FACTA UNIVERSITATIS (NI

S)
SER.: ELEC. ENERG. vol. 19, no. 3, December 2006, 453-464
Closed-FormVariance Formula of the RPHD Single-Tone
Frequency Estimator
Liviu Toma, Aldo De Sabata, and Septimiu Mischie
Abstract: In this paper, we provide a very convenient, closed-form formula for the
variance of the single tone frequency estimator in the reformed Pisarenko harmonic
decomposer method. Several computer experiments show that calculated and mea-
sured variances are in excellent agreement. Our formula is useful when a small num-
ber of signal samples are used for frequency estimation, as asymptotic forms can lead
to important errors.
Keywords: noisy real sinusoid, frequency estimation, RPHD method, variance ana-
lysis
1 Introduction
The problem of the frequency estimation of a single real sinusoid from a nite
number of noisy data samples is relevant for various applications, such as speech
analysis, radar, sonar, communication systems, measurements, and adaptive control
[13]. To solve this problem, many frequency estimation techniques have been
proposed and analyzed, for example the modied covariance (MC) method [4, 5]
and the Pisarenko harmonic decomposer (PHD) method [68].
In a recently proposed method, called reformed Pisarenko harmonic decom-
poser (RPHD) [911], a closed-form, asymptotically unbiased frequency estimator
is proposed and analyzed, based on the linear prediction (LP) property of sinusoidal
signals, and on a modied least-squares (LS) cost function. The same frequency
estimator is derived in [12] from a constrained notch-lter point of view.
Manuscript received May 26, 2006.
The authors are with the Politehnica University of Timisoara, Faculty of Electro-nics
and Telecommunications, Bd. Vasile Parvan nr. 2, 300223 Timisoara, Romania (e-mails:
[liviu.toma, aldo.desabata, septimiu.mischie]@etc.upt.ro).
453
454 L. Toma, A. De Sabata, and S. Mischie:
In this paper we derive a closed-form expression of the RPHD variance, based
on the same variance analysis technique used in [11, 13]. We obtained a much more
convenient formula as compared to [11].
In section 2, the RPHD method for single-tone frequency estimation is re-
viewed based on [911]. The variance analysis technique, the main steps and
results of the closed-form formula derivation for RPHD variance are presented in
Section 3. Details of this derivation are developed in the Appendix. Results of com-
puter experiments, presented in Section 4, show that the measured RPHD variance
agrees with the analytical calculations based on the derived formula. Conclusions
are drawn in the last section.
2 RPHD method
In single frequency estimation, the following signal model is used [11]:
x(n) = s(n) +q(n) = cos(
0
n+) +q(n), n = 0, 1, 2, ...N (1)
where >0,
0
(0, ), and are the unknown amplitude, frequency and phase of
the sinusoid, and q(n) is a zero-mean white noise, which we will suppose Gaussian.
The sinusoid is linearly predictable from the past samples:
s(n) = 2cos(
0
)s(n1) s(n2), (2)
a fact that allows to dene an error function
e(n) = x(n) 2cos()x(n1) +x(n2), (3)
where is the parameter to be determined.
In order to obtain an asymptotically unbiased frequency estimator, a modied
error function is dened in [11] as follows:
(n) =
e(n)
_
2
_
2+cos(2)
_
. (4)
The corresponding LS cost function is
J
N
() =
N

n=3
e
2
(n)
2
_
2+cos(2)
_. (5)
Solving
dJ
N
()
d
= 0 yields [911]:
2A
N
cos
2
() B
N
cos() A
N
= 0 (6)
Closed-FormVariance Formula of the RPHD Single-Tone Frequency Estimator455
where
A
N
=
N

n=3
_
x(n) +x(n2)
_
x(n1) (7)
and
B
N
=
N

n=3
_
_
x(n) +x(n2)
_
2
2x
2
(n1)
_
. (8)
The root of (6) that provides the estimate is

=
B
N
+
_
B
2
N
+8A
2
N
4A
N
. (9)
The frequency estimate, which is denoted by
0
, is computed as

