See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/235004789

Derivations of the Phase Noise Spectra of Lasers and of Lasers Passing Through
Interferometers

Article · January 2013

CITATIONS READS

0 4,504

1 author:

Robert Steed
European-Mediterranean Seismological Centre
45 PUBLICATIONS 636 CITATIONS

SEE PROFILE

All content following this page was uploaded by Robert Steed on 05 June 2014.

The user has requested enhancement of the downloaded file.

 Derivations of the Phase Noise Spectra of Lasers and of
Lasers Passing Through Interferometers

Robert J. Steed

20/01/2013

Contents

I Oscillator Phase Noise and Spectral Lineshapes 2

1 Introduction: A Phasor derivation of Laser Spectral Lineshape 2

2 Laser Phase Noise models 3

3 Deriving the Power Spectral Density 4

4 Derivation using real notation 6

II Unbalanced Interferometers & Self-Heterodyne Measurements 7

5 Derivation of theory for heterodyne case 7

6 Implications 10

7 Self-heterodyne linewidth measurements 11

8 Fitting some heterodyne phase noise measurements 12

9 Homodyne Interferometry 15

III Appendices 19

A Wiener–Khinchin theorem 19

B Wiener Processes 19

C Random Variables 21

References 23

1
 Abstract
These theory notes first provide a fairly standard derivation connecting laser phase
noise to its measured power spectum, then other derivations are given for the power
spectra found after a laser has passed through an interferometer for both the heterodyne
and homodyne cases. The heterodyne self-interometric theoretical result is also shown
to compare well with some measured data.

Introduction
Lasers can have many outstanding features such as high intensity, high spatial coherence
or short pulse lengths. Here we will consider another of their characteristics, namely their
narrow linewidths or equivalently their long coherence times. The fundamental limit of a
laser linewidth originates in the noise from spontaneous emission (although other sources of
noise often obscure this limit) [1, 2, 3]. Hence one approach to modelling laser linewidths can
be to study a stochastical model of a laser [1, 4, 5, 6]. In the approach that we will concentrate
upon here, it is noted that since amplitude noise is normally negligible in laser systems, we
can normally say that laser linewidths are due to phase noise in the output [3, 2]. This means
that we can link temporal coherence directly to laser linewidth and that we will talk about
phase noise measurements interchangeably with linewidth measurements. It also means
that we can borrow basic theory from the field of electrical oscillators where the same is true
and where many of these concepts originate[7, 8, 9, 10, 11].
In this note, we will derive the linewidth of a laser signal from its phase noise in Part I.
Part II then derives the phase noise spectrum for an unbalanced interferometer and shows
that it exhibits characteristic oscillations. We then compare this theory with some real meas-
urements to show that the effect is real; in fact, this effect was limiting our linewidth meas-
urements and meant that we had to adjust our setup in order to get a more accurate result.

Part I
Oscillator Phase Noise and Spectral
Lineshapes
1 Introduction: A Phasor derivation of Laser Spectral Lineshape
Here is a derivation of spectral line profiles from temporal phase noise. This is mostly an
expanded version of the notes found in Yariv’s books ([3] or [2]).
A pseudo monochromatic signal can be described by

A0 (1 + α(t)) eiω0 t+iφ(t) (1)

where α is amplitude noise and φ is phase noise. For a typical laser amplitude noise can be
ignored (due to a lasing mode always operating in gain saturated conditions). So we have

A0 eiω0 t+iφ(t) (2)

2
 h i
Imagine that we knew how to calculate the Fourier transform directly, then F eiφ(t) =
Ψ(ω ) and so h i
F A0 eiω0 t+iφ(t) = E0 Ψ(ω − ω0 )
and the power spectrum would be
2
P(ω ) ∝ | A0 Ψ(ω − ω0 )| (3)

However, we will find that we can’t do this step directly when dealing with a noisy variable.
In order to make progress with this problem, we first have to think about what kind of noise
best describes the laser’s phase noise. There are many kinds of noise; they even have their
own colours depending on their spectral content. White noise has noise at all frequencies,
pink noise has a logarithmic reduction in power density with increasing frequency, brown
noise has a spectrum that reduces at 6dB/octave, etc. Moreover, to specify a noise source
requires both its spectral content (aka. colour) and a probability distribution of the values
that it can take. So, a variable that can take values from a Gaussian distribution with no
statistical correlation between values is called white Gaussian noise.

2 Laser Phase Noise models

The fundamental source of noise in a laser is from spontaneous emission of photons. The
spontaneously emitted photons cause the phase of the laser output to shift randomly over
time but they do not have a large affect on the amplitude of the laser output, to such an
extent that amplitude noise is often ignored. Since the phase of the spontaneously emitted
photons is completely random, they are a source of white noise but their effect on the total
phase is incremental which leads to the total phase taking a random walk. In the study of
noise, such a process is known as a Weiner process (see [12] and appendix B) or as Brown
noise (short for Brownian noise).
What can we say about φ(t)? It is a noisy process but we will assume that it is unbiased
in its direction and so it’s expectation is zero: E [φ(t)] = 0. We also assume that the signal
is mostly monochromatic and so variations in φ(t) should be much slower than ω0 t i.e. φ(t)
can’t be white noise. We’re expecting the phase to perform some kind of random walk. That
means that variance of φ(t + τ ) − φ(t) increases with increasing τ. We also want our problem
to be ’stationary’ or independent of absolute time, if we’re considering c.w. lasers then this
is a reasonable assumption.
Although we are not too concerned with the details for this derivation; when laser beha-
viour is modelled, we find a fundamental limit to the narrowness of the linewidth which is
given by the Schawlow-Townes limit [1]
∆vc
∆vST = (4)
P̄
where ∆vc is the linewidth of the cavity and P̄ is the average number of photons in the cavity.
However, a diode laser will also have a “linewidth multiplication factor” due to interactions
between the laser light and the effective refractive index of the gain medium which will
increase the linewidth by a factor of ∼ 25 or more[1, 5]. Naturally, a real laser will normally
have an larger linewidth due to noise from the current driving the gain medium or thermal
or mechanical noise etc.

