K. C. Chen, for NTUEE Course “Mobile Communications” Only 1 2 Frequency Estimation Once the receiver lowers the frequency into baseband range, the immediate challenge is to achieve synchronization so that other signal processing and detection/decoding can proceed. As the received waveform has ambiguity in time, frequency/phase/carrier, we have to decide what to do first to effectively extract useful signal. This is a critical and fundamental step to design receiver architecture and algorithms. In case the sampling period is Ts and we suppose the overall phase θ T (t ) = Ωt + θ , the maximal frequency ambiguity can be represented as Ω max . Frequency offset/ambiguity can come from oscillator error or instability between transmitter and receiver, LNB, Doppler spread due to fading and mobility, etc. We may come out a rule of thumb: in case | ΩT |<< 1 , we usually first recover the timing and perform frequency estimation in a time-directed (TD) mode. Otherwise, we estimate frequency first to recover carrier, then recover timing, which is known as non-time-directed (NTD) mode. Along with well known data-aided (DA) (i.e. with pilot tones/signals or preamble as reference signal) and non-data-aided (NDA), we can proceed frequency estimation in different combination of ways. 2.1 NDA-NTD Frequency Control The most fundamental frequency estimation might be non-data-aided and non-time-directed (NDA-NTD), which has been used in a wide range of communications, broadcasting, and other application scenarios. In this section, we are introducing a few technological milestones in automatic frequency control (or tracking) in digital communications. 2.1.1 Costas Loop The most well known frequency/carrier recovery scheme might be Costas loop. K. C. Chen, for NTUEE Course “Mobile Communications” Only 2 Costas Loop LPF cos VCO Loop Filter Data Decision 90o phase shift sin Mobile Communications LPF KC Chen, NTU EE 1 Figure 2-1: Costas Loop The major application constraint for Costas loop is small loop bandwidth like the same reason for typical phase-locked loop (PLL), which suggests small pull-in and locking frequency range. In case the maximum possible frequency deviation is only a small fraction of signal bandwidth, Costas loop is simple and effective. 2.1.2 Automatic Frequency Control (AFC) Algorithms Generally speaking, automatic frequency control (AFC) or frequency estimation or tracking can be considered as the general structure of the following figure. K. C. Chen, for NTUEE Course “Mobile Communications” Only 3 I X H(f) D(f) X AFC Discriminator (Frequency Detector) & Filtering r(t) X H(f) D(f) + + - e(t) X Q VCO Loop Filter Figure 2-2 General AFC Loop Configuration The key of design is how to realize the AFC discriminator. Traditionally, there are a few common approaches: n n n Delay-and-multiply Cross-product Discrete Fourier transform In many cases, synchronizers adopt non-linearity to remove modulated signal information and to generate desirable spectrum features. Delay-and-multiply is a well known structure for this purpose. The S-curve of the differentiate-and-multiply AFC (AFC-Di) has pull-in region around 0.75 BLP and linear contraction region larger than 0.5 BLP , where BLP is the bandwidth of low pass filter. In digital realization, we are likely to use interpolation of samples to animate differentiation, which would result in some reduction of pull-in region and lock-in region. K. C. Chen, for NTUEE Course “Mobile Communications” Only 4 I X LPF d/dt X + + - r(t) X LPF d/dt e(t) X Q VCO Loop Filter Figure 2-3 Differentiating AFC A more practical digital realization is to replace differentiators by delays. Since the frequency is the differentiation