An Idea Whose Time Has Come
ltra-wideband (UWB) radio is a fast emerging technology with uniquely attractive features inviting major advances in wireless communications, networking, radar, imaging, and positioning systems. By its rule- making proposal in 2002, the Federal Communications Commission (FCC) in the United States essen- tially unleashed huge “new bandwidth’’ (3.6–10.1 GHz) at the noise floor, where UWB radios Uoverlaying coexistent RF systems can operate using low-power ultra-short information bearing pulses. With similar regulatory processes currently under way in many countries worldwide, industry, government agencies, and academic institutions responded to this FCC ruling with rapidly growing research efforts targeting a host of exciting UWB applications: short-range very high-speed broadband access to the Internet, covert communication links, localization at centimeter-level accuracy, high-resolution ground-penetrating radar, through-wall imaging, precision navigation and asset tracking, just to name a few. This tutorial focuses on UWB wireless communications at the physical layer. It overviews the state-of-the-art in channel modeling, transmitters, and receivers of UWB radios, and outlines research directions and challenges to be overcome. As signal processing expertise is expected to have major impact in research and development of UWB systems, emphasis is placed on DSP aspects.
Introduction UWB characterizes transmission systems with instantaneous spectral occu- pancy in excess of 500 MHz
Liuqing Yang and Georgios B. Giannakis
2626 IEEEEEE SIGNALSIGNAL PROCESSINGPROCESSING MAGAZINEMAGAZINE NONOVEMBERVEMBER 20042004 1053-5888/04/$20.00053-5888/04/$20.00©2004IEEE2004IEEE or a fractional bandwidth of more than 20%. (The frac- dia transmissions. Similar to the frequency reuse princi- tional bandwidth is defined as B/fc , where ple exploited by wireless cellular architectures, low- B := fH − fL denotes the −10 dB bandwidth and cen- power, short-range UWB communications are also ter frequency fc := ( fH + fL )/2 with fH being the potentially capable of providing high spatial capacity, in upper frequency of the −10 dB emission point, and fL terms of bits per second per square meter. In addition, the lower frequency of the −10 dB emission point. UWB connectivity is expected to offer a rich set of soft- According to [12], UWB systems with fc > 2.5 GHz ware-controllable parameters that can be used to design need to have a −10 dB bandwidth of at least 500 location-aware communication networks flexible to MHz, while UWB systems with fc < 2.5 GHz need to scale in rates and power requirements. have fractional bandwidth at least 0.20.) Such systems To fulfill these expectations, however, UWB research rely on ultra-short (nanosecond scale) waveforms that and development has to cope with formidable chal- can be free of sine-wave carriers and do not require IF lenges that limit their bit error rate (BER) perform- processing because they can operate at baseband. As ance, capacity, throughput, and network flexibility. information-bearing pulses with ultra-short duration Those include high sensitivity to synchronizing the have UWB spectral occupancy, UWB radios come with reception of ultra-short pulses, optimal exploitation of unique advantages that have long been appreciated by fading propagation effects with pronounced frequency- the radar and communications communities: i) enhanced selectivity, low-complexity constraints in decoding capability to penetrate through obstacles; ii) ultra high high-performance multiple access protocols, and strict precision ranging at the centimeter level; iii) potential for power limitations imposed by the desire to minimize very high data rates along with a commensurate increase interference among UWB communicators, and with in user capacity; and iv) potentially small size and pro- coexisting legacy systems, particularly GPS, unmanned cessing power. Despite these attractive features, interest air vehicles (UAVs), aircraft radar, and WLANs. These in UWB devices prior to 2001 was primarily limited to challenges call for advanced digital signal processing radar systems, mainly for military applications. With (DSP) expertise to accomplish tasks such as synchro- bandwidth resources becoming increasingly scarce, nization, channel estimation and equalization, multi- UWB radio was “a midsummer night’s dream’’ waiting user detection, high-rate high-precision low-power to be fulfilled. But things changed drastically in the analog/digital conversion (ADC), and suppression of spring of 2002, when the FCC released a spectral mask aggregate interference arising from coexisting legacy allowing (even commercial) operation of UWB radios at systems. As DSP theory, algorithms, and hardware the noise floor, but over an enormous bandwidth (up to advanced narrowband and broadband technology, DSP 7.5 GHz). is expected to play a similar role in pushing the fron- This huge “new bandwidth” opens the door for an tiers of emerging UWB applications. To this end, it is unprecedented number of bandwidth-demanding posi- important to understand features and challenges that tion-critical low-power applications in wireless commu- are unique to UWB signaling from a DSP perspective. nications, networking, radar imaging, and localization systems [64]. It also explains the rapidly increasing Regulatory Issues efforts undertaken by several research institutions, and Motivating Applications industry, and government agencies to assess and exploit Despite its renewed interest during the past decade, the potential of UWB radios in various areas. These UWB has a history as long as radio. When invented by include short-range, high-speed access to the