1

SNR walls for signal detection

Rahul Tandra Anant Sahai

[email protected] [email protected]

Dept. of Electrical Engineering and Computer Sciences, U C Berkeley

Abstract

This paper considers the detection of the presence/absence of signals in uncertain low SNR environments.
Small modeling uncertainties are unavoidable in any practical system and so robustness to them is a fundamentally
important performance metric. The impact of these modeling uncertainties can be quantified by the position of the
“SNR wall” below which any particular detector will fail to be robust, no matter how long it can observe the channel.
We propose simple mathematical models for the uncertainty in the noise distribution and the fading process and
show what aspects of the model lead to SNR walls for differing levels of knowledge of the signal to be detected.
These results have implications for wireless spectrum regulators. The context is opportunistically sharing spec-
trum with primary users that must be detected in order to avoid causing harmful interference on a channel. Ideally,
a secondary system would be able to detect primaries robustly without having to know much about their signaling
strategies. We argue that the tension between primary and secondary users is captured by the technical question of
computing the optimal tradeoff between capacity and robustness as quantified by the SNR wall. This is an open
problem, but we compute this tradeoff for some simple detectors.

I. I NTRODUCTION

In order to recycle underutilized spectrum, the operation of unlicensed secondary devices within licensed primary
bands such as “unused” television broadcast bands has been proposed [1]. The main constraint for opportunistic
secondary devices/cognitive radios is guaranteeing non-interference to the primary system. One possibility is to
dynamically align the secondary transmissions such that the resulting interference is mostly orthogonal to the
primary’s received signal [2] or by having the secondary user counteract its own interference in the direction of
the primary signals [3]. Such design strategies face several practical problems. The two systems must be jointly
engineered and may require a huge overhead in terms of the amount of coordination between the primary and
secondary systems. Furthermore, even simple phase uncertainties can significantly lower the performance of such
collaborative designs [4].
Another possible strategy is for the secondary user to sense for primary signals and opportunistically use a primary
channel only if sensing declares the channel to be vacant. This approach has the advantage of minimal coordination
 2

with primary systems and hence is more flexible, but this flexibility comes at the cost of lower performance when
many primaries are present. The opportunistic approach is also fair in the sense that the onus is on the secondary
system to sense for the primary if it wants to use the spectrum. However, in order to guarantee non-interference
with potentially hidden primary receivers, the secondary system needs to be able to detect the presence/absence of
very weak primary signals [5], [6]. The IEEE 802.22 is the first international standards process for a cognitive-radio
based PHY/MAC/air interface for use in spectrum that is allocated to Television as the primary user. Under the
currently understood requirements, secondary devices are required to sense TV transmissions as low as −116 dBm
(SN R = −22 dB) [7], [8].
In spectrum sensing, the goal is to meet a given ‘receiver operating characteristic’ (ROC) constraint at very
low SNR. Classical detection theory suggests that degradation in the ROC due to reduced SNR can be countered
by increasing the sensing time [9], [10]. Hence, the sensitivity is considered to be limited by higher layer design
considerations. For instance, the QoS requirements of the application drive the protocol layer design, which in turn
dictates the time available for sensing by the physical layer. This traditional perspective implies that a cognitive
radio system can always be engineered at the cost of low enough QoS.
In a real-world physical system, parameters are never known to infinite precision. To list just a few, real-world
background “noise” is not perfectly Gaussian, it is not perfectly white, nor is it perfectly stationary. The channel
fading is neither flat nor is it constant over time. Real-world filters are not ideal, A/D converters have finite precision
and dynamic range, I and Q signal pathways in a receiver are never perfectly matched and local oscillators are
never perfect sine-waves. This paper argues that as a result of these uncertainties, the degradation in QoS may be
complete once the target primary SNR is low enough. These model uncertainties impose fundamental limitations
on detection performance. The limitations cannot be countered by increasing the sensing duration. At very low
SNRs, the ergodic view of the world is no longer valid, i.e., one cannot count on infinite averaging to combat the
relevant uncertainties.
To illustrate the limitations caused by model uncertainties, consider a simple thought experiment. Imagine a
cognitive radio running a detection algorithm to sense the presence/absence of a primary signal. Assume that the
detector computes a test-statistic and compares it to a threshold to decide if the primary is present/absent. The
threshold is set so the target false alarm probability ( PF A ) is met using the nominal model. Now, we take this
detector and deploy it in physically distinct scenarios with their own slight variations in local noise and fading
characteristics. An example using the energy detector is illustrated in Figure 1. The actual performance achieved
when the detector is deployed in the field can deviate significantly from the prediction. In fact, below a certain
SNR threshold, at least one of the error probabilities can become worse than 12 . We call this sort of failure a lack
of robustness in the detector. The nominal SNR threshold below which this phenomenon manifests is called the
 3

SNR wall for the detector.

ROC curves for energy detection (N=2000, SNR = −10 dB, x=.25)
0.5

0.45

Probability of mis−detection PMD

0.4

0.35

0.3

0.25 Target
Performance
0.2

0.15

0.1

0.05

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Probability of false alarm PFA

Fig. 1. Error probabilities of energy detection under noise uncertainty. The point marked by an ‘X’ is the target theoretical performance
under a nominal SNR of −10. The shaded area illustrates the range of performance achieved with actual SN R’s varying between −9.9 dB
and −10.1 dB (x = .1 in Sec III-A2).

Such robustness limits were first shown in the context of radiometric (energy) detection of spread spectrum
signals [11]. To make progress for general classes of signals and detection algorithms, it is important to distill the
relevant uncertainties into tractable mathematical models that enable us to focus on the uncertainties that are most
critical. The basic sensing problem is posed in Section II with the explicit models for practical wireless system
uncertainties in noise and fading being introduced gradually throughout this paper. Section III then explores the
robustness of two extreme cases — non-coherent detectors and pilot-detection. The former corresponds to very
limited knowledge about the primary signaling scheme whereas the latter corresponds to complete knowledge of
a part of the primary’s signal. Both of these detectors suffer from SNR walls. In particular, the best possible
non-coherent detector is essentially as non-robust as the energy detector and this non-robustness arises from the
distributional uncertainty in the background noise. In the coherent case, the SNR wall is pushed back, but only to
an extent limited by the finite coherence time of the fading process. It is important to note that these SNR walls
are not modeling artifacts. They have been experimentally verified in [12] using controlled experiments that were
carefully calibrated to limit uncertainties.
Once it is clear that uncertainties limit detector performance, it is natural to consider detection strategies that
attempt to learn the characteristics of the environment at run-time and thereby reduce the uncertainty. We call
this approach noise calibration and it is explored in Section IV. First, pilot detection is revisited. The key idea is
that the pilot tone does not fill all degrees of freedom at the receiver and so the other degrees of freedom can be
used to learn the noise model. Such strategies can improve performance only so far, with the limit coming from
a combination of the non-whiteness of the background noise process and the finite coherence-time of the fading
 4

process. Frequency is not the only way to think about degrees of freedom. So next, noise calibration is explored in
time domain by considering primary signals whose pulses have only a 50% duty cycle. In this example, the limits
to noise calibration are due to the uncertain frequency selectivity of the fading and the concept of delay coherence
time is introduced to capture this.
The robustness results in this paper have significant implications for policymakers. Suppose that a band of
spectrum is opened for cognitive use. The rules governing the band must be flexible enough to allow interoperability
with future technologies while also being compatible with high performance. One can think of a simple and flexible
rule: the primary can transmit any signal within a given spectral mask, similarly the secondary should be able to
robustly sense the primary with a high probability of detection. Under these rules, the primary could potentially
use all its degrees of freedom by transmitting a ‘white’ signal. As shown in Section III-A, the secondary would
be forced to non-coherently detect for the primary, and such detection algorithms are highly non-robust. Complete
flexibility for the primary user comes at a cost of inefficient spectral usage overall. The secondary systems face
SNR walls and hence are forced to presumptively shut-up since they cannot robustly detect the absence of potential
primary signals. But, if the rules are written so that the primary is mandated to transmit a known pilot tone at a
certain power, then the secondary systems can operate more easily at the cost of potentially lower performance
for the primary user. The general tradeoff can be posed technically as the capacity-robustness tradeoff. We briefly
discuss this tradeoff in Section V for some of the detectors considered here.

