal of Combinatorial Theory, Series A, vol. 73, no. 1, pp. 172–173, 1996. [23] E. Porat and A. Rothschild, “Explicit nonadaptive combinatorial group testing schemes,” IEEE Transactions on Information Theory, vol. 57, no. 12, pp. 7982–7989, 2011. [24] A. Sebő, “On two random search problems,” Journal of Statistical Planning and Inference, vol. 11, no. 1, pp. 23–31, 1985. [25] J. Scarlett and V. Cevher, “Limits on support recovery with probabilistic models: An information-theoretic framework,” in 2015 IEEE International Symposium on Information Theory (ISIT). IEEE, 2015, pp. 2331–2335. [26] ——, “Phase transitions in group testing,” in Proceedings of the TwentySeventh Annual ACM-SIAM Symposium on Discrete Algorithms. SIAM, 2016, pp. 40–53. [27] M. Aldridge, “The capacity of bernoulli nonadaptive group testing,” IEEE Transactions on Information Theory, 2017. [28] V. Y. Tan and G. K. Atia, “Strong impossibility results for sparse signal processing,” IEEE Signal Processing Letters, vol. 21, no. 3, pp. 260–264, 2014. [29] O. Johnson, “Strong converses for group testing in the finite blocklength regime,” arXiv preprint arXiv:1509.06188, 2015. [30] J. Scarlett and V. Cevher, “Converse bounds for noisy group testing with arbitrary measurement matrices,” arXiv preprint arXiv:1602.00875, 2016. [31] M. Aldridge, “Adaptive group testing as channel coding with feedback,” in Information Theory Proceedings (ISIT), 2012 IEEE International Symposium on. IEEE, 2012, pp. 1832–1836. [32] H.-B. Chen and F. K. Hwang, “A survey on nonadaptive group testing algorithms through the angle of decoding,” Journal of Combinatorial Optimization, vol. 15, no. 1, pp. 49–59, 2008. [33] A. De Bonis, L. Gasieniec, and U. Vaccaro, “Optimal two-stage algorithms for group testing problems,” SIAM Journal on Computing, vol. 34, no. 5, pp. 1253–1270, 2005. [34] P. Indyk, H. Q. Ngo, and A. Rudra, “Efficiently decodable non-adaptive group testing,” in Proceedings of the twenty-first annual ACM-SIAM symposium on Discrete Algorithms. Society for Industrial and Applied Mathematics, 2010, pp. 1126–1142. [35] C. L. Chan, S. Jaggi, V. Saligrama, and S. Agnihotri, “Non-adaptive group testing: Explicit bounds and novel algorithms,” Information Theory, IEEE Transactions on, vol. 60, no. 5, pp. 3019–3035, 2014. [36] M. Aldridge, L. Baldassini, and O. Johnson, “Group testing algorithms: bounds and simulations,” IEEE Transactions on Information Theory, vol. 60, no. 6, pp. 3671–3687, 2014. [37] C. E. Shannon, “Communication theory of secrecy systems,” Bell system technical journal, vol. 28, no. 4, pp. 656–715, 1949. [38] L. H. Ozarow and A. D. Wyner, “Wire-tap channel II,” AT&T Bell Laboratories technical journal, vol. 63, no. 10, pp. 2135–2157, 1984. [39] I. Csiszár and P. Narayan, “Secrecy capacities for multiple terminals,” IEEE Transactions on Information Theory, vol. 50, no. 12, pp. 3047– 3061, 2004. [40] ——, “Secrecy capacities for multiterminal channel models,” IEEE Transactions on Information Theory, vol. 54, no. 6, pp. 2437–2452, 2008. [41] U. M. Maurer, “Secret key agreement by public discussion from common information,” Information Theory, IEEE Transactions on, vol. 39, no. 3, pp. 733–742, 1993. [42] P. Kocher, “Timing attacks on implementations of Diffie-Hellman, RSA, DSS, and other systems,” in Advances in CryptologyCRYPTO96. Springer, 1996, pp. 104–113. [43] D. R. Stinson, Cryptography: theory and practice. CRC press, 2005. [44] A. W. Dent, “Fundamental problems in provable security and cryptography,” Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, vol. 364, no. 1849, pp. 3215–3230, 2006. [45] D. J. Bernstein, T. Lange, and C. Peters, “Attacking and defending the mceliece cryptosystem,” in International Workshop on Post-Quantum Cryptography. Springer, 2008, pp. 31–46. [46] P. Damaschke and A. S. Muhammad, “Competitive group testing and learning hidden vertex covers with minimum adaptivity,” Discrete Mathematics, Algorithms and Applications, vol. 2, no. 03, pp. 291–311, 2010. [47] P. Damaschke and A. Muhammad, “Bounds for nonadaptive group tests to estimate the amount of defectives,” Combinatorial Optimization and Applications, pp. 117–130, 2010. [48] R. G. Gallager, Information theory and reliable communication. Springer, 1968, vol. 2. [49] Y. Polyanskiy, H. V. Poor, and S. Verdú, “Channel coding rate in the finite blocklength regime,” IEEE Transactions on Information Theory, vol. 56, no. 5, pp. 2307–2359, 2010. [50] G. K. Atia, V. Saligrama, and C. Aksoylar, “Correction to “boolean compressed sensing and noisy group testing”,” Information Theory, IEEE Transactions on, vol. 61, no. 3, pp. 1507–1507, 2015.
7cs.IT
1 Random Ensembles of Lattices from Generalized Reductions arXiv:1702.00608v3 [cs.IT] 5 Feb 2018 Antonio Campello, Member, IEEE Abstract—We propose a general framework to study constructions of Euclidean lattices from linear codes over finite fields. In particular, we prove general conditions for an ensemble constructed using linear codes to contain dense lattices (i.e., with packing density comparable to the Minkowski-Hlawka lower bound). Specializing to number field lattices, we obtain a number of interesting corollaries - for instance, the best known packing density of ideal lattices, and an elementary coding-theoretic construction of asymptotically dense Hurwitz lattices. All results are algorithmically effective, in the sense that, for any dimension, a finite family containing dense lattices is exhibited. For suitable constructions based on Craig’s lattices, this family is smaller, in terms of alphabet-size, than previous ensembles in the literature. Keywords: Lattices, sphere packings, random codes, ideal lattices, codes over matrix rings the lower bound from integer lattices constructed from linear codes in Fnp , in the limit when p → ∞, with randomcoding arguments. Improvements of the lower bound, in turn, strongly rely on additional (algebraic) structure. For instance, Vance [11] showed that the best quaternionic lattice in dimension m (with real equivalent in dimension 4m) has density at least 3mζ(4m)/e24m−3 ≤ ∆4m . Using lattices built from cyclotomic number fields, Venkatesh [12] established the bound ∆2ϕ(m) ≥ m/22ϕ(m)−1, where ϕ(m) is Euler’s totient function (since m can grow as fast as a constant times ϕ(m) log log ϕ(m) this provides the first super-linear improvement). From another perspective, Gaborit and Zémor [13], and Moustrou [14] exploited additional coding-theoretic and algebraic structures to significantly reduce the family size of ensembles containing dense lattices. I. I NTRODUCTION There has been a renewed interest in the search for new constructions of lattices from error-correcting codes due to their various recent applications, such as coding for fading wiretap channels [1], Gaussian relay networks [2], compound fading channels [3] and index codes [4], to name only a few. For the applications considered in these works, it is desirable to lift a code over a finite field into a lattice that possesses a rich algebraic structure, often inherited from the properties of number fields. In the present work we provide an unified analysis of these constructions and investigate “random-coding”-like results for such lattices. Our focus is on the problem of finding dense structured lattice packings, although our techniques have a much broader scope of applications (as discussed in Section VII). Indeed, finding the densest packing is a central subject in the Geometry of Numbers, with a variety of well-established connections to Coding Theory. Let ∆n denote the best possible sphere packing density achievable by a Euclidean lattice of dimension n. The celebrated Minkowski-Hlawka theorem (e.g. [5], [6]) gives the lower bound ∆n ≥ ζ(n)/2n−1 for all n ≥ 2, where ζ(n) = 1 + 1/2n + 1/3n + . . . is the Riemann zeta function. Up to very modest asymptotic improvements, this is still, to date, the best lower bound for the best packing density in high dimensions. Typical methods for establishing the theorem depend on the construction of random ensembles of lattices and on meanvalue results [7], [8]. Rush [9] and Loeliger [10] obtained This work was supported in part by the São Paulo Research Foundation (fellowship 2014/20602-8). † Current address: Department of Electrical and Electronic Engineering, Imperial College London, South Kensington Campus, London SW7 2AZ, United Kingdom (e mail: [email protected]). Main Contributions. In this work we investigate general random lattices obtained from error correcting codes. The objective of this study is twofold: we provide unified analyses and coding-theoretic proofs of the aforementioned results, as well as a simple condition to verify if any new construction can be used to build ensembles containing dense lattices. We start from the fairly general definition of a reduction, i.e. a mapping that takes a lattice into the space Fnp . For a general reduction we prove the following: Theorem 1. Let φp : Λ → Fnp be a family of reductions (surjective homomorphisms), where Λ is a lattice of rank m in the Euclidean space. Consider the ensemble  Lp = βφp−1 (C) : C is a k − dimensional code in Fnp , for an appropriate normalization factor β so that all lattices have volume V . Denote by NΛ′ (r) = #(Br ∩ Λ′ ) the number of primitive points of Λ inside a ball of radius r. If the first minimum of Λp = ker φp satisfies   λ1 (Λp ) lim inf > 0, then (1) p→∞ pn/m lim ELp [NΛ′ (r)] = (ζ(m)V )−1 vol Br , p→∞ where the average is with respect to the uniform distribution on Lp . A slightly stronger version of the above result is precisely stated in Theorem 2. We shall refer to a family of reductions that satisfies condition (1) as non-degenerate. Non-degeneracy is indeed a very mild condition, and is satisfied, for instance, if Λp has non-vanishing Hermite parameter (e.g., Λp = pZn ). 2 Non-degenerate constructions immediately yield lattices satisfying the Minkowski-Hlawka lower bound (see the discussion in the end of Section II). By choosing specific suitable families of non-degenerate reductions, we can further improve this density and obtain a number of interesting corollaries. We highlight one of them: Corollary 1. Let OK be the ring of integers of a degree n number field K containing r(K) roots of unity. For any integer t ≥ 2, there exists an OK -lattice with dimension t and packing density r(K)tζ(tn)(1 − ε) ∆≥ , e(1 − e−t )2tn for any ε > 0. This proves for instance, the existence of ideal lattices in any dimension with density better, by a linear factor, than the Minkowski-Hlawka lower bound. This also recovers, for t = 2 and a judicious choice of number field and degree, the density in [12, Thm. 1] and [14, Thm. 2]. By allowing reductions to codes over matrix rings (rather than the field Fp ), we provide, in Section V, a coding-theoretic proof of the existence of dense Hurwitz lattices, as in [11]. Here is how Theorem 1 may be interpreted: the density of the kernel (coarse) lattice Λp is improved by “adjoining” a code C to it, through the reduction φp . Now if Λp itself has a reasonable density, we can improve it up to the MinkowskiHlawka bound. Building on this idea, we show in Section VI that if we start from a suitable reduction so that the base (fine) lattice Λ is not so thick (in terms of its covering density) and such that the kernels are not so sparse (in terms of packing density), we can bound the required size of p to be within a finite range. For instance, we show that by starting from the family of Craig’s lattices [15], we can √ build dense lattices from codes with alphabet-size p = O((n log n)1+ν ), where ν > 0 is any (small) positive constant. This improves significantly the size of codes required by usual constructions [9] [10] (where p grows at least as ω(n3/2 )). As observed in [15, pp. 18-19], the works of Rush (and Loeliger) already significantly reduce the family sizes of typical proofs of the Minkowski-Hlawka lower bound. It is worth mentioning that, in terms of absolute family size, the best result is achieved by [13] by restricting the average to double-circulant codes or [14] using cyclotomic lattices. For instance, while the logarithm of the search space for Craig’s lattices reductions has size n2 log log n, the family based on double-circulant codes has size n log n (we refer the reader to Table 1 for more details). We leave it as an open question to whether coupling a good reduction with a smaller family of codes can further reduce the overall search space. Organization. This work is organized as follows. In Section II we describe some basic definitions and notation. In Section III we establish our main result on general reductions and several corollaries. In Section IV we consider reductions induced by quotients of ideals in the ring of integers of a number field, proving the main corollaries. In Section V we construct random Hurwitz and Lipschitz lattices from codes over matrix rings. In Section VI, we discuss an “algorithmic” version of the main theorem and draw the final conclusions. II. P RELIMINARIES AND N OTATION m The Euclidean norm of x ∈ R is denoted by kxk = (x21 + . . . + x2m )1/2 . The ball of radius r in Rm is denoted by Br = {x ∈ Rm : kxk ≤ r}. A lattice Λ is a discrete additive subgroup of Rm . Denote by span Λ the minimal subspace of Rm that contains Λ. The rank of Λ is defined to be the dimension of span Λ. The quotient (span Λ)/Λ is compact, and its volume, denoted by V (Λ), is said to be the volume of Λ. The first minimum λ1 (Λ) of Λ is the shortest Euclidean norm of non-zero vectors in Λ. In general, the i-th minima of Λ are defined as λi (Λ) = min {r : dim (span {Br ∩ Λ}) = i} . The packing density of a rank m lattice Λ is defined as ∆(Λ) = vol Bλ1 /2 . V (Λ) We say that a point in x ∈ Λ is primitive if the intersection between Λ and the open line segment {αx : α ∈ (0, 1)} is the empty set. The set of all primitive points in Λ is denoted by Λ′ . Theorem 1 implies the Minkowski-Hlawka lower bound in the following fashion. From the average result, it follows that it must exist at least one Λ ∈ L such that NΛ (r) ≤ (ζ(m)V )−1 vol Brm . Now if we force the right-hand side to be equal to 2(1 − ε), for some small ε > 0, then, since a lattice has at least two minimum vectors, we must have NΛ (r) = 0. Therefore Λ can pack balls of radius r/2; rearanging the terms gives us, up to ε, the Minkowski-Hlawka bound. If Λ is a lattice with guaranteed number of minimum vectors (say, L) we can, by similar arguments, achieve density L(1 − ε)/2m . A k-dimensional vector subspace C ⊂ Fnp is called a (linear) code with parameters (n, k, p) (or simply an (n, k, p)-code). Throughout the paper we use the standard “big-O”, “bigomega” and “little-omega” notations, e.g. f (x) = Ω(g(x)) if lim supx→∞ |g(x)/f (x)| < +∞, and f (x) = ω(g(x)) if limx→∞ g(x)/f (x) = 0. III. G ENERALIZED R EDUCTIONS From now on, let Λ be a rank m lattice and let n ≤ m be an integer. Definition 1. Let φp : Λ → Fnp be a surjective homomorphism. Given a linear code C, its associated lattice via φp is defined as Λp (C) , φ−1 p (C). A surjective homomorphism as in the above definition will, from now on, be called a reduction. We shall see that Λp (C) is indeed a lattice of rank m. First observe that Λp (C) is a subgroup of Λp (Fnp ) = Λ. Since the quotient Λ/ker(φp ) ≃ Fnp is finite, ker(φp ) = Λp ({0}) , Λp is a sub-lattice of Λ, of rank m. From the inclusion Λp ⊂ Λp (C) ⊂ Λ, we conclude that the three lattices have the same rank. Moreover, Λp (C)/Λp ≃ C, and therefore V (Λp (C)) = |C|−1 pn V (Λ). Remark 1. There is an off-topic connection between Definition 1 and combinatorial tilings. If in addition to being surjective, the reduction φp is a bijection when restricted to a 3 set P ⊂ Λ of cardinality pn , then P tiles Λ by translations of vectors of Λp . This framework contains classical Construction A [15], [9], the constructions in [1], and [16]. We derive sufficient conditions for this general construction to admit a MinkowskiHlawka theorem. Set β = V 1/m /(pn−k V (Λ)1/m ) and let Lp = {βΛp (C) : C is an (n, k, p) − code} Definition 2. Let f : W → R be a Riemann-integrable function. We say that f is semi-admissible if b , ∀x ∈ W (1 + kxk)m+δ 1 |Cn,k | X Remark 2. If f and its Fourier transform fˆ are semiadmissible, then f is said to be admissible. In this paper we will not be concerned about admissible functions, which play an important role in the development of the so-called linear programming bounds for packings. Theorem 2. Let (pj )∞ j=1 be an increasing sequence of prime numbers such that there exist reductions φpj : Λ → Fnpj and let f : W → R be a semi-admissible function. If the first minimum of Λpj = Λpj ({0}) satisfies   E x∈βΛpj (C)  pj →∞  lim ELpj  pj →∞ X x∈βΛ′p (C) j   f (x) = (ζ(m)V )−1 X ≤ f (βx) = x∈Λpj \{0} W (5) where the averages are taken over all βΛpj (C) in the ensemble Lpj (Equation (2)). Proof. We will prove the “refined” statement (ii). The proof of (i) is similar, except for the last step. Recall that the set f (βx) j b . (1 + kβxk)m+δ Since the lattice Λp (C) has rank m ≥ 1, the series on the righthand-side of the above inequality is absolutely convergent for any pj . Moreover, since, by assumption kβxk ≥ βλ1 (Λpj ) ≥ cV 1/m (V (Λ))−1/m pα j → ∞, each individual term of the last sum tends to zero, as pj → ∞ and, by dominated convergence, the sum tends to zero. Let γ = (pkj − 1)/(pnj − 1). For the second term of Equation (7), we have:    X    (a) X E f (βx) = γ f (βx)  x∈βΛ′ (C)  ′ x∈Λ pj φp (x)6=0 j ∞ X µ(r) rm X rm γf (rβx) x∈Λ\{0} where µ denotes the Möbius function (see, e.g. [5, Sec. VI. 3.2]). In the above, (a) follows from (6) and (b) is the Möbius function inversion formula (see, e.g., [5, Sec. VI. 3.2]). The theorem follows by using the property ∞ X µ(r) r=1 f (x)dx, X x∈Λ′p X r=1 Z (7) From the assumption on f , (ii)  pj φp (x/β)=0 j pj φp (x/β)6=0 j = f (x)dx, (4) (6) j  (b) W x∈βΛpj (C)\{0} g(v), v∈Fn p \{0}  X    +E  f (x) .  x∈βΛ′ (C)  , for some constant c, α > 0, then (i)   Z X f (x) = V −1 lim ELpj  X Fnp x∈Λ′p (C) j φp (x)=0 j x∈Λ +α pkj − 1 pnj − 1 → R. Now for f : W → R,     X  X    f (x) = E  f (x)   ′ x∈βΛ′ (C) (3) Any bounded integrable function with compact support is semi-admissible. Of particular interest are indicator functions of bounded convex sets. Notice that, for any semi-admissible function and a rank-m lattice Λ in W , X f (x) < +∞. n−k g(c) = C∈Cn,k c∈C\{0} where b > 0 and δ > 0 are positive constants. λ1 (Λpj ) ≥ cpj m X for a function g : (2) be the ensemble of all lattices associated to codes of dimension k, normalized to volume V . Suppose that a lattice in Lp is picked at random by choosing C uniformly. We shall prove a generalized version of the Minkowski-Hlawka theorem for Lp . Instead of functions with bounded support, we will consider a wider class of functions. Let W = span(Λ) be the minimal subspace of Rn containing Λ (therefore, dim(span(Λp (C))) = m). |f (x)| ≤ of Cn,k of all (n, k)-codes satisfies Loeliger’s balancedness equation [10] rm = 1 ζ(m) and observing that for the inner sum satisfies Z X rm γf (rβx) = (V −1 ζ(m)) lim pj →∞ x∈Λ\{0} f (x)dx, W by the definition of Riemann integral. Exchanging the limit and the sum is justified by dominated convergence, given the condition on f . 4 Corollary 3. For any ε > 0, there exists Λpj ∈ L such that Example 1. If Λ = Zn and φp is the reduction modulo p, we obtain mod-p lattices as in [10]. It is clear that Λp = pZn satisfies the hypothesis of Theorem 2, with m = n and α = k/n. This implies Theorem 1 of [10]. Definition 3. A sequence of surjective homomorphisms n (φj )∞ j=1 , φj : Λ → Fpj is said to be non-degenerate if m Y i=1 ∆i (Λ)1/m ≥ 2mζ(m)(1 − ε) . e(1 − e−m ) (10) Proof. The proof follows from a method of Rogers [8], choosing f (x) appropriately in the Minkowski-Hlawka theorem. We shall give a complete proof in the next section, in the context of OK -lattices. n λ(Λpj ) ≥ cpjm , IV. C ONSTRUCTIONS F ROM N UMBER F IELDS for some constants c > 0. Similarly, the sequence of associated ensembles (Equation 2) are said to be non-degenerate. It follows that if the reductions are non-degenerate, the associated ensemble admits the Minkowski-Hlawka theorem. Example 2 (“Natural reduction”). If m = n, the natural reduction to Fnp is as follows. Given a basis x1 , . . . , xn for Λ, take φp to be the linear map defined by φp (xi ) = ei ∈ Fnp , where ei the i-th canonical vector (0, . . . , 0, 1, 0, . . . , 0). It is clear that φ is surjective and ker φp = pΛ, therefore the associated sequence of reductions is non-degenerate. This provides a systematic way of constructing good sublattices of a given lattice. Taking f (x) to be the indicator function of a ball in part (ii) of Theorem 2, we recover Theorem 1. Another function 2 of interest is f (x) = e−τ kxk for τ > 0, yielding the theta series X 2 ΘΛ (τ ) = e−τ kxk . x∈Λ A corollary of part (i) of the theorem is the following: Corollary 2. The average theta series of a sequence of nondegenerate ensemble satisfies  π m/2 + 1. (8) lim ELpj [ΘΛ (τ )] = V −1 pj →∞ τ Corollary 2 can, for instance, be applied to the construction of sufficiently flat Gaussian measures for secure communications (cf [17]). Remark 3. The condition for non-degeneracy can be rewritten as lim inf γ(Λpj ) > 0, pj →∞ where γ(Λ) = λ(Λ)/V (Λ)1/m is the Hermite parameter of Λ. In other words, non-degeneracy is equivalent to non-vanishing Hermite parameter of the sequence of kernel lattices. We close this section with another consequence of Theorem 2. We shall refer to each ratio ∆i (Λ) = vol Bλi /2 , i = 1, . . . , m, V (Λ) (9) as the i-th successive density of a lattice S∞ Λ. For a sequence of non-degenerate ensemble, put L = j=1 Lpj . From now on we consider constructions of random ensembles based on algebraic number theory. We refer the reader to [18] and [19] for an introduction to the theory, as well as undefined notation. Let K/Q be a number field with degree n and signature (r1 , r2 ). Denote its real embeddings by σ1 , . . . , σr1 and their pairs of complex embeddings by σr1 +1 , σr1 +1 , . . . , σr1 +r2 +1 , σr1 +r2 +1 . Let OK be the ring of integers of K and I ⊂ OK be an ideal. An ideal can be identified with a real lattice of dimension (r1 + 2r2 ) via the canonical embedding σ : OK → Rr1 +2r2 σ(x) = (σ1 (x), . . . , σr1 (x),ℜσr1 +1 (x), . . . ℜσr1 +r2 +1 (x), ℑσr1 +1 (x), . . . , ℑσr1 +r2 +1 (x)). Lattices constructed from the embedding of ideals I ⊂ OK are called ideal lattices, and appear in the study of modular forms, coding theory, and cryptography. In this section we study the Minkowski Hlawka theorem for OK -lattices and related structures. Let E = K ⊗Q R be the Euclidean space generated by K. An OK -lattice is a free OK sub-module of E t , for some t > 0. In particular, an OK lattice is closed under multiplication by elements of OK . The Euclidean norm in E is induced by the trace form. Notice that K is naturally embedded in E. In the cases when K is either totally real or a totally imaginary extension of a real number field (CM-field) some notational simplifications can be made. For instance, we can write the trace form as tr(xy) = σ1 (x)σ1 (x) + . . . + σn (x)σn (x). We discuss the average behavior of a general reduction from algebraic number theory [1], [3], [2], defined in the sequel. A prime p is said to split completely if pOK can be factored into the product of prime ideals p1 p2 · · · pn . Definition 4. Let p be a prime that splits completely, and p an ideal above p. Consider π : OK → OK /p ≃ Fp a projection onto p and σ the canonical embedding. Let Λ = σ(OK )t (the canonical embedding is applied componentwise). Take φp : Λ → Ftp φp (σ(x1 , . . . , xt )) = (π(x1 ), . . . , π(xt )) nt and define Λp (C) = φ−1 p (C) ⊂ R . 5 Lemma 1. The ensemble induced by Definition 4 is nondegenerate. Proof. The minimum algebraic norm of an element of p is greater or equal p. Hence Λp = σ(p)t has minimum √ than 1/n norm at least np , finishing the proof. A very important caveat to the previous lemma is the fact that there must exist an infinite number of primes p such that the construction above is possible. This follows from Chebotarev’s density theorem (e.g. [20] Cor. 13.6, p. 547), which implies that the natural density δ of primes that split completely in K is positive (indeed, one has 0 < δ ≤ 1/n!). Remark 4. Very similarly, it is possible to prove that the constructions in [16] are non-degenerate. Suppose that K contains r(K) roots of unity. Let µ be a root of unity contained in OK . It follows that kσ(µx)k = kσ(x)k (e.g. [21, Lem. 3.1]). Therefore, each Λ constructed as in Definition 4 contains at least r(K) minimal vectors and we automatically obtain the density r(K)(1 − ε)/2nt (this argument was used by Venkatesh [12] to prove that cyclotomic lattices (Z[µ]-lattices where µ is a root of unity) achieve density m(1 − ε)/22φ(m) ). For t-dimensional OK lattices, however, there is a loss of a linear factor of t in the enumerator. Nevertheless, we can improve the following density up to that of Corollary 1, using a method by Rogers [8], recently employed in [11] to quaternionic lattices. The basic idea is to apply to Theorem 2, rather than indicator function of a ball, to a bounded-support function that allows us to analyze the generalized densities of the ensemble. After ensuring the existence of a lattice with good generalized densities, it is possible to apply standard linear transformations to such a lattice (see e.g. [11, Thm. 2.2]) in order in order to transform it into a lattice with good packing density. t Proof of Corollary 1: . For Λ0 ⊂ OK let the i-th successive minima of Λ0 (over K) be the smallest i-such that the ball Br contains the canonical embedding of i linearly independent vectors (over K). More formally   −1 (σ(Λ0 ) ∩ Br ) = i . λK i (Λ0 ) = min r > 0 : dim spanK σ (11) K Notice that λK 1 (Λ0 ) = λ1 (σ(Λ0 )) and, in general λi (Λ0 ) ≥ λi (σ(Λ0 )). Also, if x1 , . . . , xt are linearly independent over K and achieve the sucessive minima of Λ0 , then the embeddings σ(x1 ), . . . , σ(xt ) are linearly independent and primitive in σ(Λ0 ) ⊂ Rnt . Now let f : Rnt → R be the following function with limited support:  1/n if kyk ≤ re(1−t)/tn    kyk 1 f (y) = if re(1−t)/tn ≤ kyk ≤ re1/tn nt − log r   0 otherwise. (12) We have Z e(1 − e−t )rnt vol B1 f (y)dy = . nt Rnt Choose r such that the right-hand side of this last equation is equal to r(K)V ζ(nt)(1 − ε)/n for a small ε < 1. Let φp be as in Definition 1 and Lp its induced ensemble Lp = {βΛp (C) : C is an (n, k, p) − code} (13) as in Equation (2). According to Theorem 2, it is possible to find Λ1 = βσ(Λ0 ) ∈ Lp of volume V such that, for p sufficiently large X r(K) r(K) < , (14) f (y) ≤ (1 − ε) n n ′ y∈Λ1 Let v1 , . . . , vt be linearly independent vectors in Λ0 achieving the successive minima, kβσ(vi )k = λK i (Λ0 ). We have X y∈Λ′1 f (y) ≥ t X X i=1 f (βσ(µvi )) = r(K) µ t X f (βσ(vi )), i=1 where the sum with subscript µ is over all roots of unity in 1/n−1 K. From this we conclude that, for all i, βλK i (Λ0 ) ≥ re and   tQ K 1 1 β λi (Λ0 ) < . − log n rt n Therefore, for the t successive densities (Eq. (9)): t Y i=1 ∆K i !1/t = t Y vol(BλK ) i /2 i=1 V !1/t (15) r(K)tζ(nt)(1 − ε) ≥ . e(1 − e−t )2nt But in this case, we can find Λ̃ whose packing density (or r(K)tζ(nt) ∆K 1 ) is greater or equal than e(1−e−t )2nt (1−ε) (e.g. [11, Thm. 2.2]). V. BALANCED S ETS OF C ODES OVER M ATRIX R INGS In some contexts, the “natural” underlying alphabet in the reduction φp is, rather than the field Fp , the ring Mn (Fp ) of n × n matrices with entries in Fp . Although we can 2 identify Mn (Fp ) with Fnp , the identification does not carry enough algebraic structure for our purposes. For instance, we cannot guarantee that the constructed lattices are closed under multiplication by units, which is crucial in order to obtain the full density improvements of these lattices, as in [11]. For this reason, we study in this section a version of Theorem 2 for codes over matrix rings. A. An Averaging Bound for Codes over Rings Let R be a finite ring and R∗ its units. Denote by (Rn )∗ the set of vectors in Rn such that at least one coordinate is a unit. A linear code in C ⊂ Rn is a free1 R-submodule of Rn (with the natural scalar multiplication). Following [10], we define balanced sets of codes as follows. 1 This may differ from the literature, where a linear code over a ring is simply an additive subgroup of Rn . The requirement that a linear code is a free module is necessary for Lemma 2 to hold. 6 Definition 5. Consider a non-empty set of codes Cb of same cardinality. We say that Cb is balanced if any x ∈ (Rn )∗ is contained in the same number of codes (say, L) of Cb . Let M be the cardinality of a code in Cb . From a counting argument, one can see that M |Cb | ≥ L|(Rn )∗ |. The following lemma shows how to bound averages of functions in (Rn )∗ . n + Lemma 2. Let g : R P → R be a function. For a code C, ∗ we define g (C) = c∈C∩(Rn )∗ g(c). If Cb is the set of all codes of rank k then k E [g ∗ (C)] ≤ |R| g ∗ (Rn ), |(Rn )∗ | where the expectation is with respect to the uniform distribution on Cb . Proof. For any balanced set of codes with cardinality M , we have     X X E[g ∗ (C)] = E  g(c) = E  g(x)1C (x) c∈C∩(Rn )∗ = x∈(Rn )∗ X E [g(x)1C (x)] X g(x) x∈(Rn )∗ = x∈(Rn )∗ M L ≤ g ∗ (Rn ). |Cb | |(Rn )∗ | We now need to prove that the set of all codes of rank k is balanced. Let y be any element in Rn ∗ . There exists an invertible linear map T (y) = (1, 0, . . . , 0) = e1 . Since T is rank-preserving, y ∈ C if and only if e1 ∈ T (C), where C and T (C) have same rank. This induces a bijection between the codes that contain y and the codes that contain e1 , proving the statement. B. Lipschitz and Hurwitz Lattices The quaternion skew-field H is given by H = {a + bi + (c + di)j : a, b, c, d ∈ R}, with the usual relations i2 = j 2 = −1 and ij = −ji. Vance recently [11] proved a Minkowski-Hlawka theorem for lattices in H over the Hurwitz order. Here we show how to recover a “coding-theoretic” version of this result from generalized reductions. We first explain how to deduce a slightly simpler case, for the Lipschitz order. The Lipschitz integers L ⊂ H is the (non-maximal) order L = {x + yj : x, y ∈ Z[i]} . Recall that a quaternion has matrix representation   x −y . y x Let p be an ideal in Z[i] above p that splits. Let π : Z[i] → Z[i]/p be a projection. We consider the following “singleletter” reduction: φH p : L → M2 (Fp ) φp (x + yj)H =  π(x) π(y) −π(y) π(x) 4 We have ker φH p = (pZ[i]) + (pZ[i])j. Identifying H with R in the natural way ψ(a, b, c, d) → a + bi + (c + di)j we obtain a reduction φp : Z4 → M2 (Fp ), φp (x) = H φH p (ψ(x)). By abuse of notation, we will also denote by φp m the reduction applied componentwise in vector of L , i.e., φH p (x1 + y1 j, . . . , xm + ym j) H m = (φH p (x1 + y1 j), . . . , φp (xm + ym j)) ∈ M2 (Fp ) . (16) H −1 If C ⊂ M2 (Fp )m is a linear code, then ΛH (C) p (C) = (φp ) −4m is a quaternionic lattice with volume |C|p . Let Cb be a balanced set and Lp the associated lattice ensembles  Lp = βΛH p (C) : C ∈ Cb , where β = (V /(|C|−1 p4m ))1/(4m) . The following Theorem 3 is the analogous of Theorem 1 for Lipschitz lattices. We need the following lemma Lemma 3. If φp (x + yj) is non-invertible for x + yi ∈ L, then the squared norm of x + yj is a multiple of p. Proof. If det φp (x + yj) = 0, then π(xx + yy) = 0, i.e., k(x, y)k2 ∈ p. Since the norm of a Lipschitz quaternion is an integer, and p is above p, the result follows. Theorem 3. Let Cb be a balanced set of codes with rank k > m/2. If f is a semi-admissible function then   Z X lim ELp  f (ψ(x)) ≤ (ζ(4m)V −1 ) f (x)dx. p→∞ R4m x∈βΛH p (C) (17) Proof. The proof is very similar to that of Theorem 1. Here we divide the expectation into invertible and non-invertible elements (we make the change of variable x = βy, to facilitate), i.e. X X f (ψ(βy)) f (ψ(βy)) = y∈ΛH p (C) y∈ΛH p (C) φ(y)∈(M2 (Fp )m )∗ + X f (ψ(βy)). x∈ΛH p (C) m ∗ φ(y)∈(M / 2 (Fp ) ) The first term tends to zero as p → ∞ from Lemma 3, since f is semi-admissible and √ β kψ(y)k ≥ β p = |C|1/4m p−1/2 = pk/m−1/2 → ∞, (18) as pto∞. From Lemma 2 we conclude that the second term is upper bounded by the right-hand side of (17) as p → ∞. For the maximal Hurwitz order H = {a + bi + cj + d(−1 + i + j + ij)/2 : a, b, c, d ∈ Z} ,  . the theorem follows by considering reductions from leftprime ideals P ⊳ H. For any rational prime p, there exist 7 isomorphisms H/pH ∼ Fp (i, j, k) ∼ M2 (Fp ) (e.g. Wedderburn’s Theorem [22, Thm. 6.16 Lem 9.2.1]), where noninvertible elements in H/pH have reduced norm (determinant) proportional to p. Notice that in this case we obtain a reduction φp : D4 → M2 (Fp ), where D4 is the checkerboard lattice in dimension four [15, Sec. 7.2]. An explicit realization of ring isomorphism is obtained by setting     1 0 0 −1 φ(1) = , φ(i) = 0 1 1 0   a b and φ(j) = , b −a where a and b are two integers such that a2 + b2 ≡ −1 (mod p). Notice that such an isomorphism preserves the residue class of the reduced norm, i.e. nrd(x) = det φ(x) (mod p), for any x ∈ H. VI. A LGORITHMIC E FFECTIVENESS Theorem 1 holds in the limit pj → ∞. However, for each n, under some conditions it is possible to find finite ensembles that contain dense lattices. In the literature, this is referred to as effectiveness (e.g. [15, p. 18] and [13]). We show conditions for a family of reductions to be effective. We need the following lemma, which is a special case of a classical result in the Geometry of Numbers (see [6, p. 141]) and is also valid if Br is replaced by more general sets. We include a proof here for the sake of completeness. Lemma 4. Let P be a fundamental region for Λ, an let l0 = supx∈P kxk . For r > l0 , we have (r − l0 )n Vn ≤ V (Λ)NΛ (r) ≤ (r + l0 )n Vn . (19) In particular, we can take l0 = τ (Λ) to be the covering radius of Λ. Proof. We show the set inclusion [ Br−l0 ⊂ (x + P) ⊂ Br+l0 . x∈Λ∩Br The lemma then follows from a simple volume calculation of the three sets. If y = x + p, x ∈ Λ ∩ Br , p ∈ P, then kyk ≤ kxk + kpk ≤ r + l0 , proving the second inclusion. For the first inclusion, let y ∈ Br−l0 and write it as y = x + p, with x ∈ Λ and p ∈ P (this is always true since P is a fundamental region). Then kxk ≤ kyk + kpk ≤ r. as µ(Λ) = τ (Λ)/V (Λ)1/m . Recall that ρ(Λ) = λ1 (Λ)/2 is the packing radius of Λ. A lattice satisfying the MinkowskiHlawka bound has density  1/m V (Λ) Vm ρ(Λ)m m−1 1−1/m > 1/2 ⇒ ρ(Λ) ≥ 2 , ∆= V (Λ) Vm where Vm = vol B1 is thepvolume of the unit ball in Rm . −1/m ∼ m/2πe, if V (Λ) is normalized Recalling that Vm to one, this implies that the packing radius of good lattices should scale as r m . ρ(Λ) ∼ 2 2πe Proposition 1. Let the notation be as in Theorem 1 and let ε > 0. Let δ = k/n be the rate of the underlying codes. Suppose that √ nδ/m γ(Λpj ) = Ω( m) and (i) pj n/m = ω(mµ(Λ)/γ(Λpj )). (ii) pj If m is sufficiently large, there exists a code with parameters (n, k, pj ) such that the lattice Λpj (C) has packing density greater than (1 − ε)/2m−1 . Proof. For simplicity, suppose that the volume of Λpj (C) equals 1, which can be achieved by choosing an appropriatepscaling factor. Considering the above discussion let r = m/2πe. In the notation of the proof of Theorem 1, the average lattice point enumerator (Equation (7)) becomes: h  i h  i E # βΛ′pj (C) ∩ Br = E # βΛ′pj ∩ Br i h  + E # β(Λ′pj (C)\Λp ) ∩ Br . The first term of the right-hand side zero whenever k/m pj λ1 (Λpj ) V (Λpj )1/m ≥ r. (20) The second term satisfies i pk − 1 h   j # β(Λ′ \Λpj ) ∩ Br E # β(Λ′pj (C)\Λp ) ∩ Br = n − 1 pj !m µ(Λ) Vm k−n m m 1 + (n−k)/m ≤pj (r + βτ (Λ)) = Vm r . β m V (Λ) rp j (21) Imposing the right-hand-side of (21) to be 2(1 − ε), we obtain a lattice with density (cf. Section II): !−m µ(Λ) 1−ε . (22) ∆ ≥ m−1 1 + (n−k)/m 2 rp j A. Effective families containing dense lattices The next proposition essentially says that if the base lattice Λ is sufficiently “thin” and the kernel-lattices Λpj are not so sparse, it is possible to bound pj in terms of the rate of the underlying code. The conditions are very mild (they are achievable, for instance, by Λ = Zm and Λp = pZm ). For convenience, we recall the definition of the Hermite parameter γ(Λ) = λ1 (Λ)/V (Λ)1/m and define the covering parameter Under the conditions of the theorem, the term in parenthesis tends to 1 as m → ∞. Remark 5. Similar conditions hold for the case of quaternionic lattices. In this case, in light of the proof of Theorem 3 and Equation (18), condition (i) should be replaced by p = Ω(r2m/(2k−m) ). 8 Example 3. Let m = n, Λ = Zm and φp be the “modulo-p” reduction. Conditions (i)-(ii) of Proposition 1 state that p ≥ c1 m1/2δ and p ≥ c2 m3/2+ν , where c1 , c2 are constants and ν is any small number. The optimal rate (i.e., the one that yields the smallest asymptotic behavior of p) is δ ∼ 1/3, which gives us optimal alphabetsize p = m3/2+ν , for any positive constant ν. This provides an alternative derivation of [9]. The alphabet-size in the above example can be further improved by starting the reductions with a lattice which already has a good density, as shown next. Let Λ = Aln be a Craig’s lattice [15, pp.222-224] of rank m = n, where n + 1 = q is a prime. From [23, Prop. 4.1], a Craig’s lattice is similar to the embedding of the ideal (1 − µp )l Z[µp ] in the cyclotomic field Q(µp ). A concrete realization is 1 Λ = √ σ((1 − µp )l Z[µp ]). p From this, we have Λ∗ ∼ n/2−l An , √ λ1 (Λ) 2l ≥ 1/n (2l−1)/2n V (Λ) (n − 1) √ λ1 (Λ∗ ) n − 2l ≥ . and ∗ 1/n V (Λ ) (n − 1)(n−2l−1)/2n Following [15, p. 224]’s suggestion, we consider Craig’s lattices with parameter l = ⌊n/2 log(n + 1)⌉ so that, for sufficiently large n, r r  λ1 (Λ) 2π n + o(1) ≥ log n 2πe V (Λ)1/n and √ λ1 (Λ∗ ) ≥ e + o(1). ∗ 1/n V (Λ ) From Banaszczyk’s transference bound [24]: √ τ (Λ) n √ ≤ . 1/n 2 e + o(1) V (Λ) Therefore, using a natural reduction, conditions (i) and (ii) in Proposition 1 become p p p ≥ c1 ( log n)1/δ and p ≥ c2 (n log n)1+ν , for some constants c1 , c2 and any positive ν. We can further optimize the rate by equalizing the coefficients from where we get log log n . (23) δ∼ 2 log n + log log n ⌊n/2(n+1)⌉ Corollary 4. Let Λ = An and let φp be a natural reduction, as described in Example 2. Let ε > 0, n sufficiently large and let the rate be as in (23). There exists a code C with parameters   p n, δn, O((n log n)1+ν ) , for any positive ν, such that Λp (C) has packing density arbitrarily close to (1 − ε)/2m−1 . We close this subsection with a comment on the absolute family size of a reduction. If the set of all (n, k, p) codes is considered, then the search space for a dense lattice is given by the Gaussian binomial: hni k p = k−1 Y i=1 pn − pi ∼ pk(n−k) . pk − pi (24) Plugging the bounds for p in Proposition 1 gives an upper bound on the exhaustive search complexity. On Table 1 we provide a comparison of the parameters of some constructions in the literature in terms of rank of the base lattice Λ, code parameters, alphabet-size and log of the family-size (contrary to a statement in [13], [14], the complexity of Rush’s construction is exp(cn2 log n) rather than exp(n log n), and therefore the gains of averaging over double-circulant codes/cyclotomic lattices are even higher than the ones stated). B. Packing Efficiency A cruder measure of goodness, which is suitable for coding applications, is the packing efficiency [25]. Define ρeff (Λ) = (V (Λ)/Vm )1/m as the radius of a ball whose volume is V (Λ). The Minkowski bound can be rephrased in terms of packing efficiency, as 1 ρ(Λ) ≥ . (25) ρeff (Λ) 2 A “packing-good” family of lattices is such that its packing efficiency is arbitrarily close to 1/2. As shown in [25, Sec. IV], it is possible to find families with good asymptotic packing efficiency using Loeliger’s construction, provided that p = O(m1/2+β ), for any positive small β. Similarly, we can show that Craig’s lattices constructions can achieve packing efficiency arbitrarily close to 1/2 with alphabet-size p = O((log m)1/2+β ). VII. F INAL D ISCUSSION Applications. As observed by Loeliger [10], random ensembles of lattices are not only good in terms of packing density, but are also sphere-bound achieving when used as infinite constellations for the AWGN channel. Indeed, Rush [9] and Loeliger’s [10] Construction A Z-lattices are ubiquitous in applications to information transmission over Gaussian channels and networks. However, for other communication problems, such as information transmission in the presence of fading and multiple antennas, it is desirable to enrich the lattices with some algebraic (multiplicative) structure. To this purpose, several recent works such as [2], [1], [16], [3] present different constructions that attach a linear code to an algebraic lattice, but, to date, there is no unified analysis of such ensembles. The generalized reductions described here provide a method for establishing the “goodness” of all such constructions at once. It also provides a simple condition to verify if any new construction is “good” (e.g., sphere-bound achieving). This was indeed the initial motivation of the author. Further Perspectives. The framework considered in this paper is used to provide simple alternative (coding-theoretic) proofs 9 Construction Construction A over Z [9], [10] Random double-circulant [13] Cyclotomic lattices [14] Craig’s reduction rank(Λ) m=n m=n m = 2Φ(l), l ∈ N m=n (n, k) (n, δn), δ ∼ 1/3 (n, n/2) (2, 1) log n (n, δn), δ ∼ 2 loglog n+log log n p n3/2 n2 log n 3 l (log l)Φ(l) n(log n)1/2 Log family size n2 log n n log n m log m n2 log log n Table 1. Parameters of different effective families contaning dense lattices. The rates δ are up to lower order terms, and the log family-sizes up to constants and lower order terms. Φ(l) denotes Euler’s totient function. and improvements on previous refinements on the best packing density. It not only implies the existence of dense lattices, but also of structured lattices, with the structured inherited from the underlying reduction. The question whether it is possible to improve on the cn log log n/2n−1 asymptotic behavior of cyclotomic fields by specializing the reductions (or the family of codes) appropriately is still open. Furthermore, all known lower bounds on ∆n are of the form ∆n ≥ 2−n(1+ε(n)) , with ε(n) = O(log n/n), which improves only marginally on the Minkowski-Hlawka lower bound. According to Gruber [26, p. 388], Hlawka believed that no essential improvement can be made, probably meaning that the exponent 2 is optimal. Nevertheless, the best known upper bound on ∆n , due to Kabatianskii and Leveshenstein, is of the form C −n , where C ≈ 1.51. Closing this gap is a long-standing open problem. ACKNOWLEDGMENTS The author would like to thank Cong Ling for his constant enthusiasm on the topic and for his suggestions, as well as both reviewers for pointing out inaccuracies in the first version of the manuscript. He also acknowledges Sueli Costa and JeanClaude Belfiore for fruitful discussions and for hosting him at University of Campinas and at Télécom Paristech, where part of this work was developed. This work was supported in part by FAPESP (fellowship 2014/20602-8). R EFERENCES [1] W. Kositwattanarerk, S. S. Ong, and F. Oggier. Construction A of Lattices Over Number Fields and Block Fading (Wiretap) Coding. IEEE Transactions on Information Theory, 61(5):2273–2282, May 2015. [2] Y.-C. Huang, K. R. Narayanan, and P.-C. Wang. Adaptive compute-andforward with lattice codes over algebraic integers. arXiv/1501.07740, 2015. [3] A. Campello, C. Ling, and J. C. Belfiore. Algebraic lattice codes achieve the capacity of the compound block-fading channel. In IEEE International Symposium on Information Theory (ISIT), pages 910–914, July 2016. [4] Y. C. Huang. Lattice index codes from algebraic number fields. IEEE Transactions on Information Theory, 63(4):2098–2112, April 2017. [5] J. W. S. Cassels. An Introduction to the Geometry of Numbers. SpringerVerlag, 1997. [6] P. M. Gruber and C. G. Lekkerkerker. Geometry of Numbers. NorthHolland, 1987. [7] C. L. Siegel. A Mean Value Theorem in Geometry of Numbers. Annals of Mathematics, 46(2):340–347, 1945. [8] C. A. Rogers. Packing and Covering. Cambridge University Press, March 1964. [9] J. A. Rush. A lower bound on packing density. Inventiones mathematicae, 98(3):499–509, 1989. [10] H.-A. Loeliger. Averaging bounds for lattices and linear codes. IEEE Transactions on Information Theory, 43(6):1767–1773, Nov 1997. [11] S. Vance. Improved sphere packing lower bounds from hurwitz lattices. Advances in Mathematics, 227(5):2144 – 2156, 2011. [12] A. Venkatesh. A note on sphere packings in high dimension. International Mathematics Research Notices, 2012. [13] P. Gaborit and G. Zémor. On the construction of dense lattices with a given automorphisms group. Ann. Inst. Fourier (Grenoble), 57(4):1051– 1062, 2007. [14] P. Moustrou. On the density of cyclotomic lattices constructed from codes. International Journal of Number Theory, 13(05):1261–1274, 2017. [15] J. H. Conway and N. J. A. Sloane. Sphere-packings, lattices, and groups. Springer-Verlag, New York, NY, USA, 1998. [16] R. Vehkalahti, W. Kositwattanarerk, and F. Oggier. Constructions A of lattices from number fields and division algebras. In IEEE International Symposium on Information Theory (ISIT), pages 2326–2330, June 2014. [17] C. Ling, L. Luzzi, J.-C. Belfiore, and D. Stehle. Semantically Secure Lattice Codes for the Gaussian Wiretap Channel. IEEE Transactions on Information Theory, 60(10):6399–6416, Oct 2014. [18] I. Stewart and D. Tall. Algebraic Number Theory and Fermat’s Last Theorem. CRC Press, 3rd edition, 2001. [19] F. Oggier. Introduction to Algebraic Number Theory. Lecture notes available on the author’s website http://www1.spms.ntu.edu.sg/ frederique/AA11.pdf, accessed on 26/07/2017. [20] J. Neukirch. Algebraic Number Theory, volume 322 of Grundlehren der mathematischen Wissenschaften. Springer-Verlag Berlin Heidelberg, 1999. [21] P. Autissier. Variétés abéliennes et théorème de Minkowski-Hlawka. Manuscripta Mathematica, 149(3):275–281, 2016. [22] J Voight. The arithmetic of quaternion algebras. 2017, (book available online https://math.dartmouth.edu/ jvoight/quat-book.pdf), accessed on 26/07/2017. [23] C. Bachoc and C. Batut. Étude algorithmique de réseaux construits avec la forme trace. Experiment. Math., 1(3):183–190, 1992. [24] W. Banaszczyk. New bounds in some transference theorems in the geometry of numbers. Mathematische Annalen, 296(1):625–635, 1993. [25] U. Erez, S. Litsyn, and R. Zamir. Lattices which are good for (almost) everything. IEEE Transactions on Information Theory, 51(10):3401– 3416, Oct 2005. [26] P. Gruber. Convex and Discrete Geometry. Springer Berlin Heidelberg, 2007.
7cs.IT
Twenty (simple) questions arXiv:1611.01655v3 [cs.DM] 25 Apr 2017 Yuval Dagan, Yuval Filmus, Ariel Gabizon, and Shay Moran April 26, 2017 Abstract A basic combinatorial interpretation of Shannon’s entropy function is via the “20 questions” game. This cooperative game is played by two players, Alice and Bob: Alice picks a distribution π over the numbers {1, . . . , n}, and announces it to Bob. She then chooses a number x according to π, and Bob attempts to identify x using as few Yes/No queries as possible, on average. An optimal strategy for the “20 questions” game is given by a Huffman code for π: Bob’s questions reveal the codeword for x bit by bit. This strategy finds x using fewer than H(π) + 1 questions on average. However, the questions asked by Bob could be arbitrary. In this paper, we investigate the following question: Are there restricted sets of questions that match the performance of Huffman codes, either exactly or approximately? Our first main result shows that for every distribution π, Bob has a strategy that uses only questions of the form “x < c?” and “x = c?”, and uncovers x using at most H(π) + 1 questions on average, matching the performance of Huffman codes in this sense. We also give a natural set of O(rn1/r ) questions that achieve a performance of at most H(π) + r, and show that Ω(rn1/r ) questions are required to achieve such a guarantee. Our second main result gives a set Q of 1.25n+o(n) questions such that for every distribution π, Bob can implement an optimal strategy for π using only questions from Q. We also show that 1.25n−o(n) questions are needed, for infinitely many n. If we allow a small slack of r over the optimal strategy, then roughly (rn)Θ(1/r) questions are necessary and sufficient. 1 Introduction A basic combinatorial and operational interpretation of Shannon’s entropy function, which is often taught in introductory courses on information theory, is via the “20 questions” game (see for example the well-known textbook [7]). This game is played between two players, Alice and Bob: Alice picks a distribution π over a (finite) set of objects X, and announces it to Bob. Alice then chooses an object x according to π, and Bob attempts to identify the object using as few Yes/No queries as possible, on average.What questions should Bob ask? An optimal strategy for Bob is to compute a Huffman code for π, and then follow the corresponding decision tree: his first query, for example, asks whether x lies in the left subtree of the root. While this strategy minimizes the expected number of queries, the queries themselves could be arbitrarily complex; already for the first question, Huffman’s algorithm draws from an exponentially large reservoir of potential queries (see Theorem 5.3 for more details). Therefore, it is natural to consider variants of this game in which the set of queries used is restricted; for example, it is plausible that Alice and Bob would prefer to use queries that (i) can be communicated efficiently (using as few bits as possible), and (ii) can be tested efficiently (i.e. there is an efficient encoding scheme for elements of X and a fast algorithm that given x ∈ X and a query q as input, determines whether x satisfies the query q). 1 We summarize this with the following meta-question, which guides this work: Are there “nice” sets of queries Q such that for any distribution, there is a “high quality” strategy that uses only queries from Q? Formalizing this question depends on how “nice” and “high quality” are quantified. We consider two different benchmarks for sets of queries: 1. An information-theoretical benchmark: A set of queries Q has redundancy r if for every distribution π there is a strategy using only queries from Q that finds x with at most H(π)+r queries on average when x is drawn according to π. 2. A combinatorial benchmark: A set of queries Q is r-optimal (or has prolixity r) if for every distribution π there is a strategy using queries from Q that finds x with at most Opt(π) + r queries on average when x is drawn according to π, where Opt(π) is the expected number of queries asked by an optimal strategy for π (e.g. a Huffman tree). Given a certain redundancy or prolixity, we will be interested in sets of questions achieving that performance that (i) are as small as possible, and (ii) allow efficient construction of high quality strategies which achieve the target performance. In some cases we will settle for only one of these properties, and leave the other as an open question. Information-theoretical benchmark. Let π be a distribution over X. A basic result in information theory is that every algorithm that reveals an unknown element x drawn according to π (in short, x ∼ π) using Yes/No questions must make at least H(π) queries on average. Moreover, there are algorithms that make at most H(π) + 1 queries on average, such as Huffman coding and Shannon–Fano coding. However, these algorithms may potentially use arbitrary queries. Are there restricted sets of queries that match the performance of H(π) + 1 queries on average, for every distribution π? Consider the setting in which X is linearly ordered (say X = [n], with its natural ordering: 1 < · · · < n). Gilbert and Moore [14], in a result that paved the way to arithmetic coding, showed that two-way comparison queries (“x < c?”) almost fit the bill: they achieve a performance of at most H(π) + 2 queries on average. Our first main result shows that the optimal performance of H(π) + 1 can be achieved by allowing also equality queries (“x = c?”): Theorem (restatement of Theorem 3.1). For every distribution π there is a strategy that uses only comparison and equality queries which finds x drawn from π with at most H(π) + 1 queries on average. Moreover, this strategy can be computed in time O(n log n). In a sense, this result gives an affirmative answer to our main question above. The set of comparison and equality queries (first suggested by Spuler [30]) arguably qualifies as “nice”: linearly ordered universes appear in many natural and practical settings (numbers, dates, names, IDs) in which comparison and equality queries can be implemented efficiently. Moreover, from a communication complexity perspective, Bob can communicate a comparison/equality query using just log2 n + 1 bits (since there are just 2n such queries). This is an exponential improvement over the Ω(n) bits he would need to communicate had he used Huffman coding. We extend this result to the case where X is a set of vectors of length r, ~x = (x1 , . . . , xr ), by showing that there is a strategy using entry-wise comparisons (“xi < c?”) and entry-wise equalities (“xi = c?”) that achieves redundancy r: Theorem (restatement of Theorem 4.1). Let X be the set of vectors of length r over a linearly ordered universe. For every distribution π there is a strategy that uses only entry-wise comparison queries and entry-wise equality queries and finds ~x ∼ π with at most H(π) + r queries. 2 The theorem is proved by applying the algorithm of the preceding theorem to uncover the vector ~x entry by entry. As a corollary, we are able to determine almost exactly the minimum size of a set of queries that achieves redundancy r ≥ 1. In more detail, let uH (n, r) denote the minimum size of a set of queries Q such that for every distribution π on [n] there is a strategy using only queries from Q that finds x with at most H(π) + r queries on average, when x is drawn according to π. Corollary (Theorem 4.1). For every n, r ∈ N, 1 1/r rn ≤ uH (n, r) ≤ 2rn1/r . e Obtaining this tight estimate of uH (n, r) = Θ(rn1/r ) hinges on adding equality queries; had we used only entry-wise comparison queries and the Gilbert–Moore algorithm instead, the resulting upper bound would have been uH (n, r) = O rn2/r , which is quadratically worse than the truth. Combinatorial benchmark. The analytical properties of the entropy function make H(π) a standard benchmark for the average number of bits needed to describe an element x drawn from a known distribution π, and so it is natural to use it as a benchmark for the average number of queries needed to find x when it is drawn according to π. However, there is a conceptually simpler, and arguably more natural, benchmark: Opt(π) — the average number of queries that are used by a best strategy for π (several might exist), such as one generated by Huffman’s algorithm. Can the optimal performance of Huffman codes be matched exactly? Can it be achieved without using all possible queries? Our second main result answers this in the affirmative: Theorem (restatement of Theorem 5.2 and Theorem 5.3). For every n there is a set Q of 1.25n+o(n) queries such that for every distribution over [n], there is a strategy using only queries from Q which matches the performance of the optimal (unrestricted) strategy exactly. Furthermore, for infinitely many n, at least 1.25n−o(n) queries are required to achieve this feat. One drawback of our construction is that it is randomized. Thus, we do not consider it particularly “efficient” nor “natural”. It is interesting to find an explicit set Q that achieves this bound. Our best explicit construction is: Theorem (restatement of Theorem 5.5). For every n there is an explicit set Q of O(2n/2 ) queries such that for every distribution over [n], there is a strategy using only queries from Q which matches the performance of the optimal (unrestricted) strategy exactly. Moreover, we can compute this strategy in time O(n2 ). It is natural to ask in this setting how small can a set of queries be if it is r-optimal; that is, if it uses at most Opt(π) + r questions on average when the secret element is drawn according to π, for small r > 0. Let uOpt (n, r) denote the minimum size of a set of queries Q such that for every distribution π on [n] there is a strategy using only queries from Q that finds x with at most Opt(π)+r queries on average when x is drawn from π. We show that for any fixed r > 0, significant savings can be achieved: Theorem (restatement of Theorem 6.1). For all r ∈ (0, 1): 1 16 (r · n) 4r . uOpt (n, r) . (r · n) r . Instead of the exponential number of questions needed to match Huffman’s algorithm exactly, for fixed r > 0 an r-optimal set of questions has polynomial size. In this case the upper bound is achieved by an explicit set of queries Qr . We also present an efficient randomized algorithm for computing an r-optimal strategy that uses queries from Qr . 3 Related work The “20 questions” game is the starting point of combinatorial search theory [18, 1, 3]. Combinatorial search theory considers many different variants of the game, such as several unknown elements, non-adaptive queries, non-binary queries, and a non-truthful Alice [28, 29, 2, 9]. Both average-case and worst-case complexity measures are of interest. An important topic in combinatorial search theory is combinatorial group testing [10, 11]. We are unaware of any prior work which has considered the quantities uH (n, r) or uOpt (n, r). However, several particular sets of questions have been analyzed in the literature from the perspective of redundancy and prolixity: Huffman codes [13, 17, 5, 24, 21, 23], binary search trees [14, 26, 27, 15], and comparison-based sorting algorithms [12, 25]. 2 Preliminaries Notation We use log n for the base 2 logarithm of n and [n] to denote the set {1, . . . , n}. Throughout the document, we will consider probability distributions over the set Xn = {x1 , . . . , xn } of size n. In some cases, we will think of this set as ordered: x1 ≺ · · · ≺ xn . If π is a probability distribution over Xn , we will denote the probability of xi by πi , and the probability of a set S ⊆ Xn by π(S). Information theory We use H(π) to denote the base 2 entropy of a distribution π and D(πkµ) to denote the Kullback–Leibler divergence. The binary entropy function h(p) is the entropy of a Bernoulli random variable with success probability p. When Y is a Bernoulli random variable, the chain rule takes the following form: H(X, Y ) = h(Pr[Y = 1]) + Pr[Y = 0]H(X|Y = 0) + Pr[Y = 1]H(X|Y = 1). We call this the Bernoulli chain rule. Decision trees In this paper we consider the task of revealing a secret element x from Xn by using Yes/No questions. Such a strategy will be called a decision tree or an algorithm. A decision tree is a binary tree in which the internal nodes are labeled by subsets of Xn (which we call questions or queries), each internal node has two outgoing edges labeled Yes (belongs to the question set) and No (doesn’t belong to the question set), and the leaves are labeled by distinct elements of Xn . The depth of an element xi in a decision tree T , denoted T (xi ), is the number of edges in the unique path from the root to the unique leaf labeled xi (if any). Decision trees can be thought of as annotated prefix codes: the code of an element xi is the concatenation of the labels of the edges leading to it. The mapping can also be used in the other direction: each binary prefix code of cardinality n corresponds to a unique decision tree over Xn . Given a set Q ⊆ 2Xn (a set of allowed questions), a decision tree using Q is one in which all questions belong to Q. A decision tree is valid for a distribution µ if its leaves are labeled by elements of the support of µ, and each element in the support appears as the label of some leaf. Given a distribution µ and a decision tree T valid for µ, the cost (or query complexity) P of T on µ, labeled T (µ), is the average number of questions asked on a random element, T (µ) = ni=1 µi T (xi ). Given a set Q of allowed questions and a distribution µ, the optimal cost of µ with respect to Q, denoted c(Q, µ), is the minimal cost of a valid decision tree for µ using Q. Dyadic distributions and Huffman’s algorithm Huffman’s algorithm [16] finds the optimal cost of an unrestricted decision tree for a given distribution: Opt(µ) = c(2Xn , µ). 4 We call a decision tree with this cost a Huffman tree or an optimal decision tree for µ. More generally, a decision tree is r-optimal for µ if its cost is at most Opt(µ) + r. It will be useful to consider this definition from a different point of view. Say that a distribution is dyadic if the probability of every element is either 0 or of the form 2−d for some integer d. We can associate with each decision tree T a distribution τ on the leaves of T given by τi = 2−T (xi ) . This gives a correspondence between decision trees and dyadic distributions. In the language of dyadic distributions, Huffman’s algorithm solves the following optimization problem: n X 1 [H(µ) + D(µkτ )] . min Opt(µ) = min µi log = τ dyadic τ dyadic τi supp(τ )=supp(µ) i=1 supp(τ )=supp(µ) In other words, computing Opt(µ) amounts to minimizing D(µkτ ), and thus to “rounding” µ to a dyadic distribution. We call τ a Huffman distribution for µ. The following classical inequality shows that Opt(µ) is very close to the entropy of µ: H(µ) ≤ Opt(µ) < H(µ) + 1. The lower bound follows from the non-negativity of the Kullback–Leibler divergence; it is tight exactly when µ is dyadic. The upper bound from the Shannon–Fano code, which corresponds to the dyadic sub-distribution τi = 2−⌈log µi ⌉ (in which the probabilities could sum to less than 1). Redundancy and prolixity We measure the quality of sets of questions by comparing the cost of decision trees using them to the entropy (the difference is known as redundancy) and to the cost of optimal decision trees (for the difference we coin the term prolixity). In more detail, the redundancy of a decision tree T for a distribution µ is T (µ) − H(µ), and its prolixity is T (µ) − Opt(µ). Given a set of questions Q, the redundancy r H (Q, µ) and prolixity r Opt (Q, µ) of a distribution µ are the best redundancy and prolixity achievable using questions from Q: r H (Q, µ) = c(Q, µ) − H(µ), r Opt (Q, µ) = c(Q, µ) − Opt(µ). The redundancy of a set of questions Q, denoted r H (Q), is the supremum of r H (Q, µ) over all distributions µ over Xn . The prolixity r Opt (Q) of Q is defined similarly. These quantities are closely related, as the inequality H(µ) ≤ Opt(µ) < H(µ) + 1 implies: r Opt (Q) ≤ r H (Q) ≤ r Opt (Q) + 1. A set of questions Q is optimal if r Opt (Q) = 0, and r-optimal if r Opt (Q) ≤ r. The parameters uH (n, r) and uOpt (n, r) Our main object of study in this paper are the parameters uH (n, r) and uOpt (n, r). The parameter uH (n, r) is the cardinality of the minimal set of questions Q ⊆ 2Xn such that r H (Q) ≤ r. Similarly, the parameter uOpt (n, r) is the cardinality of the minimal set of questions Q such that r Opt (Q) ≤ r. These quantities are closely related: uOpt (n, r) ≤ uH (n, r) ≤ uOpt (n, r − 1). 5 A useful lemma The following simple lemma will be used several times in the rest of the paper. Lemma 2.1. Let p1 ≥ . . . ≥ pn be a non-increasing list of numbers of the form P pi = 2−ai (for Pn −a −a integer ai ), and let P a ≤ a1 be an integer. If i=1 pi ≥ 2 then for some m we have m i=1 pi = 2 . P n n If furthermore i=1 pi is a multiple of 2−a then for some ℓ we have i=ℓ pi = 2−a . P −a Proof. Let m be the maximalP index such that m i=1 pi ≤ 2 . If m = n then we are done, so m suppose that m < n. Let S = i=1 pi . We would like to show that S = 2−a . P am+1 −ai The condition p1 ≤ · · · ≤ pn implies that am+1 ≥ · · · ≥ a1 , and so k := 2am+1 S = m i=1 2 P m+1 a −a a a −a is an integer. By assumption k ≤ 2 m+1 whereas k+1 = 2 m+1 i=1 pi > 2 m+1 . Since 2am+1 −a is itself an integer (since am+1 ≥ a1 ≥ a), we conclude that k = 2am+1 −a , and so S = 2−a . To prove the furthermore part, notice that by repeated applications of the of the lemma we can partition [n] into intervals whose probabilities are 2−a . The last such interval provides the required index ℓ. 3 Comparisons and equality tests Let π be a distribution over Xn = {x1 , . . . , xn }. A fundamental result in information theory is that the entropy of a distribution π captures the average number of queries needed to identify a random x ∼ π. More specifically, every algorithm asks at least H(π) questions in expectation, and there are algorithms that ask at most H(π) + 1 questions on average (such as Huffman coding and Shannon–Fano coding). However, these algorithms may potentially use arbitrary questions. In this section we are interested in the setting where Xn is linearly ordered: x1 ≺ x2 ≺ · · · ≺ xn . We wish to use questions that are compatible with the ordering. Perhaps the most natural question in this setting is a comparison query; namely a question of the form “x ≺ xi ?”. Gilbert and Moore [14] showed that there exists an algorithm that uses at most H(π) + 2 comparisons. Is this tight? Can comparison queries achieve the benchmark of H(π) + 1? A simple argument shows that their result is tight: let n = 3, and let π be a distribution such that π(x1 ) = ǫ/2, π(x2 ) = 1 − ǫ, π(x3 ) = ǫ/2,  for some small ǫ. Note that H(π) = O(ǫ log 1/ǫ ), and therefore any algorithm with redundancy 1 must use the query “x = x2 ?” as its first question. This is impossible if we only allow comparison queries (see Lemma 4.2.1 for a more detailed and general argument). In fact, this shows that any set of questions that achieves redundancy 1 must include all equality queries. So, we need to at least add all equality queries. Is it enough? Do comparison and equality queries achieve redundancy of at most 1? Our main result in this section gives an affirmative answer to this question: (n) (n) Theorem 3.1. Let Q= = {{xi } : 1 ≤ i ≤ n} and Q≺ = {{x1 , . . . , xi } : 1 ≤ i ≤ n − 1}. In (n) (n) other words, Q= consists of the questions “x = xi ?” for i ∈ {1, . . . , n}, and Q≺ consists of the questions “x ≺ xi ?” for i ∈ {2, . . . , n}. (Recall that x is the secret element.) (n) (n) For all n, r H (Q≺ ∪ Q= ) = 1. We prove the theorem by modifying the weight-balancing algorithm of Rissanen [27], which uses only comparison queries and achieves redundancy 2 (as shown by Horibe [15]). The original weight-balancing algorithm is perhaps the first algorithm that comes to mind: it asks the most balanced comparison query (the one that splits the distribution into two parts whose probability is as equal as possible), and recurses according to the answer. 6 Our modified algorithm, Algorithm At , first checks whether the probability of the most probable element xmax exceeds the threshold t. If so, it asks the question “x = xmax ?”. Otherwise, it proceeds as in the weight-balancing algorithm. The choice t = 0.3 results in an algorithm whose redundancy is at most 1. While a naive implementation of Algorithm At takes time O(n) to determine the next query, this can be improved to O(log n), given O(n) preprocessing time. Moreover, the entire strategy can be computed in time O(n log n). The interested reader can find the complete details in [8]. 3.1 Algorithm At Let π be a distribution over Xn . Let xmax denote the most probable element, and let πmax denote its probability. Let xmid denote a point xi that minimizes |π({x : x ≺ xi }) − 1/2| over i ∈ [n]. We call xmid the middle1 of π. Note that the query “x ≺ xmid ?” is the most balanced query among all queries of the form “x ≺ xi ?”. Let A ⊆ {x1 , . . . , xn }. We use πA to denote π(A); i.e. the probability of A. Specifically, we use π≺xi , πxi , π6=xi to denote πA when A is {x : x ≺ xi }, {x : x  xi }, {x : x 6= xi }. We use π|A to denote the restriction of π to A; i.e., the distribution derived by conditioning π on A. Specifically, we use π|≺xi , π|xi , π|6=xi to denote π|A when A is {x : x ≺ xi }, {x : x  xi }, {x : x 6= xi }. Algorithm At . Given a threshold t ∈ (0, 1), Algorithm At takes as input a distribution π over Xn and a secret element x, and determines x using only comparison and equality queries, in the following recursive fashion: 1. If π(xi ) = 1 for some element xi , then output xi . 2. If πmax ≥ t then ask whether x = xmax , and either output xmax , or continue with π|6=xmax . 3. If πmax < t, ask whether x ≺ xmid , and continue with either π|≺xmid or π|xmid . (When recursing on a domain D, we identify D with X|D| .) ⊳ The weight balancing algorithm of Rissanen [27] is the special case t = 1; in this case no equality queries are needed, and the resulting redundancy is 2, as shown by Horibe [15]. We will show that for a certain range of values of t (for example, t = 0.3), Algorithm At achieves redundancy at most 1, thus proving Theorem 3.1. Recall that At (π) is the cost of At on π, and let Rt (π) := At (π)−H(π)−1. It is more convenient to present the proof in terms of Rt (π) rather than in terms of the redundancy. Our goal, stated in these terms, is showing that there exists some t for which Rt (π) ≤ 0 for all distributions π. We next observe two simple properties of the algorithm At . The proof of Theorem 3.1 relies only on these properties. The first property is a recursive definition of Rt that is convenient to induct on. See Figure 1 for a pictorial illustration. Lemma 3.1.1. Let π be a distribution over Xn . Then   −1 Rt (π) = 1 − h(πmax ) − πmax + (1 − πmax )Rt (π|6=xmax )   1 − h(π≺xmid ) + π≺xmid Rt (π|≺xmid ) + πxmid Rt (π|xmid ) 1 Note that one of {xmid , xmid−1 } is a median. 7 if πmax = 1, if πmax ∈ [t, 1), if πmax ∈ (0, t). “x 6= xmax ?” “x ≺ xmid ?” πmax πxmid π≺xmid 1 − h(πxmid ) xmax 1 − h(π6=xmax ) 1 − πmax Rt (π|=xmax ) = −1 Rt (π|6=xmax ) Rt (π|xmid ) Rt (π|≺xmid ) (a) πmax ∈ (0, t) (b) πmax ∈ [t, 1) Figure 1: Recursive definition of Rt Proof. If πmax = 1 then At (π) = 0 and H(π) = 0, so Rt (π) = −1. If πmax ∈ [t, 1) then Rt (π) = At (π) − H(π) − 1 i h i i h h = 1 + (1 − πmax )At (π|6=xmax ) − h(πmax ) + (1 − πmax )H(π|6=xmax ) − πmax + (1 − πmax ) i h = 1 − h(πmax ) − πmax + (1 − πmax ) At (π|6=xmax ) − H(π|6=xmax ) − 1 = 1 − h(πmax ) − πmax + (1 − πmax )Rt (π6=xmax ). If πmax ∈ (0, t) then Rt (π) = At (π) − H(π) − 1 h i = 1 + π≺xmid At (π|≺xmid ) + πxmid At (π|xmid ) i h h i − h(π≺xmid ) + π≺xmid H(π|≺xmid ) + πxmid H(π|xmid ) − π≺xmid + πxmid h i h i = 1 − h(π≺xmid ) + π≺xmid At (π|≺xmid ) − H(π|≺xmid ) − 1 + πxmid At (π|xmid ) − H(π|xmid ) − 1 = 1 − h(π≺xmid ) + π≺xmid Rt (π|≺xmid ) + πxmid Rt (π|xmid ). The second property is that whenever At uses a comparison query (i.e. when πmax < t), then this question is balanced: Lemma 3.1.2. Let π be a distribution over Xn . Then π≺xmid , πxmid ∈ [ 1−π2max , 1+π2max ]. Proof. By the definition of xmid , it suffices to show that there exists some j with π≺j ∈ [ 1−π2max , 1+π2max ]. Indeed, for all j, π≺j+1 −π≺j = πj ≤ πmax , and therefore, if j is the maximum element with π≺j < 12 (possibly j = 0), then either π≺j or π≺j+1 are in [ 1−π2max , 1+π2max ]. 3.2 Game Gt We will use the properties described in Lemma 3.1.2 to bound Rt (π). Since this involves quantifying over all distributions π, it is convenient to introduce a game that simulates At on a distribution 8 π. The game involves one player, Alice, who we think of as an adversary that chooses the input distribution π (in fact, she only chooses πmax ), and wins a revenue of Rt (π) (thus, her objective is to maximize the redundancy). This reduces our goal to showing that Alice’s optimum revenue is nonpositive. The definition of the game is tailored to the properties stated in Lemma 3.1.1 and Lemma 3.1.2. We first introduce Gt , and then relate it to the redundancy of At (see Lemma 3.3.1 below). Game Gt . Let t ≤ 31 , and let f, s : (0, 1] → R be f (x) := 1 − h(x) − x and s(x) := 1 − h(x). The game Gt consists of one player called Alice, whose objective is to maximize her revenue. The game Gt begins at an initial state p ∈ (0, 1], and proceeds as follows. 1. If p ∈ [t, 1], the game ends with revenue f (p).  2p 2p  , 1−p and recursively plays Gt with initial 2. If p ∈ (0, t), then Alice chooses a state2 p′ ∈ 1+p ′ ′ state p . Let r denote her revenue in the game that begins at p′ . Alice’s final revenue is   p p s ′ + ′ · r′. ⊳ p p Note that given any initial state p and a strategy for Alice, the game Gt always terminates: indeed, if p = p0 , p1 , p2 , . . . is the sequence of states chosen by Alice, then as long as pi < t, it holds 2pi 2pi > 1+1/3 = 32 pi (the second inequality is since pi < t ≤ 13 ). So the sequence of that pi+1 ≥ 1+p i states grows at an exponential rate, which means that for ℓ = O(log(1/p0 )), the state pℓ exceeds the threshold t and the game terminates. For p ∈ (0, 1], let rt (p) denote the supremum of Alice’s revenue in Gt when the initial state is p, the supremum ranging over all possible strategies for Alice. Our next step is using the game Gt to prove Theorem 3.1: We will show that t = 0.3 satisfies: (i) rt (p) ≤ 0 for all p ∈ (0, 1], and (ii) Rt (π) ≤ rt (πmax ) for all π. Note that (i) and (ii) imply Theorem 3.1. Before establishing (i) and (ii), we state and prove three simple lemmas regarding Gt that are useful to this end. The first lemma will be used in the proof of Lemma 3.3.1, which shows that if rt (p) ≤ 0 for all p ∈ (0, 1], then Rt (π) ≤ rt (πmax ). Its proof follows directly from the definition of rt .  2p 2p  , 1−p : Lemma 3.2.1. For all p < t and all p′ ∈ 1+p   p p + ′ · rt (p′ ). rt (p) ≥ s ′ p p The next two lemmas will be used in the proof of Lemma 3.3.2, which shows that rt (p) ≤ 0 for all p ∈ (0, 1] when t = 0.3. The first one gives a tighter estimate on the growth of the sequence of states: Lemma 3.2.2. Let p0 , p1 , . . . , pk be the sequence of states chosen by Alice. For every i ≤ k − 1: pk−1−i < 2 Note that p′ ∈ (0, 1], since p < t ≤ 1 3 implies that  2i (1 t . − t) + t 2p , 2p 1+p 1−p 9  ⊆ (0, 1]. Proof. We prove the bound by induction on i. The case i = 0 follows from pk−1 being a non-final state, and thereforepk−1 < t as required. Assume now that i > 0. By the of Gt it follows  definition pk−1−(i−1)  pk−1−(i−1) 2pk−1−i  2p , , which implies that p ∈ that pk−1−(i−1) ∈ 1+pk−1−i k−1−i 2+pk−1−(i−1) , 2−pk−1−(i−1) . k−1−i 1−pk−1−i Therefore, pk−1−(i−1) pk−1−i ≤ 2 − pk−1−(i−1)  t/ 2i−1 (1 − t) + t  (by induction hypothesis on i − 1) < 2 − t/ 2i−1 (1 − t) + t t = i . 2 (1 − t) + t The second lemma gives a somewhat more explicit form of the revenue of Gt : Lemma 3.2.3. Let p0 , p1 , . . . , pk be the sequence of states chosen by Alice. Let r(p0 , . . . , pk ) denote the revenue obtained by choosing these states. Then   k−1 X p0 pi p0 r(p0 , . . . , pk ) = s + f (pk ). pi pi+1 pk i=0 Proof. We prove the formula by induction on k. If k = 0 then pk ≥ t and r(pk ) = f (pk ) = ppk0 f (pk ). When k ≥ 1:   p0 p0 · r(p1 , . . . , pk ) (by definition of Gt ) + r(p0 , . . . , pk ) = s p1 p1     k−1  p0 p0 pi p0 X p1 p1 = s · s (by induction hypothesis) + + f (pk ) p0 p1 p1 p pi+1 pk i=1 i   k−1 X pi p0 p0 s + f (pk ). = pi pi+1 pk i=0 3.3 Relating At to Gt Next, we relate the revenue in Gt to the redundancy of At by linking rt and Rt . We first reduce the Theorem 3.1 to showing that there exists some t ∈ (0, 1] such that rt (p) ≤ 0 for all p ∈ (0, 1] (Lemma 3.3.1), and then show that t = 0.3 satisfies this condition (Lemma 3.3.2). Lemma 3.3.1. Let t be such that rt (p) ≤ 0 for all p ∈ (0, 1]. For every distribution π, Rt (π) ≤ rt (πmax ). In particular, such t satisfies Rt (π) ≤ 0 for all π. Proof. We proceed by induction on the size of supp(π) = {xi : π(xi ) 6= 0}. If supp(π) = 1 then πmax = 1, and therefore Rt (π) = −1, rt (πmax ) = 0, and indeed Rt (π) ≤ rt (πmax ). Assume now that supp(π) = k > 1. Since k > 1, it follows that πmax < 1. There are two cases, according to whether πmax ∈ (0, t) or πmax ∈ [t, 1). If πmax ∈ (0, t) then At asks whether x ≺ xmid , and continues accordingly with π|≺xmid or π|xmid . Let σ := π|≺xmid and τ := π|xmid . By Lemma 3.1.1: Rt (π) = 1 − h(π≺xmid ) + π≺xmid Rt (σ) + πxmid Rt (τ ) ≤ 1 − h(π≺xmid ) + π≺xmid rt (σmax ) + πxmid rt (τmax ) 10 (since πmax ∈ (0, t)) (by induction hypothesis) Without loss of generality, assume that xmax ≺ xmid . Therefore σmax = πmax /π≺xmid , and by Lemma 3.1.2:   2πmax 2πmax σmax ∈ , . (1) 1 + πmax 1 − πmax Thus, Rt (π) ≤ 1 − h(π≺xmid ) + π≺xmid rt (σmax )   πmax πmax rt (σmax ) =1−h + σmax σmax   πmax πmax =s rt (σmax ) + σmax σmax ≤ rt (πmax ). (since rt (τmax ) ≤ 0) (σmax = πmax π≺xmid ) (by definition of s) (by (1) and Lemma 3.2.1) The analysis when πmax ∈ [t, 1) is very similar. In this case At asks whether x = xmax , and continues with π|6=xmax if x 6= xmax . Let σ := π|≤xmax . By Lemma 3.1.1, Rt (π) = 1 − h(πmax ) − πmax + (1 − πmax )Rt (σ) ≤ 1 − h(πmax ) − πmax + (1 − πmax )rt (σmax ) (since πmax ∈ (t, 1)) (by induction hypothesis) ≤ 1 − h(πmax ) − πmax (since rt (σmax ) ≤ 0) = f (πmax ) (by definition of f ) (by definition of rt , since πmax ≥ t) = rt (πmax ). The following lemma shows that t = 0.3 satisfies rt (p) ≤ 0 for all p ∈ (0, 1], completing the proof of Theorem 3.1. It uses some technical results, proved below in Lemma 3.3.3. Lemma 3.3.2. Let t = 0.3. Then rt (p) ≤ 0 for all p ∈ (0, 1]. Proof. Let p ∈ (0, 1]. We consider two cases: (i) p ≥ t, and (ii) p < t. In each case we derive a constraint on t that suffices for ensuring that rt (p) ≤ 0, and conclude the proof by showing that t = 0.3 satisfies both constraints. Consider the case t ≤ p. Here rt (p) = f (p) = 1 − h(p) − p, and calculation shows that f (p) is non-positive on [0.23, 1]; therefore, rt (p) ≤ 0 for all p ≥ t, as long as t ≥ 0.23. Consider the case p < t. Here, we are not aware of an explicit formula for rt (p); instead, we derive the following upper bound, for all p < t: ∞ rt (p) X S ≤ p n=0  t n 2 (1 − t) + t where    2t  , + max F (t), F 1−t  (2)  1+x f (x) 2 S(x) = , F (x) = . x x With (2) in hand we are done: indeed, calculation shows that the right hand side of (2) is nonpositive in some neighborhood of 0.3 (e.g. it is −0.0312 when t = 0.3, it is −0.0899 when t = 0.294); thus, as these values of t also satisfy the constraint from (i), this finishes the proof. s 11 It remains to prove (2). Let p = p0 , p1 , p2 , . . . , pk be a sequence  of states  chosen by Alice. It P∞ r(p0 ,...,pk ) t 2t ≤ n=0 S 2n (1−t)+t + max F (t), F 1−t . By Lemma 3.2.3: suffices to show that p0   k−1 X p0 pi p0 r(p0 , . . . , pk ) = s + f (pk ) p pi+1 pk i=0 i   k−1 X p0 1 + pi p0 ≤ s + f (pk ) pi 2 pk i=0 = p0 k−1 X i=0  S(pi ) + F (pk ) , i 1+pi where in the second line we used that pi+1 ∈ [ 1−p 2 , 2 ], and the fact that s(x) = 1 − h(x) is symmetric around x = 0.5 and increases with |x − 0.5|. Therefore, pi k−1 r(p0 , . . . , pk ) X S(pi ) + F (pk ) ≤ p0 i=0  k−1  X t S ≤ + F (pk ) 2i (1 − t) + t i=0    ∞  2t  X t , + max F (t), F S ≤ 2i (1 − t) + t 1−t i=0 where in the second last inequality we used that pk−1−i < t 2i (1−t)+t (Lemma 3.2.2) and that S(x) 2t ) and that is monotone (Lemma 3.3.3 below), and in the last inequality we used that pk ∈ [t, 1−t F (x) is convex (Lemma 3.3.3 below). The following technical lemma completes the proof of Lemma 3.3.2. Lemma 3.3.3. The function S(x) = convex. ) 1−h( 1+x 2 x is monotone, and the function F (x) = 1−h(x)−x x is Proof. The function h( 1−x 2 ) is equal to its Maclaurin series for x ∈ (−1, +1): ∞ 1 + x X log2 e =1− · x2k . h 2 2k(2k − 1) k=1 Therefore, S(x) = ∞ X k=1 and F (x) = ∞ X k=1 log2 e · x2k−1 , 2k(2k − 1) log2 e (1 − 2x)2k  − 1. · 2k(2k − 1) x Now, each of the functions x2k−1 is monotone, and each of the functions (0, ∞): its second derivative is (1−2x)2k x is convex on 2(1 − 2x)2k−2 (1 + 4(k − 1)x + 4(k − 1)(2k − 1)x2 ) > 0. x3 Therefore, S(x) is monotone as a non-negative combination of monotone functions, and F (x) is convex as a non-negative combination of convex functions. 12 4 Information theoretical benchmark — Shannon’s entropy In this section we study the minimum number of questions that achieve redundancy of at most r, for a fixed r ≥ 1. Note that r = 1 is the optimal redundancy: the distribution π on X2 given by π1 = 1 − ǫ, π2 = ǫ has redundancy 1 − Õ(ǫ) (that is, 1 − O(ǫ log(1/ǫ)) even without restricting the set of allowed questions. In the previous section we have shown that the optimal redundancy of r = 1 can be achieved with just 2n comparison and equality queries (in fact, as we show below, there are only 2n − 3 of these queries). It is natural to ask how small the number of questions can be if we allow for a larger r. Note that at least log n questions are necessary to achieve any finite redundancy. Indeed, a smaller set of questions is not capable of specifying all elements even if all questions are being asked. The main result of this section is that the minimum number of questions that are sufficient for achieving redundancy r is roughly r · n1/⌊r⌋ : Theorem 4.1. For every r ≥ 1 and n ∈ N, 1 ⌊r⌋n1/⌊r⌋ ≤ uH (n, r) ≤ 2⌊r⌋n1/⌊r⌋ . e  In particular, uH (n, r) = Θ ⌊r⌋n1/⌊r⌋ . 4.1 Upper bound The upper bound in Theorem 4.1 is based on the following corollary of Theorem 3.1: Theorem 4.1.1. Let Y be a linearly ordered set, and let Z = Y k (we don’t think of Z as ordered). For any distribution π on Z there is an algorithm that uses only questions of the form (i) “~xi ≺ y?” and (ii) “~xi = y?”, where i ∈ [k] and y ∈ Y , whose cost is at most H(π) + k. Proof. Let Z1 Z2 . . . Zk ∼ π. Consider the algorithm which determines Z1 , . . . , Zk in order, where Zi is determined by applying the algorithm from Theorem 3.1 on “Zi |Z1 . . . Zi−1 ”, which is the conditional distribution of Zi given the known values of Z1 , . . . , Zi−1 . The expected number of queries is at most    H(Z1 ) + 1 + H(Z2 |Z1 ) + 1 + · · · + H(Zk |Z1 . . . Zk−1 ) + 1 = H(Z1 . . . Zk ) + k, using the chain rule. We use this theorem to construct a set of questions of size at most 2⌊r⌋n1/⌊r⌋ that achieves redundancy r for any distribution over Xn .  ⌊r⌋ Note that n ≤ n1/⌊r⌋ . Therefore every element x ∈ Xn can be represented by a vector  ⌊r⌋ . Let Q be the set containing all questions of the form (i) “~xi = y?” and ~x ∈ 1, . . . , ⌈n1/⌊r⌋ ⌉ (ii) “~xi ≺ y?”. By Theorem 4.1.1, r(Q) = ⌊r⌋. The following questions from Q are redundant, and can be removed from Q without increasing its redundancy: (i) “~xi ≺ 1?” (corresponds to the empty set and therefore provides no information), (ii) “~xi ≺ 2?” (equivalent to the question “~xi = 1?”), and (iii) “~xi ≺ ⌈n1/⌊r⌋ ⌉?” (equivalent to the question “~xi = ⌈n1/⌊r⌋ ⌉?”). The number of remaining questions is       ⌊r⌋ · 2 n1/⌊r⌋ − 3 ≤ 2⌊r⌋ · n1/⌊r⌋ . This proves the upper bound in Theorem 4.1. 13 4.2 Lower bound The crux of the proof of the lower bound in Theorem 4.1 is that if Q is a set of questions whose redundancy is at most r then every x ∈ Xn can be identified by at most ⌊r⌋ questions from Q. We say that the questions q1 , . . . , qT identify x if for every y 6= x there is some i ≤ T such that qi (x) 6= qi (y). Define t(n, r) to be the minimum cardinality of a set Q of questions such that every x ∈ X has at most r questions in Q that identify it. The quantity t(n, r) can be thought of as a non-deterministic version of u(n, r): it is the minimal size of a set of questions so that every element can be “verified” using at most r questions. The lower bound on u(n, r) follows from Lemma 4.2.1 and Lemma 4.2.2 below. Lemma 4.2.1. For all n, r, u(n, r) ≥ t(n, ⌊r⌋). Proof. It suffices to show that for every set of questions Q with redundancy at most r, every x ∈ X has at most ⌊r⌋ questions in Q that identify it. Consider the distribution π given by π(x) = 1 − ǫ and π(y) = ǫ/(n − 1) for y 6= x. Thus H(π) = Õ(ǫ). Consider an algorithm for π with redundancy r that uses only questions from Q. Let T be the number of questions it uses to find x. The cost of the algorithm is at least (1 − ǫ)T , ǫ )r. For small enough ǫ > 0, and so (1 − ǫ)T ≤ H(π) + r = Õ(ǫ) + r, implying T ≤ Õ(ǫ) + (1 + 1−ǫ the right-hand side is smaller than ⌊r⌋ + 1, and so T ≤ ⌊r⌋. Lemma 4.2.1 says that in order to lower bound u(n, r), it suffices to lower bound t(n, ⌊r⌋), which is easier to handle. For example, the following straightforward argument shows that t(n, R) ≥ 1 1/R , for every R, n ∈ N. Assume Q is a set of questions of size u(n, R) so that every x is 2e Rn identified by at most R questions. This implies an encoding (i.e. a one-to-one mapping) of x ∈ Xn by the R questions identifying it, and by the bits indicating whether x satisfies each of these questions. Therefore   |Q| R  2e|Q| R , n≤ 2 ≤ R ≤R k m for all m, k. This implies that t(n, ⌊r⌋) ≥ where in the last inequality we used that ≤k ≤ em k 1 1 1/⌊r⌋ 1/⌊r⌋ . The constant 2e in front of ⌊r⌋n can be increased to 1e , using an algebraic argument: 2e ⌊r⌋n Lemma 4.2.2. For all n, R ∈ N: 1 R · n1/R . e Proof. We use the so-called polynomial method. Let Q be a set of questions such that each x ∈ X can be identified by atmost R queries. For each x ∈ X, let ux be the |Q|-dimensional vector ux = q1 (x), . . . , q|Q| (x) , and let U = {ux : x ∈ X} ⊆ {0, 1}|Q| . We will show that every function F : U → F2 can be represented as a multilinear polynomial of degree at most R in |Q| variables. Since the dimension over F2 of all such functions is n, whereas the dimension of the space of all |Q|  multilinear polynomials of degree at most R is ≤R , the bound follows:    |Q| e|Q| R 1 n≤ ≤ =⇒ n ≥ R · n1/R . ≤R R e t(n, R) ≥ It is enough to show that for any ux ∈ U , the corresponding “delta function” δx : U → F2 , defined as δx (ux ) = 1 and δx (v) = 0 for ux 6= v ∈ U , can be represented as a polynomial of degree at most d. Suppose that qi1 , . . . , qiT are T ≤ R questions that identify x. Consider the polynomial P (y1 , . . . , y|Q| ) = (yi1 − qi1 (x) + 1) · · · (yir − qir (x) + 1). 14 Clearly P (ux ) = 1. On the other hand, if P (uy ) = 1 then qij (y) = qij (x) for all j, showing that y = x. So P = δx , completing the proof. Our proof of the lower bound in Theorem 4.1 is based on t(n, R), which is the minimum cardinality of a set of queries such that each element can be identified by at most R questions. This quantity is closely related to witness codes [22, 6]; see [8] for more details. 5 Combinatorial benchmark — Huffman codes Section 3 shows that the optimal redundancy, namely 1, can be achieved using only O(n) questions. However, it is natural to ask for an instance-optimal algorithm? That is, we are looking for a set of questions which matches the performance of minimum redundancy codes such as Huffman codes. Let us repeat the definition of an optimal set of questions that is central in this section. Definition 5.1. A set Q of subsets of Xn is an optimal set of questions over Xn if for all distributions µ on Xn , c(Q, µ) = Opt(µ). Using the above definition, uOpt (n, 0) is equal to the minimal size of an optimal set of questions over Xn . Perhaps surprisingly, the trivial upper bound of 2n−1 on uOpt (n, 0) can be exponentially improved: Theorem 5.2. We have uOpt (n, 0) ≤ 1.25n+o(n) . We prove a similar lower bound, which is almost tight for infinitely many n: Theorem 5.3. For n of the form n = 5 · 2m , √ uOpt (n, 0) ≥ 1.25n /O( n). For all n, √ uOpt (n, 0) ≥ 1.232n /O( n). Corollary 5.4. We have lim sup n→∞ log uOpt (n, 0) = log 1.25. n Unfortunately, the construction in Theorem 5.2 is not explicit. A different construction, which √ n uses O( 2 ) questions, is not only explicit, but can also be implemented efficiently: Theorem 5.5. Consider the set of questions Q = {A ⊆ Xn : A ⊆ X⌈n/2⌉ or A ⊇ X⌈n/2⌉ }. The set Q consists of 2⌈n/2⌉ + 2⌊n/2⌋ questions and satisfies the following properties: 1. There is an indexing scheme Q = {Qq : q ∈ {0, 1}⌈n/2⌉+1 } such that given an index q and an element xi ∈ Xn , we can decide whether xi ∈ Qq in time O(n). 2. Given a distribution π, we can construct an optimal decision tree for π using Q in time O(n2 ). 3. Given a distribution π, we can implement an optimal decision tree for π in an online fashion in time O(n) per question, after O(n log n) preprocessing. 15 Section organization. Section 5.1 shows that a set of questions is optimal if and only if it is a dyadic hitter, that is, contains a question splitting every non-constant dyadic distribution into two equal halves. Section 5.2 discusses a relation to hitting sets for maximal antichains, and proves Theorem 5.5. Section 5.3 shows that the optimal size of a dyadic hitter is controlled by the minimum value of another parameter, the maximum relative density. We upper bound the minimum value in Section 5.4, thus proving Theorem 5.3, and lower bound it in Section 5.5, thus proving Theorem 5.2. 5.1 Reduction to dyadic hitters The purpose of this subsection is to give a convenient combinatorial characterization of optimal sets of questions. Before presenting this characterization, we show that in this context it suffices to look at dyadic distributions. Lemma 5.1.1. A set Q of questions over Xn is optimal if and only if c(Q, µ) = Opt(µ) for all dyadic distributions µ. Proof. Suppose that Q is optimal for all dyadic distributions, and let π be an arbitrary distribution over Xn . Let µ be a dyadic distribution such that Opt(π) = n X πi log i=1 1 . µi By assumption, Q is optimal for µ. Let T be an optimal decision tree for µ using questions from Q only, and let τ be the corresponding dyadic distribution, given by τi = 2−T (xi ) (recall that T (xi ) is the depth of xi ). Since τ minimizes T (µ) = H(µ) + D(µkτ ) over dyadic distributions, necessarily τ = µ. Thus n n X X 1 1 πi log = T (π) = πi log = Opt(π), τi µi i=1 i=1 showing that Q is optimal for µ. Given a dyadic distribution µ on Xn , we will be particularly interested in the collection of subsets of Xn that have probability exactly half under µ. Definition 5.1.2 (Dyadic hitters). Let µ be a non-constant dyadic distribution. A set A ⊆ Xn splits µ if µ(A) = 1/2. We denote the collection of all sets splitting µ by Spl(µ). We call a set of the form Spl(µ) a dyadic set. We call a set of questions Q a dyadic hitter in Xn if it intersects Spl(µ) for all non-constant dyadic distributions µ. (Lemma 2.1 implies that Spl(µ) is always non-empty.) A dyadic hitter is precisely the object we are interested in: Lemma 5.1.3. A set Q of subsets of Xn is an optimal set of questions if and only if it is a dyadic hitter in Xn . Proof. Let Q be a dyadic hitter in Xn . We prove by induction on 1 ≤ m ≤ n that for a dyadic distribution µ on Xn with support size m, c(Q, µ) = H(µ). Since Opt(µ) = H(µ), Lemma 5.1.1 implies that Q is an optimal set of questions. The base case, m = 1, is trivial. Suppose therefore that µ is a dyadic distribution whose support has size m > 1. In particular, µ is not constant, and so Q contains some set S ∈ Spl(µ). Let α = µ|S and β = µ|S , and note that α, β are both dyadic. The induction hypothesis shows 16 that c(Q, α) = H(α) and c(Q, β) = H(β). A decision tree which first queries S and then uses the implied algorithms for α and β has cost 1 1 1 + H(α) + H(β) = h(µ(S)) + µ(S)H(µ|S ) + µ(S)H(µ|S ) = H(µ), 2 2 using the Bernoulli chain rule; here µ|S is the restriction of µ to the elements in S. Conversely, suppose that Q is not a dyadic hitter, and let µ be a non-constant dyadic distribution such that Spl(µ) is disjoint from Q. Let T be any decision tree for µ using Q, and let S be its first question. The cost of T is at least 1 + µ(S)H(µ|S ) + µ(S)H(µ|S ) > h(µ(S)) + µ(S)H(µ|S ) + µ(S)H(µ|S ) = H(µ), since µ(S) 6= 21 . Thus c(Q, µ) > Opt(µ), and so Q is not an optimal set of questions. 5.2 Dyadic sets as antichains There is a surprising connection between dyadic hitters and hitting sets for maximal antichains. We start by defining the latter: Definition 5.2.1. A fibre in Xn is a subset of 2Xn which intersects every maximal antichain in Xn . Fibres were defined by Lonc and Rival [20], who also gave a simple construction, via cones: Definition 5.2.2. The cone C(S) of a set S consists of all subsets and all supersets of S. Any cone C(S) intersects any maximal antichain A, since otherwise A ∪ {S} is also an antichain. By choosing S of size ⌊n/2⌋, we obtain a fibre of size 2⌊n/2⌋ + 2⌈n/2⌉ − 1 = Θ(2n/2 ). Our goal now is to show that every fibre is a dyadic hitter: Theorem 5.2.3. every fibre is a dyadic hitter. This shows that every cone is a dyadic hitter, and allows us to give a simple algorithm for constructing an optimal decision tree using any cone. We start with a simple technical lemma which will also be used in Section 5.4: Definition 5.2.4. Let µ be a dyadic distribution over Xn . The tail of µ is the largest set of elements T ⊆ Xn such that for some a ≥ 1, (i) The elements in T have probabilities 2−a−1 , 2−a−2 , . . . , 2−a−(|T |−1) , 2−a−(|T |−1) . (ii) Every element not in T has probability at least 2−a . Lemma 5.2.5. Suppose that µ is a non-constant dyadic distribution with non-empty tail T . Every set in Spl(µ) either contains T or is disjoint from T . Proof. The proof is by induction on |T |. If |T | = 2 then there exist an integer a ≥ 1 and two elements, without loss of generality x1 , x2 , of probability 2−a−1 , such that all other elements have probability at least 2−a . Suppose that S ∈ Spl(µ) contains exactly one of x1 , x2 . Then 2a−1 = X xi ∈S X 2a µ(xi ) = xi ∈S\{x1 ,x2 } 17 1 2a µ(xi ) + . 2 However, the left-hand side is an integer while the right-hand side is not. We conclude that S must contain either both of x1 , x2 or none of them. For the induction step, let the elements in the tail T of µ have probabilities 2−a−1 , 2−a−2 , . . . , −a−(|T |−1) , 2−a−(|T |−1) . Without loss of generality, suppose that x 2 n−1 , xn are the elements whose −a−(|T |−1) probability is 2 . The same argument as before shows that every dyadic set in Spl(µ) must contain either both of xn−1 , xn or neither. Form a new dyadic distribution ν on Xn−1 by merging the elements xn−1 , xn into xn−1 , and note that Spl(µ) can be obtained from Spl(ν) by replacing xn−1 with xn−1 , xn . The distribution ν has tail T ′ = T \ {xn }, and so by induction, every set in Spl(ν) either contains T ′ or is disjoint from T ′ . This implies that every set in Spl(µ) either contains T or is disjoint from T . The first step in proving Theorem 5.2.3 is a reduction to dyadic distributions having full support: Lemma 5.2.6. A set of questions is a dyadic hitter in Xn if and only if it intersects Spl(µ) for all non-constant full-support dyadic distributions µ on Xn . Proof. A dyadic hitter clearly intersects Spl(µ) for all non-constant full-support dyadic distributions on Xn . For the other direction, suppose that Q is a set of questions that intersects Spl(µ) for every non-constant full-support dyadic distribution µ. Let ν be a non-constant dyadic distribution on Xn which doesn’t have full support. Let xmin be an element in the support of ν with minimal probability, which we denote νmin . Arrange the elements in supp(ν) in some arbitrary order xi1 , . . . , xim . Consider the distribution µ given by: • µ(xi ) = ν(xi ) if xi ∈ supp(µ) and xi 6= xmin . • µ(xmin ) = νmin /2. • µ(xij ) = νmin /2j+1 for j < m. • µ(xim ) = νmin /2m . In short, we have replaced ν(xmin ) = νmin with a tail xmin , xi1 , . . . , xim of the same total probability. It is not hard to check that µ is a non-constant dyadic distribution having full support on Xn . We complete the proof by showing that Q intersects Spl(ν). By assumption, Q intersects Spl(µ), say at a set S. Lemma 5.2.5 shows that S either contains all of {xmin } ∪ supp(ν), or none of these elements. In both cases, ν(S) = µ(S) = 1/2, and so Q intersects Spl(ν). We complete the proof of Theorem 5.2.3 by showing that dyadic sets corresponding to fullsupport distributions are maximal antichains: Lemma 5.2.7. Let µ be a non-constant dyadic distribution over Xn with full support, and let D = Spl(µ). Then D is a maximal antichain which is closed under complementation (i.e. A ∈ D =⇒ X \ A ∈ D). Proof. (i) That D is closed under complementation follows since if A ∈ D then µ(X \ A) = 1 − µ(A) = 1/2. (ii) That D is an antichain follows since if A strictly contains B then µ(A) > µ(B) (because µ has full support). (iii) It remains to show that D is maximal. By (i) it suffices to show that every B with µ(B) > 1/2 contains some A ∈ D. This follows from applying Lemma 2.1 on the probabilities of the elements in B. 18 Cones allow us to prove Theorem 5.5: Proof of Theorem 5.5. Let S = {x1 , . . . , x⌊n/2⌋ }. The set of questions Q is the cone C(S), whose size is 2⌊n/2⌋ + 2⌈n/2⌉ − 1 < 2⌈n/2⌉+1 . An efficient indexing scheme for Q divides the index into a bit b, signifying whether the set is a subset of S or a superset of S, and ⌊n/2⌋ bits (in the first case) or ⌈n/2⌉ bits (in the second case) for specifying the subset or superset. To prove the other two parts, we first solve an easier question. Suppose that µ is a non-constant dyadic distribution whose sorted order is known. We show how to find a set in Spl(µ) ∩ Q in time O(n). If µ(S) = 1/2 then S ∈ Spl(µ). If µ(S) > 1/2, go over the elements in S in non-decreasing order. According to Lemma 2.1, some prefix will sum to 1/2 exactly. If µ(S) < 1/2, we do the same with S, and then complement the result. Suppose now that π is a non-constant distribution. We can find a Huffman distribution µ for π and compute the sorted order of π in time O(n log n). The second and third part now follow as in the proof of Lemma 5.1.3. 5.3 Reduction to maximum relative density Our lower bound on the size of a dyadic hitter, proved in the following subsection, will be along the following lines. For appropriate values of n, we describe a dyadic distribution µ, all of whose splitters have a certain size i or n − i. Moreover, only a ρ fraction of sets of size i split µ. We then consider all possible permutations of µ. Each set of size i splits a ρ fraction of these, and so any dyadic hitter must contain at least 1/ρ sets. This lower bound argument prompts the definition of maximum relative density (MRD), which corresponds to the parameter ρ above; in the general case we will also need to optimize over i. We think of the MRD as a property of dyadic sets rather than dyadic distributions; indeed, the concept of MRD makes sense for any collection of subsets of Xn . If a dyadic set has MRD ρ then any dyadic hitter must contain at least 1/ρ questions, due to the argument outlined above. Conversely, using the probabilistic method we will show that roughly 1/ρmin (n) questions suffice, where ρmin (n) is the minimum MRD of a dyadic set on Xn . Definition 5.3.1 (Maximum relative density). Let D be a collection of subsets of Xn . For 0 ≤ i ≤ n, let  S ∈ D : |S| = i  . ρi (D) := n i We define the maximum relative density (MRD) of D, denoted ρ(D), as ρ(D) := max i∈{1,...,n−1} ρi (D). We define ρmin (n) to be the minimum of ρ(D) over all dyadic sets. That is, ρmin (n) is the smallest possible maximum relative density of a set of the form Spl(µ). The following theorem shows that uOpt (n, 0) is controlled by ρmin (n), up to polynomial factors. Theorem 5.3.1. Fix an integer n, and denote M := 1 ρmin (n) . Then M ≤ uOpt (n, 0) ≤ n2 log n · M. 19 Proof. Note first that according to Lemma 5.1.3, uOpt (n, 0) is equal to the minimal size of a dyadic hitter in Xn , and thus it suffices to lower- and upper-bound this size. Let σ be a uniformly random permutation on Xn . If S is any set of size i then σ −1 (S) is a uniformly random set of size i, and so ρi (D) = Pr σ∈Sym(Xn ) [σ −1 (S) ∈ D] = Pr [S ∈ σ(D)]. σ∈Sym(Xn ) (Here Sym(Xn ) is the group of permutations of Xn .) Fix a dyadic set D on Xn with ρ(D) = ρmin (n). The formula for ρi (D) implies that for any subset S of Xn (of any size), Pr σ∈Sym(Xn ) [S ∈ σ(D)] ≤ ρmin (n). Let Q be a collection of subsets of Xn with |Q| < M . A union bound shows that Pr [Q ∩ σ(D) 6= ∅] ≤ |Q|ρmin (n) < 1. σ∈Sym(Xn ) Thus, there exists a permutation σ such that Q ∩ σ(D) = ∅. Since σ(D) is also a dyadic set, this shows that Q is not a dyadic hitter. We deduce that any dyadic hitter must contain at least M questions. For the upper bound on uOpt (n, 0), construct a set of subsets Q containing, for each i ∈ {1, . . . , n − 1}, M n log n uniformly chosen sets S ⊆ Xn of size i. We show that with positive probability, Q is a dyadic hitter. Fix any dyadic set D, and let i ∈ {1, . . . , n − 1} be such that ρi (D) = ρ(D) ≥ ρmin (n). The probability that a random set of size i doesn’t belong to D is at most 1 − ρ(D) ≤ 1 − ρmin (n). Therefore the probability that Q is disjoint from D is at most (1 − ρmin (n))M n log n ≤ e−ρmin (n)M n log n = e−n log n < n−n . As we show below in Claim 5.3.2, there are at most nn non-constant dyadic distributions, and so a union bound implies that with positive probability, Q is indeed a dyadic hitter. In order to complete the proof of Theorem 5.3.1, we bound the number of non-constant dyadic distributions: Claim 5.3.2. There are at most nn non-constant dyadic distributions on Xn . Proof. Recall that dyadic distributions correspond to decision trees in which an element of probability 2−ℓ is a leaf at depth ℓ. Clearly the maximal depth of a leaf is n − 1, and so the probability of each element in a non-constant dyadic distribution is one of the n values 0, 2−1 , . . . , 2−(n−1) . The claim immediately follows. Krenn and Wagner [19] showed that the number of full-support dyadic distributions on Xn is asymptotic to αγ n−1 n!, where α ≈ 0.296 and γ ≈ 1.193, implying that the number of dyadic distributions on Xn is asymptotic to αe1/γ γ n−1 n!. Boyd [4] showed that the number of monotone full-support dyadic distributions on Xn is asymptotic to βλn , where β ≈ 0.142 and λ ≈ 1.794, implying that the number of monotone dyadic distributions on Xn is asymptotic to β(1 + λ)n . The proof of Theorem 5.3.1 made use of two properties of dyadic sets: 1. Any permutation of a dyadic set is a dyadic set. 20 2. There are en O(1) dyadic sets. If F is any collection of subsets of 2Xn satisfying the first property then the proof of Theorem 5.3.1 generalizes to show that the minimal size U of a hitting set for F satisfies M ≤ U ≤ M n log |F|, 5.4 where M = 1 minD∈F ρ(D) . Upper bounding ρmin (n) Theorem 5.3 will ultimately follow from the following lemma, by way of Theorem 5.3.1: Lemma 5.4.1. Fix 0 < β ≤ 1/2. There exists an infinite sequence of positive integers n (namely, √ 2a those of the form ⌊ 2β ⌋ for integer a) such that some dyadic set D in Xn satisfies ρ(D) ≤ O( n)2−(h(β)−2β)n . Proof. We prove the lemma under the simplifying assumption that 1/β is an integer (our most important application of the lemma has β := 1/5). Extending the argument for general β is straightforward and left to the reader. 2a , for a positive integer a. Note that for n of this form, Let n be an integer of the form 2β a−1 βn = 2 is a power of two. Let t = βn, and construct a dyadic distribution µ on Xn as follows: 1. For i ∈ [2t − 1], µ(xi ) = 2−a = 1 2t . 2. For i ∈ [n − 1] \ [2t − 1], µ(xi ) = µ(xi−1 )/2 = 2−(a+i−2t+1) . 3. µ(xn ) = µ(xn−1 ). The corresponding decision tree is obtained by taking a complete binary tree of depth a and replacing one of the leaves by a “path” of length n − 2a ; see Figure 2. Alternatively, in the terminology of Definition 5.2.4 we form µ by taking the uniform distribution on X2t and replacing x2t with a tail on x2t , . . . , xn . We claim that D := Spl(µ) contains only two types of sets: 1. Subsets of size t of X2t−1 . 2. Subsets of size n − t containing t − 1 elements of X2t−1 and all the elements x2t , . . . , xn . It is immediate that any such set S is in D. On the other hand, Lemma 5.2.5 shows that every set S ∈ D either contains the tail x2t , . . . , xn or is disjoint from it. If S is disjoint from the tail then it must be of the first form, and if S contains the tail then it must be of the second form. √ n Using the estimate βn ≥ 2h(β)n /O( n) (see for example [31]), we see that 2t−1 √ √ 22t 22t ρt (D) = ρn−t (D) = nt  ≤ n  ≤ O( n) h(β)n = O( n)2(2β−h(β))n . 2 t βn For i ∈ {1, . . . , n − 1} \ {t, n − t} we have ρi (D) = 0. Thus indeed √ ρ(D) ≤ O( n)2(2β−h(β))n . Theorem 5.3 can now be easily derived. The first step is determining the optimal value of β: Claim 5.4.2. We have max 2h(β)−2β = 1.25, β∈[0,1] and the maximum is attained (uniquely) at β = 1/5. 21 Complete binary tree on 2βn vertices Path of length (1 − 2β)n Figure 2: The hard distribution used to prove Lemma 5.4.1, in decision tree form Proof. Let f (β) = h(β) − 2β. Calculation shows that the derivative f ′ (β) is equal to   1−β ′ f (β) = log − 2, β which is decreasing for 0 < β < 1 and vanishes at β = 1/5. Thus f (β) achieves a unique maximum over β ∈ (0, 1) at β = 1/5, where 2f (1/5) = 2h(1/5)−2·1/5 = 1.25. Proof of Theorem 5.3: a 2 for a Proof. Let β := 1/5. Claim 5.4.2 shows that 2−(2β−h(β)) = 1.25. Fix any n of the form n = 2β positive integer a. It follows from Lemma 5.4.1 together with the first inequality in Theorem 5.3.1 √ that uOpt (n, 0) ≥ 1.25n /O( n). 2a for a positive integer a and 1/4 ≤ β ≤ 1/2. A general n can be written in the form n = 2β Lemma 5.4.1 and Theorem 5.3.1 show that for any integer ℓ ≥ 0, ℓ )−2β/2ℓ ]n uOpt (n, 0) ≥ 2[h(β/2 √ /O( n). Calculation shows that when β ≤ β0 ≈ 0.27052059413118146, this is maximized at ℓ = 0, and √ otherwise this is maximized at ℓ = 1. Denote the resulting lower bound by L(β)n /O( n), the minimum of L(β) is attained at β0 , at which point its value is L(β0 ) ≈ 1.23214280723432. 5.5 Lower bounding ρmin (n) We will derive Theorem 5.2 from the following lemma: 22 Lemma 5.5.1. For every non-constant dyadic distribution µ there exists 0 < β < 1 such that ρ(Spl(µ)) ≥ 2(2β−h(β))n = 2(2β−h(β))n−o(n) . √ O( n)O(log n) Proof. Assume without loss of generality that the probabilities in µ are non-increasing: µ1 ≥ µ2 ≥ · · · ≥ µn . The idea is to find a partition of Xn of the form Xn = γ [ i=1 (Di ∪ Ei ) which satisfies the following properties: 1. Di consists of elements having the same probability pi . 2. If Di has an even number of elements then Ei = ∅. 3. If Di has an odd number of elements then µ(Ei ) = pi . 4. γ = O(log n). (In fact, γ = o(n/ log n) would suffice.) We will show later how to construct such a partition. The conditions imply that µ(Di ∪ Ei ) is an even integer multiple of pi , say µ(Di ∪ Ei ) = 2ci pi . It is not hard to check that ci = ⌈|Di |/2⌉. Given such a partition, we show how to lower bound theS maximum relative density of Spl(µ). If Si ⊆ Di is a set of size ci for each i ∈ [γ] then the set S = i Si splits µ: µ(S) = γ X i=1 Defining c = is Pγ i=1 ci , i=1 we see that each such set S contains c elements, and the number of such sets  γ  Y |Di | i=1 using the estimate γ 1X 1 ci pi = µ(Di ∪ Ei ) = . 2 2 ci  γ Y 22c 22ci √ = ≥ √ O(log n) , O( n) O( n) i=1 m ⌈m/2⌉  =Θ 22⌈m/2⌉ √ m ! , which follows from Stirling’s approximation. In order to obtain an estimate on the maximum relative density of Spl(µ), we use the following n 3 folklore upper bound on c :   n ≤ 2h(c/n)n . c 3 Here is a quick proof: Let Y be a uniformly random subset of Xn of size c, and let Yi indicate the event xi ∈ Y . Then log nc = H(Y ) ≤ nH(Y1 ) = nh(c/n). 23 We conclude that the maximum relative density of Spl(µ) is at least Qγ |Di | 22c−h(c/n)n i=1 ci  . ρ(µ) ≥ ρc (µ) ≥ ≥ √ n O( n)O(log n) c To obtain the expression in the statement of the lemma, take β := c/n. We now show how to construct the partition of Xn . We first explain the idea behind the construction, and then provide full details; the reader who is interested only in the construction itself can skip ahead. Proof idea Let q1 , . . . , qγ be the different probabilities of elements in µ. We would like to put all elements of probability qi in the set Di , but there are two difficulties: 1. There might be an odd number of elements whose probability is qi . 2. There might be too many distinct probabilities, that is, γ could be too large. (We need γ = o(n/ log n) for the argument to work.) The second difficulty is easy to solve: we let D1 = {x1 }, and use Lemma 2.1 to find an index ℓ such that µ(E1 ) := µ({xℓ , . . . , xn }) = µ1 . A simple argument shows that all remaining elements have probability at least µ1 /n, and so the number of remaining distinct probabilities is O(log n). (The reader should observe the resemblance between E1 and the tail of the hard distribution constructed in Lemma 5.4.1.) Lemma 2.1 also allows us to resolve the first difficulty. The idea is as follows. Suppose that the current set under construction, Di , has an odd number of elements, each of probability qi . We use Lemma 2.1 to find a set of elements whose total probability is qi , and put them in Ei . Detailed proof Let N be the maximal index such that µN > 0. Since µ is non-constant, µ1 ≤ 1/2, and so Lemma 2.1 proves the existence of an index M such that µ({xM +1 , . . . , xN }) = µ1 (we use the furthermore part of the lemma, and M = ℓ − 1). We take D1 := {x1 }, E1 := {xM +1 , . . . , xn }. Thus µ(D1 ) = µ(E1 ) = µ1 , and so µ({x2 , . . . , xM }) = 1 − 2µ1 (possibly M = 1, in which case the construction is complete). By construction nµM > µ(E1 ) = µ1 , and so µM < µ1 /n. In particular, the number of distinct probabilities among µ2 , . . . , µM is at most log n. This will guarantee that γ ≤ log n + 1, as will be evident from the construction. The construction now proceeds in steps. At step i, we construct the sets Di and Ei , given the set of available elements {xαi , . . . , xM }, where possibly αi = M + 1; in the latter case, we have completed the construction. We will maintain the invariant that µ({xαi , . . . , xM }) is an even multiple of µαi ; initially α2 := 2, and µ({xαi , . . . , xM }) = (1/µ1 − 2)µ1 is indeed an even multiple of µ2 . Let βi be the maximal index such that µβi = µαi (possibly βi = αi ). We define Di := {xαi , . . . , xβi }. Suppose first that |Di | is even. In this case we define Ei := ∅, and αi+1 := βi + 1. Note that µ({xαi+1 , . . . , xM }) = µ({xαi , . . . , xM }) − |Di |µαi , 24 and so the invariant is maintained. Suppose next that |Di | is odd. In this case µ({xβi +1 , . . . , xM }) ≥ xαi , since µ({xβi +1 , . . . , xM }) is an odd multiple of µαi . Therefore we can use Lemma 2.1 to find an index γi such that µ({xβi +1 , . . . , xγi }) = µαi . We take Ei := {xβi +1 , . . . , xγi } and αi+1 := γi + 1. Note that µ({xαi+1 , . . . , xM }) = µ({xαi , . . . , xM }) − (|Di | + 1)µαi , and so the invariant is maintained. The construction eventually terminates, say after step γ. The construction ensures that µα2 > µα3 > · · · > µαγ . Since there are at most log n distinct probabilities among the elements {xα2 , . . . , xM }, γ ≤ log n + 1, completing the proof. Theorem 5.2 follows immediately from the second inequality in Theorem 5.3.1 together with the following lemma: Lemma 5.5.2. Fix an integer n and let D be a dyadic set in Xn . Then ρ(D) ≥ 1.25−n−o(n) , and thus ρmin (n) ≥ 1.25−n−o(n) . Proof. Fix a dyadic set D in Xn . Lemma 5.5.1 implies that there exists 0 < β < 1 such that ρ(D) ≥ 2(2β−h(β))n−o(n) . Using Claim 5.4.2 we have 22β−h(β) ≥ 54 , and so ρ(D) ≥ (4/5)n · 2−o(n) = 1.25−n−o(n) . 6 Combinatorial benchmark with prolixity In the previous section we studied the minimum size of a set Q of questions with the property that for every distribution, there is an optimal decision tree using only questions from Q. In this section we relax this requirement by allowing the cost to be slightly worse than the optimal cost. More formally, recall that uOpt (n, r) is the minimum size of a set of questions Q such that for every distribution π there exists a decision tree that uses only questions from Q with cost at most Opt(π) + r. In a sense, uOpt (n, r) is an extension of uH (n, r) for r ∈ (0, 1): indeed, uH (n, r) is not defined for r < 1 since for some distributions π there is no decision tree with cost less than H(π) + 1 (see Section 3). Moreover, Opt(π), which is the benchmark used by uOpt (n, r), is precisely the optimal cost, whereas H(π), the benchmark used by uH (n, r) is a convex surrogate of Opt(π). We focus here on the range r ∈ (0, 1). We prove the following bounds on uOpt (n, r), establishing that uOpt (n, r) ≈ (r · n)Θ(1/r) . Theorem 6.1. For all r ∈ (0, 1), and for all n > 1/r: 16 1 1 (r · n) 4r ≤ uOpt (n, r) ≤ n2 (3r · n) r . n As a corollary, we get that the threshold of exponentiality is 1/n: 25 Corollary 6.2. If r = ω(1/n) then uOpt (n, r) = 2o(n) . Conversely, if r = O(1/n) then uOpt (n, r) = 2Ω(n) . For larger r, the following theorem is a simple corollary of Theorem 4.1 and the bound uOpt (n, r) ≤ uH (n, r) ≤ uOpt (n, r − 1): Theorem 6.3. For every r ≥ 1 and n ∈ N, 1 ⌊r + 1⌋n1/⌊r+1⌋ ≤ uOpt (n, r) ≤ 2⌊r⌋n1/⌊r⌋ . e Theorem 6.1 is implied by the following lower and upper bounds, which provide better bounds when r ∈ (0, 1) is a negative power of 2. Theorem 6.4 (Lower bound). For every r of the form 1/2k , where k ≥ 1 is an integer, and n > 2k : 1 uOpt (n, r) ≥ (r · n) 2r −1 . Theorem 6.5 (Upper bound). For every r of the form 4/2k , where k ≥ 3 is an integer, and n > 2k : uOpt (n, r + r 2 ) ≤ n2 4 3e r · n r. 4 These results imply Theorem 6.1, due to the monotonicity of uOpt (n, r), as follows. Let r ∈ (0, 1). For the lower bound, pick the smallest t ≥ r of the form 1/2k . Note that t ≤ 2r, and thus: 1 1 1 1 uOpt (n, r) ≥ uOpt (n, t) ≥ (t · n) 2t −1 ≥ (r · n) 4r −1 ≥ (r · n) 4r . n For the upper bound, pick the largest t of the form 4/2k , k ≥ 3 such that t + t2 ≤ r. Note that t ≥ r/4 (since s = r/2 satisfies s + s2 ≤ 2s ≤ r), and thus uOpt (n, r) ≤ uOpt (n, t + t2 ) ≤ n2 6.1 4 16 3e t · n t ≤ n2 (3r · n) r . 4 Lower bound Pick a sufficiently small δ > 0 (as we will soon see, δ < r 2 suffices), and consider a distribution µ with 2k − 1 “heavy” elements (this many elements exist since n > 1/r), each of probability 21−δ k −1 , k and n − (2 − 1) “light” elements with total probability of δ. Recall that a decision tree is r-optimal if its cost is at most Opt(µ) + r. The proof proceeds by showing that if T is an r-optimal tree, then the first question in T has the following properties: (i) it separates the heavy elements to two sets of almost equal sizes (2k−1 and 2k−1 − 1), and (ii) it does not distinguish between the light elements.  The result then follows since there are 2kn−1 such distributions σ (the number of ways to choose  k−1 the light elements), and each question can serve as a first question to at most n−(22k−1 −1) of them. To establish these properties, we first prove a more general result (cf. Lemma 5.2.5): Lemma 6.1.1.  Let µ be a distribution over a finite set X, and let A ⊆ X be such that for every x∈ / A, µ {x} > µ(A) + ǫ. Then every decision tree T which is ǫ-optimal with respect to µ has a subtree T ′ whose set of leaves is A. 26 x x y a′ y a′′ (a) Original tree a′ a′′ (b) Transformed tree   Figure 3: The transformation in Lemma 6.1.1. The cost decreases by µ {x} − µ {a′ , a′′ } > ǫ. Proof. By induction on |A|. The case |A| = 1 follows since any leaf is a subtree. Assume |A| > 1. Let T be a decision tree which is ǫ-optimal with respect to µ. Let x, y be two siblings of maximal depth. Note that it suffices to show that x, y ∈ A, since then, merging x, y to a new element z with µ({z}) = µ({x}) + µ({y}) and applying the induction hypothesis yields that A ∪ {z} \ {x, y} is the set of leaves of a subtree of T with x, y removed. This finishes the proof since x, y are the children of z. It remains to show that x, y ∈ A. Let d denote the depth of x and y. Assume toward contradiction that x ∈ / A. Pick a′ , a′′ ∈ A, with depths d′ , d′′ (this is possible since |A| > 1). If d′ < d or ′′ d < d then replacing a′ with x or a′′ with x improves the cost of T by more than ǫ, contradicting its optimality. Therefore, it must be that d′ = d′′ = d, and we perform the following transformation (see Figure 3): the parent of x and y becomes a leaf with label x (decreasing the depth of x by 1), y takes the place of a′ (the depth of y does not change), and a′′ becomes aninternal node with two children labeled by a′ , a′′ (increasing the depths of a′ , a′′ by 1). Since µ {x} − µ {a′ , a′′ } > ǫ, this transformation improves the cost of T by more than ǫ, contradicting its ǫ-optimality. Corollary 6.1.2. Let µ be a distribution over X, and let A ⊆ X be such that for every x ∈ / A,  µ {x} > µ(A). Then every optimal tree T with respect to µ has a subtree T ′ whose set of leaves is A. Property (ii) follows from Lemma 6.1.1, which implies that if δ is sufficiently small then all light elements are clustered together as the leaves of some subtree. Indeed, by Lemma 6.1.1, this happens if the probability of a single heavy element (which is 21−δ k −1 ) exceeds the total probability of all light elements (which is δ) by at least r. A simple calculation shows that setting δ smaller than r 2 suffices. We summarize this in the following claim: Claim 6.1.3 (light elements). Every r-optimal tree has a subtree whose set of leaves is the set of light elements. The next claim concerns the other property: Claim 6.1.4 (heavy elements). In every r-optimal decision tree, the first question partitions the heavy elements into a set of size 2k−1 and a set of size 2k−1 − 1. 27 Proof. When k = 2, it suffices to prove that an r-optimal decision tree cannot have a first question which separates the heavy elements from the light elements. Indeed, the heavy elements in such a tree reside at depths 2, 3, 3. Exchanging one of the heavy elements at depth 2 with the subtree consisting of all light elements (which is at depth 1) decreases the cost by 21−δ k −1 − δ > r, showing that the tree wasn’t r-optimal. Suppose that some r-optimal decision tree T contradicts the statement of the claim, for some k ≥ 3. The first question in T leads to two subtrees T1 , T2 , one of which (say T1 ) contains at least 2k−1 + 1 heavy elements, and the other (say T2 ) contain at most 2k−1 − 2. One of the subtrees also contains a subtree T ′ whose leaves are all the light elements. For the sake of the argument, we replace the subtree T ′ with a new element y. We claim that T1 contains an internal node v at depth D(v) ≥ k − 1 which has at least two heavy descendants. To see this, first remove y if it is present in T1 , by replacing its parent by its sibling. The possibly modified tree T1′ contains at least 2k−1 + 1 leaves, and in particular some leaf at depth at least k. Its parent v has depth at least k − 1 and at least two heavy descendants, in both T1′ and T1 . In contrast, T2 contains at least two leaves (since 2k−1 − 2 ≥ 2), and the two shallowest ones must have depth at most k − 2. At least one of these is some heavy element xℓ . Exchanging v and xℓ results in a tree T ∗ whose cost c(T ∗ ) is at most c(T ∗ ) ≤ c(T ) + (D(v) − D(xℓ ))(2 − 1) 1−δ 1−δ ≤ c(T ∗ ) − k < c(T ) − r, k 2 −1 2 −1 contradicting the assumption that T is r-optimal. (That tion 21−δ k −1 > δ + r.) 1−δ 2k −1 > r follows from the earlier assump- By the above claims, there are two types of first questions for µ, depending on which of the two subtrees of the root contains the light elements: • Type 1: questions that split the elements into a part with 2k−1 elements, and a part with n − 2k−1 elements. • Type 2: questions that split the elements into a part with 2k−1 − 1 elements, and a part with n − (2k−1 − 1) elements. If we identify a question with its smaller part (i.e. the part of size 2k−1 or the part of size 2k−1 − 1), we deduce that any set of questions with redundancy r must contain a family F such that (i) every set in F has size 2k−1 or 2k−1 − 1, and (ii) for every set of size n − (2k − 1), there exists some set in F that is disjoint from it. It remains to show that any such family F is large. Indeed, there are 2kn−1 sets of size n − (2k − 1), and since every set in F has size at least k−1 −1) n−(2k−1 −1) = of them. Thus 2k−1 − 1, it is disjoint from at most n−(2 k 2k−1 n−(2 −1) |F| ≥ 6.2 n  2k −1 n−(2k−1 −1) 2k−1 = 1 n 2k−1 −1 n(n − 1) · · · (n − (2k−1 − 1) + 1) ≥ = (r · n) 2r −1 . (2k − 1)(2k − 2) · · · (2k−1 + 1) 2k Upper bound The set of questions. In order to describe the set of queries it is convenient to assign a cyclic order on Xn : x1 ≺ x2 ≺ · · · ≺ xn ≺ x1 ≺ · · · . The set of questions Q consists of all cyclic intervals, 28 with up to 2k elements added or removed. Since r = 4 · 2−k , the number of questions is plainly at most   4 2 n 2k 2 3e r , n r · n 3 ≤ n 2k 4  d using the inequality nd ≤ en . d High level of the proof. Let π be an arbitrary distribution on Xn , and let r ∈ (0, 1) be of the form 4 · 2−k , with k ≥ 3. Let µ be a Huffman distribution for π; we remind the reader that µ is a dyadic distribution corresponding to some optimal decision tree for π. We construct a decision tree T that uses only queries from Q, with cost T (π) ≤ Opt(π) + r + r 2 = X π(x) log x∈Xn 1 + r + r2. µ(x) The construction is randomized: we describe a randomized decision tree TR (‘R’ denotes the randomness that determines the tree) which uses queries from Q and has the property that for every x ∈ Xn , the expected number of queries TR uses to find x satisfies the inequality E[TR (x)] ≤ log R 1 + r + r2, µ(x) (3) where TR (x) is the depth of x. This implies the existence of a deterministic tree with cost Opt(µ) + r + r 2 : indeed, when x ∼ µ, the expected cost of TR is E [TR (x)] ≤ x∼π;R X x∈Xn   1 + r + r 2 = Opt(π) + r + r 2 . π(x) µ(x) Since the randomness of the tree is independent from the randomness of π, it follows that there is a choice of R such that the cost of the (deterministic) decision tree TR is at most Opt(π) + r + r 2 . The randomized decision tree. The randomized decision tree maintains a dyadic sub-distribution µ(i) that is being updated after each query. A dyadic sub-distribution is a measure on Xn such P (i) (i) (i) that (i) µ (x) is either 0 or a power of 2, and (ii) µ (Xn ) = x∈Xn µ (x) ≤ 1. A natural (i) interpretation of µ (x) is as a dyadic sub-estimate of the probability that x is the secret element, conditioned on the answers to the first i queries. The analysis hinges on the following properties: 1. µ(0) = µ,  2. µ(i) (x) ∈ 2µ(i−1) (x), µ(i−1) (x), 0 for all x ∈ Xn , 3. if x is the secret element then almost always µ(i) (x) is doubled; that is, µ(i) (x) > 0 for all i, and the expected number of i’s for which µ(i) (x) = µ(i−1) (x) is at most r + r 2 . These properties imply (3), which implies Theorem 6.5. Next, we describe the randomized decision tree and establish these properties. The algorithm distinguishes between light and heavy elements. An element x ∈ Xn is light if µ(i) (x) < 2−k . Otherwise it is heavy. The algorithm is based on the following win-win-win situation: (i) If the total mass of the heavy elements is at least 1/2 then by Lemma 2.1, there is a set I of heavy elements whose mass is exactly 1/2. Since the number of heavy elements is at most 2k , 29 the algorithm can ask whether x ∈ I and recurse by doubling the sub-probabilities of the elements that are consistent with the answer (and setting the others to zero). (ii) Otherwise, the mass of the heavy elements is less than 1/2. If the mass of the light elements is also less than 1/2 (this could happen since µ(i) is a sub-distribution), then we ask whether x is a heavy element or a light element, and accordingly recurse with either the heavy or the light elements, with their sub-probabilities doubled (in this case the “true” probabilities conditioned on the answers become larger than the sub-probabilities). (iii) The final case is when the mass of the light elements is larger than 1/2. In this case we query a random cyclic interval of light elements of mass ≈ 1/2, and recurse; there are two light elements in the recursion whose sub-probability is not doubled (the probabilities of the rest are doubled). Elements whose probability is not doubled occur only in case (iii). The randomized decision tree: formal description. The algorithm gets as input a subset y1 , . . . , ym of Xn whose order is induced by that of Xn , and a dyadic sub-distribution q1 , . . . , qm . Initially, the input is x1 , . . . , xn , and qi = µi . We say that an element is heavy is qi ≥ 2−k ; otherwise it is light. There are at most 2k heavy elements. The questions asked by the algorithm are cyclic intervals in y1 , . . . , ym , with some heavy elements added or removed. Since each cyclic interval in y1 , . . . , ym corresponds to a (not necessarily unique) cyclic interval in Xn (possibly including elements outside of y1 , . . . , ym ), these questions belong to Q. Algorithm TR . 1. If m = 1, return y1 . Otherwise, continue to Step 2. 2. If the total mass of heavy elements is at least 1/2 then find (using Lemma 2.1) a subset I whose mass is exactly 1/2, and ask whether x ∈ I. Recurse with either {2qi : yi ∈ I} or {2qi : yi ∈ / I}, according to the answer. Otherwise, continue to Step 3. 3. Let S be the set of all light elements, and let σ be their total mass. If σ ≤ 1/2 then ask whether x ∈ S, and recurse with either {2qi : yi ∈ S} or {2qi : yi ∈ / S}, according to the answer. Otherwise, continue to Step 4. 4. Arrange all light elements according to their cyclic order on a circle of circumference σ, by assigning each light element xi an arc Ai of length qi of the circle. Pick an arc of length 1/2 uniformly at random (e.g. by picking uniformly a point on the circle and taking an arc of length 1/2 directed clockwise from it), which we call the window. Let K ⊆ S consist of all light elements whose midpoints are contained in the window, and let B consist of the light elements whose arcs are cut by the boundary of the window (so |B| ≤ 2); we call these elements boundary elements. Ask whether x ∈ K; note that K is a cyclic interval in y1 , . . . , ym with some heavy elements removed. If x ∈ K, recurse with {2qi : yi ∈ K \ B} ∪ {qi : yi ∈ K ∩ B}. The sum of these dyadic probabilities is at most 1 since the window contains at least qi /2 of the arc Ai for each yi ∈ K ∩ B. If x ∈ / K, recurse with {2qi : yi ∈ K \ B} ∪ {qi : yi ∈ K ∩ B}. As in the preceding case, the total mass of light elements in the recursion is at most 2(σ − 1/2) (since the complement of the window contains at least qi /2 of the arc Ai for each yi ∈ K ∩ B), and the total mass of heavy elements is 2(1 − σ), for a total of at most (2σ − 1) + (2 − 2σ) = 1. ⊳ 30 Analysis. We now finish the proof by establishing the three properties of the randomized decision tree that are stated above. The first two properties follow immediately from the description of the algorithm, and it thus remains to establish the third property. Fix some x ∈ Xn , and let d ∈ N be such that µ(x) = 2−d . We need to show that the expected number of questions that are asked when the secret element is x is at most d + r + r 2 . Let q = q (i) denote the sub-probability of x after the i’th question; note that q ∈ {2−j : j ≤ d}. Lemma 6.2.1. If q ≥ 2−k then q doubles (that is, q (i+1) = 2q (i) ). Otherwise, the expected number 1 . of questions until q doubles is at most 1−4q Proof. From the description of the algorithm, it is clear that the only case in which the subprobability of x is not doubled is when x is one of the two boundary elements in Step 4. This only happens when x is a light element (i.e. q < 2−k ). The probability that x is one of the boundary elements is at most 2q/σ ≤ 4q, where σ ≥ 1/2 is the total mass of light elements: indeed, the probability that a given endpoint of the window lies inside the arc corresponding to q is q/σ, since each endpoint is distributed uniformly on the circle of circumference σ. It follows that the distribution of the number of questions that pass until q doubles is dominated by the geometric distribution with failure probability 4q, and so the expected number of questions 1 . until q doubles is at most 1−4q The desired bound on the expected number of questions needed to find x follows from Lemma 6.2.1: as long as q, the sub-probability associated with x, is smaller than 2−k , it takes an expected number 1 questions until it doubles. Once q ≥ 2−k , it doubles after every question. Thus, by linearity of 1−4q of expectation, the expected total number of questions is at most: k+ d X j=k+1 d X 1 [1 + 4 · 2−j + 2(4 · 2−j )2 ] <k+ 1 − 4 · 2−j j=k+1 =d+ d X [4 · 2−j + 2(4 · 2−j )2 ] j=k+1 2 < d + 4 · 2−k + (4 · 2−k )2 3 1 < log + r + r2. µ(x) 7 Open questions Our work suggests many open questions, some of which are: 1. The main results of Section 5 show that when n = 5 · 2m , uOpt (n, 0) = 1.25n±o(n) . We conjecture that there exists a function G : [1, 2] → R such that for n = α2m , uOpt (n, 0) = G(α)n±o(n) . Our results show that 1.232 ≤ G(α) ≤ 1.25 and that G(1.25) = 1.25. What is the function G? 2. Theorem 5.2 constructs an optimal set of questions of size 1.25n+o(n) , but this set is not explicit. √ n In contrast, Theorem 5.5 constructs explicitly an optimal set of questions of size O( 2 ), which furthermore supports efficient indexing and efficient construction of optimal strategies. Can we construct such an explicit set of optimal size 1.25n+o(n) ? 31 3. The results of Section 3 show that n ≤ uH (n, 1) ≤ 2n − 3. We conjecture that the limit H exists. What is the value of β? β = limn→∞ u (n,1) n An interesting suggestion for future research is to generalize the entire theory to d-way questions. References [1] Rudolf Ahlswede and Ingo Wegener. Search problems. John Wiley & Sons, Inc., New York, 1987. [2] Javad A. Aslam and Aditi Dhagat. Searching in the presence of linearly bounded errors. In Proceedings of the twenty-third annual ACM symposium on Theory of computing (STOC ’91), pages 486–493, 1991. [3] Harout Aydinian, Ferdinando Cicalese, and Christian Deppe, editors. Information Theory, Combinatorics, and Search Theory. Springer-Verlag Berlin Heidelberg, 2013. [4] David W. Boyd. The asymptotic number of solutions of a diophantine equation from coding theory. Journal of Combinatorial Theory, Series A, 18:210–215, 1975. [5] Renato M. Capocelli, Raffaele Giancarlo, and Indeer Jeet Taneja. Bounds on the redundancy of Huffman codes. IEEE Transactions on Information Theory, IT-32(6):854–857, 1986. [6] Gérard Cohen, Hugues Randriam, and Gilles Zémor. Witness sets. In Coding Theory and Applications (ICMCTA 2008), volume 5228 of LNCS. Springer, 2008. [7] Thomas M. Cover and Joy A. Thomas. Elements of information theory (2. ed.). Wiley, 2006. [8] Yuval Dagan. Twenty questions game using restricted sets of questions. Master’s thesis, 2017. [9] Aditi Dhagat, Peter Gács, and Peter Winkler. On playing “twenty questions” with a liar. In Proceedings of 3rd Symposium on Discrete Algorithms (SODA’92), pages 16–22, 1992. [10] Robert Dorfman. The detection of defective members of large populations. The Annals of Mathematical Statistics, 14(4):436–440, 1943. [11] Ding-Zhu Du and Frank K. Hwang. Combinatorial Group Testing and Its Applications, volume 12 of Series on Applied Mathematics. World Scientific, 2nd edition, 1999. [12] Michael L. Fredman. How good is the information theory bound in sorting? Computer Science, 1(4):355–361, 1976. Theoretical [13] Robert G. Gallager. Variations on a theme by Huffman. IEEE Transactions on Information Theory, IT-24(6):668–674, 1978. [14] E. N. Gilbert and E. F. Moore. Variable-length binary encodings. Bell System Technical Journal, 38:933–967, 1959. [15] Yasuichi Horibe. An improved bound for weight-balanced tree. Information and Control, 34(2):148–151, 1977. [16] David A. Huffman. A method for the construction of minimum-redundancy codes. In Proceedings of the I.R.E., pages 1098–1103, 1952. 32 [17] Ottar Johnsen. On the redundancy of binary Huffman codes. IEEE Transactions on Information Theory, IT-26(2):220–222, 1980. [18] Gyula O. H. Katona. Combinatorial search problems. In J. N. Srivastava et al., editor, A Survery of Combinatorial Theory. North-Holland Publishing Company, 1973. [19] Daniel Krenn and Stephan Wagner. Compositions into powers of b: asymptotic enumeration and parameters. Algorithmica, 75(4):606–631, August 2016. [20] Zbigniew Lonc and Ivan Rival. Chains, antichains, and fibres. Journal of Combinatorial Theory, Series A, 44:207–228, 1987. [21] Dietrich Manstetten. Tight bounds on the redundancy of Huffman codes. IEEE Transactions on Information Theory, IT-38(1):144–151, 1992. [22] Roy Meshulam. On families of faces in discrete cubes. Graphs and Combinatorics, 8:287–289, 1992. [23] Soheil Mohajer, Payam Pakzad, and Ali Kakhbod. Tight bounds on the redundancy of Huffman codes. In Information Theory Workshop (ITW ’06), pages 131–135, 2006. [24] Bruce L. Montgomery and Julia Abrahams. On the redundancy of optimal binary prefixcondition codes for finite and infinite sources. IEEE Transactions on Information Theory, IT-33(1):156–160, 1987. [25] Shay Moran and Amir Yehudayoff. A note on average-case sorting. Order, 33(1):23–28, 2016. [26] Narao Nakatsu. Bounds on the redundancy of binary alphabetical codes. IEEE Transactions on Information Theory, IT-37(4):1225–1229, 1991. [27] Jorma Rissanen. Bounds for weight balanced trees. IBM Journal of Research and Development, 17:101–105, 1973. [28] Ronald L. Rivest, Albert R. Meyer, Daniel J. Kleitman, Karl Winklmann, and Joel Spencer. Coping with errors in binary search procedures. Journal of Computer and System Sciences, 20:396–404, 1980. [29] Joel Spencer and Peter Winkler. Three thresholds for a liar. Combinatorics, Probability and Computing, 1(1):81–93, 1992. [30] David Spuler. Optimal search trees using two-way key comparisons. Acta Informatica, 31:729– 740, 1994. [31] Neal Young. Reverse Chernoff bound. Theoretical Computer Science Stack Exchange, 2012. URL:http://cstheory.stackexchange.com/q/14476 (version: 2012-11-26). 33
8cs.DS
arXiv:1801.08267v1 [cs.CV] 25 Jan 2018 Visual Weather Temperature Prediction Wei-Ta Chu National Chung Cheng University Kai-Chia Ho National Chung Cheng University [email protected] [email protected] Ali Borji Center for Research in Computer Vision, University of Central Florida [email protected] Abstract In this paper, we attempt to employ convolutional recurrent neural networks for weather temperature estimation using only image data. We study ambient temperature estimation based on deep neural networks in two scenarios a) estimating temperature of a single outdoor image, and b) predicting temperature of the last image in an image sequence. In the first scenario, visual features are extracted by a convolutional neural network trained on a large-scale image dataset. We demonstrate that promising performance can be obtained, and analyze how volume of training data influences performance. In the second scenario, we consider the temporal evolution of visual appearance, and construct a recurrent neural network to predict the temperature of the last image in a given image sequence. We obtain better prediction accuracy compared to the state-of-the-art models. Further, we investigate how performance varies when information is extracted from different scene regions, and when images are captured in different daytime hours. Our approach further reinforces the idea of using only visual information for cost efficient weather prediction in the future. 1. Introduction Visual attributes of images have been heavily studied for years. Most previous works have been focused on recognizing “explicit visual attributes” in images, such as object’s texture and color distributions [4], and semantic categories [10]. With the advance of computer vision and machine learning technologies, more and more works have been proposed to study discovering “subtle attributes” in images. These subtle attributes may not be well-formulated in explicit forms, but are usually recognizable by human beings. For example, Lu et al. [13] proposed a method to recognize whether an image was cap- tured on a sunny day or on a cloudy day. Hays and Efros [6] proposed to estimate geographic information from a single image (a.k.a IM2GPS). Recent research has demonstrated that deep learning approaches are effective for recognizing painting styles [11][2]. Among various subtle attributes, weather properties of images have attracted increasing attention. The earliest investigation of the relationship between vision and weather conditions dates back to early 2000s [16]. Thanks to the development of more advanced visual analysis and deep learning methods, a new wave of works studying the correlation between visual appearance and ambient temperature or other weather properties has recently emerged [5][20]. Inspired by these works, here we mainly focus on ambient temperature prediction from a single image or a sequence of images. We will investigate different deep learning approaches and demonstrate that the state-of-the-art performance can be obtained. Given an outdoor image, we aim to estimate ambient temperature based on visual information extracted from the image. We consider two application scenarios. The first one regards estimating the temperature of any given image regardless of temporal changes in the weather. In this case, only visual information extracted from the image is used. The second scenario is when several images of the same location over time are available and the goal is to “forecast” the temperature in the near future. In the first scenario, we extract visual features using a convolutional neural network (CNN), followed by a regression layer outputting the estimated temperature. In the second scenario, we also extract features using CNNs, but further consider temporal evolution by recruiting a long-short term memory (LSTM) network, which outputs the the estimated temperature of the last image in the given image sequence. Our contributions in this work are as follows. • We introduce a CNN-based approach to do temperature prediction for a single image and achieve promis- Figure 1. Sample scene images from the SkyFinder dataset [15], all captured at 11 am. ing prediction performance. • We introduce a recurrent neural network (RNN) to forecast temperature of the last image in a given image sequence. To the best of our knowledge, this is the first deep model considering temporal evolution of appearance for temperature prediction. We show that our RNN-based model outperforms existing approaches. • We provide a comprehensive analysis of images captured in different scenes. Interesting discussions are provided on the basis of various experimental settings and their corresponding results. 2. Related Works As a pioneering work studying visual manifestations of different weather conditions, Narasimhan and Nayar [16] firstly discussed the relationship between visual appearance and weather conditions. Since then several works have been proposed to target the weather type classification task (i.e., for weather types such as sunny, cloudy, foggy, snowy, and rainy). Roser and Moosmann [17] analyzed the images captured by the camera mounted on vehicles, and constructed an SVM classifier to categorize images into clear, light rain, and heavy rain weather conditions. In [13], five weather features including sky, shadow, reflection, contrast, and haze were extracted, and a collaborative learning framework was proposed to classify images into sunny or cloudy. In [3], a random forest classifier was proposed to integrate various types of visual features and classify images into one of five weather conditions including sunny, cloudy, snowy, rainy, or foggy. In addition to weather condition classification, more weather properties have also been investigated. Jacobs and his colleagues [9] initiated a project for collecting outdoor scene images captured by static webcams over a long period of time. The collected images form the Archive of Many Outdoor Scene (AMOS) dataset [9]. Based on the AMOS dataset, they proposed that webcams installed across the earth can be viewed as image sensors and can enable us to understand weather patterns and variations over time [8]. More specifically, they adopted principal component analysis and canonical correlation analysis to predict wind velocity and vapor pressure from a sequence of images. Recently, Laffont et al. [12] estimated scene attributes like lighting, weather conditions, and seasons for images captured by webcams based on a set of regressors. Glasner et al. [5] studied the correlation between pixel intensity/camera motion and temperature and found a moderate correlation. With this observation, a regression model considering pixel intensity was constructed to predict temperature. Following the discussion in [5], Volokitin et al. [19] showed that, with appropriate fine tuning, deep features can be promising for temperature prediction. Zhou et al. [20] proposed a selective comparison learning scheme and showed that the state-of-the-art temperature prediction performance can be obtained by a CNN-based approach. In this work, we aim at predicting temperature from a single image, as well as forecasting the temperature of the last image in a given image sequence. Deep learning approaches will be developed to consider temporal evolution of visual appearance, and promising performance will be shown. Compared with [5], [19], and [20], we particularly advocate the importance of modeling temporal evolution with designed deep networks. 3. Data Collection and Processing 3.1. Datasets Two tasks are considered in this work. One is predicting temperature from a single image, and another is predicting the temperature of the last image in an image sequence. These two tasks are proceeded based on two datasets. For the first task, the SkyFinder dataset [15] consisting of a large scale webcam images annotated with sky regions is adopted. This dataset contains roughly 90,000 labeled outdoor images captured by 53 cameras in a wide range of weather and illumination conditions. By filtering out unlinked data and data without temperature information, we finally retain images captured by 44 cameras in the experiments. Currently, for each camera, only the images captured around 11am on each day are selected. Such selection decreases sunlight variations due to capturing at different time instants on a day. Finally, we are left with 35,417 images in total. We denote this image set as Dataset 1 in what follows. Fig. 1 shows five sample images from this set. For the second task, the scene images mentioned in the Glasner dataset [5] are used as the seed. The Glasner dataset consists of images continuously captured by 10 cameras in 10 different environments for two consecutive years. Figure 2 shows snapshots of 3 environments (3 for each). These cameras are in fact a small subset of the AMOS dataset [9], classes, represented as a 70D vector. Each dimension in this vector corresponds to a specific degree. That is, the first dimension encodes −20◦ C, the second dimension encodes −19◦ C, and so on. Given an image, we attempt to classify it into one of the 70 classes. We utilize two ways to encode class labels as follows. • One-hot encoding: Only the dimension representing the correct temperature is set to 1, and other dimensions are set to 0. For example, the 70D label vector y = (0, 0, 1, 0, ..., 0) denotes that the corresponding image has temperature of −18◦ C. • Local distribution encoding (LDE): Motivated by [1], for the image with temperature corresponding to the ith dimension, we set the value tj of the jth dimension of the label vector by a Gaussian distribution tj = Figure 2. Sample scenes from Dataset 2. First two rows correspond to the same environments in different seasons (Jan. & Aug.). and from each of these ten cameras, the Glasner dataset only contains one image captured closest to 11am local time on each day. To build the proposed model, we need more data for training and testing. Therefore, according to the camera IDs mentioned in the Glasner dataset, we collect the entire set of its corresponding images from the AMOS dataset. In addition, according to the geographical information and the timestamp associated with each image, we obtain the temperature of each image from the cli-MATE website1 . Overall, we collect 53,378 images from 9 cameras2 in total. We denote this dataset as Dataset 2 in what follows. The first two rows of Fig. 2 shows two different image sequences from the same scene in Dataset 2. The first row shows images captured on Jan. 12, Jan. 13, and Jan. 14, while the second row shows images captured on Aug. 14, Aug. 15, and Aug. 16. We see that visual appearance of images captured at the same scene may drastically vary due to climate changes. In addition, we clearly observe the temporal continuity of images captured on consecutive days. The third and fourth rows of Fig. 2 show two more image sequences captured at different scenes. 3.2. Label Encoding Regardless of the temperature prediction task (i.e., for a single image or an image sequence), we formulate temperature estimation as a classification problem. We divide the considered temperature range (−20◦ C to 49◦ C) into 70 1 http://mrcc.isws.illinois.edu/CLIMATE/ 2 One camera’s information is incorrect, and we could not successfully collect the corresponding temperature values (j − i)2 1 √ exp( ). 2σ 2 σ 2π (1) We will compare two encoding schemes in the evaluation section. 4. Proposed Methods This section presents details of the proposed models for temperature prediction. We are interested in predicting the temperature for a given still image as well as predicting the temperature of a location from temperature data of the same location at the same time on previous days (i.e., next day prediction). 4.1. Temperature Estimation from a Single Image To estimate ambient temperature of a given outdoor image, we construct a convolutional neural network (CNN) trained from scratch based on the SkyFinder dataset. Table 1 shows detailed configurations of our CNN architecture. There are totally 4 convolutional layers, followed by 4 fully-connected layers. The model’s output is a 70D vector indicating the probabilities of different temperatures (i.e., classes). To train the model, the activation function of each layer is ReLU, the loss function is cross entropy, the optimization algorithm is Adam, and learning rate is 0.001. Different batch sizes are tested to achieve the best performance for different experimental settings. 80% of the SkyFinder dataset is used for training, and the remaining 20% is used for testing. We do not use a validation set and train the network for 90 epochs. 4.2. Temperature Estimation from a Sequence of Images Given an image sequence I1 , I2 , ..., In captured at the same scene, and assuming that the corresponding temperature values of the first n − 1 images, i.e., t1 , t2 , ..., tn−1 Table 1. Detailed configurations of the convolutional neural network. Conv2D(32, 3, 3) Conv2D(64, 3, 3) Dense(512) Dense(70) Conv2D(32, 3, 3) MaxPooling2D(2, 2) Conv2D(64, 3, 3) MaxPooling2D(2, 2) Flatten Dropout(0.5) Softmax Dropout(0.25) Dropout(0.25) 5. Empirical Evaluation 5.1. Implementation Details Figure 3. Structure of the proposed LSTM network for temperature prediction. are available, we would like to predict the temperature tn of the image In . In addition to the visual information of In , temporal evolution of the visual appearance of I1 to In−1 may provide information regarding future temperatures. Therefore, we construct a long-short term memory network (LSTM) [7] to successively propagate visual information over time to predict temperature. Fig. 3 shows the model structure. Each image in the sequence is first fed to a CNN to extract visual features. This CNN has the same structure as mentioned in in Sec. 4.1, without the last softmax layer. The extracted feature vector of the image Ii is then fed to an LSTM layer, which not only processes the current input, but also considers the information propagated from the intermediate result for the image Ii−1 . Similarly, the intermediate result for the image Ii will be sent to the LSTM layer for processing the image Ii+1 . The output of the LSTM layer would be input to an embedding layer that transforms the input vector into a 70D vector y i indicating the probabilities of different temperatures. To train the RNN, the loss function is defined as the mean square error between the ground truth and the predicted vector. That is, n X L= ky i − ti k2 . (2) i=1 Fig. 3 shows that information is only propagated from I1 to In . In addition to this forward propagation scheme, we also try to construct a bi-directional LSTM model to consider both forward propagation and backward propagation. Detailed performance comparison will be shown later. To estimate ambient temperature of a single image, the CNN model mentioned in Sec. 4.1 is constructed based on 80% of the images in Dataset 1, while the remaining 20% of the images are used for testing. We repeat this process five times with the five-fold cross validation scheme, and report the average estimation error. In this task, the standard deviation σ mentioned in the LDE scheme is empirically set as 3.5. More details for predicting the temperature of the last image of an image sequence are provided as follows. For a (t) (t) (t) camera, let I08 , I09 , ..., I17 denote the images with timestamps closest to 8am, 9am, ..., 17pm, on day t. To train the RNN mentioned in Sec. 4.2, we select a sequence consisting of n images that were captured around the same time (t) (t+1) (t+n−1) on n consecutive days, like (I08 , I08 , ..., I08 ) and (t) (t+1) (t+n−1) (I15 , I15 , ..., I15 ). We always keep images captured closest to 11am as the testing data, and exclude them in the training data set. The training image sequences collected from all cameras are put together to be the generic training dataset, and the (generic) RNN is trained based on it. In this task, the standard deviation σ mentioned in the LDE scheme is empirically set as 4. (t) (t+1) (t+n−1) Given a test image sequence (I11 , I11 , ..., I11 ), assuming that we have known the temperature values of (t) (t+1) (t+n−2) (I11 , I11 , ..., I11 ), we would like to predict the (t+n−1) temperature values of I11 . In our experiments, we found that the best performance can be obtained when n is set as three. That is, given the images captured on day t and day t + 1, we predict the temperature of the image captured on day t + 2. 5.2. Performance of Temperature Estimation for Single Images Similar to [5], [19], and [20], we evaluate the prediction performance based on the root mean squared error, RMSE, between the estimated temperature and the ground truth. The dataset used in [5], [19], and [20] is relatively small, i.e., images captured by ten cameras, and from each day Table 2. Root mean square errors when different percentages of data in Dataset 1 used during training. Avg. RMSE 20% 5.34 40% 4.86 60% 4.55 80% 4.28 Table 3. Temperature estimation errors, in terms of RMSE, for each of the 9 scenes in the Glasner dataset. Method ↓ \ Scene → Local Regression [5] LR Temporal Win. [5] Global Ridge Reg. [5] CNN [5] Fine-tuned VGG-16 [19] Selective Comparison Learning [20] Our CNN (LDE) Our CNN (one hot) (a) 6.85 7.52 18.16 8.55 7.79 6.26 4.38 6.28 (b) 7.24 6.86 5.74 5.59 4.87 4.57 3.79 5.48 (c) 6.03 5.82 35.02 5.96 5.03 4.92 4.34 5.78 only one image is selected. This may impede the development of a robust model. For example, in [5], the authors report that the CNN-based approach does not work very well, but in [19], the authors show that more promising performance can be obtained by appropriately fine-tuning a pretrained CNN model, i.e., VGG-16 [18]. Volume of training data. In this evaluation, we first focus on studying how the volume of training data influences temperature estimation performance, based on a larger-scale dataset, i.e., Dataset 1. We train 4 CNN models on 20%, 40%, 60%, and 80% of the data in Dataset 1, respectively, and use the same remaining 20% of data for testing. As shown in Table 2, unsurprisingly more training data yield better performance. Performance comparison. To fairly compare our CNN model with existing works, we train the proposed CNN model on the Glasner dataset. Note that volume of the Glasner dataset is smaller than Dataset 1. Table 3 shows comprehensive performance comparison, where ten cameras in the Glasner dataset are referred to as scenes (a)-(j). We find that results of [19] are superior to [5], which demonstrates fine-tuning a deep model yields performance gain. Comparing our CNN (with the LDE scheme) with [19], our CNN is much simpler than VGG-16 and achieves even better results. The performance gain might be due to the LDE scheme, which may provide more flexibility to model temperature fluctuation. To verify this, we show performance obtained by our CNN model with the one-hot encoding scheme in the last row of Table 3. By comparing the last two rows, we clearly see the benefits brought by the LDE scheme. Performance of the work in [20] remains the best. A two-stream network was proposed to take an image pair captured at the same place and on the same day as input. Each network stream extracts features from one image, and estimates the temperature of each image. To train this two-stream network, in addition to the softmax losses given by each stream, a ranking loss representing the temperature difference between two images is jointly considered. Our CNN model with LDE fares on par with [20] but is much simpler. Further, the idea of ranking loss can be integrated in our model to make performance gain in the future. (d) 4.53 4.23 11.37 6.17 4.45 3.80 4.60 6.69 (e) 5.77 5.39 43.51 7.30 4.22 3.98 4.47 6.36 (f) 3.19 2.77 3.84 2.90 3.14 2.53 5.83 7.69 (g) 7.63 7.35 5.54 8.48 6.61 5.20 5.36 6.78 (h) 5.09 5.22 13.86 4.88 4.72 3.95 3.73 6.02 (i) 3.68 3.67 3.41 2.93 2.70 2.48 2.81 3.25 (j) 7.77 7.85 8.91 7.12 6.01 5.81 6.47 7.54 Avg. 5.78 5.67 14.94 5.99 4.96 4.35 4.58 6.19 Table 4. Root mean square errors when different lengths of image sequences are used. Avg. RMSE n=2 2.97 n=3 2.82 n=4 2.95 n=5 2.87 n=6 2.86 n=7 2.98 5.3. Performance of Temperature Estimation for Image Sequences Length of the image sequence. We first evaluate estimation performance when different lengths of image sequences are used for training and testing, and show the average RMSEs in Table 4. As can be seen, the best estimation performance can be obtained when n is set as 3. That is, we estimate the temperature of the day t + 2 based on day t + 1 and day t. This result is not surprising because this setting appropriately considers information of previous days, and prevents blunt updates when too many days are considered. Therefore, we set n = 3 in the subsequent experiments. Performance comparison. Table 5 shows performance of the proposed single LSTM and bidirectional LSTM, and compares performance with the state of the art methods. Previously only the frequency decomposition method proposed in [5] takes temporal evolution into account. Note that the data used in [5] are from the Glasner dataset. Our LSTM models are trained and tested on Dataset 2 that is also from the Glasner dataset but includes much more images to capture temporal evolution. We discarded scene (d) as its data have errors in the geographical information and we could not access corresponding temperature information. As can be seen from Table 5, excluding camera (d), the RNN-based approach achieves better performance over eight of the nine scenes3 . The proposed RNN-based approach outperforms the state of the art. The frequency decomposition method attempts to model the temporal evolution of the entire year, while we consider local temporal evolution and achieve much better performance. The structure shown in Fig. 3 shows that the LSTM model works based on features extracted by the CNN de3 The average RMSE of the frequency decomposition method is calculated based on ten scenes, while that of our LSTM-based methods is calculated based on nine scenes. Table 5. Temperature estimation errors (in terms of RMSE) for each scene in the Glasner dataset. Method ↓ \ Scene → Freq. Decom. [5] Our single LSTM Our bidirectional LSTM Handcrafted + our bi-LSTM (a) 5.05 6.01 7.09 10.84 (b) 6.15 2.32 2.54 6.60 (c) 4.61 2.53 2.49 11.79 signed like Table 1. To demonstrate the influence of the CNN, we replace the CNN features by a handcrafted feature, say Gabor wavelet texture feature [14], and train a bidirectional LSTM based on it. The performance obtained based on the handcrafted features is shown in the last row of Table 5. The average RMSE is 8.93, which is much higher than 2.80 and evidently demonstrates the effectiveness of CNN features. Overall, Table 5 shows that temporal evolution of visual appearance provides rich information in estimating ambient temperature, and yields substantially better estimation performance. To our knowledge, this result is the first work considering temporal evolution by a deep neural network to do ambient temperature estimation. Sky only vs. ground only. The results shown in Table 5 are obtained when visual information is extracted from the entire image. Intuitively, one may think that the sky region provides more information in estimating weather properties. This is why the works [13] and [3] extracted cloud features from the sky region to do weather condition classification. On the other hand, the first two rows of Fig. 2 show that ground appearance sometimes provides important clues regarding temperature, e.g., whenever there is snow on the ground, it is colder. To quantify the influences of sky and ground regions on temperature estimation, first we manually segment the sky region for each scene. Then, the minimum bounding box of the sky region is fed to the proposed bidirectional LSTM model to obtain the performance yielded by sky-only information. We also feed the minimum bounding box of the ground region to obtain corresponding performance. Fig. 4 shows samples of one scene, where the top row shows the original scene image and the corresponding sky mask, and the bottom row shows the minimum bounding boxes of the sky region and the ground region, respectively. Notice that because the sky region is irregular, the minimum bounding box may include parts of the ground, and vice versa. Table 6 shows performance variations when visual information is extracted from different regions of the images in image sequences. Comparing the performance yielded by sky only with ground only, we found that in fact the ground region provides more information for temperature estimation. Visual variations in the sky region may just represent texture of cloud and intensity, and provide relatively fewer clues for estimating temperature. According to this exper- (d) 4.10 – – – (e) 5.27 2.26 1.93 8.91 (f) 2.51 2.46 2.61 11.71 (g) 4.51 2.93 2.36 7.93 (h) 4.50 2.04 1.89 9.47 (i) 2.81 1.62 1.57 3.50 (j) 5.16 3.02 2.93 9.58 Avg. 4.47 2.80 2.82 8.93 Figure 4. Top row: one sample scene image (left), and the corresponding mask of the sky region (right). Bottom row: the minimum bounding box of the sky region (left), and the minimum bounding box of the ground region (right). Table 6. Root mean square errors when visual information is extracted from different regions. Avg. RMSE Sky only 3.42 Ground only 2.93 Entire Image 2.82 iment, we see that the best performance can be obtained when the entire image is considered. This conforms to the setting of previous works [5] and [20]. Variations of different daytime hours. As mentioned in Sec. 5.1, only the image sequences captured at 11am are tested to obtain performance shown above. The main reason to use such setting is to align the settings of [5] and [20] to make a fair comparison. Besides, we are interested in whether image sequences captured in different daytime hours would yield performance variations. To show this, we respectively use image sequences at h as the testing data, and the remaining is used for training. The hour h is from 8am to 17pm. Fig. 5 shows variations of average RMSEs for image sequences captured in different daytime hours. Interestingly, the variation is significant when images captured in different daytime hours are processed. The best performance is obtained for images captured at 11 pm, which conforms to the selection of [5] and [20]. On the other hand, estimation errors at 8pm and 17am are much larger than others. This may be because the sunlight is maximal around the noon, and more robust visual information can be extracted. Sample predictions. Fig. 6 shows ground truth temperatures (red curve) and the estimated temperatures (blue curve) of images captured at 11am by the camera (f) mentioned in Table 5 over two years (Jan. 2013 to Dec. 2014). Figure 5. Variations of average RMSEs for image sequences captured in different daytime hours. Basically, we see that two curves correlate well with each other, conforming to the promising performance shown in Table 5. We also clearly observe the trend of climate change, i.e., higher temperature in July and August, and lower temperature in January and February. Fig. 7 and Fig. 8 show “successful estimation cases” and “failure estimation cases”, respectively. Looking only at one image may not allow us to figure out why temperature of some images are easier to estimate, and some are not. After a deeper inspection, we found that usually the failure cases occur when the temperature values of consecutive days fluctuate significantly. This is expectable. We also found some ground truths are obviously incorrect, e.g., a snowy scene has the truth temperature of 42◦ C. These errors need to be fixed manually in the future to elaborate more robust performance evaluation. It is interesting to know what visual clues are more important in temperature prediction. To get some insights for this issue, we take all images captured closest to 11am by a camera, and divide each image into non-overlapping 5 × 5 block. Based on pixels’ RGB colors in blocks corresponding to the same position, e.g., the ith block in images captured on the first day, the second day, and so on, we calculate the standard deviations of R, G, and B, and then average these three standard deviations as a value ρ. The value ρ is then normalized into the range from 0 to 255 by ρ − minρ ρ̂ = × 255, where maxρ and minρ are the maxρ − minρ maximum ρ’s and minimum ρ’s of all blocks. Finally, we visualize the value ρ̂ of each block. A block with a larger ρ̂ value indicates higher color variations on different days, and probably conveys more information in predicting temperature. Fig. 9 shows the visualization results corresponding to the scenes shown in Fig. 7. We see that, for the first scene, blocks on tree leaves are more salient, and may be more informative in temperature prediction. For the second scene, blocks on some parts of building facades are more salient. Figure 6. Ground truth temperatures and the estimated temperatures of images captured at 11am by the camera (f) over two years. Figure 7. Sample images of “successful cases”. For the left image, the truth temperature is 19◦ C, and the proposed bidirectional LSTM model estimates it as 19◦ C. For the right image, the truth temperature and the estimated one are 10◦ C and 11◦ C, respectively. Figure 8. Samples image of “failure cases”. For the left image, the truth temperature and the estimated one are 21◦ C and 38◦ C, respectively. For the right image, the truth temperature and the estimated one are 13◦ C and 21◦ C, respectively. Figure 9. Saliency maps showing the importance of different pixels on temperature prediction. These results are similar to the discussion mentioned in [5], and could be interesting clues for future studies. 6. Conclusion and Discussion In this work, we present deep models to estimate ambient temperature of a single image or the last image in an image sequence. For the first task, we verify that the CNNbased approach is promising, and when more training data is available better performance can be obtained. For the second task, we consider temporal evolution of visual appearance, and propose an RNN-based approach to “forecast” the temperature of the last image. State-of-the-art performance can be obtained by the proposed method. We also discuss how performance varies when information is just extracted from the sky region or from the ground region, and how performance varies when images captured in different daytime hours are processed. In the future, more weather properties like humidity and wind speed may be estimated by the proposed methods. Furthermore, we can investigate which region in a scene provides more clues in temperature estimation, based on the currently emerged attention networks. Exploring the relationship between weather properties and vision would be interesting in a wide range of works and applications. Acknowledgement This work was partially supported by the Ministry of Science and Technology of Taiwan under the grant MOST 1052628-E-194-001-MY2 and MOST 106-3114-E-002-009. References [1] G. Antipov, M. Baccouche, S.-A. Berrania, and J.-L. Dugelay. Effective training of convolutional neural networks for face-based gender and age prediction. Pattern Recognition, 72:15–26, 2017. [2] W.-T. Chu and Y.-L. Wu. Deep correlation features for image style classification. In Proceedings of ACM International Conference on Multimedia, pages 402–406, 2016. [3] W.-T. Chu, X.-Y. Zheng, and D.-S. Ding. Camera as weather sensor: Estimating weather information from single images. Journal of Visual Communication and Image Representation, 46:233–249, 2017. [4] V. Ferrari and A. Zisserman. Learning visual attributes. In Proceedings of Advances in Neural Information Processing Systems, pages 433–440, 2007. [5] D. Glasner, P. Fua, T. Zickler, and L. Zelnik-Manor. Hot or not: Exploring correlations between appearance and temperature. In Proceedings of IEEE International Conference on Computer Vision, 2015. [6] J. Hays and A.Efros. Im2gps: Estimating geographic information from a single image. In Proceedings of IEEE International Conference on Computer Vision and Patten Recognition, 2008. [7] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, 1997. [8] N. Jacobs, W. Burgin, R. Speyer, D. Ross, and R. Pless. Adventures in archiving and using three years of webcam images. In Proceedings of IEEE CVPR Workshop on Internet Vision, pages 39–46, 2009. [9] N. Jacobs, N. Roman, and R. Pless. Consistent temporal variations in many outdoor scenes. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition, 2007. [10] D. Jayaraman, F. Sha, and K. Grauman. Decorrelating semantic visual attributes by resisting the urge to share. In Proceedings of IEEE International Conference on Computer Vision and Patten Recognition, 2014. [11] S. Karayev, M. Trentacoste, H. Han, A. Agarwala, T. Darrell, A. Hertzmann, and H. Winnemoeller. Recognizing image style. In Proceedings of British Machine Vision Conference, 2014. [12] P. Laffont, Z. Ren, X. Tao, C. Qian, and J. Hays. Transient attributes for high-level understanding and editing of outdoor scenes. ACM Transactions on Graphics (Proceedings of SIGGRAPH 2014), 33(4), 2014. [13] C. Lu, D. Lin, J. Jia, and C.-K. Tang. Two-class weather classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(12):2510–2524, 2017. [14] B. Manjunath and W. Ma. Texture features for browsing and retrieval of image data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 18(8):837–842, 1996. [15] R. P. Mihail, S. Workman, Z. Bessinger, and N. Jacobs. Sky segmentation in the wild: An empirical study. In Proceedings of IEEE Winter Conference on Applications of Computer Vision, pages 1–6, 2016. [16] S. Narasimhan and S. Nayar. Vision and atmosphere. International Journal of Computer Vision, 48(3):233–354, 2002. [17] M. Roser and F. Moosmann. Classification of weather situations on single color images. In Proceedings of IEEE Intelligent Vehicles Symposium, pages 798–803, 2008. [18] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In arXiv preprint arXiv:1409.1556, 2014. [19] A. Volokitin, R. Timofte, and L. Van Gool. Deep features or not: Temperature and time prediction in outdoor scenes. In Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition Workshop, pages 63– 71, 2016. [20] H.-Y. Zhou, B.-B. Gao, and J. Wu. Sunrise or sunset: Selective comparison learning for subtle attribute recognition. In Proceedings of British Machine Vision Conference, 2017.