Internet, Guglielmo Marconi more than a century ago, radio accurate personnel and asset tracking for increased safe- communications utilized enormous bandwidth as infor- ty and security, precision navigation, imaging of steel mation was conveyed using spark-gap transmitters. The reinforcement bars in concrete or pipes hidden inside next milestone of UWB technology came in the late walls, surveillance, and medical monitoring of the 1960s, when the high sensitivity to scatterers and low heart’s actual contractions. power consumption motivated the introduction of For wireless communications in particular, the FCC UWB radar systems [5], [45], [46]. Ross’ patent in regulated power levels are very low (below −41.3 dBm), 1973 set up the foundation for UWB communications. which allows UWB technology to overlay already avail- Readers are referred to [5] for an interesting and able services such as the global positioning system informative review of pioneer works in UWB radar and (GPS) and the IEEE 802.11 wireless local area net- communications. works (WLANs) that coexist in the 3.6−10.1 GHz In 1989, the U.S. Department of Defense (DoD) band. Although UWB signals can propagate greater coined the term “ultra wideband” for devices occupy- distances at higher power levels, current FCC regula- ing at least 1.5 GHz, or a −20 dB fractional bandwidth tions enable high-rate (above 110 MB/s) data trans- exceeding 25% [37]. Similar definitions were also missions over a short range (10−15 m) at very low adopted by the FCC notice of proposed rule making power. Major efforts are currently under way by the that regulated UWB recently. The rule making of UWB IEEE 802.15 Working Group for standardizing UWB was opened by FCC in 1998. The resulting First COMSTOCK
© wireless radios for indoor (home and office) multime- Report and Order (R&O) that permitted deployment
NOVEMBER 2004 IEEE SIGNAL PROCESSING MAGAZINE 27 of UWB devices was announced on 14 February and ing a vast quantity of sensory data in a timely manner. released in April 2002 [12]. Three types of UWB sys- Typically, energy is more limited in sensor networks tems are defined in this R&O: imaging systems, com- than in WPANs because of the nature of the sensing munication and measurement systems, and vehicular devices and the difficulty in recharging their batteries. radar systems. Spectral masks assigned to these applica- Studies have shown that current commercial Bluetooth tions are listed in Table 1. In particular, the FCC devices are less suitable for sensor network applications assigned bandwidth and spectral mask for indoor com- because of their energy requirements [62] and higher munications is illustrated in Figure 1. expected cost [2]. In addition, exploiting the precise Although currently only the United States permits localization capability of UWB promises wireless sensor operation of UWB devices, regulatory efforts are under networks with improved positioning accuracy. This is way both in Europe and in Japan. Market drivers for especially useful when GPSs are not available, e.g., due UWB technology are many even at this early stage, and to obstruction. are expected to include new applications in the next ▲ Imaging systems: Different from conventional radar few years. We outline here application trends where sig- systems where targets are typically considered as point nal processing tools will probably have considerable scatterers, UWB radar pulses are shorter than the target impact in UWB system development. dimensions. UWB reflections off the target exhibit not ▲ Wireless personal area networks (WPANs): Also known only changes in amplitude and time shift but also as in-home networks, WPANs address short-range (gen- changes in the pulse shape. As a result, UWB wave- erally within 10−20 m) ad hoc connectivity among forms exhibit pronounced sensitivity to scattering rela- portable consumer electronic and communication tive to conventional radar signals. This property has devices. They are envisioned to provide high-quality been readily adopted by radar systems (see e.g., [5] and real-time video and audio distribution, file exchange references therein) and can be extended to additional among storage systems, and cable replacement for home applications, such as underground, through-wall and entertainment systems. UWB technology emerges as a ocean imaging, as well as medical diagnostics and bor- promising physical layer candidate for WPANs, because der surveillance devices [55], [57]. it offers high-rates over short range, with low cost, high ▲ Vehicular radar systems: UWB-based sensing has the power efficiency, and low duty cycle. potential to improve the resolution of conventional ▲ Sensor networks: Sensor networks consist of a large proximity and motion sensors. Relying on the high number of nodes spread across a geographical area. ranging accuracy and target differentiation capability The nodes can be static, if deployed for, e.g., avalanche enabled by UWB, intelligent collision-avoidance and monitoring and pollution tracking, or mobile, if cruise-control systems can be envisioned. These sys- equipped on soldiers, firemen, or robots in military and tems can also improve airbag deployment and adapt emergency response situations. Key requirements for suspension/braking systems depending on road condi- sensor networks operating in challenging environments tions. UWB technology can also be integrated into include low cost, low power, and multifunctionality. vehicular entertainment and navigation systems by High data-rate UWB communication systems are well downloading high-rate data from airport off ramp, motivated for gathering and disseminating or exchang- road-side, or gas station UWB transmitters.
