You are on page 1of 13

IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO.

7, JULY 2013 2635

Randomized Masking in Cognitive Radio Networks


Kamyar Moshksar, Member, IEEE, and Amir K. Khandani, Member, IEEE

Abstract—A decentralized network of one Primary User (PU) significance, many researchers have addressed the problem of
and several Secondary Users (SU) is studied. PU is licensed to resource allocation in wireless networks. One major challenge
exploit the resources, while the party of SUs intend to share the in wireless networks is the destructive effect of multiuser
resources with PU. Each SU must guarantee to not disturb the
performance of PU beyond a certain level, while maintaining interference, which degrades the performance when multiple
a satisfactory quality of service for itself. It is proposed that users share the spectrum. As such, an efficient and low com-
each secondary transmitter adopts a Randomized Masking (RM) plexity resource allocation scheme that maximizes the quality
strategy with full average transmission power where it remains of service and mitigates the impact of multiuser interference
silent or transmits a symbol in its codeword independently from is desirable. Due to the complexity of adapting to the network
transmission slot to transmission slot. We consider a setup where
the primary transmitter is unaware of channel coefficients, code- structure (e.g. number of active users), centralized resource
books of secondary users and the number of secondary users. allocation schemes are usually designed for a fixed network
SUs are anonymous to each other, i.e, they are unaware of each structure. This makes inefficient usage of resources because,
others’ code-books, however, each SU is smart in the sense that in most cases, the number of active users may be considerably
it is aware of the code-book of PU, channel coefficients and less than the value assumed in the design process. Therefore,
the number of active SUs. Invoking the concept of ε-outage
capacity, we define the (ε, ν)-admissible region as the set of it is of interest to devise an efficient and low-complexity
masking probabilities for each SU such that the probability of decentralized resource allocation scheme. In decentralized
outage for PU is maintained under a threshold ε in a case where schemes, decisions concerning network resources are made
PU sets its transmission rate at a fraction ν of its ε-outage by individual nodes based on their local information. Most of
capacity as if there were no SUs in the network. The masking decentralized schemes reported in the literature rely on either
probability of SUs is designed through maximizing the average
(with respect to channel coefficients) achievable rate per SU over game-theoretic approaches or cognitive radios.
the (ε, ν)-admissible region. In our analysis, the primary receiver Distributed strategies based on game-theoretic arguments
treats interference as noise, however, each secondary receiver have already attracted a great deal of attention [1]–[4]. In
has the option to decode and cancel the interference caused by
PU, while treating the signals of other SUs as noise. In another [1], the authors introduce a non-cooperative game-theoretic
approach, referred to as Continuous Transmission with Power framework to investigate the spectral efficiency issue when
Control (CTPC), each SU transmits continuously (no masking is several users compete over an unlicensed band with no central
applied), however, it adjusts its transmission power in order to controller. Reference [2] offers a brief overview of game-
yield the largest value for average achievable rate per SU. The theoretic dynamic spectrum sharing. Repeated non-cooperstive
schemes RM and CTPC are compared for different values of
transmission power for each SU and PU and distance between market game methods are adopted in [3] for resource alloca-
different users. It is observed that neither of RM or CTPC tion in a decentralized network. In [4] the interaction between
always outperforms the other in various scenarios in terms of the multiple random access networks has been considered from
underlying system parameters. A combination of RM and CTPC a game-theoretic point of view. Another scenario is sharing
referred to as Randomized Masking with Power Control (RMPC) the spectrum by a certain number of users competing over
is also investigated where each SU controls both its probability
of masking and average transmission power. It is demonstrated a certain open bandwidth [5]–[8]. In [7], through an asyn-
through simulations that RMPC can outperform both RM and chronous distributed pricing scheme, users exchange signals
CTPC. that indicate the negative effect of interference at the receivers.
Index Terms—Cognitive radio, decentralized resource alloca- In [8], users affected by the mobility event, self-organize into
tion, outage capacity, power control, randomized masking. bargaining groups and adapt their spectrum assignment to
approximate a new optimal assignment.
I. I NTRODUCTION Cognitive radios [9] are flexible devices that have the ability
A. Related Work to sense the unoccupied portion of the available spectrum
and use this information in resource allocation. In its prime
I NCREASING demand for wireless applications on one
hand, and the limited available resources on the other hand,
provoke more efficient usage of such resources. Due to its
definition, a cognitive transmitter (also known as SU) was
assumed to only transmit over the white-spaces, i.e., the
cognitive transmitter remains inactive until the so-called PUs
Manuscript received June 30, 2012; revised February 18, 2013. The editor become silent. The model suggested in [10] is a landmark
coordinating the review of this paper and approving it for publication was E. in the literature on cognitive radios where it is proposed
Larsson.
The authors are with the Department of Electrical and Computer that in order to start its transmission, a cognitive transmitter
Engineering, University of Waterloo, Waterloo, ON, Canada (e-mail: does not need to wait for PU to become idle. In fact, the
kmoshksa@uwaterloo.ca, khandani@shannon2.uwaterloo.ca). cognitive transmitter can transmit simultaneously with PU as
A preliminary version of this paper was presented in Asilomar Conference
on Signals, Systems and Computers, Asilomar 2011. See [28]. long as it adopts an elegant transmission scheme and designs
Digital Object Identifier 10.1109/TCOMM.2013.050813.120450 its code-book properly. In [10] the network of one PU and one
0090-6778/13$31.00 
c 2013 IEEE
2636 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 7, JULY 2013

