A mathematical model for the atomic clock

error in case of jumps

arXiv:1506.01080v1 [math.PR] 2 Jun 2015

C Zucca
Department of Mathematics “G. Peano”,
University of Torino,
Via Carlo Alberto 10, 10123 Turin, Italy
[email protected]

P Tavella
Istituto Nazionale di Ricerca Metrologica (INRiM),
Strada delle Cacce 91,
10135 Turin, Italy

Abstract
We extend the mathematical model based on stochastic differential
equations describing the error gained by an atomic clock to the cases
of anomalous behavior including jumps and an increase of instability.
We prove an exact iterative solution that can be useful for clock simu-
lation, prediction, and interpretation, as well as for the understanding
of the impact of clock error in the overall system in which clocks may
be inserted as, for example, the Global Satellite Navigation Systems.

Keywords: Atomic clock, Clock model, Clock jump, Anomalies.

1 Introduction
Atomic clocks are usually the heart of complex scientific and technological
systems and clock behavior has a direct impact on the overall performance of

1
the system. For the study, simulation, analysis, characterization, and inter-
pretation of all these systems it is therefore mandatory to have a consistent
mathematical clock model able to represent its typical behavior. Several
studies have dealt with this matter [1, 2, 3, 4, 5, 6, 7].
Recently, particularly in space applications such as the Global Satellite
Navigation Systems (GNSS), anomalous clock behaviors have been experi-
mentally observed [8, 9, 10, 11, 12, 13, 14]: phase or frequency jumps, increase
of instability, general non stationary trends. Such anomalous behaviors have
an important impact on the global system and may affect the overall per-
formance very seriously. This could have, in case of GNSS, dramatic conse-
quences even security of life.
In this paper we propose a mathematical model including the typical
anomalies of atomic clocks and we show how the stochastic multidimen-
sional differential equation describing such behaviors can be exactly solved
obtaining the complete description of clock states. In addition, the solution
can be written in an iterative form.
We therefore provide a clock model including anomalies with an exact
iterative form that may be useful for simulations as well as for the insertion
of the clock model in further processing as Kalman filtering, orbit and clock
estimation algorithms, time scale algorithms, or other types of estimation
and filtering. Moreover we provide an exact solution that allows the predic-
tion and evaluation of the clock error at a certain time after synchronization
also in presence of anomalies and as a function of the epoch of the anomaly
occurrence. This knowledge is of fundamental importance for GNSS evalua-
tions as the clock time deviation directly impact the user solution accuracy
[15, 16].
In Section 2 we recall the main features of the mathematical model
based on stochastic differential equations including phase, frequency, and
drift jumps and we introduce in the model a possible sequence of jumps and
the increase of instability. In Section 3 we illustrate through examples the
features captured by the mathematical model. In Section 4 we explicitly ad-
dress the example of a space Rubidium clock on board of a navigation satellite
and we evaluate the prediction error at a certain time after synchronization
in presence of a frequency jump.
To ease the use of this model in simulating clock behavior, we also add
the availability of a Matlab code that can be freely downloaded from the
INRIM web site.

2
2 The model of the atomic clock error with
anomalies
It is experimentally observed that the signal of an atomic clock can be af-
fected by several anomalies. Our aim is to extend the model proposed in [7]
introducing anomalies. In the forthcoming we will consider three different
types of anomalies that involve jumps at deterministic times or changes in
the variance.

2.1 Instantaneous jumps

Let us consider the three-state clock model that we proposed in [7]. The
three-state clock model X = {(X1 , X2 , X3 )(t), t ≥ 0} is described by the
three dimensional stochastic differential equation:


 dX1 (t) = (X2 (t) + µ1 )dt + σ1 dW1 (t)
dX2 (t) = (X3 (t) + µ2 )dt + σ2 dW2 (t) (1)


dX3 (t) = µ3 dt + σ3 dW3 (t)

with initial condition X(0) = (c1 , c2 , c3 ), where {Wi (t), t ≥ 0}, i = 1, 2, 3, are
three indipendent, one-dimensional standard Wiener processes.
The variable X1 (t) represents the time deviation, the derivative Ẋ1 (t)
represents the frequency deviation of which X2 (t) is only a component (i.e.,
what is generally called the random walk component). Finally X3 (t) repre-
sents the frequency drift or aging.
The metrological interpretation of the model, the solution of (1) and
its iterative form can be found in [7]. Here we recall that the probability
distribution function of the solution X at time t is Normal with mean [7]
2 3
c1 + (c2 + µ1 )t + (c3 + µ2 ) t2 + µ3 t6
 