Table 1. FCC spectral mask for UWB systems.
Equivalent Isotropically Radiated Power (EIRP) [dBm] Frequency Indoor Hand Held Low Freq. High Freq. Med. Freq. Vehicular [MHz] Comm. Comm. Imaging Imaging Imaging Radar <960 15.209 limits 15.209 limits 15.209 limits 15.209 limits 15.209 limits 15.209 limits 960–1610 –75.3 –75.3 –65.3 –65.3 –53.3 –75.3 1610–1900 –53.3 –63.3 –53.3 –53.3 –51.3 1900–1990 –61.3 1990–3100 –51.3 –51.3 –41.3 –61.3 3100–10,600 –41.3 –41.3 –41.3 10,600–22,000 –51.3 22,000–29,000 –51.3 –61.3 –51.3 –51.3 –41.3 20,000–31,000 –51.3 >31,000 –61.3
Indoor/handled communication systems: fL , fH [3.1, 10.6] GHz; low-frequency imaging systems: fL , fH < 960 MHz; high-frequency imaging systems:
fL , fH [3.1, 10.6] GHz; medium-frequency imaging systems: fL , fH [1.99, 10.6] GHz; vehicular radar systems: fc > 24.075 GHz, fL , fH [22, 29] GHz.
28 IEEE SIGNAL PROCESSING MAGAZINE NOVEMBER 2004 UWB Communications at the Physical Layer −40 In this section, we outline physical layer issues of UWB −45 communication systems, including transmitter/receiver designs, synchronization, channel estimation, and mul- −50 tiple access schemes. In addition to the conventional 3.1 10.6 − single-band UWB transmissions, we will also discuss 55 1.99 recent multiband alternatives for power-efficient adher- −60 ence to FCC’s spectral mask, mitigation of narrowband interference (NBI), and relaxed sampling requirements. −65 Indoor Limit To appreciate UWB system designs, however, it is GPS Part 15 Limit − important to understand first the propagation charac- 70 Band UWB EIRP Emission Level in dBm teristics of the transmitted ultra-short waveform and −75 establish a realistic channel model. 0.96 1.61 100 101 Channel Modeling Frequency in GHz Since more than 80% of the envisioned commercial ▲ UWB applications will be indoor communications, we 1. FCC spectral mask for indoor commercial systems. will focus on indoor channels. The well-known Saleh- Valenzuela (S-V) indoor channel model was established firmed, the standardized channel model is basically a back in 1987 [50], based on measurements utilizing modified version of the S-V model [50]. To reach an low power ultra-short pulses (of width 10 ns and center analytically tractable channel model, the total number frequency 1.5 GHz) in a medium-size, two-story office of paths is defined as the number of multipath arrivals building. In the S-V model, multipath components with expected power within 10 dB from that of the arrive at the receiver in groups (clusters). Cluster strongest arrival. The Rayleigh distribution in the S-V arrivals are Poisson distributed with rate . Within channel model is replaced by the log-normal distribu- each cluster, subsequent arrivals are also Poisson dis- tion. The phases θm,n are also constrained to take values tributed with rate λ> . With αm,n denoting the gain 0 or π with equal probability to account for signal of the nth multipath component of the mth cluster, inversion due to reflection, yielding a real-valued chan- having phase θm,n, the channel impulse response can be nel model. With path gains normalized to have unit expressed as energy, a log-normal random variable is introduced to account for shadowing. Model parameters correspon- +∞ ding to several ranges are also provided in [14], for h(t) = α δ(t − τ ) l l both line-of-sight (LOS) and nonline-of-sight (NLOS) l =0 +∞ +∞ scenarios. The standardized UWB channel model in jθm,n [14] is claimed to better match the measurements. = αm,ne δ(t − Tm − τm,n), (1) m=0 n=0 However, the log-normal distribution and the shadow- ing factor render this model less tractable for theoreti- where Tm + τm,n (τm,0 = 0) denotes the arrival time of cal performance analysis and quantification of the the nth multipath component of the mth cluster, θm,n channel-induced diversity and coding gains. are independent uniform random variables over A typical realization of the channel impulse response [0, 2π), and αm,n are independent Rayleigh random generated using the channel model 2 [14] is shown in − /γ −τ /γ {α2 }= {α2 } Tm m,n variables with power E m,n E 0,0 e e , Figure 2. The number of multipath components is 315 where >γ. The number of clusters and multipath in this realization, spanning over a delay spread of about components may theoretically extend over infinite 50 ns. Let p(t) denote the transmitted pulse of duration time. However, the terms of the double sum in (1) Tp . After multipath propagation, the received waveform practically vanish for sufficiently large (m, n) with an is given by the convolution of p(t) with the physical ( ) exponentially decaying power profile. channel h t , and contains multiple delayed copies of ( ) ( ) = ( )( ) = L α ( − τ ) Lately, efforts have been