1cs.CV
Higher-Order Program Verification via HFL Model Checking arXiv:1710.08614v2 [cs.PL] 28 Feb 2018 Naoki Kobayashi, Takeshi Tsukada, and Keiichi Watanabe The University of Tokyo Abstract. There are two kinds of higher-order extensions of model checking: HORS model checking and HFL model checking. Whilst the former has been applied to automated verification of higher-order functional programs, applications of the latter have not been well studied. In the present paper, we show that various verification problems for functional programs, including may/must-reachability, trace properties, and linear-time temporal properties (and their negations), can be naturally reduced to (extended) HFL model checking. The reductions yield a sound and complete logical characterization of those program properties. Compared with the previous approaches based on HORS model checking, our approach provides a more uniform, streamlined method for higher-order program verification.1 1 Introduction There are two kinds of higher-order extensions of model checking in the literature: HORS model checking [18, 34] and HFL model checking [46]. The former is concerned about whether the tree generated by a given higher-order tree grammar called a higher-order recursion scheme (HORS) satisfies the property expressed by a given modal µ-calculus formula (or a tree automaton), and the latter is concerned about whether a given finite state system satisfies the property expressed by a given formula of higher-order modal fixpoint logic (HFL), a higher-order extension of the modal µ-calculus. Whilst HORS model checking has been applied to automated verification of higher-order functional programs [19, 20, 25, 28, 35, 45, 47], there have been few studies on applications of HFL model checking to program/system verification. Despite that HFL has been introduced more than 10 years ago, we are only aware of applications to assume-guarantee reasoning [46] and process equivalence checking [30]. In the present paper, we show that various verification problems for higherorder functional programs can actually be reduced to (extended) HFL model checking in a rather natural manner. We briefly explain the idea of our reduction below.2 We translate a program to an HFL formula that says “the program has a valid behavior” (where the validity of a behavior depends on each verification 1 2 A shorter version of this article is published in Proceedings of ESOP 2018. In this section, we use only a fragment of HFL that can be expressed in the modal µ-calculus. Some familiarity with the modal µ-calculus [27] would help. problem). Thus, a program is actually mapped to a property, and a program property is mapped to a system to be verified; this has been partially inspired by the recent work of Kobayashi et al. [21], where HORS model checking problems have been translated to HFL model checking problems by switching the roles of models and properties. For example, consider a simple program fragment read(x); close(x) that reads and then closes a file (pointer) x. The transition system in Figure 1 shows a valid access protocol to read-only files. Then, the property that a read operation is allowed in the current state can be expressed by a formula of the form hreadiϕ, which says that the current state has a read-transition, after which ϕ is satisfied. Thus, the program read(x); close(x) being valid is expressed as hreadihcloseitrue,3 which is indeed satisfied by the initial state q0 of the transition system in Figure 1. Here, we have just replaced the operations read and close of the program with the corresponding modal operators hreadi and hclosei. We can also naturally deal with branches and recursions. For example, consider the program close(x)(read(x); close(x)), where e1 e2 represents a non-deterministic choice between e1 and e2 . Then the property that the program always accesses x in a valid manner can be expressed by (hcloseitrue) ∧ (hreadihcloseitrue). Note that we have just replaced the non-deterministic branch with the logical conjunction, as we wish here to require that the program’s behavior is valid in both branches. We can also deal with conditional branches if HFL is extended with predicates; if b then close(x) else (read(x); close(x)) can be translated to (b ⇒ hcloseitrue) ∧ (¬b ⇒ hreadihcloseitrue). Let us also consider the recursive function f defined by: f x = close(x)(read(x); read(x); f x), Then, the program f x being valid can be represented by using a (greatest) fixpoint formula: νF.(hcloseitrue) ∧ (hreadihreadiF). If the state q0 satisfies this formula (which is indeed the case), then we know that all the file accesses made by f x are valid. So far, we have used only the modal µ-calculus formulas. If we wish to express the validity of higher-order programs, we need HFL formulas; such examples are given later. We generalize the above idea and formalize reductions from various classes of verification problems for simply-typed higher-order functional programs with recursion, integers and non-determinism – including verification of may/must-reachability, trace properties, and linear-time temporal properties (and their negations) – to (extended) HFL model checking where HFL is extended with integer predicates, and prove soundness and completeness of the reductions. Extended HFL model checking problems obtained by the reductions are (necessarily) undecidable in general, but for finite-data programs (i.e., 3 Here, for the sake of simplicity, we assume that we are interested in the usage of the single file pointer x, so that the name x can be ignored in HFL formulas; usage of multiple files can be tracked by using the technique of [19]. 2 read ?>=< 89:; q0 t close ?>=< q1 4 89:; Fig. 1. File access protocol programs that consist of only functions and data from finite data domains such as Booleans), the reductions yield pure HFL model checking problems, which are decidable [46]. Our reductions provide sound and complete logical characterizations of a wide range of program properties mentioned above. Nice properties of the logical characterizations include: (i) (like verification conditions for Hoare triples,) once the logical characterization is obtained as an HFL formula, purely logical reasoning can be used to prove or disprove it (without further referring to the program semantics); for that purpose, one may use theorem provers with various degrees of automation, ranging from interactive ones like Coq, semiautomated ones requiring some annotations, to fully automated ones (though the latter two are yet to be implemented), (ii) (unlike the standard verification condition generation for Hoare triples using invariant annotations) the logical characterization can automatically be computed, without any annotations,4 (iii) standard logical reasoning can be applied based on the semantics of formulas; for example, co-induction and induction can be used for proving ν- and µ-formulas respectively, and (iv) thanks to the completeness, the set of program properties characterizable by HFL formula is closed under negations; for example, from a formula characterizing may-reachability, one can obtain a formula characterizing non-reachability by just taking the De Morgan dual. Compared with previous approaches based on HORS model checking [20, 25, 28, 35, 41], our approach based on (extended) HFL model checking provides more uniform, streamlined methods for higher-order program verification. HORS model checking provides sound and complete verification methods for finite-data programs [19, 20], but for infinite-data programs, other techniques such as predicate abstraction [25] and program transformation [29, 33] had to be combined to obtain sound (but incomplete) reductions to HORS model checking. Furthermore, the techniques were different for each of program properties, such as reachability [25], termination [29], non-termination [28], fair termination [33], and fair non-termination [47]. In contrast, our reductions are sound and complete even for infinite-data programs. Although the obtained HFL model checking problems are undecidable in general, the reductions allow us to treat various program properties uniformly; all the verifications are boiled down to the issue of how to prove µ- and ν-formulas (and as remarked above, we can 4 This does not mean that invariant discovery is unnecessary; invariant discovery is just postponed to the later phase of discharging verification conditions, so that it can be uniformly performed among various verification problems. 3 use induction and co-induction to deal with them). Technically, our reduction to HFL model checking may actually be considered an extension of HORS model checking in the following sense. HORS model checking algorithms [23, 34] usually consist of two phases, one for computing a kind of higher-order “procedure summaries” in the form of variable profiles [34] or intersection types [23], and the other for nested least/greatest fixpoint computations. Our reduction from program verification to extended HFL model checking (the reduction given in Section 7, in particular) can be regarded as an extension of the first phase to deal with infinite data domains, where the problem for the second phase is expressed in the form of extended HFL model checking: see Appendix H for more details. The rest of this paper is structured as follows. Section 2 introduces HFL extended with integer predicates and defines the HFL model checking problem. Section 3 informally demonstrates some examples of reductions from program verification problems to HFL model checking. Section 4 introduces a functional language used to formally discuss the reductions in later sections. Sections 5, 6, and 7 consider may/must-reachability, trace properties, and temporal properties respectively, and present (sound and complete) reductions from verification of those properties to HFL model checking. Section 8 discusses related work, and Section 9 concludes the paper. Proofs are found in Appendices. 2 (Extended) HFL In this section, we introduce an extension of higher-order modal fixpoint logic (HFL) [46] with integer predicates (which we call HFLZ ; we often drop the subscript and just write HFL, as in Section 1), and define the HFLZ model checking problem. The set of integers can actually be replaced by another infinite set X of data (like the set of natural numbers or the set of finite trees) to yield HFLX . 2.1 Syntax For a map f , we write dom( f ) and codom( f ) for the domain and codomain of f respectively. We write Z for the set of integers, ranged over by the meta-variable n below. We assume a set Pred of primitive predicates on integers, ranged over by p. We write arity(p) for the arity of p. We assume that Pred contains standard integer predicates such as = and <, and also assume that, for each predicate p ∈ Pred, there also exists a predicate ¬p ∈ Pred such that, for any integers n1 , . . . , nk , p(n1 , . . . , nk ) holds if and only if ¬p(n1 , . . . , nk ) does not hold; thus, ¬p(n1 , . . . , nk ) should be parsed as (¬p)(n1, . . . , nk ), but can semantically be interpreted as ¬(p(n1 , . . . , nk )). The syntax of HFLZ formulas is given by: ϕ (formulas) ::= n | ϕ1 op ϕ2 | true | false | p(ϕ1 , . . . , ϕk ) | ϕ1 ∨ ϕ2 | ϕ1 ∧ ϕ2 | X | haiϕ | [a]ϕ | µXτ .ϕ | νXτ .ϕ | λX : σ.ϕ | ϕ1 ϕ2 τ (types) ::= • | σ → τ σ (extended types) ::= τ | int 4 Here, op ranges over a set of binary operations on integers, such as +, and X ranges over a denumerable set of variables. We have extended the original HFL [46] with integer expressions (n and ϕ1 op ϕ2 ), and atomic formulas p(ϕ1 , . . . , ϕk ) on integers (here, the arguments of integer operations or predicates will be restricted to integer expressions by the type system introduced below). Following [21], we have omitted negations, as any formula can be transformed to an equivalent negation-free formula [32]. We explain the meaning of each formula informally; the formal semantics is given in Section 2.2. Like modal µ-calculus [11, 27], each formula expresses a property of a labeled transition system. The first line of the syntax of formulas consists of the standard constructs of predicate logics. On the second line, as in the standard modal µ-calculus, haiϕ means that there exists an a-labeled transition to a state that satisfies ϕ. The formula [a]ϕ means that after any a-labeled transition, ϕ is satisfied. The formulas µXτ .ϕ and νXτ .ϕ represent the least and greatest fixpoints respectively (the least and greatest X that X = ϕ) respectively; unlike the modal µ-calculus, X may range over not only propositional variables but also higher-order predicate variables (of type τ). The λ-abstractions λX : σ.ϕ and applications ϕ1 ϕ2 are used to manipulate higher-order predicates. We often omit type annotations in µXτ .ϕ, νXτ .ϕ and λX : σ.ϕ, and just write µX.ϕ, νX.ϕ and λX.ϕ. Example 1. Consider ϕab ϕ where ϕab = µX•→• .λY : •.Y ∨ hai(X(hbiY)). We can expand the formula as follows: ϕab ϕ = (λY. • .Y ∨ hai(ϕab(hbiY)))ϕ = ϕ ∨ hai(ϕab(hbiϕ)) = ϕ ∨ hai(hbiϕ ∨ hai(ϕab(hbihbiϕ))) = · · · , and obtain ϕ ∨ (haihbiϕ) ∨ (haihaihbihbiϕ) ∨ · · ·. Thus, the formula means that there is a transition sequence of the form an bn for some n ≥ 0 that leads to a state satisfying ϕ. Following [21], we exclude out unmeaningful formulas such as (haitrue) + 1 by using a simple type system.5 The types •, int, and σ → τ describe propositions, integers, and (monotonic) functions from σ to τ, respectively. Note that the integer type int may occur only in an argument position; this restriction is required to ensure that least and greatest fixpoints are well-defined. The typing rules for formulas are given in Figure 2. In the figure, ∆ denotes a type environment, which is a finite map from variables to (extended) types. Below we consider only well-typed formulas, i.e., formulas ϕ such that ∆ ⊢H ϕ : τ for some ∆ and τ. 2.2 Semantics and HFLZ Model Checking We now define the formal semantics of HFLZ formulas. A labeled transition system (LTS) is a quadruple L = (U, A, −→, sinit), where U is a finite set of states, A is a 5 The original type system of [46] was more complex due to the presence of negations. 5 ∆ ⊢H n : int ∆ ⊢H ϕi : int for each i ∈ {1, 2} ∆ ⊢H ϕ1 op ϕ2 : int (HT-Int) ∆ ⊢H ϕi : • for each i ∈ {1, 2} (HT-Op) ∆ ⊢ H ϕ1 ∧ ϕ2 : • ∆ ⊢H ϕ : • ∆ ⊢H haiϕ : • ∆ ⊢H ϕ : • (HT-True) ∆ ⊢H true : • ∆ ⊢H [a]ϕ : • ∆, X : τ ⊢H ϕ : τ (HT-False) ∆ ⊢H false : • arity(p) = k ∆ ⊢H ϕi : int for each i ∈ {1, . . . , k} ∆ ⊢H µXτ . ϕ : τ ∆, X : τ ⊢H ϕ : τ ∆ ⊢H p(ϕ1 , . . . , ϕk ) : • ∆ ⊢H νXτ . ϕ : τ ∆, X : σ ⊢H ϕ : τ ∆, X : σ ⊢H X : σ ∆ ⊢H ϕi : • for each i ∈ {1, 2} (HT-Pred) (HT-Var) ∆ ⊢H λX : σ. ϕ : σ → τ ∆ ⊢ H ϕ1 : σ → τ ∆ ⊢ H ϕ2 : σ (HT-Or) ∆ ⊢ H ϕ1 ∨ ϕ2 : • ∆ ⊢ H ϕ1 ϕ2 : τ (HT-And) (HT-Some) (HT-All) (HT-Mu) (HT-Nu) (HT-Abs) (HT-App) Fig. 2. Typing Rules for HFLZ Formulas finite set of actions, −→ ⊆ U × A× U is a labeled transition relation, and sinit ∈ U a is the initial state. We write s1 −→ s2 when (s1 , a, s2 ) ∈ −→. For an LTS L = (U, A, −→, sinit) and an extended type σ, we define the partially ordered set (DL,σ , ⊑L,σ) inductively by: DL,• = 2U ⊑L,• =⊆ DL,int = Z ⊑L,int= {(n, n) | n ∈ Z} DL,σ→τ = { f ∈ DL,σ → DL,τ | ∀x, y.(x ⊑L,σ y ⇒ f x ⊑L,τ f y)} ⊑L,σ→τ = {( f, g) | ∀x ∈ DL,σ . f (x) ⊑L,τ g(x)} Note that (DL,τ , ⊑L,τ ) forms a complete lattice (but (DL,int, ⊑L,int) does not). We write ⊥L,τ and ⊤L,τ for the least and greatest elements of DL,τ (which are λe x.∅ and λe x.U) respectively. We sometimes omit the subscript L below. Let ~∆L be the set of functions (called valuations) that maps X to an element of DL,σ for each X : σ ∈ ∆. For an HFL formula ϕ such that ∆ ⊢H ϕ : σ, we define ~∆ ⊢H ϕ : σL as a map from ~∆L to Dσ , by induction on the derivation6 of ∆ ⊢H ϕ : σ, as follows. ~∆ ⊢H n : intL (ρ) = n ~∆ ⊢H true : •L (ρ) = U ~∆ ⊢H false : •L (ρ) = ∅ ~∆ ⊢H ϕ1 op ϕ2 : intL (ρ) = (~∆ ⊢H ϕ1 : intL (ρ))~op(~∆ ⊢H ϕ2 : intL (ρ)) ~∆ ⊢H p(ϕ1 , . . . , ϕk ) : •L (ρ) = ( U if (~∆ ⊢H ϕ1 : intL (ρ), . . . , ~∆ ⊢H ϕk : intL (ρ)) ∈ ~p ∅ otherwise ~∆, X : σ ⊢H X : σL (ρ) = ρ(X) ~∆ ⊢H ϕ1 ∨ ϕ2 : •L (ρ) = ~∆ ⊢H ϕ1 : •L (ρ) ∪ ~∆ ⊢H ϕ2 : •L (ρ) ~∆ ⊢H ϕ1 ∧ ϕ2 : •L (ρ) = ~∆ ⊢H ϕ1 : •L (ρ) ∩ ~∆ ⊢H ϕ2 : •L (ρ) 6 Note that the derivation of each judgment ∆ ⊢H ϕ : σ is unique if there is any. 6 a ~∆ ⊢H haiϕ : •L (ρ) = {s | ∃s′ ∈ ~∆ ⊢H ϕ : •L (ρ). s −→ s′ } a ~∆ ⊢H [a]ϕ : •L (ρ) = {s | ∀s′ ∈ U. (s −→ s′ implies s′ ∈ ~∆ ⊢H ϕ : •L (ρ))} ~∆ ⊢H µXτ .ϕ : τL (ρ) = lfpL,τ (~∆ ⊢H λX : τ. ϕ : τ → τL (ρ)) ~∆ ⊢H νXτ .ϕ : τL (ρ) = gfpL,τ (~∆ ⊢H λX : τ. ϕ : τ → τL (ρ)) ~∆ ⊢H λX : σ. ϕ : σ → τL (ρ) = {(v, ~∆, X : σ ⊢H ϕ : τL (ρ[X 7→ v])) | v ∈ DL,σ } ~∆ ⊢H ϕ1 ϕ2 : τL (ρ) = ~∆ ⊢H ϕ1 : σ → τL (ρ)(~∆ ⊢H ϕ2 : σL (ρ)) Here, ~op denotes the binary function on integers represented by op and ~p denotes the k-ary relation on integers represented by p. The least/greatest fixpoint operators lfpL,τ and gfpL,τ are defined by: F  lfpL,τ ( f ) = L,τ {x ∈ DL,τ | f (x) ⊑L,τ x} gfpL,τ ( f ) = L,τ {x ∈ DL,τ | x ⊑L,τ f (x)} F  Here, L,τ and L,τ respectively denote the least upper bound and the greatest lower bound with respect to ⊑L,τ . We often omit the subscript L and write ~∆ ⊢H ϕ : σ for ~∆ ⊢H ϕ : σL . For a closed formula, i.e., a formula well-typed under the empty type environment ∅, we often write ~ϕL or just ~ϕ for ~∅ ⊢H ϕ : σL (∅). Example 2. For the LTS Lfile in Figure 1, we have: ~νX• .(hcloseitrue ∧ hreadiX) = gfpL,• (λx ∈ DL,• .~X : • ⊢ hcloseitrue ∧ hreadiX : •({X 7→ x})) = {q0 }. In fact, x = {q0 } ∈ DL,• satisfies the equation: ~X : • ⊢ hcloseitrue ∧ hreadiX : •L ({X 7→ x}) = x, and x = {q0 } ∈ DL,• is the greatest such element. Consider the LTS L1 in Figure 3 and ϕab (hcitrue) where ϕab is the one introduced in Example 1. Then, ~ϕab (hcitrue)L1 = {q0 , q2 }. Consider the formula ϕ2 = ϕeven (hcitrue) 0, where ϕeven is: νX•→• .λY : •.λZ : int.(even(Z) ∧ Y) ∨ hai(X (hbiY) (Z + 1)). Here, even ∈ Pred is a unary predicate on integers such that even(n) if and only if n is even. Then, ϕ2 denotes the set of states from which there is a transition sequence of the form a2n b2n to a state where a c-labeled transition is possible. Thus, ~ϕ2 L1 = {q2 }. 89:; ?>=< q0 a * 89:; ?>=< b q1 j q2 c Fig. 3. LTS L1 7 * ?>=< 89:; Definition 1 (HFLZ model checking). For a closed formula ϕ of type •, we write L, s |= ϕ if s ∈ ~ϕL , and write L |= ϕ if sinit ∈ ~ϕL . HFLZ model checking is the problem of, given L and ϕ, deciding whether L |= ϕ holds. The HFLZ model checking problem is undecidable, due to the presence of integers; in fact, the semantic domain DL,σ is not finite for σ that contains int. The undecidability is obtained as a corollary of the soundness and completeness of the reduction from the may-reachability problem to HFL model checking discussed in Section 5. For the fragment of pure HFL (i.e., HFLZ without integers, which we write HFL∅ below), the model checking problem is decidable [46]. ? The order of an HFLZ model checking problem L |= ϕ is the highest order of types of subformulas of ϕ, where the order of a type is defined by: order(•) = order(int) = 0 and order(σ → τ) = max(order(σ)+1, order(τ)). The complexity of order-k HFL∅ model checking is k-EXPTIME complete [1], but polynomial time in the size of HFL formulas under the assumption that the other parameters (the size of LTS and the largest size of types used in formulas) are fixed [21]. Remark 1. Though we do not have quantifiers on integers as primitives, we can encode them using fixpoint operators. Given a formula ϕ : int → •, we can express ∃x:int.ϕ(x) and ∀x:int.ϕ(x) by (µXint→• .λx:int.ϕ(x)∨X(x−1)∨X(x+1))0 and (νXint→• .λx : int.ϕ(x) ∧ X(x − 1) ∧ X(x + 1))0 respectively. 2.3 HES As in [21], we often write an HFLZ formula as a sequence of fixpoint equations, called a hierarchical equation system (HES). Definition 2. An (extended) hierarchical equation system (HES) is a pair (E, ϕ) where E is a sequence of fixpoint equations, of the form: X1τ1 =α1 ϕ1 ; · · · ; Xnτn =αn ϕn . Here, αi ∈ {µ, ν}. We assume that X1 : τ1 , . . . , Xn : τn ⊢H ϕi : τi holds for each i ∈ {1, . . . , n}, and that ϕ1 , . . . , ϕn , ϕ do not contain any fixpoint operators. The HES Φ = (E, ϕ) represents the HFLZ formula toHFL(E, ϕ) defined inductively by: toHFL(ǫ, ϕ) = ϕ toHFL(E; Xτ =α ϕ′ , ϕ) = toHFL([αXτ .ϕ′ /X]E, [αXτ .ϕ′ /X]ϕ) Conversely, every HFLZ formula can be easily converted to an equivalent HES. In the rest of the paper, we often represent an HFLZ formula in the form of HES, and just call it an HFLZ formula. We write ~Φ for ~toHFL(Φ). An HES (X1τ1 =α1 ϕ1 ; · · · ; Xnτn =αn ϕn , ϕ) can be normalized to (X0τ0 =ν ϕ; X1τ1 =α1 ϕ1 ; · · · ; Xnτn =αn ϕn , X0 ) where τ0 is the type of ϕ. Thus, we sometimes call just a sequence of equations X0τ0 =ν ϕ; X1τ1 =α1 ϕ1 ; · · · ; Xnτn =αn ϕn an HES, with the understanding that “the main formula” is the first variable X0 . Also, we often write Xτ x1 · · · xk =α ϕ for the equation Xτ =α λx1 . · · · λxk .ϕ. We often omit type annotations and just write X =α ϕ for Xτ =α ϕ. 8 Example 3. The formula ϕ2 = ϕab (hcitrue) in Example 2 is expressed as the following HES:   X =µ λY : •.Y ∨ hai(X(hbiY)), X (hcitrue) . The formula νX.µY.hbiX ∨ haiY (which means that the current state has a transition sequence of the form (a∗ b)ω ) is expressed as the following HES:   (X =ν Y; Y =µ hbiX ∨ haiY), X .   Note that the order of the equations matters. (Y =µ hbiX ∨ haiY); X =ν Y, X represents the HFLZ formula νX.µY.hbiνX.Y ∨ haiY ≡ µY.hbiY ∨ haiY, which is equivalent to false. 3 Warming Up To help readers get more familiar with HFLZ and the idea of reductions, we give here some variations of the examples of verification of file-accessing programs in Section 1, which are instances of the “resource usage verification problem” [16]. General reductions will be discussed in Sections 5–7, after the target language is set up in Section 4. Consider the following OCaml-like program, which uses exceptions. let readex x = read x; (if * then () else raise Eof) in let rec f x = readex x; f x in let d = open_in "foo" in try f d with Eof -> close d Here, * represents a non-deterministic boolean value. The function readex reads the file pointer x, and then non-deterministically raises an end-of-file (Eof) exception. The main expression (on the third line) first opens file “foo”, calls f to read the file repeatedly, and closes the file upon an end-of-file exception. Suppose, as in the example of Section 1, we wish to verify that the file “foo” is accessed following the protocol in Figure 1. First, we can remove exceptions by representing an exception handler as a special continuation [6]: let readex x h k = read x; (if * then k() else h()) in let rec f x h k = readex x h (fun _ -> f x h k) in let d = open_in "foo" in f d (fun _ -> close d) (fun _ -> ()) Here, we have added to each function two parameters h and k, which represent an exception handler and a (normal) continuation respectively. Let Φ be (E, F true (λr.hcloseitrue) (λr.true)) where E is: Readex x h k =ν hreadi(k true ∧ h true); F x h k =ν Readex x h (λr.F x h k). 9 Here, we have just replaced read/close operations with the modal operators hreadi and hclosei, non-deterministic choice with a logical conjunction, and the unit value ( ) with true. Then, Lfile |= Φ if and only if the program performs only valid accesses to the file (e.g., it does not access the file after a close operation), where Lfile is the LTS shown in Figure 1. The correctness of the reduction can be informally understood by observing that there is a close correspondence between reductions of the program and those of the HFL formula above, and when the program reaches a read command read x, the corresponding formula is of the form hreadi · · ·, meaning that the read operation is valid in the current state; a similar condition holds also for close operations. We will present a general translation and prove its correctness in Section 6. Let us consider another example, which uses integers: let rec f y x k = if y=0 then (close x; k()) else (read x; f (y-1) x k) in let d = open_in "foo" in f n d (fun _ -> ()) Here, n is an integer constant. The function f reads x y times, and then calls the continuation k. Let L′file be the LTS obtained by adding to Lfile a new state q2 and end the transition q1 −→ q2 (which intuitively means that a program is allowed to terminate in the state q1 ), and let Φ′ be (E′ , F n true (λr.henditrue)) where E′ is: F y x k =µ (y = 0 ⇒ hclosei(k true)) ∧ (y , 0 ⇒ hreadi(F (y − 1) x k)). Here, p(ϕ1 , . . . , ϕk ) ⇒ ϕ is an abbreviation of ¬p(ϕ1 , . . . , ϕk ) ∨ ϕ. Then, L′file |= Φ′ if and only if (i) the program performs only valid accesses to the file, (ii) it eventually terminates, and (iii) the file is closed when the program terminates. Notice the use of µ instead of ν above; by using µ, we can express liveness properties. The property L′file |= Φ′ indeed holds for n ≥ 0, but not for n < 0. In fact, F n x k is equivalent to false for n < 0, and hreadin hclosei(k true) for n ≥ 0. 4 Target Language This section sets up, as the target of program verification, a call-by-name7 higherorder functional language extended with events. The language is essentially the same as the one used by Watanabe et al. [47] for discussing fair non-termination. 4.1 Syntax and Typing We assume a finite set Ev of names called events, ranged over by a, and a denumerable set of variables, ranged over by x, y, . . .. Events are used to express temporal properties of programs. We write e x (e t, resp.) for a sequence of variables (terms, resp.), and write |e x| for the length of the sequence. 7 Call-by-value programs can be handled by applying the CPS transformation before applying the reductions to HFL model checking. 10 x1 = A program is a pair (D, t) consisting of a set D of function definitions { f1 e xn = tn } and a term t. The set of terms, ranged over by t, is defined by: t1 , . . . , fn e t ::= ( ) | x | n | t1 op t2 | event a; t | if p(t′1 , . . . , t′k ) then t1 else t2 | t1 t2 | t1 t2 . Here, n and p range over the sets of integers and integer predicates as in HFL formulas. The expression event a; t raises an event a, and then evaluates t. Events are used to encode program properties of interest. For example, an assertion assert(b) can be expressed as if b then ( ) else (event fail; Ω), where fail is an event that expresses an assertion failure and Ω is a non-terminating term. If program termination is of interest, one can insert “event end” to every termination point and check whether an end event occurs. The expression t1 t2 evaluates t1 or t2 in a non-deterministic manner; it can be used to model, e.g., unknown inputs from an environment. We use the meta-variable P for programs. When xn = tn }, we write funs(P) for { f1 , . . . , fn } x1 = t1 , . . . , fn e P = (D, t) with D = { f1 e (i.e., the set of function names defined in P). Using λ-abstractions, we sometimes write f = λe x.t for the function definition f e x = t. We also regard D as a map from function names to terms, and write dom(D) for { f1 , . . . , fn } and D( fi ) for λe xi .ti . Any program (D, t) can be normalized to (D ∪ {main = t}, main) where main is a name for the “main” function. We sometimes write just D for a program (D, main), with the understanding that D contains a definition of main. We restrict the syntax of expressions using a type system. The set of simple types, ranged over by κ, is defined by: κ ::= ⋆ | η → κ η ::= κ | int. The types ⋆, int, and η → κ describe the unit value, integers, and functions from η to κ respectively. Note that int is allowed to occur only in argument positions. We defer typing rules to Appendix A, as they are standard, except that we require that the righthand side of each function definition must have type ⋆; this restriction, as well as the restriction that int occurs only in argument positions, does not lose generality, as those conditions can be ensured by applying CPS transformation. We consider below only well-typed programs. 4.2 Operational Semantics ℓ We define the labeled transition relation t −→D t′ , where ℓ is either ǫ or an event name, as the least relation closed under the rules in Figure 4. We implicitly assume that the program (D, t) is well-typed, and this assumption is maintained throughout reductions by the standard type preservation property (which we omit to prove). In the rules for if-expressions, ~t′i  represents the integer value denoted by t′i ; note that the well-typedness of (D, t) guarantees that t′i must be arithmetic expressions consisting of integers and integer operations; thus, ~t′i  is well defined. We often omit the subscript D when it is clear from the context. 11 |e x| = |e t| fe x=u∈D a ǫ (~t′1 , . . . , ~t′k ) ∈ ~p ǫ fe t −→D [e t/e x]u event a; t −→D t if p(t′1 , . . . , t′k ) then t1 else t2 −→D t1 (~t′1 , . . . , ~t′k ) < ~p i ∈ {1, 2} ǫ ǫ if p(t′1 , . . . , t′k ) then t1 else t2 −→D t2 t1 t2 −→D ti Fig. 4. Labeled Transition Semantics ℓ1 ···ℓk ℓ1 ℓk We write t −→ ∗D t′ if t −→D · · · −→D t′ . Here, ǫ is treated as an empty sequence; ab a ǫ ǫ b thus, for example, we write t −→∗D t′ if t −→D −→D −→D −→D t′ . For a program P = (D, t0), we define the set Traces(P)(⊆ Ev∗ ∪ Evω ) of traces by: ℓi Traces(D, t0 ) = {ℓ0 · · · ℓn−1 ∈ ({ǫ} ∪ Ev)∗ | ∀i ∈ {0, . . . , n − 1}.ti −→D ti+1 } ℓi ∪{ℓ0 ℓ1 · · · ∈ ({ǫ} ∪ Ev)ω | ∀i ∈ ω.ti −→D ti+1 }. Note that since the label ǫ is regarded as an empty sequence, ℓ0 ℓ1 ℓ2 = aa if ℓ0 = ℓ2 = a and ℓ1 = ǫ, and an element of ({ǫ} ∪ Ev)ω is regarded as that of Ev∗ ∪ Evω . We write FinTraces(P) and InfTraces(P) for Traces(P) ∩ Ev∗ and Traces(P) ∩ Evω respectively. The set of full traces FullTraces(D, t0)(⊆ Ev∗ ∪ Evω ) is defined as: ℓi {ℓ0 · · · ℓn−1 ∈ ({ǫ} ∪ Ev)∗ | tn = ( ) ∧ ∀i ∈ {0, . . . , n − 1}.ti −→D ti+1 } ℓi ∪{ℓ0 ℓ1 · · · ∈ ({ǫ} ∪ Ev)ω | ∀i ∈ ω.ti −→D ti+1 }. Example 4. The last example in Section 1 is modeled as Pfile = (D, f ( )), where D = { f x = (event close; ( ))(event read; event read; f x)}. We have: Traces(P) = {readn | n ≥ 0} ∪ {read2n close | n ≥ 0} ∪ {readω } FinTraces(P) = {readn | n ≥ 0} ∪ {read2n close | n ≥ 0} InfTraces(P) = {readω } FullTraces(P) = {read2n close | n ≥ 0} ∪ {readω }. 5 May/Must-Reachability Verification Here we consider the following problems: – May-reachability: “Given a program P and an event a, may P raise a?” – Must-reachability: “Given a program P and an event a, must P raise a?” Since we are interested in a particular event a, we restrict here the event set Ev to a singleton set of the form {a}. Then, the may-reachability is formalized as ? a ∈ Traces(P), whereas the must-reachability is formalized as “does every trace in FullTraces(P) contain a?” We encode both problems into the validity of HFLZ formulas (without any modal operators hai or [a]), or the HFLZ model checking 12 of those formulas against a trivial model (which consists of a single state without any transitions). Since our reductions are sound and complete, the characterizations of their negations –non-reachability and may-non-reachability– can also be obtained immediately. Although these are the simplest classes of properties among those discussed in Sections 5–7, they are already large enough to accommodate many program properties discussed in the literature, including lack of assertion failures/uncaught exceptions [25] (which can be characterized as nonreachability; recall the encoding of assertions in Section 4), termination [29, 31] (characterized as must-reachability), and non-termination [28] (characterized as may-non-reachability). 5.1 May-Reachability As in the examples in Section 3, we translate a program to a formula that says “the program may raise an event a” in a compositional manner. For example, event a; t can be translated to true (since the event will surely be raised immediately), and t1 t2 can be translated to t†1 ∨t†2 where t†i is the result of the translation of ti (since only one of t1 and t2 needs to raise an event). Definition 3. Let P = (D, t) be a program. ΦP,may is the HES (D†may , t†may ), where D†may and t†may are defined by:   † xn =µ tn †may x1 =µ t1 †may ; · · · ; fn e xn = tn } may = f1 e x1 = t1 , . . . , fn e { f1 e ( )†may = false x†may = x n†may = n (t1 op t2 )†may = t1 †may op t2 †may †may ′ ′ = (if p(t1 , . . . , tk ) then t1 else t2 ) (p(t′1 †may , . . . , t′k †may ) ∧ t1 †may ) ∨ (¬p(t′1†may , . . . , t′k †may ) ∧ t2 †may ) (event a; t)†may = true (t1 t2 )†may = t1 †may t2 †may (t1 t2 )†may = t1 †may ∨ t2 †may . Note that, in the definition of D†may , the order of function definitions in D does not matter (i.e., the resulting HES is unique up to the semantic equality), since all the fixpoint variables are bound by µ. Example 5. Consider the program: Ploop = ({loop x = loop x}, loop(event a; ( ))). It is translated to the HES Φloop = (loop x =µ loop x, loop(true)). Since loop ≡ µloop.λx.loop x is equivalent to λx.false, Φloop is equivalent to false. In fact, Ploop never raises an event a (recall that our language is call-by-name). Example 6. Consider the program Psum = (Dsum , main) where Dsum is: main = sum n (λr.assert(r ≥ n)) sum x k = if x = 0 then k 0 else sum (x − 1) (λr.k(x + r)) Here, n is some integer constant, and assert(b) is the macro introduced in Section 4. We have used λ-abstractions for the sake of readability. The function sum 13 is a CPS version of a function that computes the summation of integers from 1 to x. The main function computes the sum r = 1 + · · · + n, and asserts r ≥ n. It is translated to the HES ΦP2 ,may = (Esum , main) where Esum is: main =µ sum n (λr.(r ≥ n ∧ false) ∨ (r < n ∧ true)); sum x k =µ (x = 0 ∧ k 0) ∨ (x , 0 ∧ sum (x − 1) (λr.k(x + r))). Here, n is treated as a constant. Since the shape of the formula does not depend on the value of n, the property “an assertion failure may occur for some n” can be expressed by ∃n.ΦP2 ,may . Thanks to the completeness of the encoding (Theorem 1 below), the lack of assertion failures can be characterized by ∀n.Φ, where Φ is the De Morgan dual of the above HES: main =ν sum n (λr.(r < n ∨ true) ∧ (r ≥ n ∨ false)) sum x k =ν (x , 0 ∨ k 0) ∧ (x = 0 ∨ sum (x − 1) (λr.k(x + r))). ⊔ ⊓ The following theorem states that ΦP,may is a complete characterization of the may-reachability of P. Theorem 1. Let P be a program. Then, a ∈ Traces(P) if and only if L0 |= ΦP,may for L0 = ({s⋆ }, ∅, ∅, s⋆). To prove the theorem, we first show the theorem for recursion-free programs and then lift it to arbitrary programs by using the continuity of functions represented in the fixpoint-free fragment of HFLZ formulas. To show that the theorem holds for recursion-free programs, See Appendix B.1 for a concrete proof. 5.2 Must-Reachability The characterization of must-reachability can be obtained by an easy modification of the characterization of may-reachability: we just need to replace branches with logical conjunction. Definition 4. Let P = (D, t) be a program. ΦP,must is the HES (D†must , t†must ), where D†must and t†must are defined by:   † xn =µ tn †must x1 =µ t1 †must ; · · · ; fn e xn = tn } must = f1 e x1 = t1 , . . . , fn e { f1 e ( )†must = false x†must = x n†must = n (t1 op t2 )†must = t1 †must op t2 †must †must ′ ′ (if p(t1 , . . . , tk ) then t1 else t2 ) = (p(t′1 †must , . . . , t′k †must ) ⇒ t1 †must ) ∧ (¬p(t′1 †must , . . . , t′k †must ) ⇒ t2 †must ) (event a; t)†must = true (t1 t2 )†must = t1 †must t2 †must (t1 t2 )†must = t1 †must ∧ t2 †must . Here, p(ϕ1 , . . . , ϕk ) ⇒ ϕ is a shorthand for ¬p(ϕ1 , . . . , ϕk ) ∨ ϕ. Example 7. Consider Ploop = (D, loop m n) where D is: loop x y = if x ≤ 0 ∨ y ≤ 0 then (event end; ( )) else (loop (x − 1) (y ∗ y))(loop x (y − 1)) 14 Here, the event end is used to signal the termination of the program. The function loop non-deterministically updates the values of x and y until either x or y becomes non-positive. The must-termination of the program is characterized by ΦPloop,must = (E, loop m n) where E is: loop x y =µ (x ≤ 0 ∨ y ≤ 0 ⇒ true) ∧(¬(x ≤ 0 ∨ y ≤ 0) ⇒ (loop (x − 1) (y ∗ y)) ∧ (loop x (y − 1))). We write Musta (P) if every π ∈ FullTraces(P) contains a. The following theorem, which can be proved in a manner similar to Theorem 1, guarantees that ΦP,must is indeed a sound and complete characterization of the must-reachability. Theorem 2. Let P be a program. Then, Musta (P) if and only if L0 |= ΦP,must for L0 = ({s⋆ }, ∅, ∅, s⋆). The proof is given in Appendix B.2. 6 Trace Properties Here we consider the verification problem: “Given a (non-ω) regular language L and a program P, does every finite event sequence of P belong to L? (i.e. ? FinTraces(P) ⊆ L)” and reduce it to an HFLZ model checking problem. The verification of file-accessing programs considered in Section 3 may be considered an instance of the problem.8 Here we assume that the language L is closed under the prefix operation; this does not lose generality because FinTraces(P) is also closed under the prefix operation. We write AL = (Q, Σ, δ, q0 , F) for the minimal, deterministic automaton with no dead states (hence the transition function δ may be partial). Since L is prefix-closed and the automaton is minimal, w ∈ L if and only if δ̂(q0 , w) is defined (where δ̂ is defined by: δ̂(q, ǫ) = q and δ̂(q, aw) = δ̂(δ(q, a), w)). We use the corresponding LTS LL = (Q, Σ, {(q, a, q′) | δ(q, a) = q′ }, q0 ) as the model of the reduced HFLZ model checking problem. Given the LTS LL above, whether an event sequence a1 · · · ak belongs ? to L can be expressed as LL |= ha1 i · · · hak itrue. Whether all the event sequences in {a j,1 · · · a j,k j | j ∈ {1, . . . , n}} belong to L can be expressed as ? V LL |= j∈{1,...,n} ha j,1 i · · · ha j,k j itrue. We can lift these translations for event sequences to the translation from a program (which can be considered a description of a set of event sequences) to an HFLZ formula, as follows. 8 The last example in Section 3 is actually a combination with the must-reachability problem. 15 Definition 5. Let P = (D, t) be a program. ΦP,path is the HES (D†path , t†path ), where D†path and t†path are defined by:   † xn =ν tn †path x1 =ν t1 †path ; · · · ; fn e xn = tn } path = f1 e x1 = t1 , . . . , fn e { f1 e ( )†path = true x†path = x n†path = n (t1 op t2 )†path = t1 †path op t2 †path †path ′ ′ (if p(t1 , . . . , tk ) then t1 else t2 ) = †path ′ †path ′ †path (p(t1 , . . . , tk ) ⇒ t1 ) ∧ (¬p(t′1 †path , . . . , t′k †path ) ⇒ t2 †path ) (event a; t)†path = hait†path (t1 t2 )†path = t1 †path t2 †path (t1 t2 )†path = t1 †path ∧ t2 †path . Example 8. The last program discussed in Section 3 is modeled as P2 = (D2 , f m g), where m is an integer constant and D2 consists of: f y k = if y = 0 then (event close; k ( )) else (event read; f (y − 1) k) g r = event end; ( ) Here, we have modeled accesses to the file, and termination as events. Then, ΦP2 ,path = (EP2 ,path , f m g) where EP2 ,path is:9 f n k =ν (n = 0 ⇒ hclosei(k true)) ∧ (n , 0 ⇒ hreadi( f (n − 1) k)) g r =ν henditrue. Let L be the prefix-closure of read∗ · close · end. Then LL is L′file in Section 3, and FinTraces(P2 )⊆L can be verified by checking LL |=ΦP2 ,path . ⊔ ⊓ Theorem 3. Let P be a program and L be a regular, prefix-closed language. Then, FinTraces(P) ⊆ L if and only if LL |= ΦP,path . As in Section 5, we first prove the theorem for programs in normal form, and then lift it to recursion-free programs by using the preservation of the semantics of HFLZ formulas by reductions, and further to arbitrary programs by using the (co-)continuity of the functions represented by fixpoint-free HFLZ formulas. The proof is given in Appendix C. 7 Linear-Time Temporal Properties This section considers the following problem: “Given a program P and an ωregular word language L, does InfTraces(P)∩L = ∅ hold?” From the viewpoint of program verification, L represents the set of “bad” behaviors. This can be considered an extension of the problems considered in the previous sections.10 9 10 Unlike in Section 3, the variables are bound by ν since we are not concerned with the termination property here. Note that finite traces can be turned into infinite ones by inserting a dummy event for every function call and replacing each occurrence of the unit value ( ) with loop( ) where loop x = event dummy; loop x. 16 b ?>=< qa j 4 89:; b * ?>=< 89:; t qb a a Fig. 5. LTS for (a∗ b)ω The reduction to HFL model checking is more involved than those in the previous sections. To see the difficulty, consider the program P0 : { f = if c then (event a; f ) else (event b; f )},  f , where c is some boolean expression. Let L be the complement of (a∗ b)ω , i.e., the set of infinite sequences that contain only finitely many b’s. Following Section 6 (and noting that InfTraces(P)∩L = ∅ is equivalent to InfTraces(P) ⊆ (a∗ b)ω in this case), one may be tempted to prepare an LTS like the one in Figure 5 (which corresponds to the transition function of a (parity) word automaton accepting (a∗ b)ω ), and translate the program to an HES ΦP0 of the form: f =α (c ⇒ hai f ) ∧ (¬c ⇒ hbi f ),  f , where α is µ or ν. However, such a translation would not work. If c = true, then InfTraces(P0 ) = aω , hence InfTraces(P0 ) ∩ L , ∅; thus, α should be µ for ΦP0 to be unsatisfied. If c = false, however, InfTraces(P0 ) = bω , hence InfTraces(P0 ) ∩ L = ∅; thus, α must be ν for ΦP0 to be satisfied. The example above suggests that we actually need to distinguish between the two occurrences of f in the body of f ’s definition. Note that in the then- and else-clauses respectively, f is called after different events a and b. This difference is important, since we are interested in whether b occurs infinitely often. We thus duplicate f , and replace the program with the following program Pdup : ({ fb = if c then (event a; fa ) else (event b; fb ), fa = if c then (event a; fa ) else (event b; fb )}, fb ). For checking InfTraces(P0 ) ∩ L = ∅, it is now sufficient to check that fb is recursively called infinitely often. We can thus obtain the following HES: (( fb =ν (c ⇒ hai fa ) ∧ (¬c ⇒ hbi fb ); fa =µ (c ⇒ hai fa ) ∧ (¬c ⇒ hbi fb )), fb ). Note that fb and fa are bound by ν and µ respectively, reflecting the fact that b should occur infinitely often, but a need not. If c = true, the formula is equivalent to ν fb .haiµ fa .hai fa , which is false. If c = false, then the formula is equivalent to ν fb .hbi fb , which is satisfied by by the LTS in Figure 5. The general translation is more involved due to the presence of higher-order functions, but, as in the example above, the overall translation consists of two 17 steps. We first replicate functions according to what events may occur between ? two recursive calls, and reduce the problem InfTraces(P) ∩ L = ∅ to a problem of analyzing which functions are recursively called infinitely often, which we call a call-sequence analysis. We can then reduce the call-sequence analysis to HFL model checking in a rather straightforward manner (though the proof of the correctness is non-trivial). The resulting HFL formula actually does not contain modal operators.11 So, as in Section 5, the resulting problem is the validity checking of HFL formulas without modal operators. In the rest of this section, we first introduce the call-sequence analysis problem and its reduction to HFL model checking in Section 7.1. We then show how ? to reduce the temporal verification problem InfTraces(P) ∩ L = ∅ to an instance of the call-sequence analysis problem in Section 7.2. 7.1 Call-sequence analysis We define the call-sequence analysis and reduce it to an HFL model-checking problem. As mentioned above, in the call-sequence analysis, we are interested in analyzing which functions are recursively called infinitely often. Here, we say e ℓ ǫ t, where f e x = tf ∈ D that g is recursively called from f , if f e s −→D [e s/e x]t f −→∗D g e and g “originates from” t f (a more formal definition will be given in Definition 6 below). For example, consider the following program Papp , which is a twisted version of Pdup above. ({app h x = h x, fb x = if x > 0 then (event a; app fa (x − 1)) else (event b; app fb 5), fa x = if x > 0 then (event a; app fa (x − 1)) else (event b; app fb 5)}, fb 5). a ǫ Then fa is “recursively called” from fb in fb 5 −→∗D app fa 4 −→∗D fa 4 (and so is app). We are interested in infinite chains of recursive calls f0 f1 f2 · · ·, and which functions may occur infinitely often in each chain. For instance, the program above has the unique infinite chain ( fb fa5 )ω , in which both fa and fb occur infinitely often. (Besides the infinite chain, the program has finite chains like fb app; note that the chain cannot be extended further, as the body of app does not have any occurrence of recursive functions: app, fa and fb .) We define the notion of “recursive calls” and call-sequences formally below. s) be a proDefinition 6 (recursive call relation, call sequences). Let P = (D, f1 e gram, with D = { fi x̃i = ui }1≤i≤n . We define D♯ := D ∪ { fi♯ x̃ = ui }1≤i≤n where f1♯ , . . . , fn♯ are fresh symbols. (Thus, D♯ has two copies of each function symbol, one of which is marked by ♯.) For the terms e ti and e t j that do not contain marked symbols, we write ti fi e 11 tj D fje e ℓ t′j and (ii) e t j is obtained by erasing if (i) [e ti /e xi ][ f1♯ / f1 , . . . , fn♯ / fn ]ui −→∗D♯ f j♯ e In the example above, we can actually remove hai and hbi, as information about events has been taken into account when f was duplicated. 18 all the marks in e t′j . We write Callseq(P) for the set of (possibly infinite) sequences of function symbols: s { f1 g1 g2 · · · | f1 e t1 D g1 e t2 D g2 e D · · · }. We write InfCallseq(P) for the subset of Callseq(P) consisting of infinite sequences, i.e., Callseq(P) ∩ { f1 , . . . , fn }ω . For example, for Papp above, Callseq(P) is the prefix closure of {( fb fa5 )ω } ∪ {s · app | s is a non-empty finite prefix of ( fb fa5 )ω }, and InfCallseq(P) is the singleton set {( fb fa5 )ω }. Definition 7 (Call-sequence analysis). A priority assignment for a program P is a function Ω : funs(P) → N from the set of function symbols of P to the set N of natural numbers. We write |=csa (P, Ω) if every infinite call-sequence g0 g1 g2 · · · ∈ InfCallseq(P) satisfies the parity condition w.r.t. Ω, i.e., the largest number occurring infinitely often in Ω(g0 )Ω(g1 )Ω(g2 ) . . . is even. Call-sequence analysis is the problem of, given a program P with a priority assignment Ω, deciding whether |=csa (P, Ω) holds. For example, for Papp and the priority assignment Ωapp = {app 7→ 3, fa 7→ 1, fb 7→ 2}, |=csa (Papp , Ωapp ) holds. The call-sequence analysis can naturally be reduced to HFL model checking against the trivial LTS L0 = ({s⋆ }, ∅, ∅, s⋆) (or validity checking). Definition 8. Let P = (D, t) be a program and Ω be a priority assignment for P. The HES Φ(P,Ω),csa is (D†csa , t†csa ), where D†csa and t†csa are defined by:   † xn =αn tn †csa x1 =α1 t1 †csa ; · · · ; fn e xn = tn } csa = f1 e x1 = t1 , . . . , fn e { f1 e ( )†csa = true x†csa = x n†csa = n (t1 op t2 )†csa = t1 †csa op t2 †csa (if p(t′1 , . . . , t′k ) then t1 else t2 )†csa = (p(t′1 †csa , . . . , t′k †csa ) ⇒ t1 †csa ) ∧ (¬p(t′1†csa , . . . , t′k †csa ) ⇒ t2 †csa ) (event a; t)†csa = t†csa (t1 t2 )†csa = t1 †csa ∧ t2 †csa . (t1 t2 )†csa = t1 †csa t2 †csa Here, we assume that Ω( fi ) ≥ Ω( fi+1 ) for each i ∈ {1, . . . , n − 1}, and αi = ν if Ω( fi ) is even and µ otherwise. The following theorem states the soundness and completeness of the reduction. See Appendix E.3 for a proof. Theorem 4. Let P be a program and Ω be a priority assignment for P. Then |=csa (P, Ω) if and only if L0 |= Φ(P,Ω),csa . Example 9. For Papp and Ωapp above, (Papp , Ωapp )†csa = (E, fb 5), where: E is: app h x =µ h x; fb x =ν (x > 0 ⇒ app fa (x − 1)) ∧ (x ≤ 0 ⇒ app fb 5); fa x =µ (x > 0 ⇒ app fa (x − 1)) ∧ (x ≤ 0 ⇒ app fb 5). Note that L0 |= (Papp , Ωapp )†csa holds. 19 7.2 From temporal verification to call-sequence analysis This subsection shows a reduction from the temporal verification problem ? ? InfTraces(P) ∩ L = ∅ to a call-sequence analysis problem |=csa (P′ , Ω). For the sake of simplicity, we assume without loss of generality12 that every program P = (D, t) in this section is non-terminating and every infinite reduction sequence produces infinite events, so that FullTraces(P) = InfTraces(P) holds. We also assume that the ω-regular language L for the temporal verification problem is specified by using a non-deterministic, parity word automaton [11]. We recall the definition of non-deterministic, parity word automata below. Definition 9 (Parity automaton). A non-deterministic parity word automaton (NPW)13 is a quintuple A = (Q, Σ, δ, qI , Ω) where (i) Q is a finite set of states; (ii) Σ is a finite alphabet; (iii) δ, called a transition function, is a total map from Q × Σ to 2Q ; (iv) qI ∈ Q is the initial state; and (v) Ω ∈ Q → N is the priority function. A run of A on an ω-word a0 a1 · · · ∈ Σω is an infinite sequence of states ρ = ρ(0)ρ(1) · · · ∈ Qω such that (i) ρ(0) = qI , and (ii) ρ(i + 1) ∈ δ(ρ(i), ai ) for each i ∈ ω. An ω-word w ∈ Σω is accepted by A if, there exists a run ρ of A on w such that max{Ω(q) | q ∈ Inf(ρ)} is even, where Inf(ρ) is the set of states that occur infinitely often in ρ. We write L(A) for the set of ω-words accepted by A. For technical convenience, we assume below that δ(q, a) , ∅ for every q ∈ Q and a ∈ Σ; this does not lose generality since if δ(q, a) = ∅, we can introduce a new “dead” state qdead (with priority 1) and change δ(q, a) to {qdead }. Given a parity automaton A, we refer to each component of A by QA , ΣA , δA , qI,A and ΩA . Example 10. Consider the automaton Aab = ({qa , qb }, {a, b}, δ, qa, Ω), where δ is as given in Figure 5, Ω(qa ) = 0, and Ω(qb ) = 1. Then, L(Aab ) = (a∗ b)ω = (a∗ b)∗ aω . The goal of this subsection is, given a program P and a parity word automaton A, to construct another program P′ and a priority assignment Ω for P′ , such that InfTraces(P) ∩ L(A) = ∅ if and only if |=csa (P′ , Ω). Note that a necessary and sufficient condition for InfTraces(P) ∩ L(A) = ∅ is that no trace in InfTraces(P) has a run whose priority sequence satisfies the parity condition; in other words, for every sequence in InfTraces(P), and 12 13 As noted at the beginning of this section, every finite trace can be turned into an infinite trace by inserting (fresh) dummy events. Then, InfTraces(P) ∩ L = ∅ holds if and only if InfTraces(P′ )∩L′ = ∅, where P′ is the program obtained from P by inserting dummy events, and L′ is the set of all event sequences obtained by inserting dummy events into a sequence in L. Note that non-deterministic Büchi automata may be viewed as instances of nondeterministic parity word automata, where there are only two priorities 1 and 2, and accepting and non-accepting states have priorities 2 and 1 respectively. We also note that the classes of deterministic parity, non-deteterministic parity, and non-deteterministic Büchi word automata accept the same class of ω-regular languages; here we opt for non-deteterministic parity word automata, because the translations from the others to NPW are trivial but the other directions may blow up the size of automata. 20 for every run for the sequence, the largest priority that occurs in the associated priority sequence is odd. As explained at the beginning of this section, we reduce this condition to a call sequence analysis problem by appropriately duplicating functions in a given program. For example, recall the program P0 :  { f = if c then (event a; f ) else (event b; f )}, f . It is translated to P′0 : ({ fb = if c then (event a; fa ) else (event b; fb ), fa = if c then (event a; fa ) else (event b; fb )}, fb ), where c is some (closed) boolean expression. Since the largest priorities encountered before calling fa and fb (since the last recursive call) respectively are 0 and 1, we assign those priorities plus 1 (to flip odd/even-ness) to fa and fb respectively. Then, the problem of InfTraces(P0 ) ∩ L(A) = ∅ is reduced to |=csa (P′0 , { fa 7→ 1, fb 7→ 2}). Note here that the priorities of fa and fb represent summaries of the priorities (plus one) that occur in the run of the automaton until fa and fb are respectively called since the last recursive call; thus, the largest priority of states that occur infinitely often in the run for an infinite trace is equivalent to the largest priority that occurs infinitely often in the sequence of summaries (Ω( f1 ) − 1)(Ω( f2 ) − 1)(Ω( f3 ) − 1) · · · computed from a corresponding call sequence f1 f2 f3 · · ·. Due to the presence of higher-order functions, the general reduction is more complicated than the example above. First, we need to replicate not only function symbols, but also arguments. For example, consider the following variation P1 of P0 above:  {g k = if c then (event a; k) else (event b; k), f = g f }, f . Here, we have just made the calls to f indirect, by preparing the function g. Obviously, the two calls to k in the body of g must be distinguished from each other, since different priorities are encountered before the calls. Thus, we duplicate the argument k, and obtain the following program P′1 : ({g ka kb = if c then (event a; ka ) else (event b; kb ), fa = g fa fb , fb = g fa fb }, fa ). Then, for the priority assignment Ω = { fa 7→ 1, fb 7→ 2, g 7→ 1}, InfTraces(P1 ) ∩ L(Aab ) = ∅ if and only if |=csa (P′1 , Ω). Secondly, we need to take into account not only the priorities of states visited by A, but also the states themselves. For example, if we have a function definition f h = h(event a; f h), the largest priority encountered before f is recursively called in the body of f depends on the priorities encountered inside h, and also the state of A when h uses the argument event a; f (because the state after the a event depends on the previous state in general). We, therefore, use intersection types (a la Kobayashi and Ong’s intersection types for HORS model checking [23]) to represent summary information on how each function traverses states of the automaton, and replicate 21 each function and its arguments for each type. We thus formalize the translation as an intersection-type-based program transformation; related transformation techniques are found in [8, 12, 13, 22, 42]. Definition 10. Let A = (Q, Σ, δ, qI , Ω) be a non-deterministic parity word automaton. Let q and m range over Q and the set codom(Ω) of priorities respectively. The set TypesA of intersection types, ranged over by θ, is defined by: V θ ::= q | ρ → θ ρ ::= int | 1≤i≤k (θi , mi ) V We assume a certain total order < on TypesA × N, and require that in 1≤i≤k (θi , mi ), (θi , mi ) < (θ j , m j ) holds for each i < j. V We often write (θ1 , m1 ) ∧ · · · ∧ (θk , mk ) for 1≤i≤k (θi , mi ), and ⊤ when k = 0. Intuitively, the type q describes expressions of simple type ⋆, which may be evaluated when the automaton A is in the state q (here, we have in mind an execution of the product of a program and the automaton, where the V latter takes events produced by the program and changes its states). The type ( 1≤i≤k (θi , mi )) → θ describes functions that take an argument, use it according to types θ1 , . . . , θk , and return a value of type θ. Furthermore, the part mi describes that the argument may be used as a value of type θi only when the largest priority visited since the function is called is mi . For example, given the automaton in Example 10, the function λx.(event a; x) may have types (qa , 0) → qa and (qa , 0) → qb , because the body may be executed from state qa or qb (thus, the return type may be any of them), but x is used only when the automaton is in state qa and the largest priority visited is 1. In contrast, λx.(event b; x) have types (qb , 1) → qa and (qb , 1) → qb . Using the intersection types above, we shall define a type-based transformation relation of the form Γ ⊢A t : θ ⇒ t′ , where t and t′ are the source and target terms of the transformation, and Γ, called an intersection type environment, is a finite set of type bindings of the form x : int or x : (θ, m, m′). We allow multiple type bindings for a variable x except for x : int (i.e. if x : int ∈ Γ, then this must be the unique type binding for x in Γ). The binding x : (θ, m, m′) means that x should be used as a value of type θ when the largest priority visited is m; m′ is auxiliary information used to record the largest priority encountered so far. The transformation relation Γ ⊢A t : θ ⇒ t′ is inductively defined by the rules in Figure 6. (For technical convenience, we have extended terms with λ-abstractions; they may occur only at top-level function definitions.) In the figure, [k] denotes the set {i ∈ N | 1 ≤ i ≤ k}. The operation Γ ↑ m used in the figure is defined by: Γ ↑ m = {x : int | x : int ∈ Γ} ∪ {x : (θ, m1 , max(m2 , m)) | x : (θ, m1 , m2 ) ∈ Γ} The operation is applied when the priority m is encountered, in which case the largest priority encountered is updated accordingly. The key rules are IT-Var, IT-Event, IT-App, and IT-Abs. In IT-Var, the variable x is replicated for each type; in the target of the translation, xθ,m and xθ′ ,m′ are treated as different variables if 22 (IT-Unit) Γ ⊢A ( ) : q ⇒ ( ) Γ, x : int ⊢A x : int ⇒ xint (IT-VarInt) Γ, x : (θ, m, m) ⊢A x : θ ⇒ xθ,m (IT-Var) (IT-Int) Γ ⊢A n : int ⇒ n Γ ⊢A t1 : int ⇒ t′1 Γ ⊢A t2 : int ⇒ t′2 Γ ⊢A t1 op t2 : int ⇒ t′1 op t′2 (IT-Op) Γ ⊢A ti : int ⇒ t′i (for each i ∈ [k]) Γ ⊢A tk+1 : q ⇒ t′k+1 Γ ⊢A tk+2 : q ⇒ t′k+2 e e t = t1 , . . . , tk t′ = t′1 , . . . , t′k Γ ⊢A t1 : q ⇒ t′1 Γ ⊢A t2 : q ⇒ t′2 Γ ⊢A t1 t2 : q ⇒ t′1 t′2 (IT-AppInt) Γ ⊢A t1 t2 : θ ⇒ t′1 t′2 V Γ ⊢A t1 : 1≤i≤k (θi , mi ) → θ ⇒ t′1 Γ ↑ mi ⊢A t2 : θi ⇒ t′2,i (for each i ∈ [k]) Γ ⊢A t1 t2 : θ ⇒ t′1 t′2,1 . . . t′2,k Γ, x : int ⊢A t : θ ⇒ t′ (IT-If) Γ ⊢A if p(e t) then tk+1 else tk+2 : q ⇒ if p(e t′ ) then t′k+1 else t′k+2 δA (q, a) = {q1 , . . . , qk } Γ ↑ ΩA (qi ) ⊢A t : qi ⇒ t′i (for each i ∈ [k]) (IT-NonDet) Γ ⊢A t1 : int → θ ⇒ t′1 Γ ⊢A t2 : int ⇒ t′2 (IT-App) x < dom(Γ) Γ ⊢A λx.t : int → θ ⇒ λxint .t′ (IT-AbsInt) Γ ∪ {x : (θi , mi , 0) | i ∈ [k]} ⊢A t : θ′ ⇒ t′ x < dom(Γ) V Γ ⊢A λx.t : 1≤i≤k (θi , mi ) → θ′ ⇒ λxθ1 ,m1 . . . xθk ,mk .t′ (IT-Abs) Γ ⊢A (event a; t) : q ⇒ (event a; t′1  · · · t′k ) (IT-Event) Fig. 6. Type-based Transformation Rules for Terms (θ, m) , (θ′ , m′ ). The rule IT-Event reflects the state change caused by the event a to the type and the type environment. Since the state change may be nondeterministic, we transform t for each of the next states q1 , . . . , qn , and combine the resulting terms with non-deterministic choice. The rule IT-App and IT-Abs replicates function arguments for each type. In addition, in IT-App, the operation Γ ↑ mi reflects the fact that t2 is used as a value of type θi after the priority mi is encountered. The other rules just transform terms in a compositional manner. If target terms are ignored, the entire rules are close to those of Kobayashi and Ong’s type system for HORS model checking [23]. We now define the transformation for programs. A top-level type environment Ξ is a finite set of type bindings of the form x : (θ, m). Like intersection type environments, Ξ may have more than one binding for each variable. We write Ξ ⊢A t : θ to mean {x : (θ, m, 0) | x : (θ, m) ∈ Ξ} ⊢A t : θ. For a set D of function definitions, we write Ξ ⊢A D ⇒ D′ if dom(D′ ) = { fθ,m | f : (θ, m) ∈ Ξ } and Ξ ⊢A D( f ) : θ ⇒ D′ ( fθ,m ) for every f : (θ, m) ∈ Ξ. For a program P = (D, t), we write Ξ ⊢A P ⇒ (P′ , Ω′ ) if P′ = (D′ , t′ ), Ξ ⊢A D ⇒ D′ and Ξ ⊢A t : qI ⇒ t′ , with Ω′ ( fθ,m ) = m + 1 for each fθ,m ∈ dom(D′ ). We just write ⊢A P ⇒ (P′ , Ω′ ) if Ξ ⊢A P ⇒ (P′ , Ω′ ) holds for some Ξ. 23 Example 11. Consider the automaton Aab in Example 10, and the program P2 = (D2 , f 5) where D2 consists of the following function definitions: g k = (event a; k)(event b; k), f x = if x > 0 then g ( f (x − 1)) else (event b; f 5). Let Ξ be: g : ((qa , 0) ∧ (qb , 1) → qa , 0), g : ((qa , 0) ∧ (qb , 1) → qb , 0), f : (int → qa , 0), f : (int → qb , 1) Then, Ξ ⊢A P1 ⇒ ((D′2 , fint→qa ,0 5), Ω) where: D′2 = {g(qa ,0)∧(qb ,1)→qa ,0 kqa ,0 kqb ,1 = t g , g(qa ,0)∧(qb ,1)→qb ,0 kqa ,0 kqb ,1 = t g , fint→qa ,0 xint = t f,qa , fint→qb ,1 xint = t f,qb } t g = (event a; kqa ,0 )(event b; kqb ,1 ), t f,q = if xint > 0 then g(qa ,0)∧(qb ,1)→q,0 ( fint→qa ,0 (xint − 1)) ( fint→qb ,1 (xint − 1)) else (event b; fint→qb ,1 5), (for each q ∈ {qa , qb }) Ω = {g(qa ,0)∧(qb ,1)→qa ,0 7→ 1, g(qa ,0)∧(qb ,1)→qb ,0 7→ 1, fint→qa ,0 7→ 1, fint→qb ,1 7→ 2}. Appendix F shows how t g and t f are derived. Notice that f , g, and the arguments of g have been duplicated. Furthermore, whenever fθ,m is called, the largest priority that has been encountered since the last recursive call is m. For example, in the then-clause of fint→qa ,0 , fint→qb ,1 (x−1) may be called through g(qa ,0)∧(qb ,1)→qa ,0 . Since g(qa ,0)∧(qb ,1)→qa ,0 uses the second argument only after an event b, the largest priority encountered is 1. This property is important for the correctness of our reduction. The following theorem claims the soundness and completeness of our reduction. See Appendix E for a proof. Theorem 5. Let P be a program and A be a parity automaton. Suppose that Ξ ⊢A P ⇒ (P′ , Ω). Then InfTraces(P) ∩ L(A) = ∅ if and only if |=csa (P′ , Ω). Furthermore, one can effectively find an appropriate transformation. Theorem 6. For every P and A, one can effectively construct Ξ, P′ and Ω such that Ξ ⊢A P ⇒ (P′ , Ω). See Appendix E.5 for a proof sketch. A proof of the above theorem is given in Appendix E.5. The proof also implies that the reduction from temporal property verification to call-sequence analysis can be performed in polynomial time. Combined with the reduction from call-sequence analysis to HFL model checking, we have thus obtained a polynomial-time reduction from the temporal ? verification problem InfTraces(P) ⊆ L(A) to HFL model checking. 24 8 Related Work As mentioned in Section 1, our reduction from program verification problems to HFL model checking problems has been partially inspired by the translation of Kobayashi et al. [21] from HORS model checking to HFL model checking. As in their translation (and unlike in previous applications of HFL model checking [30, 46]), our translation switches the roles of properties and models (or programs) to be verified. Although a combination of their translation with Kobayashi’s reduction from program verification to HORS model checking [19, 20] yields an (indirect) translation from finite-data programs to pure HFL model checking problems, the combination does not work for infinite-data programs. In contrast, our translation is sound and complete even for infinite-data programs. Among the translations in Sections 5–7, the translation in Section 7.2 shares some similarity to their translation, in that functions and their arguments are replicated for each priority. The actual translations are however quite different; ours is type-directed and optimized for a given automaton, whereas their translation is not. This difference comes from the difference of the goals: the goal of [21] was to clarify the relationship between HORS and HFL, hence their translation was designed to be independent of an automaton. The proof of the correctness of our translation in Section 7 is much more involved (cf. Appendix D and E), due to the need for dealing with integers. Whilst the proof of [21] could reuse the type-based characterization of HORS model checking [23], we had to generalize arguments in both [23] and [21] to work on infinite-data programs. Lange et al. [30] have shown that various process equivalence checking problems (such as bisimulation and trace equivalence) can be reduced to (pure) HFL model checking problems. The idea of their reduction is quite different from ours. They reduce processes to LTSs, whereas we reduce programs to HFL formulas. Major approaches to automated or semi-automated higher-order program verification have been HORS model checking [19, 20, 25, 29, 33, 35, 47], (refinement) type systems [15, 26, 36, 39, 40, 43, 45, 50], Horn clause solving [2, 7], and their combinations. As already discussed in Section 1, compared with the HORS model checking approach, our new approach provides more uniform, streamlined methods. Whilst the HORS model checking approach is for fully automated verification, our approach enables various degrees of automation: after verification problems are automatically translated to HFLZ formulas, one can prove them (i) interactively using a proof assistant like Coq (see Appendix G), (ii) semi-automatically, by letting users provide hints for induction/co-induction and discharging the rest of proof obligations by (some extension of) an SMT solver, or (iii) fully automatically by recasting the techniques used in the HORSbased approach; for example, to deal with the ν-only fragment of HFLZ , we can reuse the technique of predicate abstraction [25]. For a more technical comparison between the HORS-based approach and our HFL-based approach, see Appendix H. 25 As for type-based approaches [15, 26, 36, 39, 40, 43, 45, 50], most of the refinement type systems are (i) restricted to safety properties, and/or (ii) incomplete. A notable exception is the recent work of Unno et al. [44], which provides a relatively complete type system for the classes of properties discussed in Section 5. Our approach deals with a wider class of properties (cf. Sections 6 and 7). Their “relative completeness” property relies on Godel coding of functions, which cannot be exploited in practice. The reductions from program verification to Horn clause solving have recently been advocated [2–4] or used [36, 43] (via refinement type inference problems) by a number of researchers. Since Horn clauses can be expressed in a fragment of HFL without modal operators, fixpoint alternations (between ν and µ), and higher-order predicates, our reductions to HFL model checking may be viewed as extensions of those approaches. Higher-order predicates and fixpoints over them allowed us to provide sound and complete characterizations of properties of higher-order programs for a wider class of properties. Bjørner et al. [4] proposed an alternative approach to obtaining a complete characterization of safety properties, which defunctionalizes higher-order programs by using algebraic data types and then reduces the problems to (first-order) Horn clauses. A disadvantage of that approach is that control flow information of higher-order programs is also encoded into algebraic data types; hence even for finite-data higher-order programs, the Horn clauses obtained by the reduction belong to an undecidable fragment. In contrast, our reductions yield pure HFL model checking problems for finite-data programs. Burn et al. [7] have recently advocated the use of higher-order (constrained) Horn clauses for verification of safety properties (i.e., which correspond to the negation of may-reachability properties discussed in Section 5.1 of the present paper) of higher-order programs. They interpret recursion using the least fixpoint semantics, so their higher-order Horn clauses roughly corresponds to a fragment of the HFLZ without modal operators and fixpoint alternations. They have not shown a general, concrete reduction from safety property verification to higher-order Horn clause solving. The characterization of the reachability problems in Section 5 in terms of formulas without modal operators is a reminiscent of predicate transformers [9, 14] used for computing the weakest preconditions of imperative programs. In particular, [5] and [14] respectively used least fixpoints to express weakest preconditions for while-loops and recursions. 9 Conclusion We have shown that various verification problems for higher-order functional programs can be naturally reduced to (extended) HFL model checking problems. In all the reductions, a program is mapped to an HFL formula expressing the property that the behavior of the program is correct. For developing verification tools for higher-order functional programs, our reductions allow us to focus on the development of (automated or semi-automated) HFLZ model checking tools (or, even more simply, theorem provers for HFLZ without modal operators, 26 as the reductions of Section 5 and 7 yield HFL formulas without modal operators). To this end, we have developed a prototype model checker for pure HFL (without integers), which will be reported in a separate paper. Work is under way to develop HFLZ model checkers by recasting the techniques [25, 28, 29, 47] developed for the HORS-based approach, which, together with the reductions presented in this paper, would yield fully automated verification tools. We have also started building a Coq library for interactively proving HFLZ formulas, as briefly discussed in Appendix G. As a final remark, although one may fear that our reductions may map program verification problems to “harder” problems due to the expressive power of HFLZ , it is actually not the case at least for the classes of problems in Section 5 and 6, which use the only alternation-free fragment of HFLZ . The model checking problems for µ-only or ν-only HFLZ are semi-decidable and co-semi-decidable respectively, like the source verification problems of may/must-reachability and their negations of closed programs. Acknowledgment We would like to thank anonymous referees for useful comments. This work was supported by JSPS KAKENHI Grant Number JP15H05706 and JP16K16004. References 1. Axelsson, R., Lange, M., Somla, R.: The complexity of model checking higher-order fixpoint logic. Logical Methods in Computer Science 3(2) (2007) 2. Bjørner, N., Gurfinkel, A., McMillan, K.L., Rybalchenko, A.: Horn clause solvers for program verification. In: Fields of Logic and Computation II - Essays Dedicated to Yuri Gurevich on the Occasion of His 75th Birthday. LNCS, vol. 9300, pp. 24–51. Springer (2015) 3. Bjørner, N., McMillan, K.L., Rybalchenko, A.: Program verification as satisfiability modulo theories. In: SMT 2012. EPiC Series in Computing, vol. 20, pp. 3–11. EasyChair (2012) 4. Bjørner, N., McMillan, K.L., Rybalchenko, A.: Higher-order program verification as satisfiability modulo theories with algebraic data-types. CoRR abs/1306.5264 (2013) 5. Blass, A., Gurevich, Y.: Existential fixed-point logic. In: Computation Theory and Logic, In Memory of Dieter Rödding. LNCS, vol. 270, pp. 20–36. Springer (1987) 6. Blume, M., Acar, U.A., Chae, W.: Exception handlers as extensible cases. In: Proceedings of APLAS 2008. LNCS, vol. 5356, pp. 273–289. Springer (2008) 7. Burn, T.C., Ong, C.L., Ramsay, S.J.: Higher-order constrained horn clauses for verification. PACMPL 2(POPL), 11:1–11:28 (2018) 8. Carayol, A., Serre, O.: Collapsible pushdown automata and labeled recursion schemes: Equivalence, safety and effective selection. In: LICS 2012. pp. 165–174. IEEE (2012) 9. Dijkstra, E.W.: Guarded commands, nondeterminacy and formal derivation of programs. Commun. ACM 18(8), 453–457 (1975) 10. Filliâtre, J.C., Paskevich, A.: Why3 — where programs meet provers. In: Felleisen, M., Gardner, P. (eds.) Proceedings of ESOP 2013. LNCS, vol. 7792, pp. 125–128. Springer (2013) 27 11. Grädel, E., Thomas, W., Wilke, T. (eds.): Automata, Logics, and Infinite Games: A Guide to Current Research, LNCS, vol. 2500. Springer (2002) 12. Grellois, C., Melliès, P.: Relational semantics of linear logic and higher-order model checking. In: Proceedings of CSL 2015. LIPIcs, vol. 41, pp. 260–276 (2015) 13. Haddad, A.: Model checking and functional program transformations. In: Proceedings of FSTTCS 2013. LIPIcs, vol. 24, pp. 115–126 (2013) 14. Hesselink, W.H.: Predicate-transformer semantics of general recursion. Acta Inf. 26(4), 309–332 (1989) 15. Hofmann, M., Chen, W.: Abstract interpretation from Büchi automata. In: Proceedings of CSL-LICS ’14. pp. 51:1–51:10. ACM (2014) 16. Igarashi, A., Kobayashi, N.: Resource usage analysis. ACM Trans. Prog. Lang. Syst. 27(2), 264–313 (2005) 17. Jurdzinski, M.: Small progress measures for solving parity games. In: Proceeding of STACS 2000. LNCS, vol. 1770, pp. 290–301. Springer (2000) 18. Knapik, T., Niwinski, D., Urzyczyn, P.: Higher-order pushdown trees are easy. In: FoSSaCS 2002. LNCS, vol. 2303, pp. 205–222. Springer (2002) 19. Kobayashi, N.: Types and higher-order recursion schemes for verification of higherorder programs. In: Proceedings of POPL. pp. 416–428. ACM Press (2009) 20. Kobayashi, N.: Model checking higher-order programs. Journal of the ACM 60(3) (2013) 21. Kobayashi, N., Lozes, É., Bruse, F.: On the relationship between higher-order recursion schemes and higher-order fixpoint logic. In: Proceedings of POPL 2017. pp. 246–259 (2017) 22. Kobayashi, N., Matsuda, K., Shinohara, A., Yaguchi, K.: Functional programs as compressed data. Higher-Order and Symbolic Computation (2013) 23. Kobayashi, N., Ong, C.H.L.: A type system equivalent to the modal mu-calculus model checking of higher-order recursion schemes. In: Proceedings of LICS 2009. pp. 179–188 (2009) 24. Kobayashi, N., Ong, C.H.L.: A type system equivalent to the modal mu-calculus model checking of higher-order recursion schemes. http://www-kb.is.s.u-tokyo. ac.jp/˜koba/tmp/lics09-full.pdf. A longer version of [23] (2012) 25. Kobayashi, N., Sato, R., Unno, H.: Predicate abstraction and CEGAR for higher-order model checking. In: Proc. of PLDI. pp. 222–233. ACM Press (2011) 26. Koskinen, E., Terauchi, T.: Local temporal reasoning. In: Proceedings of CSL-LICS ’14. pp. 59:1–59:10. ACM (2014) 27. Kozen, D.: Results on the propositional µ-calculus. Theor. Comput. Sci. 27, 333–354 (1983) 28. Kuwahara, T., Sato, R., Unno, H., Kobayashi, N.: Predicate abstraction and CEGAR for disproving termination of higher-order functional programs. In: Proceedings of CAV 2015. LNCS, vol. 9207, pp. 287–303. Springer (2015) 29. Kuwahara, T., Terauchi, T., Unno, H., Kobayashi, N.: Automatic termination verification for higher-order functional programs. In: Proceedings of ESOP 2014. Lecture Notes in Computer Science, vol. 8410, pp. 392–411. Springer (2014) 30. Lange, M., Lozes, É., Guzmán, M.V.: Model-checking process equivalences. Theor. Comput. Sci. 560, 326–347 (2014) 31. Ledesma-Garza, R., Rybalchenko, A.: Binary reachability analysis of higher order functional programs. In: SAS 2012. LNCS, vol. 7460, pp. 388–404. Springer (2012) 32. Lozes, É.: A type-directed negation elimination. In: Proceedings FICS 2015. EPTCS, vol. 191, pp. 132–142 (2015) 33. Murase, A., Terauchi, T., Kobayashi, N., Sato, R., Unno, H.: Temporal verification of higher-order functional programs. In: Proceedings of POPL 2016. pp. 57–68 (2016) 28 34. Ong, C.H.L.: On model-checking trees generated by higher-order recursion schemes. In: LICS 2006. pp. 81–90. IEEE Computer Society Press (2006) 35. Ong, C.H.L., Ramsay, S.: Verifying higher-order programs with pattern-matching algebraic data types. In: Proceedings of POPL. pp. 587–598. ACM Press (2011) 36. Rondon, P.M., Kawaguchi, M., Jhala, R.: Liquid types. In: PLDI 2008. pp. 159–169 (2008) 37. Salvati, S., Walukiewicz, I.: Krivine machines and higher-order schemes. Information and Computation 239(Supplement C), 340 – 355 (2014), http://www. sciencedirect.com/science/article/pii/S0890540114000984 38. Sangiorgi, D.: Introduction to Bisimulation and Coinduction. Cambridge University Press (2012) 39. Skalka, C., Smith, S.F., Horn, D.V.: Types and trace effects of higher order programs. J. Funct. Program. 18(2), 179–249 (2008) 40. Terauchi, T.: Dependent types from counterexamples. In: Proceedings of POPL. pp. 119–130. ACM (2010) 41. Tobita, Y., Tsukada, T., Kobayashi, N.: Exact flow analysis by higher-order model checking. In: Proceedings of FLOPS 2012. LNCS, vol. 7294, pp. 275–289. Springer (2012) 42. Tsukada, T., Ong, C.L.: Compositional higher-order model checking via ω-regular games over böhm trees. In: Proceedings of CSL-LICS ’14. pp. 78:1–78:10. ACM (2014) 43. Unno, H., Kobayashi, N.: Dependent type inference with interpolants. In: PPDP 2009. pp. 277–288. ACM (2009) 44. Unno, H., Satake, Y., Terauchi, T.: Relatively complete refinement type system for verification of higher-order non-deterministic programs. PACMPL 2(POPL), 12:1– 12:29 (2018) 45. Unno, H., Terauchi, T., Kobayashi, N.: Automating relatively complete verification of higher-order functional programs. In: POPL 2013. pp. 75–86. ACM (2013) 46. Viswanathan, M., Viswanathan, R.: A higher order modal fixed point logic. In: CONCUR. Lecture Notes in Computer Science, vol. 3170, pp. 512–528. Springer (2004) 47. Watanabe, K., Sato, R., Tsukada, T., Kobayashi, N.: Automatically disproving fair termination of higher-order functional programs. In: Proceedings of ICFP 2016. pp. 243–255. ACM (2016) 48. Winskel, G.: The Formal Semantics of Programming Languages: An Introduction. The MIT Press (1993) 49. Winskel, G.: Prime algebraicity. Theor. Comput. Sci. 410(41), 4160–4168 (2009) 50. Zhu, H., Nori, A.V., Jagannathan, S.: Learning refinement types. In: Proceedings of ICFP 2015. pp. 400–411. ACM (2015) 29 Appendix A Typing Rules for Programs The type judgments for expressions and programs are of the form K ⊢ t : η and K ⊢ P, where K is a finite map from variables to types. The typing rules are shown in Figure 7. We write ⊢ P if K ⊢ P for some K . K ⊢ () : ⋆ (LT-Unit) K, x : η ⊢ x : η K ⊢ n : int K ⊢ t1 : int (LT-Var) arity(p) = k K ⊢ t′i : int for each i ∈ {1, . . . , k} K ⊢ t j : ⋆ for each j ∈ {1, 2} K ⊢ if p(t′1 , . . . , t′k ) then t1 else t2 : ⋆ (LT-Int) K ⊢ t1 : η → κ K ⊢ t2 : η K ⊢ t1 t2 : κ K ⊢ t2 : int K ⊢ t1 : ⋆ K ⊢ t1 op t2 : int (LT-If) (LT-App) K ⊢ t2 : ⋆ K ⊢ t1 t2 : ⋆ (LT-Op) (LT-NonDet) K ⊢t:⋆ K ⊢ event a; t : ⋆ (LT-Ev) K = f1 : κ1 , . . . , fn : κn K ⊢t:⋆ K,e xi : e η i ⊢ ti : ⋆ κi = e ηi → ⋆ for each i ∈ {1, . . . , n} xn = tn }, t) x1 = t1 , . . . , fn e K ⊢ ({ f1 e (LT-Prog) Fig. 7. Typing Rules for Expressions and Programs B Proofs for Section 5 B.1 Proofs for Section 5.1 To prove the theorem, we define the reduction relation t −→D t′ as given in Figure 8. It differs from the labeled transition semantics in that  and event a; · · · are not eliminated; this semantics is more convenient for establishing the relationship between a program and a corresponding HFLZ formula. It should be clear that a ∈ Traces(D, t) if and only if t −→∗D E[event a; t′ ] for some t′ . We shall first prove the theorem for recursion-free programs. Here, a program P = (D, t) is recursion-free if the transitive closure of the relation {( fi , f j ) ∈ dom(D) × dom(D) | f j occurs in D( fi )} is irreflexive. To this end, we prepare a few lemmas. The following lemma says that the semantics of HFLZ formulas is preserved by reductions of the corresponding programs. 30 E(evaluation contexts) ::= [ ] | Et | tE | event a; E fe x=u∈D |e x| = |e t| E[ f e t] −→D E[[e t/e x]u] (R-Fun) E[if (~t′1 , . . . , ~t′k ) < ~p ′ p(t1 , . . . , t′k ) then t1 else t2 ] −→D E[t2 ] (R-IfF) (~t′1 , . . . , ~t′k ) ∈ ~p E[if p(t′1 , . . . , t′k ) then t1 else t2 ] −→D E[t1 ] (R-IfT) Fig. 8. Reduction Semantics Lemma 1. Let (D, t) be a program and L be an LTS. If t −→D t′ , then ~(D, t)†may L = ~(D, t′ )†may L . xn = tn }, and (F1 , . . . , Fn ) be the least fixpoint of x1 = t1 , . . . , fn e Proof. Let D = { f1 e λ(X1 , . . . , Xn ).(~λe x1 .t1 †may ({ f1 7→ X1 , . . . , fn 7→ Xn }), . . . , ~λe xn .tn †may ({ f1 7→ X1 , . . . , fn 7→ Xn })). By the Bekić property, ~(D, t)†may  = ~t†may ({ f1 7→ F1 , . . . , fn 7→ Fn }). Thus, it suffices to show that t −→D t′ implies ~t†may (ρ) = ~t′ †may (ρ) for ρ = { f1 7→ F1 , . . . , fn 7→ Fn }. We show it by case analysis on the rule used for deriving t −→D t′ . s] and t′ = E[[e s/e xi ]ti ]. Since (F1 , . . . , Fn ) is a – Case R-Fun: In this case, t = E[ fi e fixpoint, we have: s(ρ) = Fi (~e s(ρ)) ~ fi e = ~λe xi .ti (ρ)(~e s(ρ)) = ~[e s/e xi ]ti (ρ) Thus, we have ~t†may (ρ) = ~t′ †may (ρ) as required. – Case R-IfT: In this case, t = E[if p(s′1 , . . . , s′k ) then s1 else s2 ] and t′ = E[s1 ] with (~s′1 , . . . , ~s′k ) ∈ ~p. Thus, t†may = E†may [(p(s′1 , . . . , s′k ) ∧ s1 †may ) ∨ (¬p(s′1 , . . . , s′k ) ∧ s2 †may )]. Since (~s′1 , . . . , ~s′k ) ∈ ~p, (~s′1 , . . . , ~s′k ) < ~¬p. Thus, ~t†may (ρ) = ~E†may [(true ∧ s1 †may ) ∨ (false ∧ s2 †may )](ρ) = ~E†may [s1 †may ](ρ). We have thus ~t†may (ρ) = ~t′ †may (ρ) as required. – Case R-IfF: Similar to the above case. ⊔ ⊓ The following lemma says that Theorem 1 holds for programs in normal form. Lemma 2. Let (D, t) be a program and t −→ 6 D . Then, L0 |= (D, t)†may if and only if ′ t = E[event a; t ] for some evaluation context E and t′ . Proof. The proof proceeds by induction on the structure of t. By the condition t −→ 6 D and the (implicit) assumption that ⊢ (D, t), t is generated by the following grammar: t ::= ( ) | event a; t′ | t1 t2 . 31 – Case t = ( ): The result follows immediately, as t is not of the form E[event a; t′ ], and t†may = false. – Case t = event a; t′ : The result follows immediately, as t is of the form E[event a; t′ ], and t†may = true. – Case t1 t2 : Because (t1 t2 )†may = t1 †may ∨ t2 †may , L0 |= (D, t)†may if and only if L0 |= (D, ti )†may for some i ∈ {1, 2}. By the induction hypothesis, the latter is equivalent to the property that ti is of the form E[event a; t′ ] for some i ∈ {1, 2}, which is equivalent to the property that t is of the form E′ [event a; t′ ]. ⊔ ⊓ The following lemma says that Theorem 1 holds for recursion-free programs; this is an immediate corollary of Lemmas 1 and 2, and the strong normalization property of the simply-typed λ-calculus. Lemma 3. Let P be a recursion-free program. Then, a ∈ Traces(P) if and only if L0 |= ΦP,may for L0 = ({s⋆ }, ∅, ∅, s⋆). Proof. Since P = (D, t0) is recursion-free, there exists a finite, normalizing reduction sequence t0 −→∗D t −→ 6 D . We show the required property by induction on the length n of this reduction sequence. – Case n = 0: Since t0 −→ 6 D , a ∈ Traces(P) if and only if t0 = E[event a; t] for some E and t. Thus, the result follows immediately from Lemma 2. – Case n > 0: In this case, t0 −→D t1 −→∗D t. By the induction hypothesis, a ∈ Traces(D, t1 ) if and only if L0 |= Φ(D,t1 ),may . Thus, by the definition of the reduction semantics and Lemma 1, a ∈ Traces(D, t0 ) if and only if a ∈ Traces(D, t1 ), if and only if L0 |= Φ(D,t1 ),may , if and only if L0 |= Φ(D,t0 ),may . ⊔ ⊓ To prove Theorem 1 for arbitrary programs, we use the fact that the semantics † of P†may may be approximated by P(i) may , where P(i) is the recursion-free program obtained by unfolding recursion functions i times (a more formal definition will be given later). To guarantee the correctness of this finite approximation, we need to introduce a slightly non-standard notion of (ω-)continuous functions below. Definition 11. For an LTS L = (U, A, −→, sinit) and a type σ, the set of continuous elements ContL,σ ⊆ DL,σ and the equivalence relation =cont,L,σ ⊆ ContL,σ × ContL,σ are defined by induction on σ as follows. ContL,• = DL,• ContL,int = DL,int ContL,σ→τ = { f ∈ DL,σ→τ | ∀x1 , x2 ∈ ContL,σ .(x1 =cont,L,σ x2 ⇒ f (x1 ) =cont,L,τ f (x2 )) F F ∧∀{yi }i∈ω ∈ ContL,σ (↑ω) . f ( i∈ω yi ) =cont,L,τ i∈ω f (yi )}. =cont,L,• = {(x, x) | x ∈ ContL,• } =cont,L,int = {(x, x) | x ∈ ContL,int } =cont,L,σ→τ = {( f1 , f2 ) | f1 , f2 ∈ ContL,σ→τ ∧∀x1 , x2 ∈ ContL,σ .(x1 =cont,L,σ x2 ⇒ f1 (x1 ) =cont,L,τ f2 (x2 )}. Here, ContL,σ (↑ω) denotes the set of increasing infinite sequences a0 ⊑ a1 ⊑ a2 ⊑ · · · consisting of elements of ContL,σ . We just write =cont for =cont,L,σ when L and σ are clear from the context. 32 Remark 2. Note that we require that a continuous function returns a continuous element only if its argument is. To see the need for this requirement, consider an LTS with a singleton state set {s}, and the function: f = λx : ((int → •) → •).λy : (int → •).x y. One may expect that f is continuous in the usual sense (i.e., f preserves the limit), but for the function g defined by ( {s} if s ∈ p(n) for every n ≥ 0 , g(p) = ∅ otherwise f (g) is not continuous. In fact, let pi be {(n, {s}) | 0 ≤ n ≤ i} ∪ {(n, ∅) | n < 0 ∨ n > i}. Then f (g)(pi ) = ∅ for every i but f (g)(⊔i∈ω pi ) = {s}. The (non-continuous) function g above can be expressed by (νX.λn.λp.(p(n) ∧ (X (n + 1) p)))0. ⊔ ⊓ (↑ω) and xi =cont,L,σ yi for each i ∈ ω, then Lemma 4. If {xF i }i∈ω , {yi }i∈ω ∈ ContL,σ F y . x = i i cont,L,σ i∈ω i∈ω Proof. The proof proceeds by induction on σ. The base case, where σ = • or σ = int, is trivial, as ContL,σ = DL,σ and =cont,L,σ is the identity relation. F Let us consider the induction step, where σ = σ1 → τ. We first check that i∈ω xi ∈ ContL,σ . To this end, suppose z1 =cont,L,σ1 z2 . By the continuity of xi and the assumption z1 =cont,L,σF z2 , we have xi z1F =cont,L,τ xi z2 for each i. By the induction 1 hypothesis, we have i∈ω (xi z1 ) =cont,L,τ i∈ω (xi z2 ). Therefore, we have: G G G G (xi z1 ) =cont,L,τ ( xi )z1 = (xi z2 ) = ( xi )z2 i∈ω i∈ω i∈ω i∈ω F as required. To check the second condition for i∈ω xi ∈ ContL,σ , suppose that F F F F {zi }i∈ω ∈ ContL,σ1 (↑ω) . We need to show ( i∈ω xi )( j∈ω z j ) =cont,L,τ j∈ω (( i∈ω xi )z j ). By the continuity of xi and the induction hypothesis, we have indeed: F F F F ( i∈ω xi )( j∈ω z j ) = i∈ω (xi ( j∈ω z j )) F F F F F F =cont,L,τ i∈ω ( j∈ω (xi z j )) = j∈ω ( i∈ω (xi z j )) = j∈ω (( i∈ω xi )z j ). F F Thus, we have proved i∈ω xi ∈ ContL,σ . The proof of i∈ω yi ∈ ContL,σ is the same. F F It remains to check that z =cont,L,σ1 w implies ( i∈ω xi )z =cont,L,τ ( i∈ω yi )w. Suppose z =cont,L,σ1 w. Then we have: F F F F ( i∈ω xi )z = i∈ω (xi z) =cont,L,τ i∈ω (yi w) = ( i∈ω yi )w. as required. Note that xi z =cont,L,τ yi w follows from the assumptions xi =cont,L,σ yi and and then we have applied the induction hypothesis to obtain F z =cont,L,σ1 w, F ⊔ i∈ω (yi w). This completes the proof for the induction step. ⊓ i∈ω (xi z) =cont,L,τ The following lemma guarantees the continuity of the functions expressed by fixpoint-free HFLZ formulas. Lemma 5 (continuity of fixpoint-free functions). Let L be an LTS. If ϕ is a closed, fixpoint-free HFLZ formula of type τ, then ~ϕ ∈ ContL,τ . 33 Proof. We write ~∆cont for the set of valuations: {ρ ∈ ~∆ | ρ( f ) ∈ ContL,σ for each f : σ ∈ ∆}, and =cont,L,∆ for: {(ρ1 , ρ2 ) ∈ ~∆cont × ~∆cont | ρ1 (x) =cont,L,σ ρ2 (x) for every x : σ ∈ ∆}. We show the following property by induction on the derivation of ∆ ⊢ ϕ : σ: If ϕ is fixpoint-free and ∆ ⊢ ϕ : σ, then (i) ρ1 =cont,L,∆ ρ2 imply ~∆ ⊢ ϕ : σ(ρ1 ) =cont,L,σ ~∆ ⊢ ϕ : σ(ρ2 ). (ii) For any increasing sequence of valuations ρ0 ⊑ ρ1 ⊑ ρ2 ⊑F· · · such that ρi ∈ ~∆cont for each i ∈ ω, ~∆ ⊢ ϕ : σ(⊔i∈ω ρi ) =cont,L,σ i∈ω ~∆ ⊢ ϕ : σ(ρi ). Then, the lemma would follow as a special case, where ∆ = ∅. We perform case analysis on the last rule used for deriving ∆ ⊢ ϕ : σ. We discuss only the main cases; the other cases are similar or straightforward. – Case HT-Var: In this case, ϕ = X and ∆ = ∆′ , X : σ. The condition (i) follows immediately by: ~∆ ⊢ ϕ : σ(ρ1 ) = ρ1 (X) =cont,L,σ ρ2 (X) = ~∆ ⊢ ϕ : σ(ρ2 ). To see (ii), suppose ρ0 ⊑ ρ1 ⊑ ρ2 ⊑ · · ·, with ρi ∈ ~∆cont for each i ∈ ω. Then, we have ~∆ ⊢ ϕ : σ(⊔i∈ω ρi ) = (⊔i∈ω ρi )(X) = ⊔i∈ω (ρi (X)) = ⊔i∈ω (~∆ ⊢ ϕ : σ(ρi )). By Lemma 4, ⊔i∈ω (ρi (X)) ∈ ContL,σ . We have thus ~∆ ⊢ ϕ : σ(⊔i∈ω ρi ) =cont,L,σ ⊔i∈ω (~∆ ⊢ ϕ : σ(ρi )) as required. – Case HT-Some: In this case, ϕ = haiϕ′ with ∆ ⊢ ϕ : • and σ = •. The condition (i) is trivial since ContL,• = DL,• . We also have the condition (ii) by: a ~∆ ⊢ haiϕ′ : σ(⊔i∈ω ρi ) = {s | ∃s′ ∈ ~∆ ⊢H ϕ′ : •(⊔i∈ω ρi ). s −→ s′ } a = {s | ∃s′ ∈ ⊔i∈ω (~∆ ⊢H ϕ′ : •(ρi )). s −→ s′ } (by induction hypothesis) a ′ ′ ′ = ⊔i∈ω {s | ∃s ∈ ~∆ ⊢H ϕ : •(ρi ). s −→ s } = ⊔i∈ω ~∆ ⊢H haiϕ′ : •(ρi ). – Case HT-All: In this case, ϕ = [a]ϕ′ with ∆ ⊢ ϕ : • and σ = •. The condition (i) is trivial. The condition (ii) follows by: a ~∆ ⊢ [a]ϕ′ : σ(⊔i∈ω ρi ) = {s | ∀s′ ∈ U.s −→ s′ ⇒ s′ ∈ ~∆ ⊢H ϕ′ : •(⊔i∈ω ρi )} a = {s | ∀s′ ∈ U.s −→ s′ ⇒ s′ ∈ ⊔i∈ω (~∆ ⊢H ϕ′ : •(ρi ))} (by induction hypothesis) a = ⊔i∈ω {s | ∀s′ ∈ U.s −→ s′ ⇒ s′ ∈ ~∆ ⊢H ϕ′ : •(ρi )} (*) = ⊔i∈ω ~∆ ⊢H [a]ϕ′ : •(ρi ). a To see the direction ⊆ in step (*), suppose ∀s′ ∈ U.s −→ s′ ⇒ s′ ∈ ⊔i∈ω (~∆ ⊢H a ϕ′ : •(ρi )) holds. Since U is finite, the set {s′ | s −→ s′ } is a finite set {s1 , . . . , sk }. For each j ∈ {1, . . . , k}, there exists i j such that s j ∈ ~∆ ⊢H ϕ′ : •(ρi j ). Let i′ = max(i1 , . . . , ik ). Then we have s′ ∈ ~∆ ⊢H ϕ′ : •(ρi′ ) for every 34 a s′ ∈ {s1 , . . . , sk }. We have thus ∀s′ ∈ U.s −→ s′ ⇒ s′ ∈ ~∆ ⊢H ϕ′ : •(ρi′ ), which implies s belongs to the set in the righthand side of (*). a To see the converse (i.e., ⊇), suppose ∀s′ ∈ U.s −→ s′ ⇒ s′ ∈ ~∆ ⊢H ϕ′ : •(ρi ) a for some i ∈ ω. Then, ∀s′ ∈ U.s −→ s′ ⇒ s′ ∈ ⊔i∈ω (~∆ ⊢H ϕ′ : •(ρi )) follows immediately. – Case HT-Abs: In this case, ϕ = λX : σ1 .ϕ′ , with ∆, X : σ1 ⊢ ϕ′ : τ and σ = σ1 → τ. To prove the condition (i), suppose ρ1 =cont,L,∆ ρ2 . Let f j = ~∆ ⊢ λX : σ1 .ϕ′ : σ1 → τ(ρ j ) for j ∈ {1, 2}. We first check f j ∈ ContL,σ1 →τ . Suppose x1 =cont,L,σ1 x2 . Then, by the induction hypothesis, we have f j x1 = ~∆, X:σ1 ⊢ ϕ′ : τ(ρ j {X 7→ x1 }) =cont,L,τ ~∆, X:σ1 ⊢ ϕ′ : τ(ρ j {X 7→ x2 }) = f j x2 . To check the second condition for f j ∈ ContL,σ , let {yi }i∈ω ∈ ContL,σ1 (↑ω) . Then we have: f j (⊔i∈ω yi ) = ~∆, X : σ1 ⊢ ϕ′ : τ(ρ j {X 7→ ⊔i∈ω yi }) = ~∆, X : σ1 ⊢ ϕ′ : τ(⊔i∈ω (ρ j {X 7→ yi })) =cont ⊔i∈ω ~∆, X : σ1 ⊢ ϕ′ : τ(ρ j {X 7→ yi }) (by the induction hypothesis) = ⊔i∈ω f j (yi ) as required. To show f1 =cont,L,σ1→τ f2 , assume again that x1 =cont,L,σ1 x2 . Then, ρ1 {X 7→ x1 } =cont,∆,X:σ1 ρ2 {X 7→ x2 }. Therefore, by the induction hypothesis, we have f1 (x) = ~∆, X:σ1 ⊢ ϕ′ : τ(ρ1 {X 7→ x1 }) =cont,L,τ ~∆, X:σ1 ⊢ ϕ′ : τ(ρ2 {X 7→ x2 }) = f2 (x). This completes the proof of the condition (i). To prove the condition (ii), suppose ρ0 ⊑ ρ1 ⊑ ρ2 ⊑ · · · with ρi ∈ ~∆cont for each i ∈ ω. We need to show ~∆ ⊢ ϕ : σ(⊔i∈ω ρi ) =cont,L,σ ⊔i∈ω ~∆ ⊢ ϕ : σ(ρi ). We first check that both sides of the equality belong to ContL,σ . By Lemma 4, ⊔i∈ω ρi ∈ ~∆cont . Thus, by the condition (i) (where we set both ρ1 and ρ2 to ⊔i∈ω ρi ), we have ~∆ ⊢ ϕ : σ(⊔i∈ω ρi ) =cont,L,σ ~∆ ⊢ ϕ : σ(⊔i∈ω ρi ), which implies ~∆ ⊢ ϕ : σ(⊔i∈ω ρi ) ∈ ContL,σ . For the righthand side, by the condition (i), we have ~∆ ⊢ ϕ : σ(ρi ) ∈ ContL,σ for each i. By Lemma 4, we have ⊔i∈ω ~∆ ⊢ ϕ : σ(ρi ) as required. It remains to check that x =cont,L,σ1 y implies ~∆ ⊢ ϕ : σ(⊔i∈ω ρi )(x) =cont,L,τ ⊔i∈ω ~∆ ⊢ ϕ : σ(ρi )(y). Suppose x =cont,L,σ1 y. Then we have ~∆ ⊢ ϕ : σ(⊔i∈ω ρi )(x) = ~∆, X : σ1 ⊢ ϕ′ : σ((⊔i∈ω ρi ){X 7→ x}) = ~∆, X : σ1 ⊢ ϕ′ : σ(⊔i∈ω (ρi {X 7→ x})) =cont ~∆, X : σ1 ⊢ ϕ′ : σ(⊔i∈ω (ρi {X 7→ y})) (by the induction hypothesis, (i)) =cont ⊔i∈ω ~∆, X : σ1 ⊢ ϕ′ : σ(ρi {X 7→ y}) (by the induction hypothesis, (ii)) = ⊔i∈ω ~∆ ⊢ ϕ : σ(ρi )(y) as required. 35 – Case HT-App: In this case, ϕ = ϕ1 ϕ2 and σ = τ, with ∆ ⊢ ϕ1 : σ2 → τ, ∆ ⊢ ϕ2 : σ2 . The condition (i) follows immediately from the induction hypothesis. To prove the condition (ii), suppose ρ0 ⊑ ρ1 ⊑ ρ2 ⊑ · · · with ρi ∈ ~∆cont for each i ∈ ω. Then we have ~∆ ⊢ ϕ1 ϕ2 : τ(⊔i∈ω ρi ) = ~∆ ⊢ ϕ1 : σ2 → τ(⊔i∈ω ρi )~∆ ⊢ ϕ2 : σ2 (⊔i∈ω ρi ) =cont (⊔i∈ω ~∆ ⊢ ϕ1 : σ2 → τ(ρi ))(⊔ j∈ω ~∆ ⊢ ϕ2 : σ2 (ρ j )) (by the induction hypothesis) = ⊔i∈ω (~∆ ⊢ ϕ1 : σ2 → τ(ρi )(⊔ j∈ω ~∆ ⊢ ϕ2 : σ2 (ρ j )) =cont ⊔i∈ω ⊔ j∈ω (~∆ ⊢ ϕ1 : σ2 → τ(ρi )(~∆ ⊢ ϕ2 : σ2 (ρ j )) (by the continuity of ~∆ ⊢ ϕ1 : σ2 → τ(ρi ) and Lemma 4) = ⊔i∈ω (~∆ ⊢ ϕ1 : σ2 → τ(ρi )(~∆ ⊢ ϕ2 : σ2 (ρi )) = ⊔i∈ω (~∆ ⊢ ϕ1 ϕ2 : τ(ρi ) as required. ⊔ ⊓ The following is an immediate corollary of the lemma above (see, e.g., [38]). Lemma 6 (fixpointF of continuous functions). Let L be an LTS. If f ∈ ContL,τ→τ , then lfpτ ( f ) =cont,L,τ i∈ω f i (⊥L,τ ). F F Proof. By the continuity of f , we have f ( i∈ω f i (⊥L,τ )) =cont,L,τ i∈ω f i (⊥L,τ ). Thus, F by (transfinite) induction, we have f β (⊥) =cont,L,τ i∈ω f i (⊥L,τ ). for every ordinal β ≥ ω. Since lfpτ ( f ) = f β (⊥) for some ordinal β [38], we have the required result. The detail of the transfinite induction is given as follows. For a given ordinal number β, let us define f β (⊥L,τ ) by transfinite induction. IfFβ = β′ + 1, then ′ ′ f β (⊥L,τ ) = f ( f β (⊥L,τ )). If β is a limit ordinal, then f β (⊥L,τ ) = β′ <β f β (⊥L,τ ). We prove that f ω (⊥L,τ ) =cont f β (⊥L,τ ) for every β > ω by transfinite induction. We have proved the claim for β = ω + 1, i.e. f ω (⊥L,τ ) =cont f ω+1 (⊥L,τ ). ′ – If β = β′ + 1, by the induction hypothesis, f ω (⊥L,τ ) =cont f β . Since f is continuous, ′ f β (⊥L,τ ) = f ( f β (⊥L,τ )) =cont f ( f ω (⊥L,τ )) =cont f ω (⊥L,τ ). – Assume that β is a limit ordinal and τ = σ1 → · · · → σk → •. Assume also xi =cont,L,σi yi for each 1 ≤ i ≤ k. By the induction hypothesis, for every β′ < β, we have ′ f β (⊥L,τ ) =cont f ω (⊥L,τ ) and thus ′ f β (⊥L,τ ) x1 . . . xk =cont f ω (⊥L,τ ) y1 . . . yk . Since =cont on ContL,• is the standard equivalence =, for every β′ < β, we have ′ f β (⊥L,τ ) x1 . . . xk = f ω (⊥L,τ ) y1 . . . yk . 36 Since the limit of a function is defined pointwise,    G ′ G ′  β   x1 . . . xk = ( f β (⊥L,τ ) x1 . . . xk ) f (⊥ ) L,τ   β′ <β β′ <β G = ( f ω (⊥L,τ ) y1 . . . yk ) β′ <β = f ω (⊥L,τ ) y1 . . . yk . Thus, we have f β (⊥L,τ ) =cont f ω . Since lfpτ ( f ) = f β (⊥L,τ ) for some ordinal β, we have G f i (⊥L,τ ) = f ω (⊥L,τ ) =cont f β (⊥L,τ ) = lfpτ ( f ) i∈ω ⊔ ⊓ as required. We are now ready to prove Theorem 1. Below we extend continuous functions to those on tuples, and make use of Bekić property between a simultaneous recursive definition of multiple functions and a sequence of recursive definitions; see, e.g., [48], Chapter 10. Proof (of Theorem 1). Given a program P = (D, t) with D = { f1 xe1 = t1 , . . . , fn xen = (i) (i) tn }, we write P(i) for the recursion-free program (D(i) , [ f1 / f1 , . . . , fn / fn ]t) where: (j+1) (j) (j) e D(i) = { fk xk = [ f1 / f1 , . . . , fn / fn ]tk | j ∈ {1, . . . , i − 1}, k ∈ {1, . . . , n}} (0) ∪{ fk e xk = ( ) | k ∈ {1, . . . , n}}. Then, we obtain the required result as follows. L0 |= ΦP,may ⇔ s⋆ ∈ ~t†may ({( f1 , . . . , fn ) 7→ lfp(λ( f1 , . . . , fn ).(λe x1 .~t1 †may , . . . , λxen .~tn †may ))}) (Bekić property) ⇔ s⋆ ∈ ~t†may ({( f , . . . , f ) → 7 1 n F x1 .∅, . . . , λxen .∅)}) x1 .~t1 †may , . . . , λxen .~tn †may ))i (λe i∈ω (λ( f1 , . . . , fn ).(λe (Lemmas 5 and 6) ⇔ s⋆ ∈ ~t†may ({( f1 , . . . , fn ) 7→ x1 .∅, . . . , λxen .∅)}) (λ( f1 , . . . , fn ).(λe x1 .~t1 †may , . . . , λxen .~tn †may ))i (λe for some i (Lemmas 5) ⇔ L0 |= ΦP(i) ,may for some i (by the definition of P(i) ) (i) ⇔ a ∈ Traces(P ) for some i (Lemma 3) ⇔ a ∈ Traces(P) ⊔ ⊓ B.2 Proofs for Section 5.2 We first prepare lemmas corresponding to Lemmas 1–3. 37 Lemma 7. Let (D, t) be a program. If t −→D t′ , then ~(D, t)†must  = ~(D, t′ )†must . ⊔ ⊓ Proof. Almost the same as the proof of Lemma 1. †must Lemma 8. Let (D, t) be a program and t −→ 6 D . Then, L0 |= (D, t) if and only if t is of the form C[event a; t1 , . . . , event a; tk ], where C is a (multi-hole) context generated by the syntax: C ::= [ ]i | C1 C2 . Proof. The proof proceeds by induction on the structure of t. By the condition t −→ 6 D , t is generated by the following grammar: t ::= ( ) | event a; t′ | t1 t2 . – Case t = ( ): The result follows immediately, as t is not of the form C[event a; t1 , . . . , event a; tk ] and t†must = false. – Case t = event a; t′ : The result follows immediately, as t is of the form C[event a; t1 , . . . , event a; tk ] (where C = [ ]1 and k = 1) and t†must = true. – Case t1 t2 : Because (t1 t2 )†must = t1 †must ∧ t2 †must , L0 |= (D, t)†must if and only if L0 |= (D, ti )†must for each i ∈ {1, 2}. By the induction hypothesis, the latter is equivalent to the property that ti is of the form Ci [event a; ti,1 , . . . , event a; ti,ki ] for each i ∈ {1, 2}, which is equivalent to the property that t is of the form C[event a; t1 , . . . , event a; tk ] (where C = C1 C2 ). ⊔ ⊓ Lemma 9. Let P be a recursion-free program. Then, Musta (P) if and only if L0 |= ΦP,must for L0 = ({sinit}, ∅, ∅, sinit). Proof. Since P = (D, t0) is recursion-free, there exists a finite, normalizing reduction sequence t0 −→∗D t −→ 6 D . We show the required property by induction on the length n of this reduction sequence. – Case n = 0: Since t0 −→ 6 D , the result follows immediately from Lemma 8. – Case n > 0: In this case, t0 −→D t1 −→∗P t. Thus, by the induction hypothesis, the definition of the reduction semantics and Lemma 7, Musta (D, t0) if and only if Musta (D, t1 ), if and only if L0 |= Φ(D,t1 ),must , if and only if L0 |= Φ(D,t0 ),must . Proof (Theorem 2). Let P(i) be the the recursion-free program defined in the proof of Theorem 1. Then, we obtain the required result as follows. L0 |= ΦP,must ⇔ s⋆ ∈ ~t†must ({( f1 , . . . , fn ) 7→ lfp(λ( f1 , . . . , fn ).(λe x1 .~t1 †must , . . . , λxen .~tn †must ))}) (Bekić property) ⇔ s⋆ ∈ ~t†must ({( f , . . . , f ) → 7 n F 1 x1 .∅, . . . , λxen .∅)}) x1 .~t1 †must , . . . , λxen .~tn †must ))i (λe i∈ω (λ( f1 , . . . , fn ).(λe ((Lemmas 5 and 6) ⇔ s⋆ ∈ ~t†must ({( f1 , . . . , fn ) 7→ x1 .∅, . . . , λxen .∅)}) for some i (λ( f1 , . . . , fn ).(λe x1 .~t1 †must , . . . , λxen .~tn †must ))i (λe ⇔ L0 |= ΦP(i) ,must for some i (by the definition of P(i) ) (i) ⇔ Musta (P ) for some i (Lemma 9) ⇔ Musta (P). 38 ⊔ ⊓ C Proofs for Section 6 We first modify the reduction semantics in Figure 8 by adding the following rule for distributing events with respect to : E[event a; (t1 t2 )] −→D E[(event a; t1 )(event a; t2 )] (R-Dist) We write −→D,dist for this modified version of the reduction relation. We call an evaluation context event-free if it is generated only by the following syntax: E ::= [ ] | Et | tE. Lemma 10. Let P = (D, t) be a program such that t−→ 6 D,dist. Then, FinTraces(P) ⊆ L if and only if LL |= ΦP,path . Proof. Since t−→ 6 D,dist, t must be of the form: i∈{1,...,m} (event ai,1 ; · · · event ai,ni ; ( )) (where i∈{1,...,m} ti denotes a combination of t1 , . . . , tm with ). Thus, FinTraces(P) ⊆ L ⇔ ai,1 · · · ai,ni ∈ L for every i ∈ {1, . . . , m} ⇔ LL |= hai,1 i · · · hai,ni itrue for every i ∈ {1, . . . , m} ⇔ LL |= ΦP,path (by the construction of LL ) (by the definition of ΦP,path ). ⊔ ⊓ Lemma 11. Let (D, t) be a program, and L be a prefix-closed regular language. If t −→D,dist t′ , then ~(D, t)†path LL = ~(D, t′ )†path LL . xn = tn }, and (F1 , . . . , Fn ) be the greatest fixpoint x1 = t1 , . . . , fn e Proof. Let D = { f1 e of λ(X1 , . . . , Xn ).(~λe x1 .t1 †path LL ({ f1 7→ X1 , . . . , fn 7→ Xn }), . . . , ~λe xn .tn †path LL ({ f1 7→ X1 , . . . , fn 7→ Xn })). By the Bekić property, ~(D, t)†path LL = ~t†path LL ({ f1 7→ F1 , . . . , fn 7→ Fn }). Thus, it suffices to show that t −→D t′ implies ~t†path LL (ρ) = ~t′ †path LL (ρ) for ρ = { f1 7→ F1 , . . . , fn 7→ Fn }. We show it by case analysis on the rule used for deriving t −→D,dist t′ . t] and t′ = E[[e t/e xi ]s] with D( fi ) = λe xi .s. – Case R-Fun: In this case, t = E[ fi e Since (F1 , . . . , Fn ) is a fixpoint, we have: tLL (ρ) = Fi (~e tLL (ρ) ~ fi e = ~λe xi .sLL (ρ)(~e tLL (ρ)) = ~[e t/e xi ]sLL (ρ) Thus, we have ~t†path LL (ρ) = ~t′ †path LL (ρ) as required. 39 – Case R-IfT: In this case, t = E[if p(t′1 , . . . , t′k ) then t1 else t2 ] and t′ = E[t1 ] with (~t′1 LL , . . . , ~t′k LL ) ∈ ~p. Thus,t†path = (p(t′1 , . . . , t′k ) ⇒ t1 †path ) ∧ (¬p(t′1 , . . . , t′k ) ⇒ t2 †path ). Since (~t′1 LL , . . . , ~t′k LL ) ∈ ~p, (~t′1 LL , . . . , ~t′k LL ) < ~¬p. Thus, ~(if p(t′1 , . . . , t′k ) then t1 else t2 )†path LL (ρ) = ~(true ⇒ t1 †path ) ∧ (false ⇒ t2 †path )LL (ρ) = ~t1 †path LL (ρ). We have, therefore, ~t†path LL (ρ) = ~t′ †path LL (ρ) as required. – Case R-IfF: Similar to the above case. – Case R-Dist: In this case, t = E[event a; (t1 t2 )] and t′ = E[(event a; t1 )(event a; t2 )]. Thus, it suffices to show that ~(event a; (t1 t2 ))†path LL (ρ) = ~((event a; t1 )(event a; t2 ))†path LL (ρ). We have: (event a; (t1 t2 ))†path = hai(t1†path ∧ t2 †path ) ((event a; t1 )(event a; t2 ))†path = (hait1†path ) ∧ (hait2 †path ) Since LL has at most one a-transition from each state, both formulas are equivalent, i.e., ~(event a; (t1 t2 ))†path LL (ρ) = ~((event a; t1 )(event a; t2 ))†path LL (ρ). ⊔ ⊓ Lemma 12. Let P be a recursion-free program and L be a regular, prefix-closed language. Then, FinTraces(P) ⊆ L if and only if LL |= ΦP,path . Proof. Since P = (D, t0 ) is recursion-free, there exists a finite, normalizing reduction sequence t0 −→∗D,dist t−→ 6 D,dist. We show the required property by induction on the length n of this reduction sequence. The base case follows immediately from Lemma 10. For the induction step (where n > 0), we have: t0 −→D t1 −→∗D t. By the induction hypothesis, FinTraces(D, t1 ) ⊆ L if and only if LL |= Φ(D,t1 ),path . Thus, by the definition of the reduction semantics and Lemma 11, FinTraces(D, t0 ) ⊆ L, if and only if FinTraces(D, t1 ) ⊆ L, if and only if LL |= Φ(D,t1 ),path , if and only if LL |= Φ(D,t0 ),path . ⊔ ⊓ To prove Theorem 3, we introduce the (slightly non-standard) notion of cocontinuity, which is dual of the continuity in Definition 11. Definition 12. For an LTS L = (U, A, −→, sinit) and a type σ, the set of cocontinuous elements CocontL,σ ⊆ DL,σ and the equivalence relation =cocont,L,σ⊆ CocontL,σ × CocontL,σ are defined by induction on σ as follows. CocontL,• = DL,• CocontL,int = DL,int CocontL,σ→τ = { f ∈ DL,σ→τ | ∀x1 , x2 ∈ CocontL,σ .(x1 =cocont,L,σ x2 ⇒ f (x1 ) =cocont,L,τ f (x2 ))   ∧∀{yi }i∈ω ∈ CocontL,σ (↓ω) . f ( i∈ω yi ) =cocont,L,τ i∈ω f (yi )}. =cocont,L,• = {(x, x) | x ∈ CocontL,• } =cocont,L,int = {(x, x) | x ∈ CocontL,int} =cocont,L,σ→τ = {( f1 , f2 ) | f1 , f2 ∈ CocontL,σ→τ ∧ ∀x1 , x2 ∈ CocontL,σ .x1 =cocont,L,σ x2 ⇒ f1 (x1 ) =cocont,L,τ f2 (x2 )}. Here, CocontL,σ (↓ω) denotes the set of decreasing infinite sequences a0 ⊒ a1 ⊒ a2 ⊒ · · · consisting of elements of CocontL,σ . We just write =cocont for =cocont,L,σ when L and σ are clear from the context. 40 The following lemma is analogous to Lemma 5. Lemma 13 (cocontinuity of fixpoint-free functions). Let L be an LTS. If ϕ is a closed, fixpoint-free HFL formula of type τ, then ~ϕL ∈ CocontL,τ . Proof. The proof is almost the same as that of Lemma 5. We write ~∆cocont for the set of valuations: {ρ ∈ ~∆L | ρ( f ) ∈ CocontL,σ for each f : σ ∈ ∆}. We show the following property by induction on the derivation of ∆ ⊢ ϕ : σ: If ϕ is fixpoint-free and ∆ ⊢ ϕ : σ, then (i) ~∆ ⊢ ϕ : σL (ρ) ∈ CocontL,σ for every ρ ∈ ~∆cocont ; and (ii) For any decreasing sequence of interpretations ρ0 ⊒ρ1 ⊒ ρ2 ⊒ · · · such  that ρi ∈ ~∆cont for each i ∈ ω, ~∆ ⊢ ϕ : σL ( i∈ω ρi ) =cocont i∈ω ~∆ ⊢ ϕ : σL (ρi ). Then, the lemma will follow as a special case, where ∆ = ∅. We perform case analysis on the last rule used for deriving ∆ ⊢ ϕ : σ. We discuss only two cases below, as the proof is almost the same as the corresponding proof for Lemma 5. – Case HT-Some: In this case, ϕ = haiϕ′ with ∆ ⊢ ϕ : • and σ = •. The condition (i) is trivial since CocontL,• = DL,• . We also have the condition (ii) by:   a ~∆ ⊢ haiϕ′ : σL ( i∈ω ρi ) = {s | ∃s′ ∈ ~∆ ⊢H ϕ′ : •L ( i∈ω ρi ). s −→ s′ }  a = {s | ∃s′ ∈ i∈ω (~∆ ⊢H ϕ′ : •L (ρi )). s −→ s′ } (by induction hypothesis)  a = i∈ω {s | ∃s′ ∈ ~∆ ⊢H ϕ′ : •L (ρi ). s −→ s′ } (*) = i∈ω ~∆ ⊢H haiϕ′ : •L (ρi ).  ′ The step (*) is obtained as follows. Suppose ∃s′ ∈ i∈ω .~∆ ⊢H ϕ :  a •L (ρi ). s −→ s′ . Since i∈ω .~∆ ⊢H ϕ′ : •L (ρi ) ⊆ ~∆ ⊢H ϕ′ : •L (ρi ) for a every i, we have ∃s′ ∈ ~∆ ⊢H ϕ′ : •L (ρi ). s −→ s′ for every i; hence we have s∈  i∈ω {s a | ∃s′ ∈ ~∆ ⊢H ϕ′ : •L (ρi ). s −→ s′ }. Conversely, suppose s∈  i∈ω {s a | ∃s′ ∈ ~∆ ⊢H ϕ′ : •L (ρi ). s −→ s′ }, a i.e., for every i, Si = {s′ ∈ ~∆ ⊢H ϕ′ : •L (ρi ) | s −→ s′ } must be non-empty. T a Since the set {s′ | s −→ s′ } is finite, and Si decreases monotonically, i∈ω Si must be non-empty. Thus, we have ∃s′ ∈  i∈ω ~∆ a ⊢H ϕ′ : •L (ρi ). s −→ s′ , as required. 41 – Case HT-All: In this case, ϕ = [a]ϕ′ with ∆ ⊢ ϕ : • and σ = •. The condition (i) is trivial. The condition (ii) follows by:   a ~∆ ⊢ [a]ϕ′ : σL ( i∈ω ρi ) = {s | ∀s′ ∈ U.s −→ s′ ⇒ s′ ∈ ~∆ ⊢H ϕ′ : •L ( i∈ω ρi )}  a = {s | ∀s′ ∈ U.s −→ s′ ⇒ s′ ∈ i∈ω (~∆ ⊢H ϕ′ : •L (ρi ))} (by induction hypothesis)  a ′ ′ ′ ′ = i∈ω {s | ∀s ∈ U.s −→ s ⇒ s ∈ ~∆ ⊢H ϕ : •L (ρi )} = i∈ω ~∆ ⊢H [a]ϕ′ : •L (ρi ). ⊔ ⊓ The following lemma states a standard property of cocontinuous functions [38], which can be proved in the same manner as Lemma 6. Lemma 14 (fixpoint of cocontinuous functions). Let L be an LTS. If f ∈  CocontL,τ→τ , then gfpτ ( f ) =cocont,L,τ i∈ω f i (⊤L,τ ). Proof (Proof of Theorem 3). The result follows by: LL 6|= ΦP,path ⇔ q0 < ~t†path LL ({( f1 , . . . , fn ) 7→ gfp(λ( f1 , . . . , fn ).(λe x1 .~t1 †path LL , . . . , λxen .~tn †path LL ))}) (Bekić property) †path ⇔ q0< ~t LL ({( f1 , . . . , fn ) 7→ x1 .Q, . . . , λxen .Q)}) x1 .~t1 †path LL , . . . , λxen .~tn †path LL ))i (λe i∈ω (λ( f1 , . . . , fn ).(λe (Lemmas 13 and 14) ⇔ q0 < ~t†path LL ({( f1 , . . . , fn ) 7→ x1 .Q, . . . , λxen .Q)}) (λ( f1 , . . . , fn ).(λe x1 .~t1 †path LL , . . . , λxen .~tn †path LL ))i (λe for some i (Lemma 13) ⇔ LL 6|= ΦP(i) ,path for some i ⇔ Traces(P(i) ) * L for some i (Lemma 12) ⇔ Traces(P) * L ⊔ ⊓ D Game-based characterization of HES Fix an LTS L and let E = (X1τ1 =α1 ϕ1 ; · · · ; Xnτn =αn ϕn ). The goal of this section is to construct a parity game GL,E characterizing the semantic interpretation of the HES E (or equivalently (E, X1 )) over the LTS L. The game-based characterization will be used to prove some results in Section 7. D.1 Preliminary: Parity Game A parity game G is a tuple (VP , VO , E, Ω) where – VP and VO are disjoint sets of Proponent and Opponent nodes, respectively, – E ⊆ (VP ∪ VO ) × (VP ∪ VO ) is the set of edges, and – Ω : (VP ∪ VO ) → N is a priority function whose image is bounded. 42 We write V for VP ∪ VO . A play of a parity game is a (finite or infinite) sequence v1 v2 . . . of nodes in V such that (vi , vi+1 ) ∈ E for every i. We write · for the concatenation operation. An infinite play v1 v2 . . . is said to satisfy the parity condition if max Inf(Ω(v1 )Ω(v2 ) . . . ) is even, where Inf(m1 m2 · · · ) is the set of numbers that occur infinitely often in m1 m2 · · ·. A play is maximal if either – it is a finite sequence v1 · · · vn and the last node vn has no successor (i.e., {v | (vn , v) ∈ E} = ∅), or – it is infinite. A maximal play is P-winning (or simply winning) if either – it is finite and the last node is of Opponent, or – it is infinite and satisfies the parity condition. Let S : V ∗ VP ⇀ V be a partial function that respects E in the sense that (vk , S(v1 . . . vk )) ∈ E (if S(v1 . . . vk ) is defined). A play v1 v2 . . . vk is said to conform with S if, for every 1 ≤ i < k with vi ∈ VP , S(v1 . . . vi ) is defined and S(v1 . . . vi ) = vi+1 . An infinite play conforms with S if so does every finite prefix of the play. The partial function S is a P-strategy on V0 ⊆ V (or simply a strategy on V0 ) if it is defined on every play that conforms with S, starts from a node in V0 and ends with a node in VP . A strategy is P-winning on V0 ⊆ V (or simply winning on V0 ) if every maximal play that conforms with the strategy and starts from V0 is P-winning. We say that Proponent wins the game G on V0 ⊆ V if there exists a P-winning strategy of G on V0 . An O-strategy and an O-winning strategy is defined similarly. A strategy S is memoryless if S(v1 . . . vk ) = f (vk ) for some f : VP → V. We shall consider only games of limited shape, which we call bipartite games. A game is bipartite if E ⊆ (VP × VO ) ∪ (VO × VP ) and Ω(vO ) = 0 for every vO ∈ VO . Parity progress measure [17] is a useful tool to show that a strategy is winning. We give a modified version of the definition, applicable only to bipartite games. Definition 13 (Parity progress measure). Let γ be an ordinal number. Given (β1 , . . . , βn ), (β′1 , . . . , β′n ) ∈ γn , we write (β1 , . . . , βn ) ≥ j (β′1 , . . . , β′n ) if (β1 , . . . , βn− j+1 ) ≥ (β′1 , . . . , β′n− j+1 ) by the lexicographic ordering. The strict inequality > j ( j = 1, . . . , n) is defined by an analogous way. Let G be a bipartite parity game and S be a strategy of G on V0 . Let n be the maximum priority in G. A (partial) mapping ̟ : VO ⇀ γn is a parity progress measure of S on V0 if it satisfies the following condition: For every finite play ṽ · vO · vP · v′O (vO , v′O ∈ VO and vP ∈ VP ) that starts from V0 and conforms with the strategy S, both ̟(vO ) and ̟(v′O ) are defined and ̟(vO ) ≥Ω(vP ) ̟(v′O ). Furthermore if Ω(vP ) is odd, then ̟(vO ) >Ω(vP ) ̟(v′O ). Lemma 15 ([17]). Let G be a bipartite parity game and S be a strategy of G on V0 ⊆ V. If there exists a parity progress measure of S on V0 , then S is a winning strategy on V0 . 43 Proof. Let ̟ be a parity progress measure of S on V0 . We prove by contradiction. Assume an infinite play v1 v2 . . . that conforms S, starts from V0 and violates the parity condition. Assume that v1 ∈ VO ; the other case can be proved similarly. Since G is bipartite, vi ∈ VO if and only if i is odd. Furthermore Ω(vi ) = 0 for every odd i. Let ℓ = max Inf(Ω(v1 )Ω(v2 ) . . . ). Then there exists an even number k such that, for every even number i ≥ k, we have Ω(vi ) ≤ ℓ. By definition, ̟(vi−1 ) ≥Ω(vi ) ̟(vi+1 ) for every even i. Since j1 ≤ j2 implies (≥ j1 ) ⊆ (≥ j2 ), we have ̟(vi−1 ) ≥ℓ ̟(vi+1 ) for every even i ≥ k. Hence we have an infinite decreasing chain ̟(vk−1 ) ≥ℓ ̟(vk+1 ) ≥ℓ ̟(vk+3 ) ≥ℓ · · ·. Furthermore this inequality is strict if Ω(vi ) = ℓ because ℓ is odd. Since {i | i ≥ k, Ω(vi ) = ℓ} is infinite, we have an infinite strictly decreasing chain, a contradiction. ⊔ ⊓ D.2 Preliminary: Complete-Prime Algebraic Lattice W Definition 14. Let (A, ≤) be a complete lattice. For U ⊆ A, we write U for the least upper bound W of U in A. An element p ∈ A is a complete prime if (1) p , ⊥, and (2) p ≤ U implies p ≤ x for some x W ∈ U (for every U ⊆ A). A complete lattice is complete-prime algebraic [49] if x = { p ≤ x | p : complete prime } for every x. The following is a basic property about complete-prime algebraic lattices, which will be used later. Lemma 16. Let (A, ≤) be a complete prime algebraic lattice, and x, y ∈ A. Then the followings are equivalent. (i) x ≤ y. (ii) p ≤ x implies p ≤ y for every complete prime p. (iii) p  y implies p  x for every complete prime p. Proof. (i) implies (ii): Trivial from the fact p ≤ x ≤ y implies p ≤ y. (ii) implies (i): Assume (ii). Then, {p | p ≤ x, p: complete prime} ⊆ {p | p ≤ y, p: complete prime}. W W Since A is complete-prime algebraic, x = {p | p ≤ x, p: complete prime} ≤ {p | p ≤ y, p: complete prime} = y. (iii) is just a contraposition of (ii). ⊔ ⊓ Let (A, ≤) be a complete lattice and f : A → A be a monotone function. We write ⊥A for the least element of A. Then A has the least fixed point of f , which can be computed by iterative application of f to ⊥A as follows. Let γ be an ordinal number greater than the cardinality of A. We define a family {ai }i<γ of elements in A by: _ a0 := ⊥A aβ+1 := f (aβ ) aβ := aβ′ (if β is a limit ordinal). β′ <β Then aγ is the least fixed point of f . We shall write aβ as f β (⊥A ). Lemma 17. Let (A, ≤) be a complete lattice and f : A → A be a monotone function. For every complete prime p ∈ A, the minimum ordinal β such that p ≤ f β (⊥) is a successor ordinal. 44 Proof.WAssume that the minimum ordinal β is a limit ordinal. By definition, ′ p ≤ β′ <β f β (⊥). Since p is a complete prime, there exists β′ < β such that ′ p ≤ f β (⊥). Hence β is not the minimum, a contradiction. ⊔ ⊓ Lemma 18. Let L be an LTS. Then (DL,τ , ⊑L,τ) is complete-prime algebraic for every τ. Proof. If τ = •, then DL,τ = 2U is ordered by the set inclusion and the complete primes are singleton sets { q } (q ∈ U). For function types, DL,σ→τ is the set of monotone functions ordered by the pointwise ordering. Given an element d ∈ DL,σ and a complete prime p ∈ DL,τ , consider the function fd,p defined by    p (if d ≤ x) fd,p (x) =   ⊥ (otherwise). These functions are complete primes of DL,σ→τ . It is not difficult to see that every element is the least upper bound of a subset of complete primes. ⊔ ⊓ D.3 Parity Game for HES Let E = (X1τ1 =α1 ϕ1 ; · · · ; Xnτn =αn ϕn ) be an HES (with “the main formula” X1 ) and L be an LTS. Definition 15 (Parity game for HES). The parity game GL,E is defined by the following data: VP := { (p, Xi) | p ∈ DL,τi , p: complete prime } VO := { (x1 , . . . , xn ) | x1 ∈ DL,τ1 , . . . , xn ∈ DL,τn } E := { ((p, Xi), (x1 , . . . , xn )) | p ⊑ ~ϕi ([X1 7→ x1 , . . . , Xn 7→ xn ]) } ∪ { ((x1, . . . , xn ), (p, Xi)) | p ⊑ xi , p: complete prime }. The priority of the opponent node is 0; the priority of node (p, Xi ) is 2(n − i) if αi = ν and 2(n − i) + 1 if αi = µ. The parity game defined above is analogous to the typability game for HES defined by [21]. A position (p, Xi ) ∈ VP represents the state where Proponent tries to show that p ⊑ ~(E, Xi ). To show p ⊑ ~(E, Xi ), Proponent picks a valuation ρ = [X1 7→ x1 , . . . , Xn 7→ xn ] such that p ⊑ ~ϕi (ρ) indeed holds. A position (x1 , . . . , xn ) ∈ VO represents such a valuation, and Opponent challenges Proponent’s assumption that ρ = [X1 7→ x1 , . . . , Xn 7→ xn ] is a valid valuation, i.e., xi ⊑ ~(E, Xi ) holds for each i. To this end, Opponent chooses i, picks a complete prime p such that p ⊑ xi , and asks why p ⊑ ~(E, Xi ) holds, as represented by the edge ((x1 , . . . , xn ), (p, Xi)). (Note that Lemma 18 implies that xi ⊑ ~(E, Xi ) if and only if p ⊑ ~(E, Xi ) for every complete prime such that p ⊑ xi ; therefore, it is sufficient for Opponent to consider only complete primes.) As in the typability game for HES defined by [21], a play may continue indefinitely, in which case, the winner is determined by the largest priority of (p, Xi ) visited infinitely often. 45 The goal of this section is to show that p ⊑ ~EL if and only if Proponent wins GL,E on (p, X1 ) (Theorem 7 given at the end of this subsection). We start from an alternative description of the interpretation of an HES. Let us define the family {ϑi }i=1,...,n of substitutions by induction on n − i as follows: ϑn := (the identity substitution) ϑi := [Xi+1 7→ αi+1 Xi+1 .(ϑi+1 ϕi+1 )] ◦ ϑi+1 (if i < n). The family {ψi }i=1,...,n of formulas is defined by ψi := αi Xi .(ϑi ϕi ). Then ϑi = [Xi+1 7→ ψi+1 ] ◦ ϑi+1 . The substitution ϑi maps Xi+1 , . . . , Xn to closed formulas and thus ϑi ϕ j ( j ≤ i) is a formula with free variables X1 , . . . , Xi . The formula ψi has free variables X1 , . . . , Xi−1 . In particular, ψ1 = toHFL(E, X1 ). Given 1 ≤ i ≤ j ≤ n and dk ∈ DL,τk (1 ≤ k ≤ n), let us define fi, j (d1 , . . . , d j ) := ~ϑ j ϕi ([X1 7→ d1 , . . . , X j 7→ d j ]) gi (d1 , . . . , di−1 ) := ~ψi ([X1 7→ d1 , . . . , Xi−1 7→ di−1 ]). Lemma 19. We have fi,n (d1 , . . . , dn ) = ~ϕi ([X1 7→ d1 , . . . , Xn 7→ dn ]) fi, j (d1 , . . . , d j ) = fi, j+1 (d1 , . . . , d j , g j+1 (d1 , . . . , d j )) (if j < n) gi (d1 , . . . , di−1 ) = αi y. fi,i (d1 , . . . , di−1 , y). g1 ( ) = ~E. Proof. By induction on n − i. For each i, the claim is proved by induction on the structure of formulas ϕi . The last claim follows from the fact that ψ1 = toHFL(E, X1 ) and: g1 ( ) = α1 y. f1,1 (y) = α1 y.~ϑ1 ϕ1 ([X1 7→ y]) = ~α1 X1 .ϑ1 ϕ1  = ~ψ1 . ⊔ ⊓ We first prove the following lemma, which implies that the game-based characterization is complete, i.e., that p ⊑ ~EL only if Proponent wins GL,E on (p, X1 ). Lemma 20. Let (d1 , . . . , dn ) be an opponent node. If d j ⊑ g j (d1 , . . . , d j−1 ) for every j ∈ {1, . . . , n}, then Proponent wins GL,E on (d1 , . . . , dn ). Proof. Let (⋆) be the following condition on O-nodes (d1 , . . . , dn ): d j ⊑ g j (d1 , . . . , d j−1 ) for every j ∈ {1, . . . , n}. Let V0 = {(d1 , . . . , dn ) ∈ VO | (d1 , . . . , dn ) satisfies (⋆)}. Let us first define a strategy S of GL,E on V0 . It is defined for plays of the form e v · (d1 , . . . , dn ) · (p, Xi ) where (d1 , . . . , dn ) ∈ V0 . The next node is determined by the last two nodes as follows. 46 – Case αi = ν: Then the next node is (d1 , . . . , di−1 , ci , ci+1 , . . . , cn ) where c j = g j (d1 , . . . , di−1 , ci , ci+1 , . . . , c j−1 ) for j ≥ i. By the assumption di ⊑ gi (d1 , . . . , di−1 ) and Lemma 19, p ⊑ di ⊑ gi (d1 , . . . , di−1 ) = νy. fi,i (d1 , . . . , di−1 , y) = fi,i (d1 , . . . , di−1 , νy. fi,i (d1 , . . . , di−1 , y)) = fi,i (d1 , . . . , di−1 , gi (d1 , . . . , di−1 )) = fi,i (d1 , . . . , di−1 , ci ) = fi,i+1 (d1 , . . . , di−1 , ci , gi+1 (d1 , . . . , di−1 , ci )) = fi,i+1 (d1 , . . . , di−1 , ci , ci+1 ) .. . = fi,n (d1 , . . . , di−1 , ci , ci+1 , . . . , cn ) = ~ϕi ([X1 7→ d1 , . . . , Xi−1 7→ di−1 , Xi 7→ ci , . . . , Xn 7→ cn ]). – Case αi = µ: Then the next node is (d1 , . . . , di−1 , e, ci+1 , . . . , cn ) where c j = g j (d1 , . . . , di−1 , e, ci+1 , . . . , c j−1 ) for j > i and e is defined as follows. By the condition (⋆) and Lemma 19, di ⊑ gi (d1 , . . . , di−1 ) = µy. fi,i (d1 , . . . , di−1 , y) and thus di ⊑ (λy. fi,i (d1 , . . . , di−1 , y))β (⊥) for some ordinal β. By the definition of edges of GL,E , we have p ⊑ di and thus p ⊑ (λy. fi,i (d1 , . . . , di−1 , y))β (⊥) for some β. Consider the minimum ordinal among those satisfying this condition; let β0 be this ordinal. Since p is complete prime, β0 is a successor ordinal by Lemma 17, i.e. β0 = β′0 + 1 for some β′0 . Then we define e = ′ (λy. fi,i (x1 , . . . , xi−1 , y))β (⊥). Now we have p ⊑ (λy. fi,i (d1 , . . . , di−1 , y))(e) = fi,i (d1 , . . . , di−1 , e) = fi,i+1 (d1 , . . . , di−1 , e, gi+1 (d1 , . . . , di−1 , e)) = fi,i+1 (d1 , . . . , di−1 , e, ci+1 ) .. . = fi,n (d1 , . . . , di−1 , e, ci+1 , . . . , cn ) = ~ϕi ([X1 7→ d1 , . . . , Xi−1 7→ di−1 , Xi 7→ e, Xi+1 7→ ci+1 , . . . , Xn 7→ cn ]). 47 It is not difficult to see that S is indeed a strategy on V0 . We prove that S is winning by giving a parity progress measure of S on V0 . Let γ be an ordinal greater than the cardinality of DL,τℓ for every ℓ ∈ {1, . . . , n}. The (partial) mapping VO ∋ (d1 , . . . , dn ) 7→ (β1 , . . . , βn ) ∈ γn is defined by:    min{β < γ | di ≤ (λy. fi,i (d1 , . . . , di−1 , y))β (⊥)} (if αi = µ) βi :=   0 (if αi = ν). This is well-defined on V0 . It is not difficult to see that this mapping is a parity progress measure of the strategy S. ⊔ ⊓ Next, we prepare lemmas used for proving that the game-based characterization is sound, i.e., that existence of a winning strategy of GL,E on (p, X1 ) implies p ⊑ ~EL . We first prove this result for a specific class of strategies, which we call stable strategies. Then we show that every winning strategy can be transformed to a stable one. Definition 16 (Stable strategy). A strategy S of GL,E is stable if for every play e v · (d1 , . . . , dn ) · (p, Xi ) · (c1 , . . . , cn ) that conforms with S, one has d1 = c1 , . . . , di−1 = ci−1 and di ⊒ ci . An important property of a stable winning strategy is as follows. 1 ,...,dℓ for the subgame of Lemma 21. Given d1 ∈ DL,τ1 , . . . , dℓ ∈ DL,τℓ , we write GdL,E GL,E consisting of nodes { (p, Xi ) ∈ VP | ℓ < i } ∪ { (d1 , . . . , dℓ , cℓ+1 , . . . , cn ) | cℓ+1 ∈ DL,τℓ+1 , . . . , cn = DL,τn }. If a stable strategy S wins on an opponent node (d1 , . . . , dn ), then its restriction is a 1 ,...,dℓ winning strategy of the subgame GdL,E on (d1 , . . . , dn ). 1 ,...,dℓ . Since Proof. It suffices to show that S does not get stuck in the subgame GdL,E proponent nodes are restricted to (p, Xi ) with ℓ < i, by the definition of stability, S does not change the first ℓ components. ⊔ ⊓ Lemma 22. Let S be a stable winning strategy of GL,E on V0 ⊆ V. For every 1 ≤ j ≤ n, we have the following propositions: – (P j ): For every 1 ≤ i ≤ j, if ṽ · (p, Xi ) · (d1 , . . . , dn ) conforms with S and starts from V0 , then p ⊑ fi, j (d1 , . . . , d j ). – (Q j ): If e v · (p, X j ) · (d1 , . . . , dn ) conforms with S and starts from V0 , then p ⊑ g j (d1 , . . . , d j−1 ). Proof. By induction on the order Pn < Qn < Pn−1 < Qn−1 < · · · < P1 < Q1 . Let S be a stable winning strategy of GL,E on V0 ⊆ V. (Pn ) By the definition of the game GL,E and Lemma 19, we have p ⊑ ~ϕi ([X1 7→ d1 , . . . , Xn 7→ dn ]) = fi,n (d1 , . . . , dn ). 48 (P j with j < n) We first show that d j+1 ⊑ g j+1 (d1 , . . . , d j ). Since DL,τ j+1 is complete-prime algebraic, it suffices to show q ⊑ g j+1 (d1 , . . . , d j ) for every complete prime q ⊑ d j+1 . Given a complete prime q ⊑ d j+1 , ṽ · (p, Xi ) · (d1 , . . . , dn ) · (q, X j+1 ) is a valid play that conforms with S. Let (c1 , . . . , cn ) = S( ṽ · (p, Xi ) · (d1 , . . . , dn ) · (q, X j+1 ) ). Since S is stable, d1 = c1 , . . . , d j = c j . By the induction hypothesis, q ⊑ gi+1 (c1 , . . . , c j ) = gi+1 (d1 , . . . , d j ). Since q ⊑ d j+1 is an arbitrary complete prime, we conclude that d j+1 ⊑ g j+1 (d1 , . . . , d j ). Then by Lemma 19, we have p ⊑ fi, j+1 (d1 , . . . , d j , d j+1 ) ⊑ fi, j+1 (d1 , . . . , d j , g j+1 (d1 , . . . , d j )) = fi, j (d1 , . . . , d j ), where the first inequality follows from the induction hypothesis and the second from d j+1 ⊑ g j+1 (d1 , . . . , d j ) and monotonicity of fi, j+1 . (Q j with α j = ν) For every complete prime q ⊑ d j , we have a play ṽ · (p, Xi ) · (d1 , . . . , dn ) · (q, X j ) that conforms with S. Let (cq,1 , . . . , cq,n ) = S( ṽ · (p, Xi ) · (d1 , . . . , dn ) · (q, X j ) ). Since S is stable, d1 = cq,1 , . . . , d j−1 = cq, j−1 and d j ⊒ cq, j . By the induction hypothesis and monotonicity of f j, j , we have q ⊑ f j, j (cq,1 , . . . , cq, j−1 , cq, j ) ⊑ f j, j (d1 , . . . , d j−1 , d j ). Since the complete prime q ⊑ d j is arbitrary and DL,E is complete-prime algebraic, we have G q ⊑ f j, j (d1 , . . . , d j−1 , d j ). dj = q⊑d j Hence d j is a prefixed point of λy. f j, j (d1 , . . . , d j−1 , y). So d j ⊑ νy. f j, j (d1 , . . . , d j−1 , y). By the induction hypothesis and Lemma 19, we have p ⊑ f j, j (d1 , . . . , d j−1 , d j ) ⊑ f j, j (d1 , . . . , d j−1 , νy. f j, j (d1 , . . . , d j−1 , y)) = νy. f j, j (d1 , . . . , d j−1 , y) = g j (d1 , . . . , d j ). (Q j with α j = µ) Let ṽ · (p, X j) · (d1, . . . , dn ) be a play that conforms with S and starts from V0 . Given opponent nodes O1 , O2 ∈ VO , we write O1 ≻ O2 if there exists a play of the form e v · (p, X j ) · e v1 · O1 · e v2 · O2 49 that conforms with S such that every proponent node in e v1 or e v2 is of the form (q, X j ). Every play of the form ṽ · (p, X j ) · (d1 , . . . , dn ) · (p1 , X j ) · (c1,1 , . . . , c1,n ) · (p2 , X j ) · (c2,1 , . . . , c2,n ) . . . that conforms with S eventually terminates because S is a winning strategy and the priority of (pi , X j ) is odd (recall that opponent node Oi has priority 0). Hence the relation ≻ defined above is well-founded. We prove the following claim by induction on the well-founded relation ≻: Let ṽ · (p, X j) · (d1 , . . . , dn ) · ṽ1 · (c1 , . . . , cn ) be a play that conforms with S and starts from V0 . Suppose that every proponent node in ṽ1 is of the form (q, X j ). Then c j ⊑ µy. f j, j (d1 , . . . , d j−1 , y). It suffices to show that q ⊑ µy. f j, j (d1 , . . . , d j−1 , y) for every complete prime q ⊑ c j . Consider the play ṽ · (p, X j ) · (d1 , . . . , dn ) · ṽ1 · (c1 , . . . , cn ) · (q, X j ) · (c′1 , . . . , c′n ) that conforms with S and starts from V0 . Because S is stable and ṽ1 does not contain a node of the form (r, Xk ) with k < j, we have d1 = c1 = c′1 , . . . , d j−1 = c j−1 = c′j−1 . By the induction hypothesis of the above claim, c′j ⊑ µy. f j, j (d1 , . . . , d j−1 , y). By the induction hypothesis of the lemma, q ⊑ f j, j (c′1 , . . . , c′j−1 , c′j ). Hence q ⊑ f j, j (c′1 , . . . , c′j−1 , c′j ) ⊑ f j, j (c′1 , . . . , c′j−1 , µy. f j, j (d1 , . . . , d j−1 , y)) ⊑ f j, j (d1 , . . . , d j−1 , µy. f j, j (d1 , . . . , d j−1 , y)) = µy. f j, j (d1 , . . . , d j−1 , y). So q ⊑ µy. f j, j (d1 , . . . , d j−1 , y) for every complete prime q ⊑ c j and thus c j ⊑ µy. f j, j (d1 , . . . , d j−1 , y)). By the same argument, we have d j ⊑ µy. f j, j (d1 , . . . , d j−1 , y). This completes the proof of the above claim. By the induction hypothesis, the above claim and Lemma 19, we have p ⊑ f j, j (d1 , . . . , d j−1 , d j ) (induction hypothesis) ⊑ f j, j (d1 , . . . , d j−1 , µy. f j, j (d1 , . . . , d j−1 , y)) ⊑ µy. f j, j (d1 , . . . , d j−1 , y) = g j (d1 , . . . , d j−1 ) (the above claim) (Lemma 19). ⊔ ⊓ Lemma 23. If there exists a winning strategy of GL,E on an Opponent node (d1, . . . , dn ), there exists a stable winning strategy on a node (d′1 , . . . , d′n ) with d1 ⊑ d′1 , . . . , dn ⊑ d′n . Proof. Let S be a winning strategy on the node (d1 , . . . , dn ). Since GL,E is a parity game, we can assume without loss of generality that S is memoryless. Given 50 nodes v and v′ of the game and j ∈ { 1, . . . , n }, we say v′ is reachable from v following S if there exists a play vv1 . . . vℓ v′ that conforms with S. We say that v′ is j-reachable from v following S if furthermore, for every proponent node (p, Xi ) in v1 . . . vk , we have j < i. For every k = 1, . . . , n, let G d′k = dk ⊔ { ek | (e1 , . . . , en ) is reachable from (d1 , . . . , dn ) following S }. Obviously di ⊑ d′i . Let S′ be a partial function V ∗ VP ⇀ V defined by where c′k = G e v · (c1 , . . . , cn ) · (q, X j ) 7→ (c1 , . . . , c j−1 , c′j , . . . , c′n ) { ek | (e1 , . . . , en ) is j-reachable from (q, X j ) following S }. We show that S′ is indeed a strategy on (d′1 , . . . , d′n ). Let v0 = (d′1 , . . . , d′n ) and assume that v0 v1 . . . vℓ (ℓ > 0) is a play that conforms with S′ and ends with an opponent node vℓ ∈ VO (hence ℓ > 0 is even). Let vℓ = (c1 , . . . , cn ). We first prove that for every k = 1, . . . n, G ck ⊒ { ek | (e1 , . . . , en ) is k-reachable from vℓ−1 following S } by induction on ℓ > 0. Let vℓ−1 = (p, X j ). – If k ≥ j, then by the definition of S′ , G ck = { ek | (e1 , . . . , en ) is j-reachable from (p, X j ) following S } G ⊒ { ek | (e1 , . . . , en ) is k-reachable from (p, X j ) following S } because the k-reachable set is a subset of the j-reachable set. – If k < j, then ck is the k-th component of vℓ−2 by the definition of S′ . • If ℓ = 2, then G { ek | (e1 , . . . , en ) is reachable from (d1 , . . . , dn ) following S } ck = d′k = dk ⊔ G ⊒ { ek | (e1 , . . . , en ) is k-reachable from v1 following S } since v1 is reachable from (d1 , . . . , dn ) following S. • Assume that ℓ > 2. By the induction hypothesis for the subsequence v0 v1 . . . vℓ−2 , we have G ck ⊒ { ek | (e1 , . . . , en ) is k-reachable from vℓ−3 following S } G ⊒ { ek | (e1 , . . . , en ) is k-reachable from vℓ−1 following S } since vℓ−1 is k-reachable from vℓ−3 following S. 51 Let v0 · v1 · · · · · vℓ · (p, Xi ) be a play that conforms with S′ and starts from v0 = (d′1 , . . . , d′n ). Let (c1 , . . . , cn ) = S′ (v0 · v1 · · · · · vℓ · (p, Xi )) and (e1 , . . . , en ) = S( (p, Xi ) ). Since (e1 , . . . , en ) is reachable from (p, Xi ) following S, the above claim shows that em ⊑ cm for every 1 ≤ m ≤ n. Since p ⊑ ~ϕi ([X1 7→ e1 , . . . , Xn 7→ en ]), we have p ⊑ ~ϕi ([X1 7→ c1 , . . . , Xn 7→ cn ]). Hence S′ is indeed a strategy. By definition and the above claim, it is easy to see that S′ is a stable strategy. We show that the strategy S′ is winning on (d′1 , . . . , d′n ). Assume that v0 v1 . . . be an infinite play that conforms with S′ and starts from (d′1 , . . . , d′n ). Let j be the minimum index of variables appearing infinitely often in the play. Then the play can be split into e0 · (p1 , X j ) · w e1 · (c1,1 , . . . , c1,n ) · (p2 , X j ) · w e2 · (c2,1 , . . . , c2,n ) · (p3 , X j ) · . . . w where w̃k is a (possible empty) sequence of nodes, which furthermore consists of VO ∪ { (q, Xℓ ) | j < ℓ } if k ≥ 1. For every m = 1, 2, . . ., we have cm, j = G { e j | (e1 , . . . , en ) is j-reachable from (pm , X j ) following S } em . By the defsince j-th component of opponent nodes is unchanged during w inition of edges of the game, pm+1 ⊑ cm, j . Since pm+1 is a complete prime, there exists (e1 , . . . , en ) that is j-reachable from (pm , X j ) following S and such that e′m pm+1 ⊑ e j . By definition of j-reachability, there exists a sequence of nodes w ′ ′ e e such that (pm , X j ) · wm · (pm+1 , X j ) conforms with S. Furthermore wm consists of VO ∪ { (q, Xℓ ) | j < ℓ } if k ≥ 1. It is easy to see that (p1 , X j ) is reachable from v′0 = (d1 , . . . , dn ) following S. Hence we have a sequence e′0 · (p1 , X j ) · w e′1 · (p2 , X j ) · w e′2 · . . . v′0 · w that conforms with S and starts from v′0 = (d1 , . . . , dn ). Since S is winning on v′0 , this sequence satisfies the parity condition, i.e. the priority of X j is even. Hence the original play, which conforms with S′ , also satisfies the parity condition. ⊓ ⊔ We are now ready to prove soundness and completeness of the type-based characterization. Theorem 7. p ⊑ ~EL if and only if Proponent wins GL,E on (p, X1 ). Proof. (⇒) By Lemma 20, Proponent wins GL,E on (p, ⊥, . . . , ⊥). Since (p, ⊥, . . . , ⊥)· (p, X1 ) is a valid play, Proponent also wins on (p, X1 ). (⇐) If Proponent wins on (p, X1), then Proponent also wins on (q, X1 ) for every q ⊑ p. Hence Proponent wins on (p, ⊥, . . . , ⊥). By Lemma 23, we have a stable winning strategy S of GL,E on (d1 , . . . , dn ) with p ⊑ d1 . Then (d1 , . . . , dn ) · (p, X1 ) is a play that conforms with S and starts from (d1 , . . . , dn ). Hence, by Lemma 22, p ⊑ g1 (). By Lemma 19, we have g1 () = ~EL . Therefore, we have p ⊑ ~EL as required. ⊔ ⊓ 52 D.4 The Opposite Parity Game This subsection defines a game in which Proponent tries to disprove p ⊑ ~EL by giving an upper-bound ~EL ⊑ q. A similar construction can be found in Salvati and Walukiewicz [37]; our construction can be seen as an infinite variant of them. We first define the notion of complete coprimes, which is the dual of complete primes. Definition 17 (Complete coprime). Let (A, ≤) be a complete lattice.VAn element p ∈ A is a complete coprime if (1) p , ⊤ and (2) for every U ⊆ A, if ( x∈U x) ≤ p, then x ≤ p for some x ∈ U. Lemma 24. If (A, ≤) is complete-prime algebraic, then for every x ∈ A, V x = { p | x ≤ p, p: complete coprime }. V Proof. Obviously x ≤ { p | x ≤ p, p: complete coprime }. We show that x ≥ V { p | x ≤ p, p: complete coprime }. Since (A, ≤) is complete-prime algebraic, by Lemma 16, it suffices to show that, for every complete prime q such that q  x, V we have q  { p | x ≤ p, p: complete coprime }, i.e., there exists a complete coprime p ≥ x such that q  p. Let q be a complete prime such that q  x. We show that W p = { q′ | q  q′ , q′ : complete prime } is indeed a complete coprime p ≥ x such W that q  p. We first check q  p. If it were the case that q ≤ p, then q ≤ p = { q′ | q  q′ , q′ : complete prime }, which would imply q ≤ q′ for some q′ with q  q′ (as q is a complete prime), a contradiction. It remains toV check that p is a complete coprime. Since q  p, we have p , ⊤. Suppose that y∈U y ≤ p. If q ≤ y for every y ∈ U, then q ≤ p, a contradiction. Hence q  y for some y ∈ U. Now W y = { q′′ | q′′ ≤ y, q′′ : complete prime } W ′′ ≤ { q | q  q′′ , q′′ : complete prime } = p. Here the second inequality comes from the fact that q′′ ≤ y implies q  q′′ ; in fact, if q′′ ≤ y and q ≤ q′′ , then q ≤ q′′ ≤ y, hence a contradiction. ⊔ ⊓ Definition 18 (Opposite parity game). The parity game GL,E is defined by the following data: VP := { (p, Xi) | p ∈ DL,τi , p: complete coprime } VO := { (d1 , . . . , dn ) | d1 ∈ DL,τ1 , . . . , dn ∈ DL,τn } E := { ((p, Xi), (d1 , . . . , dn )) | p ⊒ ~ϕi ([X1 7→ d1 , . . . , Xn 7→ dn ]) } ∪ { ((d1, . . . , dn ), (p, Xi)) | p ⊒ di , p: complete coprime }. The priority of the opponent node is 0; the priority of node (p, Xi ) is 2(n − i) + 1 if αi = ν and 2(n − i) + 2 if αi = µ. 53 The game GL,E is obtained by replacing complete primes with complete coprimes and ⊑ with ⊒. The position (p, Xi ) is the state where Proponent ties to show that ~(E, Xi ) ⊑ p; to this end, Proponent picks a valuation [X1 7→ d1 , . . . , Xn 7→ dn ] such that (p ⊒ ~ϕi ([X1 7→ d1 , . . . , Xn 7→ dn ]). Opponent challenges the validity of the valuation [X1 7→ d1 , . . . , Xn 7→ dn ], by picking i and a complete coprime p′ such that p′ ⊒ di , and asking why ~(E, Xi ) ⊑ p′ holds. Lemma 25. If Proponent wins GL,E on (p, Xi ) (where p is a complete coprime), then for every complete prime p′ ∈ DL,τi such that p′ @ p, Opponent wins GL,E on (p′ , Xi ). Proof. Let S be a P-winning strategy of GL,E on (p, Xi ) (where p is a complete coprime). We can assume without loss of generality that S is memoryless. Let p′ ∈ DL,τi be a complete prime such that p′ @ p. Let (q, X j ) be a Proponent node in GL,E and (q′ , X j′ ) be a Proponent node in GL,E . We say that (q, X j ) is inconsistent with (q′ , X j′ ) if j = j′ and q′ @ q. Similarly, given Opponent nodes (d1 , . . . , dn ) in GL,E and (d′1 , . . . , d′n ) in GL,E , those nodes are inconsistent if d′j @ d j for some 1 ≤ j ≤ n. Let (p′ , Xi ) · v′1 · v′2 · · · · · v′k be a play of GL,E . This play is admissible (with respect to S) if there exists a play (p, Xi ) · v1 · v2 · · · · · vk of GL,E that conforms with S such that v j is inconsistent with v′j for every j ∈ {1, . . . , k}. For each admissible play (p′ , Xi ) · v′1 · v′2 · · · · · v′k of GL,E , we choose such a play (p, Xi ) · v1 · v2 · · · · · vk of GL,E . Suppose that our choice satisfies the following property: If (p, Xi )·v1 ·v2 ·· · ··vk is the chosen play corresponding to (p′ , Xi )·v′1 ·v′2 ·· · ·· v′k , then (p, Xi ) ·v1 ·v2 ·· · ··vk−1 is the chosen play for (p′ , Xi ) ·v′1 ·v′2 ·· · ··v′k−1 . (A way to achieve this is to introduce a well-ordering on nodes of GL,E and to choose the minimum play w.r.t. the lexicographic ordering from those that satisfy the requirement.) We define an Opponent strategy S for GL,E defined on admissible plays starting from (p′ , Xi ). Let (p′ , Xi ) · v′1 · v′2 · · · · · v′k · (d′1 , . . . , d′n ) be an admissible play of GL,E . Let (p, Xi ) · v1 · v2 · · · · · vk · (d1 , . . . , dn ) be the chosen play of GL,E corresponding to the above play. So v j and v′j are inconsistent for every j ∈ {1, . . . , k} and (d1 , . . . , dn ) is inconsistent with (d′1 , . . . , d′n ). By definition, d′ℓ @ dℓ for some 1 ≤ ℓ ≤ n. Since DL,τℓ is complete-prime algebraic, by Lemma 24, F d′ℓ = { q′ ∈ DL,τℓ | q′ ⊑ d′ℓ , q′ : complete prime } and dℓ =  { q ∈ DL,τℓ | dℓ ⊑ q, q: complete coprime }. 