3
3 Deriving the Power Spectral Density
We are going to calculate the power spectral density using the Wiener-Khinchin theorem
(appendix A). So the autocorrelation is

R(τ ) = E [ A(t + τ ) A(t)∗ ] (5)

h i
= | A0 |2 eiω0 τ E ei(φ(t+τ )−φ(t)) (6)

In section 2, we decided that the phase noise is a Wiener process and so we know that
φ(t + τ ) − φ(t) = ∆τ φ, which is a random variable and in particular is a form of Gaussian
white noise. ∆τ φ has a Gaussian probability distribution of

1
e −∆τ φ
2 / (2σ2 )
g(∆τ φ) = p τ (7)
2πστ2

and the variance στ2 is proportional to |τ |. We will define στ2 = b |τ |.

Now we can use the standard formula for statistical expectation
ˆ
E [ f ( x )] = f ( x ) g( x ) dx (8)

where g( x ) is the distribution of x. So

h i ˆ
E e i∆τ φ
= ei∆τ φ g(∆τ φ)d(∆τ φ)
ˆ
1
e −∆τ φ
2 / (2σ2 )
⇒ ei∆τ φ p τ d(∆τ φ)
2πστ2

This integral looks a lot like a Fourier transform of a Gaussian (which gives a Gaussian).
r
h
− ax2
i π − ω2
F e → e 4a
a

and so we get h i 2
E ei∆τ φ = e−στ /2 = e−b|τ |/2 (9)

Therefore the auto-correlation is

R(τ ) = | A0 |2 eiω0 τ e−b|τ |/2 (10)

The Fourier transform of the autocorrelation gives the power spectrum of the signal via
the Wiener–Khinchin theorem (appendix A). The Fourier transform of this expression is a
Lorentzian centred at ω0 . But since it will be useful later, I will perform the integration

4
explicitly1 .
ˆ ∞
F [ R(τ )]
= eiω0 τ e−b|τ |/2 e−iωτ dτ
| A0 |2 −∞
ˆ ∞
= eiω0 τ −b|τ |/2−iωτ dτ
−∞
ˆ ∞ 
−bτ/2−i (ω −ω0 )τ
= 2< e dτ
0
" #∞
e−bτ/2−i(ω −ω0 )τ
= 2<
−b/2 − i (ω − ω0 )
  0
1
= 2<
b/2 + i (ω − ω0 )
" #
b/2 − i (ω − ω0 )
= 2<
b2 /4 + (ω − ω0 )2
b
=
b2 /4 + (ω − ω0 )2

Define L(w, λ) which is the Lorentzian distribution, centred at 0 and with a (FWHM) linewidth
of λ.
1 λ/2
L(w, λ) = (11)
π λ /4 + (ω − ω0 )2
2

Therefore the power spectral density of the laser signal is

nε 0 c
P(ω ) = | A0 |2 2πL(w − ω0 , b) (12)
2
where the factor at the front comes from the definition of irradiance for complex notation
nε 0 c
I (t) = | E0 |2 (13)
2
The factor of 2π present in eqn.12 is initially confusing but can be explained using Par-
seval’s theorem which states that
ˆ ∞ ˆ ∞
2 1
| x (t)| dt = | X (ω )|2 dω
−∞ 2π −∞

when using angular frequency (ω) and so the 2π factor is expected.

Often discussions of phase noise get a bit confused between the power spectrum of the
phase noise and the power spectrum of the signal. This arises due to the two spectra be-
coming almost equivalent some distance from the central frequency of the signal. Since the
phase noise is a Wiener process, the power spectrum of the phase noise is

b
Pφ (ω ) = (14)
ω2
1 I’m going to do all these derivations with the non-unitary form of the Fourier transform and using natural
´∞
frequencies. ie.F [ f (t)] (ω ) = −∞ f (t)e−iωt dt

5
as shown in appendix B.
In a normal ’phase noise’ measurement of an oscillator, it will actually be the power
spectral density that is measured. It will be assumed that amplitude noise is minimal and
that therefore the spectrum can be taken to be the phase noise spectrum. Although this is
only really accurate for frequencies greater than the linewidth of the oscillator.

4 Derivation using real notation

Finally, we ask whether complex notation fails for signals near 0 Hz. Consider

R0 (τ ) = E A20 cos(ω0 t + ω0 τ + φ (t + τ )) cos(ω0 t + φ (t))

 
(15)
0
⇒ R (τ ) = A20 /2 E [cos(ω0 τ + φ (t + τ ) − φ (t)) + cos(2ω0 t + ω0 τ + φ (t + τ ) + φ (t))]

we neglect the second term since it looks like it will average to zero

⇒ R0 (τ ) = A20 /2 E [cos (ω0 τ + ∆τ φ)]

= A20 /4 E [exp (iω0 τ + i∆τ φ) + exp (−iω0 τ − i∆τ φ)]
= A20 /4 {exp (iω0 τ − b |τ | /2) + exp (−iω0 τ − b |τ | /2)} (16)

!
A2 b b
⇒ P0 (w) = nε 0 c 0 +
4 b2 /4 + (ω − ω0 )2 b2 /4 + (ω + ω0 )2
A20
= nε 0 c (2πL(ω − ω0 , b) + 2πL(ω + ω0 , b)) (17)
4
For this case, we used the definition of irradiance for electric field

I (t) = nε 0 c E(t)2 = nε 0 cE E(t)2

 
(18)

(the second form is slightly confusing due to a conflict in notation between electric field and
expectation value). There is a factor of 2 difference between the real and complex derived
results. This might be explained if the measured power must be the sum over both positive
and negative frequencies2 i.e.

P0 (ω ) = P(ω ) + P(−ω ) (19)

Importantly though, we see that aliasing will occur around 0 Hz and if ω0 ≈ 0 Hz then
the signal will be 3 dB higher due to this aliasing (whether there is a factor of 2 missing in
the result or we must count both positive and negative frequencies, this result is the same).

2 Please note that I am not entirely certain of this.

6
Part II
Unbalanced Interferometers &
Self-Heterodyne Measurements
We will now derive the power spectrum for a laser passing through an interferometer with
and without a frequency shift in one of the paths. Similar derivations can be found in [13,
14, 15, 16, 17, 18, 19, 3]

5 Derivation of theory for heterodyne case

Consider a general equation for an interferometer.