cognitive user is modelled by a two-user Interference Channel as if the other users were absent. Some restrictive assump-
(IC) such that the cognitive user is aware of the message of tions are made to simplify the presentation, e.g., the channel
PU. Two scenarios are considered, referred to as non-causal coefficients are symmetric and the cognitive users do not
and causal cognitive networks. interfere with each other. Also, [22] studies the SNR scaling
In the non-causal case, the cognitive transmitter is aware of sum rate in an IC with multiple users where each cognitive
of PU’s message non-causally through the aid of a “genie”. transmitter only knows the messages of other surrounding
In this case, a combined signaling strategy is proposed by cognitive transmitters that are located in its vicinity. The so-
invoking Gel’fand-Pinker Coding [11] (or its counterpart in called clustered local processing is allowed at the receivers
the Gaussian IC called Dirty Paper Coding (DPC) [12]) and where each receiver has access to the signals received by the
the Han-Kobayashi scheme [13] that is commonly used in nearby users.
ICs. In fact, the cognitive transmitter utilizes its knowledge of
PU’s message to mitigate the interference on its receiver, while B. Contributions
relaying a replica of PU’s message to the primary receiver.
Before proceeding to this section, let us provide a list of
In the causal case, users go through two different phases.
notations used throughout the paper:
In phase I, the cognitive transmitter is in the “listening mode”
Notations- The set of real numbers is shown by R. Random
where it discovers the message of PU as PU broadcasts its
quantities are shown in bold such as x with realization x.
message to its receiver and the cognitive transmitter. In phase
Vectors are shown by an arrow on top such as x. The transpose
II, the cognitive transmitter proceeds as in the non-causal case.
of a vector x is denoted by x t . The probability of an event
Following the model proposed in [10], the authors in [14]
E is shown by P(E). The Probability Density Function (PDF)
find the largest achievable rate for the cognitive user under the
and the expectation of a random variable x are shown by px (·)
following conditions:
and E[x], respectively. The differential entropy of a continuous
• The primary transmitter must achieve a transmission rate random variable x is denoted by h(x). The mutual information
as large as the capacity of its direct link as if the cognitive between two random variables x and y is shown by I(x; y).
user was inactive. The binary entropy function is shown by hb (x)  −x log x −
• PU does not perform multiuser decoding, i.e., PU is
(1 − x) log(1 − x) for x ∈ [0, 1]. A Bernoulli random variable
unaware of SU’s code-book. x taking values in {0, 1} such that P(x = 1) = a is shown
This is an interesting setup in the sense that it is in agreement by Ber(a). A uniform random variable x taking values in the
with more realistic scenarios, because one may not expect any interval [a, b] for a, b ∈ R is shown by Unif(a, b). A circularly
cooperation between the primary and cognitive users. How- symmetric Gaussian random variable x with zero mean and
ever, the assumption that the cognitive user is aware of PU’s variance σ 2 is shown by CN(0, σ 2 ). The PDF of this Gaussian
|x|2
message non-causally still demands some sort of cooperation random variable is denoted by ϕ(x; σ 2 )  πσ1 2 e− σ2 . A real
between the two users. In some scenarios if one can justify Gaussian random variable x with zero mean and variance σ 2 is
that the message sent by PU is delayed along its direct link, shown by N(0, σ 2 ). Throughout the paper, C(x)  log(1 + x).
then it is reasonable to assume that the cognitive transmitter We consider a network of one PU and several SUs modelled
can discover PU’s message before it is received at the primary by a Gaussian IC. We assume the number of SUs is fixed along
receiver. Note that the observation made in [14] is a capacity the communication period of interest once it is set first. PU is
result only in the weak interference regime where the cognitive ignorant in the sense that
transmitter is closer to its affiliated receiver than the primary
• The primary transmitter is unaware of channel coeffi-
receiver. In fact, it is shown that multiuser decoding at the
cients .
primary receiver does not improve the set of achievable rates
• The primary transmitter is unaware of the number of SUs.
attained by the users. Moreover, the authors demonstrate that
• PU is unaware of the code-book of each SU.
interference decoding at the primary receiver can improve the
set of achievable rates in the strong interference regime. Each SU is smart in the sense that
The result in [14] is also observed in a more general setup • It is aware of the code-book of PU.
in a concurrent paper [15] where the authors look at the • It is aware of the number of SUs.
cognitive IC as a special case of an IC with Degraded Message • It is aware of its direct channel coefficient and the coeffi-
Sets (IC-DMS). The capacity region of IC-DMS in the strong cient of channels connecting other secondary transmitters
interference regime is studied in [16], [17]. and the primary transmitter to its receiver.
Motivated by the fact that SUs must attempt to not disturb Note that we only assume each SU is aware of PU’s
the performance of PUs, a novel approach is taken in [18] code-book and not its message. This is in contrast to the
where the capacity region of wireless networks is studied assumption made in some literature on non-causal cognitive
under constraints on the received Signal-to-Noise Ratio (SNR) radios where the secondary transmitter is aware of PU’s
rather than constraints on the power at the transmitters. The message. Therefore, in the setup considered in this paper, the
results in [18] are applied to fading environments in [19]. secondary transmitters are unable to perform DPC, however,
Cognitive ICs with more than one cognitive user have rarely the secondary receivers are capable of multiuser decoding.
been studies. Reference [20] considers a cognitive IC with This paper studies a decentralized cognitive IC where SUs
several cognitive users. Adopting a coding technique based sneak into the network and have to make sure that PU’s
on nested lattices [21], it is established that in the strong performance is always maintained at a satisfactory level,
interference regime, it is possible for all users to achieve rates while the quality of service per SU is convincing as well.
MOSHKSAR and KHANDANI: RANDOMIZED MASKING IN COGNITIVE RADIO NETWORKS 2637

It is proposed that each secondary transmitter follows the ε. In fact, if I 1 is the largest achievable rate1 of PU, the outage
Randomized Masking (RM) scheme [26] with full average event for PU happens if the actual transmission rate R1 of PU
transmission power where it transmits a Gaussian symbol in is larger that I 1 , i.e., I 1 < R1 . Since R  1 ≤ I 1 , we have
its codeword with a probability of θ or remains silent with 
P(I 1 < R1 ) ≤ P(R1 < R1 ) < ε. Therefore, P(I 1 < R1 ) is
a probability of 1 − θ independently from transmission slot guaranteed to be less than ε.
to transmission slot. The parameter θ is called the activity For any 2 ≤ k ≤ K + 1, let Rk be the set of all tuples
factor for each SU. This is to provide PU with a partially (R1 , Rk ) such that at least one of the following happens:
interference-free reception. It is worth mentioning that the RM • The rate tuple (R1 , Rk ) is in the capacity region of
scheme in [26] is proposed as a means of coexistence in a the Multiple Access Channel (MAC) consisting of the
decentralized network of several users with no hierarchy. In primary transmitter, the transmitter of SUk−1 and the
fact, users are all similar and they are not grouped as primary receiver of SUk−1 .
and secondary users. In this work the design of θ is not solely • The transmission rate Rk is achievable for SUk−1 if
based on improving the performance of SUs. Controlling the SUk−1 treats the whole interference as noise regardless
level of quality of service for PU is also an important factor of the value of R1 .
in designing θ. We define Rk by
Throughout the paper, PU is labeled as user 1 and its actual
Rk  sup{Rk : (R1 , Rk ) ∈ Rk }.
transmission rate is denoted by R1 . The primary receiver treats
interference as noise, because it is unaware of the code-book of As channel coefficients are realizations of independent
SUs. However, each secondary receiver has two options, i.e., random variables with known distributions and noting that
it can treat the whole interference as noise or it can perform each Rk is a function of the number of SUs and the channel
multiuser decoding by jointly decoding its message and the coefficients, Rk can be thought as a realization of a random
message of PU while treating other SUs as additive noise. variable Rk . Thereafter, the activity factor θ is selected based
Assume there are K secondary users labeled as user 2 to on the rule  
user K + 1. We refer to the SUs by SU1 , · · · , SUK with θ = arg max E Rk ,
(RM)
actual transmission rates R2 , · · · , RK+1 , respectively. The θ∈Aε,ν
static and non frequency-selective channel coefficient from where the expectation is with respect to the channel coeffi-
f
each transmitter to each receiver is given by 1+d α . Here, f is a cients. This design procedure is known as maximizing average
realization of a CN(0, 1) random variable f , the parameter 2α achievable rate per SU over the admissible region and is
is called the path loss exponent and d is the distance between completely distributed, i.e., the resulting θ is computable by
the transmitter and receiver which is a realization of a uniform all SUs. Finally, SUk−1 sets its actual transmission rate at
random variable d to be specified in section II. We emphasize Rk = Rk θ=θ.
that the distances between any transmitter and any receiver We also examine an alternative to the RM scheme called
are not necessarily the same. Note that the proposed model Continuous Transmission with Power Control (CTPC). In this
1
for channel coefficients depends on distance as 1+d α and is approach, each secondary transmitter continuously transmits
known as the bounded path loss model [27]. This is in contrast (no masking is applied), however, the average transmission
to d1α which is the model often adopted in literature. power for SU is adjusted such that the average (with respect
The primary transmitter continuously transmits the Gaus- to channel coefficients) transmission rate per SU is maximized,
sian symbols in its codewords. Since the primary transmitter while keeping the probability of outage for PU below the
is unaware of the channel coefficients and the number of SUs, threshold ε. It is demonstrated through extended simulations
outage probability is an appropriate measure to assess quality in section V that none of the RM and CTPC schemes offers a
of service for this user. Let cε be the largest transmission rate better performance than the other universally for all different
for PU such that its probability of outage in the absence of values of the underlying system parameters. The possibility
SUs is ε. We require the activity factor θ per SU be such of combining RM and CTPC into one single scheme is also
that the probability of outage for PU is less than the same considered where the average transmission power and the
threshold ε if PU sets its actual transmission rate at R1 = νcε masking probability per SU are both optimized in the design
for some ν ∈ (0, 1) in the presence of SUs. In practice, ν process. This is referred to as Randomized Masking with
must be selected close to 1 to guarantee that the quality of Power control (RMPC). It is demonstrated that RMPC can
service for PU is only slightly degraded. Let us define the offer an improvement over the performances of both RM and
(RM)
(ε, ν)-admissible region Aε,ν by CTPC.
It is remarkable that CTPC is already studied in a number
  of valuable papers including [23], [24] and the references
A(RM)
ε,ν  θ : P(R 1 < R1 ) < ε , therein. The authors in [23] consider optimal power allocation
in a network of one PU and one SU where SU is aware of
all channel coefficients from the two transmitters to the two
where R 1 is an achievable rate for PU. Note that R  1 is a receivers. The so-called ergodic and outage capacities for SU
function of the number of SUs and the channel coefficients are maximized under a constraint on the probability of outage
and hence, it is a random variable as long as the primary 1 I is the largest transmission rate for PU such that reliable decoding
transmitter is concerned. The inequality P(R  1 < R1 ) < ε 1
at PU’s receiver is guaranteed under maximum likelihood decoding and
ensures that the probability of PU being in outage is less than sufficiently large code block length.
2638 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 7, JULY 2013

for PU and average or peak power constraints at the secondary


transmitter. Invoking convex optimization techniques, it is
shown how SU regulates its transmission power depending on
the channel state information. In contrast to [23], the current
paper studies a scenario with an arbitrary number of SUs Dps
where each SU has only partial knowledge over channel coef-
Tx1
ficients. Moreover, SUs can perform interference cancellation,
Rx2
while in [23] all receivers treat interference as noise. The
Rx3
idea of multiuser decoding and interference cancellation also
appears in [24] where the authors consider the performance of
Tx3 Tx2
iterative water-filling algorithm [25] in a multi-carrier based PU
two-user interference channel. Rx1