II. P ROBLEM FORMULATION

Let X(t) denote the band-limited signal we are trying to sense, let H denote the fading process, and let the
additive noise process be W (t). Although this paper focuses on real-valued signals, the analysis easily extends to
complex signals. The discrete-time version is obtained by sampling the received signal at the appropriate rate. The
sensing problem can be formulated as a binary hypothesis testing problem with the following hypotheses:

H0 : Y [n] = W [n]

H1 : Y [n] = HX[n] + W [n] (1)

Here X[n] are the samples of the signal of interest, H is a linear time-varying operator, W [n] are samples of
noise and Y [n] are the received signal samples. Throughout this paper, the signal is assumed to be independent
of both the noise and the fading process. Random processes are also assumed to be stationary and ergodic unless
otherwise specified.
All three aspects of the system (W, H, X) admit statistical models, but it is unrealistic to assume complete
knowledge of their parameters to infinite precision. To understand the issue of robustness to uncertainty, we assume
 5

knowledge of their distributions within some bounds and are interested in the worst case performance of detection
algorithms over the uncertain distributions. 1
Background noise is an aggregation of various sources like thermal noise, leakage of signals from other bands due
to receiver non-linearity, aliasing from imperfect front end filters, interference due to transmissions from licensed
users far away, interference from other opportunistic systems in the vicinity, etc. A stationary white Gaussian
assumption is only an approximation2 and so the noise samples W [n] are modeled as having any distribution W
from a set of possible distributions Wx . This set is called the noise uncertainty set and specific models for it
are considered in subsequent sections. The goal is to introduce only as much uncertainty as needed to show that
detection becomes non-robust.
Fading is modeled with the same philosophy. The fading process H ∈ Hf is considered a linear time varying
filter and models for it will be developed as needed. The general idea is that filter coefficients vary on the time-scale
given by the channel coherence time, but it is unreasonable to assume that the true channel coherence time is known
exactly. All we can hope for are bounds on it.
Finally, uncertainty arises due to intentional under-modeling of system parameters. For instance, the primary
signal X can be modeled by imposing a cap on its power spectral density, instead of actually modeling its specific
signal constellation, waveform, etc. The advantage of under-modeling is two-fold. Firstly, it leads to less complex
systems. Secondly, intentional under-modeling keeps the model flexible. For example, a simple power spectral
density cap on a chunk of spectrum is advantageous because it gives the primary user flexibility to choose from a
diverse set of signaling strategies. This is useful for ease of system evolution.
Any detection strategy/algorithm can be written as a possibly random function F : RN → {0, 1}, where F maps
the N dimensional received vector Y = (Y [1], Y [2], · · · , Y [N ]) onto the set {0, 1}. Here ‘0’ stands for the decision
that the received signal is noise and ‘1’ stands for the decision that the received signal is signal plus noise. For
each choice of a noise distribution W ∈ Wx and fading model H ∈ Hf , the error probabilities are

   
PF A (W, H) = EW,H 1{F =1} |H0 ; PM D (W, H) = EW,H 1{F =0} |H1 .

• A decision strategy robustly achieves a given target probability of false alarm, PF A , and probability of missed
detection, PM D if the algorithm satisfies

sup PF A (W, H) ≤ PF A ; sup PM D (W, H) ≤ PM D .

W ∈Wx ,H∈Hf W ∈Wx ,H∈Hf

1
If a Bayesian perspective is taken and a prior is assumed over these parameters, then nothing much changes. The prior turns into a set
of possible parameter values once a desired probability of missed detection and false alarm are set. Having a prior just makes the notation
and definitions more complicated and so we do not assume priors here for system parameters.
2
In information theory, it turns out that the Gaussian distribution is a saddle point for many interesting optimization problems. Therefore,
we are sometimes safe in ignoring the distributional uncertainty and just overbounding with a worst case Gaussian. However, this turns out
not to be true in our case.
 6

• The average signal to noise ratio is defined as

N
P 1 X
SN R = 2 , where P = lim X[n]2 .
σn N →∞ N
n=1

Although the actual noise variance might vary over distributions in the set Wx , we assume that there is a
single nominal noise variance σn2 associated with the noise uncertainty set Wx .
• A detection algorithm is non-robust at a fixed SN R if the algorithm cannot robustly achieve any pair
1
(PF A , PM D ), where 0 ≤ PF A < 2 and 0 ≤ PM D < 12 , irrespective of the sampling duration N .
• Suppose there exists a signal to noise ratio threshold, SN Rt , such that the detector is non-robust for all
SN R ≤ SN Rt . The maximum value of such a threshold is defined as the SN R wall for the detector.

SN Rwall = sup{SN Rt , s.t., the detector is non-robust for all SN R ≤ SN Rt }.

An equivalent condition for testing robustness of a detector can be given in terms of the test statistic T (Y)
for the detector. The detector is non-robust iff the sets of means of T (Y) under both hypotheses overlap [13],
[14]. Mathematically, for a fixed N , define AN := {EW,H [T (Y)|H0 ] : W ∈ Wx , H ∈ Hf } and BN :=
{EW,H [T (Y)|H1 ] : W ∈ Wx , H ∈ Hf }. The detector is non-robust iff AN ∩ BN 6= ∅ for all N > 0. Furthermore,
if the set of distributions for the observations Y under both hypotheses overlap then every detection algorithm will
be non-robust. This criterion is used throughout this paper to verify if a given detection algorithm is non-robust.
The goal of this paper is to analyze detection algorithms and prove the existence of SN R walls.

III. D ETECTOR ROBUSTNESS : TWO EXTREME CASES

A. Unknown signal structure

Consider a situation where the spectrum sensor knows very little about the primary signal. Assume that the
primary signaling scheme is unknown, except with a known power within the band of interest. This corresponds to
a primary licensee that has absolute freedom to choose its signaling strategy with only the bandwidth and power
specified in the license. To robustly detect such a primary user, a single detector must be able to detect the presence
of any possible primary signal that satisfies the power and bandwidth constraint.
Under this limited information, the intuitively hardest signal to detect is a zero-mean white signal in the frequency
band of interest.
1) Radiometer robustness: It is useful to review the radiometer (energy detector) under noise level uncertainty
P
[11]. The test statistic is given by T (Y) = N1 N 2
n=1 Y [n] . If there is no uncertainty and the noise variance is

completely known, the central limit theorem (see [15]) gives the following approximations:

1 4 1
T (Y)|H0 ∼ N (σ 2 , 2σ ), T (Y)|H1 ∼ N (P + σ 2 , 2(P + σ 2 )2 ),
N N
 7

where P is the average signal power and σ 2 is the noise variance. Using these approximations gives

PF A = P rob (T (Y) > γ|H0 )

 
γ−σ  2
= Q q , (2)
2 2
N σ

where γ is the detector threshold. Similarly, the probability of detection is given by

 
2
γ − (P + σ ) 
PD = Q  q (3)
2 2
N (P + σ )

Eliminating γ from (2) and (3), we get N = 2[Q−1 (PF A ) − Q−1 (PD )(1 + SN R)]2 SN R−2 . At low SN R we can
approximate 1 + SN R ≈ 1 and this gives a sample complexity N = O(SN R−2 ). This shows that if the noise
statistics are completely known, then we can detect signals at arbitrarily low SN R’s by increasing the sensing time
N.
Now, consider the case with uncertainty in the noise model. Since the radiometer only sees energy, the distribu-
tional uncertainty of noise can be summarized in a single interval [ ρ1 σn2 , ρσn2 ] where σn2 is the nominal noise power
and ρ > 1 is a parameter that quantifies the size of the uncertainty.
To see the sample complexity required to achieve a target PF A and PD robustly, (2) and (3) are modified to get
   
γ − σ 2 γ − ρσ 2
n
PF A = max Q q  = Q q ,
σ 2 ∈[ ρ1 σn
2 ,ρσ 2 ] 2 2 2 2
n
N σ N ρσ n
   
1 2
2
γ − (P + σ )  γ − (P + σ )
ρ n 
PD = min Q q = Q q . (4)
σ 2 ∈[ ρ1 σn
2 ,ρσ 2 ] 2 2 2 1 2
N (P + σ ) N (P + ρ σn )
n

Approximating 1 + SN R ≈ 1 gives

2[Q−1 (PF A ) − Q−1 (PD )]2

N= h  i2 . (5)
1
SN R − ρ − ρ
 
From the above expression it is clear that N → ∞ as SN R ↓ ρ − ρ1 , and this is illustrated in Figure 2. The
figure gives rise to the descriptive term “SNR Wall.” Figure 3 illustrates that the SNR Wall reflects the fact that
the radiometer will not be able to robustly detect the signal iff the signal power is less than the uncertainty in the
noise power, i.e., P ≤ (ρ − ρ1 )σn2 . The presence of the signal is indistinguishable from a slightly larger value for
the noise. Hence,

energy ρ2 − 1
SN Rwall = . (6)
ρ

Figure 4 plots the SN R wall for the radiometer as a function of the noise uncertainty ρ = 10x/10 , as expressed
 8

in dB terms. Extensive simulations results showing the limits on radiometer performance under noise uncertainty
have been done in [16].

Sample complexity of energy detection under noise uncertainty

16

14 x=1 dB
x =0.001 dB

12

10 x=0.1 dB
log10 N

0
−40 −35 −30 −25 −20 −15 −10 −5 0

SNR in dB

Fig. 2. This figure shows how the sample complexity varies for energy detection as the SN R approaches the SNR wall (See Equation (5)).

Test statistic
Target
Sensitivity

Signal Finite sample

present
2
uncertainty
ρ σn

1/ρ σn
2 } Intrensic Noise power
Uncertainty
Zone
Impossible

Fig. 3. This figure helps understand noise uncertainty in the case of a radiometer. The shaded area in the figure represents the uncertainty
in the noise power.

2) Optimal non-coherent detection: When both the primary signal and the noise are modeled as white Gaussian
processes, then clearly nothing can perform better than the radiometer. However, real-world primary signals are
decidedly not Gaussian. We can attempt to exploit knowledge of the underlying signal constellation, even if the
dependence through time is unknown and hence allowed to be white.
 9

Position of SNR wall for radiometer

2

−2

−3.3 dB
−4

(in dB)
−6

wall
SNR
−8

−10

−12

−14
0 0.5 1 1.5 2 2.5 3
Noise uncertainty x (in dB)

energy
Fig. 4. This figure plots the SN Rwall in (6) as a function of noise level uncertainty x where ρ = 10x/10 . The point marked by a “x”
on the plot corresponds to x=1 dB of device level noise uncertainty.

Even in this case, it turns out that all detectors suffer from an SN R wall if the noise uncertainty set is made
sufficiently rich. To keep the analysis tractable, we propose a bounded moment noise uncertainty model that captures
the idea of approximately Gaussian. See Figure 5 for a pictorial description of the noise uncertainty set. A white
noise distribution Wa ∈ Wx if:

• The noise process is symmetric EWa2k−1 = 0, ∀k = 1, 2, . . ..