Aout = A1 eiω0 (t−T )+iφ(t−T ) + A2 ei(ω0 +∆ω )t+iφ(t)+iφ (t)

0
(20)
This equation can account for various setups, for instance a Mach-Zender interferometer, or
a self-heterodyne measurement where the path lengths greatly differ. However, note that
the following derivation will assume that ∆ω 6= 0 Hz (see sec.9 for the homodyne case). The
equation has also included the possibility that one path could introduce its own phase noise
φ0 (t). The two paths have time difference of T (dispersion is not considered in this equation).
The intensity will be

I ∝ | Aout |2 (21)
2 2 0
∝ | A1 | + | A2 | + 2A1 A2 cos ∆ωt + φ(t) − φ (t − T ) + φ (t) + ω0 T


(22)

Measured by a fast enough detector, this gives a photocurrent (in complex notation)

i (t) = i1 + i2 + 2 i1 i2 ei(∆ωt+φ(t)−φ(t−T )+φ (t)+ω0 T + ϕ)

p 0
(23)

where ϕ is a fixed phase difference due to the interferometer’s couplers/beam-splitters (which

was hidden in A1 A2 before). To work out the spectrum of this current, we calculate the auto-
correlation as before

R(τ ) = E i (t + τ ) i (t)∗
 
(24)
This calculation of the autocorrelation actually misses some terms for the case where ∆ω = 0
but we consider this case later in sec.9. Nonetheless, continuing the derivation we get

R ( τ ) = ( i1 + i2 )2
h 
+ E 2 (i1 + i2 ) i1 i2 ei(∆ω (t+τ )+φ(t+τ )−φ(t+τ −T )+φ (t+τ )+ω0 T + ϕ)
p 0


+e−i(∆ωt+φ(t)−φ(t−T )+φ (t)+ω0 T + ϕ)
0

i
+4i1 i2 ei(∆ωτ +φ(t+τ )−φ(t+τ −T )−φ(t)+φ(t−T )+φ (t+τ )−φ (t))
0 0
(25)

When working with expectations, we can use

E [ X + Y ] ⇒ E [ X ] + E [Y ]

7
 Δφ Δφ
T T

t-T t+τ-T t t+τ τ< T

t-T t t+τ-T t+τ τ>T

Δτφ Δτφ

Figure 1: Representation of the possible statistically independent (non-overlapping) inter-

vals in the phase noise time line for different sizes of τ wrt. T. Only positive values of τ are
considered here but negative values should be considered also.

even if X and Y are not statistically independent; therefore we can consider each term separ-
ately. The second and third terms look like they will average to zero due to the ei∆ωt term and
even if ∆ω = 0, the terms will be independent of τ and so can only contribute to the DC term
of the spectral density (in the second term a substitution of ϕ = t + τ shows its independence
from τ) . So consider the fourth term and the phase error terms in its exponent:

φ(t + τ ) − φ (t + τ − T ) − φ(t) + φ (t − T )

The phase noises can be modelled as Wiener processes and so this can be simplified to two
Gaussian white noise variables (appendix B) and there are two ways of doing this
either
∆ T φ | τ − ∆ T φ |0
or
∆ τ φ |0 − ∆ τ φ | − T
where the absolute time of the variables is shown explicitly in order to distinguish between
the variables. If we want to further combine these two random variables, we need to know
whether they are correlated or statistically independent. From the properties of Wiener pro-
cesses, we know that the intervals will be independent as long as they don’t overlap. In
fig.1 we see that if τ ≥ T then the first expression ensures that the two terms are statistically
independent but if τ ≤ T then the second expression should be used to get independent
statistical quantities. Now, where the terms are statistically independent then we can write
√
∆ T φ|τ − ∆ T φ|0 = 2∆ T φ = N (0, 2b | T |) |τ | ≥ T (26)

and √
∆ τ φ |0 − ∆ τ φ | − T = 2∆τ φ = N (0, 2b |τ |) |τ | ≤ T (27)
where we see that combining noisy variables uses a different kind of mathematical rules (see
appendix C). It is not completely clear here but the same final results are found when τ is
negative, hence we have included modulus signs around τ in the above expressions.

8
 Now we note that in the case that X and Y are statistically independent

E [ XY ] ⇒ E [ X ] E [Y ]

This means that the forth term can become3

 h √ i
h 0
i  E exp i 2∆ τ φ τ≤T
4i1 i2 ei∆ωτ E ei∆τ φ h √ i
 E exp i 2∆ T φ τ≥T

where ∆τ φ0 = φ0 (t + τ ) − φ0 (t) is the contribution of the extra phase noise that has been
introduced onto one of the interferometer paths; which we will assume is also Brown noise.
We have worked out these expectations before (sec.3) and so can go directly to
(
−b|τ | τ ≤ T
i∆ωτ −b0 |τ |/2 e
R(τ ) = R0 + 4i1 i2 e e (28)
e−b| T | τ ≥ T

where R0 = (i1 + i2 )2 or something else for d.c (i.e. ∆ω = 0) due to the neglected second and
third terms.
This can be written as
0
  τ  
R(τ ) = R0 + 4i1 i2 ei∆ωτ e−b |τ |/2 e−b|T | + rect e−b|τ | − e−b| T | (29)
2T
where rect( x ) is the rectangular function which is equal to 1 for 0.5 < x < 0.5 and zero
elsewhere. We now Fourier transform the autocorrelation to get the spectral density4
(ˆ )
T  
F [ R(τ )] = R0 2πδ(ω ) + 4i1 i2 2πL(ω − ∆ω, b0 )e−b|T | + 4i1 i2 L(ω − ∆ω, b0 ) ∗ e−b|τ | − e−b|T | e−iωτ dτ
−T
(30)
where L(w, λ) is a Lorentzian function, centred at 0 and with a (FWHM) linewidth of λ.