54 Hence d′ℓ @ dℓ implies that there exist a complete prime q′ ⊑ d′ℓ and a complete coprime dℓ ⊑ q such that q′ @ q. The Opponent strategy S chooses (q′ , Xℓ ) as the next node in GL,E . We show that S is indeed an O-strategy of GL,E on (p′ , Xi ). It suffices to show that every play (p′ , Xi ) · v′1 · · · · · v′k that conforms with the Opponent strategy S is admissible. If k = 0, the play is obviously admissible. Let (p′ , Xi ) · v′1 · · · · · v′k be a play that conforms with S and assume that it is admissible. If v′k is an Opponent node, the next node v′k+1 is determined by S and then (p′ , Xi ) · v′1 · · · · · v′k · v′k+1 is admissible by the definition of S. Consider the case that v′k = (p′k , Xik ) is an Proponent node. Since this play is admissible, we have the chosen play (p, Xi ) · v1 · v2 · · · · · vk of GL,E corresponding to the above play. Then v j and v′j are inconsistent for every j = 1, . . . , k. Since vk is inconsistent with v′k , vk = (pk , Xik ) with p′k @ pk . Let v′k+1 = (d′1 , . . . , d′n ) be an arbitrary Opponent node in GL,E such that (vk , vk+1 ) is valid in GL,E. Let vk+1 = (d1 , . . . , dn ) be the Opponent node in GL,E determined by S. By definition of edges, p′k ⊑ ~ϕi ([X1 7→ d′1 , . . . , Xn 7→ d′n ]) and ~ϕi ([X1 7→ d1 , . . . , Xn 7→ dn ]) ⊑ pk . Since ~ϕi  is monotone, d′j ⊑ d j for every j = 1, . . . , n would imply p′k ⊑ pk , a contradiction. Hence d′j @ d j for some 1 ≤ j ≤ n, that means, (d′1 , . . . , d′n ) is inconsistent with (d1 , . . . , dn ). Therefore the play (p, Xi ) · v1 · v2 · · · · · vk · (d′1 , . . . , d′n ) is admissible for every choice of (d′1 , . . . , d′n ). So S is defined for every play that conforms with S and starts from (p′ , Xi ). We show that S is O-winning. As S does not get stuck, it suffices to show that every infinite play following S does not satisfy the parity condition. Assume an infinite play (p′ , Xi ) · v′1 · (p′1 , Xi′1 ) · v′2 · (p′2 , Xi′2 ) · · · · · v′ℓ · (p′ℓ , Xi′ℓ ) · . . . following S. By the definition of S, every finite prefix of the play is admissible. Let (p′ , Xi ) · vk,1 · · · · · vk,k be the chosen play for the prefix of the above play of length k. Then, by the assumption on the choice, the choice of each node does not depend on the length of the prefix, i.e. vk,i = vk′ ,i for every i ≥ 1 and k, k′ ≥ i. Hence the above infinite play is “admissible” in the sense that there exists an infinite play in GL,E (p, Xi ) · v1 · (p1 , Xi1 ) · v2 · (p2 , Xi2 ) · · · · · vℓ · (pℓ , Xiℓ ) · . . . that conforms with S and such that (pℓ , Xiℓ ) and (p′ℓ , Xi′ℓ ) are inconsistent for every ℓ. Hence iℓ = i′ℓ for every ℓ. Because (priority of (pℓ , Xiℓ ) in GL,E ) = (priority of (p′ℓ , Xiℓ ) in GL,E ) + 1, the latter play satisfies the parity condition if and only if the former play does not satisfy the parity condition. Since S is a winning strategy, we conclude that the former play does not satisfy the parity condition. ⊔ ⊓ 55 E Proof for Section 7 E.1 Some definitions Here we introduce some notions and notations, which are useful in proofs of results of Section 7. Definition 19 (Choice sequences). Let π be an infinite sequence consisting of L and R, called a choice sequence. We define a reduction relation for pairs of terms and choice sequences by ǫ ǫ (t1 t2 ; Lπ) −→D (t1 ; π) (t1 t2 ; Rπ) −→D (t2 ; π) ℓ ℓ and (u; π) −→D (u′ ; π) if u , u1 u2 and u −→D u′ . We sometimes omit labels ℓ and just write (u; π)−→D (u′ ; π) for (u; π) −→D (u′ ; π). A choice sequence resolves nondeterminism of a program P; a reduction sequence is completely determined by its initial term, its length and a choice sequence. We write −→kD for the k-step reduction relation. If (t0 , π0 ) −→D (t1 , π1 ) −→D · · · −→D (tk , πk ), then we write π0 t0 −→D t1 −→D · · · −→D tk and say that the reduction sequence t0 −→D t1 −→D · · · −→D tk follows π0 . We introduce a slightly more elaborate notion of the recursive call relation below. Note that the relation f e t Dg e u in Definition 6 coincides with f e t Dg e u redefined below. Definition 20 (Recursive call relation, call sequences). Let P = (D, main) be ♯ xi = ui }1≤i≤n . We define D♯ := D ∪ { fi x̃ = ui }1≤i≤n where a program, with D = { fi e f1♯ , . . . , fn♯ are fresh symbols. So D♯ has two copies of each function symbol, one of which is marked. For the terms e ti and e t j that do not contain marked symbols, we write k+1 ′ t j ; π ) if ti ; π) D ( f j e ( fi e ♯ ♯ ′ ♯ ([e ti /e xi ][ f1 / f1 , . . . , fn / fn ]ui ; π) −→kD♯ ( f j e t j ; π′ ) (then e t j is obtained by erasing all the marks in e t′j ). If there exists a (finite or infinite) sequence ( f s̃; π) k2 k1 D (g1 ũ1 ; π1 ) k2 D (g2 ũ2 ; π2 ) k3 D · · ·, then we write π f s̃ k1 D k3 g1 ũ1 D g2 ũ2 D · · · and call it a call sequence of f s̃ following π. We often omit the number k j of steps and the choice sequence π. Given a program P = (D, main), the set of call sequences of P is a subset of (finite or infinite) sequences of function symbols u2 D · · · }. u1 D g2 e defined by Callseq(P) := { main g1 g2 · · · | main D g1 e k ti ; π) −→kD t j ; π′ ) implies ( fi e ti ; π) Note that by the definition above, ( fi e D ( fje t j ; π′ ). We write t♯ for the term obtained by marking all function symbols in ( fje t, i.e. [ f1♯ / f1 , . . . , fn♯ / fn ]t. 56 E.2 Existence of a unique infinite call-sequence Let P = (D, main) be a program and ℓ1 π ℓ2 main −→D t1 −→D · · · be the infinite reduction sequence following π. This subsection proves that there exists a unique infinite call-sequence π f ũ D g1 ũ1 D g2 ũ2 D ··· of this reduction sequence. Lemma 26. Let P = (D, main) be a program and π ∈ {L, R}ω be a choice sequence. Suppose we have an infinite reduction sequence π main −→D t1 −→D t2 −→D . . . . Then there exists an infinite call-sequence π main D g1 ũ1 D g2 ũ2 D ··· . Proof. Kobayashi and Ong [24, Appendix B] have proved a similar result for simply-typed programs without integers. We remove integers from D as follows: – First, replace each conditional branching if p(t′1 , . . . , t′k ) then t1 else t2 with the nondeterministic branching t1 t2 . – Then replace each integer expression with the unit value. Let us write D′ for the resulting function definitions. Then there exists a choice π′ such that π′ main −→D′ t′1 −→D′ t′2 −→D′ · · · and, for every i ≥ 1, t′i is obtained from ti by the above translation. By the result of Kobayashi and Ong, this sequence has an infinite call-sequence. This call-sequence can be transformed into a call-sequence of the original reduction sequence. ⊔ ⊓ We prove uniqueness. v; π′ ). Given an occurrence of g in f e u, we Suppose that ( f e u; π) −→kD (g e ′ e′ write f u for the term obtained by replacing the occurrence of g with g♯ . We say that the head-occurrence g in g e v is a copy of the occurrence of g in f e u if (f′ e u′ ; π) −→kD♯ (g♯ e v′ ; π). An occurrence of h in f e u is an ancestor of the headoccurrence of g in g e v if e1 ; π1 ) (f e u; π) −→kD0 (h1 w k1 D e2 ; π2 ) (h2 w k2 D ... kℓ D eℓ ; πℓ ) = (g e v; π′ ) (hℓ w P where ℓ ≥ 1, k = ℓj=0 k j and h1 is a copy of the occurrence of h in f e u (hence h1 = h as function symbols). It is not difficult to see that, given a reduction sequence (f e u; π) −→kD (g e v; π′ ), there exists a unique ancestor of g in f e u. 57 Lemma 27. For every term t and choice sequence π ∈ {L, R}ω, there exists at most one infinite call-sequence of t following π. Proof. Let π t0 −→D t1 −→D t2 −→D · · · be the infinite reduction sequence following π. In this proof, we shall consider only sequences following π and hence we shall omit π . Assume that there exist difference infinite call-sequences, say, t0 D g1 ũ1 D ··· D gk ũk D h1 ṽ1 D h2 ṽ2 D ··· t0 D g1 ũ1 D ··· D gk ũk D k1 w̃1 D k2 w̃2 D ··· and where h1 and k1 are different. Since both call sequences are infinite, the reduction sequence has infinite switching of the sequence, i.e. the reduction sequence must be of the following form: t0 −→∗D hi1 ṽi1 −→∗D k j1 w̃ j1 −→∗D hi2 ṽi2 −→∗D k j2 w̃ j2 −→∗D · · · . Since hi1 is not an ancestor of k j1 , an ancestor of k j1 must appear in ṽi1 , say, in the ℓ1 -th argument vi1 ,ℓ1 . Since k j1 appears in the reduction sequence, vi1 ,ℓ1 must have a head occurrence and the reduction sequence is of the form t0 −→∗D hi1 ṽi1 −→∗D vi1 ,ℓ1 s̃1 −→∗D k j1 w̃ j1 −→∗D hi2 ṽi2 −→∗D · · · . Since hi1 is an ancestor of hi2 , there is no ancestor of hi2 in ṽi1 , in particular, in vi1 ,ℓ1 . So an ancestor of hi2 must appear in s̃1 , say, s1,ℓ2 , and the reduction sequence is of the form t0 −→∗D hi1 ṽi1 −→∗D vi1 ,ℓ1 s̃1 −→∗D s1,ℓ2 s̃2 −→∗D hi2 ṽi2 −→∗D k j2 w̃ j2 −→∗D · · · . Since an ancestor of k j2 appears in vi1 ,ℓ1 , there is no ancestor in s̃1 , in particular, in s1,ℓ2 . Hence there is an ancestor of k j2 in s̃2 , say, s2,ℓ3 , and the reduction sequence is t0 −→∗D hi1 ṽi1 −→∗D vi1 ,ℓ1 s̃1 −→∗D s1,ℓ2 s̃2 −→∗D s2,ℓ3 s̃3 −→∗D k j2 w̃ j2 −→∗D hi3 ṽi3 −→∗D · · · . By repeatedly applying the above argument, we have the following decomposition of the reduction sequence: t0 −→∗D hi1 ṽi1 −→∗D vi1 ,ℓ1 s̃1 −→∗D s1,ℓ2 s̃2 −→∗D s2,ℓ3 s̃3 −→∗D s3,ℓ4 s̃4 −→∗D s4,ℓ5 s̃5 −→∗D · · · . Then the order of sm,ℓm is greater than sm+1,ℓm+1 for every m since sm+1,ℓm+1 is an argument of sm,ℓm . This contradicts the assumption that the program D is simply typed. ⊔ ⊓ Corollary 1. Let P = (D, main) be a program and π main −→D t1 −→D · · · be the infinite reduction sequence following π. Then there exists a unique infinite call-sequence of this reduction sequence, i.e. π main D g1 ũ1 58 D g2 ũ2 D ··· . E.3 Proof of Theorem 4 Let L0 = ({s⋆ }, ∅, ∅, s⋆) be the transition system with one state and no transition. This section shows that given a program P = (D, main) with priority Ω, |=csa (P, Ω) holds if and only if L0 |= Φ(P,Ω),csa . Without loss of generality, we assume below that P is of the form (D, main); note that, if P = (D, t), then we can replace it with (D ∪ {main = t}, main). Fix a program P = (D, main) with priority Ω. Suppose that D = { f1 x̃1 = t1 , . . . , fn x̃n = tn } and let κi be the simple type of fi . We can assume without loss of generality that main = f1 and that Ω( fi ) ≥ Ω( f j ) if i ≤ j. Let    2(n − i) Ω′ ( fi ) =   2(n − i) + 1 (if Ω( fi ) is even) (if Ω( fi ) is odd). The priority assignments Ω and Ω′ are equivalent in the sense that an infinite sequence g1 g2 . . . of function symbols satisfies the parity condition with respect to Ω if and only if so does to Ω′ . Hence we can assume without loss of generality that 2(n − i) ≤ Ω( fi ) ≤ 2(n − i) + 1. Given a type κ (resp. η) of the target language, we define a type κ†csa (resp. η†csa ) for HFLs as follows: ⋆†csa := • (η → κ)†csa := η†csa → κ†csa int†csa := int. Simply (−)†csa replaces ⋆ with •. The operation (−)†csa is pointwise extended for type environments by: (x1 : η1 , . . . , xℓ : ηℓ )†csa = x1 : η1 †csa , . . . , xℓ : ηℓ †csa Given an LTS L, let us define DL,κ := DL,κ†csa . Definition 21 (Semantic interpretation of a program). Let K ⊢ t :κ. A valuation of K is a mapping ρ such that dom(ρ) = dom(K ) and ρ(x) ∈ DL,K (x) for each x ∈ dom(K ). The set of valuations of K is ordered by the point-wise ordering. The interpretation of a type judgment K ⊢ t : η is a (monotonic) function from the valuations of K to DL,κ inductively defined by: ~K ⊢ ( ) : ⋆(ρ) := {s⋆ } ~K ⊢ x : κ(ρ) := ρ(x) ~K ⊢ n : int(ρ) := n ~K ⊢ t1 op t2 : int(ρ) := (~K ⊢ t1 : int(ρ))~op(~K ⊢ t2 : int(ρ))   ~K ⊢ t1 : ⋆(ρ) (if (~t′ (ρ), . . . , ~t′ (ρ)) ∈ ~p)  1 k ~K ⊢ if p(t′1 , . . . , t′k ) then t1 else t2 : ⋆(ρ) :=  ~K ⊢ t : ⋆(ρ) (if (~t′ (ρ), . . . , ~t′ (ρ)) < ~p)  2 1 k ~K ⊢ event a; t : ⋆(ρ) := ~K ⊢ t : ⋆(ρ) ~K ⊢ t1 t2 : κ(ρ) := (~K ⊢ t1 : η → κ(ρ))(~K ⊢ t2 : η(ρ)) ~K ⊢ t1 t2 : ⋆(ρ) := ~K ⊢ t1 : ⋆(ρ) ⊓ ~K ⊢ t2 : ⋆(ρ) ~K ⊢ λx.t : η → κ(ρ) := {(d, ~K , x : η ⊢ t : κ(ρ[x 7→ d])) | d ∈ DL,η }. 59 We often write just ~t for ~K ⊢ t : η (on the assumption that the simple type environment for t is implicitly determined). Lemma 28. ~K ⊢ t : ηL0 ([x1 7→ d1 , . . . , xn 7→ dn ]) = ~K †csa ⊢ t†csa : η†csa L0 ([x1 7→ d1 , . . . , xn 7→ dn ]). Proof. By induction on the structure of t. ⊔ ⊓ Completeness We show that |=csa (P, Ω) implies L0 |= Φ(P,Ω),csa . We use gamebased characterization of L0 |= Φ(P,Ω),csa . Definition 22 (Parity game for a program). The parity game GL,(P,Ω) is defined as follows: VP := { (p, fi ) | p ∈ DL,κi , p: complete prime } VO := { (d1 , . . . , dn ) | d1 ∈ DL,κ1 , . . . , dn ∈ DL,κn } E := { ((p, fi ), (d1, . . . , dn )) | p ⊑ ~λx̃i .ti ([ f1 7→ d1 , . . . , fn 7→ dn ]) } ∪ { ((d1, . . . , dn ), (p, fi )) | p ⊑ di , p: complete prime }. The priority of (d1 , . . . , dn ) is 0 and that of (p, fi ) is the priority Ω( fi ) of fi . Corollary 2. GL0 ,(P,Ω) is isomorphic to GL0 ,Φ(P,Ω),csa , i.e. there exists a bijection of nodes that preserves the owner and the priority of each node. Proof. The bijection on nodes is given by (d1 , . . . , dn ) ↔ (d1 , . . . , dn ) and (p, fi ) ↔ (p, fi ). This bijection preserves edges because of Lemma 28. ⊔ ⊓ We define a strategy of GL0 ,(P,Ω) on ({s⋆ }, f1 ) by inspecting the reduction sequences of P. The technique used here is the same as [23]. In this construction, we need to track occurrences of a function symbol or a term in a reduction sequence. This can be achieved by marking symbols and terms if needed as in Definition 6. In the sequel, we shall not introduce explicit marking and use the convention that, if the same metavariable appears in a reduction sequence (e.g. f u −→∗ u e v), then the first one is the origin of all others. The occurrence of t in t u1 . . . uk (k ≥ 0) is called a head occurrence. Given a head occurrence of t of a term reachable from main (i.e. main −→∗D t ũ for some ũ), we assign an element dt,ũ ∈ DL0 ,κ where κ is the simple type of t by induction on the order of κ. – Case e u is empty: Then t : ⋆ and we define dt,ǫ = {s⋆ }. – Case t ũ = t u1 . . . uk : Note that the order of u j (1 ≤ j ≤ k) is less than that of t. For every 1 ≤ j ≤ k, we define e j as follows. • Case u j : int: Then e j = ~u j . (Note that u j has no free variable and no function symbol.) • Case u j : κ: Let e j be the element defined by G ej = { du j ,ṽ | t ũ −→∗D u j ṽ }. 60 Then dt,ũ is the function defined by    {s⋆ } (if e j ⊑ x j for all j ∈ {1, . . . , k}) dt,ũ (x1 , . . . , xk ) :=   { } (otherwise). Note that dt,ũ is the minimum function such that dt,ũ (e1 , . . . , ek ) = {s⋆ }. Let θ = [s1 /x1 , . . . , sk /xk ] and assume main −→∗D (θt0 ) ũ. We define a mapping from function symbols and variables to elements of their semantic domains as follows. – For a function symbol fi : ρt0 ,ũ,θ ( fi ) = G { d fi ,ṽ | (θt0 )e u −→∗D fi ṽ, fi originates from t0 } – For a variable xi : int: ρt0 ,ũ,θ (xi ) = ~si  – For a variable xi : κ: G ρt0 ,ũ,θ (xi ) = { dsi,ṽ | (θt0 )e u −→∗D si ṽ, si originates from θ(xi ) }. Lemma 29. Let θ = [s1 /x1 , . . . , sk /xk ] and assume main −→∗D (θt0 ) ũ. Let v be an arithmetic expression (i.e. a term of type int) appearing in t0 . Then ~θv = ~v(ρt0 ,ũ,θ ). Proof. By induction on v. – Case v = xi : Then ~θxi  = ~si  = ρt0 ,ũ,θ (xi ) = ~xi (ρt0 ,ũ,θ ). – Case v = v1 opv2 : By the induction hypothesis, ~θv j  = ~v j (ρt0 ,ũ,θ ) for j = 1, 2. Then ~θ(v1 op v2 ) = ~θv1 ~op~θv2  = ~v1 (ρt0 ,ũ,θ )~op~v2 (ρt0 ,ũ,θ ) = ~v1 op v2 (ρt0 ,ũ,θ ). ⊔ ⊓ Lemma 30. Let θ = [s1 /x1 , . . . , sk /xk ] and assume main −→∗D (θt0 ) ũ. Then dθt0 ,ũ ⊑ ~t0 (ρt0 ,ũ,θ ). Proof. By induction on the structure of t0 . – Case t0 = ( ): Then ~t0  = {s⋆ } = d( ),ǫ . – Case t0 = xi for some 1 ≤ i ≤ k: Then (θt0 )ũ = si ũ. By definition, dsi ,ũ = ρt0 ,ũ,θ (xi ). – Case t0 = fi : Then (θt0 )ũ = fi ũ. Since this occurrence of fi originates from t0 , we have d fi ,ũ = ρt0 ,ũ,θ ( fi ). – Case t0 = n or t0 = t1 op t2 : Never occurs because the type of t0 must be of the form η1 → · · · → ηℓ → ⋆. 61 – Case t0 = if p(t′1 , . . . , t′k ) then t1 else t2 : Then ũ is an empty sequence. Assume that (~θt′1 , . . . , ~θt′k ) ∈ ~p. The other case can be proved in a similar manner. Then we have θt0 −→D θt1 . By definition, dθt0 ,ǫ = dθt1 ,ǫ = {s⋆ }. Here, −→+D is the transitive closure of −→D . Since every reduction sequence θt0 −→+D s ṽ can be factored into θt0 −→D θt1 −→∗D s ṽ, we have ρt0,ǫ,θ = ρt1 ,ǫ,θ . By the induction hypothesis, {s⋆ } ⊑ ~t1 (ρt1 ,ǫ,θ ). Hence {s⋆ } ⊑ ~t1 (ρt0 ,ǫ,θ ). By Lemma 29, ~θt′j  = ~t′j (ρt0 ,ũ,θ ) for every j = 1, . . . , k. Hence (~θt′1 , . . . , ~θt′k ) ∈ ~p implies (~t′1 (ρt0 ,ũ,θ ), . . . , ~t′k (ρt0 ,ũ,θ )) ∈ ~p. By definition, ~t0 (ρt0 ,ũ,θ ) = ~t1 (ρt0 ,ũ,θ ) ⊒ {s⋆ }. – Case t0 = event a; t: Then t0 has type ⋆ and ũ is the empty sequence. We have θt0 = event a; (θt). Since every reduction sequence θt0 −→+D s ṽ can be factored into θt0 −→D θt −→∗D s ṽ, we have ρt0 ,ǫ,θ = ρt,ǫ,θ . By definition, dθt0 ,ǫ = dθt,ǫ = {s⋆ } and ~t0 (ρt0 ,ǫ,θ ) = ~t(ρt0 ,ǫ,θ ) = ~t(ρt,ǫ,θ ). By the induction hypothesis, dθt,ǫ ⊑ ~t(ρt,ǫ,θ ). Hence dθt0 ,ǫ ⊑ ~t0 (ρt0 ,ǫ,θ ). – Case t0 = t1 t2 : Suppose that (θt1 )(θt2 )ũ −→∗D (θt2 ) ṽ. We first show that ρt2 ,ṽ,θ (x) ⊑ ρt1 t2 ,ũ,θ (x) for every x ∈ { f1 , . . . , fn } ∪ {x1 , . . . , xk }. • Case x = xi : int: Then ρt2 ,ṽ,θ (xi ) = ρt1 t2 ,ũ,θ (xi ) = ~si . • Case x = xi : κ: Then G ρt2 ,ṽ,θ (xi ) = { dsi ,w̃ | (θt2 ) ṽ −→∗D si w̃, si originates from θ(xi ) } G ⊑ { dsi ,w̃ | (θ(t1 t2 )) ũ −→∗D si w̃, si originates from θ(xi ) } = ρt1 t2 ,ũ,θ (xi ) because (θt2 ) ṽ −→∗D si w̃ (where si originates from θxi ) implies (θ(t1 t2 )) ũ −→∗D (θt2 )ṽ −→∗D si w̃ (where si originates from θxi ). • Case x = fi : G ρt2 ,ṽ,θ ( fi ) = { d fi ,w̃ | (θt2 ) ṽ −→∗D fi w̃, fi originates from t2 } G ⊑ { d fi ,w̃ | (θ(t1 t2 )) ũ −→∗D fi w̃, fi originates from t1 t2 } = ρt1 t2 ,ũ,θ ( fi ) because (θt2 ) ṽ −→∗D fi w̃ (where fi originates from t2 ) implies (θ(t1 t2 )) ũ −→∗D (θt2 )ṽ −→∗D fi w̃ (where fi originates from t1 t2 ). It is easy to see that ρt1 ,(θt2 ) ũ,θ ⊑ ρt1 t2 ,ũ,θ . By the induction hypothesis, dθt1 ,(θt2 ) ũ ⊑ ~t1 (ρt1 ,(θt2 ) ũ,θ ) ⊑ ~t1 (ρt1 t2 ,ũ,θ ). By the definition of dθt1 ,(θt2 ) ũ , G  {dθt2 ,ṽ | (θt1 ) (θt2 ) ũ −→∗D (θt2 )ṽ} = d(θt1 ) (θt2 ), ũ . dθt1 ,(θt2 ) ũ By the induction hypothesis, for every reduction sequence (θt1 ) (θt2 ) ũ −→∗D (θt2 )ṽ, we have dθt2 ,ṽ ⊑ ~t2 (ρt2 ,ṽ,θ ). Since ρt2 ,ṽ,θ (x) ⊑ ρt1 t2 ,ũ,θ (x) for every x, we have dθt2 ,ṽ ⊑ ~t2 (ρt2 ,ṽ,θ ) ⊑ ~t2 (ρt1 t2 ,ũ,θ ). 62 Because the reduction sequence (θt1 ) (θt2) ũ −→∗D (θt2 )ṽ is arbitrary, G  {dθt2 ,ṽ | (θt1 ) (θt2 ) ũ −→∗D (θt2 )ṽ} ⊑ ~t2 (ρt1 t2 ,ũ,θ ). Therefore, by monotonicity, G {dθt2 ,ṽ | (θt1 ) (θt2 ) ũ −→∗D (θt2 )ṽ}   ⊑ ~t1 (ρt1 t2 ,ũ,θ ) ~t2 (ρt1 t2 ,ũ,θ ) d(θt1 ) (θt2 ), ũ = dθt1 ,(θt2 ) ũ  = ~t1 t2 (ρt1 t2 ,ũ,θ ). – Case t0 = t1 t2 : Then ũ = ǫ. We have θt0 −→D θt1 and θt0 −→D θt2 . By the definition of dθt1 ,ǫ and the induction hypothesis, {s⋆ } = dθt1 ,ǫ ⊑ ~t1 (ρt1 ,ǫ,θ ). Similarly {s⋆ } = dθt2 ,ǫ ⊑ ~t2 (ρt2 ,ǫ,θ ). Because θt0 −→D θt1 , we have ρt1 ,ǫ,θ ⊑ ρt0 ,ǫ,θ . Similarly ρt2 ,ǫ,θ ⊑ ρt0 ,ǫ,θ . By monotonicity of interpretation, {s⋆ } = dθt1 ,ǫ ⊑ ~t1 (ρt1 ,ǫ,θ ) ⊑ ~t1 (ρt0 ,ǫ,θ ) and {s⋆ } = dθt2 ,ǫ ⊑ ~t2 (ρt2 ,ǫ,θ ) ⊑ ~t2 (ρt0 ,ǫ,θ ). Hence {s⋆ } ⊑ ~t1 (ρt0 ,ǫ,θ ) ⊓ ~t2 (ρt0 ,ǫ,θ ) = ~t0 (ρt0 ,ǫ,θ ). ⊔ ⊓ The strategy SP of GL0 ,P on ({s⋆ }, main) is defined as follows. Each play in the domain of SP ({s⋆ }, main) · O1 · (p1 , g1 ) · O2 · · · · · Ok · (pk , gk ) is associated with a call-sequence main m1 D g1 ũ1 m2 D g2 ũ2 m3 D ... mk−1 D gk−1 ũk−1 mk D gk ũk such that p j ⊑ d g j ,ũ j for every j = 1, 2, . . . , k. Let ξi be the finite sequence over mi {L, R} describing the choice made during gi−1 ũi−1 D gi ũi (where g0 ũ0 = main). The canonical associated call-sequence of the play is the minimum one ordered by the lexicographic ordering on (m1 , ξ1 , m2 , ξ2 , . . . , mk , ξk ). Assume that the above call-sequence is canonical. The next step of this reduction sequence is [ũk /x̃k ]t if (gk e xk = t) ∈ D. Let ϑ = [ũk /x̃k ]. In this situation, the strategy SP chooses (ρt,ǫ,ϑ ( f1 ), . . . , ρt,ǫ,ϑ ( fn )) as the next node. This is a valid choice, i.e.: 63 Lemma 31. pk ⊑ ~D(gk )([ f1 7→ ρt,ǫ,ϑ ( f1 ), . . . , fn 7→ ρt,ǫ,ϑ ( fn )]). Proof. By Lemma 30, we have: dϑt,ǫ = {s⋆ } ⊑ ~t(ρt,ǫ,ϑ ) = (~λe xk .t([ f1 7→ ρt,ǫ,ϑ ( f1 ), . . . , fn 7→ ρt,ǫ,ϑ ( fn )]))(ρt,ǫ,ϑ(e xk )). By definition, d gk ,ũk is the least element such that d gk ,ũk (ρt,ǫ,ϑ (e xk )) = {s⋆ }. Therefore d gk ,ũk ⊑ ~λe xk .t([ f1 7→ ρt,ǫ,ϑ ( f1 ), . . . , fn 7→ ρt,ǫ,ϑ ( fn )]). Since pk ⊑ d gk ,ũk , we have pk ⊑ d gk ,ũk ⊑ ~λx̃k .t([ f1 7→ ρt,ǫ,ϑ ( f1 ), . . . , fn 7→ ρt,ǫ,ϑ ( fn )]). ⊔ ⊓ Now we have a play ({s⋆ }, main) · O1 · (p1 , g1 ) · O2 · · · · · Ok · (pk , gk ) · (ρt,ǫ,θ ( f1 ), . . . , ρt,ǫ,θ ( fn )) associated with the call-sequence main m1 D g1 ũ1 m2 D g2 ũ2 m3 D ... mk−1 D gk−1 ũk−1 mk D gk ũk Let (pk+1 , gk+1 ) be the next opponent move. By definition of the game, G pk+1 ⊑ ρt,ǫ,θ (gk+1 ) = { d gk+1 ,ṽ | θt −→∗D gk+1 ṽ, gk+1 originates from t }. This can be more simply written as pk+1 ⊑ ρt,ǫ,θ (gk+1 ) = G { d gk+1 ,ṽ | gk e uk gk+1 e v }. Since pk+1 is a complete prime, pk+1 ⊑ d gk+1 ,ṽ for some ṽ with gk e uk we have an associated call-sequence main m1 D g1 ũ1 m2 D g2 ũ2 m3 D ... mk−1 D gk−1 ũk−1 mk D gk ũk gk+1 e v. Then mk+1 D gk+1 ũk+1 of ({s⋆ }, main) · O1 · (p1 , g1 ) · O2 · · · · · Ok · (pk , gk ) · (ρt,ǫ,θ ( f1 ), . . . , ρt,ǫ,θ ( fn )) · (pk+1 , gk+1 ). Lemma 32. SP is a winning strategy of GL0 ,(P,Ω) on ({s⋆ }, main) if every infinite call-sequence of P satisfies the parity condition. Proof. The above argument shows that SP is a strategy of GL0,(P,Ω) on ({s⋆ }, main). We prove that it is winning. Assume an infinite play ({s⋆ }, main) · v1 · (p1 , g1 ) · v2 · (p2 , g2 ) · . . . 64 that conforms with SP and starts from ({s⋆ }, main). Then each odd-length prefix is associated with the canonical call-sequence πk main mk,1 D uk,1 g1 e mk,2 D uk,2 g2 e mk,3 D ... mk,k D gk e uk,k . mk,i uk,i−1 D Let ξk,i be the sequence of {L, R} describing the choice made during gi−1 e ui . By the definitions of SP and canonical call-sequence (which is the minimum gi e with respect to the lexicographic ordering on (mk,1 , ξk,1 , . . . , mk,k , ξk,k )), a prefix of the canonical call-sequence is the canonical call-sequence of the prefix. In other words, kk,i = kk′ ,i and ξk,i = ξk′ ,i for every i ≥ 1 and k, k′ ≤ i. Let us write ki for ki,i = ki+1,i = · · · and e ui for e ui,i = e ui+1,i = · · ·. Let π = ξ1,1 ξ2,2 ξ3,3 . . .. Then we have π main m1 D u1 g1 e m2 D u2 g2 e m3 D .... Since every infinite call-sequence satisfies the parity condition, the infinite play is P-winning. ⊔ ⊓ Soundness The proof of soundness (i.e. L0 |= Φ(P,Ω),csa implies |=csa (P, Ω)) can be given in a manner similar to the proof of completeness. To show the contraposition, assume that there exists an infinite call-sequence that violates the parity condition. Let π be the choice of this call-sequence. We construct a strategy SP,π of the opposite game GL0 ,Φ(P,Ω),csa by inspecting the reduction sequence. Then, by Lemma 25, there is no winning strategy for GL0 ,Φ(P,Ω),csa , which implies L0 6|= Φ(P,Ω),csa by Theorem 7. Definition 23 (Opposite parity game for a program). The parity game GL,(P,Ω) is defined as follows: VP := { (p, fi ) | p ∈ DL,κi , p: complete coprime } VO := { (d1 , . . . , dn ) | d1 ∈ DL,κ1 , . . . , dn ∈ DL,κn } E := { ((p, fi ), (d1, . . . , dn )) | p ⊒ ~λx̃i .ti ([ f1 7→ d1 , . . . , fn 7→ dn ]) } ∪ { ((d1, . . . , dn ), (p, fi )) | p ⊒ di , p: complete coprime }. The priority of (d1 , . . . , dn ) is 0 and that of (p, fi ) is Ω( fi ) + 1. Corollary 3. GL0 ,(P,Ω) is isomorphic to GL0 ,Φ(P,Ω),csa , i.e. there exists a bijection of nodes that preserves the owner and the priority of each node. Proof. The bijection on nodes is given by (d1 , . . . , dn ) ↔ (d1 , . . . , dn ) and (p, fi ) ↔ (p, fi ). This bijection preserves edges because of Lemma 28. ⊔ ⊓ Let π0 ∈ {L, R}ω. Given a head occurrence of t of a term reachable from main following π0 (i.e. (main; π0 ) −→∗D (t ũ; π) for some ũ), we assign an element d¯t,ũ,π ∈ DL0 ,κ where κ is the simple type of t by induction on the order of κ. – Case e u is empty: Then t : ⋆ and we define d¯t,ǫ,π = { }. 65 – Case e u = u1 . . . uk : Then t ũ = t u1 . . . uk . Note that the order of u j (1 ≤ j ≤ k) is less than that of t. For every 1 ≤ j ≤ k, we define ē j as follows. • Case u j : int: Then ē j = ~u j . (Note that u j has no free variable and no function symbol.) • Case u j : κ: Let ē j be the element defined by  ē j = { d¯u j ,ṽ,π′ | (t ũ; π) −→∗D (u j ṽ; π′ ) }. (By the convention, u j in the right-hand side originates from the j-th argument of t.) Then d¯t,ũ,π is the function defined by    (if x̄ j ⊑ ē j for all j ∈ {1, . . . , k}) { } d¯t,ũ,π (x1 , . . . , xk ) :=   {s⋆ } (otherwise). Note that d¯t,ũ,π is the maximum function such that d¯t,ũ,π (ē1 , . . . , ēk ) = { }. Let θ = [s1 /x1 , . . . , sk /xk ] and assume (main; π0 ) −→∗D ((θt0 ) ũ; π). We define a mapping from function symbols and variables to elements of their semantic domains as follows. – For a function symbol fi :  ρ̄t0 ,ũ,θ,π ( fi ) = { d¯fi,ṽ,π′ | ((θt0 )e u; π) −→∗D ( fi ṽ; π′ ), fi originates from t0 } – For a variable xi : int: ρ̄t0 ,ũ,θ,π (xi ) = ~si  – For a variable xi : κ:  ρ̄t0 ,ũ,θ,π (xi ) = { d¯si,ṽ,π′ | ((θt0 )e u; π) −→∗D (si ṽ; π′ ), si originates from θ(xi ) }. Lemma 33. Let θ = [s1 /x1 , . . . , sk /xk ] and assume (main, π0 ) −→∗D ((θt0 ) ũ; π). Let v be an arithmetic expression (i.e. a term of type int) appearing in t0 . Then ~θv = ~v(ρ̄t0 ,ũ,θ,π). Proof. Similar to the proof of Lemma 29. ⊔ ⊓ Lemma 34. Let θ = [s1 /x1 , . . . , sk /xk ] and π0 ∈ {L, R}ω. Assume that the unique reduction sequence starting from (main; π0 ) does not terminate and that (main; π0 ) −→∗D ((θt0 ) ũ, π). Then d¯θt0 ,ũ,π ⊒ ~t0 (ρ̄t0 ,ũ,θ,π ). Proof. By induction on the structure of t0 . – Case t0 = ( ): This case never occurs since the reduction sequence does not terminate. – Case t0 = xi for some 1 ≤ i ≤ k: Then (θt0 )ũ = si ũ. By definition, d¯si ,ũ,π = ρ̄t0 ,ũ,θ,π (xi ). 66 – Case t0 = fi : Then (θt0 )ũ = fi ũ. Since this occurrence of fi originates from t0 , we have d¯fi ,ũ,π ⊒ ρ̄t0 ,ũ,θ,π ( fi ). – Case t0 = n or t1 op t2 : Never occurs because the type of t0 must be of the form η1 → · · · → ηℓ → ⋆. – Case t0 = if p(t′1 , . . . , t′k ) then t1 else t2 : Then ũ is the empty sequence. Assume that (~θt′1 , . . . , ~θt′k ) ∈ ~p. The other case can be proved by a similar way. Then we have (θt0 ; π) −→D (θt1 ; π). By definition, d¯θt0 ,ǫ,π = d¯θt1 ,ǫ,π = { }. Since every reduction sequence (θt0 ; π) −→+D (s ṽ; π′ ) can be factored into (θt0 ; π) −→D (θt1 ; π) −→∗D (s ṽ; π′ ), we have ρ̄t0 ,ǫ,θ,π = ρ̄t1 ,ǫ,θ,π. By the induction hypothesis, { } ⊒ ~t1 (ρ̄t1 ,ǫ,θ,π). Hence { } ⊒ ~t1 (ρ̄t0 ,ǫ,θ,π ). By Lemma 33, ~θt′j  = ~t′j (ρ̄t0 ,ũ,θ,π ) for every j = 1, . . . , k. Hence (~θt′1 , . . . , ~θt′k ) ∈ ~p implies (~t′1 (ρ̄t0 ,ũ,θ,π ), . . . , ~t′k (ρ̄t0 ,ũ,θ,π )) ∈ ~p. By definition, ~t0 (ρ̄t0 ,ũ,θ,π ) = ~t1 (ρ̄t0 ,ũ,θ,π ) ⊑ { }. – Case t0 = event a; t: Then t0 has type ⋆ and ũ is the empty sequence. We have θt0 = event a; (θt). Since every reduction sequence (θt0 ; π) −→+D (s ṽ; π′ ) can be factored into (θt0 ; π) −→D (θt; π) −→∗D (s ṽ; π′ ), we have ρ̄t0 ,ǫ,θ,π = ρ̄t,ǫ,θ,π . By definition, d¯θt0 ,ǫ,π = d¯θt,ǫ,π = { } and ~t0 (ρ̄t0 ,ǫ,θ,π ) = ~t(ρ̄t0 ,ǫ,θ,π ) = ~t(ρ̄t,ǫ,θ ). By the induction hypothesis, d¯θt,ǫ,π ⊒ ~t(ρ̄t,ǫ,θ,π ). Hence d¯θt0 ,ǫ,π ⊒ ~t0 (ρ̄t0 ,ǫ,θ,π ). – Case t0 = t1 t2 : Suppose that ((θt1 )(θt2 )ũ; π) −→∗D ((θt2 ) ṽ; π′ ). We first show that ρ̄t2 ,ṽ,θ,π′ (x) ⊒ ρ̄t1 t2 ,ũ,θ,π (x) for every x ∈ { f1 , . . . , fn } ∪ {x1 , . . . , xk }. • Case x = xi : int: Then ρ̄t2 ,ṽ,θ,π′ (xi ) = ρ̄t1 t2 ,ũ,θ,π (xi ) = ~si . • Case x = xi : κ: Then  { d¯si,w̃,π′′ | ((θt2 ) ṽ; π′ ) −→∗D (si w̃; π′′ ), si originates from θ(xi ) } ρ̄t2 ,ṽ,θ,π′ (xi ) =  ⊒ { d¯si,w̃,π′′ | ((θ(t1 t2 )) ũ; π) −→∗D (si w̃; π′′ ), si originates from θ(xi ) } = ρ̄t1 t2 ,ũ,θ,π (xi ) because ((θt2 ) ṽ; π′ ) −→∗D (si w̃; π′′ ) (where si originates from θ(xi )) implies ((θ(t1 t2 )) ũ; π) −→∗D ((θt2 )ṽ; π′ ) −→∗D (si w̃; π′′ ) (where si originates from θ(xi )). • Case x = fi :  { d¯fi,w̃,π′′ | ((θt2 ) ṽ; π′ ) −→∗D ( fi w̃; π′′ ), fi originates from t2 } ρ̄t2 ,ṽ,θ,π′ ( fi ) =  ⊒ { d¯fi,w̃,π′′ | ((θ(t1 t2 )) ũ; π) −→∗D ( fi w̃; π′′ ), fi originates from t1 t2 } = ρ̄t1 t2 ,ũ,θ,π ( fi ) because ((θt2 ) ṽ; π′ ) −→∗D ( fi w̃; π′′ ) (where fi originates from t2 ) implies ((θ(t1 t2 )) ũ; π) −→∗D ((θt2 )ṽ; π′ ) −→∗D ( fi w̃; π′′ ) (where fi originates from t1 t2 ). It is easy to see that ρ̄t1 ,(θt2 ) ũ,θ,π ⊒ ρ̄t1 t2 ,ũ,θ,π . By the induction hypothesis, d¯θt1 ,(θt2 ) ũ,π ⊒ ~t1 (ρ̄t1 ,(θt2 ) ũ,θ,π ) ⊒ ~t1 (ρ̄t1 t2 ,ũ,θ,π ). 67 By the definition of d¯θt1 ,(θt2 ) ũ,π ,   {d¯θt2 ,ṽ,π′ | ((θt1 ) (θt2 ) ũ; π) −→∗D ((θt2 )ṽ; π′ )} = d¯(θt1 ) (θt2 ), ũ,π . d¯θt1 ,(θt2 ) ũ,π By the induction hypothesis, for every reduction sequence ((θt1 ) (θt2 ) ũ; π) −→∗D ((θt2 )ṽ; π′ ), we have d¯θt2 ,ṽ,π′ ⊒ ~t2 (ρ̄t2 ,ṽ,θ,π′ ). Since ρ̄t2 ,ṽ,θ,π′ (x) ⊒ ρ̄t1 t2 ,ũ,θ,π′ (x) for every x, we have d¯θt2 ,ṽ,π′ ⊒ ~t2 (ρ̄t2 ,ṽ,θ,π′ ) ⊒ ~t2 (ρ̄t1 t2 ,ũ,θ,π ). Because the reduction sequence ((θt1 ) (θt2 ) ũ; π) −→∗D ((θt2 )ṽ; π) is arbitrary,   {d¯θt2 ,ṽ,π′ | ((θt1 ) (θt2 ) ũ; π) −→∗D ((θt2 )ṽ; π′ )} ⊒ ~t2 (ρ̄t1 t2 ,ũ,θ,π ). Therefore, by monotonicity of interpretation,   d¯(θt1 )(θt2 ), ũ,π = d¯θt1 ,(θt2 ) ũ,π {d¯θt2 ,ṽ,π′ | ((θt1 ) (θt2) ũ; π) −→∗D ((θt2 )ṽ; π′ )}   ⊒ ~t1 (ρ̄t1 t2 ,ũ,θ,π) ~t2 (ρ̄t1 t2 ,ũ,θ,π ) = ~t1 t2 (ρ̄t1 t2 ,ũ,θ,π ). – Case t0 = t1 t2 : Then ũ = ǫ. Assume that π = Lπ′ ; the other case can be proved similarly. Then we have (θt0 ; π) −→D (θt1 ; π′ ). By the definition of d¯θt1 ,ǫ,π′ and the induction hypothesis, { } = d¯θt1 ,ǫ,π′ ⊒ ~t1 (ρ̄t1 ,ǫ,θ,π′ ). Because (θt0 ; π) −→D (θt1 ; π′ ), we have ρ̄t1 ,ǫ,θ,π′ ⊒ ρ̄t0 ,ǫ,θ,π. By monotonicity of interpretation, { } = d¯θt1 ,ǫ,π′ ⊒ ~t1 (ρ̄t1 ,ǫ,θ,π′ ) ⊒ ~t1 (ρ̄t0 ,ǫ,θ,π). Hence { } ⊒ ~t1 (ρ̄t0 ,ǫ,θ,π ) ⊓ ~t2 (ρ̄t0 ,ǫ,θ,π) = ~t0 (ρ̄t0 ,ǫ,θ,π ). ⊔ ⊓ The strategy SP,π0 of GL0 ,P on ({s⋆ }, main) is defined as follows. Each play in the domain of S ({s⋆ }, main) · O1 · (p1 , g1 ) · O2 · · · · · Ok · (pk , gk ) is associated with a call-sequence (main; π0 ) m1 D (g1 ũ1 ; π1 ) m2 D (g2 ũ2 ; π2 ) m3 D ... mk−1 D (gk−1 ũk−1 ; πk−1 ) mk D (gk ũk ; πk ) such that p j ⊒ d¯g j ,ũ j for every j = 1, 2, . . . , k. The canonical associated call-sequence of the play is the minimum one ordered by the lexicographic ordering of reduction steps (m1 , m2 , . . . , mk ). Assume that the above call-sequence is canonical. The next step of this reduction sequence is ([ũk /x̃k ]t; πk ) if (gk e xk = t) ∈ D. Let ϑ = [ũk /x̃k ]. In this situation, the strategy SP,π chooses (ρ̄t,ǫ,ϑ,πk ( f1 ), . . . , ρ̄t,ǫ,ϑ,πk ( fn )) as the next node. This is a valid choice, i.e.: 68 Lemma 35. pk ⊒ ~D(gk )([ f1 7→ ρ̄t,ǫ,ϑ,πk ( f1 ), . . . , fn 7→ ρ̄t,ǫ,ϑ,πk ( fn )]). Proof. By Lemma 34, we have: d¯ϑt,ǫ,πk = { } ⊒ ~t(ρ̄t,ǫ,ϑ,πk ) = ~λe xk .t([ f1 7→ ρ̄t,ǫ,ϑ,πk ( f1 ), . . . , fn 7→ ρ̄t,ǫ,ϑ,πk ( fn )])(ρ̄t,ǫ,ϑ,π(e xk )). By definition, d¯gk ,ũk ,πk is the greatest element such that d¯gk ,ũk ,πk (ρ̄t,ǫ,ϑ,π (e xk )) = { }. Therefore d¯gk ,ũk ⊒ ~λe xk .t([ f1 7→ ρ̄t,ǫ,ϑ,πk ( f1 ), . . . , fn 7→ ρ̄t,ǫ,ϑ,πk ( fn )]). Since pk ⊒ d¯gk ,ũk ,πk , we have pk ⊒ d¯gk ,ũk ,πk ⊒ ~λx̃k .t([ f1 7→ ρ̄t,ǫ,ϑ,πk ( f1 ), . . . , fn 7→ ρ̄t,ǫ,ϑ,πk ( fn )]). ⊔ ⊓ Now we have a play ({s⋆ }, main) · O1 · (p1 , g1 ) · O2 · · · · · Ok · (pk , gk ) · (ρ̄t,ǫ,θ,πk ( f1 ), . . . , ρ̄t,ǫ,θ,πk ( fn )) associated with the call sequence (main; π0 ) m1 D (g1 ũ1 ; π1 ) m2 D (g2 ũ2 ; π2 ) m3 D ... mk−1 D (gk−1 ũk−1 ; πk−1 ) mk D (gk ũk ; πk ) Let (pk+1 , gk+1 ) be the next opponent move. By definition of the game,  pk+1 ⊒ ρ̄t,ǫ,θ,π (gk+1 ) = { d¯gk+1 ,ṽ,π | (θt; πk ) −→∗P (gk+1 ṽ; π), gk+1 originates from t }  = { d¯gk+1 ,ṽ,π | (gk ũk ; πk ) D (gk+1 ṽ; π) }. Since pk+1 is a complete coprime, pk+1 ⊒ d¯gk+1 ,ṽ,π for some ṽ and π with (gk ũk ; πk ) D (gk+1 ṽ; π). Hence we have an associated call-sequence (main; π0 ) m1 D (g1 ũ1 ; π1 ) m2 D (g2 ũ2 ; π2 ) m3 D ... mk D (gk ũk ; πk ) mk+1 (gk+1 ũ; π) of ({s⋆ }, main)·O1 ·(p1 , g1 )·O2 ·· · ··Ok ·(pk , gk )·(ρ̄t,ǫ,θ,πk ( f1 ), . . . , ρ̄t,ǫ,θ,πk ( fn ))·(pk+1 , gk+1 ) as desired. Lemma 36. If the (unique) call sequence following π0 does not satisfy the parity condition, then SP,π0 is an winning strategy of GL0 ,(P,Ω) on ({}, main). Proof. The above argument shows that SP,π0 is a strategy of GL0 ,(P,Ω) on ({}, main). We prove that this is winning. Assume an infinite play ({}, main) · v1 · (p1 , g1 ) · v2 · (p2 , g2 ) · . . . 69 that conforms with SP,π and starts from ({s⋆ }, main). Then each odd-length prefix is associated with the canonical call-sequence (main; π0 ) km,1 D um,1 ; πm,1 ) (g1 e km,2 D um,2 ; πm,2 ) (g2 e km,3 D ... km,m D um,m ; πm,m ). (gm e By the definitions of SP,π0 and canonical call-sequence (which is the minimum with respect to the lexicographic ordering on (km,1 , . . . , km,m )), a prefix of the canonical call-sequence is the canonical call-sequence of the prefix. In other words, km,i = km′ ,i for every i ≥ 1 and m, m′ ≥ i. Since the reduction sequence is completely determined by a choice π0 , the initial term main and the number of steps, we have πm,i = πm′ ,i for every i ≥ 1 and m, m′ ≥ i. Let us write ki for ki,i = ki+1,i = · · ·, e ui for e ui,i = e ui+1,i = · · · and πi for πi,i = πi+1,i = · · ·. Now we have an infinite call-sequence π0 main k1 D u1 g1 e k2 D u2 g2 e k3 D ... following π0 . By Corollary 1, this is the unique call-sequence following π0 . Hence, by the assumption, this infinite call-sequence does not satisfy the parity condition. This means that this play is P-winning (recall the definition of the priorities of GL0 ,(P,Ω) ). ⊔ ⊓ Proof (Proof of Theorem 4). We prove the following result: |=csa (P, Ω) if and only if L0 |= Φ(P,Ω),csa . (⇒) Suppose that |=csa (P, Ω). Then by Lemma 32, there exists a winning strategy of GL0 ,(P,Ω) on ({s⋆ }, main). Then by Corollary 2, there also exists a winning strategy of GL0 ,Φ(P,Ω),csa on ({s⋆ }, main). By Theorem 7, this implies s⋆ ∈ ~Φ(P,Ω),csa . By definition, L0 |= Φ(P,Ω),csa . (⇐) We prove the contraposition. Suppose that |=csa (P, Ω) does not hold. Then there exist a choice π0 and an infinite call-sequence following π0 that does not satisfy the parity condition. Then by Lemma 36, SP,π0 is a winning strategy of GL0 ,(P,Ω) on ({}, main). By Corollary 3, there also exists a winning strategy of GL0 ,Φ(P,Ω),csa on ({}, main). Then by Lemma 25, Opponent wins GL0 ,Φ(P,Ω),csa on ({s⋆ }, main). Hence by Theorem 7, s⋆ < ~Φ(P,Ω),csa . ⊔ ⊓ E.4 Proof of Theorem 5 Assume a total order on the set of states of the automaton A, fixed in the sequel. Recall (IT-Event) rule in Fig. 6: δ(q, a) = {q1 , . . . , qn } Γ ↑ Ω(qi ) ⊢A t : qi ⇒ t′i (for each i ∈ {1, . . . , n}) Γ ⊢A (event a; t) : q ⇒ (event a; t′1  · · · t′n ) (IT-Event) Since the order of t′1 , . . . , t′n is not important, we can assume without loss of generality that q1 < q2 < · · · < qn holds for every instance of the above rule used in a derivation. 70 Here we prove Theorem 5 for the translation using the following rule IT-Event’ instead of IT-Event. δ(q, a) = {q1 , . . . , qn } Γ ↑ Ω(qi ) ⊢A t : qi ⇒ t′i q1 < q2 < · · · < qn (for each i ∈ {1, . . . , n}) Γ ⊢A (event a; t) : q ⇒ (event a; t′1  · · · t′n ) (IT-Event’) V Notations Given an intersection type ρ = 1≤i≤k (θi , mi ) and a variable x, we abbreviate the sequence xθ1 ,m1 xθ2 ,m2 . . . xθk ,mk as dup(x, ρ). For ρ = int, we write dup(x, ρ) for xint. V For an intersection type ρ = 1≤i≤k (θi , mi ), we write [x : ρ] for the intersection type environment {x : (θi , mi , 0) | 1 ≤ i ≤ k}. If ρ = int, [x : ρ] means {x : int}. Similarly, for a top-level environment Ξ, we write [Ξ] to mean {x : (θ, m, 0) | x : (θ, m) ∈ Ξ}. Given an intersection type environment Γ, we write Γ♯ for {x♯ : (θ, m, m′) | x : (θ, m, m′) ∈ Γ} ∪ {x♯ : int | x : int ∈ Γ}. Similarly, for a top-level environment Ξ, we write Ξ♯ for {x♯ : (θ, m) | x : (θ, m) ∈ Ξ}. For an intersection type environment Γ and a top-level environment Ξ, we write Γ ⊳ Ξ if f : (θ, m) ∈ Ξ for each ( f : (θ, m, m′ )) ∈ Γ. Modified type-based translation Given programs P and P′ with Ξ ⊢A P ⇒ (P′ , Ω), the first step to prove Theorem 5 is to compare reduction sequences of P and P′ . There is a little gap between reduction sequences of P and P′ since a reduction sequence of P′ expresses a reduction sequence of P together with a run of A over the events generated by the reduction sequence of P. In particular a nondeterministic branch in t′ comes from either – a non-deterministic branch t1 t2 in P, or – non-determinism of the transition rule of the automaton A. To fill the gap, we shall distinguish between the two kinds of non-deterministic branches, by using ◦ for the latter. Formally let us introduce a new binary construct ◦ to the syntax of terms. Here we use the convention that  and ◦ are right associative, i.e. t1 t2 t3 (resp. t1 ◦ t2 ◦ t3 ) means t1 (t2 t3 ) (resp. t1 ◦ (t2 ◦ t3 )). The operational behavior of ◦ is the same as that of , i.e., ǫ ǫ (t1 ◦ t2 ; Lπ) −→D (t1 ; π) (t1 ◦ t2 ; Rπ) −→D (t2 ; π) where π ∈ {L, R}ω is a choice sequence. Hence we have a ((event a; (t1 ◦ . . . ◦ tn )); R . . . R Lπ) −→∗D (ti ; π) |{z} i−1 for 1 ≤ i < n and a ((event a; (t1 ◦ . . . ◦ tn )); R . . . R π) −→∗D (tn ; π). |{z} n−1 71 δ(q, a) = {q1 , . . . , qn } (q1 < q2 < · · · < qn ) Γ ↑ Ω(qi ) ⊢A t : qi ⇛ t′i (for each i ∈ {1, . . . , n}) Γ ⊢A (event a; t) : q ⇛ (event a; t′1 ◦ · · · ◦ t′n ) (IT-Event-Alt) Fig. 9. Modified type-based transformation rules. Other rules are obtained by replacing ⇒ with ⇛ in the rules in Figure 6 except for (IT-Event). Definition 24 (Modified type-based transformation). The modified type-based transformation judgment Γ ⊢A t : θ ⇛ t′ is a quadruple where t is a term without ◦ and t′ is a term possibly having ◦. The modified type-based transformation relation is defined by the rules in Fig. 9. The translation ⊢A P ⇛ (P′ , Ω) of programs is defined in the same way as ⊢A P ⇒ (P′ , Ω). The next lemma establishes the connection between the original and modified transformations. Given a program P′ with ◦, we write [/◦]P′ for the program obtained by replacing ◦ with . Lemma 37. If Ξ ⊢A P ⇛ (P′ , Ω), then Ξ ⊢A P ⇒ ([/◦]P′, Ω). Conversely, if Ξ ⊢A P ⇒ (P′ , Ω), then there exists P′′ such that Ξ ⊢A P ⇛ (P′′ , Ω) and P′ = [/◦]P′′ . ⊔ ⊓ Proof. Straightforward induction. Suppose that Ξ ⊢A P ⇛ (P′ , Ω). Since P′ simulates both P and the automaton A, an infinite reduction sequence of P′ induces a pair of an infinite event sequence e ℓ and a run R of A over e ℓ. Definition 25 (Induced event sequence and run). Let D′ be a function definition e ℓ,R (s′ , π′ , q′ ), meaning that the possibly containing ◦. The reduction relation (s, π, q) −→N D′ N-step reduction following π from s to s′ generates events e ℓ associated with a run qR ending with q′ , is defined by the following rules. ǫ ′ ′ (s, π) −→N D′ (s , π ) ǫ,ǫ (s, π, q) −→N (s′ , π′ , q) D′ δ(q, a) = {q1 , . . . , qn } (q1 < q2 < · · · < qn ) e ℓ,R (si , π, qi ) −→N (s′ , π′ , q′ ) D′ ae ℓ,qi R (event a; (s1 ◦ . . . ◦ sn ), R . . . R Lπ, q) −→ N+i+1 (s′ , π′ , q′ ) D′ |{z} i−1 δ(q, a) = {q1 , . . . , qn } (q1 < q2 < · · · < qn ) e ℓ,R (sn , π, qn ) −→N (s′ , π′ , q′ ) D′ ae ℓ,qn R (s′ , π′ , q′ ) (event a; (s1 ◦ . . . ◦ sn ), R . . . R π, q) −→ N+n D′ |{z} n−1 72 e ℓ1 ,R1 e ℓ2 ,R2 1 2 If (s0 , π0 , q0 ) −→N (s1 , π1 , q1 ) −→N . . ., we write π0 D′ D′ e ℓ1 ,R1 If the number of steps (resp. π) is not important, we write as π e ℓ,R e ℓ2 ,R2 1 2 (s0 , q0 ) −→N (s1 , q1 ) −→N . . .. D′ D′ e ℓ,R (s, q) −→∗D′ (s′ , q′ ) e ℓ,R (resp. (s, q)−→N (s′ , q′ )). Other notations such as (s, q)−→∗D′ (s′ , q′ ) are defined similarly. D′ e ℓ,R It is easy to show that, if (s, q) −→∗D′ (s′ , q′ ), then qR is a run of A over e ℓ. Obviously π e ℓ,R e ℓ (s, q) −→∗D′ (s′ , q′ ) implies π s −→∗D′ s′ . As proved later in e ℓ Lemma 41, the converse also holds: Given state q and s −→∗D′ s′ (with a mild e ℓ,R condition), there exist R and q′ such that (s, q) −→∗D′ (s′ , q′ ). Basic properties of the type system and transformation We prove Weakening Lemma and Substitution Lemma. Lemma 38 (Weakening). If Γ ⊢A t : θ ⇛ t′ and Γ ⊆ Γ′ , then Γ′ ⊢A t : θ ⇛ t′ . Proof. Straightforward induction on structure of the derivations. We discuss only one case below. – Case for IT-Event-Alt: Then we have θ=q δ(q, a) = {q1 , . . . , qn } t = event a; s t′ = event a; (s′1 ◦ . . . ◦ s′n ) ∀i ∈ {1, . . . , n}.Γ ↑ Ω(qi ) ⊢A s : θ ⇛ s′i . Since Γ ⊆ Γ′ , for each i ∈ {1, . . . , n}, we have Γ ↑ Ω(qi ) ⊆ Γ′ ↑ Ω(qi ). By the induction hypothesis, ∀i ∈ {1, . . . , n}.Γ′ ↑ Ω(qi ) ⊢A s : θ ⇛ s′i . By using IT-Event-Alt, we obtain Γ′ ⊢A t : θ ⇛ t′ . ⊔ ⊓ In order to simplify the statement V of the substitution lemma, we introduce the following abbreviation. For ρ = 1≤i≤l (θi , mi ), we write Γ ⊢A t : ρ ⇛ e s def ⇔ ∀i ∈ {1, . . . , l}.Γ ↑ mi ⊢A t : θi ⇛ si . where e s = s1 s2 . . . sl . If ρ = int, the judgment Γ ⊢A t : ρ ⇛ e s has the obvious meaning (in this case, e s is of length 1). By using this notation, the application rule can be simply written as Γ ⊢A t1 : ρ → θ ⇛ t′1 Γ ⊢A t 2 : ρ ⇛ e t′2 Γ ⊢A t1 t2 : θ ⇛ t′1 e t′2 73 Lemma 39 (Substitution). Assume that Γ ⊢A u : ρ ⇛ e u′ ((Γ ∪ [x : ρ]) ↑ n) ∪ Γ′ ⊢A t : θ ⇛ t′ (x < dom(Γ)). Then (Γ ↑ n) ∪ Γ′ ⊢A [u/x]t : θ ⇛ [e u′ /dup(x, ρ)]t′. Proof. The proof proceeds by induction on the derivation of ((Γ ∪ [x : ρ]) ↑ n) ∪ Γ′ ⊢A t : θ ⇛ t′ , with case analysis on the last rule used. – Case for IT-Var: The V case where t , x is trivial. Assume that t = x. The type ρ is either int or 1≤i≤l (θi , mi ). The former case is easy; we prove the latter case. V Assume that ρ = 1≤i≤l (θi , mi ). Since ((Γ ∪ [x : ρ]) ↑ n) ∪ Γ′ ⊢A x : θ ⇛ t′ , there exists k ∈ {1, . . . , l} such that θ = θk , t′ = xθk ,mk and n = mk . By the assumption Γ ⊢A u : ρ ⇛ e u′ , we have Γ ↑ mk ⊢A u : θk ⇛ u′k . Since [u/x]x = u ′ and [e u /dup(x, ρ)]xθk,mk = u′k , we have Γ ↑ mk ⊢A [u/x]x : θk ⇛ [ue′ /dup(x, ρ)]xθk,mk . By Weakening (Lemma 38), we have (Γ ↑ mk ) ∪ Γ′ ⊢A [u/x]x : θk ⇛ [ue′ /dup(x, ρ)]xθk,mk . as required. – Case for IT-Event-Alt: Then t = event a; t1 and we have θ=q δ(q, a) = {q1 , . . . , qk } q1 < q2 < · · · < qk ′ t = event a; (t′1 ◦ . . . ◦ t′k ) ∀j ∈ {1, . . . , k}.(((Γ ∪ [x : ρ]) ↑ n) ∪ Γ′ ) ↑ Ω(q j ) ⊢A t1 : q j ⇛ t′j . For each j ∈ {1, . . . , k}, we have (((Γ ∪ [x : ρ]) ↑ n) ∪ Γ′ ) ↑ Ω(q j ) = ((Γ ∪ [x : ρ]) ↑ n ↑ Ω(q j )) ∪ (Γ′ ↑ Ω(q j )) ((Γ ↑ n) ∪ Γ′ ) ↑ Ω(q j ) = (Γ ↑ n ↑ Ω(q j )) ∪ (Γ′ ↑ Ω(q j )). Thus, one can apply the induction hypotheses, obtaining ∀j ∈ {1, . . . , k}.((Γ ↑ n) ∪ Γ′ ) ↑ Ω(q j ) ⊢A [u/x]t1 : q j ⇛ [ue′ /dup(x, ρ)]t′j. By using IT-Event-Alt, we have the following as required: (Γ ↑ n) ∪ Γ′ ⊢A [u/x](event a; t1 ) : q ⇛ [ue′ /dup(x, ρ)](event a; (t′1 ◦ . . . ◦ t′k )). 74 – Case for IT-App: Then t = t1 t2 and we have t′ = t′1 t′2,1 . . . t′2,k ((Γ ∪ [x : ρ]) ↑ n) ∪ Γ′ ⊢A t1 : ^ (θ′j , m′j ) → θ ⇛ t′1 1≤ j≤k ∀j ∈ {1, . . . , k}.(((Γ ∪ [x : ρ]) ↑ n) ∪ Γ′ ) ↑ m′j ⊢A t2 : θ′j ⇛ t′2, j . By the induction hypothesis, we have ^ (Γ ↑ n) ∪ Γ′ ⊢A [u/x]t1 : (θ′j , m′j ) → θ ⇛ [ue′ /dup(x, ρ)]t′1 1≤ j≤k ′ ∀j ∈ {1, . . . , k}.((Γ ↑ n) ∪ Γ ) ↑ m′j By using IT-App, we obtain: ⊢A [u/x]t2 : θ′j ⇛ [ue′ /dup(x, ρ)]t′2, j. (Γ ↑ n) ∪ Γ′ ⊢A [u/x](t1 t2 ) : θ ⇛ [ue′ /dup(x, ρ)](t′1 t′2,1 . . . t′2,k ) as required. – Case for IT-Abs: Then t = λy.s and we have ^ θ= (θ′j , m′j ) → θ′0 t′ = λyθ′1 ,m′1 . . . yθ′k ,m′k .s′ 1≤ j≤k ((Γ ∪ [x : ρ]) ↑ n) ∪ Γ′ ∪ [y : y < dom((Γ ∪ [x : ρ]) ↑ n) ^ (θ′j , m′j )] ⊢A s : θ′0 ⇛ s′ 1≤ j≤k By the induction hypothesis, we have ^ (Γ ↑ n) ∪ Γ′ ∪ [y : (θ′j , m′j )] ⊢A [u/x]s : θ′0 ⇛ [ue′ /dup(x, ρ)]s′. 1≤ j≤k By using IT-Abs, we obtain: ^ (Γ ↑ n) ∪ Γ′ ⊢A [u/x]λy.s : (θ′j , m′j ) → θ′0 ⇛ [ue′ /dup(x, ρ)]λyθ′1,m′1 . . . yθ′k ,m′k .s′ 1≤ j≤k as required. ⊔ ⊓ Simulations in both directions Given Ξ ⊢A P ⇛ (P′ , Ω), we show that the following two data are equivalent: 1. A reduction sequence of P together with a run of A over the generated event sequence. 2. A reduction sequence of P′ . We first prove the direction (1) ⇒ (2). Lemma 40. Assume 75 – Ξ ⊢A D ⇛ D ′ , – Γ ⊢A t : q ⇛ s, – Γ ⊳ Ξ, and e ℓ – t −→∗D t′ . Let qR be an arbitrary run of A over e ℓ and q′ be the last state of qR. Then there exist a ′ ′ term s and a type environment Γ ⊳ Ξ such that e ℓ,R (s, q) −→∗D′ (s′ , q′ ) and Γ′ ⊢A t′ : q′ ⇛ s′ . e ℓ Proof. By induction on the length of the reduction sequence t −→∗D t′ . The claim trivially holds if the length is 0; we assume that the length is not 0. The proof proceeds by case analysis on the shape of t. e ℓ – Case t = event a; t1 : Then the reduction sequence t −→∗D t′ is of the form e ℓ′ a event a; t1 −→D t1 −→∗D t′ with e ℓ = aℓe′ . Since the last rule used to derive Γ ⊢A t : q ⇛ s is IT-Event-Alt, we have: δ(q, a) = {q1 , . . . , qk } ∀i ∈ {1, . . . , k}.Γ ↑ Ω(qi ) ⊢A t1 : qi ⇛ si (q1 < q2 < · · · < qk ) s = event a; (s1 ◦ . . . ◦ sk ). Since qR is a run over e ℓ = ae ℓ′ , it must be of the form qR = qqi R′ for some 1 ≤ i ≤ k, where qi R′ is a run over e ℓ′ . By applying the induction hypothesis e ℓ′ to t1 −→∗D t′ , there exists s′ such that e ℓ′ ,R′ Γ′ ⊢A t′ : q′ ⇛ s′ (si , qi ) −→∗D′ (s′ , q′ ) for some Γ′ ⊳ Ξ. Then we have a,qi e ℓ′ ,R′ (event a; (s1 ◦ . . . ◦ sk ), q) −→∗D′ (si , qi ) −→∗D′ s′ . ǫ e ℓ – Case t = t1 t2 : Suppose that the reduction sequence is t −→D t1 −→∗D t′ ; the ǫ e ℓ case that t −→D t2 −→∗D t′ can be proved similarly. Since the last rule used to derive Γ ⊢A t : q ⇛ s is IT-NonDet, we have: s = s1 s2 Γ ⊢A t1 : q ⇛ s1 76 Γ ⊢A t2 : q ⇛ s2 e ℓ By applying the induction hypothesis to t1 −→∗D t′ , we have s′ such that e ℓ,R Γ′ ⊢A t′ : q′ ⇛ s′ (s1 , q) −→∗D′ (s′ , q′ ) for some Γ′ ⊳ Ξ. Then e ℓ,R ǫ,ǫ (s1 s2 , q) −→∗D′ (s1 , q) −→∗D′ (s′ , q′ ). – Case t = if p(t1 , . . . , tn ) then tn+1 else tn+2 : Suppose (~t1 , . . . , ~tn ) ∈ ~p. e ℓ ǫ Then the reduction sequence is t −→D tn+1 −→∗D t′ . Since the last rule used to derive Γ ⊢A t : q ⇛ s is IT-IF, we have: s = if p(s1 , . . . , sn ) then sn+1 else sn+2 ∀i ∈ {1, . . . , n}.Γ ⊢A ti : int ⇛ si Γ ⊢A tn+1 : q ⇛ sn+1 Γ ⊢A tn+2 : q ⇛ sn+2 By the well-typedness of the program, for every i ∈ {1, . . . , n}, ti consists of only integers and integer operations. Thus, we have ~ti  = ~si  for each ǫ i ∈ {1, . . . , n} and s −→D′ sn+1 . e ℓ By applying the induction hypothesis to tn+1 −→∗D t′ , we have e ℓ,R (sn+1 , q) −→∗D′ (s′ , q′ ) Γ′ ⊢A t′ : q′ ⇛ s′ for some Γ′ ⊳ Ξ and s′ . Then e ℓ,R ǫ,ǫ (s, q) −→∗D′ (sn+1 , q) −→∗D′ (s′ , q′ ). The case where (~t1 , . . . , ~tn ) < ~p is similar. – Case t = f t1 . . . tn : Suppose D( f ) = λx1 . . . xn .t0 . In this case, the reduction e ℓ ǫ sequence is t −→D [t1 /x1 ] . . . [tn /xn ]t0 −→∗D t′ . Since Γ ⊢A t : q ⇛ s, we have f : (θ, m, m) ∈ Γ such that θ = ρ1 → . . . ρn → q ∀i ∈ {1, . . . , n}. Γ ⊢A ti : ρi ⇛ e si s1 . . . e sn . s = fθ,m e Since Ξ ⊢A D ⇛ D′ , we have D′ ( fθ,m ) = λdup(x1 , ρ1 ) . . . dup(xn , ρn ).s0 [Ξ] ∪ [x1 : ρ1 ] ∪ · · · ∪ [xn : ρn ] ⊢A t0 : q ⇛ s0 By using Weakening (Lemma 38), we have ∀i ∈ {1, . . . , n}. [Ξ] ∪ Γ ⊢A ti : ρi ⇛ e si [Ξ] ∪ Γ ∪ [x1 : ρ1 ] ∪ · · · ∪ [xn : ρn ] ⊢A t0 : q ⇛ s0 77 Since ([Ξ] ∪ Γ ∪ [x1 : ρ1 ] ∪ · · · ∪ [xn : ρn ]) ↑ 0 = [Ξ] ∪ Γ ∪ [x1 : ρ1 ] ∪ · · · ∪ [xn : ρn ], by using Substitution Lemma 39 repeatedly, we have [Ξ] ∪ Γ ⊢A [t1 /x1 ] . . . [tn /xn ]t0 : q ⇛ [e s1 /dup(x1 , ρ1 )] . . . [e sn /dup(xn , ρn )]s0 . Since ([Ξ] ∪ Γ) ⊳ Ξ, we can apply the induction hypothesis to e ℓ [t1 /x1 ] . . . [tn /xn ]t0 −→∗D t′ ; we have e ℓ,R ′ ′ ′ (s′′ , q) −→N D′ (s , q ) Γ′ ⊢A t′ : q′ ⇛ s′ for some Γ′ ⊳ Ξ and s′ , where s′′ = [e s1 /dup(x1 , ρ1 )] . . . [e sn /dup(xn , ρn )]s0 . e ℓ,R ǫ,ǫ We have (s, q) −→∗D′ (s′′ , q) −→∗D′ (s′ , q′ ) as desired. ⊔ ⊓ An infinite analogue of this lemma can be obtained as a corollary. Corollary 4. Assume Ξ ⊢A D ⇛ D ′ Γ ⊢A t0 : q0 ⇛ s0 Γ⊳Ξ and an infinite reduction sequence ℓ1 ℓ1 ℓ2 t0 −→D t1 −→D t2 −→D · · · generating an infinite event sequence e ℓ = ℓ1 ℓ2 · · ·. Let q0 R be an infinite run of A over e the infinite sequence ℓ. Then there exist {(qi , Ri , si )}i∈ω such that ℓ1 ,R1 ℓ2 ,R2 ℓ3 ,R3 (s0 , q0 ) −→∗D′ (s1 , q1 ) −→∗D′ (s2 , q2 ) −→∗D′ · · · and R = R1 R2 . . .. Proof. By using Lemma 40, one can construct by induction on i > 0 a family {(qi , Ri , si , Γi )}i∈ω that satisfies ℓi ,Ri (si−1 , qi−1 )−→∗D′ (si , qi ) Γi ⊳Ξ Γi ⊢A ti :qi ⇛ si R1 R2 . . . Ri is a prefix of R. Since the length of R1 R2 . . . Ri is equivalent to that of ℓ1 ℓ2 . . . ℓi , the sequence R1 R2 . . . is indeed an infinite sequence and thus equivalent to R. ⊔ ⊓ We show the converse in a bit stronger form. 78 Lemma 41. Assume Ξ ⊢A D ⇛ D ′ Γ1 ⊳ Ξ Γ2 ⊳ Ξ♯ Γ1 ∪ Γ2 ⊢A t : q ⇛ s e ℓ (s, π) −→N (s′ , π′ ) D′♯ where t does not contain λ-abstraction and s′ is not of the form s′1 ◦ s′2 . Then there exist a run qR over e ℓ and a state q′ such that e ℓ,R (s, π, q) −→N (s′ , π′ , q′ ) D′♯ Furthermore there exist a term t′ and a type environment Γ′1 such that Γ′1 ⊳ Ξ Γ′1 ∪ (Γ2 ↑ max(Ω(R))) ⊢A t′ : q′ ⇛ s′ e ℓ t −→∗D♯ t′ . Proof. By induction on the length of the reduction sequence. The claim trivially holds if the length is 0. Assume that the length is greater than 0. The proof proceeds by case analysis on the shape of s. By the assumption e ℓ s −→∗D′♯ s′ , it suffices to consider only the cases where the shape of s matches the lefthand side of a transition rule. – Case s = event a; s0 : By the shape of s, the last rule used on the derivation of Γ ∪ Γ′ ⊢A t : q ⇛ s is either IT-Event-Alt or IT-App (IT-Abs is not applicable since t is assumed to have no abstraction). By induction on the derivation, one can prove that t = (event a; t0 ) t1 . . . tk for some k. By the simple-type system, one cannot apply event a; t0 to a term; hence k = 0 (i.e. t = event a; t0 ) and the last rule used on the derivation is IT-Event-Alt. Then we have: δ(q, a) = {q1 , . . . , qn } (q1 < q2 < · · · < qn ) ∀i ∈ {1, . . . , n}.(Γ1 ∪ Γ2 ) ↑ Ω(qi ) ⊢A t0 : qi ⇛ si s0 = (s1 ◦ . . . ◦ sl ) Recall that s is not of the form s′1 ◦ s′2 ; hence the reduction sequence ′ e ℓ (s, π) −→N (s′ , π′ ) D′♯ must be of the form e ℓ′ a (s, π) −→kD′♯ (si , π′′ ) −→N−k (s′ , π′ ) D′♯ where k = i + 1 if i < n and k = i if i = n. By the induction hypothesis, there exist a run qi R′ over ℓ′ , a term t′ and a type environment Γ′ with Γ′1 ⊳ Ξ such that e ℓ′ ,R′ (si , π′′ , qi ) −→N−k (s′ , π′ , q′ ) D′♯ and Γ′1 ∪ (Γ2 ↑ ΩA (qi ) ↑ max(ΩA (R′ ))) ⊢A t′ : q′ ⇛ s′ . We have (Γ2 ↑ ΩA (qi ) ↑ max(ΩA (R′ ))) = (Γ2 ↑ max(ΩA (qi R′ ))) and a,qi e ℓ′ ,R′ (s, π, q) −→kD′♯ (si , π′′ , qi ) −→N−k (s′ , π′ , q′ ). By construction, qqi R′ is a run over D′♯ ae ℓ′ . 79 – Case s = s1 s2 : As with the previous case, the last rule used on the derivation is IT-NonDet. Then we have (Γ1 ∪ Γ2 ) ⊢A t1 : q ⇛ s1 (Γ1 ∪ Γ2 ) ⊢A t2 : q ⇛ s2 t = t1 t2 Suppose π = Lπ′′ ; the other case can be proved by a similar way. Then the e ℓ reduction sequence (s, π) −→N (s′ , π′ ) must be of the form D′♯ e ℓ ǫ (s, π) −→D′♯ (s1 , π′′ ) −→N−1 (s′ , π′ ). D′♯ Hence, by applying the induction hypothesis to Γ1 ∪ Γ2 ⊢A t1 : q ⇛ s1 , we obtain the desired result. – Case s = if p(s1 , . . . , sn ) then sn+1 else sn+2 : As with the previous cases, the last rule used on the derivation is IT-If. Hence we have t = if p(t1 , . . . , tn ) then tn+1 else tn+2 ∀i ∈ {1, . . . , n}.Γ1 ∪ Γ2 ⊢A ti : int ⇛ si Γ1 ∪ Γ2 ⊢A tn+1 : q ⇛ sn+1 Γ1 ∪ Γ2 ⊢A tn+2 : q ⇛ sn+2 Recall that the result type of a function cannot be the integer type in our language. This implies that ti (1 ≤ i ≤ n) consists only of constants and numerical operations, and thus ti = si for every 1 ≤ i ≤ n. Suppose that (~s1 , . . . , ~sn ) = (~t1 , . . . , ~tn ) ∈ ~p. Then the reduction e ℓ (s′ , π′ ) must be of the form sequence (s, π) −→N D′♯ e ℓ ǫ (s, π) −→D′♯ (sn+1 , π) −→N−1 (s′ , π′ ). D′♯ We obtain the desired result by applying the induction hypothesis to Γ1 ∪ ǫ Γ2 ⊢A tn+1 : q ⇛ sn+1 ; the corresponding reduction sequence is t −→D♯ e ℓ tn+1 −→∗D♯ t′ . The case that (~s1 , . . . , ~sn ) < ~p can be proved similarly. – Case s = g s1 . . . sn with g ∈ dom(D′ ): By the shape of s, the last rule used on the derivation is IT-App or IT-AppInt. By induction on the derivation, we s′1 . . . e s′k and have t = f t1 . . . tk for some k. Then s = fθ,m e f : (θ, m, m) ∈ Γ1 θ = ρ1 → . . . ρk → q ∀i ∈ {1, . . . , k}. Γ1 ∪ Γ2 ⊢A ti : ρi ⇛ e s′i . Assume that D( f ) = λx1 . . . xk .tbody and D′ ( fθ,m ) = λy1 . . . yn .sbody . Since Ξ ⊢A D ⇛ D′ and f : (θ, m, m) ∈ Γ1 ⊳ Ξ, we have [Ξ] ⊢A λx1 . . . xk .tbody : θ ⇛ λy1 . . . yn .sbody . Hence y1 . . . yn = dup(x1 , ρ1 ) . . . dup(xk , ρk ) and [Ξ] ∪ [x1 : ρ1 ] ∪ · · · ∪ [xk : ρk ] ⊢A tbody : q ⇛ sbody . 80 By using Weakening (Lemma 38) and Substitution Lemma (Lemma 39) repeatedly, we have [Ξ] ∪ Γ1 ∪ Γ2 ⊢A [t1 /x1 ] . . . [tk /xk ]tbody : q ⇛ [ue1 /dup(x1 , ρ1 )] . . . [uek /dup(xk , ρk )]sbody . e ℓ The reduction sequence (s, π) −→N (s′ , π′ ) must be of the form D′♯ e ℓ ǫ (s, π) −→D′♯ ([ue1 /dup(x1 , ρ1 )] . . . [uek /dup(xk , ρk )]sbody , π) −→N−1 (s′ , π′ ). D′♯ By applying the induction hypothesis to the above judgment, we ǫ complete the proof; the corresponding reduction sequence is t −→D♯ e ℓ [t1 /x1 ] . . . [tk /xk ]tbody −→∗D♯ t′ where t′ is the term obtained by the induction hypothesis. – Case s = g♯ s1 . . . sn with g ∈ dom(D′ ): Similar to the above case. ⊔ ⊓ Corollary 5. Assume Ξ ⊢A D ⇛ D ′ Γ⊳Ξ Γ ⊢A t0 : q0 ⇛ s0 and an infinite reduction sequence ℓ1 ℓ2 ℓ3 1 2 3 (s0 , π0 ) −→N (s1 , π1 ) −→N (s2 , π2 ) −→N ··· . D′ D′ D′ Suppose that, for every i, si is not of the form si1 ◦ si2 . Then there exist {(qi , Ri , ti )}i∈ω such that ℓ1 ,R1 1 ℓ2 ,R2 2 ℓ3 ,R3 3 (s0 , π0 , q0 ) −→N (s1 , π1 , q1 ) −→N (s2 , π2 , q2 ) −→N ··· D′ D′ D′ and ℓ1 ℓ2 ℓ3 t0 −→∗D t1 −→∗D t2 −→∗D · · · . Furthermore q0 R1 R2 . . . is an infinite run of A over ℓ1 ℓ2 . . .. Proof. By using Lemma 41, one can define a family {(qi , Ri , ti , Γi )}i∈ω that satisfies ℓi ,Ri i (si−1 , πi−1 , qi−1 ) −→N (si , πi , qi ) D′♯ Γi ⊳ Ξ Γi ⊢A ti : qi ⇛ si e ℓ ti−1 −→∗D♯ ti by induction on i > 0. Since t0 and s0 do not contain marked symbols as well as the bodies of function definitions in D and D′ , ti and si dose not have marked ℓi ,Ri e ℓ i symbols for every i. Hence (si−1 , πi−1 , qi−1 ) −→N (s , πi , qi ) and ti−1 −→∗D ti for every D′ i i > 0. Furthermore q0 R1 . . . Ri is a run over ℓ1 . . . ℓi for every i. Hence the infinite sequence q0 R1 R2 . . . is an infinite run over ℓ1 ℓ2 . . .. So {(qi , Ri , ti )}i∈ω satisfies the requirements. ⊔ ⊓ 81 Lemma 42. Assume that Ξ ⊢A P ⇛ (P′ , Ω). Let P = (main, D) and P′ = (main′ , D′ ). The following conditions are equivalent. 1. ∃e ℓ ∈ InfTraces(P).∃R : run of A over e ℓ. maxInf(ΩA (R)) is even. 2. There exist π and an infinite reduction sequence ℓ1 ,R1 ℓ2 ,R2 (main′ , qI ) = (s0 , q0 ) −→∗D′ (s1 , q1 ) −→∗D′ · · · π such that R1 R2 · · · is an infinite sequence and maxInf(ΩA (R1 R2 · · · )) is even. Proof. ((1) ⇒ (2)) Assume an infinite reduction sequence ℓ1 π ℓ3 ℓ2 main = t0 −→D t1 −→D t2 −→D · · · and an infinite run qI R of A over ℓ1 ℓ2 . . . (here we implicitly assume that ℓ1 ℓ2 . . . is an infinite sequence). Since Ξ ⊢A P ⇛ (P′ , Ω), we have [Ξ] ⊢A main : qI ⇛ main′ . Then by Corollary 4, there exist {(qi , Ri , si )}i∈ω such that ℓ1 ,R1 ℓ2 ,R2 ℓ3 ,R3 (main′ , qI ) = (s0 , q0 ) −→∗D′ (s1 , q1 ) −→∗D′ (s2 , q2 ) −→∗D′ · · · and R = R1 R2 . . .. Then maxInf(ΩA (R1 R2 · · · )) = maxInf(ΩA (qI R1 R2 · · · )) = maxInf(ΩA (R)) is even by the assumption. ((2) ⇒ (1)) Assume an infinite reduction sequence π ℓ1 ,R1 ℓ2 ,R2 (main′ , qI ) = (s0 , q0 ) −→∗D′ (s1 , q1 ) −→∗D′ · · · such that R1 R2 · · · is an infinite sequence and maxInf(ΩA (R1 R2 · · · )) is even. By Corollary 5, there exists {ti }i such that ℓ1 ℓ2 main −→∗D t1 −→∗D · · · . By construction, R = qI R1 R2 · · · is an infinite run over ℓ1 ℓ2 · · ·. Since R is infinite by the assumption, ℓ1 ℓ2 · · · is also infinite. Then maxInf(ΩA (R)) = maxInf(ΩA (qI R1 R2 · · · )) = maxInf(ΩA (R1 R2 · · · )) is even by the assumption. ⊔ ⊓ Lemma 43. Assume that Ξ ⊢A D ⇛ D ′ Γ0 ⊳ Ξ Γ0 ⊢A t : q ⇛ ge s (ge s; π) e ℓ,R e ′ ′ and (ge s, π, q) −→N D′ (hθ,m u, π , q ). Then m = max(Ω(R)). Proof. Since Γ0 ⊢A t : q ⇛ ge s, there exist θ0 and m0 such that t = f t1 . . . tn e s =e s1 . . . e sn f : (θ0 , m0 , m0 ) ∈ Γ0 θ0 = ρ1 → · · · → ρn → q ∀i ∈ {1, . . . , n}. Γ0 ⊢A ti : ρi ⇛ e si . 82 N D′ u; π′ ) (hθ,m e Suppose D( f ) = λx1 . . . xn .t0 . Since Ξ ⊢A D ⇛ D′ , we have D′ ( fθ0 ,m0 ) = λe x1 . . . e xn .s0 [Ξ] ∪ [x1 : ρ1 ] ∪ · · · ∪ [xn : ρn ] ⊢A t0 : q ⇛ s0 where e xi = dup(xi , ρi ) for each i ∈ {1, . . . , n}. By easy induction on the structure of the derivation, we have [Ξ] ∪ [Ξ♯ ] ∪ [x1 : ρ1 ] ∪ · · · ∪ [xn : ρn ] ⊢A t♯0 : q ⇛ s♯0 . By using Weakening (Lemma 38) and Substitution Lemma (Lemma 39), we have [Ξ] ∪ [Ξ♯ ] ∪ Γ0 ⊢A [t1 /x1 ] . . . [tn /xn ]t♯0 : q ⇛ [e s1 /e x1 ] . . . [e sn /e xn ]s♯0 . Now, by the definition of call sequence, we have e ℓ ([e s1 /e x1 ] . . . [e sn /e xn ]s♯0 , π) −→N−1 (h♯θ,m ue′ , π′ ). D′♯ By Lemma 41, there exist R′ , t′ and Γ′ ⊳ Ξ such that e and ′ ℓ,R ([e s1 /e x1 ] . . . [e sn /e xn ]s♯0 , π, q) −→N−1 (h♯θ,m ue′ , π′ , q′ ) D′♯ Γ′ ∪ ([Ξ♯ ] ↑ max(ΩA (R′ ))) ⊢A t′ : q′ ⇛ h♯θ,m ue′ where q′ is the last state in R′ . e ℓ,R e ′ ′ By (ge s, π, q) −→N D′ (hθ,m u, π , q ), we have e ℓ,R ǫ,ǫ u, π′ , q′ ). s1 /e x1 ] . . . [e sn /e xn ]s0 , π, q) −→N−1 (ge s, π, q) −→D′ ([e D′ (hθ,m e Since the mark does not affect the induced run of the automaton, we have e ℓ,R ′′ ′ ′ ([e s1 /e x1 ] . . . [e sn /e xn ]s♯0 , π, q) −→N−1 D′ (s , π , q ) u except for marks). Comparing this for some s′′ (which is equivalent to hθ,m e reduction sequence with e ′ ℓ,R ([e s1 /e x1 ] . . . [e sn /e xn ]s♯0 , π, q) −→N−1 (h♯θ,m ue′ , π′ , q′ ) D′♯ given above, we conclude that s′′ = h♯θ,m ue′ and R = R′ . Thus Γ′ ∪ ([Ξ♯ ] ↑ max(Ω(R))) ⊢A t′ : q′ ⇛ h♯θ,m ue′ . Hence h♯ : (θ, m, m) ∈ ([Ξ♯] ↑ max(Ω(R))) and thus m = max(Ω(R)) by definition of [Ξ♯ ]. ⊔ ⊓ 83 Lemma 44. Assume that Ξ ⊢A P ⇛ (P′ , Ω). Let P = (main, D) and P′ = (main′ , D′ ). For each choice sequence π ∈ {L, R}ω, the following conditions are equivalent. 1. There exists an infinite reduction sequence π ℓ1 ,R1 ℓ2 ,R2 (main′ , qI ) = (s0 , q0 ) −→∗D′ (s1 , q1 ) −→∗D′ · · · such that R1 R2 · · · is an infinite sequence and maxInf(ΩA (R1 R2 · · · )) is even. 2. There exists an infinite call-sequence π main′ = g0θ0 ,m0 ũ0 k0 D′ g1θ1 ,m1 ũ1 k1 D′ g2θ2 ,m2 ũ2 k2 D′ ··· . such that maxInf(Ω(e g)) is odd. Proof. ((1) ⇒ (2)) Assume an infinite reduction sequence π ℓ1 ,R1 ℓ2 ,R2 (main′ , qI ) = (s0 , q0 ) −→∗D′ (s1 , q1 ) −→∗D′ · · · . By Corollary 1, we have a (unique) infinite call-sequence π main′ = g0θ0 ,m0 ũ0 k0 D′ g1θ1 ,m1 ũ1 k1 D′ g2θ2 ,m2 ũ2 k2 D′ ··· . So the given reduction sequence can be rewritten as π e ℓ1′ ,R′1 e ℓ3′ ,R′3 e ℓ2′ ,R′2 (main, qI ) = (g0θ0 ,m0 ũ0 , q′0 ) −→kD0′ (g1θ1 ,m1 ũ1 , q′1 ) −→kD1′ (g2θ2,m2 ũ2 , q′2 ) −→kD2′ · · · . Note that R1 R2 · · · = R′1 R′2 . . .. By Lemma 43, mi = max(ΩA (R′i )). Since maxInf(ΩA (R1 R2 · · · )) is even, maxInf(Ω(e g)) = maxInf(m1 + 1, m2 + 1, . . . ) = maxInf(max(ΩA (R′1 )) + 1, max(Ω(R′2 )) + 1, . . . ) = maxInf(max(ΩA (R′1 )), max(ΩA (R′2 )), . . . ) + 1 = maxInf(ΩA (R′1 R′2 . . . )) + 1 = maxInf(ΩA (R1 R2 . . . )) + 1 is odd. ((2) ⇒ (1)) Assume an infinite call-sequence π main′ = g0θ0 ,m0 ũ0 k0 D′ g1θ1 ,m1 ũ1 k1 D′ g2θ2 ,m2 ũ2 k2 D′ ··· . This is an infinite reduction sequence π e ℓ2 e ℓ1 e ℓ3 main′ = g0θ0 ,m0 ũ0 −→kD0′ g1θ1 ,m1 ũ1 −→kD1′ g2θ2 ,m2 ũ2 −→kD2′ · · · . 84 ℓ2 . . . is an infinite sequence. Then by By the assumption on the program, e ℓ1 e Corollary 5, π e ℓ1 ,R1 e ℓ2 ,R2 e ℓ3 ,R3 (main′ , q0 ) = (g0θ0 ,m0 ũ0 , q0 ) −→kD0′ (g1θ1 ,m1 ũ1 , q1 ) −→kD1′ (g2θ2 ,m2 ũ2 , q2 ) −→kD2′ · · · for some {(qi , Ri )}i . By Lemma 43, mi = max(ΩA (Ri )). Hence maxInf(ΩA (R1 R2 . . . )) = maxInf(max(ΩA (R1 )), max(ΩA (R2 )), . . . ) = maxInf(m1 m2 . . . ) = maxInf(Ω(e g)) − 1 Since maxInf(Ω(e g)) is odd, maxInf(ΩA (R1 R2 . . . )) is even. The sequence R1 R2 . . . e is infinite since ℓ1 e ℓ2 . . . is. ⊔ ⊓ Proof of Theorem 5 Assume that Ξ ⊢A P ⇒ (P0 , Ω). Then by Lemma 37, there exists P1 such that Ξ ⊢A P ⇛ (P1 , Ω) and P0 = [/◦]P1. Obviously |=csa (P1 , Ω) if and only if |=csa (P0 , Ω). Let P1 = (D1 , main1 ). We prove the lemma by establishing the equivalence of the following propositions: 1. InfTraces(P) ∩ L(A) , ∅. 2. ∃e ℓ ∈ InfTraces(P).∃R : run of A over e ℓ. maxInf(ΩA (R)) is even. 3. There exist π and an infinite reduction sequence π ℓ1 ,R1 ℓ2 ,R2 (main1 , qI ) = (s0 , q0 ) −→∗D1 (s1 , q1 ) −→∗D1 · · · such that R1 R2 · · · is an infinite sequence and maxInf(ΩA (R1 R2 · · · )) is even. 4. There exist π and an infinite call-sequence π main1 = g0θ0 ,m0 ũ0 k0 D1 g1θ1 ,m1 ũ1 k1 D1 g2θ2 ,m2 ũ2 k2 D1 ··· . such that maxInf(Ω(e g)) is odd. 5. ¬(|=csa (P1 , Ω)). 6. ¬(|=csa (P0 , Ω)). (1) ⇔ (2): By definition. (2) ⇔ (3): By Lemma 42. (3) ⇔ (4): By Lemma 44. (4) ⇔ (5): By definition (5) ⇔ (6): Obvious. E.5 Proof of Theorem 6 Here we provide a proof of Theorem 6. The proof also implies that the reduction from the temporal verification to HFLZ model checking can be performed in polynomial time. 85 For each simple type η, we define the set Θη of canonical intersection types by: Θ⋆ = {q | q ∈ QA } Θint = {int} Θint→κ = {int → θ | θ ∈ Θκ } V Θκ1 →κ2 = { {(θ1 , m) | θ1 ∈ Θκ1 , m ∈ {0, . . . , M}} → θ2 | θ2 ∈ Θκ2 } Here, M is the largest priority used in A. Note that, for any simple type κ (, int), the size of the set Θη is |Q|. For a simple type environment K , we define the canonical type environment ΓK ,m as: {x : int | x : int ∈ K } ∪ {x : (θ, m′ , m) | x : κ ∈ K , θ ∈ Θκ , m′ ∈ {0, . . . , M}}. Lemma 45. Suppose K ⊢ t : η. Then, for any m ∈ {0, . . . , M} and θ ∈ Θη , there exists t′ such that ΓK ,m ⊢A t : θ ⇒ t′ . Furthermore, t′ can be effectively constructed. Proof. The proof proceeds by induction on the derivation of K ⊢ t : η, with case analysis on the last rule used. – Case LT-Unit: In this case, t = ( ) and η = ⋆. By rule IT-Unit, we have ΓK ,m ⊢A t : θ ⇒ t′ for t′ = ( ). – Case LT-Var: In this case, t = x and K = K ′ , x : θ. If η = int, then θ = int and ΓK ,m = ΓK ′ ,m , x : int. By rule IT-VarInt, we have ΓK ,m ⊢A t : θ ⇒ t′ for t′ = xint, as required. If η , int, then ΓK ,m = ΓK ′ ,m ∪ {x : (θ′ , m′ , m) | θ′ ∈ Θη , m′ ∈ {0, . . . , M}} ∋ x : (θ, m, m). Therefore, by rule IT-Var, we have ΓK ,m ⊢A t : θ ⇒ t′ for t′ = xθ,m , as required. – Case LT-Int: In this case, t = n and η = int. By rule IT-Int, we have ΓK ,m ⊢A t : θ ⇒ t′ for t′ = n, as required. – Case LT-Op: In this case, t = t1 op t2 and θ = int, with K ⊢ ti : int for each i ∈ {1, 2}. By the induction hypothesis, there exists t′i such that ΓK ,m ⊢A ti : θ ⇒ t′i for each i. By rule IT-Op, we have ΓK ,m ⊢A t : θ ⇒ t′ for t′ = t′1 op t′2 , as required. – Case LT-Ev: In this case, t = event a; t1 and η = ⋆, with K ⊢ t1 : ⋆ and θ = q ∈ QA . Let δA (q, a) = {q1 , . . . , qn } and mi = max(m, ΩA (qi )). By the induction hypothesis, we have ΓK ,mi ⊢A t1 : qi ⇒ t′i for each i ∈ {1, . . . , n}. Since ΓK ,mi = ΓK ,m ↑ ΩA (qi ), by applying rule IT-Event, we obtain ΓK ,m ⊢A t : θ ⇒ t′ for t′ = event a; (t′1  · · · t′n ), as required. – Case LT-If: In this case, t = if p(t1 , . . . , tk ) then tk+1 else tk+2 and η = ⋆, with (i) K ⊢ ti : int for each i ∈ {1, . . . , k}, (ii) K ⊢ ti : ⋆ for each i ∈ {k+1, k+2}, and (iii) θ = q ∈ QA . By the induction hypothesis, we have: ΓK ,m ⊢A ti : int ⇒ t′i for each i ∈ {1, . . . , k}, and ΓK ,m ⊢A ti : θ ⇒ t′i for each i ∈ {k+1, k+2}. Thus, by rule IT-If, we have ΓK ,m ⊢A t : θ ⇒ t′ for t′ = if p(t′1 , . . . , t′k ) then t′k+1 else t′k+2 as required. – Case LT-App: In this case, t = t1 t2 with K ⊢ t1 : η2 → η and K ⊢ t2 : η2 . Let {(θ1 , m1 ), . .V . , (θk , mk )} = Θη2 × {0, . . . , M} with (θ1 , m1 ) < · · · < (θk , mk ), and let θ′ be 1≤i≤k (θi , mi ) → θ. By the induction hypothesis, we have ΓK ,m ⊢A t1 : θ′ ⇒ t′1 and ΓK ,m ↑ mi ⊢A t2 : θi ⇒ t′2,i for each i ∈ {1, . . . , k}. Thus, by rule IT-App, we have ΓK ,m ⊢A t : θ ⇒ t′ for t′ = t′1 t′2,1 · · · t′2,k as required. 86 – Case LT-NonDet: In this case, t = t1 t2 and η = ⋆, with K ⊢ ti : ⋆ for each i ∈ {1, 2}. By the assumption θ ∈ Θη , θ ∈ QA . By the induction hypothesis, we have ΓK ,m ⊢A ti : θ ⇒ t′i for each i. By rule IT-NonDet, we have ΓK ,m ⊢A t : θ ⇒ t′ for t′ = t′1 t′2 as required. ⊔ ⊓ Theorem 6 is an immediate corollary of the lemma above. Proof (Theorem 6). Let P = ({ f1 = t1 , . . . , fn = tn }, t), and K be a simple type environment for P, i.e., K ⊢ P. Let Ξ be: { fi : (θ, m) | i ∈ {1, . . . , n}, θ ∈ ΘK ( fi ) }. For each fi : (θ, m) ∈ Ξ, let t′i,θ and t′ be the terms given by Lemma 45, such that ΓK ,0 ⊢A ti : θ ⇒ ti,θ and ΓK ,0 ⊢A t : qI ⇒ t′ . Let P′ be: ({ fi,θ,m = t′i,θ | fi : (θ, m) ∈ Ξ}, t′), and Ω be { fi,θ,m 7→ m + 1 | fi : (θ, m) ∈ Ξ}. Then Ξ ⊢A P ⇒ (P′ , Ω) holds, and Ξ, P′ , Ω can be effectively constructed as described above. ⊓ ⊔ Any program P = ({ f1 = t1 , . . . , fn = tn }, t) can be normalized (with an at most polynomial blow-up of the size), so that each of ti and t is one of the following forms: – – – – () y (z1 x1,1 · · · x1,ℓ1 ) · · · (zk xk,1 · · · xk,ℓk ) event a; (y x1 · · · xk ) if p(t1 , . . . , tk ) then (y x1 · · · xk ) else (y′ x′1 · · · x′k′ ) where k, k′ , ℓi may be 0. The normalization can be performed by adding auxiliary functions: See [20], Section 4.3 for a similar normalization. For a normalized program P, the size of P′ obtained by the transformation in the proof above is polynomial in the size of P and |QA |. Furthermore, P′ and Ω can be constructed in polynomial time. Thus, the whole reduction from temporal property verification to call-sequence analysis (hence also to HFLZ model checking) can be performed in polynomial time. F An Example of the Translation of Section 7 Here we show derivation trees for the translation in Example 11. The body of g is translated as follows (where we omit irrelevant type bindings), where Γ0 = k : (qa , 0, 0), k : (qb , 1, 0). k : (qa , 0, 0), k : (qb , 1, 0) ⊢A k : qa ⇒ kqa ,0 k : (qa , 0, 1), k : (qb , 1, 1) ⊢A k : qb ⇒ kqb ,1 Γ0 ⊢A (event a; k) : qa ⇒ (event a; kqa ,0 ) Γ0 ⊢A (event b; k) : qa ⇒ (event a; kqb ,1 ) Γ0 ⊢A (event a; k)(event b; k) : qa ⇒ (event a; kqa ,0 )(event b; kqb ,1 ) 87 The body of f is translated as follows. Γ1 , x : int ⊢A x : int ⇒ xint Γ1 , x : int ⊢A 0 : int ⇒ 0 π1 π2 Γ1 , x : int ⊢A if x > 0 then g ( f (x − 1)) else (event b; f 5) : qa ⇒ t f,qa Here, Γ1 is: g : ((qa , 0) ∧ (qb , 1) → qa , 0, 0), g : ((qa , 0) ∧ (qb , 1) → qb , 0, 0), f : (int → qa , 0, 0), f : (int → qb , 1, 0) and π1 and π2 are: Γ1 , x : int ⊢A g : (qa , 0) ∧ (qb , 1) → qa ⇒ g(qa ,0)∧(qb ,1)→qa ,0 π1 = π3 π4 Γ1 , x : int ⊢A g ( f (x − 1)) : qa ⇒ g(qa ,0)∧(qb ,1)→qa ,0 ( fint→qa ,0 (xint − 1)) ( fint→qb ,1 (xint − 1)) π2 = Γ1 ↑ 1, x : int ⊢A f : int → qb ⇒ fint→qb ,1 Γ1 ↑ 1, x : int ⊢A 5 : int ⇒ 5 Γ1 ↑ 1, x : int ⊢A f 5 : qa ⇒ fint→qb ,1 5 Γ1 , x : int ⊢A event b; f 5 : qa ⇒ event b; fint→qb ,1 5 π3 = π4 = G Γ1 ↑ 0, x : int ⊢A f : int → qa ⇒ fint→qa ,0 ··· Γ1 ↑ 0, x : int ⊢A x − 1 : int ⇒ xint − 1 Γ1 ↑ 0, x : int ⊢A f (x − 1) : qa ⇒ fint→qa ,0 (xint − 1) Γ1 ↑ 1, x : int ⊢A f : int → qb ⇒ fint→qb ,1 ··· Γ1 ↑ 1, x : int ⊢A x − 1 : int ⇒ xint − 1 Γ1 ↑ 1, x : int ⊢A f (x − 1) : qb ⇒ fint→qb ,1 (xint − 1) Proving HFLZ formulas in Coq Given program verification problems, the reductions presented in Sections 5–7 yield HFLZ model checking problems, which may be thought as a kind of “verification conditions” (like those for Hoare triples). Though we plan to develop automated/semi-automated tools for discharging the “verification conditions”, we demonstrate here that it is also possible to use an interactive theorem prover to do so. Here we use Coq proof assistant, and consider HFLNat (HFL extended with natural numbers) instead of HFLZ . Let us consider the termination of the following program: let sum n k = if n<=0 then k 0 else sum (n-1) (fun r -> k(n+r)) in sum m (fun r->()) 88 Here, we assume m ranges over the set natural numbers. The translation in Section 5.2 yields the following HFLNat formulas. (µsum.λn.λk.(n ≤ 0 ⇒ k 0) ∧ (n > 0 ⇒ sum (n − 1) λr.k(n + r))) m (λr.true). The goal is to prove that for every m, the formula is satisfied by the trivial model L0 = ({s⋆ }, ∅, ∅, s⋆). In order to avoid the clumsy issue of representing variable bindings in Coq, we represent the semantics of the above formula in Coq. The following definitions correspond to those of DL,τ in Section 2. (* syntax of simple types: "arint t" and "ar t1 t2" represent nat->t and t1->t2 respectively *) Inductive ty: Set := o: ty | arint: ty -> ty | ar: ty -> ty -> ty. (* definition of semantic domains, minus the monotonicity condition *) Fixpoint dom (t:ty): Type := match t with o => Prop | arint t’ => nat -> dom t’ | ar t1 t2 => (dom t1) -> (dom t2) end. Here, we use Prop as the semantic domain DL0 ,• = {∅, {s⋆}} and represent {s⋆ } as a proposition True. Above, we have deliberately omitted the monotonicity condition, which is separately defined by induction on simple types, as follows. Fixpoint ord (t:ty) {struct t}: dom t -> dom t -> Prop := match t with o => fun x: dom o => fun y: dom o => (x -> y) | arint t’ => fun x: dom (arint t’) => fun y: dom (arint t’) => forall z:nat, ord t’ (x z) (y z) | ar t1 t2 => fun x: dom (ar t1 t2) => fun y: dom (ar t1 t2) => forall z w:dom t1, ord t1 z z -> ord t1 w w -> ord t1 z w -> ord t2 (x z) (y w) end. Definition mono (t: ty) (f:dom t) := ord t f f. 89 Here, ord τ corresponds to ⊑L0 ,τ in Section 2,14 and the monotonicity condition on f is expressed as the reflexivity condition ord τ f f . We can then state the claim that the sum program is terminating for every m as the following theorem: Definition sumt := arint (ar (arint o) o). Definition sumgen := fun sum: dom sumt => fun n:nat=> fun k:nat->Prop => (n<=0 -> k 0) /\ (n>0 -> sum (n-1) (fun r:nat=>k(r+n))). Theorem sum_is_terminating: forall sum: dom sumt, forall FPsum: (* sum is a fixpoint of sumgen *) (forall n:nat, forall k:nat->Prop, sum n k <-> sumgen sum n k), forall LFPsum: (* sum is the least fixpoint of sumgen *) (forall x:dom sumt, mono sumt x -> ord sumt (sumgen x) x -> ord sumt sum x), forall m:nat, sum m (fun r:nat => True). (* can be automatically generated up to this point *) Proof. (* this part should be filled by a user *) ... Qed. Here, sumgen is the semantics of the argument of the µ-operator (i.e., λsum.λn.λk.(n ≤ 0 ⇒ k 0) ∧ (n > 0 ⇒ sum (n − 1) λr.k(n + r))), and the first three “forall ...” assumes that sum is the least fixpoint of it, and the last line says that sum m λx.True is equivalent to True for every m. Note that except the proof (the part “...”), all the above script can be automatically generated based on the development in the paper (like Why3[10], but without any invariant annotations). The following is a proof of the above theorem. Proof. intros. (* apply induction on m *) induction m; apply FPsum; unfold sumgen; simpl;auto. split; auto. omega. (* induction step *) 14 Note, however, that since dom t may be inhabited by non-monotonic functions, “ord τ” is not reflexive. 90 assert (m-0=m); try omega. rewrite H; auto. Qed. More examples are found at http://www-kb.is.s.u-tokyo.ac.jp/˜koba/ papers/hfl_in_coq.zip. The Coq proofs for some of those examples are much longer. For each example, however, there are only a few places where human insights are required (like “induction m” above). We, therefore, expect that the proofs can be significantly shortened by preparing appropriate libraries. H HORS- vs HFL-based Approaches to Program Verification In this section, we provide a more detailed comparison between our new HFLbased approach and HORS-based approaches [19, 20, 25, 29, 33, 35, 47] to program verification. Some familiarity with HORS-based approaches may be required to fully understand the comparison. HORS model checking algorithms [23, 34] usually consist of two phases, one for computing a kind of higher-order “procedure summaries” in the form of variable profiles [34] or intersection types [23] (that summarize which states are visited between two function calls), and the other for solving games, which consists in nested least/greatest fixpoint computations. In the case of finite-data programs, the combination of the two phases provides a sound and complete verification algorithm [23, 34]. To deal with infinite-data programs, however, the HORS-based approaches [19, 20, 25, 29, 33, 35, 47] had to apply various transformations to map verification problems to HORS model checking problems in a sound but incomplete manner (incompleteness is inevitable because, in the presence of values from infinite data domains, the former is undecidable whereas the latter is decidable), as illustrated on the lefthand side of Figure 10. A problem about this approach is that the second phase of HORS model checking – nested least/greatest fixpoint computations – actually does not help much to solve the original problem, because least fixpoint computations are required for proving liveness (such as termination), but the liveness of infinite-data programs usually depends on properties about infinite data domains such as “there is no infinite decreasing sequence of natural numbers,” which are not available after the transformations to HORS. For this reason, the previous techniques for proving termination [29] and fair termination [33] used HORS model checking only as a backend of a safety property checker, where only greatest fixpoint computations are performed in the second phase; reasoning about liveness was performed during the transformation to HORS model checking. As shown on the righthand side of Figure 10, in our new HFL-based approach, we have extended the first phase of HORS model checking – the computation of higher-order procedure summaries – to deal with infinite data programs, and moved it up front; we then formalized the remaining problems as HFL model checking problems. That can be viewed as the essence of the reduction in Section 7, and the reductions in Sections 5 and 6 are those for degenerate cases. Advantages of the new approach include: (i) necessary information on 91 fair termination verification problems for arbitrary Z-regular (linear-time) properties termination problem-specific transformations fair non-termination non-termination safety computing higherorder summaries (for infinite-data programs) predicate abstractions HFLZ model checking (nested fixpoint computations over infinite data domains) computing higherorder summaries game solving (nested fixpoint computations over finite data domains) HORS model checking Fig. 10. Comparison between HORS- and HFL-based approaches. In the HORS-based approaches (shown on the lefthand side), there are actually some feedback loops (due to counterexample-guided abstraction refinement, etc.), which are omitted. The right hand side shows the approach based on the reduction in Section 7. In the reductions in Section 5 and 6, the first phase is optimized for degenerate cases. 92 infinite data is available in the second phase of least/greatest computations (cf. the discussion above on the HORS-based approach), and (ii) various verification problems boil down to the issue of how to prove least/greatest fixpoint formulas; thus we can reuse and share the techniques developed for different verification problems. The price to pay is that the first phase (especially a proof of its correctness) is technically more involved, because now we need to deal with infinite-data programs instead of HORS’s (which are essentially finitedata functional programs). That explains long proofs in Appendices (especially Appendix D and E). 93
6cs.PL
Hybrid models of the cell cycle molecular machinery Vincent Noel IRMAR UMR 6625, University of Rennes 1, Rennes, France Dima Grigoriev CNRS, Mathématiques, Université de Lille, 59655, Villeneuve d’Ascq, France Sergei Vakulenko Saint Petersburg State University of Technology and Design, St.Petersburg, Russia Ovidiu Radulescu DIMNP UMR CNRS 5235, University of Montpellier 2, Montpellier, France Piecewise smooth hybrid systems, involving continuous and discrete variables, are suitable models for describing the multiscale regulatory machinery of the biological cells. In hybrid models, the discrete variables can switch on and off some molecular interactions, simulating cell progression through a series of functioning modes. The advancement through the cell cycle is the archetype of such an organized sequence of events. We present an approach, inspired from tropical geometry ideas, allowing to reduce, hybridize and analyse cell cycle models consisting of polynomial or rational ordinary differential equations. 1 Introduction Hybrid systems are widely used in automatic control theory to cope with situations arising when a finitestate machine is coupled to mechanisms that can be modeled by differential equations [13]. It is the case of robots, plant controllers, computer disk drives, automated highway systems, flight control, etc. The general behavior of such systems is to pass from one type of smooth dynamics (mode) described by one set of differential equations to another smooth dynamics (mode) described by another set of differential equations. The command of the modes can be performed by changing one or several discrete variables. The mode change can be accompanied or not by jumps (discontinuities) of the trajectories. Depending on how the discrete variables are changed, there may be several types of hybrid systems: switched systems [22], multivalued differential automata [26], piecewise smooth systems [9]. Notice that in the last case, the mode changes when the trajectory attains some smooth manifolds. In these examples, the changes of discrete variables and the evolution of continuous variables are deterministic. The class of hybrid systems can be extended by considering stochastic dynamics of both continuous and discrete variables, leading to piece-wise deterministic processes, switched diffusions or diffusions with jumps [19, 6, 5, 23, 4]. Hybrid, differential, or stochastic Petri nets provide equivalent descriptions of the dynamics and were also used in this context [7]. The use of hybrid models in systems biology can be justified by the temporal and spatial multiscaleness of biological processes, and by the need to combine qualitative and quantitative approaches to study dynamics of cell regulatory networks. Furthermore, hybrid modelling offers a good compromise between realistic description of mechanisms of regulation and possibility of testing the model in terms of state reachability and temporal logics [12, 14]. Threshold dynamics of gene regulatory networks [2, 20] or of excitable signaling systems [28] has been modelled by piecewise-linear and piecewiseaffine models. These models have relatively simple structure and can, in certain cases, be identified from Ezio Bartocci and Luca Bortolussi (Eds.): HSB 2012 EPTCS 92, 2012, pp. 88–105, doi:10.4204/EPTCS.92.7 V. Noel, D. Grigoriev, S. Vakulenko, & O. Radulescu 89 data [18, 8].Some methods were proposed for computing the set of reachable states of piecewise affine models [3]. Among the applications of hybrid modeling, one of the most important is the cell cycle regulation. The machinery of the cell cycle, leading to cell division and proliferation, combines slow growth, spatiotemporal re-organisation of the cell, and rapid changes of regulatory proteins concentrations induced by post-translational modifications. The advancement through the cell cycle is a well defined sequence of stages, separated by checkpoint transitions. This justifies hybrid modelling approaches, such as Tyson’s hybrid model of the mammalian cell cycle [24]. This model is based on a Boolean automaton whose discrete transitions trigger changes of kinetic parameters in a set of ODEs. The model has been used to reproduce flow cytometry data. Instead of building the hybrid model from scratch, another strategy is to identify hybrid models from experimental or artificial time series [16, 17, 1]. The resulting cell cycle hybrid models can be used for hypothesis testing, as they are or as parts of larger, integrated models. In this paper we develop ideas first introduced in [15]. We discuss how a given model of the cell cycle, based on ODEs, can be hybridized. The hybridization, based on a tropical geometry heuristics, unravels commonalities of cell cycle models. These systems combine quasi-equilibrium states, represented by slow invariant manifolds and excitability, represented by rapid transitions to and from these manifolds. With respect to [15] we introduce several new concepts and provide rigorous justification of the procedures. Two general hybridization procedures called tropicalizations are introduced in section 2. The tropicalized dynamics is guaranteed to be a good approximation for polynomial or rational systems with well separated terms and that satisfy a condition called permanency. In subsection 2.2 we introduce the tropical equilibration as a method to test permanency. In section 3 we apply these methods to a cell cycle biochemical network model. 2 Tropical geometry and hybridization 2.1 General settings In chemical kinetics, the reagent concentrations satisfy ordinary differential equations: dxi = Fi (x), 1 ≤ i ≤ n. dt (1) Rather generally, the rates are rational functions of the concentrations and read Fi (x) = Pi (x)/Qi (x), (2) where Pi(x) = ∑α ∈Ai ai,α xα , Qi (x) = ∑β ∈Bi bi,β xβ , are multivariate polynomials. Here xα = x1α1 x2α2 . . . xαn n , β β β xβ = x1 1 x2 2 . . . xn n , ai,α , bi,β , are nonzero real numbers, and Ai , Bi are finite subsets of Nn . Special case are represented by Fi (x) = Pi+ (x) − Pi− (x), (3) α ± where Pi+ (x), Pi− (x) are Laurent polynomials with positive coefficients, Pi± (x) = ∑α ∈A±i a± i,α x , ai,α > n 0, A± i are finite subsets of Z . In multiscale biochemical systems, the various monomials of the Laurent polynomials have different orders, and at a given time, there is only one or a few dominating terms. Therefore, it could make sense to replace Laurent polynomials with positive real coefficients ∑α ∈A aα xα , by max-plus polynomials maxα ∈A {log(aα )+ < log(x), α >}. Hybrid models of the cell cycle 90 This heuristic can be used to associate a piecewise-smooth hybrid model to the system of rational ODEs (1), in two different ways. The first method was proposed in [15] and can be applied to any rational ODE system defined by (1),(2): Definition 2.1 We call complete tropicalization of the smooth ODE system (1),(2) the following piecewisesmooth system: dxi = DomPi (x)/DomQi (x), (4) dt where Dom{ai,α xα }α ∈Ai = sign(ai,αmax )exp[maxα ∈Ai {log(|ai,α |)+ < u, α >}]. u = (logx1 , . . . , logxn ), and ai,αmax , αmax ∈ Ai denote the coefficient of the monomial for which the maximum is attained. In simple words, Dom renders the monomial of largest absolute value, with its sign. The second method,proposed in [21], applies to the systems (1),(3). Definition 2.2 We call two terms tropicalization of the smooth ODE system (1),(3) the following piecewisesmooth system: dxi = DomPi+ (x) − DomPi−(x), (5) dt The two-terms tropicalization was used in [21] to analyse the dependence of steady states on the model parameters. The complete tropicalization was used for the study of the model dynamics and for the model reduction [15]. For both tropicalization methods, for each occurrence of the Dom operator, one can introduce a tropical manifold, defined as the subset of Rn where the maximum in Dom is attained by at least two terms. For instance, for n = 2, such tropical manifold is made of points, segments connecting these points, and half-lines. The tropical manifolds in such an arrangement decompose the space into sectors, inside which one monomial dominates all the others in the definition of the reagent rates. The study of this arrangement give hints on the possible steady states and attractors, as well as on their bifurcations. 2.2 Justification of the tropicalization and some estimates In the general case, the tropicalization heuristic is difficult to justify by rigorous estimates, however, this is possible in some cases. We state here some results in this direction. To simplify, let us consider the class of polynomial systems, corresponding to mass action law chemical kinetics: dxi = Pi (x, ε ) = dt M ∑ Mi j (x, ε ), Mi j = Pi j (ε )xαi j (6) j=1 where αi j are multi-indices, and ε is a small parameter. So, the right hand side of (6) is a sum of monomials. We suppose that coefficients Pi j have different orders in ε : Pi j (ε ) = ε bi j P̄i j , (7) where bi j 6= bi′ j′ for (i, j) 6= (i′ , j′ ). We also suppose that the cone R> = {x : xi ≥ 0} is invariant under dynamics (6) and initial data are positive: xi (0) > δ > 0. The terms (7) can have different signs, the ones with P̄i j > 0 are production terms, and those with P̄i j < 0 are degradation terms. V. Noel, D. Grigoriev, S. Vakulenko, & O. Radulescu 91 From the biochemical point of view, the choice (7) is justified by the fact that biochemical processes have many, well separated timescales. Furthermore, we are interested in biochemical circuits that can function in a stable way. More precisely, we use the permanency concept, borrowed from ecology (the Lotka -Volterra model, see for instance [25]). Definition 2.3 The system (6) is permanent, if there are two constants C− > 0 and C+ > 0, and a function T0 , such that C− < xi (t) < C+ , f or all t > T0 (x(0)) and f or every i. (8) We assume that C± and T0 are uniform in (do not depend on) ε as ε → 0. For permanent systems, we can obtain some results justifying the two procedures of tropicalization. Proposition 2.4 Assume that system (6) is permanent. Let x, x̂ be the solutions to the Cauchy problem for (6) and (4) (or (5)), respectively, with the same initial data: x(0) = x̂(0). Then the difference y(t) = x(t) − x̂(t) satisfies the estimate |y(t)| < C1 ε γ exp(bt), γ > 0, (9) where the positive constants C1 , b are uniform in ε . If the original system (6) is structurally stable in the domain ΩC− ,C+ = {x : C− < |x| < C+ }, then the corresponding tropical systems (4) and (5) are also permanent and there is an orbital topological equivalence x̄ = hε (x) between the trajectories x(t) and x̄(t) of the corresponding Cauchy problems. The homeomorphism hε is close to the identity as ε → 0. The proof of the estimate (9) follows immediately by the Gronwall lemma. The second assertion follows directly from the definition of structural stability. Permanency property is not easy to check. In the case of systems (6) we can make a renormalization xi = ε ai x̄i (10) and suppose that (8) holds for x̄i with Ci± uniform in ε . We seek for renormalization exponents ai such that only a few terms dominate all the others, for each i-th equation (6) as ε → 0. Let us denote the number of terms with minimum degree in ε for i-th equation as mi . Naturally, 1 ≤ mi ≤ Mi . After renormalization, we remove all small terms that have smaller orders in ε as ε → 0. We can call this procedure tropical removing. The system obtained can be named tropically truncated system. Let us denote αli j the l th coefficient of the multi-index α~i j . If all mi = 1 then we have the following truncated system d x̄i i j(i) = ε µi Fi (x̄), Fi = pi j(i) x̄α , (11) dt where n µi = γi j(i) ∑ αl i j(i) al (12) f or all j 6= j(i). (13) l=1 and n µi > γi j ∑ αli j al l=1 Hybrid models of the cell cycle 92 If all mi = 2, in order to find possible renormalization exponents ai , it is necessary to resolve a family of linear programming problem. Each problem is defined by a set of pairs ( j(i), k(i)) such that j(i) 6= k(i). We define µi by n µi = γi j(i) + ∑ αl i j(i) l=1 n al = γik(i) + ∑ αl ik(i) al (14) f or all j 6= j(i), k(i). (15) l=1 and obtain the system of the following inequalities n µi ≥ γi j + ∑ αli j al l=1 The following straightforward lemma gives a necessary condition of permanency of the system (6). Lemma 2.5 Assume a tropically truncated system is permanent. Then, for each i ∈ {1, . . . , n}, the i-th equation of this system contains at least two terms. The terms should have different signs for coefficients pi j , i.e., one term should be a production one, while another term should be a degradation term. We call “tropical equilibration”, the condition in Lemma 2.5. This condition means that permanency is acquired only if at least two terms of different signs have the maximal order, for each equation of the system (6). This idea is not new, and can be traced back to Newton. The tropical equilibration condition can be used to determine the renormalization exponents, by the following algorithm. Step 1. For each i let us choose a pair ( j(i), k(i)) such that j, i ∈ {1, . . . , mi } and j < k. The sign of the corresponding terms should be different. Step 2. We resolve the linear system of algebraic equations n γi j(i) − γik(i) = − ∑ αl l=1 i j(i) n al + ∑ αl ik(i) al , (16) l=1 for al , together with the inequalities (15). Notice that although that Step 2 has polynomial complexity, the tropical equilibration problem has an exponential number of choices at Step 1. Assume that, as a result of this procedure, we obtain the system d x̄i = ε µi (Fi+ (x̄) − Fi− (x̄)), dt ij Fi± = pi j± x̄α± . (17) One can expect that, in a ”generic” case1 , all µi are mutually different, namely 0 = µ1 < µ2 < ... < µn−1 < µn . (18) We can now state a sufficient condition for permanency. Let us consider the first equation (17) with i = 1 and let us denote y = x̄1 , z = (x̄2 , ..., x̄n )tr . In this notation, the first equation becomes dy = f (y) = b1 (z)yβ1 − b2 (z)yβ2 , dt b1 , b2 > 0, βi ∈ R. (19) Since µ2 > 0, one has that z(t) is a slow function of time and thus we can suppose that bi are constants (this step can be rendered rigorous by using the concept of invariant manifold and methods from [11]). 1 supposing that multi-indices α ~i j are chosen uniformly, by generic we understand almost always except for cases of vanishing probability, see also [10] V. Noel, D. Grigoriev, S. Vakulenko, & O. Radulescu 93 The permanency property can be then checked in an elementary way. All rest points of (19) are roots of f . If f > 0, y(t) is an increasing time function and if f < 0, y(t) is a decreasing time function. A single root y1 of f within (0, +∞) is given by y1 = b2 d , b1 d= 1 . β1 − β2 (20) These properties entail the following result: Lemma 2.6 Equation 19 has the permanence property if and only if 0 < β1 < β2 , i β1 < β2 < 0, ii or or β1 < 0, β2 > 0. iii For fixed z, in these cases we have y(t, z) → y0 , as t → ∞. The generic situation described by the conditions (18) lead to trivial “chain-like” relaxation towards a point attractor, provided that we have permanency at each step. This result is the nonlinear analogue of the similar result that monomolecular networks with total separation relax as chains and can only have stable point attractors [10]. The following theorem describes a less trivial situation, when limit cycles are possible. Theorem 2.7 Assume 0 = µ1 < µ2 < ... < µn−1 ≤ µn holds. If the procedure, described above, leads to the permanency property at each step, where i = 1, 2, ..., n − 2, and the last two equations have a globally attracting hyperbolic rest point or globally attracting hyperbolic limit cycle, then the tropically truncated system is permanent and has an attractor of the same type. Moreover, for sufficiently small ε the initial system also is permanent for initial data from some appropriate domain Wε ,a,A and has an analogous attracting hyperbolic rest point (limit cycle) close to the attractor of the truncated system. If the rest point (cycle) is not globally attracting, then we can say nothing on permanency but, for sufficiently small ε , the initial system still has an analogous attracting hyperbolic rest point (limit cycle) close to the attractor of truncated system and the same topological structure. Finally, let us note that tropical equilibrations with permanency imply the existence of invariant manifolds. This allows to reduce the number of variables of the model while preserving good accuracy in the description of the dynamics. The following Lemma is useful in this aspect. Lemma 2.8 Consider the system dy = f (y) = b1 (z)yβ1 − b2 (z)yβ2 , dt dz = λ F(y, z), dt b1 , b2 > 0, βi ∈ R. (21) (22) Hybrid models of the cell cycle 94 where z ∈ Rm , λ > 0 is a parameter and the function F enjoys the following properties. This function lies in an Hölder class F ∈ C1+r , r > 0, and the corresponding norms are uniformly bounded in Ω = (0, +∞) ×W , for some open domain W ⊂ Rm : |F|C1+r (Ω ) < C2 . Assume one of conditions i, ii, iii of Lemma 2.6 holds. We also suppose that bi are smooth functions of z for all z such that |z| > δ0 > 0. Assume that z ∈ W implies |z| > δ0 . Then, for sufficiently small λ < λ0 (C2 , b1 , b2 , β1 , β2 , δ ) equations (21), (23) have a locally invariant and locally attracting manifold y = Y (z, λ ), Y ∈ C1+r (W ), (23) and Y has the asymptotics Y (z, λ ) = y1 (z) + Ỹ , where Ỹ ∈ C1+r (W ), |Ỹ (z, λ )|C1+r (W ) < Cs λ s , s > 0. (24) (25) Proof. The proof is standard, follows from Theorems in [11], Ch. 9. 3 A paradigmatic cell cycle model and its tropicalization 3.1 Description of the model We study here the cell cycle model proposed by Tyson [27]. This model mimics the interplay between cyclin and cyclin dependent kinase cdc2 (forming the maturation promoting factor MPF complex) during the progression of the cell cycle. The model demonstrates that this biochemical system can function as an oscillator, or converge to a steady state with large MPF concentration, or behave as an excitable switch. The three regimes can be associated to early embryos rapid division, metaphase arrest of unfertilized eggs, and growth controlled division of somatic cells, respectively. This model takes into account autocatalytic activity of MPF (positive feed-back). It can be described as a nonlinear cycle of biochemical reactions and corresponds to the following set of differential equations: y′1 = ε −3 k9 y2 − ε −6k8 y1 + k6 y3 , y′3 = ε 2 k4′ y4 + ε −2 k4 y4 y23 − k6 y3 , y′5 = ε 2 k1 − ε −2 k3 y2 y5 , y′2 = ε −6 k8 y1 − ε −3k9 y2 − ε −2k3 y2 y5 , y′4 = −ε 2 k4′ y4 − ε −2k4 y4 y23 + ε −2 k3 y2 y5 , (26) Here ki > 0 are rate constants, yi , i ∈ [1, 5] are concentrations of cdc2, p-cdc2 (phosphorylated kinase), cyclin-p:cdc2 complex (active MPF), cyclin-p:cdc2-p complex (inactive MPF), and cyclin, respectively. With respect to the original model we have introduced a small parameter ε > 0 to cope with the order of the rate constants (ε = 0.1 in the original model). The system (26) has the conservation law y1 (t) + y2 (t) + y3 (t) + y4 (t) = 1, where the value 1 (total initial concentration of kinase cdc2) was chosen by convenience. (27) V. Noel, D. Grigoriev, S. Vakulenko, & O. Radulescu 95 3.2 Tropical equilibrations and model reduction Let us apply the tropical equilibration principle. To this aim, we renormalize the variables, yi = ε ai ȳi . (28) Let us substitute these relations into the system of equations. As a result, we obtain ȳ′1 = ε −3+a2 −a1 k9 ȳ2 − ε −6 k8 y1 + k6 ε a3 −a1 ȳ3 , ȳ′3 = ε 2+a4 −a3 k4′ ȳ4 + ε −2+a3+a4 k4 ȳ4 ȳ23 − k6 ȳ3 , ȳ′5 = ε 2−a5 k1 − ε −2+a2 k3 ȳ2 ȳ5 . ȳ′2 = ε −6+a1 −a2 k8 ȳ1 − ε −3 k9 ȳ2 − ε −2+a5 k3 ȳ2 ȳ5 , ȳ′4 = −ε 2 k4′ ȳ4 − ε −2+2a3 k4 ȳ4 ȳ23 + ε −2+a2+a5 −a4 k3 ȳ2 ȳ5 , (29) The system (29) has the conservation law ε a1 ȳ1 (t) + ε a2 ȳ2 (t) + ε a3 ȳ3 (t) + ε a4 ȳ4 (t) = 1. (30) In order to compute the exponents ai we use tropical equilibrations together with the conservation law (30). There are 24 variants of tropical equilibrations. To our surprise, there is only one solution for the exponents values. We can show that all possible equilibrations of the variables y3 , y4 and y5 uniquely set the values of two exponents, a3 = 2, a4 = 0. Let us consider the variants with respect to the equilibrations of the variables y1 and y2 . Denoting by Ti the ith term in the equation, we have the following situations: 1) In eq. for ȳ1 : T 1 = T 2, T 3 <= T 1, In eq. for ȳ2 : T 1 = T 2, T 3 <= T 1. 2) In eq. for ȳ1 : T 1 = T 2, T 3 <= T 1, In eq. for ȳ2 : T 1 = T 3, T 2 <= T 3. 3) In eq. for ȳ1 : T 2 = T 3, T 1 <= T 2, In eq. for ȳ2 : T 1 = T 2, T 3 <= T 1. 4) In eq. for ȳ1 : T 2 = T 3, T 1 <= T 2, In eq. for ȳ2 : T 1 = T 3, T 2 <= T 1. In the variant 1 (Case I) the tropical equilibrations do not fix the values of the exponent a5 and we get a1 = 7 − a5 , a2 = 4 − a5 , a3 = 2, a4 = 0, a5 ≥ −1. (31) In the variants 2,3,4 (Case II) the exponents are uniquely determined from equilibrations and we obtain a1 = 8, a2 = 5, a3 = 2, a4 = 0, a5 = −1. (32) However, (32) and the conservation law (30) are incompatible, therefore, the case II can be rejected. In the case I the conservation law takes the form ε 7−a5 ȳ1 (t) + ε 4−a5 ȳ2 (t) + ȳ4 (t) = 1 + o(1), as ε → 0. Assuming that ȳ2 (0) = O(1) and ȳ4 6= 1 (it is reasonable, since it is a ”generic case”), we obtain a5 = 4. Thus, the only possible situation is variant 1 (Case I) and the corresponding set of exponents is: a1 = 3, a2 = 0, a3 = 2, a4 = 0, a5 = 4. (33) Let us note that the terms T 1 and T 2 in the equations for the variables y1 , y2 correspond to direct and reverse rates of a phosphorylation/dephosphorylation cycle transforming y1 into y2 and back. Thus, biochemically, (Case I) corresponds to the quasi-equilibrium of this cycle. Furthermore, the equilibration of all the variables leads to the exponents (31). In this case, 2 + a4 − a3 = −2 + a3 + a4 = 0, 2 = −2 + 2a3 = −2 + a2 + a5 − a4 , meaning that the tropical equilibrations of the variables y3 , y4 are triple (in each equation, all three terms have the same order). Hybrid models of the cell cycle 96 We finally obtain the following renormalized system ȳ′1 = ε −6 (k9 ȳ2 − k8 ȳ1 ) + k6 ε −1 ȳ3 , ȳ′3 ȳ′5 = k4′ ȳ4 + k4 ȳ4 y23 − k6 ȳ3 , = ε −2 (k1 − k3 ȳ2 ȳ5 ). ȳ′2 = ε −3 (k8 ȳ1 − k9 ȳ2 ) − ε 2 k3 ȳ2 ȳ5 , ȳ′4 = ε 2 (−k4′ ȳ4 − k4 ȳ4 ȳ23 + k3 ȳ2 ȳ5 ), (34) The structure of the system (34) emphasizes the multiple time scales of the model. The fastest variables are in order y1 , then y2 and y5 . The variables y3 and y4 are slow. Assume that (35) ȳ2 > δ > 0. This important assumption ensures the existence of an invariant manifold and will be justified, a posteriori. Then, from the last equation (34) we obtain the relation ȳ2 ȳ5 = k1 /k3 + O(ε 2 ), which represents the equation of an invariant manifold. In turn, this relation leads to the following equations for ȳ3 , ȳ4 ȳ′3 = k4′ ȳ4 + k4 ȳ4 ȳ23 − k6 ȳ3 , ȳ′4 = ε 2 (−k4′ ȳ4 − k4 ȳ4 ȳ23 + k1 ). (36) A second invariant manifold equation is defined by the equation ȳ1 = k8−1 (k9 ȳ2 + k6 ε 5 ȳ3 ). (37) Remind that ȳ3 is a slow variable. Then, for ȳ2 we have ȳ′2 = ε 2 (k6 ȳ3 − k1 ). (38) System (36) represents a two-dimensional reduced model of the initial five-dimensional system. This result shows that tropical equilibrations can be used for model reduction. The solutions of (36) either tend to the stable equilibrium ȳ4 = k1 , ′ k4 + k4 (k1 /k6 )2 ȳ3 = k1 /k6 , (39) or, if this equilibrium is unstable, to a limit cycle. Based on the general Theorem 2.7 we can assert the following: Theorem 3.1 Assume (35) holds with δ > 0. If the shorted system (36) has a stable hyperbolic limit cycle, then, under above conditions, for sufficiently small ε the five component system (26) also has a stable limit cycle. If the shorted system (36) has a stable hyperbolic equilibrium, then, under above conditions, for sufficiently small ε the five component system (26) also has a stable hyperbolic equilibrium. We have studied the system (36) analytically and numerically. The numerical simulations confirm the criteria of cycle existence both for small ε and for ε = O(1). For small epsilon the cycle has a singular structure. The amplitude of ȳ3 and the cycle period increase in ε , approximatively, as ε −2 (the assertion about period is natural since the rate of ȳ4 is O(ε 2 )). V. Noel, D. Grigoriev, S. Vakulenko, & O. Radulescu 97 Hyperbolicity can be straightforwardly checked for the rest point (39), by computing the eigenvalues of the linearized system. Denote by Y = (ȳ3 , ȳ4 ). When the rest point Y 0 = (ȳ03 , ȳ04 ) is hyperbolic and stable we have the following estimate |Y (t) −Y 0 | < C1 exp(−c1 ε 2t) (40) with some C1 , c1 > 0 holds. Integrating (38) for ȳ2 over interval [0, τ ] gives |ȳ2 (τ ) − ȳ2 (0)| < ε 2C1 c−1 1 = o(1) (41) uniformly in τ > 0 as ε → 0. This yields that ȳ2 (t) > δ if ȳ2 (0) > 2δ and therefore, ȳ2 (t) does not go to zero for large t, justifying the estimate (35) needed for the existence of an invariant manifold. The case of a limit cycle is discussed in the next subsection. 3.3 Singular limit cycle and hybrid dynamics Up to this point, the tropical ideas were used for reducing the dynamics of the model. In this section we show that the tropicalization heuristic is well adapted for decomposing the limit cycle into slow and fast modes, providing a hybrid description of the dynamics. Let us note that in a hybrid, excitable system, it is possible that not all variables are equilibrated. Also, the system can have more than two different equilibrations and associated invariant manifolds, and jump from one invariant manifold to another during the dynamics. Let us consider that the variables y1 , y2 , y5 are equilibrated as above, but now, only one among the variables y3 or y4 are equilibrated. We have four situations: 1) In eq. for ȳ3 : T 1 = T 3, T 2 <= T 1, 2) In eq. for ȳ3 : T 2 = T 3, T 1 <= T 2, 3) In eq. for ȳ4 : T 1 = T 3, T 2 <= T 1, 4) In eq. for ȳ4 : T 2 = T 3, T 1 <= T 2. Combined with the conservation law condition (30), variants 1 and 2 lead to the same triple tropical equilibration as before (Case I) and exponents (33). We denote the corresponding invariant manifold M1 . The renormalized equations are the same as (36). Variant 3 can be rejected by the general permanency criterion given by Lemma 2.6. Variant 4 (Case III) satisfies the permanency criterion and leads to a1 = 3, a2 = 0, a3 = 0, a4 = 4, a5 = 4. (42) This corresponds to a double equilibration (two equal terms) of the variable y4 , the variable y3 being not equilibrated. We denote the corresponding invariant manifold M2 . The renormalized equations read ỹ′3 = ε 6 k4′ ỹ4 + ε 2 k4 ỹ4 ỹ23 − k6 ỹ3 , ỹ′4 = −ε 2 k4′ ỹ4 + ε −2 (−k4 ỹ4 ỹ23 + k1 ). (43) We can provide a hybrid description of the cell cycle, by decomposing the periodic orbit into three modes (Fig.1). The first mode is the slowest and has the longest duration. It consists in the dynamics on the slow invariant manifold M1 at low values of y3 , and can be described by the algebraic-differential system ỹ′3 = ε 2 k1 − k6 ỹ3 , k4′ ỹ4 + k4 ỹ′4 ỹ23 − k6 ỹ3 = 0 (part between O1 and O of the orbit in Fig1c)). The exit from the invariant manifold M1 occurs at a critical point (point O). The next slowest mode corresponds to the decrease of y3 (part between O2 and O1 of the orbit in Fig1c)) and can be described by two terms truncated system ỹ′4 = ε −2 (k1 − k4 ỹ4 ỹ23 ), ỹ′3 = −k6 ỹ3 . Finally, there is a fast mode, corresponding to the fast increase of ỹ3 (part between O and O2 of the orbit in Fig1c)) and described by the truncated system ỹ′4 = −ε 2 k4 ỹ4 ỹ23 , ỹ′3 = ε −2 k4 ỹ4 ỹ23 . One can notice (Fig1c)) that this hybrid approximation is very accurate Hybrid models of the cell cycle 98 for small ε . At a distance from the tropical manifold, the hybrid orbit coincides with the one generated by the two terms or by the complete tropicalization. However, close to the tropical manifold, the two term and the complete tropicalization are less accurate than the hybrid approximation described above. Below we state rigorous estimates describing the slow movement on M1 and the fast jump towards M2 . The two terms description of the dynamics on M2 is a direct consequence of the Proposition 2.4 and Lemmas 2.6,2.8. To simplify notation, we rewrite the system (36) for y3 , y4 as x′ ′ = y = y + yx2 − k0 x, (44) ε 2 (−y − yx2 + k1 ). (45) We can obtain such a presentation by a linear variable change. Let us define the functions q k02 − k02 − 4y2 X (y) = , 2y q k02 + k02 − 4y2 , X+ (y) = 2y and the points y0 = k0 /2, x0 = X (y0 ) = 1. Lemma 3.2 Solutions of (44), (45) with initial data x(0), y(0) such that 0 < δ0 < x(0) < X+ (y0 ) − δ0 , 0 < y(0) < y0 − δ0 , (46) where δ0 is a small positive number independent on ε , satisfy |x(t) − X (y(t))| < C1 (ε + exp(−c1 δ0 t)), t < T0 (x(0), y(0), ε ) (47) this estimate holds while y(t) < y0 − δ2 , δ2 > 0. (48) Let us find some estimates of solutions at y = y0 , x = x0 . Our goal is to prove that at this point x(t) starts to increase sharply. After this, the terms ±yx2 play the main role in the equations (44),(45), and the other terms can be removed while the x-component is big. Let us introduce new variables u, v by x − x0 = v, u = y − y0 . For u, v one obtains v′ = k0 2 v + u(1 + (1 + v)2 ), 2 k0 + u)(1 + (1 + v)2 ) + k1 )) = ε 2 g(u, v). 2 Let us consider for this system the Cauchy problem with initial data u′ = ε 2 (−( v(0) = v0 , u(0) = u0 . (49) (50) (51) V. Noel, D. Grigoriev, S. Vakulenko, & O. Radulescu 99 Lemma 3.3 Consider the Cauchy problem (49), (50) and (51) under assumptions that u0 = κ > 0, |v0 | < δ5 , and k1 > k0 , (52) where δ0 , δ5 are small enough (but independent on ε ). Let A be a large positive number independent of ε . Then within some interval t ∈ (τ0 (δ5 , k0 , k1 , A), τ1 (δ5 , k0 , k1 , A)) one has u(t) > 0, v(t) ≥ A, v(τ0 ) = A. (53) This result can be reinforced. Actually, x(t) attains values of the order O(ε −2 ). Lemma 3.4 Assume x(t1 ) = A >> 1, Then for σ , A such that and t such that 4 y(t1 ) ≥ k0 /2. (54) 1 x(t) ≥ (A−1 − σ t)−1 , 2 (55) y(t) ≥ σ , (56) σ A > 2k0 , σ > k0 exp(−2σ −1 ), 1 (A−1 − σ t) ≥ ε 2 . 2 (57) (58) Conclusion We showed that tropical ideas can be usefully employed to reduce and hybridize polynomial or rational dynamical systems occurring in modelling the molecular machinery of the cell cycle. The main idea consists in keeping only the dominant monomial terms in the right hand side of the ordinary differential equations. Depending on the position in phase space, one should keep one, two, or more such terms. The places where two or more monomial terms are equal form the so-called tropical manifolds. The one term approximation is valid far from the tropical manifolds, whereas close to tropical manifolds several dominating terms of opposite signs can equilibrate each other. These “tropical equilibrations” of the dominating terms slow down the dynamics and produce attractive invariant manifolds. The possible applications of this method are multiple. Generally, the method can be used to obtain simplified models. In the example studied here, we have started with a five variables model, that has been reduced to two variables and hybridized. The modes of the hybrid model have the simple structure of monomial differential or differential-algebraic equations. Two general methods that we called complete and two terms tropicalizations provide description of the modes and of the mode changes. However, these general procedures may lead to inaccurate approximations when the full model does not satisfy permanency globally. In such cases, more thorough analysis is needed. We have shown that the model of embryonic cell cycle has essentially three modes with different timescales, namely slow accumulation of cyclin, rapid activation of MPF and intermediately rapid degradation of cyclin and inactivation of MPF. 100 Hybrid models of the cell cycle Figure 1: Limit cycle behavior of the paradigmatic cell cycle model from [27]. (a) The three main processes during the embryonic cell cycle are, in order of the timescales, fast increase of y3 (active MPF, triggering mitosis), slower decrease of y3 and very slow increase of y4 (inactive MPF). (b) Two invariant manifolds corresponding to the two slow processes are close to the tropical manifolds (blue lines) and result from equilibration of the variables (equilibration of y4 corresponds to M2 and equilibration of y3 corresponds to M1 ). (c) A three modes hybrid approximation of the cell cycle (in red) compared to the original limit cycle (black crosses). V. Noel, D. Grigoriev, S. Vakulenko, & O. Radulescu 101 The fastest mode is described by monomial ODEs, whereas the less fast modes correspond to tropical equilibrations and are described by differential-algebraic equations. Several improvements and developments are needed in order to apply these methods at a larger scale. The computation of tropical equilibrations suffers from combinatorial explosion. However, for the biochemical network used as working example, the number of solutions seems to be very small compared to the large combinatorics of monomial terms. There is hope, that once formulated in constraint logic programming, the problem of equilibrations could be efficiently computed in practice as a constraint satisfaction problem. Also, effective methods are needed to compute the transitions between modes. The main difficulty here is related to walls (segments of the tropical manifolds) crossing. Near walls, two or more terms are dominant. When these terms are equilibrated, orbits remain close to the walls and are contained in invariant manifolds. The complete or two terms tropicalizations provide general heuristic for mode transitions. These approximations mail fail close to walls. For instance, as we showed in a previous paper [15], the complete tropicalization predicts sliding modes that evolve on the wall and stay thus close to orbits of the full system. However, these sliding modes can be too long, leaving the wall when the orbits of the full system are already far away. In order to get an accurate description of the behavior near such walls we had to compute invariant manifolds. Although this is generally much simpler than integrating the full set of equations, it could become difficult for tropical equilibrations involving more than two terms. Future work will be dedicated to developing general methods for this problem. References [1] R. Alfieri, E. Bartocci, E. Merelli & L. Milanesi (2011): Modeling the cell cycle: From deterministic models to hybrid systems. Biosystems 105, pp. 34–40, doi:10.1016/j.biosystems.2011.03.002. [2] V. Baldazzi, P.T. Monteiro, M. Page, D. Ropers, J. Geiselmann & H. Jong (2011): Qualitative Analysis of Genetic Regulatory Networks in Bacteria. Understanding the Dynamics of Biological Systems, pp. 111–130, doi:10.1007/978-1-4419-7964-3_6. [3] G. Batt, H. De Jong, M. Page & J. Geiselmann (2008): Symbolic reachability analysis of genetic regulatory networks using discrete abstractions. Automatica 44(4), pp. 982–989, doi:10.1016/j.automatica.2007. 08.004. [4] L. Bortolussi & A. Policriti (2011): (Hybrid) automata and (stochastic) programs. The hybrid automata lattice of a stochastic program. Journal of Logic and Computation, doi:10.1093/logcom/exr045. [5] A. Crudu, A. Debussche, A. Muller & O. Radulescu (to appear, Arxiv preprint arXiv:1101.1431): Convergence of stochastic gene networks to hybrid piecewise deterministic processes. Annals of Applied Probability. [6] A. Crudu, A. Debussche & O. Radulescu (2009): Hybrid stochastic simplifications for multiscale gene networks. BMC Systems Biology 3(1), p. 89, doi:10.1186/1752-0509-3-89. [7] R. David & H. Alla (2008): Discrete, continuous, and hybrid Petri nets. IEEE Control Systems 28, pp. 81–84, doi:10.1109/MCS.2008.920445. [8] S. Drulhe, G. Ferrari-Trecate & H. de Jong (2008): The switching threshold reconstruction problem for piecewise-affine models of genetic regulatory networks. IEEE Transactions on Automatic Control 53, pp. 153–165, doi:10.1109/TAC.2007.911326. [9] A.F. Filippov & FM Arscott (1988): Differential equations with discontinuous righthand sides. 18, Springer. [10] AN Gorban & O. Radulescu (2008): Dynamic and static limitation in reaction networks, revisited . In David West Guy B. Marin & Gregory S. Yablonsky, editors: Advances in Chemical Engineering - Mathematics in Chemical Kinetics and Engineering, Advances in Chemical Engineering 34, Elsevier, pp. 103–173, doi:10.1016/S0065-2377(08)00002-1. Hybrid models of the cell cycle 102 [11] D. Henry (1981): Geometric theory of semilinear parabolic problems. Lecture Notes in Mathematics 840. [12] P. Lincoln & A. Tiwari (2004): Symbolic systems biology: Hybrid modeling and analysis of biological networks. Hybrid Systems: Computation and Control, pp. 147–165, doi:10.1007/978-3-540-24743-2_44. [13] A.S. Matveev & A.V. Savkin (2000): Qualitative theory of hybrid dynamical systems. Birkhauser. [14] B. Mishra (2009): Intelligently deciphering unintelligible designs: algorithmic algebraic model checking in systems biology. Journal of The Royal Society Interface 6(36), pp. 575–597, doi:10.1098/rsif.2008. 0546. [15] V. Noel, D. Grigoriev, S. Vakulenko & O. Radulescu (2012): Tropical geometries and dynamics of biochemical networks. Application to hybrid cell cycle models. Electronic Notes in Theoretical Computer Science 284, pp. 75–91, doi:10.1016/j.entcs.2012.05.016. [16] V. Noel, S. Vakulenko & O. Radulescu (2010): Piecewise smooth hybrid systems as models for networks in molecular biology. In: Proceedings of JOBIM 2010, Jobim, Montpellier. [17] V. Noel, S. Vakulenko & O. Radulescu (2011): Algorithm for identification of piecewise smooth hybrid systems: application to eukaryotic cell cycle regulation. Lecture Notes in Computer Science 6833, pp. 225– 236, doi:10.1007/978-3-642-23038-7_20. [18] R. Porreca, S. Drulhe, H. Jong & G. Ferrari-Trecate (2008): Structural identification of piecewise-linear models of genetic regulatory networks. Journal of Computational Biology 15(10), pp. 1365–1380, doi:10. 1089/cmb.2008.0109. [19] O. Radulescu, A. Muller & A. Crudu (2007): Théorèmes limites pour des processus de Markov à sauts. Synthèse des resultats et applications en biologie moleculaire. Technique et Science Informatique 26, pp. 443–469, doi:10.3166/tsi.26.443-469. [20] D. Ropers, V. Baldazzi & H. de Jong (2011): Model reduction using piecewise-linear approximations preserves dynamic properties of the carbon starvation response in Escherichia coli. IEEE/ACM Transactions on Computational Biology and Bioinformatics 8(1), pp. 166–181, doi:10.1109/TCBB.2009.49. [21] M.A. Savageau, P.M.B.M. Coelho, R.A. Fasani, D.A. Tolla & A. Salvador (2009): Phenotypes and tolerances in the design space of biochemical systems. Proceedings of the National Academy of Sciences 106(16), p. 6435, doi:10.1073/pnas.0809869106. [22] R. Shorten, F. Wirth, O. Mason, K. Wulff & C. King (2007): Stability Criteria for Switched and Hybrid Systems. SIAM Review 49(4), pp. 545–592, doi:10.1137/05063516X. [23] A. Singh & J.P. Hespanha (2010): Stochastic hybrid systems for studying biochemical processes. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences 368(1930), pp. 4995–5011, doi:10.1098/rsta.2010.0211. [24] R. Singhania, R.M. Sramkoski, J.W. Jacobberger & J.J. Tyson (2011): A hybrid model of mammalian cell cycle regulation. PLoS computational biology 7(2), p. e1001077, doi:10.1371/journal.pcbi.1001077. [25] Y. Takeuchi (1996): Global dynamical properties of Lotka-Volterra systems. World Scientific, Singapore. [26] L. Tavernini (1987): Differential automata and their discrete simulators. Nonlinear Anal. Theory Methods Applic. 11(6), pp. 665–683, doi:10.1016/0362-546X(87)90034-4. [27] J.J. Tyson (1991): Modeling the cell division cycle: cdc2 and cyclin interactions. Proceedings of the National Academy of Sciences of the United States of America 88(16), p. 7328, doi:10.1073/pnas.88.16.7328. [28] P. Ye, E. Entcheva, SA Smolka & R. Grosu (2008): Modelling excitable cells using cycle-linear hybrid automata. Systems Biology, IET 2(1), pp. 24–32, doi:10.1049/iet-syb:20070001. 5 Appendix: proofs Proof of Lemma 3.2 Let us consider the equation x′ = y(0) + y(0)x2 − k0 x = f0 (x). (59) V. Noel, D. Grigoriev, S. Vakulenko, & O. Radulescu 103 One observes that f (x) < −δ4 < 0, x ∈ (X (y(0)) + δ0 , X+ (y0 ) − δ0 ), f (x) > δ4 , x ∈ (δ0 , X (y0 ) − δ0 ), and f ′ (x)|x=X(y0 ) < −δ5 < 0. √ k2 + k02 −4y(0)2 Therefore, if 0 < x(0) < 0 2y(0) − δ0 , then the solution x(t) attains a small δ - neighborhood of X (y(0)) within a bounded time interval T1 (δ , δ0 , δ4 ): |x(t, x0 , y0 ) − X (y(0))| < δ , t = T1 (δ , δ0 , δ4 ). Within a small δ - neighborhood of X (y(0) we set u = x − X (y(0)) and then we can rewrite (59) as follows : u′ = −κ u + h(u), |h(u)| < C1 u2 , u(T1 ) = δ . (60) where κ > 0 is independent of δ . Then, if δ is small enough, we have that u(t) < δ for all t > T1 and |u(t)| < C(T0 ) exp(−κ t/2), t > T0 . Let us compare now the solution x(t) of (59) and the corresponding solution x̄(t, x0 , y0 ) of (44) with the same initial data. For x(t) − x(0) = w one has, since y(t) − y(0) < C(T0 )ε 2 on any bounded interval t ∈ [0, T0 ], wt = a(t)w + ε 2 g(x,t, w), 0 < t ≤ T0 . with a smooth function g, a(t) is bounded function. Now the Gronwall inequality implies |w(t)| < C2 (T0 )ε 2 , t ∈ [0, T0 ]. Therefore, one has |x̄(t, x0 , y0 ) − X (y(t))| < δ , t = T2 (δ , δ0 , δ4 ). Since the function X (y) defines a smooth, locally attracting (for y < y0 ) invariant manifold, this proves our assertion. One can prove this in another, elementary way. Let us define u = x̄ − X (y). Then u′ = −κ (t)u + h(u) + O(ε ), |h(u)| < C1 u2 , u(T1 ) = δ , (61) where κ (t) > κ0 while y < y0 − δ0 . Again one has u(t) < 2δ for t > T1 (while y < y0 − δ0 ). Thus, u′ ≤ − κ0 u + O(ε ) 2 that entails the need estimate 47. Proof of Lemma 3.3 Let us consider the Cauchy problem w′ = k0 2 w , 2 w(0) = v0 = v(0) It is clear, by the comparison principle, that v(t) ≥ w(t) (62) Hybrid models of the cell cycle 104 while u(t) > 0. Consequently, the assertion w(τ0 ) = A proves the lemma. Let us prove first that if δ6 is small enough (but independent on ε ), for some t = τ1 we have v(τ1 ) = δ6 . (63) Without loss of generality, we assume that v(t) < δ6 , 0 < t < τ1 , v(τ1 ) = δ6 . (64) Let δ6 > 0 be small enough such that − k0 (1 + (1 + δ6 )2 ) + k1 > 0. 2 (65) Such a choice of δ6 is possible due to (52). Assume u(t) ≥ 0 within some time interval [0, τ2 ], and u(τ2 ) = 0 for some τ2 < τ1 . Since u(0) = 0, we have τ2 > 0. We can suppose without loss of generality that τ2 is the first moment, where u(τ2 ) = 0. Then u′ (τ2 ) ≤ 0. But then we obtain a contradiction with (65) at t = τ2 , since the right hand side of this equation is positive at this time moment. Therefore, we have shown that u(t) > 0 for all t from [0, τ1 ] if v(t) < δ6 for such t. Then v(t) ≥ w(t) on this time interval. The function w can be found, and an easy computation gives w(t) = (w−1 (0) − 2k0−1t)−1 ) = (u0 − 2k0−1t). (66) Assume that v(t) < δ6 for all t. Then (66) holds for all t, but w(τ1 ) > δ6 for some τ1 . We have obtained a contradiction, thus (63) is proved. Let us prove that v(t) = A. Let us consider an interval [t1 , T ] such that |v(t)| < A, v(t1 ) = δ6 . t ∈ [t1 , T ], (67) Then (65) implies u′ ≤ ε 2 (( k0 + |u|)(1 + (1 + A)2 ) + k1 )), 2 (68) that gives, by the Gronwall lemma, u(t) < C2 u(0) exp(C1 (A)ε 2t). Within interval (t1 , T ] one has then u(t) < C4 κ , t ∈ [0, T ]. (69) (70) and, therefore, v′ ≥ k0 2 v −C4 κ , 2 t ∈ [0, T ]. (71) Suppose that C4 κ < k0 δ6 . 8 Then (71) entails v′ ≥ k0 v, 4 t ∈ [0, T ], (72) V. Noel, D. Grigoriev, S. Vakulenko, & O. Radulescu 105 This gives k0 (t − t1 )). 4 This leads to a contradiction for T large enough (however, let us remark that T is uniform in ε ). Proof of Lemma 3.4 Remark: to satisfy (57) it suffices to set σ = 2k0 A−1 with a large A. Suppose that either the estimate (55) (case A) or the second estimate (56) (case B) is violated at some T but the both inequalities hold for all t1 ≤ t < T . We can set t1 = 0. Let us consider the case A. Since (56) hold, one has x′ ≥ σ x2 , x(0) = A, t ∈ [0, T ]. v(t) ≥ δ6 exp( This implies x′ /x2 = (−1/x)′ ≥ σ , x(0) = A. Thus, x ≥ (A−1 − σ t)−1 . For t = T this result gives a contradiction with (55). Let us consider the case B. Since (55) hold, one has y′ (t) ≤ −ε 2 (y(A−1 − σ t)−1 − k1 ), y(0) = k0 /2, t ∈ [0, T ]. Z t (1 + (A−1 − σ s)−2 )ds)d τ . This implies y(t) ≤ k0 exp(−ε 2 2 Z t 0 (1 + (A−1 − σ s)−2 )ds) + ε 2 k1 0 exp(−ε 2 Z t τ for t ∈ [0, T ]. Notice Z 0 τ t (1 + (A−1 − σ s)−2 )ds = t − τ + σ −1((A−1 − σ t)−1 − (A−1 − σ τ )−1). k0 (−2/σ ). Under condition (57) Thus, for sufficiently small ε , taking into account (58) one has y(T ) ≤ exp this result gives a contradiction with (56). Proof of Theorem 2.7 i Suppose that the tropically truncated system (TTS) has a globally attracting compact invariant set A . Let be Π be an open neighborhood of this set. We can choose this neighborhood as a box that contains A . Then, for all initial data x(0), the corresponding trajectory x(t), x(0) lies in Π for all t > T0 (x0 , Π ). Therefore, our TTS is permanent. Here we do not use that the cycle (rest point) is hyperbolic. ii Permanency of the initial system follows from hyperbolicity of A . Hyperbolic sets are persistent (structurally stable) (Ruelle 1989). Since this set is globally attracting, all TTS is structurally stable (as a dynamical system). This implies that the initial system has a hyperbolic attractor close to A , since initial system is a small perturbation of TTC in Π . iii If the set A is only locally attracting, the last assertion of Theorem follows from persistency of hyperbolic sets.