2
M(t) =  c2 + (c3 + µ2 )t + µ3 t2 (2)
 

c3 + µ 3 t

and covariance matrix

 3 5 2 4 3

σ 2 t + σ22 t3 + σ32 20
t
σ22 t2 + σ32 t8 σ32 t6
 1 t2 4 3 2 
2t
Σ(t) =  2
 σ2 2 + σ3 8 σ22 t + σ32 t3 σ32 t2 
. (3)
3 2
σ32 t6 σ32 t2 2
σ3 t

3
 Let us extend the model of the clock error (1) characterized by jumps
with amplitude ai , i = 1, 2, 3 in each component: phase, frequency and drift.
To introduce a simple and treatable mathematical description, we make use
of general continuous time Markov Chain N(t) = {(N1 , N2 , N3 )(t), t ≥ 0},
i.e. a continuous-time stochastic process which takes values in some finite or
countable set and for which the time spent in each state has an exponential
distribution (cf. [17]). If we introduce these anomalies into the clock model
(1) the corresponding mathematical model of the clock becomes


 dX1 (t) = (X2 (t) + µ1 )dt + σ1 dW1 (t) + a1 dN1 (t)
dX2 (t) = (X3 (t) + µ2 )dt + σ2 dW2 (t) + a2 dN3 (t) , (4)


dX3 (t) = µ3 dt + σ3 dW3 (t) + a3 dN3 (t)

where Ni (t), i = 1, 2, 3 are independent continuous time Markov Chains.

The easiest particular case considers istantaneous jumps with amplitude
ai , i = 1, 2, 3 that take place at deterministic times θi , i = 1, 2, 3. It means
that
dNi (t) = dH(t − θi ) = δ(t − θi )dt. (5)
where H(t) is the Heaviside function and δ(t) is the Dirac delta function
[18]. Note that the derivatives of the Heaviside function is the Dirac delta
function, the integral is the so called ramp function and the multiple integrals
are given by Z Z
tn
. . . H(t) dt . . . dt = H(t). (6)
| {z }
n
n!
In Figure 1 the Heaviside function, its integral and double integral are shown.
A first generalization of this simple model would consider Ni (t), i = 1, 2, 3
independent Poisson processes, i.e. anomalies occurring on random epochs,
the intertimes between the anomalies are exponentially distributed and the
jumps are deterministic of size 1 [17]. This model will be the subject of a
future study as it would help in treating the complete clock behaviour for
example in the analysis of time scale algorithms.
Here we consider the choice (5), the system (4) becomes


 dX1 (t) = (X2 (t) + µ1 )dt + σ1 dW1 (t) + a1 dH(t − θ1 )
dX2 (t) = (X3 (t) + µ2 )dt + σ2 dW2 (t) + a2 dH(t − θ2 ) . (7)
dX3 (t) = µ3 dt + σ3 dW3 (t) + a3 dH(t − θ3 )



4
 2
∫ ∫ H(t) (dt)2
1

−2 −1.5 −1 −0.5 0 0.5 1 1.5 2

t
2
∫ H(t) dt

−2 −1.5 −1 −0.5 0 0.5 1 1.5 2

t
1.5
1
H(t)

0.5
0
−0.5
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2
t

Figure 1: Plot of the Heaviside function H(t), its integral and its double
integral (from below to above).