1 λ/2
L(w, λ) =
π λ2 /4 + (ω − ω0 )2

Now consider ˆ T  
e−b|τ | − e−b|T | e−iωτ dτ
−T
3 There is also an alternative method [15] to the derivation given here, that somehow rearranges the four phase

terms to get
1 1
φ(t + τ ) − φ (t + τ − T ) − φ(t) + φ (t − T ) = −∆φ( T ) − ∆φ (τ ) + ∆φ(τ + T ) + ∆φ (τ − T )
2 2

4 Note also that since we are using a non-unitary, angular frequency definition of the Fourier transform, we

have the following

1
F [ f (t) g(t)] → F (ω ) ∗ G (ω )
2π
and h i
F eiat → 2πδ(ω − a)

9
and treat T as positive just to simplify the notation.
"ˆ # ˆ
T T
⇒ 2< e−bτ e−iωτ dτ − e−bT e−iωτ dτ
0 −T

−iωτ  T T
" #
e−bτ e e−iωτ

−bT
⇒ 2< −e
−b − iω 0 −iω −T

e−bT e−iωT − 1 −iωT − eiωT 

  
−bT e
⇒ 2< −e
−b − iω −iω
" #
− bT − iωT


e e − 1 (−b + iω ) sin (ωT )
⇒ 2< − 2e−bT
b + ω2
2 ω
"
#
b 1 − e−bT e−iωT + iω e−bT e−iωT − 1


sin (ωT )
⇒ 2< 2 2
− 2e−bT
b +ω ω
b 1 − e−bT cos (ωT ) + ωe−bT sin (ωT )


sin (ωT )
⇒2 − 2e−bT
b2 + ω 2 ω
This can be simplified to
  
−bT b
⇒ 2πL (ω, 2b) 1 − e cos (ωT ) + sin (ωT )
ω
Therefore the power spectral density is
P(ω ) ∝ R0 2πδ(ω ) + 4i1 i2 2πL(ω − ∆ω, b0 )e−bT +
   
0 −bT b
4i1 i2 L(ω − ∆ω, b ) ∗ 2πL (ω, 2b) 1 − e cos (ωT ) + sin (ωT ) (31)
ω
Note that there will be some extra terms for a homodyne interferometer i.e. when ∆ω = 0
(see sec.9). Note also that if b0 → 0 then L(ω − ∆ω, b0 ) → δ(ω − ∆ω ).

6 Implications
If we consider eqn.31, we can see that it consists of a dc term, a term due to extra phase
noise introduced between the two paths of the interferometer and a third term that shows
interference between the two paths. This last term includes a convolution with part of the
second term which obviously serves to broaden any features by the linewidth of the extra
phase noise term. Both the second and third terms are also influenced by the coherence
between the signals from the two paths, evident through e−bT .
So for a heterodyne measurement on a interferometer that doesn’t introduce any extra
phase noise onto its paths but does shift the frequency the signal in one arm, we have the
following power spectrum.
P(ω ) ∝ R0 2πδ(ω ) + 4i1 i2 e−bT 2πδ(δω )
  
−bT b
4i1 i2 .2πL (δω, 2b) 1 − e cos (δωT ) + sin (δωT ) (32)
δω

where R0 = (i1 + i2 )2 and δω = ω − ∆ω. We plot an example in fig.2

10
Figure 2: A plot of the power spectrum for a linewidth of b=300 kHz and a difference
between the two paths of T=50 ns.

7 Self-heterodyne linewidth measurements

In a normal self-heterodyne measurement of a laser linewidth[20, 21], one path is much
longer than the other. Starting from eqn.25

h i
R(τ ) = R0 + 4i1 i2 E ei(∆ωτ +φ(t+τ )−φ(t+τ −T )−φ(t)+φ(t−T )+φ (t+τ )−φ (t))
0 0

and letting T → ∞ (and b0 = 0), we reach the simplified autocorrelation (see eqn.28)

R(τ ) = R0 + 4i1 i2 ei∆ωτ e−b|τ | (33)

for all τ, or at least for all τ of interest. This gives
P(ω ) ∝ R0 δ(ω ) + 4i1 i2 2πL(ω − ∆ω, 2b) (34)
and therefore the rf spectrum of the detector current has twice the linewidth of the laser line.
This is very useful for measuring the linewidths of telecoms lasers as it transforms the
optical linewidth into a r.f. linewidth which can be more easily measured. In a normal setup,
the laser output is split and half is sent down 4 km of optical fibre; this is so that it will be
incoherent with respect to the output of the shorter path. The two paths are recombined
and measured using a photodiode and rf analyser. It is important for one path to include an
acoustic optical modulator to shift the frequency such that this will be a heterodyne meas-
urement. This removes all of the added complications (see sec.9) and noise that occur at
d.c. An additional advantage is that slow drifts of the laser’s frequency are ignored by the
measurement. For 4 km of path difference, any changes occurring slower than 50 kHz are
not seen.
As eqn.31 is a rather complicated expression, we take this opportunity to check that it
simplifies correctly. If we let T → ∞ and b0 = 0 in eqn.31 , we find

P(ω ) ∝ R0 δ(ω ) + 4i1 i2 δ(ω − ∆ω ) ∗ {2πL(ω, 2b)}

⇒ R0 δ(ω ) + 4i1 i2 2πL(ω − ∆ω, 2b)

11
 EDFA
Master Isolator OSA
DUTT-TDeviceTUnderTTest
Laser 90%
PC 50%

10% 50% RFTanalyser

PD
coupler coupler

Figure 3: Schematic layout of the interferometric setup used in the first experiment. The
device under test (DUT) was actually a laser under the control of an optical phase lock loop.
The erbium doped fibre amplifier (EDFA) is in the path of the device being tested in order
to amplify to output signal but this introduces an excess path length of 60 m. OSA -optical
spectrum analyser, PD - photodiode, RF analyser - radio frequency analyser, PC - polarisa-
tion controller.