5cs.CE
A Small-Order-Polynomial-Sized Linear Program for Solving the Traveling Salesman Problem Moustapha Diaby OPIM Department; University of Connecticut; Storrs, CT 06268 [email protected] arXiv:1610.00353v2 [cs.DS] 4 Oct 2016 Mark H. Karwan Department of Industrial and Systems Engineering; SUNY at Buffalo; Amherst, NY 14260 [email protected] Lei Sun Praxair, Inc.; Tonawanda, NY 14150 [email protected] Abstract. We present a polynomial-sized linear program (LP) for the n-city TSP drawing upon “complex 9 8 5 flow” modeling ideas by the first two authors who used an O(n ) × O(n ) model*. Here we have only O(n ) 4 variables and O(n ) constraints. We do not model explicit cycles of the cities, and our modeling does not involve the city-to-city variables-based, traditional TSP polytope referred to in the literature as “The TSP Polytope.” Optimal TSP objective value and tours are achieved by solving our proposed LP. In the case of a unique optimum, the integral solution representing the optimal tour is obtained using any LP solver (solution algorithm). In the case of alternate optima, an LP solver (e.g., an interior-point solver) may stop with a fractional (interior-point) solution, which (we prove) is a convex combination of alternate optimal TSP tours. In such cases, one of the optimal tours can be trivially retrieved from the solution using a simple iterative elimination procedure we propose. We have solved over a million problems with up to 27 cities using the barrier methods of CPLEX, consistently obtaining all integer solutions. Since LP is polynomially solvable and we have a model which is of polynomial size in n, the paper is thus offering (although, incidentally) a proof of the equality of the computational complexity classes “P ” and “N P ”. The non-applicability and numerical refutations of existing negative extended formulations results (such as Braun et. al. (2015)∗∗ and Fiorini et al. (2015)∗∗∗ in particular) are briefly discussed in an appendix. Keywords: Linear Programming; Combinatorial Optimization; Traveling Salesman Problem; TSP; Computational Complexity. *: Advances in Combinatorial Optimization: Linear Programming Formulation of the Traveling Salesman and Other Hard Combinatorial Optimization Problems (World Scientific, January 2016). **: Braun, G., S. Fiorini, S. Pokutta, and D. Steurer (2015). Approximation Limits of Linear Programs (Beyond Hierarchies). Mathematics of Operations Research 40:3, pp. 756-772.. ***: Fiorini, S., S. Massar, S. Pokutta, H.R. Tiwary, and R. de Wolf (2015). Exponential Lower Bounds for Polytopes in Combinatorial Optimization. Journal of the ACM 62:2, Article No. 17. 1 Introduction The traveling salesman problem (TSP), with its many variants, is one of the most widely studied problems in Operations Research. A good review of natural formulations which have been proposed for the more general (asymmetric) version is given in Öncan et al. (2009). These formulations all 1 use the ‘traditional’ city-to-city modeling variables plus additional variables which are introduced in order to enforce the subtour elimination constraints and to tighten the bounds of the corresponding linear programming (LP) relaxation. Here we will introduce a new modeling approach based on the work of Diaby (2007) and Diaby and Karwan (2016a). Those authors developed a generalized framework for formulating hard combinatorial optimization problems (COPs) as polynomial-sized linear programs. The perspective they adopted was that many of the well-known hard combinatorial optimization problems (starting with the TSP) could be modeled as Assignment Problems (APs) with side/complicating-constraints (see Diaby (2010a, b, c)). Their overall idea for the TSP was to use a flow perspective over an AP-based graph with nodes representing (city, time-of-travel) pairs, and arcs representing travel legs. The novel feature in their modeling was to have variables representing doublets and triplets of travel legs, thereby “building” enough information into the modeling variables themselves that the enforcement of the “complicating constraints” could be done implicitly by simply setting appropriate variables to zero in the model. This led to an O(n9 )-variables, O(n8 )-constraints model. Using a similar modeling philosophy in this self-contained paper, we propose a simplified model with O(n5 ) variables and O(n4 ) constraints. Where Diaby (2007) and Diaby and Karwan (2016a) used doublets and triplets of arcs, we employ only combinations comprising a node (a “dot”) and an arc (a “dash”) of the AP-based graph in defining our flow variables. That is, whereas Diaby (2007) and Diaby and Karwan (2016a) respectively used three “dashes,” we accomplish the optimization task with a “dot” and a “dash” only. Unlike in Diaby (2007) and Diaby and Karwan (2016a), in the overall polytope of this new model, not all extreme points are integer. We show that a projection of the model onto the space of one of its classes of variables is integral. Hence, by expressing TSP tour costs in terms of the variables being projected to only, the proposed LP always finds an optimal TSP tour. Since LP is polynomially solvable (e.g. Bazaraa, M.S., J.J. Jarvis and H.D. Sherali (2010)), and we have a model which has polynomial size in n, we are thus offering a new proof of the equality of the computational complexity classes “P ” and “N P ” (see Garey, M.R. and D.S. Johnson (1979) for a treatise on computational complexity). The plan of this paper is as follows. We discuss a path representation of TSP tours in section 2. Section 3 develops our LP model of the path representations using the “complex flow” variables which we introduce. Special characteristics of the model structure are discussed in section 4. Problem size and our computational experiments are discussed in section 5. Finally, conclusions are discussed in section 6, and the non-applicability of existing negative extended formulations (EF) results for the “standard” TSP polytope and our software implementation of the model in this paper (which will be made publicly available) are described in Appendices A and B, respectively. 2 Path representation of TSP tours We will begin this section with statements of our basic assumptions and notation for the TSP parameters. Then, we will develop the path representation which serves as the foundation for the developments in the remainder of the paper. Assumption 1 We assume without loss of generality (w.l.o.g.) that: 1. The number of cities is greater than 5; 2. The TSP graph is complete; (Arcs on which travel is not permitted can be handled in the optimization model by associating large (“Big-M ”) costs to them.) 2 3. City “0” has been designated as the starting and ending point of the travels. Definition 2 We refer to the order in which a given city is visited after city 0 in a given TSP tour as the “time-of-travel” of that city in that TSP tour. In other words, if city i is the rth city to be visited after city 0 in a given TSP tour, then we will say that the time-of-travel of city i in the given tour is r. Notation 3 (General Notation) 1. n : Total number of cities (including the starting city); 2. ∀i, j ∈ Ω := {0, . . . , n − 1}, ci,j : Cost of travel from city i to j; 3. m := n − 1 (Number of cities to sequence); 4. M := {1, . . . , m} = {1, . . . , n − 1} (Set of cities to sequence); 5. S := {1, . . . , m} = {1, . . . , n − 1} (Index set for the times-of-travel ); 6. Ext(·): Set of extreme points of (·). Our overall approach consists of formulating the TSP as a generalized flow problem over the APbased digraph illustrated in Figure 1. We refer to this graph as the “TSP Flow Graph (TSPFG).” The nodes of the TSPFG correspond to (city, time-of-travel ) pairs. The arcs of the graph link nodes at consecutive times-of-travel and represent travel legs. The graph notation is formalized below. Notation 4 (TSPFG formalisms) 1. ∀r ∈ S, Nr := M = {1, . . . , m} (Set of cities that can be visited at time r); 2. N := {(i, r) : r ∈ S, i ∈ Nr } (Set of nodes of the TSPFG. We will, henceforth, write (i, r) ((i, r) ∈ N ) as [i, r] in order to distinguish it from other doublets); 3. ∀r ∈ S, ∀i ∈ Nr ,   ∅ if r = m Fr (i) :=  Nr+1 \{i} otherwise (Forward star of node [i, r] of the TSPFG); 4. ∀r ∈ S, ∀i ∈ Nr ,   ∅ if r = 1 Br (i) :=  Nr−1 \{i} otherwise (Backward star of node [i, r] of the TSPFG); 5. A : {(i, r, j), r ∈ R, i ∈ Nr , j ∈ Fr (i)} (Set of arcs of the TSPFG. We will, henceforth, write (i, r, j) ((i, r, j) ∈ A) as [i, r, j] in order to distinguish it from other triplets); 3 Figure 1: Illustration of the TSP Flow Graph 6. R := {1, . . . , m − 1} = {1, . . . , n − 2} (Set of stages of the TSPFG which have nodes with positive out-degrees). Remark 5 It follows directly from the definitions in Notations 4.3-4.5 that: ∀(i, j) ∈ M 2 , ∀r ∈ R, [i, r, j] ∈ A =⇒ i 6= j. That is, no arc in the TSPFG connects a city with itself between consecutive times-of-travel, as illustrated in Figure 2 below. Definition 6 (“stages,” “levels,” “TSP paths”) 1. We refer to the set of nodes of the TSPFG corresponding to a city as a “level” of the graph. (The set of levels is denoted M in Notation 3.4); 2. We refer to the nodes of the TSPFG corresponding to a given time-of-travel of the TSP as a “stage” of the graph. (The set of stages is denoted S in Notation 3.5); 3. We refer to a path of the TSPFG that includes exactly one node of each level of the graph as a “TSP path.” A TSP path is illustrated in Figure 2. Theorem 7 There exists a one-to-one correspondence between TSP paths (of the TSPFG) and TSP tours. Proof. Trivial, since each level, i, of the TSPFG can be visited only once, and the number of levels is equal to the number of stages. 4 Figure 2: Illustration of “TSP paths” of the TSPFG 3 Mathematical programming model of TSP paths We will first state the LP model. Then, we will illustrate each its classes of variables and constraints and provide further intuition on the modeling approach. Finally, we will discuss the objective function. 3.1 Modeling Variables Notation 8 (Modeling variables) 1. ∀[i, r, j] ∈ A, we define a variable y[i,r,j] . Variable y[i,r,j] may be interpreted as the amount of flow in the TSP FG that traverses arc [i, r, j] in a given (feasible) solution to our LP model (to be stated); 2. ∀[i, r] ∈ N , ∀[j, s, k] ∈ A, we define a variable x[i,r][j,s,k] . Variable x[i,r][j,s,k] may be interpreted as the amount of flow in the TSP FG that traverses both, node [i, r] and arc [j, s, k] in a given (feasible) solution to our LP model (to be stated). Hence, our modeling variables respectively are comprised of an arc (for the y-variables), and a combination of an arc and a node (a “dash” and a “dot”) of the TSPFG, as illustrated in Figure 3 below. 3.2 Model Constraints Our Linear Programming (LP) model of TSP paths consists of the five main classes of constraints described below. We will subsequently fully explain each of these and the objective function. 5 Figure 3: Illustration of the modeling variables 3.2.1 Statement of the constraints • Initial Flow m X m m X X x[i,1][j,2,k] = 1. (1) i=1 j=1 k=1 / j6=i k∈{i,j} • Mass Balance/Generalized Kirchhoff Equations (GKEs) x[i,r][i,r,j] − m X x[i,r][j,r+1,k] = 0; k=1 k∈{i,j} / ∀i, j = 1, . . . , m : i 6= j; m X x[i,r][j,r−1,i] − ∀r = 1, . . . , m − 2. (2) x[i,r][k,r−2,j] = 0; k=1 k∈{i,j} / ∀i, j = 1, . . . , m : i 6= j; m X x[i,r][k,r−1,i] − k=1 k6=i ∀i = 1, . . . , m; m X ∀r = 3, . . . , m. (3) x[i,r][i,r,k] = 0; k=1 k6=i ∀r = 2, . . . , m − 1. (4) 6 m X m X x[i,r][k,p−1,u] − k=1 k∈{i,u} / x[i,r][u,p,k] = 0; ∀i, u = 1, . . . , m : i 6= u; k=1 k∈{i,u} / ∀r = 1, . . . , m; ∀p = 2, . . . , m − 1 : p ∈ / {r − 1, r, r + 1}. (5) • Node-pair Reciprocities x[i,r][k,r+1,j] − x[j,r+2][i,r,k] = 0 ∀i, j, k = 1, . . . , m : i 6= j 6= k; m X m X x[i,r][k,s−1,j] − k=1:k∈{i,j} / ∀r = 1, . . . , m − 2. (6) x[j,s][i,r,k] = 0 k=1:k∈{i,j} / ∀i, j = 1, . . . , m : i 6= j; ∀r = 1, . . . , m − 2; ∀s = r + 3, . . . , m. (7) • Flow Consistencies y[i,r,j] − x[i,r][i,r,j] = 0; ∀i, j = 1, . . . , m : i 6= j; ∀r = 1, . . . , m − 1. (8) ∀r = 1, . . . , m − 1. (9) y[i,r,j] − x[j,r+1][i,r,j] = 0; ∀i, j = 1, . . . , m : i 6= j; • Visit Requirements for Arcs m X y[i,r,j] − x[k,s][i,r,j] = 0; k=1 k∈{i,j} / ∀i, j = 1, . . . , m : i 6= j; y[i,r,j] − m X ∀r = 1, . . . , m − 1; ∀s = 1, . . . , m : s ∈ / {r, r + 1}. (10) x[u,s][i,r,j] = 0; s=1 s∈{r,r+1} / ∀i, j, u = 1, . . . , m : i 6= j 6= u; ∀r = 1, . . . , m − 1. (11) • Visit Requirements for Nodes   m−1 r−2 X m X X    x + x + x + [i,r][u,r−1,i] [i,r][i,r,u] [i,r][u,p,k]   p=1 p=r+1 k=1 k∈{i,u} / m X x[i,r][k,p,u] k=1 k∈{i,u} /   r−2 X m m−1 m X X X   − x − x − x − x[i,r][l,q,1] = 0; [i,r][1,r−1,i] [i,r][i,r,1] [i,r][1,q,l]   q=1 q=r+1 l=1 l∈{1,i} / ∀i, u = 2, . . . , m : i 6= u; ∀r = 1, . . . , m. 7 l=1 l∈{1,i} / (12)   r−2 X m X p=1 k=2 k∈{1,u} /    m−1 X   x[1,r][u,p,k]  + x[1,r][u,r−1,1] + x[1,r][1,r,u] + p=r+1 m X x[1,r][k,p,u] k=2 k∈{1,u} /   r−2 X m m−1 m X X X − x[1,r][2,q,l]  − x[1,r][2,r−1,1] − x[1,r][1,r,2] − x[1,r][l,q,2] = 0; q=1 l=3 q=r+1 l=3 ∀u = 3, . . . , m; ∀r = 1, . . . , m. (13) • Implicit Zeros y[i,r,j] = 0 if (i = j); ∀i, j = 1, . . . , m; x[i,r][j,s,k] = 0 ∀r = 1, . . . , m − 1. (14) if ((i = k) or (j = k) or (s = r and i 6= j) or (s = r − 1 and i 6= k)) ; ∀i, j, k = 1, . . . , m; ∀r, s = 1, . . . , m. (15) • Nonnegativities y[i,r,j] ≥ 0; ∀i, j = 1, ..., m; ∀r = 1, ..., m − 1. x[i,r][j,s,k] ≥ 0; ∀i, j, k = 1, ..., m; ∀r = 1, ..., m; ∀s = 1, ..., m − 1. (16) (17) The overall logic of the model is described in eight steps below. 1. We have a model built over an Assignment flow graph. Constraint (1) creates an initial flow in x. 2. Constraints (2)-(5) are a set of Kirchoff Equations in x, keeping flow moving across stages in a connected and “balanced” mannner, as illustrated in Figure 4 below. For p > r, flow that nodes [i, r] and [u, p] have in common must leave [i, r],and enter and leave [u, p], and then leave [u, p] (via some arcs at stages (p − 1) and (p + 1),respectively) in a balanced manner. In a similar way, for p > r, flow which is “destined” to [u, p] must enter and leave [i, r] (via some arcs at stages (r − 1) and (r + 1),respectively) in a blance manner, before entering [u, p]. 3. Constraints (6)-(7) stipulate that each pair of assignments induced by a y-variable elicits a “flow in = flow out” condition over a set of arcs (via the x-variables) for the nodes comprising the variable. These are illustrated in Figures 5 and 6. 8 Figure 4: Illustrations of the GKE/Mass Balance Constraints 4. Constraints (8) and (9), together, ensure that flow on a given arc, [i, r, j], is “recognized” in the same way when viewed from either the tail node, [i, r], or the head node, [j, r + 1], of the arc. This is illustrated in Figure 7 below. 5. Each set of consecutive assignment pairs must “visit” each stage via constraints (10). Figure 8 below illustrates this condition, which is that the total amount of flow a given arc, [i, r, j], has in common with nodes at a given stage, s, is the the same for all s. (In other words, the sum of x-variables involving nodes at stage s and arc [i, r, j] equals the flow from node [i, r] to node [j, r + 1] (y-variable ) for all s). Also, flow traversing a given arc, [i, r, j], must “visit” (i.e., share value with) each level of the flow graph (city of the TSP) via constraints (11). This is illustrated in Figure 9. 6. A given node, [i, r], must share the same amount of flow with every level, u, of the TSPFG. 9 Figure 5: Illustration of Reciprocity Constraints (6) Constraints (12) enforce this condition for the case in which i 6= 1 by setting the total amount of flow shared with each u ∈ / {1, i} equal to the total amount of flow shared with level “1”. When i = 1, the condition is enforced by constraints (13) in which the total amount of flow shared with each u ∈ / {1, 2} is set equal to the total amount of flow shared with level “2”. These are illustrated in Figures 10 and 11. 7. Constraints (14)-(15) are used to restrict variables which by themselves would imply repeating a level (city) or a stage, or flow which is “broken” (by not “going out” of the node into which it “came in” at a stage involved in the variable). We call these “implicit zeros.” In implementation, these variables are just not created. 8. Finally, constraints (16) and (17) are the usual nonnegativity constraints. 3.2.2 Intuition behind the model A non-mathematical intuition for the idea of our modeling approach (to be formally described in the remainder of this paper) may be gained through the following observation. Assume that no part of a flow which traverses a given arc of the TSPFG can subsequently “re-visit” either of the levels (cities) of the graph that are involved in that arc. Then, flow can propagate forward through the graph, only in a “pattern” that consists of a convex combination of TSP paths. That is, we will show that each positive variable must lie on a TSP path. An x-variable being positive may be interpreted as a “communication” between the corresponding node and arc. The constraints are such that a node which has positive flow through it is made to “communicate” with pairs of nodes which are at consecutive stages at different parts of the network (through the x-variables) only along paths which are TSP paths of the network. As a result, feasible solutions of our proposed model are such that the total flow on each arc is a convex combination of flows along TSP paths 10 Figure 6: Illustrations of Reciprocity Constraints (7) only. This does not mean that every feasible solution to the model is a convex combination of flows along TSP paths only. The reason for this is that the model (the Overall Polytope) also has non-integral extreme points, due to the fact that the “pattern” of the y-variables may not be consistent with a single realization of the x-variables. We will show that the “flow pattern” of the y-variables is sufficient for being able to solve any TSP as a LP by utilizing an appropriate cost function which involves the y-variables only. 3.3 Objective Function Our developments to follow in the remainder of this paper show that our approach is guaranted to correctly solve the TSP only when the travel costs are captured through the y-variables only. The reason for this is that our developments only show the integrality of the y-Polytope. Many variations of the objective function may be possible. The simplest one is used here placing the cost of traveling from city i to city j on the y-variables representing the total flows on the arcs. Two additional terms are added to account for node 0 being the starting and ending point of the travels. This is accomplished by adding the cost from node 0 to the first and last arcs chosen, connecting cities at stages 1 and 2 and cities at stages m − 1 and m, respectively. The objective function we use is as follows: min m X m m−2 m X m m X m X XX X (c0i + cij )y[i,1,j] + cij y[i,r,j] + (cij + cj0 )y[i,m−1,j] i=1 j=1 4 r=2 i=1 j=1 i=1 j=1 Model structure Notation 9 11 (18) Figure 7: Illustration of the Flow Consistency Constraints 1. We denote the number of x-variables and y-variables in (1)-(17) by ξx and ξy , respectively. 2. We denote the polytope by (1)-(17)  induced    by Q, and refer  to it as the “Overall (LP) x x Polytope.” (i.e., Q := ∈ Rξx +ξy : satisfies (1)-(17) is referred to as the “Overall y y (LP) Polytope.”) 3. We denote the projection of Q onto the space of the  y-variables  by “Y ,” and  refer to it as the x “y-Polytope.” In other words, we refer to Y := y ∈ Rξy : ∃x ∈ Rξx : ∈Q as the y “y-Polytope.” Our objective in this section is to show that the y-Polytope is integral with extreme points corresponding to TSP tours. Note that this integrality does not imply the integrality of the Overall Polytope, even though the converse is true. The proof requires some preliminary notions and results which we will first discuss.   x Definition 10 Let ∈ Q. y 1. For (r, s) ∈ R2 : s > r, we say that a set of nodes at consecutive stages of the TSPFG, {ir , . . . , is }, (or the set of arcs, {[ir , r, ir+1 ], . . . , [is−1 , s − 1, is }, linking them in the TSPFG) is a “(r-to-s) TSP walk (of y)” if ir , . . . , is are pairwise distinct and yip ,p,ip+1 > 0 for p = 1, . . . , m − 1. 2. We refer to a TSP walk of y which spans the set of stages of the TSPFG as a “TSP path (of y)” (since, clearly, every such spanning TSP walk of y corresponds to exactly one TSP path of the TSPFG). 12 Figure 8: Illustrations of the “Column Visit” Requirements for the Arcs   x Remark 11 Let ∈ Q. By definitions 6 and 10.2 above, every TSP path of y corresponds to y exactly one TSP tour.   x Theorem 12 V := ∈ Q : ∃(i1 , . . . , im ) ∈ M m : ∀(p, q) ∈ R2 : p 6= q, (ip 6= iq ; and y  yip ,p,ip+1 = 1, p = 1, . . . , m − 1 ⊆ Ext(Q). Proof. We have: 1= = = m X m m X X i=1 j=1 k=1 / j6=i k∈{i,j} m m m XX X i=1 j=1 k=1 / j6=i k∈{i,j} m X m X x[i,1][j,2,k] (Using (1)) x[i,1][j,s,k] ∀s ∈ R y[j,s,k] ∀s ∈ R (Using (2)-(5) recursively) (Using (8)-(10)). j=1 k=1 k6=j 13 (19) Figure 9: Illustration of the “Row Visit” Requirements for the Arcs (19) and the nonnegativity constraints (16)-(17) imply:   x ∈ V =⇒ y  1 if j and k are the levels included in the existentially-quantified set    of the theorem for stages s and s + 1 (i.e., if j = is and k = is+1 ); ∀s ∈ R, y[j,s,k] =    0 otherwise. (20) (20) implies:   x ∈ V =⇒ y ∈ {0, 1}ξy . y (21) It is easy to verify that:    x ξy ∈ Q and y ∈ {0, 1} =⇒ x ∈ {0, 1}ξx . y (21) and (22) imply:     x x ∈ V =⇒ ∈ {0, 1}ξx +ξy . y y (22) (23) Clearly, 0 and 1 (respectively) cannot be represented as non-trivial convex combinations ofnumbers  x on the interval [0, 1]. It follows from the combination of this with (22) and (23) that a given ∈V y cannot be represented as a convex combination of other points of Q. 14 Figure 10: Illustration of the “Row Visit” Requirement for Node (i, r), When i > 1.   x Theorem 12 shows that even though a TSP path for ∈ Q is defined in terms of conditions y on y only, its (mathematical) correspondence to an extreme point of Q is preserved. Our  aim  in x this section is to show that each extreme point of the y-Polytope (Y ) is a TSP path for a ∈Q y (and therefore, corresponds to a TSP path of the TSPFG (and therefore, to a TSP tour)). Note that, as we have indicated above, the integrality of Y does not imply that of Q. The reason for this is that an extreme point of Q may project to a non-extreme point of Y . A conceptual illustration of this fact is given in Figure 12. Remark 13 An observation which can be made from Figure 12 is that the y-component of any point of “Q” (in the figure) must be a convex combination of the extreme points of “Y ” (in the figure). The y-components of “C”, “D”, “E”, and “F” in particular, must respectively be convex combinations of “A” and “H”. Hence, as can be also seen from the figure, provided the objective function in an LP over “Q” (in the figure) is expressed in terms of the y-variables only, at least one the extreme points of “Y ” (either “A” or “H” (or both) in the figure) must be optimal. In this way, for our purposes in this paper, the integrality of the projection of the Overall Polytope onto the space of a class of its variables happens to be sufficient for the optimizatin problem at hand. This sufficiency is the consequence of the combination of the fact that the projection is integral with the fact that it is possible to correctly capture TSP tour costs using the variables being “projected to” only in our modeling. The following notation is needed in order to develop the subsequent three lemmas (Lemmas 15-17) which establish some “connectivity” conditions between a node and an arc of the TSPFG which share flow in our model (through the x-variables). Lemma 17 is the “main” lemma (among the three) and shows the existence of a TSP walk for every (node, arc) pair of the TSPFG which share flow. Lemmas 15 and 16 provide the foundation for the induction used in proving Lemma 17.   x Notation 14 Let ∈ Q. We define the following: y 15 Figure 11: Illustration of the “Row Visit” Requirement for Node (i, r), When i = 1.   1. ∀ ((i, r), [j, s, k]) ∈ N , A : s > r; x[i,r][j,s,k] > 0 ,  (a) Uirjsk := u ∈ M : x[u,s−1][j,s,k] > 0 ;  (b) Virjsk := v ∈ M : x[i,r][v,s−1,j] > 0 .   2. ∀ ((i, r), [j, s, k]) ∈ N , A : s < r; x[i,r][j,s,k] > 0 ,  (a) U irjsk := u ∈ M : x[u,s+2][j,s,k] > 0 ;  (b) V irjsk := v ∈ M : x[i,r][k,s+1,v] > 0 .   x Lemma 15 Let ∈ Q. Then, the following is true: y ∀(i, r) ∈ N , ∀[j, s, k] ∈ A : s > r + 1,  x[i,r][j,s,k] > 0 ⇒ Uirjsk ∩ Virjsk 6= ∅ . Proof. Assume Uirjsk ∩ Virjsk = ∅. (24) 16 Figure 12: General Illustration of a Non-Integral Polytope Projecting to an Integral Polytope Then, we have: y[j,s,k] = m X x[u,s−1][j,s,k] (Using (10)) u=1 u∈{j,k} / = m X x[u,s−1][j,s,k] (Using (24) and the definitions in Notation 14) u∈Uirjsk X = x[u,r][j,s,k] (Using (24), (14)-(15), and applying (2)-(5) recursively) u∈(M \{i,j,k})  <  X x[u,r][j,s,k]  + x[i,r][j,s,k] = u∈(M \{i,j,k}) X x[u,r][j,s,k] u∈(M \{j,k}) (Since x[i,r][j,s,k] > 0 by assumption). (25) contradicts (8)-(10). Hence, we must have: Uirjsk ∩ Virjsk 6= ∅.   x Lemma 16 Let ∈ Q. Then, the following is true: y ∀(i, r) ∈ N , ∀[j, s, k] ∈ A : (s < r − 1) ,  x[i,r][j,s,k] > 0 ⇒ U irjsk ∩ V irjsk 6= ∅ . 17 (25) Proof. Assume U irjsk ∩ V irjsk = ∅. (26) Then, we have: m X y[j,s,k] = x[u,s+2][j,s,k] (Using (10)) u=1 u∈{j,k} / X = x[u,s+2][j,s,k] (Using (26) and the definitions in Notation 14) u∈U irjsk X = x[u,r][j,s,k] (Using (26) and (14)-(15), and applying (2)-(5) recursively) u∈(M \{i,j,k})   X < X x[u,r][j,s,k]  + x[i,r][j,s,k] = u∈(M \{i,j,k}) x[u,r][j,s,k] u∈(M \{j,k}) (Since x[i,r][j,s,k] > 0 by assumption). (27) (27) contradicts (8)-(10). Hence, we must have: U irjsk ∩ V irjsk 6= ∅.   x Lemma 17 Let ∈ Q. Then, the following are true: y 1. ∀(r, s) ∈ R2 : r ≤ s < m, x[ir ,r][is ,s,is+1 ] > 0 ⇒ ∃(it , t = r, . . . , s + 1) ∈ M s−r+2 :   ∀(p, q) ∈ {ir , . . . , is+1 }2 : p 6= q, ip 6= iq ; and yip ,p,ip+1 > 0, p = r, . . . , s . 2. ∀(r, s) ∈ R2 : s < r, x[ir ,r][is ,s,is+1 ] > 0 ⇒ ∃(it , t = s, . . . , r) ∈ M r−s+1 :   ∀(p, q) ∈ {is , . . . , ir }2 : p 6= q, ip 6= iq ; and yip ,p,ip+1 > 0, p = s, . . . , r − 1 . Proof. Observe that because of constraints (6)-(9), the two statements of the lemma are equivalent. Hence, it is suffucient to prove only one of the two statements. We will provide the proof for statement 1. Case 1: s = r. (8) and (14)-(15) =⇒:  x[ir ,r][ir ,r,ir+1 ] > 0 =⇒ ir 6= ir+1 , and y[ir ,r,ir+1 ] > 0 (28) Statement 1 of the lemma for this case follows from (28) directly. Case 2: s = r + 1. (10) and (14)-(15) =⇒:  x[ir ,r][ir+1 ,r+1,ir+2 ] > 0 =⇒ ir 6= ir+1 6= ir+2 and y[ir+1 ,r+1,ir+2 ] > 0 . 18 (29) Also, (6) implies: x[ir ,r][ir+1 ,r+1,ir+2 ] > 0 =⇒ x[ir+2 ,r+2][ir ,r,ir+1 ] > 0. (30) Using (10), we have: x[ir+2 ,r+2][ir ,r,ir+1 ] > 0 =⇒ y[ir ,r,ir+1 ] > 0. (31) Statement 1 of the lemma for this case follows from (29)-(31) directly. Case 3: s ≥ r + 2. The proof for this case is inductive. We will show that if the theorem holds for (r, r + δ) ∈ R2 , then the lemma must also hold for (r, r + δ + 1) ∈ R2 (where 1 ≤ δ ≤ m − r − 1). Assume that, for a given δ, we have the following: (a)∀(r, r + δ) ∈ R2 : r + δ < m, x[ir ,r][ir+δ ,r+δ,ir+δ+1 ] > 0 ⇒ ∃(ir+1 , . . . , ir+δ−1 ) ∈ M δ−1 : yip ,p,ip+1  ∀(p, q) ∈ {r, . . . , r + δ + 1}2 : p 6= q, ip 6= iq ;  > 0 for p = r, . . . , r + δ ; and (b)(ir , ir+δ+1 , ir+δ+2 ) ∈ M 3 are such that x[ir ,r][ir+δ+1 ,r+δ+1,ir+δ+2 ] > 0. (32) (33) We will show that there must exist a r-to-(r + δ + 2) TSP walk of y. Using (2), (5), (7), (10), and (14)-(15), (33) implies: (a) ir 6= ir+δ+1 6= ir+δ+2 ; (34) (b) y[ir+δ+1 ,r+δ+1,ir+δ+2 ] > 0; (35) (c) L := Uir ,r,ir+δ+1 ,r+δ+1,ir+δ+2 ∩ Vir ,r,ir+δ+1 ,r+δ+1,ir+δ+2 6= ∅; (36) (d) K := {k ∈ M : x[ir+δ,i (37) r+δ+2 ,r+δ+2][ir ,r,k] > 0} = 6 ∅; and (d) ∀k ∈ K, Lk := U ir+δ+2 ,r+δ+2,ir ,r,k ∩ V ir+δ+2 ,r+δ+2,ir ,r,k 6= ∅ (38) From “Case 2” and the definitions in Notation 14, we must have: ∀ur+δ ∈ L, ∃(ur+1 , . . . , ur+δ−1 ) ∈ M δ−1 : p 6= q, up 6= uq ) ; and yup ,p,up+1 ∀(p, q) ∈ {r, . . . , r + δ + 1}2 :  > 0 for p = r, . . . , r + δ . (Where ir and ir+δ+1 have been relabled as ur and ur+δ+1 , respectively). Similarly, we must also have: [ ∀vr+1 ∈ Lk , ∃(vr+2 , . . . , vr+δ ) ∈ M δ−1 : (39) ∀(p, q) ∈ {r + 1, . . . , r + δ + 2}2 : k∈K p 6= q, vp 6= vq ) ; and yvp ,p,vp+1 > 0 for p = r + 1, . . . , r + δ + 1  and ir+δ+2 have been relabled as vr+δ+1 and vr+δ+2 , respectively). 19 . (Where ir+δ+1 (40) Conclusion. By (11)-(12), at least one of the r − to − (r + δ + 1) TSP walks specified in (39) and one of the (r + 1) − to − (r + δ + 2) TSP walks specified in (40) must overlap in their components between stages (r + 1) and (r + δ + 1). (If not, constraints (11)-(13) would be violated for nodes ir and ir+δ+2 , and arc [ir+δ+1 , r + δ + 1, ir+δ+2 ] when r = 1 and δ = m − 3, in particular.) Statement 1 of the lemma follows from this directly. We will now discuss some topological properties of Q and Y which will be needed in order to prove our main result, which is the integrality of Y . Definition 18 Let λ be a scalar on the interval (0, 1].      x x ξ +ξ x y e λ := 1. We refer to Q ∈R : satisfies (41), and (2)-(17) as the “λ-scaled Overy y all Polytope,” where (41) is specified as: m X m m X X x(i,1)(j,2,k) = λ. (41) i=1 j=1 k=1 / j6=i k∈{i,j} e λ onto the space of the y-variables as the “λ-scaled y-Polytope,” 2. We refer to the projection of Q      x ξ ξ y x e e e and denote it by Yλ . In other words, we refer to Yλ := y ∈ R : ∃x ∈ R : ∈ Qλ y as the “λ-scaled y-Polytope.” Lemma 19 ∀(λ, µ) ∈ (0, 1]2 : λ 6= µ, e λ and Q e µ are homeomorphic; 1. Q 2. Yeλ and Yeµ are homeomorphic. Proof. e λ and Q e µ are homeomorphic. 1. Q It is easy to see (by observing the result of the multiplication of each of the constraints of Q by a (same) constant, for example) that:     x x eα . ∀α ∈ (0, 1], ∈ Q ⇐⇒ α · ∈Q (42) y y Hence, the point-to-point mapping     x x e α with hα hα : Q −→ Q =α· y y (43) is bijective. Also, clearly, ∀α ∈ (0, 1], hα is bicontinuous (see Gamelin and Greene (1999, pp. 26-27), or Panik (1993, pp. 267-268)). Hence, ∀α ∈ (0, 1], hα is a homeomorphism (see Gamelin and Greene (1999, pp. 27, 67), or Panik (1993, pp. 253-257)). The homeomorphism e λ and Q e µ follows directly from the equivalence property of homeomorphisms (since between Q e λ and Q e µ are respectively homeomorphic to Q). Q 20 2. Yeλ and Yeµ are homeomorphic. Clearly, the conditions expressed in (42)-(43) are applicable to the y-component vector of   x ∈ Q. The lemma follows from this, in a similar way as for Part (1) of the proof above. y A direct result of the combination of Lemmas 17 and 19 is the following. Corollary 20 ∀λ ∈ (0, 1], y ∈ Yeλ =⇒ There exists at least one TSP path of y. (44) Notation 21 ∀λ ∈ (0, 1], ∀y ∈ Yeλ , define: 1. Πλ (y) : Set of TSP paths of y; 2. Tλ (y) := {1, . . . , |Π(y)|} (Index set associated with Π(y));  3. ∀t ∈ Tλ (y), Pλ,t (y) := [iλ,t,k , k, iλ,t,k+1 ] ∈ A, k = 1, . . . , m − 1 (tth TSP path of y); 4. ∀t ∈ Tλ (y), w bλ,t (y) : Characteristic vector of Pλ,t (y) (w bλ,t (y) is obtained by setting the y-variable for each arc involved in Pλ,t (y) to “1”, and setting the y-variable for each arc not involved in Pλ,t (y) to “0”.). We have the following results. Lemma 22 Let λ ∈ (0, 1] and y ∈ Yeλ . Then, the following is true: ∀t ∈ Tλ (y), ∃ε ∈ (0, 1] : (y − ε · w bλ,t (y)) ∈ Yeλ−ε ∪ {0}. Proof. From (42) in the proof of Lemma 19 and the fact that w bλ,t (y) is a point of Y , we must have: ∀t ∈ Tλ (y), (λ · w bλ,t (y)) ∈ Yeλ . (45) By the convexity of Yeλ , (45) implies: D E ∀t ∈ Tλ (y), ∃ α ∈ (0, 1], y ∈ Yeλ ∪ {0} : hy = (α · (λ · w bλ,t (y)) + (1 − α) · y)i . (46) From (46), we have: ∀t ∈ Tλ (y), α = 1 =⇒ (1 − α) · y = y − λ · w bλ,t (y) = 0. (47) Hence, it suffices to set ε = 1 in order for the lemma to hold when y = λ · w bλ,t (y). Now, consider t ∈ Tλ (y) : y 6= λ · w bλ,t (y). Using (46), we have: D E ∃ α ∈ (0, 1), y ∈ Yeλ : hy − α · (λ · w bλ,t (y)) = (1 − α) · yi . (48) Letting αλ = ε, we get from (48) that: ∀t ∈ Tλ (y) : y 6= λ · w bλ,t (y), (y − ε · w bλ,t (y)) ∈ Yeλ(1−α) = Yeλ−ε . The lemma follows directly from the combination of (47) and (49). Our main result for this section will now be discussed. 21 (49) Theorem 23 The y-Polytope is integral with each extreme point corresponding to a TSP path of a point of Y . In other words, every extreme point of Y is integral and corresponds to a TSP path of a point of Y . Proof. Let λ ∈ (0, 1], y ∈ Yeλ . ∀t ∈ Tλ (y), let: n o fλt ,t (y t ) := max ε ∈ (0, 1] : (y − ε · w bλ,t (y)) ∈ Yeλ−ε ∪ {0} (50) Then, clearly, from Lemma 22, we must have: ∀t ∈ Tλ (y), fλt ,t (y t ) > 0. (51) Now, consider the “Iterative Elimination (IE)” procedure below: Step 0 (Initialization) (0.1) : set λ1 = 1 (0.2) : set y 1 = y (0.4) : set t = 1 Step 1 (Iterative step) (1.1) : choose a (arbitrary) TSP path of y t , P (1.2) : set Pλt ,t (y t ) = P (1.3) : compute fλt ,t (y t ); (Using (50)) (1.4) : compute y t+1 = y t − fλt ,t (y t ) · w bλt ,t (y t ); Step 2 (Stopping criterion) (2.1) : If y t+1 = 0, stop (2.2) : Otherwise: - compute λt+1 = λt − fλt ,t (y t ) - set t = t + 1 - go to Step 1. (Using Notation 21) Table 1: Summary of the “Iterative Elimination (IE)” procedure Clearly, it follows directly from the combination of (51) and Lemma 22 above that this procedure must stop. Letting the number of iterations of the procedure when it stops be ν, we must have:  ν  P   fλt ,t (y t ) = λ1 = 1; and  (i) t=1 ν  P   fλt ,t (y t ) · w bλt ,t (y t )  (ii) y = y 1 = (52) t=1 Also, clearly, we have: fλt ,t (y t ) > 0 for t = 1, . . . , ν. (53) It follows from the combination of the Minkowski-Weyl Theorem (Minskowski (1910); Weyl (1935); see also Rockafellar (1997, pp.153-172)) that (52)-(53) imply that every extreme point of Y is the characteristic vector of a TSP path of a point of Y . The theorem follows from this directly. 22 Corollary 24 The optimization model comprised of Objective Function (18) and Constraints (1)(17) correctly solves the TSP.   x Proof. For ∈ Q, write Objective Function (18) as ct · y. Denote the value of the optimizay  ∗   x x ∗ tion problem by V and let be an optimal solution. Clearly, ∈ Q implies that y is a ∗ y y convex combination of extreme points of Y . Hence, there must Pp exist p (p ≥ 1) extreme points of Y, yb1 , . . . , ybp , and scalars α1 , . . . , αp (αi ∈ (0, 1], i = 1 . . . , p; i=1 αi = 1) such that the following are true: ∗ t ∗ V =c ·y = p X αi ct · ybi  (54) i=1 = ct · ybi ∀i = 1, . . . , p. (Since we must have V∗ ≤ ct · ybi for all i.) (55) Now, by Theorem 23, each ybi (i = 1, . . . , p) in (54)-(55) corresponds to a TSP path of a point of Y, and therefore (by Theorem 7 and Remark 11) to a TSP tour. Remark 25 Clearly, because of Theorem 23, any off-the-shelf, standard LP solver can be used in order to solve our model. When the TSP at hand has a unique optimum, then any solver would stop with that optimum. On the other hand, when there are multiple optima and an interior-point method is used (or, perhaps, due to numerical issues), the solution obtained may be a convex combination of those optima. In such a case, formal procedures exist, which can be used in order to obtain a vertex optimum (which, according to Theorem 23, would correspond to exactly one of the optimal tours of the TSP at hand). In particular, the “Normal Perturbation” approach of Mangasarian (1984) can be used in order to perturb the LP costs, and thereby “make” the LP into one in which one of the alternate optima becomes the unique optimum. A polynomial-time interior-point method which stops at a vertex solution only (see Wright (1997; pp. 137-157) could be used also, perhaps in combination with the Normal Pertubation approach. In our substantial emprical testing (to be described in the next section), we used a heuristic implementation of the IE procedure described in the proof of Theorem 23 above. In all cases, this heuristic implementation was successful at “retrieving” at least one of the optimal TSP tours comprising the solution. Example 26 We illustrate Theorem 23 and Remark 25 using a 8-city TSP we solved using the barrier method of CPLEX without crossover and thus stopping on a face, as follows: • Travel costs: 23 • Optimal solution obtained: – Solution Value: 281.000. – Postive y-variables: • TSP tours “retrieved” using the IE procedure-based heuristic: 24 • Illustration on the TSPFG:  5 Model size and empirical testing We have run over a million test problems using the interior point (barrier) methods of CPLEX. In all of the cases with multiple optima, our heuristic implementation of the Iterative Elimination procedure described above (in the proof of Theorem 23) was sufficient in “retrieving” at least one of the optimal TSP tours comprising the final LP solution obtained. The optimal TSP tour lengths were also verified by using the Miller-Tucker-Zemlin (1960) model and solving the corresponding integer program using CPLEX. Through our empirical experimentations we found that either constraints (11) or constraints (12) (but not both at the same time) could be left out of the model 25 and still have it solve correctly. Since, the number of the constraints (12) is much smaller than that of constraints (11), we only used constraints (12) in all of the computational experiments we are reporting in this paper. We solved the problems using the CPLEX option of barrier without crossover as it results in much faster solution times, the reason being that even if we stop on a face our iterative procedure is much faster than having CPLEX use crossover and try to get to a vertex. The figures below represent problems solved on a Dell Precision Tower 5810 XCTO Base (210-ACQM)—-64GB (4x16GB) 2133MHz DDR4 RDIMM ECC (370-ABWB) —Integrated Intel AHCI chipset SATA controller (6 x 6.0Gb/s) SW RAID 0/1/5/10 (403-BBGV). Table 2 shows the size and solutions times of problems with 7 to 22 cities. (Although, not reported in this poaper, using a higher-memory (but slower) computer, we were able to solve problems as large as 27 cities.) Each of the times shown in Table 2 is the average of five (5) problems. These problems were symmetric euclidean distance based randomly generated problems. In detail, we used a 100 by 100 grid and modified the distances to between 90%-110% of their true values and rounded distances to be integer. Similar results were obtained using exact euclidean distances, using uniform distributions for distances, and choices of symmetric/asymmetric/integer or non-integer values. Table 2: Problem size and solution time Figures 13 and 14 below graphically display the increase in the number of variables as n (the number of cities) ranges from 7 to 22. Figure 15 shows how the number of constraints grows with respect to the number of variables. While the complexity order of size of the model is O(n5 ), we found that a regression equation using a polynomial of order n3 fits the model size with coefficients of determination which are close to 1.000 (greater that 0.999, to be more precise). This is likely due to the many “implicit zero” variables in the model. We also note that the number of constraints grows more slowly than the number of variables as can be seen in Figure 15. We certainly recognize that while this model is not computationally competitive with other approaches, it does provide motivation to explore other models/algorithms with the knowledge that the complexity classes “P ” and “N P ” are equal. Also, further research may uncover ways to streamline the proposed LP, enough to make it practical. 26 Figure 13: Graph of the number of variables versus the number of cities Figure 14: Graph of the number of constraints versus the number of cities 27 Figure 15: Graph of the number of constraints versus the number of variables The model presented here can be modified in many ways. For example, clearly, a straightforward alternate model could be obtained by substituting out the y-variables, resulting in a model described in x-variable space only. What we have shown here is the model we found easiest to describe and understand. 6 Conclusions We have presented a polynomial-sized linear program for the n-city TSP drawing upon “complex flow modeling” ideas. We used an Assignment problem-based abstraction of tours not employing the traditional city-to-city variables of the standard TSP formulation. Integer solutions were obtained in the case of unique optima, and a simple iterative procedure was used to obtain one or more of the optimal TSP routes in the case of alternate optima. We have solved more than one million problems with up to 27 cities using the barrier methods of CPLEX, consistently obtaining all integer solutions, consistent with our theoretical developments. Our work complements our earlier affirmation resolving the important “P versus N P ” question. It accomplishes this with a much smaller model than that developed previously by the first two authors. Paraphrasing/quoting from Diaby and Karwan (2016a), we conclude: ‘Our developments (and their incidental consequence of “P = N P ”) remove the exponential shift in complexity, but do not suggest a collapse of the “continuum of difficulty,” nor any change in the sequence along that continuum. In other words, our developments do not imply (or suggest) that all of the problems in the NP class have become equally “easy” to solve in practice. The suggestion is that, in theory, for NP problems, the “continuum of difficulty” actually ranges from low-degree-polynomial time complexity to increasingly-higher-degree-polynomial time complexities. However, from a theoretical perspective, we believe that these results make it necessary to reframe the computational complexity question away from: “Does there exist a polynomial algorithm for Problem X?” to (perhaps): What is the smallest-dimensional space in which Problem X has a polynomial algorithm?’ 28 References [1] Bazaraa, M.S., J.J. Jarvis and H.D. Sherali (2010). Linear Programming and Network Flows. Wiley, New York, NY. [2] Beachy, J.A. and W.D. Blair (2006). Abstract Algebra. Waveland Press, Inc., Long Grove, IL. [3] Braun, G., S. Fiorini, S. Pokutta, and D. Steurer (2015). Approximation Limits of Linear Programs (Beyond Hierarchies). Mathematics of Operations Research 40:3, pp. 756–772. [4] Diaby, M. (2007). The traveling salesman problem: A linear programming formulation. WSEAS Transactions on Mathematics 6:6, pp. 745–754. [5] Diaby, M. (2010a). Linear programming formulation of the multi-depot multiple traveling salesman problem with differentiated travel costs. In D. Davendra (ed.), Traveling Salesman Problem, Theory and Applications. InTech, New York, NY, pp. 257–282. [6] Diaby, M. (2010b). Linear programming formulation of the set partitioning problem. International Journal of Operational Research 8:4, pp. 399–427. [7] Diaby, M. (2010c). Linear programming formulation of the vertex coloring problem. International Journal of Mathematics in Operational Research 2:3, pp. 259–289. [8] Diaby, M. and M.H. Karwan (2016a). Advances in Combinatorial Optimization: Linear Programming Formulations of the Traveling Salesman and Other Hard Combinatorial Optimization Problems. World Scientific Publishing, Singapore. [9] Diaby, M. and M.H. Karwan (2016b). “Limits to the scope of applicability of extended formulations,” International Journal of Mathematics in Operational Research (Forthcoming). [10] Fiorini, S., S. Massar, S. Pokutta, H.R. Tiwary, and R. de Wolf (2011). Linear vs. Semidefinite Extended Formulations: Exponential Separation and Strong Bounds. Unpublished (Available at: http://arxiv.org/pdf/1111.0837.pdf). [11] Fiorini, S., S. Massar, S. Pokutta, H.R. Tiwary, and R. de Wolf (2012). Linear vs. Semidefinite Extended Formulations: Exponential Separation and Strong Bounds. Proceedings of the 44 th ACM Symposium on the Theory of Computing (STOC ’12 ), New York, NY, pp. 95-106. [12] Fiorini, S., S. Massar, S. Pokutta, H.R. Tiwary, and R. de Wolf (2015). Exponential Lower Bounds for Polytopes in Combinatorial Optimization. Journal of the ACM 62:2, Article No. 17. [13] Gamelin, T.W. and R.E. Greene (1999). Introduction to Topology. Dover Publications, Inc., Mineola, NY. [14] Garey, M.R. and D.S. Johnson (1979). Computers and Intractability: A Guide to the Theory of NP-Completeness. Freeman and Company, New York, NY. [15] Lawler, E.L., J.K. Lenstra, A.H.G. Rinnooy Kan, and D.B. Shmoys (1985), eds. The Traveling Salesman Problem: A Guided Tour of Combinatorial Optimization (Wiley, New York). 29 [16] Mangasarian, O.L. (1984). Normal solutions of linear programs. Mathematical Programming Study 22, pp. 206–216. [17] Miller, C.E., A.W. Tucker and R.A. Zemlin (1960). Integer programming formulations and traveling salesman problems. Journal of the Association for Computing Machinery 7:4, pp. 326–329. [18] Minkowski, H. (1910). Geometrie der Zahlen. Teubner, Leipzig. [19] Öncan, T., K. Altinel and G, Laporte (2009). A comparative analysis of several asymmetric traveling salesman problem formulations. Computers & Operations Research 36:3, pp. 637–654. [20] Panik, M.J. (1993). Fundamentals of Convex Analysis. Kluwer Academic Publishers, Boston, MA. [21] Rockaffelar, R. T. (1997). Convex Analysis. Princeton University Press. [22] Weyl, H. (1935). Elementare Theorie der konvexen Polyheder. Commentarii Math. Helvetici 7, pp. 290-306. [23] Wright, S.J. (1997). Primal-Dual Interior-Point Methods. SIAM. [24] Yannakakis, M. (1991). Expressing combinatorial optimization problems by linear programming. Journal of Computer and System Sciences 43:3, pp. 441-466. [25] Yannakakis, M. (2013). Private Email Communications. 30 Appendix A: Non-applicability of developments for “The TSP Polytope.” Our model does not require the city-to-city traditional variables that are used in describing the standard TSP Polytope (i.e., “The TSP Polytope”; see Lawler et al. (1985), or Diaby and Karwan (2016a; 2016b) for exact definition). Hence, our modeling does not involve “The TSP Polytope,” and hence, extended formulations developments which pertain to that polytope and those for other hard combinatorial problems (such as Yannakakis (1991), or Fiorini et al. (2011; 2012; 2015), in particular), do not apply/cannot be applied to the developments in this paper. The reasons for this are fully detailed in Diaby and Karwan (2016a) and Diaby and Karwan (2016b), respectively, but will be briefly overviewed in this section. Pertaining to the Fiorini et al. (2011; 2012; 2015) developments in particular, as shown in Diaby and Karwan (2016a) and Diaby and Karwan (2016b), respectively, the existence of a linear transformation between the descriptive variables of two polytopes does not necessarily imply extension relations between the external descriptions of the polytopes from which valid/meaningful inferences can be made. This important subtlety was demonstrated in Diaby and Karwan (2016a and 2016b, respectively), and illustrated using polyopes consisting of single points (i.e., of dimension zero). In this section, we present a numerical example involving polytopes of dimensions greater than zero. A.1. Alternate abstraction of the TSP Optimization Problem Theorem 27 Consider the TSP defined on the set of cities Ω := {0, . . . , n − 1}. Assume city “0” has been designated as the starting and ending point of the travels. For the sake of greater clarity, let M := Ω\{0} be the set of the remaining cities to be sequenced, and S := {1, . . . , m}, the index set of the travel legs to cities “1” through “n − 1” (where m := n − 1). Then, there exists a one-to-one correspondence between TSP tours and extreme points of ( ) X X 2 wi,t = 1 ∀i ∈ M ; wi,t = 1 ∀t ∈ S; w ≥ 0 . AP := w ∈ R(n−1) : t∈S i∈M Proof. Using the assumption that city “0” is the starting and ending point of travel, it is trivial to construct a unique TSP tour from a given extreme point of AP, and vice versa (i.e., it is trivial to construct a unique extreme point of AP from a given TSP tour). Example 28 We illustrate some of the differences between the city-stage Assignment-based modeling of this paper and the traditional (city-to-city) modeling which induces the standard TSP polytope (i.e., “The TSP polytope”) using a 6-city problem with node set {0, 1, 2, 3, 4, 5}. • Assignment Problem representation of TSP tours: 31 • Illustration of the Assignment Problem representation on the TSP Graph: • Illustration of the standard (city-to-city) representation on the TSP Graph: 32 Remark 29 1. It is not possible to make the city-to-stage Assignment representation induce TSP tours per se by adding a “dummy” city to the set of TSP cities and using it as the starting and ending point of the travels. The reason for this is that the cycle then obtained would not necessarily be a cycle in the actual TSP cities, as the begining and ending cities in that sequence may be different. Hence, the city-to-stage Assignment representation and the traditional (city-to-city) representation of TSP tours cannot be made to become equivalent using constraints expressed in the city-to-stage variables only. In other words, the equivalence (i.e., “making” the cityto-stage Assignment representation induce TSP tours per se) cannot be done mathematically while preserving the (“nice”) AP structure. 2. As concerns the extended formulations “barriers” works (Yannakakis (1991); Fiorini et al. (2011; 2012; 2015)), it is important to observe that AP is also a TSP polytope, since a oneto-one correspondence can be established between its extreme points and TSP tours, and that its mathematical charateristics are unrelated to those of “The TSP Polytope.” According to the Minkowski-Weyl Theorem (Minskowski (1910); Weyl (1935); see also Rockafellar (1997, pp.153-172)), every polytope can be equivalently described as the intersection of hyperplanes (H-representation/external description) or as a convex combination of (a finite number of) vertices (V-representation/internal description). “The TSP Polytope” is easy to state in terms of its V-representation. However, no polynomial-sized H-representation of it is known. On the other hand, the H-representation of AP is well-known to be of (low-degree) polynomial size (see Burkard et al. (2009)), and it is trivial to state its V-representation also. The vertices of AP are Assignment Problem solutions, whereas the vertices of “The TSP Polytope” are Hamiltonian cycles. Hence, even though the extreme points of AP and those of “The TSP Polytope” respectively correspond to TSP tours, the two sets of extreme points are different kinds of mathematical objects, with unrelated mathematical characterizations, and there does not exist any a priori mathematical relation between the two polytopes. In other words, AP and “The TSP Polytope” are simply alternate abstractions of TSP tours. 33 Or, put another way, AP is (simply) an alternate TSP polytope from “The TSP Polytope”, and vice versa (i.e., that the “The TSP Polytope” is (simply) an alternate TSP polytope from AP ). 3. As illustrated in Example 28 and also in the first point above, we do not attempt to induce TSP tours per se in this paper. Instead, we develop a higher-dimensional reformulation of city-to-stage Assignment representation, “building” enough information in the modeling variables themselves that the “judicious manipulation” of the costs called for in order to “complete” the abstraction from an overall optimization point-of-view is possible. Hence, we “side-step” (altogether) the issue of having to derive linear constraints that induce explicit TSP tours (which is at the crux of the difficulty in developing an H-description of “The TSP polytope”). In other words, we do not deal with the issue of describing “The TSP polytope” in this paper. Rather we simply perform a reformulation of the city-to-stage Assignment representation discussed above in a higher-dimensional space. Clearly, since we are not modeling cycles/TSP tours per se, this approach can become a complete abstraction of the overall TSP optimization problem only if a proper/judicious objective function to apply to the reformulated-AP solutions can be developed. Note that the issue of a proper objective function to apply in “The TSP polytope” (traditional) modeling approach is a trivial one. 4. Somewhat more philosophically, the most fundamental subtlety to grasp may be the fact that there exists no a priori mathematical relationship (projective or otherwise) between our proposed AP-based model and “The TSP polytope.” The associations which can be made between the two conceptualizations are cognitive only, and not “mathematical” a priori. This can be seen, for example, by considering the fact that one could not infer the existence of a (TSP) node “0” from the statement of our constraints only, as indicated in Example 28 above. From a general/common-sense perspective, if the addition of something changes conditions for what was added to, then whatever was added cannot have been “redundant,” since that would be a contradiction of the definition of the term “redundant.” In the context of this paper in particular, the extended formulations relationships that can be created through additions of redundant variables and constraints to our model can be degenerate ones only, from which it is not possible to make any meaningful inference (see Diaby and Karwan (2016a) and Diaby and Karwan (2016b), respectively, for formal developments of these).  A.2. Non-applicability of the Fiorini et al. (2015) results Extended formulations (EFs) have been the dominant theory which has been used in deciding on the validity of proposed LP models for hard combinatorial problems. All of the developments are predicated on the model being evaluated projecting to the “natural” polytope of the specific problem at hand. For the TSP that polytope is the one stated in terms of the natural, standard city-to-city variables, and is referred to in the literature as “The TSP Polytope.” We show in Diaby and Karwan (2016a) that the generalized version of the model proposed in this paper is not symmetric. Hence, among other reasons, the results in the seminal Yannakakis (1991)) work are not applicable to the model in this paper. Our developments in Diaby and Karwan (2016a and 2016b, respectively) specifically show the non-applicability of the Fiorini et al. (2011; 2012) developments to models which do not involve the variables of “The TSP Polytope” in general, and provide counter-examples using zero-dimensional polytopes (i.e., singletons) for convenience. In the latest version of their papers, Fiorini et al. (Fiorini et al. (2015)) stipulate that the polytopes 34 considered must have dimensions greater than zero. The objective of this section is to show that the counter-examples shown in Diaby and Karwan (2016a and 2016b, respectively) for the Fiorini et al. (2011; 2012) developments remain valid for the Fiorini et al. (2015) work. We do this by exhibiting polytopes of dimensions greater than zero which refute key foundation results in Fiorini et al. (2015). For convenience, we will start with a statement of the standard definition of an “extended formulation” as well as those of the alternate definitions used Fiorini et al. (2015). Then, we will discuss our numerical example. The discussions to follow apply to the developments in Braun et al. (2015) as well, since those developments are based on the same notions of extended formulations, “slack matrices”, and “extension complexity”, in particular. Definition 30 (“Standard EF Definition”) An extended formulation for a polytope X ⊆ Rp is a polyhedron U = {(x, w) ∈ Rp+q : Gx + Hw ≤ g} the projection, ϕx (U ) := {x ∈ Rp : (∃w ∈ Rq : (x, w) ∈ U )}, of which onto x-space is equal to X (where G ∈ Rm×p , H ∈ Rm×q , and g ∈ Rm ) (Yannakakis (1991)). Definition 31 (“Fiorini et al. Definition #1”) A polyhedron U = {(x, w) ∈ Rp+q : Gx + Hw ≤ g} is an extended formulation of a polytope X ⊆ Rp if there exists a linear map π : Rp+q −→ Rp such that X is the image of U under π (i.e., X = π(U ); where G ∈ Rm×p , H ∈ Rm×q , and g ∈ Rm ) (see Fiorini et al (2015; p. 17:3, lines 20-21; p. 17:9, lines 22-23)). Definition 32 (“Fiorini et al. Definition #2”) An extended formulation of a polytope X ⊆ Rp is a linear system U = {(x, w) ∈ Rp+q : Gx + Hw ≤ g} such that x ∈ X if and only if there exists w ∈ Rq such that (x, w) ∈ U. (In other words, U is an EF of X if (x ∈ X ⇐⇒ (∃ w ∈ Rq : (x, w) ∈ U ))) (where G ∈ Rm×p , H ∈ Rm×q , and g ∈ Rm ) (see Fiorini et al. (2015; p. 17:2, last paragraph; p. 17:9, line 20-21)). Our numerical example will now be discussed. Example 33 :Let x ∈ R3 and w ∈R be disjoint vectors of variables. Let X be a polytope in the   w space of x, and U, a polytope in the space of , with: x     8 12   X := Conv   10  ,  15   , and (56)   6 9    w U := ∈ R4 : 2 ≤ 0 · x + w ≤ 3 (57) x We now discuss some key results of Fiorini et al. (2015) which are refuted by X and U. 1. Refutation of the validity of Definition 31. (a) Note that the following is true for X and U :      w x ∈ X < ∃w ∈R : ∈U . x 35 (58) For example,        w 22.5   22.5   ∃w ∈R :     −50  ∈ X  .   −50  ∈ U  ; 100 100 (59) Hence, U is not an extended formulation of X according to Definition 32. (b) Observe that the following is also true for X and U :  X=         4 0 0 0 w w x ∈R3 : x = A · , ∈U , where A=  5 0 0 0  . x x 3 0 0 0 (60) In other words, X is the image of U under the linear map A. Hence, U is an extended formulation of X according to Definition 31. (c) It follows from (a) and (b) above, that Definitions 31 and 32 are in contradiction of each other with respect to X and U . Hence X and U are a refutation of the validity of Definition 31 (since it is easy to verify the equivalence of Definition 32 to Definition 30, which is the “standard” definition). 2. Refutation of “Theorem 3” (p.17:10) of Fiorini et al. (2015). The proof of the theorem (“Theorem 3”) hinges on Definition 32. The specific statement in Fiorini et al. (2015; p. 17:10, lines 26-28) is: “...Because Ax ≤ b ⇐⇒ ∃y : E = x + F = y = g = , E ≤ x + F ≤ y = g = , (61) each inequality in Ax ≤ bis valid for all points of Q. ...” The equivalent of (61) in terms of X and U is:   w x ∈ X ⇐⇒ ∃w ∈R : ∈ U. x (62) Clearly, (62) is not true, as we have illustrated in Part (1.a) above. Hence, the proof of “Theorem 3” (and therefore, “Theorem 3”) of Fiorini et al. (2015) is refuted by X and U . 3. Refutation of “Lemma 9” (p. 17:13-17:14) of Fiorini et al. (2015). The first part of the lemma is stated (in Fiorini et al. (2015)) thus: “Lemma 9. Let P, Q, and F be polytopes. Then, the following hold: (i) if F is an extension of P, then xc(F) ≥ xc(P);. . . ” The proof of this is stated as follows: “Proof. The first part is obvious because every extension of F is in particular an extension of P. . . . ” 36 The notation “xc(·)” stands for “extension complexity of (·),” which is defined as (p. 17:9, lines 24-25 of Fiorini et al. (2015)): “...the extension complexity of P is the minimum size (i.e., the number of inequalities) of an EF of P.” The refutation of these for X (as shown in (56) above) and U (as shown in (57) above) is as follows. As shown in Part (1) above, U is an extension of X according to Definition 31 (which is central in Fiorini et al. (2015)). This means that U is an extended formulation of every one of the infinitely-many possible H-descriptions of X. This would be true in particular for the H-description below for X:   x ∈R3 :                  −5x1 + 4x2 ≤ 0;                  3x − 5x = 0;   2 3         3x1 − 4x3 ≤ 0; X := . (63)               8 ≤ x ≤ 12; 1                 10 ≤ x ≤ 15; 2               6 ≤ x3 ≤ 9 Clearly, however, we have that: xc(U )  xc(X). (64) Hence, X and U are a refutation of “Lemma 9” of Fiorini et al. (2015), being that U is the extension, and X, the projection, according to definitions used in Fiorini et al. (2015).  According to Fiorini et al. (2015; p. 17:7, Section 1.4, first sentence; p. 17:11, lines 6-11; p. 17:14, lines 5-6; p.17:16, lines 13-14 after the “Fig. 4”), their “Theorem 3” and “Lemma 9” play pivotal, foundational roles in the rest of their developments. Hence, we believe we have offered a simple-yet-complete refutation of their developements when polytopes are described in terms of disjoint sets of variables, as is the case for the model in this paper in relation to “The TSP Polytope.” Hence, no part of the Fiorini et al. (2015) developments is applicable to the model developed in this paper. 37 A.3. Meaning of the existence of a linear transformation We will now provide some insights into the correct meaning/consequence of the existence of an affine map establishing a one-to-one correspondence between two polytopes that are stated in disjoint variable spaces, as brought to our attention in private e-mails by Yannakakis (2013). The linear map stipulated in Fiorini et al. (2012; 2015) in particular, is a special case of the affine map. In the case of polytopes stated in disjoint variable spaces, if the constraints expressing the affine transformation are redundant for each of the models/polytopes, the implication is that one model can be used in an “auxiliary” way, in order to solve the optimization problem over the other model, without any reference to/knowledge of the H-description of that other model. This is shown in Remark 34 below. Remark 34 • Let: – x ∈ Rp and y ∈ Rq be disjoint vectors of variables; – X := {x ∈ Rp : Ax ≤ a}; – L := {(x, y) ∈ Rp+q : Bx + Cy = b}; – Y := {y ∈ Rq : Dy ≤ d}; (Where: A ∈ Rk×p ; a ∈ Rk ; B ∈ Rm×p ; C ∈ Rm×q ; b ∈ Rm ; D ∈ Rl×q , d ∈ Rl ). • If B T B is nonsingular, then L can be re-written in the form: L = {(x, y) ∈ Rp+q : x = Cy + b}. (Where: C := −(B T B)−1 B T C, and b := (B T B)−1 B T b). (65) Hence, the linear map stipulated in Definition 31 is simply a special case of L in which b = 0 and B T B is nonsingular. • Assume that: – L 6= ∅ exists, with constraints that are redundant for X and Y , respectively; – the non-negativity requirements for x and y are included in the constraint sets of X and Y , respectively; and that: – B T B is nonsingular. (This is equivalent to assuming that the more general (affine map) version of the linear map stipulated in Definition 31 exists.) • Then, the optimization problem: Problem LP 1 : 38 Minimize: αT x Subject To: (x, y) ∈ L; y ∈ Y (where α ∈ Rp ). is equivalent to the smaller linear program: Problem LP 2 : Minimize: Subject To:  αT C y + αT b y∈Y (where α ∈ Rp ). • Hence, if L is the graph of a one-to-one correspondence between the points of X and the points of Y (see Beachy and Blair (2006, pp. 47-59)), then, the optimization of any linear function of x over X can be done by first using Problem LP 2 in order to get an optimal y, and then using Graph L to “retrieve” the corresponding x. Note that the second term of the objective function of Problem LP 2 can be ignored in the optimization process of Problem LP 2 , since that term is a constant. Hence, if L is derived from knowledge of the V-representation of X only, then this would mean that the H-representation of X is not involved in the “two-step” solution process (of using Problem LP 2 and then Graph L), but rather, that only the V-representation of X is involved.  Hence, the existence of a linear map between points of the model in this paper and points of “The TSP Polytope” would simply imply that our model can be used in an “auxiliary” way, in order to solve the TSP optimization problem without any reference to/knowledge of the H-description of “The TSP Polytope” since, as we have discussed above, our modeling is independent of the traditional variables of “The TSP Polytope” and the linear map could only be inferred from our knowledge of the V-description of “The TSP Polytope.” Hence, there does not exist any meaningful extension relationship between our constraint sets and an H-description of “The TSP Polytope.” Hence, the developments of Fiorini et al. (2012; 2015) in particular, which are predicated on the existence of such a linear map, are not applicable to the developments in this paper. Also, with respect to the developments in Fiorini et al. (2015) specifically, note that the Boolean Satisfiability Problem can be formulated, in a straightforward manner, as a linear program using the more generalized, generic version of the model in this paper developed in Diaby and Karwan (2016a). 39 Appendix B: Software Implementation • General Description and Interface A software package, TSP LP Solver, has been submitted as supplementary files to this paper. TSP LP Solver builds linear programming (LP) models for the traveling salesman problem and calls CPLEX to solve them as LPs. The interface has been designed to run multiple replications of the chosen problem and run control settings at a time. With this tool, users can: (1) randomly generate or read a TSP cost matrix in multiple ways; (2) directly solve the TSP or only build the LP models; (3) adjust CPLEX settings for different tests; (4) show solutions (optimal objective, variables, routes) in different formats. The MTZ model is available for the purposes of verifying the correctness of the solutions obtained using our LP model. It is solved as an Integer Program, and only its objective function value is displayed. A screenshot of the TSP LP Solver is shown in Figure 16. Figure 16: First Screen of the TSP LP Solver • Requirements TSP LP Solver is written in C# with .NET framework 4 and calls CPLEX 12.5 to solve models. For those with CPLEX 12.5, all functions in this program are available. For those with different 40 versions of CPLEX, problems may happen when calling CPLEX to solve. If so, users need to adjust references in the source code to rebuild the program. Users who do not wish to use CPLEX or to modify the source code can choose the Model Only button in order to build a .lp file (no requirement to use CPLEX) which they can then solve using the software of their choice. • Data There are two ways to input data to the solver: randomly generate and read cost data from files. For either way, users can check “Export all replications in XML format” to export cost files in XML format for every replication. Randomly generating data supports the testing of multiple replications of a problem in a single run. Users input # of Cities (number of cities) and # of Replications. Cost values are generated based on either Euclidean distances or uniformly distributed random numbers. If the Euclidean distance option is chosen, the program will first randomly generate coordinates within a (0, 100) x (0, 100) square plane, and then randomly generate costs within the given percentage range of Euclidean distances. If (absolute) interval limits is chosen, the program will directly randomly generate costs within the given range, not based on Euclidean distances. Other options include whether the cost matrix is asymmetric or not (checked or unchecked), whether the cost matrix is integer or not (checked or unchecked), and whether the triangle inequality holds or is not required (checked or unchecked). Reading cost files supports XML and CSV formats as input file formats. The required data format can be found in the attached Sample.xml and Sample.csv. The XML data format follows the classic TSPLIB. • Modelers and Solvers – Modeler Settings If the Model Only button is chosen, the program will build an .lp file without the requirement to use CPLEX. If the Miller-Tucker-Zemlin (MTZ) model is chosen to be solved, the program will call CPLEX to build and solve the model and display the optimal objective value for reference. If the N5 LP model is chosen to be solved, the program will call CPLEX to build and solve the model and display the solution time, optimal objective value and other solution information depending on which are chosen among the show y solutions, show x solutions and show routes options. – CPLEX Settings If users have the correct CPLEX version on their machines, they can adjust CPLEX parameters with this tool and solve the model with different algorithmic settings. For details of each adjustable parameter, please refer to a CPLEX Parameters Reference from IBM. • Results All output files are located in the “Results” subfolder of the folder containing the TSP LP Solver executable (“TSPsolvers.exe”), including the XML cost files, .lp files and solution text files. 41 The solution text file is named as “Tests on #-node random cases.txt” for randomly generated problems and ”Tests on given cost files.txt” for tests with reading cost files. This file may include the MTZ objective value, along with the N5 LP model objective value, solution time, number of variables, number of constraints, values of the non-zero y variables, values of the non-zero x variables, and optimal routes. The optimal routes are retrieved using our iterative elimination procedure. In the case that many alternate optimal TSP routes are involved in the solution (e.g. when a non-crossover interior-point method stops on a face due to alternate optima), the program will display all or a subset of these routes for reference. If it is desired that all of the alternate optimal routes in the solution be always displayed, users need to adjust the source code. We provide two versions of the iterative elimination in the source code: SplitRoutesByY() is a greedytype implementation; EnumerSplitByY() involves an enumeration scheme for “tracing” the TSP paths through the y-variables with positive values in the solution. EnumerSplitByY() is more robust to avoid failure due to numerical issues, but is time-consuming. Besides these two methods, users can also apply perturbation means (Mangasarian, 1984) to make one the alternate optima the unique LP optimum, or a polynomial-time interior-point method which stops at a vertex optimum such as described in Wright (1997). If any one of the following two cases is found, the program will create XML cost files for records for further analysis: (1) MTZ and N5 LP models result in different optimal objective values (this has never happened in our more than 1 million tests); (2) There are many optimal routes and the program only displays a subset of them. 42
8cs.DS
Maximizing the Weighted Number of Spanning Trees: arXiv:1604.01116v2 [cs.DS] 12 Apr 2016 Near-t-Optimal Graphs∗ Kasra Khosoussi† Gaurav S. Sukhatme‡ Shoudong Huang† Gamini Dissanayake† March 12, 2018 Abstract Designing well-connected graphs is a fundamental problem that frequently arises in various contexts across science and engineering. The weighted number of spanning trees, as a connectivity measure, emerges in numerous problems and plays a key role in, e.g., network reliability under random edge failure, estimation over networks and D-optimal experimental designs. This paper tackles the open problem of designing graphs with the maximum weighted number of spanning trees under various constraints. We reveal several new structures, such as the log-submodularity of the weighted number of spanning trees in connected graphs. We then exploit these structures and design a pair of efficient approximation algorithms with performance guarantees and near-optimality certificates. Our results can be readily applied to a wide verity of applications involving graph synthesis and graph sparsification scenarios. ∗ Working paper. [email protected] – https://kasra.github.io for Autonomous Systems (CAS), University of Technology Sydney. ‡ Department of Computer Science, University of Southern California. † Centre Contents 1 Introduction 3 2 Background 3 2.1 2.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Matrix-Tree Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 4 3 Tree-Connectivity 5 4 ESP: Edge Selection Problem 7 4.1 4.2 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Exhaustive Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 8 4.3 4.4 Greedy Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Convex Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 10 4.5 Certifying Near-Optimality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 4.6 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 5 Beyond k-ESP+ 5.1 5.2 13 Matroid Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1.1 Greedy Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 14 5.1.2 Convex Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Dual of k-ESP+ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 15 5.2.1 5.2.2 Greedy Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Convex Relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 16 5.2.3 Certifying Near-Optimality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 6 Conclusion 17 A Proofs 20 1 Introduction Various graph connectivity measures have been studied and used in different contexts. Among them are the combinatorial measures, such as vertex/edge-connectivity, as well as spectral notions, like algebraic connectivity [10]. As a connectivity measure, the number of spanning trees (sometimes referred to as graph complexity or tree-connectivity) stands out in this list since despite its combinatorial origin, it can also be characterized solely based on the spectrum of graph Laplacian. It has been shown that tree-connectivity is associated with D-optimal (determinant-optimal) experimental designs [8, 6, 1, 26]. The number of spanning trees also appears in the study of all-terminal network reliability under (i.i.d.) random edge failure (defined as the probability of network being connected) [16, 30]. In particular, it has been proved that for a given number of edges and vertices, the uniformly-most reliable network, upon existence, must have the maximum number of spanning trees [3, 22, 4]. The graph with the maximum number of spanning trees among a finite set of graphs (e.g., graphs with n vertices and m edges) is called t-optimal. The problem of identifying toptimal graphs under a (n, m) constraint remains open and has been solved only for specific pairs of (n, m); see, e.g., [27, 6, 15, 25]. We prove that the (weighted) number of spanning trees in connected graphs can be posed as a monotone log-submodular function. This structure enables us to design a complementary greedy-convex pair of approximate algorithms to synthesize near-t-optimal graphs under several constraints with approximation guarantees and near-optimality certificates. Notation Throughout this paper, bold lower-case and upper-case letters are reserved for real vectors and matrices, respectively. The standard basis for Rn is denoted by {eni }ni=1 , and en0 is defined to be the zero n-vector. For any n ∈ N, [n] denotes the set N≤n = {1, 2, . . . , n}. Sets are shown by upper-case letters. |X | denotes the  cardinality of set X . For any finite set W, W k is the set of all k-subsets of W. The eigenvalues of symmetric matrix M are denoted by λ1 (M) ≤ · · · ≤ λn (M). 1, I and 0 denote the vector of all ones, identity and zero matrix with appropriate sizes, respectively. S1 ≻ S2 means S1 − S2 is positive-definite. The Euclidean norm is denoted by k · k. diag(Wi )ki=1 is the block-diagonal matrix with matrices (Wi )ki=1 as blocks on its main diagonal. For any graph G, E(G) denotes the edge set of G. Finally, Sn≥0 and Sn>0 denote the set of symmetric positive semidefinite and symmetric positive definite matrices in Rn×n , respectively. 2 2.1 Background Preliminaries Let G = (V, E) be an undirected graph over V = [n] and with |E| = m edges. By assigning a positive weight to each edge of the graph through w : E → R>0 , we obtain Gw = (V, E, w). To shorten our notation let us define wuv , w(u, v) = w(v, u). As it will become clear shortly, without loss of generality we can assume G is a simple graph since (i) loops do not affect the number of spanning trees, and (ii) parallel edges can be replaced by a single edge whose weight is the sum of the weights of the parallel edges. W , diag (w(e1 ), . . . , w(em )) denotes the weight matrix in which ei ∈ E is the ith edge. The degree of vertex v ∈ V in G is denoted by deg(v). Let à be the incidence matrix of G after assigning arbitrary orientations 3 to its edges. The Laplacian matrix of G is defined as L̃ , ÃÃ⊤ . For an arbitrary choice of v0 ∈ V , let A ∈ {−1, 0, 1}(n−1)×m be the matrix obtained by removing the row that corresponds to v0 from Ã. We call A the reduced incidence matrix of G after anchoring v0 . The reduced Laplacian matrix of G is defined as L , AA⊤ . L is also known as the Dirichlet or grounded Laplacian matrix of G. Note that L can also be obtained by removing the row and column associated to the anchor from the graph Laplacian matrix. A is full column rank and consequently L is positive definite, iff G is connected. For weighted graphs, AWA⊤ is the reduced weighted Laplacian of Gw . Note that this is a natural generalization of L, and will reduce to its unweighted counterpart if all weights are equal to one (i.e., W = I). The reduced (weighted) Laplacian matrix can be decomposed into the (weighted) sum of elementary reduced Laplacian matrices: L= X w(u, v)Luv (1) {u,v}∈E in which Luv , auv a⊤ uv and auv = eu − ev is the corresponding column of A. 2.2 Matrix-Tree Theorems The spanning trees of G are spanning subgraphs of G that are also trees. Let TG and t(G) , |TG | denote the set of all spanning trees of G and its number of spanning trees, respectively. Let Tn and Kn be, respectively, an arbitrary tree and the complete graph with n vertices. The following statements hold. 1. t(G) ≥ 0, and t(G) = 0 iff G is disconnected, 2. t(Tn ) = 1, 3. t(Kn ) = nn−2 (Cayley’s formula), 4. if G is connected, then t(Tn ) ≤ t(G) ≤ t(Kn ), 5. if G1 is a spanning subgraph of G2 , then t(G1 ) ≤ t(G2 ). Therefore t(G) is a sensible measure of graph connectivity. The following theorem by Kirchhoff provides an expression for computing t(G). Theorem 2.1 (Matrix-Tree Theorem [10]). Let LG and L̃G be, respectively, the reduced Laplacian and the Laplacian matrix of any simple undirected graph G after anchoring an arbitrary vertex out of its n vertices. The following statements hold. 1. t(G) = det(LG ), 2. t(G) = 1 n Qn i=2 λi (L̃G ).1 The matrix-tree theorem can be naturally generalized to weighted graphs, where each spanning tree is “counted” according to its value defined below. 1 Recall that the Laplacian matrix of any connected graph has a zero eigenvalue with multiplicity one (see, e.g., [10]). 4 Definition 2.1. Suppose G = (V, E, w) is a weighted graph with a non-negative weight function. The value of each spanning tree of G is measured by the following function, Vw : TG → R≥0 Y w(e). T 7→ (2) (3) e∈ E(T ) Furthermore, we define the weighted number of trees as tw (G) , P T ∈ TG Vw (T ). Theorem 2.2 (Weighted Matrix-Tree Theorem [20]). For every simple weighted graph G = (V, E, w) with w : E → R>0 we have tw (G) = det AWA⊤ . Note that Theorem 2.2 reduces to Theorem 2.1 if w(e) = 1 for all e ∈ E. Therefore, in the rest of this paper we focus our attention mainly on weighted graphs. Definition 2.2. The weighted tree-connectivity of graph G is formally defined as τw (G) , 3 Tree-Connectivity  log tw (G) 0 if tw (G) > 0, (4) otherwise. Definition 3.1. Consider an arbitrary simple undirected graph G◦ . Let pi be the probability assigned to the ith edge, and p be the stacked vector of probabilities. G ∼ G(G◦ , p) indicates that 1. G is a spanning subgraph of G◦ . 2. The ith edge of G◦ appears in G with probability pi , independent of other edges. The naive procedure for computing the expected weighted number of spanning trees in such random graphs involves a summation over exponentially many terms. Theorem 3.1 offers an efficient and intuitive way of computing this expectation in terms of G◦ and p. Theorem 3.1. For any G(G◦ , p) and w : E(Kn ) → R>0 ,   EG∼G(G◦ ,p) tw (G) = twp (G◦ ), (5) where wp (ei ) , pi w(ei ) for all ei ∈ E(G◦ ). Note that this expectation can now be computed in O(n3 ) time for general G◦ . Lemma 3.1. Let G+ be the graph obtained by adding {u, v} ∈ / E with weight wuv to G = (V, E, w). Let LG be the reduced Laplacian matrix and auv be the corresponding column of the reduced incidence matrix of G after anchoring an arbitrary vertex. If G is connected, τw (G+ ) = τw (G) + log(1 + wuv ∆G uv ), ⊤ −1 where ∆G uv , auv LG auv . 5 (6) Lemma 3.2. Similar to Lemma 3.1, let G− be the graph obtained by removing {p, q} ∈ E with weight wpq from E. If G is connected, τw (G− ) = τw (G) + log(1 − wpq ∆G pq ). (7) n o Corollary 3.2. Define TGuv , T ∈ TG : {u, v} ∈ E(T ) . Then we have uv uv ∆G uv = |TG |/|TG | = |TG |/t(G). (8) Similarly, for weighted graphs we have wuv ∆G uv P = P T ∈TGuv Vw (T ) T ∈ TG Vw (T ) = P T ∈TGuv Vw (T ) tw (G) . (9) Lemmas 3.1 and 3.2 imply that wuv ∆G uv determines the change in tree-connectivity after adding or removing an edge. This term is known as the effective resistance between u and v. If G is an electrical circuit where each edge represents a resistor with a conductance equal to its weight, then wuv ∆G uv is equal to the electrical resistance across u and v. The effective resistance also emerges as a key factor in various other contexts; see, e.g., [9, 2, 19]. Note that although we derived ∆G uv using the reduced graph Laplacian, it is more common to define the effective resistance using the pseudoinverse of graph Laplacian L̃G [9]. Now, on a seemingly unrelated note, we turn our attention to structures associated to tree-connectivity when seen as a set function. Definition 3.2. Let V be a set of n ≥ 2 vertices. Denote by GE the graph (V, E) for any E ∈ E(Kn ). For any w : 2E(Kn ) → R>0 define treen,w : 2E(Kn ) → R≥0 E 7→ tw (GE ), (10) log treen,w : 2E(Kn ) → R E 7→ τw (GE ). (11) Definition 3.3 (Tree-Connectivity Gain). Suppose a connected base graph (V, Einit ) with n ≥ 2 vertices and an arbitrary positive weight function w : E(Kn ) → R>0 are given. Define logTGn,w : 2E(Kn) → R≥0 E 7→ log treen,w (E ∪ Einit ) − log treen,w (Einit ). Definition 3.4. Suppose W is a finite set. For any ξ : 2W → R, 1. ξ is called normalized iff ξ(∅) = 0. 2. ξ is called monotone if ξ(B) ≥ ξ(A) for every A and B s.t. A ⊆ B ⊆ W. 6 (12) 3. ξ is called submodular iff for every A and B s.t. A ⊆ B ⊆ W and ∀s ∈ W \ B we have, ξ(A ∪ {s}) − ξ(A) ≥ ξ(B ∪ {s}) − ξ(B). (13) 4. ξ is called supermodular iff −ξ is submodular. 5. ξ is called log-submodular iff ξ is positive and log ξ is submodular. Theorem 3.3. treen,w is normalized, monotone and supermodular. Theorem 3.4. logTGn,w is normalized, monotone and submodular. Corollary 3.5 follows directly from Theorems 3.1, 3.3 and 3.4. Corollary 3.5. The expected weighted number of spanning trees in random graphs is normalized, monotone and supermodular when seen as a set function similar to treen,w . Moreover, the expected weighted number of spanning trees can be posed as a log-submodular function similar to logTGn,w . ESP: Edge Selection Problem 4 4.1 Problem Definition Suppose a connected base graph is given. The edge selection problem (ESP) is a combinatorial optimization problem whose goal is to pick the optimal k-set of edges from a given candidate set of new edges such that the weighted number of spanning trees after adding those edges to the base graph is maximized. Problem 4.1 (ESP). Let Ginit = (V, Einit , w) be a given connected graph where w : E(Kn ) → R>0 . Consider the following scenarios. 1. k-ESP+ : For some M+ ⊆ E(Kn ) \ Einit , maximize + tw (GEinit ∪E ) subject to |E| = k. maximize tw (GEinit \E ) E⊆M (14) 2. k-ESP− : For some M− ⊆ Einit , E⊆M− subject to (15) |E| = k. Remark 1. It is easy to see that every instance of k-ESP− can be expressed as an instance of d-ESP+ problem for a different base graph, some d and a candidate set M+ (and vice versa). Remark 2. The open problem of identifying t-optimal graphs among all graphs with n vertices and m edges [4] is an instance of k-ESP+ with k = m, Einit = ∅ and M+ = E(Kn ). Remarks 1 and 2 ensure that any algorithm designed for solving k-ESP+ carries over to the other forms of ESP. Therefore, although many graph sparsification and edge pruning scenarios can be naturally stated as a k-ESP− , in the rest of this paper we focus our attention mainly on k-ESP+ . 7 4.2 Exhaustive Search The brute force algorithm for solving k-ESP+ requires computing the weighted tree-connectivity of every k-subset of the candidate set. tw (G) can be computed by performing a Cholesky decomposition on the reduced weighted Laplacian matrix which requires O(n3 ) time in general. This time may significantly reduce for sparse graphs. Let c , |M+ |. For k = O(1), the time complexity of the brute force algorithm is O(ck n3 ). If c = O(n2 ), this complexity becomes O(n2k+3 ), which clearly is not scalable beyond k ≥ 3. Moreover, for k = α · c (α < 1) the time complexity of exhaustive search becomes exponential in c. To address this problem, in the rest of this section with propose two efficient approximation algorithms with performance guarantees by exploiting the inherent structures of tree-connectivity. 4.3 Greedy Algorithm For any n ≥ 2, w : E(Kn ) → R>0 , connected (V, Einit ), and M+ ⊆ E(Kn ) define + ϕ : 2M → R≥0 E 7→ logTGn,w (E) (16) (17) Note that ϕ is essentially logTGn,w restricted to M+ . Therefore, Corollary 4.1 readily follows from Theorem 3.4. Corollary 4.1. ϕ is normalized, monotone and submodular. Consequently, k-ESP+ can be expressed as the problem of maximizing a normalized monotone submodular function subject to a cardinality constraint, i.e., maximize ϕ(E) subject to |E| = k. E⊆M+ (18) Maximizing an arbitrary monotone submodular function subject to a cardinality constraint can be NPhard in general (see e.g., the Maximum Coverage problem [13]). Therefore it is reasonable to look for reliable approximation algorithms. In this section we study the greedy algorithm described in Algorithm 1. Theorem 4.2 guarantees that Algorithm 1 is a constant-factor approximation algorithm for k-ESP+ with a factor of (1 − 1/e) ≈ 0.63. Theorem 4.2 (Nemhauser et al. [23]). The greedy algorithm attains at least (1 − 1/e)f ⋆ , where f ⋆ is the maximum of any normalized monotone submodular function subject to a cardinality constraint.2 Remark 3. Recall that ϕ is normalized by log treen,w (Einit ), and therefore reflects the tree-connectivity gain achieved by adding k new edges to the original graph (V, Einit , w). In order to avoid any confusion, from now on we denote the optimum value of (18) by OPTϕ , and use OPT to refer to the maximum achievable tree-connectivity in k-ESP+ . Note that, OPTϕ = OPT − log treen,w (Einit ). (19) 2 A generalized version of Theorem 4.2 [17] states that after ℓ ≥ k steps, the greedy algorithm is guaranteed to achieve at least (1 − e−ℓ/k )fk⋆ , where fk⋆ is the maximum of f (A) subject to |A| = k. 8 Algorithm 1 Greedy Edge Selection 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: 27: 28: 29: 30: 31: 32: function GreedyESP(Linit , M+ , k) E←∅ L ← Linit C ←Cholesky(L) while |E| < k do e⋆uv ← BestEdge(M+ \ E, C) E ← E ∪ {e⋆ } auv ← eu − ev L ← L + w(e⋆uv )auv a⊤ uv p C ← CholeskyUpdate(C, w(e⋆uv )auv ) end while return E end function ⊲ Rank-one update function BestEdge(M, C) m←0 for all e ∈ M do we ← w(e) ∆e ← Reff(e, C) if we ∆e > m then e⋆ ← e m ← we ∆e end if end for return e⋆ end function ⊲ Maximum value ⊲ Parallelizable loop function Reff(euv , C) auv ← eu − ev // solve Cxuv = auv xuv ← ForwardSolver(C, auv ) ∆uv ← kxuv k2 return ∆uv end function ⊲ Effective Resistance ⊲ Lower Triangular Let Egreedy be the set of edges picked by Algorithm 1. Define ϕgreedy , ϕ(Egreedy ). Then, according to Theorem 4.2, ϕgreedy ≥ (1 − 1/e) OPTϕ and therefore, log treen,w (Egreedy ∪ Einit ) ≥ (1 − 1/e) OPT + 1/e log treen,w (Einit ). (20) Algorithm 1 starts with an empty set of edges, and in each round picks the edge that maximizes the weighted tree-connectivity of the graph, until the cardinality requirement is met. Hence now we need a procedure for finding the edge that maximizes the weighted tree-connectivity. An efficient strategy is to use −1 Lemma 3.1 and pick the edge with the highest effective resistance wuv ∆uv . To compute ∆uv = a⊤ auv , we uv L first compute the Cholesky factor of the reduced weighted Laplacian matrix of the current graph L = CC⊤ . Next, we note that ∆uv = kxuv k2 where xuv is the solution of the triangular system Cxuv = auv . xuv can be computed by forward substitution in O(n2 ) time. The time complexity of each round is dominated by the 9 O(n3 ) time required for computing the Cholesky factor C. In the ith round, Algorithm 1 has to compute c−i effective resistances where c = |M+ |. For k = α · c (α < 1), evaluating effective resistances takes O(c2 n2 ) time. If k = O(1), this time reduces to O(c n2 ). Also, note that upon computing the Cholesky factor once in each round, xuv ’s can be computed in parallel by solving Cxuv = auv for different values of auv (see line #16 in Algorithm 1). We can avoid the O(k n3 ) time spent on repetitive Cholesky factorization by factorizing Linit once, followed by k − 1 rank-one updates, each of which takes O(n2 ) time. Therefore, the total time complexity of Algorithm 1 for k = O(1) and k = α · c will be O(n3 + c n2 ) and O(n3 + c2 n2 ), respectively. In the worst case of M+ = E(Kn ), c = O(n2 ) and therefore we get O(n4 ) and O(n6 ), respectively, for k = O(1) and k = α · c. Finally, note that for sparse graphs this complexity drops significantly given a sufficiently good fill-reducing permutation for the reduced weighted graph Laplacian. 4.4 Convex Relaxation Now we take a different approach and design an efficient approximation algorithm for k-ESP+ by means of convex relaxation. We begin by assigning an auxiliary variable 0 ≤ πi ≤ 1 to each candidate edge ei ∈ M+ . Let π , [π1 π2 . . . πc ]⊤ be the stacked vector of auxiliary variables in which c = |M+ |. Let G = (V, Einit , w) be the given base graph. Define L(π; G, M+ ) , X Lei + ei ∈Einit X ej ∈M πj Lej = AW′A⊤ , (21) + where Lek is the corresponding reduced elementary weighted Laplacian, A is the reduced incidence matrix of (V, Einit ∪ M+ ), and W′ , diag(w′ (e1 ), . . . , w′ (es )) in which s , |Einit | + |M+ | and,  π w(e ) i i w′ (ei ) , w(ei ) ei ∈ M+ , ei ∈ / M+ . (22) Lemma 4.1. L(π) is positive definite iff (V, Einit ∪ M+ ) is connected. Note that every k-subset of M+ is optimal for k-ESP+ if (V, Einit ∪ M+ ) is not connected. Therefore, if we ignore this degenerate case, we can safely assume that L(π; G, M+ ) is positive definite. With a slight abuse of notation, from now on we drop the parameters from L(π; G, M+ ) and use L(π) whenever G and M+ are clear from the context. Now consider the following optimization problem over π. maximize log det L(π) π subject to kπk0 = k, (P1 ) 0 ≤ πi ≤ 1, ∀i ∈ [c]. P1 is equivalent to our former definition of k-ESP+ . The auxiliary variables act as selectors: the ith candidate edge is selected iff πi = 1. The objective function rewards strong weighted tree-connectivity. The combinatorial difficulty of ESP here is embodied in the non-convex ℓ0 -norm constraint. It is easy to see that 10 at the optimal solution, auxiliary variables take binary values. Therefore P1 can also be expressed as maximize log det L(π) π subject to (P′1 ) kπk1 = k, πi ∈ {0, 1}, ∀i ∈ [c]. A natural choice for relaxing P′1 is to replace πi ∈ {0, 1} with 0 ≤ πi ≤ 1, i.e., maximize log det L(π) π subject to (P2 ) kπk1 = k, 0 ≤ πi ≤ 1, ∀i ∈ [c]. The feasible set of P2 contains that of P1 (or, equivalently, P′1 ), and therefore the optimum value of P2 is an upper bound for the optimum of P1 (or, equivalently, P′1 ). Note that the ℓ1 -norm here is identical to Pc i=1 πi . P2 is a convex optimization problem since the objective function (tree-connectivity) is concave and the constraints are linear and affine in π. In fact, P2 is an instance of the MAXDET problem [29] subject to additional affine constraints on π. It is worth noting that P2 can be reached also by relaxing the non-convex Pc ℓ0 -norm constraint in P1 by a convex ℓ1 -norm constraint (essentially i=1 πi = k). Furthermore, P2 is also closely related to a ℓ1 -regularalized instance of MAXDET, maximize π subject to log det L(π) − λ kπk1 (P3 ) 0 ≤ πi ≤ 1, ∀i ∈ [c]. This problem is a penalized form of P2 ; these two problems are equivalent for some positive value of λ. Problem P3 is also a convex optimization problem for non-negative λ. The ℓ1 -norm in P3 encourages sparser π, while the log-determinant rewards stronger tree-connectivity. The penalty coefficient λ is a parameter that specifies the desired degree of sparsity, i.e., larger λ yields a sparser vector of selectors π. Problem P2 (and P3 ) can be solved efficiently using interior-point methods [5]. After finding a globally optimal solution π ⋆ for the relaxed problem P2 , we ultimately need to map it into a feasible π for P1 , i.e., picking k edges from the candidate set M+ . First note that if π⋆ ∈ {0, 1}c, it means that π⋆ is already an optimal solution for k-ESP+ and P1 . However, in the more likely case of π ⋆ containing fractional values, we need a rounding procedure to set k auxiliary variables to one and others to zero. The most intuitive choice is to pick the k edges with the largest πi⋆ ’s. Another (approximate) rounding strategy (and a justification for picking the k largest πi⋆ ) emerges from interpreting πi as the probability of selecting the ith candidate edge. Theorem 4.3 provides a new interesting way of interpreting the convex relaxation of P1 by P2 . Theorem 4.3. Define E• , Einit ∪ M+ and G• , (V, E• , w). Let π • = [π1 . . . πs ]⊤ ∈ (0, 1]s such that s , |Einit | + |M+ | and πi = 1 if ei ∈ Einit . Then we have   EH∼G(G• ,π• ) tw (H) = det L(π),   X EH∼G(G• ,π• ) |E(H)| − |Einit | = πi = kπk1 . ei ∈M+ 11 (23) (24) Note that (23) and (24) appear in the objective function and the constraints of P2 , respectively. Thus P2 can be rewritten as maximize π subject to   EH∼G(G• ,π• ) tw (H)   EH∼G(G• ,π• ) |E(H)| = k + |Einit |, (P′2 ) 0 ≤ πi ≤ 1, ∀i ∈ [s]. This offers a new narrative: the objective in P2 is to find the optimal probabilities π ⋆ for sampling edges from M+ such that the weighted number of spanning trees is maximized in expectation, while the expected number of newly selected edges is equal to k. In other words, P2 can be seen as a convex relaxation of P1 at the expense of maximizing the objective and satisfying the constraint, both in expectation. This new interpretation motivates an approximate randomized rounding procedure that picks ei ∈ M+ with probability πi⋆ . According to Theorem 4.3, this randomized rounding scheme, in average, attains det L(π ⋆ ) by picking k new edges in average. Theorem 4.4. For any 0 < ǫ < 1 and δ > 0,    P |E ⋆ | < (1 − ǫ)k < exp −ǫ2 k/2 ,    P |E ⋆ | > (1 + δ)k < exp −δ 2 k/3 , (25) (26) where E ⋆ is the set of selected edges by the randomized rounding scheme defined above. Theorem 4.4 ensures that the probability of the events in which the aforementioned randomized rounding strategy picks too many/few edges (compared to k) decay exponentially. Note that this new narrative offers another intuitive justification for deterministically picking the k edges with largest πi⋆ ’s. Finally, we believe that Theorems 4.3 and 4.4 can potentially be used as building blocks to design new randomized rounding schemes. 4.5 Certifying Near-Optimality The proposed approximation algorithms also provide a posteriori lower and upper bounds for the maximum achievable tree-connectivity in ESP. Let Egreedy , Ecvx be the solutions returned by the greedy ⋆ and convex3 approximation algorithms, respectively. Let τcvx be the optimum value of P2 and define τinit , log treen,w (Einit ), τcvx , log treen,w (Ecvx ∪ Einit ) and τgreedy , log treen,w (Egreedy ∪ Einit ). Corollary 4.5. max −1 where ζ , (1 − 1/e) n o o n ⋆ τgreedy , τcvx ≤ OPT ≤ min ζτgreedy + (1 − ζ)τinit , τcvx (27) ≈ 1.58.4 Corollary 4.5 can be used as a tool to asses the quality of any suboptimal design. Let A be anoarbitrary n + ⋆ k-subset of M and τA = log treen,w (A ∪ Einit ). Define U , min ζτgreedy + (1 − ζ)τinit , τcvx . U can be computed by running the proposed greedy and convex approximation algorithms. From Corollary 4.5 it readily follows that OPT − τA ≤ U − τA and OPT/τA ≤ U/τA . Therefore, although we may not have 3 Picking the k edges with the largest πi⋆ ’s from the solution of P2 . recall that the leftmost term in (27) is bounded from below by the expression given in (20). 4 Furthermore, 12 53 52 51 110 90 80 49 150 70 150 160 165 170 Number of Edges 175 180 (a) Varying |E| for k = 5 and |V | = 20 τgreedy τcvx ⋆ τcvx 115 100 50 155 120 τgreedy τcvx ⋆ τcvx Tree−Connectivity 120 OPT τgreedy τcvx ⋆ τcvx Tree−Connectivity Tree−Connectivity 54 110 105 100 95 200 250 Number of Edges 300 (b) Varying |E| for k = 5 and |V | = 50 90 0 20 40 60 80 100 k (c) Varying k for |V | = 50 and |E| = 200 ⋆ . Figure 1: ESP on randomly generated graphs. Recall that according to Corollary 4.5, τgreedy ≤ OPT ≤ τcvx direct access to OPT, we can still certify the near-optimality of any design such as A whose δ , U − τA is sufficiently small. 4.6 Numerical Results We implemented Algorithm 1 in MATLAB. Problem P2 is modelled using CVX [12, 11] and YALMIP [18], and solved using SDPT3 [28]. Figure 1 illustrates the performance of our approximate solutions to k-ESP+ in randomly generated graphs. The search space in these experiments is M+ = E(Kn ) \ Einit . Figures 1a and 1b show tree-connectivity as a function of number of randomly generated edges for a fixed k = 5 and, respectively, |V | = 20 and |V | = 50. Our results indicate that both algorithms exhibit remarkable performances for k = 5. Note that computing OPT by exhaustive search is only feasible in small instances such as Figure 1a. Nevertheless, computing the exact OPT is not crucial for evaluating our approximate ⋆ algorithms, as it is tightly bounded in [τgreedy , τcvx ] as predicted by Corollary 4.5 (i.e., between each black · and green ×). Figure 1c shows the results obtained for varying k. The optimality gap for τcvx gradually grows as the planning horizon k increases. Our greedy algorithm, however, still yields a near-optimal approximation. 5 5.1 Beyond k-ESP+ Matroid Constraints Recall that ϕ is monotone. Therefore, except the degenerate case of (V, Einit ∪ M+ ) not being connected, replacing the cardinality constraint |E| = k in k-ESP+ with an inequality constraint |E| ≤ k does not affect the set of optimal solutions. Consider the uniform matroid [24] defined as (M+ , IU ) where  IU , A ⊆ M+ : |A| ≤ k . The inequality cardinality constraint can be expressed as E ∈ IU . + + Definition 5.1 (Partition Matroid). Let M+ 1 , . . . , Mℓ be a partition for M . Assign an integer (budget) 13 + 0 ≤ ki ≤ |M+ i | to each Mi . Define o n | ≤ k for i ∈ [ℓ] . IP , A ⊆ M+ : |A ∩ M+ i i The pair (M+ , IP ) is called a partition matroid. Now let us consider ESP under a partition matroid constraint; i.e., maximize ϕ(E) (28) subject to E ∈ IP . Note that k-ESP+ is a special case of this problem with ℓ = 1 and k1 = k. Now, by choosing different partitions for M+ and different budgets ki we can model a wide variety of graph synthesis problems. For example consider the following extension of k-ESP+ , maximize E⊆M+ ,|E|≤k subject to ϕ(E) (29) deg(v) ≤ d. n o + Define M+ , e ∈ M : v ∈ e . Now note that the constraints in (29) can be expressed as a partition v + matroid with two blocks: (i) Mv with a budget of k1 = d, and (ii) M+ \ M+ v with a budget of k2 = k − d. 5.1.1 Greedy Algorithm Theorem 5.1 (Fisher et al. [7]). The greedy algorithm attains at least (1/2)f ⋆ , where f ⋆ is the maximum of any normalized monotone submodular function subject to a matroid constraint. According to Theorem 5.1, a slightly modified version of Algorithm 1, that abides by the matroid constraint while greedily choosing the next best edge, yields a 12 -approximation [7, 17]. 5.1.2 Convex Relaxation The proposed convex relaxation of k-ESP+ can be modified to handle a partition matroid constraint. First note that (28) can be expressed as maximize π subject to log det L(π) X πi ≤ kj , ∀j ∈ [ℓ] ei ∈M+ j πi ∈ {0, 1}, ∀i ∈ [c]. 14 (P4 ) Relaxing the binary constraints on πi ’s yields maximize log det L(π) X πi ≤ kj , ∀j ∈ [ℓ] π subject to ei ∈M+ j (P5 ) 0 ≤ πi ≤ 1, ∀i ∈ [c]. P5 is a convex optimization problem and, as before, can be solved efficiently using interior-point methods. A simple rounding strategy for the solution of P5 is to pick the edges in M+ i that are associated to the ki largest πj⋆ ’s (for i ∈ [ℓ]). Moreover, the bounds in (27) (with ζ = 2) and Theorem 4.3 can also be readily generalized to handle partition matroid constraints. In particular, the optimum value of P5 gives an upper bound for the optimum value of P4 . Also, similar to Theorem 4.3, P5 can be interpreted as maximizing the expected value of the weighted number of spanning trees such that the expected number of new edges sampled from M+ i is at most ki , for i ∈ [ℓ]. 5.2 Dual of k-ESP+ The dual of k-ESP+ aims to identify and select the minimal set of new edges from a candidate set M+ such that the resulting tree-connectivity gain is at least 0 ≤ δ ≤ ϕ(M+ ) for some given δ; i.e., minimize |E| subject to ϕ(E) ≥ δ. E⊆M+ 5.2.1 (30) Greedy Algorithm The greedy algorithm for approximating the solution of (30) is outlined in Algorithm 2. The only difference between Algorithm 1 and Algorithm 2 is that the latter terminates when the δ-bound is achieved (or, alternatively, when there are no more edges left in M+ , which indicates an empty feasible set). Wolsey [31] proves several upper bounds for the ratio between the objective value achieved by the greedy algorithm and the optimum value of the following class of problems, minimize |A| subject to φ(A) ≥ φ0 , A⊆W (31) in which φ : 2W → R is an arbitrary monotone submodular function and φ0 ≤ φ(W). Note that our problem (30) is special case of (31), and therefore (some of) the bounds proved by Wolsey [31, Theorem 1] also hold for Algorithm 2. Theorem 5.2 (Wolsey [31]). Let kOPT and kgreedy be the global minimum of (30) and the objective value achieved by Algorithm 2, respectively. Also, let Ẽgreedy be the set formed by Algorithm 2 one step before termination. Then kgreedy ≤ γ kOPT in which γ , 1 + log   δ . δ − ϕ(Ẽgreedy ) 15 (32) Algorithm 2 Greedy Dual Edge Selection 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: function GreedyDualESP(Linit , M+ , δ) E←∅ L ← Linit C ←Cholesky(L) while (log det L < δ) ∧ (E 6= M+ ) do e⋆uv ← BestEdge(M+ \ E, C) E ← E ∪ {e⋆ } auv ← eu − ev L ← L + w(e⋆uv )auv a⊤ uv p C ← CholeskyUpdate(C, w(e⋆uv )auv ) end while return E end function ⊲ Rank-one update The upper bound given above and some of the other bounds in [31] are a posteriori in the sense that they can be computed only after running the greedy algorithm. 5.2.2 Convex Relaxation Let τinit , log det L(0). The dual problem can be expressed as minimize π c X πi i=1 subject to (D1 ) log det L(π) ≥ δ + τinit , πi ∈ {0, 1}, ∀i ∈ [c]. The combinatorial difficulty of the dual formulation of ESP is manifested in the binary constraints of D1 . Relaxing these constraints into 0 ≤ πi ≤ 1 yields the following convex optimization problem, minimize π c X πi i=1 subject to (D2 ) log det L(π) ≥ δ + τinit , 0 ≤ πi ≤ 1, ∀i ∈ [c]. D2 can be solved efficiently using interior-point methods. Let π ⋆ be the minimizer of D2 . ⋆ c ⋆ Pc i=1 πi⋆ is a lower bound for the optimum value of the dual ESP D1 . If π ∈ {0, 1} , π is also a globally optimal solution for D1 . Otherwise we need a rounding scheme to map π ⋆ into a feasible (suboptimal) solution for D1 . A simple deterministic rounding strategy is the following. - Step 1. Sort the edges in M+ according to π⋆ in descending order. - Step 2. Pick edges from the sorted list until log det L(π) ≥ δ + τinit . 16 Theorem 4.3 allows us to interpret D2 as finding the optimal sampling probabilities π⋆ that minimizes the expected number of new edges such that the expected weighted number of spanning trees is at least exp(δ + τinit ); i.e.,   minimize EH∼G(G• ,π• ) |E(H)| , π   (D′2 ) subject to EH∼G(G• ,π• ) tw (H) ≥ exp(δ + τinit ), 0 ≤ πi ≤ 1, ∀i ∈ [s], in which G(G• , π • ) is defined in Theorem 4.3. This narrative suggests a randomized rounding scheme in which ei ∈ M+ is selected with probability πi⋆ . The expected number of selected edges by this procedure is Pc ⋆ i=1 πi . 5.2.3 Certifying Near-Optimality Corollary 5.3. Define ζ ∗ , 1/γ where γ is the approximation factor given by Theorem 5.2. Let kcvx be the number of new edges selected by the deterministic rounding procedure described above.  ∗ max ζ kgreedy , c lX i=1 πi⋆ m o n ≤ kOPT ≤ min kgreedy , kcvx . (33) As we did before for k-ESP+ , the lower bound provided by Corollary 5.3 can be used to construct an upper bound n for the gap lbetween kmo OPT and any (feasible) suboptimal design with an objective value of kA . Pc ⋆ ∗ . L can be computed by running Algorithm 2 and solving the convex Let L , max ζ kgreedy , i=1 πi optimization problem D2 . Consequently, kA − kOPT ≤ kA − L and kA /kOPT ≤ kA /L. 6 Conclusion We studied the problem of designing near-t-optimal graphs under several types of constraints and formulations. Several new structures were revealed and exploited to design efficient approximation algorithms. In particular, we proved that the weighted number of spanning trees in connected graphs can be posed as a monotone log-submodular function of the edge set. Our approximation algorithms can find near-optimal solutions with performance guarantees. They also provide a posteriori near-optimality certificates for arbitrary designs. Our results can be readily applied to a wide verity of applications involving graph synthesis and graph sparsification scenarios. References [1] Rosemary A Bailey and Peter J Cameron. Combinatorics of optimal designs. Surveys in Combinatorics, 365:19–73, 2009. 3 [2] Prabir Barooah and Joao P Hespanha. Estimation on graphs from relative measurements. Control Systems, IEEE, 27(4):57–74, 2007. 6 17 [3] Douglas Bauer, Francis T Boesch, Charles Suffel, and R Van Slyke. On the validity of a reduction of reliable network design to a graph extremal problem. Circuits and Systems, IEEE Transactions on, 34 (12):1579–1581, 1987. 3 [4] Francis T Boesch, Appajosyula Satyanarayana, and Charles L Suffel. A survey of some network reliability analysis and synthesis results. Networks, 54(2):99–107, 2009. 3, 7 [5] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 11 [6] Ching-Shui Cheng. Maximizing the total number of spanning trees in a graph: two related problems in graph theory and optimum design theory. Journal of Combinatorial Theory, Series B, 31(2):240–248, 1981. 3 [7] Marshall L Fisher, George L Nemhauser, and Laurence A Wolsey. An analysis of approximations for maximizing submodular set functionsII. Springer, 1978. 14 [8] N Gaffke. D-optimal block designs with at most six varieties. Journal of Statistical Planning and Inference, 6(2):183–200, 1982. 3 [9] Arpita Ghosh, Stephen Boyd, and Amin Saberi. Minimizing effective resistance of a graph. SIAM review, 50(1):37–66, 2008. 6 [10] Chris Godsil and Gordon Royle. Algebraic graph theory. Graduate Texts in Mathematics Series. Springer London, Limited, 2001. ISBN 9780387952413. 3, 4 [11] Michael Grant and Stephen Boyd. Graph implementations for nonsmooth convex programs. In V. Blondel, S. Boyd, and H. Kimura, editors, Recent Advances in Learning and Control, Lecture Notes in Control and Information Sciences, pages 95–110. Springer-Verlag Limited, 2008. http://stanford.edu/~boyd/graph_dcp.html. 13 [12] Michael Grant and Stephen Boyd. CVX: Matlab software for disciplined convex programming, version 2.1. http://cvxr.com/cvx, March 2014. 13 [13] Dorit S Hochbaum. Approximation algorithms for NP-hard problems. PWS Publishing Co., 1996. 8 [14] Siddharth Joshi and Stephen Boyd. Sensor selection via convex optimization. Signal Processing, IEEE Transactions on, 57(2):451–462, 2009. 23 [15] Alexander K Kelmans. On graphs with the maximum number of spanning trees. Random Structures & Algorithms, 9(1-2):177–192, 1996. 3 [16] Alexander K Kelmans and BN Kimelfeld. Multiplicative submodularity of a matrix’s principal minor as a function of the set of its rows and some combinatorial applications. Discrete Mathematics, 44(1): 113–116, 1983. 3 [17] Andreas Krause and Daniel Golovin. Submodular function maximization. Tractability: Practical Approaches to Hard Problems, 3:19, 2012. 8, 14 18 [18] Johan Löfberg. Yalmip : A toolbox for modeling and optimization in MATLAB. In Proceedings of the CACSD Conference, Taipei, Taiwan, 2004. URL http://users.isy.liu.se/johanl/yalmip. 13 [19] László Lovász. Random walks on graphs: A survey. Combinatorics, Paul Erdős is eighty, 2(1):1–46, 1993. 6 [20] Mehran Mesbahi and Magnus Egerstedt. Graph theoretic methods in multiagent networks. Princeton University Press, 2010. 5 [21] Carl D Meyer. Matrix analysis and applied linear algebra. SIAM, 2000. 20 [22] Wendy Myrvold. Reliable network synthesis: Some recent developments. In Proceedings of International Conference on Graph Theory, Combinatorics, Algorithms, and Applications, 1996. 3 [23] George L Nemhauser, Laurence A Wolsey, and Marshall L Fisher. An analysis of approximations for maximizing submodular set functions - I. Mathematical Programming, 14(1):265–294, 1978. 8 [24] James G Oxley. Matroid theory, volume 3. Oxford university press, 2006. 13 [25] Louis Petingi and Jose Rodriguez. A new technique for the characterization of graphs with a maximum number of spanning trees. Discrete mathematics, 244(1):351–373, 2002. 3 [26] Friedrich Pukelsheim. Optimal design of experiments, volume 50. SIAM, 1993. 3 [27] DR Shier. Maximizing the number of spanning trees in a graph with n nodes and m edges. Journal Research National Bureau of Standards, Section B, 78:193–196, 1974. 3 [28] Reha H Tütüncü, Kim C Toh, and Michael J Todd. Solving semidefinite-quadratic-linear programs using sdpt3. Mathematical programming, 95(2):189–217, 2003. 13 [29] Lieven Vandenberghe, Stephen Boyd, and Shao-Po Wu. Determinant maximization with linear matrix inequality constraints. SIAM journal on matrix analysis and applications, 19(2):499–533, 1998. 11 [30] Guy Weichenberg, Vincent WS Chan, and Muriel Médard. High-reliability topological architectures for networks under stress. Selected Areas in Communications, IEEE Journal on, 22(9):1830–1845, 2004. 3 [31] Laurence A Wolsey. An analysis of the greedy algorithm for the submodular set covering problem. Combinatorica, 2(4):385–393, 1982. 15, 16 19 A Proofs Lemma A.1. For any M ∈ Sn>0 and N ∈ Sn>0 , M  N iff N−1  M−1 . Proof. Due to symmetry it suffices to prove that M  N ⇒ N−1  M−1 . Multiplying both sides of M  N 1 1 1 1 1 by N− 2 from left and right results in N− 2 MN− 2 − I  0. Therefore the eigenvalues of N− 2 MN− 2 , which 1 1 1 1 are the same as the eigenvalues of M 2 N−1 M 2 ,5 are at least 1. Therefore M 2 N−1 M 2 − I  0. Multiplying 1 both sides by M− 2 from left and right proves the lemma. Lemma A.2 (Matrix Determinant Lemma). For any non-singular M ∈ Rn×n and c, d ∈ Rn , det(M + cd⊤ ) = (1 + d⊤ M−1 c) det M. (34) Proof. See e.g., [21]. w w Lemma A.3. Let G1 be a spanning subgraph of G2 . For any w : E(K) → R≥0 , Lw G2  LG1 in which LG is the reduced weighted Laplacian matrix of G when its edges are weighted by w. Proof. From the definition of the reduced weighted Laplacian matrix we have, w Lw G2 − LG1 = X wuv auv a⊤ uv  0. (35) {u,v}∈E(G2 )\E(G1 ) Proof of Theorem 3.1. Define the following indicator function,  1 T ∈ T , G 1TG (T ) , 0 T ∈ / TG , (36) in which TG denotes the set of spanning trees of G. Now note that, hX i   EG∼G(G◦ ,p) tw (G) = EG∼G(G◦ ,p) 1TG (T )Vw (T ) (37) T ∈TG◦ = X EG∼G(G◦ ,p) T ∈TG◦ = h i 1TG (T )Vw (T ) (38) i X h P T ∈ TG Vw (T ) (39) Vp (T )Vw (T ) (40) T ∈TG◦ = X T ∈TG◦ = X Vwp (T ) (41) T ∈TG◦ = twp (G◦ ). 5 Recall that MN and NM have the same spectrum. 20 (42) Here we have used the fact the P[T ∈ TG ] is equal to the probability of existence of every edge of T in G, which is equal to Vp (T ). Proof of Lemma 3.1. Note that LG+ = LG + wuv auv a⊤ uv . Taking the determinant, applying Lemma A.2 and taking the log concludes the proof. Proof of Lemma 3.2. The proof is similar to the proof of Lemma 3.1. Proof of Theorem 3.3. First recall that Vw (T ) is positive for any T by definition. 1. Normalized: treen,w (∅) = 0 by definition. 2. Monotone: Let G , (V, E ∪ {e}). Denote by TGe the set of spanning trees of G that contain e. treen,w (E ∪ {e}) = Vw (T ) = X Vw (T ) + treen,w (E) ≥ treen,w (E). Vw (T ) + T ∈TGe T ∈ TG = X X X Vw (T ) (43) T∈ / TGe (44) T ∈TGe 3. Supermodular: treen,w is supermodular iff for all E1 ⊆ E2 ⊆ E(Kn ) and all e ∈ E(Kn ) \ E2 , treen,w (E2 ∪ {e}) − treen,w (E2 ) ≥ treen,w (E1 ∪ {e}) − treen,w (E1 ). (45) Define G1 , (V, E1 ) and G2 , (V, E2 ). As we showed in (44), treen,w (E1 ∪ {e}) − treen,w (E1 ) = X Vw (T ), X Vw (T ). T ∈T treen,w (E2 ∪ {e}) − treen,w (E2 ) = T ∈T Therefore we need to show that TGe2 . P T ∈TGe 2 Vw (T ) ≥ P T ∈TGe 1 (46) e G1 (47) e G2 Vw (T ). This inequality holds since TGe1 ⊆ Proof of Theorem 3.4. 1. Normalized: By definition logTGn,w (∅) = log treen,w (Einit ) − log treen,w (Einit ) = 0. 2. Monotone: We need to show that logTGn,w (E ∪ {e}) ≥ logTGn,w (E). This is equivalent to showing that, log treen,w (Einit ∪ E ∪ {e}) ≥ log treen,w (Einit ∪ E). (48) Now note that (V, Einit ∪ E) is connected since (V, Einit ) was assumed to be connected. Therefore we can apply Lemma 3.1 on the LHS of (48); i.e., log treen,w (Einit ∪ E ∪ {e}) = log treen,w (Einit ∪ E) + log(1 + we ∆e ). 21 (49) Therefore it sufficies to show that log(1 + we ∆e ) is non-negative. Since (V, Einit ) is connected, L is −1 positive definite. Consequently we ∆e = we a⊤ ae > 0 and hence log(1 + we ∆e ) > 0. e L 3. Submodular: logTGn,w is submodular iff for all E1 ⊆ E2 ⊆ E(Kn ) and all e ∈ E(Kn ) \ E2 , logTGn,w (E1 ∪ {e}) − logTGn,w (E1 ) ≥ logTGn,w (E2 ∪ {e}) − logTGn,w (E2 ). (50) After canceling log treen,w (Einit ) we need to show that, log treen,w (E1 ∪Einit ∪{e})−log treen,w (E1 ∪Einit ) ≥ log treen,w (E2 ∪Einit ∪{e})−log treen,w (E2 ∪Einit ). (51) / Einit . To shorten If e ∈ Einit , both sides of (51) become zero. Hence we can safely assume that e ∈ ∗ our notation let us define Ei , Ei ∪ Einit for i = 1, 2. Therefore (51) can be rewritten as, log treen,w (E1∗ ∪ {e}) − log treen,w (E1∗ ) ≥ log treen,w (E2∗ ∪ {e}) − log treen,w (E2∗ ). (52) Recall that by assumption (V, Einit ) is connected. Thus (V, Ei∗ ) is connected for i = 1, 2, and we can apply Lemma 3.1 on both sides of (52). After doing so we have to show that G2 1 log(1 + we ∆G e ) ≥ log(1 + we ∆e ) (53) G2 1 where Gi , (V, Ei ∪ Einit , w) for i = 1, 2. It is easy to see that (53) holds iff ∆G e ≥ ∆e . Now note that −1 −1 ⊤ 2 ∆eG1 − ∆G e = ae (LG1 − LG2 ) ae ≥ 0 (54) −1 since LG2  LG1 (G1 is a spanning subgraph of G2 ), and therefore according to Lemma A.1 L−1 G1  LG2 . Proof of Theorem 4.3. First note that (23) directly follows from Theorem 3.1 since L(π) is the reduced weighted Laplacian matrix of G• after scaling its edge weights by the sampling probabilities π1 , . . . , πs . To prove (24) consider the following indicator function,  1 e ∈ E(H), 1E(H) (e) = 0 e ∈ / E(H). 22 (55) Now note that 1E(H) (ei ) ∼ Bern(πi ) for i = 1, . . . , s. Therefore, s hX i   EH∼G(G• ,π • ) |E(H)| = EH∼G(G• ,π• ) 1E(H) (ei ) (56) i=1 = = s X i=1 s X EH∼G(G• ,π• ) h i 1E(H) (ei ) (57) πi (58) i=1 = X πi + X 1 (59) ej ∈Einit ei ∈M+ = kπk1 + |Einit |. (60) Proof of Theorem 4.4. This theoreom is a direct application of Chernoff bounds for Poisson trials of independently sampling edges from M+ with probabilities specified by π⋆ . Generalizing Theorems 3.1 and 4.3 The following theorem generalizes Theorem 3.1 (and, consequently, Theorem 4.3). Theorem A.1 provides a similar interpretation for the convex relaxation approach designed by Joshi and Boyd [14] for the sensor selection problem with linear measurement models. n Theorem A.1. Let {(yi , zi )}m i=1 be a collection of m pairs of vectors in R such that m ≥ n. Furthermore, let s1 , . . . , sm be a collection of m independent random variables such that si ∼ Bern(pi ) for some pi ∈ [0, 1]. Then we have, E si ∼Bern(pi ) ∀i∈[m] Proof. Let Sn , have [m] n  h det m X si yi z⊤ i i=1 ! i = det pi yi z⊤ i i=1 ! . (61) be the set of all n-subsets of [m]. According to the Cauchy-Binet (C-B) formula we E si ∼Bern(pi ) ∀i∈[m] h det m X i=1 si yi z⊤ i ! i C-B = E si ∼Bern(pi ) ∀i∈[m] = X Q∈Sn Now note that |Q| = n and rank(yi z⊤ i ) = 1. Therefore det Thus for every Q ∈ Sn , det m X X i∈Q si yi z⊤ i ! si yi z⊤ i Q∈Sn X i∈Q ! h X si yi z⊤ i ! h X E si ∼Bern(pi ) ∀i∈[m] det det ⊤ i∈Q si yi zi P i∈Q  i (62) i . (63) is non-zero iff si = 1 for all i ∈ Q.   Q ⊤ d , det P with probability pQ , i∈Q pi , Q i∈Q yi zi = 0 with probability 1 − pQ . 23 (64) Taking the expectation yields E si ∼Bern(pi ) ∀i∈[m] h det X si yi z⊤ i i∈Q ! i = pQ dQ . (65) Replacing (65) in (63) results in E si ∼Bern(pi ) ∀i∈[m] h det m X i=1 si yi z⊤ i ! i = X Q∈Sn pQ dQ = X Noting that the RHS in (66) is the Cauchy-Binet expansion of det 24 det Q∈S X i∈Q Pm i=1 pi yi z⊤ i ! .  concludes the proof. pi yi z⊤ i (66)
8cs.DS
arXiv:1604.07736v1 [math.GR] 26 Apr 2016 BOUNDARY ACTION OF AUTOMATON GROUPS WITHOUT SINGULAR POINTS AND WANG TILINGS DANIELE D’ANGELI, THIBAULT GODIN, INES KLIMANN, MATTHIEU PICANTIN, AND EMANUELE RODARO Abstract. We study automaton groups without singular points, that is, points in the boundary for which the map that associates to each point its stabilizer, is not continuous. This is motivated by the problem of finding examples of infinite bireversible automaton groups with all trivial stabilizers in the boundary, raised by Grigorchuk and Savchuk. We show that, in general, the set of singular points has measure zero. Then we focus our attention on several classes of automata. We characterize those contracting automata generating groups without singular points, and apply this characterization to the Basilica group. We prove that potential examples of reversible automata generating infinite groups without singular points are necessarily bireversible. Then we provide some necessary conditions for such examples to exist, and study some dynamical properties of their Schreier graphs in the boundary. Finally we relate some of those automata with aperiodic tilings of the discrete plane via Wang tilings. This has a series of consequences from the algorithmic and dynamical points of view, and is related to a problem of Gromov regarding the searching for examples of CAT(0) complexes whose fundamental groups are not hyperbolic and contain no subgroup isomorphic to Z2 . Contents 1. Introduction 2. Preliminaries 2.1. Mealy automata 2.2. Automaton groups 3. The considered classes 3.1. Contracting automata 3.2. (Bi)reversible automata 3.3. Automata with sink 4. Topological properties of the action on the boundary 4.1. Action on the boundary 4.2. Singular points 4.3. The contracting case 4.4. The bireversible case 5. Commuting pairs and dynamics on the boundary 5.1. Commuting pairs 5.2. The reversible-invertible case 6. Dynamics and Wang tilings 6.1. Wang tiles vs cross-diagrams 6.2. Commuting pairs on a restricted stateset 6.3. Aperiodic tilings and singular points 7. Some open problems References 2 3 3 5 6 6 7 9 10 10 11 13 17 18 19 20 26 26 29 34 38 38 Date: March 17, 2018. Key words and phrases. automaton groups, singular points, critical points, Schreier graphs, boundary continuity, Wang tilings, commuting pairs, helix graphs. 1 2 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO 1. Introduction The motivation comes from the study of the dynamical system (G, ∂T, µ) given by the measure µ preserving action of a group G on the boundary ∂T of a rooted tree T . By considering one orbit of this action (i.e. a Schreier graph), one may ask if it is possible to recover the information about the original dynamics in terms of the information contained in a typical orbit. This problem may be rephrased as follows: what conditions have to be imposed on the dynamical system (G, ∂T, µ) in order to guarantee that, for a typical point ξ ∈ ∂T , this dynamical system is isomorphic to the system (G, StabhM i (ξ), ν), for some measure ν concentrated on the closure StabhM i (ξ)? (Problem 8.2 in [15]). This problem has been studied by Y. Vorobets in the special case of the Grigorchuk group [36]. He showed that for this group it is possible to reconstruct the action of the general dynamical system on the boundary starting from the study of one orbit. His method uses the study of the map St that associates to any point in the boundary of the tree its stabilizer subgroup in the automaton group. Motivated by these ideas, we examine the dynamical and algorithmic implications of the continuity of the map St in the context of automaton groups, and how some combinatorial properties of the generating automaton reflect into the continuity of this map. In particular, we focus our attention on several classes of Mealy automata: the contracting case (see [25]), the reversible case, the bireversible case, and finally the case of automata with a sink-state which is accessible from every state (henceforth denoted by Sa ). We first show that, in general, the measure of the set of the points in which St is not continuous (henceforth called singular points) is zero. In the bireversible case singular points are exactly points with non-trivial stabilizers. This reproves the well known fact that bireversible automata give rise to essentially free actions on the boundary [32, Corollary 2.10]. Driven by these facts and the question raised by Grigorchuk and Savchuk in [16] regarding the existence of singular points for the action of a bireversible automaton generating an infinite group, we generalize the previous open problem into the study of examples of automaton groups without singular points. In the case of contracting groups we provide a characterization for such automata in terms of languages recognized by Büchi automata that we call stable automata. In the examples that we present, we show that the situations may be very different. For instance, in contrast with the Hanoi Towers group case, we show that the Basilica group has no singular points. In both the class of reversible invertible automata and the class Sa , by using the notion of helix graph, we reduce this problem to the existence of certain pairs of words, called commuting pairs, that is, two words, one on the stateset the other one on the alphabet, that commute with respect to the induced actions. Using this fact we prove a series of results. For instance, it turns out that the existence of singular points is always guaranteed for the reversible invertible automata that are not bireversible. This shows that, in the class of reversible invertible automata, the core of the problem of finding examples of group automata without singular points is reduced to the class of bireversible automata. We present some necessary conditions for such examples to exist. For instance, the generated group is necessarily fully positive, that is, it is defined by relators that do not contain negative occurrences of the generators. Furthermore, we prove that if a bireversible automaton generates an infinite non-torsion group, then having all stabilizers in the boundary that are BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 3 torsion groups (like in the situation of not having singular elements) is equivalent to have in the dual automaton all Schreier graphs in the boundary which are either finite, or acyclic multigraphs (just considering the edges without their inverse). The study of commuting pairs also leads to a connection with periodic tessellations of the discrete plane using Wang tilings, and it is also related to the so-called Gromov’s problem (the reader is referred to the paper [21] for more details). This connection has been pointed out to us by I. Bondarenko [5]. Using the helix graph one can easily show that any automaton group has a commuting pair, whence the associated tileset has always a periodic tiling. However this commuting pair may involve a trivial word. For instance in the class Sa there is always a trivial commuting pair involving the sink-state. This fact leads to the notions of non-elementary commuting pair and reduced tileset of an automaton group. Using a result by [23] we first show that the problem of finding non-elementary commuting pairs is undecidable. Further, the notion of non-elementary commuting pair is strictly related to the existence of periodic singular points in the boundary. From this connection, we start a study of the relationship between non-periodic tessellations of the discrete plane, and algebraic and dynamical properties of the associated automaton group. Indeed, we first provide conditions for the associated reduced tileset to tile the discrete plane. Then, we pinpoint the algebraic and dynamical properties that an automaton group from Sa has to possess so that the associated reduced tileset generates just aperiodic tilings. Finally, we characterize the existence of aperiodic tilings à la Kari-Papasoglu with some properties of the group generated by an automaton and its set of singular points. 2. Preliminaries 2.1. Mealy automata. We first start with some vocabulary on words, then introduce our main tool — Mealy automata. Let Q be a finite set, as usual, Qn , Q≤n , Q<n , Q≥n , Q∗ , and Qω denote respectively the set of words of length n, of length less than or equal to n, of length less than n, of length greater than or equal to n, of finite length, and the set of right-infinite words on Q . For two words u, v ∈ Q∗ with u = vv ′ (u = v ′ v) for some v ′ ∈ Q∗ , we say that v is a prefix (suffix ), denoted by v ≤p u (respectively, v ≤s u). If ξ = x1 x2 · · · ∈ Qω , then ξ[n] = xn is the n-th letter of ξ, and ξ[: n] = x1 · · · xn its initial prefix of length n. Similarly, for m ≤ n we denote by ξ[m : n] = xm · · · xn the factor of ξ of length n − m + 1 between the m-th and the n-th letter of ξ, and by ξ[m :] = xm xm+1 · · · its tail. Two infinite sequences ξ, η ∈ Qω are said to be cofinal (written ξ ∝ η) if there exists an integer k such that ξ[k :] = η[k :]. By Q̃ = Q ∪ Q−1 we denote the involutive set where Q−1 is the set of formal inverses of Q. The operator −1 : Q → Q−1 sending q 7→ q −1 is extended to an e ∗ through involution on the free monoid Q e ∗ ). 1−1 = 1, (q −1 )−1 = q, (uv)−1 = v −1 u−1 (q ∈ Q; u, v ∈ Q e ∗ generated by the relation set {(qq −1 , 1) | q ∈ Q}. e Let ∼ be the congruence on Q ∗ ∗ e e The quotient FQ = Q / ∼ is the free group on Q, and let σ : Q → FQ be the e ∗ may be compactly canonical homomorphism. The set of all reduced words on Q 4 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO written as RQ = Q̃∗ r [ Q̃∗ qq −1 Q̃∗ . q∈Q̃ e ∗ , we denote by u ∈ RQ the (unique) reduced word ∼-equivalent For each u ∈ Q to u. With a slight abuse in the notation we often identify the elements of FQ with their reduced representatives, i.e. σ(u) = u; this clearly extends to subsets e∗ . σ(L) = L for L ⊆ Q A Mealy automaton is a tuple M = (Q, Σ, ·, ◦) where Q and Σ are finite set respectively called the stateset and the alphabet, and · , ◦ are functions from Q × Σ to, respectively, Q and Σ called the transition and the production function. This automaton can be seen as a complete, deterministic, letter-to-letter transducer with same input and output alphabet or, following [10], as a labelled digraph. The graphical representation is standard (see Fig. 2 for instance) and one displays transitions as follows: a|b q −−→ p ∈ M ⇐⇒ q·a = p, q◦a = b . It can be seen that the stateset Q and the alphabet Σ play a symmetric role, hence we can define a new Mealy automaton: the dual of the automaton M = (Q, Σ, ·, ◦) q|p is the automaton dM = (Σ, Q, ◦, ·) where we have the transition a−−→b whenever a|b q −−→p is a transition in M (see Fig. 2). a|q◦a For each automaton transition q −−→ q·a, we associate the cross-transition depicted in the following way: a q·a, q q◦a see also Fig. 1. If the functions (Σ → Σ : a 7→ q◦a)q∈Q are permutations the automaton is said to be invertible. On the other hand, when the functions (Q → Q : q 7→ q·a)a∈Σ are permutations the automaton is called reversible. Note that when an automaton is reversible its dual is reversible and the other way around. Mealy automata that are both invertible and reversible are called reversible invertible automata, or RIautomata for short. Other classes of automata will be described in Section 3. ◦ A Mealy automaton M = (Q, Σ, ·, ◦) defines inductively an action Q y Σ∗ of Q on Σ∗ by q◦(a1 · · · an ) = (q◦a1 ) ((q·a1 )◦(a2 · · · an )) , that can also be depicted by a cross-diagram by gluing cross-transitions (see Glasner and Mozes [13], or [1]) representing the action of a word of states on a word of letters (or vice-versa): a2 · · · an a1 (q·a1 )·(a2 · · · an ) . q·a1 q q◦a1 (q·a1 )◦(a2 · · · an ) BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 5 · In a dual way, this Mealy automaton defines also an action Q∗ x Σ. Both actions naturally extend to words, respectively in Q∗ and Σ∗ with the convention hg◦a = h◦(g◦a) and g·ab = (g·a)·b . In addition to these descriptions of a Mealy automaton, we are going to use another visualization, the helix graphs (introduced in [1]). The helix graph Hn,k of a Mealy automaton M = graph with nodes Qn × Σk and  (Q, Σ, ·, ◦) is thendirected k arcs (u, v) −→ u·v, u◦v for all (u, v) ∈ Q × Σ (see Fig. 1). 1 0 0|1 1|0 y x y x 1 0 1|1 0|0 x x x, 0 x, 1 y y, 0 y, 1 0 1 y x y 0 1 Figure 1. The Mealy automaton L generating the lamplighter group, the set of its cross-transitions, and its helix graph H1,1 (L). ◦ 2.2. Automaton groups. From the algebraic point of view, the action Q∗ y Σ∗ gives rise to a semigroup hM i+ generated by the endomorphisms q ∈ Q of the regular rooted tree identified with Σ∗ defined by q : u 7→ q◦u for u ∈ Σ∗ . Groups generated by invertible automata play an important role in group theory (for more details we refer the reader to [25]). In this framework all the maps q : u 7→ q◦u, q ∈ Q, are automorphisms of the regular rooted tree Σ∗ , and the group generated by these automorphisms is denoted by hM i (with identity 1). Note that ◦ · ◦ the actions Q∗ y Σ∗ and Q∗ x Σ∗ extend naturally to the actions hM i y Σ∗ · and hM i x Σ∗ , respectively. There is a natural way to factorize these actions using the wreath product [25, 3]. Let M = (Q, Σ, ·, ◦) be an invertible Mealy automaton. The inverse of the automorphism q is denoted by q −1 ∈ Q−1 = {q −1 : q ∈ Q}. There is an explicit way to express the actions of the inverses by considering the inverse automaton M −1 havb|a a|b ing Q−1 as stateset, and a transition q −1 −−→p−1 whenever q −−→p is a transition in M (see Fig. 2). 1|0 0|1 1|0 y x 1|1 0|1 −1 x|y y x −1 1|1 y|x 0 1 0|0 0|0 x|x L L−1 dL y|y Figure 2. The lamplighter automaton L, its inverse automaton L−1 , and its dual automaton dL. The action of the group hM i on Σ∗ , in case M is invertible (or of the semigroup hM i+ in a more general case), may be naturally extended on the boundary Σω of the tree. 6 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO This action gives rise to the so-called orbital graph. In general, given a finitely generated semigroup S, with set of generators Q, that acts on the left of a set X ◦ according to S y X, if π : Q∗ → S denotes the canonical map, then the orbital graph Γ(S, Q, X) is defined as the Q-digraph with set of vertices X, and there is a an edge x−−→y whenever π(a) ◦ x = y. When we want to pinpoint the connected component containing the element y ∈ X we use the shorter notation Γ(S, Q, X, y) instead of (Γ(S, Q, X), y). Note that in the realm of groups, this notion corresponds to the notion of Schreier graph. In particular for an invertible Mealy automaton M = (Q, Σ, ·, ◦) and a word v ∈ Σ∗ ⊔ Σω , if StabhM i (v) = {g ∈ hM i : g◦v = v} e corresponds to the is the stabilizer of v, the Schreier graph Sch(StabhM i (v), Q) connected component pinpointed by v of the orbital graph: e ≃ Γ(hM i, Q, e Σ∗ ⊔ Σω , v) . Sch(StabhM i (v), Q) This simply corresponds to consider the orbit of v as the vertex set and the edges given by the action of the generators of the group (in our context the state of the generating automaton). Henceforth, when the automaton group is clear from the context we will use the e more compact notation Sch(v) when we deal with Sch(StabhM i (v), Q). 3. The considered classes Throughout the paper we focus mainly on four classes of automata: contracting automata, reversible automata, bireversible automata, and automata with a sink. 3.1. Contracting automata. The notion of contracting automata has been introduced by V. Nekrashevych in [25]. For a Mealy automaton M = (Q, Σ, ·, ◦), the group hM i is said to be contracting if there exists a finite set ℵ ⊂ hM i such that, for any g ∈ hM i there exists an integer n = n(g) such that g·v ∈ ℵ, for any v ∈ Σ≥n . The set ℵ is called a nucleus of hM i. Graphically it means that, from any element of the group, a long enough path leads to the nucleus. This enables the construction of a finite automaton Mℵ with stateset ℵ, alphabet Σ, a|g◦a and transitions g −−−−→g·a. By extension, an automaton generating a contracting group is said to be contracting itself. Examples of such automata are depicted on Fig. 3 and 4. Note that if the contracting automaton M has a sink-state, this state necessarily belongs to ℵ. The importance of the notion of contracting automata refers to the beautiful and surprising connection with complex dynamics established by V. Nekrashevych [25]. With every contracting group one may associate a topological space called limit space, that is encoded by the set of left-infinite words on Σ modulo the equivalence relation given by the action of the nucleus, i.e., two left infinite sequences ξ = · · · ξn ξn−1 · · · ξ1 and η = · · · ηn ηn−1 · · · η1 are equivalent if for any n ≥ 1 there exists gn ∈ ℵ satisfying gn ◦ξn ξn−1 · · · ξ1 = ηn ηn−1 · · · η1 . It turns out that the iterated monodromy group IMG(f ) of a post-singularly finite rational function is contracting and its limit space is homeomorphic to the Julia set of f . This discovery puts in strict relation the dynamics of the map f and the algebraic properties of IMG(f ). As an example, this powerful correspondence has allowed L. Bartholdi BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 7 and V. Nekrashevych to solve a classical problem in complex dynamics, the so-called Hubbard Twisted Rabbit Problem by using algebraic methods [2]. a 2|2 0|0 1|1 a 0|1 0|1 0|2 1|0 1|1 b 0|0, 1|1, 2|2 b 2|0 1 2|1 1|2 1|0 1 c 1|1, 0|0 0|0 Figure 3. The contracting automata generating the Basilica group (on the left) and the Hanoi Towers group H (3) (on the right). 3.2. (Bi)reversible automata. The classes of reversible and bireversible Mealy automata are also interesting. We recall that a Mealy automaton is reversible whenever each input letter induces a permutation of the stateset, i.e. simultanea|b a|c ous transitions q −−→p and q ′ −−→p are forbidden. In the context of groups, we are especially interested in reversible invertible automata (called henceforward RIautomata). Moreover such a reversible automaton is bireversible if in addition each output leta|b ter induces a permutation of the stateset, i.e. simultaneous transitions q −−→p and ′ c|b q −−→p are also forbidden. In this case it is necessarily invertible. An interesting feature of an RI-automaton is that the dual of such an automaton is still reversible and invertible. The following lemma, which will be useful later, may be easily deduced from [32] or [10, Theorem 2]. Lemma 3.1. Let M be an RI-automaton with hM i ≃ FQ /N . Then, the following facts hold: (i) if g ∈ Q∗ is such that g·a ∈ N for some a ∈ Σ∗ , then g ∈ N ; (ii) let g, g ′ , g ′′ ∈ Q∗ with g·a = g ′ for some a ∈ Σ∗ , then there is a h ∈ Q∗ satisfying (hg)·a = g ′′ g ′ . e in (i). Furthermore, if M is bireversible then Q can be replaced by Q Proof. Let us prove point (i). First note that any state reachable from a state in N also belongs to N . Indeed let g ′ ∈ N and a′ , b′ ∈ Σ∗ , we have: a′ g ′ b′ ′ g ·a a ′ ′ b′ 8 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO So for any word b′ ∈ Σ∗ , g ′ ◦a′ b′ = a′ b′ , hence (g ′ ·a′ )◦b′ = b′ , i.e., g ′ ·a′ ∈ N . Let g ′ = g·a ∈ N . By the reversibility of the automaton there is an a′ ∈ Σ∗ such that g ′ ·a′ = g and we can conclude that g belongs to N . Property (ii) follows by observing that by reversibility there is an h ∈ Q∗ such that h·(g◦a) = g ′′ holds (remember that the words in Q∗ are written from right to left). In terms of cross-diagramms we obtain: a g ′ = g·a g g◦a h . ′′ g = h·(g◦a) The last statement follows by applying (i) to M ⊔ M −1 that is reversible by the bireversibility of M .  The structure of the groups generated by reversible or bireversible automata is far from being understood. For instance, for a long time the only known examples of groups generated by bireversible automata were finite, free, or free products of finite groups [34, 35, 25]. Recently examples of bireversible automata generating non-finitely presented groups have been exhibited in [7] and in [22, 29]. In this regard, we now provide an embedding result of any group generated by a bireversible automaton whose dual does not generate a free group, into the outer automorphism group of a free group of infinite rank. This fact may give some extra insight on the kind of groups that are generated considering bireversibility. Proposition 3.2. Let M be a bireversible automaton such that hM i is infinite and hdM i is not free. Then, there is a monomorphism φ : hM i ֒→ Out(F∞ ). Proof. Let M = (Q, Σ, ·, ◦). By [8, Theorem 4], if−1we−1 consider the enriched automaa |b a|b ton M − obtained from M by adding the edge p −−→ q for any edge q −−→p of M , then hM − i ≃ hM i. By [10, Theorem 2] we may express the group hdM i as the ◦ e ∗. quotient FΣ /N where N is the maximal subset invariant under the action Q y Σ We may regard N as a normal subgroup of the free group FΣ , in particular note that [FΣ : N ] < ∞ if and only if hdM i is finite, and so hM i is also finite [25, 30]. Therefore, [FΣ : N ] = ∞, whence N ≃ F∞ , since N is free by Nielsen’s theorem. Let us first prove that there is an embedding hM i ֒→ Aut(N ). By the stability ◦ e ∗ and the invertibility of M − , we have that for any of N under the action Q y Σ g ∈ hM i the map ψg : n 7→ g◦n for n ∈ N is a bijection of N that is also a homoe ∗ and n ∈ N . Hence, ψg ∈ Aut(N ). morphism since w·n = w holds for any w ∈ Q Furthermore, the map φ : hM i → Aut(N ) that sends g to ψg is a homomorphism since equality ψg′ g = ψg′ ◦ψg holds. This map is also injective. Indeed, assume that ψg1 = ψg2 for some g1 , g2 ∈ hM i. Since N is normal, then for any u ∈ Σ∗ there is a reduced element n ∈ N such that u is a prefix of n (take a suitable conjugate of a reduced non-trivial element of N ). Hence, ψg1 = ψg2 implies that g1 ◦u = g2 ◦u for any u ∈ Σ∗ , hence g1 = g2 . Hence we get the claim φ : hM i ֒→ Aut(N ). We now show that each automorphism ψg is not inner. Indeed, assume contrary to our claim, that ψg (n) = unu−1 , n ∈ N , for some reduced non-empty element u ∈ N . BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 9 Then, for any n ∈ N there is an integer ℓ(n) such that g ℓ(n) ◦n = n, whence we have n = ψgℓ(n) (n) = ψgℓ(n) (n) = uℓ(n) nu−ℓ(n) from which we get nuℓ(n) = uℓ(n) n. We consider the subgroup generated by n and uℓ(n) . By Nielsen’s theorem this subgroup is free, and so both n and uℓ(n) belongs to the same cyclic subgroup hhi for some h ∈ N . In particular, there is e ∗ common to both u and n. However, since n ∈ N a non-empty prefix h′ ∈ Σ e ∗ appears is arbitrary and N is normal by the same argument above each u ∈ Σ ′ as a prefix of some non-trivial reduced element n ∈ N , a contradiction. Hence, φ : hM i ֒→ Out(N ) ≃ Out(F∞ ).  Unfortunately, the condition of having a free group of infinite rank appears to be mandatory in Proposition 3.2. Indeed, the next proposition shows that the embedding of an automaton group generated by a bireversible automaton into the group of length preserving automorphisms of the free group Fm for some 1 < m < ∞, characterizes finite groups. In what follows we call an element ψ ∈ Aut(Fm ) length preserving if given any w ∈ Fm one has |w| = |ψ(w)|. We denote by Autℓp (Fm ) the subgroup of Aut(Fm ) formed by the length preserving automorphisms. Proposition 3.3. Let M be a bireversible automaton. There is a monomorphism φ : hM i ֒→ Autℓp (Fm ) if and only if hM i is finite. Proof. Let M = (Q, Σ, ·, ◦). Suppose hM i finite and let N be the maximal invariant ◦ e ∗ (as in Proposition 3.2), then [FΣ : N ] < ∞ and so subset for the action Q y Σ N ≃ Fm for some m. We proceed as in the proof of the previous proposition to show that φ : hM i → Autℓp (Fm ) is a monomorphism. Conversely, let consider an embedding φ : hM i ֒→ Autℓp (Fm ) for some m, and let R = {x1 , . . . , xm } be minimal set of generators of Fm . For any g ∈ hM i, let ψg be the corresponding automorphism in Autℓp (Fm ). Since R is finite and the automorphisms preserve the length, the set [ ψg (xi ) Ω= g∈hM i,i=1,...,m is clearly finite. Further, there is a natural homomorphism of hM i into Sym(Ω). Let us prove that it is actually a monomorphism. Indeed, let g 6= g ′ in hM i. Then ψg 6= ψg′ holds in Autℓp (Fm ). Since R is a generating set, then ψg (xik ) 6= ψg′ (xik ) for some xik ∈ R. We deduce hM i ֒→ Sym(Ω), and so hM i is finite.  For similar results that link automaton groups defined by bireversible Mealy automata and the group of automorphisms of a free group, the reader is referred to [24]. 3.3. Automata with sink. In the complement of the class of the RI-automata there is another interesting class that, in some sense, represents the opposite case: the class Sa of all the invertible Mealy automata with a sink-state e which is accessible from every state (the index “a” standing for accessible). We recall that a sink-state of a Mealy automaton is a special state e such that e·a = e and e◦a = a for any a ∈ A. Note that in this setting the sink-state is unique. The reason we require that the sink-state is accessible from every state will be clear 10 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO in Section 6. This class is rather broad and it contains many known classes of Mealy automata like automata with polynomial state activity [31]. Furthermore, in [10, Proposition 6] it is shown that this class is essentially formed by those automata for which every element g in the generated group has a g-regular element in the boundary (for the notion of g-regular element see for instance [26]). Moreover, this class is also included into the broader class of synchronizing automata for which some results on automaton groups can be found in [9]. The connection with synchronizing automata will also be crucial in Section 6 in characterizing automata whose associated set of reduced tiles do not tile the plane. In [8] the problem of finding free groups generated by automata in Sa is tackled. Indeed,until recently, all known free automaton groups were generated by bireversible automata. This led to the question whether or not it is possible to generate a free group by means of automata with a sink-state. In [8] a series of examples of automata from Sa generating free groups is exhibited. However, in this case the resulting free groups do not act transitively on the corresponding tree, so this leaves open the question of finding a free group generated by an automaton from Sa acting transitively on the rooted tree. This problem is also connected with the interesting combinatorial notion of fragile word introduced in [8]. 4. Topological properties of the action on the boundary In this section we describe some topological properties of the action of an automaton group on the boundary of a rooted tree. In particular, we consider the problem of continuity of the map that associates with any point in the boundary the corresponding Schreier graph. We prove that the set of those points where this function is not continuous has zero measure. Moreover, we provide a characterization of contracting automata whose action on the boundary is continuous everywhere. In the reversible case, we prove that examples of automata generating groups with all continuous points in the boundary are necessarily bireversible, and in this case, this condition may be rephrased in terms of triviality of the stabilizers in the boundary. 4.1. Action on the boundary. Let M = (Q, Σ, ·, ◦) be an invertible automaton e denote and Sub(hM i) denote the space of all subgroups of hM i and let Sch(hM i, Q) the space of marked Schreier graphs of hM i (i.e. Schreier graphs in which we have chosen a special vertex, the marked vertex) contained in the space of all marked labeled graphs and put ∂T = Σω . Both spaces may be endowed with a natural topology (also induced by an opportune metric). We endow the space Sub(hM i) with the Tikhonov topology of the space {0, 1}hM i in such a way that any subgroup H may be identified with its characteristic function. Given a finite subset F of hM i the F −neighborhood of a subgroup H contains all subgroups K such that H ∩ F = K ∩ F . Roughly speaking we say that two subgroups H and K of hM i are close if they share many elements. On the other hand, two marked Schreier graphs Sch(ξ) and Sch(η) are close when the subgraphs given by the balls of large radius around ξ and η are isomorphic, and two points ξ and η in ∂T are close if they share a long common prefix. Notice that in our notation, Sch(ξ) corresponds BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 11 e Σ∗ ⊔ Σω , ξ). to the graph Γ(hM i, Q, Vorobets studied the map F : ∂T −→ Sch(hM i, S) e ξ 7−→ Sch(StabhM i (ξ), Q) in the case where M is the Grigorchuk automaton [36]. His results may be summarized as follows: the closure F (∂T ) of the image of the boundary of the binary tree into the space of marked labeled Schreier graphs consists of a countable set of points (the one-ended boundary graphs) and another component containing all two-ended Schreier graphs. The Grigorchuk group acts on the compact component given by F (∂T ) without these isolated points by shifting the marked vertex of the graph and such action is minimal (every orbit is dense) and uniquely ergodic (there is a unique Borel probability measure on this set that is invariant under the action of the group). 4.2. Singular points. Let M = (Q, Σ, ·, ◦) invertible. We define the map St : Σω −→ Sub(hM i) ξ 7−→ StabhM i (ξ). The neighborhood stabilizer Stab0hM i (ξ) of ξ is the set of all g ∈ hM i that fix the point ξ together with its neighborhood (that may depend on g). One may check that Stab0hM i (ξ) is a normal subgroup of StabhM i (ξ). A point ξ ∈ Σω is called singular if the map St is not continuous at ξ. The set of singular points is denoted by κ. The following lemma clarifies the connection between the continuity of the map St and the dynamics in the boundary. Lemma 4.1. [36, Lemma 5.4] St is continuous at the point ξ if and only if the stabilizer of ξ under the action coincides with its neighborhood stabilizer, i.e.: ξ ∈ κ ⇐⇒ Stab0hM i (ξ) 6= StabhM i (ξ) . The following lemma characterizes continuous points in terms of restrictions. Lemma 4.2. Let M be an invertible automaton and let ξ be an element in Σω . The following are equivalent. (i) ξ is not singular; (ii) St is continuous at ξ; (iii) For any g ∈ StabhM i (ξ) there exists n such that g·ξ[: n] = 1; Proof. (i)⇔(ii), (iii)⇒(ii), follow from definition. (ii)⇒(iii). Let us prove ¬(iii)⇒¬(ii). Suppose that there exists g ∈ StabhM i (ξ) such that g·ξ[: n] 6= 1 for all n ≥ 0. Then we can find, for any n, a letter xn ∈ Σ such that (g·ξ[: n])◦xn = x′n 6= xn . Hence if we put ζn = ξ[: n]xn ξ[n + 2 :] ∈ Σω we get g◦ζn = g◦ξ[: n]xn ξ[n + 2 :] = ξ[: n]x′n ξ ′ 6= ξ. Hence, since we can construct a ζn in any neigbourhood of ξ, Stab0hM i (v ω ) 6= StabhM i (v ω ), and St is not continuous.  Moreover we can characterize continuous points by looking only to periodic points Lemma 4.3. Let M be an invertible automaton. The following are equivalent. 12 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO (i) There is no singular point in Σω ; (ii) There is no singular periodic point in Σω . Proof. (i) ⇒ (ii) is obvious, let us prove the converse, by contraposition. Assume that St is not continuous at some ξ ∈ Σω and let g ∈ StabhM i (ξ) r Stab0hM i (ξ). If there exists k such that g·ξ[: k] = 1 then g stabilizes some neighborhood U of ξ and is contained in Stab0hM i (ξ). Hence, for any n ≥ 1, g·ξ[: n] is a non-trivial element in hM i. The set { g·ξ[: k], k ∈ N } is finite, this implies that there exist m and n such that n > m > 0 and g ′ := g·ξ[: m] = g·ξ[: n] 6= 1. Therefore g ′ ·ξ[m + 1 : n] = g ′ and g ′ ◦ξ[m + 1 : n] = ξ[m + 1 : n]. Put v = ξ[m + 1 : n]: g ′ ∈ StabhM i (v ω ). In order to prove that g ′ 6∈ Stab0hM i (ξ) we notice that, since g ′ is not trivial, there exists w ∈ Σ∗ such that g ′ ◦w = w′ 6= w. Consider the sequence wk := v k wv ω for k ≥ 0. Clearly, for any neighborhood U of v ω there exists n such that wn ∈ U . But g ′ ◦wn = g ′ ◦v n (g ′ ·v n )◦(wv ω ) = v n w′ v ′ 6= wn for some v ′ ∈ Σω . Therefore g ′ ∈ StabhM i (v ω ) 6= Stab0hM i (v ω ).  In the following theorem we prove that the measure of the set κ of singular points is zero. For the sake of completeness we recall that a subset of a topological space X is nowhere dense if its closure has an empty interior. A subset is meager in X if it is a union of countably many nowhere dense subsets. A Baire space, as Σω with the usual topology, cannot be given by the countable union of disjoint nowhere dense sets. In general the notion of nowhere dense and meager set do not coincide with the notion of zero-measure. In [36] it is proven that κ is meager. e ∗ , denote by Fix(u) the set consisting in the vertices ξ ∈ Σω fixed by Given u ∈ Q the action of u. If w ∈ Σk is an element stabilized by u we write w ∈ Fixk (u). By using the ideas developed in [20] we are able to give the following characterization. Theorem 4.4. For any invertible automaton, the set measure zero. κ of singular points has Proof. The proof is heavily based on the ideas contained in Proposition 4.1 and e ∗ , k ≥ 1, consider the following sets: Theorem 4.2 of [20]. For u ∈ Q χ(u) = {ξ ∈ Σω : ξ ∈ Fix(u) and π(u·ξ[: j]) 6= 1  χk (u) = w ∈ Σk : w ∈ Fixk (u) and π(u·ξ[: j]) 6= 1 By Lemma 4.2, the set of singular points is [ χ(u). κ= ∀j ≥ 0} , 0≤j≤k . e∗ u∈Q e ∗ . Since Let us prove µ(χ(u)) = 0, for any u ∈ Q \ χ(u) = (χk (u)Σω ) , k≥1 we get: |χk (u)| . k→∞ k→∞ |Σ|k We now show that this limit is 0. Indeed, as proved below, there is an integer p such that µ(χ(u)) = lim µ(χk (u)Σω ) = lim (Ek ) |χpk (u)| ≤ (|Σ|p − 1)k , for all k ≥ 1 . BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 13 e i : π(v) 6= 1}. Since H|u| is finite, there is an integer p such that Let Hi = {v ∈ Q no element of H|u| induces the identity on Σp : take it for Equation (Ek ). Use an induction on k ≥ 1. For k = 1: χp (u) ⊆ Σp and Fixp (u) 6= Σp by the choice of p. Suppose that |χp(k−1) (u)| ≤ (|Σ|p − 1)k−1 . Since u·h ∈ H|u| for any h ∈ χp(k−1) (u), there is a v ∈ Σp that is not fixed by u·h, whence |χp·k (u)| ≤ |χp(k−1) (u)|(|Σ|p − 1) ≤ (|Σ|p − 1)k .  4.3. The contracting case. The property of being contracting allows us to characterize the set κ of singular points in terms of a language recognized by an automaton. Let M = (Q, Σ, ·, ◦) be a Mealy automaton. We define its stable automaton as the automaton B(M ) on infinite words where for a ∈ Σ we have: a q −−→p ∈ B(M ) ⇐⇒ a|a q −−→p ∈ M . Given a Büchi acceptance condition (i.e. a set of states that has to be visited infinity often), such an automaton recognizes a language of right-infinite words (see for instance [33]). See Figs. 4 and 5. Theorem 4.5. Let M be a invertible automaton admitting a finite nucleus automaton N . Then the set κ of singular points is included in the set of word cofinal with a word in the language recognized by the Büchi automaton B(N ) with every state but the sink-state accepting. Conversely any word recognized by B(N ) is singular. In particular if the language B(N ) is empty then κ = ∅. Proof. Let e be the sink-state of the nucleus and suppose that there exists ξ ∈ Σω such that St is not continuous on ξ. From Lemma 4.2 it follows that there exists g ∈ StabhM i (ξ) such that g·ξ[: k] 6= e for every k > 0. Since hM i is contracting there exists gn ∈ N such that gn = g·ξ[: n] ∈ N . Notice that gn+k := g·ξ[: n + k] = gn ·ξ[n + 1 : n + k] ∈ N r {e} and gn+k ◦ξ[n + k] = ξ[n + k]. Since B(N ) has the same set of states as N and admits a transition exactly when g◦i = i holds, there is an infinite path ξ[n + k] ξ[n + 1] ξ[n] gn −−→gn+1 −−→ · · · gn+k −−→ · · · that avoids e. Furthermore since the stateset is finite there is a state that is infinitely visited, hence the run is accepted by B(N ). To prove that ξ is cofinal to a word recognized by B(N ) notice that since ξ[n :] is recognized and N is finite, there exists an integer m such that g·ξ[: m] = gm+1 belongs to a strongly connected component C of N . Hence, by reading transition in C backward, we can extend ξ[m :] on ′ the left to obtain a word ξ ′ = ξ ′ [1] · · · ξ ′ [m − 1]ξ[m :] and a sequence g ′ , g1′ , . . . , gm ′ ′ ′ ′ ′ of elements of C such that there exist g ·ξ [: k] = gk+1 for k ≤ m, g ·ξ [: m] = gm+1 and g ′ ◦ξ ′ [: k] = ξ. Then B(N ) recognizes ξ ′ , and the points ξ ′ and ξ are cofinal. ik i1 On the other hand if a run g1 −−→g 2 · · · gk −−→gk+1 · · · is accepted by B(N ) then it is infinite and it does not visit e (since it is a sink, i.e. an absorbing state). Hence g1 stabilizes ξ = i1 i2 · · · and satisfies π(g1 ) 6= 1. So by Lemma 4.2, St is not continuous.  Notice that if ξ ∈ Σω is singular, so is each element of its orbit. Indeed if g◦ξ = ξ with g 6= 1 then we have hgh−1 ◦(h◦ξ) = h◦ξ with hgh−1 6= 1 for all h ∈ hM i. So the previous characterization is exact when the set of words cofinal to a word in the language B(N ) coincide with the orbit of words in the language B(N ). 14 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO In particular this occurs for self-replicating automata. A Mealy automaton is called self-replicating (also called fractal ) whenever for any word u ∈ Σ∗ and any g ∈ hM i, there exists h ∈ StabhM i (u) satisfying h·u = g. Proposition 4.6. Let M be a contracting, self-replicating Mealy automaton. Then κ is the set of words cofinal to a word in the language recognized by B(N ). Proof. The first inclusion comes directly from Theorem 4.5. For the other one, let ξ be recognized by B(M ) and uξ[n :] be cofinal to ξ. Since ξ[n :] is recognized by B(N ) it is singular. So there exists g ∈ hM i satisfying g◦ξ[n :] = ξ[n :] with g·ξ[n : n + k] 6= 1 for all k ≥ 0. Now, since M is self-replicating there exists h ∈ hM i satisfying the following cross diagram: u h ξ[n :] h·u = g u ξ[n :] Hence uξ is singular.  Note that, despite of these strong requirements, the class of contracting and selfreplicating automata is wide. It contains in particular the Grigorchuk automaton, the Hanoi Tower automaton or the basilica automaton. One can notice that all states in a connected component have same type (accepting or rejecting). In that case the automaton is said to be weak Büchi [27]. It follows that the language of B(N ) is closed and regular. It is also (infinite) suffix closed. Moreover we get informations about the topology of the border F (∂T ). Proposition 4.7. Let M be a Mealy automaton and a be a letter in its alphabet. If there exists a state q such that q◦a = a and q·a = q, and p·a 6= p for p 6= q ∈ Q, then F (aω ) = Sch(aω ) is an isolated point in the closure of F (∂T ). Proof. Recall that two rooted graphs are close if they have balls of large radius around the roots that are isomorphic. Let us prove StabhM i (aω ) 6= Stab0hM i (aω ). The Schreier graph associated with aω contains a loop rooted at aω and labeled by q. Any other element in ∂T that is not cofinal to aω , contains at some position a letter other than a, so that it is not fixed by q. This implies that F (aω ) is isolated, since no other graph contains a loop labeled by q. The same argument works for the vertices of the orbit of aω , that are the only ones having at finite distance a vertex with a loop labeled by q.  Note that this is still true when one consider words on Σk instead of single letters. We now apply the previous characterizations to two examples of contracting, self-replicating automata, namely the Basilica automaton et the Hanoi Towers automaton. The Basilica group, introduced in [18], is generated by the automorphisms a and b having the following self-similar form: a = (b, 1), b = (a, 1)(01), where (01) denotes the nontrivial permutation of the symmetric group on {0, 1, 2} and, with a slightly abuse of notation, 1 denotes the sink-state (see Fig. 3). In Fig. 4 the nucleus automaton associated with this group is presented. Since the stable automaton accepts no infinite word we obtain: BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 15 0|1 ba−1 ab−1 ba−1 ab−1 1|0 b 0|0 1|0 1|1 0|1 a 1|0 0|1 1|1 1 1|1, 0|0 a−1 1|0 0|1 1 b 0|0 1 0 1 a b−1 a−1 1, 0 0 b−1 Figure 4. The nucleus automaton (on the left) and the stable automaton (on the right), associated with the Basilica automaton. Corollary 4.8. For the Basilica group, the set κ of singular points is empty. We now consider the case of the Hanoi Towers group H (3) (see Fig. 3) introduced in [17]. This group is generated by the automorphisms of the ternary rooted tree having the following self-similar form: a = (1, 1, a)(01) b = (1, b, 1)(02), and c = (c, 1, 1)(12), where (01), (02), and (12) are elements of the symmetric group on {0, 1, 2}. Observe that a, b, c are involutions. Corollary 4.9. For the Hanoi Towers group H (3) , the set κ of singular points is a countable set consisting of the (disjoint union of the) orbits of the three points 0ω , 1ω , and 2ω . Moreover F (iω ) = Sch(iω ) is an isolated point in the closure of F (∂T ) for i ∈ {0, 1, 2}. 2 1 a 0, 1, 2 b 1 c 0 Figure 5. The stable automaton B(H (3) ) for the Hanoi Towers automaton. 16 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO Since Sch (iω ) is isolated in F (∂T ) for i ∈ {0, 1, 2}, one can ask about the behaviour of the sequence of Schreier graphs of finite words converging to iω . It turns out that this sequence is converging, but the limit is not a Schreier graph for the Hanoi Towers group.  Let ηni n∈N , i = 0, 1, 2 be a sequence of elements of Σ∗ , such that |ηni | = n, converging to iω . Recall that Sch(iω ) contains only one loop rooted at iω . Let Υi , i = 0, 1, 2 be the graph obtained from Sch(iω ) as follows: (1) Take two copies of Sch(iω ) and let gi ∈ {a, b, c} be the label of the loop at iω . (2) Erase the loop at iω in each copy of Sch(iω ). (3) Join the two copies by an edge labeled gi and connecting the vertices iω of each copy and choose one of these iω as marked vertex. c c 02k−1 0α 02k−1 1α b b a a 2k 0α a 2k 1α c c 12k−1 0α 12k−1 1α b b c 02ω b a a 2ω c 12ω b Figure 6. Part of the (finite) Schreier graphs in the Hanoi Towers group H (3) of 2k 0α (above) and of the (infinite) Schreier graph of 2ω (below). See proof of Theorem 4.10. Theorem 4.10. F (ηni ) converges to Υi for i ∈ {0, 1, 2} as n → ∞. BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 17  Proof. Let zni n∈N be a sequence of natural numbers such that zni is the position of the first letter in ηni different from i. It comes from the structure of finite Schreier graphs of H (3) (see Fig. 6) that the balls of radius zni − 1 in Υi rooted at iω , and in Sch(ηni ) rooted at ηni are isomorphic. Since zni goes to infinity as n does, we have the assertion.  It can be shown that the infinite Schreier graphs of the Hanoi Tower group are all one ended, see [6]. Hence, since Υi is clearly two ended it follows that it is not an infinite Schreier graph of H (3) . More precisely there is no ξ ∈ ∂T = Σω such that the orbital Schreier graph Sch(ξ) is isomorphic to Υi , even if the graph is considered non marked. 0|0 d0 0|0 1|1 b 0|0 1|1 c0 1|1 d1 1|1 c1 1|1 0|0 a 0|0 0|1 1|0 1 Figure 7. A twisted version of the Grigorchuk automaton. Note that Grigorchuk automaton can be twisted in order to obtain a contracting automaton generating a fractal group with non countable κ, see Fig 7. However one can ask if there exist minimal automata where two singular points have isomorphic Shreier graphs, see Problem 2. 4.4. The bireversible case. In this section, we focus on the bireversible case. We start with the following alternative result which shows that, in the class of bireversible automata, the problem of finding examples with all continuous points in the boundary is equivalent to look for automata with all trivial stabilizers in the boundary. Proposition 4.11. Let M be a bireversible automaton. For any ξ ∈ Σω , we have: ξ ∈ κ ⇔ StabhM i (ξ) 6= {1}. Proof. Let g ∈ StabhM i (ξ). We have only two possibilities: either there exists n ≥ 1 such that g·ξ[: n] = 1, or for every n ≥ 1 one has g·ξ[: n] 6= 1. In the first case, since M is bireversible we may apply Lemma 3.1, which implies g = 1. On the other hand, suppose g·ξ[: n] = gn ∈ hM i r {1}. We have already remarked that St : ∂T −→ Sub(hM i) is not continuous at ξ if Stab0hM i (ξ) 6= StabhM i (ξ). Since gn 6= 1, for every n there exists wn ∈ Σ∗ such that gn ◦wn 6= wn . Let η (n) := ξ[: n]wn ξ[n + |wn | + 1 :] ∈ Σω . Notice that η (n) → ξ, so that for every 18 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO neighborhood U of ξ there exists k = kU ≥ 1 such that η (k) ∈ U . By hypothesis g ∈ StabhM i (ξ) and g·η (k) [: k] = g·ξ[: k] = gk . The sequence η (n) converges to ξ since they have the same prefix of length n, so given U we find k such that η (k) ∈ U . We have g◦η (k) = g◦ξ[: k] ((g·η[: k])◦wk ξ[n + |wn | + 1 :]) = ξ[: k](gk ◦wk ) ((gk ·wk )◦ξ[n + |wn | + 1 :]) 6= η (k) . This implies that Stab0hM i (x) 6= StabhM i (x), which means x ∈ κ.  We recall that the action of hM i on Σω is essentially free if µ({x ∈ Σω : StabhM i (x) 6= {1}}) = 0, where µ is the uniform measure on Σω (see [15]). One may prove that groups generated by bireversible automata give rise to essentially free actions on Σω [32]. Equivalently, this fact may be deduced by Proposition 4.11 and Lemma 4.2. The next proposition characterizes in terms of Schreier graphs those automaton groups having all trivial stabilizers in the boundary, which by the previous proposition are those with κ = ∅. Proposition 4.12. Let M be a bireversible automaton. The set κ of singular points is empty if and only if any two words in a same orbit cannot be cofinal. Proof. First, suppose that there exist ξ ∈ Σω and 1 6= g ∈ hM i such that η := g◦ξ ∝ ξ. Let us prove that there is an element of Σω whose stabilizer is not trivial. Since η ∝ ξ, then there exists N > 0 such that ξ[N :] = η[N :]. Since M is bireversible, from Lemma 3.1 gk := g·ξ[: k − 1] 6= 1 for any k. Since g◦ξ = η, one gets gk ◦ξ[k :] = (g·ξ[: k − 1])◦ξ[k :] = η[k :]. In particular, for k = N we get gN ◦ξ[N :] = gN ◦η[N :] = η[N :] = ξ[N :]. This implies that ξ ′ := ξ[N :] is stabilized by gN 6= 1 and so it is a boundary point with a non-trivial stabilizer, so it is a singular point. On the other hand, let hM i admit a boundary point ξ with a non-trivial stabilizer. Then, there exists 1 6= g ∈ hM i such that g·ξ = g. Since g 6= 1 and M is bireversible, there exist v ∈ Σ∗ and f ∈ Q∗ , such that f ◦v 6= v and f ·v = g. Hence vξ and f ◦(vξ) are two distinct cofinal words.  5. Commuting pairs and dynamics on the boundary Theorem 4.4 from Section 4 states that the set κ of all singular points has measure zero. In this section we are interested in seeking for examples of automata with κ = ∅. Note that by Proposition 4.11, in the bireversible case this property is equivalent to have all trivial stabilizers in the boundary [16]. On the other hand, no example of such dynamics is known in this class (which seems to be the most difficult case), unless the generated group is finite. In this section we focus our attention on the class of reversible invertible automata. BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 19 5.1. Commuting pairs. Given an automaton M = (Q, Σ, ·, ◦), we say that v ∈ Σ∗ , u ∈ Q∗ commute whenever u·v = u and u◦v = v ; in this case we say that (u, v) is a commuting pair. The previous definition considers words v ∈ Σ∗ , u ∈ Q∗ . However, we may consider commuting pairs in M ⊔ M −1 = e Σ, ·, ◦). The importance of commuting pairs stems in the connection with the (Q, stabilizers of periodic points on the boundary. For example a classical way to prove that an automaton is not contracting has an interpretation in terms of commuting pairs: if there exists a commuting pair (u, v) such that u has infinite order then the automaton cannot be contracting [11]. We put ∗ Stab+ hM i (ξ) := StabhM i+ (ξ) = π(Q ) ∩ StabhM i (ξ) as the set of “positive” stabilizers. We have the following proposition that clarifies the connection between commuting pairs and stabilizers. e ∗ (u ∈ Q∗ ) Proposition 5.1. Let M be an invertible automaton. If v ∈ Σ∗ , u ∈ Q + ω ω commute, then π(u) ∈ StabhM i (v ) (respectively, StabhM i (v )). e ∗ (π(u) ∈ Q∗ ) such that π(u) ∈ StabhM i (v ω ) Conversely, for any v ∈ Σ∗ , π(u) ∈ Q ω j n (respectively, u ∈ Stab+ hM i (v )) then there are integers n, j ≥ 1 such that u·v , v commute. Moreover if M is an RI-automaton then one can take j = 0, i.e. u, v n commute. Proof. The first statement follows from u◦v i = v i , u·v i = u, for all i ≥ 1, i.e. ω ∗ π(u) ∈ StabhM i (v ω ) (π(u) ∈ Stab+ hM i (v ) in case u ∈ Q ). On the other hand, since the stateset is finite there exists k such that, ∀j ≥ k, u·v j is in the strongly connected component of u·v k . Then by strong connectivity there exist n and j ≥ k such that (u·v j )·v n = u·v j . Moreover (u·v j )◦v n = v n , hence (u·v j , v n ) is a commuting pair. In particular if M is in addition reversible then every connected component is in fact strongly connected, hence one can choose j = 0.  