Integrating (7) and recalling

Z t
aδ(s − t0 )ds = aH(t − t0 )
0
Z t
aH(s − t0 )ds = a(t − t0 )H(t − t0 ) (8)
0
Z t
a(t − t0 )2
a(s − t0 )H(s − t0 )ds = H(t − t0 ),
0 2

5
 the solution can be written in closed form
2 3



 X1 (t) = c1 + (c2 + µ1 )t + (c3 + µ2 ) t2 + µ3 t6 + σ1 W1 (t)
2
+σ2 0t (t − s)dW2 (s) + σ3 0t (t−s)
 R R
dW3 (s)


2



+a1 H(t − θ1 ) + a2 (t − θ2 )H(t − θ2 )




2

 +a3 (t−θ2 3 ) H(t − t3 ) . (9)

t2 Rt




 X2 (t) = c2 + (c3 + µ2 )t + µ3 2 + σ2 W2 (t) + σ3 0 (t − s)dW3 (s)
+a2 H(t − θ2 ) + a3 (t − θ3 )H(t − θ3 )





X3 (t) = c3 + µ3 t + σ3 W3 (t) + a3 H(t − θ3 )



The probability distribution function of the solution X at time t is Normal

with mean MI (t) and covariance matrix Σ(t), where the mean depends on
the mean value of the process without anomalies (2)
2
a1 H(t − θ1 ) + a2 (t − θ2 )H(t − θ2 ) + a3 (t−θ2 3 ) H(t − t3 )
 
I
M (t) = M(t) +  a2 H(t − θ2 ) + a3 (t − θ3 )H(t − θ3 )
 
 (10)
a3 H(t − θ3 )

while its covariance matrix (3) does not change.

Let us now consider a fixed time interval [0, T ] and an equally spaced
partition 0 ≡ t0 < t1 < . . . < tN ≡ T and let us denote with τ = tk+1 − tk ,
k = 0, 1, . . . , N − 1 the resulting discretization time step. For simplicity
we hypothesize that the jumps can occur only in the discretization epochs.
The sampling time can be as small as necessary to ensure this assumption.
Moreover, observing that

H(tk+1 − θ) − H(tk − θ) = δ(tk+1 −θ) , (11)

and consequently

(tk+1 − θ)[H(tk+1 − θ) − H(tk − θ)] = 0, (12)

where δ(tk+1 −θ) is the Kronecker delta [1], we can express the solution at time
tk+1 in terms of the solution at time tk as:

2 3



 X1 (tk+1 ) = X1 (tk ) + (µ1 + X2 (tk ))τ + (µ2 + X3 (tk )) τ2 + µ3 τ6 + Jk,1
+a1 δ(tk+1 −θ1 )


2 ,(13)


 X2 (tk+1 ) = X2 (tk ) + (µ2 + X3 (tk ))τ + µ3 τ2 + Jk,2 + a2 δ(tk+1 −θ2 )
X3 (tk+1 ) = X3 (tk ) + µ3 τ + Jk,3 + a3 δ(tk+1 −θ3 )



6
 where Jk = (Jk,1 , Jk,2 , Jk,3 ) can be interpreted as an innovation
 Rt 
σ1 (W1 (tk+1 ) − W1 (tk )) + σ2 tkk+1 (W2 (s) − W2 (tk ))ds
Rt
+σ3 tkk+1 (tk+1 − s)(W3 (s) − W3 (tk ))ds
 
 
Jk =  Rt . (14)

 σ2 (W2 (tk+1 ) − W2 (tk )) + σ3 tkk+1 (W3 (s) − W3 (tk ))ds 

σ3 (W3 (tk+1 ) − W3 (tk ))

and is Normal distributed with zero mean and covariance matrix Q given by
 3 5 2 4 3

σ 2 τ + σ22 τ3 + σ32 τ20 σ22 τ2 + σ32 τ8 σ32 τ6
 1 τ2
2 τ4 3 2 
Q= 2
 σ2 2 + σ3 8 σ22 τ + σ32 τ3 σ32 τ2 . (15)
3 2
σ32 τ6 σ32 τ2 2
σ3 h

In the iterative form, we note that the jump acts only once on the corre-
sponding component, then the iterative process carries the effect in the other
components.