8 Fitting some heterodyne phase noise measurements

We took some measurements which heterodyned the output of an optical phase locked loop
(OPLL) with its master laser (Fig.3 and [22]). An OPLL is a system that locks the phase of a
laser output to phase of an input signal. This might seem to pointlessly replicate the input
signal but a good OPLL will amplify in the input while rejecting noise at frequencies away
from the desired signal. Also an OPLL can simultaneously introduce a frequency difference
between the input master signal and the output, which is useful for low noise measurements.
Noise is always worse near 0 Hz and so measuring a signal at higher frequency gives better
signal to noise ratios. RF engineers call this heterodyning as apposed to homodyning.
Unfortunately, the path with the OPLL was longer than the other path . This meant that
we weren’t measuring the true phase noise of the heterodyne between the OPLL output
and the master laser. We will see whether the previously derived unbalanced interferometer
equations can describe our results. The results were a phase noise measurement (really a
power/Hz vs frequency measurement), therefore the equation to fit to the data is
    
0 −bT 0 −bT b
A0 L(δω, b )e + L(δω, b ) ∗ 2πL (ω, 2b) 1 − e cos (ωT ) + sin (ωT )
ω
(35)
where A0 is a constant, δω = ω − ∆ω and L(w, λ) is a Lorentzian function, centred at 0 and
with a (FWHM) linewidth of λ.

1 λ/2
L(w, λ) =
π λ2 /4 + (ω − ω0 )2

If we assume that b0 is quite small and in particular, is smaller than 2π/T then we might
approximate the equation with
   
b
A0 L(δω, b0 )e−bT + 2πL (δω, 2b) 1 − e−bT cos (δωT ) + sin (δωT ) (36)
δω

This equation fitted the data that was taken very well, showing that the result was limited
by the excess path length of 60 m in one the optical paths that was introduced via the erbium
doped amplifier in that path (Fig.4).

12
 heterodyne
b b'
heterodyne
heterodyne

Figure 4: Here the data is shown with eqn.36 plotted with some hand chosen values. The fit
describes the results very well including the oscillations around 10 MHz.

Master DUTT-TDeviceTUnderTTest Isolator OSA

Laser 90%
PC 50%
EDFA

10% 50% RFTanalyser

PC PD
coupler PLM coupler
Attenuator

Figure 5: Schematic layout of the interferometric setup used in the second experiment[22].
The DUT was actually an optical phase lock loop. The erbium doped fibre amplifier has been
moved and the path lengths balanced. PLM - path length matching.

13
Figure 6: The new phase noise measurement (black) compared with the old measurements
(green). The new phase noise measurements show that the excess path length was limiting
our measurement of the true phase noise of the system. The new phase noise measurements
were reported in [22]

The experiment was then repeated but the erbium doped amplifier was moved and the
two paths of the measurement were balanced in length to better than 1 metre (Fig.5). The
new phase noise measurements reveal a much lower noise floor and improved phase noise
variance (Fig.6).

14
9 Homodyne Interferometry
In this section, we will consider what happens when ∆ω = 0, i.e. when there is no frequency
shifter in either arm of the interferometer. We will also consider what happens around 0 Hz
and so we will have to consider the 2nd and 3rd terms that we discarded from eqn.25. Re-
turning to eqn.22, we start by

i (t) = i1 + i2 + 2 i1 i2 cos φ(t) − φ (t − T ) + φ0 (t) + ω0 T + ϕ

p

and as before the autocorrelation is given by

R(τ ) = E i (t + τ ) i (t)∗
 

however, this time we use the real functions rather than their complex substitutes. To keep
things simple, we redefine the photocurrent to be

i (t) = i a + 2ib cos (Σ(t) + ω0 T + ϕ) (37)

√
where i a = i1 + i2 , ib = i1 i2 and Σ(t) = φ(t) − φ (t − T ) + φ0 (t).
So the autocorrelation becomes

R(τ ) = i2a + 2i a ib E [cos (Σ(t + τ ) + ω0 T + ϕ)] +

2i a ib E [cos (Σ(t) + ω0 T + ϕ)] +
4i2b E [cos (Σ(t + τ ) + ω0 T + ϕ) cos (Σ(t) + ω0 T + ϕ)] (38)

The detailed derivation Eqn.38 has four terms. Let’s consider the second and third terms.
Do we know how to calculate terms like E [cos (φ0 (t))] or E [exp ( jφ0 (t))]? I suspect that the
expectation is zero but I haven’t proven it, instead I will assume that φ0 (t) = 0 to keep things
simple. The second and third terms now become

(2nd + 3rd) ⇒ 2i a ib E [cos (N (0, b | T |) + ω0 T + ϕ)] + 2i a ib E cos N 0 (0, b | T |) + ω0 T + ϕ




where N (0, b | T |) is a white gaussian variable with a mean value of zero and a variance of
b | T |. We worked this type of expression out before using E [exp (i N (0, b | T |))] = exp (−b | T | /2)
and so we get
(2nd + 3rd) ⇒ 4i a ib exp (−b | T | /2) cos (ω0 T + ϕ)
Therefore the second and third terms contribute a constant term (with respect to τ) show-
ing a decaying oscillation with increasing path difference. However this is just the normal
interference that one would see with a Mach-Zender! We can also see this from finding the
expectation of i (t) (for ∆ω = 0).
p
E [i (t)] = i1 + i2 + 2 i1 i2 exp (−b | T | /2) cos (ω0 T + ϕ)

Now we consider the forth term; first it can be written as a sum of two cosines.
1
(4th) ⇒ 4i2b E [cos (Σ(t + τ ) − Σ(t)) + cos (Σ(t + τ ) + Σ(t) + 2ω0 T + 2ϕ)]
2

15
The second part of this expression wasn’t part of the proof in sec.5, when using the com-
plex form for the photocurrent this term never appears. However, if we had left ∆ω in the
expressions up to this point then we would found our expression to be

1
(4th) ⇒ 4i2b E [cos (∆ωτ + Σ(t + τ ) − Σ(t)) + cos (∆ω (2t + τ ) + Σ(t + τ ) + Σ(t) + 2ω0 T + 2ϕ)]
2
Hence, if ∆ω 6= 0, then we neglect the second term due to the fact that it oscillates and so its
expectation tends to zero. We can deal with the first part of the term exactly like we did in
sec.5. Once we have expanded the cosine into complex exponentials, the 4hphase  error terms
√ i
are combined into a single gaussian noise term and applying the relation E exp −i 2∆τ φ =
h √ i
E exp i 2∆τ φ helps simplify the expression. So we get
(
−b0 |τ |/2 e−b|τ | τ≤T
(4th ( a)) ⇒ 2i2b e
e−b| T | τ≥T