• Even moments of the noise must be close to the nominal noise moments in that EWa2k ∈ [ ρ1k EWn2k , ρk EWn2k ],
where Wn ∼ N (0, σn2 ) is a nominal Gaussian noise random variable and ρ = 10x/10 > 1.

As before, the parameter x is used to quantify the amount of uncertainty in the noise power, i.e., we allow for
x dB of uncertainty in the noise variance. The above model is reasonable for the following reasons:

• Suppose the nominal noise distribution is believed to be Wn ∼ N (0, σn2 ). Then, any reasonable uncertainty
model must include Gaussian distributions with variance slightly different than σn2 , i.e., any Gaussian with
variance σ 2 ∈ ( ρ1 σn2 , ρσn2 ), for some ρ > 1.
• Background noise also includes weak signals from interferers and other random disturbances. This means that
assuming perfect Gaussianity is unrealistic and the noise uncertainty set should include other distributions that
are close to Gaussian. For mathematical tractability, closeness to Gaussian distributions is expressed in terms
of the moments. It is clear that similar results hold if the noise uncertainty set is modeled as a small ball (in
the variational distance) around the nominal Gaussian CDF for Wn .

Theorem 1: Consider the robust hypothesis testing problem defined in (1) and the above noise uncertainty model
with ρ > 1. Assume that there is no fading and the primary signal model allows signals X[n] satisfying the following
 10

Fig. 5. An example of a noise uncertainty set Wx , for x = 3dB. The shaded region is a set of possible pdfs for the noise random
variable. The bold curve inside the region is the pdf of a nominal Gaussian, Wn . The dashed curve is the pdf of X + Wn for BPSK X at
SN R = −6dB. This illustrates that at low signal powers, it is impossible to robustly distinguish the signal from noise.

properties:

1) X[n] are independent of the noise samples W [n].

2) X[n] are independent and identically distributed as a random variable X taking values in a bounded set X .
3) X is zero-mean and all its odd moments are zero, i.e., E[X 2k−1 ] = 0 for all k = 1, 2, · · ·
supx∈X x2 ∗
Define SN Rpeak = σn2 . Under these assumptions, there exists a threshold SN Rwall such that robust
∗
detection is impossible if SN Rpeak ≤ SN Rwall .
Furthermore, there are easy to compute bounds on this SNR wall:

∗ ρ2 − 1
ρ − 1 ≤ SN Rwall ≤ . (7)
ρ

Proof: See Appendix A.

It is important to note that Theorem 1 does not demand that fading be the identity or that the primary signals be
white. It merely requires that these be possible. Clearly any reasonable model for uncertain fading will include the
identity as a possibility. While truly white primary signals seem unlikely in practice, it is important to remember
that they would be allowed by regulations saying that the primary can transmit any signal within a given spectral
mask. If the spectrum sensor must detect the presence of any legitimate primary user under such regulations, then
it must be able to detect truly white primary signals as well. The only way around the limitations induced by
Theorem 1 is to impose additional structure on the primary signal or to reduce the noise uncertainty.
 11

B. Unknown signal structure + deterministic pilot

The other extreme signal detection scenario is when the entire signal X(t) is known. Equivalently, assume that the
primary signal allocates a fraction of its total power to transmit a known and deterministic pilot tone and just focus
on that part for detection. This model covers many practical communication schemes that use pilot tones/training
sequences for data frame synchronization and timing acquisition generally. For example, a digital television (ATSC)
signal has a pilot tone that is -11 dB weaker than the average signal power [17].
√ p
For simplicity, assume the primary’s signal X[n] is given by θXp [n] + (1 − θ)Xd [n]. Here Xp [n] is a
known pilot tone with θ being the fraction of the total power allocated to the pilot tone. Continue to model
Xd [n], W [n] as zero-mean iid processes as in the previous section. The test statistic for the matched filter is
P
given by T (Y) = √1N N b b
n=1 Y [n]Xp [n], where Xp is a unit vector in the direction of the pilot tone. In the

case of completely known noise statistics, the matched filter is asymptotically optimal. It achieves a dwell time
N ≈ [Q−1 (PD ) − Q−1 (PF A )]2 θ−1 SN R−1 [18].
It is easy to see that the matched filter is robust to uncertainties in the noise distribution alone. Notice that
P
E[T (Y)|H0 ] = 0 and E[T (Y)|H1 ] = √1N N b
n=1 Xp [n]Xp [n] 6= 0. This shows that the distributional uncertainty

classes of T (Y) under both hypotheses do not overlap and hence the detector can robustly achieve any (PF A , PM D )
pair for sufficiently large N .
The SNR wall in this case is a consequence of fading. The fading process H is modeled as acting on the signal
P
by ∞ l=0 hl [n]X[n − l] where hl [n] are the time-varying multipath fading coefficients. In this case, it is clear that

we cannot reap the gains of coherent signal processing forever. As soon as the channel taps assume independent
realizations, we can no longer gain from coherent signal processing.
Suppose that a genie gives the detector a lower-bound Nc on the coherence time and the further guarantee that
the fading filter h can only change at integer multiples of Nc . Since this information is not available in practice,
making this assumption can only improve robustness. By considering time in units of coherence times, it is clear
h i2
1 PM −1 √1 PNc bp [nNc + k] , where M is the
that a good test statistic in this case is T (Y) = M n=0 N k=1 Y [nN c + k] X
c

number of coherent blocks over which we can listen for the primary.
This detector can be visualized as a combination of two detectors. First, the signal is coherently combined within
each coherence time Nc . Coherent processing gain boosts the signal power by Nc while the noise uncertainty
is unchanged. Second, the new boosted signal is detected at the receiver by passing it through a radiometer. The
radiometer aspect remains non-robust to noise uncertainties, in spite of the boost in the signal strength. The effective
SN R of the coherently combined signal is given by SN Ref f = SN R · θ · Nc .
 12

Hence, the modified matched filter will be non-robust if

 
ρ2 − 1
SN Ref f ≤
ρ
 2 
ρ −1
⇒ SN R · θ · Nc ≤
ρ
 2 
mf 1 ρ −1
⇒ SN Rwall = . (8)
Nc · θ ρ

Coherent processing gains could also be limited due to implementation complexity. The clock-instability of
both the sensing radio and the primary transmitter imposes a limit on the coherent processing time. For instance,
suppose that there is 1000 Hz of frequency uncertainty and we are doing coherent processing over 1 ms. Since the
pilot frequency is uncertain, suppose that we need to search over 4 bins to achieve a target probability of missed
detection. If we want to do coherent processing over 10 ms, we now need to search over 40 frequency bins. Since
searching over frequency bins involves computing FFTs, it is clear that the coherent processing time is limited
by the complexity of the spectrum sensor. Finally, coherent processing gains might be limited due to the lengths
of the pilots themselves. In packet-based primary systems, the pilots are embedded within each packet and are
only coherent for a packet duration since the inter-packet arrival times are usually random on the time-scale of the
carrier frequency. No further coherent processing gain is available. Simulations for evaluating the performance of
coherent detectors under noise uncertainty have been done in the IEEE 802.22 community. Extensive simulation
results evaluating performance of coherent detectors for captured DTV signal data can be found in [19], [20].

IV. N OISE CALIBRATION

So far we have shown the existence of SN R walls for detection algorithms that do not attempt to explicitly
do anything about the noise uncertainty. The key question is whether it is possible to obtain accurate calibration
of uncertain statistical quantities like noise at run-time. The focus on run-time is important because both the
fading process and the noise/interference are likely to be at least mildly nonstationary in practice. This means that
calibration will get stale (leaving substantial residual uncertainty) if it is done far in advance.
This leads to a tension. Ideally, we would like access to H0 in a parallel universe that is identical to our own
except that the primary user is certainly absent. However, our detector is necessarily confined to our own universe
where the presence or absence of the primary user is unknown. Any run-time calibration that aims to reduce the
noise uncertainty must use data that might be corrupted by the presence of the primary signal. When the primary
user is allowed to be white, then Theorem 1 can be interpreted as telling us that noise-calibration is impossible.
Intuitively, there is no degree of freedom that is guaranteed to be uncorrupted by the primary.
Therefore, a prerequisite for getting noise-calibration gains is that the primary signal must be constrained to
discriminate among the available degrees of freedom. The case of pilot-detection provides a natural first case to
 13

explore since the pilot signal is confined to a specific frequency, leaving measurements at other frequencies available
to do noise calibration. After finishing the discussion of primaries that discriminate on the basis of frequency, we
consider models for primary users that discriminate in time.

A. Noise calibration: narrowband pilots

Consider the basic model of Section III-B, except with the additional assumption that the pilot Xp [n] is a
√
narrowband pilot-tone given by Xp [n] = 2 P sin 2π ff0s n, where P is the average signal power. Rather than modeling
the fading process using a block-fading model with coherence time Nc , it is convenient to think about the fading
B
process as being arbitrary but bandlimited to total bandwidth Nc where B is the original width of the channel. The
coherent processing of the previous section can be reinterpreted as an ideal bandpass filter that selects a bandwidth
1
that is Nc as wide as the total primary band and hence gives a reduction in noise power by the same factor.
The approach of the previous section throws away all the information outside that bandpass filter. Noise calibration
is about using that information to improve detection robustness.