of phase, it is also very reasonable for such a substitution. The LPFs of AFC-Di are implemented by integrate-and-dump circuits that can be considered as a sort of LPF, and we call this scheme as cross-product AFC (AFC-CP). AFC-CP is very suitable for digital implementation and suitable for tracking CW or FSK signals. However, the effective loop bandwidth to pull-in and to lock-in signal is less than that of AFC-Di. Typical discriminator curve can effectively pull-in and lock-in around 0.2 / δ . To more fit QPSK signals, we may introduce an angle doubling mechanism by squaring operation and other techniques discussed in later section. DFT approach requires timing information to execute DFT (or FFT in practical implementation) and we will discuss in NDA-TD section. K. C. Chen, for NTUEE Course “Mobile Communications” Only I kδ 2BLP∫ X (k−1)δ Delay δ dt 5 X + + - r(t) X Q kδ 2BLP∫ (k−1)δ Delay δ dt X N ∑ Loop Filter VCO e(t) 1 Figure 2-4 AFC-CP 2.1.3 Quadri-correlator Frequency Control In [1], two frequency detector structures (quadriccorrelator and rotational frequency detector) are described, and later modified into widely applied AFC for possible large frequency deviation, which will be discussed hereafter. FDD I X H(f) D(f) X + + - r(t) X H(f) D(f) Q VCO Figure 2-5 Loop Filter X e(t) K. C. Chen, for NTUEE Course “Mobile Communications” Only 6 Quadricorrelator AFC has a balanced structure and is very similar to AFC-CP, except different filter design. Such an AFC family is effective and robust in tracking frequency. Compared with Costas loop, such a mechanism has wider frequency pull-in and locking range, at half of the signal bandwidth (or symbol rate more precisely). 2.1.4 Maximum Likelihood Frequency Estimation In certain wireless communication systems (most likely in satellite communications), frequency deviation can be significant and even larger than signal bandwidth. Effective frequency estimation is greatly needed to ensure coherent communications usually with better performance. In the presence of frequency offset Ω , phase process could be modeled as θ T (t ) = Ωt + θ . It is required that the frequency range | ω |≤ 2πB + | Ω max | (2.10) where B is 1-side signal bandwidth and ± Ω max is the maximum frequency uncertainty. The sampling rate must satisfy |Ω | 1 > B + max 2Ts 2π (2.11) s f (t , Ω) = s (t ) ⊗ c (t ) ⊗ f (t )e jΩt is the frequency-translated signal s f (t )e jΩt , and N −1 s f (t , Ω) = ∑ an g (t − nT − εT )e j (Ωt +θ ) (2.11) n =0 The matched filter output z n (ε , Ω) = ∞ ∑r k = −∞ f (kTs )e − jΩkTs g MF (nT + εT − kTs ) (2.12) K. C. Chen, for NTUEE Course “Mobile Communications” Only 7 Modulator s(t )e jθT ( t ) C ( f ) or c(t ) Channel Pre-filtering F(f) or f(t) ˆ εˆ gMF (kTs ) X kT s Detection e − jθ Interpolator Decimator X ˆ e − jΩkTs Figure 2-6 Frequency Estimation For NTD case, we start from low SNR approximation of likelihood function. L(ε , Ω) = L ∑ | z (lT + εT , Ω) | l =− L 2 = c0 + 2 Re{c1e j 2πε } + random disturbanc e (2.13) The value Ω̂ which maximizes the coefficient c0 (Ω) is an unbiased estimate, as the following proposition. Proposition 2.1: Under the conditions, (i) Sampling rate fulfills 1 / Ts > 2(1 + α ) / T (i.e. twice the rate required for data (ii) (iii) path) T / Ts = M s is an integer Data {an } i.i.d. Then, ˆ = arg max Ω Ω LM s −1 ∑ | z (lT , Ω) | l = − LM s defines an unbiased estimate of frequency. s 2 (2.14) K. C. Chen, for NTUEE Course “Mobile Communications” Only 8 Consequently, our exploration of maximum likelihood (ML) frequency estimator is similar to the ML timing estimator. We usually do NOT implement the ML frequency estimator. Instead, we adopt the following PLL-type