For consistancy with the second and third terms, I will discard the additional phase noise
terms at this point ie. b0 = 0. So we have
(
e−b|τ | τ ≤ T
(4th ( a)) ⇒ 2i2b −b|T |
e τ≥T

Now we have to consider the second part of the forth term

(4th (b)) ⇒ 2i2b E [cos (2ω0 T + 2ϕ + Σ(t + τ ) + Σ(t))]

Firstly, we will drop the φ0 (t) terms (from inside the Σ(t) terms) because these should send
the expectation towards zero (see above) which isn’t very interesting. Now, we can use a
similar trick to that used before in order to combine the difference phase noise terms. We do
(
∆τ φ|0 + 2 ∆ T −|τ | φ τ −T + ∆τ φ|−T |τ | ≤ T
Σ(t + τ ) + Σ(t) ⇒
∆ T φ|τ −T + ∆ T φ|−T |τ | ≥ T

which makes a collection of statistically independent gaussian white noise variables. We can
combine these variables to get
(
N (0, 2b |τ | + 4b ( T − |τ |)) |τ | ≤ T
Σ(t + τ ) + Σ(t) ⇒
N (0, 2bT ) |τ | ≥ T
see appendices C and B. Note also that T is always considered to be positive while τ can be
both positive and negative. So our latest term becomes
(
2 E [exp (i N (0, 4bT − 2b |τ |))] |τ | ≤ T
(4th (b)) ⇒ ib cos (2ω0 T + 2ϕ)
E [exp (i N (0, 2bT ))] |τ | ≥ T
(
e−b(2T −|τ |) |τ | ≤ T
(4th (b)) ⇒ i2b cos (2ω0 T + 2ϕ)
e−bT |τ | ≥ T

16
leading to
  τ  
(4th (b)) ⇒ i2b cos (2ω0 T + 2ϕ) e−bT + rect eb|τ |−2bT − e−bT
2T
when we Fourier transform the above expression, we will need to consider
ˆ T  
eb|τ |−2bT − e−bT e−iωτ dτ
−T
"ˆ # ˆ
T T
⇒ 2< ebτ −2bT e−iωτ dτ − e−bT e−iωτ dτ
0 −T
"ˆ #
T
sin (ωT )
⇒ 2e−2bT < ebτ e−iωτ dτ − 2e−bT
0 ω
" #
ebT e−iωT − 1 (b + iω )


−2bT sin (ωT )
⇒ 2e < 2 2
− 2e−bT
b +ω ω
b ebT cos(ωT ) − 1 + ωebT sin(ωT )


sin (ωT )
⇒ 2e−2bT 2 2
− 2e−bT
 b +ω  ω
b b
⇒2 2 e−bT cos(ωT ) − e−bT − sin (ωT )
b + ω2 ω

The final autocorrelation expression Putting all of our terms together, we finally have the
full autocorrelation expression
R(τ ) = i2a + 4i a ib exp (−b | T | /2) cos (ω0 T + ϕ)
  τ  
+ 2i2b e−b|T | + rect e−b|τ | − e−b| T |
  2T
τ   b|τ |−2bT 
−b| T |
2
+ 2ib e + rect e − e−b|T | cos (2ω0 T + 2ϕ) (39)
2T
for ∆ω = 0 and φ0 (t) = 0. The power spectrum can now be calculated
P(ω ) ∝ 2πδ(ω ) i2a + 4i a ib exp (−b | T | /2) cos (ω0 T + ϕ)

  
2 −bT bT b
+2ib e 2πδ(ω ) + 2πL (ω, 2b) e − cos (ωT ) − sin (ωT ) (40)
ω
  
2 −bT −bT b
+2ib e 2πδ(ω ) + 2πL (ω, 2b) −e + cos(ωT ) − sin (ωT ) cos (2ω0 T + 2ϕ)
ω
comparing this result to eqn.31, we can see some extra terms compared to the heterodyne
phase noise spectrum5 . In particular the last line of the equation is entirely new. Reorgan-
ising the equation gives us
n o
P(ω ) ∝ 2πδ(ω ) i2a + 4i a ib e−bT/2 cos (ω0 T + ϕ) + 2i2b e−bT (1 + cos (2ω0 T + 2ϕ))
+2i2b 2πL (ω, 2b) e−bT f (ω )
5 Asin sec.4 there is a difference of 2 between the complex and real proofs (even after the difference of 2
between their intensity definitions are accounted for). Is it that in the real case the measured power is the sum
over both positive and negative frequencies i.e
P0 (ω ) = P(ω ) + P(−ω )
while for the complex case, we only consider positive frequencies? I’m not sure.

17
where
    
bT b −bT b
f (ω ) = e − cos (ωT ) + sin (ωT ) + cos (2ω0 T + 2ϕ) −e − − cos(ωT ) + sin (ωT )
ω ω

Now we use some simply trigonometric relations to reorganise the equation

 
2 bT −bT 2b
⇒ f (ω ) = cos (ω0 T + ϕ) e − e − sin (ωT )
ω
 
+ sin2 (ω0 T + ϕ) ebT + e−bT − 2 cos (ωT )

 
2 b
⇒ f (ω ) = 2 cos (ω0 T + ϕ) sinh(bT ) − sin (ωT )
ω
+2 sin2 (ω0 T + ϕ) (cosh(bT ) − cos (ωT )) (41)

If we put ϕ = 0, b = τ2c where τc is the just another way to define the laser linewidth, then
we can see that we have rederived eqn.8c in [18]. For this homodyne setup, we can see that
we have an extra degree of freedom in the system since we can choose between the two
terms of f (ω ) by relatively small changes in the path difference or ϕ. The two terms lead
to quite different power spectra although we also affect whether we are at a maximum or
minimum for the d.c. component of the interferometer. Knowledge of these formulae can
help in optimising the sensitivity of an interferometer (see [18] for details).