0
= cos
1
(

). (10)
3 RPHD variance development
The variance analysis technique [11] is based on dening a second order polyno-
mial
f () = 2A
N

2
B
N
A
N
(11)
and utilizes the following formulas
var{

}
E{f
2
()}
_
E{f

()}
_
2

=cos(
0
)
(12)
var(
0
)
var(

)
sin
2
(
0
)
. (13)
By using
= cos(
0
) (14)
and (11) there results
f
2
()

=cos(
0
)
= cos
2
(2
0
)A
2
N
2cos(
0
)cos(2
0
)A
N
B
N
+cos
2
(
0
)B
2
N
(15)
f

()|
=cos(
0
)
= 4cos(
0
)A
N
B
N
. (16)
In order to compute the variance of
0
using (12) and (13), the values of E{A
N
},
E{B
N
}, E{A
2
N
}, E{B
2
N
}, and E{A
N
B
N
} are required. The main steps that we fol-
lowed in computing these terms are presented in the Appendix. We obtained:
456 L. Toma, A. De Sabata, and S. Mischie:
E{A
N
} =
2
_
N2+(2, N 1)
_
cos(
0
) (17)
E{B
N
} =
2
_
N2+(2, N 1)
_
cos(2
0
) (18)
E{A
2
N
} =
4
_
N2+(2, N 1)
_
2
cos
2
(
0
)
+
2

2
q
_
4N 10+(4N 12)cos(2
0
) +(2, N 1)
_
2+cos(2
0
)
_
+2(2, N 2) +2(3, N 1) +(3, N 2)

+
4
q
(4N 10) (19)
E{B
2
N
} =
4
_
N2+(2, N 1)
_
2
cos
2
(2
0
)
+
2

2
q
_
4N +4(N 4)cos(4
0
) +8(2, N 1)
_
2+cos(2
0
)
_
16(2, N 2) 16(3, N 1) +8(3, N 2)
_
1+cos(2
0
)
_
+
4
q
(4N 8) (20)
E{A
N
B
N
} =
4
_
N2+(2, N 1)
_
2
cos(
0
)cos
2
(2
0
)
+
2

2
q
_
(4N 14)
_
cos(
0
) +cos(3
0
)
_
+2
_
(2, N 2) +(3, N1)
_
cos(2
0
)
cos(
0
)
+4(3, N 2)cos(
0
)

(21)
where
(k
1
, k
2
) =
k
2

n=k
1
cos
_
2(
0
n+)
_
=
sin
_

0
(k
2
k
1
+1)
_
cos
_

0
(k
2
+k
1
) +2
_
sin(
0
)
(22)
By using (12)...(21), and after denoting the signal-to-noise ratio
SNR =

2
2
2
q
, (23)
we obtained the closed-form variance formula of the RPHD single-tone frequency
estimator as
var{
0
}

2+2
_
2+cos(2
0
)
_
(2, N 1) 2(2, N 2) 2(3, N 1) +(3, N 2)
2SNR
_
N 2+(2, N1)
_
2
sin
2
(
0
)
+
(2N 5)cos
2
(2
0
) +(2N 4)cos
2
(
0
)
2SNR
2
_
N2+(2, N 1)
_
2
_
2+cos(2
0
)
_
2
sin
2
(
0
)
.
(24)
Closed-FormVariance Formula of the RPHD Single-Tone Frequency Estimator457
The shape of the variance in (24) is more convenient and it has a simpler struc-
ture when compared to the original results presented in [11]. In that paper, an
asymptotic form of the variance is also considered. That form can be approached
by making = 0 in (24):
var
asympt
{
0
}
1
SNR(N2)
2
sin
2
(
0
)
+
(2N 5)cos
2
(2
0
) +(2N 4)cos
2
(
0
)
2SNR
2
(N 2)
2
_
2+cos(2
0
)
_
2
sin
2
(
0
)
(25)
Minor differences are probably caused by small calculation errors in the original
work.
4 Numerical examples
In order to conrm the expression (24) for the variance, and to evaluate its asymp-
totic form (25), we have performed some computer experiments. RPHD variances
have been measured for data sequences we have generated using (1), with =