ek,n . The following proposition The helix graph of M ⊔ M −1 is denoted by H gives a way to build all pairs of commuting pairs by looking at the “labels” of the cycles of the helix graphs. Lemma 5.2. Let M be an invertible automaton. Let (u0 , v0 )−−→(u1 , v1 )−−→ · · · −−→(um , vm )−−→(u0 , v0 ) e k,n ). Let u = um · · · u0 and v = v0 · · · vm , then be a cycle in the helix graph Hk,n (H u, v commute. e ∗ ) there is a helix graph Conversely, for any commuting pair v ∈ Σ∗ , u ∈ Q∗ (u ∈ Q ek,n ) such that (u, v)−−→(u, v). Hk,n (H e k ), v0 ∈ Σn there is a path Furthermore, for any u0 ∈ Qk (u0 ∈ Q (u0 , v0 )−−→(u1 , v1 )−−→ · · · −−→(uℓ , vℓ )−−→ · · · −−→(uℓ , vℓ ) ek,n ), for some ℓ ≥ 0. in the helix graph Hk,n (H Proof. From the definition of the helix graph we get: (um · · · u0 )◦(v0 · · · vm ) = v0 · · · vm , (um · · · u0 )·(v0 · · · vm ) = um · · · u0 20 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO that is, u, v commute. On the other hand, if u·v = u and u◦v = v, then the path (u, v)−−→(u, v) is a cycle in Hk,n for k = |u|, n = |v|. The last statement is a consequence of the determinism of an automaton and the fact that any pair (u, v) has out-degree one, i.e., for (u, v) there is exactly one pair (u′ , v ′ ) such that (u, v)−−→(u′ , v ′ ) is an edge in the helix graph.  5.2. The reversible-invertible case. In this section we prove that those examples of RI-automata having all continuous points (for the map St) in the boundary (if any) are located in the class of bireversible automata. In other words an RIautomaton where the set of singular points is empty is necessarily bireversible. We also show a series of equivalences and connections with the property of having all trivial stabilizers in the boundary. Examples of RI-automata with the aforementioned property are strictly related to a particular class of groups. Given a group G presented by hX|Ri, we say that G is fully positive whenever R ⊆ X ∗ , and for any x ∈ X there is a word u ∈ R such that x is a prefix of u. Note that the last property is equivalent to the fact that each x−1 ∈ X −1 may be expressed as a positive element ux ∈ X ∗ . These groups have the following alternative (apparently stronger) definition. e ∗ → G be Proposition 5.3. Let G be a group presented by hX|Ri, and let π : X the natural map. The following are equivalent. (i) G is fully positive; (ii) For any u ∈ X ∗ there is v ∈ X ∗ such that π(uv) = 1 in G; (iii) For any u ∈ X ∗ there is v ∈ X ∗ such that π(vu) = 1 in G. Proof. (ii)⇒(i). Follows by substituting every negative occurrence of a generator in a relator by a positive word. (i)⇒(ii). Let us prove the statement by induction on the length |u|. The case |u| = 1 follows from the fact that G is fully positive. Therefore, suppose that the statement holds for |u| < n and let us prove it for |u| = n. Consider any u = au′ ∈ X n with |u′ | = n − 1, for some u′ ∈ X ∗ and a ∈ X. By the induction hypothesis there is v ′ ∈ X ∗ such that π(u′ v ′ ) = 1. By the definition there is a defining relation ah ∈ R starting with a ∈ X. Take v = uv ′ h ∈ X ∗ , then it is easy to check that π(uv ′ h) = π(v) = 1 holds. Equivalence (ii)⇔(iii) follows by conjugation.  Note that all torsion groups are fully positive. The following lemma is a direct consequence of Proposition 15 in [10]. Lemma 5.4. Let M = (Q, Σ, ·, ◦) be an invertible automaton. Then, for any ξ ∈ Σω and π(u) ∈ Stab+ hM i (ξ) (π(u) ∈ StabhM i (ξ)) there exist i > j ≥ 1 such that + π(u·ξ[: j]) ∈ StabhM i (ξ[j : i]ω ) (respectively, π(u·ξ[: j]) ∈ StabhM i (ξ[j : i]ω )). Proof. Let π(u) ∈ Stab+ hM i (ξ), then we have u◦ξ[: n] = ξ[: n] for all n ≥ 1. Furthermore, by the finiteness of Q|u| , and since {u·ξ[: k]}k>0 is infinite, there are two indices i > j ≥ 1 such that u·ξ[: i] = u·ξ[: j]. From which it follows that (u·ξ[: j])◦(ξ[j : i]ω ) = ξ[j : i]ω ω i.e., π(u·ξ[: j]) ∈ Stab+ hM i (ξ[j : i] ). The general case π(u) ∈ StabhM i (ξ) is treated analogously considering M ⊔ M −1 instead of M .  BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 21 From which we derive the two following consequences. We have the following theorem. Theorem 5.5. Let M = (Q, Σ, ·, ◦) be an RI-automaton such that, for all ξ ∈ Σω , Stab+ hM i (ξ) contains only torsion elements. Then hM i is fully positive. ω ∗ Proof. In particular, Stab+ hM i (v ) contains only torsion elements for all v ∈ Σ . ∗ Take any arbitrary u0 ∈ Q , we show that there is a relation having u0 as a suffix, whence the statement follows by Proposition 5.3. By Lemma 5.2 there is path (u0 , v0 ) −→ · · · −→ (uℓ , vℓ ) −→ (uℓ+1 , vℓ+1 ) −→ · · · −→ (uℓ+k , vℓ+k ) = (uℓ , vℓ ) in the helix graph H|u0 |,n for some v0 ∈ Σn , ℓ ≥ 0. Since we have a loop around the vertex (uℓ , vℓ ), by Lemma 5.2 u = uℓ+1 · · · uℓ+k and v = vℓ+1 · · · vℓ+k are a ω m commuting pair. Proposition 5.1 implies π(u) ∈ Stab+ hM i (v ), thus π(u ) = 1 for some m ≥ 1. Hence uℓ is the suffix of some relation. Since u0 and uℓ belong to the same connected component, by using Lemma 3.1, we get a relation ending with u0 .  ω Note that, in particular, if Stab+ hM i (ξ) = {1} for all ξ ∈ Σ , then hM i is fully positive. Here are two very similar results on stabilizers. One when stabilizers are trivial, the other when they contain only torsion elements. Proposition 5.6. Let M = (Q, Σ, ·, ◦) be an RI-automaton. The following are equivalent. ω (i) Stab+ hM i (ξ) = {1} for all ξ ∈ Σ ; + ω (ii) StabhM i (v ) = {1} for all v ∈ Σ∗ ; (iii) StabhM i (v ω ) = {1} for all v ∈ Σ∗ ; (iv) StabhM i (ξ) = {1} for all ξ ∈ Σω . Proof. (i)⇔(ii). The implication is trivial. Let us look at the reciprocal implication: let π(u) ∈ Stab+ hM i (ξ). By Lemma 5.4, there are integers i > j ≥ 1 such that π(u·ξ[: + j]) ∈ StabhM i (ξ[j : i]ω ), which is trivial by hypothesis. Hence π(u·ξ[: j]) = 1, and so, by reversibility and Lemma 3.1, π(u) = 1. (ii)⇔(iii). The converse (iii)⇒(ii) is trivial. Conversely, if (ii) holds, then by e we can Theorem 5.5 we have that hM i is fully positive. Therefore, for any u ∈ Q + ∗ + construct a word u ∈ Q such that π(u ) = π(u). The result follows. (iii)⇔(iv) follows from [10, Proposition 15].  Proposition 5.7. Let M = (Q, Σ, ·, ◦) be an RI-automaton. The following are equivalent. ω (i) Stab+ hM i (ξ) is formed by torsion elements, for all ξ ∈ Σ ; + ω (ii) StabhM i (v ) is formed by torsion elements for all v ∈ Σ∗ ; (iii) StabhM i (v ω ) is a torsion group for all v ∈ Σ∗ ; Furthermore, if M is bireversible, following is also equivalent: (iv) StabhM i (ξ) is a torsion group for all ξ ∈ Σω . Proof. (i)⇔(ii). The implication is trivial. Let us look at the reciprocal implication: let π(u) ∈ Stab+ hM i (ξ). By Lemma 5.4, there are integers i > j ≥ 1 such that ω π(u·ξ[: j]) ∈ Stab+ hM i (ξ[j : i] ), which contains only torsion elements by hypothesis. 22 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO Hence, there is an integer ℓ ≥ 1 such that π (u·ξ[: j])ℓ π(u) ∈ Stab+ hM i (ξ) the following equality:  = 1. Further, since (u·ξ[: j])ℓ = uℓ ·ξ[: j] holds for any j ≥ 1.  Thus, by Lemma 3.1, we deduce π(u) = 1 and we get π(uℓ ) = π (u·ξ[: j])ℓ = 1, i.e., π(u) is torsion. (ii)⇔(iii). The proof is similar as the one in Proposition 5.6. In the bireversible case, the implication (iv)⇒(iii) is trivial, while the converse is proven similarly to Proposition 5.6. Take π(u) ∈ StabhM i (ξ). By Lemma 5.4, there are integers i > j ≥ 1 such that π(u·ξ[: j]) ∈ StabhM i (ξ[j : i]ω ). Since ω the group Stab  hM i (ξ[j : i] ) is torsion, then there is an integer ℓ ≥ 1 such ℓthat ℓ π (u·ξ[: j]) = 1. Hence, by the bireversibility and Lemma 3.1, we get π(u ) =  π(uℓ ·ξ[: j]) = π (u·ξ[: j])ℓ = 1. In particular, note that the previous propositions imply the existence of a nontrivial “positive” stabilizer whenever the action on the boundary for an RI-automaton has at least one non-trivial stabilizer. Let M = (Q, Σ, ·, ◦) be an invertible automaton, and let hM i = FQ /N . The set of “positive relations” of hM i is P(M ) = Q+ ∩ ψ −1 (N ) e ∗ → FQ is the canonical homomorphism. Note that P(M ) = ∅ implies where ψ : Q + that hM i is torsion-free and therefore infinite. ω ∗ Lemma 5.8. If P(M ) = ∅, then Stab+ hM i (v ) 6= {1} for some v ∈ Σ . Proof. Let (u0 , v0 ) −→ · · · −→ (uℓ , vℓ ) −→ (u0 , v0 ) be a cycle in the helix graph H1,1 . Then by Lemma 5.2 (u0 . . . uℓ , v0 . . . vℓ ) is a ω commuting pair. Hence π(u0 . . . uℓ ) ∈ Stab+ hM i ((v0 . . . vℓ ) ), and since P(M ) = ∅, + ω we conclude StabhM i ((v0 · · · vℓ ) ) 6= {1}.  The next theorem shows that either hM i or hdM i may have all trivial stabilizers in the boundary. We first recall the following proposition. Proposition 5.9. [8, Corollary 5] Let M = (Q,   Σ, ·, ◦) be an RI-automaton with hM i infinite. Then the index hM i : StabhM i (y ω ) is infinite for all y ∈ Σ∗ , if and only if P(dM ) = ∅. The following simple proposition is technical: Proposition 5.10. Let M = (Q, Σ, ·, ◦) be an RI-automaton with hM i infinite. ω ∗ If Stab+ hM i (y ) = {1} for all y ∈ Σ , we have P(dM ) = ∅. ω ∗ Proof. Suppose Stab+ hM i (y ) = {1} for all y ∈ Σ . Let us prove   hM i : StabhM i (y ω ) = ∞   for all y ∈ Σ∗ . Indeed, if hM i : StabhM i (y ω ) < ∞ for some y ∈ Σ∗ , then the following property holds:   ω k (1) ∃k ≤ hM i : StabhM i (y ω ) such that ∀g ∈ Stab+ hM i (y ) : g = 1 . In particular, by Proposition 5.3 we get that hM i is fully positive. We claim that every element of hM i is torsion. Indeed, since hM i is fully positive we have that BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 23 e ∗, for any q ∈ Q there is a uq ∈ Q∗ such that π(q −1 ) = π(uq ). Now take any w ∈ Q −1 −1 by substituting each negative occurrence q ∈ Q appearing in w with uq , we obtain a “positive” word w b ∈ Q∗ such that π(w) b = π(w). Thus, by (1) we have: π(w)k = π(w) b k = π(w bk ) = 1, that is, hM i is torsion. Being a residually finite group with uniformly bounded torsion, we deduce from [37,  38] that hM i is finite, a contradiction. Hence, we obtain hM i : StabhM i (y ω ) = ∞ for all y ∈ Σ∗ , and so, by Proposition 5.9, ω P(dM ) = ∅. In particular, by Corollary 5.8, P(dM ) = ∅ implies Stab+ D (z ) 6= {1} ∗ for some z ∈ Q , hence the last statement holds.  The latter admits a partial converse in the case where there exists an aperiodic element in hM i: Proposition 5.11. Let M be a RI Mealy automaton. If hM i is not torsion and P(dM ) 6= ∅ then ∃y ∈ Σ∗ , Stab+ hM i (y) 6= {1} e ∗ such that π(u) is aperiodic. Since P(dM ) is not empty there Proof. Let u ∈ Q e )|u|,|v| , there is a path exists y such that σ(y) = 1. Then in the helix graph H(M (u, y) −→ (u, y1 ) −→ · · · −→ (u, yk ) −→ (u, yk+1 ) −→ · · · −→ (u, yℓ ) −→ (u, yk ) . Hence ul−k and yk · · · yl commutes. Then π(ul−k ) ∈ StabhM i ((yk · · · yl )ω ) and π(ul−k ) 6= 1. We conclude using Prop. 5.6.  Theorem 5.12. Let M = (Q, Σ, ·, ◦) be an RI-automaton with hM i infinite. We + ω ∗ ω have either Stab+ hM i (y ) 6= {1} for some y ∈ Σ , or StabhdM i (z ) 6= {1} for ∗ some z ∈ Q . Now, we focus on the possible consequences of being bireversible or not. We first recall the following proposition. Proposition 5.13. [8, 14] Let M be an RI-automaton. If M does not contain a bireversible connected component, then P(M ) = ∅. From which we have the following proposition. Proposition 5.14. Let M = (Q, Σ, ·, ◦) be a non bireversible RI-automaton. Then ∗ ω Stab+ hM i (v ) 6= {1} for some v ∈ Σ . Furthermore, the set κ of singular points is not empty. Proof. The first statement follows by applying Proposition 5.13 and Corollary 5.8 to a non bireversible component of M . Let us prove the last claim of the proposition. ω We show that St is actually not continuous at v ω . Indeed, let π(u) ∈ Stab+ hM i (v ) with π(u) 6= 1, for some u ∈ Q∗ . If St were continuous at v ω then, by Lemma 4.3, π(u·v j ) = 1 for some j ≥ 1. Thus, by Lemma 3.1, we get π(u) = 1, a contradiction.  As an immediate consequence of the previous proposition we obtain the following result. Corollary 5.15. If there exists an RI-automaton M generating a group without singular points, then necessarily M is bireversible. 24 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO Let us have a look at the case of bireversible automata without singular points. Proposition 5.16. Let M = (Q, Σ, ·, ◦) be a bireversible automaton with hM i infinite and no singular points. Then for any u ∈ Q∗ the following are equivalent: (i) the Schreier graph centered at uω is finite; (ii) there is an integer ℓ > 0 such that π(uℓ ) = 1; (iii) Stab+ (uω ) 6= {1}. hd M i Proof. Recall that, for bireversible automata M , having no singular points is equivω ∗ alent to Stab+ hM i (v ) = {1} for all v ∈ Σ , by Propositions 5.7 and 4.11. (i)⇔(ii) follows from [8, Theorem 6]. e ∗ → hdM i is (iii)⇒(ii): Let v ∈ Σ∗ such that σ(v) ∈ Stab+ (v ω ), where σ : Σ hd M i the natural map. By Proposition 5.1 there is an integer ℓ > 0 such that v and uℓ ℓ ω commute, hence π(uℓ ) ∈ Stab+ hM i (v ) = {1}, i.e., π(u ) = 1. (i)⇒(iii): If the Schreier graph centered at uω is finite, then, for any v ∈ Σ∗ we get σ(v k ) ∈ StabhdM i (uω ) for k = [hdM i : StabhdM i (uω )]. Furthermore, σ(v k ) 6= 1 (uω ) 6= {1}. since, by Proposition 5.10, P(dM ) = ∅, whence Stab+  hdM i The equivalence (ii)⇐⇒(iii) of Prop. 5.16 links being torsion and having nontrivial stabilizers. We name it for future references in this paper: (TS) ℓ ω ∀u ∈ Σ∗ we have Stab+ hM i (u ) 6= {1} ⇐⇒ σ(u ) = 1 , for some ℓ ≥ 1 . We recall that an acyclic multidigraph is a multidigraph without cycles. We have the following geometrical description in terms of some algebraic conditions. Proposition 5.17. Let M be an RI-automaton. If P(M ) = ∅ and (TS) then the orbital graphs Γ(hM i, Q, Σω , v ω ) of periodic points v ω for v ∈ Σ∗ , are either finite or acyclic multidigraphs. Proof. Suppose that both P(M ) = ∅ and Condition (TS) hold. Let u ∈ Σ∗ . By [8, Lemma 2] it follows that every vertex of Γ(hM i, Q, Σω , uω ) is a periodic point hω ω for some h ∈ Σ∗ . Assume that there exists h ∈ Σ∗ such that Stab+ hM i (h ) 6= {1}, ω ω then by condition (TS) and [8, Theorem 6], Γ(hM i, Q, Σ , h ) is finite and so is Γ(hM i, Q, Σω , uω ). ω ω Otherwise, Stab+ hM i (p) = {1} for each vertex p of Γ(hM i, Q, Σ , u ). Thus, by condition P(M ) = ∅, we deduce that, for any vertex p of Γ(hM i, Q, Σω , uω ) there is v no cycle p−−→p for any v ∈ Q∗ , i.e., Γ(hM i, Q, Σω , uω ) is an acyclic multidigraph.  Note that the converse of the previous proposition holds if one assumes the existence of an infinite orbital graph rooted at some periodic point. Gathering Theorem 5.12 and the equivalence (TS), we obtain the following corollary, and so by Proposition 5.17 a description of the Schreier graphs of the dual in the case hM i has no singular points. Corollary 5.18. Let M be a bireversible automaton with hdM i infinite and no singular points in the dual: ∀ξ ∈ Qω , StabhdM i (ξ) = {1}. Then (TS) holds. We can now obtain a lower bound on the growth of the Schreier graphs pointed at periodic points. BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 25 Proposition 5.19. Let M be a bireversible automaton with hM i infinite and no singular points. Then for any u ∈ Q∗ with π(u) aperiodic, we have h i ∀m ≥ 1, hdM i : StabhdM i (um ) > log|Σ| (m). Proof. Fix some u ∈ Q∗ with π(u) aperiodic. We claim that for any v ∈ Σ∗ , (um ) r {1}, then m < |Σ||v| . For 0 ≤ i < m, we have if σ(v) ∈ Stab+ hdM i σ(ui ◦v) ∈ Stab+ (u). By the reversibility of M , there is an integer k ≤ |Σ||v| hdM i such that uk ◦v = v. If m ≥ |Σ||v| , then k ≤ m and σ(v) ∈ Stab+ (uω ), hdM i contradicting Corollary 5.18. Therefore, m < |Σ||v| . In particular, taking any a ∈ Σ, we get that there is an integer j ≤ [hdM i : StabhdM i (um )] such that σ(aj ) ∈ StabhdM i (um ), from which we obtain h i hdM i : StabhdM i (um ) ≥ j > log|Σ| (m) and this concludes the proof.  We obtain the following geometrical characterization. Theorem 5.20. Let M be a bireversible automaton with hM i infinite, with at least one aperiodic element. The following are equivalent. (i) StabhM i (ξ) is a torsion group, for all ξ ∈ Σω ; (ii) P(dM ) = ∅ and (TS) hold. (iii) the orbital graphs Γ(hdM i, Σ, Qω , v ω ) of periodic points are either finite or acyclic multidigraphs. Proof. (i)⇒(ii): From (i), the subgroup StabhM i (y ω ) is torsion for any y ∈ Σ∗ . Let  us prove P(dM ) = ∅. Assume hM i : StabhM i (y ω ) < ∞ for some y ∈ Σ∗ . Then, the subgroup \ g StabhM i (y ω )g −1 N= g∈hM i is a finite index torsion normal subgroup of hM i. Thus, since hM i/N isfinite, the  group hM i is torsion, a contradiction. We deduce hM i : StabhM i (y ω ) = ∞ for all y ∈ Σ∗ , whence by Proposition 5.9, P(dM ) = ∅. Now, to prove (TS) we essentially repeat the proof of the equivalence (iii)⇔(ii) of Proposition 5.16; the only point where the torsion hypothesis is used, is in the implication (iii)⇒(ii), while the other parts may be repeat “verbatim”. ω ∗ (ii)⇒(i): By Proposition 5.7 it is enough to prove that Stab+ hM i (y ), y ∈ Σ , are ω ∗ ∗ formed by torsion elements. Thus, let π(u) ∈ Stab+ hM i (y ) for some u ∈ Q , y ∈ Σ . n By Proposition 5.1, u, y is a commuting pair for some integer n ≥ 1, and by the same proposition σ(y n ) ∈ StabhdM i (uω ). Therefore, since P(dM ) = ∅, we find Stab+ (uω ) 6= {1}, whence π(uℓ ) = 1, by Proposition 5.16, i.e. π(u) is torsion. hdM i (ii)⇒(iii): Direct by Proposition 5.17. (iii)⇒(ii): Orbital graphs Γ(hdM i, Σ, Qω , v ω ), v ∈ Q∗ cannot be all finite. Indeed, if v ∈ Q∗ , the connected components of its powers have sizes bounded by the size of Γ(hdM i, Σ, Qω , v ω ). Thus, by [22, Proposition 7] π(v) is of finite order. And therefore, hM i is torsion, which contradicts the fact that hM i has an aperiodic element. Hence there is an infinite (acyclic by hypothesis) orbital graph Γ(hdM i, Σ, Qω , hω ), for some h ∈ Q∗ . Now, if P(dM ) 6= ∅, there is a non-trivial cycle hω −−→hω in 26 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO the previous acyclic multidigraph, which is impossible. Finally, let us show (TS): in fact, when the orbital graph is finite, then both sides of the equivalence are true, otherwise neither of the sides holds. Let u ∈ Q∗ . If Γ(hdM i, Σ, Qω , uω ) is finite, then clearly Stab+ (uω ) 6= {1} hdM i holds and, by the previous argument, π(u) is of finite order. On the other hand, if Γ(hdM i, Σ, Qω , uω ) is infinite, then π(u) is aperiodic according to [22, Proposition 7]. Further, by hypothesis, the infinite orbital graph Γ(hdM i, Σ, Qω , uω ) is acyclic, whence Stab+  (uω ) = {1}, and this concludes the proof. hdM i In this section we have constructed tools to find singular points, especially in the case of bireversible automata. It is still unknown whether there exists examples of infinite groups generated by RI automata without singular point. In particular we connected singular points with helix graphs and we proved that search can be narrow down to fully positive groups (Theorem 5.5). We also discuss the connection between the existence of non-trivial stabilizer of a group generated by a RI-automaton and the stabilizers of its dual (Theorem 5.12). This analysis shows that the possible existence of a RI automaton generating a group with all trivial stabilizers can be restricted to the class of bireversible. However evidences suggest that if G, generated by a RI automaton is infinite then it admits at least a non-trivial boundary. stabilizers. 6. Dynamics and Wang tilings There is an interesting connection between commuting pairs and Wang tilings observed by I. Bondarenko [5]. Given a Mealy automaton, one may associate a set of Wang tiles reflecting the action of the automaton on the stateset and the alphabet. The existence of periodic tilings corresponds to the notion of commuting words that generate elements of the stabilizers of infinite periodic words. Note that this problem – called the domino problem – is undecidable in general [4, 21]. In this section we prove that the domino problem is decidable for some family of tilesets linked to Mealy automata. On the other hand, we show that the problem of determining whether or not an automaton has commuting words on a restricted stateset is undecidable. 6.1. Wang tiles vs cross-diagrams. We recall that a Wang tile is a unit square tile with a color on each edge. Formally, it is a quadruple t = (tw , ts , te , tn ) ∈ C 4 where C is a finite set of colors (see Fig. 8 for a typical depiction of a Wang tile). A tileset is a finite set T of Wang tiles, and for each t ∈ T and d ∈ {n, s, e, w}, we tn tw te ts Figure 8. A Wang tile. put td for the color of the edge in the d-side. Given a tileset T , a tiling of the discrete plane is a map f : Z2 → T that associates to each point in the discrete plane a tile from T such that adjacent tiles share the same color on their common edge, BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 27 i.e., for any (x, y) ∈ Z2 , f (x, y)e = f (x + 1, y)w and f (x, y)n = f (x, y + 1)s . For a rectangle [x, x′ ] × [y, y ′ ] ⊆ Z2 , and d ∈ {n, s, e, w} we denote by f ([x, x′ ] × [y, y ′ ])d the word in C ∗ labelling the d-side of the square. For instance for d = w we have: f ([x, x′ ] × [y, y ′ ])w = f (x, y)w f (x, y + 1)w · · · f (x, y ′ )w . This notion extends naturally for rectangles of the form [x, ∞] × [y, ∞] ⊆ Z2 . A tiling f is periodic if there exists a periodicity vector v ∈ Z2 such that f (t + v) = f (t) for t ∈ Z2 . A tiling is bi-periodic if there are two linearly independent vectors v1 , v2 for which f is periodic. It is a well known fact belonging to the folklore that a tileset admits a periodic tiling if and only if it admits a bi-periodic tiling if and only if it admits vertical and horizontal periods. In case f : Z2 → T is a periodic tiling with a vertical period py and a horizontal period px , a fundamental domain of f is given by the square [0, px − 1] × [0, py − 1]. Following [23] we say that the tileset T is cd-deterministic with (c, d) ∈ {(e, n), (e, s), (w, n), (w, s)} if each tile t ∈ T is uniquely determined by its pair (tc , td ) of colors. Whenever T is cd-deterministic for each (c, d) ∈ {(e, n), (e, s), (w, n), (w, s)}, we say that T is 4-way deterministic. There is a natural way to associate to a Mealy automaton M = (Q, Σ, ·, ◦) a a|b tileset T (M ): for each transition q −−→p we associate the Wang tile (q, a, p, b) with colors on C = Q ⊔ Σ, see Figure 9. This point of view is just a reformulation of the cross-diagram defined previously (and is completely different from the one in [12]). The following lemma links properties of the automaton M with properties of the associated Wang tileset T (M ). v′ u v u v ′ ∈T ⇔ u′ ∈ M u ⇔ v|v ′ u−−→u′ ∈ M v′ Figure 9. Tileset T (M ) associated with an automaton M . Lemma 6.1. [5] The tileset associated to a Mealy automaton is necessarily wsdeterministic. Furthermore, for a Mealy automaton M we have the following. • T (M ) is es-deterministic if and only if M is reversible; • T (M ) is wn-deterministic if and only if M is invertible; • T (M ) is 4-way deterministic if and only if M is bireversible. There is a clear correspondence between tilings and cross diagrams, as illustrated in Figure 10. This easily extends to infinite words. The following proposition links periodicity of a tiling and existence of a commuting pair for a Mealy automaton. Proposition 6.2. Let M = (Q, Σ, ·, ◦) be an automaton. The following are equivalent. (i) T (M ) admits a periodic tiling; (ii) M admits a commuting pair: v ∈ Σ+ , u ∈ Q+ with u◦v = v, u·v = u. 28 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO ′ v1 ′ v2 uk ... .. . .. . .. . .. . v1 v2 ... u1 ... u1 u′k .. . vℓ v1 vℓ′ ... uk u′1 vℓ .. . .. . ⇔ u′1 v1′ u′k vℓ′ u1 · · · uk ·v1 · · · vℓ = v1′ · · · vℓ′ ⇔ u1 · · · uk ◦v1 · · · vℓ = u′1 · · · u′k Figure 10. Correspondence between partial tilings, diagrams and actions of the automaton. cross- Proof. If T (M ) admits a periodic tiling, then there is a bi-periodic tiling f : Z2 → T (M ) with horizontal and vertical periods px , py , respectively. Therefore, the fundamental domain [0, px − 1] × [0, py − 1] pinpoints a rectangle in the tiling f with colors given by the string f ([0, px − 1] × [0, py − 1])n = f ([0, px − 1] × [0, py − 1])s = v ∈ Σ∗ in both the south and north edge and f ([0, px − 1] × [0, py − 1])w = f ([0, px − 1] × [0, py − 1])e = u ∈ Q∗ in both the west and east edge. Hence, by Figure 10: u◦v = v, u·v = u. On the other hand, if there is a commuting pair u ∈ Q∗ , v ∈ Σ∗ such that u◦v = v, u·v = u then by Figure 10 there is a partial tiling f : [0, |v| − 1] × [0, |u| − 1] → T (M ) with f ([0, |u| − 1] × [0, |v| − 1])n = f ([0, |u| − 1] × [0, |v| − 1])s = u, f ([0, |u| − 1] × [0, |v| − 1])w = f ([0, |u| − 1] × [0, |v| − 1])e = v. This tiling extends naturally to a bi-periodic tiling f : Z2 → T (M ) of the whole discrete plane.  As a consequence we obtain a decidability result for the domino problem with particular sets of tilesets. Starting from the tileset T we construct a letter-to-letter transducer M (T ) = (Q, Σ·, ◦) in a natural way as follows: Let Q = ∪t∈T {tw , te }, Σ = ∪t∈T {ts , tn }, and ts |tn transitions given by tw −−→te ∈ M (T ) whenever (tw , ts , te , tn ) ∈ T . Note that in general M (T ) is not a Mealy automaton – we will deal with these cases in the next section. However when it is indeed a Mealy automaton the following holds: Corollary 6.3. Let T be a Wang tileset. If M (T ) = (Q, Σ, ·, ◦) is a Mealy automaton then T tiles the plane periodically, in an effective way. BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 29 Proof. It is decidable if M (T ) is a Mealy automaton in time O(|Q||Σ|). Now construct the helix graph H1,1 (once again in time O(|Q||Σ|)). Since the stateset and the alphabet are finite, and each pair (u, v) ∈ H1,1 has exactly one successor, there exists a cycle (u0 , v0 )−−→(u1 , v1 )−−→ · · · −−→(um , vm )−−→(u0 , v0 ) in H1,1 . Hence by Lemma 5.2 there is a commuting pair (u, v) ∈ Q+ × Σ+ , so by Proposition 6.2, T tiles periodically the plane.  6.2. Commuting pairs on a restricted stateset. With Corollary 6.3, we have seen that each tileset T (M ) associated with a Mealy automaton M = (Q, Σ, ·, ◦) admits a periodic tiling. However (u, v) could commute because the word u acts like the identity, i.e. π(u) = 1; so this commuting pair will not help us to find singular points, and we may want to avoid it. In this spirit we now consider commuting pair on a restricted stateset, that is commuting pair in Q′∗ × Σ∗ , with Q′ Q. We now consider an automaton M = (Q, Σ, ·, ◦) and a subset Q′ Q. We consider the restricted tileset T (M , Q′ ) formed by the tiles of T (M ) whose colors are in the set Q′ ∪Σ. Note that T (M , Q′ ) is the tileset associated with the (partial) automaton obtained from M by erasing all the transitions to or from states in Q r Q′ . In particular when the automaton has a sink-state1 e, we define non-elementary commuting pairs, that are commuting pairs restricted to the stateset Q r {e}, and the non-elementary tileset T (M , Q r {e}) will be denoted by T (M ). We have the following proposition. Proposition 6.4. Let M be a Mealy automaton with a sink-state. Then M admits a non-elementary commuting pair if and only if T (M ) has a periodic tiling. Proof. The proof is similar to the one presented in Proposition 6.2. Indeed if e were to appear in the cross diagramm, then since it a sink it will stay present every step after, preventing the tiling to by periodic: ∀u ∈ Q∗ eQ∗ , ∀v ∈ Σ∗ , u·v ∈ Q∗ eQ∗ .  This raises the problem of finding restricted commuting pairs, since we can no longer apply Corollary 6.3. We obtain the following decision problem: RESTRICTED COMMUTING PAIRS: • Input: M = (Q, Σ, ·, ◦), Q′ Q. • Output: Does M have commuting pairs restricted to the stateset Q′ ? It turns out that this problem is undecidable. To prove this we are going to reduce a known undecidable problem to it, namely the existence of a periodic tiling for a 4-way deterministic tileset (see [23]). For the reduction from the periodic 4-way deterministic problem to our problem we show that, given a 4-way deterministic tileset, a periodic tiling can be determined from a non-trivial commuting pair of some Mealy automaton. 1One can recognize trivial states in linear time (note that we can also find trivial components through minimization, with cost O( |Σ||Q| log |Q| ), see [3]). In the following we assume that the automaton has at most one sink-state, which we denote e. 30 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO a 1 1|2 2|3 a 1 2 b 1|2 2|3 3 =⇒ b 3 a b 1 a a b 2 b 2|2 3|1 2 a =⇒ 3|1 1|1 2|2 3|3 b 2|2 3|1 1|3 e Figure 11. Construction of an invertible Mealy automaton from a 4-way deterministic tileset. Let T be a 4-way deterministic tileset. We build M (T ) as in Corollary 6.3. This transducer is not in general complete, however, we may add an extra sink-state e and extra transitions to this sink-state in order to make the automaton complete and invertible. We call MS (T ) the invertible automaton associated with T . Since the extra transitions go to the sink-state it is clear that T = T (MS (T )). We are now in position to prove the following result. Theorem 6.5. The problem RESTRICTED COMMUTING PAIRS is undecidable. Proof. We prove that a particular instance of RESTRICTED COMMUTING PAIRS is undecidable: NON-ELEMENTARY COMMUTING PAIRS: • Input: M a Mealy automaton with a sink-state. • Output: Does M have non-elementary commuting pairs? By [23] checking whether a 4-way deterministic tileset admits a periodic tiling is undecidable. By Proposition 6.4 checking whether T admits a periodic tiling is equivalent to check whether MS (T ) has a non-elementary commuting pair, hence NON-ELEMENTARY COMMUTING PAIRS is undecidable, and so is RESTRICTED COMMUTING PAIRS.  Note that in this context we deal with inverses of Q, and it is not difficult to e ∗ may generate non-elementary commuting pair. check that reduced words u ∈ Q Hence, in this context a non-elementary pair of commuting words is a pair of words ∗ u ∈ Q^ r {e} , v ∈ Σ∗ such that u is reduced and u, v commute. Proposition 6.6. With the above notation. Let M ∈ Sa . The following are equivalent. (i) v ω is a singular point, v ∈ Σ∗ ; ∗ (ii) there is a non-elementary commuting pair u ∈ Q^ r {e} , v ∈ Σ∗ with π(u) 6= 1. Proof. (i)⇒(ii) If v ω ∈ κ then the map St is not continuous at vω if and only if ∗ StabhM i (v ω ) 6= Stab0hM i (v ω ). This implies that there exists u ∈ Q^ r {e} satisfy0 ω ω k ing π(u) ∈ StabhM i (v ) r StabhM i (v ). Note that π(u·v ) 6= 1 for any k ≥ 0 (from Lemma 4.2). Since |u| is finite, there exist n > m > 0 such that u·v m = u·v n . Then ∗ um := u·v m and v n−m form a non-elementary commuting pair with um ∈ Q^ r {e} . ∗ (ii)⇒(i) Let u ∈ Q^ r {e} , v ∈ Σ∗ be a non-elementary commuting pair. Since BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 31 π(u) 6= 1, there is a word w ∈ Σ∗ such that u◦w 6= w. Consider the sequence wn := v n wv ω and proceed as in the proof of Lemma 4.3 to show that π(u) ∈ StabhM i (v ω ) r Stab0hM i (v ω ).  Remark 6.7. Note that Proposition 4.11 implies that Proposition 6.6 holds also for bireversible automata. As we have already noted, the tileset T (M ) associated with a Mealy automaton M always admits a periodic tiling, however, in case we consider automata from Sa , it is interesting, and useful for the sequel, to understand when the nonelementary tileset of T (M ) tiles the discrete plane. We need first the following lemma regarding inverse X-digraphs, that are digraph Γ with edges labelled by a a−1 element of X, such that if p−−→q is an edge of Γ, so is q −−→p (involutive), for e there is exactly one edge starting from v any vertex v ∈ Γ and any label x ∈ X, and labelled with x. In inverse digraph the natural quasimetric d is symmetric and so it is a distance. e Note that an inverse X-digraph has in- and out-degree |X|. Lemma 6.8. Let Γ be an infinite connected inverse X-digraph, |X| < ∞, and let e ω such that v be a vertex of Γ. Then there is a right-infinite word θ ∈ X lim d(v, vi ) = ∞ , i→∞ θ[:i] where vi is defined by v −−→vi . Proof. Since Γ is an inverse digraph it is symmetric. Consider Γ the non-directed a a−1 graph obtained by gluing edges p−−→q and q −−→p in Γ. Note that the distances in Γ and in Γ are the same and we will denote both of them by d(., .). Since Γ has finite degree, is infinite and connected, we can apply König’s lemma that claims that Γ contains an infinitely long simple path (that is, a path with no repeated vertices). Let ν = v1 · · · vi · · · denote such a path and θ the associated infinite word. Since Γ is connected there exists a path from any vertex v to v1 : let k = d(v, v1 ). Since the graph Γ has bounded degree, the ball Bk (v) of radius k centred in v is finite and since ν is simple there is an index ik for which vj ∈ / Bk (v) for j ≥ ik . The same holds for arbitrary large k, which concludes the proof.  Consider the map λ from the set of edges of Sch(ξ) to the integers, that associate q to each edge f = η −−→η ′ the integer ( min{n : q·η[: n] = e} if ∃m : q·η[: m] = e λ(f ) = ∞ otherwise. Let d(η, η ′ ) denote the usual metric on Sch(ξ). Since Sch(ξ) is a regular digraph e for each integer n we consider the set of edges that admit with finite out-degree Q at least one vertex inside the ball of radius n centered at ξ: q Bξ→(n) = {η −−→η ′ : d(ξ, η) ≤ n ∨ d(ξ, η ′ ) ≤ n} . Note that Bξ→(n) is clearly finite and we may define  ψξ (n) = max λ(f ) : f ∈ Bξ→(n) ∧ λ(f ) < ∞ . n o We remark that λ(f ) : f ∈ Bξ→(n) ∧ λ(f ) < ∞ is non-empty in view of the exise tence of the trivial edges of type η −−→η. Thus, ψξ is a monotonically increasing 32 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO function. We have thus two cases: either limn→∞ ψξ (n) = ∞, or limn→∞ ψξ (n) = ℓ for some ℓ ≥ 1. The following proposition provides sufficient conditions on the dynamics of the boundary for the non-elementary tileset to tile the discrete plane. Proposition 6.9. Let M = (Q, Σ, ·, ◦) ∈ Sa . If there is a point ξ ∈ Σω whose Schreier graph Sch(ξ) is infinite and such that limn→∞ ψξ (n) = ℓ for some integer ℓ ≥ 1, then T (M ) tiles the discrete plane. Conversely, if T (M ) admits just aperiodic tilings, then there is an infinite Schreier graph Sch(ξ) for some ξ ∈ Σω . Proof. It belongs to the folklore that a tileset tiles the whole discrete plane if and only if it tiles the first quadrant of the discrete plane (see for instance [28]). Let us prove that there is a tiling f : [1, ∞] × [1, ∞] → T (M ) of the first quadrant. First there is a tiling h : [1, ∞] × [1, ∞] → T (M ) of the first quadrant such that, putting η = h ([1, ∞] × [1, ∞])s and σ = h ([1, ∞] × [1, ∞])w , we have: (1) σ is not cofinal with eω , and (2) if σ[j] 6= e, for some j ≥ 1, then σ[j]·(σ[: j − 1]◦η[: i]) 6= e for all i ≥ 1. e σ θ words eω or e free ξ η ℓ Figure 12. Construction of a non-elementary tilling when limn→∞ ψξ (n) = ℓ. Intuitively, the last property means that if the west edge of the leftmost tile of a band is not colored by e, then also none of the west and east edges of all in this band are colored by e. Indeed, since e is a sink, if the west edge of the leftmost tile of a band is colored by e, then clearly all the west and east edges of the all the tiles in this band are colored by e as well, see Fig. 12. Let us first find a tiling fulfilling Condition 2: If there exists an element q ∈ Q of infinite order, take q ω as σ and some word of infinite orbit under the action of q as η. Otherwise, since Sch(ξ) is infinite we may find, by Lemma 6.8, a right-infinite path starting at ξ and e ω such that limi→∞ d(ξ, θ[: i]◦ξ) = ∞. labelled by some right-infinite word θ ∈ Q Since every element of Q has finite order we may remplace each negative occurence of a letter by the suitable power of this letter, hence without loss of generality we can suppose θ ∈ Qω . By Figure 10 consider the tiling h : [1, ∞] × [1, ∞] → T (M ) associated with ξ, θ, i.e., such that ξ = h ([1, ∞] × [1, ∞])s , θ = h ([1, ∞] × [1, ∞])w . BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 33 Since limn→∞ ψξ (n) = ℓ, then for any vertex ξ ′ of Sch(ξ) and q ∈ Q the following property: (2) if q·ξ ′ [: i] 6= e for all 1 ≤ i ≤ ℓ, then q·ξ ′ [: i] 6= e for all i ≥ ℓ holds. Since e is a sink-state, Equation (2) can be written equivalently: (2) if q·ξ ′ [: ℓ] 6= e, then q·ξ ′ [: i] 6= e for all i ≥ ℓ . Let us restrict the tiling h to the quadrant [ℓ, ∞] × [1, ∞]. Take η = h([ℓ, ∞] × [1, ∞])s = ξ[ℓ :] , σ = h([ℓ, ∞] × [1, ∞])w = θ·ξ[: ℓ − 1] . Property 2 is obtained from Equation (2). Let us prove now Property 1. Two elements ξ ′ , ξ ′′ ∈ Σω are m-cofinal if ξ ′ [m :] = ξ ′′ [m :]. Note that for fixed ξ ′ ∈ Σω and m ≥ 1, the number of vertices ξ ′′ that are m-cofinal with ξ ′ is finite. Suppose that σ is cofinal with eω : σ = σ[: k]eω for some k ≥ 1. Hence the sequence (σ[: i]◦η)i is ultimately constant. But σ[: i]◦η is the suffix of θ[: i]◦ξ, whence the set {θ[: k + j]◦ξ : j ≥ 1} formed ℓ-cofinal vertices is finite. Which contradicts limi→∞ d(ξ, θ[: i]◦ξ) = ∞, and σ cannot be cofinite with eω . Whence we have obtained a tiling of the first quadrant with σ and η as west and south borders respectively, and such that, for all i, either σ[i]·η = eω or σ[i]·η contains no e. We now obtain a tiling in T (M ) by deleting those line eω . Let us prove the converse in our proposition. Let ξ = f ([1, ∞] × [1, ∞])s , θ = f ([1, ∞] × [1, ∞])w where f : Z × Z → T (M ) is an aperiodic tiling. Since f is aperiodic, we get θ[: i]◦ξ 6= θ[: j]◦ξ for i 6= j (since periodicity in one direction is equivalent to periodicity in both directions). Therefore, {θ[: i]◦ξ : i ≥ 0} is infinite and Sch(ξ) as well.  We recall that an automaton M = (Q, Σ, ·) is synchronizing whenever there is a word w ∈ Σ∗ such that q·w = p·w for any q, p ∈ Q. A Mealy automaton M = (Q, Σ, ·, ◦) is synchronizing whenever (Q, Σ, ·) is synchronizing. The set of synchronizing words is denoted by Syn(M ). Note that an automaton in Sa is always synchronizing. For further details on synchronizing automata and some connections with automaton groups see [9]. The next proposition characterizes automata in Sa whose set of reduced tiles does not tile the discrete plane in terms of a stability property regarding synchronization. Consider the prefix-closed language N SynQ (v) = {u ∈ Q<|v|−1 | ∀j ≤ |u|, u[: j]◦v ∈ / Syn(M )} . and the property Maximal non Synchronizing : there exists an integer m such that, for all v ∈ Σm , any u maximum in N SynQ (v) for the prefix relation satisfies uq◦v ∈ Syn(M ), for all q ∈ Q. Proposition 6.10. Let M = (Q, Σ, ·, ◦) ∈ Sa . Then T (M ) does not tile the discrete plane if and only if property Maximal non Synchronizing holds for M . 34 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO Proof. Suppose that T (M ) does not tile the discrete plane. Hence there exists an integer m such that T (M ) cannot tile a square of size m. Let v ∈ Σm and u ∈ N SynQ (v). First suppose |u| < m − 2: if for some q ∈ Q we have uq◦v ∈ / Syn(M ) then uq ∈ N SynQ (v), hence u cannot be maximal. Now suppose |u| = m − 2: if uq◦v ∈ / Syn(M ), for some q ∈ Q, then there is a q ′ ∈ Q ′ such that q ·(uq◦v)[: j] 6= e for all j ≤ m. Hence, by Figure 10 there is a tiling of the square of size m associated with the two words v, uqq ′ , a contradiction. Conversely, if T (M ) tiles a square f : [1, m] × [1, m] → T (M ), then if we put h = f ([1, m] × [1, m])s , v = f ([1, m] × [1, m])w we get that v[: j]◦h ∈ / Syn(M ) for all j ≤ m − 1. In particular, u = v[: m − 2] is maximal in N SynQ (v). However, (uv[m − 1])◦h = v[: m − 1]◦h ∈ / Syn(M ), hence the condition of the statement is not satisfied.  6.3. Aperiodic tilings and singular points. In Subsection 5.2 we have seen that RI-automata, with empty set of singular points are necessarily bireversible with all trivial stabilizers in the boundary (see Corollary 5.15 and Proposition 4.11). Furthermore, Corollary 5.8 implies that if a bireversible automaton M has no positive relations, then it must have a non-trivial stabilizer in the boundary. In particular, for the class of bireversible automata it is not possible to have simultaneously no positive relations and no singular points. This fact no longer holds in the class Sa : there exist automata in Sa without positive relations and with no singular points. However, we need some precaution in defining the set of “positive relations” for an automaton M ∈ Sa . Indeed, for such an automaton, the set P(M ) is always non-empty since the sink-state e acts like the identity. In order to characterize tilings without non-trivial commuting pairs, we need to define a special set of relations that are in some sense non elementary. Such relations intuitively correspond to words with the property that some of their restrictions do not become the trivial word. For w ∈ Q∗ , |w|e denotes the number of occurrences of the letter e in w. The set E(M ) of the non-elementary relations is defined by: E(M ) = {u ∈ (Q r {e})∗ : π(u) = 1, ∃v ∈ Σ∗ , |u·v n |e < |u|, ∀n ≥ 1} . By a compactness argument note that the complement of the non-elementary relations, the set of elementary relations, may be described as the set of words u ∈ (Q r {e})∗ with π(u) = 1, such that there exists some n ≥ 1 for which u·v = e|v| for every v ∈ Σ≥n . Geometrically, automata in the class Sa with no singular points and whose eventual relations are elementary relations, possess helix graphs with a particular shape, as proved in Proposition 6.11. We say that the helix graph Hk,n of an automaton M ∈ Sa is singular whenever each connected component of Hk,n has a unique cycle which is necessarily of the form (ek , v)−−→(ek , v) for some v ∈ Σn . Proposition 6.11. Let M ∈ Sa . The following are equivalent: (i) for any ξ ∈ Σω , g ∈ Stab+ hM i(ξ), there is n ≥ 1 with g·ξ[: n] = 1 and E(M ) = ∅; ω n (ii) for any v ∈ Σ∗ , g ∈ Stab+ hM i(v ), there is n ≥ 1 with g·v = 1 and E(M ) = ∅; (iii) for any k, n ≥ 1, the helix graph Hk,n is singular; (iv) there is no non-elementary pair of commuting words. BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 35 Proof. (i)⇒(ii). Trivial. (ii)⇒(iii). Suppose that Hk,n is not singular for some k, n ≥ 1. Therefore, by Lemma 5.2 there is a commuting pair u ∈ Q∗ r {e}∗, v ∈ Σ∗ . As e acts like the identity, by erasing the (potential) occurrences of e in u, we obtain a word u′ ∈ ω (Q r {e})∗ such that u′ , v commutes. Hence Lemma 5.2 implies π(u′ ) ∈ Stab+ G (v ). ′ ′ ′ If π(u ) = 1, then, since u , v commutes, u ∈ E(M ) 6= ∅. Otherwise, we have π(u′ )·v n = π(u′ ) 6= 1 for all n ≥ 1. (iii)⇒(iv). If there is a non-elementary commuting pair u ∈ (Q r {e})∗ , v ∈ Σ∗ , then by Lemma 5.2 the helix graph H|u|,|v| contains the loop (u, v)−−→(u, v), i.e., H|u|,|v| is not singular. (iv)⇒(i). If E(M ) 6= ∅, then for any u ∈ E(M ), by definition, there exists v ∈ Σ∗ satisfying |u·v n |e < |u|, ∀n ≥ 1. By a compactness argument, there exists some m ≥ 1 with |u·v m |e = |u·v m+k |e for every k ≥ 0. Moreover, there exist indices i > j ≥ m satisfying u·v j = u·v i . As e acts like the identity, by erasing the (potential) occurrences of e in u, we obtain a word u′ ∈ (Qr{e})∗ such that u′ , v i−j is a non-elementary pair of commuting words. ω Thus, we may assume E(M ) = ∅. Now suppose g ∈ Stab+ hM i (ξ) for some ξ ∈ Σ ∗ with g·ξ[: n] 6= 1 for all n ≥ 1. Hence, there are some word u ∈ (Qr{e}) such that π(u) = g and some indices i > j ≥ 1 such that if we put u′ = u·ξ[: j] ∈ (Q r {e})∗ , then u′ , ξ[j + 1 : i] is a non-elementary commuting pair.  Note that the previous proposition provides necessary conditions on M ∈ Sa for tilings of T (M ) of the discrete plane to be aperiodic. As a result of Theorem 6.5 we immediately obtain the following undecidability result of checking the previous “continuity” condition. Theorem 6.12. Given an automaton M ∈ Sa , it is undecidable whether for any ξ ∈ Σω , g ∈ Stab+ G (ξ), there exists n ≥ 1 with g·ξ[: n] = 1 and E(M ) = ∅. Moreover, by Proposition 6.11, taking the aperiodic 4-way deterministic tileset T described in [21] and the associated automaton M ∈ Sa with T = T (M ) we get that M actually satisfies the “continuity” conditions described in Proposition 6.11. The same paper raised the problem to determining the existence of an aperiodic reflection-closed tileset. A tileset T with colored oriented edges is closed under reflection if for each tile in T the reflection of this tile along a horizontal or vertical line also belongs to T . Kari and Papasoglu considered the following tiling rule: a tiling of the plane, using tiles from a tileset T which is closed under reflection, is said to be valid if two adjacent tiles meet along an edge with the same color and orientation and two tiles that are the reflection of each other are never adjacent. In the following, such a tiling will be called of Kari-Papasoglu type. If we consider only the horizontal (vertical) symmetry we say that T is hreflection-closed (respectively, v-reflection-closed) tileset. Note that if T is hreflection-closed, then it is ws-deterministic (es-deterministic) if and only if it is wndeterministic (respectively, en-deterministic). Similarly, if T is v-reflection-closed, then it is ws-deterministic (wn-deterministic) if and only if it is es-deterministic (respectively, en-deterministic). Hence, if T is reflection-closed and xy-deterministic for some (x, y) ∈ {(e, n), (e, s), (w, n), (w, s)}, then T is necessarily 4-way deterministic. In [21] the authors raised the problem of finding a 4-way deterministic tileset which is valid, aperiodic and reflection-closed. Such a tileset would give an 36 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO example of a CAT(0) complex whose fundamental group is not hyperbolic and does not contain a subgroup isomorphic to Z2 , see [19, 21]. In this setting we can prove a statement analogous to Proposition 6.11. We will prove that the search for aperiodic h-reflection-closed tilesets that are ws- and wn-deterministic is related to the search for automata in the class Sa whose set κ of singular points is empty and which are elementary-free, in the following sense. First we need an analogous to Proposition 6.10 for h-reflexion-closed tilings (resp. 4-way deterministic tilings). Let us define the prefix-reduced relation: we say that e is smaller than u′ ∈ Q e for the prefix-reduced relation if u ≤p u′ holds and both u∈Q u and u′ are reduced. We define a property that will serve for the characterization of Kari-Papasoglu type tilings: an automaton M = (Q, Σ, ·, ◦) ∈ Sa satisfies property h-Maximal non Synchronizing (resp. 4-way Maximal non Synchronizing ) if m ), any g there exists an integer m such that, for all v ∈ Σm (resp. reduced v ∈ Σ ∗ f maximum in N Syn e (v) for the prefix-reduced relation satisfies reduced u ∈ Q Q e uq◦v ∈ Syn(M ) for all q ∈ Q. Proposition 6.13. Let M = (Q, Σ, ·, ◦) ∈ Sa . Then T admits an h-reflexionclosed tiling (4-way deterministic tiling) if and only if property h-Maximal non Synchronizing (resp. 4-way Maximal non Synchronizing ) does not hold. Proof. Similar to the proof of Proposition 6.10, but avoiding patterns xx−1 that are not allowed in the Kari-Papasoglu type tilings.  A group generated by an automaton in the class Sa is said to be elementary-free if the only relations that it contains are words whose restrictions become eventually all trivial, i.e. the set of its relations may be described as the set of words u ∈ ∗ (Q^ r {e}) with π(u) = 1, such that there exists an n ≥ 1 for which u·v = ekv for every v ∈ Σ≥n and some integer kv such that |kv | ≤ |v|. ek,n of an automaton M ∈ Sa is In this context, we say that a helix graph H strongly-singular whenever any cycle (u, v)−−→(u′ , v ′ )−−→ · · · −−→(u, v) implies ∗ g or π(u) = 1. We have the following proposition analogous to Propoeither u ∈ {e} sition 6.11. Proposition 6.14. Let M ∈ Sa . The following are equivalent: (i) hM i is elementary-free and the set κ of singular points is empty; (ii) hM i is elementary-free and for any v ∈ Σ∗ , g ∈ StabhM i (v ω ), there is n ≥ 1 with g·v ω [: n] = 1; ek,n is strongly-singular; (iii) for any k, n ≥ 1, the helix graph H ∗ (iv) there is no non-elementary pair u ∈ (Q^ r {e}) , v ∈ Σ∗ of commuting words. Proof. (i)Equivalence ⇔(ii) follows from Lemma 4.2. Equivalence (ii)⇔(iii) may be proven in an analogous way as in Proposition 6.11. Equivalence (i)⇔(iv) is a consequence of Proposition 6.6.  e ∗ , in what follows we consider this acNote that hM i also acts naturally on Σ tion. Following [8], we say that a point ξ ∈ Σω is essentially non-trivial when lim |ξ[: n]| →n→∞ +∞. Moreover we say that a helix graph is essentially-singular BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 37 helix whenever, if (u, v)−−→(u′ , v ′ )−−→ · · · −−→(u, v) is a cycle, then v ω is es∗ sentially non-trivial and either u ∈ g {e} , or π(u) = 1. We have the following proposition analogous to Proposition 6.11: Proposition 6.15. Let M ∈ Sa . The following are equivalent: (i) hM i is elementary-free and the set of essentially non-trivial singular points is empty; f∗ such that v ω is essentially (ii) hM i is elementary-free and for any v ∈ Σ ω non-trivial, g ∈ StabhM i (v ), there is n ≥ 1 with g·v ω [: n] = 1; ek,n is essentially singular; (iii) for any k, n ≥ 1, the helix graph H ∗ f∗ of commuting (iv) there is no non-elementary pair u ∈ (Q^ r {e}) , v ∈ Σ words. Proof. Similar to the proof of Prop. 6.14.  Putting together all the previous results we may characterize aperiodic h-reflectionclosed tilesets (resp. h-reflection-closed tilesets). Theorem 6.16. With the above notation. The following are equivalent: (i) there is a ws- and wn-deterministic tileset T which is h-reflection-closed (resp. a 4-way deterministic which is h- and v-reflection-closed) that tiles the discrete plane with aperiodic tilings of Kari-Papasoglu type; (ii) there is an automaton M ∈ Sa such that hM i is elementary-free, the set of singular points (resp. essentially non-trivial singular points) is empty, and, in the automaton B obtained from M ⊔ M −1 identifying the two sinks e and e−1 , property h-Maximal non Synchronizing (resp. 4-way Maximal non Synchronizing ) does not hold. Proof. If T is h-reflection-closed, we may fix a direction and divide the colors of the vertical edges of T into two distinct and disjoint sets Q ⊔ Q−1 , while we put for Σ the set of colors of the horizontal edges of the tiles in T . Note that, for e a, b ∈ Σ, the corresponding horizontally reflected tile each tile (q, a, p, b), q, p ∈ Q, is (q −1 , a, p−1 , b). The partition Q ⊔ Q−1 induces a partition T + ⊔ T − on T in the obvious way. Consider an associated automaton M ∈ Sa such that T (M ) = T + (as in Section 5). Note that T (M ⊔ M −1 ) = T . Conversely, to any automaton M ∈ Sa , the tileset T (M ⊔ M −1 ) is ws- and wn-deterministic, and it is h-reflection-closed. By an argument very similar to the proof of Propositions 6.2 and 6.4, it is not difficult to see that, in the previous cor∗ respondence, M has a non-elementary pair u ∈ (Q^ r {e}) , v ∈ Σ∗ of commuting words where u is non-trivial and reduced if and only if the corresponding tileset T (M ⊔ M −1 ) admits a periodic tiling (in the sense of Kari-Papasoglu). Hence, the equivalence in the statement follows from Proposition 6.13 for the existence of a tiling of Kari-Papasoglu type and Proposition 6.14 for its aperiodicity. The proof for the 4-way case is similar and uses Proposition 6.15.  The last theorem gives a characterization of specific Wang tilings in the language of Mealy automata. This is another motivation to further explore this connection. 38 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO 7. Some open problems Problem 1. Let M ∈ Sa be an automaton generating a free group. Is it always the case that there is a point in the boundary whose Schreier graph is infinite? Problem 2. Given that M is minimized, can two singular points have isomorphic infinite Schreier Graphs? Problem 3. Given a Mealy automaton M , is it decidable whether P(M ) = ∅? Problem 4. Are there interesting classes of automata where the non-elementary commuting pair is decidable? References [1] A. Akhavi, I. Klimann, S. Lombardy, J. Mairesse, and M. Picantin. On the finiteness problem for automaton (semi)groups. International Journal of Algebra and Computation, 22(6):1–26, 2012. [2] L. Bartholdi and V. Nekrashevych. Thurston equivalence of topological polynomials. Acta Mathematica, 197(1):1–51, 2006. [3] L. Bartholdi and P. V. Silva. Groups defined by automata. In J.-É. Pin, editor, AutoMathA Handbook. European Mathematical Society, to appear. http://arxiv.org/abs/1012.1531. [4] R. Berger. The undecidability of the domino problem. Memoirs of the American Mathematical Society, 66:72, 1966. [5] I. Bondarenko. private communication. [6] I. Bondarenko, D. D’Angeli, and T. Nagnibeda. Ends of schreier graphs and cut-points of limit spaces of self-similar groups. http://arxiv.org/abs/1601.07587. [7] I. Bondarenko, D. D’Angeli, and E. Rodaro. The lamplighter group z3 ≀ z generated by a bireversible automaton. Communications in Algebra, to appear. http://arxiv.org/abs/1502.07981. [8] D. D’Angeli and E. Rodaro. Freeness of automaton groups vs boundary dynamics. http://arxiv.org/abs/1502.07981. [9] D. D’Angeli and E. Rodaro. Groups and semigroups defined by colorings of synchronizing automata. International Journal of Algebra and Computation, 24(6):773–793, 2014. [10] D. D’Angeli and E. Rodaro. A geometric approach to (semi)-groups defined by automata via dual transducers. Geometriae Dedicata, 174:375–400, 2015. [11] N. Davis, M. Elder, and L. Reeves. Non-contracting groups generated by (3, 2)-automata. Algebra and Discrete Mathematics, 17(1):20–32, 2014. [12] P. Gillibert. The finiteness problem for automaton semigroups is undecidable. International Journal of Algebra and Computation, 24-1:1–9, 2014. [13] Y. Glasner and S. Mozes. Automata and square complexes. Geometriae Dedicata, 111:43–64, 2005. [14] Th. Godin, I. Klimann, and M. Picantin. On torsion-free semigroups generated by invertible reversible Mealy automata. In Language and automata theory and applications, volume 8977 of Lecture Notes in Comput. Sci., pages 328–339. Springer, Cham, 2015. [15] R. I. Grigorchuk. Some topics of the dynamics of group actions on rooted trees. The Proceedings of the Steklov Institute of Math., 273:1–118, 2011. [16] R. I. Grigorchuk and D. Savchuk. Self-similar groups acting essentially freely on the boundary of the binary rooted tree. In Group theory, combinatorics, and computing, volume 611 of Contemp. Math., pages 9–48. Amer. Math. Soc., Providence, RI, 2014. [17] R. I. Grigorchuk and Z. Šuniḱ. Asymptotic aspects of Schreier graphs and Hanoi Towers groups. Comptes Rendus Mathématique. Académie des Sciences. Paris, 342(8):545–550, 2006. [18] R. I. Grigorchuk and A. Żuk. On a torsion-free weakly branch group defined by a three state automaton. International Journal of Algebra and Computation, 12(1-2):223–246, 2002. International Conference on Geometric and Combinatorial Methods in Group Theory and Semigroup Theory (Lincoln, NE, 2000). [19] M. Gromov. Hyperbolic manifolds, groups and actions. In Riemann surfaces and related topics: Proceedings of the 1978 Stony Brook Conference (State Univ. New York, Stony BOUNDARY ACTION OF AUTOM. GROUPS W/O SINGULAR POINTS & WANG TILINGS 39 [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] Brook, N.Y., 1978), volume 97 of Ann. of Math. Stud., pages 183–213. Princeton Univ. Press, Princeton, N.J., 1981. M. Kambites, P. V. Silva, and B. Steinberg. The spectra of lamplighter groups and Cayley machines. Geometriae Dedicata, 120:193–227, 2006. J. Kari and P. Papasoglu. Deterministic aperiodic tile sets. Geometric and Functional Analysis, 9(2):353–369, 1999. I. Klimann, M. Picantin, and D. Savchuk. A connected 3-state reversible mealy automaton cannot generate an infinite burnside group. In Developments in Language Theory - 19th International Conference, DLT 2015, Liverpool, UK, July 27-30, 2015, Proceedings., pages 313–325, 2015. B. Le Gloannec. The 4-way deterministic periodic domino problem is undecidable. https://hal.archives-ouvertes.fr/hal-00985482, 2014. O. Macedońska, V. Nekrashevych, and V. Sushchansky. Commensurators of groups and reversible automata. Dopovı̄dı̄ Natsı̄onal′ noı̈ Akademı̄ı̈ Nauk Ukraı̈ni. Matematika. Prirodoznavstvo. Tekhnı̄chnı̄ Nauki, 12:36–39, 2000. V. Nekrashevych. Self-similar groups, volume 117 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2005. V. Nekrashevych. Free subgroups in groups acting on rooted trees. Groups, Geometry, and Dynamics, 4(4):847–862, 2010. D. Perrin and J.-É. Pin. Infinite words : automata, semigroups, logic and games. Pure and applied mathematics. Academic, London, San Diego (Calif.), 2004. R. M. Robinson. Undecidability and nonperiodicity for tilings of the plane. Invent. Math., 12:177–209, 1971. D. Savchuk and S. Sidki. Affine automorphisms of rooted trees. Geometriae Dedicata, to appear. http://arxiv.org/abs/1502.07981. D. Savchuk and Y. Vorobets. Automata generating free products of groups of order 2. Journal of Algebra, 336:53–66, 2011. S. Sidki. Automorphisms of one-rooted trees: growth, circuit structure, and acyclicity. Journal of Mathematical Sciences (New York), 100(1):1925–1943, 2000. Algebra, 12. B. Steinberg, M. Vorobets, and Y. Vorobets. Automata over a binary alphabet generating free groups of even rank. International Journal of Algebra and Computation, 21(1-2):329–354, 2011. W. Thomas. Automata on infinite objects. In Handbook of theoretical computer science, Vol. B, pages 133–191. Elsevier, Amsterdam, 1990. M. Vorobets and Y. Vorobets. On a free group of transformations defined by an automaton. Geometriae Dedicata, 124:237–249, 2007. M. Vorobets and Y. Vorobets. On a series of finite automata defining free transformation groups. Groups, Geometry, and Dynamics, 4(2):377–405, 2010. Y. Vorobets. Notes on the Schreier graphs of the Grigorchuk group. In Dynamical systems and group actions, volume 567 of Contemp. Math., pages 221–248. Amer. Math. Soc., Providence, RI, 2012. E. I. Zel′ manov. Solution of the restricted Burnside problem for groups of odd exponent. Izvestiya Akademii Nauk SSSR. Seriya Matematicheskaya, 54-1:42–59, 221, 1990. E. I. Zel′ manov. Solution of the restricted Burnside problem for 2-groups. Matematicheskiı̆ Sbornik, 182-4:568–592, 1991. 40 D. D’ANGELI, TH. GODIN, I. KLIMANN, M. PICANTIN, AND E. RODARO Institut für Diskrete Mathematik, Technische Universität Graz, Steyrergasse 30, 8010 Graz, Austria. E-mail address: [email protected] IRIF, UMR 8243 Université Paris Diderot & CNRS, Bâtiment Sophie Germain, 75205 Paris Cedex 13, France. E-mail address: [email protected] IRIF, UMR 8243 Université Paris Diderot & CNRS, Bâtiment Sophie Germain, 75205 Paris Cedex 13, France. E-mail address: [email protected] IRIF, UMR 8243 Université Paris Diderot & CNRS, Bâtiment Sophie Germain, 75205 Paris Cedex 13, France. E-mail address: [email protected] Department of mathematics, Politecnico di Milano, Piazza Leonardo da Vinci, 32, 20133 Milano, Italy E-mail address: [email protected]
4math.GR
Packrat Parsing: Simple, Powerful, Lazy, Linear Time Functional Pearl Bryan Ford arXiv:cs/0603077v1 [cs.DS] 18 Mar 2006 Massachusetts Institute of Technology Cambridge, MA [email protected] Abstract 1 Introduction Packrat parsing is a novel technique for implementing parsers in a lazy functional programming language. A packrat parser provides the power and flexibility of top-down parsing with backtracking and unlimited lookahead, but nevertheless guarantees linear parse time. Any language defined by an LL(k) or LR(k) grammar can be recognized by a packrat parser, in addition to many languages that conventional linear-time algorithms do not support. This additional power simplifies the handling of common syntactic idioms such as the widespread but troublesome longest-match rule, enables the use of sophisticated disambiguation strategies such as syntactic and semantic predicates, provides better grammar composition properties, and allows lexical analysis to be integrated seamlessly into parsing. Yet despite its power, packrat parsing shares the same simplicity and elegance as recursive descent parsing; in fact converting a backtracking recursive descent parser into a linear-time packrat parser often involves only a fairly straightforward structural change. This paper describes packrat parsing informally with emphasis on its use in practical applications, and explores its advantages and disadvantages with respect to the more conventional alternatives. There are many ways to implement a parser in a functional programming language. The simplest and most direct approach is top-down or recursive descent parsing, in which the components of a language grammar are translated more-or-less directly into a set of mutually recursive functions. Top-down parsers can in turn be divided into two categories. Predictive parsers attempt to predict what type of language construct to expect at a given point by “looking ahead” a limited number of symbols in the input stream. Backtracking parsers instead make decisions speculatively by trying different alternatives in succession: if one alternative fails to match, then the parser “backtracks” to the original input position and tries another. Predictive parsers are fast and guarantee lineartime parsing, while backtracking parsers are both conceptually simpler and more powerful but can exhibit exponential runtime. Categories and Subject Descriptors D.3.4 [Programming Languages]: Processors—Parsing; D.1.1 [Programming Techniques]: Applicative (Functional) Programming; F.4.2 [Mathematical Logic and Formal Languages]: Grammars and Other Rewriting Systems—Parsing General Terms This paper presents a top-down parsing strategy that sidesteps the choice between prediction and backtracking. Packrat parsing provides the simplicity, elegance, and generality of the backtracking model, but eliminates the risk of super-linear parse time, by saving all intermediate parsing results as they are computed and ensuring that no result is evaluated more than once. The theoretical foundations of this algorithm were worked out in the 1970s [3, 4], but the linear-time version was apparently never put in practice due to the limited memory sizes of computers at that time. However, on modern machines the storage cost of this algorithm is reasonable for many applications. Furthermore, this specialized form of memoization can be implemented very elegantly and efficiently in modern lazy functional programming languages, requiring no hash tables or other explicit lookup structures. This marriage of a classic but neglected linear-time parsing algorithm with modern functional programming is the primary technical contribution of this paper. Languages, Algorithms, Design, Performance Haskell, memoization, top-down parsing, backtracking, lexical analysis, scannerless parsing, parser combinators Packrat parsing is unusually powerful despite its linear time guarantee. A packrat parser can easily be constructed for any language described by an LL(k) or LR(k) grammar, as well as for many languages that require unlimited lookahead and therefore are not LR. This flexibility eliminates many of the troublesome restrictions imposed by parser generators of the YACC lineage. Packrat parsers are also much simpler to construct than bottom-up LR parsers, making it practical to build them by hand. This paper explores the manual construction approach, although automatic construction of packrat parsers is a promising direction for future work. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. ICFP’02, October 4-6, 2002, Pittsburgh, Pennsylvania, USA. Copyright 2002 ACM 1-58113-487-8/02/0010 ...$5.00 A packrat parser can directly and efficiently implement common disambiguation rules such as longest-match, followed-by, and notfollowed-by, which are difficult to express unambiguously in a context-free grammar or implement in conventional linear-time Keywords Additive ← Multitive ‘+’ Additive | Multitive Multitive ← Primary ‘*’ Multitive | Primary Primary ← ‘(’ Additive ‘)’ | Decimal Decimal ← ‘0’ | . . . | ‘9’ Figure 1. Grammar for a trivial language parsers. For example, recognizing identifiers or numbers during lexical analysis, parsing if-then-else statements in C-like languages, and handling do, let, and lambda expressions in Haskell inherently involve longest-match disambiguation. Packrat parsers are also more easily and naturally composable than LR parsers, making them a more suitable substrate for dynamic or extensible syntax [1]. Finally, both lexical and hierarchical analysis can be seamlessly integrated into a single unified packrat parser, and lexical and hierarchical language features can even be blended together, so as to handle string literals with embedded expressions or literate comments with structured document markup, for example. The main disadvantage of packrat parsing is its space consumption. Although its asymptotic worst-case bound is the same as those of conventional algorithms—linear in the size of the input—its space utilization is directly proportional to input size rather than maximum recursion depth, which may differ by orders of magnitude. However, for many applications such as modern optimizing compilers, the storage cost of a pacrkat parser is likely to be no greater than the cost of subsequent processing stages. This cost may therefore be a reasonable tradeoff for the power and flexibility of linear-time parsing with unlimited lookahead. The rest of this paper explores packrat parsing with the aim of providing a pragmatic sense of how to implement it and when it is useful. Basic familiarity with context-free grammars and top-down parsing is assumed. For brevity and clarity of presentation, only small excerpts of example code are included in the text. However, all of the examples described in this paper are available, as complete and working Haskell code, at: http://pdos.lcs.mit.edu/˜baford/packrat/icfp02 The paper is organized as follows. Section 2 introduces packrat parsing and describes how it works, using conventional recursive descent parsing as a starting point. Section 3 presents useful extensions to the basic algorithm, such as support for left recursion, lexical analysis, and monadic parsing. Section 4 explores in more detail the recognition power of packrat parsers in comparison with conventional linear-time parsers. Section 5 discusses the three main practical limitations of packrat parsing: determinism, statelessness, and space consumption. Section 6 presents some experimental results to demonstrate the practicality of packrat parsing for real languages. Section 7 discusses related work, Section 8 points out directions for future exploration, and Section 9 concludes. 2 Building a Parser Packrat parsing is essentially a top-down parsing strategy, and as such packrat parsers are closely related to recursive descent parsers. For this reason, we will first build a recursive descent parser for a trivial language and then convert it into a packrat parser. 2.1 Recursive Descent Parsing Consider the standard approach for constructing a recursive descent parser for a grammar such as the trivial arithmetic expression lan- guage shown in Figure 1. We define four functions, one for each of the nonterminals on the left-hand sides of the rules. Each function takes takes the string to be parsed, attempts to recognize some prefix of the input string as a derivation of the corresponding nonterminal, and returns either a “success” or “failure” result. On success, the function returns the remainder of the input string immediately following the part that was recognized, along with some semantic value computed from the recognized part. Each function can recursively call itself and the other functions in order to recognize the nonterminals appearing on the right-hand sides of its corresponding grammar rules. To implement this parser in Haskell, we first need a type describing the result of a parsing function: data Result v = Parsed v String | NoParse In order to make this type generic for different parse functions producing different kinds of semantic values, the Result type takes a type parameter v representing the type of the associated semantic value. A success result is built with the Parsed constructor and contains a semantic value (of type v) and the remainder of the input text (of type String). A failure result is represented by the simple value NoParse. In this particular parser, each of the four parse functions takes a String and produces a Result with a semantic value of type Int: pAdditive pMultitive pPrimary pDecimal :: :: :: :: String String String String -> -> -> -> Result Result Result Result Int Int Int Int The definitions of these functions have the following general structure, directly reflecting the mutual recursion expressed by the grammar in Figure 1: pAdditive pMultitive pPrimary pDecimal s s s s = = = = ... (calls itself and pMultitive) ... ... (calls itself and pPrimary) ... ... (calls pAdditive and pDecimal) ... ... For example, the pAdditive function can be coded as follows, using only primitive Haskell pattern matching constructs: -- Parse an additive-precedence expression pAdditive :: String -> Result Int pAdditive s = alt1 where -- Additive <- Multitive ’+’ Additive alt1 = case pMultitive s of Parsed vleft s’ -> case s’ of (’+’:s’’) -> case pAdditive s’’ of Parsed vright s’’’ -> Parsed (vleft + vright) s’’’ _ -> alt2 _ -> alt2 _ -> alt2 -- Additive <- Multitive alt2 = case pMultitive s of Parsed v s’ -> Parsed v s’ NoParse -> NoParse To compute the result of pAdditive, we first compute the value of alt1, representing the first alternative for this grammar rule. This alternative in turn calls pMultitive to recognize a multiplicativeprecedence expression. If pMultitive succeeds, it returns the semantic value vleft of that expression and the remaining input s’ following the recognized portion of input. We then check for a ‘+’ operator at position s’, which if successful produces the string s’’ representing the remaining input after the ‘+’ operator. Finally, we recursively call pAdditive itself to recognize another additiveprecedence expression at position s’’, which if successful yields the right-hand-side result vright and the final remainder string s’’’. If all three of these matches were successful, then we return as the result of the initial call to pAdditive the semantic value of the addition, vleft + vright, along with the final remainder string s’’’. If any of these matches failed, we fall back on alt2, the second alternative, which merely attempts to recognize a single multiplicative-precedence expression at the original input position s and returns that result verbatim, whether success or failure. The other three parsing functions are constructed similarly, in direct correspondence with the grammar. Of course, there are easier and more concise ways to write these parsing functions, using an appropriate library of helper functions or combinators. These techniques will be discussed later in Section 3.3, but for clarity we will stick to simple pattern matching for now. 2.2 Backtracking Versus Prediction The parser developed above is a backtracking parser. If alt1 in the pAdditive function fails, for example, then the parser effectively “backtracks” to the original input position, starting over with the original input string s in the second alternative alt2, regardless of whether the first alternative failed to match during its first, second, or third stage. Notice that if the input s consists of only a single multiplicative expression, then the pMultitive function will be called twice on the same string: once in the first alternative, which will fail while trying to match a nonexistent ‘+’ operator, and then again while successfully applying the second alternative. This backtracking and redundant evaluation of parsing functions can lead to parse times that grow exponentially with the size of the input, and this is the principal reason why a “naive” backtracking strategy such as the one above is never used in realistic parsers for inputs of substantial size. The standard strategy for making top-down parsers practical is to design them so that they can “predict” which of several alternative rules to apply before actually making any recursive calls. In this way it can be guaranteed that parse functions are never called redundantly and that any input can be parsed in linear time. For example, although the grammar in Figure 1 is not directly suitable for a predictive parser, it can be converted into an LL(1) grammar, suitable for prediction with one lookahead token, by “left-factoring” the Additive and Multitive nonterminals as follows: Additive AdditiveSuffix Multitive MultitiveSuffix ← ← ← ← Multitive AdditiveSuffix ‘+’ Additive | ε Primary MultitiveSuffix ‘*’ Multitive | ε Now the decision between the two alternatives for AdditiveSuffix can be made before making any recursive calls simply by checking whether the next input character is a ‘+’. However, because the prediction mechanism only has “raw” input tokens (characters in this case) to work with, and must itself operate in constant time, the class of grammars that can be parsed predictively is very restrictive. Care must also be taken to keep the prediction mechanism consistent with the grammar, which can be difficult to do manu- column C1 C2 C3 C4 C5 C6 pAdditive (7,C7) (4,C7) pMultitive (3,C5) (4,C7) (3,C5) (4,C7) (3,C5) (4,C7) ? pPrimary pDecimal input ’2’ ’*’ ’(’ ’3’ ’+’ ’4’ C7 C8 ’)’ (end) Figure 2. Matrix of parsing results for string ‘2*(3+4)’ ally and highly sensitive to global properties of the language. For example, the prediction mechanism for MultitiveSuffix would have to be adjusted if a higher-precedence exponentiation operator ‘**’ was added to the language; otherwise the exponentiation operator would falsely trigger the predictor for multiplication expressions and cause the parser to fail on valid input. Some top-down parsers use prediction for most decisions but fall back on full backtracking when more flexibility is needed. This strategy often yields a good combination of flexibility and performance in practice, but it still suffers the additional complexity of prediction, and it requires the parser designer to be intimately aware of where prediction can be used and when backtracking is required. 2.3 Tabular Top-Down Parsing As pointed out by Birman and Ullman [4], a backtracking top-down parser of the kind presented in Section 2.1 can be made to operate in linear time without the added complexity or constraints of prediction. The basic reason the backtracking parser can take super-linear time is because of redundant calls to the same parse function on the same input substring, and these redundant calls can be eliminated through memoization. Each parse function in the example is dependent only on its single parameter, the input string. Whenever a parse function makes a recursive call to itself or to another parse function, it always supplies either the same input string it was given (e.g., for the call by pAdditive to pMultitive), or a suffix of the original input string (e.g., for the recursive call by pAdditive to itself after matching a ‘+’ operator). If the input string is of length n, then there are only n + 1 distinct suffixes that might be used in these recursive calls, counting the original input string itself and the empty string. Since there are only four parse functions, there are at most 4(n + 1) distinct intermediate results that the parsing process might require. We can avoid computing any of these intermediate results multiple times by storing them in a table. The table has one row for each of the four parse functions and one column for each distinct position in the input string. We fill the table with the results of each parse function for each input position, starting at the right end of the input string and working towards the left, column by column. Within each column, we start from the bottommost cell and work upwards. By the time we compute the result for a given cell, the results of all would-be recursive calls in the corresponding parse function will already have been computed and recorded elsewhere in the table; we merely need to look up and use the appropriate results. Figure 2 illustrates a partially-completed result table for the input string ‘2*(3+4)’. For brevity, Parsed results are indicated as (v,c), where v is the semantic value and c is the column number at which the associated remainder suffix begins. Columns are labeled C1, C2, and so on, to avoid confusion with the integer semantic values. NoParse results are indicated with an X in the cell. The next cell to be filled is the one for pPrimary at column C3, indicated with a circled question mark. The rule for Primary expressions has two alternatives: a parenthesized Additive expression or a Decimal digit. If we try the alternatives in the order expressed in the grammar, pPrimary will first check for a parenthesized Additive expression. To do so, pPrimary first attempts to match an opening ‘(’ in column C3, which succeeds and yields as its remainder string the input suffix starting at column C4, namely ‘3+4)’. In the simple recursive-descent parser pPrimary would now recursively call pAdditive on this remainder string. However, because we have the table we can simply look up the result for pAdditive at column C4 in the table, which is (7,C7). This entry indicates a semantic value of 7—the result of the addition expression ‘3+4’—and a remainder suffix of ‘)’ starting in column C7. Since this match is a success, pPrimary finally attempts to match the closing parenthesis at position C7, which succeeds and yields the empty string C8 as the remainder. The result entered for pPrimary at column C3 is thus (7,C8). Although for a long input string and a complex grammar this result table may be large, it only grows linearly with the size of the input assuming the grammar has a fixed number of nonterminals. Furthermore, as long as the grammar uses only the standard operators of Backus-Naur Form [2], only a fixed number of previouslyrecorded cells in the matrix need to be accessed in order to compute each new result. Therefore, assuming table lookup occurs in constant time, the parsing process as a whole completes in linear time. Due to the “forward pointers” embedded in the results table, the computation of a given result may examine cells that are widely spaced in the matrix. For example, computing the result for pPrimary at C3 above made use of results from columns C3, C4, and C7. This ability to skip ahead arbitrary distances while making parsing decisions is the source of the algorithm’s unlimited lookahead capability, and this capability makes the algorithm more powerful than linear-time predictive parsers or LR parsers. 2.4 Packrat Parsing An obvious practical problem with the tabular right-to-left parsing algorithm above is that it computes many results that are never needed. An additional inconvenience is that we must carefully determine the order in which the results for a particular column are computed, so that parsing functions such as pAdditive and pMultitive that depend on other results from the same column will work correctly. Packrat parsing is essentially a lazy version of the tabular algorithm that solves both of these problems. A packrat parser computes results only as they are needed, in the same order as the original recursive descent parser would. However, once a result is computed for the first time, it is stored for future use by subsequent calls. A non-strict functional programming language such as Haskell provides an ideal implementation platform for a packrat parser. In fact, packrat parsing in Haskell is particularly efficient because it does not require arrays or any other explicit lookup structures other than the language’s ordinary algebraic data types. First we will need a new type to represent a single column of the parsing result matrix, which we will call Derivs (“derivations”). This type is merely a tuple with one component for each nonterminal in the grammar. Each component’s type is the result type of the corresponding parse function. The Derivs type also contains one additional component, which we will call dvChar, to represent “raw” characters of the input string as if they were themselves the results of some parsing function. The Derivs type for our example parser can be conveniently declared in Haskell as follows: data Derivs = Derivs { dvAdditive dvMultitive dvPrimary dvDecimal dvChar :: :: :: :: :: Result Result Result Result Result Int, Int, Int, Int, Char} This Haskell syntax declares the type Derivs to have a single constructor, also named Derivs, with five components of the specified types. The declaration also automatically creates a corresponding data-accessor function for each component: dvAdditive can be used as a function of type Derivs → Result Int, which extracts the first component of a Derivs tuple, and so on. Next we modify the Result type so that the “remainder” component of a success result is not a plain String, but is instead an instance of Derivs: data Result v = Parsed v Derivs | NoParse The Derivs and Result types are now mutually recursive: the success results in one Derivs instance act as links to other Derivs instances. These result values in fact provide the only linkage we need between different columns in the matrix of parsing results. Now we modify the original recursive-descent parsing functions so that each takes a Derivs instead of a String as its parameter: pAdditive pMultitive pPrimary pDecimal :: :: :: :: Derivs Derivs Derivs Derivs -> -> -> -> Result Result Result Result Int Int Int Int Wherever one of the original parse functions examined input characters directly, the new parse function instead refers to the dvChar component of the Derivs object. Wherever one of the original functions made a recursive call to itself or another parse function, in order to match a nonterminal in the grammar, the new parse function instead instead uses the Derivs accessor function corresponding to that nonterminal. Sequences of terminals and nonterminals are matched by following chains of success results through multiple Derivs instances. For example, the new pAdditive function uses the dvMultitive, dvChar, and dvAdditive accessors as follows, without making any direct recursive calls: -- Parse an additive-precedence expression pAdditive :: Derivs -> Result Int pAdditive d = alt1 where -- Additive <- Multitive ’+’ Additive alt1 = case dvMultitive d of Parsed vleft d’ -> case dvChar d’ of Parsed ’+’ d’’ -> case dvAdditive d’’ of Parsed vright d’’’ -> Parsed (vleft + vright) d’’’ _ -> alt2 _ -> alt2 _ -> alt2 -- Additive <- Multitive alt2 = dvMultitive d Finally, we create a special “top-level” function, parse, to produce instances of the Derivs type and “tie up” the recursion between all of the individual parsing functions: -- Create a result matrix for an input string parse :: String -> Derivs parse s = d where d = Derivs add mult prim dec chr add = pAdditive d mult = pMultitive d prim = pPrimary d dec = pDecimal d chr = case s of (c:s’) -> Parsed c (parse s’) [] -> NoParse The “magic” of the packrat parser is in this doubly-recursive function. The first level of recursion is produced by the parse function’s reference to itself within the case statement. This relatively conventional form of recursion is used to iterate over the input string one character at a time, producing one Derivs instance for each input position. The final Derivs instance, representing the empty string, is assigned a dvChar result of NoParse, which effectively terminates the list of columns in the result matrix. The second level of recursion is via the symbol d. This identifier names the Derivs instance to be constructed and returned by the parse function, but it is also the parameter to each of the individual parsing functions. These parsing functions, in turn, produce the rest of the components forming this very Derivs object. This form of data recursion of course works only in a non-strict language, which allow some components of an object to be accessed before other parts of the same object are available. For example, in any Derivs instance created by the above function, the dvChar component can be accessed before any of the other components of the tuple are available. Attempting to access the dvDecimal component of this tuple will cause pDecimal to be invoked, which in turn uses the dvChar component but does not require any of the other “higher-level” components. Accessing the dvPrimary component will similarly invoke pPrimary, which may access dvChar and dvAdditive. Although in the latter case pPrimary is accessing a “higher-level” component, doing so does not create a cyclic dependency in this case because it only ever invokes dvAdditive on a different Derivs object from the one it was called with: namely the one for the position following the opening parenthesis. Every component of every Derivs object produced by parse can be lazily evaluated in this fashion. Figure 3 illustrates the data structure produced by the parser for the example input text ‘2*(3+4)’, as it would appear in memory under a modern functional evaluator after fully reducing every cell. Each vertical column represents a Derivs instance with its five Result components. For results of the form ‘Parsed v d’, the semantic value v is shown in the appropriate cell, along with an arrow representing the “remainder” pointer leading to another Derivs instance in the matrix. In any modern lazy language implementation that properly preserves sharing relationships during evaluation, the arrows in the diagram will literally correspond to pointers in the heap, and a given cell in the structure will never be evaluated twice. Shaded boxes represent cells that would never be evaluated at all in dvAdditive 14 7 7 ’4’ dvMultitive 14 7 3 4 dvPrimary 2 7 3 4 dvDecimal 2 dvChar ’2’ 4 3 ’*’ ’(’ ’3’ ’+’ ’4’ ’)’ Figure 3. Illustration of Derivs data structure produced by parsing the string ‘2*(3+4)’ the likely case that the dvAdditive result in the leftmost column is the only value ultimately needed by the application. This illustration should make it clear why this algorithm can run in O(n) time under a lazy evaluator for an input string of length n. The top-level parse function is the only function that creates instances of the Derivs type, and it always creates exactly n + 1 instances. The parse functions only access entries in this structure instead of making direct calls to each other, and each function examines at most a fixed number of other cells while computing a given result. Since the lazy evaluator ensures that each cell is evaluated at most once, the critical memoization property is provided and linear parse time is guaranteed, even though the order in which these results are evaluated is likely to be completely different from the tabular, right-to-left, bottom-to-top algorithm presented earlier. 3 Extending the Algorithm The previous section provided the basic principles and tools required to create a packrat parser, but building parsers for real applications involves many additional details, some of which are affected by the packrat parsing paradigm. In this section we will explore some of the more important practical issues, while incrementally building on the example packrat parser developed above. We first examine the annoying but straightforward problem of left recursion. Next we address the issue of lexical analysis, seamlessly integrating this task into the packrat parser. Finally, we explore the use of monadic combinators to express packrat parsers more concisely. 3.1 Left Recursion One limitation packrat parsing shares with other top-down schemes is that it does not directly support left recursion. For example, suppose we wanted to add a subtraction operator to the above example and have addition and subtraction be properly left-associative. A natural approach would be to modify the grammar rules for Additive expressions as follows, and to change the parser accordingly: Additive ← | | Additive ‘+’ Multitive Additive ‘-’ Multitive Multitive In a recursive descent parser for this grammar, the pAdditive function would recursively invoke itself with the same input it was provided, and therefore would get into an infinite recursion cycle. In a packrat parser for this grammar, pAdditive would attempt to access the dvAdditive component of its own Derivs tuple—the same component it is supposed to compute—and thus would create a circular data dependency. In either case the parser fails, although the packrat parser’s failure mode might be viewed as slightly “friendlier” since modern lazy evaluators often detect circular data dependencies at run-time but cannot detect infinite recursion. Fortunately, a left-recursive grammar can always be rewritten into an equivalent right-recursive one [2], and the desired leftassociative semantic behavior is easily reconstructed using higherorder functions as intermediate parser results. For example, to make Additive expressions left-associative in the example parser, we can split this rule into two nonterminals, Additive and AdditiveSuffix. The pAdditive function recognizes a single Multitive expression followed by an AdditiveSuffix: pAdditive :: Derivs -> Result Int pAdditive d = case dvMultitive d of Parsed vl d’ -> case dvAdditiveSuffix d’ of Parsed suf d’’ -> Parsed (suf vl) d’’ _ -> NoParse _ -> NoParse The pAdditiveSuffix function collects infix operators and righthand-side operands, and builds a semantic value of type ‘Int → Int’, which takes a left-hand-side operand and produces a result: pAdditiveSuffix :: Derivs -> Result (Int -> Int) pAdditiveSuffix d = alt1 where -- AdditiveSuffix <- ’+’ Multitive AdditiveSuffix alt1 = case dvChar d of Parsed ’+’ d’ -> case dvMultitive d’ of Parsed vr d’’ -> case dvAdditiveSuffix d’’ of Parsed suf d’’’ -> Parsed (\vl -> suf (vl + vr)) d’’’ _ -> alt2 _ -> alt2 _ -> alt2 -- AdditiveSuffix <- <empty> alt3 = Parsed (\v -> v) d 3.2 Integrated Lexical Analysis Traditional parsing algorithms usually assume that the “raw” input text has already been partially digested by a separate lexical analyzer into a stream of tokens. The parser then treats these tokens as atomic units even though each may represent multiple consecutive input characters. This separation is usually necessary because conventional linear-time parsers can only use primitive terminals in their lookahead decisions and cannot refer to higher-level nonterminals. This limitation was explained in Section 2.2 for predictive top-down parsers, but bottom-up LR parsers also depend on a similar token-based lookahead mechanism sharing the same problem. If a parser can only use atomic tokens in its lookahead decisions, then parsing becomes much easier if those tokens represent whole keywords, identifiers, and literals rather than raw characters. Packrat parsing suffers from no such lookahead limitation, however. Because a packrat parser reflects a true backtracking model, decisions between alternatives in one parsing function can depend on complete results produced by other parsing functions. For this reason, lexical analysis can be integrated seamlessly into a packrat parser with no special treatment. To extend the packrat parser example with “real” lexical analysis, we add some new nonterminals to the Derivs type: data Derivs = Derivs { -- Expressions dvAdditive :: Result Int, ... -- Lexical tokens dvDigits :: Result dvDigit :: Result dvSymbol :: Result dvWhitespace :: Result (Int, Int), Int, Char, (), -- Raw input dvChar :: Result Char} The pWhitespace parse function consumes any whitespace that may separate lexical tokens: pWhitespace :: Derivs -> Result () pWhitespace d = case dvChar d of Parsed c d’ -> if isSpace c then pWhitespace d’ else Parsed () d _ -> Parsed () d In a more complete language, this function might have the task of eating comments as well. Since the full power of packrat parsing is available for lexical analysis, comments could have a complex hierarchical structure of their own, such as nesting or markups for literate programming. Since syntax recognition is not broken into a unidirectional pipeline, lexical constructs can even refer “upwards” to higher-level syntactic elements. For example, a language’s syntax could allow identifiers or code fragments embedded within comments to be demarked so the parser can find and analyze them as actual expressions or statements, making intelligent software engineering tools more effective. Similarly, escape sequences in string literals could contain generic expressions representing static or dynamic substitutions. The pWhitespace example also illustrates how commonplace longest-match disambiguation rules can be easily implemented in a packrat parser, even though they are difficult to express in a pure context-free grammar. More sophisticated decision and disambiguation strategies are easy to implement as well, including general syntactic predicates [14], which influence parsing decisions based on syntactic lookahead information without actually consuming input text. For example, the useful followed-by and notfollowed-by rules allow a parsing alternative to be used only if the text matched by that alternative is (or is not) followed by text matching some other arbitrary nonterminal. Syntactic predicates of this kind require unlimited lookahead in general and are therefore outside the capabilities of most other linear-time parsing algorithms. Continuing with the lexical analysis example, the function pSymbol recognizes “operator tokens” consisting of an operator character followed by optional whitespace: -- Parse an operator followed by optional whitespace pSymbol :: Derivs -> Result Char pSymbol d = case dvChar d of Parsed c d’ -> if c ‘elem‘ "+-*/%()" then case dvWhitespace d’ of Parsed _ d’’ -> Parsed c d’’ _ -> NoParse else NoParse _ -> NoParse Now we modify the higher-level parse functions for expressions to use dvSymbol instead of dvChar to scan for operators and parentheses. For example, pPrimary can be implemented as follows: and simple aliases cannot be assigned type classes. We must instead wrap the parsing functions with a “real” user-defined type: newtype Parser v = Parser (Derivs -> Result v) -- Parse a primary expression pPrimary :: Derivs -> Result Int pPrimary d = alt1 where -- Primary <- ’(’ Additive ’)’ alt1 = case dvSymbol d of Parsed ’(’ d’ -> case dvAdditive d’ of Parsed v d’’ -> case dvSymbol d’’ of Parsed ’)’ d’’’ -> Parsed v d’’’ _ -> alt2 _ -> alt2 _ -> alt2 We can now implement Haskell’s standard sequencing (>>=), result-producing (return), and error-producing combinators: instance Monad Parser where (Parser p1) >>= f2 = Parser pre where pre d = post (p1 d) post (Parsed v d’) = p2 d’ where Parser p2 = f2 v post (NoParse) = NoParse return x = Parser (\d -> Parsed x d) fail msg = Parser (\d -> NoParse) -- Primary <- Decimal alt2 = dvDecimal d This function demonstrates how parsing decisions can depend not only on the existence of a match at a given position for a nonterminal such as Symbol, but also on the semantic value associated with that nonterminal. In this case, even though all symbol tokens are parsed together and treated uniformly by pSymbol, other rules such as pPrimary can still distinguish between particular symbols. In a more sophisticated language with multi-character operators, identifiers, and reserved words, the semantic values produced by the token parsers might be of type String instead of Char, but these values can be matched in the same way. Such dependencies of syntax on semantic values, known as semantic predicates [14], provide an extremely powerful and useful capability in practice. As with syntactic predicates, semantic predicates require unlimited lookahead in general and cannot be implemented by conventional parsing algorithms without giving up their linear time guarantee. 3.3 Monadic Packrat Parsing A popular method of constructing parsers in functional languages such as Haskell is using monadic combinators [11, 13]. Unfortunately, the monadic approach usually comes with a performance penalty, and with packrat parsing this tradeoff presents a difficult choice. Implementing a packrat parser as described so far assumes that the set of nonterminals and their corresponding result types is known statically, so that they can be bound together in a single fixed tuple to form the Derivs type. Constructing entire packrat parsers dynamically from other packrat parsers via combinators would require making the Derivs type a dynamic lookup structure, associating a variable set of nonterminals with corresponding results. This approach would be much slower and less space-efficient. A more practical strategy, which provides most of the convenience of combinators with a less significant performance penalty, is to use monads to define the individual parsing functions comprising a packrat parser, while keeping the Derivs type and the “top-level” recursion statically implemented as described earlier. Since we would like our combinators to build the parse functions we need directly, the obvious method would be to make the combinators work with a simple type alias: Finally, for parsing we need an alternation combinator: (<|>) :: Parser v -> Parser v -> Parser v (Parser p1) <|> (Parser p2) = Parser pre where pre d = post d (p1 d) post d NoParse = p2 d post d r = r With these combinators in addition to a trivial one to recognize specific characters, the pAdditive function in the original packrat parser example can be written as follows: Parser pAdditive = (do vleft <- Parser dvMultitive char ’+’ vright <- Parser dvAdditive return (vleft + vright)) <|> (do Parser dvMultitive) It is tempting to build additional combinators for higher-level idioms such as repetition and infix expressions. However, using iterative combinators within packrat parsing functions violates the assumption that each cell in the result matrix can be computed in constant time once the results from any other cells it depends on are available. Iterative combinators effectively create “hidden” recursion whose intermediate results are not memoized in the result matrix, potentially making the parser run in super-linear time. This problem is not necessarily serious in practice, as the results in Section 6 will show, but it should be taken into account when using iterative combinators. The on-line examples for this paper include a full-featured monadic combinator library that can be used to build large packrat parsers conveniently. This library is substantially inspired by PARSEC [13], though the packrat parsing combinators are much simpler since they do not have to implement lexical analysis as a separate phase or implement the one-token-lookahead prediction mechanism used by traditional top-down parsers. The full combinator library provides a variety of “safe” constant-time combinators, as well as a few “dangerous” iterative ones, which are convenient but not necessary to construct parsers. The combinator library can be used simultaneously by multiple parsers with different Derivs types, and supports user-friendly error detection and reporting. type Parser v = Derivs -> Result v 4 Comparison with LL and LR Parsing Unfortunately, in order to take advantage of Haskell’s useful do syntax, the combinators must use a type of the special class Monad, Whereas the previous sections have served as a tutorial on how to construct a packrat parser, for the remaining sections we turn to S R A P the issue of when packrat parsing is useful in practice. This section informally explores the language recognition power of packrat parsing in more depth, and clarifies its relationship to traditional linear-time algorithms such as LL(k) and LR(k). Although LR parsing is commonly seen as “more powerful” than limited-lookahead top-down or LL parsing, the class of languages these parsers can recognize is the same [3]. As Pepper points out [17], LR parsing can be viewed simply as LL parsing with the grammar rewritten so as to eliminate left recursion and to delay all important parsing decisions as long as possible. The result is that LR provides more flexibility in the way grammars can be expressed, but no actual additional recognition power. For this reason, we will treat LL and LR parsers here as being essentially equivalent. 4.1 Lookahead The most critical practical difference between packrat parsing and LL/LR parsing is the lookahead mechanism. A packrat parser’s decisions at any point can be based on all the text up to the end of the input string. Although the computation of an individual result in the parsing matrix can only perform a constant number of “basic operations,” these basic operations include following forward pointers in the parsing matrix, each of which can skip over a large amount of text at once. Therefore, while LL and LR parsers can only look ahead a constant number of terminals in the input, packrat parsers can look ahead a constant number of terminals and nonterminals in any combination. This ability for parsing decisions to take arbitrary nonterminals into account is what gives packrat parsing its unlimited lookahead capability. To illustrate the difference in language recognition power, the following grammar is not LR(k) for any k, but is not a problem for a packrat parser: S A B ← ← ← A|B xAy|xzy xByy|xzyy Once an LR parser has encountered the ‘z’ and the first following ‘y’ in a string in the above language, it must decide immediately whether to start reducing via nonterminal A or B, but there is no way for it to make this decision until as many ‘y’s have been encountered as there were ‘x’s on the left-hand side. A packrat parser, on the other hand, essentially operates in a speculative fashion, producing derivations for nonterminals A and B in parallel while scanning the input. The ultimate decision between A and B is effectively delayed until the entire input string has been parsed, where the decision is merely a matter of checking which nonterminal has a success result at that position. Mirroring the above grammar left to right does not change the situation, making it clear that the difference is not merely some side-effect of the fact that LR scans the input left-to-right whereas packrat parsing seems to operate in reverse. 4.2 Grammar Composition The limitations of LR parsing due to fixed lookahead are frequently felt when designing parsers for practical languages, and many of these limitations stem from the fact that LL and LR grammars are not cleanly composable. For example, the following grammar represents a simple language with expressions and assignment, which only allows simple identifiers on the left side of an assignment: ← ← ← ← R | ID ‘=’ R A | A EQ A | A NE A P | P ‘+’ P | P ‘-’ P ID | ‘(’ R ‘)’ If the symbols ID, EQ, and NE are terminals—i.e., atomic tokens produced by a separate lexical analysis phase—then an LR(1) parser has no trouble with this grammar. However, if we try to integrate this tokenization into the parser itself with the following simple rules, the grammar is no longer LR(1): ID EQ NE ← ← ← ’a’ | ’a’ ID ’=’ ’=’ ’!’ ’=’ The problem is that after scanning an identifier, an LR parser must decide immediately whether it is a primary expression or the lefthand side of an assignment, based only on the immediately following token. But if this token is an ‘=’, the parser has no way of knowing whether it is an assignment operator or the first half of an ‘==’ operator. In this particular case the grammar could be parsed by an LR(2) parser. In practice LR(k) and even LALR(k) parsers are uncommon for k > 1. Recently developed extensions to the traditional left-to-right parsing algorithms improve the situation somewhat [18, 16, 15], but they still cannot provide unrestricted lookahead capability while maintaining the linear time guarantee. Even when lexical analysis is separated from parsing, the limitations of LR parsers often surface in other practical situations, frequently as a result of seemingly innocuous changes to an evolving grammar. For example, suppose we want to add simple array indexing to the language above, so that array indexing operators can appear on either the left or right side of an assignment. One possible approach is to add a new nonterminal, L, to represent left-side or “lvalue” expressions, and incorporate the array indexing operator into both types of expressions as shown below: S R A P L ← ← ← ← ← R | L ‘=’ R A | A EQ A | A NE A P | P ‘+’ P | P ‘-’ P ID | ‘(’ R ‘)’ | P ‘[’ A ‘]’ ID | ‘(’ L ‘)’ | L ‘[’ A ‘]’ Even if the ID, EQ, and NE symbols are again treated as terminals, this grammar is not LR(k) for any k, because after the parser sees an identifier it must immediately decide whether it is part of a P or L expression, but it has no way of knowing this until any following array indexing operators have been fully parsed. Again, a packrat parser has no trouble with this grammar because it effectively evaluates the P and L alternatives “in parallel” and has complete derivations to work with (or the knowledge of their absence) by the time the critical decision needs to be made. In general, grammars for packrat parsers are composable because the lookahead a packrat parser uses to make decisions between alternatives can take account of arbitrary nonterminals, such as EQ in the first example or P and L in the second. Because a packrat parser does not give “primitive” syntactic constructs (terminals) any special significance as an LL or LR parser does, any terminal or fixed sequence of terminals appearing in a grammar can be substituted with a nonterminal without “breaking” the parser. This substitution capability gives packrat parsing greater composition flexibility. 4.3 Recognition Limitations Given that a packrat parser can recognize a broader class of languages in linear time than either LL(k) or LR(k) algorithms, what kinds of grammars can’t a packrat parser recognize? Though the precise theoretical capabilities of the algorithm have not been thoroughly characterized, the following trivial and unambiguous context-free grammar provides an example that proves just as troublesome for a packrat parser as for an LL or LR parser: S ← xSx|x The problem with this grammar for both kinds of parsers is that, while scanning a string of ‘x’s—left-to-right in the LR case or rightto-left in the packrat case—the algorithm would somehow have to “know” in advance where the middle of the string is so that it can apply the second alternative at that position and then “build outwards” using the first alternative for the rest of the input stream. But since the stream is completely homogeneous, there is no way for the parser to find the middle until the entire input has been parsed. This grammar therefore provides an example, albeit contrived, requiring a more general, non-linear-time CFG parsing algorithm. 5 Practical Issues and Limitations Although packrat parsing is powerful and efficient enough for many applications, there are three main issues that can make it inappropriate in some situations. First, packrat parsing is useful only to construct deterministic parsers: parsers that can produce at most one result. Second, a packrat parser depends for its efficiency on being mostly or completely stateless. Finally, due to its reliance on memoization, packrat parsing is inherently space-intensive. These three issues are discussed in this section. 5.1 Deterministic Parsing An important assumption we have made so far is that each of the mutually recursive parsing functions from which a packrat parser is built will deterministically return at most one result. If there are any ambiguities in the grammar the parser is built from, then the parsing functions must be able to resolve them locally. In the example parsers developed in this paper, multiple alternatives have always been implicitly disambiguated by the order in which they are tested: the first alternative to match successfully is the one used, independent of whether any other alternatives may also match. This behavior is both easy to implement and useful for performing longestmatch and other forms of explicit local disambiguation. A parsing function could even try all of the possible alternatives and produce a failure result if more than one alternative matches. What parsing functions in a packrat parser cannot do is return multiple results to be used in parallel or disambiguated later by some global strategy. In languages designed for machine consumption, the requirement that multiple matching alternatives be disambiguated locally is not much of a problem in practice because ambiguity is usually undesirable in the first place, and localized disambiguation rules are preferred over global ones because they are easier for humans to understand. However, for parsing natural languages or other grammars in which global ambiguity is expected, packrat parsing is less likely to be useful. Although a classic nondeterministic top-down parser in which the parse functions return lists of results [23, 8] could be memoized in a similar way, the resulting parser would not be linear time, and would likely be comparable to existing tabular algorithms for ambiguous context-free grammars [3, 20]. Since nondeterministic parsing is equivalent in computational complexity to boolean matrix multiplication [12], a linear-time solution to this more general problem is unlikely to be found. 5.2 Stateless Parsing A second limitation of packrat parsing is that it is fundamentally geared toward stateless parsing. A packrat parser’s memoization system assumes that the parsing function for each nonterminal depends only on the input string, and not on any other information accumulated during the parsing process. Although pure context-free grammars are by definition stateless, many practical languages require a notion of state while parsing and thus are not really context-free. For example, C and C++ require the parser to build a table of type names incrementally as types are declared, because the parser must be able to distinguish type names from other identifiers in order to parse subsequent text correctly. Traditional top-down (LL) and bottom-up (LR) parsers have little trouble maintaining state while parsing. Since they perform only a single left-to-right scan of the input and never look ahead more than one or at most a few tokens, nothing is “lost” when a state change occurs. A packrat parser, in contrast, depends on statelessness for the efficiency of its unlimited lookahead capability. Although a stateful packrat parser can be constructed, the parser must start building a new result matrix each time the parsing state changes. For this reason, stateful packrat parsing may be impractical if state changes occur frequently. For more details on packrat parsing with state, please refer to my master’s thesis [9]. 5.3 Space Consumption Probably the most striking characteristic of a packrat parser is the fact that it literally squirrels away everything it has ever computed about the input text, including the entire input text itself. For this reason packrat parsing always has storage requirements equal to some possibly substantial constant multiple of the input size. In contrast, LL(k), LR(k), and simple backtracking parsers can be designed so that space consumption grows only with the maximum nesting depth of the syntactic constructs appearing in the input, which in practice is often orders of magnitude smaller than the total size of the text. Although LL(k) and LR(k) parsers for any nonregular language still have linear space requirements in the worst case, this “average-case” difference can be important in practice. One way to reduce the space requirements of the derivations structure, especially in parsers for grammars with many nonterminals, is by splitting up the Derivs type into multiple levels. For example, suppose the nonterminals of a language can be grouped into several broad categories, such as lexical tokens, expressions, statements, and declarations. Then the Derivs tuple itself might have only four components in addition to dvChar, one for each of these nonterminal categories. Each of these components is in turn a tuple containing the results for all of the nonterminals in that category. For the majority of the Derivs instances, representing character positions “between tokens,” none of the components representing the categories of nonterminals will ever be evaluated, and so only the small top-level object and the unevaluated closures for its components occupy space. Even for Derivs instances corresponding to the beginning of a token, often the results from only one or two categories will be needed depending on what kind of language construct is located at that position. On the other hand, for parsing complex modern programming languages in which the source code is usually written by humans and the top priority is the power and expressiveness of the language, the space cost of packrat parsing is probably reasonable. Standard programming practice involves breaking up large programs into modules of manageable size that can be independently compiled, and the main memory sizes of modern machines leave at least three orders of magnitude in “headroom” for expansion of a typical 10– 100KB source file during parsing. Even when parsing larger source files, the working set may still be relatively small due to the strong structural locality properties of realistic languages. Finally, since the entire derivations structure can be thrown away after parsing is complete, the parser’s space consumption is likely to be irrelevant if its result is fed into some other complex computation, such as a global optimizer, that requires as much space as the packrat parser used. Section 6 will present evidence that this space consumption can be reasonable in practice. 6 Performance Results Although a detailed empirical analysis of packrat parsing is outside the scope of this paper, it is helpful to have some idea of how a packrat parser is likely to behave in practice before committing to a new and unfamiliar parsing paradigm. For this reason, this section presents a few experimental results with realistic packrat parsers running on real source files. For more detailed results, please refer to my master’s thesis [9]. 6.1 Space Efficiency The first set of tests measure the space efficiency of a packrat parser for the Java1 programming language. I chose Java for this experiment because it has a rich and complex grammar, but nevertheless adopts a fairly clean syntactic paradigm, not requiring the parser to keep state about declared types as C and C++ parsers do, or to perform special processing between lexical and hierarchical analysis as Haskell’s layout scheme requires. The experiment uses two different versions of this Java parser. Apart from a trivial preprocessing stage to canonicalize line breaks and Java’s Unicode escape sequences, lexical analysis for both parsers is fully integrated as described in Section 3.2. One parser uses monadic combinators in its lexical analysis functions, while the other parser relies only on primitive pattern matching. Both parsers use monadic combinators to construct all higher-level parsing functions. Both parsers also use the technique described in Section 5.3 of splitting the Derivs tuple into two levels, in order to increase modularity and reduce space consumption. The parsers were compiled with the Glasgow Haskell Compiler2 version 5.04, with optimization and profiling enabled. GHC’s heap profiling system was used to measure live heap utilization, which excludes unused heap space and collectible garbage when samples are taken. 1 Java is a trademark of Sun Microsystems, Inc. 2 http://www.haskell.org/ghc/ 100 Monadic parser and scanner Monadic parser, pattern-matching scanner 80 Maximum heap size (MB) Even with such optimizations a packrat parser can consume many times more working storage than the size of the original input text. For this reason there are some application areas in which packrat parsing is probably not the best choice. For example, for parsing XML streams, which have a fairly simple structure but often encode large amounts of relatively flat, machine-generated data, the power and flexibility of packrat parsing is not needed and its storage cost would not be justified. Avg. 695:1 60 40 Avg. 301:1 20 0 0 20000 40000 60000 80000 Source file size (bytes) 100000 120000 140000 Figure 4. Maximum heap size versus input size The test suite consists of 60 unmodified Java source files from the Cryptix library3 , chosen because it includes a substantial number of relatively large Java source files. (Java source files are small on average because the compilation model encourages programmers to place each class definition in a separate file.) Figure 4 shows a plot of each parser’s maximum live heap size against the size of the input files being parsed. Because some of the smaller source files were parsed so quickly that garbage collection never occurred and the heap profiling mechanism did not yield any samples, the plot includes only 45 data points for the fully monadic parser, and 31 data points for the hybrid parser using direct pattern matching for lexical analysis. Averaged across the test suite, the fully monadic parser uses 695 bytes of live heap per byte of input, while the hybrid parser uses only 301 bytes of heap per input byte. These results are encouraging: although packrat parsing can consume a substantial amount of space, a typical modern machine with 128MB or more of RAM should have no trouble parsing source files up to 100-200KB. Furthermore, even though both parsers use some iterative monadic combinators, which can break the linear time and space guarantee in theory, the space consumption of the parsers nevertheless appears to grow fairly linearly. The use of monadic combinators clearly has a substantial penalty in terms of space efficiency. Modifying the parser to use direct pattern matching alone may yield further improvement, though the degree is difficult to predict since the cost of lexical analysis often dominates the rest of the parser. The lexical analysis portion of the hybrid parser is about twice as long as the equivalent portion of the monadic parser, suggesting that writing packrat parsers with pattern matching alone is somewhat more cumbersome but not unreasonable when efficiency is important. 6.2 Parsing Performance The second experiment measures the absolute execution time of the two packrat parsers. For this test the parsers were compiled by GHC 5.04 with optimization but without profiling, and timed on a 1.28GHz AMD Athlon processor running Linux 2.4.17. For this test I only used the 28 source files in the test suite that were larger than 10KB, because the smaller files were parsed so quickly that the Linux time command did not yield adequate precision. Figure 5 shows the resulting execution time plotted against source file size. On these inputs the fully monadic parser averaged 25.0 Kbytes 3 http://www.cryptix.org/ 6 Monadic parser and scanner Monadic parser, pattern-matching scanner Execution time (seconds) 5 Avg. 25.0Kbps 4 3 Avg. 49.8Kbps 2 1 0 0 20000 40000 60000 80000 Source file size (bytes) 100000 120000 140000 Figure 5. Execution time versus input size per second with a standard deviation of 8.6 KB/s, while the hybrid parser averaged 49.8 KB/s with a standard deviation of 16 KB/s. In order to provide a legitimate performance comparison between packrat parsing and more traditional linear-time algorithms, I converted a freely available YACC grammar for Java [5] into a grammar for Happy4 , an LR parser generator for Haskell. Unfortunately, GHC was unable to compile the 230KB Haskell source file resulting from this grammar, even without optimization and on a machine with 1GB of RAM. (This difficulty incidentally lends credibility to the earlier suggestion that, in modern compilers, the temporary storage cost of a packrat parser is likely to be exceeded by the storage cost of subsequent stages.) Nevertheless, the generated LR parser worked under the Haskell interpreter Hugs.5 Therefore, to provide a rough performance comparison, I ran five of the larger Java sources through the LR and packrat parsers under Hugs using an 80MB heap. For fairness, I only compared the LR parser against the slower, fully monadic packrat parser, because the LR parser uses a monadic lexical analyzer derived from the latter packrat parser. The lexical analysis performance should therefore be comparable and only the parsing algorithm is of primary importance. Under Hugs, the LR parser consistently performs approximately twice the number of reductions and allocates 55% more total heap storage. (I could not find a way to profile live heap utilization under Hugs instead of total allocation.) The difference in real execution time varied widely however: the LR parser took almost twice as long on smaller files but performed about the same on the largest ones. One probable reason for this variance is the effects of garbage collection. Since a running packrat parser will naturally have a much higher ratio of live data to garbage than an LR parser over time, and garbage collection both increases in overhead cost and decreases in effectiveness (i.e., frees less space) when there is more live data, garbage collection is likely to penalize a packrat parser more than an LR parser as the size of the source file increases. Still, it is encouraging that the packrat parser was able to outperform the LR parser on all but the largest Java source files. 7 Related Work This section briefly relates packrat parsing to relevant prior work. For a more detailed analysis of packrat parsing in comparison with other algorithms please refer to my master’s thesis [9]. 4 http://www.haskell.org/happy 5 http://www.haskell.org/hugs Birman and Ullman [4] first developed the formal properties of deterministic parsing algorithms with backtracking. This work was refined by Aho and Ullman [3] and classified as “top-down limited backtrack parsing,” in reference to the restriction that each parsing function can produce at most one result and hence backtracking is localized. They showed this kind of parser, formally known as a Generalized Top-Down Parsing Language (GTDPL) parser, to be quite powerful. A GTDPL parser can simulate any push-down automaton and thus recognize any LL or LR language, and it can even recognize some languages that are not context free. Nevertheless, all “failures” such as those caused by left recursion can be detected and eliminated from a GTDPL grammar, ensuring that the algorithm is well-behaved. Birman and Ullman also pointed out the possibility of constructing linear-time GTDPL parsers through tabulation of results, but this linear-time algorithm was apparently never put into practice, no doubt because main memories were much more limited at the time and compilers had to operate as streaming “filters” that could run in near-constant space. Adams [1] recently resurrected GTDPL parsing as a component of a modular language prototyping framework, after recognizing its superior composability in comparison with LR algorithms. In addition, many practical top-down parsing libraries and toolkits, including the popular ANTLR [15] and the PARSEC combinator library for Haskell [13], provide similar limited backtracking capabilities which the parser designer can invoke selectively in order to overcome the limitations of predictive parsing. However, all of these parsers implement backtracking in the traditional recursive-descent fashion without memoization, creating the danger of exponential worst-case parse time, and thereby making it impractical to rely on backtracking as a substitute for prediction or to integrate lexical analysis with parsing. The only prior known linear-time parsing algorithm that effectively supports integrated lexical analysis, or “scannerless parsing,” is the NSLR(1) algorithm originally created by Tai [19] and put into practice for this purpose by Salomon and Cormack [18]. This algorithm extends the traditional LR class of algorithms by adding limited support for making lookahead decisions based on nonterminals. The relative power of packrat parsing with respect to NSLR(1) is unclear: packrat parsing is less restrictive of rightward lookahead, but NSLR(1) can also take leftward context into account. In practice, NSLR(1) is probably more space-efficient, but packrat parsing is simpler and cleaner. Other recent scannerless parsers [22, 21] forsake linear-time deterministic algorithms in favor of more general but slower ambiguity-tolerant CFG parsing. 8 Future Work While the results presented here demonstrate the power and practicality of packrat parsing, more experimentation is needed to evaluate its flexibility, performance, and space consumption on a wider variety of languages. For example, languages that rely extensively on parser state, such as C and C++, as well as layout-sensitive languages such as ML and Haskell, may prove more difficult for a packrat parser to handle efficiently. On the other hand, the syntax of a practical language is usually designed with a particular parsing technology in mind. For this reason, an equally compelling question is what new syntax design possibilities are created by the “free” unlimited lookahead and unrestricted grammar composition capabilities of packrat parsing. Section 3.2 suggested a few simple extensions that depend on integrated lexical analysis, but packrat parsing may be even more useful in languages with extensible syntax [7] where grammar composition flexibility is important. Although packrat parsing is simple enough to implement by hand in a lazy functional language, there would still be practical benefit in a grammar compiler along the lines of YACC in the C world or Happy [10] and Mı́mico [6] in the Haskell world. In addition to the parsing functions themselves, the grammar compiler could automatically generate the static “derivations” tuple type and the top-level recursive “tie-up” function, eliminating the problems of monadic representation discussed in Section 3.3. The compiler could also reduce iterative notations such as the popular ‘+’ and ‘*’ repetition operators into a low-level grammar that uses only primitive constant-time operations, preserving the linear parse time guarantee. Finally, the compiler could rewrite left-recursive rules to make it easier to express left-associative constructs in the grammar. One practical area in which packrat parsing may have difficulty and warrants further study is in parsing interactive streams. For example, the “read-eval-print” loops in language interpreters often expect the parser to detect at the end of each line whether or not more input is needed to finish the current statement, and this requirement violates the packrat algorithm’s assumption that the entire input stream is available up-front. A similar open question is under what conditions packrat parsing may be suitable for parsing infinite streams. 9 Conclusion Packrat parsing is a simple and elegant method of converting a backtracking recursive descent parser implemented in a non-strict functional programming language into a linear-time parser, without giving up the power of unlimited lookahead. The algorithm relies for its simplicity on the ability of non-strict functional languages to express recursive data structures with complex dependencies directly, and it relies on lazy evaluation for its practical efficiency. A packrat parser can recognize any language that conventional deterministic linear-time algorithms can and many that they can’t, providing better composition properties and allowing lexical analysis to be integrated with parsing. The primary limitations of the algorithm are that it only supports deterministic parsing, and its considerable (though asymptotically linear) storage requirements. Acknowledgments I wish to thank my advisor Frans Kaashoek, my colleagues Chuck Blake and Russ Cox, and the anonymous reviewers for many helpful comments and suggestions. 10 References [1] Stephen Robert Adams. Modular Grammars for Programming Language Prototyping. PhD thesis, University of Southampton, 1991. [2] Alfred V. Aho, Ravi Sethi, and Jeffrey D. Ullman. Compilers: Principles, Techniques, and Tools. Addison-Wesley, 1986. [3] Alfred V. Aho and Jeffrey D. Ullman. The Theory of Parsing, Translation and Compiling - Vol. I: Parsing. Prentice Hall, Englewood Cliffs, N.J., 1972. [4] Alexander Birman and Jeffrey D. Ullman. Parsing algorithms with backtrack. Information and Control, 23(1):1–34, Aug 1973. [5] Dmitri Bronnikov. Free Yacc-able Java(tm) grammar, 1998. http://home.inreach.com/bronikov/grammars/java.html. [6] Carlos Camarão and Lucı́lia Figueiredo. A monadic combinator compiler compiler. In 5th Brazilian Symposium on Programming Languages, Curitiba – PR – Brazil, May 2001. Universidade Federal do Paraná. [7] Luca Cardelli, Florian Matthes, and Martı́n Abadi. Extensible syntax with lexical scoping. Technical Report 121, Digital Systems Research Center, 1994. [8] Jeroen Fokker. Functional parsers. In Advanced Functional Programming, pages 1–23, 1995. [9] Bryan Ford. Packrat parsing: a practical linear-time algorithm with backtracking. Master’s thesis, Massachusetts Institute of Technology, Sep 2002. [10] Andy Gill and Simon Marlow. Happy: The parser generator for Haskell. http://www.haskell.org/happy. [11] Graham Hutton and Erik Meijer. Monadic parsing in Haskell. Journal of Functional Programming, 8(4):437–444, Jul 1998. [12] Lillian Lee. Fast context-free grammar parsing requires fast boolean matrix multiplication. Journal of the ACM, 2002. To appear. [13] Daan Leijen. Parsec, a fast combinator parser. http://www.cs.uu.nl/˜daan. [14] Terence J. Parr and Russell W. Quong. Adding semantic and syntactic predicates to LL(k): pred-LL(k). In Computational Complexity, pages 263–277, 1994. [15] Terence J. Parr and Russell W. Quong. ANTLR: A predicatedLL(k) parser generator. Software Practice and Experience, 25(7):789–810, 1995. [16] Terence John Parr. Obtaining practical variants of LL(k) and LR(k) for k > 1 by splitting the atomic k-tuple. PhD thesis, Purdue University, Apr 1993. [17] Peter Pepper. LR parsing = grammar transformation + LL parsing: Making LR parsing more understandable and more efficient. Technical Report 99-5, TU Berlin, Apr 1999. [18] Daniel J. Salomon and Gordon V. Cormack. Scannerless NSLR(1) parsing of programming languages. In Proceedings of the ACM SIGPLAN’89 Conference on Programming Language Design and Implementation (PLDI), pages 170–178, Jul 1989. [19] Kuo-Chung Tai. Noncanonical SLR(1) grammars. ACM Transactions on Programming Languages and Systems, 1(2):295–320, Oct 1979. [20] Masaru Tomita. Efficient parsing for natural language. Kluwer Academic Publishers, 1985. [21] M.G.J. van den Brand, J. Scheerder, J.J. Vinju, and E. Visser. Disambiguation filters for scannerless generalized LR parsers. In Compiler Construction, 2002. [22] Eelco Visser. Scannerless generalized-LR parsing. Technical Report P9707, Programming Research Group, University of Amsterdam, 1997. [23] Philip Wadler. How to replace failure by a list of successes: A method for exception handling, backtracking, and pattern matching in lazy functional languages. In Functional Programming Languages and Computer Architecture, pages 113– 128, 1985.
6cs.PL
Lambda-lifting and CPS conversion in an imperative language Gabriel Kerneis Juliusz Chroboczek Université Paris Diderot, PPS, Paris, France arXiv:1202.3247v1 [cs.PL] 15 Feb 2012 February 2012 Abstract This paper is a companion technical report to the article “Continuation-Passing C: from threads to events through continuations”. It contains the complete version of the proofs of correctness of lambda-lifting and CPS-conversion presented in the article. Contents 1 Introduction 2 2 Lambda-lifting in an imperative language 2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1 Naive reduction rules . . . . . . . . . . . . . . . . . . . . 2.1.2 Lambda-lifting . . . . . . . . . . . . . . . . . . . . . . . 2.1.3 Correctness condition . . . . . . . . . . . . . . . . . . . 2.2 Optimised reduction rules . . . . . . . . . . . . . . . . . . . . . 2.2.1 Minimal stores . . . . . . . . . . . . . . . . . . . . . . . 2.2.2 Compact closures . . . . . . . . . . . . . . . . . . . . . . 2.2.3 Optimised reduction rules . . . . . . . . . . . . . . . . . 2.3 Equivalence of optimised and naive reduction rules . . . . . . . 2.3.1 Optimised and intermediate reduction rules equivalence 2.3.2 Intermediate and naive reduction rules equivalence . . . 2.4 Correctness of lambda-lifting . . . . . . . . . . . . . . . . . . . 2.4.1 Strengthened hypotheses . . . . . . . . . . . . . . . . . . 2.4.2 Overview of the proof . . . . . . . . . . . . . . . . . . . 2.4.3 Rewriting lemmas . . . . . . . . . . . . . . . . . . . . . 2.4.4 Aliasing lemmas . . . . . . . . . . . . . . . . . . . . . . 2.4.5 Proof of correctness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 3 4 5 6 6 6 7 7 8 13 21 21 22 23 26 27 3 CPS conversion 3.1 CPS-convertible form . . . . . . . 3.2 Early evaluation . . . . . . . . . 3.3 Small-step reduction . . . . . . . 3.4 CPS terms . . . . . . . . . . . . . 3.5 Correctess of the CPS-conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 32 33 35 36 37 . . . . . . . . . . . . . . . 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Introduction This paper is a companion technical report to the article “Continuation-Passing C: from threads to events through continuations” [4]. It contains the complete version of the proofs presented in the article. It does not, however, give any background or motivation for our work: please refer to the original article. 2 Lambda-lifting in an imperative language To prove the correctness of lambda-lifting in an imperative, call-by-value language when functions are called in tail position, we do not reason directly on CPC programs, because the semantics of C is too broad and complex for our purposes. The CPC translator leaves most parts of converted programs intact, transforming only control structures and function calls. Therefore, we define a simple language with restricted values, expressions and terms, that captures the features we are most interested in (Section 2.1). The reduction rules for this language (Section 2.1.1) use a simplified memory model without pointers and enforce that local variables are not accessed outside of their scope, as ensured by our boxing pass. This is necessary since lambda-lifting is not correct in general in the presence of extruded variables. It turns out that the “naive” reduction rules defined in Section 2.1.1 do not provide strong enough invariants to prove this correctness theorem by induction, mostly because we represent memory with a store that is not invariant with respect to lambda-lifting. Therefore, in Section 2.2, we define an equivalent, “optimised” set of reduction rules which enforces more regular stores and closures. The proof of correctness is then carried out in Section 2.4 using these optimised rules. We first define the invariants needed for the proof and formulate a strengthened version of the correctness theorem (Theorem 2.28, Section 2.4.1). A comprehensive overview of the proof is then given in Section 2.4.2. The proof is fully detailed in Section 2.4.5, with the help of a number of lemmas to keep the main proof shorter (Sections 2.4.3 and 2.4.4). The main limitation of this proof is that Theorems 2.9 and 2.28 are implications, not equivalences: we do not prove that if a term does not reduce, it will not reduce once lifted. For instance, this proof does not ensure that lambda-lifting does not break infinite loops. 2.1 Definitions In this section, we define the terms (Definition 2.1), the reduction rules (Section 2.1.1) and the lambda-lifting transformation itself (Section 2.1.2) for our small imperative language. With these preliminary definitions, we are then able to characterise liftable parameters (Definition 2.8) and state the main correctness theorem (Theorem 2.9, Section 2.1.3). Definition 2.1 (Values, expression and terms). Values are either boolean and integer constants or 1, a special value for functions returning void. v ::= 1 | true | false | n ∈ N Expressions are either values or variables. We deliberately omit arithmetic and boolean operators, with the sole concern of avoiding boring cases in the proofs. e ::= v | x | ... 2 Terms are consist of assignments, conditionals, sequences, recursive functions definitions and calls. T ::= e | x := T | if T then T else T | T ; T | letrec f (x1 . . . xn ) = T in T | f (T, . . . , T ) Our language focuses on the essential details affected by the transformations: recursive functions, conditionals and memory accesses. Loops, for instance, are ignored because they can be expressed in terms of recursive calls and conditional jumps — and that is, in fact, how the splitting pass translates them. Since lambda-lifting happens after the splitting pass, our language need to include inner functions (although they are not part of the C language), but it can safely exclude goto statements. 2.1.1 Naive reduction rules Environments and stores Handling inner functions requires explicit closures in the reduction rules. We need environments, written ρ, to bind variables to locations, and a store, written s, to bind locations to values. Environments and stores are partial functions, equipped with a single operator which extends and modifies a partial function: · + {· 7→ ·}. Definition 2.2. The modification (or extension) f ′ of a partial function f , written f ′ = f +{x 7→ y}, is defined as follows: ( y when t = x f ′ (t) = f (t) otherwise dom(f ′ ) = dom(f ) ∪ {x} Definition 2.3 (Environments of variables and functions). Environments of variables are defined inductively by ρ ::= ε | (x, l) · ρ, i.e. the empty domain function and ρ + {x 7→ l} (respectively). Environments of functions associate function names to closures: F : {f, g, h, . . . } → {[λx1 . . . xn .T, ρ, F ]}. Note that although we have a notion of locations, which correspond roughly to memory addresses in C, there is no way to copy, change or otherwise manipulate a location directly in the syntax of our language. This is on purpose, since adding this possibility would make lambdalifting incorrect: it translates the fact, ensured by the boxing pass in the CPC translator, that there are no extruded variables in the lifted terms. Reduction rules We use classical big-step reduction rules for our language (Figure 1, p. 4). In the (call) rule, we need to introduce fresh locations for the parameters of the called function. This means that we must choose locations that are not already in use, in particular in the environments ρ′ and F . To express this choice, we define two ancillary functions, Env and Loc, to extract the environments and locations contained in the closures of a given environment of functions F . 3 (val) ρ s v −−−→ v (var) s ρ x = l ∈ dom s ρ x s −−−→ s l s F ρ ′ a s −−−→ v s ρ ρ a −−−→ true (if-t.) s′ F b s′ ρ −−−→ v F ρ ′′ F ρ ′′ a ; b s −−−→ v ′ s F ρ ρ ′ b s −−−→ v ′ s F (seq) ′ x := a s −−−→ 1 s +{l7→v} s ′ a s −−−→ v s ρ x = l ∈ dom s′ F (assign) F F s′′ ρ s a −−−→ false ′ ρ ′′ c s −−−→ v s F ρ F (if-f.) ′′ if a then b else c s −−−→ v s s′ ′′ if a then b else c s −−−→ v s F F ρ ′ b s −−−− → vs ′ F (letrec) F ′ = F + {f 7→ [λx1 . . . xn .a, ρ, F ]} ρ ′ letrec f (x1 . . . xn ) = a in b s −−−→ v s F F f = [λx1 . . . xn .b, ρ′ , F ′ ] ∀i, aisi (call) ρ −−−→ F ρ′′ = (x1 , l1 ) · . . . · (xn , ln ) s vi i+1 b sn+1 +{li 7→vi } li fresh and distinct ρ′′ ·ρ′ ′ −−−− −−−−−−−→ v s ′ F +{f 7→F f } f (a1 . . . an ) s1 ρ −−−→ v s′ F Figure 1: “Naive” reduction rules Definition 2.4 (Set of environments, set of locations). [ Env(F ) = {ρ, ρ′ | [λx1 . . . xn .M, ρ, F ′ ] ∈ Im(F ), ρ′ ∈ Env(F ′ )} Loc(F ) = [ {Im(ρ) | ρ ∈ Env(F )} A location l is said to appear in F iff l ∈ Loc(F ). These functions allow us to define fresh locations. Definition 2.5 (Fresh location). In the (call) rule, a location is fresh when: – l∈ / dom(sn+1 ), i.e. l is not already used in the store before the body of f is evaluated, and – l doesn’t appear in F ′ + {f 7→ F f }, i.e. l will not interfere with locations captured in the environment of functions. Note that the second condition implies in particular that l does not appear in either F or ρ′ . 2.1.2 Lambda-lifting Lambda-lifting can be split into two parts: parameter lifting and block floating[2]. We will focus only on the first part here, since the second one is trivial. Parameter lifting consists in adding a free variable as a parameter of every inner function where it appears free. This step is repeated until every variable is bound in every function, and closed functions can safely be floated to top-level. Note that although the transformation is called lambda-lifting, we do not focus on a single function and try to lift all of its free variables; on the contrary, we define the lifting of a single free parameter x in every possible function. 4 Smart lambda-lifting algorithms strive to minimize the number of lifted variables. Such is not our concern in this proof: parameters are lifted in every function where they might potentially be free. Definition 2.6 (Parameter lifting in a term). Assume that x is defined as a parameter of a given function g, and that every inner function in g is called hi (for some i ∈ N). Also assume that function parameters are unique before lambda-lifting. Then the lifted form (M )∗ of the term M with respect to x is defined inductively as follows: (1)∗ = 1 (n)∗ = n (true)∗ = true (f alse)∗ = f alse (y)∗ = y and (y := a)∗ = y := (a)∗ (even if y = x) (a ; b)∗ = (a)∗ ; (b)∗ (if a then b else c)∗ = if (a)∗ then (b)∗ else (c)∗ ( letrec f (x1 . . . xn x) = (a)∗ in (b)∗ if f = hi (letrec f (x1 . . . xn ) = a in b)∗ = letrec f (x1 . . . xn ) = (a)∗ in (b)∗ otherwise ( f ((a1 )∗ , . . . , (an )∗ , x) if f = hi for some i (f (a1 . . . an ))∗ = f ((a1 )∗ , . . . , (an )∗ ) otherwise 2.1.3 Correctness condition We show that parameter lifting is correct for variables defined in functions whose inner functions are called exclusively in tail position. We call these variables liftable parameters. We first define tail positions as usual [1]: Definition 2.7 (Tail position). Tail positions are defined inductively as follows: 1. M and N are in tail position in if P then M else N . 2. N is in tail position in N and M ; N and letrec f (x1 . . . xn ) = M in N . A parameter x defined in a function g is liftable if every inner function in g is called exclusively in tail position. Definition 2.8 (Liftable parameter). A parameter x is liftable in M when: – x is defined as the parameter of a function g, – inner functions in g, named hi , are called exclusively in tail position in g or in one of the hi . Our main theorem states that performing parameter-lifting on a liftable parameter preserves the reduction: Theorem 2.9 (Correctness of lambda-lifting). If x is a liftable parameter in M , then ε ε ε ′ ∃t, M ε −−−→ v t implies ∃t′ , (M )∗ −−−→ v t . ε ε Note that the resulting store t′ changes because lambda-lifting introduces new variables, hence new locations in the store, and changes the values associated with lifted variables; Section 2.4 is devoted to the proof of this theorem. To maintain invariants during the proof, we need to use an equivalent, “optimised” set of reduction rules; it is introduced in the next section. 5 2.2 Optimised reduction rules The naive reduction rules (Section 2.1.1) are not well-suited to prove the correctness of lambdalifting. Indeed, the proof is by induction and requires a number of invariants on the structure of stores and environments. Rather than having a dozen of lemmas to ensure these invariants during the proof of correctness, we translate them as constraints in the reduction rules. To this end, we introduce two optimisations — minimal stores (Section 2.2.1) and compact closures (Section 2.2.2) — which lead to the definition of an optimised set of reduction rules (Figure 2, Section 2.2.3). The equivalence between optimised and naive reduction rules is shown in Section 2.3. 2.2.1 Minimal stores In the naive reduction rules, the store grows faster when reducing lifted terms, because each function call adds to the store as many locations as it has function parameters. This yields stores of different sizes when reducing the original and the lifted term, and that difference cannot be accounted for locally, at the rule level. Consider for instance the simplest possible case of lambda-lifting: letrec g(x) = (letrec h() = x in h()) in g(1) letrec g(x) = (letrec h(y) = y in h(x)) in g(1) (original) (lifted) At the end of the reduction, the store for the original term is {lx 7→ 1} whereas the store for the lifted term is {lx 7→ 1; ly 7→ 1}. More complex terms would yield even larger stores, with many out-of-date copies of lifted variables. To keep the store under control, we need to get rid of useless variables as soon as possible during the reduction. It is safe to remove a variable x from the store once we are certain that it will never be used again, i.e. as soon as the term in tail position in the function which defines x has been evaluated. This mechanism is analogous to the deallocation of a stack frame when a function returns. To track the variables whose location can be safely reclaimed after the reduction of some term M , we introduce split environments. Split environments are written ρT |ρ, where ρT is called the tail environment and ρ the non-tail one; only the variables belonging to the tail environment may be safely reclaimed. The reduction rules build environments so that a variable x belongs to ρT if and only if the term M is in tail position in the current function f and x is a parameter of f . In that case, it is safe to discard the locations associated to all of the parameters of f , including x, after M has been reduced because we are sure that the evaluation of f is completed (and there are no first-class functions in the language to keep references on variables beyond their scope of definition). We also define a cleaning operator, · \ ·, to remove a set of variables from the store. Definition 2.10 (Cleaning of a store). The store s cleaned with respect to the variables in ρ, written s \ ρ, is defined as s \ ρ = s|dom(s)\Im(ρ) . 2.2.2 Compact closures Another source of complexity with the naive reduction rules is the inclusion of useless variables in closures. It is safe to remove from the environments of variables contained in closures the variables that are also parameters of the function: when the function is called, and the environment restored, these variables will be hidden by the freshly instantiated parameters. 6 This is typically what happens to lifted parameters: they are free variables, captured in the closure when the function is defined, but these captured values will never be used since calling the function adds fresh parameters with the same names. We introduce compact closures in the optimised reduction rules to avoid dealing with this hiding mechanism in the proof of lambda-lifting. A compact closure is a closure that does not capture any variable which would be hidden when the closure is called because of function parameters having the same name. Definition 2.11 (Compact closure and environment). A closure [λx1 . . . xn .M, ρ, F ] is compact if ∀i, xi ∈ / dom(ρ) and F is compact. An environment is compact if it contains only compact closures. We define a canonical mapping from any environment F to a compact environment F∗ , restricting the domains of every closure in F . Definition 2.12 (Canonical compact environment). The canonical compact environment F∗ is the unique environment with the same domain as F such that 2.2.3 ∀f ∈ dom(F ), F f = [λx1 . . . xn .M, ρ, F ′ ]   implies F∗ f = λx1 . . . xn .M, ρ|dom(ρ)\{x1 ...xn } , F ′ ∗ . Optimised reduction rules Combining both optimisations yields the optimised reduction rules (Figure 2, p. 8), used Section 2.4 for the proof of lambda-lifting. We ensure minimal stores by cleaning them in the (val), (var) and (assign) rules, which correspond to tail positions; split environments are introduced in the (call) rule to distinguish fresh parameters, to be cleaned, from captured variables, which are preserved. Tail positions are tracked in every rule through split environments, to avoid cleaning variables too early, in a non-tail branch. We also build compact closures in the (letrec) rule by removing the parameters of f from the captured environment ρ′ . Theorem 2.13 (Equivalence between naive and optimised reduction rules). Optimised and naive reduction rules are equivalent: every reduction in one set of rules yields the same result in the other. It is necessary, however, to take care of locations left in the store by the naive reduction: ε|ε M ε ====⇒ v ε ε iff ε ∃s, M ε −−−→ v s ε We prove this theorem in Section 2.3. 2.3 Equivalence of optimised and naive reduction rules This section is devoted to the proof of equivalence between the optimised naive reduction rules (Theorem 2.13). To clarify the proof, we introduce intermediate reduction rules (Figure 3, p. 9), with only one of the two optimisations: minimal stores, but not compact closures. The proof then consists in proving that optimised and intermediate rules are equivalent (Lemma 2.15 and Lemma 2.16, Section 2.3.1), then that naive and intermediate rules are equivalent (Lemma 2.21 and Lemma 2.22, Section 2.3.2). Lemma 2.22 Lemma 2.15 Lemma 2.21 Lemma 2.16 Naive rules −========= − Intermediate rules −========= − Optimised rules 7 (val) (var) ρT |ρ s v =====⇒ v ρT · ρ x = l ∈ dom s ρT |ρ s\ρT x s =====⇒ s l s\ρT F s |ρT ·ρ a =====⇒ v s′ ρT · ρ x = l ∈ dom s F (assign) F ρT |ρ |ρT ·ρ s a =====⇒ true s′ b F (if-t.) s′ ρT |ρ =====⇒ v F ρT |ρ s if a then b else c =====⇒ v ρT |ρ ′′ F s′′ |ρT ·ρ s a =====⇒ false s′ ′ ρT |ρ ′′ c s =====⇒ v s F (if-f.) s′′ s F ρT |ρ ′′ if a then b else c =====⇒ v s F F ρT |ρ s b ====′=⇒ v s′ F ′ (letrec) ′′ F a ; b s =====⇒ v ′ s F |ρT ·ρ ρT |ρ ′ b s =====⇒ v ′ s F (seq) ′ x := a s =====⇒ 1 s +{l7→v}\ρT s ′ a =====⇒ v s ′ F ′ = F + {f 7→ [λx1 . . . xn .a, ρ′ , F ]} ρ = ρT · ρ|dom(ρT ·ρ)\{x1 ...xn } ρT |ρ ′ letrec f (x1 . . . xn ) = a in b s =====⇒ v s F F f = [λx1 . . . xn .b, ρ′ , F ′ ] ∀i, aisi (call) |ρT ·ρ =====⇒ F ρ′′ = (x1 , l1 ) · . . . · (xn , ln ) s vi i+1 b f (a1 . . . an ) sn+1 +{li 7→vi } li fresh and distinct ρ′′ |ρ′ ′ ==== =======⇒ v s ′ F +{f 7→F f } s1 ρT |ρ =====⇒ v s′ \ρT F Figure 2: Optimised reduction rules 2.3.1 Optimised and intermediate reduction rules equivalence In this section, we show that optimised and intermediate reduction rules are equivalent: Lemma 2.15 Intermediate rules −========= − Optimised rules Lemma 2.16 We must therefore show that it is correct to use compact closures in the optimised reduction rules. Compact closures carry the implicit idea that some variables can be safely discarded from the environments when we know for sure that they will be hidden. The following lemma formalises this intuition. Lemma 2.14 (Hidden variables elimination). ρT ·(x,l)|ρ ′ ∀l, l′ , M s ≡≡≡≡≡≡≡≡⇛ v s iff ′ ∀l, l′ , M s ========⇒ v s F ′ F F ρT ·(x,l)|ρ ρT ·(x,l)|(x,l′ )·ρ M s ≡≡≡≡≡≡≡≡≡≡≡≡⇛ v s iff ρT ·(x,l)|(x,l′ )·ρ ′ M s ============⇒ v s F Moreover, both derivations have the same height. Proof. The exact same proof holds for both intermediate and optimised reduction rules. By induction on the structure of the derivation. The proof relies solely on the fact that ρT · (x, l) · ρ = ρT · (x, l) · (x, l′ ) · ρ. 8 (val) (var) ρT |ρ ρT · ρ x = l ∈ dom s ρT |ρ v s ≡≡≡≡≡⇛ v s\ρT x s ≡≡≡≡≡⇛ s l s\ρT F |ρT ·ρ ′ a s ≡≡≡≡≡⇛ v s |ρT ·ρ ′′ a ; b s ≡≡≡≡≡⇛ v ′ s F F ′ ρT |ρ ′ |ρT ·ρ ′′ F F ρT |ρ ′ if a then b else c s ≡≡≡≡≡⇛ v s F ρT |ρ F ρT |ρ ′ b s ≡≡≡≡≡⇛ v s (letrec) F′ F ′ = F + {f 7→ [λx1 . . . xn .a, ρ, F ]} ρT |ρ ′ letrec f (x1 . . . xn ) = a in b s ≡≡≡≡≡⇛ v s F F f = [λx1 . . . xn .b, ρ′ , F ′ ] |ρT ·ρ ∀i, aisi ≡≡≡≡≡⇛ F (call) ρ′′ = (x1 , l1 ) · . . . · (xn , ln ) li fresh and distinct ρ′′ |ρ′ s vi i+1 ′ b sn+1 +{li 7→vi } ≡≡≡≡≡≡≡≡≡≡≡⇛ v s F ′ +{f 7→F ρT |ρ f} ′ f (a1 . . . an ) s1 ≡≡≡≡≡⇛ v s \ρT F Figure 3: Intermediate reduction rules (seq) ρT · (x, l) · ρ = ρT · (x, l) · (x, l′ ) · ρ. So, |ρT ·(x,l)·(x,l′ )·ρ |ρT ·(x,l)·ρ ′ a s ≡≡≡≡≡≡≡≡≡≡≡≡⇛ v s ′ iff a s ≡≡≡≡≡≡≡≡≡⇛ v s F F Moreover, by the induction hypotheses, ′ ρT ·(x,l)|(x,l′ )·ρ ′′ ′ b s ≡≡≡≡≡≡≡≡≡≡≡≡⇛ v ′ s ρT ·(x,l)|ρ ′′ iff b s ≡≡≡≡≡≡≡≡⇛ v ′ s F F Hence, ρT ·(x,l)|(x,l′ )·ρ ρT ·(x,l)|ρ ′′ a ; b s ≡≡≡≡≡≡≡≡≡≡≡≡⇛ v ′ s ′′ iff a ; b s ≡≡≡≡≡≡≡≡⇛ v ′ s F F The other cases are similar. ρT ·(x,l)|ρ (val) v s ≡≡≡≡≡≡≡≡⇛ v s\ρT ·(x,l) F (var) iff ρT ·(x,l)|(x,l′ )·ρ v s ≡≡≡≡≡≡≡≡≡≡≡≡⇛ v s\ρT ·(x,l) F ρT · (x, l) · ρ = ρT · (x, l) · (x, l′ ) · ρ so, with l′′ = ρT · (x, l) · ρ y, ρT ·(x,l)|ρ ′′ if a then b else c s ≡≡≡≡≡⇛ v s F ρ′ = ρT · ρ ′′ c s ≡≡≡≡≡⇛ v s F (if-f.) ′′ ρT |ρ ′ a s ≡≡≡≡≡⇛ false s b s ≡≡≡≡≡⇛ v s a s ≡≡≡≡≡⇛ true s (if-t.) ′′ F ρT |ρ ′ x := a s ≡≡≡≡≡⇛ 1 s +{l7→v}\ρT |ρT ·ρ ρT |ρ ′ b s ≡≡≡≡≡⇛ v ′ s F (seq) ρT |ρ ′ a s ≡≡≡≡≡⇛ v s ρT · ρ x = l ∈ dom s′ F (assign) F y s ≡≡≡≡≡≡≡≡⇛ s l′′ s\ρT ·(x,l) ρT ·(x,l)|(x,l′ )·ρ y s ≡≡≡≡≡≡≡≡≡≡≡≡⇛ s l′′ s\ρT ·(x,l) iff F F 9 (assign) ρT · (x, l) · ρ = ρT · (x, l) · (x, l′ ) · ρ. So, |ρT ·(x,l)·(x,l′ )·ρ |ρT ·(x,l)·ρ ′ ′ iff a s ≡≡≡≡≡≡≡≡≡⇛ v s a s ≡≡≡≡≡≡≡≡≡≡≡≡⇛ v s F F Hence, with l′′ = ρT · (x, l) · ρ y, ρT ·(x,l)|ρ ′ y := a s ≡≡≡≡≡≡≡≡⇛ 1 s +{l ′′ 7→v}\ρT ·(x,l) F ρT ·(x,l)|(x,l′ )·ρ ′ y := a s ≡≡≡≡≡≡≡≡≡≡≡≡⇛ 1 s +{l iff ′′ 7→v}\ρT ·(x,l) F (if-true) and (if-false) are proved similarly to (seq). (letrec) ρT · (x, l) · ρ = ρT · (x, l) · (x, l′ ) · ρ = ρ′ . Moreover, by the induction hypotheses, ρT ·(x,l)|(x,l′ )·ρ ρT ·(x,l)|ρ ′ ′ iff b s ≡≡≡≡≡≡≡≡⇛ v s b s ≡≡≡≡≡≡≡≡≡≡≡≡⇛ v s F′ F′ Hence, ρT ·(x,l)|(x,l′ )·ρ ′ letrec f (x1 . . . xn ) = a in b s ≡≡≡≡≡≡≡≡≡≡≡≡⇛ v s iff F ρT ·(x,l)|ρ ′ letrec f (x1 . . . xn ) = a in b s ≡≡≡≡≡≡≡≡⇛ v s F (call) ρT · (x, l) · ρ = ρT · (x, l) · (x, l′ ) · ρ. So, |ρT ·(x,l)·(x,l′ )·ρ |ρT ·(x,l)·ρ s s iff aisi ≡≡≡≡≡≡≡≡≡⇛ vi i+1 ∀i, aisi ≡≡≡≡≡≡≡≡≡≡≡≡⇛ vi i+1 F F Hence, ρT ·(x,l)|(x,l′ )·ρ ′ f (a1 . . . an ) s1 ≡≡≡≡≡≡≡≡≡≡≡≡⇛ v s \ρT ·(x,l) iff ρT ·(x,l)|ρ ′ f (a1 . . . an ) s1 ≡≡≡≡≡≡≡≡⇛ v s \ρT ·(x,l) . F F Now we can show the required lemmas and prove the equivalence between the intermediate and optimised reduction rules. Lemma 2.15 (Intermediate implies optimised). ρT |ρ ρT |ρ ′ ′ If M s ≡≡≡≡≡⇛ v s then M s =====⇒ v s . F∗ F Proof. By induction on the structure of the derivation. The interesting cases are (letrec) and (call), where compact environments are respectively built and used. (letrec) By the induction hypotheses, ρT |ρ ′ b s ==== =⇒ v s ′ F ∗ Since we defined canonical compact environments so as to match exactly the way compact environments are built in the optimised reduction rules, the constraints of the (letrec) rule are fulfilled: F ′ ∗ = F∗ + {f 7→ [λx1 . . . xn .a, ρ′ , F∗ ]}, hence: ρT |ρ ′ letrec f (x1 . . . xn ) = a in b s =====⇒ v s F∗ 10 (call) By the induction hypotheses, |ρT ·ρ s ∀i, aisi =====⇒ vi i+1 F∗ and ρ′′ |ρ′ ′ b sn+1 +{li 7→vi } ====′=========⇒ v s (F +{f 7→F f })∗ Lemma 2.14 allows to remove hidden variables, which leads to ρ′′ |ρ′| dom(ρ′ )\{x ′ ...xn } ======1====⇒ v s b sn+1 +{li 7→vi } ====== ′ (F +{f 7→F f })∗ Besides, i h F∗ f = λx1 . . . xn .b, ρ′| dom(ρ′ )\{x1 ...xn } , F ′ ∗ and (F ′ + {f 7→ F f })∗ = F ′ ∗ + {f 7→ F∗ f } Hence ρT |ρ ′ f (a1 . . . an ) s1 =====⇒ v s \ρT . F∗ ρT |ρ (val) v s =====⇒ v s\ρT F∗ (var) ρT |ρ x s =====⇒ s l s\ρT F∗ |ρT ·ρ ′ (assign) By the induction hypotheses, a s =====⇒ v s . Hence, F∗ ρT |ρ ′ x := a s =====⇒ 1 s +{l7→v}\ρT F∗ (seq) By the induction hypotheses, |ρT ·ρ ′ ρT |ρ ′ a s =====⇒ v s ′′ b s =====⇒ v ′ s F∗ F∗ Hence, ρT |ρ ′′ a ; b s =====⇒ v ′ s F∗ (if-true) and (if-false) are proved similarly to (seq). Lemma 2.16 (Optimised implies intermediate). ρT |ρ ρT |ρ ′ ′ If M s =====⇒ v s then ∀G such that G∗ = F , M s ≡≡≡≡≡⇛ v s . F G Proof. First note that, since G∗ = F , F is necessarily compact. By induction on the structure of the derivation. The interesting cases are (letrec) and (call), where non-compact environments are respectively built and used. 11 (letrec) Let G such as G∗ = F . Remember that ρ′ = ρT · ρ|dom(ρT ·ρ)\{x1 ...xn } . Let G ′ = G + {f 7→ [λx1 . . . xn .a, ρT · ρ, F ]} which leads, since F is compact (F∗ = F ), to G ′ ∗ = F + {f 7→ [λx1 . . . xn .a, ρ′ , F ]} = F′ By the induction hypotheses, ρT |ρ ′ b s ≡≡≡≡≡⇛ v s G′ Hence, ρT |ρ ′ letrec f (x1 . . . xn ) = a in b s ≡≡≡≡≡⇛ v s G (call) Let G such as G∗ = F . By the induction hypotheses, |ρT ·ρ s ∀i, aisi ≡≡≡≡≡⇛ vi i+1 G Moreover, since G∗ f = F f , G f = [λx1 . . . xn .b, (xi , li ) · . . . · (xj , lj )ρ′ , G ′ ] where G ′ ∗ = F ′ , and the li are some locations stripped out when compacting G to get F . By the induction hypotheses, ρ′′ |ρ′ ′ b sn+1 +{li 7→vi } ≡≡≡≡≡≡≡≡≡≡≡⇛ v s G ′ +{f 7→G Lemma 2.14 leads to f} ρ′′ |(xi ,li )·...·(xj ,lj )ρ′ ′ b sn+1 +{li 7→vi } ≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡≡⇛ v s G ′ +{f 7→G f } Hence, ρT |ρ ′ f (a1 . . . an ) s1 ≡≡≡≡≡⇛ v s \ρT . G ρT |ρ ′ (val) ∀G such as G∗ = F , v s ≡≡≡≡≡⇛ v s G (var) ρT |ρ ∀G such as G∗ = F , x s ≡≡≡≡≡⇛ s l s\ρT G |ρT ·ρ ′ (assign) Let G such as G∗ = F . By the induction hypotheses, a s ≡≡≡≡≡⇛ v s . Hence, G ρT |ρ ′ x := a s ≡≡≡≡≡⇛ 1 s +{l7→v}\ρT G 12 (seq) Let G such as G∗ = F . By the induction hypotheses, |ρT ·ρ ′ ρT |ρ ′ ′′ a s ≡≡≡≡≡⇛ v s b s ≡≡≡≡≡⇛ v ′ s G G Hence ρT |ρ ′′ a ; b s ≡≡≡≡≡⇛ v ′ s G (if-true) and (if-false) are proved similarly to (seq). 2.3.2 Intermediate and naive reduction rules equivalence In this section, we show that the naive and intermediate reduction rules are equivalent: Lemma 2.22 Naive rules −========= − Intermediate rules Lemma 2.21 We must therefore show that it is correct to use minimal stores in the intermediate reduction rules. We first define a partial order on stores: Definition 2.17 (Store extension). s ⊑ s′ s′ |dom(s) = s iff Property 2.18. Store extension (⊑) is a partial order over stores. The following operations preserve this order: · \ ρ and · + {l 7→ v}, for some given ρ, l and v. Proof. Immediate when considering the stores as function graphs: ⊑ is the inclusion, · \ ρ a relative complement, and · + {l 7→ v} a disjoint union (preceded by · \ (l, v ′ ) when l is already bound to some v ′ ). Before we prove that using minimal stores is equivalent to using full stores, we need an alphaconversion lemma, which allows us to rename locations in the store, provided the new location does not already appear in the store or the environments. It is used when choosing a fresh location for the (call) rule in proofs by induction. ρT |ρ ′ Lemma 2.19 (Alpha-conversion). If M s ≡≡≡≡≡⇛ v s then, for all l, for all l′ appearing neither F in s nor in F nor in ρ · ρT , ′ ρT [l′ /l]|ρ[l′ /l] ′ ′ M s[l /l] ≡≡≡≡≡≡≡≡≡≡≡⇛ v s [l /l] . F [l′ /l] Moreover, both derivations have the same height. Proof. By induction on the height of the derivation. For the (call) case, we must ensure that the fresh locations li do not clash with l′ . In case they do, we conclude by applying the induction hypotheses twice: first to rename the clashing li into a fresh li′ , then to rename l into l′ . Two preliminary elementary remarks. First, provided l′ appears neither in ρ or ρT , nor in s, (s \ ρ)[l′ /l] = (s[l′ /l]) \ (ρ[l′ /l]) and (ρT · ρ)[l′ /l] = ρT [l′ /l] · ρ[l′ /l]. 13 ρT |ρ ′ Moreover, if M s ≡≡≡≡≡⇛ v s , then dom(s′ ) = dom(s) \ ρT (straightforward by induction). F This leads to: ρT = ε ⇒ dom(s′ ) = dom(s). By induction on the height of the derivation, because the induction hypothesis must be applied twice in the case of the (call) rule. (call) ∀i, dom(si ) = dom(si+1 ). Thus, ∀i, l′ ∈ / dom(si ). This leads, by the induction hypotheses, to s [l′ /l] ∀i, ai i |(ρT ·ρ)[l′ /l] s ≡≡≡≡≡≡≡≡≡⇛ vi i+1 [l′ /l] F [l′ /l] F Moreover , F ′ is part of F . As a result, since l′ does not appear in F , it does not appear in F ′ , nor in F ′ + {f 7→ F f }. It does not appear in ρ′ either (since ρ′ is part of F ′ ). On the other hand, there might be some j such that lj = l′ , so l′ might appear in ρ′′ . In that case, we apply the induction hypotheses a first time to rename lj in some lj′ 6= l′ . One can chose lj′ such that it does not appear in sn+1 , F ′ + {f 7→ F f } nor in ρ′′ · ρ. As a result, lj′ is fresh. Since lj is fresh too, and does not appear in dom(s′ ) (because of our preliminary remarks), this leads to a mere substitution in ρ′′ : ρ′′ [l′j /lj ]|ρ′ ′ ′ b sn+1 +{li [lj /lj ]7→vi } ≡≡≡≡≡≡≡≡≡≡≡⇛ v s F ′ +{f 7→F f} Once this (potentially) disturbing lj has been renamed (we ignore it in the rest of the proof), we apply the induction hypotheses a second time to rename l to l′ : ρ′′ [l′ /l]|ρ′ [l′ /l] ′ ′ ′ b (sn+1 +{li 7→vi })[l /l] ≡≡≡≡≡≡≡≡≡≡≡⇛ v s [l /l] F ′ +{f 7→F f} Now, (sn+1 + {li 7→ vi })[l′ /l] = sn+1 [l′ /l] + {li 7→ vi }. Moreover, F [l′ /l] f = [λx1 . . . xn .b, ρ′ [l′ /l], F ′ [l′ /l]] and (F ′ + {f 7→ F f })[l′ /l] = F ′ [l′ /l] + {f 7→ F [l′ /l] f } Finally, ρ′′ [l′ /l] = ρ′′ . Hence: ρT [l′ /l]|ρ[l′ /l] ′ ′ ′ ′ f (a1 . . . an ) s1 [l /l] ≡≡≡≡≡≡≡≡≡≡≡⇛ v s [l /l]\ρT [l /l] . F [l′ /l] ′ ρT [l′ /l]|ρ[l′ /l] ′ ′ (val) v s[l /l] ≡≡≡≡≡≡≡≡≡≡≡⇛ v s[l /l]\ρT [l /l] F [l′ /] (var) s[l′ /l](ρT [l′ /l] · ρ[l′ /l] x) = s(ρT · ρ x) = v implies ′ ρT [l′ /l]|ρ[l′ /l] ′ ′ x s[l /l] ≡≡≡≡≡≡≡≡≡≡≡⇛ v s[l /l]\ρT [l /l] F [l′ /] 14 (assign) By the induction hypotheses, |(ρT ·ρ)[l′ /l] ′ ′ ′ a s[l /l] ≡≡≡≡≡≡≡≡≡⇛ v s [l /] F [l′ /] Let s′′ = s′ + {ρT · ρ x 7→ v}. Then, s′ [l′ /l] + {(ρT · ρ)[l′ /l] x 7→ v} = s′′ [l′ /l] Hence ρT [l′ /l]|ρ[l′ /l] ′ ′′ x := a s[l /l] ≡≡≡≡≡≡≡≡≡≡≡⇛ 1 s [l′ /l]\ρT [l′ /l] F [l′ /] (seq) By the induction hypotheses, |(ρT ·ρ)[l′ /l] ′ ′ ′ a s[l /l] ≡≡≡≡≡≡≡≡≡⇛ v s [l /l] F [l′ /] Besides, dom(s′ ) = dom(s), therefore l′ ∈ / dom(s′ ). Then, by the induction hypotheses, ′ ρT [l′ /l]|ρ[l′ /l] ′ ′′ b s [l /l] ≡≡≡≡≡≡≡≡≡≡≡⇛ v ′ s [l′ /l] F [l′ /] Hence ′ ρT [l′ /l]|ρ[l′ /l] ′′ a ; b s[l /l] ≡≡≡≡≡≡≡≡≡≡≡⇛ v ′ s [l′ /l] F [l′ /] (if-true) and (if-false) are proved similarly to (seq). (letrec) Since l′ appears neither in ρ′ nor in F , it does not appear in F ′ either. By the induction hypotheses, ρT [l′ /l]|ρ[l′ /l] ′ ′ ′ b s[l /l] ≡≡≡≡≡≡≡≡≡≡≡⇛ v s [l /l] F ′ [l′ /l] Moreover, F ′ [l′ /l] = F [l′ /l] + {f 7→ [λx1 . . . xn .a, ρ′ [l′ /l], F ]} Hence ρT [l′ /l]|ρ[l′ /l] ′ letrec f (x1 . . . xn ) = a in b s ≡≡≡≡≡≡≡≡≡≡≡⇛ v s F [l′ /] To prove that using minimal stores is correct, we need to extend them so as to recover the full stores of naive reduction. The following lemma shows that extending a store before an (intermediate) reduction extends the resulting store too: Lemma 2.20 (Extending a store in a derivation). ρT |ρ ρT |ρ ′ ′ Given the reduction M s ≡≡≡≡≡⇛ v s , then ∀t ⊒ s, ∃t′ ⊒ s′ , M t ≡≡≡≡≡⇛ v t . F F Moreover, both derivations have the same height. Proof. By induction on the height of the derivation. The most interesting case is (call), which requires alpha-converting a location (hence the induction on the height rather than the structure of the derivation). (var), (val) and (assign) are straightforward by the induction hypotheses and Property 2.18; (seq), (if-true), (if-false) and (letrec) are straightforward by the induction hypotheses. 15 (call) Let t1 ⊒ s1 . By the induction hypotheses, |ρT ·ρ ∃t2 ⊒ s2 , a1t1 ≡≡≡≡≡⇛ v1t2 F |ρT ·ρ t ∃ti+1 ⊒ si+1 , aiti ≡≡≡≡≡⇛ vi i+1 F |ρT ·ρ ∃tn+1 ⊒ sn+1 , antn ≡≡≡≡≡⇛ vntn+1 F The locations li might belong to dom(tn+1 ) and thus not be fresh. By alpha-conversion (Lemma 2.19), we chose fresh li′ (not in Im(ρ′ ) and dom(s′ )) such that (l′i ,vi )|ρ′ ′ ′ b sn+1 +{li 7→vi } ≡≡≡≡≡≡≡≡≡≡≡⇛ v s F ′ +{f 7→F By Property 2.18, tn+1 + {li′ 7→ vi } ⊒ sn+1 + {li′ f} 7→ vi }. By the induction hypotheses, (l′i ,vi )|ρ′ ′ ′ ∃t′ ⊒ s′ , b tn+1 +{li 7→vi } ≡≡≡≡≡≡≡≡≡≡≡⇛ v t F ′ +{f 7→F ′ f} ′ Moreover, t \ ρT ⊒ s \ ρT . Hence, ρT |ρ ′ f (a1 . . . an ) t1 ≡≡≡≡≡⇛ v t \ρT . F (var) ρT |ρ Let t ⊒ s. v t ≡≡≡≡≡⇛ v t\ρT and ∃t′ = t \ ρT ⊒ s \ ρT = s′ (Property 2.18). F ρT |ρ (val) Let t ⊒ s. x t ≡≡≡≡≡⇛ t l t\ρT and ∃t′ = t \ ρT ⊒ s \ ρT = s′ (Property 2.18). Moreover, F t l = s l because l ∈ dom(s) and t|dom(s) = s. (assign) Let t ⊒ s. By the induction hypotheses, |ρT ·ρ ′ ∃t′ ⊒ s′ , a t ≡≡≡≡≡⇛ v t F Hence, ρT |ρ ′ x := a t ≡≡≡≡≡⇛ 1 t +{l7→v}\ρT F concludes, since t′ + {l 7→ v} \ ρT ⊒ t′ + {l 7→ v} \ ρT (Property 2.18). (seq) Let t ⊒ s. By the induction hypotheses, |ρT ·ρ ′ ∃t′ ⊒ s′ , a t ≡≡≡≡≡⇛ v t F ′ ρT |ρ ′′ ∃t′′ ⊒ s′′ , b t ≡≡≡≡≡⇛ v ′ t F Hence, ρT |ρ ′′ ∃t′′ ⊒ s′′ , a ; b t ≡≡≡≡≡⇛ v ′ t F 16 (if-true) and (if-false) are proved similarly to (seq). (letrec) Let t ⊒ s. By the induction hypotheses, ρT |ρ ′ ∃t′ ⊒ s′ , b s ≡≡≡≡≡⇛ v s F′ Hence, ρT |ρ ′ ∃t′ ⊒ s′ , letrec f (x1 . . . xn ) = a in b s ≡≡≡≡≡⇛ v t F Now we can show the required lemmas and prove the equivalence between the intermediate and naive reduction rules. Lemma 2.21 (Intermediate implies naive). ρT |ρ ρT ·ρ ′ ′ If M s ≡≡≡≡≡⇛ v s then ∃t′ ⊒ s′ , M s −−−−−→ v t . F F Proof. By induction on the height of the derivation, because some stores are modified during the proof. The interesting cases are (seq) and (call), where Lemma 2.20 is used to extend intermediary stores. Other cases are straightforward by Property 2.18 and the induction hypotheses. (seq) By the induction hypotheses, ρ ′ ∃t′ ⊒ s′ , a s −−−→ v t . F Moreover, ρT |ρ ′ ′′ b s ≡≡≡≡≡⇛ v ′ s . F Since t′ ⊒ s′ , Lemma 2.20 leads to: ′ ρT |ρ ∃t ⊒ s′′ , b t ≡≡≡≡≡⇛ v ′ t F and the height of the derivation is preserved. By the induction hypotheses, ρ ′ ′′ ∃t′′ ⊒ t, b t −−−→ v ′ t F Hence, since ⊑ is transitive (Property 2.18), ρ ′′ ∃t′′ ⊒ s′′ , a ; b s −−−→ v ′ t . F (call) Similarly to the (seq) case, we apply the induction hypotheses and Lemma 2.20: ρ ∃t2 ⊒ s2 , a1s1 −−−→ v1t2 (Induction) F |ρT ·ρ t′ ∃t′i+1 ⊒ si+1 , aiti ≡≡≡≡≡⇛ vi i+1 (Lemma 2.20) F ρ t ∃ti+1 ⊒ t′i+1 ⊒ si+1 , aiti −−−→ vi i+1 F |ρT ·ρ t′ ∃t′n+1 ⊒ sn+1 , antn ≡≡≡≡≡⇛ vnn+1 (Induction) (Lemma 2.20) F ρ ∃tn+1 ⊒ t′n+1 ⊒ sn+1 , antn −−−→ vntn+1 F 17 (Induction) The locations li might belong to dom(tn+1 ) and thus not be fresh. By alpha-conversion (Lemma 2.19), we choose a set of fresh li′ (not in Im(ρ′ ) and dom(s′ )) such that (l′i ,vi )|ρ′ ′ ′ b sn+1 +{li 7→vi } ≡≡≡≡≡≡≡≡≡≡≡⇛ v s . F ′ +{f 7→F f } By Property 2.18, tn+1 + {li′ 7→ vi } ⊒ sn+1 + {li′ 7→ vi }. Lemma 2.20 leads to, ′ ∃t ⊒ s , b tn+1 +{l′i 7→vi } (l′i ,vi )|ρ′ ≡≡≡≡≡≡≡≡≡≡≡⇛ v t . F ′ +{f 7→F f } By the induction hypotheses, (l′ ,vi )·ρ′ ′ ′ −i−−−−−−→ v t . ∃t′ ⊒ t ⊒ s′ , b tn+1 +{li 7→vi } −−−− ′ F +{f 7→F f } Moreover, t′ \ ρT ⊒ s′ \ ρT . Hence, ρ ′ f (a1 . . . an ) s1 −−−→ v t \ρT . F ρ ′ (val) v s −−−→ v t with t′ = s ⊒ s \ ρT = s′ . F (var) ρ ′′ x s −−−→ s l s with t′ = s ⊒ s \ ρT = s′ . F (assign) By the induction hypotheses, ρ ′ ∃s′′ ⊒ s′ , a s −−−→ v t F Hence, ρ ′ x := a s −−−→ 1 t +{l7→v} F concludes since t′ + {l 7→ v} ⊒ s′ + {l 7→ v} (Property 2.18). (if-true) and (if-false) are proved similarly to (seq). (letrec) By the induction hypotheses, ρ ′ ∃t′ ⊒ s′ , b s −−−− → vs . ′ F Hence, ρ ′ ∃t′ ⊒ s′ , letrec f (x1 . . . xn ) = a in b s −−−→ v t . F The proof of the converse property — i.e. if a term reduces in the naive reduction rules, it reduces in the intermediate reduction rules too — is more complex because the naive reduction rules provide very weak invariants about stores and environments. For that reason, we add an hypothesis to ensure that every location appearing in the environments ρ, ρT and F also appears in the store s: Im(ρT · ρ) ∪ Loc(F ) ⊂ dom(s). Moreover, since stores are often larger in the naive reduction rules than in the intermediate ones, we need to generalise the induction hypothesis. 