II. S YSTEM M ODEL AND A SSUMPTIONS Dss

We consider a wireless communication network consisting (a) Dss < Dps


of one primary and a number K of secondary transmitter-
receiver pairs. We label PU as user 1 and the SUs as user 2 to
user K + 1. The SUs are referred to by SU1 , · · · , SUK . The
static and non-frequency selective channel coefficient from the Tx1

transmitter of user k to the receiver of user l is denoted by Tx2

hk,l given by Rx2

fk,l PU
hk,l  , d1,1 = 0, (1) Rx1
1 + dαk,l Rx3

for any 1 ≤ k, l ≤ K + 1 where fk,l is a realization of


Tx3
a random variable f k,l ∼ CN(0, 1) representing the fading
coefficient2 and dk,l is a realization of a random variable dk,l
representing the distance between the transmitter of user k
and the receiver of user l. The parameter α can be set as any
Dps
nonnegative real number. The quantity 2α is referred to as the
path loss exponent. The collection of fading random variables (b) Dss = Dps
is independent of the collection of distance random variables.
Fig. 1. Two snapshots of a network with one PU and K = 3 SUs. The two
Note that PU’s receiver is assumed to be close enough to ends of PU are assumed to be sufficiently close so that the effect of path loss
PU’s transmitter such that the path loss effect is neglected, from primary transmitter to primary receiver is negligible.
i.e., fading is the only important factor in modelling the direct
link of PU3 . From PU’s point of view, the random variables
dk,1 are independent Unif(0, Dps ) random variables for any and
k > 1 and from SUk−1 ’s point of view, d1,k ∼ Unif(0, Dps ) (K)

K+1

and dl,k are independent Unif(0, Dss ) random variables for gk\1  |hl,k |2 , k ≥ 2. (4)
any k, l > 1. Fig. 1 depicts two snapshots of a network for l=2,l=k

Dps > Dss and Dps = Dss . The code-book of user k consists
Denoting the transmitted signal of user k in a typical
of 2T Rk  random Gaussian codewords. User k transmits a
transmission slot by xk , the signal received at the receiver
codeword in T consecutive transmission slots, referred to as a
of user k is given by
block. Hence, the transmission rate of user k becomes Rk as
T tends to infinity. Each SU is capable of synchronizing itself
K+1
with PU at the symbol and block level. As such, the K + 1 y k = hk,k xk + hl,k xl + z k . (5)
transmitter-receiver pairs represent a slotted and synchronous l=1,l=k
Gaussian IC. Throughout the paper, hk is a vector that contains
the channel coefficients related to user k, i.e., The random variable z k ∼ CN(0, 1) is the ambient noise at
  the receiver side of user k. We show the average transmission
hk  h1,k · · · hK+1,k t . (2) powers for PU and each SU by Qp and Qs , respectively. The
signal xk must satisfy the power constraint
We also define
   

K+1 E |x1 |2 = Qp ≤ Pp , E |xk |2 = Qs ≤ Ps , 2 ≤ k ≤ K+1.
(K)
gk  |hl,k |2 , k ≥ 1 (3) (6)
l=1,l=k PU is ignorant in the sense that
2 Note that |f k,l | is a Rayleigh random variable.
• It is not aware of the code-book of each SU.
3 This assumption is adopted merely to simplify the calculations related to • The primary transmitter is unaware of the presence of
probability of outage in section V and can be easily relaxed. SUs, i.e., K is unknown to PU.
MOSHKSAR and KHANDANI: RANDOMIZED MASKING IN COGNITIVE RADIO NETWORKS 2639

• The primary transmitter is unaware of hk,1 for 1 ≤ k ≤ Any rate R1 in the range
K+1. However, the primary receiver is aware of the value
of K and the channel coefficients hk,1 for 1 ≤ k ≤ K +1. R1 < I(x1 ; y 1 ) (9)
PU transmits with full power Qp = Pp , i.e., x1 is a CN(0, Pp )
is achievable for PU in the conventional sense [29]. Let us
random variable that represents a symbol in a codeword of PU.
denote the noise plus interference at the primary receiver by
This is an optimal scheme for PU in the absence of SUs.
w 1 , i.e.,
Moreover, each SU is smart in the sense that
• It is aware of PU’s code-book. However, it does not know

K+1
w1  hk,1 ak sk + z 1 . (10)
the codeword transmitted by the primary transmitter in k=2
the current block. Also, SUs are anonymous to each other,
i.e., they are unaware of each others’ code-books. Since w1 is a circularly symmetric complex mixed Gaussian
• It is aware of the number K of active SUs in the network. random variable, the quantity I(x1 ; y 1 ) has no closed ex-
• For any 2 ≤ k ≤ K + 1, SUk−1 is only aware of hl,k pression. However, we are able to develop a lower bound on
for all 1 ≤ l ≤ K + 1. I(x1 ; y 1 ). The following Lemma is essential to this purpose:
In appendix C, we describe how each receiver in the
network can estimate the number of SUs and the channel Lemma 1 Let u and v be independent random variables
gains. where u is a CN(0, σ 2 ) random variable and v is a zero-
Next, we describe the RM scheme that enables SUs to mean circularly symmetric complex mixed Gaussian random

L
coexist and control the amount of interference imposed on variable such
that p v (v) = p
l=1 l ϕ(v; σl2 ) and σ1 < · · · <
PU. 2
σL . Then C 
L 2
σ
2p1 2(1−p1 ) L −p
 is a lower
min l=1 pl σl ,σ1 σL Πl=1 pl l
III. R ANDOMIZED M ASKING AT THE S ECONDARY bound on I(u; u + v).
T RANSMITTERS
A. Randomized Masking Proof: See appendix A.

2
Motivated by randomized resource allocation schemes in- In the context of Lemma 1, let pw1 (w) = L l=1 pl ϕ(w; σl )
troduced in [26], SUs follow a Randomized Masking (RM) where σ1 < · · · < σL . Note that the coefficients hk,1 in
scheme with average transmission power Qs ≤ Ps where the expression of w1 are realizations of independent and
each SU transmits a Gaussian symbol of power Qθs in its continuous random variables. Hence, the number of Gaussian
codeword with a probability of θ ∈ (0, 1) and remains silent components in the PDF of w1 is L = 2K almost surely.
with a probability of 1 − θ independently from transmission These different Gaussian components correspond to different
slot to transmission slot. As such, only a fraction θ of T realizations of the Bernoulli random variables ak for 2 ≤ k ≤
Gaussian symbols in a randomly generated codeword are K +1. Therefore, the probabilities p1 , · · · , p2K are exactly the
 − θ)
numbers θi (1 for 0 ≤ i ≤ K where θi (1 − θ)K−i is
K−i
transmitted in a block of T consecutive transmission slots.
The parameter θ is called the activity factor of SUs. Hence, repeated i times among the numbers p1 , · · · , p2K for each
K

for any 2 ≤ k ≤ K + 1, i. This yields


xk = ak sk (7) 2

K
K

where ak and sk are Ber(θ) and CN(0, Qθs ) random variables, log Π2l=1 p−p
l
l
=− pl log pl
l=1
respectively. We assume the receiver of SUk−1 is aware of ak ,
K

i.e., the masking pattern applied by its associated transmitter is K
= − θi (1 − θ)K−i log(θi (1 − θ)K−i )
already revealed to the receiver of SUk−1 . Each SU is unaware i
i=0
of the masking pattern of other SUs. In appendix C, it is shown K
that if a secondary receiver is not aware of the masking pattern K i
= − log θ i θ (1 − θ)K−i
of its affiliated transmitter, it still can decode its message, i=0
i
however, this lack of knowledge is at the cost of a loss of at
K
most 1 bits/sec/hz in its achievable rate. K i
− log(1 − θ) (K − i) θ (1 − θ)K−i
i=0
i
B. An Achievable Rate Region (a)
= −Kθ log θ − K(1 − θ) log(1 − θ)
Since PU is unaware of the code-books of SUs, multiuser
(b)
decoding is not performed at the primary receiver, i.e., the = Khb (θ), (11)
primary receiver treats the signals transmitted by the secondary K  i