2,
and several values for
2
q
. In every experiment we computed the frequency esti-
mate using (9) and (10) for 500 independent runs, and we evaluated the measured
frequency variance of the RPHD method in terms of the mean square frequency
errors.
We used for evaluation purposes the Cramer-Rao lower bound (CRLB) for the
frequency estimator of a single sinusoid [3]
CRLB =
24
2
q
N(N
2
1)
2
. (26)
The measured frequency variances with the RPHD method, the theoretical fre-
quency variances calculated with (24), the asymptotic forms calculated with (25),
and the CRLB are shown in every gure that illustrates the experimental results.
In Fig. 1 and Fig. 2, the frequency variances versus
0
are represented, for
N = 20, SNR = 20 dB, = 0, and = /4 respectively. The measured variances
are very close to the variances calculated with (24), and uctuate in function of
frequency and phase around the curve representing the asymptotic variance (they
are not always symmetric around
0
= /2; (k
1
, k
2
) is symmetric around
0
=
/2 for = 0, but it is not for = /4).
In Fig. 3 and Fig. 4, the frequency variances versus SNR are represented, at

0
= 0.2, N = 20, = 0, and = /4 respectively. It can be noticed that, like
for all methods, errors increase at low values of the SNR.
458 L. Toma, A. De Sabata, and S. Mischie:
0 0.2 0.4 0.6 0.8 1
50
45
40
35
30
25
20

0
/
f
r
e
q
u
e
n
c
y

v
a
r
i
a
n
c
e

(
d
B

r
a
d
2
)


RPHD
variance by (24)
asympt. variance by (25)
CRLB
Fig. 1. Frequency variances versus
0
at SNR=
20 dB, N = 20 and = 0.
0 0.2 0.4 0.6 0.8 1
50
45
40
35
30
25

0
/
f
r
e
q
u
e
n
c
y

v
a
r
i
a
n
c
e

(
d
B

r
a
d
2
)


RPHD
variance by (24)
asympt. variance by (25)
CRLB
Fig. 2. Frequency variances versus
0
at SNR=
20 dB, N = 20 and = /4.
0 5 10 15 20 25 30 35 40
70
60
50
40
30
20
10
0
SNR (dB)
f
r
e
q
u
e
n
c
y

v
a
r
i
a
n
c
e

(
d
B

r
a
d
2
)


RPHD
variance by (24)
asympt. variance by (25)
CRLB
Fig. 3. Frequency variances versus SNR at
0
=
0.2, N = 20 and = 0.
0 5 10 15 20 25 30 35 40
70
60
50
40
30
20
10
0
SNR (dB)
f
r
e
q
u
e
n
c
y

v
a
r
i
a
n
c
e

(
d
B

r
a
d
2
)


RPHD
variance by (24)
asympt. variance by (25)
CRLB
Fig. 4. Frequency variances versus SNR at
0
=
0.2, N = 20 and = /4.
Fig. 5 and Fig. 6 show the frequency variances in function of N for
0
= 0.2,
SNR = 20 dB, = 0, and = /4 respectively. The coincidence between the
measured and the calculated variances is remarkable again. It can be seen from
(24) and (25) that, while the calculated variance depends on , its asymptotic form
does not. Figures 5 and 6 show that important differences between the two can
occur in function of , for the small values of N that are used in practice and in our
experiments, and attenuate relatively when N becomes large.
5 Conclusions
In this contribution we provided a very convenient, closed-form expression for the
variance of the frequency estimator in the RPHD method of single-tone frequency
estimation from a nite set of data samples. The signal model consisted in a sam-
Closed-FormVariance Formula of the RPHD Single-Tone Frequency Estimator459
10 20 30 40 50
65
60
55
50
45
40
35
30
N
f
r
e
q
u
e
n
c
y

v
a
r
i
a
n
c
e
s

(
d
B

r
a
d
2
)