18 Lemma 2.22 (Naive implies intermediate). Assume Im(ρT · ρ) ∪ Loc(F ) ⊂ dom(s). Then, ′ ρT ·ρ M s −−−−−→ v s implies F ∀t ⊑ s such that Im(ρT · ρ) ∪ Loc(F ) ⊂ dom(t), ρT |ρ ′ M t ≡≡≡≡≡⇛ v s |dom(t)\Im(ρT ) . F Proof. By induction on the structure of the derivation. (val) Let t ⊑ s. Then t \ ρT = s|dom(t)\Im(ρT ) because s|dom(t) = t ′ because s′ = s = s |dom(t)\Im(ρT ) Hence, ρT |ρ v t ≡≡≡≡≡⇛ v t\ρT . F (var) Let t ⊑ s such that Im(ρT · ρ) ∪ Loc(F ) ⊂ dom(t). Note that l ∈ Im(ρT · ρ) ⊂ dom(t) implies t l = s l. Then, t \ ρT = s|dom(t)\Im(ρT ) = s′ |dom(t)\Im(ρT ) because s|dom(t) = t because s′ = s Hence, ρT |ρ x t ≡≡≡≡≡⇛ t l t\ρT . F (assign) Let t ⊑ s such that Im(ρT · ρ) ∪ Loc(F ) ⊂ dom(t). By the induction hypotheses, since Im(ε) = ∅, |ρT ·ρ ′ a t ≡≡≡≡≡⇛ v s |dom(t) F Note that l ∈ Im(ρT · ρ) ⊂ dom(t) implies l ∈ dom(s′ |dom(t) ). Then (s′ |dom(t) + {l 7→ v}) \ ρT = (s′ + {l 7→ v})|dom(t) \ ρT because l ∈ dom(s′ |dom(t) ) ′ = (s + {l 7→ v})|dom(t)\Im(ρT ) Hence, ρT |ρ ′ x := a s ≡≡≡≡≡⇛ 1 (s |dom(t) +{l7→v})\ρT . F (seq) Let t ⊑ s such that Im(ρT · ρ) ∪ Loc(F ) ⊂ dom(t). By the induction hypotheses, since Im(ε) = ∅, |ρT ·ρ ′ a t ≡≡≡≡≡⇛ v s |dom(t) F Moreover, s′ |dom(t) ⊑ s′ and Im(ρT · ρ) ∪ Loc(F ) ⊂ dom(s′ |dom(t) ) = dom(t). By the induction hypotheses, this leads to: ′ ρT |ρ b s |dom(t) ≡≡≡≡≡⇛ v ′ s′′ |dom(s′ | dom(t) )\Im(ρT ) F 19 . Hence, with dom(s′ |dom(t) ) = dom(t), ρT |ρ ′′ a ; b t ≡≡≡≡≡⇛ v ′ s |dom(t)\Im(ρT ) . F (if-true) and (if-false) are proved similarly to (seq). (letrec) Let t ⊑ s such that Im(ρT · ρ) ∪ Loc(F ) ⊂ dom(t). Loc(F ′ ) = Loc(F ) ∪ Im(ρT · ρ) implies Im(ρT · ρ) ∪ Loc(F ′ ) ⊂ dom(t). Then, by the induction hypotheses, ρT |ρ ′ b t ≡≡≡≡≡⇛ v s |dom(t)\Im(ρT ) . F′ Hence, ρT |ρ ′ letrec f (x1 . . . xn ) = a in b t ≡≡≡≡≡⇛ v s |dom(t)\Im(ρT ) . F (call) Let t ⊑ s1 such that Im(ρT · ρ) ∪ Loc(F ) ⊂ dom(t). Note the following equalities: s1 |dom(t) = t s2 |dom(t) ⊑ s2 Im(ρT · ρ) ∪ Loc(F ) ⊂ dom(s2 |dom(t) ) = dom(t) s3 |dom(s2 |dom(t) ) = s3 |dom(t) By the induction hypotheses, they yield: |ρT ·ρ s2 |dom(t) a1t ≡≡≡≡≡⇛ v1 F s2 |dom(t) a2 |ρT ·ρ s3 |dom(t) ≡≡≡≡≡⇛ v1 F si |dom(t) ∀i, ai |ρT ·ρ si+1 |dom(t) ≡≡≡≡≡⇛ vi F Moreover, sn+1 |dom(t) ⊑ sn+1 implies sn+1 |dom(t) + {li 7→ vi } ⊑ sn+1 + {li 7→ vi } (Property 2.18) and: Im(ρ′′ · ρ′ ) ∪ Loc(F ′ + {f 7→ F f }) = Im(ρ′′ ) ∪ (Im(ρ′ ) ∪ Loc(F ′ )) ⊂ {li } ∪ Loc(F ) ⊂ {li } ∪ dom(t) ⊂ dom(sn+1 |dom(t) + {li 7→ vi }) Then, by the induction hypotheses, ρ′′ |ρ′ b sn+1 |dom(t) +{li 7→vi } ≡≡≡≡≡≡≡≡≡≡≡⇛ v F ′ +{f 7→F f} 20 s′ |dom(sn+1 | ′′ dom(t) +{li 7→vi })\Im(ρ ) Finally, s′ |dom(sn+1 |dom(t) +{li 7→vi })\Im(ρ′′ ) \ ρT = s′ |dom(t)∪{li }\{li } \ ρT = s′ |dom(t) \ ρT = (s′ \ ρT )|dom(t)\Im(ρT ) (by definition of · \ ·) Hence, ρT |ρ ′ f (a1 . . . an ) t ≡≡≡≡≡⇛ v (s \ρT )|dom(t)\Im(ρT ) . F 2.4 Correctness of lambda-lifting In this section, we prove the correctness of lambda-lifting (Theorem 2.9, p. 5) by induction on the height of the optimised reduction. Section 2.4.1 defines stronger invariants and rewords the correctness theorem with them. Section 2.4.2 gives an overview of the proof. Sections 2.4.3 and 2.4.4 prove a few lemmas needed for the proof. Section 2.4.5 contains the actual proof of correctness. 2.4.1 Strengthened hypotheses We need strong induction hypotheses to ensure that key invariants about stores and environments hold at every step. For that purpose, we define aliasing-free environments, in which locations may not be referenced by more than one variable, and local positions. They yield a strengthened version of liftable parameters (Definition 2.25). We then define lifted environments (Definition 2.26) to mirror the effect of lambda-lifting in lifted terms captured in closures, and finally reformulate the correctness of lambda-lifting in Theorem 2.28 with hypotheses strong enough to be provable directly by induction. Definition 2.23 (Aliasing). A set of environments E is aliasing-free when: ∀ρ, ρ′ ∈ E, ∀x ∈ dom(ρ), ∀y ∈ dom(ρ′ ), ρ x = ρ′ y ⇒ x = y. By extension, an environment of functions F is aliasing-free when Env(F ) is aliasing-free. The notion of aliasing-free environments is not an artifact of our small language, but translates a fundamental property of the C semantics: distinct function parameters or local variables are always bound to distinct memory locations (Section 6.2.2, paragraph 6 in ISO/IEC 9899 [3]). A local position is any position in a term except inner functions. Local positions are used to distinguish functions defined directly in a term from deeper nested functions, because we need to enforce Invariant 3 (Definition 2.25) on the former only. Definition 2.24 (Local position). Local positions are defined inductively as follows: 1. M is in local position in M , x := M , M ; M , if M then M else M and f (M, . . . , M ). 2. N is in local position in letrec f (x1 . . . xn ) = M in N . We extend the notion of liftable parameter (Definition 2.8, p. 5) to enforce invariants on stores and environments. Definition 2.25 (Extended liftability). The parameter x is liftable in (M, F , ρT , ρ) when: 1. x is defined as the parameter of a function g, either in M or in F , 2. in both M and F , inner functions in g, named hi , are defined and called exclusively: (a) in tail position in g, or (b) in tail position in some hj (with possibly i = j), or 21 3. 4. 5. 6. (c) in tail position in M , for all f defined in local position in M , x ∈ dom(ρT · ρ) ⇔ ∃i, f = hi , moreover, if hi is called in tail position in M , then x ∈ dom(ρT ), in F , x appears necessarily and exclusively in the environments of the hi ’s closures, F contains only compact closures and Env(F ) ∪ {ρ, ρT } is aliasing-free. We also extend the definition of lambda-lifting (Definition 2.6, p. 5) to environments, in order to reflect changes in lambda-lifted parameters captured in closures. Definition 2.26 (Lifted form of an environment). If F f = [λx1 . . . xn .b, ρ′ , F ′ ] then (  ′ λx1 . . . xn x. (b)∗ , ρ |dom(ρ′ )\{x} , (F ′ )∗ (F )∗ f = [λx1 . . . xn . (b)∗ , ρ′ , (F ′ )∗ ] when f = hi for some i otherwise Lifted environments are defined such that a liftable parameter never appears in them. This property will be useful during the proof of correctness. Lemma 2.27. If x is a liftable parameter in (M, F , ρT , ρ), then x does not appear in (F )∗ . Proof. Since x is liftable in (M, F , ρT , ρ), it appears exclusively in the environments of hi . By definition, it is removed when building (F )∗ . These invariants and definitions lead to a correctness theorem with stronger hypotheses. Theorem 2.28 (Correctness of lambda-lifting). If x is a liftable parameter in (M, F , ρT , ρ), then ρT |ρ ρT |ρ ′ ′ M s =====⇒ v s implies (M )∗s =====⇒ v s F (F )∗ Since naive and optimised reductions rules are equivalent (Theorem 2.13, p. 7), the proof of Theorem 2.9 (p. 5) is a direct corollary of this theorem. Corollary 2.29. If x is a liftable parameter in M , then ε ε ε ′ ∃t, M ε −−−→ v t implies ∃t′ , (M )∗ −−−→ v t . ε 2.4.2 ε Overview of the proof With the enhanced liftability definition, we have invariants strong enough to perform a proof by induction of the correctness theorem. This proof is detailed in Section 2.4.5. The proof is not by structural induction but by induction on the height of the derivation. This is necessary because, even with the stronger invariants, we cannot apply the induction hypotheses directly to the premises in the case of the (call) rule: we have to change the stores and environments, which means rewriting the whole derivation tree, before using the induction hypotheses. To deal with this most difficult case, we distinguish between calling one of the lifted functions (f = hi ) and calling another function (either g, where x is defined, or any other function outside of g). Only the former requires rewriting; the latter follows directly from the induction hypotheses. In the (call) rule with f = hi , issues arise when reducing the body b of the lifted function. During this reduction, indeed, the store contains a new location l′ bound by the environment to the lifted variable x, but also contains the location l which contains the original value of x. Our 22 goal is to show that the reduction of b implies the reduction of (b)∗ , with store and environments fulfilling the constraints of the (call) rule. To obtain the reduction of the lifted body (b)∗ , we modify the reduction of b in a series of steps, using several lemmas: – the location l of the free variable x is moved to the tail environment (Lemma 2.30); – the resulting reduction meets the induction hypotheses, which we apply to obtain the reduction of the lifted body (b)∗ ; – however, this reduction does not meet the constraints of the optimised reduction rules because the location l is not fresh: we rename it to a fresh location l′ to hold the lifted variable (Lemma 2.31); – finally, since we renamed l to l′ , we need to reintroduce a location l to hold the original value of x (Lemmas 2.32 and 2.33). The rewriting lemmas used in the (call) case are shown in Section 2.4.3. For every other case, the proof consists in checking thoroughly that the induction hypotheses apply, in particular that x is liftable in the premises. These verifications consist in checking Invariants 3 to 6 of the extended liftability definition (Definition 2.25) — Invariants 1 and 2 are obvious enough not to be detailed. To keep the main proof as compact as possible, the most difficult cases of liftability, related to aliasing, are proven in some preliminary lemmas (Section 2.4.4). One last issue arises during the induction when one of the premises does not contain the lifted variable x. In that case, the invariants do not hold, since they assume the presence of x. But it turns out that in this very case, the lifting function is the identity (since there is no variable to lift) and lambda-lifting is trivially correct. 2.4.3 Rewriting lemmas Calling a lifted function has an impact on the resulting store: new locations are introduced for the lifted parameters and the earlier locations, which are not modified anymore, are hidden. Because of these changes, the induction hypotheses do not apply directly in the case of the (call) rule for a lifted function hi . We use the following four lemmas to obtain, through several rewriting steps, a reduction of lifted terms meeting the induction hypotheses. – Lemma 2.30 shows that moving a variable from the non-tail environment ρ to the tail environment ρT does not change the result, but restricts the domain of the store. It is used transform the original free variable x (in the non-tail environment) to its lifted copy (which is a parameter of hi , hence in the tail environment). – Lemma 2.31 handles alpha-conversion in stores and is used when choosing a fresh location. – Lemmas 2.32 and 2.33 finally add into the store and the environment a fresh location, bound to an arbitrary value. It is used to reintroduce the location containing the original value of x, after it has been alpha-converted to l′ . ρT |(x,l)·ρ ′ Lemma 2.30 (Switching to tail environment). If M s ========⇒ v s and x ∈ / dom(ρT ) then F ρT ·(x,l)|ρ ′ M s ========⇒ v s |dom(s′ )\{l} . Moreover, both derivations have the same height. F Proof. By induction on the structure of the derivation. For the (val), (var), (assign) and (call) cases, we use the fact that s \ ρT · (x, l) = s′ |dom(s′ )\{l} when s′ = s \ ρT . ρT ·(x,l)|ρ (val) v s ========⇒ v s\ρT ·(x,l) and s \ ρT · (x, l) = s′ |dom(s′ )\{l} with s′ = s \ ρT . F 23 (var) ρT ·(x,l)|ρ y s ========⇒ s l′ s\ρT ·(x,l) and s \ ρT · (x, l) = s′ |dom(s′ )\{l} , with l′ = ρT · (x, l) · ρ y and s ′ = s \ ρT . F |ρT ·(x,l)·ρ ρT ·(x,l)|ρ ′ ′ ′ (assign) By hypothesis, a s =========⇒ v s hence y := a s ========⇒ 1 s +{l 7→v}\ρT ·(x,l) and F F s′ + {l′ 7→ v} \ ρT · (x, l) = s′ |dom(s′ )\{l} with l′ = ρT · (x, l) · ρ y and s′ = s′ + {l′ 7→ v} \ ρT . |ρT ·(x,l)·ρ ′ ′ ρT ·(x,l)|ρ (seq) By hypothesis, a s =========⇒ v s and, by the induction hypotheses, b s ========⇒ ′′ vs F |dom(s′′ )\{l} F hence ρT ·(x,l)|ρ ′′ a ; b s ========⇒ v s |dom(s′′ )\{l} F . (if-true) and (if-false) are proved similarly to (seq). (letrec) By the induction hypotheses, ρT ·(x,l)|ρ ′ b s ====== ==⇒ v s |dom(s′ )\{l} ′ F hence ρT ·(x,l)|ρ ′ letrec f (x1 . . . xn ) = a in b s ========⇒ v s |dom(s′ )\{l} F (call) The hypotheses do not change, and the conclusion becomes: ρT ·(x,l)|ρ ′ f (a1 . . . an ) s1 ========⇒ v s \ρT ·(x,l) F as expected, since s′ \ ρT · (x, l) = s′′ |dom(s′′ )\{l} with s′′ = s′ \ ρT ρT |ρ ′ Lemma 2.31 (Alpha-conversion). If M s =====⇒ v s then, for all l, for all l′ appearing neither F in s nor in F nor in ρ · ρT , ′ ρT [l′ /l]|ρ[l′ /l] ′ ′ M s[l /l] ======= ====⇒ v s [l /l] ′ F [l /l] Moreover, both derivations have the same height. Proof. See Lemma 2.19, p. 2.19. ρT |ρ ′ Lemma 2.32 (Spurious location in store). If M s =====⇒ v s and k does not appear in either F ρT |ρ ′ s, F or ρT · ρ, then, for all value u, M s+{k7→u} =====⇒ v s +{k7→u} . Moreover, both derivations F have the same height. Proof. By induction on the height of the derivation. The key idea is to add (k, u) to every store in the derivation tree. A collision might occur in the (call) rule, if there is some j such that lj = k. In that case, we need to rename lj to some fresh variable lj′ 6= k (by alpha-conversion) before applying the induction hypotheses. 24 (call) By the induction hypotheses, s +{k7→u} ∀i, ai i |ρT ·ρ s =====⇒ vi i+1 +{k7→u} F Because k does not appear in F , k∈ / Loc(F ′ + {f 7→ F f }) ⊂ Loc(F ) For the same reason, it does not appear in ρ′ . On the other hand, there might be a j such that lj = k, so k might appear in ρ′′ . In that case, we rename lj in some fresh lj′ 6= k, appearing in neither sn+1 , nor F ′ or ρ′′ · ρ′ (Lemma 2.31). After this alpha-conversion, k does not appear in either ρ′′ · ρ′ , F ′ + {f 7→ F f }, or sn+1 + {li 7→ vi }. By the induction hypotheses, ρ′′ |ρ′ ′ b sn+1 +{li 7→vi }+{k7→u} ==== =======⇒ v s +{k7→u} ′ F +{f 7→F f } Moreover, s′ + {k 7→ u} \ ρT = s′ \ ρT + {k 7→ u} (since k does not appear in ρT ). Hence ρT |ρ ′ f (a1 . . . an ) s1 +{k7→u} =====⇒ v s +{k7→u}\ρT . F ρT |ρ (val) v s+{k7→u} =====⇒ v s+{k7→u}\ρT and s + {k 7→ u} \ ρT = s \ ρT + {k 7→ u} since k does not appear in ρT . (var) F ρT |ρ x s+{k7→u} =====⇒ (s + {k 7→ u}) l s+{k7→u}\ρT , with s + {k 7→ u} \ ρT = s \ ρT + {k 7→ u} F since k does not appear in ρT , and (s + {k 7→ u}) l = s l since k 6= l (k does not appear in s). |ρT ·ρ ′ (assign) By the induction hypotheses, a s+{k7→u} =====⇒ v s +{k7→u} . And k 6= l (since k does F not appear in s) then s′ + {k 7→ u} + {l 7→ v} = s′ + {l 7→ v} + {k 7→ u}. Moreover, k does not appear in ρT then s′ + {l 7→ v} + {k 7→ u} \ ρT = s′ + {l 7→ v} \ ρT + {k 7→ u}. Hence ρT |ρ ′ x := a s+{k7→u} =====⇒ 1 s +{l7→v}\ρT +{k7→u} F (seq) By the induction hypotheses, |ρT ·ρ ′ a s+{k7→u} =====⇒ true s +{k7→u} F ′ ρT |ρ ′′ b s +{k7→u} =====⇒ v ′ s +{k7→u} F Hence ρT |ρ ′′ a ; b s+{k7→u} =====⇒ v ′ s F (if-true) and (if-false) are proved similarly to (seq). 25 +{k7→u} (letrec) The location k does not appear in F ′ , because it does not appear in either F or ρ′ ⊂ ρT · ρ (F ′ = F + {f 7→ [λx1 . . . xn .a, ρ′ , F ]}). Then, by the induction hypotheses, ρT |ρ ′ b s+{k7→u} ====′=⇒ v s +{k7→u} F Hence ρT |ρ ′ letrec f (x1 . . . xn ) = a in b s+{k7→u} =====⇒ v s +{k7→u} . F Lemma 2.33 (Spurious variable in environments). ρT ·(x,l)|ρ ′ ∀l, l′ , M s ========⇒ v s F iff ρT ·(x,l)|(x,l′ )·ρ ′ M s ============⇒ v s F Moreover, both derivations have the same height. Proof. See Lemma 2.14, p. 2.14. 2.4.4 Aliasing lemmas We need three lemmas to show that environments remain aliasing-free during the proof by induction in Section 2.4.5. The first lemma states that concatenating two environments in an aliasing-free set yields an aliasing-free set. The other two prove that the aliasing invariant (Invariant 6, Definition 2.25) holds in the context of the (call) and (letrec) rules, respectively. Lemma 2.34 (Concatenation). If E ∪ {ρ, ρ′ } is aliasing-free then E ∪ {ρ · ρ′ } is aliasing-free. Proof. By exhaustive check of cases. We want to prove ∀ρ1 , ρ2 ∈ E ∪ {ρ · ρ′ }, ∀x ∈ dom(ρ1 ), ∀y ∈ dom(ρ2 ), ρ1 x = ρ2 y ⇒ x = y. given that ∀ρ1 , ρ2 ∈ E ∪ {ρ, ρ′ }, ∀x ∈ dom(ρ1 ), ∀y ∈ dom(ρ2 ), ρ1 x = ρ2 y ⇒ x = y. If ρ1 ∈ E and ρ2 ∈ E, immediate. If ρ1 ∈ {ρ · ρ′ }, ρ1 x = ρ x or ρ′ x. This is the same for ρ2 . Then ρ1 x = ρ2 y is equivalent to ρ x = ρ′ y (or some other combination, depending on x, y, ρ1 and ρ2 ) which leads to the expected result. Lemma 2.35 (Aliasing in (call) rule). Assume that, in a (call) rule, – F f = [λx1 . . . xn .b, ρ′ , F ′ ], – Env(F ) is aliasing-free, and – ρ′′ = (x1 , l1 ) · . . . · (xn , ln ), with fresh and distinct locations li . Then Env(F ′ + {f 7→ F f }) ∪ {ρ′ , ρ′′ } is also aliasing-free. Proof. Let E = Env(F ′ + {f 7→ F f }) ∪ {ρ′ }. We know that E ⊂ Env(F ) so E is aliasing-free We want to show that adding fresh and distinct locations from ρ′′ preserves this lack of freedom. More precisely, we want to show that ∀ρ1 , ρ2 ∈ E ∪ {ρ′′ }, ∀x ∈ dom(ρ1 ), ∀y ∈ dom(ρ2 ), ρ1 x = ρ2 y ⇒ x = y given that ∀ρ1 , ρ2 ∈ E, ∀x ∈ dom(ρ1 ), ∀y ∈ dom(ρ2 ), ρ1 x = ρ2 y ⇒ x = y. We reason by checking of all cases. If ρ1 ∈ E and ρ2 ∈ E, immediate. If ρ1 = ρ2 = ρ′′ then ρ′′ x = ρ′′ y ⇒ x = y holds because the locations of ρ′′ are distinct. If ρ1 = ρ′′ and ρ2 ∈ E then ρ1 x = ρ2 y ⇒ x = y holds because ρ1 x 6= ρ2 y (by freshness hypothesis). 26 Lemma 2.36 (Aliasing in (letrec) rule). If Env(F ) ∪ {ρ, ρT } is aliasing free, then, for all xi , Env(F ) ∪ {ρ, ρT } ∪ {ρT · ρ |dom(ρT ·ρ)\{x1 ...xn } } is aliasing free. Proof. Let E = Env(F ) ∪ {ρ, ρT } and ρ′′ = ρT · ρ|dom(ρT ·ρ)\{x1 ...xn } . Adding ρ′′ , a restricted concatenation of ρT and ρ, to E preserves aliasing freedom, as in the proof of Lemma 2.34. If ρ1 ∈ E and ρ2 ∈ E, immediate. If ρ1 ∈ {ρ′′ }, ρ1 x = ρ x or ρ′ x. This is the same for ρ2 . Then ρ1 x = ρ2 y is equivalent to ρ x = ρ′ y (or some other combination, depending on x, y, ρ1 and ρ2 ) which leads to the expected result. 2.4.5 Proof of correctness We finally show Theorem 2.28. Theorem 2.28. If x is a liftable parameter in (M, F , ρT , ρ), then ρT |ρ ′ s ρT |ρ ′ M s =====⇒ v s implies (M )∗ =====⇒ v s F (F )∗ Assume that x is a liftable parameter in (M, F , ρT , ρ). The proof is by induction on the height ρT |ρ ′ of the reduction of M s =====⇒ v s . To keep the proof readable, we detail only the non-trivial F cases when checking the invariants of Definition 2.25 to ensure that the induction hypotheses hold. (call) — first case First, we consider the most interesting case where there exists i such that f = hi . The variable x is a liftable parameter in (hi (a1 . . . an ), F , ρT , ρ) hence in (ai , F , ε, ρT · ρ) too. Indeed, the invariants of Definition 2.25 hold: – Invariant 3: By definition of a local position, every f defined in local position in ai is in local position in hi (a1 . . . an ), hence the expected property by the induction hypotheses. – Invariant 4: Immediate since the premise does not hold : since the ai are not in tail position in hi (a1 . . . an ), they cannot feature calls to hi (by Invariant 2). – Invariant 6: Lemma 2.34, p. 26. The other invariants hold trivially. By the induction hypotheses, we get (ai )∗ si |ρT ·ρ s =====⇒ vi i+1 . (F )∗ By definition of lifting, (hi (a1 . . . an ))∗ = hi ((a1 )∗ , . . . , (an )∗ , x). But x is not a liftable parameter in (b, F ′ , ρ′′ , ρ′ ) since the Invariant 4 might be broken: x ∈ / dom(ρ′′ ) (x is not a parameter of hi ) but hj might appear in tail position in b. On the other hand, we have x ∈ dom(ρ′ ): since, by hypothesis, x is a liftable parameter in (hi (a1 . . . an ), F , ρT , ρ), it appears necessarily in the environments of the closures of the hi , such as ρ′ . This allows us to split ρ′ into two parts: ρ′ = (x, l) · ρ′′′ . It is then possible to move (x, l) to the tail environment, according to Lemma 2.30: ρ′′ (x,l)|ρ′′′ ′ b sn+1 +{li 7→vi } ==== =======⇒ v s |dom(s′ )\{l} ′ F +{f 7→F f } This rewriting ensures that x is a liftable parameter in (b, F ′ + {f 7→ F f }, ρ′′ · (x, l), ρ′′′ ). Indeed, the invariants of Definition 2.25 hold: 27 – Invariant 3: Every function defined in local position in b is an inner function in hi so, by Invariant 2, it is one of the hi and x ∈ dom(ρ′′ · (x, l) · ρ′′′ ). – Invariant 4: Immediate since x ∈ dom(ρ′′ · (x, l) · ρ′′′ ). – Invariant 5: Immediate since F ′ is included in F . – Invariant 6: Immediate for the compact closures. Aliasing freedom is guaranteed by Lemma 2.35 (p. 26). The other invariants hold trivially. By the induction hypotheses, ρ′′ (x,l)|ρ′′′ ′ (b)∗sn+1 +{li 7→vi } ===== ========⇒ v s |dom(s′ )\{l} ′ (F +{f 7→F f })∗ The l location is not fresh: it must be rewritten into a fresh location, since x is now a parameter of hi . Let l′ be a location appearing in neither (F ′ + {f 7→ F f })∗ , nor sn+1 + {li 7→ vi } or ρ′′ · ρT ′ . Then l′ is a fresh location, which is to act as l in the reduction of (b)∗ . We will show that, after the reduction, l′ is not in the store (just like l before the lambdalifting). In the meantime, the value associated to l does not change (since l′ is modified instead of l). Lemma 2.27 implies that x does not appear in the environments of (F )∗ , so it does not appear in the environments of (F ′ + {f 7→ F f })∗ ⊂ (F )∗ either. As a consequence, lack of aliasing implies by Definition 2.23 that the label l, associated to x, does not appear in (F ′ + {f 7→ F f })∗ either, so (F ′ + {f 7→ F f })∗ [l′ /l] = (F ′ + {f 7→ F f })∗ . Moreover, l does not appear in s′ |dom(s′ )\{l} . By alpha-conversion (Lemma 2.31, since l′ does not appear in the store or the environments of the reduction, we rename l to l′ : s (b)∗ n+1 [l′ /l]+{li 7→vi } ρ′′ (x,l′ )|ρ′′′ ′ ===== ========⇒ v s |dom(s′ )\{l} . ′ (F +{f 7→F f })∗ We want now to reintroduce l. Let vx = sn+1 l. The location l does not appear in sn+1 [l′ /l] + {li 7→ vi }, (F ′ + {f 7→ F f })∗ , or ρ′′ (x, l′ ) · ρ′′′ . Thus, by Lemma 2.32, s (b)∗ n+1 [l′ /l]+{li 7→vi }+{l7→vx } ρ′′ (x,l′ )|ρ′′′ ′ ===== ========⇒ v s |dom(s′ )\{l} +{l7→vx } . ′ (F +{f 7→F f })∗ Since sn+1 [l′ /l] + {li 7→ vi } + {l 7→ vx } = sn+1 [l′ /l] + {l 7→ vx } + {li 7→ vi } = sn+1 + {l′ 7→ vx } + {li 7→ vi } = sn+1 + {li 7→ vi } + {l′ 7→ vx } because ∀i, l 6= li because vx = sn+1 l because ∀i, l′ 6= li and s′ |dom(s′ )\{l} + {l 7→ vx } = s′ + {l 7→ vx }, we finish the rewriting by Lemma 2.33, s (b)∗ n+1 +{li 7→vi }+{l′ 7→vx } ρ′′ (x,l′ )|(x,l)·ρ′′′ ====′=========⇒ v (F +{f 7→F f })∗ 28 s′ +{l7→vx } . Hence the result: (F )∗ hi = [λx1 . . . xn x. (b)∗ , ρ′ , (F ′ )∗ ] ρ = (x1 , l1 ) · . . . · (xn , ln )(x, ρT x) l′ and li fresh and distinct ′′ si ∀i, (ai )∗ (x)∗ (call) sn+1 |ρT ·ρ s =====⇒ vxsn+1 (b)∗ n+1 (F )∗ |ρT ·ρ s =====⇒ vi i+1 (F )∗ +{li 7→vi }+{l′ 7→vx } (hi (a1 . . . an ))∗ ρ′′ (x,l′ )|ρ′ ====′=========⇒ v s′ +{l7→vx } (F +{f 7→F f })∗ s1 ρT |ρ =====⇒ v s′ +{l7→vx }\ρT (F )∗ Since l ∈ dom(ρT ) (because x is a liftable parameter in (hi (a1 . . . an ), F , ρT , ρ)), the extraneous location is reclaimed as expected: s′ + {l 7→ vx } \ ρT = s′ \ ρT . (call) — second case We now consider the case where f is not one of the hi . The variable x is a liftable parameter in (f (a1 . . . an ), F , ρT , ρ) hence in (ai , F , ε, ρT · ρ) too. Indeed, the invariants of Definition 2.25 hold: – Invariant 3: By definition of a local position, every f defined in local position in ai is in local position in f (a1 . . . an ), hence the expected property by the induction hypotheses. – Invariant 4: Immediate since the premise does not hold : the ai are not in tail position in f (a1 . . . an ) so they cannot feature calls to hi (by Invariant 2:). – Invariant 6: Lemma 2.34, p. 26. The other invariants hold trivially. By the induction hypotheses, we get (ai )∗ si |ρT ·ρ s =====⇒ vi i+1 , (F )∗ and, by Definition 2.6, (f (a1 . . . an ))∗ = f ((a1 )∗ , . . . , (an )∗ ). If x is not defined in b or F , then ()∗ is the identity function and can trivially be applied to the reduction of b. Otherwise, x is a liftable parameter in (b, F ′ + {f 7→ F f }, ρ′′ , ρ′ ). Indeed, the invariants of Definition 2.25 hold. Assume that x is defined as a parameter of some function g, in either b or F : – Invariant 3: We have to distinguish the cases where f = g (with x ∈ dom(ρ′′ )) and f 6= g (with x ∈ / dom(ρ′′ ) and x ∈ / dom(ρ′ )). In both cases, the result is immediate by the induction hypotheses. – Invariant 4: If f 6= g, the premise cannot hold (by the induction hypotheses, Invariant 2). If f = g, x ∈ dom(ρ′′ ) (by the induction hypotheses, Invariant 2). – Invariant 5: Immediate since F ′ is included in F . – Invariant 6: Immediate for the compact closures. Aliasing freedom is guaranteed by Lemma 2.35 (p. 26). The other invariants hold trivially. By the induction hypotheses, s (b)∗ n+1 +{li 7→vi } ρ′′ |ρ′ ′ ====′=========⇒ v s (F +{f 7→F f })∗ 29 hence: (F )∗ f = [λx1 . . . xn . (b)∗ , ρ′ , (F ′ )∗ ] ∀i, (ai )∗ (call) si |ρT ·ρ =====⇒ (F )∗ ρ′′ = (x1 , l1 ) · . . . · (xn , ln ) s vi i+1 s +{l 7→v } (b)∗ n+1 i i ρT |ρ li fresh and distinct ρ′′ |ρ′ ′ ====′=========⇒ v s (F +{f 7→F f })∗ ′ (f (a1 . . . an ))∗ s1 =====⇒ v s \ρT (F )∗ (letrec) The parameter x is a liftable in (letrec f (x1 . . . xn ) = a in b, F , ρT , ρ) so x is a liftable parameter in (b, F ′ , ρT , ρ) too. Indeed, the invariants of Definition 2.25 hold: – Invariants 3 and 4: Immediate by the induction hypotheses and definition of tail and local positions. – Invariant 5: By the induction hypotheses, Invariant 3 (x is to appear in the new closure if and only if f = hi ). – Invariant 6: Lemma 2.36 (p. 27). The other invariants hold trivially. By the induction hypotheses, we get s ρT |ρ ′ (b)∗ ====′==⇒ v s . (F )∗ If f 6= hi , (letrec f (x1 . . . xn ) = a in b)∗ = letrec f (x1 . . . xn ) = (a)∗ in (b)∗ hence, by definition of (F ′ )∗ , s ρT |ρ ′ (b)∗ ====′==⇒ v s (letrec) ρ′ = ρT · ρ|dom(ρT ·ρ)\{x1 ...xn } (F )∗ (F ′ )∗ = (F )∗ +{f 7→ [λx1 . . . xn . (a)∗ , ρ′ , F ]} s ρT |ρ ′ (letrec f (x1 . . . xn ) = a in b)∗ =====⇒ v s (F )∗ On the other hand, if f = hi , (letrec f (x1 . . . xn ) = a in b)∗ = letrec f (x1 . . . xn x) = (a)∗ in (b)∗ hence, by definition of (F ′ )∗ , s ρT |ρ ′ (b)∗ ====′==⇒ v s (letrec) ρ′ = ρT · ρ|dom(ρT ·ρ)\{x1 ...xn x} (F )∗ (F ′ )∗ = (F )∗ +{hi 7→ [λx1 . . . xn x. (a)∗ , ρ′ , F ]} ρT |ρ (F )∗ (val) (v)∗ = v so ′ (letrec hi (x1 . . . xn ) = a in b)∗ s =====⇒ v s (val) s ρT |ρ (v)∗ =====⇒ v s\ρT (F )∗ 30 (var) (y)∗ = y so (var) ρT · ρ y = l ∈ dom s ρT |ρ s (y)∗ =====⇒ s l s\ρT (F )∗ (assign) The parameter x is liftable in (y := a, F , ρT , ρ) so in (a, F , ε, ρT · ρ) too. Indeed, the invariants of Definition 2.25 hold: – Invariant 6: Lemma 2.34, p. 26. The other invariants hold trivially. By the induction hypotheses, we get |ρT ·ρ ′ (a)∗ s =====⇒ v s . (F )∗ Moreover (y := a)∗ = y := (a)∗ , so : |ρT ·ρ ′ (a)∗ s =====⇒ v s (assign) ρT · ρ y = l ∈ dom s′ (F )∗ ρT |ρ ′ s (y := a)∗ =====⇒ 1 s +{l7→v}\ρT (F )∗ (seq) The parameter x is liftable in (a ; b, F , ρT , ρ). If x is not defined in a or F , then ()∗ is the identity function and can trivially be applied to the reduction of a. Otherwise, x is a liftable parameter in (a, F , ε, ρT · ρ). Indeed, the invariants of Definition 2.25 hold: – Invariant 6: Lemma 2.34, p. 26. The other invariants hold trivially. If x is not defined in b or F , then ()∗ is the identity function and can trivially be applied to the reduction of b. Otherwise, x is a liftable parameter in (b, F , ρT , ρ). Indeed, the invariants of Definition 2.25 hold trivially. s |ρT ·ρ ′ By the induction hypotheses, we get (a)∗ =====⇒ v s and (b)∗ s′ (F )∗ Moreover, (a ; b)∗ = (a)∗ ; (b)∗ , hence: s |ρT ·ρ ′ (a)∗ =====⇒ v s (seq) (b)∗ (F )∗ s s′ ρT |ρ ρT |ρ (F )∗ ′′ (a ; b)∗ =====⇒ v ′ s (F )∗ (if-true) and (if-false) are proved similarly to (seq). 31 ′′ =====⇒ v ′ s ρT |ρ ′′ =====⇒ v ′ s . (F )∗ 3 CPS conversion In this section, we prove the correctness of the CPS-conversion performed by the CPC translator. This conversion is defined only on a subset of C programs that we call CPS-convertible terms (Section 3.1). We first show that the early evaluation of function parameters in CPS-convertible terms is correct (Section 3.2). To simplify the proof of correctness of CPS-conversion, we then introduce small-step reduction rules featuring contexts and early evaluation (Section 3.3). In Section 3.4, we define CPS terms, with the push and invoke operators to build and execute continuations, and the associated reduction rules. Since the syntax of CPS-terms does not ensure a correct reduction, we also define well-formed CPS-terms, which are the image of CPS-convertible terms by CPS-conversion. The proof of correctness of CPS-conversion is finally carried out in Section 3.5. It consists merely in checking that the reduction rules for CPS-convertible terms and well-formed CPS-terms execute in lock-step. 3.1 CPS-convertible form CPS conversion is not defined for every C function; instead, we restrict ourselves to a subset of functions, which we call the CPS-convertible subset. The CPS-convertible form restricts the calls to cps functions to make it straightforward to capture their continuation. In CPS-convertible form, a call to a cps function f is either in tail position, or followed by a tail call to another cps function whose parameters are non-shared variables that cannot be modified by f. In the C language, we define the CPS-convertible form as follows: Definition 3.1 (CPS-convertible form). A function h is in CPS-convertible form if every call to a cps function that it contains matches one of the following patterns, where both f and g are cps functions, e1 , ..., en are any C expressions and x, y1 , ..., yn are distinct, non-shared variables: return f(e1 , ..., en ); (1) x = f(e1 , ..., en ); return g(x, y1 , ..., yn ); f(e1 , ..., en ); return g(x, y1 , ..., yn ); (2) (3) f(e1 , ..., en ); return; f(e1 , ..., en ); g(x, y1 , ..., yn ); return; (4) (5) x = f(e1 , ..., en ); g(x, y1 , ..., yn ); return; (6) Note the use of return to explicitly mark calls in tail position. The forms (3) to (6) are only necessary to handle the cases where f and g return void; in the rest of the proof, we ignore these cases that are a syntactical detail of the C language, and focus on the essential cases (1) and (2). To prove the correctness of CPS-conversion, we need to express this definition in our small imperative language. This is done by defining CPS-convertible terms, which are a subset of the terms introduced in Definition 2.1 (Section 2.1). A program in CPS-convertible form consists of a set of mutually-recursive functions with no free variables, the body of each of which is a CPS-convertible term. A CPS-convertible term has two parts: the head and the tail. The head is a (possibly empty) sequence of assignments, possibly embedded within conditional statements. The tail is a (possibly empty) sequence of function calls in a highly restricted form: their parameters are (side-effect free) expressions, except possibly for the last one, which can be another function call of the same form. Values and expressions are left unchanged. 32 Definition 3.2 (CPS-convertible terms). v ::= 1 | true | false | n ∈ N expr ::= v | x | . . . F ::=f (expr, . . . , expr) | f (expr, . . . , expr, F ) Q ::=ǫ | Q ; F T ::= expr | x := expr ; T | if e then T else T | Q (values) (expressions) (nested function calls) (tail) (head) The essential property of CPS-convertible terms, which makes their CPS conversion immediate to perform, is the guarantee that there is no cps call outside of the tails. It makes continuations easy to represent as a series of function calls (tails) and separates them clearly from imperative blocks (heads), which are not modified by the CPC translator. The tails are a generalisation of Definition 3.1, which will be useful for the proof of correctness of CPS-conversion. Note that x = f(e1 , ..., en ); return g(x, y1 , ..., yn ) is represented by g(f (e1 . . . en ), y1 . . . yn ): this translation is correct because, contrary to C, our language guarantees a left-to-right evaluation of function parameters. Also noteworthy are the facts that: – there is no letrec construct anymore since every function is defined at top-level, – assignments, conditions and function parameters of f are restricted to expressions, to ensure that function calls only appear in tail position, – there is no need to forbid shared variables in the parameters of g because they are ruled out of our language by design. 3.2 Early evaluation In this section, we prove that correctness of early evaluation, ie. evaluating the expressions expr before F when reducing f (expr, . . . , expr, F ) in a tail. This result is necessary to show the correctness of the CPS-conversion, because function parameters are evaluated before any function call when building continuations. The reduction rules may be simplified somewhat for CPS-convertible terms. We do not need to keep an explicit environment of functions since there are no inner functions any more; for the same reason, the (letrec) rule disappears. Instead, we use a constant environment F holding every function used in the reduced term M . To account for the absence of free variables, the closures in F need not carry an environment. As a result, in the (call) rule, ρ′ = ε and F ′ = F . Early evaluation is correct for lifted terms because a lifted term can never modify the variables that are not in its environment, since it cannot access them through closures. Lemma 3.3. Let M be a lambda-lifted term. Then, ρ ′ M s −−−→ v s F implies s|dom(s)\Im(ρ) = s′ |dom(s)\Im(ρ) . Proof. By induction on the structure of the reduction. The key points are the use of ρ′ = ε in the (call) case, and the absence of (letrec) rules. (val) and (var) Trivial (s = s′ ). 33 (assign) By the induction hypotheses, s|dom(s)\Im(ρ) = s′ |dom(s)\Im(ρ) and l ∈ Im(ρ), hence s|dom(s)\Im(ρ) = (s′ + {l 7→ v})|dom(s)\Im(ρ) . (seq) By the induction hypotheses, s|dom(s)\Im(ρ) = s′ |dom(s)\Im(ρ) and s′ |dom(s′ )\Im(ρ) = s′′ |dom(s′ )\Im(ρ) . Since, dom(s) ⊂ dom(s′ ), the second equality can be restricted to s′ |dom(s)\Im(ρ) = s′′ |dom(s)\Im(ρ) . Hence, s|dom(s)\Im(ρ) = s′′ |dom(s)\Im(ρ) . (if-true) and (if-false) are proved similarly to (seq). (letrec) doesn’t occur since M is lambda-lifted. (call) By the induction hypotheses, (sn+1 + {li 7→ vi })|dom(sn+1 +{li 7→vi })\Im(ρ′′ ·ρ′ ) = s′ |dom(sn+1 +{li 7→vi })\Im(ρ′′ ·ρ′ ) Since ρ′ = ε, Im(ρ′′ ) = {li } and dom(sn+1 ) ∩ {li } = ∅ (by freshness), (sn+1 + {li 7→ vi })|dom(sn+1 ) = s′ |dom(sn+1 ) so sn+1 = s′ |dom(sn+1 ) . Since dom(s) \ Im(ρ) ⊂ dom(s) ⊂ dom(sn+1 ), sn+1 |dom(s)\Im(ρ) = s′ |dom(s)\Im(ρ) . Finally, we can prove similarly to the (seq) case that s|dom(s)\Im(ρ) = sn+1 |dom(s)\Im(ρ) . Hence, s|dom(s)\Im(ρ) = s′ |dom(s)\Im(ρ) . As a consequence, a tail of function calls cannot modify the current store, only extend it with the parameters of the called functions. Corollary 3.4. For every tail Q, ρ ′ Q s −−−→ v s implies s ⊑ s′ . F 34 Proof. We prove the corollary by induction on the structure of a tail. First remember that store extension (written ⊑) is a partial order over stores (Property 2.18), defined in Section 2.3.2 as follows: s ⊑ s′ iff s′ |dom(s) = s. The case ǫ is trivial. The case Q ; F is immediate by induction ((seq) rule), since ⊑ is transitive. Similarly, it is pretty clear that f (expr, . . . , expr, F ) follows by induction and transitivity from f (expr, . . . , expr) ((call) rule). We focus on this last case. Lemma 3.3 implies: (sn+1 + {li 7→ vi })|dom(sn+1 +{li 7→vi })\Im(ρ′′ ·ρ′ ) = s′ |dom(sn+1 +{li 7→vi })\Im(ρ′′ ·ρ′ ) . Since ρ′ = ε, Im(ρ′′ ) = {li } and dom(sn+1 ) ∩ {li } = ∅ (by freshness), (sn+1 + {li 7→ vi })|dom(sn+1 ) = s′ |dom(sn+1 ) so sn+1 = s′ |dom(sn+1 ) . The evaluation of expr parameters do not change the store: sn+1 = s. The expected result follows: s = s′ |dom(s) , hence s ⊑ s′ . This leads to the correctness of early evaluation. ρ ′ ρ ′ Theorem 3.5 (Early evaluation). For every tail Q, Q s −−−→ v s implies Q[x\s(ρ x)] s −−−→ v s F F (provided x ∈ dom(ρ) and ρ x ∈ dom(s)). Proof. Immediate induction on the structure of tails and expressions: Corollary 3.4 implies that s ⊑ s′′ and ρ x ∈ dom(s) ensures that s(ρ x) = s′′ (ρ x) in the relevant cases (namely the (seq) rule for Q ; F and the (call) rule for f (expr, . . . , expr, F )). 3.3 Small-step reduction We define the semantics of CPS-convertible terms through a set of small-step reduction rules. We distinguish three kinds of reductions: →T to reduce the head of terms, →Q to reduce the tail, and →e to evaluate expressions. These rules describe a stack machine with a store σ to keep the value of variables. Since free and shared variables have been eliminated in earlier passes, there is a direct correspondence at any point in the program between variable names and locations, with no need to dynamically maintain an extra environment. We use contexts as a compact representation for stacks. The head rules →T reduce triples made of a term, a context and a store: hT, C[ ], σi. The tail rules →Q , which merely unfold tails with no need of a store, reduce couples of a tail and a context: hQ, C[ ], i. The expression rules do not need context to reduce, thus operating on couples made of an expression and a store: he, σi. Contexts Contexts are sequences of function calls. In those sequences, function parameters shall be already evaluated: constant expressions are allowed, but not variables. As a special case, the last parameter might be a “hole” instead, written ⊖, to be filled with the return value of the next, nested function. Definition 3.6 (Contexts). Contexts are defined inductively: C ::= [ ] | C[[ ] ; f (v, . . . , v)] | C[[ ] ; f (v, . . . , v, ⊖)] 35 Definition 3.7 (CPS-convertible reduction rules). hx := expr ; T, C[ ], σi →T hT, C[ ], σ[x 7→ v]i when hexpr, σi hif expr then T1 else T2 , C[ ], σi →T hT1 , C[ ], σi →⋆e when hexpr, σi hif expr then T1 else T2 , C[ ], σi →T hT2 , C[ ], σi →⋆e when hexpr, σi →⋆e (7) v (8) true (9) false hexpr, C[[ ] ; f (v1 , . . . , vn )], σi →T hǫ, C[[ ] ; f (v1 , . . . , vn )]i hexpr, C[[ ] ; f (v1 , . . . , vn , ⊖)], σi →T hǫ, C[[ ] ; f (v1 , . . . , vn , v)]i (10) (11) when hexpr, σi →⋆e v hexpr, [ ], σi →T v when hexpr, σi →⋆e v hQ, C[ ], σi →T hQ[xi \ σ xi ], C[ ]i for every xi in dom(σ) hQ ; f (v1 , . . . , vn ), C[ ]i →Q hQ, C[[ ] ; f (v1 , . . . , vn )]i hQ ; f (v1 , . . . , vn , F ), C[ ]i →Q hQ ; F, C[[ ] ; f (v1 , . . . , vn , ⊖)]i hǫ, C[[ ] ; f (v1 , . . . , vn )]i →Q hT, C[ ], σi (12) (13) (14) (15) when f (x1 , . . . , xn ) = T and σ = {xi 7→ vi } We do not detail the rules for →e , which simply looks for variables in σ and evaluates arithmetical and boolean operators. Early evaluation Note that Rule 12 evaluates every function parameter in a tail before the evaluation of the tail itself. This is precisely the early evaluation process described above, which is correct by Theorem 3.5. We introduce early evaluation directly in the reduction rules rather than using it as a lemma to simplify the proof of correctess of the CPS-conversion. 3.4 CPS terms Unlike classical CPS conversion techniques [5], our CPS terms are not continuations, but a procedure which builds and executes the continuation of a term. Construction is performed by push, which adds a function to the current continuation, and execution by invoke, which calls the first function of the continuation, optionally passing it the return value of the current function. Definition 3.8 (CPS terms). v ::= 1 | true | false | n ∈ N (values) expr ::= v | x | . . . (expressions) Q ::= invoke | push f (expr, . . . , expr) ; Q | push f (expr, . . . , expr, ⊡) ; Q (tail) T ::= invoke expr | x := expr ; T | if e then T else T | Q (head) Continuations and reduction rules A continuation is a sequence of function calls to be performed, with already evaluated parameters. We write · for appending a function to a continuation, and ⊡ for a “hole”, i.e. an unknown parameter. 36 Definition 3.9 (Continuations). C ::= ε | f (v, . . . , v) · C | f (v, . . . , v, ⊡) · C The reduction rules for CPS terms are isomorphic to the rules for CPS-convertible terms, except that they use continuations instead of contexts. Definition 3.10 (CPS reduction rules). hx := expr ; T, C, σi →T hT, C, σ[x 7→ v]i when hexpr, σi →⋆e v (16) hif expr then T1 else T2 , C, σi →T hT1 , C, σi if hexpr, σi →⋆e true (17) hif expr then T1 else T2 , C, σi →T hT2 , C, σi if hexpr, σi →⋆e false (18) hinvoke expr, f (v1 , . . . , vn ) · C, σi →T hinvoke, f (v1 , . . . , vn ) · Ci hinvoke expr, f (v1 , . . . , vn , ⊡) · C, σi →T hinvoke, f (v1 , . . . , vn , v) · Ci (19) (20) when hexpr, σi →⋆e v hinvoke expr, ε, σi →T v when hexpr, σi →⋆e v hQ, C, σi →T hQ[xi \ σ xi ], Ci for every xi in dom(σ) (21) hpush f (v1 , . . . , vn ) ; Q, Ci →Q hQ, f (v1 , . . . , vn ) · Ci (22) hpush f (v1 , . . . , vn , ⊡) ; Q, Ci →Q hQ, f (v1 , . . . , vn , ⊡) · Ci (23) hinvoke, f (v1 , . . . , vn ) · Ci →Q hT, C, σi when f (x1 , . . . , xn ) = T and σ = {xi 7→ vi } (24) Well-formed terms Not all CPS term will lead to a correct reduction. If we push a function expecting the result of another function and invoke it immediately, the reduction blocks: hpush f (v1 , . . . , vn , ⊡) ; invoke, C, σi → hinvoke, f (v1 , . . . , vn , ⊡) · C, σi 6→ Well-formed terms avoid this behaviour. Definition 3.11 (Well-formed term). A continuation queue is well-formed if it does not end with: push f (expr, . . . , expr, ⊡) ; invoke . A term is well-formed if every continuation queue in this term is well-formed. 3.5 Correctess of the CPS-conversion We define the CPS conversion as a mapping from CPS-convertible terms to CPS terms. 37 Definition 3.12 (CPS conversion). (Q ; f (expr, . . . , expr))N = push f (expr, . . . , expr) ; QN (Q ; f (expr, . . . , expr, F ))N = push f (expr, . . . , expr, ⊡) ; (Q ; F )N ǫN = invoke (x := expr ; T )N = x := expr ; T N (if expr then T1 else T2 )N = if expr then T1N else T2N expr N = invoke expr In the rest of this section, we prove that this mapping yields an isomorphism between the reduction rules of CPS-convertible terms and well-formed CPS terms, whence the correctness of our CPS conversion (Theorem 3.17). We first prove two lemmas to show that N yields only well-formed CPS terms. This leads to a third lemma to show that N is a bijection between CPS-convertible terms and well-formed CPS terms. CPS-convertible terms have been carefully designed to make CPS conversion as simple as possible. Accordingly, the following three proofs, while long and tedious, are fairly trivial. Lemma 3.13. Let Q be a continuation queue. Then QN is well-formed. Proof. By induction on the structure of a tail. ǫN = invoke and (ǫ ; f (expr, . . . , expr))N = push f (expr, . . . , expr) ; invoke are well-formed by definition. ((Q ; F ) ; f (expr, . . . , expr))N = push f (expr, . . . , expr) ; (Q ; F )N and (Q ; f (expr, . . . , expr, F ))N = push f (expr, . . . , expr, ⊡) ; (Q ; F )N are well-formed by induction. Lemma 3.14. Let T be a CPS-convertible term. Then T N is well-formed. Proof. Induction on the structure of T , using the above lemma. Lemma 3.15. The terms. N relation is a bijection between CPS-convertible terms and well-formed CPS Proof. Consider the following mapping from well-formed CPS terms to CPS-convertible terms: (push f (expr, . . . , expr) ; Q)H = QH ; f (expr, . . . , expr) (push f (expr, . . . , expr, ⊡) ; Q)H = Q′ ; f (expr, . . . , expr, F ) with QH = Q′ ; F H invoke = ǫ (x := expr ; T )H = x := expr ; T H if expr then T1 else T2H = if expr then T1H else T2H (invoke expr)H = expr (*) The existence of Q′ is guaranteed by well-formedness: 38 (*) – ∀T, T H = ǫ ⇒ T = invoke (by disjunction on the definition of H ), – here, Q 6= invoke because (push f (expr, . . . , expr, ⊡) ; Q) is well-formed, – hence QH 6= ǫ. One checks easily that (T H )N = T and (T N )H = T . To conclude the proof of isomorphism, we also need an (obviously bijective) mapping from contexts to continuations: Definition 3.16 (Conversion of contexts). ([ ])△ = ε (C[[ ] ; f (v1 , . . . , vn )])△ = f (v1 , . . . , vn ) · C with (C[ ])△ = C (C[[ ] ; f (v1 , . . . , vn , ⊖)])△ = f (v1 , . . . , vn , ⊡) · C with (C[ ])△ = C The correctness theorem follows: Theorem 3.17 (Correctness of CPS conversion). The N and △ mappings are two bijections, the inverses of which are written H and ▽ . They yield an isomorphism between reduction rules of CPS-convertible terms and CPS terms. Proof. Lemma 3.15 ensures that N is a bijection between CPS-convertible terms and well-formed CPS terms. Moreover, △ is an obvious bijection between contexts and continuations. To complete the proof, we only need to apply N , △ , H and ▽ to CPS-convertible terms, contexts, well-formed CPS terms and continuations (respectively) in every reduction rule and check that we get a valid rule in the dual reduction system. The result is summarized in Figure 4. 39 hx := expr ; T, C[ ], σi →T hT, C[ ], σ[x 7→ v]i ⇔ hx := expr ; T, C, σi →T hT, C, σ[x 7→ v]i when hexpr, σi →⋆e v hif expr then T1 else T2 , C[ ], σi →T hT1 , C[ ], σi ⇔ hif expr then T1 else T2 , C, σi →T hT1 , C, σi if hexpr, σi →⋆e true hif expr then T1 else T2 , C[ ], σi →T hT2 , C[ ], σi ⇔ hif expr then T1 else T2 , C, σi →T hT2 , C, σi if hexpr, σi →⋆e false hexpr, C[[ ] ; f (v1 , . . . , vn )], σi →T hǫ, C[[ ] ; f (v1 , . . . , vn )]i hexpr, C[[ ] ; f (v1 , . . . , vn , ⊖)], σi →T hǫ, C[[ ] ; f (v1 , . . . , vn , v)]i ⇔ hinvoke expr, f (v1 , . . . , vn ) · C, σi →T hinvoke, f (v1 , . . . , vn ) · Ci ⇔ hinvoke expr, f (v1 , . . . , vn , ⊡) · C, σi →T hinvoke, f (v1 , . . . , vn , v) · Ci when hexpr, σi →⋆e v 40 hexpr, [ ], σi →T v ⇔ hQ, C[ ], σi →T hQ[xi \ σ xi ], C[ ]i hQ ; f (v1 , . . . , vn ), C[ ]i →Q hQ, C[[ ] ; f (v1 , . . . , vn )]i ⇔ hQ, C, σi →T hQ[xi \ σ xi ], Ci for every xi in dom(σ) ⇔ hpush f (v1 , . . . , vn ) ; Q, Ci →Q hQ, f (v1 , . . . , vn ) · Ci hQ ; f (v1 , . . . , vn , F ), C[ ]i →Q hQ ; F, C[[ ] ; f (v1 , . . . , vn , ⊖)]i ⇔ hǫ, C[[ ] ; f (v1 , . . . , vn )]i →Q hT, C[ ], σi hinvoke expr, ε, σi →T v when hexpr, σi →⋆e v ⇔ hpush f (v1 , . . . , vn , ⊡) ; Q′ , Ci →Q hQ′ , f (v1 , . . . , vn , ⊡) · Ci when Q′ = (Q ; F )N hinvoke, f (v1 , . . . , vn ) · Ci →Q hT, C, σi when f (x1 , . . . , xn ) = T and σ = {xi 7→ vi } Figure 4: Isomorphism between reduction rules References [1] William D. Clinger. Proper tail recursion and space efficiency. In Proceedings of the ACM SIGPLAN 1998 conference on Programming language design and implementation, PLDI ’98, pages 174–185, New York, NY, USA, 1998. ACM. [2] Olivier Danvy and Ulrik Schultz. Lambda-lifting in quadratic time. In Functional and Logic Programming, volume 2441 of Lecture Notes in Computer Science, pages 134–151. SpringerVerlag, Berlin, Germany, 2002. [3] International Organization for Standardization. ISO/IEC 9899:1999 “Programming Languages – C”, December 1999. [4] Gabriel Kerneis and Juliusz Chroboczek. Continuation-Passing C: from threads to events through continuations. January 2012. Submitted for publication. [5] G. D. Plotkin. Call-by-name, call-by-value and the lambda-calculus. Theoretical Computer Science, 1(2):125–159, December 1975. 41
6cs.PL
Analyzing Adaptive Cache Replacement Strategies Mario E. Consuegra2 , Wendy A. Martinez1 , Giri Narasimhan1 , Raju Rangaswami1 , Leo Shao1 , and Giuseppe Vietri1 arXiv:1503.07624v2 [cs.DS] 24 Apr 2017 1 School of Computing and Information Sciences, Florida International University, Miami, FL 33199, USA. {walem001,giri,raju,gviet001}@fiu.edu 2 Google Inc., Kirkland, WA, USA. April 25, 2017 Abstract Adaptive Replacement Cache (Arc) and CLOCK with Adaptive Replacement (Car) are state-of-theart “adaptive” cache replacement algorithms invented to improve on the shortcomings of classical cache replacement policies such as Lru, Lfu and Clock. By separating out items that have been accessed only once and items that have been accessed more frequently, both Arc and Car are able to control the harmful effect of single-access items flooding the cache and pushing out more frequently accessed items. Both Arc and Car have been shown to outperform their classical and popular counterparts in practice. Both algorithms are complex, yet popular. Even though they can be treated as online algorithms with an “adaptive” twist, a theoretical proof of the competitiveness of Arc and Car remained unsolved for over a decade. We show that the competitiveness ratio of Car (and Arc) has a lower bound of N + 1 (where N is the size of the cache) and an upper bound of 18N (4N for Arc). If the size of cache offered to Arc or Car is larger than the one provided to Opt, then we show improved competitiveness ratios. The important implication of the above results are that no “pathological” worst-case request sequences exist that could deteriorate the performance of Arc and Car by more than a constant factor as compared to Lru. 1 Introduction Megiddo and Modha [MM03,MM04] engineered an amazing cache replacement algorithm that was self-tuning and called it Adaptive Replacement Cache or Arc. Later, Bansal and Modha [BM04] designed another algorithm called Clock with Adaptive Replacement (Car). Extensive experimentation suggested that Arc and Car showed substantial improvements over previously known cache replacement algorithms, including the well-known Least Recently Used or Lru and Clock. On the theoretical side, the seminal work of Sleator and Tarjan [ST85] showed that Lru can be analyzed using the theory of online algorithms. They showed that Lru has a competitiveness ratio of N (where N is the size of the cache). More surprisingly, they also showed that with no prefetching, no online algorithm for cache replacement could achieve a competitiveness ratio less than N , suggesting that under this measure, Lru is optimal. In other words, there exist worst-case request sequences that would prevent any algorithm from being better than N -competitive. While these results are significant, they highlight the difference between theory and practice. Sleator and Tarjan’s techniques analyze online algorithms in terms of their worst-case behavior (i.e., over all possible inputs), which means that other algorithms with poorer competitiveness ratios could perform better in practice. Another way to state this is that the results assume an oblivious adversary who designs inputs for the online algorithms in a way that make them perform as poorly as possible. The upper bound on performance ratio merely guarantees that no surprises are in store, i.e., there is no input designed by an adversary that can make the algorithm perform poorly. 1 Given a fixed size cache, the cache replacement problem is that of deciding which data item to evict from the cache in order to make room for a newly requested data item with the objective of maximizing cache hits in the future. The cache replacement problem has been referred to as a fundamental and practically important online problem in computer science (see Irani and Karlin [Hoc97], Chapter 13) and a “fundamental metaphor in modern computing” [MM04]. The Lru algorithm was considered the most optimal page replacement policy for a long time, but it had the drawback of not being “scan-resistant”, i.e., items used only once could pollute the cache and diminish its performance. Furthermore, Lru is difficult to implement efficiently, since moving an accessed item to the front of the queue is an expensive operation, first requiring locating the item, and then requiring data moves that could lead to unacceptable cache contention if it is to be implemented consistently and correctly. The Clock algorithm was invented by Frank Corbató in 1968 as an efficient one-bit approximation to Lru with minimum overhead [Cor68] and continues to be used in MVS, Unix, Linux, and Windows operating systems [Fri99]. Like Lru, Clock is also not scan-resistant because it puts too much emphasis on “recency” of access and pays no attention to “frequency” of access. So there are sequences in which many other algorithms can have significantly less cost than the theoretically optimal Lru. Since then, many other cache replacement strategies have been developed and have been showed to be better than Lru in practice. These are discussed below in Section 2. An important development in this area was the invention of adaptive algorithms. While regular “online” algorithms are usually designed to respond to input requests in an optimal manner, these selftuning algorithms are capable of adapting to changes in the request sequence caused by changes in the workloads. Megiddo and Modha’s Arc [MM03] is a self-tuning algorithm that is a hybrid of Lfu and Lru. Bansal and Modha’s Car is an adaptivehybrid of Lfu and Clock [BM04]. Experiments show that Arc and Car outperform Lru and Clock for many benchmark data sets [BM04]. Versions of Arc have been deployed in commercial systems such as the IBM DS6000/DS8000, Sun Microsystems’s ZFS, and in PostgreSQL. Unfortunately, no theoretical analysis of the adaptive algorithms, Arc and Car, exist in the literature. The main open question that remained unanswered was whether or not there exist “pathological” request sequences that could force Arc or Car to perform poorly. In this document we show that these two algorithms are O(N )-competitive, suggesting that they are not much worse than the optimal Lru. We also prove a surprising lower bound on the competitiveness that is larger than N . The main contributions of this paper are as follows: 1. For completeness, we provide proofs that Lru and Clock are N -competitive. 2. We prove a lower bound on the competitiveness of Arc and Car of N + 1, proving that there are request sequences where they cannot outperform Lru and Clock. 3. We show that Arc is 4N -competitive. 4. We show that Car is 18N -competitive. 5. We obtain precise bounds for the competitiveness of all four algorithms when the sizes of the caches maintained by them are different from that maintained by Opt. 6. We show that if the size of the cache is twice that of the one allowed for the optimal offline algorithm, then the competitiveness ratio drops to a small constant. We use the method of potential functions to analyze the algorithms. However, the main challenges in solving these problems is that of carefully designing the potential function for the analysis. We discuss the role of the adaptive parameter in the potential function. The contributions of this paper are summarized in Table 1. In this table, N is the size of the cache maintained by the algorithm, while NO is the size of the cache maintained by Opt. The table provides lower bounds (LB) and upper bounds (UB) on the competitiveness ratio when the cache sizes are equal, i.e., N = NO ; it also provides upper bounds when they are not equal. 2 Algorithm Lru Arc Clock Car Compet. Ratio LB N N +1 N N +1 Compet. Ratio UB N 4N N 18N Compet. Ratio UB w/ Unequal Sizes N/(N − NO + 1) 12N/(N − NO + 1) N/(N − NO + 1) 18N/(N − NO + 1) [Ref] [ST85] This paper This paper This paper Table 1: Summary of Results After providing relevant background on cache replacement algorithms in Section 2, we discuss the lower bounds on the competitiveness ratios of Arc and Car in Section 3. Next we prove upper bounds on the competitiveness ratios of Lru, Clock, Arc, and Car in Section 4. Concluding remarks can be found in Section 5. 2 Previous Work on Cache Replacement Strategies Below we give brief descriptions of the four algorithms being discussed in this paper, after which we mention a large collection of other closely related cache replacement algorithms. The Lru Algorithm: Lru evicts the least recently used entry. It tends to perform well when there are many items that are requested more than once in a relatively short period of time, and performs poorly on “scans”. Lru is expensive to implement because it requires a queue with move-to-front operations whenever a page is requested. The Clock Algorithm: On the other hand, Clock was designed as an efficient approximation of Lru, which it achieves by avoiding the move-to-front operation. Clock’s cache is organized as a single “circular” list, instead of a queue. The algorithm maintains a pointer to the “head” of the list. The item immediately counterclockwise to it is the “tail” of the list. Each item is associated with a “mark” bit. Some of the pages in the cache are marked, and the rest are unmarked. When a page hit occurs that page is marked, but the contents of the cache remain unchanged. When a page fault occurs, in order to make room for the requested page, the head page is evicted if the page is unmarked. If the head page is marked, the page is unmarked and the head is moved forward clockwise, making the previous head as the tail of the list. After a page is evicted, the requested page is unmarked and placed at the tail of the list. Clock is inexpensive to implement, but is not scan-resistant like Lru. The Arc Algorithm To facilitate our discussion, we briefly describe the Arc algorithm. As mentioned before, it combines ideas of recency and frequency. Arc’s cache is organized into a “main” part (of size N ) and a “history” part (of size N ). The main part is further divided into two lists, T1 and T2 , both maintained as LRU lists (i.e., sorted by “recency”). T1 focuses on “recency” because it contains pages with short-term utility. Consequently, when an item is accessed for the first time from the disk, it is brought into T1 . Items “graduate” to T2 when they are accessed more than once. Thus, T2 deals with “frequency” and stores items with potential for long-term utility. Additionally, Arc maintains a history of N more items, consisting of B1 , i.e., items that have been recently deleted from T1 , and B2 , i.e., items that have been recently deleted from T2 . History lists are also organized in the order of recency of access. The unique feature of Arc is its self-tuning capability, which makes it scan-resistant. Based on a self-tuning parameter, p, the size of T1 may grow or shrink relative to the size of T2 . The details of the algorithm are fairly complex and non-intuitive. Detailed pseudocode for Arc (Figure 4 from [MM03]) is provided in the Appendix for convenience. It is worth noting that Arc is considered a “universal” algorithm in the sense that it does not use any a priori knowledge of its input, nor does it do any offline tuning. Furthermore, Arc is continuously adapting, since adaptation can potentially happen at every step. 3 It must be noted that our results on Arc assume the “learning rate”, δ, to be equal to 1, while the Arc algorithm as presented by Megiddo and Modha recommended a “faster” learning rate based on experiments on real data. The learning rate is the rate at which the adaptive parameter p is changed as and when needed. The Car Algorithm Inspired by Arc, Car’s cache is organized into two main lists, T1 and T2 , and two history lists, B1 and B2 . Inspired by Clock, both T1 and T2 are organized as “circular” lists, with each item associated with a mark bit. The history lists, B1 and B2 are maintained as simple FIFO lists. We let t1 , t2 , b1 , b2 denote the sizes of T1 , T2 , B1 , B2 , respectively. Also, let t := t1 + t2 . Let lists L1 (and L2 , resp.) be the list of size ℓ1 (ℓ2 , resp.) obtained by concatenating list B1 to the end of “linearized” T1 (concatenating B2 to the tail of T2 , resp.). Note that circular lists are linearized from head to tail. We let T10 and T20 (T11 and T21 , resp.) denote the sequence of unmarked (marked, resp.) pages in T1 and T2 , respectively. The following invariants are maintained by Car for the lists: 1. 0 ≤ t1 + t2 ≤ N 2. 0 ≤ ℓ1 = t1 + b1 ≤ N 3. 0 ≤ ℓ1 + ℓ2 = t1 + t2 + b1 + b2 ≤ 2N 4. t1 + t2 < N =⇒ b1 + b2 = 0 5. t1 + t2 + b1 + b2 ≥ N =⇒ t1 + t2 = N 6. Once t1 + t2 = N and/or ℓ1 + ℓ2 = 2N , they remain true from that point onwards. Car maintains an adaptive parameter p, which it uses as a target for t1 , the size of list T1 . Consequently, N − p is the target for t2 . Using this guiding principle, it decides whether to evict an item from T1 or T2 in the event that a miss requires one of the pages to be replaced. The replacement policy can be summarized into two main points: 1. If the number of items in T1 (barring the marked items at the head of the list) exceeds the target size, p, then evict an unmarked page from T1 , else evict an unmarked page from T2 . 2. If ℓ1 = t1 + b1 = N , then evict a history page from B1 , else evict a history page from B2 . Since the details of the algorithm are complex, the actual pseudocode is provided (Figure 2 from [BM04]) in the Appendix. Other Cache Replacement Algorithms The DuelingClock algorithm [JIPP10] is like Clock but keeps the clock hand at the newest page rather than the oldest one, which allows it to be scan-resistant. More recent algorithms try to improve over Lru by implementing multiple cache levels and leveraging history. In [OOW93] the Lru-K algorithm was introduced. Briefly, the Lru-K algorithm estimates interarrival times from observed requests, and favors retaining pages with shorter interarrival times. Experiments have shown Lru-2 performs better than Lru, and that Lru-K does not show increase in performance over Lru2 [OOW93], but has a higher implementation overhead. It was also argued that Lru-K is optimal under the independence reference model (IRM) among all algorithms A that have limited knowledge of the K most recent references to a page and no knowledge of the future [OOW93]. In essence, the Lru-K algorithm tries to efficiently approximate Least Frequently Used (Lfu) cache replacement algorithm. As K becomes larger, it gets closer and closer to Lfu. It has been argued that Lfu cannot adapt well to changing workloads because it may replace currently “hot” blocks instead of “cold” blocks that had been “hot” in the past. Lfu is implemented as a heap and takes O(log N ) time per request. Another cache replacement algorithm is Lirs [JZ02]. The Lirs algorithm evicts the page with the largest IRR (inter-reference recency). It attempts to keep a small (≈ 1%) portion of the cache for HIR (high inter-reference) pages, and a large (≈ 99%) portion of the cache for LIR (low inter-reference) pages. The Clock-Pro algorithm approximates Lirs efficiently using Clock [JCZ05]. The 2q [JS94] algorithm 4 is scan-resistant. It keeps a FIFO buffer A1 of pages that have been accessed once and a main Lru buffer Am of pages accessed more than once. 2q admits only hot pages to the main buffer. The buffer A1 is divided into a main component that keeps the pages in A1 that still reside in cache, and a history component that remembers pages that have been evicted after one access. The relative sizes of the main and history components are tunable parameters. 2q has time complexity of O(1). Another algorithm that tries to bridge the gap between recency and frequency is Lrfu [LCK+ 01]. This is a hybrid of Lru and Lfu and is adaptive to changes in workload. The time complexity ranges from O(1) for Lru to O(log n) for Lfu. 3 Lower Bounds on Competitiveness Ratio for Arc and Car This section presents our results on the lower bounds for Arc and Car. We also show that the adaptive parameter is critical to both Arc and Car by showing that their non-adaptive versions have an unbounded competitiveness ratio. 3.1 Lower Bound for Arc First, we show a lower bound on the competitiveness ratio for Arc. Theorem 1. The competitiveness ratio of Algorithm Arc has a lower bound of N + 1. Proof. We show that we can generate an unbounded request sequence that causes N + 1 page faults on Arc for every page fault on Opt. The sequence only involves 2N + 1 pages denoted by 1, . . . , 2N + 1. Our example, will take the contents of the cache managed by Arc from configurations 1 through configuration 5, which are shown in Table 2. Note that configuration 1 and configuration 5 are essentially the same to the extent that the value of p is 0 in both, and the number of pages in each of the four parts of the cache are identical. Configuration 1 2 3 4 5 p 0 0 0 1 0 T1 ∅ 2N + 1 ∅ ∅ ∅ T2 1, . . . , N 2, . . . , N 2, . . . , N, 1 3, . . . , N, 1, 2N + 1 1, 2N + 1, 2, . . . , N − 1 B1 ∅ ∅ 2N + 1 ∅ ∅ B2 N + 1, . . . , 2N N + 2, . . . , 2N, 1 N + 2, . . . , 2N N + 2, . . . , 2N, 2 N + 2, . . . , 2N, N Table 2: Example for Lower Bound on Arc’s competitiveness We note that we can obtain configuration 1 from an empty cache with the following request sequence: 2N, 2N, 2N − 1, 2N − 1, . . . , 2, 2, 1, 1. Consider the first half of the above request sequence, which contains a total of 4N requests to 2N new pages, each page requested twice in succession. The first time a page is requested from the first N new pages, it will be put in T1 . The second time the page is requested, it will get moved to T2 . In the second half, if a page not in Arc is requested, Replace will be called, which will move a page from T2 to B2 , and the new page will be placed in T1 . When the same page is requested again, it simply gets moved to T2 . The value of p remains unchanged in this process. It is clear that we get Configuration 1 as a result of the request sequence. We design our request sequence by following the steps below. 1. Make one request to a page 2N + 1 not in Arc. We will assume that this is a brand new page and therefore also causes a fault for Opt and for Arc. The page 2N + 1 will go into T1 and a page in T2 will be demoted to B2 . The contents of Arc is given by Configuration 2 in Table 2. 2. Request any page in B2 . This decreases the value of p but since p is zero it will remain unchanged. Since the size of T1 is more than p Arc will call Replace, which will act on T1 , hence 2N + 1 will be demoted to B1 . Upon requesting page 1 in B2 , we get Configuration 3 in Table 2. 5 3. The next step is to request 2N + 1 again, which will move to T2 , p is increased and a page in T2 is demoted to B2 . Configuration 4 reflects the contents of the cache at this stage. 4. Finally we make N − 2 requests to any pages from B2 . By requesting the pages 2, 3, . . . , N , we end up in Configuration 5 from Table 2. The steps outlined above cause N + 1 page faults for Arc and at most one page fault for Opt. Since we are back to the initial configuration we can repeat this process over again. This concludes the proof that the competitiveness ratio of Arc is at least N + 1. 3.2 Lower Bound for Car Now we prove a similar lower bound for Car. Theorem 2. The competitiveness ratio of Algorithm Car has a lower bound of N + 1. Proof. We show that we can generate an infinite request sequence that causes N + 1 page faults in Car for every page fault on Opt. The sequence only involves 2N + 1 pages denoted by 1, . . . , 2N + 1. Our example, will take the contents of the cache managed by Car from configurations 1 through N + 2 as shown in Table 3. Note that a superscript of 1 on any page in T1 ∪ T2 indicates that it is marked. All others are unmarked. Also note that configuration 1 and configuration N + 2 are essentially the same upon relabeling. First, we show that configuration 1 is attainable, by showing that it can be reached from an empty cache. This is formalized in the following lemma. Lemma 1. We can obtain configuration 1 starting from an empty cache with the following request sequence: 2N, 2N, 2N − 1, 2N − 1, . . . , 2, 2, 1, 1. Proof. The first half of the above request sequence calls each of the N pages 2N, 2N − 1, . . . , N + 1 twice in succession. The first time they are called, they are moved into T1 unmarked. The second time the same page is called it gets marked, but remains in T1 . At the end of the first half, all the N pages requested end up in T1 and are all marked. The next call to new page N , will trigger a call to Replace, which will move all the marked pages in T1 to T2 leaving them unmarked. It will also move one page from T2 to B2 . Finally, the requested page N will be moved to T1 and left unmarked. When requested again, it simply gets marked. When the next page, i.e., N − 1 is requested, it moves marked page N to T2 , moves one more page from T2 to B2 . As the rest of the pages from the request sequences are requested, the previous requested page gets moved to T2 , which in turn demotes one of its pages to B2 . At the end of the process, we get a marked page 1 in T1 . Pages 2, . . . , N are in T2 , unmarked, and pages N + 1, . . . , 2N end up in B2 . This is exactly what we need for configuration 1. Continuing on the proof of Theorem 2, we show what happens when, starting from configuration 1, Car processes the following request sequence. Page 2N + 1: A page in T2 is demoted to B2 , which loses a page; the marked page from T1 is moved to T2 and the new page is moved into T1 . MRU page in B2 : This should have decremented p but remains unchanged since it is already zero. Since the size of T1 is more than p Car will call Replace and 2N + 1 will be demoted to B1 , resulting in configuration 3 in Table 3. Page 2N + 1: It will move to T2 , p is increased and a page in T2 is demoted to B2 . See configuration 4 in Table 3. MRU page from B2 , repeat N − 2 times: It results in configuration N + 2 in Table 3. 6 Config. 1 2 3 4 5 ... N +2 p 0 0 0 1 0 0 0 B1 ∅ ∅ 2N + 1 ∅ ∅ ... ∅ T1 11 2N + 1 ∅ ∅ ∅ ... ∅ T2 2, . . . , N 1, . . . , N − 1 N, 1, . . . , N − 1 2N + 1, N, 1, . . . , N − 2 N − 1, 2N + 1, N, 1, . . . , N − 3 ... 2, . . . , N − 1, 2N + 1, N B2 N + 1, . . . , 2N N, . . . , 2N − 1 N + 1, . . . , 2N − 1 N − 1, N + 1, . . . , 2N − 1 N − 2, N + 1, . . . , 2N − 1 ... 1, N + 1, . . . , 2N − 1 Table 3: Example for Lower Bound on Car’s competitiveness The request sequence detailed above generates N + 1 faults for Car while only N different pages are requested. Thus, Opt could limit itself to at most one fault in this stretch. Opt will fault once during each stretch if the next page is picked to be one that is farthest used in the future. Repeating the above steps an unbounded number of times with appropriate relabeling proves that the competitiveness ratio of Car is lower bounded by N + 1. 3.3 Non-Adaptive Arc and Car are not Competitive It is particularly interesting to note that the non-adaptive version of Car and Arc (called Fixed Replacement cache) [MM03] are not competitive. The following two theorems prove that the competitiveness ratios can be unbounded in this case. Theorem 3. Algorithm Car with fixed p is not competitive. Proof. Suppose that algorithm Car has p fixed instead of being adaptive and 0 < p < N − 1. Recall that p is the target size of T1 and N − p is the target size of T2 . We design a request sequence such that with less than N pages we can generate an infinite number of page faults for Car. The sequence is described as follows: Step 1: Fill up T2 with N − p unmarked pages as described above in the proof of Theorem 2. Step 2: Request the MRU page in B2 . The requested page goes to the tail of T2 as an unmarked page. Since the size of T2 is greater than p we discard the head of T2 . Step 3: Request the MRU page in B2 which is actually the page discarded in Step 2 from T2 . This step is similar to Step 2 and we can continue to repeat this infinitely often, since the page that moves from B2 to T2 get’s unmarked and the page that moves from T2 to B2 goes to MRU. Therefore, we can cycle infinitely many times through N − p + 1 pages triggering an infinite number of faults, while Opt can avoid faults altogether during the cycle. Theorem 4. Algorithm Arc with fixed p is not competitive. Proof. Suppose that algorithm Arc has p fixed instead of being adaptive and 0 < p < N . Recall that p is the target size of T1 and N − p is the target size of T2 . We design a request sequence such that with less than N pages we can generate an infinite number of page faults for Arc. The first step is to fill up T2 (size of T2 = N − p). Next we request the MRU page in B2 . Every time we request a page from B2 , it goes into the top of T2 and thus it increases the size of T2 beyond its target size. It follows that Arc will call Replace and move a page from T2 to the MRU position in B2 . If the MRU page from B2 is repeatedly requested, we will cycle through N − p pages, every time incurring a page fault for Arc, while Opt can avoid faults altogether during the cycle. 7 4 4.1 Analyzing Lru using potential functions The generic approach The standard approach used here is as follows. First, we define a carefully crafted potential function, Φ. As per the strategy of analyzing competitiveness ratios suggested by Sleator and Tarjan [ST85], we then try to show the following inequality: CA + ∆Φ ≤ f (N ) · CO + g(N ), (1) where CA and CO are the costs incurred by the algorithm and by Opt, respectively, ∆Φ is the change in potential, f (N ) is some function of N , the size of the cache. In all of our proofs, we assume that the work involves simultaneously maintaining Opt’s cache as well as the algorithm’s cache. So we can break down the work into two steps, one where only Opt serves and one where only the algorithm serves. When only Opt serves, there are 2 cases: first when Opt has a hit and next when Opt has a miss. Next, we consider the cases when the algorithm serves, once when it has a hit and once when it has a miss. In each case, our goal is to prove the inequality (1) mentioned above, which establishes that f (N ) is the competitiveness ratio of algorithm A. There may be an additive term of g(N ) which is a function of the misses needed to get to some initial configuration for the cache. 4.2 Analyzing Lru using potential functions Assuming that the size of cache given to the competing Opt algorithm is NO ≤ N , the following result was proved by Sleator and Tarjan [ST85] (Theorem 6) for Lru.  N -competitive. Theorem 5. [ST85] Algorithm Lru is N −N O +1 Here we present a complete proof of this well-known result because we believe it is instructive for the other proofs in this paper. Proof. While this was not used in the proof in Sleator and Tarjan [ST85], a potential function that will facilitate the proof of the above theorem is: Φ= P r(x) , NL − NO + 1 x∈D (2) where D is the list of items in Lru’s cache but not in Opt’s cache, and r(x) is the rank of item x in Lru’s list with the understanding that the LRU item has rank 1, while the MRU item has rank equal to the size of the cache [Alb96]. We now show the following inequality: CA + ∆Φ ≤  N · CO + O(N ), N − NO + 1 (3) where CA and CO are the costs incurred by the algorithm and by Opt, respectively, ∆Φ is the change in potential, f (N ) is some function of N , the size of the cache. We assume that the work involves simultaneously maintaining Opt’s cache as well as Lru’s cache. So we can break down the work of Lru into two steps, one where only Opt serves and one where only Lru serves. When only Opt serves, there are 2 cases: first when Opt has a hit and next when Opt has a miss. In either case, the cost for Lru is 0, since only Opt is serving. When Opt has a hit, the cost for Opt is also 0. Furthermore, since Lru’s cache remains untouched, and no changes take place in the contents of Opt’s cache, the ranks of the items in Lru remain unchanged. Thus, ∆Φ = 0. Therefore, the inequality in (3) is trivially satisfied in this case. 8 When Opt has a miss, CO = 1, as before. The item evicted by Opt can contribute the rank of that item NL to increase at most by NL , making the increase in potential function to be bounded by NL −N . Thus, O +1 the inequality in (3) is satisfied. Next, we consider the step where Lru serves the request. As with Opt, when Lru is serving, the cost for Opt is 0. We again consider two cases: first when Lru has a hit and next when Lru has a miss. When Lru has a hit, the cost for Lru is 0. The contents of Lru’s cache may change. The item that was accessed is moved to the MRU position. However, this item is already in Opt’s cache and therefore cannot contribute to a change in potential. Several other items may move down in the cache, thus contributing to a decrease in potential of at most (N − 1). In the worst, case the increase in potential is at most 0. Therefore, the inequality in (3) is again satisfied. Finally, we consider the case when Lru has a miss. As before, CL = 1. Following the previous arguments, an item would be brought into MRU (which is already present in Opt’s cache), a bunch of items may be demoted in rank, and the Lru item will be evicted. The only action that can contribute to an increase is caused by the item that is brought into the MRU location. However, this item is already present in Opt’s cache, and hence cannot contribute to an increase. All the demotions and eviction can only decrease the potential function. Note that before the missed item is brought into Lru’s cache, the contents of Lru’s and Opt’s cache agree in at most NO − 1 items, since Opt just finished serving the request and the item that caused the miss is already in Opt’s cache. Thus there are at least NL − NO + 1 items that contribute their ranks to the potential function. These items either get demoted in rank or get evicted. Either way, the potential function will reduce by a minimum value of NL − NO + 1, although it could more if there are more items that are in Lru and that are not in Opt’s cache. Thus the total change in potential has to be at most NL − NO + 1, and we have CL + ∆Φ ≤ 1 − (NL − NO + 1) NL ≤0= · CO . (NL − NO + 1) NL − NO + 1 Summarizing the costs, we have the following: Step CL ∆Φ Opt Serves Request Opt has a hit 0 0 Opt has a miss 0 ≤ NL Lru Serves Request Lru has a hit 0 ≤0 Lru has a miss 1 ≤ NL − NO + 1 CO 0 1 0 0 The analysis of Lru states that if the sizes of Lru’s and Opt’s caches are NL and NO respectively, and NL if NL ≥ NO , then the competitiveness ratio of Lru is NL −N . Thus Lru is 2-competitive if the size of O +1 Lru’s cache is roughly twice that of Opt’s cache. 4.3 Analyzing the competitiveness of Clock Our result on the competitiveness of Clock is formalized in the following theorem. While this result appears to be known, we have not been able to locate a full proof and we believe this is of value. We therefore present it for the sake of completeness.  N Theorem 6. Algorithm Clock is N −N -competitive. O +1 Proof. Let M0 denote the subsequence of unmarked pages in Clock, ordered counterclockwise from head to tail. Let M1 denote the subsequence of marked pages in Clock, ordered counterclockwise from head to tail. Let q be any page in Clock’s cache. Let P 0 [q] denote the position of an unmarked page q in the 9 ordered sequence M0 , and let P 1 [q] denote the position of a marked page q in M1 . Finally, let R[q] denote the rank of page q defined as follows: R[q] = ( P 0 [q] if q is unmarked, P 1 [q] + |M0 | otherwise. (4) Thus, if q is an unmarked page at the head, then R[q] = 1. By the above definition, the following lemmas are obvious. Lemma 2. If q is any page in Clock’s cache, then R[q] ≤ N . Lemma 3. If a marked page q at the head of Clock’s cache is unmarked and moved to the tail, then R[q] does not change in the process. Let D be the set of pages that are in the cache maintained by Clock, but not in the cache maintained by Opt. We define the potential function as follows: Φ= X R[q] (5) q∈D We prove one more useful lemma about the ranks as defined above. Lemma 4. If an unmarked page at the head of Clock’s cache is evicted from Clock’s cache, and if there is at least one page in D, then Φ decreases by at least 1 in the process. . Proof. All pages, marked or unmarked, will move down by at least one position (reducing the rank of each by at least 1). The decrease in potential for at least one page that is in D will contribute to Φ, guaranteeing that ∆Φ ≤ −1. Let CClock and COpt be the costs incurred by the algorithms Clock and Opt, and let S = σ1 , σ2 , . . . , σm be an arbitrary request sequence. Let S ′ denote the initial subsequence of requests that take place prior to the cache becoming full. Note that exactly N faults are incurred in S ′ , after which the cache remains full. Let S ′′ be the subsequence of S that comes after S ′ . Let CClock and COpt be the cost incurred by the algorithms Clock and Opt respectively. We will prove that for every individual request, σ ∈ S ′′ : CClock (σ) + ∆Φ ≤ N ∗ COpt (σ) (6) As before, we assume that request σ is processed in two distinct steps: first when Opt services the page request and, next when Clock services the request. We will show that inequality (6) is satisfied for both the steps. When only Opt acts in this step, Cclock = 0. If Opt does not fault on this request, then COP T = 0. No change occurs to the contents of the cache maintained by Opt as well as Clock, and the clock hand does not move. Thus, ∆Φ = 0, satisfying inequality 6. If Opt faults on request σ, then COP T = 1 and CClock = 0. The contents of the cache maintained by Opt does change, which could affect the potential function. The potential could increase due to the eviction of a page in Opt. Since by Lemma 2 the rank of the evicted page cannot exceed N , the potential will change by at most N . Thus, inequality 6 is satisfied. Next we consider what happens when Clock services the request. For this case COP T = 0. If Clock does not fault, then Cclock = 0 and the requested page may change from an unmarked status to a marked one. However, since the page is already in the cache maintained by Opt it is not in D and is therefore not considered for the potential function calculations in 5. Thus, inequality 6 is satisfied. 10 Finally, we consider the case when Clock faults, in which case CClock = 1 and COpt = 0. To satisfy inequality 6, ∆Φ needs to be less or equal to -1. When Clock has a miss, if the head page happens to be marked, then Clock will repeatedly unmark the marked head page, moving it to the tail position, until an unmarked head page is encountered. The unmarked head page is then evicted. Each time a marked head page becomes an unmarked tail page, by Lemma 3 its rank does not change. When finally an unmarked head page is evicted, we know that there is at least one page in Opt’s cache that is not in Clock’s cache (i.e., the page that caused the fault). Since there are N pages in the cache maintained by Clock, at least one of those pages is guaranteed not to be part of the cache maintained by Opt. Since there is at least one page in D, by Lemma 4 it is clear that evicting an unmarked head page will decrease the potential function by at least one, which will pay for the Clock’s page fault. We have therefore showed that for every request σ, inequality 6 is satisfied. Since there can be at most N faults for the requests in S ′ , summing up the above inequality for all requests, σ ∈ S, we get CClock (S) ≤ N ∗ COpt (S) + N. This completes the proof of the theorem and the competitiveness analysis of the Clock algorithm. 4.4 Analyzing the Competitiveness of ARC In this paper, we prove two different upper bounds for the competitiveness of Arc. These two proofs use very different potential function. The first one allows for the sizes of the caches maintained by Arc and Opt to be different, while the second one does not allow for it, but provides a tighter bound. We provide both results below. Our first result on the competitiveness of Arc is formalized in the following theorem:  12N Theorem 7. Algorithm Arc is N −N -competitive. O +1 Proof. Let PX [q] be the position of page q in an arbitrary ordered sequence of pages X. When the set is obvious, we will drop the subscript and denote PX [q] simply by P [q]. The set of history pages T1 , T2 , B1 , and B2 will be treated as an ordered sequence of pages ordered from its LRU position to its MRU position. Let Opt and Car be the set of main pages stored in the caches for algorithms Opt and Arc respectively. Let D = Arc \ Opt. As before, we associate each page with a rank value R[q], which is defined as follows:  2PB1 [q]    2P [q] B2 R[q] =  4P T1 [q] + 2b1    4PT2 [q] + 2b2 if if if if q q q q ∈ B1 ∈ B2 ∈ T1 ∈ T2 (7) Finally, we define the potential function as follows: Φ = p + 2t1 + 2 P R[q]  − 3|Arc| N − NO + 1 q∈D (8) The initial value of Φ is 0. If the following inequality (9) is true for any request σ, where ∆Φ is the change in potential caused by serving the request, then when summed over all requests, it proves Theorem 7. CArc (σ) + ∆Φ ≤ 12N COpt (σ) . N − NO + 1 (9) As before, we assume that request σ is processed in two distinct steps: first when Opt serves and, next when Arc serves. We will show that inequality (9) is satisfied for each of the two steps. 11 Step 1: Opt serves request σ Since only Opt acts in this step, CArc = 0, and T1 ∪ T2 does not change. There are two possible cases: either Opt faults on σ or it does not. If Opt does not fault on this request, then it is easy to see that COpt = 0 and ∆Φ = 0, thus satisfying inequality (9). If Opt faults on request σ, then COpt = 1 and some page q, is evicted from the cache maintained by Opt will belong to D after this step and thus its rank will contribute to the potential function, which will increase by two times the rank of q. The maximal positive change in potential will occur when q is the MRU page of either T1 or T2 . In this case the rank of q is given by: R[q] = 4P [q] + b1 (R[q] = 4P [q] + b2 ). The maximum possible values for each of the terms P [q] and b1 will be N , hence the maximum possible rank of 12N q will be 4N + 2N = 6N . Therefore resulting potential change is at most N −N . O +1 Step 2: Arc serves request σ We break down the analysis into four cases. Case 2.1 deals with the case when Arc finds the page in its cache. The other three cases assume that Arc faults on this request because the item is not in T1 ∪T2 . Cases 2.2 and 2.3 assume that the missing page is found recorded in the history in lists B1 and B2 , respectively. Case 2.4 assumes that the missing page is not recorded in history. Case 2.1: Arc has a page hit Clearly, the page was found in T1 ∪ T2 , and CArc = 0. We consider the change of each of terms in the potential function individually. 1. As per the algorithm, p can only change when the page is found in history. (See lines 3 through 10 of Arc(x).) Since the page is not found in Arc’s history, ∆p = 0. 2. If the hit happens in T1 , the page will move to the top of T2 (See line 2 of Arc(x).), which will result in a decrease in t1 . If the hit happens in T2 , the size of t1 will remain the same. The overall change in t1 will be 0. 3. Since Opt has already served the page, the page is in Opt’s cache. Therefore, even if the page’s rank could change when moved from T1 to MRU position of T2 , this rank will not affect the potential since the page is not in D. We, therefore, conclude that ∆Φ = 0, satisfying inequality (9). Next we will analyze the 3 cases when the requested page is not in Arc’s cache. Since CArc = 1, the change in potential must be ≤ −1 in each case in order for inequality (9) to be satisfied. Case 2.2: Arc has a page miss and the missing page is in B1 We consider the two cases, first when Replace moves an item from T1 to B1 and second when it moves an item from T2 to B2 . 1. Case 1: We consider the change in potential function by analyzing each of the 3 terms. • Value of p will either increase by 1 or stay the same in case p = N , we will account for the worst case which is when ∆p = 1. • A new page is being added to MRU of T2 , and Replace is taking the LRU page of T1 to B1 , then 2∆t1 = −2. • The page that moved from B1 to T2 is not in D, therefore the change in its rank will P not affect the potential, the other pages will could only decrease their rank, meaning that 2∆ q∈D R[q] ≤ 0. Since p increases by at most 1 and t1 decreases by at least 2 the total change in potential is at most -1. 2. Case 2: Once again. we consider the change in potential function by analyzing each of the three terms. • Value of p will either increase by 1 or stay the same in case p = N , we will account for the worst case which is when ∆p = 1. 12 • A new page is added to MRU of T2 , and Replace moves the LRU page of T2 to B2 . Thus, there is no change in T1 . • The page that moved from B1 to T2 is not in D, therefore the change in its rank will not affect the potential. Since t1 + t2 = N , it is guaranteed that at least N − NO + 1 pages are not in Opt. For the pages that are in T1 , their ranks will decrease by at least 2 since b1 decreases by 1, and for the pages in T2 their ranks will decrease by at least 2 as well since b2 increasesPby 1 but the LRU R[q] page in T2 will move to B2 , reducing P [q] for all the pages in T2 . The term 2 N q∈D −NO +1 decreases by at least -4. P R[q] Since p increases by at most 1 and 2 N q∈D −NO +1 decreases by at least -4 the total change in potential is at most -3. Case 2.3: Arc has a page miss and the missing page is in B2 When the missing page is in B2 , Arc makes a call to Replace (Line 5) and then executes Lines 18-19. Thus, p is decremented except if it is already equal to 0. We consider two sub cases: ∆p ≤ −1 and ∆p = 0. ∆p ≤ −1: As in Case 2.2, the call to Replace has no effect on t1 . Replace will not increment the rank using a similar analysis as in 2.2 and change in p will at least be -1. The change in the potential function is at most -1. ∆p = 0: Unlike the sub case above when p decreases by 1, the change in p cannot guarantee the required reduction in the potential. We therefore need a tighter argument. We know that there is a call to Replace. Two cases arise and are discussed below. • Replace moves an item in T1 to B1 : Since the LRU page of T1 is moved to the MRU position of B1 , 2∆t1 = −2 and there is no movement of a page in D that could increase the rank. Therefore the total change in the potential function is at most -2. • Replace moves an item in T2 to B2 : p = 0 indicates that T2 has N pages, therefore is guarantee that at least N − NO + 1 pages will not be part of Opt, contributing to the change in potential. The page being moved from T2 to B2 will decrease it’s rank by at least 2, and the rest of the pages in T2 will move down one position (P [q] will decrease by 1) while B2 will remain the same, resulting in a change in the potential function of at most -4. Thus, in each case the potential function decreased by at most -2. Case 2.4: Arc has a page miss and the missing page is not in B1 ∪ B2 1. t1 + b1 = N ; t1 < N ; The LRU page in B1 is evicted. Assume Replace moves a page from T1 to B1 and a new page is brought into T1 (∆t1 = 0, ∆b1 = 0, ∆t2 = 0, ∆b2 = 0). • The term p is not affected. • The term t1 is not affected. • Since t1 + b1 = N , at least N − No + 1 pages in T1 ∪ B1 are not in Opt. If the page is in B1 \ Opt then its rank decreases by 2; if the page is in T1 \ Opt its rank decreases by 4. 2. t1 + b1 = N ; t1 < N ; The LRU page in B1 is evicted. Assume Replace moves a page from T2 to B2 and a new page is brought into T1 (∆t1 = 1, ∆b1 = −1, ∆t2 = 1, ∆b2 = 1). • The term p is not affected. • The term t1 is increased by 1. 13 • Since t1 + t2 = N , at least N − No + 1 pages in T1 ∪ T2 that are not in Opt. If a page, q, is in T1 \ Opt then its rank decreases by 2 (∆R[q] = ∆4 ∗ P [q] + ∆2 ∗ b2 = −2); if the page, q, is in T2 \ Opt its rank decreases by 2 (∆R[q] = ∆4 ∗ P [q] + ∆2 ∗ b2 = −2). 3. t1 + b1 < N ; t1 + t2 + b1 + b2 = 2N ; Assume that the LRU page in B2 is evicted and Replace moves a page from T1 to B1 and a new page is brought into T1 (∆t1 = 0, ∆b1 = 1, ∆t2 = 0, ∆b2 = −1). • The term p is not affected. • The term t1 is not affected. • Here we used the fact that t2 + b2 > N , then at least N − No + 1 pages in T2 ∪B2 are not in Opt. If a page, q, is in T2 \Opt then its rank decreases by 2 (∆R[q] = ∆4∗P [q]+∆2∗b2 = 4∗(0)+2(−1) = −2); if the page, q, is in B2 \ Opt its rank decreases by 2 (∆R[q] = ∆2 ∗ P [q] = 2 ∗ (−1) = −2). 4. t1 + b1 < N ; t1 + t2 + b1 + b2 = 2N ; Assume that the LRU page in B2 is evicted and Replace moves a page from T2 to B2 and a new page is brought into T1 (∆t1 = 1, ∆b1 = 0, ∆t2 = 1, ∆b2 = 0). • The term p is not affected. • The term t1 is increased by 1. • Here we used the fact that t2 + b2 > N , then at least N − No + 1 pages in T2 ∪B2 are not in Opt. If a page, q, is in T2 \Opt then its rank decreases by 2 (∆R[q] = ∆4∗P [q]+∆2∗b2 = 4∗(0)+2(−1) = −2); if the page, q, is in B2 \ Opt its rank decreases by 2 (∆R[q] = ∆2 ∗ P [q] = 2 ∗ (−1) = −2). 5. t1 + b1 < N ; t1 + t2 + b1 + b2 < 2N ; In this case, no pages are evicted from history. Assume that Replace moves a page from T1 to B1 and a new page is brought into T1 (∆t1 = 0, ∆b1 = 1, ∆t2 = 0, ∆b2 = 0) • The term p is not affected. • The term t1 is increased by 1. • Here we cannot say that the rank decreases. Hence the rank term is at most 0. • The term |Arc| increases by 1. 6. t1 + b1 < N ; t1 + t2 + b1 + b2 < 2N ; In this case, no pages are evicted from history. Assume Replace moves a page from T2 to B2 and a new page is brought into T1 (∆t1 = 1, ∆b1 = 0, ∆t2 = −1, ∆b2 = 1) • The term p is not affected. • The term t1 is not affected. • Here we cannot say that the rank decreases. Hence the rank term is at most 0. • The term |Arc| increases by 1. Wrapping up the proof of Theorem 7: Combining the four cases (2.1 through 2.4) proves that inequality (9) is satisfied when Arc serves request σ. This completes the proof of Theorem 7, establishing that the upper bound on the competitiveness of Arc is 12N for the cases where the sizes of Opt and Arc are the same. By analyzing cases where the size of Arc is greater than Opt we can observe that since Arc 12N the greater the size of Arc’s cache relative to the size of Opt’s cache, smaller will be the will be N −N O +1 competitiveness of Arc. 14 4.5 Alternative Analysis of Competitiveness of Arc Below, we prove an improved upper bound on the competitiveness ratio of Arc. As seen below, the potential function is considerably different. Let CA and CO be the costs incurred by the algorithms Arc and Opt. We start with some notation and definitions. If X is the set of pages in a cache, then let M RU (X) and LRU (X) be the most recently and least recently used pages from X. Let M RUk (X) and LRUk (X) be the k most recently and k least recently used pages from X. Let lists L1 (and L2 ) be the lists obtained by concatenating lists T1 and B1 (T2 and B2 , resp.). Let list L be obtained by concatenating lists L1 and L2 . We let ℓ1 , ℓ2 , t1 , t2 , b1 , b2 denote the sizes of L1 , L2 , T1 , T2 , B1 , B2 , respectively. Finally, let t := t1 + t2 and ℓ := ℓ1 + ℓ2 . At any instant of time during the parallel simulation of Opt and Arc, and for any list X, we let M RUk (X) be denoted by T OP (X), where k is the largest integer such that all pages of M RUk (X) are also in the cache maintained by OPT. We let L′1 , L′2 , T1′ , T2′ denote the T OP s of L1 , L2 , T1 , T2 , respectively, with sizes ℓ′1 , ℓ′2 , t′1 , t′2 , respectively. We let b′1 and b′2 denote the sizes of the B1′ = L′1 ∩ B1 and B2′ = L′2 ∩ B2 , respectively. Note that if b′1 > 0 (b′2 > 0, resp.), then all of T1 (T2 , resp.) is in Opt. Finally, we let ℓ′ := ℓ′1 + ℓ′2 . The Arc algorithm ensures that 0 ≤ t ≤ N , 0 ≤ ℓ ≤ 2N and 0 ≤ ℓ1 ≤ N , thus making 0 ≤ ℓ2 ≤ 2N . We assume that algorithm X being analyzed is provided an arbitrary request sequence σ = σ1 , σ2 , . . . , σm . We define the potential function as follows: Φ = p − (b′1 + 2 · t′1 + 3 · b′2 + 4 · t′2 ). (10) The main result of this section is the following theorem: Theorem 8. Algorithm ARC is 4N -competitive. We say that the cache is full if t = N and either t1 + b1 = N or t2 + b2 ≥ N . We will prove the above theorem by proving the following inequality for any request σ that is requested after the cache is full: CA (σ) + ∆Φ ≤ 4N · CO (σ) + 2N, (11) where ∆X represents the change in any quantity X. Summing up the above inequality for all requests would prove the theorem as long as the number of faults prior to the cache becoming full is bounded by the additive term 2N . We make the following useful observation about a full cache. Lemma 5. When the request sequence requests the N -th distinct page, we have t = N , and this remains an invariant from that point onward. No items are discarded from the cache (main or history) until either t1 + b1 = N or ℓ1 + ℓ2 = 2N . By the time the request sequence requests the 2N -th distinct page, we have either t1 + b1 = N or ℓ1 + ℓ2 = 2N . Proof. Once the request sequence requests the N -th distinct page, it is obvious that we will have t = N , since until then, no item is evicted from T1 ∪ T2 ∪ B1 ∪ B2 . (Note that Replace only moves items from the main part to the history, i.e., from T1 ∪ T2 to B1 ∪ B2 .) Also, until then, p does not change. From that point forward, the algorithm never evicts any item from T1 ∪ T2 without replacing it with some other item. Thus, t = N is an invariant once it is satisfied. The history remains empty until the main cache is filled, i.e., t = N. From the pseudocode it is clear that items are discarded from the cache in statements 14, 17, and 21; no discards happen from the cache until either t1 + b1 = N (statement 12) or ℓ1 + ℓ2 = 2N (statement 20). If ℓ1 + ℓ2 = 2N is reached, since t1 + b1 ≤ N , we are guaranteed that t2 + b2 ≥ N and b1 + b2 = N , both of which will remain true from that point onward. Thus, by the time the 2N -th distinct page is requested, we have reached either t1 + b1 = N or ℓ1 + ℓ2 = 2N . We assume that request σ is processed in two distinct steps: first when Opt services the page request and, next when Arc services the request. We will show that inequality (11) is satisfied for each of the two steps. 15 Step 1: Opt services request σ Since only Opt acts in this step, CA = 0, and the contents of Arc’s cache does not change. There are two possible cases: either Opt faults on σ or it does not. Assume that page x is requested on request σ. If Opt does not fault on this request, then CO = 0. Since the contents of the cache maintained by Opt does not change, and neither do the lists L1 and L2 , we have ∆Φ = 0, and CA (σ) + ∆Φ ≤ 4N · CO (σ) ≤ 0. If Opt faults on request σ, then CO = 1. The contents of the cache maintained by Opt does change, which will affect the potential function. Opt will bring in page x into its cache. Assume that it evicts page y from its cache. The entry of page x into Opt’s cache can only decrease the potential function. The exit of page y from Opt’s cache can increase the potential function by at most 4N . The reason is as follows. Since the sum of b′1 , b′2 , t′1 , t′2 cannot exceed the size of Opt’s cache, we have 0 ≤ b′1 + t′1 + b′2 + t′2 ≤ N . Since b′1 + 2t′1 + 3b′2 + 4t′2 ≤ 4(b′1 + t′1 + b′2 + t′2 ), the left hand side cannot decrease by more than 4N . Thus, CA (σ) + ∆Φ1 ≤ 4N , proving inequality (11). Step 2: Arc services request σ There are four possible cases, which correspond to the four cases in Arc’s replacement algorithm. Case 1 deals with the case when Arc finds the page in its cache. The other three cases assume that Arc faults on this request because the item is not in T1 ∪ T2 . Cases 2 and 3 assume that the missing page is found recorded in the history in lists B1 and B2 , respectively. Case 4 assumes that the missing page is not recorded in history. Case I: Arc has a page hit. Clearly, CA = 0. We consider several subcases. In each case, the requested page will be moved to M RU (T2 ) while shifting other pages in T2 down. Case I.1 If the requested page is in T1′ , the move of this page from T1′ to T2′ implies ∆t′1 = −1; ∆t′2 = +1 and ∆Φ = −(2 · ∆t′1 + 4 · ∆t′2 ) = −2. Case I.2 If the requested page is in T2′ , the move of this page to M RU (T2 ) does not change the set of items in T2′ . Thus, ∆t′1 = ∆t′2 = 0 and ∆Φ = 0. Case I.3 If the requested page is in T1 − T1′ , then ∆t′1 = 0; ∆t′2 = +1 and ∆Φ = −4. One subtle point to note is that moving x from T1 − T1′ could potentially increase t′1 if the following conditions are met: x is located just below T1′ in T1 , it is not in Opt’s cache, and the items in T1 immediately below it are in Opt. However, x is already in Opt’s cache and there must be some item above it in T1 that is not in Opt. Case I.4 If the requested page is in T2 − T2′ , then ∆t′2 = +1 and ∆Φ = −4. The subtle point mentioned in Case I.3 also applies here. Next we will analyze the three cases when the requested page is not in Arc’s cache. Since CA = 1, the change in potential must be at most -1 in order for inequality (11) to be satisfied. We make the following useful observations in the form of lemmas. Lemma 6. If Arc has a miss and if the page is not in Arc’s history, we have ℓ′ = t′1 + t′2 + b′1 + b′2 < N . Consequently, we also have ℓ′1 < N and ℓ′2 < N . Proof. Since Opt has just finished serving the request, the page is present in the cache maintained by Opt just before Arc starts to service the request. If Arc has a miss, there is at least one page in the cache maintained by Opt that is not present in the cache maintained by Arc, implying that l′ < N . By definition, ℓ′ = ℓ′1 + ℓ′2 = t′1 + t′2 + b′1 + b′2 . Thus, the lemma holds. Lemma 7. A call to procedure Replace either causes an element to be moved from T1 to B1 or from T2 to B2 . In either case, the change in potential due to Replace, denoted by ∆ΦR , has an upper bound of 1. 16 Proof. Procedure Replace is only called when Arc has a page miss. Clearly, it causes an item to be moved from T1 to B1 or from T2 to B2 . If that item is in T1′ (or T2′ ), then T1 = T1′ (T2 = T2′ , resp.) and the moved item becomes part of B1′ (B2′ , resp.). Because the coefficients of b′1 and t′1 (b′2 and t′2 , resp.) differ by 1, we have ∆ΦR = +1. On the other hand, if that element is in T1 − T1′ (T2 − T2′ , resp.), then B1′ (B2′ , resp.) was empty before the move and remains empty after the move, and thus, ∆ΦR = 0. Lemma 8. On an Arc miss after phase P (0), if T1 = T1′ then the Replace step will not move a page from T2′ to B2 . On the other hand, if T2 = T2′ then Replace will not move a page from T1′ to B1 . Proof. In an attempt to prove by contradiction, let us assume that T1 = T1′ and T2 = T2′ are simultaneously true and Arc has a miss. By Lemma 5, we know that after phase, we have t = t1 + t2 = N , which by our assumption means that t′1 + t′2 = N ; this is impossible by Lemma 6. Thus, if T1 = T1′ , then T2 6= T2′ . Consequently, if LRU (T2 ) is moved to B2 , this item cannot be from T2′ . By a symmetric argument, if T2′ = T2 , then T1 6= T1′ , and LRU (T1 ) is not in T1′ . Case II: Arc has a miss and the missing page is in B1 Note that in this case the value of p will change by +1, unless its value equals N , in which case it has no change. Thus ∆p ≤ 1. If the missing item is in B1′ , then ∆b′1 = −1 and ∆t′2 = +1. Adding the change due to Replace, we get ∆Φ ≤ 1 − (∆b′1 + 4 · ∆t′2 ) + ∆ΦR ≤ −1 If the missing item is in B1 − B1′ , then we have ∆t′2 = 1 and ∆b′1 = 0. Thus, we have ∆Φ ≤ 1 − (∆b′1 + 4 · ∆t′2 ) + ∆ΦR ≤ −2 Case III: Arc has a miss and the missing page is in B2 . Note that in this case the value of p will change by -1, if its value was positive, otherwise it has no change. Thus ∆p ≤ 0. If the requested item is in B2′ , then ∆t′2 = 1, and ∆b′2 = −1. Thus, we have ∆Φ = ∆p − (3 · ∆b′2 + 4 · ∆t′2 ) + ∆ΦR ≤ 0 But this is not good enough since we need the potential change to be at most -1. When ∆p = −1, then we get the required inequality ∆Φ ≤ −1. Clearly, the difficulty is when ∆p = 0, which happens when p = 0. Since the missing item is from b′2 , it implies that B2′ is non-empty and T2′ = T2 . By Lemma 8 above, there must be at least one item in T1 − T1′ , which means that means that t1 > 0. As per the algorithm, since T1 is non-empty and p = 0, we are guaranteed to replace LRU (T1 ), and not an element from T1′ . Therefore, Replace will leave t′1 and b′1 unchanged, implying that ∆ΦR = 0. Thus, we have ∆Φ = ∆p − (3 · ∆b′2 + 4 · ∆t′2 ) + ∆ΦR ≤ −1 If the requested item is from B2 − B2′ , then ∆t′2 = 1, and ∆b′2 = 0. Thus, we have ∆Φ ≤ ∆p − (4 · ∆t′2 ) + ∆ΦR ≤ −3 Case IV: Arc has a miss and the missing page is not in B1 ∪ B2 17 We consider two cases. First, when ℓ1 = N , Arc will evict the LRU (L1 ). Since by Lemma 6, ℓ′1 < N , we know that for this case, b′1 remains unchanged at 0 and ∆t′1 = +1. Thus, ∆Φ ≤ ≤ −(2 · ∆t′1 ) + ∆ΦR −1 On the other hand, if ℓ1 < N , then Arc will evict the LRU (L2 ). Again, if the cache is full (i.e., t1 + t2 = N and ℓ1 + ℓ2 = 2N ), then we know that ℓ2 > N , which means that L′2 6= L2 and LRU (L2 ) is not in L′2 . Thus, deletion of LRU (L2 ) = LRU (B2 ) will not affect b′2 or any of the other quantities in the potential function. Then comes the Replace step, for which a bound has been proved earlier. Finally, a new item is brought in and placed in M RU (T1 ). Thus ∆t′1 ≤ 1. Putting it all together, we have ∆Φ ≤ −(2 · ∆t′1 ) + ∆ΦR ≤ −1 Wrapping up the proof of Theorem 8 Tying it all up, we have shown that inequality (11) holds for every request made after the cache is full, i.e., CA (σ) + ∆Φ ≤ 4N · CO (σ). If we assume that the caches started empty, then the initial potential is 0, while the final potential can be at most 4N . Thus, we have CA (σ) ≤ 4N · CO (σ) + 4N, thus proving Theorem 8. 4.6 Analyzing the Competitiveness of CAR Next, we analyze the competitiveness of Car. The main result of this section is the following: Theorem 9. Algorithm Car is 18N -competitive. Proof. Let PX [q] be the position of page q in an arbitrary ordered sequence of pages X. When the set is obvious, we will drop the subscript and denote PX [q] simply by P [q]. The set of history pages B1 and B2 will be treated as an ordered sequence of pages ordered from its LRU position to its MRU position. The set of main pages T10 (resp., T20 , T11 , and T21 ) will be treated as an ordered sequence of unmarked (resp., unmarked, marked, and marked) pages in T1 (resp, T2 , T1 , and T2 ) ordered from head to tail. Let Opt and Car be the set of (main and history) pages stored in the caches for algorithms Opt and Car respectively. Let D = (T1 ∪ T2 ∪ B1 ∪ B2 ) \ Opt. Thus D consists of pages in Car but not in Opt. We associate each page with a rank value R[q], which is defined as follows:  PB1 [q]      PB2 [q]    2P 0 [q] + b 1 T1 R[q] =  0 2P [q] + b 2  T2    3N + 2PT11 [q] + b1    3N + 2PT21 [q] + b2 if if if if if if q q q q q q ∈ B1 ∈ B2 ∈ T10 ∈ T20 ∈ T11 ∈ T21 (12) Finally, we define the potential function as follows: Φ= X  1 (p + 2(b1 + t1 ) + 3 R[q]) N − NO + 1 q∈D 18 (13) The initial value of Φ is 0. If the following inequality (14) is true for any request σ, where ∆Φ is the change in potential caused by serving the request, then when summed over all requests, it proves Theorem 9. CCar (σ) + ∆Φ ≤  18N COpt (σ). N − NO + 1 (14) As before, we assume that request σ is processed in two distinct steps: first when Opt serves and, next when Car serves. We will show that inequality (14) is satisfied for each of the two steps. Step 1: Opt serves request σ Since only Opt acts in this step, CCar = 0, and T1 ∪ T2 does not change. There are two possible cases: either Opt faults on σ or it does not. If Opt does not fault on this request, then it is easy to see that COpt = 0 and ∆Φ = 0, thus satisfying inequality (14). If Opt faults on request σ, then COpt = 1 and some page, q, is evicted from the cache maintained by Opt. If q is maintained by Car then it follows that q will belong to D after this step and thus its rank will contribute to the potential function, which will increase by three times the rank of q. The maximal positive change in potential will occur when q is the marked head page in T2 . In this case the rank of q is given by: R[q] = 3N + 2P [q] + b2 . The maximal possible values for each of the terms P [q] and b2 will be N , hence the maximum possible rank of q will be 3N + 2N + N = 6N . Therefore resulting potential change is at most 3(6N ) = 18N . Step 2: Car serves request σ We break down the analysis into four cases. Case 2.1 deals with the case when Car finds the page in its cache. The other three cases assume that Car faults on this request because the item is not in T1 ∪T2 . Cases 2.2 and 2.3 assume that the missing page is found recorded in the history in lists B1 and B2 , respectively. Case 2.4 assumes that the missing page is not recorded in history. Case 2.1: Car has a page hit Clearly, the page was found in T1 ∪ T2 , and CCar = 0. We consider the change of each of terms in the potential function individually. 1. As per the algorithm, p can only change when the page is found in history. (See lines 14 through 20 of Car(x).) Since the page is not found in Car’s history, ∆p = 0. 2. Neither the cache nor the history lists maintained by Car will change. Thus, the contribution to the second term in Φ, i.e., 2(b1 + t1 ) does not change. 3. Since Opt has already served the page, the page is in Opt’s cache. Therefore, even if the page gets marked during this hit, its rank value does not change. Thus, the contribution to the last term in Φ, also remains unchanged. We, therefore, conclude that ∆Φ = 0, satisfying inequality (14). Next we will analyze the three cases when the requested page is not in Car’s cache. Since CCar = 1, the change in potential must be at most −1 in each case in order for inequality (14) to be satisfied. Before tackling the three cases, the following lemmas (9 and 10) Pare useful for understanding the potential change caused by the last term in the potential function, i.e., q∈D R[q]. It is worth pointing out that a call to Replace moves either an item from T1 to B1 or from T2 to B2 , which is exactly the premise of Lemma 9 below. Lemma 9. When a page is moved from T1 to B1 (or from T2 to B2 ) its rank decreases by at least 1. Proof. Let q be any page in T1 . In order for q to be moved from T1 to B1 it must have been unmarked and located at the head of T1 . Since PT1 [q] = 1, the rank of q prior to the move must have been R[q] = 2PT1 [q] + b1 = b1 + 2, where b1 is the size of B1 prior to moving q. 19 After q is moved to the MRU position of B1 , R[q] = PB1 [q] = b1 + 1. Thus its rank decreased by 1. The arguments for the move from T2 to B2 are identical with the appropriate changes in subscripts. P Lemma 10. When Car has a page miss, the term q∈D R[q] in the potential function Φ cannot increase. Proof. We examine the rank change based on the original location of the page(s) whose ranks changed and in each case show that the rank change is never positive. Wherever appropriate we have provided references to line numbers in Pseudocode Car(x) from Appendix. Case A: q ∈ B1 ∪ B2 The rank of q ∈ B1 , which is simply its position in B1 , can change in one of three different ways. 1. Some page x less recently used than q (i.e., PB1 [x] < PB1 [q]) was evicted (Line 7). In this case, it is clear that PB1 [q] decreases by at least 1. 2. The page q is the requested page and is moved to T2 (Line 16). In this case, q ∈ Opt and hence its rank cannot affect the potential function. 3. Some page x is added to MRU of B1 (Line 27). Since pages are ordered from LRU to MRU, the added page cannot affect the rank of q. Using identical arguments for q ∈ B2 , we conclude that a miss will not increase the rank of any page in B1 ∪ B2 . Case B: q ∈ T10 ∪ T20 The rank of page q ∈ T10 , defined as R[q] = 2PT10 [q] + b1 , may be affected in four different ways. 1. If page q is the head of T1 and gets moved to B1 (Line 27), by lemma 9, the change in rank of q is at most −1. 2. If an unmarked page x is added to the tail of T1 (Line 13), then since the ordering is from head to tail, it does not affect the position of page q. Since there was no change in b1 , it is clear that the change in R[q] is 0. 3. If the unmarked page x 6= q at the head of T1 is marked and moved to tail of T2 (Line 29), then P [q] decreases by at least 1. Since the content of B1 is unchanged, the change in R[q] = 2P [q] + b1 is at most -2. 4. If the unmarked page x 6= q at the head of T1 is moved to B1 (Line 29), then P [q] decreases by at least 1, and b1 increases by 1. Hence the change in R[q] = 2P [q] + b1 is at most -1. The arguments are identical for q ∈ T20 . In each case, we have shown that a miss will not increase the rank of any page in T10 ∪ T20 . Case C: q ∈ T11 The rank of page q ∈ T11 , defined as R[q] = 3N + 2PT11 [q] + b1 , may be affected in four different ways. 1. If an unmarked page x is added to the tail of T1 (Line 13), then since the ordering is from head to tail, it does not affect the position of page q. Since there was no change in b1 , it is clear that the change in R[q] is 0. 2. If the unmarked page x 6= q at the head of T1 is marked and moved to tail of T2 (Line 29), then P [q] decreases by at least 1. Since B1 is unchanged, the change in R[q] = 3N + 2P [q] + b1 is at most -2. 3. If the unmarked page x 6= q at the head of T1 is moved to B1 (Line 29), then P [q] decreases by at least 1, and b1 increases by 1. Hence the change in R[q] = 3N + 2P [q] + b1 is at most -1. 20 4. Next, we consider the case when the marked page q is the head of T1 and gets unmarked and moved to T2 (Line 29). Prior to the move, the rank of q is given by R[q] = 3N + 2PT11 [q] + b1 . Since B1 could be empty, we know that R[q] ≥ 3N + 2. After page q is unmarked and moved to T2 , its rank is given by R[q] = 2PT20 [q] + b2 . Since P [q] ≤ N and b2 ≤ N , we know that the new R[q] ≤ 3N . Thus, the rank of page q does not increase. In each case, we have shown that a miss will not increase the rank of any page in T11 . Case D: q ∈ T21 The rank of page q ∈ T21 , defined as R[q] = 3N + 2PT21 [q] + b2 , may be affected in four different ways. 1. If an unmarked page x is added to the tail of T2 (Lines 16, 19, or 29), and if b2 does not change, it is once again clear that the change in R[q] is 0. 2. If a marked page x 6= q at the head of T2 gets unmarked and moved to the tail of T2 (Line 36), the position of q will decrease by 1 and there is no change in b2 . Thus R[q] changes by at most -2. 3. If an unmarked page x at the head of T2 is moved to B2 (Line 34), P [q] decreases by 1 and b2 increases by 1. Thus R[q] changes by at most -1. 4. Finally, we consider the case when the marked page q is the head of T2 and gets unmarked and moved to the tail of T2 (Line 36). Prior to the move, the rank of q is given by R[q] = 3N + 2PT21 [q] + b2 . Even if B2 is empty, we know that R[q] ≥ 3N + 2. After page q is unmarked and moved to T2 , its rank is given by R[q] = 2PT20 [q] + b2 . Since P [q] ≤ N and b2 ≤ N , we know that the new R[q] ≤ 3N . Thus, the rank of page q does not increase. In each case, we have shown that a miss will not increase the rank of any page in T21 . The four cases (A through D) together complete the proof of Lemma 10. We continue with the remaining cases for the proof of Theorem 9. Case 2.2: Car has a page miss and the missing page is in B1 We consider the change in the potential function (defined in Eq. 13) by analyzing each of its three terms. 1. Value of p increases by 1, except when it is equal to N , in which case it remains unchanged. (See Line 15.) Thus, the first term increases by at most 1. 2. The call to Replace has no effect on the value of (t1 + b1 ) because an item is moved either from T1 to B1 or from T2 to B2 . Since the requested page in B1 is moved to T2 , (t1 + b1 ) decreases by 1. 3. By Lemma 10, we already know that the last term increases by at most 0. Since p increases by at most 1 and the term 2(t1 +b1 ) decreases by at least 2, the total change in the potential function, is at most -1. Case 2.3: Car has a page miss and the missing page is in B2 When the missing page is in B2 , Car makes a call to Replace (Line 5) and then executes Lines 18-19. Thus, p is decremented except if it is already equal to 0. We consider two subcases: ∆p < 0 and ∆p = 0. ∆p < 0: As in Case 2.2, the call to Replace has no effect on (t1 + b1 ). Since, Lines 18-19 do not affect T1 ∪ B1 , the second term does not change. By Lemma 10, we know that the last term increases by at most 0. Since ∆p ≤ −1, the total change in the potential function, ∆p + ∆2(t1 + b1 ) is at most -1. 21 ∆p = 0: Unlike the subcase above when p decreases by 1, the change in p cannot guarantee the required reduction in the potential. We therefore need a tighter argument. We know that there is a call to Replace. Three cases arise and are discussed below. • If T1 is empty, then T2 must have N pages, at least one of which must be in D. Also, Replace must act on T2 , eventually evicting an unmarked page from head of T2 , causing the rank of any page from T2 \ Opt to decrease by 1. • If T1 is not empty and has at least one page from D, then the condition in Line 24 passes and Replace must act on T1 , eventually evicting an unmarked page from head of T1 , causing the rank of at least one page from T1 \ Opt to decrease by 1. • Finally, if T1 is not empty and all its pages are in Opt, then T2 must have a page q ∈ D. Since the requested page x was found in B2 and is moved to the tail of T2 , even though the position of q in T2 does not change, b2 decreased by 1 and consequently the rank of q decreases by 1. Thus, in each case, even though neither p nor the quantity (t1 + b1 ) changed, the third term involving ranks, and consequently, the potential function decreased by at least 3. The following two lemmas are useful for Case 2.4, when the missing page is not in T1 ∪ T2 ∪ B1 ∪ B2 . Lemma 11. We make two claims: 1. If t1 + b1 = N and the LRU page of B1 is evicted from the cache on Line 7, then decrease by at least one. 2. If t2 + b2 > N , and the LRU page of B2 , is evicted from the cache on Line 9, then decrease by at least one. P q∈D P q∈D R[q] will R[q] will Proof. We tacke the first claim. Assume that y is the LRU page of B1 that is being evicted on Line 7. Then Car must have had a page miss on x 6∈ B1 ∪ B2 , and the requested page x is added to the tail of T1 . Since t1 + b1 = N , there is at least one page q ∈ T1 ∪ B1 that is not in Opt’s cache and whose rank contributes to the potential function. First, we assume that q ∈ T1 \ Opt, whose rank is given by: R[q] = 2 ∗ P [q] + b1 . For each of the three cases, we show that the potential function does decrease by at least 1. • If Replace acts on T1 and the unmarked head of T1 , different from q, is moved to B1 then the size of B1 remains the same (because a page gets added to B1 while another page is evicted) but the position of q in T1 decreases by one. Therefore R[q] decreases by 2. • If Replace acts on T1 and q itself is moved to B1 then by Lemma 9, R[q] decreases by at least 1. • If Replace acts on T2 , then we use the fact that a page is evicted from B1 , and the b1 term in R[q] must decrease by 1. Next, we assume that q ∈ B1 \ Opt. Since LRU (B1 ) is evicted, the position of the page q will decrease by one. Thus R[q] = PB1 [q] must decrease by at least 1, completing the proof of the first claim in the lemma. The proof of the second claim is very similar and only requires appropriate changes to the subscripts. Next we tackle the last case in the proof of Theorem 9. Case 2.4: Car has a page miss and the missing page is not in B1 ∪ B2 We assume that Car’s cache is full (i.e., l1 + l2 = 2N ). We consider two cases below – first, if l1 = N and the next when l1 < N . If l1 = t1 + b1 = N , Car will call Replace, evict LRU (B1 ) and then add the requested page to the tail of T1 . Below, we analyze the changes to the three terms in the potential function. • Since p is not affected, the first term does not change. • Since a page is added to T1 and a page is evicted from B1 , the net change in the second term is 0. • Since the conditions of Lemma 11 apply, the total rank will decrease by at least 1. 22 Adding up all the changes, we conclude that the potential function decreases by at least 3. If l1 < N , Car will call Replace, evict LRU (B2 ) and then add a page to the tail of T1 . As above, we analyze the changes to the three terms in the potential function. • Since p is not affected, the first term does not change. • A page is added to T1 and a page is evicted from B2 hence (t1 + b1 ) increases by 1. • Since l2 > N , the conditions of Lemma 11 apply, the total rank will decrease by at least 1. Adding up all the changes, we conclude that the potential function decreases by at least 1, thus completing Case 2.4. Wrapping up the proof of Theorem 9: Combining the four cases (2.1 through 2.4) proves that inequality (14) is satisfied when Car serves request σ. This completes the proof of Theorem 9, establishing that the upper bound on the competitiveness of Car is 18N . 5 Conclusions and Future Work Adaptive algorithms are tremendously important in situations where inputs are infinite online sequences and no single optimal algorithm exists for all inputs. Thus, different portions of the input sequence require different algorithms to provide optimal responses. Consequently, it is incumbent upon the algorithm to sense changes in the nature of the input sequence and adapt to these changes. Unfortunately, these algorithms are harder to analyze. We present the analysis of two important adaptive algorithms called Arc and Car and show that they are competitive along with proving good lower bounds on the competitiveness ratios. Two important open questions remain unanswered. Given that there is a gap between the lower and upper bounds on the competitiveness ratios of the two adaptive algorithms, Arc and Car, what is the true ratio? More importantly, is there an “expected” competitiveness ratio for request sequences that come from real applications? The second question would help explain why Arc and Car perform better in practice than Lru and Clock, respectively. Acknowledgments This work was partly supported by two NSF Grants (CNS-1018262 and CNS-1563883) and the NSF Graduate Research Fellowship (DGE-1038321). We are grateful to Kirk Pruhs for suggesting enhancing our results with the assumption of unequal cache sizes. 23 References [Alb96] S. Albers. Competitive online algorithms. Technical report, BRICS Lecture Series, Computer Science Department, University of Aarhus, 1996. [BM04] S. Bansal and D. S. Modha. CAR: CLOCK with adaptive replacement. In Proceedings of the 3rd USENIX Conference on File and Storage Technologies, FAST ’04, pages 187–200, Berkeley, CA, USA, 2004. USENIX Association. [Cor68] F. J. Corbato. A paging experiment with the MULTICS system. Technical report, DTIC Document, 1968. [Fri99] M. B. Friedman. Windows NT page replacement policies. In Proceedings of the Intl. CMG Conference, pages 234–244, 1999. [Hoc97] D. S. Hochbaum, editor. Approximation algorithms for NP-hard problems. PWS Publishing Co., Boston, MA, USA, 1997. [JCZ05] S. Jiang, F. Chen, and X. Zhang. CLOCK-Pro: An effective improvement of the CLOCK replacement. In USENIX Annual Technical Conference, General Track, pages 323–336, 2005. [JIPP10] A. Janapsatya, A. Ignjatovic, J. Peddersen, and S. Parameswaran. Dueling CLOCK: adaptive cache replacement policy based on the CLOCK algorithm. In Design, Automation & Test in Europe Conference & Exhibition (DATE), 2010, pages 920–925. IEEE, 2010. [JS94] T. Johnson and D. Shasha. 2Q: A low overhead high performance buffer management replacement algorithm. In Proc. of VLDB, pages 297–306, 1994. [JZ02] S. Jiang and X. Zhang. LIRS: An efficient low inter-reference recency set replacement policy to improve buffer cache performance. In Proc. ACM Sigmetrics Conf., pages 297–306. ACM Press, 2002. [LCK+ 01] D. Lee, J. Choi, J. H. Kim, S. H. Noh, S. L. Min, Y. Cho, and C. S. Kim. LRFU: A spectrum of policies that subsumes the least recently used and least frequently used policies. IEEE Trans. Comput., 50(12):1352–1361, December 2001. [MM03] N. Megiddo and D. S. Modha. ARC: A self-tuning, low overhead replacement cache. In Proceedings of the 2nd USENIX Conference on File and Storage Technologies, FAST ’03, pages 115–130, Berkeley, CA, USA, 2003. USENIX Association. [MM04] N. Megiddo and D. S. Modha. Outperforming LRU with an adaptive replacement cache algorithm. IEEE Computer, 37(4):58–65, 2004. [OOW93] E. J. O’Neil, P. E. O’Neil, and G. Weikum. The LRU-K page replacement algorithm for database disk buffering. SIGMOD Rec., 22(2):297–306, June 1993. [ST85] D. D. Sleator and R. E. Tarjan. Amortized efficiency of list update and paging rules. Commun. ACM, 28(2):202–208, February 1985. 24 6 Appendix We reproduce the pseudocode for Arc and Car below. Pseudocode: Arc(x) INPUT: The requested page x INITIALIZATION: Set p = 0 and set lists T1 , B1 , T2 , and B2 to empty 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: if (x is in T1 ∪ T2 ) then Move x to the top of T2 else if (x is in B1 ) then Adaptation: Update p = min{p + 1, N } Replace() Fetch x and move to the top of T2 else if (x is in B2 ) then Adaptation: Update: p = max{p − 1, 0} Replace() Fetch x and move to the top of T2 else if (t1 + b1 = N ) then if (t1 < N ) then Discard LRU item in B1 Replace() else Discard LRU page in T1 and remove from cache end if else if ((t1 + b1 < N ) and (t1 + t2 + b1 + b2 ≥ N )) then if (t1 + t2 + b1 + b2 = 2N ) then Discard LRU item in B2 end if Replace() end if Fetch x and move to the top of T1 end if Replace() 26: if ((t1 ≥ 1) and ((x ∈ B2 and t1 = p) or (t1 > p))) then 27: Discard LRU page in T1 and insert as MRU history item in B1 28: else 29: Discard LRU page in T2 and insert as MRU history item in B2 30: end if 25 ⊲ cache hit ⊲ cache history hit ⊲ learning rate = 1 ⊲ make space in T1 or T2 ⊲ cache history hit ⊲ learning rate = 1 ⊲ make space in T1 or T2 ⊲ cache and history miss ⊲ make space in T1 or T2 ⊲ make space in T1 or T2 Pseudocode: Car(x) INPUT: The requested page x INITIALIZATION: Set p = 0 and set lists T1 , B1 , T2 , and B2 to empty 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: if (x is in T1 ∪ T2 ) then ⊲ cache hit Mark page x else ⊲ cache miss if (t1 + t2 = N ) then ⊲ cache full, replace a page from cache Replace() ⊲ make space in T1 or T2 if ((x 6∈ B1 ∪ B2 ) and (t1 + b1 = N )) then Discard LRU page in B1 else if ((x 6∈ B1 ∪ B2 ) and (t1 + t2 + b1 + b2 = 2N )) then Discard LRU page in B2 . end if end if if (x 6∈ B1 ∪ B2 ) then ⊲ cache miss Insert x at the tail of T1 ; Unmark page x else if (x ∈ B1 ) then ⊲ cache history hit Adaptation: Update p = min{p + 1, N } ⊲ learning rate = 1 Move x to the tail of T2 ; Unmark page x else ⊲ cache history hit Adaptation: Update: p = max{p − 1, 0} ⊲ learning rate = 1 Move x to the tail of T2 ; Unmark page x end if end if Replace() 22: found = false 23: repeat 24: if (t1 ≥ max{1, p}) then 25: if (head page in T1 is unmarked) then 26: found = true 27: Discard head page in T1 and insert as MRU history item in B1 28: else 29: Unmark head page in T1 , move page as tail page in T2 , and move head of T1 clockwise 30: end if 31: else 32: if (head page in T2 is unmarked) then 33: found = true 34: Discard head page in T2 and insert as MRU history item in B2 35: else 36: Unmark head page in T2 , and move head of T2 clockwise 37: end if 38: end if 39: until (found) 26
8cs.DS
arXiv:1712.01542v1 [math.RA] 5 Dec 2017 CAPABLE LIE ALGEBRAS WITH THE DERIVED SUBALGEBRA OF DIMENSION TWO OVER AN ARBITRARY FIELD PEYMAN NIROOMAND, FARANGIS JOHARI, AND MOHSEN PARVIZI Abstract. In this paper, we classify all capable nilpotent Lie algebras with the derived subalgebra of dimension 2 over an arbitrary field. Moreover, the explicit structure of such Lie algebras of class 3 is given. 1. Introduction and Motivation The concept of capability was introduced by P. Hall in [9]. Recall that a group G is called capable if there exists some group E such that G ∼ = E/Z(E), where Z(E) denotes the center of E. There are some fundamental known results concerning capability of p-groups. For instance, in [3, Corollary 4.16], it is shown the only capable extra-special p-groups (the p-groups with Z(G) = G′ and |G′ | = p) are those of order p3 and exponent p. In the case that G′ = Z(G) and Z(G) is elementary abelian p-group of rank 2, Heineken in [11] proved that the capable ones has order at most p7 . By some results due to Lazard, we may associate a p-group to a Lie algebra. Therefore some results of Lie algebras and p-groups have similarities in the structure. But in this way not every thing are the same and there are differences between groups and Lie algebras, so that most of time the proofs are different. Similar to the concept of the capability for groups, a Lie algebra is called capable provided that L∼ = H/Z(H) for a Lie algebra H. Beyl et al. in [3] introduced the epicenter Z ∗ (G) of a group G that plays an important role in the capability of G. The analogous notion of the epicenter, Z ∗ (L) for a Lie algebra L was defined in [19]. It is shown that L is capable if and only if Z ∗ (L) = 0. Another notion having relation to the capability is the concept of exterior square of Lie algebras, L ∧ L, which was introduced in [6]. Our approach is on the concept of the exterior center Z ∧ (L), the set of all elements l of L for which l ∧l′ = 0L∧L for all l′ ∈ L. Niroomand et al. in [17] showed Z ∧ (L) = Z ∗ (L) for any finite dimensional Lie algebra L. In [17], the last two authors obtained the structure of a capable nilpotent Lie algebra L when dim L2 ≤ 1. It developed the result of [3, Corollary 4.16] for groups to the class of Lie algebras. Recall that from [18], a Lie algebra H is called generalized Heisenberg of rank n if H 2 = Z(H) and dim H 2 = n. If n = 1, then H is a Heisenberg Lie algebra that is more well-known. Such algebras are odd dimension and have the presentation H(m) ∼ = ha1 , b1 , . . . , am , bm , z [al , bl ] = z, 1 ≤ l ≤ mi. Date: December 6, 2017. Key words and phrases. Capability, Schur multiplier, generalized Heisenberg Lie algebras, stem Lie algebras. Mathematics Subject Classification 2010. Primary 17B30; Secondary 17B05, 17B99. 1 2 P. NIROOMAND, F. JOHARI, AND M. PARVIZI Recently, Niroomand et al. in [18] proved the capable generalized Heisenberg Lie algebras of rank 2 have dimension at most 7 over a filed of characteristic not equal to 2. They developed the result of Heineken [11] for groups to the area of Lie algebras. They also characterized the structure of all capable nilpotent Lie algebras of class two when dim L2 = 2. In virtue of the recent results in [18], in this paper we intend to classify the structure of all capable nilpotent Lie algebras of class two with the derived subalgebra of dimension 2 over an arbitrary filed. Furthermore, we determine the structure of all nilpotent Lie algebras of class 3 with the derived subalgebra of dimension 2 and then we specify which one of them are capable. 2. Preliminaries All Lie algebras in this paper are finite dimensional. The Schur multiplier of a Lie algebra L, M(L), is defined as M(L) ∼ = F/R and F = R ∩ F 2 /[R, F ] where L ∼ is a free Lie algebra. It can be shown that the Lie algebra M(L) is abelian and independent of the choice of the free Lie algebra F (see [1, 2, 4, 10, 13, 14, 15, 16, 17] for more information on this topics). Throughout the paper ⊗ used to denote the operator of usual tensor product of algebras. For a Lie algebra L, we denote the factor Lie algebra L/L2 by L(ab) . Also we denote an abelian Lie algebra of dimension n by A(n). The following proposition plays a key role in detecting the capability of Lie algebras. Proposition 2.1. Let L be a finite dimensional Lie algebra with a central ideal I. Then (i) dim(M(L)) ≥ dim(M(L/I)) − dim(L2 ∩ I), (ii) dim(M(L)) = dim(M(L/I)) − dim(L2 ∩ I) if and only if I ⊆ Z ∗ (L). Proof. The result follows from [19, Proposition 4.1(iii) and Theorem 4.4 ].  The following lemma from [19] is a useful instrument in the next investigations. Lemma 2.2. Let I be an ideal of a Lie algebras L such that L/I is capable. Then Z ∗ (L) ⊆ I. The next corollary shows that the epicenter of a Lie algebra is contained in its derived subalgebra. Corollary 2.3. Let L be a finite dimensional non-abelian nilpotent Lie algebra. Then Z ∗ (L) ⊆ L2 . Proof. Since dim L/L2 ≥ 2, L/L2 is capable, using [17, Theorem 3.3]. Therefore the result follows by Lemma 2.2.  We need the notion of a central product of Lie algebras, it is defined as follows. Definition 2.4. The Lie algebra L is a central product of A and B, if L = A + B, where A and B are ideals of L such that [A, B] = 0 and A ∩ B ⊆ Z(L). We denote the central product of two Lie algebras A and B by A ∔ B. The Heisenberg Lie algebra can be presented in terms of central products. Lemma 2.5. [12, Lemma 3.3] Let L be a Heisenberg Lie algebra of dimension 2m + 1. Then L is central products of its ideals Bj for all 1 ≤ j ≤ m such that Bj is the Heisenberg Lie algebra of dimension 3. LIE ALGEBRAS WITH THE DERIVED SUBALGEBRA OF DIMENSION TWO 3 It is not an easy matter to determine the capability of a central product in general. The next result gives the answer in a particular case. Proposition 2.6. [18, Proposition 2.2] Let L be a Lie algebra such that L = A ∔ B with A2 ∩ B 2 6= 0. Then A2 ∩ B 2 ⊆ Z ∧ (L) and so L is non-capable. Following [18, Propositions 2.4 and 2.6], for determining the capable nilpotent Lie algebras of class 2 with the derived subalgebra of dimension 2, it is enough to consider a generalized Heisenberg Lie algebra H when 5 ≤ dim H ≤ 7. Throughout the paper multiplication tables with respect to a fixed basis with trivial products of the form [x, y] = 0 omitted, when x, y belongs to the Lie algebra. Nilpotent Lie algebras of dimension at most 5 can be described uniformly over all fields in [5, 7, 8]. For the dimension 6, 7 over an algebraic closed filed the structure are known in [5, 7, 8]. Using notation and terminology of [5, 7, 8], we list the generalized Heisenberg Lie algebras of rank two of dimension at most 6 and 7 over field of characteristic 2 and characteristic different 2, respectively. Recall that in a field F of characteristic 2, ω denotes a fixed element from F \ {x2 + x|x ∈ F}. Theorem 2.7. Let H be a generalized Heisenberg Lie algebra of rank 2. Then (i) over a field F of characteristic 2, the list of the isomorphism types of generalized Heisenberg Lie algebras of dimension at most 6 are the fol(2) lowing: L5,8 = hx1 , . . . , x5 [x1 , x2 ] = x4 , [x1 , x3 ] = x5 i and L6,7 (η) = hx1 , . . . , x6 |[x1 , x2 ] = x5 , [x1 , x3 ] = x6 , [x2 , x4 ] = ηx6 , [x3 , x4 ] = x5 + x6 i where η ∈ {0, ω}, (ii) over a field F of characteristic different from 2, the list of the isomorphism types of generalized Heisenberg Lie algebras of dimension at most 7 are L5,8 = hx1 , . . . , x5 [x1 , x2 ] = x4 , [x1 , x3 ] = x5 i, L6,22 (ǫ) = hx1 , . . . , x6 [x1 , x2 ] = x5 = [x3 , x4 ], [x1 , x3 ] = x6 , [x2 , x4 ] = ǫx6 i, ∗ where ǫ ∈ F/(∼) and char F 6= 2, L1 = 27A = hx1 , . . . , x7 [x1 , x2 ] = x6 = [x3 , x4 ], [x1 , x5 ] = x7 = [x2 , x3 ]i, L2 = 27B = hx1 , . . . , x7 [x1 , x2 ] = x6 , [x1 , x4 ] = x7 = [x3 , x5 ]i. 3. capable generalized Heisenberg Lie algebras of rank two Here, we are going to determine the structures of capable generalized Heisenberg Lie algebras of rank two. By [18, Proposition 2.6], generalized Heisenberg Lie algebras with the derived subalgebra of dimension 2 are capable if their dimension lies between 5 and 7. According to Theorem 2.7, we have the presentation of all capable generalized Heisenberg Lie algebras of rank two over a filed F with char F 6= 2. But when char F = 2 the structure of them is unknown. Therefore, at first we intend to find the structure of them on an arbitrary filed and then we determine which ones are capable. Theorem 3.1. Let L be a 7-dimensional generalized Heisenberg Lie algebra over a filed F with char F = 2 of rank two. Then L ∼ = hx1 , . . . , x7 [x1 , x2 ] = x6 = [x3 , x4 ], [x1 , x5 ] = x7 = [x2 , x3 ]i or L ∼ = hx1 , . . . , x7 [x1 , x2 ] = x6 , [x1 , x4 ] = x7 = [x3 , x5 ]i. 4 P. NIROOMAND, F. JOHARI, AND M. PARVIZI Proof. Let L2 = hz1 i ⊕ hz2 i. By [17, Theorem 3.6], we have L/hz2 i ∼ = H(2) ⊕ A(1) or L/hz2 i ∼ = H(1) ⊕ A(3). First suppose that L/hz2 i ∼ = H(2) ⊕ A(1). There exist two ideals I1 /hz2 i and I2 /hz2 i of L/hz2 i such that I1 /hz2 i ∼ = H(2) and I2 /hz2 i ∼ = A(1). Clearly, L = I1 + I2 , I1 ∩ I2 = hz2 i, [I1 , I2 ] ⊆ hz2 i and [L, I2 ] ⊆ hz2 i. Thus I1 /hz2 i = hx1 + hz2 i, y1 + hz2 i, x2 + hz2 i, y2 + hz2 i, z1 + hz2 i|[x1 , y1 ] + hz2 i = [x2 , y2 ] + hz2 i = z1 + hz2 ii and I2 = hqi ⊕ hz2 i ∼ = A(2) for some q ∈ L. Hence the set {x1 , y1 , x2 , y2 , z1 , z2 , q} is a basis of L and [x1 , y1 ] = z1 + α1 z2 , [x2 , y2 ] = z1 + α2 z2 , [x1 , y2 ] = α3 z2 , [x1 , x2 ] = α4 z2 , [x2 , y1 ] = α5 z2 , [y1 , y2 ] = α6 z2 , [x1 , q] = α7 z2 , [y1 , q] = α8 z2 , [x2 , q] = α9 z2 , [y2 , q] = α10 z2 . By changing variable, we assume that α1 = α2 = 0. Since q ∈ / L2 = Z(L) = hz1 i⊕hz2 i, q is not central. Thus [L, I2 ] = [hqi, I2 ] = hz2 i. Without loss of generality, assume that α7 6= 0. By [17, Theorem 3.6], we have L/hz1 i ∼ = H(2) ⊕ A(1) or L/hz1 i ∼ = H(2) ⊕ A(1). There exist two = H(1) ⊕ A(3). First suppose that L/hz1 i ∼ ideals I3 /hz1 i and I4 /hz1 i of L/hz1 i such that ∼ H(2) and I4 /hz1 i = ∼ A(1). I3 /hz1 i = Clearly, L = I3 + I4 and [I3 , I4 ] ⊆ hz1 i. We claim that q, x1 ∈ I3 and [x2 , q] = [y1 , q] = [y2 , q] = [x1 , x2 ] = [x1 , y2 ] = 0. Let a+hz1 i ∈ I3 /hz1 i and I4 /hz1 i = hb+hz1 ii such that q+hz1 i = (a+hz1 i)+(αb+hz1 i). If a + hz1 i = 0, then since [L, I2 ] = [hqi, I2 ] = hz2 i, we have [hqi, L] ∈ hz1 i ∩ hz2 i = 0. So q ∈ Z(L) = L2 = hz1 i ⊕ hz2 i. It is a contradiction. Thus q − a − αb ∈ hz1 i and a + hz1 i 6= 0. We have a = η1 x1 + η2 x2 + η3 y1 + η4 y2 + η5 z1 + η6 z2 + η7 q, αb = η1′ x1 + η2′ x2 + η3′ y1 + η4′ y2 + η5′ z1 + η6′ z2 + η7′ q, and so q = a + αb + γz1 = η1 x1 + η2 x2 + η3 y1 + η4 y2 + η5 z1 + η6 z2 + η7 q+ η1′ x1 + η2′ x2 + η3′ y1 + η4′ y2 + η5′ z1 + η6′ z2 + η7′ q + γz1 . Since the set {x1 , y1 , x2 , y2 , z1 , z2 , q} is linearly independent and L is a Lie algebra on filed over characteristic two, η1 = η1′ , η2 = η2′ , η3 = η3′ , η4 = η4′ , η5 = −η5′ −γ, η6 = η6′ , η7 = 1 − η7′ . Thus q + hz1 i = a + hz1 i + αb + hz1 i = (η1 + η1′ )x1 + (η2 + η2′ )x2 + (η3 + η3′ )y1 + (η4 + η4′ )y2 + (η6 + η6′ )z2 + (η7 + η7′ )q + hz1 i. We conclude that q + hz1 i = (η7 + η7′ )q + hz1 i so η7 6= 0 or η7′ 6= 0. Since q + hz1 i ∈ / I4 /hz1 i, we have η7 6= 0 and η7′ = 0. Thus q + hz1 i ∈ I3 /hz1 i. Hence q ∈ I3 . Now, we prove that x1 ∈ I3 . By contrary, assume that x1 ∈ / I3 . We may assume that x1 + hz1 i = a1 + hz1 i + α′ b + hz1 i, where a1 + hz1 i ∈ I3 /hz1 i and I4 /hz1 i = hb + hz1 ii. Now, if a1 + hz1 i = 0, then x1 + hz1 i = α′ b + hz1i ∈ I3 /hz1 i and so [q, x1 ] + hz1 i = 0. Since [q, x1 ] = z2 , we have [q, x1 ] ∈ hz1 i ∩ hz2 i = 0 and hence [q, x1 ] = 0. It is a contradiction. Thus a1 + hz1 i = 6 0 and x1 − a1 − α′ b ∈ hz1 i. Taking into account the LIE ALGEBRAS WITH THE DERIVED SUBALGEBRA OF DIMENSION TWO 5 basis of L, we can write a1 = β1 x1 + β2 x2 + β3 y1 + β4 y2 + β5 z1 + β6 z2 + β7 q and α′ b = β1′ x1 + β2′ x2 + β3′ y1 + β4′ y2 + β5′ z1 + β6′ z2 + β7′ q. Therefore x1 = a1 + α′ b + γz1 = (β1 + β1′ )x1 + (β2 + β2′ )x2 + (β3 + β3′ )y1 + (β4 + β4′ )y2 + (β5 + β5′ )z1 + (β6 + β6′ )z2 + (β7 + β7′ )q. Now, since the set {x1 , y1 , x2 , y2 , z1 , z2 , q} is linearly independent and the characteristic of the field is 2, we have β1 = 1 − β1′ , β2 = β2′ , β3 = β3′ , β4 = β4′ , β5 = β5′ + γ, β6 = β6′ and β7 = β7′ . Now we have x1 + hz1 i = a1 + hz1 i + α′ b + hz1 i = (β1 +β1′ )x1 +(β2 +β2′ )x2 +(β3 +β3′ )y1 +(β4 +β4′ )y2 +(β6 +β6′ )z2 +(β7 +β7′ )q +hz1 i = (β1 + β1′ )x1 + hz1 i. Therefore β1 6= 0 or β1′ 6= 0. But x1 + hz1 i ∈ / I4 /hz1 i and hence β1 6= 0 and β1′ = 0 which implies x1 + hz1 i ∈ I3 /hz1 i. Thus x1 ∈ I3 . Now, we want to show [x1 , x2 ] = [x2 , q] = [y2 , q] = [y1 , q] = [x1 , y2 ] = 0. By contrary, let α9 6= 0. Using the same procedure as used in the previous step, we can show x2 ∈ I3 . Because I3 /hz1 i ∼ = H(2), we have [x2 , q] + hz1 i = 0 and so [x2 , q] ∈ hz1 i ∩ hz2 i = 0 which is a contradiction. Similarly, one can show that each of the mentioned brackets vanishes. Here, we want to prove that exactly one of the brackets [x2 , y1 ] or [y1 , y2 ] vanish. Note that the case [x2 , y1 ] = [y1 , y2 ] = 0 leads to L/hz1 i ∼ = H(1) ⊕ A(3) which is a contradiction since L/hz1 i ∼ = H(2) ⊕ A(1). Now, without loss of generality, assume that [y1 , y2 ] 6= 0. Using the same process, we may prove [x2 , y1 ] = 0 and y1 , y2 ∈ I3 and x2 ∈ I4 . Thus L = hx1 , y1 , x2 , y2 , z1 , z2 , q|[y1 , y2 ] = z2 = [x1 , q], [x1 , y1 ] = z1 = [x2 , y2 ]i. Similarly if either L/hz2 i ∼ = H(2) ⊕ A(1) and L/hz1 i ∼ = H(1) ⊕ A(3) or ∼ ∼ L/hz2 i = H(1) ⊕ A(3) and L/hz1 i = H(1) ⊕ A(3), then L ∼ = hx1 , . . . , x7 [x1 , x2 ] = x6 , [x1 , x4 ] = x7 = [x3 , x5 ]i. The result follows.  Corollary 3.2. Let L be a 7-dimensional generalized Heisenberg Lie algebra over any filed F. Then L ∼ = hx1 , . . . , x7 [x1 , x2 ] = x6 = [x3 , x4 ], [x1 , x5 ] = x7 = [x2 , x3 ]i ∼ = L1 or L ∼ = L2 . = hx1 , . . . , x7 [x1 , x2 ] = x6 , [x1 , x4 ] = x7 = [x3 , x5 ]i ∼  Proof. The result follows from Theorems 2.7 (ii) and 3.1. The following result gives the Schur multipliers of Lie algebras L6,22 (ǫ) and It helps to determine the capability of these Lie algebras in the next proposition. (2) L6,7 (η). (2) Proposition 3.3. The Schur multiplier of Lie algebras L6,22 (ǫ) and L6,7 (η) are abelian Lie algebras of dimension 8. Proof. Using the method of Hardy and Stitzinger in [10], we can obtain that in both cases, the dimension of the Schur multipliers are 8.  In the characterizing of capable generalized Heisenberg Lie algebra of rank two of dimension 6 in [18, Theorem 2.12], the Lie algebra L6,22 (ǫ) is missing. Here, we improve this result as bellow. (2) Proposition 3.4. L6,22 (ǫ) and L6,7 (η) are capable. Proof. Let L ∼ = L6,22 (ǫ) is capable. Then by using Theorem 2.7, we have L6,22 (ǫ)2 = Z(L6,22 (ǫ)) = hx5 i ⊕ hx6 i. By Proposition 2.1 (ii) and [19, Corollary 4.6], it is enough to show that dim M(L6,22 (ǫ)/hxi i) − 1 < dim M(L6,22 (ǫ)) for i = 5, 6. Clearly, L6,22 (ǫ)/hxi i ∼ = H(2) or L6,22 (ǫ)/hxi i ∼ = H(1) ⊕ A(2) for i = 5, 6, by using [17, Theorem 3.6]. Thus for i = 5, 6, we have dim M(L6,22 (ǫ)/hxi i) = 5 or 8, by [17, Lemma 2.6 and Theorem 2.7]. Since dim M(L6,22 (ǫ)) = 8, by Proposition 3.3, 6 P. NIROOMAND, F. JOHARI, AND M. PARVIZI we conclude that dim M(L6,22 (ǫ)/hxi i) − 1 < dim M(L6,22 (ǫ)) for i = 5, 6. (2) Therefore L6,22 (ǫ) is capable. By a similar way, we can see that L6,7 (η) is also capable. The proof is completed.  The next result is useful. Lemma 3.5. [18, Lemma 2.11] L5,8 and L1 are capable while L2 is not. We are ready to summarize our results to show that which ones of generalized Heisenberg Lie algebras of rank 2 is capable. Theorem 3.6. Let H be an n-dimensional generalized Heisenberg Lie algebra of rank 2. Then H is capable if and only if H is isomorphic to one of Lie algebras (2) L5,8 , L6,22 (ǫ), L6,7 (η) or L1 . Proof. Let H be capable. Then [18, Proposition 2.6] implies 5 ≤ dim H ≤ 7. Using Corollary 3.2 and Theorem 2.7, H is isomorphic to one of Lie algebras (2) L5,8 , L6,22 (ǫ), L6,7 (η), L1 or L2 . By Proposition 3.4 and Lemma 3.5, all of them are capable while L2 is non-capable. The converse is held by Proposition 3.4 and Lemma 3.5.  4. Stem nilpotent Lie algebras of class 3 with the derived subalgebra of dimension 2 We know that every nilpotent Lie algebra with the derived subalgebra of dimension 2 is of class two and three. In this section, we are going to obtain the structure of stem Lie algebras of class 3 with the derived subalgebra of dimension 2. Then we determine which of them is capable. Moreover, we show that all such Lie algebras of dimension greater than 6 are unicentral. Recall that an n-dimensional nilpotent Lie algebra L is said to be nilpotent of maximal class if the class of L is n − 1. If L is of maximal class, then dim(L/L2 ) = 2, Zi (L) = Ln−i and dim(Lj /Lj+1 ) = 1 for all 0 ≤ i ≤ n − 1 and 2 ≤ j ≤ n − 1 (see [4] for more information). From [8], the only Lie algebra of maximal class of dimension 4 is isomorphic to L4,3 = hx1 , . . . , x4 |[x1 , x2 ] = x3 , [x1 , x3 ] = x4 i. We say a Lie algebra L is a semidirect sum of an ideal I and a subalgebra K if L = I + K, I ∩ K = 0. The semidirect sum of an ideal I and a subalgebra K is denoted by K ⋉ I. Let cl(L) denotes the nilpotency class of a Lie algebra L. The following two lemmas characterize the structure of all stem Lie algebras L of dimensions 5 and 6, when cl(L) = 3 and dim L2 = 2. Lemma 4.1. Let L be a nilpotent stem Lie algebra of dimension 5 such that dim L2 = 2 and cl(L) = 3. Then ∼ L5,5 = hx1 , . . . , x5 |[x1 , x2 ] = x3 , [x1 , x3 ] = x5 , [x2 , x4 ] = x5 i. L= Moreover, L5,5 = I ⋊ hx4 i, where I = hx1 , x2 , x3 , x5 |[x1 , x2 ] = x3 , [x1 , x3 ] = x5 i ∼ = L4,3 , and [I, hx4 i] = hx5 i. LIE ALGEBRAS WITH THE DERIVED SUBALGEBRA OF DIMENSION TWO 7 Proof. By the classification of 5-dimensional nilpotent Lie algebras in [8], we get L∼ = L5,5 . It is easy to check that L5,5 = I⋊hx4 i such that I = hx1 , x2 , x3 , x5 |[x1 , x2 ] = x3 , [x1 , x3 ] = x5 i ∼  = L4,3 and [I, hx4 i] = hx5 i. Lemma 4.2. Let L be a nilpotent stem Lie algebra of dimension 6 such that dim L2 = 2 and cl(L) = 3. Then L ∼ = L6,10 = hx1 , . . . , x6 |[x1 , x2 ] = x3 , [x1 , x3 ] = x6 , [x4 , x5 ] = x6 i. Moreover, L6,10 = I ∔ hx4 , x5 , x6 |[x4 , x5 ] = x6 i = I ∔ K such that I = hx1 , x2 , x3 , x6 |[x1 , x2 ] = x3 , [x1 , x3 ] = x6 i ∼ = L4,3 and K = hx4 , x5 , x6 |[x4 , x5 ] = x6 i ∼ = H(1). Proof. By the classification of 6-dimensional nilpotent Lie algebras in [8], we get L∼ = L6,10 . Clearly Z(L) = hx6 i and L6,10 = I+K, where I = hx1 , x2 , x3 , x6 |[x1 , x2 ] = x3 , [x1 , x3 ] = x6 i ∼ = L4,3 and K = hx4 , x5 , x6 |[x4 , x5 ] = x6 i ∼ = H(1). Since I ∩ K = hx6 i = Z(I) = Z(L) and [I, K] = 0, we can see L6,10 = I ∔ K.  The following proposition is an useful instrument in the next. Proposition 4.3. Let L be an n-dimensional nilpotent stem Lie algebra of class 3 (n ≥ 5) and dim L2 = 2 such that L = I + K, where I and K are two subalgebras of L, I ∼ = L4,3 is the maximal class Lie algebra of dimension 4 and [I, K] ⊆ Z(I) = Z(L). Then (i) If K is a non-trivial abelian Lie algebra such that K ∩ I = 0, then [I, K] = Z(L) and K ∼ = A(1). Moreover, L = I ⋊ K ∼ = L5,5 . (ii) Assume dim K 2 = 1 and I ∩ K = K 2 = Z(L). (a) If K 2 = Z(K), then L = I ∔ K, where n = 2m + 4. Moreover, for m = 1, we have L = I ∔ K ∼ = L6,10 , where K ∼ = H(1). For m ≥ 2, we ∼ have L = I ∔ K = I1 ∔ I2 , where I1 ∼ = H(m − 1). = L6,10 and I2 ∼ (b) If K 2 6= Z(K), then L = (I ⋊A)∔K, where K ∼ = H(m) and A ∼ = A(1), [I, A] = Z(L) = Z(I) = K 2 and n = 2m + 5. Moreover, I ⋊ A ∼ = L5,5 . Proof. (i) We have [I, K] ⊆ Z(I) = Z(L), so I is an ideal of L. Since K ∩I = 0 and I ∼ = L4,3 , dim K = dim L − dim I = n − 4 and so L = I ⋊ K, where K ∼ = A(n − 4). We claim that [I, K] = Z(I). By contrary, assume that [I, K] = 0. Since K is abelian, we have K ⊆ Z(I) = Z(L) ⊆ I. Now I∩K 6= 0, so we have a contradiction. Thus [I, K] = Z(I) = Z(L). We know that I is a Lie algebra of maximal class of dimension 4 so Z(I) = Z(L) = I 3 . We also have dim L2 = dim I 2 = 2. Therefore L2 = I 2 . We claim that K∼ = A(1). First assume that n = 5. Lemma 4.1 implies L ∼ = I ⋊K ∼ = L5,5 and since I ∼ = L4,3 , we have I = hx1 , . . . , x4 |[x1 , x2 ] = x3 , [x1 , x3 ] = x4 i and Ln−4 Z(I) = hx4 i = I 3 . Now let n ≥ 6 and K = i=1 hzi i. In this case we show that K ∩ I 6= 0, which is a contradiction. So this case does not occur. By using Jacobian identity, for all 1 ≤ i ≤ n − 4, we have [zi , x3 ] = [zi , [x1 , x2 ]] = [zi , x1 , x2 ] + [x2 , zi , x1 ] = 0 since [zi , x1 ] and [x2 , zi ] are central. Thus [zi , x3 ] = 0 for all 1 ≤ i ≤ n − 4. Now, let [zi , x1 ] = αi x4 with αi 6= 0. Putting zi′ = zi + αi x3 , we have ′ [zi , x1 ] = [zi + αi x3 , x1 ] = [zi , x1 ] + αi [x3 , x1 ] = αi x4 − αi x4 = 0. So we also obtain [zi′ , x3 ] = 0. Thus [zi′ , x3 ] = [zi′ , x1 ] = 0 for all i, 1 ≤ i ≤ n − 4. Now let [zi′ , x2 ] = αx4 and [zj′ , x2 ] = βx4 , in which α 6= 0 and β 6= 0 for i 6= j and fixed i, j. Put di = βzi′ − αzj′ . We have [di , x2 ] = [βzi′ − αzj′ , x2 ] = 8 P. NIROOMAND, F. JOHARI, AND M. PARVIZI βαx4 −βαx4 = 0 and so [di , x2 ] = 0. On the other hand, [di , x1 ] = [di , x2 ] = [di , x3 ] = 0. Therefore [di , I] = 0 and hence di ∈ Z(L) = Z(I) = hx4 i. Since di = βzi′ − αzj′ = β(zi − αi x3 ) − α(zj − αj x3 ) = βzi − αzj + (ααj − βαi )x3 ∈ Z(I) so 0 6= βzi − αzj ∈ K ∩ I = 0, which is a contradiction. Thus n = 5, K ∼ = A(1) and L = I ⋊ hz1 i and [x2 , z1 ] = x4 , as required. Considering the classification of nilpotent Lie algebras of dimension 5 with dim L2 = 2 given in [8] and Lemma 4.1, we should have L ∼ = L5,5 . (ii) Since I ∩ K = K 2 = Z(L) = Z(I) ∼ = A(1), dim(K) = dim(L) − dim(I) + dim(I ∩ K) = n − 4 + 1 = n − 3. We know that dim K 2 = 1, so [17, Theorem 3.6] implies K ∼ = H(m) and A ∼ = A(n − 2m − 4). If = K1 ⊕ A, in which K1 ∼ ∼ A = 0, then K = H(m). Since I ∩K = K 2 = Z(L) = Z(I) = Z(K) ∼ = A(1), dim(K) = dim(L) − dim(I) + dim(I ∩ K) = n − 4 + 1 = n − 3. Now since 2m + 1 = dim(K) = n − 3, we have n = 2m + 4. We are going to show that [I, K] = 0. In fact, we show that there exists I1 ∼ = L3,4 and K2 ∼ = H(m) with [I1 , K2 ] = 0 and L = I1 ∔ K2 . First let m = 1. We have dim L = 6 and K = hx, y, x4 |[x, y] = x4 i, since K ∼ = H(1). By looking the classification of nilpotent Lie algebras of dimension 6 with dim L2 = 2 given in [5] and Lemma 4.2, we should have L ∼ = L6,10 . Now, let m ≥ 2 and H(m) = ha1 , b1 , . . . , am , bm , z [al , bl ] = z, 1 ≤ l ≤ mi. Lemma 2.5 implies that H(m) = T1 ∔ . . . ∔ Tm , in which Ti ∼ = H(1) for all 1 ≤ i ≤ m. With the same procedure as case in (i) and changing the variables we can see that [Ti , I] = 0 for all i, 1 ≤ i ≤ m. So [I, K] = 0 and hence L = I ∔ K. Since m ≥ 2, we have L = (I ∔ T1 ) ∔ (T2 ∔ . . . ∔ Tm ) such that I ∔ T1 ∼ = L6,10 and T2 ∔ . . . ∔ Tm ∼ = H(m − 1), as required. The case (a) is completed. Now, let A 6= 0 and so n 6= 2m − 4. Thus L = I + (K1 ⊕ A) such that [I, K] ⊆ Z(L) = Z(I). We are going to show that A ∼ = A(1), [I, K1 ] = 0 and [I, A] = Z(I) = Z(L). Similar to the part (ii), we can see that [I, K1 ] = 0. We claim that [I, A] 6= 0. By contrary, let [A, K1 ] = [I, A] = 0 and so A ⊆ Z(L) = Z(I). Since A ∩ I = 0, we have A = 0, which is a contradiction. So we have [I, A] = Z(L). We claim that dim A = 1. Let dim A ≥ 2. Similar to the proof of the part (i), we have [a1 , x1 ] = [a2 , x1 ] = [a1 , x3 ] = [a2 , x3 ] = 0 where a1 , a2 ∈ A and a1 6= a2 . Now let [a1 , x2 ] = αx4 and [a2 , x2 ] = βx4 such that α 6= 0 and β 6= 0. Putting a′1 = βa1 − αa2 . We have [a′1 , x2 ] = [βa1 − αa2 , x2 ] = βαx4 − βαx4 = 0 and so [a′1 , x2 ] = 0. Hence [a′1 , x1 ] = [a′1 , x2 ] = [a′1 , x3 ] = 0. Therefore [a′1 , I] = 0 and hence a′1 ∈ Z(L) = Z(I) = hx4 i = K12 . So a′1 ∈ K1 and since K1 ∩ A = 0, we have a contradiction. Hence A ∼ = A(1) and so n = 2m + 5. Thus L = (I ⋊ A) ∔ K1 such that [I, A] = Z(L) = Z(I). By part (i), we have I ⋊ A ∼ = L5,5 . The case (b) is completed. The result follows.  We need the following lemma for the next investigation. Lemma 4.4. [20, Lemma 1] Let L be a nilpotent Lie algebra and H be a subalgebra of L such that L2 = H 2 + L3 . Then Li = H i for all i ≥ 2. Moreover, H is an ideal of L. LIE ALGEBRAS WITH THE DERIVED SUBALGEBRA OF DIMENSION TWO 9 In the following, we determine the central factor of all stem Lie algebras T such that cl(T ) = 3 and dim T 2 = 2. Lemma 4.5. Let T be an n-dimensional stem Lie algebra such that cl(T ) = 3 and dim T 2 = 2. Then Z(T ) = T 3 ∼ = A(1) and T /Z(T ) ∼ = H(1) ⊕ A(n − 4). Proof. Since T is stem, we have A(1) ∼ = A(1). = T 3 ⊆ Z(T ) $ T 2 . Thus Z(T ) = T 3 ∼ 2 ∼ This follows T /Z(T ) = A(1). Since T /Z(T ) is capable, [17, Theorem 3.6] implies that T /Z(T ) ∼  = H(1) ⊕ A(n − 4). It completes the proof. In the following theorem, we determine the structure of all stem Lie algebras of class 3 with the derived subalgebra of dimension 2. Theorem 4.6. Let T be an n-dimensional stem Lie algebra such that cl(T ) = 3 and dim T 2 = 2. Then one of the following holds (a) T ∼ = L4,3 . (b) T ∼ =I ⋊K ∼ = L5,5 where K ∼ = A(1), I ∼ = L4,3 and Z2 (T ) = Z2 (I) ⋊ K. ∼ ∼ (c) T = I ∔I1 , where I1 = H(m), Z2 (T ) = Z2 (I)+I1 , I ∼ = L4,3 and n = 2m+4. Moreover if m ≥ 2, then L ∼ = H(m − 1), if m = 1, = L6,10 ∔ I2 , where I2 ∼ then L ∼ = L6,10 . (d) T ∼ = H(m), K ∼ = A(1), I ∼ = L4,3 , = L5,5 ∔ I1 , where I1 ∼ = (I ⋊ K) ∔ I1 ∼ Z2 (T ) = (Z2 (I) ⋊ K) ∔ I1 and n = 2m + 5. Moreover, in the cases (b), (c) and (d), Z(T ) = Z(I) = I12 = [I, K]. Proof. Since cl(T ) = 3, we have dim T ≥ 4. If dim T = 4, then T must be a Lie algebra of maximal class and hence T ∼ = L4,3 . Assume that dim T ≥ 5. We have T /Z(T ) ∼ = A(1), by Lemma 4.5. There exist = H(1) ⊕ A(n − 4) and Z(T ) = T 3 ∼ ideals I1 /Z(T ) and I2 /Z(T ) of T /Z(T ) such that I1 /Z(T ) ∼ = H(1) and I2 /Z(T ) ∼ = A(n − 4).  2 2 2 Since T /Z(T ) = I1 +Z(T ) /Z(T ), we have T = I12 +Z(T ) and Z(T ) = T 3 . Using Lemma 4.4, we have T 2 = I12 and so cl(T ) = cl(I1 ) = 3. Hence I1 is a Lie algebra of maximal class and since dim I1 = 4, we have I1 ∼ = L4,3 . Now, Z(T ) = Z(I1 ) because Z(T ) ∩ I1 ⊆ Z(I1 ) and dim Z(T ) = 1. Since Z(T ) ⊆ I1 ∩ I2 ⊆ Z(T ), we have I1 ∩ I2 = Z(T ) = Z(I1 ). Now we are going to determine the structure of I2 . We have I2 /Z(T ) ∼ = A(n − 4) so I22 ⊆ Z(T ) ∼ = A(1), and hence cl(I2 ) ≤ 2 and [I1 , I2 ] ⊆ I1 ∩ I2 = Z(T ) = Z(I1 ) ∼ = A(1). We have dim T /Z(T ) ≥ 4 and so dim I2 ≥ 2. Let cl(I2 ) = 1. Therefore [I1 , I2 ] = I1 ∩ I2 = Z(T ), otherwise [I1 , I2 ] = 0 and since I2 is abelian, I2 ⊆ Z(T ) ∼ = A(1). It is a contradiction, since dim I2 ≥ 2. Hence I2 = Z(T ) ⊕ A, where A ∼ = A(n − 4) and [I1 , I2 ] = Z(T ). Now Z(T ) ⊆ I1 , A ∩ I1 = 0 and I1 ∩ I2 = Z(T ) so T = I1 + I2 = I1 + Z(T ) + A = I1 ⋊ A. Using the proof of Proposition 4.3 (i), we have T ∼ = I1 ⋊K ∼ = L5,5 in which K ∼ = A(1) and [K, I1 ] = Z(T ). This is the case (b). Now, let cl(I2 ) = 2. Since I22 = I1 ∩ I2 = Z(T ) = Z(I1 ) ∼ = A(1), by [17, Theorem 3.6], we have I2 ∼ = H(m) ⊕ A(n − 2m − 4). First assume that A(n − 2m − 4) = 0. Then n = 2m − 4 and I2 ∼ = H(m). Using Proposition 4.3 (ii)(a), we can similarly prove [I1 , I2 ] = 0 and T = I1 ∔ I2 where I2 ∼ = H(m). This is the case (c). Now, let A(n − 2m − 4) 6= 0. Then n 6= 2m − 4 and hence T = I1 + (K ⊕ A) where K ∼ = H(m) and A ∼ = A(n − 2m − 4) and [I1 , K ⊕ A] ⊆ Z(T ) = Z(I1 ). Similar to the case (c), we have [I1 , K] = 0. Now we claim that [I1 , A] = Z(T ) ∼ = A(1). Let [I1 , A] = 0. Since [K, A] = 0, we have A ⊆ Z(T ) = Z(I1 ) = Z(K) = K 2 ∼ = A(1). It is a contradiction, 10 P. NIROOMAND, F. JOHARI, AND M. PARVIZI since A ∩ K = 0. Therefore [I1 , A] = Z(T ) and hence T ∼ = (I1 ⋊ A) ∔ K where A∼ = A(n − 2m − 4) and [I1 , A] = Z(T ) = Z(I1 ). Similar to the case (b), one can obtain that A ∼ = A(1), so n − 2m − 4 = 1, n = 2m + 5 and [I1 , A] = Z(T ). So T = (I1 ⋊ A) ∔ K in which A ∼ = A(1) and [I1 , A] = Z(T ). This is the case (d). Now, we have Z2 (T )/Z(T ) = Z(T /Z(T )) = Z(I1 /Z(T )) ⊕ I2 /Z(T ) and Z(T ) = Z(I1 ) also Z(I1 /Z(T )) = I12 /Z(T ), so Z2 (T )/Z(T ) = I12 /Z(T ) ⊕ I2 /Z(T ). Since I1 is maximal class of dimension 4, we have Z2 (T ) = Z2 (I1 ) + I2 = I12 + I2 . The result follows.  In the following theorem, we classify all non-capable stem Lie algebras of class 3 with the derived subalgebra of dimension 2. Theorem 4.7. Let T be an n-dimensional stem Lie algebra such that cl(T ) = dim T 2 = 2 and n ≥ 6. Then T ∼ = (I ⋊ K) ∔ H or T ∼ = I ∔ H such that H ∼ ∼ H(m), K = A(1), I = L4,3 and [K, I] = Z(T ) = Z(I) = H 2 . Moreover, T non-capable. 3, ∼ = is Proof. By Theorem 4.6 (c) and (d), we obtain (I ⋊ A) ∔ H or I ∔ H such that H ∼ = H(m), A ∼ = A(1), I ∼ = L4,3 and [A, I] = Z(T ) = Z(I) = H 2 . By using Proposition 2.6, T is non-capable. The result follows.  The capable stem Lie algebras of class 3 with the derived subalgebra of dimension 2 are characterized as following. Lemma 4.8. L4,3 and L5,5 are capable. Proof. From [8], let L5,7 = hx1 , . . . , x5 |[x1 , x2 ] = x3 , [x1 , x3 ] = x4 , [x1 , x4 ] = x5 i and L6,13 = hx1 , . . . , x6 |[x1 , x2 ] = x3 , [x1 , x3 ] = x5 , [x2 , x4 ] = x5 , [x1 , x5 ] = x6 , [x3 , x4 ] = x6 i. We have Z(L5,7 ) = hx5 i and Z(L6,13 ) = hx6 i, so L5,7 /hx5 i ∼ = L4,3 and L6,13 /hx6 i ∼  = L5,5 . Thus L4,3 and L5,5 are capable. We are in a position to characterize the capability of an n-dimensional stem Lie algebra T such that cl(T ) = 3 and dim T 2 = 2. Theorem 4.9. Let T be an n-dimensional stem Lie algebra such that cl(T ) = 3 and dim T 2 = 2. Then T is capable if and only if T ∼ = L5,5 . = L4,3 or T ∼ Proof. Let T be capable. By Theorems 4.6, 4.7 and Lemma 4.8, T is isomorphic to L4,3 or L5,5 . The converse holds by Lemma 4.8.  The next theorem gives a necessary and sufficient condition for detecting the capability of stem Lie algebras of class 3 with the derived subalgebra of dimension 2. Theorem 4.10. Let T be an n-dimensional stem Lie algebra such that cl(T ) = 3 and dim T 2 = 2. Then T is capable if and only if 3 ≤ dim(T /Z(T )) ≤ 4. Proof. The result follows from Lemma 4.5 and Theorem 4.9.  Recall that a Lie algebra L is called unicentral if Z ∗ (L) = Z(L). Corollary 4.11. Let T be an n-dimensional stem Lie algebra such that cl(T ) = 3 and dim T 2 = 2. Then T is non-capable if and only if n ≥ 6. Moreover, T is unicentral. Proof. The result follows from Theorems 4.7 and 4.9.  LIE ALGEBRAS WITH THE DERIVED SUBALGEBRA OF DIMENSION TWO 11 5. Nilpotent Lie Algebras with the derived subalgebra of dimension two In this section, we are going to determine all capable nilpotent Lie algebras with the derived subalgebra of dimension 2. At first we show that every finite dimensional nilpotent Lie algebra of class 3 with derived subalgebra of dimension 2 can be considered as a direct sum of a non-abelian stem Lie algebra of class 3 and an abelian Lie algebra. The following result shows that the capability of the direct product of a non-abelian Lie algebra and an abelian Lie algebra depends only on the capability of its nonabelian factor. Theorem 5.1. Let L be a finite dimensional nilpotent Lie algebra of class 3 and dim L2 = 2. Then L = T ⊕ A such that Z(T ) = L2 ∩ Z(L) = L3 = T 3 and Z ∗ (L) = Z ∗ (T ), where A is an abelian Lie algebra. Proof. By using [12, Proposition 3.1], L = T ⊕ A such that Z(T ) = L2 ∩ Z(L) and Z ∗ (L) = Z ∗ (T ), where A is an abelian Lie algebra. Since T is stem, Lemma 4.5 implies Z(T ) = T 3 , as required.  In the following theorem, all capable nilpotent Lie algebras of class 2 with the derived subalgebra of dimension 2 are classified. Theorem 5.2. Let L be an n-dimensional nilpotent Lie algebra of nilpotency class 2 and dim L2 = 2. Then L is capable if and only if L ∼ = = L5,8 ⊕ A(n − 5), L ∼ (2) ∼ ∼ L6,22 (ǫ) ⊕ A(n − 6), L = L6,7 (η) ⊕ A(n − 6) or L = L1 ⊕ A(n − 7). Proof. This is immediately obtained from [18, Propositions 2.4 and 2.6] and Theorem 3.6.  We are ready to determine all capable Lie algebras of class 3 when its derived subalgebra is of dimension 2. Theorem 5.3. Let L be an n-dimensional Lie algebra such that cl(L) = 3 and dim L2 = 2. Then L is capable if and only if L ∼ = L4,3 ⊕ A(n − 4) or L ∼ = L5,5 ⊕ A(n − 5). ∼ T ⊕ A, where A is an abelian Lie algebra and Proof. Theorem 5.1 implies L = ∼ A(1) and Z ∗ (L) = Z ∗ (T ). Now the result follows Z(T ) = T 2 ∩ Z(L) = L3 = T 3 = from Theorem 4.9.  The following result is obtained from Theorems 4.10 and 5.3. Corollary 5.4. Let L be a finite dimensional Lie algebra of class 3 and dim L2 = 2. Then L is capable if and only if 3 ≤ dim(L/Z(L)) ≤ 4. We summarize all results to classify all capable nilpotent Lie algebras with the derived subalgebra of dimension at most two. Theorem 5.5. Let L be an n-dimensional nilpotent Lie algebra with dim L2 ≤ 2. Then L is capable is if and only if L is isomorphic to one the following Lie algebras. (i) If dim L2 = 0, then L ∼ = A(n) and n > 1. (ii) If dim L2 = 1, then L ∼ = H(1) ⊕ A(n − 3). (2) 2 (iii) If dim L = 2 and cl(L) = 2, then L ∼ = L5,8 ⊕ A(n − 5), L = L6,7 (η) ⊕ A(n − ∼ ∼ 6), L = L6,22 (ǫ) ⊕ A(n − 6), or L = L1 ⊕ A(n − 7). 12 P. NIROOMAND, F. JOHARI, AND M. PARVIZI (iv) If dim L2 = 2 and cl(L) = 3, then L ∼ = L4,3 ⊕A(n−4) or L ∼ = L5,5 ⊕A(n−5). Proof. The result follows from [17, Theorems 3.3 and 3.6], Theorems 5.2 and 5.3.  References [1] P. Batten, K. Moneyhun, E. Stitzinger, On characterizing nilpotent Lie algebras by their multipliers. Comm. Algebra 24 (1996) 4319-4330. [2] P. Batten, E. Stitzinger, On covers of Lie algebras, Comm. Algebra 24 (1996) 4301-4317. [3] F. R. Beyl, U. Felgner, and P. Schmid, On groups occurring as center factor groups, J. Algebra 61 (1970) 161-177. [4] L. Bosko, On Schur multiplier of Lie algebras and groups of maximal class, Internat. J. Algebra Comput. 20 (2010) 807-821. [5] S. Cicalò, W. A. de Graaf, C. Schneider, Six-dimensional nilpotent Lie algebras, Linear Algebra Appl. 436 (2012), no. 1, 163-189. [6] G. Ellis, A non-abelian tensor product of Lie algebras, Glasg. Math. J. 39 (1991) 101-120. [7] M. P. Gong, Classification of nilpotent Lie Algebras of dimension 7 (over Algebraically closed fields and R), A thesis in Waterloo, Ontario, Canada, 1998. [8] W. A. de Graaf, Classification of 6-dimensional nilpotent Lie algebras over fields of characteristic not 2, Algebra 309 (2007) 640-653. [9] P. Hall, The classification of prime power groups, J. Reine Angew. Math. 182, (1940) 130-141. [10] P. Hardy, E. Stitzinger, On characterizing nilpotent Lie algebras by their multipliers t(L) = 3; 4; 5; 6; Comm. Algebra, 1998, 26(11), 3527-3539. [11] H. Heineken, Nilpotent groups of class 2 that can appear as central quotient groups, Rend. Sem. Mat. Univ. Padova 84 (1990), 241-248. [12] F. Johari, M. Parvizi, P. Niroomand, Capability and Schur multiplier of a pair of Lie algebras, J. Geometry Phys 114 (2017), 184-196. [13] K. Moneyhun, Isoclinisms in Lie algebras. Algebras Groups Geom. 11 (1994) 9-22. [14] P. Niroomand, F. G. Russo, A note on the Schur multiplier of a nilpotent Lie algebra, Comm. Algebra 39 (2011) 1293-1297. [15] P. Niroomand, F. G. Russo, A restriction on the Schur multiplier of nilpotent Lie algebras, Electron. J. Linear Algebra 22 (2011) 1-9. [16] P. Niroomand, On the dimension of the Schur multiplier of nilpotent Lie algebras, Cent. Eur. J. Math. 9 (2011) 57-64. [17] P. Niroomand, M. Parvizi, F. G. Russo, Some criteria for detecting capable Lie algebras, J. Algebra 384 (2013) 36-44. [18] P. Niroomand, F. Johari, M. Parvizi, On the capability and Schur multiplier of nilpotent Lie algebra of class two, Proc. Amer. Math. Soc. 144 (2016), 4157-4168. [19] A. R. Salemkar, V. Alamian and H. Mohammadzadeh, Some properties of the Schur multiplier and covers of Lie Algebras, Comm. Algebra 36 (2008) 697-707. [20] L. M. Zack, Nilpotent Lie algebras with a small second derived quotient, Comm. Algebra, 36 (2008) 460-4619. School of Mathematics and Computer Science, Damghan University, Damghan, Iran E-mail address: p [email protected], [email protected] Department of Pure Mathematics, Ferdowsi University of Mashhad, Mashhad, Iran E-mail address: [email protected], [email protected] Department of Pure Mathematics, Ferdowsi University of Mashhad, Mashhad, Iran E-mail address: [email protected]
0math.AC
