Constrained Pole-Zero Linear Prediction:
Optimization of Cascaded Biquadratic Notch Filters for Multi-Tone Frequency Estimation
Toon van Waterschoot and Marc Moonen
ESAT/SCD
Abstract
Constrained pole-zero linear prediction (CPZLP) is a new method for parametric frequency estimation of multiple real sinusoids in noise. The method is based on a signal model that consists of a cascade of constrained biquadratic models, thereby exploiting the linear prediction property of sinusoidal signals. The signal model is parametrized directly with the unknown frequencies, which are then estimated using a numerical optimization approach. By independently optimizing each biquadratic stage in the cascade model, a computationally efficient algorithm is obtained which has linear complexity. The linear complexity allows for using relatively long data records, leading to high accuracy even at low signal-to-noise ratios (SNR).
Multi-Tone Frequency Estimation
Problem definition: Given a length-L data sequence y (t) =
N X n=1
α n cos(ω n t + φ n ) + r(t), t = 1, . . . , L
estimate the unknown frequencies ω n , n = 1, . . . , N .
State-of-the-art:
Non-parametric methods:
• typically FFT-based
• limited accuracy: resolution ∼ L
• reasonable complexity: O(L log L) Parametric methods:
• ML, nonlinear LS, total LS, subspace (MUSIC, ESPRIT), ...
• high accuracy: CRLB is achieved ∀ L > L min (SNR)
• high complexity: O (L 2 ) or higher
Proposed method:
• parametric method based on constrained pole-zero model
• reasonable accuracy: CRLB is approached ∀ L > L min (SNR)
• low complexity: O (L)
Constrained Pole-Zero Linear Prediction
Signal model: cascade of N constrained biquadratic models y (t) =
N Y n=1
1 − 2ρ cos θ n z − 1 + ρ 2 z − 2 1 − 2 cos θ n z − 1 + z − 2
e(t)
Goal:
θ n → ω n , n = 1, . . . , N Global objective:
min θ
1 L
L X t=1
e 2 (t, θ)
Decoupled objectives:
min θ n
1 L
L X t=1
e 2 n (t, θ n ), n = 1, . . . , N
Im j
− 1
− j
e jθ
n1 Re θ n
ρe jθ
ne − jθ
nρe − jθ
nwith θ n = h θ 1 . . . θ n i T and e n (t, θ n ) =
n Y l=1
1 − 2 cos θ l z − 1 + z − 2 1 − 2ρ cos θ l z − 1 + ρ 2 z − 2
y(t)
Decoupled optimization: If N subproblems are solved consec- utively, then estimates of θ 1 , . . . , θ n−1 are available when solving subproblem n ⇒ each subproblem reduces to a scalar problem!
Line search algorithm for nth subproblem:
θ n (k+1) = θ n (k) + µ k p (k)
• step length µ k ⇒ backtracking with Armijo’s sufficient de- crease condition
• search direction p (k) ⇒ quasi-Newton with damped BFGS up- dating:
p (k) = −B k − 1 ∂
∂θ n V n θ ˆ (k)
n
B k +1 = max
v k
s k , γB k
= scalar BFGS update
with s k = ˆ θ n (k+1) − ˆ θ n (k) and v k = ∂θ ∂
n V n θ ˆ (k+1)
n
− ∂
∂θ n V n θ ˆ (k)
n
Performance Evaluation
Computational complexity: number of multiplications M BFGS = ¯ κN h (13 + 3 ¯ β )L + (17 + 5 ¯ β ) i
¯
κ: average no. iterations per subproblem
β ¯ : average no. backtracking steps per iteration per subproblem
Frequency variance:
averaged over 100 Monte Carlo runs
10
210
3−120
−100
−80
−60
−40
−20
Frame length N (samples)
F re q u en cy va ri an ce (d B )
θ
1θ
2θ
3CRLB( ω
1) CRLB(ω
2) CRLB( ω
3)
−10 −5 0 5 10 15 20 25 30 35 40
−140
−120
−100
−80
−60
−40
−20 0
SNR (dB)
F re q u en cy va ri an ce (d B )
θ
1θ
2θ
3CRLB(ω
1) CRLB(ω
2