K
transmitters as additive noise. However, each secondary trans- where (a) follows by the fact that i=0 i i θ (1 − θ)K−i =
mitter has the choice to perform multiuser decoding (decoding Kθ and (b) is by definition of the binary entropy function hb (·)
its own message and the message of PU), while treating other presented in Notations in section I. The smallest and largest
SUs as additive noise. variances of the Gaussian components in the PDF of w1 are
(K)
The received signal at PU can be written as σ12 = 1 and σL2
= 1 + 1θ g1 Qs which are obtained for a2 =
· · · = aK+1 = 0 and a2 = · · · = aK+1 = 1, respectively.

K+1
y 1 = h1,1 x1 + hk,1 ak sk + z 1 . (8) Moreover, the corresponding probabilities of these Gaussian
k=2 components are p1 = (1 − θ)K and pL = θK . Note that the
2640 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 7, JULY 2013


2 2 R1
term L l=1 pl σl appearing in Lemma 1 is exactly E[|w 1 | ].
We can write

K+1
E[|w 1 |2 ] = E[|z 1 |2 ] + |hk,1 |2 E[|xk |2 ]
k=2

K+1
= 1+ |hk,1 |2 E[a2k ]E[|sk |2 ] κ−ξ
k=2

K+1
Qs
= 1+ |hk,1 |2 θ ×
θ
k=2
(K)
= 1+ g1 Qs . (12) κ−ω

Using these observations in Lemma 1, we get (13). By


definition of (h1 ; θ, Qs ) in (13), it is easy to see that
ξ ω Rk
R1 ≤ (h1 ; θ, Qs ) (14)
(a)
is an inner bound to the region in (9). For 2 ≤ k ≤ K + 1,
the received signal at the receiver of SUk−1 is given by Fig. 2. Sketch of the region Rk (hk ; θ, Qs ) described in (25).


K+1
y k = hk,k ak sk + h1,k x1 + hl,k al sl + z k . (15)
Deriving a lower bound on I(sk ; y k |ak ), the first term on
l=2,l=k
the right side of (21), is very similar to that developed in
The knowledge about the code-book of PU enables the sec- (13). For completeness, the details are brought in appendix B
ondary receiver to perform interference decoding. Using the where we derive (22). Similarly, one can find a lower bound
achievable rate region for a memoryless MAC [29], SUk−1 is on I(x1 ; y k |sk , ak ), the second term on the right side of (21)
able to jointly decode its message and the message of PU if as shown in (23). By (21), (22) and (23),

⎨ R1 < I(x1 ; y k |ak , sk ) I(sk , x1 ; y k |ak ) ≥ κ(hk ; θ, Qs )  ξ(hk ; θ, Qs )+ζ(hk ; θ, Qs ).
Rk < I(sk ; y k |ak , x1 ) . (16)
⎩ (24)
R1 + Rk < I(sk , x1 ; y k |ak ) By (20), (22), (23) and (24), we obtain an achievable region
The first constraint in (16) is irrelevant. In fact, it is a constraint for the rate tuple (R1 , Rk ) as
that guarantees there is no error in decoding PU’s message, 
while the message sent by the transmitter of SUk−1 is already Rk ≤ ω(hk ; θ, Qs )
or Rk ≤ ξ(hk ; θ, Qs ). (25)
decoded successfully. Dropping the first constraint in (16), we R1 + Rk ≤ κ(hk ; θ, Qs )
get 
Rk < I(sk ; y k |ak , x1 ) We denote this region by Rk (hk ; θ, Qs ). Note that
. (17) Rk (hk ; θ, Qs ) is included in the achievable region given in
R1 + Rk < I(sk , x1 ; y k |ak )
(19). Moreover, the inequalities in (13), (20), (22) and (23)
Another approach for the receiver of SUk−1 is to treat the
turn into equalities for θ = 1. This is due to the fact that the
whole interference as noise. In this case, any rate in the range
noise plus interference at each receiver is Gaussian for θ = 1.
Rk < I(sk ; y k |ak ) (18)
is achievable by SU. Considering the two criteria in (17) and
IV. S YSTEM D ESIGN
(18), the achievable region for the rate tuple (R1 , Rk ) can be
written as In this section, we present two different approaches to
 system design. First, we assume all SUs adopt RM with
Rk < I(sk ; y k |ak , x1 )
or Rk < I(sk ; y k |ak ). full average transmission power Ps , simply referred to as
R1 + Rk < I(sk , x1 ; y k |ak )
(19) RM. Next, we consider the CTPC scenario where all SUs
Due to the fact that the noise plus interference at the receiver continuously transmit (no masking is performed), however,
of SUk−1 is mixed Gaussian, there is no closed formula for the the average transmission power Qs can be set at any value
terms I(sk ; y k |ak , x1 ), I(sk , x1 ; y k |ak ) and I(sk ; y k |ak ). less than or equal to Ps .
However, one can apply Lemma 1 to obtain lower bounds on
these quantities. As for I(sk ; y k |ak , x1 ), we have the thread A. Randomized masking under full power transmission
in (20) where (a) is by the fact that x1 is independent of the
The primary transmitter has no knowledge about channel
tuple (sk , y k − h1,k x1 , ak ), (b) is due to independence of sk
coefficients and the number of active users in the network. An
and y k −h1,k x1 under the assumption ak = 0 and (c) follows
appropriate measure to assess the performance of this user is
by Lemma 1 and the same lines of reasoning that led us to
the so-called outage probability. We define the outage event
(13). As for I(sk , x1 ; y k |ak ), one can write
for PU by
I(sk , x1 ; y k |ak ) = I(sk ; y k |ak ) + I(x1 ; y k |sk , ak ). (21) O(R1 )  {I(x1 ; y1 ) < R1 } , (26)
MOSHKSAR and KHANDANI: RANDOMIZED MASKING IN COGNITIVE RADIO NETWORKS 2641

⎛ ⎞
⎜ |h1,1 | Pp 2 ⎟
I(x1 ; y 1 ) ≥ (h1 ; θ, Qs )  C ⎜
⎝   1−(1−θ)K  ⎟
⎠. (13)
(K) (K)
min 1 + g1 Qs , 2Khb (θ) 1 + 1θ g1 Qs

I(sk ; y k |ak , x1 ) = I(sk ; y k − h1,k x1 |x1 , ak )


(a)
= I(sk ; y k − h1,k x1 |ak )
= I(sk ; y k − h1,k x1 |ak = 0)P(ak = 0) + I(sk ; y k − h1,k x1 |ak = 1)P(ak = 1)
(b)
= θI(sk ; y k − h1,k x1 |ak = 1)

K+1 

= θI sk ; hk,k sk + hl,k al sl + z k ak = 1
l=2,l=k
⎛ ⎞
(c) ⎜ |hk,k | Qs 2 ⎟
≥ ω(hk ; θ, Qs )  θC ⎜
⎝   1−(1−θ)K−1  ⎟
⎠. (20)
(K) (K)
θ min 1 + gk\1 Qs , 2(K−1)hb (θ) 1 + 1θ gk\1 Qs

I(sk ; y k |ak ) ≥ ξ(hk ; θ, Qs )


⎛ ⎞
⎜ ⎟
⎜ |hk,k |2 Qs ⎟
 θC ⎜
⎜  1−(1−θ)K−1 ⎟
 ⎟.
⎝ (K)
1 (K)
θ gk\1 Qs ⎠
θ min 1 + |h1,k |2 Pp + gk\1 Qs , 2(K−1)hb (θ) (1 + |h1,k |2 Pp ) 1+ 1+|h1,k |2 Pp