arXiv:1701.07097v1 [math.RT] 24 Jan 2017 THE BRAUER TREES OF UNIPOTENT BLOCKS DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER Abstract. In this paper we complete the determination of the Brauer trees of unipotent blocks (with cyclic defect groups) of finite groups of Lie type. These trees were conjectured by the first author in [19]. As a consequence, the Brauer trees of principal ℓ-blocks of finite groups are known for ℓ > 71. Contents 1. Introduction 2. Notation 3. Modular representations and geometry 3.1. Deligne–Lusztig varieties 3.1.1. Unipotent blocks 3.1.2. Harish-Chandra induction and restriction 3.1.3. Complex of cohomology and Frobenius action 3.1.4. Simple modules in the cohomology of Deligne–Lusztig varieties 3.2. Compactifications 3.3. Steinberg representation 3.4. Coxeter orbits 3.5. Generic theory 3.5.1. Reflection data 3.5.2. d-Harish-Chandra theory 4. Unipotent blocks with cyclic defect groups 4.1. Blocks with cyclic defect groups 4.1.1. Brauer trees and folding 4.1.2. Brauer tree of a block with cyclic defect 4.1.3. Folding 4.2. Structure of unipotent blocks with cyclic defect groups 4.2.1. Centre 4.2.2. Local subgroups and characters 2 5 6 6 6 7 9 10 11 13 14 15 15 16 16 16 17 17 18 18 18 20 Date: January 26, 2017. The first author is supported by a Royal Society Research Fellowship. The second author gratefully acknowledges financial support by the ANR, Project No ANR-16-CE40-0010-01. The third author is partly supported by the NSF (grant DMS-1161999) and by a grant from the Simons Foundation (#376202). 1 2 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER 4.2.3. Genericity 4.2.4. Determination of the trees 4.3. Properties of the trees 4.3.1. Harish-Chandra branches 4.3.2. Real stem 4.3.3. Exceptional vertex 4.3.4. Stable equivalence 4.3.5. Coxeter orbits 4.4. Summary of the algebraic methods 5. Determination of the trees 5.1. Groups of type E7 5.1.1. d = 14 5.1.2. d = 9 5.1.3. d = 10 5.2. Groups of type E8 5.2.1. d = 9 5.2.2. d = 12 5.2.3. d = 18 5.2.4. d = 15 5.2.5. d = 20 5.2.6. d = 24 5.3. Other exceptional groups 5.3.1. E6 (q), 2 E6 (q), F4 (q) and G2 (q) 5.3.2. 2 B2 (q 2 ) and 2 G2 (q 2 ) 5.3.3. 2 F4 (q 2 ) Appendix A. Brauer trees for E7 (q) and E8 (q) References 21 22 24 24 25 26 27 27 31 31 32 32 33 34 34 34 35 35 39 41 45 48 48 48 48 52 58 1. Introduction A basic problem in the modular representation theory of finite groups is to determine decomposition matrices. Ever since the theory of blocks with cyclic defect groups that originated with Brauer [5] and was extended by Dade [21], the prospect of determining all Brauer trees associated to finite groups was a fundamental challenge in modular representation theory. Those trees determine the decomposition matrix as well as the Morita equivalence class of the block. In 1984, Feit [32, Theorem 1.1] proved that, up to unfolding — broadly speaking, taking a graph consisting of several copies of a given Brauer tree and then identifying all exceptional vertices – the collection of Brauer trees of all finite groups coincides with that of the quasisimple groups. THE BRAUER TREES OF UNIPOTENT BLOCKS 3 For alternating groups and their double covers, the Brauer trees are known [61], and for all but the two largest sporadic groups all Brauer trees are known (see [50] for most of the trees). The remaining quasisimple groups, indeed the ‘majority’ of quasisimple groups, are groups of Lie type G(q): if ℓ is a prime dividing |G(q)| then either ℓ ∤ q or ℓ | q — in the latter case, for there to be an ℓ-block with cyclic defect group we must have that G/Z(G) = PSL2 (ℓ) and the Brauer tree is a line. Thus the major outstanding problem is to determine the Brauer trees of ℓ-blocks of groups of Lie type when ℓ ∤ q. Here we make substantial progress on this problem by constructing the trees for the unipotent blocks of G(q). We determine in particular the trees occurring in principal blocks. Our main theorem is the following. Theorem 1.1. Let G be a finite group of Lie type and let ℓ be a prime distinct from the defining characteristic. If B is a unipotent ℓ-block of G with cyclic defect groups then the planar-embedded Brauer tree of B is known. Furthermore, the labelling of the vertices by unipotent characters in terms of Lusztig’s parametrization is known. Theorem 1.1 has the following corollary. Corollary 1.2. Let G be a finite group with cyclic Sylow ℓ-subgroups. If ℓ 6= 29, 41, 47, 59, 71, then the (unparametrized) Brauer tree of the principal ℓ-block of G is known. Conjecturally, every block of a finite group of Lie type is Morita equivalent to a unipotent block. This would extend the previous corollary to all blocks with cyclic defect groups of all finite groups (for ℓ > 71 so that no sporadic groups are involved). Non-unipotent blocks with cyclic defect are currently under investigation by the first author and Radha Kessar. A basic method to determine decomposition matrices of finite groups is to induce projective modules from proper subgroups. In the case of modular representations of finite groups of Lie type in non-defining characteristic, Harish-Chandra induction from standard Levi subgroups has similarly been a very useful tool to produce projective modules. Here, we introduce a new method, based on the construction, via Deligne–Lusztig induction, of bounded complexes of projective modules with few non-zero cohomology groups. This is powerful enough to allow us to determine the decomposition matrices of all unipotent blocks with cyclic defect groups of finite groups of Lie type. In [30], the second and third authors used Deligne–Lusztig varieties associated to Coxeter elements to analyse representations modulo ℓ, where the order d of q modulo ℓ is the Coxeter number. Here, we consider cases where that order is not the Coxeter number, but we use nevertheless the geometry of Coxeter Deligne–Lusztig varieties, as they are the best understood, and have certain remarkable properties not shared by other Deligne–Lusztig varieties. 4 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER Our main result is the proof of the first author’s conjecture [19], in the case of blocks with cyclic defect groups. That conjecture is about the existence of a perverse equivalence with a specific perversity function. Using the algorithm that determines the Brauer tree from the perversity function [17], the first author had proposed conjectural Brauer trees and proved that his conjecture held in many cases. We complete here the proof of that conjecture. The methods we use for determining the Brauer trees are a combination of standard arguments and more recent methods developed in [28, 29, 30]. We start with the subtrees corresponding to various Harish-Chandra series, giving a disjoint union of lines providing a first approximation of the tree. The difficulty lies in connecting those lines with edges labelled by cuspidal modules. Many possibilities can be ruled out by looking at the degrees of the characters and of some of their tensor products. These algebraic methods have proved to be efficient for determining most of the Brauer trees of unipotent blocks (see for instance [48, 49]), but were not sufficient for groups of type E7 and E8 . We overcome this problem by using the mod-ℓ cohomology of Deligne–Lusztig varieties and their smooth compactifications. This is done by analysing well-chosen Frobenius eigenspaces on the cohomology complexes of these varieties and extracting • projective covers of cuspidal modules, giving the missing edges in the tree, • Ext-spaces between simple modules, yielding the planar-embedded tree. This strategy requires some control on the torsion part of the cohomology groups, and for that reason we must focus on small-dimensional Deligne–Lusztig varieties only (often associated with Coxeter elements). We briefly describe the structure of the article. Section 3 is devoted to general results on unipotent blocks of modular representations of finite groups of Lie type, using algebraic and geometrical methods. In Section 4, we deal specifically with unipotent blocks with cyclic defect groups. After recalling in §4.1 the basic theory of Brauer trees, we consider in §4.2 the local structure of the blocks. In §4.3, we establish general properties of the trees, and in particular we relate properties of the complex of cohomology of Coxeter Deligne–Lusztig varieties with properties of the Brauer tree. A key result is Lemma 4.20 about certain perfect complexes for blocks with cyclic defect groups with only two non-zero rational cohomology groups. In §5 we complete the determination of the trees, which are collected in the appendix. The most complicated issues arise from differentiating the cuspidal modules E8 [θ] and E8 [θ2 ] when d = 18 (§5.2.3) and ordering cuspidal edges around the Steinberg vertex for d = 20 (§5.2.5). Acknowledgements: We thank Cédric Bonnafé, Frank Lübeck and Jean Michel for some useful discussions, and Gunter Malle for his comments on a preliminary version of the manuscript. THE BRAUER TREES OF UNIPOTENT BLOCKS 5 2. Notation Let R be a commutative ring. Given two elements a and r of R with r prime, we denote by ar the largest power of r that divides a. If M is an R-module and R′ is a commutative R-algebra, we write R′ M = R′ ⊗R M. Let ℓ be a prime number, O the ring of integers of a finite extension K of Qℓ and k its residue field. We assume that K is large enough so that the representations of finite groups considered are absolutely irreducible over K, and the Frobenius eigenvalues on the cohomology groups over K considered are in K. Given a ring A, we denote by A -mod the category of finitely generated A-modules, by A -proj the category of finitely generated projective A-modules and by Irr(A) the set of isomorphism classes of simple A-modules. When A is a finite-dimensional algebra over a field, we identify K0 (A -mod) with Z Irr(A) and we denote by [M] the class of an A-module M. Given two complexes C and C ′ of A-modules, we denote L by Hom•A (C, C ′) = i,j HomA (C i , C ′j ) the total Hom-complex. Let Λ be either k, O or K and let A be a symmetric Λ-algebra: A is finitely generated and free as a Λ-module and A∗ is isomorphic to A as an (A, A)-bimodule. An A-lattice is an A-module that is free of finite rank as a Λ-module. Given M ∈ A -mod, we denote by PM a projective cover of M. We denote by Ω(M) the kernel of a surjective map PM ։ M and we define inductively Ωi (M) = Ω(Ωi−1 (M)) for i ≥ 1, where Ω0 (M) is a minimal submodule of M such that M/Ω0 (M) is projective. Note that Ωi (M) is well defined up to isomorphism. When M is an A-lattice, we define Ω−i (M) as (Ωi (M ∗ )∗ , using the right A-module structure on M ∗ = HomΛ (M, Λ). We denote by Hob (A) and D b (A) the homotopy and derived categories of bounded complexes of finitely generated A-modules. Given a bounded complex of finitely generated A-modules C, there is a complex C red of A-modules, well-defined up to (non-unique) isomorphism, such that that C is homotopy equivalent to C red and C red has no non-zero direct summand that is homotopy equivalent to 0. Suppose that Λ = k. We denote by A -stab the stable category of A -mod, i.e., the additive quotient by the full subcategory of finitely generated projective Amodules. Note that the canonical functor A -mod → D b (A) induces an equivalence from A -stab to the quotient of D b (A) by the thick subcategory of perfect complexes of A-modules, making A -stab into a triangulated category with translation functor Ω−1 . Suppose that Λ = O. We denote by d : K0 (KA) → K0 (kA) the decomposition map. It is characterized by the property d([KM]) = [kM] for an A-lattice M. Let G be a finite group and A = KG. We identify Irr(A) with the set of Kvalued irreducible characters of G. Given χ ∈ Irr(KG), we denotePby bχ the block 1 idempotent of OG that is not in the kernel of χ. We put eG = |G| g∈G g. 6 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER Let Q be an ℓ-subgroup of G. We denote by BrQ : P OG -mod → kNG (Q) -mod  the Q Brauer functor: BrQ (M) is the image of M in M/ x∈Q,m∈M O(m − x(m)) . We Q denote by brQ : (OG) → kCG (Q) the algebra morphism that is the resriction of the linear map defined by g 7→ δg∈CG (Q) g. 3. Modular representations and geometry 3.1. Deligne–Lusztig varieties. 3.1.1. Unipotent blocks. Let G be a connected reductive algebraic group defined over an algebraic closure of a finite field of characteristic p, together with an endomorphism F , a power of which is a Frobenius endomorphism. In other words, there exists a positive integer δ such that F δ defines a split Fqδ -structure on G for a certain power q δ of p, where q ∈ R>0 . We will assume that δ is minimal for this property. Given an F -stable closed subgroup H of G, we will denote by H the finite group of fixed points HF . The group G is a finite group of Lie type. We are interested in the modular representation theory of G in non-defining characteristic, so that we shall always work under the assumption ℓ 6= p. Let T ⊂ B be a maximal torus contained in a Borel subgroup of G, both of which are assumed to be F -stable. Let W = NG (T)/T be the Weyl group of G and S be the set of simple reflections of W associated to B. We denote by r = rG the F -semisimple rank of (G, F ), i.e., the number of F -orbits on S. Given w ∈ W , the Deligne–Lusztig variety associated to w is  XG (w) = X(w) = gB ∈ G/B | g −1F (g) ∈ BwB . It is a smooth quasi-projective variety endowed with a left action of G by left multiplication. Let Λ be either K or k. Recall that a simple ΛG-module is unipotent if it is a composition factor of Hic (X(w), Λ) for some w ∈ W and i ≥ 0. We denote by Uch(G) ⊂ Irr(KG) the set of unipotent irreducible KG-modules (up to isomorphism). A unipotent block of OG is a block containing at least one unipotent character. Given P a parabolic subgroup of G with unipotent radical U and an F -stable Levi complement L, we have a Deligne–Lusztig variety YG (L ⊂ P) = {gU ∈ G/U | g −1F (g) ∈ U · F (U)}, a variety with a free left action of G and a free right action of L by multiplication. The Deligne–Lusztig induction is defined by X G RL⊂P : Z Irr(KL) → Z Irr(KG), [M] 7→ (−1)i [Hic (YG (L ⊂ P)) ⊗KL M]. i≥0 G We also write RLG = RL⊂P . We denote by ∗ RG L⊂P : Z Irr(KG) → Z Irr(KL) the G adjoint map. We have RL⊂P (Uch(L)) ⊂ ZUch(G) and ∗ RG L⊂P (Uch(G)) ⊂ ZUch(L). THE BRAUER TREES OF UNIPOTENT BLOCKS 7 Let w ∈ W and let h ∈ G such that h−1 F (h)T = w. The maximal torus L = hTh−1 is F -stable. It is contained in the Borel subgroup P = hBh−1 with unipotent radical U. In that case, the map gU 7→ gUh = gh(h−1 Uh) identifies Y(L ⊂ P) with the variety  YG (w) = Y(w) = gV ∈ G/V | g −1 F (g) ∈ VẇV where V = h−1 Uh is the unipotent radical of B and ẇ = h−1 F (h) ∈ NG (T). Furthermore, there is a morphism of varieties Y(w) → X(w), gV 7→ gB corresponding to the quotient by TwF ≃ L. 3.1.2. Harish-Chandra induction and restriction. Given an F -stable subset I of S, we denote by WI the subgroup of W generated by I and by PI and LI the standard parabolic subgroup and standard Levi subgroup respectively of G corresponding to I. In that case, the maps RLGI and ∗ RG LI are induced by the usual Harish-Chandra induction and restriction functors. A simple ΛG-module V is cuspidal if ∗ RG LI (V ) = 0 for all proper subsets I of S. The following result is due to Lusztig when L is a torus [53, Corollary 2.19]. The same proof applies, using Mackey’s formula for the Deligne–Lusztig restriction to a torus. Proposition 3.1. Let L be an F -stable Levi subgroup of G and ψ an irreducible character of L such that (−1)rG +rL RLG (ψ) is an irreducible character of G. If ψ is cuspidal and L is not contained in a proper F -stable parabolic subgroup of G, then (−1)rG +rL RLG (ψ) is cuspidal. Proof. Let T be a maximal torus contained in a proper F -stable parabolic subgroup P of G. The Mackey formula (see [22, 7.1]) provides a decomposition 1 X ∗ xL x ∗ G G RT RL (ψ) = RT (ψ ). |L| x∈G T⊂x L Let x ∈ G with T ⊂ x L. By assumption, x L * P, hence T lies in the proper F -stable parabolic subgroup x L ∩ P of x L. Since ψ is cuspidal, ψ x is a cuspidal x character of x L, hence ∗ RTL (ψ x ) = 0 by [53, Proposition 2.18]. It follows that ∗ G RT ((−1)rG +rL RLG (ψ)) = 0, hence (−1)rG +rL RLG (ψ) is cuspidal by [53, Proposition 2.18].  Let A = OGb be a block of OG. Let P be an F -stable parabolic subgroup of G with unipotent radical U and an F -stable Levi complement L. Let A′ = OLb′ be a block of OL. We say that A is relatively Harish-Chandra A′ -projective if the multiplication map bOGeU b′ ⊗OL eU b′ OGb → OGb is a split surjection as a morphism of (A, A)-bimodules. 8 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER The first part of the following lemma follows from [24, Proposition 1.11], while the second part is immediate. Lemma 3.2. Let P be an F -stable parabolic subgroup of G with unipotent radical U and an F -stable Levi complement L. Let Q be an ℓ-subgroup of L. Then P ∩ CG (Q)◦ is a parabolic subgroup of CG (Q)◦ with unipotent radical V = U ∩ CG (Q) and Levi complement L ∩ CG (Q)◦ . Given b and b′ , block idempotents of OG and OL respectively, we have an isomorphism of (kCG (Q), kCL (Q))-bimodules Br∆Q (bOGeU b′ ) ≃ brQ (b)kCG (Q)eV brQ (b′ ). ◦ Let D be a defect group of A and let H = CG (D). Assume that H = CG (D). Let λ be the character of H that is trivial on Z(H) and such that brD (b) = bλ . Lemma 3.3. Let P be an F -stable parabolic subgroup of G with unipotent radical U and an F -stable Levi complement L. ′ Assume that D ≤ L and let λ′ be a character of CL (D) such that h∗ RH 6 H∩L (λ), λ i = ′ ′ 0 and such that λ comes from a defect 0 character of CL (D)/Z(D). Let A be the block of OL corresponding to kCL (D)bλ′ . The block A is relatively Harish-Chandra A′ -projective. Proof. Let V be the unipotent radical of H ∩ P and let M = H ∩ L, a Levi complement of V in H ∩ P. Note that D ⊂ M = CL (D). By assumption, the multiplication map bλ k(H/Z(D))eV bλ′ ⊗kM/Z(D) eV bλ′ k(H/Z(D))bλ → k(H/Z(D))bλ is surjective. It follows from Nakayama’s Lemma that the multiplication map bλ kHbλ′ ⊗kM kHbλ′ bλ → kHbλ is also surjective. There is a unique block idempotent b′ with defect group D of kL such that brD (b′ ) = bλ′ . By [10, Theorem 3.2], this is a unipotent block. Since brD (eU ) = eV , the commutativity of the diagram kH ⊗kM kH   can / (kG ⊗kL kG)∆D  mult can / Br∆D (kG ⊗kL kG) (kG)∆D c can Br∆D (mult) /  Br∆D (kG) can O ∼ mult 1 kH shows that the multiplication map induces a surjection Br∆D (bkGeU b′ ⊗kL b′ eU kGb) ։ Br∆D (bkG). We deduce from [1, Lemma A.1] that the multiplication map gives a split surjective morphism of (kGb, kGb)-bimodules bkGeU b′ ⊗kL b′ eU kGb ։ bkG.  THE BRAUER TREES OF UNIPOTENT BLOCKS 9 3.1.3. Complex of cohomology and Frobenius action. Following [30, Theorem 1.14], given a variety X defined over Fqδ with the action of a finite group H, there is a e c (X, O) of O(H × hF δ i)-modules with the following properties: bounded complex RΓ e c (X, O) is well defined up to isomorphism in the quotient of the homotopy • RΓ category of complexes of O(H × hF δ i)-modules by the thick subcategory of complexes whose restriction to OH is homotopic to 0; e c (X, O) are direct summands of finite direct sums of • the terms of ResOH RΓ modules of the form O(H/L), where L is the stabilizer in H of a point of X; e c (X, O) in the derived category of O(H × hF δ i) is the usual • the image of RΓ complex RΓc (X, O). Note that in [30] such a complex was constructed over k instead of O, but the same methods lead to a complex over O. Indeed, note first that there is a bounded complex of O(H × hF δ i)-modules C constructed in [62, §2.5.2], whose restriction to OH has terms that are direct summands of possibly infinite direct sums of modules of the form O(H/L), where L is the stabilizer in L of a point of X. Furthermore, that restriction is homotopy equivalent to a bounded complex whose terms are direct summands of finite direct sums of modules of the form O(H/L), where L is the stabilizer in H of a point of X. One can then proceed as in [30] to construct e c (X, O). RΓ Given λ ∈ k × , we denote by L(λ) the inverse image of λ in O. Given an OhF δ imodule M that is finitely generated as an O-module, we denote by M(λ) = {m ∈ M | ∃λ1 , . . . , λN ∈ L(λ) such that (F δ − λ1 ) · · · (F δ − λN )(m) = 0} the ‘generalized λ-eigenspace mod ℓ’ of F δ . e c (X, k)(λ) in D b (kH) will be denoted by RΓc (X, k)(λ) and we will The image of RΓ refer to it as the generalized λ-eigenspace of F δ on the cohomology complex of X. When ℓ ∤ |TwF |, the stabilizers of points of X(w) under the action of G are ℓ′ e c (X(w), O) are projective. groups and the terms of the complex of OG-modules RΓ Lemma 3.4. Let w ∈ W such that TwF has cyclic Sylow ℓ-subgroups. Given ζ ∈ k × , we have RΓc (X(w), k)(q−δ ζ) ≃ RΓc (X(w), k)(ζ)[2] in kG -stab . Proof. Recall (§3.1.1) that there is a variety Y(w) acted on freely by GF on the left and acted on freely by TwF on the right such that Y(w)/TwF ≃ X(w). We have RΓc (Y(w), k) ⊗LkTwF k ≃ RΓc (X(w), k). Since TwF has cyclic Sylow ℓ-subgroups, there is an exact sequence of kTwF -modules 0 → k → kTwF → kTwF → k → 0. This sequence is equivariant for the action of F δ , where the action is trivial on the rightmost term k, while it is given by multiplication by q δ on the leftmost term k. The exact sequence induces an F δ -equivariant distinguished triangle k → k[2] → 10 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER C in D b (kTwF ), where C is perfect. Applying RΓc (Y(w), k)⊗LkTwF , we obtain a distinguished triangle in D b (kGF ), equivariant for the action of F δ RΓc (X(w), k) → RΓc (X(w), k)[2] → C ′ , where the action of F δ on the middle term is twisted by q δ and C ′ is perfect. The lemma follows by taking generalized ζ-eigenspaces  3.1.4. Simple modules in the cohomology of Deligne–Lusztig varieties. Every simple unipotent kG-module occurs in the cohomology of some Deligne–Lusztig variety X(w) [46, Theorem 3.1]. If w is minimal for the Bruhat order, this module only occurs in middle degree. This will be an important property to compute the cohomology of X(w) over O from the cohomology over K. Let us now recall the precise result [3, Propositions 8.10 and 8.12]. We adapt the result to the varieties X(w). Recall that there is a pairing K0 (kG -proj) × K0 (kG -mod) → Z. The Cartan map K0 (kG -proj) → K0 (kG -mod) is injective and we identify K0 (kG -proj) with its image. Let K0 (kG -mod)vp be the subgroup of K0 (kG -mod) of elements f such that Nf ∈ K0 (kG -proj) for some positive integer N. The pairing above can be extended to a pairing K0 (kG -mod)vp × K0 (kG -mod) → Q. Proposition 3.5. Let M be a simple unipotent kG-module and let w ∈ W . The following properties are equivalent: (a) w is minimal such that RHom•kG (RΓc (X(w), k), M) 6= 0; (b) w is minimal such that RHom•kG (M, RΓc (X(w), k)) 6= 0; (c) w is minimal such that h[RΓc (X(w), k)], [M]i = 6 0. ℓ(w) Assume that w is such a minimal element. We have HomkG (M, Hc (X(w), k)) 6= 0. If ℓ ∤ |TwF |, then HomDb (kG) (RΓc (X(w), k), M[−i]) = HomDb (kG) (M, RΓc (X(w), k)[i]) = 0 for i 6= ℓ(w). Proof. We use the variety Y(w) as in the proof of Lemma 3.4. Let TwF be the ℓ′ subgroup of elements of TwF of order prime to ℓ and X 1 bw = wF t |Tℓ′ | wF t∈Tℓ′ be the principal block idempotent of kTwF . All composition factors of bw kTwF are trivial, hence RΓc (Y(w), k)bw is an extension of N = |TwF |ℓ copies of RΓc (X(w), k). As a consequence, [RΓc (Y(w), k)bw ] = N · [RΓc (X(w), k)], hence [RΓc (X(w), k)] ∈ K0 (kG -mod)vp . We deduce that h[RΓc (X(w), k)], [M]i = 6 0 if and only if h[RΓc (Y(w), k)bw ], [M]i = 6 0. It follows also that an integer r is minimal such that HomDb (kG) (M, RΓc (X(w), k)[r]) 6= 0 if and only if it is minimal such that HomDb (kG) (M, RΓc (Y(w), k)bw [r]) 6= 0. It follows that w is minimal such that RHom•kG (M, RΓc (Y(w), k)bw ) 6= 0 if and only THE BRAUER TREES OF UNIPOTENT BLOCKS 11 if (b) holds. Similarly, w is minimal such that RHom•kG (RΓc (Y(w), k)bw , M) 6= 0 if and only if (a) holds. Note also that the statements above with RΓc (Y(w), k)bw are equivalent to the same statements with RΓc (Y(w), k) since M is unipotent. The equivalence between (a), (b) and (c) follows now from [3, Proposition 8.12]. Suppose that w is minimal with the equivalent properties (a), (b) and (c). It follows from [3, Proposition 8.10] that the cohomology of RHom•kG (M, RΓc (Y(w), k)bw ) is concentrated in degree ℓ(w). The last assertions of the lemma follow.  Proposition 3.5 shows that for a minimal w, if ℓ ∤ |TwF |, then the complex of kGe c (X(w), k)red is isomorphic to a bounded complex of projective modules modules RΓ such that a projective cover PM of M appears only in degree ℓ(w) as a direct summand of a term of this complex. ∼ 3.2. Compactifications. Let S be a set together with a bijection S − → S, s 7−→ s. Given s ∈ S, we put BsB = BsB ∪ B. The generalized Deligne–Lusztig variety associated to a sequence (t1 , . . . , td ) of elements of S ∪ S is X(t1 , . . . , td ) =  (g0 B, . . . , gdB) ∈ (G/B) d+1 gi−1 gi+1 ∈ Bti B for i = 0, . . . d − 1 gd−1 F (g0 ) ∈ Btd B If w = s1 · · · sd is a reduced expression of w ∈ W then X(s1 , . . . , sd ) is isomorphic to X(w) and X(s1 , . . . , sd ) is a smooth compactification of X(w). It will be denoted by X(w) (even though it depends on the choice of a reduced expression of w). Remark 3.6. Proposition 3.5 also holds for X(w) replaced by X(w) (and the assertions for X(w) are equivalent to the ones for X(w)), with the assumption ‘ℓ ∤ |TwF |’ replaced by ‘ℓ ∤ |TvF | for all v ≤ w’ for the last statement. This follows from the fact that RHom•kG (RΓc (X(w), k), M) ≃ RHom•kG (RΓc (X(w), k), M) whenever RHom•kG (RΓc (X(v), k), M) = 0 for all v < w. The cohomology of X(w) over K is known [25]. We provide here some partial information in the modular setting. Proposition 3.7. Let w, w ′ ∈ W . If ℓ ∤ |TvF | for all v ≤ w or for all v ≤ w ′ , then H∗c (X(w) ×G X(w ′ ), O) is torsion-free. Proof. Given w, w ′ ∈ W , Lusztig defined in [55] a decomposition of X(w) × X(w ′) as a disjoint union of locally closed subvarieties Za stable under the diagonal action of G. The quotient by G of each of these varieties has the same cohomology as an affine space. More precisely, given a, there exists:  . 12 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER • a finite group T , isomorphic to TvF for some v ≤ w and to Tv F for some v ′ ≤ w ′ (F -conjugate to v), and a quasi-projective variety Z0 acted on by G × T , where T acts freely, together with a G-equivariant isomorphism ∼ Z0 /T − → Za ; • a quasi-projective variety Z1 acted on freely by G and T , such that RΓc (G\Z1 , O)[2 dim Z1 ] ≃ O; ∼ • a (G×T )-equivariant quasi-isomorphism RΓc (Z0 , O)[2 dim Za ] − → RΓc (Z1 , O)[2 dim Z1 ]. From these properties and [3, Lemma 3.2] we deduce that if T is an ℓ′ -group then ′ RΓc (G\Za , O) ≃ RΓc (G\Z0 , O) ⊗OT O ≃ RΓc (G\Z1 , O) ⊗OT O[2 dim Z1 − 2 dim Za ] ≃ O[−2 dim Za ]. As a consequence, the cohomology groups of X(w) ×G X(w ′ ) are the direct sums of the cohomology groups of the varieties G\Za and the proposition follows.  Proposition 3.8. Let I be an F -stable subset of S such that ℓ ∤ |LI |. If M is a simple kG-module such that ∗ RG LI (M) 6= 0, then M is not a composition factor in ∗ the torsion of Hc (X(w), O) for any w ∈ W . Proof. Let V be a simple kLI -module such that HomkLI (V, ∗ RG LI (M)) 6= 0. Let v ∈ WI be minimal such that V ∗ occurs as a composition factor, or equivalently as a direct summand, of H∗c (XLI (v), k). By Remark 3.6, it follows that V ∗ occurs only ℓ(v) in Hc (XLI (v), k). Since OLI -mod is a hereditary category, it follows that there is a projective OLI -module V ′ such that V ′ ⊗O k ≃ V ∗ and V ′ is a direct summand of RΓc (XLI (v), O). Since a projective cover PM ∗ of M ∗ occurs as a direct summand of RLGI (V ∗ ), we deduce that it occurs as a direct summand of RLGI (RΓc (XLI (v), O)) ≃ RΓc (X(v), O). By Proposition 3.7 applied to X(v) ×G X(w), we deduce that the cohomology of ∗ PM ⊗OG RΓc (X(w), O) is torsion-free, and hence M does not appear as a composition  factor of the torsion of H∗c (X(w), O). Remark 3.9. Note two particular cases of the previous proposition: • if G is an ℓ′ -group then so is every subgroup, therefore H∗c (X(w), O) is torsionfree; • if ℓ ∤ |LI | for all F -stable I ( S, then the torsion in H∗c (X(w), O) is cuspidal. Recall that two elements w, w ′ ∈ W are F -conjugate if there exists v ∈ W such that w ′ = v −1 wF (v). Lemma 3.10. Let Λ be one of k, O and K. Let J be a subset of W such that if w ∈ J and w ′ < w, then w ′ ∈ J and such that given w ∈ W and s ∈ S with l(sw) > l(w) and l(wF (s)) > l(w), then sw ∈ J if and only if wF (s) ∈ J. THE BRAUER TREES OF UNIPOTENT BLOCKS 13 Let Z be a thick subcategory of D b (ΛG) such that RΓc (X(v), Λ) ∈ Z for all elements v ∈ J that are of minimal length in their F -conjugacy class. We have RΓc (X(v), Λ) ∈ Z for all v ∈ J and RΓc (X(w), Λ) ∈ Z for all w ∈ J. Proof. Consider s ∈ S and v, v ′ ∈ W with v = sv ′ F (s) and v 6= v ′ . Assume that ℓ(v) = ℓ(v ′ ), and furthermore that ℓ(sv) < ℓ(v). We have v = sv ′′ where ℓ(v) = ℓ(v ′′ ) + 1 and v ′ = v ′′ F (s). The G-varieties X(v) and X(v ′ ) have the same étale site, hence isomorphic complexes of cohomology [22, Theorem 1.6]. If ℓ(sv) > ℓ(v), then ℓ(vF (s)) < ℓ(v) [51, Lemma 7.2] and v = v ′′ F (s) with ℓ(v) = ℓ(v ′′ ) + 1 and v ′ = sv ′′ . We conclude as above. Assume now that ℓ(v) = ℓ(v ′ ) + 2. It follows from [25, Proposition 3.2.10] that there is a distinguished triangle RΓc (X(sv ′ ), Λ)[−2] ⊕ RΓc (X(sv ′ ), Λ)[−1] → RΓc (X(v), Λ) → RΓc (X(v ′ ), Λ)[−2] . So, if RΓc (X(v ′ ), Λ) ∈ Z and RΓc (X(sv ′ ), Λ) ∈ Z, then RΓc (X(v), Λ) ∈ Z. By [43, 45], any element v ∈ W can be reduced to an element of minimal length in its F -conjugacy class by applying one of the transformations v → v ′ above. Note that if v ∈ J, then v ′ ∈ J. The lemma follows from the discussion above.  3.3. Steinberg representation. We denote by U the unipotent radical of the Borel subgroup B. Let ψ be a regular character of U (see [4, §2.1] ), eψ be the corresponding central idempotent in OU and Γψ = IndG U (eψ OU) be the Gelfand-Graev module. It is a self-dual projective OG-module independent of ψ, up to isomorphism. The projection of Γψ onto the sum of unipotent blocks is indecomposable [46, Theorem 3.2]. Its unique simple quotient Stℓ is called the modular Steinberg representation. It is cuspidal if ℓ ∤ |LI | for all F -stable I ( S [42, Theorem 4.2]. The unipotent part of KΓψ is simple: this is the usual Steinberg representation, which we denote by St. Statement (i) of the proposition below is a result of [27]. Proposition 3.11. Let t1 , . . . , td be elements of S ∪ S.  (i) If ti ∈ S for all i, then Hom•OG Γψ , RΓc (X(t1 , . . . , td ), O) ≃ O[−ℓ(w)] in D b (O), and hence Stℓ does not occur as a composition factor of Hic (X(t1 , . . . , td ), k) for i 6= ℓ(w).  (ii) If ti 6∈S for some i, then Hom•OG Γψ , RΓc (X(t1 , . . . , td ), O) is acyclic, and hence Stℓ does not occur as a composition factor of H∗c (X(t1 , . . . , td ), k). (iii) Stℓ does not occur as a composition factor of the torsion part of H∗c (X(t1 , . . . , td ), O). Proof. (i) follows from [27] when t1 · · · td is reduced, and the general case follows by changing G and F as in [25, Proposition 2.3.3]. Assume now that ti ∈ S for all i. Using the decomposition of X(t1 , . . . , td ) into Deligne–Lusztig varieties associated to sequences of elements of S, we deduce from  the first part of the proposition that the cohomology of Hom•kG kΓψ , RΓc (X(t1 , . . . , td ), k) 14 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER is zero outside degrees 0, . . . , d. Since X(t1 , . . . , td ) is a smooth projective variety and Γψ is self-dual, we deduce that the cohomology is also zero outside the degrees d, . . . , 2d and therefore it is concentrated in  degree d. As a consequence, the cohomology of Hom•OG Γψ , RΓc (X(t1 , . . . , td ), O) is free over O and concentrated in degree d. By [25, Proposition 3.3.15], we have Hom•KG  KΓψ , RΓc (X(t1 , . . . , td ), K) = 0, • and hence HomOG OΓψ , RΓc (X(t1 , . . . , td ), O) = 0. (ii) follows now by induction on the number of i such that ti is in S: if one of the ti is in S, say t1 , we use the distinguished triangle RΓc (X(t1 , t2 , . . . , td ), O) −→ RΓc (X(t1 , t2 , . . . , td ), O) −→ RΓc (X(t2 , . . . , td ), O) and use induction. Note that the assumption that one of the ti is in S ensures that we never reach X(1) = G/B. Note finally that (iii) follows from (i) and (ii).  Proposition 3.12. If ℓ ∤ |LI | for all F -stable I ( S, then K ⊗O Ωr O ≃ St. L Proof. Given i ∈ {1, . . . , r}, let Mi = I RLGI ◦ ∗ RLGI (O), where I runs over F -stable subsets of S such that |I/F | = i. By the Solomon–Tits Theorem [20, Theorem 66.33], there is an exact sequence of OG-modules 0 → M → M 0 → · · · → M r → 0, where KM ≃ St. By assumption, M i is projective for i 6= r, while M r = O. We deduce that M ≃ Ωr O.  3.4. Coxeter orbits. Let s1 , . . . , sr be a set of representatives of F -orbits of simple reflections. The product c = s1 · · · sr is a Coxeter element of (W, F ). Throughout e c (X(c), O) is a this section and §4.3.5, we will assume that ℓ ∤ |TcF |, and hence RΓ bounded complex of finitely generated projective OG-modules. If v ∈ W satisfies ℓ(v) < ℓ(c) then v lies in a proper F -stable parabolic subgroup, forcing Hom•kG (RΓc (X(v), k), M) to be zero for every cuspidal module kG-module M. Therefore Proposition 3.5 has the following corollary for Coxeter elements. Corollary 3.13. Let c be a Coxeter element and M be a cuspidal kG-module. If ℓ ∤ |TcF |, then the cohomology of RHom•kG (RΓc (X(c), k), M) and of RHom•kG (M, RΓc (X(c), k)) vanishes outside degree r. e c (X(c), O) Lemma 3.14. Assume that ℓ ∤ |TcF |. Let C be a direct summand of RΓ b in Ho (OG -mod) such that (i) the torsion part of H∗ (C) is cuspidal, and (ii) Hi (KC) = 0 for i 6= r. Then Hr (C) is a projective OG-module and Hi (C) = 0 for i 6= r. THE BRAUER TREES OF UNIPOTENT BLOCKS 15 Proof. Since r = ℓ(c), the complex C can be chosen, up to isomorphism in Hob (OG -mod), to be complex with projective terms in degrees r, . . . , 2r and zero terms outside those degrees. Let i be maximal such that Hi (C) 6= 0 (or equivalently such that Hi (kC) 6= 0). There is a non-zero map kC −→ Hi (kC)[−i] in D b (kG). From Corollary 3.13 and the assumption (i) we deduce that i = r. It follows that the cohomology of C is concentrated in degree r. Since C is a bounded complex of projective modules, it follows that Hr (C) is projective.  Proposition 3.15. Let I ⊂ S be an F -stable subset and let cI be a Coxeter element of WI . (i) If ℓ ∤ |LI |, then H∗c (X(cI ), O) is torsion-free. (ii) If H∗c (X(cI ), O) is torsion-free, then the torsion of H∗c (X(c), O) is killed by ∗ RG LI . Proof. The first statement follows from [29, Corollary 3.3] using H∗c (X(cI ), O) = RLGI (H∗c (XLI (cI ), O)). ∗ ∗ The image by ∗ RG LI of the torsion of Hc (X(c), O) is the torsion of Hc (UI \ X(c), O). By [54, Corollary 2.10], the variety UI \ X(c) is isomorphic to (Gm )r−rI × XLI (cI ). The second statement follows.  3.5. Generic theory. We recall here constructions of [7, 8, 9], the representation theory part being based on Lusztig’s theory. 3.5.1. Reflection data. Let K = Q(q) and V = K ⊗Z Y , where Y is the cocharacter group of T. We denote by ϕ the finite order automorphism of V induced by the action of q −1 F . Q V dj We denote by |(W, ϕ)| = N dim i=1 (x − ζj ) the polynomial order of (W, ϕ). Here, N is the number of reflections of W and we have fixed a decomposition into a direct sum of (Gm × hϕi)-stable lines L1 ⊕ · · · ⊕ Ldim V of the tangent space at 0 of V /W , so that dj is the weight of the action of Gm on Lj and ζj is the eigenvalue of ϕ on Lj . Recall that there is some combinatorial data associated with W (viewed as a reflection group on V ) and ϕ: • a finite set Uch(W, ϕ); • a map Deg : Uch(W, ϕ) → Q[x]. We endow ZUch(W, ϕ) with a symmetric bilinear form making Uch(W, ϕ) an orthonormal basis. In addition, given W ′ a parabolic subgroup of W and w ∈ W such that ad(w)ϕ(W ′) = W,ϕ ′ W ′ , there is a linear map RW ′ ,ad(w)ϕ : ZUch(W , ad(w)ϕ) → ZUch(W, ϕ). W,ϕ We will denote by ∗ RW,ϕ W ′ ,ad(w)ϕ the adjoint map to RW ′ ,ad(w)ϕ . The data associated with W and ϕ depends only on the class of ϕ in GL(V )/W . The corresponding pair G = (W, W ϕ) is called a reflection datum. A pair L = (W ′ , ad(w)ϕ) as above is called a Levi subdatum of (W, ϕ). We put WL = W ′ . 16 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER There is a bijection ∼ Uch(G) − → Uch(G), χ 7→ χq such that Deg(χ)(q) = χq (1). There is a bijection from the set of W -conjugacy classes of Levi subdata of G to the set of G-conjugacy classes of F -stable Levi subgroups of G. Those bijections have the property that given L an F -stable Levi subgroup of G with associated Levi subdatum L = (W ′ , ad(w)ϕ), we have (RLG (χ))q = RLG (χq ) for all χ ∈ Uch(W, ϕ) (assuming q > 2 if (G, F ) has a component of type 2 E 6 , E7 or E8 , in order for the Mackey formula to be known to hold [2]). 3.5.2. d-Harish-Chandra theory. Let Φ be a cyclotomic polynomial over K, i.e., a prime divisor of X n − 1 in K[X] for some n ≥ 1. Let V ′ be a subspace of V and let w ∈ W such that wϕ stabilizes V ′ and the characteristic polynomial of wϕ acting on V ′ is a power of Φ. Let W ′ = CW (V ′ ). Then (W ′ , ad(w)ϕ) is called a Φ-split Levi subdatum of (W, ϕ). An element χ ∈ Uch(W, ϕ) is Φ-cuspidal if ∗ RG L (χ) = 0 for all proper Φ-split Levi subdata L of G (when G is semisimple, this is equivalent to the requirement that Deg(χ)Φ = |G|Φ ). Given a Φ-cuspidal pair (L, λ), we denote by Uch(G, (L, λ)) the set of χ ∈ Uch(G) such that hRLG (λ), χi = 6 0. We denote by WG (L, λ) = NW (WL )/WL the relative Weyl group. The Φ-Harish-Chandra theory states that: • Uch(G) is the disjoint union of the sets Uch(G, (L, λ)), where (L, λ) runs over W -conjugacy classes of Φ-cuspidal pairs; • there is an isometry ∼ G I(L,λ) : Z Irr(WG (L, λ)) − → ZUch(G, (L, λ)); W (L,λ) G M G • those isometries have the property that RM I(L,λ) = I(L,λ) IndWMG (L,λ) for all Levi subdata M of G containing L. The sets Uch(G, (L, λ)) are called the Φ-blocks of G. The defect of the Φ-block Uch(G, (L, λ)) is the integer i ≥ 0 such that the common value of Deg(χ)Φ for χ ∈ Uch(G, (L, λ)) is Φi (see §4.2.3 for the relation with unipotent ℓ-blocks). 4. Unipotent blocks with cyclic defect groups 4.1. Blocks with cyclic defect groups. We recall some basic facts on blocks with cyclic defect groups (cf. [33] and [32] for the folding). THE BRAUER TREES OF UNIPOTENT BLOCKS 17 4.1.1. Brauer trees and folding. Definition 4.1. A Brauer tree is a planar tree T with at least one edge together with a positive integer m (the ‘multiplicity’) and, if m ≥ 2, the data of a vertex vx , the ‘exceptional vertex’. Note that the data of an isomorphism class of planar trees is the same as the data of a tree together with a cyclic ordering of the vertices containing a given vertex. Let d > 1 be a divisor of m. We define a new Brauer tree ∧d T . It has a vertex ṽx , and the oriented graph (∧d T ) \ {ṽx } is the disjoint union (T \ {vx }) × Z/d of d copies of T \ {vx }. Let l1 , . . . , lr be the edges of T containing vx , in the cyclic ordering. The edges of the tree ∧d T containing ṽx are, in the cyclic ordering, (l1 , 0), . . . , (l1 , d − 1), (l2 , 0), . . . , (l2 , d − 1), . . . , (lr , 0), . . . , (lr , d − 1). Finally, for every i ∈ Z/d, we have an embedding of oriented trees of T in ∧d T given on edges by l 7→ (l, i), on non-exceptional vertices by v 7→ (v, i) and finally vx 7→ ṽx . The multiplicity of ∧d T is m/d. When m6=d, the exceptional vertex of ∧d T is ṽx . There is an automorphism σ of ∧d T given by σ(ṽx ) = ṽx and σ(v, i) = (v, i + 1) for v ∈ T \ {vx }. Let X be the group of automorphisms of ∧d T generated by σ. There is an isomorphism of planar trees ∼ κ : (∧d T )/X − → T, ṽx 7→ vx , X · (v, i) 7→ v for v ∈ T \ {vx }. In particular the Brauer tree T ′ = ∧d T together with the automorphism group X determine T . Remark 4.2. Given another planar embedding T ′ of Λd T compatible with the auto∼ morphism σ above and such that κ induces an isomorphism of planar trees T ′ /X − → ∼ T , then there is an isomorphism of planar trees T ′ − → Λd T compatible with σ. 4.1.2. Brauer tree of a block with cyclic defect. Let H be a finite group and A = bOH be a block of OH. Let D be a defect group of A and let bD be a block idempotent of the Brauer correspondent of b in ONH (D). We assume D is cyclic and non-trivial. Let E = NH (D, bD )/CH (D), a cyclic subgroup of Aut(D) of order e dividing ℓ − 1. When e = 1, the block A is Morita equivalent to OD. We will be discussing Brauer trees only when e > 1, an assumption we make for the remainder of §4.1. We define a Brauer tree T associated to A. We put m = (|D|−1)/e. An irreducible character χ of KA is called non-exceptional if d(χ)6=d(χ′ ) for all χ′ ∈ Irr(KA) for χ′ 6=χ. When m > 1, we denote by χx the sum of the exceptional irreducible characters of KA (those that are not non-exceptional). We define the set of vertices of T as the union of the non-exceptional characters together, when m > 1, with an exceptional vertex corresponding to χx . The set of edges is defined to be Irr(kA). An edge φ has vertices χ and χ′ if χ + χ′ is the character of the projective cover of the simple A-module with Brauer character φ. Note that the tree T has e edges. 18 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER The cyclic ordering of the edges containing a given vertex is defined as follows: the edge φ2 comes immediately after the edge φ1 if Ext1A (L1 , L2 )6=0, where Li is the simple A-module with Brauer character φi . Recall that the full subgraph of T with vertices the real-valued non-exceptional irreducible characters and the exceptional vertex if m > 1 is a line (the ‘real stem’ of the tree). There is an embedding of the tree T in C where the intersection of T with the real line is the real stem and taking duals of irreducible charcters corresponds to reflection with respect to the real line. 4.1.3. Folding. Let H ′ be a finite group containing H as a normal subgroup and let b′ be a block idempotent of OH such that bb′ 6= 0. We put A′ = b′ OH and we denote by T ′ the Brauer tree of A′ , with multiplicity m′ . We assume D is a defect group of b′ . Let bD′ be the block idempotent of ONH ′ (D) that is the Brauer correspondent of b′ and let E ′ = NH ′ (D, b′D )/CH ′ (D), an ℓ′ -subgroup of Aut(D). Note that E is a subgroup of E ′ . Let Hb′ be the stabilizer of b in H ′ . We have Hb′ = HNH ′ (D, b′D ) and there is a Morita equivalence between bOHb′ and b′ OH ′ induced by the bimodule bOH ′ b′ . Suppose that E ′ 6= E, i.e., m′ 6= m, since [E ′ : E] = m/m′ . The group X of 1-dimensional characters of E ′ /E ≃ Hb′ /H acts on Irr(KA′ ) and on Irr(kA′ ) and this induces an action on T ′ , the Brauer tree associated to A′ . The result below is a consequence of [32, proof of Lemma 4.3] (the planar embedding part follows from Remark 4.2). ∼ Proposition 4.3. There is an isomorphism of Brauer trees ∧d T − → T ′ such that (χ, i) maps to a lift of χ, for χ a non-exceptional vertex. The previous proposition shows that the data of T ′ and of the action of X on T determine the tree T (up to parametrization). 4.2. Structure of unipotent blocks with cyclic defect groups. We assume in §4.2 that the simple factors of [G, G] are F -stable. Note that every finite reductive group can be realized as GF for such a G. From now on, we assume that ℓ is an odd prime. 4.2.1. Centre. We show here that a Brauer tree of a unipotent block of a finite reductive group (in non-describing characteristic) is isomorphic to one coming from a simple simply connected algebraic group. Lemma 4.4. Assume that G is simple and simply connected. Let A be a unipotent block of kGF whose image in k(GF /(Z(G)F )ℓ ) has cyclic defect. Then, A has cyclic defect and Z(G)Fℓ = 1. Proof. Since ℓ is odd, it divides |Z(G)F | only in the following cases [59, Corollary 24.13]: THE BRAUER TREES OF UNIPOTENT BLOCKS 19 • (G, F ) = SLn (q), n ≥ 2 and ℓ | (n, q − 1); • (G, F ) = SUn (q), n ≥ 3 and ℓ | (n, q + 1); • (G, F ) = E6 (q) and ℓ | (3, q − 1); • (G, F ) = 2 E 6 (q) and ℓ | (3, q + 1). Let H = GF /(Z(G)F )ℓ . Suppose that the image of A in kH has non-trivial defect groups. Assume that (G, F ) = SLn (q), n ≥ 2 and ℓ | (n, q − 1) or (G, F ) = SUn (q), n ≥ 3 and ℓ | (n, q + 1). In those cases, the only unipotent block A is the principal block [14, Theorem 13], so H has cyclic Sylow ℓ-subgroups: this is impossible. Assume that (G, F ) = E6 (q) and ℓ | (3, q−1). Note that A cannot be the principal block, as H does not have cyclic Sylow 3-subgroups. There is a unique non-principal unipotent block b, and its unipotent characters are the ones in the Harish-Chandra series with Levi subgroup L of type D4 [31, “Données cuspidales 7,8,9”, p.352–353]. Those three unipotent characters are trivial on Z(G)F . It is easily seen that there is no equality between their degrees nor is the sum of two degrees equal the third one. As a consequence, they cannot belong to a block of kH with cyclic defect and inertial index at most 2. The same method (replacing q by −q) shows also that b cannot have cyclic defect when (G, F ) = 2 E 6 (q) and ℓ | (3, q + 1).  Let H be a finite simple group of Lie type. Then there is a simple simply connected reductive algebraic group G endowed with an isogeny F such that H ≃ GF /Z(G)F , unless H is the Tits group, (G, F ) = 2 F4 (2) and we have H = [GF /Z(G)F , GF /Z(G)F ], a subgroup of index 2 of GF /Z(G)F . The previous lemma shows that if the image in kH of a unipotent ℓ-block of kGF has cyclic defect groups, then the block of kGF already has cyclic defect groups. By folding (§4.1.3), the Brauer tree of a unipotent block of OGF determines the Brauer tree of the corresponding block of OH. Proposition 4.5. Let A be a unipotent block of OGF with cyclic defect group D. ◦ ◦ ◦ We have CG (x) = CG (D) and CG (x) = CG (x)F for all non trivial elements x ∈ D. Furthermore, one of the two following statements hold: • D is the Sylow ℓ-subgroup of Z ◦ (G)F and there is a finite subgroup H of G containing [G, G]F such that G = D × H; • |Z(G)F |ℓ = 1, D 6= 1 and A is Morita equivalent to a unipotent block of a simple factor of G/Z(G) with cyclic defect groups isomorphic to D. In particular, Z(G)F /Z ◦ (G)F is an ℓ′ -group. Proof. Let H be a simple factor of [G, G]. Consider a simply connected cover Hsc of H. The restriction of unipotent characters in A to H and then to HFsc are sums of unipotent characters, and the blocks that contain them have a defect group that is cyclic modulo Z(Hsc )F . It follows from Lemma 4.4 that ℓ ∤ Z(Hsc )F , and therefore ℓ ∤ Z(Gsc ), where Gsc is a simply connected cover of [G, G]. Note that 20 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER as a consequence, both (Z(G)/Z ◦ (G))F and (Z(G∗ )/Z ◦ (G∗ ))F are ℓ′ -groups, where G∗ is a Langlands dual of G. Let Gad = G/Z(G). By [16, Theorem 17.7], we have A ≃ OZ(G)Fℓ ⊗ A′ , where ′ A is the unipotent block of Gad containing the unipotent characters of A. Also, D ≃ Z(G)Fℓ × D ′ , where D ′ is a defect group of A′ . So, if ℓ divides |Z(G)F |, then ℓ divides |Z ◦ (G)F |, D ′ = 1 and D is the Sylow ℓ-subgroup of Z(G)F . Otherwise, consider a decomposition Gad = G1 ×· · ·×Gr where the Gi are simple and F -stable factors. There is a corresponding decomposition A′ = A1 ⊗ · · · ⊗ Ar where Ai is a unipotent block of Gi . So, there is a unique i such that Ai does not have trivial defect groups, and A is Morita equivalent to Ai . ◦ Let us now prove the first statement of the proposition. We have CG (x)F = CG (x) by [16, Proposition 13.16]. The block idempotent brx (b) gives a (nilpotent) block of OCG (x) with defect group D. By [10, Theorem 3.2], this is a unipotent block. ◦ We deduce from the other part of the proposition that D ⊂ Z(CG (x))F , and hence ◦ ◦ CG (x) = CG (D).  4.2.2. Local subgroups and characters. Let A be a unipotent block of OG with a non-trivial cyclic defect group D. Let (D, bD ) be a maximal b-subpair as in §4.1.2 and let E = NG (D, bD )/CG (D). Recall that we assume that ℓ is odd. ◦ Let Q be the subgroup of order ℓ of D and let L = CG (Q). ◦ Theorem 4.6. • L = CG (D) is a Levi subgroup of G. • D is the Sylow ℓ-subgroup of Z ◦ (L)F and L = D × H for some subgroup H of L containing [L, L]F . P • There is a (unique) unipotent character λ of L such that RLG (λ) = χ∈Uch(KA) εχ χ for some εχ ∈ {±1}. • We have |Uch(KA)| = |E| and Irr(KA) is the disjoint union of Uch(KA) and of {(−1)rG +rL RLG (λ ⊗ ξ)}ξ∈(Irr(KD)−{1})/E . • If |E| = 6 |D| − 1, then Uch(KA) is the set of non-exceptional characters of A. Proof. Let A′ be the block of OL corresponding to A. This is a unipotent block with defect group D. By Proposition 4.5, we have Q ≤ Z ◦ (L) 6= 1, hence L is a Levi subgroup of G, since it is the centralizer of the torus Z ◦ (L). Also, D is the Sylow ℓ-subgroup of Z ◦ (L)F and L = D × H for some subgroup H of L containing [L, L]F . There is a (unique) unipotent irreducible representation λ in Irr(KA′ ) and Irr(KA′ ) = {λ ⊗ ξ}ξ∈Irr(KD) . Let ξ ∈ Irr(KD) −{1}. The character χξ = (−1)rG +rL RLG (λ ⊗ξ) is irreducible and it depends only on IndD⋊E ξ [23, Theorem 13.25]. Furthermore, χξ = χξ′ implies D ′ ξ ∈ E · ξ. Assume that |E| = 6 |D| − 1. There are at least two W -orbits on the set of non-trivial characters of D, so the χξ for ξ ∈ (Irr(KD) − {1})/E are exceptional THE BRAUER TREES OF UNIPOTENT BLOCKS 21 characters. Since A and A′ have the same number of exceptional characters, we have found all exceptional characters of A. Let χ1 = (−1)rG +rL RLG (λ). We have d(χP 1 ) = d(χξ ) for any ξ ∈ Irr(KD) − {1}. There are integers nχ ∈ Z such that χ1 = ψ∈Uch(KA) nψ ψ. The restriction of the decomposition map to ZUch(KA) is injective, since we have removed exceptional characters (if |E| = 6 |D| − 1, otherwise one character) from Irr(KA). It follows that χ1 is the unique linear combination of unipotent characters of A such that d(χ1 ) = d(ξ) for some ξ ∈ Irr(D) − {1}. On the other hand, this unique solution satisfies nψ = ±1 and the number of unipotent characters in A′ is |E|.  ∼ Remark 4.7. Choose a bijection Irr(KE) − → Uch(KA′ ), φ 7→ χφ . Define I : ∼ Z Irr(KD ⋊ E) − → Z Irr(KA′ ) by I(IndD⋊E ξ) = RLG (ξ) if ξ ∈ Irr(KD) − {1} and D I(φ) = εχφ χφ for φ ∈ Irr(KE). The proof of Theorem 4.6 above shows that I ∼ is an isotypy, with local isometries Ix : Z Irr(KD) − → Z Irr(KA′ ), ξ 7→ λ ⊗ ξ for x ∈ D − {1}. 4.2.3. Genericity. We assume in §4.2.3 that F is a Frobenius endomorphism. Let A be a unipotent block of OG with a non-trivial cyclic defect group D and let ◦ L = CG (D). Let d be the order of q modulo ℓ. Note that ℓ divides Φe (q) if and only if e = dℓj for some j ≥ 0. Broué–Michel [11] and Cabanes–Enguehard [15] showed that under a mild additional assumption on ℓ (for quasisimple groups not of type A, ℓ good is enough), unipotent characters in ℓ-blocks with abelian defect groups are Φd -blocks. We show below that this results holds for ℓ-blocks with cyclic defect groups without assumptions on ℓ. Using the knowledge of generic degrees, the unipotent Φd -blocks with defect 1 for simple G can be easily determined, using for example Chevie [60]. Theorem 4.8. With the notations of §4.2.2, we have the following assertions: • L is a Φd -split Levi subgroup of G; • D has order |Φd (q)|ℓ ; • λ = λq for a unipotent Φd -cuspidal character λ of L and there is a bijection ∼ Uch(G, (L, λ)) − → Uch(KA) given by χ 7→ χq ; • the Φd -block Uch(G, (L, λ)) has defect 1; • if ℓ is a bad prime for G or ℓ = 3 and (G, F ) has type 3 D 4 , then we are in one of the cases listed in Table 1. Proof. By Proposition 4.5, we can assume that G is simple and simply connected. When ℓ is good and different from ℓ = 3 for type 3 D 4 , the theorem is [15]. Otherwise, the result follows from [31, Théorème A], by going through the list of d-cuspidal pairs with ℓ-central defect and checking if the defect groups given in [31, §3.2] are cyclic. We list the unipotent blocks with cyclic defect for ℓ bad in Table 1, following [31, §3.2]. Note that in [31, p.358, No 29], ‘E7 [±ξ]’ should be replaced by ‘φ512,11 , φ512,12 ’.  22 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER (G, F ) ℓ E6 (q) 3 2 E 6 (q) 3 E8 (q) 3, 5 E8 (q) 3, 5 d (L, λ) 2 (A5 (q) · (q − 1), φ321 ) 1 (2 A5 (q) · (q + 1), φ321 ) 1 (E7 (q) · (q − 1), E7 [±i]) 2 (E7 (q) · (q + 1), φ512,11 or φ512,12 ) Table 1. Unipotent blocks with cyclic defect for ℓ bad Broué [6] conjectured that there is a parabolic subgroup P with an F -stable Levi complement L such that bRΓc (YG (L ⊂ P), O) induces a derived equivalence between A and the corresponding block of ONG (D, bD ). In [17], it is conjectured that such an equivalence should be perverse. It is further shown there how the Brauer tree of A could then be combinatorially constructed from the perversity function. The perversity function can be encoded in the data a function π : Uch(KA) → Z that describes the (conjecturally) unique i such that V ∈ Uch(KA) occurs in Hic (YG (L ⊂ P), K). In [19], the first author gave a conjectural description γ of the function π, depending on Φd and not on ℓ (this is defined for Φd -blocks with arbitrary defect). Using this function, and the combinatorial procedure to recover a Brauer tree from a perversity function, [19] associates a generic Brauer tree to a Φd -block of defect 1. This is a planar-embedded tree, together with an exceptional vertex (but no multiplicity) and the non-exceptional vertices are parametrized by the unipotent characters in the given Φd -block. The Brauer tree of A is conjectured in [19] to be obtained from the generic Brauer tree, by associating the appropriate multiplicity if it is greater than 1, and turning the exceptional vertex into a non-exceptional one if the multiplicity is 1. Our main theorem is a proof of that conjecture. Theorem 4.9. Let A be a unipotent ℓ-block with cyclic defect of G. Then the unipotent characters of KA form a unipotent Φd -block and the Brauer tree of A is obtained from the generic Brauer tree of that Φd -block. Corollary 4.10. There is a perverse derived equivalence between A and D ⋊ E with perversity function γ. 4.2.4. Determination of the trees. Let us now discuss the known Brauer trees. The Brauer trees for classical groups were determined by Fong and Srinivasan [35, 36]. The Brauer trees for the following exceptional groups are known: Burkhart [13] for 2 B2 , Shamash [63] for G2 , Geck [38] for 3 D4 , Hiss [47] for 2 G2 and 2 F4 , Hiss– Lübeck [48] for F4 and 2 E6 (building on earlier work on F4 by Wings [65]) and Hiss–Lübeck–Malle [49] for E6 . More recently, the second and third authors determined in [30] the Brauer trees of the principal Φh -block of E7 and E8 for h the Coxeter number, using new geometric THE BRAUER TREES OF UNIPOTENT BLOCKS G 2 d ([L, L], λ) 23 ℓ | |LI | E6 12 E7 E8 9 10 14 9 2 ( A2 (q), φ21 ) E6 D6 (A2 (q), φ3 ) E6 (A2 (q), φ21) E6 (A2 (q), φ13 ) E6 12 (3 D4 (q), 3 D4 [1]) E6 , D7 15 18 (2 A2 (q), φ21 ) E7 20 24 Table 2. Blocks with unknown Brauer tree methods which are also at the heart of this paper. Also, the first author determined in [19] the Brauer trees of several unipotent blocks with cyclic defect of E7 and E8 . We determine the remaining unknown trees. They correspond to certain unipotent blocks of 2 E6 (cf. Remark 5.1), E7 (§5.1) and E8 (§5.2). We list in Table 4.2.4 the group G, the order d of q modulo ℓ and the d-cuspidal pair (when the block is not principal) associated to each of these blocks. We also indicate the type of the minimal proper standard F -stable Levi subgroups LI with ℓ | |LI |. Let us note that the Brauer trees of other blocks of exceptional groups were determined up to choices of fields of character values in each block. Using Lusztig’s parametrization of unipotent characters we can remove this ambiguity by choosing appropriate roots of unity in Qℓ with respect to q. Corollary 4.11. Let G be a finite group with cyclic Sylow ℓ-subgroups. If ℓ 6= 29, 41, 47, 59, 71, then the (unparametrized) Brauer tree of the principal ℓ-block of G is known. Proof. Let G be a finite group with a non-trivial cyclic Sylow ℓ-subgroup. Since the principal block of G is isomorphic to that of G/Oℓ′ (G), we can assume that Oℓ′ (G) = 1. If G has a normal Sylow ℓ-subgroup, then the Brauer tree is a star. So, we assume G does not have a normal Sylow ℓ-subgroup. It follows from the classification of finite simple groups [34, §5] that G has a normal simple subgroup H with G/H a Hall ℓ′ -subgroup of Out(H). If H is an alternating group, the Brauer trees are foldings of those of symmetric groups, which are lines as all characters are real. If H is a sporadic group, then 24 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER the Brauer tree of the principal block of H is known under the assumptions of ℓ, [50, 18]. Assume now H is a finite group of Lie type. If ℓ is the defining characteristic, then H = PSL2 (Fℓ ) and the Brauer tree of the principal block is well known. Otherwise, the Brauer tree is known by Theorem 4.9.  4.3. Properties of the trees. We assume here that G is simple and we denote by A a unipotent block with cyclic defect group D of OG. Let E = NG (D, bD )/CG (D), where (D, bD ) is a maximal A-subpair. We assume |E| > 1. We denote by T the Brauer tree of A. Recall (Theorem 4.6) that its |E| unipotent vertices are nonexceptional. We define the non-unipotent vertex of T to be the one corresponding to the sum of the non-unipotent characters in KA. It is exceptional if |E| = 6 |D| − 1. 4.3.1. Harish-Chandra branches. Let I be an F -stable subset of S and X be a cuspidal simple unipotent KLI -module with central ℓ-defect, i.e., such that (dim X)ℓ = [LI : Z(LI )]ℓ . Since the centre acts trivially on simple unipotent modules, the ℓblock bI of LI containing X has central defect group, and X is the unique unipotent simple module in bI . This yields the following three facts. e such that X ≃ (a) There exists a unique (up to isomorphism) OLI -lattice X e The kLI -module k X e is irreducible. K X. e In particular NG (LI , X) = (b) X is the unique unipotent module that lifts k X. e NG (LI , k X). e then KKer(P ։ X) e has only non-unipotent (c) If P is a projective cover of X, G G e have no irreconstituents, therefore RLI (X) and KKer(RLI (P ) ։ RLGI (X)) ducible constituents in common. Under the properties (a) and (b), Geck showed in [37, 2.6.9] that the endomorphism e is reduction-stable, i.e. algebra EndOG (RLGI (X)) e ≃ EndkG (RG (k X)). e kEndOG (RLGI (X)) LI Property (c) was used by Dipper (see [26, 4.10]) to show that the decomposition e embeds in the decomposition matrix of b. matrix of EndOG (RLGI (X)) It follows from [39] that the full subgraph of T whose vertices are in the HarishChandra series defined by (LI , X) is a union of lines. Note that [39] proves a corresponding result for blocks of Hecke algebras at roots of unity, in characteristic 0. The fact that the tree does not change when reducing modulo ℓ follows from the following two facts: • a symmetric algebra over a discrete valuation ring that is an (indecomposable) Brauer tree algebra over the field of fractions and over the residue field has the same Brauer tree over those two fields; e correspond to blocks of the • the blocks of the Hecke algebra EndOG (RLGI (X)) Hecke algebra in characteristic 0 for a suitable specialization at roots of unity. THE BRAUER TREES OF UNIPOTENT BLOCKS 25 Each such line in T is called a Harish-Chandra branch. In particular, the principal series part of T is the full subgraph whose vertices are in the Harish-Chandra series (T, 1). Proposition 4.12. Let N be an edge of T and let V1 and V2 be its vertices. Let I be a minimal F -stable subset of S such that ∗ RG LI (N) 6= 0. If ℓ ∤ |LI |, then given i ∈ {1, 2}, the F -stable subset I is also minimal with respect to the property that ∗ RG LI (Vi ) 6= 0. Proof. Let M be an OLI -lattice such that KM is simple and N is a quotient of RLGI (M). Note that M is projective, hence it follows by Harish-Chandra theory that KM is cuspidal. Since RLGI (M) is projective, it follows that V1 and V2 are direct summands of KRLGI (M). The proposition follows by Harish-Chandra theory.  Corollary 4.13. Suppose that ℓ ∤ |LI | for all F -stable I ( S. Then the edges that are not in a Harish-Chandra branch are cuspidal. The following result is a weak form of [46, Theorem 3.5]. Proposition 4.14. If St is a vertex of T , then the edge corresponding to Stℓ connects St and the non-unipotent vertex. Proof. Recall that bΓψ is the projective cover of Stℓ . Since St is the unique unipotent component of KΓψ , the proposition follows.  Proposition 4.15. Assume that A is the principal block and ℓ ∤ |LI | for any F stable I ( S. Let L be the full subgraph of T whose vertices are at distance at most r from 1. Then L is a line whose leaves are 1 and St. Proof. The tables in [44, Appendix F] show that the Brauer tree of the principal block of the Hecke algebra EndOG (RTG (O)) is a line with r + 1 vertices, with leaves corresponding to the trivial and sign characters. So, T has a full subgraph L that is a line with r + 1 vertices and with leaves 1 and St. Using Proposition 3.12 and duality, we deduce that all vertices at distance at most r from 1 are in L.  4.3.2. Real stem. Let V be a unipotent irreducible KG-module. Let w ∈ W such that V occurs in Hic (X(w), K). The eigenvalues of F δ on the V -isotypic component of Hic (X(w), K) are of the form λV q δj where λV is a root of unity (depending only on V , not on w nor i), for some j ∈ 21 Z. Note that λV ∗ = λ−1 V . So, the vertices of the real stem of T consist of the non-unipotent vertex and the unipotent vertices corresponding to the V such that λV = ±1. For classical groups, all unipotent characters have this property, and are real valued, and for exceptional groups, the unipotent characters with this property are principal-series characters and D4 -series characters, which are real-valued by [41, Proposition 5.6], and cuspidal characters G[±1], which are rational-valued by [41, Table 1]. 26 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER 4.3.3. Exceptional vertex. By Theorem 4.6, a non-unipotent character is obtained by Deligne–Lusztig induction from an irreducible non-unipotent character. We give here a condition for that character to be cuspidal. Proposition 4.16. Assume that λ is cuspidal and L is not contained in any proper F -stable parabolic subgroup of G. Let P′ be a proper F -stable parabolic subgroup of G with unipotent radical U′ and an F -stable Levi complement L′ . The (OG, OL′ )bimodule bOGeU ′ is projective and its restriction to OG is a direct sum of projective indecomposable A-modules corresponding to edges that do not contain the non-unipotent vertex. In particular, the non-unipotent characters of A are cuspidal. Proof. Let Q be the subgroup of order ℓ of D and let g ∈ G such that Qg ≤ L. Let ∆g Q = {(x, g −1 xg)|x ∈ Q}. We have Br∆g Q (bOGeU ′ ) ≃ Br∆Q (bOGegU ′ g−1 ) = bλ kLeV where V = gU ′ g −1 ∩ L (Lemma 3.2). By assumption, λ is cuspidal and P′ ∩ L is a proper F -stable parabolic subgroup of L, hence bλ kLeV = 0, hence Br∆g Q (bOGeU ′ ) = 0. Since the ((OG) ⊗ (OL′ )opp )-module bOG is a direct sum of indecomposable modules with vertices trivial or containing ∆g Q for some g ∈ G, we deduce that that the (OG, OLI )-bimodule bOGeU ′ is projective. Let ξ ∈ Irr(KD) − {1}. Since ResL[L,L]F (λ ⊗ ξ) = ResL[L,L]F (λ), it follows that λ ⊗ ξ is cuspidal. Theorem 4.6 shows that every non-unipotent character of b is of the form (−1)rG +rL (RLG (λ ⊗ ξ)) for some ξ ∈ Irr(KD) − {1}. Proposition 3.1 shows that such a character is cuspidal.  The assumptions of Proposition 4.16 are satisfied in the following cases: • L = T contains a Sylow Φd -torus of G and d is not a reflection degree of a proper parabolic subgroup of W (e.g. G = E7 (q) and d = 14 or G = E8 (q) and d ∈ {15, 20, 24}). In that case the trivial character of L is cuspidal, and no proper F -stable parabolic subgroup of G can contain a Sylow Φd -torus. • G = E8 (q), d = 12 and ([L, L]F , λ) = (3 D4 (q), 3D4 [1]) or d = 18 and ([L, L]F , λ) = (2 A2 (q), φ21 ). Lemma 4.17. Let w ∈ W and let M be a simple A-module corresponding to an edge containing χexc . If w has minimal length such that RHom•kG (RΓc (X(w), k), M) 6= 0, then ℓ | |TwF |. If ℓ ∤ |TvF | for all v ≤ w, then RHom•kG (RΓc (X(w), k), M) = 0. • Proof. Let M be as in the lemma and w be minimal such that RHom PkG (RΓc (X(w), k), M) 6= wF ℓ(w) 0. Assume that ℓ ∤ |T |. We have (−1) [bRΓc (X(w), k)] = η aη [Pη ], where η runs over the edges of T and aη ∈ Z. By Proposition 3.5 we have aµ > 0 where µ is the edge corresponding to M. Since χexc does not occur in [RΓc (X(w), K)], it follows that there is an edge ν containing χexc such that aν < 0. Let N be the simple A-module corresponding to ν. The complex RHom•kG (RΓc (X(w), k), N) has non-zero cohomology in a degree other than −ℓ(w), hence there is v < w such that THE BRAUER TREES OF UNIPOTENT BLOCKS 27 RHom•kG (RΓc (X(v), k), N) 6= 0 by Proposition 3.5, a contradiction. The lemma follows.  4.3.4. Stable equivalence. Assume in §4.3.4 that δ = 1 and L is a maximal torus of G. This is a Φd -torus. Let w ∈ W be a d-regular element. The next result follows from [30, Corollary 2.11 and its proof]. Proposition 4.18. Let m ∈ {0, . . . , d − 1}. The complex RΓc (X(w), k)(qδm ) is isomorphic in kG -stab to Ω2m k. Remark 4.19. If TvF is an ℓ′ -group for all v < w, then Proposition 4.18 holds with X(w) replaced by X(w). 4.3.5. Coxeter orbits. The following lemma holds for general symmetric O-algebras A such that kA is a Brauer tree algebra. Lemma 4.20. Let C be a bounded complex of finitely generated projective A-modules. Assume that T has a subtree of the form Vt St−1 Vt−1 ❴ ❴ ❴ ❴ ❴ ❴ V2 S1 V1 S0 V0 all of whose vertices are non-exceptional, and: (i) KHi (C) = 0 for i 6∈ {0, −t}, Hi (kC) = 0 for i < −t and KH0 (C) ≃ V0 ; (ii) given an edge M of T , given an integer i < t and given a map f ∈ HomDb (A) (C, M[i]), the induced map from the torsion part of H−i (C) to M vanishes; (iii) letting M be an edge of T that contains Vi , and assuming that M is strictly between Si−1 and Si in the cyclic ordering of edges at Vi (for 0 < i ≤ t − 1) or M 6= S0 (for i = 0), then HomDb (A) (C, M[j]) = 0 for j ∈ {i, i + 1} ∩ {0, . . . , t − 1}; (iv) Si is not a composition factor of the torsion part of H−i+1 (C) for 1 ≤ i ≤ t−1. Then C is homotopy equivalent to δt−1 δ 1 → P S0 → 0 0 → P → PSt−1 −−→ PSt−2 → · · · − where P is a projective A-module in degree −t with KP ≃ KH−t (C) ⊕ Vt and HomA (PSi , PSi−1 ) = Oδi . Furthermore, given i ∈ {0, . . . , t − 2}, the composition factors of the torsion part of H−i (C) correspond to the edges strictly between Si and Si+1 in the cyclic ordering of edges at Vi . If V = KH−t (C) is simple and distinct from Vt−1 , then there is an edge St between V and Vt and P ≃ PSt . Furthermore, the composition factors of the torsion part of H−t+1 (C) correspond to the edges strictly between St−1 and St in the cyclic ordering of edges at Vt . Proof. We can assume that C has no non-zero direct summand homotopic to zero. Since H<−t (kC) = 0, it follows that C <−t = 0. Let m be maximal such that C m 6= 0. Suppose that m > 0. By (i), Hm (C) is a non-zero torsion A-module. Let M be a 28 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER simple quotient of Hm (C). Assumption (ii) gives a contradiction. We deduce that m = 0. Let M be a simple quotient of C 0 such that V0 occurs in KPM . By (iii), we have ∼ M ≃ S0 . It follows that there is an isomorphism PS0 ⊕ Q − → C 0 such that the composite map KQ → KC 0 → KH0 (C) vanishes. Let N be the image of PS0 in H0 (C). Suppose that there is a simple quotient M of H0 (C) vanishing on N. Then M is a quotient of the torsion part of H0 (C) and the composite map Q → H0 (C) → M is non-zero. We deduce that this map induces a non-zero map from the torsion part of H0 (C) to M, which contradicts (ii). Given 1 ≤ i ≤ t − 1, fix δi : PSi → PSi−1 such that HomA (PSi , PSi−1 ) = Oδi . We put δ0 = 0 : PS0 → 0. We prove by induction on i ∈ {0, . . . , t − 1} that δi 0 → C −i → C −i+1 → · · · is isomorphic to the complex 0 → PSi − → PSi−1 → · · · → δ 1 P S1 − → PS0 → 0, where PS0 is in degree 0. This holds for i = 0 and we assume now this holds for some i ≤ t − 2. We have dim HomkA (kPSi+1 , kPSi ) = 1 and we denote by N be the image of a non-zero map PSi+1 → PSi . It is contained in k ker δi . Let M be a composition factor of (k ker δi )/N. If i = 0, then the edge corresponding to M contains V0 and M6≃S0 or it contains V1 and is strictly between S0 and S1 in the cyclic ordering of edges at V1 . If i > 0, then the edge corresponding to M contains Vi and is strictly between Si−1 and Si in the cyclic ordering of edges at Vi or it contains Vi+1 and is strictly between Si and Si+1 in the cyclic ordering of edges at Vi+1 . By (iii), PM is not a direct summand of C −i−1 . It follows from ∼ (iv) that there is an isomorphism PSi+1 ⊕ Q − → C −i−1 such that the composition Q → C −i−1 → C −i vanishes. Let M be a simple quotient of Q. Then M occurs as a quotient of H −i−1(C), which is torsion by (i). So (ii) gives a contradiction. We deduce that C −i−1 ≃ PSi+1 and the differential kC −i−1 → kC −i is not zero. This shows that the induction statement holds for i + 1. We deduce that C is isomorphic to δt−1 δ 1 → P S0 → 0 0 → P → PSt−1 −−→ PSt−2 → · · · → PS1 − for some projective A-module P in degree −t. We have [KP ] = (−1)t [KC] + [KPSt−1 ]−[KPSt−2 ]+· · ·+(−1)t [PS0 ] = [KH−t (C)]+[Vt ], hence KP ≃ KH−t (C)⊕Vt . If V = KH−t (C) is simple, then KP ≃ V ⊕Vt , hence P ≃ PSt where St is the edge containing V and Vt . The last statement follows from the fact that the differential kP → kPSt−1 is non-zero.  e c (X(c), O) that have The following theorem deals with direct summands of RΓ exactly two non-zero cohomology groups over K. Extra assumptions on the block are needed here. e c (X(c), O) Theorem 4.21. Assume that ℓ ∤ |TcF |. Let C be a direct summand of bRΓ b ′ in Ho (OG -mod). Suppose that there are r ≥ r and t > 0 such that THE BRAUER TREES OF UNIPOTENT BLOCKS 29 (i) the torsion part in H∗ (C) is cuspidal, ′ ′ (ii) Hi (KC) = 0 for i6∈{r ′ , r ′ + t} and V0 = Hr +t (KC) and V ′ = Hr (KC) are simple KG-modules, and (iii) T has a subgraph with non-exceptional vertices and non-cuspidal edges Vt St−1 Vt−1 ❴ ❴ ❴ ❴ ❴ ❴ V2 S1 S1 V1 S0 V0 S0 V1 V0 is a connected component of such that Vt−1 ❴ ❴ ❴ ❴ ❴ ❴ V2 the subgraph of T obtained by removing the edge St−1 and all cuspidal edges. Then: • there is an edge St between Vt and V ′ and C is homotopy equivalent to C ′ = 0 −→ PSt −→ PSt−1 −→ · · · −→ PS0 −→ 0 with PSt in degree r ′ ; • the complex C ′ is, up to isomorphism, the unique complex such that the differential PSi → PSi−1 generates the O-module Hom(PSi , PSi−1 ) for 1 ≤ i ≤ t; ′ • the composition factors of the torsion part of Hr +t−i (C) correspond to the edges strictly between Si and Si+1 in the cyclic ordering of edges at Vi+1 (for 0 ≤ i ≤ t − 1). In particular, the edges between St−1 and St around Vt are also cuspidal. Proof. We apply Lemma 4.20 to C[r ′ + t]. Assumptions (i), (ii) and (iv) of the lemma follow from the assumptions of the theorem. By Corollary 3.13, we have HomDb (A) (C, M[i]) = 0 for i > r and M cuspidal. If M is simple non-cuspidal and not in {S0 , . . . , St−1 }, then M does not occur as a composition factor of Hi (kC) for i > r ′ . This shows that Assumption (iii) of the lemma holds. The theorem follows.  Assumption (iii) in Theorem 4.21 may look rather difficult to check if only part of the tree is known. However, it will be satisfied for most of the Brauer trees we will consider, thanks to the following proposition. Proposition 4.22. Let V be a simple unipotent KA-module. Assume that • ℓ ∤ |TcF |, • V is leaf of T , i.e. V remains irreducible after ℓ-reduction, • the Harish-Chandra branch of V has at least t edges, and • ℓ ∤ |LI | for all F -stable subsets I ( S. Then assumptions (i) and (iii) in Theorem 4.21 are satisfied. Proof. Assumption (i) is satisfied by Proposition 3.15, while assumption (iii) is satisfied by Corollary 4.13.  30 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER Bringing Theorem 4.21, Corollary 4.22 and Propositions 4.15 and 3.11(iii) together, we have the following special case. Corollary 4.23. Let b be the block idempotent of the principal block of OG. Assume that ℓ ∤ |TcF |. Assume that • KHi (bRΓc (X(c), O)(qr ) ) = 0 for i 6∈ {r, r+2r} and V = KHr (bRΓc (X(c), O)(qr ) ) is simple, and • ℓ ∤ |LI | for all F -stable subsets I ( S. Let T ′ be the full subgraph of T with vertices at distance at most r + 1 of the trivial character. The real stem of T ′ is a line with leaves 1 and the non-unipotent vertex, and the edge Stℓ has vertices St and the non-unipotent vertex. Any non-real vertex of T ′ is connected to St by an edge. Furthermore, V is non-real and the edge connecting V and St comes between the one connecting St to a unipotent vertex and Stℓ in the cyclic ordering of edges at St. V Stℓ St k 1 V By Remark 3.9, the previous results have a counterpart for the compactification. Proposition 4.24. Assume that ℓ ∤ |TvF | for all v ≤ c. Then Lemma 3.14, Theorem 4.21, Proposition 4.22 and Corollary 4.23 hold with X(c) instead of X(c). Remark 4.25. We have |TcF | = (q + 1)(q 6 − q 3 + 1) = Φ2 (q)Φ18 (q) for G simple of type E7 (q), and |TcF | = q 8 + q 7 − q 5 − q 4 − q 3 + q + 1 = Φ30 (q) for G simple of type E8 (q). In particular, when ℓ is good and d ∈ / {2, h}, the condition ℓ ∤ |TcF | will always be satisfied for E7 (q) and E8 (q). Remark 4.26. We can easily read off the cohomology of a complex C as in Theorem 4.21 from the Brauer tree. As a consequence of Theorem 4.9, one can check that the cohomology of C is concentrated in degrees r and r + t (and irreducible in degree r + t), and is torsion-free. Other calculations in §5 give a strong evidence that the cohomology of a variety associated to a Coxeter element is always torsion-free. By [4] this holds for groups of type A. Such a statement does not hold for more 2ℓ(w)−1 general Deligne–Lusztig varieties: if Hc (X(w), K) = 0 and ℓ divides |TwF |, then 2ℓ(w)−1 Hc (X(w), k) = H1 (X(w), k)∗ is non-zero since the connected Galois covering Y(ẇ) ։ X(w) yields non-trivial connected abelian ℓ-coverings. Therefore by the THE BRAUER TREES OF UNIPOTENT BLOCKS 31 2ℓ(w)−1 universal coefficient theorem, Hc (X(w), O) is a torsion module. However, one can ask whether the property ℓ ∤ |TwF | forces the cohomology to be torsion-free (see also Proposition 3.8). 4.4. Summary of the algebraic methods. We summarize here some facts and arguments about Brauer trees that we shall use throughout §5. We consider a unipotent block with a cyclic defect group and non-trivial automizer. We also assume that the block is real (this is the case for all the unipotent blocks we will consider). (Parity) The distance between two unipotent vertices is even if and only if they degrees are congruent modulo ℓ. (Real stem) The collection of unipotent vertices V with λV = ±1, together with the non-unipotent vertex, form a subgraph of the Brauer tree in the shape of a line, called the real stem. Taking duals of characters corresponds to a reflection of the tree in the real stem. (Hecke) The union of the full subgraphs of T obtained by considering unipotent characters in a given Harish-Chandra series is a collection of lines, which is known. (Degree) The dimension of the simple module corresponding to an edge is the alternating sum of the degrees of the vertices in a minimal path from the edge to a leaf. This dimension is a positive integer, and this can be used to show that certain configurations are not possible. Broadly speaking, the effect of this condition is to force the degrees of the unipotent characters, as polynomials in q, to increase towards the non-unipotent node. (Steinberg) The vertices of the edge Stℓ are St and the non-unipotent vertex. If the proper standard Levi subgroups of G are ℓ′ -groups, then the full subgraph of T whose vertices are at distance at most r from 1 is a line whose leaves are 1 and St and the edge Stℓ is cuspidal. Our strategy is to first study the ‘mod-ℓ generalized eigenspaces’ of F on the complex of cohomology of a Coxeter Deligne–Lusztig variety (or its compactification), for those eigenvalues corresponding to unipotent cuspidal KG-modules. This gives information about the location of the corresponding vertex with respect to the real stem. A second step is required if there are cuspidal unipotent KG-modules in the block that do not occur in the cohomology of a Coxeter Deligne–Lusztig variety. In that case, we consider the eigenspaces in the complex of cohomology of a Deligne–Lusztig variety associated to a d-regular element, which is minimal for the property that this module occurs in the cohomology. 5. Determination of the trees We now determine the Brauer trees of the blocks from Table 4.2.4. The edges corresponding to cuspidal simple modules will be drawn as double lines. 32 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER Throughout this section, A denotes a block of OG with cyclic defect and b is the corresponding block idempotent. 5.1. Groups of type E7 . For groups of type E7 , we need to consider the principal Φd -blocks for d = 9, 14 and the Φ10 -block corresponding to the d-cuspidal pair (2 A2 (q).(q 5 + 1), φ21 ). 5.1.1. d = 14. In that case, the proper Levi subgroups of G are ℓ′ -groups. Let us determine the Brauer tree of the principal Φ14 -block of E7 (q). Using (Hecke), (Degree) and (Steinberg) arguments, we obtain the real stem as shown in Figure 11 (see the Appendix). The difficult part is to locate the two complex conjugate e c (X(c), O)(−1) be the generalized ‘−1 cuspidal unipotent characters. Let C = bRΓ (mod ℓ)-eigenspace’ of F . By [54, Table 7.3], we have KC ≃ (E7 [i])[−7] ⊕ K[−14], where E7 [i] is defined as the unipotent cuspidal KG-module that appears with an eigenvalue of F congruent to −1 modulo ℓ in H7c (X(c), K). Corollary 4.23 shows that E7 [i] is connected to St and that it is the first edge coming after the edge S6 in the cyclic ordering of edges containing St. This completes the determination of the tree. Let us describe more explicitely the minimal representative of the complex C. Let k = S0 , S1 , . . . , S6 be the non-cuspidal modules forming the path from the characters 1 to St in the tree (see Figure 1). E7 [i] St φ27,37 S6 φ105,26 S5 φ189,17 S4 φ189,10 S3 φ105,5 S2 φ27,2 S1 1 S0 E7 [−i] Figure 1. Right-hand side of the Brauer tree of the principal Φ14 block of E7 (q) THE BRAUER TREES OF UNIPOTENT BLOCKS 33 e c (X(c), k)red is given as follows: The complex bRΓ (−1) E7 [i] S6 Stℓ E7 [i] S5 S4 k E [−i] S5 −→ S6 S4 −→ S5 S3 −→ · · · −→ S1 −→ 0. 0 −→ 7 −→ Stℓ S6 E7 [−i] S5 S4 k S6 E7 [i] Remark 5.1. This argument applies to many other trees, especially to those associated to the principal Φd -block when d is the largest degree of W distinct from the Coxeter number (in that case the assumptions on ℓ in Proposition 4.22 are satisfied). This shows for example that the Brauer tree of the principal Φ12 -block of 2 E6 (q) given in [48] is valid without any restriction on q. It is also worth mentioning that it gives not only the planar embedding but also the labelling of the vertices with respect to Lusztig’s classification of unipotent characters (in terms of eigenvalues of Frobenius). In the previous example Ext1kG (E7 [i], Stℓ ) 6= 0 whereas Ext1kG (E7 [−i], Stℓ ) = 0. 5.1.2. d = 9. Let θ ∈ O be the unique third root of unity with image q 3 in k. It follows from Lemma 3.3 that A is Harish-Chandra projective relatively to the principal block of E6 (q), hence A has no cuspidal simple modules. The real stem gives most of the Brauer tree of the principal ℓ-block (see Figure 9). It remains to locate the pairs of complex conjugate characters {E6 [θ]ε , E6 [θ2 ]ε } and {E6 [θ]1 , E6 [θ2 ]1 }. To this end we use the homological information contained in the cohomology of the Coxeter variety X(c). Let I be a proper subset of S. If LI is not a group of type E6 , then LI is an ℓ′ -group and the cohomology of X(cI ) is torsion-free by [29, Proposition 3.1]. This remains true when LI has type E6 . Lemma 5.2. If q has order 9 modulo ℓ, the cohomology of the Coxeter variety in a simple group of type E6 is torsion-free. Proof. Denote by X the Coxeter variety of E6 (q). By Proposition 3.15, the torsion of H∗c (X, O) is cuspidal. Let λ ∈ k × and let Cλ = RΓc (X, O)(λ) . Assume that λ 6∈ {1, q 6}. The cohomology of H∗ (KCλ ) is an irreducible module V corresponding to a block idempotent bλ of defect zero. If V is cuspidal, then it occurs in degree 6 in H∗ (KCλ ), hence H∗ (Cλ ) is torsionfree by Lemma 3.14. If V is not cuspidal, then H∗ (bλ Cλ ) has no torsion. On the other hand, H∗ ((1 − bλ )Cλ ) is torsion and cuspidal, hence 0 by Lemma 3.14. Assume now that λ = 1. We have H6 (KC1 ) = St ⊕E6 [θ2 ] and Hi (KC1 ) = 0 for i 6= 6, so H∗ (C1 ) is torsion-free by Lemma 3.14 (so, St + E6 [θ2 ] is a projective character of E6 (q), as was shown in [49]). Assume finally that λ = q 6 . We have H6 (KCλ ) = E6 [θ], H12 (KCλ ) = 1 and Hi (KCλ ) = 0 for i 6∈ {6, 12}. Corollary 4.23 shows that H∗ (Cλ ) is torsion-free.  34 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER From this lemma together with Proposition 3.15, we deduce that the torsion of H∗c (X(c), O) is cuspidal, hence the principal block part of H∗c (X(c), O) is torsion-free. In particular, the complexes Dλ = bRΓc (X(c), O)(λ) for λ ∈ {q 6 , q 7 } have no torsion in their cohomology. We have KDq6 ≃ E6 [θ]ε [−7] ⊕ φ7,1 [−13], KDq7 ≃ E6 [θ]1 [−8] ⊕ 1[−14]. Theorem 4.21 gives the planar-embedded Brauer tree as shown in Figure 9. 5.1.3. d = 10. For the Φ10 -block the situation is similar: there is a unique proper F -stable subset I of S such that LI is not an ℓ′ -group. This Levi subgroup LI has type D6 . Since the Coxeter number of D6 is 10, [29, Theorem] asserts that H∗c (XLI (cI ), O) is torsion-free. Let E7 [i] be the unipotent cuspidal KG-module that appears with eigenvalue congruent to q 6 modulo ℓ in H7c (X(c), K). Theorem 4.21 e c (X(c), k)(q6 ) gives the planar-embedded Brauer tree as shown applied to C = bRΓ in Figure 10. 5.2. Groups of type E8 . The blocks we need to consider are: • the three Φ9 -blocks associated to the d-cuspidal pairs (A2 .(q 6 + q 3 + 1), φ) for φ = φ3 , φ21 and φ13 ; • the Φ12 -block associated to the d-cuspidal pair (3 D4 (q).(q 4 + q 2 + 1), 3 D4 [1]); • the Φ18 -block associated to the d-cuspidal pair (2 A2 (q).(q 6 − q 3 + 1), φ21); • the principal Φd -blocks for d = 15, 20 and 24. In those cases, the proper Levi subgroups of G are ℓ′ -groups. 5.2.1. d = 9. There are three unipotent blocks with non-trivial cyclic defect. Let θ be the third root of unity in O× with image q 3 in k. The real stem is given by Figure 9, where we have given the correspondence with vertices of the E7 tree. For each of the three trees, there are two pairs of complex conjugate characters that need to be located, namely: (1) {E6 [θ]φ1,0 , E6 [θ2 ]φ1,0 } and {E6 [θ]φ′′1,3 , E6 [θ2 ]φ′′1,3 } for the block b1 associated to the d-cuspidal pair (A2 , φ3 ); (2) {E6 [θ]φ2,1 , E6 [θ2 ]φ2,1 } and {E6 [θ]φ2,2 , E6 [θ2 ]φ2,2 } for the block b2 associated to the d-cuspidal pair (A2 , φ21 ); (3) {E6 [θ]φ1,6 , E6 [θ2 ]φ1,6 } and {E6 [θ]φ′1,3 , E6 [θ2 ]φ′1,3 } for the block b3 associated to the d-cuspidal pair (A2 , φ13 ). To this end we again use the cohomology of the Coxeter variety X(c), which we first show to be torsion-free on each block bi . Lemma 3.3 shows that all three unipotent blocks are Harish-Chandra projective relative to the principal block of E6 (q). By Lemma 5.2, the cohomology of the Coxeter variety of E6 is torsion-free. Therefore by Proposition 3.15 the cohomology of X(c), cut by the sum of the bi , is torsion-free. THE BRAUER TREES OF UNIPOTENT BLOCKS 35 We can now use the same argument as for the principal Φ9 -block of E7 : Theorem 4.21 shows that the part of the tree to the right of the non-unipotent node in Figure 9 is correct. We consider the standard Levi subgroup LI of semisimple type E7 and we use the Harish-Chandra induction of the isomorphism E6 [θ]1 ≃ Ω7 O in OLI -mod. It gives E6 [θ]1 ⊕ E6 [θ]φ2,1 ⊕ E6 [θ]φ2,2 ⊕ E6 [θ]φ′1,3 ≃ Ω7 (O ⊕ φ8,1 ⊕ φ35,2 ⊕ φ112,3 ) in OG-stab. By cutting by each bi and using the information above (on the part of the tree to the right of the non-unipotent node) we get E6 [θ]φ2,2 ≃ Ω7 φ35,2 and E6 [θ]φ′1,3 ≃ Ω7 φ112,3 . The same procedure starting with the isomorphism E6 [θ]ε ≃ Ω7 φ7,1 yields E6 [θ]φ′′1,3 ≃ Ω7 φ160,7 . This completes the determination of the three planar-embedded trees. 5.2.2. d = 12. Let θ ∈ O× be the third root of unity with image q 4 in k. The real stem is as given in Figure 12, therefore knowing the tree amounts to locating the cuspidal character E8 [−θ2 ]. e c (X(c), O)(q6 ) . The non-cuspidal simple A-modules are in the prinLet C = bRΓ cipal series, hence they cannot occur in the torsion of H∗c (X(c), O). It follows that Assumption (i) of Theorem 4.21 is satisfied. Assumption (iii) follows from the knowledge of the real stem of the tree. Finally, Assumption (ii) follows from the decomposition bH∗c (X(c), K)(q6 ) ≃ E8 [−θ2 ][−8] ⊕ φ28,8 [−14]. Theorem 4.21 shows that Figure 12 gives the correct planar-embedded Brauer tree. 5.2.3. d = 18. Let θ ∈ O× be the third root of unity congruent to q 6 modulo ℓ. The real stem is as in Figure 14. • Step 1: position of E8 [−θ2 ]. The only proper standard F -stable Levi subgroup LI with ℓ | |LI | has type E7 . It follows from Proposition 4.16 that bRLGI (M) is projective for any M ∈ OLI -mod. Since 18 is the Coxeter number of E7 , [29, §4.3] shows that the cohomology of the perfect complex bRΓc (X(cI ), O) is torsion-free. It follows from Proposition 3.15 that the torsion of bH∗c (X(c), O) is cuspidal. Since bKH∗c (X(c), O)(q7 ) = (E8 [−θ2 ])[−8] ⊕ φ8,1 [−15], Proposition 4.22 and Theorem 4.21 show that there is an edge between φ35,74 and E8 [−θ2 ], and that edge comes between the edges φ35,74 −−− φ300,44 and φ35,74 −−− φ8,91 in the cyclic ordering of edges around φ35,74 . • Step 2: E8 [θ] is connected to the non-unipotent node. From the Brauer tree of the principal Φ18 -block given in [19], we know that Ω12 k lifts to an OG-lattice of character E6 [θ]1 . Now, if E8 [θ] is not connected to the non-unipotent node, then Ω12 φ8,1 lifts to an OG-lattice of character D4,φ′′1,12 or φ8,91 depending on whether E8 [θ] is connected to the D4 -series or the principal series. 36 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER Since the degree of φ8,1 ⊗ E6 [θ]1 is smaller than that of φ8,91 and of D4,φ′′1,12 , we obtain a contradiction. This proves that E8 [θ] and E8 [θ2 ] are connected to the nonunipotent node, and we obtain the planar-embedded Brauer tree up to swapping these two characters (see Figure 14). • Step 3: description of bRΓc (X(c), O)(q) . e c (X(c), O)red . Its cohomology over K is given by Let C = bRΓ (q) ⊕14 ⊕10 ⊕4 KC ≃ (φ⊕7 8,1 ⊕ φ35,2 ⊕ φ300,8 ⊕ φ840,13 )[−2] ⊕ (E8 [−θ])[−8]. Corollary 3.13 (or rather its analogue for compactifications, which holds since ℓ ∤ TvF for all v ≤ c, see also Remark 3.6) shows that the terms of C are projective and do not involve the projective cover of a cuspidal module, except possibly in degree 8. The character of KC shows that only the projective cover of E8 [−θ] can occur, and it occurs once in degree 8. In addition, the torsion of the cohomology of C must be cuspidal by Proposition 3.8 (there are no modules lying in an E7 -series in b). Let i < 2 be minimal such that Hi (kC) 6= 0. Then Hi (kC) is cuspidal and (kC)i contains an injective hull of Hi (kC), a contradiction. So, Hi (kC) = 0 for i < 2. Let P0 , . . . , P7 be the projective indecomposable modules lying in the principal series of A, with [P0 ] = φ8,1 + φ35,2 and Hom(Pi , Pi+1 ) 6= 0 for 0 ≤ i < 7, so that [P7 ] = φ35,74 + φ8,91 . It follows from Lemma 4.20 that (5.1) C ≃ 0 → P → P2 → · · · → P6 → PE8 [−θ] → 0 where P ≃ P0⊕7 ⊕ P1⊕7 ⊕ P2⊕4 is in degree 2. • Step 4: the torsion part of bRΓc (X(w), O)(q) is cuspidal. Let w ∈ W be the unique (up to conjugation) element of minimal length for which E8 [θ] occurs in H∗c (X(w)). Here ℓ(w) = 14. Let us consider R = bRΓc (X(w), O)(q) . Using the trace formula (see [25, Corollaire 3.3.8]), we find that ⊕6 ⊕3 2 KR ≃ (φ⊕4 8,1 ⊕ φ35,2 ⊕ φ300,8 ⊕ φ840,13 )[−2] ⊕ (E8 [θ ])[−14]. By Proposition 3.8, the torsion-part of the cohomology in R is either cuspidal or in an E7 -series. Since there are no modules in E7 -series in A, we deduce that the torsion part is cuspidal. In particular, if j = 4, 5, 6 then Hom•kG (Pj , kR) ≃ 0, and if j = 0, . . . , 3 the cohomology of Hom•kG (Pj , kR) vanishes outside degree 2. Note that Hom•kG (Pj , kR) ≃ Pj ⊗kG kR where Pj is viewed as a right kG-module via the anti-automorphism g 7→ g −1 of G, since Pj is self-dual. • Step 5: E8 [−θ2 ] is not a composition factor of H∗ (kR). Let C ′ be the cone of the canonical map PE8 [−θ] [−8] → kC. By (5.1) it is homotopic to a complex involving only projective modules in the principal series. Tensoring by kR gives a distinguished triangle PE8 [−θ] [−8] ⊗kG kR → kC ⊗kG kR → C ′ ⊗kG kR . THE BRAUER TREES OF UNIPOTENT BLOCKS 37 From the explicit representative of C ′ and step 4 above we know that the cohomology of C ′ ⊗kG kR vanishes outside the degrees 4, 5 and 6. Proposition 3.7 shows that the cohomology of kC ⊗kG kR vanishes outside degree 4. The previous distinguished triangle shows that the cohomology of PE8 [−θ] ⊗kG kR vanishes outside the degrees −4, . . . , −1. Since Hi (kR) = 0 for i < 0 this proves that E8 [−θ2 ] is not a composition factor of H∗ (kR). • Step 6: E8 [−θ] is not a composition factor of Hi (kR) for i 6= 6, 7, 8. The same method as in Step 5 with D = bRΓc (X(c), O)(q7 ) ≃ C ∗ [−16] and D ′ = Cone(kD → PE8 [−θ2 ] [−8]) yields a distinguished triangle kD ⊗kG kR → PE8 [−θ2 ] [−8] ⊗kG kR → D ′ ⊗kG kR . Here kD ⊗kG R has non-zero cohomology only in degree 16 by Proposition 3.7. As for D ′ ⊗kG kR, its cohomology vanishes outside the degrees 14, 15 and 16, and we deduce from the distinguished triangle that E8 [−θ] is not a composition factor of Hi (kR) for i 6= 6, 7, 8. If v < w and ℓ | |TvF |, then v is conjugate to cI , the Coxeter element of type E7 . Since bRΓc (X(cI ), O) is perfect, it follows from Lemma 3.10 that bRΓc (X(v), O) is perfect for all v < w. • Step 7: given v < w, the complex bRΓc (X(v), k) is quasi-isomorphic to a bounded complex of projective modules whose indecomposable summands correspond to edges that do not contain the non-unipotent vertex. Consider v < w. If v is not conjugate to a Coxeter element cI of E7 , then ℓ ∤ |TvF | and bRΓc (X(v), k) is perfect and quasi-isomorphic to a bounded complex of projective modules whose indecomposable summands correspond to edges that do not contain the non-unipotent vertex by Lemma 4.17. If v = cI , the perfectness has been shown in Step 1 and the second part holds, because the edges that contain the non-unipotent vertex are cuspidal. We deduce that the statement of Step 7 holds for all v < w by Lemma 3.10. • Step 8: H>14 (R) = 0. Steps 4, 5 and 6 show that the composition factors of Hi (kR) for i > 14 are cuspidal modules M corresponding to an edge containing the non-unipotent vertex. Let M be a simple module corresponding to an edge containing the non-unipotent vertex. Step 7 shows that the canonical map RΓ(X(w), k) → RΓ(X(w), k) induces an isomorphism RHom•kG (RΓ(X(w), k), M) ≃ RHom•kG (RΓ(X(w), k), M). Let M0 = Hi0 (kR) be the non-zero cohomology group of kR of largest degree. We have HomDb (kG) (RΓ(X(w), k), M0 [−i0 ]) 6= 0. Since RΓ(X(w), k) has a representative with terms in degrees 0, . . . , ℓ(w) = 14, we deduce from the previous isomorphism that i0 ≤ 14. 38 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER • Step 9: E8 [θ] and E8 [θ2 ] do not occur as composition factors of the torsion part of H∗ (R) and E8 [θ2 ] is a direct summand of H14 (kR). Step 7 shows that E8 [θ] and E8 [θ2 ] are not composition factors of H∗c (X(v), k) for v < w. It follows that if M is any of the simple modules E8 [θ] or E8 [θ2 ], then the canonical map H∗c (X(w), k) → H∗c (X(w), k) induces an isomorphism (5.2)   ∼ HomkG PM , Hic (X(w), k) → HomkG PM , Hic (X(w), k) . Since Hic (X(w), k) = 0 for i < 14, we deduce that E8 [θ] and E8 [θ2 ] do not occur as composition factors of Hic (X(w), k) for i < 14. By Poincarà c duality and the isomorphism (5.2), it follows that E8 [θ] and E8 [θ2 ] cannot occur as composition factors of Hic (X(w), k) for i > 14. On the other hand, E8 [θ] does not occur in [KRΓc (X(w), O)(q) ] (nor does χexc ), hence E8 [θ] does not occur as a composition 14 2 factor of H14 c (X(w), k)(q) or Hc (X(w), k)(q) . Similarly, E8 [θ ] occurs with multiplic14 ity 1 as a composition factor of H14 c (X(w), k)(q) and of Hc (X(w), k)(q) . Proposition 3.5 shows that E8 [θ2 ] is actually a submodule of H14 c (X(w), k)(q) and hence of 14 14 Hc (X(w), k)(q) by (5.2). Since KHc (X(w), O)(q) = E8 [θ2 ], it follows that E8 [θ2 ] is a quotient of H14 c (X(w), k)(q) , hence it is a direct summand. • Step 10: (E8 [θ2 ])[−14] is a direct summand of kR in D b (kG). Let Z be the cone of the canonical map RΓc (X(w), k)(q) → RΓc (X(w), k)(q) . Step 7 shows that Z can be chosen (up to isomorphism in D b (kG)) to be a bounded complex of projective modules that do not involve edges containing the non-unipotent vertex. The complex kR is quasi-isomorphic to the cone D ′ of a map Z[−1] → bRΓc (X(w), k)(q) , hence to the truncation τ ≤14 (D ′ ), a complex N with N i = 0 for i < 0 and i > 14 and with N i a direct sum of projective modules corresponding to edges that do not contain the non-unipotent vertex for i ≤ 13. Note that E8 [θ] and E8 [θ2 ] are not composition factors of N 13 , hence E8 [θ] is not a composition factor of N 14 while E8 [θ2 ] is a composition factor of N 14 with multiplicity 1 (see Step 9 above). Consider a non-zero morphism PE8 [θ2 ] → N 14 and let U be its image. Since E8 [θ2 ] is a direct summand of H14 (N), it follows that the image of U in H14 (N) is E8 [θ2 ]. On the other hand, the simple modules corresponding to edges containing the non-unipotent vertex but not E8 [θ] nor E8 [θ2 ] are not quotients of N 13 , hence U ≃ E8 [θ2 ] embeds in H14 (N). It follows that U[−14] is a direct summand of N. • Step 11: RΓc (X(w), O)(q−2 ) ≃ V [−14], where V is an OG-lattice such that the simple factors of KV are in the D4 -series. Let R′ = RΓc (X(w), O)(q−2) . We have KHi (R′ ) = 0 for i 6= 14 and the simple factors of KH14 (R′ ) are in the D4 -series. As in Step 4, one shows that the torsion of RΓc (X(w), O)(q−2 ) is cuspidal. We show as in Steps 5 and 6 that E8 [−θ] and E8 [−θ2 ] are not composition factors of H∗ (kR′ ). Furthermore, H>14 (R′ ) = 0 as in Step 8. Proceeding as in Step 8, one sees that the canonical map RΓc (X(w), k) → THE BRAUER TREES OF UNIPOTENT BLOCKS 39 RΓc (X(w), k) induces an isomorphism RHom•kG (M, RΓc (X(w), k)) ≃ RHom•kG (M, RΓc (X(w), k)). Let i0 be minimal such that Hi0 (kR′ ) 6= 0, and suppose that i0 < 14. Then Hi0 (kR′ ) is cuspidal and HomDb (kG) (Hi0 (kR′ ), RΓc (X(w), k)[i0 ]) ≃ RHomDb (kG) (Hi0 (kR′ ), RΓc (X(w), k)[i0 ]) 6= 0. This contradicts the fact that RΓc (X(w), k) has no cohomology in degrees less than 14. Thus, Hi (kR′ ) = 0 for i 6= 14. • Step 12: conclusion. Lemma 3.4 shows that RΓc (X(w), k)(q−2 ) ≃ RΓc (X(w), k)(q) [6] in kG -stab. By Step 11, we deduce that kV has a direct summand isomorphic to Ω−6 (E8 [θ2 ]) in kG -stab. If the Brauer tree is not the one given in Figure 14 (i.e., E8 [θ] and E8 [θ2 ] need to be swapped), then Ω−6 (E8 [θ2 ]) is the reduction of a lattice in φ300,44 , which cannot be a direct summand of kV . We deduce that the planar-embedded tree is as shown in Figure 14. Remark 5.3. We use the determination of the tree to obtain a character-theoretic statement that will be needed in the study of the case d = 15. The Brauer tree of the principal Φ18 -block of G is given in [30, Remark 3.11]. In particular, E6 [θ2 ]1 ≃ Ω24 k. Since Ω24 φ8,1 ≃ E8 [θ2 ], we deduce that φ8,1 ⊗ E6 [θ2 ]1 is isomorphic to E8 [θ2 ] plus a projective OG-module P . If E8 [θ] occurs in the character of P , then the non-unipotent vertex occurs as well. As the degree of the non-unipotent vertex is larger than the degree of φ8,1 ⊗ E6 [θ2 ]1 , we obtain a contradiction. So, the character of E8 [θ] is not a constituent of φ8,1 ⊗ E6 [θ2 ]1 . 5.2.4. d = 15. Choose the roots of unity θ of order 3 and ζ of order 5 so that θ = q 5 and ζ = q 3 in k. The real stem is known and comprises the principal series characters in the principal ℓ-block. A (Hecke) argument also gives the two subtrees consisting of characters in the E6 [θ]-series and the E6 [θ2 ]-series as shown in Figure 13. Except for the two characters E8 [θ] and E8 [θ2 ], each Harish-Chandra series lying in the principal Φ15 -block has a character which appears in the cohomology of the Coxeter variety. The generalized (λ)-eigenspaces on the cohomology of the Coxeter variety are given by bH∗c (X(c), K)(q8 ) ≃ E6 [θ]ε [−8] ⊕ K[−16], bH∗c (X(c), K)(q10 ) ≃ (E8 [ζ 2 ])[−8] ⊕ E6 [θ]1 [−10], bH∗c (X(c), K)(q7 ) ≃ (E8 [ζ])[−8] ⊕ φ8,1 [−15]. Corollary 4.23 applied to λ = q 8 shows that there is an edge between St and E6 [θ]ε , and this edge comes between the one containing φ84,64 and Stℓ in the cyclic ordering of edges around St. Only cuspidal characters remain to be located, and since they have a larger degree than E6 [θ]1 , we deduce that E6 [θ]1 remains irreducible modulo ℓ. 40 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER From Proposition 4.22 and Theorem 4.21 applied to C = bRΓc (X(c), O)(q10 ) , we deduce that there is an edge between E8 [ζ 2 ] and E6 [θ]ε . Similarly, using C = bRΓc (X(c), O)(q7 ) , we deduce that there is an edge between φ112,63 and E8 [ζ], and this edge comes between the one containing φ1400,37 and the one containing φ8,91 in the cyclic ordering of edges around φ112,63 . Consequently, the trees in Figures 2 and 3 are subtrees of the Brauer tree T (although as of yet we cannot fix the planar embedding around E6 [θ]ε ). We claim that E8 [θ] and E8 [θ2 ] are not connected to the subtree shown in Figure 2. Let us assume otherwise. By a (Parity) argument, they are not connected to the non-unipotent node. Let w ∈ W be an element of minimal length such that E8 [θ2 ] appears in the cohomology of X(w) (we have ℓ(w) = 14). We have [bH∗c (X(w), K)q2 ] = [E8 [θ2 ]] + [φ8,91 ] = ([φ8,91 ] + [χexc ]) − ([St] + [χexc ]) + η, where η = [KP ] and P is a projective bOG-module whose character does not involve χexc . Therefore there is an odd integer i such that HomDb (kG) (RΓc (X(w), k), Stℓ [i]) 6= 6 0 for some 0. Since i 6= 14, it follows from Proposition 3.5 that h[H∗c (X(v), k)], Stℓ i = v < w. One easily checks on the character of H∗c (X(v), K) that this is impossible. As a consequence, E8 [θ] and E8 [θ2 ] are connected to the subtree shown in Figure 3. We next return to the planar embedding of the edges around the node E6 [θ]ε . If the embedding is not as in Figure 2 then Ω−13 k would lift to an OG-lattice of character E6 [θ2 ]1 , and as Ω30 k lifts to φ8,1 , we get that Ω30 k ⊗ Ω−13 k lifts to an OG-lattice with character the sum of the non-unipotent character plus a projective OG-module. The sum of the degrees of the non-unipotent characters is (q 6 − 1)(q 8 − 1)(q 10 − 1)(q 12 − 1)(q 14 − 1)(q 18 − 1)(q 20 − 1)(q 24 − 1)Φ30 . Since this is larger than the degree of E6 [θ2 ]1 ⊗ φ8,1 , we deduce that Ω−13 k does not lift to E6 [θ2 ]1 , and we obtain the planar-embedded Brauer tree as in Figure 2. E6 [θ]1 E8 [ζ 2 ] E6 [θ]φ2,2 E6 [θ]ε St E8 [ζ 3 ] φ84,64 E6 [θ2 ]ε E6 [θ2 ]φ2,2 φ1344,38 φ4096,26 φ5670,18 φ4096,12 φ1344,8 E6 [θ2 ]1 Figure 2. Subtree of the principal Φ15 -block of E8 (q) φ84,4 1 THE BRAUER TREES OF UNIPOTENT BLOCKS 41 E8 [ζ 4 ] φ8,1 φ112,3 φ1400,7 φ4096,11 φ5600,19 φ4096,27 φ1400,37 φ112,63 φ8,91 E8 [ζ] Figure 3. Subtree of the principal Φ15 -block of E8 (q) It remains to locate E8 [θ] and E8 [θ2 ]. If they were not connected to φ8,91 , then Ω k would lift to an OG-lattice of character φ112,63 , although Ω30 k ⊗ Ω−11 k lifts to a lattice of character φ8,1 ⊗ E6 [θ2 ]1 plus a projective module. Since that tensor product has a degree smaller than φ112,63 , we obtain a contradiction. Consequently, we obtain the planar-embedded Brauer tree given in Figure 13, up to swapping E8 [θ] and E8 [θ2 ]. Assume the planar embedded tree shown in Figure 13 is not correct. Then Ω19 k lifts to a lattice of character E8 [θ]. Since Ω30 k ⊗ Ω−11 k lifts to a lattice of character φ8,1 ⊗ E6 [θ2 ]1 , we deduce that E8 [θ] occurs as a constituent of that tensor product, contradicting Remark 5.3. Consequently, the tree in Figure 13 is correct. 19 5.2.5. d = 20. The real stem of the tree is easily determined (see Figure 15). Let i and ζ in O× be the fourth and fifth roots of unity congruent to q 5 and q 4 modulo ℓ respectively. The difficult part is to locate the six cuspidal characters in the block. We have bH∗c (X(c), K)(q8 ) ≃ (E8 [ζ])[−8] ⊕ K[−16], bH∗c (X(c), K)(q16 ) ≃ (E8 [ζ 3 ])[−8] ⊕ D4,1 [−12]. Proposition 4.24 and Corollary 4.23 show that there is an edge connecting E8 [ζ] to St and that this edge comes between the edge containing φ112,63 and the one containing Stℓ in the cyclic ordering of edges containing St. Also, there is no cuspidal edge connected to a principal series character other than St and we have (5.3) bRΓc (X(c), k)(q8 ) ≃ 0 → PE8 [ζ] → P7 → · · · → P0 → 0, where P0 is in degree 16 and P0 , . . . , P7 are projective indecomposable modules labelling the principal series edges from 1 to St. Proposition 4.24 and Theorem 4.21 show that there is an edge connecting E8 [ζ 3] and D4,ε . We now want to locate the characters E8 [i] and E8 [−i]. A (Parity) argument shows that they are not connected to the non-unipotent vertex. The smallest Deligne– Lusztig variety in which they appear is associated to a 24-regular element w of 42 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER length 10. Note that ℓ ∤ |TvF | for all v ≤ w. In particular, the character η = [bH∗c (X(w), K)(1) ] = [St] + [E8 [−i]] is virtually projective. It follows from Lemma 4.17 that χexc +D4,ε does not occur in the decomposition of η in the basis of projective indecomposable. As a consequence, E8 [−i] is not connected by an edge to the D4 series, hence E8 [i] and E8 [−i] are connected to the Steinberg character. We are therefore left with determining the planar embedding around D4,ε and St. Assume that we are in the case shown in Figure 4. Let S0 , . . . , S4 be the E8 [ζ 3 ] D4,φ9,10 D4,ε E8 [ζ 2 ] Figure 4. Wrong planar embedding for the principal Φ20 -block of E8 (q) simple modules labelling the edges from D4,1 to the non-unipotent node so that e c (X(c), k)(q16 ) is given by [PS4 ] = χexc + D4,ε . A minimal representative of bRΓ E8 [ζ 3 ] S3 S3 S2 S1 S0 E8 [ζ 2 ] 2 D = 0 −→ E8 [ζ ] −→ S4 S2 −→ S3 S1 −→ S2 S0 −→ S1 S2 S1 S4 S0 E8 [ζ 3 ] 3 E8 [ζ ] S3 −→ 0, where the cohomology groups (represented by the boxes) are non-zero in degrees 8, 9 and 12 only. A non-zero map PE8 [ζ 2] → PE8 [ζ 3 ] gives a non-zero element of HomDb (kG) (D ∗ [−16], D). Consequently, H16 (D ⊗kG D) 6= 0. We have bRΓc (X(c), K)(q16 ) ⊗KG bRΓc (X(c), K)(q16 ) ≃ K[−24], and Proposition 3.7 shows that the cohomology of bRΓc (X(c), O)(q16 ) ⊗OG bRΓc (X(c), O)(q16 ) is torsion-free, hence Hi (bRΓc (X(c), k)(q16 ) ⊗kG bRΓc (X(c), k)(q16 ) ) = 0 for i 6= 24: this gives a contradiction. We now turn to the four possibilities for the planar embedding around the node labelled by the Steinberg character. We need to rule out the three of them shown in Figure 5. Recall that w denotes a 24-regular element. As in the case of X(c), Proposition 3.8 ensures that the torsion part in the cohomology of bRΓc (X(w), O) THE BRAUER TREES OF UNIPOTENT BLOCKS E8 [−i] E8 [ζ] E8 [ζ] St E8 [i] E8 [i] E8 [ζ] St E8 [ζ 4 ] E8 [ζ 4 ] 43 E8 [−i] St E8 [−i] E8 [ζ 4 ] E8 [i] Figure 5. Wrong planar embeddings for the principal Φ20 -block of E8 (q) e c (X(w), O)(q10 ) is cuspidal. Let C = bRΓ than 0 and greater than 20. We have red , a complex with 0 terms in degrees less KC ≃ E8 [i][−10] ⊕ K[−20]. We will describe completely the complex C, and rule out the wrong planar embeddings. We will proceed in a number of steps. • Step 1: the only non-cuspidal simple module that can appear as a composition factor of H∗ (kC) is K, and it can only appear in H20 (kC). The simple modules Stℓ , E8 [ζ 2] and E8 [ζ 3 ] do not occur as composition factors of H∗ (kC). The first statement follows from the discussion above. As a consequence, we have PSi ⊗kG kC ≃ 0 for i = 0, . . . , 3 and therefore PE8 [ζ 3 ] [−8] ⊗kG kC ≃ bRΓc (X(c), k)(q16 ) ⊗kG kC. The latter is a direct summand of RΓc (X(c) ×G X(w)), which by Proposition 3.7 has no torsion in its cohomology. We deduce that PE8 [ζ 3 ] ⊗kG kC is quasi-isomorphic to zero, which means that E8 [ζ 2 ] does not occur as a composition factor in H∗ (kC). The same result can be shown to hold for E8 [ζ 3], by replacing bRΓc (X(c), k)(q16 ) by (bRΓc (X(c), k)(q16 ) )∗ [−16] ≃ bRΓc (X(c), k)(q4 ) . The statement about Stℓ follows from Proposition 3.11. • Step 2: E8 [ζ 4 ] does not occur as a composition factor of H∗ (kC) and E8 [ζ] does not occur as a composition factor of Hi (kC) for i6∈{12, 13}. We have bRΓc (X(c), k)(1) ⊗kG kC ≃ k[−20] and RΓc (X(c), k)(q8 ) ⊗kG kC ≃ k[−36]. Moreover, Pi ⊗kG kC ≃ 0 for i = 1, . . . , 7 but P0 ⊗kG kC ≃ k[−20], so we obtain from (5.3) a distinguished triangle PE8 [ζ] [−9] ⊗kG kC → k[−36] → k[−36] ∗ Using RΓc (X(c), k)(q12 ) ≃ RΓc (X(c), k)(q8 ) [−16] instead of RΓc (X(c), k)(q8 ) , we obtain a distinguished triangle k[−20] → k[−20] → PE8 [ζ 4] [−7] ⊗kG kC 44 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER The variety X(w) has dimension 10, and therefore its cohomology vanishes outside the degrees 0, . . . , 20. Therefore PE8 [ζ] ⊗kG kC ≃ 0. We also deduce that PE8 [ζ 4] ⊗kG kC is quasi-isomorphic to either 0 or k[−12] ⊕ k[−13]. • Step 3: PS4 , PStℓ and PE8 [−i] and do not occur in C, while PE8 [i] occurs with multiplicity 1 in C (and this is in C 10 ). The statements about PE8 [±i] are clear using Proposition 3.5, while the other two statements follow from Lemma 4.17. We have now enough information to determine C and rule out the planar embeddings given in Figure 5. • Step 4: C i = 0 for i < 10. Let i be the smallest degree for which Hi (C) has non-zero torsion. Assume that i ≤ 10. The cohomology Hi−1 (kC) is cuspidal with socle in {S4 , E8 [i], E8 [−i]}. On the other hand, kC <(i−1) = 0 and the injective hulls of S4 and E8 [±i] do not occur as direct summands of kC i−1 , a contradiction. It follows that Hi (C) = 0 for i < 10 and H10 (C) is torsion-free. So, Hi (kC) = 0 for i < 10, hence (kC)i = 0 for i < 10. • Step 5: We have Hi (C) = 0 for 14 ≤ i ≤ 19 and H20 (C) = O. Lemma 4.20 applied to the stupid truncation C 13 → C 14 → · · · → C 20 (viewed in degrees −7, . . . , 0) shows that C ≃ 0 → C 10 → C 11 → C 12 → C 13 → P6 → P5 → · · · → P1 → P0 → 0, Hi (C) = 0 for 14 ≤ i ≤ 19 and H20 (C) = O. • Step 6: H10 (kC) = E8 [i] and H11 (C) = 0. By the universal coefficient theorem H10 (kC) is an extension of L = Tork1 (H11 (C), k) by kH10 (C) = E8 [i]. Since Ext1 (M, E8 [i]) = 0 for all kG-modules M with composition factors in {S4 , E8 [i], E8 [−i]}, it follows that the kG-module L is a direct summand of H10 (kC), hence C 10 has an injective hull of L as a direct summand of kC 10 . This shows that C 10 = PE8 [i] and L = 0. • Step 7: Ext1 (Stℓ , E8 [i]) = 0. The differential C 10 → C 11 induces an injective map Ω−1 E8 [i] ֒→ C 11 . Since PStℓ is not a direct summand of C 11 , it follows that Stℓ does not occur in the socle of C 11 , hence not in the socle of Ω−1 E8 [i]. This rules out the first possibility of the planar embedding around St in Figure 5. • Step 8: H12 (C) = 0. Let L = Tork1 (H12 (C), k). The kG-module Ω−2 E8 [i] has no composition factors isomorphic to S4 , E8 [i] or E8 [−i], hence Hom(L, Ω−2 E8 [i]). It follows that Ext1 (L, Ω−1 E8 [i]) = 0, hence an injective hull of L is a direct summand of kC 11 , which forces L = 0, hence H12 (C) = 0. • Step 9: C 13 ≃ P7 . We have C 13 ≃ P7 ⊕ R for some projective kG-module R, whose head is in 13 H (kC). It follows from Steps 1-3 that R ≃ PE⊕n for some n ≥ 0. Assume that 8 [ζ] THE BRAUER TREES OF UNIPOTENT BLOCKS 45 n > 0. Since Stℓ does not occur as a composition factor of H13 (kC), it follows that E8 [εi] must occur immediately after E8 [ζ] in the cyclic ordering around St for some ε ∈ {+, −} and PE8 [εi] occurs as a direct summand of C 12 : this is a contradiction. We deduce that C 13 ≃ P7 . • Step 10: conclusion. Assume that the configuration around St is the second one in Figure 5. Then −3 Ω E8 [i] is an extension of S6 by S5 . Since S5 does not occur as a composition factor of H12 (kC), it follows that P5 is a direct summand of C 13 , a contradiction. Assume now the configuration is the third one in Figure 5. The socle of Ω−3 E8 [i] is Stℓ . Since Stℓ does not occur as a composition factor of H12 (kC) and a projective cover does not occur as a direct summand of C 13 , we obtain a contradiction. This concludes the determination of the Brauer tree. Note that now Ω−2 E8 [i] = E8 [ζ], C 12 ≃ PE8 [ζ] and H12 (kC) = 0. In particular, H∗ (C) is torsion-free and C is 0 → PE8 [i] → PE8 [ζ] → PE8 [ζ] → P7 → P6 → · · · → P0 → 0. 5.2.6. d = 24. Suppose that d = 24. Let θ and i be the third and fourth roots of unity in O× such that θ = q 8 and i = q 6 in k. Several Harish-Chandra series lie in the principal Φd -block, and a (Hecke) argument gives the corresponding subtrees, as well as the real stem, as shown in Figure 16. • Step 1: cuspidal modules E8 [−θ] and E8 [−θ2 ]. The two cuspidal characters E8 [−θ] and E8 [−θ2 ] appear in the cohomology of a Coxeter variety X(c). To locate them on the Brauer tree we shall look at the cohomology of a compactification X(c) and proceed as in the beginning of §5.2.5. We have bRΓc (X(c), K)(q8 ) ≃ (E8 [−θ2 ])[−8] ⊕ K[−16]. we deduce from Corollary 4.23 and Theorem 4.21 (see Proposition 4.24) that (5.4) bRΓc (X(c), O)(q8 ) ≃ 0 → PE8 [−θ2 ] → P7 → · · · → P0 → 0, where P1 , . . . , P7 is the unique path of projective covers of non-cuspidal simple modules corresponding to edges from k to St in the Brauer tree, and the tree in Figure 6 is a subtree of T . Furthermore, the only principal series vertex connected by an edge to a non-principal series vertex is St. • Step 2: E6 -series. We now locate the E6 -series characters. By a (Degree) argument, E6 [θ]φ′′1,3 and E6 [θ2 ]φ′′1,3 are not leaves in the tree, so, by a (Parity) argument, must be connected to one of the non-unipotent vertex, D4,φ′′8,9 or D4,φ′8,3 , and they are connected to the same node. Note that E6 [θ±1 ]φ′′1,3 is connected to exactly two characters: E6 [θ±1 ]φ2,2 and the real character above (by a (Parity) argument, it cannot be connected to E8 [±i]). For all q, the degree of  [D4,φ′′8,9 ] − [E6 [θ]φ′′1,3 ] − [E6 [θ]φ2,2 ] + [E6 [θ2 ]φ′′1,3 ] − [E6 [θ2 ]φ2,2 ] 46 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER E8 [−θ] St φ35,74 φ160,55 φ350,38 φ448,25 φ350,14 φ160,7 φ35,2 1 E8 [−θ2 ] Figure 6. Subtree of the principal Φ24 -block of E8 (q) is negative, hence it cannot be the class of a kG-module. As a consequence, E6 [θ±1 ]φ′′1,3 is not connected to D4,φ′′8,9 . The same statement holds for D4,φ′8,3 , hence E6 [θ±1 ]φ′′1,3 is connected to the non-unipotent node. Again, (Parity) and (Degree) arguments show that the characters E8 [±i] are connected to the non-unipotent node, or one of the nodes D4,φ′′8,9 , E6 [θ]φ2,2 or E6 [θ2 ]φ2,2 . Note that, from the subtree constructed so far and that E8 [±i] and E6 [θ±1 ]φ′1,3 have the same parity, they must both be leaves in the tree and so remain irreducible modulo ℓ. Let w ∈ W be a regular element of order 24 and length 10. We have bRΓc (X(w), K)(q11 ) ≃ E8 [i][−10], bRΓc (X(w), K)(q14 ) ≃ E6 [θ]φ′1,3 [−12]. • Step 3: E8 [−θ] and E8 [−θ2 ] do not occur in H∗c (X(w), k)(λ) . Let λ be either q 11 or q 14 . The torsion part in bRΓc (X(w), O)(λ) is cuspidal by Proposition 3.8. Since its character has no composition factor in the principal series we have RΓc (X(w), k)(λ) ⊗kG Pi = 0 for i ∈ {0, . . . , 7}. Using Proposition 3.7 for the variety X(w)×G X(c) together with (5.4) and the dual description of bRΓc (X(c), O)(1) , we deduce that RΓc (X(w), k)(λ) ⊗kG PE8 [−θ] = RΓc (X(w), k)(λ) ⊗kG PE8 [−θ2 ] = 0. This ensures that neither E8 [−θ] nor E8 [−θ2 ] can occur as composition factors of the cohomology of bRΓc (X(w), k)(λ) . • Step 4: bRΓc (X(w), k)(q11 ) and position of E8 [±i]. Let C = bRΓc (X(w), k)(q11 ) and let M = Hi (C) be the non-zero cohomology group with largest degree. Suppose that i > 10. The module M is cuspidal and its composition factors are cuspidal modules different from E8 [−θ] and E8 [−θ2 ]. Proposition 3.5 shows that RHom•kG (RΓc (X(v), k), M) = 0 for all v < w. By the construction of the smooth compactifications, we obtain an isomorphism ∼ RHom•kG (RΓ(X(w), k), M) → RHom•kG (RΓc (X(w), k), M). THE BRAUER TREES OF UNIPOTENT BLOCKS 47 Since RΓ(X(w), k) has a representative with terms in degrees 0, . . . , ℓ(w) = 10, we deduce that HomDb (kG) (RΓc (X(w), k), M[−i]) = 0, which is impossible since C is a direct summand of RΓc (X(w), k) and the map C −→ M[−i] = Hi (C)[−i] is nonzero. This shows that Hj (C) = 0 for j > 10. Using the same argument with the isomorphism ∼ RHom•kG (M, RΓc (X(w), k)) → RHom•kG (M, RΓc (X(w), k)) and the fact that RΓc (X(w), k) has a representative with terms in degrees 10 = ℓ(w), . . . , 2ℓ(w) = 20, we deduce that Hj (C) = 0 for j < 10. Therefore C ≃ H10 (C)[−10]. Now, Proposition 4.18 and Remark 4.19 show that E8 [i] ≃ Ω12 k. We deduce that E8 [±i] are connected to the non-unipotent node and this gives the whole tree as shown in Figure 16, up to swapping the E6 [θ] and the E6 [θ2 ]-series. • Step 5: bRΓc (X(w), k)(q14 ) and conclusion. The previous argument applied to the complex D = bRΓc (X(w), k)(q14 ) shows that the cohomology of D vanishes outside the degrees 10, 11 and 12, and that H12 (D) is a module with simple head isomorphic to E6 [θ]φ′1,3 . The radical of H12 (D) is cuspidal. Since E6 [θ]φ′1,3 has no non-trivial extensions with simple cuspidal modules, we deduce that H12 (D) = E6 [θ]φ′1,3 and H12 c (X(w), O)q 14 is torsion-free. Let us denote the simple modules in the E6 [θ]-series as in Figure 7. There exists E6 [θ]φ′′1,3 S3 E6 [θ]φ2,2 S2 E6 [θ]φ′1,3 S1 Figure 7. Subtree of the principal Φ24 -block of E8 (q) a representative of D of the form D = 0 → X → P ′ ⊕ PS2 → PS1 → 0, where P ′ is a projective module with no cuspidal simple quotient except possibly E8 [−θ] or E8 [−θ2 ] (by Proposition 3.5). Since H11 (D) is a cuspidal module with no composition factor isomorphic to E8 [−θ] or E8 [−θ2 ], we deduce that the representative of D can be chosen so that P ′ = 0. By the universal coefficient theorem, we have H10 (D) ≃ H11 (D). We have H11 (D) = 0 or H11 (D) = S3 . In both cases, we find that X is a module with composition factors S2 and S3 . Proposition 4.18 and Remark 4.19 show that X ≃ Ω18 k in kG -stab. We deduce that Ω18 k lifts to an OG-lattice of character E6 [θ]φ′′1,3 , which gives the planar embedding. 48 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER 5.3. Other exceptional groups. The Brauer trees of unipotent blocks for exceptional groups other than E7 (q) and E8 (q) were determined in [13, 63, 47, 38, 48, 49] (under an assumption on q for one of the blocks in 2 E6 (q)), but only up to choice of field of values in each block. This ambiguity can be removed using Lusztig’s parametrization of unipotent characters. We achieve this by choosing carefully the roots of unity in Qℓ associated with the cuspidal characters, as we did in the previous sections. 5.3.1. E6 (q), 2 E6 (q), F4 (q) and G2 (q). For each of the exceptional groups of type E6 (q), 2 E6 (q), F4 (q) and G2 (q) there are only two blocks with cyclic defect groups whose Brauer trees are not lines. One of the blocks corresponds to the principal Φh -block with h the Coxeter number, and this case was solved in [29]. For the other one, one proceeds exactly as in §5.1.1, where only a pair of conjugate cuspidal characters lies outside the real stem (these characters appear in the cohomology of a Coxeter variety). The planar-embedded Brauer trees can be found in [19]. 5.3.2. 2 B2 (q 2 ) and 2 G2 (q 2 ). For the Suzuki groups 2 B2 (q 2 ) and the Ree groups G2 (q 2 ), the Frobenius eigenvalue corresponding to each unipotent character is known by [12]. It is enough to locate a single non-real character to fix the planar embedding. One can take this character to be a non-real cuspidal character occurring in the cohomology of the Coxeter variety and proceed as before to get the trees given in [19]. Note that for these groups the Coxeter variety is 1-dimensional, therefore its cohomology is torsion-free and Ω2 k is isomophic in kG -stab to the generalized (q 2 )-eigenspace of F 2 in RΓc (X(c), k) (when d is not the Coxeter number). 2 5.3.3. 2 F4 (q 2 ). We now consider the Ree groups 2 F4 (q 2 ), whose Brauer trees have been determined in [47] using the parametrization given in [58], but not using Lusztig’s parametrization. Here, there are three trees that are not lines. One of them corresponds to the case solved in [29], and another one is similar to §5.1.1. The√only block √ which deserves a specific treatment is the principal ℓ-block with ℓ | (q 4 + 2q 3 + q 2 + 2q + 1) (so, q is a 24-th root of unity modulo ℓ). Let η, i and θ be the roots of unity in O congruent to q 15 , q 6 and q 16 modulo ℓ respectively. A (Hecke) argument gives the real stem of the Brauer tree as shown in Figure 8 as well as the two edges for the 2 B2 -series. We consider the two generalized ‘mod- ℓ-eigenspaces’ of F 2 on the cohomology of the Coxeter variety given by  RΓc (X(c), K)(q−2 ) = 2 B2 [η 3 ]ε ⊕ 2 F4 [−θ2 ] [−2], RΓc (X(c), K)(q4 ) = 2 B2 [η 5 ]ε [−2] ⊕ K[−4]. Lemma 3.14 and Proposition 3.15 show that 2 B2 [η 3 ]ε + 2 F4 [−θ2 ] is the character of a projective module P2 F4 [−θ2 ] , hence 2 B2 [η 3 ]ε and 2 F4 [−θ2 ] are connected by an edge. Furthermore, RΓc (X(c), O)(q−2 ) ≃ P2 F4 [−θ2 ] [−2]. THE BRAUER TREES OF UNIPOTENT BLOCKS 49 Corollary 4.23 show that there is no non-real vertex connected to 1 or φ2,3 , that there is an edge S[η 5 ] connecting St and 2 B2 [η 5 ]ε , and, in the cyclic ordering of edges containing St, the edge S[η 5 ] comes after the one containing φ2,3 and before Stℓ . Furthermore, RΓc (X(c), O)(q−2 ) ≃ 0 → PS[η5 ] → P1 → Pk → 0, where Pk is in degree 4 and P1 is projective with character φ2,3 + St. We can now deduce the corresponding complexes of cohomology for X(c). For λ ∈ {q −2 , q 4 } and I an F -stable proper subset of S, we have bRΓc (X(cI ), O)(λ) = 0 unless (Li , F ) has type 2 B2 , in which case the complex has cohomology concentrated in degree 1. In addition, using duality for the case λ ∈ {q 6 , 1}, we find bRΓc (X(c), O)(q−2 ) ≃ 0 → 0 → P2 B2 [η3 ]1 → P2 F4 [−θ2 ] → 0 → 0 → 0, bRΓc (X(c), O)(q6 ) ≃ 0 → 0 → 0 → P2 F4 [−θ] → P2 B2 [η5 ]1 → 0 → 0, bRΓc (X(c), O)(q4 ) ≃ 0 → 0 → P2 B2 [η5 ]1 → PS[η5 ] → P1 → Pk → 0, bRΓc (X(c), O)(1) ≃ 0 → Pk → P1 → PS[η3 ] → P2 B2 [η3 ]1 → 0 → 0, where S[η 3 ] is the edge connecting St and 2 B2 [η 3 ]ε . Since RΓc (X(c) ×G X(c), O) is torsion-free (Proposition 3.7), we deduce that the differentials between non-zero terms of the complexes above cannot be zero. This determines uniquely the four complexes above up to isomorphism. We have bRΓc (X(c), O)(q−2 ) ⊗OG bRΓc (X(c), O)(q4 ) ≃ Hom•kG (0 → PS[η3 ] → P2 B2 [η3 ]1 → 0, 0 → P2 B2 [η3 ]1 → P2 F4 [−θ2 ] → 0)[−3]. By Proposition 3.7, this complex D has homology O concentrated in degree 2. Assume that, in the cyclic ordering of edges containing 2 B2 [η 3 ]ε , the edge containing 2 F4 [−θ2 ] comes after the edge containing 2 B2 [η 3 ]1 but before the edge containing St. Then a non-zero map kPS[η3 ] → kPF4 [−θ2 ] does not factor through kP2 B2 [η3 ]1 : so, it gives rise to a non-zero element of H4 (kD), a contradiction. It follows that the subtree obtained by removing 2 F4 [±i] is given by Figure 8. Let w ∈ W of length 6 such that wF has order 8 and let C = bRΓc (X(w), O)(−1) . It is a perfect complex; the torsion part of its cohomology is cuspidal by Proposition 3.8 and it does not involve Stℓ by Proposition 3.11. In addition, there is a representative of C that involves neither P2 F4IV [−1] nor PStℓ by Lemma 4.17. It follows that 2 F4IV [−1] does not occur as a composition factor of the cohomology of kC. Therefore the possible composition factors in the torsion part of H∗ (C) are the cuspidal simple modules 2 F4 [±i], 2 F4 [−θj ] and S[η m ]. The cohomology of KC is given by KC ≃ (F4 [i])[−6] ⊕ F4 [−θ]⊕3 [−8] ⊕ 2 B2 [η 5 ]⊕5 1 [−9] ⊕ K[−12]. 50 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER Using Proposition 3.7 one can easily compute kC ⊗LkG bRΓc (X(c), k)(λ) for the various eigenvalues λ of F 2 . With the same method as in Steps 1 and 2 of §5.2.5, the cases λ = q −2 , q 6 , q 4 , 1 show that • 2 F4 [−θ] can occur as a composition factor of H∗ (kC) only in degrees 8 or 9, because HomDb (kG (P2 B2 [η5 ]1 , kC[i]) = 0 for i 6= 9; • 2 F4 [−θ2 ] does not occur as a composition factor of H∗ (kC) because HomDb (kG (P2 B2 [η3 ]1 , kC[i]) = 0 for all i; • S[η 3 ] does not occur as a composition factor of H∗ (kC) because HomDb (kG (P2 B2 [η3 ]1 , kC[i]) = HomDb (kG (P1 , kC[i]) = 0 for all i; • S[η 5 ] can occur as a composition factor of H∗ (kC) only in degrees 9, 10 and 11 because HomDb (kG (P1 , kC[i]) = 0 for all i and HomDb (kG (P2 B2 [η5 ]1 , kC[i]) = 0 for i 6= 9. There are five distinct possible planar trees other than the one in Figure 8. One checks that for each of those five bad embeddings, • Ω−3 (2 F4 [i]) and Ω−4 (2 F4 [i]) do not contain 2 F4 [−θ] as a submodule, • Ω−4 (2 F4 [i]) does not contain 2 B2 [η 5 ]1 as a submodule, • Ω−4 (2 F4 [i]), Ω−5 (2 F4 [i]) and Ω−6 (2 F4 [i]) do not contain S[η 5 ] as a submodule, • Ω−7 (2 F4 [i]) does not contain k as a submodule, or • Ω−j (2 F4 [i]) does not contain 2 F4 [i] nor 2 F4 [−i] as a submodule for 1 ≤ j ≤ 6. 6+j Since kH6 (C) ≃ 2 F4 [i], it follows that Extj+1 (kC), kH6 (C)) = 0 for j ≥ 1 kG (H 7 6 and Ext1kG (TorO 1 (k, H (C)), kH (C)) = 0. Let D be the cone of the canonical map 6 kH (C) → kC[6]. We have HomDb (kG) (D, kH6 (C)[1]) = 0, hence kH6 (C) is isomorphic to a direct summand of C. Since C is perfect and 2 F4 [i] is not projective, we have a contradiction. This proves that the tree in Figure 8 is correct. THE BRAUER TREES OF UNIPOTENT BLOCKS 2 2 2 2 2 F4 [i] St 2 2 F4 [−θ] B2 [η 5 ]ε F4IV [−1] B2 [η 3 ]ε F4 [−i] 2 51 B2 [η 5 ]1 φ2,3 1 2 B2 [η 3 ]1 F4 [−θ2 ] Figure 8. Principal ℓ-block of 2 F4 (q 2 ) with ℓ | q 4 + √ 2q 3 + q 2 + √ 2q + 1 E6 [θ2 ]ε φ56,3 φ315,7 φ512,11 φ280,17 φ35,31 St E6 [θ]ε E6 [θ]1 φ7,46 φ56,30 φ315,16 φ512,12 φ280,8 φ35,4 E6 [θ2 ]1 Figure 9. Brauer tree of the principal Φ9 -block of E7 (q) The Φ9 -blocks of E8 (q) have isomorphic trees, with bijection of vertices given as follows. E7 (q) φ7,1 φ56,3 φ315,7 φ512,11 φ280,17 φ35,31 St E6 [θ2 ]ǫ E6 [θ]ǫ E8 (q), (A2, φ3 ) φ160,7 φ1008,9 φ2800,13 φ5600,21 φ4096,27 φ560,47 φ112,63 E6 [θ2 ]φ′′1,3 E6 [θ]φ′′1,3 E8 (q), (A2 , φ21 ) φ35,2 φ700,6 φ2240,10 φ3150,18 φ2240,28 φ700,42 φ35,74 E6 [θ2 ]φ2,2 E6 [θ]φ2,2 E8 (q), (A2 , φ13 ) φ112,3 φ560,5 φ4096,11 φ5600,15 φ2800,25 φ1008,39 φ160,55 E6 [θ2 ]φ′1,3 E6 [θ]φ′1,3 E7 (q) φ7,46 φ56,30 φ315,16 φ512,12 φ280,8 φ35,4 1 E6 [θ2 ]1 E6 [θ]1 E8 (q), (A2 , φ3 ) φ28,68 φ1575,34 φ4096,26 φ3200,22 φ700,16 φ50,8 1 E6 [θ2 ]φ1,0 E6 [θ]φ1,0 E8 (q), (A2, φ21 ) φ8,91 φ400,43 φ1400,29 φ2016,19 φ1400,11 φ400,7 φ8,1 E6 [θ2 ]φ2,1 E6 [θ]φ2,1 E8 (q), (A2 , φ13 ) φ1,120 φ50,56 φ700,28 φ3200,16 φ4096,12 φ1575,10 φ28,8 E6 [θ2 ]φ1,6 E6 [θ]φ1,6 1 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER φ7,1 52 Appendix A. Brauer trees for E7 (q) and E8 (q) E7 [i] φ7,46 φ27,37 φ168,21 φ378,14 φ378,9 φ168,6 φ27,2 φ7,1 Figure 10. Brauer tree of the Φ10 -block of E7 (q) associated to (2 A2 (q).(q 5 + 1), φ21 ) THE BRAUER TREES OF UNIPOTENT BLOCKS E7 [−i] 53 D4,ε1 D4,rε1 D4,rε2 D4,ε2 St 54 E7 [i] φ27,37 φ105,26 φ189,17 φ189,10 φ27,2 φ160,7 φ28,8 E7 [−i] Figure 11. Brauer tree of the principal Φ14 -block of E7 (q) E8 [−θ2 ] E8 [−1] φ28,68 φ160,55 φ300,44 φ840,26 φ1344,19 φ840,14 φ300,8 E8 [−θ] Figure 12. Brauer tree of the Φ12 -block of E8 (q) associated to (3 D4 (q), 3 D4 [1]) 1 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER φ105,5 12,3 E8 [ζ 2 ] φ112,63 φ8,91 E8 [θ] St φ84,64 φ1344,38 φ4096,26 φ5670,18 φ4096,12 φ1344,8 φ84,4 1 E6 [θ2 ]ε E6 [θ2 ]φ2,2 E8 [ζ 3 ] E6 [θ2 ]1 Figure 13. Brauer tree of the principal Φ15 -block of E8 (q) E8 [θ] E8 [−θ2 ] φ8,91 φ35,74 D4,φ′1,12 D4,φ′4,7 D4,φ′6,6 D4,φ′′4,7 D4,φ′′1,12 E8 [θ2 ] φ300,44 φ840,31 φ1134,20 φ840,13 φ300,8 φ35,2 φ8,1 THE BRAUER TREES OF UNIPOTENT BLOCKS E8 [ζ] E6 [θ]1 E6 [θ]ε E8 [ζ 4 ] E8 [θ2 ] φ1400,7 φ4096,11 φ5600,19 φ4096,27 φ1400,37 E6 [θ]φ2,2 E8 [−θ] Figure 14. Brauer tree of the Φ18 -block of E8 (q) associated to (2 A2 (q), φ21 ) 55 56 E8 [ζ 2] D4,φ9,2 D4,φ16,5 D4,φ9,10 D4,ε E8 [ζ 3] E8 [ζ] St E8 [−i] φ112,63 φ567,46 φ1296,33 φ1680,22 φ1296,13 E8 [ζ 4] Figure 15. Brauer tree of the principal Φ20 -block of E8 (q) φ567,6 φ112,3 1 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER D4,1 E8 [i] E6 [θ]φ′1,3 E6 [θ]φ2,2 E6 [θ]φ′′1,3 D4,φ′2,4 D4,φ′8,3 D4,φ12,4 D4,φ′′8,9 D4,φ′′2,16 E8 [i] St E8 [−θ2 ] φ35,74 φ160,55 φ350,38 φ448,25 φ350,14 E6 [θ2 ]φ2,2 E6 [θ2 ]φ′1,3 Figure 16. Brauer tree of the principal Φ24 -block of E8 (q) φ35,2 1 THE BRAUER TREES OF UNIPOTENT BLOCKS E6 [θ2 ]φ′′1,3 E [−i] E [−θ] 8 8 φ160,7 57 58 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER References [1] C. Bonnafé, J.-F.Dat and R. Rouquier, Derived categories and Deligne-Lusztig varieties II. Ann. Math., to appear. [2] C. Bonnafé and J. Michel. Computational proof of the Mackey formula for q > 2. J. Algebra 327 (2011), 506–526. [3] C. Bonnafé and R. Rouquier. Catégories dérivées et variétés de Deligne-Lusztig. Publ. Math. Inst. Hautes Études Sci. 97 (2003), 1–59. [4] C. Bonnafé and R. Rouquier. Coxeter orbits and modular representations. Nagoya Math. J. 183 (2006), 1–34. [5] R. Brauer. Investigations on group characters. Ann. of Math. 42 (1941), 936–958. [6] M. Broué. Isométries parfaites, types de blocs, catégories dérivées. Astérisque 181–182 (1990), 61–92. [7] M. Broué and G. Malle. Théorèmes de Sylow génériques pour les groupes réductifs sur les corps finis. Math. Ann. 292 (1992), 241–262. [8] M. Broué, G. Malle and J. Michel, Generic blocks of finite reductive groups. Astérisque 212 (1993), 7–92. [9] M. Broué, G. Malle and J. Michel. Towards spetses. I. Transform. Groups 4 (1999), 157–218. [10] M. Broué and J. Michel. Blocs et séries de Lusztig dans un groupe réductif fini. J. Reine Angew. Math. 395 (1989), 56–67. [11] M. Broué and J. Michel. Blocs à groupe de défaut abélien. Astérisque 212 (1993), 93–118. [12] O. Brunat. The Shintani descents of Suzuki groups and their consequences. J. Algebra 303 (2006), 869–890. [13] R. Burkhardt. Über die Zerlegungszahlen der Suzukigruppen Sz(q). J. Algebra 59 (1979), 421–433. [14] M. Cabanes and M. Enguehard. Unipotent blocks of finite reductive groups of a given type. Math. Zeitschrift 213 (1993), 479–490. [15] M. Cabanes and M. Enguehard. On unipotent blocks and their ordinary characters. Invent. Math. 117 (1994), 149–164. [16] M. Cabanes and M. Enguehard. Representation theory of finite reductive groups. Cambridge Univ. Press, 2004. [17] J. Chuang and R. Rouquier. Perverse equivalences and Calabi-Yau algebras. In preparation. [18] G. Cooperman, G. Hiß, K. Lux and J. Müller. The Brauer tree of the principal 19-block of the sporadic simple Thompson group. Experiment. Math. 6 (1997), 293–300. [19] D. Craven. Perverse equivalences and Broué’s conjecture II: The cyclic case. submitted. [20] C.W. Curtis and I. Reiner. Methods of representation theory, vol. II. John Wiley, 1987. [21] E. C. Dade. Blocks with cyclic defect groups. Ann. of Math. 84 (1966), 20–48. [22] P. Deligne and G. Lusztig. Duality for representations of a reductive group over a finite field, II. J. Algebra 81 (1983), 540–545. [23] F. Digne and J. Michel. Representations of finite groups of Lie type. Cambridge University Press, 1991. [24] F. Digne and J. Michel. Groupes réductifs non connexes. Ann. Sci. ENS 27 (1994), 345–406. [25] F. Digne, J. Michel and R. Rouquier. Cohomologie des variétés de Deligne-Lusztig. Adv. Math. 209 (2007), 749–822. [26] R. Dipper. On quotient of Hom-functors and representations of finite general linear groups. I. J. Algebra 130 (1990), 20–52. [27] O. Dudas. Deligne-Lusztig restriction of a Gelfand-Graev module. Ann. Sci. Éc. Norm. Supér. 42 (2009), 653–674. [28] O. Dudas. Coxeter Orbits and Brauer trees. Adv. Math. 229 (2012), 3398–3435. THE BRAUER TREES OF UNIPOTENT BLOCKS 59 [29] O. Dudas. Coxeter Orbits and Brauer trees II. Int. Math. Res. Not. 15 (2014), 4100–4123. [30] O. Dudas and R. Rouquier. Coxeter Orbits and Brauer trees III. J. Amer. Math. Soc. 27 (2014), 1117–1145. [31] M. Enguehard. Sur les ℓ-blocs des groupes réductifs finis quand ℓ est mauvais. Journal of Algebra 230 (2000), 334–377. [32] W. Feit. Possible Brauer trees. Illinois J. Math. 28 (1984), 43–56. [33] W. Feit. The representation theory of finite groups. North-Holland, 1982. [34] P. Fong and M. Harris. On perfect isometries and isotypies in finite groups. Invent. Math. 114 (1993), 139–191. [35] P. Fong and B. Srinivasan. Brauer trees in GLn (q). Math. Z. 187 (1984), 81–88. [36] P. Fong and B. Srinivasan. Brauer trees in classical groups. J. Algebra 131 (1990), 179–225. [37] M. Geck. Verallgemeinerte Gelfand–Graev Charaktere und Zerlegungszahlen endlicher Gruppen. Dissertation, Aachen (1990). [38] M. Geck. Generalized Gelfand–Graev characters for Steinberg’s triality groups and their applications. Comm. Algebra 19 (1991), 3249–3269. [39] M. Geck. Brauer trees of Hecke algebras. Comm. Algebra 20 (1992), 2937–2973. [40] M. Geck. Kazhdan-Lusztig cells and decomposition numbers. Represent. Theory 2 (1998), 264–277. [41] M. Geck. Character values, Schur indices and character sheaves. Represent. Theory 7 (2003), 19–55. [42] M. Geck, G. Hiss and G. Malle. Cuspidal unipotent Brauer characters. J. Algebra 168 (1994), 182–220. [43] M. Geck and G. Pfeiffer. On the irreducible characters of Hecke algebras. Adv. Math. 102 (1993), 79–94. [44] M. Geck and G. Pfeiffer. Characters of finite Coxeter groups and Iwahori–Hecke algebras. The Clarendon Press Oxford University Press, 2000. [45] M. Geck and G. Pfeiffer and S. Kim. Minimal length elements in twisted conjugacy classes of finite Coxeter groups. J. Algebra 229 (2000), 570–600. [46] G. Hiss. Regular and semisimple blocks of finite reductive groups. J. London Math. Soc. 41 (1990), 63–68. [47] G. Hiss. The Brauer trees of the Ree groups. Comm. Algebra 19 (1991), 871–888. [48] G. Hiss and F. Lübeck. The Brauer trees of the exceptional Chevalley groups of types F4 and 2 E6 . Arch. Math. (Basel) 70 (1998), 16–21. [49] G. Hiss, F. Lübeck and G. Malle. The Brauer trees of the exceptional Chevalley groups of type E6 . Manuscripta Math. 87 (1995), 131–144. [50] G. Hiss and K. Lux. Brauer trees of sporadic groups. Oxford University Press, 1989. [51] J. E. Humphreys. Reflection groups and Coxeter groups. Cambridge Univ. Press, 1990. [52] M. Linckelmann. Le centre d’un bloc à groupes de défaut cycliques. Comptes Rendus de l’Académie des Sciences 306 (1988), 727–730. [53] G. Lusztig. Representations of finite Chevalley groups. Regional Conf. Series in Math. 39, Amer. Math. Soc. 1978. [54] G. Lusztig. Coxeter orbits and eigenspaces of Frobenius. Invent. Math. 38 (1976), 101–159. [55] G. Lusztig. Homology bases arising from reductive groups over a finite field. In Algebraic groups and their representations (Cambridge, 1997), volume 517 of NATO Adv. Sci. Inst. Ser. C Math. Phys. Sci., pp. 53–72. [56] G. Lusztig. Rationality properties of unipotent representations. J. Algebra 258 (2002), 1–22. [57] G. Lusztig and N. Spaltenstein. Induced unipotent classes. J. London Math. Soc. 19 (1979), 41–52. 60 DAVID A. CRAVEN, OLIVIER DUDAS, AND RAPHAËL ROUQUIER [58] G. Malle. Die unipotenten Charaktere von 2 F4 (q 2 ). Comm. Algebra 18 (1990), 2361–2381. [59] G. Malle and D. Testerman Linear algebraic groups and finite groups of Lie type. Cambridge University Press, 2011. [60] J. Michel. The development version of the CHEVIE package of GAP3. J. Algebra 435 (2015), 308–336. [61] J. Müller. Brauer trees for the Schur cover of the symmetric group. J. Algebra 266 (2003), 427–445. [62] R. Rouquier. Complexes de chaînes étales et courbes de Deligne-Lusztig. J. Algebra 257 (2002), 482–508. [63] J. Shamash. Brauer trees for blocks of cyclic defect in the groups G2 (q) for primes dividing q 2 ± q + 1. J. Algebra 123 (1989), 378–396. [64] H. N. Ward. On Ree’s series of simple groups. Trans. Amer. Math. Soc. 121 (1966), 62–89. [65] E. Wings. Über die unipotenten Charaktere der Gruppen F4 (q). Dissertation, Lehrstuhl D für Mathematik, RWTH Aachen, 1995. School of Mathematics, University of Birmingham, Edgbaston, Birmingham B15 2TT, UK. E-mail address: [email protected] Université Paris Diderot, UFR de Mathématiques, Bâtiment Sophie Germain, 5 rue Thomas Mann, 75205 Paris CEDEX 13, France. E-mail address: [email protected] UCLA Mathematics Department, Los Angeles, CA 90095-1555, USA. E-mail address: [email protected]
4math.GR
Short-term Predictable Power of a Wind Farm via Distributed Control of Wind Generators with Integrated Storage arXiv:1602.04261v1 [cs.SY] 12 Feb 2016 Stefanos Baros1 Abstract— In this paper we introduce a Leader-follower consensus protocol and study its stability properties with and without communication delays. On the practical side, we explore its application on coordinating a group of wind Double-Fed Induction Generators (DFIGs) with integrated storage. To begin with, we establish asymptotic stability of the consensus protocol by employing singular perturbation theory. Subsequently, we establish asymptotic stability of the protocol under communication delays using a Lyapunov-Krasovskii functional. Lastly, we use the proposed protocol to design a methodology that can be adopted by a fleet of state-of-the-art wind generators (WGs). The objective is that the WGs selforganize and control their storage devices such that WF total power output is tracking a reference while equal contribution from each storage device is attained i.e equal power output from each storage. We demonstrate the effectiveness of our results and the corresponding approach via simulations on the IEEE 24-bus RT system. I. I NTRODUCTION According to a US Department of Energy (DOE) study [1], wind generators (WGs) are going to produce one of the largest shares of renewable energy in the future. Specifically, the same study argues that, by 2020, 10% of the US electricity demand is expected to be produced by wind generators, offshore and onshore [1]. Clearly, US stake their renewable energy future on wind generators for meeting their national goals regarding the reduction of the carbon dioxide emissions. A similar status is evident in most of the countries around the world. However, when power systems accommodate high-levels of wind integration they have to face a crucial challenge. That is, to guarantee their stability, their reliability and their dynamical performance which becomes highly dependent on the control methods that are adopted by the various WGs. The problem of coordinating and controlling a group of WGs in order for the WF power output to track a reference had been studied in [2],[3],[4]. The authors in [2] introduced a centralized two-layer constant power control scheme for a WF comprised with wind DFIGs with integrated storage. We emphasize that, wind DFIGs with energy storage are the gold standard on WGs technology, able to offer increased flexibility for short-term scheduling and power output smoothing [5],[6]. On the high-layer [2], the wind farm supervisory controller computed the power references for the low-layer controllers of the DFIGs, where on the low-layer, each DFIG controller generated power equal to the reference using the storage. It is important to underline that centralized control 1 Stefanos Baros is with the ECE department of Carnegie Mellon University, Pittsburgh, PA, 15213 USA e-mail: [email protected] schemes as the one in [2], require communication links from each WG to a central wind farm controller and extensive computations centrally. As a result, centralized approaches are plagued by expensive communication network costs, single-point failures, high computational cost and delays [7]. In the future, the number of large WFs with hundreds of WGs is going to increase all over the world. This prediction rests on the numerous renewable energy national plans around the world that have the expected wind generation dominating the total generation share coming from the renewable energy resources. Nevertheless, in high-wind-integration settings, the disadvantages of using centralized approaches for the coordination and control of large wind farms will become more pronounced. Consider that, with centralized approaches, the performance and the reliability of a WF in providing services can be severely compromised. On the other hand, distributed methods for controlling WFs with state-of-the-art WGs are becoming very appealing due to the reduced communication infrastructure they require, the low computational costs and the increased reliability that characterizes them [7], [4]. In fact, such methods can enable future WFs to perform robustly and reliably toward achieving certain objectives e.g rapidly track a power reference point, providing frequency and inertial response or alleviating intermittency. Distributed approaches for addressing the problem of coordination and control of WGs were studied in [3],[4]. Particularly, in [3], the authors proposed a strategy, based on multi-agent theory, to coordinate the WGs in a microgrid. The total available wind power and the total demand were retrieved in a distributed way, using agents at each bus and two average consensus protocols that were executed in parallel. Subsequently, the total demand and the wind power were used to compute the set points of the WGs. In the same spirit, the authors in [4], derived a distributed controller to regulate the set-points of various WGs. The control objectives for these controllers were to generate a specific total power output and at the same time minimize the fatigue on the wind turbines. In [3],[4] distributed control of DFIGs with no integrated storage is concerned. Also, since regulation of the WGs setpoints was aimed, only control of the Rotor-side-converter (RSC) is studied. In this work, we are focusing on the particular problem of controlling the storage devices of state-of-the-art WGs in a fair-sharing fashion such that the WF total power matches a committed power output, that can be constant or quasi-constant. The only work related to this problem is the work in [2], which considered a centralized approach. We underline that state-of-the-art WGs 1 are expected to be widely deployed in the future. In light of that, deriving new distributed methods for their efficient control and coordination becomes both an emerging and important challenge. (1, 2) 2 A. Our Contributions Motivated by multi-agent systems theory, we make the following contributions toward addressing the problem of distributed coordination and control of state-of-the-art WGs: a) We propose a distributed leader-follower consensus protocol that requires minimal communication links. Using singular perturbation arguments we prove that this protocol converges to its asymptotically stable equilibrium point, under certain conditions. b) Further, we prove that the proposed protocol possesses a delay-independent stability property which induces certain robustness to it with respect to timedelays that usually arise in communication networks c) On the practical side, we adopt the consensus protocol in a WF set-up and show that it can achieve the following objectives. It can effectively coordinate the storage devices of a group of WGs such that the WF total power matches a reference while consensus on the power output of the storage devices is reached. We outline the rest of the paper as follows. In Section II, we provide some preliminares and introduce the leaderfollower consensus protocol. In Section III, stability analysis of the proposed protocol is studied, using singular perturbation arguments. In the same Section, stability analysis under time-delays using Lyapunov-Krasovskii functionals is presented in detail. In Section IV, we apply the proposed protocol in a WF, to address the problem of fair utilization of WGs storage devices for short-term power output tracking. In Section V, we validate our results with simulations on the modified IEEE-RTS 24-bus system. Finally, Section VI concludes the paper. II. P ROBLEM F ORMULATION A. Notation In this paper we use standard notation. We denote by R the set of reals and by C the set of complex numbers. Also, we denote by R+ the set of non-negative real numbers and with R++ the set of positive reals. The m-dimensional Euclidean space is denoted by Rm . We denote vectors and matrices with bold characters. With A ∈ Rm×n we denote a m × n matrix of reals. With A> we denote the transpose of A and with [a]ij the (i, j)-entry of the matrix A. With A  0 (A  0) we denote that the matrix A is positive definite (semidefinite). The spectrum of the matrix A (set of eigenvalues) is denoted by σ(A). A n×n diagonal matrix B is denoted by B = diag[bi ]ni=1 . Define a n-dimensional vector a as a = [ai ]ni=1 . The operator k·k2 denotes the standard Euclidean norm L2 . The maximum value of the vector a is denoted by ā. Similarly the maximum value of a scalar quantity z is given by z̄. With In we denote the n × n identity matrix and the with 0n×1 and 1n×1 a n × 1 column vector of zeros and ones respectively. With x being a variable, we denote its (i 1, i) (n 1, n) i n Fig. 1: Graph G = (V, E) denoting the information exchange between the agents time derivative with ẋ ( dx dt ). The operator Re(·) returns the real part of an imaginary number · ∈ C. B. Leader-Follower Consensus Protocol Consider a set of agents denoted by V , {1, ..., n} and indexed by indexed by i ∈ V. The exchange of information between the agents can be represented as a graph G = (V, E) where V is the set of nodes and n o E , (i, j) | j = i+1, i ∈ V\{n}, j ∈ V\{1} , E ⊆ V×V is the set of edges corresponding to the allowable communication between the agents as shown in Fig. 1. The edge (i, j) denotes information exchange between agent i and agent j. Without loss of generality we assign as leader l the agent l , 1. Observe that, each agent exchanges information with one neighboring agent as shown in Fig. 1. Now, let V̄ denote the set of agents without the leader. In our case, V̄ , {2, ..., n} where V̄ ⊂ V. In the above set-up, consider the case where the agents desire to reach consensus on a relevant state-variable zi , i ∈ V. The only constraint is that the sum of their consensus states has to be z ∗ . For addressing this problem under the above set-up, we propose the following Leader-follower Consensus Protocol [7]. Protocol P1 Leader agent  X  dξh , z∗ − zi , dt i∈V dzl , −kα,l (zl − ξh ) , dt zl , z1 , ξh ∈ R (1a) (1b) zl ∈ R Agent i dzi , −kα,i (zi − zi−1 ) , zi ∈ R, i ∈ V̄ (1c) dt The variables kα,l , kα,i ∈ R++ are the gains of the consensus protocol, ξh is the auxiliary state of the leader agent and zl its consensus state, whereas, zi , ∀i ∈ V̄, the consensus states of all the other agents. Let the following assumption hold. P Assumption 1: The information i∈V̄ zi can be retrieved by the leader agent l. The above information can be obtained by the leader agent through indirect information passing from each agent to the leader (even without direct connection). III. S TABILITY A NALYSIS OF THE L EADER -F OLLOWER C ONSENSUS P ROTOCOL and of the equilibrium point of the slow sub-system, when these systems are considered decoupled from each other. Succeeding that step, we use Theorem 11.4 [8] to directly deduce that a maximum bound ε∗ on the small parameter ε exists under which stability of the full coupled system is guaranteed. The equilibrium point of (2a)-(2c) is: A. Two-time-scales Property In the protocol P1 dynamics a two-time-scales property emerges for certain values of the gains kα,l , kα,i . To prove that, we show that the system given by equations (1a)(1c) attains a standard singular perturbed form [8]. Let the vectors z and z̃ be defined as z , [z2 ... zn ]> , z ∈ Rn−1 z̃ , [zl z> ]> , z̃ ∈ Rn respectively. Then, the following proposition holds. Proposition 1: The protocol P1 given by Eq. (1a)-(1c) acquires a two-time-scales property for ka,l = ka,i  1, ∀i. Proof: Without loss of generality we consider that kα,l = kα,i , ∀i ∈ V̄. Accordingly, we define the time-scale 1 1 = kα,i ∈ R, ∀i ∈ V. With separation parameters as ε = kα,l those, equations (1a)-(1c) become:  X dξh , z∗ − zi ) (2a) dt ξh0 = z ∗ /n (4a) zl0 = ξh0 (4b) zi0 = ξh0 , ∀i ∈ V̄ (4c) Define the consensus subspace as: S , {z̃ ∈ Rn | z̃ = β · 1n×1 , β ∈ R} (5) and let z̃0 denote the equilibrium point of z̃. Notice that, consensus between the agent state-variables zi , i ∈ V can be reached, whenever the next two conditions are satisfied: Condition 1: z̃0 ∈ S Condition 2: z̃0 is asymptotically stable Observing equations (4a)-(4c), we readily have that Con∗ dition 1 is met with β = zn . We are left to establish that Condition 2 is met. For this purpose, we define the augmented variables yl , (zl − ξh ), yi , (zi − ξh ), ψh , (ξh − ξh0 ) that move the equilibrium point (ξh0 , z̃> 0 ) to the origin. Using those, equations (2a)-(2c) become: i∈V dzl ε , −(zl − ξh ) , l , 1 (2b) dt dzi ε , −(zi − zi−1 ) , i ∈ V̄ (2c) dt Then, equations (2a)-(2c) can be written in the next form: dξh , gh dt dz̃ ε , g̃ dt (3a) (3b) where dz̃ dzl dzi dzn > ,[ ... ... ] dt dt Xdt dt gh , (z ∗ − zi ) ∈ R (3c) (3d) i∈V g̃ , −[(zl − ξh )...(zn − zn−1 )]> (3e) n with dz̃ dt , g̃ ∈ R being smooth vector fields and ε the corresponding small positive parameter. The equations (3a),(3b) are in the standard singularly perturbed form with two time scales, namely t and τ = εt . That, completes the proof. B. Asymptotic Stability In this section, we rely on the two-time-scales property stated above and singular perturbation theory for proving asymptotic stability of the equilibrium point of P1 in a compartmental fashion. The above analysis reveals that the slow quasi-steady state is the augmented state-variable ξh of the leader agent l, and the fast boundary-layer states, are the consensus state-variables z̃. We first establish the stability of the equilibrium point of the fast sub-system Slow sub-system X dψh , −nψh − yi dt (6a) i∈V Fast boundary-layer sub-system dyl dψh , −yl − ε , y l , y1 dt dt dyi dψh ε , −(yi − yi−1 ) − ε , ∀i ∈ V̄ dt dt ε (6b) (6c) Define the following vectors; y = [y2 ... yi ... yn ]> , y ∈ Rn−1 and ỹ = [yl y> ]> , ỹ ∈ Rn and let the variable τ = εt denote the fast time-scale. On this time scale, the slow statevariable ψh appears as “frozen” and therefore we can assume h that dψ dτ ≈ 0. Under this approximation, the fast decoupled sub-system becomes: dyl = −yl (7a) dτ dyi = −(yi − yi−1 ), ∀i ∈ V̄ (7b) dτ The following lemma establishes asymptotic stability of the fast sub-system equilibrium point . Lemma 1: The equilibrium point of the system (7a)-(7b), (ỹ0 = 0n×1 ), is asymptotically stable. Proof: In matrix form we have:  dyl     −1 0 · · · 0 0 yl dτ  ..   ..   . . .. ..   ...   . = .  dyn dτ 0 | 0 ··· {z Af 1 −1 yn } Denoting the eigenvalues of Af with λ = [λl λ2 · · · λn ]> and noticing that Af is a lower triangular matrix [9], gives that λ = −1n×1 . That is equivalent to Af being a Hurwitz matrix and to the origin of the system (7a), (7b) being asymptotically stable [8]. That, concludes the proof of Lemma 1. According to Theorem 4.6 [8] the above statements are valid if and only if for any Q  0, ∃P  0 that satisfies: PAf + A> f P = −Q, Q, P ∈ Rn×n (8) with stability certificate the Lyapunov function given by: > Vf = ỹ Pỹ (9) dψh = −nψh (10) dt Lemma 2: The equilibrium point of the slow sub-system (10), (ψh0 = 0), is asymptotically stable. Proof: Take a candidate Lyapunov function as: Vh > 0, ∀ψh ∈ Dψh \ {0} dỹ = A0 ỹ + A1 ỹ(τ − r) dτ (12) where A0 , −In and A1 defined as:  >  0 0 A1 , (n−1)×1 I(n−1) 0(n−1)×1 (13) We construct a Lyapunov-Krasovskii functional [10] as: Z τ > V1 = ỹ(τ ) P1 ỹ(τ ) + ỹ(η)> Q1 ỹ(η)dη (14) τ −r We now turn our focus on the slow time scale t and notice that Lemma 1 establishes limτ →∞ (yl ) = 0, limτ →∞ (yi ) = 0, ∀i ∈ V̄. With these, the slow sub-system (10) can be approximated by: Vh = ψh2 , The time-delayed version of (7a)-(7b) is: (11) Computing its time derivative along the trajectories of the 2 h system in (10) leads to dV dt = −2n(ψh ) with n > 0 and dVh dt < 0, ∀ψh ∈ Dψh \ {0} which finally yields that ψh0 = 0 is asymptotically stable. Lemmas 1, 2 establish asymptotic stability of the equilibrium points of the fast and the slow reduced sub-systems respectively. Nonetheless, asymptotic stability of the full system (6a)-(6c) equilibrium point cannot be seamlessly inferred since the decoupled system dynamics serve as an approximation of the full system dynamics. To grapple with this challenge, we resort to a methodology, described in [8], whose sketch we briefly provide next. To prove asymptotic stability of the equilibrium point of the full system, a candidate Lyapunov function Vc can be first formulated from the linear combination of the Lyapunov functions Vf and Vh . Further, by deriving the time derivative of this composite candidate Lyapunov function along the trajectories of the full system we can impose negative definetess conditions in it. That, will lead a maximum bound ε∗ (Theorem 11.4, [8]) on the small parameter ε which can be interpreted as the maximum value of the time-scale separation ratio for which asymptotic stability of the full system is still assured. C. Delay-Independent Asymptotic Stability In practice, when distributed protocols are implemented, communication delays in the exchange of information arise. In this section, we study the asymptotic stability property of the fast sub-system equilibrium point under a specific type of delays, namely the fixed-time delays. Denoting the delays with r ∈ R+ we emphasize that, while the time-delays can be h arbitrary, the approximation dψ dτ ≈ 0 has to be valid so that the slow state-variable can still be considered as “frozen”. n×n where P1 , Q1 ∈ R and P1 , Q1  0. Direct differentiation with respect to τ yields: dV1 = ỹd> Q̃1 ỹd dτ (15) where  Q̃1 , P1 A0 + A> 0 P1 + Q1 A> 1 P1 P1 A1 −Q1  (16) and ỹd> , (ỹ(τ )> ỹ(τ − r)> ), ỹd ∈ R2n (17) From the Lyapunov-Krasovskii Stability Theorem [10], we have that the delay-independent asymptotic stability of the equilibrium point is guaranteed whenever dV1 < 0, ∀ỹd ∈ D1 \ {0(2n×1) } (18) dτ To prove (18), we introduce the next Lemma. Lemma 3: ∃P1 , Q1  0 diagonal matrices such that for Q̃1 , given by (16), it holds: Q̃1 ≺ 0 (19) Proof: Without loss of generality we take P1 , Q1 to be positive definite diagonal matrices. The Schur complement conditions on the matrix Q̃1 , give that Q̃1 ≺ 0 is equivalent to P1 A0 + A> 0 P1 + Q1 ≺ 0 together with the Schur complement matrix (of Q̃1 ), S1 , satisfying S1 ≺ 0. Performing basic matrix calculations gives: P1 A0 +A> P1 + Q1 =  0 q1 − 2p1 0  0 q − 2pi i    0 0 0 0 ··· ··· .. . · 0 0 0 qn − 2pn      which is negative definite (≺ 0) when: qi − 2pi < 0, ∀i ∈ V The Schur complement is obtained as:  p22 −q1 − 2p2 −q 0 2  p2  0 −qi − 2pi+1i+1 −qi+1  S1 =   0 0 0 0 (20a) ··· ··· .. . ··· 0   0    0  −qn which is negative definite (≺ 0) when: p2i+1 < 0, ∀i ∈ V \ {n} 2pi+1 − qi+1 −qn < 0 (20b) (20c) W F Bus −qi − 2 1 n 3 Restricting P1 , Q1 to be positive definite yields: qi > 0 , ∀i ∈ V (20d) (20e) Now, we have to prove that we can find pi , qi , ∀i ∈ V, that satisfy the inequalities (20a)-(20e) i.e LMI feasibility. Our proof for the LMI feasibility is constructive; we introduce a methodology for finding pi , qi to meet the inequalities. We begin by choosing pi , ∀i ∈ V s.t (20d) holds. Next, choose the qn s.t (20e),(20a) together hold and each qi , ∀i ∈ V \{n} s.t both (20a), (20b) are satisfied. Using this methodology, we can always find pi , qi , ∀i ∈ V that meet the inequalities (20a)-(20e). Hence, the LMI problem in (19) is indeed feasible and we conclude the proof. Theorem 1: The equilibrium point (ỹ0 = 0n×1 ) of the system (12) is delay-independent asymptotically stable. Proof: From Lemma 3 and Lyapunov-Krasovskii Stability Theorem [10] we conclude the proof. Intuitively, Theorem 1 guarantees that the fast consensus state-variables zi , i ∈ V converge to the slow state-variable ξh independently of any time-delays, as long as ψh remains “frozen”. That is of practical interest, since it certifies that the performance of the protocol P1 is robust with respect to time-delays that are inherent in communication channels. IV. WF S HORT- TERM P REDICTABLE P OWER V IA D ISTRIBUTED C ONTROL OF THE WG S In this Section, we apply the proposed protocol for coordinating a group of WGs storage devices in a WF toward achieving a common objective. The group objective is shortterm predictable power from the WF i.e provide constant power or track a reference with load sharing between the WGs. Let the set G , {1, ..., n} denote the WGs with integrated storage. To ease the notation, we simply denote a wind generator of this type by WG and index it by i, where i ∈ G. Without loss of generality we take l , 1. Let Ḡ denote the set of WGs without the leader WG. In our case, Ḡ , {2, ..., n}. In addition, we assume that each WG is an agent and we have WGs exchanging information between each other. Let Pe,i , Pg,i , Pr,i and Pst,i denote the electric power produced by the stator, the grid-side converter (GSC), the rotor-side converter (RSC) and the storage respectively. All these can be seen in Fig.2c. In general, the WFs obtain a reference power output, Pd from a system operator and this corresponds to the WF committed output. The committed power output is obtained as an outcome of a forecasting method and an Economic Dispatch process, conducted by the system operator taking into account local wind conditions. Consider the dynamics of the capacitor that interfaces the Rotor Side Converter (RSC) and the GSC of each WG, as W F Supervisory Controller (a) pi > 0 , ∀i ∈ V 2 1 n 3 (b) Pe,i Gearbox Pg,i Pr,i Capacitor RSC GSC Pst,i Consensus state-variable Energy Storage (c) Fig. 2: a) Physical topology of the WF b) Communication links with integrated storage c) WG (Fig.2c): dVdc,i , (Pr,i + Pst,i − Pg,i ), i ∈ G (21) dt where Cdc,i , Vdc,i ∈ R are the capacitance and the voltage potential of the capacitor respectively. When the storage is inactive, at the equilibrium we have: (Cdc,i Vdc,i ) Pg,i = Pr,i , i ∈ G (22) Taking that into account, we have that the total WF power output is equal to the total available mechanical power that is coming from the wind i.e: X X Pm,i ≈ (Pe,i + Pr,i ) (23) i∈G i∈G P The total mechanical power i∈G Pm,i is highly variable and since it depends on the wind speed conditions, it varies minute by minute. As a result, for a WF we might have that: X X Pm,i ≈ (Pe,i + Pr,i ) < Pd (24) i∈G i∈G The opposite case, where more wind power is available does not pose any problem since the WF can “spill” wind. On the other hand, the reasons for which the WF cannot generate Pd are twofold. Firstly, due to time delays between the time that the system operator conducts the wind forecasting and defines Pd until the time that this set-point is communicated to the WF. In this case, the WF available mean power can actually be less than Pd , making it not possible for the WF to generate the required amount of power. Secondly, due to the continuous wind fluctuations, the WF can only generate Pd on average. Although the WF cannot always extract (from the wind) power equal to Pd , it can generate a total power Pd when the WGs have storage devices, able to contribute. When the storage is active i.e Pst,i > 0 or Pst,i < 0, at the equilibrium of (21), we get: Pst,i = (Pg,i − Pr,i ), i ∈ G (25) The additional flexibility offered by the integrated storage can enable the WGs to actually generate the required Pd , given that the storage devices have stored enough energy. Hence, the WGs can provide short-term predictable power and balance wind intermittency as well. With storage, we have: X (Pe,i + Pr,i + Pst,i ) = Pd (26) i∈G The storage devices can provide or draw the excess power that is required by the WF to meet its committed power output in a short time-scale. That, can be translated into the next condition for the total storage power. Condition 3: X X lim Pst,i = Pd − (Pe,i + Pr,i ) (27) t→∞ i∈G i∈G For the scope of utilizing the available resources efficiently, the storage devices can be controlled such that they contribute equally to the excess power required by the WF to reach Pd . That, is translated into the following condition. Condition 4: lim Pst,i = lim Pst,j , t→∞ t→∞ ∀i, j ∈ G (28) Consider the following definition. Definition 1: Any utilization scenario between the energy storage devices of the WGs for which Conditions 3,4 hold, is called a fair utilization scenario. The problem we seek to address can now be formulated as. Problem 1: To coordinate the energy storages of various state-of-the-art WGs in a completely distributed way i.e each WG exchanges information with only its neighbor-WGs, such that Conditions 3,4 hold. Next, we introduce a methodology based on the Leaderfollower Consensus Protocol P1 to address Problem 1. A. Proposed Methodology In this section, we propose a methodology that addresses Problem 1. In our set-up, the physical and the communication topology of the WF is as shown in Fig. 2a, 2b. Notice that, the communication topology is exactly identical to the one in Fig. 1 with each WG exchanging information with one neighboring WG. Also, the physical topology is in alignment with the communication topology. Using Eq. (25), Condition 4 yields to: Condition 5: lim zi = lim zj , i, j ∈ G t→∞ (29) t→∞ where zi , (Pg,i − Pr,i ), ∀i ∈ G, zi ∈ R. It springs from (29) that, we can pose Problem 1 as a constrained consensus agreement problem. Further, we can address this problem by requiring each WG to exhange its zi information with one of its neighbors according to Fig.2b, until they reach consensus on their zi0 s. In this case, we would ensure that Eq. (29) holds. To realize this process, we propose the protocol given below. Definition 2 (Protocol P2 ): Leader WG   X dξh , Pd − (Pe,i + zi + Pr,i ) , dt i∈G dzl , −kα,l (zl − ξh ), dt zl , z1 ξh ∈ R zl ∈ R (30a) (30b) WG i dzi , −kα,i (zi − zi−1 ) , zi ∈ R, i ∈ Ḡ (30c) dt where ξh is the additional state-variable of the leader WG. P Assumption 2: The information i∈Ḡ (Pe,i + zi + Pg,i ) can be retrieved bythe leader WG.  P By defining z ∗ = Pd − i∈G (Pe,i + Pr,i ) , the Protocol P2 takes the form of the Protocol P1 . Thus, the preceding stability analysis of P1 still applies. Notice that, P2 guarantees that the total power output of the WF will be Pd h since at the equilibrium, equation dξ dt = 0 is identical to (26). Summarizing, the Protocol P2 effectively addresses the Problem 1. To implement P2 the GSC and the storage controllers have to be designed appropriately Fig.2c. We briefly describe how this can be done ommiting the full derivation due to space limitation. The GSC controller for the leader and the rest of the WGs can be designed using equations (30b) and (30c). Besides that, the storage controllers are critical to be designed such that by the time the GSCs reach consensus on the zi0 s, for the storage power, it holds: xi = zi , ∀i ∈ G (31) where xi = Pst,i , ∀i. A storage controller that achieves this objective can be found in our previous work, [11], [12]. As a general guideline, we emphasize that the storage controller has to guarantee that xi is tracking the consensus state zi , regulated by the GSC, in a much faster time-scale than the time-scale of the consensus protocol. This is very important for having: Condition 5 is true ⇒ Condition 4 is true That, results to guaranteed consensus on Pst,i , ∀i ∈ G, whenever consensus on the zi , ∀i ∈ G is reached. P Pd 0.7  18    21    22   0.6  23    17   16   (p:u) 0.5  20    19   3 3.5 0.5 1 4 i2G (Pe;i + Pg;i ) 0.4 0.3  15   0.2  14    13   0.1  11    24    12   1.5 2 2.5 3 3.5 4 4.5 time(s)  3    9    10   (a) Total WF power output  6    4   x1 x2 x3 x4 x5 x6 x7 x8 x9 x10 0.06  5    8   0.04  2    7   (p:u)  1   Fig. 3: Modified IEEE 24-bus RT system Under Scenario 1, the WF dynamics evolve as shown in Fig. 4. More specifically, we observe from Fig. 4a that the total power of the WF is tracking the reference with good transient response i.e no oscillations, no overshoot. Hence, the WF total power output is reaching the committed reference rapidly. Additionally, from Fig. 4b we can see that the power coming from each storage is identical. This, verifies that the storage power outputs reached consensus. Direct comparison of the responses in Fig. 4c and Fig. 4b gives that, the storage power outputs and the consensus states are one-to-one identical to each other. This, verifies the effectiveness of the time-scale-separation-based design of the GSC and storage controllers which we explain as follows. The storage devices continuously generate the power that is required by the GSCs, whereas, the GSCs via the consensus protocol, regulate their power to match the total WF power, 0 -0.02 1 2 3 4 time(s) (b) Storage power of each WG, Pst,i 9h z1 z2 z3 z4 z5 z6 z7 z8 z9 z10 0.06 0.04 (p:u) V. P ERFORMANCE E VALUATION To validate our results, we apply our methodology on the modified IEEE 24-bus RT system. The goal is the WF power output to track a varying reference when is suddenly ordered to and when the available wind power is not sufficient. Also, the storage power outputs are desired to contribute to this power mismatch in an equal-sharing manner. At bus 22 of the 24-bus power grid, we have a WF comprised of 10 WGs where each of them has a storage encapsulated into its scheme Fig.2c. The physical as well as the communication structure of the WF is the same as in Fig.2a, 2b. We verify our proposed methodology, by conducting simulations under the next 2 critical scenarios. • Scenario 1: Step-wise changes on the reference Pd and no time-delays. • Scenario 2: Step-wise changes on the reference Pd and fixed delays r = 5ms. 0.02 0.02 0 -0.02 1 2 3 4 time(s) (c) Consensus state of each WG zi Fig. 4: WF response under Scenario 1 Pd . In the second scenario, we regarded time-delays on the exchange of information between the WGs. The WF response under this scenario can be seen in Fig. 5. It is obvious that the time-delays deteriorated the tracking response, Fig.5a, but the control objectives were still achieved. Specifically, the WF power output was tracking a reference and consensus on the power output of each storage device Fig.5b, was attained. These results verify the time-delay independent stability property of the proposed protocol, proved in Section IIIC. We conclude from the above results that, our proposed methodology is able to coordinate a fleet of WGs such that the total power from a WF is tracking a reference while using the available storage devices efficiently i.e addressing Problem 1. P !d 0.6 0.5 G i∈W of state-of-the-art WGs toward achieving a group objective. In our case the group objective was the WF power output to track a reference using the storage devices contributing in a fairly manner. Finally, we used the IEEE 24-bus RT system to verify our theoretical results and to demonstrate the effectiveness of our proposed methodology. In future work, we would like to explore how our methodology can be advanced such that the WFs can provide more services to the grid e.g frequency and inertial response, with efficient coordination of its available WGs. Pe,i + Pg,i [p.u] 0.4 0.3 0.2 0.1 0 0.5 1 1.5 2 2.5 3 3.5 4 time[s] (a) Total WF power output 20 ×10 -3 Pst,1 Pst,2 Pst,3 Pst,4 Pst,5 Pst,6 Pst,7 Pst,8 Pst,9 Pst,10 15 [p.u] 10 5 0 0.0105 0.01 3.06 3.08 3.1 3.12 3.14 -5 0.5 1 1.5 2 2.5 3 3.5 4 time[s] (b) Storage power of each WG, Pst,i 0.025 ξh z1 z2 z3 z4 z5 z6 z7 z8 z9 z10 0.02 [p.u] 0.015 0.01 0.005 0 0.0112 0.011 0.0108 3.02 3.04 3.06 3.08 3.1 3.12 -0.005 0.5 1 1.5 2 2.5 3 3.5 4 time[s] (c) Consensus state of each WG zi Fig. 5: WF response under Scenario 2 VI. C ONCLUDING R EMARKS AND F UTURE W ORK In this paper, we introduced a Leader-follower Consensus Protocol under a specific communication set-up. We first analyzed and proved its stability properties using singular perturbation arguments. Further, we extended these results by proving that its stability property is not altered under time-delays i.e its stability property is delay-independent. On the practical side, we showed how this protocol can find applications in power grids by proposing a methodology that can be adopted by WFs to address an emerging challenge. In particular the challenge is to coordinate, through appropriate communication and control, the storage devices of a fleet R EFERENCES [1] U.S Deparment of Energy. WindVision: A New Era for Wind Power in the United States. Technical report, U.S Deparment of Energy, March 2015. [2] L.Qu and W.Qiao. Constant Power Control of Dfig Wind Turbines With Supercapacitor Energy Storage. IEEE Transactions on Industry Applications, 47(1):359–367, January/February 2011. [3] W.Zhang, Y.Xu, W.Liu, F.Ferrese, and L.Liu. Fully Distributed Coordination of Multiple DFIGs in a Microgrid for Load Sharing. IEEE Transactions on Smart Grid, 4(2):806–815, June 2013. [4] Biegel B., Madjidian D., Spudic V., Rantzer A., and Stoustrup J. Distributed low-complexity controller for wind power plant in derated operation. In Proceeding of 2013 IEEE International Conference on Control Applications (CCA), pages 146–151. IEEE, August 2013. [5] C.Abbey and G.Joos. Integration of Energy Storage with a DoublyFed Induction Machine for Wind Power Applications. In 35th Annual Power Electronics Specialists Conference, volume 3, pages 1964– 1968. IEEE, June 2004. [6] N.W.Miller. GE Experience with Turbine Integrated Battery Energy Storage. In IEEE PES General Meeting, 2014. [7] H.Xin, Z.Qu, J.Seuss, and A.Maknouninejad. A Self-Organizing Strategy for Power Flow Control of Photovoltaic Generators in a Distribution Network. IEEE Transactions on Power Systems, 26(3):1462– 1473, August 2011. [8] H.K.Khalil. Nonlinear Systems. Prentice Hall, 3rd edition, 2002. [9] R.A. Horn and C.R. Johnson. Matrix Analysis. Cambridge University Press, 2013. [10] Keqin Gu, Vladimir L.Kharitonov, and Jie Chen. Stability of TimeDelay Systems. Springer Science+Business Media,LLC, 2003. [11] S.Baros. A Novel Ectropy-Based Control Scheme for a DFIG Driven by a Wind Turbine with an Integrated Energy Storage. In American Control Conference, July 2015. [12] S.Baros and M. Ilic. Robust Ectropy-based Cooperative Control of a Wind DFIG for Transient Stabilization and Mppt. In IEEE PES General Meeting, July 2015. [13] S.Baros and M.Ilic. Multi-objective Lyapunov-based Control of a STATCOM/BESS. In Innovative Smart Grid Technologies Conference (ISGT),, Feb 2015. [14] R. Olfati Saber, J. A.Fax, and R.M. Murray. Consensus and Cooperation in Networked Multi-Agent Systems. Proceedings of the IEEE, 95(1):215–233, January 2007. [15] A.D.Dominguez Garcia and C.N.Hadjicostis. Coordination and Control of Distributed Energy Resources for Provision of Ancillary Services. In IEEE International Conference on Smart Grid Communications (SmartGridComm), pages 537–542, October 2010. [16] S.Baros and M.Ilic. On Transient Stability and Voltage Regulation Through MIMO Feedback Linearizing Control of Generator and Energy Storage. In American Control Conference, July 2015. [17] W.Qiao. Dynamic Modeling and Control of Doubly Fed Induction Generators Driven by Wind Turbines. In PSCE, 2009, pages 1–8. IEEE, March 2009. [18] C.Abbey and G.Joos. Supercapacitor Energy Storage for Wind Energy Applications. IEEE Transactions on Industry Applications, 43(3):769– 776, May/June 2007. [19] F.A.Bhuiyan and A.Yazdani. Multimode Control of a DFIG-Based Wind-Power Unit for Remote Applications. IEEE Transactions on Power Delivery, 24(4):2079–2089, October 2009.