18
Part III
Appendices
A Wiener–Khinchin theorem
The Wiener–Khinchin theorem states that the power spectral density is the Fourier transform
of the autocorrelation. This can be summarised as

P(ω ) = Fτ [ R(τ )] (42)

where R(τ ) is the autocorrelation. Autocorrelation is defined by

ˆ ˆ
∗
R f (τ ) = f (t + τ ) f (t) dt = f (t) f ∗ (t − τ ) dt = f ? f

In fact, autocorrelation has several definitions, some of which normalise the quantities being
multiplied by their variances and means. Expect to be confused by operations variously
defined as covariance, autocovariance, autocorrelation etc.
The Wiener–Khinchin theorem can be easily proven for normal functions via the convo-
lution theorem (F [ f (t) ∗ g(t)] = F (ω ) G (ω )) since

f ? f = f (t) ∗ f (−t)∗

and
F [ f (−t)∗ ] = F (ω )∗
therefore
F [ f ? f ] = | F (ω )|2
However the theorem is more powerful than this implies since it also applies to functions
that can not normally be Fourier transformed or integrated. Phrases often encountered are
’wide sense stationary process’ and ’ergodic’. In this case, the autocorrelation can be re-
defined as
R f (τ ) = E [ f (t) f (t − τ )∗ ] (43)
where E [] signifies expectation. This proves very useful.

B Wiener Processes
In the study of noise known as stochastics, a Wiener process is an integration of a Gaus-
sian white noise variable. Although this is not a standard integration, as when working
with noisy variables the normal rules of calculus must be adjusted leading to Itō calculus.
A Wiener process is also known as Brown noise (short for Brownian noise) and is directly
related to the random walks in Brownian motion 6 . It also turns out to be a good theoretical
model for phase noise (see part I). A Wiener process is characterised by [12]

1. W (0) = 0
6 These paragraphs contain many good keywords for searching for more material

19
 2. W (t) is almost surely continuous

3. W (t) has independent increments with W (t) − W (s) = N (0, t − s) (for 0 ≤ s ≤ t)

where N (µ, σ2 ) indicates a random variable with a normal distribution with expected value
µ and variance σ2 .
1 2 2
N =√ e−(x−µ) /(2σ )
2πσ 2

A few properties of the Wiener process:

• E [W (t)] = 0

• covariance: cov (W (t), W (s)) = min (s, t)

q
min(s,t)
• correlation: corr (W (t), W (s)) = max(s,t)
h i
• E (W (t) − W (s))2 = |t − s|

• increments of a Wiener process that don’t overlap are independent variables.

• A Wiener function can’t strictly be differentiated although in actuality this can be got
around (somehow).

• The Fourier transform of a Wiener process is F [W (t)] =?

1 σ2
• The spectral density of a Wiener process is w2
or in general ω2
(as expected for Brown
noise).

To understand the power spectral density remember that the Wiener process is the integra-
tion of Gaussian white noise (using Itō calculus). White noise has a power spectral density
of σ2 Watts/Hz across all frequencies; this implies infinite power and so is physically im-
possible but nonetheless we find that white noise is a useful approximation to real noise. If
we also recall the following property of Fourier transforms:
 n 
d g
F = (iω ) n G (ω )
dtn

or more applicably ˆ 
G (ω )
F gdt =
iω
We can then see that the Wiener process will have a power spectrum of ω12 (since σ2 = 1 in
the pure Wiener process).
The last proof feels a bit like a cheat, although only as I don’t know whether these rules
apply to Itō calculus. So we look for another proof to find the power density spectrum. This
time we calculate the autocorrelation function of the Wiener process

R(τ ) = E [W (t + τ ) W (t)]
= min (t + τ, t)

20
(temporarily ignore the fact that we don’t known the value of t) This can be rewritten as

R(τ ) = t − Ramp (−τ )

where
x + |x|
Ramp( x ) = = xH ( x )
2
where H ( x ) is the Heaviside step function. The Fourier transform of the ramp function is
(non-unitary, natural frequency)

dδ (w) 1
F [Ramp(t)] = iπ − 2
dt ω
and therefore the power spectrum is given by

1 dδ (w)
PW (ω ) = 2
+ iπ + t2πδ (ω )
ω dt
see sec.A. We now ignore all of the strange things that happen at ω = 0 and find

1
PW (ω ) =
ω2
as expected. √
To transform a random variable, people will often write things like N (0, 1) t → N (0, t)
which is confusing when you look at the definition and think in terms of normal functions
but what they mean is that if you multiply a random variable then the mean is increased by
that factor and variance is increased by the square of that factor. ie.

aX + c ∼ N X ( aµ + c, a2 σ2 )

Therefore to model the phase noise with a Wiener process, we define the phase noise as
√
φ(t) = bW (t)

where b is the variance (σ2 ) of the phase noise. (Intuitively one might have done √1 W ( bt )
b
but actually this is just equal to W (t)).
We can now say that the phase noise spectral density is given by
h i b
Φ(ω ) = E |F [φ(t)]|2 = 2 (44)
ω
The expectation notation isn’t strictly necessary since the Fourier transform is over all time
but in a realistic situation, the Fourier transform will be replaced by an approximation.

C Random Variables
If we are dealing with a Gaussian random variables then we must remember that the rules of
addition and subtraction are different from normal. Consider a random variable N X (µ, σ2 )

21
where N (µ, σ2 ) indicates a random variable with a normal distribution with expected value
µ and variance σ2 .
1 2 2
g(N X ) = √ e−(x−µ) /(2σ )
2πσ2
If we transform the variable X → aX + c, we can write this equivalently as N X ( aµ + c, a2 σ2 ).
Now consider two statistically independent variables X and Y

X+Y = Z
≡ N X (µ X , σX2 ) + NY (µY , σY2 ) = N Z (µ X + µY , σX2 + σY2 )

X−Y = Z
≡ N X (µ X , σX2 ) − NY (µY , σY2 ) = N Z (µ X − µY , σX2 + σY2 )

XY = Z
≡ N X (µ X , σX2 )NY (µY , σY2 ) = N Z (µ X µY , µY2 σX2 + µ2X σY2 )