(22)


K+1
I(x1 ; y k |sk , ak ) = I x1 ; h1,k x1 + hl,k al sl + z k
l=2,l=k
⎛ ⎞
⎜ |h1,k |2 Pp ⎟
≥ ζ(hk ; θ, Qs )  C ⎜
⎝   1−(1−θ)K−1  ⎟
⎠. (23)
(K) (K−1)h (θ) 1 (K)
min 1 + gk\1 Qs , 2 b 1 + θ gk\1 Qs

where R1 is the actual transmission rate of PU. Note that where (a) is due to the fact that |f 1,1 |2 is an exponential
I(x1 ; y 1 ) is a random variable with realization I(x1 ; y 1 ) as random variable with parameter 1. We can develop an upper
it depends on the channel coefficients 
h1 . It is required that the bound on P (O(νcε )) as shown in (28) where (a) holds
probability of outage for PU be less than a certain threshold as the event {I(x1 ; y 1 ) < νcε } is a subset of the event
ε, i.e., P (O(R1 )) < ε. It is assumed that R1 = νcε where { (h 1 ; θ, Ps ) < νcε } by (13), (b) is due to the tower property
ν ∈ (0, 1) and cε is the ε-outage capacity of PU in the absence of conditional expectations [30] and (c) is due to the fact that
of SUs. This implies that the resulting drop in the transmission |f 1,1 |2 is an exponential random variable with parameter 1. If
rate for PU is only a fraction 1 − ν of its transmission rate as the right side of (28) is less than ε, then P (O(νcε )) is also
if there were no SUs in the network. We have guaranteed to be less than ε. Let us define the (ε, ν)-admissible
region by
A(RM)
ε,ν  {θ : the right side of (28) < ε}. (29)
2
cε = sup{r : P(log(1 + |f 1,1 | Pp ) < r) < ε} (RM)
  One can express Aε,ν as
2r − 1  
= sup r : P |f 1,1 |2 < <ε
Pp A(RM)
ε,ν = 0, θ∗ , (30)
 
(a) 2r −1 ∗
− where θ is the solution for θ in the nonlinear equation given
= sup r : 1 − e Pp < ε
in (31). For 2 ≤ k ≤ K + 1, the transmitter of SUk−1 has
= log(1 − Pp ln(1 − ε)), (27) complete knowledge of K and the realization of  hk . Hence,
2642 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 7, JULY 2013

(a)  
P (O(νcε )) ≤ P (h 1 ; θ, Ps ) < νcε
   
(b)
= E P (h  1 ; θ, Ps ) < νcε g (K)
1
!  1−(1−θ)K   "#
2 1 (K) Khb (θ) 1 (K)  (K)
= E P |f 1,1 | ≤ (2 − 1) min 1 + g 1 Ps , 2
νcε
1 + g 1 Ps g 1
Pp θ 
!  1−(1−θ)K "#
(c) 1 (K) 1 (K)
= 1 − E exp − (2νcε − 1) min 1 + g 1 Ps , 2Khb (θ) 1 + g 1 Ps . (28)
Pp θ

!  1−(1−θ)K "#
1 νcε (K) Khb (θ) 1 (K)
1 − E exp − (2 − 1) min 1 + g 1 Ps , 2 1 + g 1 Ps = ε. (31)
Pp θ

SUk−1 can adjust its transmission rate as the largest Rk such Denoting the actual transmission power of SUs by Q  s , we
that (R1 , Rk ) = (νcε , Rk ) ∈ R(h  k ; θ, Ps ). We denote this have  
value of Rk by R(h  k ; θ, Ps ), i.e.,  s = arg
Q sup E R(h k ; 1, Qs ) , (39)
(CTPC)
  Qs ∈Aε,ν
R(h k ; θ, Ps )  sup Rk : (νcε , Rk ) ∈ Rk (h  k ; θ, Ps ) .
where R(h  k ; 1, Qs ) is given in (33). Recall from section III
(32)
According to (25), R(h  k ; θ, Ps ) can be written explicitly as that for θ = 1 all the lower  bounds on the achievable rates
are tight. Therefore, E R(h  s ) is the exact value of
 k ; 1, Q
in (33). It is proposed that all SUs fix their activity factor at
θ where  
average achievable rate per SU that can be attained under  the
 
CTPC scheme. This is in contrast to E R(hk ; θ, Ps ) which
θ = arg sup E R(h  k ; θ, Ps ) . (34)
(RM)
θ∈Aε,ν is only a lower bound to the exact value of average achievable
rate per SU under the RM strategy.
After θ is obtained, SUk−1 uses its knowledge of K and hk One may consider the possibility of controlling the trans-
to adjust its transmission rate at mission power Qs as well as θ at each SU. If the full transmis-
sion power of each SU is extremely large, the admissible range
 Ps ).
Rk = R(hk ; θ, (35) [0, θ∗ ] will be very small. This is due to the fact that under a
large average transmission power by SU, the only way to keep
the probability of outage for PU under the desired threshold
B. Power control at secondary transmitters
is to strictly limit the range of θ. This costs each SU with
An alternative to RM is that each SU transmits the Gaussian a negligible achievable transmission rate. Fig. 3(a) presents
symbols in its codeword continuously, however, the actual probability of outage for PU under the RM scheme in terms
transmission power of SUs is designed such that the average of θ for different values of Ps in a network with K = 3, α = 1,
achievable rate per SU is maximized under the constraint that ε = 0.05, ν = 0.8, Dps = 100, Dss = 1 and Pp = 20 dB. It is
probability of outage for PU is kept under a certain threshold seen that for Ps = 30 dB, θ∗ is as small as 0.08. As for CTPC,
ε. We refer to this scheme as Continuous Transmission with the solution Q∗s for Qs in (38) might be much smaller than Ps ,
Power Control (CTPC). The design approach is along the lines and hence, each SU only transmits at a negligible transmission
as it was presented for RM with the difference that the quan- power. For instance, Fig. 3(b) offers the plot of probability of
tities (h  k ; θ, Ps ), κ(h
 1 ; θ, Ps ), ω(h  k ; θ, Ps ) and ξ(h
 k ; θ, Ps ) outage for PU under the CTPC scheme in terms of Qs in the
  
are replaced by (h1 ; 1, Qs ), ω(hk ; 1, Qs ), κ(hk ; 1, Qs ) and same setup as Fig. 3(a). It is seen that Q∗s = Ps if Ps ≤ 14 dB
ξ(h k ; 1, Qs ), respectively. The new (ε, ν)-admissible region is and Q∗s = 14 dB for any Ps > 14 dB. Designing both Qs and
given by θ simultaneously can highly improve the performance per SU.
    We look into this alternative in example 2 at the end of section
A(CTPC)
ε,ν  Qs ≤ Ps : P (h  1 ; 1, Qs ) < νcε < ε . (36) V.
 
The quantity P (h  1 ; 1, Qs ) < νcε is given by (28) if we V. D ESIGN E XAMPLES AND S IMULATION R ESULTS
substitute Ps and θ by Qs and 1, respectively. In fact, In this section we present two examples in order to address
  the design methods presented in section IV. In the first
A(CTPC)
ε,ν = 0, min{Ps , Q∗s } , (37)
example, we compare the two signalling schemes RM and
where Q∗s is the solution for Qs in CTPC. The measure of performance is the average achievable
$ % rate per SU under the required constraint on the probably
1 νcε (K) of outage for PU. We emphasize that the recorded data for
1 − E exp − (2 − 1)(1 + g 1 Qs ) = ε. (38)
Pp RM is only a lower bound to the exact performance of this
MOSHKSAR and KHANDANI: RANDOMIZED MASKING IN COGNITIVE RADIO NETWORKS 2643




ω(h  k ; θ, Ps ) νcε ≤ κ(h  k ; θ, Ps ) − ω(h
 k ; θ, Ps )
 k ; θ, Ps ) =
R(h  k ; θ, Ps ) − νcε
κ(h  k ; θ, Ps ) ≤ κ(h
ξ(h  k ; θ, Ps ) − νcε ≤ ω(h  k ; θ, Ps ) . (33)

⎩  k ; θ, Ps )  
ξ(h νcε ≥ κ(hk ; θ, Ps ) − ξ(hk ; θ, Ps )