measured RPHD
variance by (24)
asympt. variance by (25)
CRLB
Fig. 5. Frequency variances versus N at
0
=
0.2, SNR = 20 dB and = 0.
10 15 20 25 30 35 40 45 50
65
60
55
50
45
40
35
30
N
f
r
e
q
u
e
n
c
y

v
a
r
i
a
n
c
e
s

(
d
B

r
a
d
2
)


measured RPHD
variance by (24)
asympt. variance by (25)
CRLB
Fig. 6. Frequency variances versus N at
0
=
0.2, SNR = 20 dB and =/4.
pled sinusoid, embedded in white, gaussian noise. We also derived an asymptotic
form of the variance, which agreed with previously published results. We validated
our results by means of several computer experiments. Coincidence between the
experimental variances and those calculated with our formula (24) was remarkable
in all cases, except for very low signal-to-noise ratios. According to the experi-
mental results, the fact that the asymptotic variance (25) does not depend on the
phase of the sampled sinusoid, while (24) does, can be a serious drawback, so that
quite large errors are possible in the case of a small number of samples. This is
an argument for the usefulness of our closed-form formula (24) as, in common
applications, a number as small as possible of signal samples is desired.
Appendix
The main steps that are necessary for computing the values of E{A
N
}, E{B
N
},
E{A
2
N
}, E{B
2
N
}, and E{A
N
B
N
} are presented in this Appendix.
We use (1), (2) , (7), and (22) in order to get
A
N
= T
A1
+T
A2
+T
A3
, (A1)
where
T
A1
=
N

n=3
2
2
cos(
0
)cos
2
_

0
(n1) +
_
=
2
_
N 2+(2, N1)
_
cos(
0
), (A2)
T
A2
=
N

n=3
cos
_

0
(n1) +
__
2cos(
0
)q(n1) +q(n) +q(n2)
_
, (A3)
460 L. Toma, A. De Sabata, and S. Mischie:
and
T
A3
=
N

n=3
q(n1)
_
q(n) +q(n2)
_
. (A4)
As q(n) is white, with zero mean, we have
E{q(n)} = 0; E{q(n)q(m)} =
n
m

2
q
. (A5)
There results E{T
A2
} = E{T
A3
} = 0, E{A
N
} = E{T
A1
} = T
A1
and (17).
In order to calculate E{A
2
N
}, we note that, due to (A5), we have E{T
Ai
T
Aj
} = 0
for i = j. Therefore
E{A
2
N
} = T
2
A1
+E{T
2
A2
}+E{T
2
A3
}. (A6)
We evaluate now the second and the third term in the RHS of (A6).
E{T
2
A2
} =
N

n=3

2
cos
2
_

0
(n1) +
_

_
4cos
2
(
0
)E
__
q(n1)
_
2
_
+E
__
q(n)
_
2
_
+E
__
q(n2)
_
2
_
+2
N

n=3
N

m=n+1

2
cos
_

0
(n1) +
_
cos
_

0
(m1) +
_
E
__
2cos(
0
)q(n1) +q(n) +q(n2)
_

_
2cos(
0
)q(m1) +q(m) +q(m2)
__
.
By performing the multiplications and by using (A5) there results
E{T
2
A2
} =
N

n=3

2
1+cos
_
2
0
(n1) +2
_
2
_
4
2
q
cos
2
(
0
) +2
2
q
_
+
N

n=3
N

m=n+1

2
_
cos
_

0
(n+m2) +2
_
+cos
_

0
(mn)
_

_
2cos(
0
)E{q(n1)q(m2)}
+2cos(
0
)E{q(n)q(m1)}+E{q(n)q(m2)}

.
Closed-FormVariance Formula of the RPHD Single-Tone Frequency Estimator461
Using again (A5) we get
E{T
2
A2
} =
2