Band pass filters

Pilot

Signal Noise

111111111111111111111
000000000000000000000
000000000000000000000
111111111111111111111 1111111
0000000 1111111
0000000
000000000000000000000
111111111111111111111 0000000
1111111 0000000
1111111
000000000000000000000
111111111111111111111 0000000
1111111 0000000
1111111
000000000000000000000
111111111111111111111 0000000
1111111 0000000
1111111
000000000000000000000
111111111111111111111 0000000
1111111 0000000
1111111
000000000000000000000
111111111111111111111 0000000
1111111
0000000
1111111 0000000
1111111
0000000
1111111
000000000000000000000
111111111111111111111 0000000
1111111 0000000
1111111
000000000000000000000
111111111111111111111 f 0000000
1111111 0000000
1111111 f
fp fm fp fm

Fig. 6. Noise calibration in the case of ‘white’, stationary noise and ‘white’ signals with narrowband pilot tones.

Figure 6 pictorially describes the noise calibration technique in the case of signals with narrowband pilots. The
figure plots the power spectral density (PSD) of the primary signal plus noise. The shaded region is the noise
PSD and the solid region corresponds to the signal PSD. The sinusoidal pilot tone is shown as a Dirac delta in
the frequency domain. Let fp be the frequency of the pilot tone and fm 6= fp be any other frequency within the
channel. The noise-calibration detection algorithm can be formally described as follows:

• Pass the received signal through two identical bandpass filters centered at fp and fm respectively. For mathe-
matical convenience assume that these bandpass filters are ideal “brick-wall” filters and denote their bandwidth
by BBP F . Let the output of these filters be denoted by Yp [n] and Ym [n] respectively.
• Measure the empirical power of both outputs. Let them be denoted by Pb(fp , N ) and Pb(fm , N ) respectively.
P 1 PN
That is Pb(fp , N ) = N1 N 2 b
n=1 |Yp [n]| and P (fm , N ) = N
2
n=1 |Ym [n]| .
 14

• The test statistic is given by T (Y, N ) := [Pb(fp , N ) − Pb(fm , N )]. Compare this test statistic to a threshold
and declare the signal is present if it exceeds the threshold.

We call this detector the pilot power detector [19]. Pb(fp , N ) estimates the total power in a narrow band around
the pilot tone and this includes the pilot power itself, signal power (the solid region in Figure 6) and the noise
power. Since we are worried about the low SN R regime, the signal power in the band is much lower than the noise
power and can be safely neglected. Similarly, Pb(fm , N ) estimates the signal power plus noise power in a narrow
band around fm . If we assume that the noise is ‘white’, i.e., its power spectrum is flat, the noise power in the output
of both the band pass filters is the same. Hence, we have E[T (Y, N )|H0 ] = 0 and E[T (Y, N )|H1 ] = θP 6= 0. This
clearly implies that this detector is robust to distributional uncertainties in the marginals of a white noise process.

B. Colored noise uncertainty model

The robustness of the noise-calibration detector described in the previous section critically depends on the absolute
‘whiteness’ of noise. It is unreasonable to assume that noise is completely white. Since “noise” includes thermal
noise and interference from other transmitters sharing the same band of spectrum, leakage from adjacent bands,
etc, it has some a priori unknown color to its spectrum. Real world noise also contains seemingly arbitrarily placed
spurs and other terms that can change quite rapidly across the channel [21].
Figure 7(a) illustrates the set of possible power spectral densities under the earlier noise uncertainty model of
Section III. The noise power spectral density is flat, but its variance can vary within the bounds as shown in the
figure. Figure 7(b) plots the set of possible power spectral densities under the colored noise uncertainty model. The
two solid (red) curves correspond to two different nominal levels around which the noise could have some ripples.
cx . A noise distribution Wa ∈ W
The new model of the possibly colored noise uncertainty set is denoted W cx if

the marginals satisfy the conditions from Section III-A2, but are not necessarily white. We do assume that the noise
1
process is wide sense stationary, and its power spectral density SWa (f ) satisfies λ Na ≤ SWa (f ) ≤ λNa for some
λ > 0. Here Na is some constant satisfying ρ1 N0 ≤ Na ≤ ρN0 , where N0 B = σn2 and B is the total bandwidth of
the signals we are trying to detect. Intuitively, the noise power spectral density is approximately flat at a level Na ,
but allowed to ripple arbitrarily around the Na level.

C. Noise calibration: matched filter with colored noise

Consider the detector in Section IV-A under our new colored noise uncertainty model. Recall that the test statistic
for the detector is T (Y, N ) = Pb(fp , N ) − Pb(fm , N ), i.e., the test statistic measures the difference in powers at
the pilot frequency fp and the noise calibration frequency fm . Since the noise can be colored, it is clear that the
difference in noise powers in the two pass bands (centered at fp and fm ) need not be zero. The gains from noise
calibration are limited and the detector faces SN R wall limitations.
 15

λ Na −1/ λ Na
Actual noise PSD Two possible noise PSD’s

ρN ρN
0 0

1/ ρ N 0 1/ ρ N 0

f f
Band of interest Band of interest

(a) (b)
Fig. 7. Figure (a) describes the power spectral density of noise under the ‘white’ noise uncertainty model, and Figure (b) describes the
power spectral density of noise under the colored noise uncertainty model.

√ p
Theorem 2: Consider the detection problem in (1). Here the nominal X[n] = θXp [n] + (1 − θ)Xd [n],
1
the fading H is an arbitrary bandlimited process with total bandwidth and unit power, and W [n] are noise
Nc
√
cx . Assume Xp [n] = 2 P sin 2π f0 n and
samples whose distribution lies in the colored noise uncertainty set W fs
1 PN 2
P = limN →∞ N n=1 |Xd [n]| is the average signal power.
In this case the pilot power detector defined in Section IV-A is non-robust if
   
P ρ λ2 − 1 2BBP F
SN R := 2 ≤ · · , (9)
σn θ λ B

e
where B is the total bandwidth of the signal X[n] and BBP F is the bandwidth of the band pass filter in the
algorithm.
Proof: The test statistic in Section IV-A is T (Y, N ) = Pb(fp , N ) − Pb(fm , N ). This detector is non-robust if

min EW [T (Y, N )|H1 ] ≤ max EW [T (Y, N )|H0 ]

cx
W ∈W cx
W ∈W
(a)
⇒ θP + min EW [T (W, N )] ≤ max EW [T (W, N )]
cx
W ∈W W ∈Wc
   x 
(b) 1 1
⇒ θP + Na BBP F − λNa BBP F ≤ λNa BBP F − Na BBP F
λ λ
 2     
(c) P 1 λ −1 2BBP F Na
⇒ SN R := 2 ≤ · · ·
σn θ λ B N0
 2   
ρ λ −1 2BBP F
≤ · ·
θ λ B

• Step (a) above follows from the fact that the test statistic under hypothesis H1 is the sum of the faded pilot
h i
power and the noise power3 in the bandpass filter. This can be seen from: E |HXp [n]|2 = E[|H|2 ]|Xp |2 = θP .
Here we have used that the fact that the unknown bandlimited fading process H is restricted to have unit power.
cx .
In this step, W is an N dimensional vector of noise samples drawn from the distribution W ∈ W
3
This statement has two implicit assumptions. Firstly we are assuming that the band pass filter is wide enough that it captures all the pilot
power after it is spread in frequency by the Doppler spread of the channel. Secondly, we are neglecting the power of the data part of the
signal (Xd [n]) in the passband. This assumption is reasonable because we are interested in the regime where P ≪ σn2 .
 16

• Step (b) is obtained by using the fact that the minimum difference in noise powers in the two bands is
1


λ N0 BBP F − λN0 BBP F . Similarly, the maximum possible difference in noise powers in the two bands is


λN0 BBP F − λ1 N0 BBP F .
Na
• (c) follows from the fact that N0 B = σn2 . The final factor of ρ is just present to account for the fact that N0

could be as big as ρ. This is an artifact of the fact that we are expressing the wall in terms of the nominal
SN R and the uncertainty is modeled multiplicatively.

BBP F


From Theorem 2, it is clear that the pilot power detector’s SN R wall goes to zero as B → 0. So, can
the bandwidth BBP F of the bandpass filter be reduced arbitrarily? Recall that in the proof of Theorem 2, we
assumed that the pilot power under the bandpass filter is θP . This assumption is true only if the bandpass filter
has a bandwidth greater than the bandwidth of the fading process. This is because the pilot tone is a Dirac delta
in frequency domain and is smeared out due to the fading process. The pilot power is spread over the bandwidth
1
of the fading process in some unknown way. Therefore, the result of the theorem is true only if BBP F ≥ Nc . In
fact, tightening the bandpass filter any further could result in the loss of all the pilot power and missed detection.
 2   
Hence the best possible SN R wall for the pilot power detector is given by ρθ · λ λ−1 · N2c .
Notice that in this model of coloring uncertainty, there is no real benefit from calibrating the noise using nearby
frequencies. All frequencies within the channel are considered equally good. This makes physical sense if the
dominant coloring uncertainty is coming from potential local spurs and nulls in the noise. Otherwise, it is easy to
alter the uncertainty model to include some sense of continuity for the noise PSD. In that case, the λ uncertainty
1
parameter would also be proportional to Nc since it is the bandwidth of the fading process that limits how close in
frequency the calibration filter can come. That would result in the SNR Wall improving as O( N12 ) with increasing
c

coherence time Nc .
Finally, it is easy to see that the above noise calibration strategy does not really depend on the width of the
bandpass filter used for calibration. Even if it used all the information outside of a tight window containing the
pilot signal, a detector cannot completely calibrate the noise process within that tight window due to the unknown
color. Within that tight window, all detectors essentially face the same problem as Theorem 1 — an arbitrary (and
hence possibly white) signal in the presence of noise with a residual unknown level coming from the color ripple.