0.2 0.3

Average Achievable Rate per SU (bits/sec/hz)


Ps=20 dB RM
CTPC
0.18 Ps=30 dB
0.28
Ps=40 dB
0.16 Ps=50 dB
0.26
ε=0.05
Probability of Outage

0.14
0.24
0.12

0.22
0.1

0.08 0.2

0.06
0.18

0.04
0.16
0 0.2 0.4 0.6 0.8 1 0 10 20 30 40 50
θ Ps (dB)

(a) (a) Dps = 1


0.09 0.75

Average Achievable Rate per SU (bits/sec/hz)


RM
Probability of Outage
0.7 CTPC
ε=0.05
0.08
0.65
Probability of Outage

0.07 0.6

0.55
0.06
0.5

0.05 0.45

0.4
0.04
0.35

0.03
−10 −5 0 5 10 15 20 0 10 20 30 40 50
Qs (dB) Ps (dB)

(b) (b) Dps = 100


Fig. 3. Plots of probability of outage in terms of θ (Fig. (a)) and in terms of Fig. 4. Plots of average achievable rate per SU in terms of Ps for different
Qs (Fig. (b)) under RM and CTPC schemes, respectively. The parameters in values of Dps . It is assumed that K = 3, α = 1, ε = 0.05, ν = 0.8,
the network are given by K = 3, ε = 0.05, ν = 0.8, Dps = 100, Dss = 1 Dss = 1 and Pp = 20dB.
and Pp = 20 dB.

at the end of section IV, since Q∗s does not depend on


scheme, while the recorded data for CTPC represent the exact Ps , the performance of CTPC saturates by increasing Ps .
value of performance. In the second example, we consider Also, according to Fig. 3, θ∗ considerably decreases as
the possibility of designing both Qs and θ in a scheme that Ps increases. This translates to a drop in the performance
is referred to as Randomized Masking with Power Control of RM as observed in Fig. 4.
(RMPC). • Effect of primary-secondary distance Dps together with
Example 1- Let us consider a scenario with K = 3 SUs in the transmission power Pp :
the network. We fix α = 1, ν = 0.8 and Dss = 1 throughout Let ε = 0.05 and Ps = 30 dB. Fig. 5 offers plots of
this example. average achievable rate per SU in terms of Pp in the
• Effect of primary-secondary distance Dps together with range 0 dB to 50 dB for different values of Dps = 1, 100.
the transmission power Ps : As Pp increases, both θs∗ and Q∗s increase. On one hand,
Let ε = 0.05 and Pp = 20 dB. Fig. 4 presents plots this implies an increase in performance per SU under
of average achievable rate per SU in terms of Ps in the both RM and CTPC strategies. On the other hand, one
range 0 dB to 50 dB for different values of Dps = 1, 100. expects that the increase in PU’s power Pp would lead
If Dps = 1, RM outperforms CTPC for all value of to a drop in performance per SU, but, this is only true
Ps . However, as Dps increases, there are values of Ps if each SU treats PU’s signal as noise. More precisely,
such that CTPC outperforms RM. One may justify the according to Fig. 2, as long as νcε < ζ = κ − ξ, SU can
observations in Fig. 4 as follows. As explained earlier jointly decode its own and PU’s message and achieve
2644 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 7, JULY 2013

0.8 0.17

Average Achievable Rate per SU (bits/sec/hz)


RM

Probability of Treating PU as Noise by SU


CTPC 0.16
0.7
0.15
0.6
0.14

0.5 RM
0.13
CTPC

0.4 0.12

0.11
0.3
0.1
0.2
0.09

0.1 0.08
0 10 20 30 40 50 0 10 20 30 40 50
Pp (dB) Pp (dB)

(a) Dps = 1 (a) Dps = 1


0.9 1.002
Average Achievable Rate per SU (bits/sec/hz)

RM RM

Probability of Treating PU as Noise by SU


0.85 CTPC 1 CTPC

0.8 0.998

0.75 0.996

0.7 0.994

0.65 0.992

0.6 0.99

0.55 0.988

0.5 0.986

0.45 0.984
0 10 20 30 40 50 0 10 20 30 40 50
Pp (dB) Pp (dB)

(b) Dps = 100 (b) Dps = 100

Fig. 5. Plots of average achievable rate per SU in terms of Pp for different Fig. 6. Plots of p(RM) and p(CTPC) in terms of Pp for different values of
values of Dps . It is assumed that K = 3, α = 1, ε = 0.05, ν = 0.8, Dps . It is assumed that K = 3, α = 1, ε = 0.05, ν = 0.8, Dss = 1 and
Dss = 1 and Ps = 30dB. Ps = 30dB.

higher data rates. However, if νcε > ζ, PU’s signal is where Ψ(θ, Qs ) is the expression on the right side of (28)
treated as noise by each SU and this drops the rate per with Ps replaced by Qs as shown in (43). Let us set the
SU. To gain more insight on this point, let us define system parameters at K = 3, α = 1, ε = 0.05, ν = 0.8,
 
p(RM)  P νcε > ζ(h  Ps )
 k ; θ, (40) Dss = 1, Dps = 10 and Pp = 20 dB. Fig. 7(a) demonstrates
(RMPC)
Aε,ν for Ps = 30 dM. Fig. 7(b) presents plots of average
and   achievable rate per SU in terms of Ps for RMPC, RM and
p(CTPC)  P νcε > ζ(h s) .
 k ; 1, Q (41) CTPC strategies. It is seen that if Ps > 15 dB, RMPC
outperforms both RM and CTPC.
Note that p(RM) and p(CTPC) are the probabilities that
any SU treats PU’s signal as noise under the RM and
VI. C ONCLUSION
CTPC schemes, respectively. Fig. 6 demonstrates plots of
p(RM) and p(CTPC) corresponding to each plot in Fig. 5. We addressed a decentralized network of one PU and
For Dps = 100, the probabilities p(RM) and p(CTPC) are several SUs. It was proposed that each secondary transmitter
close to 1 implying that for most realizations of channel adopts the RM strategy with full average transmission power
coefficients, SU treats PU’s signal as noise. Therefore, where it remained silent or transmitted a symbol in its code-
increasing Pp results in a drop in average achievable rate word independently from transmission slot to transmission
per SU as observed Fig. 5. slot. Invoking the concept of ε-outage capacity, we defined
Example 2- In this example, we consider the possibility the (ε, ν)-admissible region as the set of masking probabilities
of optimizing the average rate per SU over both Qs and θ. for each SU such that the probability of outage for PU was
As mentioned earlier, this is referred to as the RMPC scheme. maintained under a threshold ε in a case where PU set its
The (ε, ν)-admissible region in this case is the set of all tuples transmission rate at a fraction ν of its ε-outage capacity as if
(θ, Qs ) such that probability of outage for PU is less than ε. there were no SUs in the network. The masking probability of
(RMPC) SUs was designed through maximizing the average achievable
Denoting this region by Aε,ν , we have
rate per SU over the (ε, ν)-admissible region. The primary
A(RMPC)
ε,ν = {(θ, Qs ) : Ψ(θ, Qs ) ≤ ε} , (42) receiver treated interference as noise, however, each secondary
MOSHKSAR and KHANDANI: RANDOMIZED MASKING IN COGNITIVE RADIO NETWORKS 2645

!
 1−(1−θ)K "#
1 νcε (K) nhb (θ) 1 (K)
Ψ(θ, Qs )  1 − E exp − (2 − 1) min 1 + g 1 Qs , 2 1 + g 1 Qs . (43)
Pp θ

1
Admissible Region A PPENDIX A
0.9
By entropy power inequality4 [29],
0.8  
0.7 I(u; u + v) ≥ C 2h(u)−h(v) . (44)
0.6
As u ∼ CN(0, σ 2 ), then h(u) = log(πeσ 2 ). By maximum
0.5
entropy Lemma [29] and noting that E[v] = 0,
θ

0.4

L
0.3
h(v) ≤ log(πeE[|v|2 ]) = log(πe) + log pl σl2 . (45)
0.2 l=1
0.1 By the proof of Lemma 1 in [26],
0
0 5 10 15
Q (dB)
20 25 30
L
s h(v) ≤ pl log(πeσl2 ) − pl log pl
(a) l=1 l=1
(a)
p1 log(πeσ12 ) + (p2 + · · · + pL ) log(πeσL
2
0.36
≤ )
Average Achievable Rate per SU (bits/sec/hz)

0.34 L
− pl log pl
0.32 l=1

L
0.3
RMPC = p1 log(πeσ12 ) + (1 − p1 ) log(πeσL
2
)− pl log pl
RM l=1
0.28
CTPC

L

0.26 = log(πe) + p1 log σ12 + (1 − p1 ) log σL


2
− pl log pl .
l=1
0.24 (46)
0.22 where (a) is due to σ1 < σ2 < · · · < σL . By (45) and (46),
0 10 20 30 40 50
P (dB)
s Therefore, we arrive at (47) which together with (44) yield
(b)
the desired bound on I(u; u + v).