2
q
_
N 2+(2, N1)
__
cos(2
0
) +2
_
+
2

2
q
N1

n=3
4cos(
0
)
_
cos
_
(2n1)
0
+2
_
+cos(
0
)
_
+
2

2
q
N2

n=3
_
cos(2
0
n+2) +2cos(2
0
)
_
=
2

2
q
[4N 10+(4N12)cos(2
0
)
+(2, N 1)
_
2+cos(2
0
)
_
+2(2, N 2)
+2(3, N 1) +(3, N 2)]. (A7)
Then
E{T
2
A3
} = E
_
N

n=3
__
q(n1)q(n)
_
2
+2q
2
(n1)q(n)q(n2)
+
_
q(n1)q(n2)
_
2
_
+2E
_
N

n=3
N

m=n+1
q(n1)q(m1)

_
q(m) +q(m2)
__
q(n) +q(n2)
__
= 2(N 2)
4
q
+2
N1

n=3
E{q
2
(n)q
2
(n1)}
= (4N 10)
4
q
. (A8)
With the use of (A2), (A6), (A7), and (A8) we obtain (19).
From (1), (2), (8), and (22) we get
B
N
= T
B1
+T
B2
+T
B3
(A9)
where
T
B1
=
2
_
N 2+(2, N 1)
_
cos(2
0
), (A10)
T
B2
=4
N

n=3
cos
_

0
(n1) +
_

_
cos(
0
)
_
q(n) +q(n2)
_
q(n1)
_
, (A11)
and
T
B3
=
N

n=3
_
q
2
(n) +q
2
(n2) +2q(n)q(n2) 2q
2
(n1)
_
. (A12)
462 L. Toma, A. De Sabata, and S. Mischie:
The application of (A5) yields as before E{B
N
} = T
B1
, (18) and
E{B
2
N
} = T
2
B1
+E{T
2
B2
}+E{T
2
B3
}. (A13)
We have
E{T
2
B2
} = 16
2
N

n=3
cos
2
_

0
(n1) +
__
2
2
q
cos
2
(
0
) +
4
q
_
+32
2
N

n=3
N

m=n+1
cos
_

0
(n1) +
_
cos
_

0
(m1) +
_
E
__
cos(
0
)
_
q(n) +q(n2)
_
q(n1)
_

_
cos(
0
)
_
q(m) +q(m2)
_
q(m1)
__
.
By performing the multiplications and by using again (A5) we get
E{T
2
B2
} = 16
2

2
q
_
2cos
2
(
0
) +1
_
N

n=3
1+cos
_
2
0
(n1) +2
_
2
+16
2
N

n=3
N

m=n+1
_
cos
_

0
(m+n2) +2
_
+cos
_

0
(mn)
_
_

_
cos
2
(
0
)E{q(n)q(m2)}
cos(
0
)
_
E{q(n)q(m1)}+E{q(n1)q(m2)}
_
= 8
2

2
q
_
N 2+(2, N 1)
__
cos(2
0
) +2
_
+16
2

2
q
N2

n=3
_
cos(2
0
n+2) +cos(2
0
)
_
cos
2
(
0
)
32
2

2
q
N1

n=3
_
cos
_

0
(2n1) +2
_
+cos(
0
_
_
cos(
0
).
After some manipulations there results
E{T
2
B2
} =
2

2
q
_
4N +4(N 4)cos(4
0
) +8(2, N 1)
_
2+cos(2
0
)
_
16(2, N 2) 16(3, N 1) +8(3, N 2)
_
1+cos(2
0
)
_
.
(A14)
For the last term in (A9) we have
E{T
2
B3
} = 4(N 2)
4
q
+2
N

n=3
N

m=n+1
E
__
q
2
(n) +q
2
(n2) +2q(n)q(n2) 2q
2
(n1)
_

_
q
2
(m) +q
2
(m2) +2q(m)q(m2) 2q
2
(m1)
__
= 4(N 2)
4
q
. (A15)
Closed-FormVariance Formula of the RPHD Single-Tone Frequency Estimator463
With the use of (A10), (A13), (A14), and (A15) we obtain (20).
We now start from (A1)...(A4) and A(9)...A(12) in order to obtain
E{A
N
B
N
} = T
C1
+T
C2
+T
C3
(A16)
where
T
C1
=
4
cos(
0
)cos(2
0
)
_
N 2+(2, N1)
_
2
, (A17)
T
C2
=
2
N