2.2 Effect of two equal and opposite jumps

Let us consider the model of the clock error characterized by an anomaly
on the second component, a temporary frequency jump. The anomaly is
composed of two istantaneous jumps at times θ0 and θ1 > θ0 . At time θ0
we register a positive jump of amplitude a/∆ and at time θ1 we register a
negative jump of amplitude −a/∆, where ∆ = θ1 − θ0 . The atomic clock
error can be described by the following stochastic differential equation


 dX1 (t) = (X2 (t) + µ1 )dt + σ1 dW1 (t)
a a
dX2 (t) = (X3 (t) + µ2 )dt + σ2 dW2 (t) + ∆
dH(t − θ0 ) − ∆
dH(t − θ1 ) (16)


dX3 (t) = µ3 dt + σ3 dW3 (t).

Integrating (16) and recalling (8) we get the solution

2 3



 X1 (t) = c1 + (c2 + µ1 )t + (c3 + µ2 ) t2 + µ3 t6 + σ1 W1 (t)
2
+σ2 0t (t − s)dW2 (s) + σ3 0t (t−s)
 R R
dW3 (s)


2


 a

+aH(t − θ1 ) + ∆ (t − θ0 )I[θ0 ,θ1 ) (t)
2 , (17)
X2 (t) = c2 + (c3 + µ2 )t + µ3 t2 + σ2 W2 (t) + σ3 0t (t − s)dW3 (s)
R



+ ∆a I[θ0 ,θ1 ) (t)







X3 (t) = c3 + µ3 t + σ3 W3 (t)

7
 where I[a,b) (t) is the window function defined as

I[a,b) (t) = H(t − a) − H(t − b), (18)

and the term +aH(t − θ1 ) takes care of the phase deviation accumulated
until time θ1 .
The probability distribution function of the solution X at time t is Normal
with mean MII (t) and covariance matrix Σ(t). Note that the mean can be
decomposed in two terms where the first corresponds to the mean value of
the process without anomalies (2)
a
 
aH(t − θ1 ) + ∆
(t − θ0 )I[θ0 ,θ1 ) (t)
MII (t) = M(t) +  a
 ∆ I[θ0 ,θ1 ) (t)

 (19)
0

while its covariance matrix (3) does not change.

The solution in iterative form becomes:
2 3



 X1 (tk+1 ) = X1 (tk ) + (µ1 + X2 (tk ))τ + (µ2 + X3 (tk )) τ2 h+ µ3 τ6 + Jk,1 i
τ2 a
X (t
 2 k+1
) = X (t
2 k ) + (µ 2 + X (t
3 k ))τ + µ 3 2 + Jk,2 + ∆
δ (tk+1 −θ0 ) − δ(tk+1 −θ1 ) , (20)
 X (t
3 k+1 ) = X3 (tk ) + µ3 τ + Jk,3


where Jk = (Jk,1 , Jk,2 , Jk,3 ) is (14). Again we note that the effect of the
jump is present only in the affected component, X2 in this case, and then
the iterative process takes care of the effect on the other components.

Remark 2.1 Note that the algoritms can be easily extended to n determin-
istic jumps.

2.3 Variance increase

The atomic clock dynamics can also show changes in the variance. Let us
model an atomic clock error with a change of the variance values in the
interval [θ0 , θ1 ]. The process for t ∈
/ [θ0 , θ1 ] is described as the model without
anomalies (1) and its probability distribution is Normal with mean M(t) and
covariance matrix Σ(t). At time t ∈ [θ0 , θ1 ] its distribution changes, while
the mean M(t) is unchanged (2) its covariance matrix becomes ΣIII (t) that
is the same matrix (3) but with parameters σ1 , σ2 and σ3 substituted by σ10 ,
σ20 and σ30 .