Fig. 7. The system parameters are set at K = 3, α = 1, ε = 0.05, ν = 0.8, A PPENDIX B


(RMPC)
Dss = 1, Dps = 10 and Pp = 20 dB. The left plot demonstrates Aε,ν
for Ps = 30 dM. The right plot compares average achievable rate per SU in Note that
terms of Ps for RMPC, RM and CTPC strategies.
I(sk ; y k |ak )
= (1 − θ)I(sk ; y k |ak = 0) + θI(sk ; y k |ak = 1)
receiver had the option to decode and cancel the interference (a)
caused by PU, while treating the signals of other SUs as noise. = θI(sk ; y k |ak = 1)
In another approach, referred to as CTPC, each SU transmitted = θI(sk ; y k |ak = 1)
continuously, however, it adjusted its transmission power in
K+1
order to yield the largest value for average achievable rate per = θI sk ; hk,k sk + h1,k x1 + hl,k al sl + z k ,
SU. The schemes RM and CTPC were compared for different l=2,l=k
values of transmission power for each SU and PU and distance (48)
between different users. It was observed that neither of RM
where (a) is by the fact that under ak = 0, sk and y k are
or CTPC always outperformed the other in various scenarios
independent. Since the channel coefficients are realizations of
in terms of the underlying system parameters. A combination
independent and continuous random variables, the PDF of the

K+1
of RM and CTPC referred to as RMPC was also investigated
mixed Gaussian random variable h1,k x1 + l=2,l=k hl,k al sl
where each SU adjusted both its probability of masking and
average transmission power. It was demonstrated that RMPC has almost surely 2K−1 Gaussian components with vari-
can outperform both RM and CTPC. ances σ12 < · · · < σ22K−1 and corresponding probabilities
p1 , · · · , p2K−1 , respectively. Similar to (11),
K−1
Π2l=1 p−p
l
l
= 2(K−1)hb (θ) . (49)
4 Note that the signals are complex, i.e., they can be considered as real
vectors in R2 .
2646 IEEE TRANSACTIONS ON COMMUNICATIONS, VOL. 61, NO. 7, JULY 2013

 

L
L
2
h(u) − h(v) ≥ log σ − min log pl σl2 , p1 log σ12 + (1 − 2
p1 ) log σL − pl log pl . (47)
l=1 l=1

Moreover, for any arbitrary integer N . However, SUK is not aware of K.


By setting K = K  , SUK assumes there are already K  − 1
p1 = (1 − θ)K−1
active SUs in the network and the masking parameter is set
  − 1). Using the multinomial identity, SUK calculates
on θ(K
σ12 = 1 + |h1,k |2 Pp (50) E[|y K+1 [0]|2n ] as shown in (54) where6
(K) E[xn1 1 ] = (n1 − 1)!!Ppn1
σ22K−1 = 1 + |h1,k |2 Pp + 1θ gk\1 Qs (55)

and and
⎡ 2 ⎤
K−1   E[anl l xnl l ] = E[al ]E[xnl l ]
2
⎢
K+1
 ⎥ ! "nl
pl σl2 = E ⎣h1,k x1 + hl,k al sl + z k  ⎦ Ps
  = θK  −1 × (nl − 1)!!
l=1 l=2,l=k
θK  −1
(K) (nl − 1)!!Psnl
= 1 + |h1,k |2 Pp + gk\1 Qs . (51) = . (56)
θ nl −1
K −1
Using Lemma 1, we are led to (22).
By (53) and (54) and by setting N = K  − 1, SUK obtains
K  − 1 equations for K  − 1 variables h2,K  +1 , · · · , hK  ,K  +1 .
A PPENDIX C
After finding these parameters, several other equations in (53),
A. Estimating the number K of active SUs say for n = K  , K  + 1, · · · , are checked to see if there exists
In practice SUs do not become active at the same time. a consistency. In this case, the candidate K  for K is correct.
Assume there are already K −1 active SUs in the network. We In practice, the new SU starts by setting K  = 0 and increases
refer to these users as SU1 , · · · , SUK−1 . A new SU shown by K  until a consistency is found in the underlying system of
SUK plans to enter the network. In the following, we describe equations in (53). The first value of K  that guarantees such
how SUK estimates K (phase 1) and how the K − 1 already consistency is considered to be the true value of K.
active SUs learn that there is a new active SU that plans to 2) Phase 2: After SUK finds K, it starts transmitting a
join the network (phase 2). Throughout this section, we show sequence of spikes (a training sequence of pulses) by which
θ by θK  in a network with K  active SUs5 . This quantity can the already active SUs learn that a new SU is about to become
be calculated by SUK for any value of K  . This is due to the active. Finally, all K SUs update their activity factors to θK .
fact that θK  does not depend on the realizations of channel
coefficients. Moreover, for simplicity of presentation, we as- B. Knowledge of the masking sequence at the receiver side of
sume all codeword symbols and the background noise samples each SU
are real Gaussian random variables and channel coefficients Assume both ends of each secondary user have access
are realizations of real random variables. Extending the result to two synchronized pseudorandom noise generators such as
to the complex case is straightforward. linear feedback shift registers. Then each 0 is interpreted as
1) Phase 1: At this phase, SUK does not transmit and only “mask” and each 1 is interpreted as “no mask”. However,
examines the signal received at it receiver side. Let us show this assumption can be relaxed, i.e., one can assume each
this signal in the tth transmission slot by y K+1 [t]. We have secondary receiver is unaware of the masking pattern of
its corresponding transmitter and is only aware the mask-

K
y K+1 [t] = h1,K+1 x1 [t] + hk,K+1 al [t]xl [t] + z K+1 [t], ing parameter θ. This lack of knowledge at the receiver
l=2
side costs each secondary user with a loss of at most 1
(52) bits/sec/hz in its achievable rate. To see this, note that the
for t = 0, 1, · · · where x1 [t] ∼ N(0, Pp ), xl [t] ∼ mutual information between the input and output of SUk−1
N(0,  Ps ) for 2 ≤ l ≤ K and z K+1 [t] ∼ N(0, 1). is given by I(xk ; y k |ak ) and I(xk ; y k ) in the presence and
θK−1
By the strong law of large numbers [30], as T  grows in the absence of knowledge about ak at the receiver of

 SUk−1 , respectively. Clearly, I(xk ; y k |ak ) − I(xk ; y k ) ≥ 0.
to infinity, T1 Tt=0−1 |y K+1 [t]|2n converges almost surely
to E[|y K+1 [0]|2n ] for any positive integer n. For suffi- Moreover, by independence of xk and ak ,
ciently  large T  , let us denote the recorded value for I(xk ; y k |ak ) = I(xk , ak ; y k ). (57)
1

T −1 2n
T t=0 |y K+1 [t]| by τn . As such, SUK can obtain a set
of equations One can expand I(xk , ak ; y k ) as
I(xk , ak ; y k ) = I(xk ; y k ) + I(ak ; y k |xk ). (58)
E[|y K+1 [0]|2n ] = τn , n = 1, · · · , N, (53)
6 For any odd number n, the double factorial notation is defined by n!! 
5 See 
(34) for the definition of θ. 1 × 3 × 5 × · · · n.
MOSHKSAR and KHANDANI: RANDOMIZED MASKING IN COGNITIVE RADIO NETWORKS 2647