D. Noise calibration: 50% duty cycle pulse amplitude modulated signals

It is immediately obvious that Theorem 2 also generalize to cases where the primary signal restricts itself
to use a β fraction of the bandwidth. This gives a processing gain corresponding to the bandwidth restriction
1
BBP F = (β + Nc )B and sets θ = 1, but otherwise introduces no qualitatively new concerns. In that case, the
 17

coherence time plays no significant role in the SNR wall. The noise calibration gains add to the bandwidth reduction
gains in a straightforward way.
We now explore another approach to using only some of the underlying degrees of freedom. To be specific,
consider an example signal that uses only half of the available degrees of freedom in time rather than frequency.
 √
 ± 2P w. p. 1 if n is odd,
2
X[n] = (10)
 0 if n is even.

The average power of this signal is P and this example is a caricature of practical signals whose amplitude is
modulated using time domain pulses. Such pulse amplitude modulated signals come under the general category
of cyclostationary signals [22]–[24]. Feature detectors are a general class of detectors that have been proposed to
distinguish cyclostationary signals from noise. These detectors were first proposed by William Gardner [25], [26].
Robustness of feature detectors is an active topic of research in this area. Fundamental results on robustness of
single cycle feature detectors are shown in [27], [28], but here we illustrate the basic ideas in the context of this
simple caricature.
As in (1), the received signal under hypothesis H1 can be written as Y [n] = HX[n]+W [n], where W [n]’s are the
noise samples whose distribution lies in the uncertainty set Wx . In this example, the average power measurements
at ‘even’ time samples can be used to calibrate the noise in the ‘odd’ time samples.
1) No fading: Suppose that there was no fading. The natural test statistic is given by
" N N
#
1 X 1 X
T (Y) = |Y [2n − 1]|2 − |Y [2n]|2 . (11)
N N
n=1 n=1
PN PN
In this case it is easy to verify that E[T (Y)|H0 ] = E[ N1 n=1 |W [2n − 1]|2 ] − E[ N1 n=1 |W [2n]|
2] = 0 and
E[T (Y)|H1 ] = 2P 6= 0. Therefore, it is clear that we can obtain infinite gains from noise calibration and hence
there is no SN R wall if the noise is guaranteed to be white. The somewhat surprising fact here is that the same is
true even if the the noise is colored. The mean of the test statistic is still the difference of two expectations. The
noise contribution to each expectation is identical by stationarity, not whiteness.
2) Flat fading: Assume that the channel fading is flat (i.e H is just multiplication by a bandlimited scalar signal).
For pilot-detection, this is enough to cause an SNR wall since it spreads the pilot tone over a finite bandwidth and
thereby stops infinite coherent processing gains. In this case however, the scalar multiplication can do nothing to
the zero samples of the signal! Even time-samples continue to provide a clean view of the noise. Therefore, by the
arguments above the noise calibration becomes perfect. Noise color is not enough to stop this since the stationarity
P 1 PN
and presumed ergodicity of the noise guarantees that the calibration term N1 N 2
n=1 |Y [2n]| = N n=1 |W [2n]|
2

P
converges with probability one to N1 N 2
n=1 |W [2n − 1]| as N gets large.
 18

3) Frequency-selective fading: The previous result is somewhat surprising since the matched filter for pilot
tones is not-robust to this kind of flat-fading uncertainty, even with noise calibration when the noise is colored.
However, frequency-selective fading does introduce an SNR wall for this detector. To illustrate the idea, consider
the simplest possible frequency-selective fading: a delay. Let the received signal under hypothesis H1 be Y [n] =
X[n − l[n]] + W [n].
Here, the traditional notion of a bandlimited fading process is inadequate to capture the size of the relevant
uncertainty. After all, the arguments of the previous section suggest that HX[n] = h[n]X[n − l[n]] would behave
the same as just X[n − l[n]] while the apparent bandwidth of the fading process might be dominated by a fast flat
component given by h[n]. Instead, consider a block-fading model in which the delay l[n] is piecewise constant for
Dc time steps before taking on an independent realization. Call Dc the lower bound on the delay coherence time.
To understand the nature of the delay coherence time Dc , it is useful to consider two cases:

• Assume that the delay spread of the channel is significantly larger than 1 sample and all of the channel taps
arise from the sum of many paths. (There is no dominant line-of-sight path) As these path-lengths change over
time, all of the taps will get different values. In such a case, the delay coherence time Dc and the traditional
amplitude coherence time Nc (or equivalently, the reciprocal of the bandwidth of the fading process) are the
same since the relative phase response can change significantly in this time. This is the model that is likely to
hold for wideband primary users being detected at significant distances from their transmitters. This basically
corresponds to when ISI is significant.
• Now suppose there is only one dominant line-of-sight path that is arriving at the spectrum sensor, but is subject
to local reflections in the near neighborhood of the sensor. (e.g. A user in an airplane) The delay spread of the
channel can be smaller than 1 sample. In this case, there will just be a single tap in the filter. The amplitude
might change rapidly since the local paths can go in and out of phase with each other, but the overall delay
is not going to be changing as fast. In such cases, the delay coherence time Dc > Nc and the dominant effect
may very well be the clock skew of the sensor’s local oscillator relative to the primary oscillator. In such cases,
the factor difference between Dc and Nc is comparable to the factor difference between the signal bandwidth
and the carrier frequency.

For this simplest example of a fading process with finite Dc , the strategy of (11) is doomed to failure if the sums
are taken beyond one coherence time. The signal will corrupt both the even and odd time samples eventually due
to the random delay. By analogy with the case of matched filtering, a natural detection strategy is to compute the
square of the test statistic within each coherence time and average that over multiple coherence times. The modified
 19

test statistic is given by

 2 
 
1
M
X −1
 2 DXc /2
 
Te(Y, Dc ) = 
 Dc
2 2  
|Y [kM + 2n − 1]| − |Y [kM + 2n]|  , (12)
M  
k=0 n=1

where N = M Dc .
Theorem 3: Let the received signal under hypothesis H1 in (1) be given by Y [n] = X[n − l[m]] + W [n], with
X[n] given in (10), and W [n] stationary uncertain noise independent of the signal. Assume that the distribution of
the noise arises from the noise uncertainty set Wx . Suppose the delay fading satisfies:

• l[m] is constant within a delay coherence block and iid across different delay coherence blocks.
• l[m] is uniformly distributed on {0, 1}.
P
Define, SN R = 2 .
σn The detector defined in (12) is non-robust if
q
−2
ρ + 2 ρ12 + (2 + Dc )(ρ2 − 1
ρ2 ) x
SN R ≤ , where ρ = 10 10 .
2 + Dc

Proof: From (12) it is clear that

 2 
 
 2 DXc /2
  
E[Te(Y, Dc )|Hi ] = E  |Y [2n − 1]|2 − |Y [2n]|2  Hi  , for i = 0, 1. (13)
 Dc n=1 

The following lemma gives the desired quantities.

Lemma 1:

16σ 4 16σ 4 + 16P σ 2 + 8P 2

E[Te(Y, Dc )|H0 ] = , E[Te(Y, N )|H1 ] = 4P 2 + , (14)
Dc Dc

where W [n] ∼ N (0, σ 2 ).

Proof: See Appendix C.
Using Lemma 1, it is clear that the detector in (12) is non-robust if the set of means under both hypotheses
overlap. This condition is equivalent to

16σ 4 + 16P σ 2 + 8P 2 16σ 4

min 4P 2 + ≤ max
1 2
σ ≤σ 2 ≤ρσn
ρ n
2 Dc 1 2
σ ≤σ 2 ≤ρσn
ρ n
2 Dc

2
4 ρ12 σn4 + 4P ρ1 σn2 + 2P 2 4ρ2 σn4
⇒P + ≤
Dc Dc
P
The last inequality in SN R := σn
2 can be solved to give the required result by the quadratic formula.
This theorem suggests that it is the uncertainty corresponding to the frequency selectivity of the fading that induces
limits on the detection of very weak signals that do not fully utilize all the degrees of freedom in time-domain.
There are two interesting aspects of this result:
 20

• The dependence of the SNR wall for this detector on the delay coherence time is O( √1D ). This gain from
c

increasing delay coherence time is slower than the O( N1c ) gain for the matched filter, but it is better than the
lack of any gain with coherence time for a primary signal that only uses half the degrees of freedom in the
frequency domain.
• The SNR wall result so far does not depend on the color of the noise! The proposed detector in (12), while
being intuitively very natural, is non-robust in the presence of frequency-selective fading, even when the noise
is truly white. A consequence of this is that the SNR wall depends only on ρ, not on the potentially much
smaller λ. As the next section shows, this is an artifact of the detection strategy used.

E. Noise calibration: color revisited

The intuitively natural detector from (12) seems to be doing noise calibration, but Theorem 3 shows that it gets
no improvements in the SNR Wall from reduced λ ripples in the noise color. This prompts the question whether
the detector in (12) is even close to the best possible detector for exploiting noise calibration gains?
While the question remains open in general, it turns out that detectors based on cyclostationary feature detection
behave similarly to the detector (12) and thus do not seem to be able to fully exploit the promise of noise calibration
[27], [28]. However, it is possible to give an example of a better detector that is motivated from information-theoretic
arguments coming from the method of types [29]. Consider the following detection algorithm:

• Within each delay coherence block, compute the empirical average power of the odd samples and even samples.
That is, compute (Pbk,odd (Dc ), Pbk,even (Dc )), where
Dc Dc

2 X2
2 X2

Pbk,odd (Dc ) = |Y [kDc + 2n − 1]|2 , Pbk,even (Dc ) = |Y [kDc + 2n]|2

Dc Dc
n=1 n=1

• bk,odd (Dc ) = 1
Quantize these average powers into two bins using a global quantization threshold T — i.e. set B
if Pbk,odd (Dc ) ≥ T and zero otherwise. Similarly for B
bk,even (Dc ).