8
 This anomaly can be modeled introducing a new innovation matrix in
the time interval [θ0 , θ1 ]. The corresponding mathematical iterative model
becomes
 2 3


 X1 (tk+1 ) = X1 (tk ) + (µ1 + X2 (tk ))τ + (µ2 + X3 (tk )) τ2 + µ3 τ6 + J˜k,1
2
X2 (tk+1 ) = X2 (tk ) + (µ2 + X3 (tk ))τ + µ3 τ2 + J˜k,2 , (21)
˜

 X (t
3 k+1 ) = X3 (tk ) + µ3 τ + Jk,3


where the innovation is

(
N (0, Q0 ) if tk+1 ∈ [θ0 , θ1 ]
J̃k = (J˜k,1 , J˜k,2 , J˜k,3 ) ∼ (22)
N (0, Q) if tk+1 ∈ / [θ0 , θ1 ]

where Q is the covariance matrix (15) while Q0 is the same matrix where the
parameters σ1 , σ2 and σ3 are substituted by σ10 , σ20 and σ30 .

Remark 2.2 In its present form the model accounts for anomalies at deter-
ministic times. In reality anomalies occur at random times. It is possible to
extend the model and the corresponding algorithms proposed to random times
θi , i = 0, 1, 2, 3 under the hypothesis that the random variables θi are inde-
pendent for i = 0, 1, 2, 3 and that are independent from the process X without
anomalies. Under these hypothesis the distribution of the solution can be
computed conditioning on the time of the anomalies. The iterative solution
with deteriministic times can still be used for simulations after generating
random times for the anomalies epochs according to their given distribution.

3 Simulations and examples

The simulation of the models introduced in the previous paragraph can be
easily handled using the iterative solutions (13), (20) and (21).
In all the cases it is necessary to simulate the innovations Jk that, at any
instant tk , are independent identical distributed three dimensional Normal
random variabiles with zero mean and covariance matrix Q given by (15).
Since the matrix Q is not diagonal, the three components of the vector Jk
are dependent. Many softwares have routines to generate multidimensional
Normal random variables.
To better understand the correlation mechanism, we explicitely describe
the procedure for generating Jk in our three dimensional case. It is well

9
known that the vector Jk may be written as Jk = AZ where Z is a standard
Normal random vector with zero mean and identity covariance matrix I,
i.e. Z is a vector of three standard Normal independent random variables
[19]. The matrix A transforms the three components of Z introducing a
dependency that is captured in the correlation matrix Q, where Q = AAT .
Here the superscript T denotes the transposed matrix. The values of A can
be computed using for example Cholesky decomposition method [20] that
gives
 √ 
q11 r
0 0
 2
q11 q22 −q21

 √q21
A= q11 q11
0 
 (23)
 r 
2
q31
√q32 q11 −q31 q212 − √q32 q11 −q31 q212
 q 
√ 31
q11
q33 − q11
q11 (q11 q22 −q21 ) q11 (q11 q22 −q21 )

where qi,j , i, j = 1, 2, 3 are the entry of Q.

Note that the iterative solutions, useful for simulations, are exact solu-
tions and do not introduce any approximation error. Exact simulations are
helpful to understand the evolution of the clock behavior or to obtain esti-
mates of functionals of clock states that cannot be computed theoretically.
The Matlab code can be freely downloaded from the INRIM website.

3.1 Simulation of anomalous behavior

Here we report some examples to illustrate the effect of the different types
of anomalies of the process described in the previous section.
In Figure 2 the model (7) with an instantaneous jump on the third com-
ponent is considered and a sample path of the process X(t) is plotted. The
parameters considered are µ1 = µ2 = µ3 = 0, σ1 = σ2 = σ3 = 1, a1 = a2 = 0,
a3 = 3 and the times of the jump is θ3 = 2. The figure clearly shows the
jump in the third component, the linear trend of the second component and
the quadratic trend of the first one, as expected by the solution (9).
In Figure 3 the model (7) with instantaneous jumps in all the three com-
ponents is considered and a sample path of the process X(t) is plotted.
The parameters considered are µ1 = µ2 = µ3 = 0, σ1 = σ2 = σ3 = 1,
a1 = a2 = a3 = 3 and the times of the jumps are θ1 = 6, θ2 = 4, θ3 = 2,
respectively. The figure clearly shows the jumps in the trajectories and the
increasing trend of each component sample path after the jumps.
In Figure 4 the three components of the process X(t) are plotted in case
of two equal and opposite frequency jumps. The parameters considered are