(2n)! 
E[|y K+1 [0]|2n ]
n 
= hn1  hn2  · · · hKK ,K  +1
n1 ! · · · nK  +1 ! 1,K +1 2,K +1
n1 , · · · , nK  +1 nonnegative even numbers
n1 +···+nK  +1 =2n

n n
E[xn1 1 ]E[an2 2 xn2 2 ] · · · E[aKK  xKK  ] . (54)

However, I(ak ; y k |xk ) is never larger than the entropy of ak [18] M. Gastpar, “On capacity under receive and spatial spectrum-sharing
which is hb (θ), i.e., constraints,” IEEE Trans. Inf. Theory, vol. 53, no. 2, pp. 471–487, Feb.
2007.
I(ak ; y k |xk ) ≤ hb (θ) ≤ 1. (59) [19] A. Ghasemi and E. S. Sousa, “Fundamental limits of spectrum sharing
in fading environments,” IEEE Trans. Wireless Commun., vol. 6, no. 2,
By (57), (58) and (59), pp. 649–658, Feb. 2007.
[20] A. Jafarian and S. Vishwanath, “On the capacity of multiuser cognitive
0 ≤ I(xk ; y k |ak ) − I(xk ; y k ) ≤ 1. (60) radio networks,” 2009 IEEE International Symp. on Inf. Theory.
[21] U. Erez, S. Shamai (Shitz), and R. Zamir, “Capacity and lattice strategies
for cancelling known interference,” IEEE Trans. Inf. Theory, vol. 51, no.
ACKNOWLEDGMENT 11, pp. 3820–3833, Nov. 2005.
[22] A. Lapidoth, N. Levy, S. Shamai (Shitz), and M. A. Wigger, “A cognitive
The authors would like to thank the associate editor and the network with clustered decoding,” 2009 IEEE International Symp. on Inf.
anonymous reviewers for their valuable comments that led to Theory.
improvements in the final presentation of the paper. [23] X. Kang, R. Zhang, Y. C. Liang, and H. K. Garg, “Optimal power
allocation strategies for fading cognitive radio channels with primary user
outage constraint,” IEEE J. Sel. Areas Commun., vol. 29, no. 2, pp. 374–
R EFERENCES 383, Feb. 2011.
[1] R. Etkin, A. P. Parekh, and D. Tse, “Spectrum sharing for unlicensed [24] R. Zhang and J. M. Ciofi, “Iterative spectrum sharing with opportunistic
bands,” IEEE J. Sel. Areas Commun., vol. 25, no. 3, pp. 517–528, Apr. multiuser detection,” IEEE Trans. Commun., vol. 60, no. 6, pp. 1680–
2007. 1691, June 2012.
[2] Z. Ji and K. J. R. Liu, “Dynamic spectrum sharing: a game-theoretical [25] W. Yu, G. Ginis, and J. Cioffi, “Distributed multiuser power control for
overview,” IEEE Commun. Mag., pp. 88–94, May 2007. digital subscriber lines,” IEEE J. Sel. Areas Commun., vol. 20, no. 5, pp.
[3] R. T. Maheswaran and T. Baser, “Decentralized network resource allo- 1105–1115, June 2002.
cation as a repeated noncooperative market game,” in Proc. 2001 IEEE [26] K. Moshksar, A. Bayesteh, and A. K. Khandani, “A model for random-
Conference on Decision and Control. ized recourse allocation in decentralized wireless networks,” IEEE Trans.
[4] L. Grokop and D. N. C. Tse, “Spectrum sharing between wireless Inf. Theory, vol. 57, no. 4, pp. 2115–2142, Apr. 2011.
networks,” 2008 IEEE Infocom Conference. [27] H. Inaltekin, M. Chiang, H. V. Poor, and S. B. Wicker, “On unbounded
[5] S. Hayashi and Z. Q. Luo, “Dynamic spectrum management: when is path-loss models: effects of singularity on wireless network performance,”
FDMA sum-rate optimal?” in Proc. 2007 IEEE International Conf. on IEEE J. Sel. Areas Commun., vol. 27, no. 7, pp. 1078–1092, Sept. 2009.
Acoustics, Speech and Signal Processing, vol. 3, pp. 609–612. [28] K. Moshksar and A. K. Khandani, “Decentralized cognition via random-
[6] A. P. Hulbert, “Spectrum sharing through beacons,” in Proc. 2005 IEEE ized masking,” in Proc. 2011 Asilomar Conference on Signals, Systems
International Symposium on Personal, Indoor and Mobile Radio Comm., and Computers, pp. 1742–1746.
vol. 2, pp. 989–993. [29] T. M. Cover and J. A. Thomas, Elements of Information Theory. John
[7] J. Huang, R. A. Berry, and M. L. Honig, “Spectrum sharing with Wiley and Sons, Inc., 1991.
distributed interference compensation,” in Proc. 2005 IEEE International [30] R. M. Dudley, Real Analysis and Probability. Cambridge University
Symposium on New Frontiers in Dynamic Spectrum Access Networks, pp. Press, 2002.
88–93.
[8] L. Cao and H. Zheng, “Distributed spectrum allocation via local bargain- Kamyar Moshksar was born in Shiraz, Iran. He
ing,” in Proc. 2005 IEEE Conf. on Sensor and Ad-Hoc Networking, pp. received the B.Sc. and M.Sc. degrees, both in
475–486. electrical engineering, from Shiraz University and
[9] J. Mitola, “Cognitive radio: an integrated agent architecture for software Sharif University of Technology, respectively, and
defined radio,” Ph.D. dissertation, KTH, Stockholm, Sweden, Dec. 2000. the Ph.D. degree in 2011 from the Department of
[10] N. Devroye, P. Mitran, and V. Tarokh, “Achievable rates in cognitive Electrical and Computer Engineering, University of
radio channels,” IEEE Trans. Inf. Theory, vol. 52, no. 5, pp. 1813–1827, Waterloo, Waterloo, ON, Canada. He is currently a
May 2006. postdoctoral researcher and lecturer at the University
[11] S. Gel’fand and M. Pinsker, “Coding for channels with random param- of Waterloo. His interests lie in the broad area of
eters,” Problems Contr. and Inf. Theory, vol. 9, no. 1, pp. 19–31, Jan. applied probability.
1980.
[12] M. Costa, “Writing on a dirty paper,” IEEE Trans. Inf. Theory, vol. 29,
no. 3, pp. 439–441, May 1983. Amir K. Khandani received the M.A.Sc. degree
[13] T. S. Han and K. Kobayashi, “A new achievable rate region for the from the Uni- versity of Tehran, Tehran, Iran, in
interference channel,” IEEE Trans. Inf. Theory, vol. 27, no. 1, pp. 49–60, 1985, and the Ph.D. degree from McGill Uni- ver-
Jan. 1981. sity, Montreal, QC, Canada, in 1992. He was a
[14] A. Jovicic and P. Viswanath, “Cognitive radio: an information theoretic Research Associate with INRS Telecommunications,
perspective,” IEEE Trans. Inf. Theory, vol. 55, no. 9, pp. 3945–3958, Apr. Quebec Uni- versity, Montreal, for one year. Since
2006. 1993, he has been with the Department of Electrical
[15] W. Wu, S. Vishwanath, and A. Arapostathis, “Capacity of a class of and Computer Engineering, University of Waterloo,
cognitive radio channels: interference channels with degraded message Waterloo, ON, Canada, where he is currently a
sets,” IEEE Trans. Inf. Theory, vol. 53, no. 11, pp. 4391–4399, Nov. Professor. He currently holds an NSERC In- dus-
2007. trial Research Chair (funded by NSERC and Nortel
[16] I. Maric, R. D. Yates, and G. Kramer, “Capacity of interference channels Networks) on Advanced Telecommunications Technologies and a Canada
with partial transmitter cooperation,” IEEE Trans. Inf. Theory, vol. 53, Research Chair (Tier I) on Wireless Systems. Dr. Khandani is serving as
no. 10, pp. 3536–3548, Oct. 2007. an Associate Editor for the IEEE T RANSACTIONS ON C OMMUNICATION in
[17] J. Jiang and Y. Xin, “On the achievable rate regions for interference the area of coding and communication theory.
channels with degraded message sets,” IEEE Trans. Inf. Theory, vol. 54,
no. 10, pp. 4707–4712, Oct. 2008.

You might also like