structure to (sub-)optimally realize ML frequency estimator. Signal Matched Filter r(t) FDD Error Generator X e(k) Frequency Matched Filter fˆk ,θˆk VCO Loop Filter Figure 2-7 Above structure is based on the maximum-likelihood principle. The key challenge is how to implement the filters. In [1], rotational frequency detector (FD) was suggested to realize, which is more suitable for digital realizations. Rotational FD has no filtering functions, and decides by observing transitions of relative phases. If the received waveform carrier has frequency f c and estimated frequency fˆc . When fˆc = f c , the transitions of f c maintain a fixed relationship to fˆ those of fˆc . Otherwise, the angle of rotation is shown to be 2π ( c − 1) , which is fc counterclockwise if fˆc > f c and clockwise if fˆc < f c . By “measuring” phase change, we can track the frequency. K. C. Chen, for NTUEE Course “Mobile Communications” Only The useful range of the FD is | ∆ω |< πf c = 9 ωc and the largest FD (linearity) output is 2 ωc / 4 . 2.2 Digital Signal Processing Carrier Frequency Recovery Schemes that we discussed in earlier section are basically with the closed loop structure to recover carrier frequency. With advance of microprocessors and digital signal processors, a new thinking based on processing digital samples of the received waveform has been introduced into digital communication transceiver design, and has been widely applied in many modern digital communication systems. As packet communications becomes a main stream in packet switching networks (such as Internet), bursty digital transmission based on packets or frames suggests importance of open loop carrier frequency estimation. 2.2.1 Nonlinear Estimation of Modulated Carrier Viterbi and Viterbi [8] developed a non-linear estimation of PSK modulated carrier phase for such a purpose. I ( n +1 / 2 ) T xsn ( n −1 / 2 )T Optimal Estimate Function For Certain y sn Modulation ∫ X r(t) X Q VCO Figure 2-8 ∫ ( n +1 / 2 ) T ( n −1 / 2 )T xn N X 1 ∑ 2 N + 1 n=− N 1 Y tan −1 m X yn N 1 Y ∑ 2 N + 1 n=− N 1 Y θˆ(m) = tan −1 for m − PSK m X θˆ K. C. Chen, for NTUEE Course “Mobile Communications” Only 10 In above figure, xn = x sn and y n = y sn represents m = 1 and ∆f = 0 for un-modulated 1 Y carrier. This estimator θˆ(m) = tan −1 for m − PSK m X variance of the phase estimator is var(θˆ) = is unbiased and the 1 [1 + O(2 N + 1) −1 ] (2 N + 1)(2 ES / N 0 ) while the variance corresponds to the Cramer-Rao lower bound. This linear case is not useful for m-PSK modulated waveform, and we shall have a nonlinear function as typical synchronization solutions, which can be modeled as xn + jy n = F ( ρ n )e jmϕ n where ρ n2 = xn2 + y n2 and ϕ n = tan −1 ( y n / x n ) . Multiplying the phase by m, along with the final operation of dividing tan −1 by m, gives m-fold phase ambiguity of carrier phase estimator. A practical way to adjust this phase ambiguity is through the (error correcting) coding, rather than data itself. Typical selections are F ( ρ ) = 1, ρ 2 , ρ 4 . Finally, we may note the similarity between the selection of F(.) and square-loop etc. 2.2.2 Frequency Offset Estimation Let us consider the ML estimation of frequency ∆f of a complex-valued oscillation by observing the sampled signal [9] rk = e j ( 2π∆fkTS +θ ) + υ k , 1 ≤ k ≤ N where TS ≤ 1 /( 2∆f ) is the sampling interval, and θ is the unknown random phase uniformly distributed in (0, 2π ) , each component of complex υ k is independent Gaussian with autocorrelation Rυ (k ) = σ 2δ k . This is a well-known problem to detect frequency in modern spectral estimation, and it leads to the maximization of equivalent likelihood function K. C. Chen, for NTUEE