3cs.SY
Subexponential parameterized algorithms for planar and apex-minor-free graphs via low treewidth pattern covering∗ Fedor V. Fomin† Daniel Lokshtanov‡ arXiv:1604.05999v1 [cs.DS] 20 Apr 2016 Michal Pilipczukk Dániel Marx§ Marcin Pilipczuk¶ Saket Saurabh∗∗ Abstract We prove the following theorem. Given a planar graph G and an integer k, it is possible in polynomial time to randomly sample a subset A of vertices of G with the following properties: √ • A induces a subgraph of G of treewidth O( k log k), and • for every connected subgraph H of G on at√most k vertices, the probability that A covers 2 the whole vertex set of H is at least (2O( k log k) · nO(1) )−1 , where n is the number of vertices of G. Together with standard dynamic programming techniques for graphs of bounded treewidth, this result gives a versatile technique for obtaining (randomized) subexponential parameterized √ 2 algorithms for problems on planar graphs, usually with running time bound 2O( k log k) nO(1) . The technique can be applied to problems expressible as searching for a small, connected pattern with a prescribed property in a large host graph; examples of such problems include Directed k-Path, Weighted k-Path, Vertex Cover Local Search, and Subgraph Isomorphism, among others. Up to this point, it was open whether these problems can be solved in subexponential parameterized time on planar graphs, because they are not amenable to the classic technique of bidimensionality. Furthermore, all our results hold in fact on any class of graphs that exclude a fixed apex graph as a minor, in particular on graphs embeddable in any fixed surface. ∗ The research leading to these results has received funding from the European Research Council under the European Union’s Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreements no. 267959 (F. V. Fomin), no. 280152 (D. Marx), and no. 306992 (S. Saurabh). The work of D. Marx is also supported by Hungarian Scientific Research Fund (OTKA) grant NK105645. The research of Ma. Pilipczuk is supported by Polish National Science Centre grant UMO-2013/09/B/ST6/03136. The research of Mi. Pilipczuk is supported by Polish National Science Centre grant UMO-2013/11/D/ST6/03073. Mi. Pilipczuk is also supported by the Foundation for Polish Science (FNP) via the START stipend programme. Part of the research of F. Fomin, D. Marx, and Ma. Pilipczuk has been done when they were participating in the “Fine-grained complexity and algorithm design” program at the Simons Institute for Theory of Computing in Berkeley. † Department of Informatics, University of Bergen, Norway, [email protected]. ‡ Department of Informatics, University of Bergen, Norway, [email protected]. § Institute for Computer Science and Control, Hungarian Academy of Sciences (MTA SZTAKI), Hungary, [email protected]. ¶ Institute of Informatics, University of Warsaw, Poland, [email protected]. k Institute of Informatics, University of Warsaw, Poland, [email protected]. ∗∗ Institute of Mathematical Sciences, India, [email protected], and Department of Informatics, University of Bergen, Norway, [email protected]. 1 Introduction Most of natural NP-hard problems on graphs remain NP-hard even when the input graph is restricted to be planar. However, it was realized already in the dawn of algorithm design that the planarity of the input can be exploited algorithmically. Using the classic planar separator theorem of Lipton and√Tarjan [24], √ one can design algorithms working in subexponential time, usually of the form O( n) O( n log n) , for a wide variety of problems that behave well with respect to separators; 2 or 2 such running time cannot be achieved on general graph unless the Exponential Time Hypothesis (ETH) fails [22]. From the modern perspective, the planar separator theorem implies that a planar √ graph on n vertices has treewidth O( n), and the obtained tree decomposition can be used to run a Divide&Conquer algorithm or a dynamic programming subroutine. The idea of exploiting short separators plays a crucial role in modern algorithm design on planar graphs and related graph classes, including polynomial-time, approximation, and parameterized algorithmic paradigms. Let us take a closer look at the area of parameterized complexity. While for most parameterized NP-hard problems the exponential dependence on the parameter is the best we can hope for, under ETH, there are plenty of problems admitting subexponential parameterized algorithms when restricted to planar graphs. This was first observed in 2000 by Alber et al. [1], who obtained a subexponential parameterized algorithm √ for deciding whether a given n-vertex planar graph contains O( k) nO(1) . It turned out that the phenomenon is much more gena dominating set of size k in time 2 eral. A robust framework explaining why various problems like Feedback Vertex Set, Vertex Cover, Dominating Set or Longest Path admit subexponential parameterized algorithms on √ √ planar graphs, usually with running times of the form 2O( k) · nO(1) or 2O( k log k) · nO(1) , is provided by the bidimensionality theory of Demaine et al. [10]. Roughly speaking, the idea is to exploit the relation between the treewidth of a planar graph and the size of a grid minor that can be found in it. More√precisely, the following win/win approach is implemented. Either the treewidth of the graph is O( and the problem can be solved using dynamic programming on a tree decomposition, or √ k) √ a c k × c k grid minor can be found, for some large constant c, which immediately implies that we are working with a yes- or with a no-instance of the problem. Furthermore, it turns out that for a large majority of √problems, the running time yielded by bidimensionality are essentially optimal under ETH: no 2o( k) · nO(1) -time algorithm can be expected. We refer to the survey [11], as well as the textbook [7, Chapter 7] for an overview of bidimensionality and its applications. While the requirement that the problem can be solved efficiently on bounded treewidth graphs is usually not restrictive, the assumption that uncovering any large grid minor provides a meaningful insight into the instance considerably limits the applicability of the bidimensionality methodology. Therefore, while bidimensionality can be extended to more general classes, like excluding some fixed graph as a minor [10, 12], map graphs [8], or unit disk graphs [19], there are many problems that are “ε-close” to being bidimensional, and yet their parameterized complexity remains open for years. One example where such √ √ a situation occurs is the Directed Longest Path problem. While the existence of a k × k grid minor in an undirected graph immediately implies the existence of an undirected path on k vertices, the same principle cannot be applied in the directed setting: even if we uncover a large grid minor in the underlying undirected graph, there is no guarantee that a long directed path can be found, because we do not control the orientation of the arcs. Thus, Directed Longest Path can be solved in time 2O(k) nO(1) on general graphs using color coding√[2], but no substantially faster algorithms on planar graphs were known. On the other hand, no 2o( k) · nO(1) -time algorithm on planar graphs can be expected under ETH, which leaves a large gap between the known upper and lower bounds. Closing this embarrassing gap was mentioned as an open problem in [5,15,32,33]. A similar situation happens for Weighted Longest Path, where 1 we are looking for a k-path of minimum weight in an edge-weighted planar graph; the question about the complexity of this problem was raised in [5, 31]. Another example is k-Cycle: deciding whether a given planar graph contains a cycle of length exactly k. While√ the property of admitting a cycle of length at least k is bidimensional, and therefore admits a 2O( k) nO(1) -time algorithm on planar graphs, the technique fails for the variant when we ask for length exactly k. This question was asked in [5]. We will mention more problems with this behavior later on. The theme of “subexponential algorithms beyond bidimensionality” has recently been intensively investigated, with various success. For a number of specific problems such algorithms were found; these include planar variants of Steiner Tree parameterized by the size of the tree [27, 28], Subset TSP [23], or Max Leaf Outbranching [15]. In all these cases, the algorithms were technically very involved and depended heavily on the combinatorics of the problem at hand. A more systematic approach is offered by the work of Dorn et al. [15] and Tazari [32, 33], who obtained “almost” subexponential algorithm for Directed Longest Path on planar, and more generally, apex-minor-free graphs. More precisely, they proved that for any ε > 0, there is δ such that the Directed Longest Path problem is solvable in time O((1 + ε)k · nδ ) on planar directed graphs, and more generally, on directed graphs whose underlying undirected graph excludes a fixed apex graph as a minor. This technique can be extended to other problems that can be characterized as searching for a connected pattern in a large host graph, which suggests that some more robust methodology is hiding just beyond the frontier of our understanding. Main result. In this paper, we introduce a versatile technique for solving such problems in subexponential parameterized time, by proving the following theorem. Theorem 1. Let C be a class of graphs that exclude a fixed apex graph as a minor. Then there exists a randomized polynomial-time algorithm that, given an n-vertex graph G from C and an integer k, samples a vertex subset A ⊆ V (G) with the following properties: √ (P1) The induced subgraph G[A] has treewidth O( k log k). (P2) For every vertex subset X ⊆ V (G) with |X| ≤ k that induces a connected subgraph of G, the √ 2 O( k log k) probability that X is covered by A, that is X ⊆ A, is at least (2 · nO(1) )−1 . Here, by an apex graph we mean a graph that can be made planar by removing one vertex. Note that Theorem 1 in particular applies to planar graphs, and to graphs embeddable in a fixed surface. Applications. Similarly as in the case of bidimensionality, Theorem 1 provides a simple recipe for obtaining subexponential parameterized algorithms: check how fast the considered problem can be solved on graphs of bounded treewidth, and then combine the treewidth-based algorithm with Theorem 1. We now show how Theorem 1 can be used to obtain randomized subexponential parameterized algorithms for a variety of problems on apex-minor-free classes; for these problems, the existence of such algorithms so far was open even for planar graphs. We only list the most interesting examples to showcase possible applications. Directed and weighted paths and cycles. As mentioned earlier, the question about the existence of subexponential parameterized algorithms for Directed Longest Path and Weighted Longest Path on planar graphs was asked in [5, 15, 32, 33]. Let us observe that on a graph of treewidth1 t, both Directed Longest Path and Weighted Longest Path, as well as their different combinations, like finding a maximum or minimum weight directed path or cycle on k vertices, are solvable in time 2O(t log t) nO(1) by the standard dynamic programming; see e.g. [7, Chapter 7]. This running time can be improved to single-exponential time 2O(t) nO(1) [3, 16, 20]. 1 For Directed Longest Path we speak about the treewidth of the underlying undirected graph. 2 In order to obtain a subexponential parameterized algorithm for, say, Directed Longest Path on planar directed graphs, we do the following. Let G be the given planar directed graph, and let U (G) be its underlying undirected graph. Apply the algorithm of Theorem 1 to U (G), √ which in polynomial time samples a subset A ⊆ V (U (G)) such that G[A] has treewidth at most O( k log k), and the prob√ O( k log2 k) · nO(1) )−1 . ability that A covers some directed k-path in G, provided it exists, is at least (2 Then, we verify whether G[A] admits a directed k-path using standard dynamic programming in time √ 2O( k log k) ·nO(1) . Provided directed k-path exists in the graph, algorithm will find one with √ some √ this 2 2 O( k log k) O(1) −1 O( k log k) probability at least (2 ) . Thus, by making 2 ·n · nO(1) independent runs of the algorithm, we√ can reduce the error probability to at most 1/2. All in all, the obtained algorithm 2 runs in time 2O( k log k) · nO(1) and can only report false negatives with probability at most 1/2 Note that in order to apply the dynamic programming algorithm, we need to construct a suitable tree decomposition of G[A]. However, a variety of standard algorithms, e.g., the classic 4-approximation of Robertson and Seymour [30], can be used to construct such an approximate tree decomposition within the same asymptotic running time. Actually, a closer look into the proof of Theorem 1 reveals that the algorithm can construct, within the same running time, a tree decomposition of G[A] certifying the upper bound on its treewidth. Observe that the same approach works also for any apex-minor-free class C, and can be applied also to Weighted Longest Path and k-Cycle. We obtain the following corollary. Corollary 2. Let C be a class of graphs that exclude some fixed apex graph as a minor. Then problems: Weighted Longest Path, k-Cycle, and Directed Longest Path, are all solvable √ 2 O( k log k) in randomized time 2 · nO(1) on graphs from C. In case of Directed Longest Path, we mean that the underlying undirected graph of the input graph belongs to C. Note here that the approach presented above works in the same way for various combinations and extensions of problems in Corollary 2, like weighted, colored, or directed variants, possibly with some constraints on in- and out-degrees, etc. In essence, the only properties that we need is that the sought pattern persists in the subgraph induced by the covering set A, and that it can be efficiently found using dynamic programming on a tree decomposition. To give one more concrete example, Sau and Thilikos in [31] studied the problem of finding a connected k-edge subgraph with all vertices of degree at most some integer ∆; for ∆ = 2 this corresponds to finding a k-path or a k-cycle. For fixed ∆ they gave a subexponential algorithm on (unweighted) graphs excluding some fixed graph as a minor, and asked if the weighted version of this problem can be solved in subexponential parameterized time. Theorem 1 immediately implies that for fixed ∆, the weighted √ 2 O( k log k) · nO(1) on apex-minor-free graphs. variant of the problem is solvable in randomized time 2 Subgraph Isomorphism. Subgraph Isomorphism is a fundamental problem, where we are given two graphs: an n-vertex host graph G and a k-vertex pattern graph P . The task is to decide whether P is isomorphic to a subgraph of G. Eppstein [17] gave an algorithm solving Subgraph Isomorphism on planar graphs in time k O(k) n, which was subsequently improved by Dorn [14] to 2O(k) n. The first implication of our main result for Subgraph Isomorphism concerns the case when the maximum degree of P is bounded by a constant. Matoušek and Thomas [25] proved that if a tree decomposition of the host graph G of width t is given, and the pattern graph P is connected and of maximum degree at most some constant ∆, then deciding whether P is isomorphic to a subgraph of G can be done in time O(k t+1 n). By combining this with Theorem 1 as before, we obtain the following. Corollary 3. Let C be a class of graphs that exclude some fixed apex graph as a minor, and let ∆ be a fixed constant. Then, given a connected graph P with at most k vertices and maximum degree not exceeding ∆, and a graph G ∈ C on n vertices, it is possible to decide whether P is isomorphic √ 2 O( k log k) to a subgraph of G in randomized time 2 · nO(1) . 3 In a very recent, unpublished work, Bodlaender et al. [4] proved that Subgraph Isomorphism on planar graphs cannot be solved in time 2o(n/ log n) unless ETH fails. The lower bound of Bodlaender et al. holds for two very special cases. The first case is when the pattern graph P is a tree and has only one vertex of super-constant degree. Then second case is when P is not connected, but its maximum degree is a constant. Thus, the results of Bodlaender et al. show that both the connectivity and the bounded degree constraints on pattern P in Corollary 3 are necessary to keep the square root dependence on k in the exponent. However, a possibility of solving Subgraph Isomorphism in time 2O(k/ log k) · nO(1) , which is still parameterized subexponential, is not ruled out by the work of Bodlaender et al. Interestingly enough, Bodlaender et al. [4], also give a matching dynamic programming algorithm that can be combined with our theorem. Theorem 4 ( [4]). Let H be a fixed graph, and let us fix any ε > 0. Given a pattern graph P on at most k vertices, and an H-minor-free host graph G of treewidth at most O(k 1−ε ), it is possible to decide whether P is isomorphic to a subgraph of G in time 2O(k/ log k) · nO(1) . We remark that Bodlaender et al. [4] claim only a subexponential exact algorithm with running time 2O(n/ log n) , but the dynamic programming subroutine underlying this result actually gives the stronger algorithmic result as stated above. By combining Theorem 4 with Theorem 1 in the same way as before, we obtain the following. Corollary 5. Let C be a class of graphs that exclude some fixed apex graph as a minor. Then, given a connected graph P with at most k vertices, and a graph G ∈ C on n vertices, it is possible to decide whether P is isomorphic to a subgraph of G in randomized time 2O(k/ log k) · nO(1) . Let us stress here that the lower bounds of Bodlaender et al. [4] show that the running time given by Corollary 5 is tight: no 2o(k/ log k) · nO(1) -time algorithm can be expected under ETH. Local search. Fellows et al. [18] studied the following parameterized local search problem on apexminor-free graphs. In the LS Vertex Cover problem we are given an n-vertex graph G, a vertex cover S in G, and an integer k. The task is to decide whether G contains a vertex cover S 0 , such that |S 0 | < |S| and the Hamming distance |S4S 0 | between sets S and S 0 is at most k. In other words, for a given vertex cover, we ask if there is a smaller vertex cover which is k-close to the given one, in terms of Hamming (edit) distance. Fellows et al. [18] gave an algorithm solving LS Vertex Cover in time 2O(k) · nO(1) on planar graphs. The question whether this can be improved to subexponential parameterized time was raised in [9, 18]. The crux of the approach of Fellows et al. [18] is the following observation. If there is solution to LS Vertex Cover, then there is a solution S 0 , such that S4S 0 induces a connected subgraph in G. Since S4S 0 contains at most k vertices and is connected, √ our Theorem 1 can be used to sample a vertex subset A that induces a subgraph of treewidth O( k log k) and covers S4S 0 with high probability. Thus, by applying the same principle of independent repetition of the algorithm, we basically have search for suitable sets S \ S 0 and S 0 \ S in the subgraph of G induced by A. We should be, however, careful here: there can be edges between A and its complement, and these edges also need to be covered by S 0 , so we cannot just restrict our attention to G[A]. To handle this, we apply the following preprocessing. For every vertex v ∈ A, if v is adjacent to some vertex outside of A that is not included S, then v must be in S and needs also to remain in S 0 . Hence, we delete all such vertices from G[A], and it is easy to see that now the problem boils down to looking for feasible S \ S 0 and S 0 \√ S within the obtained induced subgraph. This can be easily done in time O(t) O(1) 2 ·n , where t ≤ O( k log k) is the treewidth of this subgraph; hence we obtain the following: Corollary 6. Let C be a class of graphs that exclude some fixed apex graph as a minor. Then LS √ 2 O( k log k) · nO(1) . Vertex Cover on graphs from C can be solved in randomized time 2 4 Steiner tree. Steiner Tree is a fundamental network design problem: for a graph G with a prescribed set of terminal vertices S, and an integer k, we ask whether there is a tree on at most k edges that spans all terminal vertices. Pilipczuk et al. [27] gave an algorithm for this problem 2/3 with running time 2O((k log k) ) · n on planar graphs and on graphs of bounded genus. With much √ more additional work, the running time was improved to 2O( k log k) · n in [28]. Again, by combining the standard dynamic programming solving Steiner Tree on graphs of treewidth t in time 2O(t log t) nO(1) , see e.g. [6], with Theorem 1, we immediately obtain the following. Corollary 7. Let C be a class of graphs that exclude some fixed apex√ graph as a minor. Then 2 Steiner Tree on graphs from C can be solved in randomized time 2O( k log k) · nO(1) . Contrary to the much more involved algorithm of Pilipczuk et al. [28], the algorithm above can equally easily handle various variants of the problem. For instance, we can look for a Steiner tree on k edges that minimizes the total weight of the edges, or we can ask for a Steiner out-branching in a directed graph, or we can put additional constraints on vertex degrees in the tree, and so on. Outline. In Section 2 we give an informal overview of the proof of Theorem 1 for the case of planar graphs. We try rather to focus on main intuitions and concepts, than describe technical details necessary for a formal reasoning. Then, in Sections 3 and 4 we recall the standard concepts and introduce auxiliary technical results. The full proof of Theorem 1 is contained in Section 5. We conclude in Section 6 by listing open problems raised by our work. The appendix outlines possible extensions to multiple components of the pattern and arbitrary proper minor-closed graph classes. 2 Overview of the proof of Theorem 1 We now give an informal overview of the proof of Theorem 1 in the case of planar graphs. In fact, the only two properties of planar graphs which are essential to the proof are (a) planar graphs are minor-closed, and (b) they have locally bounded treewidth by a linear function, that is, there exists a constant ctw such that every planar graph of radius k has treewidth at most ctw · k. In fact, for planar graphs one can take ctw = 3 [29], and as shown in [13], the graph classes satisfying both (a) and (b) are exactly graph classes excluding a fixed apex graph as a minor. However, in a planar graph we can rely on some topological intuition, making the presentation more intuitive. In the description we assume familiarity with tree decompositions; see Section 3 for a formal definition. Locally bounded treewidth of planar graphs. As a warm-up, let us revisit a proof that planar graphs have locally bounded treewidth. The considered proof yields a worse constant than ctw = 3, but it is insightful for our argumentation. Let G0 be a graph of radius k, that is, there exists a root vertex r0 such that every vertex of G0 is within distance at most k from r0 . As with most proofs showing that a graph in question has bounded treewidth, we will recursively construct a tree decomposition of bounded width. To this end, we need to carefully define the state of the recursion. We do it as follows: the recursive step aims at decomposing a subgraph G of the input graph G0 , with some chosen set of terminals T ⊆ V (G) on the outer face of G. The terminals T represent connections between G and the rest of G0 . In order to be able to glue back the obtained decompositions from the recursive step, our goal is to provide a tree decomposition of G with T contained in the root bag of the decomposition, so that later we can connect this bag to decompositions of other pieces of the graph that also contain the vertices of T . During the process, we keep the invariant that |T | ≤ 8(k + 2), allowing us to bound the width of the decomposition. Furthermore, the assumption that G0 is of bounded radius projects onto the recursive subinstances by the following invariant: every vertex of G is within distance at most k from some terminal. 5 In the recursive step, if T = V (G), |T | < 8(k + 2), or G is not connected, then we can perform some simple steps that we do not discuss here. The interesting case is when |T | = 8(k + 2). We partition T along the outer face into four parts of size 2(k + 2) each, called north, east, south, and west terminals. We compute minimum vertex cuts between the north and the the south terminals, and between the east and the west terminals. If, in any of these directions, a cut W of size strictly smaller than 2(k + 2) is found, then we can make a divide-and-conquer step: for every connected component D of G − (T ∪ W ) we recurse on the graph G[N [D]] with terminals N (D), obtaining a tree decomposition TD . Finally, we attach all the obtained tree decompositions below a fresh root node with bag T ∪ W , which is of size less than 10(k + 2). The crux is that such a separator W is always present in the graph. Indeed, otherwise there would exist 2(k + 2) disjoint paths between the north and the south terminals and 2(k + 2) disjoint paths between the east and the west terminals. Consider the region bounded by the two middle north-south and the two middle east-west paths: the vertices contained in this region are within distance larger than k from the outer face, on which all terminals lie (see Fig. 1a). This contradicts our invariant. Our recursion. In our case, we use a similar, but much more involved recursion scheme. In the recursive step, we are given a minor G of the input graph G0 , with some light terminals T li ⊆ V (G) on the outer face, and some heavy terminals T he ⊆ V (G)\T li lying anywhere in the graph. As before, the terminals represent connections to the other parts of the graph. We require that the terminals T := T li ∪T he need to be contained in the root bag of the tree decomposition that √is going to be constructed e k), in order to bound the in this recursion step. Moreover, we maintain an invariant that |T | = O( width of the decomposition. The graph G is a minor of the input graph G0 , since we often prefer to contract some edges instead of deleting them; thus we maintain some distance properties of G. In our recursion, the light terminals originate from cutting the graph in a similar fashion as in the proof that planar graphs have locally bounded treewidth, presented above. Hence, we keep the invariant that light terminals lie on the outer face. We sometimes need to cut deeply inside G. The produced terminals are heavy, but every such step corresponds to a significant progress in detecting the pattern, and hence such steps will be rare. In every such step, we artificially provide connectivity of the subinstances through the heavy terminals; this is technical and omitted in this overview. Recall that our goal is to preserve a connected k-vertex pattern from the input graph. Here, the pattern can become disconnected by recursing on subsequent separations, but such cuttings will always be along light terminals. Therefore, we define a pattern in a subinstance (G, T li , T he ) solved in the recursion as a set X ⊆ V (G) of size at most k such that every connected component of G[X] contains a light terminal. Hence, compared to the presented proof of planarity implying √ e k), locally bounded treewidth, we aim at more restricted width of the decomposition, namely O( but we can contract or delete parts of G, as long as the probability of spoiling a fixed, but unknown k-vertex pattern X remains inverse subexponential in k. Clustering. Upon deleting a vertex or an edge, some distance properties that we rely on can be broken. We need the following sampling procedure that partitions the graph into connected components of bounded radii, such that the probability of spoiling a particular pattern is small. Theorem 8. There exists a randomized polynomial-time algorithm that, given a graph G on n > 1 vertices and a positive integer k, returns a vertex subset B ⊆ V (G) with the following properties: (a) The radius of each connected component of G[B] is less than 9k 2 log n. (b) For each vertex subset X ⊆ V (G) of size at most k, the probability that X ⊆ B is at least 1 − k1 . Proof sketch. Start with H := G and iteratively, as long as V (H) 6= ∅, perform the following procedure. Pick arbitrary v ∈ V (H), and choose a radius r as follows. Start with r = 1 and 6 east west north M south (b) (a) M z z (c) (d) z z M M (e) Figure 1: Illustrations for Section 2. (a) The proof that planarity implies locally bounded treewidth. The vertices in the gray area are too far from the terminals. (b) Standard partitioning step. √ The e margin is gray and the islands are separated by dashed lines. The blue separator consists of O( k) islands and vertices of the margin. If the√blue islands are disjoint from the solution, we delete them and e k). (c) The situation if the standard partitioning step cannot obtain a balanced separator of size O( be applied: we have a component of the pattern (green) stretched between a light terminal and a vertex z inside the margin. (d) Chain of z-T li separators: a sparse separator that partitions the pattern in a balanced fashion is highlighted. (e) Contraction of a path Pi (blue) onto its public vertices (blue circles). A significant number of the vertices of the pattern become much closer to the light terminals. iteratively, given current radius r, with probability p := (2k 2 )−1 accept r, and with probability 1 − p increase it by one and continue (i.e., choose r according to the geometric distribution with success probability p). Given an accepted radius r, put all vertices within distance less than r from v into B, and delete from H all vertices within distance at most r from v. Since the procedure performs at most n steps, by union bound the probability of some radius 1 exceeding 9k 2 lg n is at most 2k . Fix a vertex x ∈ V (G). We have x ∈ / B only if at some point x ∈ V (H) and the distance between v and x in H is exactly r when the radius r gets accepted. However, in this case, if the radius r is increased, x is put into B regardless of subsequent random 7 choices. Consequently, for a fixed vertex x ∈ V (G), the probability that x ∈ / B is at most p. By 1 union bound, the probability that X 6⊆ B is at most |X|p, which is at most 2k for |X| ≤ k. Standard divide&conquer step. We would like to apply a similar divide&conquer step as in the presented proof that planar graphs have √ locally bounded treewidth. The problem is that we e can only afford a separator W of size O( k), however the radius of the graph can√be much larger. √ e k) Let us define the margin M to be the set of vertices within distance at most 2000 k lg k = O( from any light terminal. Intuitively, our case should be easy if every vertex of the pattern X is in the margin: we could then just throw away all vertices of G − M and use locally bounded treewidth, as the light terminals lie on the outer face. However, we cannot just branch (guess) whether this is the case: the information that G − M contains a vertex of the pattern is not directly useful. Instead, we make a localized analog of this guess: we identify a relatively compact set of vertices of G − M that prohibit us from making a single step of the recursion sketched above. First, we apply the clustering procedure (Theorem 8) to the graph G − M , so that we can assume that every connected component of G − M , henceforth called an island, is of radius bounded polynomially in k and lg n. Second, we construct an auxiliary graph H by contracting every island C into a single √ √ e k) vertex uC . Note that now in H every vertex is√within distance at most 2000 k lg k + 1 = O( e k). By standard arguments, we can find a balanced from a light terminal. Thus H has treewidth O( √ e k) such that every connected component of separator WH in H, that is, a separator of size O( H − WH , after lifting it back to G by reversing contractions, contains (a) at most |T |/2 terminals from G, and (b) at most |V (G) \ T |/2 non-terminal vertices of G. √ e k) The separator WH√can be similarly lifted to a separator W in G that corresponds to O( e k) islands. Now it is useful to make a guess if some vertex of an island in vertices of M and O( W (i.e., a vertex of W \ M ) belongs to the solution: if this is not the case, we can delete the whole W \ M from the graph, and apply the procedure recursively to connected components √ of G − W ; if e k) components this is the case, the area to search for such a vertex of the pattern is limited to O( of radius polynomial in k and lg n. Therefore, with some probability q we decide to assume that W \ M contains a vertex of the pattern, and with the remaining probability 1 − q we decide that this is not the case.√In the latter case, we remove W \ M from the graph, and recurse using W ∩ M , e k), as a separator; see Fig. 1b. The fact that every connected component of which has size O( √ e k). G − W contains at most |T |/2 terminals allows us to keep the invariant that |T | = O( Let us now analyze what probability q we can afford. Observe that in every subinstance solved recursively, the number of nonterminal vertices is halved. Thus, every vertex x of the pattern X is contained in G only in O(lg n) subinstances in the whole recursion tree; here we exclude the subinstances where x is a light terminal, because then its treatment is determined by the output specification of the recursive procedure. Consequently, we care about correct choices only in O(k lg n) steps of the recursion. In these steps, we want not to make a mistake during the clustering procedure (1/k failure probability) and to correctly guess that W \ M is disjoint with the pattern, provided this is actually the case (q failure probability). Thus, if we put q = 1/k, then the probability that we succeed in all O(k lg n) steps we care about is inverse-polynomial in n; this is sufficient for our needs. Island with a vertex of the pattern. We are left with the second case, where some island C ⊆ W intersects the pattern. We have q = 1/k probability of guessing correctly that this is the case, and independently we have (1 − 1/k) probability of not making a mistake √ in the clustering step. e The bound on the radii of the islands, as well as the fact that only O( k) islands are contained in W , allow us to localize this vertex of the pattern even closer. Recall that the radius of each island is bounded by 9k 2 lg n. For the rest of this overview we assume that lg n is bounded polynomially in k, and hence the radius of each island is polynomial in k. Intuitively, this is because if, say, we 8 100 100 had lg n > 100 · k 100 , then n > 2100k and with 2100k allowed factor in the running time bound √ e k) O( we can apply a variety of other algorithmic techniques. More formally, we observe that (lg n) √ e is bounded by 2O( k) · no(1) , which is sufficient to make sure that all the experiments whose success √ e k) O( probability depend on lg n succeed simultaneously with probability at least (2 · no(1) )−1 . We first guess (by sampling at random) an island C ⊆ W that contains a vertex of the pattern. Then, we pick an arbitrary vertex z ∈ C, and guess (by sampling at random) the distance d in C between z and the closest vertex of the pattern in C. By contracting all vertices within distance less than d from z, with success probability inverse-polynomial in k, we arrive at the following situation: (see Fig. 1c) we have a vertex z ∈ / M such that either z or a neighbor of z belongs to the pattern X. Chain of separators. Hence, one of the components of G[X] is stretched across the margin M , between a light terminal on the outer face and the vertex z inside the margin. Our idea now is to use this information to cut X in a balanced fashion. Note that we have already introduced an inverse-polynomial in k multiplicative factor in the success probability. Hence, to maintain the overall inverse-subexponential dependency on k in the success probability, we should aim at √ e k). a progress that will allow us to bound the number of such steps by O( li Unfortunately, it √ is not obvious how to find such a separation. It is naive to hope for a z-T e k), and a larger separator seems useless, if there is only one. However, we can separator of size O( aim at a Baker-style argument: if we find a chain of p pairwise disjoint z-T li separators C1 , C2 , . . . , Cp (see Fig. 1d), we could guess a “sparse one”, and separate along it. Since the separators are pairwise disjoint, there exists a “sparse” separator Ci with at most k/p vertices of the pattern. On the other hand, since the pattern contains a component stretched from z to a light terminal, every Ci intersects the pattern. If we omit the first and the last p/4 separators, and look for a sparse separator in between with at most 2k/p by means of guessing only 2k/p vertices of X. If all separators are of size bounded e 2/3 ) O(k polynomially in k, the optimal choice is p ∼ k 2/3 , which leads to success probability . √ inverse in 2 However, we can apply a bit smarter counting argument. Take p = 120 k lg k. Look at Cp/2 and assume that at most half of the vertices of X lie on the side of Cp/2 with separators Ci for i < p/2; the other case is symmetric. The crucial observation is the following: there exists √ an index i ≤ p/2 such that if |Ci ∩ X| = α then Ci partitions X into two parts of size at least α k/10 each. Indeed, otherwise we have that for every i ≤ p/2 it holds that 10 X |X ∩ Ci | ≥ √ · |X ∩ Cj |. k j<i √ √ S S This implies |X ∩ j≤i Cj | ≥ (1 + 10/ k)i , and |X ∩ j≤p/2 Cj | > k for p = 120 k lg k. Hence, we guess (by sampling at random) such an index i, the value of α = |X ∩ Ci |, and the set −O(α) we partition X ∩ Ci . If the size of Ci is bounded polynomially √ in k, with success probability k the pattern into two parts of size at least α k/10 each. A simple amortization argument shows √ e k) − O( that all these guessings incur only the promised 2 multiplicative factor in the overall success probability. Furthermore, as such a step creates α heavy terminals, it can be easily seen that the √ e k). total number of heavy terminals will never grow beyond O( However, the above argumentation assumes we are given such a chain of separators Ci : they are not only pairwise disjoint, but also of size polynomial in k. Let us now inspect how to find them. Duality. In the warm-up proof of planar graphs having locally bounded treewidth, the separator W is obtained from the classic Menger’s maximum flow/minimum cut duality. Here, we aim at a 9 chain of separators, but we require that their sizes are polynomial in k. It turns out that we can find such a chain by formulating a maximum flow of minimum cost problem, and extracting the separator chain in question from the optimum solution to its (LP) dual. Theorem 9. There is a polynomial-time algorithm that given a connected graph G, a pair s, t ∈ V (G) of different vertices, and positive integers p, q, outputs one of the following structures in G: (a) A chain (C1 , . . . , Cp ) of (s, t)-separators with |Cj | ≤ 2q for each j ∈ [p]. S (b) A sequence (P1 , . . . , Pq ) of (s, t)-paths with |(V (Pi ) ∩ i0 6=i V (Pi0 )) \ {s, t}| ≤ 4p for each i ∈ [q]. Proof sketch. We formulate the second outcome as a maximum flow of minimum cost problem in an auxiliary graph, where every vertex v ∈ V (G) \ {s, t} is duplicated into two copies: one of cost 0 and capacity 1, and one of cost 1 and infinite capacity. We ask for a minimum-cost flow of size 2q from s to t. If the cost of such flow is at most 2pq, the projection onto G of the cheapest q flow paths gives the second output. Otherwise, we read the desired chain of separators (C1 , . . . , Cp ) as distance layers from s in the graph with distances imposed by the solution to the dual linear program. Since all light terminals lie on the outer face, we can attach an auxiliary √ root vertex r0 adjacent to all light terminals, and apply Theorem 9 to (s, t) = (r0 , z), p = 120 k lg k and q = poly(k). If the algorithm of Theorem 9 returns a chain of separators, we proceed as described before. Thus, we are left with the second output: q = poly(k) nearly-disjoint paths from T li to z. Nearly-disjoint paths. The vertex set of every path Pi can be partitioned into public vertices Pub(Pi ), the ones used √ also by other paths, and the remaining private vertices Prv(Pi ). We have |Pub(Pi )| ≤ 4p = 480 k lg k, and the sets Prv(Pi ) are pairwise disjoint. We can assume q > k, so there exists a path Pi such that Prv(Pi ) is disjoint with the pattern X. By incurring an additional 1/k multiplicative factor in the success probability, we can guess, by sampling at random, such index i. How can we use such a path Pi ? Clearly, we can delete the private vertices of Pi , because they can be assumed not to be used by the patter X. However, we choose a different way: we contract them onto neighboring public vertices along Pi , reducing Pi to a path with vertex set Pub(Pi ). Observe that by this operation the vertex z changes its location in G: from a vertex deeply√inside G, namely not within the margin M , it is moved to a place within distance |Pub(Pi )| ≤ 480 k lg k from the light terminals, which is less than a quarter of the width of the margin; see Fig. 1e. Furthermore, by the connectivity assumptions on the pattern X, the vertex z drags along a number of vertices of X that are close to it. More precisely, √ if Q is a path in G[X] connecting z or a neighbor of z with a light terminal, then the first 500 k lg k vertices on Q are moved from √ being√within distance at least 1500 k lg k from all light terminals, to being within distance at most 1000 k lg k from some light √ terminal. Hence, if we define that a vertex x ∈ X is far if it is within distance larger than 1000 k lg k (i.e., half of the width of the margin) from all light terminals, √ and close otherwise, then by contracting the private vertices of Pi as described above, at least 500 k lg k vertices of k change their status from far to close. By a careful implementation of all separation steps, we can ensure that no close vertex of X √ e k) times. becomes far again. Consequently, we ensure that the above step can happen only O( Since the probability of succeeding in all guessings within this step is inverse-polynomial in k, this √ e incurs only a 2−O( k) multiplicative factor in the overall success probability. This finishes the overview of the proof of Theorem 1. We invite the reader to the next sections for a fully formal proof, which is moreover conducted for an arbitrary apex-minor-free class. 10 3 Preliminaries Notation. We use standard graph notation; see, e.g., [7] for a reference. All graphs considered in this paper are undirected and simple (without loops or multiple edges), unless explicitly stated. For a vertex u of a graph G, by NG (u) = {v : uv ∈ E(G)} and NG [u] = {u} ∪ NG (u) we denote the open and S closed neighborhood of u, respectively. Similarly, for a vertex subset X ⊆ V (G), by NG [X] = u∈X NG [u] and NG (X) = NG [X] \ X we denote the closed and open neighborhood of X, respectively. The subscript is dropped whenever it is clear from the context. For an undirected graph G and an edge uv ∈ E(G), by contracting uv we mean the following operation: remove u and v from the graph, and replace them with a new vertex that is adjacent to exactly those vertices that were neighbors of u or v in G. Note that this definition preserves the simplicity of the graph. By contracting v onto u we mean the operation of contracting the edge uv and renaming the obtained vertex as u. More generally, if X is a subset of vertices with G[X] being connected, and u ∈ / X is such that u has a neighbor in X, then by contracting X onto u we mean the operation of exhaustively contracting a neighbor of u in X onto u up to the point when X becomes empty. Note that due to the connectivity of G[X] such outcome will always be achieved. We say that H is a minor of G if H can be obtained from G by means of vertex deletions, edge deletions, and edge contractions. An apex graph is a graph that can be made planar by removing one of its vertices. For a positive integer k, we denote [k] = {1, . . . , k}. We denote lg x = log2 x. Notation log is used only under the O(·)-notation, where multiplicative constants are hidden anyways. We also denote exp[t] = et , where e is the base of the natural logarithm. Tree decompositions. Let G be an undirected graph. A tree decomposition T of G is a rooted tree T with a bag β(x) ⊆ V (G) associated with each its node x, which satisfies the following conditions: (T1) For each u ∈ V (G) there is some x ∈ V (T ) with u ∈ β(x). (T2) For each uv ∈ E(G) there is some x ∈ V (T ) with {u, v} ⊆ β(x). (T3) For each u ∈ V (G), the node subset {x ∈ V (T ) : u ∈ β(x)} induces a connected subtree of T . The width of a tree decomposition T is maxx∈V (T ) |β(x)| − 1, and the treewidth of G is equal to the minimum possible width of a tree decomposition of G. We assume reader’s familiarity with basic combinatorics of tree decompositions, and hence we often omit a formal verification that some constructed object is indeed a tree decomposition of some graph. The following fact about the existence of balanced separators in graphs of bounded treewidth is well-known, see e.g. [7, Lemma 7.19]. Lemma 10. Let G be a graph of treewidth at most t and let w : V (G) → R≥0 be a nonnegative weight function on vertices of G. Then there exists a subset of vertices X ⊆ V (G) of size at most t + 1 such that for every connected component C of G − X the following holds: w(V (C)) ≤ w(V (G))/2, where w(A) = P u∈A w(u) for any vertex subset A. Locally bounded treewidth and apex-minor-freeness. For the whole proof we fix a class C of graphs that satisfies the following properties: (a) C is closed under taking minors. 11 (b) C has locally bounded treewidth. That is, there exists a function f : N → N such that any connected graph G from C of radius r has treewidth bounded by f (r). Demaine and Hajiaghayi [13] proved that any class C that satisfies properties (a) and (b) is actually apex-minor-free. More precisely, there exists some apex graph H such that no graph from C admits H as a minor. However, Grohe [21] proved that every apex-minor-free class of graphs has locally bounded treewidth, where moreover the function f is linear. As observed by Demaine and Hajiaghayi [13], these fact combined yield the following. Proposition 11. If a class of graphs C satisfies properties (a) and (b), then there is a constant ctw such that for every connected graph G from C of radius r, the treewidth of G is bounded by ctw · r. The results of Grohe [21] in particular show that if C is a class of graphs that exclude a fixed apex graph as a minor, then the closure of C under taking minors satisfies properties (a) and (b), and hence also the property implied by Proposition 11. Hence, it suffices to prove Theorem 1 for any graph class C that satisfies properties (a) and (b). From now on, we fix the constant ctw yielded by Proposition 11 for the class C. To avoid corner cases, we assume without loss of generality that ctw ≥ 10. 4 Auxiliary tools In this section we introduce auxiliary technical tools that will be needed in the proof: a clustering procedure that reduces the radius of the graph, and a duality result concerning almost disjoint paths and chains of separators between a pair of vertices. 4.1 Clustering procedure The following result was stated as Theorem 8 in Section 2. Theorem 12. There exists a randomized polynomial-time algorithm that, given a graph G on n > 1 vertices and a positive integer k, returns a vertex subset B ⊆ V (G) with the following properties: (a) The radius of each connected component of G[B] is less than 9k 2 lg n. (b) For each vertex subset X ⊆ V (G) of size at most k, the probability that X ⊆ B is at least 1 − k1 . Proof. Consider the following iterative procedure which constructs a set B0 . We start with V0 = V (G). In the step i, given a set Vi−1 ⊆ V (G), we terminate the procedure if Vi−1 = ∅. Otherwise we pick an arbitrary vertex vi ∈ Vi−1 and we randomly select a radius ri according to the geometric distribution with success probability p := 2k12 . Given vi and ri , we insert BallG[Vi−1 ] (vi , ri − 1) into B0 and define Vi := Vi−1 \ BallG[Vi−1 ] (vi , ri ). That is, we delete from the graph vertices within distance at most ri from vi in G[Vi−1 ], and put the vertices within distance less than ri into B0 . Finally, at the end of the procedure, if any of the selected radii ri is larger than 9k 2 lg n, we return B = ∅, and otherwise we return B = B0 . Clearly, the procedure runs in polynomial time, as at every step at least the vertex vi is removed from Vi−1 , and hence at most n iterations are executed. For the radii of the connected components of G[B0 ], note that the fact that we insert into B0 vertices within distance less than ri from vi , but delete from Vi vertices within distance at most ri from vi , ensures that at every step i of the iteration we have NG [B0 ] ∩ Vi = ∅. Consequently, BallG[Vi−1 ] (vi , ri − 1) induces a connected component of G[B0 ], and is of radius less than ri . Since we return B = ∅ instead of B = B0 if any of the selected radii ri exceeds 9k 2 lg n, the upper bound on the radii of the connected components of G[B] follows. 12 It remains to argue that any fixed k-vertex subset X ⊆ V (G) survives in B with high probability. First, note that for fixed i we have that P(ri > 9k 2 lg n) ≤ (1 − p)9k 2 lg n ≤ e−4.5 lg n < n−3 < 1 ; 2kn here, we use the fact that p = 2k12 and the inequality 1 − x ≤ e−x . Thus, as there are at most n 1 iterations, with probability less than 2k the algorithm returns B = ∅ because of some ri exceeding 2 the limit of 9k lg n. Second, we analyze the probability that X ⊆ B0 . Let us fix some x ∈ X. The only moment where the vertex x could be deleted from the graph, but not put into B0 , is when x is within distance exactly ri from the vertex vi in an iteration i. It is now useful to think of the choice of ri in the iteration i as follows: we start with ri := 1 and then, iteratively, with probability p accept the current radius, and with probability 1 − p increase the radius ri by one and repeat. However, in the aforementioned interpretation of the geometric distribution, when ri = distG[Vi−1 ] (vi , x), with probability p the radius ri is accepted (and x is deleted but not put in B0 ), but with probability (1 − p) the radius ri is increased, and the vertex x is included in the ball BallG[Vi−1 ] (vi , ri − 1) ⊆ B0 . Consequently, the probability that a fixed vertex x ∈ X is not put into B0 is at most p. By the union bound, we infer that 1 1 the probability that X 6⊆ B0 is at most kp = 2k . Together with the 2k upper bound on the probability 2 that the maximum radius among ri exceeds 9k lg n, which results in putting B = ∅ instead of B = B0 , we have that the probability that X 6⊆ B is at most k1 . This concludes the proof. 4.2 Duality We start with a few standard definitions. Definition 13. For a graph H and its vertex u, by reach(u, H) we denote the set of vertices of H reachable from u in H. Suppose G is a connected graph, and s, t are its different vertices. An (s, t)-separator is a subset C of vertices of G such that s, t ∈ / C and t ∈ / reach(s, G − C). An (s, t)-separator is minimal if no its proper subset is also an (s, t)-separator. Definition 14. A sequence (C1 , C2 , . . . , Ck ) of minimal (s, t)-separators is called an (s, t)-separator chain if all of them are pairwise disjoint and for each 1 ≤ j < j 0 ≤ k, the following holds: Cj ⊆ reach(s, G − Cj 0 ) and Cj 0 ⊆ reach(t, G − Cj ). We now state and prove the main duality result, that is, Theorem 9 from Section 2. Theorem 15. There is a polynomial-time algorithm that given a connected graph G, a pair s, t ∈ V (G) of different vertices, and positive integers p, q, outputs one of the following structures in G: (a) An (s, t)-separator chain (C1 , . . . , Cp ) with |Cj | ≤ 2q for each j ∈ [p]. S (b) A sequence (P1 , . . . , Pq ) of (s, t)-paths with |(V (Pi ) ∩ i0 6=i V (Pi0 )) \ {s, t}| ≤ 4p for each i ∈ [q]. Proof. Our approach is as follows: we formulate the second output as a min-cost max-flow problem in an auxiliary graph H. If the cost of the computed flow is not too large, a simple averaging argument yields the desired paths Pi from the flow paths. If the cost is large, we look at the dual of the min-cost max-flow problem, expressed as a linear program, which is in fact a distance LP. Then we read the separators Ci as layers of distance from the vertex s. Let us now proceed with formal argumentation. We define a graph H as follows. Starting with H := G, we replace every vertex v ∈ V (G) \ {s, t} with two copies v0 and v1 : the copy v0 has capacity 1 and cost 0, while the copy v1 has infinite 13 capacity and cost 1. The vertex s is a source of size 2q and cost 0, and the vertex t is a sink of size 2q and cost 0. The edges of H are defined naturally: every edge uv of G gives rise to up to four edges in H, between the copies of u and the copies of v. In the graph H, we ask for a minimum-cost vertex-capacitated flow from s to t of size 2q. Clearly, such a flow exists for connected graphs G, as every vertex v1 is of infinite capacity. Since all the costs and capacities are integral or infinite, in polynomial time we can find a minimum-cost solution that 0 that carry unit flow each. Let C be the total cost of this decomposes into 2q flow paths P10 , P20 , . . . , P2q 0 flow. Every path Pi induces a walk Pi in G: whenever Pi0 traverses a vertex v0 or v1 , the path Pi traverses the corresponding vertex v. By shortcutting if necessary, we may assume that each Pi is a path. In this proof, we consider every path P from s to t (either in G or in H) as oriented from s towards t; thus, the notions of a predecessor/successor on P or the relation of lying before/after on P are well-defined. Let us define the cost of a path Pi , denoted c(Pi ), as the number of internal vertices that Pi shares with other paths. That is, [ c(Pi ) := |(V (Pi ) ∩ V (Pj )) \ {s, t}|. j6=i Observe that, due to the capacity constraints, if a vertex v ∈ / {s, t} lies on h > 1 paths Pi1 , Pi2 , . . . , Pih , then all but one of the paths Pi0j have to use the vertex v1 , inducing total cost h − 1 for the minimumcost flow. Since h − 1 ≥ h/2 for h > 1, it follows that 2q X c(Pi ) ≤ 2C. i=1 P We infer that if C ≤ 2pq, then 2q i=1 c(Pi ) ≤ 4pq. A standard averaging argument now shows that for at least q paths Pi we have c(Pi ) ≤ 4p. This yields the second desired output. Thus we are left with the case when C > 2pq. Our goal is to find a separator chain suitable as the first desired output. To this end, we formulate the discussed minimum-cost flow problem as a linear program, and we analyze its dual. The precise formulations can be found in Figures 2 and 3. In the dual formulation, the value ya − ys can be interpreted as a distance of a from s, where traveling through a vertex v1 costs 1 and traveling through a vertex v0 costs zv . The goal Pis to maximize the distance from s to t with weight 2q, while paying as little as possible in the sum v∈V (G)\{s,t} zv . Let {zv : v ∈ V (G); ya : a ∈ V (H)} be an optimum solution to the dual LP. Since the primal program is a minimum-cost flow problem with integral coefficients, in polynomial time we can find such values zv and ya that are additionally integral. Observe that the dual is invariant under adding a constant to every variable ya ; hence, we can assume ys = 0. Since traveling through a vertex v1 incurs distance 1, and v0 is a twin of v1 , the optimum solution never uses values zv greater than 1; hence, zv ∈ {0, 1} for every v ∈ V (G) \ {s, t}. On one hand, C, as the optimum value of both the primal and the dual LP, is assumed to be larger than 2pq. On the other hand, we have that zv ≥ 0 for every v ∈ V (G) \ {s, t}. We infer that 2q(yt − ys ) ≥ C > 2pq, and hence yt > p. We define for every 1 ≤ j ≤ p the set Cj = {v ∈ V (G) : zv = 1 ∧ yv0 = j}. We claim that C1 , C2 , . . . , Cp is the desired separator chain. Clearly, the sets are pairwise disjoint and do not contain neither s nor t. We now show that they form a separator chain. 14 X min X f (v1 , a) v∈V (G)\{s,t} a∈NH (v1 ) s.t. X f (a, b) − f (b, a) = 0 ∀ a ∈ V (H) \ {s, t} b∈NH (a) X f (s, a) − f (a, s) = 2q a∈NH (s) X f (t, a) − f (a, t) = −2q a∈NH (t) X f (v0 , a) ≤ 1 ∀ v ∈ V (G) \ {s, t} a∈NH (v0 ) f (a, b) ≥ 0 ∀ ab ∈ E(H) Figure 2: The minimum-cost flow problem used in the proof of Theorem 15. In the flow problem, the variables f (a, b) correspond to the amount of flow pushed from a to b along an edge ab. max 2q(yt − ys ) − X zv v∈V (G)\{s,t} s.t. yv0 ≤ ya + zv ∀ v ∈ V (G) \ {s, t}, a ∈ NH (v0 ) yv1 ≤ ya + 1 ∀ v ∈ V (G) \ {s, t}, a ∈ NH (v1 ) ys ≤ ya ∀ a ∈ NH (s) yt ≤ ya ∀ a ∈ NH (t) zv ≥ 0 ∀ v ∈ V (G) \ {s, t} Figure 3: The dual of the minimum-cost flow problem from Figure 2, used in the proof of Theorem 15. Claim 16. For each 1 ≤ j ≤ p, the set Cj is an (s, t)-separator. Proof. Consider a path P from s to t. Let P 0 be the corresponding path in H that traverses a vertex v0 whenever v ∈ V (G)\{s, t} lies on P . Since zv ∈ {0, 1} for every v ∈ V (G)\{s, t}, we have yb ≤ ya +1 for every ab ∈ E(H). As yt > p, there exists a vertex b on P 0 with yb = j; let b be the first such vertex and let a be its predecessor on P 0 . Note that b 6= s as ys = 0 and b 6= t as yt > p, hence b = v0 for some v ∈ V (G) \ {s, t}. Since b is the first vertex on P 0 with yb = j, we have ya = j − 1, and, consequently, zv = 1. Thus v ∈ Cj . Since the choice of P is arbitrary, Cj is an (s, t)-separator, as desired. y Consider a path Pi0 . By complementary slackness conditions, whenever the path Pi0 traverses an edge ab ∈ E(H) from a to b, the corresponding distance inequality of the dual LP is tight: yb = ya + 1 if b = v1 for some v, yb = ya + zv if b = v0 , and yb = ya if b = s or b = t. Thus, Pi0 is a shortest path from s to t in the graph H with vertex weights 0 for s, t, zv for every v0 and 1 for every v1 . In particular, always ya ≤ yb ≤ ya + 1 for b being a successor of a on Pi0 . 15 Claim 17. For each 1 ≤ j ≤ p and each 1 ≤ i ≤ 2q, we have that |V (Pi ) ∩ V (Cj )| = 1. Proof. We first prove that the cardinality of this intersection is at most 1. Assume b ∈ Pi0 such that b = v0 or b = v1 for some v ∈ Cj and let a be the predecessor of b on Pi0 . Since zv = 1, we have yb = ya + 1 = j, that is, b must be the first vertex on Pi0 with yb = j. Consequently, there exists at most one vertex b on Pi0 that projects to a vertex of Cj in G, which proves that |V (Pi ) ∩ V (Cj )| ≤ 1. To prove the converse inequality, we show that the projection onto G of the first vertex b on 0 Pi such that yb = j belongs to Cj . First, observe that such a vertex b exists since ys = 0, yt > p, and yb ≤ ya + 1 for every ab ∈ E(H). Let b = v0 or b = v1 for some v ∈ V (G) \ {s, t}; we claim that zv = 1, yv0 = j, and hence v ∈ Cj . If b = v0 , the claim is immediate as yb = ya +1 for the predecessor a of b on Pi0 . By contradiction, let us assume b = v1 but zv = 0. Then by replacing b = v1 with v0 on Pi0 we obtain a shorter path from s to t in H, contradiction to the fact that Pi0 is a shortest path from s to t. Thus zv = 1 and, consequently, yv0 = yv1 = ya + 1, so v belongs to Cj , as claimed. y Let us denote the unique vertex of V (Pi ) ∩ V (Cj ) as wi,j . Note that wi,j and wi0 ,j may coincide for different indices i, i0 . By the complementary slackness conditions again, if zv = 1 for a vertex v ∈ V (G) \ {s, t}, then there exists a flow path Pi0 that passes through v0 . Consequently, for every 1 ≤ j ≤ p and every v ∈ Cj , there exists a flow path Pi0 that passes through v0 . It follows that v = wi,j , and thus Cj = {wi,j : 1 ≤ i ≤ 2q}. In particular, we have |Cj | ≤ 2q for every 1 ≤ j ≤ p. Moreover, since each vertex of Cj lies on some path Pi , and is the unique vertex of V (Pi ) ∩ V (Cj ), we infer the Cj is a minimal separator. We are left with verifying the inclusion of reachability sets. Since Pi0 is a shortest path from s to t in H, we have that the vertex wi,j lies before the vertex wi,j 0 on the path Pi , whenever j < j 0 . As the separators Cj and Cj 0 respectively consist only of vertices wi,j and wi,j 0 , already the reachability within paths Pi certifies that that Cj ⊆ reach(s, G − Cj 0 ) and Cj 0 ⊆ reach(t, G − Cj ) for every 1 ≤ j < j 0 ≤ p, as requested. This concludes the proof of the theorem. 5 Proof of the main result In this section we give a formal proof of our main result, i.e. Theorem 1. Throughout the proof we assume that k ≥ max(10, 2ctw ), because otherwise the result is trivial: as k is bounded by a constant, we can just sample k vertices of the graph uniformly and independently at random, and return them as A. Hence, we may assume that k ≥ 10 and lg k ≥ ctw . 5.1 Recursive scheme and potentials Let G0 ∈ C be the input graph, and let k be the requested upper bound on the sizes of patterns X that we need to cover. The algorithm constructs the set A by means of a recursive procedure that roughly partitions the graph into smaller and smaller pieces, at each point making some random decisions. For the analysis, we fix some pattern X, that is, a subset X of vertices such that G[X] is connected and |X| ≤ k. Recall that our goal is to construct A in such a manner that the probability √ 2 O( k log k) that X is covered by A is at least the inverse of 2 · nO(1) . The steps taken by the algorithm obviously will not depend on X, but at each random step we argue about the success probability: the probability that the taken decision is compliant with the target pattern X, i.e., leads to its coverage. As usual with recursive algorithms, we need to consider a more general problem, which will be supplied with a few potential measures. Formally, an instance of the general problem is a tuple consisting of: (i) A connected graph G that is a minor of the original graph G0 ; hence in particular G ∈ C. 16 (ii) A specified vertex r ∈ V (G) called the root. (iii) Two disjoint vertex subsets T li , T he ⊆ V (G), called light terminals and heavy terminals, respectively. We require that the root vertex is a light terminal, that is, r ∈ T li . By T := T li ∪ T he we will denote the sets of terminals. (iv) A subset of non-terminals R ⊆ V (G) \ T , called ghost vertices. (v) A nonnegative integer λ, called the credit. Intuitively, terminals represent the boundary via which the currently considered piece communicates with the rest of the original graph, whereas ghost vertices represent maximal connected parts of the original graph lying outside of the currently considered piece, each contracted to one vertex. The reader can think of ghost vertices as emulation of hyperedges on their neighborhoods, or rather as emulation of the torso operation that would turn their neighborhoods into cliques. This torso operation cannot be performed directly because of the necessity of staying within class C, and therefore we resort to introducing ghost vertices. In the course of the algorithm, we shall maintain the following invariants: (a) Every light terminal is at distance at most 3 from the root. √ (b) It holds that |T | ≤ 16014ctw k lg k + λ. We say that a subset X ⊆ V (G) \ R is a pattern in instance I if (i) The root r is contained in X; (ii) Every vertex of X can be reached from r by a path that traverses only vertices of X ∪ R. √ (iii) |X| ≤ k − 10 k · λ. In particular, every pattern has at most k vertices, but we may consider only smaller patterns if the credit is positive. Note that the ghost vertices provide free connectivity for the pattern. For a graph G equipped with ghost vertices R, by GhRi we define the graph obtained from G by eliminating each ghost vertex, that is, removing it and turning its neighborhood into a clique. Note that GhRi does not necessarily belong to C. For two vertices x, y in G, we define the distance function distG (x, y) as the distance between x and y as the minimum possible number of non-ghost vertices on a path connecting x and y, minus 1. In other words, when a graph is equipped with ghost vertices, nonghost vertices have cost 1 of traversing them, whereas ghost vertices have cost 0. Note that if x and y are non-ghost vertices, then distG (x, y) is equal to the (normal) distance between x and y in GhRi. For an instance I, we define the subset FarI (X) of far vertices as follows: √ FarI (X) := {u ∈ X : distG (u, r0 ) > 1000 k lg k}. If a vertex of X is not far, it is said to be close. Obviously, by invariant (a) we have that no far vertex is a light terminal, that is, FarI (X) ∩ T li = ∅. For a pattern X, we define the following potentials. Pattern potential ΠI (X) := |X \ T li | Graph potential ΓI := |V (G) \ (T li ∪ R)| Distance potential ΦI (X) := |FarI (X)| 17 We drop the subscript I whenever the instance I is clear from the context. Our goal in the general problem is to compute a subset of non-ghost vertices A ⊆ V (G) \ R with the following properties: (1) It holds √ that T li ⊆ A, and the graph G[A] admits a tree decomposition of width at most 24022ctw k lg k, where T ∩ A is contained in the root bag. (2) For every pattern X in instance I, we require that     lg k + lg lg n 1 c2 Π(X) lg Γ √ P(X ⊆ A) ≥ exp −c1 · · (Π(X) lg Π(X) + Φ(X)) · 1 − , k k for some constants c1 , c2 , where n is the total number of vertices of the graph. (1) The constants c1 , c2 will be fixed while explaining the proof; actually, we will fix c1 = 2 and c2 = 2. For convenience, by LB(n, Π(X), Γ, Φ(X)) we denote the right-hand side of (1), regarded as a function of the potentials and the number of vertices n. Note that both X and A reside in the graph with the ghost vertices removed. The intuition is that ghost vertices do not belong to the piece of the graph that we are currently decomposing, but we cannot forget about them completely because they provide connectivity for the pattern. Applying the general problem. We now argue that an algorithm for the problem stated above implies the result claimed in Theorem 1. For this, we need the following simple claim. Claim 18. The following holds: √ 2 k lg k lg lg n √ Proof. The left hand size is equal to (lg n) √ (lg n) k lg k √ ≤2 k lg k . ≤k k lg2 k · no(1) Suppose first that n ≤ 2k . Then √ k lg k √ =2 k lg2 k . Suppose second that n > 2k . Then √ (lg n) k lg k √ ≤ (lg n) lg n lg lg n Hence in both cases we are done. √ =2 lg n·(lg lg n)2 = no(1) . y In order to obtain an algorithm as stated in Theorem 1, we can sample one vertex r of the input graph G0 , define G to be the connected component of G0 that contains r, and apply the algorithm for the general problem to the instance I0 := (G, r, {r}, ∅, ∅, 0). Fix some subset X ⊆ V (G0 ) with G0 [X] being connected and |X| ≤ k. Conditioned on the event that r ∈ X, which happens with probability at least 1/|V (G0 )|, the algorithm for the general problem returns a suitable subset A that covers X with probability lower bounded by LB(n, ΠI0 (X), ΓI0 , ΦI0 (X)), where n = |V (G)|. Observe that ΠI0 (X) ≤ k, ΦI0 (X) ≤ k and ΓI0 ≤ n. Thus, we have that  h i  √ 1 c2 ·k lg n P(X ⊆ A) ≥ exp −c1 k lg k(lg k + lg lg n) · 1 − k   √ √ 2 2 ≥ exp −c1 k lg k − c1 k lg k lg lg n − · c2 k lg n k h i √ ≥ exp −2c1 k lg2 k − o(lg n) − 2c2 lg n h √ i−1 2 = 2O( k log k) · nO(1) . 18 Here, the second inequality follows from the fact that 1 − x ≥ e−2x for x ∈ [0, 1/2], which can be checked by simple differentiation. The third inequality follows from Claim 18. Together with the 1/n probability that indeed r ∈ X, we obtain the success probability as required in Theorem 1. 5.2 Solving the general problem We now proceed to the description of the recursive procedure for the general problem. Throughout the description we fix the considered instance I0 = (G, r, T li , T he , R, 0) and denote n = |V (G)|. We assume that I0 satisfies the invariants stated in the previous section. Moreover, we will assume the following: (a) Ghost vertices are pairwise non-adjacent and no ghost vertex is adjacent to the root. √ (b) It holds that λ ≤ k/10. (c) There is at least one vertex that is neither a terminal nor a ghost vertex. (d) We restrict our attention only to patterns X for which it holds that ΠI0 (X) > 0. Note that assumption (b), together with the invariant about the number of terminals, in particular √ implies that |T | ≤ 16015ctw k lg k. We now explain how the assumptions above can be guaranteed. For (a), observe that whenever two ghost vertices are adjacent, we can contract the edge connecting them without changing the family of patterns in the instance, or any of its potentials. The same happens when a ghost vertex is adjacent to the root: we can contract it onto the root. Hence, whenever (a) is not satisfied, we can just apply these operations exhaustively, and work further on the instance obtained in this manner. √ For (b), observe that if the credit exceeds k/10 in any considered instance, then the upper bound on the sizes of considered patterns becomes negative, and the instance of the general problem may be solved trivially by returning A = T with a trivial tree decomposition consisting of one root √ node with bag equal to A. This decomposition has width at most |T | ≤ 16014ctw k lg k +λ; however, we so far do not have any concrete upper bound on λ. Nevertheless, it will be the case √ throughout the algorithm that we will apply the algorithm only to instances √ with credit at most k/5. √Hence, √ it is the case that this decomposition has width at most 16014ctw k lg k + k/5 < 24022ctw k lg k. For (c), observe that in this case we may again just output A = T with a trivial tree decomposition consisting of√one root node with bag √ equal to A. Again, this decomposition has width at most |T | ≤ 16014ctw k lg k + λ < 24022ctw k lg k. This provides the base case for our recursion. Finally, for (d), if ΠI0 (X) = 0, then every vertex of X is a light terminal, so in particular X ⊆ T li . Hence, whatever choice the algorithm does, as long as it outputs a vertex subset A that contains T li , we will have that X ⊆ A with probability 1. Therefore, we can focus only on the case when ΠI0 (X) is positive. Having these assumptions in place, let us define √ M := {u ∈ V (G) : distG (r, u) ≤ 2000 k lg k}. Obviously, M can be computed in linear time using a breadth-first search from r (here we need a trivial modification to traverse ghost vertices at cost 0). Every connected component of G−M will be called an island. Note that, by the definition of M , every vertex of an island that neighbors some vertex of M cannot be a ghost vertex. Hence, in particular every island contains some non-ghost vertex. The first step of the algorithm is to apply the clustering procedure of Theorem 12 for parameter k to all the islands. More precisely, we apply the algorithm of Theorem 12 to the graph K := (G − M )hR \ M i, that is, to G − M with all the ghost vertices eliminated. This algorithm works in randomized polynomial time, and returns a subset B 0 ⊆ V (K) with the following properties: 19 • each connected component of K[B 0 ] has radius at most 9k 2 lg n, and • with probability at least 1 − 1/k we have that X \ M is contained in B 0 . Now define B to be B 0 extended by adding all ghost vertices of R \ M that have at least one neighbor in B 0 . Then each connected component of G[B] has radius at most 9k 2 lg n, computed according to the distance measure distG that treats ghost vertices as traversed for free. Moreover, with probability at least 1 − 1/k we have that X \ M ⊆ B. We henceforth assume that this event happens, i.e., indeed X ⊆ M ∪ B, keeping in mind the multiplicative factor of 1 − 1/k in the success probability. Thus, we can restrict ourselves to the graph G0 defined as the connected component of G[M ∪ B] that contains r; indeed, since each vertex of X can be reached from r using only ghost vertices and other vertices of X, by the construction of B it must be entirely contained in this connected 0 component. Note that G[M ] is connected by √ definition, so M ⊆ V (G ). Observe that as we remove only vertices at distance larger than 2000 k lg k from r in G, every vertex of X that was close or far, remains close or far respectively. These observations are formalized in the following claim, whose proof follows as explained above. Claim 19. Consider a new instance I = (G0 , r, T li ∩ V (G0 ), T he ∩ V (G0 ), R ∩ V (G0 ), λ). Provided X ⊆ M ∪ B, X remains a pattern in I0 and FarI (X) = FarI0 (X). From now on we focus on the instance I. Observe that, by the claim above, the following holds: ΓI0 ≥ ΓI , ΠI0 (X) = ΠI (X), ΦI0 (X) = ΦI (X), so this restriction can only lower the potentials and make a better lower bound on the success probability. By slightly abusing the notation, we redefine the islands to be the connected components of G0 − M . Also, if in I there is no non-ghost, non-terminal vertex, then we conclude by outputting A = T . Hence from now on we assume that I has a non-ghost, non-terminal vertex, which in particular implies that ΓI > 0. Construct an auxiliary graph H from G0 by contracting each island C ∈ cc(G0 − M ) into a single (non-ghost) vertex uC ; let I = {uC : C ∈ cc(G0 − M )}. By ι : V (G0 ) → V (H) we denote the function that assigns to each vertex of G0 its image under the contraction; i.e., ι is identity on M and ι(V (C)) = {uC } for each island C. Obviously H is a minor of G0 , which in√turn is a subgraph √ of G, and hence H ∈ C. Moreover, each vertex of H is at distance at most 2000 k lg k + 1 ≤ 2001 k lg k from r. Hence, if we measure the distance √ in H normally, without eliminating the ghost vertices, then each vertex is at distance at most 4002 k lg k; this follows from the invariant that no two √ ghost vertices are adjacent. From Proposition 11 we infer that the treewidth of H is at most 4002ctw · k lg k. Define two weight functions w1 (u) and w2 (u) on V (H) as follows. For a vertex u ∈ / I, we put w1 (u) = 1 if u ∈ T and w1 (u) = 0 otherwise. However, if u = uC for some island C, then we put w1 (uC ) = |V (C) ∩ T |. Similarly, for u ∈ / I, we put w2 (u) = 1 if u ∈ V (G) \ (T li ∪ R) and w2 (u) = 0 otherwise, whereas for u = uC , we put w2 (u) = |V (C) \ (T li ∪ R)|. In other words, w1 (u) and w2 (u) are characteristic functions of T and of V (G) \ (T li ∪ R), where all vertices contained in one island C contribute to the weight of the collapsed vertex uC . By applying Lemma 10 √ to these weight functions, √ we infer that there exist sets Z1 and Z2 , each of size at most 4002ctw · k lg k + 1 ≤ 4003ctw · k lg k, such that each connected component of H − Z1 has w1 -weight at most |T |/2, and every connected component of H − Z2 has w2 (u)-weight at most |V (G) \ (T li ∪ R)|/2. We define Z := Z1 ∪ Z2 ∪ {r} ; W := ι−1 (Z). 20 √ √ Clearly, |Z| ≤ 8006ctw k lg k + 1 ≤ 8007ctw k lg k. Also, from the definition of weight functions w1 and w2 , and the properties of Z1 and Z2 , we immediately obtain the following. Claim 20. For each connected component D of G0 − W , we have |V (D) \ (T li ∪ R)| ≤ |V (G) \ (T li ∪ R)|/2 and |V (D) ∩ T | ≤ |T |/2. Let us define Wisl := ι−1 (Z ∩ I); Wnrm := ι−1 (Z \ I). In other words, (Wisl , Wnrm ) is the partition of W with respect to which vertices√are in the islands and which are not. Since ι is identity on M , we have that |Wnrm | ≤ |Z| ≤ 8007ctw k lg k and r ∈ Wnrm . The algorithm now branches into two cases based on a random decision, as follows. With probability 1 − 1/k the algorithm assumes that Wisl is disjoint with the sought pattern X; that is, no island that is contracted onto a vertex of Z intersects X. In the remaining event, which happens with probability 1/k, the algorithm assumes that X intersects some island that is contracted onto a vertex of Z. We now describe what steps are taken next in each of these cases, supposing that the algorithm made a correct assumption. The success probability analysis will be explained at the end of each case. 5.2.1 Case when Wisl is disjoint with the pattern Here, we suppose that the algorithm assumed that Wisl ∩ X = ∅, and that this assumption is correct. Hence, it is safe for the algorithm to restrict attention to the graph G00 := G0 − Wisl . More precisely, it still holds that X is contained in G00 , and that every vertex of X can be reached from r in G00 by a path traversing only vertices of X and ghost vertices; the last claim is because vertices of islands neighboring M have to be non-ghost vertices. Note that since G00 is obtained from G0 only by removing some selection of islands, it still holds that G00 is connected. We will consider the connected components of G0 − W = G00 − Wnrm ; let CC := cc(G00 − Wnrm ) be their set. Consider any D ∈ CC. By Claim 20 we have that √ |V (D) ∩ T | ≤ |T |/2 ≤ 8007ctw k lg k + λ; √ the last inequality follows from the invariant that |T | ≤ 16014ctw k lg k + λ. Now, for each component D ∈ CC, we shall construct an instance ID = (GD , r, TDli , TDhe , RD , λ). We begin by defining the graph GD , which will be constructed from G00 by a series of contractions. Define VD = NG00 [V (D)] ∪ {r}, and consider the connected components of graph G00 − VD . For each component Q ∈ cc(G00 − VD ) that contains a neighbor of the root vertex r, contract the whole Q onto r. For each component Q ∈ cc(G00 − VD ) that has no neighbor of r, contract the whole Q into a new vertex gQ and declare it a ghost vertex. We define GD to be the graph obtained from G00 by applying such contraction for each connected component Q of G00 − VD . Thus, the set RD of ghost vertices in graph GD consists of R ∩ VD , plus we add gQ to RD for each Q ∈ cc(G00 − VD ) that has no neighbor of r. Obviously, GD is still connected and contains the root vertex r. Next, we need to define terminals in the instance ID . For this, create an auxiliary graph L from G00 by contracting every connected component D ∈ CC into a single vertex wD . Obviously, L is connected and r persists in L; we treat L as a graph without any ghost vertices, so all distances 21 are computed normally. Run a breadth-first search (BFS) in L starting from r. For every vertex v ∈ Wnrm , we say that v is charged to component D ∈ CC if the parent of v in the BFS tree exists and is equal to wD . Note that a vertex of Wnrm can be not charged to any component if v = r or the parent of v in the BFS tree also belongs to Wnrm . With this definition, we are ready to define terminals in instance ID . We put: TDli = (T li ∩ VD ) ∪ {v ∈ NG00 (V (D)) \ (T li ∪ R) : v is not charged to D}; TDhe = (T he ∩ V (D)) ∪ {v ∈ NG00 (V (D)) \ (T li ∪ R) : v is charged to D}. In other words, the terminals in the instance ID comprise terminals that were originally in component D, plus we add also all non-ghost vertices from the boundary NG00 (V (D)) as terminals. These new terminals are partitioned into light and heavy depending on whether they are charged to D. From the definition it readily follows that TDli and TDhe are disjoint, as required. Note that each vertex v ∈ Wnrm is charged to at most one component D, so it can be declared as a heavy terminal in at most one instance ID . √ From Claim 20 and the facts that |T | ≤ 16014ctw k lg k + λ and NG00 (V (D)) ⊆ Wnrm , we immediately obtain that √ √ |TD | ≤ |T ∩ VD | + |Wnrm | ≤ |T |/2 + 8007ctw k lg k ≤ 16014ctw k lg k + λ. Hence, invariant (b) is satisfied in the instance ID . We are left with invariant (a), whose satisfaction is proved in a separate claim. Claim 21. In graph GD , every vertex from TDli is at distance at most 3 from r. Proof. Fix some v ∈ TDli . Suppose first that v ∈ T li ∩ VD . Then, by the assumption on invariant (a) holding in the instance I, v was already at distance at most 3 from r in graph√G. During the construction of G00 from G we removed only vertices at distance more than 2000ctw k lg k from r, hence v is still at distance at most 3 from r in G00 . Graph GD was obtained from G00 by means of edge contractions, which can only decrease the distances. Hence, v is at distance at most 3 from r also in GD . Consider now the remaining case when v ∈ NG00 (V (D)) \ (T li ∪ R) ⊆ Wnrm \ (T li ∪ R). If v = r then we are done, so assume otherwise. Let P be the path from r to v in the BFS tree in graph L used for the definition of charging. Since v ∈ TDli , we have that v is not charged to D, and hence the parent of v in this BFS tree is not equal to wD . As wD is a neighbor of v in L, due to v ∈ NG00 (V (D)), this implies that wD cannot lie on path P . Indeed, otherwise we could shortcut path P by using the edge wD v, which contradicts the fact that P is a shortest path from r to v in L (due to using BFS for its construction). Let P 0 be the path P lifted to graph G00 in the following manner: for every vertex wD0 ∈ V (L) visited on P , we replace the visit of this vertex by traversal of an appropriate path within the connected component D0 . Then P 0 is a path from r to v in G00 that does not traverse any vertex of D. Let v 0 be the first vertex on P 0 that belongs to NG00 (V (D)); clearly v 0 is well-defined because v ∈ NG00 (V (D)). Observe that the prefix of P from r to v 0 , excluding v 0 , gets entirely contracted onto r in the construction of GD , so v 0 is a neighbor of r in GD . Finally, consider the suffix of P from v 0 to v. Observe that both v and v 0 are neighbors of wD in the graph L, so since P is a shortest path, we infer that the distance between v 0 and v on P is at most 2. It is now easy to see that the suffix of P 0 between v and v 0 gets contracted to a path of length at most 2 in GD . Hence, we have uncovered a walk of length at most 3 from r to v in GD , which concludes the proof. y 22 Thus, we constructed an instance ID = (GD , r, TDli , TDhe , RD , λ) for every connected component D ∈ CC, and we verified that this instance satisfies the requested invariants. Note that the constructed instances have the same credit λ as the original one. Finally, for the pattern X, we can define its projection XD onto instance ID by simply putting: XD := X ∩ VD . Observe that each vertex of XD can be reached from r in GD by a path that traverses only vertices from XD ∪ RD . This is because it could be reached from r in G00 by some path traversing only X ∪ R, and parts of this path lying outside of VD has been either contracted to edges, or replaced by ghost vertices during the construction of GD from G00 . Therefore, XD is a pattern in ID . We now verify how the potentials behave in instances ID . Claim 22. The following holds: ΠI (X) ≥ X ΠID (XD ) (2) ΓID (3) ΓI /2 ≥ ΓID for each D ∈ CC X ΦI (X) ≥ ΦID (XD ) (4) D∈CC ΓI ≥ X D∈CC (5) D∈CC Proof. Take any vertex u ∈ V (G) \ (T li ∪ R). By the construction of instances ID , and in particular the mentioned fact that each vertex of Wnrm can be declared a heavy terminal in at most one instance ID , it readily follows that there is at most one instance ID for which u ∈ V (GD ) \ (TDli ∪ RD ). From this observation we immediately obtain statements (2) and (3). Statement (5) follows similarly, but one needs to additionally observe the following: GD is obtained from G00 by means of edge contractions that can only decrease the distances from r, so if a vertex u ∈ XD is far in the instance ID , then it was also far in the original instance I. Finally, (4) follows directly from Claim 20. y Thus, Claim 22(4) certifies that the graph potential drops significantly in each new instance. Intuitively, this drop will be responsible for amortizing the (1 − 1/k)2 multiplicative factor in the success probability incurred by the preliminary clustering step, and by the random choice of the assumption on the considered case. Note that, by the assumption that there is at least one non-ghost, non-terminal vertex, we have that ΓI > 0. Hence, by Claim 22, statement (4) in particular, each of the instances ID will have strictly fewer vertices than I. Therefore, we may apply the algorithm recursively to each instance ID . This yields subsets of vertices {AD : D ∈ cc(G00 − Wnrm )} with the following properties: √ • AD ⊇ TDli and GD [AD ] admits a tree decomposition TD of width at most 24022ctw k lg k with TD ∩ AD contained in the root bag. • The probability that XD ⊆ AD is at least LB(nD , ΠID (XD ), ΓID , ΦID (XD )), where nD = |V (GD )|. Let us define now the set A. • First, for every D ∈ CC, we put V (D) ∩ AD into A. 23 • Second, for every v ∈ Wnrm \ R, we put v into A if for every D ∈ CC such that v ∈ NG00 (V (D)) we have v ∈ AD . In particular, if there is no D ∈ CC for which v ∈ NG00 (V (D)), we also include v in A. Observe that for every D ∈ CC, we have A ∩ VD ⊆ AD , but not necessarily A ∩ VD = AD . To get more intuition about the above step, in particular the universal quantification in the second step, let us observe the following. Consider a vertex v ∈ Wnrm \ R. This vertex is a terminal in every instance ID for which v ∈ NG00 (V (D)); it is a heavy terminal in at most one such instance, and a light terminal in all other such instances. If v ∈ TDli , then clearly v ∈ AD ; thus, we have v∈ / A if and only if the (unique) instance ID where v ∈ TDhe exists and, furthermore, v ∈ / AD for this instance. Hence, intuitively, we allow this particular instance ID where v ∈ TDhe to exclude the vertex v from AD if it is deemed necessary; all other instances are required to keep it in the set AD . On the other hand, note that the vertex v in the instance ID where v ∈ TDhe contributes to the potential ΠID (XD ), and does not contribute to this potential in the other instances. We now formally verify that A defined S in this way has the required properties. First, note that every vertex v ∈ Wnrm \ R such that v ∈ / D∈CC NG00 (V (D)) (i.e., the universal quantification in the second step of the construction of the set A is done over an empty set of choices) is included in the set A. This, together with the fact that T li ∩ VD ⊆ TDli ⊆ AD for every D ∈ CC, shows that T li ⊆ A. Moreover, from this and the way we defined A it follows that if XD ⊆ AD for every D ∈ CC, then we have X ⊆ A, as XD = X ∩ VD by definition. We now check that G[A] indeed admits a suitable tree decomposition. √ Claim 23. The subgraph G[A] admits a tree decomposition of width at most 24022ctw k with T ∩ A contained in the root bag. Proof. Create the root node with bag A ∩ (T ∪ (Wnrm \ R)) associated with it. Next, for each D ∈ CC, restrict the decomposition TD to the vertices of A ∩ VD ; that is, remove all the vertices of AD \ (A ∩ VD ) from all the bags, thus constructing a tree decomposition TD0 of GD [A ∩ VD ]. Then, for each D ∈ CC, attach the obtained decomposition TD0 below the root node by making its root a child of the root node. Observe whenever a vertex v ∈ / R is shared between multiple instances ID , we have that v is a vertex of v ∈ Wnrm \ R that is a terminal in all of them (i.e. belongs to TD ), and moreover v ∈ A only if v ∈ AD for every instance ID where v is present. Since the root bag of each TD0 contains A ∩ TD , it is now easy to verify that in this manner we obtain a tree decomposition of G[A]. Observe now that |A ∩ (T ∪ (Wnrm \ R))| ≤ |T ∪ (Wnrm \ R)| ≤ |T | + |Wnrm | √ √ √ ≤ 16014ctw k lg k + λ + 8007ctw k lg k = 24022ctw k lg k; √ here, the last √ inequality follows from the fact that λ ≤ k/10. Since each TD has width at most 24022c √ tw k lg k, it follows that the obtained tree decomposition of G[A] has width at most 24022ctw k lg k. y Finally, we analyze the success probability. For this, we fix c2 := 2. Claim 24. Supposing X ∩Wisl = ∅, the algorithm outputs a set A with X ⊆ A with probability at least LB(n, ΠI (X), ΓI , ΦI (X)). This includes the (1−1/k) probability of success of the preliminary clustering step, and (1 − 1/k) probability that the algorithm makes the correct assumption that X ∩ Wisl = ∅. Proof. The preliminary clustering step is correct (that is, the set of removed vertices is disjoint with X) with probability at least 1 − 1/k. Then, the algorithm makes the correct assumption 24 with probability 1 − 1/k. We have already argued that if XD ⊆ AD for each D ∈ CC, then also X ⊆ A. The event XD ⊆ AD holds with probability at least LB(nD , ΠID (XD ), ΓID , ΦID (XD )), where nD ≤ n is the number of vertices in instance ID . Hence, we have   1 2 Y P(X ⊆ A) ≥ 1 − · LB(nD , ΠID (XD ), ΓID , ΦID (XD )). k (6) D∈CC From Claim 22, equations (2) and (5), and convexity of function t → t lg t we infer the following:     Y lg k + lg lg n lg k + lg lg nD √ √ exp −c1 · · ΠI (X) lg ΠI (X) ≤ exp −c1 · · ΠID (XD ) lg ΠID (XD(7) ) k k D∈CC     Y lg k + lg lg n lg k + lg lg nD √ √ exp −c1 · · Φ(X) ≤ exp −c1 · · Φ(XD ) (8) k k D∈CC To estimate the last factor in each LB(nD , ΠID (XD ), ΓID , ΦID (XD )), we use Claim 22, and equations (2) and (4). Recall that for any D ∈ CC, we have ΓI /2 ≥ ΓID . Hence, if we pick c2 = 2, then we obtain the following: (1 − 1/k)c2 ΠI (X) lg ΓI = (1 − 1/k)c2 ΠI (X) · (1 − 1/k)c2 ΠI (X) lg(ΓI /2) Y ≤ (1 − 1/k)c2 ΠI (X) · (1 − 1/k)c2 ΠID (XD ) lg(ΓI /2) D∈CC ≤ (1 − 1/k) c2 ΠI (X) · Y (1 − 1/k)c2 ΠID (XD ) lg ΓID D∈CC Recall that we assumed that ΠI (X) = ΠI0 > 0. Therefore, this yields: Y (1 − 1/k)c2 ΠI (X) lg ΓI ≤ (1 − 1/k)2 · (1 − 1/k)c2 ΠID (XD ) lg ΓID . (9) D∈CC Putting (6), (7), (8), and (9) together, we obtain that Y P(X ⊆ A) ≥ (1 − 1/k)2 · LB(nD , ΠID (XD ), ΓID , ΦID (XD )) D∈CC  lg k + lg lg n √ · (ΠI (X) lg ΠI (X) + ΦI (X)) · (1 − 1/k)c2 ΠI (X) lg ΓI ≥ exp −c1 · k = LB(ΠI (X), ΓI , ΦI (X)).  This concludes the proof. y Claims 23 and 24 verify that the behaviour of the algorithm is as required, provided X ∩ Wisl = ∅. 5.2.2 Case when Wisl intersects the pattern We are left with describing the steps taken by the algorithm after taking the assumption that Wisl intersects the pattern X; recall that the algorithm takes this decision with probability k1 . Henceforth, we assume that this assumption is correct. 25 First, the algorithm chooses uniformly at random a vertex uC of Z ∩ I, and it assumes that X intersects √ the island C = ι−1 (uC ). Since at least one such island is intersected by X and |Z√ ∩I| ≤ |Z| ≤ 8007ctw k lg k, the algorithm makes a correct choice with probability at least (8007ctw k lg k)−1 ≥ k −7 . Keeping this success probability in mind, from now on we assume that the choice was indeed correct; thus, the algorithm knows one island C about which it can assume that V (C) ∩ X 6= ∅. Recall that, due to the preliminary clustering step, island C has radius bounded by 9k 2 lg n, where the radius is counted w.r.t. the distance measure that regards ghost vertices as traversed for free. Select a non-ghost vertex z of C such that distC (u, z) ≤ 9k 2 lg n for each u ∈ V (C). Let d := inf{distC (z, u) : u ∈ V (C) ∩ X}. Since V (C) ∩ X 6= ∅, we have that 0 ≤ d ≤ dmax , where dmax = inf{distC (z, u) : u ∈ V (C)}, which is not larger than 1 + 9k 2 lg n. The algorithm now samples an integer value between 0 and dmax , and assumes henceforth that the sampled value is equal to d. This assumption holds with probability at least 1/(1 + dmax ) ≥ (10k 2 lg n)−1 . Keeping this success probability in mind, we proceed with the assumption that the algorithm knows the correct value of d. Let S := {u ∈ V (C) \ R : distC (u, z) < max(d, 1)} ∪ {u ∈ V (C) ∩ R : distC (u, z) < max(d, 1) − 1}. That is, S contains all the vertices of C that are at distance less than max(d, 1) from z, but we exclude ghost vertices at distance exactly max(d, 1)−1. From the definition it readily follows that the induced subgraph C[S] contains z and is connected. Construct graph G00 (this graph is different than G00 considered in the previous section) by contracting the whole subgraph G0 [S] onto z; the contracted vertex of G00 will be also denoted as z. Note that if d ≤ 1, then in fact no contraction has been made and G00 = G0 . Observe that, provided the sampled value of d is correct, the set S is disjoint with X (unless d = 0 when S = {z} and no contraction is made). Thus X ⊆ V (G00 ). Moreover, observe that each vertex of X can be still reached from r in G00 by a path that uses only ghost vertices and vertices of X. Indeed, by the definition of S, if S is disjoint with X, then S also does not contain any ghost vertex traversed on the aforementioned path. Hence, X can be still regarded as a pattern in G00 , where the ghost vertices in G00 are inherited from G0 . On the other hand, by the definition of d it follows that some vertex of X is at distance at most 1 from z in G00 (so either z, in case d = 0 and z ∈ X, or a neighbor of z, or a neighbor via one ghost vertex). Also, C is disjoint with M , √ observe that since 00 that is, all vertices of C are at distance larger than 2000 k lg k from r in G , we have the following √ distG00 (r, z) > 2000 k lg k. (10) Now we would like to apply the duality theorem, i.e., Theorem 15. Consider graph L := G00 hR ∩ V (G00 )i (this is a different graph than L considered in the previous section), a pair of vertices (s, t) = (r, z) of L and the following parameters: √ and q = k. p = d120 k lg ke By applying Theorem 15 to these, in polynomial time we can compute one of the following structures: (a) An (r, z)-separator chain (C1 , . . . , Cp ) with |Cj | ≤ 2k for each j ∈ [p]. S (b) A sequence (Q1 , . . . , Qk ) of (s, t)-paths with |(V (Qi ) ∩ i0 6=i V (Qi0 )) \ {s, t}| ≤ 4p for each i ∈ [k]. The behavior of the algorithm now differs depending on which structure has been uncovered. We start with the simpler case when the algorithm of Theorem 15 yielded a sequence of paths. 26 Subcase: a sequence of radial paths. Suppose that the algorithm of Theorem 15 returned a sequence (Q1 , . . . , Qk ) of (r, z)-paths, where each path contains only at most 4p vertices that also belong to other paths, not including z and r. These are paths in graph G00 hR ∩ V (G00 )i, but we can lift then to (r, z)-paths P1 , . . . , Pk in G00 in a natural manner as follows: whenever some Qi traverses an edge obtained from eliminating some vertex g, we replace the usage of this edge by a path of length 2 traversing g. If we obtain a walk in this manner, i.e., some ghost vertex is used more than once, we shortcut the segment of the walk between the visits of this ghost vertex; thus we obtain again a simple path. All in all, we are obtain (r, z)-paths P1 , . . . , Pk in G00 with the following property: for each i ∈ [k], there can be at most 4p non-ghost vertices on Pi that are traversed by some other paths Pi0 for i0 6= i. Note that every ghost vertex √ can be used by many paths. By (10), we have that each Pi has length larger than 2000 k lg k (measured with ghost vertices contributing 0 to the length). For i ∈ [k], define [ [ V (Pi0 ))\(R ∪{r, z}) and Prv(Pi ) = (V (Pi )\ V (Pi0 ))\(R ∪{r, z}). Pub(Pi ) = (V (Pi )∩ i0 6=i i0 6=i We have that |Pub(Pi )| ≤ 4p for all i ∈ [k] and, by definition, sets Prv(Pi ) are pairwise disjoint. Pattern X has at most k vertices, out of which one is equal to r. Hence, there is at least one index i ∈ [k] for which Prv(Pi ) is disjoint with X. The algorithm samples one index i from [k] and assumes henceforth that the sampled index has the property stated above. Note that this holds with probability at least 1/k; keeping this success probability in mind, we proceed with the assumption that the algorithm made a correct choice of i. Now that Prv(Pi ) is assumed to be disjoint with the sought pattern X, we can get rid of it in the following manner. Consider the consecutive vertices of Pi , traversed in the direction from r to z. Let v0 be the last light terminal on Pi ; vertex v0 is well defined because r is a light terminal itself. Let P 0 be the suffix of Pi from v0 to z (both inclusive). Observe that since v0√ , being a light terminal, is at distance at most 3 from r,√whereas z is at distance more than 2000 k lg k from r, √ we have that P 0 has length at most 2000 k lg k − 3 ≥ 1997 k lg k (here, the distance is measured as the number of non-ghost vertices traversed, minus 1). 0 Let v0 , v1 , . . . , v` = z be the vertices of (Pub(P i )∩V (P ))∪{v0 , z} in the order of their appearance √ on P 0 . Then clearly ` ≤ |Pub(Pi )| + 1 ≤ 485 k lg k. For each j = 0, 1, . . . , ` − 1, inspect the segment of P 0 lying between vj and vj+1 . If this segment contains some ghost vertex gj , then contract it entirely onto gj ; in case there are multiple ghost vertices in the segment, select any of them as gj . Otherwise, if there are no ghost vertices within the segment, contract this whole segment onto vertex vj ; see Fig. 4 for a visualization. Observe that, by the definition of v0 , no light terminal gets contracted in this manner. Denote the obtained graph by H; this graph is equipped with a set RH = R ∩ V (H) of ghost vertices naturally inherited from G00 . Since we assume that X is disjoint with Prv(Pi ), no contracted vertex belonged to X. Hence, it can be easily seen that X is still a pattern in H. Observe that H has strictly less vertices than G for the following reason: path P 0 had at least √ 1997√ k lg k non-ghost vertices at the beginning, but after the contraction it got shortened to at most 485 k lg k non-ghost vertices. Observe also that, since v0 is at distance at most √ 3 from r as√a light terminal, we have that the distance between r and z in H is at most 3 + 485 k lg k ≤ 488 k lg k. Define a new instance I 0 = (H, r, THli , THhe , RH , λ) by giving it the same credit λ, and taking li TH = T li and THhe = T he ∩ V (H). Observe here that we use the fact that during the construction of H we did not contract any light terminal, and hence all vertices of T li persist in H. Clearly THli and THhe are disjoint, and observe that invariants (a) and (b) trivially hold for I 0 , because H was obtained 27 z z v3 v3 v2 v2 v0 v1 v1 v0 v0 ≤3 ≤3 r r Figure 4: The contraction procedure applied on the path Pi . Vertices v1 , v2 , . . . , v`−1 are depicted in yellow, and small gray rhombi depict ghost vertices. from G00 by means of edge contractions. Since no contracted vertex belongs to X, we have that ΠI 0 (X) = ΠI (X). (11) √ Observe that during the construction of H we contracted at least 1511 k lg k non-ghost vertices. This means that the total number of vertices that are not ghost or light terminals, strictly decreases, so ΓI 0 < ΓI . (12) This also implies that nH < n, where nH is the number of vertices of H. We are left with analyzing the distance potential, which is factor on which we gain in this step. More precisely, the crucial observation is that the performed contraction significantly reduces the number of far vertices. Claim 25. The following holds: FarI 0 (X) ⊆ FarI (X) and √ |FarI (X) \ FarI 0 (X)| ≥ 511 k lg k. Proof. Observe that each close vertex in I is contained in M , which remains intact in G00 . Moreover, H is obtained from G00 only by means of edge contractions, which can only decrease the distances. Hence, every vertex of X that is close in I, remains close in I 0 . It follows that FarI 0 (X) ⊆ FarI (X). For the second claim, recall that in G00 , there is some vertex z 0 that belongs to X and is at distance √ at most 1 from z (possibly z 0 = z). On the other hand, by (10) we have distG00 (r, z) > 2000 k lg k, √ so distG00 (r, z 0 ) > 1999 to z 0 whose vertices belong to X or R. Since √ k lg k Let P be a path from r √ 0 dist√ G00 (r, z ) > 1999 k lg k, P contains at least 1999 k lg k vertices of X, among which the last 511 k lg k vertices to FarI (X) √ have to belong √ √ for the following reason: their distance from 0r is larger than 1999 k lg k − 511 k lg k > 1000 k lg k, by the triangle√ inequality. However, in I we have that the distance between r and z is shortened to at most 488 k lg k, and hence all of them become close. y 28 It follows that √ ΦI 0 (X) ≤ ΦI 0 (X) − 511 k lg k. (13) Having analyzed the decrease in all the potentials, we are ready to finalize the case. Apply the algorithm recursively to the instance I 0 = (H, r, THli , THhe , RH , λ). As discussed earlier, 0 I satisfies the requested invariants and has strictly fewer vertices, so this recursive call is correctly defined. The application of the algorithm yields a subset A0 of vertices of H with the following properties: √ • A0 ⊇ THli = T li and H[A0 ] admits a tree decomposition T 0 of width at most 24022ctw k lg k with TH ∩ A0 contained in the root bag; • the probability that X ⊆ A0 is at least LB(nH , ΠI 0 (X), ΓI 0 , ΦI 0 (X)). The algorithm returns the set A := A0 ; we now verify that A has all the required properties. Clearly we already have that A = A0 ⊇ THli = T li , so let us check that G[A] admits a suitable tree decomposition. √ Claim 26. The subgraph G[A] admits a tree decomposition of width at most 24022ctw k lg k with A ∩ T contained in the root bag. Proof. We observe that decomposition T := T 0 is suitable. First, it can be easily verified that it is also a tree decomposition of G[A], not only H[A], because G[A] is a graph on the same vertex set as H[A] and every edge of G[A] is also present in H[A]. Second, from the recursive call we have that the root bag of T 0 contains A ∩ TH , but A ∩ TH = A ∩ T , because A contains only vertices that are present in H. Consequently, the root bag of T contains A ∩ T . Finally, from the recursive √ 0 call we obtain that the width of T = T is at most 24022 k lg k. y We are left with analyzing the success probability. For this, we assume c1 ≥ 1. Claim 27. Assume c1 ≥ 1. Supposing X ∩ Wisl = 6 ∅ and the subroutine of Theorem 15 returned a sequence of paths, the algorithm outputs a set A with X ⊆ A with probability at least LB(n, Π(X), Γ, Φ(X)). This includes the (1 − 1/k) probability of success of the preliminary clustering step, the 1/k probability that the algorithm makes the correct assumption that X ∩ Wisl = ∅, the k −7 probability of correctly choosing the island C that intersects the pattern, the (10k 2 lg n)−1 probability of choosing the right distance d, and the 1/k probability of choosing the right path index i. Proof. By the bound on the success probability of the recursive call, and the assumption that k ≥ 10, we have that   1 P(X ⊆ A) ≥ 1− · k −8 · (10k 2 lg n)−1 · LB(n0 , ΠI 0 (X), ΓI 0 , ΦI 0 (X)) k ≥ k −12 · (lg n)−1 · LB(n0 , ΠI 0 (X), ΓI 0 , ΦI 0 (X)). (14) By (11) and the fact that nH < n, we have:     lg k + lg lg nH lg k + lg lg n 0 0 √ √ exp −c1 · · ΠI (X) lg ΠI (X) ≤ exp −c1 · · ΠI (X) lg ΠI (X) . (15) k k By (13) and the facts that c2 ≥ 1 and nH < n, we have     lg k + lg lg n lg k + lg lg nH √ √ · ΦI (X) ≤ exp −c1 · · ΦI 0 (X) · exp −c1 · k k exp [−c1 · 511 lg k(lg k + lg lg n)]   lg k + lg lg nH √ ≤ exp −c1 · · ΦI 0 (X) · k 12 · lg n. k 29 (16) Finally, by (11) and (12) we infer that   1 c2 ΠI (X) lg ΓI 1− k   1 c2 ΠI 0 (X) lg ΓI 0 ≤ 1− . k (17) By multiplying (15), (16), and (17), and applying the obtained bound in (14), we infer that: P(X ⊆ A) ≥ k −12 · (lg n)−1 · LB(n, ΠI (X), ΓI , ΦI (X)) · k 12 · lg n = LB(n, ΠI (X), ΓI , ΦI (X)). This concludes the proof. y Claim 26 ensures that the output of the algorithm has the required properties, whereas Claim 27 yields the sought lower bound on the success probability. Subcase: nested chain of circular separators. In this case, the algorithm of √ Theorem 15 returned an (r, z)-separator chain (C1 , . . . , Cp ) in L = G00 hR ∩ V (G00 )i, where p = d120 k lg ke and |Ci | ≤ 2k for each i ∈ [p]. Obviously, by the definition of L we have that (C1 , . . . , Cp ) is also an (r, z)-separator chain in G00 , and no vertex of any Ci is a ghost vertex. Recall that this means that all separators Ci are pairwise disjoint and reach(r, G00 − Ci ) ⊆ reach(r, G00 − Cj ) whenever 1 ≤ i < j ≤ p. By invariant (a), at most 3 first√separators may include a light terminal, hence after excluding them we are left with at least d117 k lg ke separators without any light terminals. We restrict our attention to these separators. Thus, by slightly abusing√the notation, from now on we work with a (r, z)-separation chain (C1 , . . . , Cp0 ), where p0 = d117 k lg ke such that each Ci is disjoint with T li ∪ R and, in fact, T li ⊆ reach(r, G00 − C1 ). For i ∈ [p0 ], we define the following sets: Viin = reach(r, G00 − Ci ) Viout = V (G00 ) \ (Ci ∪ Viin ). and Thus, (Viin , Ci , Viout ) is a partition of V (G00 ). Without loss of generality we can assume that each separator Ci is inclusion-wise minimal, which implies that each vertex of Ci has a neighbor in Viin and a neighbor in Viout . We now prove that one of the separators Ci has the property that it splits X in a balanced way, relatively to the number of vertices of X it contains. Claim 28. There is an index i ∈ [p0 ] for which the following holds: √ 10 k · |X ∩ Ci | ≤ min(|(X ∩ Viin ) \ T li |, |(X ∩ Viout ) \ T li |). Proof. For i ∈ [p0 ], let ai = |(X ∩ Viin ) \ T li |, and bi = |(X ∩ Viout ) \ T li |, and ci = |X ∩ Ci |. Observe that since all light terminals are within reach(r, G00 −C1 ), for each i ∈ [p0 ] the following holds: X ai ≥ cj ; (18) j<i bi ≥ X cj . (19) j>i Observe that each separator Ci has to contain a vertex of X, because X contains a non-ghost vertex at distance at most 1 from z, and this vertex can be reached from r by a path that uses only ghost 30 vertices and vertices of X. We conclude that ci ≥ 1 for each i ∈ [p0 ]. Consequently, ai ≥ 1 for each i ≥ 2, and bi ≥ 1 for each i ≤ p0 − 1. For the sake of contradiction, suppose that ci > min(ai , bi ) √ 10 k for all i. (20) Obviously, the sequence (ai )i∈[p0 ] is non-decreasing and the sequence (bi )i∈[p0 ] is non-increasing, Let i0 be the smallest index such that ai0 > bi0 ;√possibly i0 = p0 + 1 if the condition ai ≤ bi is satisfied for all i ∈ [p0 ]. We claim that in fact i0 ≤√53 k lg k; suppose otherwise. By assumption (20) and the definition of i0 we have that ci > ai /(10 k) for all i < i0 , . Therefore, by combining this with (18), we obtain that 1 X ai > √ aj 10 k j<i for all i < i0 . Equivalently, X  aj > 1 + j≤i 1 √ 10 k  X · aj . j<i Since a2 ≥ 1, we infer by a trivial induction that X  aj ≥ 1 + j<i 1 √ 10 k i−2 , for all 2 ≤ i < i0 . Therefore, we conclude that   53√k lg k−3 50√k lg k 1 1 1 1 √ · 1+ √ ≥ √ · 1+ √ ai0 −1 > 10 k 10 k 10 k 10 k 1 √ · e5 lg k > k. ≥ 10 k √ This is a contradiction with |X| ≤ k. Hence, we have that indeed i0 ≤ 53 k lg k. By applying a symmetric reasoning for the last separators and numbers bi , instead of the √ first and numbers ai , we obtain that if i1 is the largest index such that ai1 < bi1 , then i1 ≥ 64 k lg k. However, this means that i0 < i1 , which is a contradiction with the fact that sequences (ai )i∈[p0 ] and (bi )i∈[p0 ] are non-increasing and non-decreasing, respectively. y √ Observe that if an index i satisfies the property given by Claim 28, then |X ∩Ci | ≤ k/10. Indeed, otherwise we would have that min(|X ∩ Viin |, |X ∩ Viout |) > k, which is a contradiction with |X| ≤ k. The algorithm performs random sampling as follows: • First, it samples an index i ∈ [p] uniformly at random, and assumes that this index i satisfies the property given by Claim 28. √ • Then, it samples an integer α between 1 and k/10, and assumes that the sampled number α is equal to |X ∩ Ci |. • Finally, the algorithm samples a subset Q ⊆ Ci of size α uniformly at random, and assumes Q to be equal to X ∩ Ci . 31 As |Ci | ≤ 2k, we observe that the assumptions stated above are correct with probability at least   −1 2k 1 10 1 5 · √ · |C | ≥ k · ≥ k −2α−5 , i p α k α where α = |X ∩ Ci |. Keeping this success probability assumption in mind, we proceed further with the supposition that the sampled objects are indeed as assumed. The algorithm now defines two subinstances Iout and Iin as follows. li , T he , R First, we define Iout = (Gout ,√r, Tout out , λ + α); note that the guessed size of Q is added out √ to the credit. Note √ that λ + α ≤ k/10 + k/10, so the instance on which we shall recurse will have credit at most k/5. Define Gout to be the graph constructed as follows: take G00 , and contract the whole subgraph induced by Viin ∪ (Ci \ Q) onto r. Observe that since G00 [Viin ] is connected by definition, and each vertex of Ci has a neighbor in Viin , the contracted subgraph is indeed connected. The ghost vertices are just inherited from the original instance: we put Rout = R ∩ V (Gout ). The li and T he are defined as follows. First, heavy terminals are inhersets of light and heavy terminals Tout out he = T he ∩(V (G ited, but we remove all heavy terminals that reside in Q: we put Tout out )\Q). Second, as li = {r} ∪ Q. Recall that T li ⊆ V in ⊆ V in , so all the light terminals we put r plus the whole set Q: Tout 1 i light terminals of the original instance, apart from r, got contracted onto r during the construction of he and T li defined in this Gout ; this is why we do not need to inherit any of them in Iout . Clearly, √ Tout out manner are disjoint. Note that we have that |T | ≤ |T |+|Q| ≤ 16014c k lg k+λ+|Q| and |Q| = α, out tw √ so we indeed have that |Tout | ≤ 16014ctw k lg k + (λ + α); that is, invariant (b) is satisfied in the new instance. Invariant (a) is also satisfied, because all new light terminals are adjacent to the root r. Finally, we define Xout = X ∩ V (Gout ). Since Gout was obtained from G00 only by contracting some vertices onto the root, it still holds that every vertex of Xout can be reached from r by a path traversing only ghost vertices and vertices of Xout . Observe also that Claim 28 implies that √ at least 10 k · α vertices of X that are not light terminals are contained in Viin . These vertices do not remain in X out , and hence: √ √ √ √ |Xout | ≤ |X| − 10 k · α ≤ k − 10 k · λ − 10 k · α = k − 10 k · (λ + α). Therefore, we conclude that Xout is a pattern in the instance Iout . By applying the algorithm recursively to the instance Iout , we obtain a subset of vertices Aout with the following properties: √ li and G • Aout ⊇ Tout out [Aout ] admits a tree decomposition Tout of width at most 24022ctw k lg k with Aout ∩ Tout contained in the root bag. • The probability that Xout ⊆ Aout is at least LB(nout , ΠIout (Xout ), ΓIout , ΦIout (Xout )), where nout is the number of vertices on Gout . Second, we define Iin = (Gin , √ r, Tinli , Tinhe√ , Rin , λ + α); again the guessed size of Q is added to the credit. Note that, again, λ + α ≤ k/10 + k/10, so the instance on which we shall recurse will have √ credit at most k/5. Graph Gin is constructed from G00 as follows. Inspect the connected components of the graph G00 − (Viin ∪ Q). For each such component D, contract it onto a new vertex gD that is declared to be a ghost vertex. That is, we define Rin to be (R ∩ Viin ) ∪ {gD : D ∈ cc(G00 − (Viin ∪ Q))}. Next, we define the terminal sets Tinli , Tinhe . Recall that T li ⊆ Viin , so all the original light terminals persist in the graph Gin . Hence, the light terminals are defined as simply inherited from the original instance: Tinli = T li . For the heavy terminals, we take all the ones inherited from the original instance, plus we add all vertices of Q explicitly: Tinhe = (T he ∩ V (Gin )) ∪ Q. Note that Tinli and Tinhe are disjoint, because there was no light terminal in Q; that is, Q ∩ T li = ∅. Again, 32 √ we have that |T√ in | ≤ |T | + |Q| ≤ 16014ctw k lg k + λ + |Q| and |Q| = α, so we indeed have that |Tin | ≤ 16014ctw k lg k + (λ + α); that is, invariant (b) is satisfied in the new instance. Invariant (a) is also satisfied, since edge contractions could only have made the light terminals closer to the root. Finally, we take Xin = X ∩ V (Gin ). We observe that each vertex of Xin can be reached from r in Gin by a path that uses only ghost vertices and vertices of Xin . Indeed, there is such a path in G00 , and its parts that lie outside of V (Gin ) must be contained in the connected components of G00 − (Viin ∪ Q), so they can be replaced by the traversal of the ghost vertices into which these connected components are collapsed. Next, from Claim 28 we infer that √ √ √ √ |X in | ≤ |X| − 10 k · α ≤ k − 10 k · λ − 10 k · α = k − 10 k · (λ + α). Hence, we conclude that X in is a pattern in Iin . Again, we apply the algorithm recursively to instance Iin , thus obtaining a subset of vertices Ain with the following properties: √ • Ain ⊇ Tinli and Gin [Ain ] admits a tree decomposition Tin of width at most 24022ctw k lg k with Ain ∩ Tin contained in the root bag. • The probability that Xin ⊆ Ain is at least LB(nin , ΠIin (Xin ), ΓIin , ΦIin (Xin )), where nin is the number of vertices on Gin . Observe that the sets of non-ghost vertices of Gout and Gin are contained in the vertex set of and hence we can treat Aout and Ain also as subsets of non-ghost vertices of G00 . Hence, let us define A := (Aout \ Q) ∪ Ain and declare that the algorithm returns A as the answer. Note that here, again as in the case of Section 5.2.1, we formally allow the instance Iin to exclude the vertices of Q from Ain , since they are heavy terminals there. We now verify that A has the required properties. First, since we have that Tinli = T li , then A ⊇ Ain ⊇ Tinli = T li , so A indeed covers all light terminals. We now check that G[A] admits a suitable tree decomposition. √ Claim 29. The subgraph G[A] admits a tree decomposition of width at most 24022ctw k lg k with A ∩ T contained in the root bag. √ Proof. Construct √the root node and associate with it the bag (T ∪Q)∩A. Since√|T | ≤ 16014ctw k lg k+ λ and λ, |Q| ≤ k/10, it follows that this bag has size at most 16015ctw k lg k. Let us restrict decomposition Tout to the vertex set A ∩ Aout ; that is, remove all vertices of Aout \ A from all bags of 0 of G Tout , thus obtaining a tree decomposition Tout out [Aout ∩ A]. We have that Ain ⊆ A, so there is no 0 as children of the need of restricting decomposition Tin . Finally, attach decompositions Tin and Tout root bag. It can be easily verified that √ in this manner we obtain a tree decomposition of G[A], and its width is clearly at most 24022ctw k lg k. Finally, the root bag contains A ∩ T by its definition. y G00 , We are left with estimating the success probability. Before we proceed with the final calculation, let us analyze each of the potentials. Note that graphs Gin and Gout intersect only at Q ∪ {r}, and each vertex of Q is a heavy terminal in Iin and a light terminal in Iout . This observation will be crucial in the forthcoming analysis. Recall also that Q, as a subset of Ci , contains no original light terminal, i.e., Q ∩ T li = ∅. Firstly, in Iout we contracted all vertices of Viin ∪ (Ci \ Q), and in Iin we contracted all vertices of out Vi ∪ (Ci \ Q). Among the vertices shared by the instances, being {r} ∪ Q, r is a light terminal in both instances, whereas the vertices of Q are heavy terminals only in Iin . From this it immediately follows that ΓI ≥ ΓIout + ΓIin . (21) 33 Observe that both Gout and Gin are constructed from G by means of edge contractions only, which can only decrease the distances from r. Hence, a vertex of X that was close in the original instance I, remains close in the instance Iin or Iout to which it belongs. The vertices of Q are adjacent to the root in Iout , and hence none of them can be a far vertex of Xout . Hence it follows that ΦI (X) ≥ ΦIout (Xout ) + ΦIin (Xin ). (22) Finally, the vertices of Q—shared among the instance—are declared light terminals in Iin , and hence the same analysis yields that ΠI (X) ≥ ΠIout (Xout ) + ΠIin (Xin ). (23) We now perform a finer analysis of the behaviour of potential Π. For this, we use Claim 28 as follows. Claim 30. The following holds: √ ΠI (X) lg ΠI (X) ≥ ΠIout (Xout ) lg ΠIout (Xout ) + ΠIin (Xin ) lg ΠIin (Xin ) + 10 k · α. (24) Proof. Observe that ΠIout (Xout ) = |(X ∩ Viout ) \ T li |; ΠIin (Xin ) = |(X ∩ Viin ) \ T li | + |X ∩ Ci |. Thus we have ΠIout (Xout ) + ΠIin (Xin ) = |(X ∩ Viout ) \ T li | + |(X ∩ Viin ) \ T li | + |X ∩ Ci | = |X \ T li | = ΠI (X). (25) Suppose first that ΠIout (Xout ) ≤ ΠIin (Xin ); the second case is symmetric. Combining this with (25) yields the following: ΠIout (Xout ) ≤ ΠI (X)/2 and ΠIin (Xin ) ≤ ΠI (X). By Claim 28, we infer that √ 10 k · α ≤ |(X ∩ Viout ) \ T li | ≤ ΠIout (Xout ). Putting all these together, we observe that: ΠI (X) lg ΠI (X) ≥ ΠIout (Xout ) lg ΠI (X) + ΠIin (Xin ) lg ΠI (X) ≥ ΠIout (Xout ) · (1 + lg ΠIout (Xout )) + ΠIin (Xin ) lg ΠIin (Xin ) ≥ ΠIout (Xout ) lg ΠIout (Xout ) + ΠIin (Xin ) lg ΠIin (Xin ) + ΠIout (Xout ) √ ≥ ΠIout (Xout ) lg ΠIout (Xout ) + ΠIin (Xin ) lg ΠIin (Xin ) + 10 k · α. This is exactly the claimed inequality. As mentioned before, the case when ΠIout (Xout ) ≥ ΠIin (Xin ) is symmetric. y Finally, we can proceed with the final success probability analysis. For this, we can take any c1 ≥ 2. Claim 31. Assume c1 ≥ 2. Supposing X ∩Wisl = 6 ∅ and the subroutine of Theorem 15 returned a separator chain, the algorithm outputs a set A with X ⊆ A with probability at least LB(n, Π(X), Γ, Φ(X)). This includes the (1−1/k) probability of success of the preliminary clustering step, the 1/k probability that the algorithm makes the correct assumption that X ∩ Wisl = ∅, the k −7 probability of correctly choosing the island C that intersects the pattern, the (102 lg n)−1 probability of choosing the right distance d, and k −2α−5 probability of correctly sampling the i, α, and set Q. 34 Proof. We denote by nin and nout the numbers of vertices in Gin and Gout , respectively; note that nin , nout ≤ n. From the probability of the success of recursive calls, we infer that   1 · k −8 · (10k 2 lg n)−1 · k −2α−5 · P(X ⊆ A) ≥ 1− k LB(nout , ΠIout (Xout ), ΓIout , ΦIout (Xout )) · LB(nin , ΠIin (Xin ), ΓIin , ΦIin (Xin )) ≥ k −2α−17 · (lg n)−1 · LB(nout , ΠIout (Xout ), ΓIout , ΦIout (Xout )) · LB(nin , ΠIin (Xin ), ΓIin , ΦIin (Xin )) (26) From (21) and (23) we infer that   1 c2 ΠI (X) lg ΓI 1− k ≤     1 c2 ΠIout (Xout ) lg ΓIout 1 c2 ΠIin (Xin ) lg ΓIin 1− · 1− . k k Similarly, from (22) and the fact that nin , nout ≤ n, we infer that:     lg k + lg lg nout lg k + lg lg n √ √ · ΦI (X) ≤ exp −c1 · · ΦIout (Xout ) · exp −c1 · k k   lg k + lg lg nin √ exp −c1 · · ΦIout (Xin ) . k (27) (28) Finally, from Claim 30 we have that     lg k + lg lg n lg k + lg lg nout √ √ exp −c1 · · ΠI (X) lg ΠI (X) ≤ exp −c1 · · ΠIout (Xout ) lg ΠIout (Xout ) · k k   lg k + lg lg nin √ · ΠIin (Xin ) lg ΠIin (Xin ) · exp −c1 · k   √ lg k + lg lg n √ exp −c1 · · 10α k . (29) k Let us analyze the last factor of the right hand side of (29), keeping in mind that c1 ≥ 2.   √ lg k + lg lg n √ exp −c1 · · 10α k = exp [−c1 · 10α · lg k − c1 · 10α · lg lg n] k ≤ k −20α · (lg n)−1 ≤ k −2α−17 · (lg n)−1 . (30) Finally, by plugging (27), (28), (29), and (30) into (26) with LB function expanded, and recognizing the expression LB(n, Π(X), Γ, Φ(X)), we obtain P(X ⊆ A) ≥ LB(n, Π(X), Γ, Φ(X)), which is exactly what we needed to prove. y Claim 29 ensures that the output of the algorithm has the required properties, whereas Claim 31 yields the sought lower bound on the success probability. 35 6 Conclusions In this work we have laid foundations for a new tool for obtaining subexponential parameterized algorithms for problems on planar graphs, and more generally on graphs that exclude a fixed apex graph as a minor. The technique is applicable to problems that can be expressed as searching for a small, connected pattern in a large host graph. Using the new approach, we designed, in a generic manner, a number of subexponential parameterized algorithms for problems for which the existence of such algorithms was open. We believe, however, that this work provides only the basics of a new methodology for the design of parameterized algorithms on planar and apex-minor-free graphs. This methodology goes beyond the paradigm of bidimensionality and is yet to be developed. An immediate question raised by our work is whether the technique can be derandomized. Note that our main result, Theorem 1, immediately yields the following combinatorial statement. Theorem 32. Let C be a class of graphs that exclude a fixed apex graph as a minor. Suppose G is an n-vertex graph from C and k is a positive integer. Then there exists a family F of subsets of vertices G, with the following properties satisfied: √ (P1) For each A ∈ F, the treewidth of G[A] is at most O( k log k). (P2) For each vertex subset X ⊆ V (G) such that G[X] is connected and |X| ≤ k, there exists some A ∈ F for which X ⊆ A. √ (P3) It holds that |F| ≤ 2O( √ k log2 k) · nO(1) . 2 Proof. Let f (n, k) ∈ 2O( k log k) ·nO(1) be the inverse of the lower bound on the success probability of the algorithm of Theorem 1. Repeat the algorithm of Theorem 1 f (n, k) · 2k ln n times, and consider the list of obtained vertex subsets as a candidate for F. Let us fix some X ⊆ V (G) such that |X| ≤ k and G[X] is connected, and consider the probability that there is some A ∈ F for which X ⊆ A. For one particular run of the algorithm of Theorem 1, this holds with probability at least f (n, k)−1 . As the runs are independent, the probability that no element of F covers X is upper bounded by  1− 1 f (n, k) f (n,k)·2k ln n ≤ e−2k ln n = n−2k . As the number of k-vertex subsets of V (G) is upper bounded by nk , we infer that the expected value of the number of sets X for which there is no element of F covering them, is upper bounded by n−k < 1. Hence, there is a run of the described experiment for which this number is zero; this run yields the desired family F. The above proof of Theorem 32 gives only a randomized algorithm constructing a family F that indeed covers all small patterns with high probability. We conjecture that the algorithm can be derandomized; is, that the family whose existence is asserted by Theorem 32 can be computed √ that 2 O( k log k) in time 2 · nO(1) . So far we are able to derandomize most of the components of the algorithm, primarily using standard constructions based on splitters and perfect hash families [26]. However, one part of the reasoning with which we still struggle is the clustering step (Theorem 12). Q1. Is it possible to construct a family with properties described in Theorem 32 in √ c deterministic time 2O( k log k) · nO(1) , for some constant c? In Appendix A we attempt to generalize our technique to the case when the pattern is disconnected, and when the class only excludes some fixed (but arbitrary) graph H as a minor. In the 36 case of disconnected patterns, we were able to prove a suitable generalization of Theorem 1, however the success probability of the algorithm depends inverse-exponentially on number of connected components of the pattern. In the case of general H-minor-free classes, we needed to assume that the pattern admits a spanning tree of constant maximum degree. So far we do not see any reason for any of these constraints to be necessary. Q2. Is it possible to prove Theorem 1 without the assumption that the subgraph induced by X has to be connected? Q3. Is it possible to prove Theorem 1 only under the assumption that all graphs from C exclude some fixed (but arbitrary) graph H as a minor? Our next question concerns local search problems in the spirit of the LS Vertex Cover problem considered in Section 1. Apart from this problem, Fellows et al. [18] designed FPT algorithms also for the local search for a number of other problems on apex-minor-free classes, including LS Dominating Set and its distance-d generalization. Here, we are given a dominating set S in a graph G from some apex-minor-free class C, and we ask whether there exists a strictly smaller dominating set S 0 that is at Hamming distance at most k from S. Again, the approach of Fellows et al. [18] is based on the observation that if there is some solution, then there is also a solution S 0 such that S4S 0 can be connected using at most k additional vertices. Thus, we need to search for a connected pattern of size 2k, instead of k, in which suitable sets S \S 0 and S 0 \S are to be found. Unfortunately, now the preprocessing step fails: vertices outside A may require to be dominated from within A, which poses additional constraints that are not visible in the graph G[A] only. Hence, we cannot just focus on the graph G[A]. Observe, however, that the whole reasoning would go through if A covered not just S4S 0 , but also its neighborhood. More generally, if the considered problem concerns domination at distance d, then we should cover the distance-d neighborhood of S4S 0 . This motivates the following question. Q4. Fix some positive constant d. Is it possible to prove a stronger version of Theorem 1, where the sampled set A is required to cover the whole distance-d neighborhood of the set X with the same asymptotic lower bound on the success probability? Finally, so far we do not know whether the connectivity condition in Theorem 5 is necessary. Q5. Is it possible to solve Subgraph Isomorphism on planar graphs in time 2O(k/ log k) ·nO(1) , even without the assumption that the pattern graph is connected? Note that a positive answer to Q2 implies a positive answer here as well, as the algorithm of Theorem 4 does not require the pattern graph to be connected. Acknowledgements. The authors thank Felix Reidl, Fernando Sánchez Villaamil, Somnath Sikdar, and Yngve Villanger for some preliminary discussions on the topic of this paper. References [1] J. Alber, H. L. Bodlaender, H. Fernau, T. Kloks, and R. Niedermeier. Fixed parameter algorithms for dominating set and related problems on planar graphs. Algorithmica, 33(4):461–493, 2002. [2] N. Alon, R. Yuster, and U. Zwick. Color-coding. J. ACM, 42(4):844–856, 1995. [3] H. L. Bodlaender, M. Cygan, S. Kratsch, and J. Nederlof. Deterministic single exponential time algorithms for connectivity problems parameterized by treewidth. Inf. Comput., 243:86–111, 2015. [4] H. L. Bodlaender, J. Nederlof, and T. van der Zanden. Private communication, see also slides at http://www.lorentzcenter.nl/lc/web/2015/701/presentations/Bodlaender.pdf. [5] G. Borradaile, P. Klein, D. Marx, and C. Mathieu. Algorithms for Optimization Problems in Planar Graphs (Dagstuhl Seminar 13421). Dagstuhl Reports, 3(10):36–57, 2014. 37 [6] M. Chimani, P. Mutzel, and B. Zey. Improved Steiner tree algorithms for bounded treewidth. J. Discrete Algorithms, 16:67–78, 2012. [7] M. Cygan, F. V. Fomin, L. Kowalik, D. Lokshtanov, D. Marx, M. Pilipczuk, M. Pilipczuk, and S. Saurabh. Parameterized Algorithms. Springer, 2015. [8] E. D. Demaine, F. V. Fomin, M. Hajiaghayi, and D. M. Thilikos. Fixed-parameter algorithms for (k, r)-center in planar graphs and map graphs. ACM Transactions on Algorithms, 1(1):33–47, 2005. [9] E. D. Demaine, F. V. Fomin, M. Hajiaghayi, and D. M. Thilikos. Bidimensional Structures: Algorithms, Combinatorics and Logic (Dagstuhl Seminar 13121). Dagstuhl Reports, 3(3):51–74, 2013. [10] E. D. Demaine, F. V. Fomin, M. T. Hajiaghayi, and D. M. Thilikos. Subexponential parameterized algorithms on bounded-genus graphs and H-minor-free graphs. J. ACM, 52(6):866–893, 2005. [11] E. D. Demaine and M. Hajiaghayi. The bidimensionality theory and its algorithmic applications. Comput. J., 51(3):292–302, 2008. [12] E. D. Demaine and M. Hajiaghayi. Linearity of grid minors in treewidth with applications through bidimensionality. Combinatorica, 28(1):19–36, 2008. [13] E. D. Demaine and M. T. Hajiaghayi. Equivalence of local treewidth and linear local treewidth and its algorithmic applications. In Proceedings of the 15th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2004, pages 840–849. SIAM, 2004. [14] F. Dorn. Planar Subgraph Isomorphism revisited. In Proceedings of the 27th International Symposium on Theoretical Aspects of Computer Science, STACS 2010, volume 5 of Leibniz International Proceedings in Informatics (LIPIcs), pages 263–274, Dagstuhl, Germany, 2010. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik. [15] F. Dorn, F. V. Fomin, D. Lokshtanov, V. Raman, and S. Saurabh. Beyond bidimensionality: Parameterized subexponential algorithms on directed graphs. Information and Computation, 233:60–70, 2013. [16] F. Dorn, F. V. Fomin, and D. M. Thilikos. Catalan structures and dynamic programming in H-minor-free graphs. J. Computer and System Sciences, 78(5):1606–1622, 2012. [17] D. Eppstein. Subgraph isomorphism in planar graphs and related problems. J. Graph Algorithms and Applications, 3:1–27, 1999. [18] M. R. Fellows, F. V. Fomin, D. Lokshtanov, F. A. Rosamond, S. Saurabh, and Y. Villanger. Local search: Is brute-force avoidable? J. Comput. Syst. Sci., 78(3):707–719, 2012. [19] F. V. Fomin, D. Lokshtanov, and S. Saurabh. Bidimensionality and geometric graphs. In Proceedings of the 22nd Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2012, pages 1563–1575. SIAM, 2012. [20] F. V. Fomin, D. Lokshtanov, and S. Saurabh. Efficient computation of representative sets with applications in parameterized and exact algorithms. In Proceedings of the 24th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2014, pages 142–151. SIAM, 2014. [21] M. Grohe. Local tree-width, excluded minors, and approximation algorithms. Combinatorica, 23(4):613–632, 2003. [22] R. Impagliazzo and R. Paturi. On the complexity of k-SAT. J. Computer and System Sciences, 62(2):367–375, 2001. [23] P. N. Klein and D. Marx. A subexponential parameterized algorithm for subset TSP on planar graphs. In Proceedings of the 24th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2014, pages 1812–1830. SIAM, 2014. [24] R. J. Lipton and R. E. Tarjan. Applications of a planar separator theorem. SIAM J. Computing, 9:615–627, 1980. [25] J. Matoušek and R. Thomas. On the complexity of finding iso- and other morphisms for partial k-trees. Discrete Mathematics, 108(1-3):343–364, 1992. [26] M. Naor, L. J. Schulman, and A. Srinivasan. Splitters and near-optimal derandomization. In Proceedings of the 36th IEEE Annual Symposium on Foundations of Computer Science, FOCS 1995, pages 182–191. IEEE, 1995. [27] M. Pilipczuk, M. Pilipczuk, P. Sankowski, and E. J. van Leeuwen. Subexponential-time parameterized algorithm for steiner tree on planar graphs. In Proceedings of the 30th International Symposium on Theoretical Aspects of Computer Science, STACS 2013, volume 20 of Leibniz International Proceedings in Informatics (LIPIcs), pages 353–364, Dagstuhl, Germany, 2013. Schloss Dagstuhl–Leibniz-Zentrum 38 fuer Informatik. [28] M. Pilipczuk, M. Pilipczuk, P. Sankowski, and E. J. van Leeuwen. Network sparsification for Steiner problems on planar and bounded-genus graphs. In Proceedings of the 55th IEEE Annual Symposium on Foundations of Computer Science, FOCS 2014, pages 276–285. IEEE Computer Society, 2014. [29] N. Robertson and P. D. Seymour. Graph minors. III. Planar tree-width. J. Combinatorial Theory Ser. B, 36:49–64, 1984. [30] N. Robertson and P. D. Seymour. Graph minors. XIII. The disjoint paths problem. J. Combinatorial Theory Ser. B, 63(1):65–110, 1995. [31] I. Sau and D. M. Thilikos. Subexponential parameterized algorithms for degree-constrained subgraph problems on planar graphs. J. Discrete Algorithms, 8(3):330–338, 2010. [32] S. Tazari. Algorithmic graph minor theory: approximation, parameterized complexity, and practical aspects. PhD thesis, Humboldt University of Berlin, 2010. [33] S. Tazari. Faster approximation schemes and parameterized algorithms on (odd-)H-minor-free graphs. Theor. Comput. Sci., 417:95–107, 2012. 39 A Extension to multiple components of the pattern and to Hminor-free graphs In this section we develop the following extension of Theorem 1 for graph classes excluding a fixed minor, at the cost of a bound on the maximum degree of the pattern. By a proper minor-closed graph class we mean a graph class that is minor-closed and does not contain all graphs. Theorem 33. Let C be a proper minor-closed graph class, and let ∆ be a fixed constant. Then there exists a randomized polynomial-time algorithm that, given an n-vertex graph G from C and an integer k, samples a vertex subset A ⊆ V (G) with the following properties: √ • The induced subgraph G[A] has treewidth O( k log k). • For every vertex subset X ⊆ V (G) with |X| ≤ k such that G[X] is connected and has a spanning tree of degree ∆, the probability that X is covered by A, that is X ⊆ A, √ maximum 2 k log k) O(1) O( ·n )−1 . is at least (2 To see why the above assumption seems necessary with our techniques, let us look at the following example. Let G be a graph that contains a universal vertex v0 (i.e., adjacent to all vertices of V (G) \ {v0 }) such that G − v0 is planar. It is easy to see that, since G − v0 is K5 -minor-free, we have that G is K6 -minor-free. Let H be a connected pattern in G: a connected subgraph on k vertices. If H contains v0 , then H − v0 is a not necessarily connected pattern (subgraph) of G − v0 . Hence, finding a connected k-vertex pattern in G boils down to finding a not necessarily connected (k − 1)-vertex pattern in G − v0 . However, if we bound the maximum degree of the pattern, the (k − 1)-vertex pattern H − v0 in the graph G − v0 has bounded number of connected components, making the situation much more similar to the connected planar (or apex-minor-free) case. We do not know how to handle arbitrary disconnected patterns (subgraphs) with our techniques. As we show √ in this section, we are able to do it in a limited fashion, namely we can handle up to roughly k/ lg k connected components without increasing the asymptotic bound in the exponential factor in the success probability. The proof of the following generalization of Theorem 1 is described in Section A.1. Theorem 34. Let C be a class of graphs that exclude a fixed apex graph as a minor. Then there exists a randomized polynomial-time algorithm that, given an n-vertex graph G from C and an integer k, samples a vertex subset A ⊆ V (G) with the following properties: √ • The induced subgraph G[A] has treewidth O( k log k). √ • For every vertex subset X ⊆ V (G) with |X| ≤ k such that G[X] has O( k/ log k) connected the probability that X is covered by A, that is X ⊆ A, is at least √ components, 2 O( k log k) O(1) −1 (2 ·n ) . After proving Theorem 34, in Section A.2 we show how to use this extension for a bounded number of connected components in order to handle connected patterns in graph classes excluding a fixed minor. To this end, we use the Robertson-Seymour decomposition theorem that provides a tree decomposition for any graph that exclude a fixed minor. Roughly speaking, in this decomposition every bag corresponds to a graph almost embeddable into a fixed surface, and every adhesion (intersection of neighboring bags) has bounded size. By a result of Grohe [21], one can delete a bounded number of vertices from an almost embeddable graph to get an apex-minor-free graph. If the pattern we are looking for is connected and has bounded degree, deleting a bounded number of vertices can split it only into a bounded number of connected components. Thus, the algorithm 40 for graph classes excluding a fixed minor boils down to an application of either Theorem 34 or a simple Baker-style argument to every bag, after turning it into an apex-minor-free graph. A.1 Extension to bounded number of components In this section we prove Theorem 34, that is, we extend Theorem 1 to handle bounded number of connected components of the pattern. We describe it as a series of modifications to the proof of Theorem 1 from Section 5. As in Section 5, in a recursive step we are given an instance I consisting of a minor G of the input graph G0 , a root r ∈ V (G), two disjoint sets of light and heavy terminals T li , T he ⊆ V (G) with r ∈ T li , a set R ⊆ V (G) \ T of ghost vertices representing connectivity in other parts of the input graph, and credit λ. We maintain the same invariants regarding terminals: every light terminal √ is within distance at most 3 from the root, and the number of terminals is bounded by 16014ctw k lg k + λ. The first significant difference is with regards to the definition of a pattern. We start with the following definition. Definition 35. Let X ⊆ V (G) \ R be a set of vertices. Two vertices x, y ∈ X are connected if they belong to the same connected component of G[X ∪ R]. A component of the set X is an equivalence class in the relation of being connected (i.e., a set of vertices from X from a connected component of G[X ∪ R] that contains at least one vertex of X). A component Y is rooted if it contains a vertex within distance at most 3 from the root, and free otherwise. For a set X ⊆ V (G) \ R in an instance I, we introduce the following component potential as the fourth potential: Component potential ΛI (X) := number of free components of X. We can now formally define a pattern. A set X ⊆ V (G) \ R is a pattern if r ∈ X and √ √ |X| ≤ k − 10 k · λ − 486 k lg k · ΛI (X). That is, we drop the assumption of the connectivity of X (possibly with help of some ghost vertices), but every free component imposes a penalty on the allowed size of the pattern. Note that every pattern contains at least one rooted component (the one containing the root r), and an arbitrary number of free components. A.1.1 Potentials Let us now proceed to the description of the potentials. Apart from introducing the component potential, we extend the set of far vertices: every vertex in a free component is far, regardless of its distance from the root r. √ FarI (X) := {u ∈ X : distG (u, r) > 1000 k lg k or u is in a free component} As before, every vertex of the pattern that is not far is called close. Intuitively, every free component of the pattern decreases the success probability of the algorithm by a factor inverse-quasipolynomial in k and lg n. Formally, we define c LB(n, Π(X), Γ, Φ(X), Λ(X)) :=    lg n lg k 2 LB(n, Π(X), Γ, Φ(X)) · exp −c3 · Λ(X) · lg k(lg k + lg lg n) + √ k 41 for some constant c3 that will be determined later in this section. Here, again, n denotes the total number of vertices of the graph, and we omit the subscript I whenever the instance is clear from the context. Our goal is to compute a subset of non-ghost vertices A ⊆ V (G) \ R with the following properties: (1) It holds √ that T li ⊆ A, and the graph G[A] admits a tree decomposition of width at most 24022ctw k lg k, where T ∩ A is contained in the root bag. (2) For every pattern X in instance I, we require that c P(X ⊆ A) ≥ LB(n, Π(X), Γ, Φ(X), Λ(X)). A.1.2 (31) Operations on the instance One of the crucial property of the algorithm of Section 5 is that it modifies the input graph in a limited fashion. Namely, every subinstance is created by means of the following operations: 1. Edge contraction. Furthermore, if one of the contracted vertices is a ghost vertex, the new vertex remains a ghost vertex or the contraction is made onto the root. 2. Other modifications such√as vertex/edge deletion/addition, but only involving vertices within distance larger than 2000 k lg k from r, and not involving vertices in the pattern X nor ghost vertices essential for the connectivity relation within the pattern (assuming that the algorithm made correct random choices). The analysis of Section 5 used the above properties to ensure that the algorithm never turns a close vertex into a far vertex, assuming that the algorithm makes correct random choices. Here, we observe that neither of the above modifications can create a new component. Furthermore, a component that is rooted remains rooted, and a vertex belonging to a rooted component remains in a rooted component. As a result, these modifications cannot turn a close vertex into a far vertex under the new definition of the far vertices, nor create a new free component. In particular, whenever we construct a pattern in a subinstance by projecting the original pattern in √ the natural way, the projection remains a pattern in the new instance. This is because the 486 k lg k · Λ(X) penalty in the upper bound on the size of the pattern does not increase. A.1.3 Solving the general problem First, note that we can make the same assumptions (a)–(d) as in Section 5. The general structure and the main steps of the algorithm are the same as in √ Section 5: we define the margin M to be the set of vertices of G within distance at most 2000 k lg k from the root r, and apply the clustering procedure to the graph (G − M )hR \ M i. Note that the clustering procedure does not use the assumption of the connectivity of the pattern. Thus, we can assume that every island — every connected component of G − M — has radius at most 9k 2 lg n (where ghost vertices are traversed for free), at the cost of a (1 − 1/k) multiplicative factor in the success probability. By slightly abusing the notation, we redefine G to be the graph obtained from the clustering procedure; this graph was named G0 in Section 5. By the same arguments as in Section 5, by locally bounded treewidth we obtain sets Z and W ⊆ V (G) with the following properties: (1) Z consists of a selection of vertices of M and islands of G − M , and we have that r ∈ Z; √ (2) |Z| ≤ 8007ctw k lg k; 42 (3) W consists of vertices of Z ∩ M , denoted Wnrm , and the vertices that belong to the islands contained in Z, denoted Wisl ; (4) every connected component of G − W contains at most |T |/2 terminals and at most |V (G) \ (T li ∪ R)|/2 vertices that are neither light terminals nor ghost vertices. As before, we randomly select a branch we pursue: with probability (1 − 1/k) we assume that the pattern is disjoint with Wisl , and with the remaining probability 1/k we assume otherwise. Thus, we have two cases: when Wisl is assumed to be disjoint with the pattern, and when we suppose that Wisl intersects the pattern. A.1.4 Case when Wisl is disjoint with the pattern The crux in this case is to observe that nothing new happens, mostly because the argumentation of Section 5.2.1 does not rely on the connectivity of the pattern. That is, we argue that the algorithm as described in Section 5 works also in our setting. Recall that in this case we first delete Wisl from G; let the obtained graph be named G00 , as in Section 5. Then recurse into instances ID created for every connected component D of G − W = G00 − Wnrm , defined as in Section 5. In the instance ID we look for pattern XD := X ∩ VD , where VD = NG00 [D] ∪ {r}. We denote by CC the set of connected components of G − W . We now need to analyze the behavior of the free components in the process of recursion. We start with the following observation that follows directly from the discussion of Section A.1.2. Claim 36. Let Y be a component of X in I, and let D ∈ CC be such that Y ∩ VD 6= ∅. Then Y ∩ VD is contained in a single component of XD in ID . Second, we observe that the rooted components cannot give rise to any new free components. Claim 37. Let Y be a rooted component of X in I, and let D ∈ CC be such that Y ∩ VD = 6 ∅. Then Y ∩ VD is contained in a rooted component of XD in ID . Proof. Let w be a vertex of Y that is within distance at most 3 from the root r. By the definition of a component, there exists a path P in G00 [Y ∪R] between w and a vertex v ∈ Y ∩VD such that no vertex of P −{v} belongs to VD , except for possibly a neighbor v 0 of v on P if v 0 is a ghost vertex in NG00 (D). Consequently, in the process of construction of ID , the path P − {v} is contracted either onto the root or onto a ghost vertex. As the distance between w and r is at most three in G (hence also in G00 ), and ghost vertices are traversed for free in our distance measure, we have that v is within distance at most 3 from the root in ID . Consequently, Y ∩ VD is contained in a rooted component of ID . y Third, we observe that a free component cannot split into multiple free components. Claim 38. Let Y be a free component of X in I. Then there exists a component D0 ∈ CC such that for every D ∈ CC such that D 6= D0 and Y ∩ VD = 6 ∅, the set Y ∩ VD is contained in a rooted component of XD in ID . Proof. We say that a component D ∈ CC is touched if Y ∩ VD = 6 ∅. We consider all paths in G00 between the root r and a vertex w ∈ NG00 [D] for a touched component D, and pick Q0 to be a shortest such path. Let w ∈ NG00 [D0 ] be the second endpoint of Q0 , where D0 is a touched component. By the minimality of Q0 , no vertex of Q0 − {w} belongs to NG00 [D] for a touched component D. Let Q1 be a shortest path between w and a vertex v ∈ Y ∩ VD0 with all internal vertices in D0 ; such a path exists by the connectivity of G00 [D0 ], and by the minimality of Q1 no vertex of Q1 − {v} belongs to Y . 43 Consider a touched component D ∈ CC different than D0 . By the definition of a component, there exists a path Q2 in G00 [Y ∪ R] between v and a vertex u ∈ Y ∩ VD such that no vertex of Q2 −{u} belongs to VD , except for possibly a neighbor u0 of u on Q2 that is a ghost vertex in NG00 (D). Observe that from the walk being the concatenation of the paths Q0 , Q1 , and Q2 , only the root r and the vertices w, u, and u0 may potentially belong to VD . Consequently, in ID , the vertex u is within distance at most 2 from the root (recall that u0 is a ghost vertex if it belongs to VD ). We infer that Y ∩ VD is contained in a rooted component of XD in ID . This finishes the proof of the claim. y Claims 36–38 justify the following. Claim 39. The following holds: ΛI (X) ≥ X ΛID (XD ). (32) D∈CC Furthermore, Claims 36–38 ensure that a close vertex of X in I cannot become a far vertex in any of the instances ID . Consequently, the potential analysis of Claim 22 holds also in our case. Using Claim 39, the following claim follows along the same lines as Claim 24, finishing the analysis of this subcase. Claim 40. Supposing X ∩ Wisl = ∅, the algorithm outputs a set A with X ⊆ A with probability c at least LB(n, ΠI (X), ΓI , ΦI (X), ΛI (X)). This includes the (1 − 1/k) probability of success of the preliminary clustering step, and (1 − 1/k) probability that the algorithm makes the correct assumption that X ∩ Wisl = ∅. A.1.5 Case when Wisl intersects the pattern Just as in Section 5, in the second case we (a) guess (by sampling at random) the island uC ∈ Z for which C intersects the pattern, (b) take z to be a non-ghost vertex of C that is at distance at most 9k 2 lg n from all vertices of C within C, (c) guess (by sampling at random) the distance d from z to the pattern within the island C, and (d) contract the vertices of C within distance less than d from z onto z. In step (d), we perform the same distinction as in Section 5 between non-ghost vertices (for which we use distance less than max(d, 1)) and ghost vertices (for which we use distance less than max(d, 1)−1). As a result, by incurring a multiplicative factor 1 · |Z|−1 · (10k 2 lg n)−1 ≥ k −11 (lg n)−1 k √ in the success probability, we can assume that we have a vertex z ∈ / R with distG (r, z) > 2000 k lg k such that there exists a vertex of the pattern within distance at most 1 from z. By slightly abusing the notation, G denotes the graph after the modifications. In this case, as before, we apply the duality theorem (Theorem 15) to the graph GhRi, pair of vertices (s, t) = (r, z), and parameters: √ p = d120 k lg ke and q = k. The further behavior of the algorithm, as well as its analysis, depends on the output of the duality theorem. Thus, we need to consider two subcases: either the duality theorem returns a family of paths or a nested chain of circular separators. 44 Subcase: a sequence of radial paths. Following the argumentation of Section 5, in this section we are working with the following objects: √ • A vertex z ∈ V (G) \ R with distG (z, r) > 2000 k lg k, such that some vertex of X is within distance at most 1 from z. • A sequence P1 , P2 , . . . , Pk of (r, z)-paths in G, such that for every i ∈ [k] the set V (Pi ) can be partitioned as as V (Pi ) = {r, z} ] (V (Pi ) ∩ R) ] Pub(Pi ) ] Prv(Pi ), √ where the sets Prv(Pi ) are pairwise disjoint and |Pub(Pi )| ≤ 480 k lg k. The success probability so far in this case is at least k −11 (lg n)−1 . As in Section 5, we randomly pick an index i ∈ [k] and assume further that X ∩ Prv(Pi ) = ∅. Such an index i exists as |X| ≤ k, the sets Prv(Pi ) are pairwise disjoint, and r ∈ X but r ∈ / Prv(Pi ) for every i. Hence, the success probability of this step is at least 1/k. We reduce Pi in the same way as in Section 5. Let v0 be the last light terminal on Pi (it exists as r is a light terminal), let P 0 be the suffix of Pi from v0 to z, and let v0 , v1 , . . . , v` = z be the vertices of (Pub(Pi ) ∩ V (P 0 )) ∪ {v0 , z} in the order of their appearance on P 0 . For every 0 ≤ j < `, we inspect the segment of P 0 between vj and vj+1 . If this segment contains some ghost vertex gj , then contract it entirely onto gj ; if there is more than one ghost vertex, choose an arbitrary one as gj . Otherwise, if there are no ghost vertices on the segment, contract the whole segment onto vertex vj . Let H be the resulting graph, and construct the instance I 0 as in√ Section 5. We have √ ` ≤ |Pub(Pi )| + 1 ≤ 485 k lg k and distG (r, v0 ) ≤ 3, thus distH (r, z) ≤ 488 k lg k. However, the proof of Claim 25 fails if the vertex of the pattern within distance at most 1 from z belongs to a free component. The crux here is that if this is the case, then we can turn this free component into close by adding {v0 , v1 , . . . , v` } to the pattern X, providing a gain in the potential Λ(X). Let X 0 := X ∪ {v0 , v1 , . . . , v` }. Claim 41. We have FarI 0 (X) ⊆ FarI (X) and ΛI 0 (X) ≤ ΛI (X). Furthermore, one of the following holds: √ • |FarI (X) \ FarI 0 (X)| ≥ 511 k lg k, or • X 0 is a pattern in I 0 and ΛI 0 (X 0 ) < ΛI (X). Proof. The first part of the claim follows directly from the discussion of Section A.1.2 and the fact H is created from G by means of edge contractions, in the same manner as in Section 5. For the second part, let v ∈ X be a vertex within distance at most 1 from z in G (possibly v = z). If v belongs to a rooted component of X in I, the analysis of Claim 25 remains valid. That is, G[X ∪ R] contains √ a path P from v to a vertex w that is within distance at most 3 from the root, and the first 511 k lg k vertices of X of this path belong to FarI (X). These vertices become close √ in H, as distH (r, v0 ) ≤ 488 k lg k. Hence, we are left with the case when v belongs to some free component Y of X in I. The crucial observation is that in H, the vertex v belongs to a rooted component of X 0 , as v0 ∈ T li (and thus is within distance at most 3 from the root) and v0 , v1 , . . . , v` is a path in H. By the discussion in Section A.1.2, no new free component is created in I 0 , while Y stops to be free and becomes part 45 of a rooted component in I 0 . Hence, ΛI 0 (X 0 ) < ΛI (X). Furthermore, |X 0 | ≤ (` + 1) + |X|   √ √ √ ≤ 486 k lg k + k − 10 k · λ − 486 k lg k · ΛI (X) √ √ ≤ k − 10 k · λ − 486 k lg k · ΛI 0 (X 0 ). Consequently, X 0 is a pattern in I 0 , and the claim is proven. y The potential gains in Claim 41 allow us to conclude with the analog of Claim 27. For its proof, we need the following estimate. Claim 42. For every x, y > 0 it holds that (x + y) lg(x + y) − x lg x ≤ y(2 + lg(x + y)). Proof. We have (x + y) lg(x + y) − x lg x = y lg(x + y) + x lg(1 + y/x) ≤ y lg(x + y) + 2y, where in the last inequality we have used the fact that lg(1 + t) ≤ 2t for every t > 0. y Claim 43. Assume c1 ≥ 1 and c3 is sufficiently large. Supposing X ∩ Wisl 6= ∅ and the subroutine of Theorem 15 returned a sequence of paths, the algorithm outputs a set A with X ⊆ A with c probability at least LB(n, Π(X), Γ, Φ(X), Λ(X)). This includes the (1 − 1/k) probability of success of the preliminary clustering step, the 1/k probability that the algorithm makes the correct assumption that X ∩ Wisl = ∅, the k −7 probability of correctly choosing the vertex z, the (10k 2 lg n)−1 probability of choosing the right distance d, and the 1/k probability of choosing the right path index i. Proof. The proof follows the same lines as the proof of Claim 27. Note that we have ΛI 0 (X) ≤ ΛI (X) and ΦI 0 (X) ≤ ΦI (X). If the first option of Claim 41 happens (i.e., the drop in the potential Φ(X)), then the analysis is the same as in Claim 27. It remains to analyze the second case. Note that here we need to focus on all potentials, as we will analyze pattern X 0 in the instance I 0 . √ 0 Since |X \ X| ≤ 486 k lg k, we have (using Claim 42 and |X 0 | ≤ k for the first inequality): √ √ ΠI 0 (X 0 ) lg ΠI 0 (X 0 ) − ΠI (X) lg ΠI (X) ≤ 486 k lg k · (2 + lg k) ≤ 972 k lg2 k √ ΦI 0 (X 0 ) − ΦI (X) ≤ 486 k lg k √ ΠI 0 (X 0 ) lg ΓI 0 − ΠI (X) lg ΓI ≤ 486 k lg k lg n ΛI 0 (X 0 ) − ΛI (X) ≤ −1. Thus, a straightforward computation shows that c 0 , ΠI 0 (X 0 ), ΓI 0 , ΦI 0 (X 0 ), ΛI 0 (X 0 ))   LB(n ≥ exp −1458 · c1 · lg2 k(lg k + lg lg n) · c LB(n, ΠI (X), ΓI , ΦI (X), ΛI (X)) √   1 c2 ·486 k lg k lg n 1− · k    lg k lg n 2 √ exp c3 lg k(lg k + lg lg n) + . k For sufficiently large c3 , namely c3 ≥ 1458 · c1 + 486 lg e · c2 + 12, the right hand side in the inequality above is at least k 12 lg n, required to offset the success probability of the random choices, similarly as in the proof of Claim 27. This finishes the proof of the claim and the analysis of this subcase. y 46 Subcase: nested chain of circular separators. Following the argumentation of Section 5, in this section we are working with the following objects: √ • A vertex z ∈ V (G) \ R with distG (z, r) > 2000 k lg k, such that some vertex of X is within at most distance 1 from z. • An (r, z)-separator chain (C1 , . . . , Cp ) in GhRi with |Cj | ≤ 2k for each j ∈ [p], where √ p = d120 k lg ke. The success probability so far in this case is at least k −11 (lg n)−1 . As in Section 5, we treat (C1 , . . . , Cp ) as √ a separator chain in G, and drop the first three separators. In this manner, we can assume p ≥ d117 k lg ke, every Ci is disjoint with T li ∪ R, and all vertices within distance at most 3 from the root, including all light terminals, are in reach(r, G−Ci ) for every i. Recall √ that in this case the algorithm of Section 5 samples an index i ∈ [p], an integer α between 1 and k/10, and a set Q ⊆ Ci of size α. The intuition is that we hope for Q = Ci ∩ X, and Ci being a sparse separator in the sense of Claim 28. In our case, we additionally allow the value α = 0 in the above sampling, and whenever there exists an index i ∈ [p] with Ci ∩ X = ∅, we require from the algorithm to sample such an index i and sample α = 0. If no such index exists, the analysis of the algorithm of Section 5 remains applicable: √ √ Claim−128 still −1 holds, and only the probability of choosing the correct α drops from ( k/10) to (1+ k/10) ; however, in both cases it is at least k −1 , and the total probability of making correct random choices is at least k −2α−5 , as in Section 5. Furthermore, we have the following with regards to the potential Λ(X): Claim 44. If for every separator Ci we have Ci ∩ X 6= ∅, and the algorithm made correct random choices, then the following holds: ΛI (X) ≥ ΛIin (Xin ) + ΛIout (Xout ). Proof. Let Y be a component of X in ID . As discussed in Section A.1.2, Y ∩ Xin is either empty or contained in a single component Yin of Xin in Iin . Similarly, Y ∩ Xout is empty or contained in a single connected component Yout of Xout in Iout . We show that if Y is a rooted component of X in I, then both Yin and Yout are rooted in their respective instances if they exist, and if Y is free, then at most one of these components is free. We first note that if Y ∩ Ci 6= ∅, then Yout exists and is rooted, as all vertices of X ∩ Ci are neighbors of the root in Iout . Assume first that Y is a rooted component of I. Then Yin is a rooted component of Iin , as Iin is created from I by edge contractions only. Furthermore, if Y ∩ Ci = ∅, then Y ∩ Xout = ∅, and otherwise as discussed above Yout is a rooted component of Iout . Now assume that Y is a free component of I. If Y ∩ Ci = ∅, then Y ∩ Xin or Y ∩ Xout is empty. Otherwise, as already discussed, Yout is a rooted component of Iout . This finishes the proof of the claim. y Thus, we are left with the case when there exists a separator Ci that is disjoint with X, and we assume that the algorithm correctly guessed such an index i and guessed α = 0. The probability of making a correct choice is at least √ 1 · (1 + k/10)−1 ≥ k −5 = k −2α−5 . p 47 Recall that for i ∈ [p] we defined the following partition V (G) = Viin ] Ci ] Viout : Viin = reach(r, G − Ci ) and Viout = V (G) \ (Ci ∪ Viin ). The intuition is as follows: since Ci ∩ X = ∅ and Ci does not contain any ghost terminal, we can independently recurse on Viin and Viout . In Viin , we need to replace the components of G − Viin with ghost vertices to keep the potentials bounded. On the other hand, Viout does not contain any vertex within distance at most 3 from the root, and thus every vertex of the pattern X in Viout is in a free component and thus far. Hence, we can freely choose a new root in this subcase: by proclaiming z the new root, we make the component containing a vertex of X within distance 1 from z close, making a gain in the Λ(X) potential. Let us now proceed with formal argumentation. The algorithm defines two subinstances Iout and Iin as follows. The instance Iin is defined in the same way as in Section 5 for α = 0 and Q = ∅. That is, we take Iin = (Gin , r, Tinli , Tinhe , Rin , λ). Graph Gin is constructed from G as follows. Inspect the connected components of the graph G − Viin . For each such component D, contract it onto a new vertex gD that is declared to be a ghost vertex. That is, we define Rin to be (R ∩ Viin ) ∪ {gD : D ∈ cc(G − Viin )}. For the terminal sets Tinli , Tinhe , recall that T li ⊆ Viin , so all the original light terminals persist in the graph Gin . Thus we take Tinli = T li . For the heavy terminals, we inherit them: Tinhe = T he ∩V (Gin ). Finally, we take Xin = X ∩ V (Gin ). By the same arguments as in Section 5, Iin is a valid instance with pattern Xin . li , T he , R Second, we define Iout = (Gout , z, Tout out , λ); that is, we take the vertex z to be the out new root in the instance Iout . Recall that Viout does not contain any light terminal. We define li = {z}, T he = T he ∩ V out , and R out Gout = G[Viout ], Tout out = R ∩ Vi . In other words, we inherit out i all terminals and ghost vertices from I, and additionally proclaim z the root and a light terminal. Since r ∈ T li , we have |T | ≥ |Tout | and, consequently, Iout is a valid instance. Finally, we take Xout = (X ∩ Viout ) ∪ {z}. Note that |Xout | ≤ |X|, as r ∈ X \ Xout . Since Ci ∩ (X ∪ R) = ∅, every component of X in I lies either entirely in Viin or entirely in Viout . Furthermore, since every vertex within distance at most 3 from the root is in Viin , every component of X lying in Viout is free, and, consequently, all vertices of Viout ∩ X are far. Let x ∈ X be a vertex within distance at most 1 from z; clearly, x ∈ Viout . The component Y of X in I containing x is free, but Y ∪ {z} is contained in a rooted component of Xout in Iout . This, together with the discussion of Section A.1.2 applied to the instance Iin , proves the following claim. Claim 45. The following holds: ΠI (X) ≥ ΠIin (Xin ) + ΠIout (Xout ) ΓI ≥ ΓIin + ΓIout ΦI (X) ≥ ΦIin (Xin ) + ΦIout (Xout ) ΛI (X) ≥ 1 + ΛIin (Xin ) + ΛIout (Xout ). In particular, the last inequality of Claim 45 shows that Xout is a pattern in Iout . We apply the algorithm recursively to instances Iin and Iout , obtaining sets Ain and Aout . We have T li = Tinli ⊆ Ain and z ∈ Aout . We take A := Ain ∪ Aout . Clearly, T li ⊆ A.√ The fact that G[A] admits a tree decomposition of width at most 24022ctw k lg k with A ∩ T in the root bag is straightforward: since Ain and Aout are separated by Ci , there are no edges between these two sets, and we can just take a root bag A ∩ T , and attach as children the decompositions of G[Ain ] and G[Aout ]. We are left with analyzing the success probability. All the necessary observations have already been made in Claim 45, so we can conclude with the following claim. 48 Claim 46. Assume c1 ≥ 2 and c3 ≥ 17. Supposing X ∩ Wisl = 6 ∅ and the subroutine of Theorem 15 returned a separator chain, the algorithm outputs a set A with X ⊆ A with probability at least c Π(X), Γ, Φ(X), Λ(X)). This includes the (1−1/k) probability of success of the preliminary clusLB(n, tering step, the 1/k probability that the algorithm makes the correct assumption that X ∩Wisl = ∅, the k −7 probability of correctly choosing the island C that intersects the pattern, the (10k 2 lg n)−1 probability of choosing the right distance d, and k −2α−5 probability of correctly sampling the i, α, and set Q. Proof. The case α > 0 has been already discussed, and is the same as in Section 5, with the help of Claim 44 to control the split of the potential Λ(X). For α = 0, Claim 45 ensures that c in , ΠI (Xin ), ΓI , ΦI (Xin ), ΛI (Xin )) · LB(n c out , ΠI (Xout ), ΓI , ΦI (Xout ), ΛI (Xout )) LB(n out out out out in in in in   2 c ≥ LB(n, ΠI (X), ΓI , ΦI (X), ΛI (X)) · exp c3 lg k(lg k + lg lg n) . Hence,   1 P(X ⊆ A) ≥ 1− · k −8 · (10k 2 lg n)−1 · k −5 · k c out , ΠI (Xout ), ΓI , ΦI (Xout ), ΛI (Xout )) · LB(n out out out out c LB(nin , ΠIin (Xin ), ΓIin , ΦIin (Xin ), ΛIin (Xin ))   c ≥ k −17 · (lg n)−1 · exp c3 lg2 k(lg k + lg lg n) · LB(n, ΠI (X), ΓI , ΦI (X), ΛI (X)) c ≥ LB(n, ΠI (X), ΓI , ΦI (X), ΛI (X)) This finishes the proof of the claim and concludes the description of the third and last subcase. A.1.6 y Wrap-up: a multiple-component version of Theorem 1 Let us now take a step back and use the developed recursive algorithm to obtain a multiple-component version of Theorem 1, namely Theorem 34, following the lines of the reasoning of Section 5.1. Assume we are given an n-vertex graph G0 from a minor closed graph class C that excludes some apex graph, and we are looking for a pattern X ⊆ V (G) of size k such that G[X] has d connected components. Similarly as in Section 5.1, we can guess (by sampling at random) one vertex x ∈ X and create an instance I of our recursive problem with G = G0 , r0 = x, T li = {x}, and T he = R = ∅. Since G[X] has d connected components, we have ΛI (X) ≤ d − 1. However, X may not be a pattern in I due to the size penalty incurred by multiple connected √ components. Instead, provided we assume that d ≤ c · k/ lg k for some constant c ≥ 1, we can recursive subproblem with the parameter k 0 := (105 · c)3 k = O(k). Then, as √ invoke the 2/3 √ αk lg(αk) ≤ α k lg k for every k ≥ 10 and α ≥ 105 , we have: √ k 0 − d · 486 · k 0 lg k 0 ≥ (105 · c)3 k − 486 · 1010 c3 k > k ≥ |X|. Consequently, X is a pattern in I for the parameter k 0 = O(k). Let us now look at the contribution of the term including the potential ΛI (X) into the success probability. Using the estimate on LB(n, Π(X), Γ, Φ(X)) from Section 5.1, we obtain that c LB(n, Π(X), Γ, Φ(X), Λ(X)) =   lg n lg k 0 2 0 0 = LB(n, Π(X), Γ, Φ(X)) · exp −c3 · Λ(X) · lg k (lg k + lg lg n) + √ k0  √ 2 −1 h √ i ≥ 2O( k lg k) nO(1) · exp −c3 · c · k 0 lg k 0 (lg k 0 + lg lg n) + lg n  √ 2 −1 ≥ 2O( k lg k) nO(1) .  49 Note that we used Claim 18 in the last inequality. This finishes the proof of Theorem 34. A.2 The excluded minor case Using the multiple-components variant of the previous section, we now prove Theorem 33. Let us fix a graph G and an integer k as in the theorem statement. Furthermore, let X be a pattern in G such that G[X] admits a spanning tree S of maximum degree ∆, for a fixed constant ∆. The Robertson-Seymour decomposition theorem. As announced at the beginning of this section, we use the decomposition theorem of Robertson and Seymour for graphs excluding a fixed minor. To make use of locally bounded treewidth, we will use the variant of Grohe [21]. To formulate this decomposition, we need some notation. Recall that we use a notation T = (T, β) for a tree decomposition, where T is a tree and β : V (T ) → 2V (G) are bags. The set β(t) ∩ β(t0 ) for an edge tt0 ∈ E(T ) is called an adhesion of tt0 . For a node t ∈ V (T ), the torso of the node t, denoted by torso(t), is the graph G[β(t)] with every adhesion β(t) ∩ β(t0 ) for t0 ∈ NT (t) turned into a clique. Recall that we consider rooted tree decompositions; that is, the tree T is rooted in one node. For a non-root node t ∈ V (T ), by parent(t) we denote the parent of t in T . Furthermore, we denote: ( ∅ if t is the root of T σ(t) = β(t) ∩ β(parent(t)) otherwise. We are now ready to formulate the variant of the Robertson-Seymour decomposition of Grohe that we use. Theorem 47 ( [21]). For every proper minor-closed graph class C there exists a constant h and an apex-minor-free graph class C 0 such that the following holds. Given a graph G ∈ C, one can in polynomial time compute a tree decomposition T = (T, β) of G together with a family of sets (Zt )t∈V (T ) such that every adhesion of T has size at most h, and for every t ∈ V (T ) the set Zt is a subset of β(t) of size at most h, and the graph torso(t) − Zt belongs to C 0 . Furthermore, we need the following variant of Baker’s shifting technique. Theorem 48 ( [21]). Let C be a proper minor-closed graph class. Given a graph G ∈ C and an integer `, one can in polynomial time compute a partition of V (G) into ` sets L1 , L2 , . . . , L` , such that for every 1 ≤ i ≤ ` the graph G − Li is of treewidth O(`). The algorithm. Let us now proceed to the description of the algorithm. Given a graph G ∈ C, we compute its tree decomposition T = (T, β) and sets (Zt )t∈V (T ) using Theorem 47. Paying 1/n in the success probability, we guess an arbitrary vertex r ∈ X, and root the decomposition T in a bag tr such that r ∈ β(tr ). By slightly abusing the notation, we proclaim σ(tr ) = {r}. Since we can assume that G is connected, we can henceforth assume that every adhesion is nonempty. For every t ∈ V (T ), we create an instance It = (Gt , rt , Ttli , Tthe , Rt , λt ) of the recursive problem as follows. We first take Gt = torso(t) − Zt . If σ(t) 6⊆ Zt , we define rt to be an arbitrary vertex of σ(t) \ Zt ; otherwise, we create a new vertex rt and make it adjacent to an arbitrary vertex of Gt . We set Ttli = {rt } ∪ (σ(t) \ Zt ), Tthe = ∅, Rt = ∅, and λt = 0. Furthermore, we define Xt◦ = X ∩ β(t) and Xt = (Xt◦ \ Zt ) ∪ {rt }. Note that torso(t)[Xt◦ ] is connected, as G[X] is connected. Furthermore, we claim that torso(t)[Xt◦ ] admits a spanning tree of bounded degree. 50 Claim 49. There exists a spanning tree of torso(t)[Xt◦ ] of maximum degree 2∆. Proof. We construct a connected spanning subgraph St of torso(t)[Xt◦ ] of maximum degree 2∆ as follows. First, we take V (St ) = Xt◦ and E(St ) = E(S)∩E(torso(t)[Xt◦ ]. Second, for every t0 ∈ NT (t), we perform the following operation. Let leg(t, t0 ) be the set of those vertices v ∈ Xt◦ ∩ β(t) ∩ β(t0 ) for which v is incident to an edge uv ∈ E(S) with u ∈ β(t0 ) \ β(t). If |leg(t, t0 )| ≥ 2, we add to E(St ) edges of an arbitrary path on vertex set leg(t, t0 ); such a path exists in torso(t) as the adhesion β(t) ∩ β(t0 ) is turned into a clique. Let us now show that St is connected. To this end, consider a maximal path P in S between two vertices of Xt◦ such that no edge or internal vertex of P belongs to torso(t). Let v1 , v2 be the endpoints of P . By the properties of a tree decomposition, there exists a node t0 ∈ NT (t) such that v1 , v2 ∈ β(t)∩ β(t0 ) and the first and last edges of P are v1 u1 and v2 u2 with u1 , u2 ∈ β(t0 ) \ β(t) (possibly u1 = u2 ). However, then v1 , v2 ∈ leg(t, t0 ), and they remain connected in St . This shows that St is connected. To bound the maximum degree of St , note that for every t0 ∈ NT (t) and every v ∈ leg(t, t0 ), at most two edges incident to v are added to St when considering t0 . These two edges can be charged to the edge vu ∈ E(S) with u ∈ β(t0 ) \ β(t) that certifies that v ∈ leg(t, t0 ). We have vu ∈ E(S) \ E(St ), and every edge vu can be charged at most once. Since S has maximum degree at most ∆ by assumption, the bound on the maximum degree of St follows. y We define k 0 := (105 · ∆h)3 · k = O(k). We will use the machinery of Section A.1, in particular all the potentials, using the parameter k 0 instead of the input parameter k. The reason for this is that we need to pay the penalty in the size of the pattern for multiple connected components of Xt in It . The following claim verifies that it suffices to inflate k by a constant factor. Claim 50. The graph Gt [Xt ] has at most 2∆h + 1 connected components and ΛIt (Xt ) ≤ 2∆h. Furthermore, the set Xt is a pattern in It with respect to the parameter k 0 . Proof. Since the maximum degree of St is at most 2∆ and |Zt | ≤ h, there are at most 2∆h connected components of torso(t)[Xt◦ ]−Zt . Consequently, Gt [Xt ] has at most 2∆h+1 connected components and ΛIt (Xt ) ≤ 2∆h. Then, as √ αk lg(αk) ≤ α2/3 k for every k ≥ 10 and α ≥ 105 , we have: √ k 0 − 486 k 0 lg k 0 · ΛIt (Xt ) ≥ (105 · ∆h)3 k − 486 · (105 · ∆h)2 · 2∆h > 2k ≥ |Xt |. Thus, Xt satisfies the size bound for a pattern in It . y For every node t ∈ V (T ), we are going to look for the pattern Xt in the instance It , using k 0 as the parameter. Observe that the first three potentials partition well between the instances. Claim 51. The following holds, if we measure the potentials with respect to the parameter k 0 : X X X k≥ ΠIt (Xt ) and n≥ ΓIt and k≥ ΦIt (Xt ). t∈V (T ) t∈V (T ) t∈V (T ) Proof. The crucial observation is that for an edge between t and parent(t) in T , every vertex v ∈ σ(t) = β(t) ∩ β(parent(t)) is either in Zt or a light terminal in It . Consequently, every vertex v ∈ V (G) is present but not a light terminal in at most one instance It . y 51 However, the potentials ΛIt (Xt ) do not behave as nicely as the other potentials in Claim 51: they are bounded by 2∆h by Claim 50, but may be positive in Ω(k) instances. Thus, we cannot afford to apply the algorithm of the previous section to every instance It separately. Instead, for every instance It , we make a random choice. With probability 1/k, we proclaim It interesting and apply the recursive algorithm to It (that belongs to an apex-minor-free graph class C 0 promised by Theorem 47), obtaining a set At ⊆ (β(t) \ Zt ) ∪ {rt } such that Gt [At ] is of treewidth √ √ 0 0 O( k lg k ) = O( k lg k). Furthermore, note that σ(t) ⊆ At ∪ Zt . With the remaining probability, we proclaim It standard, and proceed √ as follows. First, we apply the algorithm of Theorem 48 to the graph Gt with ` := dc3 k 0 lg k 0 e obtaining a partition Lt1 , Lt2 , . . . , Lt` . Second, we pick a random index 1 ≤ it ≤ `. Third, we √set At := (V (Gt ) \ Ltit ) ∪ {rt } ∪ (σ(t) \ Zt ). Note that in this case also Gt [At ] has treewidth O( k lg k) as √ ` = O( k lg k) and |σ(t)| ≤ h = O(1). Furthermore, we have again σ(t) ⊆ At ∪ Zt . We define [ (At ∪ Zt ) \ (σ(t) ∪ {rt }). A := {r} ∪ t∈V (T ) We claim that A satisfies the desired properties. The treewidth bound is easy. √ Claim 52. G[A] is of treewidth O( k lg k). Proof. √ Since |Zt | ≤ h = O(1) for every t ∈ V (T ), we have that torso(t)[At ∪ Zt ] is of treewidth O( k lg k). Since σ(t) ⊆ At ∪ Zt for every t ∈ V (T ), we have that G[A] can be constructed from torso(t)[At ∪ Zt ] by vertex deletions and clique sums along cliques of size at most h, and the claim follows. y Finally, we check the probability that X ⊆ A. To this end, we need the following simple estimate. Claim 53. LetPa, b be positive integers, and a1 , a2 , . . . , ar be integers such that 0 ≤ ai < a for every 1 ≤ i ≤ r and ri=1 ai ≤ b. Then r  Y i=1 1− ai  ≥ a−2b/a−1 . a Proof. We use the following local improvement argument: whenever we have two indices 1 ≤ i < j ≤ r such that ai + aj < a, we can replace ai and aj with ai + aj , since  1− aj  ai + aj ai   1− ≥1− . a a a Thus, we can assume that for every 1 ≤ i < j ≤ r we have that ai + aj ≥ a. In particular, every index i satisfies ai ≥ a/2, apart from at most one. We infer that r ≤ 2b/a + 1. Since 1 − aai ≥ a−1 for each i due to a, ai being integers, the claim follows. y √ Claim 54. The probability that X ⊆ A is at least (2O( k lg2 k) · nO(1) )−1 . Proof. Note that we have the following partition: V (G) = {r} ] ] β(t) \ σ(t). t∈V (T ) By the definition of the set A, we have r ∈ A, and for every t ∈ V (T ) and for every v ∈ β(t) \ σ(t) it holds that v ∈ A if and only if v ∈ At ∪ Zt . Consequently, by the definition of Xt and Gt , if for 52 every t ∈ V (T ) we have Xt ⊆ At , then we have X ⊆ A. In what follows we will argue that with sufficient probability it holds that for every t ∈ V (T ) we have Xt ⊆ At . First, observe that this assertion is clearly true for every node t where Xt ⊆ {rt } ∪ (σ(t) \ Zt ) as both in standard and interesting nodes we have σ(t) ⊆ At ∪ Zt and rt ∈ At . If this is not the case for a node t (i.e., Xt 6⊆ {rt }∪(σ(t)\Zt )), we call the node t touched. Note that we have ΠIt (Xt ) > 0 for a touched node t. Hence, there are k touched nodes. We require that √ at most 0 0 a touched node t is proclaimed interesting√if ΠIt (Xt ) ≥ c3 k lg k , and standard otherwise. Note that Claim 51 implies that we require at most k 0 /(c3 lg k 0 ) nodes to be interesting and at most k nodes√to be standard. Consequently, the probability that we proclaim touched nodes as requested is 2−O( k) . In every standard touched node t we have Xt ⊆ At if Xt ∩ Ltit ⊆ Ttli , as Ttli = {rt } ∪ (σ(t) \ Zt ). We have Xt ∩ Ltit ⊆ Ttli with probability at least 1 − |Xt \ Ttli |/` = 1 − ΠIt (Xt )/`. Recall that √ ` = dc3 k 0 lg k 0 e but ΠIt (Xt ) < ` in a standard node t. Consequently, by Claim 53, the probability that in every standard node we have Xt ⊆ At ∪ {rt } is at most √ 0 `−2k /`−1 = 2−O( k) . √ Let us now consider an interesting node t, that√is, a node t with ΠIt (Xt ) ≥ c3 k 0 lg k 0 . Let Wint be the set of these nodes; note that |Wint | ≤ k 0 /(c3 lg k 0 ). Since Xt is a pattern in It , we c have Xt ⊆ At with probability at least LB(n, ΠIt (Xt ), ΓIt , ΦIt (Xt ), ΛIt (Xt )) (with respect to the 0 parameter k ). By Claims 50 and 51 we have that Y c LB(n, ΠIt (Xt ), ΓIt , ΦIt (Xt ), ΛIt (Xt )) t∈Wint ≤ Y LB(n, ΠIt (Xt ), ΓIt , ΦIt (Xt )) t∈Wint    lg n lg k 0 2 0 0 · exp −c3 · ΛIt (Xt ) · lg k (lg k + lg lg n) + √ k0 t∈Wint −1  √ 2 ≤ 2O( k lg k) nO(1)    lg n lg k 0 · exp −|Wint | · c3 · 2∆h · lg2 k 0 (lg k 0 + lg lg n) + √ k0  √ 2 −1 h √ i ≤ 2O( k lg k) nO(1) · exp −2∆h k 0 lg k 0 (lg k 0 + lg lg n) + lg n  √ 2 −1 ≤ 2O( k lg k) nO(1) . Y Here, we estimated the product of the terms LB(n, ΠIt (Xt ), ΓIt , ΦIt (Xt )) using Claim 51 as in Section 5, and in the last inequality we used Claim 18. y This concludes the proof of Theorem 33. 53
8cs.DS
Training a Fully Convolutional Neural Network to Route Integrated Circuits Kye Okabe † Stanford University arXiv:1706.08948v2 [cs.CV] 11 Sep 2017 Sambhav R. Jain ∗ † Oracle America Abstract We present a deep, fully convolutional neural network that learns to route a circuit layout ‘net’ with appropriate choice of metal tracks and wire class combinations. Inputs to the network are the encoded layouts containing spatial location of pins to be routed. After 15 fully convolutional stages followed by a score comparator, the network outputs 8 layout layers (corresponding to 4 route layers, 3 via layers and an identity-mapped pin layer) which are then decoded to obtain the routed layouts. We formulate this as a binary segmentation problem on a per-pixel per-layer basis, where the network is trained to correctly classify pixels in each layout layer to be ‘on’ or ‘off’. To demonstrate learnability of layout design rules, we train the network on a dataset of 50,000 train and 10,000 validation samples that we generate based on certain pre-defined layout constraints. Precision, recall and F1 score metrics are used to track the training progress. Our network achieves F1 ≈ 97% on the train set and F1 ≈ 92% on the validation set. We use PyTorch for implementing our model. Code is made publicly available1 . 1. Introduction Routing is a complex spatial optimization problem in the physical design of integrated circuits (ICs) that is known to be NP-complete in most cases [20, 15, 10, 22]. The task is to optimally connect circuit segments spanning multiple layout hierarchies and multiple wire classes, while complying to a strict set of design rules dictated by the foundry’s process design kit (PDK). The quality of routing determines circuit performance (frequency), reliability, and can also impact area. Depending on the type of circuit (ASIC / memory / processor), the routing objective may prioritize one over the other, but in general the expectation is to: • • • • Figure 1. Training samples from our dataset. Top row shows input data containing pins for a single net case. Bottom row shows input labels (decoded to 8-bit RGB) containing pins, routes, and vias. Pins are routed using two wire classes (branch-leg). Wire class color coding: M3 =green, M4 =red, M5 =grey, M6 =blue. Formally, a circuit layout consists of several sub-blocks (or cells) with input/output ports (or pins). Connections between pins use channels (or tracks) in several metal layers that run orthogonal to one another, separated by insulating layers with connecting vias. If we treat cells as edges, then a node in a circuit graph corresponds to a net in the layout, which connects the driver’s output pin with the input pin(s) of the receiver(s). Traditionally, when the circuit sizes were tractable, routing was primarily a manual task. However, in the past couple decades with exploding circuit sizes and more than a few billion transistors to be routed, the use of Minimize path delay (resistance/capacitance) Minimize congestion (wire crowding, hot spots) Maximize routability (availability of tracks) Maximize repeatability ∗ Corresponding author: [email protected] equal contribution. 1 Code: https://github.com/sjain-stanford/deep-route † Indicates 1 auto-routers is indispensable. Shrinking technology nodes and more stringent design rules further impact routability. Commercial electronic design automation (EDA) tools try to tackle this non-linear optimization problem using various algorithmic approaches [20] such as (i) exponential algorithms to exhaust the search space for a solution, (ii) heuristic algorithms, (iii) optimal algorithms for special cases of the problem, or (iv) approximation techniques. These approaches are iterative in nature and rely on a continually evolving/changing set of design rules. The routed layouts may further require significant manual effort and fine tuning to improve the non-optimal routes. We explore a learning-based approach wherein we train a deep, fully convolutional network (FCN) to route a layout net while relying on its ability to learn implicit design rules from the training data. In order to demonstrate learnability of layout design rules by the network, we pre-define a set of basic constraints (not specific to any technology node) that are embedded in the ground truth layouts generated for a single net case. The network is trained on our dataset of 50,000 train and 10,000 validation samples. Input data (containing pins) and labels (containing pins and routes) are first encoded on a binary basis per-pixel per-layer (detailed in Section 3), before feeding to the 15 stage FCN. FCN outputs go through a score comparator to give 8-layer encoded layouts (4 route layers, 3 via layers, and 1 pin layer). We decode this to 8-bit RGB for visualization. The various architectural choices of the FCN are covered in depth in Sections 4 and 5. The network shows ability to learn (i) the identity-map for pins, (ii) optimal track locations for routes, (iii) suitable wire class combination, and (iv) via locations. One of the very first ideas of using neural networks for circuit routing came from Green and Noakes [6], who divided the routing task into several stages and combined multiple small back propagation networks to form a complex neural system. They proposed to restrict the routing task to predefined areas and slide the context window to complete routes in sections. This divide and conquer approach helped reduce the complexity of their artificial neural network router. Our work takes inspiration from the recent advances in convolutional neural networks (CNN) [3, 12, 11] which are better suited for visual tasks as they preserve spatial information in the inputs. Similar to [6], we fix the routing window to a predefined layout size, however, in contrast to [6], we develop a single, end-to-end, deep network using convolutions, which holistically learns multiple design rules during training, and is able to route using different wire class combinations, depending on the spatial spread of the pins. To our best knowledge, this is the first attempt at routing a circuit layout net using a convolutional neural network. 3. Dataset Overview. We develop our own dataset owing to the lack of a publicly available layout dataset, and the need to use simplistic design rules to assess learnability by the network (feasibility). In this section we discuss the design choices and constraints used when generating 50,000 layouts for training and 10,000 layouts for validation. Each layout sample contains both data (pins only) and labels (pins, routes, vias) for a single net (see Figure 1). The image is pixelwise binary encoded in 8 layers of the layout, viz. [pin, M3 , V ia3 , M4 , V ia4 , M5 , V ia5 , M6 ]. So each pixel in a given layer is either ’1’ or ’0’, indicating the presence or absence of the layer at that spatial location. Our reasoning behind this encoding scheme is discussed in Section 4.1. Thus training data are stored as tensors of shape N × 1 × H × W , and labels as tensors of shape N × 8 × H × W , where H = W = 32. The layouts were sized 32 × 32 pixels, which we found computationally feasible. We use chunked HDF5 [24] when storing and loading this large dataset to avoid memory bottlenecks. 2. Related Work The task of producing routes using orthogonal layers of parallel channels and vias at intersections is often tackled heuristically, since an optimal solution does not exist for this NP-complete problem [20, 15, 10, 22]. Previous work in this field [27, 17, 2, 26, 25, 28] mostly rely on explicit rule-based algorithms to tackle parts of this complex task. For instance, Zajc et al. [28] proposed using hierarchical algorithms for automatic routing of interconnects in two layer channels. Rivest et al. [17] show a ”greedy” heuristic channel router assuming all pins and wiring sit on a common grid. The other class of routers, which have received attention lately, are objective function based. Alpert et al. [1] proposed a combination of minimum spanning tree (Prim) and shortest path tree (Dijkstra) objectives for performance driven global routing. Constructing routing tree with different objectives was also explored, such as timing objective [21], buffer insertion [13] and wire sizing objective [23], congestion and thermal objective [18, 8]. A completely different objective for diagonal routing as opposed to the orthogonal (Manhattan) routing was also proposed [19]. Design rules. These design rules are fundamental to traditional layout design and do not correspond to any specific technology node. 1. M2n tracks run horizontally. 2. M2n+1 tracks run vertically. 3. Mn contacts Mn+1 through V ian . Design choices. We made the following design choices to reduce the complexity of network implementation. These could be scaled as needed given sufficient resources. 1. Maximum npins = 5. 2. Layout context window = 11µm × 11µm. 2 4.1. Binary Encoding Scheme 3. Total allowed route layers = 4; total via layers = 3. 4. Routes are limited to two wire classes (branch-leg) (e.g. [M3 , M4 ] or [M4 , M5 ] or [M5 , M6 ]). 5. Metal tracks use pixel grid (one track per pixel). 6. Pin layer is lower than all route layers, and routes need not drop vias down to the pin layer. In typical image generation problems such as generative adversarial nets (GAN) [4], at the point of weight initialization, networks are intrinsically capable to generate an arbitrary color at any pixel coordinate. Subsequently through the course of training, meaningful color/coordinate combinations must be learned. The standard 8-bit RGB color gamut has 224 color representations (3 color channels, 8 bits each) which is significantly larger than what we need. Since our data only deals with 8 layout layers (4 route, 3 via, 1 pin) we can immediately set an upper bound on the representation combinations per pixel to 28 . To take advantage of this insight, we choose to encode our data on a binary basis per-pixel per-layer. Hence instead of representing images as tensors of shape 3 × H × W with range: {0 − 255} (8-bit RGB), we encode our dataset as tensors of shape 8×H ×W with a binary range: {0, 1}. By doing so, we are able to formulate this as a layer-wise binary segmentation task, with cross entropy loss on the scores (16 × H × W ) to let the network make a binary decision of the presence of each of the 8 layers at every pixel in the context window H × W . Higher metals normally have less wire resistance per unit length compared to lower metals, due to their large cross sectional area and/or better material properties. However, higher metals also require additional vias to jump up/down to/from higher layers, which adds to the total resistance. As a result, there is a break-even route distance above which a higher metal is preferred. Moreover, for a given layout size, depending on wire performance data there are only a limited set of wire classes which make sense for routing. Using higher or lower wire classes than necessary would be non-optimal. In our case, we use empirical wire and via resistance data2 for four such wire classes (say M3 to M6 ). We select 11µm as the size of the context window as it gave us a balanced dataset among the three wire class combinations, viz. [M3 , M4 ], [M4 , M5 ], [M5 , M6 ]. By defining a context window for routing, we let the possibility open to have it slide over the entire layout to complete routing in segments, similar to [6], but we leave the slide-and-route implementation to future work. 4.2. Network Architecture Figure 2 illustrates our model with the activation volumes after each stage. The FCN has a total of 15 convolutional3 stages (see Section 5 for further explanation on the choice of network depth). All convolutions except the last one are followed by batch normalization (BN) [7] and leaky rectified linear unit (leaky ReLU) [14]. The last convolution stage (L15 ) outputs the scores, hence is not followed by BN and leaky ReLU as they would not affect the relative scores. Encoded inputs to the FCN are tensors of shape N ×1×H×W , where N is the mini-batch size and H = W = 32 (spatial dimensions). The first stage (L1 ) uses 33 × 33 convolution with 16 filters, whereas stages L2 through L15 use 3×3 convolution with 16 filters. We believe the large receptive field at the head of the network allows a fast grasp of the overall pin locations and helps the network learn spatial information better, as will be explored in Section 5. Strides and padding for convolutions are such as to preserve the spatial dimension of the feature maps at each stage. Thus the activations after each stage are of shape N × 16 × 32 × 32, or more generally N × (α.β) × H × W where α is the number of layout layers to be learned, and β is the number of segmentation classes (two in our case). Activations from the last FCN stage (scores) are passed through the score comparator to pick the class with a higher score. Thus scores of shape N × (α.β) × H × W are reduced to layout maps of shape N × α × H × W with each pixel being either ‘1’ or ‘0’ indicating the presence or absence of a layer at that spatial location. For visualization of the routes, we decode the score comparator outputs to 8-bit RGB. Route algorithm. For each layout we first sample (random uniform) npins from {2, 3, 4, 5} and (x, y) co-ordinates for each pin from {0 − 31} (since pixel grid is 32 × 32). For a given pin configuration, the direction with the largest spread of pins is chosen as dominant. The dominant direction uses a branch, while the non-dominant direction uses legs to connect individual pins to the branch. The choice of metals for branch and legs is done so as to optimize the combined wire and via resistance, as explained earlier. From Figure 1, we can qualitatively see that shorter routes use lower metals while longer routes use higher metals. We can also see that a branch is always assigned to the dominant direction, and legs to the other direction. For visualization, the routed layouts are decoded to 8-bit RGB as follows: M3 as green, M4 as red, M5 as grey and M6 as blue. 4. Model Overview. Our proposed model consists of a single, endto-end, deep network using convolutions, which takes pin locations for a net as inputs and generates routes in one of the three wire class combinations, using layout design rules it learns from training data. We first present a binary scheme of encoding the input data in Section 4.1. The FCN model (see Figure 2) and its architectural details are covered in Section 4.2. Then in Sections 4.3 and 4.4 we describe the loss function and the details of training respectively. 2 Not 3 Fully released due to proprietary reasons 3 connected layers or excessively deep layers were avoided. M3 pin Encoded Inputs Nx1x32x32 Range: 0 or 1 L1 Nx16x32x32 L2 L15 (scores) Nx16x32x32 Nx16x32x32 Nx(α.β)xHxW Range: ℝ M4 Via3 M5 M6 Via4 Via5 Score Comparator Nx8x32x32 Nx(α)xHxW Range: 0 or 1 Decoded Outputs Nx3x32x32 Range: 0~255 (8 bit RGB) Fully Convolutional Network (FCN) L1 Conv (F=33, P=16, S=1)  BN  Leaky ReLU L2~14 Conv (F=3, P=1, S=1)  BN  Leaky ReLU L15 Conv (F=3, P=1, S=1) Figure 2. Model overview. The encoded layouts (with pins only) are fed to a 15 stage FCN which performs convolutions while preserving spatial dimensions. The filter size (F), padding (P) and stride (S) used in each stage is indicated. Scores at the end of the FCN are processed by a score comparator to generate the routed layouts (with pins, routes, and vias) which are visualized after decoding to 8-bit RGB. 4.3. Loss Function 4.4. Training and optimization To implement the network objective as a layer-wise binary segmentation task wherein every pixel in every layer is classified into one of the two classes (y = 0 or y = 1), we perform simple but critical reshape operations as follows. At train time, scores from the FCN (tensor) of shape N × (α.β) × H × W are first reshaped to a matrix of shape N.H.W.α × β. The corresponding labels (tensor) of shape N × α × H × W are reshaped to a vector of length N.H.W.α. We then use averaged cross entropy loss (combination of negative log likelihood and softmax) over the predicted score matrix and label vector, to train our model (see Eq. 1). We train the model from scratch, starting with default weight initialization. We use Adam [9] with β1 = 0.9, β2 = 0.999, and  = 1 × 10−8 to train the weights of the FCN components. We experiment with two different minibatch sizes, viz. N = 10 and N = 100 with learning rates of 5 × 10−5 and 5 × 10−4 respectively (see Section 5). A mini-batch of 10 runs in approximately 90ms (∼7.5 min per epoch) and a mini-batch of 100 takes about 600ms (∼5 min per epoch) on a Tesla K80 GPU. We use PyTorch [16] to train and implement the network. L= 1 N.H.W.α ( X ky e − log P β N.H.W.α +λ Sy XXX 5. Experiments F1 score as accuracy metric. Given the nature of our training data which is massively unbalanced towards one class (y = 0), a raw metric comparing pixels of predicted and actual (ground truth) layouts will easily present an unreasonably high pixel-wise accuracy (∼98% in our case) even if the model incorrectly predicts all pixels to be background. We instead use the F1 score metric that uses an equally weighted harmonic mean of precision and recall based confusion matrix. Precision is a measure of true positives among all pixels predicted positive. Recall is a measure of true positives among all pixels ground truth positive. ) eSβ (1) wn2 α×β n α×β Class imbalance. Since a majority of pixels in the input labels are background (y = 0) with very few active pixels (y = 1), we observe that the network quickly learns to classify all pixels as background and struggles to learn further. To mitigate such sparse learning difficulties due to class imbalance, we use weighted cross entropy loss with weights k y=1 = 3 and k y=0 = 1. Regularization. An L2 regularization term was added to the loss to improve generalization.PWe sum over squared weights of 16 convolution filters ( α×β ) from each stage P ( n ) and use a regularization coefficient λ = 1 × 10−5 . ### Pseudocode # tp = true positives, tn = true negatives # fp = false positives, fn = false negatives precision = tp / (tp+fp) recall = tp / (tp+fn) accuracy = (tp+tn) / (tp+tn+fp+fn) f1_score = 2 * (precision*recall) / (precision+recall) 4 Figure 3. First stage experiment. We compare training curves with the first FCN stage using filters of size (i) F=3 (left) and (ii) F=33 (right). Small receptive field at the head of the network prevents the model from learning, as seen from the low train accuracy (F1 ≈ 35%) on the left compared to the one on the right (F1 ≈ 85%) after 200 epochs. Network depth and receptive field. We choose the FCN depth such as to have the overall receptive field of the network cover the entire input image (32 × 32 pixels). Assuming all stages use 3 × 3 unit-strided convolutions, 3 + 2 × (nstages − 1) > 32, we would need at least nstages = 15 to allow the network to learn reasonably well. We observe that reducing the FCN depth below 15 makes it difficult for the model to perfectly overfit (F1 = 100%) even a tiny dataset of 4 train samples. Note that our model uses 33 × 33 convolutions only in the first stage (see Figure 2). Final training with different mini-batches. Taking inspiration from [5], we conduct the final training using two different mini-batch sizes, viz. N = 10 and N = 100. The linear scaling rule [5] suggests adjusting the learning rate linearly as a function of mini-batch size. This makes intuitive sense because we make fewer iterative steps per epoch with a larger mini-batch, hence the step size (learning rate) needs to be proportionally larger. We set the learning rates to 5 × 10−5 and 5 × 10−4 for mini-batches of 10 and 100 respectively, which worked reasonably well in our hyperparameter tuning experiments. To overcome overfitting seen in Figure 3, we now use our complete dataset of 50,000 train and 10,000 validation samples for training. Figure 4 shows the loss and accuracy curves (on both train and validation sets) for the two mini-batches. The overall trends look comparable. With mini-batch of 10 (left plot), the model achieves accuracies F1 ≈ 90% and F1 ≈ 82% on the train and validation sets after 200 epochs, and takes about 45 epochs to reach validation accuracy of 80%. In contrast, with mini-batch of 100 (right plot) the model achieves F1 ≈ 97% and F1 ≈ 92% on the train and validation sets after 200 epochs, and takes only 20 epochs to reach validation accuracy of 80%. Total train time significantly improved from 25 hours (∼7.5 min per epoch) with minibatch of 10 to 16 hours (∼5 min per epoch) with mini-batch of 100, on a Tesla K80 GPU. The curves show good generalization of the model on the validation set with only a small gap (< 5%) between training and validation accuracies. Also included are the precision and recall curves for the second case. Significance of F = 33 in the first FCN stage. As discussed in Section 4.2, we use 33 × 33 convolutional filters in the first FCN stage as we believe this large receptive field at the head of the network helps the model quickly learn the correspondence between spatial spread of pins in the input and crucial route decisions such as wire class combinations, track usage, and branch-leg assignment. To demonstrate the significance of this architecture, we compare two models differing only in their first FCN stages, viz. (i) F = 3, and (ii) F = 33. Rest of the model is unchanged. For this experiment we train the two models on a subset of our dataset (1000 train and 200 validation samples), using mini-batches of 10, learning rate of 5 × 10−5 and regularization strength of 1 × 10−5 . From Figure 3 we see that learning stagnates around F1 = 35% after 200 epochs with F = 3, whereas the model is able to quickly overfit the small dataset fairly well (F1 = 85%) with F = 33. Interestingly, the increasing gap between training and validation accuracies is due to the small dataset size used for this experiment, causing the model to overfit to it. 5 Figure 4. Loss and accuracy curves on training and validation sets from final training. Left shows the case with mini-batch of 10, learning rate of 5 × 10−5 . Right shows the case with mini-batch of 100, learning rate of 5 × 10−4 (linear scaling). Precision and recall curves are included for the second case. Best accuracy was recorded at F1 ≈ 97% on train set and F1 ≈ 92% on validation set within 200 epochs. 6. Results after 191 epochs of training. As seen from the actual (ground truth) and predicted layouts, the network does well in (i) learning the identity mapping for pins, (ii) assigning vias to connect metals in adjacent layers, (iii) identifying the correct wire class combination from the overall pin configuration, and (iv) choosing the optimal track positions for branch and legs. It uses lower metals to route pins that are closer, and higher metals for widespread pins. In some cases, however, the routing is not perfect and the model misses connections or adds routes at undesired locations. We typically notice a higher error rate when layouts have more pins. This could likely be improved if we increase the ratio of training samples in the dataset containing more pins. While there is room for further improvement, the FCN model shows good overall ability to learn the layout design rules intrinsic to the dataset used for training. Future work. Due to the complexity involved in routing real IC layouts, there are several requirements that need to be addressed. A few worth mentioning are (1) routing multiple nets in the presence of previously occupied tracks, (2) using a finer grid of valid metal tracks customized per wire class, (3) training on more complex route configurations such as trunk-branch-leg, (4) adding driver / receiver awareness to pins, (5) adding dedicated pin layers for each wire class, (6) supporting routing of a bigger layout in segments using a sliding context window, (7) integrating timing models for timing-driven routing, (8) training on industry standard layouts after converting to our layer-encoded binary standard. Some of these requirements may be implemented by direct scaling of the model and/or training with dense To gain insight into the learning process, we show a routed net example from the validation set with two pins (see Figure 5). The left image shows the actual routed layout (ground truth), the center and right images show the predicted model outputs after 141 and 151 epochs of training respectively. We first notice that the model grasps orthogonality of adjacent metal layers, assigning M3 (green) to vertical tracks and M4 (red) to horizontal tracks only. Second, the model learns to connect different wire classes using vias at intersections. Third, the model learns to assign a branch to the dominant direction and legs to the non-dominant direction, however since the pins are roughly evenly spaced in either direction, the model attempts a vertical branch and horizontal legs. Eventually (after 10 epochs) it learns a more optimal way by using a horizontal branch and vertical legs to route, thus matching our ground truth expectation. Figure 5. Learning example from the validation set. Left image shows actual layout (ground truth), center and right images show model predicted layouts after 141 and 151 epochs respectively. Figure 6 presents some randomly sampled examples from the training and validation sets, routed by our model 6 Figure 6. Examples from training (top row) and validation (bottom row) datasets showing the actual (ground truth) layouts on the left and the corresponding predicted (model routed) layouts on the right, after 191 epochs of training. The FCN model demonstrates good ability to learn layout design rules intrinsic to the dataset. Wire class color coding: M3 =green, M4 =red, M5 =grey, M6 =blue. datasets containing more design rules of interest. However, it is plausible that other requirements may warrant more sophisticated architectures, possibly combining several neural models to construct a complex neural system to route ICs. defined layout constraints. We then implemented an encoding scheme to efficiently represent inputs to the model. The proposed FCN architecture efficiently learns to route a single net under set design constraints. Our model achieves good performance with training accuracy of F1 ≈ 97% and validation accuracy of F1 ≈ 92% within 200 epochs. 7. Conclusion Inspired by the challenges facing circuit layout routing and optimization, and the recent advances in the field of convolutional neural networks, we introduced a unique approach to routing ICs using deep, fully convolutional networks. To explore learnability of layout design rules by our FCN model, we created our own dataset based on pre- 8. Acknowledgements We thank Nishith Khandwala and Wenbin Huang for helpful comments and discussion. We gratefully acknowledge CS231N staff (Stanford) and Google Cloud for the educational GPU credits used towards this work. 7 References [14] A. L. Maas, A. Y. Hannun, and A. Y. Ng. Rectifier nonlinearities improve neural network acoustic models. In Proc. ICML, volume 30, 2013. [15] R. Pal. Multi-layer Channel Routing Complexity and Algorithms. Complexity and Algorithms. Alpha Science, 2000. [16] Facebook AI Research (FAIR). Pytorch. https:// github.com/pytorch, 2016. [17] R. L. Rivest and C. M. Fiduccia. A ”greedy” channel router. In 19th Design Automation Conference, pages 418– 424, June 1982. [18] D. Roy, P. Ghosal, and N. Das. A thermal and congestion driven global router for 3d integrated circuits. In Students’ Technology Symposium (TechSym), 2014 IEEE, pages 303– 308, Feb 2014. [19] T. Samanta, P. Ghosal, H. Rahaman, and P. Dasgupta. A heuristic method for constructing hexagonal steiner minimal trees for routing in vlsi. In 2006 IEEE International Symposium on Circuits and Systems, pages 4 pp.–, May 2006. [20] N. Sherwani, S. Bhingarde, and A. Panyam. Routing in the Third Dimension: From VLSI Chips to MCMs. IEEE Press Series on Microelectronic Systems. Wiley, 1995. [21] J. Song, C. Jang, K. Cho, S. Go, and J. Chong. Timing driven global router with a pin partition method for 3d stacked integrated circuits. In The 18th IEEE International Symposium on Consumer Electronics (ISCE 2014), pages 1–2, June 2014. [22] T. G. Szymanski. Dogleg channel routing is np-complete. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 4(1):31–41, January 1985. [23] X. Tang, R. Tian, H. Xiang, and D. F. Wong. A new algorithm for routing tree construction with buffer insertion and wire sizing under obstacle constraints. In IEEE/ACM International Conference on Computer Aided Design. ICCAD 2001. IEEE/ACM Digest of Technical Papers (Cat. No.01CH37281), pages 49–56, Nov 2001. [24] The HDF Group. Hierarchical data format version 5. http: //www.hdfgroup.org/HDF5, 2000-2010. [25] J. S. Wang and R. C. T. Lee. An efficient channel routing algorithm to yield an optimal solution. IEEE Transactions on Computers, 39(7):957–962, Jul 1990. [26] T. Yoshimura. An efficient channel router. In 21st Design Automation Conference Proceedings, pages 38–44, June 1984. [27] T. Yoshimura and E. S. Kuh. Efficient algorithms for channel routing. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 1(1):25–35, January 1982. [28] B. Zajc, D. Tomasevic, and A. Zemva. Automatic routing of interconnections between cells in integrated circuits. In [1991 Proceedings] 6th Mediterranean Electrotechnical Conference, pages 238–241 vol.1, May 1991. [1] C. J. Alpert, T. C. Hu, J. H. Huang, and A. B. Kahng. A direct combination of the prim and dijkstra constructions for improved performance-driven global routing. In 1993 IEEE International Symposium on Circuits and Systems, pages 1869–1872 vol.3, May 1993. [2] M. Burstein and R. Pelavin. Hierarchical channel router. In 20th Design Automation Conference Proceedings, pages 591–597, June 1983. [3] K. Fukushima. Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position. Biological Cybernetics, 36(4):193–202, 1980. [4] I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Proceedings of the 27th International Conference on Neural Information Processing Systems, NIPS’14, pages 2672–2680, Cambridge, MA, USA, 2014. MIT Press. [5] P. Goyal, P. Dollár, R. Girshick, P. Noordhuis, L. Wesolowski, A. Kyrola, A. Tulloch, Y. Jia, and K. He. Accurate, large minibatch SGD: Training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. [6] A. D. P. Green and P. D. Noakes. Neural networks-their use for the routing of integrated circuits. In Proceedings of the 32nd Midwest Symposium on Circuits and Systems,, pages 501–504 vol.1, Aug 1989. [7] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. [8] A. Ivanov, P. Hallschmid, and Z. Zhou. Local congestion and blockage aware routability analysis using adaptive flexible modeling. In 2016 IEEE International Conference on Electronics, Circuits and Systems (ICECS), pages 438–439, Dec 2016. [9] D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. In Proceedings of the 3rd International Conference on Learning Representations (ICLR), 2014. [10] M. R. Kramer and J. van Leeuwen. Wire-routing is NP-complete. http://www.cs.uu.nl/research/ techreps/repo/CS-1982/1982-04.pdf, 1982. [11] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 25, pages 1097–1105. Curran Associates, Inc., 2012. [12] Y. Lecun, L. D. Jackel, H. A. Eduard, N. Bottou, C. Cartes, J. S. Denker, H. Drucker, E. Sackinger, P. Simard, and V. Vapnik. Learning algorithms for classification: A comparison on handwritten digit recognition. In Neural Networks: The Statistical Mechanics Perspective, pages 261– 276. World Scientific, 1995. [13] J. Lillis, C.-K. Cheng, and T.-T. Y. Lin. Simultaneous routing and buffer insertion for high performance interconnect. In Proceedings of the Sixth Great Lakes Symposium on VLSI, pages 148–153, Mar 1996. 8
1cs.CV