n=3
N

m=3
4cos
_

0
(n1) +
_
cos
_

0
(m1) +
_
E
_
_
2cos(
0
)q(n1) +q(n) +q(n2)
_

_
cos(
0
)
_
q(m) +q(m2) q(m1)
_
__
, (A18)
and
T
C3
= E
_
N

n=3
N

m=3
q(n1)
_
q(n) +q(n2)
_

_
q
2
(m) +q
2
(m2) +2q(m)q(m2) 2q
2
(n1)
__
= 0. (A19)
The last equality in (A19) follows from (A5).
The quantity dened in (A18) can be calculated as follows
T
C2
= 2
2
N

n=3
N

m=3
_
cos
_

0
(m+n2) +2
_
+cos
_

0
(mn)
_
_

_
_
4cos
2
(
0
) 2
_
q(n1)q(m) +2cos(
0
)q(n1)q(m)
_
= 4
2

2
q
N1

m=3
cos(2
0
)
_
cos
_

0
(2m1) +2
_
+cos(
0
)
_
+4
2

2
q
N2

m=3
cos(
0
)
_
cos(2
0
m+2) +cos(2
0
)
_
.
After some straightforward transformations we get
T
C2
=
2

2
q
_
(4N 14)
_
cos(
0
) +cos(3
0
)
_
+2
_
(2, N 2) +(3, N1)
_
cos(2
0
)
cos(
0
)
+4(3, N 2)cos(
0
)

. (A20)
With the use of (A16), (A17), (A19), and (A20) we obtain (21).
464 L. Toma, A. De Sabata, and S. Mischie:
References
[1] S. M. Kay, Model Spectral Estimation: Theory and Application. Englewood Cliffs,
NJ: Prentice Hall, 1988.
[2] P. Stoica, List of references on spectral line analysis, Signal Process., vol. 31, no. 3,
pp. 329340, Apr. 1993.
[3] S. M. Kay, Fundamentals of Statistical Signal Processing: Estimation Theory. En-
glewood Cliffs, NJ: Prentice Hall, 1993.
[4] D. V. Tufts and P. D. Fiore, Simple effective estimation of frequency based on
Pronys method, in Proc. Int. Conf. Acoust., Speech, Signal Process., vol. 5, Atlanta,
GA, May 1996, pp. 28012804.
[5] L. B. Fertig and J. H. McClellan, Instantaneous frequency estimation using linear
prediction with comparisons to the DESAs, IEEE Signal Processing Lett., vol. 3,
pp. 5456, Feb. 1996.
[6] V. F. Pisarenko, The retrieval of harmonics by linear prediction, Geophis. J. R.
Astron. Soc., vol. 33, pp. 347366, 1973.
[7] H. Sakai, Statistical analysis of Pisarenkos method for sinusoidal frequency esti-
mation, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-32, pp. 95101,
Feb. 1984.
[8] A. Eriksson and P. Stoica, On statistical analysis of Pisarenko tone frequency esti-
mator, Signal Process., vol. 31, no. 3, pp. 349353, Apr. 1993.
[9] H. C. So, A closed form frequency estimator for a noisy sinusoid, in 45th IEEE
Midwest Symp. Circuits Systems, Tulsa, OK, 2002.
[10] H. C. So and S. K. Ip, A novel frequency estimator and its comparative performances
for short record lengths, in 11th Eur. Signal Processing Conf., Toulouse, France,
2002.
[11] H. C. So and K. W. Chan, Reformulation of Pisarenko harmonic decomposition
method for single-tone frequency estimation, IEEE Trans. Signal Process., vol. 52,
no. 4, pp. 11281135, Apr. 2004.
[12] S. M. Savaresi, S.Bittanti, and H. C. So, Closed-formunbiased frequency estimation
of a noisy sinusoid using notch lters, IEEE Trans. Automat. Contr., vol. 48, no. 7,
pp. 12851292, July 2003.
[13] K. W. Chan and H. C. So, An exact analysis of Pisarenkos single-tone frequency
estimation algorithm, Signal Process., vol. 83, no. 3, pp. 685690, Mar. 2003.