X/Y = Z
σX2 X2 2
≡ N X (µ X , σX2 )NY (µY , σY2 ) ≈ N Z (µ X /µY , + σ )
Y2 Y4 Y
The last case is not strictly true since a division of two Gaussian variables leads to a Cauchy
distributed random variable. In general for a function of statistically independent Gaussian
white noise variables, we have
f ( X, Y, Z )
δf 2 2
   2  2
δf δf
σ2f ≈ σX + σY2 + σZ2
δX δY δZ
However, if the variables are correlated, then we should use for example

aX ± bY = Z

σZ2 = a2 σX2 + b2 σY2 ± 2ab cov AB

where cov AB = E [( A − Ā) ( B − B̄)]. We can now consider the case

X+X = Z

and find that

σZ2 = σX2 + σX2 + 2σX2 = 4σX2
which is what we would expect for
2X = Z
In the derivations of this note, we most often add or subtract two statistically independ-
ent variables with means of zero and the same variance

N X (0, στ2 ) + NY (0, στ2 ) = NZ (0, 2στ2 )

22
which leads to the slightly contra-intuitive statement
√
N X (0, στ2 ) + NY (0, στ2 ) = 2NZ (0, στ2 )

This topic is known as the ’propagation of uncertainty’ and can be found in many books.
In the most general case, we would need to construct a matrix of the correlations between
the variables in order to calculate the error accurately.

References
[1] E. Rosencher and B. Vinter. Optoelectronics. Cambridge University Press, 2002.

[2] A. Yariv. Optical Electronics. Oxford Series in Electrical and Computer Engineering.
Oxford University Press, 1990.

[3] A. Yariv. Quantum electronics. Wiley, 1989.

[4] Melvin Lax. Theory of laser noise. In Rajarshi Roy, editor, Theory of laser noise, volume
1376, pages 2–20. SPIE, 1991.

[5] C. Henry. Theory of the linewidth of semiconductor lasers. Quantum Electronics, IEEE
Journal of, 18(2):259–264, 1982.

[6] K. Vahala and A. Yariv. Semiclassical theory of noise in semiconductor lasers–part ii.
Quantum Electronics, IEEE Journal of, 19(6):1102–1109, Jun 1983.

[7] T.H. Lee and A. Hajimiri. Oscillator phase noise: A tutorial. Solid-State Circuits, IEEE
Journal of, 35(3):326–336, 2000.

[8] A. Hajimiri and T.H. Lee. A general theory of phase noise in electrical oscillators. Solid-
State Circuits, IEEE Journal of, 33(2):179–194, 1998.

[9] Hp archive - seminar notes. http://www.hparchive.com/seminar_notes.htm.

[10] Enrico Rubiola. Phase noise. www.rubiola.com.

[11] N.J. Kasdin. Discrete simulation of colored noise and stochastic processes and 1/f<
sup> α</sup> power law noise generation. Proceedings of the IEEE, 83(5):802–827, 1995.

[12] Wiener process [wikipedia]. http://en.wikipedia.org/wiki/Wiener_process. 2011.

[13] P. Gallion, FJ Mendieta, and R. Leconte. Single-frequency laser phase-noise limitation

in single-mode optical-fiber coherent-detection systems with correlated fields. JOSA,
72(9):1167–1170, 1982.

[14] P. Gallion and G. Debarge. Quantum phase noise and field correlation in single fre-
quency semiconductor laser systems. Quantum Electronics, IEEE Journal of, 20(4):343–
349, 1984.

[15] Petar Pepeljogoski. Some useful formulas for analysis of interferometric noise.
http://www.ieee802.org/3/ae/public/adhoc/serial_pmd/documents/useful_IN_
formulas.pdf.

23
 [16] K. Petermann and E. Weidel. Semiconductor laser noise in an interferometer system.
IEEE Journal of Quantum Electronics, QE-17:1251–1256, 1981.

[17] JA Armstrong. Theory of interferometric analysis of laser phase noise. JOSA, 56(8):1024–
1028, 1966.

[18] B. Moslehi. Noise power spectra of optical two-beam interferometers induced by the
laser phase noise. Lightwave Technology, Journal of, 4(11):1704–1710, 1986.

[19] B. Moslehi. Analysis of optical phase noise in fiber-optic systems employing a laser
source with arbitrary coherence time. Lightwave Technology, Journal of, 4(9):1334–1351,
1986.

[20] T. Okoshi, K. Kikuchi, and A. Nakayama. Novel method for high resolution measure-
ment of laser output spectrum. Electronics letters, 16(16):630–631, 1980.

[21] Peter Horak and Wei H. Loh. On the delayed self-heterodyne interferometric technique
for determining the linewidth of fiber lasers. Opt. Express, 14(9):3923–3928, May 2006.

[22] Robert J. Steed, Francesca Pozzi, Martyn J. Fice, et al. Monolithically integrated hetero-
dyne optical phase-lock loop with RF XOR phase detector. Opt. Express, 19(21):20048–
20053, Oct 2011.

[23] Hanne Ludvigsen, Mika Tossavainen, and Matti Kaivola. Laser linewidth measure-
ments using self-homodyne detection with short delay. Optics Communications, 155:180
– 186, 1998.

[24] E. Milotti. Plnoise: a package for exact numerical simulation of power-law noises. Com-
puter physics communications, 175(3):212–225, 2006.

[25] L. Richter, H. Mandelberg, M. Kruger, and P. McGrath. Linewidth determination from

self-heterodyne measurements with subcoherence delay times. Quantum Electronics,
IEEE Journal of, 22(11):2070 – 2074, nov 1986.

[26] S.C. Terry, B.J. Blalock, J.M. Rochelle, M.N. Ericson, and S.D. Caylor. Time-domain
noise analysis of linear time-invariant and linear time-variant systems using matlab
and hspice. Nuclear Science, IEEE Transactions on, 52(3):805–812, 2005.

[27] L. Yang, J. Yue, and H. Ehm. Time domain phase noise synthesis for system simulation.
In Circuits and Systems, 2006. MWSCAS’06. 49th IEEE International Midwest Symposium
on, volume 1, pages 414–418. IEEE, 2006.

24

View publication stats