made to characterize UWB p t ; i.e., g t : p h t l =0 l p t l , where channels with bandwidths exceeding 2 GHz. To come denotes convolution. Notice from Figure 2 that the {τ }L up with a statistical model, channel realizations are spacing among multipath delays l l =0 is in the order identified either in the frequency domain by frequency of nanoseconds. These delayed copies of p(t) in g (t) sweeping or in the time domain using impulsive signals. can be resolved from each other only if Tp is sufficiently In November 2002, the channel modeling subcommit- small, that is, if the transmit bandwidth B ≈ 1/Tp is tee of the IEEE 802.15.3a Task Group recommended sufficiently large. The transmitted and received wave- a channel model which captures the aforementioned forms corresponding to two Tp values are shown in works, as well as recent refinements [14]. Because the Figure 3, where the received waveforms were generated clustering phenomenon has been experimentally con- using the multipath delays and amplitudes of Figure 2.
NOVEMBER 2004 IEEE SIGNAL PROCESSING MAGAZINE 29 With Tp = 91 ns, we have B ≈ 11 MHz, and the Transmitter Design received waveform consists of a single distorted version Generally adopted spectrum shapers p(t) for UWB of p(t). With Tp = 0.55 ns, we have B ≈ 2 GHz, and communications include the Gaussian pulse, the the received waveform contains multiple resolvable Gaussian monocycle (first derivative of Gaussian pulse), copies of p(t). If the channel is known at the receiver, and the second derivative of the Gaussian pulse, as these resolvable copies can be combined coherently to depicted in Figure 4, along with their Fourier trans- provide multipath diversity. But before discussing forms (FTs). The reason behind the popularity of these receiver designs that collect this diversity, let us first pulses is twofold: i) Gaussian pulses come with the introduce UWB transmission schemes. smallest possible time-bandwidth product of 0.5, which maximizes range-rate resolution and ii) the Gaussian pulses are readily available from the antenna pattern [51]. With Tp at the subnanosecond 0.25 ( ) 0.2 scale, p t occupies UWB with 0.15 bandwidth B ≈ 1/Tp . As mentioned 0.1 before, and illustrated in Figure 0.05 3(b), such an ultra-short p(t) also 0 −0.05 gives rise to multiple resolvable
Amplitude −0.1 copies, and thus enables rich multi- −0.15 path diversity. −0.2 In a typical UWB system, each − 0.25 information-conveying symbol is 5101520253035404550 represented by a number of (N f ) Delay (ns) pulses, each transmitted per frame of duration Tf Tp . Having N f ▲ 2. A typical realization of the channel impulse response generated using the IEEE frames, over which a single symbol 802.15.3 a channel model 2: NLOS channels with 0−4 m transmitter-receiver distance. is spread, reverses the commonly used terminology where a frame consists of multiple symbols (here multiple frames comprise a symbol). With M -ary modulation, log M 1 2 0.8 Transmitted Pulse message bits are transmitted during 0.6 Received Waveform a signaling interval of duration 0.4 Ts = N f Tf s that corresponds to a 0.2 bit rate R = log M/T . 0 b 2 s −0.2 Being real, baseband UWB trans- Amplitude −0.4 missions neither have to entail fre- −0.6 quency modulation nor phase − 0.8 modulation with M > 2. Conse- 1 050100 150 quently, symbol values can be trans- Time (ns) mitted by modulating the position and/or the amplitude of p(t). In (a) M -ary pulse position modulation 1 (PPM), M distinctly delayed pulses Transmitted Pulse { ( − )}M −1 0.8 p t m m=0 are employed, each 0.6 Received Waveform representing one symbol value. 0.4 Generally, the so-termed modula- 0.2 0 tion indices m are chosen such that −0.2 m = m with ≥ Tp , which cor- Amplitude −0.4 responds to an orthogonal PPM. In −0.6 binary PPM, the delay can also be −0.8 −1 chosen to minimize the correlation ( ) ( − ) 012345678910 p t p t dt [52]. As band- Time (ns) width efficiency drops with increas- (b) ing modulation size M , PPM is suitable for power-limited applica-
▲ 3. The transmitted and received waveforms corresponding to (a) Tp = 91 ns and (b) tions. In fact, PPM was almost
Tp = 0.55 ns. For illustration purposes, only the first 10 ns are shown in (b). exclusively adopted in the early
30 IEEE SIGNAL PROCESSING MAGAZINE NOVEMBER 2004 10 Gaussian Pulse 1 Gaussian Pulse Gaussian Monocycle Gaussian Monocycle 0 Scholtz Monocycle 0.8 Scholtz Monocycle 0.6 −10
0.4 −20 0.2 − 0 30
−0.2 Power (dB) −40
Amplitude (dB) −0.4 −50 −0.6 −0.8 −60 −1 −70 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 100 101 Time (ns) Frequency (GHz) (a) (b)
▲ 4. (a) Generally adopted pulse shapes in UWB communications; (b) Fourier transform of several pulse shapes. Pulse width: 0.7 ns.