• bk,odd (Dc ), B
Collect these quantized pairs (B bk,odd (Dc )) for k = 1, 2, · · · , K , for a very large number of delay

coherence blocks K .
• bk,odd (Dc ), B
Compute the empirical distribution of the unordered binary pairs {B bk,odd (Dc )}, k = 1, 2, · · · , K .

Denote this empirical probability distribution (b

p11 (K), pb10 (K), pb00 (K)).
• Choose the global threshold T so that pb11 (K) = 14 . From a practical point of view, this can be thought of as
a kind of automatic gain control (AGC) that is set so that the quantizer sees the pair (1, 1) a fourth of the
time. In a sense, this AGC is implicitly doing calibration to the noise.
• The test statistic is pb00 (K) and the detector declares hypothesis H0 if | limK→∞ pb00 (K) − 14 | ≤ δ for some
small threshold δ > 0, and declares H1 otherwise.
 21

Note that under either hypothesis Hi , i = 0, 1, the empirical probabilities (b

p11 (K), pb10 (K), pb00 (K)) converge to
their true probabilities as K → ∞. The key question is what are the possible limiting probability distributions as
the noise distribution varies over the noise uncertainty set Wx ?
1) Under hypothesis H0 and white noise: The distributions of both the odd and even samples are independent
bk,odd (Dc ) is also independent and identical to B
and identical. This implies that B bk,even (Dc ). This forces the limiting

empirical distribution to have the following form:

p11 (K), pb10 (K), pb00 (K)) = (p2 , 2p(1 − p), (1 − p)2 ),
lim (b
K→∞

where p = Prob(Pbk,odd (Dc ) ≥ T |H0 ) = Prob(Pbk,even (Dc ) ≥ T |H0 ). Furthermore, since the AGC is set such that
the probability of a (1, 1) is 41 , we must have p = 21 , and hence the limiting distribution is ( 41 , 12 , 14 ).
2) Under hypothesis H1 and white noise: In this case, the distributions of the odd and even samples are
independent but not identical. This is because one of the two has more energy than the other (since it has both
signal plus noise). Therefore, the limiting distribution is of the form

p11 (K), pb10 (K), pb00 (K)) = (qr, q(1 − r) + r(1 − q), (1 − q)(1 − r)),
lim (b
K→∞

for some 0 ≤ q ≤ 1, 0 ≤ r ≤ 1 given by

q = Prob(Pbk,odd (Dc ) ≥ T |H1 ), r = Prob(Pbk,even (Dc ) ≥ T |H1 ).

where the delay l[n] = 0 for the purposes of calculating q, r under H1 . Furthermore, the AGC is set so that qr = 14 .
1
To have the hypotheses overlap, we would have to be able to also set q(1 − r) + r(1 − q) = 2 which implies that
1
q + r = 1. But the geometric mean and arithmetic means of q, r can only agree if q = r = 2 which is not possible
if there is any signal present.
This immediately shows that the limiting empirical probability distributions under both hypotheses are disjoint
if the noise is truly white. Hence, we can always separate the two hypotheses and this detector is robust to ‘white’
noise level uncertainties regardless of how short the delay coherence time Dc ≥ 2 is. This result is consistent with
our intuition that ‘white’ noise level uncertainties are insufficient to limit noise calibration gains.
3) Under colored noise: For ease of analysis, consider a specific colored Gaussian noise uncertainty model. Let
2 ∈ [ 1 σ 2 , ρσ 2 ]. Define W [n] = M [n]−αM [n−1]
M [n] be an iid white Gaussian noise process, with the variance σm ρ n n

for some α > 0. Suppose that just the α is unknown to within some small range 0 ≤ α ≤ αmax .
√ √
b Dc Pbk,even (Dc ) Dc
The two-dimensional central limit theorem can be applied to (G1 , G2 ) := ( Pk,odd2σ
(Dc )
2 , 2σm2 ). Stan-
m

dard calculations reveal that this approaches a jointly normal random variable with variances (1 + α2 )2 and covari-
 22

√ √ √ √ √
Dc 2 ) Dc ) Dc 2 ) Dc
ance4 2α2 . The means under H0 are ((1 + α2 ) 2 , (1 + α 2 and are ((1 + α2 ) 2 , (1 + α 2 + SN R Dc )
2 .
under H1 . The variances and covariance do not change substantially in H1 when P is small relative to σm
The limiting distribution for the test statistic is thus

1
lim (b
p11 (K), pb10 (K), pb00 (K)) = ( , aα , bα ),
K→∞ 4

for some aα + bα = 34 .
1
For the noise-only hypothesis, if the noise is white, then clearly a0 = 2 and b0 = 41 . Otherwise for the noise-only
1
case, it is clear5 that bα ≈ 4 + K1 α2 for some constant K1 when α is small. When the signal is present, then
1
bα ≈ 4+ K1 α2 − K2 SN R2 Dc with K2 being another constant6 when α is small. Thus, the presence or absence
 
of the signal is robustly detectable only if SN R > K3 α√max
D
for some constant K 3 . This recovers the O √1
D
c c

scaling in the SNR wall, but also includes a clear dependence on the coloring parameter7 αmax .
We conjecture that this detector is order optimal in terms of the dependence on the coloring parameter and the
delay coherence time. However it is clear that the noise-calibration gains are intimately tied up with the features
of the signaling itself.

V. T HE CAPACITY- ROBUSTNESS TRADEOFF

It is natural to wonder ‘Who must pay the price for robust detection?’ The goal of the primary transmitter is to
maximize its data rate to the primary receiver, while the secondary user wants to be able to design a robust algorithm
to detect the presence of the primary. The relevant tradeoff is between the performance loss to the primary system
and the robustness gains to the secondary system. The primary performance can be quantified using capacity and
this paper’s results indicate that robustness can be quantified in terms of the SNR wall.
For a target SNR wall, the ultimate goal is to find a primary signaling scheme that maximizes the data rate
achievable to the primary system while still being robustly detectable above the wall. Conversely, given a target
primary data rate, the goal is to find the primary signaling scheme that maximizes the robustness of secondary
detection. If the primary were only concerned with a given bandwidth and power constraint, then information
theory reveals that the optimal signaling should look white [29] and Theorem 1 indicates that this gives very poor
robustness for the secondary. Thus, it appears that this tradeoff is fundamental from a system design point of view
and can be called the ‘capacity-robustness tradeoff’.
4
It is clear that even under a general model for colored wide-sense stationary noise, the covariance here is going to be a function of
the amount of coloring since it captures a particular way in which the noise deviates from iid. The positive correlation here obscures the
separation of the two quantities if the signal is present.
5
Space reasons force us to omit the proof. Intuitively, just think of the joint Gaussian as being the sum of a pair of two iid Gaussian
random variables (N1 , N2 ) with an independent pair distributed like (+Nα , +Nα ). The variance of Nα is proportional to α2 .
6
The constant acts negatively since having a signal present tends to make the two move in opposite directions and hence reduces the
probability that both cross the threshold.
7
A quick look at the Fourier transform of the filter [1, −αmax ] reveals the close relationship of αmax to the ripple factor λ in the PSD
of the noise process.
 23

It is a non-trivial open problem to determine the optimal tradeoff. Instead, here the tradeoff is explored in the
context of a given sensing algorithm and class of primary signals.

A. Capacity-robustness tradeoff: energy detector

Let the channel between the primary transmitter and primary receiver be a classic bandlimited AWGN channel.
The capacity of this channel is completely characterized by the channel bandwidth B and the signal to noise ratio
PP
SN RP = N0 B , where PP is the received signal power at the primary receiver and N0 is the noise power spectral
density. The capacity is given by C = B log (1 + SN RP ) assuming complex signaling. Assume that the spectrum
energy ρ2 −1
sensor uses a radiometer to detect the primary signal. As seen in Section III-A, the relevant SN Rwall = ρ .
PS
The SN R at the sensor can be written as SN RS = N0 B , where PS is the received primary power at the spectrum
sensor.
It is clear that the common parameter that affects both the primary data rate and the sensing SN R is the bandwidth
B . The primary transmitter can tradeoff its rate in exchange for robustness at the spectrum sensor by reducing the
fraction of degrees of freedom used. Let Bactual = βB , where β ∈ (0, 1). We also assume that the fading at the
B
spectrum sensor is bandlimited and has a bandwidth of Nc .
In this case, the total bandwidth potentially occupied
 2 
by the primary signal at the spectrum sensor is βB + NBc . The effective SNR wall is given by (β + N1c ) ρ ρ−1 .
Hence the capacity-robustness tradeoff for the simple radiometer is given by the parametric curve
    
1 ρ2 − 1 PP
(robustness, capacity) : (β + ) , βB log 1 + , (15)
Nc ρ βN0 B

where β ∈ [0, 1]. The factor β in the robustness is the processing gain obtained because the primary does not use
all the degrees for freedom.
One can also obtain noise calibration gains by taking measurements in the frequencies over which the primary
does not spread its signal. Using the model in Section IV-B, the uncertainty in the noise is reduced from ρ to some
smaller number λ. Therefore, the tradeoff with noise calibration is given by:
  2   
1 λ −1 PP
(robustness, capacity) : (β + )ρ , βB log 1 + . (16)
Nc λ βN0 B