10
 100
X1(t)
50

0
0 1 2 3 4 5 6 7 8 9 10

20
X2(t)

10

0 1 2 3 4 5 6 7 8 9 10

5
X3(t)

−5
0 1 2 3 4 5 6 7 8 9 10
t

Figure 2: Simulation of the three components of the process X(t) with pa-
rameters µ1 = µ2 = µ3 = 0, σ1 = σ2 = σ3 = 1, a1 = a2 = 0, a3 = 3 and
θ3 = 2. The times of the jump is highlighted by an arrow (arbitrary units).

µ1 = µ2 = µ3 = 0, σ1 = σ2 = σ3 = 1, a = 4 and the time interval of the

jumps is [θ0 , θ1 ] = [4, 6]. It is evident the linear trend of the first component
in the interval [4, 6].
In Figure 5 the three components of the process X(t) are plotted in case
of variance increase. The parameters considered are µ1 = µ2 = µ3 = 0,
σ1 = σ2 = σ3 = 1, σ10 = σ20 = σ30 = 8, and the time interval of the change
of variance is [4, 8]. The figure shows the effect of an increase of the noise in
the selected interval.
The previous figures show that the model captures the main features of
the evolution of the atomic clock error.

11
 80
X (t) 60
40
1

20
0
0 2 4 6 8 10

30
20
X (t)
2

10
0
0 2 4 6 8 10

6
4
X3(t)

2
0
−2
0 2 4 6 8 10
t

Figure 3: Simulation of the three components of the process X(t) with pa-
rameters µ1 = µ2 = µ3 = 0, σ1 = σ2 = σ3 = 1, a1 = a2 = a3 = 3 and
θ1 = 6, θ2 = 4, θ3 = 2. The times of the jumps are highlighted by an arrow
(arbitrary units).

4 Application to a space GNSS clock

In this section we tune the parameters of the model in order to study a case of
interest concerning a Rubidium Atomic Frequency Standard (RAFS) placed
on board of a satellite of a Global Navigation Satellite System (GNSS). We
evaluate the effect of a clock anomaly on the clock prediction error, which
is a crucial problem in navigation system. As an example we consider the
standard parameters describing the behavior of a RAFS on board on the
experimental satellites of the Galileo system [21].
The Allan deviation [22] indicating the stochastic short term noise of
such RAFS is usually due to white frequency noise and it is assumed to be
σy (τ ) = 5 · 10−12 τ −1/2 . The relationship between the Allan deviation and the

12
 10

X1(t) 0

−10
0 1 2 3 4 5 6 7 8 9 10

5
X2(t)

−5
0 1 2 3 4 5 6 7 8 9 10

5
X3(t)

−5
0 1 2 3 4 5 6 7 8 9 10
t

Figure 4: Simulation of the three components of the process X(t) with pa-
rameters µ1 = µ2 = µ3 = 0, σ1 = σ2 = σ3 = 1, a = 4 and the time interval
of the jumps is [4, 6]. The times of the jumps are highlighted by an arrow
(arbitrary units).

diffusion coefficient of matrix Q is known [7] and, since

σ1
σy (τ ) = √ , (24)
τ
we can infer σ1 = 5 · 10−12 . The other noises are negligible for this example,
we set σ2 = 10−22 , σ3 = 10−22 for numerical reasons only. We assume that
all the polynomial, linear or quadratic slopes are zero, either because they
are really negligible or because they have been estimated and appropriately
compensated. We also assume that the clock is perfectly synchronized at
epoch zero, therefore all the initial time constants are set to zero.
With this set of parameter we check which is the time deviation gained
by such a clock a certain time after synchronization. Note that in the Galileo
system the clocks are expected to be re-synchronized every 100 min = 6000
seconds [21].