development of UWB radios because negating ultra- b u(k) = 0, (2) models M -ary PAM. With binary sym- short pulses were difficult to implement. Another mod- bols, and b u(k) = 0, au(k) = 2su(k) − 1 corresponds to ulation scheme that does not require pulse negation is BPSK, and au(k) = su(k) corresponds to OOK [32], the so termed on-off keying (OOK), where symbol “1” [43], [69]. is represented by transmitting a pulse, and “0” by With TH codes, MA is achieved by altering the transmitting nothing. pulse position from frame to frame, according to the TH( ) As pulse negation became easier to implement, pulse sequence c u n . MA can also be enabled by modify- amplitude modulation (PAM) attracted more attention. ing the pulse amplitude from frame to frame. The latter In particular, when M = 2, antipodal pulses are used to allows for many other choices of alternative spreading represent binary symbols, as in binary phase shift key- codes which, individually or in combination with TH ing (BPSK) or bipolar signaling. Biorthogonal signal- codes, give rise to the following transmitted waveform ing by combining orthogonal PPM with binary PAM as [c.f. (2)]: well as orthogonal waveform and block orthogonal +∞ modulation schemes have also been reported [43]. ( ) = E / ( ) To allow for multi-user access (MA) to the UWB vu t u au n N f c u n = channel, time hopping (TH) was introduced early in n 0 · − − TH( ) − / , [52]. With TH, each pulse is positioned within each p t nTf c u n Tc b u n N f ( ) frame duration Tf according to a user-specific TH 3 TH( ) sequence c u n . Specifically, dividing each frame into Nc chips each of duration Tc , the uth user’s TH code where c u(n) is the user-specific “amplitude code” dur- TH( ) ∈ , − c u n [0 Nc 1] corresponds to a time shift of ing the nth frame. Depending on the spreading codes TH( ) c u n Tc during the nth frame [72]. Consequently, employed, the UWB systems are termed TH-UWB the uth user’s transmitted waveform is given by [52], direct-sequence (DS)-UWB [15], or baseband single-carrier/multicarrier (SC/MC)-UWB [66], [77], +∞ just to name a few. ( ) = E / vu t u au n N f In addition to facilitating multiple access, spreading = n 0 codes also shape the transmit spectrum. Analytical · − − TH( ) − / , p t nTf c u n Tc b u n N f expressions and simulated power spectral density (2) (PSD) for UWB transmissions are pursued in [44] and [70]. A PSD expression for TH-UWB allowing where Eu is the uth user’s energy per pulse at the trans- for both short and long spreading codes is derived mitter end. With su(k) ∈ [0, M − 1] denoting the M - in [44]. The effects of timing jitter on PSD of UWB ary information symbol transmitted by the uth user transmissions utilizing random TH codes and/or during the kth symbol duration, (2) subsumes several M -ary modulation can be found in [70]. In particu- modulation schemes. When b u(k) = su(k), and lar, let us consider “short” spreading codes that are au(k) = 1, (2) describes UWB transmissions with M - symbol-periodic with period N f . Let us now define ary PPM; when au(k) = 2su(k) + 1 − M , and the symbol level transmitted waveform for user u as
NOVEMBER 2004 IEEE SIGNAL PROCESSING MAGAZINE 31 − ( ) (n) ( ) = N f 1 ( − − TH( ) ) φ n = δ(n), and φ ( f ) = 1, ∀n pT,u t : n=0 p t nTf c u n Tc for TH- aa bb N f −1 UWB, p , (t) := c (n)p(t − nT ) for DS/SC/ T u n=0 u − f ( ) = N f 1 ( − ) MC-UWB, and pT,u t : n=0 p t nTf when no for binary PAM and spreading code is involved. Equation (2) then becomes √ +∞ (n) vu(t) = Eu = au(k)pT,u(t − kTs − b u(k) ). Along φ = 1, ∀n, and k 0 aa the lines of [42, Chapter 4], it can be shown that the 1, n = 0 ( ) (n) PSD of vu t is φ ( f ) = 1 + cos(2π f ) bb , n = 0 2 E ∞ u 2 (n) (n) − j2π fnT ( ) = | , ( )| φ φ ( ) s ,( ) vv f PT u f aa bb f e 4 for binary PPM. Accordingly, the resultant PSD of Ts =−∞ 2 n vu(t) becomes [c.f. (4)]: vv( f ) = (Eu/Ts )|PT,u( f )| ( ) = (E / )| ( )|2 ( − for PAM and vv f u Ts P T,u f [ 1 φ(n) = { ( ) ( + )},φ(n)( ) = ( π ))/ + ( + ( π ))/( ) ∞ δ( − where aa : E au k au k n bb f : cos 2 f 2 1 cos 2 f 2Ts n=−∞ f − j2π f (b u (k)−b u (k+n)) E{e }, and PT,u( f ) := F{pT,u(t)} is k/Ts )] for PPM. the FT of pT,u(t). In particular, for independent identi- The shape of the vv( f ) for both modulations is cally distributed (i.i.d.) equiprobable binary symbols determined by PT,u( f ), which may contain spectral su(k) ∈{0, 1}, we have au(k) = 2su(k)−1, and b u(k) = 0 lines due to the repeated use of N f pulses in forming with PAM; and au(k) = 1, and b u(k) = su(k) with each symbol. Different from PAM, which is a zero- PPM. It then follows that mean linear modulation, additional spectral lines
10 10
0 0 − 10 −10 − 20 −20 − 30 −30
Power (dB) − 40 Power (dB) −40 − 50 −50 − 60 −60
−70 −70 024681012 024 6 81012 Frequency (GHz) Frequency (GHz) (a) (b)
10
0
−10
−20
−30
Power (dB) −40
−50
−60
−70 024681012 Frequency (GHz) (c)
▲ 5. Transmit PSD corresponding to (a) no spreading code, (b) random DS codes, and (c) random TH codes. Gaussian pulse with = . = = = Tp 0 4 ns. Nf 64, Tf 100 ns, Tc 2 ns.
32 IEEE SIGNAL PROCESSING MAGAZINE NOVEMBER 2004 emerge with PPM. Although the bandwidth of PT,u( f ) where is determined by the pulse shaper p(t), its shape, by L definition, relies not only on p(t), but also on the p (t) := α p (t − τ , ). (5) spreading codes utilized. The transmit PSD correspon- R l T l 0 l =0 ding to no spreading, random DS spreading, and ran- dom TH spreading with binary PAM are depicted in TH Figure 5. Notice how pronounced spectral lines are in Let us select Tf ≥ τL ,0 + Tp and c (0) ≥ TH Figure 5 where no spreading code is used. Figure 5(b) c (N f − 1) to confine the duration of pR(t) within and (c) confirms that both DS and TH spreading [0, Ts ) and thus avoid intersymbol interference (ISI). smooth the transmit PSD. It is also worth mentioning Without loss of generality, let us also confine the tim- that although the spectral lines cause “noise-like’’ inter- ing offset τ0 within a symbol duration, i.e., τ0 ∈ [0, Ts ). ference to narrowband services coexisting with UWB Notice that at the timing synchronization stage, the systems, they introduce cyclostationarity that one can multipath channel is unknown, and so is pR(t). exploit for timing offset estimation simply because Consequently, even with known training symbols a(k), these spectral lines originate from the underlying perio- the traditional approach of peak-picking the correlation dicity of the transmission. of r(t) with pR(t) to estimate τ0 is not applicable. Instead of pR(t), one could correlate r(t) with pT (t) and look for the maximum. Evidently, this approach is Timing Synchronization not only suboptimum in the presence of dense multi- As defined in (1), the physical channel h(t) not only path, but also results in unacceptably slow acquisition captures multipath effects but also includes a delay speed and has prohibitive complexity when one has to (uncertainty) on the first arrival time τ0. For coherent perform exhaustive search over thousands of demodulation, timing synchronization is the first task bins/chips. There is clearly a need for low-complexity to be performed at the receiver to acquire and track the timing estimation methods in multipath-rich propaga- timing offset τ0. tion settings. To this end, a number of attempts have In fact, one of the major challenges at the UWB been proposed to improve acquisition speed and/or physical layer is the accuracy and speed of this synchro- performance in UWB radios. These attempts include: a nization step. As UWB systems employ low-power coarse bin reversal search for the noiseless case [22]; a ultra-short pulses (in the order of nanoseconds), timing coded beacon sequence in conjunction with a bank of requirements are stringent because even minor mis- correlators in the context of data-aided localization in alignments may result in lack of energy capture which the absence of multipath [13]; a ranging system that renders symbol detection impossible [59]. To better