B. Capacity-robustness tradeoff: coherent detector

Assume that the spectrum sensor uses a coherent detector and the primary signal has a deterministic pilot
tone as in Section III-B. Any power spent on the pilot improves robustness but reduces the power available for
communicating messages. Let θ be the fraction of total power spent in the pilot tone. The primary data rate is
   2 
bounded by RP = B log 1 + (1−θ)P
N0 B
P
. From (8), the SNR wall is given by SN R mf
wall = 1
Nc ·θ
ρ −1
ρ , where
 24

θ ∈ [0, 1] and Nc is the channel coherence time. Thus, the capacity-robustness tradeoff for the matched filter is
given by the parametric curve
    
1 ρ2 − 1 (1 − θ)PP
(robustness, capacity) : , B log 1 + , (17)
Nc · θ ρ N0 B

where θ ∈ [0, 1]. The gains in robustness given in (17) are coherent processing gains. As discussed in Section IV-C,
noise calibration can also be used to reduce the uncertainty down to the λ level. Hence, the capacity-robustness
tradeoff with noise calibration is given by
    
ρ λ2 − 1 (1 − θ)PP
(robustness, capacity) : , B log 1 + , (18)
Nc · θ λ N0 B

C. Numerical examples

Capacity robustness tradeoffs

Matched filter
6 Matched filter with noise calibration
Energy detector
Energy detection with noise calibration
Primary data rate (in Mbps)

0
−15 −10 −5 0 5 10
SNR wall (in dB)

Fig. 8. The capacity-robustness tradeoff curves are plotted for energy and coherent detectors where the primary is operating at high spectral
efficiency.

Figure 8 compares the capacity-robustness tradeoff curves for energy detection and coherent detection. There
curves were plotted using Equations (15), (16), (17) and (18). These plots correspond to a bandwidth B of 1 MHz,
the signal to noise ratio of the primary channel SN RP = 20 dB, the channel coherence time Nc = 100, the noise
uncertainty parameters ρ = 10 (to account for potentially significant interference) and λ = 100.1 . Qualitatively, one
can see that coherent detection is much better than energy detection, i.e., for a given loss in primary data rate, it
gives better robustness gains for the secondary system. Essentially, the primary system is not power limited. So,
sacrificing some power to the pilot tone is not a significant cost, but it gives significant robustness gains. However,
shrinking the bandwidth and using fewer degrees of freedom to improve radiometer robustness is painful since it
is degrees of freedom that are the critical resource for high SNR primary users.
 25

Capacity robustness tradeoffs

0.15
Matched filter
Matched filter with noise calibration
Energy detector
Energy detection with noise calibration

Primary data rate (in Mbps)

0.1

0.05

0
−15 −10 −5 0 5 10
SNR wall (in dB)

Fig. 9. The capacity-robustness tradeoff curves are plotted for energy and coherent detectors where the primary is operating at low spectral
efficiency.

Figure 9 compares the capacity-robustness tradeoffs for energy and coherent detection when the primary SN R
is low. These tradeoffs were plotted for SN RP = −10 dB. Since the primary is power limited, the loss of power
to the pilot has a non-trivial rate penalty. Meanwhile, giving up degrees of freedom incurs less of a penalty to help
the radiometer. The robustness curves for the two strategies come closer to each other, but deterministic pilot tones
and coherent detection still win out.

VI. C ONCLUSIONS

When a system attempts to detect the presence/absence of weak signals in noise, the model uncertainties cannot
be ignored since they can be comparable to the desired signals. So while traditionally the performance metric
used to compare various detection algorithms is the sample complexity, at low SNR, robustness becomes a more
significant issue. The robustness can be quantified in terms of the SNR wall giving the threshold below which
weak signals cannot be detected reliably no matter how many samples are taken. The SNR wall is quantified for
different detectors and the results show that there is a tradeoff between the amount of knowledge assumed about
the primary and the level of robustness. If the primary signal does not use all the degrees of freedom equally,
detection robustness can be improved by explicitly accounting for model uncertainty and trying to reduce it at
run-time through noise calibration. While this mitigates the SNR wall, it does not eliminate it.
 26

A PPENDIX A
P ROOF OF T HEOREM 1

For k ≥ 1, define f (k, SN R) to be the following uni-variate degree k polynomial

" k     #
X 2k 1 1 · 3 · 5 · · · (2k − 2i − 1) i
f (k, SN R) := SN R . (19)
2k − 2i ρk−i 1 · 3 · 5 · · · (2k − 1)
i=0

Note that all the coefficients of f (k, SN R) are non-negative and hence f (k, SN R) is monotonically increasing
(2k)
in SN R. This means that the equation f (k, SN R) = ρk has a unique real valued root SN Rwall . Because f (k, 0) =
1 (2k)
ρk < ρk since ρ > 1, the root SN Rwall > 0 for all k . Define SN Rwall
∗ to be

∗ (2k)
SN Rwall := inf SN Rwall . (20)
k=1,2,3,···∞

Let Wn ∼ N (0, σn2 ), W1 ∼ √1 Wn . Let W2 := X + W1 . If W1 ∈ Wx , and W2 ∈ Wx , then it is clear that robust

ρ

detection is impossible. We now prove that W2 ∈ Wx by showing that EW22k−1 = 0 and 1 2k

ρk EWn ≤ EW22k ≤
ρk EWn2k ∀ k .
Since E[X 2k−1 ] = 0 and W2 =: X + W1 , we have EW22k−1 = 0 and EW22k ≥ 1 2k
ρk EWn . Therefore, W2 ∈ Wx

⇔ EW22k ≤ ρk EWn2k
E(X + W1 )2k
⇔ ≤ ρk
EWn2k
Xk  
2k EW 2k−2i
⇔ 1
EX 2i ≤ ρk
2k − 2i EWn2k
i=0
Xk    
2k 1 1 · 3 · 5 · · · (2k − 2i − 1) σn2k−2i
⇔ EX 2i ≤ ρk
2k − 2i ρk−i 1 · 3 · 5 · · · (2k − 1) σn2k
i=0
Xk    
2k 1 1 · 3 · 5 · · · (2k − 2i − 1) EX 2i
⇔ ≤ ρk ∀ k = 1, 2, · · · (21)
2k − 2i ρk−i 1 · 3 · 5 · · · (2k − 1) σn2i
i=0

So, we just need to show that the last inequality in (21) is true.
EX 2i i ∗
By definition of ‘peak SN R’, we have σn
2i ≤ SN Rpeak . Now, if SN Rpeak ≤ SN Rwall , by the monotonicity
of f (k, SN R), we have
k 
X   
2k 1 1 · 3 · 5 · · · (2k − 2i − 1) EX 2i
2k − 2i ρk−i 1 · 3 · 5 · · · (2k − 1) σn2i
i=0
Xk    
2k 1 1 · 3 · 5 · · · (2k − 2i − 1) i
≤ SN Rpeak
2k − 2i ρk−i 1 · 3 · 5 · · · (2k − 1)
i=0
= f (k, SN Rpeak )
∗
≤ f (k, SN Rwall ) = ρk .
 27

which is the last inequality in (21).

∗ (2) ρ2 −1
In terms of the upper-bound in (7), from (20) it is clear that SN Rwall ≤ SN Rwall = ρ . However, the entire
result is vacuous if the infimum in (20) is zero. Getting a nontrivial lower bound is more interesting.
(2k)
Recall that SN Rwall is the root of the polynomial f (k, SN R) − ρk . The coefficient of SN Ri in f (k, SN R) − ρk
2k

 1   1·3·5···(2k−2i−1)   
is 2k−2i ρ k−i 1·3·5···(2k−1) . Replace ρ
1
k−i by 1 to get a new polynomial, fe(k, SN R) − ρk . Since ρ > 1,
 
we have ρk−i 1
≤ 1 for all i = 1, 2, · · · , k , and hence the unique positive root of fe(k, SN R) − ρk = 0 must be
(2k)
(2k) ^
smaller than SN Rwall . Call this root SN Rwall .
It is thus clear that

(2k)
∗
SN Rwall ≥ inf ^
SN Rwall (22)
k=1,2,··· ,∞

(2k)
^
where SN Rwall is the unique positive root of
" k    #
X 2k 1 · 3 · 5 · · · (2k − 2i − 1)
fe(k, SN R) := SN R = ρk .
i
(23)
2k − 2i 1 · 3 · 5 · · · (2k − 1)
i=0

This polynomial has a natural upper bound given in the following lemma.
Lemma 2:

fe(k, SN R) ≤ (1 + SN R)k , ∀ k ≥ 1, SN R ≥ 0

Proof: See Appendix B.

Using the above lemma in (23) gives

(2k)
^
ρk = fe(k, SN Rwall )
(2k)
^
≤ (1 + SN Rwall )k .

This implies that

(2k)
^
SN Rwall ≥ ρ − 1, ∀ k ≥ 1.

Substituting the above inequality in (22) gives the desired lower bound.
 28

A PPENDIX B
P ROOF OF LEMMA 2

Fix k ≥ 1 and SN R ≥ 0. From the definition of fe(k, SN R) in (23), we get

" k    #
X 2k 1 · 3 · 5 · · · (2k − 2i − 1)
fe(k, SN R) = SN R i
2k − 2i 1 · 3 · 5 · · · (2k − 1)
i=0
" k   #
X (2k)! 1 · 3 · 5 · · · (2k − 2i − 1)
= SN Ri
(2k − 2i)! · (2i)! 1 · 3 · 5 · · · (2k − 1)
i=0
" k #
X (2k) · (2k − 1) · · · (2k − 2i + 1)  1 · 3 · 5 · · · (2k − 2i − 1) 
i
= SN R
(2i) · (2i − 1) · · · 1 1 · 3 · 5 · · · (2k − 1)
i=0
" k   #
(a) X k (2k − 1) · (2k − 3) · · · (2k − 2i + 1) 1 · 3 · 5 · · · (2k − 2i − 1)
= · SN Ri
i (2i − 1) · (2i − 3) · · · 1 1 · 3 · 5 · · · (2k − 1)
i=0
" k   #
(b) X k 1 i
= SN R
i (2i − 1) · (2i − 3) · · · 1
i=0
" k   #
X k
≤ SN Ri
i
i=0
= (1 + SN R)k .