Course “Mobile Communications” Only 11 N N N ~ ~ ~ Λ(∆f ) =| ∑ ri e − j 2π∆fTS |2 = ∑∑ rk rl* e − j 2π∆fTS ( k −l ) i =1 (2.20) k =1 l =1 ~ where ∆f is a tentative value for ∆f . Above maximization leads to ML estimate ∆f ML that is consistent and asymptotically efficient. For N →∞ and ρ = 1 /(2σ ) as carrier-to-noise ratio, the Cramer-Rao lower bound is 2 2 σ CR = 3 1 2 2 2π TS ρN ( N 2 − 1) Unfortunately, (2.20) does not have a closed form solution. Considering only ~ necessary condition and taking derivative of (2.20) with respect to ∆f , we can get N N ∑∑ (k − l )r r e k =1 l =1 ~ * − j 2π∆f TS ( k −l ) k l =0 By defining R(k ) as the estimated autocorrelation of rk , that is, R(k ) = N 1 ri ri*−k 0 ≤ k ≤ N − 1 ∑ N − k i =k +1 The necessary condition is therefore ~ N −1 Im∑ k ( N − k ) R(k )e − j 2π∆fkTS = 0 k =1 (2.24) Please note that the possibility of “false maxima” can be avoided by appropriate restricting the operating range of the estimator. We may further note that (2.24) can be considered as the DFT of R(k ) , weighted by a parabolic windowing function w(k ) = k ( N − k ) . Via replacing a sub-optimal window with uniform weights, we can have ~ M Im∑ R(k )e − j 2π∆fkTS = 0 k =1 (2.25) ~ For an ideal noiseless channel, R(k ) = e − j 2π∆fkTS and ∆f = ∆f is still a trivial solution K. C. Chen, for NTUEE Course “Mobile Communications” Only 12 of (2.25). When the noise is present, with proper choice of M, the mean squared distance is expected to be negligible as CNR increases. For high CNR and low frequency deviation ( M∆fT << 1 ), replacing the exponential in (2.25) by its Taylor series expansion truncated to the linear term, we can reach M ∆fˆ ≅ 1 2πTS ∑ Im{R(k )} k =1 M ∑ k Re{R(k )} (2.26) k =1 If we consider assumptions and use Taylor expansion (to linear term with small noise) of R(k ) , we can argue M R k ≅ Im { ( ) } arg ∑ R(k ) ∑ k =1 k =1 M M ( M + 1) k Re{R(k )} ≅ ∑ 2 k =1 1 M M (2.27) It leads to a simple estimate ∆fˆ ≅ 1 M arg∑ R(k ) πTS ( M + 1) k =1 (2.28) It is true only when the first equation of (2.27) not exceeding π . Consequently, the operating range of this frequency estimation is limited as | ∆f |< 1 TS ( M + 1) (2.29) References: [1] D.G. Messerschmidt, “Frequency Detectors for PLL Acquisition in Timing and Carrier Recovery”, IEEE Tr. On Communications, Sep. 1979, pp.1288-1295. [2] F.D. Natali. “AFC Tracking Algorithms”, IEEE Tr. On Communications, Aug. 1984, pp.935-947. [3] T. Alberty, V. Hespelt, “A New Pattern Jitter Free Frequency Error Detector”, IEEE Tr. On Communications, Feb. 1989, pp.159-163. K. C. Chen, for NTUEE Course “Mobile Communications” Only 13 [4] A.N. D’Andrea, U. Mengali, “Performance of a Quadricorrelator Driven by Modulated Signals”, IEEE Tr. On Communications, Nov. 1990, pp.1952-1957. [5] ------, “Design of a Quadricorrelator for Automatic Frequency Control Systems”, IEEE Tr. On Communications, June 1993, pp.988-997. [6] ------, “Noise Performance of two Frequency-Error Detectors Derived from Maximum Likelihood Estimation Methods”, IEEE Tr. On Communications, Feb/Mar/Apr 1994, pp.793-802. [7] G. Karam, I. Jeanclaude, H. Sari, “A reduced-Complexity Frequency Detector Derived from the Maximum-Likelihood Principle”, IEEE Tr. On Communications, Oct. 1995, pp.2641-2650. [8] A. Viterbi, A. Viterbi, “Nonlinear Estimation of PSK-Modulated Carrier Phase with Application to Burst Digital Transmission”, IEEE Trans. On Information Theory, no. 4, vol.29, July 1983. [9] M. Luise, R. Reggianni, “Carrier Frequency Recovery in All-Digital Modems for Burst-Mode Transmissions”, IEEE Trans. On Communications, no. 2/3/4, vol. 43, Feb/Mar/Apr 1995. [10] H. Meyr, Digital Communication Receiver, 1999.