requires knowledge of the strongest path [26]; a non- understand the unique challenges in UWB timing, let data aided (a.k.a. blind) timing estimator that relies on us consider a single-user link with zero-mean, i.i.d. the cyclostationarity that arises in UWB transmissions M -ary PAM symbols a(k) having zero-mean and vari- with slow TH and sufficiently dense multipath [61]; σ 2 ance s . (Generalization to PPM is also possible.) and a data-aided maximum likelihood (ML) timing Dropping temporarily the subscript u, the received sig- algorithm using symbol- and frame-rate samples [60]. √nal after multipath propagation is [c.f. (3)]: r(t) = E ∞ ( ) L α ( − − τ ) + η( ) k=0 a k l =0 l pT t kTs l t , where UWB Timing with Dirty Templates (L + 1) is the number of effectively nonzero paths, To relax the rather restrictive assumptions in these each with amplitude αl and delay τl satisfying approaches, we have developed timing algorithms τl <τl +1 , ∀l . We assume that the channel is quasi- based on the novel concept of “dirty templates’’ [79]. {α }L {τ }L static, meaning that l l =0 and l l =0 remain invari- The idea is to rely on pairs of successive symbol-long ant per burst but may change from burst to burst; a(k), segments of r(t) taken at candidate time-shifts {α }L {τ }L η( ) τ ∈ , ) l l =0 and l l =0 are independent of the noise t [0 Ts and have one segment in each pair serve as a which is assumed zero-mean, wide-sense stationary but template for the other. Specifically, integrate-and-dump not necessarily white and/or Gaussian, as it consists of operations are performed on products of such segments both ambient noise and multi-user interference (MUI). to obtain symbol-rate samples: With respect to the first arrival time (timing offset) τ0, T other path delays can be uniquely described as: s xk(τ) = r(t + kTs + τ)r(t + (k − 1)Ts + τ)dt, τl ,0 := τl − τ0, ∀l ∈ [0, L ]. It is convenient to express 0 ( ) ( ) r t in terms of the aggregate pulse pR t at the receiver (6) which encompasses the transmit pulse, spreading codes and multipath effects: ∀k ∈ [1, +∞) and τ ∈ [0, Ts ). The symbol-long seg- ments r(t + kTs + τ) and r(t + (k − 1)Ts + τ), for ∈ , ) √ ∞ t [0 Ts are “dirty templates’’ because: i) they are r(t) = E a(k)pR(t − kTs − τ0) + η(t), noisy, ii) they are distorted by the unknown channel, and k=0 iii) they are subject to the unknown offset τ0. The latter
NOVEMBER 2004 IEEE SIGNAL PROCESSING MAGAZINE 33 2(τ) cancel the cross terms in xk we used that { ( ) ( )}= { 2( ) 2( )}=σ 4 = UWB characterizes transmission E a k a l 0 and E a k a l s , for k l . E (τ)˜ + E (τ)˜ = Ts 2 ( ) = E Now notice that A B 0 pR t dt : R is systems with instantaneous the constant energy of the unknown aggregate tem- E (τ)˜ − E (τ)˜ spectral occupancy in excess plate at the receiver and also that B A is uniquely maximized at τ˜ = 0, since EA (τ)˜ is minimized of 500 MHz or a fractional at τ˜ = 0 and EB (τ)˜ is maximized at τ˜ = 0, by definition. { 2(τ)} τ˜ = bandwidth of more than 20%. So, E xk is maximized when 0, or equivalently, τ = τ0. Compactly written, the approach for nondata τ = τ∈ , ) { 2(τ)} aided TDT yields: 0 arg max [0 Ts E xk . constitutes a major difference between the Timing tech- As usual, the ensemble mean must be replaced in nique based on dirty templates (TDT) and the transmit- practice by its consistent sample mean estimator ted reference (TR) approach [21] for channel estimation obtained from K symbol-long pairs of received seg- −1 K 2 (τ) and symbol demodulation, as we will detail in the ensu- ments: K k=1 x2k−1 . The number of samples K ing sections. To grasp the gist of TDT, let x¯k(τ) and required for reliable estimation can be reduced marked- r¯(t) denote the noise-free parts of xk(τ) and r(t). ly if a data aided approach is pursued [79]. The training Applying Cauchy-Schwartz’s inequality to (6) yields sequence {a(k)} for data-aided TDT comprises a repeated pattern (s, s, −s, −s ) with s being any M -ary PAM sym- Ts bol. This pattern is