13
 50

X1(t) 0

−50
0 1 2 3 4 5 6 7 8 9 10

20
X2(t)

−20
0 1 2 3 4 5 6 7 8 9 10

10
X3(t)

−10
0 1 2 3 4 5 6 7 8 9 10
t

Figure 5: Simulation of the three components of the process X(t) with pa-
rameters µ1 = µ2 = µ3 = 0, σ1 = σ2 = σ3 = 1, σ10 = σ20 = σ30 = 8, and the
time interval of the change of variance is [4, 8]. The time interval of variation
of the variance is highlighted by arrows (arbitrary units).

Using the clock model without anomalies (1) the time deviation X1 at
time t = 6000 s is a Normal random variable with mean value M1 (t) =
q √
E[X1 (t)] = 0 and standard deviation approximately Σ11 (t) ' σ1 t =
√
5 · 10−12 q6000 ' 4 · 10−10 s = 0.4 ns. In this case a time deviation interval
of ±1, 96 Σ11 (t) ' ±759 · 10−12 s = ±0.8 ns, with 95% confidence level,
after 6000 s, represents the prediction error and hence the contribution on
the positioning error due to unpredictable clock time deviation. Note that
the polynomial behavior of the three components has not been considered in
this example.
To check the effect of a clock anomaly, we consider a frequency jump
of relative value 10−12 . Frequency jumps are sometimes observed on space
clock on board of navigation satellites [10]. In our notation a frequency
jump corresponds to (9) when we choose a2 = 10−12 and a1 = a3 = 0. We

14
 2.5

2
Probability density

1.5

0.5

0
0
1000

3000

6000 10
8
6
4
9000 2
0
t [s] −2
X1(t) [s]

Figure 6: Probability density function and the 95% confidence interval of

the time deviation X1 at time t = [1000, 3000, 6000, 9000] s, with parameters
µ1 = µ2 = µ3 = 0, σ1 = 5 · 10−12 , σ2 = σ3 = 10−22 and jump a2 = 10−12 at
time θ = 100 compared with the case without anomalies.

suppose that the frequency jump occurs at epoch θ and the effect on the time
deviation at epoch t depends on the interval (t−θ). Looking to (9) we expect
the time deviation X1 to be a Normal random q variable √
with mean value
√
M 1 (t) = a2 (t − θ) and standard deviation Σ11 (t) = σ1 t = 5 · 10−12 t.
I

Let us deal with anomaly occurring at θ = 100 s or at θ = 5000 s. In the

first case the anomaly occurs immediately after the synchronization and at
t = 6000 s its effect is quite impacting. The 95%qconfidence level for the time
deviation X1 gives the interval a2 (t − θ) ± 1.96 Σ11 (t) = 5.9 ± 0.8 ns.
In the second case, the anomaly occurs just at the end of the predictive
q the prediction error at 95% confidence level in that case is a2 (t −
period, and
θ) ± 1.96 Σ11 (t) = 1 ± 0.8 ns, therefore the effect is much smaller, very close
to the case without anomalies discussed above.

15
 In Figure 6 is shown the probability density function and the 95% con-
fidence interval of the time deviation X1 at time t = [1000, 3000, 6000, 9000]
s, with jump a2 = 10−12 at time θ = 100 compared with the case without
anomalies.
As we see from (10), the impact of the frequency jump on the time devi-
ation and hence on the time prediction error is linearly depending on q t − θ,
the worst case being at θ = 0 giving a confidence interval a2 t ± 1.96 Σ11 (t),
q
while the best case of no anomaly gives ±1.96 Σ11 (t) at 95% confidence
level.

5 Conclusion
A previous mathematical model based on multidimensional stochastic dif-
ferential equation describing the atomic clock error has been extended to
include the presence of clock anomalies experimentally observed on clocks,
particularly in GNSS space clocks. The model includes phase, frequency,
and drift jumps, as well as changes on the random noise power. Examples
of such trends are presented and the case of a GNSS space clock is discussed
by evaluating the time deviation gained at a certain time after synchroniza-
tion with or without the effect of a frequency jump. Our results show the
ability of the proposed model to describe clock anomalies and their effect on
complex systems.

Acknowledgment
Work partially supported by Turin University: Stochastic Processes/ ZUCC01CT11.