• (a) follows from the following calculation.

   
(2k) · (2k − 1) · · · (2k − 2i + 1) (2k) · (2k − 2) · · · (2k − 2i + 2) (2k − 1) · (2k − 3) · · · (2k − 2i + 1)
= ·
(2i) · (2i − 1) · · · 1 (2i) · (2i − 2) · · · 2 (2i − 1) · (2i − 3) · · · 1
 i   
2 · (k) · (k − 1) · · · (k − i + 1) (2k − 1) · (2k − 3) · · · (2k − 2i + 1)
= ·
2i · (i) · (i − 1) · · · 1 (2i − 1) · (2i − 3) · · · 1
  
k (2k − 1) · (2k − 3) · · · (2k − 2i + 1)
= .
i (2i − 1) · (2i − 3) · · · 1

• (b) follows from the following calculation.

 
(2k − 1) · (2k − 3) · · · (2k − 2i + 1) 1 · 3 · 5 · · · (2k − 2i − 1)
·
(2i − 1) · (2i − 3) · · · 1 1 · 3 · 5 · · · (2k − 1)
   
1 1 · 3 · 5 · · · (2k − 1) 1
= · = .
(2i − 1) · (2i − 3) · · · 1 1 · 3 · 5 · · · (2k − 1) (2i − 1) · (2i − 3) · · · 1

A PPENDIX C
P ROOF OF LEMMA 1

Under hypothesis H0 : Y [n] = W [n] ∼ N (0, σ 2 ) for all n > 0. Using this in (13), we get
 2 
 Dc /2 
 2 X  
e 
E[T (Y, Dc )|H0 ] = E  2 2  
|W [2n − 1]| − |W [2n]|  . (24)
 Dc n=1 
 29

|W [2n − 1]|2 − |W [2n]|2 are iid random variables with zero mean and variance 4σ 4 . Using the Central Limit
Theorem [15], we get
Dc /2
1 X
√ |W [2n − 1]|2 − |W [2n]|2 ∼ N (0, 4σ 4 ).
Dc n=1

Using this approximation in (24), we get E[Te(Y, Dc )|H0 ] = 16σ 4

Dc .

Under hypothesis H1 : To compute the expectation of the modified test-statistic in (13) we can assume l[m] = 0
without loss of generality. So, we have Y [2n − 1] = X[2n − 1] + W [2n − 1] ∼ N (0, P + σ 2 ) and Y [2n] =
X[2n] + W [2n] = W [2n] ∼ N (0, σ 2 ) for all n > 0. This implies that |Y [2n − 1]|2 − |Y [2n]|2 are iid random
variables with mean P and variance 4σ 4 + 4P σ 2 + 2P 2 . Again using the Central Limit Theorem [15], we get
Dc /2
1 X p
√ |Y [2n − 1]|2 − |Y [2n]|2 ∼ N ( Dc P, 4σ 4 + 4P σ 2 + 2P 2 ).
Dc n=1

Using this approximation we get

 2 
 
 2 DXc /2
  
E[Te(Y, Dc )|H1 ] = E  |Y [2n − 1]|2 − |Y [2n]|2  H1 
D
 c n=1 
 2 
 Dc /2
X 
4  1    
= E √ |Y [2n − 1]|2 − |Y [2n]|2  H1 
Dc  Dc n=1 
4
= [4σ 4 + 4P σ 2 + 2P 2 + Dc P 2 ]
Dc
16σ 4 + 16P σ 2 + 8P 2
= 4P 2 +
Dc

ACKNOWLEDGMENTS

We thank the NSF ITR program (ANI-0326503) for funding this research. The work is also indebted to ongoing
interactions with Mubaraq Mishra, Danijela Cabric, and Niels Hoven. The anonymous reviewers and issue editor
Akbar Sayeed were immensely helpful in shaping the presentation and scope of this paper.

R EFERENCES

[1] FCC, “FCC 04-113,” May 2004. [Online]. Available: http://hraunfoss.fcc.gov/edocs public/attachmatch/FCC-04-113A1.pdf
[2] V. R. Cadambe and S. A. Jafar, “Interference alignment and the degrees of freedom for the K user interference channel,” Submitted to
the IEEE Transactions on Information theory, 2007.
[3] N. Devroye, P. Mitran, and V. Tarokh, “Achievable rates in cognitive radio,” IEEE Trans. Inform. Theory, May 2006.
[4] P. Grover and A. Sahai, “On the need for knowledge of the phase in exploring known primary transmissions,” in Proc. of 2nd IEEE
International Symposium on New Frontiers in Dynamic Spectrum Access Networks, April 2007.
[5] A. Sahai, N. Hoven, and R. Tandra, “Some fundamental limits on cognitive radio,” in Forty-second Allerton Conference on
Communication, Control, and Computing, Monticello, IL, Oct. 2004.
 30

[6] N. Hoven, “On the feasibility of cognitive radio,” Master’s thesis, University of California, Berkeley, 2005.
[7] S. Shellhammer, “Numerical spectrum sensing requirements,” in IEEE 802.22-06/0088r0, June 2006.
[8] C. R. Stevenson, C. Cordeiro, E. Sofer, and G. Chouinard, “Functional requirements for the 802.22 WRAN standard,” in IEEE 802.22-
05/0007r46, September 2005.
[9] S. M. Kay, Fundamentals of statistical signal processing: Detection theory. Printice Hall PTR, 1998, vol. 2.
[10] H. V. Poor, An Introduction to Signal Detection and Estimation, 2nd ed. Springer, March 1998.
[11] A. Sonnenschein and P. M. Fishman, “Radiometric detection of spread-spectrum signals in noise of uncertain power,” IEEE Trans.
Aerosp. Electron. Syst., vol. 28, pp. 654–660, Jul. 1992.
[12] D. Cabric, A. Tkachenko, and R. W. Brodersen, “Experimental study of spectrum sensing based on energy detection and network
cooperation,” in the proceedings of IEEE MILCOM 2006.
[13] P. J. Huber, “A robust version of the probability ratio test,” The Annals of Mathematical Statistics, vol. 36, pp. 1753–1758, 1965.
[14] ——, Robust Statistics. Wiley-Interscience, February 1981.
[15] R. Durrett, Probability: Theory and Examples, 3rd ed. Belmont: Duxbury Press, 2004.
[16] S. Shellhammer and R. Tandra, “Performance of the power detector with noise uncertainty,” in IEEE
802.22-06/0134r0, July 2006. [Online]. Available: ”http://ieee802.org/22/Meeting documents/2006 July/22-06-0134-00-0000
Performance-of-the-power-detector-with-Noise-Uncertainty.ppt”
[17] ATSC Digital Television Standard, ATSC Std. A/53, 2007. [Online]. Available: ”http://www.atsc.org/standards.html”
[18] R. Tandra, “Fundamental limits on detection in low SNR,” Master’s thesis, University of California, Berkeley, 2005. [Online].
Available: http://www.eecs.berkeley.edu/∼tandra/pub1.htm
[19] S. Shellhammer and R. Tandra, “An evaluation of DTV pilot power detection,” in IEEE 802.22-06/0188r0, September 2006. [Online].
Available: ”http://ieee802.org/22/Meeting documents/2006 Sept/22-06-0188-00-0000-An-Evaluation-of-DTV-Pilot-Power-Detection.
ppt”
[20] S. Mathur and S. Shellhammer, “An evaluation of the PN sequence based detection of DTV signals in the
draft,” in IEEE 802.22-06/0189r0, September 2006. [Online]. Available: ”http://ieee802.org/22/Meeting documents/2006 Sept/
22-06-0189-00-0000-An-Evaluation-of-the-PN-Sequence-based-detection-of-DTV-signals.ppt”
[21] N. Agarwal, A. Sahai, and J. Tsitsiklis, “Narrowband noise mitigation in location-determining signal processing,” U.S. Patent 7 177 614,
February, 2007.
[22] W. A. Gardner, “The spectral correlation theory of cyclostationary time-series,” Signal Processing, vol. 11, pp. 13–36, July 1986.
[23] ——, Statistical Spectral Analysis: A Nonprobabilistic Theory. Englewood Cliffs, NJ: Prentice-Hall, 1987.
[24] ——, Introduction to Random Processes with Applications to Signals and Systems. New York: Macmillan, 1985.
[25] W. Gardner, “Signal interception: A unifying theoretical framework for feature detection,” IEEE Trans. Commun., vol. 36, pp. 897–906,
Aug. 1988.
[26] ——, “Signal interception: Performance advantages of cyclic-feature detectors,” IEEE Trans. Commun., vol. 40, pp. 149–159, Jan.
1992.
[27] R. Tandra and A. Sahai, “SNR walls for feature detectors,” in Proc. of 2nd IEEE International Symposium on New Frontiers in Dynamic
Spectrum Access Networks, April 2007, pp. 559–570.
[28] ——, “SNR walls for cyclostationary feature detectors,” In preparation for submission to IEEE Transactions on Signal Processing,
2007.
[29] T. M. Cover and J. A. Thomas, Elements of Information Theory. Wiley-Interscience, 1991.