particularly attractive because it simpli- ¯ 2(τ) ≤ ¯2( + + τ) (τ) = (− )k 2 −E (τ)˜ + E (τ)˜ + ξ( ) xk r t kTs dt fies (7) to xk 1 s [ A B ] k . As 0 −1 K 2 a result, K = x − (τ) converges faster to Ts k 1 2k 1 2 E{x 2(τ)}=s 4[E (τ)˜ − E (τ)˜ ]2 + σ 2 , because now one × r¯ (t + (k − 1)Ts + τ)dt, k A B ξ σ 4 0 obviates convergence to s , that is necessary in the blind approach to remove unknown symbol effects where the equality holds ∀k if and only if (self-noise). The benefit with data-aided TDT is very r¯(t + kTs + τ) = λr¯(t + (k − 1)Ts + τ). But the latter rapid acquisition since only K = 1 pair of received is true ∀t ∈ [0, Ts ) if and only if τ = τ0. In words, the symbol-long segments carrying as few as four training cross correlation of successive symbol-long received seg- symbols, is sufficient; see also Figure 6. Summarizing, ments reaches a unique maximum magnitude if and consistent (non)data-aided TDT can be accomplished only if these segments are scaled versions of each other, in the absence of ISI even when TH codes are present which is achieved only at the correct timing; i.e., when and the UWB multipath channel is unknown, using τ = τ0. In its simplicity, this neat observation offers a “dirty’’ Ts -long segments of the received waveform as distinct criterion for timing synchronization that, for follows [79]: years, has relied on the idea that the autocorrelation of the noise-free template has a unique maximum at the K 2 correct timing; the latter has been the principle behind 2kTs τˆ0 = arg max r(t + τ)r(t + τ − Ts ) dt . τ∈[0,T ) ( − ) all existing narrowband and UWB timing schemes, s k=1 2k 1 Ts including the popular early-late gate algorithm [42]. (8) The TDT approach is different from these conventional synchronizers. It has fundamental implications to UWB and also to a number of other applications such as time- Both training and blind modes have low-complexity as delay and displacement estimation through unknown they require only symbol rate samples, but the data- time-dispersive media, where there is no undistorted aided mode enjoys also rapid acquisition relying on as template to rely on, even in the noise-free case. few as four training symbols [s, s, −s, −s ]. To establish its validity even in the presence of noise, The estimator in (8) enables timing synchronization we start by recalling that pR(t) has finite nonzero sup- at any desirable resolution constrained only by the port [0, Ts ). This implies that the noisy xk(τ) can be affordable complexity: i) coarse timing with low com- expressed as [c.f. (5) and (6)]: plexity, e.g., by picking the maximum over N f candi- date offsets τ = nTf , where integer n ∈ [0,N f ); ii) fine x (τ) = a(k) [a(k − 1)E (τ)˜ + a(k + 1)E (τ)˜ ] + ξ(k), timing with higher complexity at the chip resolution k A B τ = ∈ , ) ( ) with iTc , i [0 Nc ; and iii) adaptive timing 7 (tracking) with voltage-controlled clock (VCC) circuits; the preliminary tests in Figure 6 correspond to frame where ξ(k ) denotes noise after correlation, level timing resolution. As the TDT algorithms only Ts Ts −˜τ E (τ)˜ := E p2 (t)dt,E (τ)˜ := E p2 (t)dt and require zero ISI, the condition T ≥ τ , + T can be A Ts −˜τ R B 0 R f L 0 p τ˜ := (τ0 − τ) mod Ts . The mean-square of xk(τ) in relaxed to allow for higher data rates, as long as guard ( ) { 2(τ)}=σ 4 E2 (τ)˜ + E2(τ)˜ + σ 2 = (σ 4/ ) 7 is E xk s [ A B ] ξ s 2 frames are inserted between symbols to avoid ISI, much 2 2 2 ×{[EA (τ)˜ + EB (τ)˜ ] + [EA (τ)˜ − EB (τ)˜ ] }+σξ , where to like zero-padding in narrowband systems [67].
34 IEEE SIGNAL PROCESSING MAGAZINE NOVEMBER 2004 0 10 100 K = 1 K = 1 K = 2 K = 2 K = 16 −1 K = 16 − 1 10 10 K = 64 K = 64 No Timing 10−2
10−2 Perfect − 10 3 Timing
Normalized MSE With
Average BER Timing 10−3 10−4 Acquisition
−5 10−4 10 −50 5101520 051015 20 25 30 35 SNR Per Pulse (dB) SNR Per Symbol (dB) (a) (b)
▲ 6. (a) Normalized MSE of τˆ0 in (8) and (b) average BER using TDT algorithm. Dashed (solid) curves correspond to (non) data-aided mode.