References
[1] Arnold L., Stochastic Differential Equations: theory and applications,
John Wiley and Sons, New York (1974).

[2] Audoin C. and Dimarcq N., Stochastic models of stable frequency and
time sources and their relationship, IEEE Trans. Instrum. Meas. 42 (3),
682–688 (1993).

16
 [3] Galleani L., A tutorial on the two-state model of the atomic clock noise,
Metrologia, 45 (6), 175–182 (2008).

[4] Galleani L., Sacerdote L., Tavella P. and Zucca C., The mathematical
model of for the atomic clock error: an overview Metrologia, 40 (3),
S257–S264 (2003).

[5] VanDierendonck A. and Brown R., Relationship Between Allan Vari-

ances and Kalman Filter Parameters, Proc. of 16th Annual PTTI Meet-
ing, Greenbelt, MD (1984) .

[6] Vernotte F., Delporte J., Brunet M. and Tournier T., Uncertainties of
drift coefficients and extrapolation errors: Application to clock error
prediction,Metrologia 38, 325–342 (2001).

[7] Zucca C. and Tavella P., The Clock Model and its Relationship with the
Allan and related Variances IEEE Trans. Ultra. Ferro. Freq. Contr. 52
(2), 289–296 (2005).

[8] Brown K.R., Hutsell S., and Shank C., gL-band anomaly detection in
GPS, Proc. Institute of Navigation 51st Annu. Meeting, 107–111 (1995).

[9] Camparo J. , Frequency Equilibration and the Light-Shift Effect for

Block IIR GPS Rubidium clocks, Proc. of 36th Annual PTTI Meeting,
Washington, (2004).

[10] Cernigliaro A., Fantino G., Sesia I., Galleani L. and Tavella P., Nonsta-
tionarities in space clocks: Investigations on Experimental Data Proc.
28th European Frequency and Time Forum (EFTF), June 2014, Neuch-
tel, Switzerland, 126–129 (2014).

[11] Droz F., Barmaverain G. , Wang Q., Rochat P., Emma F. and Waller
P., Galileo Rubidium Standard - Lifetime data and GIOVE-A related
telemetries, Proceedings of Frequency Control Symposium, 2007 joint
with the 21st European Frequency and Time Forum, 1122–1126 (2007).

[12] Galleani L. and Tavella P., Detection of Atomic Clock Frequency Jumps
with the Kalman Filter, IEEE Trans. Ultra. Ferro. Freq. Contr., 59(3),
504–509 (2012).

17
[13] Galleani L. and Tavella P., Detection and identification of atomic clock
anomalies, Metrologia 45 (6), S127-S133 (2008).

[14] Weiss M., Shome P., and Beard R., GPS signal integrity dependencies
on atomic clocks, in Proc of 38th Annual PTTI Meeting, 439–448 (2006).

[15] Panfilo G. and Tavella P., Atomic clock prediction based on stochastic
differential equations, Metrologia 45, S108–S116 (2008).

[16] Tavella P., Statistical and mathematical tools for atomic clocks, Metrolo-
gia 45 (6), S183-S192 (2008).

[17] Stroock D.W.,An Introduction to Markov Processes, Springer, (2005).

[18] Abramowitz M. and Stegun I.A., Handbook of Mathematical Functions:

with Formulas, Graphs, and Mathematical Tables (Dover Books on
Mathematics) Paperback, (1965).

[19] Jacod J. and Protter P. Probability Essentials, Springer, Berlin, (2004).

[20] Horn R.A. and Johnson C.R.,Matrix Analysis, Cambridge University

Press, New York, (2013).

[21] Waller P., Gonzalez F., Binda S., Sesia I., Hidalgo I., Tobias G., and
Tavella P., The In-Orbit Performances of GIOVE Clocks IEEE Trans.
Ultra. Ferro. Freq. Contr., 57 (3), 738–745 (2010).

[22] Allan D.W., Time and frequency (time-domain) characterization, esti-

mation, and prediction of precision clocks and oscillators, IEEE Trans.
Ultra. Ferro. Freq. Contr., 34, 647–654 (1987).

18